Perfbook-1c 2023 06 11a

Is Parallel Programming Hard, And, If So,
What Can You Do About It?
Edited by:
Paul E. McKenney
Facebook
paulmck@kernel.org
June 11, 2023

Release v2023.06.11a
v2023.06.11a
ii
Legal Statement
This work represents the views of the editor and the authors and does not necessarily
represent the view of their respective employers.
Trademarks:
• IBM, z Systems, and PowerPC are trademarks or registered trademarks of Inter-
national Business Machines Corporation in the United States, other countries, or
both.
• Linux is a registered trademark of Linus Torvalds.
• Intel, Itanium, Intel Core, and Intel Xeon are trademarks of Intel Corporation or its
subsidiaries in the United States, other countries, or both.
• Arm is a registered trademark of Arm Limited (or its subsidiaries) in the US and/or
elsewhere.
• SPARC is a registered trademark of SPARC International, Inc. Products bearing
SPARC trademarks are based on an architecture developed by Sun Microsystems,
Inc.
• Other company, product, and service names may be trademarks or service marks
of such companies.
The non-source-code text and images in this document are provided under the terms
of the Creative Commons Attribution-Share Alike 3.0 United States license.1 In brief,
you may use the contents of this document for any purpose, personal, commercial, or
otherwise, so long as attribution to the authors is maintained. Likewise, the document
may be modified, and derivative works and translations made available, so long as
such modifications and derivations are offered to the public on equal terms as the
non-source-code text and images in the original document.
Source code is covered by various versions of the GPL.2 Some of this code is
GPLv2-only, as it derives from the Linux kernel, while other code is GPLv2-or-later.
See the comment headers of the individual source files within the CodeSamples directory
in the git archive3 for the exact licenses. If you are unsure of the license for a given
code fragment, you should assume GPLv2-only.
Combined work © 2005–2023 by Paul E. McKenney. Each individual contribution is
copyright by its contributor at the time of contribution, as recorded in the git archive.
1 https://creativecommons.org/licenses/by-sa/3.0/us/
2 https://www.gnu.org/licenses/gpl-2.0.html
3 git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/perfbook.
git
v2023.06.11a
Contents
1 How To Use This Book 1

1.1 Roadmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Quick Quizzes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Alternatives to This Book . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Sample Source Code . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Whose Book Is This? . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Introduction 9
2.1 Historic Parallel Programming Difficulties . . . . . . . . . . . . . . . 9
2.2 Parallel Programming Goals . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.2 Productivity . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.3 Generality . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Alternatives to Parallel Programming . . . . . . . . . . . . . . . . . . 17
2.3.1 Multiple Instances of a Sequential Application . . . . . . . . 17
2.3.2 Use Existing Parallel Software . . . . . . . . . . . . . . . . . 18
2.3.3 Performance Optimization . . . . . . . . . . . . . . . . . . . 18
2.4 What Makes Parallel Programming Hard? . . . . . . . . . . . . . . . 19
2.4.1 Work Partitioning . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.2 Parallel Access Control . . . . . . . . . . . . . . . . . . . . 20
2.4.3 Resource Partitioning and Replication . . . . . . . . . . . . . 21
2.4.4 Interacting With Hardware . . . . . . . . . . . . . . . . . . . 21
2.4.5 Composite Capabilities . . . . . . . . . . . . . . . . . . . . 22
2.4.6 How Do Languages and Environments Assist With These Tasks? 22
2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3 Hardware and its Habits 25

3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.1 Pipelined CPUs . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.2 Memory References . . . . . . . . . . . . . . . . . . . . . . 28
3.1.3 Atomic Operations . . . . . . . . . . . . . . . . . . . . . . . 29
3.1.4 Memory Barriers . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.5 Thermal Throttling . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.6 Cache Misses . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.7 I/O Operations . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Overheads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.1 Hardware System Architecture . . . . . . . . . . . . . . . . 34
3.2.2 Costs of Operations . . . . . . . . . . . . . . . . . . . . . . 35
iii
v2023.06.11a
iv CONTENTS
3.2.3 Hardware Optimizations . . . . . . . . . . . . . . . . . . . . 38

3.3 Hardware Free Lunch? . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.1 3D Integration . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3.2 Novel Materials and Processes . . . . . . . . . . . . . . . . . 41
3.3.3 Light, Not Electrons . . . . . . . . . . . . . . . . . . . . . . 41
3.3.4 Special-Purpose Accelerators . . . . . . . . . . . . . . . . . 41
3.3.5 Existing Parallel Software . . . . . . . . . . . . . . . . . . . 42
3.4 Software Design Implications . . . . . . . . . . . . . . . . . . . . . . 43
4 Tools of the Trade 45

4.1 Scripting Languages . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 POSIX Multiprocessing . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.2.1 POSIX Process Creation and Destruction . . . . . . . . . . . 46
4.2.2 POSIX Thread Creation and Destruction . . . . . . . . . . . 48
4.2.3 POSIX Locking . . . . . . . . . . . . . . . . . . . . . . . . 50
4.2.4 POSIX Reader-Writer Locking . . . . . . . . . . . . . . . . 53
4.2.5 Atomic Operations (GCC Classic) . . . . . . . . . . . . . . . 56
4.2.6 Atomic Operations (C11) . . . . . . . . . . . . . . . . . . . 57
4.2.7 Atomic Operations (Modern GCC) . . . . . . . . . . . . . . 58
4.2.8 Per-Thread Variables . . . . . . . . . . . . . . . . . . . . . . 58
4.3 Alternatives to POSIX Operations . . . . . . . . . . . . . . . . . . . 58
4.3.1 Organization and Initialization . . . . . . . . . . . . . . . . . 59
4.3.2 Thread Creation, Destruction, and Control . . . . . . . . . . 59
4.3.3 Locking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.3.4 Accessing Shared Variables . . . . . . . . . . . . . . . . . . 62
4.3.5 Atomic Operations . . . . . . . . . . . . . . . . . . . . . . . 72
4.3.6 Per-CPU Variables . . . . . . . . . . . . . . . . . . . . . . . 73
4.4 The Right Tool for the Job: How to Choose? . . . . . . . . . . . . . . 74
5 Counting 77
5.1 Why Isn’t Concurrent Counting Trivial? . . . . . . . . . . . . . . . . 78
5.2 Statistical Counters . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.2.1 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.2.2 Array-Based Implementation . . . . . . . . . . . . . . . . . 81
5.2.3 Per-Thread-Variable-Based Implementation . . . . . . . . . . 83
5.2.4 Eventually Consistent Implementation . . . . . . . . . . . . 84
5.2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3 Approximate Limit Counters . . . . . . . . . . . . . . . . . . . . . . 87
5.3.1 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.3.2 Simple Limit Counter Implementation . . . . . . . . . . . . 88
5.3.3 Simple Limit Counter Discussion . . . . . . . . . . . . . . . 94
5.3.4 Approximate Limit Counter Implementation . . . . . . . . . 94
5.3.5 Approximate Limit Counter Discussion . . . . . . . . . . . . 95
5.4 Exact Limit Counters . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.4.1 Atomic Limit Counter Implementation . . . . . . . . . . . . 95
5.4.2 Atomic Limit Counter Discussion . . . . . . . . . . . . . . . 100
5.4.3 Signal-Theft Limit Counter Design . . . . . . . . . . . . . . 101
5.4.4 Signal-Theft Limit Counter Implementation . . . . . . . . . . 102
5.4.5 Signal-Theft Limit Counter Discussion . . . . . . . . . . . . 106
5.4.6 Applying Exact Limit Counters . . . . . . . . . . . . . . . . 107
v2023.06.11a
CONTENTS v
5.5 Parallel Counting Discussion . . . . . . . . . . . . . . . . . . . . . . 108

5.5.1 Parallel Counting Validation . . . . . . . . . . . . . . . . . . 108
5.5.2 Parallel Counting Performance . . . . . . . . . . . . . . . . 109
5.5.3 Parallel Counting Specializations . . . . . . . . . . . . . . . 110
5.5.4 Parallel Counting Lessons . . . . . . . . . . . . . . . . . . . 110
6 Partitioning and Synchronization Design 113

6.1 Partitioning Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.1.1 Dining Philosophers Problem . . . . . . . . . . . . . . . . . 114
6.1.2 Double-Ended Queue . . . . . . . . . . . . . . . . . . . . . 117
6.1.3 Partitioning Example Discussion . . . . . . . . . . . . . . . 126
6.2 Design Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.3 Synchronization Granularity . . . . . . . . . . . . . . . . . . . . . . 129
6.3.1 Sequential Program . . . . . . . . . . . . . . . . . . . . . . 129
6.3.2 Code Locking . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.3.3 Data Locking . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.3.4 Data Ownership . . . . . . . . . . . . . . . . . . . . . . . . 135
6.3.5 Locking Granularity and Performance . . . . . . . . . . . . . 136
6.4 Parallel Fastpath . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.4.1 Reader/Writer Locking . . . . . . . . . . . . . . . . . . . . . 140
6.4.2 Hierarchical Locking . . . . . . . . . . . . . . . . . . . . . . 140
6.4.3 Resource Allocator Caches . . . . . . . . . . . . . . . . . . 141
6.5 Beyond Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.5.1 Work-Queue Parallel Maze Solver . . . . . . . . . . . . . . . 147
6.5.2 Alternative Parallel Maze Solver . . . . . . . . . . . . . . . . 150
6.5.3 Maze Validation . . . . . . . . . . . . . . . . . . . . . . . . 151
6.5.4 Performance Comparison I . . . . . . . . . . . . . . . . . . 152
6.5.5 Alternative Sequential Maze Solver . . . . . . . . . . . . . . 155
6.5.6 Performance Comparison II . . . . . . . . . . . . . . . . . . 156
6.5.7 Future Directions and Conclusions . . . . . . . . . . . . . . 157
6.6 Partitioning, Parallelism, and Optimization . . . . . . . . . . . . . . . 158
7 Locking 159
7.1 Staying Alive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
7.1.1 Deadlock . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
7.1.2 Livelock and Starvation . . . . . . . . . . . . . . . . . . . . 171
7.1.3 Unfairness . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
7.1.4 Inefficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
7.2 Types of Locks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.2.1 Exclusive Locks . . . . . . . . . . . . . . . . . . . . . . . . 175
7.2.2 Reader-Writer Locks . . . . . . . . . . . . . . . . . . . . . . 176
7.2.3 Beyond Reader-Writer Locks . . . . . . . . . . . . . . . . . 177
7.2.4 Scoped Locking . . . . . . . . . . . . . . . . . . . . . . . . 178
7.3 Locking Implementation Issues . . . . . . . . . . . . . . . . . . . . . 181
7.3.1 Sample Exclusive-Locking Implementation Based on Atomic
Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.3.2 Other Exclusive-Locking Implementations . . . . . . . . . . 182
7.4 Lock-Based Existence Guarantees . . . . . . . . . . . . . . . . . . . 184
7.5 Locking: Hero or Villain? . . . . . . . . . . . . . . . . . . . . . . . 186
7.5.1 Locking For Applications: Hero! . . . . . . . . . . . . . . . 187
v2023.06.11a
vi CONTENTS
7.5.2 Locking For Parallel Libraries: Just Another Tool . . . . . . 187

7.5.3 Locking For Parallelizing Sequential Libraries: Villain! . . . 191
7.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8 Data Ownership 195

8.1 Multiple Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8.2 Partial Data Ownership and pthreads . . . . . . . . . . . . . . . . . . 196
8.3 Function Shipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
8.4 Designated Thread . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
8.5 Privatization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
8.6 Other Uses of Data Ownership . . . . . . . . . . . . . . . . . . . . . 198
9 Deferred Processing 201

9.1 Running Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
9.2 Reference Counting . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
9.3 Hazard Pointers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
9.4 Sequence Locks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
9.5 Read-Copy Update (RCU) . . . . . . . . . . . . . . . . . . . . . . . 218
9.5.1 Introduction to RCU . . . . . . . . . . . . . . . . . . . . . . 219
9.5.2 RCU Fundamentals . . . . . . . . . . . . . . . . . . . . . . 228
9.5.3 RCU Linux-Kernel API . . . . . . . . . . . . . . . . . . . . 238
9.5.4 RCU Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
9.5.5 RCU Related Work . . . . . . . . . . . . . . . . . . . . . . . 277
9.6 Which to Choose? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
9.6.1 Which to Choose? (Overview) . . . . . . . . . . . . . . . . . 281
9.6.2 Which to Choose? (Details) . . . . . . . . . . . . . . . . . . 282
9.6.3 Which to Choose? (Production Use) . . . . . . . . . . . . . . 285
9.7 What About Updates? . . . . . . . . . . . . . . . . . . . . . . . . . . 287
10 Data Structures 289

10.1 Motivating Application . . . . . . . . . . . . . . . . . . . . . . . . . 290
10.2 Partitionable Data Structures . . . . . . . . . . . . . . . . . . . . . . 290
10.2.1 Hash-Table Design . . . . . . . . . . . . . . . . . . . . . . . 290
10.2.2 Hash-Table Implementation . . . . . . . . . . . . . . . . . . 291
10.2.3 Hash-Table Performance . . . . . . . . . . . . . . . . . . . . 293
10.3 Read-Mostly Data Structures . . . . . . . . . . . . . . . . . . . . . . 295
10.3.1 RCU-Protected Hash Table Implementation . . . . . . . . . . 296
10.3.2 RCU-Protected Hash Table Validation . . . . . . . . . . . . . 297
10.3.3 RCU-Protected Hash Table Performance . . . . . . . . . . . 298
10.3.4 RCU-Protected Hash Table Discussion . . . . . . . . . . . . 303
10.4 Non-Partitionable Data Structures . . . . . . . . . . . . . . . . . . . 304
10.4.1 Resizable Hash Table Design . . . . . . . . . . . . . . . . . 304
10.4.2 Resizable Hash Table Implementation . . . . . . . . . . . . . 306
10.4.3 Resizable Hash Table Discussion . . . . . . . . . . . . . . . 312
10.4.4 Other Resizable Hash Tables . . . . . . . . . . . . . . . . . . 314
10.5 Other Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . 317
10.6 Micro-Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
10.6.1 Specialization . . . . . . . . . . . . . . . . . . . . . . . . . 318
10.6.2 Bits and Bytes . . . . . . . . . . . . . . . . . . . . . . . . . 319
10.6.3 Hardware Considerations . . . . . . . . . . . . . . . . . . . 320
v2023.06.11a
CONTENTS vii
10.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
11 Validation 323
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
11.1.1 Where Do Bugs Come From? . . . . . . . . . . . . . . . . . 324
11.1.2 Required Mindset . . . . . . . . . . . . . . . . . . . . . . . 325
11.1.3 When Should Validation Start? . . . . . . . . . . . . . . . . 328
11.1.4 The Open Source Way . . . . . . . . . . . . . . . . . . . . . 329
11.2 Tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
11.3 Assertions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
11.4 Static Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
11.5 Code Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
11.5.1 Inspection . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
11.5.2 Walkthroughs . . . . . . . . . . . . . . . . . . . . . . . . . . 334
11.5.3 Self-Inspection . . . . . . . . . . . . . . . . . . . . . . . . . 334
11.6 Probability and Heisenbugs . . . . . . . . . . . . . . . . . . . . . . . 336
11.6.1 Statistics for Discrete Testing . . . . . . . . . . . . . . . . . 337
11.6.2 Statistics Abuse for Discrete Testing . . . . . . . . . . . . . . 339
11.6.3 Statistics for Continuous Testing . . . . . . . . . . . . . . . . 339
11.6.4 Hunting Heisenbugs . . . . . . . . . . . . . . . . . . . . . . 341
11.7 Performance Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 346
11.7.1 Benchmarking . . . . . . . . . . . . . . . . . . . . . . . . . 347
11.7.2 Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
11.7.3 Differential Profiling . . . . . . . . . . . . . . . . . . . . . . 348
11.7.4 Microbenchmarking . . . . . . . . . . . . . . . . . . . . . . 348
11.7.5 Isolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
11.7.6 Detecting Interference . . . . . . . . . . . . . . . . . . . . . 350
11.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
12 Formal Verification 357

12.1 State-Space Search . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
12.1.1 Promela and Spin . . . . . . . . . . . . . . . . . . . . . . . 358
12.1.2 How to Use Promela . . . . . . . . . . . . . . . . . . . . . . 362
12.1.3 Promela Example: Locking . . . . . . . . . . . . . . . . . . 365
12.1.4 Promela Example: QRCU . . . . . . . . . . . . . . . . . . . 368
12.1.5 Promela Parable: dynticks and Preemptible RCU . . . . . . . 376
12.1.6 Validating Preemptible RCU and dynticks . . . . . . . . . . . 381
12.2 Special-Purpose State-Space Search . . . . . . . . . . . . . . . . . . 402
12.2.1 Anatomy of a Litmus Test . . . . . . . . . . . . . . . . . . . 403
12.2.2 What Does This Litmus Test Mean? . . . . . . . . . . . . . . 404
12.2.3 Running a Litmus Test . . . . . . . . . . . . . . . . . . . . . 405
12.2.4 PPCMEM Discussion . . . . . . . . . . . . . . . . . . . . . 406
12.3 Axiomatic Approaches . . . . . . . . . . . . . . . . . . . . . . . . . 407
12.3.1 Axiomatic Approaches and Locking . . . . . . . . . . . . . . 409
12.3.2 Axiomatic Approaches and RCU . . . . . . . . . . . . . . . 410
12.4 SAT Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
12.5 Stateless Model Checkers . . . . . . . . . . . . . . . . . . . . . . . . 414
12.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
12.7 Choosing a Validation Plan . . . . . . . . . . . . . . . . . . . . . . . 417
v2023.06.11a
viii CONTENTS
13 Putting It All Together 421

13.1 Counter Conundrums . . . . . . . . . . . . . . . . . . . . . . . . . . 421
13.1.1 Counting Updates . . . . . . . . . . . . . . . . . . . . . . . 421
13.1.2 Counting Lookups . . . . . . . . . . . . . . . . . . . . . . . 421
13.2 Refurbish Reference Counting . . . . . . . . . . . . . . . . . . . . . 422
13.2.1 Implementation of Reference-Counting Categories . . . . . . 424
13.2.2 Counter Optimizations . . . . . . . . . . . . . . . . . . . . . 428
13.3 Hazard-Pointer Helpers . . . . . . . . . . . . . . . . . . . . . . . . . 429
13.3.1 Scalable Reference Count . . . . . . . . . . . . . . . . . . . 429
13.3.2 Long-Duration Accesses . . . . . . . . . . . . . . . . . . . . 429
13.4 Sequence-Locking Specials . . . . . . . . . . . . . . . . . . . . . . . 430
13.4.1 Dueling Sequence Locks . . . . . . . . . . . . . . . . . . . . 430
13.4.2 Correlated Data Elements . . . . . . . . . . . . . . . . . . . 430
13.4.3 Atomic Move . . . . . . . . . . . . . . . . . . . . . . . . . . 431
13.4.4 Upgrade to Writer . . . . . . . . . . . . . . . . . . . . . . . 433
13.5 RCU Rescues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
13.5.1 RCU and Per-Thread-Variable-Based Statistical Counters . . . 433
13.5.2 RCU and Counters for Removable I/O Devices . . . . . . . . 436
13.5.3 Array and Length . . . . . . . . . . . . . . . . . . . . . . . 437
13.5.4 Correlated Fields . . . . . . . . . . . . . . . . . . . . . . . . 438
13.5.5 Update-Friendly Traversal . . . . . . . . . . . . . . . . . . . 439
13.5.6 Scalable Reference Count Two . . . . . . . . . . . . . . . . . 439
13.5.7 Retriggered Grace Periods . . . . . . . . . . . . . . . . . . . 440
13.5.8 Long-Duration Accesses Two . . . . . . . . . . . . . . . . . 441
14 Advanced Synchronization 445

14.1 Avoiding Locks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
14.2 Non-Blocking Synchronization . . . . . . . . . . . . . . . . . . . . . 446
14.2.1 Simple NBS . . . . . . . . . . . . . . . . . . . . . . . . . . 447
14.2.2 Applicability of NBS Benefits . . . . . . . . . . . . . . . . . 450
14.2.3 NBS Discussion . . . . . . . . . . . . . . . . . . . . . . . . 454
14.3 Parallel Real-Time Computing . . . . . . . . . . . . . . . . . . . . . 455
14.3.1 What is Real-Time Computing? . . . . . . . . . . . . . . . . 455
14.3.2 Who Needs Real-Time? . . . . . . . . . . . . . . . . . . . . 461
14.3.3 Who Needs Parallel Real-Time? . . . . . . . . . . . . . . . . 462
14.3.4 Implementing Parallel Real-Time Systems . . . . . . . . . . 463
14.3.5 Implementing Parallel Real-Time Operating Systems . . . . . 465
14.3.6 Implementing Parallel Real-Time Applications . . . . . . . . 478
14.3.7 Real Time vs. Real Fast: How to Choose? . . . . . . . . . . . 483
15 Advanced Synchronization: Memory Ordering 485

15.1 Ordering: Why and How? . . . . . . . . . . . . . . . . . . . . . . . 485
15.1.1 Why Hardware Misordering? . . . . . . . . . . . . . . . . . 487
15.1.2 How to Force Ordering? . . . . . . . . . . . . . . . . . . . . 490
15.1.3 Basic Rules of Thumb . . . . . . . . . . . . . . . . . . . . . 492
15.2 Tricks and Traps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
15.2.1 Variables With Multiple Values . . . . . . . . . . . . . . . . 495
15.2.2 Memory-Reference Reordering . . . . . . . . . . . . . . . . 498
15.2.3 Address Dependencies . . . . . . . . . . . . . . . . . . . . . 501
15.2.4 Data Dependencies . . . . . . . . . . . . . . . . . . . . . . . 503
v2023.06.11a
CONTENTS ix
15.2.5 Control Dependencies . . . . . . . . . . . . . . . . . . . . . 504

15.2.6 Cache Coherence . . . . . . . . . . . . . . . . . . . . . . . . 505
15.2.7 Multicopy Atomicity . . . . . . . . . . . . . . . . . . . . . . 506
15.2.8 A Counter-Intuitive Case Study . . . . . . . . . . . . . . . . 519
15.3 Compile-Time Consternation . . . . . . . . . . . . . . . . . . . . . . 523
15.3.1 Memory-Reference Restrictions . . . . . . . . . . . . . . . . 523
15.3.2 Address- and Data-Dependency Difficulties . . . . . . . . . . 524
15.3.3 Control-Dependency Calamities . . . . . . . . . . . . . . . . 528
15.4 Higher-Level Primitives . . . . . . . . . . . . . . . . . . . . . . . . . 533
15.4.1 Memory Allocation . . . . . . . . . . . . . . . . . . . . . . 533
15.4.2 Locking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
15.4.3 RCU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539
15.4.4 Higher-Level Primitives: Discussion . . . . . . . . . . . . . 549
15.5 Hardware Specifics . . . . . . . . . . . . . . . . . . . . . . . . . . . 550
15.5.1 Alpha . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552
15.5.2 Armv7-A/R . . . . . . . . . . . . . . . . . . . . . . . . . . . 555
15.5.3 Armv8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 556
15.5.4 Itanium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
15.5.5 MIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 558
15.5.6 POWER / PowerPC . . . . . . . . . . . . . . . . . . . . . . 559
15.5.7 SPARC TSO . . . . . . . . . . . . . . . . . . . . . . . . . . 560
15.5.8 x86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561
15.5.9 z Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562
15.5.10 Hardware Specifics: Discussion . . . . . . . . . . . . . . . . 562
15.6 Memory-Model Intuitions . . . . . . . . . . . . . . . . . . . . . . . 562
15.6.1 Transitive Intuitions . . . . . . . . . . . . . . . . . . . . . . 563
15.6.2 Rules of Thumb . . . . . . . . . . . . . . . . . . . . . . . . 567
16 Ease of Use 569

16.1 What is Easy? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
16.2 Rusty Scale for API Design . . . . . . . . . . . . . . . . . . . . . . . 570
16.3 Shaving the Mandelbrot Set . . . . . . . . . . . . . . . . . . . . . . . 571
17 Conflicting Visions of the Future 575

17.1 The Future of CPU Technology Ain’t What it Used to Be . . . . . . . 575
17.1.1 Uniprocessor Über Alles . . . . . . . . . . . . . . . . . . . . 578
17.1.2 Multithreaded Mania . . . . . . . . . . . . . . . . . . . . . . 579
17.1.3 More of the Same . . . . . . . . . . . . . . . . . . . . . . . 579
17.1.4 Crash Dummies Slamming into the Memory Wall . . . . . . 579
17.1.5 Astounding Accelerators . . . . . . . . . . . . . . . . . . . . 582
17.2 Transactional Memory . . . . . . . . . . . . . . . . . . . . . . . . . 582
17.2.1 Outside World . . . . . . . . . . . . . . . . . . . . . . . . . 583
17.2.2 Process Modification . . . . . . . . . . . . . . . . . . . . . . 587
17.2.3 Synchronization . . . . . . . . . . . . . . . . . . . . . . . . 592
17.2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 596
17.3 Hardware Transactional Memory . . . . . . . . . . . . . . . . . . . . 599
17.3.1 HTM Benefits WRT Locking . . . . . . . . . . . . . . . . . 600
17.3.2 HTM Weaknesses WRT Locking . . . . . . . . . . . . . . . 601
17.3.3 HTM Weaknesses WRT Locking When Augmented . . . . . 607
17.3.4 Where Does HTM Best Fit In? . . . . . . . . . . . . . . . . . 610
v2023.06.11a
x CONTENTS
17.3.5 Potential Game Changers . . . . . . . . . . . . . . . . . . . 612

17.3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 615
17.4 Formal Regression Testing? . . . . . . . . . . . . . . . . . . . . . . . 616
17.4.1 Automatic Translation . . . . . . . . . . . . . . . . . . . . . 616
17.4.2 Environment . . . . . . . . . . . . . . . . . . . . . . . . . . 617
17.4.3 Overhead . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
17.4.4 Locate Bugs . . . . . . . . . . . . . . . . . . . . . . . . . . 619
17.4.5 Minimal Scaffolding . . . . . . . . . . . . . . . . . . . . . . 620
17.4.6 Relevant Bugs . . . . . . . . . . . . . . . . . . . . . . . . . 621
17.4.7 Formal Regression Scorecard . . . . . . . . . . . . . . . . . 621
17.5 Functional Programming for Parallelism . . . . . . . . . . . . . . . . 623
17.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624
18 Looking Forward and Back 625
A Important Questions 629

A.1 Why Aren’t Parallel Programs Always Faster? . . . . . . . . . . . . . 629
A.2 Why Not Remove Locking? . . . . . . . . . . . . . . . . . . . . . . . 630
A.3 What Time Is It? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 630
A.4 What Does “After” Mean? . . . . . . . . . . . . . . . . . . . . . . . 632
A.5 How Much Ordering Is Needed? . . . . . . . . . . . . . . . . . . . . 635
A.5.1 Where is the Defining Data? . . . . . . . . . . . . . . . . . . 636
A.5.2 Consistent Data Used Consistently? . . . . . . . . . . . . . . 637
A.5.3 Is the Problem Partitionable? . . . . . . . . . . . . . . . . . 637
A.5.4 None of the Above? . . . . . . . . . . . . . . . . . . . . . . 637
A.6 What is the Difference Between “Concurrent” and “Parallel”? . . . . . 638
A.7 Why Is Software Buggy? . . . . . . . . . . . . . . . . . . . . . . . . 639
B “Toy” RCU Implementations 641

B.1 Lock-Based RCU . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
B.2 Per-Thread Lock-Based RCU . . . . . . . . . . . . . . . . . . . . . . 642
B.3 Simple Counter-Based RCU . . . . . . . . . . . . . . . . . . . . . . 643
B.4 Starvation-Free Counter-Based RCU . . . . . . . . . . . . . . . . . . 644
B.5 Scalable Counter-Based RCU . . . . . . . . . . . . . . . . . . . . . . 647
B.6 Scalable Counter-Based RCU With Shared Grace Periods . . . . . . . 650
B.7 RCU Based on Free-Running Counter . . . . . . . . . . . . . . . . . 652
B.8 Nestable RCU Based on Free-Running Counter . . . . . . . . . . . . 654
B.9 RCU Based on Quiescent States . . . . . . . . . . . . . . . . . . . . 656
B.10 Summary of Toy RCU Implementations . . . . . . . . . . . . . . . . 659
C Why Memory Barriers? 661

C.1 Cache Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 661
C.2 Cache-Coherence Protocols . . . . . . . . . . . . . . . . . . . . . . . 664
C.2.1 MESI States . . . . . . . . . . . . . . . . . . . . . . . . . . 664
C.2.2 MESI Protocol Messages . . . . . . . . . . . . . . . . . . . 664
C.2.3 MESI State Diagram . . . . . . . . . . . . . . . . . . . . . . 665
C.2.4 MESI Protocol Example . . . . . . . . . . . . . . . . . . . . 667
C.3 Stores Result in Unnecessary Stalls . . . . . . . . . . . . . . . . . . . 668
C.3.1 Store Buffers . . . . . . . . . . . . . . . . . . . . . . . . . . 668
C.3.2 Store Forwarding . . . . . . . . . . . . . . . . . . . . . . . . 669
v2023.06.11a
CONTENTS xi
C.3.3 Store Buffers and Memory Barriers . . . . . . . . . . . . . . 671

C.4 Store Sequences Result in Unnecessary Stalls . . . . . . . . . . . . . 674
C.4.1 Invalidate Queues . . . . . . . . . . . . . . . . . . . . . . . 674
C.4.2 Invalidate Queues and Invalidate Acknowledge . . . . . . . . 674
C.4.3 Invalidate Queues and Memory Barriers . . . . . . . . . . . 675
C.5 Read and Write Memory Barriers . . . . . . . . . . . . . . . . . . . 678
C.6 Example Memory-Barrier Sequences . . . . . . . . . . . . . . . . . . 679
C.6.1 Ordering-Hostile Architecture . . . . . . . . . . . . . . . . . 679
C.6.2 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 680
C.6.3 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 681
C.6.4 Example 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 681
C.7 Are Memory Barriers Forever? . . . . . . . . . . . . . . . . . . . . . 682
C.8 Advice to Hardware Designers . . . . . . . . . . . . . . . . . . . . . 682
D Style Guide 685

D.1 Paul’s Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . 685
D.2 NIST Style Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . 686
D.2.1 Unit Symbol . . . . . . . . . . . . . . . . . . . . . . . . . . 686
D.2.2 NIST Guide Yet To Be Followed . . . . . . . . . . . . . . . 688
D.3 LATEX Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . 688
D.3.1 Monospace Font . . . . . . . . . . . . . . . . . . . . . . . . 688
D.3.2 Cross-reference . . . . . . . . . . . . . . . . . . . . . . . . . 695
D.3.3 Non Breakable Spaces . . . . . . . . . . . . . . . . . . . . . 695
D.3.4 Hyphenation and Dashes . . . . . . . . . . . . . . . . . . . . 696
D.3.5 Punctuation . . . . . . . . . . . . . . . . . . . . . . . . . . . 697
D.3.6 Floating Object Format . . . . . . . . . . . . . . . . . . . . 699
D.3.7 Improvement Candidates . . . . . . . . . . . . . . . . . . . . 699
E Answers to Quick Quizzes 705

E.1 How To Use This Book . . . . . . . . . . . . . . . . . . . . . . . . . 705
E.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707
E.3 Hardware and its Habits . . . . . . . . . . . . . . . . . . . . . . . . . 713
E.4 Tools of the Trade . . . . . . . . . . . . . . . . . . . . . . . . . . . . 719
E.5 Counting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 729
E.6 Partitioning and Synchronization Design . . . . . . . . . . . . . . . . 750
E.7 Locking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 759
E.8 Data Ownership . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 769
E.9 Deferred Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 772
E.10 Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 798
E.11 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 805
E.12 Formal Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . 815
E.13 Putting It All Together . . . . . . . . . . . . . . . . . . . . . . . . . 826
E.14 Advanced Synchronization . . . . . . . . . . . . . . . . . . . . . . . 832
E.15 Advanced Synchronization: Memory Ordering . . . . . . . . . . . . 836
E.16 Ease of Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 854
E.17 Conflicting Visions of the Future . . . . . . . . . . . . . . . . . . . . 854
E.18 Important Questions . . . . . . . . . . . . . . . . . . . . . . . . . . 863
E.19 “Toy” RCU Implementations . . . . . . . . . . . . . . . . . . . . . . 864
E.20 Why Memory Barriers? . . . . . . . . . . . . . . . . . . . . . . . . . 873
v2023.06.11a
xii CONTENTS
Glossary 881
Bibliography 893
Credits 945
LATEX Advisor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 945
Reviewers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 945
Machine Owners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 946
Original Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 946
Figure Credits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 947
Other Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 949
Acronyms 951
Index 953
API Index 957
v2023.06.11a
If you would only recognize that life is hard, things
would be so much easier for you.
Louis D. Brandeis
Chapter 1
How To Use This Book
The purpose of this book is to help you program shared-memory parallel systems
without risking your sanity.1 Nevertheless, you should think of the information in this
book as a foundation on which to build, rather than as a completed cathedral. Your
mission, if you choose to accept, is to help make further progress in the exciting field of
parallel programming—progress that will in time render this book obsolete.
Parallel programming in the 21st century is no longer focused solely on science,
research, and grand-challenge projects. And this is all to the good, because it means
that parallel programming is becoming an engineering discipline. Therefore, as befits
an engineering discipline, this book examines specific parallel-programming tasks and
describes how to approach them. In some surprisingly common cases, these tasks can
be automated.
This book is written in the hope that presenting the engineering discipline underlying
successful parallel-programming projects will free a new generation of parallel hackers
from the need to slowly and painstakingly reinvent old wheels, enabling them to instead
focus their energy and creativity on new frontiers. However, what you get from this
book will be determined by what you put into it. It is hoped that simply reading this
book will be helpful, and that working the Quick Quizzes will be even more helpful.
However, the best results come from applying the techniques taught in this book to
real-life problems. As always, practice makes perfect.
But no matter how you approach it, we sincerely hope that parallel programming
brings you at least as much fun, excitement, and challenge that it has brought to us!
1 Or, perhaps more accurately, without much greater risk to your sanity than that incurred
by non-parallel programming. Which, come to think of it, might not be saying all that much.
v2023.06.11a
2 CHAPTER 1. HOW TO USE THIS BOOK
1.1 Roadmap
Cat: Where are you going?

Alice: Which way should I go?
Cat: That depends on where you are going.
Alice: I don’t know.
Cat: Then it doesn’t matter which way you go.
Lewis Carroll, Alice in Wonderland
This book is a handbook of widely applicable and heavily used design techniques, rather
than a collection of optimal algorithms with tiny areas of applicability. You are currently
reading Chapter 1, but you knew that already. Chapter 2 gives a high-level overview of
parallel programming.
Chapter 3 introduces shared-memory parallel hardware. After all, it is difficult to
write good parallel code unless you understand the underlying hardware. Because
hardware constantly evolves, this chapter will always be out of date. We will nevertheless
do our best to keep up. Chapter 4 then provides a very brief overview of common
shared-memory parallel-programming primitives.
Chapter 5 takes an in-depth look at parallelizing one of the simplest problems
imaginable, namely counting. Because almost everyone has an excellent grasp of
counting, this chapter is able to delve into many important parallel-programming issues
without the distractions of more-typical computer-science problems. My impression is
that this chapter has seen the greatest use in parallel-programming coursework.
Chapter 6 introduces a number of design-level methods of addressing the issues
identified in Chapter 5. It turns out that it is important to address parallelism at the
design level when feasible: To paraphrase Dĳkstra [Dĳ68], “retrofitted parallelism
considered grossly suboptimal” [McK12c].
The next three chapters examine three important approaches to synchronization.
Chapter 7 covers locking, which is still not only the workhorse of production-quality
parallel programming, but is also widely considered to be parallel programming’s worst
villain. Chapter 8 gives a brief overview of data ownership, an often overlooked but
remarkably pervasive and powerful approach. Finally, Chapter 9 introduces a number
of deferred-processing mechanisms, including reference counting, hazard pointers,
sequence locking, and RCU.
Chapter 10 applies the lessons of previous chapters to hash tables, which are heavily
used due to their excellent partitionability, which (usually) leads to excellent performance
and scalability.
As many have learned to their sorrow, parallel programming without validation is a
sure path to abject failure. Chapter 11 covers various forms of testing. It is of course
impossible to test reliability into your program after the fact, so Chapter 12 follows up
with a brief overview of a couple of practical approaches to formal verification.
Chapter 13 contains a series of moderate-sized parallel programming problems. The
difficulty of these problems vary, but should be appropriate for someone who has
mastered the material in the previous chapters.
Chapter 14 looks at advanced synchronization methods, including non-blocking
synchronization and parallel real-time computing, while Chapter 15 covers the advanced
topic of memory ordering. Chapter 16 follows up with some ease-of-use advice. Chap-
ter 17 looks at a few possible future directions, including shared-memory parallel system
v2023.06.11a
1.2. QUICK QUIZZES 3
design, software and hardware transactional memory, and functional programming for
parallelism. Finally, Chapter 18 reviews the material in this book and its origins.
This chapter is followed by a number of appendices. The most popular of these
appears to be Appendix C, which delves even further into memory ordering. Appendix E
contains the answers to the infamous Quick Quizzes, which are discussed in the next
section.
1.2 Quick Quizzes

Undertake something difficult, otherwise you will
never grow.
Abbreviated from Ronald E. Osburn
“Quick quizzes” appear throughout this book, and the answers may be found in
Appendix E starting on page 705. Some of them are based on material in which that
quick quiz appears, but others require you to think beyond that section, and, in some
cases, beyond the realm of current knowledge. As with most endeavors, what you get
out of this book is largely determined by what you are willing to put into it. Therefore,
readers who make a genuine effort to solve a quiz before looking at the answer find their
effort repaid handsomely with increased understanding of parallel programming.
Quick Quiz 1.1: Where are the answers to the Quick Quizzes found?
Quick Quiz 1.2: Some of the Quick Quiz questions seem to be from the viewpoint of the
reader rather than the author. Is that really the intent?
Quick Quiz 1.3: These Quick Quizzes are just not my cup of tea. What can I do about it?
In short, if you need a deep understanding of the material, then you should invest
some time into answering the Quick Quizzes. Don’t get me wrong, passively reading
the material can be quite valuable, but gaining full problem-solving capability really
does require that you practice solving problems. Similarly, gaining full code-production
capability really does require that you practice producing code.
Quick Quiz 1.4: If passively reading this book doesn’t get me full problem-solving and
code-production capabilities, what on earth is the point???
I learned this the hard way during coursework for my late-in-life Ph.D. I was studying
a familiar topic, and was surprised at how few of the chapter’s exercises I could answer
off the top of my head.2 Forcing myself to answer the questions greatly increased
my retention of the material. So with these Quick Quizzes I am not asking you to do
anything that I have not been doing myself.
Finally, the most common learning disability is thinking that you already understand
the material at hand. The quick quizzes can be an extremely effective cure.
2 So I suppose that it was just as well that my professors refused to let me waive that
class!
v2023.06.11a
1.3 Alternatives to This Book

Between two evils I always pick the one I never tried
before.
Mae West
As Knuth learned the hard way, if you want your book to be finite, it must be focused.
This book focuses on shared-memory parallel programming, with an emphasis on
software that lives near the bottom of the software stack, such as operating-system
kernels, parallel data-management systems, low-level libraries, and the like. The
programming language used by this book is C.
If you are interested in other aspects of parallelism, you might well be better served
by some other book. Fortunately, there are many alternatives available to you:
1. If you prefer a more academic and rigorous treatment of parallel programming,

you might like Herlihy’s and Shavit’s textbook [HS08, HSLS20]. This book starts
with an interesting combination of low-level primitives at high levels of abstraction
from the hardware, and works its way through locking and simple data structures
including lists, queues, hash tables, and counters, culminating with transactional
memory, all in Java. Michael Scott’s textbook [Sco13] approaches similar material
with more of a software-engineering focus, and, as far as I know, is the first formally
published academic textbook with section devoted to RCU.
Herlihy, Shavit, Luchangco, and Spear did catch up in their second edi-
tion [HSLS20] by adding short sections on hazard pointers and on RCU, with the
latter in the guise of EBR.3 They also include a brief history of both, albeit with
an abbreviated history of RCU that picks up almost a year after it was accepted
into the Linux kernel and more than 20 years after Kung’s and Lehman’s landmark
paper [KL80]. Those wishing a deeper view of the history may find it in this
book’s Section 9.5.5.
However, readers who might otherwise suspect a hostile attitude towards RCU
on the part of this textbook’s first author should refer to the last full sentence on
the first page of one of his papers [BGHZ16]. This sentence reads “QSBR [a
particular class of RCU implementations] is fast and can be applied to virtually
any data structure.” These are clearly not the words of someone who is hostile
towards RCU.
2. If you would like an academic treatment of parallel programming from a program-

ming-language-pragmatics viewpoint, you might be interested in the concurrency
chapter from Scott’s textbook [Sco06, Sco15] on programming-language pragmat-
ics.
3. If you are interested in an object-oriented patternist treatment of parallel pro-
gramming focussing on C++, you might try Volumes 2 and 4 of Schmidt’s POSA
series [SSRB00, BHS07]. Volume 4 in particular has some interesting chapters
applying this work to a warehouse application. The realism of this example is
attested to by the section entitled “Partitioning the Big Ball of Mud”, in which
3 Albeit an implementation that contains a reader-preemption bug noted by Richard
Bornat.
v2023.06.11a
1.3. ALTERNATIVES TO THIS BOOK 5
the problems inherent in parallelism often take a back seat to getting one’s head
around a real-world application.
4. If you want to work with Linux-kernel device drivers, then Corbet’s, Rubini’s, and
Kroah-Hartman’s “Linux Device Drivers” [CRKH05] is indispensable, as is the
Linux Weekly News web site (https://lwn.net/). There is a large number of
books and resources on the more general topic of Linux kernel internals.
5. If your primary focus is scientific and technical computing, and you prefer a
patternist approach, you might try Mattson et al.’s textbook [MSM05]. It covers
Java, C/C++, OpenMP, and MPI. Its patterns are admirably focused first on design,
then on implementation.
6. If your primary focus is scientific and technical computing, and you are interested
in GPUs, CUDA, and MPI, you might check out Norm Matloff’s “Programming
on Parallel Machines” [Mat17]. Of course, the GPU vendors have quite a bit of
additional information [AMD20, Zel11, NVi17a, NVi17b].
7. If you are interested in POSIX Threads, you might take a look at David R. Butenhof’s
book [But97]. In addition, W. Richard Stevens’s book [Ste92, Ste13] covers UNIX
and POSIX, and Stewart Weiss’s lecture notes [Wei13] provide an thorough and
accessible introduction with a good set of examples.
8. If you are interested in C++11, you might like Anthony Williams’s “C++ Concur-
rency in Action: Practical Multithreading” [Wil12, Wil19].
9. If you are interested in C++, but in a Windows environment, you might try Herb
Sutter’s “Effective Concurrency” series in Dr. Dobbs Journal [Sut08]. This series
does a reasonable job of presenting a commonsense approach to parallelism.
10. If you want to try out Intel Threading Building Blocks, then perhaps James
Reinders’s book [Rei07] is what you are looking for.
11. Those interested in learning how various types of multi-processor hardware cache
organizations affect the implementation of kernel internals should take a look at
Curt Schimmel’s classic treatment of this subject [Sch94].
12. If you are looking for a hardware view, Hennessy’s and Patterson’s classic
textbook [HP17, HP11] is well worth a read. A “Readers Digest” version of this
tome geared for scientific and technical workloads (bashing big arrays) may be
found in Andrew Chien’s textbook [Chi22]. If you are looking for an academic
textbook on memory ordering from a more hardware-centric viewpoint, that of
Daniel Sorin et al. [SHW11, NSHW20] is highly recommended. For a memory-
ordering tutorial from a Linux-kernel viewpoint, Paolo Bonzini’s LWN series is a
good place to start [Bon21a, Bon21e, Bon21c, Bon21b, Bon21d, Bon21f].
13. Those wishing to learn about the Rust language’s support for low-level concurrency
should refer to Mara Bos’s book [Bos23].
14. Finally, those using Java might be well-served by Doug Lea’s textbooks [Lea97,
GPB+ 07].
However, if you are interested in principles of parallel design for low-level software,
especially software written in C, read on!
v2023.06.11a
Listing 1.1: Creating an Up-To-Date PDF

git clone git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/perfbook.git
cd perfbook
# You may need to install a font. See item 1 in FAQ.txt.
make # -jN for parallel build
evince perfbook.pdf & # Two-column version
make perfbook-1c.pdf
evince perfbook-1c.pdf & # One-column version for e-readers
make help # Display other build options
1.4 Sample Source Code
Use the source, Luke!

Unknown Star Wars fan
This book discusses its fair share of source code, and in many cases this source code
may be found in the CodeSamples directory of this book’s git tree. For example, on
UNIX systems, you should be able to type the following:
find CodeSamples -name rcu_rcpls.c -print
This command will locate the file rcu_rcpls.c, which is called out in Appendix B.
Non-UNIX systems have their own well-known ways of locating files by filename.
1.5 Whose Book Is This?
If you become a teacher, by your pupils you’ll be

taught.
Oscar Hammerstein II
As the cover says, the editor is one Paul E. McKenney. However, the editor does accept
contributions via the perfbook@vger.kernel.org email list. These contributions
can be in pretty much any form, with popular approaches including text emails, patches
against the book’s LATEX source, and even git pull requests. Use whatever form
works best for you.
To create patches or git pull requests, you will need the LATEX source to
the book, which is at git://git.kernel.org/pub/scm/linux/kernel/git/
paulmck/perfbook.git, or, alternatively, https://git.kernel.org/pub/scm/
linux/kernel/git/paulmck/perfbook.git. You will of course also need git
and LATEX, which are available as part of most mainstream Linux distributions. Other
packages may be required, depending on the distribution you use. The required list
of packages for a few popular distributions is listed in the file FAQ-BUILD.txt in the
LATEX source to the book.
To create and display a current LATEX source tree of this book, use the list of Linux
commands shown in Listing 1.1. In some environments, the evince command that
displays perfbook.pdf may need to be replaced, for example, with acroread. The
git clone command need only be used the first time you create a PDF, subsequently,
you can run the commands shown in Listing 1.2 to pull in any updates and generate an
v2023.06.11a
1.5. WHOSE BOOK IS THIS? 7
Listing 1.2: Generating an Updated PDF

git remote update
git checkout origin/master
make # -jN for parallel build
evince perfbook.pdf & # Two-column version
make perfbook-1c.pdf
evince perfbook-1c.pdf & # One-column version for e-readers
updated PDF. The commands in Listing 1.2 must be run within the perfbook directory
created by the commands shown in Listing 1.1.
PDFs of this book are sporadically posted at https://kernel.org/pub/
linux/kernel/people/paulmck/perfbook/perfbook.html and at http://
www.rdrop.com/users/paulmck/perfbook/.
The actual process of contributing patches and sending git pull requests is similar to
that of the Linux kernel, which is documented here: https://www.kernel.org/doc/
html/latest/process/submitting-patches.html. One important requirement
is that each patch (or commit, in the case of a git pull request) must contain a valid
Signed-off-by: line, which has the following format:
Signed-off-by: My Name <myname@example.org>
Please see https://lkml.org/lkml/2007/1/15/219 for an example patch with

a Signed-off-by: line. Note well that the Signed-off-by: line has a very specific
meaning, namely that you are certifying that:
(a) The contribution was created in whole or in part by me and I have the right to
submit it under the open source license indicated in the file; or
(b) The contribution is based upon previous work that, to the best of my knowledge, is
covered under an appropriate open source license and I have the right under that
license to submit that work with modifications, whether created in whole or in part
by me, under the same open source license (unless I am permitted to submit under
a different license), as indicated in the file; or
(c) The contribution was provided directly to me by some other person who certified
(a), (b) or (c) and I have not modified it.
(d) I understand and agree that this project and the contribution are public and that
a record of the contribution (including all personal information I submit with
it, including my sign-off) is maintained indefinitely and may be redistributed
consistent with this project or the open source license(s) involved.
This is quite similar to the Developer’s Certificate of Origin (DCO) 1.1 used by the
Linux kernel. You must use your real name: I unfortunately cannot accept pseudonymous
or anonymous contributions.
The language of this book is American English, however, the open-source nature
of this book permits translations, and I personally encourage them. The open-source
licenses covering this book additionally allow you to sell your translation, if you wish. I
do request that you send me a copy of the translation (hardcopy if available), but this
is a request made as a professional courtesy, and is not in any way a prerequisite to
the permission that you already have under the Creative Commons and GPL licenses.
Please see the FAQ.txt file in the source tree for a list of translations currently in
v2023.06.11a
progress. I consider a translation effort to be “in progress” once at least one chapter has
been fully translated.
There are many styles under the “American English” rubric. The style for this
particular book is documented in Appendix D.
As noted at the beginning of this section, I am this book’s editor. However, if you
choose to contribute, it will be your book as well. In that spirit, I offer you Chapter 2,
our introduction.
v2023.06.11a
If parallel programming is so hard, why are there so
many parallel programs?
Unknown
Chapter 2
Introduction
Parallel programming has earned a reputation as one of the most difficult areas a
hacker can tackle. Papers and textbooks warn of the perils of deadlock, livelock, race
conditions, non-determinism, Amdahl’s-Law limits to scaling, and excessive realtime
latencies. And these perils are quite real; we authors have accumulated uncounted years
of experience along with the resulting emotional scars, grey hairs, and hair loss.
However, new technologies that are difficult to use at introduction invariably become
easier over time. For example, the once-rare ability to drive a car is now commonplace
in many countries. This dramatic change came about for two basic reasons: (1) Cars
became cheaper and more readily available, so that more people had the opportunity to
learn to drive, and (2) Cars became easier to operate due to automatic transmissions,
automatic chokes, automatic starters, greatly improved reliability, and a host of other
technological improvements.
The same is true for many other technologies, including computers. It is no longer
necessary to operate a keypunch in order to program. Spreadsheets allow most non-
programmers to get results from their computers that would have required a team of
specialists a few decades ago. Perhaps the most compelling example is web-surfing
and content creation, which since the early 2000s has been easily done by untrained,
uneducated people using various now-commonplace social-networking tools. As
recently as 1968, such content creation was a far-out research project [Eng68], described
at the time as “like a UFO landing on the White House lawn” [Gri00].
Therefore, if you wish to argue that parallel programming will remain as difficult as
it is currently perceived by many to be, it is you who bears the burden of proof, keeping
in mind the many centuries of counter-examples in many fields of endeavor.
2.1 Historic Parallel Programming Difficulties
Not the power to remember, but its very opposite,

the power to forget, is a necessary condition for our
existence.
Sholem Asch
As indicated by its title, this book takes a different approach. Rather than complain
about the difficulty of parallel programming, it instead examines the reasons why
v2023.06.11a
10 CHAPTER 2. INTRODUCTION
parallel programming is difficult, and then works to help the reader to overcome these
difficulties. As will be seen, these difficulties have historically fallen into several
categories, including:
1. The historic high cost and relative rarity of parallel systems.
2. The typical researcher’s and practitioner’s lack of experience with parallel systems.
3. The paucity of publicly accessible parallel code.
4. The lack of a widely understood engineering discipline of parallel programming.
5. The high overhead of communication relative to that of processing, even in tightly

coupled shared-memory computers.
Many of these historic difficulties are well on the way to being overcome. First, over
the past few decades, the cost of parallel systems has decreased from many multiples of
that of a house to that of a modest meal, courtesy of Moore’s Law [Moo65]. Papers
calling out the advantages of multicore CPUs were published as early as 1996 [ONH+ 96].
IBM introduced simultaneous multi-threading into its high-end POWER family in 2000,
and multicore in 2001. Intel introduced hyperthreading into its commodity Pentium
line in November 2000, and both AMD and Intel introduced dual-core CPUs in 2005.
Sun followed with the multicore/multi-threaded Niagara in late 2005. In fact, by 2008,
it was becoming difficult to find a single-CPU desktop system, with single-core CPUs
being relegated to netbooks and embedded devices. By 2012, even smartphones were
starting to sport multiple CPUs. By 2020, safety-critical software standards started
addressing concurrency.
Second, the advent of low-cost and readily available multicore systems means that the
once-rare experience of parallel programming is now available to almost all researchers
and practitioners. In fact, parallel systems have long been within the budget of students
and hobbyists. We can therefore expect greatly increased levels of invention and
innovation surrounding parallel systems, and that increased familiarity will over time
make the once prohibitively expensive field of parallel programming much more friendly
and commonplace.
Third, in the 20th century, large systems of highly parallel software were almost
always closely guarded proprietary secrets. In happy contrast, the 21st century has
seen numerous open-source (and thus publicly available) parallel software projects,
including the Linux kernel [Tor03], database systems [Pos08, MS08], and message-
passing systems [The08, Uni08a]. This book will draw primarily from the Linux kernel,
but will provide much material suitable for user-level applications.
Fourth, even though the large-scale parallel-programming projects of the 1980s and
1990s were almost all proprietary projects, these projects have seeded other communities
with cadres of developers who understand the engineering discipline required to develop
production-quality parallel code. A major purpose of this book is to present this
engineering discipline.
Unfortunately, the fifth difficulty, the high cost of communication relative to that
of processing, remains largely in force. This difficulty has been receiving increasing
attention during the new millennium. However, according to Stephen Hawking,
the finite speed of light and the atomic nature of matter will limit progress in this
area [Gar07, Moo03]. Fortunately, this difficulty has been in force since the late 1980s,
so that the aforementioned engineering discipline has evolved practical and effective
v2023.06.11a
2.2. PARALLEL PROGRAMMING GOALS 11
strategies for handling it. In addition, hardware designers are increasingly aware of
these issues, so perhaps future hardware will be more friendly to parallel software, as
discussed in Section 3.3.
Quick Quiz 2.1: Come on now!!! Parallel programming has been known to be exceedingly
hard for many decades. You seem to be hinting that it is not so hard. What sort of game are you
playing?
However, even though parallel programming might not be as hard as is commonly

advertised, it is often more work than is sequential programming.
Quick Quiz 2.2: How could parallel programming ever be as easy as sequential programming?
It therefore makes sense to consider alternatives to parallel programming. However,

it is not possible to reasonably consider parallel-programming alternatives without
understanding parallel-programming goals. This topic is addressed in the next section.
2.2 Parallel Programming Goals

If you don’t know where you are going, you will end
up somewhere else.
Yogi Berra
The three major goals of parallel programming (over and above those of sequential
programming) are as follows:
1. Performance.
2. Productivity.
3. Generality.
Unfortunately, given the current state of the art, it is possible to achieve at best two of
these three goals for any given parallel program. These three goals therefore form the
iron triangle of parallel programming, a triangle upon which overly optimistic hopes all
too often come to grief.1
Quick Quiz 2.3: Oh, really??? What about correctness, maintainability, robustness, and so
on?
Quick Quiz 2.4: And if correctness, maintainability, and robustness don’t make the list, why
do productivity and generality?
Quick Quiz 2.5: Given that parallel programs are much harder to prove correct than are
sequential programs, again, shouldn’t correctness really be on the list?
Quick Quiz 2.6: What about just having fun?
Each of these goals is elaborated upon in the following sections.
1 Kudos to Michael Wong for naming the iron triangle.
v2023.06.11a
10000
CPU Clock Frequency / MIPS

1000
100
10
0.1
1975
1980
1985
1990
1995
2000
2005
2010
2015
2020
Year
Figure 2.1: MIPS/Clock-Frequency Trend for Intel CPUs
2.2.1 Performance
Performance is the primary goal behind most parallel-programming effort. After all, if
performance is not a concern, why not do yourself a favor: Just write sequential code,
and be happy? It will very likely be easier and you will probably get done much more
quickly.
Quick Quiz 2.7: Are there no cases where parallel programming is about something other
than performance?
Note that “performance” is interpreted broadly here, including for example scalability
(performance per CPU) and efficiency (performance per watt).
That said, the focus of performance has shifted from hardware to parallel software.
This change in focus is due to the fact that, although Moore’s Law continues to deliver
increases in transistor density, it has ceased to provide the traditional single-threaded
performance increases. This can be seen in Figure 2.1,2 which shows that writing
single-threaded code and simply waiting a year or two for the CPUs to catch up may
no longer be an option. Given the recent trends on the part of all major manufacturers
towards multicore/multithreaded systems, parallelism is the way to go for those wanting
to avail themselves of the full performance of their systems.
Quick Quiz 2.8: Why not instead rewrite programs from inefficient scripting languages to C
or C++?
Even so, the first goal is performance rather than scalability, especially given that the
easiest way to attain linear scalability is to reduce the performance of each CPU [Tor01].
2 This plot shows clock frequencies for newer CPUs theoretically capable of retiring one
or more instructions per clock, and MIPS (millions of instructions per second, usually from
the old Dhrystone benchmark) for older CPUs requiring multiple clocks to execute even the
simplest instruction. The reason for shifting between these two measures is that the newer
CPUs’ ability to retire multiple instructions per clock is typically limited by memory-system
performance. Furthermore, the benchmarks commonly used on the older CPUs are obsolete,
and it is difficult to run the newer benchmarks on systems containing the old CPUs, in part
because it is hard to find working instances of the old CPUs.
v2023.06.11a
Given a four-CPU system, which would you prefer? A program that provides 100
transactions per second on a single CPU, but does not scale at all? Or a program that
provides 10 transactions per second on a single CPU, but scales perfectly? The first
program seems like a better bet, though the answer might change if you happened to
have a 32-CPU system.
That said, just because you have multiple CPUs is not necessarily in and of itself
a reason to use them all, especially given the recent decreases in price of multi-CPU
systems. The key point to understand is that parallel programming is primarily a
performance optimization, and, as such, it is one potential optimization of many. If your
program is fast enough as currently written, there is no reason to optimize, either by
parallelizing it or by applying any of a number of potential sequential optimizations.3
By the same token, if you are looking to apply parallelism as an optimization to a
sequential program, then you will need to compare parallel algorithms to the best
sequential algorithms. This may require some care, as far too many publications ignore
the sequential case when analyzing the performance of parallel algorithms.
2.2.2 Productivity
Quick Quiz 2.9: Why all this prattling on about non-technical issues??? And not just any
non-technical issue, but productivity of all things? Who cares?
Productivity has been becoming increasingly important in recent decades. To see

this, consider that the price of early computers was tens of millions of dollars at a time
when engineering salaries were but a few thousand dollars a year. If dedicating a team
of ten engineers to such a machine would improve its performance, even by only 10 %,
then their salaries would be repaid many times over.
One such machine was the CSIRAC, the oldest still-intact stored-program computer,
which was put into operation in 1949 [Mus04, Dep06]. Because this machine was built
before the transistor era, it was constructed of 2,000 vacuum tubes, ran with a clock
frequency of 1 kHz, consumed 30 kW of power, and weighed more than three metric
tons. Given that this machine had but 768 words of RAM, it is safe to say that it did
not suffer from the productivity issues that often plague today’s large-scale software
projects.
Today, it would be quite difficult to purchase a machine with so little computing
power. Perhaps the closest equivalents are 8-bit embedded microprocessors exemplified
by the venerable Z80 [Wik08], but even the old Z80 had a CPU clock frequency more
than 1,000 times faster than the CSIRAC. The Z80 CPU had 8,500 transistors, and
could be purchased in 2008 for less than $2 US per unit in 1,000-unit quantities. In
stark contrast to the CSIRAC, software-development costs are anything but insignificant
for the Z80.
The CSIRAC and the Z80 are two points in a long-term trend, as can be seen in
Figure 2.2. This figure plots an approximation to computational power per die over the
past four decades, showing an impressive six-order-of-magnitude increase over a period
of forty years. Note that the advent of multicore CPUs has permitted this increase to
continue apace despite the clock-frequency wall encountered in 2003, albeit courtesy of
dies supporting more than 50 hardware threads each.
3 Of course, if you are a hobbyist whose primary interest is writing parallel software,
that is more than enough reason to parallelize whatever software you are interested in.
v2023.06.11a
1x106
100000
10000
MIPS per Die

1000
100
10
0.1
1975
1980
1985
1990
1995
2000
2005
2010
2015
2020
Year
Figure 2.2: MIPS per Die for Intel CPUs
One of the inescapable consequences of the rapid decrease in the cost of hardware is
that software productivity becomes increasingly important. It is no longer sufficient
merely to make efficient use of the hardware: It is now necessary to make extremely
efficient use of software developers as well. This has long been the case for sequential
hardware, but parallel hardware has become a low-cost commodity only recently.
Therefore, only recently has high productivity become critically important when
creating parallel software.
Quick Quiz 2.10: Given how cheap parallel systems have become, how can anyone afford to
pay people to program them?
Perhaps at one time, the sole purpose of parallel software was performance. Now,
however, productivity is gaining the spotlight.
2.2.3 Generality
One way to justify the high cost of developing parallel software is to strive for maximal
generality. All else being equal, the cost of a more-general software artifact can be
spread over more users than that of a less-general one. In fact, this economic force
explains much of the maniacal focus on portability, which can be seen as an important
special case of generality.4
Unfortunately, generality often comes at the cost of performance, productivity, or
both. For example, portability is often achieved via adaptation layers, which inevitably
exact a performance penalty. To see this more generally, consider the following popular
parallel programming environments:
C/C++ “Locking Plus Threads”: This category, which includes POSIX Threads
(pthreads) [Ope97], Windows Threads, and numerous operating-system kernel
environments, offers excellent performance (at least within the confines of a
single SMP system) and also offers good generality. Pity about the relatively low
productivity.
4 Kudos to Michael Wong for pointing this out.
v2023.06.11a
Productivity
Application
Middleware (e.g., DBMS)
Performance
Generality
System Libraries
Container
Operating System Kernel

Hypervisor
Firmware
Hardware
Figure 2.3: Software Layers and Performance, Productivity, and Generality
Java: This general purpose and inherently multithreaded programming environment

is widely believed to offer much higher productivity than C or C++, courtesy of
the automatic garbage collector and the rich set of class libraries. However, its
performance, though greatly improved in the early 2000s, lags that of C and C++.
MPI: This Message Passing Interface [MPI08] powers the largest scientific and technical
computing clusters in the world and offers unparalleled performance and scalability.
In theory, it is general purpose, but it is mainly used for scientific and technical
computing. Its productivity is believed by many to be even lower than that of
C/C++ “locking plus threads” environments.
OpenMP: This set of compiler directives can be used to parallelize loops. It is thus
quite specific to this task, and this specificity often limits its performance. It is,
however, much easier to use than MPI or C/C++ “locking plus threads.”
SQL: Structured Query Language [Int92] is specific to relational database queries.
However, its performance is quite good as measured by the Transaction Processing
Performance Council (TPC) benchmark results [Tra01]. Productivity is excellent;
in fact, this parallel programming environment enables people to make good
use of a large parallel system despite having little or no knowledge of parallel
programming concepts.
The nirvana of parallel programming environments, one that offers world-class

performance, productivity, and generality, simply does not yet exist. Until such a
nirvana appears, it will be necessary to make engineering tradeoffs among performance,
productivity, and generality. One such tradeoff is depicted by the green “iron triangle”5
shown in Figure 2.3, which shows how productivity becomes increasingly important
at the upper layers of the system stack, while performance and generality become
increasingly important at the lower layers of the system stack. The huge development
costs incurred at the lower layers must be spread over equally huge numbers of users
5 Kudos to Michael Wong for coining “iron triangle.”
v2023.06.11a
Special−Purpose
User 1 Env Productive User 2
for User 1
HW / Special−Purpose
Abs Environment
Productive for User 2
User 3
General−Purpose User 4
Environment
Special−Purpose Environment
Special−Purpose
Environment
Figure 2.4: Tradeoff Between Productivity and Generality
(hence the importance of generality), and performance lost in lower layers cannot easily
be recovered further up the stack. In the upper layers of the stack, there might be very
few users for a given specific application, in which case productivity concerns are
paramount. This explains the tendency towards “bloatware” further up the stack: Extra
hardware is often cheaper than extra developers. This book is intended for developers
working near the bottom of the stack, where performance and generality are of greatest
concern.
It is important to note that a tradeoff between productivity and generality has existed
for centuries in many fields. For but one example, a nailgun is more productive than
a hammer for driving nails, but in contrast to the nailgun, a hammer can be used for
many things besides driving nails. It should therefore be no surprise to see similar
tradeoffs appear in the field of parallel computing. This tradeoff is shown schematically
in Figure 2.4. Here, users 1, 2, 3, and 4 have specific jobs that they need the computer to
help them with. The most productive possible language or environment for a given user is
one that simply does that user’s job, without requiring any programming, configuration,
or other setup.
Quick Quiz 2.11: This is a ridiculously unachievable ideal! Why not focus on something that
is achievable in practice?
Unfortunately, a system that does the job required by user 1 is unlikely to do user 2’s job.
In other words, the most productive languages and environments are domain-specific,
and thus by definition lacking generality.
Another option is to tailor a given programming language or environment to the
hardware system (for example, low-level languages such as assembly, C, C++, or Java)
or to some abstraction (for example, Haskell, Prolog, or Snobol), as is shown by the
circular region near the center of Figure 2.4. These languages can be considered to be
general in the sense that they are equally ill-suited to the jobs required by users 1, 2, 3,
and 4. In other words, their generality comes at the expense of decreased productivity
when compared to domain-specific languages and environments. Worse yet, a language
that is tailored to a given abstraction is likely to suffer from performance and scalability
problems unless and until it can be efficiently mapped to real hardware.
v2023.06.11a
2.3. ALTERNATIVES TO PARALLEL PROGRAMMING 17
Is there no escape from iron triangle’s three conflicting goals of performance,

productivity, and generality?
It turns out that there often is an escape, for example, using the alternatives to parallel
programming discussed in the next section. After all, parallel programming can be a
great deal of fun, but it is not always the best tool for the job.
2.3 Alternatives to Parallel Programming
Experiment is folly when experience shows the way.

Roger M. Babson
In order to properly consider alternatives to parallel programming, you must first decide
on what exactly you expect the parallelism to do for you. As seen in Section 2.2, the
primary goals of parallel programming are performance, productivity, and generality.
Because this book is intended for developers working on performance-critical code near
the bottom of the software stack, the remainder of this section focuses primarily on
performance improvement.
It is important to keep in mind that parallelism is but one way to improve performance.
Other well-known approaches include the following, in roughly increasing order of
difficulty:
1. Run multiple instances of a sequential application.
2. Make the application use existing parallel software.
3. Optimize the serial application.
These approaches are covered in the following sections.
2.3.1 Multiple Instances of a Sequential Application

Running multiple instances of a sequential application can allow you to do parallel
programming without actually doing parallel programming. There are a large number
of ways to approach this, depending on the structure of the application.
If your program is analyzing a large number of different scenarios, or is analyzing a
large number of independent data sets, one easy and effective approach is to create a
single sequential program that carries out a single analysis, then use any of a number of
scripting environments (for example the bash shell) to run a number of instances of
that sequential program in parallel. In some cases, this approach can be easily extended
to a cluster of machines.
This approach may seem like cheating, and in fact some denigrate such programs
as “embarrassingly parallel”. And in fact, this approach does have some potential
disadvantages, including increased memory consumption, waste of CPU cycles recom-
puting common intermediate results, and increased copying of data. However, it is
often extremely productive, garnering extreme performance gains with little or no added
effort.
v2023.06.11a
2.3.2 Use Existing Parallel Software

There is no longer any shortage of parallel software environments that can present
a single-threaded programming environment, including relational databases [Dat82],
web-application servers, and map-reduce environments. For example, a common design
provides a separate process for each user, each of which generates SQL from user queries.
This per-user SQL is run against a common relational database, which automatically
runs the users’ queries concurrently. The per-user programs are responsible only for
the user interface, with the relational database taking full responsibility for the difficult
issues surrounding parallelism and persistence.
In addition, there are a growing number of parallel library functions, particularly for
numeric computation. Even better, some libraries take advantage of special-purpose
hardware such as vector units and general-purpose graphical processing units (GPGPUs).
Taking this approach often sacrifices some performance, at least when compared to
carefully hand-coding a fully parallel application. However, such sacrifice is often well
repaid by a huge reduction in development effort.
Quick Quiz 2.12: Wait a minute! Doesn’t this approach simply shift the development effort
from you to whoever wrote the existing parallel software you are using?
2.3.3 Performance Optimization

Up through the early 2000s, CPU clock frequencies doubled every 18 months. It was
therefore usually more important to create new functionality than to carefully optimize
performance. Now that Moore’s Law is “only” increasing transistor density instead of
increasing both transistor density and per-transistor performance, it might be a good
time to rethink the importance of performance optimization. After all, new hardware
generations no longer bring significant single-threaded performance improvements.
Furthermore, many performance optimizations can also conserve energy.
From this viewpoint, parallel programming is but another performance optimization,
albeit one that is becoming much more attractive as parallel systems become cheaper and
more readily available. However, it is wise to keep in mind that the speedup available
from parallelism is limited to roughly the number of CPUs (but see Section 6.5 for an
interesting exception). In contrast, the speedup available from traditional single-threaded
software optimizations can be much larger. For example, replacing a long linked list with
a hash table or a search tree can improve performance by many orders of magnitude. This
highly optimized single-threaded program might run much faster than its unoptimized
parallel counterpart, making parallelization unnecessary. Of course, a highly optimized
parallel program would be even better, aside from the added development effort required.
Furthermore, different programs might have different performance bottlenecks. For
example, if your program spends most of its time waiting on data from your disk drive,
using multiple CPUs will probably just increase the time wasted waiting for the disks.
In fact, if the program was reading from a single large file laid out sequentially on a
rotating disk, parallelizing your program might well make it a lot slower due to the
added seek overhead. You should instead optimize the data layout so that the file can be
smaller (thus faster to read), split the file into chunks which can be accessed in parallel
from different drives, cache frequently accessed data in main memory, or, if possible,
reduce the amount of data that must be read.
Quick Quiz 2.13: What other bottlenecks might prevent additional CPUs from providing
additional performance?
v2023.06.11a
2.4. WHAT MAKES PARALLEL PROGRAMMING HARD? 19
Performance Productivity
Work
Partitioning
Resource
Parallel
Partitioning and
Access Control Replication
Interacting
With Hardware
Generality
Figure 2.5: Categories of Tasks Required of Parallel Programmers
Parallelism can be a powerful optimization technique, but it is not the only such
technique, nor is it appropriate for all situations. Of course, the easier it is to
parallelize your program, the more attractive parallelization becomes as an optimization.
Parallelization has a reputation of being quite difficult, which leads to the question
“exactly what makes parallel programming so difficult?”
2.4 What Makes Parallel Programming Hard?
Real difficulties can be overcome; it is only the

imaginary ones that are unconquerable.
Theodore N. Vail
It is important to note that the difficulty of parallel programming is as much a human-

factors issue as it is a set of technical properties of the parallel programming problem.
We do need human beings to be able to tell parallel systems what to do, otherwise known
as programming. But parallel programming involves two-way communication, with
a program’s performance and scalability being the communication from the machine
to the human. In short, the human writes a program telling the computer what to do,
and the computer critiques this program via the resulting performance and scalability.
Therefore, appeals to abstractions or to mathematical analyses will often be of severely
limited utility.
In the Industrial Revolution, the interface between human and machine was evaluated
by human-factor studies, then called time-and-motion studies. Although there have
been a few human-factor studies examining parallel programming [ENS05, ES05,
HCS+ 05, SS94], these studies have been extremely narrowly focused, and hence unable
to demonstrate any general results. Furthermore, given that the normal range of
programmer productivity spans more than an order of magnitude, it is unrealistic
to expect an affordable study to be capable of detecting (say) a 10 % difference in
productivity. Although the multiple-order-of-magnitude differences that such studies
can reliably detect are extremely valuable, the most impressive improvements tend to
be based on a long series of 10 % improvements.
We must therefore take a different approach.
v2023.06.11a
One such approach is to carefully consider the tasks that parallel programmers must
undertake that are not required of sequential programmers. We can then evaluate how
well a given programming language or environment assists the developer with these
tasks. These tasks fall into the four categories shown in Figure 2.5, each of which is
covered in the following sections.
2.4.1 Work Partitioning

Work partitioning is absolutely required for parallel execution: If there is but one “glob”
of work, then it can be executed by at most one CPU at a time, which is by definition
sequential execution. However, partitioning the work requires great care. For example,
uneven partitioning can result in sequential execution once the small partitions have
completed [Amd67]. In less extreme cases, load balancing can be used to fully utilize
available hardware and restore performance and scalability.
Although partitioning can greatly improve performance and scalability, it can also
increase complexity. For example, partitioning can complicate handling of global errors
and events: A parallel program may need to carry out non-trivial synchronization in
order to safely process such global events. More generally, each partition requires
some sort of communication: After all, if a given thread did not communicate at all,
it would have no effect and would thus not need to be executed. However, because
communication incurs overhead, careless partitioning choices can result in severe
performance degradation.
Furthermore, the number of concurrent threads must often be controlled, as each such
thread occupies common resources, for example, space in CPU caches. If too many
threads are permitted to execute concurrently, the CPU caches will overflow, resulting in
high cache miss rate, which in turn degrades performance. Conversely, large numbers
of threads are often required to overlap computation and I/O so as to fully utilize I/O
devices.
Quick Quiz 2.14: Other than CPU cache capacity, what might require limiting the number of
concurrent threads?
Finally, permitting threads to execute concurrently greatly increases the program’s

state space, which can make the program difficult to understand and debug, degrading
productivity. All else being equal, smaller state spaces having more regular structure
are more easily understood, but this is a human-factors statement as much as it is a
technical or mathematical statement. Good parallel designs might have extremely large
state spaces, but nevertheless be easy to understand due to their regular structure, while
poor designs can be impenetrable despite having a comparatively small state space. The
best designs exploit embarrassing parallelism, or transform the problem to one having
an embarrassingly parallel solution. In either case, “embarrassingly parallel” is in fact
an embarrassment of riches. The current state of the art enumerates good designs; more
work is required to make more general judgments on state-space size and structure.
2.4.2 Parallel Access Control

Given a single-threaded sequential program, that single thread has full access to all of
the program’s resources. These resources are most often in-memory data structures,
but can be CPUs, memory (including caches), I/O devices, computational accelerators,
files, and much else besides.
v2023.06.11a
2.4. WHAT MAKES PARALLEL PROGRAMMING HARD? 21
The first parallel-access-control issue is whether the form of access to a given

resource depends on that resource’s location. For example, in many message-passing
environments, local-variable access is via expressions and assignments, while remote-
variable access uses an entirely different syntax, usually involving messaging. The
POSIX Threads environment [Ope97], Structured Query Language (SQL) [Int92],
and partitioned global address-space (PGAS) environments such as Universal Parallel
C (UPC) [EGCD03, CBF13] offer implicit access, while Message Passing Interface
(MPI) [MPI08] offers explicit access because access to remote data requires explicit
messaging.
The other parallel-access-control issue is how threads coordinate access to the
resources. This coordination is carried out by the very large number of synchronization
mechanisms provided by various parallel languages and environments, including
message passing, locking, transactions, reference counting, explicit timing, shared
atomic variables, and data ownership. Many traditional parallel-programming concerns
such as deadlock, livelock, and transaction rollback stem from this coordination.
This framework can be elaborated to include comparisons of these synchronization
mechanisms, for example locking vs. transactional memory [MMW07], but such
elaboration is beyond the scope of this section. (See Sections 17.2 and 17.3 for more
information on transactional memory.)
Quick Quiz 2.15: Just what is “explicit timing”???
2.4.3 Resource Partitioning and Replication

The most effective parallel algorithms and systems exploit resource parallelism, so much
so that it is usually wise to begin parallelization by partitioning your write-intensive
resources and replicating frequently accessed read-mostly resources. The resource in
question is most frequently data, which might be partitioned over computer systems,
mass-storage devices, NUMA nodes, CPU cores (or dies or hardware threads), pages,
cache lines, instances of synchronization primitives, or critical sections of code. For
example, partitioning over locking primitives is termed “data locking” [BK85].
Resource partitioning is frequently application dependent. For example, numerical
applications frequently partition matrices by row, column, or sub-matrix, while com-
mercial applications frequently partition write-intensive data structures and replicate
read-mostly data structures. Thus, a commercial application might assign the data for a
given customer to a given few computers out of a large cluster. An application might
statically partition data, or dynamically change the partitioning over time.
Resource partitioning is extremely effective, but it can be quite challenging for
complex multilinked data structures.
2.4.4 Interacting With Hardware

Hardware interaction is normally the domain of the operating system, the compiler,
libraries, or other software-environment infrastructure. However, developers working
with novel hardware features and components will often need to work directly with such
hardware. In addition, direct access to the hardware can be required when squeezing
the last drop of performance out of a given system. In this case, the developer may
need to tailor or configure the application to the cache geometry, system topology, or
interconnect protocol of the target hardware.
v2023.06.11a
Performance Productivity
Work
Partitioning
Resource
Parallel
Partitioning and
Interacting
With Hardware
Generality
Figure 2.6: Ordering of Parallel-Programming Tasks
In some cases, hardware may be considered to be a resource which is subject to

partitioning or access control, as described in the previous sections.
2.4.5 Composite Capabilities
Although these four capabilities are fundamental, good engineering practice uses
composites of these capabilities. For example, the data-parallel approach first partitions
the data so as to minimize the need for inter-partition communication, partitions the code
accordingly, and finally maps data partitions and threads so as to maximize throughput
while minimizing inter-thread communication, as shown in Figure 2.6. The developer
can then consider each partition separately, greatly reducing the size of the relevant state
space, in turn increasing productivity. Even though some problems are non-partitionable,
clever transformations into forms permitting partitioning can sometimes greatly enhance
both performance and scalability [Met99].
2.4.6 How Do Languages and Environments Assist With These

Tasks?
Although many environments require the developer to deal manually with these tasks,
there are long-standing environments that bring significant automation to bear. The poster
child for these environments is SQL, many implementations of which automatically
parallelize single large queries and also automate concurrent execution of independent
queries and updates.
These four categories of tasks must be carried out in all parallel programs, but that
of course does not necessarily mean that the developer must manually carry out these
tasks. We can expect to see ever-increasing automation of these four tasks as parallel
systems continue to become cheaper and more readily available.
Quick Quiz 2.16: Are there any other obstacles to parallel programming?
v2023.06.11a
2.5. DISCUSSION 23
2.5 Discussion
Until you try, you don’t know what you can’t do.
Henry James
This section has given an overview of the difficulties with, goals of, and alternatives to
parallel programming. This overview was followed by a discussion of what can make
parallel programming hard, along with a high-level approach for dealing with parallel
programming’s difficulties. Those who still insist that parallel programming is impossibly
difficult should review some of the older guides to parallel programmming [Seq88,
Bir89, BK85, Inm85]. The following quote from Andrew Birrell’s monograph [Bir89]
is especially telling:
Writing concurrent programs has a reputation for being exotic and difficult.
I believe it is neither. You need a system that provides you with good primitives
and suitable libraries, you need a basic caution and carefulness, you need an
armory of useful techniques, and you need to know of the common pitfalls.
I hope that this paper has helped you towards sharing my belief.
The authors of these older guides were well up to the parallel programming challenge
back in the 1980s. As such, there are simply no excuses for refusing to step up to the
parallel-programming challenge here in the 21st century!
We are now ready to proceed to the next chapter, which dives into the relevant
properties of the parallel hardware underlying our parallel software.
v2023.06.11a
v2023.06.11a
Premature abstraction is the root of all evil.
A cast of thousands
Chapter 3
Hardware and its Habits
Most people intuitively understand that passing messages between systems is more
expensive than performing simple calculations within the confines of a single system.
But it is also the case that communicating among threads within the confines of a single
shared-memory system can also be quite expensive. This chapter therefore looks at the
cost of synchronization and communication within a shared-memory system. These
few pages can do no more than scratch the surface of shared-memory parallel hardware
design; readers desiring more detail would do well to start with a recent edition of
Hennessy’s and Patterson’s classic text [HP17, HP95].
Quick Quiz 3.1: Why should parallel programmers bother learning low-level properties of
the hardware? Wouldn’t it be easier, better, and more elegant to remain at a higher level of
abstraction?
3.1 Overview
Mechanical Sympathy: Hardware and software
working together in harmony.
Martin Thompson
Careless reading of computer-system specification sheets might lead one to believe that
CPU performance is a footrace on a clear track, as illustrated in Figure 3.1, where the
race always goes to the swiftest.
Although there are a few CPU-bound benchmarks that approach the ideal case shown
in Figure 3.1, the typical program more closely resembles an obstacle course than a
race track. This is because the internal architecture of CPUs has changed dramatically
over the past few decades, courtesy of Moore’s Law. These changes are described in the
following sections.
3.1.1 Pipelined CPUs

In the 1980s, the typical microprocessor fetched an instruction, decoded it, and executed
it, typically taking at least three clock cycles to complete one instruction before even
starting the next. In contrast, the CPU of the late 1990s and of the 2000s execute
many instructions simultaneously, using pipelines; superscalar techniques; out-of-order
25
v2023.06.11a
26 CHAPTER 3. HARDWARE AND ITS HABITS
CPU Benchmark Trackmeet
Figure 3.1: CPU Performance at its Best
4.0 GHz clock, 20 MB L3

cache, 20 stage pipeline...
The only pipeline I need

is to cool off that hot-
headed brat.
Figure 3.2: CPUs Old and New
v2023.06.11a
3.1. OVERVIEW 27
ON
TI
DIC
RE
SP
PI MI
PE H
LI
NE A NC
ER BR
RO
R
Figure 3.3: CPU Meets a Pipeline Flush
instruction and data handling; speculative execution, and more [HP17, HP11] in order
to optimize the flow of instructions and data through the CPU. Some cores have more
than one hardware thread, which is variously called simultaneous multithreading (SMT)
or hyperthreading (HT) [Fen73], each of which appears as an independent CPU to
software, at least from a functional viewpoint. These modern hardware features can
greatly improve performance, as illustrated by Figure 3.2.
Achieving full performance with a CPU having a long pipeline requires highly
predictable control flow through the program. Suitable control flow can be provided
by a program that executes primarily in tight loops, for example, arithmetic on large
matrices or vectors. The CPU can then correctly predict that the branch at the end of
the loop will be taken in almost all cases, allowing the pipeline to be kept full and the
CPU to execute at full speed.
However, branch prediction is not always so easy. For example, consider a program
with many loops, each of which iterates a small but random number of times. For
another example, consider an old-school object-oriented program with many virtual
objects that can reference many different real objects, all with different implementations
for frequently invoked member functions, resulting in many calls through pointers. In
these cases, it is difficult or even impossible for the CPU to predict where the next
branch might lead. Then either the CPU must stall waiting for execution to proceed
far enough to be certain where that branch leads, or it must guess and then proceed
using speculative execution. Although guessing works extremely well for programs
with predictable control flow, for unpredictable branches (such as those in binary search)
the guesses will frequently be wrong. A wrong guess can be expensive because the
CPU must discard any speculatively executed instructions following the corresponding
branch, resulting in a pipeline flush. If pipeline flushes appear too frequently, they
drastically reduce overall performance, as fancifully depicted in Figure 3.3.
This gets even worse in the increasingly common case of hyperthreading (or SMT, if
you prefer), especially on pipelined superscalar out-of-order CPU featuring speculative
execution. In this increasingly common case, all the hardware threads sharing a core also
share that core’s resources, including registers, cache, execution units, and so on. The
instructions are often decoded into micro-operations, and use of the shared execution
v2023.06.11a
Thread 0 Thread 1
Instructions Instructions
Decode and
Translate
Micro-Op
Scheduler
Registers
Execution (100s!)
Units
Figure 3.4: Rough View of Modern Micro-Architecture
units and the hundreds of hardware registers is often coordinated by a micro-operation

scheduler. A rough diagram of such a two-threaded core is shown in Figure 3.4, and
more accurate (and thus more complex) diagrams are available in textbooks and scholarly
papers.1 Therefore, the execution of one hardware thread can and often is perturbed by
the actions of other hardware threads sharing that core.
Even if only one hardware thread is active (for example, in old-school CPU designs
where there is only one thread), counterintuitive results are quite common. Execution
units often have overlapping capabilities, so that a CPU’s choice of execution unit can
result in pipeline stalls due to contention for that execution unit from later instructions.
In theory, this contention is avoidable, but in practice CPUs must choose very quickly
and without the benefit of clairvoyance. In particular, adding an instruction to a tight
loop can sometimes actually cause execution to speed up.
Unfortunately, pipeline flushes and shared-resource contention are not the only
hazards in the obstacle course that modern CPUs must run. The next section covers the
hazards of referencing memory.
3.1.2 Memory References

In the 1980s, it often took less time for a microprocessor to load a value from memory
than it did to execute an instruction. More recently, microprocessors might execute
hundreds or even thousands of instructions in the time required to access memory. This
disparity is due to the fact that Moore’s Law has increased CPU performance at a much
greater rate than it has decreased memory latency, in part due to the rate at which
memory sizes have grown. For example, a typical 1970s minicomputer might have 4 KB
(yes, kilobytes, not megabytes, let alone gigabytes or terabytes) of main memory, with
single-cycle access.2 Present-day CPU designers still can construct a 4 KB memory
with single-cycle access, even on systems with multi-GHz clock frequencies. And
in fact they frequently do construct such memories, but they now call them “level-0
caches”, plus they can be quite a bit bigger than 4 KB.
1 Here is one example for a late-2010s Intel core: https://en.wikichip.org/wiki/
intel/microarchitectures/skylake_(server).
2 It is only fair to add that each of these single cycles lasted no less than 1.6 microseconds.
v2023.06.11a
3.1. OVERVIEW 29
Figure 3.5: CPU Meets a Memory Reference
Although the large caches found on modern microprocessors can do quite a bit to help
combat memory-access latencies, these caches require highly predictable data-access
patterns to successfully hide those latencies. Unfortunately, common operations such as
traversing a linked list have extremely unpredictable memory-access patterns—after
all, if the pattern was predictable, us software types would not bother with the pointers,
right? Therefore, as shown in Figure 3.5, memory references often pose severe obstacles
to modern CPUs.
Thus far, we have only been considering obstacles that can arise during a given CPU’s
execution of single-threaded code. Multi-threading presents additional obstacles to the
CPU, as described in the following sections.
3.1.3 Atomic Operations

One such obstacle is atomic operations. The problem here is that the whole idea of an
atomic operation conflicts with the piece-at-a-time assembly-line operation of a CPU
pipeline. To hardware designers’ credit, modern CPUs use a number of extremely clever
tricks to make such operations look atomic even though they are in fact being executed
piece-at-a-time, with one common trick being to identify all the cachelines containing
the data to be atomically operated on, ensure that these cachelines are owned by the
CPU executing the atomic operation, and only then proceed with the atomic operation
while ensuring that these cachelines remained owned by this CPU. Because all the data
is private to this CPU, other CPUs are unable to interfere with the atomic operation
despite the piece-at-a-time nature of the CPU’s pipeline. Needless to say, this sort of
trick can require that the pipeline must be delayed or even flushed in order to perform
the setup operations that permit a given atomic operation to complete correctly.
In contrast, when executing a non-atomic operation, the CPU can load values from
cachelines as they appear and place the results in the store buffer, without the need to
wait for cacheline ownership. Although there are a number of hardware optimizations
v2023.06.11a
Figure 3.6: CPU Meets an Atomic Operation
that can sometimes hide cache latencies, the resulting effect on performance is all too
often as depicted in Figure 3.6.
Unfortunately, atomic operations usually apply only to single elements of data.
Because many parallel algorithms require that ordering constraints be maintained
between updates of multiple data elements, most CPUs provide memory barriers. These
memory barriers also serve as performance-sapping obstacles, as described in the next
section.
Quick Quiz 3.2: What types of machines would allow atomic operations on multiple data
elements?
3.1.4 Memory Barriers

Memory barriers will be considered in more detail in Chapter 15 and Appendix C. In
the meantime, consider the following simple lock-based critical section:
1 spin_lock(&mylock);
2 a = a + 1;
3 spin_unlock(&mylock);
If the CPU were not constrained to execute these statements in the order shown, the
effect would be that the variable “a” would be incremented without the protection of
“mylock”, which would certainly defeat the purpose of acquiring it. To prevent such
destructive reordering, locking primitives contain either explicit or implicit memory
barriers. Because the whole purpose of these memory barriers is to prevent reorderings
that the CPU would otherwise undertake in order to increase performance, memory
barriers almost always reduce performance, as depicted in Figure 3.7.
As with atomic operations, CPU designers have been working hard to reduce
memory-barrier overhead, and have made substantial progress.
3.1.5 Thermal Throttling

One increasingly common frustrating experience is to carefully micro-optimize a critical
code path, greatly reducing the number of clock cycles consumed by that code path,
only to find that the wall-clock time consumed by that code has actually increased.
v2023.06.11a
3.1. OVERVIEW 31
Memory
Barrier
Figure 3.7: CPU Meets a Memory Barrier
Figure 3.8: CPU Encounters Thermal Throttling
v2023.06.11a
CACHE-
MISS
TOLL
BOOTH
Figure 3.9: CPU Meets a Cache Miss
Welcome to modern thermal throttling.

If you reduced the number of clock cycles by making more effective use of the CPU’s
functional units, you will have increased the power consumed by that CPU. This will
in turn increase the amount of heat dissipated by that CPU. If this heat dissipation
exceeds the cooling system’s capacity, the system will thermally throttle that CPU, for
example, by reducing its clock frequency, as fancifully depicted by the snow penguin in
Figure 3.8.
If performance is of the essence, the proper fix is improved cooling, an approach loved
by serious gamers and by overclockers.3 But if you cannot modify your computer’s
cooling system, perhaps because you are renting it from a cloud provider, then you
will need to take some other optimization approach. For example, you might need
to apply algorithmic optimizations instead of hardware-centric micro-optimizations.
Alternatively, perhaps you can parallelize your code, spreading the work (and thus the
heat) over multiple CPU cores.
3.1.6 Cache Misses

An additional multi-threading obstacle to CPU performance is the “cache miss”. As
noted earlier, modern CPUs sport large caches in order to reduce the performance
penalty that would otherwise be incurred due to high memory latencies. However, these
caches are actually counter-productive for variables that are frequently shared among
CPUs. This is because when a given CPU wishes to modify the variable, it is most
likely the case that some other CPU has modified it recently. In this case, the variable
will be in that other CPU’s cache, but not in this CPU’s cache, which will therefore
incur an expensive cache miss (see Appendix C.1 for more detail). Such cache misses
form a major obstacle to CPU performance, as shown in Figure 3.9.
Quick Quiz 3.3: So have CPU designers also greatly reduced the overhead of cache misses?
3 Some of whom make good use of liquid nitrogen.
v2023.06.11a
3.2. OVERHEADS 33
TELE Please stay on the

line. Your call is very
important to us...
Figure 3.10: CPU Waits for I/O Completion
3.1.7 I/O Operations

A cache miss can be thought of as a CPU-to-CPU I/O operation, and as such is one
of the cheapest I/O operations available. I/O operations involving networking, mass
storage, or (worse yet) human beings pose much greater obstacles than the internal
obstacles called out in the prior sections, as illustrated by Figure 3.10.
This is one of the differences between shared-memory and distributed-system
parallelism: Shared-memory parallel programs must normally deal with no obstacle
worse than a cache miss, while a distributed parallel program will typically incur the
larger network communication latencies. In both cases, the relevant latencies can be
thought of as a cost of communication—a cost that would be absent in a sequential
program. Therefore, the ratio between the overhead of the communication to that of
the actual work being performed is a key design parameter. A major goal of parallel
hardware design is to reduce this ratio as needed to achieve the relevant performance
and scalability goals. In turn, as will be seen in Chapter 6, a major goal of parallel
software design is to reduce the frequency of expensive operations like communications
cache misses.
Of course, it is one thing to say that a given operation is an obstacle, and quite another
to show that the operation is a significant obstacle. This distinction is discussed in the
following sections.
3.2 Overheads
Don’t design bridges in ignorance of materials, and
don’t design low-level software in ignorance of the
underlying hardware.
Unknown
This section presents actual overheads of the obstacles to performance listed out in the
previous section. However, it is first necessary to get a rough view of hardware system
architecture, which is the subject of the next section.
v2023.06.11a
CPU 0 CPU 1 CPU 2 CPU 3

Cache Cache Cache Cache
Interconnect Interconnect
Memory System Interconnect Memory
Speed−of−Light Round−Trip Distance in Vacuum

for 1.8 GHz Clock Period (8 cm)
Figure 3.11: System Hardware Architecture
3.2.1 Hardware System Architecture

Figure 3.11 shows a rough schematic of an eight-core computer system. Each die has a
pair of CPU cores, each with its cache, as well as an interconnect allowing the pair of
CPUs to communicate with each other. The system interconnect allows the four dies to
communicate with each other and with main memory.
Data moves through this system in units of “cache lines”, which are power-of-two
fixed-size aligned blocks of memory, usually ranging from 32 to 256 bytes in size.
When a CPU loads a variable from memory to one of its registers, it must first load
the cacheline containing that variable into its cache. Similarly, when a CPU stores a
value from one of its registers into memory, it must also load the cacheline containing
that variable into its cache, but must also ensure that no other CPU has a copy of that
cacheline.
For example, if CPU 0 were to write to a variable whose cacheline resided in CPU 7’s
cache, the following over-simplified sequence of events might ensue:
1. CPU 0 checks its local cache, and does not find the cacheline. It therefore records
the write in its store buffer.
2. A request for this cacheline is forwarded to CPU 0’s and 1’s interconnect, which
checks CPU 1’s local cache, and does not find the cacheline.
3. This request is forwarded to the system interconnect, which checks with the other
three dies, learning that the cacheline is held by the die containing CPU 6 and 7.
4. This request is forwarded to CPU 6’s and 7’s interconnect, which checks both
CPUs’ caches, finding the value in CPU 7’s cache.
5. CPU 7 forwards the cacheline to its interconnect, and also flushes the cacheline
from its cache.
6. CPU 6’s and 7’s interconnect forwards the cacheline to the system interconnect.
7. The system interconnect forwards the cacheline to CPU 0’s and 1’s interconnect.
v2023.06.11a
3.2. OVERHEADS 35
8. CPU 0’s and 1’s interconnect forwards the cacheline to CPU 0’s cache.
9. CPU 0 can now complete the write, updating the relevant portions of the newly
arrived cacheline from the value previously recorded in the store buffer.
Quick Quiz 3.4: This is a simplified sequence of events? How could it possibly be any more
complex?
Quick Quiz 3.5: Why is it necessary to flush the cacheline from CPU 7’s cache?
This simplified sequence is just the beginning of a discipline called cache-coherency

protocols [HP95, CSG99, MHS12, SHW11], which is discussed in more detail in
Appendix C. As can be seen in the sequence of events triggered by a CAS operation,
a single instruction can cause considerable protocol traffic, which can significantly
degrade your parallel program’s performance.
Fortunately, if a given variable is being frequently read during a time interval during
which it is never updated, that variable can be replicated across all CPUs’ caches.
This replication permits all CPUs to enjoy extremely fast access to this read-mostly
variable. Chapter 9 presents synchronization mechanisms that take full advantage of
this important hardware read-mostly optimization.
3.2.2 Costs of Operations

The overheads of some common operations important to parallel programs are displayed
in Table 3.1. This system’s clock period rounds to 0.5 ns. Although it is not unusual
for modern microprocessors to be able to retire multiple instructions per clock period,
the operations’ costs are nevertheless normalized to a clock period in the third column,
labeled “Ratio”. The first thing to note about this table is the large values of many of
the ratios.
The same-CPU compare-and-swap (CAS) operation consumes about seven nanosec-
onds, a duration more than ten times that of the clock period. CAS is an atomic operation
in which the hardware compares the contents of the specified memory location to a
specified “old” value, and if they compare equal, stores a specified “new” value, in
which case the CAS operation succeeds. If they compare unequal, the memory location
keeps its (unexpected) value, and the CAS operation fails. The operation is atomic in
that the hardware guarantees that the memory location will not be changed between the
compare and the store. CAS functionality is provided by the lock;cmpxchg instruction
on x86.
The “same-CPU” prefix means that the CPU now performing the CAS operation on a
given variable was also the last CPU to access this variable, so that the corresponding
cacheline is already held in that CPU’s cache. Similarly, the same-CPU lock operation
(a “round trip” pair consisting of a lock acquisition and release) consumes more than
fifteen nanoseconds, or more than thirty clock cycles. The lock operation is more
expensive than CAS because it requires two atomic operations on the lock data structure,
one for acquisition and the other for release.
On-core operations involving interactions between the hardware threads sharing a
single core are about the same cost as same-CPU operations. This should not be too
surprising, given that these two hardware threads also share the full cache hierarchy.
In the case of the blind CAS, the software specifies the old value without looking at
the memory location. This approach is appropriate when attempting to acquire a lock.
v2023.06.11a
Table 3.1: CPU 0 View of Synchronization Mechanisms on 8-Socket System With

Intel Xeon Platinum 8176 CPUs @ 2.10 GHz
Ratio
Operation Cost (ns) (cost/clock) CPUs
Clock period 0.5 1.0
Same-CPU 0
CAS 7.0 14.6
lock 15.4 32.3
On-Core 224
Blind CAS 7.2 15.2
CAS 18.0 37.7
Off-Core 1–27
Blind CAS 47.5 99.8 225–251
CAS 101.9 214.0
Off-Socket 28–111
Blind CAS 148.8 312.5 252–335
CAS 442.9 930.1
Cross-Interconnect 112–223
Blind CAS 336.6 706.8 336–447
CAS 944.8 1,984.2
Off-System
Comms Fabric 5,000 10,500
Global Comms 195,000,000 409,500,000
If the unlocked state is represented by zero and the locked state is represented by the
value one, then a CAS operation on the lock that specifies zero for the old value and one
for the new value will acquire the lock if it is not already held. The key point is that
there is only one access to the memory location, namely the CAS operation itself.
In contrast, a normal CAS operation’s old value is derived from some earlier load. For
example, to implement an atomic increment, the current value of that location is loaded
and that value is incremented to produce the new value. Then in the CAS operation, the
value actually loaded would be specified as the old value and the incremented value
as the new value. If the value had not been changed between the load and the CAS,
this would increment the memory location. However, if the value had in fact changed,
then the old value would not match, causing a miscompare that would result in the
CAS operation failing. The key point is that there are now two accesses to the memory
location, the load and the CAS.
Thus, it is not surprising that on-core blind CAS consumes only about seven
nanoseconds, while on-core CAS consumes about 18 nanoseconds. The non-blind
case’s extra load does not come for free. That said, the overhead of these operations are
similar to same-CPU CAS and lock, respectively.
Quick Quiz 3.6: Table 3.1 shows CPU 0 sharing a core with CPU 224. Shouldn’t that instead
be CPU 1???
v2023.06.11a
3.2. OVERHEADS 37
Table 3.2: Cache Geometry for 8-Socket System With Intel Xeon Platinum 8176 CPUs
@ 2.10 GHz
Level Scope Line Size Sets Ways Size

L0 Core 64 64 8 32K
L1 Core 64 64 8 32K
L2 Core 64 1024 16 1024K
L3 Socket 64 57,344 11 39,424K
A blind CAS involving CPUs in different cores but on the same socket consumes
almost fifty nanoseconds, or almost one hundred clock cycles. The code used for this
cache-miss measurement passes the cache line back and forth between a pair of CPUs,
so this cache miss is satisfied not from memory, but rather from the other CPU’s cache.
A non-blind CAS operation, which as noted earlier must look at the old value of the
variable as well as store a new value, consumes over one hundred nanoseconds, or
more than two hundred clock cycles. Think about this a bit. In the time required to
do one CAS operation, the CPU could have executed more than two hundred normal
instructions. This should demonstrate the limitations not only of fine-grained locking,
but of any other synchronization mechanism relying on fine-grained global agreement.
If the pair of CPUs are on different sockets, the operations are considerably more
expensive. A blind CAS operation consumes almost 150 nanoseconds, or more than three
hundred clock cycles. A normal CAS operation consumes more than 400 nanoseconds,
or almost one thousand clock cycles.
Worse yet, not all pairs of sockets are created equal. This particular system appears
to be constructed as a pair of four-socket components, with additional latency penalties
when the CPUs reside in different components. In this case, a blind CAS operation
consumes more than three hundred nanoseconds, or more than seven hundred clock
cycles. A CAS operation consumes almost a full microsecond, or almost two thousand
clock cycles.
Quick Quiz 3.7: Surely the hardware designers could be persuaded to improve this situation!
Why have they been content with such abysmal performance for these single-instruction
operations?
Quick Quiz 3.8: Table E.1 in the answer to Quick Quiz 3.7 on page 716 says that on-core CAS
is faster than both of same-CPU CAS and on-core blind CAS. What is happening there?
Unfortunately, the high speed of within-core and within-socket communication does

not come for free. First, there are only two CPUs within a given core and only 56 within
a given socket, compared to 448 across the system. Second, as shown in Table 3.2, the
on-core caches are quite small compared to the on-socket caches, which are in turn
quite small compared to the 1.4 TB of memory configured on this system. Third, again
referring to the figure, the caches are organized as a hardware hash table with a limited
number of items per bucket. For example, the raw size of the L3 cache (“Size”) is
almost 40 MB, but each bucket (“Line”) can only hold 11 blocks of memory (“Ways”),
each of which can be at most 64 bytes (“Line Size”). This means that only 12 bytes of
memory (admittedly at carefully chosen addresses) are required to overflow this 40 MB
cache. On the other hand, equally careful choice of addresses might make good use of
the entire 40 MB.
v2023.06.11a
Spatial locality of reference is clearly extremely important, as is spreading the data

across memory.
I/O operations are even more expensive. As shown in the “Comms Fabric” row,
high performance (and expensive!) communications fabric, such as InfiniBand or any
number of proprietary interconnects, has a latency of roughly five microseconds for an
end-to-end round trip, during which time more than ten thousand instructions might
have been executed. Standards-based communications networks often require some
sort of protocol processing, which further increases the latency. Of course, geographic
distance also increases latency, with the speed-of-light through optical fiber latency
around the world coming to roughly 195 milliseconds, or more than 400 million clock
cycles, as shown in the “Global Comms” row.
Quick Quiz 3.9: These numbers are insanely large! How can I possibly get my head around
them?
3.2.3 Hardware Optimizations

It is only natural to ask how the hardware is helping, and the answer is “Quite a bit!”
One hardware optimization is large cachelines. This can provide a big performance
boost, especially when software is accessing memory sequentially. For example, given
a 64-byte cacheline and software accessing 64-bit variables, the first access will still be
slow due to speed-of-light delays (if nothing else), but the remaining seven can be quite
fast. However, this optimization has a dark side, namely false sharing, which happens
when different variables in the same cacheline are being updated by different CPUs,
resulting in a high cache-miss rate. Software can use the alignment directives available
in many compilers to avoid false sharing, and adding such directives is a common step
in tuning parallel software.
A second related hardware optimization is cache prefetching, in which the hardware
reacts to consecutive accesses by prefetching subsequent cachelines, thereby evading
speed-of-light delays for these subsequent cachelines. Of course, the hardware must use
simple heuristics to determine when to prefetch, and these heuristics can be fooled by
the complex data-access patterns in many applications. Fortunately, some CPU families
allow for this by providing special prefetch instructions. Unfortunately, the effectiveness
of these instructions in the general case is subject to some dispute.
A third hardware optimization is the store buffer, which allows a string of store
instructions to execute quickly even when the stores are to non-consecutive addresses
and when none of the needed cachelines are present in the CPU’s cache. The dark side
of this optimization is memory misordering, for which see Chapter 15.
A fourth hardware optimization is speculative execution, which can allow the hardware
to make good use of the store buffers without resulting in memory misordering. The
dark side of this optimization can be energy inefficiency and lowered performance if the
speculative execution goes awry and must be rolled back and retried. Worse yet, the
advent of Spectre and Meltdown [Hor18] made it apparent that hardware speculation
can also enable side-channel attacks that defeat memory-protection hardware so as to
allow unprivileged processes to read memory that they should not have access to. It is
clear that the combination of speculative execution and cloud computing needs more
than a bit of rework!
A fifth hardware optimization is large caches, allowing individual CPUs to operate
on larger datasets without incurring expensive cache misses. Although large caches
v2023.06.11a
3.3. HARDWARE FREE LUNCH? 39
Figure 3.12: Hardware and Software: On Same Side
can degrade energy efficiency and cache-miss latency, the ever-growing cache sizes on
production microprocessors attests to the power of this optimization.
A final hardware optimization is read-mostly replication, in which data that is
frequently read but rarely updated is present in all CPUs’ caches. This optimization
allows the read-mostly data to be accessed exceedingly efficiently, and is the subject of
Chapter 9.
In short, hardware and software engineers are really on the same side, with both trying
to make computers go fast despite the best efforts of the laws of physics, as fancifully
depicted in Figure 3.12 where our data stream is trying its best to exceed the speed of
light. The next section discusses some additional things that the hardware engineers
might (or might not) be able to do, depending on how well recent research translates to
practice. Software’s contribution to this noble goal is outlined in the remaining chapters
of this book.
3.3 Hardware Free Lunch?
The great trouble today is that there are too many

people looking for someone else to do something for
them. The solution to most of our troubles is to be
found in everyone doing something for themselves.
Henry Ford, updated
The major reason that concurrency has been receiving so much focus over the past few
years is the end of Moore’s-Law induced single-threaded performance increases (or
“free lunch” [Sut08]), as shown in Figure 2.1 on page 12. This section briefly surveys a
few ways that hardware designers might bring back the “free lunch”.
However, the preceding section presented some substantial hardware obstacles to
exploiting concurrency. One severe physical limitation that hardware designers face is
the finite speed of light. As noted in Figure 3.11 on page 34, light can manage only
about an 8-centimeters round trip in a vacuum during the duration of a 1.8 GHz clock
period. This distance drops to about 3 centimeters for a 5 GHz clock. Both of these
distances are relatively small compared to the size of a modern computer system.
To make matters even worse, electric waves in silicon move from three to thirty times
more slowly than does light in a vacuum, and common clocked logic constructs run still
v2023.06.11a
70 um
3 cm 1.5 cm
Figure 3.13: Latency Benefit of 3D Integration
more slowly, for example, a memory reference may need to wait for a local cache lookup
to complete before the request may be passed on to the rest of the system. Furthermore,
relatively low speed and high power drivers are required to move electrical signals
from one silicon die to another, for example, to communicate between a CPU and main
memory.
Quick Quiz 3.10: But individual electrons don’t move anywhere near that fast, even in
conductors!!! The electron drift velocity in a conductor under semiconductor voltage levels is
on the order of only one millimeter per second. What gives???
There are nevertheless some technologies (both hardware and software) that might
help improve matters:
1. 3D integration,
2. Novel materials and processes,
3. Substituting light for electricity,
4. Special-purpose accelerators, and
5. Existing parallel software.
Each of these is described in one of the following sections.
3.3.1 3D Integration
3-dimensional integration (3DI) is the practice of bonding very thin silicon dies to
each other in a vertical stack. This practice provides potential benefits, but also poses
significant fabrication challenges [Kni08].
Perhaps the most important benefit of 3DI is decreased path length through the system,
as shown in Figure 3.13. A 3-centimeter silicon die is replaced with a stack of four
1.5-centimeter dies, in theory decreasing the maximum path through the system by a
factor of two, keeping in mind that each layer is quite thin. In addition, given proper
attention to design and placement, long horizontal electrical connections (which are
both slow and power hungry) can be replaced by short vertical electrical connections,
which are both faster and more power efficient.
However, delays due to levels of clocked logic will not be decreased by 3D integration,
and significant manufacturing, testing, power-supply, and heat-dissipation problems
must be solved for 3D integration to reach production while still delivering on its
promise. The heat-dissipation problems might be solved using semiconductors based
v2023.06.11a
3.3. HARDWARE FREE LUNCH? 41
on diamond, which is a good conductor for heat, but an electrical insulator. That said, it
remains difficult to grow large single diamond crystals, to say nothing of slicing them
into wafers. In addition, it seems unlikely that any of these technologies will be able to
deliver the exponential increases to which some people have become accustomed. That
said, they may be necessary steps on the path to the late Jim Gray’s “smoking hairy golf
balls” [Gra02].
3.3.2 Novel Materials and Processes

Stephen Hawking is said to have claimed that semiconductor manufacturers have but
two fundamental problems: (1) The finite speed of light and (2) The atomic nature of
matter [Gar07]. It is possible that semiconductor manufacturers are approaching these
limits, but there are nevertheless a few avenues of research and development focused on
working around these fundamental limits.
One workaround for the atomic nature of matter are so-called “high-K dielectric”
materials, which allow larger devices to mimic the electrical properties of infeasibly
small devices. These materials pose some severe fabrication challenges, but nevertheless
may help push the frontiers out a bit farther. Another more-exotic workaround stores
multiple bits in a single electron, relying on the fact that a given electron can exist at a
number of energy levels. It remains to be seen if this particular approach can be made
to work reliably in production semiconductor devices.
Another proposed workaround is the “quantum dot” approach that allows much
smaller device sizes, but which is still in the research stage.
One challenge is that many recent hardware-device-level breakthroughs require very
tight control of which atoms are placed where [Kel17]. It therefore seems likely that
whoever finds a good way to hand-place atoms on each of the billions of devices on a
chip will have most excellent bragging rights, if nothing else!
3.3.3 Light, Not Electrons

Although the speed of light would be a hard limit, the fact is that semiconductor devices
are limited by the speed of electricity rather than that of light, given that electric waves
in semiconductor materials move at between 3 % and 30 % of the speed of light in a
vacuum. The use of copper connections on silicon devices is one way to increase the
speed of electricity, and it is quite possible that additional advances will push closer
still to the actual speed of light. In addition, there have been some experiments with
tiny optical fibers as interconnects within and between chips, based on the fact that the
speed of light in glass is more than 60 % of the speed of light in a vacuum. One obstacle
to such optical fibers is the inefficiency conversion between electricity and light and
vice versa, resulting in both power-consumption and heat-dissipation problems.
That said, absent some fundamental advances in the field of physics, any exponential
increases in the speed of data flow will be sharply limited by the actual speed of light in
a vacuum.
3.3.4 Special-Purpose Accelerators

A general-purpose CPU working on a specialized problem is often spending significant
time and energy doing work that is only tangentially related to the problem at hand. For
example, when taking the dot product of a pair of vectors, a general-purpose CPU will
normally use a loop (possibly unrolled) with a loop counter. Decoding the instructions,
v2023.06.11a
incrementing the loop counter, testing this counter, and branching back to the top of the
loop are in some sense wasted effort: The real goal is instead to multiply corresponding
elements of the two vectors. Therefore, a specialized piece of hardware designed
specifically to multiply vectors could get the job done more quickly and with less energy
consumed.
This is in fact the motivation for the vector instructions present in many commodity
microprocessors. Because these instructions operate on multiple data items simultane-
ously, they would permit a dot product to be computed with less instruction-decode and
loop overhead.
Similarly, specialized hardware can more efficiently encrypt and decrypt, compress
and decompress, encode and decode, and many other tasks besides. Unfortunately, this
efficiency does not come for free. A computer system incorporating this specialized
hardware will contain more transistors, which will consume some power even when not
in use. Software must be modified to take advantage of this specialized hardware, and
this specialized hardware must be sufficiently generally useful that the high up-front
hardware-design costs can be spread over enough users to make the specialized hardware
affordable. In part due to these sorts of economic considerations, specialized hardware
has thus far appeared only for a few application areas, including graphics processing
(GPUs), vector processors (MMX, SSE, and VMX instructions), and, to a lesser extent,
encryption. And even in these areas, it is not always easy to realize the expected
performance gains, for example, due to thermal throttling [Kra17, Lem18, Dow20].
Unlike the server and PC arena, smartphones have long used a wide variety of
hardware accelerators. These hardware accelerators are often used for media decoding,
so much so that a high-end MP3 player might be able to play audio for several minutes—
with its CPU fully powered off the entire time. The purpose of these accelerators is
to improve energy efficiency and thus extend battery life: Special purpose hardware
can often compute more efficiently than can a general-purpose CPU. This is another
example of the principle called out in Section 2.2.3: Generality is almost never free.
Nevertheless, given the end of Moore’s-Law-induced single-threaded performance
increases, it seems safe to assume that increasing varieties of special-purpose hardware
will appear.
3.3.5 Existing Parallel Software
Although multicore CPUs seem to have taken the computing industry by surprise, the
fact remains that shared-memory parallel computer systems have been commercially
available for more than a quarter century. This is more than enough time for significant
parallel software to make its appearance, and it indeed has. Parallel operating systems
are quite commonplace, as are parallel threading libraries, parallel relational database
management systems, and parallel numerical software. Use of existing parallel software
can go a long ways towards solving any parallel-software crisis we might encounter.
Perhaps the most common example is the parallel relational database management
system. It is not unusual for single-threaded programs, often written in high-level
scripting languages, to access a central relational database concurrently. In the resulting
highly parallel system, only the database need actually deal directly with parallelism. A
very nice trick when it works!
v2023.06.11a
3.4. SOFTWARE DESIGN IMPLICATIONS 43
3.4 Software Design Implications
One ship drives east and another west

While the self-same breezes blow;
’Tis the set of the sail and not the gail
That bids them where to go.
Ella Wheeler Wilcox
The values of the ratios in Table 3.1 are critically important, as they limit the efficiency
of a given parallel application. To see this, suppose that the parallel application uses
CAS operations to communicate among threads. These CAS operations will typically
involve a cache miss, that is, assuming that the threads are communicating primarily
with each other rather than with themselves. Suppose further that the unit of work
corresponding to each CAS communication operation takes 300 ns, which is sufficient
time to compute several floating-point transcendental functions. Then about half of the
execution time will be consumed by the CAS communication operations! This in turn
means that a two-CPU system running such a parallel program would run no faster than
a sequential implementation running on a single CPU.
The situation is even worse in the distributed-system case, where the latency of a
single communications operation might take as long as thousands or even millions
of floating-point operations. This illustrates how important it is for communications
operations to be extremely infrequent and to enable very large quantities of processing.
Quick Quiz 3.11: Given that distributed-systems communication is so horribly expensive,
why does anyone bother with such systems?
The lesson should be quite clear: Parallel algorithms must be explicitly designed with
these hardware properties firmly in mind. One approach is to run nearly independent
threads. The less frequently the threads communicate, whether by atomic operations,
locks, or explicit messages, the better the application’s performance and scalability will
be. This approach will be touched on in Chapter 5, explored in Chapter 6, and taken to
its logical extreme in Chapter 8.
Another approach is to make sure that any sharing be read-mostly, which allows the
CPUs’ caches to replicate the read-mostly data, in turn allowing all CPUs fast access.
This approach is touched on in Section 5.2.4, and explored more deeply in Chapter 9.
In short, achieving excellent parallel performance and scalability means striving for
embarrassingly parallel algorithms and implementations, whether by careful choice of
data structures and algorithms, use of existing parallel applications and environments,
or transforming the problem into an embarrassingly parallel form.
Quick Quiz 3.12: OK, if we are going to have to apply distributed-programming techniques to
shared-memory parallel programs, why not just always use these distributed techniques and
dispense with shared memory?
So, to sum up:
1. The good news is that multicore systems are inexpensive and readily available.
2. More good news: The overhead of many synchronization operations is much lower
than it was on parallel systems from the early 2000s.
v2023.06.11a
3. The bad news is that the overhead of cache misses is still high, especially on large
systems.
The remainder of this book describes ways of handling this bad news.
In particular, Chapter 4 will cover some of the low-level tools used for parallel
programming, Chapter 5 will investigate problems and solutions to parallel counting,
and Chapter 6 will discuss design disciplines that promote performance and scalability.
v2023.06.11a
You are only as good as your tools, and your tools
are only as good as you are.
Unknown
Chapter 4
Tools of the Trade
This chapter provides a brief introduction to some basic tools of the parallel-programming
trade, focusing mainly on those available to user applications running on operating
systems similar to Linux. Section 4.1 begins with scripting languages, Section 4.2
describes the multi-process parallelism supported by the POSIX API and touches on
POSIX threads, Section 4.3 presents analogous operations in other environments, and
finally, Section 4.4 helps to choose the tool that will get the job done.
Quick Quiz 4.1: You call these tools??? They look more like low-level synchronization
primitives to me!
Please note that this chapter provides but a brief introduction. More detail is available
from the references (and from the Internet), and more information will be provided in
later chapters.
4.1 Scripting Languages

The supreme excellence is simplicity.
Henry Wadsworth Longfellow, simplified
The Linux shell scripting languages provide simple but effective ways of managing
parallelism. For example, suppose that you had a program compute_it that you needed
to run twice with two different sets of arguments. This can be accomplished using
UNIX shell scripting as follows:
1 compute_it 1 > compute_it.1.out &
3 wait
4 cat compute_it.1.out
Lines 1 and 2 launch two instances of this program, redirecting their output to two
separate files, with the & character directing the shell to run the two instances of the
program in the background. Line 3 waits for both instances to complete, and lines 4
and 5 display their output. The resulting execution is as shown in Figure 4.1: The two
instances of compute_it execute in parallel, wait completes after both of them do,
and then the two instances of cat execute sequentially.
45
v2023.06.11a
46 CHAPTER 4. TOOLS OF THE TRADE
compute_it 1 > compute_it 2 >

compute_it.1.out & compute_it.2.out &
wait
cat compute_it.1.out
cat compute_it.2.out
Figure 4.1: Execution Diagram for Parallel Shell Execution
Quick Quiz 4.2: But this silly shell script isn’t a real parallel program! Why bother with such
trivia???
Quick Quiz 4.3: Is there a simpler way to create a parallel shell script? If so, how? If not,
why not?
For another example, the make software-build scripting language provides a -j option
that specifies how much parallelism should be introduced into the build process. Thus,
typing make -j4 when building a Linux kernel specifies that up to four build steps be
executed concurrently.
It is hoped that these simple examples convince you that parallel programming need
not always be complex or difficult.
Quick Quiz 4.4: But if script-based parallel programming is so easy, why bother with anything
else?
4.2 POSIX Multiprocessing

A camel is a horse designed by committee.
Unknown
This section scratches the surface of the POSIX environment, including pthreads [Ope97],
as this environment is readily available and widely implemented. Section 4.2.1 provides
a glimpse of the POSIX fork() and related primitives, Section 4.2.2 touches on thread
creation and destruction, Section 4.2.3 gives a brief overview of POSIX locking, and,
finally, Section 4.2.4 describes a specific lock which can be used for data that is read by
many threads and only occasionally updated.
4.2.1 POSIX Process Creation and Destruction

Processes are created using the fork() primitive, they may be destroyed using the
kill() primitive, they may destroy themselves using the exit() primitive. A process
v2023.06.11a
4.2. POSIX MULTIPROCESSING 47
Listing 4.1: Using the fork() Primitive

1 pid = fork();
2 if (pid == 0) {
3 /* child */
4 } else if (pid < 0) {
5 /* parent, upon error */
6 perror("fork");
7 exit(EXIT_FAILURE);
8 } else {
9 /* parent, pid == child ID */
10 }
Listing 4.2: Using the wait() Primitive

1 static __inline__ void waitall(void)
2 {
3 int pid;
4 int status;
5
6 for (;;) {
7 pid = wait(&status);
8 if (pid == -1) {
9 if (errno == ECHILD)
10 break;
11 perror("wait");
13 }
14 }
15 }
executing a fork() primitive is said to be the “parent” of the newly created process.
A parent may wait on its children using the wait() primitive.
Please note that the examples in this section are quite simple. Real-world applications
using these primitives might need to manipulate signals, file descriptors, shared memory
segments, and any number of other resources. In addition, some applications need to
take specific actions if a given child terminates, and might also need to be concerned
with the reason that the child terminated. These issues can of course add substantial
complexity to the code. For more information, see any of a number of textbooks on the
subject [Ste92, Wei13].
If fork() succeeds, it returns twice, once for the parent and again for the child.
The value returned from fork() allows the caller to tell the difference, as shown in
Listing 4.1 (forkjoin.c). Line 1 executes the fork() primitive, and saves its return
value in local variable pid. Line 2 checks to see if pid is zero, in which case, this is the
child, which continues on to execute line 3. As noted earlier, the child may terminate
via the exit() primitive. Otherwise, this is the parent, which checks for an error return
from the fork() primitive on line 4, and prints an error and exits on lines 5–7 if so.
Otherwise, the fork() has executed successfully, and the parent therefore executes
line 9 with the variable pid containing the process ID of the child.
The parent process may use the wait() primitive to wait for its children to complete.
However, use of this primitive is a bit more complicated than its shell-script counterpart,
as each invocation of wait() waits for but one child process. It is therefore customary
to wrap wait() into a function similar to the waitall() function shown in Listing 4.2
(api-pthreads.h), with this waitall() function having semantics similar to the
shell-script wait command. Each pass through the loop spanning lines 6–14 waits on
one child process. Line 7 invokes the wait() primitive, which blocks until a child
process exits, and returns that child’s process ID. If the process ID is instead −1, this
indicates that the wait() primitive was unable to wait on a child. If so, line 9 checks
v2023.06.11a
Listing 4.3: Processes Created Via fork() Do Not Share Memory

1 int x = 0;
2
3 int main(int argc, char *argv[])
4 {
5 int pid;
6
7 pid = fork();
8 if (pid == 0) { /* child */
9 x = 1;
10 printf("Child process set x=1\n");
11 exit(EXIT_SUCCESS);
12 }
13 if (pid < 0) { /* parent, upon error */
14 perror("fork");
16 }
17
18 /* parent */
19
20 waitall();
21 printf("Parent process sees x=%d\n", x);
22
23 return EXIT_SUCCESS;
24 }
for the ECHILD errno, which indicates that there are no more child processes, so that
line 10 exits the loop. Otherwise, lines 11 and 12 print an error and exit.
Quick Quiz 4.5: Why does this wait() primitive need to be so complicated? Why not just
make it work like the shell-script wait does?
It is critically important to note that the parent and child do not share memory. This
is illustrated by the program shown in Listing 4.3 (forkjoinvar.c), in which the child
sets a global variable x to 1 on line 9, prints a message on line 10, and exits on line 11.
The parent continues at line 20, where it waits on the child, and on line 21 finds that its
copy of the variable x is still zero. The output is thus as follows:
Child process set x=1

Parent process sees x=0
Quick Quiz 4.6: Isn’t there a lot more to fork() and wait() than discussed here?
The finest-grained parallelism requires shared memory, and this is covered in

Section 4.2.2. That said, shared-memory parallelism can be significantly more complex
than fork-join parallelism.
4.2.2 POSIX Thread Creation and Destruction

To create a thread within an existing process, invoke the pthread_create() primitive,
for example, as shown on lines 16 and 17 of Listing 4.4 (pcreate.c). The first argument
is a pointer to a pthread_t in which to store the ID of the thread to be created, the
second NULL argument is a pointer to an optional pthread_attr_t, the third argument
is the function (in this case, mythread()) that is to be invoked by the new thread, and
the last NULL argument is the argument that will be passed to mythread().
In this example, mythread() simply returns, but it could instead call pthread_
exit().
v2023.06.11a
Listing 4.4: Threads Created Via pthread_create() Share Memory

1 int x = 0;
2
3 void *mythread(void *arg)
4 {
5 x = 1;
6 printf("Child process set x=1\n");
7 return NULL;
8 }
9
11 {
12 int en;
13 pthread_t tid;
14 void *vp;
15
16 if ((en = pthread_create(&tid, NULL,
17 mythread, NULL)) != 0) {
18 fprintf(stderr, "pthread_create: %s\n", strerror(en));
20 }
21
22 /* parent */
23
24 if ((en = pthread_join(tid, &vp)) != 0) {
25 fprintf(stderr, "pthread_join: %s\n", strerror(en));
27 }
28 printf("Parent process sees x=%d\n", x);
29
30 return EXIT_SUCCESS;
31 }
Quick Quiz 4.7: If the mythread() function in Listing 4.4 can simply return, why bother
with pthread_exit()?
The pthread_join() primitive, shown on line 24, is analogous to the fork-join

wait() primitive. It blocks until the thread specified by the tid variable completes
execution, either by invoking pthread_exit() or by returning from the thread’s
top-level function. The thread’s exit value will be stored through the pointer passed as
the second argument to pthread_join(). The thread’s exit value is either the value
passed to pthread_exit() or the value returned by the thread’s top-level function,
depending on how the thread in question exits.
The program shown in Listing 4.4 produces output as follows, demonstrating that
memory is in fact shared between the two threads:
Child process set x=1

Parent process sees x=1
Note that this program carefully makes sure that only one of the threads stores a value
to variable x at a time. Any situation in which one thread might be storing a value to a
given variable while some other thread either loads from or stores to that same variable
is termed a data race. Because the C language makes no guarantee that the results of
a data race will be in any way reasonable, we need some way of safely accessing and
modifying data concurrently, such as the locking primitives discussed in the following
section.
But your data races are benign, you say? Well, maybe they are. But please do
everyone (yourself included) a big favor and read Section 4.3.4.1 very carefully. As
compilers optimize more and more aggressively, there are fewer and fewer truly benign
data races.
v2023.06.11a
Quick Quiz 4.8: If the C language makes no guarantees in presence of a data race, then why
does the Linux kernel have so many data races? Are you trying to tell me that the Linux kernel
is completely broken???
4.2.3 POSIX Locking

The POSIX standard allows the programmer to avoid data races via “POSIX locking”.
POSIX locking features a number of primitives, the most fundamental of which are
pthread_mutex_lock() and pthread_mutex_unlock(). These primitives operate
on locks, which are of type pthread_mutex_t. These locks may be declared statically
and initialized with PTHREAD_MUTEX_INITIALIZER, or they may be allocated dynami-
cally and initialized using the pthread_mutex_init() primitive. The demonstration
code in this section will take the former course.
The pthread_mutex_lock() primitive “acquires” the specified lock, and the
pthread_mutex_unlock() “releases” the specified lock. Because these are “exclusive”
locking primitives, only one thread at a time may “hold” a given lock at a given time.
For example, if a pair of threads attempt to acquire the same lock concurrently, one of
the pair will be “granted” the lock first, and the other will wait until the first thread
releases the lock. A simple and reasonably useful programming model permits a given
data item to be accessed only while holding the corresponding lock [Hoa74].
Quick Quiz 4.9: What if I want several threads to hold the same lock at the same time?
This exclusive-locking property is demonstrated using the code shown in Listing 4.5
(lock.c). Line 1 defines and initializes a POSIX lock named lock_a, while line 2
similarly defines and initializes a lock named lock_b. Line 4 defines and initializes a
shared variable x.
Lines 6–33 define a function lock_reader() which repeatedly reads the shared
variable x while holding the lock specified by arg. Line 12 casts arg to a pointer
to a pthread_mutex_t, as required by the pthread_mutex_lock() and pthread_
mutex_unlock() primitives.
Quick Quiz 4.10: Why not simply make the argument to lock_reader() on line 6 of
Listing 4.5 be a pointer to a pthread_mutex_t?
Quick Quiz 4.11: What is the READ_ONCE() on lines 20 and 47 and the WRITE_ONCE() on
line 47 of Listing 4.5?
Lines 14–18 acquire the specified pthread_mutex_t, checking for errors and exiting
the program if any occur. Lines 19–26 repeatedly check the value of x, printing the
new value each time that it changes. Line 25 sleeps for one millisecond, which allows
this demonstration to run nicely on a uniprocessor machine. Lines 27–31 release
the pthread_mutex_t, again checking for errors and exiting the program if any
occur. Finally, line 32 returns NULL, again to match the function type required by
pthread_create().
Quick Quiz 4.12: Writing four lines of code for each acquisition and release of a pthread_
mutex_t sure seems painful! Isn’t there a better way?
Lines 35–56 of Listing 4.5 show lock_writer(), which periodically updates the
shared variable x while holding the specified pthread_mutex_t. As with lock_
reader(), line 39 casts arg to a pointer to pthread_mutex_t, lines 41–45 acquire
v2023.06.11a
Listing 4.5: Demonstration of Exclusive Locks

1 pthread_mutex_t lock_a = PTHREAD_MUTEX_INITIALIZER;
2 pthread_mutex_t lock_b = PTHREAD_MUTEX_INITIALIZER;
3
4 int x = 0;
5
6 void *lock_reader(void *arg)
7 {
8 int en;
9 int i;
10 int newx = -1;
11 int oldx = -1;
12 pthread_mutex_t *pmlp = (pthread_mutex_t *)arg;
13
14 if ((en = pthread_mutex_lock(pmlp)) != 0) {
15 fprintf(stderr, "lock_reader:pthread_mutex_lock: %s\n",
16 strerror(en));
18 }
19 for (i = 0; i < 100; i++) {
20 newx = READ_ONCE(x);
21 if (newx != oldx) {
22 printf("lock_reader(): x = %d\n", newx);
23 }
24 oldx = newx;
25 poll(NULL, 0, 1);
26 }
27 if ((en = pthread_mutex_unlock(pmlp)) != 0) {
28 fprintf(stderr, "lock_reader:pthread_mutex_unlock: %s\n",
29 strerror(en));
31 }
32 return NULL;
33 }
34
35 void *lock_writer(void *arg)
36 {
37 int en;
38 int i;
39 pthread_mutex_t *pmlp = (pthread_mutex_t *)arg;
40
41 if ((en = pthread_mutex_lock(pmlp)) != 0) {
42 fprintf(stderr, "lock_writer:pthread_mutex_lock: %s\n",
43 strerror(en));
45 }
46 for (i = 0; i < 3; i++) {
47 WRITE_ONCE(x, READ_ONCE(x) + 1);
49 }
50 if ((en = pthread_mutex_unlock(pmlp)) != 0) {
51 fprintf(stderr, "lock_writer:pthread_mutex_unlock: %s\n",
52 strerror(en));
54 }
55 return NULL;
56 }
v2023.06.11a
Listing 4.6: Demonstration of Same Exclusive Lock

1 printf("Creating two threads using same lock:\n");
2 en = pthread_create(&tid1, NULL, lock_reader, &lock_a);
3 if (en != 0) {
6 }
7 en = pthread_create(&tid2, NULL, lock_writer, &lock_a);
8 if (en != 0) {
11 }
12 if ((en = pthread_join(tid1, &vp)) != 0) {
15 }
19 }
Listing 4.7: Demonstration of Different Exclusive Locks

1 printf("Creating two threads w/different locks:\n");
2 x = 0;
3 en = pthread_create(&tid1, NULL, lock_reader, &lock_a);
4 if (en != 0) {
7 }
8 en = pthread_create(&tid2, NULL, lock_writer, &lock_b);
9 if (en != 0) {
12 }
16 }
20 }
the specified lock, and lines 50–54 release it. While holding the lock, lines 46–49
increment the shared variable x, sleeping for five milliseconds between each increment.
Finally, lines 50–54 release the lock.
Listing 4.6 shows a code fragment that runs lock_reader() and lock_writer() as
threads using the same lock, namely, lock_a. Lines 2–6 create a thread running lock_
reader(), and then lines 7–11 create a thread running lock_writer(). Lines 12–19
wait for both threads to complete. The output of this code fragment is as follows:
Creating two threads using same lock:

lock_reader(): x = 0
Because both threads are using the same lock, the lock_reader() thread cannot
see any of the intermediate values of x produced by lock_writer() while holding the
lock.
Quick Quiz 4.13: Is “x = 0” the only possible output from the code fragment shown in
Listing 4.6? If so, why? If not, what other output could appear, and why?
v2023.06.11a
Listing 4.7 shows a similar code fragment, but this time using different locks: lock_
a for lock_reader() and lock_b for lock_writer(). The output of this code
fragment is as follows:
Creating two threads w/different locks:
Because the two threads are using different locks, they do not exclude each other, and
can run concurrently. The lock_reader() function can therefore see the intermediate
values of x stored by lock_writer().
Quick Quiz 4.14: Using different locks could cause quite a bit of confusion, what with
threads seeing each others’ intermediate states. So should well-written parallel programs restrict
themselves to using a single lock in order to avoid this kind of confusion?
Quick Quiz 4.15: In the code shown in Listing 4.7, is lock_reader() guaranteed to see all
the values produced by lock_writer()? Why or why not?
Quick Quiz 4.16: Wait a minute here!!! Listing 4.6 didn’t initialize shared variable x, so why
does it need to be initialized in Listing 4.7?
Although there is quite a bit more to POSIX exclusive locking, these primitives
provide a good start and are in fact sufficient in a great many situations. The next section
takes a brief look at POSIX reader-writer locking.
4.2.4 POSIX Reader-Writer Locking

The POSIX API provides a reader-writer lock, which is represented by a pthread_
rwlock_t. As with pthread_mutex_t, pthread_rwlock_t may be statically initial-
ized via PTHREAD_RWLOCK_INITIALIZER or dynamically initialized via the pthread_
rwlock_init() primitive. The pthread_rwlock_rdlock() primitive read-acquires
the specified pthread_rwlock_t, the pthread_rwlock_wrlock() primitive write-
acquires it, and the pthread_rwlock_unlock() primitive releases it. Only a single
thread may write-hold a given pthread_rwlock_t at any given time, but multiple
threads may read-hold a given pthread_rwlock_t, at least while there is no thread
currently write-holding it.
As you might expect, reader-writer locks are designed for read-mostly situations. In
these situations, a reader-writer lock can provide greater scalability than can an exclusive
lock because the exclusive lock is by definition limited to a single thread holding the
lock at any given time, while the reader-writer lock permits an arbitrarily large number
of readers to concurrently hold the lock. However, in practice, we need to know how
much additional scalability is provided by reader-writer locks.
Listing 4.8 (rwlockscale.c) shows one way of measuring reader-writer lock
scalability. Line 1 shows the definition and initialization of the reader-writer lock, line 2
shows the holdtime argument controlling the time each thread holds the reader-writer
lock, line 3 shows the thinktime argument controlling the time between the release of
the reader-writer lock and the next acquisition, line 4 defines the readcounts array
into which each reader thread places the number of times it acquired the lock, and line 5
defines the nreadersrunning variable, which determines when all reader threads have
started running.
v2023.06.11a
Listing 4.8: Measuring Reader-Writer Lock Scalability

1 pthread_rwlock_t rwl = PTHREAD_RWLOCK_INITIALIZER;
2 unsigned long holdtime = 0;
3 unsigned long thinktime = 0;
4 long long *readcounts;
5 int nreadersrunning = 0;
6
7 #define GOFLAG_INIT 0
8 #define GOFLAG_RUN 1
9 #define GOFLAG_STOP 2
10 char goflag = GOFLAG_INIT;
11
12 void *reader(void *arg)
13 {
14 int en;
15 int i;
16 long long loopcnt = 0;
17 long me = (long)arg;
18
19 __sync_fetch_and_add(&nreadersrunning, 1);
20 while (READ_ONCE(goflag) == GOFLAG_INIT) {
21 continue;
22 }
23 while (READ_ONCE(goflag) == GOFLAG_RUN) {
24 if ((en = pthread_rwlock_rdlock(&rwl)) != 0) {
25 fprintf(stderr,
26 "pthread_rwlock_rdlock: %s\n", strerror(en));
28 }
29 for (i = 1; i < holdtime; i++) {
30 wait_microseconds(1);
31 }
32 if ((en = pthread_rwlock_unlock(&rwl)) != 0) {
33 fprintf(stderr,
34 "pthread_rwlock_unlock: %s\n", strerror(en));
36 }
37 for (i = 1; i < thinktime; i++) {
38 wait_microseconds(1);
39 }
40 loopcnt++;
41 }
42 readcounts[me] = loopcnt;
43 return NULL;
44 }
v2023.06.11a
10
ideal 10000us
1
Critical Section Performance

1000us
0.1
100us
0.01
10us
0.001
1us
0.0001
0 50 100 150 200 250 300 350 400 450
Number of CPUs (Threads)
Figure 4.2: Reader-Writer Lock Scalability vs. Microseconds in Critical Section on

8-Socket System With Intel Xeon Platinum 8176 CPUs @ 2.10GHz
Lines 7–10 define goflag, which synchronizes the start and the end of the test. This
variable is initially set to GOFLAG_INIT, then set to GOFLAG_RUN after all the reader
threads have started, and finally set to GOFLAG_STOP to terminate the test run.
Lines 12–44 define reader(), which is the reader thread. Line 19 atomically
increments the nreadersrunning variable to indicate that this thread is now running,
and lines 20–22 wait for the test to start. The READ_ONCE() primitive forces the
compiler to fetch goflag on each pass through the loop—the compiler would otherwise
be within its rights to assume that the value of goflag would never change.
Quick Quiz 4.17: Instead of using READ_ONCE() everywhere, why not just declare goflag
as volatile on line 10 of Listing 4.8?
Quick Quiz 4.18: READ_ONCE() only affects the compiler, not the CPU. Don’t we also need
memory barriers to make sure that the change in goflag’s value propagates to the CPU in a
timely fashion in Listing 4.8?
Quick Quiz 4.19: Would it ever be necessary to use READ_ONCE() when accessing a per-thread
variable, for example, a variable declared using GCC’s __thread storage class?
The loop spanning lines 23–41 carries out the performance test. Lines 24–28 acquire
the lock, lines 29–31 hold the lock for the specified number of microseconds, lines 32–36
release the lock, and lines 37–39 wait for the specified number of microseconds before
re-acquiring the lock. Line 40 counts this lock acquisition.
Line 42 moves the lock-acquisition count to this thread’s element of the
readcounts[] array, and line 43 returns, terminating this thread.
Figure 4.2 shows the results of running this test on a 224-core Xeon system with two
hardware threads per core for a total of 448 software-visible CPUs. The thinktime
parameter was zero for all these tests, and the holdtime parameter set to values ranging
from one microsecond (“1us” on the graph) to 10,000 microseconds (“10000us” on the
v2023.06.11a
graph). The actual value plotted is:

𝐿𝑁
(4.1)
𝑁 𝐿1
where 𝑁 is the number of threads in the current run, 𝐿 𝑁 is the total number of lock
acquisitions by all 𝑁 threads in the current run, and 𝐿 1 is the number of lock acquisitions
in a single-threaded run. Given ideal hardware and software scalability, this value will
always be 1.0.
As can be seen in the figure, reader-writer locking scalability is decidedly non-ideal,
especially for smaller sizes of critical sections. To see why read-acquisition can be
so slow, consider that all the acquiring threads must update the pthread_rwlock_t
data structure. Therefore, if all 448 executing threads attempt to read-acquire the
reader-writer lock concurrently, they must update this underlying pthread_rwlock_t
one at a time. One lucky thread might do so almost immediately, but the least-lucky
thread must wait for all the other 447 threads to do their updates. This situation will
only get worse as you add CPUs. Note also the logscale y-axis. Even though the
10,000 microsecond trace appears quite ideal, it has in fact degraded by about 10 % from
ideal.
Quick Quiz 4.20: Isn’t comparing against single-CPU throughput a bit harsh?
Quick Quiz 4.21: But one microsecond is not a particularly small size for a critical section.
What do I do if I need a much smaller critical section, for example, one containing only a few
instructions?
Quick Quiz 4.22: The system used is a few years old, and new hardware should be faster. So
why should anyone worry about reader-writer locks being slow?
Despite these limitations, reader-writer locking is quite useful in many cases, for
example when the readers must do high-latency file or network I/O. There are alternatives,
some of which will be presented in Chapters 5 and 9.
4.2.5 Atomic Operations (GCC Classic)

Figure 4.2 shows that the overhead of reader-writer locking is most severe for the smallest
critical sections, so it would be nice to have some other way of protecting tiny critical
sections. One such way uses atomic operations. We have seen an atomic operation
already, namely the __sync_fetch_and_add() primitive on line 19 of Listing 4.8.
This primitive atomically adds the value of its second argument to the value referenced
by its first argument, returning the old value (which was ignored in this case). If a pair
of threads concurrently execute __sync_fetch_and_add() on the same variable, the
resulting value of the variable will include the result of both additions.
The GNU C compiler offers a number of additional atomic operations, includ-
ing __sync_fetch_and_sub(), __sync_fetch_and_or(), __sync_fetch_and_
and(), __sync_fetch_and_xor(), and __sync_fetch_and_nand(), all of which
return the old value. If you instead need the new value, you can instead use the __
sync_add_and_fetch(), __sync_sub_and_fetch(), __sync_or_and_fetch(),
__sync_and_and_fetch(), __sync_xor_and_fetch(), and __sync_nand_and_
fetch() primitives.
Quick Quiz 4.23: Is it really necessary to have both sets of primitives?
v2023.06.11a
Listing 4.9: Compiler Barrier Primitive (for GCC)

#define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))
#define READ_ONCE(x) \
({ typeof(x) ___x = ACCESS_ONCE(x); ___x; })
#define WRITE_ONCE(x, val) \
do { ACCESS_ONCE(x) = (val); } while (0)
#define barrier() __asm__ __volatile__("": : :"memory")
The classic compare-and-swap operation is provided by a pair of primitives, __

sync_bool_compare_and_swap() and __sync_val_compare_and_swap(). Both
of these primitives atomically update a location to a new value, but only if its prior value
was equal to the specified old value. The first variant returns 1 if the operation succeeded
and 0 if it failed, for example, if the prior value was not equal to the specified old value.
The second variant returns the prior value of the location, which, if equal to the specified
old value, indicates that the operation succeeded. Either of the compare-and-swap
operation is “universal” in the sense that any atomic operation on a single location can
be implemented in terms of compare-and-swap, though the earlier operations are often
more efficient where they apply. The compare-and-swap operation is also capable of
serving as the basis for a wider set of atomic operations, though the more elaborate of
these often suffer from complexity, scalability, and performance problems [Her90].
Quick Quiz 4.24: Given that these atomic operations will often be able to generate single
atomic instructions that are directly supported by the underlying instruction set, shouldn’t they
be the fastest possible way to get things done?
The __sync_synchronize() primitive issues a “memory barrier”, which constrains

both the compiler’s and the CPU’s ability to reorder operations, as discussed in Chapter 15.
In some cases, it is sufficient to constrain the compiler’s ability to reorder operations,
while allowing the CPU free rein, in which case the barrier() primitive may be used.
In some cases, it is only necessary to ensure that the compiler avoids optimizing away a
given memory read, in which case the READ_ONCE() primitive may be used, as it was on
line 20 of Listing 4.5. Similarly, the WRITE_ONCE() primitive may be used to prevent
the compiler from optimizing away a given memory write. These last three primitives
are not provided directly by GCC, but may be implemented straightforwardly as shown
in Listing 4.9, and all three are discussed at length in Section 4.3.4. Alternatively,
READ_ONCE(x) has much in common with the GCC intrinsic __atomic_load_n(&x,
__ATOMIC_RELAXED) and WRITE_ONCE() has much in common with the GCC intrinsic
__atomic_store_n(&x, v, __ATOMIC_RELAXED).
Quick Quiz 4.25: What happened to ACCESS_ONCE()?
4.2.6 Atomic Operations (C11)

The C11 standard added atomic operations, including loads (atomic_load()),
stores (atomic_store()), memory barriers (atomic_thread_fence() and
atomic_signal_fence()), and read-modify-write atomics. The read-modify-
write atomics include atomic_fetch_add(), atomic_fetch_sub(), atomic_
fetch_and(), atomic_fetch_xor(), atomic_exchange(), atomic_compare_
exchange_strong(), and atomic_compare_exchange_weak(). These operate
in a manner similar to those described in Section 4.2.5, but with the addition of memory-
order arguments to _explicit variants of all of the operations. Without memory-order
v2023.06.11a
arguments, all the atomic operations are fully ordered, and the arguments permit weaker
orderings. For example, “atomic_load_explicit(&a, memory_order_relaxed)”
is vaguely similar to the Linux kernel’s “READ_ONCE()”.1
4.2.7 Atomic Operations (Modern GCC)

One restriction of the C11 atomics is that they apply only to special atomic types,
which can be problematic. The GNU C compiler therefore provides atomic in-
trinsics, including __atomic_load(), __atomic_load_n(), __atomic_store(),
__atomic_store_n(), __atomic_thread_fence(), etc. These intrinsics offer the
same semantics as their C11 counterparts, but may be used on plain non-atomic objects.
Some of these intrinsics may be passed a memory-order argument from this list: __
ATOMIC_RELAXED, __ATOMIC_CONSUME, __ATOMIC_ACQUIRE, __ATOMIC_RELEASE,
__ATOMIC_ACQ_REL, and __ATOMIC_SEQ_CST.
4.2.8 Per-Thread Variables

Per-thread variables, also called thread-specific data, thread-local storage, and other
less-polite names, are used extremely heavily in concurrent code, as will be explored
in Chapters 5 and 8. POSIX supplies the pthread_key_create() function to create
a per-thread variable (and return the corresponding key), pthread_key_delete()
to delete the per-thread variable corresponding to key, pthread_setspecific() to
set the value of the current thread’s variable corresponding to the specified key, and
pthread_getspecific() to return that value.
A number of compilers (including GCC) provide a __thread specifier that may be
used in a variable definition to designate that variable as being per-thread. The name
of the variable may then be used normally to access the value of the current thread’s
instance of that variable. Of course, __thread is much easier to use than the POSIX
thead-specific data, and so __thread is usually preferred for code that is to be built
only with GCC or other compilers supporting __thread.
Fortunately, the C11 standard introduced a _Thread_local keyword that can be
used in place of __thread. In the fullness of time, this new keyword should combine
the ease of use of __thread with the portability of POSIX thread-specific data.
4.3 Alternatives to POSIX Operations

The strategic marketing paradigm of Open Source is
a massively parallel drunkard’s walk filtered by a
Darwinistic process.
Bruce Perens
Unfortunately, threading operations, locking primitives, and atomic operations were

in reasonably wide use long before the various standards committees got around
to them. As a result, there is considerable variation in how these operations are
supported. It is still quite common to find these operations implemented in assembly
language, either for historical reasons or to obtain better performance in specialized
circumstances. For example, GCC’s __sync_ family of primitives all provide full
1 Memory ordering is described in more detail in Chapter 15 and Appendix C.
v2023.06.11a
4.3. ALTERNATIVES TO POSIX OPERATIONS 59
Listing 4.10: Thread API

int smp_thread_id(void)
thread_id_t create_thread(void *(*func)(void *), void *arg)
for_each_thread(t)
for_each_running_thread(t)
void *wait_thread(thread_id_t tid)
void wait_all_threads(void)
memory-ordering semantics, which in the past motivated many developers to create

their own implementations for situations where the full memory ordering semantics are
not required. The following sections show some alternatives from the Linux kernel and
some historical primitives used by this book’s sample code.
4.3.1 Organization and Initialization

Although many environments do not require any special initialization code, the code
samples in this book start with a call to smp_init(), which initializes a mapping from
pthread_t to consecutive integers. The userspace RCU library2 similarly requires a
call to rcu_init(). Although these calls can be hidden in environments (such as that
of GCC) that support constructors, most of the RCU flavors supported by the userspace
RCU library also require each thread invoke rcu_register_thread() upon thread
creation and rcu_unregister_thread() before thread exit.
In the case of the Linux kernel, it is a philosophical question as to whether the kernel
does not require calls to special initialization code or whether the kernel’s boot-time
code is in fact the required initialization code.
4.3.2 Thread Creation, Destruction, and Control

The Linux kernel uses struct task_struct pointers to track kthreads, kthread_
create() to create them, kthread_should_stop() to externally suggest that they
stop (which has no POSIX equivalent),3 kthread_stop() to wait for them to stop,
and schedule_timeout_interruptible() for a timed wait. There are quite a few
additional kthread-management APIs, but this provides a good start, as well as good
search terms.
The CodeSamples API focuses on “threads”, which are a locus of control.4 Each
such thread has an identifier of type thread_id_t, and no two threads running at a
given time will have the same identifier. Threads share everything except for per-thread
local state,5 which includes program counter and stack.
The thread API is shown in Listing 4.10, and members are described in the following
section.
4.3.2.1 API Members

create_thread()
The create_thread() primitive creates a new thread, starting the new thread’s
2 See Section 9.5 for more information on RCU.

3 POSIX environments can work around the lack of kthread_should_stop() by using
a properly synchronized boolean flag in conjunction with pthread_join().
4 There are many other names for similar software constructs, including “process”, “task”,
“fiber”, “event”, “execution agent”, and so on. Similar design principles apply to all of them.
5 How is that for a circular definition?
v2023.06.11a
execution at the function func specified by create_thread()’s first argument,

and passing it the argument specified by create_thread()’s second argument.
This newly created thread will terminate when it returns from the starting function
specified by func. The create_thread() primitive returns the thread_id_t
corresponding to the newly created child thread.
This primitive will abort the program if more than NR_THREADS threads are created,
counting the one implicitly created by running the program. NR_THREADS is a
compile-time constant that may be modified, though some systems may have an
upper bound for the allowable number of threads.
smp_thread_id()
Because the thread_id_t returned from create_thread() is system-dependent,
the smp_thread_id() primitive returns a thread index corresponding to the thread
making the request. This index is guaranteed to be less than the maximum number
of threads that have been in existence since the program started, and is therefore
useful for bitmasks, array indices, and the like.
for_each_thread()
The for_each_thread() macro loops through all threads that exist, including all
threads that would exist if created. This macro is useful for handling the per-thread
variables introduced in Section 4.2.8.
for_each_running_thread()
The for_each_running_thread() macro loops through only those threads that
currently exist. It is the caller’s responsibility to synchronize with thread creation
and deletion if required.
wait_thread()
The wait_thread() primitive waits for completion of the thread specified by the
thread_id_t passed to it. This in no way interferes with the execution of the
specified thread; instead, it merely waits for it. Note that wait_thread() returns
the value that was returned by the corresponding thread.
wait_all_threads()
The wait_all_threads() primitive waits for completion of all currently running
threads. It is the caller’s responsibility to synchronize with thread creation and
deletion if required. However, this primitive is normally used to clean up at the
end of a run, so such synchronization is normally not needed.
4.3.2.2 Example Usage

Listing 4.11 (threadcreate.c) shows an example hello-world-like child thread. As
noted earlier, each thread is allocated its own stack, so each thread has its own private
arg argument and myarg variable. Each child simply prints its argument and its
smp_thread_id() before exiting. Note that the return statement on line 7 terminates
the thread, returning a NULL to whoever invokes wait_thread() on this thread.
The parent program is shown in Listing 4.12. It invokes smp_init() to initialize the
threading system on line 6, parses arguments on lines 8–15, and announces its presence
on line 16. It creates the specified number of child threads on lines 18–19, and waits
for them to complete on line 21. Note that wait_all_threads() discards the threads
return values, as in this case they are all NULL, which is not very interesting.
v2023.06.11a
Listing 4.11: Example Child Thread

1 void *thread_test(void *arg)
2 {
3 int myarg = (intptr_t)arg;
4
5 printf("child thread %d: smp_thread_id() = %d\n",
6 myarg, smp_thread_id());
7 return NULL;
8 }
Listing 4.12: Example Parent Thread

2 {
3 int i;
4 int nkids = 1;
5
6 smp_init();
7
8 if (argc > 1) {
9 nkids = strtoul(argv[1], NULL, 0);
10 if (nkids > NR_THREADS) {
11 fprintf(stderr, "nkids = %d too large, max = %d\n",
12 nkids, NR_THREADS);
13 usage(argv[0]);
14 }
15 }
16 printf("Parent thread spawning %d threads.\n", nkids);
17
18 for (i = 0; i < nkids; i++)
19 create_thread(thread_test, (void *)(intptr_t)i);
20
21 wait_all_threads();
22
23 printf("All spawned threads completed.\n");
24
25 exit(0);
26 }
Listing 4.13: Locking API

void spin_lock_init(spinlock_t *sp);
void spin_lock(spinlock_t *sp);
int spin_trylock(spinlock_t *sp);
void spin_unlock(spinlock_t *sp);
Quick Quiz 4.26: What happened to the Linux-kernel equivalents to fork() and wait()?
4.3.3 Locking
A good starting subset of the Linux kernel’s locking API is shown in Listing 4.13,
each API element being described in the following section. This book’s CodeSamples
locking API closely follows that of the Linux kernel.
4.3.3.1 API Members

spin_lock_init()
The spin_lock_init() primitive initializes the specified spinlock_t variable,
and must be invoked before this variable is passed to any other spinlock primitive.
spin_lock()
The spin_lock() primitive acquires the specified spinlock, if necessary, waiting
v2023.06.11a
Listing 4.14: Living Dangerously Early 1990s Style

1 ptr = global_ptr;
2 if (ptr != NULL && ptr < high_address)
3 do_low(ptr);
Listing 4.15: C Compilers Can Invent Loads

1 if (global_ptr != NULL &&
2 global_ptr < high_address)
3 do_low(global_ptr);
until the spinlock becomes available. In some environments, such as pthreads, this
waiting will involve blocking, while in others, such as the Linux kernel, it might
involve a CPU-bound spin loop.
The key point is that only one thread may hold a spinlock at any given time.
spin_trylock()
The spin_trylock() primitive acquires the specified spinlock, but only if it is
immediately available. It returns true if it was able to acquire the spinlock and
false otherwise.
spin_unlock()
The spin_unlock() primitive releases the specified spinlock, allowing other
threads to acquire it.
4.3.3.2 Example Usage

A spinlock named mutex may be used to protect a variable counter as follows:
spin_lock(&mutex);
counter++;
spin_unlock(&mutex);
Quick Quiz 4.27: What problems could occur if the variable counter were incremented
without the protection of mutex?
However, the spin_lock() and spin_unlock() primitives do have performance

consequences, as will be seen in Chapter 10.
4.3.4 Accessing Shared Variables

It was not until 2011 that the C standard defined semantics for concurrent read/write
access to shared variables. However, concurrent C code was being written at least
a quarter century earlier [BK85, Inm85]. This raises the question as to what today’s
greybeards did back in long-past pre-C11 days. A short answer to this question is “they
lived dangerously”.
At least they would have been living dangerously had they been using 2021 compilers.
In (say) the early 1990s, compilers did fewer optimizations, in part because there were
fewer compiler writers and in part due to the relatively small memories of that era.
Nevertheless, problems did arise, as shown in Listing 4.14, which the compiler is within
its rights to transform into Listing 4.15. As you can see, the temporary on line 1 of
Listing 4.14 has been optimized away, so that global_ptr will be loaded up to three
times.
v2023.06.11a
Quick Quiz 4.28: What is wrong with loading Listing 4.14’s global_ptr up to three times?
Section 4.3.4.1 describes additional problems caused by plain accesses, Sec-

tions 4.3.4.2 and 4.3.4.3 describe some pre-C11 solutions. Of course, where practical,
direct C-language memory references should be replaced by the primitives described in
Section 4.2.5 or (especially) Section 4.2.6. Use these primitives to avoid data races, that
is, ensure that if there are multiple concurrent C-language accesses to a given variable,
all of those accesses are loads.
4.3.4.1 Shared-Variable Shenanigans

Given code that does plain loads and stores,6 the compiler is within its rights to assume
that the affected variables are neither accessed nor modified by any other thread. This
assumption allows the compiler to carry out a large number of transformations, including
load tearing, store tearing, load fusing, store fusing, code reordering, invented loads,
invented stores, store-to-load transformations, and dead-code elimination, all of which
work just fine in single-threaded code. But concurrent code can be broken by each of
these transformations, or shared-variable shenanigans, as described below.
Load tearing occurs when the compiler uses multiple load instructions for a single
access. For example, the compiler could in theory compile the load from global_
ptr (see line 1 of Listing 4.14) as a series of one-byte loads. If some other thread
was concurrently setting global_ptr to NULL, the result might have all but one
byte of the pointer set to zero, thus forming a “wild pointer”. Stores using such
a wild pointer could corrupt arbitrary regions of memory, resulting in rare and
difficult-to-debug crashes.
Worse yet, on (say) an 8-bit system with 16-bit pointers, the compiler might have
no choice but to use a pair of 8-bit instructions to access a given pointer. Because
the C standard must support all manner of systems, the standard cannot rule out
load tearing in the general case.
Store tearing occurs when the compiler uses multiple store instructions for a single
access. For example, one thread might store 0x12345678 to a four-byte integer
variable at the same time another thread stored 0xabcdef00. If the compiler
used 16-bit stores for either access, the result might well be 0x1234ef00, which
could come as quite a surprise to code loading from this integer. Nor is this a
strictly theoretical issue. For example, there are CPUs that feature small immediate
instruction fields, and on such CPUs, the compiler might split a 64-bit store
into two 32-bit stores in order to reduce the overhead of explicitly forming the
64-bit constant in a register, even on a 64-bit CPU. There are historical reports
of this actually happening in the wild (e.g. [KM13]), but there is also a recent
report [Dea19].7
Of course, the compiler simply has no choice but to tear some stores in the general
case, given the possibility of code using 64-bit integers running on a 32-bit system.
6 That is, normal loads and stores instead of C11 atomics, inline assembly, or volatile
accesses.
7 Note that this tearing can happen even on properly aligned and machine-word-sized
accesses, and in this particular case, even for volatile stores. Some might argue that this
behavior constitutes a bug in the compiler, but either way it illustrates the perceived value of
store tearing from a compiler-writer viewpoint.
v2023.06.11a
Listing 4.16: Inviting Load Fusing

1 while (!need_to_stop)
2 do_something_quickly();
Listing 4.17: C Compilers Can Fuse Loads

1 if (!need_to_stop)
2 for (;;) {
19 }
But for properly aligned machine-sized stores, WRITE_ONCE() will prevent store
tearing.
Load fusing occurs when the compiler uses the result of a prior load from a given
variable instead of repeating the load. Not only is this sort of optimization just fine
in single-threaded code, it is often just fine in multithreaded code. Unfortunately,
the word “often” hides some truly annoying exceptions.
For example, suppose that a real-time system needs to invoke a function named
do_something_quickly() repeatedly until the variable need_to_stop was set,
and that the compiler can see that do_something_quickly() does not store to
need_to_stop. One (unsafe) way to code this is shown in Listing 4.16. The
compiler might reasonably unroll this loop sixteen times in order to reduce the
per-invocation of the backwards branch at the end of the loop. Worse yet, because
the compiler knows that do_something_quickly() does not store to need_to_
stop, the compiler could quite reasonably decide to check this variable only once,
resulting in the code shown in Listing 4.17. Once entered, the loop on lines 2–19
will never exit, regardless of how many times some other thread stores a non-zero
value to need_to_stop. The result will at best be consternation, and might well
also include severe physical damage.
The compiler can fuse loads across surprisingly large spans of code. For example,
in Listing 4.18, t0() and t1() run concurrently, and do_something() and
do_something_else() are inline functions. Line 1 declares pointer gp, which
C initializes to NULL by default. At some point, line 5 of t0() stores a non-NULL
pointer to gp. Meanwhile, t1() loads from gp three times on lines 10, 12, and 15.
Given that line 13 finds that gp is non-NULL, one might hope that the dereference
on line 15 would be guaranteed never to fault. Unfortunately, the compiler is within
its rights to fuse the read on lines 10 and 15, which means that if line 10 loads
NULL and line 12 loads &myvar, line 15 could load NULL, resulting in a fault.8
8 Will Deacon reports that this happened in the Linux kernel.
v2023.06.11a
Listing 4.18: C Compilers Can Fuse Non-Adjacent Loads

1 int *gp;
2
3 void t0(void)
4 {
5 WRITE_ONCE(gp, &myvar);
6 }
7
8 void t1(void)
9 {
10 p1 = gp;
11 do_something(p1);
12 p2 = READ_ONCE(gp);
13 if (p2) {
14 do_something_else();
15 p3 = *gp;
16 }
17 }
Listing 4.19: C Compilers Can Fuse Stores

1 void shut_it_down(void)
2 {
3 status = SHUTTING_DOWN; /* BUGGY!!! */
4 start_shutdown();
5 while (!other_task_ready) /* BUGGY!!! */
6 continue;
7 finish_shutdown();
8 status = SHUT_DOWN; /* BUGGY!!! */
10 }
11
12 void work_until_shut_down(void)
13 {
14 while (status != SHUTTING_DOWN) /* BUGGY!!! */
15 do_more_work();
16 other_task_ready = 1; /* BUGGY!!! */
17 }
Note that the intervening READ_ONCE() does not prevent the other two loads from
being fused, despite the fact that all three are loading from the same variable.
Quick Quiz 4.29: Why does it matter whether do_something() and do_something_
else() in Listing 4.18 are inline functions?
Store fusing can occur when the compiler notices a pair of successive stores to a given
variable with no intervening loads from that variable. In this case, the compiler is
within its rights to omit the first store. This is never a problem in single-threaded
code, and in fact it is usually not a problem in correctly written concurrent code.
After all, if the two stores are executed in quick succession, there is very little
chance that some other thread could load the value from the first store.
However, there are exceptions, for example as shown in Listing 4.19. The function
shut_it_down() stores to the shared variable status on lines 3 and 8, and
so assuming that neither start_shutdown() nor finish_shutdown() access
status, the compiler could reasonably remove the store to status on line 3.
Unfortunately, this would mean that work_until_shut_down() would never exit
its loop spanning lines 14 and 15, and thus would never set other_task_ready,
which would in turn mean that shut_it_down() would never exit its loop spanning
lines 5 and 6, even if the compiler chooses not to fuse the successive loads from
other_task_ready on line 5.
v2023.06.11a
Listing 4.20: Inviting an Invented Store

1 if (condition)
2 a = 1;
3 else
4 do_a_bunch_of_stuff(&a);
And there are more problems with the code in Listing 4.19, including code
reordering.
Code reordering is a common compilation technique used to combine common
subexpressions, reduce register pressure, and improve utilization of the many
functional units available on modern superscalar microprocessors. It is also
another reason why the code in Listing 4.19 is buggy. For example, suppose that
the do_more_work() function on line 15 does not access other_task_ready.
Then the compiler would be within its rights to move the assignment to other_
task_ready on line 16 to precede line 14, which might be a great disappointment
for anyone hoping that the last call to do_more_work() on line 15 happens before
the call to finish_shutdown() on line 7.
It might seem futile to prevent the compiler from changing the order of accesses in
cases where the underlying hardware is free to reorder them. However, modern
machines have exact exceptions and exact interrupts, meaning that any interrupt or
exception will appear to have happened at a specific place in the instruction stream.
This means that the handler will see the effect of all prior instructions, but won’t see
the effect of any subsequent instructions. READ_ONCE() and WRITE_ONCE() can
therefore be used to control communication between interrupted code and interrupt
handlers, independent of the ordering provided by the underlying hardware.9
Invented loads were illustrated by the code in Listings 4.14 and 4.15, in which the
compiler optimized away a temporary variable, thus loading from a shared variable
more often than intended.
Invented loads can be a performance hazard. These hazards can occur when a load
of variable in a “hot” cacheline is hoisted out of an if statement. These hoisting
optimizations are not uncommon, and can cause significant increases in cache
misses, and thus significant degradation of both performance and scalability.
Invented stores can occur in a number of situations. For example, a compiler emitting
code for work_until_shut_down() in Listing 4.19 might notice that other_
task_ready is not accessed by do_more_work(), and stored to on line 16. If
do_more_work() was a complex inline function, it might be necessary to do a
register spill, in which case one attractive place to use for temporary storage is
other_task_ready. After all, there are no accesses to it, so what is the harm?
Of course, a non-zero store to this variable at just the wrong time would result
in the while loop on line 5 terminating prematurely, again allowing finish_
shutdown() to run concurrently with do_more_work(). Given that the entire
point of this while appears to be to prevent such concurrency, this is not a good
thing.
9 That said, the various standards committees would prefer that you use atomics or
variables of type sig_atomic_t, instead of READ_ONCE() and WRITE_ONCE().
v2023.06.11a
Listing 4.21: Compiler Invents an Invited Store

1 a = 1;
2 if (!condition) {
3 a = 0;
4 do_a_bunch_of_stuff(&a);
5 }
Listing 4.22: Inviting a Store-to-Load Conversion

1 r1 = p;
2 if (unlikely(r1))
3 do_something_with(r1);
4 barrier();
5 p = NULL;
Listing 4.23: Compiler Converts a Store to a Load

1 r1 = p;
2 if (unlikely(r1))
3 do_something_with(r1);
4 barrier();
5 if (p != NULL)
6 p = NULL;
Using a stored-to variable as a temporary might seem outlandish, but it is permitted

by the standard. Nevertheless, readers might be justified in wanting a less outlandish
example, which is provided by Listings 4.20 and 4.21.
A compiler emitting code for Listing 4.20 might know that the value of a is
initially zero, which might be a strong temptation to optimize away one branch by
transforming this code to that in Listing 4.21. Here, line 1 unconditionally stores 1
to a, then resets the value back to zero on line 3 if condition was not set. This
transforms the if-then-else into an if-then, saving one branch.
Quick Quiz 4.30: Ouch! So can’t the compiler invent a store to a normal variable pretty
much any time it likes?
Finally, pre-C11 compilers could invent writes to unrelated variables that happened
to be adjacent to written-to variables [Boe05, Section 4.2]. This variant of invented
stores has been outlawed by the prohibition against compiler optimizations that
invent data races.
Store-to-load transformations can occur when the compiler notices that a plain store
might not actually change the value in memory. For example, consider Listing 4.22.
Line 1 fetches p, but the “if” statement on line 2 also tells the compiler that the
developer thinks that p is usually zero.10 The barrier() statement on line 4
forces the compiler to forget the value of p, but one could imagine a compiler
choosing to remember the hint—or getting an additional hint via feedback-directed
optimization. Doing so would cause the compiler to realize that line 5 is often an
expensive no-op.
Such a compiler might therefore guard the store of NULL with a check, as shown
on lines 5 and 6 of Listing 4.23. Although this transformation is often desirable, it
could be problematic if the actual store was required for ordering. For example,
a write memory barrier (Linux kernel smp_wmb()) would order the store, but
10 The unlikely() function provides this hint to the compiler, and different compilers
provide different ways of implementing unlikely().
v2023.06.11a
not the load. This situation might suggest use of smp_store_release() over
smp_wmb().
Dead-code elimination can occur when the compiler notices that the value from a
load is never used, or when a variable is stored to, but never loaded from. This
can of course eliminate an access to a shared variable, which can in turn defeat
a memory-ordering primitive, which could cause your concurrent code to act in
surprising ways. Experience thus far indicates that relatively few such surprises
will be at all pleasant. Elimination of store-only variables is especially dangerous
in cases where external code locates the variable via symbol tables: The compiler
is necessarily ignorant of such external-code accesses, and might thus eliminate a
variable that the external code relies upon.
Reliable concurrent code clearly needs a way to cause the compiler to preserve the
number, order, and type of important accesses to shared memory, a topic taken up by
Sections 4.3.4.2 and 4.3.4.3, which are up next.
4.3.4.2 A Volatile Solution

Although it is now much maligned, before the advent of C11 and C++11 [Bec11], the
volatile keyword was an indispensable tool in the parallel programmer’s toolbox. This
raises the question of exactly what volatile means, a question that is not answered
with excessive precision even by more recent versions of this standard [Smi19].11
This version guarantees that “Accesses through volatile glvalues are evaluated
strictly according to the rules of the abstract machine”, that volatile accesses are side
effects, that they are one of the four forward-progress indicators, and that their exact
semantics are implementation-defined. Perhaps the clearest guidance is provided by
this non-normative note:
volatile is a hint to the implementation to avoid aggressive optimization

involving the object because the value of the object might be changed by means
undetectable by an implementation. Furthermore, for some implementations,
volatile might indicate that special hardware instructions are required to
access the object. See 6.8.1 for detailed semantics. In general, the semantics
of volatile are intended to be the same in C++ as they are in C.
This wording might be reassuring to those writing low-level code, except for the
fact that compiler writers are free to completely ignore non-normative notes. Parallel
programmers might instead reassure themselves that compiler writers would like to
avoid breaking device drivers (though perhaps only after a few “frank and open”
discussions with device-driver developers), and device drivers impose at least the
following constraints [MWPF18]:
1. Implementations are forbidden from tearing an aligned volatile access when

machine instructions of that access’s size and type are available.12 Concurrent
code relies on this constraint to avoid unnecessary load and store tearing.
11 JF Bastien thoroughly documented the history and use cases for the volatile keyword
in C++ [Bas18].
12 Note that this leaves unspecified what to do with 128-bit loads and stores on CPUs
having 128-bit CAS but not 128-bit loads and stores.
v2023.06.11a
Listing 4.24: Avoiding Danger, 2018 Style

1 ptr = READ_ONCE(global_ptr);
2 if (ptr != NULL && ptr < high_address)
3 do_low(ptr);
Listing 4.25: Preventing Load Fusing

1 while (!READ_ONCE(need_to_stop))
Listing 4.26: Preventing Store Fusing and Invented Stores

2 {
3 WRITE_ONCE(status, SHUTTING_DOWN); /* BUGGY!!! */
4 start_shutdown();
5 while (!READ_ONCE(other_task_ready)) /* BUGGY!!! */
6 continue;
8 WRITE_ONCE(status, SHUT_DOWN); /* BUGGY!!! */
10 }
11
13 {
14 while (READ_ONCE(status) != SHUTTING_DOWN) /* BUGGY!!! */
15 do_more_work();
16 WRITE_ONCE(other_task_ready, 1); /* BUGGY!!! */
17 }
2. Implementations must not assume anything about the semantics of a volatile

access, nor, for any volatile access that returns a value, about the possible set
of values that might be returned.13 Concurrent code relies on this constraint to
avoid optimizations that are inapplicable given that other processors might be
concurrently accessing the location in question.
3. Aligned machine-sized non-mixed-size volatile accesses interact naturally with
volatile assembly-code sequences before and after. This is necessary because some
devices must be accessed using a combination of volatile MMIO accesses and
special-purpose assembly-language instructions. Concurrent code relies on this
constraint in order to achieve the desired ordering properties from combinations of
volatile accesses and other means discussed in Section 4.3.4.3.
Concurrent code also relies on the first two constraints to avoid undefined behavior
that could result due to data races if any of the accesses to a given object was either
non-atomic or non-volatile, assuming that all accesses are aligned and machine-sized.
The semantics of mixed-size accesses to the same locations are more complex, and are
left aside for the time being.
So how does volatile stack up against the earlier examples?
Using READ_ONCE() on line 1 of Listing 4.14 avoids invented loads, resulting in the
code shown in Listing 4.24.
As shown in Listing 4.25, READ_ONCE() can also prevent the loop unrolling in
Listing 4.17.
READ_ONCE() and WRITE_ONCE() can also be used to prevent the store fusing and
invented stores that were shown in Listing 4.19, with the result shown in Listing 4.26.
13 This is strongly implied by the implementation-defined semantics called out above.
v2023.06.11a
Listing 4.27: Disinviting an Invented Store

1 if (condition)
2 WRITE_ONCE(a, 1);
3 else
4 do_a_bunch_of_stuff();
Listing 4.28: Preventing C Compilers From Fusing Loads

1 while (!need_to_stop) {
2 barrier();
4 barrier();
5 }
However, this does nothing to prevent code reordering, which requires some additional
tricks taught in Section 4.3.4.3.
Finally, WRITE_ONCE() can be used to prevent the store invention shown in List-
ing 4.20, with the resulting code shown in Listing 4.27.
To summarize, the volatile keyword can prevent load tearing and store tearing in
cases where the loads and stores are machine-sized and properly aligned. It can also
prevent load fusing, store fusing, invented loads, and invented stores. However, although
it does prevent the compiler from reordering volatile accesses with each other, it
does nothing to prevent the CPU from reordering these accesses. Furthermore, it does
nothing to prevent either compiler or CPU from reordering non-volatile accesses with
each other or with volatile accesses. Preventing these types of reordering requires
the techniques described in the next section.
4.3.4.3 Assembling the Rest of a Solution

Additional ordering has traditionally been provided by recourse to assembly language, for
example, GCC asm directives. Oddly enough, these directives need not actually contain
assembly language, as exemplified by the barrier() macro shown in Listing 4.9.
In the barrier() macro, the __asm__ introduces the asm directive, the __
volatile__ prevents the compiler from optimizing the asm away, the empty string
specifies that no actual instructions are to be emitted, and the final "memory" tells
the compiler that this do-nothing asm can arbitrarily change memory. In response,
the compiler will avoid moving any memory references across the barrier() macro.
This means that the real-time-destroying loop unrolling shown in Listing 4.17 can be
prevented by adding barrier() calls as shown on lines 2 and 4 of Listing 4.28. These
two lines of code prevent the compiler from pushing the load from need_to_stop into
or past do_something_quickly() from either direction.
However, this does nothing to prevent the CPU from reordering the references. In
many cases, this is not a problem because the hardware can only do a certain amount
of reordering. However, there are cases such as Listing 4.19 where the hardware must
be constrained. Listing 4.26 prevented store fusing and invention, and Listing 4.29
further prevents the remaining reordering by addition of smp_mb() on lines 4, 8, 10,
18, and 21. The smp_mb() macro is similar to barrier() shown in Listing 4.9, but
with the empty string replaced by a string containing the instruction for a full memory
barrier, for example, "mfence" on x86 or "sync" on PowerPC.
Quick Quiz 4.31: But aren’t full memory barriers very heavyweight? Isn’t there a cheaper
way to enforce the ordering needed in Listing 4.29?
v2023.06.11a
Listing 4.29: Preventing Reordering

2 {
3 WRITE_ONCE(status, SHUTTING_DOWN);
4 smp_mb();
5 start_shutdown();
6 while (!READ_ONCE(other_task_ready))
7 continue;
8 smp_mb();
10 smp_mb();
11 WRITE_ONCE(status, SHUT_DOWN);
13 }
14
16 {
17 while (READ_ONCE(status) != SHUTTING_DOWN) {
18 smp_mb();
19 do_more_work();
20 }
21 smp_mb();
22 WRITE_ONCE(other_task_ready, 1);
23 }
Ordering is also provided by some read-modify-write atomic operations, some of

which are presented in Section 4.3.5. In the general case, memory ordering can be quite
subtle, as discussed in Chapter 15. The next section covers an alternative to memory
ordering, namely limiting or even entirely avoiding data races.
4.3.4.4 Avoiding Data Races

“Doctor, it hurts my head when I think about concurrently accessing shared
variables!”
“Then stop concurrently accessing shared variables!!!”
The doctor’s advice might seem unhelpful, but one time-tested way to avoid con-
currently accessing shared variables is access those variables only when holding a
particular lock, as will be discussed in Chapter 7. Another way is to access a given
“shared” variable only from a given CPU or thread, as will be discussed in Chapter 8. It
is possible to combine these two approaches, for example, a given variable might be
modified only by a given CPU or thread while holding a particular lock, and might be
read either from that same CPU or thread on the one hand, or from some other CPU or
thread while holding that same lock on the other. In all of these situations, all accesses
to the shared variables may be plain C-language accesses.
Here is a list of situations allowing plain loads and stores for some accesses to a
given variable, while requiring markings (such as READ_ONCE() and WRITE_ONCE())
for other accesses to that same variable:
1. A shared variable is only modified by a given owning CPU or thread, but is read by
other CPUs or threads. All stores must use WRITE_ONCE(). The owning CPU or
thread may use plain loads. Everything else must use READ_ONCE() for loads.
2. A shared variable is only modified while holding a given lock, but is read by
code not holding that lock. All stores must use WRITE_ONCE(). CPUs or threads
holding the lock may use plain loads. Everything else must use READ_ONCE() for
loads.
v2023.06.11a
3. A shared variable is only modified while holding a given lock by a given owning
CPU or thread, but is read by other CPUs or threads or by code not holding that
lock. All stores must use WRITE_ONCE(). The owning CPU or thread may use
plain loads, as may any CPU or thread holding the lock. Everything else must use
READ_ONCE() for loads.
4. A shared variable is only accessed by a given CPU or thread and by a signal or

interrupt handler running in that CPU’s or thread’s context. The handler can use
plain loads and stores, as can any code that has prevented the handler from being
invoked, that is, code that has blocked signals and/or interrupts. All other code
must use READ_ONCE() and WRITE_ONCE().
5. A shared variable is only accessed by a given CPU or thread and by a signal or

interrupt handler running in that CPU’s or thread’s context, and the handler always
restores the values of any variables that it has written before return. The handler
can use plain loads and stores, as can any code that has prevented the handler from
being invoked, that is, code that has blocked signals and/or interrupts. All other
code can use plain loads, but must use WRITE_ONCE() to prevent store tearing,
store fusing, and invented stores.
Quick Quiz 4.32: What needs to happen if an interrupt or signal handler might itself be
interrupted?
In most other cases, loads from and stores to a shared variable must use READ_
ONCE() and WRITE_ONCE() or stronger, respectively. But it bears repeating that neither
READ_ONCE() nor WRITE_ONCE() provide any ordering guarantees other than within
the compiler. See the above Section 4.3.4.3 or Chapter 15 for information on such
guarantees.
Examples of many of these data-race-avoidance patterns are presented in Chapter 5.
4.3.5 Atomic Operations

The Linux kernel provides a wide variety of atomic operations, but those defined on
type atomic_t provide a good start. Normal non-tearing reads and stores are provided
by atomic_read() and atomic_set(), respectively. Acquire load is provided by
smp_load_acquire() and release store by smp_store_release().
Non-value-returning fetch-and-add operations are provided by atomic_add(),
atomic_sub(), atomic_inc(), and atomic_dec(), among others. An atomic
decrement that returns a reached-zero indication is provided by both atomic_dec_
and_test() and atomic_sub_and_test(). An atomic add that returns the new value
is provided by atomic_add_return(). Both atomic_add_unless() and atomic_
inc_not_zero() provide conditional atomic operations, where nothing happens unless
the original value of the atomic variable is different than the value specified (these are
very handy for managing reference counters, for example).
An atomic exchange operation is provided by atomic_xchg(), and the celebrated
compare-and-swap (CAS) operation is provided by atomic_cmpxchg(). Both of these
return the old value. Many additional atomic RMW primitives are available in the Linux
kernel, see the Documentation/atomic_t.txt file in the Linux-kernel source tree.14
This book’s CodeSamples API closely follows that of the Linux kernel.
14 As of Linux kernel v5.11.
v2023.06.11a
Listing 4.30: Per-Thread-Variable API

DEFINE_PER_THREAD(type, name)
DECLARE_PER_THREAD(type, name)
per_thread(name, thread)
__get_thread_var(name)
init_per_thread(name, v)
4.3.6 Per-CPU Variables

The Linux kernel uses DEFINE_PER_CPU() to define a per-CPU variable, this_
cpu_ptr() to form a reference to this CPU’s instance of a given per-CPU variable,
per_cpu() to access a specified CPU’s instance of a given per-CPU variable, along
with many other special-purpose per-CPU operations.
Listing 4.30 shows this book’s per-thread-variable API, which is patterned after the
Linux kernel’s per-CPU-variable API. This API provides the per-thread equivalent of
global variables. Although this API is, strictly speaking, not necessary,15 it can provide
a good userspace analogy to Linux kernel code.
Quick Quiz 4.33: How could you work around the lack of a per-thread-variable API on
systems that do not provide it?
4.3.6.1 API Members

DEFINE_PER_THREAD()
The DEFINE_PER_THREAD() primitive defines a per-thread variable. Unfortu-
nately, it is not possible to provide an initializer in the way permitted by the Linux
kernel’s DEFINE_PER_CPU() primitive, but there is an init_per_thread()
primitive that permits easy runtime initialization.
DECLARE_PER_THREAD()
The DECLARE_PER_THREAD() primitive is a declaration in the C sense, as opposed
to a definition. Thus, a DECLARE_PER_THREAD() primitive may be used to access
a per-thread variable defined in some other file.
per_thread()
The per_thread() primitive accesses the specified thread’s variable.
__get_thread_var()
The __get_thread_var() primitive accesses the current thread’s variable.
init_per_thread()
The init_per_thread() primitive sets all threads’ instances of the specified
variable to the specified value. The Linux kernel accomplishes this via normal C
initialization, relying in clever use of linker scripts and code executed during the
CPU-online process.
4.3.6.2 Usage Example

Suppose that we have a counter that is incremented very frequently but read out quite
rarely. As will become clear in Section 5.2, it is helpful to implement such a counter
using a per-thread variable. Such a variable can be defined as follows:
15 You could instead use __thread or _Thread_local.
v2023.06.11a
DEFINE_PER_THREAD(int, counter);
The counter must be initialized as follows:

init_per_thread(counter, 0);
A thread can increment its instance of this counter as follows:

p_counter = &__get_thread_var(counter);
WRITE_ONCE(*p_counter, *p_counter + 1);
The value of the counter is then the sum of its instances. A snapshot of the value of
the counter can thus be collected as follows:
for_each_thread(t)
sum += READ_ONCE(per_thread(counter, t));
Again, it is possible to gain a similar effect using other mechanisms, but per-thread
variables combine convenience and high performance, as will be shown in more detail
in Section 5.2.
Quick Quiz 4.34: What do you do if you need a per-thread (not per-CPU!) variable in the
Linux kernel?
4.4 The Right Tool for the Job: How to Choose?

If you get stuck, change your tools; it may free your
thinking.
Paul Arden, abbreviated
As a rough rule of thumb, use the simplest tool that will get the job done. If you
can, simply program sequentially. If that is insufficient, try using a shell script to
mediate parallelism. If the resulting shell-script fork()/exec() overhead (about 480
microseconds for a minimal C program on an Intel Core Duo laptop) is too large,
try using the C-language fork() and wait() primitives. If the overhead of these
primitives (about 80 microseconds for a minimal child process) is still too large, then
you might need to use the POSIX threading primitives, choosing the appropriate locking
and/or atomic-operation primitives. If the overhead of the POSIX threading primitives
(typically sub-microsecond) is too great, then the primitives introduced in Chapter 9
may be required. Of course, the actual overheads will depend not only on your hardware,
but most critically on the manner in which you use the primitives. Furthermore,
always remember that inter-process communication and message-passing can be good
alternatives to shared-memory multithreaded execution, especially when your code
makes good use of the design principles called out in Chapter 6.
Quick Quiz 4.35: Wouldn’t the shell normally use vfork() rather than fork()?
Because concurrency was added to the C standard several decades after the C language
was first used to build concurrent systems, there are a number of ways of concurrently
v2023.06.11a
4.4. THE RIGHT TOOL FOR THE JOB: HOW TO CHOOSE? 75
accessing shared variables. All else being equal, the C11 standard operations described
in Section 4.2.6 should be your first stop. If you need to access a given shared variable
both with plain accesses and atomically, then the modern GCC atomics described in
Section 4.2.7 might work well for you. If you are working on an old codebase that
uses the classic GCC __sync API, then you should review Section 4.2.5 as well as
the relevant GCC documentation. If you are working on the Linux kernel or similar
codebase that combines use of the volatile keyword with inline assembly, or if you
need dependencies to provide ordering, look at the material presented in Section 4.3.4
as well as that in Chapter 15.
Whatever approach you take, please keep in mind that randomly hacking multi-
threaded code is a spectacularly bad idea, especially given that shared-memory parallel
systems use your own intelligence against you: The smarter you are, the deeper a hole
you will dig for yourself before you realize that you are in trouble [Pok16]. Therefore, it
is necessary to make the right design choices as well as the correct choice of individual
primitives, as will be discussed at length in subsequent chapters.
v2023.06.11a
v2023.06.11a
As easy as 1, 2, 3!
Unknown
Chapter 5
Counting
Counting is perhaps the simplest and most natural thing a computer can do. However,
counting efficiently and scalably on a large shared-memory multiprocessor can be quite
challenging. Furthermore, the simplicity of the underlying concept of counting allows
us to explore the fundamental issues of concurrency without the distractions of elaborate
data structures or complex synchronization primitives. Counting therefore provides an
excellent introduction to parallel programming.
This chapter covers a number of special cases for which there are simple, fast, and
scalable counting algorithms. But first, let us find out how much you already know
about concurrent counting.
Quick Quiz 5.1: Why should efficient and scalable counting be hard??? After all, computers
have special hardware for the sole purpose of doing counting!!!
Quick Quiz 5.2: Network-packet counting problem. Suppose that you need to collect
statistics on the number of networking packets transmitted and received. Packets might be
transmitted or received by any CPU on the system. Suppose further that your system is capable
of handling millions of packets per second per CPU, and that a systems-monitoring package
reads the count every five seconds. How would you implement this counter?
Quick Quiz 5.3: Approximate structure-allocation limit problem. Suppose that you need
to maintain a count of the number of structures allocated in order to fail any allocations once
the number of structures in use exceeds a limit (say, 10,000). Suppose further that the structures
are short-lived, the limit is rarely exceeded, and a “sloppy” approximate limit is acceptable.
Quick Quiz 5.4: Exact structure-allocation limit problem. Suppose that you need to
maintain a count of the number of structures allocated in order to fail any allocations once the
number of structures in use exceeds an exact limit (again, say 10,000). Suppose further that
these structures are short-lived, and that the limit is rarely exceeded, that there is almost always
at least one structure in use, and suppose further still that it is necessary to know exactly when
this counter reaches zero, for example, in order to free up some memory that is not required
unless there is at least one structure in use.
Quick Quiz 5.5: Removable I/O device access-count problem. Suppose that you need to
maintain a reference count on a heavily used removable mass-storage device, so that you can
tell the user when it is safe to remove the device. As usual, the user indicates a desire to remove
the device, and the system tells the user when it is safe to do so.
77
v2023.06.11a
78 CHAPTER 5. COUNTING
Listing 5.1: Just Count!

1 unsigned long counter = 0;
2
3 static __inline__ void inc_count(void)
4 {
5 WRITE_ONCE(counter, READ_ONCE(counter) + 1);
6 }
7
8 static __inline__ unsigned long read_count(void)
9 {
10 return READ_ONCE(counter);
11 }
Section 5.1 shows why counting is non-trivial. Sections 5.2 and 5.3 investigate
network-packet counting and approximate structure-allocation limits, respectively.
Section 5.4 takes on exact structure-allocation limits. Finally, Section 5.5 presents
performance measurements and discussion.
Sections 5.1 and 5.2 contain introductory material, while the remaining sections are
more advanced.
5.1 Why Isn’t Concurrent Counting Trivial?
Seek simplicity, and distrust it.

Alfred North Whitehead
Let’s start with something simple, for example, the straightforward use of arithmetic
shown in Listing 5.1 (count_nonatomic.c). Here, we have a counter on line 1, we
increment it on line 5, and we read out its value on line 10. What could be simpler?
Quick Quiz 5.6: One thing that could be simpler is ++ instead of that concatenation of
READ_ONCE() and WRITE_ONCE(). Why all that extra typing???
This approach has the additional advantage of being blazingly fast if you are doing
lots of reading and almost no incrementing, and on small systems, the performance is
excellent.
There is just one large fly in the ointment: This approach can lose counts. On my
six-core x86 laptop, a short run invoked inc_count() 285,824,000 times, but the final
value of the counter was only 35,385,525. Although approximation does have a large
place in computing, loss of 87 % of the counts is a bit excessive.
Quick Quiz 5.7: But can’t a smart compiler prove that line 5 of Listing 5.1 is equivalent to the
++ operator and produce an x86 add-to-memory instruction? And won’t the CPU cache cause
this to be atomic?
Quick Quiz 5.8: The 8-figure accuracy on the number of failures indicates that you really did
test this. Why would it be necessary to test such a trivial program, especially when the bug is
easily seen by inspection?
The straightforward way to count accurately is to use atomic operations, as shown in

Listing 5.2 (count_atomic.c). Line 1 defines an atomic variable, line 5 atomically
increments it, and line 10 reads it out. Because this is atomic, it keeps perfect count.
v2023.06.11a
5.1. WHY ISN’T CONCURRENT COUNTING TRIVIAL? 79
Listing 5.2: Just Count Atomically!

1 atomic_t counter = ATOMIC_INIT(0);
2
4 {
5 atomic_inc(&counter);
6 }
7
8 static __inline__ long read_count(void)
9 {
10 return atomic_read(&counter);
11 }
100000
Time Per Increment (ns)

10000
1000
100
10
1
1
10
100
Figure 5.1: Atomic Increment Scalability on x86
However, it is slower: On my six-core x86 laptop, it is more than twenty times slower
than non-atomic increment, even when only a single thread is incrementing.1
This poor performance should not be a surprise, given the discussion in Chapter 3,
nor should it be a surprise that the performance of atomic increment gets slower as
the number of CPUs and threads increase, as shown in Figure 5.1. In this figure, the
horizontal dashed line resting on the x axis is the ideal performance that would be
achieved by a perfectly scalable algorithm: With such an algorithm, a given increment
would incur the same overhead that it would in a single-threaded program. Atomic
increment of a single global variable is clearly decidedly non-ideal, and gets multiple
orders of magnitude worse with additional CPUs.
Quick Quiz 5.9: Why doesn’t the horizontal dashed line on the x axis meet the diagonal line
at 𝑥 = 1?
Quick Quiz 5.10: But atomic increment is still pretty fast. And incrementing a single variable
in a tight loop sounds pretty unrealistic to me, after all, most of the program’s execution should
be devoted to actually doing work, not accounting for the work it has done! Why should I care
about making this go faster?
1 Interestingly enough, non-atomically incrementing a counter will advance the counter
more quickly than atomically incrementing the counter. Of course, if your only goal is to
make the counter increase quickly, an easier approach is to simply assign a large value to the
counter. Nevertheless, there is likely to be a role for algorithms that use carefully relaxed
notions of correctness in order to gain greater performance and scalability [And91, ACMS03,
Rin13, Ung11].
v2023.06.11a

Figure 5.2: Data Flow For Global Atomic Increment
One one thousand.

Two one thousand.
Three one thousand...
Figure 5.3: Waiting to Count
For another perspective on global atomic increment, consider Figure 5.2. In order
for each CPU to get a chance to increment a given global variable, the cache line
containing that variable must circulate among all the CPUs, as shown by the red arrows.
Such circulation will take significant time, resulting in the poor performance seen in
Figure 5.1, which might be thought of as shown in Figure 5.3. The following sections
discuss high-performance counting, which avoids the delays inherent in such circulation.
Quick Quiz 5.11: But why can’t CPU designers simply ship the addition operation to the data,
avoiding the need to circulate the cache line containing the global variable being incremented?
v2023.06.11a
5.2. STATISTICAL COUNTERS 81
Listing 5.3: Array-Based Per-Thread Statistical Counters

1 DEFINE_PER_THREAD(unsigned long, counter);
2
4 {
5 unsigned long *p_counter = &__get_thread_var(counter);
6
7 WRITE_ONCE(*p_counter, *p_counter + 1);
8 }
9
11 {
12 int t;
13 unsigned long sum = 0;
14
15 for_each_thread(t)
16 sum += READ_ONCE(per_thread(counter, t));
17 return sum;
18 }
5.2 Statistical Counters

Facts are stubborn things, but statistics are pliable.
Mark Twain
This section covers the common special case of statistical counters, where the count
is updated extremely frequently and the value is read out rarely, if ever. These will be
used to solve the network-packet counting problem posed in Quick Quiz 5.2.
5.2.1 Design
Statistical counting is typically handled by providing a counter per thread (or CPU, when
running in the kernel), so that each thread updates its own counter, as was foreshadowed
in Section 4.3.6 on page 73. The aggregate value of the counters is read out by simply
summing up all of the threads’ counters, relying on the commutative and associative
properties of addition. This is an example of the Data Ownership pattern that will be
introduced in Section 6.3.4 on page 135.
Quick Quiz 5.12: But doesn’t the fact that C’s “integers” are limited in size complicate things?
5.2.2 Array-Based Implementation

One way to provide per-thread variables is to allocate an array with one element per
thread (presumably cache aligned and padded to avoid false sharing).
Quick Quiz 5.13: An array??? But doesn’t that limit the number of threads?
Such an array can be wrapped into per-thread primitives, as shown in Listing 5.3
(count_stat.c). Line 1 defines an array containing a set of per-thread counters of
type unsigned long named, creatively enough, counter.
Lines 3–8 show a function that increments the counters, using the __get_thread_
var() primitive to locate the currently running thread’s element of the counter
array. Because this element is modified only by the corresponding thread, non-atomic
increment suffices. However, this code uses WRITE_ONCE() to prevent destructive
v2023.06.11a

Figure 5.4: Data Flow For Per-Thread Increment
compiler optimizations. For but one example, the compiler is within its rights to use a
to-be-stored-to location as temporary storage, thus writing what would be for all intents
and purposes garbage to that location just before doing the desired store. This could
of course be rather confusing to anything attempting to read out the count. The use of
WRITE_ONCE() prevents this optimization and others besides.
Quick Quiz 5.14: What other nasty optimizations could GCC apply?
Lines 10–18 show a function that reads out the aggregate value of the counter, using
the for_each_thread() primitive to iterate over the list of currently running threads,
and using the per_thread() primitive to fetch the specified thread’s counter. This
code also uses READ_ONCE() to ensure that the compiler doesn’t optimize these loads
into oblivion. For but one example, a pair of consecutive calls to read_count()
might be inlined, and an intrepid optimizer might notice that the same locations were
being summed and thus incorrectly conclude that it would be simply wonderful to sum
them once and use the resulting value twice. This sort of optimization might be rather
frustrating to people expecting later read_count() calls to account for the activities of
other threads. The use of READ_ONCE() prevents this optimization and others besides.
Quick Quiz 5.15: How does the per-thread counter variable in Listing 5.3 get initialized?
Quick Quiz 5.16: How is the code in Listing 5.3 supposed to permit more than one counter?
This approach scales linearly with increasing number of updater threads invoking
inc_count(). As is shown by the green arrows on each CPU in Figure 5.4, the reason
for this is that each CPU can make rapid progress incrementing its thread’s variable,
without any expensive cross-system communication. As such, this section solves the
network-packet counting problem presented at the beginning of this chapter.
Quick Quiz 5.17: The read operation takes time to sum up the per-thread values, and
during that time, the counter could well be changing. This means that the value returned by
read_count() in Listing 5.3 will not necessarily be exact. Assume that the counter is being
incremented at rate 𝑟 counts per unit time, and that read_count()’s execution consumes 𝛥
units of time. What is the expected error in the return value?
However, many implementations provide cheaper mechanisms for per-thread data

that are free from arbitrary array-size limits. This is the topic of the next section.
v2023.06.11a
Listing 5.4: Per-Thread Statistical Counters

1 unsigned long _Thread_local counter = 0;
2 unsigned long *counterp[NR_THREADS] = { NULL };
3 unsigned long finalcount = 0;
4 DEFINE_SPINLOCK(final_mutex);
5
6 static inline void inc_count(void)
7 {
8 WRITE_ONCE(counter, counter + 1);
9 }
10
11 static inline unsigned long read_count(void)
12 {
13 int t;
14 unsigned long sum;
15
16 spin_lock(&final_mutex);
17 sum = finalcount;
19 if (counterp[t] != NULL)
20 sum += READ_ONCE(*counterp[t]);
21 spin_unlock(&final_mutex);
22 return sum;
23 }
24
25 void count_register_thread(unsigned long *p)
26 {
27 int idx = smp_thread_id();
28
30 counterp[idx] = &counter;
32 }
33
34 void count_unregister_thread(int nthreadsexpected)
35 {
37
39 finalcount += counter;
40 counterp[idx] = NULL;
42 }
5.2.3 Per-Thread-Variable-Based Implementation

The C language, since C11, features a _Thread_local storage class that provides per-
thread storage.2 This can be used as shown in Listing 5.4 (count_end.c) to implement
a statistical counter that not only scales well and avoids arbitrary thread-number limits,
but that also incurs little or no performance penalty to incrementers compared to simple
non-atomic increment.
Lines 1–4 define needed variables: counter is the per-thread counter variable,
the counterp[] array allows threads to access each others’ counters, finalcount
accumulates the total as individual threads exit, and final_mutex coordinates between
threads accumulating the total value of the counter and exiting threads.
Quick Quiz 5.18: Doesn’t that explicit counterp array in Listing 5.4 reimpose an arbitrary
limit on the number of threads? Why doesn’t the C language provide a per_thread() interface,
similar to the Linux kernel’s per_cpu() primitive, to allow threads to more easily access each
others’ per-thread variables?
2 GCC provides its own __thread storage class, which was used in previous versions of
this book. The two methods for specifying a thread-local variable are interchangeable when
using GCC.
v2023.06.11a
The inc_count() function used by updaters is quite simple, as can be seen on

lines 6–9.
The read_count() function used by readers is a bit more complex. Line 16 acquires
a lock to exclude exiting threads, and line 21 releases it. Line 17 initializes the sum to
the count accumulated by those threads that have already exited, and lines 18–20 sum
the counts being accumulated by threads currently running. Finally, line 22 returns the
sum.
Quick Quiz 5.19: Doesn’t the check for NULL on line 19 of Listing 5.4 add extra branch
mispredictions? Why not have a variable set permanently to zero, and point unused counter-
pointers to that variable rather than setting them to NULL?
Quick Quiz 5.20: Why on earth do we need something as heavyweight as a lock guarding the
summation in the function read_count() in Listing 5.4?
Lines 25–32 show the count_register_thread() function, which must be called

by each thread before its first use of this counter. This function simply sets up this
thread’s element of the counterp[] array to point to its per-thread counter variable.
Quick Quiz 5.21: Why on earth do we need to acquire the lock in count_register_
thread() in Listing 5.4? It is a single properly aligned machine-word store to a location that
no other thread is modifying, so it should be atomic anyway, right?
Lines 34–42 show the count_unregister_thread() function, which must be

called prior to exit by each thread that previously called count_register_thread().
Line 38 acquires the lock, and line 41 releases it, thus excluding any calls to read_
count() as well as other calls to count_unregister_thread(). Line 39 adds this
thread’s counter to the global finalcount, and then line 40 NULLs out its counterp[]
array entry. A subsequent call to read_count() will see the exiting thread’s count in
the global finalcount, and will skip the exiting thread when sequencing through the
counterp[] array, thus obtaining the correct total.
This approach gives updaters almost exactly the same performance as a non-atomic
add, and also scales linearly. On the other hand, concurrent reads contend for a single
global lock, and therefore perform poorly and scale abysmally. However, this is not
a problem for statistical counters, where incrementing happens often and readout
happens almost never. Of course, this approach is considerably more complex than the
array-based scheme, due to the fact that a given thread’s per-thread variables vanish
when that thread exits.
Quick Quiz 5.22: Fine, but the Linux kernel doesn’t have to acquire a lock when reading out
the aggregate value of per-CPU counters. So why should user-space code need to do this???
Both the array-based and _Thread_local-based approaches offer excellent update-

side performance and scalability. However, these benefits result in large read-side
expense for large numbers of threads. The next section shows one way to reduce
read-side expense while still retaining the update-side scalability.
5.2.4 Eventually Consistent Implementation

One way to retain update-side scalability while greatly improving read-side performance
is to weaken consistency requirements. The counting algorithm in the previous section
is guaranteed to return a value between the value that an ideal counter would have
taken on near the beginning of read_count()’s execution and that near the end of
v2023.06.11a
Listing 5.5: Array-Based Per-Thread Eventually Consistent Counters

1 DEFINE_PER_THREAD(unsigned long, counter);
2 unsigned long global_count;
3 int stopflag;
4
6 {
7 unsigned long *p_counter = &__get_thread_var(counter);
8
9 WRITE_ONCE(*p_counter, *p_counter + 1);
10 }
11
13 {
14 return READ_ONCE(global_count);
15 }
16
17 void *eventual(void *arg)
18 {
19 int t;
21
22 while (READ_ONCE(stopflag) < 3) {
23 sum = 0;
25 sum += READ_ONCE(per_thread(counter, t));
26 WRITE_ONCE(global_count, sum);
28 if (READ_ONCE(stopflag))
29 smp_store_release(&stopflag, stopflag + 1);
30 }
31 return NULL;
32 }
33
34 void count_init(void)
35 {
36 int en;
37 pthread_t tid;
38
39 en = pthread_create(&tid, NULL, eventual, NULL);
40 if (en != 0) {
43 }
44 }
45
46 void count_cleanup(void)
47 {
48 WRITE_ONCE(stopflag, 1);
49 while (smp_load_acquire(&stopflag) < 3)
51 }
read_count()’s execution. Eventual consistency [Vog09] provides a weaker guarantee:

In absence of calls to inc_count(), calls to read_count() will eventually return an
accurate count.
We exploit eventual consistency by maintaining a global counter. However, updaters
only manipulate their per-thread counters. A separate thread is provided to transfer
counts from the per-thread counters to the global counter. Readers simply access the
value of the global counter. If updaters are active, the value used by the readers will be
out of date, however, once updates cease, the global counter will eventually converge on
the true value—hence this approach qualifies as eventually consistent.
The implementation is shown in Listing 5.5 (count_stat_eventual.c). Lines 1–2
show the per-thread variable and the global variable that track the counter’s value, and
line 3 shows stopflag which is used to coordinate termination (for the case where
v2023.06.11a
we want to terminate the program with an accurate counter value). The inc_count()
function shown on lines 5–10 is similar to its counterpart in Listing 5.3. The read_
count() function shown on lines 12–15 simply returns the value of the global_count
variable.
However, the count_init() function on lines 34–44 creates the eventual() thread
shown on lines 17–32, which cycles through all the threads, summing the per-thread
local counter and storing the sum to the global_count variable. The eventual()
thread waits an arbitrarily chosen one millisecond between passes.
The count_cleanup() function on lines 46–51 coordinates termination. The call to
smp_load_acquire() here and the call to smp_store_release() in eventual()
ensure that all updates to global_count are visible to code following the call to
count_cleanup().
This approach gives extremely fast counter read-out while still supporting linear
counter-update scalability. However, this excellent read-side performance and update-
side scalability comes at the cost of the additional thread running eventual().
Quick Quiz 5.23: Why doesn’t inc_count() in Listing 5.5 need to use atomic instructions?
After all, we now have multiple threads accessing the per-thread counters!
Quick Quiz 5.24: Won’t the single global thread in the function eventual() of Listing 5.5
be just as severe a bottleneck as a global lock would be?
Quick Quiz 5.25: Won’t the estimate returned by read_count() in Listing 5.5 become
increasingly inaccurate as the number of threads rises?
Quick Quiz 5.26: Given that in the eventually-consistent algorithm shown in Listing 5.5 both
reads and updates have extremely low overhead and are extremely scalable, why would anyone
bother with the implementation described in Section 5.2.2, given its costly read-side code?
Quick Quiz 5.27: What is the accuracy of the estimate returned by read_count() in
Listing 5.5?
5.2.5 Discussion
These three implementations show that it is possible to obtain near-uniprocessor
performance for statistical counters, despite running on a parallel machine.
Quick Quiz 5.28: What fundamental difference is there between counting packets and counting
the total number of bytes in the packets, given that the packets vary in size?
Quick Quiz 5.29: Given that the reader must sum all the threads’ counters, this counter-read
operation could take a long time given large numbers of threads. Is there any way that the
increment operation can remain fast and scalable while allowing readers to also enjoy not only
reasonable performance and scalability, but also good accuracy?
Given what has been presented in this section, you should now be able to answer the
Quick Quiz about statistical counters for networking near the beginning of this chapter.
v2023.06.11a
5.3. APPROXIMATE LIMIT COUNTERS 87
5.3 Approximate Limit Counters

An approximate answer to the right problem is worth
a good deal more than an exact answer to an
approximate problem.
John Tukey
Another special case of counting involves limit-checking. For example, as noted in

the approximate structure-allocation limit problem in Quick Quiz 5.3, suppose that
you need to maintain a count of the number of structures allocated in order to fail any
allocations once the number of structures in use exceeds a limit, in this case, 10,000.
Suppose further that these structures are short-lived, that this limit is rarely exceeded,
and that this limit is approximate in that it is OK either to exceed it sometimes by some
bounded amount or to fail to reach it sometimes, again by some bounded amount. See
Section 5.4 if you instead need the limit to be exact.
5.3.1 Design
One possible design for limit counters is to divide the limit of 10,000 by the number
of threads, and give each thread a fixed pool of structures. For example, given 100
threads, each thread would manage its own pool of 100 structures. This approach is
simple, and in some cases works well, but it does not handle the common case where
a given structure is allocated by one thread and freed by another [MS93]. On the one
hand, if a given thread takes credit for any structures it frees, then the thread doing most
of the allocating runs out of structures, while the threads doing most of the freeing
have lots of credits that they cannot use. On the other hand, if freed structures are
credited to the CPU that allocated them, it will be necessary for CPUs to manipulate
each others’ counters, which will require expensive atomic instructions or other means
of communicating between threads.3
In short, for many important workloads, we cannot fully partition the counter. Given
that partitioning the counters was what brought the excellent update-side performance
for the three schemes discussed in Section 5.2, this might be grounds for some pessimism.
However, the eventually consistent algorithm presented in Section 5.2.4 provides an
interesting hint. Recall that this algorithm kept two sets of books, a per-thread counter
variable for updaters and a global_count variable for readers, with an eventual()
thread that periodically updated global_count to be eventually consistent with the
values of the per-thread counter. The per-thread counter perfectly partitioned the
counter value, while global_count kept the full value.
For limit counters, we can use a variation on this theme where we partially partition
the counter. For example, consider four threads with each having not only a per-thread
counter, but also a per-thread maximum value (call it countermax).
But then what happens if a given thread needs to increment its counter, but counter
is equal to its countermax? The trick here is to move half of that thread’s counter
value to a globalcount, then increment counter. For example, if a given thread’s
counter and countermax variables were both equal to 10, we do the following:
1. Acquire a global lock.

3 That said, if each structure will always be freed by the same CPU (or thread) that
allocated it, then this simple partitioning approach works extremely well.
v2023.06.11a
2. Add five to globalcount.
3. To balance out the addition, subtract five from this thread’s counter.
4. Release the global lock.
5. Increment this thread’s counter, resulting in a value of six.
Although this procedure still requires a global lock, that lock need only be acquired
once for every five increment operations, greatly reducing that lock’s level of contention.
We can reduce this contention as low as we wish by increasing the value of countermax.
However, the corresponding penalty for increasing the value of countermax is reduced
accuracy of globalcount. To see this, note that on a four-CPU system, if countermax
is equal to ten, globalcount will be in error by at most 40 counts. In contrast, if
countermax is increased to 100, globalcount might be in error by as much as 400
counts.
This raises the question of just how much we care about globalcount’s deviation
from the aggregate value of the counter, where this aggregate value is the sum of
globalcount and each thread’s counter variable. The answer to this question depends
on how far the aggregate value is from the counter’s limit (call it globalcountmax).
The larger the difference between these two values, the larger countermax can be
without risk of exceeding the globalcountmax limit. This means that the value of a
given thread’s countermax variable can be set based on this difference. When far from
the limit, the countermax per-thread variables are set to large values to optimize for
performance and scalability, while when close to the limit, these same variables are set
to small values to minimize the error in the checks against the globalcountmax limit.
This design is an example of parallel fastpath, which is an important design pattern
in which the common case executes with no expensive instructions and no interactions
between threads, but where occasional use is also made of a more conservatively
designed (and higher overhead) global algorithm. This design pattern is covered in
more detail in Section 6.4.
5.3.2 Simple Limit Counter Implementation

Listing 5.6 shows both the per-thread and global variables used by this implementation.
The per-thread counter and countermax variables are the corresponding thread’s local
counter and the upper bound on that counter, respectively. The globalcountmax vari-
able on line 3 contains the upper bound for the aggregate counter, and the globalcount
variable on line 4 is the global counter. The sum of globalcount and each thread’s
counter gives the aggregate value of the overall counter. The globalreserve vari-
able on line 5 is at least the sum of all of the per-thread countermax variables. The
relationship among these variables is shown by Figure 5.5:
1. The sum of globalcount and globalreserve must be less than or equal to

globalcountmax.
2. The sum of all threads’ countermax values must be less than or equal to
globalreserve.
3. Each thread’s counter must be less than or equal to that thread’s countermax.
v2023.06.11a
Listing 5.6: Simple Limit Counter Variables

1 unsigned long __thread counter = 0;
2 unsigned long __thread countermax = 0;
3 unsigned long globalcountmax = 10000;
4 unsigned long globalcount = 0;
5 unsigned long globalreserve = 0;
7 DEFINE_SPINLOCK(gblcnt_mutex);
countermax 3
globalreserve
counter 3
countermax 2 counter 2
globalcountmax
countermax 1 counter 1
countermax 0
counter 0
globalcount
Figure 5.5: Simple Limit Counter Variable Relationships
Each element of the counterp[] array references the corresponding thread’s

counter variable, and, finally, the gblcnt_mutex spinlock guards all of the global
variables, in other words, no thread is permitted to access or modify any of the global
variables unless it has acquired gblcnt_mutex.
Listing 5.7 shows the add_count(), sub_count(), and read_count() functions
(count_lim.c).
Quick Quiz 5.30: Why does Listing 5.7 provide add_count() and sub_count() instead of
the inc_count() and dec_count() interfaces show in Section 5.2?
Lines 1–18 show add_count(), which adds the specified value delta to the counter.
Line 3 checks to see if there is room for delta on this thread’s counter, and, if so,
line 4 adds it and line 5 returns success. This is the add_counter() fastpath, and it
does no atomic operations, references only per-thread variables, and should not incur
any cache misses.
Quick Quiz 5.31: What is with the strange form of the condition on line 3 of Listing 5.7?
Why not the more intuitive form of the fastpath shown in Listing 5.8?
If the test on line 3 fails, we must access global variables, and thus must acquire
gblcnt_mutex on line 7, which we release on line 11 in the failure case or on line 16
in the success case. Line 8 invokes globalize_count(), shown in Listing 5.9,
which clears the thread-local variables, adjusting the global variables as needed, thus
simplifying global processing. (But don’t take my word for it, try coding it yourself!)
v2023.06.11a
Listing 5.7: Simple Limit Counter Add, Subtract, and Read

1 static __inline__ int add_count(unsigned long delta)
2 {
3 if (countermax - counter >= delta) {
4 WRITE_ONCE(counter, counter + delta);
5 return 1;
6 }
7 spin_lock(&gblcnt_mutex);
8 globalize_count();
9 if (globalcountmax -
10 globalcount - globalreserve < delta) {
11 spin_unlock(&gblcnt_mutex);
12 return 0;
13 }
14 globalcount += delta;
15 balance_count();
17 return 1;
18 }
19
20 static __inline__ int sub_count(unsigned long delta)
21 {
22 if (counter >= delta) {
23 WRITE_ONCE(counter, counter - delta);
24 return 1;
25 }
28 if (globalcount < delta) {
30 return 0;
31 }
32 globalcount -= delta;
33 balance_count();
35 return 1;
36 }
37
39 {
40 int t;
42
44 sum = globalcount;
45 for_each_thread(t) {
48 }
50 return sum;
51 }
Listing 5.8: Intuitive Fastpath

3 if (counter + delta <= countermax) {
5 return 1;
6 }
v2023.06.11a
Lines 9 and 10 check to see if addition of delta can be accommodated, with the
meaning of the expression preceding the less-than sign shown in Figure 5.5 as the
difference in height of the two red (leftmost) bars. If the addition of delta cannot
be accommodated, then line 11 (as noted earlier) releases gblcnt_mutex and line 12
returns indicating failure.
Otherwise, we take the slowpath. Line 14 adds delta to globalcount, and then
line 15 invokes balance_count() (shown in Listing 5.9) in order to update both the
global and the per-thread variables. This call to balance_count() will usually set this
thread’s countermax to re-enable the fastpath. Line 16 then releases gblcnt_mutex
(again, as noted earlier), and, finally, line 17 returns indicating success.
Quick Quiz 5.32: Why does globalize_count() zero the per-thread variables, only to
later call balance_count() to refill them in Listing 5.7? Why not just leave the per-thread
variables non-zero?
Lines 20–36 show sub_count(), which subtracts the specified delta from the
counter. Line 22 checks to see if the per-thread counter can accommodate this
subtraction, and, if so, line 23 does the subtraction and line 24 returns success. These
lines form sub_count()’s fastpath, and, as with add_count(), this fastpath executes
no costly operations.
If the fastpath cannot accommodate subtraction of delta, execution proceeds to the
slowpath on lines 26–35. Because the slowpath must access global state, line 26 acquires
gblcnt_mutex, which is released either by line 29 (in case of failure) or by line 34 (in
case of success). Line 27 invokes globalize_count(), shown in Listing 5.9, which
again clears the thread-local variables, adjusting the global variables as needed. Line 28
checks to see if the counter can accommodate subtracting delta, and, if not, line 29
releases gblcnt_mutex (as noted earlier) and line 30 returns failure.
Quick Quiz 5.33: Given that globalreserve counted against us in add_count(), why
doesn’t it count for us in sub_count() in Listing 5.7?
Quick Quiz 5.34: Suppose that one thread invokes add_count() shown in Listing 5.7, and
then another thread invokes sub_count(). Won’t sub_count() return failure even though
the value of the counter is non-zero?
If, on the other hand, line 28 finds that the counter can accommodate subtracting
delta, we complete the slowpath. Line 32 does the subtraction and then line 33 invokes
balance_count() (shown in Listing 5.9) in order to update both global and per-thread
variables (hopefully re-enabling the fastpath). Then line 34 releases gblcnt_mutex,
and line 35 returns success.
Quick Quiz 5.35: Why have both add_count() and sub_count() in Listing 5.7? Why not
simply pass a negative number to add_count()?
Lines 38–51 show read_count(), which returns the aggregate value of the counter. It
acquires gblcnt_mutex on line 43 and releases it on line 49, excluding global operations
from add_count() and sub_count(), and, as we will see, also excluding thread
creation and exit. Line 44 initializes local variable sum to the value of globalcount,
and then the loop spanning lines 45–48 sums the per-thread counter variables. Line 50
then returns the sum.
Listing 5.9 shows a number of utility functions used by the add_count(), sub_
count(), and read_count() primitives shown in Listing 5.7.
v2023.06.11a
Listing 5.9: Simple Limit Counter Utility Functions

1 static __inline__ void globalize_count(void)
2 {
3 globalcount += counter;
4 counter = 0;
5 globalreserve -= countermax;
6 countermax = 0;
7 }
8
9 static __inline__ void balance_count(void)
10 {
11 countermax = globalcountmax -
12 globalcount - globalreserve;
13 countermax /= num_online_threads();
14 globalreserve += countermax;
15 counter = countermax / 2;
16 if (counter > globalcount)
17 counter = globalcount;
18 globalcount -= counter;
19 }
20
21 void count_register_thread(void)
22 {
24
28 }
29
31 {
33
38 }
Lines 1–7 show globalize_count(), which zeros the current thread’s per-thread
counters, adjusting the global variables appropriately. It is important to note that this
function does not change the aggregate value of the counter, but instead changes how
the counter’s current value is represented. Line 3 adds the thread’s counter variable to
globalcount, and line 4 zeroes counter. Similarly, line 5 subtracts the per-thread
countermax from globalreserve, and line 6 zeroes countermax. It is helpful to
refer to Figure 5.5 when reading both this function and balance_count(), which is
next.
Lines 9–19 show balance_count(), which is roughly speaking the inverse of
globalize_count(). This function’s job is to set the current thread’s countermax
variable to the largest value that avoids the risk of the counter exceeding the
globalcountmax limit. Changing the current thread’s countermax variable of course
requires corresponding adjustments to counter, globalcount and globalreserve,
as can be seen by referring back to Figure 5.5. By doing this, balance_count()
maximizes use of add_count()’s and sub_count()’s low-overhead fastpaths. As with
globalize_count(), balance_count() is not permitted to change the aggregate
value of the counter.
Lines 11–13 compute this thread’s share of that portion of globalcountmax that
is not already covered by either globalcount or globalreserve, and assign the
computed quantity to this thread’s countermax. Line 14 makes the corresponding
adjustment to globalreserve. Line 15 sets this thread’s counter to the middle of
v2023.06.11a
globalize_count() balance_count()
cm 3
globalreserve
c 3
globalreserve
cm 3 cm 3
globalreserve
c 3 c 3
cm 2
c 2
cm 2 cm 2
c 2 c 2
cm 1 c 1
cm 1 c 1 cm 1 c 1
cm 0
c 0
cm 0 c 0
globalcount
globalcount
globalcount
Figure 5.6: Schematic of Globalization and Balancing
the range from zero to countermax. Line 16 checks to see whether globalcount can
in fact accommodate this value of counter, and, if not, line 17 decreases counter
accordingly. Finally, in either case, line 18 makes the corresponding adjustment to
globalcount.
Quick Quiz 5.36: Why set counter to countermax / 2 in line 15 of Listing 5.9? Wouldn’t
it be simpler to just take countermax counts?
It is helpful to look at a schematic depicting how the relationship of the counters

changes with the execution of first globalize_count() and then balance_count(),
as shown in Figure 5.6. Time advances from left to right, with the leftmost configuration
roughly that of Figure 5.5. The center configuration shows the relationship of these
same counters after globalize_count() is executed by thread 0. As can be seen from
the figure, thread 0’s counter (“c 0” in the figure) is added to globalcount, while the
value of globalreserve is reduced by this same amount. Both thread 0’s counter
and its countermax (“cm 0” in the figure) are reduced to zero. The other three threads’
counters are unchanged. Note that this change did not affect the overall value of the
counter, as indicated by the bottommost dotted line connecting the leftmost and center
configurations. In other words, the sum of globalcount and the four threads’ counter
variables is the same in both configurations. Similarly, this change did not affect the
sum of globalcount and globalreserve, as indicated by the upper dotted line.
The rightmost configuration shows the relationship of these counters after balance_
count() is executed, again by thread 0. One-quarter of the remaining count, denoted
by the vertical line extending up from all three configurations, is added to thread 0’s
countermax and half of that to thread 0’s counter. The amount added to thread 0’s
counter is also subtracted from globalcount in order to avoid changing the overall
value of the counter (which is again the sum of globalcount and the three threads’
v2023.06.11a
Listing 5.10: Approximate Limit Counter Variables

8 #define MAX_COUNTERMAX 100
counter variables), again as indicated by the lowermost of the two dotted lines
connecting the center and rightmost configurations. The globalreserve variable
is also adjusted so that this variable remains equal to the sum of the four threads’
countermax variables. Because thread 0’s counter is less than its countermax,
thread 0 can once again increment the counter locally.
Quick Quiz 5.37: In Figure 5.6, even though a quarter of the remaining count up to the limit
is assigned to thread 0, only an eighth of the remaining count is consumed, as indicated by the
uppermost dotted line connecting the center and the rightmost configurations. Why is that?
Lines 21–28 show count_register_thread(), which sets up state for newly

created threads. This function simply installs a pointer to the newly created thread’s
counter variable into the corresponding entry of the counterp[] array under the
protection of gblcnt_mutex.
Finally, lines 30–38 show count_unregister_thread(), which tears down state
for a soon-to-be-exiting thread. Line 34 acquires gblcnt_mutex and line 37 releases
it. Line 35 invokes globalize_count() to clear out this thread’s counter state, and
line 36 clears this thread’s entry in the counterp[] array.
5.3.3 Simple Limit Counter Discussion

This type of counter is quite fast when aggregate values are near zero, with some overhead
due to the comparison and branch in both add_count()’s and sub_count()’s fastpaths.
However, the use of a per-thread countermax reserve means that add_count() can
fail even when the aggregate value of the counter is nowhere near globalcountmax.
Similarly, sub_count() can fail even when the aggregate value of the counter is
nowhere near zero.
In many cases, this is unacceptable. Even if the globalcountmax is intended to be
an approximate limit, there is usually a limit to exactly how much approximation can be
tolerated. One way to limit the degree of approximation is to impose an upper limit on
the value of the per-thread countermax instances. This task is undertaken in the next
section.
5.3.4 Approximate Limit Counter Implementation

Because this implementation (count_lim_app.c) is quite similar to that in the previous
section (Listings 5.6, 5.7, and 5.9), only the changes are shown here. Listing 5.10 is
identical to Listing 5.6, with the addition of MAX_COUNTERMAX, which sets the maximum
permissible value of the per-thread countermax variable.
Similarly, Listing 5.11 is identical to the balance_count() function in Listing 5.9,
with the addition of lines 6 and 7, which enforce the MAX_COUNTERMAX limit on the
per-thread countermax variable.
v2023.06.11a
5.4. EXACT LIMIT COUNTERS 95
Listing 5.11: Approximate Limit Counter Balancing

1 static void balance_count(void)
2 {
3 countermax = globalcountmax -
4 globalcount - globalreserve;
6 if (countermax > MAX_COUNTERMAX)
7 countermax = MAX_COUNTERMAX;
13 }
5.3.5 Approximate Limit Counter Discussion

These changes greatly reduce the limit inaccuracy seen in the previous version, but
present another problem: Any given value of MAX_COUNTERMAX will cause a workload-
dependent fraction of accesses to fall off the fastpath. As the number of threads increase,
non-fastpath execution will become both a performance and a scalability problem.
However, we will defer this problem and turn instead to counters with exact limits.
5.4 Exact Limit Counters

Exactitude can be expensive. Spend wisely.
Unknown
To solve the exact structure-allocation limit problem noted in Quick Quiz 5.4, we need a
limit counter that can tell exactly when its limits are exceeded. One way of implementing
such a limit counter is to cause threads that have reserved counts to give them up. One
way to do this is to use atomic instructions. Of course, atomic instructions will slow
down the fastpath, but on the other hand, it would be silly not to at least give them a try.
5.4.1 Atomic Limit Counter Implementation

Unfortunately, if one thread is to safely remove counts from another thread, both threads
will need to atomically manipulate that thread’s counter and countermax variables.
The usual way to do this is to combine these two variables into a single variable, for
example, given a 32-bit variable, using the high-order 16 bits to represent counter and
the low-order 16 bits to represent countermax.
Quick Quiz 5.38: Why is it necessary to atomically manipulate the thread’s counter and
countermax variables as a unit? Wouldn’t it be good enough to atomically manipulate them
individually?
The variables and access functions for a simple atomic limit counter are shown in
Listing 5.12 (count_lim_atomic.c). The counter and countermax variables in
earlier algorithms are combined into the single variable counterandmax shown on
line 1, with counter in the upper half and countermax in the lower half. This variable
is of type atomic_t, which has an underlying representation of int.
Lines 2–6 show the definitions for globalcountmax, globalcount,
globalreserve, counterp, and gblcnt_mutex, all of which take on roles simi-
v2023.06.11a
Listing 5.12: Atomic Limit Counter Variables and Access Functions

1 atomic_t __thread counterandmax = ATOMIC_INIT(0);
2 unsigned long globalcountmax = 1 << 25;
5 atomic_t *counterp[NR_THREADS] = { NULL };
7 #define CM_BITS (sizeof(atomic_t) * 4)
8 #define MAX_COUNTERMAX ((1 << CM_BITS) - 1)
9
10 static __inline__ void
11 split_counterandmax_int(int cami, int *c, int *cm)
12 {
13 *c = (cami >> CM_BITS) & MAX_COUNTERMAX;
14 *cm = cami & MAX_COUNTERMAX;
15 }
16
17 static __inline__ void
18 split_counterandmax(atomic_t *cam, int *old, int *c, int *cm)
19 {
20 unsigned int cami = atomic_read(cam);
21
22 *old = cami;
23 split_counterandmax_int(cami, c, cm);
24 }
25
26 static __inline__ int merge_counterandmax(int c, int cm)
27 {
28 unsigned int cami;
29
30 cami = (c << CM_BITS) | cm;
31 return ((int)cami);
32 }
lar to their counterparts in Listing 5.10. Line 7 defines CM_BITS, which gives the
number of bits in each half of counterandmax, and line 8 defines MAX_COUNTERMAX,
which gives the maximum value that may be held in either half of counterandmax.
Quick Quiz 5.39: In what way does line 7 of Listing 5.12 violate the C standard?
Lines 10–15 show the split_counterandmax_int() function, which, when given

the underlying int from the atomic_t counterandmax variable, splits it into its
counter (c) and countermax (cm) components. Line 13 isolates the most-significant
half of this int, placing the result as specified by argument c, and line 14 isolates the
least-significant half of this int, placing the result as specified by argument cm.
Lines 17–24 show the split_counterandmax() function, which picks up the
underlying int from the specified variable on line 20, stores it as specified by the old
argument on line 22, and then invokes split_counterandmax_int() to split it on
line 23.
Quick Quiz 5.40: Given that there is only one counterandmax variable, why bother passing
in a pointer to it on line 18 of Listing 5.12?
Lines 26–32 show the merge_counterandmax() function, which can be thought

of as the inverse of split_counterandmax(). Line 30 merges the counter and
countermax values passed in c and cm, respectively, and returns the result.
Quick Quiz 5.41: Why does merge_counterandmax() in Listing 5.12 return an int rather
than storing directly into an atomic_t?
Listing 5.13 shows the add_count() and sub_count() functions.
v2023.06.11a
Listing 5.13: Atomic Limit Counter Add and Subtract

1 int add_count(unsigned long delta)
2 {
3 int c;
4 int cm;
5 int old;
6 int new;
7
8 do {
9 split_counterandmax(&counterandmax, &old, &c, &cm);
10 if (delta > MAX_COUNTERMAX || c + delta > cm)
11 goto slowpath;
12 new = merge_counterandmax(c + delta, cm);
13 } while (atomic_cmpxchg(&counterandmax,
14 old, new) != old);
15 return 1;
16 slowpath:
19 if (globalcountmax - globalcount -
20 globalreserve < delta) {
21 flush_local_count();
25 return 0;
26 }
27 }
29 balance_count();
31 return 1;
32 }
33
34 int sub_count(unsigned long delta)
35 {
36 int c;
37 int cm;
38 int old;
39 int new;
40
41 do {
43 if (delta > c)
44 goto slowpath;
45 new = merge_counterandmax(c - delta, cm);
46 } while (atomic_cmpxchg(&counterandmax,
48 return 1;
49 slowpath:
56 return 0;
57 }
58 }
60 balance_count();
62 return 1;
63 }
v2023.06.11a
Listing 5.14: Atomic Limit Counter Read

1 unsigned long read_count(void)
2 {
3 int c;
4 int cm;
5 int old;
6 int t;
8
12 if (counterp[t] != NULL) {
13 split_counterandmax(counterp[t], &old, &c, &cm);
14 sum += c;
15 }
16 }
18 return sum;
19 }
Lines 1–32 show add_count(), whose fastpath spans lines 8–15, with the remainder
of the function being the slowpath. Lines 8–14 of the fastpath form a compare-and-swap
(CAS) loop, with the atomic_cmpxchg() primitive on lines 13–14 performing the
actual CAS. Line 9 splits the current thread’s counterandmax variable into its counter
(in c) and countermax (in cm) components, while placing the underlying int into old.
Line 10 checks whether the amount delta can be accommodated locally (taking care
to avoid integer overflow), and if not, line 11 transfers to the slowpath. Otherwise,
line 12 combines an updated counter value with the original countermax value into
new. The atomic_cmpxchg() primitive on lines 13–14 then atomically compares this
thread’s counterandmax variable to old, updating its value to new if the comparison
succeeds. If the comparison succeeds, line 15 returns success, otherwise, execution
continues in the loop at line 8.
Quick Quiz 5.42: Yecch! Why the ugly goto on line 11 of Listing 5.13? Haven’t you heard
of the break statement???
Quick Quiz 5.43: Why would the atomic_cmpxchg() primitive at lines 13–14 of Listing 5.13
ever fail? After all, we picked up its old value on line 9 and have not changed it!
Lines 16–31 of Listing 5.13 show add_count()’s slowpath, which is protected by

gblcnt_mutex, which is acquired on line 17 and released on lines 24 and 30. Line 18
invokes globalize_count(), which moves this thread’s state to the global counters.
Lines 19–20 check whether the delta value can be accommodated by the current
global state, and, if not, line 21 invokes flush_local_count() to flush all threads’
local state to the global counters, and then lines 22–23 recheck whether delta can be
accommodated. If, after all that, the addition of delta still cannot be accommodated,
then line 24 releases gblcnt_mutex (as noted earlier), and then line 25 returns failure.
Otherwise, line 28 adds delta to the global counter, line 29 spreads counts to the
local state if appropriate, line 30 releases gblcnt_mutex (again, as noted earlier), and
finally, line 31 returns success.
Lines 34–63 of Listing 5.13 show sub_count(), which is structured similarly to
add_count(), having a fastpath on lines 41–48 and a slowpath on lines 49–62. A
line-by-line analysis of this function is left as an exercise to the reader.
Listing 5.14 shows read_count(). Line 9 acquires gblcnt_mutex and line 17
releases it. Line 10 initializes local variable sum to the value of globalcount, and
v2023.06.11a
Listing 5.15: Atomic Limit Counter Utility Functions 1

1 static void globalize_count(void)
2 {
3 int c;
4 int cm;
5 int old;
6
8 globalcount += c;
9 globalreserve -= cm;
10 old = merge_counterandmax(0, 0);
11 atomic_set(&counterandmax, old);
12 }
13
14 static void flush_local_count(void)
15 {
16 int c;
17 int cm;
18 int old;
19 int t;
20 int zero;
21
22 if (globalreserve == 0)
23 return;
24 zero = merge_counterandmax(0, 0);
26 if (counterp[t] != NULL) {
27 old = atomic_xchg(counterp[t], zero);
28 split_counterandmax_int(old, &c, &cm);
29 globalcount += c;
30 globalreserve -= cm;
31 }
32 }
the loop spanning lines 11–16 adds the per-thread counters to this sum, isolating each
per-thread counter using split_counterandmax on line 13. Finally, line 18 returns
the sum.
Listings 5.15 and 5.16 show the utility functions globalize_count(), flush_
local_count(), balance_count(), count_register_thread(), and count_
unregister_thread(). The code for globalize_count() is shown on lines 1–12
of Listing 5.15, and is similar to that of previous algorithms, with the addition of line 7,
which is now required to split out counter and countermax from counterandmax.
The code for flush_local_count(), which moves all threads’ local counter state
to the global counter, is shown on lines 14–32. Line 22 checks to see if the value of
globalreserve permits any per-thread counts, and, if not, line 23 returns. Otherwise,
line 24 initializes local variable zero to a combined zeroed counter and countermax.
The loop spanning lines 25–31 sequences through each thread. Line 26 checks to see if
the current thread has counter state, and, if so, lines 27–30 move that state to the global
counters. Line 27 atomically fetches the current thread’s state while replacing it with
zero. Line 28 splits this state into its counter (in local variable c) and countermax (in
local variable cm) components. Line 29 adds this thread’s counter to globalcount,
while line 30 subtracts this thread’s countermax from globalreserve.
Quick Quiz 5.44: What stops a thread from simply refilling its counterandmax variable
immediately after flush_local_count() on line 14 of Listing 5.15 empties it?
Quick Quiz 5.45: What prevents concurrent execution of the fastpath of either add_count()
or sub_count() from interfering with the counterandmax variable while flush_local_
count() is accessing it on line 27 of Listing 5.15?
v2023.06.11a
Listing 5.16: Atomic Limit Counter Utility Functions 2

2 {
3 int c;
4 int cm;
5 int old;
6 unsigned long limit;
7
8 limit = globalcountmax - globalcount -
9 globalreserve;
10 limit /= num_online_threads();
11 if (limit > MAX_COUNTERMAX)
12 cm = MAX_COUNTERMAX;
13 else
14 cm = limit;
15 globalreserve += cm;
16 c = cm / 2;
17 if (c > globalcount)
18 c = globalcount;
19 globalcount -= c;
20 old = merge_counterandmax(c, cm);
21 atomic_set(&counterandmax, old);
22 }
23
25 {
27
29 counterp[idx] = &counterandmax;
31 }
32
34 {
36
41 }
Lines 1–22 on Listing 5.16 show the code for balance_count(), which refills
the calling thread’s local counterandmax variable. This function is quite similar
to that of the preceding algorithms, with changes required to handle the merged
counterandmax variable. Detailed analysis of the code is left as an exercise for the
reader, as it is with the count_register_thread() function starting on line 24 and
the count_unregister_thread() function starting on line 33.
Quick Quiz 5.46: Given that the atomic_set() primitive does a simple store to the specified
atomic_t, how can line 21 of balance_count() in Listing 5.16 work correctly in face of
concurrent flush_local_count() updates to this variable?
The next section qualitatively evaluates this design.
5.4.2 Atomic Limit Counter Discussion

This is the first implementation that actually allows the counter to be run all the way
to either of its limits, but it does so at the expense of adding atomic operations to the
fastpaths, which slow down the fastpaths significantly on some systems. Although some
workloads might tolerate this slowdown, it is worthwhile looking for algorithms with
better write-side performance. One such algorithm uses a signal handler to steal counts
v2023.06.11a
IDLE
need no
flushed
flush count
!counting
REQ READY
done
counting
counting
ACK
Figure 5.7: Signal-Theft State Machine
from other threads. Because signal handlers run in the context of the signaled thread,
atomic operations are not necessary, as shown in the next section.
Quick Quiz 5.47: But signal handlers can be migrated to some other CPU while running.
Doesn’t this possibility require that atomic instructions and memory barriers are required to
reliably communicate between a thread and a signal handler that interrupts that thread?
5.4.3 Signal-Theft Limit Counter Design

Even though per-thread state will now be manipulated only by the corresponding thread,
there will still need to be synchronization with the signal handlers. This synchronization
is provided by the state machine shown in Figure 5.7.
The state machine starts out in the IDLE state, and when add_count() or sub_
count() find that the combination of the local thread’s count and the global count
cannot accommodate the request, the corresponding slowpath sets each thread’s theft
state to REQ (unless that thread has no count, in which case it transitions directly to
READY). Only the slowpath, which holds the gblcnt_mutex lock, is permitted to
transition from the IDLE state, as indicated by the green color.4 The slowpath then sends
a signal to each thread, and the corresponding signal handler checks the corresponding
thread’s theft and counting variables. If the theft state is not REQ, then the signal
handler is not permitted to change the state, and therefore simply returns. Otherwise, if
the counting variable is set, indicating that the current thread’s fastpath is in progress,
the signal handler sets the theft state to ACK, otherwise to READY.
If the theft state is ACK, only the fastpath is permitted to change the theft state,
as indicated by the blue color. When the fastpath completes, it sets the theft state to
READY.
Once the slowpath sees a thread’s theft state is READY, the slowpath is permitted
to steal that thread’s count. The slowpath then sets that thread’s theft state to IDLE.
4 For those with black-and-white versions of this book, IDLE and READY are green,
REQ is red, and ACK is blue.
v2023.06.11a
Listing 5.17: Signal-Theft Limit Counter Data

1 #define THEFT_IDLE 0
2 #define THEFT_REQ 1
3 #define THEFT_ACK 2
4 #define THEFT_READY 3
5
6 int __thread theft = THEFT_IDLE;
7 int __thread counting = 0;
14 unsigned long *countermaxp[NR_THREADS] = { NULL };
15 int *theftp[NR_THREADS] = { NULL };
17 #define MAX_COUNTERMAX 100
Quick Quiz 5.48: In Figure 5.7, why is the REQ theft state colored red?
Quick Quiz 5.49: In Figure 5.7, what is the point of having separate REQ and ACK theft
states? Why not simplify the state machine by collapsing them into a single REQACK state?
Then whichever of the signal handler or the fastpath gets there first could set the state to READY.
5.4.4 Signal-Theft Limit Counter Implementation

Listing 5.17 (count_lim_sig.c) shows the data structures used by the signal-theft
based counter implementation. Lines 1–7 define the states and values for the per-thread
theft state machine described in the preceding section. Lines 8–17 are similar to earlier
implementations, with the addition of lines 14 and 15 to allow remote access to a
thread’s countermax and theft variables, respectively.
Listing 5.18 shows the functions responsible for migrating counts between per-thread
variables and the global variables. Lines 1–7 show globalize_count(), which is
identical to earlier implementations. Lines 9–16 show flush_local_count_sig(),
which is the signal handler used in the theft process. Lines 11 and 12 check to see if the
theft state is REQ, and, if not returns without change. Line 13 sets the theft state
to ACK, and, if line 14 sees that this thread’s fastpaths are not running, line 15 uses
smp_store_release() to set the theft state to READY, further ensuring that any
change to counter in the fastpath happens before this change of theft to READY.
Quick Quiz 5.50: In Listing 5.18, doesn’t flush_local_count_sig() need stronger
memory barriers?
Lines 18–47 show flush_local_count(), which is called from the slowpath to

flush all threads’ local counts. The loop spanning lines 23–32 advances the theft state
for each thread that has local count, and also sends that thread a signal. Line 24 skips
any non-existent threads. Otherwise, line 25 checks to see if the current thread holds
any local count, and, if not, line 26 sets the thread’s theft state to READY and line 27
skips to the next thread. Otherwise, line 29 sets the thread’s theft state to REQ and
line 30 sends the thread a signal.
Quick Quiz 5.51: In Listing 5.18, why is it safe for line 25 to directly access the other thread’s
countermax variable?
v2023.06.11a
Listing 5.18: Signal-Theft Limit Counter Value-Migration Functions

1 static void globalize_count(void)
2 {
3 globalcount += counter;
4 counter = 0;
5 globalreserve -= countermax;
6 countermax = 0;
7 }
8
9 static void flush_local_count_sig(int unused)
10 {
11 if (READ_ONCE(theft) != THEFT_REQ)
12 return;
13 WRITE_ONCE(theft, THEFT_ACK);
14 if (!counting)
15 smp_store_release(&theft, THEFT_READY);
16 }
17
18 static void flush_local_count(void)
19 {
20 int t;
21 thread_id_t tid;
22
23 for_each_tid(t, tid) {
24 if (theftp[t] != NULL) {
25 if (*countermaxp[t] == 0) {
26 WRITE_ONCE(*theftp[t], THEFT_READY);
27 continue;
28 }
29 WRITE_ONCE(*theftp[t], THEFT_REQ);
30 pthread_kill(tid, SIGUSR1);
31 }
32 }
33 for_each_tid(t, tid) {
34 if (theftp[t] == NULL)
35 continue;
36 while (smp_load_acquire(theftp[t]) != THEFT_READY) {
38 if (READ_ONCE(*theftp[t]) == THEFT_REQ)
39 pthread_kill(tid, SIGUSR1);
40 }
41 globalcount += *counterp[t];
42 *counterp[t] = 0;
43 globalreserve -= *countermaxp[t];
44 *countermaxp[t] = 0;
45 smp_store_release(theftp[t], THEFT_IDLE);
46 }
47 }
48
50 {
51 countermax = globalcountmax - globalcount -
52 globalreserve;
54 if (countermax > MAX_COUNTERMAX)
55 countermax = MAX_COUNTERMAX;
61 }
v2023.06.11a
Listing 5.19: Signal-Theft Limit Counter Add Function

1 int add_count(unsigned long delta)
2 {
3 int fastpath = 0;
4
5 WRITE_ONCE(counting, 1);
6 barrier();
7 if (smp_load_acquire(&theft) <= THEFT_REQ &&
8 countermax - counter >= delta) {
10 fastpath = 1;
11 }
12 barrier();
14 barrier();
15 if (READ_ONCE(theft) == THEFT_ACK)
17 if (fastpath)
18 return 1;
27 return 0;
28 }
29 }
31 balance_count();
33 return 1;
34 }
Quick Quiz 5.52: In Listing 5.18, why doesn’t line 30 check for the current thread sending
itself a signal?
Quick Quiz 5.53: The code shown in Listings 5.17 and 5.18 works with GCC and POSIX.
What would be required to make it also conform to the ISO C standard?
The loop spanning lines 33–46 waits until each thread reaches READY state, then steals
that thread’s count. Lines 34–35 skip any non-existent threads, and the loop spanning
lines 36–40 waits until the current thread’s theft state becomes READY. Line 37
blocks for a millisecond to avoid priority-inversion problems, and if line 38 determines
that the thread’s signal has not yet arrived, line 39 resends the signal. Execution reaches
line 41 when the thread’s theft state becomes READY, so lines 41–44 do the thieving.
Line 45 then sets the thread’s theft state back to IDLE.
Quick Quiz 5.54: In Listing 5.18, why does line 39 resend the signal?
Lines 49–61 show balance_count(), which is similar to that of earlier examples.

Listing 5.19 shows the add_count() function. The fastpath spans lines 5–18, and
the slowpath lines 19–33. Line 5 sets the per-thread counting variable to 1 so that
any subsequent signal handlers interrupting this thread will set the theft state to ACK
rather than READY, allowing this fastpath to complete properly. Line 6 prevents the
compiler from reordering any of the fastpath body to precede the setting of counting.
Lines 7 and 8 check to see if the per-thread data can accommodate the add_count()
and if there is no ongoing theft in progress, and if so line 9 does the fastpath addition
and line 10 notes that the fastpath was taken.
v2023.06.11a
Listing 5.20: Signal-Theft Limit Counter Subtract Function

1 int sub_count(unsigned long delta)
2 {
3 int fastpath = 0;
4
6 barrier();
7 if (smp_load_acquire(&theft) <= THEFT_REQ &&
8 counter >= delta) {
9 WRITE_ONCE(counter, counter - delta);
10 fastpath = 1;
11 }
12 barrier();
14 barrier();
15 if (READ_ONCE(theft) == THEFT_ACK)
17 if (fastpath)
18 return 1;
25 return 0;
26 }
27 }
29 balance_count();
31 return 1;
32 }
Listing 5.21: Signal-Theft Limit Counter Read Function

2 {
3 int t;
5
11 }
13 return sum;
14 }
In either case, line 12 prevents the compiler from reordering the fastpath body to follow
line 13, which permits any subsequent signal handlers to undertake theft. Line 14 again
disables compiler reordering, and then line 15 checks to see if the signal handler deferred
the theft state-change to READY, and, if so, line 16 uses smp_store_release() to
set the theft state to READY, further ensuring that any CPU that sees the READY
state also sees the effects of line 9. If the fastpath addition at line 9 was executed, then
line 18 returns success.
Otherwise, we fall through to the slowpath starting at line 19. The structure of the
slowpath is similar to those of earlier examples, so its analysis is left as an exercise
to the reader. Similarly, the structure of sub_count() on Listing 5.20 is the same as
that of add_count(), so the analysis of sub_count() is also left as an exercise for the
reader, as is the analysis of read_count() in Listing 5.21.
v2023.06.11a
Listing 5.22: Signal-Theft Limit Counter Initialization Functions

2 {
3 struct sigaction sa;
4
5 sa.sa_handler = flush_local_count_sig;
6 sigemptyset(&sa.sa_mask);
7 sa.sa_flags = 0;
8 if (sigaction(SIGUSR1, &sa, NULL) != 0) {
9 perror("sigaction");
11 }
12 }
13
15 {
17
20 countermaxp[idx] = &countermax;
21 theftp[idx] = &theft;
23 }
24
26 {
28
32 countermaxp[idx] = NULL;
33 theftp[idx] = NULL;
35 }
Lines 1–12 of Listing 5.22 show count_init(), which set up flush_local_

count_sig() as the signal handler for SIGUSR1, enabling the pthread_kill() calls
in flush_local_count() to invoke flush_local_count_sig(). The code for
thread registry and unregistry is similar to that of earlier examples, so its analysis is left
as an exercise for the reader.
5.4.5 Signal-Theft Limit Counter Discussion

The signal-theft implementation runs more than eight times as fast as the atomic
implementation on my six-core x86 laptop. Is it always preferable?
The signal-theft implementation would be vastly preferable on Pentium-4 systems,
given their slow atomic instructions, but the old 80386-based Sequent Symmetry
systems would do much better with the shorter path length of the atomic implementation.
However, this increased update-side performance comes at the prices of higher read-side
overhead: Those POSIX signals are not free. If ultimate performance is of the essence,
you will need to measure them both on the system that your application is to be deployed
on.
Quick Quiz 5.55: Not only are POSIX signals slow, sending one to each thread simply does
not scale. What would you do if you had (say) 10,000 threads and needed the read side to be
fast?
This is but one reason why high-quality APIs are so important: They permit
implementations to be changed as required by ever-changing hardware performance
characteristics.
v2023.06.11a
Quick Quiz 5.56: What if you want an exact limit counter to be exact only for its lower limit,
but to allow the upper limit to be inexact?
5.4.6 Applying Exact Limit Counters

Although the exact limit counter implementations presented in this section can be very
useful, they are not much help if the counter’s value remains near zero at all times, as it
might when counting the number of outstanding accesses to an I/O device. The high
overhead of such near-zero counting is especially painful given that we normally don’t
care how many references there are. As noted in the removable I/O device access-count
problem posed by Quick Quiz 5.5, the number of accesses is irrelevant except in those
rare cases when someone is actually trying to remove the device.
One simple solution to this problem is to add a large “bias” (for example, one billion)
to the counter in order to ensure that the value is far enough from zero that the counter
can operate efficiently. When someone wants to remove the device, this bias is subtracted
from the counter value. Counting the last few accesses will be quite inefficient, but the
important point is that the many prior accesses will have been counted at full speed.
Quick Quiz 5.57: What else had you better have done when using a biased counter?
Although a biased counter can be quite helpful and useful, it is only a partial solution
to the removable I/O device access-count problem called out on page 77. When
attempting to remove a device, we must not only know the precise number of current
I/O accesses, we also need to prevent any future accesses from starting. One way to
accomplish this is to read-acquire a reader-writer lock when updating the counter, and to
write-acquire that same reader-writer lock when checking the counter. Code for doing
I/O might be as follows:
1 read_lock(&mylock);
2 if (removing) {
3 read_unlock(&mylock);
4 cancel_io();
5 } else {
6 add_count(1);
7 read_unlock(&mylock);
8 do_io();
9 sub_count(1);
10 }
Line 1 read-acquires the lock, and either line 3 or 7 releases it. Line 2 checks to see
if the device is being removed, and, if so, line 3 releases the lock and line 4 cancels
the I/O, or takes whatever action is appropriate given that the device is to be removed.
Otherwise, line 6 increments the access count, line 7 releases the lock, line 8 performs
the I/O, and line 9 decrements the access count.
Quick Quiz 5.58: This is ridiculous! We are read-acquiring a reader-writer lock to update the
counter? What are you playing at???
The code to remove the device might be as follows:
1 write_lock(&mylock);
2 removing = 1;
3 sub_count(mybias);
4 write_unlock(&mylock);
5 while (read_count() != 0)
v2023.06.11a
Table 5.1: Statistical/Limit Counter Performance on x86
Exact?
Reads (ns)
Algorithm Updates
(count_*.c) Section (ns) 1 CPU 8 CPUs 64 CPUs 420 CPUs
stat 5.2.2 6.3 294 303 315 612
stat_eventual 5.2.4 6.4 1 1 1 1
end 5.2.3 2.9 301 6,309 147,594 239,683
end_rcu 13.5.1 2.9 454 481 508 2,317
lim 5.3.2 N 3.2 435 6,678 156,175 239,422
lim_app 5.3.4 N 2.4 485 7,041 173,108 239,682
lim_atomic 5.4.1 Y 19.7 513 7,085 199,957 239,450
lim_sig 5.4.4 Y 4.7 519 6,805 120,000 238,811
6 poll(NULL, 0, 1);
7 remove_device();
Line 1 write-acquires the lock and line 4 releases it. Line 2 notes that the device is
being removed, and the loop spanning lines 5–6 waits for any I/O operations to complete.
Finally, line 7 does any additional processing needed to prepare for device removal.
Quick Quiz 5.59: What other issues would need to be accounted for in a real system?
5.5 Parallel Counting Discussion

This idea that there is generality in the specific is of
far-reaching importance.
Douglas R. Hofstadter
This chapter has presented the reliability, performance, and scalability problems with
traditional counting primitives. The C-language ++ operator is not guaranteed to
function reliably in multithreaded code, and atomic operations to a single variable
neither perform nor scale well. This chapter therefore presented a number of counting
algorithms that perform and scale extremely well in certain special cases.
It is well worth reviewing the lessons from these counting algorithms. To that end,
Section 5.5.1 overviews requisite validation, Section 5.5.2 summarizes performance and
scalability, Section 5.5.3 discusses the need for specialization, and finally, Section 5.5.4
enumerates lessons learned and calls attention to later chapters that will expand on these
lessons.
5.5.1 Parallel Counting Validation

Many of the algorithms in this section are quite simple, so much so that it is tempting
to declare them to be correct by construction or by inspection. Unfortunately, it
is all too easy for those carrying out the construction or the inspection to become
overconfident, tired, confused, or just plain sloppy, all of which can result in bugs. And
early implementations of these limit counters have in fact contained bugs, in some cases
v2023.06.11a
5.5. PARALLEL COUNTING DISCUSSION 109
aided and abetted by the complexities inherent in maintaining a 64-bit count on a 32-bit
system. Therefore, validation is not optional, even for the simple algorithms presented
in this chapter.
The statistical counters are tested for acting like counters (“counttorture.h”), that
is, that the aggregate sum in the counter changes by the sum of the amounts added by
the various update-side threads.
The limit counters are also tested for acting like counters (“limtorture.h”), and
additionally checked for their ability to accommodate the specified limit.
Both of these test suites produce performance data that is used in Section 5.5.2.
Although this level of validation is good and sufficient for textbook implementations
such as these, it would be wise to apply additional validation before putting similar
algorithms into production. Chapter 11 describes additional approaches to testing,
and given the simplicity of most of these counting algorithms, most of the techniques
described in Chapter 12 can also be quite helpful.
5.5.2 Parallel Counting Performance

The top half of Table 5.1 shows the performance of the four parallel statistical counting
algorithms. All four algorithms provide near-perfect linear scalability for updates. The
per-thread-variable implementation (count_end.c) is significantly faster on updates
than the array-based implementation (count_stat.c), but is slower at reads on large
numbers of core, and suffers severe lock contention when there are many parallel readers.
This contention can be addressed using the deferred-processing techniques introduced in
Chapter 9, as shown on the count_end_rcu.c row of Table 5.1. Deferred processing
also shines on the count_stat_eventual.c row, courtesy of eventual consistency.
Quick Quiz 5.60: On the count_stat.c row of Table 5.1, we see that the read-side scales
linearly with the number of threads. How is that possible given that the more threads there are,
the more per-thread counters must be summed up?
Quick Quiz 5.61: Even on the fourth row of Table 5.1, the read-side performance of these
statistical counter implementations is pretty horrible. So why bother with them?
The bottom half of Table 5.1 shows the performance of the parallel limit-counting
algorithms. Exact enforcement of the limits incurs a substantial update-side performance
penalty, although on this x86 system that penalty can be reduced by substituting signals
for atomic operations. All of these implementations suffer from read-side lock contention
in the face of concurrent readers.
Quick Quiz 5.62: Given the performance data shown in the bottom half of Table 5.1, we
should always prefer signals over atomic operations, right?
Quick Quiz 5.63: Can advanced techniques be applied to address the lock contention for
readers seen in the bottom half of Table 5.1?
In short, this chapter has demonstrated a number of counting algorithms that perform
and scale extremely well in a number of special cases. But must our parallel counting
be confined to special cases? Wouldn’t it be better to have a general algorithm that
operated efficiently in all cases? The next section looks at these questions.
v2023.06.11a
5.5.3 Parallel Counting Specializations

The fact that these algorithms only work well in their respective special cases might
be considered a major problem with parallel programming in general. After all, the
C-language ++ operator works just fine in single-threaded code, and not just for special
cases, but in general, right?
This line of reasoning does contain a grain of truth, but is in essence misguided.
The problem is not parallelism as such, but rather scalability. To understand this, first
consider the C-language ++ operator. The fact is that it does not work in general, only
for a restricted range of numbers. If you need to deal with 1,000-digit decimal numbers,
the C-language ++ operator will not work for you.
Quick Quiz 5.64: The ++ operator works just fine for 1,000-digit numbers! Haven’t you heard
of operator overloading???
This problem is not specific to arithmetic. Suppose you need to store and query data.
Should you use an ASCII file? XML? A relational database? A linked list? A dense
array? A B-tree? A radix tree? Or one of the plethora of other data structures and
environments that permit data to be stored and queried? It depends on what you need
to do, how fast you need it done, and how large your data set is—even on sequential
systems.
Similarly, if you need to count, your solution will depend on how large of numbers
you need to work with, how many CPUs need to be manipulating a given number
concurrently, how the number is to be used, and what level of performance and scalability
you will need.
Nor is this problem specific to software. The design for a bridge meant to allow people
to walk across a small brook might be a simple as a single wooden plank. But you would
probably not use a plank to span the kilometers-wide mouth of the Columbia River, nor
would such a design be advisable for bridges carrying concrete trucks. In short, just
as bridge design must change with increasing span and load, so must software design
change as the number of CPUs increases. That said, it would be good to automate this
process, so that the software adapts to changes in hardware configuration and in workload.
There has in fact been some research into this sort of automation [AHS+ 03, SAH+ 03],
and the Linux kernel does some boot-time reconfiguration, including limited binary
rewriting. This sort of adaptation will become increasingly important as the number of
CPUs on mainstream systems continues to increase.
In short, as discussed in Chapter 3, the laws of physics constrain parallel software
just as surely as they constrain mechanical artifacts such as bridges. These constraints
force specialization, though in the case of software it might be possible to automate the
choice of specialization to fit the hardware and workload in question.
Of course, even generalized counting is quite specialized. We need to do a great
number of other things with computers. The next section relates what we have learned
from counters to topics taken up later in this book.
5.5.4 Parallel Counting Lessons

The opening paragraph of this chapter promised that our study of counting would
provide an excellent introduction to parallel programming. This section makes explicit
connections between the lessons from this chapter and the material presented in a
number of later chapters.
v2023.06.11a
5.5. PARALLEL COUNTING DISCUSSION 111
The examples in this chapter have shown that an important scalability and performance
tool is partitioning. The counters might be fully partitioned, as in the statistical counters
discussed in Section 5.2, or partially partitioned as in the limit counters discussed in
Sections 5.3 and 5.4. Partitioning will be considered in far greater depth in Chapter 6,
and partial parallelization in particular in Section 6.4, where it is called parallel fastpath.
Quick Quiz 5.65: But if we are going to have to partition everything, why bother with
shared-memory multithreading? Why not just partition the problem completely and run as
multiple processes, each in its own address space?
The partially partitioned counting algorithms used locking to guard the global data,
and locking is the subject of Chapter 7. In contrast, the partitioned data tended to be fully
under the control of the corresponding thread, so that no synchronization whatsoever
was required. This data ownership will be introduced in Section 6.3.4 and discussed in
more detail in Chapter 8.
Because integer addition and subtraction are extremely cheap compared to typical
synchronization operations, achieving reasonable scalability requires synchronization
operations be used sparingly. One way of achieving this is to batch the addition and
subtraction operations, so that a great many of these cheap operations are handled by
a single synchronization operation. Batching optimizations of one sort or another are
used by each of the counting algorithms listed in Table 5.1.
Finally, the eventually consistent statistical counter discussed in Section 5.2.4 showed
how deferring activity (in that case, updating the global counter) can provide substantial
performance and scalability benefits. This approach allows common case code to use
much cheaper synchronization operations than would otherwise be possible. Chapter 9
will examine a number of additional ways that deferral can improve performance,
scalability, and even real-time response.
Summarizing the summary:
1. Partitioning promotes performance and scalability.
2. Partial partitioning, that is, partitioning applied only to common code paths, works
almost as well.
3. Partial partitioning can be applied to code (as in Section 5.2’s statistical counters’
partitioned updates and non-partitioned reads), but also across time (as in Sec-
tion 5.3’s and Section 5.4’s limit counters running fast when far from the limit, but
slowly when close to the limit).
4. Partitioning across time often batches updates locally in order to reduce the number
of expensive global operations, thereby decreasing synchronization overhead, in
turn improving performance and scalability. All the algorithms shown in Table 5.1
make heavy use of batching.
5. Read-only code paths should remain read-only: Spurious synchronization writes
to shared memory kill performance and scalability, as seen in the count_end.c
row of Table 5.1.
6. Judicious use of delay promotes performance and scalability, as seen in Sec-
tion 5.2.4.
7. Parallel performance and scalability is usually a balancing act: Beyond a certain
point, optimizing some code paths will degrade others. The count_stat.c and
count_end_rcu.c rows of Table 5.1 illustrate this point.
v2023.06.11a
Batch
Work
Partitioning
Resource
Parallel
Partitioning and
Interacting
With Hardware
Weaken Partition
Figure 5.8: Optimization and the Four Parallel-Programming Tasks
8. Different levels of performance and scalability will affect algorithm and data-
structure design, as do a large number of other factors. Figure 5.1 illustrates this
point: Atomic increment might be completely acceptable for a two-CPU system,
but nevertheless be completely inadequate for an eight-CPU system.
Summarizing still further, we have the “big three” methods of increasing performance
and scalability, namely (1) partitioning over CPUs or threads, (2) batching so that more
work can be done by each expensive synchronization operation, and (3) weakening
synchronization operations where feasible. As a rough rule of thumb, you should
apply these methods in this order, as was noted earlier in the discussion of Figure 2.6
on page 22. The partitioning optimization applies to the “Resource Partitioning and
Replication” bubble, the batching optimization to the “Work Partitioning” bubble,
and the weakening optimization to the “Parallel Access Control” bubble, as shown
in Figure 5.8. Of course, if you are using special-purpose hardware such as digital
signal processors (DSPs), field-programmable gate arrays (FPGAs), or general-purpose
graphical processing units (GPGPUs), you may need to pay close attention to the
“Interacting With Hardware” bubble throughout the design process. For example, the
structure of a GPGPU’s hardware threads and memory connectivity might richly reward
very careful partitioning and batching design decisions.
In short, as noted at the beginning of this chapter, the simplicity of counting have
allowed us to explore many fundamental concurrency issues without the distraction of
complex synchronization primitives or elaborate data structures. Such synchronization
primitives and data structures are covered in later chapters.
v2023.06.11a
Divide and rule.
Philip II of Macedon
Chapter 6
Partitioning and
Synchronization Design
This chapter describes how to design software to take advantage of modern commodity
multicore systems by using idioms, or “design patterns” [Ale79, GHJV95, SSRB00], to
balance performance, scalability, and response time. Correctly partitioned problems lead
to simple, scalable, and high-performance solutions, while poorly partitioned problems
result in slow and complex solutions. This chapter will help you design partitioning
into your code, with some discussion of batching and weakening as well. The word
“design” is very important: You should partition first, batch second, weaken third, and
code fourth. Changing this order often leads to poor performance and scalability along
with great frustration.1
This chapter will also look at some specific problems, including:
1. Constraints on the classic Dining Philosophers problem requiring that all the
philophers be able to dine concurrently.
2. Lock-based double-ended queue implementations that provide concurrency between
operations on both ends of a given queue when there are many elements in the
queue, but still work correctly when the queue contains only a few elements. (Or,
for that matter, no elements.)
3. Summarizing the rough quality of a concurrent algorithm with only a few numbers.
4. Selecting the right granularity of partitioning.

5. Comcurrent designs for applications that do not fully partition.
6. Obtaining more than 2x speedup from two CPUs.
To this end, Section 6.1 presents partitioning exercises, Section 6.2 reviews parti-
tionability design criteria, Section 6.3 discusses synchronization granularity selection,
Section 6.4 overviews important parallel-fastpath design patterns that provide speed
and scalability on common-case fastpaths while using simpler less-scalable “slow
path” fallbacks for unusual situations, and finally Section 6.5 takes a brief look beyond
partitioning.
1 That other great dodge around the Laws of Physics, read-only replication, is covered in
Chapter 9.
113
v2023.06.11a
114 CHAPTER 6. PARTITIONING AND SYNCHRONIZATION DESIGN
P1
P5 P2
P4 P3
Figure 6.1: Dining Philosophers Problem
6.1 Partitioning Exercises

Whenever a theory appears to you as the only
possible one, take this as a sign that you have
neither understood the theory nor the problem
which it was intended to solve.
Karl Popper
Although partitioning is more widely understood than it was in the early 2000s, its value
is still underappreciated. Section 6.1.1 therefore takes more highly parallel look at the
classic Dining Philosophers problem and Section 6.1.2 revisits the double-ended queue.
6.1.1 Dining Philosophers Problem

Figure 6.1 shows a diagram of the classic Dining Philosophers problem [Dĳ71]. This
problem features five philosophers who do nothing but think and eat a “very difficult
kind of spaghetti” which requires two forks to eat.2 A given philosopher is permitted to
use only the forks to his or her immediate right and left, but will not put a given fork
down until sated.
The object is to construct an algorithm that, quite literally, prevents starvation. One
starvation scenario would be if all of the philosophers picked up their leftmost forks
simultaneously. Because none of them will put down their fork until after they finished
eating, and because none of them may pick up their second fork until at least one of
them has finished eating, they all starve. Please note that it is not sufficient to allow
at least one philosopher to eat. As Figure 6.2 shows, starvation of even a few of the
philosophers is to be avoided.
Dĳkstra’s solution used a global semaphore, which works fine assuming negligible
communications delays, an assumption that became invalid in the late 1980s or early
2 But feel free to instead think in terms of chopsticks.
v2023.06.11a
6.1. PARTITIONING EXERCISES 115
Figure 6.2: Partial Starvation Is Also Bad
P1
5 1
P5 P2
4 2
P4 P3
Figure 6.3: Dining Philosophers Problem, Textbook Solution
v2023.06.11a
1990s.3 More recent solutions number the forks as shown in Figure 6.3. Each philosopher
picks up the lowest-numbered fork next to his or her plate, then picks up the other fork.
The philosopher sitting in the uppermost position in the diagram thus picks up the
leftmost fork first, then the rightmost fork, while the rest of the philosophers instead
pick up their rightmost fork first. Because two of the philosophers will attempt to pick
up fork 1 first, and because only one of those two philosophers will succeed, there will
be five forks available to four philosophers. At least one of these four will have two
forks, and will thus be able to eat.
This general technique of numbering resources and acquiring them in numerical
order is heavily used as a deadlock-prevention technique. However, it is easy to imagine
a sequence of events that will result in only one philosopher eating at a time even though
all are hungry:
1. P2 picks up fork 1, preventing P1 from taking a fork.

2. P3 picks up fork 2.
5. P5 picks up fork 5 and eats.
6. P5 puts down forks 4 and 5.
7. P4 picks up fork 4 and eats.
In short, this algorithm can result in only one philosopher eating at a given time, even
when all five philosophers are hungry, despite the fact that there are more than enough
forks for two philosophers to eat concurrently. It should be possible to do better than
this!
One approach is shown in Figure 6.4, which includes four philosophers rather than five
to better illustrate the partition technique. Here the upper and rightmost philosophers
share a pair of forks, while the lower and leftmost philosophers share another pair of
forks. If all philosophers are simultaneously hungry, at least two will always be able to
eat concurrently. In addition, as shown in the figure, the forks can now be bundled so
that the pair are picked up and put down simultaneously, simplifying the acquisition and
release algorithms.
Quick Quiz 6.1: Is there a better solution to the Dining Philosophers Problem?
Quick Quiz 6.2: How would you valididate an algorithm alleged to solve the Dining
Philosophers Problem?
This is an example of “horizontal parallelism” [Inm85] or “data parallelism”, so

named because there is no dependency among the pairs of philosophers. In a horizontally
parallel data-processing system, a given item of data would be processed by only one of
a replicated set of software components.
Quick Quiz 6.3: And in just what sense can this “horizontal parallelism” be said to be
“horizontal”?
3 It is all too easy to denigrate Dĳkstra from the viewpoint of the year 2021, more than
50 years after the fact. If you still feel the need to denigrate Dĳkstra, my advice is to publish
something, wait 50 years, and then see how well your ideas stood the test of time.
v2023.06.11a
P1
P4 P2
P3
Figure 6.4: Dining Philosophers Problem, Partitioned
6.1.2 Double-Ended Queue

A double-ended queue is a data structure containing a list of elements that may be
inserted or removed from either end [Knu73]. It has been claimed that a lock-based
implementation permitting concurrent operations on both ends of the double-ended
queue is difficult [Gro07]. This section shows how a partitioning design strategy can
result in a reasonably simple implementation, looking at three general approaches in the
following sections. But first, how should we validate a concurrent double-ended queue?
6.1.2.1 Double-Ended Queue Validation

A good place to start is with invariants. For example, if elements are pushed onto one
end of a double-ended queue and popped off of the other, the order of those elements
must be preserved. Similarly, if elements are pushed onto one end of the queue and
popped off of that same end, the order of those elements must be reversed. Any element
popped from the queue must have been most recently pushed onto that queue, and if the
queue is emptied, all elements pushed onto it must have already been popped from it.
The beginnings of a test suite for concurrent double-ended queues (“deqtorture.h”)
provides the following checks:
1. Element-ordering checks provided by CHECK_SEQUENCE_PAIR().

2. Checks that elements popped were most recently pushed, provided by melee().
3. Checks that elements pushed are popped before the queue is emptied, also provided
by melee().
This suite includes both sequential and concurrent tests. Although this suite is good
and sufficient for textbook code, you should test considerably more thoroughly for code
intended for production use. Chapters 11 and 12 cover a large array of validation tools
and techniques.
But with a prototype test suite in place, we are ready to look at the double-ended-queue
algorithms in the next sections.
v2023.06.11a
Lock L Lock R
Header L Header R
Lock L Lock R
Header L 0 Header R
Lock L Lock R
Header L 0 1 Header R
Lock L Lock R
Header L 0 1 2 Header R
Lock L Lock R
Header L 0 1 2 3 Header R
Figure 6.5: Double-Ended Queue With Left- and Right-Hand Locks
Lock L Lock R
DEQ L DEQ R
Figure 6.6: Compound Double-Ended Queue
6.1.2.2 Left- and Right-Hand Locks

One seemingly straightforward approach would be to use a doubly linked list with a
left-hand lock for left-hand-end enqueue and dequeue operations along with a right-hand
lock for right-hand-end operations, as shown in Figure 6.5. However, the problem
with this approach is that the two locks’ domains must overlap when there are fewer
than four elements on the list. This overlap is due to the fact that removing any given
element affects not only that element, but also its left- and right-hand neighbors. These
domains are indicated by color in the figure, with blue with downward stripes indicating
the domain of the left-hand lock, red with upward stripes indicating the domain of the
right-hand lock, and purple (with no stripes) indicating overlapping domains. Although
it is possible to create an algorithm that works this way, the fact that it has no fewer than
five special cases should raise a big red flag, especially given that concurrent activity at
the other end of the list can shift the queue from one special case to another at any time.
It is far better to consider other designs.
6.1.2.3 Compound Double-Ended Queue

One way of forcing non-overlapping lock domains is shown in Figure 6.6. Two separate
double-ended queues are run in tandem, each protected by its own lock. This means
that elements must occasionally be shuttled from one of the double-ended queues to
the other, in which case both locks must be held. A simple lock hierarchy may be used
v2023.06.11a
DEQ 0 DEQ 1 DEQ 2 DEQ 3
Lock 0 Lock 1 Lock 2 Lock 3
Index L Index R
Lock L Lock R
Figure 6.7: Hashed Double-Ended Queue
to avoid deadlock, for example, always acquiring the left-hand lock before acquiring
the right-hand lock. This will be much simpler than applying two locks to the same
double-ended queue, as we can unconditionally left-enqueue elements to the left-hand
queue and right-enqueue elements to the right-hand queue. The main complication
arises when dequeuing from an empty queue, in which case it is necessary to:
1. If holding the right-hand lock, release it and acquire the left-hand lock.
2. Acquire the right-hand lock.
3. Rebalance the elements across the two queues.
4. Remove the required element if there is one.
5. Release both locks.
Quick Quiz 6.4: In this compound double-ended queue implementation, what should be done
if the queue has become non-empty while releasing and reacquiring the lock?
The resulting code (locktdeq.c) is quite straightforward. The rebalancing operation

might well shuttle a given element back and forth between the two queues, wasting time
and possibly requiring workload-dependent heuristics to obtain optimal performance.
Although this might well be the best approach in some cases, it is interesting to try for
an algorithm with greater determinism.
6.1.2.4 Hashed Double-Ended Queue

One of the simplest and most effective ways to deterministically partition a data structure
is to hash it. It is possible to trivially hash a double-ended queue by assigning each
element a sequence number based on its position in the list, so that the first element
left-enqueued into an empty queue is numbered zero and the first element right-enqueued
into an empty queue is numbered one. A series of elements left-enqueued into an
otherwise-idle queue would be assigned decreasing numbers (−1, −2, −3, . . .), while
a series of elements right-enqueued into an otherwise-idle queue would be assigned
increasing numbers (2, 3, 4, . . .). A key point is that it is not necessary to actually
represent a given element’s number, as this number will be implied by its position in the
queue.
Given this approach, we assign one lock to guard the left-hand index, one to guard
the right-hand index, and one lock for each hash chain. Figure 6.7 shows the resulting
data structure given four hash chains. Note that the lock domains do not overlap, and
v2023.06.11a
R1
Index L Index R
Enq 3R
R4 R1 R2 R3
Index L Index R
Enq 3L1R
R4 R5 R2 R3
L0 R1 L −2 L −1
Index L Index R
Figure 6.8: Hashed Double-Ended Queue After Insertions
v2023.06.11a
R4 R5 R6 R7
L0 R1 R2 R3
L−4 L−3 L−2 L−1
L−8 L−7 L−6 L−5
Figure 6.9: Hashed Double-Ended Queue With 16 Elements
Listing 6.1: Lock-Based Parallel Double-Ended Queue Data Structure

1 struct pdeq {
2 spinlock_t llock;
3 int lidx;
4 spinlock_t rlock;
5 int ridx;
6 struct deq bkt[PDEQ_N_BKTS];
7 };
that deadlock is avoided by acquiring the index locks before the chain locks, and by
never acquiring more than one lock of a given type (index or chain) at a time.
Each hash chain is itself a double-ended queue, and in this example, each holds
every fourth element. The uppermost portion of Figure 6.8 shows the state after a
single element (“R1 ”) has been right-enqueued, with the right-hand index having been
incremented to reference hash chain 2. The middle portion of this same figure shows
the state after three more elements have been right-enqueued. As you can see, the
indexes are back to their initial states (see Figure 6.7), however, each hash chain is
now non-empty. The lower portion of this figure shows the state after three additional
elements have been left-enqueued and an additional element has been right-enqueued.
From the last state shown in Figure 6.8, a left-dequeue operation would return element
“L−2 ” and leave the left-hand index referencing hash chain 2, which would then contain
only a single element (“R2 ”). In this state, a left-enqueue running concurrently with a
right-enqueue would result in lock contention, but the probability of such contention
can be reduced to arbitrarily low levels by using a larger hash table.
Figure 6.9 shows how 16 elements would be organized in a four-hash-bucket parallel
double-ended queue. Each underlying single-lock double-ended queue holds a one-
quarter slice of the full parallel double-ended queue.
Listing 6.1 shows the corresponding C-language data structure, assuming an existing
struct deq that provides a trivially locked double-ended-queue implementation. This
data structure contains the left-hand lock on line 2, the left-hand index on line 3,
the right-hand lock on line 4 (which is cache-aligned in the actual implementation),
the right-hand index on line 5, and, finally, the hashed array of simple lock-based
double-ended queues on line 6. A high-performance implementation would of course
use padding or special alignment directives to avoid false sharing.
Listing 6.2 (lockhdeq.c) shows the implementation of the enqueue and dequeue
functions.4 Discussion will focus on the left-hand operations, as the right-hand
operations are trivially derived from them.
Lines 1–13 show pdeq_pop_l(), which left-dequeues and returns an element if
possible, returning NULL otherwise. Line 6 acquires the left-hand spinlock, and line 7
computes the index to be dequeued from. Line 8 dequeues the element, and, if line 9
4 One could easily create a polymorphic implementation in any number of languages,
but doing so is left as an exercise for the reader.
v2023.06.11a
Listing 6.2: Lock-Based Parallel Double-Ended Queue Implementation

1 struct cds_list_head *pdeq_pop_l(struct pdeq *d)
2 {
3 struct cds_list_head *e;
4 int i;
5
6 spin_lock(&d->llock);
7 i = moveright(d->lidx);
8 e = deq_pop_l(&d->bkt[i]);
9 if (e != NULL)
10 d->lidx = i;
11 spin_unlock(&d->llock);
12 return e;
13 }
14
15 struct cds_list_head *pdeq_pop_r(struct pdeq *d)
16 {
18 int i;
19
20 spin_lock(&d->rlock);
21 i = moveleft(d->ridx);
22 e = deq_pop_r(&d->bkt[i]);
23 if (e != NULL)
24 d->ridx = i;
25 spin_unlock(&d->rlock);
26 return e;
27 }
28
29 void pdeq_push_l(struct cds_list_head *e, struct pdeq *d)
30 {
31 int i;
32
34 i = d->lidx;
35 deq_push_l(e, &d->bkt[i]);
36 d->lidx = moveleft(d->lidx);
38 }
39
40 void pdeq_push_r(struct cds_list_head *e, struct pdeq *d)
41 {
42 int i;
43
45 i = d->ridx;
46 deq_push_r(e, &d->bkt[i]);
47 d->ridx = moveright(d->ridx);
49 }
v2023.06.11a
finds the result to be non-NULL, line 10 records the new left-hand index. Either way,
line 11 releases the lock, and, finally, line 12 returns the element if there was one, or
NULL otherwise.
Lines 29–38 show pdeq_push_l(), which left-enqueues the specified element.
Line 33 acquires the left-hand lock, and line 34 picks up the left-hand index. Line 35
left-enqueues the specified element onto the double-ended queue indexed by the left-hand
index. Line 36 then updates the left-hand index and line 37 releases the lock.
As noted earlier, the right-hand operations are completely analogous to their left-
handed counterparts, so their analysis is left as an exercise for the reader.
Quick Quiz 6.5: Is the hashed double-ended queue a good solution? Why or why not?
6.1.2.5 Compound Double-Ended Queue Revisited

This section revisits the compound double-ended queue, using a trivial rebalancing
scheme that moves all the elements from the non-empty queue to the now-empty queue.
Quick Quiz 6.6: Move all the elements to the queue that became empty? In what possible
universe is this brain-dead solution in any way optimal???
In contrast to the hashed implementation presented in the previous section, the

compound implementation will build on a sequential implementation of a double-ended
queue that uses neither locks nor atomic operations.
Listing 6.3 shows the implementation. Unlike the hashed implementation, this
compound implementation is asymmetric, so that we must consider the pdeq_pop_l()
and pdeq_pop_r() implementations separately.
Quick Quiz 6.7: Why can’t the compound parallel double-ended queue implementation be
symmetric?
The pdeq_pop_l() implementation is shown on lines 1–16 of the figure. Line 5

acquires the left-hand lock, which line 14 releases. Line 6 attempts to left-dequeue an
element from the left-hand underlying double-ended queue, and, if successful, skips
lines 8–13 to simply return this element. Otherwise, line 8 acquires the right-hand
lock, line 9 left-dequeues an element from the right-hand queue, and line 10 moves any
remaining elements on the right-hand queue to the left-hand queue, line 11 initializes
the right-hand queue, and line 12 releases the right-hand lock. The element, if any, that
was dequeued on line 9 will be returned.
The pdeq_pop_r() implementation is shown on lines 18–38 of the figure. As
before, line 22 acquires the right-hand lock (and line 36 releases it), and line 23
attempts to right-dequeue an element from the right-hand queue, and, if successful,
skips lines 25–35 to simply return this element. However, if line 24 determines that there
was no element to dequeue, line 25 releases the right-hand lock and lines 26–27 acquire
both locks in the proper order. Line 28 then attempts to right-dequeue an element from
the right-hand list again, and if line 29 determines that this second attempt has failed,
line 30 right-dequeues an element from the left-hand queue (if there is one available),
line 31 moves any remaining elements from the left-hand queue to the right-hand queue,
and line 32 initializes the left-hand queue. Either way, line 34 releases the left-hand
lock.
Quick Quiz 6.8: Why is it necessary to retry the right-dequeue operation on line 28 of
Listing 6.3?
v2023.06.11a
Listing 6.3: Compound Parallel Double-Ended Queue Implementation

1 struct cds_list_head *pdeq_pop_l(struct pdeq *d)
2 {
4
6 e = deq_pop_l(&d->ldeq);
7 if (e == NULL) {
9 e = deq_pop_l(&d->rdeq);
10 cds_list_splice(&d->rdeq.chain, &d->ldeq.chain);
11 CDS_INIT_LIST_HEAD(&d->rdeq.chain);
13 }
15 return e;
16 }
17
18 struct cds_list_head *pdeq_pop_r(struct pdeq *d)
19 {
21
23 e = deq_pop_r(&d->rdeq);
24 if (e == NULL) {
28 e = deq_pop_r(&d->rdeq);
29 if (e == NULL) {
30 e = deq_pop_r(&d->ldeq);
31 cds_list_splice(&d->ldeq.chain, &d->rdeq.chain);
32 CDS_INIT_LIST_HEAD(&d->ldeq.chain);
33 }
35 }
37 return e;
38 }
39
40 void pdeq_push_l(struct cds_list_head *e, struct pdeq *d)
41 {
43 deq_push_l(e, &d->ldeq);
45 }
46
47 void pdeq_push_r(struct cds_list_head *e, struct pdeq *d)
48 {
50 deq_push_r(e, &d->rdeq);
52 }
v2023.06.11a
Quick Quiz 6.9: Surely the left-hand lock must sometimes be available!!! So why is it
necessary that line 25 of Listing 6.3 unconditionally release the right-hand lock?
The pdeq_push_l() implementation is shown on lines 40–45 of Listing 6.3. Line 42

acquires the left-hand spinlock, line 43 left-enqueues the element onto the left-hand
queue, and finally line 44 releases the lock. The pdeq_push_r() implementation
(shown on lines 47–52) is quite similar.
Quick Quiz 6.10: But in the case where data is flowing in only one direction, the algorithm
shown in Listing 6.3 will have both ends attempting to acquire the same lock whenever the
consuming end empties its underlying double-ended queue. Doesn’t that mean that sometimes
this algorithm fails to provide concurrent access to both ends of the queue even when the queue
contains an arbitrarily large number of elements?
6.1.2.6 Double-Ended Queue Discussion

The compound implementation is somewhat more complex than the hashed variant
presented in Section 6.1.2.4, but is still reasonably simple. Of course, a more intelligent
rebalancing scheme could be arbitrarily complex, but the simple scheme shown here
has been shown to perform well compared to software alternatives [DCW+ 11] and even
compared to algorithms using hardware assist [DLM+ 10]. Nevertheless, the best we
can hope for from such a scheme is 2x scalability, as at most two threads can be holding
the dequeue’s locks concurrently. This limitation also applies to algorithms based on
non-blocking synchronization, such as the compare-and-swap-based dequeue algorithm
of Michael [Mic03].5
Quick Quiz 6.11: Why are there not one but two solutions to the double-ended queue problem?
In fact, as noted by Dice et al. [DLM+ 10], an unsynchronized single-threaded double-

ended queue significantly outperforms any of the parallel implementations they studied.
Therefore, the key point is that there can be significant overhead enqueuing to or
dequeuing from a shared queue, regardless of implementation. This should come as no
surprise in light of the material in Chapter 3, given the strict first-in-first-out (FIFO)
nature of these queues.
Furthermore, these strict FIFO queues are strictly FIFO only with respect to lin-
earization points [HW90]6 that are not visible to the caller, in fact, in these examples,
the linearization points are buried in the lock-based critical sections. These queues
are not strictly FIFO with respect to (say) the times at which the individual operations
started [HKLP12]. This indicates that the strict FIFO property is not all that valuable in
concurrent programs, and in fact, Kirsch et al. present less-strict queues that provide
improved performance and scalability [KLP12].7 All that said, if you are pushing all
5 This paper is interesting in that it showed that special double-compare-and-swap
(DCAS) instructions are not needed for lock-free implementations of double-ended queues.
Instead, the common compare-and-swap (e.g., x86 cmpxchg) suffices.
6 In short, a linearization point is a single point within a given function where that
function can be said to have taken effect. In this lock-based implementation, the linearization
points can be said to be anywhere within the critical section that does the work.
7 Nir Shavit produced relaxed stacks for roughly the same reasons [Sha11]. This situation
leads some to believe that the linearization points are useful to theorists rather than developers,
and leads others to wonder to what extent the designers of such data structures and algorithms
were considering the needs of their users.
v2023.06.11a
the data used by your concurrent program through a single queue, you really need to
rethink your overall design.
6.1.3 Partitioning Example Discussion

The optimal solution to the dining philosophers problem given in the answer to the
Quick Quiz in Section 6.1.1 is an excellent example of “horizontal parallelism” or “data
parallelism”. The synchronization overhead in this case is nearly (or even exactly)
zero. In contrast, the double-ended queue implementations are examples of “vertical
parallelism” or “pipelining”, given that data moves from one thread to another. The
tighter coordination required for pipelining in turn requires larger units of work to obtain
a given level of efficiency.
Quick Quiz 6.12: The tandem double-ended queue runs about twice as fast as the hashed
double-ended queue, even when I increase the size of the hash table to an insanely large number.
Why is that?
Quick Quiz 6.13: Is there a significantly better way of handling concurrency for double-ended
queues?
These two examples show just how powerful partitioning can be in devising parallel
algorithms. Section 6.3.5 looks briefly at a third example, matrix multiply. However, all
three of these examples beg for more and better design criteria for parallel programs, a
topic taken up in the next section.
6.2 Design Criteria

One pound of learning requires ten pounds of
commonsense to apply it.
Persian proverb
One way to obtain the best performance and scalability is to simply hack away until
you converge on the best possible parallel program. Unfortunately, if your program is
other than microscopically tiny, the space of possible parallel programs is so huge that
convergence is not guaranteed in the lifetime of the universe. Besides, what exactly is
the “best possible parallel program”? After all, Section 2.2 called out no fewer than
three parallel-programming goals of performance, productivity, and generality, and
the best possible performance will likely come at a cost in terms of productivity and
generality. We clearly need to be able to make higher-level choices at design time in
order to arrive at an acceptably good parallel program before that program becomes
obsolete.
However, more detailed design criteria are required to actually produce a real-world
design, a task taken up in this section. This being the real world, these criteria often
conflict to a greater or lesser degree, requiring that the designer carefully balance the
resulting tradeoffs.
As such, these criteria may be thought of as the “forces” acting on the design, with
particularly good tradeoffs between these forces being called “design patterns” [Ale79,
GHJV95].
The design criteria for attaining the three parallel-programming goals are speedup,
contention, overhead, read-to-write ratio, and complexity:
v2023.06.11a
6.2. DESIGN CRITERIA 127
Speedup: As noted in Section 2.2, increased performance is the major reason to go to

all of the time and trouble required to parallelize it. Speedup is defined to be the
ratio of the time required to run a sequential version of the program to the time
required to run a parallel version.
Contention: If more CPUs are applied to a parallel program than can be kept busy by
that program, the excess CPUs are prevented from doing useful work by contention.
This may be lock contention, memory contention, or a host of other performance
killers.
Work-to-Synchronization Ratio: A uniprocessor, single-threaded, non-preemptible,
and non-interruptible8 version of a given parallel program would not need any
synchronization primitives. Therefore, any time consumed by these primitives (in-
cluding communication cache misses as well as message latency, locking primitives,
atomic instructions, and memory barriers) is overhead that does not contribute
directly to the useful work that the program is intended to accomplish. Note that
the important measure is the relationship between the synchronization overhead
and the overhead of the code in the critical section, with larger critical sections
able to tolerate greater synchronization overhead. The work-to-synchronization
ratio is related to the notion of synchronization efficiency.
Read-to-Write Ratio: A data structure that is rarely updated may often be replicated
rather than partitioned, and furthermore may be protected with asymmetric
synchronization primitives that reduce readers’ synchronization overhead at the
expense of that of writers, thereby reducing overall synchronization overhead.
Corresponding optimizations are possible for frequently updated data structures,
as discussed in Chapter 5.
Complexity: A parallel program is more complex than an equivalent sequential program
because the parallel program has a much larger state space than does the sequential
program, although large state spaces having regular structures can in some cases
be easily understood. A parallel programmer must consider synchronization
primitives, messaging, locking design, critical-section identification, and deadlock
in the context of this larger state space.
This greater complexity often translates to higher development and maintenance
costs. Therefore, budgetary constraints can limit the number and types of modi-
fications made to an existing program, since a given degree of speedup is worth
only so much time and trouble. Worse yet, added complexity can actually reduce
performance and scalability.
Therefore, beyond a certain point, there may be potential sequential optimizations
that are cheaper and more effective than parallelization. As noted in Section 2.2.1,
parallelization is but one performance optimization of many, and is furthermore an
optimization that applies most readily to CPU-based bottlenecks.
These criteria will act together to enforce a maximum speedup. The first three criteria are
deeply interrelated, so the remainder of this section analyzes these interrelationships.9
8 Either by masking interrupts or by being oblivious to them.

9 A real-world parallel system will be subject to many additional design criteria, such as
data-structure layout, memory size, memory-hierarchy latencies, bandwidth limitations, and
I/O issues.
v2023.06.11a
Note that these criteria may also appear as part of the requirements specification,
and further that they are one solution to the problem of summarizing the quality of
a concurrent algorithm from page 113. For example, speedup may act as a relative
desideratum (“the faster, the better”) or as an absolute requirement of the workload (“the
system must support at least 1,000,000 web hits per second”). Classic design pattern
languages describe relative desiderata as forces and absolute requirements as context.
An understanding of the relationships between these design criteria can be very
helpful when identifying appropriate design tradeoffs for a parallel program.
1. The less time a program spends in exclusive-lock critical sections, the greater the
potential speedup. This is a consequence of Amdahl’s Law [Amd67] because only
one CPU may execute within a given exclusive-lock critical section at a given time.
More specifically, for unbounded linear scalability, the fraction of time that the
program spends in a given exclusive critical section must decrease as the number
of CPUs increases. For example, a program will not scale to 10 CPUs unless it
spends much less than one tenth of its time in the most-restrictive exclusive-lock
critical section.
2. Contention effects consume the excess CPU and/or wallclock time when the actual
speedup is less than the number of available CPUs. The larger the gap between the
number of CPUs and the actual speedup, the less efficiently the CPUs will be used.
Similarly, the greater the desired efficiency, the smaller the achievable speedup.
3. If the available synchronization primitives have high overhead compared to the

critical sections that they guard, the best way to improve speedup is to reduce the
number of times that the primitives are invoked. This can be accomplished by
batching critical sections, using data ownership (see Chapter 8), using asymmetric
primitives (see Chapter 9), or by using a coarse-grained design such as code
locking.
4. If the critical sections have high overhead compared to the primitives guarding
them, the best way to improve speedup is to increase parallelism by moving to
reader/writer locking, data locking, asymmetric, or data ownership.
5. If the critical sections have high overhead compared to the primitives guarding them
and the data structure being guarded is read much more often than modified, the
best way to increase parallelism is to move to reader/writer locking or asymmetric
primitives.
6. Many changes that improve SMP performance, for example, reducing lock con-
tention, also improve real-time latencies [McK05c].
Quick Quiz 6.14: Don’t all these problems with critical sections mean that we should just
always use non-blocking synchronization [Her90], which don’t have critical sections?
It is worth reiterating that contention has many guises, including lock contention,
memory contention, cache overflow, thermal throttling, and much else besides. This
chapter looks primarily at lock and memory contention.
v2023.06.11a
6.3. SYNCHRONIZATION GRANULARITY 129
Sequential
Program
Partition Batch
Code
Locking
Partition Batch
Data
Locking
Own Disown
Data
Ownership
Figure 6.10: Design Patterns and Lock Granularity
6.3 Synchronization Granularity

Doing little things well is a step toward doing big
things better.
Harry F. Banks
Figure 6.10 gives a pictorial view of different levels of synchronization granularity, each
of which is described in one of the following sections. These sections focus primarily
on locking, but similar granularity issues arise with all forms of synchronization.
6.3.1 Sequential Program

If the program runs fast enough on a single processor, and has no interactions with
other processes, threads, or interrupt handlers, you should remove the synchronization
primitives and spare yourself their overhead and complexity. Some years back, there
were those who would argue that Moore’s Law would eventually force all programs
into this category. However, as can be seen in Figure 6.11, the exponential increase in
single-threaded performance halted in about 2003. Therefore, increasing performance
will increasingly require parallelism.10 Given that back in 2006 Paul typed the first
version of this sentence on a dual-core laptop, and further given that many of the
graphs added in 2020 were generated on a system with 56 hardware threads per socket,
parallelism is well and truly here. It is also important to note that Ethernet bandwidth is
continuing to grow, as shown in Figure 6.12. This growth will continue to motivate
multithreaded servers in order to handle the communications load.
Please note that this does not mean that you should code each and every program in a
multi-threaded manner. Again, if a program runs quickly enough on a single processor,
spare yourself the overhead and complexity of SMP synchronization primitives. The
10 This plot shows clock frequencies for newer CPUs theoretically capable of retiring
one or more instructions per clock, and MIPS for older CPUs requiring multiple clocks to
execute even the simplest instruction. The reason for taking this approach is that the newer
CPUs’ ability to retire multiple instructions per clock is typically limited by memory-system
performance.
v2023.06.11a
10000
CPU Clock Frequency / MIPS

1000
100
10
0.1
1975
1980
1985
1990
1995
2000
2005
2010
2015
2020
Year
Figure 6.11: MIPS/Clock-Frequency Trend for Intel CPUs
1x106
100000 Ethernet
Relative Performance
10000
1000
100 x86 CPUs
10
0.1
1970
1975
1980
1985
1990
1995
2000
2005
2010
2015
2020
Year
Figure 6.12: Ethernet Bandwidth vs. Intel x86 CPU Performance
v2023.06.11a
Listing 6.4: Sequential-Program Hash Table Search

1 struct hash_table
2 {
3 long nbuckets;
4 struct node **buckets;
5 };
6
7 typedef struct node {
8 unsigned long key;
9 struct node *next;
10 } node_t;
11
12 int hash_search(struct hash_table *h, long key)
13 {
14 struct node *cur;
15
16 cur = h->buckets[key % h->nbuckets];
17 while (cur != NULL) {
18 if (cur->key >= key) {
19 return (cur->key == key);
20 }
21 cur = cur->next;
22 }
23 return 0;
24 }
simplicity of the hash-table lookup code in Listing 6.4 underscores this point.11 A key
point is that speedups due to parallelism are normally limited to the number of CPUs.
In contrast, speedups due to sequential optimizations, for example, careful choice of
data structure, can be arbitrarily large.
Quick Quiz 6.15: What should you do to validate a hash table?
On the other hand, if you are not in this happy situation, read on!
6.3.2 Code Locking

Code locking is quite simple due to the fact that it uses only global locks.12 It is
especially easy to retrofit an existing program to use code locking in order to run it on a
multiprocessor. If the program has only a single shared resource, code locking will even
give optimal performance. However, many of the larger and more complex programs
require much of the execution to occur in critical sections, which in turn causes code
locking to sharply limits their scalability.
Therefore, you should use code locking on programs that spend only a small fraction
of their execution time in critical sections or from which only modest scaling is required.
In addition, programs that primarily use the more scalable approaches described in later
sections often use code locking to handle rare error cases or significant state transitions.
In these cases, code locking will provide a relatively simple program that is very similar
to its sequential counterpart, as can be seen in Listing 6.5. However, note that the
simple return of the comparison in hash_search() in Listing 6.4 has now become
three statements due to the need to release the lock before returning.
Note that the hash_lock acquisition and release statements on lines 19, 24, and 29
are mediating ownership of the hash table among the CPUs wishing to concurrently
11 The examples in this section are taken from Hart et al. [HMB06], adapted for clarity by
gathering related code from multiple files.

12 If your program instead has locks in data structures, or, in the case of Java, uses classes
with synchronized instances, you are instead using “data locking”, described in Section 6.3.3.
v2023.06.11a
Listing 6.5: Code-Locking Hash Table Search

1 spinlock_t hash_lock;
2
3 struct hash_table
4 {
5 long nbuckets;
7 };
8
12 } node_t;
13
15 {
17 int retval;
18
19 spin_lock(&hash_lock);
23 retval = (cur->key == key);
24 spin_unlock(&hash_lock);
25 return retval;
26 }
27 cur = cur->next;
28 }
29 spin_unlock(&hash_lock);
30 return 0;
31 }
access that hash table. Another way of looking at this is that hash_lock is partitioning
time, thus giving each requesting CPU its own partition of time during which it owns this
hash table. In addition, in a well-designed algorithm, there should be ample partitions
of time during which no CPU owns this hash table.
Quick Quiz 6.16: “Partitioning time”? Isn’t that an odd turn of phrase?
Unfortunately, code locking is particularly prone to “lock contention”, where multiple

CPUs need to acquire the lock concurrently. SMP programmers who have taken care of
groups of small children (or groups of older people who are acting like children) will
immediately recognize the danger of having only one of something, as illustrated in
Figure 6.13.
One solution to this problem, named “data locking”, is described in the next section.
6.3.3 Data Locking

Many data structures may be partitioned, with each partition of the data structure having
its own lock. Then the critical sections for each part of the data structure can execute
in parallel, although only one instance of the critical section for a given part could
be executing at a given time. You should use data locking when contention must be
reduced, and where synchronization overhead is not limiting speedups. Data locking
reduces contention by distributing the instances of the overly-large critical section across
multiple data structures, for example, maintaining per-hash-bucket critical sections in a
hash table, as shown in Listing 6.6. The increased scalability again results in a slight
increase in complexity in the form of an additional data structure, the struct bucket.
In contrast with the contentious situation shown in Figure 6.13, data locking helps
promote harmony, as illustrated by Figure 6.14—and in parallel programs, this almost
v2023.06.11a
toy
Figure 6.13: Lock Contention
Listing 6.6: Data-Locking Hash Table Search

1 struct hash_table
2 {
3 long nbuckets;
4 struct bucket **buckets;
5 };
6
7 struct bucket {
8 spinlock_t bucket_lock;
9 node_t *list_head;
10 };
11
15 } node_t;
16
18 {
19 struct bucket *bp;
21 int retval;
22
23 bp = h->buckets[key % h->nbuckets];
24 spin_lock(&bp->bucket_lock);
25 cur = bp->list_head;
29 spin_unlock(&bp->bucket_lock);
30 return retval;
31 }
32 cur = cur->next;
33 }
35 return 0;
36 }
v2023.06.11a
toy
yot
toy
Figure 6.14: Data Locking
always translates into increased performance and scalability. For this reason, data
locking was heavily used by Sequent in its kernels [BK85, Inm85, Gar90, Dov90, MD92,
MG92, MS93].
Another way of looking at this is to think of each ->bucket_lock as mediating
ownership not of the entire hash table as was done for code locking, but only for the
bucket corresponding to that ->bucket_lock. Each lock still partitions time, but
the per-bucket-locking technique also partitions the address space, so that the overall
technique can be said to partition spacetime. If the number of buckets is large enough,
this partitioning of space should with high probability permit a given CPU immediate
access to a given hash bucket.
However, as those who have taken care of small children can again attest, even
providing enough to go around is no guarantee of tranquillity. The analogous situation
can arise in SMP programs. For example, the Linux kernel maintains a cache of files
and directories (called “dcache”). Each entry in this cache has its own lock, but the
entries corresponding to the root directory and its direct descendants are much more
likely to be traversed than are more obscure entries. This can result in many CPUs
contending for the locks of these popular entries, resulting in a situation not unlike that
shown in Figure 6.15.
In many cases, algorithms can be designed to reduce the instance of data skew, and in
some cases eliminate it entirely (for example, in the Linux kernel’s dcache [MSS04,
Cor10a, Bro15a, Bro15b, Bro15c]). Data locking is often used for partitionable data
structures such as hash tables, as well as in situations where multiple entities are each
represented by an instance of a given data structure. The Linux-kernel task list is an
example of the latter, each task structure having its own alloc_lock and pi_lock.
A key challenge with data locking on dynamically allocated structures is ensuring
that the structure remains in existence while the lock is being acquired [GKAS99]. The
code in Listing 6.6 finesses this challenge by placing the locks in the statically allocated
hash buckets, which are never freed. However, this trick would not work if the hash
table were resizeable, so that the locks were now dynamically allocated. In this case,
v2023.06.11a
toy
toy
toy
toy
Figure 6.15: Data Locking and Skew
there would need to be some means to prevent the hash bucket from being freed during
the time that its lock was being acquired.
Quick Quiz 6.17: What are some ways of preventing a structure from being freed while its
lock is being acquired?
6.3.4 Data Ownership

Data ownership partitions a given data structure over the threads or CPUs, so that
each thread/CPU accesses its subset of the data structure without any synchronization
overhead whatsoever. However, if one thread wishes to access some other thread’s data,
the first thread is unable to do so directly. Instead, the first thread must communicate
with the second thread, so that the second thread performs the operation on behalf of
the first, or, alternatively, migrates the data to the first thread.
Data ownership might seem arcane, but it is used very frequently:
1. Any variables accessible by only one CPU or thread (such as auto variables in C
and C++) are owned by that CPU or process.
2. An instance of a user interface owns the corresponding user’s context. It is very
common for applications interacting with parallel database engines to be written as
if they were entirely sequential programs. Such applications own the user interface
and his current action. Explicit parallelism is thus confined to the database engine
itself.
3. Parametric simulations are often trivially parallelized by granting each thread
ownership of a particular region of the parameter space. There are also computing
frameworks designed for this type of problem [Uni08a].
If there is significant sharing, communication between the threads or CPUs can result
in significant complexity and overhead. Furthermore, if the most-heavily used data
happens to be that owned by a single CPU, that CPU will be a “hot spot”, sometimes
with results resembling that shown in Figure 6.15. However, in situations where no
sharing is required, data ownership achieves ideal performance, and with code that
v2023.06.11a
can be as simple as the sequential-program case shown in Listing 6.4. Such situations
are often referred to as “embarrassingly parallel”, and, in the best case, resemble the
situation previously shown in Figure 6.14.
Another important instance of data ownership occurs when the data is read-only, in
which case, all threads can “own” it via replication.
Where data locking partitions both the address space (with one hash buckets per
partition) and time (using per-bucket locks), data ownership partitions only the address
space. The reason that data ownership need not partition time is because a given thread
or CPU is assigned permanent ownership of a given address-space partition.
Quick Quiz 6.18: But won’t system boot and shutdown (or application startup and shutdown)
be partitioning time, even for data ownership?
Data ownership will be presented in more detail in Chapter 8.
6.3.5 Locking Granularity and Performance

This section looks at locking granularity and performance from a mathematical syn-
chronization-efficiency viewpoint. Readers who are uninspired by mathematics might
choose to skip this section.
The approach is to use a crude queueing model for the efficiency of synchronization
mechanism that operate on a single shared global variable, based on an M/M/1 queue.
M/M/1 queuing models are based on an exponentially distributed “inter-arrival rate”
𝜆 and an exponentially distributed “service rate” 𝜇. The inter-arrival rate 𝜆 can be
thought of as the average number of synchronization operations per second that the
system would process if the synchronization were free, in other words, 𝜆 is an inverse
measure of the overhead of each non-synchronization unit of work. For example, if each
unit of work was a transaction, and if each transaction took one millisecond to process,
excluding synchronization overhead, then 𝜆 would be 1,000 transactions per second.
The service rate 𝜇 is defined similarly, but for the average number of synchronization
operations per second that the system would process if the overhead of each transaction
was zero, and ignoring the fact that CPUs must wait on each other to complete
their synchronization operations, in other words, 𝜇 can be roughly thought of as
the synchronization overhead in absence of contention. For example, suppose that
each transaction’s synchronization operation involves an atomic increment instruction,
and that a computer system is able to do a private-variable atomic increment every
5 nanoseconds on each CPU (see Figure 5.1).13 The value of 𝜇 is therefore about
200,000,000 atomic increments per second.
Of course, the value of 𝜆 increases as increasing numbers of CPUs increment a shared
variable because each CPU is capable of processing transactions independently (again,
ignoring synchronization):
𝜆 = 𝑛𝜆0 (6.1)
Here, 𝑛 is the number of CPUs and 𝜆0 is the transaction-processing capability of a
single CPU. Note that the expected time for a single CPU to execute a single transaction
in the absence of contention is 1/𝜆0 .
13 Of course, if there are 8 CPUs all incrementing the same shared variable, then each
CPU must wait at least 35 nanoseconds for each of the other CPUs to do its increment before
consuming an additional 5 nanoseconds doing its own increment. In fact, the wait will be
longer due to the need to move the variable from one CPU to another.
v2023.06.11a
Synchronization Efficiency
1
0.9
0.8
0.7
100
0.6
0.5 75
0.4 50
0.3 25
0.2 10
0.1
10
20
30
40
50
60
70
80
90
100
Figure 6.16: Synchronization Efficiency
Because the CPUs have to “wait in line” behind each other to get their chance to
increment the single shared variable, we can use the M/M/1 queueing-model expression
for the expected total waiting time:
1
𝑇= (6.2)
𝜇−𝜆
Substituting the above value of 𝜆:
1
𝑇= (6.3)
𝜇 − 𝑛𝜆0
Now, the efficiency is just the ratio of the time required to process a transaction
in absence of synchronization (1/𝜆0 ) to the time required including synchronization
(𝑇 + 1/𝜆0 ):
1/𝜆0
𝑒= (6.4)
𝑇 + 1/𝜆0
Substituting the above value for 𝑇 and simplifying:
𝜇
𝜆0 −𝑛
𝑒= 𝜇 (6.5)
𝜆0 − (𝑛 − 1)
But the value of 𝜇/𝜆0 is just the ratio of the time required to process the transaction
(absent synchronization overhead) to that of the synchronization overhead itself (absent
contention). If we call this ratio 𝑓 , we have:
𝑓 −𝑛
𝑒= (6.6)
𝑓 − (𝑛 − 1)
Figure 6.16 plots the synchronization efficiency 𝑒 as a function of the number of
CPUs/threads 𝑛 for a few values of the overhead ratio 𝑓 . For example, again using the
5-nanosecond atomic increment, the 𝑓 = 10 line corresponds to each CPU attempting
an atomic increment every 50 nanoseconds, and the 𝑓 = 100 line corresponds to
each CPU attempting an atomic increment every 500 nanoseconds, which in turn
v2023.06.11a
1.2
Matrix Multiply Efficiency

1
1024
0.8 256
0.6
128 512
0.4
0.2
64
0
1 10 100
Figure 6.17: Matrix Multiply Efficiency
corresponds to some hundreds (perhaps thousands) of instructions. Given that each trace
drops off sharply with increasing numbers of CPUs or threads, we can conclude that
synchronization mechanisms based on atomic manipulation of a single global shared
variable will not scale well if used heavily on current commodity hardware. This is an
abstract mathematical depiction of the forces leading to the parallel counting algorithms
that were discussed in Chapter 5. Your real-world mileage may differ.
Nevertheless, the concept of efficiency is useful, and even in cases having little or no
formal synchronization. Consider for example a matrix multiply, in which the columns
of one matrix are multiplied (via “dot product”) by the rows of another, resulting in
an entry in a third matrix. Because none of these operations conflict, it is possible to
partition the columns of the first matrix among a group of threads, with each thread
computing the corresponding columns of the result matrix. The threads can therefore
operate entirely independently, with no synchronization overhead whatsoever, as is done
in matmul.c. One might therefore expect a perfect efficiency of 1.0.
However, Figure 6.17 tells a different story, especially for a 64-by-64 matrix multiply,
which never gets above an efficiency of about 0.3, even when running single-threaded,
and drops sharply as more threads are added.14 The 128-by-128 matrix does better,
but still fails to demonstrate much performance increase with added threads. The
256-by-256 matrix does scale reasonably well, but only up to a handful of CPUs. The
512-by-512 matrix multiply’s efficiency is measurably less than 1.0 on as few as 10
threads, and even the 1024-by-1024 matrix multiply deviates noticeably from perfection
at a few tens of threads. Nevertheless, this figure clearly demonstrates the performance
and scalability benefits of batching: If you must incur synchronization overhead, you
may as well get your money’s worth, which is the solution to the problem of deciding
on granularity of synchronization put forth on page 113.
Quick Quiz 6.19: How can a single-threaded 64-by-64 matrix multiple possibly have an
efficiency of less than 1.0? Shouldn’t all of the traces in Figure 6.17 have efficiency of exactly
1.0 when running on one thread?
14 In contrast to the smooth traces of Figure 6.16, the wide error bars and jagged traces of
Figure 6.17 gives evidence of its real-world nature.
v2023.06.11a
6.4. PARALLEL FASTPATH 139
Reader/Writer
Locking
RCU
Parallel
Fastpath
Hierarchical
Locking
Allocator
Caches
Figure 6.18: Parallel-Fastpath Design Patterns
Given these inefficiencies, it is worthwhile to look into more-scalable approaches

such as the data locking described in Section 6.3.3 or the parallel-fastpath approach
discussed in the next section.
Quick Quiz 6.20: How are data-parallel techniques going to help with matrix multiply? It is
already data parallel!!!
Quick Quiz 6.21: What did you do to validate this matrix multiply algorithm?
6.4 Parallel Fastpath
There are two ways of meeting difficulties: You alter

the difficulties, or you alter yourself to meet them.
Phyllis Bottome
Fine-grained (and therefore usually higher-performance) designs are typically more

complex than are coarser-grained designs. In many cases, most of the overhead is
incurred by a small fraction of the code [Knu73]. So why not focus effort on that small
fraction?
This is the idea behind the parallel-fastpath design pattern, to aggressively parallelize
the common-case code path without incurring the complexity that would be required to
aggressively parallelize the entire algorithm. You must understand not only the specific
algorithm you wish to parallelize, but also the workload that the algorithm will be
subjected to. Great creativity and design effort is often required to construct a parallel
fastpath.
Parallel fastpath combines different patterns (one for the fastpath, one elsewhere) and
is therefore a template pattern. The following instances of parallel fastpath occur often
enough to warrant their own patterns, as depicted in Figure 6.18:
1. Reader/Writer Locking (described below in Section 6.4.1).
v2023.06.11a
Listing 6.7: Reader-Writer-Locking Hash Table Search

1 rwlock_t hash_lock;
2
3 struct hash_table
4 {
5 long nbuckets;
7 };
8
12 } node_t;
13
15 {
17 int retval;
18
19 read_lock(&hash_lock);
24 read_unlock(&hash_lock);
25 return retval;
26 }
27 cur = cur->next;
28 }
29 read_unlock(&hash_lock);
30 return 0;
31 }
2. Read-copy update (RCU), which may be used as a high-performance replacement

for reader/writer locking, is introduced in Section 9.5. Other alternatives include
hazard pointers (Section 9.3) and sequence locking (Section 9.4). These alternatives
will not be discussed further in this chapter.
3. Hierarchical Locking ([McK96a]), which is touched upon in Section 6.4.2.
4. Resource Allocator Caches ([McK96a, MS93]). See Section 6.4.3 for more detail.
6.4.1 Reader/Writer Locking

If synchronization overhead is negligible (for example, if the program uses coarse-
grained parallelism with large critical sections), and if only a small fraction of the
critical sections modify data, then allowing multiple readers to proceed in parallel can
greatly increase scalability. Writers exclude both readers and each other. There are
many implementations of reader-writer locking, including the POSIX implementation
described in Section 4.2.4. Listing 6.7 shows how the hash search might be implemented
using reader-writer locking.
Reader/writer locking is a simple instance of asymmetric locking. Snaman [ST87]
describes a more ornate six-mode asymmetric locking design used in several clustered
systems. Locking in general and reader-writer locking in particular is described
extensively in Chapter 7.
6.4.2 Hierarchical Locking

The idea behind hierarchical locking is to have a coarse-grained lock that is held only
long enough to work out which fine-grained lock to acquire. Listing 6.8 shows how
v2023.06.11a
Listing 6.8: Hierarchical-Locking Hash Table Search

1 struct hash_table
2 {
3 long nbuckets;
4 struct bucket **buckets;
5 };
6
7 struct bucket {
8 spinlock_t bucket_lock;
9 node_t *list_head;
10 };
11
13 spinlock_t node_lock;
16 } node_t;
17
19 {
20 struct bucket *bp;
22 int retval;
23
24 bp = h->buckets[key % h->nbuckets];
25 spin_lock(&bp->bucket_lock);
26 cur = bp->list_head;
29 spin_lock(&cur->node_lock);
32 spin_unlock(&cur->node_lock);
33 return retval;
34 }
35 cur = cur->next;
36 }
38 return 0;
39 }
our hash-table search might be adapted to do hierarchical locking, but also shows the
great weakness of this approach: We have paid the overhead of acquiring a second lock,
but we only hold it for a short time. In this case, the data-locking approach would be
simpler and likely perform better.
Quick Quiz 6.22: In what situation would hierarchical locking work well?
6.4.3 Resource Allocator Caches

This section presents a simplified schematic of a parallel fixed-block-size memory
allocator. More detailed descriptions may be found in the literature [MG92, MS93,
BA01, MSK01, Eva11, Ken20] or in the Linux kernel [Tor03].
6.4.3.1 Parallel Resource Allocation Problem

The basic problem facing a parallel memory allocator is the tension between the need to
provide extremely fast memory allocation and freeing in the common case and the need
to efficiently distribute memory in face of unfavorable allocation and freeing patterns.
To see this tension, consider a straightforward application of data ownership to this
problem—simply carve up memory so that each CPU owns its share. For example,
suppose that a system with 12 CPUs has 64 gigabytes of memory, for example, the
v2023.06.11a
Global Pool
Overflow
Overflow
(Code Locked)
Empty
Empty
CPU 0 Pool CPU 1 Pool
(Owned by CPU 0) (Owned by CPU 1)
Allocate/Free
Figure 6.19: Allocator Cache Schematic
laptop I am using right now. We could simply assign each CPU a five-gigabyte region
of memory, and allow each CPU to allocate from its own region, without the need for
locking and its complexities and overheads. Unfortunately, this scheme fails when CPU 0
only allocates memory and CPU 1 only frees it, as happens in simple producer-consumer
workloads.
The other extreme, code locking, suffers from excessive lock contention and over-
head [MS93].
6.4.3.2 Parallel Fastpath for Resource Allocation
The commonly used solution uses parallel fastpath with each CPU owning a modest
cache of blocks, and with a large code-locked shared pool for additional blocks. To
prevent any given CPU from monopolizing the memory blocks, we place a limit on the
number of blocks that can be in each CPU’s cache. In a two-CPU system, the flow of
memory blocks will be as shown in Figure 6.19: When a given CPU is trying to free a
block when its pool is full, it sends blocks to the global pool, and, similarly, when that
CPU is trying to allocate a block when its pool is empty, it retrieves blocks from the
global pool.
6.4.3.3 Data Structures
The actual data structures for a “toy” implementation of allocator caches are shown
in Listing 6.9 (“smpalloc.c”). The “Global Pool” of Figure 6.19 is implemented
by globalmem of type struct globalmempool, and the two CPU pools by the per-
thread variable perthreadmem of type struct perthreadmempool. Both of these
data structures have arrays of pointers to blocks in their pool fields, which are filled
from index zero upwards. Thus, if globalmem.pool[3] is NULL, then the remainder
of the array from index 4 up must also be NULL. The cur fields contain the index of the
highest-numbered full element of the pool array, or −1 if all elements are empty. All
v2023.06.11a
Listing 6.9: Allocator-Cache Data Structures

1 #define TARGET_POOL_SIZE 3
2 #define GLOBAL_POOL_SIZE 40
3
4 struct globalmempool {
5 spinlock_t mutex;
6 int cur;
7 struct memblock *pool[GLOBAL_POOL_SIZE];
8 } globalmem;
9
10 struct perthreadmempool {
11 int cur;
12 struct memblock *pool[2 * TARGET_POOL_SIZE];
13 };
14
15 DEFINE_PER_THREAD(struct perthreadmempool, perthreadmem);
(Empty) −1
Figure 6.20: Allocator Pool Schematic
elements from globalmem.pool[0] through globalmem.pool[globalmem.cur]

must be full, and all the rest must be empty.15
The operation of the pool data structures is illustrated by Figure 6.20, with the six boxes
representing the array of pointers making up the pool field, and the number preceding
them representing the cur field. The shaded boxes represent non-NULL pointers, while
the empty boxes represent NULL pointers. An important, though potentially confusing,
invariant of this data structure is that the cur field is always one smaller than the number
of non-NULL pointers.
6.4.3.4 Allocation Function

The allocation function memblock_alloc() may be seen in Listing 6.10. Line 7 picks
up the current thread’s per-thread pool, and line 8 checks to see if it is empty.
If so, lines 9–16 attempt to refill it from the global pool under the spinlock acquired
on line 9 and released on line 16. Lines 10–14 move blocks from the global to the
per-thread pool until either the local pool reaches its target size (half full) or the global
pool is exhausted, and line 15 sets the per-thread pool’s count to the proper value.
15 Both pool sizes (TARGET_POOL_SIZE and GLOBAL_POOL_SIZE) are unrealistically
small, but this small size makes it easier to single-step the program in order to get a feel for
its operation.
v2023.06.11a
Listing 6.10: Allocator-Cache Allocator Function

1 struct memblock *memblock_alloc(void)
2 {
3 int i;
4 struct memblock *p;
5 struct perthreadmempool *pcpp;
6
7 pcpp = &__get_thread_var(perthreadmem);
8 if (pcpp->cur < 0) {
9 spin_lock(&globalmem.mutex);
10 for (i = 0; i < TARGET_POOL_SIZE &&
11 globalmem.cur >= 0; i++) {
12 pcpp->pool[i] = globalmem.pool[globalmem.cur];
13 globalmem.pool[globalmem.cur--] = NULL;
14 }
15 pcpp->cur = i - 1;
16 spin_unlock(&globalmem.mutex);
17 }
18 if (pcpp->cur >= 0) {
19 p = pcpp->pool[pcpp->cur];
20 pcpp->pool[pcpp->cur--] = NULL;
21 return p;
22 }
23 return NULL;
24 }
Listing 6.11: Allocator-Cache Free Function

1 void memblock_free(struct memblock *p)
2 {
3 int i;
4 struct perthreadmempool *pcpp;
5
6 pcpp = &__get_thread_var(perthreadmem);
7 if (pcpp->cur >= 2 * TARGET_POOL_SIZE - 1) {
8 spin_lock(&globalmem.mutex);
9 for (i = pcpp->cur; i >= TARGET_POOL_SIZE; i--) {
10 globalmem.pool[++globalmem.cur] = pcpp->pool[i];
11 pcpp->pool[i] = NULL;
12 }
13 pcpp->cur = i;
14 spin_unlock(&globalmem.mutex);
15 }
16 pcpp->pool[++pcpp->cur] = p;
17 }
In either case, line 18 checks for the per-thread pool still being empty, and if not,
lines 19–21 remove a block and return it. Otherwise, line 23 tells the sad tale of memory
exhaustion.
6.4.3.5 Free Function
Listing 6.11 shows the memory-block free function. Line 6 gets a pointer to this thread’s
pool, and line 7 checks to see if this per-thread pool is full.
If so, lines 8–15 empty half of the per-thread pool into the global pool, with lines 8
and 14 acquiring and releasing the spinlock. Lines 9–12 implement the loop moving
blocks from the local to the global pool, and line 13 sets the per-thread pool’s count to
the proper value.
In either case, line 16 then places the newly freed block into the per-thread pool.
Quick Quiz 6.23: Doesn’t this resource-allocator design resemble that of the approximate
limit counters covered in Section 5.3?
v2023.06.11a
30
Allocations/Frees Per Microsecond

25
20
15
10
0
0 5 10 15 20 25
Allocation Run Length
Figure 6.21: Allocator Cache Performance
6.4.3.6 Performance
Rough performance results16 are shown in Figure 6.21, running on a dual-core Intel
x86 running at 1 GHz (4300 bogomips per CPU) with at most six blocks allowed in
each CPU’s cache. In this micro-benchmark, each thread repeatedly allocates a group
of blocks and then frees all the blocks in that group, with the number of blocks in the
group being the “allocation run length” displayed on the x-axis. The y-axis shows the
number of successful allocation/free pairs per microsecond—failed allocations are not
counted. The “X”s are from a two-thread run, while the “+”s are from a single-threaded
run.
Note that run lengths up to six scale linearly and give excellent performance, while run
lengths greater than six show poor performance and almost always also show negative
scaling. It is therefore quite important to size TARGET_POOL_SIZE sufficiently large,
which fortunately is usually quite easy to do in actual practice [MSK01], especially
given today’s large memories. For example, in most systems, it is quite reasonable to
set TARGET_POOL_SIZE to 100, in which case allocations and frees are guaranteed to
be confined to per-thread pools at least 99 % of the time.
As can be seen from the figure, the situations where the common-case data-ownership
applies (run lengths up to six) provide greatly improved performance compared to the
cases where locks must be acquired. Avoiding synchronization in the common case will
be a recurring theme through this book.
Quick Quiz 6.24: In Figure 6.21, there is a pattern of performance rising with increasing run
length in groups of three samples, for example, for run lengths 10, 11, and 12. Why?
Quick Quiz 6.25: Allocation failures were observed in the two-thread tests at run lengths of
19 and greater. Given the global-pool size of 40 and the per-thread target pool size 𝑠 of three,
number of threads 𝑛 equal to two, and assuming that the per-thread pools are initially empty
with none of the memory in use, what is the smallest allocation run length 𝑚 at which failures
16 This data was not collected in a statistically meaningful way, and therefore should be
viewed with great skepticism and suspicion. Good data-collection and -reduction practice is
discussed in Chapter 11. That said, repeated runs gave similar results, and these results match
more careful evaluations of similar algorithms.
v2023.06.11a
can occur? (Recall that each thread repeatedly allocates 𝑚 block of memory, and then frees the
𝑚 blocks of memory.) Alternatively, given 𝑛 threads each with pool size 𝑠, and where each
thread repeatedly first allocates 𝑚 blocks of memory and then frees those 𝑚 blocks, how large
must the global pool size be? Note: Obtaining the correct answer will require you to examine
the smpalloc.c source code, and very likely single-step it as well. You have been warned!
6.4.3.7 Validation
Validation of this simple allocator spawns a specified number of threads, with each
thread repeatedly allocating a specified number of memory blocks and then deallocating
them. This simple regimen suffices to exercise both the per-thread caches and the global
pool, as can be seen in Figure 6.21.
Much more aggressive validation is required for memory allocators that are to be used
in production. The test suites for tcmalloc [Ken20] and jemalloc [Eva11] are instructive,
as are the tests for the Linux kernel’s memory allocator.
6.4.3.8 Real-World Design
The toy parallel resource allocator was quite simple, but real-world designs expand on
this approach in a number of ways.
First, real-world allocators are required to handle a wide range of allocation sizes, as
opposed to the single size shown in this toy example. One popular way to do this is to
offer a fixed set of sizes, spaced so as to balance external and internal fragmentation,
such as in the late-1980s BSD memory allocator [MK88]. Doing this would mean that
the “globalmem” variable would need to be replicated on a per-size basis, and that the
associated lock would similarly be replicated, resulting in data locking rather than the
toy program’s code locking.
Second, production-quality systems must be able to repurpose memory, meaning that
they must be able to coalesce blocks into larger structures, such as pages [MS93]. This
coalescing will also need to be protected by a lock, which again could be replicated on a
per-size basis.
Third, coalesced memory must be returned to the underlying memory system, and
pages of memory must also be allocated from the underlying memory system. The
locking required at this level will depend on that of the underlying memory system, but
could well be code locking. Code locking can often be tolerated at this level, because
this level is so infrequently reached in well-designed systems [MSK01].
Concurrent userspace allocators face similar challenges [Ken20, Eva11].
Despite this real-world design’s greater complexity, the underlying idea is the same—
repeated application of parallel fastpath, as shown in Table 6.1.
And “parallel fastpath” is one of the solutions to the non-partitionable application
problem put forth on page 113.
v2023.06.11a
6.5. BEYOND PARTITIONING 147
Table 6.1: Schematic of Real-World Parallel Allocator
Level Locking Purpose
Per-thread pool Data ownership High-speed

allocation
Global block pool Data locking Distributing blocks
among threads
Coalescing Data locking Combining blocks
into pages
System memory Code locking Memory from/to
system
6.5 Beyond Partitioning
It is all right to aim high if you have plenty of

ammunition.
Hawley R. Everhart
This chapter has discussed how data partitioning can be used to design simple linearly
scalable parallel programs. Section 6.3.4 hinted at the possibilities of data replication,
which will be used to great effect in Section 9.5.
The main goal of applying partitioning and replication is to achieve linear speedups,
in other words, to ensure that the total amount of work required does not increase
significantly as the number of CPUs or threads increases. A problem that can be
solved via partitioning and/or replication, resulting in linear speedups, is embarrassingly
parallel. But can we do better?
To answer this question, let us examine the solution of labyrinths and mazes. Of
course, labyrinths and mazes have been objects of fascination for millennia [Wik12],
so it should come as no surprise that they are generated and solved using computers,
including biological computers [Ada11], GPGPUs [Eri08], and even discrete hard-
ware [KFC11]. Parallel solution of mazes is sometimes used as a class project in
universities [ETH11, Uni10] and as a vehicle to demonstrate the benefits of parallel-
programming frameworks [Fos10].
Common advice is to use a parallel work-queue algorithm (PWQ) [ETH11, Fos10].
This section evaluates this advice by comparing PWQ against a sequential algorithm
(SEQ) and also against an alternative parallel algorithm, in all cases solving ran-
domly generated square mazes. Section 6.5.1 discusses PWQ, Section 6.5.2 discusses
an alternative parallel algorithm, Section 6.5.4 analyzes its anomalous performance,
Section 6.5.5 derives an improved sequential algorithm from the alternative parallel algo-
rithm, Section 6.5.6 makes further performance comparisons, and finally Section 6.5.7
presents future directions and concluding remarks.
6.5.1 Work-Queue Parallel Maze Solver

PWQ is based on SEQ, which is shown in Listing 6.12 (pseudocode for maze_seq.c).
The maze is represented by a 2D array of cells and a linear-array-based work queue
named ->visited.
v2023.06.11a
Listing 6.12: SEQ Pseudocode

1 int maze_solve(maze *mp, cell sc, cell ec)
2 {
3 cell c = sc;
4 cell n;
5 int vi = 0;
6
7 maze_try_visit_cell(mp, c, c, &n, 1);
8 for (;;) {
9 while (!maze_find_any_next_cell(mp, c, &n)) {
10 if (++vi >= mp->vi)
11 return 0;
12 c = mp->visited[vi].c;
13 }
14 do {
15 if (n == ec) {
16 return 1;
17 }
18 c = n;
19 } while (maze_find_any_next_cell(mp, c, &n));
20 c = mp->visited[vi].c;
21 }
22 }
1 2 3
2 3 4
3 4 5
Figure 6.22: Cell-Number Solution Tracking
Line 7 visits the initial cell, and each iteration of the loop spanning lines 8–21 traverses
passages headed by one cell. The loop spanning lines 9–13 scans the ->visited[]
array for a visited cell with an unvisited neighbor, and the loop spanning lines 14–19
traverses one fork of the submaze headed by that neighbor. Line 20 initializes for the
next pass through the outer loop.
The pseudocode for maze_try_visit_cell() is shown on lines 1–12 of Listing 6.13
(maze.c). Line 4 checks to see if cells c and t are adjacent and connected, while line 5
checks to see if cell t has not yet been visited. The celladdr() function returns the
address of the specified cell. If either check fails, line 6 returns failure. Line 7 indicates
the next cell, line 8 records this cell in the next slot of the ->visited[] array, line 9
indicates that this slot is now full, and line 10 marks this cell as visited and also records
the distance from the maze start. Line 11 then returns success.
The pseudocode for maze_find_any_next_cell() is shown on lines 14–28 of
Listing 6.13 (maze.c). Line 17 picks up the current cell’s distance plus 1, while lines 19,
21, 23, and 25 check the cell in each direction, and lines 20, 22, 24, and 26 return
true if the corresponding cell is a candidate next cell. The prevcol(), nextcol(),
prevrow(), and nextrow() each do the specified array-index-conversion operation.
If none of the cells is a candidate, line 27 returns false.
The path is recorded in the maze by counting the number of cells from the starting
point, as shown in Figure 6.22, where the starting cell is in the upper left and the
ending cell is in the lower right. Starting at the ending cell and following consecutively
decreasing cell numbers traverses the solution.
v2023.06.11a
Listing 6.13: SEQ Helper Pseudocode

1 int maze_try_visit_cell(struct maze *mp, cell c, cell t,
2 cell *n, int d)
3 {
4 if (!maze_cells_connected(mp, c, t) ||
5 (*celladdr(mp, t) & VISITED))
6 return 0;
7 *n = t;
8 mp->visited[mp->vi] = t;
9 mp->vi++;
10 *celladdr(mp, t) |= VISITED | d;
11 return 1;
12 }
13
14 int maze_find_any_next_cell(struct maze *mp, cell c,
15 cell *n)
16 {
17 int d = (*celladdr(mp, c) & DISTANCE) + 1;
18
19 if (maze_try_visit_cell(mp, c, prevcol(c), n, d))
20 return 1;
21 if (maze_try_visit_cell(mp, c, nextcol(c), n, d))
22 return 1;
23 if (maze_try_visit_cell(mp, c, prevrow(c), n, d))
24 return 1;
25 if (maze_try_visit_cell(mp, c, nextrow(c), n, d))
26 return 1;
27 return 0;
28 }
1
0.9
0.8
0.7 PWQ
Probability
0.6
0.5 SEQ
0.4
0.3
0.2
0.1
0
0 20 40 60 80 100 120 140
CDF of Solution Time (ms)
Figure 6.23: CDF of Solution Times For SEQ and PWQ
v2023.06.11a
Listing 6.14: Partitioned Parallel Solver Pseudocode

1 int maze_solve_child(maze *mp, cell *visited, cell sc)
2 {
3 cell c;
4 cell n;
5 int vi = 0;
6
7 myvisited = visited; myvi = &vi;
8 c = visited[vi];
9 do {
10 while (!maze_find_any_next_cell(mp, c, &n)) {
11 if (visited[++vi].row < 0)
12 return 0;
13 if (READ_ONCE(mp->done))
14 return 1;
15 c = visited[vi];
16 }
17 do {
18 if (READ_ONCE(mp->done))
19 return 1;
20 c = n;
21 } while (maze_find_any_next_cell(mp, c, &n));
22 c = visited[vi];
23 } while (!READ_ONCE(mp->done));
24 return 1;
25 }
The parallel work-queue solver is a straightforward parallelization of the algorithm

shown in Listings 6.12 and 6.13. Line 10 of Listing 6.12 must use fetch-and-add, and
the local variable vi must be shared among the various threads. Lines 5 and 10 of
Listing 6.13 must be combined into a CAS loop, with CAS failure indicating a loop
in the maze. Lines 8–9 of this listing must use fetch-and-add to arbitrate concurrent
attempts to record cells in the ->visited[] array.
This approach does provide significant speedups on a dual-CPU Lenovo W500
running at 2.53 GHz, as shown in Figure 6.23, which shows the cumulative distribution
functions (CDFs) for the solution times of the two algorithms, based on the solution of
500 different square 500-by-500 randomly generated mazes. The substantial overlap of
the projection of the CDFs onto the x-axis will be addressed in Section 6.5.4.
Interestingly enough, the sequential solution-path tracking works unchanged for
the parallel algorithm. However, this uncovers a significant weakness in the parallel
algorithm: At most one thread may be making progress along the solution path at any
given time. This weakness is addressed in the next section.
6.5.2 Alternative Parallel Maze Solver

Youthful maze solvers are often urged to start at both ends, and this advice has been
repeated more recently in the context of automated maze solving [Uni10]. This advice
amounts to partitioning, which has been a powerful parallelization strategy in the
context of parallel programming for both operating-system kernels [BK85, Inm85] and
applications [Pat10]. This section applies this strategy, using two child threads that start
at opposite ends of the solution path, and takes a brief look at the performance and
scalability consequences.
The partitioned parallel algorithm (PART), shown in Listing 6.14 (maze_part.c), is
similar to SEQ, but has a few important differences. First, each child thread has its own
visited array, passed in by the parent as shown on line 1, which must be initialized to
all [−1, −1]. Line 7 stores a pointer to this array into the per-thread variable myvisited
to allow access by helper functions, and similarly stores a pointer to the local visit
v2023.06.11a
Listing 6.15: Partitioned Parallel Helper Pseudocode

1 int maze_try_visit_cell(struct maze *mp, int c, int t,
2 int *n, int d)
3 {
4 cell_t t;
5 cell_t *tp;
6 int vi;
7
8 if (!maze_cells_connected(mp, c, t))
9 return 0;
10 tp = celladdr(mp, t);
11 do {
12 t = READ_ONCE(*tp);
13 if (t & VISITED) {
14 if ((t & TID) != mytid)
15 mp->done = 1;
16 return 0;
17 }
18 } while (!CAS(tp, t, t | VISITED | myid | d));
19 *n = t;
20 vi = (*myvi)++;
21 myvisited[vi] = t;
22 return 1;
23 }
index. Second, the parent visits the first cell on each child’s behalf, which the child
retrieves on line 8. Third, the maze is solved as soon as one child locates a cell that has
been visited by the other child. When maze_try_visit_cell() detects this, it sets
a ->done field in the maze structure. Fourth, each child must therefore periodically
check the ->done field, as shown on lines 13, 18, and 23. The READ_ONCE() primitive
must disable any compiler optimizations that might combine consecutive loads or that
might reload the value. A C++1x volatile relaxed load suffices [Smi19]. Finally, the
maze_find_any_next_cell() function must use compare-and-swap to mark a cell
as visited, however no constraints on ordering are required beyond those provided by
thread creation and join.
The pseudocode for maze_find_any_next_cell() is identical to that shown in
Listing 6.13, but the pseudocode for maze_try_visit_cell() differs, and is shown
in Listing 6.15. Lines 8–9 check to see if the cells are connected, returning failure if not.
The loop spanning lines 11–18 attempts to mark the new cell visited. Line 13 checks to
see if it has already been visited, in which case line 16 returns failure, but only after
line 14 checks to see if we have encountered the other thread, in which case line 15
indicates that the solution has been located. Line 19 updates to the new cell, lines 20
and 21 update this thread’s visited array, and line 22 returns success.
Performance testing revealed a surprising anomaly, shown in Figure 6.24. The median
solution time for PART (17 milliseconds) is more than four times faster than that of
SEQ (79 milliseconds), despite running on only two threads.
The first reaction to such a dramatic performance anomaly is to check for bugs, which
suggests stringent validation be applied. This is the topic of the next section.
6.5.3 Maze Validation

Much of the validation effort comprised consistency checks, which can be located by
searching for ABORT() in CodeSamples/SMPdesign/maze/*.c. Examples checks
include:
1. Maze solution steps that end up outside of the maze.
v2023.06.11a
1
0.9
0.8 PART
0.7 PWQ
Probability
0.6
0.5 SEQ
0.4
0.3
0.2
0.1
0
0 20 40 60 80 100 120 140
CDF of Solution Time (ms)
Figure 6.24: CDF of Solution Times For SEQ, PWQ, and PART
2. Mazes that suddenly have zero or fewer rows or columns.

3. Newly created mazes with unreachable cells.
4. Mazes that have no solution.
5. Discontinuous maze solutions.
6. Attempts to start the maze solver outside of the maze.
7. Mazes whose solution path is longer than the number of cells in the maze.
8. Subsolutions by different threads cross each other.
9. Memory-allocation failure.
10. System-call failure.
Additional manual validation was applied by Paul’s wife, who greatly enjoys solving
puzzles.
However, if this maze software was to be used in production, whatever that might
mean, it would be wise to construct an independent maze fsck program. Nevertheless,
the mazes and solutions all proved to be quite valid. The next section therefore more
deeply analyzes the scalability anomaly called out in Section 6.5.2.
6.5.4 Performance Comparison I

Although the algorithms were in fact finding valid solutions to valid mazes, the plot
of CDFs in Figure 6.24 assumes independent data points. This is not the case: The
performance tests randomly generate a maze, and then run all solvers on that maze. It
therefore makes sense to plot the CDF of the ratios of solution times for each generated
maze, as shown in Figure 6.25, greatly reducing the CDFs’ overlap. This plot reveals
that for some mazes, PART is more than forty times faster than SEQ. In contrast, PWQ
is never more than about two times faster than SEQ. A forty-times speedup on two
threads demands explanation. After all, this is not merely embarrassingly parallel, where
partitionability means that adding threads does not increase the overall computational
cost. It is instead humiliatingly parallel: Adding threads significantly reduces the overall
computational cost, resulting in large algorithmic superlinear speedups.
v2023.06.11a
1
0.9
0.8
0.7
Probability
0.6
0.5 SEQ/PWQ SEQ/PART
0.4
0.3
0.2
0.1
0
0.1 1 10 100
CDF of Speedup Relative to SEQ
Figure 6.25: CDF of SEQ/PWQ and SEQ/PART Solution-Time Ratios
Figure 6.26: Reason for Small Visit Percentages
140
120
100
Solution Time (ms)
SEQ
80
60 PWQ
40
20 PART
0
0 10 20 30 40 50 60 70 80 90 100
Percent of Maze Cells Visited
Figure 6.27: Correlation Between Visit Percentage and Solution Time
v2023.06.11a
Figure 6.28: PWQ Potential Contention Points
Further investigation showed that PART sometimes visited fewer than 2 % of the
maze’s cells, while SEQ and PWQ never visited fewer than about 9 %. The reason for
this difference is shown by Figure 6.26. If the thread traversing the solution from the
upper left reaches the circle, the other thread cannot reach the upper-right portion of the
maze. Similarly, if the other thread reaches the square, the first thread cannot reach the
lower-left portion of the maze. Therefore, PART will likely visit a small fraction of the
non-solution-path cells. In short, the superlinear speedups are due to threads getting
in each others’ way. This is a sharp contrast with decades of experience with parallel
programming, where workers have struggled to keep threads out of each others’ way.
Figure 6.27 confirms a strong correlation between cells visited and solution time
for all three methods. The slope of PART’s scatterplot is smaller than that of SEQ,
indicating that PART’s pair of threads visits a given fraction of the maze faster than can
SEQ’s single thread. PART’s scatterplot is also weighted toward small visit percentages,
confirming that PART does less total work, hence the observed humiliating parallelism.
This humiliating parallelism also provides more than 2x speedup on two CPUs, as put
forth in page 113.
The fraction of cells visited by PWQ is similar to that of SEQ. In addition, PWQ’s
solution time is greater than that of PART, even for equal visit fractions. The reason for
this is shown in Figure 6.28, which has a red circle on each cell with more than two
neighbors. Each such cell can result in contention in PWQ, because one thread can
enter but two threads can exit, which hurts performance, as noted earlier in this chapter.
In contrast, PART can incur such contention but once, namely when the solution is
located. Of course, SEQ never contends.
Quick Quiz 6.26: Given that a 2D maze achieved 4x speedup on two CPUs, would a 3D maze
achieve an 8x speedup on two CPUs?
Although PART’s speedup is impressive, we should not neglect sequential optimiza-

tions. Figure 6.29 shows that SEQ, when compiled with -O3, is about twice as fast as
unoptimized PWQ, approaching the performance of unoptimized PART. Compiling all
three algorithms with -O3 gives results similar to (albeit faster than) those shown in
Figure 6.25, except that PWQ provides almost no speedup compared to SEQ, in keeping
with Amdahl’s Law [Amd67]. However, if the goal is to double performance compared
to unoptimized SEQ, as opposed to achieving optimality, compiler optimizations are
quite attractive.
Cache alignment and padding often improves performance by reducing false sharing.
However, for these maze-solution algorithms, aligning and padding the maze-cell array
degrades performance by up to 42 % for 1000x1000 mazes. Cache locality is more
important than avoiding false sharing, especially for large mazes. For smaller 20-by-20
v2023.06.11a
1
0.9
PART
0.8
0.7
Probability
0.6
0.5 PWQ
0.4
0.3
0.2
0.1 SEQ -O3
0
0.1 1 10 100
CDF of Speedup Relative to SEQ
Figure 6.29: Effect of Compiler Optimization (-O3)
1
COPART
0.9
PWQ
0.8 PART
0.7
Probability
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1 1 10 100
CDF of Speedup Relative to SEQ (-O3)
Figure 6.30: Partitioned Coroutines
or 50-by-50 mazes, aligning and padding can produce up to a 40 % performance

improvement for PART, but for these small sizes, SEQ performs better anyway because
there is insufficient time for PART to make up for the overhead of thread creation and
destruction.
In short, the partitioned parallel maze solver is an interesting example of an algorithmic
superlinear speedup. If “algorithmic superlinear speedup” causes cognitive dissonance,
please proceed to the next section.
6.5.5 Alternative Sequential Maze Solver

The presence of algorithmic superlinear speedups suggests simulating parallelism via
co-routines, for example, manually switching context between threads on each pass
through the main do-while loop in Listing 6.14. This context switching is straightforward
because the context consists only of the variables c and vi: Of the numerous ways to
achieve the effect, this is a good tradeoff between context-switch overhead and visit
percentage. As can be seen in Figure 6.30, this coroutine algorithm (COPART) is quite
effective, with the performance on one thread being within about 30 % of PART on two
threads (maze_2seq.c).
v2023.06.11a
12
Speedup Relative to SEQ (-O3)

10
2 PART PWQ
0
10 100 1000
Maze Size
Figure 6.31: Varying Maze Size vs. SEQ
1.8
Speedup Relative to COPART (-O3)
1.6
1.4
1.2
0.8 PART
0.6
PWQ
0.4
0.2
0
10 100 1000
Maze Size
Figure 6.32: Varying Maze Size vs. COPART
6.5.6 Performance Comparison II

Figures 6.31 and 6.32 show the effects of varying maze size, comparing both PWQ
and PART running on two threads against either SEQ or COPART, respectively, with
90-percent-confidence error bars. PART shows superlinear scalability against SEQ and
modest scalability against COPART for 100-by-100 and larger mazes. PART exceeds
theoretical energy-efficiency breakeven against COPART at roughly the 200-by-200
maze size, given that power consumption rises as roughly the square of the frequency
for high frequencies [Mud01], so that 1.4x scaling on two threads consumes the same
energy as a single thread at equal solution speeds. In contrast, PWQ shows poor
scalability against both SEQ and COPART unless unoptimized: Figures 6.31 and 6.32
were generated using -O3.
Figure 6.33 shows the performance of PWQ and PART relative to COPART. For
PART runs with more than two threads, the additional threads were started evenly
spaced along the diagonal connecting the starting and ending cells. Simplified link-state
routing [BG87] was used to detect early termination on PART runs with more than two
threads (the solution is flagged when a thread is connected to both beginning and end).
PWQ performs quite poorly, but PART hits breakeven at two threads and again at five
threads, achieving modest speedups beyond five threads. Theoretical energy efficiency
breakeven is within the 90-percent-confidence interval for seven and eight threads. The
v2023.06.11a
3.5
Mean Speedup Relative to COPART (-O3)

3
2.5
1.5
PART
1
0.5
PWQ
0
1 2 3 4 5 6 7 8
Number of Threads
Figure 6.33: Mean Speedup vs. Number of Threads, 1000x1000 Maze
reasons for the peak at two threads are (1) the lower complexity of termination detection
in the two-thread case and (2) the fact that there is a lower probability of the third
and subsequent threads making useful forward progress: Only the first two threads are
guaranteed to start on the solution line. This disappointing performance compared to
results in Figure 6.32 is due to the less-tightly integrated hardware available in the larger
and older Xeon system running at 2.66 GHz.
Quick Quiz 6.27: Why place the third, fourth, and so on threads on the diagonal? Why not
instead distribute them evenly around the maze?
6.5.7 Future Directions and Conclusions

Much future work remains. First, this section applied only one technique used by
human maze solvers. Others include following walls to exclude portions of the maze
and choosing internal starting points based on the locations of previously traversed
paths. Second, different choices of starting and ending points might favor different
algorithms. Third, although placement of the PART algorithm’s first two threads is
straightforward, there are any number of placement schemes for the remaining threads.
Optimal placement might well depend on the starting and ending points. Fourth, study
of unsolvable mazes and cyclic mazes is likely to produce interesting results. Fifth, the
lightweight C++11 atomic operations might improve performance. Sixth, it would be
interesting to compare the speedups for three-dimensional mazes (or of even higher-order
mazes). Finally, for mazes, humiliating parallelism indicated a more-efficient sequential
implementation using coroutines. Do humiliatingly parallel algorithms always lead to
more-efficient sequential implementations, or are there inherently humiliatingly parallel
algorithms for which coroutine context-switch overhead overwhelms the speedups?
This section demonstrated and analyzed parallelization of maze-solution algorithms.
A conventional work-queue-based algorithm did well only when compiler optimizations
were disabled, suggesting that some prior results obtained using high-level/overhead
languages will be invalidated by advances in optimization.
This section gave a clear example where approaching parallelism as a first-class
optimization technique rather than as a derivative of a sequential algorithm paves
the way for an improved sequential algorithm. High-level design-time application of
parallelism is likely to be a fruitful field of study. This section took the problem of
v2023.06.11a
solving mazes from mildly scalable to humiliatingly parallel and back again. It is hoped
that this experience will motivate work on parallelism as a first-class design-time whole-
application optimization technique, rather than as a grossly suboptimal after-the-fact
micro-optimization to be retrofitted into existing programs.
6.6 Partitioning, Parallelism, and Optimization

Knowledge is of no value unless you put it into
practice.
Anton Chekhov
Most important, although this chapter has demonstrated that applying parallelism at
the design level gives excellent results, this final section shows that this is not enough.
For search problems such as maze solution, this section has shown that search strategy
is even more important than parallel design. Yes, for this particular type of maze,
intelligently applying parallelism identified a superior search strategy, but this sort of
luck is no substitute for a clear focus on search strategy itself.
As noted back in Section 2.2, parallelism is but one potential optimization of many.
A successful design needs to focus on the most important optimization. Much though I
might wish to claim otherwise, that optimization might or might not be parallelism.
However, for the many cases where parallelism is the right optimization, the next
section covers that synchronization workhorse, locking.
v2023.06.11a
Locking is the worst general-purpose
synchronization mechanism except for all those
other mechanisms that have been tried from time to
time.
With apologies to the memory of Winston Churchill
and to whoever he was quoting
Chapter 7
Locking
In recent concurrency research, locking often plays the role of villain. Locking stands
accused of inciting deadlocks, convoying, starvation, unfairness, data races, and all
manner of other concurrency sins. Interestingly enough, the role of workhorse in
production-quality shared-memory parallel software is also played by locking. This
chapter will look into this dichotomy between villain and hero, as fancifully depicted in
Figures 7.1 and 7.2.
There are a number of reasons behind this Jekyll-and-Hyde dichotomy:
1. Many of locking’s sins have pragmatic design solutions that work well in most
cases, for example:
(a) Use of lock hierarchies to avoid deadlock.
(b) Deadlock-detection tools, for example, the Linux kernel’s lockdep facil-
ity [Cor06a].
(c) Locking-friendly data structures, such as arrays, hash tables, and radix trees,
which will be covered in Chapter 10.
2. Some of locking’s sins are problems only at high levels of contention, levels
reached only by poorly designed programs.
3. Some of locking’s sins are avoided by using other synchronization mechanisms in
concert with locking. These other mechanisms include statistical counters (see
Chapter 5), reference counters (see Section 9.2), hazard pointers (see Section 9.3),
sequence-locking readers (see Section 9.4), RCU (see Section 9.5), and simple
non-blocking data structures (see Section 14.2).
4. Until quite recently, almost all large shared-memory parallel programs were
developed in secret, so that it was not easy to learn of these pragmatic solutions.
5. Locking works extremely well for some software artifacts and extremely poorly for
others. Developers who have worked on artifacts for which locking works well
can be expected to have a much more positive opinion of locking than those who
have worked on artifacts for which locking works poorly, as will be discussed in
Section 7.5.
6. All good stories need a villain, and locking has a long and honorable history
serving as a research-paper whipping boy.
159
v2023.06.11a
160 CHAPTER 7. LOCKING
XXXX
Figure 7.1: Locking: Villain or Slob?
Figure 7.2: Locking: Workhorse or Hero?
Quick Quiz 7.1: Just how can serving as a whipping boy be considered to be in any way
honorable???
This chapter will give an overview of a number of ways to avoid locking’s more
serious sins.
v2023.06.11a
7.1. STAYING ALIVE 161
Lock 1
Thread A Lock 2
Lock 3 Thread B
Thread C Lock 4
Figure 7.3: Deadlock Cycle
7.1 Staying Alive

I work to stay alive.
Bette Davis
Given that locking stands accused of deadlock and starvation, one important concern
for shared-memory parallel developers is simply staying alive. The following sections
therefore cover deadlock, livelock, starvation, unfairness, and inefficiency.
7.1.1 Deadlock
Deadlock occurs when each member of a group of threads is holding at least one lock
while at the same time waiting on a lock held by a member of that same group. This
happens even in groups containing a single thread when that thread attempts to acquire
a non-recursive lock that it already holds. Deadlock can therefore occur even given but
one thread and one lock!
Without some sort of external intervention, deadlock is forever. No thread can acquire
the lock it is waiting on until that lock is released by the thread holding it, but the thread
holding it cannot release it until the holding thread acquires the lock that it is in turn
waiting on.
We can create a directed-graph representation of a deadlock scenario with nodes for
threads and locks, as shown in Figure 7.3. An arrow from a lock to a thread indicates
that the thread holds the lock, for example, Thread B holds Locks 2 and 4. An arrow
from a thread to a lock indicates that the thread is waiting on the lock, for example,
Thread B is waiting on Lock 3.
A deadlock scenario will always contain at least one deadlock cycle. In Figure 7.3,
this cycle is Thread B, Lock 3, Thread C, Lock 4, and back to Thread B.
Quick Quiz 7.2: But the definition of lock-based deadlock only said that each thread was
holding at least one lock and waiting on another lock that was held by some thread. How do
you know that there is a cycle?
Although there are some software environments such as database systems that can
recover from an existing deadlock, this approach requires either that one of the threads
v2023.06.11a
be killed or that a lock be forcibly stolen from one of the threads. This killing and
forcible stealing works well for transactions, but is often problematic for kernel and
application-level use of locking: Dealing with the resulting partially updated structures
can be extremely complex, hazardous, and error-prone.
Therefore, kernels and applications should instead avoid deadlocks. Deadlock-
avoidance strategies include locking hierarchies (Section 7.1.1.1), local locking hierar-
chies (Section 7.1.1.2), layered locking hierarchies (Section 7.1.1.3), temporal locking
hierarchies (Section 7.1.1.4), strategies for dealing with APIs containing pointers to
locks (Section 7.1.1.5), conditional locking (Section 7.1.1.6), acquiring all needed locks
first (Section 7.1.1.7), single-lock-at-a-time designs (Section 7.1.1.8), and strategies for
signal/interrupt handlers (Section 7.1.1.9). Although there is no deadlock-avoidance
strategy that works perfectly for all situations, there is a good selection of tools to choose
from.
7.1.1.1 Locking Hierarchies

Locking hierarchies order the locks and prohibit acquiring locks out of order. In
Figure 7.3, we might order the locks numerically, thus forbidding a thread from
acquiring a given lock if it already holds a lock with the same or a higher number.
Thread B has violated this hierarchy because it is attempting to acquire Lock 3 while
holding Lock 4. This violation permitted the deadlock to occur.
Again, to apply a locking hierarchy, order the locks and prohibit out-of-order lock
acquisition. For different types of locks, it is helpful to have a carefully considered
hierarchy from one type to the next. For many instances of the same type of lock, for
example, a per-node lock in a search tree, the traditional approach is to carry out lock
acquisition in order of the addresses of the locks to be acquired. Either way, in large
program, it is wise to use tools such as the Linux-kernel lockdep [Cor06a] to enforce
your locking hierarchy.
7.1.1.2 Local Locking Hierarchies

However, the global nature of locking hierarchies makes them difficult to apply to
library functions. After all, when a program using a given library function has not
yet been written, how can the poor library-function implementor possibly follow the
yet-to-be-defined locking hierarchy?
One special (but common) case is when the library function does not invoke any of
the caller’s code. In this case, the caller’s locks will never be acquired while holding any
of the library’s locks, so that there cannot be a deadlock cycle containing locks from
both the library and the caller.
Quick Quiz 7.3: Are there any exceptions to this rule, so that there really could be a deadlock
cycle containing locks from both the library and the caller, even given that the library code
never invokes any of the caller’s functions?
But suppose that a library function does invoke the caller’s code. For example,
qsort() invokes a caller-provided comparison function. Now, normally this comparison
function will operate on unchanging local data, so that it need not acquire locks, as
shown in Figure 7.4. But maybe someone is crazy enough to sort a collection whose
keys are changing, thus requiring that the comparison function acquire locks, which
might result in deadlock, as shown in Figure 7.5. How can the library function avoid
this deadlock?
v2023.06.11a
Application
Lock A Lock B
foo() bar() cmp()
Library
Lock C
qsort()
Figure 7.4: No qsort() Compare-Function Locking
Application
Lock A Lock B Lock B
foo() bar() cmp()
DEADLOCK
Library
Lock C
qsort()
Figure 7.5: Without qsort() Local Locking Hierarchy
The golden rule in this case is “Release all locks before invoking unknown code.”
To follow this rule, the qsort() function must release all of its locks before invoking
the comparison function. Thus qsort() will not be holding any of its locks while the
comparison function acquires any of the caller’s locks, thus avoiding deadlock.
Quick Quiz 7.4: But if qsort() releases all its locks before invoking the comparison function,
how can it protect against races with other qsort() threads?
To see the benefits of local locking hierarchies, compare Figures 7.5 and 7.6. In
both figures, application functions foo() and bar() invoke qsort() while holding
Locks A and B, respectively. Because this is a parallel implementation of qsort(),
it acquires Lock C. Function foo() passes function cmp() to qsort(), and cmp()
acquires Lock B. Function bar() passes a simple integer-comparison function (not
shown) to qsort(), and this simple function does not acquire any locks.
Now, if qsort() holds Lock C while calling cmp() in violation of the golden
release-all-locks rule above, as shown in Figure 7.5, deadlock can occur. To see this,
suppose that one thread invokes foo() while a second thread concurrently invokes
bar(). The first thread will acquire Lock A and the second thread will acquire Lock B.
v2023.06.11a
Application
Lock A Lock B Lock B
foo() bar() cmp()
Library
Lock C
qsort()
Figure 7.6: Local Locking Hierarchy for qsort()
If the first thread’s call to qsort() acquires Lock C, then it will be unable to acquire
Lock B when it calls cmp(). But the first thread holds Lock C, so the second thread’s
call to qsort() will be unable to acquire it, and thus unable to release Lock B, resulting
in deadlock.
In contrast, if qsort() releases Lock C before invoking the comparison function,
which is unknown code from qsort()’s perspective, then deadlock is avoided as shown
in Figure 7.6.
If each module releases all locks before invoking unknown code, then deadlock is
avoided if each module separately avoids deadlock. This rule therefore greatly simplifies
deadlock analysis and greatly improves modularity.
Nevertheless, this golden rule comes with a warning. When you release those locks,
any state that they protect is subject to arbitrary changes, changes that are all too
easy for the function’s caller to forget, resulting in subtle and difficult-to-reproduce
bugs. Because the qsort() comparison function rarely acquires locks, let’s switch to a
different example.
Consider the recursive tree iterator in Listing 7.1 (rec_tree_itr.c). The iterator
visits every node in the tree, invoking a user’s callback function. The tree lock is
released before the invocation and re-acquired after return. This code makes dangerous
assumptions: (1) The number of children of the current node has not changed, (2) The
ancestors stored on the stack by the recursion are still there, and (3) The visited node
itself has not been removed and freed. A few of these hazards can be encountered if one
thread calls tree_add() while another thread releases the tree’s lock to run a callback
function.
Quick Quiz 7.5: So the iterating thread may or may not observe the added child. What is the
big deal?
One strategy is to ensure that state is preserved despite the lock being released, for
example, by acquiring a reference on a node to prevent it from being freed. Alternatively,
the state can be re-initialized once the lock is re-acquired after the callback function
returns.
v2023.06.11a
Listing 7.1: Recursive Tree Iterator

1 struct node {
2 int data;
3 int nchildren;
4 struct node **children;
5 };
6
7 struct tree {
8 spinlock_t s;
9 struct node *root;
10 };
11
12 void tree_for_each_rec(struct tree *tr, struct node *nd,
13 void (*callback)(struct node *))
14 {
15 struct node **itr;
16
17 spin_unlock(&tr->s);
18 callback(nd);
19 spin_lock(&tr->s);
20
21 itr = nd->children;
22 for (int i = 0; i < nd->nchildren; i++) {
23 tree_for_each_rec(tr, *itr, callback);
24 itr++;
25 }
26 }
27
28 void tree_for_each(struct tree *tr,
29 void (*callback)(struct node *))
30 {
32 tree_for_each_rec(tr, tr->root, callback);
34 }
35
36 void tree_add(struct tree *tr, struct node *parent,
37 struct node *new_child)
38 {
40 parent->nchildren++;
41 parent->children = realloc(parent->children,
42 sizeof(struct node *) *
43 parent->nchildren);
44 parent->children[parent->nchildren - 1] = new_child;
46 }
v2023.06.11a
Application
Lock A Lock B
foo() bar()
Library
Lock C
qsort()
Lock D
cmp()
Figure 7.7: Layered Locking Hierarchy for qsort()
7.1.1.3 Layered Locking Hierarchies
Unfortunately, it might be infeasible to preserve state on the one hand or to re-initialize

it on the other, thus ruling out a local locking hierarchy where all locks are released
before invoking unknown code. However, we can instead construct a layered locking
hierarchy, as shown in Figure 7.7. Here, the cmp() function uses a new Lock D that is
acquired after all of Locks A, B, and C, avoiding deadlock. We therefore have three
layers to the global deadlock hierarchy, the first containing Locks A and B, the second
containing Lock C, and the third containing Lock D.
Please note that it is not typically possible to mechanically change cmp() to use the
new Lock D. Quite the opposite: It is often necessary to make profound design-level
modifications. Nevertheless, the effort required for such modifications is normally a
small price to pay in order to avoid deadlock. More to the point, this potential deadlock
should preferably be detected at design time, before any code has been generated!
For another example where releasing all locks before invoking unknown code is
impractical, imagine an iterator over a linked list, as shown in Listing 7.2 (locked_
list.c). The list_start() function acquires a lock on the list and returns the first
element (if there is one), and list_next() either returns a pointer to the next element
in the list or releases the lock and returns NULL if the end of the list has been reached.
Listing 7.3 shows how this list iterator may be used. Lines 1–4 define the list_ints
element containing a single integer, and lines 6–17 show how to iterate over the list.
Line 11 locks the list and fetches a pointer to the first element, line 13 provides a pointer
to our enclosing list_ints structure, line 14 prints the corresponding integer, and
line 15 moves to the next element. This is quite simple, and hides all of the locking.
That is, the locking remains hidden as long as the code processing each list element
does not itself acquire a lock that is held across some other call to list_start() or
v2023.06.11a
Listing 7.2: Concurrent List Iterator

1 struct locked_list {
2 spinlock_t s;
3 struct cds_list_head h;
4 };
5
6 struct cds_list_head *list_start(struct locked_list *lp)
7 {
8 spin_lock(&lp->s);
9 return list_next(lp, &lp->h);
10 }
11
12 struct cds_list_head *list_next(struct locked_list *lp,
13 struct cds_list_head *np)
14 {
15 struct cds_list_head *ret;
16
17 ret = np->next;
18 if (ret == &lp->h) {
19 spin_unlock(&lp->s);
20 ret = NULL;
21 }
22 return ret;
23 }
Listing 7.3: Concurrent List Iterator Usage

1 struct list_ints {
2 struct cds_list_head n;
3 int a;
4 };
5
6 void list_print(struct locked_list *lp)
7 {
8 struct cds_list_head *np;
9 struct list_ints *ip;
10
11 np = list_start(lp);
12 while (np != NULL) {
13 ip = cds_list_entry(np, struct list_ints, n);
14 printf("\t%d\n", ip->a);
15 np = list_next(lp, np);
16 }
17 }
v2023.06.11a
list_next(), which results in deadlock. We can avoid the deadlock by layering the
locking hierarchy to take the list-iterator locking into account.
This layered approach can be extended to an arbitrarily large number of layers, but
each added layer increases the complexity of the locking design. Such increases in
complexity are particularly inconvenient for some types of object-oriented designs, in
which control passes back and forth among a large group of objects in an undisciplined
manner.1 This mismatch between the habits of object-oriented design and the need to
avoid deadlock is an important reason why parallel programming is perceived by some
to be so difficult.
Some alternatives to highly layered locking hierarchies are covered in Chapter 9.
7.1.1.4 Temporal Locking Hierarchies

One way to avoid deadlock is to defer acquisition of one of the conflicting locks. This
approach is used in Linux-kernel RCU, whose call_rcu() function is invoked by the
Linux-kernel scheduler while holding its locks. This means that call_rcu() cannot
always safely invoke the scheduler to do a wakeup, for example, in order to wake up
an RCU kthread in order to start the new grace period that is required by the callback
queued by call_rcu().
Quick Quiz 7.6: What do you mean “cannot always safely invoke the scheduler”? Either
call_rcu() can or cannot safely invoke the scheduler, right?
However, grace periods last for many milliseconds, so waiting another millisecond
before starting a new grace period is not normally a problem. Therefore, if call_rcu()
detects a possible deadlock with the scheduler, it arranges to start the new grace period
later, either within a timer handler or within the scheduler-clock interrupt handler,
depending on configuration. Because no scheduler locks are held across either handler,
deadlock is successfully avoided.
The overall approach is thus to adhere to a locking hierarchy by deferring lock
acquisition to an environment in which no locks are held.
7.1.1.5 Locking Hierarchies and Pointers to Locks

Although there are some exceptions, an external API containing a pointer to a lock
is very often a misdesigned API. Handing an internal lock to some other software
component is after all the antithesis of information hiding, which is in turn a key design
principle.
Quick Quiz 7.7: Name one common situation where a pointer to a lock is passed into a
function.
One exception is functions that hand off some entity, where the caller’s lock must
be held until the handoff is complete, but where the lock must be released before the
function returns. One example of such a function is the POSIX pthread_cond_wait()
function, where passing a pointer to a pthread_mutex_t prevents hangs due to lost
wakeups.
Quick Quiz 7.8: Doesn’t the fact that pthread_cond_wait() first releases the mutex and
then re-acquires it eliminate the possibility of deadlock?
1 One name for this is “object-oriented spaghetti code.”
v2023.06.11a
Listing 7.4: Protocol Layering and Deadlock

1 spin_lock(&lock2);
2 layer_2_processing(pkt);
3 nextlayer = layer_1(pkt);
4 spin_lock(&nextlayer->lock1);
5 spin_unlock(&lock2);
7 spin_unlock(&nextlayer->lock1);
Listing 7.5: Avoiding Deadlock Via Conditional Locking

1 retry:
4 nextlayer = layer_1(pkt);
5 if (!spin_trylock(&nextlayer->lock1)) {
7 spin_lock(&nextlayer->lock1);
9 if (layer_1(pkt) != nextlayer) {
12 goto retry;
13 }
14 }
In short, if you find yourself exporting an API with a pointer to a lock as an argument
or as the return value, do yourself a favor and carefully reconsider your API design. It
might well be the right thing to do, but experience indicates that this is unlikely.
7.1.1.6 Conditional Locking

But suppose that there is no reasonable locking hierarchy. This can happen in real life,
for example, in some types of layered network protocol stacks where packets flow in
both directions, for example, in implementations of distributed lock managers. In the
networking case, it might be necessary to hold the locks from both layers when passing
a packet from one layer to another. Given that packets travel both up and down the
protocol stack, this is an excellent recipe for deadlock, as illustrated in Listing 7.4. Here,
a packet moving down the stack towards the wire must acquire the next layer’s lock out
of order. Given that packets moving up the stack away from the wire are acquiring the
locks in order, the lock acquisition in line 4 of the listing can result in deadlock.
One way to avoid deadlocks in this case is to impose a locking hierarchy, but when
it is necessary to acquire a lock out of order, acquire it conditionally, as shown in
Listing 7.5. Instead of unconditionally acquiring the layer-1 lock, line 5 conditionally
acquires the lock using the spin_trylock() primitive. This primitive acquires the
lock immediately if the lock is available (returning non-zero), and otherwise returns
zero without acquiring the lock.
If spin_trylock() was successful, line 16 does the needed layer-1 processing.
Otherwise, line 6 releases the lock, and lines 7 and 8 acquire them in the correct order.
Unfortunately, there might be multiple networking devices on the system (e.g., Ethernet
and WiFi), so that the layer_1() function must make a routing decision. This decision
might change at any time, especially if the system is mobile.2 Therefore, line 9 must
recheck the decision, and if it has changed, must release the locks and start over.
2 And, in contrast to the 1900s, mobility is the common case.
v2023.06.11a
Quick Quiz 7.9: Can the transformation from Listing 7.4 to Listing 7.5 be applied universally?
Quick Quiz 7.10: But the complexity in Listing 7.5 is well worthwhile given that it avoids
deadlock, right?
7.1.1.7 Acquire Needed Locks First

In an important special case of conditional locking, all needed locks are acquired before
any processing is carried out, where the needed locks might be identified by hashing
the addresses of the data structures involved. In this case, processing need not be
idempotent: If it turns out to be impossible to acquire a given lock without first releasing
one that was already acquired, just release all the locks and try again. Only once all
needed locks are held will any processing be carried out.
However, this procedure can result in livelock, which will be discussed in Section 7.1.2.
Quick Quiz 7.11: When using the “acquire needed locks first” approach described in
Section 7.1.1.7, how can livelock be avoided?
A related approach, two-phase locking [BHG87], has seen long production use in
transactional database systems. In the first phase of a two-phase locking transaction, locks
are acquired but not released. Once all needed locks have been acquired, the transaction
enters the second phase, where locks are released, but not acquired. This locking
approach allows databases to provide serializability guarantees for their transactions,
in other words, to guarantee that all values seen and produced by the transactions are
consistent with some global ordering of all the transactions. Many such systems rely on
the ability to abort transactions, although this can be simplified by avoiding making
any changes to shared data until all needed locks are acquired. Livelock and deadlock
are issues in such systems, but practical solutions may be found in any of a number of
database textbooks.
7.1.1.8 Single-Lock-at-a-Time Designs

In some cases, it is possible to avoid nesting locks, thus avoiding deadlock. For example,
if a problem is perfectly partitionable, a single lock may be assigned to each partition.
Then a thread working on a given partition need only acquire the one corresponding
lock. Because no thread ever holds more than one lock at a time, deadlock is impossible.
However, there must be some mechanism to ensure that the needed data structures
remain in existence during the time that neither lock is held. One such mechanism is
discussed in Section 7.4 and several others are presented in Chapter 9.
7.1.1.9 Signal/Interrupt Handlers

Deadlocks involving signal handlers are often quickly dismissed by noting that it is
not legal to invoke pthread_mutex_lock() from within a signal handler [Ope97].
However, it is possible (though often unwise) to hand-craft locking primitives that can
be invoked from signal handlers. Besides which, almost all operating-system kernels
permit locks to be acquired from within interrupt handlers, which are analogous to
signal handlers.
The trick is to block signals (or disable interrupts, as the case may be) when acquiring
any lock that might be acquired within a signal (or an interrupt) handler. Furthermore,
v2023.06.11a
if holding such a lock, it is illegal to attempt to acquire any lock that is ever acquired
outside of a signal handler without blocking signals.
Quick Quiz 7.12: Suppose Lock A is never acquired within a signal handler, but Lock B
is acquired both from thread context and by signal handlers. Suppose further that Lock A
is sometimes acquired with signals unblocked. Why is it illegal to acquire Lock A holding
Lock B?
If a lock is acquired by the handlers for several signals, then each and every one of
these signals must be blocked whenever that lock is acquired, even when that lock is
acquired within a signal handler.
Quick Quiz 7.13: How can you legally block signals within a signal handler?
Unfortunately, blocking and unblocking signals can be expensive in some operating

systems, notably including Linux, so performance concerns often mean that locks
acquired in signal handlers are only acquired in signal handlers, and that lockless
synchronization mechanisms are used to communicate between application code and
signal handlers.
Or that signal handlers are avoided completely except for handling fatal errors.
Quick Quiz 7.14: If acquiring locks in signal handlers is such a bad idea, why even discuss
ways of making it safe?
7.1.1.10 Discussion
There are a large number of deadlock-avoidance strategies available to the shared-
memory parallel programmer, but there are sequential programs for which none of them
is a good fit. This is one of the reasons that expert programmers have more than one
tool in their toolbox: Locking is a powerful concurrency tool, but there are jobs better
addressed with other tools.
Quick Quiz 7.15: Given an object-oriented application that passes control freely among a
group of objects such that there is no straightforward locking hierarchy,a layered or otherwise,
how can this application be parallelized?
a Also known as “object-oriented spaghetti code.”
Nevertheless, the strategies described in this section have proven quite useful in many
settings.
7.1.2 Livelock and Starvation

Although conditional locking can be an effective deadlock-avoidance mechanism, it
can be abused. Consider for example the beautifully symmetric example shown in
Listing 7.6. This example’s beauty hides an ugly livelock. To see this, consider the
following sequence of events:
1. Thread 1 acquires lock1 on line 4, then invokes do_one_thing().

2. Thread 2 acquires lock2 on line 18, then invokes do_a_third_thing().
3. Thread 1 attempts to acquire lock2 on line 6, but fails because Thread 2 holds it.
4. Thread 2 attempts to acquire lock1 on line 20, but fails because Thread 1 holds it.
v2023.06.11a
Listing 7.6: Abusing Conditional Locking

1 void thread1(void)
2 {
3 retry:
5 do_one_thing();
6 if (!spin_trylock(&lock2)) {
8 goto retry;
9 }
10 do_another_thing();
13 }
14
16 {
17 retry:
19 do_a_third_thing();
22 goto retry;
23 }
24 do_a_fourth_thing();
27 }
5. Thread 1 releases lock1 on line 7, then jumps to retry at line 3.

6. Thread 2 releases lock2 on line 21, and jumps to retry at line 17.
7. The livelock dance repeats from the beginning.
Quick Quiz 7.16: How can the livelock shown in Listing 7.6 be avoided?
Livelock can be thought of as an extreme form of starvation where a group of threads

starves, rather than just one of them.3
Livelock and starvation are serious issues in software transactional memory implemen-
tations, and so the concept of contention manager has been introduced to encapsulate
these issues. In the case of locking, simple exponential backoff can often address
livelock and starvation. The idea is to introduce exponentially increasing delays before
each retry, as shown in Listing 7.7.
Quick Quiz 7.17: What problems can you spot in the code in Listing 7.7?
For better results, backoffs should be bounded, and even better high-contention results
are obtained via queued locking [And90], which is discussed more in Section 7.3.2.
Of course, best of all is to use a good parallel design that avoids these problems by
maintaining low lock contention.
7.1.3 Unfairness
Unfairness can be thought of as a less-severe form of starvation, where a subset of
threads contending for a given lock are granted the lion’s share of the acquisitions. This
3 Try not to get too hung up on the exact definitions of terms like livelock, starvation,
and unfairness. Anything that causes a group of threads to fail to make adequate forward
progress is a bug that needs to be fixed, and debating names doesn’t fix bugs.
v2023.06.11a
Listing 7.7: Conditional Locking and Exponential Backoff

2 {
3 unsigned int wait = 1;
4 retry:
6 do_one_thing();
9 sleep(wait);
10 wait = wait << 1;
11 goto retry;
12 }
16 }
17
19 {
20 unsigned int wait = 1;
21 retry:
26 sleep(wait);
27 wait = wait << 1;
28 goto retry;
29 }
33 }

Speed−of−Light Round−Trip Distance in Vacuum

for 1.8 GHz Clock Period (8 cm)
Figure 7.8: System Architecture and Lock Unfairness
v2023.06.11a
Figure 7.9: Saw Kerf
can happen on machines with shared caches or NUMA characteristics, for example, as
shown in Figure 7.8. If CPU 0 releases a lock that all the other CPUs are attempting to
acquire, the interconnect shared between CPUs 0 and 1 means that CPU 1 will have an
advantage over CPUs 2–7. Therefore CPU 1 will likely acquire the lock. If CPU 1 holds
the lock long enough for CPU 0 to be requesting the lock by the time CPU 1 releases it
and vice versa, the lock can shuttle between CPUs 0 and 1, bypassing CPUs 2–7.
Quick Quiz 7.18: Wouldn’t it be better just to use a good parallel design so that lock contention
was low enough to avoid unfairness?
7.1.4 Inefficiency
Locks are implemented using atomic instructions and memory barriers, and often
involve cache misses. As we saw in Chapter 3, these instructions are quite expensive,
roughly two orders of magnitude greater overhead than simple instructions. This can be
a serious problem for locking: If you protect a single instruction with a lock, you will
increase the overhead by a factor of one hundred. Even assuming perfect scalability,
one hundred CPUs would be required to keep up with a single CPU executing the same
code without locking.
This situation is not confined to locking. Figure 7.9 shows how this same principle
applies to the age-old activity of sawing wood. As can be seen in the figure, sawing a
board converts a small piece of that board (the width of the saw blade) into sawdust.
Of course, locks partition time instead of sawing wood,4 but just like sawing wood,
using locks to partition time wastes some of that time due to lock overhead and (worse
yet) lock contention. One important difference is that if someone saws a board into
too-small pieces, the resulting conversion of most of that board into sawdust will be
immediately obvious. In contrast, it is not always obvious that a given lock acquisition
is wasting excessive amounts of time.
And this situation underscores the importance of the synchronization-granularity
tradeoff discussed in Section 6.3, especially Figure 6.16: Too coarse a granularity will
limit scalability, while too fine a granularity will result in excessive synchronization
overhead.
Acquiring a lock might be expensive, but once held, the CPU’s caches are an effective
performance booster, at least for large critical sections. In addition, once a lock is held,
the data protected by that lock can be accessed by the lock holder without interference
from other threads.
4 That is, locking is temporal synchronization. Mechanisms that synchronize both
temporally and spatially are described in Chapter 9.
v2023.06.11a
7.2. TYPES OF LOCKS 175
Quick Quiz 7.19: How might the lock holder be interfered with?
The Rust programming language takes lock/data association a step further by allowing
the developer to make a compiler-visible association between a lock and the data that
it protects [JJKD21]. When such an association has been made, attempts to access
the data without the benefit of the corresponding lock will result in a compile-time
diagnostic. The hope is that this will greatly reduce the frequency of this class of bugs.
Of course, this approach does not apply straightforwardly to cases where the data to be
locked is distributed throughout the nodes of some data structure or when that which is
locked is purely abstract, for example, when a small subset of state-machine transitions
is to be protected by a given lock. For this reason, Rust allows locks to be associated
with types rather than data items or even to be associated with nothing at all. This last
option permits Rust to emulate traditional locking use cases, but is not popular among
Rust developers. Perhaps the Rust community will come up with other mechanisms
tailored to other locking use cases.
7.2 Types of Locks
Only locks in life are what you think you know, but
don’t. Accept your ignorance and try something new.
Dennis Vickers
There are a surprising number of types of locks, more than this short chapter can
possibly do justice to. The following sections discuss exclusive locks (Section 7.2.1),
reader-writer locks (Section 7.2.2), multi-role locks (Section 7.2.3), and scoped locking
(Section 7.2.4).
7.2.1 Exclusive Locks

Exclusive locks are what they say they are: Only one thread may hold the lock at a time.
The holder of such a lock thus has exclusive access to all data protected by that lock,
hence the name.
Of course, this all assumes that this lock is held across all accesses to data purportedly
protected by the lock. Although there are some tools that can help (see for example
Section 12.3.1), the ultimate responsibility for ensuring that the lock is always acquired
when needed rests with the developer.
Quick Quiz 7.20: Does it ever make sense to have an exclusive lock acquisition immediately
followed by a release of that same lock, that is, an empty critical section?
It is important to note that unconditionally acquiring an exclusive lock has two

effects: (1) Waiting for all prior holders of that lock to release it and (2) Blocking any
other acquisition attempts until the lock is released. As a result, at lock acquisition
time, any concurrent acquisitions of that lock must be partitioned into prior holders
and subsequent holders. Different types of exclusive locks use different partitioning
strategies [Bra11, GGL+ 19], for example:
1. Strict FIFO, with acquisitions starting earlier acquiring the lock earlier.
v2023.06.11a
2. Approximate FIFO, with acquisitions starting sufficiently earlier acquiring the lock
earlier.
3. FIFO within priority level, with higher-priority threads acquiring the lock earlier
than any lower-priority threads attempting to acquire the lock at about the same
time, but so that some FIFO ordering applies for threads of the same priority.
4. Random, so that the new lock holder is chosen randomly from all threads attempting
acquisition, regardless of timing.
5. Unfair, so that a given acquisition might never acquire the lock (see Section 7.1.3).
Unfortunately, locking implementations with stronger guarantees typically incur

higher overhead, motivating the wide variety of locking implementations in production
use. For example, real-time systems often require some degree of FIFO ordering within
priority level, and much else besides (see Section 14.3.5.1), while non-realtime systems
subject to high contention might require only enough ordering to avoid starvation, and
finally, non-realtime systems designed to avoid contention might not need fairness at all.
7.2.2 Reader-Writer Locks

Reader-writer locks [CHP71] permit any number of readers to hold the lock concurrently
on the one hand or a single writer to hold the lock on the other. In theory, then, reader-
writer locks should allow excellent scalability for data that is read often and written
rarely. In practice, the scalability will depend on the reader-writer lock implementation.
The classic reader-writer lock implementation involves a set of counters and flags
that are manipulated atomically. This type of implementation suffers from the same
problem as does exclusive locking for short critical sections: The overhead of acquiring
and releasing the lock is about two orders of magnitude greater than the overhead of a
simple instruction. Of course, if the critical section is long enough, the overhead of
acquiring and releasing the lock becomes negligible. However, because only one thread
at a time can be manipulating the lock, the required critical-section size increases with
the number of CPUs.
It is possible to design a reader-writer lock that is much more favorable to readers
through use of per-thread exclusive locks [HW92]. To read, a thread acquires only its
own lock. To write, a thread acquires all locks. In the absence of writers, each reader
incurs only atomic-instruction and memory-barrier overhead, with no cache misses,
which is quite good for a locking primitive. Unfortunately, writers must incur cache
misses as well as atomic-instruction and memory-barrier overhead—multiplied by the
number of threads.
In short, reader-writer locks can be quite useful in a number of situations, but each type
of implementation does have its drawbacks. The canonical use case for reader-writer
locking involves very long read-side critical sections, preferably measured in hundreds
of microseconds or even milliseconds.
As with exclusive locks, a reader-writer lock acquisition cannot complete until all
prior conflicting holders of that lock have released it. If a lock is read-held, then read
acquisitions can complete immediately, but write acquisitions must wait until there are
no longer any readers holding the lock. If a lock is write-held, then all acquisitions
must wait until the writer releases the lock. Again as with exclusive locks, different
reader-writer lock implementations provide different degrees of FIFO ordering to readers
on the one hand and to writers on the other.
v2023.06.11a
Table 7.1: VAX/VMS Distributed Lock Manager Policy
Concurrent Write
Concurrent Read
Null (Not Held)
Protected Write
Protected Read
Exclusive
Null (Not Held)
Concurrent Read X
Concurrent Write X X X
Protected Read X X X
Protected Write X X X X
Exclusive X X X X X
But suppose a large number of readers hold the lock and a writer is waiting to acquire
the lock. Should readers be allowed to continue to acquire the lock, possibly starving the
writer? Similarly, suppose that a writer holds the lock and that a large number of both
readers and writers are waiting to acquire the lock. When the current writer releases the
lock, should it be given to a reader or to another writer? If it is given to a reader, how
many readers should be allowed to acquire the lock before the next writer is permitted
to do so?
There are many possible answers to these questions, with different levels of complexity,
overhead, and fairness. Different implementations might have different costs, for example,
some types of reader-writer locks incur extremely large latencies when switching from
read-holder to write-holder mode. Here are a few possible approaches:
1. Reader-preference implementations unconditionally favor readers over writers,

possibly allowing write acquisitions to be indefinitely blocked.
2. Batch-fair implementations ensure that when both readers and writers are acquiring
the lock, both have reasonable access via batching. For example, the lock might
admit five readers per CPU, then two writers, then five more readers per CPU, and
so on.
3. Writer-preference implementations unconditionally favor writers over readers,

possibly allowing read acquisitions to be indefinitely blocked.
Of course, these distinctions matter only under conditions of high lock contention.
Please keep the waiting/blocking dual nature of locks firmly in mind. This will
be revisited in Chapter 9’s discussion of scalable high-performance special-purpose
alternatives to locking.
7.2.3 Beyond Reader-Writer Locks

Reader-writer locks and exclusive locks differ in their admission policy: Exclusive
locks allow at most one holder, while reader-writer locks permit an arbitrary number
of read-holders (but only one write-holder). There is a very large number of possible
admission policies, one of which is that of the VAX/VMS distributed lock manager
v2023.06.11a
(DLM) [ST87], which is shown in Table 7.1. Blank cells indicate compatible modes,
while cells containing “X” indicate incompatible modes.
The VAX/VMS DLM uses six modes. For purposes of comparison, exclusive locks
use two modes (not held and held), while reader-writer locks use three modes (not held,
read held, and write held).
The first mode is null, or not held. This mode is compatible with all other modes,
which is to be expected: If a thread is not holding a lock, it should not prevent any other
thread from acquiring that lock.
The second mode is concurrent read, which is compatible with every other mode except
for exclusive. The concurrent-read mode might be used to accumulate approximate
statistics on a data structure, while permitting updates to proceed concurrently.
The third mode is concurrent write, which is compatible with null, concurrent read,
and concurrent write. The concurrent-write mode might be used to update approximate
statistics, while still permitting reads and concurrent updates to proceed concurrently.
The fourth mode is protected read, which is compatible with null, concurrent read, and
protected read. The protected-read mode might be used to obtain a consistent snapshot
of the data structure, while permitting reads but not updates to proceed concurrently.
The fifth mode is protected write, which is compatible with null and concurrent
read. The protected-write mode might be used to carry out updates to a data structure
that could interfere with protected readers but which could be tolerated by concurrent
readers.
The sixth and final mode is exclusive, which is compatible only with null. The
exclusive mode is used when it is necessary to exclude all other accesses.
It is interesting to note that exclusive locks and reader-writer locks can be emulated
by the VAX/VMS DLM. Exclusive locks would use only the null and exclusive modes,
while reader-writer locks might use the null, protected-read, and protected-write modes.
Quick Quiz 7.21: Is there any other way for the VAX/VMS DLM to emulate a reader-writer
lock?
Although the VAX/VMS DLM policy has seen widespread production use for
distributed databases, it does not appear to be used much in shared-memory applications.
One possible reason for this is that the greater communication overheads of distributed
databases can hide the greater overhead of the VAX/VMS DLM’s more-complex
admission policy.
Nevertheless, the VAX/VMS DLM is an interesting illustration of just how flexible
the concepts behind locking can be. It also serves as a very simple introduction to the
locking schemes used by modern DBMSes, which can have more than thirty locking
modes, compared to VAX/VMS’s six.
7.2.4 Scoped Locking

The locking primitives discussed thus far require explicit acquisition and release
primitives, for example, spin_lock() and spin_unlock(), respectively. Another
approach is to use the object-oriented resource-acquisition-is-initialization (RAII)
pattern [ES90].5 This pattern is often applied to auto variables in languages like C++,
where the corresponding constructor is invoked upon entry to the object’s scope, and
the corresponding destructor is invoked upon exit from that scope. This can be applied
to locking by having the constructor acquire the lock and the destructor free it.
5 Though more clearly expressed at https://www.stroustrup.com/bs_faq2.
html#finally.
v2023.06.11a
Root rcu_node
Structure
Leaf rcu_node Leaf rcu_node

Structure 0 Structure N
CPU 0
CPU 1
CPU m
CPU m * (N − 1)
CPU m * (N − 1) + 1
CPU m * N − 1
Figure 7.10: Locking Hierarchy
This approach can be quite useful, in fact in 1990 I was convinced that it was the only
type of locking that was needed.6 One very nice property of RAII locking is that you
don’t need to carefully release the lock on each and every code path that exits that scope,
a property that can eliminate a troublesome set of bugs.
However, RAII locking also has a dark side. RAII makes it quite difficult to encapsulate
lock acquisition and release, for example, in iterators. In many iterator implementations,
you would like to acquire the lock in the iterator’s “start” function and release it in the
iterator’s “stop” function. RAII locking instead requires that the lock acquisition and
release take place in the same level of scoping, making such encapsulation difficult or
even impossible.
Strict RAII locking also prohibits overlapping critical sections, due to the fact that
scopes must nest. This prohibition makes it difficult or impossible to express a number
of useful constructs, for example, locking trees that mediate between multiple concurrent
attempts to assert an event. Of an arbitrarily large group of concurrent attempts, only
one need succeed, and the best strategy for the remaining attempts is for them to fail as
quickly and painlessly as possible. Otherwise, lock contention becomes pathological on
large systems (where “large” is many hundreds of CPUs). Therefore, C++17 [Smi19]
has escapes from strict RAII in its unique_lock class, which allows the scope of the
critical section to be controlled to roughly the same extent as can be achieved with
explicit lock acquisition and release primitives.
Example strict-RAII-unfriendly data structures from Linux-kernel RCU are shown
in Figure 7.10. Here, each CPU is assigned a leaf rcu_node structure, and each
rcu_node structure has a pointer to its parent (named, oddly enough, ->parent), up to
the root rcu_node structure, which has a NULL ->parent pointer. The number of child
rcu_node structures per parent can vary, but is typically 32 or 64. Each rcu_node
structure also contains a lock named ->fqslock.
6 My later work with parallelism at Sequent Computer Systems very quickly disabused
me of this misguided notion.
v2023.06.11a
Listing 7.8: Conditional Locking to Reduce Contention

1 void force_quiescent_state(struct rcu_node *rnp_leaf)
2 {
3 int ret;
4 struct rcu_node *rnp = rnp_leaf;
5 struct rcu_node *rnp_old = NULL;
6
7 for (; rnp != NULL; rnp = rnp->parent) {
8 ret = (READ_ONCE(gp_flags)) ||
9 !raw_spin_trylock(&rnp->fqslock);
10 if (rnp_old != NULL)
11 raw_spin_unlock(&rnp_old->fqslock);
12 if (ret)
13 return;
14 rnp_old = rnp;
15 }
16 if (!READ_ONCE(gp_flags)) {
17 WRITE_ONCE(gp_flags, 1);
18 do_force_quiescent_state();
19 WRITE_ONCE(gp_flags, 0);
20 }
21 raw_spin_unlock(&rnp_old->fqslock);
22 }
The general approach is a tournament, where a given CPU conditionally acquires

its leaf rcu_node structure’s ->fqslock, and, if successful, attempt to acquire that of
the parent, then release that of the child. In addition, at each level, the CPU checks
a global gp_flags variable, and if this variable indicates that some other CPU has
asserted the event, the first CPU drops out of the competition. This acquire-then-release
sequence continues until either the gp_flags variable indicates that someone else won
the tournament, one of the attempts to acquire an ->fqslock fails, or the root rcu_node
structure’s ->fqslock has been acquired. If the root rcu_node structure’s ->fqslock
is acquired, a function named do_force_quiescent_state() is invoked.
Simplified code to implement this is shown in Listing 7.8. The purpose of this
function is to mediate between CPUs who have concurrently detected a need to invoke
the do_force_quiescent_state() function. At any given time, it only makes
sense for one instance of do_force_quiescent_state() to be active, so if there are
multiple concurrent callers, we need at most one of them to actually invoke do_force_
quiescent_state(), and we need the rest to (as quickly and painlessly as possible)
give up and leave.
To this end, each pass through the loop spanning lines 7–15 attempts to advance up
one level in the rcu_node hierarchy. If the gp_flags variable is already set (line 8) or
if the attempt to acquire the current rcu_node structure’s ->fqslock is unsuccessful
(line 9), then local variable ret is set to 1. If line 10 sees that local variable rnp_old
is non-NULL, meaning that we hold rnp_old’s ->fqs_lock, line 11 releases this lock
(but only after the attempt has been made to acquire the parent rcu_node structure’s
->fqslock). If line 12 sees that either line 8 or 9 saw a reason to give up, line 13
returns to the caller. Otherwise, we must have acquired the current rcu_node structure’s
->fqslock, so line 14 saves a pointer to this structure in local variable rnp_old in
preparation for the next pass through the loop.
If control reaches line 16, we won the tournament, and now holds the root rcu_node
structure’s ->fqslock. If line 16 still sees that the global variable gp_flags is zero,
line 17 sets gp_flags to one, line 18 invokes do_force_quiescent_state(), and
line 19 resets gp_flags back to zero. Either way, line 21 releases the root rcu_node
structure’s ->fqslock.
v2023.06.11a
7.3. LOCKING IMPLEMENTATION ISSUES 181
Listing 7.9: Sample Lock Based on Atomic Exchange

1 typedef int xchglock_t;
2 #define DEFINE_XCHG_LOCK(n) xchglock_t n = 0
3
4 void xchg_lock(xchglock_t *xp)
5 {
6 while (xchg(xp, 1) == 1) {
7 while (READ_ONCE(*xp) == 1)
8 continue;
9 }
10 }
11
12 void xchg_unlock(xchglock_t *xp)
13 {
14 (void)xchg(xp, 0);
15 }
Quick Quiz 7.22: The code in Listing 7.8 is ridiculously complicated! Why not conditionally
acquire a single global lock?
Quick Quiz 7.23: Wait a minute! If we “win” the tournament on line 16 of Listing 7.8, we get
to do all the work of do_force_quiescent_state(). Exactly how is that a win, really?
This function illustrates the not-uncommon pattern of hierarchical locking. This

pattern is difficult to implement using strict RAII locking,7 just like the iterator
encapsulation noted earlier, and so explicit lock/unlock primitives (or C++17-style
unique_lock escapes) will be required for the foreseeable future.
7.3 Locking Implementation Issues

When you translate a dream into reality, it’s never a
full implementation. It is easier to dream than to do.
Shai Agassi
Developers are almost always best-served by using whatever locking primitives are
provided by the system, for example, the POSIX pthread mutex locks [Ope97, But97].
Nevertheless, studying sample implementations can be helpful, as can considering the
challenges posed by extreme workloads and environments.
7.3.1 Sample Exclusive-Locking Implementation Based on Atomic

Exchange
This section reviews the implementation shown in Listing 7.9. The data structure for
this lock is just an int, as shown on line 1, but could be any integral type. The initial
value of this lock is zero, meaning “unlocked”, as shown on line 2.
Quick Quiz 7.24: Why not rely on the C language’s default initialization of zero instead of
using the explicit initializer shown on line 2 of Listing 7.9?
7 Which is why many RAII locking implementations provide a way to leak the lock out of
the scope that it was acquired and into the scope in which it is to be released. However, some
object must mediate the scope leaking, which can add complexity compared to non-RAII
explicit locking primitives.
v2023.06.11a
Lock acquisition is carried out by the xchg_lock() function shown on lines 4–10.
This function uses a nested loop, with the outer loop repeatedly atomically exchanging
the value of the lock with the value one (meaning “locked”). If the old value was already
the value one (in other words, someone else already holds the lock), then the inner loop
(lines 7–8) spins until the lock is available, at which point the outer loop makes another
attempt to acquire the lock.
Quick Quiz 7.25: Why bother with the inner loop on lines 7–8 of Listing 7.9? Why not simply
repeatedly do the atomic exchange operation on line 6?
Lock release is carried out by the xchg_unlock() function shown on lines 12–15.
Line 14 atomically exchanges the value zero (“unlocked”) into the lock, thus marking it
as having been released.
Quick Quiz 7.26: Why not simply store zero into the lock word on line 14 of Listing 7.9?
This lock is a simple example of a test-and-set lock [SR84], but very similar
mechanisms have been used extensively as pure spinlocks in production.
7.3.2 Other Exclusive-Locking Implementations

There are a great many other possible implementations of locking based on atomic
instructions, many of which are reviewed in the classic paper by Mellor-Crummey and
Scott [MCS91]. These implementations represent different points in a multi-dimensional
design tradeoff [GGL+ 19, Gui18, McK96b]. For example, the atomic-exchange-based
test-and-set lock presented in the previous section works well when contention is low
and has the advantage of small memory footprint. It avoids giving the lock to threads
that cannot use it, but as a result can suffer from unfairness or even starvation at high
contention levels.
In contrast, ticket lock [MCS91], which was once used in the Linux kernel, avoids
unfairness at high contention levels. However, as a consequence of its strict FIFO
discipline, it can grant the lock to a thread that is currently unable to use it, perhaps
due to that thread being preempted or interrupted. On the other hand, it is important to
avoid getting too worried about the possibility of preemption and interruption. After all,
in many cases, this preemption and interruption could just as well happen just after the
lock was acquired.8
All locking implementations where waiters spin on a single memory location,
including both test-and-set locks and ticket locks, suffer from performance problems at
high contention levels. The problem is that the thread releasing the lock must update the
value of the corresponding memory location. At low contention, this is not a problem:
The corresponding cache line is very likely still local to and writeable by the thread
holding the lock. In contrast, at high levels of contention, each thread attempting to
acquire the lock will have a read-only copy of the cache line, and the lock holder will
need to invalidate all such copies before it can carry out the update that releases the
lock. In general, the more CPUs and threads there are, the greater the overhead incurred
when releasing the lock under conditions of high contention.
This negative scalability has motivated a number of different queued-lock imple-
mentations [And90, GT90, MCS91, WKS94, Cra93, MLH94, TS93], some of which
8 Besides, the best way of handling high lock contention is to avoid it in the first place!
There are nevertheless some situations where high lock contention is the lesser of the available
evils, and in any case, studying schemes that deal with high levels of contention is a good
mental exercise.
v2023.06.11a
7.3. LOCKING IMPLEMENTATION ISSUES 183
are used in recent versions of the Linux kernel [Cor14b]. Queued locks avoid high
cache-invalidation overhead by assigning each thread a queue element. These queue
elements are linked together into a queue that governs the order that the lock will be
granted to the waiting threads. The key point is that each thread spins on its own queue
element, so that the lock holder need only invalidate the first element from the next
thread’s CPU’s cache. This arrangement greatly reduces the overhead of lock handoff at
high levels of contention.
More recent queued-lock implementations also take the system’s architecture into
account, preferentially granting locks locally, while also taking steps to avoid starva-
tion [SSVM02, RH03, RH02, JMRR02, MCM02]. Many of these can be thought of as
analogous to the elevator algorithms traditionally used in scheduling disk I/O.
Unfortunately, the same scheduling logic that improves the efficiency of queued locks
at high contention also increases their overhead at low contention. Beng-Hong Lim and
Anant Agarwal therefore combined a simple test-and-set lock with a queued lock, using
the test-and-set lock at low levels of contention and switching to the queued lock at high
levels of contention [LA94], thus getting low overhead at low levels of contention and
getting fairness and high throughput at high levels of contention. Browning et al. took
a similar approach, but avoided the use of a separate flag, so that the test-and-set fast
path uses the same sequence of instructions that would be used in a simple test-and-set
lock [BMMM05]. This approach has been used in production.
Another issue that arises at high levels of contention is when the lock holder is delayed,
especially when the delay is due to preemption, which can result in priority inversion,
where a low-priority thread holds a lock, but is preempted by a medium priority
CPU-bound thread, which results in a high-priority process blocking while attempting to
acquire the lock. The result is that the CPU-bound medium-priority process is preventing
the high-priority process from running. One solution is priority inheritance [LR80],
which has been widely used for real-time computing [SRL90, Cor06b], despite some
lingering controversy over this practice [Yod04a, Loc02].
Another way to avoid priority inversion is to prevent preemption while a lock is held.
Because preventing preemption while locks are held also improves throughput, most
proprietary UNIX kernels offer some form of scheduler-conscious synchronization
mechanism [KWS97], largely due to the efforts of a certain sizable database vendor.
These mechanisms usually take the form of a hint that preemption should be avoided in
a given region of code, with this hint typically being placed in a machine register. These
hints frequently take the form of a bit set in a particular machine register, which enables
extremely low per-lock-acquisition overhead for these mechanisms. In contrast, Linux
avoids these hints. Instead, the Linux kernel community’s response to requests for
scheduler-conscious synchronization was a mechanism called futexes [FRK02, Mol06,
Ros06, Dre11].
Interestingly enough, atomic instructions are not strictly needed to implement
locks [Dĳ65, Lam74]. An excellent exposition of the issues surrounding locking
implementations based on simple loads and stores may be found in Herlihy’s and Shavit’s
textbook [HS08, HSLS20]. The main point echoed here is that such implementations
currently have little practical application, although a careful study of them can be both
entertaining and enlightening. Nevertheless, with one exception described below, such
study is left as an exercise for the reader.
Gamsa et al. [GKAS99, Section 5.3] describe a token-based mechanism in which a
token circulates among the CPUs. When the token reaches a given CPU, it has exclusive
v2023.06.11a
access to anything protected by that token. There are any number of schemes that may
be used to implement the token-based mechanism, for example:
1. Maintain a per-CPU flag, which is initially zero for all but one CPU. When a
CPU’s flag is non-zero, it holds the token. When it finishes with the token, it zeroes
its flag and sets the flag of the next CPU to one (or to any other non-zero value).
2. Maintain a per-CPU counter, which is initially set to the corresponding CPU’s
number, which we assume to range from zero to 𝑁 − 1, where 𝑁 is the number
of CPUs in the system. When a CPU’s counter is greater than that of the next
CPU (taking counter wrap into account), the first CPU holds the token. When it is
finished with the token, it sets the next CPU’s counter to a value one greater than
its own counter.
Quick Quiz 7.27: How can you tell if one counter is greater than another, while accounting
for counter wrap?
Quick Quiz 7.28: Which is better, the counter approach or the flag approach?
This lock is unusual in that a given CPU cannot necessarily acquire it immediately,
even if no other CPU is using it at the moment. Instead, the CPU must wait until the
token comes around to it. This is useful in cases where CPUs need periodic access
to the critical section, but can tolerate variances in token-circulation rate. Gamsa et
al. [GKAS99] used it to implement a variant of read-copy update (see Section 9.5), but
it could also be used to protect periodic per-CPU operations such as flushing per-CPU
caches used by memory allocators [MS93], garbage-collecting per-CPU data structures,
or flushing per-CPU data to shared storage (or to mass storage, for that matter).
The Linux kernel now uses queued spinlocks [Cor14b], but because of the complexity
of implementations that provide good performance across the range of contention
levels, the path has not always been smooth [Mar18, Dea18]. As increasing numbers
of people gain familiarity with parallel hardware and parallelize increasing amounts
of code, we can continue to expect more special-purpose locking primitives to appear,
see for example Guerraoui et al. [GGL+ 19, Gui18]. Nevertheless, you should carefully
consider this important safety tip: Use the standard synchronization primitives whenever
humanly possible. The big advantage of the standard synchronization primitives over
roll-your-own efforts is that the standard primitives are typically much less bug-prone.9
7.4 Lock-Based Existence Guarantees

Existence precedes and rules essence.
Jean-Paul Sartre
A key challenge in parallel programming is to provide existence guarantees [GKAS99],

so that attempts to access a given object can rely on that object being in existence
throughout a given access attempt.
In some cases, existence guarantees are implicit:
9 And yes, I have done at least my share of roll-your-own synchronization primitives.
However, you will notice that my hair is much greyer than it was before I started doing that
sort of work. Coincidence? Maybe. But are you really willing to risk your own hair turning
prematurely grey?
v2023.06.11a
7.4. LOCK-BASED EXISTENCE GUARANTEES 185
Listing 7.10: Per-Element Locking Without Existence Guarantees (Buggy!)

1 int delete(int key)
2 {
3 int b;
4 struct element *p;
5
6 b = hashfunction(key);
7 p = hashtable[b];
8 if (p == NULL || p->key != key)
9 return 0;
10 spin_lock(&p->lock);
11 hashtable[b] = NULL;
12 spin_unlock(&p->lock);
13 kfree(p);
14 return 1;
15 }
1. Global variables and static local variables in the base module will exist as long as
the application is running.
2. Global variables and static local variables in a loaded module will exist as long as
that module remains loaded.
3. A module will remain loaded as long as at least one of its functions has an active
instance.
4. A given function instance’s on-stack variables will exist until that instance returns.
5. If you are executing within a given function or have been called (directly or
indirectly) from that function, then the given function has an active instance.
These implicit existence guarantees are straightforward, though bugs involving

implicit existence guarantees really can happen.
Quick Quiz 7.29: How can relying on implicit existence guarantees result in a bug?
But the more interesting—and troublesome—guarantee involves heap memory: A

dynamically allocated data structure will exist until it is freed. The problem to be
solved is to synchronize the freeing of the structure with concurrent accesses to that
same structure. One way to do this is with explicit guarantees, such as locking. If a
given structure may only be freed while holding a given lock, then holding that lock
guarantees that structure’s existence.
But this guarantee depends on the existence of the lock itself. One straightforward
way to guarantee the lock’s existence is to place the lock in a global variable, but global
locking has the disadvantage of limiting scalability. One way of providing scalability
that improves as the size of the data structure increases is to place a lock in each element
of the structure. Unfortunately, putting the lock that is to protect a data element in the
data element itself is subject to subtle race conditions, as shown in Listing 7.10.
Quick Quiz 7.30: What if the element we need to delete is not the first element of the list on
To see one of these race conditions, consider the following sequence of events:
1. Thread 0 invokes delete(0), and reaches line 10 of the listing, acquiring the lock.
2. Thread 1 concurrently invokes delete(0), reaching line 10, but spins on the lock
because Thread 0 holds it.
v2023.06.11a
Listing 7.11: Per-Element Locking With Lock-Based Existence Guarantees

2 {
3 int b;
5 spinlock_t *sp;
6
8 sp = &locktable[b];
9 spin_lock(sp);
10 p = hashtable[b];
11 if (p == NULL || p->key != key) {
12 spin_unlock(sp);
13 return 0;
14 }
15 hashtable[b] = NULL;
16 spin_unlock(sp);
17 kfree(p);
18 return 1;
19 }
3. Thread 0 executes lines 11–14, removing the element from the hashtable, releasing
the lock, and then freeing the element.
4. Thread 0 continues execution, and allocates memory, getting the exact block of
memory that it just freed.
5. Thread 0 then initializes this block of memory as some other type of structure.
6. Thread 1’s spin_lock() operation fails due to the fact that what it believes to be
p->lock is no longer a spinlock.
Because there is no existence guarantee, the identity of the data element can change
while a thread is attempting to acquire that element’s lock on line 10!
One way to fix this example is to use a hashed set of global locks, so that each hash
bucket has its own lock, as shown in Listing 7.11. This approach allows acquiring
the proper lock (on line 9) before gaining a pointer to the data element (on line 10).
Although this approach works quite well for elements contained in a single partitionable
data structure such as the hash table shown in the listing, it can be problematic if a
given data element can be a member of multiple hash tables or given more-complex
data structures such as trees or graphs. Not only can these problems be solved,
but the solutions also form the basis of lock-based software transactional memory
implementations [ST95, DSS06]. However, Chapter 9 describes simpler—and faster—
ways of providing existence guarantees.
7.5 Locking: Hero or Villain?

You either die a hero or you live long enough to see
yourself become the villain.
Aaron Eckhart as Harvey Dent
As is often the case in real life, locking can be either hero or villain, depending on
how it is used and on the problem at hand. In my experience, those writing whole
applications are happy with locking, those writing parallel libraries are less happy, and
those parallelizing existing sequential libraries are extremely unhappy. The following
sections discuss some reasons for these differences in viewpoints.
v2023.06.11a
7.5. LOCKING: HERO OR VILLAIN? 187
7.5.1 Locking For Applications: Hero!

When writing an entire application (or entire kernel), developers have full control of
the design, including the synchronization design. Assuming that the design makes
good use of partitioning, as discussed in Chapter 6, locking can be an extremely
effective synchronization mechanism, as demonstrated by the heavy use of locking in
production-quality parallel software.
Nevertheless, although such software usually bases most of its synchronization design
on locking, such software also almost always makes use of other synchronization
mechanisms, including special counting algorithms (Chapter 5), data ownership (Chap-
ter 8), reference counting (Section 9.2), hazard pointers (Section 9.3), sequence locking
(Section 9.4), and read-copy update (Section 9.5). In addition, practitioners use tools for
deadlock detection [Cor06a], lock acquisition/release balancing [Cor04b], cache-miss
analysis [The11], hardware-counter-based profiling [EGMdB11, The12b], and many
more besides.
Given careful design, use of a good combination of synchronization mechanisms,
and good tooling, locking works quite well for applications and kernels.
7.5.2 Locking For Parallel Libraries: Just Another Tool

Unlike applications and kernels, the designer of a library cannot know the locking
design of the code that the library will be interacting with. In fact, that code might not
be written for years to come. Library designers therefore have less control and must
exercise more care when laying out their synchronization design.
Deadlock is of course of particular concern, and the techniques discussed in Sec-
tion 7.1.1 need to be applied. One popular deadlock-avoidance strategy is therefore
to ensure that the library’s locks are independent subtrees of the enclosing program’s
locking hierarchy. However, this can be harder than it looks.
One complication was discussed in Section 7.1.1.2, namely when library functions
call into application code, with qsort()’s comparison-function argument being a case
in point. Another complication is the interaction with signal handlers. If an application
signal handler is invoked from a signal received within the library function, deadlock
can ensue just as surely as if the library function had called the signal handler directly.
A final complication occurs for those library functions that can be used between a
fork()/exec() pair, for example, due to use of the system() function. In this case,
if your library function was holding a lock at the time of the fork(), then the child
process will begin life with that lock held. Because the thread that will release the
lock is running in the parent but not the child, if the child calls your library function,
deadlock will ensue.
The following strategies may be used to avoid deadlock problems in these cases:
1. Don’t use either callbacks or signals.
2. Don’t acquire locks from within callbacks or signal handlers.
3. Let the caller control synchronization.
4. Parameterize the library API to delegate locking to caller.
5. Explicitly avoid callback deadlocks.
6. Explicitly avoid signal-handler deadlocks.
v2023.06.11a
7. Avoid invoking fork().
Each of these strategies is discussed in one of the following sections.
7.5.2.1 Use Neither Callbacks Nor Signals

If a library function avoids callbacks and the application as a whole avoids signals,
then any locks acquired by that library function will be leaves of the locking-hierarchy
tree. This arrangement avoids deadlock, as discussed in Section 7.1.1.1. Although this
strategy works extremely well where it applies, there are some applications that must
use signal handlers, and there are some library functions (such as the qsort() function
discussed in Section 7.1.1.2) that require callbacks.
The strategy described in the next section can often be used in these cases.
7.5.2.2 Avoid Locking in Callbacks and Signal Handlers

If neither callbacks nor signal handlers acquire locks, then they cannot be involved
in deadlock cycles, which allows straightforward locking hierarchies to once again
consider library functions to be leaves on the locking-hierarchy tree. This strategy
works very well for most uses of qsort, whose callbacks usually simply compare the
two values passed in to them. This strategy also works wonderfully for many signal
handlers, especially given that acquiring locks from within signal handlers is generally
frowned upon [Gro01],10 but can fail if the application needs to manipulate complex
data structures from a signal handler.
Here are some ways to avoid acquiring locks in signal handlers even if complex data
structures must be manipulated:
1. Use simple data structures based on non-blocking synchronization, as will be

discussed in Section 14.2.1.
2. If the data structures are too complex for reasonable use of non-blocking syn-
chronization, create a queue that allows non-blocking enqueue operations. In
the signal handler, instead of manipulating the complex data structure, add an
element to the queue describing the required change. A separate thread can then
remove elements from the queue and carry out the required changes using normal
locking. There are a number of readily available implementations of concurrent
queues [KLP12, Des09b, MS96].
This strategy should be enforced with occasional manual or (preferably) automated

inspections of callbacks and signal handlers. When carrying out these inspections, be
wary of clever coders who might have (unwisely) created home-brew locks from atomic
operations.
7.5.2.3 Caller Controls Synchronization

Letting the caller control synchronization works extremely well when the library
functions are operating on independent caller-visible instances of a data structure, each
of which may be synchronized separately. For example, if the library functions operate
on a search tree, and if the application needs a large number of independent search trees,
10 But the standard’s words do not stop clever coders from creating their own home-brew
locking primitives from atomic operations.
v2023.06.11a
then the application can associate a lock with each tree. The application then acquires
and releases locks as needed, so that the library need not be aware of parallelism at all.
Instead, the application controls the parallelism, so that locking can work very well, as
was discussed in Section 7.5.1.
However, this strategy fails if the library implements a data structure that requires
internal concurrency, for example, a hash table or a parallel sort. In this case, the library
absolutely must control its own synchronization.
7.5.2.4 Parameterize Library Synchronization

The idea here is to add arguments to the library’s API to specify which locks to acquire,
how to acquire and release them, or both. This strategy allows the application to take on
the global task of avoiding deadlock by specifying which locks to acquire (by passing in
pointers to the locks in question) and how to acquire them (by passing in pointers to lock
acquisition and release functions), but also allows a given library function to control its
own concurrency by deciding where the locks should be acquired and released.
In particular, this strategy allows the lock acquisition and release functions to block
signals as needed without the library code needing to be concerned with which signals
need to be blocked by which locks. The separation of concerns used by this strategy can
be quite effective, but in some cases the strategies laid out in the following sections can
work better.
That said, passing explicit pointers to locks to external APIs must be very carefully
considered, as discussed in Section 7.1.1.5. Although this practice is sometimes the
right thing to do, you should do yourself a favor by looking into alternative designs first.
7.5.2.5 Explicitly Avoid Callback Deadlocks

The basic rule behind this strategy was discussed in Section 7.1.1.2: “Release all locks
before invoking unknown code.” This is usually the best approach because it allows
the application to ignore the library’s locking hierarchy: The library remains a leaf or
isolated subtree of the application’s overall locking hierarchy.
In cases where it is not possible to release all locks before invoking unknown code,
the layered locking hierarchies described in Section 7.1.1.3 can work well. For example,
if the unknown code is a signal handler, this implies that the library function block
signals across all lock acquisitions, which can be complex and slow. Therefore, in
cases where signal handlers (probably unwisely) acquire locks, the strategies in the next
section may prove helpful.
7.5.2.6 Explicitly Avoid Signal-Handler Deadlocks

Suppose that a given library function is known to acquire locks, but does not block
signals. Suppose further that it is necessary to invoke that function both from within
and outside of a signal handler, and that it is not permissible to modify this library
function. Of course, if no special action is taken, then if a signal arrives while that
library function is holding its lock, deadlock can occur when the signal handler invokes
that same library function, which in turn attempts to re-acquire that same lock.
Such deadlocks can be avoided as follows:
1. If the application invokes the library function from within a signal handler, then
that signal must be blocked every time that the library function is invoked from
outside of a signal handler.
v2023.06.11a
2. If the application invokes the library function while holding a lock acquired within
a given signal handler, then that signal must be blocked every time that the library
function is called outside of a signal handler.
These rules can be enforced by using tools similar to the Linux kernel’s lockdep lock
dependency checker [Cor06a]. One of the great strengths of lockdep is that it is not
fooled by human intuition [Ros11].
7.5.2.7 Library Functions Used Between fork() and exec()

As noted earlier, if a thread executing a library function is holding a lock at the time that
some other thread invokes fork(), the fact that the parent’s memory is copied to create
the child means that this lock will be born held in the child’s context. The thread that
will release this lock is running in the parent, but not in the child, which means that
although the parent’s copy of this lock will be released, the child’s copy never will be.
Therefore, any attempt on the part of the child to invoke that same library function (thus
acquiring that same lock) will result in deadlock.
A pragmatic and straightforward way of solving this problem is to fork() a child
process while the process is still single-threaded, and have this child process remain
single-threaded. Requests to create further child processes can then be communicated
to this initial child process, which can safely carry out any needed fork() and exec()
system calls on behalf of its multi-threaded parent process.
Another rather less pragmatic and straightforward solution to this problem is to have
the library function check to see if the owner of the lock is still running, and if not,
“breaking” the lock by re-initializing and then acquiring it. However, this approach has
a couple of vulnerabilities:
1. The data structures protected by that lock are likely to be in some intermediate
state, so that naively breaking the lock might result in arbitrary memory corruption.
2. If the child creates additional threads, two threads might break the lock concurrently,
with the result that both threads believe they own the lock. This could again result
in arbitrary memory corruption.
The pthread_atfork() function is provided to help deal with these situations. The
idea is to register a triplet of functions, one to be called by the parent before the fork(),
one to be called by the parent after the fork(), and one to be called by the child after
the fork(). Appropriate cleanups can then be carried out at these three points.
Be warned, however, that coding of pthread_atfork() handlers is quite subtle in
general. The cases where pthread_atfork() works best are cases where the data
structure in question can simply be re-initialized by the child. Which might be one
reason why the POSIX standard forbids use of any non-async-signal-safe functions
between the fork() and the exec(), which rules out acquisition of locks during that
time.
Other alternatives to fork()/exec() include posix_spawn() and io_uring_
spawn() [Tri22, Edg22].
7.5.2.8 Parallel Libraries: Discussion

Regardless of the strategy used, the description of the library’s API must include a clear
description of that strategy and how the caller should interact with that strategy. In short,
v2023.06.11a
constructing parallel libraries using locking is possible, but not as easy as constructing
a parallel application.
7.5.3 Locking For Parallelizing Sequential Libraries: Villain!

With the advent of readily available low-cost multicore systems, a common task is
parallelizing an existing library that was designed with only single-threaded use in mind.
This all-too-common disregard for parallelism can result in a library API that is severely
flawed from a parallel-programming viewpoint. Candidate flaws include:
1. Implicit prohibition of partitioning.

2. Callback functions requiring locking.
3. Object-oriented spaghetti code.
These flaws and the consequences for locking are discussed in the following sections.
7.5.3.1 Partitioning Prohibited

Suppose that you were writing a single-threaded hash-table implementation. It is easy
and fast to maintain an exact count of the total number of items in the hash table, and
also easy and fast to return this exact count on each addition and deletion operation. So
why not?
One reason is that exact counters do not perform or scale well on multicore systems,
as was seen in Chapter 5. As a result, the parallelized implementation of the hash table
will not perform or scale well.
So what can be done about this? One approach is to return an approximate count,
using one of the algorithms from Chapter 5. Another approach is to drop the element
count altogether.
Either way, it will be necessary to inspect uses of the hash table to see why the
addition and deletion operations need the exact count. Here are a few possibilities:
1. Determining when to resize the hash table. In this case, an approximate count
should work quite well. It might also be useful to trigger the resizing operation
from the length of the longest chain, which can be computed and maintained in a
nicely partitioned per-chain manner.
2. Producing an estimate of the time required to traverse the entire hash table. An
approximate count works well in this case, also.
3. For diagnostic purposes, for example, to check for items being lost when transferring
them to and from the hash table. This clearly requires an exact count. However,
given that this usage is diagnostic in nature, it might suffice to maintain the lengths
of the hash chains, then to infrequently sum them up while locking out addition
and deletion operations.
It turns out that there is now a strong theoretical basis for some of the constraints that
performance and scalability place on a parallel library’s APIs [AGH+ 11a, AGH+ 11b,
McK11b]. Anyone designing a parallel library needs to pay close attention to those
constraints.
Although it is all too easy to blame locking for what are really problems due to a
concurrency-unfriendly API, doing so is not helpful. On the other hand, one has little
v2023.06.11a
choice but to sympathize with the hapless developer who made this choice in (say)
1985. It would have been a rare and courageous developer to anticipate the need for
parallelism at that time, and it would have required an even more rare combination of
brilliance and luck to actually arrive at a good parallel-friendly API.
Times change, and code must change with them. That said, there might be a huge
number of users of a popular library, in which case an incompatible change to the API
would be quite foolish. Adding a parallel-friendly API to complement the existing
heavily used sequential-only API is usually the best course of action.
Nevertheless, human nature being what it is, we can expect our hapless developer
to be more likely to complain about locking than about his or her own poor (though
understandable) API design choices.
7.5.3.2 Deadlock-Prone Callbacks

Sections 7.1.1.2, 7.1.1.3, and 7.5.2 described how undisciplined use of callbacks can
result in locking woes. These sections also described how to design your library function
to avoid these problems, but it is unrealistic to expect a 1990s programmer with no
experience in parallel programming to have followed such a design. Therefore, someone
attempting to parallelize an existing callback-heavy single-threaded library will likely
have many opportunities to curse locking’s villainy.
If there are a very large number of uses of a callback-heavy library, it may be wise to
again add a parallel-friendly API to the library in order to allow existing users to convert
their code incrementally. Alternatively, some advocate use of transactional memory in
these cases. While the jury is still out on transactional memory, Section 17.2 discusses
its strengths and weaknesses. It is important to note that hardware transactional memory
(discussed in Section 17.3) cannot help here unless the hardware transactional memory
implementation provides forward-progress guarantees, which few do. Other alternatives
that appear to be quite practical (if less heavily hyped) include the methods discussed in
Sections 7.1.1.6 and 7.1.1.7, as well as those that will be discussed in Chapters 8 and 9.
7.5.3.3 Object-Oriented Spaghetti Code

Object-oriented programming went mainstream sometime in the 1980s or 1990s, and as
a result there is a huge amount of single-threaded object-oriented code in production.
Although object orientation can be a valuable software technique, undisciplined use of
objects can easily result in object-oriented spaghetti code. In object-oriented spaghetti
code, control flits from object to object in an essentially random manner, making the
code hard to understand and even harder, and perhaps impossible, to accommodate a
locking hierarchy.
Although many might argue that such code should be cleaned up in any case, such
things are much easier to say than to do. If you are tasked with parallelizing such a beast,
you can reduce the number of opportunities to curse locking by using the techniques
described in Sections 7.1.1.6 and 7.1.1.7, as well as those that will be discussed in
Chapters 8 and 9. This situation appears to be the use case that inspired transactional
memory, so it might be worth a try as well. That said, the choice of synchronization
mechanism should be made in light of the hardware habits discussed in Chapter 3. After
all, if the overhead of the synchronization mechanism is orders of magnitude more than
that of the operations being protected, the results are not going to be pretty.
And that leads to a question well worth asking in these situations: Should the code
remain sequential? For example, perhaps parallelism should be introduced at the process
v2023.06.11a
7.6. SUMMARY 193
level rather than the thread level. In general, if a task is proving extremely hard, it is
worth some time spent thinking about not only alternative ways to accomplish that
particular task, but also alternative tasks that might better solve the problem at hand.
7.6 Summary
Achievement unlocked.
Unknown
Locking is perhaps the most widely used and most generally useful synchronization
tool. However, it works best when designed into an application or library from the
beginning. Given the large quantity of pre-existing single-threaded code that might
need to one day run in parallel, locking should therefore not be the only tool in your
parallel-programming toolbox. The next few chapters will discuss other tools, and how
they can best be used in concert with locking and with each other.
v2023.06.11a
v2023.06.11a
It is mine, I tell you. My own. My precious. Yes, my
precious.
Gollum in The Fellowship of the Ring, J.R.R. Tolkien
Chapter 8
Data Ownership
One of the simplest ways to avoid the synchronization overhead that comes with locking
is to parcel the data out among the threads (or, in the case of kernels, CPUs) so that a
given piece of data is accessed and modified by only one of the threads. Interestingly
enough, data ownership covers each of the “big three” parallel design techniques: It
partitions over threads (or CPUs, as the case may be), it batches all local operations, and
its elimination of synchronization operations is weakening carried to its logical extreme.
It should therefore be no surprise that data ownership is heavily used: Even novices use
it almost instinctively. In fact, it is so heavily used that this chapter will not introduce
any new examples, but will instead refer back to those of previous chapters.
Quick Quiz 8.1: What form of data ownership is extremely difficult to avoid when creating
shared-memory parallel programs (for example, using pthreads) in C or C++?
There are a number of approaches to data ownership. Section 8.1 presents the
logical extreme in data ownership, where each thread has its own private address space.
Section 8.2 looks at the opposite extreme, where the data is shared, but different threads
own different access rights to the data. Section 8.3 describes function shipping, which
is a way of allowing other threads to have indirect access to data owned by a particular
thread. Section 8.4 describes how designated threads can be assigned ownership of a
specified function and the related data. Section 8.5 discusses improving performance
by transforming algorithms with shared data to instead use data ownership. Finally,
Section 8.6 lists a few software environments that feature data ownership as a first-class
citizen.
8.1 Multiple Processes

A man’s home is his castle
Ancient Laws of England
Section 4.1 introduced the following example:

3 wait
195
v2023.06.11a
196 CHAPTER 8. DATA OWNERSHIP
This example runs two instances of the compute_it program in parallel, as separate
processes that do not share memory. Therefore, all data in a given process is owned by that
process, so that almost the entirety of data in the above example is owned. This approach
almost entirely eliminates synchronization overhead. The resulting combination of
extreme simplicity and optimal performance is obviously quite attractive.
Quick Quiz 8.2: What synchronization remains in the example shown in Section 8.1?
Quick Quiz 8.3: Is there any shared data in the example shown in Section 8.1?
This same pattern can be written in C as well as in sh, as illustrated by Listings 4.1
and 4.2.
It bears repeating that these trivial forms of parallelism are not in any way cheating
or ducking responsibility, but are rather simple and elegant ways to make your code
run faster. It is fast, scales well, is easy to program, easy to maintain, and gets the job
done. In addition, taking this approach (where applicable) allows the developer more
time to focus on other things whether these things might involve applying sophisticated
single-threaded optimizations to compute_it on the one hand, or applying sophisticated
parallel-programming patterns to portions of the code where this approach is inapplicable.
What is not to like?
The next section discusses the use of data ownership in shared-memory parallel
programs.
8.2 Partial Data Ownership and pthreads
Give thy mind more to what thou hast than to what

thou hast not.
Marcus Aurelius Antoninus
Concurrent counting (see Chapter 5) uses data ownership heavily, but adds a twist.
Threads are not allowed to modify data owned by other threads, but they are permitted to
read it. In short, the use of shared memory allows more nuanced notions of ownership
and access rights.
For example, consider the per-thread statistical counter implementation shown in
Listing 5.4 on page 83. Here, inc_count() updates only the corresponding thread’s
instance of counter, while read_count() accesses, but does not modify, all threads’
instances of counter.
Quick Quiz 8.4: Does it ever make sense to have partial data ownership where each thread
reads only its own instance of a per-thread variable, but writes to other threads’ instances?
Partial data ownership is also common within the Linux kernel. For example, a given
CPU might be permitted to read a given set of its own per-CPU variables only with
interrupts disabled, another CPU might be permitted to read that same set of the first
CPU’s per-CPU variables only when holding the corresponding per-CPU lock. Then
that given CPU would be permitted to update this set of its own per-CPU variables if
it both has interrupts disabled and holds its per-CPU lock. This arrangement can be
thought of as a reader-writer lock that allows each CPU very low-overhead access to its
own set of per-CPU variables. There are a great many variations on this theme.
v2023.06.11a
8.3. FUNCTION SHIPPING 197
For its own part, pure data ownership is also both common and useful, for example,
the per-thread memory-allocator caches discussed in Section 6.4.3 starting on page 141.
In this algorithm, each thread’s cache is completely private to that thread.
8.3 Function Shipping
If the mountain will not come to Muhammad, then

Muhammad must go to the mountain.
Essays, Francis Bacon
The previous section described a weak form of data ownership where threads reached
out to other threads’ data. This can be thought of as bringing the data to the functions
that need it. An alternative approach is to send the functions to the data.
Such an approach is illustrated in Section 5.4.3 beginning on page 101, in particular the
flush_local_count_sig() and flush_local_count() functions in Listing 5.18
on page 103.
The flush_local_count_sig() function is a signal handler that acts as the shipped
function. The pthread_kill() function in flush_local_count() sends the signal—
shipping the function—and then waits until the shipped function executes. This shipped
function has the not-unusual added complication of needing to interact with any
concurrently executing add_count() or sub_count() functions (see Listing 5.19 on
page 104 and Listing 5.20 on page 105).
Quick Quiz 8.5: What mechanisms other than POSIX signals may be used for function
shipping?
8.4 Designated Thread
Let a man practice the profession which he best

knows.
Cicero
The earlier sections describe ways of allowing each thread to keep its own copy or its
own portion of the data. In contrast, this section describes a functional-decomposition
approach, where a special designated thread owns the rights to the data that is required to
do its job. The eventually consistent counter implementation described in Section 5.2.4
provides an example. This implementation has a designated thread that runs the
eventual() function shown on lines 17–32 of Listing 5.5. This eventual() thread
periodically pulls the per-thread counts into the global counter, so that accesses to the
global counter will, as the name says, eventually converge on the actual value.
Quick Quiz 8.6: But none of the data in the eventual() function shown on lines 17–32
of Listing 5.5 is actually owned by the eventual() thread! In just what way is this data
ownership???
v2023.06.11a
8.5 Privatization
There is, of course, a difference between what a man

seizes and what he really possesses.
Pearl S. Buck
One way of improving the performance and scalability of a shared-memory parallel

program is to transform it so as to convert shared data to private data that is owned by a
particular thread.
An excellent example of this is shown in the answer to one of the Quick Quizzes in
Section 6.1.1, which uses privatization to produce a solution to the Dining Philosophers
problem with much better performance and scalability than that of the standard textbook
solution. The original problem has five philosophers sitting around the table with one
fork between each adjacent pair of philosophers, which permits at most two philosophers
to eat concurrently.
We can trivially privatize this problem by providing an additional five forks, so
that each philosopher has his or her own private pair of forks. This allows all five
philosophers to eat concurrently, and also offers a considerable reduction in the spread
of certain types of disease.
In other cases, privatization imposes costs. For example, consider the simple limit
counter shown in Listing 5.7 on page 90. This is an example of an algorithm where
threads can read each others’ data, but are only permitted to update their own data.
A quick review of the algorithm shows that the only cross-thread accesses are in
the summation loop in read_count(). If this loop is eliminated, we move to the
more-efficient pure data ownership, but at the cost of a less-accurate result from
read_count().
Quick Quiz 8.7: Is it possible to obtain greater accuracy while still maintaining full privacy of
the per-thread data?
Partial privatization is also possible, with some synchronization requirements, but

less than in the fully shared case. Some partial-privatization possibilities were explored
in Section 4.3.4.4. Chapter 9 will introduce a temporal component to data ownership by
providing ways of safely taking public data structures private.
In short, privatization is a powerful tool in the parallel programmer’s toolbox, but it
must nevertheless be used with care. Just like every other synchronization primitive, it
has the potential to increase complexity while decreasing performance and scalability.
8.6 Other Uses of Data Ownership
Everything comes to us that belongs to us if we

create the capacity to receive it.
Rabindranath Tagore
Data ownership works best when the data can be partitioned so that there is little or no
need for cross thread access or update. Fortunately, this situation is reasonably common,
and in a wide variety of parallel-programming environments.
Examples of data ownership include:
v2023.06.11a
8.6. OTHER USES OF DATA OWNERSHIP 199
1. All message-passing environments, such as MPI [MPI08] and BOINC [Uni08a].

2. Map-reduce [Jac08].
3. Client-server systems, including RPC, web services, and pretty much any system
with a back-end database server.
4. Shared-nothing database systems.
5. Fork-join systems with separate per-process address spaces.
6. Process-based parallelism, such as the Erlang language.
7. Private variables, for example, C-language on-stack auto variables, in threaded

environments.
8. Many parallel linear-algebra algorithms, especially those well-suited for GPGPUs.1
9. Operating-system kernels adapted for networking, where each connection (also
called flow [DKS89, Zha89, McK90]) is assigned to a specific thread. One recent
example of this approach is the IX operating system [BPP+ 16]. IX does have some
shared data structures, which use synchronization mechanisms to be described in
Section 9.5.
Data ownership is perhaps the most underappreciated synchronization mechanism

in existence. When used properly, it delivers unrivaled simplicity, performance, and
scalability. Perhaps its simplicity costs it the respect that it deserves. Hopefully a greater
appreciation for the subtlety and power of data ownership will lead to greater level of
respect, to say nothing of leading to greater performance and scalability coupled with
reduced complexity.
1 But note that a great many other classes of applications have also been ported to
GPGPUs [Mat17, AMD20, NVi17a, NVi17b].
v2023.06.11a
v2023.06.11a
All things come to those who wait.
Violet Fane
Chapter 9
Deferred Processing
The strategy of deferring work goes back before the dawn of recorded history. It has
occasionally been derided as procrastination or even as sheer laziness. However, in
the last few decades workers have recognized this strategy’s value in simplifying and
streamlining parallel algorithms [KL80, Mas92]. Believe it or not, “laziness” in parallel
programming often outperforms and out-scales industriousness! These performance
and scalability benefits stem from the fact that deferring work can enable weakening of
synchronization primitives, thereby reducing synchronization overhead.
Those who are willing and able to read and understand this chapter will uncover many
mysteries, including:
1. The reference-counting trap that awaits unwary developers of concurrent code.
2. A concurrent reference counter that avoids not only this trap, but also avoids
expensive atomic read-modify-write accesses, and in addition avoids as well as
writes of any kind to the data structure being traversed.
3. The under-appreciated restricted form of software transactional memory that is

used heavily within the Linux kernel.
4. A synchronization primitive that allows a concurrently updated linked data structure

to be traversed using exactly the same sequence of machine instructions that might
be used to traverse a sequential implementation of that same data structure.
5. A synchronization primitive whose use cases are far more conceptually more
complex than is the primitive itself.
6. How to choose among the various deferred-processing primitives.
General approaches of work deferral include reference counting (Section 9.2), hazard
pointers (Section 9.3), sequence locking (Section 9.4), and RCU (Section 9.5). Finally,
Section 9.6 describes how to choose among the work-deferral schemes covered in this
chapter and Section 9.7 discusses updates. But first, Section 9.1 will introduce an
example algorithm that will be used to compare and contrast these approaches.
201
v2023.06.11a
202 CHAPTER 9. DEFERRED PROCESSING
route_list
->addr=42 ->addr=56 ->addr=17

->iface=1 ->iface=3 ->iface=7
Figure 9.1: Pre-BSD Packet Routing List
9.1 Running Example

An ounce of application is worth a ton of abstraction.
Booker T. Washington
This chapter will use a simplified packet-routing algorithm to demonstrate the value
of these approaches and to allow them to be compared. Routing algorithms are used
in operating-system kernels to deliver each outgoing TCP/IP packet to the appropriate
network interface. This particular algorithm is a simplified version of the classic 1980s
packet-train-optimized algorithm used in BSD UNIX [Jac88], consisting of a simple
linked list.1 Modern routing algorithms use more complex data structures, however a
simple algorithm will help highlight issues specific to parallelism in a straightforward
setting.
We further simplify the algorithm by reducing the search key from a quadruple
consisting of source and destination IP addresses and ports all the way down to a simple
integer. The value looked up and returned will also be a simple integer, so that the data
structure is as shown in Figure 9.1, which directs packets with address 42 to interface 1,
address 56 to interface 3, and address 17 to interface 7. This list will normally be
searched frequently and updated rarely. In Chapter 3 we learned that the best ways to
evade inconvenient laws of physics, such as the finite speed of light and the atomic
nature of matter, is to either partition the data or to rely on read-mostly sharing. This
chapter applies read-mostly sharing techniques to Pre-BSD packet routing.
Listing 9.1 (route_seq.c) shows a simple single-threaded implementation corre-
sponding to Figure 9.1. Lines 1–5 define a route_entry structure and line 6 defines
the route_list header. Lines 8–20 define route_lookup(), which sequentially
searches route_list, returning the corresponding ->iface, or ULONG_MAX if there is
no such route entry. Lines 22–33 define route_add(), which allocates a route_entry
structure, initializes it, and adds it to the list, returning -ENOMEM in case of memory-
allocation failure. Finally, lines 35–47 define route_del(), which removes and frees
the specified route_entry structure if it exists, or returns -ENOENT otherwise.
This single-threaded implementation serves as a prototype for the various concurrent
implementations in this chapter, and also as an estimate of ideal scalability and
performance.
1 In other words, this is not OpenBSD, NetBSD, or even FreeBSD, but none other than
Pre-BSD.
v2023.06.11a
9.1. RUNNING EXAMPLE 203
Listing 9.1: Sequential Pre-BSD Routing Table

1 struct route_entry {
2 struct cds_list_head re_next;
3 unsigned long addr;
4 unsigned long iface;
5 };
6 CDS_LIST_HEAD(route_list);
7
8 unsigned long route_lookup(unsigned long addr)
9 {
10 struct route_entry *rep;
11 unsigned long ret;
12
13 cds_list_for_each_entry(rep, &route_list, re_next) {
14 if (rep->addr == addr) {
15 ret = rep->iface;
16 return ret;
17 }
18 }
19 return ULONG_MAX;
20 }
21
22 int route_add(unsigned long addr, unsigned long interface)
23 {
25
26 rep = malloc(sizeof(*rep));
27 if (!rep)
28 return -ENOMEM;
29 rep->addr = addr;
30 rep->iface = interface;
31 cds_list_add(&rep->re_next, &route_list);
32 return 0;
33 }
34
35 int route_del(unsigned long addr)
36 {
38
41 cds_list_del(&rep->re_next);
42 free(rep);
43 return 0;
44 }
45 }
46 return -ENOENT;
47 }
v2023.06.11a
9.2 Reference Counting

I am never letting you go!
Unknown
Reference counting tracks the number of references to a given object in order to prevent
that object from being prematurely freed. As such, it has a long and honorable history
of use dating back to at least an early 1960s Weizenbaum paper [Wei63]. Weizenbaum
discusses reference counting as if it was already well-known, so it likely dates back to
the 1950s or even to the 1940s. And perhaps even further, given that people repairing
large dangerous machines have long used a mechanical reference-counting technique
implemented via padlocks. Before entering the machine, each worker locks a padlock
onto the machine’s on/off switch, thus preventing the machine from being powered
on while that worker is inside. Reference counting is thus an excellent time-honored
candidate for a concurrent implementation of Pre-BSD routing.
To that end, Listing 9.2 shows data structures and the route_lookup() function
and Listing 9.3 shows the route_add() and route_del() functions (all at route_
refcnt.c). Since these algorithms are quite similar to the sequential algorithm shown
in Listing 9.1, only the differences will be discussed.
Starting with Listing 9.2, line 2 adds the actual reference counter, line 6 adds
a ->re_freed use-after-free check field, line 9 adds the routelock that will be
used to synchronize concurrent updates, and lines 11–15 add re_free(), which sets
->re_freed, enabling route_lookup() to check for use-after-free bugs. In route_
lookup() itself, lines 29–30 release the reference count of the prior element and free
it if the count becomes zero, and lines 34–42 acquire a reference on the new element,
with lines 35 and 36 performing the use-after-free check.
Quick Quiz 9.1: Why bother with a use-after-free check?
In Listing 9.3, lines 11, 15, 24, 32, and 39 introduce locking to synchronize concurrent
updates. Line 13 initializes the ->re_freed use-after-free-check field, and finally
lines 33–34 invoke re_free() if the new value of the reference count is zero.
Quick Quiz 9.2: Why doesn’t route_del() in Listing 9.3 use reference counts to protect the
traversal to the element to be freed?
Figure 9.2 shows the performance and scalability of reference counting on a read-
only workload with a ten-element list running on an eight-socket 28-core-per-socket
hyperthreaded 2.1 GHz x86 system with a total of 448 hardware threads (hps.2019.
12.02a/lscpu.hps). The “ideal” trace was generated by running the sequential code
shown in Listing 9.1, which works only because this is a read-only workload. The
reference-counting performance is abysmal and its scalability even more so, with the
“refcnt” trace indistinguishable from the x-axis. This should be no surprise in view of
Chapter 3: The reference-count acquisitions and releases have added frequent shared-
memory writes to an otherwise read-only workload, thus incurring severe retribution
from the laws of physics. As well it should, given that all the wishful thinking in the
world is not going to increase the speed of light or decrease the size of the atoms used
in modern digital electronics.
Quick Quiz 9.3: Why the break in the “ideal” line at 224 CPUs in Figure 9.2? Shouldn’t it be
a straight line?
v2023.06.11a
9.2. REFERENCE COUNTING 205
Listing 9.2: Reference-Counted Pre-BSD Routing Table Lookup (BUGGY!!!)

2 atomic_t re_refcnt;
3 struct route_entry *re_next;
6 int re_freed;
7 };
8 struct route_entry route_list;
9 DEFINE_SPINLOCK(routelock);
10
11 static void re_free(struct route_entry *rep)
12 {
13 WRITE_ONCE(rep->re_freed, 1);
14 free(rep);
15 }
16
18 {
19 int old;
20 int new;
22 struct route_entry **repp;
24
25 retry:
26 repp = &route_list.re_next;
27 rep = NULL;
28 do {
29 if (rep && atomic_dec_and_test(&rep->re_refcnt))
30 re_free(rep);
31 rep = READ_ONCE(*repp);
32 if (rep == NULL)
34 do {
35 if (READ_ONCE(rep->re_freed))
36 abort();
37 old = atomic_read(&rep->re_refcnt);
38 if (old <= 0)
39 goto retry;
40 new = old + 1;
41 } while (atomic_cmpxchg(&rep->re_refcnt,
43 repp = &rep->re_next;
44 } while (rep->addr != addr);
46 if (atomic_dec_and_test(&rep->re_refcnt))
47 re_free(rep);
48 return ret;
49 }
2.5x107
7
2x10
Lookups per Millisecond
ideal
7
1.5x10
1x107
5x106
refcnt
0
0 50 100 150 200 250 300 350 400 450
Figure 9.2: Pre-BSD Routing Table Protected by Reference Counting
v2023.06.11a
Listing 9.3: Reference-Counted Pre-BSD Routing Table Add/Delete (BUGGY!!!)

2 {
4
6 if (!rep)
7 return -ENOMEM;
8 atomic_set(&rep->re_refcnt, 1);
9 rep->addr = addr;
11 spin_lock(&routelock);
12 rep->re_next = route_list.re_next;
13 rep->re_freed = 0;
14 route_list.re_next = rep;
15 spin_unlock(&routelock);
16 return 0;
17 }
18
20 {
23
26 for (;;) {
27 rep = *repp;
28 if (rep == NULL)
29 break;
31 *repp = rep->re_next;
33 if (atomic_dec_and_test(&rep->re_refcnt))
34 re_free(rep);
35 return 0;
36 }
38 }
40 return -ENOENT;
41 }
Quick Quiz 9.4: Shouldn’t the refcnt trace in Figure 9.2 be at least a little bit off of the
x-axis???
But it gets worse.

Running multiple updater threads repeatedly invoking route_add() and route_
del() will quickly encounter the abort() statement on line 36 of Listing 9.2, which
indicates a use-after-free bug. This in turn means that the reference counts are not only
profoundly degrading scalability and performance, but also failing to provide the needed
protection.
One sequence of events leading to the use-after-free bug is as follows, given the list
shown in Figure 9.1:
1. Thread A looks up address 42, reaching line 32 of route_lookup() in Listing 9.2.

In other words, Thread A has a pointer to the first element, but has not yet acquired
a reference to it.
2. Thread B invokes route_del() in Listing 9.3 to delete the route entry for
address 42. It completes successfully, and because this entry’s ->re_refcnt field
was equal to the value one, it invokes re_free() to set the ->re_freed field and
to free the entry.
v2023.06.11a
9.3. HAZARD POINTERS 207
3. Thread A continues execution of route_lookup(). Its rep pointer is non-NULL,

but line 35 sees that its ->re_freed field is non-zero, so line 36 invokes abort().
The problem is that the reference count is located in the object to be protected, but
that means that there is no protection during the instant in time when the reference
count itself is being acquired! This is the reference-counting counterpart of a locking
issue noted by Gamsa et al. [GKAS99]. One could imagine using a global lock or
reference count to protect the per-route-entry reference-count acquisition, but this
would result in severe contention issues. Although algorithms exist that allow safe
reference-count acquisition in a concurrent environment [Val95], they are not only
extremely complex and error-prone [MS95], but also provide terrible performance and
scalability [HMBW07].
In short, concurrency has most definitely reduced the usefulness of reference counting!
Of course, as with other synchronization primitives, reference counts also have well-
known ease-of-use shortcomings. These can result in memory leaks on the one hand or
premature freeing on the other.
And this is the reference-counting trap that awaits unwary developers of concurrent
code, noted back on page 201.
Quick Quiz 9.5: If concurrency has “most definitely reduced the usefulness of reference
counting”, why are there so many reference counters in the Linux kernel?
It is sometimes helpful to look at a problem in an entirely different way in order to

successfully solve it. To this end, the next section describes what could be thought of as
an inside-out reference count that provides decent performance and scalability.
9.3 Hazard Pointers

If in doubt, turn it inside out.
Zara Carpenter
One way of avoiding problems with concurrent reference counting is to implement the
reference counters inside out, that is, rather than incrementing an integer stored in the
data element, instead store a pointer to that data element in per-CPU (or per-thread)
lists. Each element of these lists is called a hazard pointer [Mic04a].2 The value of a
given data element’s “virtual reference counter” can then be obtained by counting the
number of hazard pointers referencing that element. Therefore, if that element has been
rendered inaccessible to readers, and there are no longer any hazard pointers referencing
it, that element may safely be freed.
Of course, this means that hazard-pointer acquisition must be carried out quite carefully
in order to avoid destructive races with concurrent deletion. One implementation is
shown in Listing 9.4, which shows hp_try_record() on lines 1–16, hp_record()
on lines 18–27, and hp_clear() on lines 29–33 (hazptr.h).
The hp_try_record() macro on line 16 is simply a casting wrapper for the _h_t_
r_impl() function, which attempts to store the pointer referenced by p into the hazard
pointer referenced by hp. If successful, it returns the value of the stored pointer. If it
fails due to that pointer being NULL, it returns NULL. Finally, if it fails due to racing
with an update, it returns a special HAZPTR_POISON token.
2 Also independently invented by others [HLM02].
v2023.06.11a
Listing 9.4: Hazard-Pointer Recording and Clearing

1 static inline void *_h_t_r_impl(void **p,
2 hazard_pointer *hp)
3 {
4 void *tmp;
5
6 tmp = READ_ONCE(*p);
7 if (!tmp || tmp == (void *)HAZPTR_POISON)
8 return tmp;
9 WRITE_ONCE(hp->p, tmp);
10 smp_mb();
11 if (tmp == READ_ONCE(*p))
12 return tmp;
13 return (void *)HAZPTR_POISON;
14 }
15
16 #define hp_try_record(p, hp) _h_t_r_impl((void **)(p), hp)
17
18 static inline void *hp_record(void **p,
19 hazard_pointer *hp)
20 {
21 void *tmp;
22
23 do {
24 tmp = hp_try_record(p, hp);
25 } while (tmp == (void *)HAZPTR_POISON);
26 return tmp;
27 }
28
29 static inline void hp_clear(hazard_pointer *hp)
30 {
31 smp_mb();
32 WRITE_ONCE(hp->p, NULL);
33 }
Quick Quiz 9.6: Given that papers on hazard pointers use the bottom bits of each pointer to
mark deleted elements, what is up with HAZPTR_POISON?
Line 6 reads the pointer to the object to be protected. If line 8 finds that this pointer
was either NULL or the special HAZPTR_POISON deleted-object token, it returns the
pointer’s value to inform the caller of the failure. Otherwise, line 9 stores the pointer
into the specified hazard pointer, and line 10 forces full ordering of that store with the
reload of the original pointer on line 11. (See Chapter 15 for more information on
memory ordering.) If the value of the original pointer has not changed, then the hazard
pointer protects the pointed-to object, and in that case, line 12 returns a pointer to that
object, which also indicates success to the caller. Otherwise, if the pointer changed
between the two READ_ONCE() invocations, line 13 indicates failure.
Quick Quiz 9.7: Why does hp_try_record() in Listing 9.4 take a double indirection to the
data element? Why not void * instead of void **?
The hp_record() function is quite straightforward: It repeatedly invokes hp_try_

record() until the return value is something other than HAZPTR_POISON.
Quick Quiz 9.8: Why bother with hp_try_record()? Wouldn’t it be easier to just use the
failure-immune hp_record() function?
The hp_clear() function is even more straightforward, with an smp_mb() to force

full ordering between the caller’s uses of the object protected by the hazard pointer and
the setting of the hazard pointer to NULL.
Once a hazard-pointer-protected object has been removed from its linked data
structure, so that it is now inaccessible to future hazard-pointer readers, it is passed to
v2023.06.11a
Listing 9.5: Hazard-Pointer Scanning and Freeing

1 int compare(const void *a, const void *b)
2 {
3 return ( *(hazptr_head_t **)a - *(hazptr_head_t **)b );
4 }
5
6 void hazptr_scan()
7 {
8 hazptr_head_t *cur;
9 int i;
10 hazptr_head_t *tmplist;
11 hazptr_head_t **plist = gplist;
12 unsigned long psize;
13
14 if (plist == NULL) {
15 psize = sizeof(hazptr_head_t *) * K * NR_THREADS;
16 plist = (hazptr_head_t **)malloc(psize);
17 BUG_ON(!plist);
18 gplist = plist;
19 }
20 smp_mb();
21 psize = 0;
22 for (i = 0; i < H; i++) {
23 uintptr_t hp = (uintptr_t)READ_ONCE(HP[i].p);
24
25 if (!hp)
26 continue;
27 plist[psize++] = (hazptr_head_t *)(hp & ~0x1UL);
28 }
29 smp_mb();
30 qsort(plist, psize, sizeof(hazptr_head_t *), compare);
31 tmplist = rlist;
32 rlist = NULL;
33 rcount = 0;
34 while (tmplist != NULL) {
35 cur = tmplist;
36 tmplist = tmplist->next;
37 if (bsearch(&cur, plist, psize,
38 sizeof(hazptr_head_t *), compare)) {
39 cur->next = rlist;
40 rlist = cur;
41 rcount++;
42 } else {
43 hazptr_free(cur);
44 }
45 }
46 }
47
48 void hazptr_free_later(hazptr_head_t *n)
49 {
50 n->next = rlist;
51 rlist = n;
52 rcount++;
53 if (rcount >= R) {
54 hazptr_scan();
55 }
56 }
v2023.06.11a
Listing 9.6: Hazard-Pointer Pre-BSD Routing Table Lookup

2 struct hazptr_head hh;
6 int re_freed;
7 };
10 hazard_pointer __thread *my_hazptr;
11
13 {
14 int offset = 0;
17
18 retry:
20 do {
21 rep = hp_try_record(repp, &my_hazptr[offset]);
22 if (!rep)
24 if ((uintptr_t)rep == HAZPTR_POISON)
25 goto retry;
29 abort();
30 return rep->iface;
31 }
hazptr_free_later(), which is shown on lines 48–56 of Listing 9.5 (hazptr.c).

Lines 50 and 51 enqueue the object on a per-thread list rlist and line 52 counts the
object in rcount. If line 53 sees that a sufficiently large number of objects are now
queued, line 54 invokes hazptr_scan() to attempt to free some of them.
The hazptr_scan() function is shown on lines 6–46 of the listing. This function
relies on a fixed maximum number of threads (NR_THREADS) and a fixed maximum
number of hazard pointers per thread (K), which allows a fixed-size array of hazard
pointers to be used. Because any thread might need to scan the hazard pointers, each
thread maintains its own array, which is referenced by the per-thread variable gplist.
If line 14 determines that this thread has not yet allocated its gplist, lines 15–18
carry out the allocation. The memory barrier on line 20 ensures that all threads see the
removal of all objects by this thread before lines 22–28 scan all of the hazard pointers,
accumulating non-NULL pointers into the plist array and counting them in psize.
The memory barrier on line 29 ensures that the reads of the hazard pointers happen
before any objects are freed. Line 30 then sorts this array to enable use of binary search
below.
Lines 31 and 32 remove all elements from this thread’s list of to-be-freed objects,
placing them on the local tmplist and line 33 zeroes the count. Each pass through the
loop spanning lines 34–45 processes each of the to-be-freed objects. Lines 35 and 36
remove the first object from tmplist, and if lines 37 and 38 determine that there is a
hazard pointer protecting this object, lines 39–41 place it back onto rlist. Otherwise,
line 43 frees the object.
The Pre-BSD routing example can use hazard pointers as shown in Listing 9.6
for data structures and route_lookup(), and in Listing 9.7 for route_add() and
route_del() (route_hazptr.c). As with reference counting, the hazard-pointers
v2023.06.11a
implementation is quite similar to the sequential algorithm shown in Listing 9.1 on

page 203, so only differences will be discussed.
Starting with Listing 9.6, line 2 shows the ->hh field used to queue objects pending
hazard-pointer free, line 6 shows the ->re_freed field used to detect use-after-free
bugs, and line 21 invokes hp_try_record() to attempt to acquire a hazard pointer. If
the return value is NULL, line 23 returns a not-found indication to the caller. If the call
to hp_try_record() raced with deletion, line 25 branches back to line 18’s retry
to re-traverse the list from the beginning. The do–while loop falls through when the
desired element is located, but if this element has already been freed, line 29 terminates
the program. Otherwise, the element’s ->iface field is returned to the caller.
Note that line 21 invokes hp_try_record() rather than the easier-to-use hp_
record(), restarting the full search upon hp_try_record() failure. And such
restarting is absolutely required for correctness. To see this, consider a hazard-pointer-
protected linked list containing elements A, B, and C that is subjected to the following
sequence of events:
1. Thread 0 stores a hazard pointer to element B (having presumably traversed to

element B from element A).
2. Thread 1 removes element B from the list, which sets the pointer from element B
to element C to the special HAZPTR_POISON value in order to mark the deletion.
Because Thread 0 has a hazard pointer to element B, it cannot yet be freed.
3. Thread 1 removes element C from the list. Because there are no hazard pointers
referencing element C, it is immediately freed.
4. Thread 0 attempts to acquire a hazard pointer to now-removed element B’s successor,
but hp_try_record() returns the HAZPTR_POISON value, forcing the caller to
restart its traversal from the beginning of the list.
Which is a very good thing, because B’s successor is the now-freed element C, which
means that Thread 0’s subsequent accesses might have resulted in arbitrarily horrible
memory corruption, especially if the memory for element C had since been re-allocated
for some other purpose. Therefore, hazard-pointer readers must typically restart the
full traversal in the face of a concurrent deletion. Often the restart must go back to
some global (and thus immortal) pointer, but it is sometimes possible to restart at some
intermediate location if that location is guaranteed to still be live, for example, due to
the current thread holding a lock, a reference count, etc.
Quick Quiz 9.9: Readers must “typically” restart? What are some exceptions?
Because algorithms using hazard pointers might be restarted at any step of their
traversal through the linked data structure, such algorithms must typically take care to
avoid making any changes to the data structure until after they have acquired all the
hazard pointers that are required for the update in question.
Quick Quiz 9.10: But don’t these restrictions on hazard pointers also apply to other forms of
reference counting?
These hazard-pointer restrictions result in great benefits to readers, courtesy of the

fact that the hazard pointers are stored local to each CPU or thread, which in turn allows
traversals to be carried out without any writes to the data structures being traversed.
Referring back to Figure 5.8 on page 112, hazard pointers enable the CPU caches to
v2023.06.11a
Listing 9.7: Hazard-Pointer Pre-BSD Routing Table Add/Delete

2 {
4
6 if (!rep)
7 return -ENOMEM;
8 rep->addr = addr;
15 return 0;
16 }
17
19 {
22
25 for (;;) {
26 rep = *repp;
27 if (rep == NULL)
28 break;
31 rep->re_next = (struct route_entry *)HAZPTR_POISON;
33 hazptr_free_later(&rep->hh);
34 return 0;
35 }
37 }
39 return -ENOENT;
40 }
do resource replication, which in turn allows weakening of the parallel-access-control

mechanism, thus boosting performance and scalability.
Another advantage of restarting hazard pointers traversals is a reduction in minimal
memory footprint: Any object not currently referenced by some hazard pointer may
be immediately freed. In contrast, Section 9.5 will discuss a mechanism that avoids
read-side retries (and minimizes read-side overhead), but which can result in a much
larger memory footprint.
The route_add() and route_del() functions are shown in Listing 9.7. Line 10
initializes ->re_freed, line 31 poisons the ->re_next field of the newly removed
object, and line 33 passes that object to the hazptr_free_later() function, which
will free that object once it is safe to do so. The spinlocks work the same as in Listing 9.3.
Figure 9.3 shows the hazard-pointers-protected Pre-BSD routing algorithm’s perfor-
mance on the same read-only workload as for Figure 9.2. Although hazard pointers
scale far better than does reference counting, hazard pointers still require readers to
do writes to shared memory (albeit with much improved locality of reference), and
also require a full memory barrier and retry check for each object traversed. Therefore,
hazard-pointers performance is still far short of ideal. On the other hand, unlike naive
approaches to concurrent reference-counting, hazard pointers not only operate correctly
for workloads involving concurrent updates, but also exhibit excellent scalability. Addi-
v2023.06.11a
9.4. SEQUENCE LOCKS 213
2.5x107
2x107

ideal
7
1.5x10
1x107
5x106
hazptr
0
0 50 100 150 200 250 300 350 400 450
Figure 9.3: Pre-BSD Routing Table Protected by Hazard Pointers
tional performance comparisons with other mechanisms may be found in Chapter 10

and in other publications [HMBW07, McK13, Mic04a].
Quick Quiz 9.11: Figure 9.3 shows no sign of hyperthread-induced flattening at 224 threads.
Why is that?
Quick Quiz 9.12: The paper “Structured Deferral: Synchronization via Procrastina-
tion” [McK13] shows that hazard pointers have near-ideal performance. Whatever happened in
Figure 9.3???
And hazard pointers are the concurrent reference counter mentioned on page 201.
The next section attempts to improve on hazard pointers by using sequence locks, which
avoid both read-side writes and per-object memory barriers.
9.4 Sequence Locks

It’ll be just like starting over.
John Lennon
The published sequence-lock record [Eas71, Lam77] extends back as far as that of
reader-writer locking, but sequence locks nevertheless remain in relative obscurity.
Sequence locks are used in the Linux kernel for read-mostly data that must be seen in
a consistent state by readers. However, unlike reader-writer locking, readers do not
exclude writers. Instead, like hazard pointers, sequence locks force readers to retry
an operation if they detect activity from a concurrent writer. As can be seen from
Figure 9.4, it is important to design code using sequence locks so that readers very
rarely need to retry.
Quick Quiz 9.13: Why isn’t this sequence-lock discussion in Chapter 7, you know, the one on
locking?
The key component of sequence locking is the sequence number, which has an even
value in the absence of updaters and an odd value if there is an update in progress.
Readers can then snapshot the value before and after each access. If either snapshot has
v2023.06.11a
Ah, I finally got

done reading!
No, you didn't!

Start over!
Figure 9.4: Reader And Uncooperative Sequence Lock
Listing 9.8: Sequence-Locking Reader

1 do {
2 seq = read_seqbegin(&test_seqlock);
3 /* read-side access. */
4 } while (read_seqretry(&test_seqlock, seq));
Listing 9.9: Sequence-Locking Writer

1 write_seqlock(&test_seqlock);
2 /* Update */
3 write_sequnlock(&test_seqlock);
an odd value, or if the two snapshots differ, there has been a concurrent update, and the
reader must discard the results of the access and then retry it. Readers therefore use
the read_seqbegin() and read_seqretry() functions shown in Listing 9.8 when
accessing data protected by a sequence lock. Writers must increment the value before
and after each update, and only one writer is permitted at a given time. Writers therefore
use the write_seqlock() and write_sequnlock() functions shown in Listing 9.9
when updating data protected by a sequence lock.
As a result, sequence-lock-protected data can have an arbitrarily large number of
concurrent readers, but only one writer at a time. Sequence locking is used in the Linux
kernel to protect calibration quantities used for timekeeping. It is also used in pathname
traversal to detect concurrent rename operations.
A simple implementation of sequence locks is shown in Listing 9.10 (seqlock.h).
The seqlock_t data structure is shown on lines 1–4, and contains the sequence number
along with a lock to serialize writers. Lines 6–10 show seqlock_init(), which, as
the name indicates, initializes a seqlock_t.
Lines 12–19 show read_seqbegin(), which begins a sequence-lock read-side
critical section. Line 16 takes a snapshot of the sequence counter, and line 17 orders
this snapshot operation before the caller’s critical section. Finally, line 18 returns the
value of the snapshot (with the least-significant bit cleared), which the caller will pass
to a later call to read_seqretry().
Quick Quiz 9.14: Why not have read_seqbegin() in Listing 9.10 check for the low-order
bit being set, and retry internally, rather than allowing a doomed read to start?
v2023.06.11a
Listing 9.10: Sequence-Locking Implementation

1 typedef struct {
2 unsigned long seq;
3 spinlock_t lock;
4 } seqlock_t;
5
6 static inline void seqlock_init(seqlock_t *slp)
7 {
8 slp->seq = 0;
9 spin_lock_init(&slp->lock);
10 }
11
12 static inline unsigned long read_seqbegin(seqlock_t *slp)
13 {
14 unsigned long s;
15
16 s = READ_ONCE(slp->seq);
17 smp_mb();
18 return s & ~0x1UL;
19 }
20
21 static inline int read_seqretry(seqlock_t *slp,
22 unsigned long oldseq)
23 {
24 unsigned long s;
25
26 smp_mb();
27 s = READ_ONCE(slp->seq);
28 return s != oldseq;
29 }
30
31 static inline void write_seqlock(seqlock_t *slp)
32 {
33 spin_lock(&slp->lock);
34 ++slp->seq;
35 smp_mb();
36 }
37
38 static inline void write_sequnlock(seqlock_t *slp)
39 {
40 smp_mb();
41 ++slp->seq;
42 spin_unlock(&slp->lock);
43 }
v2023.06.11a
Listing 9.11: Sequence-Locked Pre-BSD Routing Table Lookup (BUGGY!!!)

5 int re_freed;
6 };
8 DEFINE_SEQ_LOCK(sl);
9
11 {
15 unsigned long s;
16
17 retry:
18 s = read_seqbegin(&sl);
20 do {
21 rep = READ_ONCE(*repp);
22 if (rep == NULL) {
23 if (read_seqretry(&sl, s))
24 goto retry;
26 }
30 abort();
32 if (read_seqretry(&sl, s))
33 goto retry;
34 return ret;
35 }
Lines 21–29 show read_seqretry(), which returns true if there was at least one
writer since the time of the corresponding call to read_seqbegin(). Line 26 orders the
caller’s prior critical section before line 27’s fetch of the new snapshot of the sequence
counter. Line 28 checks whether the sequence counter has changed, in other words,
whether there has been at least one writer, and returns true if so.
Quick Quiz 9.15: Why is the smp_mb() on line 26 of Listing 9.10 needed?
Quick Quiz 9.16: Can’t weaker memory barriers be used in the code in Listing 9.10?
Quick Quiz 9.17: What prevents sequence-locking updaters from starving readers?
Lines 31–36 show write_seqlock(), which simply acquires the lock, increments
the sequence number, and executes a memory barrier to ensure that this increment is
ordered before the caller’s critical section. Lines 38–43 show write_sequnlock(),
which executes a memory barrier to ensure that the caller’s critical section is ordered
before the increment of the sequence number on line 41, then releases the lock.
Quick Quiz 9.18: What if something else serializes writers, so that the lock is not needed?
Quick Quiz 9.19: Why isn’t seq on line 2 of Listing 9.10 unsigned rather than unsigned
long? After all, if unsigned is good enough for the Linux kernel, shouldn’t it be good enough
for everyone?
v2023.06.11a
Listing 9.12: Sequence-Locked Pre-BSD Routing Table Add/Delete (BUGGY!!!)

2 {
4
6 if (!rep)
7 return -ENOMEM;
8 rep->addr = addr;
11 write_seqlock(&sl);
14 write_sequnlock(&sl);
15 return 0;
16 }
17
19 {
22
23 write_seqlock(&sl);
25 for (;;) {
26 rep = *repp;
27 if (rep == NULL)
28 break;
32 smp_mb();
34 free(rep);
35 return 0;
36 }
38 }
40 return -ENOENT;
41 }
So what happens when sequence locking is applied to the Pre-BSD routing table?
Listing 9.11 shows the data structures and route_lookup(), and Listing 9.12 shows
route_add() and route_del() (route_seqlock.c). This implementation is once
again similar to its counterparts in earlier sections, so only the differences will be
highlighted.
In Listing 9.11, line 5 adds ->re_freed, which is checked on lines 29 and 30. Line 8
adds a sequence lock, which is used by route_lookup() on lines 18, 23, and 32, with
lines 24 and 33 branching back to the retry label on line 17. The effect is to retry any
lookup that runs concurrently with an update.
In Listing 9.12, lines 11, 14, 23, 31, and 39 acquire and release the sequence lock,
while lines 10 and 33 handle ->re_freed. This implementation is therefore quite
straightforward.
It also performs better on the read-only workload, as can be seen in Figure 9.5, though
its performance is still far from ideal. Worse yet, it suffers use-after-free failures. The
problem is that the reader might encounter a segmentation violation due to accessing an
already-freed structure before read_seqretry() has a chance to warn of the concurrent
update.
v2023.06.11a
2.5x107
2x107

ideal
7
1.5x10
1x107
seqlock
5x106
hazptr
0
0 50 100 150 200 250 300 350 400 450
Figure 9.5: Pre-BSD Routing Table Protected by Sequence Locking
Quick Quiz 9.20: Can this bug be fixed? In other words, can you use sequence locks as
the only synchronization mechanism protecting a linked list supporting concurrent addition,
deletion, and lookup?
As hinted on page 201, both the read-side and write-side critical sections of a sequence
lock can be thought of as transactions, and sequence locking therefore can be thought
of as a limited form of transactional memory, which will be discussed in Section 17.2.
The limitations of sequence locking are: (1) Sequence locking restricts updates and
(2) Sequence locking does not permit traversal of pointers to objects that might be freed
by updaters. These limitations are of course overcome by transactional memory, but can
also be overcome by combining other synchronization primitives with sequence locking.
Sequence locks allow writers to defer readers, but not vice versa. This can result in
unfairness and even starvation in writer-heavy workloads.3 On the other hand, in the
absence of writers, sequence-lock readers are reasonably fast and scale linearly. It is only
human to want the best of both worlds: Fast readers without the possibility of read-side
failure, let alone starvation. In addition, it would also be nice to overcome sequence
locking’s limitations with pointers. The following section presents a synchronization
mechanism with exactly these properties.
9.5 Read-Copy Update (RCU)
“Free” is a very good price!

Tom Peterson
All of the mechanisms discussed in the preceding sections used one of a number of
approaches to defer specific actions until they may be carried out safely. The reference
counters discussed in Section 9.2 use explicit counters to defer actions that could disturb
readers, which results in read-side contention and thus poor scalability. The hazard
pointers covered by Section 9.3 uses implicit counters in the guise of per-thread lists
3 Dmitry Vyukov describes one way to reduce (but, sadly, not eliminate) reader
starvation: http://www.1024cores.net/home/lock-free-algorithms/reader-
writer-problem/improved-lock-free-seqlock.
v2023.06.11a
9.5. READ-COPY UPDATE (RCU) 219
of pointer. This avoids read-side contention, but requires readers to do stores and
conditional branches, as well as either full memory barriers in read-side primitives or
real-time-unfriendly inter-processor interrupts in update-side primitives.4 The sequence
lock presented in Section 9.4 also avoids read-side contention, but does not protect
pointer traversals and, like hazard pointers, requires either full memory barriers in
read-side primitives, or inter-processor interrupts in update-side primitives. These
schemes’ shortcomings raise the question of whether it is possible to do better.
This section introduces read-copy update (RCU), which provides an API that allows
readers to be associated with regions in the source code, rather than with expensive
updates to frequently updated shared data. The remainder of this section examines RCU
from a number of different perspectives. Section 9.5.1 provides the classic introduction
to RCU, Section 9.5.2 covers fundamental RCU concepts, Section 9.5.3 presents the
Linux-kernel API, Section 9.5.4 introduces some common RCU use cases, and finally
Section 9.5.5 covers recent work related to RCU.
Although RCU has gained a reputation for being subtle and difficult, when used
properly, it is quite straightforward. In fact, no less an authority than Butler Lampson
classifies it as easy concurrency [AH22, Chapter 3].
9.5.1 Introduction to RCU

The approaches discussed in the preceding sections have provided good scalability but
decidedly non-ideal performance for the Pre-BSD routing table. Therefore, in the spirit
of “only those who have gone too far know how far you can go”,5 we will go all the
way, looking into algorithms in which concurrent readers might well execute exactly the
same sequence of assembly language instructions as would a single-threaded lookup,
despite the presence of concurrent updates. Of course, this laudable goal might raise
serious implementability questions, but we cannot possibly succeed if we don’t even try!
And should we succeed, we will have uncovered yet another of the mysteries set forth
on page 201.
9.5.1.1 Minimal Insertion and Deletion

To minimize implementability concerns, we focus on a minimal data structure, which
consists of a single global pointer that is either NULL or references a single structure.
Minimal though it might be, this data structure is heavily used in production [RH18].
A classic approach for insertion is shown in Figure 9.6, which shows four states with
time advancing from top to bottom. The first row shows the initial state, with gptr
equal to NULL. In the second row, we have allocated a structure which is uninitialized,
as indicated by the question marks. In the third row, we have initialized the structure.
Finally, in the fourth and final row, we have updated gptr to reference the newly
allocated and initialized element.
We might hope that this assignment to gptr could use a simple C-language assignment
statement. Unfortunately, Section 4.3.4.1 dashes these hopes. Therefore, the updater can-
not use a simple C-language assignment, but must instead use smp_store_release()
as shown in the figure, or, as will be seen, rcu_assign_pointer().
4 In some important special cases, this extra work can be avoided by using link counting
as exemplified by the UnboundedQueue and ConcurrentHashMap data structures implemented

in Folly open-source library (https://github.com/facebook/folly).
5 With apologies to T. S. Eliot.
v2023.06.11a
(1) gptr
kmalloc()
p
->addr=?
(2) gptr
->iface=?
initialization
p
->addr=42
(3) gptr
->iface=1
smp_store_release(&gptr, p);
p
->addr=42
(4) gptr
->iface=1
Figure 9.6: Insertion With Concurrent Readers
Similarly, one might hope that readers could use a single C-language assignment
to fetch the value of gptr, and be guaranteed to either get the old value of NULL or
to get the newly installed pointer, but either way see a valid result. Unfortunately,
Section 4.3.4.1 dashes these hopes as well. To obtain this guarantee, readers must
instead use READ_ONCE(), or, as will be seen, rcu_dereference(). However, on
most modern computer systems, each of these read-side primitives can be implemented
with a single load instruction, exactly the instruction that would normally be used in
single-threaded code.
Reviewing Figure 9.6 from the viewpoint of readers, in the first three states all readers
see gptr having the value NULL. Upon entering the fourth state, some readers might
see gptr still having the value NULL while others might see it referencing the newly
inserted element, but after some time, all readers will see this new element. At all times,
all readers will see gptr as containing a valid pointer. Therefore, it really is possible
to add new data to linked data structures while allowing concurrent readers to execute
the same sequence of machine instructions that is normally used in single-threaded
code. This no-cost approach to concurrent reading provides excellent performance and
scalability, and also is eminently suitable for real-time use.
Insertion is of course quite useful, but sooner or later, it will also be necessary to
delete data. As can be seen in Figure 9.7, the first step is easy. Again taking the lessons
from Section 4.3.4.1 to heart, smp_store_release() is used to NULL the pointer, thus
moving from the first row to the second in the figure. At this point, pre-existing readers
see the old structure with ->addr of 42 and ->iface of 1, but new readers will see a
v2023.06.11a
->addr=42 Readers?
(1) gptr
->iface=1
1 Version
smp_store_release(&gptr, NULL);
->addr=42
Readers?
(2) gptr
->iface=1
2 Versions
wait
gptr = for readers???
NULL; /*almost*/
->addr=42
Readers?
(3) gptr
->iface=1
1 Version
gptr = NULL; /*almost*/

free()
(4) gptr 1 Version
Figure 9.7: Deletion With Concurrent Readers
NULL pointer, that is, concurrent readers can disagree on the state, as indicated by the
“2 Versions” in the figure.
Quick Quiz 9.21: Why does Figure 9.7 use smp_store_release() given that it is storing a
NULL pointer? Wouldn’t WRITE_ONCE() work just as well in this case, given that there is no
structure initialization to order against the store of the NULL pointer?
Quick Quiz 9.22: Readers running concurrently with each other and with the procedure
outlined in Figure 9.7 can disagree on the value of gptr. Isn’t that just a wee bit problematic???
We get back to a single version simply by waiting for all the pre-existing readers to
complete, as shown in row 3. At that point, all the pre-existing readers are done, and
no later reader has a path to the old data item, so there can no longer be any readers
referencing it. It may therefore be safely freed, as shown on row 4.
Thus, given a way to wait for pre-existing readers to complete, it is possible to both
add data to and remove data from a linked data structure, despite the readers executing
the same sequence of machine instructions that would be appropriate for single-threaded
execution. So perhaps going all the way was not too far after all!
But how can we tell when all of the pre-existing readers have in fact completed? This
question is the topic of Section 9.5.1.3. But first, the next section defines RCU’s core
API.
9.5.1.2 Core RCU API

The full Linux-kernel API is quite extensive, with more than one hundred API members.
However, this section will confine itself to six core RCU API members, which suffices
v2023.06.11a
Table 9.1: Core RCU API
Primitive Purpose
Readers rcu_read_lock() Start an RCU read-side critical section.

rcu_read_unlock() End an RCU read-side critical section.
rcu_dereference() Safely load an RCU-protected pointer.
Updaters synchronize_rcu() Wait for all pre-existing RCU read-side critical

sections to complete.
call_rcu() Invoke the specified function after all pre-existing
RCU read-side critical sections complete.
rcu_assign_pointer() Safely update an RCU-protected pointer.
for the upcoming sections introducing RCU and covering its fundamentals. The full
API is covered in Section 9.5.3.
Three members of the core APIs are used by readers. The rcu_read_lock() and
rcu_read_unlock() functions delimit RCU read-side critical sections. These may
be nested, so that one rcu_read_lock()–rcu_read_unlock() pair can be enclosed
within another. In this case, the nested set of RCU read-side critical sections act as
one large critical section covering the full extent of the nested set. The third read-side
API member, rcu_dereference(), fetches an RCU-protected pointer. Conceptually,
rcu_dereference() simply loads from memory, but we will see in Section 9.5.2.1
that rcu_dereference() must prevent the compiler and (in one case) the CPU from
reordering its load with later memory operations that dereference this pointer.
Quick Quiz 9.23: What is an RCU-protected pointer?
The other three members of the core APIs are used by updaters. The synchronize_
rcu() function implements the “wait for readers” operation from Figure 9.7. The
call_rcu() function is the asynchronous counterpart of synchronize_rcu() by
invoking the specified function after all pre-existing RCU readers have completed.
Finally, the rcu_assign_pointer() macro is used to update an RCU-protected
pointer. Conceptually, this is simply an assignment statement, but we will see in
Section 9.5.2.1 that rcu_assign_pointer() must prevent the compiler and the CPU
from reordering this assignment to precede any prior assignments used to initialize the
pointed-to structure.
Quick Quiz 9.24: What does synchronize_rcu() do if it starts at about the same time as
an rcu_read_lock()?
The core RCU API is summarized in Table 9.1 for easy reference. With that, we are
ready to continue this introduction to RCU with the key RCU operation, waiting for
readers.
9.5.1.3 Waiting for Readers

It is tempting to base the reader-waiting functionality of synchronize_rcu() and
call_rcu() on a reference counter updated by rcu_read_lock() and rcu_read_
unlock(), but Figure 5.1 in Chapter 5 shows that concurrent reference counting results
in extreme overhead. This extreme overhead was confirmed in the specific case of
v2023.06.11a
reference counters in Figure 9.2 on page 205. Hazard pointers profoundly reduce this
overhead, but, as we saw in Figure 9.3 on page 213, not to zero. Nevertheless, many
RCU implementations use counters with carefully controlled cache locality.
A second approach observes that memory synchronization is expensive, and therefore
uses registers instead, namely each CPU’s or thread’s program counter (PC), thus
imposing no overhead on readers, at least in the absence of concurrent updates. The
updater polls each relevant PC, and if that PC is not within read-side code, then the
corresponding CPU or thread is within a quiescent state, in turn signaling the completion
of any reader that might have access to the newly removed data element. Once all
CPU’s or thread’s PCs have been observed to be outside of any reader, the grace
period has completed. Please note that this approach poses some serious challenges,
including memory ordering, functions that are sometimes invoked from readers, and
ever-exciting code-motion optimizations. Nevertheless, this approach is said to be used
in production [Ash15].
A third approach is to simply wait for a fixed period of time that is long enough to
comfortably exceed the lifetime of any reasonable reader [Jac93, Joh95]. This can work
quite well in hard real-time systems [RLPB18], but in less exotic settings, Murphy says
that it is critically important to be prepared even for unreasonably long-lived readers.
To see this, consider the consequences of failing do so: A data item will be freed while
the unreasonable reader is still referencing it, and that item might well be immediately
reallocated, possibly even as a data item of some other type. The unreasonable reader
and the unwitting reallocator would then be attempting to use the same memory for two
very different purposes. The ensuing mess will be exceedingly difficult to debug.
A fourth approach is to wait forever, secure in the knowledge that doing so will
accommodate even the most unreasonable reader. This approach is also called “leaking
memory”, and has a bad reputation due to the fact that memory leaks often require
untimely and inconvenient reboots. Nevertheless, this is a viable strategy when the
update rate and the uptime are both sharply bounded. For example, this approach could
work well in a high-availability cluster where systems were periodically crashed in order
to ensure that cluster really remained highly available.6 Leaking the memory is also a
viable strategy in environments having garbage collectors, in which case the garbage
collector can be thought of as plugging the leak [KL80]. However, if your environment
lacks a garbage collector, read on!
A fifth approach avoids the period crashes in favor of periodically “stopping the
world”, as exemplified by the traditional stop-the-world garbage collector. This approach
was also heavily used during the decades before ubiquitous connectivity, when it was
common practice to power systems off at the end of each working day. However, in
today’s always-connected always-on world, stopping the world can gravely degrade
response times, which has been one motivation for the development of concurrent
garbage collectors [BCR03]. Furthermore, although we need all pre-existing readers to
complete, we do not need them all to complete at the same time.
This observation leads to the sixth approach, which is stopping one CPU or thread
at a time. This approach has the advantage of not degrading reader response times at
all, let alone gravely. Furthermore, numerous applications already have states (termed
quiescent states) that can be reached only after all pre-existing readers are done. In
transaction-processing systems, the time between a pair of successive transactions might
6 The program that forces the periodic crashing is sometimes known as a “chaos monkey”:
https://netflix.github.io/chaosmonkey/. However, it might also be a mistake to

neglect chaos caused by systems running for too long.
v2023.06.11a
be a quiescent state. In reactive systems, the state between a pair of successive events
might be a quiescent state. Within non-preemptive operating-systems kernels, a context
switch can be a quiescent state [MS98a]. Either way, once all CPUs and/or threads have
passed through a quiescent state, the system is said to have completed a grace period, at
which point all readers in existence at the start of that grace period are guaranteed to
have completed. As a result, it is also guaranteed to be safe to free any removed data
items that were removed prior to the start of that grace period.7
Within a non-preemptive operating-system kernel, for context switch to be a valid
quiescent state, readers must be prohibited from blocking while referencing a given
instance data structure obtained via the gptr pointer shown in Figures 9.6 and 9.7. This
no-blocking constraint is consistent with similar constraints on pure spinlocks, where a
CPU is forbidden from blocking while holding a spinlock. Without this constraint, all
CPUs might be consumed by threads spinning attempting to acquire a spinlock held
by a blocked thread. The spinning threads will not relinquish their CPUs until they
acquire the lock, but the thread holding the lock cannot possibly release it until one of
the spinning threads relinquishes a CPU. This is a classic deadlock situation, and this
deadlock is avoided by forbidding blocking while holding a spinlock.
Again, this same constraint is imposed on reader threads dereferencing gptr: Such
threads are not allowed to block until after they are done using the pointed-to data
item. Returning to the second row of Figure 9.7, where the updater has just completed
executing the smp_store_release(), imagine that CPU 0 executes a context switch.
Because readers are not permitted to block while traversing the linked list, we are
guaranteed that all prior readers that might have been running on CPU 0 will have
completed. Extending this line of reasoning to the other CPUs, once each CPU has
been observed executing a context switch, we are guaranteed that all prior readers
have completed, and that there are no longer any reader threads referencing the newly
removed data element. The updater can then safely free that data element, resulting in
the state shown at the bottom of Figure 9.7.
This approach is termed quiescent-state-based reclamation (QSBR) [HMB06]. A
QSBR schematic is shown in Figure 9.8, with time advancing from the top of the
figure to the bottom. The cyan-colored boxes depict RCU read-side critical sections,
each of which begins with rcu_read_lock() and ends with rcu_read_unlock().
CPU 1 does the WRITE_ONCE() that removes the current data item (presumably having
previously read the pointer value and availed itself of appropriate synchronization), then
waits for readers. This wait operation results in an immediate context switch, which is a
quiescent state (denoted by the pink circle), which in turn means that all prior reads
on CPU 1 have completed. Next, CPU 2 does a context switch, so that all readers on
CPUs 1 and 2 are now known to have completed. Finally, CPU 3 does a context switch.
At this point, all readers throughout the entire system are known to have completed, so
the grace period ends, permitting synchronize_rcu() to return to its caller, in turn
permitting CPU 1 to free the old data item.
Quick Quiz 9.25: In Figure 9.8, the last of CPU 3’s readers that could possibly have access
to the old data item ended before the grace period even started! So why would anyone bother
waiting until CPU 3’s later context switch???
7 It is possible to do much more with RCU than simply defer reclamation of memory, but
deferred reclamation is RCU’s most common use case, and is therefore an excellent place to
start. For an example of the more general case of deferred execution, please see phased state
change in Section 9.5.4.3.
v2023.06.11a
WRITE_ONCE(gptr, NULL);
synchronize_rcu()
CPU 1 CPU 2 CPU 3
Context Switch
Reader
Grace Period
free()
Figure 9.8: QSBR: Waiting for Pre-Existing Readers
9.5.1.4 Toy Implementation

Although production-quality QSBR implementations can be quite complex, a toy
non-preemptive Linux-kernel implementation is quite simple:
1 void synchronize_rcu(void)
2 {
3 int cpu;
4
5 for_each_online_cpu(cpu)
6 sched_setaffinity(current->pid, cpumask_of(cpu));
7 }
The for_each_online_cpu() primitive iterates over all CPUs, and the sched_
setaffinity() function causes the current thread to execute on the specified CPU,
which forces the destination CPU to execute a context switch. Therefore, once the
for_each_online_cpu() has completed, each CPU has executed a context switch,
which in turn guarantees that all pre-existing reader threads have completed.
Please note that this approach is not production quality. Correct handling of a
number of corner cases and the need for a number of powerful optimizations mean that
production-quality implementations are quite complex. In addition, RCU implementa-
tions for preemptible environments require that readers actually do something, which
in non-real-time Linux-kernel environments can be as simple as defining rcu_read_
lock() and rcu_read_unlock() as preempt_disable() and preempt_enable(),
respectively.8 However, this simple non-preemptible approach is conceptually complete,
8 Some toy RCU implementations that handle preempted read-side critical sections are
shown in Appendix B.
v2023.06.11a
Listing 9.13: Insertion and Deletion With Concurrent Readers

1 struct route *gptr;
2
3 int access_route(int (*f)(struct route *rp))
4 {
5 int ret = -1;
6 struct route *rp;
7
8 rcu_read_lock();
9 rp = rcu_dereference(gptr);
10 if (rp)
11 ret = f(rp);
12 rcu_read_unlock();
13 return ret;
14 }
15
16 struct route *ins_route(struct route *rp)
17 {
18 struct route *old_rp;
19
20 spin_lock(&route_lock);
21 old_rp = gptr;
22 rcu_assign_pointer(gptr, rp);
23 spin_unlock(&route_lock);
24 return old_rp;
25 }
26
27 int del_route(void)
28 {
29 struct route *old_rp;
30
31 spin_lock(&route_lock);
32 old_rp = gptr;
33 RCU_INIT_POINTER(gptr, NULL);
34 spin_unlock(&route_lock);
35 synchronize_rcu();
36 free(old_rp);
37 return !!old_rp;
38 }
and demonstrates that it really is possible to provide read-side synchronization at zero

cost, even in the face of concurrent updates. In fact, Listing 9.13 shows how reading
(access_route()), Figure 9.6’s insertion (ins_route()) and Figure 9.7’s deletion
(del_route()) can be implemented. (A slightly more capable routing table is shown
in Section 9.5.4.1.)
Quick Quiz 9.26: What is the point of rcu_read_lock() and rcu_read_unlock() in
Listing 9.13? Why not just let the quiescent states speak for themselves?
Quick Quiz 9.27: What is the point of rcu_dereference(), rcu_assign_pointer()

and RCU_INIT_POINTER() in Listing 9.13? Why not just use READ_ONCE(), smp_store_
release(), and WRITE_ONCE(), respectively?
Referring back to Listing 9.13, note that route_lock is used to synchronize between
concurrent updaters invoking ins_route() and del_route(). However, this lock is
not acquired by readers invoking access_route(): Readers are instead protected by
the QSBR techniques described in Section 9.5.1.3.
Note that ins_route() simply returns the old value of gptr, which Figure 9.6
assumed would always be NULL. This means that it is the caller’s responsibility to figure
out what to do with a non-NULL value, a task complicated by the fact that readers might
still be referencing it for an indeterminate period of time. Callers might use one of the
following approaches:
v2023.06.11a
1. Use synchronize_rcu() to safely free the pointed-to structure. Although this

approach is correct from an RCU perspective, it arguably has software-engineering
leaky-API problems.
2. Trip an assertion if the returned pointer is non-NULL.
3. Pass the returned pointer to a later invocation of ins_route() to restore the earlier
value.
In contrast, del_route() uses synchronize_rcu() and free() to safely free the

newly deleted data item.
Quick Quiz 9.28: But what if the old structure needs to be freed, but the caller of ins_
route() cannot block, perhaps due to performance considerations or perhaps because the
caller is executing within an RCU read-side critical section?
This example shows one general approach to reading and updating RCU-protected
data structures, however, there is quite a variety of use cases, several of which are
covered in Section 9.5.4.
In summary, it is in fact possible to create concurrent linked data structures that can
be traversed by readers executing the same sequence of machine instructions that would
be executed by single-threaded readers. The next section summarizes RCU’s high-level
properties.
9.5.1.5 RCU Properties
A key RCU property is that reads need not wait for updates. This property enables
RCU implementations to provide low-cost or even no-cost readers, resulting in low
overhead and excellent scalability. This property also allows RCU readers and updaters
to make useful concurrent forward progress. In contrast, conventional synchronization
primitives must enforce strict mutual exclusion using expensive instructions, thus
increasing overhead and degrading scalability, but also typically prohibiting readers and
updaters from making useful concurrent forward progress.
Quick Quiz 9.29: Doesn’t Section 9.4’s seqlock also permit readers and updaters to make
useful concurrent forward progress?
As noted earlier, RCU delimits readers with rcu_read_lock() and rcu_read_

unlock(), and ensures that each reader has a coherent view of each object (see
Figure 9.7) by maintaining multiple versions of objects and using update-side primitives
such as synchronize_rcu() to ensure that objects are not freed until after the
completion of all readers that might be using them. RCU uses rcu_assign_pointer()
and rcu_dereference() to provide efficient and scalable mechanisms for publishing
and reading new versions of an object, respectively. These mechanisms distribute the
work among read and update paths in such a way as to make read paths extremely fast,
using replication and weakening optimizations in a manner similar to hazard pointers,
but without the need for read-side retries. In some cases, including CONFIG_PREEMPT=n
Linux kernels, RCU’s read-side primitives have zero overhead.
But are these properties actually useful in practice? This question is taken up by the
next section.
v2023.06.11a
20000
18000
16000
# RCU API Uses

14000
12000
10000
8000
6000
4000
2000
0
2000
2005
2010
2015
2020
2025
Year
Figure 9.9: RCU Usage in the Linux Kernel
9.5.1.6 Practical Applicability

RCU has been used in the Linux kernel since October 2002 [Tor02]. Use of the RCU
API has increased substantially since that time, as can be seen in Figure 9.9. RCU
has enjoyed heavy use both prior to and since its acceptance in the Linux kernel, as
discussed in Section 9.5.5. In short, RCU enjoys wide practical applicability.
The minimal example discussed in this section is a good introduction to RCU.
However, effective use of RCU often requires that you think differently about your
problem. It is therefore useful to examine RCU’s fundamentals, a task taken up by the
following section.
9.5.2 RCU Fundamentals

This section re-examines the ground covered in the previous section, but independent of
any particular example or use case. People who prefer to live their lives very close to
the actual code may wish to skip the underlying fundamentals presented in this section.
RCU is made up of three fundamental mechanisms, the first being used for insertion,
the second being used for deletion, and the third being used to allow readers to tolerate
concurrent insertions and deletions. Section 9.5.2.1 describes the publish-subscribe
mechanism used for insertion, Section 9.5.2.2 describes how waiting for pre-existing
RCU readers enables deletion, and Section 9.5.2.3 discusses how maintaining multiple
versions of recently updated objects permits concurrent insertions and deletions. Finally,
Section 9.5.2.4 summarizes RCU fundamentals.
9.5.2.1 Publish-Subscribe Mechanism

Because RCU readers are not excluded by RCU updaters, an RCU-protected data
structure might change while a reader accesses it. The accessed data item might be
moved, removed, or replaced. Because the data structure does not “hold still” for the
reader, each reader’s access can be thought of as subscribing to the current version of
v2023.06.11a
ins_route() access_route()
Allocate
Pre-initialization
garbage
Initialize
Valid route structure
Subscribe to
Publish pointer
pointer
Not OK
Valid route structure Dereference pointer
OK
Surprising, but OK
Figure 9.10: Publication/Subscription Constraints
the RCU-protected data item. For their part, updaters can be thought of as publishing
new versions.
Unfortunately, as laid out in Section 4.3.4.1 and reiterated in Section 9.5.1.1, it is
unwise to use plain accesses for these publication and subscription operations. It is
instead necessary to inform both the compiler and the CPU of the need for care, as can
be seen from Figure 9.10, which illustrates interactions between concurrent executions
of ins_route() (and its caller) and access_route() from Listing 9.13.
The ins_route() column from Figure 9.10 shows ins_route()’s caller allocating
a new route structure, which then contains pre-initialization garbage. The caller then
initializes the newly allocated structure, and then invokes ins_route() to publish a
pointer to the new route structure. Publication does not affect the contents of the
structure, which therefore remain valid after publication.
The access_route() column from this same figure shows the pointer being sub-
scribed to and dereferenced. This dereference operation absolutely must see a valid
route structure rather than pre-initialization garbage because referencing garbage could
result in memory corruption, crashes, and hangs. As noted earlier, avoiding such
garbage means that the publish and subscribe operations must inform both the compiler
and the CPU of the need to maintain the needed ordering.
Publication is carried out by rcu_assign_pointer(), which ensures that ins_
route()’s caller’s initialization is ordered before the actual publication operation’s
store of the pointer. In addition, rcu_assign_pointer() must be atomic in the sense
that concurrent readers see either the old value of the pointer or the new value of the
pointer, but not some mash-up of these two values. These requirements are met by the
C11 store-release operation, and in fact in the Linux kernel, rcu_assign_pointer()
is defined in terms of smp_store_release(), which is similar to C11 store-release.
Note that if concurrent updates are required, some sort of synchronization mechanism
will be required to mediate among multiple concurrent rcu_assign_pointer() calls
on the same pointer. In the Linux kernel, locking is the mechanism of choice, but
v2023.06.11a
pretty much any synchronization mechanism may be used. An example of a particularly

lightweight synchronization mechanism is Chapter 8’s data ownership: If each pointer
is owned by a particular thread, then that thread may execute rcu_assign_pointer()
on that pointer with no additional synchronization overhead.
Quick Quiz 9.30: Wouldn’t use of data ownership for RCU updaters mean that the updates
could use exactly the same sequence of instructions as would the corresponding single-threaded
code?
Subscription is carried out by rcu_dereference(), which orders the subscrip-

tion operation’s load from the pointer is before the dereference. Similar to rcu_
assign_pointer(), rcu_dereference() must be atomic in the sense that the value
loaded must be that from a single store, for example, the compiler must not tear the
load.9 Unfortunately, compiler support for rcu_dereference() is at best a work in
progress [MWB+ 17, MRP+ 17, BM18]. In the meantime, the Linux kernel relies on
volatile loads, the details of the various CPU architectures, coding restrictions [McK14e],
and, on DEC Alpha [Cor02], a memory-barrier instruction. However, on other architec-
tures, rcu_dereference() typically emits a single load instruction, just as would the
equivalent single-threaded code. The coding restrictions are described in more detail in
Section 15.3.2, however, the common case of field selection (“->”) works quite well.
Software that does not require the ultimate in read-side performance can instead use
C11 acquire loads, which provide the needed ordering and more, albeit at a cost. It
is hoped that lighter-weight compiler support for rcu_dereference() will appear in
due course.
In short, use of rcu_assign_pointer() for publishing pointers and use of rcu_
dereference() for subscribing to them successfully avoids the “Not OK” garbage
loads depicted in Figure 9.10. These two primitives can therefore be used to add new
data to linked structures without disrupting concurrent readers.
Quick Quiz 9.31: But suppose that updaters are adding and removing multiple data items
from a linked list while a reader is iterating over that same list. Specifically, suppose that a list
initially contains elements A, B, and C, and that an updater removes element A and then adds
a new element D at the end of the list. The reader might well see {A, B, C, D}, when that
sequence of elements never actually ever existed! In what alternate universe would that qualify
as “not disrupting concurrent readers”???
Adding data to a linked structure without disrupting readers is a good thing, as are the
cases where this can be done with no added read-side cost compared to single-threaded
readers. However, in most cases it is also necessary to remove data, and this is the
subject of the next section.
9.5.2.2 Wait For Pre-Existing RCU Readers

In its most basic form, RCU is a way of waiting for things to finish. Of course, there
are a great many other ways of waiting for things to finish, including reference counts,
reader-writer locks, events, and so on. The great advantage of RCU is that it can wait
for each of (say) 20,000 different things without having to explicitly track each and
every one of them, and without having to worry about the performance degradation,
scalability limitations, complex deadlock scenarios, and memory-leak hazards that are
inherent in schemes using explicit tracking.
9 That is, the compiler must not break the load into multiple smaller loads, as described
under “load tearing” in Section 4.3.4.1.
v2023.06.11a
P0()
rcu_read_lock()
r1 = x; Given this ordering ...
r2 = y;
rcu_read_unlock() x = 1;
synchronize_rcu()
.... RCU guarantees this ordering. y = 1;
P1()
Figure 9.11: RCU Reader and Later Grace Period
In RCU’s case, each of the things waited on is called an RCU read-side critical
section. As noted in Table 9.1, an RCU read-side critical section starts with an
rcu_read_lock() primitive, and ends with a corresponding rcu_read_unlock()
primitive. RCU read-side critical sections can be nested, and may contain pretty much
any code, as long as that code does not contain a quiescent state. For example, within
the Linux kernel, it is illegal to sleep within an RCU read-side critical section because
a context switch is a quiescent state.10 If you abide by these conventions, you can
use RCU to wait for any pre-existing RCU read-side critical section to complete, and
synchronize_rcu() uses indirect means to do the actual waiting [DMS+ 12, McK13].
The relationship between an RCU read-side critical section and a later RCU grace
period is an if-then relationship, as illustrated by Figure 9.11. If any portion of a given
critical section precedes the beginning of a given grace period, then RCU guarantees
that all of that critical section will precede the end of that grace period. In the figure,
P0()’s access to x precedes P1()’s access to this same variable, and thus also precedes
the grace period generated by P1()’s call to synchronize_rcu(). It is therefore
guaranteed that P0()’s access to y will precede P1()’s access. In this case, if r1’s final
value is 0, then r2’s final value is guaranteed to also be 0.
Quick Quiz 9.32: What other final values of r1 and r2 are possible in Figure 9.11?
The relationship between an RCU read-side critical section and an earlier RCU grace
period is also an if-then relationship, as illustrated by Figure 9.12. If any portion of
a given critical section follows the end of a given grace period, then RCU guarantees
that all of that critical section will follow the beginning of that grace period. In the
10 However, a special form of RCU called SRCU [McK06] does permit general sleeping
in SRCU read-side critical sections.
v2023.06.11a
P1()
x = 1;
P0()
this ordering.
rcu_read_lock() synchronize_rcu()
r1 = x; ... RCU guarantees y = 1;
r2 = y; Given this ordering ...
rcu_read_unlock()
Figure 9.12: RCU Reader and Earlier Grace Period
figure, P0()’s access to y follows P1()’s access to this same variable, and thus follows
the grace period generated by P1()’s call to synchronize_rcu(). It is therefore
guaranteed that P0()’s access to x will follow P1()’s access. In this case, if r2’s final
value is 1, then r1’s final value is guaranteed to also be 1.
Quick Quiz 9.33: What would happen if the order of P0()’s two accesses was reversed in
Figure 9.12?
Finally, as shown in Figure 9.13, an RCU read-side critical section can be completely
overlapped by an RCU grace period. In this case, r1’s final value is 1 and r2’s final
value is 0.
However, it cannot be the case that r1’s final value is 0 and r2’s final value is 1.
This would mean that an RCU read-side critical section had completely overlapped
a grace period, which is forbidden (or at the very least constitutes a bug in RCU).
RCU’s wait-for-readers guarantee therefore has two parts: (1) If any part of a given
RCU read-side critical section precedes the beginning of a given grace period, then
the entirety of that critical section precedes the end of that grace period. (2) If any
part of a given RCU read-side critical section follows the end of a given grace period,
then the entirety of that critical section follows the beginning of that grace period. This
definition is sufficient for almost all RCU-based algorithms, but for those wanting more,
simple executable formal models of RCU are available as part of Linux kernel v4.17
and later, as discussed in Section 12.3.2. In addition, RCU’s ordering properties are
examined in much greater detail in Section 15.4.3.
Quick Quiz 9.34: What would happen if P0()’s accesses in Figures 9.11–9.13 were stores?
Although RCU’s wait-for-readers capability really is sometimes used to order the

assignment of values to variables as shown in Figures 9.11–9.13, it is more frequently
used to safely free data elements removed from a linked structure, as was done in
Section 9.5.1. The general process is illustrated by the following pseudocode:
v2023.06.11a
P1()
x = 1;
P0()
r1 = x; Given this ordering ...
r2 = y; ... this can happen
rcu_read_unlock()
y = 1;
Figure 9.13: RCU Reader Within Grace Period
1. Make a change, for example, remove an element from a linked list.

2. Wait for all pre-existing RCU read-side critical sections to completely finish (for
example, by using synchronize_rcu()).
3. Clean up, for example, free the element that was replaced above.
This more abstract procedure requires a more abstract diagram than Figures 9.11–9.13,
which are specific to a particular litmus test. After all, an RCU implementation must
work correctly regardless of the form of the RCU updates and the RCU read-side critical
sections. Figure 9.14 fills this need, showing the four possible scenarios, with time
advancing from top to bottom within each scenario. Within each scenario, an RCU
reader is represented by the left-hand stack of boxes and RCU updater by the right-hand
stack.
In the first scenario, the reader starts execution before the updater starts the removal,
so it is possible that this reader has a reference to the removed data element. Therefore,
the updater must not free this element until after the reader completes. In the second
scenario, the reader does not start execution until after the removal has completed. The
reader cannot possibly obtain a reference to the already-removed data element, so this
element may be freed before the reader completes. The third scenario is like the second,
but illustrates that even when the reader cannot possibly obtain a reference to an element,
it is still permissible to defer the freeing of that element until after the reader completes.
In the fourth and final scenario, the reader starts execution before the updater starts
removing the data element, but this element is (incorrectly) freed before the reader
completed. A correct RCU implementation will not allow this fourth scenario to occur.
This diagram thus illustrates RCU’s wait-for-readers functionality: Given a grace period,
each reader ends before the end of that grace period, starts after the beginning of that
grace period, or both, in which case it is wholly contained within that grace period.
v2023.06.11a
rcu_read_lock()
Remove
rcu_read_unlock() synchronize_rcu()
Free Old Memory
(1) Reader precedes removal
Remove
Free Old Memory
rcu_read_unlock()
(2) Removal precedes reader
Remove
rcu_read_lock()
synchronize_rcu()
rcu_read_unlock()
Free Old Memory
(3) Reader within grace period
rcu_read_lock()
Remove
synchronize_rcu()
Free Old Memory
rcu_read_unlock()
BUG!!!
(4) Grace period within reader (BUG!!!)
Figure 9.14: Summary of RCU Grace-Period Ordering Guarantees
v2023.06.11a
Reader { A }
1. A B C D
Reader { A, B }
2. A B C D
Reader { A, B }
3. B C D
Reader { A, B }
4. B C D E
Reader { A, B, C, D, E }
5. B C D E
Figure 9.15: Multiple RCU Data-Structure Versions
Because RCU readers can make forward progress while updates are in progress,
different readers might disagree about the state of the data structure, a topic taken up by
the next section.
9.5.2.3 Maintain Multiple Versions of Recently Updated Objects
This section discusses how RCU accommodates synchronization-free readers by main-

taining multiple versions of data. Because these synchronization-free readers provide
very weak temporal synchronization, RCU users compensate via spatial synchronization.
Spatial synchronization was discussed in Chapter 6, and is heavily used in practice to
obtain good performance and scalability. In this section, spatial synchronization will be
used to attain a weak (but useful) form of correctness as well as excellent performance
and scalability.
Figure 9.7 in Section 9.5.1.1 showed a simple variant of spatial synchronization,
in which different readers running concurrently with del_route() (see Listing 9.13)
might see the old route structure or an empty list, but either way get a valid result. Of
course, a closer look at Figure 9.6 shows that calls to ins_route() can also result in
concurrent readers seeing different versions: Either the initial empty list or the newly
inserted route structure. Note that both reference counting (Section 9.2) and hazard
pointers (Section 9.3) can also cause concurrent readers to see different versions, but
RCU’s lightweight readers make this more likely.
v2023.06.11a
However, maintaining multiple weakly consistent versions can provide some surprises.
For example, consider Figure 9.15, in which a reader is traversing a linked list that is
concurrently updated.11 In the first row of the figure, the reader is referencing data
item A, and in the second row, it advances to B, having thus far seen A followed by B.
In the third row, an updater removes element A and in the fourth row an updater adds
element E to the end of the list. In the fifth and final row, the reader completes its
traversal, having seeing elements A through E.
Except that there was no time at which such a list existed. This situation might
be even more surprising than that shown in Figure 9.7, in which different concurrent
readers see different versions. In contrast, in Figure 9.15 the reader sees a version that
never actually existed!
One way to resolve this strange situation is via weaker semanitics. A reader traversal
must encounter any data item that was present during the full traversal (B, C, and D),
and might or might not encounter data items that were present for only part of the
traversal (A and E). Therefore, in this particular case, it is perfectly legitimate for the
reader traversal to encounter all five elements. If this outcome is problematic, another
way to resolve this situation is through use of stronger synchronization mechanisms,
such as reader-writer locking, or clever use of timestamps and versioning, as discussed
in Section 9.5.4.11. Of course, stronger mechanisms will be more expensive, but then
again the engineering life is all about choices and tradeoffs.
Strange though this situation might seem, it is entirely consistent with the real world.
As we saw in Section 3.2, the finite speed of light cannot be ignored within a computer
system, and it most certainly cannot be ignored outside of this system. This in turn
means that any data within the system representing state in the real world outside of
the system is always and forever outdated, and thus inconsistent with the real world.
Therefore, it is quite possible that the sequence {A, B, C, D, E} occurred in the real
world, but due to speed-of-light delays was never represented in the computer system’s
memory. In this case, the reader’s surprising traversal would correctly reflect reality.
As a result, algorithms operating on real-world data must account for inconsistent
data, either by tolerating inconsistencies or by taking steps to exclude or reject them. In
many cases, these algorithms are also perfectly capable of dealing with inconsistencies
within the system.
The pre-BSD packet routing example laid out in Section 9.1 is a case in point.
The contents of a routing list is set by routing protocols, and these protocols feature
significant delays (seconds or even minutes) to avoid routing instabilities. Therefore,
once a routing update reaches a given system, it might well have been sending packets
the wrong way for quite some time. Sending a few more packets the wrong way for the
few microseconds during which the update is in flight is clearly not a problem because
the same higher-level protocol actions that deal with delayed routing updates will also
deal with internal inconsistencies.
Nor is Internet routing the only situation tolerating inconsistencies. To repeat, any
algorithm in which data within a system tracks outside-of-system state must tolerate
inconsistencies, which includes security policies (often set by committees of humans),
storage configuration, and WiFi access points, to say nothing of removable hardware such
as microphones, headsets, cameras, mice, printers, and much else besides. Furthermore,
the large number of Linux-kernel RCU API uses shown in Figure 9.9, combined with
the Linux kernel’s heavy use of reference counting and with increasing use of hazard
11 RCU linked-list APIs may be found in Section 9.5.3.
v2023.06.11a
pointers in other projects, demonstrates that tolerance for such inconsistencies is more
common than one might imagine.
One root cause of this common-case tolerance of inconsistencies is that single-item
lookups are much more common in practice than are full-data-structure traversals. After
all, full-data-structure traversals are much more expensive than single-item lookups, so
developers are motivated to avoid such traversals. Not only are concurrent updates less
likely to affect a single-item lookup than they are a full traversal, but it is also the case
that an isolated single-item lookup has no way of detecting such inconsistencies. As a
result, in the common case, such inconsistencies are not just tolerable, they are in fact
invisible.
In such cases, RCU readers can be considered to be fully ordered with updaters,
despite the fact that these readers might be executing the exact same sequence of
machine instructions that would be executed by a single-threaded program, as hinted on
page 201. For example, referring back to Listing 9.13 on page 226, suppose that each
reader thread invokes access_route() exactly once during its lifetime, and that there
is no other communication among reader and updater threads. Then each invocation
of access_route() can be ordered after the ins_route() invocation that produced
the route structure accessed by line 11 of the listing in access_route() and ordered
before any subsequent ins_route() or del_route() invocation.
In summary, maintaining multiple versions is exactly what enables the extremely
low overheads of RCU readers, and as noted earlier, many algorithms are unfazed by
multiple versions. However, there are algorithms that absolutely cannot handle multiple
versions. There are techniques for adapting such algorithms to RCU [McK04], for
example, the use of sequence locking described in Section 13.4.2.
Exercises These examples assumed that a mutex was held across the entire update
operation, which would mean that there could be at most two versions of the list active
at a given time.
Quick Quiz 9.35: How would you modify the deletion example to permit more than two
versions of the list to be active?
Quick Quiz 9.36: How many RCU versions of a given list can be active at any given time?
Quick Quiz 9.37: How can the per-update overhead of RCU be reduced?
9.5.2.4 Summary of RCU Fundamentals

This section has described the three fundamental components of RCU-based algorithms:
1. A publish-subscribe mechanism for adding new data featuring rcu_assign_
pointer() for update-side publication and rcu_dereference() for read-side
subscription,
2. A way of waiting for pre-existing RCU readers to finish based on readers being
delimited by rcu_read_lock() and rcu_read_unlock() on the one hand and
updaters waiting via synchronize_rcu() or call_rcu() on the other (see
Section 15.4.3 for a formal description), and
3. A discipline of maintaining multiple versions to permit change without harming or
unduly delaying concurrent RCU readers.
v2023.06.11a
Quick Quiz 9.38: How can RCU updaters possibly delay RCU readers, given that neither
rcu_read_lock() nor rcu_read_unlock() spin or block?
These three RCU components allow data to be updated in the face of concurrent
readers that might be executing the same sequence of machine instructions that would
be used by a reader in a single-threaded implementation. These RCU components can
be combined in different ways to implement a surprising variety of different types of
RCU-based algorithms, a number of which are presented in Section 9.5.4. However, it
is usually better to work at higher levels of abstraction. To this end, the next section
describes the Linux-kernel API, which includes simple data structures such as lists.
9.5.3 RCU Linux-Kernel API

This section looks at RCU from the viewpoint of its Linux-kernel API.12 Section 9.5.3.2
presents RCU’s wait-to-finish APIs, Section 9.5.3.3 presents RCU’s publish-subscribe
and version-maintenance APIs, Section 9.5.3.4 presents RCU’s list-processing APIs,
Section 9.5.3.5 presents RCU’s diagnostic APIs, and Section 9.5.3.6 describes in which
contexts RCU’s various APIs may be used. Finally, Section 9.5.3.7 presents concluding
remarks.
Readers who are not excited about kernel internals may wish to skip ahead to
Section 9.5.4 on page 251, but preferably after reviewing the next section covering
software-engineering considerations.
9.5.3.1 RCU API and Software Engineering

Readers who have looked ahead to Tables 9.2, 9.3, 9.4, and 9.5 might have noted that
the full list of Linux-kernel APIs sports more than 100 members. This is in sharp (and
perhaps dismaying) contrast to the mere six API members shown in Table 9.1. This
situation clearly raises the question “Why so many???”
This question is answered more thoroughly in the following sections, but in the
meantime the rest of this section summarizes the motivations.
There is a wise old saying to the effect of “To err is human.” This means that purpose
of a significant fraction of the RCU API is to provide diagnostics, most notably in
Table 9.5, but elsewhere as well.
Important causes of human error are the limits of the human brain, for example, the
limited capacity of short-term memory. The toy examples shown in this book do not
stress these limits. This is out of necessity: Many readers push their cognitive limits
while learning new material, so the examples need to be kept simple.
These examples therefore keep rcu_dereference() invocations in the same function
as the enclosing rcu_read_lock() and rcu_read_unlock() calls. In contrast, real-
world software must frequently invoke these API members from different functions, and
even from different translation units. The Linux kernel RCU API has therefore expanded
to accommodate lockdep, which allows rcu_dereference() and friends to complain
if it is not protected by rcu_read_lock(). Linux-kernel RCU also checks for some
double-free errors, infinite loops in RCU read-side critical sections, and attempts to
invoke quiescent states within RCU read-side critical sections.
Another way that real-world software accommodates the limits of human cognition is
through abstraction. The Linux-kernel API therefore includes members that operate on
12 Userspace RCU’s API is documented elsewhere [MDJ13f].
v2023.06.11a
lists in addition to the pointer-oriented core API of Table 9.1. The Linux kernel itself
also provides RCU-protected hash tables and search trees.
Operating-systems kernels such as Linux operate near the bottom of the “iron triangle”
of the software stack shown in Figure 2.3, where performance is critically important.
There are thus specialized variants of a number of RCU APIs for use on fastpaths,
for example, as discussed in Section 9.5.3.3, RCU_INIT_POINTER() may be used
in place of rcu_assign_pointer() in cases where the RCU-protected pointer is
being assigned to NULL or when that pointer is not yet accessible by readers. Use of
RCU_INIT_POINTER() allows the compiler more leeway in selecting instructions and
carrying out optimizations, thus increasing performance.
On the other hand, when used incorrectly RCU_INIT_POINTER() can result in silent
memory corruption, so please be careful! Yes, in some cases, the kernel can check for
inappropriate use of RCU API members from a given kernel context, but the constraints
of RCU_INIT_POINTER() use are not yet checkable.
Finally, within the Linux kernel, the aforementioned limits of human cognition are
compounded by the variety and severity of workloads running on Linux. As of v5.16,
this has given rise to no fewer than five flavors of RCU, each designed to provide
different performance, scalability, response-time, and energy efficiency tradeoffs to
RCU readers and writers. These RCU flavors are the subject of the next section.
9.5.3.2 RCU has a Family of Wait-to-Finish APIs

The most straightforward answer to “what is RCU” is that RCU is an API. For example,
the RCU implementation used in the Linux kernel is summarized by Table 9.2, which
shows the wait-for-readers portions of the RCU, “sleepable” RCU (SRCU), Tasks RCU,
and generic APIs, respectively, and by Table 9.3, which shows the publish-subscribe
portions of the API [McK19b].13
If you are new to RCU, you might consider focusing on just one of the columns in
Table 9.2, each of which summarizes one member of the Linux kernel’s RCU API
family. For example, if you are primarily interested in understanding how RCU is used
in the Linux kernel, “RCU” would be the place to start, as it is used most frequently. On
the other hand, if you want to understand RCU for its own sake, “Tasks RCU” has the
simplest API. You can always come back for the other columns later.
If you are already familiar with RCU, these tables can serve as a useful reference.
Quick Quiz 9.39: Why do some of the cells in Table 9.2 have exclamation marks (“!”)?
The “RCU” column corresponds to the consolidation of the three Linux-kernel RCU
implementations [McK19c, McK19a], in which RCU read-side critical sections start
with rcu_read_lock(), rcu_read_lock_bh(), or rcu_read_lock_sched() and
end with rcu_read_unlock(), rcu_read_unlock_bh(), or rcu_read_unlock_
sched(), respectively. Any region of code that disables bottom halves, interrupts,
or preemption also acts as an RCU read-side critical section. RCU read-side criti-
cal sections may be nested. The corresponding synchronous update-side primitives,
synchronize_rcu() and synchronize_rcu_expedited(), along with their syn-
onym synchronize_net(), wait for any type of currently executing RCU read-side
critical sections to complete. The length of this wait is known as a “grace period”, and
synchronize_rcu_expedited() is designed to reduce grace-period latency at the
13 This citation covers v4.20 and later. Documetation for earlier versions of the Linux-
kernel RCU API may be found elsewhere [McK08e, McK14f].
v2023.06.11a
v2023.06.11a
Table 9.2: RCU Wait-to-Finish APIs
CHAPTER 9. DEFERRED PROCESSING
RCU: Original SRCU: Sleeping readers Tasks RCU: Free tracing Tasks RCU Rude: Free idle-task Tasks RCU Trace: Protect sleepable
trampolines tracing trampolines BPF programs
Initialization and DEFINE_SRCU()
Cleanup DEFINE_STATIC_SRCU()
init_srcu_struct()
cleanup_srcu_struct()
Read-side rcu_read_lock() ! srcu_read_lock() Voluntary context switch Voluntary context switch and rcu_read_lock_trace()
critical-section rcu_read_unlock() ! srcu_read_unlock() preempt-enable regions of code rcu_read_unlock_trace()
markers rcu_read_lock_bh()
rcu_read_unlock_bh()
rcu_read_lock_sched()
rcu_read_unlock_sched()
(Plus anything disabing bottom
halves, preemption, or interrupts.)
Update-side primitives synchronize_rcu() synchronize_srcu() synchronize_rcu_tasks() synchronize_rcu_tasks_rude() synchronize_rcu_tasks_trace()
(synchronous) synchronize_net() synchronize_srcu_expedited()
synchronize_rcu_expedited()
Update-side primitives call_rcu() ! call_srcu() call_rcu_tasks() call_rcu_tasks_rude() call_rcu_tasks_trace()
(asynchronous /
callback)
Update-side primitives rcu_barrier() srcu_barrier() rcu_barrier_tasks() rcu_barrier_tasks_rude() rcu_barrier_tasks_trace()
(wait for callbacks)
Update-side primitives get_state_synchronize_rcu()
(initiate / wait) cond_synchronize_rcu()
Update-side primitives kfree_rcu()
(free memory)
Type-safe memory SLAB_TYPESAFE_BY_RCU
Read side constraints No blocking (only preemption) No synchronize_srcu() with No voluntary context switch Neither blocking nor preemption No RCU tasks trace grace period
same srcu_struct
Read side overhead CPU-local accesses (barrier() Simple instructions, memory Free CPU-local accesses (free on CPU-local accesses
on PREEMPT=n) barriers PREEMPT=n)
Asynchronous sub-microsecond sub-microsecond sub-microsecond sub-microsecond sub-microsecond
update-side overhead
Grace-period latency 10s of milliseconds Milliseconds Seconds Milliseconds 10s of milliseconds
Expedited 10s of microseconds Microseconds N/A N/A N/A
240
grace-period latency
expense of increased CPU overhead and IPIs. The asynchronous update-side primitive,
call_rcu(), invokes a specified function with a specified argument after a subsequent
grace period. For example, call_rcu(p,f); will result in the “RCU callback” f(p)
being invoked after a subsequent grace period. There are situations, such as when
unloading a Linux-kernel module that uses call_rcu(), when it is necessary to wait for
all outstanding RCU callbacks to complete [McK07e]. The rcu_barrier() primitive
does this job.
Quick Quiz 9.40: How do you prevent a huge number of RCU read-side critical sections from
indefinitely blocking a synchronize_rcu() invocation?
Quick Quiz 9.41: The synchronize_rcu() API waits for all pre-existing interrupt handlers
to complete, right?
Quick Quiz 9.42: What is the difference between synchronize_rcu() and rcu_barrier()?
Finally, RCU may be used to provide type-safe memory [GC96], as described in

Section 9.5.4.5. In the context of RCU, type-safe memory guarantees that a given data
element will not change type during any RCU read-side critical section that accesses
it. To make use of RCU-based type-safe memory, pass SLAB_TYPESAFE_BY_RCU to
kmem_cache_create().
The “SRCU” column in Table 9.2 displays a specialized RCU API that permits general
sleeping in SRCU read-side critical sections [McK06] delimited by srcu_read_lock()
and srcu_read_unlock(). However, unlike RCU, SRCU’s srcu_read_lock()
returns a value that must be passed into the corresponding srcu_read_unlock().
This difference is due to the fact that the SRCU user allocates an srcu_struct for each
distinct SRCU usage, so that there is no convenient place to store a per-task reader-nesting
count. (Keep in mind that although the Linux kernel provides dynamically allocated
per-CPU storage, there is not yet dynamically allocated per-task storage.)
A given srcu_struct structure may be defined as a global variable with DEFINE_
SRCU() if the structure must be used in multiple translation units, or with DEFINE_
STATIC_SRCU() otherwise. For example, DEFINE_SRCU(my_srcu) would create
a global variable named my_srcu that could be used by any file in the program.
Alternatively, an srcu_struct structure may be either an on-stack variable or a
dynamically allocated region of memory. In both of these non-global-variable cases,
the memory must be initialized using init_srcu_struct() prior to its first use and
cleaned up using cleanup_srcu_struct() after its last use (but before the underlying
storage disappears).
However they are created, these distinct srcu_struct structures prevent
SRCU read-side critical sections from blocking unrelated synchronize_srcu()
and synchronize_srcu_expedited() invocations. Of course, use of either
synchronize_srcu() or synchronize_srcu_expedited() within an SRCU read-
side critical section can result in self-deadlock, so should be avoided. As with RCU,
SRCU’s synchronize_srcu_expedited() decreases grace-period latency compared
to synchronize_srcu(), but at the expense of increased CPU overhead.
Quick Quiz 9.43: Under what conditions can synchronize_srcu() be safely used within
an SRCU read-side critical section?
Similar to normal RCU, self-deadlock can be avoided using the asynchronous call_
srcu() function. However, special care must be taken when using call_srcu()
v2023.06.11a
because a single task could register SRCU callbacks very quickly. Given that SRCU
allows readers to block for arbitrary periods of time, this could consume an arbitrarily
large quantity of memory. In contrast, given the synchronous synchronize_srcu()
interface, a given task must finish waiting for a given grace period before it can start
waiting for the next one.
Also similar to RCU, there is an srcu_barrier() function that waits for all prior
call_srcu() callbacks to be invoked.
In other words, SRCU compensates for its extremely weak forward-progress guarantees
by permitting the developer to restrict its scope.
The “Tasks RCU” column in Table 9.2 displays a specialized RCU API that mediates
freeing of the trampolines used in Linux-kernel tracing. These trampolines are used to
transfer control from a point in the code being traced to the code doing the actual tracing.
It is of course necessary to ensure that all code executing within a given trampoline has
finished before freeing that trampoline.
Changes to the code being traced are typically limited to a single jump or call
instruction, and thus cannot accommodate the sequence of code required to implement
rcu_read_lock() and rcu_read_unlock(). Nor can the trampoline contain these
calls to rcu_read_lock() and rcu_read_unlock(). To see this, consider a CPU
that is just about to start executing a given trampoline. Because it has not yet executed
the rcu_read_lock(), that trampoline could be freed at any time, which would
come as a fatal surprise to this CPU. Therefore, trampolines cannot be protected by
synchronization primitives executed in either the traced code or in the trampoline itself.
Which does raise the question of exactly how the trampoline is to be protected.
The key to answering this question is to note that trampoline code never contains
code that either directly or indirectly does a voluntary context switch. This code might
be preempted, but it will never directly or indirectly invoke schedule(). This suggests
a variant of RCU having voluntary context switches and idle execution as its only
quiescent states. This variant is Tasks RCU.
Tasks RCU is unusual in having no read-side marking functions, which is good given
that its main use case has nowhere to put such markings. Instead, calls to schedule()
serve directly as quiescent states. Updates can use synchronize_rcu_tasks() to wait
for all pre-existing trampoline execution to complete, or they can use its asynchronous
counterpart, call_rcu_tasks(). There is also an rcu_barrier_tasks() that waits
for completion of callbacks corresponding to all prior invocations of call_rcu_
tasks(). There is no synchronize_rcu_tasks_expedited() because there has
not yet been a request for it, though implementing a useful variant of it would not be
free of challenges.
Quick Quiz 9.44: In a kernel built with CONFIG_PREEMPT_NONE=y, won’t synchronize_

rcu() wait for all trampolines, given that preemption is disabled and that trampolines never
directly or indirectly invoke schedule()?
The “Tasks RCU Rude” column provides a more effective variant of the toy imple-
mentation presented in Section 9.5.1.4. This variant causes each CPU to execute a
context switch, so that any voluntary context switch or any preemptible region of code
can serve as a quiescent state. The Tasks RCU Rude variant uses the Linux-kernel
workqueues facility to force concurrent context switches, in contrast to the serial
CPU-by-CPU approach taken by the toy implementation. The API mirrors that of Tasks
RCU, including the lack of explicit read-side markers.
v2023.06.11a
Table 9.3: RCU Publish-Subscribe and Version Maintenance APIs
Category Primitives Overhead
Pointer publish rcu_assign_pointer() Memory barrier

rcu_replace_pointer() Memory barrier (two of them on Alpha)
rcu_pointer_handoff() Simple instructions
RCU_INIT_POINTER() Simple instructions
RCU_POINTER_INITIALIZER() Compile-time constant
Pointer subscribe (traversal) rcu_access_pointer() Simple instructions

rcu_dereference() Simple instructions (memory barrier on Alpha)
rcu_dereference_check() Simple instructions (memory barrier on Alpha)
rcu_dereference_protected() Simple instructions
rcu_dereference_raw() Simple instructions (memory barrier on Alpha)
rcu_dereference_raw_notrace() Simple instructions (memory barrier on Alpha)
Finally, the “Tasks RCU Trace” column provides an RCU implementation with
functionality similar to that of SRCU, except with much faster read-side markers.14
However, this speed is a consequence of the fact that these markers do not execute
memory-barrier instructions, which means that Tasks RCU Trace grace periods must
often send IPIs to all CPUs and must always scan the entire task list, thus degrading
real-time response and consuming considerable CPU time. Nevertheless, in the absence
of readers, the resulting grace-period latency is reasonably short, rivaling that of RCU.
9.5.3.3 RCU has Publish-Subscribe and Version-Maintenance APIs

Fortunately, the RCU publish-subscribe and version-maintenance primitives shown in
Table 9.3 apply to all of the variants of RCU discussed above. This commonality can
allow more code to be shared, and reduces API proliferation. The original purpose of
the RCU publish-subscribe APIs was to bury memory barriers into these APIs, so that
Linux kernel programmers could use RCU without needing to become expert on the
memory-ordering models of each of the 20+ CPU families that Linux supports [Spr01].
These primitives operate directly on pointers, and are useful for creating RCU-
protected linked data structures, such as RCU-protected arrays and trees. The special
case of linked lists is handled by a separate set of APIs described in Section 9.5.3.4.
The first category publishes pointers to new data items. The rcu_assign_
pointer() primitive ensures that any prior initialization remains ordered before the
assignment to the pointer on weakly ordered machines. The rcu_replace_pointer()
primitive updates the pointer just like rcu_assign_pointer() does, but also returns
the previous value, just like rcu_dereference_protected() (see below) would,
including the lockdep expression. This replacement is convenient when the updater
must both publish a new pointer and free the structure referenced by the old pointer.
Quick Quiz 9.45: Normally, any pointer subject to rcu_dereference() must always be
updated using one of the pointer-publish functions in Table 9.3, for example, rcu_assign_
pointer().
What is an exception to this rule?
14 And thus is unusual for the Tasks RCU family for having explicit read-side markers!
v2023.06.11a
Quick Quiz 9.46: Are there any downsides to the fact that these traversal and update primitives
can be used with any of the RCU API family members?
The rcu_pointer_handoff() primitive simply returns its sole argument, but is

useful to tooling checking for pointers being leaked from RCU read-side critical sections.
Use of rcu_pointer_handoff() indicates to such tooling that protection of the
structure in question has been handed off from RCU to some other mechanism, such as
locking or reference counting.
The RCU_INIT_POINTER() macro can be used to initialize RCU-protected pointers
that have not yet been exposed to readers, or alternatively, to set RCU-protected
pointers to NULL. In these restricted cases, the memory-barrier instructions provided by
rcu_assign_pointer() are not needed. Similarly, RCU_POINTER_INITIALIZER()
provides a GCC-style structure initializer to allow easy initialization of RCU-protected
pointers in structures.
The second category subscribes to pointers to data items, or, alternatively, safely
traverses RCU-protected pointers. Again, simply loading these pointers using C-language
accesses could result in seeing pre-initialization garbage in the pointed-to data. Similarly,
loading these pointer by any means outside of an RCU read-side critical section could
result in the pointed-to object being freed at any time. However, if the pointer is merely
to be tested and not dereferenced, the freeing of the pointed-to object is not necessarily
a problem. In this case, rcu_access_pointer() may be used. Normally, however,
RCU read-side protection is required, and so the rcu_dereference() primitive uses
the Linux kernel’s lockdep facility [Cor06a] to verify that this rcu_dereference()
invocation is under the protection of rcu_read_lock(), srcu_read_lock(), or some
other RCU read-side marker. In contrast, the rcu_access_pointer() primitive does
not involve lockdep, and thus will not provoke lockdep complaints when used outside
of an RCU read-side critical section.
Another situation where protection is not required is when update-side code accesses
the RCU-protected pointer while holding the update-side lock. The rcu_dereference_
protected() API member is provided for this situation. Its first parameter is the
RCU-protected pointer, and the second parameter takes a lockdep expression describing
which locks must be held in order for the access to be safe. Code invoked both from
readers and updaters can use rcu_dereference_check(), which also takes a lockdep
expression, but which may also be invoked from read-side code not holding the locks.
In some cases, the lockdep expressions can be very complex, for example, when
using fine-grained locking, any of a very large number of locks might be held, and
it might be quite difficult to work out which applies. In these (hopefully rare) cases,
rcu_dereference_raw() provides protection but does not check for being invoked
within a reader or with any particular lock being held. The rcu_dereference_raw_
notrace() API member acts similarly, but cannot be traced, and may therefore be
safely used by tracing code.
Although pretty much any linked structure can be accessed by manipulating pointers,
higher-level structures can be quite helpful. The next section therefore looks at various
sorts of RCU-protected linked lists used by the Linux kernel.
9.5.3.4 RCU has List-Processing APIs

Although rcu_assign_pointer() and rcu_dereference() can in theory be used
to construct any conceivable RCU-protected data structure, in practice it is often
better to use higher-level constructs. Therefore, the rcu_assign_pointer() and
v2023.06.11a
next next next next

prev prev prev prev
A B C
Figure 9.16: Linux Circular Linked List (list)
A B C
Figure 9.17: Linux Linked List Abbreviated
first next next next

pprev pprev pprev
A B C
Figure 9.18: Linux Linear Linked List (hlist)
rcu_dereference() primitives have been embedded in special RCU variants of

Linux’s list-manipulation API. Linux has four variants of doubly linked list, the circu-
lar struct list_head and the linear struct hlist_head/struct hlist_node,
struct hlist_nulls_head/struct hlist_nulls_node, and struct hlist_
bl_head/struct hlist_bl_node pairs. The former is laid out as shown in Fig-
ure 9.16, where the green (leftmost) boxes represent the list header and the blue
(rightmost three) boxes represent the elements in the list. This notation is cumbersome,
and will therefore be abbreviated as shown in Figure 9.17, which shows only the
non-header (blue) elements.
Linux’s hlist15 is a linear list, which means that it needs only one pointer for the
header rather than the two required for the circular list, as shown in Figure 9.18. Thus,
use of hlist can halve the memory consumption for the hash-bucket arrays of large hash
tables. As before, this notation is cumbersome, so hlist structures will be abbreviated
in the same way list_head-style lists are, as shown in Figure 9.17.
A variant of Linux’s hlist, named hlist_nulls, provides multiple distinct NULL
pointers, but otherwise uses the same layout as shown in Figure 9.18. In this variant, a
->next pointer having a zero low-order bit is considered to be a pointer. However, if
the low-order bit is set to one, the upper bits identify the type of NULL pointer. This type
of list is used to allow lockless readers to detect when a node has been moved from one
list to another. For example, each bucket of a hash table might use its index to mark its
NULL pointer. Should a reader encounter a NULL pointer not matching the index of the
bucket it started from, that reader knows that an element it was traversing was moved
to some other bucket during the traversal, taking that reader with it. The reader can
use the is_a_nulls() function (which returns true if passed an hlist_nulls NULL
pointer) to determine when it reaches the end of a list, and the get_nulls_value()
function (which returns its argument’s NULL-pointer identifier) to fetch the type of NULL
15 The “h” stands for hashtable, in which it reduces memory use by half compared to
Linux’s double-pointer circular linked list.
v2023.06.11a
pointer. When get_nulls_value() returns an unexpected value, the reader can take
corrective action, for example, restarting its traversal from the beginning.
Quick Quiz 9.47: But what if an hlist_nulls reader gets moved to some other bucket and
then back again?
More information on hlist_nulls is available in the Linux-kernel source tree, with

helpful example code provided in the rculist_nulls.rst file (rculist_nulls.txt
in older kernels).
Another variant of Linux’s hlist incorporates bit-locking, and is named hlist_bl.
This variant uses the same layout as shown in Figure 9.18, but reserves the low-order bit
of the head pointer (“first” in the figure) to lock the list. This approach also reduces
memory usage, as it allows what would otherwise be a separate spinlock to be stored
with the pointer itself.
The API members for these linked-list variants are summarized in Table 9.4. More
information is available in the Documentation/RCU directory of the Linux-kernel
source tree and at Linux Weekly News [McK19b].
However, the remainder of this section expands on the use of list_replace_rcu(),
given that this API member gave RCU its name. This API member is used to carry out
more complex updates in which an element in the middle of the list having multiple
fields is atomically updated, so that a given reader sees either the old set of values or the
new set of values, but not a mixture of the two sets. For example, each node of a linked
list might have integer fields ->a, ->b, and ->c, and it might be necessary to update a
given node’s fields from 5, 6, and 7 to 5, 2, and 3, respectively.
The code implementing this atomic update is straightforward:
15 q = kmalloc(sizeof(*p), GFP_KERNEL);
16 *q = *p;
17 q->b = 2;
18 q->c = 3;
19 list_replace_rcu(&p->list, &q->list);
21 kfree(p);
The following discussion walks through this code, using Figure 9.19 to illustrate
the state changes. The triples in each element represent the values of fields ->a,
->b, and ->c, respectively. The red-shaded elements might be referenced by readers,
and because readers do not synchronize directly with updaters, readers might run
concurrently with this entire replacement process. Please note that backwards pointers
and the link from the tail to the head are omitted for clarity.
The initial state of the list, including the pointer p, is the same as for the deletion
example, as shown on the first row of the figure.
The following text describes how to replace the 5,6,7 element with 5,2,3 in such a
way that any given reader sees one of these two values.
Line 15 allocates a replacement element, resulting in the state as shown in the second
row of Figure 9.19. At this point, no reader can hold a reference to the newly allocated
element (as indicated by its green shading), and it is uninitialized (as indicated by the
question marks).
Line 16 copies the old element to the new one, resulting in the state as shown in the
third row of Figure 9.19. The newly allocated element still cannot be referenced by
readers, but it is now initialized.
v2023.06.11a
Table 9.4: RCU-Protected List APIs
list
list: Circular doubly linked list hlist
hlist: Linear doubly linked list hlist_nulls
hlist_nulls: Linear doubly linked list hlist_bl
hlist_bl: Linear doubly linked list
with marked NULL pointer, with up to with bit locking
31 bits of marking
Structures
struct list_head struct hlist_head struct hlist_nulls_head struct hlist_bl_head
struct hlist_node struct hlist_nulls_node struct hlist_bl_node
Initialization
INIT_LIST_HEAD_RCU()
Full traversal
list_for_each_entry_rcu() hlist_for_each_entry_rcu() hlist_nulls_for_each_entry_rcu() hlist_bl_for_each_entry_rcu()
list_for_each_entry_lockless() hlist_for_each_entry_rcu_bh() hlist_nulls_for_each_entry_safe()
hlist_for_each_entry_rcu_notrace()
Resume traversal
9.5. READ-COPY UPDATE (RCU)
list_for_each_entry_continue_rcu() hlist_for_each_entry_continue_rcu()
list_for_each_entry_from_rcu() hlist_for_each_entry_continue_rcu_bh()
hlist_for_each_entry_from_rcu()
Stepwise traversal
list_entry_rcu() hlist_first_rcu() hlist_nulls_first_rcu() hlist_bl_first_rcu()
list_entry_lockless() hlist_next_rcu() hlist_nulls_next_rcu()
list_first_or_null_rcu() hlist_pprev_rcu()
list_next_rcu()
list_next_or_null_rcu()
Add
list_add_rcu() hlist_add_before_rcu() hlist_nulls_add_head_rcu() hlist_bl_add_head_rcu()
list_add_tail_rcu() hlist_add_behind_rcu() hlist_bl_set_first_rcu()
hlist_add_head_rcu()
hlist_add_tail_rcu()
Delete
list_del_rcu() hlist_del_rcu() hlist_nulls_del_rcu() hlist_bl_del_rcu()
hlist_del_init_rcu() hlist_nulls_del_init_rcu() hlist_bl_del_init_rcu()
Replace
247
list_replace_rcu() hlist_replace_rcu()
Splice
list_splice_init_rcu() list_splice_tail_init_rcu()
v2023.06.11a
1,2,3 5,6,7 11,4,8
Allocate
?,?,?
1,2,3 5,6,7 11,4,8
Copy
5,6,7
1,2,3 5,6,7 11,4,8
Update
5,2,3
1,2,3 5,6,7 11,4,8
list_replace_rcu()
5,2,3
1,2,3 5,6,7 11,4,8
synchronize_rcu()
5,2,3
1,2,3 5,6,7 11,4,8
kfree()
1,2,3 5,2,3 11,4,8
Figure 9.19: RCU Replacement in Linked List
v2023.06.11a
Table 9.5: RCU Diagnostic APIs
Category Primitives
Mark RCU pointer __rcu
Debug-object support init_rcu_head()

destroy_rcu_head()
init_rcu_head_on_stack()
destroy_rcu_head_on_stack()
Stall-warning control rcu_cpu_stall_reset()
Callback checking rcu_head_init()

rcu_head_after_call_rcu()
lockdep support rcu_read_lock_held()

rcu_read_lock_bh_held()
rcu_read_lock_sched_held()
srcu_read_lock_held()
rcu_is_watching()
RCU_LOCKDEP_WARN()
RCU_NONIDLE()
rcu_sleep_check()
Line 17 updates q->b to the value “2”, and line 18 updates q->c to the value “3”, as
shown on the fourth row of Figure 9.19. Note that the newly allocated structure is still
inaccessible to readers.
Now, line 19 does the replacement, so that the new element is finally visible to readers,
and hence is shaded red, as shown on the fifth row of Figure 9.19. At this point, as
shown below, we have two versions of the list. Pre-existing readers might see the 5,6,7
element (which is therefore now shaded yellow), but new readers will instead see the
5,2,3 element. But any given reader is guaranteed to see one set of values or the other,
not a mixture of the two.
After the synchronize_rcu() on line 20 returns, a grace period will have elapsed,
and so all reads that started before the list_replace_rcu() will have completed. In
particular, any readers that might have been holding references to the 5,6,7 element are
guaranteed to have exited their RCU read-side critical sections, and are thus prohibited
from continuing to hold a reference. Therefore, there can no longer be any readers
holding references to the old element, as indicated its green shading in the sixth row of
Figure 9.19. As far as the readers are concerned, we are back to having a single version
of the list, but with the new element in place of the old.
After the kfree() on line 21 completes, the list will appear as shown on the final
row of Figure 9.19.
Despite the fact that RCU was named after the replacement case, the vast majority
of RCU usage within the Linux kernel relies on the simple independent insertion and
deletion, as was shown in Figure 9.15 in Section 9.5.2.3.
The next section looks at APIs that assist developers in debugging their code that
makes use of RCU.
9.5.3.5 RCU Has Diagnostic APIs

Table 9.5 shows RCU’s diagnostic APIs.
v2023.06.11a
The __rcu tag marks an RCU-protected pointer, for example,

“struct foo __rcu *p;”. Pointers that might be passed to rcu_dereference()
can be marked, but pointers holding values returned from rcu_dereference() should
not be. Providing these markings on variables, structure fields, function parameters,
and return values allows the Linux kernel’s sparse tool to detect situations where
RCU-protected pointers are incorrectly accessed using plain C-language loads and
stores.
Debug-object support is automatic for any rcu_head structures that are part of a struc-
ture obtained from the Linux kernel’s memory allocators, but those building their own
special-purpose memory allocators can use init_rcu_head() and destroy_rcu_
head() at allocation and free time, respectively. Those using rcu_head structures allo-
cated on the function-call stack (it happens!) may use init_rcu_head_on_stack()
before first use and destroy_rcu_head_on_stack() after last use, but before return-
ing from the function. Debug-object support allows detection of bugs involving passing
the same rcu_head structure to call_rcu() and friends in quick succession, which is
the call_rcu() counterpart to the infamous double-free class of memory-allocation
bugs.
Stall-warning control is provided by rcu_cpu_stall_reset(), which allows the
caller to suppress RCU CPU stall warnings for the remainder of the current grace
period. RCU CPU stall warnings help pinpoint situations where an RCU read-side
critical section runs for an excessive length of time, and it is useful for things like kernel
debuggers to be able to suppress them, for example, when encountering a breakpoint.
Callback checking is provided by rcu_head_init() and rcu_head_after_call_
rcu(). The former is invoked on an rcu_head structure before it is passed to call_
rcu(), and then rcu_head_after_call_rcu() will check to see if the callback has
been invoked with the specified function.
Support for lockdep [Cor06a] includes rcu_read_lock_held(), rcu_read_lock_
bh_held(), rcu_read_lock_sched_held(), and srcu_read_lock_held(), each
of which returns true if invoked within the corresponding type of RCU read-side
critical section.
Quick Quiz 9.48: Why isn’t there a rcu_read_lock_tasks_held() for Tasks RCU?
Because rcu_read_lock() cannot be used from the idle loop, and because energy-
efficiency concerns have caused the idle loop to become quite ornate, rcu_is_
watching() returns true if invoked in a context where use of rcu_read_lock() is
legal. Note again that srcu_read_lock() may be used from idle and even offline
CPUs, which means that rcu_is_watching() does not apply to SRCU.
RCU_LOCKDEP_WARN() emits a warning if lockdep is enabled and if its argument
evaluates to true. For example, RCU_LOCKDEP_WARN(!rcu_read_lock_held())
would emit a warning if invoked outside of an RCU read-side critical section.
RCU_NONIDLE() may be used to force RCU to watch when executing the statement
that is passed in as the sole argument. For example, RCU_NONIDLE(WARN_ON(!rcu_
is_watching())) would never emit a warning. However, changes in the 2020–2021
timeframe extend RCU’s reach deeper into the idle loop, which should greatly reduce or
even eliminate the need for RCU_NONIDLE().
Finally, rcu_sleep_check() emits a warning if invoked within an RCU, RCU-bh,
or RCU-sched read-side critical section.
v2023.06.11a
NMI
RCU List Traversal

rcu_read_unlock()
rcu_dereference()
rcu_read_lock()
rcu_assign_pointer()
RCU List Mutation

IRQ
call_rcu()
Process synchronize_rcu()
Figure 9.20: RCU API Usage Constraints
9.5.3.6 Where Can RCU’s APIs Be Used?

Figure 9.20 shows which APIs may be used in which in-kernel environments. The
RCU read-side primitives may be used in any environment, including NMI, the RCU
mutation and asynchronous grace-period primitives may be used in any environment
other than NMI, and, finally, the RCU synchronous grace-period primitives may
be used only in process context. The RCU list-traversal primitives include list_
for_each_entry_rcu(), hlist_for_each_entry_rcu(), etc. Similarly, the RCU
list-mutation primitives include list_add_rcu(), hlist_del_rcu(), etc.
Note that primitives from other families of RCU may be substituted, for example,
srcu_read_lock() may be used in any context in which rcu_read_lock() may be
used.
9.5.3.7 So, What is RCU Really?

At its core, RCU is nothing more nor less than an API that supports publication and
subscription for insertions, waiting for all RCU readers to complete, and maintenance
of multiple versions. That said, it is possible to build higher-level constructs on top of
RCU, including the reader-writer-locking, reference-counting, and existence-guarantee
constructs listed in Section 9.5.4. Furthermore, I have no doubt that the Linux community
will continue to find interesting new uses for RCU, just as they do for any of a number
of synchronization primitives throughout the kernel.
Of course, a more-complete view of RCU would also include all of the things you
can do with these APIs.
However, for many people, a complete view of RCU must include sample RCU
implementations. Appendix B therefore presents a series of “toy” RCU implementations
of increasing complexity and capability, though others might prefer the classic “User-
Level Implementations of Read-Copy Update” [DMS+ 12]. For everyone else, the next
section gives an overview of some RCU use cases.
9.5.4 RCU Usage

This section answers the question “What is RCU?” from the viewpoint of the uses to
which RCU can be put. Because RCU is most frequently used to replace some existing
mechanism, we look at it primarily in terms of its relationship to such mechanisms, as
v2023.06.11a
Table 9.6: RCU Usage
Mechanism RCU Replaces Page
RCU for pre-BSD routing 252

Wait for pre-existing things to finish 254
Phased state change 257
Add-only list (publish/subscribe) 259
Type-safe memory 259
Existence Guarantee 260
Light-weight garbage collector 261
Delete-only list 262
Quasi reader-writer lock 262
Quasi reference count 272
Quasi multi-version concurrency control (MVCC) 274
2.5x107
2x107
ideal
1.5x107
RCU
1x107
seqlock
5x106
hazptr
0
0 50 100 150 200 250 300 350 400 450
Figure 9.21: Pre-BSD Routing Table Protected by RCU
listed in Table 9.6 and as displayed in Figure 9.23. Following the sections listed in this
table, Section 9.5.4.12 provides a summary.
9.5.4.1 RCU for Pre-BSD Routing

In contrast to the later sections, this section focuses on a very specific use case for the
purpose of comparison with other mechanisms.
Listings 9.14 and 9.15 show code for an RCU-protected Pre-BSD routing table
(route_rcu.c). The former shows data structures and route_lookup(), and the
latter shows route_add() and route_del().
In Listing 9.14, line 2 adds the ->rh field used by RCU reclamation, line 6 adds
the ->re_freed use-after-free-check field, lines 16, 22, and 26 add RCU read-side
protection, and lines 20 and 21 add the use-after-free check. In Listing 9.15, lines 11,
13, 30, 34, and 39 add update-side locking, lines 12 and 33 add RCU update-side
protection, line 35 causes route_cb() to be invoked after a grace period elapses, and
lines 17–24 define route_cb(). This is minimal added code for a working concurrent
implementation.
v2023.06.11a
Listing 9.14: RCU Pre-BSD Routing Table Lookup

2 struct rcu_head rh;
3 struct cds_list_head re_next;
6 int re_freed;
7 };
8 CDS_LIST_HEAD(route_list);
10
12 {
15
16 rcu_read_lock();
17 cds_list_for_each_entry_rcu(rep, &route_list, re_next) {
21 abort();
23 return ret;
24 }
25 }
28 }
Listing 9.15: RCU Pre-BSD Routing Table Add/Delete

2 {
4
6 if (!rep)
7 return -ENOMEM;
8 rep->addr = addr;
12 cds_list_add_rcu(&rep->re_next, &route_list);
14 return 0;
15 }
16
17 static void route_cb(struct rcu_head *rhp)
18 {
20
21 rep = container_of(rhp, struct route_entry, rh);
22 WRITE_ONCE(rep->re_freed, 1);
23 free(rep);
24 }
25
27 {
29
33 cds_list_del_rcu(&rep->re_next);
35 call_rcu(&rep->rh, route_cb);
36 return 0;
37 }
38 }
40 return -ENOENT;
41 }
v2023.06.11a
2.5x107
2x107 RCU-QSBR

1.5x10
7 ideal
RCU
1x107
seqlock
5x106
hazptr
0
0 50 100 150 200 250 300 350 400 450
Figure 9.22: Pre-BSD Routing Table Protected by RCU QSBR
Figure 9.21 shows the performance on the read-only workload. RCU scales quite
well, and offers nearly ideal performance. However, this data was generated using
the RCU_SIGNAL flavor of userspace RCU [Des09b, MDJ13f], for which rcu_read_
lock() and rcu_read_unlock() generate a small amount of code. What happens for
the QSBR flavor of RCU, which generates no code at all for rcu_read_lock() and
rcu_read_unlock()? (See Section 9.5.1, and especially Figure 9.8, for a discussion
of RCU QSBR.)
The answer to this is shown in Figure 9.22, which shows that RCU QSBR’s perfor-
mance and scalability actually exceeds that of the ideal synchronization-free workload.
Quick Quiz 9.49: Wait, what??? How can RCU QSBR possibly be better than ideal? Just
what rubbish definition of ideal would fail to be the best of all possible results???
Quick Quiz 9.50: Given RCU QSBR’s read-side performance, why bother with any other
flavor of userspace RCU?
Although Pre-BSD routing is an excellent RCU use case, it is worthwhile looking at

the relationships betweeen the wider spectrum of use cases shown in Figure 9.23. This
task is taken up by the following sections.
While reading these sections, please ask yourself which of these use cases best
describes Pre-BSD routing.
9.5.4.2 Wait for Pre-Existing Things to Finish

As noted in Section 9.5.2 an important component of RCU is a way of waiting for RCU
readers to finish. One of RCU’s great strength is that it allows you to wait for each of
thousands of different things to finish without having to explicitly track each and every
one of them, and without incurring the performance degradation, scalability limitations,
complex deadlock scenarios, and memory-leak hazards that are inherent in schemes
that use explicit tracking.
In this section, we will show how synchronize_sched()’s read-side counterparts
(which include anything that disables preemption, along with hardware operations and
primitives that disable interrupts) permit you to interaction with non-maskable interrupt
(NMI) handlers, which is quite difficult using locking. This approach has been called
“Pure RCU” [McK04], and it is used in a few places in the Linux kernel.
v2023.06.11a
Quasi Reader-Writer Lock
+ Readers as read-held reader-writer lock

+ Spatial as well as temporal synchronization
+ Optional read-to-write upgrade
+ Optional bridging to per-object lock or reference
+ Optionally ignore deleted objects
Quasi Reference Count
+ Readers as individual or bulk unconditional references

+ Optional bridging to per-object lock or reference
Quasi Multi-Version Consistency Control
+ Readers include some sort of snapshot operation

+ Constraints on readers and writers:
+ (1) single object, (2) sequence locks, (3) version number(s),
+ (4) Issaquah challenge, and/or (5) many other approaches
Light-Weight Garbage Collector for Delete-Only List

Non-Blocking Synchronization (NBS)
- Publish/subscribe
+ NBS
Type-Safe Memory Existence Guarantee
+ Slab allocator + Heap allocator

+ Deferred slab reclamation + Deferred reclamation
Publish/Subscribe Add-Only Phased State Change

For Linked Structure List
+ Checked state variable
rcu_assign_pointer() &
rcu_dereference()
Wait for Pre-Existing Things to Finish
rcu_read_lock() & rcu_read_unlock()

vs. synchronize_rcu()
Figure 9.23: Relationships Between RCU Use Cases
v2023.06.11a
Listing 9.16: Using RCU to Wait for NMIs to Finish

1 struct profile_buffer {
2 long size;
3 atomic_t entry[0];
4 };
5 static struct profile_buffer *buf = NULL;
6
7 void nmi_profile(unsigned long pcvalue)
8 {
9 struct profile_buffer *p = rcu_dereference(buf);
10
11 if (p == NULL)
12 return;
13 if (pcvalue >= p->size)
14 return;
15 atomic_inc(&p->entry[pcvalue]);
16 }
17
18 void nmi_stop(void)
19 {
20 struct profile_buffer *p = buf;
21
22 if (p == NULL)
23 return;
24 rcu_assign_pointer(buf, NULL);
25 synchronize_sched();
26 kfree(p);
27 }
The basic form of such “Pure RCU” designs is as follows:
1. Make a change, for example, to the way that the OS reacts to an NMI.
2. Wait for all pre-existing read-side critical sections to completely finish (for example,
by using the synchronize_sched() primitive).16 The key observation here is
that subsequent RCU read-side critical sections are guaranteed to see whatever
change was made.
3. Clean up, for example, return status indicating that the change was successfully
made.
The remainder of this section presents example code adapted from the Linux kernel.
In this example, the nmi_stop() function in the now-defunct oprofile facility uses
synchronize_sched() to ensure that all in-flight NMI notifications have completed
before freeing the associated resources. A simplified version of this code is shown in
Listing 9.16.
Lines 1–4 define a profile_buffer structure, containing a size and an indefinite
array of entries. Line 5 defines a pointer to a profile buffer, which is presumably
initialized elsewhere to point to a dynamically allocated region of memory.
Lines 7–16 define the nmi_profile() function, which is called from within an
NMI handler. As such, it cannot be preempted, nor can it be interrupted by a normal
interrupt handler, however, it is still subject to delays due to cache misses, ECC errors,
and cycle stealing by other hardware threads within the same core. Line 9 gets a local
pointer to the profile buffer using the rcu_dereference() primitive to ensure memory
ordering on DEC Alpha, and lines 11 and 12 exit from this function if there is no profile
buffer currently allocated, while lines 13 and 14 exit from this function if the pcvalue
16 In Linux kernel v5.1 and later, synchronize_sched() has been subsumed into
synchronize_rcu().
v2023.06.11a
argument is out of range. Otherwise, line 15 increments the profile-buffer entry indexed
by the pcvalue argument. Note that storing the size with the buffer guarantees that the
range check matches the buffer, even if a large buffer is suddenly replaced by a smaller
one.
Lines 18–27 define the nmi_stop() function, where the caller is responsible for
mutual exclusion (for example, holding the correct lock). Line 20 fetches a pointer to the
profile buffer, and lines 22 and 23 exit the function if there is no buffer. Otherwise, line 24
NULLs out the profile-buffer pointer (using the rcu_assign_pointer() primitive to
maintain memory ordering on weakly ordered machines), and line 25 waits for an RCU
Sched grace period to elapse, in particular, waiting for all non-preemptible regions of
code, including NMI handlers, to complete. Once execution continues at line 26, we
are guaranteed that any instance of nmi_profile() that obtained a pointer to the old
buffer has returned. It is therefore safe to free the buffer, in this case using the kfree()
primitive.
Quick Quiz 9.51: Suppose that the nmi_profile() function was preemptible. What would
need to change to make this example work correctly?
In short, RCU makes it easy to dynamically switch among profile buffers (you just
try doing this efficiently with atomic operations, or at all with locking!). This is a rare
use of RCU in its pure form. RCU is normally used at higher levels of abstraction, as
will be shown in the following sections.
9.5.4.3 Phased State Change
Figure 9.24 shows a timeline for an example phased state change to efficiently handle
maintenance operations. If there is no maintenance operation in progress, common-case
operations must proceed quickly, for example, without acquiring a reader-writer lock.
However, if there is a maintenance operation in progress, the common-case operations
must be undertaken carefully, taking into account added complexities due to their
running concurrently with that maintenance operation. This means that common-case
operations will incur higher overhead during maintenance operations, which is one
reason that maintenance operations are normally scheduled to take place during times
of low load.
In the figure, these apparently conflicting requirements are resolved by having a
prepare phase prior to the maintenance operation and a cleanup phase after it, during
which the common-case operations can proceed either quickly or carefully.
Example pseudo-code for this phased state change is shown in Listing 9.17. The
common-case operations are carried out by cco() within an RCU read-side critical
section extending from line 5 to line 10. Here, line 6 checks a global be_careful flag,
invoking cco_carefully() or cco_quickly(), as indicated.
This allows the maint() function to set the be_careful flag on line 15 and wait for
an RCU grace period on line 16. When control reaches line 17, all cco() functions that
saw a false value of be_careful (and thus which might invoke the cco_quickly()
function) will have completed their operations, so that all currently executing cco()
functions will be invoking cco_carefully(). This means that it is safe for the
do_maint() function to be invoked. Line 18 then waits for all cco() functions that
might have run concurrently with do_maint() to complete, and finally line 19 sets the
be_careful flag back to false.
v2023.06.11a
Common-Case Maintenance
Operations Operations
Time
Quickly
Either Prepare
Carefully Maintenance
Either Clean up
Quickly
Figure 9.24: Phased State Change for Maintenance Operation
Listing 9.17: Phased State Change for Maintenance Operations

1 bool be_careful;
2
3 void cco(void)
4 {
5 rcu_read_lock();
6 if (READ_ONCE(be_careful))
7 cco_carefully();
8 else
9 cco_quickly();
11 }
12
13 void maint(void)
14 {
15 WRITE_ONCE(be_careful, true);
17 do_maint();
19 WRITE_ONCE(be_careful, false);
20 }
v2023.06.11a
Quick Quiz 9.52: What is the point of the second call to synchronize_rcu() in function
maint() in Listing 9.17? Isn’t it OK for any cco() invocations in the clean-up phase to invoke
either cco_carefully() or cco_quickly()?
Quick Quiz 9.53: How can you be sure that the code shown in maint() in Listing 9.17 really
works?
Phased state change allows frequent operations to use light-weight checks, without
the need for expensive lock acquisitions or atomic read-modify-write operations, and is
used in the Linux kernel in the guise of rcu_sync [NZ13] to implement a variant of
reader-writer semaphores with lightweight readers. Phased state change adds only a
checked state variable to the wait-to-finish use case (Section 9.5.4.2), thus also residing
at a rather low level of abstraction.
9.5.4.4 Add-Only List

Add-only data structures, exemplified by the add-only list, can be used for a surprisingly
common set of use cases, perhaps most commonly the logging of changes. Add-only
data structures are a pure use of RCU’s underlying publish/subscribe mechanism.
An add-only variant of a pre-BSD routing table can be derived from Listings 9.14
and 9.15. Because there is no deletion, the route_del() and route_cb() functions
may be dispensed with, along with the ->rh and ->re_freed fields of the route_
entry structure, the rcu_read_lock(), the rcu_read_unlock() invocations in the
route_lookup() function, and all uses of the ->re_freed field in all remaining
functions.
Of course, if there are many concurrent invocations of the route_add() function,
there will be heavy contention on routelock, and if lockless techniques are used, heavy
memory contention on routelist. The usual way to avoid this contention is to use a
concurrency-friendly data structure such as a hash table (see Chapter 10). Alternatively,
per-CPU data structures might be periodically merged into a single global data structure.
On the other hand, if there is never any deletion, extended time periods featuring
many concurrent invocations of route_add() will eventually consume all available
memory. Therefore, most RCU-protected data structures also implement deletion.
9.5.4.5 Type-Safe Memory

A number of lockless algorithms do not require that a given data element keep the
same identity through a given RCU read-side critical section referencing it—but only if
that data element retains the same type. In other words, these lockless algorithms can
tolerate a given data element being freed and reallocated as the same type of structure
while they are referencing it, but must prohibit a change in type. This guarantee,
called “type-safe memory” in academic literature [GC96], is weaker than the existence
guarantees discussed in Section 9.5.4.6, and is therefore quite a bit harder to work
with. Type-safe memory algorithms in the Linux kernel make use of slab caches,
specially marking these caches with SLAB_TYPESAFE_BY_RCU so that RCU is used
when returning a freed-up slab to system memory. This use of RCU guarantees that any
in-use element of such a slab will remain in that slab, thus retaining its type, for the
duration of any pre-existing RCU read-side critical sections.
Quick Quiz 9.54: But what if there is an arbitrarily long series of RCU read-side critical
sections in multiple threads, so that at any point in time there is at least one thread in the
v2023.06.11a
system executing in an RCU read-side critical section? Wouldn’t that prevent any data from a
SLAB_TYPESAFE_BY_RCU slab ever being returned to the system, possibly resulting in OOM
events?
It is important to note that SLAB_TYPESAFE_BY_RCU will in no way prevent kmem_

cache_alloc() from immediately reallocating memory that was just now freed via
kmem_cache_free()! In fact, the SLAB_TYPESAFE_BY_RCU-protected data structure
just returned by rcu_dereference() might be freed and reallocated an arbitrarily
large number of times, even when under the protection of rcu_read_lock(). Instead,
SLAB_TYPESAFE_BY_RCU operates by preventing kmem_cache_free() from returning
a completely freed-up slab of data structures to the system until after an RCU grace
period elapses. In short, although a given RCU read-side critical section might see a
given SLAB_TYPESAFE_BY_RCU data element being freed and reallocated arbitrarily
often, the element’s type is guaranteed not to change until that critical section has
completed.
These algorithms therefore typically use a validation step that checks to make sure
that the newly referenced data structure really is the one that was requested [LS86,
Section 2.5]. These validation checks require that portions of the data structure remain
untouched by the free-reallocate process. Such validation checks are usually very hard
to get right, and can hide subtle and difficult bugs.
Therefore, although type-safety-based lockless algorithms can be extremely helpful in
a very few difficult situations, you should instead use existence guarantees where possible.
Simpler is after all almost always better! On the other hand, type-safety-based lockless
algorithms can provide improved cache locality, and thus improved performance. This
improved cache locality is provided by the fact that such algorithms can immediately
reallocate a newly freed block of memory. In contrast, algorithms based on existence
guarantees must wait for all pre-existing readers before reallocating memory, by which
time that memory may have been ejected from CPU caches.
As can be seen in Figure 9.23, RCU’s type-safe-memory use case combines both the
wait-to-finish and publish-subscribe components, but in the Linux kernel also includes
the slab allocator’s deferred reclamation specified by the SLAB_TYPESAFE_BY_RCU
flag.
9.5.4.6 Existence Guarantee

Gamsa et al. [GKAS99] discuss existence guarantees and describe how a mechanism
resembling RCU can be used to provide these existence guarantees (see Section 5 on
page 7 of the PDF), and Section 7.4 discusses how to guarantee existence via locking,
along with the ensuing disadvantages of doing so. The effect is that if any RCU-protected
data element is accessed within an RCU read-side critical section, that data element is
guaranteed to remain in existence for the duration of that RCU read-side critical section.
Listing 9.18 demonstrates how RCU-based existence guarantees can enable per-
element locking via a function that deletes an element from a hash table. Line 6
computes a hash function, and line 7 enters an RCU read-side critical section. If line 9
finds that the corresponding bucket of the hash table is empty or that the element present
is not the one we wish to delete, then line 10 exits the RCU read-side critical section
and line 11 indicates failure.
Quick Quiz 9.55: What if the element we need to delete is not the first element of the list on
v2023.06.11a
Listing 9.18: Existence Guarantees Enable Per-Element Locking

2 {
4 int b;
5
7 rcu_read_lock();
8 p = rcu_dereference(hashtable[b]);
9 if (p == NULL || p->key != key) {
11 return 0;
12 }
13 spin_lock(&p->lock);
14 if (hashtable[b] == p && p->key == key) {
16 rcu_assign_pointer(hashtable[b], NULL);
19 kfree(p);
20 return 1;
21 }
24 return 0;
25 }
Otherwise, line 13 acquires the update-side spinlock, and line 14 then checks that the
element is still the one that we want. If so, line 15 leaves the RCU read-side critical
section, line 16 removes it from the table, line 17 releases the lock, line 18 waits for
all pre-existing RCU read-side critical sections to complete, line 19 frees the newly
removed element, and line 20 indicates success. If the element is no longer the one we
want, line 22 releases the lock, line 23 leaves the RCU read-side critical section, and
line 24 indicates failure to delete the specified key.
Quick Quiz 9.56: Why is it OK to exit the RCU read-side critical section on line 15 of
Listing 9.18 before releasing the lock on line 17?
Quick Quiz 9.57: Why not exit the RCU read-side critical section on line 23 of Listing 9.18
before releasing the lock on line 22?
Quick Quiz 9.58: The RCU-based algorithm shown in Listing 9.18 locks very similar to that
in Listing 7.11, so why should the RCU-based approach be any better?
Alert readers will recognize this as only a slight variation on the original wait-to-finish
theme (Section 9.5.4.2), adding publish/subscribe, linked structures, a heap allocator
(typically), and deferred reclamation, as shown in Figure 9.23. They might also note the
deadlock-immunity advantages over the lock-based existence guarantees discussed in
Section 7.4.
9.5.4.7 Light-Weight Garbage Collector

A not-uncommon exclamation made by people first learning about RCU is “RCU is sort
of like a garbage collector!” This exclamation has a large grain of truth, but it can also
be misleading.
Perhaps the best way to think of the relationship between RCU and automatic
garbage collectors (GCs) is that RCU resembles a GC in that the timing of collection is
automatically determined, but that RCU differs from a GC in that: (1) The programmer
v2023.06.11a
must manually indicate when a given data structure is eligible to be collected and (2) The
programmer must manually mark the RCU read-side critical sections where references
might be held.
Despite these differences, the resemblance does go quite deep. In fact, the first
RCU-like mechanism I am aware of used a reference-count-based garbage collector
to handle the grace periods [KL80], and the connection between RCU and garbage
collection has been noted more recently [SWS16].
The light-weight garbage collector use case is very similar to the existence-guarantee
use case, adding only the desired non-blocking algorithm to the mix. This light-weight
garbage collector use case can also be used in conjunction with the existence guarantees
described in the next section.
9.5.4.8 Delete-Only List

The delete-only list is the less-popular counterpart to the add-only list covered in
Section 9.5.4.4, and can be thought of as the existence-guarantee use case, but without
the publish/subscribe component, as shown in Figure 9.23. A delete-only list can be
used when the universe of possible members of the list is known at initialization, and
where members can be removed. For example, elements of the list might represent
hardware elements of the system that are subject to failure, but cannot be repaired or
replaced without a reboot.
An delete-only variant of a pre-BSD routing table can be derived from Listings 9.14
and 9.15. Because there is no addition, the route_add() function may be dispensed
with, or, alternatively, its use might be restricted to initialization time. In theory, the
route_lookup() function can use a non-RCU iterator, though in the Linux kernel this
will result in complaints from debug code. In addition, the incremental cost of an RCU
iterator is usually negligible.
As a result, delete-only situations typically use algorithms and data structures that
are designed for addition as well as deletion.
9.5.4.9 Quasi Reader-Writer Lock

Perhaps the most common use of RCU within the Linux kernel is as a replacement
for reader-writer locking in read-intensive situations. Nevertheless, this use of RCU
was not immediately apparent to me at the outset. In fact, I chose to implement a
lightweight reader-writer lock [HW92]17 before implementing a general-purpose RCU
implementation back in the early 1990s. Each and every one of the uses I envisioned for
the lightweight reader-writer lock was instead implemented using RCU. In fact, it was
more than three years before the lightweight reader-writer lock saw its first use. Boy,
did I feel foolish!
The key similarity between RCU and reader-writer locking is that both have read-side
critical sections that can execute concurrently. In fact, in some cases, it is possible to
mechanically substitute RCU API members for the corresponding reader-writer lock
API members. But first, why bother?
Advantages of RCU include performance, deadlock immunity, and realtime latency.
There are, of course, limitations to RCU, including the fact that readers and updaters run
concurrently, that low-priority RCU readers can block high-priority threads waiting for a
grace period to elapse, and that grace-period latencies can extend for many milliseconds.
These advantages and limitations are discussed in the following paragraphs.
17 Similar to brlock in the 2.4 Linux kernel and to lglock in more recent Linux kernels.
v2023.06.11a
10000
Nanoseconds per operation

1000 rwlock
100
10
RCU
1
0.1
1 10 100
Figure 9.25: Performance Advantage of RCU Over Reader-Writer Locking
10000
1000
rwlock
100
10 RCU
1
1 10 100
Figure 9.26: Performance Advantage of Preemptible RCU Over Reader-Writer Locking
Performance The read-side performance advantages of Linux-kernel RCU over

reader-writer locking are shown in Figure 9.25, which was generated on a 448-CPU
2.10 GHz Intel x86 system.
Quick Quiz 9.59: WTF? How the heck do you expect me to believe that RCU can have less
than a 300-picosecond overhead when the clock period at 2.10 GHz is almost 500 picoseconds?
Quick Quiz 9.60: Didn’t an earlier edition of this book show RCU read-side overhead way
down in the sub-picosecond range? What happened???
Quick Quiz 9.61: Why is there such large variation for the RCU trace in Figure 9.25?
Note that reader-writer locking is more than an order of magnitude slower than RCU
on a single CPU, and is more than four orders of magnitude slower on 192 CPUs. In
contrast, RCU scales quite well. In both cases, the error bars cover the full range of the
measurements from 30 runs, with the line being the median.
v2023.06.11a
100000

10000 rwlock 100 CPUs
10 CPUs
1000
1 CPU RCU
100
100 1000 10000
Critical-Section Duration (nanoseconds)
Figure 9.27: Comparison of RCU to Reader-Writer Locking as Function of Critical-

Section Duration, 192 CPUs
A more moderate view may be obtained from a CONFIG_PREEMPT kernel, though

RCU still beats reader-writer locking by between a factor of seven on a single CPU
and by three orders of magnitude on 192 CPUs, as shown in Figure 9.26, which was
generated on the same 448-CPU 2.10 GHz x86 system. Note the high variability of
reader-writer locking at larger numbers of CPUs. The error bars span the full range of
data.
Quick Quiz 9.62: Given that the system had no fewer than 448 hardware threads, why only
192 CPUs?
Of course, the low performance of reader-writer locking in Figures 9.25 and 9.26 is
exaggerated by the unrealistic zero-length critical sections. The performance advantages
of RCU decrease as the overhead of the critical sections increase, as shown in Figure 9.27,
which was run on the same system as the previous plots. Here, the y-axis represents
the sum of the overhead of the read-side primitives and that of the critical section and
the x-axis represents the critical-section overhead in nanoseconds. But please note
the logscale y axis, which means that the small separations between the traces still
represent significant differences. This figure shows non-preemptible RCU, but given
that preemptible RCU’s read-side overhead is only about three nanoseconds, its plot
would be nearly identical to Figure 9.27.
Quick Quiz 9.63: Why the larger error ranges for the submicrosecond durations in Figure 9.27?
There are three traces for reader-writer locking, with the upper trace being for
100 CPUs, the next for 10 CPUs, and the lowest for 1 CPU. The greater the number of
CPUs and the shorter the critical sections, the greater is RCU’s performance advantage.
These performance advantages are underscored by the fact that 100-CPU systems are
no longer uncommon and that a number of system calls (and thus any RCU read-side
critical sections that they contain) complete within microseconds.
In addition, as is discussed in the next paragraph, RCU read-side primitives are almost
entirely deadlock-immune.
Deadlock Immunity Although RCU offers significant performance advantages for

read-mostly workloads, one of the primary reasons for creating RCU in the first place
v2023.06.11a
was in fact its immunity to read-side deadlocks. This immunity stems from the fact that
RCU read-side primitives do not block, spin, or even do backwards branches, so that
their execution time is deterministic. It is therefore impossible for them to participate in
a deadlock cycle.
Quick Quiz 9.64: Is there an exception to this deadlock immunity, and if so, what sequence of
events could lead to deadlock?
An interesting consequence of RCU’s read-side deadlock immunity is that it is

possible to unconditionally upgrade an RCU reader to an RCU updater. Attempting
to do such an upgrade with reader-writer locking results in deadlock. A sample code
fragment that does an RCU read-to-update upgrade follows:
1 rcu_read_lock();
2 list_for_each_entry_rcu(p, &head, list_field) {
3 do_something_with(p);
4 if (need_update(p)) {
5 spin_lock(my_lock);
6 do_update(p);
7 spin_unlock(&my_lock);
8 }
9 }
Note that do_update() is executed under the protection of the lock and under RCU
read-side protection.
Another interesting consequence of RCU’s deadlock immunity is its immunity to a
large class of priority inversion problems. For example, low-priority RCU readers cannot
prevent a high-priority RCU updater from acquiring the update-side lock. Similarly, a
low-priority RCU updater cannot prevent high-priority RCU readers from entering an
RCU read-side critical section.
Quick Quiz 9.65: Immunity to both deadlock and priority inversion??? Sounds too good to
be true. Why should I believe that this is even possible?
Realtime Latency Because RCU read-side primitives neither spin nor block, they
offer excellent realtime latencies. In addition, as noted earlier, this means that they are
immune to priority inversion involving the RCU read-side primitives and locks.
However, RCU is susceptible to more subtle priority-inversion scenarios, for example,
a high-priority process blocked waiting for an RCU grace period to elapse can be
blocked by low-priority RCU readers in -rt kernels. This can be solved by using RCU
priority boosting [McK07d, GMTW08].
However, use of RCU priority boosting requires that rcu_read_unlock() do
deboosting, which entails acquiring scheduler locks. Some care is therefore required
within the scheduler and RCU to avoid deadlocks, which as of the v5.15 Linux kernel
requires RCU to avoid invoking the scheduler while holding any of RCU’s locks.
This in turn means that rcu_read_unlock() is not always lockless when RCU
priority boosting is enabled. However, rcu_read_unlock() will still be lockless if
its critical section was not priority-boosted. Furthermore, critical sections will not be
priority boosted unless they are preempted, or, in -rt kernels, they acquire non-raw
spinlocks. This means that rcu_read_unlock() will normally be lockless from the
perspective of the highest priority task running on any given CPU.
v2023.06.11a
rwlock reader spin rwlock reader

spin rwlock writer
RCU reader RCU reader RCU reader

RCU updater
Time
Update Received
Figure 9.28: Response Time of RCU vs. Reader-Writer Locking
RCU Readers and Updaters Run Concurrently Because RCU readers never spin
nor block, and because updaters are not subject to any sort of rollback or abort semantics,
RCU readers and updaters really can run concurrently. This means that RCU readers
might access stale data, and might even see inconsistencies, either of which can render
conversion from reader-writer locking to RCU non-trivial.
However, in a surprisingly large number of situations, inconsistencies and stale data
are not problems. The classic example is the networking routing table. Because routing
updates can take considerable time to reach a given system (seconds or even minutes),
the system will have been sending packets the wrong way for quite some time when
the update arrives. It is usually not a problem to continue sending updates the wrong
way for a few additional milliseconds. Furthermore, because RCU updaters can make
changes without waiting for RCU readers to finish, the RCU readers might well see the
change more quickly than would batch-fair reader-writer-locking readers, as shown in
Figure 9.28.
Quick Quiz 9.66: But how many other algorithms really tolerate stale and inconsistent data?
Once the update is received, the rwlock writer cannot proceed until the last reader
completes, and subsequent readers cannot proceed until the writer completes. However,
these subsequent readers are guaranteed to see the new value, as indicated by the green
shading of the rightmost boxes. In contrast, RCU readers and updaters do not block
each other, which permits the RCU readers to see the updated values sooner. Of course,
because their execution overlaps that of the RCU updater, all of the RCU readers might
well see updated values, including the three readers that started before the update.
Nevertheless only the green-shaded rightmost RCU readers are guaranteed to see the
updated values.
Reader-writer locking and RCU simply provide different guarantees. With reader-
writer locking, any reader that begins after the writer begins is guaranteed to see new
values, and any reader that attempts to begin while the writer is spinning might or
might not see new values, depending on the reader/writer preference of the rwlock
implementation in question. In contrast, with RCU, any reader that begins after the
updater completes is guaranteed to see new values, and any reader that completes after
the updater begins might or might not see new values, depending on timing.
v2023.06.11a
The key point here is that, although reader-writer locking does indeed guarantee
consistency within the confines of the computer system, there are situations where
this consistency comes at the price of increased inconsistency with the outside world,
courtesy of the finite speed of light and the non-zero size of atoms. In other words,
reader-writer locking obtains internal consistency at the price of silently stale data with
respect to the outside world.
Note that if a value is computed while read-holding a reader-writer lock, and then
that value is used after that lock is released, then this reader-writer-locking use case
is using stale data. After all, the quantities that this value is based on could change
at any time after that lock is released. This sort of reader-writer-locking use case is
often easy to convert to RCU, as will be shown in Listings 9.19, 9.20, and 9.21 and the
accompanying text.
Low-Priority RCU Readers Can Block High-Priority Reclaimers In Realtime

RCU [GMTW08] or SRCU [McK06], a preempted reader will prevent a grace period
from completing, even if a high-priority task is blocked waiting for that grace period
to complete. Realtime RCU can avoid this problem by substituting call_rcu() for
synchronize_rcu() or by using RCU priority boosting [McK07d, GMTW08]. It
might someday be necessary to augment SRCU and RCU Tasks Trace with priority
boosting, but not before a clear real-world need is demonstrated.
Quick Quiz 9.67: If Tasks RCU Trace might someday be priority boosted, why not also Tasks
RCU and Tasks RCU Rude?
RCU Grace Periods Extend for Many Milliseconds With the exception of userspace
RCU [Des09b, MDJ13f], expedited grace periods, and several of the “toy” RCU
implementations described in Appendix B, RCU grace periods extend milliseconds.
Although there are a number of techniques to render such long delays harmless, including
use of the asynchronous interfaces (call_rcu() and call_rcu_bh()) or of the polling
interfaces (get_state_synchronize_rcu(), start_poll_synchronize_rcu(),
and poll_state_synchronize_rcu()), this situation is a major reason for the rule
of thumb that RCU be used in read-mostly situations.
As noted in Section 9.5.3, within the Linux kernel, shorter grace periods may be
obtained via expedited grace periods, for example, by invoking synchronize_rcu_
expedited() instead of synchronize_rcu(). Expedited grace periods can reduce
delays to as little as a few tens of microseconds, albeit at the expense of higher CPU
utilization and IPIs. The added IPIs can be especially unwelcome in some real-time
workloads.
Code: Reader-Writer Locking vs. RCU In the best case, the conversion from
reader-writer locking to RCU is quite simple, as shown in Listings 9.19, 9.20, and 9.21,
all taken from Wikipedia [MPA+ 06].
However, the transformation is not always this straightforward. This is because neither
the spin_lock() nor the synchronize_rcu() in Listing 9.21 exclude the readers in
Listing 9.20. First, the spin_lock() does not interact in any way with rcu_read_
lock() and rcu_read_unlock(), thus not excluding them. Second, although both
write_lock() and synchronize_rcu() wait for pre-existing readers, only write_
lock() prevents subsequent readers from commencing.18 Thus, synchronize_rcu()
18 Kudos to whoever pointed this out to Paul.
v2023.06.11a
Listing 9.19: Converting Reader-Writer Locking to RCU: Data

1 struct el { 1 struct el {
2 struct list_head lp; 2 struct list_head lp;
3 long key; 3 long key;
4 spinlock_t mutex; 4 spinlock_t mutex;
5 int data; 5 int data;
6 /* Other data fields */ 6 /* Other data fields */
7 }; 7 };
8 DEFINE_RWLOCK(listmutex); 8 DEFINE_SPINLOCK(listmutex);
9 LIST_HEAD(head); 9 LIST_HEAD(head);
Listing 9.20: Converting Reader-Writer Locking to RCU: Search

1 int search(long key, int *result) 1 int search(long key, int *result)
2 { 2 {
3 struct el *p; 3 struct el *p;
4 4
5 read_lock(&listmutex); 5 rcu_read_lock();
6 list_for_each_entry(p, &head, lp) { 6 list_for_each_entry_rcu(p, &head, lp) {
7 if (p->key == key) { 7 if (p->key == key) {
8 *result = p->data; 8 *result = p->data;
9 read_unlock(&listmutex); 9 rcu_read_unlock();
10 return 1; 10 return 1;
11 } 11 }
12 } 12 }
13 read_unlock(&listmutex); 13 rcu_read_unlock();
15 } 15 }
Listing 9.21: Converting Reader-Writer Locking to RCU: Deletion

1 int delete(long key) 1 int delete(long key)
2 { 2 {
3 struct el *p; 3 struct el *p;
4 4
5 write_lock(&listmutex); 5 spin_lock(&listmutex);
6 list_for_each_entry(p, &head, lp) { 6 list_for_each_entry(p, &head, lp) {
7 if (p->key == key) { 7 if (p->key == key) {
8 list_del(&p->lp); 8 list_del_rcu(&p->lp);
9 write_unlock(&listmutex); 9 spin_unlock(&listmutex);
10 kfree(p); 11 kfree(p);
12 } 13 }
13 } 14 }
14 write_unlock(&listmutex); 15 spin_unlock(&listmutex);
16 } 17 }
v2023.06.11a
cannot exclude readers. Nevertheless, a great many situations using reader-writer

locking can be converted to RCU.
More-elaborate cases of replacing reader-writer locking with RCU may be found
elsewhere [Bro15a, Bro15b].
Semantics: Reader-Writer Locking vs. RCU Expanding on the previous section,

reader-writer locking semantics can be roughly and informally summarized by the
following three temporal constraints:
1. Write-side acquisitions wait for any read-holders to release the lock.
2. Writer-side acquisitions wait for any write-holder to release the lock.
3. Read-side acquisitions wait for any write-holder to release the lock.
RCU dispenses entirely with constraint #3 and weakens the other two as follows:
1. Writers wait for any pre-existing read-holders before progressing to the destructive
phase of their update (usually the freeing of memory).
2. Writers synchronize with each other as needed.
It is of course this weakening that permits RCU implementations to attain excellent

performance and scalability. It also allows RCU to implement the aforementioned
unconditional read-to-write upgrade that is so attractive and so deadlock-prone in reader-
writer locking. Code using RCU can compensate for this weakening in a surprisingly
large number of ways, but most commonly by imposing spatial constraints:
1. New data is placed in newly allocated memory.
2. Old data is freed, but only after:
(a) That data has been unlinked so as to be inaccessible to later readers, and
(b) A subsequent RCU grace period has elapsed.
Of course, there are some reader-writer-locking use cases for which RCU’s weakened
semantics are inappropriate, but experience in the Linux kernel indicates that more than
80% of reader-writer locks can in fact be replaced by RCU. For example, a common
reader-writer-locking use case computes some value while holding the lock and then
uses that value after releasing that lock. This use case results in stale data, and therefore
often accommodates RCU’s weaker semantics.
This interaction of temporal and spatial constraints is illustrated by the RCU singleton
data structure illustrated in Figures 9.6 and 9.7. This structure is defined on lines 1–4 of
Listing 9.22, and contains two integer fields, ->a and ->b (singleton.c). The current
instance of this structure is referenced by the curconfig pointer defined on line 4.
The fields of the current structure are passed back through the cur_a and cur_b
parameters to the get_config() function defined on lines 6–20. These two fields can
be slightly out of date, but they absolutely must be consistent with each other. The
get_config() function provides this consistency within the RCU read-side critical
section starting on line 10 and ending on either line 13 or line 18, which provides the
needed temporal synchronization. Line 11 fetches the pointer to the current myconfig
structure. This structure will be used regardless of any concurrent changes due to
v2023.06.11a
Listing 9.22: RCU Singleton Get

1 struct myconfig {
2 int a;
3 int b;
4 } *curconfig;
5
6 int get_config(int *cur_a, int *cur_b)
7 {
8 struct myconfig *mcp;
9
10 rcu_read_lock();
11 mcp = rcu_dereference(curconfig);
12 if (!mcp) {
14 return 0;
15 }
16 *cur_a = mcp->a;
17 *cur_b = mcp->b;
19 return 1;
20 }
Listing 9.23: RCU Singleton Set

1 void set_config(int cur_a, int cur_b)
2 {
3 struct myconfig *mcp;
4
5 mcp = malloc(sizeof(*mcp));
6 BUG_ON(!mcp);
7 mcp->a = cur_a;
8 mcp->b = cur_b;
9 mcp = xchg(&curconfig, mcp);
10 if (mcp) {
12 free(mcp);
13 }
14 }
calls to the set_config() function, thus providing the needed spatial synchronization.
If line 12 determines that the curconfig pointer was NULL, line 14 returns failure.
Otherwise, lines 16 and 17 copy out the ->a and ->b fields and line 19 returns success.
These ->a and ->b fields are from the same myconfig structure, and the RCU read-side
critical section prevents this structure from being freed, thus guaranteeing that these
two fields are consistent with each other.
The structure is updated by the set_config() function shown in Listing 9.23.
Lines 5–8 allocate and initialize a new myconfig structure. Line 9 atomically exchanges
a pointer to this new structure with the pointer to the old structure in curconfig, while
also providing full memory ordering both before and after the xchg() operation, thus
providing the needed updater/reader spatial synchronization on the one hand and the
needed updater/updater synchronization on the other. If line 10 determines that the
pointer to the old structure was in fact non-NULL, line 11 waits for a grace period (thus
providing the needed reader/updater temporal synchronization) and line 12 frees the old
structure, safe in the knowledge that there are no longer any readers still referencing it.
Figure 9.29 shows an abbreviated representation of get_config() on the left and
right and a similarly abbreviated representation of set_config() in the middle. Time
advances from top to bottom, and the address space of the objects referenced by
curconfig advances from left to right. The boxes with comma-separated numbers
each represent a myconfig structure, with the constraint that ->b is the square of ->a.
Each blue dash-dotted arrow represents an interaction with the old structure (on the left,
v2023.06.11a
Time
Address Space
5,
5,25 curconfig
5, 5,
9,81
Readers
rcu_read_lock();
mcp = ...
*cur_a = mcp->a; (5) mcp = kmalloc(...)
mcp = xchg(&curconfig, mcp);
*cur_b = mcp->b; (25) synchronize_rcu();
Grace
Period
rcu_read_unlock(); ... rcu_read_lock();

... mcp = ...
kfree(mcp); *cur_a = mcp->a; (9)

Readers
*cur_b = mcp->b; (81)

rcu_read_unlock();
Figure 9.29: RCU Spatial/Temporal Synchronization
containing “5,25”) and each green dashed arrow represents an interaction with the new
structure (on the right, containing “9,81”).
The black dotted arrows represent temporal relationships between RCU readers
on the left and right and the RCU grace period at center, with each arrow pointing
from an older event to a newer event. The call to synchronize_rcu() followed the
leftmost rcu_read_lock(), and therefore that synchronize_rcu() invocation must
not return until after the corresponding rcu_read_unlock(). In contrast, the call to
synchronize_rcu() precedes the rightmost rcu_read_lock(), which allows the
return from that same synchronize_rcu() to ignore the corresponding rcu_read_
unlock(). These temporal relationships prevent the myconfig structures from being
freed while RCU readers are still accessing them.
The two horizontal grey dashed lines represent the period of time during which
different readers get different results, however, each reader will see one and only one of
the two objects. All readers that end before the first horizontal line will see the leftmost
myconfig structure, and all readers that start after the second horizontal line will see
the rightmost structure. Between the two lines, that is, during the grace period, different
readers might see different objects, but as long as each reader loads the curconfig
pointer only once, each reader will see a consistent view of its myconfig structure.
Quick Quiz 9.68: But doesn’t the RCU grace period start sometime after the call to
synchronize_rcu() rather than in the middle of that xchg() statement?
In short, when operating on a suitable linked data structure, RCU combines temporal
and spatial synchronization in order to approximate reader-writer locking, with RCU
read-side critical sections acting as the reader-writer-locking reader, as shown in
Figures 9.23 and 9.29. RCU’s temporal synchronization is provided by the read-side
markers, for example, rcu_read_lock() and rcu_read_unlock(), as well as the
update-side grace-period primitives, for example, synchronize_rcu() or call_
rcu(). The spatial synchronization is provided by the read-side rcu_dereference()
family of primitives, each of which subscribes to a version published by rcu_assign_
v2023.06.11a
pointer().19 RCU’s combining of temporal and spatial synchronization contrasts to

the schemes presented in Sections 6.3.2, 6.3.3, and 7.1.4, in which temporal and spatial
synchronization are provided separately by locking and by static data-structure layout,
respectively.
Quick Quiz 9.69: Is RCU the only synchronization mechanism that combines temporal and
spatial synchronization in this way?
9.5.4.10 Quasi Reference Count

Because grace periods are not allowed to complete while there is an RCU read-side
critical section in progress, the RCU read-side primitives may be used as a restricted
reference-counting mechanism. For example, consider the following code fragment:
1 rcu_read_lock(); /* acquire reference. */

2 p = rcu_dereference(head);
3 /* do something with p. */
4 rcu_read_unlock(); /* release reference. */
The combination of the rcu_read_lock() and rcu_dereference() primitives

can be thought of as acquiring a reference to p, because a grace period starting after
the rcu_dereference() assignment to p cannot possibly end until after we reach the
matching rcu_read_unlock(). This reference-counting scheme is restricted in that it
is forbidden to wait for RCU grace periods within RCU read-side critical sections, and
also forbidden to hand off an RCU read-side critical section’s references from one task
to another.
Regardless of these restrictions, the following code can safely delete p:
2 p = head;
3 rcu_assign_pointer(head, NULL);
5 /* Wait for all references to be released. */
7 kfree(p);
The assignment to head prevents any future references to p from being acquired, and
the synchronize_rcu() waits for any previously acquired references to be released.
Quick Quiz 9.70: But wait! This is exactly the same code that might be used when thinking
of RCU as a replacement for reader-writer locking! What gives?
Of course, RCU can also be combined with traditional reference counting, as discussed
in Section 13.2.
But why bother? Again, part of the answer is performance, as shown in Figures 9.30
and 9.31, again showing data taken on a 448-CPU 2.1 GHz Intel x86 system for non-
preemptible and preemptible Linux-kernel RCU, respectively. Non-preemptible RCU’s
advantage over reference counting ranges from more than an order of magnitude at one
CPU up to about four orders of magnitude at 192 CPUs. Preemptible RCU’s advantage
ranges from about a factor of three at one CPU up to about three orders of magnitude at
192 CPUs.
19 Preferably with both rcu_dereference() and rcu_assign_pointer() being
embedded in higher-level APIs.
v2023.06.11a
10000

1000
refcnt
100
10
RCU
1
0.1
1 10 100
Figure 9.30: Performance of RCU vs. Reference Counting
10000
1000
refcnt
100
10 RCU
1
1 10 100
Figure 9.31: Performance of Preemptible RCU vs. Reference Counting
100000
10000 refcnt 100 CPUs
10 CPUs
1000 RCU
1 CPU
100
100 1000 10000
Critical-Section Duration (nanoseconds)
Figure 9.32: Response Time of RCU vs. Reference Counting, 192 CPUs
v2023.06.11a
However, as with reader-writer locking, the performance advantages of RCU are most
pronounced for short-duration critical sections and for large numbers of CPUs, as shown
in Figure 9.32 for the same system. In addition, as with reader-writer locking, many
system calls (and thus any RCU read-side critical sections that they contain) complete
in a few microseconds.
Although traditional reference counters are usually associated with a specific data
structure, or perhaps a specific group of data structures, this approach does have some
disadvantages. For example, maintaining a single global reference counter for a large
variety of data structures typically results in bouncing the cache line containing the
reference count. As we saw in Figures 9.30–9.32, such cache-line bouncing can severely
degrade performance.
In contrast, RCU’s lightweight rcu_read_lock(), rcu_dereference(), and rcu_
read_unlock() read-side primitives permit extremely frequent read-side usage with
negligible performance degradation. Except that the calls to rcu_dereference() are
not doing anything specific to acquire a reference to the pointed-to object. The heavy
lifting is instead done by the rcu_read_lock() and rcu_read_unlock() primitives
and their interactions with RCU grace periods.
And ignoring those calls to rcu_dereference() permits RCU to be thought of as a
“bulk reference-counting” mechanism, where each call to rcu_read_lock() obtains
a reference on each and every RCU-protected object, and with little or no overhead.
However, the restrictions that go with RCU can be quite onerous. For example, in many
cases, the Linux-kernel prohibition against sleeping while in an RCU read-side critical
section would defeat the entire purpose. Such cases might be better served by the hazard
pointers mechanism described in Section 9.3. Cases where code rarely sleeps have been
handled by using RCU as a reference count in the common non-sleeping case and by
bridging to an explicit reference counter when sleeping is necessary.
Alternatively, situations where a reference must be held by a single task across a
section of code that sleeps may be accommodated with Sleepable RCU (SRCU) [McK06].
This fails to cover the not-uncommon situation where a reference is “passed” from one
task to another, for example, when a reference is acquired when starting an I/O and
released in the corresponding completion interrupt handler. Again, such cases might be
better handled by explicit reference counters or by hazard pointers.
Of course, SRCU brings restrictions of its own, namely that the return value from
srcu_read_lock() be passed into the corresponding srcu_read_unlock(), and that
no SRCU primitives be invoked from hardware interrupt handlers or from non-maskable
interrupt (NMI) handlers. The jury is still out as to how much of a problem is presented
by this restriction, and as to how it can best be handled.
However, in the common case where references are held within the confines of
a single CPU or task, RCU can be used as high-performance and highly scalable
reference-counting mechanism.
As shown in Figure 9.23, quasi reference counts add RCU readers as individual or
bulk reference counts, possibly also bridging to reference counters in corner cases.
9.5.4.11 Quasi Multi-Version Concurrency Control
RCU can also be thought of as a simplified multi-version concurrency control (MVCC)

mechanism with weak consistency criteria. The multi-version aspects were touched
upon in Section 9.5.2.3. However, in its native form, RCU provides version consistency
only within a given RCU-protected data element.
v2023.06.11a
Nevertheless, there are situations where consistency and fresh data are required
across multiple data elements. Fortunately, there are a number of approaches that avoid
inconsistency and stale data, including the following:
1. Enclose RCU readers within sequence-locking readers, forcing the RCU readers to
be retried should an update occur, as described in Section 13.4.2 and Section 13.4.3.
2. Place the data that must be consistent into a single element of a linked data structure,
and refrain from updating those fields within any element visible to RCU readers.
RCU readers gaining a reference to any such element are then guaranteed to see
consistent values. See Section 13.5.4 for additional details.
3. Use a per-element lock that guards a “deleted” flag to allow RCU readers to reject
stale data [McK04, ACMS03].
4. Provide an existence flag that is referenced by all data elements whose update is to
appear atomic to RCU readers [McK14d, McK14a, McK15b, McK16b, McK16a].
5. Use one of a wide range of counter-based methods [McK08a, McK10, MW11,

McK14b, MSFM15, KMK+ 19]. In these approaches, updaters maintain a version
number and maintain links to old versions of a given piece of data. Readers take a
snapshot of the current version number, and, if necessary, traverse the links to find
a version consistent with that snapshot.
In short, when using RCU to approximate multi-version concurrency control, you

only pay for the level of consistency that you actually need.
As shown in Figure 9.23, quasi multi-version concurrency control is based on existence
guarantees, adding read-side snapshot operations and constraints on readers and writers,
the exact form of the constraint being dictated by the consistency requirements, as
summarized above.
9.5.4.12 RCU Usage Summary

At its core, RCU is nothing more nor less than an API that provides:
1. A publish-subscribe mechanism for adding new data,
2. A way of waiting for pre-existing RCU readers to finish, and
3. A discipline of maintaining multiple versions to permit change without harming or

unduly delaying concurrent RCU readers.
That said, it is possible to build higher-level constructs on top of RCU, including the
various use cases described in the earlier sections. Furthermore, I have no doubt that
new use cases will continue to be found for RCU, as well as for any of a number of other
synchronization primitives. And so it is that RCU’s use cases are conceptually more
complex than is RCU itself, as hinted on page 201.
Quick Quiz 9.71: Which of these use cases best describes the Pre-BSD routing example in
Section 9.5.4.1?
In the meantime, Figure 9.33 shows some rough rules of thumb on where RCU is
most helpful.
v2023.06.11a
Pre-BSD Routing Table
Stale and Inconsistent Data OK
& CU
Re con ork
In
ad sist s G
(R
-M en
Ne U M
Ne
os t D reat
100% Reads
100% Writes
(R
W
Re ns
ed
Re ons Be
ed U W
tly at !!!)
(R
C
ad iste Wel
,S a
ad iste OK
C ght
Co ork
Ne
-M
-W nt
tal OK
ed
W ons ot B
os t D
i
e
rite Da ..)
rite ist
C
tly ata
(R
,
-M ent st)*
n
,
CU
s
os Da
tly
N
ta
l)
, ta
.
e
Need Fully Fresh and Consistent Data
* 1. RCU provides ABA protection for update-friendly synchronization mechanisms
* 2. RCU provides bounded wait-free read-side primitives for real-time use
Figure 9.33: RCU Areas of Applicability
As shown in the blue box in the upper-right corner of the figure, RCU works best
if you have read-mostly data where stale and inconsistent data is permissible (but see
below for more information on stale and inconsistent data). The canonical example
of this case in the Linux kernel is routing tables. Because it may have taken many
seconds or even minutes for the routing updates to propagate across the Internet, the
system has been sending packets the wrong way for quite some time. Having some
small probability of continuing to send some of them the wrong way for a few more
milliseconds is almost never a problem.
If you have a read-mostly workload where consistent data is required, RCU works
well, as shown by the green “read-mostly, need consistent data” box. One example
of this case is the Linux kernel’s mapping from user-level System-V semaphore IDs
to the corresponding in-kernel data structures. Semaphores tend to be used far more
frequently than they are created and destroyed, so this mapping is read-mostly. However,
it would be erroneous to perform a semaphore operation on a semaphore that has
already been deleted. This need for consistency is handled by using the lock in the
in-kernel semaphore data structure, along with a “deleted” flag that is set when deleting
a semaphore. If a user ID maps to an in-kernel data structure with the “deleted” flag set,
the data structure is ignored, so that the user ID is flagged as invalid.
Although this requires that the readers acquire a lock for the data structure representing
the semaphore itself, it allows them to dispense with locking for the mapping data
structure. The readers therefore locklessly traverse the tree used to map from ID to
data structure, which in turn greatly improves performance, scalability, and real-time
response.
As indicated by the yellow “read-write” box, RCU can also be useful for read-write
workloads where consistent data is required, although usually in conjunction with a
number of other synchronization primitives. For example, the directory-entry cache in
recent Linux kernels uses RCU in conjunction with sequence locks, per-CPU locks, and
per-data-structure locks to allow lockless traversal of pathnames in the common case.
Although RCU can be very beneficial in this read-write case, the corresponding code is
often more complex than that of the read-mostly cases.
Finally, as indicated by the red box in the lower-left corner of the figure, update-mostly
workloads requiring consistent data are rarely good places to use RCU, though there
are some exceptions [DMS+ 12]. For example, as noted in Section 9.5.4.5, within
v2023.06.11a
the Linux kernel, the SLAB_TYPESAFE_BY_RCU slab-allocator flag provides type-safe

memory to RCU readers, which can greatly simplify non-blocking synchronization and
other lockless algorithms. In addition, if the rare readers are on critical code paths
on real-time systems, use of RCU for those readers might provide real-time response
benefits that more than make up for the increased update-side overhead, as discussed in
Section 14.3.6.5.
In short, RCU is an API that includes a publish-subscribe mechanism for adding
new data, a way of waiting for pre-existing RCU readers to finish, and a discipline of
maintaining multiple versions to allow updates to avoid harming or unduly delaying
concurrent RCU readers. This RCU API is best suited for read-mostly situations,
especially if stale and inconsistent data can be tolerated by the application.
9.5.5 RCU Related Work

The first known mention of anything resembling RCU took the form of a bug report
from Donald Knuth [Knu73, page 413 of Fundamental Algorithms] against Joseph
Weizenbaum’s SLIP list-processing facility for FORTRAN [Wei63]. Knuth was justified
in reporting the bug, as SLIP had no notion of any sort of grace-period guarantee.
The first known non-bug-report mention of anything resembling RCU appeared in
Kung’s and Lehman’s landmark paper [KL80]. There was some additional use of
this technique in academia [ML82, ML84, Lis88, Pug90, And91, PAB+ 95, CAK+ 96,
RSB+ 97, GKAS99], but much of the work in this area was instead carried out by
practitioners [RTY+ 87, HOS89, Jac93, Joh95, SM95, SM97, SM98, MS98a].
Quick Quiz 9.72: Garbage collectors? Passive serialization? System reference points?
Quiescent states? Aging? Generations? Why on earth couldn’t the knuckleheads working on
these early papers bring themselves to agree on a common terminology???
By the year 2000, the initiative had passed to open-source projects, most notably the
Linux kernel community [Rus00a, Rus00b, MS01, MAK+ 01, MSA+ 02, ACMS03].20
RCU was accepted into the Linux kernel in late 2002, with many subsequent improve-
ments for scalability, robustness, real-time response, energy efficiency, and specialized
use cases. As of 2023, Linux-kernel RCU is still under active development.
Quick Quiz 9.73: Why didn’t Kung’s and Lehman’s paper result in immediate use of RCU?
However, in the mid 2010s, there was a welcome upsurge in RCU research and
development across a number of communities and institutions [Kaa15]. Section 9.5.5.1
describes uses of RCU, Section 9.5.5.2 describes RCU implementations (as well as work
that both creates and uses an implementation), and finally, Section 9.5.5.3 describes
verification and validation of RCU and its uses.
9.5.5.1 RCU Uses

Phil Howard and Jon Walpole of Portland State University (PSU) have applied RCU to
red-black trees [How12, HW11] combined with updates synchronized using software
transactional memory. Josh Triplett and Jon Walpole (again of PSU) applied RCU
to resizable hash tables [Tri12, TMW11, Cor14c, Cor14d]. Other RCU-protected
resizable hash tables have been created by Herbert Xu [Xu10] and by Mathieu Desnoy-
ers [MDJ13c].
20 A list of citations with well over 200 entries may be found in bib/RCU.bib in the
LATEX source for this book.
v2023.06.11a
Austin Clements, Frans Kaashoek, and Nickolai Zeldovich of MIT created an RCU-
optimized balanced binary tree (Bonsai) [CKZ12], and applied this tree to the Linux
kernel’s VM subsystem in order to reduce read-side contention on the Linux kernel’s
mmap_sem. This work resulted in order-of-magnitude speedups and scalability up to
at least 80 CPUs for a microbenchmark featuring large numbers of minor page faults.
This is similar to a patch developed earlier by Peter Zĳlstra [Zĳ14], and both were
limited by the fact that, at the time, filesystem data structures were not safe for RCU
readers. Clements et al. avoided this limitation by optimizing the page-fault path for
anonymous pages only. More recently, filesystem data structures have been made safe
for RCU readers [Cor10a, Cor11], so perhaps this work can be implemented for all page
types, not just anonymous pages—Peter Zĳlstra has, in fact, recently prototyped exactly
this, and Laurent Dufour Michel Lespinasse have continued work along these lines.
For their part, Matthew Wilcox and Liam Howlett are working towards use of RCU to
enable fine-grained locking of and lockless access to other memory-management data
structures.
Yandong Mao and Robert Morris of MIT and Eddie Kohler of Harvard University
created another RCU-protected tree named Masstree [MKM12] that combines ideas
from B+ trees and tries. Although this tree is about 2.5x slower than an RCU-protected
hash table, it supports operations on key ranges, unlike hash tables. In addition, Masstree
supports efficient storage of objects with long shared key prefixes and, furthermore,
provides persistence via logging to mass storage.
The paper notes that Masstree’s performance rivals that of memcached, even given that
Masstree is persistently storing updates and memcached is not. The paper also compares
Masstree’s performance to the persistent datastores MongoDB, VoltDB, and Redis,
reporting significant performance advantages for Masstree, in some cases exceeding two
orders of magnitude. Another paper [TZK+ 13], by Stephen Tu, Wenting Zheng, Barbara
Liskov, and Samuel Madden of MIT and Kohler, applies Masstree to an in-memory
database named Silo, achieving 700K transactions per second (42M transactions per
minute) on a well-known transaction-processing benchmark. Interestingly enough, Silo
guarantees linearizability without incurring the overhead of grace periods while holding
locks.
Maya Arbel and Hagit Attiya of Technion took a more rigorous approach [AA14]
to an RCU-protected search tree that, like Masstree, allows concurrent updates. This
paper includes a proof of correctness, including proof that all operations on this
tree are linearizable. Unfortunately, this implementation achieves linearizability by
incurring the full latency of grace-period waits while holding locks, which degrades
scalability of update-only workloads. One way around this problem is to abandon
linearizability [HKLP12, McK14d], however, Arbel and Attiya instead created an RCU
variant that reduces low-end grace-period latency. Of course, nothing comes for free,
and this RCU variant appears to hit a scalability limit at about 32 CPUs. Although
there is much to be said for dropping linearizability, thus gaining both performance
and scalability, it is very good to see academics experimenting with alternative RCU
implementations.
9.5.5.2 RCU Implementations

Timothy Harris created a time-based user-space RCU [Har01] that improves on those
created previously by Jacobson [Jac93] and John [Joh95]. These prior two time-based
approaches each assume a sharp upper bound on reader duration, which can work
correctly in hard real-time systems. In non-real-time systems, this type of approach
v2023.06.11a
is subject to failure when readers are interrupted, preempted, or otherwise delayed.

However, the fact that such a failure-prone implementation would be independently
invented twice shows the depth of the need for RCU-like mechanisms. Timothy Harris
improves upon these two earlier efforts by requiring each reader to take a snapshot of a
global timebase before starting its read-side traversal. Freeing a reader-visible object is
then deferred until all processes’ reader snapshots indicate a time following that of the
removal of that object. However, global timebases can be expensive and inaccurate on
some systems.
Keir Fraser created a user-space RCU named EBR for use in non-blocking syn-
chronization and software transactional memory [Fra03, Fra04, FH07]. This work
improves on that of Timothy Harris by replacing the global clock with a software
counter, thus eliminating much of the expense and all of the inaccuracy associated with
commodity-system global clocks of that time. Interestingly enough, this work cites
Linux-kernel RCU on the one hand, but also inspired the name QSBR for the original
non-preemptible Linux-kernel RCU implementation.
Mathieu Desnoyers created a user-space RCU for use in tracing [Des09b, Des09a,
DMS+ 12, MDJ13f, MDJ13c, MDJ13b, MDJ13d, MDJ13e, MDJ13h, MDJT13b,
MDJ13g, MDJ13a, MDJT13a], which has seen use in a number of projects [BD13].
Researchers at Charles University in Prague have also been working on RCU
implementations, including dissertations by Andrej Podzimek [Pod10] and Adam
Hraska [Hra13].
Yujie Liu (Lehigh University), Victor Luchangco (Oracle Labs), and Michael Spear
(also Lehigh) [LLS13] pressed scalable non-zero indicators (SNZI) [ELLM07] into
service as a grace-period mechanism. The intended use is to implement software
transactional memory (see Section 17.2), which imposes linearizability requirements,
which in turn seems to limit scalability.
RCU-like mechanisms are also finding their way into Java. Sivaramakrishnan
et al. [SZJ12] use an RCU-like mechanism to eliminate the read barriers that are
otherwise required when interacting with Java’s garbage collector, resulting in significant
performance improvements.
Ran Liu, Heng Zhang, and Haibo Chen of Shanghai Jiao Tong University cre-
ated a specialized variant of RCU that they used for an optimized “passive reader-
writer lock” [LZC14], similar to those created by Gautham Shenoy [She06] and
Srivatsa Bhat [Bha14]. The Liu et al. paper is interesting from a number of perspec-
tives [McK14g].
Mike Ash posted [Ash15] a description of an RCU-like primitive in Apple’s Objective-
C runtime. This approach identifies read-side critical sections via designated code
ranges, thus qualifying as another method of achieving zero read-side overhead, albeit
one that poses some interesting practical challenges for large read-side critical sections
that span multiple functions.
Pedro Ramalhete and Andreia Correia [RC15] produced “Poor Man’s RCU”, which,
despite using a pair of reader-writer locks, manages to provide lock-free forward-progress
guarantees to readers [MP15a].
Maya Arbel and Adam Morrison [AM15] produced “Predicate RCU”, which works
hard to reduce grace-period duration in order to efficiently support algorithms that hold
update-side locks across grace periods. This results in reduced batching of updates into
grace periods and reduced scalability, but does succeed in providing short grace periods.
Quick Quiz 9.74: Why not just drop the lock before waiting for the grace period, or using
something like call_rcu() instead of waiting for a grace period?
v2023.06.11a
Alexander Matveev (MIT), Nir Shavit (MIT and Tel-Aviv University), Pascal Felber
(University of Neuchâtel), and Patrick Marlier (also University of Neuchâtel) [MSFM15]
produced an RCU-like mechanism that can be thought of as software transactional
memory that explicitly marks read-only transactions. Their use cases require holding
locks across grace periods, which limits scalability [MP15a, MP15b]. This appears to
be the first academic RCU-related work to make good use of the rcutorture test suite,
and also the first to have submitted a performance improvement to Linux-kernel RCU,
which was accepted into v4.4.
Alexander Matveev’s RLU was followed up by MV-RLU from Jaeho Kim et
al. [KMK+ 19]. This work improves scalability over RLU by permitting multiple
concurrent updates, by avoiding holding locks across grace periods, and by using asyn-
chronous grace periods, for example, call_rcu() instead of synchronize_rcu().
This paper also made some interesting performance-evaluation choices that are discussed
further in Section 17.2.3.3 on page 594.
Adam Belay et al. created an RCU implementation that guards the data structures used
by TCP/IP’s address-resolution protocol (ARP) in their IX operating system [BPP+ 16].
Geoff Romer and Andrew Hunter (both at Google) proposed a cell-based API for
RCU protection of singleton data structures for inclusion in the C++ standard [RH18].
Dimitrios Siakavaras et al. have applied HTM and RCU to search trees [SNGK17,
SBN+ 20], Christina Giannoula et al. have used HTM and RCU to color graphs [GGK18],
and SeongJae Park et al. have used HTM and RCU to optimize high-contention locking
on NUMA systems.
Alex Kogan et al. applied RCU to the construction of range locking for scalable
address spaces [KDI20].
Production uses of RCU are listed in Section 9.6.3.3.
9.5.5.3 RCU Validation

In early 2017, it is commonly recognized that almost any bug is a potential security
exploit, so validation and verification are first-class concerns.
Researchers at Stony Brook University have produced an RCU-aware data-race
detector [Dug10, Sey12, SRK+ 11]. Alexey Gotsman of IMDEA, Noam Rinetzky of
Tel Aviv University, and Hongseok Yang of the University of Oxford have published a
paper [GRY12] expressing the formal semantics of RCU in terms of separation logic,
and have continued with other aspects of concurrency.
Joseph Tassarotti (Carnegie-Mellon University), Derek Dreyer (Max Planck Institute
for Software Systems), and Viktor Vafeiadis (also MPI-SWS) [TDV15] produced a
manual formal proof of correctness of the quiescent-state-based reclamation (QSBR)
variant of userspace RCU [Des09b, DMS+ 12]. Lihao Liang (University of Oxford), Paul
E. McKenney (IBM), Daniel Kroening, and Tom Melham (both also Oxford) [LMKM16]
used the C bounded model checker (CBMC) [CKL04] to produce a mechanical proof of
correctness of a significant portion of Linux-kernel Tree RCU. Lance Roy [Roy17] used
CBMC to produce a similar proof of correctness for a significant portion of Linux-kernel
sleepable RCU (SRCU) [McK06]. Finally, Michalis Kokologiannakis and Konstantinos
Sagonas (National Technical University of Athens) [KS17a, KS19] used the Nighugg
tool [LSLK14] to produce a mechanical proof of correctness of a somewhat larger
portion of Linux-kernel Tree RCU.
None of these efforts located any bugs other than bugs injected into RCU specifically
to test the verification tools. In contrast, Alex Groce (Oregon State University), Iftekhar
Ahmed, Carlos Jensen (both also OSU), and Paul E. McKenney (IBM) [GAJM15]
v2023.06.11a
9.6. WHICH TO CHOOSE? 281
Table 9.7: Which Deferred Technique to Choose? (Overview)
Property Reference Counting Hazard Pointers Sequence Locks RCU
Readers Slow and unscalable Fast and scalable Fast and scalable Fast and scalable
Memory Overhead Counter per object Pointer per No protection None
reader per object
Duration of Protection Can be long Can be long No protection User must bound
duration
Need for Traversal If object deleted If object deleted If any update Never
Retries
automatically mutated Linux-kernel RCU’s source code to test the coverage of the
rcutorture test suite. The effort found several holes in this suite’s coverage, one of
which was hiding a real bug (since fixed) in Tiny RCU.
With some luck, all of this validation work will eventually result in more and better
tools for validating concurrent code.
9.6 Which to Choose?
Choose always the way that seems the best, however

rough it may be; custom will soon render it easy and
agreeable.
Pythagoras
Section 9.6.1 provides a high-level overview and then Section 9.6.2 provides a more
detailed view of the differences between the deferred-processing techniques presented
in this chapter. This discussion assumes a linked data structure that is large enough
that readers do not hold references from one traversal to another, and where elements
might be added to and removed from the structure at any location and at any time.
Section 9.6.3 then points out a few publicly visible production uses of hazard pointers,
sequence locking, and RCU. This discussion should help you to make an informed
choice between these techniques.
9.6.1 Which to Choose? (Overview)

Table 9.7 shows a few high-level properties that distinguish the deferred-reclamation
techniques from one another.
The “Readers” row summarizes the results presented in Figure 9.22, which shows
that all but reference counting enjoy reasonably fast and scalable readers.
The “Memory Overhead” row evaluates each technique’s need for external storage
with which to record reader protection. RCU relies on quiescent states, and thus needs
no storage to represent readers, whether within or outside of the object. Reference
counting can use a single integer within each object in the structure, and no additional
storage is required. Hazard pointers require external-to-object pointers be provisioned,
and that there be sufficient pointers for each CPU or thread to track all the objects
being referenced at any given time. Given that most hazard-pointer-based traversals
require only a few hazard pointers, this is not normally a problem in practice. Of course,
v2023.06.11a
sequence locks provides no pointer-traversal protection, which is why it is normally

used on static data.
Quick Quiz 9.75: Why can’t users dynamically allocate the hazard pointers as they are needed?
The “Duration of Protection” describes constraints (if any) on how long a period
of time a user may protect a given object. Reference counting and hazard pointers
can both protect objects for extended time periods with no untoward side effects, but
maintaining an RCU reference to even one object prevents all other RCU from being
freed. RCU readers must therefore be relatively short in order to avoid running the
system out of memory, with special-purpose implementations such as SRCU, Tasks
RCU, and Tasks Trace RCU being exceptions to this rule. Again, sequence locks provide
no pointer-traversal protection, which is why it is normally used on static data.
The “Need for Traversal Retries” row tells whether a new reference to a given object
may be acquired unconditionally, as it can with RCU, or whether the reference acquisition
can fail, resulting in a retry operation, which is the case for reference counting, hazard
pointers, and sequence locks. In the case of reference counting and hazard pointers,
retries are only required if an attempt to acquire a reference to a given object while
that object is in the process of being deleted, a topic covered in more detail in the
next section. Sequence locking must of course retry its critical section should it run
concurrently with any update.
Quick Quiz 9.76: But don’t Linux-kernel kref reference counters allow guaranteed uncondi-
tional reference acquisition?
Of course, different rows will have different levels of importance in different situations.
For example, if your current code is having read-side scalability problems with hazard
pointers, then it does not matter that hazard pointers can require retrying reference
acquisition because your current code already handles this. Similarly, if response-time
considerations already limit the duration of reader traversals, as is often the case in
kernels and low-level applications, then it does not matter that RCU has duration-limit
requirements because your code already meets them. In the same vein, if readers must
already write to the objects that they are traversing, the read-side overhead of reference
counters might not be so important. Of course, if the data to be protected is in statically
allocated variables, then sequence locking’s inability to protect pointers is irrelevant.
Finally, there is some work on dynamically switching between hazard pointers and
RCU based on dynamic sampling of delays [BGHZ16]. This defers the choice between
hazard pointers and RCU to runtime, and delegates responsibility for the decision to the
software.
Nevertheless, this table should be of great help when choosing between these
techniques. But those wishing more detail should continue on to the next section.
9.6.2 Which to Choose? (Details)

Table 9.8 provides more-detailed rules of thumb that can help you choose among the
four deferred-processing techniques presented in this chapter.
As shown in the “Existence Guarantee” row, if you need existence guarantees for
linked data elements, you must use reference counting, hazard pointers, or RCU.
Sequence locks do not provide existence guarantees, instead providing detection of
updates, retrying any read-side critical sections that do encounter an update.
v2023.06.11a
Table 9.8: Which Deferred Technique to Choose? (Details)
Property Reference Counting Hazard Sequence RCU

Pointers Locks
Existence Guarantees Complex Yes No Yes

Updates and Readers Yes Yes No Yes
Progress Concurrently
Contention Among High None None None
Readers
Reader Per-Critical- N/A N/A Two Ranges from none
Section Overhead smp_mb() to two smp_mb()
Reader Per-Object Read-modify-write atomic smp_mb()* None, but None (volatile
Traversal Overhead operations, memory-barrier unsafe accesses)
instructions, and cache
misses
Reader Forward Progress Lock free Lock free Blocking Bounded wait free
Guarantee
Reader Reference Can fail (conditional) Can fail Unsafe Cannot fail
Acquisition (conditional) (unconditional)
Memory Footprint Bounded Bounded Bounded Unbounded
Reclamation Forward Lock free Lock free N/A Blocking
Progress
Automatic Reclamation Yes Use Case N/A Use Case
Lines of Code 94 79 79 73
* Thissmp_mb() can be downgraded to a compiler barrier() by using the Linux-kernel membarrier()
system call.
Of course, as shown in the “Updates and Readers Progress Concurrently” row, this
detection of updates implies that sequence locking does not permit updaters and readers
to make forward progress concurrently. After all, preventing such forward progress is
the whole point of using sequence locking in the first place! This situation points the
way to using sequence locking in conjunction with reference counting, hazard pointers,
or RCU in order to provide both existence guarantees and update detection. In fact, the
Linux kernel combines RCU and sequence locking in this manner during pathname
lookup.
The “Contention Among Readers”, “Reader Per-Critical-Section Overhead”, and
“Reader Per-Object Traversal Overhead” rows give a rough sense of the read-side
overhead of these techniques. The overhead of reference counting can be quite large,
with contention among readers along with a fully ordered read-modify-write atomic
operation required for each and every object traversed. Hazard pointers incur the
overhead of a memory barrier for each data element traversed, and sequence locks
incur the overhead of a pair of memory barriers for each attempt to execute the critical
section. The overhead of RCU implementations vary from nothing to that of a pair
of memory barriers for each read-side critical section, thus providing RCU with the
best performance, particularly for read-side critical sections that traverse many data
elements. Of course, the read-side overhead of all deferred-processing variants can be
reduced by batching, so that each read-side operation covers more data.
Quick Quiz 9.77: But didn’t the answer to one of the quick quizzes in Section 9.3 say that
pairwise asymmetric barriers could eliminate the read-side smp_mb() from hazard pointers?
v2023.06.11a
The “Reader Forward Progress Guarantee” row shows that only RCU has a bounded
wait-free forward-progress guarantee, which means that it can carry out a finite traversal
by executing a bounded number of instructions.
The “Reader Reference Acquisition” row indicates that only RCU is capable of
unconditionally acquiring references. The entry for sequence locks is “Unsafe” because,
again, sequence locks detect updates rather than acquiring references. Reference
counting and hazard pointers both require that traversals be restarted from the beginning
if a given acquisition fails. To see this, consider a linked list containing objects A, B, C,
and D, in that order, and the following series of events:
1. A reader acquires a reference to object B.
2. An updater removes object B, but refrains from freeing it because the reader holds
a reference. The list now contains objects A, C, and D, and object B’s ->next
pointer is set to HAZPTR_POISON.
3. The updater removes object C, so that the list now contains objects A and D.
Because there is no reference to object C, it is immediately freed.
4. The reader tries to advance to the successor of the object following the now-removed
object B, but the poisoned ->next pointer prevents this. Which is a good thing,
because object B’s ->next pointer would otherwise point to the freelist.
5. The reader must therefore restart its traversal from the head of the list.
Thus, when failing to acquire a reference, a hazard-pointer or reference-counter

traversal must restart that traversal from the beginning. In the case of nested linked data
structures, for example, a tree containing linked lists, the traversal must be restarted
from the outermost data structure. This situation gives RCU a significant ease-of-use
advantage.
However, RCU’s ease-of-use advantage does not come for free, as can be seen in the
“Memory Footprint” row. RCU’s support of unconditional reference acquisition means
that it must avoid freeing any object reachable by a given RCU reader until that reader
completes. RCU therefore has an unbounded memory footprint, at least unless updates
are throttled. In contrast, reference counting and hazard pointers need to retain only
those data elements actually referenced by concurrent readers.
This tension between memory footprint and acquisition failures is sometimes resolved
within the Linux kernel by combining use of RCU and reference counters. RCU is
used for short-lived references, which means that RCU read-side critical sections can be
short. These short RCU read-side critical sections in turn mean that the corresponding
RCU grace periods can also be short, which limits the memory footprint. For the few
data elements that need longer-lived references, reference counting is used. This means
that the complexity of reference-acquisition failure only needs to be dealt with for those
few data elements: The bulk of the reference acquisitions are unconditional, courtesy of
RCU. See Section 13.2 for more information on combining reference counting with
other synchronization mechanisms.
The “Reclamation Forward Progress” row shows that hazard pointers can provide
non-blocking updates [Mic04a, HLM02]. Reference counting might or might not,
depending on the implementation. However, sequence locking cannot provide non-
blocking updates, courtesy of its update-side lock. RCU updaters must wait on readers,
which also rules out fully non-blocking updates. However, there are situations in which
v2023.06.11a
the only blocking operation is a wait to free memory, which results in a situation that,
for many purposes, is as good as non-blocking [DMS+ 12].
As shown in the “Automatic Reclamation” row, only reference counting can automate
freeing of memory, and even then only for non-cyclic data structures. Certain use
cases for hazard pointers and RCU can provide automatic reclamation using link counts,
which can be thought of as reference counts, but applying only to incoming links from
other parts of the data structure [Mic18].
Finally, the “Lines of Code” row shows the size of the Pre-BSD Routing Table
implementations, giving a rough idea of relative ease of use. That said, it is important to
note that the reference-counting and sequence-locking implementations are buggy, and
that a correct reference-counting implementation is considerably more complex [Val95,
MS95]. For its part, a correct sequence-locking implementation requires the addition of
some other synchronization mechanism, for example, hazard pointers or RCU, so that
sequence locking detects concurrent updates and the other mechanism provides safe
reference acquisition.
As more experience is gained using these techniques, both separately and in combi-
nation, the rules of thumb laid out in this section will need to be refined. However, this
section does reflect the current state of the art.
9.6.3 Which to Choose? (Production Use)

This section points out a few publicly visible production uses of hazard pointers, sequence
locking, and RCU. Reference counting is omitted, not because it is unimportant, but
rather because it is not only used pervasively, but heavily documented in textbooks
going back a half century. One of the hoped-for benefits of listing production uses of
these other techniques is to provide examples to study—or to find bugs in, as the case
may be.21
9.6.3.1 Production Uses of Hazard Pointers

In 2010, Keith Bostic added hazard pointers to WiredTiger [Bos10]. MongoDB 3.0,
released in 2015, included WiredTiger and thus hazard pointers.
In 2011, Samy Al Bahra added hazard pointers to the Concurrency Kit li-
brary [Bah11b].
In 2014, Maxim Khizhinsky added hazard pointers to libcds [Khi14].
In 2015, David Gwynne introduced shared reference pointers, a form of hazard
pointers, to OpenBSD [Gwy15].
In 2017–2018, the Rust-language arc-swap [Van18] and conc [cut17] crates rolled
their own implementations of hazard pointers.
In 2018, Maged Michael added hazard pointers to Facebook’s Folly library [Mic18],
where it is used heavily.
9.6.3.2 Production Uses of Sequence Locking

The Linux kernel added sequence locking to v2.5.60 in 2003 [Cor03], having been gener-
alized from an ad-hoc technique used in x86’s implementation of the gettimeofday()
system call.
21 Kudos to Mathias Stearn, Matt Wilson, David Goldblatt, LiveJournal user fanf, Nadav
Har’El, Avi Kivity, Dmitry Vyukov, Raul Guitterez S., Twitter user @peo3, Paolo Bonzini,
and Thomas Monjalon for locating a great many of these use cases.
v2023.06.11a
In 2011, Samy Al Bahra added sequence locking to the Concurrency Kit li-
brary [Bah11c].
Paolo Bonzini added a simple sequence-lock to the QEMU emulator in 2013 [Bon13].
Alexis Menard abstracted a sequence-lock implementation in Chromium in
2016 [Men16].
A simple sequence locking implementation was added to jemalloc() in
2018 [Gol18a]. The eigen library also has a special-purpose queue that is managed by a
mechanism resembling sequence locking.
9.6.3.3 Production Uses of RCU

IBM’s VM/XA is adopted passive serialization, a mechanism similar to RCU, some
time in the 1980s [HOS89].
DYNIX/ptx adopted RCU in 1993 [MS98a, SM95].
The Linux kernel adopted Dipankar Sarma’s implementation of RCU in 2002 [Tor02].
The userspace RCU project started in 2009 [Des09b].
The Knot DNS project started using the userspace RCU library in 2010 [Slo10]. That
same year, the OSv kernel added an RCU implementation [Kiv13], later adding an
RCU-protected linked list [Kiv14b] and an RCU-protected hash table [Kiv14a].
In 2011, Samy Al Bahra added epochs (a form of RCU [Fra04, FH07]) to the
Concurrency Kit library [Bah11a].
NetBSD began using the aforementioned passive serialization with v6.0 in
2012 [The12a]. Among other things, passive serialization is used in NetBSD packet
filter (NPF) [Ras14].
Paolo Bonzini added RCU support to the QEMU emulator in 2015 via a friendly fork
of the userspace RCU library [BD13, Bon15].
In 2015, Maxim Khizhinsky added RCU to libcds [Khi15].
Mindaugas Rasiukevicius implemented libqsbr in 2016, which features QSBR and
epoch-based reclamation (EBR) [Ras16], both of which are types of implementations
of RCU.
Sheth et al. [SWS16] demonstrated the value of leveraging Go’s garbage collector to
provide RCU-like functionality, and the Go programming language provides a Value
type that can provide this functionality.22
Matt Klein describes an RCU-like mechanism that is used in the Envoy Proxy [Kle17].
Honnappa Nagarahalli added an RCU library to the Data Plane Development Kit
(DPDK) in 2018 [Nag18].
Stjepan Glavina merged an epoch-based RCU implementation into the crossbeam
set of concurrency-support “crates” for the Rust language [Gla18].
Jason Donenfeld produced an RCU implementations as part of his port of WireGuard
to Windows NT kernel [Don21].
Finally, any garbage-collected concurrent language (not just Go!) gets the update side
of an RCU implementation at zero incremental cost.
9.6.3.4 Summary of Production Uses

Perhaps the time will come when sequence locking, hazard pointers, and RCU are all as
heavily used and as well known as are reference counters. Until that time comes, the
current production uses of these mechanisms should help guide the choice of mechanism
22 See https://golang.org/pkg/sync/atomic/#Value, particularly the “Example
(ReadMostly)”.
v2023.06.11a
9.7. WHAT ABOUT UPDATES? 287
as well as showing how best to apply each of them. And with that, we have uncovered
the last of the mysteries put forth on page 201.
The next section discusses updates, a ticklish issue for many of the read-mostly
mechanisms described in this chapter.
9.7 What About Updates?

The only thing constant in life is change.
François de la Rochefoucauld
The deferred-processing techniques called out in this chapter are most directly applicable
to read-mostly situations, which begs the question “But what about updates?” After all,
increasing the performance and scalability of readers is all well and good, but it is only
natural to also want great performance and scalability for writers.
We have already seen one situation featuring high performance and scalability
for writers, namely the counting algorithms surveyed in Chapter 5. These algorithms
featured partially partitioned data structures so that updates can operate locally, while the
more-expensive reads must sum across the entire data structure. Silas Boyd-Wickhizer
has generalized this notion to produce OpLog, which he has applied to Linux-kernel
pathname lookup, VM reverse mappings, and the stat() system call [BW14].
Another approach, called “Disruptor”, is designed for applications that process
high-volume streams of input data. The approach is to rely on single-producer-single-
consumer FIFO queues, minimizing the need for synchronization [Sut13]. For Java
applications, Disruptor also has the virtue of minimizing use of the garbage collector.
And of course, where feasible, fully partitioned or “sharded” systems provide excellent
performance and scalability, as noted in Chapter 6.
The next chapter will look at updates in the context of several types of data structures.
v2023.06.11a
v2023.06.11a
Bad programmers worry about the code. Good
programmers worry about data structures and their
relationships.
Linus Torvalds
Chapter 10
Data Structures
Serious discussions of algorithms include time complexity of their data struc-

tures [CLRS01]. However, for parallel programs, the time complexity includes
concurrency effects because these effects can be overwhelmingly large, as shown
in Chapter 3. In other words, a good programmer’s data-structure relationships include
those aspects related to concurrency.
This chapter will expose a number of complications:
1. Data structures designed in full accordance with the good advice given in Chapter 6
can nonetheless abjectly fail to scale on some types of systems.
2. Data structures designed in full accordance with the good advice given in both
Chapter 6 and Chapter 9 can still abjectly fail to scale on some types of systems.
3. Even read-only synchronization-free data-structure traversal can fail to scale on

some types of systems.
4. Data-structure traverals avoiding the aforementioned complications can still be

impeded by concurrent updates.
Section 10.1 presents the motivating application for this chapter’s data structures.
Chapter 6 showed how partitioning improves scalability, so Section 10.2 discusses
partitionable data structures. Chapter 9 described how deferring some actions can
greatly improve both performance and scalability, a topic taken up by Section 10.3.
Section 10.4 looks at a non-partitionable data structure, splitting it into read-mostly and
partitionable portions, which improves both performance and scalability. Because this
chapter cannot delve into the details of every concurrent data structure, Section 10.5
surveys a few of the important ones. Although the best performance and scalability
results from design rather than after-the-fact micro-optimization, micro-optimization is
nevertheless necessary for the absolute best possible performance and scalability, as
described in Section 10.6. Finally, Section 10.7 presents a summary of this chapter.
289
v2023.06.11a
290 CHAPTER 10. DATA STRUCTURES
10.1 Motivating Application

The art of doing mathematics consists in finding that
special case which contains all the germs of
generality.
David Hilbert
We will use the Schrödinger’s Zoo application to evaluate performance [McK13].

Schrödinger has a zoo containing a large number of animals, and he would like to track
them using an in-memory database with each animal in the zoo represented by a data
item in this database. Each animal has a unique name that is used as a key, with a variety
of data tracked for each animal.
Births, captures, and purchases result in insertions, while deaths, releases, and sales
result in deletions. Because Schrödinger’s zoo contains a large quantity of short-lived
animals, including mice and insects, the database must handle high update rates.
Those interested in Schrödinger’s animals can query them, and Schrödinger has noted
suspiciously query rates for his cat, so much so that he suspects that his mice might be
checking up on their nemesis. Whatever their source, Schrödinger’s application must
handle high query rates to a single data element.
As we will see, this simple application can be a challenge to concurrent data structures.
10.2 Partitionable Data Structures

Finding a way to live the simple life today is the
most complicated task.
Henry A. Courtney, updated
There are a huge number of data structures in use today, so much so that there are
multiple textbooks covering them. This section focuses on a single data structure,
namely the hash table. This focused approach allows a much deeper investigation of
how concurrency interacts with data structures, and also focuses on a data structure that
is heavily used in practice. Section 10.2.1 overviews the design, and Section 10.2.2
presents the implementation. Finally, Section 10.2.3 discusses the resulting performance
and scalability.
10.2.1 Hash-Table Design

Chapter 6 emphasized the need to apply partitioning in order to attain respectable
performance and scalability, so partitionability must be a first-class criterion when
selecting data structures. This criterion is well satisfied by that workhorse of parallelism,
the hash table. Hash tables are conceptually simple, consisting of an array of hash
buckets. A hash function maps from a given element’s key to the hash bucket that this
element will be stored in. Each hash bucket therefore heads up a linked list of elements,
called a hash chain. When properly configured, these hash chains will be quite short,
permitting a hash table to access its elements extremely efficiently.
Quick Quiz 10.1: But chained hash tables are but one type of many. Why the focus on chained
hash tables?
v2023.06.11a
10.2. PARTITIONABLE DATA STRUCTURES 291
Listing 10.1: Hash-Table Data Structures

1 struct ht_elem {
2 struct cds_list_head hte_next;
3 unsigned long hte_hash;
4 };
5
6 struct ht_bucket {
7 struct cds_list_head htb_head;
8 spinlock_t htb_lock;
9 };
10
11 struct hashtab {
12 unsigned long ht_nbuckets;
13 int (*ht_cmp)(struct ht_elem *htep, void *key);
14 struct ht_bucket ht_bkt[0];
15 };
struct hashtab
−>ht_nbuckets = 4
−>ht_cmp
−>ht_bkt[0] struct ht_elem struct ht_elem
−>htb_head −>hte_next −>hte_next
−>htb_lock −>hte_hash −>hte_hash
−>ht_bkt[1]
−>htb_head
−>htb_lock
−>ht_bkt[2] struct ht_elem
−>htb_head −>hte_next
−>htb_lock −>hte_hash
−>ht_bkt[3]
−>htb_head
−>htb_lock
Figure 10.1: Hash-Table Data-Structure Diagram
In addition, each bucket has its own lock, so that elements in different buckets of the
hash table may be added, deleted, and looked up completely independently. A large
hash table with a large number of buckets (and thus locks), with each bucket containing
a small number of elements should therefore provide excellent scalability.
10.2.2 Hash-Table Implementation

Listing 10.1 (hash_bkt.c) shows a set of data structures used in a simple fixed-sized
hash table using chaining and per-hash-bucket locking, and Figure 10.1 diagrams how
they fit together. The hashtab structure (lines 11–15 in Listing 10.1) contains four ht_
bucket structures (lines 6–9 in Listing 10.1), with the ->ht_nbuckets field controlling
the number of buckets and the ->ht_cmp field holding the pointer to key-comparison
function. Each such bucket contains a list header ->htb_head and a lock ->htb_lock.
The list headers chain ht_elem structures (lines 1–4 in Listing 10.1) through their ->
hte_next fields, and each ht_elem structure also caches the corresponding element’s
hash value in the ->hte_hash field. The ht_elem structure is included in a larger
structure which might contain a complex key.
Figure 10.1 shows bucket 0 containing two elements and bucket 2 containing one.
Listing 10.2 shows mapping and locking functions. Lines 1 and 2 show the macro
HASH2BKT(), which maps from a hash value to the corresponding ht_bucket structure.
v2023.06.11a
Listing 10.2: Hash-Table Mapping and Locking

1 #define HASH2BKT(htp, h) \
2 (&(htp)->ht_bkt[h % (htp)->ht_nbuckets])
3
4 static void hashtab_lock(struct hashtab *htp,
5 unsigned long hash)
6 {
7 spin_lock(&HASH2BKT(htp, hash)->htb_lock);
8 }
9
10 static void hashtab_unlock(struct hashtab *htp,
12 {
13 spin_unlock(&HASH2BKT(htp, hash)->htb_lock);
14 }
Listing 10.3: Hash-Table Lookup

1 struct ht_elem *
2 hashtab_lookup(struct hashtab *htp, unsigned long hash,
3 void *key)
4 {
5 struct ht_bucket *htb;
6 struct ht_elem *htep;
7
8 htb = HASH2BKT(htp, hash);
9 cds_list_for_each_entry(htep, &htb->htb_head, hte_next) {
10 if (htep->hte_hash != hash)
11 continue;
12 if (htp->ht_cmp(htep, key))
13 return htep;
14 }
15 return NULL;
16 }
This macro uses a simple modulus: If more aggressive hashing is required, the caller
needs to implement it when mapping from key to hash value. The remaining two
functions acquire and release the ->htb_lock corresponding to the specified hash
value.
Listing 10.3 shows hashtab_lookup(), which returns a pointer to the element with
the specified hash and key if it exists, or NULL otherwise. This function takes both a hash
value and a pointer to the key because this allows users of this function to use arbitrary
keys and arbitrary hash functions. Line 8 maps from the hash value to a pointer to the
corresponding hash bucket. Each pass through the loop spanning lines 9–14 examines
one element of the bucket’s hash chain. Line 10 checks to see if the hash values match,
and if not, line 11 proceeds to the next element. Line 12 checks to see if the actual key
matches, and if so, line 13 returns a pointer to the matching element. If no element
matches, line 15 returns NULL.
Quick Quiz 10.2: But isn’t the double comparison on lines 10–13 in Listing 10.3 inefficient in
the case where the key fits into an unsigned long?
Listing 10.4 shows the hashtab_add() and hashtab_del() functions that add and
delete elements from the hash table, respectively.
The hashtab_add() function simply sets the element’s hash value on line 4, then
adds it to the corresponding bucket on lines 5 and 6. The hashtab_del() function
simply removes the specified element from whatever hash chain it is on, courtesy of the
doubly linked nature of the hash-chain lists. Before calling either of these two functions,
the caller is required to ensure that no other thread is accessing or modifying this same
bucket, for example, by invoking hashtab_lock() beforehand.
v2023.06.11a
10.2. PARTITIONABLE DATA STRUCTURES 293
Listing 10.4: Hash-Table Modification

1 void hashtab_add(struct hashtab *htp, unsigned long hash,
2 struct ht_elem *htep)
3 {
4 htep->hte_hash = hash;
5 cds_list_add(&htep->hte_next,
6 &HASH2BKT(htp, hash)->htb_head);
7 }
8
9 void hashtab_del(struct ht_elem *htep)
10 {
11 cds_list_del_init(&htep->hte_next);
12 }
Listing 10.5: Hash-Table Allocation and Free

1 struct hashtab *
2 hashtab_alloc(unsigned long nbuckets,
3 int (*cmp)(struct ht_elem *htep, void *key))
4 {
5 struct hashtab *htp;
6 int i;
7
8 htp = malloc(sizeof(*htp) +
9 nbuckets * sizeof(struct ht_bucket));
10 if (htp == NULL)
11 return NULL;
12 htp->ht_nbuckets = nbuckets;
13 htp->ht_cmp = cmp;
14 for (i = 0; i < nbuckets; i++) {
15 CDS_INIT_LIST_HEAD(&htp->ht_bkt[i].htb_head);
16 spin_lock_init(&htp->ht_bkt[i].htb_lock);
17 }
18 return htp;
19 }
20
21 void hashtab_free(struct hashtab *htp)
22 {
23 free(htp);
24 }
Listing 10.5 shows hashtab_alloc() and hashtab_free(), which do hash-table

allocation and freeing, respectively. Allocation begins on lines 8–9 with allocation of the
underlying memory. If line 10 detects that memory has been exhausted, line 11 returns
NULL to the caller. Otherwise, lines 12 and 13 initialize the number of buckets and
the pointer to key-comparison function, and the loop spanning lines 14–17 initializes
the buckets themselves, including the chain list header on line 15 and the lock on
line 16. Finally, line 18 returns a pointer to the newly allocated hash table. The
hashtab_free() function on lines 21–24 is straightforward.
10.2.3 Hash-Table Performance

The performance results for a single 28-core socket of a 2.1 GHz Intel Xeon system
using a bucket-locked hash table with 262,144 buckets are shown in Figure 10.2. The
performance does scale nearly linearly, but it falls a far short of the ideal performance
level, even at only 28 CPUs. Part of this shortfall is due to the fact that the lock
acquisitions and releases incur no cache misses on a single CPU, but do incur misses on
two or more CPUs.
And things only get worse with more CPUs, as can be seen in Figure 10.3. We do
not need to show ideal performance: The performance for 29 CPUs and beyond is all
v2023.06.11a
6
1.4x10
6
Total Lookups per Millisecond

1.2x10
1x106
800000
ideal
600000
400000
200000
bucket
0
5 10 15 20 25
Figure 10.2: Read-Only Hash-Table Performance For Schrödinger’s Zoo
250000
200000
150000
100000
50000
0
0 50 100 150 200 250 300 350 400 450
Figure 10.3: Read-Only Hash-Table Performance For Schrödinger’s Zoo, 448 CPUs
too clearly worse than abysmal. This clearly underscores the dangers of extrapolating
performance from a modest number of CPUs.
Of course, one possible reason for the collapse in performance might be that more
hash buckets are needed. We can test this by increasing the number of hash buckets.
Quick Quiz 10.3: Instead of simply increasing the number of hash buckets, wouldn’t it be
better to cache-align the existing hash buckets?
However, as can be seen in Figure 10.4, changing the number of buckets has almost no
effect: Scalability is still abysmal. In particular, we still see a sharp dropoff at 29 CPUs
and beyond, clearly demonstrating the complication put forward on page 289. And just
as clearly, something else is going on.
The problem is that this is a multi-socket system, with CPUs 0–27 and 225–251
mapped to the first socket as shown in Figure 10.5. Test runs confined to the first
28 CPUs therefore perform quite well, but tests that involve socket 0’s CPUs 0–27 as
well as socket 1’s CPU 28 incur the overhead of passing data across socket boundaries.
This can severely degrade performance, as was discussed in Section 3.2.1. In short, large
multi-socket systems require good locality of reference in addition to full partitioning.
The remainder of this chapter will discuss ways of providing good locality of reference
v2023.06.11a
10.3. READ-MOSTLY DATA STRUCTURES 295
250000

200000
150000
100000
50000
0
0 50 100 150 200 250 300 350 400 450
Figure 10.4: Read-Only Hash-Table Performance For Schrödinger’s Zoo, Varying

Buckets
within the hash table itself, but in the meantime please note that one other way to provide
good locality of reference would be to place large data elements in the hash table. For
example, Schrödinger might attain excellent cache locality by placing photographs or
even videos of his animals in each element of the hash table. But for those needing hash
tables containing small data elements, please read on!
Quick Quiz 10.4: Given the negative scalability of the Schrödinger’s Zoo application across
sockets, why not just run multiple copies of the application, with each copy having a subset of
the animals and confined to run on a single socket?
One key property of the Schrödinger’s-zoo runs discussed thus far is that they are all
read-only. This makes the performance degradation due to lock-acquisition-induced
cache misses all the more painful. Even though we are not updating the underlying hash
table itself, we are still paying the price for writing to memory. Of course, if the hash
table was never going to be updated, we could dispense entirely with mutual exclusion.
This approach is quite straightforward and is left as an exercise for the reader. But
even with the occasional update, avoiding writes avoids cache misses, and allows the
read-mostly data to be replicated across all the caches, which in turn promotes locality
of reference.
The next section therefore examines optimizations that can be carried out in read-
mostly cases where updates are rare, but could happen at any time.
10.3 Read-Mostly Data Structures
Adapt the remedy to the disease.

Chinese proverb
Although partitioned data structures can offer excellent scalability, NUMA effects can
result in severe degradations of both performance and scalability. In addition, the
need for read-side synchronization can degrade performance in read-mostly situations.
However, we can achieve both performance and scalability by using RCU, which
was introduced in Section 9.5. Similar results can be achieved using hazard pointers
v2023.06.11a
Hyperthread
Socket 0 1
0 0–27 224–251
1 28–55 252–279
2 56–83 280–307
3 84–111 308–335
4 112–139 336–363
5 140–167 364–391
6 168–195 392–419
7 196–223 420–447
Figure 10.5: NUMA Topology of System Under Test
Listing 10.6: RCU-Protected Hash-Table Read-Side Concurrency Control

1 static void hashtab_lock_lookup(struct hashtab *htp,
3 {
4 rcu_read_lock();
5 }
6
7 static void hashtab_unlock_lookup(struct hashtab *htp,
9 {
11 }
(hazptr.c) [Mic04a], which will be included in the performance results shown in this
section [McK13].
10.3.1 RCU-Protected Hash Table Implementation

For an RCU-protected hash table with per-bucket locking, updaters use locking as shown
in Section 10.2, but readers use RCU. The data structures remain as shown in Listing 10.1,
and the HASH2BKT(), hashtab_lock(), and hashtab_unlock() functions remain
as shown in Listing 10.2. However, readers use the lighter-weight concurrency-control
embodied by hashtab_lock_lookup() and hashtab_unlock_lookup() shown in
Listing 10.6.
Listing 10.7 shows hashtab_lookup() for the RCU-protected per-bucket-locked
hash table. This is identical to that in Listing 10.3 except that cds_list_for_
each_entry() is replaced by cds_list_for_each_entry_rcu(). Both of these
primitives traverse the hash chain referenced by htb->htb_head but cds_list_for_
each_entry_rcu() also correctly enforces memory ordering in case of concurrent
insertion. This is an important difference between these two hash-table implementations:
Unlike the pure per-bucket-locked implementation, the RCU protected implementation
allows lookups to run concurrently with insertions and deletions, and RCU-aware
primitives like cds_list_for_each_entry_rcu() are required to correctly handle
this added concurrency. Note also that hashtab_lookup()’s caller must be within an
RCU read-side critical section, for example, the caller must invoke hashtab_lock_
lookup() before invoking hashtab_lookup() (and of course invoke hashtab_
unlock_lookup() some time afterwards).
v2023.06.11a
Listing 10.7: RCU-Protected Hash-Table Lookup

1 struct ht_elem *hashtab_lookup(struct hashtab *htp,
2 unsigned long hash,
3 void *key)
4 {
5 struct ht_bucket *htb;
7
8 htb = HASH2BKT(htp, hash);
9 cds_list_for_each_entry_rcu(htep,
10 &htb->htb_head,
11 hte_next) {
12 if (htep->hte_hash != hash)
13 continue;
15 return htep;
16 }
17 return NULL;
18 }
Listing 10.8: RCU-Protected Hash-Table Modification

1 void hashtab_add(struct hashtab *htp,
2 unsigned long hash,
3 struct ht_elem *htep)
4 {
5 htep->hte_hash = hash;
6 cds_list_add_rcu(&htep->hte_next,
7 &HASH2BKT(htp, hash)->htb_head);
8 }
9
10 void hashtab_del(struct ht_elem *htep)
11 {
12 cds_list_del_rcu(&htep->hte_next);
13 }
Quick Quiz 10.5: But if elements in a hash table can be removed concurrently with lookups,
doesn’t that mean that a lookup could return a reference to a data element that was removed
immediately after it was looked up?
Listing 10.8 shows hashtab_add() and hashtab_del(), both of which are quite
similar to their counterparts in the non-RCU hash table shown in Listing 10.4. The
hashtab_add() function uses cds_list_add_rcu() instead of cds_list_add()
in order to ensure proper ordering when an element is added to the hash table at the
same time that it is being looked up. The hashtab_del() function uses cds_list_
del_rcu() instead of cds_list_del_init() to allow for the case where an element
is looked up just before it is deleted. Unlike cds_list_del_init(), cds_list_del_
rcu() leaves the forward pointer intact, so that hashtab_lookup() can traverse to the
newly deleted element’s successor.
Of course, after invoking hashtab_del(), the caller must wait for an RCU grace
period (e.g., by invoking synchronize_rcu()) before freeing or otherwise reusing
the memory for the newly deleted element.
10.3.2 RCU-Protected Hash Table Validation

Although the topic of validation is covered in detail in Chapter 11, the fact is that a hash
table with lockless RCU-protected lookups needs special attention to validation sooner
rather than later.
v2023.06.11a
1x108

1x107
ideal U
RC zptr
1x106 ha
100000 bucket
10000 global
1000
1 10 100
Figure 10.6: Read-Only RCU-Protected Hash-Table Performance For Schrödinger’s

Zoo
The test suite (“hashtorture.h”) contains a smoketest() function that verifies

that a specific series of single-threaded additions, deletions, and lookups give the
expected results.
Concurrent test runs put each updater thread in control of its portion of the elements,
which allows assertions checking for the following issues:
1. A just-now-to-be-added element already being in the table according to hastab_

lookup().
2. A just-now-to-be-added element being marked as being in the table by its ->in_
table flag.
3. A just-now-to-be-deleted element not being in the table according to hastab_
lookup().
4. A just-now-to-be-deleted element being marked as not being in the table by its
->in_table flag.
In addition, concurrent test runs run lookups concurrently with updates in order to
catch all manner of data-structure corruption problems. Some runs also continually
resize the hash table concurrently with both lookups and updates to verify correct
behavior, and also to verify that resizes do not unduly delay either readers or updaters.
Finally, the concurrent tests output statistics that can be used to track down performance
and scalabilty issues, which provides the raw data used by Section 10.3.3.
Quick Quiz 10.6: The hashtorture.h file contains more than 1,000 lines! Is that a
comprehensive test or what???
All code requires significant validation effort, and high-performance concurrent code
requires more validation than most.
10.3.3 RCU-Protected Hash Table Performance

Figure 10.6 shows the read-only performance of RCU-protected and hazard-pointer-
protected hash tables against the previous section’s per-bucket-locked implementation.
v2023.06.11a
2.2x107
2x107

7
1.8x10
1.6x107
7
1.4x10 ideal
1.2x107
1x107
8x106
6
6x10
4x106 RCU
6
2x10 hazptr
0
0 50 100 150 200 250 300 350 400 450

Zoo, Linear Scale
As you can see, both RCU and hazard pointers perform and scale much better than
per-bucket locking because read-only replication avoids NUMA effects. The difference
increases with larger numbers of threads. Results from a globally locked implementation
are also shown, and as expected the results are even worse than those of the per-bucket-
locked implementation. RCU does slightly better than hazard pointers.
Figure 10.7 shows the same data on a linear scale. This drops the global-locking
trace into the x-axis, but allows the non-ideal performance of RCU and hazard pointers
to be more readily discerned. Both show a change in slope at 224 CPUs, and this is due
to hardware multithreading. At 224 and fewer CPUs, each thread has a core to itself. In
this regime, RCU does better than does hazard pointers because the latter’s read-side
memory barriers result in dead time within the core. In short, RCU is better able to
utilize a core from a single hardware thread than is hazard pointers.
This situation changes above 224 CPUs. Because RCU is using more than half of
each core’s resources from a single hardware thread, RCU gains relatively little benefit
from the second hardware thread in each core. The slope of the hazard-pointers trace
also decreases at 224 CPUs, but less dramatically, because the second hardware thread
is able to fill in the time that the first hardware thread is stalled due to memory-barrier
latency. As we will see in later sections, this second-hardware-thread advantage depends
on the workload.
But why is RCU’s performance a factor of five less than ideal? One possibility is that
the per-thread counters manipulated by rcu_read_lock() and rcu_read_unlock()
are slowing things down. Figure 10.8 therefore adds the results for the QSBR variant
of RCU, whose read-side primitives do nothing. And although QSBR does perform
slightly better than does RCU, it is still about a factor of five short of ideal.
Figure 10.9 adds completely unsynchronized results, which works because this is
a read-only benchmark with nothing to synchronize. Even with no synchronization
whatsoever, performance still falls far short of ideal, thus demonstrating two more
complications on page 289.
The problem is that this system has sockets with 28 cores, which have the modest
cache sizes shown in Table 3.2 on page 37. Each hash bucket (struct ht_bucket)
occupies 56 bytes and each element (struct zoo_he) occupies 72 bytes for the RCU
and QSBR runs. The benchmark generating Figure 10.9 used 262,144 buckets and
v2023.06.11a
2.2x107
2x107

1.8x107
1.6x107
7
1.4x10 ideal
1.2x107
1x107
8x106
6
6x10
4x106 QSBR,RCU
2x106 hazptr
0
0 50 100 150 200 250 300 350 400 450

Zoo including QSBR, Linear Scale
2.2x107
2x107
1.8x107
1.6x107
1.4x107 ideal
1.2x107
1x107
8x106
6x106
unsync,QSBR,RCU
6
4x10
2x106 hazptr
0
0 50 100 150 200 250 300 350 400 450

Zoo including QSBR and Unsynchronized, Linear Scale
v2023.06.11a
1x107
1x106
Cat Lookups per Millisecond

RCU
100000 hazptr
10000
bucket
1000
100 global
10
1 10
Number of CPUs Looking Up The Cat
Figure 10.10: Read-Side Cat-Only RCU-Protected Hash-Table Performance For

Schrödinger’s Zoo at 64 CPUs
up to 262,144 elements, for a total of 33,554,448 bytes, which not only overflows the
1,048,576-byte L2 caches by more than a factor of thirty, but is also uncomfortably
close to the L3 cache size of 40,370,176 bytes, especially given that this cache has
only 11 ways. This means that L2 cache collisions will be the rule and also that L3
cache collisions will not be uncommon, so that the resulting cache misses will degrade
performance. In this case, the bottleneck is not in the CPU, but rather in the hardware
memory system.
Additional evidence for this memory-system bottleneck may be found by examining
the unsynchronized code. This code does not need locks, so each hash bucket occupies
only 16 bytes compared to the 56 bytes for RCU and QSBR. Similarly, each hash-table
element occupies only 56 bytes compared to the 72 bytes for RCU and QSBR. So it is
unsurprising that the single-CPU unsynchronized run performs up to about half again
faster than that of either QSBR or RCU.
Quick Quiz 10.7: How can we be so sure that the hash-table size is at fault here, especially
given that Figure 10.4 on page 295 shows that varying hash-table size has almost no effect?
Might the problem instead be something like false sharing?
What if the memory footprint is reduced still further? Figure E.5 on page 788 shows
that RCU attains very nearly ideal performance on the much smaller data structure
represented by the pre-BSD routing table.
Quick Quiz 10.8: The memory system is a serious bottleneck on this big system. Why bother
putting 448 CPUs on a system without giving them enough memory bandwidth to do something
useful???
As noted earlier, Schrödinger is surprised by the popularity of his cat [Sch35], but
recognizes the need to reflect this popularity in his design. Figure 10.10 shows the
results of 64-CPU runs, varying the number of CPUs that are doing nothing but looking
up the cat. Both RCU and hazard pointers respond well to this challenge, but bucket
locking scales negatively, eventually performing as badly as global locking. This should
not be a surprise because if all CPUs are doing nothing but looking up the cat, the lock
corresponding to the cat’s bucket is for all intents and purposes a global lock.
This cat-only benchmark illustrates one potential problem with fully partitioned
sharding approaches. Only the CPUs associated with the cat’s partition is able to access
v2023.06.11a
7
1x10
RCU
6
1x10

hazptr
100000 bucket
10000
1000 global
100
1 10 100
Number of CPUs Doing Updates
Figure 10.11: Read-Side RCU-Protected Hash-Table Performance For Schrödinger’s

Zoo in the Presence of Updates
the cat, limiting the cat-only throughput. Of course, a great many applications have
good load-spreading properties, and for these applications sharding works quite well.
However, sharding does not handle “hot spots” very well, with the hot spot exemplified
by Schrödinger’s cat being but one case in point.
If we were only ever going to read the data, we would not need any concurrency
control to begin with. Figure 10.11 therefore shows the effect of updates on readers.
At the extreme left-hand side of this graph, all but one of the CPUs are doing lookups,
while to the right all 448 CPUs are doing updates. For all four implementations, the
number of lookups per millisecond decreases as the number of updating CPUs increases,
of course reaching zero lookups per millisecond when all 448 CPUs are updating. Both
hazard pointers and RCU do well compared to per-bucket locking because their readers
do not increase update-side lock contention. RCU does well relative to hazard pointers
as the number of updaters increases due to the latter’s read-side memory barriers, which
incur greater overhead, especially in the presence of updates, and particularly when
execution involves more than one socket. It therefore seems likely that modern hardware
heavily optimizes memory-barrier execution, greatly reducing memory-barrier overhead
in the read-only case.
Where Figure 10.11 showed the effect of increasing update rates on lookups, Fig-
ure 10.12 shows the effect of increasing update rates on the updates themselves. Again,
at the left-hand side of the figure all but one of the CPUs are doing lookups and at the
right-hand side of the figure all 448 CPUs are doing updates. Hazard pointers and RCU
start off with a significant advantage because, unlike bucket locking, readers do not
exclude updaters. However, as the number of updating CPUs increases, update-side
overhead starts to make its presence known, first for RCU and then for hazard pointers.
Of course, all three of these implementations beat global locking.
It is quite possible that the differences in lookup performance observed in Figure 10.11
are affected by the differences in update rates. One way to check this is to artificially
throttle the update rates of per-bucket locking and hazard pointers to match that of
RCU. Doing so does not significantly improve the lookup performance of per-bucket
locking, nor does it close the gap between hazard pointers and RCU. However,
removing the read-side memory barriers from hazard pointers (thus resulting in an
unsafe implementation) does nearly close the gap between hazard pointers and RCU.
v2023.06.11a
6
1x10
bucket
100000
Updates per Millisecond

hazptr
10000 RCU
1000
global
100
10
1 10 100
Number of CPUs Doing Updates
Figure 10.12: Update-Side RCU-Protected Hash-Table Performance For Schrödinger’s

Zoo
Although this unsafe hazard-pointer implementation will usually be reliable enough for
benchmarking purposes, it is absolutely not recommended for production use.
Quick Quiz 10.9: The dangers of extrapolating from 28 CPUs to 448 CPUs was made quite
clear in Section 10.2.3. Would extrapolating up from 448 CPUs be any safer?
And this situation exposes yet another of the complications listed on page 289.
10.3.4 RCU-Protected Hash Table Discussion

One consequence of the RCU and hazard-pointer implementations is that a pair of
concurrent readers might disagree on the state of the cat. For example, one of the readers
might have fetched the pointer to the cat’s data structure just before it was removed,
while another reader might have fetched this same pointer just afterwards. The first
reader would then believe that the cat was alive, while the second reader would believe
that the cat was dead.
This situation is completely fitting for Schrödinger’s cat, but it turns out that it is quite
reasonable for normal non-quantum cats as well. After all, it is impossible to determine
exactly when an animal is born or dies.
To see this, let’s suppose that we detect a cat’s death by heartbeat. This raise the
question of exactly how long we should wait after the last heartbeat before declaring
death. It is clearly ridiculous to wait only one millisecond, because then a healthy living
cat would have to be declared dead—and then resurrected—more than once per second.
It is equally ridiculous to wait a full month, because by that time the poor cat’s death
would have made itself very clearly known via olfactory means.
Because an animal’s heart can stop for some seconds and then start up again, there
is a tradeoff between timely recognition of death and probability of false alarms. It is
quite possible that a pair of veterinarians might disagree on the time to wait between the
last heartbeat and the declaration of death. For example, one veterinarian might declare
death thirty seconds after the last heartbeat, while another might insist on waiting a full
minute. In this case, the two veterinarians would disagree on the state of the cat for the
second period of thirty seconds following the last heartbeat, as fancifully depicted in
Figure 10.13.
v2023.06.11a
Figure 10.13: Even Veterinarians Disagree!
Heisenberg taught us to live with this sort of uncertainty [Hei27], which is a good
thing because computing hardware and software acts similarly. For example, how do
you know that a piece of computing hardware has failed? Often because it does not
respond in a timely fashion. Just like the cat’s heartbeat, this results in a window of
uncertainty as to whether or not the hardware has really failed, as opposed to just being
slow.
Furthermore, most computing systems are intended to interact with the outside world.
Consistency with the outside world is therefore of paramount importance. However,
as we saw in Figure 9.28 on page 266, increased internal consistency can come at the
expense of degraded external consistency. Techniques such as RCU and hazard pointers
give up some degree of internal consistency to attain improved external consistency.
In short, internal consistency is not necessarily a natural part of all problem domains,
and often incurs great expense in terms of performance, scalability, consistency with
the outside world [HKLP12, HHK+ 13, Rin13], or all of the above.
10.4 Non-Partitionable Data Structures

Don’t be afraid to take a big step if one is indicated.
You can’t cross a chasm in two small steps.
David Lloyd George
Fixed-size hash tables are perfectly partitionable, but resizable hash tables pose parti-
tioning challenges when growing or shrinking, as fancifully depicted in Figure 10.14.
However, it turns out that it is possible to construct high-performance scalable RCU-
protected hash tables, as described in the following sections.
10.4.1 Resizable Hash Table Design

In happy contrast to the situation in the early 2000s, there are now no fewer than three
different types of scalable RCU-protected hash tables. The first (and simplest) was
developed for the Linux kernel by Herbert Xu [Xu10], and is described in the following
sections. The other two are covered briefly in Section 10.4.4.
v2023.06.11a
10.4. NON-PARTITIONABLE DATA STRUCTURES 305
Figure 10.14: Partitioning Problems
Bucket 0 Bucket 1
Links 0 Links 0 Links 0 Links 0
A B C D
Figure 10.15: Growing a Two-List Hash Table, State (a)
The key insight behind the first hash-table implementation is that each data element
can have two sets of list pointers, with one set currently being used by RCU readers (as
well as by non-RCU updaters) and the other being used to construct a new resized hash
table. This approach allows lookups, insertions, and deletions to all run concurrently
with a resize operation (as well as with each other).
The resize operation proceeds as shown in Figures 10.15–10.18, with the initial
two-bucket state shown in Figure 10.15 and with time advancing from figure to figure.
The initial state uses the zero-index links to chain the elements into hash buckets. A
Bucket 0 Bucket 1
A B C D
Bucket 0 Bucket 1 Bucket 2 Bucket 3
Figure 10.16: Growing a Two-List Hash Table, State (b)
v2023.06.11a
Bucket 0 Bucket 1
A B C D
Figure 10.17: Growing a Two-List Hash Table, State (c)
A B C D
Figure 10.18: Growing a Two-List Hash Table, State (d)
four-bucket array is allocated, and the one-index links are used to chain the elements
into these four new hash buckets. This results in state (b) shown in Figure 10.16, with
readers still using the original two-bucket array.
The new four-bucket array is exposed to readers and then a grace-period operation
waits for all readers, resulting in state (c), shown in Figure 10.17. In this state, all
readers are using the new four-bucket array, which means that the old two-bucket array
may now be freed, resulting in state (d), shown in Figure 10.18.
This design leads to a relatively straightforward implementation, which is the subject
of the next section.
10.4.2 Resizable Hash Table Implementation

Resizing is accomplished by the classic approach of inserting a level of indirection,
in this case, the ht structure shown on lines 11–20 of Listing 10.9 (hash_resize.c).
The hashtab structure shown on lines 27–30 contains only a pointer to the current
ht structure along with a spinlock that is used to serialize concurrent attempts to
resize the hash table. If we were to use a traditional lock- or atomic-operation-based
implementation, this hashtab structure could become a severe bottleneck from both
performance and scalability viewpoints. However, because resize operations should be
relatively infrequent, we should be able to make good use of RCU.
The ht structure represents a specific size of the hash table, as specified by the
->ht_nbuckets field on line 12. The size is stored in the same structure containing
the array of buckets (->ht_bkt[] on line 19) in order to avoid mismatches between
the size and the array. The ->ht_resize_cur field on line 13 is equal to −1 unless
a resize operation is in progress, in which case it indicates the index of the bucket
whose elements are being inserted into the new hash table, which is referenced by the
v2023.06.11a
Listing 10.9: Resizable Hash-Table Data Structures

1 struct ht_elem {
2 struct rcu_head rh;
3 struct cds_list_head hte_next[2];
4 };
5
6 struct ht_bucket {
7 struct cds_list_head htb_head;
8 spinlock_t htb_lock;
9 };
10
11 struct ht {
12 long ht_nbuckets;
13 long ht_resize_cur;
14 struct ht *ht_new;
15 int ht_idx;
16 int (*ht_cmp)(struct ht_elem *htep, void *key);
17 unsigned long (*ht_gethash)(void *key);
18 void *(*ht_getkey)(struct ht_elem *htep);
19 struct ht_bucket ht_bkt[0];
20 };
21
22 struct ht_lock_state {
23 struct ht_bucket *hbp[2];
24 int hls_idx[2];
25 };
26
27 struct hashtab {
28 struct ht *ht_cur;
29 spinlock_t ht_lock;
30 };
->ht_new field on line 14. If there is no resize operation in progress, ->ht_new is NULL.
Thus, a resize operation proceeds by allocating a new ht structure and referencing it
via the ->ht_new pointer, then advancing ->ht_resize_cur through the old table’s
buckets. When all the elements have been added to the new table, the new table is linked
into the hashtab structure’s ->ht_cur field. Once all old readers have completed, the
old hash table’s ht structure may be freed.
The ->ht_idx field on line 15 indicates which of the two sets of list pointers are
being used by this instantiation of the hash table, and is used to index the ->hte_next[]
array in the ht_elem structure on line 3.
The ->ht_cmp(), ->ht_gethash(), and ->ht_getkey() fields on lines 16–18
collectively define the per-element key and the hash function. The ->ht_cmp() function
compares a specified key with that of the specified element, the ->ht_gethash()
calculates the specified key’s hash, and ->ht_getkey() extracts the key from the
enclosing data element.
The ht_lock_state shown on lines 22–25 is used to communicate lock state from
a new hashtab_lock_mod() to hashtab_add(), hashtab_del(), and hashtab_
unlock_mod(). This state prevents the algorithm from being redirected to the wrong
bucket during concurrent resize operations.
The ht_bucket structure is the same as before, and the ht_elem structure differs
from that of previous implementations only in providing a two-element array of list
pointer sets in place of the prior single set of list pointers.
In a fixed-sized hash table, bucket selection is quite straightforward: Simply transform
the hash value to the corresponding bucket index. In contrast, when resizing, it is also
necessary to determine which of the old and new sets of buckets to select from. If the
bucket that would be selected from the old table has already been distributed into the
new table, then the bucket should be selected from the new table as well as from the old
v2023.06.11a
Listing 10.10: Resizable Hash-Table Bucket Selection

1 static struct ht_bucket *
2 ht_get_bucket(struct ht *htp, void *key,
3 long *b, unsigned long *h)
4 {
5 unsigned long hash = htp->ht_gethash(key);
6
7 *b = hash % htp->ht_nbuckets;
8 if (h)
9 *h = hash;
10 return &htp->ht_bkt[*b];
11 }
12
13 static struct ht_elem *
14 ht_search_bucket(struct ht *htp, void *key)
15 {
16 long b;
18 struct ht_bucket *htbp;
19
20 htbp = ht_get_bucket(htp, key, &b, NULL);
21 cds_list_for_each_entry_rcu(htep,
22 &htbp->htb_head,
23 hte_next[htp->ht_idx]) {
25 return htep;
26 }
27 return NULL;
28 }
table. Conversely, if the bucket that would be selected from the old table has not yet
been distributed, then the bucket should be selected from the old table.
Bucket selection is shown in Listing 10.10, which shows ht_get_bucket() on
lines 1–11 and ht_search_bucket() on lines 13–28. The ht_get_bucket() func-
tion returns a reference to the bucket corresponding to the specified key in the specified
hash table, without making any allowances for resizing. It also stores the bucket index
corresponding to the key into the location referenced by parameter b on line 7, and the
corresponding hash value corresponding to the key into the location referenced by pa-
rameter h (if non-NULL) on line 9. Line 10 then returns a reference to the corresponding
bucket.
The ht_search_bucket() function searches for the specified key within the speci-
fied hash-table version. Line 20 obtains a reference to the bucket corresponding to the
specified key. The loop spanning lines 21–26 searches that bucket, so that if line 24
detects a match, line 25 returns a pointer to the enclosing data element. Otherwise, if
there is no match, line 27 returns NULL to indicate failure.
Quick Quiz 10.10: How does the code in Listing 10.10 protect against the resizing process
progressing past the selected bucket?
This implementation of ht_get_bucket() and ht_search_bucket() permits

lookups and modifications to run concurrently with a resize operation.
Read-side concurrency control is provided by RCU as was shown in Listing 10.6, but
the update-side concurrency-control functions hashtab_lock_mod() and hashtab_
unlock_mod() must now deal with the possibility of a concurrent resize operation as
shown in Listing 10.11.
The hashtab_lock_mod() spans lines 1–25 in the listing. Line 10 enters an RCU
read-side critical section to prevent the data structures from being freed during the
traversal, line 11 acquires a reference to the current hash table, and then line 12 obtains a
reference to the bucket in this hash table corresponding to the key. Line 13 acquires that
v2023.06.11a
Listing 10.11: Resizable Hash-Table Update-Side Concurrency Control

1 static void
2 hashtab_lock_mod(struct hashtab *htp_master, void *key,
3 struct ht_lock_state *lsp)
4 {
5 long b;
6 unsigned long h;
7 struct ht *htp;
9
10 rcu_read_lock();
11 htp = rcu_dereference(htp_master->ht_cur);
12 htbp = ht_get_bucket(htp, key, &b, &h);
13 spin_lock(&htbp->htb_lock);
14 lsp->hbp[0] = htbp;
15 lsp->hls_idx[0] = htp->ht_idx;
16 if (b > READ_ONCE(htp->ht_resize_cur)) {
17 lsp->hbp[1] = NULL;
18 return;
19 }
20 htp = rcu_dereference(htp->ht_new);
25 }
26
27 static void
28 hashtab_unlock_mod(struct ht_lock_state *lsp)
29 {
30 spin_unlock(&lsp->hbp[0]->htb_lock);
31 if (lsp->hbp[1])
32 spin_unlock(&lsp->hbp[1]->htb_lock);
34 }
bucket’s lock, which will prevent any concurrent resizing operation from distributing
that bucket, though of course it will have no effect if that bucket has already been
distributed. Lines 14–15 store the bucket pointer and pointer-set index into their
respective fields in the ht_lock_state structure, which communicates the information
to hashtab_add(), hashtab_del(), and hashtab_unlock_mod(). Line 16 then
checks to see if a concurrent resize operation has already distributed this bucket across
the new hash table, and if not, line 17 indicates that there is no already-resized hash
bucket and line 18 returns with the selected hash bucket’s lock held (thus preventing
a concurrent resize operation from distributing this bucket) and also within an RCU
read-side critical section. Deadlock is avoided because the old table’s locks are always
acquired before those of the new table, and because the use of RCU prevents more than
two versions from existing at a given time, thus preventing a deadlock cycle.
Otherwise, a concurrent resize operation has already distributed this bucket, so
line 20 proceeds to the new hash table, line 21 selects the bucket corresponding to
the key, and line 22 acquires the bucket’s lock. Lines 23–24 store the bucket pointer
and pointer-set index into their respective fields in the ht_lock_state structure,
which again communicates this information to hashtab_add(), hashtab_del(), and
hashtab_unlock_mod(). Because this bucket has already been resized and because
hashtab_add() and hashtab_del() affect both the old and the new ht_bucket
structures, two locks are held, one on each of the two buckets. Additionally, both
elements of each array in ht_lock_state structure are used, with the [0] element
pertaining to the old ht_bucket structure and the [1] element pertaining to the new
structure. Once again, hashtab_lock_mod() exits within an RCU read-side critical
section.
v2023.06.11a
Listing 10.12: Resizable Hash-Table Access Functions

1 struct ht_elem *
2 hashtab_lookup(struct hashtab *htp_master, void *key)
3 {
4 struct ht *htp;
6
8 htep = ht_search_bucket(htp, key);
9 return htep;
10 }
11
12 void hashtab_add(struct ht_elem *htep,
14 {
15 struct ht_bucket *htbp = lsp->hbp[0];
16 int i = lsp->hls_idx[0];
17
18 cds_list_add_rcu(&htep->hte_next[i], &htbp->htb_head);
19 if ((htbp = lsp->hbp[1])) {
20 cds_list_add_rcu(&htep->hte_next[!i], &htbp->htb_head);
21 }
22 }
23
24 void hashtab_del(struct ht_elem *htep,
26 {
28
29 cds_list_del_rcu(&htep->hte_next[i]);
30 if (lsp->hbp[1])
31 cds_list_del_rcu(&htep->hte_next[!i]);
32 }
The hashtab_unlock_mod() function releases the lock(s) acquired by hashtab_

lock_mod(). Line 30 releases the lock on the old ht_bucket structure. In the unlikely
event that line 31 determines that a resize operation is in progress, line 32 releases the
lock on the new ht_bucket structure. Either way, line 33 exits the RCU read-side
critical section.
Quick Quiz 10.11: Suppose that one thread is inserting an element into the hash table during
a resize operation. What prevents this insertion from being lost due to a subsequent resize
operation completing before the insertion does?
Now that we have bucket selection and concurrency control in place, we are ready to
search and update our resizable hash table. The hashtab_lookup(), hashtab_add(),
and hashtab_del() functions are shown in Listing 10.12.
The hashtab_lookup() function on lines 1–10 of the listing does hash lookups.
Line 7 fetches the current hash table and line 8 searches the bucket corresponding to the
specified key. Line 9 returns a pointer to the searched-for element or NULL when the
search fails. The caller must be within an RCU read-side critical section.
Quick Quiz 10.12: The hashtab_lookup() function in Listing 10.12 ignores concurrent
resize operations. Doesn’t this mean that readers might miss an element that was previously
added during a resize operation?
The hashtab_add() function on lines 12–22 of the listing adds new data elements
to the hash table. Line 15 picks up the current ht_bucket structure into which the
new element is to be added, and line 16 picks up the index of the pointer pair. Line 18
adds the new element to the current hash bucket. If line 19 determines that this bucket
has been distributed to a new version of the hash table, then line 20 also adds the new
element to the corresponding new bucket. The caller is required to handle concurrency,
v2023.06.11a
Listing 10.13: Resizable Hash-Table Resizing

1 int hashtab_resize(struct hashtab *htp_master,
2 unsigned long nbuckets,
3 int (*cmp)(struct ht_elem *htep, void *key),
4 unsigned long (*gethash)(void *key),
5 void *(*getkey)(struct ht_elem *htep))
6 {
7 struct ht *htp;
8 struct ht *htp_new;
9 int i;
10 int idx;
13 struct ht_bucket *htbp_new;
14 long b;
15
16 if (!spin_trylock(&htp_master->ht_lock))
17 return -EBUSY;
18 htp = htp_master->ht_cur;
19 htp_new = ht_alloc(nbuckets,
20 cmp ? cmp : htp->ht_cmp,
21 gethash ? gethash : htp->ht_gethash,
22 getkey ? getkey : htp->ht_getkey);
23 if (htp_new == NULL) {
24 spin_unlock(&htp_master->ht_lock);
25 return -ENOMEM;
26 }
27 idx = htp->ht_idx;
28 htp_new->ht_idx = !idx;
29 rcu_assign_pointer(htp->ht_new, htp_new);
31 for (i = 0; i < htp->ht_nbuckets; i++) {
32 htbp = &htp->ht_bkt[i];
34 cds_list_for_each_entry(htep, &htbp->htb_head, hte_next[idx]) {
35 htbp_new = ht_get_bucket(htp_new, htp_new->ht_getkey(htep), &b, NULL);
36 spin_lock(&htbp_new->htb_lock);
37 cds_list_add_rcu(&htep->hte_next[!idx], &htbp_new->htb_head);
38 spin_unlock(&htbp_new->htb_lock);
39 }
40 WRITE_ONCE(htp->ht_resize_cur, i);
41 spin_unlock(&htbp->htb_lock);
42 }
43 rcu_assign_pointer(htp_master->ht_cur, htp_new);
45 spin_unlock(&htp_master->ht_lock);
46 free(htp);
47 return 0;
48 }
for example, by invoking hashtab_lock_mod() before the call to hashtab_add()

and invoking hashtab_unlock_mod() afterwards.
The hashtab_del() function on lines 24–32 of the listing removes an existing
element from the hash table. Line 27 picks up the index of the pointer pair and line 29
removes the specified element from the current table. If line 30 determines that this
bucket has been distributed to a new version of the hash table, then line 31 also removes
the specified element from the corresponding new bucket. As with hashtab_add(),
the caller is responsible for concurrency control and this concurrency control suffices
for synchronizing with a concurrent resize operation.
Quick Quiz 10.13: The hashtab_add() and hashtab_del() functions in Listing 10.12
can update two hash buckets while a resize operation is progressing. This might cause poor
performance if the frequency of resize operation is not negligible. Isn’t it possible to reduce the
cost of updates in such cases?
v2023.06.11a
The actual resizing itself is carried out by hashtab_resize, shown in Listing 10.13
on page 311. Line 16 conditionally acquires the top-level ->ht_lock, and if this
acquisition fails, line 17 returns -EBUSY to indicate that a resize is already in progress.
Otherwise, line 18 picks up a reference to the current hash table, and lines 19–22
allocate a new hash table of the desired size. If a new set of hash/key functions have
been specified, these are used for the new table, otherwise those of the old table are
preserved. If line 23 detects memory-allocation failure, line 24 releases ->ht_lock
and line 25 returns a failure indication.
Line 27 picks up the current table’s index and line 28 stores its inverse to the new
hash table, thus ensuring that the two hash tables avoid overwriting each other’s linked
lists. Line 29 then starts the bucket-distribution process by installing a reference to the
new table into the ->ht_new field of the old table. Line 30 ensures that all readers who
are not aware of the new table complete before the resize operation continues.
Each pass through the loop spanning lines 31–42 distributes the contents of one of
the old hash table’s buckets into the new hash table. Line 32 picks up a reference to the
old table’s current bucket and line 33 acquires that bucket’s spinlock.
Quick Quiz 10.14: In the hashtab_resize() function in Listing 10.13, what guarantees
that the update to ->ht_new on line 29 will be seen as happening before the update to ->ht_
resize_cur on line 40 from the perspective of hashtab_add() and hashtab_del()? In
other words, what prevents hashtab_add() and hashtab_del() from dereferencing a NULL
pointer loaded from ->ht_new?
Each pass through the loop spanning lines 34–39 adds one data element from the
current old-table bucket to the corresponding new-table bucket, holding the new-table
bucket’s lock during the add operation. Line 40 updates ->ht_resize_cur to indicate
that this bucket has been distributed. Finally, line 41 releases the old-table bucket lock.
Execution reaches line 43 once all old-table buckets have been distributed across
the new table. Line 43 installs the newly created table as the current one, and line 44
waits for all old readers (who might still be referencing the old table) to complete. Then
line 45 releases the resize-serialization lock, line 46 frees the old hash table, and finally
line 47 returns success.
Quick Quiz 10.15: Why is there a WRITE_ONCE() on line 40 in Listing 10.13?
10.4.3 Resizable Hash Table Discussion

Figure 10.19 compares resizing hash tables to their fixed-sized counterparts for 262,144
and 2,097,152 elements in the hash table. The figure shows three traces for each
element count, one for a fixed-size 262,144-bucket hash table, another for a fixed-size
524,288-bucket hash table, and a third for a resizable hash table that shifts back and
forth between 262,144 and 524,288 buckets, with a one-millisecond pause between each
resize operation.
The uppermost three traces are for the 262,144-element hash table.1 The dashed trace
corresponds to the two fixed-size hash tables, and the solid trace to the resizable hash
table. In this case, the short hash chains cause normal lookup overhead to be so low
that the overhead of resizing dominates over most of the range. In particular, the entire
hash table fits into L3 cache.
1 You see only two traces? The dashed one is composed of two traces that differ only
slightly, hence the irregular-looking dash pattern.
v2023.06.11a
1x107

1x106
262,144
100000
10000 2,097,152
1000
1 10 100
Figure 10.19: Overhead of Resizing Hash Tables Between 262,144 and 524,288 Buckets
vs. Total Number of Elements
The lower three traces are for the 2,097,152-element hash table. The upper dashed
trace corresponds to the 262,144-bucket fixed-size hash table, the solid trace in the
middle for low CPU counts and at the bottom for high CPU counts to the resizable
hash table, and the other trace to the 524,288-bucket fixed-size hash table. The fact that
there are now an average of eight elements per bucket can only be expected to produce
a sharp decrease in performance, as in fact is shown in the graph. But worse yet, the
hash-table elements occupy 128 MB, which overflows each socket’s 39 MB L3 cache,
with performance consequences analogous to those described in Section 3.2.2. The
resulting cache overflow means that the memory system is involved even for a read-only
benchmark, and as you can see from the sublinear portions of the lower three traces, the
memory system can be a serious bottleneck.
Quick Quiz 10.16: How much of the difference in performance between the large and small
hash tables shown in Figure 10.19 was due to long hash chains and how much was due to
memory-system bottlenecks?
Referring to the last column of Table 3.1, we recall that the first 28 CPUs are in
the first socket, on a one-CPU-per-core basis, which explains the sharp decrease in
performance of the resizable hash table beyond 28 CPUs. Sharp though this decrease
is, please recall that it is due to constant resizing back and forth. It would clearly be
better to resize once to 524,288 buckets, or, even better, do a single eight-fold resize to
2,097,152 elements, thus dropping the average number of elements per bucket down to
the level enjoyed by the runs producing the upper three traces.
The key point from this data is that the RCU-protected resizable hash table performs
and scales almost as well as does its fixed-size counterpart. The performance during
an actual resize operation of course suffers somewhat due to the cache misses causes
by the updates to each element’s pointers, and this effect is most pronounced when the
memory system becomes a bottleneck. This indicates that hash tables should be resized
by substantial amounts, and that hysteresis should be applied to prevent performance
degradation due to too-frequent resize operations. In memory-rich environments,
hash-table sizes should furthermore be increased much more aggressively than they are
decreased.
v2023.06.11a
Another key point is that although the hashtab structure is non-partitionable, it is

also read-mostly, which suggests the use of RCU. Given that the performance and
scalability of this resizable hash table is very nearly that of RCU-protected fixed-sized
hash tables, we must conclude that this approach was quite successful.
Finally, it is important to note that insertions, deletions, and lookups can proceed
concurrently with a resize operation. This concurrency is critically important when
resizing large hash tables, especially for applications that must meet severe response-time
constraints.
Of course, the ht_elem structure’s pair of pointer sets does impose some memory
overhead, which is taken up in the next section.
10.4.4 Other Resizable Hash Tables

One shortcoming of the resizable hash table described earlier in this section is memory
consumption. Each data element has two pairs of linked-list pointers rather than just
one. Is it possible to create an RCU-protected resizable hash table that makes do with
just one pair?
It turns out that the answer is “yes”. Josh Triplett et al. [TMW11] produced a
relativistic hash table that incrementally splits and combines corresponding hash chains
so that readers always see valid hash chains at all points during the resizing operation.
This incremental splitting and combining relies on the fact that it is harmless for a reader
to see a data element that should be in some other hash chain: When this happens, the
reader will simply ignore the extraneous data element due to key mismatches.
The process of shrinking a relativistic hash table by a factor of two is shown in
Figure 10.20, in this case shrinking a two-bucket hash table into a one-bucket hash table,
otherwise known as a linear list. This process works by coalescing pairs of buckets
in the old larger hash table into single buckets in the new smaller hash table. For this
process to work correctly, we clearly need to constrain the hash functions for the two
tables. One such constraint is to use the same underlying hash function for both tables,
but to throw out the low-order bit when shrinking from large to small. For example,
the old two-bucket hash table would use the two top bits of the value, while the new
one-bucket hash table could use the top bit of the value. In this way, a given pair of
adjacent even and odd buckets in the old large hash table can be coalesced into a single
bucket in the new small hash table, while still having a single hash value cover all of the
elements in that single bucket.
The initial state is shown at the top of the figure, with time advancing from top to
bottom, starting with initial state (a). The shrinking process begins by allocating the
new smaller array of buckets, and having each bucket of this new smaller array reference
the first element of one of the buckets of the corresponding pair in the old large hash
table, resulting in state (b).
Then the two hash chains are linked together, resulting in state (c). In this state,
readers looking up an even-numbered element see no change, and readers looking up
elements 1 and 3 likewise see no change. However, readers looking up some other odd
number will also traverse elements 0 and 2. This is harmless because any odd number
will compare not-equal to these two elements. There is some performance loss, but on
the other hand, this is exactly the same performance loss that will be experienced once
the new small hash table is fully in place.
Next, the new small hash table is made accessible to readers, resulting in state (d).
Note that older readers might still be traversing the old large hash table, so in this state
both hash tables are in use.
v2023.06.11a
even 0 2
(a)
odd 1 3
even 0 2
(b)
odd 1 3
all
even 0 2
(c)
odd 1 3
all
even 0 2
(d)
odd 1 3
all
even 0 2
(e)
odd 1 3
all
(f) all 1 3 0 2
Figure 10.20: Shrinking a Relativistic Hash Table
The next step is to wait for all pre-existing readers to complete, resulting in state (e).
In this state, all readers are using the new small hash table, so that the old large hash
table’s buckets may be freed, resulting in the final state (f).
Growing a relativistic hash table reverses the shrinking process, but requires more
grace-period steps, as shown in Figure 10.21. The initial state (a) is at the top of this
figure, with time advancing from top to bottom.
We start by allocating the new large two-bucket hash table, resulting in state (b).
Note that each of these new buckets references the first element destined for that bucket.
These new buckets are published to readers, resulting in state (c). After a grace-period
operation, all readers are using the new large hash table, resulting in state (d). In this
state, only those readers traversing the even-values hash bucket traverse element 0,
which is therefore now colored white.
At this point, the old small hash buckets may be freed, although many implementations
use these old buckets to track progress “unzipping” the list of items into their respective
new buckets. The last even-numbered element in the first consecutive run of such
elements now has its pointer-to-next updated to reference the following even-numbered
element. After a subsequent grace-period operation, the result is state (e). The vertical
arrow indicates the next element to be unzipped, and element 1 is now colored black to
indicate that only those readers traversing the odd-values hash bucket may reach it.
v2023.06.11a
(a) all 0 1 2 3
even
(b) odd
all 0 1 2 3
even
(c) odd
all 0 1 2 3
even
(d) odd
all 0 1 2 3
even 0 1 2 3
(e)
odd
even 0 1 2 3
(f)
odd
even 0 2
(g)
odd 1 3
Figure 10.21: Growing a Relativistic Hash Table
v2023.06.11a
10.5. OTHER DATA STRUCTURES 317
Next, the last odd-numbered element in the first consecutive run of such elements
now has its pointer-to-next updated to reference the following odd-numbered element.
After a subsequent grace-period operation, the result is state (f). A final unzipping
operation (including a grace-period operation) results in the final state (g).
In short, the relativistic hash table reduces the number of per-element list pointers
at the expense of additional grace periods incurred during resizing. These additional
grace periods are usually not a problem because insertions, deletions, and lookups may
proceed concurrently with a resize operation.
It turns out that it is possible to reduce the per-element memory overhead from a pair
of pointers to a single pointer, while still retaining O (1) deletions. This is accomplished
by augmenting split-order list [SS06] with RCU protection [Des09b, MDJ13c]. The
data elements in the hash table are arranged into a single sorted linked list, with each
hash bucket referencing the first element in that bucket. Elements are deleted by setting
low-order bits in their pointer-to-next fields, and these elements are removed from the
list by later traversals that encounter them.
This RCU-protected split-order list is complex, but offers lock-free progress guarantees
for all insertion, deletion, and lookup operations. Such guarantees can be important
in real-time applications. An implementation is available from recent versions of the
userspace RCU library [Des09b].
10.5 Other Data Structures
All life is an experiment. The more experiments you

make the better.
Ralph Waldo Emerson
The preceding sections have focused on data structures that enhance concurrency due
to partitionability (Section 10.2), efficient handling of read-mostly access patterns
(Section 10.3), or application of read-mostly techniques to avoid non-partitionability
(Section 10.4). This section gives a brief review of other data structures.
One of the hash table’s greatest advantages for parallel use is that it is fully partitionable,
at least while not being resized. One way of preserving the partitionability and the
size independence is to use a radix tree, which is also called a trie. Tries partition
the search key, using each successive key partition to traverse the next level of the
trie. As such, a trie can be thought of as a set of nested hash tables, thus providing
the required partitionability. One disadvantage of tries is that a sparse key space can
result in inefficient use of memory. There are a number of compression techniques that
may be used to work around this disadvantage, including hashing the key value to a
smaller keyspace before the traversal [ON07]. Radix trees are heavily used in practice,
including in the Linux kernel [Pig06].
One important special case of both a hash table and a trie is what is perhaps the oldest
of data structures, the array and its multi-dimensional counterpart, the matrix. The fully
partitionable nature of matrices is exploited heavily in concurrent numerical algorithms.
Self-balancing trees are heavily used in sequential code, with AVL trees and red-
black trees being perhaps the most well-known examples [CLRS01]. Early attempts
to parallelize AVL trees were complex and not necessarily all that efficient [Ell80],
however, more recent work on red-black trees provides better performance and scalability
v2023.06.11a
by using RCU for readers and hashed arrays of locks2 to protect reads and updates,
respectively [HW11, HW14]. It turns out that red-black trees rebalance aggressively,
which works well for sequential programs, but not necessarily so well for parallel use.
Recent work has therefore made use of RCU-protected “bonsai trees” that rebalance less
aggressively [CKZ12], trading off optimal tree depth to gain more efficient concurrent
updates.
Concurrent skip lists lend themselves well to RCU readers, and in fact represents an
early academic use of a technique resembling RCU [Pug90].
Concurrent double-ended queues were discussed in Section 6.1.2, and concurrent
stacks and queues have a long history [Tre86], though not normally the most impressive
performance or scalability. They are nevertheless a common feature of concurrent
libraries [MDJ13d]. Researchers have recently proposed relaxing the ordering constraints
of stacks and queues [Sha11], with some work indicating that relaxed-ordered queues
actually have better ordering properties than do strict FIFO queues [HKLP12, KLP12,
HHK+ 13].
It seems likely that continued work with concurrent data structures will produce novel
algorithms with surprising properties.
10.6 Micro-Optimization
The devil is in the details.
Unknown
The data structures shown in this chapter were coded straightforwardly, with no adaptation
to the underlying system’s cache hierarchy. In addition, many of the implementations
used pointers to functions for key-to-hash conversions and other frequent operations.
Although this approach provides simplicity and portability, in many cases it does give
up some performance.
The following sections touch on specialization, memory conservation, and hardware
considerations. Please do not mistake these short sections for a definitive treatise on
this subject. Whole books have been written on optimizing to a specific CPU, let alone
to the set of CPU families in common use today.
10.6.1 Specialization
The resizable hash table presented in Section 10.4 used an opaque type for the key.
This allows great flexibility, permitting any sort of key to be used, but it also incurs
significant overhead due to the calls via of pointers to functions. Now, modern hardware
uses sophisticated branch-prediction techniques to minimize this overhead, but on the
other hand, real-world software is often larger than can be accommodated even by
today’s large hardware branch-prediction tables. This is especially the case for calls via
pointers, in which case the branch prediction hardware must record a pointer in addition
to branch-taken/branch-not-taken information.
This overhead can be eliminated by specializing a hash-table implementation to a given
key type and hash function, for example, by using C++ templates. Doing so eliminates
the ->ht_cmp(), ->ht_gethash(), and ->ht_getkey() function pointers in the ht
2 In the guise of swissTM [DFGG11], which is a variant of software transactional
memory in which the developer flags non-shared accesses.
v2023.06.11a
10.6. MICRO-OPTIMIZATION 319
structure shown in Listing 10.9 on page 307. It also eliminates the corresponding calls
through these pointers, which could allow the compiler to inline the resulting fixed
functions, eliminating not only the overhead of the call instruction, but the argument
marshalling as well.
Quick Quiz 10.17: How much do these specializations really save? Are they really worth it?
All that aside, one of the great benefits of modern hardware compared to that available
when I first started learning to program back in the early 1970s is that much less
specialization is required. This allows much greater productivity than was possible back
in the days of four-kilobyte address spaces.
10.6.2 Bits and Bytes

The hash tables discussed in this chapter made almost no attempt to conserve memory.
For example, the ->ht_idx field in the ht structure in Listing 10.9 on page 307 always
has a value of either zero or one, yet takes up a full 32 bits of memory. It could be
eliminated, for example, by stealing a bit from the ->ht_resize_key field. This works
because the ->ht_resize_key field is large enough to address every byte of memory
and the ht_bucket structure is more than one byte long, so that the ->ht_resize_key
field must have several bits to spare.
This sort of bit-packing trick is frequently used in data structures that are highly
replicated, as is the page structure in the Linux kernel. However, the resizable hash
table’s ht structure is not all that highly replicated. It is instead the ht_bucket
structures we should focus on. There are two major opportunities for shrinking the
ht_bucket structure: (1) Placing the ->htb_lock field in a low-order bit of one of
the ->htb_head pointers and (2) Reducing the number of pointers required.
The first opportunity might make use of bit-spinlocks in the Linux kernel, which
are provided by the include/linux/bit_spinlock.h header file. These are used in
space-critical data structures in the Linux kernel, but are not without their disadvantages:
1. They are significantly slower than the traditional spinlock primitives.
2. They cannot participate in the lockdep deadlock detection tooling in the Linux
kernel [Cor06a].
3. They do not record lock ownership, further complicating debugging.
4. They do not participate in priority boosting in -rt kernels, which means that
preemption must be disabled when holding bit spinlocks, which can degrade
real-time latency.
Despite these disadvantages, bit-spinlocks are extremely useful when memory is at a

premium.
One aspect of the second opportunity was covered in Section 10.4.4, which presented
resizable hash tables that require only one set of bucket-list pointers in place of the pair
of sets required by the resizable hash table presented in Section 10.4. Another approach
would be to use singly linked bucket lists in place of the doubly linked lists used in this
chapter. One downside of this approach is that deletion would then require additional
overhead, either by marking the outgoing pointer for later removal or by searching the
bucket list for the element being deleted.
v2023.06.11a
Listing 10.14: Alignment for 64-Byte Cache Lines

1 struct hash_elem {
2 struct ht_elem e;
3 long __attribute__ ((aligned(64))) counter;
4 };
In short, there is a tradeoff between minimal memory overhead on the one hand, and
performance and simplicity on the other. Fortunately, the relatively large memories
available on modern systems have allowed us to prioritize performance and simplicity
over memory overhead. However, even though the year 2022’s pocket-sized smartphones
sport many gigabytes of memory and its mid-range servers sport terabytes, it is sometimes
necessary to take extreme measures to reduce memory overhead.
10.6.3 Hardware Considerations

Modern computers typically move data between CPUs and main memory in fixed-sized
blocks that range in size from 32 bytes to 256 bytes. These blocks are called cache
lines, and are extremely important to high performance and scalability, as was discussed
in Section 3.2. One timeworn way to kill both performance and scalability is to place
incompatible variables into the same cacheline. For example, suppose that a resizable
hash table data element had the ht_elem structure in the same cacheline as a frequently
incremented counter. The frequent incrementing would cause the cacheline to be present
at the CPU doing the incrementing, but nowhere else. If other CPUs attempted to
traverse the hash bucket list containing that element, they would incur expensive cache
misses, degrading both performance and scalability.
One way to solve this problem on systems with 64-byte cache line is shown in
Listing 10.14. Here GCC’s aligned attribute is used to force the ->counter and the
ht_elem structure into separate cache lines. This would allow CPUs to traverse the
hash bucket list at full speed despite the frequent incrementing.
Of course, this raises the question “How did we know that cache lines are 64
bytes in size?” On a Linux system, this information may be obtained from the
/sys/devices/system/cpu/cpu*/cache/ directories, and it is even possible to
make the installation process rebuild the application to accommodate the system’s
hardware structure. However, this would be more difficult if you wanted your application
to also run on non-Linux systems. Furthermore, even if you were content to run only
on Linux, such a self-modifying installation poses validation challenges. For example,
systems with 32-byte cachelines might work well, but performance might suffer on
systems with 64-byte cachelines due to false sharing.
Fortunately, there are some rules of thumb that work reasonably well in practice,
which were gathered into a 1995 paper [GKPS95].3 The first group of rules involve
rearranging structures to accommodate cache geometry:
1. Place read-mostly data far from frequently updated data. For example, place
read-mostly data at the beginning of the structure and frequently updated data at
the end. Place data that is rarely accessed in between.
2. If the structure has groups of fields such that each group is updated by an independent
code path, separate these groups from each other. Again, it can be helpful to place
3 A number of these rules are paraphrased and expanded on here with permission from
Orran Krieger.
v2023.06.11a
10.7. SUMMARY 321
rarely accessed data between the groups. In some cases, it might also make sense to
place each such group into a separate structure referenced by the original structure.
3. Where possible, associate update-mostly data with a CPU, thread, or task. We saw
several very effective examples of this rule of thumb in the counter implementations
in Chapter 5.
4. Going one step further, partition your data on a per-CPU, per-thread, or per-task
basis, as was discussed in Chapter 8.
There has been some work towards automated trace-based rearrangement of structure
fields [GDZE10]. This work might well ease one of the more painstaking tasks required
to get excellent performance and scalability from multithreaded software.
An additional set of rules of thumb deal with locks:
1. Given a heavily contended lock protecting data that is frequently modified, take
one of the following approaches:
(a) Place the lock in a different cacheline than the data that it protects.
(b) Use a lock that is adapted for high contention, such as a queued lock.
(c) Redesign to reduce lock contention. (This approach is best, but is not always
trivial.)
2. Place uncontended locks into the same cache line as the data that they protect. This
approach means that the cache miss that brings the lock to the current CPU also
brings its data.
3. Protect read-mostly data with hazard pointers, RCU, or, for long-duration critical
sections, reader-writer locks.
Of course, these are rules of thumb rather than absolute rules. Some experimentation
is required to work out which are most applicable to a given situation.
10.7 Summary
There’s only one thing more painful than learning

from experience, and that is not learning from
experience.
Archibald MacLeish
This chapter has focused primarily on hash tables, including resizable hash tables, which
are not fully partitionable. Section 10.5 gave a quick overview of a few non-hash-table
data structures. Nevertheless, this exposition of hash tables is an excellent introduction
to the many issues surrounding high-performance scalable data access, including:
1. Fully partitioned data structures work well on small systems, for example, single-
socket systems.
2. Larger systems require locality of reference as well as full partitioning.
v2023.06.11a
3. Read-mostly techniques, such as hazard pointers and RCU, provide good locality
of reference for read-mostly workloads, and thus provide excellent performance
and scalability even on larger systems.
4. Read-mostly techniques also work well on some types of non-partitionable data
structures, such as resizable hash tables.
5. Large data structures can overflow CPU caches, reducing performance and scala-
bility.
6. Additional performance and scalability can be obtained by specializing the data
structure to a specific workload, for example, by replacing a general key with a
32-bit integer.
7. Although requirements for portability and for extreme performance often conflict,
there are some data-structure-layout techniques that can strike a good balance
between these two sets of requirements.
That said, performance and scalability are of little use without reliability, so the next
chapter covers validation.
v2023.06.11a
If it is not tested, it doesn’t work.
Unknown
Chapter 11
Validation
I have had a few parallel programs work the first time, but that is only because I have
written an extremely large number parallel programs over the past few decades. And I
have had far more parallel programs that fooled me into thinking that they were working
correctly the first time than actually were working the first time.
I thus need to validate my parallel programs. The basic trick behind validation, is to
realize that the computer knows what is wrong. It is therefore your job to force it to tell
you. This chapter can therefore be thought of as a short course in machine interrogation.
But you can leave the good-cop/bad-cop routine at home. This chapter covers much
more sophisticated and effective methods, especially given that most computers couldn’t
tell a good cop from a bad cop, at least as far as we know.
A longer course may be found in many recent books on validation, as well as at least
one older but valuable one [Mye79]. Validation is an extremely important topic that
cuts across all forms of software, and is worth intensive study in its own right. However,
this book is primarily about concurrency, so this chapter will do little more than scratch
the surface of this critically important topic.
Section 11.1 introduces the philosophy of debugging. Section 11.2 discusses tracing,
Section 11.3 discusses assertions, and Section 11.4 discusses static analysis. Section 11.5
describes some unconventional approaches to code review that can be helpful when the
fabled 10,000 eyes happen not to be looking at your code. Section 11.6 overviews the
use of probability for validating parallel software. Because performance and scalability
are first-class requirements for parallel programming, Section 11.7 covers these topics.
Finally, Section 11.8 gives a fanciful summary and a short list of statistical traps to
avoid.
But never forget that the three best debugging tools are a thorough understanding of
the requirements, a solid design, and a good night’s sleep!
323
v2023.06.11a
324 CHAPTER 11. VALIDATION
11.1 Introduction
Debugging is like being the detective in a crime
movie where you are also the murderer.
Filipe Fortes
Section 11.1.1 discusses the sources of bugs, and Section 11.1.2 overviews the mindset
required when validating software. Section 11.1.3 discusses when you should start
validation, and Section 11.1.4 describes the surprisingly effective open-source regimen
of code review and community testing.
11.1.1 Where Do Bugs Come From?

Bugs come from developers. The basic problem is that the human brain did not evolve
with computer software in mind. Instead, the human brain evolved in concert with
other human brains and with animal brains. Because of this history, the following three
characteristics of computers often come as a shock to human intuition:
1. Computers lack common sense, despite huge sacrifices at the altar of artificial
intelligence.
2. Computers fail to understand user intent, or more formally, computers generally

lack a theory of mind.
3. Computers cannot do anything useful with a fragmentary plan, instead requiring

that every detail of all possible scenarios be spelled out in full.
The first two points should be uncontroversial, as they are illustrated by any number
of failed products, perhaps most famously Clippy and Microsoft Bob. By attempting to
relate to users as people, these two products raised common-sense and theory-of-mind
expectations that they proved incapable of meeting. Perhaps the set of software assistants
are now available on smartphones will fare better, but as of 2021 reviews are mixed.
That said, the developers working on them by all accounts still develop the old way:
The assistants might well benefit end users, but not so much their own developers.
This human love of fragmentary plans deserves more explanation, especially given
that it is a classic two-edged sword. This love of fragmentary plans is apparently due
to the assumption that the person carrying out the plan will have (1) common sense
and (2) a good understanding of the intent and requirements driving the plan. This
latter assumption is especially likely to hold in the common case where the person
doing the planning and the person carrying out the plan are one and the same: In this
case, the plan will be revised almost subconsciously as obstacles arise, especially when
that person has the a good understanding of the problem at hand. In fact, the love of
fragmentary plans has served human beings well, in part because it is better to take
random actions that have a some chance of locating food than to starve to death while
attempting to plan the unplannable. However, the usefulness of fragmentary plans in
the everyday life of which we are all experts is no guarantee of their future usefulness in
stored-program computers.
Furthermore, the need to follow fragmentary plans has had important effects on the
human psyche, due to the fact that throughout much of human history, life was often
difficult and dangerous. It should come as no surprise that executing a fragmentary
v2023.06.11a
11.1. INTRODUCTION 325
plan that has a high probability of a violent encounter with sharp teeth and claws
requires almost insane levels of optimism—a level of optimism that actually is present
in most human beings. These insane levels of optimism extend to self-assessments
of programming ability, as evidenced by the effectiveness of (and the controversy
over) code-interviewing techniques [Bra07]. In fact, the clinical term for a human
being with less-than-insane levels of optimism is “clinically depressed”. Such people
usually have extreme difficulty functioning in their daily lives, underscoring the perhaps
counter-intuitive importance of insane levels of optimism to a normal, healthy life.
Furtheremore, if you are not insanely optimistic, you are less likely to start a difficult
but worthwhile project.1
Quick Quiz 11.1: When in computing is it necessary to follow a fragmentary plan?
An important special case is the project that, while valuable, is not valuable enough
to justify the time required to implement it. This special case is quite common, and
one early symptom is the unwillingness of the decision-makers to invest enough to
actually implement the project. A natural reaction is for the developers to produce an
unrealistically optimistic estimate in order to be permitted to start the project. If the
organization is strong enough and its decision-makers ineffective enough, the project
might succeed despite the resulting schedule slips and budget overruns. However, if
the organization is not strong enough and if the decision-makers fail to cancel the
project as soon as it becomes clear that the estimates are garbage, then the project might
well kill the organization. This might result in another organization picking up the
project and either completing it, canceling it, or being killed by it. A given project
might well succeed only after killing several organizations. One can only hope that
the organization that eventually makes a success of a serial-organization-killer project
maintains a suitable level of humility, lest it be killed by its next such project.
Quick Quiz 11.2: Who cares about the organization? After all, it is the project that is
important!
Important though insane levels of optimism might be, they are a key source of bugs
(and perhaps failure of organizations). The question is therefore “How to maintain the
optimism required to start a large project while at the same time injecting enough reality
to keep the bugs down to a dull roar?” The next section examines this conundrum.
11.1.2 Required Mindset

When carrying out any validation effort, keep the following definitions firmly in mind:
1. The only bug-free programs are trivial programs.

2. A reliable program has no known bugs.
From these definitions, it logically follows that any reliable non-trivial program
contains at least one bug that you do not know about. Therefore, any validation effort
undertaken on a non-trivial program that fails to find any bugs is itself a failure. A good
validation is therefore an exercise in destruction. This means that if you are the type of
person who enjoys breaking things, validation is just job for you.
1 There are some famous exceptions to this rule of thumb. Some people take on difficult
or risky projects in order to at least a temporarily escape from their depression. Others have
nothing to lose: The project is literally a matter of life or death.
v2023.06.11a
Quick Quiz 11.3: Suppose that you are writing a script that processes the output of the time
command, which looks as follows:
real 0m0.132s
user 0m0.040s
sys 0m0.008s
The script is required to check its input for errors, and to give appropriate diagnostics if fed
erroneous time output. What test inputs should you provide to this program to test it for use
with time output generated by single-threaded programs?
But perhaps you are a super-programmer whose code is always perfect the first time
every time. If so, congratulations! Feel free to skip this chapter, but I do hope that you
will forgive my skepticism. You see, I have too many people who claimed to be able to
write perfect code the first time, which is not too surprising given the previous discussion
of optimism and over-confidence. And even if you really are a super-programmer, you
just might find yourself debugging lesser mortals’ work.
One approach for the rest of us is to alternate between our normal state of insane
optimism (Sure, I can program that!) and severe pessimism (It seems to work, but I just
know that there have to be more bugs hiding in there somewhere!). It helps if you enjoy
breaking things. If you don’t, or if your joy in breaking things is limited to breaking
other people’s things, find someone who does love breaking your code and have them
help you break it.
Another helpful frame of mind is to hate it when other people find bugs in your code.
This hatred can help motivate you to torture your code beyond all reason in order to
increase the probability that you will be the one to find the bugs. Just make sure to
suspend this hatred long enough to sincerely thank anyone who does find a bug in your
code! After all, by so doing, they saved you the trouble of tracking it down, and possibly
at great personal expense dredging through your code.
Yet another helpful frame of mind is studied skepticism. You see, believing that you
understand the code means you can learn absolutely nothing about it. Ah, but you know
that you completely understand the code because you wrote or reviewed it? Sorry, but
the presence of bugs suggests that your understanding is at least partially fallacious.
One cure is to write down what you know to be true and double-check this knowledge,
as discussed in Sections 11.2–11.5. Objective reality always overrides whatever you
might think you know.
One final frame of mind is to consider the possibility that someone’s life depends on
your code being correct. One way of looking at this is that consistently making good
things happen requires a lot of focus on a lot of bad things that might happen, with an
eye towards preventing or otherwise handling those bad things.2 The prospect of these
bad things might also motivate you to torture your code into revealing the whereabouts
of its bugs.
This wide variety of frames of mind opens the door to the possibility of multiple
people with different frames of mind contributing to the project, with varying levels of
optimism. This can work well, if properly organized.
Some people might see vigorous validation as a form of torture, as depicted in
Figure 11.1.3 Such people might do well to remind themselves that, Tux cartoons aside,
2 For more on this philosophy, see the chapter entitled “The Power of Negative Thinking”
from Chris Hadfield’s excellent book entitled “An Astronaut’s Guide to Life on Earth.”
3 The cynics among us might question whether these people are afraid that validation
will find bugs that they will then be required to fix.
v2023.06.11a
Figure 11.1: Validation and the Geneva Convention
Figure 11.2: Rationalizing Validation
v2023.06.11a
they are really torturing an inanimate object, as shown in Figure 11.2. Rest assured that
those who fail to torture their code are doomed to be tortured by it!
However, this leaves open the question of exactly when during the project lifetime
validation should start, a topic taken up by the next section.
11.1.3 When Should Validation Start?

Validation should start exactly when the project starts.
To see this, consider that tracking down a bug is much harder in a large program than
in a small one. Therefore, to minimize the time and effort required to track down bugs,
you should test small units of code. Although you won’t find all the bugs this way, you
will find a substantial fraction, and it will be much easier to find and fix the ones you
do find. Testing at this level can also alert you to larger flaws in your overall design,
minimizing the time you waste writing code that is broken by design.
But why wait until you have code before validating your design?4 Hopefully reading
Chapters 3 and 4 provided you with the information required to avoid some regrettably
common design flaws, but discussing your design with a colleague or even simply
writing it down can help flush out additional flaws.
However, it is all too often the case that waiting to start validation until you have a
design is waiting too long. Mightn’t your natural level of optimism caused you to start
the design before you fully understood the requirements? The answer to this question
will almost always be “yes”. One good way to avoid flawed requirements is to get to
know your users. To really serve them well, you will have to live among them.
Quick Quiz 11.4: You are asking me to do all this validation BS before I even start coding???
That sounds like a great way to never get started!!!
First-of-a-kind projects often use different methodologies such as rapid prototyping or

agile. Here, the main goal of early prototypes are not to create correct implementations,
but rather to learn the project’s requirements. But this does not mean that you omit
validation; it instead means that you approach it differently.
One such approach takes a Darwinian view, with the validation suite eliminating code
that is not fit to solve the problem at hand. From this viewpoint, a vigorous validation
suite is essential to the fitness of your software. However, taking this approach to its
logical conclusion is quite humbling, as it requires us developers to admit that our
carefully crafted changes to the codebase are, from a Darwinian standpoint, random
mutations. On the other hand, this conclusion is supported by long experience indicating
that seven percent of fixes introduce at least one bug [BJ12].
How vigorous should your validation suite be? If the bugs it finds aren’t threatening
the very foundations of your software design, then it is not yet vigorous enough. After
all, your design is just as prone to bugs as is your code, and the earlier you find and fix
the bugs in your design, the less time you will waste coding those design bugs.
Quick Quiz 11.5: Are you actually suggesting that it is possible to test correctness into
software??? Everyone knows that is impossible!!!
It is worth reiterating that this advice applies to first-of-a-kind projects. If you are
instead doing a project in a well-explored area, you would be quite foolish to refuse
to learn from previous experience. But you should still start validating right at the
4 The old saying “First we must code, then we have incentive to think” notwithstanding.
v2023.06.11a
beginning of the project, but hopefully guided by others’ hard-won knowledge of both
requirements and pitfalls.
An equally important question is “When should validation stop?” The best answer
is “Some time after the last change.” Every change has the potential to create a bug,
and thus every change must be validated. Furthermore, validation development should
continue through the full lifetime of the project. After all, the Darwinian perspective
above implies that bugs are adapting to your validation suite. Therefore, unless you
continually improve your validation suite, your project will naturally accumulate hordes
of validation-suite-immune bugs.
But life is a tradeoff, and every bit of time invested in validation suites as a bit of time
that cannot be invested in directly improving the project itself. These sorts of choices
are never easy, and it can be just as damaging to overinvest in validation as it can be to
underinvest. But this is just one more indication that life is not easy.
Now that we have established that you should start validation when you start the
project (if not earlier!), and that both validation and validation development should
continue throughout the lifetime of that project, the following sections cover a number
of validation techniques and methods that have proven their worth.
11.1.4 The Open Source Way

The open-source programming methodology has proven quite effective, and includes a
regimen of intense code review and testing.
I can personally attest to the effectiveness of the open-source community’s intense
code review. One of my first patches to the Linux kernel involved a distributed filesystem
where one node might write to a given file that another node has mapped into memory.
In this case, it is necessary to invalidate the affected pages from the mapping in order to
allow the filesystem to maintain coherence during the write operation. I coded up a first
attempt at a patch, and, in keeping with the open-source maxim “post early, post often”,
I posted the patch. I then considered how I was going to test it.
But before I could even decide on an overall test strategy, I got a reply to my posting
pointing out a few bugs. I fixed the bugs and reposted the patch, and returned to thinking
out my test strategy. However, before I had a chance to write any test code, I received a
reply to my reposted patch, pointing out more bugs. This process repeated itself many
times, and I am not sure that I ever got a chance to actually test the patch.
This experience brought home the truth of the open-source saying: Given enough
eyeballs, all bugs are shallow [Ray99].
However, when you post some code or a given patch, it is worth asking a few questions:
1. How many of those eyeballs are actually going to look at your code?
2. How many will be experienced and clever enough to actually find your bugs?
3. Exactly when are they going to look?
I was lucky: There was someone out there who wanted the functionality provided by
my patch, who had long experience with distributed filesystems, and who looked at my
patch almost immediately. If no one had looked at my patch, there would have been
no review, and therefore none of those bugs would have been located. If the people
looking at my patch had lacked experience with distributed filesystems, it is unlikely
that they would have found all the bugs. Had they waited months or even years to look, I
v2023.06.11a
likely would have forgotten how the patch was supposed to work, making it much more
difficult to fix them.
However, we must not forget the second tenet of the open-source development, namely
intensive testing. For example, a great many people test the Linux kernel. Some test
patches as they are submitted, perhaps even yours. Others test the -next tree, which is
helpful, but there is likely to be several weeks or even months delay between the time
that you write the patch and the time that it appears in the -next tree, by which time the
patch will not be quite as fresh in your mind. Still others test maintainer trees, which
often have a similar time delay.
Quite a few people don’t test code until it is committed to mainline, or the master
source tree (Linus’s tree in the case of the Linux kernel). If your maintainer won’t
accept your patch until it has been tested, this presents you with a deadlock situation:
Your patch won’t be accepted until it is tested, but it won’t be tested until it is accepted.
Nevertheless, people who test mainline code are still relatively aggressive, given that
many people and organizations do not test code until it has been pulled into a Linux
distro.
And even if someone does test your patch, there is no guarantee that they will be
running the hardware and software configuration and workload required to locate your
bugs.
Therefore, even when writing code for an open-source project, you need to be prepared
to develop and run your own test suite. Test development is an underappreciated and
very valuable skill, so be sure to take full advantage of any existing test suites available
to you. Important as test development is, we must leave further discussion of it to books
dedicated to that topic. The following sections therefore discuss locating bugs in your
code given that you already have a good test suite.
11.2 Tracing
The machine knows what is wrong. Make it tell you.

Unknown
When all else fails, add a printk()! Or a printf(), if you are working with user-mode
C-language applications.
The rationale is simple: If you cannot figure out how execution reached a given point
in the code, sprinkle print statements earlier in the code to work out what happened. You
can get a similar effect, and with more convenience and flexibility, by using a debugger
such as gdb (for user applications) or kgdb (for debugging Linux kernels). Much more
sophisticated tools exist, with some of the more recent offering the ability to rewind
backwards in time from the point of failure.
These brute-force testing tools are all valuable, especially now that typical systems
have more than 64K of memory and CPUs running faster than 4 MHz. Much has been
written about these tools, so this chapter will add only a little more.
However, these tools all have a serious shortcoming when you need a fastpath to tell
you what is going wrong, namely, these tools often have excessive overheads. There are
special tracing technologies for this purpose, which typically leverage data ownership
techniques (see Chapter 8) to minimize the overhead of runtime data collection. One
example within the Linux kernel is “trace events” [Ros10b, Ros10c, Ros10d, Ros10a],
which uses per-CPU buffers to allow data to be collected with extremely low overhead.
v2023.06.11a
11.3. ASSERTIONS 331
Even so, enabling tracing can sometimes change timing enough to hide bugs, resulting
in heisenbugs, which are discussed in Section 11.6 and especially Section 11.6.4. In the
kernel, BPF can do data reduction in the kernel, reducing the overhead of transmitting
the needed information from the kernel to userspace [Gre19]. In userspace code, there
is a huge number of tools that can help you. One good starting point is Brendan Gregg’s
blog.5
Even if you avoid heisenbugs, other pitfalls await you. For example, although the
machine really does know all, what it knows is almost always way more than your head
can hold. For this reason, high-quality test suites normally come with sophisticated
scripts to analyze the voluminous output. But beware—scripts will only notice what you
tell them to. My rcutorture scripts are a case in point: Early versions of those scripts
were quite satisfied with a test run in which RCU grace periods stalled indefinitely. This
of course resulted in the scripts being modified to detect RCU grace-period stalls, but
this does not change the fact that the scripts will only detect problems that I make them
detect. But note well that unless you have a solid design, you won’t know what your
script should check for!
Another problem with tracing and especially with printk() calls is that their
overhead can rule out production use. In such cases, assertions can be helpful.
11.3 Assertions
No man really becomes a fool until he stops asking
questions.
Charles P. Steinmetz
Assertions are usually implemented in the following manner:
1 if (something_bad_is_happening())
2 complain();
This pattern is often encapsulated into C-preprocessor macros or language intrinsics,

for example, in the Linux kernel, this might be represented as WARN_ON(something_
bad_is_happening()). Of course, if something_bad_is_happening() quite
frequently, the resulting output might obscure reports of other problems, in which case
WARN_ON_ONCE(something_bad_is_happening()) might be more appropriate.
Quick Quiz 11.6: How can you implement WARN_ON_ONCE()?
In parallel code, one bad something that might happen is that a function expecting
to be called under a particular lock might be called without that lock being held.
Such functions sometimes have header comments stating something like “The caller
must hold foo_lock when calling this function”, but such a comment does no good
unless someone actually reads it. An executable statement carries far more weight.
The Linux kernel’s lockdep facility [Cor06a, Ros11] therefore provides a lockdep_
assert_held() function that checks whether the specified lock is held. Of course,
lockdep incurs significant overhead, and thus might not be helpful in production.
An especially bad parallel-code something is unexpected concurrent access to data.
The kernel concurrency sanitizer (KCSAN) [Cor16a] uses existing markings such as
5 http://www.brendangregg.com/blog/
v2023.06.11a
READ_ONCE() and WRITE_ONCE() to determine which concurrent accesses deserve

warning messages. KCSAN has a significant false-positive rate, especially from the
viewpoint of developers thinking in terms of C as assembly language with additional
syntax. KCSAN therefore provides a data_race() construct to forgive known-benign
data races, and also the ASSERT_EXCLUSIVE_ACCESS() and ASSERT_EXCLUSIVE_
WRITER() assertions to explicitly check for data races [EMV+ 20a, EMV+ 20b].
So what can be done in cases where checking is necessary, but where the overhead
of runtime checking cannot be tolerated? One approach is static analysis, which is
discussed in the next section.
11.4 Static Analysis
A lot of automation isn’t a replacement of humans

but of mind-numbing behavior.
Summarized from Stewart Butterfield
Static analysis is a validation technique where one program takes a second program as
input, reporting errors and vulnerabilities located in this second program. Interestingly
enough, almost all programs are statically analyzed by their compilers or interpreters.
These tools are far from perfect, but their ability to locate errors has improved immensely
over the past few decades, in part because they now have much more than 64K bytes of
memory in which to carry out their analyses.
The original UNIX lint tool [Joh77] was quite useful, though much of its function-
ality has since been incorporated into C compilers. There are nevertheless lint-like tools
in use to this day. The sparse static analyzer [Cor04b] finds higher-level issues in the
Linux kernel, including:
1. Misuse of pointers to user-space structures.
2. Assignments from too-long constants.
3. Empty switch statements.
4. Mismatched lock acquisition and release primitives.
5. Misuse of per-CPU primitives.
6. Use of RCU primitives on non-RCU pointers and vice versa.
Although it is likely that compilers will continue to increase their static-analysis

capabilities, the sparse static analyzer demonstrates the benefits of static analysis outside
of the compiler, particularly for finding application-specific bugs. Sections 12.4–12.5
describe more sophisticated forms of static analysis.
v2023.06.11a
11.5. CODE REVIEW 333
11.5 Code Review
If a man speaks of my virtues, he steals from me; if

he speaks of my vices, then he is my teacher.
Chinese proverb
Code review is a special case of static analysis with human beings doing the analysis.
This section covers inspection, walkthroughs, and self-inspection.
11.5.1 Inspection
Traditionally, formal code inspections take place in face-to-face meetings with formally
defined roles: Moderator, developer, and one or two other participants. The developer
reads through the code, explaining what it is doing and why it works. The one or two
other participants ask questions and raise issues, hopefully exposing the author’s invalid
assumptions, while the moderator’s job is to resolve any resulting conflicts and take
notes. This process can be extremely effective at locating bugs, particularly if all of the
participants are familiar with the code at hand.
However, this face-to-face formal procedure does not necessarily work well in the
global Linux kernel community. Instead, individuals review code separately and provide
comments via email or IRC. The note-taking is provided by email archives or IRC logs,
and moderators volunteer their services as required by the occasional flamewar. This
process also works reasonably well, particularly if all of the participants are familiar with
the code at hand. In fact, one advantage of the Linux kernel community approach over
traditional formal inspections is the greater probability of contributions from people not
familiar with the code, who might not be blinded by the author’s invalid assumptions,
and who might also test the code.
Quick Quiz 11.7: Just what invalid assumptions are you accusing Linux kernel hackers of
harboring???
It is quite likely that the Linux kernel community’s review process is ripe for
improvement:
1. There is sometimes a shortage of people with the time and expertise required to
carry out an effective review.
2. Even though all review discussions are archived, they are often “lost” in the sense
that insights are forgotten and people fail to look up the discussions. This can
result in re-insertion of the same old bugs.
3. It is sometimes difficult to resolve flamewars when they do break out, especially

when the combatants have disjoint goals, experience, and vocabulary.
Perhaps some of the needed improvements will be provided by continuous-integration-

style testing, but there are many bugs more easily found by review than by testing. When
reviewing, therefore, it is worthwhile to look at relevant documentation in commit logs,
bug reports, and LWN articles. This documentation can help you quickly build up the
required expertise.
v2023.06.11a
11.5.2 Walkthroughs
A traditional code walkthrough is similar to a formal inspection, except that the group
“plays computer” with the code, driven by specific test cases. A typical walkthrough
team has a moderator, a secretary (who records bugs found), a testing expert (who
generates the test cases) and perhaps one to two others. These can be extremely effective,
albeit also extremely time-consuming.
It has been some decades since I have participated in a formal walkthrough, and I
suspect that a present-day walkthrough would use single-stepping debuggers. One could
imagine a particularly sadistic procedure as follows:
1. The tester presents the test case.

2. The moderator starts the code under a debugger, using the specified test case as
input.
3. Before each statement is executed, the developer is required to predict the outcome
of the statement and explain why this outcome is correct.
4. If the outcome differs from that predicted by the developer, this is taken as a
potential bug.
5. In parallel code, a “concurrency shark” asks what code might execute concurrently
with this code, and why such concurrency is harmless.
Sadistic, certainly. Effective? Maybe. If the participants have a good understanding

of the requirements, software tools, data structures, and algorithms, then walkthroughs
can be extremely effective. If not, walkthroughs are often a waste of time.
11.5.3 Self-Inspection
Although developers are usually not all that effective at inspecting their own code,
there are a number of situations where there is no reasonable alternative. For example,
the developer might be the only person authorized to look at the code, other qualified
developers might all be too busy, or the code in question might be sufficiently bizarre
that the developer is unable to convince anyone else to take it seriously until after
demonstrating a prototype. In these cases, the following procedure can be quite helpful,
especially for complex parallel code:
1. Write design document with requirements, diagrams for data structures, and
rationale for design choices.
2. Consult with experts, updating the design document as needed.
3. Write the code in pen on paper, correcting errors as you go. Resist the temptation
to refer to pre-existing nearly identical code sequences, instead, copy them.
4. At each step, articulate and question your assumptions, inserting assertions or
constructing tests to check them.
5. If there were errors, copy the code in pen on fresh paper, correcting errors as you
go. Repeat until the last two copies are identical.
6. Produce proofs of correctness for any non-obvious code.
v2023.06.11a
11.5. CODE REVIEW 335
7. Use a source-code control system. Commit early; commit often.

8. Test the code fragments from the bottom up.
9. When all the code is integrated (but preferably before), do full-up functional and
stress testing.
10. Once the code passes all tests, write code-level documentation, perhaps as an
extension to the design document discussed above. Fix both the code and the test
code as needed.
When I follow this procedure for new RCU code, there are normally only a few
bugs left at the end. With a few prominent (and embarrassing) exceptions [McK11a], I
usually manage to locate these bugs before others do. That said, this is getting more
difficult over time as the number and variety of Linux-kernel users increases.
Quick Quiz 11.8: Why would anyone bother copying existing code in pen on paper??? Doesn’t
that just increase the probability of transcription errors?
Quick Quiz 11.9: This procedure is ridiculously over-engineered! How can you expect to get
a reasonable amount of software written doing it this way???
Quick Quiz 11.10: What do you do if, after all the pen-on-paper copying, you find a bug while
typing in the resulting code?
The above procedure works well for new code, but what if you need to inspect code
that you have already written? You can of course apply the above procedure for old
code in the special case where you wrote one to throw away [FPB79], but the following
approach can also be helpful in less desperate circumstances:
1. Using your favorite documentation tool (LATEX, HTML, OpenOffice, or straight

ASCII), describe the high-level design of the code in question. Use lots of diagrams
to illustrate the data structures and how these structures are updated.
2. Make a copy of the code, stripping away all comments.
3. Document what the code does statement by statement.
4. Fix bugs as you find them.
This works because describing the code in detail is an excellent way to spot
bugs [Mye79]. This second procedure is also a good way to get your head around
someone else’s code, although the first step often suffices.
Although review and inspection by others is probably more efficient and effective,
the above procedures can be quite helpful in cases where for whatever reason it is not
feasible to involve others.
At this point, you might be wondering how to write parallel code without having to
do all this boring paperwork. Here are some time-tested ways of accomplishing this:
1. Write a sequential program that scales through use of available parallel library
functions.
2. Write sequential plug-ins for a parallel framework, such as map-reduce, BOINC,
or a web-application server.
v2023.06.11a
3. Fully partition your problems, then implement sequential program(s) that run in
parallel without communication.
4. Stick to one of the application areas (such as linear algebra) where tools can
automatically decompose and parallelize the problem.
5. Make extremely disciplined use of parallel-programming primitives, so that the

resulting code is easily seen to be correct. But beware: It is always tempting to
break the rules “just a little bit” to gain better performance or scalability. Breaking
the rules often results in general breakage. That is, unless you carefully do the
paperwork described in this section.
But the sad fact is that even if you do the paperwork or use one of the above ways to
more-or-less safely avoid paperwork, there will be bugs. If nothing else, more users and
a greater variety of users will expose more bugs more quickly, especially if those users
are doing things that the original developers did not consider. The next section describes
how to handle the probabilistic bugs that occur all too commonly when validating
parallel software.
Quick Quiz 11.11: Wait! Why on earth would an abstract piece of software fail only
sometimes???
11.6 Probability and Heisenbugs

With both heisenbugs and impressionist art, the
closer you get, the less you see.
Unknown
So your parallel program fails sometimes. But you used techniques from the earlier
sections to locate the problem and now have a fix in place! Congratulations!!!
Now the question is just how much testing is required in order to be certain that you
actually fixed the bug, as opposed to just reducing the probability of it occurring on
the one hand, having fixed only one of several related bugs on the other hand, or made
some ineffectual unrelated change on yet a third hand. In short, what is the answer to
the eternal question posed by Figure 11.3?
Unfortunately, the honest answer is that an infinite amount of testing is required to
attain absolute certainty.
Quick Quiz 11.12: Suppose that you had a very large number of systems at your disposal. For
example, at current cloud prices, you can purchase a huge amount of CPU time at low cost.
Why not use this approach to get close enough to certainty for all practical purposes?
But suppose that we are willing to give up absolute certainty in favor of high
probability. Then we can bring powerful statistical tools to bear on this problem.
However, this section will focus on simple statistical tools. These tools are extremely
helpful, but please note that reading this section is not a substitute for statistics classes.6
For our start with simple statistical tools, we need to decide whether we are doing
discrete or continuous testing. Discrete testing features well-defined individual test runs.
6 Which I most highly recommend. The few statistics courses I have taken have provided
value far beyond that of the time I spent on them.
v2023.06.11a
11.6. PROBABILITY AND HEISENBUGS 337
Hooray! I passed
the stress test!
Ha. You just got lucky
Figure 11.3: Passed on Merits? Or Dumb Luck?
For example, a boot-up test of a Linux kernel patch is an example of a discrete test: The
kernel either comes up or it does not. Although you might spend an hour boot-testing
your kernel, the number of times you attempted to boot the kernel and the number of
times the boot-up succeeded would often be of more interest than the length of time you
spent testing. Functional tests tend to be discrete.
On the other hand, if my patch involved RCU, I would probably run rcutorture,
which is a kernel module that, strangely enough, tests RCU. Unlike booting the kernel,
where the appearance of a login prompt signals the successful end of a discrete test,
rcutorture will happily continue torturing RCU until either the kernel crashes or until
you tell it to stop. The duration of the rcutorture test is usually of more interest than
the number of times you started and stopped it. Therefore, rcutorture is an example
of a continuous test, a category that includes many stress tests.
Statistics for discrete tests are simpler and more familiar than those for continuous
tests, and furthermore the statistics for discrete tests can often be pressed into service
for continuous tests, though with some loss of accuracy. We therefore start with discrete
tests.
11.6.1 Statistics for Discrete Testing

Suppose a bug has a 10 % chance of occurring in a given run and that we do five runs.
How do we compute the probability of at least one run failing? Here is one way:
1. Compute the probability of a given run succeeding, which is 90 %.

2. Compute the probability of all five runs succeeding, which is 0.9 raised to the fifth
power, or about 59 %.
3. Because either all five runs succeed, or at least one fails, subtract the 59 % expected
success rate from 100 %, yielding a 41 % expected failure rate.
For those preferring formulas, call the probability of a single failure 𝑓 . The probability
of a single success is then 1 − 𝑓 and the probability that all of 𝑛 tests will succeed is 𝑆 𝑛 :
v2023.06.11a
𝑆 𝑛 = (1 − 𝑓 ) 𝑛 (11.1)
The probability of failure is 1 − 𝑆 𝑛 , or:
𝐹𝑛 = 1 − (1 − 𝑓 ) 𝑛 (11.2)
Quick Quiz 11.13: Say what??? When I plug the earlier five-test 10 %-failure-rate example
into the formula, I get 59,050 % and that just doesn’t make sense!!!
So suppose that a given test has been failing 10 % of the time. How many times do
you have to run the test to be 99 % sure that your supposed fix actually helped?
Another way to ask this question is “How many times would we need to run the test to
cause the probability of failure to rise above 99 %?” After all, if we were to run the test
enough times that the probability of seeing at least one failure becomes 99 %, if there
are no failures, there is only 1 % probability of this “success” being due to dumb luck.
And if we plug 𝑓 = 0.1 into Eq. 11.2 and vary 𝑛, we find that 43 runs gives us a 98.92 %
chance of at least one test failing given the original 10 % per-test failure rate, while 44
runs gives us a 99.03 % chance of at least one test failing. So if we run the test on our
fix 44 times and see no failures, there is a 99 % probability that our fix really did help.
But repeatedly plugging numbers into Eq. 11.2 can get tedious, so let’s solve for 𝑛:
𝐹𝑛 = 1 − (1 − 𝑓 ) 𝑛 (11.3)
1 − 𝐹𝑛 = (1 − 𝑓 ) 𝑛 (11.4)
log (1 − 𝐹𝑛 ) = 𝑛 log (1 − 𝑓 ) (11.5)
Finally the number of tests required is given by:
log (1 − 𝐹𝑛 )
𝑛= (11.6)
log (1 − 𝑓 )
Plugging 𝑓 = 0.1 and 𝐹𝑛 = 0.99 into Eq. 11.6 gives 43.7, meaning that we need 44
consecutive successful test runs to be 99 % certain that our fix was a real improvement.
This matches the number obtained by the previous method, which is reassuring.
Quick Quiz 11.14: In Eq. 11.6, are the logarithms base-10, base-2, or base-e?
Figure 11.4 shows a plot of this function. Not surprisingly, the less frequently each
test run fails, the more test runs are required to be 99 % confident that the bug has been
fixed. If the bug caused the test to fail only 1 % of the time, then a mind-boggling 458
test runs are required. As the failure probability decreases, the number of test runs
required increases, going to infinity as the failure probability goes to zero.
The moral of this story is that when you have found a rarely occurring bug, your
testing job will be much easier if you can come up with a carefully targeted test with a
much higher failure rate. For example, if your targeted test raised the failure rate from
1 % to 30 %, then the number of runs required for 99 % confidence would drop from
458 to a more tractable 13.
But these thirteen test runs would only give you 99 % confidence that your fix had
produced “some improvement”. Suppose you instead want to have 99 % confidence that
your fix reduced the failure rate by an order of magnitude. How many failure-free test
runs are required?
v2023.06.11a
1000
Number of Runs for 99% Confidence

100
10
1
0 0.2 0.4 0.6 0.8 1
Per-Run Failure Probability
Figure 11.4: Number of Tests Required for 99 Percent Confidence Given Failure Rate
An order of magnitude improvement from a 30 % failure rate would be a 3 % failure

rate. Plugging these numbers into Eq. 11.6 yields:
log (1 − 0.99)
𝑛= = 151.2 (11.7)
log (1 − 0.03)
So our order of magnitude improvement requires roughly an order of magnitude more
testing. Certainty is impossible, and high probabilities are quite expensive. This is
why making tests run more quickly and making failures more probable are essential
skills in the development of highly reliable software. These skills will be covered in
Section 11.6.4.
11.6.2 Statistics Abuse for Discrete Testing

But suppose that you have a continuous test that fails about three times every ten hours,
and that you fix the bug that you believe was causing the failure. How long do you have
to run this test without failure to be 99 % certain that you reduced the probability of
failure?
Without doing excessive violence to statistics, we could simply redefine a one-hour
run to be a discrete test that has a 30 % probability of failure. Then the results of in the
previous section tell us that if the test runs for 13 hours without failure, there is a 99 %
probability that our fix actually improved the program’s reliability.
A dogmatic statistician might not approve of this approach, but the sad fact is that the
errors introduced by this sort of statistical abuse are usually quite small compared to
the errors in your failure-rate estimates. Nevertheless, the next section takes a more
rigorous approach.
11.6.3 Statistics for Continuous Testing

The fundamental formula for failure probabilities is the Poisson distribution:
𝜆 𝑚 −𝜆
𝐹𝑚 = e (11.8)
𝑚!
v2023.06.11a
Here 𝐹𝑚 is the probability of 𝑚 failures in the test and 𝜆 is the expected failure
rate per unit time. A rigorous derivation may be found in any advanced probability
textbook, for example, Feller’s classic “An Introduction to Probability Theory and Its
Applications” [Fel50], while a more intuitive derivation may be found in the first edition
of this book [McK14c, Equations 11.8–11.26].
Let’s try reworking the example from Section 11.6.2 using the Poisson distribution.
Recall that this example involved a test with a 30 % failure rate per hour, and that the
question was how long the test would need to run error-free on a alleged fix to be 99 %
certain that the fix actually reduced the failure rate. In this case, 𝑚 is zero, so that
Eq. 11.8 reduces to:
𝐹0 = e−𝜆 (11.9)
Solving this requires setting 𝐹0 to 0.01 and solving for 𝜆, resulting in:
𝜆 = − ln 0.01 = 4.6 (11.10)

Because we get 0.3 failures per hour, the number of hours required is 4.6/0.3 = 14.3,
which is within 10 % of the 13 hours calculated using the method in Section 11.6.2.
Given that you normally won’t know your failure rate to anywhere near 10 %, the simpler
method described in Section 11.6.2 is almost always good and sufficient.
More generally, if we have 𝑛 failures per unit time, and we want to be 𝑃 % certain
that a fix reduced the failure rate, we can use the following formula:
1 100 − 𝑃
𝑇 = − ln (11.11)
𝑛 100
Quick Quiz 11.15: Suppose that a bug causes a test failure three times per hour on average.
How long must the test run error-free to provide 99.9 % confidence that the fix significantly
reduced the probability of failure?
As before, the less frequently the bug occurs and the greater the required level of
confidence, the longer the required error-free test run.
Suppose that a given test fails about once every hour, but after a bug fix, a 24-hour
test run fails only twice. Assuming that the failure leading to the bug is a random
occurrence, what is the probability that the small number of failures in the second run
was due to random chance? In other words, how confident should we be that the fix
actually had some effect on the bug? This probability may be calculated by summing
Eq. 11.8 as follows:
𝑚
∑︁ 𝜆𝑖
𝐹0 + 𝐹1 + · · · + 𝐹𝑚−1 + 𝐹𝑚 = e−𝜆 (11.12)
𝑖=0
𝑖!
This is the Poisson cumulative distribution function, which can be written more
compactly as:
𝑚
∑︁ 𝜆𝑖
𝐹𝑖 ≤𝑚 = e−𝜆 (11.13)
𝑖=0
𝑖!
Here 𝑚 is the actual number of errors in the long test run (in this case, two) and 𝜆
is expected number of errors in the long test run (in this case, 24). Plugging 𝑚 = 2
and 𝜆 = 24 into this expression gives the probability of two or fewer failures as about
v2023.06.11a
1.2 × 10−8 , in other words, we have a high level of confidence that the fix actually had
some relationship to the bug.7
Quick Quiz 11.16: Doing the summation of all the factorials and exponentials is a real pain.
Isn’t there an easier way?
Quick Quiz 11.17: But wait!!! Given that there has to be some number of failures (including
the possibility of zero failures), shouldn’t Eq. 11.13 approach the value 1 as 𝑚 goes to infinity?
The Poisson distribution is a powerful tool for analyzing test results, but the fact is
that in this last example there were still two remaining test failures in a 24-hour test
run. Such a low failure rate results in very long test runs. The next section discusses
counter-intuitive ways of improving this situation.
11.6.4 Hunting Heisenbugs

This line of thought also helps explain heisenbugs: Adding tracing and assertions can
easily reduce the probability of a bug appearing, which is why extremely lightweight
tracing and assertion mechanism are so critically important.
The term “heisenbug” was inspired by the Heisenberg Uncertainty Principle from
quantum physics, which states that it is impossible to exactly quantify a given particle’s
position and velocity at any given point in time [Hei27]. Any attempt to more accurately
measure that particle’s position will result in increased uncertainty of its velocity and
vice versa. Similarly, attempts to track down the heisenbug causes its symptoms to
radically change or even disappear completely.8 Of course, adding debugging overhead
can and sometimes does make bugs more probable. But developers are more likely to
remember the frustration of a disappearing heisenbug than the joy inspired by the bug
becoming more easily reproduced!
If the field of physics inspired the name of this problem, it is only fair that the field of
physics should inspire the solution. Fortunately, particle physics is up to the task: Why
not create an anti-heisenbug to annihilate the heisenbug? Or, perhaps more accurately,
to annihilate the heisen-ness of the heisenbug? Although producing an anti-heisenbug
for a given heisenbug is more an art than a science, the following sections describe a
number of ways to do just that:
1. Add delay to race-prone regions (Section 11.6.4.1).

2. Increase workload intensity (Section 11.6.4.2).
3. Isolate suspicious subsystems (Section 11.6.4.3).

4. Simulate unusual events (Section 11.6.4.4).
5. Count near misses (Section 11.6.4.5).
These are followed by discussion in Section 11.6.4.6.
7 Of course, this result in no way excuses you from finding and fixing the bug(s) resulting
in the remaining two failures!

8 The term “heisenbug” is a misnomer, as most heisenbugs are fully explained by the
observer effect from classical physics. Nevertheless, the name has stuck.
v2023.06.11a
11.6.4.1 Add Delay
Consider the count-lossy code in Section 5.1. Adding printf() statements will likely
greatly reduce or even eliminate the lost counts. However, converting the load-add-store
sequence to a load-add-delay-store sequence will greatly increase the incidence of lost
counts (try it!). Once you spot a bug involving a race condition, it is frequently possible
to create an anti-heisenbug by adding delay in this manner.
Of course, this begs the question of how to find the race condition in the first place.
Although very lucky developers might accidentally create delay-based anti-heisenbugs
when adding debug code, this is in general a dark art. Nevertheless, there are a number
of things you can do to find your race conditions.
One approach is to recognize that race conditions often end up corrupting some
of the data involved in the race. It is therefore good practice to double-check the
synchronization of any corrupted data. Even if you cannot immediately recognize
the race condition, adding delay before and after accesses to the corrupted data might
change the failure rate. By adding and removing the delays in an organized fashion (e.g.,
binary search), you might learn more about the workings of the race condition.
Quick Quiz 11.18: How is this approach supposed to help if the corruption affected some
unrelated pointer, which then caused the corruption???
Another important approach is to vary the software and hardware configuration and
look for statistically significant differences in failure rate. For example, back in the
1990s, it was common practice to test on systems having CPUs running at different
clock rates, which tended to make some types of race conditions more probable. One
way of getting a similar effect today is to test on multi-socket systems, thus incurring
the large delays described in Section 3.2.
However you choose to add delays, you can then look more intensively at the code
implicated by those delays that make the greatest difference in failure rate. It might be
helpful to test that code in isolation, for example.
One important aspect of software configuration is the history of changes, which is
why git bisect is so useful. Bisection of the change history can provide very valuable
clues as to the nature of the heisenbug, in this case presumably by locating a commit
that shows a change in the software’s response to the addition or removal of a given
delay.
Quick Quiz 11.19: But I did the bisection, and ended up with a huge commit. What do I do
now?
Once you locate the suspicious section of code, you can then introduce delays to
attempt to increase the probability of failure. As we have seen, increasing the probability
of failure makes it much easier to gain high confidence in the corresponding fix.
However, it is sometimes quite difficult to track down the problem using normal
debugging techniques. The following sections present some other alternatives.
11.6.4.2 Increase Workload Intensity
It is often the case that a given test suite places relatively low stress on a given subsystem,
so that a small change in timing can cause a heisenbug to disappear. One way to create
an anti-heisenbug for this case is to increase the workload intensity, which has a good
chance of increasing the bug’s probability. If the probability is increased sufficiently, it
v2023.06.11a
may be possible to add lightweight diagnostics such as tracing without causing the bug
to vanish.
How can you increase the workload intensity? This depends on the program, but here
are some things to try:
1. Add more CPUs.

2. If the program uses networking, add more network adapters and more or faster
remote systems.
3. If the program is doing heavy I/O when the problem occurs, either (1) add more
storage devices, (2) use faster storage devices, for example, substitute SSDs for
disks, or (3) use a RAM-based filesystem to substitute main memory for mass
storage.
4. Change the size of the problem, for example, if doing a parallel matrix multiply,
change the size of the matrix. Larger problems may introduce more complexity,
but smaller problems often increase the level of contention. If you aren’t sure
whether you should go large or go small, just try both.
However, it is often the case that the bug is in a specific subsystem, and the structure
of the program limits the amount of stress that can be applied to that subsystem. The
next section addresses this situation.
11.6.4.3 Isolate Suspicious Subsystems

If the program is structured such that it is difficult or impossible to apply much stress to
a subsystem that is under suspicion, a useful anti-heisenbug is a stress test that tests
that subsystem in isolation. The Linux kernel’s rcutorture module takes exactly this
approach with RCU: Applying more stress to RCU than is feasible in a production
environment increases the probability that RCU bugs will be found during testing rather
than in production.9
In fact, when creating a parallel program, it is wise to stress-test the components
separately. Creating such component-level stress tests can seem like a waste of time, but
a little bit of component-level testing can save a huge amount of system-level debugging.
11.6.4.4 Simulate Unusual Events

Heisenbugs are sometimes due to unusual events, such as memory-allocation failure,
conditional-lock-acquisition failure, CPU-hotplug operations, timeouts, packet losses,
and so on. One way to construct an anti-heisenbug for this class of heisenbug is to
introduce spurious failures.
For example, instead of invoking malloc() directly, invoke a wrapper function that
uses a random number to decide whether to return NULL unconditionally on the one hand,
or to actually invoke malloc() and return the resulting pointer on the other. Inducing
spurious failures is an excellent way to bake robustness into sequential programs as well
as parallel programs.
Quick Quiz 11.20: Why don’t conditional-locking primitives provide this spurious-failure
functionality?
9 Though sadly not increased to probability one.
v2023.06.11a
call_rcu()
Grace-Period Start
Near Miss
Reader
Reader Error
Time
Grace-Period End
Callback Invocation
Figure 11.5: RCU Errors and Near Misses
11.6.4.5 Count Near Misses

Bugs are often all-or-nothing things, so that a bug either happens or not, with nothing in
between. However, it is sometimes possible to define a near miss where the bug does
not result in a failure, but has likely manifested. For example, suppose your code is
making a robot walk. The robot’s falling down constitutes a bug in your program, but
stumbling and recovering might constitute a near miss. If the robot falls over only once
per hour, but stumbles every few minutes, you might be able to speed up your debugging
progress by counting the number of stumbles in addition to the number of falls.
In concurrent programs, timestamping can sometimes be used to detect near misses.
For example, locking primitives incur significant delays, so if there is a too-short delay
between a pair of operations that are supposed to be protected by different acquisitions
of the same lock, this too-short delay might be counted as a near miss.10
For example, a low-probability bug in RCU priority boosting occurred roughly once
every hundred hours of focused rcutorture testing. Because it would take almost
500 hours of failure-free testing to be 99 % certain that the bug’s probability had been
significantly reduced, the git bisect process to find the failure would be painfully
slow—or would require an extremely large test farm. Fortunately, the RCU operation
being tested included not only a wait for an RCU grace period, but also a previous wait
for the grace period to start and a subsequent wait for an RCU callback to be invoked
after completion of the RCU grace period. This distinction between an rcutorture
error and near miss is shown in Figure 11.5. To qualify as a full-fledged error, an RCU
read-side critical section must extend from the call_rcu() that initiated a grace period,
through the remainder of the previous grace period, through the entirety of the grace
period initiated by the call_rcu() (denoted by the region between the jagged lines),
and through the delay from the end of that grace period to the callback invocation, as
indicated by the “Error” arrow. However, the formal definition of RCU prohibits RCU
read-side critical sections from extending across a single grace period, as indicated by
the “Near Miss” arrow. This suggests using near misses as the error condition, however,
this can be problematic because different CPUs can have different opinions as to exactly
where a given grace period starts and ends, as indicated by the jagged lines.11 Using the
10 Of course, in this case, you might be better off using whatever lock_held() primitive
is available in your environment. If there isn’t a lock_held() primitive, create one!

11 In real life, these lines can be much more jagged because idle CPUs can be completely
unaware of a great many recent grace periods.
v2023.06.11a
near misses as the error condition could therefore result in false positives, which need
to be avoided in the automated rcutorture testing.
By sheer dumb luck, rcutorture happens to include some statistics that are sensitive
to the near-miss version of the grace period. As noted above, these statistics are subject
to false positives due to their unsynchronized access to RCU’s state variables, but these
false positives turn out to be extremely rare on strongly ordered systems such as the
IBM mainframe and x86, occurring less than once per thousand hours of testing.
These near misses occurred roughly once per hour, about two orders of magnitude
more frequently than the actual errors. Use of these near misses allowed the bug’s root
cause to be identified in less than a week and a high degree of confidence in the fix to
be built in less than a day. In contrast, excluding the near misses in favor of the real
errors would have required months of debug and validation time.
To sum up near-miss counting, the general approach is to replace counting of
infrequent failures with more-frequent near misses that are believed to be correlated
with those failures. These near-misses can be considered an anti-heisenbug to the real
failure’s heisenbug because the near-misses, being more frequent, are likely to be more
robust in the face of changes to your code, for example, the changes you make to add
debugging code.
11.6.4.6 Heisenbug Discussion

The alert reader might have noticed that this section was fuzzy and qualitative, in stark
contrast to the precise mathematics of Sections 11.6.1, 11.6.2, and 11.6.3. If you love
precision and mathematics, you may be disappointed to learn that the situations to which
this section applies are far more common than those to which the preceding sections
apply.
In fact, the common case is that although you might have reason to believe that your
code has bugs, you have no idea what those bugs are, what causes them, how likely
they are to appear, or what conditions affect their probability of appearance. In this
all-too-common case, statistics cannot help you.12 That is to say, statistics cannot help
you directly. But statistics can be of great indirect help—if you have the humility
required to admit that you make mistakes, that you can reduce the probability of these
mistakes (for example, by getting enough sleep), and that the number and type of
mistakes you made in the past is indicative of the number and type of mistakes that you
are likely to make in the future. For example, I have a deplorable tendency to forget to
write a small but critical portion of the initialization code, and frequently get most or
even all of a parallel program correct—except for a stupid omission in initialization.
Once I was willing to admit to myself that I am prone to this type of mistake, it was
easier (but not easy!) to force myself to double-check my initialization code. Doing this
allowed me to find numerous bugs ahead of time.
When your quick bug hunt morphs into a long-term quest, it is important to log
everything you have tried and what happened. In the common case where the software
is changing during the course of your quest, make sure to record the exact version of the
software to which each log entry applies. From time to time, reread the entire log in
order to make connections between clues encountered at different times. Such rereading
is especially important upon encountering a surprising test result, for example, I reread
my log upon realizing that what I thought was a failure of the hypervisor to schedule
12 Although if you know what your program is supposed to do and if your program is small
enough (both less likely that you might think), then the formal-verification tools described in
Chapter 12 can be helpful.
v2023.06.11a
a vCPU was instead an interrupt storm preventing that vCPU from making forward
progress on the interrupted code. If the code you are debugging is new to you, this log is
also an excellent place to document the relationships between code and data structures.
Keeping a log when you are furiously chasing a difficult bug might seem like needless
paperwork, but it has on many occasions saved me from debugging around and around
in circles, which can waste far more time than keeping a log ever could.
Using Taleb’s nomenclature [Tal07], a white swan is a bug that we can reproduce. We
can run a large number of tests, use ordinary statistics to estimate the bug’s probability,
and use ordinary statistics again to estimate our confidence in a proposed fix. An
unsuspected bug is a black swan. We know nothing about it, we have no tests that
have yet caused it to happen, and statistics is of no help. Studying our own behavior,
especially the number and types of mistakes we make, can turn black swans into grey
swans. We might not know exactly what the bugs are, but we have some idea of their
number and maybe also of their type. Ordinary statistics is still of no help (at least not
until we are able to reproduce one of the bugs), but robust13 testing methods can be of
great help. The goal, therefore, is to use experience and good validation practices to
turn the black swans grey, focused testing and analysis to turn the grey swans white, and
ordinary methods to fix the white swans.
That said, thus far, we have focused solely on bugs in the parallel program’s
functionality. However, performance is a first-class requirement for a parallel program.
Otherwise, why not write a sequential program? To repurpose Kipling, our goal when
writing parallel code is to fill the unforgiving second with sixty minutes worth of
distance run. The next section therefore discusses a number of performance bugs that
would be happy to thwart this Kiplingesque goal.
11.7 Performance Estimation

There are lies, damn lies, statistics, and benchmarks.
Unknown
Parallel programs usually have performance and scalability requirements, after all, if
performance is not an issue, why not use a sequential program? Ultimate performance
and linear scalability might not be necessary, but there is little use for a parallel program
that runs slower than its optimal sequential counterpart. And there really are cases
where every microsecond matters and every nanosecond is needed. Therefore, for
parallel programs, insufficient performance is just as much a bug as is incorrectness.
Quick Quiz 11.21: That is ridiculous!!! After all, isn’t getting the correct answer later than
one would like better than getting an incorrect answer???
Quick Quiz 11.22: But if you are going to put in all the hard work of parallelizing an
application, why not do it right? Why settle for anything less than optimal performance and
linear scalability?
Validating a parallel program must therfore include validating its performance. But
validating performance means having a workload to run and performance criteria with
which to evaluate the program at hand. These needs are often met by performance
benchmarks, which are discussed in the next section.
13 That is to say brutal.
v2023.06.11a
11.7. PERFORMANCE ESTIMATION 347
11.7.1 Benchmarking
Frequent abuse aside, benchmarks are both useful and heavily used, so it is not helpful
to be too dismissive of them. Benchmarks span the range from ad hoc test jigs to
international standards, but regardless of their level of formality, benchmarks serve four
major purposes:
1. Providing a fair framework for comparing competing implementations.
2. Focusing competitive energy on improving implementations in ways that matter to

users.
3. Serving as example uses of the implementations being benchmarked.
4. Serving as a marketing tool to highlight your software against your competitors’
offerings.
Of course, the only completely fair framework is the intended application itself. So
why would anyone who cared about fairness in benchmarking bother creating imperfect
benchmarks rather than simply using the application itself as the benchmark?
Running the actual application is in fact the best approach where it is practical.
Unfortunately, it is often impractical for the following reasons:
1. The application might be proprietary, and you might not have the right to run the
intended application.
2. The application might require more hardware than you have access to.
3. The application might use data that you cannot access, for example, due to privacy
regulations.
4. The application might take longer than is convenient to reproduce a performance
or scalability problem.14
Creating a benchmark that approximates the application can help overcome these
obstacles. A carefully constructed benchmark can help promote performance, scalability,
energy efficiency, and much else besides. However, be careful to avoid investing too
much into the benchmarking effort. It is after all important to invest at least a little into
the application itself [Gra91].
11.7.2 Profiling
In many cases, a fairly small portion of your software is responsible for the majority of
the performance and scalability shortfall. However, developers are notoriously unable
to identify the actual bottlenecks by inspection. For example, in the case of a kernel
buffer allocator, all attention focused on a search of a dense array which turned out to
represent only a few percent of the allocator’s execution time. An execution profile
collected via a logic analyzer focused attention on the cache misses that were actually
responsible for the majority of the problem [MS93].
An old-school but quite effective method of tracking down performance and scalability
bugs is to run your program under a debugger, then periodically interrupt it, recording
14 Microbenchmarks can help, but please see Section 11.7.4.
v2023.06.11a
the stacks of all threads at each interruption. The theory here is that if something is
slowing down your program, it has to be visible in your threads’ executions.
That said, there are a number of tools that will usually do a much better job of helping
you to focus your attention where it will do the most good. Two popular choices are
gprof and perf. To use perf on a single-process program, prefix your command with
perf record, then after the command completes, type perf report. There is a lot
of work on tools for performance debugging of multi-threaded programs, which should
make this important job easier. Again, one good starting point is Brendan Gregg’s
blog.15
11.7.3 Differential Profiling

Scalability problems will not necessarily be apparent unless you are running on very
large systems. However, it is sometimes possible to detect impending scalability
problems even when running on much smaller systems. One technique for doing this is
called differential profiling.
The idea is to run your workload under two different sets of conditions. For example,
you might run it on two CPUs, then run it again on four CPUs. You might instead
vary the load placed on the system, the number of network adapters, the number
of mass-storage devices, and so on. You then collect profiles of the two runs, and
mathematically combine corresponding profile measurements. For example, if your
main concern is scalability, you might take the ratio of corresponding measurements,
and then sort the ratios into descending numerical order. The prime scalability suspects
will then be sorted to the top of the list [McK95, McK99].
Some tools such as perf have built-in differential-profiling support.
11.7.4 Microbenchmarking
Microbenchmarking can be useful when deciding which algorithms or data structures
are worth incorporating into a larger body of software for deeper evaluation.
One common approach to microbenchmarking is to measure the time, run some
number of iterations of the code under test, then measure the time again. The difference
between the two times divided by the number of iterations gives the measured time
required to execute the code under test.
Unfortunately, this approach to measurement allows any number of errors to creep in,
including:
1. The measurement will include some of the overhead of the time measurement.
This source of error can be reduced to an arbitrarily small value by increasing the
number of iterations.
2. The first few iterations of the test might incur cache misses or (worse yet) page
faults that might inflate the measured value. This source of error can also be
reduced by increasing the number of iterations, or it can often be eliminated entirely
by running a few warm-up iterations before starting the measurement period. Most
systems have ways of detecting whether a given process incurred a page fault,
and you should make use of this to reject runs whose performance has been thus
impeded.
15 http://www.brendangregg.com/blog/
v2023.06.11a
3. Some types of interference, for example, random memory errors, are so rare that
they can be dealt with by running a number of sets of iterations of the test. If the
level of interference was statistically significant, any performance outliers could be
rejected statistically.
4. Any iteration of the test might be interfered with by other activity on the system.
Sources of interference include other applications, system utilities and daemons,
device interrupts, firmware interrupts (including system management interrupts, or
SMIs), virtualization, memory errors, and much else besides. Assuming that these
sources of interference occur randomly, their effect can be minimized by reducing
the number of iterations.
5. Thermal throttling can understate scalability because increasing CPU activity

increases heat generation, and on systems without adequate cooling (most of
them!), this can result in the CPU frequency decreasing as the number of CPUs
increases.16 Of course, if you are testing an application to evaluate its expected
behavior when run in production, such thermal throttling is simply a fact of life.
Otherwise, if you are interested in theoretical scalability, use a system with adequate
cooling or reduce the CPU clock rate to a level that the cooling system can handle.
The first and fourth sources of interference provide conflicting advice, which is one
sign that we are living in the real world.
Quick Quiz 11.23: But what about other sources of error, for example, due to interactions
between caches and memory layout?
The following sections discuss ways of dealing with these measurement errors, with
Section 11.7.5 covering isolation techniques that may be used to prevent some forms of
interference, and with Section 11.7.6 covering methods for detecting interference so as
to reject measurement data that might have been corrupted by that interference.
11.7.5 Isolation
The Linux kernel provides a number of ways to isolate a group of CPUs from outside
interference.
First, let’s look at interference by other processes, threads, and tasks. The POSIX
sched_setaffinity() system call may be used to move most tasks off of a given
set of CPUs and to confine your tests to that same group. The Linux-specific user-
level taskset command may be used for the same purpose, though both sched_
setaffinity() and taskset require elevated permissions. Linux-specific control
groups (cgroups) may be used for this same purpose. This approach can be quite
effective at reducing interference, and is sufficient in many cases. However, it does have
limitations, for example, it cannot do anything about the per-CPU kernel threads that
are often used for housekeeping tasks.
One way to avoid interference from per-CPU kernel threads is to run your test at
a high real-time priority, for example, by using the POSIX sched_setscheduler()
system call. However, note that if you do this, you are implicitly taking on responsibility
for avoiding infinite loops, because otherwise your test can prevent part of the kernel
from functioning. This is an example of the Spiderman Principle: “With great power
comes great responsibility.” And although the default real-time throttling settings often
16 Systems with adequate cooling tend to look like gaming systems.
v2023.06.11a
address such problems, they might do so by causing your real-time threads to miss their
deadlines.
These approaches can greatly reduce, and perhaps even eliminate, interference
from processes, threads, and tasks. However, it does nothing to prevent interfer-
ence from device interrupts, at least in the absence of threaded interrupts. Linux
allows some control of threaded interrupts via the /proc/irq directory, which
contains numerical directories, one per interrupt vector. Each numerical direc-
tory contains smp_affinity and smp_affinity_list. Given sufficient permis-
sions, you can write a value to these files to restrict interrupts to the specified
set of CPUs. For example, either “echo 3 > /proc/irq/23/smp_affinity” or
“echo 0-1 > /proc/irq/23/smp_affinity_list” would confine interrupts on
vector 23 to CPUs 0 and 1, at least given sufficient privileges. You can use “cat
/proc/interrupts” to obtain a list of the interrupt vectors on your system, how many
are handled by each CPU, and what devices use each interrupt vector.
Running a similar command for all interrupt vectors on your system would confine
interrupts to CPUs 0 and 1, leaving the remaining CPUs free of interference. Or mostly
free of interference, anyway. It turns out that the scheduling-clock interrupt fires on
each CPU that is running in user mode.17 In addition you must take care to ensure that
the set of CPUs that you confine the interrupts to is capable of handling the load.
But this only handles processes and interrupts running in the same operating-system
instance as the test. Suppose that you are running the test in a guest OS that is itself
running on a hypervisor, for example, Linux running KVM? Although you can in theory
apply the same techniques at the hypervisor level that you can at the guest-OS level, it is
quite common for hypervisor-level operations to be restricted to authorized personnel.
In addition, none of these techniques work against firmware-level interference.
Quick Quiz 11.24: Wouldn’t the techniques suggested to isolate the code under test also affect
that code’s performance, particularly if it is running within a larger application?
Of course, if it is in fact the interference that is producing the behavior of interest,

you will instead need to promote interference, in which case being unable to prevent it
is not a problem. But if you really do need interference-free measurements, then instead
of preventing the interference, you might need to detect the interference as described in
the next section.
11.7.6 Detecting Interference

If you cannot prevent interference, perhaps you can detect it and reject results from any
affected test runs. Section 11.7.6.1 describes methods of rejection involving additional
measurements, while Section 11.7.6.2 describes statistics-based rejection.
11.7.6.1 Detecting Interference Via Measurement

Many systems, including Linux, provide means for determining after the fact whether
some forms of interference have occurred. For example, process-based interference
results in context switches, which, on Linux-based systems, are visible in /proc/
<PID>/sched via the nr_switches field. Similarly, interrupt-based interference can
be detected via the /proc/interrupts file.
17 Frederic Weisbecker leads up a NO_HZ_FULL adaptive-ticks project that allows
scheduling-clock interrupts to be disabled on CPUs that have only one runnable task.
As of 2021, this is largely complete.
v2023.06.11a
Listing 11.1: Using getrusage() to Detect Context Switches

1 #include <sys/time.h>
2 #include <sys/resource.h>
3
4 /* Return 0 if test results should be rejected. */
5 int runtest(void)
6 {
7 struct rusage ru1;
8 struct rusage ru2;
9
10 if (getrusage(RUSAGE_SELF, &ru1) != 0) {
11 perror("getrusage");
12 abort();
13 }
14 /* run test here. */
15 if (getrusage(RUSAGE_SELF, &ru2 != 0) {
16 perror("getrusage");
17 abort();
18 }
19 return (ru1.ru_nvcsw == ru2.ru_nvcsw &&
20 ru1.runivcsw == ru2.runivcsw);
21 }
Opening and reading files is not the way to low overhead, and it is possible to get the
count of context switches for a given thread by using the getrusage() system call, as
shown in Listing 11.1. This same system call can be used to detect minor page faults
(ru_minflt) and major page faults (ru_majflt).
Unfortunately, detecting memory errors and firmware interference is quite system-
specific, as is the detection of interference due to virtualization. Although avoidance is
better than detection, and detection is better than statistics, there are times when one
must avail oneself of statistics, a topic addressed in the next section.
11.7.6.2 Detecting Interference Via Statistics

Any statistical analysis will be based on assumptions about the data, and performance
microbenchmarks often support the following assumptions:
1. Smaller measurements are more likely to be accurate than larger measurements.
2. The measurement uncertainty of good data is known.

3. A reasonable fraction of the test runs will result in good data.
The fact that smaller measurements are more likely to be accurate than larger
measurements suggests that sorting the measurements in increasing order is likely to be
productive.18 The fact that the measurement uncertainty is known allows us to accept
measurements within this uncertainty of each other: If the effects of interference are
large compared to this uncertainty, this will ease rejection of bad data. Finally, the fact
that some fraction (for example, one third) can be assumed to be good allows us to
blindly accept the first portion of the sorted list, and this data can then be used to gain
an estimate of the natural variation of the measured data, over and above the assumed
measurement error.
The approach is to take the specified number of leading elements from the beginning
of the sorted list, and use these to estimate a typical inter-element delta, which in turn
may be multiplied by the number of elements in the list to obtain an upper bound on
18 To paraphrase the old saying, “Sort first and ask questions later.”
v2023.06.11a
Listing 11.2: Statistical Elimination of Interference

1 div=3
2 rel=0.01
3 tre=10
4 while test $# -gt 0
5 do
6 case "$1" in
7 --divisor)
8 shift
9 div=$1
10 ;;
11 --relerr)
12 shift
13 rel=$1
14 ;;
15 --trendbreak)
16 shift
17 tre=$1
18 ;;
19 esac
20 shift
21 done
22
23 awk -v divisor=$div -v relerr=$rel -v trendbreak=$tre '{
24 for (i = 2; i <= NF; i++)
25 d[i - 1] = $i;
26 asort(d);
27 i = int((NF + divisor - 1) / divisor);
28 delta = d[i] - d[1];
29 maxdelta = delta * divisor;
30 maxdelta1 = delta + d[i] * relerr;
31 if (maxdelta1 > maxdelta)
32 maxdelta = maxdelta1;
33 for (j = i + 1; j < NF; j++) {
34 if (j <= 2)
35 maxdiff = d[NF - 1] - d[1];
36 else
37 maxdiff = trendbreak * (d[j - 1] - d[1]) / (j - 2);
38 if (d[j] - d[1] > maxdelta && d[j] - d[j - 1] > maxdiff)
39 break;
40 }
41 n = sum = 0;
42 for (k = 1; k < j; k++) {
43 sum += d[k];
44 n++;
45 }
46 min = d[1];
47 max = d[j - 1];
48 avg = sum / n;
49 print $1, avg, min, max, n, NF - 1;
50 }'
permissible values. The algorithm then repeatedly considers the next element of the list.
If it falls below the upper bound, and if the distance between the next element and the
previous element is not too much greater than the average inter-element distance for the
portion of the list accepted thus far, then the next element is accepted and the process
repeats. Otherwise, the remainder of the list is rejected.
Listing 11.2 shows a simple sh/awk script implementing this notion. Input consists
of an x-value followed by an arbitrarily long list of y-values, and output consists of one
line for each input line, with fields as follows:
1. The x-value.
2. The average of the selected data.
3. The minimum of the selected data.
v2023.06.11a
4. The maximum of the selected data.
5. The number of selected data items.
6. The number of input data items.
This script takes three optional arguments as follows:
--divisor: Number of segments to divide the list into, for example, a divisor of four
means that the first quarter of the data elements will be assumed to be good. This
defaults to three.
--relerr: Relative measurement error. The script assumes that values that differ by
less than this error are for all intents and purposes equal. This defaults to 0.01,
which is equivalent to 1 %.
--trendbreak: Ratio of inter-element spacing constituting a break in the trend of the

data. For example, if the average spacing in the data accepted so far is 1.5, then if
the trend-break ratio is 2.0, then if the next data value differs from the last one by
more than 3.0, this constitutes a break in the trend. (Unless of course, the relative
error is greater than 3.0, in which case the “break” will be ignored.)
Lines 1–3 of Listing 11.2 set the default values for the parameters, and lines 4–21 parse
any command-line overriding of these parameters. The awk invocation on line 23 sets
the values of the divisor, relerr, and trendbreak variables to their sh counterparts.
In the usual awk manner, lines 24–50 are executed on each input line. The loop spanning
lines 24 and 25 copies the input y-values to the d array, which line 26 sorts into increasing
order. Line 27 computes the number of trustworthy y-values by applying divisor and
rounding up.
Lines 28–32 compute the maxdelta lower bound on the upper bound of y-values. To
this end, line 29 multiplies the difference in values over the trusted region of data by the
divisor, which projects the difference in values across the trusted region across the
entire set of y-values. However, this value might well be much smaller than the relative
error, so line 30 computes the absolute error (d[i] * relerr) and adds that to the
difference delta across the trusted portion of the data. Lines 31 and 32 then compute
the maximum of these two values.
Each pass through the loop spanning lines 33–40 attempts to add another data value to
the set of good data. Lines 34–39 compute the trend-break delta, with line 34 disabling
this limit if we don’t yet have enough values to compute a trend, and with line 37
multiplying trendbreak by the average difference between pairs of data values in the
good set. If line 38 determines that the candidate data value would exceed the lower
bound on the upper bound (maxdelta) and that the difference between the candidate
data value and its predecessor exceeds the trend-break difference (maxdiff), then
line 39 exits the loop: We have the full good set of data.
Lines 41–49 then compute and print statistics.
Quick Quiz 11.25: This approach is just plain weird! Why not use means and standard
deviations, like we were taught in our statistics classes?
Quick Quiz 11.26: But what if all the y-values in the trusted group of data are exactly zero?
Won’t that cause the script to reject any non-zero value?
v2023.06.11a
Figure 11.6: Choose Validation Methods Wisely
Although statistical interference detection can be quite useful, it should be used only
as a last resort. It is far better to avoid interference in the first place (Section 11.7.5), or,
failing that, detecting interference via measurement (Section 11.7.6.1).
11.8 Summary
To err is human! Stop being human‼!

Ed Nofziger
Although validation never will be an exact science, much can be gained by taking
an organized approach to it, as an organized approach will help you choose the right
validation tools for your job, avoiding situations like the one fancifully depicted in
Figure 11.6.
A key choice is that of statistics. Although the methods described in this chapter
work very well most of the time, they do have their limitations, courtesy of the Halting
Problem [Tur37, Pul00]. Fortunately for us, there is a huge number of special cases in
which we can not only work out whether a program will halt, but also estimate how long
it will run before halting, as discussed in Section 11.7. Furthermore, in cases where a
given program might or might not work correctly, we can often establish estimates for
what fraction of the time it will work correctly, as discussed in Section 11.6.
Nevertheless, unthinking reliance on these estimates is brave to the point of fool-
hardiness. After all, we are summarizing a huge mass of complexity in code and data
structures down to a single solitary number. Even though we can get away with such
bravery a surprisingly large fraction of the time, abstracting all that code and data away
will occasionally cause severe problems.
One possible problem is variability, where repeated runs give wildly different results.
This problem is often addressed using standard deviation, however, using two numbers
to summarize the behavior of a large and complex program is about as brave as using
only one number. In computer programming, the surprising thing is that use of the
mean or the mean and standard deviation are often sufficient. Nevertheless, there are no
guarantees.
One cause of variation is confounding factors. For example, the CPU time consumed
by a linked-list search will depend on the length of the list. Averaging together runs
with wildly different list lengths will probably not be useful, and adding a standard
deviation to the mean will not be much better. The right thing to do would be control for
v2023.06.11a
11.8. SUMMARY 355
list length, either by holding the length constant or to measure CPU time as a function
of list length.
Of course, this advice assumes that you are aware of the confounding factors, and
Murphy says that you will not be. I have been involved in projects that had confounding
factors as diverse as air conditioners (which drew considerable power at startup, thus
causing the voltage supplied to the computer to momentarily drop too low, sometimes
resulting in failure), cache state (resulting in odd variations in performance), I/O errors
(including disk errors, packet loss, and duplicate Ethernet MAC addresses), and even
porpoises (which could not resist playing with an array of transponders, which could be
otherwise used for high-precision acoustic positioning and navigation). And this is but
one reason why a good night’s sleep is such an effective debugging tool.
In short, validation always will require some measure of the behavior of the system.
To be at all useful, this measure must be a severe summarization of the system, which in
turn means that it can be misleading. So as the saying goes, “Be careful. It is a real
world out there.”
But what if you are working on the Linux kernel, which as of 2017 was estimated to
have more than 20 billion instances running throughout the world? In that case, a bug
that occurs once every million years on a single system will be encountered more than
50 times per day across the installed base. A test with a 50 % chance of encountering
this bug in a one-hour run would need to increase that bug’s probability of occurrence
by more than ten orders of magnitude, which poses a severe challenge to today’s testing
methodologies. One important tool that can sometimes be applied with good effect
to such situations is formal verification, the subject of the next chapter, and, more
speculatively, Section 17.4.
The topic of choosing a validation plan, be it testing, formal verification, or both, is
taken up by Section 12.7.
v2023.06.11a
v2023.06.11a
Beware of bugs in the above code; I have only proved
it correct, not tried it.
Donald Knuth
Chapter 12
Formal Verification
Parallel algorithms can be hard to write, and even harder to debug. Testing, though
essential, is insufficient, as fatal race conditions can have extremely low probabilities
of occurrence. Proofs of correctness can be valuable, but in the end are just as prone
to human error as is the original algorithm. In addition, a proof of correctness cannot
be expected to find errors in your assumptions, shortcomings in the requirements,
misunderstandings of the underlying software or hardware primitives, or errors that you
did not think to construct a proof for. This means that formal methods can never replace
testing. Nevertheless, formal methods can be a valuable addition to your validation
toolbox.
It would be very helpful to have a tool that could somehow locate all race conditions.
A number of such tools exist, for example, Section 12.1 provides an introduction to
the general-purpose state-space search tools Promela and Spin, Section 12.2 similarly
introduces the special-purpose ppcmem tool, Section 12.3 looks at an example axiomatic
approach, Section 12.4 briefly overviews SAT solvers, Section 12.5 briefly overviews
stateless model checkers, Section 12.6 sums up use of formal-verification tools for
verifying parallel algorithms, and finally Section 12.7 discusses how to decide how
much and what type of validation to apply to a given software project.
12.1 State-Space Search
Follow every byway / Every path you know.

Climb Every Mountain, Rodgers & Hammerstein
This section features the general-purpose Promela and Spin tools, which may be used
to carry out a full state-space search of many types of multi-threaded code. They are
used to verifying data communication protocols. Section 12.1.1 introduces Promela and
Spin, including a couple of warm-up exercises verifying both non-atomic and atomic
increment. Section 12.1.2 describes use of Promela, including example command
lines and a comparison of Promela syntax to that of C. Section 12.1.3 shows how
Promela may be used to verify locking, Section 12.1.4 uses Promela to verify an unusual
implementation of RCU named “QRCU”, and finally Section 12.1.5 applies Promela to
early versions of RCU’s dyntick-idle implementation.
357
v2023.06.11a
358 CHAPTER 12. FORMAL VERIFICATION
12.1.1 Promela and Spin

Promela is a language designed to help verify protocols, but which can also be used to
verify small parallel algorithms. You recode your algorithm and correctness constraints
in the C-like language Promela, and then use Spin to translate it into a C program that
you can compile and run. The resulting program carries out a full state-space search of
your algorithm, either verifying or finding counter-examples for assertions that you can
associate with in your Promela program.
This full-state search can be extremely powerful, but can also be a two-edged sword.
If your algorithm is too complex or your Promela implementation is careless, there
might be more states than fit in memory. Furthermore, even given sufficient memory, the
state-space search might well run for longer than the expected lifetime of the universe.
Therefore, use this tool for compact but complex parallel algorithms. Attempts to
naively apply it to even moderate-scale algorithms (let alone the full Linux kernel) will
end badly.
Promela and Spin may be downloaded from https://spinroot.com/spin/
whatispin.html.
The above site also gives links to Gerard Holzmann’s excellent book [Hol03]
on Promela and Spin, as well as searchable online references starting at: https:
//www.spinroot.com/spin/Man/index.html.
The remainder of this section describes how to use Promela to debug parallel
algorithms, starting with simple examples and progressing to more complex uses.
12.1.1.1 Warm-Up: Non-Atomic Increment

Listing 12.1 demonstrates the textbook race condition resulting from non-atomic
increment. Line 1 defines the number of processes to run (we will vary this to see the
effect on state space), line 3 defines the counter, and line 4 is used to implement the
assertion that appears on lines 29–39.
Lines 6–13 define a process that increments the counter non-atomically. The argument
me is the process number, set by the initialization block later in the code. Because
simple Promela statements are each assumed atomic, we must break the increment into
the two statements on lines 10–11. The assignment on line 12 marks the process’s
completion. Because the Spin system will fully search the state space, including all
possible sequences of states, there is no need for the loop that would be used for
conventional stress testing.
Lines 15–40 are the initialization block, which is executed first. Lines 19–28 actually
do the initialization, while lines 29–39 perform the assertion. Both are atomic blocks in
order to avoid unnecessarily increasing the state space: Because they are not part of the
algorithm proper, we lose no verification coverage by making them atomic.
The do-od construct on lines 21–27 implements a Promela loop, which can be
thought of as a C for (;;) loop containing a switch statement that allows expressions
in case labels. The condition blocks (prefixed by ::) are scanned non-deterministically,
though in this case only one of the conditions can possibly hold at a given time. The
first block of the do-od from lines 22–25 initializes the i-th incrementer’s progress cell,
runs the i-th incrementer’s process, and then increments the variable i. The second
block of the do-od on line 26 exits the loop once these processes have been started.
The atomic block on lines 29–39 also contains a similar do-od loop that sums up
the progress counters. The assert() statement on line 38 verifies that if all processes
have been completed, then all counts have been correctly recorded.
v2023.06.11a
12.1. STATE-SPACE SEARCH 359
Listing 12.1: Promela Code for Non-Atomic Increment

1 #define NUMPROCS 2
2
3 byte counter = 0;
4 byte progress[NUMPROCS];
5
6 proctype incrementer(byte me)
7 {
8 int temp;
9
10 temp = counter;
11 counter = temp + 1;
12 progress[me] = 1;
13 }
14
15 init {
16 int i = 0;
17 int sum = 0;
18
19 atomic {
20 i = 0;
21 do
22 :: i < NUMPROCS ->
23 progress[i] = 0;
24 run incrementer(i);
25 i++;
26 :: i >= NUMPROCS -> break;
27 od;
28 }
29 atomic {
30 i = 0;
31 sum = 0;
32 do
33 :: i < NUMPROCS ->
34 sum = sum + progress[i];
35 i++
36 :: i >= NUMPROCS -> break;
37 od;
38 assert(sum < NUMPROCS || counter == NUMPROCS);
39 }
40 }
You can build and run this program as follows:
spin -a increment.spin # Translate the model to C

cc -DSAFETY -o pan pan.c # Compile the model
./pan # Run the model
This will produce output as shown in Listing 12.2. The first line tells us that our
assertion was violated (as expected given the non-atomic increment!). The second line
that a trail file was written describing how the assertion was violated. The “Warning”
line reiterates that all was not well with our model. The second paragraph describes the
type of state-search being carried out, in this case for assertion violations and invalid
end states. The third paragraph gives state-size statistics: This small model had only 45
states. The final line shows memory usage.
The trail file may be rendered human-readable as follows:
spin -t -p increment.spin
This gives the output shown in Listing 12.3. As can be seen, the first portion of the
init block created both incrementer processes, both of which first fetched the counter,
then both incremented and stored it, losing a count. The assertion then triggered, after
which the global state is displayed.
v2023.06.11a
Listing 12.2: Non-Atomic Increment Spin Output

pan:1: assertion violated
((sum<2)||(counter==2)) (at depth 22)
pan: wrote increment.spin.trail
(Spin Version 6.4.8 -- 2 March 2018)

Warning: Search not completed
+ Partial Order Reduction
Full statespace search for:

never claim - (none specified)
assertion violations +
cycle checks - (disabled by -DSAFETY)
invalid end states +
State-vector 48 byte, depth reached 24, errors: 1

45 states, stored
13 states, matched
58 transitions (= stored+matched)
53 atomic steps
hash conflicts: 0 (resolved)
Stats on memory usage (in Megabytes):

0.003 equivalent memory usage for states
(stored*(State-vector + overhead))
0.290 actual memory usage for states
128.000 memory used for hash table (-w24)
0.534 memory used for DFS stack (-m10000)
128.730 total actual memory usage
Table 12.1: Memory Usage of Increment Model
# incrementers # states total memory usage (MB)
1 11 128.7
2 52 128.7
3 372 128.7
4 3,496 128.9
5 40,221 131.7
6 545,720 174.0
7 8,521,446 881.9
12.1.1.2 Warm-Up: Atomic Increment
It is easy to fix this example by placing the body of the incrementer processes in an
atomic block as shown in Listing 12.4. One could also have simply replaced the pair of
statements with counter = counter + 1, because Promela statements are atomic.
Either way, running this modified model gives us an error-free traversal of the state
space, as shown in Listing 12.5.
Table 12.1 shows the number of states and memory consumed as a function of number
of incrementers modeled (by redefining NUMPROCS):
Running unnecessarily large models is thus subtly discouraged, although 882 MB is
well within the limits of modern desktop and laptop machines.
With this example under our belt, let’s take a closer look at the commands used to
analyze Promela models and then look at more elaborate examples.
v2023.06.11a
Listing 12.3: Non-Atomic Increment Error Trail

using statement merging
1: proc 0 (:init::1) increment.spin:21 (state 1) [i = 0]
2: proc 0 (:init::1) increment.spin:23 (state 2) [((i<2))]
2: proc 0 (:init::1) increment.spin:24 (state 3) [progress[i] = 0]
Starting incrementer with pid 1
3: proc 0 (:init::1) increment.spin:25 (state 4) [(run incrementer(i))]
4: proc 0 (:init::1) increment.spin:26 (state 5) [i = (i+1)]
5: proc 0 (:init::1) increment.spin:24 (state 3) [progress[i] = 0]
Starting incrementer with pid 2
6: proc 0 (:init::1) increment.spin:25 (state 4) [(run incrementer(i))]
8: proc 0 (:init::1) increment.spin:27 (state 6) [((i>=2))]
9: proc 0 (:init::1) increment.spin:22 (state 10) [break]
10: proc 2 (incrementer:1) increment.spin:11 (state 1) [temp = counter]
11: proc 1 (incrementer:1) increment.spin:11 (state 1) [temp = counter]
12: proc 2 (incrementer:1) increment.spin:12 (state 2) [counter = (temp+1)]
13: proc 2 (incrementer:1) increment.spin:13 (state 3) [progress[me] = 1]
14: proc 2 terminates
15: proc 1 (incrementer:1) increment.spin:12 (state 2) [counter = (temp+1)]
16: proc 1 (incrementer:1) increment.spin:13 (state 3) [progress[me] = 1]
17: proc 1 terminates
18: proc 0 (:init::1) increment.spin:31 (state 12) [i = 0]
18: proc 0 (:init::1) increment.spin:32 (state 13) [sum = 0]
19: proc 0 (:init::1) increment.spin:35 (state 15) [sum = (sum+progress[i])]
20: proc 0 (:init::1) increment.spin:35 (state 15) [sum = (sum+progress[i])]
21: proc 0 (:init::1) increment.spin:37 (state 17) [((i>=2))]
22: proc 0 (:init::1) increment.spin:33 (state 21) [break]
spin: increment.spin:39, Error: assertion violated
spin: text of failed assertion: assert(((sum<2)||(counter==2)))
23: proc 0 (:init::1) increment.spin:39 (state 22) [assert(((sum<2)||(counter==2)))]
spin: trail ends after 23 steps
#processes: 1
counter = 1
progress[0] = 1
progress[1] = 1
23: proc 0 (:init::1) increment.spin:41 (state 24) <valid end state>
3 processes created
Listing 12.4: Promela Code for Atomic Increment

1 proctype incrementer(byte me)
2 {
3 int temp;
4
5 atomic {
6 temp = counter;
7 counter = temp + 1;
8 }
9 progress[me] = 1;
10 }
v2023.06.11a
Listing 12.5: Atomic Increment Spin Output



52 states, stored
21 states, matched
68 atomic steps

unreached in proctype incrementer

(0 of 5 states)
unreached in init
(0 of 24 states)
12.1.2 How to Use Promela

Given a source file qrcu.spin, one can use the following commands:
spin -a qrcu.spin
Create a file pan.c that fully searches the state machine.
cc -DSAFETY [-DCOLLAPSE] [-DMA=N] -o pan pan.c

Compile the generated state-machine search. The -DSAFETY generates opti-
mizations that are appropriate if you have only assertions (and perhaps never
statements). If you have liveness, fairness, or forward-progress checks, you may
need to compile without -DSAFETY. If you leave off -DSAFETY when you could
have used it, the program will let you know.
The optimizations produced by -DSAFETY greatly speed things up, so you should
use it when you can. An example situation where you cannot use -DSAFETY is
when checking for livelocks (AKA “non-progress cycles”) via -DNP.
The optional -DCOLLAPSE generates code for a state vector compression mode.
Another optional flag -DMA=N generates code for a slow but aggressive state-space
memory compression mode.
./pan [-mN] [-wN]

This actually searches the state space. The number of states can reach into the
tens of millions with very small state machines, so you will need a machine with
large memory. For example, qrcu.spin with 3 updaters and 2 readers required
10.5 GB of memory even with the -DCOLLAPSE flag.
If you see a message from ./pan saying: “error: max search depth too
small”, you need to increase the maximum depth by a -mN option for a complete
search. The default is -m10000.
v2023.06.11a
The -wN option specifies the hashtable size. The default for full state-space search
is -w24.1
If you aren’t sure whether your machine has enough memory, run top in one
window and ./pan in another. Keep the focus on the ./pan window so that you
can quickly kill execution if need be. As soon as CPU time drops much below
100 %, kill ./pan. If you have removed focus from the window running ./pan,
you may wait a long time for the windowing system to grab enough memory to do
anything for you.
Another option to avoid memory exhaustion is the -DMEMLIM=N compiler flag.
-DMEMLIM=2000 would set the maximum of 2 GB.
Don’t forget to capture the output, especially if you are working on a remote
machine.
If your model includes forward-progress checks, you will likely need to enable
“weak fairness” via the -f command-line argument to ./pan. If your forward-
progress checks involve accept labels, you will also need the -a argument.
spin -t -p qrcu.spin
Given trail file output by a run that encountered an error, output the sequence
of steps leading to that error. The -g flag will also include the values of changed
global variables, and the -l flag will also include the values of changed local
variables.
12.1.2.1 Promela Peculiarities

Although all computer languages have underlying similarities, Promela will provide
some surprises to people used to coding in C, C++, or Java.
1. In C, “;” terminates statements. In Promela it separates them. Fortunately, more

recent versions of Spin have become much more forgiving of “extra” semicolons.
2. Promela’s looping construct, the do statement, takes conditions. This do statement

closely resembles a looping if-then-else statement.
3. In C’s switch statement, if there is no matching case, the whole statement is
skipped. In Promela’s equivalent, confusingly called if, if there is no matching
guard expression, you get an error without a recognizable corresponding error
message. So, if the error output indicates an innocent line of code, check to see if
you left out a condition from an if or do statement.
4. When creating stress tests in C, one usually races suspect operations against each
other repeatedly. In Promela, one instead sets up a single race, because Promela
will search out all the possible outcomes from that single race. Sometimes you do
need to loop in Promela, for example, if multiple operations overlap, but doing so
greatly increases the size of your state space.
5. In C, the easiest thing to do is to maintain a loop counter to track progress and
terminate the loop. In Promela, loop counters must be avoided like the plague
1 As of Spin Version 6.4.6 and 6.4.8. In the online manual of Spin dated 10 July 2011,
the default for exhaustive search mode is said to be -w19, which does not meet the actual
behavior.
v2023.06.11a
because they cause the state space to explode. On the other hand, there is no
penalty for infinite loops in Promela as long as none of the variables monotonically
increase or decrease—Promela will figure out how many passes through the loop
really matter, and automatically prune execution beyond that point.
6. In C torture-test code, it is often wise to keep per-task control variables. They are
cheap to read, and greatly aid in debugging the test code. In Promela, per-task
control variables should be used only when there is no other alternative. To see
this, consider a 5-task verification with one bit each to indicate completion. This
gives 32 states. In contrast, a simple counter would have only six states, more than
a five-fold reduction. That factor of five might not seem like a problem, at least
not until you are struggling with a verification program possessing more than 150
million states consuming more than 10 GB of memory!
7. One of the most challenging things both in C torture-test code and in Promela is
formulating good assertions. Promela also allows never claims that act like an
assertion replicated between every line of code.
8. Dividing and conquering is extremely helpful in Promela in keeping the state space
under control. Splitting a large model into two roughly equal halves will result
in the state space of each half being roughly the square root of the whole. For
example, a million-state combined model might reduce to a pair of thousand-state
models. Not only will Promela handle the two smaller models much more quickly
with much less memory, but the two smaller algorithms are easier for people to
understand.
12.1.2.2 Promela Coding Tricks

Promela was designed to analyze protocols, so using it on parallel programs is a bit
abusive. The following tricks can help you to abuse Promela safely:
1. Memory reordering. Suppose you have a pair of statements copying globals x and
y to locals r1 and r2, where ordering matters (e.g., unprotected by locks), but
where you have no memory barriers. This can be modeled in Promela as follows:
1 if
2 :: 1 -> r1 = x;
3 r2 = y
4 :: 1 -> r2 = y;
5 r1 = x
6 fi
The two branches of the if statement will be selected nondeterministically, since

they both are available. Because the full state space is searched, both choices will
eventually be made in all cases.
Of course, this trick will cause your state space to explode if used too heavily. In
addition, it requires you to anticipate possible reorderings.
2. State reduction. If you have complex assertions, evaluate them under atomic.
After all, they are not part of the algorithm. One example of a complex assertion
(to be discussed in more detail later) is as shown in Listing 12.6.
There is no reason to evaluate this assertion non-atomically, since it is not actually
part of the algorithm. Because each statement contributes to state, we can reduce
v2023.06.11a
Listing 12.6: Complex Promela Assertion

1 i = 0;
2 sum = 0;
3 do
4 :: i < N_QRCU_READERS ->
5 sum = sum + (readerstart[i] == 1 &&
6 readerprogress[i] == 1);
7 i++
8 :: i >= N_QRCU_READERS ->
9 assert(sum == 0);
10 break
11 od
Listing 12.7: Atomic Block for Complex Promela Assertion

1 atomic {
2 i = 0;
3 sum = 0;
4 do
8 i++
10 assert(sum == 0);
11 break
12 od
13 }
the number of useless states by enclosing it in an atomic block as shown in

Listing 12.7.
3. Promela does not provide functions. You must instead use C preprocessor macros.
However, you must use them carefully in order to avoid combinatorial explosion.
Now we are ready for further examples.
12.1.3 Promela Example: Locking

Since locks are generally useful, spin_lock() and spin_unlock() macros are
provided in lock.h, which may be included from multiple Promela models, as shown
in Listing 12.8. The spin_lock() macro contains an infinite do-od loop spanning
lines 2–11, courtesy of the single guard expression of “1” on line 3. The body of this
loop is a single atomic block that contains an if-fi statement. The if-fi construct is
similar to the do-od construct, except that it takes a single pass rather than looping. If
the lock is not held on line 5, then line 6 acquires it and line 7 breaks out of the enclosing
do-od loop (and also exits the atomic block). On the other hand, if the lock is already
held on line 8, we do nothing (skip), and fall out of the if-fi and the atomic block so
as to take another pass through the outer loop, repeating until the lock is available.
The spin_unlock() macro simply marks the lock as no longer held.
Note that memory barriers are not needed because Promela assumes full ordering.
In any given Promela state, all processes agree on both the current state and the order
of state changes that caused us to arrive at the current state. This is analogous to the
“sequentially consistent” memory model used by a few computer systems (such as 1990s
MIPS and PA-RISC). As noted earlier, and as will be seen in a later example, weak
memory ordering must be explicitly coded.
These macros are tested by the Promela code shown in Listing 12.9. This code is
similar to that used to test the increments, with the number of locking processes defined
v2023.06.11a
Listing 12.8: Promela Code for Spinlock

1 #define spin_lock(mutex) \
2 do \
3 :: 1 -> atomic { \
4 if \
5 :: mutex == 0 -> \
6 mutex = 1; \
7 break \
8 :: else -> skip \
9 fi \
10 } \
11 od
12
13 #define spin_unlock(mutex) \
14 mutex = 0
Listing 12.9: Promela Code to Test Spinlocks

1 #include "lock.h"
2
3 #define N_LOCKERS 3
4
5 bit mutex = 0;
6 bit havelock[N_LOCKERS];
7 int sum;
8
9 proctype locker(byte me)
10 {
11 do
12 :: 1 ->
13 spin_lock(mutex);
14 havelock[me] = 1;
15 havelock[me] = 0;
16 spin_unlock(mutex)
17 od
18 }
19
20 init {
21 int i = 0;
22 int j;
23
24 end: do
25 :: i < N_LOCKERS ->
26 havelock[i] = 0;
27 run locker(i);
28 i++
29 :: i >= N_LOCKERS ->
30 sum = 0;
31 j = 0;
32 atomic {
33 do
34 :: j < N_LOCKERS ->
35 sum = sum + havelock[j];
36 j = j + 1
37 :: j >= N_LOCKERS ->
38 break
39 od
40 }
41 assert(sum <= 1);
42 break
43 od
44 }
v2023.06.11a
Listing 12.10: Output for Spinlock Test



576 states, stored
929 states, matched
368 atomic steps

unreached in proctype locker

lock.spin:19, state 20, "-end-"
(1 of 20 states)
unreached in init
(0 of 22 states)
by the N_LOCKERS macro definition on line 3. The mutex itself is defined on line 5, an
array to track the lock owner on line 6, and line 7 is used by assertion code to verify that
only one process holds the lock.
The locker process is on lines 9–18, and simply loops forever acquiring the lock on
line 13, claiming it on line 14, unclaiming it on line 15, and releasing it on line 16.
The init block on lines 20–44 initializes the current locker’s havelock array entry on
line 26, starts the current locker on line 27, and advances to the next locker on line 28.
Once all locker processes are spawned, the do-od loop moves to line 29, which checks
the assertion. Lines 30 and 31 initialize the control variables, lines 32–40 atomically
sum the havelock array entries, line 41 is the assertion, and line 42 exits the loop.
We can run this model by placing the two code fragments of Listings 12.8 and 12.9
into files named lock.h and lock.spin, respectively, and then running the following
commands:
spin -a lock.spin
cc -DSAFETY -o pan pan.c
./pan
The output will look something like that shown in Listing 12.10. As expected, this
run has no assertion failures (“errors: 0”).
Quick Quiz 12.1: Why is there an unreached statement in locker? After all, isn’t this a full
state-space search?
Quick Quiz 12.2: What are some Promela code-style issues with this example?
v2023.06.11a
Listing 12.11: QRCU Global Variables

1 #include "lock.h"
2
3 #define N_QRCU_READERS 2
4 #define N_QRCU_UPDATERS 2
5
6 bit idx = 0;
7 byte ctr[2];
8 byte readerprogress[N_QRCU_READERS];
9 bit mutex = 0;
12.1.4 Promela Example: QRCU

This final example demonstrates a real-world use of Promela on Oleg Nesterov’s
QRCU [Nes06a, Nes06b], but modified to speed up the synchronize_qrcu() fastpath.
But first, what is QRCU?
QRCU is a variant of SRCU [McK06] that trades somewhat higher read overhead
(atomic increment and decrement on a global variable) for extremely low grace-period
latencies. If there are no readers, the grace period will be detected in less than a
microsecond, compared to the multi-millisecond grace-period latencies of most other
RCU implementations.
1. There is a qrcu_struct that defines a QRCU domain. Like SRCU (and unlike
other variants of RCU) QRCU’s action is not global, but instead focused on the
specified qrcu_struct.
2. There are qrcu_read_lock() and qrcu_read_unlock() primitives that delimit
QRCU read-side critical sections. The corresponding qrcu_struct must be
passed into these primitives, and the return value from qrcu_read_lock() must
be passed to qrcu_read_unlock().
For example:
idx = qrcu_read_lock(&my_qrcu_struct);
/* read-side critical section. */
qrcu_read_unlock(&my_qrcu_struct, idx);
3. There is a synchronize_qrcu() primitive that blocks until all pre-existing QRCU

read-side critical sections complete, but, like SRCU’s synchronize_srcu(),
QRCU’s synchronize_qrcu() need wait only for those read-side critical sections
that are using the same qrcu_struct.
For example, synchronize_qrcu(&your_qrcu_struct) would not need to
wait on the earlier QRCU read-side critical section. In contrast, synchronize_
qrcu(&my_qrcu_struct) would need to wait, since it shares the same qrcu_
struct.
A Linux-kernel patch for QRCU has been produced [McK07c], but is unlikely to ever
be included in the Linux kernel.
Returning to the Promela code for QRCU, the global variables are as shown in
Listing 12.11. This example uses locking and includes lock.h. Both the number of
readers and writers can be varied using the two #define statements, giving us not one
but two ways to create combinatorial explosion. The idx variable controls which of
the two elements of the ctr array will be used by readers, and the readerprogress
variable allows an assertion to determine when all the readers are finished (since a QRCU
v2023.06.11a
Listing 12.12: QRCU Reader Process

1 proctype qrcu_reader(byte me)
2 {
3 int myidx;
4
5 do
6 :: 1 ->
7 myidx = idx;
8 atomic {
9 if
10 :: ctr[myidx] > 0 ->
11 ctr[myidx]++;
12 break
13 :: else -> skip
14 fi
15 }
16 od;
17 readerprogress[me] = 1;
18 readerprogress[me] = 2;
19 atomic { ctr[myidx]-- }
20 }
Listing 12.13: QRCU Unordered Summation

1 #define sum_unordered \
2 atomic { \
3 do \
4 :: 1 -> \
5 sum = ctr[0]; \
6 i = 1; \
7 break \
8 :: 1 -> \
9 sum = ctr[1]; \
10 i = 0; \
11 break \
12 od; \
13 } \
14 sum = sum + ctr[i]
update cannot be permitted to complete until all pre-existing readers have completed
their QRCU read-side critical sections). The readerprogress array elements have
values as follows, indicating the state of the corresponding reader:
0: Not yet started.

1: Within QRCU read-side critical section.
2: Finished with QRCU read-side critical section.
Finally, the mutex variable is used to serialize updaters’ slowpaths.

QRCU readers are modeled by the qrcu_reader() process shown in Listing 12.12.
A do-od loop spans lines 5–16, with a single guard of “1” on line 6 that makes it an
infinite loop. Line 7 captures the current value of the global index, and lines 8–15
atomically increment it (and break from the infinite loop) if its value was non-zero
(atomic_inc_not_zero()). Line 17 marks entry into the RCU read-side critical
section, and line 18 marks exit from this critical section, both lines for the benefit of the
assert() statement that we shall encounter later. Line 19 atomically decrements the
same counter that we incremented, thereby exiting the RCU read-side critical section.
The C-preprocessor macro shown in Listing 12.13 sums the pair of counters so as
to emulate weak memory ordering. Lines 2–13 fetch one of the counters, and line 14
fetches the other of the pair and sums them. The atomic block consists of a single
do-od statement. This do-od statement (spanning lines 3–12) is unusual in that it
v2023.06.11a
contains two unconditional branches with guards on lines 4 and 8, which causes Promela
to non-deterministically choose one of the two (but again, the full state-space search
causes Promela to eventually make all possible choices in each applicable situation).
The first branch fetches the zero-th counter and sets i to 1 (so that line 14 will fetch the
first counter), while the second branch does the opposite, fetching the first counter and
setting i to 0 (so that line 14 will fetch the second counter).
Quick Quiz 12.3: Is there a more straightforward way to code the do-od statement?
With the sum_unordered macro in place, we can now proceed to the update-side
process shown in Listing 12.14. The update-side process repeats indefinitely, with
the corresponding do-od loop ranging over lines 7–57. Each pass through the loop
first snapshots the global readerprogress array into the local readerstart array on
lines 12–21. This snapshot will be used for the assertion on line 53. Line 23 invokes
sum_unordered, and then lines 24–27 re-invoke sum_unordered if the fastpath is
potentially usable.
Lines 28–40 execute the slowpath code if need be, with lines 30 and 38 acquiring and
releasing the update-side lock, lines 31–33 flipping the index, and lines 34–37 waiting
for all pre-existing readers to complete.
Lines 44–56 then compare the current values in the readerprogress array to those
collected in the readerstart array, forcing an assertion failure should any readers that
started before this update still be in progress.
Quick Quiz 12.4: Why are there atomic blocks at lines 12–21 and lines 44–56, when the
operations within those atomic blocks have no atomic implementation on any current production
microprocessor?
Quick Quiz 12.5: Is the re-summing of the counters on lines 24–27 really necessary?
All that remains is the initialization block shown in Listing 12.15. This block simply
initializes the counter pair on lines 5–6, spawns the reader processes on lines 7–14, and
spawns the updater processes on lines 15–21. This is all done within an atomic block to
reduce state space.
12.1.4.1 Running the QRCU Example
To run the QRCU example, combine the code fragments in the previous section into a
single file named qrcu.spin, and place the definitions for spin_lock() and spin_
unlock() into a file named lock.h. Then use the following commands to build and
run the QRCU model:
spin -a qrcu.spin
cc -DSAFETY [-DCOLLAPSE] -o pan pan.c
./pan [-mN]
The output shows that this model passes all of the cases shown in Table 12.2. It
would be nice to run three readers and three updaters, however, simple extrapolation
indicates that this will require about half a terabyte of memory. What to do?
It turns out that ./pan gives advice when it runs out of memory, for example, when
attempting to run three readers and three updaters:
v2023.06.11a
Listing 12.14: QRCU Updater Process

1 proctype qrcu_updater(byte me)
2 {
3 int i;
4 byte readerstart[N_QRCU_READERS];
5 int sum;
6
7 do
8 :: 1 ->
9
10 /* Snapshot reader state. */
11
12 atomic {
13 i = 0;
14 do
16 readerstart[i] = readerprogress[i];
17 i++
19 break
20 od
21 }
22
23 sum_unordered;
24 if
25 :: sum <= 1 -> sum_unordered
26 :: else -> skip
27 fi;
28 if
29 :: sum > 1 ->
30 spin_lock(mutex);
31 atomic { ctr[!idx]++ }
32 idx = !idx;
33 atomic { ctr[!idx]-- }
34 do
35 :: ctr[!idx] > 0 -> skip
36 :: ctr[!idx] == 0 -> break
37 od;
38 spin_unlock(mutex);
39 :: else -> skip
40 fi;
41
42 /* Verify reader progress. */
43
44 atomic {
45 i = 0;
46 sum = 0;
47 do
51 i++
53 assert(sum == 0);
54 break
55 od
56 }
57 od
58 }
v2023.06.11a
Listing 12.15: QRCU Initialization Process

1 init {
2 int i;
3
4 atomic {
5 ctr[idx] = 1;
6 ctr[!idx] = 0;
7 i = 0;
8 do
10 readerprogress[i] = 0;
11 run qrcu_reader(i);
12 i++
13 :: i >= N_QRCU_READERS -> break
14 od;
15 i = 0;
16 do
17 :: i < N_QRCU_UPDATERS ->
18 run qrcu_updater(i);
19 i++
20 :: i >= N_QRCU_UPDATERS -> break
21 od
22 }
23 }
Table 12.2: Memory Usage of QRCU Model
updaters readers # states depth memory (MB)a
1 1 376 95 128.7
1 2 6,177 218 128.9
1 3 99,728 385 132.6
2 1 29,399 859 129.8
2 2 1,071,181 2,352 169.6
2 3 33,866,736 12,857 1,540.8
3 1 2,749,453 53,809 236.6
3 2 186,202,860 328,014 10,483.7
a Obtained with the compiler flag -DCOLLAPSE specified.
hint: to reduce memory, recompile with

-DCOLLAPSE # good, fast compression, or
-DMA=96 # better/slower compression, or
-DHC # hash-compaction, approximation
-DBITSTATE # supertrace, approximation
Let’s try the suggested compiler flag -DMA=N, which generates code for aggressive
compression of the state space at the cost of greatly increased search overhead. The
required commands are as follows:
spin -a qrcu.spin
cc -DSAFETY -DMA=96 -O2 -o pan pan.c
./pan -m20000000
Here, the depth limit of 20,000,000 is an order of magnitude larger than the expected
depth deduced from simple extrapolation. Although this increases up-front memory
usage, it avoids wasting a long run due to incomplete search resulting from a too-tight
depth limit. This run took a little more than 3 days on a POWER9 server. The result is
shown in Listing 12.16. This Spin run completed successfully with a total memory usage
v2023.06.11a
Listing 12.16: 3 Readers 3 Updaters QRCU Spin Output with -DMA=96

(Spin Version 6.4.6 -- 2 December 2016)
+ Graph Encoding (-DMA=96)


MA stats: -DMA=84 is sufficient
Minimized Automaton: 56420520 nodes and 1.75128e+08 edges
9.6647071e+09 states, stored
9.7503813e+09 states, matched
1.9415088e+10 transitions (= stored+matched)
7.2047951e+09 atomic steps

(compression: 0.48%)
1.619 memory lost to fragmentation
unreached in proctype qrcu_reader

(0 of 18 states)
unreached in proctype qrcu_updater
qrcu.spin:102, state 82, "-end-"
(1 of 82 states)
unreached in init
(0 of 23 states)
pan: elapsed time 2.72e+05 seconds

pan: rate 35500.523 states/second
of only 6.5 GB, which is almost two orders of magnitude lower than the -DCOLLAPSE
usage of about half a terabyte.
Quick Quiz 12.6: A compression rate of 0.48 % corresponds to a 200-to-1 decrease in memory
occupied by the states! Is the state-space search really exhaustive???
For reference, Table 12.3 summarizes the Spin results with -DCOLLAPSE and -DMA=N
compiler flags. The memory usage is obtained with minimal sufficient search depths
and -DMA=N parameters shown in the table. Hashtable sizes for -DCOLLAPSE runs are
tweaked by the -wN option of ./pan to avoid using too much memory hashing small
state spaces. Hence the memory usage is smaller than what is shown in Table 12.2,
where the hashtable size starts from the default of -w24. The runtime is from a POWER9
server, which shows that -DMA=N suffers up to about an order of magnitude higher CPU
overhead than does -DCOLLAPSE, but on the other hand reduces memory overhead by
well over an order of magnitude.
So far so good. But adding a few more updaters or readers would exhaust memory,
even with -DMA=N.2 So what to do? Here are some possible approaches:
1. See whether a smaller number of readers and updaters suffice to prove the general
case.
2. Manually construct a proof of correctness.
2 Alternatively, the CPU consumption would become excessive.
v2023.06.11a
Table 12.3: QRCU Spin Result Summary
-DCOLLAPSE -DMA=N
updaters readers # states depth reached -wN memory (MB) runtime (s) N memory (MB) runtime (s)
1 1 376 95 12 0.10 0.00 40 0.29 0.00

1 2 6,177 218 12 0.39 0.01 47 0.59 0.02
1 3 99,728 385 16 4.60 0.14 54 3.04 0.45
2 1 29,399 859 16 2.30 0.03 55 0.70 0.13
2 2 1,071,181 2,352 20 49.24 1.45 62 7.77 5.76
2 3 33,866,736 12,857 24 1,540.70 62.5 69 111.66 326
3 1 2,749,453 53,809 21 125.25 4.01 70 11.41 19.5
3 2 186,202,860 328,014 28 10,482.51 390 77 222.26 2,560
3 3 9,664,707,100 2,055,621 84 5,557.02 266,000
3. Use a more capable tool.
4. Divide and conquer.
The following sections discuss each of these approaches.
12.1.4.2 How Many Readers and Updaters Are Really Needed?

One approach is to look carefully at the Promela code for qrcu_updater() and notice
that the only global state change is happening under the lock. Therefore, only one
updater at a time can possibly be modifying state visible to either readers or other
updaters. This means that any sequences of state changes can be carried out serially by
a single updater due to the fact that Promela does a full state-space search. Therefore, at
most two updaters are required: One to change state and a second to become confused.
The situation with the readers is less clear-cut, as each reader does only a single
read-side critical section then terminates. It is possible to argue that the useful number
of readers is limited, due to the fact that the fastpath must see at most a zero and a one in
the counters. This is a fruitful avenue of investigation, in fact, it leads to the full proof
of correctness described in the next section.
12.1.4.3 Alternative Approach: Proof of Correctness

An informal proof [McK07c] follows:
1. For synchronize_qrcu() to exit too early, then by definition there must have
been at least one reader present during synchronize_qrcu()’s full execution.
2. The counter corresponding to this reader will have been at least 1 during this time
interval.
3. The synchronize_qrcu() code forces at least one of the counters to be at least 1

at all times.
4. The above two items imply that if the counter corresponding to this reader is exactly
one, then the other counter must be greater than or equal to one. Similarly, if the
other counter is equal to zero, then the counter corresponding to the reader must
be greater than or equal to two.
v2023.06.11a
5. Therefore, at any given point in time, either one of the counters will be at least 2,
or both of the counters will be at least one.
6. However, the synchronize_qrcu() fastpath code can read only one of the
counters at a given time. It is therefore possible for the fastpath code to fetch the
first counter while zero, but to race with a counter flip so that the second counter is
seen as one.
7. There can be at most one reader persisting through such a race condition, as
otherwise the sum would be two or greater, which would cause the updater to take
the slowpath.
8. But if the race occurs on the fastpath’s first read of the counters, and then again on
its second read, there have to have been two counter flips.
9. Because a given updater flips the counter only once, and because the update-side
lock prevents a pair of updaters from concurrently flipping the counters, the only
way that the fastpath code can race with a flip twice is if the first updater completes.
10. But the first updater will not complete until after all pre-existing readers have
completed.
11. Therefore, if the fastpath races with a counter flip twice in succession, all pre-
existing readers must have completed, so that it is safe to take the fastpath.
Of course, not all parallel algorithms have such simple proofs. In such cases, it may
be necessary to enlist more capable tools.
12.1.4.4 Alternative Approach: More Capable Tools

Although Promela and Spin are quite useful, much more capable tools are available,
particularly for verifying hardware. This means that if it is possible to translate your
algorithm to the hardware-design VHDL language, as it often will be for low-level
parallel algorithms, then it is possible to apply these tools to your code (for example,
this was done for the first realtime RCU algorithm). However, such tools can be quite
expensive.
Although the advent of commodity multiprocessing might eventually result in powerful
free-software model-checkers featuring fancy state-space-reduction capabilities, this
does not help much in the here and now.
As an aside, there are Spin features that support approximate searches that require
fixed amounts of memory, however, I have never been able to bring myself to trust
approximations when verifying parallel algorithms.
Another approach might be to divide and conquer.
12.1.4.5 Alternative Approach: Divide and Conquer

It is often possible to break down a larger parallel algorithm into smaller pieces, which
can then be proven separately. For example, a 10-billion-state model might be broken
into a pair of 100,000-state models. Taking this approach not only makes it easier for
tools such as Promela to verify your algorithms, it can also make your algorithms easier
to understand.
v2023.06.11a
12.1.4.6 Is QRCU Really Correct?
Is QRCU really correct? We have a Promela-based mechanical proof and a by-hand

proof that both say that it is. However, a paper by Alglave et al. [AKT13] says otherwise
(see Section 5.1 of the paper at the bottom of page 12). Which is it?
It turns out that both are correct! When QRCU was added to a suite of formal-
verification benchmarks, its memory barriers were omitted, thus resulting in a buggy
version of QRCU. So the real news here is that a number of formal-verification tools
incorrectly proved this buggy QRCU correct. And this is why formal-verification tools
themselves should be tested using bug-injected versions of the code being verified. If a
given tool cannot find the injected bugs, then that tool is clearly untrustworthy.
Quick Quiz 12.7: But different formal-verification tools are often designed to locate particular
classes of bugs. For example, very few formal-verification tools will find an error in the
specification. So isn’t this “clearly untrustworthy” judgment a bit harsh?
Therefore, if you do intend to use QRCU, please take care. Its proofs of correctness
might or might not themselves be correct. Which is one reason why formal verification
is unlikely to completely replace testing, as Donald Knuth pointed out so long ago.
Quick Quiz 12.8: Given that we have two independent proofs of correctness for the QRCU
algorithm described herein, and given that the proof of incorrectness covers what is known to
be a different algorithm, why is there any room for doubt?
12.1.5 Promela Parable: dynticks and Preemptible RCU

In early 2008, a preemptible variant of RCU was accepted into mainline Linux in
support of real-time workloads, a variant similar to the RCU implementations in the -rt
patchset [Mol05] since August 2005. Preemptible RCU is needed for real-time workloads
because older RCU implementations disable preemption across RCU read-side critical
sections, resulting in excessive real-time latencies.
However, one disadvantage of the older -rt implementation was that each grace
period requires work to be done on each CPU, even if that CPU is in a low-power
“dynticks-idle” state, and thus incapable of executing RCU read-side critical sections.
The idea behind the dynticks-idle state is that idle CPUs should be physically powered
down in order to conserve energy. In short, preemptible RCU can disable a valuable
energy-conservation feature of recent Linux kernels. Although Josh Triplett and Paul
McKenney had discussed some approaches for allowing CPUs to remain in low-power
state throughout an RCU grace period (thus preserving the Linux kernel’s ability to
conserve energy), matters did not come to a head until Steve Rostedt integrated a new
dyntick implementation with preemptible RCU in the -rt patchset.
This combination caused one of Steve’s systems to hang on boot, so in October,
Paul coded up a dynticks-friendly modification to preemptible RCU’s grace-period
processing. Steve coded up rcu_irq_enter() and rcu_irq_exit() interfaces called
from the irq_enter() and irq_exit() interrupt entry/exit functions. These rcu_
irq_enter() and rcu_irq_exit() functions are needed to allow RCU to reliably
handle situations where a dynticks-idle CPU is momentarily powered up for an interrupt
handler containing RCU read-side critical sections. With these changes in place, Steve’s
system booted reliably, but Paul continued inspecting the code periodically on the
assumption that we could not possibly have gotten the code right on the first try.
v2023.06.11a
Paul reviewed the code repeatedly from October 2007 to February 2008, and almost
always found at least one bug. In one case, Paul even coded and tested a fix before
realizing that the bug was illusory, and in fact in all cases, the “bug” turned out to be
illusory.
Near the end of February, Paul grew tired of this game. He therefore decided to enlist
the aid of Promela and Spin. The following presents a series of seven increasingly
realistic Promela models, the last of which passes, consuming about 40 GB of main
memory for the state space.
More important, Promela and Spin did find a very subtle bug for me!
Quick Quiz 12.9: Yeah, that’s just great! Now, just what am I supposed to do if I don’t happen
to have a machine with 40 GB of main memory???
Still better would be to come up with a simpler and faster algorithm that has a smaller
state space. Even better would be an algorithm so simple that its correctness was
obvious to the casual observer!
Sections 12.1.5.1–12.1.5.4 give an overview of preemptible RCU’s dynticks interface,
followed by Section 12.1.6’s discussion of the validation of the interface.
12.1.5.1 Introduction to Preemptible RCU and dynticks

The per-CPU dynticks_progress_counter variable is central to the interface be-
tween dynticks and preemptible RCU. This variable has an even value whenever the
corresponding CPU is in dynticks-idle mode, and an odd value otherwise. A CPU exits
dynticks-idle mode for the following three reasons:
1. To start running a task,
2. When entering the outermost of a possibly nested set of interrupt handlers, and
3. When entering an NMI handler.
Preemptible RCU’s grace-period machinery samples the value of the dynticks_

progress_counter variable in order to determine when a dynticks-idle CPU may
safely be ignored.
The following three sections give an overview of the task interface, the interrupt/NMI
interface, and the use of the dynticks_progress_counter variable by the grace-
period machinery as of Linux kernel v2.6.25-rc4.
12.1.5.2 Task Interface

When a given CPU enters dynticks-idle mode because it has no more tasks to run, it
invokes rcu_enter_nohz():
1 static inline void rcu_enter_nohz(void)

2 {
3 mb();
4 __get_cpu_var(dynticks_progress_counter)++;
5 WARN_ON(__get_cpu_var(dynticks_progress_counter) &
6 0x1);
7 }
This function simply increments dynticks_progress_counter and checks that

the result is even, but first executing a memory barrier to ensure that any other CPU that
v2023.06.11a
sees the new value of dynticks_progress_counter will also see the completion of
any prior RCU read-side critical sections.
Similarly, when a CPU that is in dynticks-idle mode prepares to start executing a
newly runnable task, it invokes rcu_exit_nohz():
1 static inline void rcu_exit_nohz(void)
2 {
3 __get_cpu_var(dynticks_progress_counter)++;
4 mb();
5 WARN_ON(!(__get_cpu_var(dynticks_progress_counter) &
6 0x1));
7 }
This function again increments dynticks_progress_counter, but follows it with

a memory barrier to ensure that if any other CPU sees the result of any subsequent RCU
read-side critical section, then that other CPU will also see the incremented value of
dynticks_progress_counter. Finally, rcu_exit_nohz() checks that the result of
the increment is an odd value.
The rcu_enter_nohz() and rcu_exit_nohz() functions handle the case where
a CPU enters and exits dynticks-idle mode due to task execution, but does not handle
interrupts, which are covered in the following section.
12.1.5.3 Interrupt Interface

The rcu_irq_enter() and rcu_irq_exit() functions handle interrupt/NMI entry
and exit, respectively. Of course, nested interrupts must also be properly accounted
for. The possibility of nested interrupts is handled by a second per-CPU variable,
rcu_update_flag, which is incremented upon entry to an interrupt or NMI handler (in
rcu_irq_enter()) and is decremented upon exit (in rcu_irq_exit()). In addition,
the pre-existing in_interrupt() primitive is used to distinguish between an outermost
or a nested interrupt/NMI.
Interrupt entry is handled by the rcu_irq_enter() shown below:
1 void rcu_irq_enter(void)
2 {
3 int cpu = smp_processor_id();
4
5 if (per_cpu(rcu_update_flag, cpu))
6 per_cpu(rcu_update_flag, cpu)++;
7 if (!in_interrupt() &&
8 (per_cpu(dynticks_progress_counter,
9 cpu) & 0x1) == 0) {
10 per_cpu(dynticks_progress_counter, cpu)++;
11 smp_mb();
12 per_cpu(rcu_update_flag, cpu)++;
13 }
14 }
Line 3 fetches the current CPU’s number, while lines 5 and 6 increment the rcu_
update_flag nesting counter if it is already non-zero. Lines 7–9 check to see whether
we are the outermost level of interrupt, and, if so, whether dynticks_progress_
counter needs to be incremented. If so, line 10 increments dynticks_progress_
counter, line 11 executes a memory barrier, and line 12 increments rcu_update_flag.
As with rcu_exit_nohz(), the memory barrier ensures that any other CPU that sees
the effects of an RCU read-side critical section in the interrupt handler (following the
rcu_irq_enter() invocation) will also see the increment of dynticks_progress_
counter.
v2023.06.11a
Quick Quiz 12.10: Why not simply increment rcu_update_flag, and then only increment
dynticks_progress_counter if the old value of rcu_update_flag was zero???
Quick Quiz 12.11: But if line 7 finds that we are the outermost interrupt, wouldn’t we always
need to increment dynticks_progress_counter?
Interrupt exit is handled similarly by rcu_irq_exit():

1 void rcu_irq_exit(void)
2 {
3 int cpu = smp_processor_id();
4
5 if (per_cpu(rcu_update_flag, cpu)) {
6 if (--per_cpu(rcu_update_flag, cpu))
7 return;
8 WARN_ON(in_interrupt());
9 smp_mb();
10 per_cpu(dynticks_progress_counter, cpu)++;
11 WARN_ON(per_cpu(dynticks_progress_counter,
12 cpu) & 0x1);
13 }
14 }
Line 3 fetches the current CPU’s number, as before. Line 5 checks to see if the
rcu_update_flag is non-zero, returning immediately (via falling off the end of the
function) if not. Otherwise, lines 6 through 12 come into play. Line 6 decrements
rcu_update_flag, returning if the result is not zero. Line 8 verifies that we are indeed
leaving the outermost level of nested interrupts, line 9 executes a memory barrier,
line 10 increments dynticks_progress_counter, and lines 11 and 12 verify that this
variable is now even. As with rcu_enter_nohz(), the memory barrier ensures that
any other CPU that sees the increment of dynticks_progress_counter will also see
the effects of an RCU read-side critical section in the interrupt handler (preceding the
rcu_irq_exit() invocation).
These two sections have described how the dynticks_progress_counter variable
is maintained during entry to and exit from dynticks-idle mode, both by tasks and by
interrupts and NMIs. The following section describes how this variable is used by
preemptible RCU’s grace-period machinery.
12.1.5.4 Grace-Period Interface

Of the four preemptible RCU grace-period states shown in Figure 12.1, only the rcu_
try_flip_waitack_state and rcu_try_flip_waitmb_state states need to wait
for other CPUs to respond.
Of course, if a given CPU is in dynticks-idle state, we shouldn’t wait for it. Therefore,
just before entering one of these two states, the preceding state takes a snapshot of each
CPU’s dynticks_progress_counter variable, placing the snapshot in another per-
CPU variable, rcu_dyntick_snapshot. This is accomplished by invoking dyntick_
save_progress_counter(), shown below:
1 static void dyntick_save_progress_counter(int cpu)
2 {
3 per_cpu(rcu_dyntick_snapshot, cpu) =
4 per_cpu(dynticks_progress_counter, cpu);
5 }
The rcu_try_flip_waitack_state state invokes rcu_try_flip_waitack_

needed(), shown below:
v2023.06.11a
rcu_try_flip_idle_state
Still no activity
(No RCU activity)
Increment grace−period counter

Request counter−flip acknowledgement
rcu_try_flip_waitack_state
(Wait for acknowledgements)
Memory barrier
rcu_try_flip_waitzero_state
(Wait for RCU read−side
critical sections to complete)
Request memory barriers
rcu_try_flip_waitmb_state
(Wait for memory barriers)
Figure 12.1: Preemptible RCU State Machine
1 static inline int

2 rcu_try_flip_waitack_needed(int cpu)
3 {
4 long curr;
5 long snap;
6
7 curr = per_cpu(dynticks_progress_counter, cpu);
8 snap = per_cpu(rcu_dyntick_snapshot, cpu);
9 smp_mb();
10 if ((curr == snap) && ((curr & 0x1) == 0))
11 return 0;
12 if ((curr - snap) > 2 || (snap & 0x1) == 0)
13 return 0;
14 return 1;
15 }
Lines 7 and 8 pick up current and snapshot versions of dynticks_progress_

counter, respectively. The memory barrier on line 9 ensures that the counter checks
in the later rcu_try_flip_waitzero_state follow the fetches of these counters.
Lines 10 and 11 return zero (meaning no communication with the specified CPU is
required) if that CPU has remained in dynticks-idle state since the time that the snapshot
was taken. Similarly, lines 12 and 13 return zero if that CPU was initially in dynticks-idle
state or if it has completely passed through a dynticks-idle state. In both these cases,
there is no way that the CPU could have retained the old value of the grace-period
counter. If neither of these conditions hold, line 14 returns one, meaning that the CPU
needs to explicitly respond.
For its part, the rcu_try_flip_waitmb_state state invokes rcu_try_flip_
waitmb_needed(), shown below:
1 static inline int

2 rcu_try_flip_waitmb_needed(int cpu)
3 {
v2023.06.11a
4 long curr;
5 long snap;
6
9 smp_mb();
11 return 0;
12 if (curr != snap)
13 return 0;
14 return 1;
15 }
This is quite similar to rcu_try_flip_waitack_needed(), the difference being

in lines 12 and 13, because any transition either to or from dynticks-idle state executes
the memory barrier needed by the rcu_try_flip_waitmb_state state.
We now have seen all the code involved in the interface between RCU and the
dynticks-idle state. The next section builds up the Promela model used to verify this
code.
Quick Quiz 12.12: Can you spot any bugs in any of the code in this section?
12.1.6 Validating Preemptible RCU and dynticks

This section develops a Promela model for the interface between dynticks and RCU step
by step, with each of Sections 12.1.6.1–12.1.6.7 illustrating one step, starting with the
process-level code, adding assertions, interrupts, and finally NMIs.
Section 12.1.6.8 lists lessons (re)learned during this effort, and Sections 12.1.6.9–
12.1.6.15 present a simpler solution to RCU’s dynticks problem.
12.1.6.1 Basic Model

This section translates the process-level dynticks entry/exit code and the grace-period
processing into Promela [Hol03]. We start with rcu_exit_nohz() and rcu_enter_
nohz() from the 2.6.25-rc4 kernel, placing these in a single Promela process that
models exiting and entering dynticks-idle mode in a loop as follows:
1 proctype dyntick_nohz()
2 {
3 byte tmp;
4 byte i = 0;
5
6 do
7 :: i >= MAX_DYNTICK_LOOP_NOHZ -> break;
8 :: i < MAX_DYNTICK_LOOP_NOHZ ->
9 tmp = dynticks_progress_counter;
10 atomic {
11 dynticks_progress_counter = tmp + 1;
12 assert((dynticks_progress_counter & 1) == 1);
13 }
15 atomic {
18 }
19 i++;
20 od;
21 }
v2023.06.11a
Lines 6 and 20 define a loop. Line 7 exits the loop once the loop counter i has
exceeded the limit MAX_DYNTICK_LOOP_NOHZ. Line 8 tells the loop construct to execute
lines 9–19 for each pass through the loop. Because the conditionals on lines 7 and 8
are exclusive of each other, the normal Promela random selection of true conditions
is disabled. Lines 9 and 11 model rcu_exit_nohz()’s non-atomic increment of
dynticks_progress_counter, while line 12 models the WARN_ON(). The atomic
construct simply reduces the Promela state space, given that the WARN_ON() is not
strictly speaking part of the algorithm. Lines 14–18 similarly model the increment and
WARN_ON() for rcu_enter_nohz(). Finally, line 19 increments the loop counter.
Each pass through the loop therefore models a CPU exiting dynticks-idle mode (for
example, starting to execute a task), then re-entering dynticks-idle mode (for example,
that same task blocking).
Quick Quiz 12.13: Why isn’t the memory barrier in rcu_exit_nohz() and rcu_enter_
nohz() modeled in Promela?
Quick Quiz 12.14: Isn’t it a bit strange to model rcu_exit_nohz() followed by rcu_enter_
nohz()? Wouldn’t it be more natural to instead model entry before exit?
The next step is to model the interface to RCU’s grace-period processing. For this, we
need to model dyntick_save_progress_counter(), rcu_try_flip_waitack_
needed(), rcu_try_flip_waitmb_needed(), as well as portions of rcu_try_
flip_waitack() and rcu_try_flip_waitmb(), all from the 2.6.25-rc4 kernel. The
following grace_period() Promela process models these functions as they would be
invoked during a single pass through preemptible RCU’s grace-period processing.
1 proctype grace_period()
2 {
3 byte curr;
4 byte snap;
5
6 atomic {
7 printf("MDLN = %d\n", MAX_DYNTICK_LOOP_NOHZ);
8 snap = dynticks_progress_counter;
9 }
10 do
11 :: 1 ->
12 atomic {
13 curr = dynticks_progress_counter;
14 if
15 :: (curr == snap) && ((curr & 1) == 0) ->
16 break;
17 :: (curr - snap) > 2 || (snap & 1) == 0 ->
18 break;
19 :: 1 -> skip;
20 fi;
21 }
22 od;
24 do
25 :: 1 ->
26 atomic {
28 if
29 :: (curr == snap) && ((curr & 1) == 0) ->
30 break;
31 :: (curr != snap) ->
32 break;
33 :: 1 -> skip;
34 fi;
35 }
36 od;
37 }
v2023.06.11a
Lines 6–9 print out the loop limit (but only into the “.trail” file in case of error) and
models a line of code from rcu_try_flip_idle() and its call to dyntick_save_
progress_counter(), which takes a snapshot of the current CPU’s dynticks_
progress_counter variable. These two lines are executed atomically to reduce state
space.
Lines 10–22 model the relevant code in rcu_try_flip_waitack() and its call to
rcu_try_flip_waitack_needed(). This loop is modeling the grace-period state
machine waiting for a counter-flip acknowledgement from each CPU, but only that part
that interacts with dynticks-idle CPUs.
Line 23 models a line from rcu_try_flip_waitzero() and its call to dyntick_
save_progress_counter(), again taking a snapshot of the CPU’s dynticks_
progress_counter variable.
Finally, lines 24–36 model the relevant code in rcu_try_flip_waitack() and its
call to rcu_try_flip_waitack_needed(). This loop is modeling the grace-period
state-machine waiting for each CPU to execute a memory barrier, but again only that
part that interacts with dynticks-idle CPUs.
Quick Quiz 12.15: Wait a minute! In the Linux kernel, both dynticks_progress_counter
and rcu_dyntick_snapshot are per-CPU variables. So why are they instead being modeled
as single global variables?
The resulting model (dyntickRCU-base.spin), when run with the runspin.sh

script, generates 691 states and passes without errors, which is not at all surprising
given that it completely lacks the assertions that could find failures. The next section
therefore adds safety assertions.
12.1.6.2 Validating Safety

A safe RCU implementation must never permit a grace period to complete before the
completion of any RCU readers that started before the start of the grace period. This is
modeled by a grace_period_state variable that can take on three states as follows:
1 #define GP_IDLE 0
2 #define GP_WAITING 1
3 #define GP_DONE 2
4 byte grace_period_state = GP_DONE;
The grace_period() process sets this variable as it progresses through the grace-
period phases, as shown below:
2 {
3 byte curr;
4 byte snap;
5
6 grace_period_state = GP_IDLE;
7 atomic {
10 grace_period_state = GP_WAITING;
11 }
12 do
13 :: 1 ->
14 atomic {
16 if
17 :: (curr == snap) && ((curr & 1) == 0) ->
18 break;
v2023.06.11a
19 :: (curr - snap) > 2 || (snap & 1) == 0 ->

20 break;
21 :: 1 -> skip;
22 fi;
23 }
24 od;
25 grace_period_state = GP_DONE;
27 atomic {
30 }
31 do
32 :: 1 ->
33 atomic {
35 if
36 :: (curr == snap) && ((curr & 1) == 0) ->
37 break;
38 :: (curr != snap) ->
39 break;
40 :: 1 -> skip;
41 fi;
42 }
43 od;
45 }
Lines 6, 10, 25, 26, 29, and 44 update this variable (combining atomically with
algorithmic operations where feasible) to allow the dyntick_nohz() process to verify
the basic RCU safety property. The form of this verification is to assert that the value of
the grace_period_state variable cannot jump from GP_IDLE to GP_DONE during a
time period over which RCU readers could plausibly persist.
Quick Quiz 12.16: Given there are a pair of back-to-back changes to grace_period_state
on lines 25 and 26, how can we be sure that line 25’s changes won’t be lost?
The dyntick_nohz() Promela process implements this verification as shown below:
2 {
3 byte tmp;
4 byte i = 0;
5 bit old_gp_idle;
6
7 do
11 atomic {
13 old_gp_idle = (grace_period_state == GP_IDLE);
15 }
16 atomic {
18 assert(!old_gp_idle ||
19 grace_period_state != GP_DONE);
20 }
21 atomic {
24 }
25 i++;
26 od;
27 }
v2023.06.11a
Line 13 sets a new old_gp_idle flag if the value of the grace_period_state

variable is GP_IDLE at the beginning of task execution, and the assertion at lines 18
and 19 fire if the grace_period_state variable has advanced to GP_DONE during task
execution, which would be illegal given that a single RCU read-side critical section
could span the entire intervening time period.
The resulting model (dyntickRCU-base-s.spin), when run with the runspin.sh
script, generates 964 states and passes without errors, which is reassuring. That said,
although safety is critically important, it is also quite important to avoid indefinitely
stalling grace periods. The next section therefore covers verifying liveness.
12.1.6.3 Validating Liveness

Although liveness can be difficult to prove, there is a simple trick that applies here. The
first step is to make dyntick_nohz() indicate that it is done via a dyntick_nohz_done
variable, as shown on line 27 of the following:
2 {
3 byte tmp;
4 byte i = 0;
5 bit old_gp_idle;
6
7 do
11 atomic {
15 }
16 atomic {
20 }
21 atomic {
24 }
25 i++;
26 od;
27 dyntick_nohz_done = 1;
28 }
With this variable in place, we can add assertions to grace_period() to check for
unnecessary blockage as follows:
2 {
3 byte curr;
4 byte snap;
5 bit shouldexit;
6
8 atomic {
10 shouldexit = 0;
13 }
14 do
15 :: 1 ->
16 atomic {
v2023.06.11a
17 assert(!shouldexit);
18 shouldexit = dyntick_nohz_done;
20 if
21 :: (curr == snap) && ((curr & 1) == 0) ->
22 break;
23 :: (curr - snap) > 2 || (snap & 1) == 0 ->
24 break;
25 :: else -> skip;
26 fi;
27 }
28 od;
31 atomic {
32 shouldexit = 0;
35 }
36 do
37 :: 1 ->
38 atomic {
40 shouldexit = dyntick_nohz_done;
42 if
43 :: (curr == snap) && ((curr & 1) == 0) ->
44 break;
45 :: (curr != snap) ->
46 break;
47 :: else -> skip;
48 fi;
49 }
50 od;
52 }
We have added the shouldexit variable on line 5, which we initialize to zero on

line 10. Line 17 asserts that shouldexit is not set, while line 18 sets shouldexit
to the dyntick_nohz_done variable maintained by dyntick_nohz(). This assertion
will therefore trigger if we attempt to take more than one pass through the wait-for-
counter-flip-acknowledgement loop after dyntick_nohz() has completed execution.
After all, if dyntick_nohz() is done, then there cannot be any more state changes to
force us out of the loop, so going through twice in this state means an infinite loop,
which in turn means no end to the grace period.
Lines 32, 39, and 40 operate in a similar manner for the second (memory-barrier)
loop.
However, running this model (dyntickRCU-base-sl-busted.spin) results in
failure, as line 23 is checking that the wrong variable is even. Upon failure, spin writes
out a “trail” file (dyntickRCU-base-sl-busted.spin.trail), which records the
sequence of states that lead to the failure. Use the “spin -t -p -g -l dyntickRCU-
base-sl-busted.spin” command to cause spin to retrace this sequence of states,
printing the statements executed and the values of variables (dyntickRCU-base-sl-
busted.spin.trail.txt). Note that the line numbers do not match the listing above
due to the fact that spin takes both functions in a single file. However, the line numbers
do match the full model (dyntickRCU-base-sl-busted.spin).
We see that the dyntick_nohz() process completed at step 34 (search for “34:”),
but that the grace_period() process nonetheless failed to exit the loop. The value of
curr is 6 (see step 35) and that the value of snap is 5 (see step 17). Therefore the first
condition on line 21 above does not hold because “curr != snap”, and the second
v2023.06.11a
condition on line 23 does not hold either because snap is odd and because curr is only
one greater than snap.
So one of these two conditions has to be incorrect. Referring to the comment block
in rcu_try_flip_waitack_needed() for the first condition:
If the CPU remained in dynticks mode for the entire time and didn’t take any
interrupts, NMIs, SMIs, or whatever, then it cannot be in the middle of an
rcu_read_lock(), so the next rcu_read_lock() it executes must use the
new value of the counter. So we can safely pretend that this CPU already
acknowledged the counter.
The first condition does match this, because if “curr == snap” and if curr is even,
then the corresponding CPU has been in dynticks-idle mode the entire time, as required.
So let’s look at the comment block for the second condition:
If the CPU passed through or entered a dynticks idle phase with no active
irq handlers, then, as above, we can safely pretend that this CPU already
acknowledged the counter.
The first part of the condition is correct, because if curr and snap differ by two,
there will be at least one even number in between, corresponding to having passed
completely through a dynticks-idle phase. However, the second part of the condition
corresponds to having started in dynticks-idle mode, not having finished in this mode.
We therefore need to be testing curr rather than snap for being an even number.
The corrected C code is as follows:
1 static inline int

3 {
4 long curr;
5 long snap;
6
9 smp_mb();
11 return 0;
12 if ((curr - snap) > 2 || (curr & 0x1) == 0)
13 return 0;
14 return 1;
15 }
Lines 10–13 can now be combined and simplified, resulting in the following. A
similar simplification can be applied to rcu_try_flip_waitmb_needed().
1 static inline int

3 {
4 long curr;
5 long snap;
6
9 smp_mb();
10 if ((curr - snap) >= 2 || (curr & 0x1) == 0)
11 return 0;
12 return 1;
13 }
v2023.06.11a
Making the corresponding correction in the model (dyntickRCU-base-sl.spin)

results in a correct verification with 661 states that passes without errors. However,
it is worth noting that the first version of the liveness verification failed to catch this
bug, due to a bug in the liveness verification itself. This liveness-verification bug was
located by inserting an infinite loop in the grace_period() process, and noting that
the liveness-verification code failed to detect this problem!
We have now successfully verified both safety and liveness conditions, but only for
processes running and blocking. We also need to handle interrupts, a task taken up in
the next section.
12.1.6.4 Interrupts
There are a couple of ways to model interrupts in Promela:
1. Using C-preprocessor tricks to insert the interrupt handler between each and every
statement of the dynticks_nohz() process, or
2. Modeling the interrupt handler with a separate process.
A bit of thought indicated that the second approach would have a smaller state space,
though it requires that the interrupt handler somehow run atomically with respect to the
dynticks_nohz() process, but not with respect to the grace_period() process.
Fortunately, it turns out that Promela permits you to branch out of atomic statements.
This trick allows us to have the interrupt handler set a flag, and recode dynticks_
nohz() to atomically check this flag and execute only when the flag is not set. This
can be accomplished with a C-preprocessor macro that takes a label and a Promela
statement as follows:
1 #define EXECUTE_MAINLINE(label, stmt) \
2 label: skip; \
3 atomic { \
4 if \
5 :: in_dyntick_irq -> goto label; \
6 :: else -> stmt; \
7 fi; \
8 }
One might use this macro as follows:
EXECUTE_MAINLINE(stmt1,
tmp = dynticks_progress_counter)
Line 2 of the macro creates the specified statement label. Lines 3–8 are an atomic
block that tests the in_dyntick_irq variable, and if this variable is set (indicating
that the interrupt handler is active), branches out of the atomic block back to the label.
Otherwise, line 6 executes the specified statement. The overall effect is that mainline
execution stalls any time an interrupt is active, as required.
12.1.6.5 Validating Interrupt Handlers

The first step is to convert dyntick_nohz() to EXECUTE_MAINLINE() form, as follows:
2 {
3 byte tmp;
v2023.06.11a
4 byte i = 0;
5 bit old_gp_idle;
6
7 do
10 EXECUTE_MAINLINE(stmt1,
11 tmp = dynticks_progress_counter)
15 assert((dynticks_progress_counter & 1) == 1))
19 grace_period_state != GP_DONE))
22 assert((dynticks_progress_counter & 1) == 0))
23 i++;
24 od;
25 dyntick_nohz_done = 1;
26 }
It is important to note that when a group of statements is passed to EXECUTE_

MAINLINE(), as in lines 12–15, all statements in that group execute atomically.
Quick Quiz 12.17: But what would you do if you needed the statements in a single
EXECUTE_MAINLINE() group to execute non-atomically?
Quick Quiz 12.18: But what if the dynticks_nohz() process had “if” or “do” statements
with conditions, where the statement bodies of these constructs needed to execute non-atomically?
The next step is to write a dyntick_irq() process to model an interrupt handler:
1 proctype dyntick_irq()
2 {
3 byte tmp;
4 byte i = 0;
5 bit old_gp_idle;
6
7 do
8 :: i >= MAX_DYNTICK_LOOP_IRQ -> break;
9 :: i < MAX_DYNTICK_LOOP_IRQ ->
10 in_dyntick_irq = 1;
11 if
12 :: rcu_update_flag > 0 ->
13 tmp = rcu_update_flag;
14 rcu_update_flag = tmp + 1;
15 :: else -> skip;
16 fi;
17 if
18 :: !in_interrupt &&
19 (dynticks_progress_counter & 1) == 0 ->
24 :: else -> skip;
25 fi;
26 tmp = in_interrupt;
27 in_interrupt = tmp + 1;
32 in_interrupt = tmp - 1;
v2023.06.11a
33 if
34 :: rcu_update_flag != 0 ->
36 rcu_update_flag = tmp - 1;
37 if
38 :: rcu_update_flag == 0 ->
41 :: else -> skip;
42 fi;
43 :: else -> skip;
44 fi;
45 atomic {
47 i++;
48 }
49 od;
50 dyntick_irq_done = 1;
51 }
The loop from lines 7–49 models up to MAX_DYNTICK_LOOP_IRQ interrupts, with

lines 8 and 9 forming the loop condition and line 47 incrementing the control variable.
Line 10 tells dyntick_nohz() that an interrupt handler is running, and line 46
tells dyntick_nohz() that this handler has completed. Line 50 is used for liveness
verification, just like the corresponding line of dyntick_nohz().
Quick Quiz 12.19: Why are lines 46 and 47 (the “in_dyntick_irq = 0;” and the “i++;”)
executed atomically?
Lines 11–25 model rcu_irq_enter(), and lines 26 and 27 model the relevant
snippet of __irq_enter(). Lines 28–30 verify safety in much the same manner as
do the corresponding lines of dynticks_nohz(). Lines 31 and 32 model the relevant
snippet of __irq_exit(), and finally lines 33–44 model rcu_irq_exit().
Quick Quiz 12.20: What property of interrupts is this dynticks_irq() process unable to
model?
The grace_period() process then becomes as follows:
2 {
3 byte curr;
4 byte snap;
5 bit shouldexit;
6
8 atomic {
10 printf("MDLI = %d\n", MAX_DYNTICK_LOOP_IRQ);
11 shouldexit = 0;
14 }
15 do
16 :: 1 ->
17 atomic {
19 shouldexit = dyntick_nohz_done && dyntick_irq_done;
21 if
22 :: (curr - snap) >= 2 || (curr & 1) == 0 ->
23 break;
24 :: else -> skip;
25 fi;
26 }
27 od;
v2023.06.11a
30 atomic {
31 shouldexit = 0;
34 }
35 do
36 :: 1 ->
37 atomic {
39 shouldexit = dyntick_nohz_done && dyntick_irq_done;
41 if
42 :: (curr != snap) || ((curr & 1) == 0) ->
43 break;
44 :: else -> skip;
45 fi;
46 }
47 od;
49 }
The implementation of grace_period() is very similar to the earlier one. The only
changes are the addition of line 10 to add the new interrupt-count parameter, changes
to lines 19 and 39 to add the new dyntick_irq_done variable to the liveness checks,
and of course the optimizations on lines 22 and 42.
This model (dyntickRCU-irqnn-ssl.spin) results in a correct verification with
roughly half a million states, passing without errors. However, this version of the model
does not handle nested interrupts. This topic is taken up in the next section.
12.1.6.6 Validating Nested Interrupt Handlers

Nested interrupt handlers may be modeled by splitting the body of the loop in dyntick_
irq() as follows:
2 {
3 byte tmp;
4 byte i = 0;
5 byte j = 0;
6 bit old_gp_idle;
7 bit outermost;
8
9 do
10 :: i >= MAX_DYNTICK_LOOP_IRQ &&
11 j >= MAX_DYNTICK_LOOP_IRQ -> break;
13 atomic {
14 outermost = (in_dyntick_irq == 0);
16 }
17 if
21 :: else -> skip;
22 fi;
23 if
30 :: else -> skip;
31 fi;
v2023.06.11a
34 atomic {
35 if
36 :: outermost ->
38 :: else -> skip;
39 fi;
40 }
41 i++;
42 :: j < i ->
43 atomic {
44 if
45 :: j + 1 == i ->
48 :: else -> skip;
49 fi;
50 }
53 if
57 if
61 :: else -> skip;
62 fi;
63 :: else -> skip;
64 fi;
65 atomic {
66 j++;
67 in_dyntick_irq = (i != j);
68 }
69 od;
71 }
This is similar to the earlier dynticks_irq() process. It adds a second counter

variable j on line 5, so that i counts entries to interrupt handlers and j counts exits.
The outermost variable on line 7 helps determine when the grace_period_state
variable needs to be sampled for the safety checks. The loop-exit check on lines 10
and 11 is updated to require that the specified number of interrupt handlers are exited
as well as entered, and the increment of i is moved to line 41, which is the end of the
interrupt-entry model. Lines 13–16 set the outermost variable to indicate whether this
is the outermost of a set of nested interrupts and to set the in_dyntick_irq variable
that is used by the dyntick_nohz() process. Lines 34–40 capture the state of the
grace_period_state variable, but only when in the outermost interrupt handler.
Line 42 has the do-loop conditional for interrupt-exit modeling: As long as we
have exited fewer interrupts than we have entered, it is legal to exit another interrupt.
Lines 43–50 check the safety criterion, but only if we are exiting from the outermost
interrupt level. Finally, lines 65–68 increment the interrupt-exit count j and, if this is
the outermost interrupt level, clears in_dyntick_irq.
This model (dyntickRCU-irq-ssl.spin) results in a correct verification with a
bit more than half a million states, passing without errors. However, this version of the
model does not handle NMIs, which are taken up in the next section.
v2023.06.11a
12.1.6.7 Validating NMI Handlers

We take the same general approach for NMIs as we do for interrupts, keeping in mind
that NMIs do not nest. This results in a dyntick_nmi() process as follows:
1 proctype dyntick_nmi()
2 {
3 byte tmp;
4 byte i = 0;
5 bit old_gp_idle;
6
7 do
8 :: i >= MAX_DYNTICK_LOOP_NMI -> break;
9 :: i < MAX_DYNTICK_LOOP_NMI ->
10 in_dyntick_nmi = 1;
11 if
15 :: else -> skip;
16 fi;
17 if
24 :: else -> skip;
25 fi;
33 if
37 if
41 :: else -> skip;
42 fi;
43 :: else -> skip;
44 fi;
45 atomic {
46 i++;
47 in_dyntick_nmi = 0;
48 }
49 od;
50 dyntick_nmi_done = 1;
51 }
Of course, the fact that we have NMIs requires adjustments in the other components.
For example, the EXECUTE_MAINLINE() macro now needs to pay attention to the NMI
handler (in_dyntick_nmi) as well as the interrupt handler (in_dyntick_irq) by
checking the dyntick_nmi_done variable as follows:
1 #define EXECUTE_MAINLINE(label, stmt) \

2 label: skip; \
3 atomic { \
4 if \
5 :: in_dyntick_irq || \
6 in_dyntick_nmi -> goto label; \
v2023.06.11a
8 fi; \
9 }
We will also need to introduce an EXECUTE_IRQ() macro that checks in_dyntick_

nmi in order to allow dyntick_irq() to exclude dyntick_nmi():
1 #define EXECUTE_IRQ(label, stmt) \

2 label: skip; \
3 atomic { \
4 if \
5 :: in_dyntick_nmi -> goto label; \
7 fi; \
8 }
It is further necessary to convert dyntick_irq() to EXECUTE_IRQ() as follows:
2 {
3 byte tmp;
4 byte i = 0;
5 byte j = 0;
6 bit old_gp_idle;
7 bit outermost;
8
9 do
10 :: i >= MAX_DYNTICK_LOOP_IRQ &&
11 j >= MAX_DYNTICK_LOOP_IRQ -> break;
13 atomic {
14 outermost = (in_dyntick_irq == 0);
16 }
17 stmt1: skip;
18 atomic {
19 if
20 :: in_dyntick_nmi -> goto stmt1;
21 :: !in_dyntick_nmi && rcu_update_flag ->
22 goto stmt1_then;
23 :: else -> goto stmt1_else;
24 fi;
25 }
26 stmt1_then: skip;
27 EXECUTE_IRQ(stmt1_1, tmp = rcu_update_flag)
28 EXECUTE_IRQ(stmt1_2, rcu_update_flag = tmp + 1)
29 stmt1_else: skip;
30 stmt2: skip; atomic {
31 if
33 :: !in_dyntick_nmi &&
34 !in_interrupt &&
36 goto stmt2_then;
38 fi;
39 }
41 EXECUTE_IRQ(stmt2_1,
43 EXECUTE_IRQ(stmt2_2,
44 dynticks_progress_counter = tmp + 1)
46 EXECUTE_IRQ(stmt2_4, rcu_update_flag = tmp + 1)
48 EXECUTE_IRQ(stmt3, tmp = in_interrupt)
49 EXECUTE_IRQ(stmt4, in_interrupt = tmp + 1)
50 stmt5: skip;
51 atomic {
52 if
v2023.06.11a
54 :: !in_dyntick_nmi && outermost ->

56 :: else -> skip;
57 fi;
58 }
59 i++;
60 :: j < i ->
61 stmt6: skip;
62 atomic {
63 if
65 :: !in_dyntick_nmi && j + 1 == i ->
68 :: else -> skip;
69 fi;
70 }
71 EXECUTE_IRQ(stmt7, tmp = in_interrupt);
72 EXECUTE_IRQ(stmt8, in_interrupt = tmp - 1);
73 stmt9: skip;
74 atomic {
75 if
77 :: !in_dyntick_nmi && rcu_update_flag != 0 ->
78 goto stmt9_then;
80 fi;
81 }
84 EXECUTE_IRQ(stmt9_2, rcu_update_flag = tmp - 1)
85 stmt9_3: skip;
86 atomic {
87 if
88 :: in_dyntick_nmi -> goto stmt9_3;
89 :: !in_dyntick_nmi && rcu_update_flag == 0 ->
90 goto stmt9_3_then;
91 :: else -> goto stmt9_3_else;
92 fi;
93 }
94 stmt9_3_then: skip;
95 EXECUTE_IRQ(stmt9_3_1,
97 EXECUTE_IRQ(stmt9_3_2,
98 dynticks_progress_counter = tmp + 1)
99 stmt9_3_else:
101 atomic {
102 j++;
103 in_dyntick_irq = (i != j);
104 }
105 od;
107 }
Note that we have open-coded the “if” statements (for example, lines 17–29). In
addition, statements that process strictly local state (such as line 59) need not exclude
dyntick_nmi().
Finally, grace_period() requires only a few changes:
2 {
3 byte curr;
4 byte snap;
5 bit shouldexit;
6
8 atomic {
9 printf("MDL_NOHZ = %d\n", MAX_DYNTICK_LOOP_NOHZ);
10 printf("MDL_IRQ = %d\n", MAX_DYNTICK_LOOP_IRQ);
11 printf("MDL_NMI = %d\n", MAX_DYNTICK_LOOP_NMI);
v2023.06.11a
12 shouldexit = 0;
15 }
16 do
17 :: 1 ->
18 atomic {
20 shouldexit = dyntick_nohz_done &&
21 dyntick_irq_done &&
22 dyntick_nmi_done;
24 if
25 :: (curr - snap) >= 2 || (curr & 1) == 0 ->
26 break;
27 :: else -> skip;
28 fi;
29 }
30 od;
33 atomic {
34 shouldexit = 0;
37 }
38 do
39 :: 1 ->
40 atomic {
42 shouldexit = dyntick_nohz_done &&
43 dyntick_irq_done &&
44 dyntick_nmi_done;
46 if
47 :: (curr != snap) || ((curr & 1) == 0) ->
48 break;
49 :: else -> skip;
50 fi;
51 }
52 od;
54 }
We have added the printf() for the new MAX_DYNTICK_LOOP_NMI parameter on

line 11 and added dyntick_nmi_done to the shouldexit assignments on lines 22
and 44.
The model (dyntickRCU-irq-nmi-ssl.spin) results in a correct verification with
several hundred million states, passing without errors.
Quick Quiz 12.21: Does Paul always write his code in this painfully incremental manner?
12.1.6.8 Lessons (Re)Learned
This effort provided some lessons (re)learned:
1. Promela and Spin can verify interrupt/NMI-handler interactions.
2. Documenting code can help locate bugs. In this case, the documentation effort
located a misplaced memory barrier in rcu_enter_nohz() and rcu_exit_
nohz(), as shown by the following patch [McK08d].
v2023.06.11a
static inline void rcu_enter_nohz(void)

{
+ mb();
__get_cpu_var(dynticks_progress_counter)++;
- mb();
}
static inline void rcu_exit_nohz(void)

{
- mb();
__get_cpu_var(dynticks_progress_counter)++;
+ mb();
}
3. Validate your code early, often, and up to the point of destruction. This effort
located one subtle bug in rcu_try_flip_waitack_needed() that would have
been quite difficult to test or debug, as shown by the following patch [McK08c].
- if ((curr - snap) > 2 || (snap & 0x1) == 0)

+ if ((curr - snap) > 2 || (curr & 0x1) == 0)
4. Always verify your verification code. The usual way to do this is to insert a
deliberate bug and verify that the verification code catches it. Of course, if the
verification code fails to catch this bug, you may also need to verify the bug itself,
and so on, recursing infinitely. However, if you find yourself in this position, getting
a good night’s sleep can be an extremely effective debugging technique. You will
then see that the obvious verify-the-verification technique is to deliberately insert
bugs in the code being verified. If the verification fails to find them, the verification
clearly is buggy.
5. Use of atomic instructions can simplify verification. Unfortunately, use of the
cmpxchg atomic instruction would also slow down the critical IRQ fastpath, so
they are not appropriate in this case.
6. The need for complex formal verification often indicates a need to re-think
your design.
To this last point, it turns out that there is a much simpler solution to the dynticks
problem, which is presented in the next section.
12.1.6.9 Simplicity Avoids Formal Verification

The complexity of the dynticks interface for preemptible RCU is primarily due to the
fact that both IRQs and NMIs use the same code path and the same state variables.
This leads to the notion of providing separate code paths and variables for IRQs and
NMIs, as has been done for hierarchical RCU [McK08b] as indirectly suggested by
Manfred Spraul [Spr08]. This work was pulled into mainline kernel during the v2.6.29
development cycle [McK08f].
12.1.6.10 State Variables for Simplified Dynticks Interface

Listing 12.17 shows the new per-CPU state variables. These variables are grouped into
structs to allow multiple independent RCU implementations (e.g., rcu and rcu_bh) to
conveniently and efficiently share dynticks state. In what follows, they can be thought
of as independent per-CPU variables.
v2023.06.11a
Listing 12.17: Variables for Simple Dynticks Interface

1 struct rcu_dynticks {
2 int dynticks_nesting;
3 int dynticks;
4 int dynticks_nmi;
5 };
6
7 struct rcu_data {
8 ...
9 int dynticks_snap;
10 int dynticks_nmi_snap;
11 ...
12 };
The dynticks_nesting, dynticks, and dynticks_snap variables are for the

IRQ code paths, and the dynticks_nmi and dynticks_nmi_snap variables are for
the NMI code paths, although the NMI code path will also reference (but not modify)
the dynticks_nesting variable. These variables are used as follows:
dynticks_nesting
This counts the number of reasons that the corresponding CPU should be monitored
for RCU read-side critical sections. If the CPU is in dynticks-idle mode, then this
counts the IRQ nesting level, otherwise it is one greater than the IRQ nesting level.
dynticks
This counter’s value is even if the corresponding CPU is in dynticks-idle mode and
there are no IRQ handlers currently running on that CPU, otherwise the counter’s
value is odd. In other words, if this counter’s value is odd, then the corresponding
CPU might be in an RCU read-side critical section.
dynticks_nmi
This counter’s value is odd if the corresponding CPU is in an NMI handler, but
only if the NMI arrived while this CPU was in dyntick-idle mode with no IRQ
handlers running. Otherwise, the counter’s value will be even.
dynticks_snap
This will be a snapshot of the dynticks counter, but only if the current RCU grace
period has extended for too long a duration.
dynticks_nmi_snap
This will be a snapshot of the dynticks_nmi counter, but again only if the current
RCU grace period has extended for too long a duration.
If both dynticks and dynticks_nmi have taken on an even value during a given
time interval, then the corresponding CPU has passed through a quiescent state during
that interval.
Quick Quiz 12.22: But what happens if an NMI handler starts running before an IRQ handler
completes, and if that NMI handler continues running until a second IRQ handler starts?
12.1.6.11 Entering and Leaving Dynticks-Idle Mode

Listing 12.18 shows the rcu_enter_nohz() and rcu_exit_nohz(), which enter and
exit dynticks-idle mode, also known as “nohz” mode. These two functions are invoked
from process context.
v2023.06.11a
Listing 12.18: Entering and Exiting Dynticks-Idle Mode

1 void rcu_enter_nohz(void)
2 {
3 unsigned long flags;
4 struct rcu_dynticks *rdtp;
5
6 smp_mb();
7 local_irq_save(flags);
8 rdtp = &__get_cpu_var(rcu_dynticks);
9 rdtp->dynticks++;
10 rdtp->dynticks_nesting--;
11 WARN_ON(rdtp->dynticks & 0x1);
12 local_irq_restore(flags);
13 }
14
15 void rcu_exit_nohz(void)
16 {
17 unsigned long flags;
19
20 local_irq_save(flags);
22 rdtp->dynticks++;
23 rdtp->dynticks_nesting++;
24 WARN_ON(!(rdtp->dynticks & 0x1));
25 local_irq_restore(flags);
26 smp_mb();
27 }
Line 6 ensures that any prior memory accesses (which might include accesses from
RCU read-side critical sections) are seen by other CPUs before those marking entry to
dynticks-idle mode. Lines 7 and 12 disable and reenable IRQs. Line 8 acquires a pointer
to the current CPU’s rcu_dynticks structure, and line 9 increments the current CPU’s
dynticks counter, which should now be even, given that we are entering dynticks-idle
mode in process context. Finally, line 10 decrements dynticks_nesting, which
should now be zero.
The rcu_exit_nohz() function is quite similar, but increments dynticks_
nesting rather than decrementing it and checks for the opposite dynticks polarity.
12.1.6.12 NMIs From Dynticks-Idle Mode

Listing 12.19 shows the rcu_nmi_enter() and rcu_nmi_exit() functions, which
inform RCU of NMI entry and exit, respectively, from dynticks-idle mode. However, if
the NMI arrives during an IRQ handler, then RCU will already be on the lookout for
RCU read-side critical sections from this CPU, so lines 6 and 7 of rcu_nmi_enter()
and lines 18 and 19 of rcu_nmi_exit() silently return if dynticks is odd. Otherwise,
the two functions increment dynticks_nmi, with rcu_nmi_enter() leaving it with
an odd value and rcu_nmi_exit() leaving it with an even value. Both functions
execute memory barriers between this increment and possible RCU read-side critical
sections on lines 10 and 20, respectively.
12.1.6.13 Interrupts From Dynticks-Idle Mode

Listing 12.20 shows rcu_irq_enter() and rcu_irq_exit(), which inform RCU
of entry to and exit from, respectively, IRQ context. Line 6 of rcu_irq_enter()
increments dynticks_nesting, and if this variable was already non-zero, line 7
silently returns. Otherwise, line 8 increments dynticks, which will then have an odd
value, consistent with the fact that this CPU can now execute RCU read-side critical
v2023.06.11a
Listing 12.19: NMIs From Dynticks-Idle Mode

1 void rcu_nmi_enter(void)
2 {
4
6 if (rdtp->dynticks & 0x1)
7 return;
8 rdtp->dynticks_nmi++;
9 WARN_ON(!(rdtp->dynticks_nmi & 0x1));
10 smp_mb();
11 }
12
13 void rcu_nmi_exit(void)
14 {
16
18 if (rdtp->dynticks & 0x1)
19 return;
20 smp_mb();
21 rdtp->dynticks_nmi++;
22 WARN_ON(rdtp->dynticks_nmi & 0x1);
23 }
Listing 12.20: Interrupts From Dynticks-Idle Mode

1 void rcu_irq_enter(void)
2 {
4
6 if (rdtp->dynticks_nesting++)
7 return;
8 rdtp->dynticks++;
9 WARN_ON(!(rdtp->dynticks & 0x1));
10 smp_mb();
11 }
12
13 void rcu_irq_exit(void)
14 {
16
18 if (--rdtp->dynticks_nesting)
19 return;
20 smp_mb();
21 rdtp->dynticks++;
22 WARN_ON(rdtp->dynticks & 0x1);
23 if (__get_cpu_var(rcu_data).nxtlist ||
24 __get_cpu_var(rcu_bh_data).nxtlist)
25 set_need_resched();
26 }
sections. Line 10 therefore executes a memory barrier to ensure that the increment of
dynticks is seen before any RCU read-side critical sections that the subsequent IRQ
handler might execute.
Line 18 of rcu_irq_exit() decrements dynticks_nesting, and if the result is
non-zero, line 19 silently returns. Otherwise, line 20 executes a memory barrier to
ensure that the increment of dynticks on line 21 is seen after any RCU read-side critical
sections that the prior IRQ handler might have executed. Line 22 verifies that dynticks
is now even, consistent with the fact that no RCU read-side critical sections may appear
in dynticks-idle mode. Lines 23–25 check to see if the prior IRQ handlers enqueued any
RCU callbacks, forcing this CPU out of dynticks-idle mode via a reschedule API if so.
v2023.06.11a
Listing 12.21: Saving Dyntick Progress Counters

1 static int
2 dyntick_save_progress_counter(struct rcu_data *rdp)
3 {
4 int ret;
5 int snap;
6 int snap_nmi;
7
8 snap = rdp->dynticks->dynticks;
9 snap_nmi = rdp->dynticks->dynticks_nmi;
10 smp_mb();
11 rdp->dynticks_snap = snap;
12 rdp->dynticks_nmi_snap = snap_nmi;
13 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
14 if (ret)
15 rdp->dynticks_fqs++;
16 return ret;
17 }
12.1.6.14 Checking For Dynticks Quiescent States

Listing 12.21 shows dyntick_save_progress_counter(), which takes a snapshot
of the specified CPU’s dynticks and dynticks_nmi counters. Lines 8 and 9 snapshot
these two variables to locals, line 10 executes a memory barrier to pair with the memory
barriers in the functions in Listings 12.18, 12.19, and 12.20. Lines 11 and 12 record the
snapshots for later calls to rcu_implicit_dynticks_qs(), and line 13 checks to see
if the CPU is in dynticks-idle mode with neither IRQs nor NMIs in progress (in other
words, both snapshots have even values), hence in an extended quiescent state. If so,
lines 14 and 15 count this event, and line 16 returns true if the CPU was in a quiescent
state.
Listing 12.22 shows rcu_implicit_dynticks_qs(), which is called to check
whether a CPU has entered dyntick-idle mode subsequent to a call to dynticks_save_
progress_counter(). Lines 9 and 11 take new snapshots of the corresponding CPU’s
dynticks and dynticks_nmi variables, while lines 10 and 12 retrieve the snapshots
saved earlier by dynticks_save_progress_counter(). Line 13 then executes a
memory barrier to pair with the memory barriers in the functions in Listings 12.18,
12.19, and 12.20. Lines 14–15 then check to see if the CPU is either currently in
a quiescent state (curr and curr_nmi having even values) or has passed through
a quiescent state since the last call to dynticks_save_progress_counter() (the
values of dynticks and dynticks_nmi having changed). If these checks confirm that
the CPU has passed through a dyntick-idle quiescent state, then line 16 counts that
fact and line 17 returns an indication of this fact. Either way, line 19 checks for race
conditions that can result in RCU waiting for a CPU that is offline.
Quick Quiz 12.23: This is still pretty complicated. Why not just have a cpumask_t with
per-CPU bits, clearing the bit when entering an IRQ or NMI handler, and setting it upon exit?
Linux-kernel RCU’s dyntick-idle code has since been rewritten yet again based on a
suggestion from Andy Lutomirski [McK15c], but it is time to sum up and move on to
other topics.
12.1.6.15 Discussion
A slight shift in viewpoint resulted in a substantial simplification of the dynticks
interface for RCU. The key change leading to this simplification was minimizing of
v2023.06.11a
Listing 12.22: Checking Dyntick Progress Counters

1 static int
2 rcu_implicit_dynticks_qs(struct rcu_data *rdp)
3 {
4 long curr;
5 long curr_nmi;
6 long snap;
7 long snap_nmi;
8
9 curr = rdp->dynticks->dynticks;
10 snap = rdp->dynticks_snap;
11 curr_nmi = rdp->dynticks->dynticks_nmi;
12 snap_nmi = rdp->dynticks_nmi_snap;
13 smp_mb();
14 if ((curr != snap || (curr & 0x1) == 0) &&
15 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
16 rdp->dynticks_fqs++;
17 return 1;
18 }
19 return rcu_implicit_offline_qs(rdp);
20 }
sharing between IRQ and NMI contexts. The only sharing in this simplified interface is
references from NMI context to IRQ variables (the dynticks variable). This type of
sharing is benign, because the NMI functions never update this variable, so that its value
remains constant through the lifetime of the NMI handler. This limitation of sharing
allows the individual functions to be understood one at a time, in happy contrast to the
situation described in Section 12.1.5, where an NMI might change shared state at any
point during execution of the IRQ functions.
Verification can be a good thing, but simplicity is even better.
12.2 Special-Purpose State-Space Search
Jack of all trades, master of none.

Unknown
Although Promela and Spin allow you to verify pretty much any (smallish) algorithm,
their very generality can sometimes be a curse. For example, Promela does not
understand memory models or any sort of reordering semantics. This section therefore
describes some state-space search tools that understand memory models used by
production systems, greatly simplifying the verification of weakly ordered code.
For example, Section 12.1.4 showed how to convince Promela to account for weak
memory ordering. Although this approach can work well, it requires that the developer
fully understand the system’s memory model. Unfortunately, few (if any) developers
fully understand the complex memory models of modern CPUs.
Therefore, another approach is to use a tool that already understands this memory
ordering, such as the PPCMEM tool produced by Peter Sewell and Susmit Sarkar at the
University of Cambridge, Luc Maranget, Francesco Zappa Nardelli, and Pankaj Pawan
at INRIA, and Jade Alglave at Oxford University, in cooperation with Derek Williams
of IBM [AMP+ 11]. This group formalized the memory models of Power, Arm, x86, as
well as that of the C/C++11 standard [Smi19], and produced the PPCMEM tool based
on the Power and Arm formalizations.
v2023.06.11a
12.2. SPECIAL-PURPOSE STATE-SPACE SEARCH 403
Listing 12.23: PPCMEM Litmus Test

1 PPC SB+lwsync-RMW-lwsync+isync-simple
2 ""
3 {
4 0:r2=x; 0:r3=2; 0:r4=y; 0:r10=0; 0:r11=0; 0:r12=z;
5 1:r2=y; 1:r4=x;
6 }
7 P0 | P1 ;
8 li r1,1 | li r1,1 ;
9 stw r1,0(r2) | stw r1,0(r2) ;
10 lwsync | sync ;
11 | lwz r3,0(r4) ;
12 lwarx r11,r10,r12 | ;
13 stwcx. r11,r10,r12 | ;
14 bne Fail1 | ;
15 isync | ;
16 lwz r3,0(r4) | ;
17 Fail1: | ;
18
19 exists
20 (0:r3=0 /\ 1:r3=0)
Quick Quiz 12.24: But x86 has strong memory ordering, so why formalize its memory model?
The PPCMEM tool takes litmus tests as input. A sample litmus test is presented
in Section 12.2.1. Section 12.2.2 relates this litmus test to the equivalent C-language
program, Section 12.2.3 describes how to apply PPCMEM to this litmus test, and
Section 12.2.4 discusses the implications.
12.2.1 Anatomy of a Litmus Test

An example PowerPC litmus test for PPCMEM is shown in Listing 12.23. The ARM
interface works the same way, but with Arm instructions substituted for the Power
instructions and with the initial “PPC” replaced by “ARM”.
In the example, line 1 identifies the type of system (“ARM” or “PPC”) and contains the
title for the model. Line 2 provides a place for an alternative name for the test, which
you will usually want to leave blank as shown in the above example. Comments can be
inserted between lines 2 and 3 using the OCaml (or Pascal) syntax of (* *).
Lines 3–6 give initial values for all registers; each is of the form P:R=V, where P
is the process identifier, R is the register identifier, and V is the value. For example,
process 0’s register r3 initially contains the value 2. If the value is a variable (x, y, or z
in the example) then the register is initialized to the address of the variable. It is also
possible to initialize the contents of variables, for example, x=1 initializes the value of
x to 1. Uninitialized variables default to the value zero, so that in the example, x, y,
and z are all initially zero.
Line 7 provides identifiers for the two processes, so that the 0:r3=2 on line 4 could
instead have been written P0:r3=2. Line 7 is required, and the identifiers must be of
the form Pn, where n is the column number, starting from zero for the left-most column.
This may seem unnecessarily strict, but it does prevent considerable confusion in actual
use.
Quick Quiz 12.25: Why does line 8 of Listing 12.23 initialize the registers? Why not instead
initialize them on lines 4 and 5?
Lines 8–17 are the lines of code for each process. A given process can have empty
lines, as is the case for P0’s line 11 and P1’s lines 12–17. Labels and branches are
v2023.06.11a
Listing 12.24: Meaning of PPCMEM Litmus Test

1 void P0(void)
2 {
3 int r3;
4
5 x = 1; /* Lines 8 and 9 */
6 atomic_add_return(&z, 0); /* Lines 10-15 */
7 r3 = y; /* Line 16 */
8 }
9
10 void P1(void)
11 {
12 int r3;
13
14 y = 1; /* Lines 8-9 */
15 smp_mb(); /* Line 10 */
16 r3 = x; /* Line 11 */
17 }
permitted, as demonstrated by the branch on line 14 to the label on line 17. That said,
too-free use of branches will expand the state space. Use of loops is a particularly good
way to explode your state space.
Lines 19–20 show the assertion, which in this case indicates that we are interested in
whether P0’s and P1’s r3 registers can both contain zero after both threads complete
execution. This assertion is important because there are a number of use cases that
would fail miserably if both P0 and P1 saw zero in their respective r3 registers.
This should give you enough information to construct simple litmus tests. Some
additional documentation is available, though much of this additional documentation
is intended for a different research tool that runs tests on actual hardware. Perhaps
more importantly, a large number of pre-existing litmus tests are available with the
online tool (available via the “Select ARM Test” and “Select POWER Test” buttons at
https://www.cl.cam.ac.uk/~pes20/ppcmem/). It is quite likely that one of these
pre-existing litmus tests will answer your Power or Arm memory-ordering question.
12.2.2 What Does This Litmus Test Mean?

P0’s lines 8 and 9 are equivalent to the C statement x=1 because line 4 defines P0’s
register r2 to be the address of x. P0’s lines 12 and 13 are the mnemonics for load-linked
(“load register exclusive” in Arm parlance and “load reserve” in Power parlance) and
store-conditional (“store register exclusive” in Arm parlance), respectively. When
these are used together, they form an atomic instruction sequence, roughly similar to
the compare-and-swap sequences exemplified by the x86 lock;cmpxchg instruction.
Moving to a higher level of abstraction, the sequence from lines 10–15 is equivalent to the
Linux kernel’s atomic_add_return(&z, 0). Finally, line 16 is roughly equivalent
to the C statement r3=y.
P1’s lines 8 and 9 are equivalent to the C statement y=1, line 10 is a memory barrier,
equivalent to the Linux kernel statement smp_mb(), and line 11 is equivalent to the C
statement r3=x.
Quick Quiz 12.26: But whatever happened to line 17 of Listing 12.23, the one that is the
Fail1: label?
Putting all this together, the C-language equivalent to the entire litmus test is as shown
in Listing 12.24. The key point is that if atomic_add_return() acts as a full memory
barrier (as the Linux kernel requires it to), then it should be impossible for P0()’s and
P1()’s r3 variables to both be zero after execution completes.
v2023.06.11a
12.2. SPECIAL-PURPOSE STATE-SPACE SEARCH 405
Listing 12.25: PPCMEM Detects an Error

./ppcmem -model lwsync_read_block \
-model coherence_points filename.litmus
...
States 6
0:r3=0; 1:r3=0;
0:r3=0; 1:r3=1;
0:r3=1; 1:r3=0;
0:r3=1; 1:r3=1;
0:r3=2; 1:r3=0;
0:r3=2; 1:r3=1;
Ok
Condition exists (0:r3=0 /\ 1:r3=0)
Hash=e2240ce2072a2610c034ccd4fc964e77
Observation SB+lwsync-RMW-lwsync+isync Sometimes 1
Listing 12.26: PPCMEM on Repaired Litmus Test

./ppcmem -model lwsync_read_block \
-model coherence_points filename.litmus
...
States 5
0:r3=0; 1:r3=1;
0:r3=1; 1:r3=0;
0:r3=1; 1:r3=1;
0:r3=2; 1:r3=0;
0:r3=2; 1:r3=1;
No (allowed not found)
Condition exists (0:r3=0 /\ 1:r3=0)
Hash=77dd723cda9981248ea4459fcdf6097d
Observation SB+lwsync-RMW-lwsync+sync Never 0 5
The next section describes how to run this litmus test.
12.2.3 Running a Litmus Test

As noted earlier, litmus tests may be run interactively via https://www.cl.cam.ac.
uk/~pes20/ppcmem/, which can help build an understanding of the memory model.
However, this approach requires that the user manually carry out the full state-space
search. Because it is very difficult to be sure that you have checked every possible
sequence of events, a separate tool is provided for this purpose [McK11d].
Because the litmus test shown in Listing 12.23 contains read-modify-write instructions,
we must add -model arguments to the command line. If the litmus test is stored in
filename.litmus, this will result in the output shown in Listing 12.25, where
the ... stands for voluminous making-progress output. The list of states includes
0:r3=0; 1:r3=0;, indicating once again that the old PowerPC implementation of
atomic_add_return() does not act as a full barrier. The “Sometimes” on the last
line confirms this: The assertion triggers for some executions, but not all of the time.
The fix to this Linux-kernel bug is to replace P0’s isync with sync, which results in
the output shown in Listing 12.26. As you can see, 0:r3=0; 1:r3=0; does not appear
in the list of states, and the last line calls out “Never”. Therefore, the model predicts
that the offending execution sequence cannot happen.
Quick Quiz 12.27: Does the Arm Linux kernel have a similar bug?
Quick Quiz 12.28: Does the lwsync on line 10 in Listing 12.23 provide sufficient ordering?
v2023.06.11a
12.2.4 PPCMEM Discussion

These tools promise to be of great help to people working on low-level parallel primitives
that run on Arm and on Power. These tools do have some intrinsic limitations:
1. These tools are research prototypes, and as such are unsupported.

2. These tools do not constitute official statements by IBM or Arm on their respective
CPU architectures. For example, both corporations reserve the right to report a
bug at any time against any version of any of these tools. These tools are therefore
not a substitute for careful stress testing on real hardware. Moreover, both the
tools and the model that they are based on are under active development and might
change at any time. On the other hand, this model was developed in consultation
with the relevant hardware experts, so there is good reason to be confident that it is
a robust representation of the architectures.
3. These tools currently handle a subset of the instruction set. This subset has been
sufficient for my purposes, but your mileage may vary. In particular, the tool
handles only word-sized accesses (32 bits), and the words accessed must be properly
aligned.3 In addition, the tool does not handle some of the weaker variants of the
Arm memory-barrier instructions, nor does it handle arithmetic.
4. The tools are restricted to small loop-free code fragments running on small numbers
of threads. Larger examples result in state-space explosion, just as with similar
tools such as Promela and Spin.
5. The full state-space search does not give any indication of how each offending
state was reached. That said, once you realize that the state is in fact reachable, it
is usually not too hard to find that state using the interactive tool.
6. These tools are not much good for complex data structures, although it is possible
to create and traverse extremely simple linked lists using initialization statements
of the form “x=y; y=z; z=42;”.
7. These tools do not handle memory mapped I/O or device registers. Of course,
handling such things would require that they be formalized, which does not appear
to be in the offing.
8. The tools will detect only those problems for which you code an assertion. This
weakness is common to all formal methods, and is yet another reason why testing
remains important. In the immortal words of Donald Knuth quoted at the beginning
of this chapter, “Beware of bugs in the above code; I have only proved it correct,
not tried it.”
That said, one strength of these tools is that they are designed to model the full range
of behaviors allowed by the architectures, including behaviors that are legal, but which
current hardware implementations do not yet inflict on unwary software developers.
Therefore, an algorithm that is vetted by these tools likely has some additional safety
margin when running on real hardware. Furthermore, testing on real hardware can only
find bugs; such testing is inherently incapable of proving a given usage correct. To
appreciate this, consider that the researchers routinely ran in excess of 100 billion test
3 But recent work focuses on mixed-size accesses [FSP+ 17].
v2023.06.11a
12.3. AXIOMATIC APPROACHES 407
Listing 12.27: IRIW Litmus Test

1 PPC IRIW.litmus
2 ""
3 (* Traditional IRIW. *)
4 {
5 0:r1=1; 0:r2=x;
6 1:r1=1; 1:r4=y;
7 2:r2=x; 2:r4=y;
8 3:r2=x; 3:r4=y;
9 }
10 P0 | P1 | P2 | P3 ;
11 stw r1,0(r2) | stw r1,0(r4) | lwz r3,0(r2) | lwz r3,0(r4) ;
12 | | sync | sync ;
13 | | lwz r5,0(r4) | lwz r5,0(r2) ;
14
15 exists
16 (2:r3=1 /\ 2:r5=0 /\ 3:r3=1 /\ 3:r5=0)
runs on real hardware to validate their model. In one case, behavior that is allowed
by the architecture did not occur, despite 176 billion runs [AMP+ 11]. In contrast, the
full-state-space search allows the tool to prove code fragments correct.
It is worth repeating that formal methods and tools are no substitute for testing. The
fact is that producing large reliable concurrent software artifacts, the Linux kernel for
example, is quite difficult. Developers must therefore be prepared to apply every tool at
their disposal towards this goal. The tools presented in this chapter are able to locate
bugs that are quite difficult to produce (let alone track down) via testing. On the other
hand, testing can be applied to far larger bodies of software than the tools presented in
this chapter are ever likely to handle. As always, use the right tools for the job!
Of course, it is always best to avoid the need to work at this level by designing your
parallel code to be easily partitioned and then using higher-level primitives (such as
locks, sequence counters, atomic operations, and RCU) to get your job done more
straightforwardly. And even if you absolutely must use low-level memory barriers and
read-modify-write instructions to get your job done, the more conservative your use of
these sharp instruments, the easier your life is likely to be.
12.3 Axiomatic Approaches
Theory helps us to bear our ignorance of facts.

George Santayana
Although the PPCMEM tool can solve the famous “independent reads of independent
writes” (IRIW) litmus test shown in Listing 12.27, doing so requires no less than
fourteen CPU hours and generates no less than ten gigabytes of state space. That said,
this situation is a great improvement over that before the advent of PPCMEM, where
solving this problem required perusing volumes of reference manuals, attempting proofs,
discussing with experts, and being unsure of the final answer. Although fourteen hours
can seem like a long time, it is much shorter than weeks or even months.
However, the time required is a bit surprising given the simplicity of the litmus test,
which has two threads storing to two separate variables and two other threads loading
from these two variables in opposite orders. The assertion triggers if the two loading
threads disagree on the order of the two stores. Even by the standards of memory-order
litmus tests, this is quite simple.
v2023.06.11a
Listing 12.28: Expanded IRIW Litmus Test

1 PPC IRIW5.litmus
2 ""
3 (* Traditional IRIW, but with five stores instead of *)
4 (* just one. *)
5 {
6 0:r1=1; 0:r2=x;
7 1:r1=1; 1:r4=y;
8 2:r2=x; 2:r4=y;
9 3:r2=x; 3:r4=y;
10 }
11 P0 | P1 | P2 | P3 ;
13 addi r1,r1,1 | addi r1,r1,1 | sync | sync ;
15 addi r1,r1,1 | addi r1,r1,1 | | ;
16 stw r1,0(r2) | stw r1,0(r4) | | ;
17 addi r1,r1,1 | addi r1,r1,1 | | ;
18 stw r1,0(r2) | stw r1,0(r4) | | ;
19 addi r1,r1,1 | addi r1,r1,1 | | ;
20 stw r1,0(r2) | stw r1,0(r4) | | ;
21
22 exists
23 (2:r3=1 /\ 2:r5=0 /\ 3:r3=1 /\ 3:r5=0)
One reason for the amount of time and space consumed is that PPCMEM does a
trace-based full-state-space search, which means that it must generate and evaluate
all possible orders and combinations of events at the architectural level. At this level,
both loads and stores correspond to ornate sequences of events and actions, resulting
in a very large state space that must be completely searched, in turn resulting in large
memory and CPU consumption.
Of course, many of the traces are quite similar to one another, which suggests that
an approach that treated similar traces as one might improve performace. One such
approach is the axiomatic approach of Alglave et al. [AMT14], which creates a set of
axioms to represent the memory model and then converts litmus tests to theorems that
might be proven or disproven over this set of axioms. The resulting tool, called “herd”,
conveniently takes as input the same litmus tests as PPCMEM, including the IRIW
litmus test shown in Listing 12.27.
However, where PPCMEM requires 14 CPU hours to solve IRIW, herd does so in
17 milliseconds, which represents a speedup of more than six orders of magnitude.
That said, the problem is exponential in nature, so we should expect herd to exhibit
exponential slowdowns for larger problems. And this is exactly what happens, for
example, if we add four more writes per writing CPU as shown in Listing 12.28, herd
slows down by a factor of more than 50,000, requiring more than 15 minutes of CPU
time. Adding threads also results in exponential slowdowns [MS14].
Despite their exponential nature, both PPCMEM and herd have proven quite useful
for checking key parallel algorithms, including the queued-lock handoff on x86 systems.
The weaknesses of the herd tool are similar to those of PPCMEM, which were described
in Section 12.2.4. There are some obscure (but very real) cases for which the PPCMEM
and herd tools disagree, and as of 2021 many but not all of these disagreements was
resolved.
It would be helpful if the litmus tests could be written in C (as in Listing 12.24) rather
than assembly (as in Listing 12.23). This is now possible, as will be described in the
following sections.
v2023.06.11a
Listing 12.29: Locking Example

1 C Lock1
2
3 {}
4
5 P0(int *x, spinlock_t *sp)
6 {
7 spin_lock(sp);
8 WRITE_ONCE(*x, 1);
10 spin_unlock(sp);
11 }
12
14 {
15 int r1;
16
17 spin_lock(sp);
18 r1 = READ_ONCE(*x);
19 spin_unlock(sp);
20 }
21
22 exists (1:r1=1)
Listing 12.30: Broken Locking Example

1 C Lock2
2
3 {}
4
5 P0(int *x, spinlock_t *sp1)
6 {
7 spin_lock(sp1);
10 spin_unlock(sp1);
11 }
12
13 P1(int *x, spinlock_t *sp2) // Buggy!
14 {
15 int r1;
16
17 spin_lock(sp2);
19 spin_unlock(sp2);
20 }
21
22 exists (1:r1=1)
12.3.1 Axiomatic Approaches and Locking

Axiomatic approaches may also be applied to higher-level languages and also to higher-
level synchronization primitives, as exemplified by the lock-based litmus test shown in
Listing 12.29 (C-Lock1.litmus). This litmus test can be modeled by the Linux kernel
memory consistency model (LKMM) [AMM+ 18, MS18]. As expected, the herd tool’s
output features the string Never, correctly indicating that P1() cannot see x having a
value of one.4
Quick Quiz 12.29: What do you have to do to run herd on litmus tests like that shown in
Listing 12.29?
Of course, if P0() and P1() use different locks, as shown in Listing 12.30 (C-
Lock2.litmus), then all bets are off. And in this case, the herd tool’s output features
4 The output of the herd tool is compatible with that of PPCMEM, so feel free to look at
Listings 12.25 and 12.26 for examples showing the output format.
v2023.06.11a
Listing 12.31: Canonical RCU Removal Litmus Test

1 C C-RCU-remove
2
3 {
4 int z=1;
5 int y=2;
6 int *x=y;
7 }
8
9 P0(int **x, int *y, int *z)
10 {
11 rcu_assign_pointer(*x, z);
13 WRITE_ONCE(*y, 0);
14 }
15
16 P1(int **x, int *y, int *z)
17 {
18 int *r1;
19 int r2;
20
21 rcu_read_lock();
22 r1 = rcu_dereference(*x);
23 r2 = READ_ONCE(*r1);
25 }
26
27 locations [1:r1; x; y; z]
28 exists (1:r2=0)
the string Sometimes, correctly indicating that use of different locks allows P1() to
see x having a value of one.
Quick Quiz 12.30: Why bother modeling locking directly? Why not simply emulate locking
with atomic operations?
But locking is not the only synchronization primitive that can be modeled directly:
The next section looks at RCU.
12.3.2 Axiomatic Approaches and RCU

Axiomatic approaches can also analyze litmus tests involving RCU [AMM+ 18]. To that
end, Listing 12.31 (C-RCU-remove.litmus) shows a litmus test corresponding to the
canonical RCU-mediated removal from a linked list. As with the locking litmus test,
this RCU litmus test can be modeled by LKMM, with similar performance advantages
compared to modeling emulations of RCU. Line 6 shows x as the list head, initially
referencing y, which in turn is initialized to the value 2 on line 5.
P0() on lines 9–14 removes element y from the list by replacing it with element z
(line 11), waits for a grace period (line 12), and finally zeroes y to emulate free()
(line 13). P1() on lines 16–25 executes within an RCU read-side critical section
(lines 21–24), picking up the list head (line 22) and then loading the next element
(line 23). The next element should be non-zero, that is, not yet freed (line 28). Several
other variables are output for debugging purposes (line 27).
The output of the herd tool when running this litmus test features Never, indicating
that P0() never accesses a freed element, as expected. Also as expected, removing
line 12 results in P0() accessing a freed element, as indicated by the Sometimes in the
herd output.
A litmus test for a more complex example proposed by Roman Penyaev [Pen18] is
shown in Listing 12.32 (C-RomanPenyaev-list-rcu-rr.litmus). In this example,
v2023.06.11a
Listing 12.32: Complex RCU Litmus Test

1 C C-RomanPenyaev-list-rcu-rr
2
3 {
4 int *z=1;
5 int *y=z;
6 int *x=y;
7 int *w=x;
8 int *v=w;
9 int *c=w;
10 }
11
12 P0(int **c, int **v)
13 {
14 int *r1;
15 int *r2;
16 int *r3;
17 int *r4;
18
19 rcu_read_lock();
20 r1 = READ_ONCE(*c);
21 if (r1 == 0) {
22 r1 = READ_ONCE(*v);
23 }
24 r2 = rcu_dereference(*(int **)r1);
25 smp_store_release(c, r2);
27 rcu_read_lock();
29 if (r3 == 0) {
30 r3 = READ_ONCE(*v);
31 }
32 r4 = rcu_dereference(*(int **)r3);
33 smp_store_release(c, r4);
35 }
36
37 P1(int **c, int **v, int **w, int **x, int **y)
38 {
39 int *r1;
40
41 rcu_assign_pointer(*w, y);
44 if ((int **)r1 == x) {
45 WRITE_ONCE(*c, 0);
47 }
48 smp_store_release(x, 0);
49 }
50
51 locations [1:r1; c; v; w; x; y]
52 exists (0:r1=0 \/ 0:r2=0 \/ 0:r3=0 \/ 0:r4=0)
v2023.06.11a
readers (modeled by P0() on lines 12–35) access a linked list in a round-robin fashion by
“leaking” a pointer to the last list element accessed into variable c. Updaters (modeled
by P1() on lines 37–49) remove an element, taking care to avoid disrupting current or
future readers.
Quick Quiz 12.31: Wait!!! Isn’t leaking pointers out of an RCU read-side critical section a
critical bug???
Lines 4–8 define the initial linked list, tail first. In the Linux kernel, this would be a
doubly linked circular list, but herd is currently incapable of modeling such a beast.
The strategy is instead to use a singly linked linear list that is long enough that the end is
never reached. Line 9 defines variable c, which is used to cache the list pointer between
successive RCU read-side critical sections.
Again, P0() on lines 12–35 models readers. This process models a pair of successive
readers traversing round-robin through the list, with the first reader on lines 19–26 and
the second reader on lines 27–34. Line 20 fetches the pointer cached in c, and if line 21
sees that the pointer was NULL, line 22 restarts at the beginning of the list. In either case,
line 24 advances to the next list element, and line 25 stores a pointer to this element back
into variable c. Lines 27–34 repeat this process, but using registers r3 and r4 instead
of r1 and r2. As with Listing 12.31, this litmus test stores zero to emulate free(), so
line 52 checks for any of these four registers being NULL, also known as zero.
Because P0() leaks an RCU-protected pointer from its first RCU read-side critical
section to its second, P1() must carry out its update (removing x) very carefully. Line 41
removes x by linking w to y. Line 42 waits for readers, after which no subsequent reader
has a path to x via the linked list. Line 43 fetches c, and if line 44 determines that
c references the newly removed x, line 45 sets c to NULL and line 46 again waits for
readers, after which no subsequent reader can fetch x from c. In either case, line 48
emulates free() by storing zero to x.
Quick Quiz 12.32: In Listing 12.32, why couldn’t a reader fetch c just before P1() zeroed it
on line 45, and then later store this same value back into c just after it was zeroed, thus defeating
the zeroing operation?
The output of the herd tool when running this litmus test features Never, indicating
that P0() never accesses a freed element, as expected. Also as expected, removing
either synchronize_rcu() results in P1() accessing a freed element, as indicated by
Sometimes in the herd output.
Quick Quiz 12.33: In Listing 12.32, why not have just one call to synchronize_rcu()
immediately before line 48?
Quick Quiz 12.34: Also in Listing 12.32, can’t line 48 be WRITE_ONCE() instead of
smp_store_release()?
These sections have shown how axiomatic approaches can successfully model
synchronization primitives such as locking and RCU in C-language litmus tests. Longer
term, the hope is that the axiomatic approaches will model even higher-level software
artifacts, producing exponential verification speedups. This could potentially allow
axiomatic verification of much larger software systems, perhaps incorporating spatial-
synchronization techniques from separation logic [GRY13, ORY01]. Another alternative
is to press the axioms of boolean logic into service, as described in the next section.
v2023.06.11a
12.4. SAT SOLVERS 413
C Code
CBMC
Logic Expression
SAT Solver
Trace Generation
(If Counterexample
Located)
Verification Result
Figure 12.2: CBMC Processing Flow
12.4 SAT Solvers
Live by the heuristic, die by the heuristic.

Unknown
Any finite program with bounded loops and recursion can be converted into a logic
expression, which might express that program’s assertions in terms of its inputs. Given
such a logic expression, it would be quite interesting to know whether any possible
combinations of inputs could result in one of the assertions triggering. If the inputs are
expressed as combinations of boolean variables, this is simply SAT, also known as the
satisfiability problem. SAT solvers are heavily used in verification of hardware, which
has motivated great advances. A world-class early 1990s SAT solver might be able to
handle a logic expression with 100 distinct boolean variables, but by the early 2010s
million-variable SAT solvers were readily available [KS08].
In addition, front-end programs for SAT solvers can automatically translate C code
into logic expressions, taking assertions into account and generating assertions for error
conditions such as array-bounds errors. One example is the C bounded model checker,
or cbmc, which is available as part of many Linux distributions. This tool is quite easy
to use, with cbmc test.c sufficing to validate test.c, resulting in the processing flow
shown in Figure 12.2. This ease of use is exceedingly important because it opens the
door to formal verification being incorporated into regression-testing frameworks. In
contrast, the traditional tools that require non-trivial translation to a special-purpose
language are confined to design-time verification.
More recently, SAT solvers have appeared that handle parallel code. These solvers
operate by converting the input code into single static assignment (SSA) form, then
generating all permitted access orders. This approach seems promising, but it remains
v2023.06.11a
to be seen how well it works in practice. One encouraging sign is work in 2016
applying cbmc to Linux-kernel RCU [LMKM16, LMKM18, Roy17]. This work used
minimal configurations of RCU, and verified scenarios using small numbers of threads,
but nevertheless successfully ingested Linux-kernel C code and produced a useful
result. The logic expressions generated from the C code had up to 90 million variables,
450 million clauses, occupied tens of gigabytes of memory, and required up to 80 hours
of CPU time for the SAT solver to produce the correct result.
Nevertheless, a Linux-kernel hacker might be justified in feeling skeptical of a claim
that his or her code had been automatically verified, and such hackers would find
many fellow skeptics going back decades [DMLP79]. One way to productively express
such skepticism is to provide bug-injected versions of the allegedly verified code. If
the formal-verification tool finds all the injected bugs, our hacker might gain more
confidence in the tool’s capabilities. Of course, tools that find valid bugs of which
the hacker was not yet aware will likely engender even more confidence. And this is
exactly why there is a git archive with a 20-branch set of mutations, with each branch
potentially containing a bug injected into Linux-kernel RCU [McK17]. Anyone with a
formal-verification tool is cordially invited to try that tool out on this set of verification
challenges.
Currently, cbmc is able to find a number of injected bugs, however, it has not yet been
able to locate a bug that RCU’s maintainer was not already aware of. Nevertheless, there
is some reason to hope that SAT solvers will someday be useful for finding concurrency
bugs in parallel code.
12.5 Stateless Model Checkers
He’s making a list, he’s permuting it twice. . .

with apologies to Haven Gillespie and J. Fred Coots
The SAT-solver approaches described in the previous section are quite convenient and
powerful, but the full tracking of all possible executions, including state, can incur
substantial overhead. In fact, the memory and CPU-time overheads can sharply limit
the size of programs that can be feasibly verified, which raises the question of whether
less-exact approaches might find bugs in larger programs.
Although the jury is still out on this question, stateless model checkers such as
Nidhugg [LSLK14] have in some cases handled larger programs [KS17b], and with
similar ease of use, as illustrated by Figure 12.3. In addition, Nidhugg was more than
an order of magnitude faster than was cbmc for some Linux-kernel RCU verification
scenarios. Of course, Nidhugg’s speed and scalability advantages are tied to the fact
that it does not handle data non-determinism, but this was not a factor in these particular
verification scenarios.
Nevertheless, as with cbmc, Nidhugg has not yet been able to locate a bug that
Linux-kernel RCU’s maintainer was not already aware of. However, it was able to
demonstrate that one historical bug in Linux-kernel RCU was fixed by a different commit
than the maintainer thought, which gives some additional hope that stateless model
checkers like Nidhugg might someday be useful for finding concurrency bugs in parallel
code.
v2023.06.11a
12.6. SUMMARY 415
C Code
Nidhugg
LLVM Internal
Representation
Dynamic Partial
Order Reduction
(DPOR) Algorithm
Trace Generation
(If Counterexample
Located)
Verification Result
Figure 12.3: Nidhugg Processing Flow
12.6 Summary
Western thought has focused on True-False; it is

high time to shift to Robust-Fragile.
Nassim Nicholas Taleb, summarized
The formal-verification techniques described in this chapter are very powerful tools
for validating small parallel algorithms, but they should not be the only tools in your
toolbox. Despite decades of focus on formal verification, testing remains the validation
workhorse for large parallel software systems [Cor06a, Jon11, McK15d].
It is nevertheless quite possible that this will not always be the case. To see this,
consider that there is estimated to be more than twenty billion instances of the Linux
kernel as of 2017. Suppose that the Linux kernel has a bug that manifests on average
every million years of runtime. As noted at the end of the preceding chapter, this
bug will be appearing more than 50 times per day across the installed base. But the
fact remains that most formal validation techniques can be used only on very small
codebases. So what is a concurrency coder to do?
Think in terms of finding the first bug, the first relevant bug, the last relevant bug,
and the last bug.
The first bug is normally found via inspection or compiler diagnostics. Although the
increasingly sophisticated compiler diagnostics comprise a lightweight sort of formal
verification, it is not common to think of them in those terms. This is in part due to an
odd practitioner prejudice which says “If I am using it, it cannot be formal verification”
on the one hand, and a large gap between compiler diagnostics and verification research
on the other.
v2023.06.11a
Although the first relevant bug might be located via inspection or compiler diagnostics,
it is not unusual for these two steps to find only typos and false positives. Either way,
the bulk of the relevant bugs, that is, those bugs that might actually be encountered in
production, will often be found via testing.
When testing is driven by anticipated or real use cases, it is not uncommon for the last
relevant bug to be located by testing. This situation might motivate a complete rejection
of formal verification, however, irrelevant bugs have an annoying habit of suddenly
becoming relevant at the least convenient moment possible, courtesy of black-hat attacks.
For security-critical software, which appears to be a continually increasing fraction of
the total, there can thus be strong motivation to find and fix the last bug. Testing is
demonstrably unable to find the last bug, so there is a possible role for formal verification,
assuming, that is, that formal verification proves capable of growing into that role. As
this chapter has shown, current formal verification systems are extremely limited.
Quick Quiz 12.35: But shouldn’t sufficiently low-level software be for all intents and purposes
immune to being exploited by black hats?
Please note that formal verification is often much harder to use than is testing. This is
in part a cultural statement, and there is reason to hope that formal verification will be
perceived to be easier with increased familiarity. That said, very simple test harnesses
can find significant bugs in arbitrarily large software systems. In contrast, the effort
required to apply formal verification seems to increase dramatically as the system size
increases.
I have nevertheless made occasional use of formal verification for almost 30 years
by playing to formal verification’s strengths, namely design-time verification of small
complex portions of the overarching software construct. The larger overarching software
construct is of course validated by testing.
Quick Quiz 12.36: In light of the full verification of the L4 microkernel, isn’t this limited
view of formal verification just a little bit obsolete?
One final approach is to consider the following two definitions from Section 11.1.2
and the consequence that they imply:
Definition: Bug-free programs are trivial programs.
Definition: Reliable programs have no known bugs.
Consequence: Any non-trivial reliable program contains at least one as-yet-unknown
bug.
From this viewpoint, any advances in validation and verification can have but two
effects: (1) An increase in the number of trivial programs or (2) A decrease in the
number of reliable programs. Of course, the human race’s increasing reliance on
multicore systems and software provides extreme motivation for a very sharp increase
in the number of trivial programs.
However, if your code is so complex that you find yourself relying too heavily
on formal-verification tools, you should carefully rethink your design, especially if
your formal-verification tools require your code to be hand-translated to a special-
purpose language. For example, a complex implementation of the dynticks interface
for preemptible RCU that was presented in Section 12.1.5 turned out to have a much
simpler alternative implementation, as discussed in Section 12.1.6.9. All else being
equal, a simpler implementation is much better than a proof of correctness for a complex
implementation.
v2023.06.11a
12.7. CHOOSING A VALIDATION PLAN 417
And the open challenge to those working on formal verification techniques and
systems is to prove this summary wrong! To assist in this task, Verification Challenge 6
is now available [McK17]. Have at it!!!
12.7 Choosing a Validation Plan
Science is a first-rate piece of furniture for one’s

upper chamber, but only given common sense on the
ground floor.
Oliver Wendell Holmes, updated
What sort of validation should you use for your project?

As is often the case in software in particular and in engineering in general, the answer
is “it depends”.
Note that neither running a test nor undertaking formal verification will change
your project. At best, such effort have an indirect effect by locating a bug that is later
fixed. Nevertheless, fixing a bug might prevent inconvenience, monetary loss, property
damage, or even loss of life. Clearly, this sort of indirect effect can be extremely
valuable.
Unfortunately, as we have seen, it is difficult to predict whether or not a given
validation effort will find important bugs. It is therefore all too easy to invest too
little—or even to fail to invest at all, especially if development estimates proved overly
optimistic or budgets unexpectedly tight, conditions which almost always come into
play in real-world software projects.
The decision to nevertheless invest in validation is often forced by experienced people
with forceful personalities. But this is no guarantee, given that other stakeholders might
also have forceful personalities. Worse yet, these other stakeholders might bring stories
of expensive validation efforts that nevertheless allowed embarrassing bugs to escape to
the end users. So although a scarred, grey-haired, and grouchy veteran might carry the
day, a more organized approach would perhaps be more useful.
Fortunately, there is a strictly financial analog to investments in validation, and that is
the insurance policy.
Both insurance policies and validation efforts require consistent up-front investments,
and both defend against disasters that might or might not ever happen. Furthermore,
both have exclusions of various types. For example, insurance policies for coastal areas
might exclude damages due to tidal waves, while on the other hand we have seen that
there is not yet any validation methodology that can find each and every bug.
In addition, it is possible to over-invest in both insurance and in validation. For but
one example, a validation plan that consumed the entire development budget would be
just as pointless as would an insurance policy that covered the Sun going nova.
One approach is to devote a given fraction of the software budget to validation, with
that fraction depending on the criticality of the software, so that safety-critical avionics
software might grant a larger fraction of its budget to validation than would a homework
assignment. Where available, experience from prior similar projects should be brought
to bear. However, it is necessary to structure the project so that the validation investment
starts when the project does, otherwise the inevitable overruns in spending on coding
will crowd out the validation effort.
v2023.06.11a
35000 50
RCU
30000 RCU Test
% Test 40
25000
20000 30
% Test
LoC 15000 20
10000
10
5000
0 v2.6.12 0
v2.6.16
v2.6.20
v2.6.24
v2.6.28
v2.6.32
v2.6.36
v3.0
v3.4
v3.8
v3.12
v3.16
v4.0
v4.4
v4.8
v4.12
v4.16
v5.0
v5.4
v5.8
v5.12
v5.16
v6.0
v6.3
Linux Release
Figure 12.4: Linux-Kernel RCU Test Code
Staffing start-up projects with experienced people can result in overinvestment in

validation efforts. Just as it is possible to go broke buying too much insurance, it is
possible to kill a project by investing too much in testing. This is especially the case
for first-of-a-kind projects where it is not yet clear which use cases will be important,
in which case testing for all possible use cases will be a possibly fatal waste of time,
energy, and funding.
However, as the tasks supported by a start-up project become more routine, users often
become less forgiving of failures, thus increasing the need for validation. Managing this
shift in investment can be extremely challenging, especially in the all-too-common case
where the users are unwilling or unable to disclose the exact nature of their use case. It
then becomes critically important to reverse-engineer the use cases from bug reports
and from discussions with the users. As these use cases are better understood, use of
continuous integration can help reduce the cost of finding and fixing any bugs located.
One example evolution of a software project’s use of validation is shown in Figure 12.4.
As can be seen in the figure, Linux-kernel RCU didn’t have any validation code
whatsoever until Linux kernel v2.6.15, which was released more than two years after
RCU was accepted into the kernel. The test suite achieved its peak fraction of the total
lines of code in Linux kernel v2.6.19–v2.6.21. This fraction decreased sharply with the
acceptance of preemptible RCU for real-time applications in v2.6.25. This decrease was
due to the fact that the RCU API was identical in the preemptible and non-preemptible
variants of RCU. This in turn meant that the existing test suite applied to both variants,
so that even though the Linux-kernel RCU code expanded significantly, there was no
need to expand the tests.
Subsequent bars in Figure 12.4 show that the RCU code base expanded significantly,
but that the corresponding validation code expanded even more dramatically. Linux
kernel v3.5 added tests for the rcu_barrier() API, closing a long-standing hole
in test coverage. Linux kernel v3.14 added automated testing and analysis of test
results, moving RCU towards continuous integration. Linux kernel v4.7 added a
performance validation suite for RCU’s update-side primitives. Linux kernel v4.12
added Tree SRCU, featuring improved update-side scalability, and v4.13 removed the
old less-scalable SRCU implementation. Linux kernel v5.0 briefly hosted the nolibc
library within the rcutorture scripting directory before it moved to its long-term home in
tools/include/nolibc. Linux kernel v5.8 added the Tasks Trace and Rude flavors
of RCU. Linux kernel v5.9 added the refscale.c suite of read-side performance
v2023.06.11a
12.7. CHOOSING A VALIDATION PLAN 419
tests. Linux kernels v5.12 and v5.13 started adding the ability to change a given CPU’s
callback-offloading status at runtime and also added the torture.sh acceptance-test
script. Linux kernel v5.14 added distributed rcutorture. Linux kernel v5.15 added
demonic vCPU placement in rcutorture testing, which was successful in locating a
number of race conditions.5 Linux kernel v5.17 removed the RCU_FAST_NO_HZ Kconfig
option. Numerous other changes may be found in the Linux kernel’s git archives.
We have established that the validation budget varies from one project to the next, and
also over the lifetime of any given project. But how should the validation investment be
split between testing and formal verification?
This question is being answered naturally as compilers adopt increasingly aggressive
formal-verification techniques into their diagnostics and as formal-verification tools
continue to mature. In addition, the Linux-kernel lockdep and KCSAN tools illustrate
the advantages of combining formal verification techniques with run-time analysis, as
discussed in Section 11.3. Other combined techniques analyze traces gathered from
executions [dOCdO19]. For the time being, the best practice is to focus first on testing
and to reserve explicit work on formal verification for those portions of the project
that are not well-served by testing, and that have exceptional needs for robustness. For
example, Linux-kernel RCU relies primarily on testing, but has made occasional use of
formal verification as discussed in this chapter.
In short, choosing a validation plan for concurrent software remains more an art than
a science, let alone a field of engineering. However, there is every reason to expect that
increasingly rigorous approaches will continue to become more prevalent.
5 The trick is to place one pair of vCPUs within the same core on one socket, while
placing another pair within the same core on some other socket. As you might expect
from Chapter 3, this produces different memory latencies between different pairs of vCPUs
(https://paulmck.livejournal.com/62071.html).
v2023.06.11a
v2023.06.11a
You don’t learn how to shoot and then learn how to
launch and then learn to do a controlled spin—you
learn to launch-shoot-spin.
Ender’s Shadow, Orson Scott Card
Chapter 13
Putting It All Together
This chapter gives some hints on concurrent-programming puzzles. Section 13.1

considers counter conundrums, Section 13.2 refurbishes reference counting, Section 13.3
helps with hazard pointers, Section 13.4 surmises on sequence-locking specials, and
finally Section 13.5 reflects on RCU rescues.
13.1 Counter Conundrums
Ford carried on counting quietly. This is about the

most aggressive thing you can do to a computer, the
equivalent of going up to a human being and saying
“Blood . . . blood . . . blood . . . blood . . .”
Douglas Adams
This section outlines solutions to counter conundrums.
13.1.1 Counting Updates

Suppose that Schrödinger (see Section 10.1) wants to count the number of updates for
each animal, and that these updates are synchronized using a per-data-element lock.
How can this counting best be done?
Of course, any number of counting algorithms from Chapter 5 might qualify, but
the optimal approach is quite simple. Just place a counter in each data element, and
increment it under the protection of that element’s lock!
If readers access the count locklessly, then updaters should use WRITE_ONCE() to
update the counter and lockless readers should use READ_ONCE() to load it.
13.1.2 Counting Lookups

Suppose that Schrödinger also wants to count the number of lookups for each animal,
where lookups are protected by RCU. How can this counting best be done?
One approach would be to protect a lookup counter with the per-element lock, as
discussed in Section 13.1.1. Unfortunately, this would require all lookups to acquire
this lock, which would be a severe bottleneck on large systems.
421
v2023.06.11a
422 CHAPTER 13. PUTTING IT ALL TOGETHER
Another approach is to “just say no” to counting, following the example of the
noatime mount option. If this approach is feasible, it is clearly the best: After all,
nothing is faster than doing nothing. If the lookup count cannot be dispensed with, read
on!
Any of the counters from Chapter 5 could be pressed into service, with the statistical
counters described in Section 5.2 being perhaps the most common choice. However,
this results in a large memory footprint: The number of counters required is the number
of data elements multiplied by the number of threads.
If this memory overhead is excessive, then one approach is to keep per-core or
even per-socket counters rather than per-CPU counters, with an eye to the hash-table
performance results depicted in Figure 10.3. This will require that the counter increments
be atomic operations, especially for user-mode execution where a given thread could
migrate to another CPU at any time.
If some elements are looked up very frequently, there are a number of approaches that
batch updates by maintaining a per-thread log, where multiple log entries for a given
element can be merged. After a given log entry has a sufficiently large increment or
after sufficient time has passed, the log entries may be applied to the corresponding data
elements. Silas Boyd-Wickizer has done some work formalizing this notion [BW14].
13.2 Refurbish Reference Counting
Counting is the religion of this generation. It is its

hope and its salvation.
Gertrude Stein
Although reference counting is a conceptually simple technique, many devils hide in

the details when it is applied to concurrent software. After all, if the object was not
subject to premature disposal, there would be no need for the reference counter in the
first place. But if the object can be disposed of, what prevents disposal during the
reference-acquisition process itself?
There are a number of ways to refurbish reference counters for use in concurrent
software, including:
1. A lock residing outside of the object must be held while manipulating the reference
count.
2. The object is created with a non-zero reference count, and new references may be
acquired only when the current value of the reference counter is non-zero. If a
thread does not have a reference to a given object, it might seek help from another
thread that already has a reference.
3. In some cases, hazard pointers may be used as a drop-in replacement for reference
counters.
4. An existence guarantee is provided for the object, thus preventing it from being
freed while some other entity might be attempting to acquire a reference. Existence
guarantees are often provided by automatic garbage collectors, and, as is seen in
Sections 9.3 and 9.5, by hazard pointers and RCU, respectively.
v2023.06.11a
13.2. REFURBISH REFERENCE COUNTING 423
Table 13.1: Synchronizing Reference Counting
Release
Reference Hazard
Acquisition Locks RCU
Counts Pointers
Locks − CAM M CA
Reference
A AM M A
Counts
Hazard
M M M M
Pointers
RCU CA MCA M CA
5. A type-safety guarantee is provided for the object. An additional identity check

must be performed once the reference is acquired. Type-safety guarantees can
be provided by special-purpose memory allocators, for example, by the SLAB_
TYPESAFE_BY_RCU feature within the Linux kernel, as is seen in Section 9.5.
Of course, any mechanism that provides existence guarantees by definition also

provides type-safety guarantees. This results in four general categories of reference-
acquisition protection: Reference counting, hazard pointers, sequence locking, and
RCU.
Quick Quiz 13.1: Why not implement reference-acquisition using a simple compare-and-swap
operation that only acquires a reference if the reference counter is non-zero?
Given that the key reference-counting issue is synchronization between acquisition of

a reference and freeing of the object, we have nine possible combinations of mechanisms,
as shown in Table 13.1. This table divides reference-counting mechanisms into the
following broad categories:
1. Simple counting with neither atomic operations, memory barriers, nor alignment
constraints (“−”).
2. Atomic counting without memory barriers (“A”).
3. Atomic counting, with memory barriers required only on release (“AM”).
4. Atomic counting with a check combined with the atomic acquisition operation,
and with memory barriers required only on release (“CAM”).
5. Atomic counting with a check combined with the atomic acquisition operation
(“CA”).
6. Simple counting with a check combined with full memory barriers (“M”).
7. Atomic counting with a check combined with the atomic acquisition operation,
and with memory barriers also required on acquisition (“MCA”).
However, because all Linux-kernel atomic operations that return a value are defined
to contain memory barriers,1 all release operations contain memory barriers, and all
checked acquisition operations also contain memory barriers. Therefore, cases “CA”
1 With atomic_read() and ATOMIC_INIT() being the exceptions that prove the rule.
v2023.06.11a
Listing 13.1: Simple Reference-Count API

1 struct sref {
2 int refcount;
3 };
4
5 void sref_init(struct sref *sref)
6 {
7 sref->refcount = 1;
8 }
9
10 void sref_get(struct sref *sref)
11 {
12 sref->refcount++;
13 }
14
15 int sref_put(struct sref *sref,
16 void (*release)(struct sref *sref))
17 {
18 WARN_ON(release == NULL);
19 WARN_ON(release == (void (*)(struct sref *))kfree);
20
21 if (--sref->refcount == 0) {
22 release(sref);
23 return 1;
24 }
25 return 0;
26 }
and “MCA” are equivalent to “CAM”, so that there are sections below for only the
first four cases and the sixth case: “−”, “A”, “AM”, “CAM”, and “M”. Later sections
describe optimizations that can improve performance if reference acquisition and release
is very frequent, and the reference count need be checked for zero only very rarely.
13.2.1 Implementation of Reference-Counting Categories

Simple counting protected by locking (“−”) is described in Section 13.2.1.1, atomic
counting with no memory barriers (“A”) is described in Section 13.2.1.2, atomic
counting with acquisition memory barrier (“AM”) is described in Section 13.2.1.3,
and atomic counting with check and release memory barrier (“CAM”) is described in
Section 13.2.1.4. Use of hazard pointers is described in Section 9.3 on page 207 and in
Section 13.3.
13.2.1.1 Simple Counting
Simple counting, with neither atomic operations nor memory barriers, can be used
when the reference-counter acquisition and release are both protected by the same
lock. In this case, it should be clear that the reference count itself may be manipulated
non-atomically, because the lock provides any necessary exclusion, memory barriers,
atomic instructions, and disabling of compiler optimizations. This is the method of
choice when the lock is required to protect other operations in addition to the reference
count, but where a reference to the object must be held after the lock is released.
Listing 13.1 shows a simple API that might be used to implement simple non-atomic
reference counting—although simple reference counting is almost always open-coded
instead.
v2023.06.11a
13.2.1.2 Atomic Counting

Simple atomic counting may be used in cases where any CPU acquiring a reference
must already hold a reference. This style is used when a single CPU creates an object for
its own private use, but must allow for accesses from other CPUs, tasks, timer handlers,
and so on. Any CPU that hands the object off must first acquire a new reference on
behalf of the recipient on the one hand, or refrain from further accesses after the handoff
on the other. In the Linux kernel, the kref primitives are used to implement this style
of reference counting, as shown in Listing 13.2.2
Atomic counting is required in this case because locking does not protect all reference-
count operations, which means that two different CPUs might concurrently manipulate
the reference count. If normal increment and decrement were used, a pair of CPUs
might both fetch the reference count concurrently, perhaps both obtaining the value “3”.
If both of them increment their value, they will both obtain “4”, and both will store this
value back into the counter. Since the new value of the counter should instead be “5”,
one of the increments has been lost. Therefore, atomic operations must be used both for
counter increments and for counter decrements.
If releases are guarded by locking, hazard pointers, or RCU, memory barriers are
not required, but for different reasons. In the case of locking, the locks provide any
needed memory barriers (and disabling of compiler optimizations), and the locks also
prevent a pair of releases from running concurrently. In the case of hazard pointers
and RCU, cleanup will be deferred, and any needed memory barriers or disabling of
compiler optimizations will be provided by the hazard-pointers or RCU infrastructure.
Therefore, if two CPUs release the final two references concurrently, the actual cleanup
will be deferred until both CPUs have released their hazard pointers or exited their RCU
read-side critical sections, respectively.
Quick Quiz 13.2: Why isn’t it necessary to guard against cases where one CPU acquires a
reference just after another CPU releases the last reference?
The kref structure itself, consisting of a single atomic data item, is shown in lines 1–3
of Listing 13.2. The kref_init() function on lines 5–8 initializes the counter to
the value “1”. Note that the atomic_set() primitive is a simple assignment, the
name stems from the data type of atomic_t rather than from the operation. The
kref_init() function must be invoked during object creation, before the object has
been made available to any other CPU.
The kref_get() function on lines 10–14 unconditionally atomically increments the
counter. The atomic_inc() primitive does not necessarily explicitly disable compiler
optimizations on all platforms, but the fact that the kref primitives are in a separate
module and that the Linux kernel build process does no cross-module optimizations has
the same effect.
The kref_sub() function on lines 16–28 atomically decrements the counter, and if
the result is zero, line 24 invokes the specified release() function and line 25 returns,
informing the caller that release() was invoked. Otherwise, kref_sub() returns
zero, informing the caller that release() was not called.
2 As of Linux v4.10. Linux v4.11 introduced a refcount_t API that improves efficiency
weakly ordered platforms, but which is functionally equivalent to the atomic_t that it
replaced.
v2023.06.11a
Listing 13.2: Linux Kernel kref API

1 struct kref {
2 atomic_t refcount;
3 };
4
5 void kref_init(struct kref *kref)
6 {
7 atomic_set(&kref->refcount, 1);
8 }
9
10 void kref_get(struct kref *kref)
11 {
12 WARN_ON(!atomic_read(&kref->refcount));
13 atomic_inc(&kref->refcount);
14 }
15
16 static inline int
17 kref_sub(struct kref *kref, unsigned int count,
18 void (*release)(struct kref *kref))
19 {
20 WARN_ON(release == NULL);
21
22 if (atomic_sub_and_test((int) count,
23 &kref->refcount)) {
24 release(kref);
25 return 1;
26 }
27 return 0;
28 }
Quick Quiz 13.3: Suppose that just after the atomic_sub_and_test() on line 22 of
Listing 13.2 is invoked, that some other CPU invokes kref_get(). Doesn’t this result in that
other CPU now having an illegal reference to a released object?
Quick Quiz 13.4: Suppose that kref_sub() returns zero, indicating that the release()
function was not invoked. Under what conditions can the caller rely on the continued existence
of the enclosing object?
Quick Quiz 13.5: Why not just pass kfree() as the release function?
13.2.1.3 Atomic Counting With Release Memory Barrier

Atomic reference counting with release memory barriers is used by the Linux kernel’s
networking layer to track the destination caches that are used in packet routing. The
actual implementation is quite a bit more involved; this section focuses on the aspects
of struct dst_entry reference-count handling that matches this use case, shown in
Listing 13.3.3
The dst_clone() primitive may be used if the caller already has a reference to the
specified dst_entry, in which case it obtains another reference that may be handed
off to some other entity within the kernel. Because a reference is already held by the
caller, dst_clone() need not execute any memory barriers. The act of handing the
dst_entry to some other entity might or might not require a memory barrier, but if
such a memory barrier is required, it will be embedded in the mechanism used to hand
the dst_entry off.
The dst_release() primitive may be invoked from any environment, and the caller
might well reference elements of the dst_entry structure immediately prior to the
3 As of Linux v4.13. Linux v4.14 added a level of indirection to permit more compre-
hensive debugging checks, but the overall effect in the absence of bugs is identical.
v2023.06.11a
Listing 13.3: Linux Kernel dst_clone API

1 static inline
2 struct dst_entry * dst_clone(struct dst_entry * dst)
3 {
4 if (dst)
5 atomic_inc(&dst->__refcnt);
6 return dst;
7 }
8
9 static inline
10 void dst_release(struct dst_entry * dst)
11 {
12 if (dst) {
13 WARN_ON(atomic_read(&dst->__refcnt) < 1);
14 smp_mb__before_atomic_dec();
15 atomic_dec(&dst->__refcnt);
16 }
17 }
call to dst_release(). The dst_release() primitive therefore contains a memory

barrier on line 14 preventing both the compiler and the CPU from misordering accesses.
Please note that the programmer making use of dst_clone() and dst_release()
need not be aware of the memory barriers, only of the rules for using these two primitives.
13.2.1.4 Atomic Counting With Check and Release Memory Barrier

Consider a situation where the caller must be able to acquire a new reference to an
object to which it does not already hold a reference, but where that object’s existence is
guaranteed. The fact that initial reference-count acquisition can now run concurrently
with reference-count release adds further complications. Suppose that a reference-count
release finds that the new value of the reference count is zero, signaling that it is now
safe to clean up the reference-counted object. We clearly cannot allow a reference-count
acquisition to start after such clean-up has commenced, so the acquisition must include
a check for a zero reference count. This check must be part of the atomic increment
operation, as shown below.
Quick Quiz 13.6: Why can’t the check for a zero reference count be made in a simple “if”
statement with an atomic increment in its “then” clause?
The Linux kernel’s fget() and fput() primitives use this style of reference counting.
Simplified versions of these functions are shown in Listing 13.4.4
Line 4 of fget() fetches the pointer to the current process’s file-descriptor table, which
might well be shared with other processes. Line 6 invokes rcu_read_lock(), which
enters an RCU read-side critical section. The callback function from any subsequent
call_rcu() primitive will be deferred until a matching rcu_read_unlock() is
reached (line 10 or 14 in this example). Line 7 looks up the file structure corresponding
to the file descriptor specified by the fd argument, as will be described later. If there
is an open file corresponding to the specified file descriptor, then line 9 attempts to
atomically acquire a reference count. If it fails to do so, lines 10–11 exit the RCU
read-side critical section and report failure. Otherwise, if the attempt is successful,
lines 14–15 exit the read-side critical section and return a pointer to the file structure.
The fcheck_files() primitive is a helper function for fget(). Line 22 uses
rcu_dereference() to safely fetch an RCU-protected pointer to this task’s current
4 As of Linux v2.6.38. Additional O_PATH functionality was added in v2.6.39, refactoring
was applied in v3.14, and mmap_sem contention was reduced in v4.1.
v2023.06.11a
Listing 13.4: Linux Kernel fget/fput API

1 struct file *fget(unsigned int fd)
2 {
3 struct file *file;
4 struct files_struct *files = current->files;
5
6 rcu_read_lock();
7 file = fcheck_files(files, fd);
8 if (file) {
9 if (!atomic_inc_not_zero(&file->f_count)) {
11 return NULL;
12 }
13 }
15 return file;
16 }
17
18 struct file *
19 fcheck_files(struct files_struct *files, unsigned int fd)
20 {
21 struct file * file = NULL;
22 struct fdtable *fdt = rcu_dereference((files)->fdt);
23
24 if (fd < fdt->max_fds)
25 file = rcu_dereference(fdt->fd[fd]);
26 return file;
27 }
28
29 void fput(struct file *file)
30 {
31 if (atomic_dec_and_test(&file->f_count))
32 call_rcu(&file->f_u.fu_rcuhead, file_free_rcu);
33 }
34
35 static void file_free_rcu(struct rcu_head *head)
36 {
37 struct file *f;
38
39 f = container_of(head, struct file, f_u.fu_rcuhead);
40 kmem_cache_free(filp_cachep, f);
41 }
file-descriptor table, and line 24 checks to see if the specified file descriptor is in range. If
so, line 25 fetches the pointer to the file structure, again using the rcu_dereference()
primitive. Line 26 then returns a pointer to the file structure or NULL in case of failure.
The fput() primitive releases a reference to a file structure. Line 31 atomically
decrements the reference count, and, if the result was zero, line 32 invokes the call_
rcu() primitives in order to free up the file structure (via the file_free_rcu()
function specified in call_rcu()’s second argument), but only after all currently-
executing RCU read-side critical sections complete, that is, after an RCU grace period
has elapsed.
Once the grace period completes, the file_free_rcu() function obtains a pointer
to the file structure on line 39, and frees it on line 40.
This code fragment thus demonstrates how RCU can be used to guarantee existence
while an in-object reference count is being incremented.
13.2.2 Counter Optimizations

In some cases where increments and decrements are common, but checks for zero
are rare, it makes sense to maintain per-CPU or per-task counters, as was discussed
in Chapter 5. For example, see the paper on sleepable read-copy update (SRCU),
v2023.06.11a
13.3. HAZARD-POINTER HELPERS 429
which applies this technique to RCU [McK06]. This approach eliminates the need for
atomic instructions or memory barriers on the increment and decrement primitives,
but still requires that code-motion compiler optimizations be disabled. In addition, the
primitives such as synchronize_srcu() that check for the aggregate reference count
reaching zero can be quite slow. This underscores the fact that these techniques are
designed for situations where the references are frequently acquired and released, but
where it is rarely necessary to check for a zero reference count.
However, it is usually the case that use of reference counts requires writing (often
atomically) to a data structure that is otherwise read only. In this case, reference counts
are imposing expensive cache misses on readers.
It is therefore worthwhile to look into synchronization mechanisms that do not require
readers to write to the data structure being traversed. One possibility is the hazard
pointers covered in Section 9.3 and another is RCU, which is covered in Section 9.5.
13.3 Hazard-Pointer Helpers

It’s the little things that count, hundreds of them.
Cliff Shaw
This section looks at some issues that can be addressed with the help of hazard pointers.
In addition, hazard pointers can sometimes be used to address the issues called out in
Section 13.5, and vice versa.
13.3.1 Scalable Reference Count

Suppose a reference count is becoming a performance or scalability bottleneck. What
can you do?
One approach is to instead use hazard pointers.
There are some differences, perhaps most notably that with hazard pointers it is
extremely expensive to determine when the corresponding reference count has reached
zero.
One way to work around this problem is to split the load between reference counters
and hazard pointers. Each data element has a reference counter that tracks the number
of other data elements referencing this element on the one hand, and readers use hazard
pointers on the other.
Making this arrangement work both efficiently and correctly can be quite chal-
lenging, and so interested readers are invited to examine the UnboundedQueue and
ConcurrentHashMap data structures implemented in Folly open-source library.5
13.3.2 Long-Duration Accesses

Suppose a reader-writer-locking reader is holding the lock for so long that updates are
excessively delayed. If that reader can reasonably be converted to use reference counting
instead of reader-writer locking, but if performance and scalability considerations
prevent use of actual reference counters, then hazard pointers provides a scalable variant
of reference counting.
5 https://github.com/facebook/folly
v2023.06.11a
The key point is that where reader-writer locking readers block all updates for that
lock, hazard pointers instead simply hang onto the data that is actually needed, while
still allowing updates to proceed.
If the reader cannot be reasonably be converted to use reference counting, the tricks
in Section 13.5.8 might be helpful.
13.4 Sequence-Locking Specials
The girl who can’t dance says the band can’t play.
Yiddish proverb
This section looks at some special uses of sequence locks.
13.4.1 Dueling Sequence Locks

The classic sequence-locking use case enables a reader to see a consistent snapshot of a
small collection of variables, for example, calibration constants for timekeeping. This
works quite well in practice because calibration constants are rarely updated and, when
updated, are updated quickly. Readers therefore almost never need to retry.
However, if the updater is delayed during the update, readers will also be delayed.
Such delays might be due to interrupts, NMIs, or even virtual-CPU preemption.
One way to prevent updater delays from causing reader delays is to maintain two sets
of calibration constants. Each set is updated in turn, but frequently enough that readers
can make good use of either set. Each set has its own sequence lock (seqlock_t
structure).
The updater alternates between the two sets, so that an delayed updater delays readers
of at most one of the sets.
Each reader attempts to access the first set, but upon retry attempts to access the
second set. If the second set also forces a retry, the reader repeats starting again from
the first set. If the updater is stuck, only one of the two sets will force readers to retry,
and therefore readers will succeed as soon as they attempt to access the other set.
Quick Quiz 13.7: Why don’t all sequence-locking use cases replicate the data in this fashion?
13.4.2 Correlated Data Elements

Suppose we have a hash table where we need correlated views of two or more of the
elements. These elements are updated together, and we do not want to see an old
version of the first element along with new versions of the other elements. For example,
Schrödinger decided to add his extended family to his in-memory database along with
all his animals. Although Schrödinger understands that marriages and divorces do not
happen instantaneously, he is also a traditionalist. As such, he absolutely does not want
his database ever to show that the bride is now married, but the groom is not, and vice
versa. Plus, if you think Schrödinger is a traditionalist, you just try conversing with
some of his family members! In other words, Schrödinger wants to be able to carry out
a wedlock-consistent traversal of his database.
v2023.06.11a
13.4. SEQUENCE-LOCKING SPECIALS 431
One approach is to use sequence locks (see Section 9.4), so that wedlock-related
updates are carried out under the protection of write_seqlock(), while reads requiring
wedlock consistency are carried out within a read_seqbegin() / read_seqretry()
loop. Note that sequence locks are not a replacement for RCU protection: Sequence
locks protect against concurrent modifications, but RCU is still needed to protect against
concurrent deletions.
This approach works quite well when the number of correlated elements is small,
the time to read these elements is short, and the update rate is low. Otherwise, updates
might happen so quickly that readers might never complete. Although Schrödinger
does not expect that even his least-sane relatives will marry and divorce quickly enough
for this to be a problem, he does realize that this problem could well arise in other
situations. One way to avoid this reader-starvation problem is to have the readers use
the update-side primitives if there have been too many retries, but this can degrade
both performance and scalability. Another way to avoid starvation is to have multiple
sequence locks, in Schrödinger’s case, perhaps one per species.
In addition, if the update-side primitives are used too frequently, poor performance
and scalability will result due to lock contention. One way to avoid this is to maintain a
per-element sequence lock, and to hold both spouses’ locks when updating their marital
status. Readers can do their retry looping on either of the spouses’ locks to gain a stable
view of any change in marital status involving both members of the pair. This avoids
contention due to high marriage and divorce rates, but complicates gaining a stable view
of all marital statuses during a single scan of the database.
If the element groupings are well-defined and persistent, which marital status is hoped
to be, then one approach is to add pointers to the data elements to link together the
members of a given group. Readers can then traverse these pointers to access all the
data elements in the same group as the first one located.
This technique is used heavily in the Linux kernel, perhaps most notably in the dcache
subsystem [Bro15b]. Note that it is likely that similar schemes also work with hazard
pointers.
This approach provides sequential consistency to successful readers, each of which
will either see the effects of a given update or not, with any partial updates resulting in
a read-side retry. Sequential consistency is an extremely strong guarantee, incurring
equally strong restrictions and equally high overheads. In this case, we saw that readers
might be starved on the one hand, or might need to acquire the update-side lock on
the other. Although this works very well in cases where updates are infrequent, it
unnecessarily forces read-side retries even when the update does not affect any of the
data that a retried reader accesses. Section 13.5.4 therefore covers a much weaker
form of consistency that not only avoids reader starvation, but also avoids any form of
read-side retry. The next section instead presents a weaker form of consistency that can
be provided with much lower probabilities of reader starvation.
13.4.3 Atomic Move

Suppose that individual data elements are moved from one data structure to another, and
that readers look up only single data structures. However, when a data element moves,
readers must must never see it as being in both structures at the same time and must
also never see it as missing from both structures at the same time. At the same time, any
reader seeing the element in its new location must never subsequently see it in its old
location. In addition, the move may be implemented by inserting a new copy of the old
data element into the destination location.
v2023.06.11a
For example, consider a hash table that supports an atomic-to-readers rename

operation. Expanding on Schrödinger’s zoo, suppose that an animal’s name changes,
for example, each of the brides in Schrödinger’s traditionalist family might change their
last name to match that of their groom.
But changing their name might change the hash value, and might also require that
the bride’s element move from one hash chain to another. The consistency set forth
above requires that if a reader successfully looks up the new name, then any subsequent
lookup of the old name by that reader must result in failure. Similarly, if a reader’s
lookup of the old name results in lookup failure, then any subsequent lookup of the
new name by that reader must succeed. In short, a given reader should not see a bride
momentarily blinking out of existence, nor should that reader lookup a bride under her
new name and then later lookup that bride under her old name.
This consistency guarantee could be enforced with a single global sequence lock as
described in Section 13.4.2, but this can result in reader starvation even for readers that
are not looking up a bride who is currently undergoing a name change. This guarantee
could also be enforced by requiring that readers acquire a per-hash-chain lock, but
reviewing Figure 10.2 shows that this results in poor performance and scalabilty, even
for single-socket systems.
Another more reader-friendly way to implement this is to use RCU and to place a
sequence lock on each element. Readers looking up a given element act as sequence-lock
readers across their full set of accesses to that element. Note that these sequence-lock
operations will order each reader’s lookups.
Renaming an element can then proceed roughly as follows:
1. Acquire a global lock protecting rename operations.
2. Allocate and initialize a copy of the element with the new name.
3. Write-acquire the sequence lock on the element with the old name, which has the
side effect of ordering this acquisition with the following insertion. Concurrent
lookups of the old name will now repeatedly retry.
4. Insert the copy of the element with the new name. Lookups of the new name will
now succeed.
5. Execute smp_wmb() to order the prior insertion with the subsequent removal.
6. Remove the element with the old name. Concurrent lookups of the old name will
now fail.
7. Write-release the sequence lock if necessary, for example, if required by lock

dependency checkers.
8. Release the global lock.
Thus, readers looking up the old name will retry until the new name is available, at
which point their final retry will fail. Any subsequent lookups of the new name will
succeed. Any reader succeeding in looking up the new name is guaranteed that any
subsequent lookup of the old name will fail, perhaps after a series of retries.
Quick Quiz 13.8: Is it possible to write-acquire the sequence lock on the new element before
it is inserted instead of acquiring that of the old element before it is removed?
v2023.06.11a
13.5. RCU RESCUES 433
Quick Quiz 13.9: Is it possible to avoid the global lock?
It is of course possible to instead implement this procedure somewhat more efficiently

using simple flags. However, this can be thought of as a simplified variant of sequence
locking that relies on the fact that a given element’s sequence lock is never write-acquired
more than once.
13.4.4 Upgrade to Writer

As discussed in Section 9.5.4.9, RCU permits readers to upgrade to writers. This
capability can be quite useful when a reader scanning an RCU-protected data structure
notices that the current element needs to be updated. What happens when you try this
trick with sequence locking?
It turns out that this sequence-locking trick is actually used in the Linux kernel, for
example, by the sdma_flush() function in drivers/infiniband/hw/hfi1/sdma.c.
The effect is to doom the enclosing reader to retry. This trick is therefore used when the
reader detects some condition that requires a retry.
13.5 RCU Rescues
With great doubts comes great understanding, with

little doubts comes little understanding.
Chinese proverb
This section shows how to apply RCU to some examples discussed earlier in this book.
In some cases, RCU provides simpler code, in other cases better performance and
scalability, and in still other cases, both.
13.5.1 RCU and Per-Thread-Variable-Based Statistical Counters

Section 5.2.3 described an implementation of statistical counters that provided excellent
performance, roughly that of simple increment (as in the C ++ operator), and linear
scalability—but only for incrementing via inc_count(). Unfortunately, threads
needing to read out the value via read_count() were required to acquire a global
lock, and thus incurred high overhead and suffered poor scalability. The code for the
lock-based implementation is shown in Listing 5.4 on page 83.
Quick Quiz 13.10: Why on earth did we need that global lock in the first place?
13.5.1.1 Design
The hope is to use RCU rather than final_mutex to protect the thread traversal in
read_count() in order to obtain excellent performance and scalability from read_
count(), rather than just from inc_count(). However, we do not want to give
up any accuracy in the computed sum. In particular, when a given thread exits, we
absolutely cannot lose the exiting thread’s count, nor can we double-count it. Such
an error could result in inaccuracies equal to the full precision of the result, in other
words, such an error would make the result completely useless. And in fact, one of the
v2023.06.11a
purposes of final_mutex is to ensure that threads do not come and go in the middle
of read_count() execution.
Therefore, if we are to dispense with final_mutex, we will need to come up with
some other method for ensuring consistency. One approach is to place the total count
for all previously exited threads and the array of pointers to the per-thread counters into
a single structure. Such a structure, once made available to read_count(), is held
constant, ensuring that read_count() sees consistent data.
13.5.1.2 Implementation
Lines 1–4 of Listing 13.5 show the countarray structure, which contains a ->total
field for the count from previously exited threads, and a counterp[] array of pointers
to the per-thread counter for each currently running thread. This structure allows a
given execution of read_count() to see a total that is consistent with the indicated set
of running threads.
Lines 6–8 contain the definition of the per-thread counter variable, the global pointer
countarrayp referencing the current countarray structure, and the final_mutex
spinlock.
Lines 10–13 show inc_count(), which is unchanged from Listing 5.4.
Lines 15–31 show read_count(), which has changed significantly. Lines 22
and 29 substitute rcu_read_lock() and rcu_read_unlock() for acquisition and
release of final_mutex. Line 23 uses rcu_dereference() to snapshot the current
countarray structure into local variable cap. Proper use of RCU will guarantee that
this countarray structure will remain with us through at least the end of the current
RCU read-side critical section at line 29. Line 24 initializes sum to cap->total, which
is the sum of the counts of threads that have previously exited. Lines 25–27 add up the
per-thread counters corresponding to currently running threads, and, finally, line 30
returns the sum.
The initial value for countarrayp is provided by count_init() on lines 33–41.
This function runs before the first thread is created, and its job is to allocate and zero
the initial structure, and then assign it to countarrayp.
Lines 43–50 show the count_register_thread() function, which is invoked by
each newly created thread. Line 45 picks up the current thread’s index, line 47 acquires
final_mutex, line 48 installs a pointer to this thread’s counter, and line 49 releases
final_mutex.
Quick Quiz 13.11: Hey!!! Line 48 of Listing 13.5 modifies a value in a pre-existing
countarray structure! Didn’t you say that this structure, once made available to read_
count(), remained constant???
Lines 52–72 show count_unregister_thread(), which is invoked by each thread

just before it exits. Lines 58–62 allocate a new countarray structure, line 63 acquires
final_mutex and line 69 releases it. Line 64 copies the contents of the current
countarray into the newly allocated version, line 65 adds the exiting thread’s counter
to new structure’s ->total, and line 66 NULLs the exiting thread’s counterp[] array
element. Line 67 then retains a pointer to the current (soon to be old) countarray
structure, and line 68 uses rcu_assign_pointer() to install the new version of the
countarray structure. Line 70 waits for a grace period to elapse, so that any threads
that might be concurrently executing in read_count(), and thus might have references
to the old countarray structure, will be allowed to exit their RCU read-side critical
v2023.06.11a
Listing 13.5: RCU and Per-Thread Statistical Counters

1 struct countarray {
2 unsigned long total;
3 unsigned long *counterp[NR_THREADS];
4 };
5
7 struct countarray *countarrayp = NULL;
9
10 __inline__ void inc_count(void)
11 {
13 }
14
16 {
17 struct countarray *cap;
18 unsigned long *ctrp;
20 int t;
21
22 rcu_read_lock();
23 cap = rcu_dereference(countarrayp);
24 sum = cap->total;
26 ctrp = READ_ONCE(cap->counterp[t]);
27 if (ctrp != NULL) sum += READ_ONCE(*ctrp);
28 }
30 return sum;
31 }
32
34 {
35 countarrayp = malloc(sizeof(*countarrayp));
36 if (countarrayp == NULL) {
37 fprintf(stderr, "Out of memory\n");
39 }
40 memset(countarrayp, '\0', sizeof(*countarrayp));
41 }
42
44 {
46
48 countarrayp->counterp[idx] = &counter;
50 }
51
53 {
54 struct countarray *cap;
55 struct countarray *capold;
57
58 cap = malloc(sizeof(*countarrayp));
59 if (cap == NULL) {
60 fprintf(stderr, "Out of memory\n");
62 }
64 *cap = *countarrayp;
65 cap->total += counter;
66 cap->counterp[idx] = NULL;
67 capold = countarrayp;
68 rcu_assign_pointer(countarrayp, cap);
71 free(capold);
72 }
v2023.06.11a
sections, thus dropping any such references. Line 71 can then safely free the old
countarray structure.
Quick Quiz 13.12: Given the fixed-size counterp array, exactly how does this code avoid a
fixed upper bound on the number of threads???
13.5.1.3 Discussion
Quick Quiz 13.13: Wow! Listing 13.5 contains 70 lines of code, compared to only 42 in
Listing 5.4. Is this extra complexity really worth it?
Use of RCU enables exiting threads to wait until other threads are guaranteed
to be done using the exiting threads’ __thread variables. This allows the read_
count() function to dispense with locking, thereby providing excellent performance
and scalability for both the inc_count() and read_count() functions. However, this
performance and scalability come at the cost of some increase in code complexity. It is
hoped that compiler and library writers employ user-level RCU [Des09b] to provide
safe cross-thread access to __thread variables, greatly reducing the complexity seen
by users of __thread variables.
13.5.2 RCU and Counters for Removable I/O Devices

Section 5.4.6 showed a fanciful pair of code fragments for dealing with counting I/O
accesses to removable devices. These code fragments suffered from high overhead on
the fastpath (starting an I/O) due to the need to acquire a reader-writer lock.
This section shows how RCU may be used to avoid this overhead.
The code for performing an I/O is quite similar to the original, with an RCU read-side
critical section being substituted for the reader-writer lock read-side critical section in
the original:
1 rcu_read_lock();
2 if (removing) {
4 cancel_io();
5 } else {
6 add_count(1);
8 do_io();
9 sub_count(1);
10 }
The RCU read-side primitives have minimal overhead, thus speeding up the fastpath,
as desired.
The updated code fragment removing a device is as follows:
2 removing = 1;
3 sub_count(mybias);
6 while (read_count() != 0) {
7 poll(NULL, 0, 1);
8 }
9 remove_device();
v2023.06.11a
Here we replace the reader-writer lock with an exclusive spinlock and add a
synchronize_rcu() to wait for all of the RCU read-side critical sections to complete.
Because of the synchronize_rcu(), once we reach line 6, we know that all remaining
I/Os have been accounted for.
Of course, the overhead of synchronize_rcu() can be large, but given that device
removal is quite rare, this is usually a good tradeoff.
13.5.3 Array and Length

Suppose we have an RCU-protected variable-length array, as shown in Listing 13.6.
The length of the array ->a[] can change dynamically, and at any given time, its length
is given by the field ->length. Of course, this introduces the following race condition:
1. The array is initially 16 characters long, and thus ->length is equal to 16.
2. CPU 0 loads the value of ->length, obtaining the value 16.
3. CPU 1 shrinks the array to be of length 8, and assigns a pointer to a new 8-character
block of memory into ->a[].
4. CPU 0 picks up the new pointer from ->a[], and stores a new value into element
12. Because the array has only 8 characters, this results in a SEGV or (worse yet)
memory corruption.
How can we prevent this?

One approach is to make careful use of memory barriers, which are covered in
Chapter 15. This works, but incurs read-side overhead and, perhaps worse, requires use
of explicit memory barriers.
A better approach is to put the value and the array into the same structure, as shown in
Listing 13.7 [ACMS03]. Allocating a new array (foo_a structure) then automatically
provides a new place for the array length. This means that if any CPU picks up a
reference to ->fa, it is guaranteed that the ->length will match the ->a[].
1. The array is initially 16 characters long, and thus ->length is equal to 16.
2. CPU 0 loads the value of ->fa, obtaining a pointer to the structure containing the
value 16 and the 16-byte array.
3. CPU 0 loads the value of ->fa->length, obtaining the value 16.
Listing 13.6: RCU-Protected Variable-Length Array

1 struct foo {
2 int length;
3 char *a;
4 };
Listing 13.7: Improved RCU-Protected Variable-Length Array

1 struct foo_a {
2 int length;
3 char a[0];
4 };
5
6 struct foo {
7 struct foo_a *fa;
8 };
v2023.06.11a
Listing 13.8: Uncorrelated Measurement Fields

1 struct animal {
2 char name[40];
3 double age;
4 double meas_1;
5 double meas_2;
6 double meas_3;
7 char photo[0]; /* large bitmap. */
8 };
Listing 13.9: Correlated Measurement Fields

1 struct measurement {
2 double meas_1;
3 double meas_2;
4 double meas_3;
5 };
6
7 struct animal {
8 char name[40];
9 double age;
10 struct measurement *mp;
12 };
4. CPU 1 shrinks the array to be of length 8, and assigns a pointer to a new foo_a
structure containing an 8-character block of memory into ->fa.
5. CPU 0 picks up the new pointer from ->a[], and stores a new value into element
12. But because CPU 0 is still referencing the old foo_a structure that contains
the 16-byte array, all is well.
Of course, in both cases, CPU 1 must wait for a grace period before freeing the old
array.
A more general version of this approach is presented in the next section.
13.5.4 Correlated Fields

Suppose that each of Schödinger’s animals is represented by the data element shown
in Listing 13.8. The meas_1, meas_2, and meas_3 fields are a set of correlated
measurements that are updated periodically. It is critically important that readers see
these three values from a single measurement update: If a reader sees an old value
of meas_1 but new values of meas_2 and meas_3, that reader will become fatally
confused. How can we guarantee that readers will see coordinated sets of these three
values?6
One approach would be to allocate a new animal structure, copy the old structure into
the new structure, update the new structure’s meas_1, meas_2, and meas_3 fields, and
then replace the old structure with a new one by updating the pointer. This does guarantee
that all readers see coordinated sets of measurement values, but it requires copying a
large structure due to the ->photo[] field. This copying might incur unacceptably
large overhead.
Another approach is to impose a level of indirection, as shown in Listing 13.9 [McK04,
Section 5.3.4]. When a new measurement is taken, a new measurement structure is
6 This situation is similar to that described in Section 13.4.2, except that here readers
need only see a consistent view of a given single data element, not the consistent view of a
group of data elements that was required in that earlier section.
v2023.06.11a
allocated, filled in with the measurements, and the animal structure’s ->mp field is
updated to point to this new measurement structure using rcu_assign_pointer().
After a grace period elapses, the old measurement structure can be freed.
Quick Quiz 13.14: But cant’t the approach shown in Listing 13.9 result in extra cache misses,
in turn resulting in additional read-side overhead?
This approach enables readers to see correlated values for selected fields, but while
incurring minimal read-side overhead. This per-data-element consistency suffices in the
common case where a reader looks only at a single data element.
13.5.5 Update-Friendly Traversal

Suppose that a statistical scan of all elements in a hash table is required. For example,
Schrödinger might wish to compute the average length-to-weight ratio over all of his
animals.7 Suppose further that Schrödinger is willing to ignore slight errors due to
animals being added to and removed from the hash table while this statistical scan is
being carried out. What should Schrödinger do to control concurrency?
One approach is to enclose the statistical scan in an RCU read-side critical section.
This permits updates to proceed concurrently without unduly impeding the scan. In
particular, the scan does not block the updates and vice versa, which allows scan of hash
tables containing very large numbers of elements to be supported gracefully, even in the
face of very high update rates.
Quick Quiz 13.15: But how does this scan work while a resizable hash table is being resized?
In that case, neither the old nor the new hash table is guaranteed to contain all the elements in
the hash table!
13.5.6 Scalable Reference Count Two

Suppose a reference count is becoming a performance or scalability bottleneck. What
can you do?
One approach is to use per-CPU counters for each reference count, somewhat similar
to the algorithms in Chapter 5, in particular, the exact limit counters described in
Section 5.4. The need to switch between per-CPU and global modes for these counters
results either in expensive increments and decrements on the one hand (Section 5.4.1)
or in the use of POSIX signals on the other (Section 5.4.3).
Another approach is to use RCU to mediate the switch between per-CPU and global
counting modes. Each update is carried out within an RCU read-side critical section,
and each update checks a flag to determine whether to update the per-CPU counters on
the one hand or the global on the other. To switch modes, update the flag, wait for a
grace period, and then move any remaining counts from the per-CPU counters to the
global counter or vice versa.
The Linux kernel uses this RCU-mediated approach in its percpu_ref style of
reference counter. Code using this reference counter must initialize the percpu_
ref structure using percpu_ref_init(), which takes as arguments a pointer to the
structure, a pointer to a function to invoke when the reference count reaches zero, a
set of mode flags, and a set of kmalloc() GFP_ flags. After normal initialization, the
structure has one reference and is in per-CPU mode.
7 Why would such a quantity be useful? Beats me! But group statistics are often useful.
v2023.06.11a
The mode flags are usually zero, but can include the PERCPU_REF_INIT_ATOMIC
bit if the counter is to start in slow non-per-CPU (that is, atomic) mode. There is
also a PERCPU_REF_ALLOW_REINIT bit that allows a given percpu_ref counter to be
reused via a call to percpu_ref_reinit() without needing to be freed and reallocated.
Regardless of how the percpu_ref structure is initialized, percpu_ref_get() may be
used to acquire a reference and percpu_ref_put() may be used to release a reference.
When in per-CPU mode, the percpu_ref structure cannot determine whether or
not its value has reached zero. When such a determination is necessary, percpu_ref_
kill() may be invoked. This function switches the structure into atomic mode and
removes the initial reference installed by the call to percpu_ref_init(). Of course,
when in atomic mode, calls to percpu_ref_get() and percpu_ref_put() are quite
expensive, but percpu_ref_put() can tell when the value reaches zero.
Readers desiring more percpu_ref information are referred to the Linux-kernel
documentation and source code.
13.5.7 Retriggered Grace Periods

There is no call_rcu_cancel(), so once an rcu_head structure is passed to call_
rcu(), there is no calling it back. It must be left alone until the callback is invoked. In
the common case, this is as it should be because the rcu_head structure is on a one-way
journey to deallocation.
However, there are use cases that combine RCU and explicit open() and close()
calls. After a close() call, readers are not supposed to begin new accesses to the data
structure, but there might well be readers completing their traversal. This situation can
be handled in the usual manner: Wait for a grace period following the close() call
before freeing the data structures.
But what if open() is called before the grace period ends?
Again, there is no call_rcu_cancel(), so another approach is to set a flag that
is checked by the callback function, which can opt out of actually freeing anything.
Problem solved!
But what if open() and then another close() are both called before the grace period
ends?
One approach is to have a second value for the flag that causes the callback to requeue
itself.
But what if there is not only a open() and then another close(), but also another
open() before the grace period ends?
In this case, the callback needs to set state to reflect that last open() still being in
effect.
Continuing this line of thought leads us to the state machine shown in Figure 13.1.
The initial state is CLOSED and the operational state is OPEN. The diamond-shaped
arrowheads denote call_rcu() invocation, while the arrows labeled “CB” denote
callback invocation.
The normal path through this state machine traverses the states CLOSED, OPEN,
CLOSING (with an invocation of call_rcu()), and back to CLOSED once the callback
has been invoked. If open() is invoked before the grace period completes, the state
machine traverses the cycle OPEN, CLOSING (with an invocation of call_rcu()),
REOPENING, and back to OPEN once the callback has been invoked. If open() and
then close() are invoked before the grace period completes, the state machine traverses
the cycle OPEN, CLOSING (with an invocation of call_rcu()), REOPENING,
RECLOSING, and back to CLOSING once the callback has been invoked.
v2023.06.11a
open()
CLOSED OPEN
CB close() CB
open()
CLOSING REOPENING
CB close() open()
RECLOSING
Figure 13.1: Retrigger-Grace-Period State Machine
Given an indefinite alternating sequence of close() and open() invocations, the

state machine would traverse OPEN, and CLOSING (with an invocation of call_
rcu()), followed by alternating sojourns in the REOPENING and RECLOSING states.
Once the grace period ends, the state machine would transition to either of the CLOSING
or the OPEN state, depending on which of the RECLOSING or REOPENING states the
callback was invoked in.
Rough pseudocode of this state machine is shown in Listing 13.10. The five states
are shown on lines 1–5, the current state is held in rtrg_status on line 7, which is
protected by the lock on line 8.
The three CB transitions (emanating from states CLOSING, REOPENING, and
RECLOSING) are implemented by the close_cb() function shown on lines 11–26.
Line 15 invokes a user-supplied close_cleanup() to take any final cleanup actions
such as freeing memory when transitioning to the CLOSED state. Line 21 contains the
call_rcu() invocation that causes a later transition to the CLOSED state.
The open() function on lines 28–42 implements the transitions to the OPEN,
CLOSING, and REOPENING states, with line 40 invoking a do_open() function to
implement any allocation and initialization of any needed data structures.
The close() function on lines 44–58 implements the transitions to the CLOSING
and RECLOSING states, with line 56 invoking a do_close() function to take any
actions that might be required to finalize this transition, for example, causing later
read-only traversals to return errors. Line 49 contains the call_rcu() invocation that
causes a later transition to the CLOSED state.
This state machine and pseudocode shows how to get the effect of a call_rcu_
cancel() in those rare situations needing such semantics.
13.5.8 Long-Duration Accesses Two

Suppose a reader-writer-locking reader is holding the lock for so long that updates are
excessively delayed. Suppose further that this reader cannot reasonably be converted to
use reference counting (otherwise, see Section 13.3.2).
v2023.06.11a
Listing 13.10: Retriggering a Grace Period (Pseudocode)

1 #define RTRG_CLOSED 0
2 #define RTRG_OPEN 1
3 #define RTRG_CLOSING 2
4 #define RTRG_REOPENING 3
5 #define RTRG_RECLOSING 4
6
7 int rtrg_status;
8 DEFINE_SPINLOCK(rtrg_lock);
9 struct rcu_head rtrg_rh;
10
11 void close_cb(struct rcu_head *rhp)
12 {
13 spin_lock(rtrg_lock);
14 if (rtrg_status = RTRG_CLOSING) {
15 close_cleanup();
16 rtrg_status = RTRG_CLOSED;
17 } else if (rtrg_status == RTRG_REOPENING) {
18 rtrg_status = RTRG_OPEN;
19 } else if (rtrg_status == RTRG_RECLOSING) {
20 rtrg_status = RTRG_CLOSING;
21 call_rcu(&rtrg_rh, close_cb);
22 } else {
23 WARN_ON_ONCE(1);
24 }
25 spin_unlock(rtrg_lock);
26 }
27
28 int open(void)
29 {
31 if (rtrg_status == RTRG_CLOSED) {
32 rtrg_status = RTRG_OPEN;
33 } else if (rtrg_status == RTRG_CLOSING ||
34 rtrg_status == RTRG_RECLOSING) {
35 rtrg_status = RTRG_REOPENING;
36 } else {
38 return -EBUSY;
39 }
40 do_open();
42 }
43
44 int close(void)
45 {
47 if (rtrg_status == RTRG_OPEN) {
48 rtrg_status = RTRG_CLOSING;
49 call_rcu(&rtrg_rh, close_cb);
50 } else if (rtrg_status == RTRG_REOPENING) {
51 rtrg_status = RTRG_RECLOSING;
52 } else {
54 return -ENOENT;
55 }
56 do_close();
58 }
v2023.06.11a
If that reader can be reasonably converted to use RCU, that might solve the problem.
The reason is that RCU readers do not completely block updates, but rather block only
the cleanup portions of those updates (including memory reclamation). Therefore, if
the system has ample memory, converting the reader-writer lock to RCU may suffice.
However, converting to RCU does not always suffice. For example, the code might
traverse an extremely large linked data structure within a single RCU read-side critical
section, which might so greatly extend the RCU grace period that the system runs out of
memory. These situations can be handled in a couple of different ways: (1) Use SRCU
instead of RCU and (2) Acquire a reference to exit the RCU reader.
13.5.8.1 Use SRCU

In the Linux kernel, RCU is global. In other words, any long-running RCU reader
anywhere in the kernel will delay the current RCU grace period. If the long-running
RCU reader is traversing a small data structure, that small amount of data is delaying
freeing of all other data structures, which can result in memory exhaustion.
One way to avoid this problem is to use SRCU for that long-running RCU reader’s
data structure, with its own srcu_struct structure. The resulting long-running SRCU
readers will then delay only that srcu_struct structure’s grace periods, and not those
of RCU, thus avoiding memory exhaustion. For more details, see the SRCU API in
Section 9.5.3.
Unfortunately, this approach does have some drawbacks. For one thing, SRCU readers
are not subject to priority boosting, which can result in additional delays to low-priority
SRCU readers on busy systems. Worse yet, defining a separate srcu_struct structure
reduces the number of RCU updaters, which in turn increases the grace-period overhead
per updater. This means that giving each current Linux-kernel RCU use case its
own srcu_struct structure could multiply system-wide grace-period overhead by the
number of such structures.
Therefore, it is often better to acquire some sort of non-RCU reference on the needed
data to permit a momentary exit from the RCU read-side critical section, as described
in the next section.
13.5.8.2 Acquire a Reference

If the RCU read-side critical section is too long, shorten it!
In some cases, this can be done trivially. For example, code that scans all of the hash
chains of a statically allocated array of hash buckets can just as easily scan each hash
chain within its own critical section.
This works because hash chains are normally quite short, and by design. When
traversing long linked structures, it is necessary to have some way of stopping in the
middle and resuming later.
For example, in Linux kernel v5.16, the khugepaged_scan_file() function checks
to see if some other task needs the current CPU using need_resched(), and if so
invokes xas_pause() to adjust the traversal’s iterator appropriately, and then invokes
cond_resched_rcu() to yield the CPU. In turn, the cond_resched_rcu() function
invokes rcu_read_unlock(), cond_resched(), and finally rcu_read_lock() to
drop out of the RCU read-side critical section in order to yield the CPU.
Of course, where feasible, another approach would be to switch to a data structure
such as a hash table that is more friendly to momentarily dropping out of an RCU
read-side critical section.
v2023.06.11a
Quick Quiz 13.16: But how would this work with a resizable hash table, such as the one
described in Section 10.4?
v2023.06.11a
If a little knowledge is a dangerous thing, just think
what you could do with a lot of knowledge!
Unknown
Chapter 14
Advanced Synchronization
This chapter covers synchronization techniques used for lockless algorithms and parallel
real-time systems.
Although lockless algorithms can be quite helpful when faced with extreme require-
ments, they are no panacea. For example, as noted at the end of Chapter 5, you should
thoroughly apply partitioning, batching, and well-tested packaged weak APIs (see
Chapters 8 and 9) before even thinking about lockless algorithms.
But after doing all that, you still might find yourself needing the advanced techniques
described in this chapter. To that end, Section 14.1 summarizes techniques used
thus far for avoiding locks and Section 14.2 gives a brief overview of non-blocking
synchronization. Memory ordering is also quite important, but it warrants its own
chapter, namely Chapter 15.
The second form of advanced synchronization provides the stronger forward-progress
guarantees needed for parallel real-time computing, which is the topic of Section 14.3.
14.1 Avoiding Locks
We are confronted with insurmountable

opportunities.
Walt Kelly
Although locking is the workhorse of parallelism in production, in many situations

performance, scalability, and real-time response can all be greatly improved through use
of lockless techniques. A particularly impressive example of such a lockless technique
is the statistical counters described in Section 5.2, which avoids not only locks, but
also read-modify-write atomic operations, memory barriers, and even cache misses for
counter increments. Other examples we have covered include:
1. The fastpaths through a number of other counting algorithms in Chapter 5.
2. The fastpath through resource allocator caches in Section 6.4.3.
3. The maze solver in Section 6.5.
4. The data-ownership techniques in Chapter 8.
445
v2023.06.11a
446 CHAPTER 14. ADVANCED SYNCHRONIZATION
5. The reference-counting, hazard-pointer, and RCU techniques in Chapter 9.

6. The lookup code paths in Chapter 10.
7. Many of the techniques in Chapter 13.
In short, lockless techniques are quite useful and are heavily used. However, it is
best if lockless techniques are hidden behind a well-defined API, such as the inc_
count(), memblock_alloc(), rcu_read_lock(), and so on. The reason for this is
that undisciplined use of lockless techniques is a good way to create difficult bugs. If
you believe that finding and fixing such bugs is easier than avoiding them, please re-read
Chapters 11 and 12.
14.2 Non-Blocking Synchronization

Never worry about theory as long as the machinery
does what it’s supposed to do.
Robert A. Heinlein
The term non-blocking synchronization (NBS) [Her90] describes eight classes of

linearizable algorithms with differing forward-progress guarantees [ACHS13], which
are as follows:
1. Bounded population-oblivious wait-free synchronization: Every thread will make

progress within a specific finite period of time, where this period of time is
independent of the number of threads [HS08]. This level is widely considered to
be even less achievable than bounded wait-free synchronization.
2. Bounded wait-free synchronization: Every thread will make progress within a
specific finite period of time [Her91]. This level is widely considered to be
unachievable, which might be why Alitarh et al. omitted it [ACHS13].
3. Wait-free synchronization: Every thread will make progress in finite time [Her93].
4. Lock-free synchronization: At least one thread will make progress in finite
time [Her93].
5. Obstruction-free synchronization: Every thread will make progress in finite time
in the absence of contention [HLM03].
6. Clash-free synchronization: At least one thread will make progress in finite time
in the absence of contention [ACHS13].
7. Starvation-free synchronization: Every thread will make progress in finite time in
the absence of failures [ACHS13].
8. Deadlock-free synchronization: At least one thread will make progress in finite
time in the absence of failures [ACHS13].
NBS class 1 was formulated some time before 2015, classes 2, 3, and 4 were first
formulated in the early 1990s, class 5 was first formulated in the early 2000s, and class 6
was first formulated in 2013. The final two classes have seen informal use for a great
many decades, but were reformulated in 2013.
v2023.06.11a
14.2. NON-BLOCKING SYNCHRONIZATION 447
Quick Quiz 14.1: Given that there will always be a sharply limited number of CPUs available,
is population obliviousness really useful?
In theory, any parallel algorithm can be cast into wait-free form, but there are a
relatively small subset of NBS algorithms that are in common use. A few of these are
listed in the following section.
14.2.1 Simple NBS

Perhaps the simplest NBS algorithm is atomic update of an integer counter using
fetch-and-add (atomic_add_return()) primitives. This section lists a few additional
commonly used NBS algorithms in roughly increasing order of complexity.
14.2.1.1 NBS Sets

One simple NBS algorithm implements a set of integers in an array. Here the array
index indicates a value that might be a member of the set and the array element indicates
whether or not that value actually is a set member. The linearizability criterion for
NBS algorithms requires that reads from and updates to the array either use atomic
instructions or be accompanied by memory barriers, but in the not-uncommon case
where linearizability is not important, simple volatile loads and stores suffice, for
example, using READ_ONCE() and WRITE_ONCE().
An NBS set may also be implemented using a bitmap, where each value that might be
a member of the set corresponds to one bit. Reads and updates must normally be carried
out via atomic bit-manipulation instructions, although compare-and-swap (cmpxchg()
or CAS) instructions can also be used.
14.2.1.2 NBS Counters

The statistical counters algorithm discussed in Section 5.2 can be considered to be
bounded-wait-free, but only by using a cute definitional trick in which the sum is
considered to be approximate rather than exact.1 Given sufficiently wide error bounds
that are a function of the length of time that the read_count() function takes to sum
the counters, it is not possible to prove that any non-linearizable behavior occurred. This
definitely (if a bit artificially) classifies the statistical-counters algorithm as bounded
wait-free. This algorithm is probably the most heavily used NBS algorithm in the Linux
kernel.
14.2.1.3 Half-NBS Queue

Another common NBS algorithm is the atomic queue where elements are enqueued
using an atomic exchange instruction [MS98b], followed by a store into the ->next
pointer of the new element’s predecessor, as shown in Listing 14.1, which shows the
userspace-RCU library implementation [Des09b]. Line 9 updates the tail pointer to
reference the new element while returning a reference to its predecessor, which is stored
in local variable old_tail. Line 10 then updates the predecessor’s ->next pointer
to reference the newly added element, and finally line 11 returns an indication as to
whether or not the queue was initially empty.
1 Citation needed. I heard of this trick verbally from Mark Moir.
v2023.06.11a
Listing 14.1: NBS Enqueue Algorithm

1 static inline bool
2 ___cds_wfcq_append(struct cds_wfcq_head *head,
3 struct cds_wfcq_tail *tail,
4 struct cds_wfcq_node *new_head,
5 struct cds_wfcq_node *new_tail)
6 {
7 struct cds_wfcq_node *old_tail;
8
9 old_tail = uatomic_xchg(&tail->p, new_tail);
10 CMM_STORE_SHARED(old_tail->next, new_head);
11 return old_tail != &head->node;
12 }
13
14 static inline bool
15 _cds_wfcq_enqueue(struct cds_wfcq_head *head,
16 struct cds_wfcq_tail *tail,
17 struct cds_wfcq_node *new_tail)
18 {
19 return ___cds_wfcq_append(head, tail,
20 new_tail, new_tail);
21 }
Although mutual exclusion is required to dequeue a single element (so that dequeue
is blocking), it is possible to carry out a non-blocking removal of the entire contents
of the queue. What is not possible is to dequeue any given element in a non-blocking
manner: The enqueuer might have failed between lines 9 and 10 of the listing, so that
the element in question is only partially enqueued. This results in a half-NBS algorithm
where enqueues are NBS but dequeues are blocking. This algorithm is nevertheless
heavily used in practice, in part because most production software is not required to
tolerate arbitrary fail-stop errors.
Quick Quiz 14.2: Wait! In order to dequeue all elements, both the ->head and ->tail
pointers must be changed, which cannot be done atomically on typical computer systems. So
how is this supposed to work???
14.2.1.4 NBS Stack

Listing 14.2 shows the LIFO push algorithm, which boasts lock-free push and bounded
wait-free pop (lifo-push.c), forming an NBS stack. The origins of this algorithm are
unknown, but it was referred to in a patent granted in 1975 [BS75]. This patent was
filed in 1973, a few months before your editor saw his first computer, which had but one
CPU.
Lines 1–4 show the node_t structure, which contains an arbitrary value and a pointer
to the next structure on the stack and line 7 shows the top-of-stack pointer.
The list_push() function spans lines 9–20. Line 11 allocates a new node and
line 14 initializes it. Line 17 initializes the newly allocated node’s ->next pointer, and
line 18 attempts to push it on the stack. If line 19 detects cmpxchg() failure, another
pass through the loop retries. Otherwise, the new node has been successfully pushed,
and this function returns to its caller. Note that line 19 resolves races in which two
concurrent instances of list_push() attempt to push onto the stack. The cmpxchg()
will succeed for one and fail for the other, causing the other to retry, thereby selecting
an arbitrary order for the two node on the stack.
The list_pop_all() function spans lines 23–34. The xchg() statement on line 25
atomically removes all nodes on the stack, placing the head of the resulting list in local
variable p and setting top to NULL. This atomic operation serializes concurrent calls to
v2023.06.11a
Listing 14.2: NBS Stack Algorithm

1 struct node_t {
2 value_t val;
3 struct node_t *next;
4 };
5
6 // LIFO list structure
7 struct node_t* top;
8
9 void list_push(value_t v)
10 {
11 struct node_t *newnode = malloc(sizeof(*newnode));
12 struct node_t *oldtop;
13
14 newnode->val = v;
15 oldtop = READ_ONCE(top);
16 do {
17 newnode->next = oldtop;
18 oldtop = cmpxchg(&top, newnode->next, newnode);
19 } while (newnode->next != oldtop);
20 }
21
22
23 void list_pop_all(void (foo)(struct node_t *p))
24 {
25 struct node_t *p = xchg(&top, NULL);
26
27 while (p) {
28 struct node_t *next = p->next;
29
30 foo(p);
31 free(p);
32 p = next;
33 }
34 }
list_pop_all(): One of them will get the list, and the other a NULL pointer, at least
assuming that there were no concurrent calls to list_push().
An instance of list_pop_all() that obtains a non-empty list in p processes this
list in the loop spanning lines 27–33. Line 28 prefetches the ->next pointer, line 30
invokes the function referenced by foo() on the current node, line 31 frees the current
node, and line 32 sets up p for the next pass through the loop.
But suppose that a pair of list_push() instances run concurrently with a list_
pop_all() with a list initially containing a single Node 𝐴. Here is one way that this
scenario might play out:
1. The first list_push() instance pushes a new Node 𝐵, executing through line 17,
having just stored a pointer to Node 𝐴 into Node 𝐵’s ->next pointer.
2. The list_pop_all() instance runs to completion, setting top to NULL and
freeing Node 𝐴.
3. The second list_push() instance runs to completion, pushing a new Node 𝐶,
but happens to allocate the memory that used to belong to Node 𝐴.
4. The first list_push() instance executes the cmpxchg() on line 18. Because new
Node 𝐶 has the same address as the newly freed Node 𝐴, this cmpxchg() succeeds
and this list_push() instance runs to completion.
Note that both pushes and the popall all ran successfully despite the reuse of Node 𝐴’s
memory. This is an unusual property: Most data structures require protection against
what is often called the ABA problem.
v2023.06.11a
But this property holds only for algorithm written in assembly language. The sad fact
is that most languages (including C and C++) do not support pointers to lifetime-ended
objects, such as the pointer to the old Node 𝐴 contained in Node 𝐵’s ->next pointer.
In fact, compilers are within their rights to assume that if two pointers (call them p and
q) were returned from two different calls to malloc(), then those pointers must not be
equal. Real compilers really will generate the constant false in response to a p==q
comparison. A pointer to an object that has been freed, but whose memory has been
reallocated for a compatibly typed object is termed a zombie pointer.
Many concurrent applications avoid this problem by carefully hiding the memory
allocator from the compiler, thus preventing the compiler from making inappropriate
assumptions. This obfuscatory approach currently works in practice, but might well one
day fall victim to increasingly aggressive optimizers. There is work underway in both
the C and C++ standards committees to address this problem [MMS19, MMM+ 20]. In
the meantime, please exercise great care when coding ABA-tolerant algorithms.
Quick Quiz 14.3: So why not ditch antique languages like C and C++ for something more
modern?
14.2.2 Applicability of NBS Benefits

The most heavily cited NBS benefits stem from its forward-progress guarantees, its
tolerance of fail-stop bugs, and from its linearizability. Each of these is discussed in
one of the following sections.
14.2.2.1 NBS Forward Progress Guarantees

NBS’s forward-progress guarantees have caused many to suggest its use in real-time
systems, and NBS algorithms are in fact used in a great many such systems. However, it
is important to note that forward-progress guarantees are largely orthogonal to those
that form the basis of real-time programming:
1. Real-time forward-progress guarantees usually have some definite time associated

with them, for example, “scheduling latency must be less than 100 microseconds.”
In contrast, the most popular forms of NBS only guarantees that progress will be
made in finite time, with no definite bound.
2. Real-time forward-progress guarantees are often probabilistic, as in the soft-real-

time guarantee that “at least 99.9 % of the time, scheduling latency must be less
than 100 microseconds.” In contrast, many of NBS’s forward-progress guarantees
are unconditional.
3. Real-time forward-progress guarantees are often conditioned on environmental

constraints, for example, only being honored: (1) For the highest-priority tasks,
(2) When each CPU spends at least a certain fraction of its time idle, and (3) When
I/O rates are below some specified maximum. In contrast, NBS’s forward-progress
guarantees are often unconditional, although recent NBS work accommodates
conditional guarantees [ACHS13].
4. An important component of a real-time program’s environment is the scheduler.

NBS algorithms assume a worst-case demonic scheduler, though for whatever
reason, not a scheduler so demonic that it simply refuses to ever run the application
v2023.06.11a
housing the NBS algorithm. In contrast, real-time systems assume that the
scheduler is doing its level best to satisfy any scheduling constraints it knows about,
and, in the absence of such constraints, its level best to honor process priorities
and to provide fair scheduling to processes of the same priority. Non-demonic
schedulers allow real-time programs to use simpler algorithms than those required
for NBS [ACHS13, Bra11].
5. NBS forward-progress guarantee classes assume that a number of underlying

operations are lock-free or even wait-free, when in fact these operations are
blocking on common-case computer systems.
6. NBS forward-progress guarantees are often achieved by subdividing operations.

For example, in order to avoid a blocking dequeue operation, an NBS algorithm
might substitute a non-blocking polling operation. This is fine in theory, but not
helpful in practice to real-world programs that require an element to propagate
through the queue in a timely fashion.
7. Real-time forward-progress guarantees usually apply only in the absence of software

bugs. In contrast, many classes of NBS guarantees apply even in the face of fail-stop
bugs.
8. NBS forward-progress guarantee classes imply linearizability. In contrast, real-time

forward progress guarantees are often independent of ordering constraints such as
linearizability.
Quick Quiz 14.4: Why does anyone care about demonic schedulers?
To reiterate, despite these differences, a number of NBS algorithms are extremely

useful in real-time programs.
14.2.2.2 NBS Underlying Operations

An NBS algorithm can be truly non-blocking only if the underlying operations that
it uses are also non-blocking. In a surprising number of cases, this is not the case in
practice.
For example, non-blocking algorithms often allocate memory. In theory, this is fine,
given the existence of lock-free memory allocators [Mic04b]. But in practice, most
environments must eventually obtain memory from operating-system kernels, which
commonly use locking. Therefore, unless all the memory that will ever be needed is
somehow preallocated, a “non-blocking” algorithm that allocates memory will not be
non-blocking when running on common-case real-world computer systems.
This same point clearly also applies to algorithms performing I/O operations or
otherwise interacting with their environment.
Perhaps surprisingly, this point also applies to ostensibly non-blocking algorithms
that do only plain loads and stores, such as the counters discussed in Section 14.2.1.2.
And at first glance, those loads and stores that can be compiled into single load and
store instructions, respectively, would seem to be not just non-blocking, but bounded
population-oblivious wait free.
Except that load and store instructions are not necessarily either fast or deterministic.
For example, as noted in Chapter 3, cache misses can consume thousands of CPU cycles.
Worse yet, the measured cache-miss latencies can be a function of the number of CPUs,
v2023.06.11a
as illustrated in Figure 5.1. It is only reasonable to assume that these latencies also
depend on the details of the system’s interconnect. In addition, given that hardware
vendors generally do not publish upper bounds for cache-miss latencies, it seems brave
to assume that memory-reference instructions are in fact wait-free in modern computer
systems. And the antique systems for which such bounds are available suffer from
profound overall slowness.
Furthermore, hardware is not the only source of slowness for memory-reference
instructions. For example, when running on typical computer systems, both loads
and stores can result in page faults. Which cause in-kernel page-fault handlers to be
invoked. Which might acquire locks, or even do I/O, potentially even using something
like network file system (NFS). All of which are most emphatically blocking operations.
Nor are page faults the only kernel-induced hazard. A given CPU might be interrupted
at any time, and the interrupt handler might run for some time. During this time, the
user-mode ostensibly non-blocking algorithm will not be running at all. This situation
raises interesting questions about the forward-progress guarantees provided by system
calls relying on interrupts, for example, the membarrier() system call.
Things do look bleak, but the non-blocking nature of such algorithms can be at least
partially redeemed using a number of approaches:
1. Run on bare metal, with paging disabled. If you are both brave and confident that
you can write code that is free of wild-pointer bugs, this approach might be for you.
2. Run on a non-blocking operating-system kernel [GC96]. Such kernels are quite
rare, in part because they have traditionally completely failed to provide the
hoped-for performance and scalability advantages over lock-based kernels. But
perhaps you should write one.
3. Use facilities such as mlockall() to avoid page faults, while also ensuring that
your program preallocates all the memory it will ever need at boot time. This can
work well, but at the expense of severe common-case underutilization of memory.
In environments that are cost-constrained or power-limited, this approach is not
likely to be feasible.
4. Use facilities such as the Linux kernel’s NO_HZ_FULL tickless mode [Cor13]. In
recent versions of the Linux kernel, this mode directs interrupts away from a
designated set of CPUs. However, this can sharply limit throughput for applications
that are I/O bound during even part of their operation.
Given these considerations, it is no surprise that non-blocking synchronization is far

more important in theory than it is in practice.
14.2.2.3 NBS Subdivided Operations

One common trick that provides a given algorithm a loftier place on the NBS ranking
is to replace blocking operations with a polling API. For example, instead of having
a reliable dequeue operation that might be merely lock-free or even blocking, instead
provide a dequeue operation that will spuriously fail in a wait-free manner rather than
exhibiting dreaded lock-free or blocking behaviors.
This can work well in theory, but a common effect in practice is to merely move the
lock-free or blocking behavior out of that specific algorithm and into the hapless code
making use of that algorithm. In such cases, not only has nothing has been gained by
this trick, but this trick has increased the complexity of all users of this algorithm.
v2023.06.11a
With concurrent algorithms as elsewhere, maximizing a specific metric is no substitute

for thinking carefully about the needs of one’s users.
14.2.2.4 NBS Fail-Stop Tolerance

Of the classes of NBS algorithms, wait-free synchronization (bounded or otherwise), lock-
free synchronization, obstruction-free synchronization, and clash-free synchronization
guarantee forward progress even in the presence of fail-stop bugs. An example fail-stop
bug might cause some thread to be preempted indefinitely. As we will see, this fail-stop-
tolerant property can be useful, but the fact is that composing a set of fail-stop-tolerant
mechanisms does not necessarily result in a fail-stop-tolerant system. To see this,
consider a system made up of a series of wait-free queues, where an element is removed
from one queue in the series, processed, and then added to the next queue.
If a thread is preempted in the midst of a queuing operation, in theory all is well
because the wait-free nature of the queue will guarantee forward progress. But in
practice, the element being processed is lost because the fail-stop-tolerant nature of the
wait-free queues does not extend to the code using those queues.
Nevertheless, there are a few applications where NBS’s rather limited fail-stop-
tolerance is useful. For example, in some network-based or web applications, a fail-stop
event will eventually result in a retransmission, which will restart any work that was lost
due to the fail-stop event. Systems running such applications can therefore be heavily
loaded, even to the point where the scheduler can no longer provide any reasonable
fairness guarantee. In constrast, if a thread fail-stops while holding a lock, the application
might need to be restarted. Nevertheless, NBS is not a panacea even within this restricted
area, due to the possibility of spurious retransmissions due to pure scheduling delays.
In some cases, it may be more efficient to reduce the load to avoid queueing delays,
which will also improve the scheduler’s ability to provide fair access, reducing or even
eliminating the fail-stop events, thus reducing the number of retry operations, in turn
further reducing the load.
14.2.2.5 NBS Linearizability

It is important to note that linearizability can be quite useful, especially when analyzing
concurrent code made up of strict locking and fully ordered atomic operations.2
Furthermore, this handling of fully ordered atomic operations automatically covers
simple NBS algorithms.
However, the linearization points of a complex NBS algorithms are often buried deep
within that algorithm, and thus not visible to users of a library function implementing
a part of such an algorithm. Therefore, any claims that users benefit from the
linearizability properties of complex NBS algorithms should be regarded with deep
suspicion [HKLP12].
It is sometimes asserted that linearizability is necessary for developers to produce
proofs of correctness for their concurrent code. However, such proofs are the exception
rather than the rule, and modern developers who do produce proofs often use modern
proof techniques that do not depend on linearizability. Furthermore, developers
frequently use modern proof techniques that do not require a full specification, given
2 For example, the Linux kernel’s value-returning atomic operations.
v2023.06.11a
that developers often learn their specification after the fact, one bug at a time. A few
such proof techniques were discussed in Chapter 12.3
It is often asserted that linearizability maps well to sequential specifications, which
are said to be more natural than are concurrent specifications [RR20]. But this assertion
fails to account for our highly concurrent objective universe. This universe can only be
expected to select for ability to cope with concurrency, especially for those participating
in team sports or overseeing small children. In addition, given that the teaching of
sequential computing is still believed to be somewhat of a black art [PBCE20], it is
reasonable to expect that teaching of concurrent computing is in a similar state of
disarray. Therefore, focusing on only one proof technique is unlikely to be a good way
forward.
Again, please understand that linearizability is quite useful in many situations. Then
again, so is that venerable tool, the hammer. But there comes a point in the field of
computing where one should put down the hammer and pick up a keyboard. Similarly,
it appears that there are times when linearizability is not the best tool for the job.
To their credit, there are some linearizability advocates who are aware of some of its
shortcomings [RR20]. There are also proposals to extend linearizability, for example,
interval-linearizability, which is intended to handle the common case of operations
that require non-zero time to complete [CnRR18]. It remains to be seen whether these
proposals will result in theories able to handle modern concurrent software artifacts,
especially given that several of the proof techniques discussed in Chapter 12 already
handle many modern concurrent software artifacts.
14.2.3 NBS Discussion

It is possible to create fully non-blocking queues [MS96], however, such queues are
much more complex than the half-NBS algorithm outlined above. The lesson here is to
carefully consider your actual requirements. Relaxing irrelevant requirements can often
result in great improvements in simplicity, performance, and scalability.
Recent research points to another important way to relax requirements. It turns
out that systems providing fair scheduling can enjoy most of the benefits of wait-
free synchronization even when running algorithms that provide only non-blocking
synchronization, both in theory [ACHS13] and in practice [AB13]. Because most
schedulers used in production do in fact provide fairness, the more-complex algorithms
providing wait-free synchronization usually provide no practical advantages over simpler
and faster non-wait-free algorithms.
Interestingly enough, fair scheduling is but one beneficial constraint that is often
respected in practice. Other sets of constraints can permit blocking algorithms to
achieve deterministic real-time response. For example, given: (1) Fair locks granted
in FIFO order within a given priority level, (2) Priority inversion avoidance (for
example, priority inheritance [TS95, WTS96] or priority ceiling), (3) A bounded
number of threads, (4) Bounded critical section durations, (5) Bounded load, and
(6) Absence of fail-stop bugs, lock-based applications can provide deterministic response
times [Bra11, SM04a]. This approach of course blurs the distinction between blocking
3 A memorable verbal discussion with an advocate of linearizability resulted in question:
“So the reason linearizability is important is to rescue 1980s proof techniques?” The advocate
immediately replied in the affirmative, then spent some time disparaging a particular modern
proof technique. Oddly enough, that technique was one of those successfully applied to
Linux-kernel RCU.
v2023.06.11a
14.3. PARALLEL REAL-TIME COMPUTING 455
and wait-free synchronization, which is all to the good. Hopefully theoretical frameworks
will continue to improve their ability to describe software actually used in practice.
Those who feel that theory should lead the way are referred to the inimitable Peter
Denning, who said of operating systems: “Theory follows practice” [Den15], or to
the eminent Tony Hoare, who said of the whole of engineering: “In all branches of
engineering science, the engineering starts before the science; indeed, without the early
products of engineering, there would be nothing for the scientist to study!” [Mor07]. Of
course, once an appropriate body of theory becomes available, it is wise to make use of
it. However, note well that the first appropriate body of theory is often one thing and
the first proposed body of theory quite another.
Quick Quiz 14.5: It seems like the various members of the NBS hierarchy are rather useless.
So why bother with them at all???
Proponents of NBS algorithms sometimes call out real-time computing as an important

NBS beneficiary. The next section looks more deeply at the forward-progress needs of
real-time systems.
14.3 Parallel Real-Time Computing
One always has time enough if one applies it well.

Johann Wolfgang von Göthe
An important emerging area in computing is that of parallel real-time computing.

Section 14.3.1 looks at a number of definitions of “real-time computing”, moving
beyond the usual sound bites to more meaningful criteria. Section 14.3.2 surveys the
sorts of applications that need real-time response. Section 14.3.3 notes that parallel real-
time computing is upon us, and discusses when and why parallel real-time computing
can be useful. Section 14.3.4 gives a brief overview of how parallel real-time systems
may be implemented, with Sections 14.3.5 and 14.3.6 focusing on operating systems
and applications, respectively. Finally, Section 14.3.7 outlines how to decide whether or
not your application needs real-time facilities.
14.3.1 What is Real-Time Computing?

One traditional way of classifying real-time computing is into the categories of hard
real time and soft real time, where the macho hard real-time applications never miss
their deadlines, but the wimpy soft real-time applications miss their deadlines quite
often.
14.3.1.1 Soft Real Time

It should be easy to see problems with this definition of soft real time. For one thing, by
this definition, any piece of software could be said to be a soft real-time application:
“My application computes million-point Fourier transforms in half a picosecond.” “No
way!!! The clock cycle on this system is more than three hundred picoseconds!” “Ah,
but it is a soft real-time application!” If the term “soft real time” is to be of any use
whatsoever, some limits are clearly required.
v2023.06.11a
Figure 14.1: Real-Time Response, Meet Hammer
We might therefore say that a given soft real-time application must meet its response-
time requirements at least some fraction of the time, for example, we might say that it
must execute in less than 20 microseconds 99.9 % of the time.
This of course raises the question of what is to be done when the application fails
to meet its response-time requirements. The answer varies with the application, but
one possibility is that the system being controlled has sufficient stability and inertia
to render harmless the occasional late control action. Another possibility is that the
application has two ways of computing the result, a fast and deterministic but inaccurate
method on the one hand and a very accurate method with unpredictable compute time
on the other. One reasonable approach would be to start both methods in parallel, and
if the accurate method fails to finish in time, kill it and use the answer from the fast
but inaccurate method. One candidate for the fast but inaccurate method is to take no
control action during the current time period, and another candidate is to take the same
control action as was taken during the preceding time period.
In short, it does not make sense to talk about soft real time without some measure of
exactly how soft it is.
14.3.1.2 Hard Real Time

In contrast, the definition of hard real time is quite definite. After all, a given system
either always meets its deadlines or it doesn’t.
Unfortunately, a strict application of this definition would mean that there can never
be any hard real-time systems. The reason for this is fancifully depicted in Figure 14.1.
And although you can always construct a more robust system, perhaps with redundancy,
your adversary can always get a bigger hammer. But don’t take my word for it: Ask the
dinosaurs.
Then again, perhaps it is unfair to blame the software for what is clearly not just a
hardware problem, but a bona fide big-iron hardware problem at that.4 This suggests
that we define hard real-time software as software that will always meet its deadlines,
but only in the absence of a hardware failure. Unfortunately, failure is not always
4 Or, given modern hammers, a big-steel problem.
v2023.06.11a
Figure 14.2: Real-Time Response: Hardware Matters
Figure 14.3: Real-Time Response: Notification Insufficient
an option, as fancifully depicted in Figure 14.2. We simply cannot expect the poor
gentleman depicted in that figure to be reassured our saying “Rest assured that if a
missed deadline results in your tragic death, it most certainly will not have been due to
a software problem!” Hard real-time response is a property of the entire system, not
just of the software.
But if we cannot demand perfection, perhaps we can make do with notification, similar
to the soft real-time approach noted earlier. Then if the Life-a-Tron in Figure 14.2 is
about to miss its deadline, it can alert the hospital staff.
Unfortunately, this approach has the trivial solution fancifully depicted in Figure 14.3.
A system that always immediately issues a notification that it won’t be able to meet its
deadline complies with the letter of the law, but is completely useless. There clearly
must also be a requirement that the system meets its deadline some fraction of the
time, or perhaps that it be prohibited from missing its deadlines on more than a certain
number of consecutive operations.
v2023.06.11a
We clearly cannot take a sound-bite approach to either hard or soft real time. The
next section therefore takes a more real-world approach.
14.3.1.3 Real-World Real Time

Although sentences like “Hard real-time systems always meet their deadlines!” are
catchy and easy to memorize, something else is needed for real-world real-time
systems. Although the resulting specifications are harder to memorize, they can simplify
construction of a real-time system by imposing constraints on the environment, the
workload, and the real-time application itself.
Environmental Constraints Constraints on the environment address the objection to

open-ended promises of response times implied by “hard real time”. These constraints
might specify permissible operating temperatures, air quality, levels and types of
electromagnetic radiation, and, to Figure 14.1’s point, levels of shock and vibration.
Of course, some constraints are easier to meet than others. Any number of people
have learned the hard way that commodity computer components often refuse to operate
at sub-freezing temperatures, which suggests a set of climate-control requirements.
An old college friend once had the challenge of operating a real-time system in an
atmosphere featuring some rather aggressive chlorine compounds, a challenge that he
wisely handed off to his colleagues designing the hardware. In effect, my colleague
imposed an atmospheric-composition constraint on the environment immediately
surrounding the computer, a constraint that the hardware designers met through use of
physical seals.
Another old college friend worked on a computer-controlled system that sputtered
ingots of titanium using an industrial-strength arc in a vacuum. From time to time, the
arc would decide that it was bored with its path through the ingot of titanium and choose
a far shorter and more entertaining path to ground. As we all learned in our physics
classes, a sudden shift in the flow of electrons creates an electromagnetic wave, with
larger shifts in larger flows creating higher-power electromagnetic waves. And in this
case, the resulting electromagnetic pulses were sufficient to induce a quarter of a volt
potential difference in the leads of a small “rubber ducky” antenna located more than 400
meters away. This meant that nearby conductors experienced higher voltages, courtesy
of the inverse-square law. This included those conductors making up the computer
controlling the sputtering process. In particular, the voltage induced on that computer’s
reset line was sufficient to actually reset the computer, mystifying everyone involved.
This situation was addressed using hardware, including some elaborate shielding and a
fiber-optic network with the lowest bitrate I have ever heard of, namely 9600 baud. Less
spectacular electromagnetic environments can often be handled by software through
use of error detection and correction codes. That said, it is important to remember that
although error detection and correction codes can reduce failure rates, they normally
cannot reduce them all the way down to zero, which can present yet another obstacle to
achieving hard real-time response.
There are also situations where a minimum level of energy is required, for example,
through the power leads of the system and through the devices through which the system
is to communicate with that portion of the outside world that is to be monitored or
controlled.
Quick Quiz 14.6: But what about battery-powered systems? They don’t require energy flowing
into the system as a whole.
v2023.06.11a
A number of systems are intended to operate in environments with impressive levels of

shock and vibration, for example, engine control systems. More strenuous requirements
may be found when we move away from continuous vibrations to intermittent shocks.
For example, during my undergraduate studies, I encountered an old Athena ballistics
computer, which was designed to continue operating normally even if a hand grenade
went off nearby.5 And finally, the “black boxes” used in airliners must continue operating
before, during, and after a crash.
Of course, it is possible to make hardware more robust against environmental shocks
and insults. Any number of ingenious mechanical shock-absorbing devices can reduce
the effects of shock and vibration, multiple layers of shielding can reduce the effects of
low-energy electromagnetic radiation, error-correction coding can reduce the effects of
high-energy radiation, various potting and sealing techniques can reduce the effect of
air quality, and any number of heating and cooling systems can counter the effects of
temperature. In extreme cases, triple modular redundancy can reduce the probability
that a fault in one part of the system will result in incorrect behavior from the overall
system. However, all of these methods have one thing in common: Although they can
reduce the probability of failure, they cannot reduce it to zero.
These environmental challenges are often met via robust hardware, however, the
workload and application constraints in the next two sections are often handled in
software.
Workload Constraints Just as with people, it is often possible to prevent a real-time

system from meeting its deadlines by overloading it. For example, if the system is being
interrupted too frequently, it might not have sufficient CPU bandwidth to handle its
real-time application. A hardware solution to this problem might limit the rate at which
interrupts were delivered to the system. Possible software solutions include disabling
interrupts for some time if they are being received too frequently, resetting the device
generating too-frequent interrupts, or even avoiding interrupts altogether in favor of
polling.
Overloading can also degrade response times due to queueing effects, so it is not
unusual for real-time systems to overprovision CPU bandwidth, so that a running system
has (say) 80 % idle time. This approach also applies to storage and networking devices.
In some cases, separate storage and networking hardware might be reserved for the sole
use of high-priority portions of the real-time application. In short, it is not unusual
for this hardware to be mostly idle, given that response time is more important than
throughput in real-time systems.
Quick Quiz 14.7: But given the results from queueing theory, won’t low utilization merely
improve the average response time rather than improving the worst-case response time? And
isn’t worst-case response time all that most real-time systems really care about?
Of course, maintaining sufficiently low utilization requires great discipline throughout

the design and implementation. There is nothing quite like a little feature creep to
destroy deadlines.
Application Constraints It is easier to provide bounded response time for some

operations than for others. For example, it is quite common to see response-time
5 Decades later, the acceptance tests for some types of computer systems involve large
detonations, and some types of communications networks must deal with what is delicately
termed “ballistic jamming.”
v2023.06.11a
specifications for interrupts and for wake-up operations, but quite rare for (say) filesystem
unmount operations. One reason for this is that it is quite difficult to bound the amount
of work that a filesystem-unmount operation might need to do, given that the unmount
is required to flush all of that filesystem’s in-memory data to mass storage.
This means that real-time applications must be confined to operations for which
bounded latencies can reasonably be provided. Other operations must either be pushed
out into the non-real-time portions of the application or forgone entirely.
There might also be constraints on the non-real-time portions of the application. For
example, is the non-real-time application permitted to use the CPUs intended for the
real-time portion? Are there time periods during which the real-time portion of the
application is expected to be unusually busy, and if so, is the non-real-time portion of
the application permitted to run at all during those times? Finally, by what amount
is the real-time portion of the application permitted to degrade the throughput of the
non-real-time portion?
Real-World Real-Time Specifications As can be seen from the preceding sections,

a real-world real-time specification needs to include constraints on the environment,
on the workload, and on the application itself. In addition, for the operations that the
real-time portion of the application is permitted to make use of, there must be constraints
on the hardware and software implementing those operations.
For each such operation, these constraints might include a maximum response time
(and possibly also a minimum response time) and a probability of meeting that response
time. A probability of 100 % indicates that the corresponding operation must provide
hard real-time service.
In some cases, both the response times and the required probabilities of meeting them
might vary depending on the parameters to the operation in question. For example, a
network operation over a local LAN would be much more likely to complete in (say)
100 microseconds than would that same network operation over a transcontinental
WAN. Furthermore, a network operation over a copper or fiber LAN might have an
extremely high probability of completing without time-consuming retransmissions,
while that same networking operation over a lossy WiFi network might have a much
higher probability of missing tight deadlines. Similarly, a read from a tightly coupled
solid-state disk (SSD) could be expected to complete much more quickly than that same
read to an old-style USB-connected rotating-rust disk drive.6
Some real-time applications pass through different phases of operation. For example,
a real-time system controlling a plywood lathe that peels a thin sheet of wood (called
“veneer”) from a spinning log must: (1) Load the log into the lathe, (2) Position the log
on the lathe’s chucks so as to expose the largest cylinder contained within that log to
the blade, (3) Start spinning the log, (4) Continuously vary the knife’s position so as
to peel the log into veneer, (5) Remove the remaining core of the log that is too small
to peel, and (6) Wait for the next log. Each of these six phases of operation might
well have its own set of deadlines and environmental constraints, for example, one
would expect phase 4’s deadlines to be much more severe than those of phase 6, as in
milliseconds rather than seconds. One might therefore expect that low-priority work
would be performed in phase 6 rather than in phase 4. In any case, careful choices of
6 Important safety tip: Worst-case response times from USB devices can be extremely
long. Real-time systems should therefore take care to place any USB devices well away from
critical paths.
v2023.06.11a
hardware, drivers, and software configuration would be required to support phase 4’s
more severe requirements.
A key advantage of this phase-by-phase approach is that the latency budgets can
be broken down, so that the application’s various components can be developed
independently, each with its own latency budget. Of course, as with any other kind of
budget, there will likely be the occasional conflict as to which component gets which
fraction of the overall budget, and as with any other kind of budget, strong leadership
and a sense of shared goals can help to resolve these conflicts in a timely fashion. And,
again as with other kinds of technical budget, a strong validation effort is required in
order to ensure proper focus on latencies and to give early warning of latency problems.
A successful validation effort will almost always include a good test suite, which might
be unsatisfying to the theorists, but has the virtue of helping to get the job done. As a
point of fact, as of early 2021, most real-world real-time system use an acceptance test
rather than formal proofs.
However, the widespread use of test suites to validate real-time systems does have
a very real disadvantage, namely that real-time software is validated only on specific
configurations of hardware and software. Adding additional configurations requires
additional costly and time-consuming testing. Perhaps the field of formal verification
will advance sufficiently to change this situation, but as of early 2021, rather large
advances are required.
Quick Quiz 14.8: Formal verification is already quite capable, benefiting from decades of
intensive study. Are additional advances really required, or is this just a practitioner’s excuse to
continue to lazily ignore the awesome power of formal verification?
In addition to latency requirements for the real-time portions of the application, there
will likely be performance and scalability requirements for the non-real-time portions
of the application. These additional requirements reflect the fact that ultimate real-time
latencies are often attained by degrading scalability and average performance.
Software-engineering requirements can also be important, especially for large appli-
cations that must be developed and maintained by large teams. These requirements
often favor increased modularity and fault isolation.
This is a mere outline of the work that would be required to specify deadlines and
environmental constraints for a production real-time system. It is hoped that this outline
clearly demonstrates the inadequacy of the sound-bite-based approach to real-time
computing.
14.3.2 Who Needs Real-Time?

It is possible to argue that all computing is in fact real-time computing. For one
example, when you purchase a birthday gift online, you expect the gift to arrive
before the recipient’s birthday. And in fact even turn-of-the-millennium web services
observed sub-second response constraints [Boh01], and requirements have not eased
with the passage of time [DHJ+ 07]. It is nevertheless useful to focus on those real-time
applications whose response-time requirements cannot be achieved straightforwardly
by non-real-time systems and applications. Of course, as hardware costs decrease and
bandwidths and memory sizes increase, the line between real-time and non-real-time
will continue to shift, but such progress is by no means a bad thing.
v2023.06.11a
Stimulus
Hard Non-Real-Time
Real-Time Strategy
Response "Reflexes" and Planning
Figure 14.4: Real-Time Reflexes
Quick Quiz 14.9: Differentiating real-time from non-real-time based on what can “be achieved
straightforwardly by non-real-time systems and applications” is a travesty! There is absolutely
no theoretical basis for such a distinction!!! Can’t we do better than that???
Real-time computing is used in industrial-control applications, ranging from manu-

facturing to avionics; scientific applications, perhaps most spectacularly in the adaptive
optics used by large Earth-bound telescopes to de-twinkle starlight; military applications,
including the afore-mentioned avionics; and financial-services applications, where the
first computer to recognize an opportunity is likely to reap most of the profit. These
four areas could be characterized as “in search of production”, “in search of life”, “in
search of death”, and “in search of money”.
Financial-services applications differ subtly from applications in the other three
categories in that money is non-material, meaning that non-computational latencies
are quite small. In contrast, mechanical delays inherent in the other three categories
provide a very real point of diminishing returns beyond which further reductions in the
application’s real-time response provide little or no benefit. This means that financial-
services applications, along with other real-time information-processing applications,
face an arms race, where the application with the lowest latencies normally wins.
Although the resulting latency requirements can still be specified as described in
Paragraph “Real-World Real-Time Specifications” on Page 460, the unusual nature
of these requirements has led some to refer to financial and information-processing
applications as “low latency” rather than “real time”.
Regardless of exactly what we choose to call it, there is substantial need for real-time
computing [Pet06, Inm07].
14.3.3 Who Needs Parallel Real-Time?

It is less clear who really needs parallel real-time computing, but the advent of
low-cost multicore systems has brought it to the fore regardless. Unfortunately, the
traditional mathematical basis for real-time computing assumes single-CPU systems,
with a few exceptions that prove the rule [Bra11]. Fortunately, there are a couple
of ways of squaring modern computing hardware to fit the real-time mathematical
circle, and a few Linux-kernel hackers have been encouraging academics to make this
transition [dOCdO19, Gle10].
One approach is to recognize the fact that many real-time systems resemble biological
nervous systems, with responses ranging from real-time reflexes to non-real-time
strategizing and planning, as depicted in Figure 14.4. The hard real-time reflexes, which
read from sensors and control actuators, run real-time on a single CPU or on special-
purpose hardware such as an FPGA. The non-real-time strategy and planning portion
of the application runs on the remaining CPUs. Strategy and planning activities might
include statistical analysis, periodic calibration, user interface, supply-chain activities,
and preparation. For an example of high-compute-load preparation activities, think
v2023.06.11a
Scripting languages 1s
100 ms
Linux 2.4 kernel
10 ms
Real-time Java (with GC)
1 ms
Linux 2.6.x/3.x kernel
Real-time Java (no GC)
Linux 4.x/5.x kernel 100 μs
Linux -rt patchset
10 μs
Specialty RTOSes (no MMU)
1 μs
Hand-coded assembly
100 ns
Custom digital hardware
10 ns
1 ns
Custom analog hardware 100 ps
Figure 14.5: Real-Time Response Regimes
back to the veneer-peeling application discussed in Paragraph “Real-World Real-Time

Specifications” on Page 460. While one CPU is attending to the high-speed real-time
computations required to peel one log, the other CPUs might be analyzing the size
and shape of the next log in order to determine how to position the next log so as to
obtain the largest cylinder of high-quality wood. It turns out that many applications have
non-real-time and real-time components [BMP08], so this approach can often be used
to allow traditional real-time analysis to be combined with modern multicore hardware.
Another trivial approach is to shut off all but one hardware thread so as to return to
the settled mathematics of uniprocessor real-time computing. However, this approach
gives up potential cost and energy-efficiency advantages. That said, obtaining these
advantages requires overcoming the parallel performance obstacles covered in Chapter 3,
and not merely on average, but instead in the worst case.
Implementing parallel real-time systems can therefore be quite a challenge. Ways of
meeting this challenge are outlined in the following section.
14.3.4 Implementing Parallel Real-Time Systems

We will look at two major styles of real-time systems, event-driven and polling. An
event-driven real-time system remains idle much of the time, responding in real time
to events passed up through the operating system to the application. Alternatively,
the system could instead be running a background non-real-time workload. A polling
real-time system features a real-time thread that is CPU bound, running in a tight loop
that polls inputs and updates outputs on each pass. This tight polling loop often executes
entirely in user mode, reading from and writing to hardware registers that have been
mapped into the user-mode application’s address space. Alternatively, some applications
place the polling loop into the kernel, for example, using loadable kernel modules.
Regardless of the style chosen, the approach used to implement a real-time system
will depend on the deadlines, for example, as shown in Figure 14.5. Starting from the
top of this figure, if you can live with response times in excess of one second, you might
well be able to use scripting languages to implement your real-time application—and
scripting languages are in fact used surprisingly often, not that I necessarily recommend
this practice. If the required latencies exceed several tens of milliseconds, old 2.4
versions of the Linux kernel can be used, not that I necessarily recommend this practice,
either. Special real-time Java implementations can provide real-time response latencies
v2023.06.11a
RTOS Process
RTOS Process
Linux Process
Linux Process
Linux Process
Linux Process
Linux Process
Linux Process
RCU read-side
Linux critical sections
Kernel Spinlock
critical sections
Interrupt handlers
Scheduling
Clock Interrupt disable
Interrupt Preempt disable
RTOS
Figure 14.6: Linux Ported to RTOS
of a few milliseconds, even when the garbage collector is used. The Linux 2.6.x and 3.x
kernels can provide real-time latencies of a few hundred microseconds if painstakingly
configured, tuned, and run on real-time-friendly hardware. Special real-time Java
implementations can provide real-time latencies below 100 microseconds if use of the
garbage collector is carefully avoided. (But note that avoiding the garbage collector
means also avoiding Java’s large standard libraries, thus also avoiding Java’s productivity
advantages.) The Linux 4.x and 5.x kernels can provide sub-hundred-microsecond
latencies, but with all the same caveats as for the 2.6.x and 3.x kernels. A Linux kernel
incorporating the -rt patchset can provide latencies well below 20 microseconds, and
specialty real-time operating systems (RTOSes) running without MMUs can provide
sub-ten-microsecond latencies. Achieving sub-microsecond latencies typically requires
hand-coded assembly or even special-purpose hardware.
Of course, careful configuration and tuning are required all the way down the stack. In
particular, if the hardware or firmware fails to provide real-time latencies, there is nothing
that the software can do to make up for the lost time. Worse yet, high-performance
hardware sometimes sacrifices worst-case behavior to obtain greater throughput. In
fact, timings from tight loops run with interrupts disabled can provide the basis for a
high-quality random-number generator [MOZ09]. Furthermore, some firmware does
cycle-stealing to carry out various housekeeping tasks, in some cases attempting to
cover its tracks by reprogramming the victim CPU’s hardware clocks. Of course, cycle
stealing is expected behavior in virtualized environment, but people are nevertheless
working towards real-time response in virtualized environments [Gle12, Kis14]. It
is therefore critically important to evaluate your hardware’s and firmware’s real-time
capabilities.
But given competent real-time hardware and firmware, the next layer up the stack is
the operating system, which is covered in the next section.
v2023.06.11a
14.3.5 Implementing Parallel Real-Time Operating Systems

There are a number of strategies that may be used to implement a real-time system. One
approach is to port a general-purpose non-real-time OS on top of a special purpose
real-time operating system (RTOS), as shown in Figure 14.6. The green “Linux Process”
boxes represent non-real-time processes running on the Linux kernel, while the yellow
“RTOS Process” boxes represent real-time processes running on the RTOS.
This was a very popular approach before the Linux kernel gained real-time capabilities,
and is still in use [xen14, Yod04b]. However, this approach requires that the application
be split into one portion that runs on the RTOS and another that runs on Linux. Although
it is possible to make the two environments look similar, for example, by forwarding
POSIX system calls from the RTOS to a utility thread running on Linux, there are
invariably rough edges.
In addition, the RTOS must interface to both the hardware and to the Linux kernel,
thus requiring significant maintenance with changes in both hardware and kernel.
Furthermore, each such RTOS often has its own system-call interface and set of system
libraries, which can balkanize both ecosystems and developers. In fact, these problems
seem to be what drove the combination of RTOSes with Linux, as this approach allowed
access to the full real-time capabilities of the RTOS, while allowing the application’s
non-real-time code full access to Linux’s open-source ecosystem.
Although pairing RTOSes with the Linux kernel was a clever and useful short-term
response during the time that the Linux kernel had minimal real-time capabilities, it
also motivated adding real-time capabilities to the Linux kernel. Progress towards this
goal is shown in Figure 14.7. The upper row shows a diagram of the Linux kernel with
preemption disabled, thus having essentially no real-time capabilities. The middle row
shows a set of diagrams showing the increasing real-time capabilities of the mainline
Linux kernel with preemption enabled. Finally, the bottom row shows a diagram
of the Linux kernel with the -rt patchset applied, maximizing real-time capabilities.
Functionality from the -rt patchset is added to mainline, hence the increasing capabilities
of the mainline Linux kernel over time. Nevertheless, the most demanding real-time
applications continue to use the -rt patchset.
The non-preemptible kernel shown at the top of Figure 14.7 is built with CONFIG_
PREEMPT=n, so that execution within the Linux kernel cannot be preempted. This means
that the kernel’s real-time response latency is bounded below by the longest code path in
the Linux kernel, which is indeed long. However, user-mode execution is preemptible,
so that one of the real-time Linux processes shown in the upper right may preempt any
of the non-real-time Linux processes shown in the upper left anytime the non-real-time
process is executing in user mode.
The middle row of Figure 14.7 shows three stages (from left to right) in the development
of Linux’s preemptible kernels. In all three stages, most process-level code within the
Linux kernel can be preempted. This of course greatly improves real-time response
latency, but preemption is still disabled within RCU read-side critical sections, spinlock
critical sections, interrupt handlers, interrupt-disabled code regions, and preempt-
disabled code regions, as indicated by the red boxes in the left-most diagram in the
middle row of the figure. The advent of preemptible RCU allowed RCU read-side
critical sections to be preempted, as shown in the central diagram, and the advent of
threaded interrupt handlers allowed device-interrupt handlers to be preempted, as shown
in the right-most diagram. Of course, a great deal of other real-time functionality was
added during this time, however, it cannot be as easily represented on this diagram. It
will instead be discussed in Section 14.3.5.1.
v2023.06.11a
RT Linux Process
RT Linux Process
RT Linux Process
Linux Process
Linux Process
Linux Process
RCU read-side
Kernel Spinlock
critical sections
Interrupt handlers
Scheduling
CONFIG_PREEMPT=n
RT Linux Process
RT Linux Process
RT Linux Process
RT Linux Process
RT Linux Process
RT Linux Process
RT Linux Process
RT Linux Process
RT Linux Process
Linux Process
Linux Process
Linux Process
Linux Process
Linux Process
Linux Process
Linux Process
Linux Process
Linux Process
RCU read-side RCU read-side RCU read-side
Linux critical sections Linux critical sections Linux critical sections
Kernel Spinlock
critical sections
Kernel Spinlock
critical sections
Kernel Spinlock
critical sections
Interrupt handlers Interrupt handlers Interrupt handlers
Scheduling Scheduling Scheduling
Clock Interrupt disable Clock Interrupt disable Clock Interrupt disable
Interrupt Preempt disable Interrupt Preempt disable Interrupt Preempt disable
CONFIG_PREEMPT=y CONFIG_PREEMPT=y CONFIG_PREEMPT=y

Pre-2008 (With preemptible RCU) (With threaded interrupts)
RT Linux Process
RT Linux Process
RT Linux Process
Linux Process
Linux Process
Linux Process
RCU read-side
Kernel Spinlock
critical sections
Interrupt handlers
Scheduling
-rt patchset
Figure 14.7: Linux-Kernel Real-Time Implementations
v2023.06.11a
RT Linux Process
RT Linux Process
RT Linux Process
Linux Process
Linux Process
Linux Process
NO_HZ_FULL Linux Process
NO_HZ_FULL Linux Process

RCU read-side
Kernel Spinlock
critical sections
Interrupt handlers
Scheduling
Figure 14.8: CPU Isolation
The bottom row of Figure 14.7 shows the -rt patchset, which features threaded
(and thus preemptible) interrupt handlers for many devices, which also allows the
corresponding “interrupt-disabled” regions of these drivers to be preempted. These
drivers instead use locking to coordinate the process-level portions of each driver with its
threaded interrupt handlers. Finally, in some cases, disabling of preemption is replaced
by disabling of migration. These measures result in excellent response times in many
systems running the -rt patchset [RMF19, dOCdO19].
A final approach is simply to get everything out of the way of the real-time process,
clearing all other processing off of any CPUs that this process needs, as shown in
Figure 14.8. This was implemented in the 3.10 Linux kernel via the CONFIG_NO_HZ_
FULL Kconfig parameter [Cor13, Wei12]. It is important to note that this approach
requires at least one housekeeping CPU to do background processing, for example
running kernel daemons. However, when there is only one runnable task on a given
non-housekeeping CPU, scheduling-clock interrupts are shut off on that CPU, removing
an important source of interference and OS jitter. With a few exceptions, the kernel
does not force other processing off of the non-housekeeping CPUs, but instead simply
provides better performance when only one runnable task is present on a given CPU. Any
number of userspace tools may be used to force a given CPU to have no more that one
runnable task. If configured properly, a non-trivial undertaking, CONFIG_NO_HZ_FULL
offers real-time threads levels of performance that come close to those of bare-metal
systems [ACA+ 18]. Frédéric Weisbecker produced a practical guide to CONFIG_NO_
HZ_FULL configuration [Wei22d, Wei22b, Wei22e, Wei22c, Wei22a, Wei22f].
There has of course been much debate over which of these approaches is best for real-
time systems, and this debate has been going on for quite some time [Cor04a, Cor04c].
As usual, the answer seems to be “It depends,” as discussed in the following sections.
Section 14.3.5.1 considers event-driven real-time systems, and Section 14.3.5.2 considers
real-time systems that use a CPU-bound polling loop.
v2023.06.11a
14.3.5.1 Event-Driven Real-Time Support

The operating-system support required for event-driven real-time applications is quite
extensive, however, this section will focus on only a few items, namely timers, threaded
interrupts, priority inheritance, preemptible RCU, and preemptible spinlocks.
Timers are clearly critically important for real-time operations. After all, if you
cannot specify that something be done at a specific time, how are you going to respond
by that time? Even in non-real-time systems, large numbers of timers are generated,
so they must be handled extremely efficiently. Example uses include retransmit timers
for TCP connections (which are almost always canceled before they have a chance to
fire),7 timed delays (as in sleep(1), which are rarely canceled), and timeouts for the
poll() system call (which are often canceled before they have a chance to fire). A
good data structure for such timers would therefore be a priority queue whose addition
and deletion primitives were fast and O (1) in the number of timers posted.
The classic data structure for this purpose is the calendar queue, which in the Linux
kernel is called the timer wheel. This age-old data structure is also heavily used in
discrete-event simulation. The idea is that time is quantized, for example, in the Linux
kernel, the duration of the time quantum is the period of the scheduling-clock interrupt.
A given time can be represented by an integer, and any attempt to post a timer at some
non-integral time will be rounded to a convenient nearby integral time quantum.
One straightforward implementation would be to allocate a single array, indexed by
the low-order bits of the time. This works in theory, but in practice systems create large
numbers of long-duration timeouts (for example, the two-hour keepalive timeouts for
TCP sessions) that are almost always canceled. These long-duration timeouts cause
problems for small arrays because much time is wasted skipping timeouts that have not
yet expired. On the other hand, an array that is large enough to gracefully accommodate
a large number of long-duration timeouts would consume too much memory, especially
given that performance and scalability concerns require one such array for each and
every CPU.
A common approach for resolving this conflict is to provide multiple arrays in a
hierarchy. At the lowest level of this hierarchy, each array element represents one unit
of time. At the second level, each array element represents 𝑁 units of time, where 𝑁 is
the number of elements in each array. At the third level, each array element represents
𝑁 2 units of time, and so on up the hierarchy. This approach allows the individual arrays
to be indexed by different bits, as illustrated by Figure 14.9 for an unrealistically small
eight-bit clock. Here, each array has 16 elements, so the low-order four bits of the time
(currently 0xf) index the low-order (rightmost) array, and the next four bits (currently
0x1) index the next level up. Thus, we have two arrays each with 16 elements, for a
total of 32 elements, which, taken together, is much smaller than the 256-element array
that would be required for a single array.
This approach works extremely well for throughput-based systems. Each timer
operation is O (1) with small constant, and each timer element is touched at most 𝑚 + 1
times, where 𝑚 is the number of levels.
Unfortunately, timer wheels do not work well for real-time systems, and for two
reasons. The first reason is that there is a harsh tradeoff between timer accuracy and timer
overhead, which is fancifully illustrated by Figures 14.10 and 14.11. In Figure 14.10,
timer processing happens only once per millisecond, which keeps overhead acceptably
7 At least assuming reasonably low packet-loss rates!
v2023.06.11a
0x x0
1x x1
2x x2
3x x3
4x x4
5x x5
6x x6
7x x7
8x x8
9x x9
ax xa
bx xb
cx xc
dx xd
ex xe
fx xf
1f
Figure 14.9: Timer Wheel
Figure 14.10: Timer Wheel at 1 kHz
v2023.06.11a
Figure 14.11: Timer Wheel at 100 kHz
low for many (but not all!) workloads, but which also means that timeouts cannot
be set for finer than one-millisecond granularities. On the other hand, Figure 14.11
shows timer processing taking place every ten microseconds, which provides acceptably
fine timer granularity for most (but not all!) workloads, but which processes timers so
frequently that the system might well not have time to do anything else.
The second reason is the need to cascade timers from higher levels to lower levels.
Referring back to Figure 14.9, we can see that any timers enqueued on element 1x in
the upper (leftmost) array must be cascaded down to the lower (rightmost) array so
that may be invoked when their time arrives. Unfortunately, there could be a large
number of timeouts waiting to be cascaded, especially for timer wheels with larger
numbers of levels. The power of statistics causes this cascading to be a non-problem for
throughput-oriented systems, but cascading can result in problematic degradations of
latency in real-time systems.
Of course, real-time systems could simply choose a different data structure, for
example, some form of heap or tree, giving up O (1) bounds on insertion and deletion
operations to gain O (log 𝑛) limits on data-structure-maintenance operations. This can
be a good choice for special-purpose RTOSes, but is inefficient for general-purpose
systems such as Linux, which routinely support extremely large numbers of timers.
The solution chosen for the Linux kernel’s -rt patchset is to differentiate between
timers that schedule later activity and timeouts that schedule error handling for low-
probability errors such as TCP packet losses. One key observation is that error handling
is normally not particularly time-critical, so that a timer wheel’s millisecond-level
granularity is good and sufficient. Another key observation is that error-handling
timeouts are normally canceled very early, often before they can be cascaded. In
addition, systems commonly have many more error-handling timeouts than they do
timer events, so that an O (log 𝑛) data structure should provide acceptable performance
for timer events.
However, it is possible to do better, namely by simply refusing to cascade timers.
Instead of cascading, the timers that would otherwise have been cascaded all the way
down the calendar queue are handled in place. This does result in up to a few percent
error for the time duration, but the few situations where this is a problem can instead
use tree-based high-resolution timers (hrtimers).
v2023.06.11a
Return From
Interrupt Interrupt Mainline
Mainline
Interrupt Handler
Code Code
Long Latency:
Degrades Response Time
Figure 14.12: Non-Threaded Interrupt Handler
Return From
Interrupt
Interrupt Interrupt Mainline
Mainline
Code Code Preemptible
IRQ Thread
Interrupt Handler
Short Latency:
Improved Response Time
Figure 14.13: Threaded Interrupt Handler
In short, the Linux kernel’s -rt patchset uses timer wheels for error-handling timeouts
and a tree for timer events, providing each category the required quality of service.
Threaded interrupts are used to address a significant source of degraded real-time

latencies, namely long-running interrupt handlers, as shown in Figure 14.12. These
latencies can be especially problematic for devices that can deliver a large number of
events with a single interrupt, which means that the interrupt handler will run for an
extended period of time processing all of these events. Worse yet are devices that can
deliver new events to a still-running interrupt handler, as such an interrupt handler might
well run indefinitely, thus indefinitely degrading real-time latencies.
One way of addressing this problem is the use of threaded interrupts shown in
Figure 14.13. Interrupt handlers run in the context of a preemptible IRQ thread, which
runs at a configurable priority. The device interrupt handler then runs for only a short
time, just long enough to make the IRQ thread aware of the new event. As shown in
the figure, threaded interrupts can greatly improve real-time latencies, in part because
interrupt handlers running in the context of the IRQ thread may be preempted by
high-priority real-time threads.
However, there is no such thing as a free lunch, and there are downsides to threaded
interrupts. One downside is increased interrupt latency. Instead of immediately running
the interrupt handler, the handler’s execution is deferred until the IRQ thread gets around
to running it. Of course, this is not a problem unless the device generating the interrupt
is on the real-time application’s critical path.
v2023.06.11a
Figure 14.14: Priority Inversion and User Input
Another downside is that poorly written high-priority real-time code might starve
the interrupt handler, for example, preventing networking code from running, in turn
making it very difficult to debug the problem. Developers must therefore take great
care when writing high-priority real-time code. This has been dubbed the Spiderman
principle: With great power comes great responsibility.
Priority inheritance is used to handle priority inversion, which can be caused by,
among other things, locks acquired by preemptible interrupt handlers [SRL90]. Suppose
that a low-priority thread holds a lock, but is preempted by a group of medium-priority
threads, at least one such thread per CPU. If an interrupt occurs, a high-priority IRQ
thread will preempt one of the medium-priority threads, but only until it decides to
acquire the lock held by the low-priority thread. Unfortunately, the low-priority thread
cannot release the lock until it starts running, which the medium-priority threads prevent
it from doing. So the high-priority IRQ thread cannot acquire the lock until after one of
the medium-priority threads releases its CPU. In short, the medium-priority threads are
indirectly blocking the high-priority IRQ threads, a classic case of priority inversion.
Note that this priority inversion could not happen with non-threaded interrupts
because the low-priority thread would have to disable interrupts while holding the lock,
which would prevent the medium-priority threads from preempting it.
In the priority-inheritance solution, the high-priority thread attempting to acquire the
lock donates its priority to the low-priority thread holding the lock until such time as
the lock is released, thus preventing long-term priority inversion.
Of course, priority inheritance does have its limitations. For example, if you can
design your application to avoid priority inversion entirely, you will likely obtain
somewhat better latencies [Yod04b]. This should be no surprise, given that priority
inheritance adds a pair of context switches to the worst-case latency. That said, priority
inheritance can convert indefinite postponement into a limited increase in latency, and
the software-engineering benefits of priority inheritance may outweigh its latency costs
in many applications.
Another limitation is that it addresses only lock-based priority inversions within the
context of a given operating system. One priority-inversion scenario that it cannot
address is a high-priority thread waiting on a network socket for a message that is
to be written by a low-priority process that is preempted by a set of CPU-bound
medium-priority processes. In addition, a potential disadvantage of applying priority
inheritance to user input is fancifully depicted in Figure 14.14.
v2023.06.11a
A final limitation involves reader-writer locking. Suppose that we have a very large
number of low-priority threads, perhaps even thousands of them, each of which read-
holds a particular reader-writer lock. Suppose that all of these threads are preempted by
a set of medium-priority threads, with at least one medium-priority thread per CPU.
Finally, suppose that a high-priority thread awakens and attempts to write-acquire this
same reader-writer lock. No matter how vigorously we boost the priority of the threads
read-holding this lock, it could well be a good long time before the high-priority thread
can complete its write-acquisition.
There are a number of possible solutions to this reader-writer lock priority-inversion
conundrum:
1. Only allow one read-acquisition of a given reader-writer lock at a time. (This is

the approach traditionally taken by the Linux kernel’s -rt patchset.)
2. Only allow 𝑁 read-acquisitions of a given reader-writer lock at a time, where 𝑁 is

the number of CPUs.
3. Only allow 𝑁 read-acquisitions of a given reader-writer lock at a time, where 𝑁 is

a number specified somehow by the developer.
4. Prohibit high-priority threads from write-acquiring reader-writer locks that are

ever read-acquired by threads running at lower priorities. (This is a variant of the
priority ceiling protocol [SRL90].)
Quick Quiz 14.10: But if you only allow one reader at a time to read-acquire a reader-writer
lock, isn’t that the same as an exclusive lock???
The no-concurrent-readers restriction eventually became intolerable, so the -rt

developers looked more carefully at how the Linux kernel uses reader-writer spinlocks.
They learned that time-critical code rarely uses those parts of the kernel that write-acquire
reader-writer locks, so that the prospect of writer starvation was not a show-stopper.
They therefore constructed a real-time reader-writer lock in which write-side acquisitions
use priority inheritance among each other, but where read-side acquisitions take absolute
priority over write-side acquisitions. This approach appears to be working well in
practice, and is another lesson in the importance of clearly understanding what your
users really need.
One interesting detail of this implementation is that both the rt_read_lock() and
the rt_write_lock() functions enter an RCU read-side critical section and both the
rt_read_unlock() and the rt_write_unlock() functions exit that critical section.
This is necessary because non-realtime kernels’ reader-writer locking functions disable
preemption across their critical sections, and there really are reader-writer locking
use cases that rely on the fact that synchronize_rcu() will therefore wait for all
pre-existing reader-writer-lock critical sections to complete. Let this be a lesson to you:
Understanding what your users really need is critically important to correct operation,
not just to performance. Not only that, but what your users really need changes over
time.
This has the side-effect that all of a -rt kernel’s reader-writer locking critical sections
are subject to RCU priority boosting. This provides at least a partial solution to the
problem of reader-writer lock readers being preempted for extended periods of time.
It is also possible to avoid reader-writer lock priority inversion by converting the
reader-writer lock to RCU, as briefly discussed in the next section.
v2023.06.11a
Listing 14.3: Preemptible Linux-Kernel RCU

1 void __rcu_read_lock(void)
2 {
3 current->rcu_read_lock_nesting++;
4 barrier();
5 }
6
7 void __rcu_read_unlock(void)
8 {
9 barrier();
10 if (!--current->rcu_read_lock_nesting)
11 barrier();
12 if (READ_ONCE(current->rcu_read_unlock_special.s)) {
13 rcu_read_unlock_special(t);
14 }
15 }
Preemptible RCU can sometimes be used as a replacement for reader-writer lock-

ing [MW07, MBWW12, McK14f], as was discussed in Section 9.5. Where it can be
used, it permits readers and updaters to run concurrently, which prevents low-priority
readers from inflicting any sort of priority-inversion scenario on high-priority updaters.
However, for this to be useful, it is necessary to be able to preempt long-running RCU
read-side critical sections [GMTW08]. Otherwise, long RCU read-side critical sections
would result in excessive real-time latencies.
A preemptible RCU implementation was therefore added to the Linux kernel. This
implementation avoids the need to individually track the state of each and every task in
the kernel by keeping lists of tasks that have been preempted within their current RCU
read-side critical sections. A grace period is permitted to end: (1) Once all CPUs have
completed any RCU read-side critical sections that were in effect before the start of the
current grace period and (2) Once all tasks that were preempted while in one of those
pre-existing critical sections have removed themselves from their lists. A simplified
version of this implementation is shown in Listing 14.3. The __rcu_read_lock()
function spans lines 1–5 and the __rcu_read_unlock() function spans lines 7–15.
Line 3 of __rcu_read_lock() increments a per-task count of the number of
nested rcu_read_lock() calls, and line 4 prevents the compiler from reordering the
subsequent code in the RCU read-side critical section to precede the rcu_read_lock().
Line 9 of __rcu_read_unlock() prevents the compiler from reordering the code in
the critical section with the remainder of this function. Line 10 decrements the nesting
count and checks to see if it has become zero, in other words, if this corresponds to the
outermost rcu_read_unlock() of a nested set. If so, line 11 prevents the compiler
from reordering this nesting update with line 12’s check for special handling. If special
handling is required, then the call to rcu_read_unlock_special() on line 13 carries
it out.
There are several types of special handling that can be required, but we will focus on
that required when the RCU read-side critical section has been preempted. In this case,
the task must remove itself from the list that it was added to when it was first preempted
within its RCU read-side critical section. However, it is important to note that these lists
are protected by locks, which means that rcu_read_unlock() is no longer lockless.
However, the highest-priority threads will not be preempted, and therefore, for those
highest-priority threads, rcu_read_unlock() will never attempt to acquire any locks.
In addition, if implemented carefully, locking can be used to synchronize real-time
software [Bra11, SM04a].
v2023.06.11a
Quick Quiz 14.11: Suppose that preemption occurs just after the load from t->rcu_read_
unlock_special.s on line 12 of Listing 14.3. Mightn’t that result in the task failing to invoke
rcu_read_unlock_special(), thus failing to remove itself from the list of tasks blocking
the current grace period, in turn causing that grace period to extend indefinitely?
Another important real-time feature of RCU, whether preemptible or not, is the ability
to offload RCU callback execution to a kernel thread. To use this, your kernel must be
built with CONFIG_RCU_NOCB_CPU=y and booted with the rcu_nocbs= kernel boot
parameter specifying which CPUs are to be offloaded. Alternatively, any CPU specified
by the nohz_full= kernel boot parameter described in Section 14.3.5.2 will also have
its RCU callbacks offloaded.
In short, this preemptible RCU implementation enables real-time response for read-
mostly data structures without the delays inherent to priority boosting of large numbers
of readers, and also without delays due to callback invocation.
Preemptible spinlocks are an important part of the -rt patchset due to the long-
duration spinlock-based critical sections in the Linux kernel. This functionality has
not yet reached mainline: Although they are a conceptually simple substitution of
sleeplocks for spinlocks, they have proven relatively controversial. In addition the
real-time functionality that is already in the mainline Linux kernel suffices for a
great many use cases, which slowed the -rt patchset’s development rate in the early
2010s [Edg13, Edg14]. However, preemptible spinlocks are absolutely necessary to the
task of achieving real-time latencies down in the tens of microseconds. Fortunately,
Linux Foundation organized an effort to fund moving the remaining code from the -rt
patchset to mainline.
Per-CPU variables are used heavily in the Linux kernel for performance reasons.
Unfortunately for real-time applications, many use cases for per-CPU variables require
coordinated update of multiple such variables, which is normally provided by disabling
preemption, which in turn degrades real-time latencies. Real-time applications clearly
need some other way of coordinating per-CPU variable updates.
One alternative is to supply per-CPU spinlocks, which as noted above are actually
sleeplocks, so that their critical sections can be preempted and so that priority inheritance
is provided. In this approach, code updating groups of per-CPU variables must acquire
the current CPU’s spinlock, carry out the update, then release whichever lock is acquired,
keeping in mind that a preemption might have resulted in a migration to some other
CPU. However, this approach introduces both overhead and deadlocks.
Another alternative, which is used in the -rt patchset as of early 2021, is to convert
preemption disabling to migration disabling. This ensures that a given kernel thread
remains on its CPU through the duration of the per-CPU-variable update, but could also
allow some other kernel thread to intersperse its own update of those same variables,
courtesy of preemption. There are cases such as statistics gathering where this is not a
problem. In the surprisingly rare case where such mid-update preemption is a problem,
the use case at hand must properly synchronize the updates, perhaps through a set of
per-CPU locks specific to that use case. Although introducing locks again introduces the
possibility of deadlock, the per-use-case nature of these locks makes any such deadlocks
easier to manage and avoid.
Closing event-driven remarks. There are of course any number of other Linux-kernel
components that are critically important to achieving world-class real-time latencies,
v2023.06.11a
for example, deadline scheduling [dO18b, dO18a], however, those listed in this section
give a good feeling for the workings of the Linux kernel augmented by the -rt patchset.
14.3.5.2 Polling-Loop Real-Time Support

At first glance, use of a polling loop might seem to avoid all possible operating-system
interference problems. After all, if a given CPU never enters the kernel, the kernel is
completely out of the picture. And the traditional approach to keeping the kernel out of
the way is simply not to have a kernel, and many real-time applications do indeed run
on bare metal, particularly those running on eight-bit microcontrollers.
One might hope to get bare-metal performance on a modern operating-system kernel
simply by running a single CPU-bound user-mode thread on a given CPU, avoiding all
causes of interference. Although the reality is of course more complex, it is becoming
possible to do just that, courtesy of the NO_HZ_FULL implementation led by Frederic
Weisbecker [Cor13, Wei12] that was accepted into version 3.10 of the Linux kernel.
Nevertheless, considerable care is required to properly set up such an environment, as
it is necessary to control a number of possible sources of OS jitter. The discussion
below covers the control of several sources of OS jitter, including device interrupts,
kernel threads and daemons, scheduler real-time throttling (this is a feature, not a bug!),
timers, non-real-time device drivers, in-kernel global synchronization, scheduling-clock
interrupts, page faults, and finally, non-real-time hardware and firmware.
Interrupts are an excellent source of large amounts of OS jitter. Unfortunately, in most
cases interrupts are absolutely required in order for the system to communicate with the
outside world. One way of resolving this conflict between OS jitter and maintaining
contact with the outside world is to reserve a small number of housekeeping CPUs, and
to force all interrupts to these CPUs. The Documentation/IRQ-affinity.txt file
in the Linux source tree describes how to direct device interrupts to specified CPUs,
which as of early 2021 involves something like the following:
$ echo 0f > /proc/irq/44/smp_affinity
This command would confine interrupt #44 to CPUs 0–3. Note that scheduling-clock
interrupts require special handling, and are discussed later in this section.
A second source of OS jitter is due to kernel threads and daemons. Individual
kernel threads, such as RCU’s grace-period kthreads (rcu_bh, rcu_preempt, and
rcu_sched), may be forced onto any desired CPUs using the taskset command, the
sched_setaffinity() system call, or cgroups.
Per-CPU kthreads are often more challenging, sometimes constraining hardware
configuration and workload layout. Preventing OS jitter from these kthreads requires
either that certain types of hardware not be attached to real-time systems, that all interrupts
and I/O initiation take place on housekeeping CPUs, that special kernel Kconfig or
boot parameters be selected in order to direct work away from the worker CPUs, or that
worker CPUs never enter the kernel. Specific per-kthread advice may be found in the
Linux kernel source Documentation directory at kernel-per-CPU-kthreads.txt.
A third source of OS jitter in the Linux kernel for CPU-bound threads running
at real-time priority is the scheduler itself. This is an intentional debugging feature,
designed to ensure that important non-realtime work is allotted at least 50 milliseconds
out of each second, even if there is an infinite-loop bug in your real-time application.
However, when you are running a polling-loop-style real-time application, you will need
to disable this debugging feature. This can be done as follows:
v2023.06.11a
$ echo -1 > /proc/sys/kernel/sched_rt_runtime_us
You will of course need to be running as root to execute this command, and you will
also need to carefully consider the aforementioned Spiderman principle. One way to
minimize the risks is to offload interrupts and kernel threads/daemons from all CPUs
running CPU-bound real-time threads, as described in the paragraphs above. In addition,
you should carefully read the material in the Documentation/scheduler directory.
The material in the sched-rt-group.rst file is particularly important, especially if
you are using the cgroups real-time features enabled by the CONFIG_RT_GROUP_SCHED
Kconfig parameter.
A fourth source of OS jitter comes from timers. In most cases, keeping a given CPU
out of the kernel will prevent timers from being scheduled on that CPU. One important
exception are recurring timers, where a given timer handler posts a later occurrence
of that same timer. If such a timer gets started on a given CPU for any reason, that
timer will continue to run periodically on that CPU, inflicting OS jitter indefinitely. One
crude but effective way to offload recurring timers is to use CPU hotplug to offline all
worker CPUs that are to run CPU-bound real-time application threads, online these
same CPUs, then start your real-time application.
A fifth source of OS jitter is provided by device drivers that were not intended for
real-time use. For an old canonical example, in 2005, the VGA driver would blank the
screen by zeroing the frame buffer with interrupts disabled, which resulted in tens of
milliseconds of OS jitter. One way of avoiding device-driver-induced OS jitter is to
carefully select devices that have been used heavily in real-time systems, and which
have therefore had their real-time bugs fixed. Another way is to confine the device’s
interrupts and all code using that device to designated housekeeping CPUs. A third way
is to test the device’s ability to support real-time workloads and fix any real-time bugs.8
A sixth source of OS jitter is provided by some in-kernel full-system synchronization
algorithms, perhaps most notably the global TLB-flush algorithm. This can be avoided by
avoiding memory-unmapping operations, and especially avoiding unmapping operations
within the kernel. As of early 2021, the way to avoid in-kernel unmapping operations is
to avoid unloading kernel modules.
A seventh source of OS jitter is provided by scheduling-clock interrrupts and RCU
callback invocation. These may be avoided by building your kernel with the NO_HZ_
FULL Kconfig parameter enabled, and then booting with the nohz_full= parameter
specifying the list of worker CPUs that are to run real-time threads. For example,
nohz_full=2-7 would designate CPUs 2, 3, 4, 5, 6, and 7 as worker CPUs, thus
leaving CPUs 0 and 1 as housekeeping CPUs. The worker CPUs would not incur
scheduling-clock interrupts as long as there is no more than one runnable task on each
worker CPU, and each worker CPU’s RCU callbacks would be invoked on one of the
housekeeping CPUs. A CPU that has suppressed scheduling-clock interrupts due to
there only being one runnable task on that CPU is said to be in adaptive ticks mode
or in nohz_full mode. It is important to ensure that you have designated enough
housekeeping CPUs to handle the housekeeping load imposed by the rest of the system,
which requires careful benchmarking and tuning.
An eighth source of OS jitter is page faults. Because most Linux implementations
use an MMU for memory protection, real-time applications running on these systems
8 If you take this approach, please submit your fixes upstream so that others can benefit.
After all, when you need to port your application to a later version of the Linux kernel, you
will be one of those “others”.
v2023.06.11a
Listing 14.4: Locating Sources of OS Jitter

1 cd /sys/kernel/debug/tracing
2 echo 1 > max_graph_depth
3 echo function_graph > current_tracer
4 # run workload
5 cat per_cpu/cpuN/trace
can be subject to page faults. Use the mlock() and mlockall() system calls to pin
your application’s pages into memory, thus avoiding major page faults. Of course, the
Spiderman principle applies, because locking down too much memory may prevent the
system from getting other work done.
A ninth source of OS jitter is unfortunately the hardware and firmware. It is therefore
important to use systems that have been designed for real-time use.
Unfortunately, this list of OS-jitter sources can never be complete, as it will change
with each new version of the kernel. This makes it necessary to be able to track down
additional sources of OS jitter. Given a CPU 𝑁 running a CPU-bound usermode thread,
the commands shown in Listing 14.4 will produce a list of all the times that this CPU
entered the kernel. Of course, the N on line 5 must be replaced with the number of the
CPU in question, and the 1 on line 2 may be increased to show additional levels of
function call within the kernel. The resulting trace can help track down the source of
the OS jitter.
As always, there is no free lunch, and NO_HZ_FULL is no exception. As noted earlier,
NO_HZ_FULL makes kernel/user transitions more expensive due to the need for delta
process accounting and the need to inform kernel subsystems (such as RCU) of the
transitions. As a rough rule of thumb, NO_HZ_FULL helps with many types of real-time
and heavy-compute workloads, but hurts other workloads that feature high rates of
system calls and I/O [ACA+ 18]. Additional limitations, tradeoffs, and configuration
advice may be found in Documentation/timers/no_hz.rst.
As you can see, obtaining bare-metal performance when running CPU-bound real-
time threads on a general-purpose OS such as Linux requires painstaking attention
to detail. Automation would of course help, and some automation has been applied,
but given the relatively small number of users, automation can be expected to appear
relatively slowly. Nevertheless, the ability to gain near-bare-metal performance while
running a general-purpose operating system promises to ease construction of some
types of real-time systems.
14.3.6 Implementing Parallel Real-Time Applications

Developing real-time applications is a wide-ranging topic, and this section can only
touch on a few aspects. To this end, Section 14.3.6.1 looks at a few software components
commonly used in real-time applications, Section 14.3.6.2 provides a brief overview
of how polling-loop-based applications may be implemented, Section 14.3.6.3 gives
a similar overview of streaming applications, and Section 14.3.6.4 briefly covers
event-based applications.
14.3.6.1 Real-Time Components
As in all areas of engineering, a robust set of components is essential to productivity and

reliability. This section is not a full catalog of real-time software components—such a
v2023.06.11a
catalog would fill multiple books—but rather a brief overview of the types of components
available.
A natural place to look for real-time software components would be algorithms
offering wait-free synchronization [Her91], and in fact lockless algorithms are very
important to real-time computing. However, wait-free synchronization only guarantees
forward progress in finite time. Although a century is finite, this is unhelpful when your
deadlines are measured in microseconds, let alone milliseconds.
Nevertheless, there are some important wait-free algorithms that do provide bounded
response time, including atomic test and set, atomic exchange, atomic fetch-and-add,
single-producer/single-consumer FIFO queues based on circular arrays, and numerous
per-thread partitioned algorithms. In addition, recent research has confirmed the
observation that algorithms with lock-free guarantees9 also provide the same latencies
in practice (in the wait-free sense), assuming a stochastically fair scheduler and absence
of fail-stop bugs [ACHS13]. This means that many non-wait-free stacks and queues are
nevertheless appropriate for real-time use.
Quick Quiz 14.12: But isn’t correct operation despite fail-stop bugs a valuable fault-tolerance
property?
In practice, locking is often used in real-time programs, theoretical concerns not-

withstanding. However, under more severe constraints, lock-based algorithms can also
provide bounded latencies [Bra11]. These constraints include:
1. Fair scheduler. In the common case of a fixed-priority scheduler, the bounded
latencies are provided only to the highest-priority threads.
2. Sufficient bandwidth to support the workload. An implementation rule supporting
this constraint might be “There will be at least 50 % idle time on all CPUs during
normal operation,” or, more formally, “The offered load will be sufficiently low to
allow the workload to be schedulable at all times.”
3. No fail-stop bugs.
4. FIFO locking primitives with bounded acquisition, handoff, and release latencies.
Again, in the common case of a locking primitive that is FIFO within priorities,
the bounded latencies are provided only to the highest-priority threads.
5. Some way of preventing unbounded priority inversion. The priority-ceiling and
priority-inheritance disciplines mentioned earlier in this chapter suffice.
6. Bounded nesting of lock acquisitions. We can have an unbounded number of locks,
but only as long as a given thread never acquires more than a few of them (ideally
only one of them) at a time.
7. Bounded number of threads. In combination with the earlier constraints, this
constraint means that there will be a bounded number of threads waiting on any
given lock.
8. Bounded time spent in any given critical section. Given a bounded number of
threads waiting on any given lock and a bounded critical-section duration, the wait
time will be bounded.
9 Wait-free algorithms guarantee that all threads make progress in finite time, while
lock-free algorithms only guarantee that at least one thread will make progress in finite time.
See Section 14.2 for more details.
v2023.06.11a
Quick Quiz 14.13: I couldn’t help but spot the word “include” before this list. Are there other
constraints?
This result opens a vast cornucopia of algorithms and data structures for use in
real-time software—and validates long-standing real-time practice.
Of course, a careful and simple application design is also extremely important. The
best real-time components in the world cannot make up for a poorly thought-out design.
For parallel real-time applications, synchronization overheads clearly must be a key
component of the design.
14.3.6.2 Polling-Loop Applications
Many real-time applications consist of a single CPU-bound loop that reads sensor data,
computes a control law, and writes control output. If the hardware registers providing
sensor data and taking control output are mapped into the application’s address space,
this loop might be completely free of system calls. But beware of the Spiderman
principle: With great power comes great responsibility, in this case the responsibility
to avoid bricking the hardware by making inappropriate references to the hardware
registers.
This arrangement is often run on bare metal, without the benefits of (or the interference
from) an operating system. However, increasing hardware capability and increasing
levels of automation motivates increasing software functionality, for example, user
interfaces, logging, and reporting, all of which can benefit from an operating system.
One way of gaining much of the benefit of running on bare metal while still having
access to the full features and functions of a general-purpose operating system is to use
the Linux kernel’s NO_HZ_FULL capability, described in Section 14.3.5.2.
14.3.6.3 Streaming Applications
One type of big-data real-time application takes input from numerous sources, processes
it internally, and outputs alerts and summaries. These streaming applications are often
highly parallel, processing different information sources concurrently.
One approach for implementing streaming applications is to use dense-array circular
FIFOs to connect different processing steps [Sut13]. Each such FIFO has only a single
thread producing into it and a (presumably different) single thread consuming from
it. Fan-in and fan-out points use threads rather than data structures, so if the output of
several FIFOs needed to be merged, a separate thread would input from them and output
to another FIFO for which this separate thread was the sole producer. Similarly, if the
output of a given FIFO needed to be split, a separate thread would input from this FIFO
and output to several FIFOs as needed.
This discipline might seem restrictive, but it allows communication among threads
with minimal synchronization overhead, and minimal synchronization overhead is
important when attempting to meet tight latency constraints. This is especially true
when the amount of processing for each step is small, so that the synchronization
overhead is significant compared to the processing overhead.
The individual threads might be CPU-bound, in which case the advice in Sec-
tion 14.3.6.2 applies. On the other hand, if the individual threads block waiting for data
from their input FIFOs, the advice of the next section applies.
v2023.06.11a
Listing 14.5: Timed-Wait Test Program

1 if (clock_gettime(CLOCK_REALTIME, &timestart) != 0) {
2 perror("clock_gettime 1");
3 exit(-1);
4 }
5 if (nanosleep(&timewait, NULL) != 0) {
6 perror("nanosleep");
7 exit(-1);
8 }
9 if (clock_gettime(CLOCK_REALTIME, &timeend) != 0) {
10 perror("clock_gettime 2");
11 exit(-1);
12 }
14.3.6.4 Event-Driven Applications
We will use fuel injection into a mid-sized industrial engine as a fanciful example for
event-driven applications. Under normal operating conditions, this engine requires that
the fuel be injected within a one-degree interval surrounding top dead center. If we
assume a 1,500-RPM rotation rate, we have 25 rotations per second, or about 9,000
degrees of rotation per second, which translates to 111 microseconds per degree. We
therefore need to schedule the fuel injection to within a time interval of about 100
microseconds.
Suppose that a timed wait was to be used to initiate fuel injection, although if you are
building an engine, I hope you supply a rotation sensor. We need to test the timed-wait
functionality, perhaps using the test program shown in Listing 14.5. Unfortunately, if
we run this program, we can get unacceptable timer jitter, even in a -rt kernel.
One problem is that POSIX CLOCK_REALTIME is, oddly enough, not intended for real-
time use. Instead, it means “realtime” as opposed to the amount of CPU time consumed
by a process or thread. For real-time use, you should instead use CLOCK_MONOTONIC.
However, even with this change, results are still unacceptable.
Another problem is that the thread must be raised to a real-time priority by using the
sched_setscheduler() system call. But even this change is insufficient, because we
can still see page faults. We also need to use the mlockall() system call to pin the
application’s memory, preventing page faults. With all of these changes, results might
finally be acceptable.
In other situations, further adjustments might be needed. It might be necessary to
affinity time-critical threads onto their own CPUs, and it might also be necessary to
affinity interrupts away from those CPUs. It might be necessary to carefully select
hardware and drivers, and it will very likely be necessary to carefully select kernel
configuration.
As can be seen from this example, real-time computing can be quite unforgiving.
14.3.6.5 The Role of RCU
Suppose that you are writing a parallel real-time application that needs to access data
that is subject to gradual change, perhaps due to changes in temperature, humidity, and
barometric pressure. The real-time response constraints on this program are so severe
that it is not permissible to spin or block, thus ruling out locking, nor is it permissible to
use a retry loop, thus ruling out sequence locks and hazard pointers. Fortunately, the
temperature and pressure are normally controlled, so that a default hard-coded set of
data is usually sufficient.
v2023.06.11a
Listing 14.6: Real-Time Calibration Using RCU

1 struct calibration {
2 short a;
3 short b;
4 short c;
5 };
6 struct calibration default_cal = { 62, 33, 88 };
7 struct calibration cur_cal = &default_cal;
8
9 short calc_control(short t, short h, short press)
10 {
11 struct calibration *p;
12
13 p = rcu_dereference(cur_cal);
14 return do_control(t, h, press, p->a, p->b, p->c);
15 }
16
17 bool update_cal(short a, short b, short c)
18 {
19 struct calibration *p;
20 struct calibration *old_p;
21
22 old_p = rcu_dereference(cur_cal);
23 p = malloc(sizeof(*p);
24 if (!p)
25 return false;
26 p->a = a;
27 p->b = b;
28 p->c = c;
29 rcu_assign_pointer(cur_cal, p);
30 if (old_p == &default_cal)
31 return true;
33 free(old_p);
34 return true;
35 }
However, the temperature, humidity, and pressure occasionally deviate too far from
the defaults, and in such situations it is necessary to provide data that replaces the
defaults. Because the temperature, humidity, and pressure change gradually, providing
the updated values is not a matter of urgency, though it must happen within a few minutes.
The program is to use a global pointer imaginatively named cur_cal that normally
references default_cal, which is a statically allocated and initialized structure that
contains the default calibration values in fields imaginatively named a, b, and c.
Otherwise, cur_cal points to a dynamically allocated structure providing the current
calibration values.
Listing 14.6 shows how RCU can be used to solve this problem. Lookups are
deterministic, as shown in calc_control() on lines 9–15, consistent with real-time
requirements. Updates are more complex, as shown by update_cal() on lines 17–35.
Quick Quiz 14.14: Given that real-time systems are often used for safety-critical applications,
and given that runtime memory allocation is forbidden in many safety-critical situations, what
is with the call to malloc()???
Quick Quiz 14.15: Don’t you need some kind of synchronization to protect update_cal()?
This example shows how RCU can provide deterministic read-side data-structure
access to real-time programs.
v2023.06.11a
Figure 14.15: The Dark Side of Real-Time Computing
Figure 14.16: The Dark Side of Real-Fast Computing
14.3.7 Real Time vs. Real Fast: How to Choose?

The choice between real-time and real-fast computing can be a difficult one. Because
real-time systems often inflict a throughput penalty on non-real-time computing, using
real-time when it is not required is unwise, as fancifully depicted by Figure 14.15.
On the other hand, failing to use real-time when it is required can also cause problems,
as fancifully depicted by Figure 14.16. It is almost enough to make you feel sorry for
the boss!
One rule of thumb uses the following four questions to help you choose:
1. Is average long-term throughput the only goal?

2. Is it permissible for heavy loads to degrade response times?
3. Is there high memory pressure, ruling out use of the mlockall() system call?
4. Does the basic work item of your application take more than 100 milliseconds to
complete?
v2023.06.11a
If the answer to any of these questions is “yes”, you should choose real-fast over
real-time, otherwise, real-time might be for you.
Choose wisely, and if you do choose real-time, make sure that your hardware, firmware,
and operating system are up to the job!
v2023.06.11a
The art of progress is to preserve order amid change
and to preserve change amid order.
Alfred North Whitehead
Chapter 15
Advanced Synchronization:
Memory Ordering
Causality and sequencing are deeply intuitive, and hackers often have a strong grasp
of these concepts. These intuitions can be quite helpful when writing, analyzing, and
debugging not only sequential code, but also parallel code that makes use of standard
mutual-exclusion mechanisms such as locking. Unfortunately, these intuitions break
down completely in code that instead uses weakly ordered atomic operations and
memory barriers. One example of such code implements the standard mutual-exclusion
mechanisms themselves, while another example implements fast paths that use weaker
synchronization. Insults to intuition notwithstanding, some argue that weakness is a
virtue [Alg13]. Virtue or vice, this chapter will help you gain an understanding of
memory ordering, that, with practice, will be sufficient to implement synchronization
primitives and performance-critical fast paths.
Section 15.1 will demonstrate that real computer systems can reorder memory
references, give some reasons why they do so, and provide some information on how
to prevent undesired reordering. Sections 15.2 and 15.3 will cover the types of pain
that hardware and compilers, respectively, can inflict on unwary parallel programmers.
Section 15.4 gives an overview of the benefits of modeling memory ordering at higher
levels of abstraction. Section 15.5 follows up with more detail on a few representative
hardware platforms. Finally, Section 15.6 provides some reliable intuitions and useful
rules of thumb.
Quick Quiz 15.1: This chapter has been rewritten since the first edition, and heavily edited
since the second edition. Did memory ordering change all that since 2014, let alone 2021?
15.1 Ordering: Why and How?

Nothing is orderly till people take hold of it.
Everything in creation lies around loose.
Henry Ward Beecher, updated
One motivation for memory ordering can be seen in the trivial-seeming litmus test in
Listing 15.1 (C-SB+o-o+o-o.litmus), which at first glance might appear to guarantee
485
v2023.06.11a
486 CHAPTER 15. ADVANCED SYNCHRONIZATION: MEMORY ORDERING
Listing 15.1: Memory Misordering: Store-Buffering Litmus Test

1 C C-SB+o-o+o-o
2
3 {}
4
5 P0(int *x0, int *x1)
6 {
7 int r2;
8
9 WRITE_ONCE(*x0, 2);
10 r2 = READ_ONCE(*x1);
11 }
12
13 P1(int *x0, int *x1)
14 {
15 int r2;
16
19 }
20
21 exists (1:r2=0 /\ 0:r2=0)
that the exists clause never triggers.1 After all, if 0:r2=0 as shown in the exists
clause,2 we might hope that Thread P0()’s load from x1 into r2 must have happened
before Thread P1()’s store to x1, which might raise further hopes that Thread P1()’s
load from x0 into r2 must happen after Thread P0()’s store to x0, so that 1:r2=2, thus
never triggering the exists clause. The example is symmetric, so similar reasoning
might lead us to hope that 1:r2=0 guarantees that 0:r2=2. Unfortunately, the lack
of memory barriers dashes these hopes. The CPU is within its rights to reorder the
statements within both Thread P0() and Thread P1(), even on relatively strongly
ordered systems such as x86.
Quick Quiz 15.2: The compiler can also reorder Thread P0()’s and Thread P1()’s memory
accesses in Listing 15.1, right?
This willingness to reorder can be confirmed using tools such as litmus7 [AMT14],
which found that the counter-intuitive ordering happened 314 times out of 100,000,000
trials on an x86 laptop. Oddly enough, the perfectly legal outcome where both loads
return the value 2 occurred less frequently, in this case, only 167 times.3 The lesson here
is clear: Increased counter-intuitivity does not necessarily imply decreased probability!
The following sections show exactly how this intuition breaks down, and then put
forward some mental models of memory ordering that can help you avoid these pitfalls.
Section 15.1.1 gives a brief overview of why hardware misorders memory accesses,
and then Section 15.1.2 gives an equally brief overview of how you can thwart such
misordering. Finally, Section 15.1.3 lists some basic rules of thumb, which will be
further refined in later sections. These sections focus on hardware reordering, but rest
assured that compilers reorder much more aggressively than hardware ever dreamed of
doing. Compiler-induced reordering will be taken up in Section 15.3.
Purists would instead insist that the exists clause is never satisfied, but we use “trigger”
1
here by analogy with assertions.

2 That is, Thread P0()’s instance of local variable r2 equals zero. See Section 12.2.1 for
documentation of litmus-test nomenclature.

3 Please note that results are sensitive to the exact hardware configuration, how heavily
the system is loaded, and much else besides. So why not try it out on your own system?
v2023.06.11a
15.1. ORDERING: WHY AND HOW? 487
128 Entry 32 KB Instruction Cache 4 µops

ITLB (8 way) Retirement Register File
96 Entry Reorder Buffer (ROB)
128 Bit
(Program Visible State)
4 µops
32 Byte Pre-Decode, Shared
32 Entry Reservation Station L2 Cache
Fetch Buffer
Instruction 6 Instructions
Port 0 Port 1 Port 5 Port 3 Port 4 Port 2 (16 way)
Fetch Unit
18 Entry SSE SSE
ALU SSE Store Store Load
ALU Shuffle ALU Shuffle
Instruction Queue ALU MUL
Branch ALU Address Data Address
256 Entry
128 Bit L2 DTLB
Micro- Complex Simple Simple Simple FMUL
128 Bit Memory Ordering Buffer
FADD
code Decoder Decoder Decoder Decoder FDIV (MOB)
4 µops 1 µop 1 µop 1 µop Store Load
Internal Results Bus 128 Bit 256
128 Bit Bit
7+ Entry µop Buffer Shared Bus
32 KB Dual Ported Data Cache 16 Entry
4 µops Interface
(8 way) DTLB
Unit
Register Alias Table
and Allocator
4 µops
Figure 15.1: Intel Core 2 Architecture
128 Entry 32 KB Instruction Cache 4 µops

ITLB (8 way) Retirement Register File
96 Entry Reorder Buffer (ROB)
128 Bit
(Program Visible State)
4 µops
32 Byte Pre-Decode, Shared
32 Entry Reservation Station L2 Cache
Fetch Buffer
Instruction 6 Instructions
Port 0 Port 1 Port 5 Port 3 Port 4 Port 2 256 (16 way)
Bit
PU
Fetch Unit
18 Entry SSE SSE
ALU SSE Store Store Load
ALU Shuffle ALU Shuffle
Instruction Queue ALU MUL
Branch ALU Address Data Address
256 Entry
128 Bit L2 DTLB
128 Bit Memory Ordering Buffer
Micro- Complex Simple Simple Simple FMUL r
code Decoder Decoder Decoder Decoder FDIV
FADD
uffe
(MOB)
eB
Cache
C
rStore
4 µops 1 µop 1 µop 1 µop
Internal Results Bus
Sto128 Bit
Load
128 Bit
7+ Entry µop Buffer Shared Bus
32 KB Dual Ported Data Cache 16 Entry
4 µops Interface
(8 way) DTLB
Unit
Register Alias Table
and Allocator
4 µops
Figure 15.2: Intel Core 2 Architecture Simplified
15.1.1 Why Hardware Misordering?
But why does memory misordering happen in the first place? Can’t CPUs keep track of
ordering on their own? Isn’t that why we have computers in the first place, to keep track
of things?
Many people do indeed expect their computers to keep track of things, but many also
insist that they keep track of things quickly. In fact, so intense is the focus on performance
that modern CPUs are extremely complex, as can be seen in the simplified block diagram
in Figure 15.1. Those needing to squeeze the last few percent of performance from their
systems will in turn need to pay close attention to the fine details of this figure when
tuning their software. Except that this close attention to detail means that when a given
CPU degrades with age, the software will no longer run quickly on it. For example, if
the leftmost ALU fails, software tuned to take full advantage of all of the ALUs might
well run more slowly than untuned software. One solution to this problem is to take
systems out of service as soon as any of their CPUs start degrading.
Another option is to recall the lessons of Chapter 3, especially the lesson that for
many important workloads, main memory cannot keep up with modern CPUs, which
can execute hundreds of instructions in the time required to fetch a single variable from
memory. For such workloads, the detailed internal structure of the CPU is irrelevant,
v2023.06.11a
CPU 0 CPU 1
Store Store
Buffer Buffer
Cache Cache
Memory
Figure 15.3: System Architecture With Store Buffers
Table 15.1: Memory Misordering: Store-Buffering Sequence of Events
CPU 0 CPU 1
Instruction Store Buffer Cache Instruction Store Buffer Cache
1 (Initial state) x1==0 (Initial state) x0==0
2 x0 = 2; x0==2 x1==0 x1 = 2; x1==2 x0==0
3 r2 = x1; (0) x0==2 x1==0 r2 = x0; (0) x1==2 x0==0
4 (Read-invalidate) x0==2 x0==0 (Read-invalidate) x1==2 x1==0
5 (Finish store) x0==2 (Finish store) x1==2
and the CPU can instead be approximated by the blue shapes in Figure 15.2 labeled
CPU, store buffer, and cache.
Because of these data-intensive workloads, CPUs sport increasingly large caches,
as was seen back in Figure 3.11, which means that although the first load by a given
CPU from a given variable will result in an expensive cache miss as was discussed in
Section 3.1.6, subsequent repeated loads from that variable by that CPU might execute
very quickly because the initial cache miss will have loaded that variable into that CPU’s
cache.
However, it is also necessary to accommodate frequent concurrent stores from
multiple CPUs to a set of shared variables. In cache-coherent systems, if the caches hold
multiple copies of a given variable, all the copies of that variable must have the same
value. This works extremely well for concurrent loads, but not so well for concurrent
stores: Each store must do something about all copies of the old value (another cache
miss!), which, given the finite speed of light and the atomic nature of matter, will be
slower than impatient software hackers would like. And these strings of stores are the
reason for the blue block labelled store buffer in Figure 15.2.
Removing the internal CPU complexity from Figure 15.2, adding a second CPU, and
showing main memory results in Figure 15.3. When a given CPU stores to a variable
not present in that CPU’s cache, then the new value is instead placed in that CPU’s store
buffer. The CPU can then proceed immediately, without having to wait for the store to
do something about all the old values of that variable residing in other CPUs’ caches.
Although store buffers can greatly increase performance, they can cause instructions
and memory references to execute out of order, which can in turn cause serious confusion,
as fancifully illustrated in Figure 15.4.
v2023.06.11a
do I out things of
Look! can order.
Figure 15.4: CPUs Can Do Things Out of Order
In particular, store buffers cause the memory misordering illustrated by Listing 15.1.
Table 15.1 shows the steps leading to this misordering. Row 1 shows the initial state,
where CPU 0 has x1 in its cache and CPU 1 has x0 in its cache, both variables having a
value of zero. Row 2 shows the state change due to each CPU’s store (lines 9 and 17 of
Listing 15.1). Because neither CPU has the stored-to variable in its cache, both CPUs
record their stores in their respective store buffers.
Quick Quiz 15.3: But wait!!! On row 2 of Table 15.1 both x0 and x1 each have two values at
the same time, namely zero and two. How can that possibly work???
Row 3 shows the two loads (lines 10 and 18 of Listing 15.1). Because the variable
being loaded by each CPU is in that CPU’s cache, each load immediately returns the
cached value, which in both cases is zero.
But the CPUs are not done yet: Sooner or later, they must empty their store buffers.
Because caches move data around in relatively large blocks called cachelines, and
because each cacheline can hold several variables, each CPU must get the cacheline into
its own cache so that it can update the portion of that cacheline corresponding to the
variable in its store buffer, but without disturbing any other part of the cacheline. Each
CPU must also ensure that the cacheline is not present in any other CPU’s cache, for
which a read-invalidate operation is used. As shown on row 4, after both read-invalidate
operations complete, the two CPUs have traded cachelines, so that CPU 0’s cache now
contains x0 and CPU 1’s cache now contains x1. Once these two variables are in
their new homes, each CPU can flush its store buffer into the corresponding cache line,
leaving each variable with its final value as shown on row 5.
Quick Quiz 15.4: But don’t the values also need to be flushed from the cache to main memory?
In summary, store buffers are needed to allow CPUs to handle store instructions
efficiently, but they can result in counter-intuitive memory misordering.
But what do you do if your algorithm really needs its memory references to be
ordered? For example, suppose that you are communicating with a driver using a pair of
flags, one that says whether or not the driver is running and the other that says whether
v2023.06.11a
Listing 15.2: Memory Ordering: Store-Buffering Litmus Test

1 C C-SB+o-mb-o+o-mb-o
2
3 {}
4
6 {
7 int r2;
8
10 smp_mb();
12 }
13
14 P1(int *x0, int *x1)
15 {
16 int r2;
17
19 smp_mb();
21 }
22
23 exists (1:r2=0 /\ 0:r2=0)
there is a request pending for that driver. The requester needs to set the request-pending
flag, then check the driver-running flag, and if false, wake the driver. Once the driver has
serviced all the pending requests that it knows about, it needs to clear its driver-running
flag, then check the request-pending flag to see if it needs to restart. This very reasonable
approach cannot work unless there is some way to make sure that the hardware processes
the stores and loads in order. This is the subject of the next section.
15.1.2 How to Force Ordering?

It turns out that there are compiler directives and synchronization primitives (such as
locking and RCU) that are responsible for maintaining the illusion of ordering through
use of memory barriers (for example, smp_mb() in the Linux kernel). These memory
barriers can be explicit instructions, as they are on Arm, POWER, Itanium, and Alpha,
or they can be implied by other instructions, as they often are on x86. Since these
standard synchronization primitives preserve the illusion of ordering, your path of least
resistance is to simply use these primitives, thus allowing you to stop reading this
section.
However, if you need to implement the synchronization primitives themselves, or
if you are simply interested in understanding how memory ordering works, read on!
The first stop on the journey is Listing 15.2 (C-SB+o-mb-o+o-mb-o.litmus), which
places an smp_mb() Linux-kernel full memory barrier between the store and load in
both P0() and P1(), but is otherwise identical to Listing 15.1. These barriers prevent
the counter-intuitive outcome from happening on 100,000,000 trials on my x86 laptop.
Interestingly enough, the added overhead due to these barriers causes the legal outcome
where both loads return the value two to happen more than 800,000 times, as opposed
to only 167 times for the barrier-free code in Listing 15.1.
These barriers have a profound effect on ordering, as can be seen in Table 15.2.
Although the first two rows are the same as in Table 15.1 and although the smp_mb()
instructions on row 3 do not change state in and of themselves, they do cause the stores
to complete (rows 4 and 5) before the loads (row 6), which rules out the counter-intuitive
outcome shown in Table 15.1. Note that variables x0 and x1 each still have more than
v2023.06.11a
Table 15.2: Memory Ordering: Store-Buffering Sequence of Events
CPU 0 CPU 1
2 x0 = 2; x0==2 x1==0 x1 = 2; x1==2 x0==0
3 smp_mb(); x0==2 x1==0 smp_mb(); x1==2 x0==0
6 r2 = x1; (2) x1==2 r2 = x0; (2) x0==2
one value on row 2, however, as promised earlier, the smp_mb() invocations straighten
things out in the end.
Although full barriers such as smp_mb() have extremely strong ordering guarantees,
their strength comes at a high price in terms of foregone hardware and compiler
optimizations. A great many situations can be handled with much weaker ordering
guarantees that use much cheaper memory-ordering instructions, or, in some case, no
memory-ordering instructions at all.
Table 15.3 provides a cheatsheet of the Linux kernel’s ordering primitives and their
guarantees. Each row corresponds to a primitive or category of primitives that might
or might not provide ordering, with the columns labeled “Prior Ordered Operation”
and “Subsequent Ordered Operation” being the operations that might (or might not) be
ordered against. Cells containing “Y” indicate that ordering is supplied unconditionally,
while other characters indicate that ordering is supplied only partially or conditionally.
Blank cells indicate that no ordering is supplied.
The “Store” row also covers the store portion of an atomic RMW operation. In
addition, the “Load” row covers the load component of a successful value-returning
_relaxed() RMW atomic operation, although the combined “_relaxed() RMW
operation” line provides a convenient combined reference in the value-returning case. A
CPU executing unsuccessful value-returning atomic RMW operations must invalidate
the corresponding variable from all other CPUs’ caches. Therefore, unsuccessful value-
returning atomic RMW operations have many of the properties of a store, which means
that the “_relaxed() RMW operation” line also applies to unsuccessful value-returning
atomic RMW operations.
The *_acquire row covers smp_load_acquire(), cmpxchg_acquire(), xchg_
acquire(), and so on; the *_release row covers smp_store_release(), rcu_
assign_pointer(), cmpxchg_release(), xchg_release(), and so on; and
the “Successful full-strength non-void RMW” row covers atomic_add_return(),
atomic_add_unless(), atomic_dec_and_test(), cmpxchg(), xchg(), and so
on. The “Successful” qualifiers apply to primitives such as atomic_add_unless(),
cmpxchg_acquire(), and cmpxchg_release(), which have no effect on either mem-
ory or on ordering when they indicate failure, as indicated by the earlier “_relaxed()
RMW operation” row.
Column “C” indicates cumulativity and propagation, as explained in Sections 15.2.7.1
and 15.2.7.2. In the meantime, this column can usually be ignored when there are at
most two threads involved.
v2023.06.11a
Table 15.3: Linux-Kernel Memory-Ordering Cheat Sheet
Prior Ordered Operation Subsequent Ordered Operation

Operation Providing Ordering C Self R W RMW Self R W DR DW RMW SV
Store, for example, WRITE_ONCE() Y Y
Load, for example, READ_ONCE() Y Y Y Y
_relaxed() RMW operation Y Y Y Y
*_dereference() Y Y Y Y
Successful *_acquire() R Y Y Y Y Y Y
Successful *_release() C Y Y Y W Y
smp_rmb() Y R Y Y R
smp_wmb() Y W Y Y W
smp_mb() and synchronize_rcu() CP Y Y Y Y Y Y Y Y
Successful full-strength non-void RMW CP Y Y Y Y Y Y Y Y Y Y Y
smp_mb__before_atomic() CP Y Y Y a a a a Y
smp_mb__after_atomic() CP a a Y Y Y Y Y Y
Key: C: Ordering is cumulative

P: Ordering propagates
R: Read, for example, READ_ONCE(), or read portion of RMW
W: Write, for example, WRITE_ONCE(), or write portion of RMW
Y: Provides the specified ordering
a: Provides specified ordering given intervening RMW atomic operation
DR: Dependent read (address dependency, Section 15.2.3)
DW: Dependent write (address, data, or control dependency, Sections 15.2.3–15.2.5)
RMW: Atomic read-modify-write operation
Self: Orders self, as opposed to accesses both before and after
SV: Orders later accesses to the same variable
Applies to Linux kernel v4.15 and later.
Quick Quiz 15.5: The rows in Table 15.3 seem quite random and confused. Whatever is the
conceptual basis of this table???
Quick Quiz 15.6: Why is Table 15.3 missing smp_mb__after_unlock_lock() and smp_
mb__after_spinlock()?
It is important to note that this table is just a cheat sheet, and is therefore in no way a
replacement for a good understanding of memory ordering. To begin building such an
understanding, the next section will present some basic rules of thumb.
15.1.3 Basic Rules of Thumb

This section presents some basic rules of thumb that are “good and sufficient” for a
great many situations. In fact, you could write a great deal of concurrent code having
excellent performance and scalability without needing anything more than these rules
of thumb. More sophisticated rules of thumb will be presented in Section 15.6.
Quick Quiz 15.7: But how can I know that a given project can be designed and coded within
the confines of these rules of thumb?
v2023.06.11a
CPU 0
Memory
.... memory barriers guarantee X0 before X1.
Reference X0
Memory
Barrier
Memory
Reference Y0
CPU 1
Given Y0 before Y1 ... Memory

Reference Y1
Memory
Barrier
Memory
Reference X1
Figure 15.5: Memory Barriers Provide Conditional If-Then Ordering
A given thread sees its own accesses in order. This rule assumes that loads and
stores from/to shared variables use READ_ONCE() and WRITE_ONCE(), respectively.
Otherwise, the compiler can profoundly scramble4 your code, and sometimes the CPU
can do a bit of scrambling as well, as discussed in Section 15.5.4.
Interrupts and signal handlers are part of a thread. Both interrupt and signal
handlers happen between a pair of adjacent instructions in a thread. This means that a
given handler appears to execute atomically from the viewpoint of the interrupted thread,
at least at the assembly-language level. However, the C and C++ languages do not define
the results of handlers and interrupted threads sharing plain variables. Instead, such
shared variables must be sig_atomic_t, lock-free atomics, or volatile.
On the other hand, because the handler executes within the interrupted thread’s
context, the memory ordering used to synchronize communication between the handler
and the thread can be extremely lightweight. For example, the counterpart of an acquire
load is a READ_ONCE() followed by a barrier() compiler directive and the counterpart
of a release store is a barrier() followed by a WRITE_ONCE(). The counterpart of a
full memory barrier is barrier(). Finally, disabling interrupts or signals (as the case
may be) within the thread excludes handlers.
Ordering has conditional if-then semantics. Figure 15.5 illustrates this for memory
barriers. Assuming that both memory barriers are strong enough, if CPU 1’s access Y1
happens after CPU 0’s access Y0, then CPU 1’s access X1 is guaranteed to happen after
CPU 0’s access X0. When in doubt as to which memory barriers are strong enough,
smp_mb() will always do the job, albeit at a price.
Quick Quiz 15.8: How can you tell which memory barriers are strong enough for a given use
case?
4 Many compiler writers prefer the word “optimize”.
v2023.06.11a
Listing 15.2 is a case in point. The smp_mb() on lines 10 and 19 serve as the barriers,
the store to x0 on line 9 as X0, the load from x1 on line 11 as Y0, the store to x1 on
line 18 as Y1, and the load from x0 on line 20 as X1. Applying the if-then rule step by
step, we know that the store to x1 on line 18 happens after the load from x1 on line 11
if P0()’s local variable r2 is set to the value zero. The if-then rule would then state
that the load from x0 on line 20 happens after the store to x0 on line 9. In other words,
P1()’s local variable r2 is guaranteed to end up with the value two only if P0()’s
local variable r2 ends up with the value zero. This underscores the point that memory
ordering guarantees are conditional, not absolute.
Although Figure 15.5 specifically mentions memory barriers, this same if-then rule
applies to the rest of the Linux kernel’s ordering operations.
Ordering operations must be paired. If you carefully order the operations in one
thread, but then fail to do so in another thread, then there is no ordering. Both threads
must provide ordering for the if-then rule to apply.5
Ordering operations almost never speed things up. If you find yourself tempted to
add a memory barrier in an attempt to force a prior store to be flushed to memory faster,
resist! Adding ordering usually slows things down. Of course, there are situations
where adding instructions speeds things up, as was shown by Figure 9.22 on page 254,
but careful benchmarking is required in such cases. And even then, it is quite possible
that although you sped things up a little bit on your system, you might well have slowed
things down significantly on your users’ systems. Or on your future system.
Ordering operations are not magic. When your program is failing due to some
race condition, it is often tempting to toss in a few memory-ordering operations in an
attempt to barrier your bugs out of existence. A far better reaction is to use higher-level
primitives in a carefully designed manner. With concurrent programming, it is almost
always better to design your bugs out of existence than to hack them down to lower
probabilities.
These are only rough rules of thumb. Although these rules of thumb cover the vast
majority of situations seen in actual practice, as with any set of rules of thumb, they do
have their limits. The next section will demonstrate some of these limits by introducing
trick-and-trap litmus tests that are intended to insult your intuition while increasing your
understanding. These litmus tests will also illuminate many of the concepts represented
by the Linux-kernel memory-ordering cheat sheet shown in Table 15.3, and can be
automatically analyzed given proper tooling [AMM+ 18]. Section 15.6 will circle back
to this cheat sheet, presenting a more sophisticated set of rules of thumb in light of
learnings from all the intervening tricks and traps.
Quick Quiz 15.9: Wait!!! Where do I find this tooling that automatically analyzes litmus
tests???
5 In Section 15.2.7.2, pairing will be generalized to cycles.
v2023.06.11a
15.2. TRICKS AND TRAPS 495
Listing 15.3: Software Logic Analyzer

1 state.variable = mycpu;
2 lasttb = oldtb = firsttb = gettb();
3 while (state.variable == mycpu) {
4 lasttb = oldtb;
5 oldtb = gettb();
6 if (lasttb - firsttb > 1000)
7 break;
8 }
15.2 Tricks and Traps
Knowing where the trap is—that’s the first step in

evading it.
Duke Leto Atreides, Dune, Frank Herbert
Now that you know that hardware can reorder memory accesses and that you can prevent
it from doing so, the next step is to get you to admit that your intuition has a problem.
This painful task is taken up by Section 15.2.1, which presents some code demonstrating
that scalar variables can take on multiple values simultaneously, and by Sections 15.2.2
through 15.2.7, which show a series of intuitively correct code fragments that fail
miserably on real hardware. Once your intuition has made it through the grieving
process, later sections will summarize the basic rules that memory ordering follows.
But first, let’s take a quick look at just how many values a single variable might have
at a single point in time.
15.2.1 Variables With Multiple Values

It is natural to think of a variable as taking on a well-defined sequence of values in a
well-defined, global order. Unfortunately, the next stop on the journey says “goodbye”
to this comforting fiction. Hopefully, you already started to say “goodbye” in response
to row 2 of Tables 15.1 and 15.2, and if so, the purpose of this section is to drive this
point home.
To this end, consider the program fragment shown in Listing 15.3. This code fragment
is executed in parallel by several CPUs. Line 1 sets a shared variable to the current
CPU’s ID, line 2 initializes several variables from a gettb() function that delivers
the value of a fine-grained hardware “timebase” counter that is synchronized among
all CPUs (not available from all CPU architectures, unfortunately!), and the loop from
lines 3–8 records the length of time that the variable retains the value that this CPU
assigned to it. Of course, one of the CPUs will “win”, and would thus never exit the
loop if not for the check on lines 6–7.
Quick Quiz 15.10: What assumption is the code fragment in Listing 15.3 making that might
not be valid on real hardware?
Upon exit from the loop, firsttb will hold a timestamp taken shortly after the
assignment and lasttb will hold a timestamp taken before the last sampling of the
shared variable that still retained the assigned value, or a value equal to firsttb if
the shared variable had changed before entry into the loop. This allows us to plot each
CPU’s view of the value of state.variable over a 532-nanosecond time period, as
shown in Figure 15.6. This data was collected in 2006 on 1.5 GHz POWER5 system
v2023.06.11a
CPU 1 1 2
CPU 2 2
CPU 3 3 2
CPU 4 4 2
0 100 200 300 400 500 (ns)
Figure 15.6: A Variable With Multiple Simultaneous Values
with 8 cores, each containing a pair of hardware threads. CPUs 1, 2, 3, and 4 recorded
the values, while CPU 0 controlled the test. The timebase counter period was about
5.32 ns, sufficiently fine-grained to allow observations of intermediate cache states.
Each horizontal bar represents the observations of a given CPU over time, with
the gray regions to the left indicating the time before the corresponding CPU’s first
measurement. During the first 5 ns, only CPU 3 has an opinion about the value of the
variable. During the next 10 ns, CPUs 2 and 3 disagree on the value of the variable,
but thereafter agree that the value is “2”, which is in fact the final agreed-upon value.
However, CPU 1 believes that the value is “1” for almost 300 ns, and CPU 4 believes
that the value is “4” for almost 500 ns.
Quick Quiz 15.11: How could CPUs possibly have different views of the value of a single
variable at the same time?
Quick Quiz 15.12: Why do CPUs 2 and 3 come to agreement so quickly, when it takes so
long for CPUs 1 and 4 to come to the party?
And if you think that the situation with four CPUs was intriguing, consider Figure 15.7,
which shows the same situation, but with 15 CPUs each assigning their number to a
single shared variable at time 𝑡 = 0. Both diagrams in the figure are drawn in the same
way as Figure 15.6. The only difference is that the unit of horizontal axis is timebase
ticks, with each tick lasting about 5.3 nanoseconds. The entire sequence therefore lasts
a bit longer than the events recorded in Figure 15.6, consistent with the increase in
number of CPUs. The upper diagram shows the overall picture, while the lower one
zooms in on the first 50 timebase ticks. Again, CPU 0 coordinates the test, so does not
record any values.
All CPUs eventually agree on the final value of 9, but not before the values 15 and 12
take early leads. Note that there are fourteen different opinions on the variable’s value at
time 21 indicated by the vertical line in the lower diagram. Note also that all CPUs see
sequences whose orderings are consistent with the directed graph shown in Figure 15.8.
Nevertheless, these figures underscore the importance of proper use of memory-ordering
operations.
How many values can a single variable take on at a single point in time? As many as
one per store buffer in the system! We have therefore entered a regime where we must
bid a fond farewell to comfortable intuitions about values of variables and the passage
of time. This is the regime where memory-ordering operations are needed.
But remember well the lessons from Chapters 3 and 6. Having all CPUs store
concurrently to the same variable is no way to design a parallel program, at least not if
performance and scalability are at all important to you.
Unfortunately, memory ordering has many other ways of insulting your intuition,
and not all of these ways conflict with performance and scalability. The next section
overviews reordering of unrelated memory reference.
v2023.06.11a
CPU 1 1 6 4 10 15 3 9
CPU 2 2 3 9
CPU 3 3 9
CPU 4 4 10 15 12 9
CPU 5 5 10 15 12 9
CPU 6 6 2 15 9
CPU 7 7 2 15 9
CPU 8 8 9
CPU 9 9
CPU 10 10 15 12 9
CPU 11 11 10 15 12 9
CPU 12 12 9
CPU 13 13 12 9
CPU 14 14 15 12 9
CPU 15 15 12 9
0 50 100 150 200 250 300 350 400 450 500 (tick)
CPU 1 1
CPU 2 2
CPU 3 3
CPU 4 4
CPU 5 5
CPU 6 6
CPU 7 7
CPU 8 8 9
CPU 9 9
CPU 10 10
CPU 11 11
CPU 12 12
CPU 13 13
CPU 14 14 15
CPU 15 15
0 5 10 15 20 25 30 35 40 45 (tick)
Figure 15.7: A Variable With More Simultaneous Values
v2023.06.11a
7 6
2 4 5 11
10 14
15 13
3 12 8
Figure 15.8: Possible Global Orders With More Simultaneous Values
Listing 15.4: Message-Passing Litmus Test (No Ordering)

1 C C-MP+o-wmb-o+o-o
2
3 {}
4
5 P0(int* x0, int* x1) {
7 smp_wmb();
9 }
10
11 P1(int* x0, int* x1) {
12 int r2;
13 int r3;
14
17 }
18
19 exists (1:r2=2 /\ 1:r3=0)
15.2.2 Memory-Reference Reordering

Section 15.1.1 showed that even relatively strongly ordered systems like x86 can reorder
prior stores with later loads, at least when the store and load are to different variables.
This section builds on that result, looking at the other combinations of loads and stores.
15.2.2.1 Load Followed By Load

Listing 15.4 (C-MP+o-wmb-o+o-o.litmus) shows the classic message-passing litmus
test, where x0 is the message and x1 is a flag indicating whether or not a message
is available. In this test, the smp_wmb() forces P0() stores to be ordered, but no
ordering is specified for the loads. Relatively strongly ordered architectures, such as x86,
do enforce ordering. However, weakly ordered architectures often do not [AMP+ 11].
Therefore, the exists clause on line 19 of the listing can trigger.
v2023.06.11a
Listing 15.5: Enforcing Order of Message-Passing Litmus Test

1 C C-MP+o-wmb-o+o-rmb-o
2
3 {}
4
5 P0(int* x0, int* x1) {
7 smp_wmb();
9 }
10
11 P1(int* x0, int* x1) {
12 int r2;
13 int r3;
14
16 smp_rmb();
18 }
19
20 exists (1:r2=2 /\ 1:r3=0)
Listing 15.6: Load-Buffering Litmus Test (No Ordering)

1 C C-LB+o-o+o-o
2
3 {}
4
6 {
7 int r2;
8
11 }
12
13 P1(int *x0, int *x1)
14 {
15 int r2;
16
19 }
20
21 exists (1:r2=2 /\ 0:r2=2)
One rationale for reordering loads from different locations is that doing so allows
execution to proceed when an earlier load misses the cache, but the values for later loads
are already present.
Quick Quiz 15.13: But why make load-load reordering visible to the user? Why not just use
speculative execution to allow execution to proceed in the common case where there are no
intervening stores, in which case the reordering cannot be visible anyway?
Thus, portable code relying on ordered loads must add explicit ordering, for example,
the smp_rmb() shown on line 16 of Listing 15.5 (C-MP+o-wmb-o+o-rmb-o.litmus),
which prevents the exists clause from triggering.
15.2.2.2 Load Followed By Store
Listing 15.6 (C-LB+o-o+o-o.litmus) shows the classic load-buffering litmus test.

Although relatively strongly ordered systems such as x86 or the IBM Mainframe do not
reorder prior loads with subsequent stores, many weakly ordered architectures really do
v2023.06.11a
Listing 15.7: Enforcing Ordering of Load-Buffering Litmus Test

1 C C-LB+o-r+a-o
2
3 {}
4
6 {
7 int r2;
8
10 smp_store_release(x0, 2);
11 }
12
13 P1(int *x0, int *x1)
14 {
15 int r2;
16
17 r2 = smp_load_acquire(x0);
19 }
20
21 exists (1:r2=2 /\ 0:r2=2)
Listing 15.8: Message-Passing Litmus Test, No Writer Ordering (No Ordering)

1 C C-MP+o-o+o-rmb-o
2
3 {}
4
5 P0(int* x0, int* x1) {
8 }
9
10 P1(int* x0, int* x1) {
11 int r2;
12 int r3;
13
15 smp_rmb();
17 }
18
19 exists (1:r2=2 /\ 1:r3=0)
allow such reordering [AMP+ 11]. Therefore, the exists clause on line 21 really can
trigger.
Although it is rare for actual hardware to exhibit this reordering [Mar17], one situation
where it might be desirable to do so is when a load misses the cache, the store buffer
is nearly full, and the cacheline for a subsequent store is ready at hand. Therefore,
portable code must enforce any required ordering, for example, as shown in Listing 15.7
(C-LB+o-r+a-o.litmus). The smp_store_release() and smp_load_acquire()
guarantee that the exists clause on line 21 never triggers.
15.2.2.3 Store Followed By Store

Listing 15.8 (C-MP+o-o+o-rmb-o.litmus) once again shows the classic message-
passing litmus test, with the smp_rmb() providing ordering for P1()’s loads, but without
any ordering for P0()’s stores. Again, the relatively strongly ordered architectures do
enforce ordering, but weakly ordered architectures do not necessarily do so [AMP+ 11],
which means that the exists clause can trigger. One situation in which such reordering
could be beneficial is when the store buffer is full, another store is ready to execute,
but the cacheline needed by the oldest store is not yet available. In this situation,
v2023.06.11a
Listing 15.9: Message-Passing Address-Dependency Litmus Test (No Ordering Before v4.15)
1 C C-MP+o-wmb-o+o-ad-o
2
3 {
4 y=1;
5 x1=y;
6 }
7
8 P0(int* x0, int** x1) {
10 smp_wmb();
11 WRITE_ONCE(*x1, x0);
12 }
13
14 P1(int** x1) {
15 int *r2;
16 int r3;
17
20 }
21
22 exists (1:r2=x0 /\ 1:r3=1)
allowing stores to complete out of order would allow execution to proceed. Therefore,
portable code must explicitly order the stores, for example, as shown in Listing 15.5,
thus preventing the exists clause from triggering.
Quick Quiz 15.14: Why should strongly ordered systems pay the performance price of
unnecessary smp_rmb() and smp_wmb() invocations? Shouldn’t weakly ordered systems
shoulder the full cost of their misordering choices???
15.2.3 Address Dependencies

An address dependency occurs when the value returned by a load instruction is used
to compute the address used by a later memory-reference instruction. This means
that the exact same sequence of instructions used to traverse a linked data structure in
single-threaded code provides weak but extremely useful ordering in concurrent code.
Listing 15.9 (C-MP+o-wmb-o+o-addr-o.litmus) shows a linked variant of the
message-passing pattern. The head pointer is x1, which initially references the int
variable y (line 5), which is in turn initialized to the value 1 (line 4). P0() updates
head pointer x1 to reference x0 (line 11), but only after initializing it to 2 (line 9) and
forcing ordering (line 10). P1() picks up the head pointer x1 (line 18), and then loads
the referenced value (line 19). There is thus an address dependency from the load on
line 18 to the load on line 19. In this case, the value returned by line 18 is exactly the
address used by line 19, but many variations are possible, including field access using
the C-language -> operator, addition, subtraction, and array indexing.6
One might hope that line 18’s load from the head pointer would be ordered before
line 19’s dereference, which is in fact the case on Linux v4.15 and later. However, prior
to v4.15, this was not the case on DEC Alpha, which could in effect use a speculated
value for the dependent load, as described in more detail in Section 15.5.1. Therefore,
on older versions of Linux, Listing 15.9’s exists clause can trigger.
Listing 15.10 shows how to make this work reliably on pre-v4.15 Linux kernels
running on DEC Alpha, by replacing READ_ONCE() on line 18 of Listing 15.9 with
6 But note that in the Linux kernel, the address dependency must be carried through the
pointer to the array, not through the array index.
v2023.06.11a
Listing 15.10: Enforced Ordering of Message-Passing Address-Dependency Litmus Test (Before

v4.15)
1 C C-MP+o-wmb-o+ld-addr-o
2
3 {
4 y=1;
5 x1=y;
6 }
7
8 P0(int* x0, int** x1) {
10 smp_wmb();
12 }
13
14 P1(int** x1) {
15 int *r2;
16 int r3;
17
18 r2 = lockless_dereference(*x1); // Obsolete
20 }
21
22 exists (1:r2=x0 /\ 1:r3=1)
Listing 15.11: S Address-Dependency Litmus Test

1 C C-S+o-wmb-o+o-addr-o
2
3 {
4 y=1;
5 x1=y;
6 }
7
8 P0(int* x0, int** x1) {
10 smp_wmb();
12 }
13
14 P1(int** x1) {
15 int *r2;
16
18 WRITE_ONCE(*r2, 3);
19 }
20
21 exists (1:r2=x0 /\ x0=2)
lockless_dereference(),7 which acts like READ_ONCE() on all platforms other

than DEC Alpha, where it acts like a READ_ONCE() followed by an smp_mb(), thereby
forcing the required ordering on all platforms, in turn preventing the exists clause
from triggering.
But what happens if the dependent operation is a store rather than a load, for
example, in the S litmus test [AMP+ 11] shown in Listing 15.11 (C-S+o-wmb-o+o-
addr-o.litmus)? Because no production-quality platform speculates stores, it is not
possible for the WRITE_ONCE() on line 9 to overwrite the WRITE_ONCE() on line 18,
meaning that the exists clause on line 21 cannot trigger, even on DEC Alpha, even in
pre-v4.15 Linux kernels.
7 Note that lockless_dereference() is not needed on v4.15 and later, and therefore
is not available in these later Linux kernels. Nor is it needed in versions of this book containing
this footnote.
v2023.06.11a
Listing 15.12: Load-Buffering Data-Dependency Litmus Test

1 C C-LB+o-r+o-data-o
2
3 {}
4
6 {
7 int r2;
8
11 }
12
13 P1(int *x0, int *x1)
14 {
15 int r2;
16
18 WRITE_ONCE(*x1, r2);
19 }
20
21 exists (1:r2=2 /\ 0:r2=2)
Quick Quiz 15.15: But how do we know that all platforms really avoid triggering the exists
clauses in Listings 15.10 and 15.11?
Quick Quiz 15.16: Why the use of smp_wmb() in Listings 15.10 and 15.11? Wouldn’t
smp_store_release() be a better choice?
Quick Quiz 15.17: SP, MP, LB, and now S. Where do all these litmus-test abbreviations come
from and how can anyone keep track of them?
However, it is important to note that address dependencies can be fragile and easily
broken by compiler optimizations, as discussed in Section 15.3.2.
15.2.4 Data Dependencies

A data dependency occurs when the value returned by a load instruction is used to
compute the data stored by a later store instruction. Note well the “data” above: If
the value returned by a load was instead used to compute the address used by a later
store instruction, that would instead be an address dependency, which was covered in
Section 15.2.3. However, the existence of data dependencies means that the exact same
sequence of instructions used to update a linked data structure in single-threaded code
provides weak but extremely useful ordering in concurrent code.
Listing 15.12 (C-LB+o-r+o-data-o.litmus) is similar to Listing 15.7, except
that P1()’s ordering between lines 17 and 18 is enforced not by an acquire load, but
instead by a data dependency: The value loaded by line 17 is what line 18 stores. The
ordering provided by this data dependency is sufficient to prevent the exists clause
from triggering.
Just as with address dependencies, data dependencies are fragile and can be easily
broken by compiler optimizations, as discussed in Section 15.3.2. In fact, data
dependencies can be even more fragile than are address dependencies. The reason
for this is that address dependencies normally involve pointer values. In contrast, as
shown in Listing 15.12, it is tempting to carry data dependencies through integral values,
which the compiler has much more freedom to optimize into nonexistence. For but one
example, if the integer loaded was multiplied by the constant zero, the compiler would
v2023.06.11a
Listing 15.13: Load-Buffering Control-Dependency Litmus Test

1 C C-LB+o-r+o-ctrl-o
2
3 {}
4
6 {
7 int r2;
8
11 }
12
13 P1(int *x0, int *x1)
14 {
15 int r2;
16
18 if (r2 >= 0)
20 }
21
22 exists (1:r2=2 /\ 0:r2=2)
know that the result was zero, and could therefore substitute the constant zero for the
value loaded, thus breaking the dependency.
Quick Quiz 15.18: But wait!!! Line 17 of Listing 15.12 uses READ_ONCE(), which marks
the load as volatile, which means that the compiler absolutely must emit the load instruction
even if the value is later multiplied by zero. So how can the compiler possibly break this data
dependency?
In short, you can rely on data dependencies only if you prevent the compiler from
breaking them.
15.2.5 Control Dependencies

A control dependency occurs when the value returned by a load instruction is tested
to determine whether or not a later store instruction is executed. In other words, a
simple conditional branch or conditional-move instruction can act as a weak but low-
overhead memory-barrier instruction. However, note well the “later store instruction”:
Although all platforms respect load-to-store dependencies, many platforms do not
respect load-to-load control dependencies.
Listing 15.13 (C-LB+o-r+o-ctrl-o.litmus) shows another load-buffering exam-
ple, this time using a control dependency (line 18) to order the load on line 17 and the
store on line 19. The ordering is sufficient to prevent the exists from triggering.
However, control dependencies are even more susceptible to being optimized out of
existence than are data dependencies, and Section 15.3.3 describes some of the rules
that must be followed in order to prevent your compiler from breaking your control
dependencies.
It is worth reiterating that control dependencies provide ordering only from loads
to stores. Therefore, the load-to-load control dependency shown on lines 14–16 of
Listing 15.14 (C-MP+o-r+o-ctrl-o.litmus) does not provide ordering, and therefore
does not prevent the exists clause from triggering.
In summary, control dependencies can be useful, but they are high-maintenance items.
You should therefore use them only when performance considerations permit no other
solution.
v2023.06.11a
Listing 15.14: Message-Passing Control-Dependency Litmus Test (No Ordering)

1 C C-MP+o-r+o-ctrl-o
2
3 {}
4
5 P0(int* x0, int* x1) {
8 }
9
10 P1(int* x0, int* x1) {
11 int r2;
12 int r3 = 0;
13
15 if (r2 >= 0)
17 }
18
19 exists (1:r2=2 /\ 1:r3=0)
Quick Quiz 15.19: Wouldn’t control dependencies be more robust if they were mandated by
language standards???
15.2.6 Cache Coherence

On cache-coherent platforms, all CPUs agree on the order of loads and stores to a
given variable. Fortunately, when READ_ONCE() and WRITE_ONCE() are used, almost
all platforms are cache-coherent, as indicated by the “SV” column of the cheat sheet
shown in Table 15.3. Unfortunately, this property is so popular that it has been named
multiple times, with “single-variable SC”,8 “single-copy atomic” [SF95], and just plain
“coherence” [AMP+ 11] having seen use. Rather than further compound the confusion
by inventing yet another term for this concept, this book uses “cache coherence” and
“coherence” interchangeably.
Listing 15.15 (C-CCIRIW+o+o+o-o+o-o.litmus) shows a litmus test that tests for
cache coherence, where “IRIW” stands for “independent reads of independent writes”.
Because this litmus test uses only one variable, P2() and P3() must agree on the order
of P0()’s and P1()’s stores. In other words, if P2() believes that P0()’s store came
first, then P3() had better not believe that P1()’s store came first. And in fact the
exists clause on line 33 will trigger if this situation arises.
Quick Quiz 15.20: But in Listing 15.15, wouldn’t be just as bad if P2()’s r1 and r2
obtained the values 2 and 1, respectively, while P3()’s r3 and r4 obtained the values 1 and 2,
respectively?
It is tempting to speculate that different-sized overlapping loads and stores to a single

region of memory (as might be set up using the C-language union keyword) would
provide similar ordering guarantees. However, Flur et al. [FSP+ 17] discovered some
surprisingly simple litmus tests that demonstrate that such guarantees can be violated on
real hardware. It is therefore necessary to restrict code to non-overlapping same-sized
aligned accesses to a given variable, at least if portability is a consideration.9
8
Recall that SC stands for sequentially consistent.
There is reason to believe that using atomic RMW operations (for example, xchg())
9
for all the stores will provide sequentially consistent ordering, but this has not yet been proven
either way.
v2023.06.11a
Listing 15.15: Cache-Coherent IRIW Litmus Test

1 C C-CCIRIW+o+o+o-o+o-o
2
3 {}
4
5 P0(int *x)
6 {
8 }
9
10 P1(int *x)
11 {
13 }
14
15 P2(int *x)
16 {
17 int r1;
18 int r2;
19
22 }
23
24 P3(int *x)
25 {
26 int r3;
27 int r4;
28
31 }
32
33 exists(2:r1=1 /\ 2:r2=2 /\ 3:r3=2 /\ 3:r4=1)
Memory Memory
Figure 15.9: Global System Bus And Multi-Copy Atomicity
Adding more variables and threads increases the scope for reordering and other
counter-intuitive behavior, as discussed in the next section.
15.2.7 Multicopy Atomicity

Threads running on a fully multicopy atomic [SF95] platform are guaranteed to agree
on the order of stores, even to different variables. A useful mental model of such a
system is the single-bus architecture shown in Figure 15.9. If each store resulted in a
message on the bus, and if the bus could accommodate only one store at a time, then any
pair of CPUs would agree on the order of all stores that they observed. Unfortunately,
building a computer system as shown in the figure, without store buffers or even caches,
would result in glacially slow computation. Most CPU vendors interested in providing
multicopy atomicity therefore instead provide the slightly weaker other-multicopy
atomicity [ARM17, Section B2.3], which excludes the CPU doing a given store from the
v2023.06.11a
Listing 15.16: WRC Litmus Test With Dependencies (No Ordering)

1 C C-WRC+o+o-data-o+o-rmb-o
2
3 {}
4
5 P0(int *x)
6 {
8 }
9
10 P1(int *x, int* y)
11 {
12 int r1;
13
15 WRITE_ONCE(*y, r1);
16 }
17
19 {
20 int r2;
21 int r3;
22
23 r2 = READ_ONCE(*y);
24 smp_rmb();
26 }
27
28 exists (1:r1=1 /\ 2:r2=1 /\ 2:r3=0)
requirement that all CPUs agree on the order of all stores.10 This means that if only a
subset of CPUs are doing stores, the other CPUs will agree on the order of stores, hence
the “other” in “other-multicopy atomicity”. Unlike multicopy-atomic platforms, within
other-multicopy-atomic platforms, the CPU doing the store is permitted to observe its
store early, which allows its later loads to obtain the newly stored value directly from
the store buffer, which improves performance.
Quick Quiz 15.21: Can you give a specific example showing different behavior for multicopy
atomic on the one hand and other-multicopy atomic on the other?
Perhaps there will come a day when all platforms provide some flavor of multi-copy
atomicity, but in the meantime, non-multicopy-atomic platforms do exist, and so software
must deal with them.
Listing 15.16 (C-WRC+o+o-data-o+o-rmb-o.litmus) demonstrates multicopy
atomicity, that is, on a multicopy-atomic platform, the exists clause on line 28 cannot
trigger. In contrast, on a non-multicopy-atomic platform this exists clause can trigger,
despite P1()’s accesses being ordered by a data dependency and P2()’s accesses being
ordered by an smp_rmb(). Recall that the definition of multicopy atomicity requires
that all threads agree on the order of stores, which can be thought of as all stores reaching
all threads at the same time. Therefore, a non-multicopy-atomic platform can have a
store reach different threads at different times. In particular, P0()’s store might reach
P1() long before it reaches P2(), which raises the possibility that P1()’s store might
reach P2() before P0()’s store does.
This leads to the question of why a real system constrained by the usual laws of
physics would ever trigger the exists clause of Listing 15.16. The cartoonish diagram
of a such a real system is shown in Figure 15.10. CPU 0 and CPU 1 share a store buffer,
10 As of early 2021, Armv8 and x86 provide other-multicopy atomicity, IBM mainframe
provides full multicopy atomicity, and PPC provides no multicopy atomicity at all. More
detail is shown in Table 15.5 on page 550.
v2023.06.11a
Store CPU 1
CPU 0 Buffer Store CPU 3
CPU 2 Buffer
Cache Cache
Memory Memory
Figure 15.10: Shared Store Buffers And Multi-Copy Atomicity
Table 15.4: Memory Ordering: WRC Sequence of Events
P0() P0() & P1() P1() P2()

Instruction Store Buffer Cache Instruction Instruction Store Buffer Cache
1 (Initial state) y==0 (Initial state) (Initial state) x==0
2 x = 1; x==1 y==0 x==0
3 (Read-Invalidate x) x==1 y==0 r1 = x (1) x==0
4 x==1 y==1 y==0 y = r1 r2 = y x==0
5 x==1 y==1 (Finish store) (Read y) x==0
6 (Respond y) x==1 y==1 (r2==1) x==0 y==1
7 x==1 y==1 smp_rmb() x==0 y==1
8 x==1 y==1 r3 = x (0) x==0 y==1
9 x==1 x==0 y==1 (Respond x) y==1
10 (Finish store) x==1 y==1 y==1
as do CPUs 2 and 3. This means that CPU 1 can load a value out of the store buffer,
thus potentially immediately seeing a value stored by CPU 0. In contrast, CPUs 2 and 3
will have to wait for the corresponding cache line to carry this new value to them.
Quick Quiz 15.22: Then who would even think of designing a system with shared store
buffers???
Table 15.4 shows one sequence of events that can result in the exists clause in
Listing 15.16 triggering. This sequence of events will depend critically on P0() and
P1() sharing both cache and a store buffer in the manner shown in Figure 15.10.
Quick Quiz 15.23: But just how is it fair that P0() and P1() must share a store buffer and a
cache, but P2() gets one each of its very own???
Row 1 shows the initial state, with the initial value of y in P0()’s and P1()’s shared
cache, and the initial value of x in P2()’s cache.
Row 2 shows the immediate effect of P0() executing its store on line 7. Because the
cacheline containing x is not in P0()’s and P1()’s shared cache, the new value (1) is
stored in the shared store buffer.
Row 3 shows two transitions. First, P0() issues a read-invalidate operation to fetch
the cacheline containing x so that it can flush the new value for x out of the shared store
buffer. Second, P1() loads from x (line 14), an operation that completes immediately
because the new value of x is immediately available from the shared store buffer.
v2023.06.11a
Row 4 also shows two transitions. First, it shows the immediate effect of P1()
executing its store to y (line 15), placing the new value into the shared store buffer.
Second, it shows the start of P2()’s load from y (line 23).
Row 5 continues the tradition of showing two transitions. First, it shows P1()
complete its store to y, flushing from the shared store buffer to the cache. Second, it
shows P2() request the cacheline containing y.
Row 6 shows P2() receive the cacheline containing y, allowing it to finish its load
into r2, which takes on the value 1.
Row 7 shows P2() execute its smp_rmb() (line 24), thus keeping its two loads
ordered.
Row 8 shows P2() execute its load from x, which immediately returns with the value
zero from P2()’s cache.
Row 9 shows P2() finally responding to P0()’s request for the cacheline containing
x, which was made way back up on row 3.
Finally, row 10 shows P0() finish its store, flushing its value of x from the shared
store buffer to the shared cache.
Note well that the exists clause on line 28 has triggered. The values of r1 and r2
are both the value one, and the final value of r3 the value zero. This strange result
occurred because P0()’s new value of x was communicated to P1() long before it was
communicated to P2().
Quick Quiz 15.24: Referring to Table 15.4, why on earth would P0()’s store take so long to
complete when P1()’s store complete so quickly? In other words, does the exists clause on
line 28 of Listing 15.16 really trigger on real systems?
This counter-intuitive result happens because although dependencies do provide

ordering, they provide it only within the confines of their own thread. This three-
thread example requires stronger ordering, which is the subject of Sections 15.2.7.1
through 15.2.7.4.
15.2.7.1 Cumulativity
The three-thread example shown in Listing 15.16 requires cumulative ordering, or
cumulativity. A cumulative memory-ordering operation orders not just any given access
preceding it, but also earlier accesses by any thread to that same variable.
Dependencies do not provide cumulativity, which is why the “C” column is blank for
the READ_ONCE() row of Table 15.3 on page 492. However, as indicated by the “C” in
their “C” column, release operations do provide cumulativity. Therefore, Listing 15.17
(C-WRC+o+o-r+a-o.litmus) substitutes a release operation for Listing 15.16’s data
dependency. Because the release operation is cumulative, its ordering applies not only
to Listing 15.17’s load from x by P1() on line 14, but also to the store to x by P0()
on line 7—but only if that load returns the value stored, which matches the 1:r1=1 in
the exists clause on line 27. This means that P2()’s load-acquire suffices to force
the load from x on line 24 to happen after the store on line 7, so the value returned is
one, which does not match 2:r3=0, which in turn prevents the exists clause from
triggering.
These ordering constraints are depicted graphically in Figure 15.11. Note also that
cumulativity is not limited to a single step back in time. If there was another load from
x or store to x from any thread that came before the store on line 7, that prior load or
store would also be ordered before the load on line 24, though only if both r1 and r2
both end up containing the value 1.
v2023.06.11a
Listing 15.17: WRC Litmus Test With Release

1 C C-WRC+o+o-r+a-o
2
3 {}
4
5 P0(int *x)
6 {
8 }
9
11 {
12 int r1;
13
15 smp_store_release(y, r1);
16 }
17
19 {
20 int r2;
21 int r3;
22
23 r2 = smp_load_acquire(y);
25 }
26
27 exists (1:r1=1 /\ 2:r2=1 /\ 2:r3=0)
CPU 0
Store x=1 ... cumulativity guarantees CPU 0's store before CPU 1's store
CPU 1
... and given this link ... Load r1=x .... memory barriers guarantee this order ...
Release store
y=r1
CPU 2
Acquire load
Given this link ...
r2=y
Memory
Barrier
Load r3=x
Figure 15.11: Cumulativity
v2023.06.11a
Listing 15.18: W+RWC Litmus Test With Release (No Ordering)

1 C C-W+RWC+o-r+a-o+o-mb-o
2
3 {}
4
5 P0(int *x, int *y)
6 {
8 smp_store_release(y, 1);
9 }
10
11 P1(int *y, int *z)
12 {
13 int r1;
14 int r2;
15
17 r2 = READ_ONCE(*z);
18 }
19
20 P2(int *z, int *x)
21 {
22 int r3;
23
24 WRITE_ONCE(*z, 1);
25 smp_mb();
27 }
28
29 exists(1:r1=1 /\ 1:r2=0 /\ 2:r3=0)
CPU 0 WRITE_ONCE(z, 1);

1
CPU 1
0
fr z=
CPU 2 z=
CPU 3 r1 = READ_ONCE(z) == 0;
Time
Figure 15.12: Load-to-Store is Counter-Temporal
In short, use of cumulative ordering operations can suppress non-multicopy-atomic

behaviors in some situations. Cumulativity nevertheless has limits, which are examined
in the next section.
15.2.7.2 Propagation
Listing 15.18 (C-W+RWC+o-r+a-o+o-mb-o.litmus) shows the limitations of cumula-
tivity and store-release, even with a full memory barrier. The problem is that although
the smp_store_release() on line 8 has cumulativity, and although that cumulativity
does order P2()’s load on line 26, the smp_store_release()’s ordering cannot
propagate through the combination of P1()’s load (line 17) and P2()’s store (line 24).
This means that the exists clause on line 29 really can trigger.
Quick Quiz 15.25: But it is not necessary to worry about propagation unless there are at least
three threads in the litmus test, right?
This situation might seem completely counter-intuitive, but keep in mind that the
speed of light is finite and computers are of non-zero size. It therefore takes time for
the effect of the P2()’s store to z to propagate to P1(), which in turn means that it is
possible that P1()’s read from z happens much later in time, but nevertheless still sees
v2023.06.11a
Listing 15.19: W+WRC Litmus Test With More Barriers

1 C C-W+RWC+o-mb-o+a-o+o-mb-o
2
3 {}
4
6 {
8 smp_mb();
10 }
11
13 {
14 int r1;
15 int r2;
16
19 }
20
22 {
23 int r3;
24
26 smp_mb();
28 }
29
30 exists(1:r1=1 /\ 1:r2=0 /\ 2:r3=0)
the old value of zero. This situation is depicted in Figure 15.12: Just because a load
sees the old value does not mean that this load executed at an earlier time than did the
store of the new value.
Note that Listing 15.18 also shows the limitations of memory-barrier pairing, given
that there are not two but three processes. These more complex litmus tests can instead
be said to have cycles, where memory-barrier pairing is the special case of a two-thread
cycle. The cycle in Listing 15.18 goes through P0() (lines 7 and 8), P1() (lines 16
and 17), P2() (lines 24, 25, and 26), and back to P0() (line 7). The exists clause
delineates this cycle: The 1:r1=1 indicates that the smp_load_acquire() on line 16
returned the value stored by the smp_store_release() on line 8, the 1:r2=0 indicates
that the WRITE_ONCE() on line 24 came too late to affect the value returned by the
READ_ONCE() on line 17, and finally the 2:r3=0 indicates that the WRITE_ONCE() on
line 7 came too late to affect the value returned by the READ_ONCE() on line 26. In
this case, the fact that the exists clause can trigger means that the cycle is said to be
allowed. In contrast, in cases where the exists clause cannot trigger, the cycle is said
to be prohibited.
But what if we need to prohibit the cycle corresponding to the exists clause on line 29
of Listing 15.18? One solution is to replace P0()’s smp_store_release() with an
smp_mb(), which Table 15.3 shows to have not only cumulativity, but also propagation.
The result is shown in Listing 15.19 (C-W+RWC+o-mb-o+a-o+o-mb-o.litmus).
Quick Quiz 15.26: But given that smp_mb() has the propagation property, why doesn’t the
smp_mb() on line 25 of Listing 15.18 prevent the exists clause from triggering?
For completeness, Figure 15.13 shows that the “winning” store among a group of
stores to the same variable is not necessarily the store that started last. This should
not come as a surprise to anyone who carefully examined Figure 15.7 on page 497.
v2023.06.11a
CPU 0 WRITE_ONCE(x, 1); 1

=
CPU 1 X
0 2
= co =
CPU 2 X X
CPU 3 WRITE_ONCE(x, 2);
Time
Figure 15.13: Store-to-Store is Counter-Temporal
Listing 15.20: 2+2W Litmus Test With Write Barriers

1 C C-2+2W+o-wmb-o+o-wmb-o
2
3 {}
4
6 {
8 smp_wmb();
10 }
11
12 P1(int *x0, int *x1)
13 {
15 smp_wmb();
17 }
18
19 exists (x0=1 /\ x1=1)
One way to rationalize the counter-temporal properties of both load-to-store and store-
to-store ordering is to clearly distinguish between the temporal order in which the
store instructions executed on the one hand, and the order in which the corresponding
cacheline visited the CPUs that executed those instructions on the other. It is the
cacheline-visitation order that defines the externally visible ordering of the actual
stores. This cacheline-visitation order is not directly visible to the code executing the
store instructions, which results in the counter-intuitive counter-temporal nature of
load-to-store and store-to-store ordering.11
Quick Quiz 15.27: But for litmus tests having only ordered stores, as shown in Listing 15.20
(C-2+2W+o-wmb-o+o-wmb-o.litmus), research shows that the cycle is prohibited, even in
weakly ordered systems such as Arm and Power [SSA+ 11]. Given that, are store-to-store really
always counter-temporal???
But sometimes time really is on our side. Read on!
15.2.7.3 Happens-Before
As shown in Figure 15.14, on platforms without user-visible speculation, if a load
returns the value from a particular store, then, courtesy of the finite speed of light
and the non-zero size of modern computing systems, the store absolutely has to have
executed at an earlier time than did the load. This means that carefully constructed
programs can rely on the passage of time itself as a memory-ordering operation.
11 In some hardware-multithreaded systems, the store would become visible to other
CPUs in that same core as soon as the store reached the shared store buffer. As a result, such
systems are non-multicopy atomic.
v2023.06.11a
CPU 0 WRITE_ONCE(x, 1);

1
CPU 1 =
X
CPU 2 0 rf
=
CPU 3
X r1 = READ_ONCE(x);
Time
Figure 15.14: Store-to-Load is Temporal
Listing 15.21: LB Litmus Test With One Acquire

1 C C-LB+a-o+o-data-o+o-data-o
2
3 {}
4
6 {
7 int r2;
8
11 }
12
13 P1(int *x1, int *x2)
14 {
15 int r2;
16
19 }
20
21 P2(int *x2, int *x0)
22 {
23 int r2;
24
27 }
28
29 exists (0:r2=2 /\ 1:r2=2 /\ 2:r2=2)
Of course, just the passage of time by itself is not enough, as was seen in Listing 15.6 on
page 499, which has nothing but store-to-load links and, because it provides absolutely
no ordering, still can trigger its exists clause. However, as long as each thread
provides even the weakest possible ordering, exists clause would not be able to trigger.
For example, Listing 15.21 (C-LB+a-o+o-data-o+o-data-o.litmus) shows P0()
ordered with an smp_load_acquire() and both P1() and P2() ordered with data
dependencies. These orderings, which are close to the top of Table 15.3, suffice to
prevent the exists clause from triggering.
Quick Quiz 15.28: Can you construct a litmus test like that in Listing 15.21 that uses only
dependencies?
An important use of time for ordering memory accesses is covered in the next section.
15.2.7.4 Release-Acquire Chains
A minimal release-acquire chain was shown in Listing 15.7 on page 500, but these chains
can be much longer, as shown in Listing 15.22 (C-LB+a-r+a-r+a-r+a-r.litmus).
The longer the release-acquire chain, the more ordering is gained from the passage
v2023.06.11a
Listing 15.22: Long LB Release-Acquire Chain

1 C C-LB+a-r+a-r+a-r+a-r
2
3 {}
4
6 {
7 int r2;
8
11 }
12
13 P1(int *x1, int *x2)
14 {
15 int r2;
16
19 }
20
21 P2(int *x2, int *x3)
22 {
23 int r2;
24
27 }
28
29 P3(int *x3, int *x0)
30 {
31 int r2;
32
35 }
36
37 exists (0:r2=2 /\ 1:r2=2 /\ 2:r2=2 /\ 3:r2=2)
of time, so that no matter how many threads are involved, the corresponding exists
clause cannot trigger.
Although release-acquire chains are inherently store-to-load creatures, it turns out that
they can tolerate one load-to-store step, despite such steps being counter-temporal, as
shown in Figure 15.12 on page 511. For example, Listing 15.23 (C-ISA2+o-r+a-r+a-
r+a-o.litmus) shows a three-step release-acquire chain, but where P3()’s final access
is a READ_ONCE() from x0, which is accessed via WRITE_ONCE() by P0(), forming a
non-temporal load-to-store link between these two processes. However, because P0()’s
smp_store_release() (line 8) is cumulative, if P3()’s READ_ONCE() returns zero,
this cumulativity will force the READ_ONCE() to be ordered before P0()’s smp_store_
release(). In addition, the release-acquire chain (lines 8, 15, 16, 23, 24, and 32) forces
P3()’s READ_ONCE() to be ordered after P0()’s smp_store_release(). Because
P3()’s READ_ONCE() cannot be both before and after P0()’s smp_store_release(),
either or both of two things must be true:
1. P3()’s READ_ONCE() came after P0()’s WRITE_ONCE(), so that the READ_

ONCE() returned the value two, so that the exists clause’s 3:r2=0 is false.
2. The release-acquire chain did not form, that is, one or more of the exists clause’s
1:r2=2, 2:r2=2, or 3:r1=2 is false.
v2023.06.11a
Listing 15.23: Long ISA2 Release-Acquire Chain

1 C C-ISA2+o-r+a-r+a-r+a-o
2
3 {}
4
6 {
9 }
10
11 P1(int *x1, int *x2)
12 {
13 int r2;
14
17 }
18
19 P2(int *x2, int *x3)
20 {
21 int r2;
22
25 }
26
27 P3(int *x3, int *x0)
28 {
29 int r1;
30 int r2;
31
34 }
35
36 exists (1:r2=2 /\ 2:r2=2 /\ 3:r1=2 /\ 3:r2=0)
Either way, the exists clause cannot trigger, despite this litmus test containing a
notorious load-to-store link between P3() and P0(). But never forget that release-
acquire chains can tolerate only one load-to-store link, as was seen in Listing 15.18.
Release-acquire chains can also tolerate a single store-to-store step, as shown in
Listing 15.24 (C-Z6.2+o-r+a-r+a-r+a-o.litmus). As with the previous example,
smp_store_release()’s cumulativity combined with the temporal nature of the
release-acquire chain prevents the exists clause on line 35 from triggering.
Quick Quiz 15.29: Suppose we have a short release-acquire chain along with one load-to-store
link and one store-to-store link, like that shown in Listing 15.25. Given that there is only one of
each type of non-store-to-load link, the exists cannot trigger, right?
But beware: Adding a second store-to-store link allows the correspondingly updated
exists clause to trigger. To see this, review Listings 15.26 and 15.27, which have
identical P0() and P1() processes. The only code difference is that Listing 15.27 has
an additional P2() that does an smp_store_release() to the x2 variable that P0()
releases and P1() acquires. The exists clause is also adjusted to exclude executions
in which P2()’s smp_store_release() precedes that of P0().
Running the litmus test in Listing 15.27 shows that the addition of P2() can totally
destroy the ordering from the release-acquire chain. Therefore, when constructing
release-acquire chains, please take care to construct them properly.
Quick Quiz 15.30: There are store-to-load links, load-to-store links, and store-to-store links.
But what about load-to-load links?
v2023.06.11a
Listing 15.24: Long Z6.2 Release-Acquire Chain

1 C C-Z6.2+o-r+a-r+a-r+a-o
2
3 {}
4
6 {
9 }
10
11 P1(int *x1, int *x2)
12 {
13 int r2;
14
17 }
18
19 P2(int *x2, int *x3)
20 {
21 int r2;
22
25 }
26
27 P3(int *x3, int *x0)
28 {
29 int r2;
30
33 }
34
35 exists (1:r2=2 /\ 2:r2=2 /\ 3:r2=2 /\ x0=2)
Listing 15.25: Z6.2 Release-Acquire Chain (Ordering?)

1 C C-Z6.2+o-r+a-o+o-mb-o
2
3 {}
4
6 {
8 smp_store_release(y, 1);
9 }
10
12 {
13 int r1;
14
17 }
18
20 {
21 int r2;
22
24 smp_mb();
26 }
27
28 exists(1:r1=1 /\ 2:r2=0 /\ z=2)
v2023.06.11a
Listing 15.26: A Release-Acquire Chain Ordering Multiple Accesses

1 C C-MP+o-r+a-o
2
3 {}
4
5 P0(int* x0, int* x1, int* x2) {
6 int r1;
7
11 }
12
13 P1(int* x0, int* x1, int* x2) {
14 int r2;
15 int r3;
16
20 }
21
22 exists (1:r2=2 /\ (1:r3=0 \/ 0:r1=2))
Listing 15.27: A Release-Acquire Chain With Added Store (Ordering?)

1 C C-MPO+o-r+a-o+o
2
3 {}
4
5 P0(int* x0, int* x1, int* x2) {
6 int r1;
7
11 }
12
13 P1(int* x0, int* x1, int* x2) {
14 int r2;
15 int r3;
16
20 }
21
22 P2(int* x2) {
24 }
25
26 exists (1:r2=3 /\ x2=3 /\ (1:r3=0 \/ 0:r1=2))
v2023.06.11a
In short, properly constructed release-acquire chains form a peaceful island of intuitive

bliss surrounded by a strongly counter-intuitive sea of more complex memory-ordering
constraints.
15.2.8 A Counter-Intuitive Case Study

This section will revisit Listing E.12 on page 845, which was presented in the answer to
Quick Quiz 15.25. This litmus test has only two threads, with the stores in P0() being
ordered by smp_wmb() and the accesses in P1() being ordered by smp_mb(). Despite
this litmus test’s small size and heavy ordering, the counter-intuitive outcome shown in
the exists clause is in fact allowed.
One way to look at this was presented in the answer to Quick Quiz 15.25, namely that
the link from P0() to P1() is a store-to-store link, and that back from P1() to P0() is a
store-to-store link. Both links are counter-temporal, thus requiring full memory barriers
in both processes. Revisiting Figures 15.13 and 15.14 shows that these counter-temporal
links give the hardware considerable latitude.
But that raises the question of exactly how hardware would go about using this
latitude to satisfy the exists clause in Listing E.12. There is no known “toy” hardware
implementation that can do this, so let us instead study the sequence of steps that the
PowerPC architecture goes through to make this happen.
The first step in this study is to translate Listing E.12 to a PowerPC assembly language
litmus test (Section 12.2.1 on page 403):
1 PPC R+lwsync+sync
2 {
3 0:r1=1; 0:r2=x; 0:r4=y;
4 1:r1=2; 1:r2=y; 1:r4=x;
5 }
6 P0 | P1 ;
7 stw r1,0(r2) | stw r1,0(r2) ;
8 lwsync | sync ;
9 stw r1,0(r4) | lwz r3,0(r4) ;
10 exists (y=2 /\ 1:r3=0)
The first line identifies the type of test (PPC) and gives the test’s name. Lines 3
and 4 initialize P0()’s and P1()’s registers, respectively. Lines 6–9 show the PowerPC
assembly statements corresponding to the C code from Listing E.12, with the first
column being the code for P0() and the second column being the code for P1().
Line 7 shows the initial WRITE_ONCE() calls in both columns; the columns of line 8
show the smp_wmb() and smp_mb() for P0() and P1(), respectively; the columns
of line 9 shows P0()’s WRITE_ONCE() and P1()’s READ_ONCE(), respectively; and
finally line 10 shows the exists clause.
In order for this exists clause to be satisfied, P0()’s stw to y must precede that of
P1(), but P1()’s later lwz from x must precede P0()’s stw to x. Seeing how this can
happen requires a rough understanding of the following PowerPC terminology.
Instruction commit:
This can be thought of as the execution of that instruction as opposed to the
memory-system consequences of having executed that instruction.
Write reaching coherence point:

This can be thought of as the value written being deposited into the corresponding
cache line.
v2023.06.11a
Figure 15.15: PPCMEM Initial R State
Partial coherence commit:

This can be thought of as the system having worked out the order in which a pair of
values written will be deposited into the corresponding cache line, but potentially
well before that cache line arrives. Some might argue that the data in Figure 15.7
suggests that real PowerPC hardware does in fact use partial coherence commits to
handle concurrent stores by multiple hardware threads within a single core.
Write propagate to thread:
This occurs when a second hardware thread becomes aware of the first hardware
thread’s write. The time at which a write propagates to a given thread might not
have any relation to cache-line movement. For example, if a pair of threads share
a store buffer, they might see each others’ writes long before the cache line gets
involved. On the other hand, if a pair of hardware threads are widely separated, the
first thread’s write’s value might have been deposited into the corresponding cache
line long before the second thread learns of that write.
Barrier propagate to thread:
Hardware threads make each other aware of memory-barrier instructions as needed
by propagating them to each other.
Acknowledge sync:
The PowerPC sync instruction implements the Linux kernel’s smp_mb() full
barrier. And one reason that the sync instruction provides such strong ordering is
that each sync is not only propagated to other hardware threads, but these other
threads must also acknowledge each sync. This two-way communication allows
the hardware threads to cooperate to produce the required strong global ordering.
We are now ready to step through the PowerPC sequence of events that satisfies the
above exists clause.
To best understand this, please follow along at https://www.cl.cam.ac.uk/
~pes20/ppcmem/index.html, carefully copying the above assembly-language litmus
test into the pane. The result should look as shown in Figure 15.15, give or take space
characters. Click on the “Interactive” button in the lower left, which, after a short delay,
should produce a display as shown in Figure 15.16. If the “Interactive” button refuses
to do anything, this usually means that there is a syntax error, for example, a spurious
newline character might have been introduced during the copy-paste operation.
This display has one clickable link in each section displaying thread state, and
as the “Commit” in each link suggests, these links commit each thread’s first stw
v2023.06.11a
Figure 15.16: PPCMEM First R Step
instruction. If you prefer, you can instead click on the corresponding links listed under
“Enabled transitions” near the bottom of the screen. Note well that some of the later
memory-system transitions will appear in the upper “Storage subsystem state” section
of this display.
The following sequence of clicks demonstrates how the exists clause can be
satisfied:
1. Commit P0()’s first stw instruction (to x).
2. Commit P1()’s stw instruction.
3. Commit P0()’s lwsync instruction.
4. Commit P0()’s second stw instruction (to y).
5. Commit P1()’s sync instruction.
6. At this point, there should be no clickable links in either of the two sections
displaying thread state, but there should be quite a few of them up in the “Storage
subsystem state”. The following steps tell you which of them to click on.
7. Partial coherence commit: c:W y=1 ->d:W y=2. This commits the sys-
tem to processing P0()’s store to y before P1()’s store even though neither store
has reached either the coherence point or any other thread. One might imagine
partial coherence commits happening within a store buffer that is shared by multiple
hardware threads that are writing to the same variable.
8. Write propagate to thread: d:W y=2 to Thread 0. This is necessary to

allow P1()’s sync instruction to propagate to P0().
v2023.06.11a
Figure 15.17: PPCMEM Final R State
9. Barrier propagate to thread: e:Sync to Thread 0.

10. Write reaching coherence point: a:W x=1.
11. Write reaching coherence point: c:W y=1.
12. Write reaching coherence point: d:W y=2. These three operations were
required in order to allow P0() to acknowledge P1()’s sync instruction.
13. Acknowledge sync: Sync e:Sync.
14. Back down in thread P1()’s state, click on Read i:W x=0, which loads the value
zero, thus satisfying the exists clause. All that remains is cleanup, which can be
carried out in any order.
15. Commit P1()’s lwz instruction.
16. Write propagate to thread: a:W x=1 to Thread 1.
17. Barrier propagate to thread: b:Lwsync to Thread 1.
At this point, you should see something like Figure 15.17. Note that the satisified
exists clause is shown in blue near the bottom, confirming that this counter-intuitive
really can happen. If you wish, you can click on “Undo” to explore other options or
click on “Reset” to start over. It can be very helpful to carry out these steps in different
orders to better understand how a non-multicopy-atomic architecture operates.
Quick Quiz 15.31: What happens if that lwsync instruction is instead a sync instruction?
Although a full understanding of how this counter-intuitive outcome happens would

require hardware details that are beyond the scope of this book, this exercise should
v2023.06.11a
15.3. COMPILE-TIME CONSTERNATION 523
provide some helpful intuitions. Or perhaps more accurately, destroy some counter-
productive intuitions.
15.3 Compile-Time Consternation

Science increases our power in proportion as it
lowers our pride.
Claude Bernard
Most languages, including C, were developed on uniprocessor systems by people

with little or no parallel-programming experience. As a result, unless explicitly told
otherwise, these languages assume that the current CPU is the only thing that is reading
or writing memory. This in turn means that these languages’ compilers’ optimizers are
ready, willing, and oh so able to make dramatic changes to the order, number, and sizes
of memory references that your program executes. In fact, the reordering carried out by
hardware can seem quite tame by comparison.
This section will help you tame your compiler, thus avoiding a great deal of compile-
time consternation. Section 15.3.1 describes how to keep the compiler from destructively
optimizing your code’s memory references, Section 15.3.2 describes how to protect
address and data dependencies, and finally, Section 15.3.3 describes how to protect
those delicate control dependencies.
15.3.1 Memory-Reference Restrictions

As noted in Section 4.3.4, unless told otherwise, compilers assume that nothing else is
affecting the variables that the code is accessing. Furthermore, this assumption is not
simply some design error, but is instead enshrined in various standards.12 It is worth
summarizing this material in preparation for the following sections.
Plain accesses, as in plain-access C-language assignment statements such as “r1 = a”
or “b = 1” are subject to the shared-variable shenanigans described in Section 4.3.4.1.
Ways of avoiding these shenanigans are described in Sections 4.3.4.2–4.3.4.4 starting
on page 68:
1. Plain accesses can tear, for example, the compiler could choose to access an
eight-byte pointer one byte at a time. Tearing of aligned machine-sized accesses
can be prevented by using READ_ONCE() and WRITE_ONCE().
2. Plain loads can fuse, for example, if the results of an earlier load from that same
object are still in a machine register, the compiler might opt to reuse the value
in that register instead of reloading from memory. Load fusing can be prevented
by using READ_ONCE() or by enforcing ordering between the two loads using
barrier(), smp_rmb(), and other means shown in Table 15.3.
3. Plain stores can fuse, so that a store can be omitted entirely if there is a later store
to that same variable. Store fusing can be prevented by using WRITE_ONCE() or
by enforcing ordering between the two stores using barrier(), smp_wmb(), and
other means shown in Table 15.3.
12 Or perhaps it is a standardized design error.
v2023.06.11a
4. Plain accesses can be reordered in surprising ways by modern optimizing compilers.

This reordering can be prevented by enforcing ordering as called out above.
5. Plain loads can be invented, for example, register pressure might cause the compiler
to discard a previously loaded value from its register, and then reload it later on.
Invented loads can be prevented by using READ_ONCE() or by enforcing ordering
as called out above between the load and a later use of its value using barrier().
6. Stores can be invented before a plain store, for example, by using the stored-to
location as temporary storage. This can be prevented by use of WRITE_ONCE().
7. Stores can be transformed into a load-check-store sequence, which can defeat

control dependencies. This can be prevented by use of smp_load_acquire().
Quick Quiz 15.32: Why not place a barrier() call immediately before a plain store to
prevent the compiler from inventing stores?
Please note that all of these shared-memory shenanigans can instead be avoided by
avoiding data races on plain accesses, as described in Section 4.3.4.4. After all, if there
are no data races, then each and every one of the compiler optimizations mentioned
above is perfectly safe. But for code containing data races, this list is subject to change
without notice as compiler optimizations continue becoming increasingly aggressive.
In short, use of READ_ONCE(), WRITE_ONCE(), barrier(), volatile, and other
primitives called out in Table 15.3 on page 492 are valuable tools in preventing the
compiler from optimizing your parallel algorithm out of existence. Compilers are starting
to provide other mechanisms for avoiding load and store tearing, for example, memory_
order_relaxed atomic loads and stores, however, work is still needed [Cor16b]. In
addition, compiler issues aside, volatile is still needed to avoid fusing and invention
of accesses, including C11 atomic accesses.
Please note that, it is possible to overdo use of READ_ONCE() and WRITE_ONCE().
For example, if you have prevented a given variable from changing (perhaps by holding
the lock guarding all updates to that variable), there is no point in using READ_ONCE().
Similarly, if you have prevented any other CPUs or threads from reading a given variable
(perhaps because you are initializing that variable before any other CPU or thread has
access to it), there is no point in using WRITE_ONCE(). However, in my experience,
developers need to use things like READ_ONCE() and WRITE_ONCE() more often than
they think that they do, and the overhead of unnecessary uses is quite low. In contrast,
the penalty for failing to use them when needed can be quite high.
15.3.2 Address- and Data-Dependency Difficulties

The low overheads of the address and data dependencies discussed in Sections 15.2.3
and 15.2.4, respectively, makes their use extremely attractive. Unfortunately, compilers
do not understand either address or data dependencies, although there are efforts
underway to teach them, or at the very least, standardize the process of teaching
them [MWB+ 17, MRP+ 17]. In the meantime, it is necessary to be very careful in order
to prevent your compiler from breaking your dependencies.
v2023.06.11a
Listing 15.28: Breakable Dependencies With Comparisons

1 int reserve_int;
2 int *gp;
3 int *p;
4
5 p = rcu_dereference(gp);
6 if (p == &reserve_int)
7 handle_reserve(p);
8 do_something_with(*p); /* buggy! */
Listing 15.29: Broken Dependencies With Comparisons

1 int reserve_int;
2 int *gp;
3 int *p;
4
5 p = rcu_dereference(gp);
6 if (p == &reserve_int) {
7 handle_reserve(&reserve_int);
8 do_something_with(reserve_int); /* buggy! */
9 } else {
10 do_something_with(*p); /* OK! */
11 }
15.3.2.1 Give your dependency chain a good start

The load that heads your dependency chain must use proper ordering, for example
rcu_dereference() or READ_ONCE(). Failure to follow this rule can have serious
side effects:
1. On DEC Alpha, a dependent load might not be ordered with the load heading the
dependency chain, as described in Section 15.5.1.
2. If the load heading the dependency chain is a C11 non-volatile memory_order_
relaxed load, the compiler could omit the load, for example, by using a value that
it loaded in the past.
3. If the load heading the dependency chain is a plain load, the compiler can omit
the load, again by using a value that it loaded in the past. Worse yet, it could load
twice instead of once, so that different parts of your code use different values—and
compilers really do this, especially when under register pressure.
4. The value loaded by the head of the dependency chain must be a pointer. In theory,
yes, you could load an integer, perhaps to use it as an array index. In practice, the
compiler knows too much about integers, and thus has way too many opportunities
to break your dependency chain [MWB+ 17].
15.3.2.2 Avoid arithmetic dependency breakage

Although it is just fine to do some arithmetic operations on a pointer in your dependency
chain, you need to be careful to avoid giving the compiler too much information. After
all, if the compiler learns enough to determine the exact value of the pointer, it can use
that exact value instead of the pointer itself. As soon as the compiler does that, the
dependency is broken and all ordering is lost.
1. Although it is permissible to compute offsets from a pointer, these offsets must not
result in total cancellation. For example, given a char pointer cp, cp-(uintptr_
t)cp will cancel and can allow the compiler to break your dependency chain.
v2023.06.11a
On the other hand, canceling offset values with each other is perfectly safe and
legal. For example, if a and b are equal, cp+a-b is an identity function, including
preserving the dependency.
2. Comparisons can break dependencies. Listing 15.28 shows how this can happen.
Here global pointer gp points to a dynamically allocated integer, but if memory
is low, it might instead point to the reserve_int variable. This reserve_int
case might need special handling, as shown on lines 6 and 7 of the listing. But the
compiler could reasonably transform this code into the form shown in Listing 15.29,
especially on systems where instructions with absolute addresses run faster than
instructions using addresses supplied in registers. However, there is clearly no
ordering between the pointer load on line 5 and the dereference on line 8. Please
note that this is simply an example: There are a great many other ways to break
dependency chains with comparisons.
Quick Quiz 15.33: Why can’t you simply dereference the pointer before comparing it to
&reserve_int on line 6 of Listing 15.28?
Quick Quiz 15.34: But it should be safe to compare two pointer variables, right? After all,
the compiler doesn’t know the value of either, so how can it possibly learn anything from the
comparison?
Note that a series of inequality comparisons might, when taken together, give the
compiler enough information to determine the exact value of the pointer, at which
point the dependency is broken. Furthermore, the compiler might be able to combine
information from even a single inequality comparison with other information to learn
the exact value, again breaking the dependency. Pointers to elements in arrays are
especially susceptible to this latter form of dependency breakage.
15.3.2.3 Safe comparison of dependent pointers

It turns out that there are several safe ways to compare dependent pointers:
1. Comparisons against the NULL pointer. In this case, all the compiler can learn is
that the pointer is NULL, in which case you are not allowed to dereference it anyway.
2. The dependent pointer is never dereferenced, whether before or after the comparison.
3. The dependent pointer is compared to a pointer that references objects that were
last modified a very long time ago, where the only unconditionally safe value of “a
very long time ago” is “at compile time”. The key point is that something other
than the address or data dependency guarantees ordering.
4. Comparisons between two pointers, each of which carries an appropriate depen-

dency. For example, you have a pair of pointers, each carrying a dependency, to
data structures each containing a lock, and you want to avoid deadlock by acquiring
the locks in address order.
5. The comparison is not-equal, and the compiler does not have enough other
information to deduce the value of the pointer carrying the dependency.
v2023.06.11a
Listing 15.30: Broken Dependencies With Pointer Comparisons

1 struct foo {
2 int a;
3 int b;
4 int c;
5 };
6 struct foo *gp1;
7 struct foo *gp2;
8
9 void updater(void)
10 {
11 struct foo *p;
12
13 p = malloc(sizeof(*p));
14 BUG_ON(!p);
15 p->a = 42;
16 p->b = 43;
17 p->c = 44;
18 rcu_assign_pointer(gp1, p);
19 WRITE_ONCE(p->b, 143);
20 WRITE_ONCE(p->c, 144);
21 rcu_assign_pointer(gp2, p);
22 }
23
24 void reader(void)
25 {
26 struct foo *p;
27 struct foo *q;
28 int r1, r2 = 0;
29
30 p = rcu_dereference(gp2);
31 if (p == NULL)
32 return;
33 r1 = READ_ONCE(p->b);
34 q = rcu_dereference(gp1);
35 if (p == q) {
36 r2 = READ_ONCE(p->c);
37 }
38 do_something_with(r1, r2);
39 }
Pointer comparisons can be quite tricky, and so it is well worth working through the
example shown in Listing 15.30. This example uses a simple struct foo shown on
lines 1–5 and two global pointers, gp1 and gp2, shown on lines 6 and 7, respectively.
This example uses two threads, namely updater() on lines 9–22 and reader() on
lines 24–39.
The updater() thread allocates memory on line 13, and complains bitterly on line 14
if none is available. Lines 15–17 initialize the newly allocated structure, and then
line 18 assigns the pointer to gp1. Lines 19 and 20 then update two of the structure’s
fields, and does so after line 18 has made those fields visible to readers. Please note
that unsynchronized update of reader-visible fields often constitutes a bug. Although
there are legitimate use cases doing just this, such use cases require more care than is
exercised in this example.
Finally, line 21 assigns the pointer to gp2.
The reader() thread first fetches gp2 on line 30, with lines 31 and 32 checking for
NULL and returning if so. Line 33 fetches field ->b and line 34 fetches gp1. If line 35
sees that the pointers fetched on lines 30 and 34 are equal, line 36 fetches p->c. Note
that line 36 uses pointer p fetched on line 30, not pointer q fetched on line 34.
But this difference might not matter. An equals comparison on line 35 might lead
the compiler to (incorrectly) conclude that both pointers are equivalent, when in fact
they carry different dependencies. This means that the compiler might well transform
v2023.06.11a
line 36 to instead be r2 = READ_ONCE(q->c), which might well cause the value 44 to

be loaded instead of the expected value 144.
Quick Quiz 15.35: But doesn’t the condition in line 35 supply a control dependency that
would keep line 36 ordered after line 34?
In short, great care is required to ensure that dependency chains in your source code
are still dependency chains in the compiler-generated assembly code.
15.3.3 Control-Dependency Calamities

The control dependencies described in Section 15.2.5 are attractive due to their low
overhead, but are also especially tricky because current compilers do not understand
them and can easily break them. The rules and examples in this section are intended to
help you prevent your compiler’s ignorance from breaking your code.
A load-load control dependency requires a full read memory barrier, not simply a
data dependency barrier. Consider the following bit of code:
1 q = READ_ONCE(x);
2 if (q) {
3 <data dependency barrier>
4 q = READ_ONCE(y);
5 }
This will not have the desired effect because there is no actual data dependency, but
rather a control dependency that the CPU may short-circuit by attempting to predict the
outcome in advance, so that other CPUs see the load from y as having happened before
the load from x. In such a case what’s actually required is:
1 q = READ_ONCE(x);
2 if (q) {
3 <read barrier>
4 q = READ_ONCE(y);
5 }
However, stores are not speculated. This means that ordering is provided for load-store
control dependencies, as in the following example:
1 q = READ_ONCE(x);
2 if (q)
3 WRITE_ONCE(y, 1);
Control dependencies pair normally with other types of ordering operations. That
said, please note that neither READ_ONCE() nor WRITE_ONCE() are optional! Without
the READ_ONCE(), the compiler might fuse the load from x with other loads from x.
Without the WRITE_ONCE(), the compiler might fuse the store to y with other stores
to y, or, worse yet, read the value, compare it, and only conditionally do the store. Any
of these can result in highly counter-intuitive effects on ordering.
Worse yet, if the compiler is able to prove (say) that the value of variable x is
always non-zero, it would be well within its rights to optimize the original example by
eliminating the “if” statement as follows:
1 q = READ_ONCE(x);
2 WRITE_ONCE(y, 1); /* BUG: CPU can reorder!!! */
v2023.06.11a
Quick Quiz 15.36: But there is a READ_ONCE(), so how can the compiler prove anything
about the value of q?
It is tempting to try to enforce ordering on identical stores on both branches of the

“if” statement as follows:
1 q = READ_ONCE(x);
2 if (q) {
3 barrier();
4 WRITE_ONCE(y, 1);
5 do_something();
6 } else {
7 barrier();
8 WRITE_ONCE(y, 1);
10 }
Unfortunately, current compilers will transform this as follows at high optimization

levels:
1 q = READ_ONCE(x);
2 barrier();
3 WRITE_ONCE(y, 1); /* BUG: No ordering!!! */
4 if (q) {
5 do_something();
6 } else {
8 }
Now there is no conditional between the load from x and the store to y, which
means that the CPU is within its rights to reorder them: The conditional is absolutely
required, and must be present in the assembly code even after all compiler optimizations
have been applied. Therefore, if you need ordering in this example, you need explicit
memory-ordering operations, for example, a release store:
1 q = READ_ONCE(x);
2 if (q) {
3 smp_store_release(&y, 1);
4 do_something();
5 } else {
6 smp_store_release(&y, 1);
8 }
The initial READ_ONCE() is still required to prevent the compiler from guessing the
value of x. In addition, you need to be careful what you do with the local variable q,
otherwise the compiler might be able to guess its value and again remove the needed
conditional. For example:
1 q = READ_ONCE(x);
2 if (q % MAX) {
3 WRITE_ONCE(y, 1);
4 do_something();
5 } else {
6 WRITE_ONCE(y, 2);
8 }
If MAX is defined to be 1, then the compiler knows that (q%MAX) is equal to zero,
in which case the compiler is within its rights to transform the above code into the
following:
v2023.06.11a
1 q = READ_ONCE(x);
2 WRITE_ONCE(y, 2);
Given this transformation, the CPU is not required to respect the ordering between
the load from variable x and the store to variable y. It is tempting to add a barrier()
to constrain the compiler, but this does not help. The conditional is gone, and the
barrier() won’t bring it back. Therefore, if you are relying on this ordering, you
should make sure that MAX is greater than one, perhaps as follows:
1 q = READ_ONCE(x);
2 BUILD_BUG_ON(MAX <= 1);
3 if (q % MAX) {
4 WRITE_ONCE(y, 1);
5 do_something();
6 } else {
7 WRITE_ONCE(y, 2);
9 }
Please note once again that the stores to y differ. If they were identical, as noted
earlier, the compiler could pull this store outside of the “if” statement.
You must also avoid excessive reliance on boolean short-circuit evaluation. Consider
this example:
1 q = READ_ONCE(x);
2 if (q || 1 > 0)
3 WRITE_ONCE(y, 1);
Because the first condition cannot fault and the second condition is always true, the
compiler can transform this example as following, defeating the control dependency:
1 q = READ_ONCE(x);
2 WRITE_ONCE(y, 1);
This example underscores the need to ensure that the compiler cannot out-guess your
code. Never forget that, although READ_ONCE() does force the compiler to actually
emit code for a given load, it does not force the compiler to use the value loaded.
In addition, control dependencies apply only to the then-clause and else-clause of the
if-statement in question. In particular, it does not necessarily apply to code following
the if-statement:
1 q = READ_ONCE(x);
2 if (q) {
3 WRITE_ONCE(y, 1);
4 } else {
5 WRITE_ONCE(y, 2);
6 }
7 WRITE_ONCE(z, 1); /* BUG: No ordering. */
It is tempting to argue that there in fact is ordering because the compiler cannot
reorder volatile accesses and also cannot reorder the writes to y with the condition.
Unfortunately for this line of reasoning, the compiler might compile the two writes to y
as conditional-move instructions, as in this fanciful pseudo-assembly language:
v2023.06.11a
Listing 15.31: LB Litmus Test With Control Dependency

1 C C-LB+o-cgt-o+o-cgt-o
2
3 {}
4
6 {
7 int r1;
8
10 if (r1 > 0)
12 }
13
15 {
16 int r2;
17
19 if (r2 > 0)
21 }
22
23 exists (0:r1=1 /\ 1:r2=1)
1 ld r1,x
2 cmp r1,$0
3 cmov,ne r4,$1
4 cmov,eq r4,$2
5 st r4,y
6 st $1,z
A weakly ordered CPU would have no dependency of any sort between the load
from x and the store to z. The control dependencies would extend only to the pair of
cmov instructions and the store depending on them. In short, control dependencies apply
only to the stores in the “then” and “else” of the “if” in question (including functions
invoked by those two clauses), and not necessarily to code following that “if”.
Finally, control dependencies do not provide cumulativity.13 This is demonstrated by
two related litmus tests, namely Listings 15.31 and 15.32 with the initial values of x
and y both being zero.
The exists clause in the two-thread example of Listing 15.31 (C-LB+o-cgt-o+o-
cgt-o.litmus) will never trigger. If control dependencies guaranteed cumulativity
(which they do not), then adding a thread to the example as in Listing 15.32 (C-WWC+o-
cgt-o+o-cgt-o+o.litmus) would guarantee the related exists clause never to
trigger.
But because control dependencies do not provide cumulativity, the exists clause in
the three-thread litmus test can trigger. If you need the three-thread example to provide
ordering, you will need smp_mb() between the load and store in P0(), that is, just
before or just after the “if” statements. Furthermore, the original two-thread example
is very fragile and should be avoided.
Quick Quiz 15.37: Can’t you instead add an smp_mb() to P1() in Listing 15.32?
The following list of rules summarizes the lessons of this section:
1. Compilers do not understand control dependencies, so it is your job to make sure

that the compiler cannot break your code.
13 Refer to Section 15.2.7.1 for the meaning of cumulativity.
v2023.06.11a
Listing 15.32: WWC Litmus Test With Control Dependency (Cumulativity?)

1 C C-WWC+o-cgt-o+o-cgt-o+o
2
3 {}
4
6 {
7 int r1;
8
10 if (r1 > 0)
12 }
13
15 {
16 int r2;
17
19 if (r2 > 0)
21 }
22
23 P2(int *x)
24 {
26 }
27
28 exists (0:r1=2 /\ 1:r2=1 /\ x=2)
2. Control dependencies can order prior loads against later stores. However, they do
not guarantee any other sort of ordering: Not prior loads against later loads, nor
prior stores against later anything. If you need these other forms of ordering, use
smp_rmb(), smp_wmb(), or, in the case of prior stores and later loads, smp_mb().
3. If both legs of the “if” statement begin with identical stores to the same variable,
then the control dependency will not order those stores, If ordering is needed,
precede both of them with smp_mb() or use smp_store_release(). Please
note that it is not sufficient to use barrier() at beginning of each leg of the “if”
statement because, as shown by the example above, optimizing compilers can
destroy the control dependency while respecting the letter of the barrier() law.
4. Control dependencies require at least one run-time conditional between the prior
load and the subsequent store, and this conditional must involve the prior load. If
the compiler is able to optimize the conditional away, it will have also optimized
away the ordering. Careful use of READ_ONCE() and WRITE_ONCE() can help to
preserve the needed conditional.
5. Control dependencies require that the compiler avoid reordering the dependency
into nonexistence. Careful use of READ_ONCE(), atomic_read(), or atomic64_
read() can help to preserve your control dependency.
6. Control dependencies apply only to the “then” and “else” of the “if” containing
the control dependency, including any functions that these two clauses call.
Control dependencies do not apply to code following the end of the “if” statement
containing the control dependency.
7. Control dependencies pair normally with other types of memory-ordering opera-

tions.
v2023.06.11a
15.4. HIGHER-LEVEL PRIMITIVES 533
8. Control dependencies do not provide cumulativity. If you need cumulativity, use

something that provides it, such as smp_store_release() or smp_mb().
Again, many popular languages were designed with single-threaded use in mind.
Successful multithreaded use of these languages requires you to pay special attention to
your memory references and dependencies.
15.4 Higher-Level Primitives
Method will teach you to win time.

Johann Wolfgang von Goethe
The answer to one of the quick quizzes in Section 12.3.1 demonstrated exponential
speedups due to verifying programs modeled at higher levels of abstraction. This section
will look into how higher levels of abstraction can also provide a deeper understanding
of the synchronization primitives themselves. Section 15.4.1 takes a look at memory
allocation, Section 15.4.2 examines the surprisingly varied semantics of locking, and
Section 15.4.3 digs more deeply into RCU.
15.4.1 Memory Allocation

Section 6.4.3.2 touched upon memory allocation, and this section expands upon the
relevant memory-ordering issues.
The key requirement is that any access executed on a given block of memory before
freeing that block must be ordered before any access executed after that same block is
reallocated. It would after all be a cruel and unusual memory-allocator bug if a store
preceding the free were to be reordered after another store following the reallocation!
However, it would also be cruel and unusual to require developers to use READ_ONCE()
and WRITE_ONCE() to access dynamically allocated memory. Full ordering must
therefore be provided for plain accesses, in spite of all the shared-variable shenanigans
called out in Section 4.3.4.1.
Of course, each CPU sees its own accesses in order and the compiler always has fully
accounted for intra-CPU shenanigans, give or take the occasional compiler bug. These
facts are what enables the lockless fastpaths in memblock_alloc() and memblock_
free(), which are shown in Listings 6.10 and 6.11, respectively. However, this is also
why the developer is responsible for providing appropriate ordering (for example, by
using smp_store_release()) when publishing a pointer to a newly allocated block
of memory. After all, in the CPU-local case, the allocator has not necessarily provided
any cross-CPU ordering.
This means that the allocator must provide ordering when rebalancing its per-
thread pools. This ordering is provided by the calls to spin_lock() and spin_
unlock() from memblock_alloc() and memblock_free(). For any block that
has migrated from one thread to another, the old thread will have executed spin_
unlock(&globalmem.mutex) after placing the block in the globalmem pool, and the
new thread will have executed spin_lock(&globalmem.mutex) before moving that
block to its per-thread pool. This spin_unlock() and spin_lock() ensures that both
the old and new threads see the old thread’s accesses as having happened before those
of the new thread.
v2023.06.11a
Quick Quiz 15.38: But doesn’t PowerPC have weak unlock-lock ordering properties within the
Linux kernel, allowing a write before the unlock to be reordered with a read after the lock?
Therefore, the ordering required by conventional uses of memory allocation can be

provided solely by non-fastpath locking, allowing the fastpath to remain synchronization-
free.
15.4.2 Locking
Locking is a well-known synchronization primitive with which the parallel-programming
community has had decades of experience. As such, locking’s semantics are quite
simple.
That is, they are quite simple until you start trying to mathematically model them.
The simple part is that any CPU or thread holding a given lock is guaranteed to see
any accesses executed by CPUs or threads while they were previously holding that
same lock. Similarly, any CPU or thread holding a given lock is guaranteed not to see
accesses that will be executed by other CPUs or threads while subsequently holding that
same lock. And what else is there?
As it turns out, quite a bit:
1. Are CPUs, threads, or compilers allowed to pull memory accesses into a given
lock-based critical section?
2. Will a CPU or thread holding a given lock also be guaranteed to see accesses
executed by CPUs and threads before they last acquired that same lock, and vice
versa?
3. Suppose that a given CPU or thread executes one access (call it “A”), releases a
lock, reacquires that same lock, then executes another access (call it “B”). Is some
other CPU or thread not holding that lock guaranteed to see A and B in order?
4. As above, but with the lock reacquisition carried out by some other CPU or thread?
5. As above, but with the lock reacquisition being some other lock?
6. What ordering guarantees are provided by spin_is_locked()?
The reaction to some or even all of these questions might well be “Why would anyone
do that?” However, any complete mathematical definition of locking must have answers
to all of these questions. Therefore, the following sections address these questions in
the context of the Linux kernel.
15.4.2.1 Accesses Into Critical Sections?

Can memory accesses be reordered into lock-based critical sections?
Within the context of the Linux-kernel memory model, the simple answer is “yes”.
This may be verified by running the litmus tests shown in Listings 15.33 and 15.34
(C-Lock-before-into.litmus and C-Lock-after-into.litmus, respectively),
both of which will yield the Sometimes result. This result indicates that the exists
clause can be satisfied, that is, that the final value of both P0()’s and P1()’s r1 variable
can be zero. This means that neither spin_lock() nor spin_unlock() are required
to act as a full memory barrier.
v2023.06.11a
Listing 15.33: Prior Accesses Into Critical Section (Ordering?)

1 C Lock-before-into
2
3 {}
4
5 P0(int *x, int *y, spinlock_t *sp)
6 {
7 int r1;
8
10 spin_lock(sp);
12 spin_unlock(sp);
13 }
14
16 {
17 int r1;
18
20 smp_mb();
22 }
23
24 exists (0:r1=0 /\ 1:r1=0)
Listing 15.34: Subsequent Accesses Into Critical Section (Ordering?)

1 C Lock-after-into
2
3 {}
4
6 {
7 int r1;
8
9 spin_lock(sp);
11 spin_unlock(sp);
13 }
14
16 {
17 int r1;
18
20 smp_mb();
22 }
23
24 exists (0:r1=0 /\ 1:r1=0)
v2023.06.11a
However, other environments might make other choices. For example, locking
implementations that run only on the x86 CPU family will have lock-acquisition
primitives that fully order the lock acquisition with any prior and any subsequent
accesses. Therefore, on such systems the ordering shown in Listing 15.33 comes for
free. There are x86 lock-release implementations that are weakly ordered, thus failing to
provide the ordering shown in Listing 15.34, but an implementation could nevertheless
choose to guarantee this ordering.
For their part, weakly ordered systems might well choose to execute the memory-
barrier instructions required to guarantee both orderings, possibly simplifying code
making advanced use of combinations of locked and lockless accesses. However, as
noted earlier, LKMM chooses not to provide these additional orderings, in part to avoid
imposing performance penalties on the simpler and more prevalent locking use cases.
Instead, the smp_mb__after_spinlock() and smp_mb__after_unlock_lock()
primitives are provided for those more complex use cases, as discussed in Section 15.5.
Thus far, this section has discussed only hardware reordering. Can the compiler also
reorder memory references into lock-based critical sections?
The answer to this question in the context of the Linux kernel is a resounding “No!”
One reason for this otherwise inexplicable favoring of hardware reordering over compiler
optimizations is that the hardware will avoid reordering a page-faulting access into a
lock-based critical section. In contrast, compilers have no clue about page faults, and
would therefore happily reorder a page fault into a critical section, which could crash
the kernel. The compiler is also unable to reliably determine which accesses will result
in cache misses, so that compiler reordering into critical sections could also result in
excessive lock contention. Therefore, the Linux kernel prohibits the compiler (but not
the CPU) from moving accesses into lock-based critical sections.
15.4.2.2 Accesses Outside of Critical Section?
If a given CPU or thread holds a given lock, it is guaranteed to see accesses executed
during all prior critical sections for that same lock. Similarly, such a CPU or thread
is guaranteed not to see accesses that will be executed during all subsequent critical
sections for that same lock.
But what about accesses preceding prior critical sections and following subsequent
critical sections?
This question can be answered for the Linux kernel by referring to Listing 15.35
(C-Lock-outside-across.litmus). Running this litmus test yields the Never result,
which means that accesses in code leading up to a prior critical section is also visible to
the current CPU or thread holding that same lock. Similarly, code that is placed after a
subsequent critical section is never visible to the current CPU or thread holding that
same lock.
As a result, the Linux kernel cannot allow accesses to be moved across the entirety of
a given critical section. Other environments might well wish to allow such code motion,
but please be advised that doing so is likely to yield profoundly counter-intuitive results.
In short, the ordering provided by spin_lock() extends not only throughout the
critical section, but also indefinitely beyond the end of that critical section. Similarly, the
ordering provided by spin_unlock() extends not only throughout the critical section,
but also indefinitely beyond the beginning of that critical section.
v2023.06.11a
Listing 15.35: Accesses Outside of Critical Sections

1 C Lock-outside-across
2
3 {}
4
6 {
7 int r1;
8
10 spin_lock(sp);
12 spin_unlock(sp);
13 }
14
16 {
17 int r1;
18
19 spin_lock(sp);
21 spin_unlock(sp);
23 }
24
25 exists (0:r1=0 /\ 1:r1=0)
15.4.2.3 Ordering for Non-Lock Holders?
Does a CPU or thread that is not holding a given lock see that lock’s critical sections as
being ordered?
This question can be answered for the Linux kernel by referring to Listing 15.36
(C-Lock-across-unlock-lock-1.litmus), which shows an example where P(0)
places its write and read in two different critical sections for the same lock. Running
this litmus test shows that the exists can be satisfied, which means that the answer
is “no”, and that CPUs can reorder accesses across consecutive critical sections. In
other words, not only are spin_lock() and spin_unlock() weaker than a full barrier
when considered separately, they are also weaker than a full barrier when taken together.
If the ordering of a given lock’s critical sections are to be observed, then either the
observer must hold that lock on the one hand or either smp_mb__after_spinlock()
or smp_mb__after_unlock_lock() must be executed just after the second lock
acquisition on the other.
But what if the two critical sections run on different CPUs or threads?
This question is answered for the Linux kernel by referring to Listing 15.37 (C-Lock-
across-unlock-lock-2.litmus), in which the first lock acquisition is executed by
P0() and the second lock acquisition is executed by P1(). Note that P1() must read x
to reject executions in which P1() executes before P0() does. Running this litmus test
shows that the exists can be satisfied, which means that the answer is “no”, and that
CPUs can reorder accesses across consecutive critical sections, even if each of those
critical sections runs on a different CPU or thread.
Quick Quiz 15.39: But if there are three critical sections, isn’t it true that CPUs not holding
the lock will observe the accesses from the first and the third critical section as being ordered?
As before, if the ordering of a given lock’s critical sections are to be observed, then
either the observer must hold that lock or either smp_mb__after_spinlock() or
smp_mb__after_unlock_lock() must be executed just after P1()’s lock acquisition.
v2023.06.11a
Listing 15.36: Accesses Between Same-CPU Critical Sections (Ordering?)

1 C Lock-across-unlock-lock-1
2
3 {}
4
6 {
7 int r1;
8
9 spin_lock(sp);
11 spin_unlock(sp);
12 spin_lock(sp);
14 spin_unlock(sp);
15 }
16
18 {
19 int r1;
20
22 smp_mb();
24 }
25
26 exists (0:r1=0 /\ 1:r1=0)
Listing 15.37: Accesses Between Different-CPU Critical Sections (Ordering?)

2
3 {}
4
6 {
7 spin_lock(sp);
9 spin_unlock(sp);
10 }
11
13 {
14 int r1;
15 int r2;
16
17 spin_lock(sp);
20 spin_unlock(sp);
21 }
22
24 {
25 int r1;
26
28 smp_mb();
30 }
31
32 exists (1:r1=1 /\ 1:r2=0 /\ 2:r1=0)
v2023.06.11a
Given that ordering is not guaranteed when both critical sections are protected by the
same lock, there is no hope of any ordering guarantee when different locks are used.
However, readers are encouraged to construct the corresponding litmus test and see this
for themselves.
This situation can seem counter-intuitive, but it is rare for code to care. This approach
also allows certain weakly ordered systems to implement locks more efficiently.
15.4.2.4 Ordering for spin_is_locked()?

The Linux kernel’s spin_is_locked() primitive returns true if the specified lock
is held and false otherwise. Note that spin_is_locked() returns true when some
other CPU or thread holds the lock, not just when the current CPU or thread holds that
lock. This raises the question of what ordering guarantees spin_is_locked() might
provide.
In the Linux kernel, the answer has varied over time. Initially, spin_is_locked()
was unordered, but a few interesting use cases motivated strong ordering. Later
discussions surrounding the Linux-kernel memory model concluded that spin_is_
locked() should be used only for debugging. Part of the reason for this is that
even a fully ordered spin_is_locked() might return true because some other CPU
or thread was just about to release the lock in question. In this case, there is little
that can be learned from that return value of true, which means that reliable use of
spin_is_locked() is surprisingly complex. Other approaches almost always work
better, for example, use of explicit shared variables or the spin_trylock() primitive.
This situation resulted in the current state, namely that spin_is_locked() provides
no ordering guarantees, except that if it returns false, the current CPU or thread cannot
be holding the corresponding lock.
Quick Quiz 15.40: But if spin_is_locked() returns false, don’t we also know that no
other CPU or thread is holding the corresponding lock?
15.4.2.5 Why Mathematically Model Locking?

Given all these possible choices, why model locking in general? Why not simply model
a simple implementation?
One reason is modeling performance, as shown in Table E.5 on page 825. Directly
modeling locking in general is orders of magnitude faster than emulating even a
trivial implementation. This should be no surprise, given the combinatorial explosion
experienced by present-day formal-verification tools with increases in the number of
memory accesses executed by the code being modeled. Splitting the modeling at API
boundaries can therefore result in combinatorial implosion.
Another reason is that a trivial implementation might needlessly constrain either real
implementations or real use cases. In contrast, modeling a platonic lock allows the
widest variety of implementations while providing specific guidance to locks’ users.
15.4.3 RCU
As described in Section 9.5.2, the fundamental property of RCU grace periods is this
straightforward two-part guarantee: (1) If any part of a given RCU read-side critical
section precedes the beginning of a given grace period, then the entirety of that critical
section precedes the end of that grace period. (2) If any part of a given RCU read-side
v2023.06.11a
If happens before ...
rcu_read_lock() call_rcu()
rcu_read_unlock()
... then happens before

... then happens before
rcu_read_lock()
callback invocation rcu_read_unlock()
If happens before ...
Figure 15.18: RCU Grace-Period Ordering Guarantees
Listing 15.38: RCU Fundamental Property

1 C C-SB+o-rcusync-o+rl-o-o-rul
2
3 {}
4
5 P0(uintptr_t *x0, uintptr_t *x1)
6 {
9 uintptr_t r2 = READ_ONCE(*x1);
10 }
11
13 {
14 rcu_read_lock();
18 }
19
20 exists (1:r2=0 /\ 0:r2=0)
critical section follows the end of a given grace period, then the entirety of that critical
section follows the beginning of that grace period. These guarantees are summarized
in Figure 15.18, where the grace period is denoted by the dashed arrow between
the call_rcu() invocation in the upper right and the corresponding RCU callback
invocation in the lower left.14
In short, an RCU read-side critical section is guaranteed never to completely overlap
an RCU grace period, as demonstrated by Listing 15.38 (C-SB+o-rcusync-o+rl-o-
o-rul.litmus). Either or neither of the r2 registers can have the final value of zero,
but at least one of them must be non-zero (that is, the cycle identified by the exists
clause is prohibited), courtesy of RCU’s fundamental grace-period guarantee, as can
14 For more detail, please see Figures 9.11–9.13 starting on page 231.
v2023.06.11a
Listing 15.39: RCU Fundamental Property and Reordering

1 C C-SB+o-rcusync-o+i-rl-o-o-rul
2
3 {}
4
6 {
10 }
11
13 {
14 rcu_read_lock();
18 }
19
20 exists (1:r2=0 /\ 0:r2=0)
be seen by running herd on this litmus test. Note that this guarantee is insensitive to
the ordering of the accesses within P1()’s critical section, so the litmus test shown in
Listing 15.3915 also forbids this same cycle.
However, this definition is incomplete, as can be seen from the following list of
questions:16
1. What ordering is provided by rcu_read_lock() and rcu_read_unlock(),

independent of RCU grace periods?
2. What ordering is provided by synchronize_rcu() and synchronize_rcu_
expedited(), independent of RCU read-side critical sections?
3. If the entirety of a given RCU read-side critical section precedes the end of a given
RCU grace period, what about accesses preceding that critical section?
4. If the entirety of a given RCU read-side critical section follows the beginning of a
given RCU grace period, what about accesses following that critical section?
5. What happens in situations involving more than one RCU read-side critical section
and/or more than one RCU grace period?
6. What happens when RCU is combined with other memory-ordering mechanisms?
These questions are addressed in the following sections.
15.4.3.1 RCU Read-Side Ordering

On their own, RCU’s read-side primitives rcu_read_lock() and rcu_read_
unlock() provide no ordering whatsoever. In particular, despite their names, they
do not act like locks, as can be seen in Listing 15.40 (C-LB+rl-o-o-rul+rl-o-o-
rul.litmus). This litmus test’s cycle is allowed: Both instances of the r1 register can
have final values of 1.
15 Dependencies can of course limit the ability to reorder accesses within RCU read-side
critical sections.
16 Several of which were introduced to Paul by Jade Alglave during early work on LKMM,
and a few more of which came from other LKMM participants [AMM+ 18].
v2023.06.11a
Listing 15.40: RCU Readers Provide No Lock-Like Ordering

1 C C-LB+rl-o-o-rul+rl-o-o-rul
2
3 {}
4
6 {
7 rcu_read_lock();
11 }
12
14 {
15 rcu_read_lock();
19 }
20
21 exists (0:r1=1 /\ 1:r1=1)
Listing 15.41: RCU Readers Provide No Barrier-Like Ordering

1 C C-LB+o-rl-rul-o+o-rl-rul-o
2
3 {}
4
6 {
8 rcu_read_lock();
11 }
12
14 {
16 rcu_read_lock();
19 }
20
21 exists (0:r1=1 /\ 1:r1=1)
Nor do these primitives have barrier-like ordering properties, at least not unless there
is a grace period in the mix, as can be seen in Listing 15.41 (C-LB+o-rl-rul-o+o-
rl-rul-o.litmus). This litmus test’s cycle is also allowed. (Try it!)
Of course, lack of ordering in both these litmus tests should be absolutely no surprise,
given that both rcu_read_lock() and rcu_read_unlock() are no-ops in the QSBR
implementation of RCU.
15.4.3.2 RCU Update-Side Ordering

In contrast with RCU readers, the RCU update-side functions synchronize_rcu()
and synchronize_rcu_expedited() provide memory ordering at least as strong as
smp_mb(),17 as can be seen by running herd on the litmus test shown in Listing 15.42.
This test’s cycle is prohibited, just as it would with smp_mb(). This should be no
surprise given the information presented in Table 15.3.
17 And also way more expensive!
v2023.06.11a
Listing 15.42: RCU Updaters Provide Full Ordering

1 C C-SB+o-rcusync-o+o-rcusync-o
2
3 {}
4
6 {
10 }
11
13 {
17 }
18
19 exists (1:r2=0 /\ 0:r2=0)
Listing 15.43: What Happens Before RCU Readers?

1 C C-SB+o-rcusync-o+o-rl-o-rul
2
3 {}
4
6 {
10 }
11
13 {
15 rcu_read_lock();
18 }
19
20 exists (1:r2=0 /\ 0:r2=0)
15.4.3.3 RCU Readers: Before and After

Before reading this section, it would be well to reflect on the distinction between
guarantees that are available and guarantees that maintainable software should rely
on. Keeping that firmly in mind, this section presents a few of the more exotic RCU
guarantees.
Listing 15.43 (C-SB+o-rcusync-o+o-rl-o-rul.litmus) shows a litmus test
similar to that in Listing 15.38, but with the RCU reader’s first access preceding the
RCU read-side critical section, rather than the more conventional (and maintainable!)
approach of being contained within it. Perhaps surprisingly, running herd on this litmus
test gives the same result as for that in Listing 15.38: The cycle is forbidden.
Why would this be the case?
Because both of P1()’s accesses are volatile, as discussed in Section 4.3.4.2, the
compiler is not permitted to reorder them. This means that the code emitted for
P1()’s WRITE_ONCE() will precede that of P1()’s READ_ONCE(). Therefore, RCU
implementations that place memory-barrier instructions in rcu_read_lock() and
rcu_read_unlock() will preserve the ordering of P1()’s two accesses all the way
down to the hardware level. On the other hand, RCU implementations that rely on
interrupt-based state machines will also fully preserve this ordering relative to the grace
v2023.06.11a
Listing 15.44: What Happens After RCU Readers?

1 C C-SB+o-rcusync-o+rl-o-rul-o
2
3 {}
4
6 {
10 }
11
13 {
14 rcu_read_lock();
18 }
19
20 exists (1:r2=0 /\ 0:r2=0)
Listing 15.45: What Happens With Empty RCU Readers?

1 C C-SB+o-rcusync-o+o-rl-rul-o
2
3 {}
4
6 {
10 }
11
13 {
15 rcu_read_lock();
18 }
19
20 exists (1:r2=0 /\ 0:r2=0)
period due to the fact that interrupts take place at a precise location in the execution of
the interrupted code.
This in turn means that if the WRITE_ONCE() follows the end of a given RCU grace
period, then the accesses within and following that RCU read-side critical section must
follow the beginning of that same grace period. Similarly, if the READ_ONCE() precedes
the beginning of the grace period, everything within and preceding that critical section
must precede the end of that same grace period.
Listing 15.44 (C-SB+o-rcusync-o+rl-o-rul-o.litmus) is similar, but instead
looks at accesses after the RCU read-side critical section. This test’s cycle is also
forbidden, as can be checked with the herd tool. The reasoning is similar to that for
Listing 15.43, and is left as an exercise for the reader.
Listing 15.45 (C-SB+o-rcusync-o+o-rl-rul-o.litmus) takes things one step
farther, moving P1()’s WRITE_ONCE() to precede the RCU read-side critical section
and moving P1()’s READ_ONCE() to follow it, resulting in an empty RCU read-side
critical section.
Perhaps surprisingly, despite the empty critical section, RCU nevertheless still
manages to forbid the cycle. This can again be checked using the herd tool. Furthermore,
v2023.06.11a
Listing 15.46: What Happens With No RCU Readers?

1 C C-SB+o-rcusync-o+o-o
2
3 {}
4
6 {
10 }
11
13 {
16 }
17
18 exists (1:r2=0 /\ 0:r2=0)
the reasoning is once again similar to that for Listing 15.43, Recapping, if P1()’s
WRITE_ONCE() follows the end of a given grace period, then P1()’s RCU read-side
critical section—and everything following it—must follow the beginning of that same
grace period. Similarly, if P1()’s READ_ONCE() precedes the beginning of a given
grace period, then P1()’s RCU read-side critical section—and everything preceding
it—must precede the end of that same grace period. In both cases, the critical section’s
emptiness is irrelevant.
Quick Quiz 15.41: Wait a minute! In QSBR implementations of RCU, no code is emitted
for rcu_read_lock() and rcu_read_unlock(). This means that the RCU read-side critical
section in Listing 15.45 isn’t just empty, it is completely nonexistent!!! So how can something
that doesn’t exist at all possibly have any effect whatsoever on ordering???
This situation leads to the question of what happens if rcu_read_lock() and rcu_
read_unlock() are omitted entirely, as shown in Listing 15.46 (C-SB+o-rcusync-
o+o-o.litmus). As can be checked with herd, this litmus test’s cycle is allowed, that
is, both instances of r2 can have final values of zero.
This might seem strange in light of the fact that empty RCU read-side critical sections
can provide ordering. And it is true that QSBR implementations of RCU would in fact
forbid this outcome, due to the fact that there is no quiescent state anywhere in P1()’s
function body, so that P1() would run within an implicit RCU read-side critical section.
However, RCU also has non-QSBR implementations, which have no implied RCU
read-side critical section, and in turn no way for RCU to enforce ordering. Therefore,
this litmus test’s cycle is allowed.
Quick Quiz 15.42: Can P1()’s accesses be reordered in the litmus tests shown in Listings 15.43,
15.44, and 15.45 in the same way that they were reordered going from Listing 15.38 to
Listing 15.39?
15.4.3.4 Multiple RCU Readers and Updaters

Because synchronize_rcu() has ordering semantics that are at least as strong as smp_
mb(), no matter how many processes there are in an SB litmus test (such as Listing 15.42),
placing synchronize_rcu() between each process’s accesses prohibits the cycle. In
addition, the cycle is prohibited in an SB test where one process uses synchronize_
rcu() and the other uses rcu_read_lock() and rcu_read_unlock(), as shown by
v2023.06.11a
Listing 15.47: One RCU Grace Period and Two Readers

1 C C-SB+o-rcusync-o+rl-o-o-rul+rl-o-o-rul
2
3 {}
4
6 {
10 }
11
13 {
14 rcu_read_lock();
18 }
19
21 {
22 rcu_read_lock();
26 }
27
28 exists (2:r2=0 /\ 0:r2=0 /\ 1:r2=0)
Listing 15.38. However, if both processes use rcu_read_lock() and rcu_read_

unlock(), the cycle will be allowed, as shown by Listing 15.40.
Is it possible to say anything general about which RCU-protected litmus tests will be
prohibited and which will be allowed? This section takes up that question.
More specifically, what if the litmus test has one RCU grace period and two RCU
readers, as shown in Listing 15.47? The herd tool says that this cycle is allowed, but it
would be good to know why.18
The key point is that even strongly ordered CPUs such as x86 can and will re-
order P1()’s and P2()’s WRITE_ONCE() and READ_ONCE(). With that reordering,
Figure 15.19 shows how the cycle forms:
1. P0()’s read from x1 precedes P1()’s write, as depicted by the dashed arrow near
the bottom of the diagram.
2. Because P1()’s write follows the end of P0()’s grace period, P1()’s read from
x2 cannot precede the beginning of P0()’s grace period.
3. P1()’s read from x2 precedes P2()’s write.
4. Because P2()’s write to x2 precedes the end of P0()’s grace period, it is completely
legal for P2()’s read from x0 to precede the beginning of P0()’s grace period.
5. Therefore, P2()’s read from x0 can precede P0()’s write, thus allowing the cycle
to form.
But what happens when another grace period is added? This situation is shown in
Listing 15.48, an SB litmus test in which P0() and P1() have RCU grace periods and
18 Especially given that Paul changed his mind several times about this particular litmus
test when working with Jade Alglave to generalize RCU ordering semantics.
v2023.06.11a
Listing 15.48: Two RCU Grace Periods and Two Readers

1 C C-SB+o-rcusync-o+o-rcusync-o+rl-o-o-rul+rl-o-o-rul
2
3 {}
4
6 {
10 }
11
13 {
17 }
18
20 {
21 rcu_read_lock();
25 }
26
28 {
29 rcu_read_lock();
33 }
34
35 exists (3:r2=0 /\ 0:r2=0 /\ 1:r2=0 /\ 2:r2=0)
P0() P1() P2()
rcu_read_lock();
r2 = READ_ONCE(x0);
WRITE_ONCE(x0, 2);
synchronize_rcu();
rcu_read_lock();
r2 = READ_ONCE(x2);
WRITE_ONCE(x2, 2);
rcu_read_unlock();
r2 = READ_ONCE(x1);
WRITE_ONCE(x1, 2);
rcu_read_unlock();
Figure 15.19: Cycle for One RCU Grace Period and Two RCU Readers
v2023.06.11a
P0() P1() P2() P3()
WRITE_ONCE(x0, 2);
synchronize_rcu(); rcu_read_lock();
r2 = READ_ONCE(x0);
r2 = READ_ONCE(x1);
WRITE_ONCE(x1, 2);
r2 = READ_ONCE(x3);
WRITE_ONCE(x3, 2);
rcu_read_unlock();
r2 = READ_ONCE(x2);
WRITE_ONCE(x2, 2);
rcu_read_unlock();
Figure 15.20: No Cycle for Two RCU Grace Periods and Two RCU Readers
P2() and P3() have RCU readers. Again, the CPUs can reorder the accesses within
RCU read-side critical sections, as shown in Figure 15.20. For this cycle to form,
P2()’s critical section must end after P1()’s grace period and P3()’s must end after the
beginning of that same grace period, which happens to also be after the end of P0()’s
grace period. Therefore, P3()’s critical section must start after the beginning of P0()’s
grace period, which in turn means that P3()’s read from x0 cannot possibly precede
P0()’s write. Therefore, the cycle is forbidden because RCU read-side critical sections
cannot span full RCU grace periods.
However, a closer look at Figure 15.20 makes it clear that adding a third reader would
allow the cycle. This is because this third reader could end before the end of P0()’s
grace period, and thus start before the beginning of that same grace period. This in turn
suggests the general rule, which is: In these sorts of RCU-only litmus tests, if there are
at least as many RCU grace periods as there are RCU read-side critical sections, the
cycle is forbidden.19
15.4.3.5 RCU and Other Ordering Mechanisms

But what about litmus tests that combine RCU with other ordering mechanisms?
The general rule is that it takes only one mechanism to forbid a cycle.
For example, refer back to Listing 15.40. Applying the general rule from the
previous section, because this litmus test has two RCU read-side critical sections and
no RCU grace periods, the cycle is allowed. But what if P0()’s WRITE_ONCE() is
19 Interestingly enough, Alan Stern proved that within the context of LKMM, the two-part
fundamental property of RCU expressed in Section 9.5.2 actually implies this seemingly more
general result, which is called the RCU axiom [AMM+ 18].
v2023.06.11a
P0() P1() P2()
WRITE_ONCE(x0, 2);
r2 = READ_ONCE(x0);
rcu_read_lock();
r2 = READ_ONCE(x1);
WRITE_ONCE(x1, 2);
smp_mb();
r2 = READ_ONCE(x2);
WRITE_ONCE(x2, 2);
rcu_read_unlock();
rcu_read_unlock();
Figure 15.21: Cycle for One RCU Grace Period, Two RCU Readers, and Memory
Barrier
replaced by an smp_store_release() and P1()’s READ_ONCE() is replaced by an

smp_load_acquire()?
RCU would still allow the cycle, but the release-acquire pair would forbid it. Because
it only takes one mechanism to forbid a cycle, the release-acquire pair would prevail,
thus forbidding the cycle.
For another example, refer back to Listing 15.47. Because this litmus test has two
RCU readers but only one grace period, its cycle is allowed. But suppose that an
smp_mb() was placed between P1()’s pair of accesses. In this new litmus test, because
of the addition of the smp_mb(), P2()’s as well as P1()’s critical sections would extend
beyond the end of P0()’s grace period, which in turn would prevent P2()’s read from
x0 from preceding P0()’s write, as depicted by the red dashed arrow in Figure 15.21. In
this case, RCU and the full memory barrier work together to forbid the cycle, with RCU
preserving ordering between P0() and both P1() and P2(), and with the smp_mb()
preserving ordering between P1() and P2().
Quick Quiz 15.43: What would happen if the smp_mb() was instead added between P2()’s
accesses in Listing 15.47?
In short, where RCU’s semantics were once purely pragmatic, they are now fully
formalized [MW05, DMS+ 12, GRY13, AMM+ 18].
15.4.4 Higher-Level Primitives: Discussion

It is quite beneficial to verify code in terms of a higher-level primitive instead of
in terms of the low-level memory accesses used in a particular implementation of
that primitive. First, this allows code using‘those primitives to be verified against an
abstract representation of those primitives, thus making that code less vulnerable to
implementation changes. Second, partitioning the verification at API boundaries results
in combinatorial implosion, greatly reducing the overhead of formal verification.
v2023.06.11a
Table 15.5: Summary of Memory Ordering
CPU Family
SPARC TSO
Armv7-A/R
z Systems
POWER
Itanium
Armv8
Alpha
MIPS
x86
Property
Memory Ordering Loads Reordered After Loads or Stores? Y Y Y Y Y Y

Stores Reordered After Stores? Y Y Y Y Y Y
Stores Reordered After Loads? Y Y Y Y Y Y Y Y Y
Atomic Instructions Reordered With
Y Y Y Y Y
Loads or Stores?
Dependent Loads Reordered? Y
Dependent Stores Reordered?
Non-Sequentially Consistent? Y Y Y Y Y Y Y Y Y
Non-Multicopy Atomic? Y Y Y Y Y Y Y Y
Non-Other-Multicopy Atomic? Y Y Y Y Y
Non-Cache Coherent? Y
Instructions Load-Acquire/Store-Release? F F i I F b
Atomic RMW Instruction Type? L L L C L L C C C
Incoherent Instruction Cache/Pipeline? Y Y Y Y Y Y Y Y Y
Key: Load-Acquire/Store-Release?
b: Lightweight memory barrier
F: Full memory barrier
i: Instruction with lightweight ordering
I: Instruction with heavyweight ordering
Atomic RMW Instruction Type?
C: Compare-and-exchange instruction
L: Load-linked/store-conditional instruction
It is hoped that verifying against detailed semantics for higher-level primitives will
greatly increase the effectiveness of static analysis and model checking.
15.5 Hardware Specifics
Rock beats paper!

Derek Williams
Each CPU family has its own peculiar approach to memory ordering, which can make
portability a challenge, as you can see in Table 15.5.
In fact, some software environments simply prohibit direct use of memory-ordering
operations, restricting the programmer to mutual-exclusion primitives that incorporate
them to the extent that they are required. Please note that this section is not intended to
be a reference manual covering all (or even most) aspects of each CPU family, but rather
v2023.06.11a
15.5. HARDWARE SPECIFICS 551
a high-level overview providing a rough comparison. For full details, see the reference
manual for the CPU of interest.
Getting back to Table 15.5, the first group of rows look at memory-ordering properties
and the second group looks at instruction properties. Please note that these properties
hold at the machine-instruction level. Compilers can and do reorder far more aggressively
than does hardware. Use marked accesses such as READ_ONCE() and WRITE_ONCE()
to constrain the compiler’s optimizations and prevent undesireable reordering.
The first three rows indicate whether a given CPU allows the four possible combinations
of loads and stores to be reordered, as discussed in Section 15.1 and Sections 15.2.2.1–
15.2.2.3. The next row (“Atomic Instructions Reordered With Loads or Stores?”)
indicates whether a given CPU allows loads and stores to be reordered with atomic
instructions.
The fifth and sixth rows cover reordering and dependencies, which was covered in
Sections 15.2.3–15.2.5 and which is explained in more detail in Section 15.5.1. The
short version is that Alpha requires memory barriers for readers as well as updaters
of linked data structures, however, these memory barriers are provided by the Alpha
architecture-specific code in v4.15 and later Linux kernels.
The next row, “Non-Sequentially Consistent”, indicates whether the CPU’s normal
load and store instructions are constrained by sequential consistency. Performance
considerations have dictated that no modern mainstream system is sequentially consistent.
The next three rows cover multicopy atomicity, which was defined in Section 15.2.7.
The first is full-up (and rare) multicopy atomicity, the second is the weaker other-
multicopy atomicity, and the third is the weakest non-multicopy atomicity.
The next row, “Non-Cache Coherent”, covers accesses from multiple threads to a
single variable, which was discussed in Section 15.2.6.
The final three rows cover instruction-level choices and issues. The first row indicates
how each CPU implements load-acquire and store-release, the second row classifies
CPUs by atomic-instruction type, and the third and final row indicates whether a given
CPU has an incoherent instruction cache and pipeline. Such CPUs require special
instructions be executed for self-modifying code.
The common “just say no” approach to memory-ordering operations can be eminently
reasonable where it applies, but there are environments, such as the Linux kernel, where
direct use of memory-ordering operations is required. Therefore, Linux provides a
carefully chosen least-common-denominator set of memory-ordering primitives, which
are as follows:
smp_mb() (full memory barrier) that orders both loads and stores. This means that
loads and stores preceding the memory barrier will be committed to memory
before any loads and stores following the memory barrier.
smp_rmb() (read memory barrier) that orders only loads.
smp_wmb() (write memory barrier) that orders only stores.
smp_mb__before_atomic() that forces ordering of accesses preceding the smp_mb__

before_atomic() against accesses following a later RMW atomic operation.
This is a noop on systems that fully order atomic RMW operations.
smp_mb__after_atomic() that forces ordering of accesses preceding an earlier RMW

atomic operation against accesses following the smp_mb__after_atomic(). This
is also a noop on systems that fully order atomic RMW operations.
v2023.06.11a
smp_mb__after_spinlock() that forces ordering of accesses preceding a lock ac-

quisition against accesses following the smp_mb__after_spinlock(). This is
also a noop on systems that fully order lock acquisitions.
mmiowb() that forces ordering on MMIO writes that are guarded by global spinlocks,
and is more thoroughly described in a 2016 LWN article on MMIO [MDR16].
The smp_mb(), smp_rmb(), and smp_wmb() primitives also force the compiler to
eschew any optimizations that would have the effect of reordering memory optimizations
across the barriers.
Quick Quiz 15.44: What happens to code between an atomic operation and an smp_mb__
after_atomic()?
These primitives generate code only in SMP kernels, however, several have UP
versions (mb(), rmb(), and wmb(), respectively) that generate a memory barrier even
in UP kernels. The smp_ versions should be used in most cases. However, these
latter primitives are useful when writing drivers, because MMIO accesses must remain
ordered even in UP kernels. In absence of memory-ordering operations, both CPUs and
compilers would happily rearrange these accesses, which at best would make the device
act strangely, and could crash your kernel or even damage your hardware.
So most kernel programmers need not worry about the memory-ordering peculiarities
of each and every CPU, as long as they stick to these interfaces and to the fully ordered
atomic operations.20 If you are working deep in a given CPU’s architecture-specific
code, of course, all bets are off.
Furthermore, all of Linux’s locking primitives (spinlocks, reader-writer locks, sem-
aphores, RCU, . . .) include any needed ordering primitives. So if you are working
with code that uses these primitives properly, you need not worry about Linux’s
memory-ordering primitives.
That said, deep knowledge of each CPU’s memory-consistency model can be very
helpful when debugging, to say nothing of when writing architecture-specific code or
synchronization primitives.
Besides, they say that a little knowledge is a very dangerous thing. Just imagine the
damage you could do with a lot of knowledge! For those who wish to understand more
about individual CPUs’ memory consistency models, the next sections describe those of
a few popular and prominent CPUs. Although there is no substitute for actually reading
a given CPU’s documentation, these sections do give a good overview.
15.5.1 Alpha
It may seem strange to say much of anything about a CPU whose end of life has long
since passed, but Alpha is interesting because it is the only mainstream CPU that reorders
dependent loads, and has thus had outsized influence on concurrency APIs, including
within the Linux kernel. The need for core Linux-kernel code to accommodate Alpha
ended with version v4.15 of the Linux kernel, and all traces of this accommodation
were removed in v5.9 with the removal of the smp_read_barrier_depends() and
read_barrier_depends() APIs. This section is nevertheless retained in the Third
Edition because here in early 2023 there are still a few Linux kernel hackers still
working on pre-v4.15 versions of the Linux kernel. In addition, the modifications to
20 For a full list, expand the patterns in Documentation/atomic_t.txt.
v2023.06.11a
Listing 15.49: Insert and Lock-Free Search (No Ordering)

1 struct el *insert(long key, long data)
2 {
3 struct el *p;
4 p = kmalloc(sizeof(*p), GFP_ATOMIC);
5 spin_lock(&mutex);
6 p->next = head.next;
7 p->key = key;
8 p->data = data;
9 smp_store_release(&head.next, p);
10 spin_unlock(&mutex);
11 }
12
13 struct el *search(long searchkey)
14 {
15 struct el *p;
16 p = READ_ONCE_OLD(head.next);
17 while (p != &head) {
18 /* Prior to v4.15, BUG ON ALPHA!!! */
19 if (p->key == searchkey) {
20 return (p);
21 }
22 p = READ_ONCE_OLD(p->next);
23 };
24 return (NULL);
25 }
READ_ONCE() that permitted these APIs to be removed have not necessarily propagated
to all userspace projects that might still support Alpha.
The dependent-load difference between Alpha and the other CPUs is illustrated by
the code shown in Listing 15.49. This smp_store_release() guarantees that the
element initialization in lines 6–8 is executed before the element is added to the list on
line 9, so that the lock-free search will work correctly. That is, it makes this guarantee
on all CPUs except Alpha.
Given the pre-v4.15 implementation of READ_ONCE(), indicated by READ_ONCE_
OLD() in the listing, Alpha actually allows the code on line 19 of Listing 15.49 to see
the old garbage values that were present before the initialization on lines 6–8.
Figure 15.22 shows how this can happen on an aggressively parallel machine with
partitioned caches, so that alternating cache lines are processed by the different partitions
of the caches. For example, the load of head.next on line 16 of Listing 15.49 might
access cache bank 0, and the load of p->key on line 19 and of p->next on line 22
might access cache bank 1. On Alpha, the smp_store_release() will guarantee that
the cache invalidations performed by lines 6–8 of Listing 15.49 (for p->next, p->key,
and p->data) will reach the interconnect before that of line 9 (for head.next), but
makes absolutely no guarantee about the order of propagation through the reading
CPU’s cache banks. For example, it is possible that the reading CPU’s cache bank 1 is
very busy, but cache bank 0 is idle. This could result in the cache invalidations for the
new element (p->next, p->key, and p->data) being delayed, so that the reading CPU
loads the new value for head.next, but loads the old cached values for p->key and
p->next. Yes, this does mean that Alpha can in effect fetch the data pointed to before
it fetches the pointer itself, strange but true. See the documentation [Com01, Pug00]
called out earlier for more information, or if you think that I am just making all this
up.21 The benefit of this unusual approach to ordering is that Alpha can use simpler
cache hardware, which in turn permitted higher clock frequencies in Alpha’s heyday.
21 Of course, the astute reader will have already recognized that Alpha is nowhere near as
mean and nasty as it could be, the (thankfully) mythical architecture in Appendix C.6.1 being
a case in point.
v2023.06.11a
p->data = key;
smp_wmb(); p = READ_ONCE_OLD(head.next);
head.next = p; BUG_ON(p && p->key != key);
Writing CPU Core Reading CPU Core

Bank 0 Bank 1
Bank 0 Bank 1 (Idle) (Busy)
head.next p->key
p->data
p->next
Figure 15.22: Why smp_read_barrier_depends() is Required in Pre-v4.15 Linux

Kernels
One could place an smp_rmb() primitive between the pointer fetch and dereference
in order to force Alpha to order the pointer fetch with the later dependent load. However,
this imposes unneeded overhead on systems (such as Arm, Itanium, PPC, and SPARC)
that respect data dependencies on the read side. A smp_read_barrier_depends()
primitive was therefore added to the Linux kernel to eliminate overhead on these systems,
but was removed in v5.9 of the Linux kernel in favor of augmenting Alpha’s definition
of READ_ONCE(). Thus, as of v5.9, core kernel code no longer needs to concern itself
with this aspect of DEC Alpha. However, it is better to use rcu_dereference() as
shown on lines 16 and 21 of Listing 15.50, which works safely and efficiently for all
recent kernel versions.
It is also possible to implement a software mechanism that could be used in place
of smp_store_release() to force all reading CPUs to see the writing CPU’s writes
in order. This software barrier could be implemented by sending inter-processor
interrupts (IPIs) to all other CPUs. Upon receipt of such an IPI, a CPU would execute a
memory-barrier instruction, implementing a system-wide memory barrier similar to
that provided by the Linux kernel’s sys_membarrier() system call. Additional logic
is required to avoid deadlocks. Of course, CPUs that respect data dependencies would
define such a barrier to simply be smp_store_release(). However, this approach
was deemed by the Linux community to impose excessive overhead [McK01], and to
their point would be completely inappropriate for systems having aggressive real-time
response requirements.
The Linux memory-barrier primitives took their names from the Alpha instructions,
so smp_mb() is mb, smp_rmb() is rmb, and smp_wmb() is wmb. Alpha is the only CPU
whose READ_ONCE() includes an smp_mb().
Quick Quiz 15.45: Why does Alpha’s READ_ONCE() include an mb() rather than rmb()?
Quick Quiz 15.46: Isn’t DEC Alpha significant as having the weakest possible memory
ordering?
For more on Alpha, see its reference manual [Cor02].
v2023.06.11a
Listing 15.50: Safe Insert and Lock-Free Search

1 struct el *insert(long key, long data)
2 {
3 struct el *p;
4 p = kmalloc(sizeof(*p), GFP_ATOMIC);
5 spin_lock(&mutex);
6 p->next = head.next;
7 p->key = key;
8 p->data = data;
9 smp_store_release(&head.next, p);
10 spin_unlock(&mutex);
11 }
12
13 struct el *search(long searchkey)
14 {
15 struct el *p;
16 p = rcu_dereference(head.next);
17 while (p != &head) {
18 if (p->key == searchkey) {
19 return (p);
20 }
21 p = rcu_dereference(p->next);
22 };
23 return (NULL);
24 }
15.5.2 Armv7-A/R
The Arm family of CPUs is popular in deep embedded applications, particularly for
power-constrained microcontrollers. Its memory model is similar to that of POWER (see
Section 15.5.6), but Arm uses a different set of memory-barrier instructions [ARM10]:
DMB (data memory barrier) causes the specified type of operations to appear to have
completed before any subsequent operations of the same type. The “type” of
operations can be all operations or can be restricted to only writes (similar to the
Alpha wmb and the POWER eieio instructions). In addition, Arm allows cache
coherence to have one of three scopes: Single processor, a subset of the processors
(“inner”) and global (“outer”).
DSB (data synchronization barrier) causes the specified type of operations to actually
complete before any subsequent operations (of any type) are executed. The “type”
of operations is the same as that of DMB. The DSB instruction was called DWB
(drain write buffer or data write barrier, your choice) in early versions of the Arm
architecture.
ISB (instruction synchronization barrier) flushes the CPU pipeline, so that all instruc-
tions following the ISB are fetched only after the ISB completes. For example, if
you are writing a self-modifying program (such as a JIT), you should execute an
ISB between generating the code and executing it.
None of these instructions exactly match the semantics of Linux’s rmb() primitive,
which must therefore be implemented as a full DMB. The DMB and DSB instructions
have a recursive definition of accesses ordered before and after the barrier, which has
an effect similar to that of POWER’s cumulativity, both of which are stronger than
LKMM’s cumulativity described in Section 15.2.7.1.
Arm also implements control dependencies, so that if a conditional branch depends
on a load, then any store executed after that conditional branch will be ordered after
the load. However, loads following the conditional branch will not be guaranteed to be
v2023.06.11a
LDLAR
Figure 15.23: Half Memory Barrier
ordered unless there is an ISB instruction between the branch and the load. Consider
the following example:
1 r1 = x;
2 if (r1 == 0)
3 nop();
4 y = 1;
5 r2 = z;
6 ISB();
7 r3 = z;
In this example, load-store control dependency ordering causes the load from x on
line 1 to be ordered before the store to y on line 4. However, Arm does not respect
load-load control dependencies, so that the load on line 1 might well happen after the
load on line 5. On the other hand, the combination of the conditional branch on line 2
and the ISB instruction on line 6 ensures that the load on line 7 happens after the load
on line 1. Note that inserting an additional ISB instruction somewhere between lines 2
and 5 would enforce ordering between lines 1 and 5.
15.5.3 Armv8
Arm’s Armv8 CPU family [ARM17] includes 64-bit capabilities, in contrast to their
32-bit-only CPU described in Section 15.5.2. Armv8’s memory model closely resembles
its Armv7 counterpart, but adds load-acquire (LDLARB, LDLARH, and LDLAR) and store-
release (STLLRB, STLLRH, and STLLR) instructions. These instructions act as “half
memory barriers”, so that Armv8 CPUs can reorder previous accesses with a later LDLAR
instruction, but are prohibited from reordering an earlier LDLAR instruction with later
accesses, as fancifully depicted in Figure 15.23. Similarly, Armv8 CPUs can reorder an
earlier STLLR instruction with a subsequent access, but are prohibited from reordering
previous accesses with a later STLLR instruction. As one might expect, this means that
these instructions directly support the C11 notion of load-acquire and store-release.
However, Armv8 goes well beyond the C11 memory model by mandating that the
combination of a store-release and load-acquire act as a full barrier under certain
v2023.06.11a
circumstances. For example, in Armv8, given a store followed by a store-release

followed a load-acquire followed by a load, all to different variables and all from a single
CPU, all CPUs would agree that the initial store preceded the final load. Interestingly
enough, most TSO architectures (including x86 and the mainframe) do not make this
guarantee, as the two loads could be reordered before the two stores.
Armv8 is one of only two architectures that needs the smp_mb__after_spinlock()
primitive to be a full barrier, due to its relatively weak lock-acquisition implementation
in the Linux kernel.
Armv8 also has the distinction of being the first CPU whose vendor publicly defined
its memory ordering with an executable formal model [ARM17].
15.5.4 Itanium
Itanium offers a weak consistency model, so that in absence of explicit memory-
barrier instructions or dependencies, Itanium is within its rights to arbitrarily reorder
memory references [Int02a]. Itanium has a memory-fence instruction named mf,
but also has “half-memory fence” modifiers to loads, stores, and to some of its
atomic instructions [Int02b]. The acq modifier prevents subsequent memory-reference
instructions from being reordered before the acq, but permits prior memory-reference
instructions to be reordered after the acq, similar to the Armv8 load-acquire instructions.
Similarly, the rel modifier prevents prior memory-reference instructions from being
reordered after the rel, but allows subsequent memory-reference instructions to be
reordered before the rel.
These half-memory fences are useful for critical sections, since it is safe to push
operations into a critical section, but can be fatal to allow them to bleed out. However, as
one of the few CPUs with this property, Itanium at one time defined Linux’s semantics
of memory ordering associated with lock acquisition and release.22 Oddly enough,
actual Itanium hardware is rumored to implement both load-acquire and store-release
instructions as full barriers. Nevertheless, Itanium was the first mainstream CPU
to introduce the concept (if not the reality) of load-acquire and store-release into its
instruction set.
Quick Quiz 15.47: Given that hardware can have a half memory barrier, why don’t locking
primitives allow the compiler to move memory-reference instructions into lock-based critical
sections?
The Itanium mf instruction is used for the smp_rmb(), smp_mb(), and smp_wmb()
primitives in the Linux kernel. Despite persistent rumors to the contrary, the “mf”
mnemonic stands for “memory fence”.
Itanium also offers a global total order for release operations, including the mf
instruction. This provides the notion of transitivity, where if a given code fragment sees
a given access as having happened, any later code fragment will also see that earlier
access as having happened. Assuming, that is, that all the code fragments involved
correctly use memory barriers.
Finally, Itanium is the only architecture supporting the Linux kernel that can reorder
normal loads to the same variable. The Linux kernel avoids this issue because READ_
ONCE() emits a volatile load, which is compiled as a ld,acq instruction, which
forces ordering of all READ_ONCE() invocations by a given CPU, including those to the
same variable.
22 PowerPC is now the architecture with this dubious privilege.
v2023.06.11a
15.5.5 MIPS
The MIPS memory model [Wav16, page 479] appears to resemble that of Arm, Itanium,
and POWER, being weakly ordered by default, but respecting dependencies. MIPS
has a wide variety of memory-barrier instructions, but ties them not to hardware
considerations, but rather to the use cases provided by the Linux kernel and the C++11
standard [Smi19] in a manner similar to the Armv8 additions:
SYNC
Full barrier for a number of hardware operations in addition to memory references,
which is used to implement the v4.13 Linux kernel’s smp_mb() for OCTEON
systems.
SYNC_WMB
Write memory barrier, which can be used on OCTEON systems to implement the
smp_wmb() primitive in the v4.13 Linux kernel via the syncw mnemonic. Other
systems use plain sync.
SYNC_MB
Full memory barrier, but only for memory operations. This may be used to
implement the C++ atomic_thread_fence(memory_order_seq_cst).
SYNC_ACQUIRE
Acquire memory barrier, which could be used to implement C++’s atomic_
thread_fence(memory_order_acquire). In theory, it could also be used
to implement the v4.13 Linux-kernel smp_load_acquire() primitive, but in
practice sync is used instead.
SYNC_RELEASE
Release memory barrier, which may be used to implement C++’s atomic_thread_
fence(memory_order_release). In theory, it could also be used to implement
the v4.13 Linux-kernel smp_store_release() primitive, but in practice sync is
used instead.
SYNC_RMB
Read memory barrier, which could in theory be used to implement the smp_rmb()
primitive in the Linux kernel, except that current MIPS implementations supported
by the v4.13 Linux kernel do not need an explicit instruction to force ordering.
Therefore, smp_rmb() instead simply constrains the compiler.
SYNCI
Instruction-cache synchronization, which is used in conjunction with other instruc-
tions to allow self-modifying code, such as that produced by just-in-time (JIT)
compilers.
Informal discussions with MIPS architects indicates that MIPS has a definition of
transitivity or cumulativity similar to that of Arm and POWER. However, it appears
that different MIPS implementations can have different memory-ordering properties, so
it is important to consult the documentation for the specific MIPS implementation you
are using.
v2023.06.11a
15.5.6 POWER / PowerPC

The POWER and PowerPC CPU families have a wide variety of memory-barrier
instructions [IBM94, LHF05]:
sync causes all preceding operations to appear to have completed before any subsequent
operations are started. This instruction is therefore quite expensive.
lwsync (lightweight sync) orders loads with respect to subsequent loads and stores,
and also orders stores. However, it does not order stores with respect to subsequent
loads. The lwsync instruction may be used to implement load-acquire and store-
release operations. Interestingly enough, the lwsync instruction enforces the same
within-CPU ordering as does x86, z Systems, and coincidentally, SPARC TSO.
However, placing the lwsync instruction between each pair of memory-reference
instructions will not result in x86, z Systems, or SPARC TSO memory ordering.
On these other systems, if a pair of CPUs independently execute stores to different
variables, all other CPUs will agree on the order of these stores. Not so on
PowerPC, even with an lwsync instruction between each pair of memory-reference
instructions, because PowerPC is non-multicopy atomic.
eieio (enforce in-order execution of I/O, in case you were wondering) causes all
preceding cacheable stores to appear to have completed before all subsequent
stores. However, stores to cacheable memory are ordered separately from stores to
non-cacheable memory, which means that eieio will not force an MMIO store
to precede a spinlock release. This instruction may well be unique in having a
five-vowel mnemonic.
isync forces all preceding instructions to appear to have completed before any subse-
quent instructions start execution. This means that the preceding instructions must
have progressed far enough that any traps they might generate have either happened
or are guaranteed not to happen, and that any side-effects of these instructions (for
example, page-table changes) are seen by the subsequent instructions. However, it
does not force all memory references to be ordered, only the actual execution of
the instruction itself. Thus, the loads might return old still-cached values and the
isync instruction does not force values previously stored to be flushed from the
store buffers.
Unfortunately, none of these instructions line up exactly with Linux’s wmb() primitive,
which requires all stores to be ordered, but does not require the other high-overhead
actions of the sync instruction. The rmb() primitive doesn’t have a matching light-
weight instruction either. But there is no choice: ppc64 versions of wmb(), rmb(), and
mb() are defined to be the heavyweight sync instruction. However, Linux’s smp_wmb()
primitive is never used for MMIO (since a driver must carefully order MMIOs in UP
as well as SMP kernels, after all), so it is defined to be the lighter weight eieio or
lwsync instruction [MDR16]. The smp_mb() primitive is also defined to be the sync
instruction, while smp_rmb() is defined to be the lighter-weight lwsync instruction.
POWER features “cumulativity”, which can be used to obtain transitivity. When
used properly, any code seeing the results of an earlier code fragment will also see the
accesses that this earlier code fragment itself saw. Much more detail is available from
McKenney and Silvera [MS09].
POWER respects control dependencies in much the same way that Arm does, with the
exception that the POWER isync instruction is substituted for the Arm ISB instruction.
v2023.06.11a
Like Armv8, POWER requires smp_mb__after_spinlock() to be a full memory

barrier. In addition, POWER is the only architecture requiring smp_mb__after_
unlock_lock() to be a full memory barrier. In both cases, this is because of the
weak ordering properties of POWER’s locking primitives, due to the use of the lwsync
instruction to provide ordering for both acquisition and release.
Many members of the POWER architecture have incoherent instruction caches, so that
a store to memory will not necessarily be reflected in the instruction cache. Thankfully,
few people write self-modifying code these days, but JITs and compilers do it all the time.
Furthermore, recompiling a recently run program looks just like self-modifying code
from the CPU’s viewpoint. The icbi instruction (instruction cache block invalidate)
invalidates a specified cache line from the instruction cache, and may be used in these
situations.
15.5.7 SPARC TSO

Although SPARC’s TSO (total-store order) is used by both Linux and Solaris, the
architecture also defines PSO (partial store order) and RMO (relaxed-memory order).
Any program that runs in RMO will also run in either PSO or TSO, and similarly, a
program that runs in PSO will also run in TSO. Moving a shared-memory parallel
program in the other direction may require careful insertion of memory barriers.
Although SPARC’s PSO and RMO modes are not used much these days, they did
give rise to a very flexible memory-barrier instruction [SPA94] that permits fine-grained
control of ordering:
StoreStore orders preceding stores before subsequent stores. (This option is used by
the Linux smp_wmb() primitive.)
LoadStore orders preceding loads before subsequent stores.
StoreLoad orders preceding stores before subsequent loads.
LoadLoad orders preceding loads before subsequent loads. (This option is used by the
Linux smp_rmb() primitive.)
Sync fully completes all preceding operations before starting any subsequent operations.
MemIssue completes preceding memory operations before subsequent memory opera-
tions, important for some instances of memory-mapped I/O.
Lookaside does the same as MemIssue, but only applies to preceding stores and
subsequent loads, and even then only for stores and loads that access the same
memory location.
So, why is “membar #MemIssue” needed? Because a “membar #StoreLoad” could
permit a subsequent load to get its value from a store buffer, which would be disastrous
if the write was to an MMIO register that induced side effects on the value to be read.
In contrast, “membar #MemIssue” would wait until the store buffers were flushed
before permitting the loads to execute, thereby ensuring that the load actually gets
its value from the MMIO register. Drivers could instead use “membar #Sync”, but
the lighter-weight “membar #MemIssue” is preferred in cases where the additional
function of the more-expensive “membar #Sync” are not required.
The “membar #Lookaside” is a lighter-weight version of “membar #MemIssue”,
which is useful when writing to a given MMIO register affects the value that will next
v2023.06.11a
be read from that register. However, the heavier-weight “membar #MemIssue” must
be used when a write to a given MMIO register affects the value that will next be read
from some other MMIO register.
SPARC requires a flush instruction be used between the time that the instruction
stream is modified and the time that any of these instructions are executed [SPA94].
This is needed to flush any prior value for that location from the SPARC’s instruction
cache. Note that flush takes an address, and will flush only that address from the
instruction cache. On SMP systems, all CPUs’ caches are flushed, but there is no
convenient way to determine when the off-CPU flushes complete, though there is a
reference to an implementation note.
But again, the Linux kernel runs SPARC in TSO mode, so all of the above membar
variants are strictly of historical interest. In particular, the smp_mb() primitive only
needs to use #StoreLoad because the other three reorderings are prohibited by TSO.
15.5.8 x86
Historically, the x86 CPUs provided “process ordering” so that all CPUs agreed on the
order of a given CPU’s writes to memory. This allowed the smp_wmb() primitive to
be a no-op for the CPU [Int04b]. Of course, a compiler directive was also required to
prevent optimizations that would reorder across the smp_wmb() primitive. In ancient
times, certain x86 CPUs gave no ordering guarantees for loads, so the smp_mb() and
smp_rmb() primitives expanded to lock;addl. This atomic instruction acts as a
barrier to both loads and stores.
But those were ancient times. More recently, Intel has published a memory model
for x86 [Int07]. It turns out that Intel’s modern CPUs enforce tighter ordering than
was claimed in the previous specifications, so this model simply mandates this modern
behavior. Even more recently, Intel published an updated memory model for x86 [Int11,
Section 8.2], which mandates a total global order for stores, although individual CPUs
are still permitted to see their own stores as having happened earlier than this total
global order would indicate. This exception to the total ordering is needed to allow
important hardware optimizations involving store buffers. In addition, x86 provides
other-multicopy atomicity, for example, so that if CPU 0 sees a store by CPU 1, then
CPU 0 is guaranteed to see all stores that CPU 1 saw prior to its store. Software may
use atomic operations to override these hardware optimizations, which is one reason
that atomic operations tend to be more expensive than their non-atomic counterparts.
It is also important to note that atomic instructions operating on a given memory
location should all be of the same size [Int16, Section 8.1.2.2]. For example, if you write
a program where one CPU atomically increments a byte while another CPU executes a
4-byte atomic increment on that same location, you are on your own.
Some SSE instructions are weakly ordered (clflush and non-temporal move
instructions [Int04a]). Code that uses these non-temporal move instructions can
also use mfence for smp_mb(), lfence for smp_rmb(), and sfence for smp_wmb().
A few older variants of the x86 CPU have a mode bit that enables out-of-order stores,
and for these CPUs, smp_wmb() must also be defined to be lock;addl.
Although newer x86 implementations accommodate self-modifying code without
any special instructions, to be fully compatible with past and potential future x86
implementations, a given CPU must execute a jump instruction or a serializing instruction
(e.g., cpuid) between modifying the code and executing it [Int11, Section 8.1.3].
v2023.06.11a
15.5.9 z Systems
The z Systems machines make up the IBM mainframe family, previously known as
the 360, 370, 390 and zSeries [Int04c]. Parallelism came late to z Systems, but
given that these mainframes first shipped in the mid 1960s, this is not saying much.
The “bcr 15,0” instruction is used for the Linux smp_mb() primitives, but compiler
constraints suffices for both the smp_rmb() and smp_wmb() primitives. It also has
strong memory-ordering semantics, as shown in Table 15.5. In particular, all CPUs will
agree on the order of unrelated stores from different CPUs, that is, the z Systems CPU
family is fully multicopy atomic, and is the only commercially available system with
this property.
As with most CPUs, the z Systems architecture does not guarantee a cache-coherent
instruction stream, hence, self-modifying code must execute a serializing instruction
between updating the instructions and executing them. That said, many actual z Systems
machines do in fact accommodate self-modifying code without serializing instruc-
tions. The z Systems instruction set provides a large set of serializing instructions,
including compare-and-swap, some types of branches (for example, the aforementioned
“bcr 15,0” instruction), and test-and-set.
15.5.10 Hardware Specifics: Discussion

There is considerable variation among these CPU families, and this section only scratched
the surface of a few families that are either heavily used or historically significant. Those
wishing more detail are invited to consult the reference manuals.
But a big benefit of the Linux-kernel memory model is that you can ignore these
details when writing architecture-independent Linux-kernel code.
15.6 Memory-Model Intuitions

Almost all people are intelligent. It is method that
they lack.
F. W. Nichol
This section revisits Table 15.3 and Section 15.1.3, summarizing the intervening
discussion with some appeals to transitive intuitions and with more sophisticated rules
of thumb.
But first, it is necessary to review the temporal and non-temporal nature of communi-
cation from one thread to another when using memory as the communications medium,
as was discussed in detail in Section 15.2.7. The key point is that although loads and
stores are conceptually simple, on real multicore hardware significant periods of time
are required for their effects to become visible to all other threads.
The simple and intuitive case occurs when one thread loads a value that some other
thread stored. This straightforward cause-and-effect case exhibits temporal behavior, so
that the software can safely assume that the store instruction completed before the load
instruction started. In real life, the load instruction might well have started quite some
time before the store instruction did, but all modern hardware must carefully hide such
cases from the software. Software will thus see the expected temporal cause-and-effect
behavior when one thread loads a value that some other thread stores, as discussed in
Section 15.2.7.3.
v2023.06.11a
15.6. MEMORY-MODEL INTUITIONS 563
Time
CPU 0
Section
Critical
Before
Lock
Section
Critical
CPU 1
Unlock
Section
Critical
Before
Section
Critical
After
Lock
Section
Critical
CPU 2
Unlock
Section
Critical
Before
Section
Critical
After
Lock
Section
Critical
Unlock
Section
Critical
Figure 15.24: Locking Intuitions After
This temporal behavior provides the basis for the next section’s transitive intuitions.
15.6.1 Transitive Intuitions

This section summarizes intuitions regarding single threads or variables, locking,
release-acquire chains, RCU, and fully ordered code.
15.6.1.1 Singular Intuitive Bliss

A program that has only one variable or only one thread will see all accesses in
order. There is quite a bit of code that can attain adequate performance when running
single-threaded on modern computer systems, but this book is primarily about software
that needs multiple CPUs. On, then, to the next section.
15.6.1.2 Locking Intuitions

Another transitive intuition involves that much-maligned workhorse, locking, described
in more detail in Section 15.4.2, to say nothing of Chapter 7. This section contains a
graphical description followed by a verbal description.
The graphical description is shown in Figure 15.24, which shows a lock being acquired
and released by CPUs 0, 1, and 2 in that order. The solid black arrows depict the
v2023.06.11a
unlock-lock ordering. The dotted lines emanating from them to the wide green arrows
show the effects on ordering. In particular:
1. The fact that CPU 0’s unlock precedes CPU 1’s lock ensures that any access
executed by CPU 0 within or before its critical section will be seen by accesses
executed by CPU 1 within and after its critical section.
In short, lock-based ordering is transitive through CPUs 0, 1, and 2. A key point
is that this ordering extends beyond the critical sections, so that everything before an
earlier lock release is seen by everything after a later lock acquisition.
For those who prefer words to diagrams, code holding a given lock will see the
accesses in all prior critical sections for that same lock, transitively. And if such code
sees the accesses in a given critical section, it will also see the accesses in all of that
CPU’s code preceding that critical section. In other words, when a CPU releases a given
lock, all of that lock’s subsequent critical sections will see the accesses in all of that
CPU’s code preceding that lock release.
Inversely, code holding a given lock will be protected from seeing the accesses in any
subsequent critical sections for that same lock, again, transitively. And if such code is
protected against seeing the accesses in a given critical section, it will also be protected
against seeing the accesses in all of that CPU’s code following that critical section.
In other words, when a CPU acquires a given lock, all of that lock’s previous critical
sections will be protected from seeing the accesses in all of that CPU’s code following
that lock acquisition.
But what does it mean to “see accesses” and exactly what accesses are seen?
To start, an access is either a load or a store, possibly occurring as part of a
read-modify-write operation.
If a CPU’s code prior to its release of a given lock contains an access A to a given
variable, then for an access B to that same variable contained in any CPU’s code
following a later acquisition of that same lock:
1. If A and B are both loads, then B will return either the same value that A did or
some later value.
2. If A is a load and B is a store, then B will overwrite either the value loaded by A or
some later value.
3. If A is a store and B is a load, then B will return either the value stored by A or
some later value.
4. If A and B are both stores, then B will overwrite either the value stored by A or
some later value.
Here, “some later value” is shorthand for “the value stored by some intervening
access”.
Locking is strongly intuitive, which is one reason why it has survived so many
attempts to eliminate it. This is also one reason why you should use it where it applies.
v2023.06.11a
Time
CPU 0
Release
Before
Release A
Release
After
CPU 1
Acquire
Before
Acquire A
CPU 2
Release B
Acquire
Before
Release
After
Acquire B
Acquire
After
Figure 15.25: Release-Acquire Intuitions
15.6.1.3 Release-Acquire Intuitions

Release-acquire chains also behave in a transitively intuitive manner not unlike that
of locking. This section also contains a graphical description followed by a verbal
description.
The graphical description is shown in Figure 15.25, which shows a release-acquire
chain extending through CPUs 0, 1, and 2. The solid black arrows depict the release-
acquire ordering. The dotted lines emanating from them to the wide green arrows show
the effects on ordering.
1. The fact that CPU 0’s release of A is read by CPU 1’s acquire of A ensures that
any accesses executed by CPU 0 prior to its release will be seen by any accesses
executed by CPU 1 after its acquire.
2. The fact that CPU 1’s release of B is read by CPU 2’s acquire of B ensures that
any accesses executed by CPU 1 prior to its release will be seen by any accesses
executed by CPU 2 after its acquire.
3. Note also that CPU 0’s release of A is read by CPU 1’s acquire of A, which
precedes CPU 1’s release of B, which is read by CPU 2’s acquire of B. Taken
together, all this ensures that any accesses executed by CPU 0 prior to its release
will be seen by any accesses executed by CPU 2 after its acquire.
This illustrates that properly constructed release-acquire ordering is transitive through

CPUs 0, 1, and 2, and in fact may be extended through as many CPUs as needed.23
23 But please note that stray stores to either A or B can break the release-acquire chain, as
illustrated by Listing 15.27.
v2023.06.11a
For those who prefer words to diagrams, when an acquire loads the value stored by a
release, discussed in Section 15.2.7.4, then the code following that release will see all
accesses preceding the acquire. More precisely, if CPU 0 does an acquire that loads the
value stored by CPU 1’s release, than all the subsequent accesses executed by CPU 0
will see the all of CPU 1’s accesses prior to its release.
Similarly, the accesses preceding that release access will be protected from seeing
the accesses following the acquire access. (More precision is left as an exercise to the
reader.)
Releases and acquires can be chained, for example CPU 0’s release stores the value
loaded by CPU 1’s acquire, a later release by CPU 1 stores the value loaded by CPU 2’s
acquire, and so on. The accesses following a given acquire will see the accesses
preceding each prior release in the chain, and, inversely, the accesses preceding a given
release will be protected from seeing the accesses following each later acquire in the
chain. Some long-chain examples are illustrated by Listings 15.22, 15.23, and 15.24.
The seeing and not seeing of accesses works the same way as described in Sec-
tion 15.6.1.2.
However, as illustrated by Listing 15.27, the acquire access must load exactly what
was stored by the release access. Any intervening store that is not itself part of that
same release-acquire chain will break the chain.
Nevertheless, properly constructed release-acquire chains are transitive, intuitive, and
useful.
15.6.1.4 RCU Intuitions

As noted in Section 9.5.2 on page 228, RCU provides a number of ordering guarantees.
The first is the publish-subscribe mechanism described in Section 9.5.2.1 on page 228.
This resembles the acquire-release chains discussed in the previous section, but sub-
stitutes a member of the rcu_dereference() family of primitives for the smp_
load_acquire(). Unlike smp_load_acquire(), the ordering implied by rcu_
dereference() applies only to subsequent accesses that dereference the pointer
returned by that rcu_dereference() as shown in Figure 9.10 on page 229.
The second guarantee says that if any part of an RCU read-side critical section
precedes the beginning of a grace period, then the entirety of that critical section
precedes the end of that grace period, as shown in Figure 9.11 on page 231.
The third guarantee says that if any part of an RCU read-side critical section follows
the end of a grace period, then the entirety of that critical section follows the beginning
of that grace period, as shown in Figure 9.12 on page 232.
Both of these two guarantees are discussed in Section 9.5.2.2 on page 230, with more
examples shown in Figures 9.13 and 9.14 on pages 233 and 234. These two guarantees
have further version-maintenance consequences that are discussed in Section 9.5.2.3 on
page 235.
These guarantees are discussed somewhat more formally in Section 15.4.3.
Much of the sophistication of RCU lies not in its guarantees, but in its use cases,
which are the subject of Section 9.5.4 starting on page 251.
15.6.1.5 Fully Ordered Intuitions

A more extreme example of transitivity places at least one smp_mb() between each pair
of accesses. All accesses seen by any given access will also be seen by all later accesses.
v2023.06.11a
The resulting program will be fully ordered, if somewhat slow. Such programs will
be sequentially consistent and much loved by formal-verification experts who specialize
in tried-and-true 1980s proof techniques. But slow or not, smp_mb() is always there
when you need it!
Nevertheless, there are situations that cannot be addressed by these intuitive ap-
proaches. The next section therefore presents a more complete, if less transitive, set of
rules of thumb.
15.6.2 Rules of Thumb

The transitive intuitions presented in the previous section are very appealing, at least
as memory models go. Unfortunately, hardware is under no obligation to provide
temporal cause-and-effect illusions when one thread’s store overwrites a value either
loaded or stored by some other thread. It is quite possible that, from the software’s
viewpoint, an earlier store will overwrite a later store’s value, but only if those two
stores were executed by different threads, as illustrated by Figure 15.13. Similarly, a
later load might well read a value overwritten by an earlier store, but again only if that
load and store were executed by different threads, as illustrated by Figure 15.12. This
counter-intuitive behavior occurs due to the need to buffer stores in order to achieve
adequate performance, as discussed in Section 15.2.7.2.
As a result, situations where one thread reads a value written by some other thread
can make do with far weaker ordering than can situations where one thread overwrites
a value loaded or stored by some other thread. These differences are captured by the
following rules of thumb.
The first rule of thumb is that memory-ordering operations are only required where
there is a possibility of interaction between at least two variables shared among at least two
threads, which underlies the singular intuitive bliss presented in Section 15.6.1.1. In light
of the intervening material, this single sentence encapsulates much of Section 15.1.3’s
basic rules of thumb, for example, keeping in mind that “memory-barrier pairing” is
a two-thread special case of “cycle”. And, as always, if a single-threaded program
will provide sufficient performance, why bother with parallelism?24 After all, avoiding
parallelism also avoids the added cost and complexity of memory-ordering operations.
The second rule of thumb involves load-buffering situations: If all thread-to-thread
communication in a given cycle use store-to-load links (that is, the next thread’s load
returns the value stored by the previous thread), minimal ordering suffices. Minimal
ordering includes dependencies and acquires as well as all stronger ordering operations.
Because a lock acquisition must load the lock-word value stored by any prior release of
that lock, this rule of thumb underlies the locking intuitions presented in Section 15.6.1.2.
The third rule of thumb involves release-acquire chains: If all but one of the links
in a given cycle is a store-to-load link, it is sufficient to use release-acquire pairs for
each of those store-to-load links, as illustrated by Listings 15.23 and 15.24. This rule
underlies the release-acquire intuitions presented in Section 15.6.1.3.
You can replace a given acquire with a dependency in environments permitting
this, keeping in mind that the C11 standard’s memory model does not fully respect
dependencies. Therefore, a dependency leading to a load must be headed by a READ_
ONCE() or an rcu_dereference(): A plain C-language load is not sufficient. In
addition, carefully review Sections 15.3.2 and 15.3.3, because a dependency broken by
24 Hobbyists and researchers should of course feel free to ignore this and many other
cautions.
v2023.06.11a
your compiler will not order anything. The two threads sharing the sole non-store-to-
load link can sometimes substitute WRITE_ONCE() plus smp_wmb() for smp_store_
release() on the one hand, and READ_ONCE() plus smp_rmb() for smp_load_
acquire() on the other. However, the wise developer will check such substitutions
carefully, for example, using the herd tool as described in Section 12.3.
Quick Quiz 15.48: Why is it necessary to use heavier-weight ordering for load-to-store and
store-to-store links, but not for store-to-load links? What on earth makes store-to-load links so
special???
The fourth and final rule of thumb identifies where full memory barriers (or stronger)
are required: If a given cycle contains two or more non-store-to-load links (that is,
a total of two or more links that are either load-to-store or store-to-store links), you
will need at least one full barrier between each pair of non-store-to-load links in that
cycle, as illustrated by Listing 15.19 as well as in the answer to Quick Quiz 15.25.
Full barriers include smp_mb(), successful full-strength non-void atomic RMW
operations, and other atomic RMW operations in conjunction with either smp_mb__
before_atomic() or smp_mb__after_atomic(). Any of RCU’s grace-period-wait
primitives (synchronize_rcu() and friends) also act as full barriers, but at far greater
expense than smp_mb(). With strength comes expense, though full barriers usually
hurt performance more than they hurt scalability. The extreme logical endpoint of this
rule of thumb underlies the fully ordered intuitions presented in Section 15.6.1.5.
Recapping the rules:
1. Memory-ordering operations are required only if at least two variables are shared
by at least two threads.
2. If all links in a cycle are store-to-load links, then minimal ordering suffices.
3. If all but one of the links in a cycle are store-to-load links, then each store-to-load
link may use a release-acquire pair.
4. Otherwise, at least one full barrier is required between each pair of non-store-to-load
links.
Note that an architecture is permitted to provide stronger guarantees, as discussed

in Section 15.5, but these guarantees may only be relied upon in code that runs only
for that architecture. In addition, more accurate memory models [AMM+ 18] may give
stronger guarantees with lower-overhead operations than do these rules of thumb, albeit
at the expense of greater complexity. In these more formal memory-ordering papers,
a store-to-load link is an example of a reads-from (rf) link, a load-to-store link is an
example of a from-reads (fr) link, and a store-to-store link is an example of a coherence
(co) link.
One final word of advice: Use of raw memory-ordering primitives is a last resort. It
is almost always better to use existing primitives, such as locking or RCU, thus letting
those primitives do the memory ordering for you.
v2023.06.11a
Creating a perfect API is like committing the perfect
crime. There are at least fifty things that can go
wrong, and if you are a genius, you might be able to
anticipate twenty-five of them.
With apologies to any Kathleen Turner fans who
might still be alive.
Chapter 16
Ease of Use
16.1 What is Easy?
When someone says “I want a programming

language in which I need only say what I wish done,”
give them a lollipop.
Alan J. Perlis, updated
If you are tempted to look down on ease-of-use requirements, please consider that an
ease-of-use bug in Linux-kernel RCU resulted in an exploitable Linux-kernel security
bug in a use of RCU [McK19a]. It is therefore clearly important that even in-kernel
APIs be easy to use.
Unfortunately, “easy” is a relative term. For example, many people would consider
a 15-hour airplane flight to be a bit of an ordeal—unless they stopped to consider
alternative modes of transportation, especially swimming. This means that creating an
easy-to-use API requires that you understand your intended users well enough to know
what is easy for them. Which might or might not have anything to do with what is easy
for you.
The following question illustrates this point: “Given a randomly chosen person among
everyone alive today, what one change would improve that person’s life?”
There is no single change that would be guaranteed to help everyone’s life. After all,
there is an extremely wide range of people, with a correspondingly wide range of needs,
wants, desires, and aspirations. A starving person might need food, but additional food
might well hasten the death of a morbidly obese person. The high level of excitement so
fervently desired by many young people might well be fatal to someone recovering from
a heart attack. Information critical to the success of one person might contribute to the
failure of someone suffering from information overload. In short, if you are working on
a software project that is intended to help people you know nothing about, you should
not be surprised when those people find fault with your project.
If you really want to help a given group of people, there is simply no substitute for
working closely with them over an extended period of time, as in years. Nevertheless,
there are some simple things that you can do to increase the odds of your users being
happy with your software, and some of these things are covered in the next section.
569
v2023.06.11a
570 CHAPTER 16. EASE OF USE
16.2 Rusty Scale for API Design

Finding the appropriate measurement is thus not a
mathematical exercise. It is a risk-taking judgment.
Peter Drucker
This section is adapted from portions of Rusty Russell’s 2003 Ottawa Linux Symposium
keynote address [Rus03, Slides 39–57]. Rusty’s key point is that the goal should not
be merely to make an API easy to use, but rather to make the API hard to misuse.
To that end, Rusty proposed his “Rusty Scale” in decreasing order of this important
hard-to-misuse property.
The following list attempts to generalize the Rusty Scale beyond the Linux kernel:
1. It is impossible to get wrong. Although this is the standard to which all API
designers should strive, only the mythical dwim()1 command manages to come
close.
2. The compiler or linker won’t let you get it wrong.
3. The compiler or linker will warn you if you get it wrong. BUILD_BUG_ON() is
your users’ friend.
4. The simplest use is the correct one.
5. The name tells you how to use it. But names can be two-edged swords. Although
rcu_read_lock() is plain enough for someone converting code from reader-
writer locking, it might cause some consternation for someone converting code
from reference counting.
6. Do it right or it will always break at runtime. WARN_ON_ONCE() is your users’
friend.
7. Follow common convention and you will get it right. The malloc() library
function is a good example. Although it is easy to get memory allocation wrong,
a great many projects do manage to get it right, at least most of the time. Using
malloc() in conjunction with Valgrind [The11] moves malloc() almost up to
the “do it right or it will always break at runtime” point on the scale.
8. Read the documentation and you will get it right.
9. Read the implementation and you will get it right.

10. Read the right mailing-list archive and you will get it right.
11. Read the right mailing-list archive and you will get it wrong.
12. Read the implementation and you will get it wrong. The original non-CONFIG_
PREEMPT implementation of rcu_read_lock() [McK07a] is an infamous exam-
ple of this point on the scale.
1 The dwim() function is an acronym that expands to “do what I mean”.
v2023.06.11a
16.3. SHAVING THE MANDELBROT SET 571
13. Read the documentation and you will get it wrong. For example, the DEC Alpha
wmb instruction’s documentation [Cor02] fooled a number of developers into
thinking that this instruction had much stronger memory-order semantics than it
actually does. Later documentation clarified this point [Com01, Pug00], moving
the wmb instruction up to the “read the documentation and you will get it right”
point on the scale.
14. Follow common convention and you will get it wrong. The printf() statement
is an example of this point on the scale because developers almost always fail to
check printf()’s error return.
15. Do it right and it will break at runtime.

16. The name tells you how not to use it.
17. The obvious use is wrong. The Linux kernel smp_mb() function is an example
of this point on the scale. Many developers assume that this function has much
stronger ordering semantics than it actually possesses. Chapter 15 contains the
information needed to avoid this mistake, as does the Linux-kernel source tree’s
Documentation and tools/memory-model directories.
18. The compiler or linker will warn you if you get it right.
19. The compiler or linker won’t let you get it right.
20. It is impossible to get right. The gets() function is a famous example of this point
on the scale. In fact, gets() can perhaps best be described as an unconditional
buffer-overflow security hole.
16.3 Shaving the Mandelbrot Set

Simplicity does not precede complexity,
but follows it.
Alan J. Perlis
The set of useful programs resembles the Mandelbrot set (shown in Figure 16.1) in
that it does not have a clear-cut smooth boundary—if it did, the halting problem would
be solvable. But we need APIs that real people can use, not ones that require a Ph.D.
dissertation be completed for each and every potential use. So, we “shave the Mandelbrot
set”,2 restricting the use of the API to an easily described subset of the full set of
potential uses.
Such shaving may seem counterproductive. After all, if an algorithm works, why
shouldn’t it be used?
To see why at least some shaving is absolutely necessary, consider a locking design
that avoids deadlock, but in perhaps the worst possible way. This design uses a circular
doubly linked list, which contains one element for each thread in the system along with
a header element. When a new thread is spawned, the parent thread must insert a new
element into this list, which requires some sort of synchronization.
2 Due to Josh Triplett.
v2023.06.11a
Figure 16.1: Mandelbrot Set (Courtesy of Wikipedia)
One way to protect the list is to use a global lock. However, this might be a bottleneck
if threads were being created and deleted frequently.3 Another approach would be to
use a hash table and to lock the individual hash buckets, but this can perform poorly
when scanning the list in order.
A third approach is to lock the individual list elements, and to require the locks for
both the predecessor and successor to be held during the insertion. Since both locks
must be acquired, we need to decide which order to acquire them in. Two conventional
approaches would be to acquire the locks in address order, or to acquire them in the
order that they appear in the list, so that the header is always acquired first when it is
one of the two elements being locked. However, both of these methods require special
checks and branches.
The to-be-shaven solution is to unconditionally acquire the locks in list order. But
what about deadlock?
Deadlock cannot occur.
To see this, number the elements in the list starting with zero for the header up to
𝑁 for the last element in the list (the one preceding the header, given that the list is
circular). Similarly, number the threads from zero to 𝑁 − 1. If each thread attempts to
lock some consecutive pair of elements, at least one of the threads is guaranteed to be
able to acquire both locks.
Why?
Because there are not enough threads to reach all the way around the list. Suppose
thread 0 acquires element 0’s lock. To be blocked, some other thread must have already
acquired element 1’s lock, so let us assume that thread 1 has done so. Similarly, for
thread 1 to be blocked, some other thread must have acquired element 2’s lock, and so
on, up through thread 𝑁 − 1, who acquires element 𝑁 − 1’s lock. For thread 𝑁 − 1 to be
blocked, some other thread must have acquired element 𝑁’s lock. But there are no more
threads, and so thread 𝑁 − 1 cannot be blocked. Therefore, deadlock cannot occur.
So why should we prohibit use of this delightful little algorithm?
The fact is that if you really want to use it, we cannot stop you. We can, however,
recommend against such code being included in any project that we care about.
But, before you use this algorithm, please think through the following Quick Quiz.
Quick Quiz 16.1: Can a similar algorithm be used when deleting elements?
3 Those of you with strong operating-system backgrounds, please suspend disbelief.
Those unable to suspend disbelief are encouraged to provide better examples.
v2023.06.11a
16.3. SHAVING THE MANDELBROT SET 573
Figure 16.2: Shaving the Mandelbrot Set
The fact is that this algorithm is extremely specialized (it only works on certain sized
lists), and also quite fragile. Any bug that accidentally failed to add a node to the list
could result in deadlock. In fact, simply adding the node a bit too late could result in
deadlock, as could increasing the number of threads.
In addition, the other algorithms described above are “good and sufficient”. For
example, simply acquiring the locks in address order is fairly simple and quick, while
allowing the use of lists of any size. Just be careful of the special cases presented by
empty lists and lists containing only one element!
Quick Quiz 16.2: Yetch! What ever possessed someone to come up with an algorithm that
deserves to be shaved as much as this one does???
In summary, we do not use algorithms simply because they happen to work. We

instead restrict ourselves to algorithms that are useful enough to make it worthwhile
learning about them. The more difficult and complex the algorithm, the more generally
useful it must be in order for the pain of learning it and fixing its bugs to be worthwhile.
Quick Quiz 16.3: Give an exception to this rule.
Exceptions aside, we must continue to shave the software “Mandelbrot set” so that
our programs remain maintainable, as shown in Figure 16.2.
v2023.06.11a
v2023.06.11a
Prediction is very difficult, especially about the
future.
Niels Bohr
Chapter 17
Conflicting Visions of the Future
This chapter presents some conflicting visions of the future of parallel programming. It
is not clear which of these will come to pass, in fact, it is not clear that any of them will.
They are nevertheless important because each vision has its devoted adherents, and if
enough people believe in something fervently enough, you will need to deal with that
thing’s existence in the form of its influence on the thoughts, words, and deeds of its
adherents. Besides which, one or more of these visions will actually come to pass. But
most are bogus. Tell which is which and you’ll be rich [Spi77]!
Therefore, the following sections give an overview of transactional memory, hardware
transactional memory, formal verification in regression testing, and parallel functional
programming. But first, a cautionary tale on prognostication taken from the early 2000s.
17.1 The Future of CPU Technology Ain’t What it Used

to Be
A great future behind him.
David Maraniss
Years past always seem so simple and innocent when viewed through the lens of many
years of experience. And the early 2000s were for the most part innocent of the
impending failure of Moore’s Law to continue delivering the then-traditional increases
in CPU clock frequency. Oh, there were the occasional warnings about the limits
of technology, but such warnings had been sounded for decades. With that in mind,
consider the following scenarios:
1. Uniprocessor Über Alles (Figure 17.1),

2. Multithreaded Mania (Figure 17.2),
3. More of the Same (Figure 17.3), and
4. Crash Dummies Slamming into the Memory Wall (Figure 17.4).
5. Astounding Accelerators (Figure 17.5).
Each of these scenarios is covered in the following sections.
575
v2023.06.11a
576 CHAPTER 17. CONFLICTING VISIONS OF THE FUTURE
Figure 17.1: Uniprocessor Über Alles
Figure 17.2: Multithreaded Mania
v2023.06.11a
17.1. THE FUTURE OF CPU TECHNOLOGY AIN’T WHAT IT USED TO BE 577
Figure 17.3: More of the Same
Figure 17.4: Crash Dummies Slamming into the Memory Wall
v2023.06.11a
Figure 17.5: Astounding Accelerators
17.1.1 Uniprocessor Über Alles

As was said in 2004 [McK04]:
In this scenario, the combination of Moore’s-Law increases in CPU clock

rate and continued progress in horizontally scaled computing render SMP
systems irrelevant. This scenario is therefore dubbed “Uniprocessor Über
Alles”, literally, uniprocessors above all else.
These uniprocessor systems would be subject only to instruction overhead,
since memory barriers, cache thrashing, and contention do not affect single-
CPU systems. In this scenario, RCU is useful only for niche applications,
such as interacting with NMIs. It is not clear that an operating system lacking
RCU would see the need to adopt it, although operating systems that already
implement RCU might continue to do so.
However, recent progress with multithreaded CPUs seems to indicate that this
scenario is quite unlikely.
Unlikely indeed! But the larger software community was reluctant to accept the
fact that they would need to embrace parallelism, and so it was some time before this
community concluded that the “free lunch” of Moore’s-Law-induced CPU core-clock
frequency increases was well and truly finished. Never forget: Belief is an emotion, not
necessarily the result of a rational technical thought process!
v2023.06.11a
17.1.2 Multithreaded Mania

Also from 2004 [McK04]:
A less-extreme variant of Uniprocessor Über Alles features uniprocessors with

hardware multithreading, and in fact multithreaded CPUs are now standard
for many desktop and laptop computer systems. The most aggressively
multithreaded CPUs share all levels of cache hierarchy, thereby eliminating
CPU-to-CPU memory latency, in turn greatly reducing the performance
penalty for traditional synchronization mechanisms. However, a multithreaded
CPU would still incur overhead due to contention and to pipeline stalls caused
by memory barriers. Furthermore, because all hardware threads share all
levels of cache, the cache available to a given hardware thread is a fraction of
what it would be on an equivalent single-threaded CPU, which can degrade
performance for applications with large cache footprints. There is also some
possibility that the restricted amount of cache available will cause RCU-based
algorithms to incur performance penalties due to their grace-period-induced
additional memory consumption. Investigating this possibility is future work.
However, in order to avoid such performance degradation, a number of
multithreaded CPUs and multi-CPU chips partition at least some of the levels
of cache on a per-hardware-thread basis. This increases the amount of cache
available to each hardware thread, but re-introduces memory latency for
cachelines that are passed from one hardware thread to another.
And we all know how this story has played out, with multiple multi-threaded cores
on a single die plugged into a single socket, with varying degrees of optimization for
lower numbers of active threads per core. The question then becomes whether or not
future shared-memory systems will always fit into a single socket.
17.1.3 More of the Same

Again from 2004 [McK04]:
The More-of-the-Same scenario assumes that the memory-latency ratios will

remain roughly where they are today.
This scenario actually represents a change, since to have more of the same,
interconnect performance must begin keeping up with the Moore’s-Law
increases in core CPU performance. In this scenario, overhead due to pipeline
stalls, memory latency, and contention remains significant, and RCU retains
the high level of applicability that it enjoys today.
And the change has been the ever-increasing levels of integration that Moore’s Law is
still providing. But longer term, which will it be? More CPUs per die? Or more I/O,
cache, and memory?
Servers seem to be choosing the former, while embedded systems on a chip (SoCs)
continue choosing the latter.
17.1.4 Crash Dummies Slamming into the Memory Wall

And one more quote from 2004 [McK04]:
v2023.06.11a
10000
Instructions per Memory Reference Time

1000
100
10
0.1
82 84 86 88 90 92 94 96 98 00 02
Year
Figure 17.6: Instructions per Local Memory Reference for Sequent Computers
spinlock
Breakeven Update Fraction
RCU
0.1
1 10 100 1000
Memory-Latency Ratio
Figure 17.7: Breakevens vs. 𝑟, 𝜆 Large, Four CPUs
v2023.06.11a
spinlock
Breakeven Update Fraction

0.1
drw
0.01
0.001 RCU
0.0001
1 10 100 1000
Memory-Latency Ratio
Figure 17.8: Breakevens vs. 𝑟, 𝜆 Small, Four CPUs
If the memory-latency trends shown in Figure 17.6 continue, then memory

latency will continue to grow relative to instruction-execution overhead.
Systems such as Linux that have significant use of RCU will find additional
use of RCU to be profitable, as shown in Figure 17.7. As can be seen in this
figure, if RCU is heavily used, increasing memory-latency ratios give RCU
an increasing advantage over other synchronization mechanisms. In contrast,
systems with minor use of RCU will require increasingly high degrees of
read intensity for use of RCU to pay off, as shown in Figure 17.8. As can be
seen in this figure, if RCU is lightly used, increasing memory-latency ratios
put RCU at an increasing disadvantage compared to other synchronization
mechanisms. Since Linux has been observed with over 1,600 callbacks per
grace period under heavy load [SM04b], it seems safe to say that Linux falls
into the former category.
On the one hand, this passage failed to anticipate the cache-warmth issues that
RCU can suffer from in workloads with significant update intensity, in part because it
seemed unlikely that RCU would really be used for such workloads. In the event, the
SLAB_TYPESAFE_BY_RCU has been pressed into service in a number of instances where
these cache-warmth issues would otherwise be problematic, as has sequence locking.
On the other hand, this passage also failed to anticipate that RCU would be used to
reduce scheduling latency or for security.
Much of the data generated for this book was collected on an eight-socket system
with 28 cores per socket and two hardware threads per core, for a total of 448 hardware
threads. The idle-system memory latencies are less than one microsecond, which are
no worse than those of similar-sized systems of the year 2004. Some claim that these
latencies approach a microsecond only because of the x86 CPU family’s relatively
strong memory ordering, but it may be some time before that particular argument is
settled.
v2023.06.11a
17.1.5 Astounding Accelerators

The potential of hardware accelerators was not quite as clear in 2004 as it is in 2021, so
this section has no quote. However, the November 2020 Top 500 list [MDSS20] features
a great many accelerators, so one could argue that this section is a view of the present
rather than of the future. The same could be said of most of the preceding sections.
Hardware accelerators are being put to many other uses, including encryption,
compression, machine learning.
In short, beware of prognostications, including those in the remainder of this chapter.
17.2 Transactional Memory

Everything should be as simple as it can be, but not
simpler.
Albert Einstein, by way of Louis Zukofsky
The idea of using transactions outside of databases goes back many decades [Lom77,
Kni86, HM93], with the key difference between database and non-database transactions
being that non-database transactions drop the “D” in the “ACID”1 properties defining
database transactions. The idea of supporting memory-based transactions, or “transac-
tional memory” (TM), in hardware is more recent [HM93], but unfortunately, support
for such transactions in commodity hardware was not immediately forthcoming, despite
other somewhat similar proposals being put forward [SSHT93]. Not long after, Shavit
and Touitou proposed a software-only implementation of transactional memory (STM)
that was capable of running on commodity hardware, give or take memory-ordering
issues [ST95]. This proposal languished for many years, perhaps due to the fact that the
research community’s attention was absorbed by non-blocking synchronization (see
Section 14.2).
But by the turn of the century, TM started receiving more attention [MT01, RG01],
and by the middle of the decade, the level of interest can only be termed “incandes-
cent” [Her05, Gro07], with only a few voices of caution [BLM05, MMW07].
The basic idea behind TM is to execute a section of code atomically, so that other
threads see no intermediate state. As such, the semantics of TM could be implemented
by simply replacing each transaction with a recursively acquirable global lock acquisition
and release, albeit with abysmal performance and scalability. Much of the complexity
inherent in TM implementations, whether hardware or software, is efficiently detecting
when concurrent transactions can safely run in parallel. Because this detection is done
dynamically, conflicting transactions can be aborted or “rolled back”, and in some
implementations, this failure mode is visible to the programmer.
Because transaction roll-back is increasingly unlikely as transaction size decreases,
TM might become quite attractive for small memory-based operations, such as linked-list
manipulations used for stacks, queues, hash tables, and search trees. However, it is
currently much more difficult to make the case for large transactions, particularly those
containing non-memory operations such as I/O and process creation. The following
sections look at current challenges to the grand vision of “Transactional Memory
Everywhere” [McK09b]. Section 17.2.1 examines the challenges faced interacting
with the outside world, Section 17.2.2 looks at interactions with process modification
1 Atomicity, consistency, isolation, and durability.
v2023.06.11a
17.2. TRANSACTIONAL MEMORY 583
primitives, Section 17.2.3 explores interactions with other synchronization primitives,

and finally Section 17.2.4 closes with some discussion.
17.2.1 Outside World

In the wise words of Donald Knuth:
Many computer users feel that input and output are not actually part of “real
programming,” they are merely things that (unfortunately) must be done in
order to get information in and out of the machine.
Whether or not we believe that input and output are “real programming”, the fact
is that software absolutely must deal with the outside world. This section therefore
critiques transactional memory’s outside-world capabilities, focusing on I/O operations,
time delays, and persistent storage.
17.2.1.1 I/O Operations

One can execute I/O operations within a lock-based critical section, while holding a
hazard pointer, within a sequence-locking read-side critical section, and from within
a userspace-RCU read-side critical section, and even all at the same time, if need be.
What happens when you attempt to execute an I/O operation from within a transaction?
The underlying problem is that transactions may be rolled back, for example, due
to conflicts. Roughly speaking, this requires that all operations within any given
transaction be revocable, so that executing the operation twice has the same effect as
executing it once. Unfortunately, I/O is in general the prototypical irrevocable operation,
making it difficult to include general I/O operations in transactions. In fact, general
I/O is irrevocable: Once you have pushed the proverbial button launching the nuclear
warheads, there is no turning back.
Here are some options for handling of I/O within transactions:
1. Restrict I/O within transactions to buffered I/O with in-memory buffers. These
buffers may then be included in the transaction in the same way that any other
memory location might be included. This seems to be the mechanism of choice,
and it does work well in many common cases of situations such as stream I/O
and mass-storage I/O. However, special handling is required in cases where
multiple record-oriented output streams are merged onto a single file from multiple
processes, as might be done using the “a+” option to fopen() or the O_APPEND
flag to open(). In addition, as will be seen in the next section, common networking
operations cannot be handled via buffering.
2. Prohibit I/O within transactions, so that any attempt to execute an I/O operation
aborts the enclosing transaction (and perhaps multiple nested transactions). This
approach seems to be the conventional TM approach for unbuffered I/O, but
requires that TM interoperate with other synchronization primitives tolerating I/O.
3. Prohibit I/O within transactions, but enlist the compiler’s aid in enforcing this
prohibition.
4. Permit only one special irrevocable transaction [SMS08] to proceed at any given
time, thus allowing irrevocable transactions to contain I/O operations.2 This works
2 In earlier literature, irrevocable transactions are termed inevitable transactions.
v2023.06.11a
in general, but severely limits the scalability and performance of I/O operations.
Given that scalability and performance is a first-class goal of parallelism, this
approach’s generality seems a bit self-limiting. Worse yet, use of irrevocability to
tolerate I/O operations seems to greatly restrict use of manual transaction-abort
operations.3 Finally, if there is an irrevocable transaction manipulating a given
data item, any other transaction manipulating that same data item cannot have
non-blocking semantics.
5. Create new hardware and protocols such that I/O operations can be pulled into the
transactional substrate. In the case of input operations, the hardware would need
to correctly predict the result of the operation, and to abort the transaction if the
prediction failed.
I/O operations are a well-known weakness of TM, and it is not clear that the problem
of supporting I/O in transactions has a reasonable general solution, at least if “reasonable”
is to include usable performance and scalability. Nevertheless, continued time and
attention to this problem will likely produce additional progress.
17.2.1.2 RPC Operations

One can execute RPCs within a lock-based critical section, while holding a hazard pointer,
within a sequence-locking read-side critical section, and from within a userspace-RCU
read-side critical section, and even all at the same time, if need be. What happens when
you attempt to execute an RPC from within a transaction?
If both the RPC request and its response are to be contained within the transaction,
and if some part of the transaction depends on the result returned by the response, then
it is not possible to use the memory-buffer tricks that can be used in the case of buffered
I/O. Any attempt to take this buffering approach would deadlock the transaction, as the
request could not be transmitted until the transaction was guaranteed to succeed, but the
transaction’s success might not be knowable until after the response is received, as is
the case in the following example:
1 begin_trans();
2 rpc_request();
3 i = rpc_response();
4 a[i]++;
5 end_trans();
The transaction’s memory footprint cannot be determined until after the RPC
response is received, and until the transaction’s memory footprint can be determined, it
is impossible to determine whether the transaction can be allowed to commit. The only
action consistent with transactional semantics is therefore to unconditionally abort the
transaction, which is, to say the least, unhelpful.
Here are some options available to TM:
1. Prohibit RPC within transactions, so that any attempt to execute an RPC operation
aborts the enclosing transaction (and perhaps multiple nested transactions). Alter-
natively, enlist the compiler to enforce RPC-free transactions. This approach does
work, but will require TM to interact with other synchronization primitives.
3 This difficulty was pointed out by Michael Factor. To see the problem, think through
what TM should do in response to an attempt to abort a transaction after it has executed an

irrevocable operation.
v2023.06.11a
2. Permit only one special irrevocable transaction [SMS08] to proceed at any given
time, thus allowing irrevocable transactions to contain RPC operations. This works
in general, but severely limits the scalability and performance of RPC operations.
Given that scalability and performance is a first-class goal of parallelism, this
approach’s generality seems a bit self-limiting. Furthermore, use of irrevocable
transactions to permit RPC operations restricts manual transaction-abort operations
once the RPC operation has started. Finally, if there is an irrevocable transaction
manipulating a given data item, any other transaction manipulating that same data
item must have blocking semantics.
3. Identify special cases where the success of the transaction may be determined
before the RPC response is received, and automatically convert these to irrevocable
transactions immediately before sending the RPC request. Of course, if several
concurrent transactions attempt RPC calls in this manner, it might be necessary
to roll all but one of them back, with consequent degradation of performance
and scalability. This approach nevertheless might be valuable given long-running
transactions ending with an RPC. This approach must still restrict manual
transaction-abort operations.
4. Identify special cases where the RPC response may be moved out of the transaction,
and then proceed using techniques similar to those used for buffered I/O.
5. Extend the transactional substrate to include the RPC server as well as its client. This
is in theory possible, as has been demonstrated by distributed databases. However,
it is unclear whether the requisite performance and scalability requirements can be
met by distributed-database techniques, given that memory-based TM has no slow
disk drives behind which to hide such latencies. Of course, given the advent of
solid-state disks, it is also quite possible that databases will need to redesign their
approach to latency hiding.
As noted in the prior section, I/O is a known weakness of TM, and RPC is simply an
especially problematic case of I/O.
17.2.1.3 Time Delays

An important special case of interaction with extra-transactional accesses involves
explicit time delays within a transaction. Of course, the idea of a time delay within a
transaction flies in the face of TM’s atomicity property, but this sort of thing is arguably
what weak atomicity is all about. Furthermore, correct interaction with memory-mapped
I/O sometimes requires carefully controlled timing, and applications often use time
delays for varied purposes. Finally, one can execute time delays within a lock-based
critical section, while holding a hazard pointer, within a sequence-locking read-side
critical section, and from within a userspace-RCU read-side critical section, and even all
at the same time, if need be. Doing so might not be wise from a contention or scalability
viewpoint, but then again, doing so does not raise any fundamental conceptual issues.
So, what can TM do about time delays within transactions?
1. Ignore time delays within transactions. This has an appearance of elegance, but
like too many other “elegant” solutions, fails to survive first contact with legacy
code. Such code, which might well have important time delays in critical sections,
would fail upon being transactionalized.
v2023.06.11a
2. Abort transactions upon encountering a time-delay operation. This is attractive,

but it is unfortunately not always possible to automatically detect a time-delay
operation. Is that tight loop carrying out a critical computation, or is it simply
waiting for time to elapse?
3. Enlist the compiler to prohibit time delays within transactions.
4. Let the time delays execute normally. Unfortunately, some TM implementations

publish modifications only at commit time, which could defeat the purpose of the
time delay.
It is not clear that there is a single correct answer. TM implementations featuring

weak atomicity that publish changes immediately within the transaction (rolling these
changes back upon abort) might be reasonably well served by the last alternative. Even
in this case, the code (or possibly even hardware) at the other end of the transaction may
require a substantial redesign to tolerate aborted transactions. This need for redesign
would make it more difficult to apply transactional memory to legacy code.
17.2.1.4 Persistence
There are many different types of locking primitives. One interesting distinction is
persistence, in other words, whether the lock can exist independently of the address
space of the process using the lock.
Non-persistent locks include pthread_mutex_lock(), pthread_rwlock_
rdlock(), and most kernel-level locking primitives. If the memory locations in-
stantiating a non-persistent lock’s data structures disappear, so does the lock. For typical
use of pthread_mutex_lock(), this means that when the process exits, all of its locks
vanish. This property can be exploited in order to trivialize lock cleanup at program
shutdown time, but makes it more difficult for unrelated applications to share locks, as
such sharing requires the applications to share memory.
Quick Quiz 17.1: But suppose that an application exits while holding a pthread_mutex_
lock() that happens to be located in a file-mapped region of memory?
Persistent locks help avoid the need to share memory among unrelated applications.
Persistent locking APIs include the flock family, lockf(), System V semaphores, or
the O_CREAT flag to open(). These persistent APIs can be used to protect large-scale
operations spanning runs of multiple applications, and, in the case of O_CREAT even
surviving operating-system reboot. If need be, locks can even span multiple computer
systems via distributed lock managers and distributed filesystems—and persist across
reboots of any or all of those computer systems.
Persistent locks can be used by any application, including applications written using
multiple languages and software environments. In fact, a persistent lock might well be
acquired by an application written in C and released by an application written in Python.
How could a similar persistent functionality be provided for TM?
1. Restrict persistent transactions to special-purpose environments designed to support

them, for example, SQL. This clearly works, given the decades-long history of
database systems, but does not provide the same degree of flexibility provided by
persistent locks.
v2023.06.11a
2. Use snapshot facilities provided by some storage devices and/or filesystems.

Unfortunately, this does not handle network communication, nor does it handle I/O
to devices that do not provide snapshot capabilities, for example, memory sticks.
3. Build a time machine.
4. Avoid the problem entirely by using existing persistent facilities, presumably

avoiding such use within transactions.
Of course, the fact that it is called transactional memory should give us pause, as
the name itself conflicts with the concept of a persistent transaction. It is nevertheless
worthwhile to consider this possibility as an important test case probing the inherent
limitations of transactional memory.
17.2.2 Process Modification

Processes are not eternal: They are created and destroyed, their memory mappings are
modified, they are linked to dynamic libraries, and they are debugged. These sections
look at how transactional memory can handle an ever-changing execution environment.
17.2.2.1 Multithreaded Transactions

It is perfectly legal to create processes and threads while holding a lock or, for that
matter, while holding a hazard pointer, within a sequence-locking read-side critical
section, and from within a userspace-RCU read-side critical section, and even all at the
same time, if need be. Not only is it legal, but it is quite simple, as can be seen from the
following code fragment:
1 pthread_mutex_lock(...);
2 for (i = 0; i < ncpus; i++)
3 pthread_create(&tid[i], ...);
4 for (i = 0; i < ncpus; i++)
5 pthread_join(tid[i], ...);
6 pthread_mutex_unlock(...);
This pseudo-code fragment uses pthread_create() to spawn one thread per CPU,
then uses pthread_join() to wait for each to complete, all under the protection of
pthread_mutex_lock(). The effect is to execute a lock-based critical section in
parallel, and one could obtain a similar effect using fork() and wait(). Of course,
the critical section would need to be quite large to justify the thread-spawning overhead,
but there are many examples of large critical sections in production software.
What might TM do about thread spawning within a transaction?
1. Declare pthread_create() to be illegal within transactions, preferably by

aborting the transaction. Alternatively, enlist the compiler to enforce pthread_
create()-free transactions.
2. Permit pthread_create() to be executed within a transaction, but only the parent

thread will be considered to be part of the transaction. This approach seems to be
reasonably compatible with existing and posited TM implementations, but seems
to be a trap for the unwary. This approach raises further questions, such as how to
handle conflicting child-thread accesses.
v2023.06.11a
3. Convert the pthread_create()s to function calls. This approach is also an

attractive nuisance, as it does not handle the not-uncommon cases where the child
threads communicate with one another. In addition, it does not permit concurrent
execution of the body of the transaction.
4. Extend the transaction to cover the parent and all child threads. This approach raises
interesting questions about the nature of conflicting accesses, given that the parent
and children are presumably permitted to conflict with each other, but not with other
threads. It also raises interesting questions as to what should happen if the parent
thread does not wait for its children before committing the transaction. Even more
interesting, what happens if the parent conditionally executes pthread_join()
based on the values of variables participating in the transaction? The answers to
these questions are reasonably straightforward in the case of locking. The answers
for TM are left as an exercise for the reader.
Given that parallel execution of transactions is commonplace in the database world,

it is perhaps surprising that current TM proposals do not provide for it. On the other
hand, the example above is a fairly sophisticated use of locking that is not normally
found in simple textbook examples, so perhaps its omission is to be expected. That
said, some researchers are using transactions to autoparallelize code [RKM+ 10], and
there are rumors that other TM researchers are investigating fork/join parallelism within
transactions, so perhaps this topic will soon be addressed more thoroughly.
17.2.2.2 The exec() System Call

One can execute an exec() system call within a lock-based critical section, while
holding a hazard pointer, within a sequence-locking read-side critical section, and from
within a userspace-RCU read-side critical section, and even all at the same time, if need
be. The exact semantics depends on the type of primitive.
In the case of non-persistent primitives (including pthread_mutex_lock(),
pthread_rwlock_rdlock(), and userspace RCU), if the exec() succeeds, the whole
address space vanishes, along with any locks being held. Of course, if the exec() fails,
the address space still lives, so any associated locks would also still live. A bit strange
perhaps, but well defined.
On the other hand, persistent primitives (including the flock family, lockf(), System
V semaphores, and the O_CREAT flag to open()) would survive regardless of whether
the exec() succeeded or failed, so that the exec()ed program might well release them.
Quick Quiz 17.2: What about non-persistent primitives represented by data structures in
mmap() regions of memory? What happens when there is an exec() within a critical section
of such a primitive?
What happens when you attempt to execute an exec() system call from within a
transaction?
1. Disallow exec() within transactions, so that the enclosing transactions abort

upon encountering the exec(). This is well defined, but clearly requires non-TM
synchronization primitives for use in conjunction with exec().
2. Disallow exec() within transactions, with the compiler enforcing this prohibition.
There is a draft specification for TM in C++ that takes this approach, allowing func-
tions to be decorated with the transaction_safe and transaction_unsafe
v2023.06.11a
attributes.4 This approach has some advantages over aborting the transaction at
runtime, but again requires non-TM synchronization primitives for use in conjunc-
tion with exec(). One disadvantage is the need to decorate a great many library
functions with transaction_safe and transaction_unsafe attributes.
3. Treat the transaction in a manner similar to non-persistent locking primitives, so
that the transaction survives if exec() fails, and silently commits if the exec()
succeeds. The case where only some of the variables affected by the transaction
reside in mmap()ed memory (and thus could survive a successful exec() system
call) is left as an exercise for the reader.
4. Abort the transaction (and the exec() system call) if the exec() system call
would have succeeded, but allow the transaction to continue if the exec() system
call would fail. This is in some sense the “correct” approach, but it would require
considerable work for a rather unsatisfying result.
The exec() system call is perhaps the strangest example of an obstacle to universal
TM applicability, as it is not completely clear what approach makes sense, and some
might argue that this is merely a reflection of the perils of real-life interaction with
exec(). That said, the two options prohibiting exec() within transactions are perhaps
the most logical of the group.
Similar issues surround the exit() and kill() system calls, as well as a longjmp()
or an exception that would exit the transaction. (Where did the longjmp() or exception
come from?)
17.2.2.3 Dynamic Linking and Loading

Lock-based critical section, code holding a hazard pointer, sequence-locking read-side
critical sections, and userspace-RCU read-side critical sections can (separately or in
combination) legitimately contain code that invokes dynamically linked and loaded
functions, including C/C++ shared libraries and Java class libraries. Of course, the
code contained in these libraries is by definition unknowable at compile time. So, what
happens if a dynamically loaded function is invoked within a transaction?
This question has two parts: (a) How do you dynamically link and load a function
within a transaction and (b) What do you do about the unknowable nature of the
code within this function? To be fair, item (b) poses some challenges for locking and
userspace-RCU as well, at least in theory. For example, the dynamically linked function
might introduce a deadlock for locking or might (erroneously) introduce a quiescent state
into a userspace-RCU read-side critical section. The difference is that while the class of
operations permitted in locking and userspace-RCU critical sections is well-understood,
there appears to still be considerable uncertainty in the case of TM. In fact, different
implementations of TM seem to have different restrictions.
So what can TM do about dynamically linked and loaded library functions? Options
for part (a), the actual loading of the code, include the following:
1. Treat the dynamic linking and loading in a manner similar to a page fault, so that
the function is loaded and linked, possibly aborting the transaction in the process.
If the transaction is aborted, the retry will find the function already present, and
the transaction can thus be expected to proceed normally.
4 Thanks to Mark Moir for pointing me at this spec, and to Michael Wong for having
pointed me at an earlier revision some time back.
v2023.06.11a
2. Disallow dynamic linking and loading of functions from within transactions.
Options for part (b), the inability to detect TM-unfriendly operations in a not-yet-
loaded function, possibilities include the following:
1. Just execute the code: If there are any TM-unfriendly operations in the function,
simply abort the transaction. Unfortunately, this approach makes it impossible for
the compiler to determine whether a given group of transactions may be safely
composed. One way to permit composability regardless is irrevocable transactions,
however, current implementations permit only a single irrevocable transaction to
proceed at any given time, which can severely limit performance and scalability.
Irrevocable transactions also to restrict use of manual transaction-abort operations.
Finally, if there is an irrevocable transaction manipulating a given data item, any
other transaction manipulating that same data item cannot have non-blocking
semantics.
2. Decorate the function declarations indicating which functions are TM-friendly.
These decorations can then be enforced by the compiler’s type system. Of
course, for many languages, this requires language extensions to be proposed,
standardized, and implemented, with the corresponding time delays, and also
with the corresponding decoration of a great many otherwise uninvolved library
functions. That said, the standardization effort is already in progress [ATS09].
3. As above, disallow dynamic linking and loading of functions from within transac-
tions.
I/O operations are of course a known weakness of TM, and dynamic linking and
loading can be thought of as yet another special case of I/O. Nevertheless, the proponents
of TM must either solve this problem, or resign themselves to a world where TM is but
one tool of several in the parallel programmer’s toolbox. (To be fair, a number of TM
proponents have long since resigned themselves to a world containing more than just
TM.)
17.2.2.4 Memory-Mapping Operations

It is perfectly legal to execute memory-mapping operations (including mmap(), shmat(),
and munmap() [Gro01]) within a lock-based critical section, while holding a hazard
pointer, within a sequence-locking read-side critical section, and from within a userspace-
RCU read-side critical section, and even all at the same time, if need be. What happens
when you attempt to execute such an operation from within a transaction? More to the
point, what happens if the memory region being remapped contains some variables
participating in the current thread’s transaction? And what if this memory region
contains variables participating in some other thread’s transaction?
It should not be necessary to consider cases where the TM system’s metadata is
remapped, given that most locking primitives do not define the outcome of remapping
their lock variables.
Here are some TM memory-mapping options:
1. Memory remapping is illegal within a transaction, and will result in all enclosing
transactions being aborted. This does simplify things somewhat, but also requires
that TM interoperate with synchronization primitives that do tolerate remapping
from within their critical sections.
v2023.06.11a
2. Memory remapping is illegal within a transaction, and the compiler is enlisted to

enforce this prohibition.
3. Memory mapping is legal within a transaction, but aborts all other transactions
having variables in the region mapped over.
4. Memory mapping is legal within a transaction, but the mapping operation will fail
if the region being mapped overlaps with the current transaction’s footprint.
5. All memory-mapping operations, whether within or outside a transaction, check
the region being mapped against the memory footprint of all transactions in the
system. If there is overlap, then the memory-mapping operation fails.
6. The effect of memory-mapping operations that overlap the memory footprint of

any transaction in the system is determined by the TM conflict manager, which
might dynamically determine whether to fail the memory-mapping operation or
abort any conflicting transactions.
It is interesting to note that munmap() leaves the relevant region of memory unmapped,
which could have additional interesting implications.5
17.2.2.5 Debugging
The usual debugging operations such as breakpoints work normally within lock-based
critical sections and from usespace-RCU read-side critical sections. However, in initial
transactional-memory hardware implementations [DLMN09] an exception within a
transaction will abort that transaction, which in turn means that breakpoints abort all
enclosing transactions.
So how can transactions be debugged?
1. Use software emulation techniques within transactions containing breakpoints. Of

course, it might be necessary to emulate all transactions any time a breakpoint is set
within the scope of any transaction. If the runtime system is unable to determine
whether or not a given breakpoint is within the scope of a transaction, then it might
be necessary to emulate all transactions just to be on the safe side. However, this
approach might impose significant overhead, which might in turn obscure the bug
being pursued.
2. Use only hardware TM implementations that are capable of handling breakpoint
exceptions. Unfortunately, as of this writing (March 2021), all such implementations
are research prototypes.
3. Use only software TM implementations, which are (very roughly speaking) more
tolerant of exceptions than are the simpler of the hardware TM implementations.
Of course, software TM tends to have higher overhead than hardware TM, so this
approach may not be acceptable in all situations.
4. Program more carefully, so as to avoid having bugs in the transactions in the first
place. As soon as you figure out how to do this, please do let everyone know the
secret!
5 This difference between mapping and unmapping was noted by Josh Triplett.
v2023.06.11a
There is some reason to believe that transactional memory will deliver productivity
improvements compared to other synchronization mechanisms, but it does seem quite
possible that these improvements could easily be lost if traditional debugging techniques
cannot be applied to transactions. This seems especially true if transactional memory is
to be used by novices on large transactions. In contrast, macho “top-gun” programmers
might be able to dispense with such debugging aids, especially for small transactions.
Therefore, if transactional memory is to deliver on its productivity promises to novice
programmers, the debugging problem does need to be solved.
17.2.3 Synchronization
If transactional memory someday proves that it can be everything to everyone, it will
not need to interact with any other synchronization mechanism. Until then, it will need
to work with synchronization mechanisms that can do what it cannot, or that work more
naturally in a given situation. The following sections outline the current challenges in
this area.
17.2.3.1 Locking
It is commonplace to acquire locks while holding other locks, which works quite well,
at least as long as the usual well-known software-engineering techniques are employed
to avoid deadlock. It is not unusual to acquire locks from within RCU read-side critical
sections, which eases deadlock concerns because RCU read-side primitives cannot
participate in lock-based deadlock cycles. It is also possible to acquire locks while
holding hazard pointers and within sequence-lock read-side critical sections. But what
happens when you attempt to acquire a lock from within a transaction?
In theory, the answer is trivial: Simply manipulate the data structure representing the
lock as part of the transaction, and everything works out perfectly. In practice, a number
of non-obvious complications [VGS08] can arise, depending on implementation details
of the TM system. These complications can be resolved, but at the cost of a 45 %
increase in overhead for locks acquired outside of transactions and a 300 % increase in
overhead for locks acquired within transactions. Although these overheads might be
acceptable for transactional programs containing small amounts of locking, they are
often completely unacceptable for production-quality lock-based programs wishing to
use the occasional transaction.
1. Use only locking-friendly TM implementations. Unfortunately, the locking-
unfriendly implementations have some attractive properties, including low over-
head for successful transactions and the ability to accommodate extremely large
transactions.
2. Use TM only “in the small” when introducing TM to lock-based programs, thereby
accommodating the limitations of locking-friendly TM implementations.
3. Set aside locking-based legacy systems entirely, re-implementing everything in
terms of transactions. This approach has no shortage of advocates, but this requires
that all the issues described in this series be resolved. During the time it takes to
resolve these issues, competing synchronization mechanisms will of course also
have the opportunity to improve.
4. Use TM strictly as an optimization in lock-based systems, as was done by
the TxLinux [RHP+ 07] group and by a great many transactional lock elision
v2023.06.11a
projects [PD11, Kle14, FIMR16, PMDY20]. This approach seems sound, but
leaves the locking design constraints (such as the need to avoid deadlock) firmly in
place.
5. Strive to reduce the overhead imposed on locking primitives.
The fact that there could possibly be a problem interfacing TM and locking came as a
surprise to many, which underscores the need to try out new mechanisms and primitives
in real-world production software. Fortunately, the advent of open source means that a
huge quantity of such software is now freely available to everyone, including researchers.
17.2.3.2 Reader-Writer Locking

It is commonplace to read-acquire reader-writer locks while holding other locks, which
just works, at least as long as the usual well-known software-engineering techniques
are employed to avoid deadlock. Read-acquiring reader-writer locks from within RCU
read-side critical sections also works, and doing so eases deadlock concerns because
RCU read-side primitives cannot participate in lock-based deadlock cycles. It is also
possible to acquire locks while holding hazard pointers and within sequence-lock
read-side critical sections. But what happens when you attempt to read-acquire a
reader-writer lock from within a transaction?
Unfortunately, the straightforward approach to read-acquiring the traditional counter-
based reader-writer lock within a transaction defeats the purpose of the reader-writer
lock. To see this, consider a pair of transactions concurrently attempting to read-acquire
the same reader-writer lock. Because read-acquisition involves modifying the reader-
writer lock’s data structures, a conflict will result, which will roll back one of the two
transactions. This behavior is completely inconsistent with the reader-writer lock’s goal
of allowing concurrent readers.
Here are some options available to TM:
1. Use per-CPU or per-thread reader-writer locking [HW92], which allows a given

CPU (or thread, respectively) to manipulate only local data when read-acquiring
the lock. This would avoid the conflict between the two transactions concurrently
read-acquiring the lock, permitting both to proceed, as intended. Unfortunately,
(1) the write-acquisition overhead of per-CPU/thread locking can be extremely
high, (2) the memory overhead of per-CPU/thread locking can be prohibitive, and
(3) this transformation is available only when you have access to the source code
in question. Other more-recent scalable reader-writer locks [LLO09] might avoid
some or all of these problems.
2. Use TM only “in the small” when introducing TM to lock-based programs, thereby
avoiding read-acquiring reader-writer locks from within transactions.
3. Set aside locking-based legacy systems entirely, re-implementing everything in
terms of transactions. This approach has no shortage of advocates, but this requires
that all the issues described in this series be resolved. During the time it takes to
resolve these issues, competing synchronization mechanisms will of course also
have the opportunity to improve.
4. Use TM strictly as an optimization in lock-based systems, as was done by the
TxLinux [RHP+ 07] group, and as has been done by more recent work using TM
to elide reader writer locks [FIMR16]. This approach seems sound, at least on
v2023.06.11a
POWER8 CPUs [LGW+ 15], but leaves the locking design constraints (such as the
need to avoid deadlock) firmly in place.
Of course, there might well be other non-obvious issues surrounding combining TM

with reader-writer locking, as there in fact were with exclusive locking.
17.2.3.3 Deferred Reclamation

This section focuses mainly on RCU. Similar issues and possible resolutions arise when
combining TM with other deferred-reclamation mechanisms such as reference counters
and hazard pointers. In the text below, known differences are specifically called out.
Reference counting, hazard pointers, and RCU are all heavily used, as noted in
Sections 9.5.5 and 9.6.3. This means that any TM implementation that chooses not
to surmount each and every challenge called out in this section needs to interoperate
cleanly and efficiently with all of these synchronization mechanisms.
The TxLinux group from the University of Texas at Austin appears to be the group to
take on the challenge of RCU/TM interoperability [RHP+ 07]. Because they applied TM
to the Linux 2.6 kernel, which uses RCU, they had no choice but to integrate TM and
RCU, with TM taking the place of locking for RCU updates. Unfortunately, although
the paper does state that the RCU implementation’s locks (e.g., rcu_ctrlblk.lock)
were converted to transactions, it is silent about what was done with those locks used by
RCU-based updates (for example, dcache_lock).
More recently, Dimitrios Siakavaras et al. have applied HTM and RCU to search
trees [SNGK17, SBN+ 20], Christina Giannoula et al. have used HTM and RCU to
color graphs [GGK18], and SeongJae Park et al. have used HTM and RCU to optimize
high-contention locking on NUMA systems [PMDY20].
It is important to note that RCU permits readers and updaters to run concurrently,
further permitting RCU readers to access data that is in the act of being updated. Of
course, this property of RCU, whatever its performance, scalability, and real-time-
response benefits might be, flies in the face of the underlying atomicity properties of
TM, although the POWER8 CPU family’s suspended-transaction facility [LGW+ 15]
makes it an exception to this rule.
So how should TM-based updates interact with concurrent RCU readers? Some
possibilities are as follows:
1. RCU readers abort concurrent conflicting TM updates. This is in fact the approach
taken by the TxLinux project. This approach does preserve RCU semantics, and
also preserves RCU’s read-side performance, scalability, and real-time-response
properties, but it does have the unfortunate side-effect of unnecessarily aborting
conflicting updates. In the worst case, a long sequence of RCU readers could
potentially starve all updaters, which could in theory result in system hangs.
In addition, not all TM implementations offer the strong atomicity required to
implement this approach, and for good reasons.
2. RCU readers that run concurrently with conflicting TM updates get old (pre-
transaction) values from any conflicting RCU loads. This preserves RCU semantics
and performance, and also prevents RCU-update starvation. However, not all TM
implementations can provide timely access to old values of variables that have
been tentatively updated by an in-flight transaction. In particular, log-based TM
implementations that maintain old values in the log (thus providing excellent TM
commit performance) are not likely to be happy with this approach. Perhaps the
v2023.06.11a
rcu_dereference() primitive can be leveraged to permit RCU to access the old

values within a greater range of TM implementations, though performance might
still be an issue. Nevertheless, there are popular TM implementations that have
been integrated with RCU in this manner [PW07, HW11, HW14].
3. If an RCU reader executes an access that conflicts with an in-flight transaction, then
that RCU access is delayed until the conflicting transaction either commits or aborts.
This approach preserves RCU semantics, but not RCU’s performance or real-time
response, particularly in presence of long-running transactions. In addition, not all
TM implementations are capable of delaying conflicting accesses. Nevertheless,
this approach seems eminently reasonable for hardware TM implementations that
support only small transactions.
4. RCU readers are converted to transactions. This approach pretty much guarantees
that RCU is compatible with any TM implementation, but it also imposes TM’s
rollbacks on RCU read-side critical sections, destroying RCU’s real-time response
guarantees, and also degrading RCU’s read-side performance. Furthermore, this
approach is infeasible in cases where any of the RCU read-side critical sections
contains operations that the TM implementation in question is incapable of handling.
This approach is more difficult to apply to hazard pointers and reference counters,
which do not have a sharply defined notion of a reader as a section of code.
5. Many update-side uses of RCU modify a single pointer to publish a new data
structure. In some of these cases, RCU can safely be permitted to see a transactional
pointer update that is subsequently rolled back, as long as the transaction respects
memory ordering and as long as the roll-back process uses call_rcu() to free up
the corresponding structure. Unfortunately, not all TM implementations respect
memory barriers within a transaction. Apparently, the thought is that because
transactions are supposed to be atomic, the ordering of the accesses within the
transaction is not supposed to matter.
6. Prohibit use of TM in RCU updates. This is guaranteed to work, but restricts use
of TM.
It seems likely that additional approaches will be uncovered, especially given the
advent of user-level RCU and hazard-pointer implementations.6 It is interesting to note
that many of the better performing and scaling STM implementations make use of
RCU-like techniques internally [Fra04, FH07, GYW+ 19, KMK+ 19].
Quick Quiz 17.3: MV-RLU looks pretty good! Doesn’t it beat RCU hands down?
17.2.3.4 Extra-Transactional Accesses

Within a lock-based critical section, it is perfectly legal to manipulate variables that are
concurrently accessed or even modified outside that lock’s critical section, with one
common example being statistical counters. The same thing is possible within RCU
read-side critical sections, and is in fact the common case.
6 Kudos to the TxLinux group, Maged Michael, and Josh Triplett for coming up with a
number of the above alternatives.
v2023.06.11a
Given mechanisms such as the so-called “dirty reads” that are prevalent in production
database systems, it is not surprising that extra-transactional accesses have received seri-
ous attention from the proponents of TM, with the concept of weak atomicity [BLM06]
being but one case in point.
Here are some extra-transactional options:
1. Conflicts due to extra-transactional accesses always abort transactions. This is

strong atomicity.
2. Conflicts due to extra-transactional accesses are ignored, so only conflicts among

transactions can abort transactions. This is weak atomicity.
3. Transactions are permitted to carry out non-transactional operations in special

cases, such as when allocating memory or interacting with lock-based critical
sections.
4. Produce hardware extensions that permit some operations (for example, addition)
to be carried out concurrently on a single variable by multiple transactions.
5. Introduce weak semantics to transactional memory. One approach is the combina-

tion with RCU described in Section 17.2.3.3, while Gramoli and Guerraoui survey
a number of other weak-transaction approaches [GG14], for example, restricted
partitioning of large “elastic” transactions into smaller transactions, thus reducing
conflict probabilities (albeit with tepid performance and scalability). Perhaps
further experience will show that some uses of extra-transactional accesses can be
replaced by weak transactions.
It appears that transactions were conceived in a vacuum, with no interaction required

with any other synchronization mechanism. If so, it is no surprise that much confusion
and complexity arises when combining transactions with non-transactional accesses.
But unless transactions are to be confined to small updates to isolated data structures, or
alternatively to be confined to new programs that do not interact with the huge body of
existing parallel code, then transactions absolutely must be so combined if they are to
have large-scale practical impact in the near term.
17.2.4 Discussion
The obstacles to universal TM adoption lead to the following conclusions:
1. One interesting property of TM is the fact that transactions are subject to rollback
and retry. This property underlies TM’s difficulties with irreversible operations,
including unbuffered I/O, RPCs, memory-mapping operations, time delays, and
the exec() system call. This property also has the unfortunate consequence of
introducing all the complexities inherent in the possibility of failure, often in a
developer-visible manner.
2. Another interesting property of TM, noted by Shpeisman et al. [SATG+ 09], is

that TM intertwines the synchronization with the data it protects. This property
underlies TM’s issues with I/O, memory-mapping operations, extra-transactional
accesses, and debugging breakpoints. In contrast, conventional synchronization
primitives, including locking and RCU, maintain a clear separation between the
synchronization primitives and the data that they protect.
v2023.06.11a
Figure 17.9: The STM Vision
3. One of the stated goals of many workers in the TM area is to ease parallelization
of large sequential programs. As such, individual transactions are commonly
expected to execute serially, which might do much to explain TM’s issues with
multithreaded transactions.
Quick Quiz 17.4: Given things like spin_trylock(), how does it make any sense at all to
claim that TM introduces the concept of failure???
What should TM researchers and developers do about all of this?

One approach is to focus on TM in the small, focusing on small transactions where
hardware assist potentially provides substantial advantages over other synchronization
primitives and on small programs where there is some evidence for increased productivity
for a combined TM-locking approach [PAT11]. Sun took the small-transaction approach
with its Rock research CPU [DLMN09]. Some TM researchers seem to agree with these
two small-is-beautiful approaches [SSHT93], others have much higher hopes for TM,
and yet others hint that high TM aspirations might be TM’s worst enemy [Att10, Section
6]. It is nonetheless quite possible that TM will be able to take on larger problems, and
this section has listed a few of the issues that must be resolved if TM is to achieve this
lofty goal.
Of course, everyone involved should treat this as a learning experience. It would seem
that TM researchers have great deal to learn from practitioners who have successfully
built large software systems using traditional synchronization primitives.
And vice versa.
Quick Quiz 17.5: What is to learn? Why not just use TM for memory-based data structures
and locking for those rare cases featuring the many silly corner cases listed in this silly section???
But for the moment, the current state of STM can best be summarized with a series
of cartoons. First, Figure 17.9 shows the STM vision. As always, the reality is a bit
v2023.06.11a
Figure 17.10: The STM Reality: Conflicts
Figure 17.11: The STM Reality: Irrevocable Operations
v2023.06.11a
17.3. HARDWARE TRANSACTIONAL MEMORY 599
Figure 17.12: The STM Reality: Realtime Response
more nuanced, as fancifully depicted by Figures 17.10, 17.11, and 17.12.7 Less fanciful
STM retrospectives are also available [Duf10a, Duf10b].
Some commercially available hardware supports restricted variants of HTM, which
are addressed in the following section.
17.3 Hardware Transactional Memory

Make sure your report system is reasonably clean
and efficient before you automate. Otherwise, your
new computer will just speed up the mess.
Robert Townsend
As of 2021, hardware transactional memory (HTM) has been available for many years
on several types of commercially available commodity computer systems [YHLR13,
Mer11, JSG12, Hay20]. This section makes an attempt to identify HTM’s place in the
parallel programmer’s toolbox.
From a conceptual viewpoint, HTM uses processor caches and speculative execution
to make a designated group of statements (a “transaction”) take effect atomically from
the viewpoint of any other transactions running on other processors. This transaction
is initiated by a begin-transaction machine instruction and completed by a commit-
transaction machine instruction. There is typically also an abort-transaction machine
instruction, which squashes the speculation (as if the begin-transaction instruction
and all following instructions had not executed) and commences execution at a failure
handler. The location of the failure handler is typically specified by the begin-transaction
7 Recent academic work-in-progress has investigated lock-based STM systems for real-
time use [And19, NA18], albeit without any performance results, and with some indications
that real-time hybrid STM/HTM systems must choose between fast common-case performance
and worst-case forward-progress guarantees [AKK+ 14, SBV10].
v2023.06.11a
instruction, either as an explicit failure-handler address or via a condition code set by

the instruction itself. Each transaction executes atomically with respect to all other
transactions.
HTM has a number of important benefits, including automatic dynamic partitioning
of data structures, reducing synchronization-primitive cache misses, and supporting a
fair number of practical applications.
However, it always pays to read the fine print, and HTM is no exception. A major
point of this section is determining under what conditions HTM’s benefits outweigh
the complications hidden in its fine print. To this end, Section 17.3.1 describes HTM’s
benefits and Section 17.3.2 describes its weaknesses. This is the same approach used in
earlier papers [MMW07, MMTW10] and also in the previous section.8
Section 17.3.3 then describes HTM’s weaknesses with respect to the combination of
synchronization primitives used in the Linux kernel (and in many user-space applications).
Section 17.3.4 looks at where HTM might best fit into the parallel programmer’s toolbox,
and Section 17.3.5 lists some events that might greatly increase HTM’s scope and
appeal. Finally, Section 17.3.6 presents concluding remarks.
17.3.1 HTM Benefits WRT Locking

The primary benefits of HTM are (1) its avoidance of the cache misses that are often
incurred by other synchronization primitives, (2) its ability to dynamically partition
data structures, and (3) the fact that it has a fair number of practical applications. I
break from TM tradition by not listing ease of use separately for two reasons. First,
ease of use should stem from HTM’s primary benefits, which this section focuses
on. Second, there has been considerable controversy surrounding attempts to test for
raw programming talent [Bor06, DBA09, PBCE20] and even around the use of small
programming exercises in job interviews [Bra07]. This indicates that we really do not
have a firm grasp on what makes programming easy or hard. Therefore, the remainder
of this section focuses on the three benefits listed above.
17.3.1.1 Avoiding Synchronization Cache Misses

Most synchronization mechanisms are based on data structures that are operated on by
atomic instructions. Because these atomic instructions normally operate by first causing
the relevant cache line to be owned by the CPU that they are running on, a subsequent
execution of the same instance of that synchronization primitive on some other CPU
will result in a cache miss. These communications cache misses severely degrade both
the performance and scalability of conventional synchronization mechanisms [ABD+ 97,
Section 4.2.3].
In contrast, HTM synchronizes by using the CPU’s cache, avoiding the need for a
separate synchronization data structure and resultant cache misses. HTM’s advantage is
greatest in cases where a lock data structure is placed in a separate cache line, in which
case, converting a given critical section to an HTM transaction can reduce that critical
section’s overhead by a full cache miss. These savings can be quite significant for the
common case of short critical sections, at least for those situations where the elided
lock does not share a cache line with an oft-written variable protected by that lock.
8 I gratefully acknowledge many stimulating discussions with the other authors, Maged
Michael, Josh Triplett, and Jonathan Walpole, as well as with Andi Kleen.
v2023.06.11a
Quick Quiz 17.6: Why would it matter that oft-written variables shared the cache line with
the lock variable?
17.3.1.2 Dynamic Partitioning of Data Structures

A major obstacle to the use of some conventional synchronization mechanisms is the
need to statically partition data structures. There are a number of data structures that are
trivially partitionable, with the most prominent example being hash tables, where each
hash chain constitutes a partition. Allocating a lock for each hash chain then trivially
parallelizes the hash table for operations confined to a given chain.9 Partitioning is
similarly trivial for arrays, radix trees, skiplists, and several other data structures.
However, partitioning for many types of trees and graphs is quite difficult, and the
results are often quite complex [Ell80]. Although it is possible to use two-phased locking
and hashed arrays of locks to partition general data structures, other techniques have
proven preferable [Mil06], as will be discussed in Section 17.3.3. Given its avoidance
of synchronization cache misses, HTM is therefore a very real possibility for large
non-partitionable data structures, at least assuming relatively small updates.
Quick Quiz 17.7: Why are relatively small updates important to HTM performance and
scalability?
17.3.1.3 Practical Value

Some evidence of HTM’s practical value has been demonstrated in a number of
hardware platforms, including Sun Rock [DLMN09], Azul Vega [Cli09], IBM Blue
Gene/Q [Mer11], Intel Haswell TSX [RD12], and IBM System z [JSG12].
Expected practical benefits include:
1. Lock elision for in-memory data access and update [MT01, RG02].
2. Concurrent access and small random updates to large non-partitionable data
structures.
However, HTM also has some very real shortcomings, which will be discussed in the
next section.
17.3.2 HTM Weaknesses WRT Locking

The concept of HTM is quite simple: A group of accesses and updates to memory
occurs atomically. However, as is the case with many simple ideas, complications arise
when you apply it to real systems in the real world. These complications are as follows:
1. Transaction-size limitations.
2. Conflict handling.
3. Aborts and rollbacks.

4. Lack of forward-progress guarantees.
9 And it is also easy to extend this scheme to operations accessing multiple hash chains
by having such operations acquire the locks for all relevant chains in hash order.
v2023.06.11a
5. Irrevocable operations.
6. Semantic differences.
Each of these complications is covered in the following sections, followed by a

summary.
17.3.2.1 Transaction-Size Limitations

The transaction-size limitations of current HTM implementations stem from the use
of the processor caches to hold the data affected by the transaction. Although this
allows a given CPU to make the transaction appear atomic to other CPUs by executing
the transaction within the confines of its cache, it also means that any transaction that
does not fit cannot commit. Furthermore, events that change execution context, such as
interrupts, system calls, exceptions, traps, and context switches either must abort any
ongoing transaction on the CPU in question or must further restrict transaction size due
to the cache footprint of the other execution context.
Of course, modern CPUs tend to have large caches, and the data required for many
transactions would fit easily in a one-megabyte cache. Unfortunately, with caches, sheer
size is not all that matters. The problem is that most caches can be thought of hash
tables implemented in hardware. However, hardware caches do not chain their buckets
(which are normally called sets), but rather provide a fixed number of cachelines per set.
The number of elements provided for each set in a given cache is termed that cache’s
associativity.
Although cache associativity varies, the eight-way associativity of the level-0 cache
on the laptop I am typing this on is not unusual. What this means is that if a given
transaction needed to touch nine cache lines, and if all nine cache lines mapped to
the same set, then that transaction cannot possibly complete, never mind how many
megabytes of additional space might be available in that cache. Yes, given randomly
selected data elements in a given data structure, the probability of that transaction being
able to commit is quite high, but there can be no guarantee [McK11c].
There has been some research work to alleviate this limitation. Fully associative victim
caches would alleviate the associativity constraints, but there are currently stringent
performance and energy-efficiency constraints on the sizes of victim caches. That said,
HTM victim caches for unmodified cache lines can be quite small, as they need to retain
only the address: The data itself can be written to memory or shadowed by other caches,
while the address itself is sufficient to detect a conflicting write [RD12].
Unbounded-transactional-memory (UTM) schemes [AAKL06, MBM+ 06] use
DRAM as an extremely large victim cache, but integrating such schemes into a
production-quality cache-coherence mechanism is still an unsolved problem. In
addition, use of DRAM as a victim cache may have unfortunate performance and
energy-efficiency consequences, particularly if the victim cache is to be fully associative.
Finally, the “unbounded” aspect of UTM assumes that all of DRAM could be used
as a victim cache, while in reality the large but still fixed amount of DRAM assigned
to a given CPU would limit the size of that CPU’s transactions. Other schemes use a
combination of hardware and software transactional memory [KCH+ 06] and one could
imagine using STM as a fallback mechanism for HTM.
However, to the best of my knowledge, with the exception of abbreviating representa-
tion of TM read sets, currently available systems do not implement any of these research
ideas, and perhaps for good reason.
v2023.06.11a
17.3.2.2 Conflict Handling

The first complication is the possibility of conflicts. For example, suppose that
transactions A and B are defined as follows:
Transaction A Transaction B
x = 1; y = 2;
y = 3; x = 4;
Suppose that each transaction executes concurrently on its own processor. If

transaction A stores to x at the same time that transaction B stores to y, neither
transaction can progress. To see this, suppose that transaction A executes its store
to y. Then transaction A will be interleaved within transaction B, in violation of the
requirement that transactions execute atomically with respect to each other. Allowing
transaction B to execute its store to x similarly violates the atomic-execution requirement.
This situation is termed a conflict, which happens whenever two concurrent transactions
access the same variable where at least one of the accesses is a store. The system is
therefore obligated to abort one or both of the transactions in order to allow execution
to progress. The choice of exactly which transaction to abort is an interesting topic that
will very likely retain the ability to generate Ph.D. dissertations for some time to come,
see for example [ATC+ 11].10 For the purposes of this section, we can assume that the
system makes a random choice.
Another complication is conflict detection, which is comparatively straightforward,
at least in the simplest case. When a processor is executing a transaction, it marks
every cache line touched by that transaction. If the processor’s cache receives a request
involving a cache line that has been marked as touched by the current transaction, a
potential conflict has occurred. More sophisticated systems might try to order the
current processors’ transaction to precede that of the processor sending the request, and
optimizing this process will likely also retain the ability to generate Ph.D. dissertations
for quite some time. However this section assumes a very simple conflict-detection
strategy.
However, for HTM to work effectively, the probability of conflict must be quite
low, which in turn requires that the data structures be organized so as to maintain
a sufficiently low probability of conflict. For example, a red-black tree with simple
insertion, deletion, and search operations fits this description, but a red-black tree that
maintains an accurate count of the number of elements in the tree does not.11 For
another example, a red-black tree that enumerates all elements in the tree in a single
transaction will have high conflict probabilities, degrading performance and scalability.
As a result, many serial programs will require some restructuring before HTM can
work effectively. In some cases, practitioners will prefer to take the extra steps (in the
red-black-tree case, perhaps switching to a partitionable data structure such as a radix
tree or a hash table), and just use locking, particularly until such time as HTM is readily
available on all relevant architectures [Cli09].
Quick Quiz 17.8: How could a red-black tree possibly efficiently enumerate all elements of
the tree regardless of choice of synchronization mechanism???
10 Liu’s and Spear’s paper entitled “Toxic Transactions” [LS11] is particularly instructive.
11 The need to update the count would result in additions to and deletions from the tree
conflicting with each other, resulting in strong non-commutativity [AGH+ 11a, AGH+ 11b,
McK11b].
v2023.06.11a
Furthermore, the potential for conflicting accesses among concurrent transactions

can result in failure. Handling such failure is discussed in the next section.
17.3.2.3 Aborts and Rollbacks

Because any transaction might be aborted at any time, it is important that transactions
contain no statements that cannot be rolled back. This means that transactions cannot
do I/O, system calls, or debugging breakpoints (no single stepping in the debugger
for HTM transactions!!!). Instead, transactions must confine themselves to accessing
normal cached memory. Furthermore, on some systems, interrupts, exceptions, traps,
TLB misses, and other events will also abort transactions. Given the number of bugs
that have resulted from improper handling of error conditions, it is fair to ask what
impact aborts and rollbacks have on ease of use.
Quick Quiz 17.9: But why can’t a debugger emulate single stepping by setting breakpoints
at successive lines of the transaction, relying on the retry to retrace the steps of the earlier
instances of the transaction?
Of course, aborts and rollbacks raise the question of whether HTM can be useful
for hard real-time systems. Do the performance benefits of HTM outweigh the costs
of the aborts and rollbacks, and if so under what conditions? Can transactions use
priority boosting? Or should transactions for high-priority threads instead preferentially
abort those of low-priority threads? If so, how is the hardware efficiently informed of
priorities? The literature on real-time use of HTM is quite sparse, perhaps because
there are more than enough problems in making HTM work well in non-real-time
environments.
Because current HTM implementations might deterministically abort a given trans-
action, software must provide fallback code. This fallback code must use some other
form of synchronization, for example, locking. If a lock-based fallback is ever used,
then all the limitations of locking, including the possibility of deadlock, reappear. One
can of course hope that the fallback isn’t used often, which might allow simpler and
less deadlock-prone locking designs to be used. But this raises the question of how
the system transitions from using the lock-based fallbacks back to transactions.12 One
approach is to use a test-and-test-and-set discipline [MT02], so that everyone holds off
until the lock is released, allowing the system to start from a clean slate in transactional
mode at that point. However, this could result in quite a bit of spinning, which might
not be wise if the lock holder has blocked or been preempted. Another approach is to
allow transactions to proceed in parallel with a thread holding a lock [MT02], but this
raises difficulties in maintaining atomicity, especially if the reason that the thread is
holding the lock is because the corresponding transaction would not fit into cache.
Finally, dealing with the possibility of aborts and rollbacks seems to put an additional
burden on the developer, who must correctly handle all combinations of possible error
conditions.
It is clear that users of HTM must put considerable validation effort into testing both
the fallback code paths and transition from fallback code back to transactional code.
Nor is there any reason to believe that the validation requirements of HTM hardware
are any less daunting.
12 The possibility of an application getting stuck in fallback mode has been termed the
“lemming effect”, a term that Dave Dice has been credited with coining.
v2023.06.11a
17.3.2.4 Lack of Forward-Progress Guarantees
Even though transaction size, conflicts, and aborts/rollbacks can all cause transactions
to abort, one might hope that sufficiently small and short-duration transactions could be
guaranteed to eventually succeed. This would permit a transaction to be unconditionally
retried, in the same way that compare-and-swap (CAS) and load-linked/store-conditional
(LL/SC) operations are unconditionally retried in code that uses these instructions to
implement atomic operations.
Unfortunately, other than low-clock-rate academic research prototypes [SBV10],
currently available HTM implementations refuse to make any sort of forward-progress
guarantee. As noted earlier, HTM therefore cannot be used to avoid deadlock on
those systems. Hopefully future implementations of HTM will provide some sort of
forward-progress guarantees. Until that time, HTM must be used with extreme caution
in real-time applications.
The one exception to this gloomy picture as of 2021 is the IBM mainframe, which
provides constrained transactions [JSG12]. The constraints are quite severe, and are
presented in Section 17.3.5.1. It will be interesting to see if HTM forward-progress
guarantees migrate from the mainframe to commodity CPU families.
17.3.2.5 Irrevocable Operations
Another consequence of aborts and rollbacks is that HTM transactions cannot accom-
modate irrevocable operations. Current HTM implementations typically enforce this
limitation by requiring that all of the accesses in the transaction be to cacheable memory
(thus prohibiting MMIO accesses) and aborting transactions on interrupts, traps, and
exceptions (thus prohibiting system calls).
Note that buffered I/O can be accommodated by HTM transactions as long as the
buffer fill/flush operations occur extra-transactionally. The reason that this works is that
adding data to and removing data from the buffer is revocable: Only the actual buffer
fill/flush operations are irrevocable. Of course, this buffered-I/O approach has the effect
of including the I/O in the transaction’s footprint, increasing the size of the transaction
and thus increasing the probability of failure.
17.3.2.6 Semantic Differences
Although HTM can in many cases be used as a drop-in replacement for locking (hence
the name transactional lock elision (TLE) [DHL+ 08]), there are subtle differences
in semantics. A particularly nasty example involving coordinated lock-based critical
sections that results in deadlock or livelock when executed transactionally was given by
Blundell [BLM06], but a much simpler example is the empty critical section.
In a lock-based program, an empty critical section will guarantee that all processes
that had previously been holding that lock have now released it. This idiom was used
by the 2.4 Linux kernel’s networking stack to coordinate changes in configuration.
But if this empty critical section is translated to a transaction, the result is a no-op.
The guarantee that all prior critical sections have terminated is lost. In other words,
transactional lock elision preserves the data-protection semantics of locking, but loses
locking’s time-based messaging semantics.
Quick Quiz 17.10: But why would anyone need an empty lock-based critical section???
v2023.06.11a
Listing 17.1: Exploiting Priority Boosting

1 void boostee(void)
2 {
3 int i = 0;
4
5 acquire_lock(&boost_lock[i]);
6 for (;;) {
7 acquire_lock(&boost_lock[!i]);
8 release_lock(&boost_lock[i]);
9 i = i ^ 1;
10 do_something();
11 }
12 }
13
14 void booster(void)
15 {
16 int i = 0;
17
18 for (;;) {
19 usleep(500); /* sleep 0.5 ms. */
20 acquire_lock(&boost_lock[i]);
21 release_lock(&boost_lock[i]);
22 i = i ^ 1;
23 }
24 }
Quick Quiz 17.11: Can’t transactional lock elision trivially handle locking’s time-based
messaging semantics by simply choosing not to elide empty lock-based critical sections?
Quick Quiz 17.12: Given modern hardware [MOZ09], how can anyone possibly expect
parallel software relying on timing to work?
One important semantic difference between locking and transactions is the priority
boosting that is used to avoid priority inversion in lock-based real-time programs. One
way in which priority inversion can occur is when a low-priority thread holding a lock
is preempted by a medium-priority CPU-bound thread. If there is at least one such
medium-priority thread per CPU, the low-priority thread will never get a chance to run.
If a high-priority thread now attempts to acquire the lock, it will block. It cannot acquire
the lock until the low-priority thread releases it, the low-priority thread cannot release
the lock until it gets a chance to run, and it cannot get a chance to run until one of the
medium-priority threads gives up its CPU. Therefore, the medium-priority threads are
in effect blocking the high-priority process, which is the rationale for the name “priority
inversion.”
One way to avoid priority inversion is priority inheritance, in which a high-priority
thread blocked on a lock temporarily donates its priority to the lock’s holder, which is
also called priority boosting. However, priority boosting can be used for things other
than avoiding priority inversion, as shown in Listing 17.1. Lines 1–12 of this listing
show a low-priority process that must nevertheless run every millisecond or so, while
lines 14–24 of this same listing show a high-priority process that uses priority boosting
to ensure that boostee() runs periodically as needed.
The boostee() function arranges this by always holding one of the two boost_
lock[] locks, so that lines 20–21 of booster() can boost priority as needed.
Quick Quiz 17.13: But the boostee() function in Listing 17.1 alternatively acquires its locks
in reverse order! Won’t this result in deadlock?
This arrangement requires that boostee() acquire its first lock on line 5 before the
system becomes busy, but this is easily arranged, even on modern hardware.
v2023.06.11a
Unfortunately, this arrangement can break down in presence of transactional lock

elision. The boostee() function’s overlapping critical sections become one infinite
transaction, which will sooner or later abort, for example, on the first time that the
thread running the boostee() function is preempted. At this point, boostee() will
fall back to locking, but given its low priority and that the quiet initialization period is
now complete (which after all is why boostee() was preempted), this thread might
never again get a chance to run.
And if the boostee() thread is not holding the lock, then the booster() thread’s
empty critical section on lines 20 and 21 of Listing 17.1 will become an empty transaction
that has no effect, so that boostee() never runs. This example illustrates some of the
subtle consequences of transactional memory’s rollback-and-retry semantics.
Given that experience will likely uncover additional subtle semantic differences,
application of HTM-based lock elision to large programs should be undertaken with
caution. That said, where it does apply, HTM-based lock elision can eliminate the
cache misses associated with the lock variable, which has resulted in tens of percent
performance increases in large real-world software systems as of early 2015. We can
therefore expect to see substantial use of this technique on hardware providing reliable
support for it.
Quick Quiz 17.14: So a bunch of people set out to supplant locking, and they mostly end up
just optimizing locking???
17.3.2.7 Summary
Although it seems likely that HTM will have compelling use cases, current imple-
mentations have serious transaction-size limitations, conflict-handling complications,
abort-and-rollback issues, and semantic differences that will require careful handling.
HTM’s current situation relative to locking is summarized in Table 17.1. As can be seen,
although the current state of HTM alleviates some serious shortcomings of locking,13
it does so by introducing a significant number of shortcomings of its own. These
shortcomings are acknowledged by leaders in the TM community [MS12].14
In addition, this is not the whole story. Locking is not normally used by itself, but is
instead typically augmented by other synchronization mechanisms, including reference
counting, atomic operations, non-blocking data structures, hazard pointers [Mic04a,
HLM02], and RCU [MS98a, MAK+ 01, HMBW07, McK12b]. The next section looks
at how such augmentation changes the equation.
17.3.3 HTM Weaknesses WRT Locking When Augmented

Practitioners have long used reference counting, atomic operations, non-blocking data
structures, hazard pointers, and RCU to avoid some of the shortcomings of locking.
13 In fairness, it is important to emphasize that locking’s shortcomings do have well-known
and heavily used engineering solutions, including deadlock detectors [Cor06a], a wealth of
data structures that have been adapted to locking, and a long history of augmentation, as
discussed in Section 17.3.3. In addition, if locking really were as horrible as a quick skim of
many academic papers might reasonably lead one to believe, where did all the large lock-based
parallel programs (both FOSS and proprietary) come from, anyway?
14 In addition, in early 2011, I was invited to deliver a critique of some of the assumptions
underlying transactional memory [McK11e]. The audience was surprisingly non-hostile,

though perhaps they were taking it easy on me due to the fact that I was heavily jet-lagged
while giving the presentation.
v2023.06.11a
Table 17.1: Comparison of Locking and HTM ( Advantage , Disadvantage , Strong

Disadvantage )
Locking Hardware Transactional Memory

Basic Idea Allow only one thread at a time to access a given Cause a given operation over a set of objects to
set of objects. execute atomically.
Scope Handles all operations. Handles revocable operations.
Irrevocable operations force fallback (typically to
locking).
Composability Limited by deadlock. Limited by irrevocable operations, transaction size,
and deadlock (assuming lock-based fallback code).
Scalability & Per- Data must be partitionable to avoid lock contention. Data must be partitionable to avoid conflicts.
formance
Partioning must typically be fixed at design time. Dynamic adjustment of partitioning carried out
automatically down to cacheline boundaries.
Partitioning required for fallbacks (less important
for rare fallbacks).
Locking primitives typically result in expensive Transactions begin/end instructions typically do
cache misses and memory-barrier instructions. not result in cache misses, but do have memory-
ordering and overhead consequences.
Contention effects are focused on acquisition and Contention aborts conflicting transactions, even if
release, so that the critical section runs at full speed. they have been running for a long time.
Privatization operations are simple, intuitive, per- Privatized data contributes to transaction size.
formant, and scalable.
Hardware Support Commodity hardware suffices. New hardware required (and is starting to become
available).
Performance is insensitive to cache-geometry de- Performance depends critically on cache geometry.
tails.
Software Support APIs exist, large body of code and experience, APIs emerging, little experience outside of DBMS,
debuggers operate naturally. breakpoints mid-transaction can be problematic.
Interaction With Long experience of successful interaction. Just beginning investigation of interaction.
Other Mechanisms
Practical Apps Yes. Yes.
Wide Applicability Yes. Jury still out.
v2023.06.11a
Table 17.2: Comparison of Locking (Augmented by RCU or Hazard Pointers) and

HTM ( Advantage , Disadvantage , Strong Disadvantage )
Locking with Userspace RCU or Hazard Pointers Hardware Transactional Memory

Basic Idea Allow only one thread at a time to access a given set Cause a given operation over a set of objects to execute
of objects. atomically.
Scope Handles all operations. Handles revocable operations.
Irrevocable operations force fallback (typically to lock-
ing).
Composability Readers limited only by grace-period-wait operations. Limited by irrevocable operations, transaction size,
and deadlock. (Assuming lock-based fallback code.)
Updaters limited by deadlock. Readers reduce dead-
lock.
Scalability & Per- Data must be partitionable to avoid lock contention Data must be partitionable to avoid conflicts.
formance among updaters.
Partitioning not needed for readers.
Partitioning for updaters must typically be fixed at Dynamic adjustment of partitioning carried out auto-
design time. matically down to cacheline boundaries.
Partitioning not needed for readers. Partitioning required for fallbacks (less important for
rare fallbacks).
Updater locking primitives typically result in expensive Transactions begin/end instructions typically do not
cache misses and memory-barrier instructions. result in cache misses, but do have memory-ordering
and overhead consequences.
Update-side contention effects are focused on acquisi- Contention aborts conflicting transactions, even if they
tion and release, so that the critical section runs at full have been running for a long time.
speed.
Readers do not contend with updaters or with each
other.
Read-side primitives are typically bounded wait-free Read-only transactions subject to conflicts and roll-
with low overhead. (Lock-free with low overhead for backs. No forward-progress guarantees other than
hazard pointers.) those supplied by fallback code.
Privatization operations are simple, intuitive, perfor- Privatized data contributes to transaction size.
mant, and scalable when data is visible only to updaters.
Privatization operations are expensive (though still
intuitive and scalable) for reader-visible data.
Hardware Support Commodity hardware suffices. New hardware required (and is starting to become
available).
Performance is insensitive to cache-geometry details. Performance depends critically on cache geometry.
Software Support APIs exist, large body of code and experience, debug- APIs emerging, little experience outside of DBMS,
gers operate naturally. breakpoints mid-transaction can be problematic.
Interaction With Long experience of successful interaction. Just beginning investigation of interaction.
Other Mechanisms
Practical Apps Yes. Yes.
Wide Applicability Yes. Jury still out.
v2023.06.11a
For example, deadlock can be avoided in many cases by using reference counts,
hazard pointers, or RCU to protect data structures, particularly for read-only critical
sections [Mic04a, HLM02, DMS+ 12, GMTW08, HMBW07]. These approaches also
reduce the need to partition data structures, as was seen in Chapter 10. RCU further
provides contention-free bounded wait-free read-side primitives [MS98a, DMS+ 12],
while hazard pointers provides lock-free read-side primitives [Mic02, HLM02, Mic04a].
Adding these considerations to Table 17.1 results in the updated comparison between
augmented locking and HTM shown in Table 17.2. A summary of the differences
between the two tables is as follows:
1. Use of non-blocking read-side mechanisms alleviates deadlock issues.
2. Read-side mechanisms such as hazard pointers and RCU can operate efficiently on
non-partitionable data.
3. Hazard pointers and RCU do not contend with each other or with updaters, allowing
excellent performance and scalability for read-mostly workloads.
4. Hazard pointers and RCU provide forward-progress guarantees (lock freedom and
bounded wait-freedom, respectively).
5. Privatization operations for hazard pointers and RCU are straightforward.
For those with good eyesight, Table 17.3 combines Tables 17.1 and 17.2.
Quick Quiz 17.15: Tables 17.1 and 17.2 state that hardware is only starting to become
available. But hasn’t HTM hardware support been widely available for almost a full decade?
Of course, it is also possible to augment HTM, as discussed in the next section.
17.3.4 Where Does HTM Best Fit In?

Although it will likely be some time before HTM’s area of applicability can be as crisply
delineated as that shown for RCU in Figure 9.33 on page 276, that is no reason not to
start moving in that direction.
HTM seems best suited to update-heavy workloads involving relatively small changes
to disparate portions of relatively large in-memory data structures running on large
multiprocessors, as this meets the size restrictions of current HTM implementations
while minimizing the probability of conflicts and attendant aborts and rollbacks. This
scenario is also one that is relatively difficult to handle given current synchronization
primitives.
Use of locking in conjunction with HTM seems likely to overcome HTM’s difficulties
with irrevocable operations, while use of RCU or hazard pointers might alleviate HTM’s
transaction-size limitations for read-only operations that traverse large fractions of the
data structure [PMDY20]. Current HTM implementations unconditionally abort an
update transaction that conflicts with an RCU or hazard-pointer reader, but perhaps future
HTM implementations will interoperate more smoothly with these synchronization
mechanisms. In the meantime, the probability of an update conflicting with a large RCU
or hazard-pointer read-side critical section should be much smaller than the probability
v2023.06.11a
Table 17.3: Comparison of Locking (Plain and Augmented) and HTM ( Advantage , Disadvantage , Strong Disadvantage )
Locking Locking with Userspace RCU or Hazard Pointers Hardware Transactional Memory
Basic Idea Allow only one thread at a time to access a given set Allow only one thread at a time to access a given set Cause a given operation over a set of objects to execute
of objects. of objects. atomically.
Scope Handles all operations. Handles all operations. Handles revocable operations.
Irrevocable operations force fallback (typically to lock-
ing).
Composability Limited by deadlock. Readers limited only by grace-period-wait operations. Limited by irrevocable operations, transaction size,
and deadlock. (Assuming lock-based fallback code.)
Updaters limited by deadlock. Readers reduce dead-
lock.
Scalability & Per- Data must be partitionable to avoid lock contention. Data must be partitionable to avoid lock contention Data must be partitionable to avoid conflicts.
formance among updaters.
Partitioning not needed for readers.
Partitioning must typically be fixed at design time. Partitioning for updaters must typically be fixed at Dynamic adjustment of partitioning carried out auto-
design time. matically down to cacheline boundaries.
Partitioning not needed for readers. Partitioning required for fallbacks (less important for
rare fallbacks).
Locking primitives typically result in expensive cache Updater locking primitives typically result in expensive Transactions begin/end instructions typically do not
misses and memory-barrier instructions. cache misses and memory-barrier instructions. result in cache misses, but do have memory-ordering
and overhead consequences.
Contention effects are focused on acquisition and re- Update-side contention effects are focused on acquisi- Contention aborts conflicting transactions, even if they
lease, so that the critical section runs at full speed. tion and release, so that the critical section runs at full have been running for a long time.
speed.
Readers do not contend with updaters or with each
17.3. HARDWARE TRANSACTIONAL MEMORY
other.
Read-side primitives are typically bounded wait-free Read-only transactions subject to conflicts and roll-
with low overhead. (Lock-free with low overhead for backs. No forward-progress guarantees other than
hazard pointers.) those supplied by fallback code.
Privatization operations are simple, intuitive, perfor- Privatization operations are simple, intuitive, perfor- Privatized data contributes to transaction size.
mant, and scalable. mant, and scalable when data is visible only to updaters.
Privatization operations are expensive (though still
intuitive and scalable) for reader-visible data.
Hardware Support Commodity hardware suffices. Commodity hardware suffices. New hardware required (and is starting to become
available).
Performance is insensitive to cache-geometry details. Performance is insensitive to cache-geometry details. Performance depends critically on cache geometry.
Software Support APIs exist, large body of code and experience, debug- APIs exist, large body of code and experience, debug- APIs emerging, little experience outside of DBMS,
611
gers operate naturally. gers operate naturally. breakpoints mid-transaction can be problematic.
Interaction With Long experience of successful interaction. Long experience of successful interaction. Just beginning investigation of interaction.
Other Mechanisms
Practical Apps Yes. Yes. Yes.
Wide Applicability Yes. Yes. Jury still out.
v2023.06.11a
of conflicting with the equivalent read-only transaction.15 Nevertheless, it is quite

possible that a steady stream of RCU or hazard-pointer readers might starve updaters
due to a corresponding steady stream of conflicts. This vulnerability could be eliminated
(at significant hardware cost and complexity) by giving extra-transactional reads the
pre-transaction copy of the memory location being loaded.
The fact that HTM transactions must have fallbacks might in some cases force static
partitionability of data structures back onto HTM. This limitation might be alleviated
if future HTM implementations provide forward-progress guarantees, which might
eliminate the need for fallback code in some cases, which in turn might allow HTM to
be used efficiently in situations with higher conflict probabilities.
In short, although HTM is likely to have important uses and applications, it is another
tool in the parallel programmer’s toolbox, not a replacement for the toolbox in its
entirety.
17.3.5 Potential Game Changers

Game changers that could greatly increase the need for HTM include the following:
1. Forward-progress guarantees.
2. Transaction-size increases.
3. Improved debugging support.
4. Weak atomicity.
These are expanded upon in the following sections.
17.3.5.1 Forward-Progress Guarantees

As was discussed in Section 17.3.2.4, current HTM implementations lack forward-
progress guarantees, which requires that fallback software is available to handle HTM
failures. Of course, it is easy to demand guarantees, but not always easy to provide them.
In the case of HTM, obstacles to guarantees can include cache size and associativity,
TLB size and associativity, transaction duration and interrupt frequency, and scheduler
implementation.
Cache size and associativity was discussed in Section 17.3.2.1, along with some
research intended to work around current limitations. However, HTM forward-progress
guarantees would come with size limits, large though these limits might one day be. So
why don’t current HTM implementations provide forward-progress guarantees for small
transactions, for example, limited to the associativity of the cache? One potential reason
might be the need to deal with hardware failure. For example, a failing cache SRAM
cell might be handled by deactivating the failing cell, thus reducing the associativity of
the cache and therefore also the maximum size of transactions that can be guaranteed
forward progress. Given that this would simply decrease the guaranteed transaction
size, it seems likely that other reasons are at work. Perhaps providing forward progress
guarantees on production-quality hardware is more difficult than one might think, an
15 It is quite ironic that strictly transactional mechanisms are appearing in shared-memory
systems at just about the time that NoSQL databases are relaxing the traditional database-
application reliance on strict transactions. Nevertheless, HTM has in fact realized the
ease-of-use promise of TM, albeit for black-hat attacks on the Linux kernel’s address-space
randomization defense mechanism [JLK16a, JLK16b].
v2023.06.11a
entirely plausible explanation given the difficulty of making forward-progress guarantees

in software. Moving a problem from software to hardware does not necessarily make it
easier to solve [JSG12].
Given a physically tagged and indexed cache, it is not enough for the transaction to fit
in the cache. Its address translations must also fit in the TLB. Any forward-progress
guarantees must therefore also take TLB size and associativity into account.
Given that interrupts, traps, and exceptions abort transactions in current HTM
implementations, it is necessary that the execution duration of a given transaction
be shorter than the expected interval between interrupts. No matter how little data
a given transaction touches, if it runs too long, it will be aborted. Therefore, any
forward-progress guarantees must be conditioned not only on transaction size, but also
on transaction duration.
Forward-progress guarantees depend critically on the ability to determine which of
several conflicting transactions should be aborted. It is all too easy to imagine an endless
series of transactions, each aborting an earlier transaction only to itself be aborted by a
later transactions, so that none of the transactions actually commit. The complexity of
conflict handling is evidenced by the large number of HTM conflict-resolution policies
that have been proposed [ATC+ 11, LS11]. Additional complications are introduced
by extra-transactional accesses, as noted by Blundell [BLM06]. It is easy to blame
the extra-transactional accesses for all of these problems, but the folly of this line of
thinking is easily demonstrated by placing each of the extra-transactional accesses into
its own single-access transaction. It is the pattern of accesses that is the issue, not
whether or not they happen to be enclosed in a transaction.
Finally, any forward-progress guarantees for transactions also depend on the scheduler,
which must let the thread executing the transaction run long enough to successfully
commit.
So there are significant obstacles to HTM vendors offering forward-progress guaran-
tees. However, the impact of any of them doing so would be enormous. It would mean
that HTM transactions would no longer need software fallbacks, which would mean
that HTM could finally deliver on the TM promise of deadlock elimination.
However, in late 2012, the IBM Mainframe announced an HTM implementation that
includes constrained transactions in addition to the usual best-effort HTM implementa-
tion [JSG12]. A constrained transaction starts with the tbeginc instruction instead of
the tbegin instruction that is used for best-effort transactions. Constrained transactions
are guaranteed to always complete (eventually), so if a transaction aborts, rather than
branching to a fallback path (as is done for best-effort transactions), the hardware instead
restarts the transaction at the tbeginc instruction.
The Mainframe architects needed to take extreme measures to deliver on this forward-
progress guarantee. If a given constrained transaction repeatedly fails, the CPU might
disable branch prediction, force in-order execution, and even disable pipelining. If
the repeated failures are due to high contention, the CPU might disable speculative
fetches, introduce random delays, and even serialize execution of the conflicting CPUs.
“Interesting” forward-progress scenarios involve as few as two CPUs or as many as one
hundred CPUs. Perhaps these extreme measures provide some insight as to why other
CPUs have thus far refrained from offering constrained transactions.
As the name implies, constrained transactions are in fact severely constrained:
1. The maximum data footprint is four blocks of memory, where each block can be
no larger than 32 bytes.
v2023.06.11a
2. The maximum code footprint is 256 bytes.
3. If a given 4K page contains a constrained transaction’s code, then that page may
not contain that transaction’s data.
4. The maximum number of assembly instructions that may be executed is 32.
5. Backwards branches are forbidden.
Nevertheless, these constraints support a number of important data structures,

including linked lists, stacks, queues, and arrays. Constrained HTM therefore seems
likely to become an important tool in the parallel programmer’s toolbox.
Note that these forward-progress guarantees need not be absolute. For example,
suppose that a use of HTM uses a global lock as fallback. Assuming that the fallback
mechanism has been carefully designed to avoid the “lemming effect” discussed in
Section 17.3.2.3, then if HTM rollbacks are sufficiently infrequent, the global lock will
not be a bottleneck. That said, the larger the system, the longer the critical sections,
and the longer the time required to recover from the “lemming effect”, the more rare
“sufficiently infrequent” needs to be.
17.3.5.2 Transaction-Size Increases

Forward-progress guarantees are important, but as we saw, they will be conditional
guarantees based on transaction size and duration. There has been some progress,
for example, some commercially available HTM implementations use approximation
techniques to support extremely large HTM read sets [RD12]. For another example,
POWER8 HTM supports suspended transations, which avoid adding irrelevant accesses
to the suspended transation’s read and write sets [LGW+ 15]. This capability has been
used to produce a high performance reader-writer lock [FIMR16].
It is important to note that even small-sized guarantees will be quite useful. For
example, a guarantee of two cache lines is sufficient for a stack, queue, or dequeue.
However, larger data structures require larger guarantees, for example, traversing a tree
in order requires a guarantee equal to the number of nodes in the tree. Therefore, even
modest increases in the size of the guarantee also increases the usefulness of HTM,
thereby increasing the need for CPUs to either provide it or provide good-and-sufficient
workarounds.
17.3.5.3 Improved Debugging Support

Another inhibitor to transaction size is the need to debug the transactions. The problem
with current mechanisms is that a single-step exception aborts the enclosing transaction.
There are a number of workarounds for this issue, including emulating the processor
(slow!), substituting STM for HTM (slow and slightly different semantics!), playback
techniques using repeated retries to emulate forward progress (strange failure modes!),
and full support of debugging HTM transactions (complex!).
Should one of the HTM vendors produce an HTM system that allows straightforward
use of classical debugging techniques within transactions, including breakpoints, single
stepping, and print statements, this will make HTM much more compelling. Some
transactional-memory researchers started to recognize this problem in 2013, with at
least one proposal involving hardware-assisted debugging facilities [GKP13]. Of course,
this proposal depends on readily available hardware gaining such facilities [Hay20,
v2023.06.11a
Int20b]. Worse yet, some cutting-edge debugging facilities are incompatible with
HTM [OHOC20].
17.3.5.4 Weak Atomicity
Given that HTM is likely to face some sort of size limitations for the foreseeable future,
it will be necessary for HTM to interoperate smoothly with other mechanisms. HTM’s
interoperability with read-mostly mechanisms such as hazard pointers and RCU would
be improved if extra-transactional reads did not unconditionally abort transactions with
conflicting writes—instead, the read could simply be provided with the pre-transaction
value. In this way, hazard pointers and RCU could be used to allow HTM to handle
larger data structures and to reduce conflict probabilities.
This is not necessarily simple, however. The most straightforward way of imple-
menting this requires an additional state in each cache line and on the bus, which is a
non-trivial added expense. The benefit that goes along with this expense is permitting
large-footprint readers without the risk of starving updaters due to continual conflicts.
An alternative approach, applied to great effect to binary search trees by Siakavaras et
al. [SNGK17], is to use RCU for read-only traversals and HTM only for the actual updates
themselves. This combination outperformed other transactional-memory techniques
by up to 220 %, a speedup similar to that observed by Howard and Walpole [HW11]
when they combined RCU with STM. In both cases, the weak atomicity is implemented
in software rather than in hardware. It would nevertheless be interesting to see what
additional speedups could be obtained by implementing weak atomicity in both hardware
and software.
17.3.6 Conclusions
Although current HTM implementations have delivered real performance benefits

in some situations, they also have significant shortcomings. The most significant
shortcomings appear to be limited transaction sizes, the need for conflict handling, the
need for aborts and rollbacks, the lack of forward-progress guarantees, the inability to
handle irrevocable operations, and subtle semantic differences from locking. There are
also reasons for lingering concerns surrounding HTM-implementation reliability [JSG12,
Was14, Int20a, Int21, Lar21, Int20c].
Some of these shortcomings might be alleviated in future implementations, but it
appears that there will continue to be a strong need to make HTM work well with the many
other types of synchronization mechanisms, as noted earlier [MMW07, MMTW10].
Although there has been some work using HTM with RCU [SNGK17, SBN+ 20, GGK18,
PMDY20], there has been little evidence of progress towards HTM work better with
RCU and with other deferred-reclamation mechanisms.
In short, current HTM implementations appear to be welcome and useful additions
to the parallel programmer’s toolbox, and much interesting and challenging work is
required to make use of them. However, they cannot be considered to be a magic wand
with which to wave away all parallel-programming problems.
v2023.06.11a
17.4 Formal Regression Testing?
Theory without experiments: Have we gone too far?

Michael Mitzenmacher
Formal verification has long proven useful in a number of production environ-

ments [LBD+ 04, BBC+ 10, Coo18, SAE+ 18, DFLO19]. However, it is an open question
as to whether hard-core formal verification will ever be included in the automated
regression-test suites used for continuous integration within complex concurrent code-
bases, such as the Linux kernel. Although there is already a proof of concept for
Linux-kernel SRCU [Roy17], this test is for a small portion of one of the simplest RCU
implementations, and has proven difficult to keep it current with the ever-changing
Linux kernel. It is therefore worth asking what would be required to incorporate formal
verification as first-class members of the Linux kernel’s regression tests.
The following list is a good start [McK15a, slide 34]:
1. Any required translation must be automated.
2. The environment (including memory ordering) must be correctly handled.
3. The memory and CPU overhead must be acceptably modest.
4. Specific information leading to the location of the bug must be provided.
5. Information beyond the source code and inputs must be modest in scope.
6. The bugs located must be relevant to the code’s users.
This list builds on, but is somewhat more modest than, Richard Bornat’s dictum:
“Formal-verification researchers should verify the code that developers write, in the
language they write it in, running in the environment that it runs in, as they write it.”
The following sections discuss each of the above requirements, followed by a section
presenting a scorecard of how well a few tools stack up against these requirements.
Quick Quiz 17.16: This list is ridiculously utopian! Why not stick to the current state of the
formal-verification art?
17.4.1 Automatic Translation

Although Promela and spin are invaluable design aids, if you need to formally
regression-test your C-language program, you must hand-translate to Promela each time
you would like to re-verify your code. If your code happens to be in the Linux kernel,
which releases every 60–90 days, you will need to hand-translate from four to six times
each year. Over time, human error will creep in, which means that the verification won’t
match the source code, rendering the verification useless. Repeated verification clearly
requires either that the formal-verification tooling input your code directly, or that there
be bug-free automatic translation of your code to the form required for verification.
PPCMEM and herd can in theory directly input assembly language and C++ code,
but these tools work only on very small litmus tests, which normally means that you
must extract the core of your mechanism—by hand. As with Promela and spin, both
v2023.06.11a
17.4. FORMAL REGRESSION TESTING? 617
PPCMEM and herd are extremely useful, but they are not well-suited for regression
suites.
In contrast, cbmc and Nidhugg can input C programs of reasonable (though still
quite limited) size, and if their capabilities continue to grow, could well become
excellent additions to regression suites. The Coverity static-analysis tool also inputs
C programs, and of very large size, including the Linux kernel. Of course, Coverity’s
static analysis is quite simple compared to that of cbmc and Nidhugg. On the other
hand, Coverity had an all-encompassing definition of “C program” that posed special
challenges [BBC+ 10]. Amazon Web Services uses a variety of formal-verification tools,
including cbmc, and applies some of these tools to regression testing [Coo18]. Google
uses a number of relatively simple static analysis tools directly on large Java code bases,
which are arguably less diverse than C code bases [SAE+ 18]. Facebook uses more
aggressive forms of formal verification against its code bases, including analysis of
concurrency [DFLO19, O’H19], though not yet on the Linux kernel. Finally, Microsoft
has long used static analysis on its code bases [LBD+ 04].
Given this list, it is clearly possible to create sophisticated formal-verification tools
that directly consume production-quality source code.
However, one shortcoming of taking C code as input is that it assumes that the
compiler is correct. An alternative approach is to take the binary produced by the C
compiler as input, thereby accounting for any relevant compiler bugs. This approach
has been used in a number of verification efforts, perhaps most notably by the SEL4
project [SM13].
Quick Quiz 17.17: Given the groundbreaking nature of the various verifiers used in the SEL4
project, why doesn’t this chapter cover them in more depth?
However, verifying directly from either the source or binary both have the advantage of
eliminating human translation errors, which is critically important for reliable regression
testing.
This is not to say that tools with special-purpose languages are useless. On the
contrary, they can be quite helpful for design-time verification, as was discussed in
Chapter 12. However, such tools are not particularly helpful for automated regression
testing, which is in fact the topic of this section.
17.4.2 Environment
It is critically important that formal-verification tools correctly model their environment.
One all-too-common omission is the memory model, where a great many formal-
verification tools, including Promela/spin, are restricted to sequential consistency. The
QRCU experience related in Section 12.1.4.6 is an important cautionary tale.
Promela and spin assume sequential consistency, which is not a good match for
modern computer systems, as was seen in Chapter 15. In contrast, one of the great
strengths of PPCMEM and herd is their detailed modeling of various CPU families
memory models, including x86, Arm, Power, and, in the case of herd, a Linux-kernel
memory model [AMM+ 18], which was accepted into Linux-kernel version v4.17.
The cbmc and Nidhugg tools provide some ability to select memory models, but
do not provide the variety that PPCMEM and herd do. However, it is likely that the
larger-scale tools will adopt a greater variety of memory models as time goes on.
In the longer term, it would be helpful for formal-verification tools to include
I/O [MDR16], but it may be some time before this comes to pass.
v2023.06.11a
Table 17.4: Emulating Locking: Performance (s)
# Threads Locking cmpxchg_acquire

2 0.004 0.022
3 0.041 0.743
4 0.374 59.565
5 4.905
Nevertheless, tools that fail to match the environment can still be useful. For example,
a great many concurrency bugs would still be bugs on a mythical sequentially consistent
system, and these bugs could be located by a tool that over-approximates the system’s
memory model with sequential consistency. Nevertheless, these tools will fail to find
bugs involving missing memory-ordering directives, as noted in the aforementioned
cautionary tale of Section 12.1.4.6.
17.4.3 Overhead
Almost all hard-core formal-verification tools are exponential in nature, which might
seem discouraging until you consider that many of the most interesting software
questions are in fact undecidable. However, there are differences in degree, even among
exponentials.
PPCMEM by design is unoptimized, in order to provide greater assurance that the
memory models of interest are accurately represented. The herd tool optimizes more
aggressively, as described in Section 12.3, and is thus orders of magnitude faster than
PPCMEM. Nevertheless, both PPCMEM and herd target very small litmus tests rather
than larger bodies of code.
In contrast, Promela/spin, cbmc, and Nidhugg are designed for (somewhat) larger bod-
ies of code. Promela/spin was used to verify the Curiosity rover’s filesystem [GHH+ 14]
and, as noted earlier, both cbmc and Nidhugg were appled to Linux-kernel RCU.
If advances in heuristics continue at the rate of the past three decades, we can look
forward to large reductions in overhead for formal verification. That said, combinatorial
explosion is still combinatorial explosion, which would be expected to sharply limit
the size of programs that could be verified, with or without continued improvements in
heuristics.
However, the flip side of combinatorial explosion is Philip II of Macedon’s timeless
advice: “Divide and rule.” If a large program can be divided and the pieces verified,
the result can be combinatorial implosion [McK11e]. One natural place to divide is on
API boundaries, for example, those of locking primitives. One verification pass can
then verify that the locking implementation is correct, and additional verification passes
can verify correct use of the locking APIs.
The performance benefits of this approach can be demonstrated using the Linux-kernel
memory model [AMM+ 18]. This model provides spin_lock() and spin_unlock()
primitives, but these primitives can also be emulated using cmpxchg_acquire()
and smp_store_release(), as shown in Listing 17.2 (C-SB+l-o-o-u+l-o-o-
*u.litmus and C-SB+l-o-o-u+l-o-o-u*-C.litmus). Table 17.4 compares the
performance and scalability of using the model’s spin_lock() and spin_unlock()
against emulating these primitives as shown in the listing. The difference is not
v2023.06.11a
Listing 17.2: Emulating Locking with cmpxchg_acquire()

1 C C-SB+l-o-o-u+l-o-o-u-C
2
3 {}
4
5 P0(int *sl, int *x0, int *x1)
6 {
7 int r2;
8 int r1;
9
10 r2 = cmpxchg_acquire(sl, 0, 1);
13 smp_store_release(sl, 0);
14 }
15
16 P1(int *sl, int *x0, int *x1)
17 {
18 int r2;
19 int r1;
20
21 r2 = cmpxchg_acquire(sl, 0, 1);
24 smp_store_release(sl, 0);
25 }
26
27 filter (0:r2=0 /\ 1:r2=0)
28 exists (0:r1=0 /\ 1:r1=0)
insignificant: At four processes, the model is more than two orders of magnitude faster
than emulation!
Quick Quiz 17.18: Why bother with a separate filter command on line 27 of Listing 17.2
instead of just adding the condition to the exists clause? And wouldn’t it be simpler to use
xchg_acquire() instead of cmpxchg_acquire()?
It would of course be quite useful for tools to automatically divide up large programs,
verify the pieces, and then verify the combinations of pieces. In the meantime,
verification of large programs will require significant manual intervention. This
intervention will preferably mediated by scripting, the better to reliably carry out
repeated verifications on each release, and preferably eventually in a manner well-suited
for continuous integration. And Facebook’s Infer tool has taken important steps towards
doing just that, via compositionality and abstraction [BGOS18, DFLO19].
In any case, we can expect formal-verification capabilities to continue to increase over
time, and any such increases will in turn increase the applicability of formal verification
to regression testing.
17.4.4 Locate Bugs

Any software artifact of any size contains bugs. Therefore, a formal-verification tool
that reports only the presence or absence of bugs is not particularly useful. What is
needed is a tool that gives at least some information as to where the bug is located and
the nature of that bug.
The cbmc output includes a traceback mapping back to the source code, similar to
Promela/spin’s, as does Nidhugg. Of course, these tracebacks can be quite long, and
analyzing them can be quite tedious. However, doing so is usually quite a bit faster and
more pleasant than locating bugs the old-fashioned way.
v2023.06.11a
In addition, one of the simplest tests of formal-verification tools is bug injection.

After all, not only could any of us write printf("VERIFIED\n"), but the plain fact is
that developers of formal-verification tools are just as bug-prone as are the rest of us.
Therefore, formal-verification tools that just proclaim that a bug exists are fundamentally
less trustworthy because it is more difficult to verify them on real-world code.
All that aside, people writing formal-verification tools are permitted to leverage
existing tools. For example, a tool designed to determine only the presence or absence
of a serious but rare bug might leverage bisection. If an old version of the program
under test did not contain the bug, but a new version did, then bisection could be used to
quickly locate the commit that inserted the bug, which might be sufficient information
to find and fix the bug. Of course, this sort of strategy would not work well for common
bugs because in this case bisection would fail due to all commits having at least one
instance of the common bug.
Therefore, the execution traces provided by many formal-verification tools will
continue to be valuable, particularly for complex and difficult-to-understand bugs. In
addition, recent work applies incorrectness-logic formalism reminiscent of the traditional
Hoare logic used for full-up correctness proofs, but with the sole purpose of finding
bugs [O’H19].
17.4.5 Minimal Scaffolding

In the old days, formal-verification researchers demanded a full specification against
which the software would be verified. Unfortunately, a mathematically rigorous
specification might well be larger than the actual code, and each line of specification is
just as likely to contain bugs as is each line of code. A formal verification effort proving
that the code faithfully implemented the specification would be a proof of bug-for-bug
compatibility between the two, which might not be all that helpful.
Worse yet, the requirements for a number of software artifacts, including Linux-kernel
RCU, are empirical in nature [McK15h, McK15e, McK15f].16 For this common type of
software, a complete specification is a polite fiction. Nor are complete specifications any
less fictional for hardware, as was made clear by the late-2017 Meltdown and Spectre
side-channel attacks [Hor18].
This situation might cause one to give up all hope of formal verification of real-world
software and hardware artifacts, but it turns out that there is quite a bit that can be done.
For example, design and coding rules can act as a partial specification, as can assertions
contained in the code. And in fact formal-verification tools such as cbmc and Nidhugg
both check for assertions that can be triggered, implicitly treating these assertions as
part of the specification. However, the assertions are also part of the code, which makes
it less likely that they will become obsolete, especially if the code is also subjected
to stress tests.17 The cbmc tool also checks for array-out-of-bound references, thus
implicitly adding them to the specification. The aforementioned incorrectness logic can
also be thought of as using an implicit bugs-not-present specification [O’H19].
This implicit-specification approach makes quite a bit of sense, particularly if you
look at formal verification not as a full proof of correctness, but rather an alternative
form of validation with a different set of strengths and weaknesses than the common
case, that is, testing. From this viewpoint, software will always have bugs, and therefore
any tool of any kind that helps to find those bugs is a very good thing indeed.
16 Or, in formal-verification parlance, Linux-kernel RCU has an incomplete specification.

17 And you do stress-test your code, don’t you?
v2023.06.11a
17.4.6 Relevant Bugs

Finding bugs—and fixing them—is of course the whole point of any type of validation
effort. Clearly, false positives are to be avoided. But even in the absence of false
positives, there are bugs and there are bugs.
For example, suppose that a software artifact had exactly 100 remaining bugs, each
of which manifested on average once every million years of runtime. Suppose further
that an omniscient formal-verification tool located all 100 bugs, which the developers
duly fixed. What happens to the reliability of this software artifact?
The answer is that the reliability decreases.
To see this, keep in mind that historical experience indicates that about 7 % of fixes
introduce a new bug [BJ12]. Therefore, fixing the 100 bugs, which had a combined
mean time to failure (MTBF) of about 10,000 years, will introduce seven more bugs.
Historical statistics indicate that each new bug will have an MTBF much less than 70,000
years. This in turn suggests that the combined MTBF of these seven new bugs will most
likely be much less than 10,000 years, which in turn means that the well-intentioned
fixing of the original 100 bugs actually decreased the reliability of the overall software.
Quick Quiz 17.19: How do we know that the MTBFs of known bugs is a good estimate of the
MTBFs of bugs that have not yet been located?
Quick Quiz 17.20: But the formal-verification tools should immediately find all the bugs
introduced by the fixes, so why is this a problem?
Worse yet, imagine another software artifact with one bug that fails once every day on
average and 99 more that fail every million years each. Suppose that a formal-verification
tool located the 99 million-year bugs, but failed to find the one-day bug. Fixing the 99
bugs located will take time and effort, decrease reliability, and do nothing at all about
the pressing each-day failure that is likely causing embarrassment and perhaps much
worse besides.
Therefore, it would be best to have a validation tool that preferentially located the
most troublesome bugs. However, as noted in Section 17.4.4, it is permissible to
leverage additional tools. One powerful tool is none other than plain old testing. Given
knowledge of the bug, it should be possible to construct specific tests for it, possibly
also using some of the techniques described in Section 11.6.4 to increase the probability
of the bug manifesting. These techniques should allow calculation of a rough estimate
of the bug’s raw failure rate, which could in turn be used to prioritize bug-fix efforts.
Quick Quiz 17.21: But many formal-verification tools can only find one bug at a time, so
that each bug must be fixed before the tool can locate the next. How can bug-fix efforts be
prioritized given such a tool?
There has been some recent formal-verification work that prioritizes executions
having fewer preemptions, under that reasonable assumption that smaller numbers of
preemptions are more likely.
Identifying relevant bugs might sound like too much to ask, but it is what is really
required if we are to actually increase software reliability.
17.4.7 Formal Regression Scorecard

Table 17.5 shows a rough-and-ready scorecard for the formal-verification tools covered
in this chapter. Shorter wavelengths are better than longer wavelengths.
v2023.06.11a
Table 17.5: Formal Regression Scorecard
Promela PPCMEM herd cbmc Nidhugg

(1) Automated
(2) Environment (MM) (MM) (MM)
(3) Overhead (SAT)
(4) Locate Bugs
(5) Minimal Scaffolding
(6) Relevant Bugs ??? ??? ??? ??? ???
Promela requires hand translation and supports only sequential consistency, so its
first two cells are red. It has reasonable overhead (for formal verification, anyway) and
provides a traceback, so its next two cells are yellow. Despite requiring hand translation,
Promela handles assertions in a natural way, so its fifth cell is green.
PPCMEM usually requires hand translation due to the small size of litmus tests that
it supports, so its first cell is orange. It handles several memory models, so its second
cell is green. Its overhead is quite high, so its third cell is red. It provides a graphical
display of relations among operations, which is not as helpful as a traceback, but is still
quite useful, so its fourth cell is yellow. It requires constructing an exists clause and
cannot take intra-process assertions, so its fifth cell is also yellow.
The herd tool has size restrictions similar to those of PPCMEM, so herd’s first cell
is also orange. It supports a wide variety of memory models, so its second cell is blue.
It has reasonable overhead, so its third cell is yellow. Its bug-location and assertion
capabilities are quite similar to those of PPCMEM, so herd also gets yellow for the
next two cells.
The cbmc tool inputs C code directly, so its first cell is blue. It supports a few memory
models, so its second cell is yellow. It has reasonable overhead, so its third cell is also
yellow, however, perhaps SAT-solver performance will continue improving. It provides
a traceback, so its fourth cell is green. It takes assertions directly from the C code, so its
fifth cell is blue.
Nidhugg also inputs C code directly, so its first cell is also blue. It supports only a
couple of memory models, so its second cell is orange. Its overhead is quite low (for
formal-verification), so its third cell is green. It provides a traceback, so its fourth cell is
green. It takes assertions directly from the C code, so its fifth cell is blue.
So what about the sixth and final row? It is too early to tell how any of the tools do at
finding the right bugs, so they are all yellow with question marks.
Quick Quiz 17.22: How would testing stack up in the scorecard shown in Table 17.5?
Quick Quiz 17.23: But aren’t there a great many more formal-verification systems than are
shown in Table 17.5?
Once again, please note that this table rates these tools for use in regression testing.
Just because many of them are a poor fit for regression testing does not at all mean that
they are useless, in fact, many of them have proven their worth many times over.18 Just
not for regression testing.
18 For but one example, Promela was used to verify the file system of none other than the
Curiosity Rover. Was your formal verification tool used on software that currently runs on
Mars???
v2023.06.11a
17.5. FUNCTIONAL PROGRAMMING FOR PARALLELISM 623
However, this might well change. After all, formal verification tools made impressive
strides in the 2010s. If that progress continues, formal verification might well become
an indispensable tool in the parallel programmer’s validation toolbox.
17.5 Functional Programming for Parallelism

The curious failure of functional programming for
parallel applications.
Malte Skarupke
When I took my first-ever functional-programming class in the early 1980s, the professor
asserted that the side-effect-free functional-programming style was well-suited to trivial
parallelization and analysis. Thirty years later, this assertion remains, but mainstream
production use of parallel functional languages is minimal, a state of affairs that might
not be entirely unrelated to professor’s additional assertion that programs should neither
maintain state nor do I/O. There is niche use of functional languages such as Erlang,
and multithreaded support has been added to several other functional languages, but
mainstream production usage remains the province of procedural languages such as C,
C++, Java, and Fortran (usually augmented with OpenMP, MPI, or coarrays).
This situation naturally leads to the question “If analysis is the goal, why not transform
the procedural language into a functional language before doing the analysis?” There
are of course a number of objections to this approach, of which I list but three:
1. Procedural languages often make heavy use of global variables, which can be
updated independently by different functions, or, worse yet, by multiple threads.
Note that Haskell’s monads were invented to deal with single-threaded global state,
and that multi-threaded access to global state inflicts additional violence on the
functional model.
2. Multithreaded procedural languages often use synchronization primitives such as

locks, atomic operations, and transactions, which inflict added violence upon the
functional model.
3. Procedural languages can alias function arguments, for example, by passing a

pointer to the same structure via two different arguments to the same invocation
of a given function. This can result in the function unknowingly updating that
structure via two different (and possibly overlapping) code sequences, which greatly
complicates analysis.
Of course, given the importance of global state, synchronization primitives, and

aliasing, clever functional-programming experts have proposed any number of attempts
to reconcile the function programming model to them, monads being but one case in
point.
Another approach is to compile the parallel procedural program into a functional
program, then to use functional-programming tools to analyze the result. But it is
possible to do much better than this, given that any real computation is a large finite-state
machine with finite input that runs for a finite time interval. This means that any real
program can be transformed into an expression, possibly albeit an impractically large
one [DHK12].
v2023.06.11a
However, a number of the low-level kernels of parallel algorithms transform into

expressions that are small enough to fit easily into the memories of modern computers.
If such an expression is coupled with an assertion, checking to see if the assertion
would ever fire becomes a satisfiability problem. Even though satisfiability problems
are NP-complete, they can often be solved in much less time than would be required to
generate the full state space. In addition, the solution time appears to be only weakly
dependent on the underlying memory model, so that algorithms running on weakly
ordered systems can also be checked [AKT13].
The general approach is to transform the program into single-static-assignment (SSA)
form, so that each assignment to a variable creates a separate version of that variable.
This applies to assignments from all the active threads, so that the resulting expression
embodies all possible executions of the code in question. The addition of an assertion
entails asking whether any combination of inputs and initial values can result in the
assertion firing, which, as noted above, is exactly the satisfiability problem.
One possible objection is that it does not gracefully handle arbitrary looping constructs.
However, in many cases, this can be handled by unrolling the loop a finite number
of times. In addition, perhaps some loops will also prove amenable to collapse via
inductive methods.
Another possible objection is that spinlocks involve arbitrarily long loops, and any
finite unrolling would fail to capture the full behavior of the spinlock. It turns out that
this objection is easily overcome. Instead of modeling a full spinlock, model a trylock
that attempts to obtain the lock, and aborts if it fails to immediately do so. The assertion
must then be crafted so as to avoid firing in cases where a spinlock aborted due to the
lock not being immediately available. Because the logic expression is independent of
time, all possible concurrency behaviors will be captured via this approach.
A final objection is that this technique is unlikely to be able to handle a full-sized
software artifact such as the millions of lines of code making up the Linux kernel. This
is likely the case, but the fact remains that exhaustive validation of each of the much
smaller parallel primitives within the Linux kernel would be quite valuable. And in
fact the researchers spearheading this approach have applied it to non-trivial real-world
code, including the Tree RCU implementation in the Linux kernel [LMKM16, KS17a].
It remains to be seen how widely applicable this technique is, but it is one of the
more interesting innovations in the field of formal verification. Although it might
well be that the functional-programming advocates are at long last correct in their
assertion of the inevitable dominance of functional programming, it is clearly the case
that this long-touted methodology is starting to see credible competition on its formal-
verification home turf. There is therefore continued reason to doubt the inevitability of
functional-programming dominance.
17.6 Summary
This chapter has taken a quick tour of a number of possible futures, including multicore,
transactional memory, formal verification as a regression test, and concurrent functional
programming. Any of these futures might come true, but it is more likely that, as in the
past, the future will be far stranger than we can possibly imagine.
v2023.06.11a
History is the sum total of things that could have
been avoided.
Konrad Adenauer
Chapter 18
Looking Forward and Back
You have arrived at the end of this book, well done! I hope that your journey was a
pleasant but challenging and worthwhile one.
For your editor and contributors, this is the end of the journey to the Second Edition,
but for those willing to join in, it is also the start of the journey to the Third Edition.
Either way, it is good to recap this past journey.
Chapter 1 covered what this book is about, along with some alternatives for those
interested in something other than low-level parallel programming.
Chapter 2 covered parallel-programming challenges and high-level approaches for
addressing them. It also touched on ways of avoiding these challenges while nevertheless
still gaining most of the benefits of parallelism.
Chapter 3 gave a high-level overview of multicore hardware, especially those aspects
that pose challenges for concurrent software. This chapter puts the blame for these
challenges where it belongs, very much on the laws of physics and rather less on
intransigent hardware architects and designers. However, there might be some things
that hardware architects and engineers can do, and this chapter discusses a few of them.
In the meantime, software architects and engineers must do their part to meet these
challenges, as discussed in the rest of the book.
Chapter 4 gave a quick overview of the tools of the low-level concurrency trade.
Chapter 5 then demonstrated use of those tools—and, more importantly, use of parallel-
programming design techniques—on the simple but surprisingly challenging task of
concurrent counting. So challenging, in fact, that a number of concurrent counting
algorithms are in common use, each specialized for a different use case.
Chapter 6 dug more deeply into the most important parallel-programming design
technique, namely partitioning the problem at the highest possible level. This chapter
also overviewed a number of points in this design space.
Chapter 7 expounded on that parallel-programming workhorse (and villain), locking.
This chapter covered a number of types of locking and presented some engineering
solutions to many well-known and aggressively advertised shortcomings of locking.
Chapter 8 discussed the uses of data ownership, where synchronization is supplied
by the association of a given data item with a specific thread. Where it applies, this
approach combines excellent performance and scalability with profound simplicity.
Chapter 9 showed how a little procrastination can greatly improve performance and
scalability, while in a surprisingly large number of cases also simplifying the code. A
number of the mechanisms presented in this chapter take advantage of the ability of
CPU caches to replicate read-only data, thus sidestepping the laws of physics that cruelly
625
v2023.06.11a
626 CHAPTER 18. LOOKING FORWARD AND BACK
limit the speed of light and the smallness of atoms. Chapter 10 looked at concurrent
data structures, with emphasis on hash tables, which have a long and honorable history
in parallel programs.
Chapter 11 dug into code-review and testing methods, and Chapter 12 overviewed
formal verification. Whichever side of the formal-verification/testing divide you might
be on, if code has not been thoroughly validated, it does not work. And that goes at
least double for concurrent code.
Chapter 13 presented a number of situations where combining concurrency mecha-
nisms with each other or with other design tricks can greatly ease parallel programmers’
lives. Chapter 14 looked at advanced synchronization methods, including lockless pro-
gramming, non-blocking synchronization, and parallel real-time computing. Chapter 15
dug into the critically important topic of memory ordering, presenting techniques and
tools to help you not only solve memory-ordering problems, but also to avoid them
completely. Chapter 16 presented a brief overview of the surprisingly important topic
of ease of use.
Last, but definitely not least, Chapter 17 expounded on a number of conflicting
visions of the future, including CPU-technology trends, transactional memory, hardware
transactional memory, use of formal verification in regression testing, and the long-
standing prediction that the future of parallel programming belongs to functional-
programming languages.
But now that we have recapped the contents of this Second Edition, how did this book
get started?
Paul’s parallel-programming journey started in earnest in 1990, when he joined
Sequent Computer Systems, Inc. Sequent used an apprenticeship-like program in which
newly hired engineers were placed in cubicles surrounded by experienced engineers,
who mentored them, reviewed their code, and gave copious quantities of advice on a
variety of topics. A few of the newly hired engineers were greatly helped by the fact
that there were no on-chip caches in those days, which meant that logic analyzers could
easily display a given CPU’s instruction stream and memory accesses, complete with
accurate timing information. Of course, the downside of this transparency was that
CPU core clock frequencies were 100 times slower than those of the twenty-first century.
Between apprenticeship and hardware performance transparency, these newly hired
engineers became productive parallel programmers within two or three months, and
some were doing ground-breaking work within a couple of years.
Sequent understood that its ability to quickly train new engineers in the mysteries of
parallelism was unusual, so it produced a slim volume that crystalized the company’s
parallel-programming wisdom [Seq88], which joined a pair of groundbreaking papers
that had been written a few years earlier [BK85, Inm85]. People already steeped in
these mysteries saluted this book and these papers, but novices were usually unable to
benefit much from them, invariably making highly creative and quite destructive errors
that were not explicitly prohibited by either the book or the papers.1 This situation of
course caused Paul to start thinking in terms of writing an improved book, but his efforts
during this time were limited to internal training materials and to published papers.
By the time Sequent was acquired by IBM in 1999, many of the world’s largest
database instances ran on Sequent hardware. But times change, and by 2001 many
of Sequent’s parallel programmers had shifted their focus to the Linux kernel. After
some initial reluctance, the Linux kernel community embraced concurrency both
enthusiastically and effectively [BWCM+ 10, McK12a], with many excellent innovations
1 “But why on earth would you do that???” “Well, why not?”
v2023.06.11a
627
and improvements from throughout the community. The thought of writing a book
occurred to Paul from time to time, but life was flowing fast, so he made no progress on
this project.
In 2006, Paul was invited to a conference on Linux scalability, and was granted the
privilege of asking the last question of panel of esteemed parallel-programming experts.
Paul began his question by noting that in the 15 years from 1991 to 2006, the price of a
parallel system had dropped from that of a house to that of a mid-range bicycle, and it was
clear that there was much more room for additional dramatic price decreases over the next
15 years extending to the year 2021. He also noted that decreasing price should result in
greater familiarity and faster progress in solving parallel-programming problems. This
led to his question: “In the year 2021, why wouldn’t parallel programming have become
routine?”
The first panelist seemed quite disdainful of anyone who would ask such an absurd
question, and quickly responded with a soundbite answer. To which Paul gave a
soundbite response. They went back and forth for some time, for example, the panelist’s
sound-bite answer “Deadlock” provoked Paul’s sound-bite response “Lock dependency
checker”.
The panelist eventually ran out of soundbites, improvising a final “People like you
should be hit over the head with a hammer!”
Paul’s response was of course “You will have to get in line for that!”
Paul turned his attention to the next panelist, who seemed torn between agreeing with
the first panelist and not wishing to have to deal with Paul’s series of responses. He
therefore have a short non-committal speech. And so it went through the rest of the
panel.
Until it was the turn of the last panelist, who was someone you might have heard of
who goes by the name of Linus Torvalds. Linus noted that three years earlier (that is,
2003), the initial version of any concurrency-related patch was usually quite poor, having
design flaws and many bugs. And even when it was cleaned up enough to be accepted,
bugs still remained. Linus contrasted this with the then-current situation in 2006, in
which he said that it was not unusual for the first version of a concurrency-related patch
to be well-designed with few or even no bugs. He then suggested that if tools continued
to improve, then maybe parallel programming would become routine by the year 2021.2
The conference then concluded. Paul was not surprised to be given wide berth by
many audience members, especially those who saw the world in the same way as did
the first panelist. Paul was also not surprised that a few audience members thanked him
for the question. However, he was quite surprised when one man came up to say “thank
you” with tears streaming down his face, sobbing so hard that he could barely speak.
You see, this man had worked several years at Sequent, and thus very well understood
parallel programming. Furthermore, he was currently assigned to a group whose job it
was to write parallel code. Which was not going well. You see, it wasn’t that they had
trouble understanding his explanations of parallel programming.
It was that they refused to listen to him at all.
In short, his group was treating this man in the same way that the first panelist
attempted to treat Paul. And so in that moment, Paul went from “I should write a book
some day” to “I will do whatever it takes to write this book”. Paul is embarrassed to
admit that he does not remember the man’s name, if in fact he ever knew it.
2 Tools have in fact continued to improve, including fuzzers, lock dependency checkers,
static analyzers, formal verification, memory models, and code-modification tools such as
coccinelle. Therefore, those who wish to assert that year-2021 parallel programming is not
routine should refer to Chapter 2’s epigraph.
v2023.06.11a
628 CHAPTER 18. LOOKING FORWARD AND BACK
Figure 18.1: The Most Important Lesson
This book is nevertheless for that man.

And this book is also for everyone else who would like to add low-level concurrency
to their skillset. If you remember nothing else from this book, let it be the lesson of
Figure 18.1.
And this book is also a salute to that unnamed panelist’s unnamed employer. Some
years later, this employer choose to appoint someone with more useful experience and
fewer sound bites. That someone was also on a panel, and during that session he looked
directly at me when he stated that parallel programming was perhaps 5% more difficult
than sequential programming.
For the rest of us, when someone tries to show us a solution to pressing problem,
perhaps we should at the very least do them the courtesy of listening!
v2023.06.11a
Ask me no questions, and I’ll tell you no fibs.
She Stoops to Conquer, Oliver Goldsmith
Appendix A
Important Questions
The following sections discuss some important questions relating to SMP programming.
Each section also shows how to avoid worrying about the corresponding question, which
can be extremely important if your goal is to simply get your SMP code working as
quickly and painlessly as possible—which is an excellent goal, by the way!
Although the answers to these questions are often less intuitive than they would be in a
single-threaded setting, with a bit of work, they are not that difficult to understand. If you
managed to master recursion, there is nothing here that should pose an overwhelming
challenge.
With that, here are the questions:
1. Why aren’t parallel programs always faster? (Appendix A.1)
2. Why not remove locking? (Appendix A.2)
3. What time is it? (Appendix A.3)
4. What does “after” mean? (Appendix A.4)
5. How much ordering is needed? (Appendix A.5)
6. What is the difference between “concurrent” and “parallel”? (Appendix A.6)
7. Why is software buggy? (Appendix A.7)
Read on to learn some answers. Improve upon these answers if you can!
A.1 Why Aren’t Parallel Programs Always Faster?

The short answer is “because parallel execution often requires communication, and
communication is not free”.
For more information on this question, see Chapter 3, Section 5.1, and especially
Chapter 6, each of which present ways of slowing down your code by ineptly parallelizing
it. Of course, much of this book deals with ways of ensuring that your parallel programs
really are faster than their sequential counterparts.
However, never forget that parallel programs can be quite fast while at the same time
being quite simple, with the example in Section 4.1 being a case in point. Also never
forget that parallel execution is but one optimization of many, and there are programs
for which other optimizations produce better results.
629
v2023.06.11a
630 APPENDIX A. IMPORTANT QUESTIONS
What time is it?
Uh. When did

you ask?
Figure A.1: What Time Is It?
A.2 Why Not Remove Locking?

There can be no doubt that many have cast locking as the evil villain of parallel
programming, and not entirely without reason. And there are important examples where
lockless code does much better than its locked counterpart, a few of which are discussed
in Section 14.2.
However, lockless algorithms are not guaranteed to perform and scale well, as
shown by Figure 5.1 on page 79. Furthermore, as a general rule, the more complex
the algorithm, the greater the advantage of combining locking with selected lockless
techniques, even with significant hardware support, as shown in Table 17.3 on page 611.
Section 14.2 looks more deeply at non-blocking synchronization, which is a popular
lockless methodology.
As a more general rule, a sound-bite approach to parallel programming is not likely
to end well. Some would argue that this is also true of many other fields of endeavor.
A.3 What Time Is It?

A key issue with timekeeping on multicore computer systems is illustrated by Figure A.1.
One problem is that it takes time to read out the time. An instruction might read from a
hardware clock, and might have to go off-core (or worse yet, off-socket) to complete
this read operation. It might also be necessary to do some computation on the value
read out, for example, to convert it to the desired format, to apply network time protocol
(NTP) adjustments, and so on. So does the time eventually returned correspond to the
beginning of the resulting time interval, the end, or somewhere in between?
Worse yet, the thread reading the time might be interrupted or preempted. Furthermore,
there will likely be some computation between reading out the time and the actual use
of the time that has been read out. Both of these possibilities further extend the interval
of uncertainty.
One approach is to read the time twice, and take the arithmetic mean of the two
readings, perhaps one on each side of the operation being timestamped. The difference
v2023.06.11a
A.3. WHAT TIME IS IT? 631
300
250
200
Frequency
150
100
50
0
-100 -80 -60 -40 -20 0 20 40 60
Nanoseconds Deviation
Figure A.2: clock_gettime(CLOCK_REALTIME) Deviation From Immediately Pre-

ceding clock_gettime(CLOCK_MONOTONIC)
between the two readings is then a measure of uncertainty of the time at which the
intervening operation occurred.
Of course, in many cases, the exact time is not necessary. For example, when
printing the time for the benefit of a human user, we can rely on slow human reflexes to
render internal hardware and software delays irrelevant. Similarly, if a server needs to
timestamp the response to a client, any time between the reception of the request and
the transmission of the response will do equally well.
There is an old saying that those who have but one clock always know the time, but
those who have several clocks can never be sure. And there was a time when the typical
low-end computer’s sole software-visible clock was its program counter, but those days
are long gone. This is not a bad thing, considering that on modern computer systems,
the program counter is a truly horrible clock [MOZ09].
In addition, different clocks provide different tradeoffs of performance, accuracy,
precision, and ordering. For example, in the Linux kernel, the jiffies counter1 provides
high-speed access to a course-grained counter (at best one-millisecond accuracy and
precision) that imposes very little ordering on either the compiler or the hardware. In
contrast, the x86 HPET hardware provides an accurate and precise clock, but at the
price of slow access. The x86 time-stamp counter (TSC) has a checkered past, but is
more recently held out as providing a good combination of precision, accuracy, and
performance. Unfortunately, for all of these counters, ordering against all effects of
prior and subsequent code requires expensive memory-barrier instructions. And this
expense appears to be an unavoidable consequence of the complex superscalar nature of
modern computer systems.
In addition, each clock source provides its own timebase. Figure A.2 shows a
histogram of the value returned by a call to clock_gettime(CLOCK_MONOTONIC)
subtracted from that returned by an immediately following clock_gettime(CLOCK_
1 The jiffies variable is a location in normal memory that is incremented by software
in response to events such as the scheduling-clock interrupt.
v2023.06.11a
Listing A.1: “After” Producer Function

1 /* WARNING: BUGGY CODE. */
2 void *producer(void *ignored)
3 {
4 int i = 0;
5
6 producer_ready = 1;
7 while (!goflag)
8 sched_yield();
9 while (goflag) {
10 ss.t = dgettimeofday();
11 ss.a = ss.c + 1;
12 ss.b = ss.a + 1;
13 ss.c = ss.b + 1;
14 i++;
15 }
16 printf("producer exiting: %d samples\n", i);
17 producer_done = 1;
18 return (NULL);
19 }
REALTIME) (timeskew.c). Because some time passes between these two function
calls, it is no surprise that there are positive deviations, but the negative deviations
should give us some pause. Nevertheless, such deviations are possible, if for no other
reason than the machinations of network time protocol (NTP) [Wei22f].
Worse yet, identical clocksources on different systems are not necessarily compatible
with that of another. For example, the jiffies counters on a pair of systems very likely
started counting at different times, and worse yet might well be counting at different
rates. This brings up the topic of synchronizing a given system’s counters with some
real-world notion of time such as the aforementioned NTP, but that topic is beyond the
scope of this book.
In short, time is a slippery topic that causes untold confusion to parallel programmers
and to their code.
A.4 What Does “After” Mean?

“After” is an intuitive, but surprisingly difficult concept. An important non-intuitive issue
is that code can be delayed at any point for any amount of time. Consider a producing
and a consuming thread that communicate using a global struct with a timestamp “t”
and integer fields “a”, “b”, and “c”. The producer loops recording the current time (in
seconds since 1970 in decimal), then updating the values of “a”, “b”, and “c”, as shown
in Listing A.1. The consumer code loops, also recording the current time, but also
copying the producer’s timestamp along with the fields “a”, “b”, and “c”, as shown in
Listing A.2. At the end of the run, the consumer outputs a list of anomalous recordings,
e.g., where time has appeared to go backwards.
Quick Quiz A.1: What SMP coding errors can you see in these examples? See time.c for
full code.
One might intuitively expect that the difference between the producer and consumer
timestamps would be quite small, as it should not take much time for the producer to
record the timestamps or the values. An excerpt of some sample output on a dual-core
1 GHz x86 is shown in Table A.1. Here, the “seq” column is the number of times
through the loop, the “time” column is the time of the anomaly in seconds, the “delta”
column is the number of seconds the consumer’s timestamp follows that of the producer
v2023.06.11a
A.4. WHAT DOES “AFTER” MEAN? 633
Listing A.2: “After” Consumer Function

1 /* WARNING: BUGGY CODE. */
2 void *consumer(void *ignored)
3 {
4 struct snapshot_consumer curssc;
5 int i = 0;
6 int j = 0;
7
8 consumer_ready = 1;
9 while (ss.t == 0.0) {
10 sched_yield();
11 }
12 while (goflag) {
13 curssc.tc = dgettimeofday();
14 curssc.t = ss.t;
15 curssc.a = ss.a;
16 curssc.b = ss.b;
17 curssc.c = ss.c;
18 curssc.sequence = curseq;
19 curssc.iserror = 0;
20 if ((curssc.t > curssc.tc) ||
21 modgreater(ssc[i].a, curssc.a) ||
22 modgreater(ssc[i].b, curssc.b) ||
23 modgreater(ssc[i].c, curssc.c) ||
24 modgreater(curssc.a, ssc[i].a + maxdelta) ||
25 modgreater(curssc.b, ssc[i].b + maxdelta) ||
26 modgreater(curssc.c, ssc[i].c + maxdelta)) {
27 i++;
28 curssc.iserror = 1;
29 } else if (ssc[i].iserror)
30 i++;
31 ssc[i] = curssc;
32 curseq++;
33 if (i + 1 >= NSNAPS)
34 break;
35 }
36 printf("consumer exited loop, collected %d items %d\n",
37 i, curseq);
38 if (ssc[0].iserror)
39 printf("0/%ld: %.6f %.6f (%.3f) %ld %ld %ld\n",
40 ssc[0].sequence,
41 ssc[j].t, ssc[j].tc,
42 (ssc[j].tc - ssc[j].t) * 1000000,
43 ssc[j].a, ssc[j].b, ssc[j].c);
44 for (j = 0; j <= i; j++)
45 if (ssc[j].iserror)
46 printf("%d/%ld: %.6f (%.3f) %ld %ld %ld\n",
47 j, ssc[j].sequence,
48 ssc[j].t, (ssc[j].tc - ssc[j].t) * 1000000,
49 ssc[j].a - ssc[j - 1].a,
50 ssc[j].b - ssc[j - 1].b,
51 ssc[j].c - ssc[j - 1].c);
52 consumer_done = 1;
53 }
v2023.06.11a
Table A.1: “After” Program Sample Output

seq time (seconds) delta a b c
17563: 1152396.251585 ( −16.928) 27 27 27

18004: 1152396.252581 ( −12.875) 24 24 24
18163: 1152396.252955 ( −19.073) 18 18 18
18765: 1152396.254449 ( −148.773) 216 216 216
19863: 1152396.256960 ( −6.914) 18 18 18
21644: 1152396.260959 ( −5.960) 18 18 18
23408: 1152396.264957 ( −20.027) 15 15 15
(where a negative value indicates that the consumer has collected its timestamp before
the producer did), and the columns labelled “a”, “b”, and “c” show the amount that
these variables increased since the prior snapshot collected by the consumer.
Why is time going backwards? The number in parentheses is the difference in
microseconds, with a large number exceeding 10 microseconds, and one exceeding even
100 microseconds! Please note that this CPU can potentially execute more than 100,000
instructions in that time.
One possible reason is given by the following sequence of events:
1. Consumer obtains timestamp (Listing A.2, line 13).
2. Consumer is preempted.
3. An arbitrary amount of time passes.
4. Producer obtains timestamp (Listing A.1, line 10).
5. Consumer starts running again, and picks up the producer’s timestamp (Listing A.2,
line 14).
In this scenario, the producer’s timestamp might be an arbitrary amount of time after
the consumer’s timestamp.
How do you avoid agonizing over the meaning of “after” in your SMP code?
Simply use SMP primitives as designed.
In this example, the easiest fix is to use locking, for example, acquire a lock in the
producer before line 10 in Listing A.1 and in the consumer before line 13 in Listing A.2.
This lock must also be released after line 13 in Listing A.1 and after line 17 in Listing A.2.
These locks cause the code segments in lines 10–13 of Listing A.1 and in lines 13–17 of
Listing A.2 to exclude each other, in other words, to run atomically with respect to each
other. This is represented in Figure A.3: The locking prevents any of the boxes of code
from overlapping in time, so that the consumer’s timestamp must be collected after the
prior producer’s timestamp. The segments of code in each box in this figure are termed
“critical sections”; only one such critical section may be executing at a given time.
This addition of locking results in output as shown in Table A.2. Here there are no
instances of time going backwards, instead, there are only cases with more than 1,000
counts difference between consecutive reads by the consumer.
Quick Quiz A.2: How could there be such a large gap between successive consumer reads?
See timelocked.c for full code.
In summary, if you acquire an exclusive lock, you know that anything you do while
holding that lock will appear to happen after anything done by any prior holder of that
v2023.06.11a
A.5. HOW MUCH ORDERING IS NEEDED? 635
Time
Producer
ss.t = dgettimeofday();
ss.a = ss.c + 1;
ss.b = ss.a + 1;
ss.c = ss.b + 1;
Consumer
curssc.tc = gettimeofday();
curssc.t = ss.t;
curssc.a = ss.a;
curssc.b = ss.b;
curssc.c = ss.c;
Producer
ss.t = dgettimeofday();
ss.a = ss.c + 1;
ss.b = ss.a + 1;
ss.c = ss.b + 1;
Figure A.3: Effect of Locking on Snapshot Collection
Table A.2: Locked “After” Program Sample Output

seq time (seconds) delta a b c
58597: 1156521.556296 (3.815) 1485 1485 1485

403927: 1156523.446636 (2.146) 2583 2583 2583
lock, at least give or take transactional lock elision (see Section 17.3.2.6). No need to
worry about which CPU did or did not execute a memory barrier, no need to worry about
the CPU or compiler reordering operations—life is simple. Of course, the fact that this
locking prevents these two pieces of code from running concurrently might limit the
program’s ability to gain increased performance on multiprocessors, possibly resulting
in a “safe but slow” situation. Chapter 6 describes ways of gaining performance and
scalability in many situations.
In short, in many parallel programs, the really important definition of “after” is
ordering of operations, which is covered in dazzling detail in Chapter 15.
However, in most cases, if you find yourself worrying about what happens before
or after a given piece of code, you should take this as a hint to make better use of the
standard primitives. Let these primitives do the worrying for you.
A.5 How Much Ordering Is Needed?

Perhaps you have carefully constructed a strongly ordered concurrent system, only to
find that it neither performs nor scales well. Or perhaps you threw caution to the wind,
only to find that your brilliantly fast and scalable software is also unreliable. Is there a
happy medium with both robust reliability on the one hand and powerful performance
augmented by scintellating scalability on the other?
The answer, as is so often the case, is “it depends”.
v2023.06.11a
One approach is to construct a strongly ordered system, then examine its performance
and scalability. If these suffice, the system is good and sufficient, and no more need
be done. Otherwise, undertake careful analysis (see Section 11.7) and attack each
bottleneck until the system’s performance is good and sufficient.
This approach can work very well, especially in contrast to the all-too-common
approach of optimizing random components of the system in the hope of achieving
significant system-wide benefits. However, starting with strong ordering can also be
quite wasteful, given that weakening ordering of the system’s bottleneck can require
that large portions of the rest of the system be redesigned and rewritten to accommodate
the weakening. Worse yet, eliminating one bottleneck often exposes another, which
in turn needs to be weakened and which in turn can result in wholesale redesigns and
rewrites of other parts of the system. Perhaps even worse is the approach, also common,
of starting with a fast but unreliable system and then playing whack-a-mole with an
endless succession of concurrency bugs, though in the latter case, Chapters 11 and 12
are always there for you.
It would be better to have design-time tools to determine which portions of the system
could use weak ordering, and at the same time, which portions actually benefit from
weak ordering. These tasks are taken up by the following sections.
A.5.1 Where is the Defining Data?

One way to do this is to keep firmly in mind that the region of consistency engendered
by strong ordering cannot extend out past the boundaries of the system.2 Portions of
the system whose role is to track the state of the outside world can usually feature
weak ordering, given that speed-of-light delays will force the within-system state to
lag that of the outside world. There is often no point in incurring large overheads
to force a consistent view of data that is inherently out of date. In these cases, the
methods of Chapter 9 can be quite helpful, as can some of the data structures described
in Chapter 10.
Nevertheless, it is wise to adopt some meaningful semantics that are visible to those
accessing the data, for example, a given function’s return value might be:
1. Some value between the conceptual value at the time of the call to the function and
the conceptual value at the time of the return from that function. For example, see
the statistical counters discussed in Section 5.2, keeping in mind that such counters
are normally monotonic, at least between consecutive overflows.
2. The actual value at some time between the call to and the return from that function.
For example, see the single-variable atomic counter shown in Listing 5.2.
3. If the values used by that function remain unchanged during the time between
that function’s call and return, the expected value, otherwise some approximation
to the expected value. Precise specification of the bounds on the approximation
can be quite challenging. For example, consider a function combining values
from different elements of an RCU-protected linked data structure, as described in
Section 10.3.
Weaker ordering usually implies weaker semantics, and you should be able to give
some sort of promise to your users as to how this weakening affects them. At the same
2 Which might well be a distributed system.
v2023.06.11a
A.5. HOW MUCH ORDERING IS NEEDED? 637
time, unless the caller holds a lock across both the function call and the use of any
values computed by that function, even fully ordered implementations normally cannot
do any better than the semantics given by the options above.
Quick Quiz A.3: But if fully ordered implementations cannot offer stronger guarantees than
the better performing and more scalable weakly ordered implementations, why bother with full
ordering?
Some might argue that useful computing deals only with the outside world, and
therefore that all computing can use weak ordering. Such arguments are incorrect. For
example, the value of your bank account is defined within your bank’s computers, and
people often prefer exact computations involving their account balances, especially
those who might suspect that any such approximations would be in the bank’s favor.
In short, although data tracking external state can be an attractive candidate for weakly
ordered access, please think carefully about exactly what is being tracked and what is
doing the tracking.
A.5.2 Consistent Data Used Consistently?

Another hint that weakening is safe can appear in the guise of data that is computed
while holding a lock, but then used after the lock is released. The computed result
clearly becomes at best an approximation as soon as the lock is released, which suggests
computing an approximate result in the first place, possibly permitting use of weaker
ordering. To this end, Chapter 5 covers numerous approximate methods for counting.
Great care is required, however. Is the use of data following lock release a hint that
weak-ordering optimizations might be helpful? Or is instead a bug in which the lock
was released too soon?
A.5.3 Is the Problem Partitionable?

Suppose that the system holds the defining instance of the data, or that using a computed
value past lock release proved to be a bug. What then?
One approach is to partition the system, as discussed in Chapter 6. Partititioning
can provide excellent scalability and in its more extreme form, per-CPU performance
rivaling that of a sequential program, as discussed in Chapter 8. Partial partitioning is
often mediated by locking, which is the subject of Chapter 7.
A.5.4 None of the Above?

The previous sections described the easier ways to gain performance and scalability,
sometimes using weaker ordering and sometimes not. But the plain fact is that multicore
systems are under no compunction to make life easy. But perhaps the advanced topics
covered in Chapters 14 and 15 will prove helpful.
But please proceed with care, as it is all too easy to destabilize your codebase
optimizing non-bottlenecks. Once again, Section 11.7 can help. It might also be worth
your time to review other portions of this book, as it contains much information on
handling a number of tricky situations.
v2023.06.11a
A.6 What is the Difference Between “Concurrent” and

“Parallel”?
From a classic computing perspective, “concurrent” and “parallel” are clearly synonyms.
However, this has not stopped many people from drawing distinctions between the two,
and it turns out that these distinctions can be understood from a couple of different
perspectives.
The first perspective treats “parallel” as an abbreviation for “data parallel”, and
treats “concurrent” as pretty much everything else. From this perspective, in parallel
computing, each partition of the overall problem can proceed completely independently,
with no communication with other partitions. In this case, little or no coordination
among partitions is required. In contrast, concurrent computing might well have tight
interdependencies, in the form of contended locks, transactions, or other synchronization
mechanisms.
Quick Quiz A.4: Suppose a portion of a program uses RCU read-side primitives as its only
synchronization mechanism. Is this parallelism or concurrency?
This of course begs the question of why such a distinction matters, which brings
us to the second perspective, that of the underlying scheduler. Schedulers come in a
wide range of complexities and capabilities, and as a rough rule of thumb, the more
tightly and irregularly a set of parallel processes communicate, the higher the level of
sophistication required from the scheduler. As such, parallel computing’s avoidance
of interdependencies means that parallel-computing programs run well on the least-
capable schedulers. In fact, a pure parallel-computing program can run successfully
after being arbitrarily subdivided and interleaved onto a uniprocessor.3 In contrast,
concurrent-computing programs might well require extreme subtlety on the part of the
scheduler.
One could argue that we should simply demand a reasonable level of competence from
the scheduler, so that we could simply ignore any distinctions between parallelism and
concurrency. Although this is often a good strategy, there are important situations where
efficiency, performance, and scalability concerns sharply limit the level of competence
that the scheduler can reasonably offer. One important example is when the scheduler is
implemented in hardware, as it often is in SIMD units or GPGPUs. Another example
is a workload where the units of work are quite short, so that even a software-based
scheduler must make hard choices between subtlety on the one hand and efficiency on
the other.
Now, this second perspective can be thought of as making the workload match the
available scheduler, with parallel workloads able to use simple schedulers and concurrent
workloads requiring sophisticated schedulers.
Unfortunately, this perspective does not always align with the dependency-based
distinction put forth by the first perspective. For example, a highly interdependent
lock-based workload with one thread per CPU can make do with a trivial scheduler
because no scheduler decisions are required. In fact, some workloads of this type can
even be run one after another on a sequential machine. Therefore, such a workload
would be labeled “concurrent” by the first perspective and “parallel” by many taking the
second perspective.
3 Yes, this does mean that data-parallel-computing programs are best-suited for sequential
execution. Why did you ask?
v2023.06.11a
A.7. WHY IS SOFTWARE BUGGY? 639
Quick Quiz A.5: In what part of the second (scheduler-based) perspective would the lock-based
single-thread-per-CPU workload be considered “concurrent”?
Which is just fine. No rule that humankind writes carries any weight against the
objective universe, not even rules dividing multiprocessor programs into categories
such as “concurrent” and “parallel”.
This categorization failure does not mean such rules are useless, but rather that you
should take on a suitably skeptical frame of mind when attempting to apply them to new
situations. As always, use such rules where they apply and ignore them otherwise.
In fact, it is likely that new categories will arise in addition to parallel, concurrent,
map-reduce, task-based, and so on. Some will stand the test of time, but good luck
guessing which!
A.7 Why Is Software Buggy?

The short answer is “because it was written by humans, and to err is human”. This
does not necessarily mean that automated code generation is the answer, because the
program that does the code generation will have been written by humans. In addition,
one of the biggest problems in producing software is working out what that software is
supposed to do, and this task has thus far proven rather resistant to automation.
Nevertheless, automation is an important part of the process of reducing the number
of bugs in software. For but one example, despite their many flaws, it is almost always
better to use a compiler than to write in assembly language.
Furthermore, careful validation can be very helpful in finding bugs, as discussed in
Chapters 11–12.
v2023.06.11a
v2023.06.11a
The only difference between men and boys is the
price of their toys.
M. Hébert
Appendix B
“Toy” RCU Implementations
The toy RCU implementations in this appendix are designed not for high performance,
practicality, or any kind of production use,1 but rather for clarity. Nevertheless, you will
need a thorough understanding of Chapters 2, 3, 4, 6, and 9 for even these toy RCU
implementations to be easily understandable.
This appendix provides a series of RCU implementations in order of increasing
sophistication, from the viewpoint of solving the existence-guarantee problem. Appen-
dix B.1 presents a rudimentary RCU implementation based on simple locking, while
Appendices B.2 through B.9 present a series of simple RCU implementations based on
locking, reference counters, and free-running counters. Finally, Appendix B.10 provides
a summary and a list of desirable RCU properties.
B.1 Lock-Based RCU

Perhaps the simplest RCU implementation leverages locking, as shown in Listing B.1
(rcu_lock.h and rcu_lock.c).
In this implementation, rcu_read_lock() acquires a global spinlock, rcu_read_
unlock() releases it, and synchronize_rcu() acquires it then immediately releases
it.
Listing B.1: Lock-Based RCU Implementation

1 static void rcu_read_lock(void)
2 {
3 spin_lock(&rcu_gp_lock);
4 }
5
6 static void rcu_read_unlock(void)
7 {
8 spin_unlock(&rcu_gp_lock);
9 }
10
12 {
15 }
1 However, production-quality user-level RCU implementations are available [Des09b,
DMS+ 12].
641
v2023.06.11a
642 APPENDIX B. “TOY” RCU IMPLEMENTATIONS
Listing B.2: Per-Thread Lock-Based RCU Implementation

2 {
3 spin_lock(&__get_thread_var(rcu_gp_lock));
4 }
5
7 {
8 spin_unlock(&__get_thread_var(rcu_gp_lock));
9 }
10
12 {
13 int t;
14
15 for_each_running_thread(t) {
16 spin_lock(&per_thread(rcu_gp_lock, t));
17 spin_unlock(&per_thread(rcu_gp_lock, t));
18 }
19 }
Because synchronize_rcu() does not return until it has acquired (and released)
the lock, it cannot return until all prior RCU read-side critical sections have completed,
thus faithfully implementing RCU semantics. Of course, only one RCU reader may be in
its read-side critical section at a time, which almost entirely defeats the purpose of RCU.
In addition, the lock operations in rcu_read_lock() and rcu_read_unlock() are
extremely heavyweight, with read-side overhead ranging from about 100 nanoseconds
on a single POWER5 CPU up to more than 17 microseconds on a 64-CPU system.
Worse yet, these same lock operations permit rcu_read_lock() to participate in
deadlock cycles. Furthermore, in absence of recursive locks, RCU read-side critical
sections cannot be nested, and, finally, although concurrent RCU updates could in
principle be satisfied by a common grace period, this implementation serializes grace
periods, preventing grace-period sharing.
Quick Quiz B.1: Why wouldn’t any deadlock in the RCU implementation in Listing B.1 also
be a deadlock in any other RCU implementation?
Quick Quiz B.2: Why not simply use reader-writer locks in the RCU implementation in
Listing B.1 in order to allow RCU readers to proceed in parallel?
It is hard to imagine this implementation being useful in a production setting, though

it does have the virtue of being implementable in almost any user-level application.
Furthermore, similar implementations having one lock per CPU or using reader-writer
locks have been used in production in the 2.4 Linux kernel.
A modified version of this one-lock-per-CPU approach, but instead using one lock
per thread, is described in the next section.
B.2 Per-Thread Lock-Based RCU

Listing B.2 (rcu_lock_percpu.h and rcu_lock_percpu.c) shows an implementa-
tion based on one lock per thread. The rcu_read_lock() and rcu_read_unlock()
functions acquire and release, respectively, the current thread’s lock. The synchronize_
rcu() function acquires and releases each thread’s lock in turn. Therefore, all RCU
read-side critical sections running when synchronize_rcu() starts must have com-
pleted before synchronize_rcu() can return.
v2023.06.11a
B.3. SIMPLE COUNTER-BASED RCU 643
This implementation does have the virtue of permitting concurrent RCU readers, and
does avoid the deadlock condition that can arise with a single global lock. Furthermore,
the read-side overhead, though high at roughly 140 nanoseconds, remains at about 140
nanoseconds regardless of the number of CPUs. However, the update-side overhead
ranges from about 600 nanoseconds on a single POWER5 CPU up to more than 100
microseconds on 64 CPUs.
Quick Quiz B.3: Wouldn’t it be cleaner to acquire all the locks, and then release them all in
the loop from lines 15–18 of Listing B.2? After all, with this change, there would be a point in
time when there were no readers, simplifying things greatly.
Quick Quiz B.4: Is the implementation shown in Listing B.2 free from deadlocks? Why or
why not?
Quick Quiz B.5: Isn’t one advantage of the RCU algorithm shown in Listing B.2 that it uses
only primitives that are widely available, for example, in POSIX pthreads?
This approach could be useful in some situations, given that a similar approach was
used in the Linux 2.4 kernel [MM00].
The counter-based RCU implementation described next overcomes some of the
shortcomings of the lock-based implementation.
B.3 Simple Counter-Based RCU

A slightly more sophisticated RCU implementation is shown in Listing B.3 (rcu_rcg.h
and rcu_rcg.c). This implementation makes use of a global reference counter rcu_
refcnt defined on line 1. The rcu_read_lock() primitive atomically increments
this counter, then executes a memory barrier to ensure that the RCU read-side critical
section is ordered after the atomic increment. Similarly, rcu_read_unlock() executes
a memory barrier to confine the RCU read-side critical section, then atomically
decrements the counter. The synchronize_rcu() primitive spins waiting for the
reference counter to reach zero, surrounded by memory barriers. The poll() on line 19
merely provides pure delay, and from a pure RCU-semantics point of view could be
omitted. Again, once synchronize_rcu() returns, all prior RCU read-side critical
sections are guaranteed to have completed.
In happy contrast to the lock-based implementation shown in Appendix B.1, this
implementation allows parallel execution of RCU read-side critical sections. In happy
contrast to the per-thread lock-based implementation shown in Appendix B.2, it also
allows them to be nested. In addition, the rcu_read_lock() primitive cannot possibly
participate in deadlock cycles, as it never spins nor blocks.
Quick Quiz B.6: But what if you hold a lock across a call to synchronize_rcu(), and then
acquire that same lock within an RCU read-side critical section?
However, this implementation still has some serious shortcomings. First, the atomic
operations in rcu_read_lock() and rcu_read_unlock() are still quite heavyweight,
with read-side overhead ranging from about 100 nanoseconds on a single POWER5
CPU up to almost 40 microseconds on a 64-CPU system. This means that the RCU
read-side critical sections have to be extremely long in order to get any real read-
side parallelism. On the other hand, in the absence of readers, grace periods elapse
in about 40 nanoseconds, many orders of magnitude faster than production-quality
implementations in the Linux kernel.
v2023.06.11a
Listing B.3: RCU Implementation Using Single Global Reference Counter

1 atomic_t rcu_refcnt;
2
4 {
5 atomic_inc(&rcu_refcnt);
6 smp_mb();
7 }
8
10 {
11 smp_mb();
12 atomic_dec(&rcu_refcnt);
13 }
14
16 {
17 smp_mb();
18 while (atomic_read(&rcu_refcnt) != 0) {
19 poll(NULL, 0, 10);
20 }
21 smp_mb();
22 }
Quick Quiz B.7: How can the grace period possibly elapse in 40 nanoseconds when
synchronize_rcu() contains a 10-millisecond delay?
Second, if there are many concurrent rcu_read_lock() and rcu_read_unlock()

operations, there will be extreme memory contention on rcu_refcnt, resulting in
expensive cache misses. Both of these first two shortcomings largely defeat a major
purpose of RCU, namely to provide low-overhead read-side synchronization primitives.
Finally, a large number of RCU readers with long read-side critical sections could
prevent synchronize_rcu() from ever completing, as the global counter might
never reach zero. This could result in starvation of RCU updates, which is of course
unacceptable in production settings.
Quick Quiz B.8: Why not simply make rcu_read_lock() wait when a concurrent
synchronize_rcu() has been waiting too long in the RCU implementation in Listing B.3?
Wouldn’t that prevent synchronize_rcu() from starving?
Therefore, it is still hard to imagine this implementation being useful in a production

setting, though it has a bit more potential than the lock-based mechanism, for example,
as an RCU implementation suitable for a high-stress debugging environment. The next
section describes a variation on the reference-counting scheme that is more favorable to
writers.
B.4 Starvation-Free Counter-Based RCU

Listing B.5 (rcu_rcpg.h) shows the read-side primitives of an RCU implementation
that uses a pair of reference counters (rcu_refcnt[]), along with a global index
that selects one counter out of the pair (rcu_idx), a per-thread nesting counter rcu_
nesting, a per-thread snapshot of the global index (rcu_read_idx), and a global lock
(rcu_gp_lock), which are themselves shown in Listing B.4.
Design It is the two-element rcu_refcnt[] array that provides the freedom from
starvation. The key point is that synchronize_rcu() is only required to wait for
v2023.06.11a
B.4. STARVATION-FREE COUNTER-BASED RCU 645
Listing B.4: RCU Global Reference-Count Pair Data

1 DEFINE_SPINLOCK(rcu_gp_lock);
2 atomic_t rcu_refcnt[2];
3 atomic_t rcu_idx;
4 DEFINE_PER_THREAD(int, rcu_nesting);
5 DEFINE_PER_THREAD(int, rcu_read_idx);
Listing B.5: RCU Read-Side Using Global Reference-Count Pair

2 {
3 int i;
4 int n;
5
6 n = __get_thread_var(rcu_nesting);
7 if (n == 0) {
8 i = atomic_read(&rcu_idx);
9 __get_thread_var(rcu_read_idx) = i;
10 atomic_inc(&rcu_refcnt[i]);
11 }
12 __get_thread_var(rcu_nesting) = n + 1;
13 smp_mb();
14 }
15
17 {
18 int i;
19 int n;
20
21 smp_mb();
23 if (n == 1) {
24 i = __get_thread_var(rcu_read_idx);
25 atomic_dec(&rcu_refcnt[i]);
26 }
27 __get_thread_var(rcu_nesting) = n - 1;
28 }
pre-existing readers. If a new reader starts after a given instance of synchronize_

rcu() has already begun execution, then that instance of synchronize_rcu() need
not wait on that new reader. At any given time, when a given reader enters its RCU
read-side critical section via rcu_read_lock(), it increments the element of the
rcu_refcnt[] array indicated by the rcu_idx variable. When that same reader exits
its RCU read-side critical section via rcu_read_unlock(), it decrements whichever
element it incremented, ignoring any possible subsequent changes to the rcu_idx value.
This arrangement means that synchronize_rcu() can avoid starvation by com-
plementing the value of rcu_idx, as in rcu_idx = !rcu_idx. Suppose that the old
value of rcu_idx was zero, so that the new value is one. New readers that arrive after
the complement operation will increment rcu_refcnt[1], while the old readers that
previously incremented rcu_refcnt[0] will decrement rcu_refcnt[0] when they
exit their RCU read-side critical sections. This means that the value of rcu_refcnt[0]
will no longer be incremented, and thus will be monotonically decreasing.2 This means
that all that synchronize_rcu() need do is wait for the value of rcu_refcnt[0] to
reach zero.
With the background, we are ready to look at the implementation of the actual
primitives.
2 There is a race condition that this “monotonically decreasing” statement ignores. This
race condition will be dealt with by the code for synchronize_rcu(). In the meantime, I
suggest suspending disbelief.
v2023.06.11a
Listing B.6: RCU Update Using Global Reference-Count Pair

2 {
3 int i;
4
5 smp_mb();
8 atomic_set(&rcu_idx, !i);
9 smp_mb();
10 while (atomic_read(&rcu_refcnt[i]) != 0) {
11 poll(NULL, 0, 10);
12 }
13 smp_mb();
14 atomic_set(&rcu_idx, i);
15 smp_mb();
16 while (atomic_read(&rcu_refcnt[!i]) != 0) {
17 poll(NULL, 0, 10);
18 }
20 smp_mb();
21 }
Implementation The rcu_read_lock() primitive atomically increments the mem-

ber of the rcu_refcnt[] pair indexed by rcu_idx, and keeps a snapshot of this index
in the per-thread variable rcu_read_idx. The rcu_read_unlock() primitive then
atomically decrements whichever counter of the pair that the corresponding rcu_read_
lock() incremented. However, because only one value of rcu_idx is remembered per
thread, additional measures must be taken to permit nesting. These additional measures
use the per-thread rcu_nesting variable to track nesting.
To make all this work, line 6 of rcu_read_lock() in Listing B.5 picks up the
current thread’s instance of rcu_nesting, and if line 7 finds that this is the outermost
rcu_read_lock(), then lines 8–10 pick up the current value of rcu_idx, save it in this
thread’s instance of rcu_read_idx, and atomically increment the selected element of
rcu_refcnt. Regardless of the value of rcu_nesting, line 12 increments it. Line 13
executes a memory barrier to ensure that the RCU read-side critical section does not
bleed out before the rcu_read_lock() code.
Similarly, the rcu_read_unlock() function executes a memory barrier at line 21 to
ensure that the RCU read-side critical section does not bleed out after the rcu_read_
unlock() code. Line 22 picks up this thread’s instance of rcu_nesting, and if line 23
finds that this is the outermost rcu_read_unlock(), then lines 24 and 25 pick up this
thread’s instance of rcu_read_idx (saved by the outermost rcu_read_lock()) and
atomically decrements the selected element of rcu_refcnt. Regardless of the nesting
level, line 27 decrements this thread’s instance of rcu_nesting.
Listing B.6 (rcu_rcpg.c) shows the corresponding synchronize_rcu() imple-
mentation. Lines 6 and 19 acquire and release rcu_gp_lock in order to prevent more
than one concurrent instance of synchronize_rcu(). Lines 7 and 8 pick up the
value of rcu_idx and complement it, respectively, so that subsequent instances of
rcu_read_lock() will use a different element of rcu_refcnt than did preceding
instances. Lines 10–12 then wait for the prior element of rcu_refcnt to reach zero,
with the memory barrier on line 9 ensuring that the check of rcu_refcnt is not
reordered to precede the complementing of rcu_idx. Lines 13–18 repeat this process,
and line 20 ensures that any subsequent reclamation operations are not reordered to
precede the checking of rcu_refcnt.
v2023.06.11a
B.5. SCALABLE COUNTER-BASED RCU 647
Quick Quiz B.9: Why the memory barrier on line 5 of synchronize_rcu() in Listing B.6
given that there is a spin-lock acquisition immediately after?
Quick Quiz B.10: Why is the counter flipped twice in Listing B.6? Shouldn’t a single
flip-and-wait cycle be sufficient?
This implementation avoids the update-starvation issues that could occur in the
single-counter implementation shown in Listing B.3.
Discussion There are still some serious shortcomings. First, the atomic operations in
rcu_read_lock() and rcu_read_unlock() are still quite heavyweight. In fact, they
are more complex than those of the single-counter variant shown in Listing B.3, with the
read-side primitives consuming about 150 nanoseconds on a single POWER5 CPU and
almost 40 microseconds on a 64-CPU system. The update-side synchronize_rcu()
primitive is more costly as well, ranging from about 200 nanoseconds on a single
POWER5 CPU to more than 40 microseconds on a 64-CPU system. This means that
the RCU read-side critical sections have to be extremely long in order to get any real
read-side parallelism.
Second, if there are many concurrent rcu_read_lock() and rcu_read_unlock()
operations, there will be extreme memory contention on the rcu_refcnt elements, re-
sulting in expensive cache misses. This further extends the RCU read-side critical-section
duration required to provide parallel read-side access. These first two shortcomings
defeat the purpose of RCU in most situations.
Third, the need to flip rcu_idx twice imposes substantial overhead on updates,
especially if there are large numbers of threads.
Finally, despite the fact that concurrent RCU updates could in principle be satisfied
by a common grace period, this implementation serializes grace periods, preventing
grace-period sharing.
Quick Quiz B.11: Given that atomic increment and decrement are so expensive, why not just
use non-atomic increment on line 10 and a non-atomic decrement on line 25 of Listing B.5?
Despite these shortcomings, one could imagine this variant of RCU being used on
small tightly coupled multiprocessors, perhaps as a memory-conserving implementation
that maintains API compatibility with more complex implementations. However, it
would not likely scale well beyond a few CPUs.
The next section describes yet another variation on the reference-counting scheme
that provides greatly improved read-side performance and scalability.
B.5 Scalable Counter-Based RCU

Listing B.8 (rcu_rcpl.h) shows the read-side primitives of an RCU implementation
that uses per-thread pairs of reference counters. This implementation is quite similar to
that shown in Listing B.5, the only difference being that rcu_refcnt is now a per-thread
array (as shown in Listing B.7). As with the algorithm in the previous section, use of
this two-element array prevents readers from starving updaters. One benefit of per-
thread rcu_refcnt[] array is that the rcu_read_lock() and rcu_read_unlock()
primitives no longer perform atomic operations.
v2023.06.11a
Listing B.7: RCU Per-Thread Reference-Count Pair Data

2 DEFINE_PER_THREAD(int [2], rcu_refcnt);
3 atomic_t rcu_idx;
Listing B.8: RCU Read-Side Using Per-Thread Reference-Count Pair

2 {
3 int i;
4 int n;
5
7 if (n == 0) {
10 __get_thread_var(rcu_refcnt)[i]++;
11 }
13 smp_mb();
14 }
15
17 {
18 int i;
19 int n;
20
21 smp_mb();
23 if (n == 1) {
25 __get_thread_var(rcu_refcnt)[i]--;
26 }
28 }
Quick Quiz B.12: Come off it! We can see the atomic_read() primitive in rcu_read_
lock()!!! So why are you trying to pretend that rcu_read_lock() contains no atomic
operations???
Listing B.9 (rcu_rcpl.c) shows the implementation of synchronize_rcu(), along

with a helper function named flip_counter_and_wait(). The synchronize_
rcu() function resembles that shown in Listing B.6, except that the repeated counter
flip is replaced by a pair of calls on lines 22 and 23 to the new helper function.
The new flip_counter_and_wait() function updates the rcu_idx variable on
line 5, executes a memory barrier on line 6, then lines 7–11 spin on each thread’s prior
rcu_refcnt element, waiting for it to go to zero. Once all such elements have gone to
zero, it executes another memory barrier on line 12 and returns.
This RCU implementation imposes important new requirements on its software
environment, namely, (1) that it be possible to declare per-thread variables, (2) that
these per-thread variables be accessible from other threads, and (3) that it is possible
to enumerate all threads. These requirements can be met in almost all software
environments, but often result in fixed upper bounds on the number of threads. More-
complex implementations might avoid such bounds, for example, by using expandable
hash tables. Such implementations might dynamically track threads, for example, by
adding them on their first call to rcu_read_lock().
v2023.06.11a
B.5. SCALABLE COUNTER-BASED RCU 649
Listing B.9: RCU Update Using Per-Thread Reference-Count Pair

1 static void flip_counter_and_wait(int i)
2 {
3 int t;
4
5 atomic_set(&rcu_idx, !i);
6 smp_mb();
8 while (per_thread(rcu_refcnt, t)[i] != 0) {
10 }
11 }
12 smp_mb();
13 }
14
16 {
17 int i;
18
19 smp_mb();
22 flip_counter_and_wait(i);
23 flip_counter_and_wait(!i);
25 smp_mb();
26 }
Quick Quiz B.13: Great, if we have 𝑁 threads, we can have 2𝑁 ten-millisecond waits (one set
per flip_counter_and_wait() invocation, and even that assumes that we wait only once for
each thread). Don’t we need the grace period to complete much more quickly?
This implementation still has several shortcomings. First, the need to flip rcu_idx
twice imposes substantial overhead on updates, especially if there are large numbers of
threads.
Second, synchronize_rcu() must now examine a number of variables that increases

linearly with the number of threads, imposing substantial overhead on applications with
large numbers of threads.
Third, as before, although concurrent RCU updates could in principle be satisfied

by a common grace period, this implementation serializes grace periods, preventing
grace-period sharing.
Finally, as noted in the text, the need for per-thread variables and for enumerating
threads may be problematic in some software environments.
That said, the read-side primitives scale very nicely, requiring about 115 nanoseconds
regardless of whether running on a single-CPU or a 64-CPU POWER5 system. As
noted above, the synchronize_rcu() primitive does not scale, ranging in overhead
from almost a microsecond on a single POWER5 CPU up to almost 200 microseconds
on a 64-CPU system. This implementation could conceivably form the basis for a
production-quality user-level RCU implementation.
The next section describes an algorithm permitting more efficient concurrent RCU
updates.
v2023.06.11a
Listing B.10: RCU Read-Side Using Per-Thread Reference-Count Pair and Shared Update Data
2 DEFINE_PER_THREAD(int [2], rcu_refcnt);
3 long rcu_idx;
Listing B.11: RCU Read-Side Using Per-Thread Reference-Count Pair and Shared Update
2 {
3 int i;
4 int n;
5
7 if (n == 0) {
8 i = READ_ONCE(rcu_idx) & 0x1;
10 __get_thread_var(rcu_refcnt)[i]++;
11 }
13 smp_mb();
14 }
15
17 {
18 int i;
19 int n;
20
21 smp_mb();
23 if (n == 1) {
25 __get_thread_var(rcu_refcnt)[i]--;
26 }
28 }
B.6 Scalable Counter-Based RCU With Shared Grace

Periods
Listing B.11 (rcu_rcpls.h) shows the read-side primitives for an RCU implementation
using per-thread reference count pairs, as before, but permitting updates to share grace
periods. The main difference from the earlier implementation shown in Listing B.8 is that
rcu_idx is now a long that counts freely, so that line 8 of Listing B.11 must mask off
the low-order bit. We also switched from using atomic_read() and atomic_set()
to using READ_ONCE(). The data is also quite similar, as shown in Listing B.10, with
rcu_idx now being a long instead of an atomic_t.
Listing B.12 (rcu_rcpls.c) shows the implementation of synchronize_rcu()
and its helper function flip_counter_and_wait(). These are similar to those in
Listing B.9. The differences in flip_counter_and_wait() include:
1. Line 6 uses WRITE_ONCE() instead of atomic_set(), and increments rather than

complementing.
2. A new line 7 masks the counter down to its bottom bit.
The changes to synchronize_rcu() are more pervasive:
1. There is a new oldctr local variable that captures the pre-lock-acquisition value
of rcu_idx on line 20.
v2023.06.11a
B.6. SCALABLE COUNTER-BASED RCU WITH SHARED GRACE PERIODS651
Listing B.12: RCU Shared Update Using Per-Thread Reference-Count Pair

1 static void flip_counter_and_wait(int ctr)
2 {
3 int i;
4 int t;
5
6 WRITE_ONCE(rcu_idx, ctr + 1);
7 i = ctr & 0x1;
8 smp_mb();
10 while (per_thread(rcu_refcnt, t)[i] != 0) {
11 poll(NULL, 0, 10);
12 }
13 }
14 smp_mb();
15 }
16
18 {
19 int ctr;
20 int oldctr;
21
22 smp_mb();
23 oldctr = READ_ONCE(rcu_idx);
24 smp_mb();
26 ctr = READ_ONCE(rcu_idx);
27 if (ctr - oldctr >= 3) {
29 smp_mb();
30 return;
31 }
32 flip_counter_and_wait(ctr);
33 if (ctr - oldctr < 2)
34 flip_counter_and_wait(ctr + 1);
36 smp_mb();
37 }
2. Line 23 uses READ_ONCE() instead of atomic_read().
3. Lines 27–30 check to see if at least three counter flips were performed by other
threads while the lock was being acquired, and, if so, releases the lock, does a
memory barrier, and returns. In this case, there were two full waits for the counters
to go to zero, so those other threads already did all the required work.
4. At lines 33–34, flip_counter_and_wait() is only invoked a second time if

there were fewer than two counter flips while the lock was being acquired. On the
other hand, if there were two counter flips, some other thread did one full wait for
all the counters to go to zero, so only one more is required.
With this approach, if an arbitrarily large number of threads invoke synchronize_

rcu() concurrently, with one CPU for each thread, there will be a total of only three
waits for counters to go to zero.
Despite the improvements, this implementation of RCU still has a few shortcomings.
First, as before, the need to flip rcu_idx twice imposes substantial overhead on updates,
especially if there are large numbers of threads.
Second, each updater still acquires rcu_gp_lock, even if there is no work to be done.
This can result in a severe scalability limitation if there are large numbers of concurrent
updates. There are ways of avoiding this, as was done in a production-quality real-time
implementation of RCU for the Linux kernel [McK07a].
v2023.06.11a
Listing B.13: Data for Free-Running Counter Using RCU

2 long rcu_gp_ctr = 0;
3 DEFINE_PER_THREAD(long, rcu_reader_gp);
4 DEFINE_PER_THREAD(long, rcu_reader_gp_snap);
Third, this implementation requires per-thread variables and the ability to enumerate
threads, which again can be problematic in some software environments.
Finally, on 32-bit machines, a given update thread might be preempted long enough
for the rcu_idx counter to overflow. This could cause such a thread to force an
unnecessary pair of counter flips. However, even if each grace period took only one
microsecond, the offending thread would need to be preempted for more than an hour,
in which case an extra pair of counter flips is likely the least of your worries.
As with the implementation described in Appendix B.3, the read-side primitives
scale extremely well, incurring roughly 115 nanoseconds of overhead regardless of the
number of CPUs. The synchronize_rcu() primitive is still expensive, ranging from
about one microsecond up to about 16 microseconds. This is nevertheless much cheaper
than the roughly 200 microseconds incurred by the implementation in Appendix B.5.
So, despite its shortcomings, one could imagine this RCU implementation being used
in production in real-life applications.
Quick Quiz B.14: All of these toy RCU implementations have either atomic operations in rcu_
read_lock() and rcu_read_unlock(), or synchronize_rcu() overhead that increases
linearly with the number of threads. Under what circumstances could an RCU implementation
enjoy lightweight implementations for all three of these primitives, all having deterministic
(O (1)) overheads and latencies?
Referring back to Listing B.11, we see that there is one global-variable access and
no fewer than four accesses to thread-local variables. Given the relatively high cost of
thread-local accesses on systems implementing POSIX threads, it is tempting to collapse
the three thread-local variables into a single structure, permitting rcu_read_lock()
and rcu_read_unlock() to access their thread-local data with a single thread-local-
storage access. However, an even better approach would be to reduce the number of
thread-local accesses to one, as is done in the next section.
B.7 RCU Based on Free-Running Counter

Listing B.14 (rcu.h and rcu.c) shows an RCU implementation based on a single
global free-running counter that takes on only even-numbered values, with data shown
in Listing B.13.
The resulting rcu_read_lock() implementation is extremely straightforward.
Lines 3 and 4 simply add the value one to the global free-running rcu_gp_ctr
variable and stores the resulting odd-numbered value into the rcu_reader_gp per-
thread variable. Line 5 executes a memory barrier to prevent the content of the
subsequent RCU read-side critical section from “leaking out”.
The rcu_read_unlock() implementation is similar. Line 10 executes a memory
barrier, again to prevent the prior RCU read-side critical section from “leaking out”.
Lines 11 and 12 then copy the rcu_gp_ctr global variable to the rcu_reader_gp
per-thread variable, leaving this per-thread variable with an even-numbered value so
that a concurrent instance of synchronize_rcu() will know to ignore it.
v2023.06.11a
B.7. RCU BASED ON FREE-RUNNING COUNTER 653
Listing B.14: Free-Running Counter Using RCU

1 static inline void rcu_read_lock(void)
2 {
3 __get_thread_var(rcu_reader_gp) =
4 READ_ONCE(rcu_gp_ctr) + 1;
5 smp_mb();
6 }
7
8 static inline void rcu_read_unlock(void)
9 {
10 smp_mb();
11 __get_thread_var(rcu_reader_gp) =
12 READ_ONCE(rcu_gp_ctr);
13 }
14
16 {
17 int t;
18
19 smp_mb();
21 WRITE_ONCE(rcu_gp_ctr, rcu_gp_ctr + 2);
22 smp_mb();
24 while ((per_thread(rcu_reader_gp, t) & 0x1) &&
25 ((per_thread(rcu_reader_gp, t) -
26 rcu_gp_ctr) < 0)) {
27 poll(NULL, 0, 10);
28 }
29 }
31 smp_mb();
32 }
Quick Quiz B.15: If any even value is sufficient to tell synchronize_rcu() to ignore a given
task, why don’t lines 11 and 12 of Listing B.14 simply assign zero to rcu_reader_gp?
Thus, synchronize_rcu() could wait for all of the per-thread rcu_reader_gp

variables to take on even-numbered values. However, it is possible to do much better
than that because synchronize_rcu() need only wait on pre-existing RCU read-side
critical sections. Line 19 executes a memory barrier to prevent prior manipulations of
RCU-protected data structures from being reordered (by either the CPU or the compiler)
to follow the increment on line 21. Line 20 acquires the rcu_gp_lock (and line 30
releases it) in order to prevent multiple synchronize_rcu() instances from running
concurrently. Line 21 then increments the global rcu_gp_ctr variable by two, so that
all pre-existing RCU read-side critical sections will have corresponding per-thread rcu_
reader_gp variables with values less than that of rcu_gp_ctr, modulo the machine’s
word size. Recall also that threads with even-numbered values of rcu_reader_gp are
not in an RCU read-side critical section, so that lines 23–29 scan the rcu_reader_gp
values until they all are either even (line 24) or are greater than the global rcu_gp_ctr
(lines 25–26). Line 27 blocks for a short period of time to wait for a pre-existing RCU
read-side critical section, but this can be replaced with a spin-loop if grace-period
latency is of the essence. Finally, the memory barrier at line 31 ensures that any
subsequent destruction will not be reordered into the preceding loop.
Quick Quiz B.16: Why are the memory barriers on lines 19 and 31 of Listing B.14 needed?
Aren’t the memory barriers inherent in the locking primitives on lines 20 and 30 sufficient?
This approach achieves much better read-side performance, incurring roughly

63 nanoseconds of overhead regardless of the number of POWER5 CPUs. Updates
v2023.06.11a
Listing B.15: Data for Nestable RCU Using a Free-Running Counter

2 #define RCU_GP_CTR_SHIFT 7
3 #define RCU_GP_CTR_BOTTOM_BIT (1 << RCU_GP_CTR_SHIFT)
4 #define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BOTTOM_BIT - 1)
5 #define MAX_GP_ADV_DISTANCE (RCU_GP_CTR_NEST_MASK << 8)
6 unsigned long rcu_gp_ctr = 0;
7 DEFINE_PER_THREAD(unsigned long, rcu_reader_gp);
incur more overhead, ranging from about 500 nanoseconds on a single POWER5 CPU
to more than 100 microseconds on 64 such CPUs.
Quick Quiz B.17: Couldn’t the update-side batching optimization described in Appendix B.6
be applied to the implementation shown in Listing B.14?
This implementation suffers from some serious shortcomings in addition to the

high update-side overhead noted earlier. First, it is no longer permissible to nest RCU
read-side critical sections, a topic that is taken up in the next section. Second, if a
reader is preempted at line 3 of Listing B.14 after fetching from rcu_gp_ctr but before
storing to rcu_reader_gp, and if the rcu_gp_ctr counter then runs through more
than half but less than all of its possible values, then synchronize_rcu() will ignore
the subsequent RCU read-side critical section. Third and finally, this implementation
requires that the enclosing software environment be able to enumerate threads and
maintain per-thread variables.
Quick Quiz B.18: Is the possibility of readers being preempted in lines 3–4 of Listing B.14 a
real problem, in other words, is there a real sequence of events that could lead to failure? If not,
why not? If so, what is the sequence of events, and how can the failure be addressed?
B.8 Nestable RCU Based on Free-Running Counter

Listing B.16 (rcu_nest.h and rcu_nest.c) shows an RCU implementation based on
a single global free-running counter, but that permits nesting of RCU read-side critical
sections. This nestability is accomplished by reserving the low-order bits of the global
rcu_gp_ctr to count nesting, using the definitions shown in Listing B.15. This is a
generalization of the scheme in Appendix B.7, which can be thought of as having a
single low-order bit reserved for counting nesting depth. Two C-preprocessor macros are
used to arrange this, RCU_GP_CTR_NEST_MASK and RCU_GP_CTR_BOTTOM_BIT. These
are related: RCU_GP_CTR_NEST_MASK=RCU_GP_CTR_BOTTOM_BIT-1. The RCU_GP_
CTR_BOTTOM_BIT macro contains a single bit that is positioned just above the bits
reserved for counting nesting, and the RCU_GP_CTR_NEST_MASK has all one bits covering
the region of rcu_gp_ctr used to count nesting. Obviously, these two C-preprocessor
macros must reserve enough of the low-order bits of the counter to permit the maximum
required nesting of RCU read-side critical sections, and this implementation reserves
seven bits, for a maximum RCU read-side critical-section nesting depth of 127, which
should be well in excess of that needed by most applications.
The resulting rcu_read_lock() implementation is still reasonably straightforward.
Line 6 places a pointer to this thread’s instance of rcu_reader_gp into the local variable
rrgp, minimizing the number of expensive calls to the pthreads thread-local-state API.
Line 7 records the current value of rcu_reader_gp into another local variable tmp,
and line 8 checks to see if the low-order bits are zero, which would indicate that this
v2023.06.11a
B.8. NESTABLE RCU BASED ON FREE-RUNNING COUNTER 655
Listing B.16: Nestable RCU Using a Free-Running Counter

2 {
3 unsigned long tmp;
4 unsigned long *rrgp;
5
6 rrgp = &__get_thread_var(rcu_reader_gp);
7 tmp = *rrgp;
8 if ((tmp & RCU_GP_CTR_NEST_MASK) == 0)
9 tmp = READ_ONCE(rcu_gp_ctr);
10 tmp++;
11 WRITE_ONCE(*rrgp, tmp);
12 smp_mb();
13 }
14
16 {
17 smp_mb();
18 __get_thread_var(rcu_reader_gp)--;
19 }
20
22 {
23 int t;
24
25 smp_mb();
27 WRITE_ONCE(rcu_gp_ctr, rcu_gp_ctr +
28 RCU_GP_CTR_BOTTOM_BIT);
29 smp_mb();
31 while (rcu_gp_ongoing(t) &&
32 ((READ_ONCE(per_thread(rcu_reader_gp, t)) -
33 rcu_gp_ctr) < 0)) {
34 poll(NULL, 0, 10);
35 }
36 }
38 smp_mb();
39 }
is the outermost rcu_read_lock(). If so, line 9 places the global rcu_gp_ctr into
tmp because the current value previously fetched by line 7 is likely to be obsolete. In
either case, line 10 increments the nesting depth, which you will recall is stored in the
seven low-order bits of the counter. Line 11 stores the updated counter back into this
thread’s instance of rcu_reader_gp, and, finally, line 12 executes a memory barrier to
prevent the RCU read-side critical section from bleeding out into the code preceding
the call to rcu_read_lock().
In other words, this implementation of rcu_read_lock() picks up a copy of the
global rcu_gp_ctr unless the current invocation of rcu_read_lock() is nested
within an RCU read-side critical section, in which case it instead fetches the contents of
the current thread’s instance of rcu_reader_gp. Either way, it increments whatever
value it fetched in order to record an additional nesting level, and stores the result in the
current thread’s instance of rcu_reader_gp.
Interestingly enough, despite their rcu_read_lock() differences, the implementa-
tion of rcu_read_unlock() is broadly similar to that shown in Appendix B.7. Line 17
executes a memory barrier in order to prevent the RCU read-side critical section from
bleeding out into code following the call to rcu_read_unlock(), and line 18 decre-
ments this thread’s instance of rcu_reader_gp, which has the effect of decrementing
the nesting count contained in rcu_reader_gp’s low-order bits. Debugging versions
v2023.06.11a
of this primitive would check (before decrementing!) that these low-order bits were
non-zero.
The implementation of synchronize_rcu() is quite similar to that shown in
Appendix B.7. There are two differences. The first is that lines 27 and 28 adds
RCU_GP_CTR_BOTTOM_BIT to the global rcu_gp_ctr instead of adding the constant
“2”, and the second is that the comparison on line 31 has been abstracted out to a separate
function, where it checks the bits indicated by RCU_GP_CTR_NEST_MASK instead of
unconditionally checking the low-order bit.
This approach achieves read-side performance almost equal to that shown in Appen-
dix B.7, incurring roughly 65 nanoseconds of overhead regardless of the number of
POWER5 CPUs. Updates again incur more overhead, ranging from about 600 nanosec-
onds on a single POWER5 CPU to more than 100 microseconds on 64 such CPUs.
Quick Quiz B.19: Why not simply maintain a separate per-thread nesting-level variable, as
was done in previous section, rather than having all this complicated bit manipulation?
This implementation suffers from the same shortcomings as does that of Appendix B.7,
except that nesting of RCU read-side critical sections is now permitted. In addition,
on 32-bit systems, this approach shortens the time required to overflow the global
rcu_gp_ctr variable. The following section shows one way to greatly increase the
time required for overflow to occur, while greatly reducing read-side overhead.
Quick Quiz B.20: Given the algorithm shown in Listing B.16, how could you double the time
required to overflow the global rcu_gp_ctr?
Quick Quiz B.21: Again, given the algorithm shown in Listing B.16, is counter overflow
fatal? Why or why not? If it is fatal, what can be done to fix it?
B.9 RCU Based on Quiescent States

Listing B.18 (rcu_qs.h) shows the read-side primitives used to construct a user-
level implementation of RCU based on quiescent states, with the data shown in
Listing B.17. As can be seen from lines 1–7 in the listing, the rcu_read_lock()
and rcu_read_unlock() primitives do nothing, and can in fact be expected to be
inlined and optimized away, as they are in server builds of the Linux kernel. This
is due to the fact that quiescent-state-based RCU implementations approximate the
extents of RCU read-side critical sections using the aforementioned quiescent states.
Each of these quiescent states contains a call to rcu_quiescent_state(), which
is shown from lines 9–15 in the listing. Threads entering extended quiescent states
(for example, when blocking) may instead call rcu_thread_offline() (lines 17–23)
when entering an extended quiescent state and then call rcu_thread_online()
(lines 25–28) when leaving it. As such, rcu_thread_online() is analogous to rcu_
read_lock() and rcu_thread_offline() is analogous to rcu_read_unlock().
In addition, rcu_quiescent_state() can be thought of as a rcu_thread_online()
immediately followed by a rcu_thread_offline().3 It is illegal to invoke rcu_
quiescent_state(), rcu_thread_offline(), or rcu_thread_online() from
an RCU read-side critical section.
3 Although the code in the listing is consistent with rcu_quiescent_state() being
the same as rcu_thread_online() immediately followed by rcu_thread_offline(),

this relationship is obscured by performance optimizations.
v2023.06.11a
B.9. RCU BASED ON QUIESCENT STATES 657
Listing B.17: Data for Quiescent-State-Based RCU

2 long rcu_gp_ctr = 0;
3 DEFINE_PER_THREAD(long, rcu_reader_qs_gp);
Listing B.18: Quiescent-State-Based RCU Read Side

2 {
3 }
4
6 {
7 }
8
9 static void rcu_quiescent_state(void)
10 {
11 smp_mb();
12 __get_thread_var(rcu_reader_qs_gp) =
13 READ_ONCE(rcu_gp_ctr) + 1;
14 smp_mb();
15 }
16
17 static void rcu_thread_offline(void)
18 {
19 smp_mb();
20 __get_thread_var(rcu_reader_qs_gp) =
21 READ_ONCE(rcu_gp_ctr);
22 smp_mb();
23 }
24
25 static void rcu_thread_online(void)
26 {
27 rcu_quiescent_state();
28 }
In rcu_quiescent_state(), line 11 executes a memory barrier to prevent any

code prior to the quiescent state (including possible RCU read-side critical sections)
from being reordered into the quiescent state. Lines 12–13 pick up a copy of the global
rcu_gp_ctr, using READ_ONCE() to ensure that the compiler does not employ any
optimizations that would result in rcu_gp_ctr being fetched more than once, and
then adds one to the value fetched and stores it into the per-thread rcu_reader_qs_gp
variable, so that any concurrent instance of synchronize_rcu() will see an odd-
numbered value, thus becoming aware that a new RCU read-side critical section has
started. Instances of synchronize_rcu() that are waiting on older RCU read-side
critical sections will thus know to ignore this new one. Finally, line 14 executes a
memory barrier, which prevents subsequent code (including a possible RCU read-side
critical section) from being re-ordered with the lines 12–13.
Quick Quiz B.22: Doesn’t the additional memory barrier shown on line 14 of Listing B.18
greatly increase the overhead of rcu_quiescent_state?
Some applications might use RCU only occasionally, but use it very heavily when
they do use it. Such applications might choose to use rcu_thread_online() when
starting to use RCU and rcu_thread_offline() when no longer using RCU. The
time between a call to rcu_thread_offline() and a subsequent call to rcu_thread_
online() is an extended quiescent state, so that RCU will not expect explicit quiescent
states to be registered during this time.
v2023.06.11a
Listing B.19: RCU Update Side Using Quiescent States

2 {
3 int t;
4
5 smp_mb();
7 WRITE_ONCE(rcu_gp_ctr, rcu_gp_ctr + 2);
8 smp_mb();
10 while (rcu_gp_ongoing(t) &&
11 ((per_thread(rcu_reader_qs_gp, t)
12 - rcu_gp_ctr) < 0)) {
13 poll(NULL, 0, 10);
14 }
15 }
17 smp_mb();
18 }
The rcu_thread_offline() function simply sets the per-thread rcu_reader_

qs_gp variable to the current value of rcu_gp_ctr, which has an even-numbered value.
Any concurrent instances of synchronize_rcu() will thus know to ignore this thread.
Quick Quiz B.23: Why are the two memory barriers on lines 11 and 14 of Listing B.18
needed?
The rcu_thread_online() function simply invokes rcu_quiescent_state(),

thus marking the end of the extended quiescent state.
Listing B.19 (rcu_qs.c) shows the implementation of synchronize_rcu(), which
is quite similar to that of the preceding sections.
This implementation has blazingly fast read-side primitives, with an rcu_read_
lock()–rcu_read_unlock() round trip incurring an overhead of roughly 50 picosec-
onds. The synchronize_rcu() overhead ranges from about 600 nanoseconds on a
single-CPU POWER5 system up to more than 100 microseconds on a 64-CPU system.
Quick Quiz B.24: To be sure, the clock frequencies of POWER systems in 2008 were
quite high, but even a 5 GHz clock frequency is insufficient to allow loops to be executed in
50 picoseconds! What is going on here?
However, this implementation requires that each thread either invoke rcu_
quiescent_state() periodically or to invoke rcu_thread_offline() for extended
quiescent states. The need to invoke these functions periodically can make this im-
plementation difficult to use in some situations, such as for certain types of library
functions.
Quick Quiz B.25: Why would the fact that the code is in a library make any difference for
how easy it is to use the RCU implementation shown in Listings B.18 and B.19?
Quick Quiz B.26: But what if you hold a lock across a call to synchronize_rcu(), and then
acquire that same lock within an RCU read-side critical section? This should be a deadlock, but
how can a primitive that generates absolutely no code possibly participate in a deadlock cycle?
In addition, this implementation does not permit concurrent calls to synchronize_

rcu() to share grace periods. That said, one could easily imagine a production-quality
RCU implementation based on this version of RCU.
v2023.06.11a
B.10. SUMMARY OF TOY RCU IMPLEMENTATIONS 659
B.10 Summary of Toy RCU Implementations

If you made it this far, congratulations! You should now have a much clearer
understanding not only of RCU itself, but also of the requirements of enclosing
software environments and applications. Those wishing an even deeper under-
standing are invited to read descriptions of production-quality RCU implementa-
tions [DMS+ 12, McK07a, McK08b, McK09a].
The preceding sections listed some desirable properties of the various RCU primitives.
The following list is provided for easy reference for those wishing to create a new RCU
implementation.
1. There must be read-side primitives (such as rcu_read_lock() and rcu_read_

unlock()) and grace-period primitives (such as synchronize_rcu() and call_
rcu()), such that any RCU read-side critical section in existence at the start of a
grace period has completed by the end of the grace period.
2. RCU read-side primitives should have minimal overhead. In particular, expensive
operations such as cache misses, atomic instructions, memory barriers, and
branches should be avoided.
3. RCU read-side primitives should have O (1) computational complexity to enable
real-time use. (This implies that readers run concurrently with updaters.)
4. RCU read-side primitives should be usable in all contexts (in the Linux kernel, they
are permitted everywhere except in the idle loop). An important special case is
that RCU read-side primitives be usable within an RCU read-side critical section,
in other words, that it be possible to nest RCU read-side critical sections.
5. RCU read-side primitives should be unconditional, with no failure returns. This
property is extremely important, as failure checking increases complexity and
complicates testing and validation.
6. Any operation other than a quiescent state (and thus a grace period) should be
permitted in an RCU read-side critical section. In particular, irrevocable operations
such as I/O should be permitted.
7. It should be possible to update an RCU-protected data structure while executing
within an RCU read-side critical section.
8. Both RCU read-side and update-side primitives should be independent of memory
allocator design and implementation, in other words, the same RCU implementation
should be able to protect a given data structure regardless of how the data elements
are allocated and freed.
9. RCU grace periods should not be blocked by threads that halt outside of RCU read-
side critical sections. (But note that most quiescent-state-based implementations
violate this desideratum.)
Quick Quiz B.27: Given that grace periods are prohibited within RCU read-side critical
sections, how can an RCU data structure possibly be updated while in an RCU read-side critical
section?
v2023.06.11a
v2023.06.11a
Order! Order in the court!
Unknown
Appendix C
Why Memory Barriers?
So what possessed CPU designers to cause them to inflict memory barriers on poor
unsuspecting SMP software designers?
In short, because reordering memory references allows much better performance,
courtesy of the finite speed of light and the non-zero size of atoms noted in Section 3.2,
and particularly in the hardware-performance question posed by Quick Quiz 3.7.
Therefore, memory barriers are needed to force ordering in things like synchronization
primitives whose correct operation depends on ordered memory references.
Getting a more detailed answer to this question requires a good understanding of how
CPU caches work, and especially what is required to make caches really work well. The
following sections:
1. Present the structure of a cache,
2. Describe how cache-coherency protocols ensure that CPUs agree on the value of
each location in memory, and, finally,
3. Outline how store buffers and invalidate queues help caches and cache-coherency
protocols achieve high performance.
We will see that memory barriers are a necessary evil that is required to enable good
performance and scalability, an evil that stems from the fact that CPUs are orders of
magnitude faster than are both the interconnects between them and the memory they are
attempting to access.
C.1 Cache Structure

Modern CPUs are much faster than are modern memory systems. A 2006 CPU might
be capable of executing ten instructions per nanosecond, but will require many tens of
nanoseconds to fetch a data item from main memory. This disparity in speed—more
than two orders of magnitude—has resulted in the multi-megabyte caches found on
modern CPUs. These caches are associated with the CPUs as shown in Figure C.1, and
can typically be accessed in a few cycles.1
1 It is standard practice to use multiple levels of cache, with a small level-one cache close
to the CPU with single-cycle access time, and a larger level-two cache with a longer access
time, perhaps roughly ten clock cycles. Higher-performance CPUs often have three or even
four levels of cache.
661
v2023.06.11a
662 APPENDIX C. WHY MEMORY BARRIERS?
CPU 0 CPU 1
Cache Cache
Interconnect
Memory
Figure C.1: Modern Computer System Cache Structure
Data flows among the CPUs’ caches and memory in fixed-length blocks called “cache
lines”, which are normally a power of two in size, ranging from 16 to 256 bytes. When a
given data item is first accessed by a given CPU, it will be absent from that CPU’s cache,
meaning that a “cache miss” (or, more specifically, a “startup” or “warmup” cache miss)
has occurred. The cache miss means that the CPU will have to wait (or be “stalled”) for
hundreds of cycles while the item is fetched from memory. However, the item will be
loaded into that CPU’s cache, so that subsequent accesses will find it in the cache and
therefore run at full speed.
After some time, the CPU’s cache will fill, and subsequent misses will likely need to
eject an item from the cache in order to make room for the newly fetched item. Such
a cache miss is termed a “capacity miss”, because it is caused by the cache’s limited
capacity. However, most caches can be forced to eject an old item to make room for a
new item even when they are not yet full. This is due to the fact that large caches are
implemented as hardware hash tables with fixed-size hash buckets (or “sets”, as CPU
designers call them) and no chaining, as shown in Figure C.2.
This cache has sixteen “sets” and two “ways” for a total of 32 “lines”, each entry
containing a single 256-byte “cache line”, which is a 256-byte-aligned block of memory.
This cache line size is a little on the large size, but makes the hexadecimal arithmetic
much simpler. In hardware parlance, this is a two-way set-associative cache, and is
analogous to a software hash table with sixteen buckets, where each bucket’s hash
chain is limited to at most two elements. The size (32 cache lines in this case) and the
associativity (two in this case) are collectively called the cache’s “geometry”. Since this
cache is implemented in hardware, the hash function is extremely simple: Extract four
bits from the memory address.
In Figure C.2, each box corresponds to a cache entry, which can contain a 256-byte
cache line. However, a cache entry can be empty, as indicated by the empty boxes in the
figure. The rest of the boxes are flagged with the memory address of the cache line that
they contain. Since the cache lines must be 256-byte aligned, the low eight bits of each
address are zero, and the choice of hardware hash function means that the next-higher
four bits match the hash line number.
The situation depicted in the figure might arise if the program’s code were located at
address 0x43210E00 through 0x43210EFF, and this program accessed data sequentially
from 0x12345000 through 0x12345EFF. Suppose that the program were now to access
location 0x12345F00. This location hashes to line 0xF, and both ways of this line are
v2023.06.11a
C.1. CACHE STRUCTURE 663
Way 0 Way 1
0x0 0x12345000
0x1 0x12345100
0x2 0x12345200
0x3 0x12345300
0x4 0x12345400
0x5 0x12345500
0x6 0x12345600
0x7 0x12345700
0x8 0x12345800
0x9 0x12345900
0xA 0x12345A00
0xB 0x12345B00
0xC 0x12345C00
0xD 0x12345D00
0xE 0x12345E00 0x43210E00
0xF
Figure C.2: CPU Cache Structure
empty, so the corresponding 256-byte line can be accommodated. If the program were
to access location 0x1233000, which hashes to line 0x0, the corresponding 256-byte
cache line can be accommodated in way 1. However, if the program were to access
location 0x1233E00, which hashes to line 0xE, one of the existing lines must be ejected
from the cache to make room for the new cache line. If this ejected line were accessed
later, a cache miss would result. Such a cache miss is termed an “associativity miss”.
Thus far, we have been considering only cases where a CPU reads a data item. What
happens when it does a write? Because it is important that all CPUs agree on the value
of a given data item, before a given CPU writes to that data item, it must first cause
it to be removed, or “invalidated”, from other CPUs’ caches. Once this invalidation
has completed, the CPU may safely modify the data item. If the data item was present
in this CPU’s cache, but was read-only, this process is termed a “write miss”. Once a
given CPU has completed invalidating a given data item from other CPUs’ caches, that
CPU may repeatedly write (and read) that data item.
Later, if one of the other CPUs attempts to access the data item, it will incur a cache
miss, this time because the first CPU invalidated the item in order to write to it. This
type of cache miss is termed a “communication miss”, since it is usually due to several
CPUs using the data items to communicate (for example, a lock is a data item that is
used to communicate among CPUs using a mutual-exclusion algorithm).
Clearly, much care must be taken to ensure that all CPUs maintain a coherent view
of the data. With all this fetching, invalidating, and writing, it is easy to imagine data
being lost or (perhaps worse) different CPUs having conflicting values for the same data
item in their respective caches. These problems are prevented by “cache-coherency
protocols”, described in the next section.
v2023.06.11a
C.2 Cache-Coherence Protocols

Cache-coherence protocols manage cache-line states so as to prevent inconsistent or
lost data. These protocols can be quite complex, with many tens of states,2 but for our
purposes we need only concern ourselves with the four-state MESI cache-coherence
protocol.
C.2.1 MESI States

MESI stands for “modified”, “exclusive”, “shared”, and “invalid”, the four states a given
cache line can take on using this protocol. Caches using this protocol therefore maintain
a two-bit state “tag” on each cache line in addition to that line’s physical address and
data.
A line in the “modified” state has been subject to a recent memory store from the
corresponding CPU, and the corresponding memory is guaranteed not to appear in any
other CPU’s cache. Cache lines in the “modified” state can thus be said to be “owned”
by the CPU. Because this cache holds the only up-to-date copy of the data, this cache is
ultimately responsible for either writing it back to memory or handing it off to some
other cache, and must do so before reusing this line to hold other data.
The “exclusive” state is very similar to the “modified” state, the single exception
being that the cache line has not yet been modified by the corresponding CPU, which in
turn means that the copy of the cache line’s data that resides in memory is up-to-date.
However, since the CPU can store to this line at any time, without consulting other
CPUs, a line in the “exclusive” state can still be said to be owned by the corresponding
CPU. That said, because the corresponding value in memory is up to date, this cache
can discard this data without writing it back to memory or handing it off to some other
CPU.
A line in the “shared” state might be replicated in at least one other CPU’s cache, so
that this CPU is not permitted to store to the line without first consulting with other
CPUs. As with the “exclusive” state, because the corresponding value in memory is up
to date, this cache can discard this data without writing it back to memory or handing it
off to some other CPU.
A line in the “invalid” state is empty, in other words, it holds no data. When new data
enters the cache, it is placed into a cache line that was in the “invalid” state if possible.
This approach is preferred because replacing a line in any other state could result in an
expensive cache miss should the replaced line be referenced in the future.
Since all CPUs must maintain a coherent view of the data carried in the cache lines,
the cache-coherence protocol provides messages that coordinate the movement of cache
lines through the system.
C.2.2 MESI Protocol Messages

Many of the transitions described in the previous section require communication among
the CPUs. If the CPUs are on a single shared bus, the following messages suffice:
Read:
The “read” message contains the physical address of the cache line to be read.
2 See Culler et al. [CSG99] pages 670 and 671 for the nine-state and 26-state diagrams
for SGI Origin2000 and Sequent (now IBM) NUMA-Q, respectively. Both diagrams are
significantly simpler than real life.
v2023.06.11a
C.2. CACHE-COHERENCE PROTOCOLS 665
Read Response:
The “read response” message contains the data requested by an earlier “read”
message. This “read response” message might be supplied either by memory or by
one of the other caches. For example, if one of the caches has the desired data in
“modified” state, that cache must supply the “read response” message.
Invalidate:
The “invalidate” message contains the physical address of the cache line to be
invalidated. All other caches must remove the corresponding data from their caches
and respond.
Invalidate Acknowledge:
A CPU receiving an “invalidate” message must respond with an “invalidate
acknowledge” message after removing the specified data from its cache.
Read Invalidate:
The “read invalidate” message contains the physical address of the cache line to
be read, while at the same time directing other caches to remove the data. Hence,
it is a combination of a “read” and an “invalidate”, as indicated by its name. A
“read invalidate” message requires both a “read response” and a set of “invalidate
acknowledge” messages in reply.
Writeback:
The “writeback” message contains both the address and the data to be written back
to memory (and perhaps “snooped” into other CPUs’ caches along the way). This
message permits caches to eject lines in the “modified” state as needed to make
room for other data.
Quick Quiz C.1: Where does a writeback message originate from and where does it go to?
Interestingly enough, a shared-memory multiprocessor system really is a message-

passing computer under the covers. This means that clusters of SMP machines that use
distributed shared memory are using message passing to implement shared memory at
two different levels of the system architecture.
Quick Quiz C.2: What happens if two CPUs attempt to invalidate the same cache line
concurrently?
Quick Quiz C.3: When an “invalidate” message appears in a large multiprocessor, every CPU
must give an “invalidate acknowledge” response. Wouldn’t the resulting “storm” of “invalidate
acknowledge” responses totally saturate the system bus?
Quick Quiz C.4: If SMP machines are really using message passing anyway, why bother with
SMP at all?
C.2.3 MESI State Diagram

A given cache line’s state changes as protocol messages are sent and received, as shown
in Figure C.3.
The transition arcs in this figure are as follows:
v2023.06.11a
a f
b c d e
g
E S
h
j k
i l
I
Figure C.3: MESI Cache-Coherency State Diagram
Transition (a):
A cache line is written back to memory, but the CPU retains it in its cache and
further retains the right to modify it. This transition requires a “writeback” message.
Transition (b):
The CPU writes to the cache line that it already had exclusive access to. This
transition does not require any messages to be sent or received.
Transition (c):
The CPU receives a “read invalidate” message for a cache line that it has modified.
The CPU must invalidate its local copy, then respond with both a “read response”
and an “invalidate acknowledge” message, both sending the data to the requesting
CPU and indicating that it no longer has a local copy.
Transition (d):
The CPU does an atomic read-modify-write operation on a data item that was not
present in its cache. It transmits a “read invalidate”, receiving the data via a “read
response”. The CPU can complete the transition once it has also received a full set
of “invalidate acknowledge” responses.
Transition (e):
The CPU does an atomic read-modify-write operation on a data item that was
previously read-only in its cache. It must transmit “invalidate” messages, and must
wait for a full set of “invalidate acknowledge” responses before completing the
transition.
Transition (f):
Some other CPU reads the cache line, and it is supplied from this CPU’s cache,
which retains a read-only copy, possibly also writing it back to memory. This
transition is initiated by the reception of a “read” message, and this CPU responds
with a “read response” message containing the requested data.
Transition (g):
Some other CPU reads a data item in this cache line, and it is supplied either from
this CPU’s cache or from memory. In either case, this CPU retains a read-only
v2023.06.11a
C.2. CACHE-COHERENCE PROTOCOLS 667
copy. This transition is initiated by the reception of a “read” message, and this
CPU responds with a “read response” message containing the requested data.
Transition (h):
This CPU realizes that it will soon need to write to some data item in this cache
line, and thus transmits an “invalidate” message. The CPU cannot complete
the transition until it receives a full set of “invalidate acknowledge” responses,
indicating that no other CPU has this cacheline in its cache. In other words, this
CPU is the only CPU caching it.
Transition (i):
Some other CPU does an atomic read-modify-write operation on a data item in a
cache line held only in this CPU’s cache, so this CPU invalidates it from its cache.
This transition is initiated by the reception of a “read invalidate” message, and
this CPU responds with both a “read response” and an “invalidate acknowledge”
message.
Transition (j):
This CPU does a store to a data item in a cache line that was not in its cache, and
thus transmits a “read invalidate” message. The CPU cannot complete the transition
until it receives the “read response” and a full set of “invalidate acknowledge”
messages. The cache line will presumably transition to “modified” state via
transition (b) as soon as the actual store completes.
Transition (k):
This CPU loads a data item in a cache line that was not in its cache. The CPU
transmits a “read” message, and completes the transition upon receiving the
corresponding “read response”.
Transition (l):
Some other CPU does a store to a data item in this cache line, but holds this cache
line in read-only state due to its being held in other CPUs’ caches (such as the
current CPU’s cache). This transition is initiated by the reception of an “invalidate”
message, and this CPU responds with an “invalidate acknowledge” message.
Quick Quiz C.5: How does the hardware handle the delayed transitions described above?
C.2.4 MESI Protocol Example

Let’s now look at this from the perspective of a cache line’s worth of data, initially
residing in memory at address 0, as it travels through the various single-line direct-
mapped caches in a four-CPU system. Table C.1 shows this flow of data, with the
first column showing the sequence of operations, the second the CPU performing the
operation, the third the operation being performed, the next four the state of each CPU’s
cache line (memory address followed by MESI state), and the final two columns whether
the corresponding memory contents are up to date (“V”) or not (“I”).
Initially, the CPU cache lines in which the data would reside are in the “invalid” state,
and the data is valid in memory. When CPU 0 loads the data at address 0, it enters the
“shared” state in CPU 0’s cache, and is still valid in memory. CPU 3 also loads the data
at address 0, so that it is in the “shared” state in both CPUs’ caches, and is still valid in
memory. Next CPU 0 loads some other cache line (at address 8), which forces the data
v2023.06.11a
Table C.1: Cache Coherence Example
CPU Cache Memory
Sequence # CPU # Operation 0 1 2 3 0 8
0 Initial State −/I −/I −/I −/I V V

1 0 Load 0/S −/I −/I −/I V V
2 3 Load 0/S −/I −/I 0/S V V
3 0 Invalidation 8/S −/I −/I 0/S V V
4 2 RMW 8/S −/I 0/E −/I V V
5 2 Store 8/S −/I 0/M −/I I V
6 1 Atomic Inc 8/S 0/M −/I −/I I V
7 1 Writeback 8/S 8/S −/I −/I V V
at address 0 out of its cache via an invalidation, replacing it with the data at address 8.
CPU 2 now does a load from address 0, but this CPU realizes that it will soon need to
store to it, and so it uses a “read invalidate” message in order to gain an exclusive copy,
invalidating it from CPU 3’s cache (though the copy in memory remains up to date).
Next CPU 2 does its anticipated store, changing the state to “modified”. The copy of
the data in memory is now out of date. CPU 1 does an atomic increment, using a “read
invalidate” to snoop the data from CPU 2’s cache and invalidate it, so that the copy in
CPU 1’s cache is in the “modified” state (and the copy in memory remains out of date).
Finally, CPU 1 reads the cache line at address 8, which uses a “writeback” message to
push address 0’s data back out to memory.
Note that we end with data in some of the CPU’s caches.
Quick Quiz C.6: What sequence of operations would put the CPUs’ caches all back into the
“invalid” state?
C.3 Stores Result in Unnecessary Stalls

Although the cache structure shown in Figure C.1 provides good performance for
repeated reads and writes from a given CPU to a given item of data, its performance
for the first write to a given cache line is quite poor. To see this, consider Figure C.4,
which shows a timeline of a write by CPU 0 to a cacheline held in CPU 1’s cache. Since
CPU 0 must wait for the cache line to arrive before it can write to it, CPU 0 must stall
for an extended period of time.3
But there is no real reason to force CPU 0 to stall for so long—after all, regardless
of what data happens to be in the cache line that CPU 1 sends it, CPU 0 is going to
unconditionally overwrite it.
C.3.1 Store Buffers

One way to prevent this unnecessary stalling of writes is to add “store buffers” between
each CPU and its cache, as shown in Figure C.5. With the addition of these store buffers,
3 The time required to transfer a cache line from one CPU’s cache to another’s is typically
a few orders of magnitude more than that required to execute a simple register-to-register
instruction.
v2023.06.11a
C.3. STORES RESULT IN UNNECESSARY STALLS 669
CPU 0 CPU 1
Write
Invalidate
Stall
Acknowledgement
Figure C.4: Writes See Unnecessary Stalls
CPU 0 can simply record its write in its store buffer and continue executing. When the
cache line does finally make its way from CPU 1 to CPU 0, the data will be moved from
the store buffer to the cache line.
Quick Quiz C.7: But then why do uniprocessors also have store buffers?
Please note that the store buffer does not necessarily operate on full cache lines. The
reason for this is that a given store-buffer entry need only contain the value stored,
not the other data contained in the corresponding cache line. Which is a good thing,
because the CPU doing the store has no idea what that other data might be! But once
the corresponding cache line arrives, any values from the store buffer that update that
cache line can be merged into it, and the corresponding entries can then be removed
from the store buffer. Any other data in that cache line is of course left intact.
Quick Quiz C.8: So store-buffer entries are variable length? Isn’t that difficult to implement
in hardware?
These store buffers are local to a given CPU or, on systems with hardware multithread-
ing, local to a given core. Either way, a given CPU is permitted to access only the store
buffer assigned to it. For example, in Figure C.5, CPU 0 cannot access CPU 1’s store
buffer and vice versa. This restriction simplifies the hardware by separating concerns:
The store buffer improves performance for consecutive writes, while the responsibility
for communicating among CPUs (or cores, as the case may be) is fully shouldered by the
cache-coherence protocol. However, even given this restriction, there are complications
that must be addressed, which are covered in the next two sections.
C.3.2 Store Forwarding

To see the first complication, a violation of self-consistency, consider the following
code with variables “a” and “b” both initially zero, and with the cache line containing
variable “a” initially owned by CPU 1 and that containing “b” initially owned by CPU 0:
v2023.06.11a
CPU 0 CPU 1
Store Store
Buffer Buffer
Cache Cache
Interconnect
Memory
Figure C.5: Caches With Store Buffers
1 a = 1;
2 b = a + 1;
3 assert(b == 2);
One would not expect the assertion to fail. However, if one were foolish enough to
use the very simple architecture shown in Figure C.5, one would be surprised. Such a
system could potentially see the following sequence of events:
1 CPU 0 starts executing the a = 1.
2 CPU 0 looks “a” up in the cache, and finds that it is missing.
3 CPU 0 therefore sends a “read invalidate” message in order to get exclusive

ownership of the cache line containing “a”.
4 CPU 0 records the store to “a” in its store buffer.
5 CPU 1 receives the “read invalidate” message, and responds by transmitting the
cache line and removing that cacheline from its cache.
6 CPU 0 starts executing the b = a + 1.
7 CPU 0 receives the cache line from CPU 1, which still has a value of zero for “a”.
8 CPU 0 loads “a” from its cache, finding the value zero.
9 CPU 0 applies the entry from its store buffer to the newly arrived cache line, setting
the value of “a” in its cache to one.
10 CPU 0 adds one to the value zero loaded for “a” above, and stores it into the cache
line containing “b” (which we will assume is already owned by CPU 0).
11 CPU 0 executes assert(b == 2), which fails.
v2023.06.11a
CPU 0 CPU 1
Store Store
Buffer Buffer
Cache Cache
Interconnect
Memory
Figure C.6: Caches With Store Forwarding
The problem is that we have two copies of “a”, one in the cache and the other in the
store buffer.
This example breaks a very important guarantee, namely that each CPU will always
see its own operations as if they happened in program order. Breaking this guarantee is
violently counter-intuitive to software types, so much so that the hardware guys took
pity and implemented “store forwarding”, where each CPU refers to (or “snoops”) its
store buffer as well as its cache when performing loads, as shown in Figure C.6. In other
words, a given CPU’s stores are directly forwarded to its subsequent loads, without
having to pass through the cache.
With store forwarding in place, item 8 in the above sequence would have found the
correct value of 1 for “a” in the store buffer, so that the final value of “b” would have
been 2, as one would hope.
C.3.3 Store Buffers and Memory Barriers

To see the second complication, a violation of global memory ordering, consider the
following code sequences with variables “a” and “b” initially zero:
1 void foo(void)
2 {
3 a = 1;
4 b = 1;
5 }
6
7 void bar(void)
8 {
9 while (b == 0) continue;
10 assert(a == 1);
11 }
Suppose CPU 0 executes foo() and CPU 1 executes bar(). Suppose further that
the cache line containing “a” resides only in CPU 1’s cache, and that the cache line
containing “b” is owned by CPU 0. Then the sequence of operations might be as follows:
v2023.06.11a
1 CPU 0 executes a = 1. The cache line is not in CPU 0’s cache, so CPU 0 places
the new value of “a” in its store buffer and transmits a “read invalidate” message.
2 CPU 1 executes while (b == 0)continue, but the cache line containing “b” is
not in its cache. It therefore transmits a “read” message.
3 CPU 0 executes b = 1. It already owns this cache line (in other words, the cache
line is already in either the “modified” or the “exclusive” state), so it stores the
new value of “b” in its cache line.
4 CPU 0 receives the “read” message, and transmits the cache line containing the
now-updated value of “b” to CPU 1, also marking the line as “shared” in its own
cache (but only after writing back the line containing “b” to main memory).
5 CPU 1 receives the cache line containing “b” and installs it in its cache.
6 CPU 1 can now finish executing while (b == 0)continue, and since it finds
that the value of “b” is 1, it proceeds to the next statement.
7 CPU 1 executes the assert(a == 1), and, since CPU 1 is working with the old
value of “a”, this assertion fails.
8 CPU 1 receives the “read invalidate” message, and transmits the cache line
containing “a” to CPU 0 and invalidates this cache line from its own cache. But it
is too late.
9 CPU 0 receives the cache line containing “a” and applies the buffered store just in
time to fall victim to CPU 1’s failed assertion.
Quick Quiz C.9: In step 1 above, why does CPU 0 need to issue a “read invalidate” rather
than a simple “invalidate”? After all, foo() will overwrite the variable a in any case, so why
should it care about the old value of a?
Quick Quiz C.10: In step 4 above, don’t systems avoid that store to memory?
Quick Quiz C.11: In step 9 above, did bar() read a stale value from a, or did its reads of b
and a get reordered?
The hardware designers cannot help directly here, since the CPUs have no idea which
variables are related, let alone how they might be related. Therefore, the hardware
designers provide memory-barrier instructions to allow the software to tell the CPU
about such relations. The program fragment must be updated to contain the memory
barrier:
1 void foo(void)
2 {
3 a = 1;
4 smp_mb();
5 b = 1;
6 }
7
8 void bar(void)
9 {
11 assert(a == 1);
12 }
v2023.06.11a
The memory barrier smp_mb() will cause the CPU to flush its store buffer before
applying each subsequent store to its variable’s cache line. The CPU could either simply
stall until the store buffer was empty before proceeding, or it could use the store buffer to
hold subsequent stores until all of the prior entries in the store buffer had been applied.
With this latter approach the sequence of operations might be as follows:
1 CPU 0 executes a = 1. The cache line is not in CPU 0’s cache, so CPU 0 places
the new value of “a” in its store buffer and transmits a “read invalidate” message.
3 CPU 0 executes smp_mb(), and marks all current store-buffer entries (namely, the
a = 1).
line is already in either the “modified” or the “exclusive” state), but there is a
marked entry in the store buffer. Therefore, rather than store the new value of “b”
in the cache line, it instead places it in the store buffer (but in an unmarked entry).
original value of “b” to CPU 1. It also marks its own copy of this cache line as
“shared”.
7 CPU 1 can now load the value of “b”, but since it finds that the value of “b” is
still 0, it repeats the while statement. The new value of “b” is safely hidden in
CPU 0’s store buffer.
8 CPU 1 receives the “read invalidate” message, and transmits the cache line
containing “a” to CPU 0 and invalidates this cache line from its own cache.
9 CPU 0 receives the cache line containing “a” and applies the buffered store, placing
this line into the “modified” state.
10 Since the store to “a” was the only entry in the store buffer that was marked by the
smp_mb(), CPU 0 can also store the new value of “b”—except for the fact that the
cache line containing “b” is now in “shared” state.
11 CPU 0 therefore sends an “invalidate” message to CPU 1.
12 CPU 1 receives the “invalidate” message, invalidates the cache line containing “b”
from its cache, and sends an “acknowledgement” message to CPU 0.
not in its cache. It therefore transmits a “read” message to CPU 0.
14 CPU 0 receives the “acknowledgement” message, and puts the cache line containing
“b” into the “exclusive” state. CPU 0 now stores the new value of “b” into the cache
line.
15 CPU 0 receives the “read” message, and transmits the cache line containing the new
value of “b” to CPU 1. It also marks its own copy of this cache line as “shared”.
v2023.06.11a
17 CPU 1 can now load the value of “b”, and since it finds that the value of “b” is 1, it
exits the while loop and proceeds to the next statement.
18 CPU 1 executes the assert(a == 1), but the cache line containing “a” is no
longer in its cache. Once it gets this cache from CPU 0, it will be working with the
up-to-date value of “a”, and the assertion therefore passes.
Quick Quiz C.12: After step 15 in Appendix C.3.3 on page 673, both CPUs might drop the
cache line containing the new value of “b”. Wouldn’t that cause this new value to be lost?
As you can see, this process involves no small amount of bookkeeping. Even
something intuitively simple, like “load the value of a” can involve lots of complex steps
in silicon.
C.4 Store Sequences Result in Unnecessary Stalls

Unfortunately, each store buffer must be relatively small, which means that a CPU
executing a modest sequence of stores can fill its store buffer (for example, if all of them
result in cache misses). At that point, the CPU must once again wait for invalidations
to complete in order to drain its store buffer before it can continue executing. This
same situation can arise immediately after a memory barrier, when all subsequent store
instructions must wait for invalidations to complete, regardless of whether or not these
stores result in cache misses.
This situation can be improved by making invalidate acknowledge messages arrive
more quickly. One way of accomplishing this is to use per-CPU queues of invalidate
messages, or “invalidate queues”.
C.4.1 Invalidate Queues

One reason that invalidate acknowledge messages can take so long is that they must
ensure that the corresponding cache line is actually invalidated, and this invalidation
can be delayed if the cache is busy, for example, if the CPU is intensively loading and
storing data, all of which resides in the cache. In addition, if a large number of invalidate
messages arrive in a short time period, a given CPU might fall behind in processing
them, thus possibly stalling all the other CPUs.
However, the CPU need not actually invalidate the cache line before sending the
acknowledgement. It could instead queue the invalidate message with the understanding
that the message will be processed before the CPU sends any further messages regarding
that cache line.
C.4.2 Invalidate Queues and Invalidate Acknowledge

Figure C.7 shows a system with invalidate queues. A CPU with an invalidate queue may
acknowledge an invalidate message as soon as it is placed in the queue, instead of having
to wait until the corresponding line is actually invalidated. Of course, the CPU must refer
to its invalidate queue when preparing to transmit invalidation messages—if an entry
for the corresponding cache line is in the invalidate queue, the CPU cannot immediately
v2023.06.11a
C.4. STORE SEQUENCES RESULT IN UNNECESSARY STALLS 675
CPU 0 CPU 1
Store Store
Buffer Buffer
Cache Cache
Invalidate Invalidate
Queue Queue
Interconnect
Memory
Figure C.7: Caches With Invalidate Queues
transmit the invalidate message; it must instead wait until the invalidate-queue entry has
been processed.
Placing an entry into the invalidate queue is essentially a promise by the CPU to
process that entry before transmitting any MESI protocol messages regarding that cache
line. As long as the corresponding data structures are not highly contended, the CPU
will rarely be inconvenienced by such a promise.
However, the fact that invalidate messages can be buffered in the invalidate queue
provides additional opportunity for memory-misordering, as discussed in the next
section.
C.4.3 Invalidate Queues and Memory Barriers
Let us suppose that CPUs queue invalidation requests, but respond to them immediately.
This approach minimizes the cache-invalidation latency seen by CPUs doing stores, but
can defeat memory barriers, as seen in the following example.
Suppose the values of “a” and “b” are initially zero, that “a” is replicated read-only
(MESI “shared” state), and that “b” is owned by CPU 0 (MESI “exclusive” or “modified”
state). Then suppose that CPU 0 executes foo() while CPU 1 executes function bar()
in the following code fragment:
v2023.06.11a
1 void foo(void)
2 {
3 a = 1;
4 smp_mb();
5 b = 1;
6 }
7
8 void bar(void)
9 {
11 assert(a == 1);
12 }
Then the sequence of operations might be as follows:

1 CPU 0 executes a = 1. The corresponding cache line is read-only in CPU 0’s
cache, so CPU 0 places the new value of “a” in its store buffer and transmits an
“invalidate” message in order to flush the corresponding cache line from CPU 1’s
cache.
3 CPU 1 receives CPU 0’s “invalidate” message, queues it, and immediately responds
to it.
4 CPU 0 receives the response from CPU 1, and is therefore free to proceed past
the smp_mb() on line 4 above, moving the value of “a” from its store buffer to its
cache line.
cache.
that the value of “b” is 1, it proceeds to the next statement.
9 CPU 1 executes the assert(a == 1), and, since the old value of “a” is still in
CPU 1’s cache, this assertion fails.
10 Despite the assertion failure, CPU 1 processes the queued “invalidate” message,
and (tardily) invalidates the cache line containing “a” from its own cache.
Quick Quiz C.13: In step 1 of the first scenario in Appendix C.4.3, why is an “invalidate” sent
instead of a ”read invalidate” message? Doesn’t CPU 0 need the values of the other variables
that share this cache line with “a”?
There is clearly not much point in accelerating invalidation responses if doing

so causes memory barriers to effectively be ignored. However, the memory-barrier
v2023.06.11a
C.4. STORE SEQUENCES RESULT IN UNNECESSARY STALLS 677
instructions can interact with the invalidate queue, so that when a given CPU executes a
memory barrier, it marks all the entries currently in its invalidate queue, and forces any
subsequent load to wait until all marked entries have been applied to the CPU’s cache.
Therefore, we can add a memory barrier to function bar as follows:
1 void foo(void)
2 {
3 a = 1;
4 smp_mb();
5 b = 1;
6 }
7
8 void bar(void)
9 {
11 smp_mb();
12 assert(a == 1);
13 }
Quick Quiz C.14: Say what??? Why do we need a memory barrier here, given that the CPU
cannot possibly execute the assert() until after the while loop completes?
With this change, the sequence of operations might be as follows:
1 CPU 0 executes a = 1. The corresponding cache line is read-only in CPU 0’s

cache, so CPU 0 places the new value of “a” in its store buffer and transmits an
“invalidate” message in order to flush the corresponding cache line from CPU 1’s
cache.
3 CPU 1 receives CPU 0’s “invalidate” message, queues it, and immediately responds
to it.
4 CPU 0 receives the response from CPU 1, and is therefore free to proceed past
the smp_mb() on line 4 above, moving the value of “a” from its store buffer to its
cache line.
cache.
that the value of “b” is 1, it proceeds to the next statement, which is now a memory
barrier.
9 CPU 1 must now stall until it processes all pre-existing messages in its invalidation
queue.
v2023.06.11a
10 CPU 1 now processes the queued “invalidate” message, and invalidates the cache
line containing “a” from its own cache.
11 CPU 1 executes the assert(a == 1), and, since the cache line containing “a” is
no longer in CPU 1’s cache, it transmits a “read” message.
12 CPU 0 responds to this “read” message with the cache line containing the new
value of “a”.
13 CPU 1 receives this cache line, which contains a value of 1 for “a”, so that the
assertion does not trigger.
With much passing of MESI messages, the CPUs arrive at the correct answer. This
section illustrates why CPU designers must be extremely careful with their cache-
coherence optimizations. The key requirement is that the memory barriers provide the
appearance of ordering to the software. As long as these appearances are maintained, the
hardware can carry out whatever queueing, buffering, marking, stallings, and flushing
optimizations it likes.
Quick Quiz C.15: Instead of all of this marking of invalidation-queue entries and stalling of
loads, why not simply force an immediate flush of the invalidation queue?
C.5 Read and Write Memory Barriers

In the previous section, memory barriers were used to mark entries in both the store
buffer and the invalidate queue. But in our code fragment, foo() had no reason to do
anything with the invalidate queue, and bar() similarly had no reason to do anything
with the store buffer.
Many CPU architectures therefore provide weaker memory-barrier instructions that
do only one or the other of these two. Roughly speaking, a “read memory barrier”
marks only the invalidate queue (and snoops entries in the store buffer) and a “write
memory barrier” marks only the store buffer, while a full-fledged memory barrier does
all of the above.
The software-visible effect of these hardware mechanisms is that a read memory
barrier orders only loads on the CPU that executes it, so that all loads preceding the
read memory barrier will appear to have completed before any load following the read
memory barrier. Similarly, a write memory barrier orders only stores, again on the
CPU that executes it, and again so that all stores preceding the write memory barrier
will appear to have completed before any store following the write memory barrier. A
full-fledged memory barrier orders both loads and stores, but again only on the CPU
executing the memory barrier.
Quick Quiz C.16: But can’t full memory barriers impose global ordering? After all, isn’t that
needed to provide the ordering shown in Listing 12.27?
If we update foo and bar to use read and write memory barriers, they appear as
follows:
v2023.06.11a
C.6. EXAMPLE MEMORY-BARRIER SEQUENCES 679
1 void foo(void)
2 {
3 a = 1;
4 smp_wmb();
5 b = 1;
6 }
7
8 void bar(void)
9 {
11 smp_rmb();
12 assert(a == 1);
13 }
Some computers have even more flavors of memory barriers, but understanding these
three variants will provide a good introduction to memory barriers in general.
C.6 Example Memory-Barrier Sequences

This section presents some seductive but subtly broken uses of memory barriers.
Although many of them will work most of the time, and some will work all the time
on some specific CPUs, these uses must be avoided if the goal is to produce code that
works reliably on all CPUs. To help us better see the subtle breakage, we first need to
focus on an ordering-hostile architecture.
C.6.1 Ordering-Hostile Architecture

A number of ordering-hostile computer systems have been produced over the decades,
but the nature of the hostility has always been extremely subtle, and understanding it
has required detailed knowledge of the specific hardware. Rather than picking on a
specific hardware vendor, and as a presumably attractive alternative to dragging the
reader through detailed technical specifications, let us instead design a mythical but
maximally memory-ordering-hostile computer architecture.4
This hardware must obey the following ordering constraints [McK05a, McK05b]:
1. Each CPU will always perceive its own memory accesses as occurring in program
order.
2. CPUs will reorder a given operation with a store only if the two operations are
referencing different locations.
3. All of a given CPU’s loads preceding a read memory barrier (smp_rmb()) will be
perceived by all CPUs to precede any loads following that read memory barrier.
4. All of a given CPU’s stores preceding a write memory barrier (smp_wmb()) will be
perceived by all CPUs to precede any stores following that write memory barrier.
4 Readers preferring a detailed look at real hardware architectures are encouraged to
consult CPU vendors’ manuals [SW95, Adv02, Int02a, IBM94, LHF05, SPA94, Int04b,
Int04a, Int04c], Gharachorloo’s dissertation [Gha95], Peter Sewell’s work [Sew], or the
excellent hardware-oriented primer by Sorin, Hill, and Wood [SHW11].
v2023.06.11a
Node 0 Node 1
Cache Cache

Message Message Message Message
Queue Queue Queue Queue
Interconnect
Memory
Figure C.8: Example Ordering-Hostile Architecture
Listing C.1: Memory Barrier Example 1

CPU 0 CPU 1 CPU 2
a = 1;
smp_wmb(); while (b == 0);
b = 1; c = 1; z = c;
smp_rmb();
x = a;
assert(z == 0 || x == 1);
5. All of a given CPU’s accesses (loads and stores) preceding a full memory barrier
(smp_mb()) will be perceived by all CPUs to precede any accesses following that
memory barrier.
Quick Quiz C.17: Does the guarantee that each CPU sees its own memory accesses in order
also guarantee that each user-level thread will see its own memory accesses in order? Why or
why not?
Imagine a large non-uniform cache architecture (NUCA) system that, in order to

provide fair allocation of interconnect bandwidth to CPUs in a given node, provided
per-CPU queues in each node’s interconnect interface, as shown in Figure C.8. Although
a given CPU’s accesses are ordered as specified by memory barriers executed by that
CPU, however, the relative order of a given pair of CPUs’ accesses could be severely
reordered, as we will see.5
C.6.2 Example 1
Listing C.1 shows three code fragments, executed concurrently by CPUs 0, 1, and 2.
Each of “a”, “b”, and “c” are initially zero.
Suppose CPU 0 recently experienced many cache misses, so that its message queue is
full, but that CPU 1 has been running exclusively within the cache, so that its message
5 Any real hardware architect or designer will no doubt be objecting strenuously, as they
just might be a bit upset about the prospect of working out which queue should handle a
message involving a cache line that both CPUs accessed, to say nothing of the many races
that this example poses. All I can say is “Give me a better example”.
v2023.06.11a
C.6. EXAMPLE MEMORY-BARRIER SEQUENCES 681

CPU 0 CPU 1 CPU 2
a = 1; while (a == 0);
smp_mb(); y = b;
b = 1; smp_rmb();
x = a;
assert(y == 0 || x == 1);

CPU 0 CPU 1 CPU 2
1 a = 1;
2 smp_wmb();
3 b = 1; while (b == 0); while (b == 0);
4 smp_mb(); smp_mb();
5 c = 1; d = 1;
6 while (c == 0);
7 while (d == 0);
8 smp_mb();
9 e = 1; assert(e == 0 || a == 1);
queue is empty. Then CPU 0’s assignment to “a” and “b” will appear in Node 0’s cache
immediately (and thus be visible to CPU 1), but will be blocked behind CPU 0’s prior
traffic. In contrast, CPU 1’s assignment to “c” will sail through CPU 1’s previously
empty queue. Therefore, CPU 2 might well see CPU 1’s assignment to “c” before it sees
CPU 0’s assignment to “a”, causing the assertion to fire, despite the memory barriers.
Therefore, portable code cannot rely on this assertion not firing, as both the compiler
and the CPU can reorder the code so as to trip the assertion.
Quick Quiz C.18: Could this code be fixed by inserting a memory barrier between CPU 1’s
“while” and assignment to “c”? Why or why not?
C.6.3 Example 2
Both “a” and “b” are initially zero.
Again, suppose CPU 0 recently experienced many cache misses, so that its message
queue is full, but that CPU 1 has been running exclusively within the cache, so that its
message queue is empty. Then CPU 0’s assignment to “a” will appear in Node 0’s cache
immediately (and thus be visible to CPU 1), but will be blocked behind CPU 0’s prior
traffic. In contrast, CPU 1’s assignment to “b” will sail through CPU 1’s previously
empty queue. Therefore, CPU 2 might well see CPU 1’s assignment to “b” before it sees
CPU 0’s assignment to “a”, causing the assertion to fire, despite the memory barriers.
In theory, portable code should not rely on this example code fragment, however, as
before, in practice it actually does work on most mainstream computer systems.
C.6.4 Example 3
All variables are initially zero.
Note that neither CPU 1 nor CPU 2 can proceed to line 5 until they see CPU 0’s
assignment to “b” on line 3. Once CPU 1 and 2 have executed their memory barriers on
line 4, they are both guaranteed to see all assignments by CPU 0 preceding its memory
barrier on line 2. Similarly, CPU 0’s memory barrier on line 8 pairs with those of
v2023.06.11a
CPUs 1 and 2 on line 4, so that CPU 0 will not execute the assignment to “e” on line 9
until after its assignment to “b” is visible to both of the other CPUs. Therefore, CPU 2’s
assertion on line 9 is guaranteed not to fire.
Quick Quiz C.19: Suppose that lines 3–5 for CPUs 1 and 2 in Listing C.3 are in an interrupt
handler, and that the CPU 2’s line 9 runs at process level. In other words, the code in all three
columns of the table runs on the same CPU, but the first two columns run in an interrupt handler,
and the third column runs at process level, so that the code in third column can be interrupted
by the code in the first two columns. What changes, if any, are required to enable the code to
work correctly, in other words, to prevent the assertion from firing?
Quick Quiz C.20: If CPU 2 executed an assert(e==0||c==1) in the example in Listing C.3,
would this assert ever trigger?
The Linux kernel’s synchronize_rcu() primitive uses an algorithm similar to that

shown in this example.
C.7 Are Memory Barriers Forever?

There have been a number of recent systems that are significantly less aggressive about
out-of-order execution in general and re-ordering memory references in particular. Will
this trend continue to the point where memory barriers are a thing of the past?
The argument in favor would cite proposed massively multi-threaded hardware
architectures, so that each thread would wait until memory was ready, with tens,
hundreds, or even thousands of other threads making progress in the meantime. In such
an architecture, there would be no need for memory barriers, because a given thread
would simply wait for all outstanding operations to complete before proceeding to the
next instruction. Because there would be potentially thousands of other threads, the
CPU would be completely utilized, so no CPU time would be wasted.
The argument against would cite the extremely limited number of applications capable
of scaling up to a thousand threads, as well as increasingly severe realtime requirements,
which are in the tens of microseconds for some applications. The realtime-response
requirements are difficult enough to meet as is, and would be even more difficult
to meet given the extremely low single-threaded throughput implied by the massive
multi-threaded scenarios.
Another argument in favor would cite increasingly sophisticated latency-hiding
hardware implementation techniques that might well allow the CPU to provide the
illusion of fully sequentially consistent execution while still providing almost all of the
performance advantages of out-of-order execution. A counter-argument would cite the
increasingly severe power-efficiency requirements presented both by battery-operated
devices and by environmental responsibility.
Who is right? We have no clue, so we are preparing to live with either scenario.
C.8 Advice to Hardware Designers

There are any number of things that hardware designers can do to make the lives of
software people difficult. Here is a list of a few such things that we have encountered in
the past, presented here in the hope that it might help prevent future such problems:
1. I/O devices that ignore cache coherence.
v2023.06.11a
C.8. ADVICE TO HARDWARE DESIGNERS 683
This charming misfeature can result in DMAs from memory missing recent changes
to the output buffer, or, just as bad, cause input buffers to be overwritten by the
contents of CPU caches just after the DMA completes. To make your system
work in face of such misbehavior, you must carefully flush the CPU caches of
any location in any DMA buffer before presenting that buffer to the I/O device.
Otherwise, a store from one of the CPUs might not be accounted for in the data
DMAed out through the device. This is a form of data corruption, which is an
extremely serious bug.
Similarly, you need to invalidate6 the CPU caches corresponding to any location in
any DMA buffer after DMA to that buffer completes. Otherwise, a given CPU
might see the old data still residing in its cache instead of the newly DMAed data
that it was supposed to see. This is another form of data corruption.
And even then, you need to be very careful to avoid pointer bugs, as even a
misplaced read to an input buffer can result in corrupting the data input! One way
to avoid this is to invalidate all of the caches of all of the CPUs once the DMA
completes, but it is much easier and more efficient if the device DMA participates
in the cache-coherence protocol, making all of this flushing and invalidating
unnecessary.
2. External busses that fail to transmit cache-coherence data.
This is an even more painful variant of the above problem, but causes groups
of devices—and even memory itself—to fail to respect cache coherence. It is
my painful duty to inform you that as embedded systems move to multicore
architectures, we will no doubt see a fair number of such problems arise. By
the year 2021, there were some efforts to address these problems with new
interconnect standards, with some debate as to how effective these standards will
really be [Won19].
3. Device interrupts that ignore cache coherence.

This might sound innocent enough—after all, interrupts aren’t memory references,
are they? But imagine a CPU with a split cache, one bank of which is extremely busy,
therefore holding onto the last cacheline of the input buffer. If the corresponding
I/O-complete interrupt reaches this CPU, then that CPU’s memory reference to the
last cache line of the buffer could return old data, again resulting in data corruption,
but in a form that will be invisible in a later crash dump. By the time the system
gets around to dumping the offending input buffer, the DMA will most likely have
completed.
4. Inter-processor interrupts (IPIs) that ignore cache coherence.
This can be problematic if the IPI reaches its destination before all of the cache
lines in the corresponding message buffer have been committed to memory.
5. Context switches that get ahead of cache coherence.
If memory accesses can complete too wildly out of order, then context switches
can be quite harrowing. If the task flits from one CPU to another before all the
memory accesses visible to the source CPU make it to the destination CPU, then
6 Why not flush? If there is a difference, then a CPU must have incorrectly stored to the
DMA buffer in the midst of the DMA operation.
v2023.06.11a
the task could easily see the corresponding variables revert to prior values, which
can fatally confuse most algorithms.
6. Overly kind simulators and emulators.
It is difficult to write simulators or emulators that force memory re-ordering, so
software that runs just fine in these environments can get a nasty surprise when it
first runs on the real hardware. Unfortunately, it is still the rule that the hardware is
more devious than are the simulators and emulators, but we hope that this situation
changes.
Again, we encourage hardware designers to avoid these practices!
v2023.06.11a
De gustibus non est disputandum.
Latin maxim
Appendix D
Style Guide
This appendix is a collection of style guides which is intended as a reference to improve

consistency in perfbook. It also contains several suggestions and their experimental
examples.
Appendix D.1 describes basic punctuation and spelling rules. Appendix D.2 explains
rules related to unit symbols. Appendix D.3 summarizes LATEX-specific conventions.
D.1 Paul’s Conventions

Following is the list of Paul’s conventions assembled from his answers to Akira’s
questions regarding perfbook’s punctuation policy.
• (On punctuations and quotations) Despite being American myself, for this sort of
book, the UK approach is better because it removes ambiguities like the following:
Type “ls -a,” look for the file “.,” and file a bug if you don’t see it.
The following is much more clear:
Type “ls -a”, look for the file “.”, and file a bug if you don’t see it.
• American English spelling: “color” rather than “colour”.
• Oxford comma: “a, b, and c” rather than “a, b and c”. This is arbitrary. Cases
where the Oxford comma results in ambiguity should be reworded, for example,
by introducing numbering: “a, b, and c and d” should be “(1) a, (2) b, and (3) c
and d”.
• Italic for emphasis. Use sparingly.
• \co{} for identifiers, \url{} for URLs, \path{} for filenames.
• Dates should use an unambiguous format. Never “mm/dd/yy” or “dd/mm/yy”, but

rather “July 26, 2016” or “26 July 2016” or “26-Jul-2016” or “2016/07/26”. I tend
to use yyyy.mm.ddA for filenames, for example.
• North American rules on periods and abbreviations. For example neither of the
following can reasonably be interpreted as two sentences:
685
v2023.06.11a
686 APPENDIX D. STYLE GUIDE
– Say hello, to Mr. Jones.

– If it looks like she sprained her ankle, call Dr. Smith and then tell her to keep
the ankle iced and elevated.
An ambiguous example:
If I take the cow, the pig, the horse, etc. George will be upset.
can be written with more words:
If I take the cow, the pig, the horse, or much of anything else, George
will be upset.
or:
If I take the cow, the pig, the horse, etc., George will be upset.
• I don’t like ampersand (“&”) in headings, but will sometimes use it if doing so
prevents a line break in that heading.
• When mentioning words, I use quotations. When introducing a new word, I use
\emph{}.
Following is a convention regarding punctuation in LATEX sources.
• Place a newline after a colon (:) and the end of a sentence. This avoids the whole
one-space/two-space food fight and also has the advantage of more clearly showing
changes to single sentences in the middle of long paragraphs.
D.2 NIST Style Guide

D.2.1 Unit Symbol
D.2.1.1 SI Unit Symbol
NIST style guide [Nat19, Chapter 5] states the following rules (rephrased for perfbook).
• When SI unit symbols such as “ns”, “MHz”, and “K” (kelvin) are used behind
numerical values, narrow spaces should be placed between the values and the
symbols.
A narrow space can be coded in LATEX by the sequence of “\,”. For example,
“2.4 GHz”, rather then “2.4GHz”.
• Even when the value is used in adjectival sense, a narrow space should be placed.
For example,
“a 10 ms interval”, rather than “a 10-ms interval” nor “a 10ms interval”.
The symbol of micro (µ:10−6 ) can be typeset easily by the help of “gensymb” LATEX
package. A macro “\micro” can be used in both text and math modes. To typeset the
symbol of “microsecond”, you can do so by “\micro s”. For example,
10 µs
Note that math mode “\mu” is italic by default and should not be used as a prefix. An
improper example:
10 𝜇s (math mode “\mu”)
v2023.06.11a
D.2. NIST STYLE GUIDE 687
D.2.1.2 Non-SI Unit Symbol

Although NIST style guide does not cover non-SI unit symbols such as “KB”, “MB”,
and “GB”, the same rule should be followed.
Example:
“A 240 GB hard drive”, rather than “a 240-GB hard drive” nor “a 240GB hard
drive”.
Strictly speaking, NIST guide requires us to use the binary prefixes “Ki”, “Mi”, or
“Gi” to represent powers of 210 . However, we accept the JEDEC conventions to use “K”,
“M”, and “G” as binary prefixes in describing memory capacity [JED].
An acceptable example:
“8 GB of main memory”, meaning “8 GiB of main memory”.
Also, it is acceptable to use just “K”, “M”, or “G” as abbreviations appended to a
numerical value, e.g., “4K entries”. In such cases, no space before an abbreviation is
required. For example,
“8K entries”, rather than “8 K entries”.
If you put a space in between, the symbol looks like a unit symbol and is confusing.
Note that “K” and “k” represent 210 and 103 , respectively. “M” can represent either
220 or 106 , and “G” can represent either 230 or 109 . These ambiguities should not be
confusing in discussing approximate order.
D.2.1.3 Degree Symbol

The angular-degree symbol (°) does not require any space in front of it. NIST style
guide clearly states so.
The symbol of degree can also be typeset easily by the help of gensymb package. A
macro “\degree” can be used in both text and math modes.
Example:
45°, rather than 45 °.
D.2.1.4 Percent Symbol

NIST style guide treats the percent symbol (%) as the same as SI unit symbols.
50 % possibility, rather than 50% possibility.
D.2.1.5 Font Style

Quote from NIST check list [Nata, #6]:
Variables and quantity symbols are in italic type. Unit symbols are in roman
type. Numbers should generally be written in roman type. These rules apply
irrespective of the typeface used in the surrounding text.
For example,
e (elementary charge)
On the other hand, mathematical constants such as the base of natural logarithms
should be roman [Natb]. For example,
e𝑥
v2023.06.11a
Table D.1: Digit-Grouping Style
Style Outputs of \num{}

NIST/SI (English) 12 345 12.345 1 234 567.89
SI (French) 12 345 12,345 1 234 567,89
English 12,345 12.345 1,234,567.89
French 12 345 12,345 1 234 567,89
Other Europe 12.345 12,345 1.234.567,89
D.2.2 NIST Guide Yet To Be Followed

There are a few cases where NIST style guide is not followed. Other English conventions
are followed in such cases.
D.2.2.1 Digit Grouping

Quote from NIST checklist [Nata, #16]:
The digits of numerical values having more than four digits on either side of
the decimal marker are separated into groups of three using a thin, fixed space
counting from both the left and right of the decimal marker. Commas are not
used to separate digits into groups of three.
NIST Example: 15 739.012 53 ms
Our convention: 15,739.01253 ms
In LATEX coding, it is cumbersome to place thin spaces as are recommended in NIST
guide. The \num{} command provided by the “siunitx” package would be of help for
us to follow this rule. It would also help us overcome different conventions. We can
select a specific digit-grouping style as a default in preamble, or specify an option to
each \num{} command as is shown in Table D.1.
As are evident in Table D.1, periods and commas used as other than decimal
markers are confusing and should be avoided, especially in documents expecting global
audiences.
By marking up constant decimal values by \num{} commands, the LATEX source
would be exempted from any particular conventions.
Because of its open-source policy, this approach should give more “portability” to
perfbook.
D.3 LATEX Conventions

Good looking LATEX documents require further considerations on proper use of font
styles, line break exceptions, etc. This section summarizes guidelines specific to LATEX.
D.3.1 Monospace Font

Monospace font (or typewriter font) is heavily used in this textbook. First policy
regarding monospace font in perfbook is to avoid directly using “\texttt” or “\tt”
macro. It is highly recommended to use a macro or an environment indicating the
reason why you want the font.
This section explains the use cases of such macros and environments.
v2023.06.11a
D.3. LATEX CONVENTIONS 689
Listing D.1: LATEX Source of Sample Code Snippet (Current)

1 \begin{listing}
2 \begin{fcvlabel}[ln:base1]
3 \begin{VerbatimL}[commandchars=\$\[\]]
4 /*
5 * Sample Code Snippet
6 */
7 #include <stdio.h>
8 int main(void)
9 {
10 printf("Hello world!\n"); $lnlbl[printf]
11 return 0; $lnlbl[return]
12 }
13 \end{VerbatimL}
14 \end{fcvlabel}
15 \caption{Sample Code Snippet}
16 \label{lst:app:styleguide:Sample Code Snippet}
17 \end{listing}
D.3.1.1 Code Snippet

Because the “verbatim” environment is a primitive way to include listings, we have
transitioned to a scheme which uses the “fancyvrb” package for code snippets.
The goal of the scheme is to extract LATEX sources of code snippets directly from code
samples under CodeSamples directory. It also makes it possible to embed line labels
in the code samples, which can be referenced within the LATEX sources. This reduces
the burden of keeping line numbers in the text consistent with those in code snippets.
Code-snippet extraction is handled by a couple of perl scripts and recipes in Makefile.
We use the escaping feature of the fancyvrb package to embed line labels as comments.
We used to use the “verbbox” environment provided by the “verbatimbox” package.
Appendix D.3.1.2 describes how verbbox can automatically generate line numbers,
but those line numbers cannot be referenced within the LATEX sources.
Let’s start by looking at how code snippets are coded in the current scheme. There are
three customized environments of “Verbatim”. “VerbatimL” is for floating snippets
within the “listing” environment. “VerbatimN” is for inline snippets with line count
enabled. “VerbatimU” is for inline snippets without line count. They are defined in the
preamble as shown below:
\DefineVerbatimEnvironment{VerbatimL}{Verbatim}%
{fontsize=\scriptsize,numbers=left,numbersep=5pt,%
xleftmargin=9pt,obeytabs=true,tabsize=2}
\AfterEndEnvironment{VerbatimL}{\vspace*{-9pt}}
\DefineVerbatimEnvironment{VerbatimN}{Verbatim}%
{fontsize=\scriptsize,numbers=left,numbersep=3pt,%
xleftmargin=5pt,xrightmargin=5pt,obeytabs=true,%
tabsize=2,frame=single}
\DefineVerbatimEnvironment{VerbatimU}{Verbatim}%
{fontsize=\scriptsize,numbers=none,xleftmargin=5pt,%
xrightmargin=5pt,obeytabs=true,tabsize=2,%
samepage=true,frame=single}
The LATEX source of a sample code snippet is shown in Listing D.1 and is typeset as
shown in Listing D.2.
Labels to lines are specified in “$lnlbl[]” command. The characters specified by
“commandchars” option to VarbatimL environment are used by the fancyvrb package
to substitute “\lnlbl{}” for “$lnlbl[]”. Those characters should be selected so that
they don’t appear elsewhere in the code snippet.
Labels “printf” and “return” in Listing D.2 can be referred to as shown below:
v2023.06.11a
Listing D.2: Sample Code Snippet

1 /*
3 */
5 int main(void)
6 {
7 printf("Hello world!\n");
8 return 0;
9 }
Listing D.3: Source of Code Sample with “snippet” Meta Command

1 //\begin{snippet}[labelbase=ln:base1,keepcomment=yes,commandchars=\$\[\]]
2 /*
4 */
6 int main(void)
7 {
8 printf("Hello world!\n"); //\lnlbl{printf}
9 return 0; //\lnlbl{return}
10 }
11 //\end{snippet}
\begin{fcvref}[ln:base1]
\Clnref{printf, return} can be referred
to from text.
\end{fcvref}
Above code results in the paragraph below:
Lines 7 and 8 can be referred to from text.
Macros “\lnlbl{}” and “\lnref{}” are defined in the preamble as follows:
\newcommand{\lnlblbase}{}
\newcommand{\lnlbl}[1]{%
\phantomsection\label{\lnlblbase:#1}}
\newcommand{\lnrefbase}{}
\newcommand{\lnref}[1]{\ref{\lnrefbase:#1}}
Environments “fcvlabel” and “fcvref” are defined as shown below:

\newenvironment{fcvlabel}[1][]{%
\renewcommand{\lnlblbase}{#1}%
\ignorespaces}{\ignorespacesafterend}
\newenvironment{fcvref}[1][]{%
\renewcommand{\lnrefbase}{#1}%
\ignorespaces}{\ignorespacesafterend}
The main part of LATEX source shown on lines 2–14 in Listing D.1 can be extracted
from a code sample of Listing D.3 by a perl script utilities/fcvextract.pl. All
the relevant rules of extraction are described as recipes in the top level Makefile and a
script to generate dependencies (utilities/gen_snippet_d.pl).
As you can see, Listing D.3 has meta commands in comments of C (C++ style). Those
meta commands are interpreted by utilities/fcvextract.pl, which distinguishes
the type of comment style by the suffix of code sample’s file name.
Meta commands which can be used in code samples are listed below:
v2023.06.11a
• \begin{snippet}[<options>]
• \end{snippet}
• \lnlbl{<label string>}
• \fcvexclude
• \fcvblank
“<options>” to the \begin{snippet} meta command is a comma-spareted list of
options shown below:
• labelbase=<label base string>
• keepcomment=yes
• gobbleblank=yes
• commandchars=\X\Y\Z
The “labelbase” option is mandatory and the string given to it will be passed to
the “\begin{fcvlabel}[<label base string>]” command as shown on line 2
of Listing D.1. The “keepcomment=yes” option tells fcvextract.pl to keep com-
ment blocks. Otherwise, comment blocks in C source code will be omitted. The
“gobbleblank=yes” option will remove empty or blank lines in the resulting snippet.
The “commandchars” option is given to the VerbatimL environment as is. At the
moment, it is also mandatory and must come at the end of options listed above. Other
types of options, if any, are also passed to the VerbatimL environment.
The “\lnlbl” commands are converted along the way to reflect the escape-character
choice.1 Source lines with “\fcvexclude” are removed. “\fcvblank” can be used to
keep blank lines when the “gobbleblank=yes” option is specified.
There can be multiple pairs of \begin{snippet} and \end{snippet} as long as
they have unique “labelbase” strings.
Our naming scheme of “labelbase” for unique name space is as follows:
ln:<Chapter/Subdirectory>:<File Name>:<Function Name>
Litmus tests, which are handled by “herdtools7” commands such as “litmus7”

and “herd7”, were problematic in this scheme. Those commands have particular rules
of where comments can be placed and restriction on permitted characters in comments.
They also forbid a couple of tokens to appear in comments. (Tokens in comments might
sound strange, but they do have such restriction.)
For example, the first token in a litmus test must be one of “C”, “PPC”, “X86”, “LISA”,
etc., which indicates the flavor of the test. This means no comment is allowed at the
beginning of a litmus test.
Similarly, several tokens such as “exists”, “filter”, and “locations” indicate
the end of litmus test’s body. Once one of them appears in a litmus test, comments
should be of OCaml style (“(* ... *)”). Those tokens keep the same meaning even
when they appear in comments!
The pair of characters “{” and “}” also have special meaning in the C flavour tests.
They are used to separate portions in a litmus test.
First pair of “{” and “}” encloses initialization part. Comments in this part should
also be in the ocaml form.
You can’t use “{” and “}” in comments in litmus tests, either.
Examples of disallowed comments in a litmus test are shown below:
1 Characters forming comments around the “\lnlbl” commands are also gobbled up
regardless of the “keepcomment” setting.
v2023.06.11a
1 // Comment at first
2 C C-sample
3 // Comment with { and } characters
4 {
5 x=2; // C style comment in initialization
6 }
7
8 P0(int *x}
9 {
10 int r1;
11
12 r1 = READ_ONCE(*x); // Comment with "exists"
13 }
14
15 [...]
16
17 exists (0:r1=0) // C++ style comment after test body
To avoid parse errors, meta commands in litmus tests (C flavor) are embedded in the
following way.
1 C C-SB+o-o+o-o
2 //\begin[snippet][labelbase=ln:base,commandchars=\%\@\$]
3
4 {
5 1:r2=0 (*\lnlbl[initr2]*)
6 }
7
8 P0(int *x0, int *x1) //\lnlbl[P0:b]
9 {
10 int r2;
11
14 } //\lnlbl[P0:e]
15
16 P1(int *x0, int *x1)
17 {
18 int r2;
19
22 }
23
24 //\end[snippet]
25 exists (1:r2=0 /\ 0:r2=0) (* \lnlbl[exists_] *)
Example above is converted to the following intermediate code by a script utilities/

reorder_ltms.pl.2 The intermediate code can be handled by the common script
utilities/fcvextract.pl.
1 // Do not edit!
2 // Generated by utillities/reorder_ltms.pl
3 //\begin{snippet}[labelbase=ln:base,commandchars=\%\@\$]
4 C C-SB+o-o+o-o
5
6 {
7 1:r2=0 //\lnlbl{initr2}
8 }
9
10 P0(int *x0, int *x1) //\lnlbl{P0:b}
11 {
12 int r2;
13
2 Currently, only C flavor litmus tests are supported.
v2023.06.11a
Listing D.4: LATEX Source of Sample Code Snippet (Obsolete)

1 \begin{listing}
2 { \scriptsize
3 \begin{verbbox}[\LstLineNo]
4 /*
6 */
8 int main(void)
9 {
11 return 0;
12 }
13 \end{verbbox}
14 }
15 \centering
16 \theverbbox
17 \caption{Sample Code Snippet (Obsolete)}
18 \label{lst:app:styleguide:Sample Code Snippet (Obsolete)}
19 \end{listing}
Listing D.5: Sample Code Snippet (Obsolete)

1 /*
3 */
5 int main(void)
6 {
8 return 0;
9 }
16 } //\lnlbl{P0:e}
17
18 P1(int *x0, int *x1)
19 {
20 int r2;
21
24 }
25
26 exists (1:r2=0 /\ 0:r2=0) \lnlbl{exists_}
27 //\end{snippet}
Note that each litmus test’s source file can contain at most one pair of
\begin[snippet] and \end[snippet] because of the restriction of comments.
D.3.1.2 Code Snippet (Obsolete)

Sample LATEX source of a code snippet coded using the “verbatimbox” package is
shown in Listing D.4 and is typeset as shown in Listing D.5.
The auto-numbering feature of verbbox is enabled by the “\LstLineNo” macro
specified in the option to verbbox (line 3 in Listing D.4). The macro is defined in the
preamble of perfbook.tex as follows:
\newcommand{\LstLineNo}
{\makebox[5ex][r]{\arabic{VerbboxLineNo}\hspace{2ex}}}
The “verbatim” environment is used for listings with too many lines to fit in a
column. It is also used to avoid overwhelming LATEX with a lot of floating objects. They
are being converted to the scheme using the VerbatimN environment.
v2023.06.11a
Table D.2: Limitation of Monospace Macro
Macro Need Escape Should Avoid
\co, \nbco \, %, {, }
\tco # %, {, }, \
D.3.1.3 Identifier
We use “\co{}” macro for inline identifiers. (“co” stands for “code”.)
By putting them into \co{}, underscore characters in their names are free of escaping
in LATEX source. It is convenient to search them in source files. Also, \co{} macro has
a capability to permit line breaks at particular sequences of letters. Current definition
permits a line break at an underscore (_), two consecutive underscores (__), a white
space, or an operator ->.
D.3.1.4 Identifier inside Table and Heading

Although \co{} command is convenient for inlining within text, it is fragile because
of its capability of line break. When it is used inside a “tabular” environment
or its derivative such as “tabularx”, it confuses column width estimation of those
environments. Furthermore, \co{} can not be safely used in section headings nor
description headings.
As a workaround, we use “\tco{}” command inside tables and headings. It has
no capability of line break at particular sequences, but still frees us from escaping
underscores.
When used in text, \tco{} permits line breaks at white spaces.
D.3.1.5 Other Use Cases of Monospace Font

For URLs, we use “\url{}” command provided by the “hyperref” package. It will
generate hyper references to the URLs.
For path names, we use “\path{}” command. It won’t generate hyper references.
Both \url{} and \path{} permit line breaks at “/”, “-”, and “.”.3
For short monospace statements not to be line broken, we use the “\nbco{}” (non-
breakable co) macro.
D.3.1.6 Limitations
There are a few cases where macros introduced in this section do not work as expected.
Table D.2 lists such limitations.
While \co{} requires some characters to be escaped, it can contain any character.
On the other hand, \tco{} can not handle “%”, “{”, “}”, nor “\” properly. If they are
escaped by a “\”, they appear in the end result with the escape character. The “\verb”
command can be used in running text if you need to use monospace font for a string
which contains many characters to escape.4
3 Overfill can be a problem if the URL or the path name contains long runs of unbreakable
characters.
4 The \verb command is not almighty though. For example, you can’t use it within a
footnote. If you do so, you will see a fatal LATEX error. A workaround would be a macro
v2023.06.11a
D.3.2 Cross-reference
Cross-references to Chapters, Sections, Listings, etc. have been expressed by combina-
tions of names and bare \ref{} commands in the following way:
1 Chapter~\ref{chp:Introduction},
2 Table~\ref{tab:app:styleguide:Digit-Grouping Style}
This is a traditional way of cross-referencing. However, it is tedious and sometimes

error-prone to put a name manually on every cross-reference. The cleveref package
provides a nicer way of cross-referencing. A few examples follow:
1 \Cref{chp:Introduction},
2 \cref{sec:intro:Parallel Programming Goals},
3 \cref{chp:app:styleguide:Style Guide},
4 \cref{tab:app:styleguide:Digit-Grouping Style}, and
5 \cref{lst:app:styleguide:Source of Code Sample} are
6 examples of cross\-/references.
Above code is typeset as follows:

Chapter 2, Section 2.2, Appendix D, Table D.1, and Listing D.3 are examples
of cross-references.
As you can see, naming of cross-references is automated. Current setting generates
capitalized names for both of \Cref{} and \cref{}, but the former should be used at
the beginning of a sentence.
We are in the middle of conversion to cleveref-style cross-referencing.
Cross-references to line numbers of code snippets can be done in a similar way by
using \Clnref{} and \clnref{} macros, which mimic cleveref. The former puts
“Line” as the name of the reference and the latter “line”.
Please refer to cleveref’s documentation for further info on its cleverness.
D.3.3 Non Breakable Spaces

In LATEX conventions, proper use of non-breakable white spaces is highly recommended.
They can prevent widowing and orphaning of single digit numbers or short variable
names, which would cause the text to be confusing at first glance.
The thin space mentioned earlier to be placed in front of a unit symbol is non
breakable.
Other cases to use a non-breakable space (“~” in LATEX source, often referred to as
“nbsp”) are the following (inexhaustive).
• Reference to a Chapter or a Section:
Please refer to Appendix D.2.
• Calling out CPU number or Thread name:
After they load the pointer, CPUs 1 and 2 will see the stored value.
• Short variable name:
The results will be stored in variables a and b.
named \VerbatimFootnotes provided by the fancyvrb package. Unfortunately, perfbook
can’t employ it due to the interference with the footnotebackref package.
v2023.06.11a
D.3.4 Hyphenation and Dashes

D.3.4.1 Hyphenation in Compound Word
In plain LATEX, compound words such as “high-frequency” can be hyphenated only at
the hyphen. This sometimes results in poor typesetting. For example:
High-frequency radio wave, high-frequency ra-

dio wave, high-frequency radio wave, high-
frequency radio wave, high-frequency radio
wave, high-frequency radio wave.
By using a shortcut “\-/” provided by the “extdash” package, hyphenation in elements

of compound words is enabled in perfbook.5
Example with “\-/”:
High-frequency radio wave, high-frequency ra-

dio wave, high-frequency radio wave, high-fre-
quency radio wave, high-frequency radio wave,
high-frequency radio wave.
D.3.4.2 Non Breakable Hyphen

We want hyphenated compound terms such as “x-coordinate”, “y-coordinate”, etc. not
to be broken at the hyphen following a single letter.
To make a hyphen unbreakable, we can use a short cut “\=/” also provided by the
“extdash” package.
Example without a shortcut:
x-, y-, and z-coordinates; x-, y-, and z-

coordinates; x-, y-, and z-coordinates; x-, y-,
and z-coordinates; x-, y-, and z-coordinates; x-,
y-, and z-coordinates;
Example with “\-/”:
x-, y-, and z-coordinates; x-, y-, and z-coordi-

nates; x-, y-, and z-coordinates; x-, y-, and z-
coordinates; x-, y-, and z-coordinates; x-, y-,
and z-coordinates;
Example with “\=/”:
x-, y-, and z-coordinates; x-, y-, and z-coor-

dinates; x-, y-, and z-coordinates; x-, y-, and
z-coordinates; x-, y-, and z-coordinates; x-, y-,
and z-coordinates;
5 In exchange for enabling the shortcut, we can’t use plain LAT X’s shortcut “\-” to
E
specify hyphenation points. Use pfhyphex.tex to add such exceptions.
v2023.06.11a
Note that “\=/” enables hyphenation in elements of compound words as the same as
“\-/” does.
D.3.4.3 Em Dash
Em dashes are used to indicate parenthetic expression. In perfbook, em dashes are
placed without spaces around it. In LATEX source, an em dash is represented by “---”.
Example (quote from Appendix C.1):
This disparity in speed—more than two orders of magnitude—has resulted in
the multi-megabyte caches found on modern CPUs.
D.3.4.4 En Dash
In LATEX convention, en dashes (–) are used for ranges of (mostly) numbers. Past
revisions of perfbook didn’t follow this rule and used plain dashes (-) for such cases.
Now that \clnrefrange, \crefrange, and their variants, which generate en dashes,
are used for ranges of cross-references, the remaining couple of tens of simple dashes
of other types of ranges have been converted to en dashes for consistency.
Example with a simple dash:
Lines 4-12 in Listing D.4 are the contents of the verbbox environment. The
box is output by the \theverbbox macro on line 16.
Example with an en dash:
Lines 4–12 in Listing D.4 are the contents of the verbbox environment. The
box is output by the \theverbbox macro on line 16.
D.3.4.5 Numerical Minus Sign

Numerical minus signs should be coded as math mode minus signs, namely “$-$”.6
For example,
−30, rather than -30.
D.3.5 Punctuation
D.3.5.1 Ellipsis
In monospace fonts, ellipses can be expressed by series of periods. For example:
Great ... So how do I fix it?
However, in proportional fonts, the series of periods is printed with tight spaces as
follows:
Great ... So how do I fix it?
Standard LATEX defines the \dots macro for this purpose. However, it has a kludge
in the evenness of spaces. The “ellipsis” package redefines the \dots macro to fix the
issue.7 By using \dots, the above example is typeset as the following:
6 This rule assumes that math mode uses the same upright glyph as text mode. Our default
font choice meets the assumption.

7 To be exact, it is the \textellipsis macro that is redefined. The behavior of \dots
macro in math mode is not affected. The “amsmath” package has another definition of \dots.
It is not used in perfbook at the moment.
v2023.06.11a
Great . . . So how do I fix it?

Note that the “xspace” option specified to the “ellipsis” package adjusts the spaces
after ellipses depending on what follows them.
For example:
• He said, “I . . . really don’t remember . . .”
• Sequence A: (one, two, three, . . .)
• Sequence B: (4, 5, . . . , 𝑛)
As you can see, extra space is placed before the comma.
\dots macro can also be used in math mode:
• Sequence C: (1, 2, 3, 5, 8, . . . )
• Sequence D: (10, 12, . . . , 20)
The \ldots macro behaves the same as the \dots macro.
D.3.5.2 Full Stop

LATEX treats a full stop in front of a white space as an end of a sentence and puts a
slightly wider skip by default (double spacing). There is an exception to this rule, i.e.
where the full stop is next to a capital letter, LATEX assumes it represents an abbreviation
and puts a normal skip.
To make LATEX use proper skips, one need to annotate such exceptions. For example,
given the following LATEX source:
\begin{quote}
Lock~1 is owned by CPU~A.
Lock~2 is owned by CPU~B. (Bad.)
Lock~1 is owned by CPU~A\@.

Lock~2 is owned by CPU~B\@. (Good.)
\end{quote}
the output will be as the following:

Lock 1 is owned by CPU A. Lock 2 is owned by CPU B. (Bad.)
Lock 1 is owned by CPU A. Lock 2 is owned by CPU B. (Good.)
On the other hand, where a full stop is following a lower case letter, e.g. as in
“Mr. Smith”, a wider skip will follow in the output unless it is properly hinted. Such
hintings can be done in one of several ways.
Given the following source,
\begin{itemize}[nosep]
\item Mr. Smith (bad)
\item Mr.~Smith (good)
\item Mr.\ Smith (good)
\item Mr.\@ Smith (good)
\end{itemize}
the result will look as follows:

• Mr. Smith (bad)
• Mr. Smith (good)
v2023.06.11a
Table D.3: Refrigeration Power Consumption

Power per watt
Situation 𝑇 (K) 𝐶P waste heat (W)
Dry Ice 195 1.990 0.5

Liquid N2 77 0.356 2.8
Liquid H2 20 0.073 13.7
Liquid He 4 0.0138 72.3
IBM Q 0.015 0.000051 19,500.0
D.3.6 Floating Object Format

D.3.6.1 Ruled Line in Table
They say that tables drawn by using ruled lines of plain LATEX look ugly.8 Vertical lines
should be avoided and horizontal lines should be used sparingly, especially in tables of
simple structure.
Table D.3 (corresponding to a table from a now-deleted section) is drawn by using
the features of “booktabs” and “xcolor” packages. Note that ruled lines of booktabs can
not be mixed with vertical lines in a table.9
D.3.6.2 Position of Caption

In LATEX conventions, captions of tables are usually placed above them. The reason
is the flow of your eye movement when you look at them. Most tables have a row of
heading at the top. You naturally look at the top of a table at first. Captions at the
bottom of tables disturb this flow. The same can be said of code snippets, which are
read from top to bottom.
For code snippets, the “ruled” style chosen for listing environment places the caption
at the top. See Listing D.2 for an example.
As for tables, the position of caption is tweaked by \floatstyle{} and
\restylefloat{} macros in preamble.
Vertical skips below captions are reduced by setting a smaller value to the
\abovecaptionskip variable, which would also affect captions to figures.
In the tables which use horizontal rules of “booktabs” package, the vertical skips
between captions and tables are further reduced by setting a negative value to the
\abovetopsep variable, which controls the behavior of \toprule.
D.3.7 Improvement Candidates

There are a few areas yet to be attempted in perfbook which would further improve its
appearance. This section lists such candidates.
8 https://www.inf.ethz.ch/personal/markusp/teaching/guides/guide-
tables.pdf
9 There is another package named “arydshln” which provides dashed lines to be used in
tables. A couple of experimental examples are presented in Appendix D.3.7.2.
v2023.06.11a
Figure D.1: Timer Wheel at 1 kHz Figure D.2: Timer Wheel at 100 kHz
Listing D.6: Message-Passing Litmus Test (by subfig)

(a) Not Enforcing Order (b) Enforcing Order
1 C C-MP+o-wmb-o+o-o.litmus 1 C C-MP+o-wmb-o+o-rmb-o.litmus
2 2
3 { 3 {
4 } 4 }
5 5
6 P0(int* x0, int* x1) { 6 P0(int* x0, int* x1) {
7 7
8 WRITE_ONCE(*x0, 2); 8 WRITE_ONCE(*x0, 2);
9 smp_wmb(); 9 smp_wmb();
10 WRITE_ONCE(*x1, 2); 10 WRITE_ONCE(*x1, 2);
11 11
12 } 12 }
13 13
14 P1(int* x0, int* x1) { 14 P1(int* x0, int* x1) {
15 15
16 int r2; 16 int r2;
17 int r3; 17 int r3;
18 18
19 r2 = READ_ONCE(*x1); 19 r2 = READ_ONCE(*x1);
20 r3 = READ_ONCE(*x0); 20 smp_rmb();
21 21 r3 = READ_ONCE(*x0);
22 } 22
23 23 }
24 24
25 exists (1:r2=2 /\ 1:r3=0) 25 exists (1:r2=2 /\ 1:r3=0)
v2023.06.11a
Table D.4: CPU 0 View of Synchronization Mechanisms on 8-Socket System With

Intel Xeon Platinum 8176 CPUs @ 2.10GHz
Ratio
Operation Cost (ns) (cost/clock)
Best-case CAS 7.0 14.6
Best-case lock 15.4 32.3
Blind CAS 7.2 15.2
CAS 18.0 37.7
Blind CAS (off-core) 47.5 99.8
CAS (off-core) 101.9 214.0
Blind CAS (off-socket) 148.8 312.5
CAS (off-socket) 442.9 930.1
Global Comms 195,000,000 409,500,000
D.3.7.1 Grouping Related Figures/Listings

To prevent a pair of closely related figures or listings from being placed in different
pages, it is desirable to group them into a single floating object. The “subfig” package
provides the features to do so.10
Two floating objects can be placed side by side by using \parbox or minipage. For
example, Figures 14.10 and 14.11 can be grouped together by using a pair of minipages
as shown in Figures D.1 and D.2.
By using subfig package, Listings 15.4 and 15.5 can be grouped together as shown in
Listing D.6 with sub-captions (with a minor change of blank line).
Note that they can not be grouped in the same way as Figures D.1 and D.2 because
the “ruled” style prevents their captions from being properly typeset.
The sub-caption can be cited by combining a \cref{} macro and a \subref{}
macro, for example, “Listing D.6 (a)”.
It can also be cited by a \cref{} macro, for example, “Listing D.6b”. Note the
difference in the resulting format. For the citing by a \cref{} to work, you need to
place the \label{} macro of the combined floating object ahead of the definition of
subfloats. Otherwise, the resulting caption number would be off by one from the actual
number.
D.3.7.2 Table Layout Experiment

This section presents some experimental tables using booktabs, xcolors, and arydshln
packages. The corresponding tables in the text have been converted using one of the
format shown here. The source of this section can be regarded as a reference to be
consulted when new tables are added in the text.
In Table D.4 (corresponding to Table 3.1), the “S” column specifiers provided by the
“siunitx” package are used to align numbers.
Table D.5 (corresponding to Table 13.1) is an example of table with a complex header.
In Table D.5, the gap in the mid-rule corresponds to the distinction which had been
represented by double vertical rules before the conversion. The legends in the frame
10 One problem of grouping figures might be the complexity in LATEX source.
v2023.06.11a
Table D.5: Synchronization and Reference Counting
Release
Reference Hazard
Acquisition Locks RCU
Counts Pointers
Locks − CAMR M CA
Reference
A AMR M A
Counts
Hazard
M M M M
Pointers
RCU CA MA CA M CA
Key: A: Atomic counting

C: Check combined with the atomic acquisition
operation
M: Full memory barriers required
MR : Memory barriers required only on release
MA : Memory barriers required on acquire
Table D.6: Cache Coherence Example
CPU Cache Memory
Sequence # CPU # Operation 0 1 2 3 0 8
0 Initial State −/I −/I −/I −/I V V

1 0 Load 0/S −/I −/I −/I V V
2 3 Load 0/S −/I −/I 0/S V V
3 0 Invalidation 8/S −/I −/I 0/S V V
4 2 RMW 8/S −/I 0/E −/I V V
5 2 Store 8/S −/I 0/M −/I I V
6 1 Atomic Inc 8/S 0/M −/I −/I I V
7 1 Writeback 8/S 8/S −/I −/I V V
box appended here explain the abbreviations used in the matrix. Two types of memory
barrier are denoted by subscripts here. The legends and subscripts are not present in
Table 13.1 since they are redundant there.
Table D.6 (corresponding to Table C.1) is a sequence diagram drawn as a table.
Table D.7 is a tweaked version of Table 9.3. Here, the “Category” column in the
original is removed and the categories are indicated in rows of bold-face font just below
the mid-rules. This change makes it easier for \rowcolors{} command of “xcolor”
package to work properly.
Table D.8 is another version which keeps original columns and colors rows only
where a category has multiple rows. This is done by combining \rowcolors{} of
“xcolor” and \cellcolor{} commands of the “colortbl” package (\cellcolor{}
overrides \rowcolors{}).
In Table 9.3, the latter layout without partial row coloring has been chosen for
simplicity.
v2023.06.11a
Table D.7: RCU Publish-Subscribe and Version Maintenance APIs
Primitives Availability Overhead

List traversal
list_for_each_entry_rcu() 2.5.59 Simple instructions (memory barrier on Alpha)
List update
list_add_rcu() 2.5.44 Memory barrier
list_add_tail_rcu() 2.5.44 Memory barrier
list_del_rcu() 2.5.44 Simple instructions
list_replace_rcu() 2.6.9 Memory barrier
list_splice_init_rcu() 2.6.21 Grace-period latency
Hlist traversal
hlist_for_each_entry_rcu() 2.6.8 Simple instructions (memory barrier on Alpha)
Hlist update
hlist_add_after_rcu() 2.6.14 Memory barrier
hlist_add_before_rcu() 2.6.14 Memory barrier
hlist_add_head_rcu() 2.5.64 Memory barrier
hlist_del_rcu() 2.5.64 Simple instructions
hlist_replace_rcu() 2.6.15 Memory barrier
Pointer traversal
rcu_dereference() 2.6.9 Simple instructions (memory barrier on Alpha)
Pointer update
rcu_assign_pointer() 2.6.10 Memory barrier
Table D.8: RCU Publish-Subscribe and Version Maintenance APIs
Category Primitives Availability Overhead
List traversal list_for_each_entry_rcu() 2.5.59 Simple instructions (mem-

ory barrier on Alpha)
List update list_add_rcu() 2.5.44 Memory barrier

list_add_tail_rcu() 2.5.44 Memory barrier
list_del_rcu() 2.5.44 Simple instructions
list_replace_rcu() 2.6.9 Memory barrier
list_splice_init_rcu() 2.6.21 Grace-period latency
Hlist traversal hlist_for_each_entry_rcu() 2.6.8 Simple instructions (mem-

Hlist update hlist_add_after_rcu() 2.6.14 Memory barrier

hlist_add_before_rcu() 2.6.14 Memory barrier
hlist_add_head_rcu() 2.5.64 Memory barrier
hlist_del_rcu() 2.5.64 Simple instructions
hlist_replace_rcu() 2.6.15 Memory barrier
Pointer traversal rcu_dereference() 2.6.9 Simple instructions (mem-

Pointer update rcu_assign_pointer() 2.6.10 Memory barrier
v2023.06.11a
Table D.9: Memory Misordering: Store-Buffering Sequence of Events
CPU 0 CPU 1
2 x0 = 2; x0==2 x1==0 x1 = 2; x1==2 x0==0
3 r2 = x1; (0) x0==2 x1==0 r2 = x0; (0) x1==2 x0==0

Power per watt
Dry Ice 195 1.990 0.5
Liquid N2 77 0.356 2.8
Liquid H2 20 0.073 13.7
Liquid He 4 0.0138 72.3
IBM Q 0.015 0.000051 19,500.0

Power per watt
Dry Ice 195 1.990 0.5

Liquid N2 77 0.356 2.8
Liquid H2 20 0.073 13.7
Liquid He 4 0.0138 72.3
IBM Q 0.015 0.000051 19,500.0
Table D.9 (corresponding to Table 15.1) is also a sequence diagram drawn as a tabular
object.
Table D.10 shows another version of Table D.3 with dashed horizontal and vertical
rules of the arydshln package.
In this case, the vertical dashed rules seems unnecessary. The one without the vertical
rules is shown in Table D.11.
D.3.7.3 Miscellaneous Candidates

Other improvement candidates are listed in the source of this section as comments.
v2023.06.11a
The Answer to the Ultimate Question of Life, The
Universe, and Everything.
The Hitchhikers Guide to the Galaxy, Douglas Adams
Appendix E
Answers to Quick Quizzes
E.1 How To Use This Book
Quick Quiz 1.1: p.3

Where are the answers to the Quick Quizzes found?
Answer:
In Appendix E starting on page 705. Hey, I thought I owed you an easy one! ❑
Quick Quiz 1.2: p.3

Some of the Quick Quiz questions seem to be from the viewpoint of the reader rather
than the author. Is that really the intent?
Answer:
Indeed it is! Many are questions that Paul E. McKenney would probably have asked if
he was a novice student in a class covering this material. It is worth noting that Paul was
taught most of this material by parallel hardware and software, not by professors. In
Paul’s experience, professors are much more likely to provide answers to verbal questions
than are parallel systems, recent advances in voice-activated assistants notwithstanding.
Of course, we could have a lengthy debate over which of professors or parallel systems
provide the most useful answers to these sorts of questions, but for the time being
let’s just agree that usefulness of answers varies widely across the population both of
professors and of parallel systems.
Other quizzes are quite similar to actual questions that have been asked during
conference presentations and lectures covering the material in this book. A few others
are from the viewpoint of the author. ❑
Quick Quiz 1.3: p.3

These Quick Quizzes are just not my cup of tea. What can I do about it?
Answer:
Here are a few possible strategies:
705
v2023.06.11a
706 APPENDIX E. ANSWERS TO QUICK QUIZZES
1. Just ignore the Quick Quizzes and read the rest of the book. You might miss out
on the interesting material in some of the Quick Quizzes, but the rest of the book
has lots of good material as well. This is an eminently reasonable approach if your
main goal is to gain a general understanding of the material or if you are skimming
through the book to find a solution to a specific problem.
2. Look at the answer immediately rather than investing a large amount of time in
coming up with your own answer. This approach is reasonable when a given
Quick Quiz’s answer holds the key to a specific problem you are trying to solve.
This approach is also reasonable if you want a somewhat deeper understanding
of the material, but when you do not expect to be called upon to generate parallel
solutions given only a blank sheet of paper.
3. If you find the Quick Quizzes distracting but impossible to ignore, you can always
clone the LATEX source for this book from the git archive. You can then run the
command make nq, which will produce a perfbook-nq.pdf. This PDF contains
unobtrusive boxed tags where the Quick Quizzes would otherwise be, and gathers
each chapter’s Quick Quizzes at the end of that chapter in the classic textbook style.
4. Learn to like (or at least tolerate) the Quick Quizzes. Experience indicates that
quizzing yourself periodically while reading greatly increases comprehension and
depth of understanding.
Note that the quick quizzes are hyperlinked to the answers and vice versa. Click
either the “Quick Quiz” heading or the small black square to move to the beginning of
the answer. From the answer, click on the heading or the small black square to move to
the beginning of the quiz, or, alternatively, click on the small white square at the end of
the answer to move to the end of the corresponding quiz. ❑
Quick Quiz 1.4: p.3

If passively reading this book doesn’t get me full problem-solving and code-production
capabilities, what on earth is the point???
Answer:
For those preferring analogies, coding concurrent software is similar to playing music
in that there are good uses for many different levels of talent and skill. Not everyone
needs to devote their entire live to becoming a concert pianist. In fact, for every such
virtuoso, there are a great many lesser pianists whose of music is welcomed by their
friends and families. But these lesser pianists are probably doing something else to
support themselves, and so it is with concurrent coding.
One potential benefit of passively reading this book is the ability to read and understand
modern concurrent code. This ability might in turn permit you to:
1. See what the kernel does so that you can check to see if a proposed use case is
valid.
2. Chase down a kernel bug.
3. Use information in the kernel to more easily chase down a userspace bug.
4. Produce a fix for a kernel bug.
v2023.06.11a
E.2. INTRODUCTION 707
5. Create a straightforward kernel feature, whether from scratch or using the modern
copy-pasta development methodology.
If you are proficient with straightforward uses of locks and atomic operations,
passively reading this book should enable you to successfully apply modern concurrency
techniques.
And finally, if your job is to coordinate the activities of developers making use of
modern concurrency techniques, passively reading this book might help you understand
what on earth they are talking about. ❑
E.2 Introduction
Quick Quiz 2.1: p.11

Come on now!!! Parallel programming has been known to be exceedingly hard for
many decades. You seem to be hinting that it is not so hard. What sort of game are
you playing?
Answer:
If you really believe that parallel programming is exceedingly hard, then you should have
a ready answer to the question “Why is parallel programming hard?” One could list any
number of reasons, ranging from deadlocks to race conditions to testing coverage, but
the real answer is that it is not really all that hard. After all, if parallel programming was
really so horribly difficult, how could a large number of open-source projects, ranging
from Apache to MySQL to the Linux kernel, have managed to master it?
A better question might be: “Why is parallel programming perceived to be so
difficult?” To see the answer, let’s go back to the year 1991. Paul McKenney was
walking across the parking lot to Sequent’s benchmarking center carrying six dual-80486
Sequent Symmetry CPU boards, when he suddenly realized that he was carrying several
times the price of the house he had just purchased.1 This high cost of parallel systems
meant that parallel programming was restricted to a privileged few who worked for
an employer who either manufactured or could afford to purchase machines costing
upwards of $100,000—in 1991 dollars US.
In contrast, in 2020, Paul finds himself typing these words on a six-core x86 laptop.
Unlike the dual-80486 CPU boards, this laptop also contains 64 GB of main memory, a
1 TB solid-state disk, a display, Ethernet, USB ports, wireless, and Bluetooth. And the
laptop is more than an order of magnitude cheaper than even one of those dual-80486
CPU boards, even before taking inflation into account.
Parallel systems have truly arrived. They are no longer the sole domain of a privileged
few, but something available to almost everyone.
The earlier restricted availability of parallel hardware is the real reason that parallel
programming is considered so difficult. After all, it is quite difficult to learn to program
even the simplest machine if you have no access to it. Since the age of rare and
expensive parallel machines is for the most part behind us, the age during which parallel
programming is perceived to be mind-crushingly difficult is coming to a close.2 ❑
1 Yes, this sudden realization did cause him to walk quite a bit more carefully. Why do
you ask?
2 Parallel programming is in some ways more difficult than sequential programming, for
example, parallel validation is more difficult. But no longer mind-crushingly difficult.
v2023.06.11a

How could parallel programming ever be as easy as sequential programming?
Answer:
It depends on the programming environment. SQL [Int92] is an underappreciated
success story, as it permits programmers who know nothing about parallelism to keep a
large parallel system productively busy. We can expect more variations on this theme as
parallel computers continue to become cheaper and more readily available. For example,
one possible contender in the scientific and technical computing arena is MATLAB*P,
which is an attempt to automatically parallelize common matrix operations.
Finally, on Linux and UNIX systems, consider the following shell command:
get_input | grep "interesting" | sort
This shell pipeline runs the get_input, grep, and sort processes in parallel. There,
that wasn’t so hard, now was it?
In short, parallel programming is just as easy as sequential programming—at least in
those environments that hide the parallelism from the user! ❑

Oh, really??? What about correctness, maintainability, robustness, and so on?
Answer:
These are important goals, but they are just as important for sequential programs as they
are for parallel programs. Therefore, important though they are, they do not belong on a
list specific to parallel programming. ❑

And if correctness, maintainability, and robustness don’t make the list, why do
productivity and generality?
Answer:
Given that parallel programming is perceived to be much harder than sequential pro-
gramming, productivity is tantamount and therefore must not be omitted. Furthermore,
high-productivity parallel-programming environments such as SQL serve a specific
purpose, hence generality must also be added to the list. ❑

Given that parallel programs are much harder to prove correct than are sequential
programs, again, shouldn’t correctness really be on the list?
Answer:
From an engineering standpoint, the difficulty in proving correctness, either formally or
informally, would be important insofar as it impacts the primary goal of productivity.
So, in cases where correctness proofs are important, they are subsumed under the
“productivity” rubric. ❑
v2023.06.11a

What about just having fun?
Answer:
Having fun is important as well, but, unless you are a hobbyist, would not normally be a
primary goal. On the other hand, if you are a hobbyist, go wild! ❑

Are there no cases where parallel programming is about something other than
performance?
Answer:
There certainly are cases where the problem to be solved is inherently parallel, for
example, Monte Carlo methods and some numerical computations. Even in these cases,
however, there will be some amount of extra work managing the parallelism.
Parallelism is also sometimes used for reliability. For but one example, triple-modulo
redundancy has three systems run in parallel and vote on the result. In extreme cases,
the three systems will be independently implemented using different algorithms and
technologies. ❑

Why not instead rewrite programs from inefficient scripting languages to C or C++?
Answer:
If the developers, budget, and time is available for such a rewrite, and if the result will
attain the required levels of performance on a single CPU, this can be a reasonable
approach. ❑

Why all this prattling on about non-technical issues??? And not just any non-technical
issue, but productivity of all things? Who cares?
Answer:
If you are a pure hobbyist, perhaps you don’t need to care. But even pure hobbyists
will often care about how much they can get done, and how quickly. After all, the
most popular hobbyist tools are usually those that are the best suited for the job, and an
important part of the definition of “best suited” involves productivity. And if someone
is paying you to write parallel code, they will very likely care deeply about your
productivity. And if the person paying you cares about something, you would be most
wise to pay at least some attention to it!
Besides, if you really didn’t care about productivity, you would be doing it by hand
rather than using a computer! ❑
v2023.06.11a

Given how cheap parallel systems have become, how can anyone afford to pay people
to program them?
Answer:
There are a number of answers to this question:
1. Given a large computational cluster of parallel machines, the aggregate cost of the
cluster can easily justify substantial developer effort, because the development cost
can be spread over the large number of machines.
2. Popular software that is run by tens of millions of users can easily justify substantial
developer effort, as the cost of this development can be spread over the tens of
millions of users. Note that this includes things like kernels and system libraries.
3. If the low-cost parallel machine is controlling the operation of a valuable piece of

equipment, then the cost of this piece of equipment might easily justify substantial
developer effort.
4. If the software for the low-cost parallel machine produces an extremely valuable
result (e.g., energy savings), then this valuable result might again justify substantial
developer cost.
5. Safety-critical systems protect lives, which can clearly justify very large developer
effort.
6. Hobbyists and researchers might instead seek knowledge, experience, fun, or glory.
So it is not the case that the decreasing cost of hardware renders software worthless, but
rather that it is no longer possible to “hide” the cost of software development within the
cost of the hardware, at least not unless there are extremely large quantities of hardware.
❑

This is a ridiculously unachievable ideal! Why not focus on something that is
achievable in practice?
Answer:
This is eminently achievable. The cellphone is a computer that can be used to make
phone calls and to send and receive text messages with little or no programming or
configuration on the part of the end user.
This might seem to be a trivial example at first glance, but if you consider it carefully
you will see that it is both simple and profound. When we are willing to sacrifice
generality, we can achieve truly astounding increases in productivity. Those who indulge
in excessive generality will therefore fail to set the productivity bar high enough to
succeed near the top of the software stack. This fact of life even has its own acronym:
YAGNI, or “You Ain’t Gonna Need It.” ❑
v2023.06.11a

Wait a minute! Doesn’t this approach simply shift the development effort from you to
whoever wrote the existing parallel software you are using?
Answer:
Exactly! And that is the whole point of using existing software. One team’s work can
be used by many other teams, resulting in a large decrease in overall effort compared to
all teams needlessly reinventing the wheel. ❑

What other bottlenecks might prevent additional CPUs from providing additional
performance?
Answer:
There are any number of potential bottlenecks:
1. Main memory. If a single thread consumes all available memory, additional threads
will simply page themselves silly.
2. Cache. If a single thread’s cache footprint completely fills any shared CPU cache(s),
then adding more threads will simply thrash those affected caches, as will be seen
in Chapter 10.
3. Memory bandwidth. If a single thread consumes all available memory band-

width, additional threads will simply result in additional queuing on the system
interconnect.
4. I/O bandwidth. If a single thread is I/O bound, adding more threads will simply
result in them all waiting in line for the affected I/O resource.
Specific hardware systems might have any number of additional bottlenecks. The fact
is that every resource which is shared between multiple CPUs or threads is a potential
bottleneck. ❑

Other than CPU cache capacity, what might require limiting the number of concurrent
threads?
Answer:
There are any number of potential limits on the number of threads:
1. Main memory. Each thread consumes some memory (for its stack if nothing else),
so that excessive numbers of threads can exhaust memory, resulting in excessive
paging or memory-allocation failures.
2. I/O bandwidth. If each thread initiates a given amount of mass-storage I/O or

networking traffic, excessive numbers of threads can result in excessive I/O queuing
delays, again degrading performance. Some networking protocols may be subject
to timeouts or other failures if there are so many threads that networking events
cannot be responded to in a timely fashion.
v2023.06.11a
3. Synchronization overhead. For many synchronization protocols, excessive numbers

of threads can result in excessive spinning, blocking, or rollbacks, thus degrading
performance.
Specific applications and platforms may have any number of additional limiting
factors. ❑

Just what is “explicit timing”???
Answer:
Where each thread is given access to some set of resources during an agreed-to slot
of time. For example, a parallel program with eight threads might be organized into
eight-millisecond time intervals, so that the first thread is given access during the first
millisecond of each interval, the second thread during the second millisecond, and so
on. This approach clearly requires carefully synchronized clocks and careful control of
execution times, and therefore should be used with considerable caution.
In fact, outside of hard realtime environments, you almost certainly want to use
something else instead. Explicit timing is nevertheless worth a mention, as it is always
there when you need it. ❑

Are there any other obstacles to parallel programming?
Answer:
There are a great many other potential obstacles to parallel programming. Here are a
few of them:
1. The only known algorithms for a given project might be inherently sequential in
nature. In this case, either avoid parallel programming (there being no law saying
that your project has to run in parallel) or invent a new parallel algorithm.
2. The project allows binary-only plugins that share the same address space, such
that no one developer has access to all of the source code for the project. Because
many parallel bugs, including deadlocks, are global in nature, such binary-only
plugins pose a severe challenge to current software development methodologies.
This might well change, but for the time being, all developers of parallel code
sharing a given address space need to be able to see all of the code running in that
address space.
3. The project contains heavily used APIs that were designed without regard to
parallelism [AGH+ 11a, CKZ+ 13]. Some of the more ornate features of the System
V message-queue API form a case in point. Of course, if your project has been
around for a few decades, and its developers did not have access to parallel hardware,
it undoubtedly has at least its share of such APIs.
4. The project was implemented without regard to parallelism. Given that there are a
great many techniques that work extremely well in a sequential environment, but
that fail miserably in parallel environments, if your project ran only on sequential
hardware for most of its lifetime, then your project undoubtably has at least its
share of parallel-unfriendly code.
v2023.06.11a
E.3. HARDWARE AND ITS HABITS 713
5. The project was implemented without regard to good software-development practice.

The cruel truth is that shared-memory parallel environments are often much less
forgiving of sloppy development practices than are sequential environments. You
may be well-served to clean up the existing design and code prior to attempting
parallelization.
6. The people who originally did the development on your project have since moved
on, and the people remaining, while well able to maintain it or add small features,
are unable to make “big animal” changes. In this case, unless you can work out a
very simple way to parallelize your project, you will probably be best off leaving
it sequential. That said, there are a number of simple approaches that you might
use to parallelize your project, including running multiple instances of it, using a
parallel implementation of some heavily used library function, or making use of
some other parallel project, such as a database.
One can argue that many of these obstacles are non-technical in nature, but that does
not make them any less real. In short, parallelization of a large body of code can be a
large and complex effort. As with any large and complex effort, it makes sense to do
your homework beforehand. ❑
E.3 Hardware and its Habits

Why should parallel programmers bother learning low-level properties of the hardware?
Wouldn’t it be easier, better, and more elegant to remain at a higher level of abstraction?
Answer:
It might well be easier to ignore the detailed properties of the hardware, but in most cases
it would be quite foolish to do so. If you accept that the only purpose of parallelism is
to increase performance, and if you further accept that performance depends on detailed
properties of the hardware, then it logically follows that parallel programmers are going
to need to know at least a few hardware properties.
This is the case in most engineering disciplines. Would you want to use a bridge
designed by an engineer who did not understand the properties of the concrete and
steel making up that bridge? If not, why would you expect a parallel programmer to be
able to develop competent parallel software without at least some understanding of the
underlying hardware? ❑

What types of machines would allow atomic operations on multiple data elements?
Answer:
One answer to this question is that it is often possible to pack multiple elements of data
into a single machine word, which can then be manipulated atomically.
A more trendy answer would be machines supporting transactional memory [Lom77,
Kni86, HM93]. By early 2014, several mainstream systems provided limited hardware
transactional memory implementations, which is covered in more detail in Section 17.3.
v2023.06.11a
The jury is still out on the applicability of software transactional memory [MMW07,
PW07, RHP+ 07, CBM+ 08, DFGG11, MS12], which is covered in Section 17.2. ❑

So have CPU designers also greatly reduced the overhead of cache misses?
Answer:
Unfortunately, not so much. There has been some reduction given constant numbers of
CPUs, but the finite speed of light and the atomic nature of matter limits their ability to
reduce cache-miss overhead for larger systems. Section 3.3 discusses some possible
avenues for possible future progress. ❑

This is a simplified sequence of events? How could it possibly be any more complex?
Answer:
This sequence ignored a number of possible complications, including:
1. Other CPUs might be concurrently attempting to perform memory-reference

operations involving this same cacheline.
2. The cacheline might have been replicated read-only in several CPUs’ caches, in
which case, it would need to be flushed from their caches.
3. CPU 7 might have been operating on the cache line when the request for it arrived,
in which case CPU 7 might need to hold off the request until its own operation
completed.
4. CPU 7 might have ejected the cacheline from its cache (for example, in order to
make room for other data), so that by the time that the request arrived, the cacheline
was on its way to memory.
5. A correctable error might have occurred in the cacheline, which would then need
to be corrected at some point before the data was used.
Production-quality cache-coherence mechanisms are extremely complicated due to

these sorts of considerations [HP95, CSG99, MHS12, SHW11]. ❑

Why is it necessary to flush the cacheline from CPU 7’s cache?
Answer:
If the cacheline was not flushed from CPU 7’s cache, then CPUs 0 and 7 might have
different values for the same set of variables in the cacheline. This sort of incoherence
greatly complicates parallel software, which is why wise hardware architects avoid it. ❑
v2023.06.11a

Table 3.1 shows CPU 0 sharing a core with CPU 224. Shouldn’t that instead be
CPU 1???
Answer:
It is easy to be sympathetic to this view, but the file /sys/devices/system/cpu/
cpu0/cache/index0/shared_cpu_list really does contain the string 0,224. There-
fore, CPU 0’s hyperthread twin really is CPU 224. Some people speculate that this
numbering allows naive applications and schedulers to perform better, citing the fact
that on many workloads the second hyperthread does not provide a huge amount of
additional performance. This speculation assumes that naive applications and schedulers
would utilize CPUs in numerical order, leaving aside the weaker hyperthread twin CPUs
until all cores are in use. ❑

Surely the hardware designers could be persuaded to improve this situation! Why
have they been content with such abysmal performance for these single-instruction
operations?
Answer:
The hardware designers have been working on this problem, and have consulted with
no less a luminary than the late physicist Stephen Hawking. Hawking’s observation was
that the hardware designers have two basic problems [Gar07]:
1. The finite speed of light, and
2. The atomic nature of matter.
The first problem limits raw speed, and the second limits miniaturization, which
in turn limits frequency. And even this sidesteps the power-consumption issue that is
currently limiting production frequencies to well below 10 GHz.
In addition, Table 3.1 on page 36 represents a reasonably large system with no fewer
than 448 hardware threads. Smaller systems often achieve better latency, as may be seen
in Table E.1, which represents a much smaller system with only 16 hardware threads.
A similar view is provided by the rows of Table 3.1 down to and including the two
“Off-Core” rows.
Furthermore, newer small-scale single-socket systems such as the laptop on which I
am typing this also have more reasonable latencies, as can be seen in Table E.2.
Alternatively, a 64-CPU system in the mid 1990s had cross-interconnect latencies
in excess of five microseconds, so even the eight-socket 448-hardware-thread monster
shown in Table 3.1 represents more than a five-fold improvement over its 25-years-prior
counterparts.
Integration of hardware threads in a single core and multiple cores on a die have
improved latencies greatly, at least within the confines of a single core or single die.
There has been some improvement in overall system latency, but only by about a factor
of two. Unfortunately, neither the speed of light nor the atomic nature of matter has
changed much in the past few years [Har16]. Therefore, spatial and temporal locality
are first-class concerns for concurrent software, even when running on relatively small
systems.
Section 3.3 looks at what else hardware designers might be able to do to ease the
plight of parallel programmers. ❑
v2023.06.11a
Table E.1: Performance of Synchronization Mechanisms on 16-CPU 2.8 GHz Intel

X5550 (Nehalem) System
Ratio
Operation Cost (ns) (cost/clock)
Same-CPU
CAS 12.2 33.8
lock 25.6 71.2
On-Core
Blind CAS 12.9 35.8
CAS 7.0 19.4
Off-Core
Blind CAS 31.2 86.6
CAS 31.2 86.5
Off-Socket
Blind CAS 92.4 256.7
CAS 95.9 266.4
Off-System
Global Comms 195,000,000 542,000,000
Table E.2: CPU 0 View of Synchronization Mechanisms on 12-CPU Intel Core

i7-8750H CPU @ 2.20 GHz
Ratio
Operation Cost (ns) (cost/clock) CPUs
Same-CPU 0
CAS 6.2 13.6
lock 13.5 29.6
On-Core 6
Blind CAS 6.5 14.3
CAS 16.2 35.6
Off-Core 1–5
Blind CAS 22.2 48.8 7–11
CAS 53.6 117.9
Off-System
Global Comms 195,000,000 429,000,000
v2023.06.11a

Table E.1 in the answer to Quick Quiz 3.7 on page 716 says that on-core CAS is faster
than both of same-CPU CAS and on-core blind CAS. What is happening there?
Answer:
I was surprised by the data I obtained and did a rigorous check of their validity. I got the
same result persistently. One theory that might explain the observation would be: The
two threads in the core are able to overlap their accesses, while the single CPU must
do everything sequentially. Unfortunately, there seems to be no public documentation
explaining why the Intel X5550 (Nehalem) system behaved like that. ❑

These numbers are insanely large! How can I possibly get my head around them?
Answer:
Get a roll of toilet paper. In the USA, each roll will normally have somewhere around
350–500 sheets. Tear off one sheet to represent a single clock cycle, setting it aside.
Now unroll the rest of the roll.
The resulting pile of toilet paper will likely represent a single CAS cache miss.
For the more-expensive inter-system communications latencies, use several rolls (or
multiple cases) of toilet paper to represent the communications latency.
Important safety tip: Make sure to account for the needs of those you live with when
appropriating toilet paper, especially in 2020 or during a similar time when store shelves
are free of toilet paper and much else besides.
Furthermore, for those working on kernel code, a CPU disabling interrupts across a
cache miss is analogous to you holding your breath while unrolling a roll of toilet paper.
How many rolls of toilet paper can you unroll while holding your breath? You might
wish to avoid disabling interrupts across that many cache misses.3 ❑

But individual electrons don’t move anywhere near that fast, even in conductors!!!
The electron drift velocity in a conductor under semiconductor voltage levels is on the
order of only one millimeter per second. What gives???
Answer:
Electron drift velocity tracks the long-term movement of individual electrons. It turns
out that individual electrons bounce around quite randomly, so that their instantaneous
speed is very high, but over the long term, they don’t move very far. In this, electrons
resemble long-distance commuters, who might spend most of their time traveling at full
highway speed, but over the long term go nowhere. These commuters’ speed might be
70 miles per hour (113 kilometers per hour), but their long-term drift velocity relative
to the planet’s surface is zero.
Therefore, we should pay attention not to the electrons’ drift velocity, but to their
instantaneous velocities. However, even their instantaneous velocities are nowhere
near a significant fraction of the speed of light. Nevertheless, the measured velocity of
electric waves in conductors is a substantial fraction of the speed of light, so we still
have a mystery on our hands.
3 Kudos to Matthew Wilcox for this holding-breath analogy.
v2023.06.11a
The other trick is that electrons interact with each other at significant distances (from
an atomic perspective, anyway), courtesy of their negative charge. This interaction is
carried out by photons, which do move at the speed of light. So even with electricity’s
electrons, it is photons doing most of the fast footwork.
Extending the commuter analogy, a driver might use a smartphone to inform other
drivers of an accident or congestion, thus allowing a change in traffic flow to propagate
much faster than the instantaneous velocity of the individual cars. Summarizing the
analogy between electricity and traffic flow:
1. The (very low) drift velocity of an electron is similar to the long-term velocity of a
commuter, both being very nearly zero.
2. The (still rather low) instantaneous velocity of an electron is similar to the

instantaneous velocity of a car in traffic. Both are much higher than the drift
velocity, but quite small compared to the rate at which changes propagate.
3. The (much higher) propagation velocity of an electric wave is primarily due to

photons transmitting electromagnetic force among the electrons. Similarly, traffic
patterns can change quite quickly due to communication among drivers. Not that
this is necessarily of much help to the drivers already stuck in traffic, any more
than it is to the electrons already pooled in a given capacitor.
Of course, to fully understand this topic, you should read up on electrodynamics. ❑

Given that distributed-systems communication is so horribly expensive, why does
anyone bother with such systems?
Answer:
There are a number of reasons:
1. Shared-memory multiprocessor systems have strict size limits. If you need more
than a few thousand CPUs, you have no choice but to use a distributed system.
2. Large shared-memory systems tend to be more expensive per unit computation

than their smaller counterparts.
3. Large shared-memory systems tend to have much longer cache-miss latencies than
do smaller system. To see this, compare Table 3.1 on page 36 with Table E.2.
4. The distributed-systems communications operations do not necessarily use much

CPU, so that computation can proceed in parallel with message transfer.
5. Many important problems are “embarrassingly parallel”, so that extremely large

quantities of processing may be enabled by a very small number of messages.
SETI@HOME [Uni08b] was but one example of such an application. These sorts
of applications can make good use of networks of computers despite extremely
long communications latencies.
v2023.06.11a
E.4. TOOLS OF THE TRADE 719
Thus, large shared-memory systems tend to be used for applications that benefit from
faster latencies than can be provided by distributed computing, and particularly for those
applications that benefit from a large shared memory.
It is likely that continued work on parallel applications will increase the number
of embarrassingly parallel applications that can run well on machines and/or clusters
having long communications latencies, reductions in cost being the driving force that
it is. That said, greatly reduced hardware latencies would be an extremely welcome
development, both for single-system and for distributed computing. ❑

OK, if we are going to have to apply distributed-programming techniques to shared-
memory parallel programs, why not just always use these distributed techniques and
dispense with shared memory?
Answer:
Because it is often the case that only a small fraction of the program is performance-
critical. Shared-memory parallelism allows us to focus distributed-programming
techniques on that small fraction, allowing simpler shared-memory techniques to be
used on the non-performance-critical bulk of the program. ❑
E.4 Tools of the Trade

You call these tools??? They look more like low-level synchronization primitives to
me!
Answer:
They look that way because they are in fact low-level synchronization primitives. And
they are in fact the fundamental tools for building low-level concurrent software. ❑

But this silly shell script isn’t a real parallel program! Why bother with such trivia???
Answer:
Because you should never forget the simple stuff!
Please keep in mind that the title of this book is “Is Parallel Programming Hard,
And, If So, What Can You Do About It?”. One of the most effective things you can do
about it is to avoid forgetting the simple stuff! After all, if you choose to do parallel
programming the hard way, you have no one but yourself to blame. ❑
v2023.06.11a

Is there a simpler way to create a parallel shell script? If so, how? If not, why not?
Answer:
One straightforward approach is the shell pipeline:
grep $pattern1 | sed -e 's/a/b/' | sort
For a sufficiently large input file, grep will pattern-match in parallel with sed editing
and with the input processing of sort. See the file parallel.sh for a demonstration
of shell-script parallelism and pipelining. ❑

But if script-based parallel programming is so easy, why bother with anything else?
Answer:
In fact, it is quite likely that a very large fraction of parallel programs in use today are
script-based. However, script-based parallelism does have its limitations:
1. Creation of new processes is usually quite heavyweight, involving the expensive

fork() and exec() system calls.
2. Sharing of data, including pipelining, typically involves expensive file I/O.
3. The reliable synchronization primitives available to scripts also typically involve

expensive file I/O.
4. Scripting languages are often too slow, but are often quite useful when coordinating
execution of long-running programs written in lower-level programming languages.
These limitations require that script-based parallelism use coarse-grained parallelism,

with each unit of work having execution time of at least tens of milliseconds, and
preferably much longer.
Those requiring finer-grained parallelism are well advised to think hard about their
problem to see if it can be expressed in a coarse-grained form. If not, they should
consider using other parallel-programming environments, such as those discussed in
Section 4.2. ❑

Why does this wait() primitive need to be so complicated? Why not just make it
work like the shell-script wait does?
Answer:
Some parallel applications need to take special action when specific children exit, and
therefore need to wait for each child individually. In addition, some parallel applications
need to detect the reason that the child died. As we saw in Listing 4.2, it is not hard to
build a waitall() function out of the wait() function, but it would be impossible to
do the reverse. Once the information about a specific child is lost, it is lost. ❑
v2023.06.11a

Isn’t there a lot more to fork() and wait() than discussed here?
Answer:
Indeed there is, and it is quite possible that this section will be expanded in future
versions to include messaging features (such as UNIX pipes, TCP/IP, and shared file
I/O) and memory mapping (such as mmap() and shmget()). In the meantime, there
are any number of textbooks that cover these primitives in great detail, and the truly
motivated can read manpages, existing parallel applications using these primitives, as
well as the source code of the Linux-kernel implementations themselves.
It is important to note that the parent process in Listing 4.3 waits until after the child
terminates to do its printf(). Using printf()’s buffered I/O concurrently to the
same file from multiple processes is non-trivial, and is best avoided. If you really need
to do concurrent buffered I/O, consult the documentation for your OS. For UNIX/Linux
systems, Stewart Weiss’s lecture notes provide a good introduction with informative
examples [Wei13]. ❑

If the mythread() function in Listing 4.4 can simply return, why bother with
pthread_exit()?
Answer:
In this simple example, there is no reason whatsoever. However, imagine a more
complex example, where mythread() invokes other functions, possibly separately
compiled. In such a case, pthread_exit() allows these other functions to end the
thread’s execution without having to pass some sort of error return all the way back up
to mythread(). ❑

If the C language makes no guarantees in presence of a data race, then why does the
Linux kernel have so many data races? Are you trying to tell me that the Linux kernel
is completely broken???
Answer:
Ah, but the Linux kernel is written in a carefully selected superset of the C language
that includes special GNU extensions, such as asms, that permit safe execution even
in presence of data races. In addition, the Linux kernel does not run on a number of
platforms where data races would be especially problematic. For an example, consider
embedded systems with 32-bit pointers and 16-bit busses. On such a system, a data
race involving a store to and a load from a given pointer might well result in the load
returning the low-order 16 bits of the old value of the pointer concatenated with the
high-order 16 bits of the new value of the pointer.
Nevertheless, even in the Linux kernel, data races can be quite dangerous and should
be avoided where feasible [Cor12]. ❑
v2023.06.11a

What if I want several threads to hold the same lock at the same time?
Answer:
The first thing you should do is to ask yourself why you would want to do such a thing.
If the answer is “because I have a lot of data that is read by many threads, and only
occasionally updated”, then POSIX reader-writer locks might be what you are looking
for. These are introduced in Section 4.2.4.
Another way to get the effect of multiple threads holding the same lock is for one
thread to acquire the lock, and then use pthread_create() to create the other threads.
The question of why this would ever be a good idea is left to the reader. ❑

Why not simply make the argument to lock_reader() on line 6 of Listing 4.5 be a
pointer to a pthread_mutex_t?
Answer:
Because we will need to pass lock_reader() to pthread_create(). Although we
could cast the function when passing it to pthread_create(), function casts are quite
a bit uglier and harder to get right than are simple pointer casts. ❑

What is the READ_ONCE() on lines 20 and 47 and the WRITE_ONCE() on line 47 of
Listing 4.5?
Answer:
These macros constrain the compiler so as to prevent it from carrying out optimizations
that would be problematic for concurrently accessed shared variables. They don’t
constrain the CPU at all, other than by preventing reordering of accesses to a given
single variable. Note that this single-variable constraint does apply to the code shown
in Listing 4.5 because only the variable x is accessed.
For more information on READ_ONCE() and WRITE_ONCE(), please see Section 4.2.5.
For more information on ordering accesses to multiple variables by multiple threads,
please see Chapter 15. In the meantime, READ_ONCE(x) has much in common with the
GCC intrinsic __atomic_load_n(&x, __ATOMIC_RELAXED) and WRITE_ONCE(x,
v) has much in common with the GCC intrinsic __atomic_store_n(&x, v, __
ATOMIC_RELAXED). ❑

Writing four lines of code for each acquisition and release of a pthread_mutex_t
sure seems painful! Isn’t there a better way?
Answer:
Indeed! And for that reason, the pthread_mutex_lock() and pthread_mutex_
unlock() primitives are normally wrapped in functions that do this error checking.
Later on, we will wrap them with the Linux kernel spin_lock() and spin_unlock()
APIs. ❑
v2023.06.11a

Is “x = 0” the only possible output from the code fragment shown in Listing 4.6? If
so, why? If not, what other output could appear, and why?
Answer:
No. The reason that “x = 0” was output was that lock_reader() acquired the lock
first. Had lock_writer() instead acquired the lock first, then the output would have
been “x = 3”. However, because the code fragment started lock_reader() first and
because this run was performed on a multiprocessor, one would normally expect lock_
reader() to acquire the lock first. Nevertheless, there are no guarantees, especially on
a busy system. ❑

Using different locks could cause quite a bit of confusion, what with threads seeing
each others’ intermediate states. So should well-written parallel programs restrict
themselves to using a single lock in order to avoid this kind of confusion?
Answer:
Although it is sometimes possible to write a program using a single global lock that both
performs and scales well, such programs are exceptions to the rule. You will normally
need to use multiple locks to attain good performance and scalability.
One possible exception to this rule is “transactional memory”, which is currently a
research topic. Transactional-memory semantics can be loosely thought of as those of a
single global lock with optimizations permitted and with the addition of rollback [Boe09].
❑

In the code shown in Listing 4.7, is lock_reader() guaranteed to see all the values
produced by lock_writer()? Why or why not?
Answer:
No. On a busy system, lock_reader() might be preempted for the entire duration of
lock_writer()’s execution, in which case it would not see any of lock_writer()’s
intermediate states for x. ❑

Wait a minute here!!! Listing 4.6 didn’t initialize shared variable x, so why does it
need to be initialized in Listing 4.7?
Answer:
See line 4 of Listing 4.5. Because the code in Listing 4.6 ran first, it could rely on
the compile-time initialization of x. The code in Listing 4.7 ran next, so it had to
re-initialize x. ❑
v2023.06.11a

Instead of using READ_ONCE() everywhere, why not just declare goflag as volatile
on line 10 of Listing 4.8?
Answer:
A volatile declaration is in fact a reasonable alternative in this particular case.
However, use of READ_ONCE() has the benefit of clearly flagging to the reader that
goflag is subject to concurrent reads and updates. Note that READ_ONCE() is especially
useful in cases where most of the accesses are protected by a lock (and thus not subject
to change), but where a few of the accesses are made outside of the lock. Using a
volatile declaration in this case would make it harder for the reader to note the special
accesses outside of the lock, and would also make it harder for the compiler to generate
good code under the lock. ❑

READ_ONCE() only affects the compiler, not the CPU. Don’t we also need memory
barriers to make sure that the change in goflag’s value propagates to the CPU in a
timely fashion in Listing 4.8?
Answer:
No, memory barriers are not needed and won’t help here. Memory barriers only enforce
ordering among multiple memory references: They absolutely do not guarantee to
expedite the propagation of data from one part of the system to another.4 This leads to a
quick rule of thumb: You do not need memory barriers unless you are using more than
one variable to communicate between multiple threads.
But what about nreadersrunning? Isn’t that a second variable used for communi-
cation? Indeed it is, and there really are the needed memory-barrier instructions buried
in __sync_fetch_and_add(), which make sure that the thread proclaims its presence
before checking to see if it should start. ❑

Would it ever be necessary to use READ_ONCE() when accessing a per-thread variable,
for example, a variable declared using GCC’s __thread storage class?
Answer:
It depends. If the per-thread variable was accessed only from its thread, and never from
a signal handler, then no. Otherwise, it is quite possible that READ_ONCE() is needed.
We will see examples of both situations in Section 5.4.4.
This leads to the question of how one thread can gain access to another thread’s
__thread variable, and the answer is that the second thread must store a pointer to
its __thread variable somewhere that the first thread has access to. One common
approach is to maintain a linked list with one element per thread, and to store the address
of each thread’s __thread variable in the corresponding element. ❑
4 There have been persistent rumors of hardware in which memory barriers actually do
expedite propagation of data, but no confirmed sightings.
v2023.06.11a

Isn’t comparing against single-CPU throughput a bit harsh?
Answer:
Not at all. In fact, this comparison was, if anything, overly lenient. A more balanced
comparison would be against single-CPU throughput with the locking primitives
commented out. ❑

But one microsecond is not a particularly small size for a critical section. What do I
do if I need a much smaller critical section, for example, one containing only a few
instructions?
Answer:
If the data being read never changes, then you do not need to hold any locks while
accessing it. If the data changes sufficiently infrequently, you might be able to checkpoint
execution, terminate all threads, change the data, then restart at the checkpoint.
Another approach is to keep a single exclusive lock per thread, so that a thread
read-acquires the larger aggregate reader-writer lock by acquiring its own lock, and
write-acquires by acquiring all the per-thread locks [HW92]. This can work quite well
for readers, but causes writers to incur increasingly large overheads as the number of
threads increases.
Some other ways of efficiently handling very small critical sections are described in
Chapter 9. ❑

The system used is a few years old, and new hardware should be faster. So why should
anyone worry about reader-writer locks being slow?
Answer:
In general, newer hardware is improving. However, it will need to improve several
orders of magnitude to permit reader-writer lock to achieve ideal performance on 448
CPUs. Worse yet, the greater the number of CPUs, the larger the required performance
improvement. The performance problems of reader-writer locking are therefore very
likely to be with us for quite some time to come. ❑

Is it really necessary to have both sets of primitives?
Answer:
Strictly speaking, no. One could implement any member of the second set using the
corresponding member of the first set. For example, one could implement __sync_
nand_and_fetch() in terms of __sync_fetch_and_nand() as follows:
tmp = v;
ret = __sync_fetch_and_nand(p, tmp);
ret = ~ret & tmp;
v2023.06.11a
It is similarly possible to implement __sync_fetch_and_add(), __sync_fetch_

and_sub(), and __sync_fetch_and_xor() in terms of their post-value counterparts.
However, the alternative forms can be quite convenient, both for the programmer and
for the compiler/library implementor. ❑

Given that these atomic operations will often be able to generate single atomic
instructions that are directly supported by the underlying instruction set, shouldn’t
they be the fastest possible way to get things done?
Answer:
Unfortunately, no. See Chapter 5 for some stark counterexamples. ❑

What happened to ACCESS_ONCE()?
Answer:
In the 2018 v4.15 release, the Linux kernel’s ACCESS_ONCE() was replaced by READ_
ONCE() and WRITE_ONCE() for reads and writes, respectively [Cor12, Cor14a, Rut17].
ACCESS_ONCE() was introduced as a helper in RCU code, but was promoted to core API
soon afterward [McK07b, Tor08]. Linux kernel’s READ_ONCE() and WRITE_ONCE()
have evolved into complex forms that look quite different than the original ACCESS_
ONCE() implementation due to the need to support access-once semantics for large
structures, but with the possibility of load/store tearing if the structure cannot be loaded
and stored with a single machine instruction. ❑

What happened to the Linux-kernel equivalents to fork() and wait()?
Answer:
They don’t really exist. All tasks executing within the Linux kernel share memory, at
least unless you want to do a huge amount of memory-mapping work by hand. ❑

What problems could occur if the variable counter were incremented without the
protection of mutex?
Answer:
On CPUs with load-store architectures, incrementing counter might compile into
something like the following:
LOAD counter,r0
INC r0
STORE r0,counter
On such machines, two threads might simultaneously load the value of counter,
each increment it, and each store the result. The new value of counter will then only
be one greater than before, despite two threads each incrementing it. ❑
v2023.06.11a

What is wrong with loading Listing 4.14’s global_ptr up to three times?
Answer:
Suppose that global_ptr is initially non-NULL, but that some other thread sets global_
ptr to NULL. Suppose further that line 1 of the transformed code (Listing 4.15) executes
just before global_ptr is set to NULL and line 2 just after. Then line 1 will conclude
that global_ptr is non-NULL, line 2 will conclude that it is less than high_address,
so that line 3 passes do_low() a NULL pointer, which do_low() just might not be
prepared to deal with.
Your editor made exactly this mistake in the DYNIX/ptx kernel’s memory allocator
in the early 1990s. Tracking down the bug consumed a holiday weekend not just for
your editor, but also for several of his colleagues. In short, this is not a new problem,
nor is it likely to go away on its own. ❑

Why does it matter whether do_something() and do_something_else() in List-
ing 4.18 are inline functions?
Answer:
Because gp is not a static variable, if either do_something() or do_something_
else() were separately compiled, the compiler would have to assume that either or
both of these two functions might change the value of gp. This possibility would force
the compiler to reload gp on line 15, thus avoiding the NULL-pointer dereference. ❑

Ouch! So can’t the compiler invent a store to a normal variable pretty much any time
it likes?
Answer:
Thankfully, the answer is no. This is because the compiler is forbidden from introducing
data races. The case of inventing a store just before a normal store is quite special:
It is not possible for some other entity, be it CPU, thread, signal handler, or interrupt
handler, to be able to see the invented store unless the code already has a data race, even
without the invented store. And if the code already has a data race, it already invokes
the dreaded spectre of undefined behavior, which allows the compiler to generate pretty
much whatever code it wants, regardless of the wishes of the developer.
But if the original store is volatile, as in WRITE_ONCE(), for all the compiler knows,
there might be a side effect associated with the store that could signal some other thread,
allowing data-race-free access to the variable. By inventing the store, the compiler
might be introducing a data race, which it is not permitted to do.
Furthermore, in Listing 4.21, the address of that variable is passed to do_a_bunch_
of_stuff(). If the compiler can see this function’s definition, and notes that a is
unconditionally stored to without any synchronization operations, then the compiler can
be quite sure that it is not introducing a data race in this case.
In the case of volatile and atomic variables, the compiler is specifically forbidden
from inventing writes. ❑
v2023.06.11a

But aren’t full memory barriers very heavyweight? Isn’t there a cheaper way to enforce
the ordering needed in Listing 4.29?
Answer:
As is often the case, the answer is “it depends”. However, if only two threads are accessing
the status and other_task_ready variables, then the smp_store_release() and
smp_load_acquire() functions discussed in Section 4.3.5 will suffice. ❑

What needs to happen if an interrupt or signal handler might itself be interrupted?
Answer:
Then that interrupt handler must follow the same rules that are followed by other
interrupted code. Only those handlers that cannot be themselves interrupted or that
access no variables shared with an interrupting handler may safely use plain accesses,
and even then only if those variables cannot be concurrently accessed by some other
CPU or thread. ❑

How could you work around the lack of a per-thread-variable API on systems that do
not provide it?
Answer:
One approach would be to create an array indexed by smp_thread_id(), and another
would be to use a hash table to map from smp_thread_id() to an array index—which
is in fact what this set of APIs does in pthread environments.
Another approach would be for the parent to allocate a structure containing fields
for each desired per-thread variable, then pass this to the child during thread creation.
However, this approach can impose large software-engineering costs in large systems.
To see this, imagine if all global variables in a large system had to be declared in a
single file, regardless of whether or not they were C static variables! ❑

What do you do if you need a per-thread (not per-CPU!) variable in the Linux kernel?
Answer:
First, needing a per-thread variable is less likely than you might think. Per-CPU
variables can often do a per-thread variable’s job. For example, if you only need to
do addition, bitwise AND, bitwise OR, exchange, or compare-and-exchange, then the
this_cpu_add(), this_cpu_add_return(), this_cpu_and(), this_cpu_or(),
this_cpu_xchg(), this_cpu_cmpxchg(), and this_cpu_cmpxchg_double() op-
erations, respectively, will do the job cheaply and atomically with respect to context
switches, interrupt handlers, and softirq handlers, but not non-maskable interrupts.
Second, within a preemption-disabled region of code, for example, one surrounded
by the preempt_disable() and preempt_enable() macros, the current task is
guaranteed to remain executing on the current CPU. Therefore, while within one
such region, any series of accesses to per-CPU variables is atomic with respect to
v2023.06.11a
E.5. COUNTING 729
context switches, though not with respect to interrupt handlers, softirq handlers, and
non-maskable interrupts. But please be aware that a preemption-disabled region of
code that runs for more than a few microseconds will not be looked upon with favor by
people attempting to construct real-time systems.
Third, a field added to the task_struct structure acts as set of per-task variables.
However, there are those who keep a close eye on the size of this structure, and these
people are likely to ask hard questions about the need for any added fields. Therefore,
if your field is being added for some facility that is only built into some kernels, you
should definitely place your new task_struct fields under an appropriate #ifdef.
Fourth and finally, your per-task variable might instead be located in some other
structure and protected by some synchronization mechanism that is already in use. For
example, if your code must hold a given lock, can accesses to this storage instead be
protected by that lock? The fact that this is at the end of the list notwithstanding, you
should look into this possibility first, not last! ❑

Wouldn’t the shell normally use vfork() rather than fork()?
Answer:
It might well do that, however, checking is left as an exercise for the reader. But in the
meantime, I hope that we can agree that vfork() is a variant of fork(), so that we
can use fork() as a generic term covering both. ❑
E.5 Counting

Why should efficient and scalable counting be hard??? After all, computers have
special hardware for the sole purpose of doing counting!!!
Answer:
Because the straightforward counting algorithms, for example, atomic operations on
a shared counter, either are slow and scale badly, or are inaccurate, as will be seen in
Section 5.1. ❑

Network-packet counting problem. Suppose that you need to collect statistics on the
number of networking packets transmitted and received. Packets might be transmitted
or received by any CPU on the system. Suppose further that your system is capable
of handling millions of packets per second per CPU, and that a systems-monitoring
package reads the count every five seconds. How would you implement this counter?
Answer:
Hint: The act of updating the counter must be blazingly fast, but because the counter is
read out only about once in five million updates, the act of reading out the counter can be
quite slow. In addition, the value read out normally need not be all that accurate—after
all, since the counter is updated a thousand times per millisecond, we should be able to
v2023.06.11a
work with a value that is within a few thousand counts of the “true value”, whatever
“true value” might mean in this context. However, the value read out should maintain
roughly the same absolute error over time. For example, a 1 % error might be just fine
when the count is on the order of a million or so, but might be absolutely unacceptable
once the count reaches a trillion. See Section 5.2. ❑

Approximate structure-allocation limit problem. Suppose that you need to maintain
a count of the number of structures allocated in order to fail any allocations once the
number of structures in use exceeds a limit (say, 10,000). Suppose further that the
structures are short-lived, the limit is rarely exceeded, and a “sloppy” approximate
limit is acceptable.
Answer:
Hint: The act of updating the counter must again be blazingly fast, but the counter is
read out each time that the counter is increased. However, the value read out need not
be accurate except that it must distinguish approximately between values below the limit
and values greater than or equal to the limit. See Section 5.3. ❑

Exact structure-allocation limit problem. Suppose that you need to maintain a
count of the number of structures allocated in order to fail any allocations once the
number of structures in use exceeds an exact limit (again, say 10,000). Suppose further
that these structures are short-lived, and that the limit is rarely exceeded, that there is
almost always at least one structure in use, and suppose further still that it is necessary
to know exactly when this counter reaches zero, for example, in order to free up some
memory that is not required unless there is at least one structure in use.
Answer:
Hint: The act of updating the counter must once again be blazingly fast, but the counter
is read out each time that the counter is increased. However, the value read out need not
be accurate except that it absolutely must distinguish perfectly between values between
the limit and zero on the one hand, and values that either are less than or equal to zero
or are greater than or equal to the limit on the other hand. See Section 5.4. ❑

Removable I/O device access-count problem. Suppose that you need to maintain a
reference count on a heavily used removable mass-storage device, so that you can tell
the user when it is safe to remove the device. As usual, the user indicates a desire to
remove the device, and the system tells the user when it is safe to do so.
Answer:
Hint: Yet again, the act of updating the counter must be blazingly fast and scalable
in order to avoid slowing down I/O operations, but because the counter is read out
only when the user wishes to remove the device, the counter read-out operation can be
extremely slow. Furthermore, there is no need to be able to read out the counter at all
unless the user has already indicated a desire to remove the device. In addition, the
value read out need not be accurate except that it absolutely must distinguish perfectly
between non-zero and zero values, and even then only when the device is in the process
v2023.06.11a
E.5. COUNTING 731
of being removed. However, once it has read out a zero value, it must act to keep the
value at zero until it has taken some action to prevent subsequent threads from gaining
access to the device being removed. See Section 5.4.6. ❑

One thing that could be simpler is ++ instead of that concatenation of READ_ONCE()
and WRITE_ONCE(). Why all that extra typing???
Answer:
See Section 4.3.4.1 on page 63 for more information on how the compiler can cause
trouble, as well as how READ_ONCE() and WRITE_ONCE() can avoid this trouble. ❑

But can’t a smart compiler prove that line 5 of Listing 5.1 is equivalent to the ++
operator and produce an x86 add-to-memory instruction? And won’t the CPU cache
cause this to be atomic?
Answer:
Although the ++ operator could be atomic, there is no requirement that it be so unless it
is applied to a C11 _Atomic variable. And indeed, in the absence of _Atomic, GCC
often chooses to load the value to a register, increment the register, then store the value
to memory, which is decidedly non-atomic.
Furthermore, note the volatile casts in READ_ONCE() and WRITE_ONCE(), which
tell the compiler that the location might well be an MMIO device register. Because
MMIO registers are not cached, it would be unwise for the compiler to assume that the
increment operation is atomic. ❑

The 8-figure accuracy on the number of failures indicates that you really did test this.
Why would it be necessary to test such a trivial program, especially when the bug is
easily seen by inspection?
Answer:
Not only are there very few trivial parallel programs, and most days I am not so sure
that there are many trivial sequential programs, either.
No matter how small or simple the program, if you haven’t tested it, it does not work.
And even if you have tested it, Murphy’s Law says that there will be at least a few bugs
still lurking.
Furthermore, while proofs of correctness certainly do have their place, they never will
replace testing, including the counttorture.h test setup used here. After all, proofs
are only as good as the assumptions that they are based on. Finally, proofs can be every
bit as buggy as are programs! ❑
v2023.06.11a

Why doesn’t the horizontal dashed line on the x axis meet the diagonal line at 𝑥 = 1?
Answer:
Because of the overhead of the atomic operation. The dashed line on the x axis
represents the overhead of a single non-atomic increment. After all, an ideal algorithm
would not only scale linearly, it would also incur no performance penalty compared to
single-threaded code.
This level of idealism may seem severe, but if it is good enough for Linus Torvalds, it
is good enough for you. ❑

But atomic increment is still pretty fast. And incrementing a single variable in a tight
loop sounds pretty unrealistic to me, after all, most of the program’s execution should
be devoted to actually doing work, not accounting for the work it has done! Why
should I care about making this go faster?
Answer:
In many cases, atomic increment will in fact be fast enough for you. In those cases,
you should by all means use atomic increment. That said, there are many real-world
situations where more elaborate counting algorithms are required. The canonical
example of such a situation is counting packets and bytes in highly optimized networking
stacks, where it is all too easy to find much of the execution time going into these sorts
of accounting tasks, especially on large multiprocessors.
In addition, as noted at the beginning of this chapter, counting provides an excellent
view of the issues encountered in shared-memory parallel programs. ❑

But why can’t CPU designers simply ship the addition operation to the data, avoiding
the need to circulate the cache line containing the global variable being incremented?
Answer:
It might well be possible to do this in some cases. However, there are a few complications:
1. If the value of the variable is required, then the thread will be forced to wait for the
operation to be shipped to the data, and then for the result to be shipped back.
2. If the atomic increment must be ordered with respect to prior and/or subsequent
operations, then the thread will be forced to wait for the operation to be shipped to
the data, and for an indication that the operation completed to be shipped back.
3. Shipping operations among CPUs will likely require more lines in the system
interconnect, which will consume more die area and more electrical power.
But what if neither of the first two conditions holds? Then you should think carefully
about the algorithms discussed in Section 5.2, which achieve near-ideal performance on
commodity hardware.
If either or both of the first two conditions hold, there is some hope for improved
hardware. One could imagine the hardware implementing a combining tree, so that the
v2023.06.11a
E.5. COUNTING 733

Figure E.1: Data Flow For Global Combining-Tree Atomic Increment
increment requests from multiple CPUs are combined by the hardware into a single
addition when the combined request reaches the hardware. The hardware could also
apply an order to the requests, thus returning to each CPU the return value corresponding
to its particular atomic increment. This results in instruction latency that varies as
O (log 𝑁), where 𝑁 is the number of CPUs, as shown in Figure E.1. And CPUs with
this sort of hardware optimization started to appear in 2011.
This is a great improvement over the O (𝑁) performance of current hardware shown in
Figure 5.2, and it is possible that hardware latencies might decrease further if innovations
such as three-dimensional fabrication prove practical. Nevertheless, we will see that in
some important special cases, software can do much better. ❑

But doesn’t the fact that C’s “integers” are limited in size complicate things?
Answer:
No, because modulo addition is still commutative and associative. At least as long as
you use unsigned integers. Recall that in the C standard, overflow of signed integers
results in undefined behavior, never mind the fact that machines that do anything other
than wrap on overflow are quite rare these days. Unfortunately, compilers frequently
carry out optimizations that assume that signed integers will not overflow, so if your
code allows signed integers to overflow, you can run into trouble even on modern
twos-complement hardware.
That said, one potential source of additional complexity arises when attempting to
gather (say) a 64-bit sum from 32-bit per-thread counters. Dealing with this added
complexity is left as an exercise for the reader, for whom some of the techniques
introduced later in this chapter could be quite helpful. ❑

An array??? But doesn’t that limit the number of threads?
Answer:
It can, and in this toy implementation, it does. But it is not that hard to come up with an
alternative implementation that permits an arbitrary number of threads, for example,
using C11’s _Thread_local facility, as shown in Section 5.2.3. ❑
v2023.06.11a

What other nasty optimizations could GCC apply?
Answer:
See Sections 4.3.4.1 and 15.3 for more information. One nasty optimization would be
to apply common subexpression elimination to successive calls to the read_count()
function, which might come as a surprise to code expecting changes in the values
returned from successive calls to that function. ❑

How does the per-thread counter variable in Listing 5.3 get initialized?
Answer:
The C standard specifies that the initial value of global variables is zero, unless they are
explicitly initialized, thus implicitly initializing all the instances of counter to zero.
Besides, in the common case where the user is interested only in differences between
consecutive reads from statistical counters, the initial value is irrelevant. ❑

How is the code in Listing 5.3 supposed to permit more than one counter?
Answer:
Indeed, this toy example does not support more than one counter. Modifying it so that it
can provide multiple counters is left as an exercise to the reader. ❑

The read operation takes time to sum up the per-thread values, and during that
time, the counter could well be changing. This means that the value returned by
read_count() in Listing 5.3 will not necessarily be exact. Assume that the counter is
being incremented at rate 𝑟 counts per unit time, and that read_count()’s execution
consumes 𝛥 units of time. What is the expected error in the return value?
Answer:
Let’s do worst-case analysis first, followed by a less conservative analysis.
In the worst case, the read operation completes immediately, but is then delayed for 𝛥
time units before returning, in which case the worst-case error is simply 𝑟 𝛥.
This worst-case behavior is rather unlikely, so let us instead consider the case where
the reads from each of the 𝑁 counters is spaced equally over the time period 𝛥. There
will be 𝑁 + 1 intervals of duration 𝑁𝛥+1 between the 𝑁 reads. The rate 𝑟 of increments
is expected to be spread evenly over the 𝑁 counters, for 𝑁𝑟 increments per unit time for
each individual counter. The error due to the delay after the read from the last thread’s
counter will be given by 𝑁 (𝑟𝑁𝛥+1) , the second-to-last thread’s counter by 𝑁 (2𝑟𝑁𝛥+1) , the
third-to-last by 𝑁 (3𝑟𝑁𝛥+1) , and so on. The total error is given by the sum of the errors due
to the reads from each thread’s counter, which is:
𝑁
𝑟𝛥 ∑︁
𝑖 (E.1)
𝑁 (𝑁 + 1) 𝑖=1
Expressing the summation in closed form yields:
v2023.06.11a
E.5. COUNTING 735
𝑟𝛥 𝑁 (𝑁 + 1)
(E.2)
𝑁 (𝑁 + 1) 2
Canceling yields the intuitively expected result:
𝑟𝛥
(E.3)
2
It is important to remember that error continues accumulating as the caller executes

code making use of the count returned by the read operation. For example, if the caller
spends time 𝑡 executing some computation based on the result of the returned count, the
worst-case error will have increased to 𝑟 ( 𝛥 + 𝑡).
The expected error will have similarly increased to:

𝛥
𝑟 +𝑡 (E.4)
2
Of course, it is sometimes unacceptable for the counter to continue incrementing

during the read operation. Section 5.4.6 discusses a way to handle this situation.
Thus far, we have been considering a counter that is only increased, never decreased.
If the counter value is being changed by 𝑟 counts per unit time, but in either direction,
we should expect the error to reduce. However, the worst case is unchanged because
although the counter could move in either direction, the worst case is when the read
operation completes immediately, but then is delayed for 𝛥 time units, during which
time all the changes in the counter’s value move it in the same direction, again giving us
an absolute error of 𝑟 𝛥.
There are a number of ways to compute the average error, based on a variety of
assumptions about the patterns of increments and decrements. For simplicity, let’s
assume that the 𝑓 fraction of the operations are decrements, and that the error of
interest is the deviation from the counter’s long-term trend line. Under this assumption,
if 𝑓 is less than or equal to 0.5, each decrement will be canceled by an increment,
so that 2 𝑓 of the operations will cancel each other, leaving 1 − 2 𝑓 of the operations
being uncanceled increments. On the other hand, if 𝑓 is greater than 0.5, 1 − 𝑓 of
the decrements are canceled by increments, so that the counter moves in the negative
direction by −1 + 2 (1 − 𝑓 ), which simplifies to 1 − 2 𝑓 , so that the counter moves an
average of 1 − 2 𝑓 per operation in either case. Therefore, that the long-term movement
of the counter is given by (1 − 2 𝑓 ) 𝑟. Plugging this into Eq. E.3 yields:
(1 − 2 𝑓 ) 𝑟 𝛥
(E.5)
2
All that aside, in most uses of statistical counters, the error in the value returned by
read_count() is irrelevant. This irrelevance is due to the fact that the time required for
read_count() to execute is normally extremely small compared to the time interval
between successive calls to read_count(). ❑
v2023.06.11a

Doesn’t that explicit counterp array in Listing 5.4 reimpose an arbitrary limit on the
number of threads? Why doesn’t the C language provide a per_thread() interface,
similar to the Linux kernel’s per_cpu() primitive, to allow threads to more easily
access each others’ per-thread variables?
Answer:
Why indeed?
To be fair, user-mode thread-local storage faces some challenges that the Linux kernel
gets to ignore. When a user-level thread exits, its per-thread variables all disappear,
which complicates the problem of per-thread-variable access, particularly before the
advent of user-level RCU (see Section 9.5). In contrast, in the Linux kernel, when a
CPU goes offline, that CPU’s per-CPU variables remain mapped and accessible.
Similarly, when a new user-level thread is created, its per-thread variables suddenly
come into existence. In contrast, in the Linux kernel, all per-CPU variables are mapped
and initialized at boot time, regardless of whether the corresponding CPU exists yet, or
indeed, whether the corresponding CPU will ever exist.
A key limitation that the Linux kernel imposes is a compile-time maximum bound
on the number of CPUs, namely, CONFIG_NR_CPUS, along with a typically tighter
boot-time bound of nr_cpu_ids. In contrast, in user space, there is not necessarily a
hard-coded upper limit on the number of threads.
Of course, both environments must handle dynamically loaded code (dynamic libraries
in user space, kernel modules in the Linux kernel), which increases the complexity of
per-thread variables.
These complications make it significantly harder for user-space environments to
provide access to other threads’ per-thread variables. Nevertheless, such access is highly
useful, and it is hoped that it will someday appear.
In the meantime, textbook examples such as this one can use arrays whose limits can
be easily adjusted by the user. Alternatively, such arrays can be dynamically allocated
and expanded as needed at runtime. Finally, variable-length data structures such as
linked lists can be used, as is done in the userspace RCU library [Des09b, DMS+ 12].
This last approach can also reduce false sharing in some cases. ❑

Doesn’t the check for NULL on line 19 of Listing 5.4 add extra branch mispredictions?
Why not have a variable set permanently to zero, and point unused counter-pointers to
that variable rather than setting them to NULL?
Answer:
This is a reasonable strategy. Checking for the performance difference is left as
an exercise for the reader. However, please keep in mind that the fastpath is not
read_count(), but rather inc_count(). ❑

Why on earth do we need something as heavyweight as a lock guarding the summation
in the function read_count() in Listing 5.4?
Answer:
Remember, when a thread exits, its per-thread variables disappear. Therefore, if we
v2023.06.11a
E.5. COUNTING 737
attempt to access a given thread’s per-thread variables after that thread exits, we will get
a segmentation fault. The lock coordinates summation and thread exit, preventing this
scenario.
Of course, we could instead read-acquire a reader-writer lock, but Chapter 9 will
introduce even lighter-weight mechanisms for implementing the required coordination.
Another approach would be to use an array instead of a per-thread variable, which, as
Alexey Roytman notes, would eliminate the tests against NULL. However, array accesses
are often slower than accesses to per-thread variables, and use of an array would imply
a fixed upper bound on the number of threads. Also, note that neither tests nor locks are
needed on the inc_count() fastpath. ❑

Why on earth do we need to acquire the lock in count_register_thread() in
Listing 5.4? It is a single properly aligned machine-word store to a location that no
other thread is modifying, so it should be atomic anyway, right?
Answer:
This lock could in fact be omitted, but better safe than sorry, especially given that this
function is executed only at thread startup, and is therefore not on any critical path. Now,
if we were testing on machines with thousands of CPUs, we might need to omit the lock,
but on machines with “only” a hundred or so CPUs, there is no need to get fancy. ❑

Fine, but the Linux kernel doesn’t have to acquire a lock when reading out the aggregate
value of per-CPU counters. So why should user-space code need to do this???
Answer:
Remember, the Linux kernel’s per-CPU variables are always accessible, even if the
corresponding CPU is offline—even if the corresponding CPU never existed and never
will exist.
One workaround is to ensure that each thread continues to exist until all threads
are finished, as shown in Listing E.1 (count_tstat.c). Analysis of this code is
left as an exercise to the reader, however, please note that it requires tweaks in
the counttorture.h counter-evaluation scheme. (Hint: See #ifndef KEEP_GCC_
THREAD_LOCAL.) Chapter 9 will introduce synchronization mechanisms that handle
this situation in a much more graceful manner. ❑

Why doesn’t inc_count() in Listing 5.5 need to use atomic instructions? After all,
we now have multiple threads accessing the per-thread counters!
Answer:
Because one of the two threads only reads, and because the variable is aligned and
machine-sized, non-atomic instructions suffice. That said, the READ_ONCE() macro is
used to prevent compiler optimizations that might otherwise prevent the counter updates
from becoming visible to eventual().5
5 A simple definition of READ_ONCE() is shown in Listing 4.9.
v2023.06.11a
Listing E.1: Per-Thread Statistical Counters With Lockless Summation

3 int finalthreadcount = 0;
5
7 {
9 }
10
12 /* need to tweak counttorture! */
13 {
14 int t;
15 unsigned long sum = 0;
16
18 if (READ_ONCE(counterp[t]) != NULL)
20 }
21 return sum;
22 }
23
25 {
26 WRITE_ONCE(counterp[smp_thread_id()], &counter);
27 }
28
30 {
32 finalthreadcount++;
34 while (READ_ONCE(finalthreadcount) < nthreadsexpected)
36 }
An older version of this algorithm did in fact use atomic instructions, kudos to
Ersoy Bayramoglu for pointing out that they are in fact unnecessary. However, note
that on a 32-bit system, the per-thread counter variables might need to be limited to
32 bits in order to sum them accurately, but with a 64-bit global_count variable to
avoid overflow. In this case, it is necessary to zero the per-thread counter variables
periodically in order to avoid overflow, which does require atomic instructions. It is
extremely important to note that this zeroing cannot be delayed too long or overflow of
the smaller per-thread variables will result. This approach therefore imposes real-time
requirements on the underlying system, and in turn must be used with extreme care.
In contrast, if all variables are the same size, overflow of any variable is harmless
because the eventual sum will be modulo the word size. ❑

Won’t the single global thread in the function eventual() of Listing 5.5 be just as
severe a bottleneck as a global lock would be?
Answer:
In this case, no. What will happen instead is that as the number of threads increases, the
estimate of the counter value returned by read_count() will become more inaccurate.
❑
v2023.06.11a
E.5. COUNTING 739

Won’t the estimate returned by read_count() in Listing 5.5 become increasingly
inaccurate as the number of threads rises?
Answer:
Yes. If this proves problematic, one fix is to provide multiple eventual() threads,
each covering its own subset of the other threads. In more extreme cases, a tree-like
hierarchy of eventual() threads might be required. ❑

Given that in the eventually-consistent algorithm shown in Listing 5.5 both reads and
updates have extremely low overhead and are extremely scalable, why would anyone
bother with the implementation described in Section 5.2.2, given its costly read-side
code?
Answer:
The thread executing eventual() consumes CPU time. As more of these eventually-
consistent counters are added, the resulting eventual() threads will eventually consume
all available CPUs. This implementation therefore suffers a different sort of scalability
limitation, with the scalability limit being in terms of the number of eventually consistent
counters rather than in terms of the number of threads or CPUs.
Of course, it is possible to make other tradeoffs. For example, a single thread could
be created to handle all eventually-consistent counters, which would limit the overhead
to a single CPU, but would result in increasing update-to-read latencies as the number
of counters increased. Alternatively, that single thread could track the update rates
of the counters, visiting the frequently-updated counters more frequently. In addition,
the number of threads handling the counters could be set to some fraction of the total
number of CPUs, and perhaps also adjusted at runtime. Finally, each counter could
specify its latency, and deadline-scheduling techniques could be used to provide the
required latencies to each counter.
There are no doubt many other tradeoffs that could be made. ❑

What is the accuracy of the estimate returned by read_count() in Listing 5.5?
Answer:
A straightforward way to evaluate this estimate is to use the analysis derived in Quick
Quiz 5.17, but set 𝛥 to the interval between the beginnings of successive runs of the
eventual() thread. Handling the case where a given counter has multiple eventual()
threads is left as an exercise for the reader. ❑

What fundamental difference is there between counting packets and counting the total
number of bytes in the packets, given that the packets vary in size?
Answer:
When counting packets, the counter is only incremented by the value one. On the other
hand, when counting bytes, the counter might be incremented by largish numbers.
v2023.06.11a
Why does this matter? Because in the increment-by-one case, the value returned
will be exact in the sense that the counter must necessarily have taken on that value at
some point in time, even if it is impossible to say precisely when that point occurred.
In contrast, when counting bytes, two different threads might return values that are
inconsistent with any global ordering of operations.
To see this, suppose that thread 0 adds the value three to its counter, thread 1 adds the
value five to its counter, and threads 2 and 3 sum the counters. If the system is “weakly
ordered” or if the compiler uses aggressive optimizations, thread 2 might find the sum
to be three and thread 3 might find the sum to be five. The only possible global orders of
the sequence of values of the counter are 0,3,8 and 0,5,8, and neither order is consistent
with the results obtained.
If you missed this one, you are not alone. Michael Scott used this question to stump
Paul E. McKenney during Paul’s Ph.D. defense. ❑

Given that the reader must sum all the threads’ counters, this counter-read operation
could take a long time given large numbers of threads. Is there any way that the
increment operation can remain fast and scalable while allowing readers to also enjoy
not only reasonable performance and scalability, but also good accuracy?
Answer:
One approach would be to maintain a global approximation to the value, similar to the
approach described in Section 5.2.4. Updaters would increment their per-thread variable,
but when it reached some predefined limit, atomically add it to a global variable, then
zero their per-thread variable. This would permit a tradeoff between average increment
overhead and accuracy of the value read out. In particular, it would allow sharp bounds
on the read-side inaccuracy.
Another approach makes use of the fact that readers often care only about certain
transitions in value, not in the exact value. This approach is examined in Section 5.3.
The reader is encouraged to think up and try out other approaches, for example, using
a combining tree. ❑

Why does Listing 5.7 provide add_count() and sub_count() instead of the inc_
count() and dec_count() interfaces show in Section 5.2?
Answer:
Because structures come in different sizes. Of course, a limit counter corresponding to a
specific size of structure might still be able to use inc_count() and dec_count(). ❑

What is with the strange form of the condition on line 3 of Listing 5.7? Why not the
more intuitive form of the fastpath shown in Listing 5.8?
Answer:
Two words. “Integer overflow.”
Try the formulation in Listing 5.8 with counter equal to 10 and delta equal to
ULONG_MAX. Then try it again with the code shown in Listing 5.7.
v2023.06.11a
E.5. COUNTING 741
A good understanding of integer overflow will be required for the rest of this example,
so if you have never dealt with integer overflow before, please try several examples to
get the hang of it. Integer overflow can sometimes be more difficult to get right than
parallel algorithms! ❑

Why does globalize_count() zero the per-thread variables, only to later call
balance_count() to refill them in Listing 5.7? Why not just leave the per-thread
variables non-zero?
Answer:
That is in fact what an earlier version of this code did. But addition and subtraction
are extremely cheap, and handling all of the special cases that arise is quite complex.
Again, feel free to try it yourself, but beware of integer overflow! ❑

Given that globalreserve counted against us in add_count(), why doesn’t it count
for us in sub_count() in Listing 5.7?
Answer:
The globalreserve variable tracks the sum of all threads’ countermax variables.
The sum of these threads’ counter variables might be anywhere from zero to
globalreserve. We must therefore take a conservative approach, assuming that
all threads’ counter variables are full in add_count() and that they are all empty in
sub_count().
But remember this question, as we will come back to it later. ❑

Suppose that one thread invokes add_count() shown in Listing 5.7, and then another
thread invokes sub_count(). Won’t sub_count() return failure even though the
value of the counter is non-zero?
Answer:
Indeed it will! In many cases, this will be a problem, as discussed in Section 5.3.3, and
in those cases the algorithms from Section 5.4 will likely be preferable. ❑

Why have both add_count() and sub_count() in Listing 5.7? Why not simply
pass a negative number to add_count()?
Answer:
Given that add_count() takes an unsigned long as its argument, it is going to be a
bit tough to pass it a negative number. And unless you have some anti-matter memory,
there is little point in allowing negative numbers when counting the number of structures
in use!
All kidding aside, it would of course be possible to combine add_count() and
sub_count(), however, the if conditions on the combined function would be more
complex than in the current pair of functions, which would in turn mean slower execution
of these fast paths. ❑
v2023.06.11a

Why set counter to countermax / 2 in line 15 of Listing 5.9? Wouldn’t it be
simpler to just take countermax counts?
Answer:
First, it really is reserving countermax counts (see line 14), however, it adjusts so that
only half of these are actually in use by the thread at the moment. This allows the thread
to carry out at least countermax / 2 increments or decrements before having to refer
back to globalcount again.
Note that the accounting in globalcount remains accurate, thanks to the adjustment
in line 18. ❑

In Figure 5.6, even though a quarter of the remaining count up to the limit is assigned
to thread 0, only an eighth of the remaining count is consumed, as indicated by the
uppermost dotted line connecting the center and the rightmost configurations. Why is
that?
Answer:
The reason this happened is that thread 0’s counter was set to half of its countermax.
Thus, of the quarter assigned to thread 0, half of that quarter (one eighth) came from
globalcount, leaving the other half (again, one eighth) to come from the remaining
count.
There are two purposes for taking this approach: (1) To allow thread 0 to use the
fastpath for decrements as well as increments and (2) To reduce the inaccuracies if all
threads are monotonically incrementing up towards the limit. To see this last point, step
through the algorithm and watch what it does. ❑

Why is it necessary to atomically manipulate the thread’s counter and countermax
variables as a unit? Wouldn’t it be good enough to atomically manipulate them
individually?
Answer:
This might well be possible, but great care is required. Note that removing counter
without first zeroing countermax could result in the corresponding thread increasing
counter immediately after it was zeroed, completely negating the effect of zeroing the
counter.
The opposite ordering, namely zeroing countermax and then removing counter,
can also result in a non-zero counter. To see this, consider the following sequence of
events:
1. Thread A fetches its countermax, and finds that it is non-zero.
2. Thread B zeroes Thread A’s countermax.
3. Thread B removes Thread A’s counter.
4. Thread A, having found that its countermax is non-zero, proceeds to add to its
counter, resulting in a non-zero value for counter.
v2023.06.11a
E.5. COUNTING 743
Again, it might well be possible to atomically manipulate countermax and counter

as separate variables, but it is clear that great care is required. It is also quite likely that
doing so will slow down the fastpath.
Exploring these possibilities are left as exercises for the reader. ❑

In what way does line 7 of Listing 5.12 violate the C standard?
Answer:
It assumes eight bits per byte. This assumption does hold for all current commodity
microprocessors that can be easily assembled into shared-memory multiprocessors, but
certainly does not hold for all computer systems that have ever run C code. (What could
you do instead in order to comply with the C standard? What drawbacks would it have?)
❑

Given that there is only one counterandmax variable, why bother passing in a pointer
to it on line 18 of Listing 5.12?
Answer:
There is only one counterandmax variable per thread. Later, we will see code that
needs to pass other threads’ counterandmax variables to split_counterandmax().
❑

Why does merge_counterandmax() in Listing 5.12 return an int rather than storing
directly into an atomic_t?
Answer:
Later, we will see that we need the int return to pass to the atomic_cmpxchg()
primitive. ❑

Yecch! Why the ugly goto on line 11 of Listing 5.13? Haven’t you heard of the break
statement???
Answer:
Replacing the goto with a break would require keeping a flag to determine whether
or not line 15 should return, which is not the sort of thing you want on a fastpath. If
you really hate the goto that much, your best bet would be to pull the fastpath into a
separate function that returned success or failure, with “failure” indicating a need for
the slowpath. This is left as an exercise for goto-hating readers. ❑
v2023.06.11a

Why would the atomic_cmpxchg() primitive at lines 13–14 of Listing 5.13 ever fail?
After all, we picked up its old value on line 9 and have not changed it!
Answer:
Later, we will see how the flush_local_count() function in Listing 5.15 might
update this thread’s counterandmax variable concurrently with the execution of the
fastpath on lines 8–14 of Listing 5.13. ❑

What stops a thread from simply refilling its counterandmax variable immediately
after flush_local_count() on line 14 of Listing 5.15 empties it?
Answer:
This other thread cannot refill its counterandmax until the caller of flush_local_
count() releases the gblcnt_mutex. By that time, the caller of flush_local_
count() will have finished making use of the counts, so there will be no problem with
this other thread refilling—assuming that the value of globalcount is large enough to
permit a refill. ❑

What prevents concurrent execution of the fastpath of either add_count() or sub_
count() from interfering with the counterandmax variable while flush_local_
count() is accessing it on line 27 of Listing 5.15?
Answer:
Nothing. Consider the following three cases:
1. If flush_local_count()’s atomic_xchg() executes before the split_

counterandmax() of either fastpath, then the fastpath will see a zero counter
and countermax, and will thus transfer to the slowpath (unless of course delta
is zero).
2. If flush_local_count()’s atomic_xchg() executes after the split_

counterandmax() of either fastpath, but before that fastpath’s atomic_
cmpxchg(), then the atomic_cmpxchg() will fail, causing the fastpath to restart,
which reduces to case 1 above.
3. If flush_local_count()’s atomic_xchg() executes after the atomic_

cmpxchg() of either fastpath, then the fastpath will (most likely) complete
successfully before flush_local_count() zeroes the thread’s counterandmax
variable.
Either way, the race is resolved correctly. ❑
v2023.06.11a
E.5. COUNTING 745

Given that the atomic_set() primitive does a simple store to the specified atomic_
t, how can line 21 of balance_count() in Listing 5.16 work correctly in face of
concurrent flush_local_count() updates to this variable?
Answer:
The caller of both balance_count() and flush_local_count() hold gblcnt_
mutex, so only one may be executing at a given time. ❑

But signal handlers can be migrated to some other CPU while running. Doesn’t this
possibility require that atomic instructions and memory barriers are required to reliably
communicate between a thread and a signal handler that interrupts that thread?
Answer:
No. If the signal handler is migrated to another CPU, then the interrupted thread is also
migrated along with it. ❑

In Figure 5.7, why is the REQ theft state colored red?
Answer:
To indicate that only the fastpath is permitted to change the theft state, and that if the
thread remains in this state for too long, the thread running the slowpath will resend the
POSIX signal. ❑

In Figure 5.7, what is the point of having separate REQ and ACK theft states? Why
not simplify the state machine by collapsing them into a single REQACK state? Then
whichever of the signal handler or the fastpath gets there first could set the state to
READY.
Answer:
Reasons why collapsing the REQ and ACK states would be a very bad idea include:
1. The slowpath uses the REQ and ACK states to determine whether the signal should
be retransmitted. If the states were collapsed, the slowpath would have no choice
but to send redundant signals, which would have the unhelpful effect of needlessly
slowing down the fastpath.
2. The following race would result:
(a) The slowpath sets a given thread’s state to REQACK.

(b) That thread has just finished its fastpath, and notes the REQACK state.
(c) The thread receives the signal, which also notes the REQACK state, and,
because there is no fastpath in effect, sets the state to READY.
(d) The slowpath notes the READY state, steals the count, and sets the state to
IDLE, and completes.
v2023.06.11a
(e) The fastpath sets the state to READY, disabling further fastpath execution for
this thread.
The basic problem here is that the combined REQACK state can be referenced by
both the signal handler and the fastpath. The clear separation maintained by the
four-state setup ensures orderly state transitions.
That said, you might well be able to make a three-state setup work correctly. If you
do succeed, compare carefully to the four-state setup. Is the three-state solution really
preferable, and why or why not? ❑

In Listing 5.18, doesn’t flush_local_count_sig() need stronger memory barriers?
Answer:
No, that smp_store_release() suffices because this code communicates only with
flush_local_count(), and there is no need for store-to-load ordering. ❑

In Listing 5.18, why is it safe for line 25 to directly access the other thread’s
countermax variable?
Answer:
Because the other thread is not permitted to change the value of its countermax variable
unless it holds the gblcnt_mutex lock. But the caller has acquired this lock, so it is not
possible for the other thread to hold it, and therefore the other thread is not permitted to
change its countermax variable. We can therefore safely access it—but not change it.
❑

In Listing 5.18, why doesn’t line 30 check for the current thread sending itself a signal?
Answer:
There is no need for an additional check. The caller of flush_local_count() has
already invoked globalize_count(), so the check on line 25 will have succeeded,
skipping the later pthread_kill(). ❑

The code shown in Listings 5.17 and 5.18 works with GCC and POSIX. What would
be required to make it also conform to the ISO C standard?
Answer:
The theft variable must be of type sig_atomic_t to guarantee that it can be safely
shared between the signal handler and the code interrupted by the signal. ❑
v2023.06.11a
E.5. COUNTING 747

In Listing 5.18, why does line 39 resend the signal?
Answer:
Because many operating systems over several decades have had the property of losing
the occasional signal. Whether this is a feature or a bug is debatable, but irrelevant. The
obvious symptom from the user’s viewpoint will not be a kernel bug, but rather a user
application hanging.
Your user application hanging! ❑

Not only are POSIX signals slow, sending one to each thread simply does not scale.
What would you do if you had (say) 10,000 threads and needed the read side to be
fast?
Answer:
One approach is to use the techniques shown in Section 5.2.4, summarizing an
approximation to the overall counter value in a single variable. Another approach would
be to use multiple threads to carry out the reads, with each such thread interacting with
a specific subset of the updating threads. ❑

What if you want an exact limit counter to be exact only for its lower limit, but to allow
the upper limit to be inexact?
Answer:
One simple solution is to overstate the upper limit by the desired amount. The limiting
case of such overstatement results in the upper limit being set to the largest value that
the counter is capable of representing. ❑

What else had you better have done when using a biased counter?
Answer:
You had better have set the upper limit to be large enough accommodate the bias, the
expected maximum number of accesses, and enough “slop” to allow the counter to work
efficiently even when the number of accesses is at its maximum. ❑

This is ridiculous! We are read-acquiring a reader-writer lock to update the counter?
What are you playing at???
Answer:
Strange, perhaps, but true! Almost enough to make you think that the name “reader-writer
lock” was poorly chosen, isn’t it? ❑
v2023.06.11a

What other issues would need to be accounted for in a real system?
Answer:
A huge number!
Here are a few to start with:
1. There could be any number of devices, so that the global variables are inappropriate,
as are the lack of arguments to functions like do_io().
2. Polling loops can be problematic in real systems, wasting CPU time and energy.
In many cases, an event-driven design is far better, for example, where the last
completing I/O wakes up the device-removal thread.
3. The I/O might fail, and so do_io() will likely need a return value.
4. If the device fails, the last I/O might never complete. In such cases, there might
need to be some sort of timeout to allow error recovery.
5. Both add_count() and sub_count() can fail, but their return values are not
checked.
6. Reader-writer locks do not scale well. One way of avoiding the high read-acquisition
costs of reader-writer locks is presented in Chapters 7 and 9. ❑

On the count_stat.c row of Table 5.1, we see that the read-side scales linearly with
the number of threads. How is that possible given that the more threads there are, the
more per-thread counters must be summed up?
Answer:
The read-side code must scan the entire fixed-size array, regardless of the number of
threads, so there is no difference in performance. In contrast, in the last two algorithms,
readers must do more work when there are more threads. In addition, the last two
algorithms interpose an additional level of indirection because they map from integer
thread ID to the corresponding _Thread_local variable. ❑

Even on the fourth row of Table 5.1, the read-side performance of these statistical
counter implementations is pretty horrible. So why bother with them?
Answer:
“Use the right tool for the job.”
As can be seen from Figure 5.1, single-variable atomic increment need not apply for
any job involving heavy use of parallel updates. In contrast, the algorithms shown in
the top half of Table 5.1 do an excellent job of handling update-heavy situations. Of
course, if you have a read-mostly situation, you should use something else, for example,
an eventually consistent design featuring a single atomically incremented variable that
can be read out using a single load, similar to the approach used in Section 5.2.4. ❑
v2023.06.11a
E.5. COUNTING 749

Given the performance data shown in the bottom half of Table 5.1, we should always
prefer signals over atomic operations, right?
Answer:
That depends on the workload. Note that on a 64-core system, you need more than one
hundred non-atomic operations (with roughly a 40-nanosecond performance gain) to
make up for even one signal (with almost a 5-microsecond performance loss). Although
there are no shortage of workloads with far greater read intensity, you will need to
consider your particular workload.
In addition, although memory barriers have historically been expensive compared
to ordinary instructions, you should check this on the specific hardware you will be
running. The properties of computer hardware do change over time, and algorithms
must change accordingly. ❑

Can advanced techniques be applied to address the lock contention for readers seen in
the bottom half of Table 5.1?
Answer:
One approach is to give up some update-side performance, as is done with scalable
non-zero indicators (SNZI) [ELLM07]. There are a number of other ways one might
go about this, and these are left as exercises for the reader. Any number of approaches
that apply hierarchy, which replace frequent global-lock acquisitions with local lock
acquisitions corresponding to lower levels of the hierarchy, should work quite well. ❑

The ++ operator works just fine for 1,000-digit numbers! Haven’t you heard of operator
overloading???
Answer:
In the C++ language, you might well be able to use ++ on a 1,000-digit number, assuming
that you had access to a class implementing such numbers. But as of 2021, the C
language does not permit operator overloading. ❑

But if we are going to have to partition everything, why bother with shared-memory
multithreading? Why not just partition the problem completely and run as multiple
processes, each in its own address space?
Answer:
Indeed, multiple processes with separate address spaces can be an excellent way to
exploit parallelism, as the proponents of the fork-join methodology and the Erlang
language would be very quick to tell you. However, there are also some advantages to
shared-memory parallelism:
1. Only the most performance-critical portions of the application must be partitioned,

and such portions are usually a small fraction of the application.
v2023.06.11a
P1
P5 P2
P4 P3
Figure E.2: Dining Philosophers Problem, Fully Partitioned
2. Although cache misses are quite slow compared to individual register-to-register in-
structions, they are typically considerably faster than inter-process-communication
primitives, which in turn are considerably faster than things like TCP/IP networking.
3. Shared-memory multiprocessors are readily available and quite inexpensive, so, in
stark contrast to the 1990s, there is little cost penalty for use of shared-memory
parallelism.
As always, use the right tool for the job! ❑
E.6 Partitioning and Synchronization Design

Is there a better solution to the Dining Philosophers Problem?
Answer:
One such improved solution is shown in Figure E.2, where the philosophers are simply
provided with an additional five forks. All five philosophers may now eat simultaneously,
and there is never any need for philosophers to wait on one another. In addition, this
approach offers greatly improved disease control.
This solution might seem like cheating to some, but such “cheating” is key to finding
good solutions to many concurrency problems, as any hungry philosopher would agree.
And this is one solution to the Dining Philosophers concurrent-consumption problem
called out on page 113. ❑

How would you valididate an algorithm alleged to solve the Dining Philosophers
Problem?
Answer:
Much depends on the details of the algorithm, but here are a couple of places to start.
v2023.06.11a
E.6. PARTITIONING AND SYNCHRONIZATION DESIGN 751
First, for algorithms in which picking up left-hand and right-hand forks are separate
operations, start with all forks on the table. Then have all philosophers attempt to pick
up their first fork. Once all philosophers either have their first fork or are waiting for
someone to put down their first fork, have each non-waiting philosopher pick up their
second fork. At this point in any starvation-free solution, at least one philosopher will
be eating. If there were any waiting philosophers, repeat this test, preferably imposing
random variations in timing.
Second, create a stress test in which philosphers start and stop eating at random times.
Generate starvation and fairness conditions and verify that these conditions are met.
Here are a couple of example starvation and fairness conditions:
1. If all other philosophers have stopped eating 𝑁 times since a given philosopher
attempted to pick up a given fork, that philosopher should have succeeded in picking
up that fork. For high-quality solutions using high-quality locking primitives (or
high-quality atomic operations), 𝑁 = 1 is doable.
2. Given an upper bound 𝑇 on the time any philosopher holds onto both forks
before putting them down, the maximum waiting time for any philosopher should
be bounded by 𝑁𝑇 for some 𝑁 that is not hugely larger than the number of
philosophers.
3. Generate some statistic representing the time from when philosophers attempt to
pick up their first fork to the time when they start eating. The smaller this statistic,
the better the solution. Mean, median, and maximum are all useful statistics, but
examining the full distribution can also be enlightening.
Readers are encouraged to actually try testing any of the solutions presented in this
book, and especially testing solutions of their own devising. ❑

And in just what sense can this “horizontal parallelism” be said to be “horizontal”?
Answer:
Inman was working with protocol stacks, which are normally depicted vertically, with
the application on top and the hardware interconnect on the bottom. Data flows up and
down this stack. “Horizontal parallelism” processes packets from different network
connections in parallel, while “vertical parallelism” handles different protocol-processing
steps for a given packet in parallel.
“Vertical parallelism” is also called “pipelining”. ❑

In this compound double-ended queue implementation, what should be done if the
queue has become non-empty while releasing and reacquiring the lock?
Answer:
In this case, simply dequeue an item from the non-empty queue, release both locks, and
return. ❑
v2023.06.11a

Is the hashed double-ended queue a good solution? Why or why not?
Answer:
The best way to answer this is to run lockhdeq.c on a number of different multiprocessor
systems, and you are encouraged to do so in the strongest possible terms. One reason
for concern is that each operation on this implementation must acquire not one but two
locks.
The first well-designed performance study will be cited.6 Do not forget to compare
to a sequential implementation! ❑

Move all the elements to the queue that became empty? In what possible universe is
this brain-dead solution in any way optimal???
Answer:
It is optimal in the case where data flow switches direction only rarely. It would of
course be an extremely poor choice if the double-ended queue was being emptied from
both ends concurrently. This of course raises another question, namely, in what possible
universe emptying from both ends concurrently would be a reasonable thing to do.
Work-stealing queues are one possible answer to this question. ❑

Why can’t the compound parallel double-ended queue implementation be symmetric?
Answer:
The need to avoid deadlock by imposing a lock hierarchy forces the asymmetry, just
as it does in the fork-numbering solution to the Dining Philosophers Problem (see
Section 6.1.1). ❑

Why is it necessary to retry the right-dequeue operation on line 28 of Listing 6.3?
Answer:
This retry is necessary because some other thread might have enqueued an element
between the time that this thread dropped d->rlock on line 25 and the time that it
reacquired this same lock on line 27. ❑

Surely the left-hand lock must sometimes be available!!! So why is it necessary that
line 25 of Listing 6.3 unconditionally release the right-hand lock?
Answer:
It would be possible to use spin_trylock() to attempt to acquire the left-hand lock
when it was available. However, the failure case would still need to drop the right-hand
6 The studies by Dalessandro et al. [DCW+ 11] and Dice et al. [DLM+ 10] are excellent
starting points.
v2023.06.11a
lock and then re-acquire the two locks in order. Making this transformation (and
determining whether or not it is worthwhile) is left as an exercise for the reader. ❑

But in the case where data is flowing in only one direction, the algorithm shown in
Listing 6.3 will have both ends attempting to acquire the same lock whenever the
consuming end empties its underlying double-ended queue. Doesn’t that mean that
sometimes this algorithm fails to provide concurrent access to both ends of the queue
even when the queue contains an arbitrarily large number of elements?
Answer:
Indeed it does!
But the same is true of other algorithms claiming this property. For example, in
solutions using software transactional memory mechanisms based on hashed arrays of
locks, the leftmost and rightmost elements’ addresses will sometimes happen to hash to
the same lock. These hash collisions will also prevent concurrent access. For another
example, solutions using hardware transactional memory mechanisms with software
fallbacks [YHLR13, Mer11, JSG12] often use locking within those software fallbacks,
and thus suffer (albeit hopefully rarely) from whatever concurrency limitations that
these locking solutions suffer from.
Therefore, as of 2021, all practical solutions to the concurrent double-ended queue
problem fail to provide full concurrency in at least some circumstances, including the
compound double-ended queue. ❑

Why are there not one but two solutions to the double-ended queue problem?
Answer:
There are actually at least three. The third, by Dominik Dingel, makes interesting use of
reader-writer locking, and may be found in lockrwdeq.c.
And so there is not one, but rather three solutions to the lock-based double-ended
queue problem on page 113! ❑

The tandem double-ended queue runs about twice as fast as the hashed double-ended
queue, even when I increase the size of the hash table to an insanely large number.
Why is that?
Answer:
The hashed double-ended queue’s locking design only permits one thread at a time at
each end, and further requires two lock acquisitions for each operation. The tandem
double-ended queue also permits one thread at a time at each end, and in the common case
requires only one lock acquisition per operation. Therefore, the tandem double-ended
queue should be expected to outperform the hashed double-ended queue.
Can you create a double-ended queue that allows multiple concurrent operations at
each end? If so, how? If not, why not? ❑
v2023.06.11a

Is there a significantly better way of handling concurrency for double-ended queues?
Answer:
One approach is to transform the problem to be solved so that multiple double-ended
queues can be used in parallel, allowing the simpler single-lock double-ended queue to
be used, and perhaps also replace each double-ended queue with a pair of conventional
single-ended queues. Without such “horizontal scaling”, the speedup is limited to 2.0. In
contrast, horizontal-scaling designs can achieve very large speedups, and are especially
attractive if there are multiple threads working either end of the queue, because in this
multiple-thread case the dequeue simply cannot provide strong ordering guarantees.
After all, the fact that a given thread removed an item first in no way implies that it will
process that item first [HKLP12]. And if there are no guarantees, we may as well obtain
the performance benefits that come with refusing to provide these guarantees.
Regardless of whether or not the problem can be transformed to use multiple queues, it
is worth asking whether work can be batched so that each enqueue and dequeue operation
corresponds to larger units of work. This batching approach decreases contention on
the queue data structures, which increases both performance and scalability, as will be
seen in Section 6.3. After all, if you must incur high synchronization overheads, be sure
you are getting your money’s worth.
Other researchers are working on other ways to take advantage of limited ordering
guarantees in queues [KLP12]. ❑

Don’t all these problems with critical sections mean that we should just always use
non-blocking synchronization [Her90], which don’t have critical sections?
Answer:
Although non-blocking synchronization can be very useful in some situations, it is
no panacea, as discussed in Section 14.2. Also, non-blocking synchronization really
does have critical sections, as noted by Josh Triplett. For example, in a non-blocking
algorithm based on compare-and-swap operations, the code starting at the initial load
and continuing to the compare-and-swap is analogous to a lock-based critical section. ❑

What should you do to validate a hash table?
Answer:
Quite a bit, actually.
See Section 10.3.2 for a good starting point. ❑

“Partitioning time”? Isn’t that an odd turn of phrase?
Answer:
Perhaps so.
v2023.06.11a
But in the next section we will be partitioning space (that is, address space) as well
as time. This nomenclature will permit us to partition spacetime, as opposed to (say)
partitioning space but segmenting time. ❑

What are some ways of preventing a structure from being freed while its lock is being
acquired?
Answer:
Here are a few possible solutions to this existence guarantee problem:
1. Provide a statically allocated lock that is held while the per-structure lock is being
acquired, which is an example of hierarchical locking (see Section 6.4.2). Of
course, using a single global lock for this purpose can result in unacceptably high
levels of lock contention, dramatically reducing performance and scalability.
2. Provide an array of statically allocated locks, hashing the structure’s address to

select the lock to be acquired, as described in Chapter 7. Given a hash function of
sufficiently high quality, this avoids the scalability limitations of the single global
lock, but in read-mostly situations, the lock-acquisition overhead can result in
unacceptably degraded performance.
3. Use a garbage collector, in software environments providing them, so that a structure

cannot be deallocated while being referenced. This works very well, removing the
existence-guarantee burden (and much else besides) from the developer’s shoulders,
but imposes the overhead of garbage collection on the program. Although garbage-
collection technology has advanced considerably in the past few decades, its
overhead may be unacceptably high for some applications. In addition, some
applications require that the developer exercise more control over the layout
and placement of data structures than is permitted by most garbage collected
environments.
4. As a special case of a garbage collector, use a global reference counter, or a global

array of reference counters. These have strengths and limitations similar to those
called out above for locks.
5. Use hazard pointers [Mic04a], which can be thought of as an inside-out reference

count. Hazard-pointer-based algorithms maintain a per-thread list of pointers, so
that the appearance of a given pointer on any of these lists acts as a reference to the
corresponding structure. Hazard pointers are starting to see significant production
use (see Section 9.6.3.1).
6. Use transactional memory (TM) [HM93, Lom77, ST95], so that each reference and
modification to the data structure in question is performed atomically. Although TM
has engendered much excitement in recent years, and seems likely to be of some use
in production software, developers should exercise some caution [BLM05, BLM06,
MMW07], particularly in performance-critical code. In particular, existence
guarantees require that the transaction covers the full path from a global reference
to the data elements being updated. For more on TM, including ways to overcome
some of its weaknesses by combining it with other synchronization mechanisms,
see Sections 17.2 and 17.3.
v2023.06.11a
7. Use RCU, which can be thought of as an extremely lightweight approximation

to a garbage collector. Updaters are not permitted to free RCU-protected data
structures that RCU readers might still be referencing. RCU is most heavily used
for read-mostly data structures, and is discussed at length in Section 9.5.
For more on providing existence guarantees, see Chapters 7 and 9. ❑

But won’t system boot and shutdown (or application startup and shutdown) be
partitioning time, even for data ownership?
Answer:
You can indeed think in these terms.
And if you are working on a persistent data store where state survives shutdown,
thinking in these terms might even be useful. ❑

How can a single-threaded 64-by-64 matrix multiple possibly have an efficiency of
less than 1.0? Shouldn’t all of the traces in Figure 6.17 have efficiency of exactly 1.0
when running on one thread?
Answer:
The matmul.c program creates the specified number of worker threads, so even the
single-worker-thread case incurs thread-creation overhead. Making the changes required
to optimize away thread-creation overhead in the single-worker-thread case is left as an
exercise to the reader. ❑

How are data-parallel techniques going to help with matrix multiply? It is already
data parallel!!!
Answer:
I am glad that you are paying attention! This example serves to show that although
data parallelism can be a very good thing, it is not some magic wand that automatically
wards off any and all sources of inefficiency. Linear scaling at full performance, even to
“only” 64 threads, requires care at all phases of design and implementation.
In particular, you need to pay careful attention to the size of the partitions. For
example, if you split a 64-by-64 matrix multiply across 64 threads, each thread gets
only 64 floating-point multiplies. The cost of a floating-point multiply is minuscule
compared to the overhead of thread creation, and cache-miss overhead also plays a role
in spoiling the theoretically perfect scalability (and also in making the traces so jagged).
The full 448 hardware threads would require a matrix with hundreds of thousands of
rows and columns to attain good scalability, but by that point GPGPUs become quite
attractive, especially from a price/performance viewpoint.
Moral: If you have a parallel program with variable input, always include a check for
the input size being too small to be worth parallelizing. And when it is not helpful to
parallelize, it is not helpful to incur the overhead required to spawn a thread, now is it?
❑
v2023.06.11a

What did you do to validate this matrix multiply algorithm?
Answer:
For this simple approach, very little.
However, the validation of production-quality matrix multiply requires great care
and attention. Some cases require careful handling of floating-point rounding errors,
others involve complex sparse-matrix data structures, and still others make use of
special-purpose arithmetic hardware such as vector units or GPGPUs. Adequate tests
for handling of floating-point rounding errors can be especially challenging. ❑

In what situation would hierarchical locking work well?
Answer:
If the comparison on line 31 of Listing 6.8 were replaced by a much heavier-weight
operation, then releasing bp->bucket_lock might reduce lock contention enough to
outweigh the overhead of the extra acquisition and release of cur->node_lock. ❑

Doesn’t this resource-allocator design resemble that of the approximate limit counters
covered in Section 5.3?
Answer:
Indeed it does! We are used to thinking of allocating and freeing memory, but the
algorithms in Section 5.3 are taking very similar actions to allocate and free “count”. ❑

In Figure 6.21, there is a pattern of performance rising with increasing run length in
groups of three samples, for example, for run lengths 10, 11, and 12. Why?
Answer:
This is due to the per-CPU target value being three. A run length of 12 must acquire the
global-pool lock twice, while a run length of 13 must acquire the global-pool lock three
times. ❑

Allocation failures were observed in the two-thread tests at run lengths of 19 and
greater. Given the global-pool size of 40 and the per-thread target pool size 𝑠 of three,
number of threads 𝑛 equal to two, and assuming that the per-thread pools are initially
empty with none of the memory in use, what is the smallest allocation run length 𝑚
at which failures can occur? (Recall that each thread repeatedly allocates 𝑚 block
of memory, and then frees the 𝑚 blocks of memory.) Alternatively, given 𝑛 threads
each with pool size 𝑠, and where each thread repeatedly first allocates 𝑚 blocks of
memory and then frees those 𝑚 blocks, how large must the global pool size be? Note:
Obtaining the correct answer will require you to examine the smpalloc.c source
code, and very likely single-step it as well. You have been warned!
v2023.06.11a
Global Pool g-i-p(n-1)
Per-Thread Pool i 0 p-m p-m
Per-Thread Allocation 0 0 m m
Figure E.3: Allocator Cache Run-Length Analysis
Answer:
This solution is adapted from one put forward by Alexey Roytman. It is based on the
following definitions:
𝑔 Number of blocks globally available.

𝑖 Number of blocks left in the initializing thread’s per-thread pool. (This is one reason
you needed to look at the code!)
𝑚 Allocation/free run length.
𝑛 Number of threads, excluding the initialization thread.
𝑝 Per-thread maximum block consumption, including both the blocks actually allocated
and the blocks remaining in the per-thread pool.
The values 𝑔, 𝑚, and 𝑛 are given. The value for 𝑝 is 𝑚 rounded up to the next multiple
of 𝑠, as follows:
l𝑚m
𝑝=𝑠 (E.6)
𝑠
The value for 𝑖 is as follows:

𝑔 (mod 2𝑠) = 0 : 2𝑠
𝑖= (E.7)
𝑔 (mod 2𝑠) ≠ 0 : 𝑔 (mod 2𝑠)
The relationships between these quantities are shown in Figure E.3. The global pool
is shown on the top of this figure, and the “extra” initializer thread’s per-thread pool
and per-thread allocations are the left-most pair of boxes. The initializer thread has no
blocks allocated, but has 𝑖 blocks stranded in its per-thread pool. The rightmost two
pairs of boxes are the per-thread pools and per-thread allocations of threads holding
the maximum possible number of blocks, while the second-from-left pair of boxes
represents the thread currently trying to allocate.
The total number of blocks is 𝑔, and adding up the per-thread allocations and per-
thread pools, we see that the global pool contains 𝑔 −𝑖 − 𝑝(𝑛 − 1) blocks. If the allocating
thread is to be successful, it needs at least 𝑚 blocks in the global pool, in other words:
𝑔 − 𝑖 − 𝑝(𝑛 − 1) ≥ 𝑚 (E.8)
The question has 𝑔 = 40, 𝑠 = 3, and 𝑛 = 2. Equation E.7 gives 𝑖 = 4, and Eq. E.6
gives 𝑝 = 18 for 𝑚 = 18 and 𝑝 = 21 for 𝑚 = 19. Plugging these into Eq. E.8 shows that
𝑚 = 18 will not overflow, but that 𝑚 = 19 might well do so.
v2023.06.11a
E.7. LOCKING 759
The presence of 𝑖 could be considered to be a bug. After all, why allocate memory
only to have it stranded in the initialization thread’s cache? One way of fixing this
would be to provide a memblock_flush() function that flushed the current thread’s
pool into the global pool. The initialization thread could then invoke this function after
freeing all of the blocks. ❑

Given that a 2D maze achieved 4x speedup on two CPUs, would a 3D maze achieve
an 8x speedup on two CPUs?
Answer:
This is an excellent question that is left to a suitably interested and industrious reader. ❑

Why place the third, fourth, and so on threads on the diagonal? Why not instead
distribute them evenly around the maze?
Answer:
There are indeed a great many ways to distribute the extra threads. Evaluation of
distribution strategies is left to a suitably interested and industrious reader. ❑
E.7 Locking

Just how can serving as a whipping boy be considered to be in any way honorable???
Answer:
The reason locking serves as a research-paper whipping boy is because it is heavily used
in practice. In contrast, if no one used or cared about locking, most research papers
would not bother even mentioning it. ❑

But the definition of lock-based deadlock only said that each thread was holding at
least one lock and waiting on another lock that was held by some thread. How do you
know that there is a cycle?
Answer:
Suppose that there is no cycle in the graph. We would then have a directed acyclic graph
(DAG), which would have at least one leaf node.
If this leaf node was a lock, then we would have a thread that was waiting on a lock
that wasn’t held by any thread, counter to the definition. In this case the thread would
immediately acquire the lock.
On the other hand, if this leaf node was a thread, then we would have a thread that
was not waiting on any lock, again counter to the definition. And in this case, the thread
would either be running or be blocked on something that is not a lock. In the first case,
in the absence of infinite-loop bugs, the thread will eventually release the lock. In the
v2023.06.11a
second case, in the absence of a failure-to-wake bug, the thread will eventually wake up
and release the lock.7
Therefore, given this definition of lock-based deadlock, there must be a cycle in the
corresponding graph. ❑

Are there any exceptions to this rule, so that there really could be a deadlock cycle
containing locks from both the library and the caller, even given that the library code
never invokes any of the caller’s functions?
Answer:
Indeed there are! Here are a few of them:
1. If one of the library function’s arguments is a pointer to a lock that this library
function acquires, and if the library function holds one of its locks while acquiring
the caller’s lock, then we could have a deadlock cycle involving both caller and
library locks.
2. If one of the library functions returns a pointer to a lock that is acquired by the
caller, and if the caller acquires one of its locks while holding the library’s lock,
we could again have a deadlock cycle involving both caller and library locks.
3. If one of the library functions acquires a lock and then returns while still holding
it, and if the caller acquires one of its locks, we have yet another way to create a
deadlock cycle involving both caller and library locks.
4. If the caller has a signal handler that acquires locks, then the deadlock cycle can
involve both caller and library locks. In this case, however, the library’s locks are
innocent bystanders in the deadlock cycle. That said, please note that acquiring a
lock from within a signal handler is a no-no in many environments—it is not just a
bad idea, it is unsupported. But if you absolutely must acquire a lock in a signal
handler, be sure to block that signal while holding that same lock in thread context,
and also while holding any other locks acquired while that same lock is held. ❑

But if qsort() releases all its locks before invoking the comparison function, how
can it protect against races with other qsort() threads?
Answer:
By privatizing the data elements being compared (as discussed in Chapter 8) or through
use of deferral mechanisms such as reference counting (as discussed in Chapter 9). Or
through use of layered locking hierarchies, as described in Section 7.1.1.3.
On the other hand, changing a key in a list that is currently being sorted is at best
rather brave. ❑
7 Of course, one type of failure-to-wake bug is a deadlock that involves not only locks,
but also non-lock resources. But the question really did say “lock-based deadlock”!
v2023.06.11a
E.7. LOCKING 761

So the iterating thread may or may not observe the added child. What is the big deal?
Answer:
There are at least two hazards in this situation.
One is indeed that the number of children may or may not be observed to have changed.
While that would be consistent with tree_add() being called either before or after
the iterator started, it is better not left to the vagaries of the compiler. A more serious
problem is that realloc() may not be able to extend the array in place, causing the
heap to free the one used by the iterator and replace it with another block of memory. If
the children pointer is not re-read then the iterating thread will access invalid memory
(either free or reclaimed). ❑

What do you mean “cannot always safely invoke the scheduler”? Either call_rcu()
can or cannot safely invoke the scheduler, right?
Answer:
It really does depend.
The scheduler locks are always held with interrupts disabled. Therefore, if call_
rcu() is invoked with interrupts enabled, no scheduler locks are held, and call_rcu()
can safely call into the scheduler. Otherwise, if interrupts are disabled, one of the
scheduler locks might be held, so call_rcu() must play it safe and refrain from calling
into the scheduler. ❑

Name one common situation where a pointer to a lock is passed into a function.
Answer:
Locking primitives, of course! ❑

Doesn’t the fact that pthread_cond_wait() first releases the mutex and then re-
acquires it eliminate the possibility of deadlock?
Answer:
Absolutely not!
Consider a program that acquires mutex_a, and then mutex_b, in that order, and
then passes mutex_a to pthread_cond_wait(). Now, pthread_cond_wait() will
release mutex_a, but will re-acquire it before returning. If some other thread acquires
mutex_a in the meantime and then blocks on mutex_b, the program will deadlock. ❑
v2023.06.11a

Can the transformation from Listing 7.4 to Listing 7.5 be applied universally?
Answer:
Absolutely not!
This transformation assumes that the layer_2_processing() function is idem-
potent, given that it might be executed multiple times on the same packet when the
layer_1() routing decision changes. Therefore, in real life, this transformation can
become arbitrarily complex. ❑

But the complexity in Listing 7.5 is well worthwhile given that it avoids deadlock,
right?
Answer:
Maybe.
If the routing decision in layer_1() changes often enough, the code will always
retry, never making forward progress. This is termed “livelock” if no thread makes any
forward progress or “starvation” if some threads make forward progress but others do
not (see Section 7.1.2). ❑

When using the “acquire needed locks first” approach described in Section 7.1.1.7,
how can livelock be avoided?
Answer:
Provide an additional global lock. If a given thread has repeatedly tried and failed to
acquire the needed locks, then have that thread unconditionally acquire the new global
lock, and then unconditionally acquire any needed locks. (Suggested by Doug Lea.) ❑

Suppose Lock A is never acquired within a signal handler, but Lock B is acquired both
from thread context and by signal handlers. Suppose further that Lock A is sometimes
acquired with signals unblocked. Why is it illegal to acquire Lock A holding Lock B?
Answer:
Because this would lead to deadlock. Given that Lock A is sometimes held outside of a
signal handler without blocking signals, a signal might be handled while holding this
lock. The corresponding signal handler might then acquire Lock B, so that Lock B is
acquired while holding Lock A. Therefore, if we also acquire Lock A while holding
Lock B, we will have a deadlock cycle. Note that this problem exists even if signals are
blocked while holding Lock B.
This is another reason to be very careful with locks that are acquired within interrupt
or signal handlers. But the Linux kernel’s lock dependency checker knows about this
situation and many others as well, so please do make full use of it! ❑
v2023.06.11a
E.7. LOCKING 763

How can you legally block signals within a signal handler?
Answer:
One of the simplest and fastest ways to do so is to use the sa_mask field of the
struct sigaction that you pass to sigaction() when setting up the signal. ❑

If acquiring locks in signal handlers is such a bad idea, why even discuss ways of
making it safe?
Answer:
Because these same rules apply to the interrupt handlers used in operating-system
kernels and in some embedded applications.
In many application environments, acquiring locks in signal handlers is frowned
upon [Ope97]. However, that does not stop clever developers from (perhaps unwisely)
fashioning home-brew locks out of atomic operations. And atomic operations are in
many cases perfectly legal in signal handlers. ❑

Given an object-oriented application that passes control freely among a group of
objects such that there is no straightforward locking hierarchy,a layered or otherwise,
how can this application be parallelized?
a Also known as “object-oriented spaghetti code.”
Answer:
There are a number of approaches:
1. In the case of parametric search via simulation, where a large number of simulations
will be run in order to converge on (for example) a good design for a mechanical or
electrical device, leave the simulation single-threaded, but run many instances of the
simulation in parallel. This retains the object-oriented design, and gains parallelism
at a higher level, and likely also avoids both deadlocks and synchronization overhead.
2. Partition the objects into groups such that there is no need to operate on objects in
more than one group at a given time. Then associate a lock with each group. This
is an example of a single-lock-at-a-time design, which discussed in Section 7.1.1.8.
3. Partition the objects into groups such that threads can all operate on objects in the
groups in some groupwise ordering. Then associate a lock with each group, and
impose a locking hierarchy over the groups.
4. Impose an arbitrarily selected hierarchy on the locks, and then use conditional
locking if it is necessary to acquire a lock out of order, as was discussed in
Section 7.1.1.6.
5. Before carrying out a given group of operations, predict which locks will be
acquired, and attempt to acquire them before actually carrying out any updates. If
the prediction turns out to be incorrect, drop all the locks and retry with an updated
prediction that includes the benefit of experience. This approach was discussed in
Section 7.1.1.7.
v2023.06.11a
6. Use transactional memory. This approach has a number of advantages and

disadvantages which will be discussed in Sections 17.2–17.3.
7. Refactor the application to be more concurrency-friendly. This would likely also
have the side effect of making the application run faster even when single-threaded,
but might also make it more difficult to modify the application.
8. Use techniques from later chapters in addition to locking. ❑

How can the livelock shown in Listing 7.6 be avoided?
Answer:
Listing 7.5 provides some good hints. In many cases, livelocks are a hint that you
should revisit your locking design. Or visit it in the first place if your locking design
“just grew”.
That said, one good-and-sufficient approach due to Doug Lea is to use conditional
locking as described in Section 7.1.1.6, but combine this with acquiring all needed
locks first, before modifying shared data, as described in Section 7.1.1.7. If a given
critical section retries too many times, unconditionally acquire a global lock, then
unconditionally acquire all the needed locks. This avoids both deadlock and livelock,
and scales reasonably assuming that the global lock need not be acquired too often. ❑

What problems can you spot in the code in Listing 7.7?
Answer:
Here are a couple:
1. A one-second wait is way too long for most uses. Wait intervals should begin with
roughly the time required to execute the critical section, which will normally be in
the microsecond or millisecond range.
2. The code does not check for overflow. On the other hand, this bug is nullified by
the previous bug: 32 bits worth of seconds is more than 50 years. ❑

Wouldn’t it be better just to use a good parallel design so that lock contention was low
enough to avoid unfairness?
Answer:
It would be better in some sense, but there are situations where it can be appropriate to
use designs that sometimes result in high lock contentions.
For example, imagine a system that is subject to a rare error condition. It might
well be best to have a simple error-handling design that has poor performance and
scalability for the duration of the rare error condition, as opposed to a complex and
difficult-to-debug design that is helpful only when one of those rare error conditions is
in effect.
That said, it is usually worth putting some effort into attempting to produce a design
that both simple as well as efficient during error conditions, for example by partitioning
the problem. ❑
v2023.06.11a
E.7. LOCKING 765

How might the lock holder be interfered with?
Answer:
If the data protected by the lock is in the same cache line as the lock itself, then attempts
by other CPUs to acquire the lock will result in expensive cache misses on the part of
the CPU holding the lock. This is a special case of false sharing, which can also occur if
a pair of variables protected by different locks happen to share a cache line. In contrast,
if the lock is in a different cache line than the data that it protects, the CPU holding the
lock will usually suffer a cache miss only on first access to a given variable.
Of course, the downside of placing the lock and data into separate cache lines is that
the code will incur two cache misses rather than only one in the uncontended case. As
always, choose wisely! ❑

Does it ever make sense to have an exclusive lock acquisition immediately followed by
a release of that same lock, that is, an empty critical section?
Answer:
Empty lock-based critical sections are rarely used, but they do have their uses. The
point is that the semantics of exclusive locks have two components: (1) The familiar
data-protection semantic and (2) A messaging semantic, where releasing a given lock
notifies a waiting acquisition of that same lock. An empty critical section uses the
messaging component without the data-protection component.
The rest of this answer provides some example uses of empty critical sections,
however, these examples should be considered “gray magic.”8 As such, empty critical
sections are almost never used in practice. Nevertheless, pressing on into this gray area
...
One historical use of empty critical sections appeared in the networking stack of the
2.4 Linux kernel through use of a read-side-scalable reader-writer lock called brlock
for “big reader lock”. This use case is a way of approximating the semantics of read-copy
update (RCU), which is discussed in Section 9.5. And in fact this Linux-kernel use case
has been replaced with RCU.
The empty-lock-critical-section idiom can also be used to reduce lock contention in
some situations. For example, consider a multithreaded user-space application where
each thread processes units of work maintained in a per-thread list, where threads are
prohibited from touching each others’ lists [McK12e]. There could also be updates that
require that all previously scheduled units of work have completed before the update
can progress. One way to handle this is to schedule a unit of work on each thread, so
that when all of these units of work complete, the update may proceed.
In some applications, threads can come and go. For example, each thread might
correspond to one user of the application, and thus be removed when that user logs
out or otherwise disconnects. In many applications, threads cannot depart atomically:
They must instead explicitly unravel themselves from various portions of the application
using a specific sequence of actions. One specific action will be refusing to accept
further requests from other threads, and another specific action will be disposing of any
remaining units of work on its list, for example, by placing these units of work in a
8 Thanks to Alexey Roytman for this description.
v2023.06.11a
global work-item-disposal list to be taken by one of the remaining threads. (Why not
just drain the thread’s work-item list by executing each item? Because a given work
item might generate more work items, so that the list could not be drained in a timely
fashion.)
If the application is to perform and scale well, a good locking design is required.
One common solution is to have a global lock (call it G) protecting the entire process
of departing (and perhaps other things as well), with finer-grained locks protecting the
individual unraveling operations.
Now, a departing thread must clearly refuse to accept further requests before disposing
of the work on its list, because otherwise additional work might arrive after the disposal
action, which would render that disposal action ineffective. So simplified pseudocode
for a departing thread might be as follows:
1. Acquire lock G.
2. Acquire the lock guarding communications.
3. Refuse further communications from other threads.
4. Release the lock guarding communications.
5. Acquire the lock guarding the global work-item-disposal list.
6. Move all pending work items to the global work-item-disposal list.
7. Release the lock guarding the global work-item-disposal list.
8. Release lock G.
Of course, a thread that needs to wait for all pre-existing work items will need to take
departing threads into account. To see this, suppose that this thread starts waiting for all
pre-existing work items just after a departing thread has refused further communications
from other threads. How can this thread wait for the departing thread’s work items to
complete, keeping in mind that threads are not allowed to access each others’ lists of
work items?
One straightforward approach is for this thread to acquire G and then the lock guarding
the global work-item-disposal list, then move the work items to its own list. The thread
then release both locks, places a work item on the end of its own list, and then wait for
all of the work items that it placed on each thread’s list (including its own) to complete.
This approach does work well in many cases, but if special processing is required
for each work item as it is pulled in from the global work-item-disposal list, the result
could be excessive contention on G. One way to avoid that contention is to acquire G
and then immediately release it. Then the process of waiting for all prior work items
look something like the following:
1. Set a global counter to one and initialize a condition variable to zero.

2. Send a message to all threads to cause them to atomically increment the global
counter, and then to enqueue a work item. The work item will atomically decrement
the global counter, and if the result is zero, it will set a condition variable to one.
3. Acquire G, which will wait on any currently departing thread to finish departing.
Because only one thread may depart at a time, all the remaining threads will have
already received the message sent in the preceding step.
v2023.06.11a
E.7. LOCKING 767
4. Release G.
5. Acquire the lock guarding the global work-item-disposal list.
6. Move all work items from the global work-item-disposal list to this thread’s list,
processing them as needed along the way.
7. Release the lock guarding the global work-item-disposal list.
8. Enqueue an additional work item onto this thread’s list. (As before, this work item
will atomically decrement the global counter, and if the result is zero, it will set a
condition variable to one.)
9. Wait for the condition variable to take on the value one.
Once this procedure completes, all pre-existing work items are guaranteed to have
completed. The empty critical sections are using locking for messaging as well as for
protection of data. ❑

Is there any other way for the VAX/VMS DLM to emulate a reader-writer lock?
Answer:
There are in fact several. One way would be to use the null, protected-read, and exclusive
modes. Another way would be to use the null, protected-read, and concurrent-write
modes. A third way would be to use the null, concurrent-read, and exclusive modes. ❑

The code in Listing 7.8 is ridiculously complicated! Why not conditionally acquire a
single global lock?
Answer:
Conditionally acquiring a single global lock does work very well, but only for relatively
small numbers of CPUs. To see why it is problematic in systems with many hundreds
of CPUs, look at Figure 5.1. ❑

Wait a minute! If we “win” the tournament on line 16 of Listing 7.8, we get to do all
the work of do_force_quiescent_state(). Exactly how is that a win, really?
Answer:
How indeed? This just shows that in concurrency, just as in life, one should take care to
learn exactly what winning entails before playing the game. ❑

Why not rely on the C language’s default initialization of zero instead of using the
explicit initializer shown on line 2 of Listing 7.9?
Answer:
Because this default initialization does not apply to locks allocated as auto variables
within the scope of a function. ❑
v2023.06.11a

Why bother with the inner loop on lines 7–8 of Listing 7.9? Why not simply repeatedly
do the atomic exchange operation on line 6?
Answer:
Suppose that the lock is held and that several threads are attempting to acquire the lock.
In this situation, if these threads all loop on the atomic exchange operation, they will
ping-pong the cache line containing the lock among themselves, imposing load on the
interconnect. In contrast, if these threads are spinning in the inner loop on lines 7–8,
they will each spin within their own caches, placing negligible load on the interconnect.
❑

Why not simply store zero into the lock word on line 14 of Listing 7.9?
Answer:
This can be a legitimate implementation, but only if this store is preceded by a memory
barrier and makes use of WRITE_ONCE(). The memory barrier is not required when the
xchg() operation is used because this operation implies a full memory barrier due to
the fact that it returns a value. ❑

How can you tell if one counter is greater than another, while accounting for counter
wrap?
Answer:
In the C language, the following macro correctly handles this:
#define ULONG_CMP_LT(a, b) \
(ULONG_MAX / 2 < (a) - (b))
Although it is tempting to simply subtract two signed integers, this should be avoided
because signed overflow is undefined in the C language. For example, if the compiler
knows that one of the values is positive and the other negative, it is within its rights
to simply assume that the positive number is greater than the negative number, even
though subtracting the negative number from the positive number might well result in
overflow and thus a negative number.
How could the compiler know the signs of the two numbers? It might be able to
deduce it based on prior assignments and comparisons. In this case, if the per-CPU
counters were signed, the compiler could deduce that they were always increasing in
value, and then might assume that they would never go negative. This assumption could
well lead the compiler to generate unfortunate code [McK12d, Reg10]. ❑

Which is better, the counter approach or the flag approach?
Answer:
The flag approach will normally suffer fewer cache misses, but a better answer is to try
both and see which works best for your particular workload. ❑
v2023.06.11a
E.8. DATA OWNERSHIP 769

How can relying on implicit existence guarantees result in a bug?
Answer:
Here are some bugs resulting from improper use of implicit existence guarantees:
1. A program writes the address of a global variable to a file, then a later instance
of that same program reads that address and attempts to dereference it. This can
fail due to address-space randomization, to say nothing of recompilation of the
program.
2. A module can record the address of one of its variables in a pointer located in some
other module, then attempt to dereference that pointer after the module has been
unloaded.
3. A function can record the address of one of its on-stack variables into a global
pointer, which some other function might attempt to dereference after that function
has returned.
I am sure that you can come up with additional possibilities. ❑

What if the element we need to delete is not the first element of the list on line 8 of
Listing 7.10?
Answer:
This is a very simple hash table with no chaining, so the only element in a given bucket
is the first element. The reader is invited to adapt this example to a hash table with full
chaining. ❑
E.8 Data Ownership

What form of data ownership is extremely difficult to avoid when creating shared-
memory parallel programs (for example, using pthreads) in C or C++?
Answer:
Use of auto variables in functions. By default, these are private to the thread executing
the current function. ❑

What synchronization remains in the example shown in Section 8.1?
Answer:
The creation of the threads via the sh & operator and the joining of thread via the sh
wait command.
Of course, if the processes explicitly share memory, for example, using the shmget()
or mmap() system calls, explicit synchronization might well be needed when acccessing
or updating the shared memory. The processes might also synchronize using any of the
following interprocess communications mechanisms:
v2023.06.11a
1. System V semaphores.
2. System V message queues.
3. UNIX-domain sockets.
4. Networking protocols, including TCP/IP, UDP, and a whole host of others.
5. File locking.
6. Use of the open() system call with the O_CREAT and O_EXCL flags.
7. Use of the rename() system call.
A complete list of possible synchronization mechanisms is left as an exercise to the

reader, who is warned that it will be an extremely long list. A surprising number of
unassuming system calls can be pressed into service as synchronization mechanisms. ❑

Is there any shared data in the example shown in Section 8.1?
Answer:
That is a philosophical question.
Those wishing the answer “no” might argue that processes by definition do not share
memory.
Those wishing to answer “yes” might list a large number of synchronization mecha-
nisms that do not require shared memory, note that the kernel will have some shared
state, and perhaps even argue that the assignment of process IDs (PIDs) constitute
shared data.
Such arguments are excellent intellectual exercise, and are also a wonderful way of
feeling intelligent and scoring points against hapless classmates or colleagues, but are
mostly a way of avoiding getting anything useful done. ❑

Does it ever make sense to have partial data ownership where each thread reads only
its own instance of a per-thread variable, but writes to other threads’ instances?
Answer:
Amazingly enough, yes. One example is a simple message-passing system where threads
post messages to other threads’ mailboxes, and where each thread is responsible for
removing any message it sent once that message has been acted on. Implementation of
such an algorithm is left as an exercise for the reader, as is identifying other algorithms
with similar ownership patterns. ❑
v2023.06.11a
E.8. DATA OWNERSHIP 771

What mechanisms other than POSIX signals may be used for function shipping?
Answer:
There is a very large number of such mechanisms, including:
1. System V message queues.
2. Shared-memory dequeue (see Section 6.1.2).

3. Shared-memory mailboxes.
4. UNIX-domain sockets.
5. TCP/IP or UDP, possibly augmented by any number of higher-level protocols,
including RPC, HTTP, XML, SOAP, and so on.
Compilation of a complete list is left as an exercise to sufficiently single-minded readers,
who are warned that the list will be extremely long. ❑

But none of the data in the eventual() function shown on lines 17–32 of Listing 5.5 is
actually owned by the eventual() thread! In just what way is this data ownership???
Answer:
The key phrase is “owns the rights to the data”. In this case, the rights in question are
the rights to access the per-thread counter variable defined on line 1 of the listing.
This situation is similar to that described in Section 8.2.
However, there really is data that is owned by the eventual() thread, namely the t
and sum variables defined on lines 19 and 20 of the listing.
For other examples of designated threads, look at the kernel threads in the Linux
kernel, for example, those created by kthread_create() and kthread_run(). ❑

Is it possible to obtain greater accuracy while still maintaining full privacy of the
per-thread data?
Answer:
Yes. One approach is for read_count() to add the value of its own per-thread variable.
This maintains full ownership and performance, but only a slight improvement in
accuracy, particularly on systems with very large numbers of threads.
Another approach is for read_count() to use function shipping, for example, in
the form of per-thread signals. This greatly improves accuracy, but at a significant
performance cost for read_count().
However, both of these methods have the advantage of eliminating cache thrashing
for the common case of updating counters. ❑
v2023.06.11a
E.9 Deferred Processing

Why bother with a use-after-free check?
Answer:
To greatly increase the probability of finding bugs. A small torture-test program
(routetorture.h) that allocates and frees only one type of structure can tolerate a
surprisingly large amount of use-after-free misbehavior. See Figure 11.4 on page 339
and the related discussion in Section 11.6.4 starting on page 341 for more on the
importance of increasing the probability of finding bugs. ❑

Why doesn’t route_del() in Listing 9.3 use reference counts to protect the traversal
to the element to be freed?
Answer:
Because the traversal is already protected by the lock, so no additional protection is
required. ❑

Why the break in the “ideal” line at 224 CPUs in Figure 9.2? Shouldn’t it be a straight
line?
Answer:
The break is due to hyperthreading. On this particular system, the first hardware thread
in each core within a socket have consecutive CPU numbers, followed by the first
hardware threads in each core for the other sockets, and finally followed by the second
hardware thread in each core on all the sockets. On this particular system, CPU numbers
0–27 are the first hardware threads in each of the 28 cores in the first socket, numbers
28–55 are the first hardware threads in each of the 28 cores in the second socket, and so
on, so that numbers 196–223 are the first hardware threads in each of the 28 cores in the
eighth socket. Then CPU numbers 224–251 are the second hardware threads in each
of the 28 cores of the first socket, numbers 252–279 are the second hardware threads
in each of the 28 cores of the second socket, and so on until numbers 420–447 are the
second hardware threads in each of the 28 cores of the eighth socket.
Why does this matter?
Because the two hardware threads of a given core share resources, and this workload
seems to allow a single hardware thread to consume more than half of the relevant
resources within its core. Therefore, adding the second hardware thread of that core
adds less than one might hope. Other workloads might gain greater benefit from each
core’s second hardware thread, but much depends on the details of both the hardware
and the workload. ❑
v2023.06.11a
E.9. DEFERRED PROCESSING 773
8
1x10
7 ideal
1x10

1x106
100000
10000 refcnt
1000
100
1 10 100
Figure E.4: Pre-BSD Routing Table Protected by Reference Counting, Log Scale

Shouldn’t the refcnt trace in Figure 9.2 be at least a little bit off of the x-axis???
Answer:
Define “a little bit.”
Figure E.4 shows the same data, but on a log-log plot. As you can see, the refcnt line
drops below 5,000 at two CPUs. This means that the refcnt performance at two CPUs is
more than one thousand times smaller than the first y-axis tick of 5 × 106 in Figure 9.2.
Therefore, the depiction of the performance of reference counting shown in Figure 9.2
is all too accurate. ❑

If concurrency has “most definitely reduced the usefulness of reference counting”,
why are there so many reference counters in the Linux kernel?
Answer:
That sentence did say “reduced the usefulness”, not “eliminated the usefulness”, now
didn’t it?
Please see Section 13.2, which discusses some of the techniques that the Linux kernel
uses to take advantage of reference counting in a highly concurrent environment. ❑

Given that papers on hazard pointers use the bottom bits of each pointer to mark
deleted elements, what is up with HAZPTR_POISON?
Answer:
The published implementations of hazard pointers used non-blocking synchronization
techniques for insertion and deletion. These techniques require that readers traversing
the data structure “help” updaters complete their updates, which in turn means that
readers need to look at the successor of a deleted element.
In contrast, we will be using locking to synchronize updates, which does away with
the need for readers to help updaters complete their updates, which in turn allows us to
v2023.06.11a
leave pointers’ bottom bits alone. This approach allows read-side code to be simpler
and faster. ❑

Why does hp_try_record() in Listing 9.4 take a double indirection to the data
element? Why not void * instead of void **?
Answer:
Because hp_try_record() must check for concurrent modifications. To do that job,
it needs a pointer to a pointer to the element, so that it can check for a modification to
the pointer to the element. ❑

Why bother with hp_try_record()? Wouldn’t it be easier to just use the failure-
immune hp_record() function?
Answer:
It might be easier in some sense, but as will be seen in the Pre-BSD routing example,
there are situations for which hp_record() simply does not work. ❑

Readers must “typically” restart? What are some exceptions?
Answer:
If the pointer emanates from a global variable or is otherwise not subject to being freed,
then hp_record() may be used to repeatedly attempt to record the hazard pointer, even
in the face of concurrent deletions.
In certain cases, restart can be avoided by using link counting as exemplified by
the UnboundedQueue and ConcurrentHashMap data structures implemented in Folly
open-source library.9 ❑

But don’t these restrictions on hazard pointers also apply to other forms of reference
counting?
Answer:
Yes and no. These restrictions apply only to reference-counting mechanisms whose
reference acquisition can fail. ❑

Figure 9.3 shows no sign of hyperthread-induced flattening at 224 threads. Why is
that?
Answer:
Modern microprocessors are complicated beasts, so significant skepticism is appropriate
for any simple answer. That aside, the most likely reason is the full memory barriers
required by hazard-pointers readers. Any delays resulting from those memory barriers
9 https://github.com/facebook/folly
v2023.06.11a
would make time available to the other hardware thread sharing the core, resulting in
greater scalability at the expense of per-hardware-thread performance. ❑

The paper “Structured Deferral: Synchronization via Procrastination” [McK13] shows
that hazard pointers have near-ideal performance. Whatever happened in Figure 9.3???
Answer:
First, Figure 9.3 has a linear y-axis, while most of the graphs in the “Structured Deferral”
paper have logscale y-axes. Next, that paper uses lightly-loaded hash tables, while
Figure 9.3’s uses a 10-element simple linked list, which means that hazard pointers
face a larger memory-barrier penalty in this workload than in that of the “Structured
Deferral” paper. Finally, that paper used an older modest-sized x86 system, while a
much newer and larger system was used to generate the data shown in Figure 9.3.
In addition, use of pairwise asymmetric barriers [Mic08, Cor10b, Cor18] has been
proposed to eliminate the read-side hazard-pointer memory barriers on systems sup-
porting this notion [Gol18b], which might improve the performance of hazard pointers
beyond what is shown in the figure.
As always, your mileage may vary. Given the difference in performance, it is clear
that hazard pointers give you the best performance either for very large data structures
(where the memory-barrier overhead will at least partially overlap cache-miss penalties)
and for data structures such as hash tables where a lookup operation needs a minimal
number of hazard pointers. ❑

Why isn’t this sequence-lock discussion in Chapter 7, you know, the one on locking?
Answer:
The sequence-lock mechanism is really a combination of two separate synchronization
mechanisms, sequence counts and locking. In fact, the sequence-count mechanism
is available separately in the Linux kernel via the write_seqcount_begin() and
write_seqcount_end() primitives.
However, the combined write_seqlock() and write_sequnlock() primitives
are used much more heavily in the Linux kernel. More importantly, many more people
will understand what you mean if you say “sequence lock” than if you say “sequence
count”.
So this section is entitled “Sequence Locks” so that people will understand what it
is about just from the title, and it appears in the “Deferred Processing” because (1) of
the emphasis on the “sequence count” aspect of “sequence locks” and (2) because a
“sequence lock” is much more than merely a lock. ❑

Why not have read_seqbegin() in Listing 9.10 check for the low-order bit being
set, and retry internally, rather than allowing a doomed read to start?
Answer:
That would be a legitimate implementation. However, if the workload is read-mostly, it
would likely increase the overhead of the common-case successful read, which could
v2023.06.11a
be counter-productive. However, given a sufficiently large fraction of updates and

sufficiently high-overhead readers, having the check internal to read_seqbegin()
might be preferable. ❑

Why is the smp_mb() on line 26 of Listing 9.10 needed?
Answer:
If it was omitted, both the compiler and the CPU would be within their rights to move the
critical section preceding the call to read_seqretry() down below this function. This
would prevent the sequence lock from protecting the critical section. The smp_mb()
primitive prevents such reordering. ❑

Can’t weaker memory barriers be used in the code in Listing 9.10?
Answer:
In older versions of the Linux kernel, no.
In very new versions of the Linux kernel, line 16 could use smp_load_acquire()
instead of READ_ONCE(), which in turn would allow the smp_mb() on line 17 to be
dropped. Similarly, line 41 could use an smp_store_release(), for example, as
follows:
smp_store_release(&slp->seq, READ_ONCE(slp->seq) + 1);
This would allow the smp_mb() on line 40 to be dropped. ❑

What prevents sequence-locking updaters from starving readers?
Answer:
Nothing. This is one of the weaknesses of sequence locking, and as a result, you should
use sequence locking only in read-mostly situations. Unless of course read-side starvation
is acceptable in your situation, in which case, go wild with the sequence-locking updates!
❑

What if something else serializes writers, so that the lock is not needed?
Answer:
In this case, the ->lock field could be omitted, as it is in seqcount_t in the Linux
kernel. ❑
v2023.06.11a

Why isn’t seq on line 2 of Listing 9.10 unsigned rather than unsigned long? After
all, if unsigned is good enough for the Linux kernel, shouldn’t it be good enough for
everyone?
Answer:
Not at all. The Linux kernel has a number of special attributes that allow it to ignore
the following sequence of events:
1. Thread 0 executes read_seqbegin(), picking up ->seq in line 16, noting that

the value is even, and thus returning to the caller.
2. Thread 0 starts executing its read-side critical section, but is then preempted for a
long time.
3. Other threads repeatedly invoke write_seqlock() and write_sequnlock(),

until the value of ->seq overflows back to the value that Thread 0 fetched.
4. Thread 0 resumes execution, completing its read-side critical section with inconsis-
tent data.
5. Thread 0 invokes read_seqretry(), which incorrectly concludes that Thread 0

has seen a consistent view of the data protected by the sequence lock.
The Linux kernel uses sequence locking for things that are updated rarely, with
time-of-day information being a case in point. This information is updated at most
once per millisecond, so that seven weeks would be required to overflow the counter. If
a kernel thread was preempted for seven weeks, the Linux kernel’s soft-lockup code
would be emitting warnings every two minutes for that entire time.
In contrast, with a 64-bit counter, more than five centuries would be required to
overflow, even given an update every nanosecond. Therefore, this implementation uses
a type for ->seq that is 64 bits on 64-bit systems. ❑

Can this bug be fixed? In other words, can you use sequence locks as the only
synchronization mechanism protecting a linked list supporting concurrent addition,
deletion, and lookup?
Answer:
One trivial way of accomplishing this is to surround all accesses, including the read-
only accesses, with write_seqlock() and write_sequnlock(). Of course, this
solution also prohibits all read-side parallelism, resulting in massive lock contention,
and furthermore could just as easily be implemented using simple locking.
If you do come up with a solution that uses read_seqbegin() and read_
seqretry() to protect read-side accesses, make sure that you correctly handle the
1. CPU 0 is traversing the linked list, and picks up a pointer to list element A.
2. CPU 1 removes element A from the list and frees it.
v2023.06.11a
3. CPU 2 allocates an unrelated data structure, and gets the memory formerly occupied
by element A. In this unrelated data structure, the memory previously used for
element A’s ->next pointer is now occupied by a floating-point number.
4. CPU 0 picks up what used to be element A’s ->next pointer, gets random bits,
and therefore gets a segmentation fault.
One way to protect against this sort of problem requires use of “type-safe memory”,
which will be discussed in Section 9.5.4.5. Roughly similar solutions are possible using
the hazard pointers discussed in Section 9.3. But in either case, you would be using
some other synchronization mechanism in addition to sequence locks! ❑

Why does Figure 9.7 use smp_store_release() given that it is storing a NULL
pointer? Wouldn’t WRITE_ONCE() work just as well in this case, given that there is
no structure initialization to order against the store of the NULL pointer?
Answer:
Yes, it would.
Because a NULL pointer is being assigned, there is nothing to order against, so there is
no need for smp_store_release(). In contrast, when assigning a non-NULL pointer,
it is necessary to use smp_store_release() in order to ensure that initialization of
the pointed-to structure is carried out before assignment of the pointer.
In short, WRITE_ONCE() would work, and would save a little bit of CPU time on some
architectures. However, as we will see, software-engineering concerns will motivate use
of a special rcu_assign_pointer() that is quite similar to smp_store_release().
❑

Readers running concurrently with each other and with the procedure outlined in
Figure 9.7 can disagree on the value of gptr. Isn’t that just a wee bit problematic???
Answer:
Not necessarily.
As hinted at in Sections 3.2.3 and 3.3, speed-of-light delays mean that a computer’s
data is always stale compared to whatever external reality that data is intended to model.
Real-world algorithms therefore absolutely must tolerate inconsistancies between
external reality and the in-computer data reflecting that reality. Many of those algorithms
are also able to tolerate some degree of inconsistency within the in-computer data.
Section 10.3.4 discusses this point in more detail.
Please note that this need to tolerate inconsistent and stale data is not limited to RCU.
It also applies to reference counting, hazard pointers, sequence locks, and even to some
locking use cases. For example, if you compute some quantity while holding a lock, but
use that quantity after releasing that lock, you might well be using stale data. After all,
the data that quantity is based on might change arbitrarily as soon as the lock is released.
So yes, RCU readers can see stale and inconsistent data, but no, this is not necessarily
problematic. And, when needed, there are RCU usage patterns that avoid both staleness
and inconsistency [ACMS03]. ❑
v2023.06.11a

What is an RCU-protected pointer?
Answer:
A pointer to RCU-protected data. RCU-protected data is in turn a block of dynamically
allocated memory whose freeing will be deferred such that an RCU grace period will
elapse between the time that there were no longer any RCU-reader-accessible pointers
to that block and the time that that block is freed. This ensures that no RCU readers will
have access to that block at the time that it is freed.
RCU-protected pointers must be handled carefully. For example, any reader that
intends to dereference an RCU-protected pointer must use rcu_dereference() (or
stronger) to load that pointer. In addition, any updater must use rcu_assign_
pointer() (or stronger) to store to that pointer. ❑

What does synchronize_rcu() do if it starts at about the same time as an rcu_
read_lock()?
Answer:
If a synchronize_rcu() cannot prove that it started before a given rcu_read_lock(),
then it must wait for the corresponding rcu_read_unlock(). ❑

In Figure 9.8, the last of CPU 3’s readers that could possibly have access to the old
data item ended before the grace period even started! So why would anyone bother
waiting until CPU 3’s later context switch???
Answer:
Because that waiting is exactly what enables readers to use the same sequence of
instructions that is appropriate for single-theaded situations. In other words, this
additional “redundant” waiting enables excellent read-side performance, scalability, and
real-time response. ❑

What is the point of rcu_read_lock() and rcu_read_unlock() in Listing 9.13?
Why not just let the quiescent states speak for themselves?
Answer:
Recall that readers are not permitted to pass through a quiescent state. For example,
within the Linux kernel, RCU readers are not permitted to execute a context switch. Use
of rcu_read_lock() and rcu_read_unlock() enables debug checks for improperly
placed quiescent states, making it easy to find bugs that would otherwise be difficult to
find, intermittent, and quite destructive. ❑
v2023.06.11a

What is the point of rcu_dereference(), rcu_assign_pointer() and RCU_
INIT_POINTER() in Listing 9.13? Why not just use READ_ONCE(), smp_store_
release(), and WRITE_ONCE(), respectively?
Answer:
The RCU-specific APIs do have similar semantics to the suggested replacements, but
also enable static-analysis debugging checks that complain if an RCU-specific API is
invoked on a non-RCU pointer and vice versa. ❑

But what if the old structure needs to be freed, but the caller of ins_route() cannot
block, perhaps due to performance considerations or perhaps because the caller is
executing within an RCU read-side critical section?
Answer:
A call_rcu() function, which is described in Section 9.5.2.2, permits asynchronous
grace-period waits. ❑

Doesn’t Section 9.4’s seqlock also permit readers and updaters to make useful
concurrent forward progress?
Answer:
Yes and no. Although seqlock readers can run concurrently with seqlock writers,
whenever this happens, the read_seqretry() primitive will force the reader to retry.
This means that any work done by a seqlock reader running concurrently with a seqlock
updater will be discarded and then redone upon retry. So seqlock readers can run
concurrently with updaters, but they cannot actually get any work done in this case.
In contrast, RCU readers can perform useful work even in presence of concurrent
RCU updaters.
However, both reference counters (Section 9.2) and hazard pointers (Section 9.3)
really do permit useful concurrent forward progress for both updaters and readers, just
at somewhat greater cost. Please see Section 9.6 for a comparison of these different
solutions to the deferred-reclamation problem. ❑

Wouldn’t use of data ownership for RCU updaters mean that the updates could use
exactly the same sequence of instructions as would the corresponding single-threaded
code?
Answer:
Sometimes, for example, on TSO systems such as x86 or the IBM mainframe where
a store-release operation emits a single store instruction. However, weakly ordered
systems must also emit a memory barrier or perhaps a store-release instruction. In
addition, removing data requires quite a bit of additional work because it is necessary to
wait for pre-existing readers before freeing the removed data. ❑
v2023.06.11a

But suppose that updaters are adding and removing multiple data items from a linked
list while a reader is iterating over that same list. Specifically, suppose that a list
initially contains elements A, B, and C, and that an updater removes element A and
then adds a new element D at the end of the list. The reader might well see {A, B, C,
D}, when that sequence of elements never actually ever existed! In what alternate
universe would that qualify as “not disrupting concurrent readers”???
Answer:
In the universe where an iterating reader is only required to traverse elements that were
present throughout the full duration of the iteration. In the example, that would be
elements B and C. Because elements A and D were each present for only part of the
iteration, the reader is permitted to iterate over them, but not obliged to. Note that this
supports the common case where the reader is simply looking up a single item, and
does not know or care about the presence or absence of other items.
If stronger consistency is required, then higher-cost synchronization mechanisms
are required, for example, sequence locking or reader-writer locking. But if stronger
consistency is not required (and it very often is not), then why pay the higher cost? ❑

What other final values of r1 and r2 are possible in Figure 9.11?
Answer:
The r1 == 0 && r2 == 0 possibility was called out in the text. Given that r1 == 0
implies r2 == 0, we know that r1 == 0 && r2 == 1 is forbidden. The following
discussion will show that both r1 == 1 && r2 == 1 and r1 == 1 && r2 == 0 are
possible. ❑

What would happen if the order of P0()’s two accesses was reversed in Figure 9.12?
Answer:
Absolutely nothing would change. The fact that P0()’s loads from x and y are in the
same RCU read-side critical section suffices; their order is irrelevant. ❑

What would happen if P0()’s accesses in Figures 9.11–9.13 were stores?
Answer:
The exact same ordering rules would apply, that is, (1) If any part of P0()’s RCU
read-side critical section preceded the beginning of P1()’s grace period, all of P0()’s
RCU read-side critical section would precede the end of P1()’s grace period, and (2) If
any part of P0()’s RCU read-side critical section followed the end of P1()’s grace
period, all of P0()’s RCU read-side critical section would follow the beginning of
P1()’s grace period.
It might seem strange to have RCU read-side critical sections containing writes, but
this capability is not only permitted, but also highly useful. For example, the Linux
kernel frequently carries out an RCU-protected traversal of a linked data structure and
v2023.06.11a
Listing E.2: Concurrent RCU Deletion

2 p = search(head, key);
3 if (p == NULL)
5 else {
6 list_del_rcu(&p->list);
9 kfree(p);
10 }
then acquires a reference to the destination data element. Because this data element
must not be freed in the meantime, that element’s reference counter must necessarily
be incremented within the traversal’s RCU read-side critical section. However, that
increment entails a write to memory. Therefore, it is a very good thing that memory
writes are permitted within RCU read-side critical sections.
If having writes in RCU read-side critical sections still seems strange, please review
Section 5.4.6, which presented a use case for writes in reader-writer locking read-side
critical sections. ❑

How would you modify the deletion example to permit more than two versions of the
list to be active?
Answer:
One way of accomplishing this is as shown in Listing E.2.
Note that this means that multiple concurrent deletions might be waiting in
synchronize_rcu(). ❑

How many RCU versions of a given list can be active at any given time?
Answer:
That depends on the synchronization design. If a semaphore protecting the update is
held across the grace period, then there can be at most two versions, the old and the new.
However, suppose that only the search, the update, and the list_replace_rcu()
were protected by a lock, so that the synchronize_rcu() was outside of that lock,
similar to the code shown in Listing E.2. Suppose further that a large number of
threads undertook an RCU replacement at about the same time, and that readers are
also constantly traversing the data structure.
Then the following sequence of events could occur, starting from the end state of
Figure 9.15:
1. Thread A traverses the list, obtaining a reference to Element C.
2. Thread B replaces Element C with a new Element F, then waits for its
synchronize_rcu() call to return.
3. Thread C traverses the list, obtaining a reference to Element F.
4. Thread D replaces Element F with a new Element G, then waits for its
v2023.06.11a
5. Thread E traverses the list, obtaining a reference to Element G.
6. Thread F replaces Element G with a new Element H, then waits for its
7. Thread G traverses the list, obtaining a reference to Element H.
8. And the previous two steps repeat quickly with additional new elements, so that all
of them happen before any of the synchronize_rcu() calls return.
Thus, there can be an arbitrary number of versions active, limited only by memory
and by how many updates could be completed within a grace period. But please note
that data structures that are updated so frequently are not likely to be good candidates
for RCU. Nevertheless, RCU can handle high update rates when necessary. ❑

How can the per-update overhead of RCU be reduced?
Answer:
The most effective way to reduce the per-update overhead of RCU is to increase the
number of updates served by a given grace period. This works because the per-grace
period overhead is nearly independent of the number of updates served by that grace
period.
One way to do this is to delay the start of a given grace period in the hope that more
updates requiring that grace period appear in the meantime. Another way is to slow
down execution of the grace period in the hope that more updates requiring an additional
grace period will accumulate in the meantime.
There are many other possible optimizations, and fanatically devoted readers are
referred to the Linux-kernel RCU implementation. ❑

How can RCU updaters possibly delay RCU readers, given that neither rcu_read_
lock() nor rcu_read_unlock() spin or block?
Answer:
The modifications undertaken by a given RCU updater will cause the corresponding CPU
to invalidate cache lines containing the data, forcing the CPUs running concurrent RCU
readers to incur expensive cache misses. (Can you design an algorithm that changes
a data structure without inflicting expensive cache misses on concurrent readers? On
subsequent readers?) ❑

Why do some of the cells in Table 9.2 have exclamation marks (“!”)?
Answer:
The API members with exclamation marks (rcu_read_lock(), rcu_read_unlock(),
and call_rcu()) were the only members of the Linux RCU API that Paul E. McKenney
was aware of back in the mid-90s. During this timeframe, he was under the mistaken
impression that he knew all that there is to know about RCU. ❑
v2023.06.11a
Listing E.3: synchronize_rcu() vs. rcu_barrier()

1 do_something_1();
2 rcu_read_lock();
3 do_something_2();
4 call_rcu(&p->rh, f);
5 do_something_3();
7 do_something_4();
8 // f(&p->rh) invoked
9 do_something_5();

How do you prevent a huge number of RCU read-side critical sections from indefinitely
blocking a synchronize_rcu() invocation?
Answer:
There is no need to do anything to prevent RCU read-side critical sections from
indefinitely blocking a synchronize_rcu() invocation, because the synchronize_
rcu() invocation need wait only for pre-existing RCU read-side critical sections. So as
long as each RCU read-side critical section is of finite duration, RCU grace periods will
also remain finite. ❑

The synchronize_rcu() API waits for all pre-existing interrupt handlers to complete,
right?
Answer:
In v4.20 and later Linux kernels, yes [McK19c, McK19a].
But not in earlier kernels, and especially not when using preemptible RCU! You
instead want synchronize_irq(). Alternatively, you can place calls to rcu_read_
lock() and rcu_read_unlock() in the specific interrupt handlers that you want
synchronize_rcu() to wait for. But even then, be careful, as preemptible RCU
will not be guaranteed to wait for that portion of the interrupt handler preceding the
rcu_read_lock() or following the rcu_read_unlock(). ❑

What is the difference between synchronize_rcu() and rcu_barrier()?
Answer:
They wait on different things. While synchronize_rcu() waits for pre-existing RCU
read-side critical sections to complete, rcu_barrier() instead waits for callbacks
from prior calls to call_rcu() to be invoked.
This distinction is illustrated by Listing E.3, which shows code being executed by a
given CPU. For simplicity, assume that no other CPU is executing rcu_read_lock(),
rcu_read_unlock(), or call_rcu().
Table E.3 shows how long each primitive must wait if invoked concurrently with each
of the do_something_*() functions, with empty cells indicating that no waiting is
necessary. As you can see, synchronize_rcu() need not wait unless it is in an RCU
read-side critical section, in which case it must wait for the rcu_read_unlock() that
ends that critical section. In contrast, RCU read-side critical sections have no effect
v2023.06.11a
Table E.3: synchronize_rcu() vs. rcu_barrier()
Must Wait Until (in Listing E.3):
Invoked at: synchronize_rcu() rcu_barrier()
do_something_1()
do_something_2() rcu_read_unlock() (line 6)
do_something_3() rcu_read_unlock() (line 6) f(&p->rh) (line 8)
do_something_4() f(&p->rh) (line 8)
do_something_5()
Listing E.4: Multistage SRCU Deadlocks

1 idx = srcu_read_lock(&ssa);
2 synchronize_srcu(&ssb);
3 srcu_read_unlock(&ssa, idx);
4
5 /* . . . */
6
7 idx = srcu_read_lock(&ssb);
8 synchronize_srcu(&ssa);
9 srcu_read_unlock(&ssb, idx);
on rcu_barrier(). However, when rcu_barrier() executes after a call_rcu()

invocation, it must wait until the corresponding RCU callback is invoked.
All that said, there is a special case where each call to rcu_barrier() can be
replaced by a direct call to synchronize_rcu(), and that is where synchronize_
rcu() is implemented in terms of call_rcu() and where there is a single global list
of callbacks. But please do not do this in portable code!!! ❑

Under what conditions can synchronize_srcu() be safely used within an SRCU
read-side critical section?
Answer:
In principle, you can use either synchronize_srcu() or synchronize_srcu_
expedited() with a given srcu_struct within an SRCU read-side critical section
that uses some other srcu_struct. In practice, however, doing this is almost certainly
a bad idea. In particular, the code shown in Listing E.4 could still result in deadlock. ❑

In a kernel built with CONFIG_PREEMPT_NONE=y, won’t synchronize_rcu() wait
for all trampolines, given that preemption is disabled and that trampolines never
directly or indirectly invoke schedule()?
Answer:
You are quite right!
In fact, in nonpreemptible kernels, synchronize_rcu_tasks() is a wrapper around
synchronize_rcu(). ❑
v2023.06.11a
Listing E.5: Diverse RCU Read-Side Nesting

1 rcu_read_lock();
2 preempt_disable();
3 p = rcu_dereference(global_pointer);
4
5 /* . . . */
6
7 preempt_enable();

Normally, any pointer subject to rcu_dereference() must always be updated
using one of the pointer-publish functions in Table 9.3, for example, rcu_assign_
pointer().
What is an exception to this rule?
Answer:
One such exception is when a multi-element linked data structure is initialized as a unit
while inaccessible to other CPUs, and then a single rcu_assign_pointer() is used to
plant a global pointer to this data structure. The initialization-time pointer assignments
need not use rcu_assign_pointer(), though any such assignments that happen after
the structure is globally visible must use rcu_assign_pointer().
However, unless this initialization code is on an impressively hot code-path, it is
probably wise to use rcu_assign_pointer() anyway, even though it is in theory
unnecessary. It is all too easy for a “minor” change to invalidate your cherished
assumptions about the initialization happening privately. ❑

Are there any downsides to the fact that these traversal and update primitives can be
used with any of the RCU API family members?
Answer:
It can sometimes be difficult for automated code checkers such as “sparse” (or indeed for
human beings) to work out which type of RCU read-side critical section a given RCU
traversal primitive corresponds to. For example, consider the code shown in Listing E.5.
Is the rcu_dereference() primitive in a vanilla RCU critical section or an RCU
Sched critical section? What would you have to do to figure this out?
But perhaps after the consolidation of the RCU flavors in the v4.20 Linux kernel we
no longer need to care! ❑

But what if an hlist_nulls reader gets moved to some other bucket and then back
again?
Answer:
One way to handle this is to always move nodes to the beginning of the destination
bucket, ensuring that when the reader reaches the end of the list having a matching NULL
pointer, it will have searched the entire list.
Of course, if there are too many move operations in a hash table with many elements
per bucket, the reader might never reach the end of a list. One way of avoiding this
v2023.06.11a
in the common case is to keep hash tables well-tuned, thus with short lists. One way
of detecting the problem and handling it is for the reader to terminate the search after
traversing some large number of nodes, acquire the update-side lock, and redo the search,
but this might introduce deadlocks. Another way of avoiding the problem entirely is for
readers to search within RCU read-side critical sections, and to wait for an RCU grace
period between successive updates. An intermediate position might wait for an RCU
grace period every 𝑁 updates, for some suitable value of 𝑁. ❑

Why isn’t there a rcu_read_lock_tasks_held() for Tasks RCU?
Answer:
Because Tasks RCU does not have read-side markers. Instead, Tasks RCU read-side
critical sections are bounded by voluntary context switches. ❑

Wait, what??? How can RCU QSBR possibly be better than ideal? Just what rubbish
definition of ideal would fail to be the best of all possible results???
Answer:
This is an excellent question, and the answer is that modern CPUs and compilers are
extremely complex. But before getting into that, it is well worth noting that RCU QSBR’s
performance advantage appears only in the one-hardware-thread-per-core regime. Once
the system is fully loaded, RCU QSBR’s performance drops back to ideal.
The RCU variant of the route_lookup() search loop actually has one more x86
instruction than does the sequential version, namely the lea in the sequence cmp, je,
mov, cmp, lea, and jne. This extra instruction is due to the rcu_head structure at the
beginning of the RCU variant’s route_entry structure, so that, unlike the sequential
variant, the RCU variant’s ->re_next.next pointer has a non-zero offset. Back in the
1980s, this additional lea instruction might have reliably resulted in the RCU variant
being slower, but we are now in the 21st century, and the 1980s are long gone.
But those of you who read Section 3.1.1 carefully already knew all of this!
These counter-intuitive results of course means that any performance result on modern
microprocessors must be subject to some skepticism. In theory, it really does not make
sense to obtain performance results that are better than ideal, but it really can happen on
modern microprocessors. Such results can be thought of as similar to the celebrated
super-linear speedups (see Section 6.5 for one such example), that is, of interest but
also of limited practical importance. Nevertheless, one of the strengths of RCU is that
its read-side overhead is so low that tiny effects such as this one are visible in real
performance measurements.
This raises the question as to what would happen if the rcu_head structure were to
be moved so that RCU’s ->re_next.next pointer also had zero offset, just the same
as the sequential variant. And the answer, as can be seen in Figure E.5, is that this
causes RCU QSBR’s performance to decrease to where it is still very nearly ideal, but
no longer super-ideal. ❑
v2023.06.11a
2.5x107
2x107 ideal

7
1.5x10
RCU
1x107
seqlock
5x106
hazptr
0
0 50 100 150 200 250 300 350 400 450
Figure E.5: Pre-BSD Routing Table Protected by RCU QSBR With Non-Initial
rcu_head

Given RCU QSBR’s read-side performance, why bother with any other flavor of
userspace RCU?
Answer:
Because RCU QSBR places constraints on the overall application that might not be
tolerable, for example, requiring that each and every thread in the application regularly
pass through a quiescent state. Among other things, this means that RCU QSBR is
not helpful to library writers, who might be better served by other flavors of userspace
RCU [MDJ13f]. ❑

Suppose that the nmi_profile() function was preemptible. What would need to
change to make this example work correctly?
Answer:
One approach would be to use rcu_read_lock() and rcu_read_unlock() in nmi_
profile(), and to replace the synchronize_sched() with synchronize_rcu(),
perhaps as shown in Listing E.6.
But why on earth would an NMI handler be preemptible??? ❑

What is the point of the second call to synchronize_rcu() in function maint() in
Listing 9.17? Isn’t it OK for any cco() invocations in the clean-up phase to invoke
either cco_carefully() or cco_quickly()?
Answer:
The problem is that there is no ordering between the cco() function’s load from be_
careful and any memory loads executed by the cco_quickly() function. Because
there is no ordering, without that second call to syncrhonize_rcu(), memory ordering
could cause loads in cco_quickly() to overlap with stores by do_maint().
v2023.06.11a
Listing E.6: Using RCU to Wait for Mythical Preemptible NMIs to Finish
1 struct profile_buffer {
2 long size;
3 atomic_t entry[0];
4 };
5 static struct profile_buffer *buf = NULL;
6
7 void nmi_profile(unsigned long pcvalue)
8 {
9 struct profile_buffer *p;
10
11 rcu_read_lock();
12 p = rcu_dereference(buf);
13 if (p == NULL) {
15 return;
16 }
17 if (pcvalue >= p->size) {
19 return;
20 }
21 atomic_inc(&p->entry[pcvalue]);
23 }
24
25 void nmi_stop(void)
26 {
27 struct profile_buffer *p = buf;
28
29 if (p == NULL)
30 return;
31 rcu_assign_pointer(buf, NULL);
33 kfree(p);
34 }
Another alternative would be to compensate for the removal of that second call to
synchronize_rcu() by changing the READ_ONCE() to smp_load_acquire() and
the WRITE_ONCE() to smp_store_release(), thus restoring the needed ordering. ❑

How can you be sure that the code shown in maint() in Listing 9.17 really works?
Answer:
By one popular school of thought, you cannot.
But in this case, those willing to jump ahead to Chapter 12 and Chapter 15 might find
a couple of LKMM litmus tests to be interesting (C-RCU-phased-state-change-
1.litmus and C-RCU-phased-state-change-2.litmus). These tests could be
argued to demonstrate that this code and a variant of it really do work. ❑

But what if there is an arbitrarily long series of RCU read-side critical sections in
multiple threads, so that at any point in time there is at least one thread in the system
executing in an RCU read-side critical section? Wouldn’t that prevent any data from a
SLAB_TYPESAFE_BY_RCU slab ever being returned to the system, possibly resulting
in OOM events?
Answer:
There could certainly be an arbitrarily long period of time during which at least one
v2023.06.11a
thread is always in an RCU read-side critical section. However, the key words in the
description in Section 9.5.4.5 are “in-use” and “pre-existing”. Keep in mind that a
given RCU read-side critical section is conceptually only permitted to gain references
to data elements that were visible to readers during that critical section. Furthermore,
remember that a slab cannot be returned to the system until all of its data elements have
been freed, in fact, the RCU grace period cannot start until after they have all been freed.
Therefore, the slab cache need only wait for those RCU read-side critical sections
that started before the freeing of the last element of the slab. This in turn means that
any RCU grace period that begins after the freeing of the last element will do—the slab
may be returned to the system after that grace period ends. ❑

What if the element we need to delete is not the first element of the list on line 9 of
Listing 9.18?
Answer:
As with the (bug-ridden) Listing 7.10, this is a very simple hash table with no chaining,
so the only element in a given bucket is the first element. The reader is again invited to
adapt this example to a hash table with full chaining. Less energetic readers might wish
to refer to Chapter 10. ❑

Why is it OK to exit the RCU read-side critical section on line 15 of Listing 9.18
before releasing the lock on line 17?
Answer:
First, please note that the second check on line 14 is necessary because some other CPU
might have removed this element while we were waiting to acquire the lock. However,
the fact that we were in an RCU read-side critical section while acquiring the lock
guarantees that this element could not possibly have been re-allocated and re-inserted
into this hash table. Furthermore, once we acquire the lock, the lock itself guarantees
the element’s existence, so we no longer need to be in an RCU read-side critical section.
The question as to whether it is necessary to re-check the element’s key is left as an
exercise to the reader. ❑

Why not exit the RCU read-side critical section on line 23 of Listing 9.18 before
releasing the lock on line 22?
Answer:
Suppose we reverse the order of these two lines. Then this code is vulnerable to the
1. CPU 0 invokes delete(), and finds the element to be deleted, executing through
line 15. It has not yet actually deleted the element, but is about to do so.
2. CPU 1 concurrently invokes delete(), attempting to delete this same element.

However, CPU 0 still holds the lock, so CPU 1 waits for it at line 13.
v2023.06.11a
3. CPU 0 executes lines 16 and 17, and blocks at line 18 waiting for CPU 1 to exit its
RCU read-side critical section.
4. CPU 1 now acquires the lock, but the test on line 14 fails because CPU 0 has
already removed the element. CPU 1 now executes line 22 (which we switched
with line 23 for the purposes of this Quick Quiz) and exits its RCU read-side
critical section.
5. CPU 0 can now return from synchronize_rcu(), and thus executes line 19,
sending the element to the freelist.
6. CPU 1 now attempts to release a lock for an element that has been freed, and,
worse yet, possibly reallocated as some other type of data structure. This is a fatal
memory-corruption error. ❑

The RCU-based algorithm shown in Listing 9.18 locks very similar to that in List-
ing 7.11, so why should the RCU-based approach be any better?
Answer:
Listing 9.18 replaces the per-element spin_lock() and spin_unlock() shown in
Listing 7.11 with a much cheaper rcu_read_lock() and rcu_read_unlock(), thus
greatly improving both performance and scalability. For more detail, please see
Section 10.3.3. ❑

WTF? How the heck do you expect me to believe that RCU can have less than a 300-
picosecond overhead when the clock period at 2.10 GHz is almost 500 picoseconds?
Answer:
First, consider that the inner loop used to take this measurement is as follows:
1 for (i = nloops; i >= 0; i--) {

2 rcu_read_lock();
4 }
Next, consider the effective definitions of rcu_read_lock() and rcu_read_

unlock():
1 #define rcu_read_lock() barrier()

2 #define rcu_read_unlock() barrier()
These definitions constrain compiler code-movement optimizations involving memory

references, but emit no instructions in and of themselves. However, if the loop variable
is maintained in a register, the accesses to i will not count as memory references.
Furthermore, the compiler can do loop unrolling, allowing the resulting code to “execute”
multiple passes through the loop body simply by incrementing i by some value larger
than the value 1.
v2023.06.11a
So the “measurement” of 267 picoseconds is simply the fixed overhead of the timing
measurements divided by the number of passes through the inner loop containing the
calls to rcu_read_lock() and rcu_read_unlock(), plus the code to manipulate i
divided by the loop-unrolling factor. And therefore, this measurement really is in error,
in fact, it exaggerates the overhead by an arbitrary number of orders of magnitude. After
all, in terms of machine instructions emitted, the actual overheads of rcu_read_lock()
and of rcu_read_unlock() are each precisely zero.
It is not just every day that a timing measurement of 267 picoseconds turns out to be
an overestimate! ❑

Didn’t an earlier edition of this book show RCU read-side overhead way down in the
sub-picosecond range? What happened???
Answer:
Excellent memory!!! The overhead in some early releases was in fact roughly 100 fem-
toseconds.
What happened was that RCU usage spread more broadly through the Linux kernel,
including into code that takes page faults. Back at that time, rcu_read_lock() and
rcu_read_unlock() were complete no-ops in CONFIG_PREEMPT=n kernels. Unfortu-
nately, that situation allowed the compiler to reorder page-faulting memory accesses
into RCU read-side critical sections. Of course, page faults can block, which destroys
those critical sections.
Nor was this a theoretical problem: A failure actually manifested in 2019. Herbert
Xu tracked down this failure down and Linus Torvalds therefore queued a commit to
upgrade rcu_read_lock() and rcu_read_unlock() to unconditionally include a
call to barrier() [Tor19]. And although barrier() emits no code, it does constrain
compiler optimizations. And so the price of widespread RCU usage is slightly higher
rcu_read_lock() and rcu_read_unlock() overhead. As such, Linux-kernel RCU
has proven to be a victim of its own success.
Of course, it is also the case that the older results were obtained on a different system
than were those shown in Figure 9.25. So which change had the most effect, Linus’s
commit or the change in the system? This question is left as an exercise to the reader. ❑

Why is there such large variation for the RCU trace in Figure 9.25?
Answer:
Keep in mind that this is a log-log plot, so those large-seeming RCU variances in reality
span only a few hundred picoseconds. And that is such a short time that anything could
cause it. However, given that the variance decreases with both small and large numbers
of CPUs, one hypothesis is that the variation is due to migrations from one CPU to
another.
Yes, these measurements were taken with interrupts disabled, but they were also
taken within a guest OS, so that preemption was still possible at the hypervisor level. In
addition, the system featured hyperthreading and a single hardware thread running this
RCU workload is able to consume more than half of the core’s resources. Therefore, the
overall throughput varies depending on how many of a given guest OS’s CPUs share
v2023.06.11a
cores. Attempting to reduce these variations by running the guest OSes at real-time
priority (as suggested by Joel Fernandes) is left as an exercise for the reader. ❑

Given that the system had no fewer than 448 hardware threads, why only 192 CPUs?
Answer:
Because the script (rcuscale.sh) that generates this data spawns a guest operating
system for each set of points gathered, and on this particular system, both qemu and
KVM limit the number of CPUs that may be configured into a given guest OS. Yes, it
would have been possible to run a few more CPUs, but 192 is a nice round number from
a binary perspective, given that 256 is infeasible. ❑

Why the larger error ranges for the submicrosecond durations in Figure 9.27?
Answer:
Because smaller disturbances result in greater relative errors for smaller measurements.
Also, the Linux kernel’s ndelay() nanosecond-scale primitive is (as of 2020) less
accurate than is the udelay() primitive used for the data for durations of a microsecond
or more. It is instructive to compare to the zero-length case shown in Figure 9.25. ❑

Is there an exception to this deadlock immunity, and if so, what sequence of events
could lead to deadlock?
Answer:
One way to cause a deadlock cycle involving RCU read-side primitives is via the
following (illegal) sequence of statements:
rcu_read_lock();
synchronize_rcu();
rcu_read_unlock();
The synchronize_rcu() cannot return until all pre-existing RCU read-side critical
sections complete, but is enclosed in an RCU read-side critical section that cannot
complete until the synchronize_rcu() returns. The result is a classic self-deadlock—
you get the same effect when attempting to write-acquire a reader-writer lock while
read-holding it.
Note that this self-deadlock scenario does not apply to RCU QSBR, because the
context switch performed by the synchronize_rcu() would act as a quiescent state
for this CPU, allowing a grace period to complete. However, this is if anything even
worse, because data used by the RCU read-side critical section might be freed as a result
of the grace period completing. Plus Linux kernel’s lockdep facility will yell at you.
In short, do not invoke synchronous RCU update-side primitives, which are listed in
Table 9.2, from within an RCU read-side critical section.
In addition, within the Linux kernel, RCU uses the scheduler and the scheduler uses
RCU. In some cases, both RCU and the scheduler must take care to avoid deadlock. ❑
v2023.06.11a

Immunity to both deadlock and priority inversion??? Sounds too good to be true.
Why should I believe that this is even possible?
Answer:
It really does work. After all, if it didn’t work, the Linux kernel would not run. ❑

But how many other algorithms really tolerate stale and inconsistent data?
Answer:
Quite a few!
Please keep in mind that the finite speed of light means that data reaching a given
computer system is at least slightly stale at the time that it arrives, and extremely stale
in the case of astronomical data. The finite speed of light also places a sharp limit on
the consistency of data arriving from different sources of via different paths.
You might as well face the fact that the laws of physics are incompatible with naive
notions of perfect freshness and consistency. ❑

If Tasks RCU Trace might someday be priority boosted, why not also Tasks RCU and
Tasks RCU Rude?
Answer:
Maybe, but these are less likely.
In the case of Tasks RCU, recall that the quiescent state is a voluntary context switch.
Thus, all tasks not blocked after a voluntary context switch might need to be boosted,
and the mechanics of deboosting would not likely be at all pretty.
In the case of Tasks RCU Rude, as was the case with the old RCU Sched, any
preemptible region of code is a quiescent state. Thus, the only tasks that might need
boosting are those currently running with preemption disabled. But boosting the priority
of a preemption-disabled task has no effect. It therefore seems doubly unlikely that
priority boosting will ever be introduced to Tasks RCU Rude, at least in its current form.
❑

But doesn’t the RCU grace period start sometime after the call to synchronize_rcu()
rather than in the middle of that xchg() statement?
Answer:
Which grace period, exactly?
The updater is required to wait for at least one grace period that starts at or some
time after the removal, in this case, the xchg(). So in Figure 9.29, the indicated
grace period starts as early as theoretically possible and extends to the return from
synchronize_rcu(). This is a perfectly legal grace period corresponding to the
change carried out by that xchg() statement. ❑
v2023.06.11a

Is RCU the only synchronization mechanism that combines temporal and spatial
synchronization in this way?
Answer:
Not at all.
Hazard pointers can be considered to combine temporal and spatial synchronization in
a similar manner. Referring to Listing 9.4, the hp_record() function’s acquisition of a
reference provides both spatial and temporal synchronization, subscribing to a version and
marking the start of a reference, respectively. This function therefore combines the effects
of RCU’s rcu_read_lock() and rcu_dereference(). Referring now to Listing 9.5,
the hp_clear() function’s release of a reference provides temporal synchronization
marking the end of a reference, and is thus similar to RCU’s rcu_read_unlock().
The hazptr_free_later() function’s retiring of a hazard-pointer-protected object
provides temporal synchronization, similar to RCU’s call_rcu(). The primitives used
to mutate a hazard-pointer-protected structure provide spatial synchronization, similar
to RCU’s rcu_assign_pointer().
Alternatively, one could instead come at hazard pointers by analogy with reference
counting. ❑

But wait! This is exactly the same code that might be used when thinking of RCU as a
replacement for reader-writer locking! What gives?
Answer:
This is an effect of the Law of Toy Examples: Beyond a certain point, the code fragments
look the same. The only difference is in how we think about the code. For example,
what does an atomic_inc() operation do? It might be acquiring another explicit
reference to an object to which we already have a reference, it might be incrementing an
often-read/seldom-updated statistical counter, it might be checking into an HPC-style
barrier, or any of a number of other things.
However, these differences can be extremely important. For but one example of
the importance, consider that if we think of RCU as a restricted reference counting
scheme, we would never be fooled into thinking that the updates would exclude the
RCU read-side critical sections.
It nevertheless is often useful to think of RCU as a replacement for reader-writer
locking, for example, when you are replacing reader-writer locking with RCU. ❑

Which of these use cases best describes the Pre-BSD routing example in Section 9.5.4.1?
Answer:
Pre-BSD routing could be argued to fit into either quasi reader-writer lock, quasi
reference count, or quasi multi-version concurrency control. The code is the same either
way. This is similar to things like atomic_inc(), another tool that can be put to a great
many uses. ❑
v2023.06.11a

Garbage collectors? Passive serialization? System reference points? Quiescent states?
Aging? Generations? Why on earth couldn’t the knuckleheads working on these early
papers bring themselves to agree on a common terminology???
Answer:
There were multiple independent inventions of mechanisms vaguely resembling RCU.
Each group of inventors was unaware of the others, so each made up its own terminology
as a matter of course. And the different terminology made it quite difficult for any one
group to find any of the others.
Sorry, but life is like that sometimes! ❑

Why didn’t Kung’s and Lehman’s paper result in immediate use of RCU?
Answer:
One reason is that Kung and Lehman were simply ahead of their time. Another reason
was that their approach, ground-breaking though it was, did not take a number of
software-engineering and performance issues into account.
To see that they were ahead of their time, consider that three years after their
paper was published, Paul was working on a PDP-11 system running BSD 2.8. This
system lacked any sort of automatic configuration, which meant that any hardware
modification, including adding a new disk drive, required hand-editing and rebuilding
the kernel. Furthermore, this was a single-CPU system, which meant that full-system
synchronization was a simple matter of disabling interrupts.
Fast-forward a number of years, and multicore systems permitting runtime changes in
hardware configuration were commonplace. This meant that the hardware configuration
data that was implicitly represented in 1980s kernel source code was now a mutable
data structure that was accessed on every I/O. Such data structures rarely change, but
could change at any time. And this read-mostly property applies to many other new-age
data structures, including those concerning networking (rare in the 1980s), security
policies (physical locks in the 1980s), software configuration (immutable at runtime in
the 1980s), and much else besides. There was thus much more opportunity for RCU to
demonstrate its benefits in the 1990s and 2000s than there was in the 1980s.
Kung’s and Lehman’s software-engineering sins included failing to mark readers (thus
presenting debugging difficulties), failing to provide a clean RCU API (thus tying their
mechanism to a specific data structure), and failing to allow for any post-grace-period
operation other than freeing memory (thus disallowing a number of RCU use cases).
Kung and Lehman presented two garbage-collection strategies. The first waited for
all processes running at a given time to terminate, which represented another software-
engineering sin that ruled out their mechanism’s use in software that runs indefinitely.
The second used per-object reference counting, which greatly complicates their read-side
code (thus representing yet another software-engineering sin), and, on modern hardware,
results in severe cache-miss overhead (thus representing a performance sin, see for
example Figures 9.30 and 9.31).
Despite this long list of software-engineering and performance sins, Kung’s and
Lehman’s paper remains a truly impressive piece of work, especially considering that
much of the later work (both independent and not) committed these same sins, plus
others as well. ❑
v2023.06.11a

Why not just drop the lock before waiting for the grace period, or using something
like call_rcu() instead of waiting for a grace period?
Answer:
The authors wished to support linearizable tree operations, so that concurrent additions
to, deletions from, and searches of the tree would appear to execute in some globally
agreed-upon order. In their search trees, this requires holding locks across grace
periods. (It is probably better to drop linearizability as a requirement in most cases, but
linearizability is a surprisingly popular (and costly!) requirement.) ❑

Why can’t users dynamically allocate the hazard pointers as they are needed?
Answer:
They can, but at the expense of additional reader-traversal overhead and, in some
environments, the need to handle memory-allocation failure. ❑

But don’t Linux-kernel kref reference counters allow guaranteed unconditional
reference acquisition?
Answer:
Yes they do, but the guarantee only applies unconditionally in cases where a reference
is already held. With this in mind, please review the paragraph at the beginning of
Section 9.6, especially the part saying “large enough that readers do not hold references
from one traversal to another”. ❑

But didn’t the answer to one of the quick quizzes in Section 9.3 say that pairwise
asymmetric barriers could eliminate the read-side smp_mb() from hazard pointers?
Answer:
Yes, it did. However, doing this could be argued to change hazard-pointers “Reclamation
Forward Progress” row (discussed later) from lock-free to blocking because a CPU
spinning with interrupts disabled in the kernel would prevent the update-side portion
of the asymmetric barrier from completing. In the Linux kernel, such blocking could
in theory be prevented by building the kernel with CONFIG_NO_HZ_FULL, designating
the relevant CPUs as nohz_full at boot time, ensuring that only one thread was ever
runnable on a given CPU at a given time, and avoiding ever calling into the kernel.
Alternatively, you could ensure that the kernel was free of any bugs that might cause
CPUs to spin with interrupts disabled.
Given that CPUs spinning in the Linux kernel with interrupts disabled seems to be
rather rare, one might counter-argue that asymmetric-barrier hazard-pointer updates are
non-blocking in practice, if not in theory. ❑
v2023.06.11a
E.10 Data Structures

But chained hash tables are but one type of many. Why the focus on chained hash
tables?
Answer:
Chained hash tables are completely partitionable, and thus well-suited to concurrent
use. There are other completely-partitionable hash tables, for example, split-ordered
list [SS06], but they are considerably more complex. We therefore start with chained
hash tables. ❑

But isn’t the double comparison on lines 10–13 in Listing 10.3 inefficient in the case
where the key fits into an unsigned long?
Answer:
Indeed it is! However, hash tables quite frequently store information with keys such
as character strings that do not necessarily fit into an unsigned long. Simplifying the
hash-table implementation for the case where keys always fit into unsigned longs is left
as an exercise for the reader. ❑

Instead of simply increasing the number of hash buckets, wouldn’t it be better to
cache-align the existing hash buckets?
Answer:
The answer depends on a great many things. If the hash table has a large number of
elements per bucket, it would clearly be better to increase the number of hash buckets.
On the other hand, if the hash table is lightly loaded, the answer depends on the
hardware, the effectiveness of the hash function, and the workload. Interested readers
are encouraged to experiment. ❑

Given the negative scalability of the Schrödinger’s Zoo application across sockets,
why not just run multiple copies of the application, with each copy having a subset of
the animals and confined to run on a single socket?
Answer:
You can do just that! In fact, you can extend this idea to large clustered systems,
running one copy of the application on each node of the cluster. This practice is called
“sharding”, and is heavily used in practice by large web-based retailers [DHJ+ 07].
However, if you are going to shard on a per-socket basis within a multisocket system,
why not buy separate smaller and cheaper single-socket systems, and then run one shard
of the database on each of those systems? ❑

But if elements in a hash table can be removed concurrently with lookups, doesn’t
v2023.06.11a
E.10. DATA STRUCTURES 799
that mean that a lookup could return a reference to a data element that was removed
immediately after it was looked up?
Answer:
Yes it can! This is why hashtab_lookup() must be invoked within an RCU read-side
critical section, and it is why hashtab_add() and hashtab_del() must also use
RCU-aware list-manipulation primitives. Finally, this is why the caller of hashtab_
del() must wait for a grace period (e.g., by calling synchronize_rcu()) before
freeing the removed element. This will ensure that all RCU readers that might reference
the newly removed element have completed before that element is freed. ❑

The hashtorture.h file contains more than 1,000 lines! Is that a comprehensive
test or what???
Answer:
What.
The hashtorture.h tests are a good start and suffice for a textbook algorithm. If
this code was to be used in production, much more testing would be required:
1. Have some subset of elements that always reside in the table, and verify that
lookups always find these elements regardless of the number and type of concurrent
updates in flight.
2. Pair an updater with one or more readers, verifying that after an element is added,
once a reader successfully looks up that element, all later lookups succeed. The
definition of “later” will depend on the table’s consistency requirements.
3. Pair an updater with one or more readers, verifying that after an element is deleted,
once a reader’s lookup of that element fails, all later lookups also fail.
There are many more tests where those came from, the exact nature of which depend
on the details of the requirements on your particular hash table. ❑

How can we be so sure that the hash-table size is at fault here, especially given that
Figure 10.4 on page 295 shows that varying hash-table size has almost no effect?
Might the problem instead be something like false sharing?
Answer:
Excellent question!
False sharing requires writes, which are not featured in the unsynchronized and RCU
runs of this lookup-only benchmark. The problem is therefore not false sharing.
Still unconvinced? Then look at the log-log plot in Figure E.6, which shows
performance for 448 CPUs as a function of the hash-table size, that is, number of buckets
and maximum number of elements. A hash-table of size 1,024 has 1,024 buckets
and contains at most 1,024 elements, with the average occupancy being 512 elements.
Because this is a read-only benchmark, the actual occupancy is always equal to the
average occupancy.
This figure shows near-ideal performance below about 8,000 elements, that is, when
the hash table comprises less than 1 MB of data. This near-ideal performance is
v2023.06.11a
1x108

ideal
1x107
unsync
6
1x10 QSBR,RCU,hazptr
100000
100
1000
10000
100000
1x106
Hash Table Size (Buckets and Maximum Elements)
Figure E.6: Read-Only RCU-Protected Hash-Table Performance For Schrödinger’s

Zoo at 448 CPUs, Varying Table Size
consistent with that for the pre-BSD routing table shown in Figure 9.21 on page 252,
even at 448 CPUs. However, the performance drops significantly (this is a log-log
plot) at about 8,000 elements, which is where the 1,048,576-byte L2 cache overflows.
Performance falls off a cliff (even on this log-log plot) at about 300,000 elements, where
the 40,370,176-byte L3 cache overflows. This demonstrates that the memory-system
bottleneck is profound, degrading performance by well in excess of an order of magnitude
for the large hash tables. This should not be a surprise, as the size-8,388,608 hash table
occupies about 1 GB of memory, overflowing the L3 caches by a factor of 25.
The reason that Figure 10.4 on page 295 shows little effect is that its data was gathered
from bucket-locked hash tables, where locking overhead and contention drowned out
cache-capacity effects. In contrast, both RCU and hazard-pointers readers avoid stores
to shared data, which means that the cache-capacity effects come to the fore.
Still not satisfied? Find a multi-socket system and run this code, making use of
whatever performance-counter hardware is available. This hardware should allow
you to track down the precise cause of any slowdowns exhibited on your particular
system. The experience gained by doing this exercise will be extremely valuable, giving
you a significant advantage over those whose understanding of this issue is strictly
theoretical.10 ❑

The memory system is a serious bottleneck on this big system. Why bother putting
448 CPUs on a system without giving them enough memory bandwidth to do something
useful???
Answer:
It would indeed be a bad idea to use this large and expensive system for a workload
consisting solely of simple hash-table lookups of small data elements. However, this
system is extremely useful for a great many workloads that feature more processing
and less memory accessing. For example, some in-memory databases run extremely
well on this class of system, albeit when running much more complex sets of queries
10 Of course, a theoretical understanding beats no understanding.
v2023.06.11a
than performed by the benchmarks in this chapter. For example, such systems might
be processing images or video streams stored in each element, providing further
performance benefits due to the fact that the resulting sequential memory accesses will
make better use of the available memory bandwidth than will a pure pointer-following
workload.
But let this be a lesson to you. Modern computer systems come in a great many
shapes and sizes, and great care is frequently required to select one that suits your
application. And perhaps even more frequently, significant care and work is required to
adjust your application to the specific computer systems at hand. ❑

The dangers of extrapolating from 28 CPUs to 448 CPUs was made quite clear in
Section 10.2.3. Would extrapolating up from 448 CPUs be any safer?
Answer:
In theory, no, it isn’t any safer, and a useful exercise would be to run these programs on
larger systems. In practice, there are only a very few systems with more than 448 CPUs,
in contrast to the huge number having more than 28 CPUs. This means that although it
is dangerous to extrapolate beyond 448 CPUs, there is very little need to do so.
In addition, other testing has shown that RCU read-side primitives offer consistent
performance and scalability up to at least 1024 CPUs. However, it is useful to review
Figure E.6 and its associated commentary. You see, unlike the 448-CPU system that
provided this data, the system enjoying linear scalability up to 1024 CPUs boasted
excellent memory bandwidth. ❑

How does the code in Listing 10.10 protect against the resizing process progressing
past the selected bucket?
Answer:
It does not provide any such protection. That is instead the job of the update-side
concurrency-control functions described next. ❑

Suppose that one thread is inserting an element into the hash table during a resize
operation. What prevents this insertion from being lost due to a subsequent resize
operation completing before the insertion does?
Answer:
The second resize operation will not be able to move beyond the bucket into which the
insertion is taking place due to the insertion holding the lock(s) on one or both of the
hash buckets in the hash tables. Furthermore, the insertion operation takes place within
an RCU read-side critical section. As we will see when we examine the hashtab_
resize() function, this means that each resize operation uses synchronize_rcu()
invocations to wait for the insertion’s read-side critical section to complete. ❑

The hashtab_lookup() function in Listing 10.12 ignores concurrent resize opera-
v2023.06.11a
Listing E.7: Resizable Hash-Table Access Functions (Fewer Updates)

1 struct ht_elem *
2 hashtab_lookup(struct hashtab *htp_master, void *key)
3 {
4 struct ht *htp;
6
8 htep = ht_search_bucket(htp, key);
9 if (htep)
10 return htep;
12 if (!htp)
13 return NULL;
14 return ht_search_bucket(htp, key);
15 }
16
17 void hashtab_add(struct ht_elem *htep,
19 {
20 struct ht_bucket *htbp = lsp->hbp[0];
22
23 htep->hte_next[!i].prev = NULL;
24 cds_list_add_rcu(&htep->hte_next[i], &htbp->htb_head);
25 }
26
27 void hashtab_del(struct ht_elem *htep,
29 {
31
32 if (htep->hte_next[i].prev) {
33 cds_list_del_rcu(&htep->hte_next[i]);
34 htep->hte_next[i].prev = NULL;
35 }
36 if (lsp->hbp[1] && htep->hte_next[!i].prev) {
37 cds_list_del_rcu(&htep->hte_next[!i]);
38 htep->hte_next[!i].prev = NULL;
39 }
40 }
tions. Doesn’t this mean that readers might miss an element that was previously added
during a resize operation?
Answer:
No. As we will see soon, the hashtab_add() and hashtab_del() functions keep the
old hash table up-to-date while a resize operation is in progress. ❑

The hashtab_add() and hashtab_del() functions in Listing 10.12 can update
two hash buckets while a resize operation is progressing. This might cause poor
performance if the frequency of resize operation is not negligible. Isn’t it possible to
reduce the cost of updates in such cases?
Answer:
Yes, at least assuming that a slight increase in the cost of hashtab_lookup() is
acceptable. One approach is shown in Listings E.7 and E.8 (hash_resize_s.c).
This version of hashtab_add() adds an element to either the old bucket if it is not
resized yet, or to the new bucket if it has been resized, and hashtab_del() removes
the specified element from any buckets into which it has been inserted. The hashtab_
lookup() function searches the new bucket if the search of the old bucket fails,
v2023.06.11a
Listing E.8: Resizable Hash-Table Update-Side Locking Function (Fewer Updates)

1 static void
2 hashtab_lock_mod(struct hashtab *htp_master, void *key,
4 {
5 long b;
6 unsigned long h;
7 struct ht *htp;
9
10 rcu_read_lock();
16 if (b > READ_ONCE(htp->ht_resize_cur)) {
17 lsp->hbp[1] = NULL;
18 return;
19 }
23 lsp->hbp[1] = lsp->hbp[0];
24 lsp->hls_idx[1] = lsp->hls_idx[0];
27 }
which has the disadvantage of adding overhead to the lookup fastpath. The alternative
hashtab_lock_mod() returns the locking state of the new bucket in ->hbp[0] and
->hls_idx[0] if resize operation is in progress, instead of the perhaps more natural
choice of ->hbp[1] and ->hls_idx[1]. However, this less-natural choice has the
advantage of simplifying hashtab_add().
Further analysis of the code is left as an exercise for the reader. ❑

In the hashtab_resize() function in Listing 10.13, what guarantees that the update
to ->ht_new on line 29 will be seen as happening before the update to ->ht_resize_
cur on line 40 from the perspective of hashtab_add() and hashtab_del()? In
other words, what prevents hashtab_add() and hashtab_del() from dereferencing
a NULL pointer loaded from ->ht_new?
Answer:
The synchronize_rcu() on line 30 of Listing 10.13 ensures that all pre-existing RCU
readers have completed between the time that we install the new hash-table reference on
line 29 and the time that we update ->ht_resize_cur on line 40. This means that any
reader that sees a non-negative value of ->ht_resize_cur cannot have started before
the assignment to ->ht_new, and thus must be able to see the reference to the new hash
table.
And this is why the update-side hashtab_add() and hashtab_del() functions
must be enclosed in RCU read-side critical sections, courtesy of hashtab_lock_mod()
and hashtab_unlock_mod() in Listing 10.11. ❑

Why is there a WRITE_ONCE() on line 40 in Listing 10.13?
v2023.06.11a
1x107

1x106
262,144
100000
10000 2,097,152
1000
1 10 100
Figure E.7: Effect of Memory-System Bottlenecks on Hash Tables
Answer:
Together with the READ_ONCE() on line 16 in hashtab_lock_mod() of Listing 10.11,
it tells the compiler that the non-initialization accesses to ->ht_resize_cur must
remain because reads from ->ht_resize_cur really can race with writes, just not in a
way to change the “if” conditions. ❑

How much of the difference in performance between the large and small hash tables
shown in Figure 10.19 was due to long hash chains and how much was due to
memory-system bottlenecks?
Answer:
The easy way to answer this question is to do another run with 2,097,152 elements, but
this time also with 2,097,152 buckets, thus bringing the average number of elements per
bucket back down to unity.
The results are shown by the triple-dashed new trace in the middle of Figure E.7. The
other six traces are identical to their counterparts in Figure 10.19 on page 313. The gap
between this new trace and the lower set of three traces is a rough measure of how much
of the difference in performance was due to hash-chain length, and the gap between
the new trace and the upper set of three traces is a rough measure of how much of that
difference was due to memory-system bottlenecks. The new trace starts out slightly
below its 262,144-element counterpart at a single CPU, showing that cache capacity
is degrading performance slightly even on that single CPU.11 This is to be expected,
given that unlike its smaller counterpart, the 2,097,152-bucket hash table does not fit
into the L3 cache. This new trace rises just past 28 CPUs, which is also to be expected.
This rise is due to the fact that the 29th CPU is on another socket, which brings with it
an additional 39 MB of cache as well as additional memory bandwidth.
But the large hash table’s advantage over that of the hash table with 524,288 buckets
(but still 2,097,152 elements) decreases with additional CPUs, which is consistent
with the bottleneck residing in the memory system. Above about 400 CPUs, the
11 Yes, as far as hardware architects are concerned, caches are part of the memory system.
v2023.06.11a
E.11. VALIDATION 805
2,097,152-bucket hash table is actually outperformed slightly by the 524,288-bucket

hash table. This should not be a surprise because the memory system is the bottleneck
and the larger number of buckets increases this workload’s memory footprint.
The alert reader will have noted the word “rough” above and might be interested in
a more detailed analysis. Such readers are invited to run similar benchmarks, using
whatever performance counters or hardware-analysis tools they might have available.
This can be a long and complex journey, but those brave enough to embark on it will be
rewarded with detailed knowledge of hardware performance and its effect on software.
❑

How much do these specializations really save? Are they really worth it?
Answer:
The answer to the first question is left as an exercise to the reader. Try specializing the
resizable hash table and see how much performance improvement results. The second
question cannot be answered in general, but must instead be answered with respect to a
specific use case. Some use cases are extremely sensitive to performance and scalability,
while others are less so. ❑
E.11 Validation

When in computing is it necessary to follow a fragmentary plan?
Answer:
There are any number of situations, but perhaps the most important situation is when no
one has ever created anything resembling the program to be developed. In this case, the
only way to create a credible plan is to implement the program, create the plan, and
implement it a second time. But whoever implements the program for the first time
has no choice but to follow a fragmentary plan because any detailed plan created in
ignorance cannot survive first contact with the real world.
And perhaps this is one reason why evolution has favored insanely optimistic human
beings who are happy to follow fragmentary plans! ❑

Who cares about the organization? After all, it is the project that is important!
Answer:
Yes, projects are important, but if you like being paid for your work, you need
organizations as well as projects. ❑

Suppose that you are writing a script that processes the output of the time command,
which looks as follows:
v2023.06.11a
real 0m0.132s
user 0m0.040s
sys 0m0.008s
The script is required to check its input for errors, and to give appropriate diagnostics
if fed erroneous time output. What test inputs should you provide to this program to
test it for use with time output generated by single-threaded programs?
Answer:
Can you say “Yes” to all the following questions?
1. Do you have a test case in which all the time is consumed in user mode by a
CPU-bound program?
2. Do you have a test case in which all the time is consumed in system mode by a
CPU-bound program?
3. Do you have a test case in which all three times are zero?
4. Do you have a test case in which the “user” and “sys” times sum to more than the
“real” time? (This would of course be completely legitimate in a multithreaded
program.)
5. Do you have a set of tests cases in which one of the times uses more than one
second?
6. Do you have a set of tests cases in which one of the times uses more than ten
seconds?
7. Do you have a set of test cases in which one of the times has non-zero minutes?
(For example, “15m36.342s”.)
8. Do you have a set of test cases in which one of the times has a seconds value of
greater than 60?
9. Do you have a set of test cases in which one of the times overflows 32 bits of
milliseconds? 64 bits of milliseconds?
10. Do you have a set of test cases in which one of the times is negative?
11. Do you have a set of test cases in which one of the times has a positive minutes
value but a negative seconds value?
12. Do you have a set of test cases in which one of the times omits the “m” or the “s”?
13. Do you have a set of test cases in which one of the times is non-numeric? (For
example, “Go Fish”.)
14. Do you have a set of test cases in which one of the lines is omitted? (For example,
where there is a “real” value and a “sys” value, but no “user” value.)
15. Do you have a set of test cases where one of the lines is duplicated? Or duplicated,
but with a different time value for the duplicate?
16. Do you have a set of test cases where a given line has more than one time value?
(For example, “real 0m0.132s 0m0.008s”.)
v2023.06.11a
17. Do you have a set of test cases containing random characters?
18. In all test cases involving invalid input, did you generate all permutations?
19. For each test case, do you have an expected outcome for that test?
If you did not generate test data for a substantial number of the above cases, you will
need to cultivate a more destructive attitude in order to have a chance of generating
high-quality tests.
Of course, one way to economize on destructiveness is to generate the tests with
the to-be-tested source code at hand, which is called white-box testing (as opposed to
black-box testing). However, this is no panacea: You will find that it is all too easy to
find your thinking limited by what the program can handle, thus failing to generate truly
destructive inputs. ❑

You are asking me to do all this validation BS before I even start coding??? That
sounds like a great way to never get started!!!
Answer:
If it is your project, for example, a hobby, do what you like. Any time you waste will
be your own, and you have no one else to answer to for it. And there is a good chance
that the time will not be completely wasted. For example, if you are embarking on a
first-of-a-kind project, the requirements are in some sense unknowable anyway. In this
case, the best approach might be to quickly prototype a number of rough solutions, try
them out, and see what works best.
On the other hand, if you are being paid to produce a system that is broadly similar
to existing systems, you owe it to your users, your employer, and your future self to
validate early and often. ❑

Are you actually suggesting that it is possible to test correctness into software???
Everyone knows that is impossible!!!
Answer:
Please note that the text used the word “validation” rather than the word “testing”. The
word “validation” includes formal methods as well as testing, for more on which please
see Chapter 12.
But as long as we are bringing up things that everyone should know, let’s remind
ourselves that Darwinian evolution is not about correctness, but rather about survival.
As is software. My goal as a developer is not that my software be attractive from a
theoretical viewpoint, but rather that it survive whatever its users throw at it.
Although the notion of correctness does have its uses, its fundamental limitation is
that the specification against which correctness is judged will also have bugs. This
means nothing more nor less than that traditional correctness proofs prove that the code
in question contains the intended set of bugs!
Alternative definitions of correctness instead focus on the lack of problematic
properties, for example, proving that the software has no use-after-free bugs, no NULL
pointer dereferences, no array-out-of-bounds references, and so on. Make no mistake,
finding and eliminating such classes of bugs can be highly useful. But the fact remains
v2023.06.11a
that the lack of certain classes of bugs does nothing to demonstrate fitness for any
specific purpose.
Therefore, usage-driven validation remains critically important.
Besides, it is also impossible to verify correctness into your software, especially given
the problematic need to verify both the verifier and the specification. ❑

How can you implement WARN_ON_ONCE()?
Answer:
If you don’t mind WARN_ON_ONCE() sometimes warning more than once, simply
maintain a static variable that is initialized to zero. If the condition triggers, check the
variable, and if it is non-zero, return. Otherwise, set it to one, print the message, and
return.
If you really need the message to never appear more than once, you can use an atomic
exchange operation in place of “set it to one” above. Print the message only if the atomic
exchange operation returns zero. ❑

Just what invalid assumptions are you accusing Linux kernel hackers of harboring???
Answer:
Those wishing a complete answer to this question are encouraged to search the Linux
kernel git repository for commits containing the string “Fixes:”. There were many
thousands of them just in the year 2020, including fixes for the following invalid
assumptions:
1. Testing for a non-zero denominator will prevent divide-by-zero errors. (Hint:

Suppose that the test uses 64-bit arithmetic but that the division uses 32-bit
arithmetic.)
2. Userspace can be trusted to zero out versioned data structures used to communicate
with the kernel. (Hint: Sometimes userspace has no idea how large the data
structure is.)
3. Outdated TCP duplicate selective acknowledgement (D-SACK) packets can be

completely ignored. (Hint: These packets might also contain other information.)
4. All CPUs are little-endian.
5. Once a data structure is no longer needed, all of its memory may be immediately
freed.
6. All devices can be initialized while in standby mode.
7. Developers can be trusted to consistently do correct hexidecimal arithmetic.
Those who look at these commits in greater detail will conclude that invalid assump-
tions are the rule, not the exception. ❑
v2023.06.11a

Why would anyone bother copying existing code in pen on paper??? Doesn’t that just
increase the probability of transcription errors?
Answer:
If you are worried about transcription errors, please allow me to be the first to introduce
you to a really cool tool named diff. In addition, carrying out the copying can be quite
valuable:
1. If you are copying a lot of code, you are probably failing to take advantage of an
opportunity for abstraction. The act of copying code can provide great motivation
for abstraction.
2. Copying the code gives you an opportunity to think about whether the code really
works in its new setting. Is there some non-obvious constraint, such as the need to
disable interrupts or to hold some lock?
3. Copying the code also gives you time to consider whether there is some better way
to get the job done.
So, yes, copy the code! ❑

This procedure is ridiculously over-engineered! How can you expect to get a reasonable
amount of software written doing it this way???
Answer:
Indeed, repeatedly copying code by hand is laborious and slow. However, when
combined with heavy-duty stress testing and proofs of correctness, this approach is also
extremely effective for complex parallel code where ultimate performance and reliability
are required and where debugging is difficult. The Linux-kernel RCU implementation
is a case in point.
On the other hand, if you are writing a simple single-threaded shell script, then you
would be best-served by a different methodology. For example, enter each command
one at a time into an interactive shell with a test data set to make sure that it does what
you want, then copy-and-paste the successful commands into your script. Finally, test
the script as a whole.
If you have a friend or colleague who is willing to help out, pair programming can
work very well, as can any number of formal design- and code-review processes.
And if you are writing code as a hobby, then do whatever you like.
In short, different types of software need different development methodologies. ❑

What do you do if, after all the pen-on-paper copying, you find a bug while typing in
the resulting code?
Answer:
The answer, as is often the case, is “it depends”. If the bug is a simple typo, fix that
typo and continue typing. However, if the bug indicates a design flaw, go back to pen
and paper. ❑
v2023.06.11a

Wait! Why on earth would an abstract piece of software fail only sometimes???
Answer:
Because complexity and concurrency can produce results that are indistinguishable
from randomness [MOZ09]. For example, a bug in Linux-kernel RCU required the
following to hold before that bug would manifest:
1. The kernel was built for HPC or real-time use, so that a given CPU’s RCU work
could be offloaded to some other CPU.
2. An offloaded CPU went offline just after generating a large quantity of RCU work.
3. A special rcu_barrier() API was invoked just at this time.
4. The RCU work from the newly offlined CPU was still being processed after
rcu_barrier() returned.
5. One of these remaining RCU work items was related to the code invoking the
rcu_barrier().
Making this bug manifest therefore required considerable luck or great testing skill. But
the testing skill could be effective only if the bug was known, which of course it was
not. Therefore, the manifesting of this bug was very well modeled as a probabilistic
process. ❑

Suppose that you had a very large number of systems at your disposal. For example, at
current cloud prices, you can purchase a huge amount of CPU time at low cost. Why
not use this approach to get close enough to certainty for all practical purposes?
Answer:
This approach might well be a valuable addition to your validation arsenal. But it does
have limitations that rule out “for all practical purposes”:
1. Some bugs have extremely low probabilities of occurrence, but nevertheless need
to be fixed. For example, suppose that the Linux kernel’s RCU implementation
had a bug that is triggered only once per million years of machine time on average.
A million years of CPU time is hugely expensive even on the cheapest cloud
platforms, but we could expect this bug to result in more than 50 failures per day
on the more than 20 billion Linux instances in the world as of 2017.
2. The bug might well have zero probability of occurrence on your particular cloud-
computing test setup, which means that you won’t see it no matter how much
machine time you burn testing it. For but one example, there are RCU bugs that
appear only in preemptible kernels, and also other RCU bugs that appear only in
non-preemptible kernels.
Of course, if your code is small enough, formal validation may be helpful, as discussed
in Chapter 12. But beware: Formal validation of your code will not find errors in your
assumptions, misunderstanding of the requirements, misunderstanding of the software
or hardware primitives you use, or errors that you did not think to construct a proof for.
❑
v2023.06.11a

Say what??? When I plug the earlier five-test 10 %-failure-rate example into the
formula, I get 59,050 % and that just doesn’t make sense!!!
Answer:
You are right, that makes no sense at all.
Remember that a probability is a number between zero and one, so that you need to
divide a percentage by 100 to get a probability. So 10 % is a probability of 0.1, which
gets a probability of 0.4095, which rounds to 41 %, which quite sensibly matches the
earlier result. ❑

In Eq. 11.6, are the logarithms base-10, base-2, or base-e?
Answer:
It does not matter. You will get the same answer no matter what base of logarithms you
use because the result is a pure ratio of logarithms. The only constraint is that you use
the same base for both the numerator and the denominator. ❑

Suppose that a bug causes a test failure three times per hour on average. How long
must the test run error-free to provide 99.9 % confidence that the fix significantly
reduced the probability of failure?
Answer:
We set 𝑛 to 3 and 𝑃 to 99.9 in Eq. 11.11, resulting in:
1 100 − 99.9
𝑇 = − ln = 2.3 (E.9)
3 100
If the test runs without failure for 2.3 hours, we can be 99.9 % certain that the fix
reduced the probability of failure. ❑

Doing the summation of all the factorials and exponentials is a real pain. Isn’t there
an easier way?
Answer:
One approach is to use the open-source symbolic manipulation program named “max-
ima”. Once you have installed this program, which is a part of many Linux distributions,
you can run it and give the load(distrib); command followed by any number of
bfloat(cdf_poisson(m,l)); commands, where the m is replaced by the desired
value of 𝑚 (the actual number of failures in actual test) and the l is replaced by the
desired value of 𝜆 (the expected number of failures in the actual test).
In particular, the bfloat(cdf_poisson(2,24)); command results in
1.181617112359357b-8, which matches the value given by Eq. 11.13.
Another approach is to recognize that in this real world, it is not all that useful to
compute (say) the duration of a test having two or fewer errors that would give a 76.8 %
confidence of a 349.2x improvement in reliability. Instead, human beings tend to focus
on specific values, for example, a 95 % confidence of a 10x improvement. People also
v2023.06.11a
Table E.4: Human-Friendly Poisson-Function Display
Improvement
Certainty (%) Any 10x 100x
90.0 2.3 23.0 230.0

95.0 3.0 30.0 300.0
99.0 4.6 46.1 460.5
99.9 6.9 69.1 690.7
greatly prefer error-free test runs, and so should you because doing so reduces your
required test durations. Therefore, it is quite possible that the values in Table E.4 will
suffice. Simply look up the desired confidence and degree of improvement, and the
resulting number will give you the required error-free test duration in terms of the
expected time for a single error to appear. So if your pre-fix testing suffered one failure
per hour, and the powers that be require a 95 % confidence of a 10x improvement, you
need a 30-hour error-free run.
Alternatively, you can use the rough-and-ready method described in Section 11.6.2.
❑

But wait!!! Given that there has to be some number of failures (including the possibility
of zero failures), shouldn’t Eq. 11.13 approach the value 1 as 𝑚 goes to infinity?
Answer:
Indeed it should. And it does.
To see this, note that e−𝜆 does not depend on 𝑖, which means that it can be pulled out
of the summation as follows:
∞
∑︁ 𝜆𝑖
e−𝜆 (E.10)
𝑖=0
𝑖!
The remaining summation is exactly the Taylor series for e𝜆 , yielding:
e−𝜆 e𝜆 (E.11)
The two exponentials are reciprocals, and therefore cancel, resulting in exactly 1, as
required. ❑

How is this approach supposed to help if the corruption affected some unrelated
pointer, which then caused the corruption???
Answer:
Indeed, that can happen. Many CPUs have hardware-debugging facilities that can help
you locate that unrelated pointer. Furthermore, if you have a core dump, you can search
the core dump for pointers referencing the corrupted region of memory. You can also
look at the data layout of the corruption, and check pointers whose type matches that
layout.
v2023.06.11a
You can also step back and test the modules making up your program more intensively,
which will likely confine the corruption to the module responsible for it. If this makes
the corruption vanish, consider adding additional argument checking to the functions
exported from each module.
Nevertheless, this is a hard problem, which is why I used the words “a bit of a dark
art”. ❑

But I did the bisection, and ended up with a huge commit. What do I do now?
Answer:
A huge commit? Shame on you! This is but one reason why you are supposed to keep
the commits small.
And that is your answer: Break up the commit into bite-sized pieces and bisect the
pieces. In my experience, the act of breaking up the commit is often sufficient to make
the bug painfully obvious. ❑

Why don’t conditional-locking primitives provide this spurious-failure functionality?
Answer:
There are locking algorithms that depend on conditional-locking primitives telling them
the truth. For example, if conditional-lock failure signals that some other thread is
already working on a given job, spurious failure might cause that job to never get done,
possibly resulting in a hang. ❑

That is ridiculous!!! After all, isn’t getting the correct answer later than one would
like better than getting an incorrect answer???
Answer:
This question fails to consider the option of choosing not to compute the answer at all,
and in doing so, also fails to consider the costs of computing the answer. For example,
consider short-term weather forecasting, for which accurate models exist, but which
require large (and expensive) clustered supercomputers, at least if you want to actually
run the model faster than the weather.
And in this case, any performance bug that prevents the model from running faster
than the actual weather prevents any forecasting. Given that the whole purpose of
purchasing the large clustered supercomputers was to forecast weather, if you cannot run
the model faster than the weather, you would be better off not running the model at all.
More severe examples may be found in the area of safety-critical real-time computing.
❑
v2023.06.11a

But if you are going to put in all the hard work of parallelizing an application, why
not do it right? Why settle for anything less than optimal performance and linear
scalability?
Answer:
Although I do heartily salute your spirit and aspirations, you are forgetting that there
may be high costs due to delays in the program’s completion. For an extreme example,
suppose that a 40 % performance shortfall from a single-threaded application is causing
one person to die each day. Suppose further that in a day you could hack together a
quick and dirty parallel program that ran 50 % faster on an eight-CPU system than the
sequential version, but that an optimal parallel program would require four months of
painstaking design, coding, debugging, and tuning.
It is safe to say that more than 100 people would prefer the quick and dirty version. ❑

But what about other sources of error, for example, due to interactions between caches
and memory layout?
Answer:
Changes in memory layout can indeed result in unrealistic decreases in execution time.
For example, suppose that a given microbenchmark almost always overflows the L0
cache’s associativity, but with just the right memory layout, it all fits. If this is a real
concern, consider running your microbenchmark using huge pages (or within the kernel
or on bare metal) in order to completely control the memory layout.
But note that there are many different possible memory-layout bottlenecks. Bench-
marks sensitive to memory bandwidth (such as those involving matrix arithmetic) should
spread the running threads across the available cores and sockets to maximize memory
parallelism. They should also spread the data across NUMA nodes, memory controllers,
and DRAM chips to the extent possible. In contrast, benchmarks sensitive to memory
latency (including most poorly scaling applications) should instead maximize locality,
filling each core and socket in turn before adding another one. ❑

Wouldn’t the techniques suggested to isolate the code under test also affect that code’s
performance, particularly if it is running within a larger application?
Answer:
Indeed it might, although in most microbenchmarking efforts you would extract the
code under test from the enclosing application. Nevertheless, if for some reason you
must keep the code under test within the application, you will very likely need to use
the techniques discussed in Section 11.7.6. ❑

This approach is just plain weird! Why not use means and standard deviations, like
we were taught in our statistics classes?
Answer:
Because mean and standard deviation were not designed to do this job. To see this, try
v2023.06.11a
E.12. FORMAL VERIFICATION 815
applying mean and standard deviation to the following data set, given a 1 % relative
error in measurement:
49,548.4 49,549.4 49,550.2 49,550.9 49,550.9 49,551.0 49,551.5 49,552.1

49,899.0 49,899.3 49,899.7 49,899.8 49,900.1 49,900.4 52,244.9 53,333.3
53,333.3 53,706.3 53,706.3 54,084.5
The problem is that mean and standard deviation do not rest on any sort of measurement-
error assumption, and they will therefore see the difference between the values near
49,500 and those near 49,900 as being statistically significant, when in fact they are
well within the bounds of estimated measurement error.
Of course, it is possible to create a script similar to that in Listing 11.2 that uses
standard deviation rather than absolute difference to get a similar effect, and this is left
as an exercise for the interested reader. Be careful to avoid divide-by-zero errors arising
from strings of identical data values! ❑

But what if all the y-values in the trusted group of data are exactly zero? Won’t that
cause the script to reject any non-zero value?
Answer:
Indeed it will! But if your performance measurements often produce a value of exactly
zero, perhaps you need to take a closer look at your performance-measurement code.
Note that many approaches based on mean and standard deviation will have similar
problems with this sort of dataset. ❑
E.12 Formal Verification

Why is there an unreached statement in locker? After all, isn’t this a full state-space
search?
Answer:
The locker process is an infinite loop, so control never reaches the end of this process.
However, since there are no monotonically increasing variables, Promela is able to
model this infinite loop with a small number of states. ❑

What are some Promela code-style issues with this example?
Answer:
There are several:
1. The declaration of sum should be moved to within the init block, since it is not
used anywhere else.
2. The assertion code should be moved outside of the initialization loop. The
initialization loop can then be placed in an atomic block, greatly reducing the state
space (by how much?).
v2023.06.11a
3. The atomic block covering the assertion code should be extended to include the
initialization of sum and j, and also to cover the assertion. This also reduces the
state space (again, by how much?). ❑

Is there a more straightforward way to code the do-od statement?
Answer:
Yes. Replace it with if-fi and remove the two break statements. ❑

Why are there atomic blocks at lines 12–21 and lines 44–56, when the operations
within those atomic blocks have no atomic implementation on any current production
microprocessor?
Answer:
Because those operations are for the benefit of the assertion only. They are not part of
the algorithm itself. There is therefore no harm in marking them atomic, and so marking
them greatly reduces the state space that must be searched by the Promela model. ❑

Is the re-summing of the counters on lines 24–27 really necessary?
Answer:
Yes. To see this, delete these lines and run the model.
Alternatively, consider the following sequence of steps:
1. One process is within its RCU read-side critical section, so that the value of ctr[0]
is zero and the value of ctr[1] is two.
2. An updater starts executing, and sees that the sum of the counters is two so that the
fastpath cannot be executed. It therefore acquires the lock.
3. A second updater starts executing, and fetches the value of ctr[0], which is zero.
4. The first updater adds one to ctr[0], flips the index (which now becomes zero),
then subtracts one from ctr[1] (which now becomes one).
5. The second updater fetches the value of ctr[1], which is now one.
6. The second updater now incorrectly concludes that it is safe to proceed on the
fastpath, despite the fact that the original reader has not yet completed. ❑
v2023.06.11a
Listing E.9: Spin Output Diff of -DCOLLAPSE and -DMA=88

@@ -1,6 +1,6 @@
(Spin Version 6.4.6 -- 2 December 2016)
- + Compression
+ + Graph Encoding (-DMA=88)

@@ -9,27 +9,22 @@

+MA stats: -DMA=77 is sufficient
+Minimized Automaton: 2084798 nodes and 6.38445e+06 edges
1.8620286e+08 states, stored
1.7759831e+08 states, matched
3.6380117e+08 transitions (= stored+matched)
1.3724093e+08 atomic steps
-hash conflicts: 1.1445626e+08 (resolved)

- 8418.559 actual memory usage for states
- (compression: 40.87%)
- state-vector as stored =
- 19 byte + 28 byte overhead
- 2048.000 memory used for hash table (-w28)
+ 204.907 actual memory usage for states
+ (compression: 0.99%)
- 1.509 memory lost to fragmentation
-10482.675 total actual memory usage
+ 222.388 total actual memory usage
-nr of templates: [ 0:globals 1:chans 2:procs ]

-collapse counts: [ 0:1021 2:32 3:1869 4:2 ]
unreached in proctype qrcu_reader
(0 of 18 states)
unreached in proctype qrcu_updater
@@ -38,5 +33,5 @@
unreached in init
(0 of 23 states)
-pan: elapsed time 369 seconds

-pan: rate 505107.58 states/second
+pan: elapsed time 2.68e+03 seconds
+pan: rate 69453.282 states/second
v2023.06.11a

A compression rate of 0.48 % corresponds to a 200-to-1 decrease in memory occupied
by the states! Is the state-space search really exhaustive???
Answer:
According to Spin’s documentation, yes, it is.
As an indirect evidence, let’s compare the results of runs with -DCOLLAPSE and with
-DMA=88 (two readers and three updaters). The diff of outputs from those runs is shown
in Listing E.9. As you can see, they agree on the numbers of states (stored and matched).
❑

But different formal-verification tools are often designed to locate particular classes
of bugs. For example, very few formal-verification tools will find an error in the
specification. So isn’t this “clearly untrustworthy” judgment a bit harsh?
Answer:
It is certainly true that many formal-verification tools are specialized in some way.
For example, Promela does not handle realistic memory models (though they can be
programmed into Promela [DMD13]), CBMC [CKL04] does not detect probabilistic
hangs and deadlocks, and Nidhugg [LSLK14] does not detect bugs involving data
nondeterminism. But this means that these tools cannot be trusted to find bugs that they
are not designed to locate.
And therefore people creating formal-verification tools should “tell the truth on
the label”, clearly calling out what classes of bugs their tools can and cannot detect.
Otherwise, the first time a practitioner finds a tool failing to detect a bug, that practitioner
is likely to make extremely harsh and extremely public denunciations of that tool. Yes,
yes, there is something to be said for putting your best foot forward, but putting it too
far forward without appropriate disclaimers can easily trigger a land mine of negative
reaction that your tool might or might not be able to recover from.
You have been warned! ❑

Given that we have two independent proofs of correctness for the QRCU algorithm
described herein, and given that the proof of incorrectness covers what is known to be
a different algorithm, why is there any room for doubt?
Answer:
There is always room for doubt. In this case, it is important to keep in mind that the
two proofs of correctness preceded the formalization of real-world memory models,
raising the possibility that these two proofs are based on incorrect memory-ordering
assumptions. Furthermore, since both proofs were constructed by the same person, it is
quite possible that they contain a common error. Again, there is always room for doubt.
❑
v2023.06.11a

Yeah, that’s just great! Now, just what am I supposed to do if I don’t happen to have a
machine with 40 GB of main memory???
Answer:
Relax, there are a number of lawful answers to this question:
1. Try compiler flags -DCOLLAPSE and -DMA=N to reduce memory consumption. See
Section 12.1.4.1.
2. Further optimize the model, reducing its memory consumption.
3. Work out a pencil-and-paper proof, perhaps starting with the comments in the code
in the Linux kernel.
4. Devise careful torture tests, which, though they cannot prove the code correct, can
find hidden bugs.
5. There is some movement towards tools that do model checking on clusters of
smaller machines. However, please note that we have not actually used such tools
myself, courtesy of some large machines that Paul has occasional access to.
6. Wait for memory sizes of affordable systems to expand to fit your problem.
7. Use one of a number of cloud-computing services to rent a large system for a short
time period. ❑

Why not simply increment rcu_update_flag, and then only increment dynticks_
progress_counter if the old value of rcu_update_flag was zero???
Answer:
This fails in presence of NMIs. To see this, suppose an NMI was received just
after rcu_irq_enter() incremented rcu_update_flag, but before it incremented
dynticks_progress_counter. The instance of rcu_irq_enter() invoked by the
NMI would see that the original value of rcu_update_flag was non-zero, and would
therefore refrain from incrementing dynticks_progress_counter. This would leave
the RCU grace-period machinery no clue that the NMI handler was executing on this
CPU, so that any RCU read-side critical sections in the NMI handler would lose their
RCU protection.
The possibility of NMI handlers, which, by definition cannot be masked, does
complicate this code. ❑

But if line 7 finds that we are the outermost interrupt, wouldn’t we always need to
increment dynticks_progress_counter?
Answer:
Not if we interrupted a running task! In that case, dynticks_progress_counter
would have already been incremented by rcu_exit_nohz(), and there would be no
need to increment it again. ❑
v2023.06.11a

Can you spot any bugs in any of the code in this section?
Answer:
Read the next section to see if you were correct. ❑

Why isn’t the memory barrier in rcu_exit_nohz() and rcu_enter_nohz() mod-
eled in Promela?
Answer:
Promela assumes sequential consistency, so it is not necessary to model memory barriers.
In fact, one must instead explicitly model lack of memory barriers, for example, as
shown in Listing 12.13 on page 369. ❑

Isn’t it a bit strange to model rcu_exit_nohz() followed by rcu_enter_nohz()?
Wouldn’t it be more natural to instead model entry before exit?
Answer:
It probably would be more natural, but we will need this particular order for the liveness
checks that we will add later. ❑

Wait a minute! In the Linux kernel, both dynticks_progress_counter and rcu_
dyntick_snapshot are per-CPU variables. So why are they instead being modeled
as single global variables?
Answer:
Because the grace-period code processes each CPU’s dynticks_progress_counter
and rcu_dyntick_snapshot variables separately, we can collapse the state onto a
single CPU. If the grace-period code were instead to do something special given
specific values on specific CPUs, then we would indeed need to model multiple CPUs.
But fortunately, we can safely confine ourselves to two CPUs, the one running the
grace-period processing and the one entering and leaving dynticks-idle mode. ❑

Given there are a pair of back-to-back changes to grace_period_state on lines 25
and 26, how can we be sure that line 25’s changes won’t be lost?
Answer:
Recall that Promela and Spin trace out every possible sequence of state changes.
Therefore, timing is irrelevant: Promela/Spin will be quite happy to jam the entire
rest of the model between those two statements unless some state variable specifically
prohibits doing so. ❑
v2023.06.11a

But what would you do if you needed the statements in a single EXECUTE_MAINLINE()
group to execute non-atomically?
Answer:
The easiest thing to do would be to put each such statement in its own EXECUTE_
MAINLINE() statement. ❑

But what if the dynticks_nohz() process had “if” or “do” statements with condi-
tions, where the statement bodies of these constructs needed to execute non-atomically?
Answer:
One approach, as we will see in a later section, is to use explicit labels and “goto”
statements. For example, the construct:
if
:: i == 0 -> a = -1;
:: else -> a = -2;
fi;
could be modeled as something like:
EXECUTE_MAINLINE(stmt1,
if
:: i == 0 -> goto stmt1_then;
:: else -> goto stmt1_else;
fi)
stmt1_then: skip;
EXECUTE_MAINLINE(stmt1_then1, a = -1; goto stmt1_end)
stmt1_else: skip;
EXECUTE_MAINLINE(stmt1_then1, a = -2)
stmt1_end: skip;
However, it is not clear that the macro is helping much in the case of the “if”
statement, so these sorts of situations will be open-coded in the following sections. ❑

Why are lines 46 and 47 (the “in_dyntick_irq = 0;” and the “i++;”) executed
atomically?
Answer:
These lines of code pertain to controlling the model, not to the code being modeled, so
there is no reason to model them non-atomically. The motivation for modeling them
atomically is to reduce the size of the state space. ❑

What property of interrupts is this dynticks_irq() process unable to model?
Answer:
One such property is nested interrupts, which are handled in the following section. ❑
v2023.06.11a

Does Paul always write his code in this painfully incremental manner?
Answer:
Not always, but more and more frequently. In this case, Paul started with the smallest
slice of code that included an interrupt handler, because he was not sure how best to
model interrupts in Promela. Once he got that working, he added other features. (But if
he was doing it again, he would start with a “toy” handler. For example, he might have
the handler increment a variable twice and have the mainline code verify that the value
was always even.)
Why the incremental approach? Consider the following, attributed to Brian W.
Kernighan:
Debugging is twice as hard as writing the code in the first place. Therefore, if
you write the code as cleverly as possible, you are, by definition, not smart
enough to debug it.
This means that any attempt to optimize the production of code should place at least
66 % of its emphasis on optimizing the debugging process, even at the expense of
increasing the time and effort spent coding. Incremental coding and testing is one way
to optimize the debugging process, at the expense of some increase in coding effort.
Paul uses this approach because he rarely has the luxury of devoting full days (let alone
weeks) to coding and debugging. ❑

But what happens if an NMI handler starts running before an IRQ handler completes,
and if that NMI handler continues running until a second IRQ handler starts?
Answer:
This cannot happen within the confines of a single CPU. The first IRQ handler cannot
complete until the NMI handler returns. Therefore, if each of the dynticks and
dynticks_nmi variables have taken on an even value during a given time interval, the
corresponding CPU really was in a quiescent state at some time during that interval. ❑

This is still pretty complicated. Why not just have a cpumask_t with per-CPU bits,
clearing the bit when entering an IRQ or NMI handler, and setting it upon exit?
Answer:
Although this approach would be functionally correct, it would result in excessive IRQ
entry/exit overhead on large machines. In contrast, the approach laid out in this section
allows each CPU to touch only per-CPU data on IRQ and NMI entry/exit, resulting in
much lower IRQ entry/exit overhead, especially on large machines. ❑
v2023.06.11a

But x86 has strong memory ordering, so why formalize its memory model?
Answer:
Actually, academics consider the x86 memory model to be weak because it can allow
prior stores to be reordered with subsequent loads. From an academic viewpoint, a
strong memory model is one that allows absolutely no reordering, so that all threads
agree on the order of all operations visible to them.
Plus it really is the case that developers are sometimes confused about x86 memory
ordering. ❑

Why does line 8 of Listing 12.23 initialize the registers? Why not instead initialize
them on lines 4 and 5?
Answer:
Either way works. However, in general, it is better to use initialization than explicit
instructions. The explicit instructions are used in this example to demonstrate their use.
In addition, many of the litmus tests available on the tool’s web site (https://www.cl.
cam.ac.uk/~pes20/ppcmem/) were automatically generated, which generates explicit
initialization instructions. ❑

But whatever happened to line 17 of Listing 12.23, the one that is the Fail1: label?
Answer:
The implementation of PowerPC version of atomic_add_return() loops when the
stwcx instruction fails, which it communicates by setting non-zero status in the
condition-code register, which in turn is tested by the bne instruction. Because actually
modeling the loop would result in state-space explosion, we instead branch to the
Fail1: label, terminating the model with the initial value of 2 in P0’s r3 register, which
will not trigger the exists assertion.
There is some debate about whether this trick is universally applicable, but I have not
seen an example where it fails. ❑

Does the Arm Linux kernel have a similar bug?
Answer:
Arm does not have this particular bug because it places smp_mb() before and after the
atomic_add_return() function’s assembly-language implementation. PowerPC no
longer has this bug; it has long since been fixed [Her11]. ❑
v2023.06.11a

Does the lwsync on line 10 in Listing 12.23 provide sufficient ordering?
Answer:
It depends on the semantics required. The rest of this answer assumes that the assembly
language for P0 in Listing 12.23 is supposed to implement a value-returning atomic
operation.
As is discussed in Chapter 15, Linux kernel’s memory consistency model requires
value-returning atomic RMW operations to be fully ordered on both sides. The ordering
provided by lwsync is insufficient for this purpose, and so sync should be used instead.
This change has since been made [Fen15] in response to an email thread discussing a
couple of other litmus tests [McK15g]. Finding any other bugs that the Linux kernel
might have is left as an exercise for the reader.
In other enviroments providing weaker semantics, lwsync might be sufficient. But
not for the Linux kernel’s value-returning atomic operations! ❑

What do you have to do to run herd on litmus tests like that shown in Listing 12.29?
Answer:
Get version v4.17 (or later) of the Linux-kernel source code, then follow the instructions
in tools/memory-model/README to install the needed tools. Then follow the further
instructions to run these tools on the litmus test of your choice. ❑

Why bother modeling locking directly? Why not simply emulate locking with atomic
operations?
Answer:
In a word, performance, as can be seen in Table E.5. The first column shows the number
of herd processes modeled. The second column shows the herd runtime when modeling
spin_lock() and spin_unlock() directly in herd’s cat language. The third column
shows the herd runtime when emulating spin_lock() with cmpxchg_acquire()
and spin_unlock() with smp_store_release(), using the herd filter clause
to reject executions that fail to acquire the lock. The fourth column is like the third,
but using xchg_acquire() instead of cmpxchg_acquire(). The fifth and sixth
columns are like the third and fourth, but instead using the herd exists clause to reject
executions that fail to acquire the lock.
Note also that use of the filter clause is about twice as fast as is use of the exists
clause. This is no surprise because the filter clause allows early abandoning of
excluded executions, where the executions that are excluded are the ones in which the
lock is concurrently held by more than one process.
More important, modeling spin_lock() and spin_unlock() directly ranges from
five times faster to more than two orders of magnitude faster than modeling emulated
locking. This should also be no surprise, as direct modeling raises the level of abstraction,
thus reducing the number of events that herd must model. Because almost everything
that herd does is of exponential computational complexity, modest reductions in the
number of events produces exponentially large reductions in runtime.
v2023.06.11a
Table E.5: Locking: Modeling vs. Emulation Time (s)
Model Emulate
# Proc.
filter exists
cmpxchg xchg cmpxchg xchg
2 0.004 0.022 0.027 0.039 0.058
3 0.041 0.743 0.968 1.653 3.203
4 0.374 59.565 74.818 151.962 500.960
5 4.905
Thus, in formal verification even more than in parallel programming itself, divide
and conquer!!! ❑

Wait!!! Isn’t leaking pointers out of an RCU read-side critical section a critical bug???
Answer:
Yes, it usually is a critical bug. However, in this case, the updater has been cleverly
constructed to properly handle such pointer leaks. But please don’t make a habit of
doing this sort of thing, and especially don’t do this without having put a lot of thought
into making some more conventional approach work. ❑

In Listing 12.32, why couldn’t a reader fetch c just before P1() zeroed it on line 45,
and then later store this same value back into c just after it was zeroed, thus defeating
the zeroing operation?
Answer:
Because the reader advances to the next element on line 24, thus avoiding storing a
pointer to the same element as was fetched. ❑

In Listing 12.32, why not have just one call to synchronize_rcu() immediately
before line 48?
Answer:
Because this results in P0() accessing a freed element. But don’t take my word for this,
try it out in herd! ❑

Also in Listing 12.32, can’t line 48 be WRITE_ONCE() instead of smp_store_
release()?
Answer:
That is an excellent question. As of late 2021, the answer is “no one knows”. Much
depends on the semantics of Armv8’s conditional-move instruction. While awaiting
clarity on these semantics, smp_store_release() is the safe choice. ❑
v2023.06.11a

But shouldn’t sufficiently low-level software be for all intents and purposes immune to
being exploited by black hats?
Answer:
Unfortunately, no.
At one time, Paul E. McKenny felt that Linux-kernel RCU was immune to such
exploits, but the advent of Row Hammer showed him otherwise. After all, if the black
hats can hit the system’s DRAM, they can hit any and all low-level software, even
including RCU.
And in 2018, this possibility passed from the realm of theoretical speculation into the
hard and fast realm of objective reality [McK19a]. ❑

In light of the full verification of the L4 microkernel, isn’t this limited view of formal
verification just a little bit obsolete?
Answer:
Unfortunately, no.
The first full verification of the L4 microkernel was a tour de force, with a large
number of Ph.D. students hand-verifying code at a very slow per-student rate. This level
of effort could not be applied to most software projects because the rate of change is
just too great. Furthermore, although the L4 microkernel is a large software artifact
from the viewpoint of formal verification, it is tiny compared to a great number of
projects, including LLVM, GCC, the Linux kernel, Hadoop, MongoDB, and a great
many others. In addition, this verification did have limits, as the researchers freely ad-
mit, to their credit: https://docs.sel4.systems/projects/sel4/frequently-
asked-questions.html#does-sel4-have-zero-bugs.
Although formal verification is finally starting to show some promise, including
more-recent L4 verifications involving greater levels of automation, it currently has no
chance of completely displacing testing in the foreseeable future. And although I would
dearly love to be proven wrong on this point, please note that such proof will be in the
form of a real tool that verifies real software, not in the form of a large body of rousing
rhetoric.
Perhaps someday formal verification will be used heavily for validation, including for
what is now known as regression testing. Section 17.4 looks at what would be required
to make this possibility a reality. ❑
E.13 Putting It All Together

Why not implement reference-acquisition using a simple compare-and-swap operation
that only acquires a reference if the reference counter is non-zero?
Answer:
Although this can resolve the race between the release of the last reference and acquisition
of a new reference, it does absolutely nothing to prevent the data structure from being
v2023.06.11a
E.13. PUTTING IT ALL TOGETHER 827
freed and reallocated, possibly as some completely different type of structure. It is quite
likely that the “simple compare-and-swap operation” would give undefined results if
applied to the differently typed structure.
In short, use of atomic operations such as compare-and-swap absolutely requires
either type-safety or existence guarantees.
But what if it is absolutely necessary to let the type change?
One approach is for each such type to have the reference counter at the same location,
so that as long as the reallocation results in an object from this group of types, all is
well. If you do this in C, make sure you comment the reference counter in each structure
in which it appears. In C++, use inheritance and templates. ❑

Why isn’t it necessary to guard against cases where one CPU acquires a reference just
after another CPU releases the last reference?
Answer:
Because a CPU must already hold a reference in order to legally acquire another
reference. Therefore, if one CPU releases the last reference, there had better not be any
CPU acquiring a new reference! ❑

Suppose that just after the atomic_sub_and_test() on line 22 of Listing 13.2 is
invoked, that some other CPU invokes kref_get(). Doesn’t this result in that other
CPU now having an illegal reference to a released object?
Answer:
This cannot happen if these functions are used correctly. It is illegal to invoke kref_
get() unless you already hold a reference, in which case the kref_sub() could not
possibly have decremented the counter to zero. ❑

Suppose that kref_sub() returns zero, indicating that the release() function was
not invoked. Under what conditions can the caller rely on the continued existence of
the enclosing object?
Answer:
The caller cannot rely on the continued existence of the object unless it knows that at
least one reference will continue to exist. Normally, the caller will have no way of
knowing this, and must therefore carefully avoid referencing the object after the call to
kref_sub().
Interested readers are encouraged to work around this limitation using RCU, in
particular, call_rcu(). ❑
v2023.06.11a

Why not just pass kfree() as the release function?
Answer:
Because the kref structure normally is embedded in a larger structure, and it is necessary
to free the entire structure, not just the kref field. This is normally accomplished by
defining a wrapper function that does a container_of() and then a kfree(). ❑

Why can’t the check for a zero reference count be made in a simple “if” statement
with an atomic increment in its “then” clause?
Answer:
Suppose that the “if” condition completed, finding the reference counter value equal
to one. Suppose that a release operation executes, decrementing the reference counter
to zero and therefore starting cleanup operations. But now the “then” clause can
increment the counter back to a value of one, allowing the object to be used after it has
been cleaned up.
This use-after-cleanup bug is every bit as bad as a full-fledged use-after-free bug. ❑

Why don’t all sequence-locking use cases replicate the data in this fashion?
Answer:
Such replication is impractical if the data is too large, as it might be in the Schrödinger’s-
zoo example described in Section 13.4.2.
Such replication is unnecessary if delays are prevented, for example, when updaters
disable interrupts when running on bare-metal hardware (that is, without the use of a
vCPU-preemption-prone hypervisor).
Alternatively, if readers can tolerate the occasional delay, then replication is again
unnecessary. Consider the example of reader-writer locking, where writers always delay
readers and vice versa.
However, if the data to be replicated is reasonably small, if delays are possible, and if
readers cannot tolerate these delays, replicating the data is an excellent approach. ❑

Is it possible to write-acquire the sequence lock on the new element before it is inserted
instead of acquiring that of the old element before it is removed?
Answer:
Yes, and the details are left as an exercise to the reader.
The term tombstone is sometimes used to refer to the element with the old name after
its sequence lock is acquired. Similarly, the term birthstone is sometimes used to refer
to the element with the new name while its sequence lock is still held. ❑
v2023.06.11a

Is it possible to avoid the global lock?
Answer:
Yes, and one way to do this would be to use per-hash-chain locks. The updater could
acquire lock(s) corresponding to both the old and the new element, acquiring them in
address order. In this case, the insertion and removal operations would of course need to
refrain from acquiring and releasing these same per-hash-chain locks. This complexity
can be worthwhile if rename operations are frequent, and of course can allow rename
operations to execute concurrently. ❑

Why on earth did we need that global lock in the first place?
Answer:
A given thread’s __thread variables vanish when that thread exits. It is therefore
necessary to synchronize any operation that accesses other threads’ __thread variables
with thread exit. Without such synchronization, accesses to __thread variable of a
just-exited thread will result in segmentation faults. ❑

Hey!!! Line 48 of Listing 13.5 modifies a value in a pre-existing countarray
structure! Didn’t you say that this structure, once made available to read_count(),
remained constant???
Answer:
Indeed I did say that. And it would be possible to make count_register_thread()
allocate a new structure, much as count_unregister_thread() currently does.
But this is unnecessary. Recall the derivation of the error bounds of read_count()
that was based on the snapshots of memory. Because new threads start with initial
counter values of zero, the derivation holds even if we add a new thread partway
through read_count()’s execution. So, interestingly enough, when adding a new
thread, this implementation gets the effect of allocating a new structure, but without
actually having to do the allocation. ❑

Given the fixed-size counterp array, exactly how does this code avoid a fixed upper
bound on the number of threads???
Answer:
You are quite right, that array does in fact reimpose the fixed upper limit. This limit may
be avoided by tracking threads with a linked list, as is done in userspace RCU [DMS+ 12].
Doing something similar for this code is left as an exercise for the reader. ❑
v2023.06.11a
Listing E.10: Localized Correlated Measurement Fields

1 struct measurement {
2 double meas_1;
3 double meas_2;
4 double meas_3;
5 };
6
7 struct animal {
8 char name[40];
9 double age;
10 struct measurement *mp;
11 struct measurement meas;
13 };

Wow! Listing 13.5 contains 70 lines of code, compared to only 42 in Listing 5.4. Is
this extra complexity really worth it?
Answer:
This of course needs to be decided on a case-by-case basis. If you need an implementation
of read_count() that scales linearly, then the lock-based implementation shown in
Listing 5.4 simply will not work for you. On the other hand, if calls to read_count()
are sufficiently rare, then the lock-based version is simpler and might thus be better,
although much of the size difference is due to the structure definition, memory allocation,
and NULL return checking.
Of course, a better question is “Why doesn’t the language implement cross-thread
access to __thread variables?” After all, such an implementation would make both the
locking and the use of RCU unnecessary. This would in turn enable an implementation
that was even simpler than the one shown in Listing 5.4, but with all the scalability and
performance benefits of the implementation shown in Listing 13.5! ❑

But cant’t the approach shown in Listing 13.9 result in extra cache misses, in turn
resulting in additional read-side overhead?
Answer:
Indeed it can.
One way to avoid this cache-miss overhead is shown in Listing E.10: Simply embed
an instance of a measurement structure named meas into the animal structure, and
point the ->mp field at this ->meas field.
Measurement updates can then be carried out as follows:
1. Allocate a new measurement structure and place the new measurements into it.
2. Use rcu_assign_pointer() to point ->mp to this new structure.
3. Wait for a grace period to elapse, for example using either synchronize_rcu()
or call_rcu().
4. Copy the measurements from the new measurement structure into the embedded
->meas field.
v2023.06.11a
5. Use rcu_assign_pointer() to point ->mp back to the old embedded ->meas

field.
6. After another grace period elapses, free up the new measurement structure.
This approach uses a heavier weight update procedure to eliminate the extra cache
miss in the common case. The extra cache miss will be incurred only while an update is
actually in progress. ❑

But how does this scan work while a resizable hash table is being resized? In that
case, neither the old nor the new hash table is guaranteed to contain all the elements
in the hash table!
Answer:
True, resizable hash tables as described in Section 10.4 cannot be fully scanned while
being resized. One simple way around this is to acquire the hashtab structure’s
->ht_lock while scanning, but this prevents more than one scan from proceeding
concurrently.
Another approach is for updates to mutate the old hash table as well as the new one
while resizing is in progress. This would allow scans to find all elements in the old hash
table. Implementing this is left as an exercise for the reader. ❑

But how would this work with a resizable hash table, such as the one described in
Section 10.4?
Answer:
In this case, more care is required because the hash table might well be resized during
the time that we momentarily exited the RCU read-side critical section. Worse yet, the
resize operation can be expected to free the old hash buckets, leaving us pointing to the
freelist.
But it is not sufficient to prevent the old hash buckets from being freed. It is also
necessary to ensure that those buckets continue to be updated.
One way to handle this is to have a reference count on each set of buckets, which is
initially set to the value one. A full-table scan would acquire a reference at the beginning
of the scan (but only if the reference is non-zero) and release it at the end of the scan.
The resizing would populate the new buckets, release the reference, wait for a grace
period, and then wait for the reference to go to zero. Once the reference was zero, the
resizing could let updaters forget about the old hash buckets and then free it.
Actual implementation is left to the interested reader, who will gain much insight
from this task. ❑
v2023.06.11a
E.14 Advanced Synchronization

Given that there will always be a sharply limited number of CPUs available, is
population obliviousness really useful?
Answer:
Given the surprisingly limited scalability of any number of NBS algorithms, population
obliviousness can be surprisingly useful. Nevertheless, the overall point of the question
is valid. It is not normally helpful for an algorithm to scale beyond the size of the largest
system it is ever going to run on. ❑

Wait! In order to dequeue all elements, both the ->head and ->tail pointers must
be changed, which cannot be done atomically on typical computer systems. So how is
this supposed to work???
Answer:
One pointer at a time!
First, atomically exchange the ->head pointer with NULL. If the return value from
the atomic exchange operation is NULL, the queue was empty and you are done. And if
someone else attempts a dequeue-all at this point, they will get back a NULL pointer.
Otherwise, atomically exchange the ->tail pointer with a pointer to the now-NULL
->head pointer. The return value from the atomic exchange operation is a pointer to
the ->next field of the eventual last element on the list.
Producing and testing actual code is left as an exercise for the interested and
enthusiastic reader, as are strategies for handling half-enqueued elements. ❑

So why not ditch antique languages like C and C++ for something more modern?
Answer:
That won’t help unless the more-modern languages proponents are energetic enough to
write their own compiler backends. The usual practice of re-using existing backends
also reuses charming properties such as refusal to support pointers to lifetime-ended
objects. ❑

Why does anyone care about demonic schedulers?
Answer:
A demonic scheduler is one way to model an insanely overloaded system. After all, if
you have an algorithm that you can prove runs reasonably given a demonic scheduler,
mere overload should be no problem, right?
On the other hand, it is only reasonable to ask if a demonic scheduler is really the
best way to model overload conditions. And perhaps it is time for more accurate models.
For one thing, a system might be overloaded in any of a number of ways. After all, an
v2023.06.11a
E.14. ADVANCED SYNCHRONIZATION 833
NBS algorithm that works fine on a demonic scheduler might or might not do well in
out-of-memory conditions, when mass storage fills, or when the network is congested.
Except that systems’ core counts have been increasing, which means that an overloaded
system is quite likely to be running more than one concurrent program.12 In that case,
even if a demonic scheduler is not so demonic as to inject idle cycles while there are
runnable tasks, it is easy to imagine such a scheduler consistently favoring the other
program over yours. If both programs could consume all available CPU, then this
scheduler might not run your program at all.
One way to avoid these issues is to simply avoid overload conditions. This is often
the preferred approach in production, where load balancers direct traffic away from
overloaded systems. And if all systems are overloaded, it is not unheard of to simply
shed load, that is, to drop the low-priority incoming requests. Nor is this approach
limited to computing, as those who have suffered through a rolling blackout can attest.
But load-shedding is often considered a bad thing by those whose load is being shed.
As always, choose wisely! ❑

It seems like the various members of the NBS hierarchy are rather useless. So why
bother with them at all???
Answer:
One advantage of the members of the NBS hierarchy is that they are reasonably simple
to define and use from a theoretical viewpoint. We can hope that work done in the
NBS arena will help lay the groundwork for analysis of real-world forward-progress
guarantees for concurrent real-time programs. However, as of 2022 it appears that
trace-based methodologies are in the lead [dOCdO19].
So why bother learning about NBS at all?
Because a great many people know of it, and are vaguely aware that it is somehow
related to real-time computing. Their response to your carefully designed real-time
constraints might well be of the form “Bah, just use wait-free algorithms!”. In the
all-too-common case where they are very convincing to your management, you will
need to understand NBS in order to bring the discussion back to reality. I hope that this
section has provided you with the required depth of understanding.
Another thing to note is that learning about the NBS hierarchy is probably no
more harmful than learning about transfinite numbers of the computational-complexity
hierarchy. In all three cases, it is important to avoid over-applying the theory. Which is
in and of itself good practice! ❑

But what about battery-powered systems? They don’t require energy flowing into the
system as a whole.
Answer:
Sooner or later, the battery must be recharged, which requires energy to flow into the
system. ❑
12 As a point of reference, back in the mid-1990s, Paul witnessed a 16-CPU system
running about 20 instances of a certain high-end proprietary database.
v2023.06.11a

But given the results from queueing theory, won’t low utilization merely improve the
average response time rather than improving the worst-case response time? And isn’t
worst-case response time all that most real-time systems really care about?
Answer:
Yes, but . . .
Those queueing-theory results assume infinite “calling populations”, which in the
Linux kernel might correspond to an infinite number of tasks. As of early 2021, no
real system supports an infinite number of tasks, so results assuming infinite calling
populations should be expected to have less-than-infinite applicability.
Other queueing-theory results have finite calling populations, which feature sharply
bounded response times [HL86]. These results better model real systems, and these
models do predict reductions in both average and worst-case response times as utiliza-
tions decrease. These results can be extended to model concurrent systems that use
synchronization mechanisms such as locking [Bra11, SM04a].
In short, queueing-theory results that accurately describe real-world real-time systems
show that worst-case response time decreases with decreasing utilization. ❑

Formal verification is already quite capable, benefiting from decades of intensive
study. Are additional advances really required, or is this just a practitioner’s excuse to
continue to lazily ignore the awesome power of formal verification?
Answer:
Perhaps this situation is just a theoretician’s excuse to avoid diving into the messy world
of real software? Perhaps more constructively, the following advances are required:
1. Formal verification needs to handle larger software artifacts. The largest verification
efforts have been for systems of only about 10,000 lines of code, and those have
been verifying much simpler properties than real-time latencies.
2. Hardware vendors will need to publish formal timing guarantees. This used to be
common practice back when hardware was much simpler, but today’s complex
hardware results in excessively complex expressions for worst-case performance.
Unfortunately, energy-efficiency concerns are pushing vendors in the direction of
even more complexity.
3. Timing analysis needs to be integrated into development methodologies and IDEs.
All that said, there is hope, given recent work formalizing the memory models of real
computer systems [AMP+ 11, AKNT13]. On the other hand, formal verification has just
as much trouble as does testing with the astronomical number of variants of the Linux
kernel that can be constructed from different combinations of its tens of thousands of
Kconfig options. Sometimes life is hard! ❑
v2023.06.11a
E.14. ADVANCED SYNCHRONIZATION 835

Differentiating real-time from non-real-time based on what can “be achieved straight-
forwardly by non-real-time systems and applications” is a travesty! There is absolutely
no theoretical basis for such a distinction!!! Can’t we do better than that???
Answer:
This distinction is admittedly unsatisfying from a strictly theoretical perspective. But
on the other hand, it is exactly what the developer needs in order to decide whether the
application can be cheaply and easily developed using standard non-real-time approaches,
or whether the more difficult and expensive real-time approaches are required. In other
words, although theory is quite important, for those of us called upon to complete
practical projects, theory supports practice, never the other way around. ❑

But if you only allow one reader at a time to read-acquire a reader-writer lock, isn’t
that the same as an exclusive lock???
Answer:
Indeed it is, other than the API. And the API is important because it allows the Linux
kernel to offer real-time capabilities without having the -rt patchset grow to ridiculous
sizes.
However, this approach clearly and severely limits read-side scalability. The Linux
kernel’s -rt patchset was long able to live with this limitation for several reasons:
(1) Real-time systems have traditionally been relatively small, (2) Real-time systems
have generally focused on process control, thus being unaffected by scalability limitations
in the I/O subsystems, and (3) Many of the Linux kernel’s reader-writer locks have been
converted to RCU.
However, the day came when it was absolutely necessary to permit concurrent readers,
as described in the text following this quiz. ❑

Suppose that preemption occurs just after the load from t->rcu_read_unlock_
special.s on line 12 of Listing 14.3. Mightn’t that result in the task failing to
invoke rcu_read_unlock_special(), thus failing to remove itself from the list of
tasks blocking the current grace period, in turn causing that grace period to extend
indefinitely?
Answer:
That is a real problem, and it is solved in RCU’s scheduler hook. If that scheduler
hook sees that the value of t->rcu_read_lock_nesting is negative, it invokes rcu_
read_unlock_special() if needed before allowing the context switch to complete.
❑

But isn’t correct operation despite fail-stop bugs a valuable fault-tolerance property?
Answer:
Yes and no.
v2023.06.11a
Yes in that non-blocking algorithms can provide fault tolerance in the face of fail-stop
bugs, but no in that this is grossly insufficient for practical fault tolerance. For example,
suppose you had a wait-free queue, and further suppose that a thread has just dequeued an
element. If that thread now succumbs to a fail-stop bug, the element it has just dequeued
is effectively lost. True fault tolerance requires way more than mere non-blocking
properties, and is beyond the scope of this book. ❑

I couldn’t help but spot the word “include” before this list. Are there other constraints?
Answer:
Indeed there are, and lots of them. However, they tend to be specific to a given situation,
and many of them can be thought of as refinements of some of the constraints listed
above. For example, the many constraints on choices of data structure will help meeting
the “Bounded time spent in any given critical section” constraint. ❑

Given that real-time systems are often used for safety-critical applications, and given
that runtime memory allocation is forbidden in many safety-critical situations, what is
with the call to malloc()???
Answer:
In early 2016, projects forbidding runtime memory allocation were also not at all
interested in multithreaded computing. So the runtime memory allocation is not an
additional obstacle to safety criticality.
However, by 2020 runtime memory allocation in multi-core real-time systems was
gaining some traction. ❑

Don’t you need some kind of synchronization to protect update_cal()?
Answer:
Indeed you do, and you could use any of a number of techniques discussed earlier in
this book. One of those techniques is use of a single updater thread, which would result
in exactly the code shown in update_cal() in Listing 14.6. ❑
E.15 Advanced Synchronization: Memory Ordering

This chapter has been rewritten since the first edition, and heavily edited since the
second edition. Did memory ordering change all that since 2014, let alone 2021?
Answer:
The earlier memory-ordering section had its roots in a pair of Linux Journal arti-
cles [McK05a, McK05b] dating back to 2005. Since then, the C and C++ memory
models [Bec11] have been formalized (and critiqued [BS14, BD14, VBC+ 15, BMN+ 15,
v2023.06.11a
E.15. ADVANCED SYNCHRONIZATION: MEMORY ORDERING 837
LVK+ 17, BGV17]), executable formal memory models for computer systems have
become the norm [MSS12, McK11d, SSA+ 11, AMP+ 11, AKNT13, AKT13, AMT14,
MS14, FSP+ 17, ARM17], and there is even a memory model for the Linux ker-
nel [AMM+ 17a, AMM+ 17b, AMM+ 18], along with a paper describing differences
between the C11 and Linux memory models [MWPF18].
The kernel concurrency sanitizer (KCSAN) [EMV+ 20a, EMV+ 20b], based in part on
RacerD [BGOS18] and implementing LKMM, has also been added to the Linux kernel
and is now heavily used.
Finally, there are now better ways of describing LKMM.
Given all this progress, substantial change was required. ❑

The compiler can also reorder Thread P0()’s and Thread P1()’s memory accesses in
Listing 15.1, right?
Answer:
In general, compiler optimizations carry out more extensive and profound reorderings
than CPUs can. However, in this case, the volatile accesses in READ_ONCE() and
WRITE_ONCE() prevent the compiler from reordering. And also from doing much else
as well, so the examples in this section will be making heavy use of READ_ONCE() and
WRITE_ONCE(). See Section 15.3 for more detail on the need for READ_ONCE() and
WRITE_ONCE(). ❑

But wait!!! On row 2 of Table 15.1 both x0 and x1 each have two values at the same
time, namely zero and two. How can that possibly work???
Answer:
There is an underlying cache-coherence protocol that straightens things out, which are
discussed in Appendix C.2. But if you think that a given variable having two values at
the same time is surprising, just wait until you get to Section 15.2.1! ❑

But don’t the values also need to be flushed from the cache to main memory?
Answer:
Perhaps surprisingly, not necessarily! On some systems, if the two variables are being
used heavily, they might be bounced back and forth between the CPUs’ caches and
never land in main memory. ❑

The rows in Table 15.3 seem quite random and confused. Whatever is the conceptual
basis of this table???
Answer:
The rows correspond roughly to hardware mechanisms of increasing power and overhead.
The WRITE_ONCE() row captures the fact that accesses to a single variable are always
fully ordered, as indicated by the “SV”column. Note that all other operations providing
ordering against accesses to multiple variables also provide this same-variable ordering.
v2023.06.11a
The READ_ONCE() row captures the fact that (as of 2021) compilers and CPUs do
not indulge in user-visible speculative stores, so that any store whose address, data, or
execution depends on a prior load is guaranteed to happen after that load completes.
However, this guarantee assumes that these dependencies have been constructed carefully,
as described in Sections 15.3.2 and 15.3.3.
The “_relaxed() RMW operation” row captures the fact that a value-returning
_relaxed() RMW has done a load and a store, which are every bit as good as a
READ_ONCE() and a WRITE_ONCE(), respectively.
The *_dereference() row captures the address and data dependency ordering
provided by rcu_dereference() and friends. Again, these dependencies must been
constructed carefully, as described in Section 15.3.2.
The “Successful *_acquire()” row captures the fact that many CPUs have special
“acquire” forms of loads and of atomic RMW instructions, and that many other CPUs
have lightweight memory-barrier instructions that order prior loads against subsequent
loads and stores.
The “Successful *_release()” row captures the fact that many CPUs have special
“release” forms of stores and of atomic RMW instructions, and that many other CPUs
have lightweight memory-barrier instructions that order prior loads and stores against
subsequent stores.
The smp_rmb() row captures the fact that many CPUs have lightweight memory-
barrier instructions that order prior loads against subsequent loads. Similarly, the
smp_wmb() row captures the fact that many CPUs have lightweight memory-barrier
instructions that order prior stores against subsequent stores.
None of the ordering operations thus far require prior stores to be ordered against
subsequent loads, which means that these operations need not interfere with store buffers,
whose main purpose in life is in fact to reorder prior stores against subsequent loads.
The lightweight nature of these operations is precisely due to their policy of store-buffer
non-interference. However, as noted earlier, it is sometimes necessary to interfere with
the store buffer in order to prevent prior stores from being reordered against later stores,
which brings us to the remaining rows in this table.
The smp_mb() row corresponds to the full memory barrier available on most platforms,
with Itanium being the exception that proves the rule. However, even on Itanium, smp_
mb() provides full ordering with respect to READ_ONCE() and WRITE_ONCE(), as
discussed in Section 15.5.4.
The “Successful full-strength non-void RMW” row captures the fact that on some
platforms (such as x86) atomic RMW instructions provide full ordering both before and
after. The Linux kernel therefore requires that full-strength non-void atomic RMW
operations provide full ordering in cases where these operations succeed. (Full-strength
atomic RMW operation’s names do not end in _relaxed, _acquire, or _release.)
As noted earlier, the case where these operations do not succeed is covered by the
“_relaxed() RMW operation” row.
However, the Linux kernel does not require that either void or _relaxed() atomic
RMW operations provide any ordering whatsoever, with the canonical example being
atomic_inc(). Therefore, these operations, along with failing non-void atomic
RMW operations may be preceded by smp_mb__before_atomic() and followed by
smp_mb__after_atomic() to provide full ordering for any accesses preceding or
following both. No ordering need be provided for accesses between the smp_mb__
before_atomic() (or, similarly, the smp_mb__after_atomic()) and the atomic
v2023.06.11a
RMW operation, as indicated by the “a” entries on the smp_mb__before_atomic()

and smp_mb__after_atomic() rows of the table.
In short, the structure of this table is dictated by the properties of the underlying
hardware, which are constrained by nothing other than the laws of physics, which were
covered back in Chapter 3. That is, the table is not random, although it is quite possible
that you are confused. ❑

Why is Table 15.3 missing smp_mb__after_unlock_lock() and smp_mb__after_
spinlock()?
Answer:
These two primitives are rather specialized, and at present seem difficult to fit into
Table 15.3. The smp_mb__after_unlock_lock() primitive is intended to be placed
immediately after a lock acquisition, and ensures that all CPUs see all accesses in
prior critical sections as happening before all accesses following the smp_mb__after_
unlock_lock() and also before all accesses in later critical sections. Here “all CPUs”
includes those CPUs not holding that lock, and “prior critical sections” includes all
prior critical sections for the lock in question as well as all prior critical sections for all
other locks that were released by the same CPU that executed the smp_mb__after_
unlock_lock().
The smp_mb__after_spinlock() provides the same guarantees as does smp_
mb__after_unlock_lock(), but also provides additional visibility guarantees for
other accesses performed by the CPU that executed the smp_mb__after_spinlock().
Given any store S performed prior to any earlier lock acquisition and any load L
performed after the smp_mb__after_spinlock(), all CPUs will see S as happening
before L. In other words, if a CPU performs a store S, acquires a lock, executes
an smp_mb__after_spinlock(), then performs a load L, all CPUs will see S as
happening before L. ❑

But how can I know that a given project can be designed and coded within the confines
of these rules of thumb?
Answer:
Much of the purpose of the remainder of this chapter is to answer exactly that question!
❑

How can you tell which memory barriers are strong enough for a given use case?
Answer:
Ah, that is a deep question whose answer requires most of the rest of this chapter. But
the short answer is that smp_mb() is almost always strong enough, albeit at some cost.
❑
v2023.06.11a

Wait!!! Where do I find this tooling that automatically analyzes litmus tests???
Answer:
Get version v4.17 (or later) of the Linux-kernel source code, then follow the instructions
in tools/memory-model/README to install the needed tools. Then follow the further
instructions to run these tools on the litmus test of your choice. ❑

What assumption is the code fragment in Listing 15.3 making that might not be valid
on real hardware?
Answer:
The code assumes that as soon as a given CPU stops seeing its own value, it will
immediately see the final agreed-upon value. On real hardware, some of the CPUs
might well see several intermediate results before converging on the final value. The
actual code used to produce the data in the figures discussed later in this section was
therefore somewhat more complex. ❑

How could CPUs possibly have different views of the value of a single variable at the
same time?
Answer:
As discussed in Section 15.1.1, many CPUs have store buffers that record the values
of recent stores, which do not become globally visible until the corresponding cache
line makes its way to the CPU. Therefore, it is quite possible for each CPU to see its
own value for a given variable (in its own store buffer) at a single point in time—and for
main memory to hold yet another value. One of the reasons that memory barriers were
invented was to allow software to deal gracefully with situations like this one.
Fortunately, software rarely cares about the fact that multiple CPUs might see multiple
values for the same variable. ❑

Why do CPUs 2 and 3 come to agreement so quickly, when it takes so long for CPUs 1
and 4 to come to the party?
Answer:
CPUs 2 and 3 are a pair of hardware threads on the same core, sharing the same cache
hierarchy, and therefore have very low communications latencies. This is a NUMA, or,
more accurately, a NUCA effect.
This leads to the question of why CPUs 2 and 3 ever disagree at all. One possible
reason is that they each might have a small amount of private cache in addition to a
larger shared cache. Another possible reason is instruction reordering, given the short
10-nanosecond duration of the disagreement and the total lack of memory-ordering
operations in the code fragment. ❑

But why make load-load reordering visible to the user? Why not just use speculative
v2023.06.11a
execution to allow execution to proceed in the common case where there are no
intervening stores, in which case the reordering cannot be visible anyway?
Answer:
They can and many do, otherwise systems containing strongly ordered CPUs would
be slow indeed. However, speculative execution does have its downsides, especially if
speculation must be rolled back frequently, particularly on battery-powered systems.
Speculative execution can also introduce side channels, which might in turn be exploited
to exfiltrate information. But perhaps future systems will be able to overcome these
disadvantages. Until then, we can expect vendors to continue producing weakly ordered
CPUs. ❑

Why should strongly ordered systems pay the performance price of unnecessary
smp_rmb() and smp_wmb() invocations? Shouldn’t weakly ordered systems shoulder
the full cost of their misordering choices???
Answer:
That is in fact exactly what happens. On strongly ordered systems, smp_rmb() and
smp_wmb() emit no instructions, but instead just constrain the compiler. Thus, in this
case, weakly ordered systems do in fact shoulder the full cost of their memory-ordering
choices. ❑

But how do we know that all platforms really avoid triggering the exists clauses in
Listings 15.10 and 15.11?
Answer:
Answering this requires identifying three major groups of platforms: (1) Total-store-order
(TSO) platforms, (2) Weakly ordered platforms, and (3) DEC Alpha.
The TSO platforms order all pairs of memory references except for prior stores against
later loads. Because the address dependency on lines 18 and 19 of Listing 15.10 is
instead a load followed by another load, TSO platforms preserve this address dependency.
They also preserve the address dependency on lines 17 and 18 of Listing 15.11 because
this is a load followed by a store. Because address dependencies must start with a
load, TSO platforms implicitly but completely respect them, give or take compiler
optimizations, hence the need for READ_ONCE().
Weakly ordered platforms don’t necessarily maintain ordering of unrelated accesses.
However, the address dependencies in Listings 15.10 and 15.11 are not unrelated: There
is an address dependency. The hardware tracks dependencies and maintains the needed
ordering.
There is one (famous) exception to this rule for weakly ordered platforms, and that
exception is DEC Alpha for load-to-load address dependencies. And this is why, in
Linux kernels predating v4.15, DEC Alpha requires the explicit memory barrier supplied
for it by the now-obsolete lockless_dereference() on line 18 of Listing 15.10.
However, DEC Alpha does track load-to-store address dependencies, which is why
line 17 of Listing 15.11 does not need a lockless_dereference(), even in Linux
kernels predating v4.15.
v2023.06.11a
To sum up, current platforms either respect address dependencies implicitly, as is

the case for TSO platforms (x86, mainframe, SPARC, . . .), have hardware tracking for
address dependencies (Arm, PowerPC, MIPS, . . .), have the required memory barriers
supplied by READ_ONCE() (DEC Alpha in Linux kernel v4.15 and later), or supplied by
rcu_dereference() (DEC Alpha in Linux kernel v4.14 and earlier). ❑

Why the use of smp_wmb() in Listings 15.10 and 15.11? Wouldn’t smp_store_
release() be a better choice?
Answer:
In most cases, smp_store_release() is indeed a better choice. However, smp_wmb()
was there first in the Linux kernel, so it is still good to understand how to use it. ❑

SP, MP, LB, and now S. Where do all these litmus-test abbreviations come from and
how can anyone keep track of them?
Answer:
The best scorecard is the infamous test6.pdf [SSA+ 11]. Unfortunately, not all of the
abbreviations have catchy expansions like SB (store buffering), MP (message passing),
and LB (load buffering), but at least the list of abbreviations is readily available. ❑

But wait!!! Line 17 of Listing 15.12 uses READ_ONCE(), which marks the load as
volatile, which means that the compiler absolutely must emit the load instruction even
if the value is later multiplied by zero. So how can the compiler possibly break this
data dependency?
Answer:
Yes, the compiler absolutely must emit a load instruction for a volatile load. But if you
multiply the value loaded by zero, the compiler is well within its rights to substitute a
constant zero for the result of that multiplication, which will break the data dependency
on many platforms.
Worse yet, if the dependent store does not use WRITE_ONCE(), the compiler could
hoist it above the load, which would cause even TSO platforms to fail to provide ordering.
❑

Wouldn’t control dependencies be more robust if they were mandated by language
standards???
Answer:
But of course! And perhaps in the fullness of time they will be so mandated. ❑
v2023.06.11a
Listing E.11: Litmus Test Distinguishing Multicopy Atomic From Other Multicopy Atomic
1 C C-MP-OMCA+o-o-o+o-rmb-o
2
3 {}
4
6 {
7 int r0;
8
11 WRITE_ONCE(*y, r0);
12 }
13
15 {
16 int r1;
17 int r2;
18
20 smp_rmb();
22 }
23
24 exists (1:r1=1 /\ 1:r2=0)

But in Listing 15.15, wouldn’t be just as bad if P2()’s r1 and r2 obtained the values 2
and 1, respectively, while P3()’s r3 and r4 obtained the values 1 and 2, respectively?
Answer:
Yes, it would. Feel free to modify the exists clause to check for that outcome and see
what happens. ❑

Can you give a specific example showing different behavior for multicopy atomic on
the one hand and other-multicopy atomic on the other?
Answer:
Listing E.11 (C-MP-OMCA+o-o-o+o-rmb-o.litmus) shows such a test.
On a multicopy-atomic platform, P0()’s store to x on line 9 must become visible
to both P0() and P1() simultaneously. Because this store becomes visible to P0()
on line 10, before P0()’s store to y on line 11, P0()’s store to x must become visible
before its store to y everywhere, including P1(). Therefore, if P1()’s load from y on
line 19 returns the value 1, so must its load from x on line 21, given that the smp_rmb()
on line 20 forces these two loads to execute in order. Therefore, the exists clause on
line 24 cannot trigger on a multicopy-atomic platform.
In contrast, on an other-multicopy-atomic platform, P0() could see its own store
early, so that there would be no constraint on the order of visibility of the two stores
from P1(), which in turn allows the exists clause to trigger. ❑

Then who would even think of designing a system with shared store buffers???
v2023.06.11a
Answer:
This is in fact a very natural design for any system having multiple hardware threads per
core. Natural from a hardware point of view, that is! ❑

But just how is it fair that P0() and P1() must share a store buffer and a cache, but
P2() gets one each of its very own???
Answer:
Presumably there is a P3(), as is in fact shown in Figure 15.10, that shares P2()’s
store buffer and cache. But not necessarily. Some platforms allow different cores to
disable different numbers of threads, allowing the hardware to adjust to the needs of the
workload at hand. For example, a single-threaded critical-path portion of the workload
might be assigned to a core with only one thread enabled, thus allowing the single
thread running that portion of the workload to use the entire capabilities of that core.
Other more highly parallel but cache-miss-prone portions of the workload might be
assigned to cores with all hardware threads enabled to provide improved throughput.
This improved throughput could be due to the fact that while one hardware thread is
stalled on a cache miss, the other hardware threads can make forward progress.
In such cases, performance requirements override quaint human notions of fairness.
❑

Referring to Table 15.4, why on earth would P0()’s store take so long to complete
when P1()’s store complete so quickly? In other words, does the exists clause on
line 28 of Listing 15.16 really trigger on real systems?
Answer:
You need to face the fact that it really can trigger. Akira Yokosawa used the litmus7
tool to run this litmus test on a POWER8 system. Out of 1,000,000,000 runs, 4 triggered
the exists clause. Thus, triggering the exists clause is not merely a one-in-a-million
occurrence, but rather a one-in-a-hundred-million occurrence. But it nevertheless really
does trigger on real systems. ❑

But it is not necessary to worry about propagation unless there are at least three threads
in the litmus test, right?
Answer:
Wrong.
Listing E.12 (C-R+o-wmb-o+o-mb-o.litmus) shows a two-thread litmus test that
requires propagation due to the fact that it only has store-to-store and load-to-store links
between its pair of threads. Even though P0() is fully ordered by the smp_wmb() and
P1() is fully ordered by the smp_mb(), the counter-temporal nature of the links means
that the exists clause on line 21 really can trigger. To prevent this triggering, the
smp_wmb() on line 8 must become an smp_mb(), bringing propagation into play twice,
once for each non-temporal link. ❑
v2023.06.11a
Listing E.12: R Litmus Test With Write Memory Barrier (No Ordering)
1 C C-R+o-wmb-o+o-mb-o
2
3 {}
4
6 {
8 smp_wmb();
10 }
11
12 P1(int *x0, int *x1)
13 {
14 int r2;
15
17 smp_mb();
19 }
20
21 exists (1:r2=0 /\ x1=2)

But given that smp_mb() has the propagation property, why doesn’t the smp_mb() on
line 25 of Listing 15.18 prevent the exists clause from triggering?
Answer:
As a rough rule of thumb, the smp_mb() barrier’s propagation property is sufficient to
maintain ordering through only one load-to-store link between processes. Unfortunately,
Listing 15.18 has not one but two load-to-store links, with the first being from the
READ_ONCE() on line 17 to the WRITE_ONCE() on line 24 and the second being from
the READ_ONCE() on line 26 to the WRITE_ONCE() on line 7. Therefore, preventing the
exists clause from triggering should be expected to require not one but two instances
of smp_mb().
As a special exception to this rule of thumb, a release-acquire chain can have one
load-to-store link between processes and still prohibit the cycle. ❑

But for litmus tests having only ordered stores, as shown in Listing 15.20 (C-2+2W+o-
wmb-o+o-wmb-o.litmus), research shows that the cycle is prohibited, even in weakly
ordered systems such as Arm and Power [SSA+ 11]. Given that, are store-to-store
really always counter-temporal???
Answer:
This litmus test is indeed a very interesting curiosity. Its ordering apparently occurs
naturally given typical weakly ordered hardware design, which would normally be
considered a great gift from the relevant laws of physics and cache-coherency-protocol
mathematics.
Unfortunately, no one has been able to come up with a software use case for this
gift that does not have a much better alternative implementation. Therefore, neither
the C11 nor the Linux kernel memory models provide any guarantee corresponding to
Listing 15.20. This means that the exists clause on line 19 can trigger.
Of course, without the barrier, there are no ordering guarantees, even on real weakly
ordered hardware, as shown in Listing E.13 (C-2+2W+o-o+o-o.litmus). ❑
v2023.06.11a
Listing E.13: 2+2W Litmus Test (No Ordering)

1 C C-2+2W+o-o+o-o
2
3 {}
4
6 {
9 }
10
11 P1(int *x0, int *x1)
12 {
15 }
16
17 exists (x0=1 /\ x1=1)
Listing E.14: LB Litmus Test With No Acquires

1 C C-LB+o-data-o+o-data-o+o-data-o
2
3 {
4 x1=1;
5 x2=2;
6 }
7
9 {
10 int r2;
11
14 }
15
16 P1(int *x1, int *x2)
17 {
18 int r2;
19
22 }
23
24 P2(int *x2, int *x0)
25 {
26 int r2;
27
30 }
31
32 exists (0:r2=2 /\ 1:r2=0 /\ 2:r2=1)

Can you construct a litmus test like that in Listing 15.21 that uses only dependencies?
Answer:
Listing E.14 shows a somewhat nonsensical but very real example. Creating a more
useful (but still real) litmus test is left as an exercise for the reader. ❑

Suppose we have a short release-acquire chain along with one load-to-store link and
v2023.06.11a
one store-to-store link, like that shown in Listing 15.25. Given that there is only one
of each type of non-store-to-load link, the exists cannot trigger, right?
Answer:
Wrong. It is the number of non-store-to-load links that matters. If there is only one
non-store-to-load link, a release-acquire chain can prevent the exists clause from
triggering. However, if there is more than one non-store-to-load link, be they store-
to-store, load-to-store, or any combination thereof, it is necessary to have at least one
full barrier (smp_mb() or better) between each non-store-to-load link. In Listing 15.25,
preventing the exists clause from triggering therefore requires an additional full barrier
between either P0()’s or P1()’s accesses. ❑

There are store-to-load links, load-to-store links, and store-to-store links. But what
about load-to-load links?
Answer:
The problem with the concept of load-to-load links is that if the two loads from the
same variable return the same value, there is no way to determine their ordering. The
only way to determine their ordering is if they return different values, in which case
there had to have been an intervening store. And that intervening store means that there
is no load-to-load link, but rather a load-to-store link followed by a store-to-load link. ❑

What happens if that lwsync instruction is instead a sync instruction?
Answer:
The counter-intuitive outcome cannot happen. (Try it!) ❑

Why not place a barrier() call immediately before a plain store to prevent the
compiler from inventing stores?
Answer:
Because it would not work. Although the compiler would be prevented from inventing a
store prior to the barrier(), nothing would prevent it from inventing a store between
that barrier() and the plain store. ❑

Why can’t you simply dereference the pointer before comparing it to &reserve_int
on line 6 of Listing 15.28?
Answer:
For first, it might be necessary to invoke handle_reserve() before do_something_
with().
But more relevant to memory ordering, the compiler is often within its rights to
hoist the comparison ahead of the dereferences, which would allow the compiler to use
&reserve_int instead of the variable p that the hardware has tagged with a dependency.
❑
v2023.06.11a
Listing E.15: Breakable Dependencies With Non-Constant Comparisons

1 int *gp1;
2 int *p;
3 int *q;
4
6 q = get_a_pointer();
7 if (p == q)
8 handle_equality(p);
9 do_something_with(*p);
Listing E.16: Broken Dependencies With Non-Constant Comparisons

1 int *gp1;
2 int *p;
3 int *q;
4
6 q = get_a_pointer();
7 if (p == q) {
8 handle_equality(q);
9 do_something_with(*q);
10 } else {
11 do_something_with(*p);
12 }

But it should be safe to compare two pointer variables, right? After all, the compiler
doesn’t know the value of either, so how can it possibly learn anything from the
comparison?
Answer:
Unfortunately, the compiler really can learn enough to break your dependency chain,
for example, as shown in Listing E.15. The compiler is within its rights to transform
this code into that shown in Listing E.16, and might well make this transformation due
to register pressure if handle_equality() was inlined and needed a lot of registers.
Line 9 of this transformed code uses q, which although equal to p, is not necessarily
tagged by the hardware as carrying a dependency. Therefore, this transformed code
does not necessarily guarantee that line 9 is ordered after line 5.13 ❑

But doesn’t the condition in line 35 supply a control dependency that would keep
line 36 ordered after line 34?
Answer:
Yes, but no. Yes, there is a control dependency, but control dependencies do not order
later loads, only later stores. If you really need ordering, you could place an smp_rmb()
between lines 35 and 36. Or better yet, have updater() allocate two structures instead
of reusing the structure. For more information, see Section 15.3.3. ❑

But there is a READ_ONCE(), so how can the compiler prove anything about the value
of q?
13 Kudos to Linus Torvalds for providing this example.
v2023.06.11a
Answer:
Given the simple if statement comparing against zero, it is hard to imagine the compiler
proving anything. But suppose that later code executed a division by q. Because division
by zero is undefined behavior, as of 2023, many compilers will assume that the value
of q must be non-zero, and will thus remove that if statement, thus unconditionally
executing the WRITE_ONCE(), in turn destroying the control dependency.
There are some who argue (correctly, in Paul’s view) that back-propagating undefined
behavior across volatile accesses constitutes a compiler bug, but many compiler writers
insist that this is not a bug, but rather a valuable optimization. ❑

Can’t you instead add an smp_mb() to P1() in Listing 15.32?
Answer:
Not given the Linux kernel memory model. (Try it!) However, you can instead replace
P0()’s WRITE_ONCE() with smp_store_release(), which usually has less overhead
than does adding an smp_mb(). ❑

But doesn’t PowerPC have weak unlock-lock ordering properties within the Linux
kernel, allowing a write before the unlock to be reordered with a read after the lock?
Answer:
Yes, but only from the perspective of a third thread not holding that lock. In contrast,
memory allocators need only concern themselves with the two threads migrating the
memory. It is after all the developer’s responsibility to properly synchronize with any
other threads that need access to the newly migrated block of memory. ❑

But if there are three critical sections, isn’t it true that CPUs not holding the lock will
observe the accesses from the first and the third critical section as being ordered?
Answer:
No.
Listing E.17 shows an example three-critical-section chain (Lock-across-unlock-
lock-3.litmus). Running this litmus test shows that the exists clause can still be
satisfied, so this additional critical section is still not sufficient to force ordering.
However, as the reader can verify, placing an smp_mb__after_spinlock() after
either P1()’s or P2()’s lock acquisition does suffice to force ordering. ❑

But if spin_is_locked() returns false, don’t we also know that no other CPU or
thread is holding the corresponding lock?
Answer:
No. By the time that the code inspects the return value from spin_is_locked(), some
other CPU or thread might well have acquired the corresponding lock. ❑
v2023.06.11a
Listing E.17: Accesses Between Multiple Different-CPU Critical Sections

2
3 {}
4
6 {
7 spin_lock(sp);
9 spin_unlock(sp);
10 }
11
12 P1(int *x, int *y, int *z, spinlock_t *sp)
13 {
14 int r1;
15
16 spin_lock(sp);
19 spin_unlock(sp);
20 }
21
22 P2(int *x, int *y, int *z, spinlock_t *sp)
23 {
24 int r1;
25 int r2;
26
27 spin_lock(sp);
30 spin_unlock(sp);
31 }
32
34 {
35 int r1;
36
38 smp_mb();
40 }
41
42 exists (1:r1=1 /\ 2:r1=1 /\ 2:r2=0 /\ 3:r1=0)

Wait a minute! In QSBR implementations of RCU, no code is emitted for rcu_
read_lock() and rcu_read_unlock(). This means that the RCU read-side critical
section in Listing 15.45 isn’t just empty, it is completely nonexistent!!! So how can
something that doesn’t exist at all possibly have any effect whatsoever on ordering???
Answer:
Because in QSBR, RCU read-side critical sections don’t actually disappear. Instead, they
are extended in both directions until a quiescent state is encountered. For example, in the
Linux kernel, the critical section might be extended back to the most recent schedule()
call and ahead to the next schedule() call. Of course, in non-QSBR implementations,
rcu_read_lock() and rcu_read_unlock() really do emit code, which can clearly
provide ordering. And within the Linux kernel, even the QSBR implementation has
a compiler barrier() in rcu_read_lock() and rcu_read_unlock(), which is
necessary to prevent the compiler from moving memory accesses that might result in
page faults into the RCU read-side critical section.
v2023.06.11a
Therefore, strange though it might seem, empty RCU read-side critical sections really
can and do provide some degree of ordering. ❑

Can P1()’s accesses be reordered in the litmus tests shown in Listings 15.43, 15.44,
and 15.45 in the same way that they were reordered going from Listing 15.38 to
Listing 15.39?
Answer:
No, because none of these later litmus tests have more than one access within their
RCU read-side critical sections. But what about swapping the accesses, for example,
in Listing 15.43, placing P1()’s WRITE_ONCE() within its critical section and the
READ_ONCE() before its critical section?
Swapping the accesses allows both instances of r2 to have a final value of zero, in
other words, although RCU read-side critical sections’ ordering properties can extend
outside of those critical sections, the same is not true of their reordering properties.
Checking this with herd and explaining why is left as an exercise for the reader. ❑

What would happen if the smp_mb() was instead added between P2()’s accesses in
Listing 15.47?
Answer:
The cycle would again be forbidden. Further analysis is left as an exercise for the reader.
❑

What happens to code between an atomic operation and an smp_mb__after_
atomic()?
Answer:
First, please don’t do this!
But if you do, this intervening code will either be ordered after the atomic operation
or before the smp_mb__after_atomic(), depending on the architecture, but not both.
This also applies to smp_mb__before_atomic() and smp_mb__after_spinlock(),
that is, both the uncertain ordering of the intervening code and the plea to avoid such
code. ❑

Why does Alpha’s READ_ONCE() include an mb() rather than rmb()?
Answer:
Alpha has only mb and wmb instructions, so smp_rmb() would be implemented by the
Alpha mb instruction in either case. In addition, at the time that the Linux kernel started
relying on dependency ordering, it was not clear that Alpha ordered dependent stores,
and thus smp_mb() was therefore the safe choice.
However, given the aforementioned v5.9 changes to READ_ONCE() and a few of
Alpha’s atomic read-modify-write operations, no Linux-kernel core code need concern
itself with DEC Alpha, thus greatly reducing Paul E. McKenney’s incentive to remove
Alpha support from the kernel. ❑
v2023.06.11a
Listing E.18: Userspace RCU Code Reordering

1 static inline int rcu_gp_ongoing(unsigned long *ctr)
2 {
3 unsigned long v;
4
5 v = LOAD_SHARED(*ctr);
6 return v && (v != rcu_gp_ctr);
7 }
8
9 static void update_counter_and_wait(void)
10 {
11 struct rcu_reader *index;
12
13 STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
14 barrier();
15 list_for_each_entry(index, &registry, node) {
16 while (rcu_gp_ongoing(&index->ctr))
17 msleep(10);
18 }
19 }
20
22 {
23 unsigned long was_online;
24
25 was_online = rcu_reader.ctr;
26 smp_mb();
27 if (was_online)
28 STORE_SHARED(rcu_reader.ctr, 0);
29 mutex_lock(&rcu_gp_lock);
30 update_counter_and_wait();
31 mutex_unlock(&rcu_gp_lock);
32 if (was_online)
33 STORE_SHARED(rcu_reader.ctr, LOAD_SHARED(rcu_gp_ctr));
34 smp_mb();
35 }

Isn’t DEC Alpha significant as having the weakest possible memory ordering?
Answer:
Although DEC Alpha does take considerable flak, it does avoid reordering reads from
the same CPU to the same variable. It also avoids the out-of-thin-air problem that
plagues the Java and C11 memory models [BD14, BMN+ 15, BS14, Boe20, Gol19,
Jef14, MB20, MJST16, Š11, VBC+ 15]. ❑

Given that hardware can have a half memory barrier, why don’t locking primitives
allow the compiler to move memory-reference instructions into lock-based critical
sections?
Answer:
In fact, as we saw in Section 15.5.3 and will see in Section 15.5.6, hardware really
does implement partial memory-ordering instructions and it also turns out that these
really are used to construct locking primitives. However, these locking primitives use
full compiler barriers, thus preventing the compiler from reordering memory-reference
instructions both out of and into the corresponding critical section.
To see why the compiler is forbidden from doing reordering that is permitted by
hardware, consider the following sample code in Listing E.18. This code is based on
the userspace RCU update-side code [DMS+ 12, Supplementary Materials Figure 5].
v2023.06.11a
Suppose that the compiler reordered lines 27 and 28 into the critical section starting
at line 29. Now suppose that two updaters start executing synchronize_rcu() at
about the same time. Then consider the following sequence of events:
1. CPU 0 acquires the lock at line 29.
2. Line 27 determines that CPU 0 was online, so it clears its own counter at line 28.
(Recall that lines 27 and 28 have been reordered by the compiler to follow line 29).
3. CPU 0 invokes update_counter_and_wait() from line 30.
4. CPU 0 invokes rcu_gp_ongoing() on itself at line 16, and line 5 sees that CPU 0 is
in a quiescent state. Control therefore returns to update_counter_and_wait(),
and line 15 advances to CPU 1.
5. CPU 1 invokes synchronize_rcu(), but because CPU 0 already holds the lock,
CPU 1 blocks waiting for this lock to become available. Because the compiler
reordered lines 27 and 28 to follow line 29, CPU 1 does not clear its own counter,
despite having been online.
6. CPU 0 invokes rcu_gp_ongoing() on CPU 1 at line 16, and line 5 sees that
CPU 1 is not in a quiescent state. The while loop at line 16 therefore never exits.
So the compiler’s reordering results in a deadlock. In contrast, hardware reordering
is temporary, so that CPU 1 might undertake its first attempt to acquire the mutex on
line 29 before executing lines 27 and 28, but it will eventually execute lines 27 and 28.
Because hardware reordering only results in a short delay, it can be tolerated. On the
other hand, because compiler reordering results in a deadlock, it must be prohibited.
Some research efforts have used hardware transactional memory to allow compilers
to safely reorder more aggressively, but the overhead of hardware transactions has thus
far made such optimizations unattractive. ❑

Why is it necessary to use heavier-weight ordering for load-to-store and store-to-store
links, but not for store-to-load links? What on earth makes store-to-load links so
special???
Answer:
Recall that load-to-store and store-to-store links can be counter-temporal, as illustrated
by Figures 15.12 and 15.13 in Section 15.2.7.2. This counter-temporal nature of
load-to-store and store-to-store links necessitates strong ordering.
In constrast, store-to-load links are temporal, as illustrated by Listings 15.12 and 15.13.
This temporal nature of store-to-load links permits use of minimal ordering. ❑
v2023.06.11a
E.16 Ease of Use

Can a similar algorithm be used when deleting elements?
Answer:
Yes. However, since each thread must hold the locks of three consecutive elements to
delete the middle one, if there are 𝑁 threads, there must be 2𝑁 + 1 elements (rather than
just 𝑁 + 1) in order to avoid deadlock. ❑

Yetch! What ever possessed someone to come up with an algorithm that deserves to
be shaved as much as this one does???
Answer:
That would be Paul.
He was considering the Dining Philosopher’s Problem, which involves a rather
unsanitary spaghetti dinner attended by five philosophers. Given that there are five
plates and but five forks on the table, and given that each philosopher requires two forks
at a time to eat, one is supposed to come up with a fork-allocation algorithm that avoids
deadlock. Paul’s response was “Sheesh! Just get five more forks!”
This in itself was OK, but Paul then applied this same solution to circular linked lists.
This would not have been so bad either, but he had to go and tell someone about it! ❑

Give an exception to this rule.
Answer:
One exception would be a difficult and complex algorithm that was the only one known
to work in a given situation. Another exception would be a difficult and complex
algorithm that was nonetheless the simplest of the set known to work in a given situation.
However, even in these cases, it may be very worthwhile to spend a little time trying to
come up with a simpler algorithm! After all, if you managed to invent the first algorithm
to do some task, it shouldn’t be that hard to go on to invent a simpler one. ❑
E.17 Conflicting Visions of the Future

But suppose that an application exits while holding a pthread_mutex_lock() that
happens to be located in a file-mapped region of memory?
Answer:
Indeed, in this case the lock would persist, much to the consternation of other processes
attempting to acquire this lock that is held by a process that no longer exists. Which is
why great care is required when using pthread_mutex objects located in file-mapped
memory regions. ❑
v2023.06.11a
E.17. CONFLICTING VISIONS OF THE FUTURE 855

What about non-persistent primitives represented by data structures in mmap() regions
of memory? What happens when there is an exec() within a critical section of such
a primitive?
Answer:
If the exec()ed program maps those same regions of memory, then this program could
in principle simply release the lock. The question as to whether this approach is sound
from a software-engineering viewpoint is left as an exercise for the reader. ❑

MV-RLU looks pretty good! Doesn’t it beat RCU hands down?
Answer:
One might get that impression from a quick read of the abstract, but more careful readers
will notice the “for a wide range of workloads” phrase in the last sentence. It turns out
that this phrase is quite important:
1. Their RCU evaluation uses synchronous grace periods, which needlessly throttle
updates, as noted in their Section 6.2.1. See Figure 10.11 page 302 of this book
to see that the venerable asynchronous call_rcu() primitive enables RCU to
perform and scale quite well with large numbers of updaters. Furthermore, in
Section 3.7 of their paper, the authors admit that asynchronous grace periods are
important to MV-RLU scalability. A fair comparison would also allow RCU the
benefits of asynchrony.
2. They use a poorly tuned 1,000-bucket hash table containing 10,000 elements. In
addition, their 448 hardware threads need considerably more than 1,000 buckets to
avoid the lock contention that they correctly state limits RCU performance in their
benchmarks. A useful comparison would feature a properly tuned hash table.
3. Their RCU hash table used per-bucket locks, which they call out as a bottleneck,
which is not a surprise given the long hash chains and small ratio of buckets to
threads. A number of their competing mechanisms instead use lockfree techniques,
thus avoiding the per-bucket-lock bottleneck, which cynics might claim sheds some
light on the authors’ otherwise inexplicable choice of poorly tuned hash tables.
The first graph in the middle row of the authors’ Figure 4 show what RCU can
achieve if not hobbled by artificial bottlenecks, as does the first portion of the
second graph in that same row.
4. Their linked-list operation permits RLU to do concurrent modifications of different
elements in the list, while RCU is forced to serialize updates. Again, RCU has
always worked just fine in conjunction with lockless updaters, a fact that has been
set forth in academic literature that the authors cited [DMS+ 12]. A fair comparison
would use the same style of update for RCU as it does for MV-RLU.
5. The authors fail to consider combining RCU and sequence locking, which is used
in the Linux kernel to give readers coherent views of multi-pointer updates.
6. The authors fail to consider RCU-based solutions to the Issaquah Chal-
lenge [McK16a], which also gives readers a coherent view of multi-pointer
updates, albeit with a weaker view of “coherent”.
v2023.06.11a
It is surprising that the anonymous reviewers of this paper did not demand an
apples-to-apples comparison of MV-RLU and RCU. Nevertheless, the authors should
be congratulated on producing an academic paper that presents an all-too-rare example
of good scalability combined with strong read-side coherence. They are also to be
congratulated on overcoming the traditional academic prejudice against asynchronous
grace periods, which greatly aided their scalability.
Interestingly enough, RLU and RCU take different approaches to avoid the inherent
limitations of STM noted by Hagit Attiya et al. [AHM09]. RCU avoids providing strict
serializability and RLU avoids providing invisible read-only transactions, both thus
avoiding the limitations. ❑

Given things like spin_trylock(), how does it make any sense at all to claim that
TM introduces the concept of failure???
Answer:
When using locking, spin_trylock() is a choice, with a corresponding failure-free
choice being spin_lock(), which is used in the common case, as in there are more
than 100 times as many calls to spin_lock() than to spin_trylock() in the v5.11
Linux kernel. When using TM, the only failure-free choice is the irrevocable transaction,
which is not used in the common case. In fact, the irrevocable transaction is not even
available in all TM implementations. ❑

What is to learn? Why not just use TM for memory-based data structures and locking
for those rare cases featuring the many silly corner cases listed in this silly section???
Answer:
The year 2005 just called, and it says that it wants its incandescent TM marketing hype
back.
In the year 2021, TM still has significant proving to do, even with the advent of HTM,
which is covered in the upcoming Section 17.3. ❑

Why would it matter that oft-written variables shared the cache line with the lock
variable?
Answer:
If the lock is in the same cacheline as some of the variables that it is protecting, then
writes to those variables by one CPU will invalidate that cache line for all the other
CPUs. These invalidations will generate large numbers of conflicts and retries, perhaps
even degrading performance and scalability compared to locking. ❑

Why are relatively small updates important to HTM performance and scalability?
Answer:
The larger the updates, the greater the probability of conflict, and thus the greater
probability of retries, which degrade performance. ❑
v2023.06.11a

How could a red-black tree possibly efficiently enumerate all elements of the tree
regardless of choice of synchronization mechanism???
Answer:
In many cases, the enumeration need not be exact. In these cases, hazard pointers or
RCU may be used to protect readers, which provides low probability of conflict with
any given insertion or deletion. ❑

But why can’t a debugger emulate single stepping by setting breakpoints at successive
lines of the transaction, relying on the retry to retrace the steps of the earlier instances
of the transaction?
Answer:
This scheme might work with reasonably high probability, but it can fail in ways that
would be quite surprising to most users. To see this, consider the following transaction:
1 begin_trans();
2 if (a) {
3 do_one_thing();
5 } else {
8 }
9 end_trans();
Suppose that the user sets a breakpoint at line 4, which triggers, aborting the transaction
and entering the debugger. Suppose that between the time that the breakpoint triggers
and the debugger gets around to stopping all the threads, some other thread sets the
value of a to zero. When the poor user attempts to single-step the program, surprise!
The program is now in the else-clause instead of the then-clause.
This is not what I call an easy-to-use debugger. ❑

But why would anyone need an empty lock-based critical section???
Answer:
See the answer to Quick Quiz 7.20 in Section 7.2.1.
However, it is claimed that given a strongly atomic HTM implementation without
forward-progress guarantees, any memory-based locking design based on empty critical
sections will operate correctly in the presence of transactional lock elision. Although I
have not seen a proof of this statement, there is a straightforward rationale for this claim.
The main idea is that in a strongly atomic HTM implementation, the results of a given
transaction are not visible until after the transaction completes successfully. Therefore,
if you can see that a transaction has started, it is guaranteed to have already completed,
which means that a subsequent empty lock-based critical section will successfully “wait”
on it—after all, there is no waiting required.
This line of reasoning does not apply to weakly atomic systems (including many STM
implementation), and it also does not apply to lock-based programs that use means other
v2023.06.11a
than memory to communicate. One such means is the passage of time (for example, in
hard real-time systems) or flow of priority (for example, in soft real-time systems).
Locking designs that rely on priority boosting are of particular interest. ❑

Can’t transactional lock elision trivially handle locking’s time-based messaging
semantics by simply choosing not to elide empty lock-based critical sections?
Answer:
It could do so, but this would be both unnecessary and insufficient.
It would be unnecessary in cases where the empty critical section was due to
conditional compilation. Here, it might well be that the only purpose of the lock was to
protect data, so eliding it completely would be the right thing to do. In fact, leaving the
empty lock-based critical section would degrade performance and scalability.
On the other hand, it is possible for a non-empty lock-based critical section to be
relying on both the data-protection and time-based and messaging semantics of locking.
Using transactional lock elision in such a case would be incorrect, and would result in
bugs. ❑

Given modern hardware [MOZ09], how can anyone possibly expect parallel software
relying on timing to work?
Answer:
The short answer is that on commonplace commodity hardware, synchronization designs
based on any sort of fine-grained timing are foolhardy and cannot be expected to operate
correctly under all conditions.
That said, there are systems designed for hard real-time use that are much more
deterministic. In the (very unlikely) event that you are using such a system, here is
a toy example showing how time-based synchronization can work. Again, do not try
this on commodity microprocessors, as they have highly nondeterministic performance
characteristics.
This example uses multiple worker threads along with a control thread. Each worker
thread corresponds to an outbound data feed, and records the current time (for example,
from the clock_gettime() system call) in a per-thread my_timestamp variable after
executing each unit of work. The real-time nature of this example results in the following
set of constraints:
1. It is a fatal error for a given worker thread to fail to update its timestamp for a time
period of more than MAX_LOOP_TIME.
2. Locks are used sparingly to access and update global state.
3. Locks are granted in strict FIFO order within a given thread priority.
When worker threads complete their feed, they must disentangle themselves from
the rest of the application and place a status value in a per-thread my_status variable
that is initialized to −1. Threads do not exit; they instead are placed on a thread pool to
accommodate later processing requirements. The control thread assigns (and re-assigns)
v2023.06.11a
worker threads as needed, and also maintains a histogram of thread statuses. The control
thread runs at a real-time priority no higher than that of the worker threads.
Worker threads’ code is as follows:
1 int my_status = -1; /* Thread local. */

2
3 while (continue_working()) {
4 enqueue_any_new_work();
5 wp = dequeue_work();
6 do_work(wp);
7 my_timestamp = clock_gettime(...);
8 }
9
10 acquire_lock(&departing_thread_lock);
11
12 /*
13 * Disentangle from application, might
14 * acquire other locks, can take much longer
15 * than MAX_LOOP_TIME, especially if many
16 * threads exit concurrently.
17 */
18 my_status = get_return_status();
19 release_lock(&departing_thread_lock);
20
21 /* thread awaits repurposing. */
The control thread’s code is as follows:
1 for (;;) {
3 ct = clock_gettime(...);
4 d = ct - per_thread(my_timestamp, t);
5 if (d >= MAX_LOOP_TIME) {
6 /* thread departing. */
7 acquire_lock(&departing_thread_lock);
8 release_lock(&departing_thread_lock);
9 i = per_thread(my_status, t);
10 status_hist[i]++; /* Bug if TLE! */
11 }
12 }
13 /* Repurpose threads as needed. */
14 }
Line 5 uses the passage of time to deduce that the thread has exited, executing lines 6
and 10 if so. The empty lock-based critical section on lines 7 and 8 guarantees that any
thread in the process of exiting completes (remember that locks are granted in FIFO
order!).
Once again, do not try this sort of thing on commodity microprocessors. After all, it
is difficult enough to get this right on systems specifically designed for hard real-time
use! ❑

But the boostee() function in Listing 17.1 alternatively acquires its locks in reverse
order! Won’t this result in deadlock?
Answer:
No deadlock will result. To arrive at deadlock, two different threads must each acquire
the two locks in opposite orders, which does not happen in this example. However,
deadlock detectors such as lockdep [Cor06a] will flag this as a false positive. ❑
v2023.06.11a

So a bunch of people set out to supplant locking, and they mostly end up just optimizing
locking???
Answer:
At least they accomplished something useful! And perhaps there will continue to be
additional HTM progress over time [SNGK17, SBN+ 20, GGK18, PMDY20]. ❑

Tables 17.1 and 17.2 state that hardware is only starting to become available. But
hasn’t HTM hardware support been widely available for almost a full decade?
Answer:
Yes and no. It appears that implementing even the HTM subset of TM in real hardware
is a bit trickier than it appears [JSG12, Was14, Int20a, Int21, Lar21]. Therefore, the sad
fact is that “starting to become available” is all too accurate as of 2021. In fact, vendors
are beginning to deprecate their HTM implementations [Int20c, Book III Appendix A].
❑

This list is ridiculously utopian! Why not stick to the current state of the formal-
verification art?
Answer:
You are welcome to your opinion on what is and is not utopian, but I will be paying more
attention to people actually making progress on the items in that list than to anyone who
might be objecting to them. This might have something to do with my long experience
with people attempting to talk me out of specific things that their favorite tools cannot
handle.
In the meantime, please feel free to read the papers written by the people who are
actually making progress, for example, this one [DFLO19]. ❑

Given the groundbreaking nature of the various verifiers used in the SEL4 project,
why doesn’t this chapter cover them in more depth?
Answer:
There can be no doubt that the verifiers used by the SEL4 project are quite capable.
However, SEL4 started as a single-CPU project. And although SEL4 has gained
multi-processor capabilities, it is currently using very coarse-grained locking that is
similar to the Linux kernel’s old Big Kernel Lock (BKL). There will hopefully come a
day when it makes sense to add SEL4’s verifiers to a book on parallel programming,
but this is not yet that day. ❑
v2023.06.11a
Table E.6: Emulating Locking: Performance Comparison (s)
cmpxchg_acquire() xchg_acquire()
# Lock filter exists filter exists
2 0.004 0.022 0.039 0.027 0.058
3 0.041 0.743 1.653 0.968 3.203
4 0.374 59.565 151.962 74.818 500.96
5 4.905

Why bother with a separate filter command on line 27 of Listing 17.2 instead of
just adding the condition to the exists clause? And wouldn’t it be simpler to use
xchg_acquire() instead of cmpxchg_acquire()?
Answer:
The filter clause causes the herd tool to discard executions at an earlier stage of
processing than does the exists clause, which provides significant speedups.
As for xchg_acquire(), this atomic operation will do a write whether or not
lock acquisition succeeds, which means that a model using xchg_acquire() will
have more operations than one using cmpxchg_acquire(), which won’t do a write
in the failed-acquisition case. More writes means more combinatorial to explode,
as shown in Table E.6 (C-SB+l-o-o-u+l-o-o-*u.litmus, C-SB+l-o-o-u+l-o-
o-u*-C.litmus, C-SB+l-o-o-u+l-o-o-u*-CE.litmus, C-SB+l-o-o-u+l-o-
o-u*-X.litmus, and C-SB+l-o-o-u+l-o-o-u*-XE.litmus). This table clearly
shows that cmpxchg_acquire() outperforms xchg_acquire() and that use of the
filter clause outperforms use of the exists clause. ❑

How do we know that the MTBFs of known bugs is a good estimate of the MTBFs of
bugs that have not yet been located?
Answer:
We don’t, but it does not matter.
To see this, note that the 7 % figure only applies to injected bugs that were subsequently
located: It necessarily ignores any injected bugs that were never found. Therefore,
the MTBF statistics of known bugs is likely to be a good approximation of that of the
injected bugs that are subsequently located.
A key point in this whole section is that we should be more concerned about bugs that
inconvenience users than about other bugs that never actually manifest. This of course
is not to say that we should completely ignore bugs that have not yet inconvenienced
users, just that we should properly prioritize our efforts so as to fix the most important
and urgent bugs first. ❑

But the formal-verification tools should immediately find all the bugs introduced by
the fixes, so why is this a problem?
v2023.06.11a
Answer:
It is a problem because real-world formal-verification tools (as opposed to those that
exist only in the imaginations of the more vociferous proponents of formal verification)
are not omniscient, and thus are only able to locate certain types of bugs. For but
one example, formal-verification tools are unlikely to spot a bug corresponding to an
omitted assertion or, equivalently, a bug corresponding to an undiscovered portion of
the specification. ❑

But many formal-verification tools can only find one bug at a time, so that each bug
must be fixed before the tool can locate the next. How can bug-fix efforts be prioritized
given such a tool?
Answer:
One approach is to provide a simple fix that might not be suitable for a production
environment, but which allows the tool to locate the next bug. Another approach is to
restrict configuration or inputs so that the bugs located thus far cannot occur. There
are a number of similar approaches, but the common theme is that fixing the bug
from the tool’s viewpoint is usually much easier than constructing and validating a
production-quality fix, and the key point is to prioritize the larger efforts required to
construct and validate the production-quality fixes. ❑

How would testing stack up in the scorecard shown in Table 17.5?
Answer:
It would be blue all the way down, with the possible exception of the third row (overhead)
which might well be marked down for testing’s difficulty finding improbable bugs.
On the other hand, improbable bugs are often also irrelevant bugs, so your mileage
may vary.
Much depends on the size of your installed base. If your code is only ever going to
run on (say) 10,000 systems, Murphy can actually be a really nice guy. Everything that
can go wrong, will. Eventually. Perhaps in geologic time.
But if your code is running on 20 billion systems, like the Linux kernel was said to
be by late 2017, Murphy can be a real jerk! Everything that can go wrong, will, and it
can go wrong really quickly!!! ❑

But aren’t there a great many more formal-verification systems than are shown in
Table 17.5?
Answer:
Indeed there are! This table focuses on those that Paul has used, but others are proving
to be useful. Formal verification has been heavily used in the seL4 project [SM13], and
its tools can now handle modest levels of concurrency. More recently, Catalin Marinas
used Lamport’s TLA tool [Lam02] to locate some forward-progress bugs in the Linux
kernel’s queued spinlock implementation. Will Deacon fixed these bugs [Dea18], and
Catalin verified Will’s fixes [Mar18].
v2023.06.11a
E.18. IMPORTANT QUESTIONS 863
Lighter-weight formal verification tools have been used heavily in produc-

tion [LBD+ 04, BBC+ 10, Coo18, SAE+ 18, DFLO19]. ❑
E.18 Important Questions
Quick Quiz A.1: p.632

What SMP coding errors can you see in these examples? See time.c for full code.
Answer:
Here are errors you might have found:
1. Missing barrier() or volatile on tight loops.
2. Missing memory barriers on update side.
3. Lack of synchronization between producer and consumer. ❑

How could there be such a large gap between successive consumer reads? See
timelocked.c for full code.
Answer:
Here are a few reasons for such gaps:
1. The consumer might be preempted for long time periods.
2. A long-running interrupt might delay the consumer.
3. Cache misses might delay the consumer.
4. The producer might also be running on a faster CPU than is the consumer (for
example, one of the CPUs might have had to decrease its clock frequency due to
heat-dissipation or power-consumption constraints). ❑

But if fully ordered implementations cannot offer stronger guarantees than the better
performing and more scalable weakly ordered implementations, why bother with full
ordering?
Answer:
Because strongly ordered implementations are sometimes able to provide greater
consistency among sets of calls to functions accessing a given data structure. For
example, compare the atomic counter of Listing 5.2 to the statistical counter of
Section 5.2. Suppose that one thread is adding the value 3 and another is adding the
value 5, while two other threads are concurrently reading the counter’s value. With
atomic counters, it is not possible for one of the readers to obtain the value 3 while the
other obtains the value 5. With statistical counters, this outcome really can happen.
In fact, in some computing environments, this outcome can happen even on relatively
strongly ordered hardware such as x86.
Therefore, if your user happen to need this admittedly unusual level of consistency,
you should avoid weakly ordered statistical counters. ❑
v2023.06.11a
Listing E.19: Deadlock in Lock-Based RCU Implementation

1 void foo(void)
2 {
3 spin_lock(&my_lock);
4 rcu_read_lock();
5 do_something();
9 }
10
11 void bar(void)
12 {
13 rcu_read_lock();
14 spin_lock(&my_lock);
15 do_some_other_thing();
17 do_whatever();
19 }

Suppose a portion of a program uses RCU read-side primitives as its only synchro-
nization mechanism. Is this parallelism or concurrency?
Answer:
Yes. ❑

In what part of the second (scheduler-based) perspective would the lock-based single-
thread-per-CPU workload be considered “concurrent”?
Answer:
The people who would like to arbitrarily subdivide and interleave the workload. Of
course, an arbitrary subdivision might end up separating a lock acquisition from the
corresponding lock release, which would prevent any other thread from acquiring that
lock. If the locks were pure spinlocks, this could even result in deadlock. ❑
E.19 “Toy” RCU Implementations
Quick Quiz B.1: p.642

Why wouldn’t any deadlock in the RCU implementation in Listing B.1 also be a
deadlock in any other RCU implementation?
Answer:
Suppose the functions foo() and bar() in Listing E.19 are invoked concurrently from
different CPUs. Then foo() will acquire my_lock() on line 3, while bar() will
acquire rcu_gp_lock on line 13.
When foo() advances to line 4, it will attempt to acquire rcu_gp_lock, which
is held by bar(). Then when bar() advances to line 14, it will attempt to acquire
my_lock, which is held by foo().
Each function is then waiting for a lock that the other holds, a classic deadlock.
v2023.06.11a
E.19. “TOY” RCU IMPLEMENTATIONS 865
Other RCU implementations neither spin nor block in rcu_read_lock(), hence

avoiding deadlocks. ❑

Why not simply use reader-writer locks in the RCU implementation in Listing B.1 in
order to allow RCU readers to proceed in parallel?
Answer:
One could in fact use reader-writer locks in this manner. However, textbook reader-writer
locks suffer from memory contention, so that the RCU read-side critical sections would
need to be quite long to actually permit parallel execution [McK03].
On the other hand, use of a reader-writer lock that is read-acquired in rcu_read_
lock() would avoid the deadlock condition noted above. ❑

Wouldn’t it be cleaner to acquire all the locks, and then release them all in the loop
from lines 15–18 of Listing B.2? After all, with this change, there would be a point in
time when there were no readers, simplifying things greatly.
Answer:
Making this change would re-introduce the deadlock, so no, it would not be cleaner. ❑

Is the implementation shown in Listing B.2 free from deadlocks? Why or why not?
Answer:
One deadlock is where a lock is held across synchronize_rcu(), and that same lock
is acquired within an RCU read-side critical section. However, this situation could
deadlock any correctly designed RCU implementation. After all, the synchronize_
rcu() primitive must wait for all pre-existing RCU read-side critical sections to
complete, but if one of those critical sections is spinning on a lock held by the thread
executing the synchronize_rcu(), we have a deadlock inherent in the definition of
RCU.
Another deadlock happens when attempting to nest RCU read-side critical sections.
This deadlock is peculiar to this implementation, and might be avoided by using recursive
locks, or by using reader-writer locks that are read-acquired by rcu_read_lock() and
write-acquired by synchronize_rcu().
However, if we exclude the above two cases, this implementation of RCU does
not introduce any deadlock situations. This is because only time some other thread’s
lock is acquired is when executing synchronize_rcu(), and in that case, the lock is
immediately released, prohibiting a deadlock cycle that does not involve a lock held
across the synchronize_rcu() which is the first case above. ❑
v2023.06.11a

Isn’t one advantage of the RCU algorithm shown in Listing B.2 that it uses only
primitives that are widely available, for example, in POSIX pthreads?
Answer:
This is indeed an advantage, but do not forget that rcu_dereference() and rcu_
assign_pointer() are still required, which means volatile manipulation for rcu_
dereference() and memory barriers for rcu_assign_pointer(). Of course, many
Alpha CPUs require memory barriers for both primitives. ❑

But what if you hold a lock across a call to synchronize_rcu(), and then acquire
that same lock within an RCU read-side critical section?
Answer:
Indeed, this would deadlock any legal RCU implementation. But is rcu_read_lock()
really participating in the deadlock cycle? If you believe that it is, then please ask
yourself this same question when looking at the RCU implementation in Appendix B.9.
❑

How can the grace period possibly elapse in 40 nanoseconds when synchronize_
rcu() contains a 10-millisecond delay?
Answer:
The update-side test was run in absence of readers, so the poll() system call was never
invoked. In addition, the actual code has this poll() system call commented out, the
better to evaluate the true overhead of the update-side code. Any production uses of this
code would be better served by using the poll() system call, but then again, production
uses would be even better served by other implementations shown later in this section.
❑

Why not simply make rcu_read_lock() wait when a concurrent synchronize_
rcu() has been waiting too long in the RCU implementation in Listing B.3? Wouldn’t
that prevent synchronize_rcu() from starving?
Answer:
Although this would in fact eliminate the starvation, it would also mean that rcu_
read_lock() would spin or block waiting for the writer, which is in turn waiting on
readers. If one of these readers is attempting to acquire a lock that the spinning/blocking
rcu_read_lock() holds, we again have deadlock.
In short, the cure is worse than the disease. See Appendix B.4 for a proper cure. ❑
v2023.06.11a

Why the memory barrier on line 5 of synchronize_rcu() in Listing B.6 given that
there is a spin-lock acquisition immediately after?
Answer:
The spin-lock acquisition only guarantees that the spin-lock’s critical section will not
“bleed out” to precede the acquisition. It in no way guarantees that code preceding the
spin-lock acquisition won’t be reordered into the critical section. Such reordering could
cause a removal from an RCU-protected list to be reordered to follow the complementing
of rcu_idx, which could allow a newly starting RCU read-side critical section to see
the recently removed data element.
Exercise for the reader: Use a tool such as Promela/spin to determine which (if any)
of the memory barriers in Listing B.6 are really needed. See Chapter 12 for information
on using these tools. The first correct and complete response will be credited. ❑

Why is the counter flipped twice in Listing B.6? Shouldn’t a single flip-and-wait cycle
be sufficient?
Answer:
Both flips are absolutely required. To see this, consider the following sequence of
events:
1 Line 8 of rcu_read_lock() in Listing B.5 picks up rcu_idx, finding its value

to be zero.
2 Line 8 of synchronize_rcu() in Listing B.6 complements the value of rcu_idx,

setting its value to one.
3 Lines 10–12 of synchronize_rcu() find that the value of rcu_refcnt[0]

is zero, and thus returns. (Recall that the question is asking what happens if
lines 13–20 are omitted.)
4 Lines 9 and 10 of rcu_read_lock() store the value zero to this thread’s instance
of rcu_read_idx and increments rcu_refcnt[0], respectively. Execution then
proceeds into the RCU read-side critical section.
5 Another instance of synchronize_rcu() again complements rcu_idx, this

time setting its value to zero. Because rcu_refcnt[1] is zero, synchronize_
rcu() returns immediately. (Recall that rcu_read_lock() incremented rcu_
refcnt[0], not rcu_refcnt[1]!)
6 The grace period that started in step 5 has been allowed to end, despite the fact
that the RCU read-side critical section that started beforehand in step 4 has not
completed. This violates RCU semantics, and could allow the update to free a data
element that the RCU read-side critical section was still referencing.
Exercise for the reader: What happens if rcu_read_lock() is preempted for a very
long time (hours!) just after line 8? Does this implementation operate correctly in that
case? Why or why not? The first correct and complete response will be credited. ❑
v2023.06.11a

Given that atomic increment and decrement are so expensive, why not just use non-
atomic increment on line 10 and a non-atomic decrement on line 25 of Listing B.5?
Answer:
Using non-atomic operations would cause increments and decrements to be lost, in turn
causing the implementation to fail. See Appendix B.5 for a safe way to use non-atomic
operations in rcu_read_lock() and rcu_read_unlock(). ❑

Come off it! We can see the atomic_read() primitive in rcu_read_lock()!!! So
why are you trying to pretend that rcu_read_lock() contains no atomic operations???
Answer:
The atomic_read() primitives does not actually execute atomic machine instructions,
but rather does a normal load from an atomic_t. Its sole purpose is to keep the
compiler’s type-checking happy. If the Linux kernel ran on 8-bit CPUs, it would also
need to prevent “store tearing”, which could happen due to the need to store a 16-bit
pointer with two eight-bit accesses on some 8-bit systems. But thankfully, it seems that
no one runs Linux on 8-bit systems. ❑

Great, if we have 𝑁 threads, we can have 2𝑁 ten-millisecond waits (one set per
flip_counter_and_wait() invocation, and even that assumes that we wait only
once for each thread). Don’t we need the grace period to complete much more quickly?
Answer:
Keep in mind that we only wait for a given thread if that thread is still in a pre-existing
RCU read-side critical section, and that waiting for one hold-out thread gives all the
other threads a chance to complete any pre-existing RCU read-side critical sections
that they might still be executing. So the only way that we would wait for 2𝑁 intervals
would be if the last thread still remained in a pre-existing RCU read-side critical section
despite all the waiting for all the prior threads. In short, this implementation will not
wait unnecessarily.
However, if you are stress-testing code that uses RCU, you might want to comment
out the poll() statement in order to better catch bugs that incorrectly retain a reference
to an RCU-protected data element outside of an RCU read-side critical section. ❑

All of these toy RCU implementations have either atomic operations in rcu_read_
lock() and rcu_read_unlock(), or synchronize_rcu() overhead that increases
linearly with the number of threads. Under what circumstances could an RCU
implementation enjoy lightweight implementations for all three of these primitives, all
having deterministic (O (1)) overheads and latencies?
v2023.06.11a
Answer:
Special-purpose uniprocessor implementations of RCU can attain this ideal [McK09a].
❑

If any even value is sufficient to tell synchronize_rcu() to ignore a given task, why
don’t lines 11 and 12 of Listing B.14 simply assign zero to rcu_reader_gp?
Answer:
Assigning zero (or any other even-numbered constant) would in fact work, but assigning
the value of rcu_gp_ctr can provide a valuable debugging aid, as it gives the developer
an idea of when the corresponding thread last exited an RCU read-side critical section.
❑

Why are the memory barriers on lines 19 and 31 of Listing B.14 needed? Aren’t the
memory barriers inherent in the locking primitives on lines 20 and 30 sufficient?
Answer:
These memory barriers are required because the locking primitives are only guaranteed
to confine the critical section. The locking primitives are under absolutely no obligation
to keep other code from bleeding in to the critical section. The pair of memory barriers
are therefore requires to prevent this sort of code motion, whether performed by the
compiler or by the CPU. ❑

Couldn’t the update-side batching optimization described in Appendix B.6 be applied
to the implementation shown in Listing B.14?
Answer:
Indeed it could, with a few modifications. This work is left as an exercise for the reader.
❑

Is the possibility of readers being preempted in lines 3–4 of Listing B.14 a real problem,
in other words, is there a real sequence of events that could lead to failure? If not, why
not? If so, what is the sequence of events, and how can the failure be addressed?
Answer:
It is a real problem, there is a sequence of events leading to failure, and there are a
number of possible ways of addressing it. For more details, see the Quick Quizzes near
the end of Appendix B.8. The reason for locating the discussion there is to (1) give you
more time to think about it, and (2) because the nesting support added in that section
greatly reduces the time required to overflow the counter. ❑
v2023.06.11a

Why not simply maintain a separate per-thread nesting-level variable, as was done in
previous section, rather than having all this complicated bit manipulation?
Answer:
The apparent simplicity of the separate per-thread variable is a red herring. This
approach incurs much greater complexity in the guise of careful ordering of operations,
especially if signal handlers are to be permitted to contain RCU read-side critical
sections. But don’t take my word for it, code it up and see what you end up with! ❑

Given the algorithm shown in Listing B.16, how could you double the time required
to overflow the global rcu_gp_ctr?
Answer:
One way would be to replace the magnitude comparison on lines 32 and 33 with an
inequality check of the per-thread rcu_reader_gp variable against rcu_gp_ctr+RCU_
GP_CTR_BOTTOM_BIT. ❑

Again, given the algorithm shown in Listing B.16, is counter overflow fatal? Why or
why not? If it is fatal, what can be done to fix it?
Answer:
It can indeed be fatal. To see this, consider the following sequence of events:
1. Thread 0 enters rcu_read_lock(), determines that it is not nested, and therefore

fetches the value of the global rcu_gp_ctr. Thread 0 is then preempted for an
extremely long time (before storing to its per-thread rcu_reader_gp variable).
2. Other threads repeatedly invoke synchronize_rcu(), so that the new value of

the global rcu_gp_ctr is now RCU_GP_CTR_BOTTOM_BIT less than it was when
thread 0 fetched it.
3. Thread 0 now starts running again, and stores into its per-thread rcu_reader_gp
variable. The value it stores is RCU_GP_CTR_BOTTOM_BIT+1 greater than that of
the global rcu_gp_ctr.
4. Thread 0 acquires a reference to RCU-protected data element A.
5. Thread 1 now removes the data element A that thread 0 just acquired a reference to.
6. Thread 1 invokes synchronize_rcu(), which increments the global rcu_gp_ctr

by RCU_GP_CTR_BOTTOM_BIT. It then checks all of the per-thread rcu_reader_
gp variables, but thread 0’s value (incorrectly) indicates that it started after thread 1’s
call to synchronize_rcu(), so thread 1 does not wait for thread 0 to complete
its RCU read-side critical section.
7. Thread 1 then frees up data element A, which thread 0 is still referencing.
v2023.06.11a
Note that scenario can also occur in the implementation presented in Appendix B.7.
One strategy for fixing this problem is to use 64-bit counters so that the time required
to overflow them would exceed the useful lifetime of the computer system. Note that
non-antique members of the 32-bit x86 CPU family allow atomic manipulation of 64-bit
counters via the cmpxchg64b instruction.
Another strategy is to limit the rate at which grace periods are permitted to occur in
order to achieve a similar effect. For example, synchronize_rcu() could record the
last time that it was invoked, and any subsequent invocation would then check this time
and block as needed to force the desired spacing. For example, if the low-order four bits
of the counter were reserved for nesting, and if grace periods were permitted to occur at
most ten times per second, then it would take more than 300 days for the counter to
overflow. However, this approach is not helpful if there is any possibility that the system
will be fully loaded with CPU-bound high-priority real-time threads for the full 300
days. (A remote possibility, perhaps, but best to consider it ahead of time.)
A third approach is to administratively abolish real-time threads from the system in
question. In this case, the preempted process will age up in priority, thus getting to run
long before the counter had a chance to overflow. Of course, this approach is less than
helpful for real-time applications.
A fourth approach would be for rcu_read_lock() to recheck the value of the global
rcu_gp_ctr after storing to its per-thread rcu_reader_gp counter, retrying if the
new value of the global rcu_gp_ctr is inappropriate. This works, but introduces
non-deterministic execution time into rcu_read_lock(). On the other hand, if your
application is being preempted long enough for the counter to overflow, you have no
hope of deterministic execution time in any case!
A fifth approach is for the grace period process to wait for all readers to become
aware of the new grace period. This works nicely in theory, but hangs if a reader blocks
indefinitely outside of an RCU read-side critical section.
A final approach is, oddly enough, to use a single-bit grace-period counter and for
each call to synchronize_rcu() to take two passes through its algorithm. This is the
approached use by userspace RCU [Des09b], and is described in detail in the journal
article and supplementary materials [DMS+ 12, Appendix D]. ❑

Doesn’t the additional memory barrier shown on line 14 of Listing B.18 greatly
increase the overhead of rcu_quiescent_state?
Answer:
Indeed it does! An application using this implementation of RCU should therefore
invoke rcu_quiescent_state sparingly, instead using rcu_read_lock() and rcu_
read_unlock() most of the time.
However, this memory barrier is absolutely required so that other threads will see the
store on lines 12–13 before any subsequent RCU read-side critical sections executed by
the caller. ❑

Why are the two memory barriers on lines 11 and 14 of Listing B.18 needed?
Answer:
The memory barrier on line 11 prevents any RCU read-side critical sections that might
v2023.06.11a
precede the call to rcu_thread_offline() won’t be reordered by either the compiler

or the CPU to follow the assignment on lines 12–13. The memory barrier on line 14 is,
strictly speaking, unnecessary, as it is illegal to have any RCU read-side critical sections
following the call to rcu_thread_offline(). ❑

To be sure, the clock frequencies of POWER systems in 2008 were quite high, but even
a 5 GHz clock frequency is insufficient to allow loops to be executed in 50 picoseconds!
What is going on here?
Answer:
Since the measurement loop contains a pair of empty functions, the compiler optimizes it
away. The measurement loop takes 1,000 passes between each call to rcu_quiescent_
state(), so this measurement is roughly one thousandth of the overhead of a single
call to rcu_quiescent_state(). ❑

Why would the fact that the code is in a library make any difference for how easy it is
to use the RCU implementation shown in Listings B.18 and B.19?
Answer:
A library function has absolutely no control over the caller, and thus cannot force the
caller to invoke rcu_quiescent_state() periodically. On the other hand, a library
function that made many references to a given RCU-protected data structure might
be able to invoke rcu_thread_online() upon entry, rcu_quiescent_state()
periodically, and rcu_thread_offline() upon exit. ❑

But what if you hold a lock across a call to synchronize_rcu(), and then acquire
that same lock within an RCU read-side critical section? This should be a deadlock,
but how can a primitive that generates absolutely no code possibly participate in a
deadlock cycle?
Answer:
Please note that the RCU read-side critical section is in effect extended beyond the
enclosing rcu_read_lock() and rcu_read_unlock(), out to the previous and next
call to rcu_quiescent_state(). This rcu_quiescent_state can be thought of as
an rcu_read_unlock() immediately followed by an rcu_read_lock().
Even so, the actual deadlock itself will involve the lock acquisition in the RCU
read-side critical section and the synchronize_rcu(), never the rcu_quiescent_
state(). ❑

Given that grace periods are prohibited within RCU read-side critical sections, how
can an RCU data structure possibly be updated while in an RCU read-side critical
section?
Answer:
This situation is one reason for the existence of asynchronous grace-period primitives
v2023.06.11a
E.20. WHY MEMORY BARRIERS? 873
such as call_rcu(). This primitive may be invoked within an RCU read-side critical
section, and the specified RCU callback will in turn be invoked at a later time, after a
grace period has elapsed.
The ability to perform an RCU update while within an RCU read-side critical
section can be extremely convenient, and is analogous to a (mythical) unconditional
read-to-write upgrade for reader-writer locking. ❑
E.20 Why Memory Barriers?
Quick Quiz C.1: p.665

Where does a writeback message originate from and where does it go to?
Answer:
The writeback message originates from a given CPU, or in some designs from a given
level of a given CPU’s cache—or even from a cache that might be shared among several
CPUs. The key point is that a given cache does not have room for a given data item, so
some other piece of data must be ejected from the cache to make room. If there is some
other piece of data that is duplicated in some other cache or in memory, then that piece
of data may be simply discarded, with no writeback message required.
On the other hand, if every piece of data that might be ejected has been modified so
that the only up-to-date copy is in this cache, then one of those data items must be copied
somewhere else. This copy operation is undertaken using a “writeback message”.
The destination of the writeback message has to be something that is able to store
the new value. This might be main memory, but it also might be some other cache.
If it is a cache, it is normally a higher-level cache for the same CPU, for example, a
level-1 cache might write back to a level-2 cache. However, some hardware designs
permit cross-CPU writebacks, so that CPU 0’s cache might send a writeback message
to CPU 1. This would normally be done if CPU 1 had somehow indicated an interest in
the data, for example, by having recently issued a read request.
In short, a writeback message is sent from some part of the system that is short of
space, and is received by some other part of the system that can accommodate the data.
❑

What happens if two CPUs attempt to invalidate the same cache line concurrently?
Answer:
One of the CPUs gains access to the shared bus first, and that CPU “wins”. The other
CPU must invalidate its copy of the cache line and transmit an “invalidate acknowledge”
message to the other CPU.
Of course, the losing CPU can be expected to immediately issue a “read invalidate”
transaction, so the winning CPU’s victory will be quite ephemeral. ❑
v2023.06.11a

When an “invalidate” message appears in a large multiprocessor, every CPU must give
an “invalidate acknowledge” response. Wouldn’t the resulting “storm” of “invalidate
acknowledge” responses totally saturate the system bus?
Answer:
It might, if large-scale multiprocessors were in fact implemented that way. Larger
multiprocessors, particularly NUMA machines, tend to use so-called “directory-based”
cache-coherence protocols to avoid this and other problems. ❑

If SMP machines are really using message passing anyway, why bother with SMP at
all?
Answer:
There has been quite a bit of controversy on this topic over the past few decades. One
answer is that the cache-coherence protocols are quite simple, and therefore can be
implemented directly in hardware, gaining bandwidths and latencies unattainable by
software message passing. Another answer is that the real truth is to be found in
economics due to the relative prices of large SMP machines and that of clusters of
smaller SMP machines. A third answer is that the SMP programming model is easier to
use than that of distributed systems, but a rebuttal might note the appearance of HPC
clusters and MPI. And so the argument continues. ❑

How does the hardware handle the delayed transitions described above?
Answer:
Usually by adding additional states, though these additional states need not be actually
stored with the cache line, due to the fact that only a few lines at a time will be
transitioning. The need to delay transitions is but one issue that results in real-world
cache coherence protocols being much more complex than the over-simplified MESI
protocol described in this appendix. Hennessy and Patterson’s classic introduction to
computer architecture [HP95] covers many of these issues. ❑

What sequence of operations would put the CPUs’ caches all back into the “invalid”
state?
Answer:
There is no such sequence, at least in absence of special “flush my cache” instructions
in the CPU’s instruction set. Most CPUs do have such instructions. ❑

But then why do uniprocessors also have store buffers?
Answer:
Because the purpose of store buffers is not just to hide acknowledgement latencies in
multiprocessor cache-coherence protocols, but to hide memory latencies in general.
v2023.06.11a
Because memory is much slower than is cache on uniprocessors, store buffers on

uniprocessors can help to hide write-miss memory latencies. ❑

So store-buffer entries are variable length? Isn’t that difficult to implement in hardware?
Answer:
Here are two ways for hardware to easily handle variable-length stores.
First, each store-buffer entry could be a single byte wide. Then an 64-bit store
would consume eight store-buffer entries. This approach is simple and flexible, but one
disadvantage is that each entry would need to replicate much of the address that was
stored to.
Second, each store-buffer entry could be double the size of a cache line, with half
of the bits containing the values stored, and the other half indicating which bits had
been stored to. So, assuming a 32-bit cache line, a single-byte store of 0x5a to the
low-order byte of a given cache line would result in 0xXXXXXX5a for the first half and
0x000000ff for the second half, where the values labeled X are arbitrary because they
would be ignored. This approach allows multiple consecutive stores corresponding to a
given cache line to be merged into a single store-buffer entry, but is space-inefficient for
random stores of single bytes.
Much more complex and efficient schemes are of course used by actual hardware
designers. ❑

In step 1 above, why does CPU 0 need to issue a “read invalidate” rather than a simple
“invalidate”? After all, foo() will overwrite the variable a in any case, so why should
it care about the old value of a?
Answer:
Because the cache line in question contains more data than just the variable a. Issuing
“invalidate” instead of the needed “read invalidate” would cause that other data to be
lost, which would constitute a serious bug in the hardware. ❑

In step 4 above, don’t systems avoid that store to memory?
Answer:
Yes, they do. But to do so, they add states beyond the MESI quadruple that this example
is working within. ❑

In step 9 above, did bar() read a stale value from a, or did its reads of b and a get
reordered?
Answer:
It could be either, depending on the hardware implementation. And it really does not
matter which. After all, the bar() function’s assert() cannot tell the difference! ❑
v2023.06.11a

After step 15 in Appendix C.3.3 on page 673, both CPUs might drop the cache line
containing the new value of “b”. Wouldn’t that cause this new value to be lost?
Answer:
It might, and that is why real hardware takes steps to avoid this problem. A traditional
approach, pointed out by Vasilevsky Alexander, is to write this cache line back to main
memory before marking the cache line as “shared”. A more efficient (though more
complex) approach is to use additional state to indicate whether or not the cache line is
“dirty”, allowing the writeback to happen. Year-2000 systems went further, using much
more state in order to avoid redundant writebacks [CSG99, Figure 8.42]. It would be
reasonable to assume that complexity has not decreased in the meantime. ❑

In step 1 of the first scenario in Appendix C.4.3, why is an “invalidate” sent instead of
a ”read invalidate” message? Doesn’t CPU 0 need the values of the other variables
that share this cache line with “a”?
Answer:
CPU 0 already has the values of these variables, given that it has a read-only copy of
the cache line containing “a”. Therefore, all CPU 0 need do is to cause the other CPUs
to discard their copies of this cache line. An “invalidate” message therefore suffices. ❑

Say what??? Why do we need a memory barrier here, given that the CPU cannot
possibly execute the assert() until after the while loop completes?
Answer:
Suppose that memory barrier was omitted.
Keep in mind that CPUs are free to speculatively execute later loads, which can have
the effect of executing the assertion before the while loop completes. Furthermore,
compilers assume that only the currently executing thread is updating the variables, and
this assumption allows the compiler to hoist the load of a to precede the loop.
In fact, some compilers would transform the loop to a branch around an infinite loop
as follows:
1 void foo(void)
2 {
3 a = 1;
4 smp_mb();
5 b = 1;
6 }
7
8 void bar(void)
9 {
10 if (b == 0)
11 for (;;)
12 continue;
13 assert(a == 1);
14 }
v2023.06.11a
Given this optimization, the code would behave in a completely different way than
the original code. If bar() observed “b == 0”, the assertion could of course not be
reached at all due to the infinite loop. However, if bar() loaded the value “1” just as
“foo()” stored it, the CPU might still have the old zero value of “a” in its cache, which
would cause the assertion to fire. You should of course use volatile casts (for example,
those volatile casts implied by the C11 relaxed atomic load operation) to prevent the
compiler from optimizing your parallel code into oblivion. But volatile casts would not
prevent a weakly ordered CPU from loading the old value for “a” from its cache, which
means that this code also requires the explicit memory barrier in “bar()”.
In short, both compilers and CPUs aggressively apply code-reordering optimizations,
so you must clearly communicate your constraints using the compiler directives and
memory barriers provided for this purpose. ❑

Instead of all of this marking of invalidation-queue entries and stalling of loads, why
not simply force an immediate flush of the invalidation queue?
Answer:
An immediate flush of the invalidation queue would do the trick. Except that the common-
case super-scalar CPU is executing many instructions at once, and not necessarily even
in the expected order. So what would “immediate” even mean? The answer is clearly
“not much”.
Nevertheless, for simpler CPUs that execute instructions serially, flushing the invali-
dation queue might be a reasonable implementation strategy. ❑

But can’t full memory barriers impose global ordering? After all, isn’t that needed to
provide the ordering shown in Listing 12.27?
Answer:
Sort of.
Note well that this litmus test has not one but two full memory-barrier instructions,
namely the two sync instructions executed by P2 and P3.
It is the interaction of those two instructions that provides the global ordering, not just
their individual execution. For example, each of those two sync instructions might stall
waiting for all CPUs to process their invalidation queues before allowing subsequent
instructions to execute.14 ❑

Does the guarantee that each CPU sees its own memory accesses in order also
guarantee that each user-level thread will see its own memory accesses in order? Why
or why not?
Answer:
No. Consider the case where a thread migrates from one CPU to another, and where
the destination CPU perceives the source CPU’s recent memory operations out of
14 Real-life hardware of course applies many optimizations to minimize the resulting

stalls.
v2023.06.11a
order. To preserve user-mode sanity, kernel hackers must use memory barriers in the
context-switch path. However, the locking already required to safely do a context switch
should automatically provide the memory barriers needed to cause the user-level task
to see its own accesses in order. That said, if you are designing a super-optimized
scheduler, either in the kernel or at user level, please keep this scenario in mind! ❑

Could this code be fixed by inserting a memory barrier between CPU 1’s “while” and
assignment to “c”? Why or why not?
Answer:
No. Such a memory barrier would only force ordering local to CPU 1. It would have no
effect on the relative ordering of CPU 0’s and CPU 1’s accesses, so the assertion could
still fail. However, all mainstream computer systems provide one mechanism or another
to provide “transitivity”, which provides intuitive causal ordering: If B saw the effects
of A’s accesses, and C saw the effects of B’s accesses, then C must also see the effects
of A’s accesses. In short, hardware designers have taken at least a little pity on software
developers. ❑

Suppose that lines 3–5 for CPUs 1 and 2 in Listing C.3 are in an interrupt handler,
and that the CPU 2’s line 9 runs at process level. In other words, the code in all
three columns of the table runs on the same CPU, but the first two columns run in
an interrupt handler, and the third column runs at process level, so that the code in
third column can be interrupted by the code in the first two columns. What changes, if
any, are required to enable the code to work correctly, in other words, to prevent the
assertion from firing?
Answer:
The assertion must ensure that the load of “e” precedes that of “a”. In the Linux kernel,
the barrier() primitive may be used to accomplish this in much the same way that
the memory barrier was used in the assertions in the previous examples. For example,
the assertion can be modified as follows:
r1 = e;
barrier();
assert(r1 == 0 || a == 1);
No changes are needed to the code in the first two columns, because interrupt handlers
run atomically from the perspective of the interrupted code. ❑

If CPU 2 executed an assert(e==0||c==1) in the example in Listing C.3, would
this assert ever trigger?
Answer:
The result depends on whether the CPU supports “transitivity”. In other words, CPU 0
stored to “e” after seeing CPU 1’s store to “c”, with a memory barrier between CPU 0’s
v2023.06.11a
load from “c” and store to “e”. If some other CPU sees CPU 0’s store to “e”, is it also
guaranteed to see CPU 1’s store?
All CPUs I am aware of claim to provide transitivity. ❑
v2023.06.11a
v2023.06.11a
Dictionaries are inherently circular in nature.
Self Reference in word definitions, David Levary et al.
Glossary
Acquire Load: A read from memory that has acquire semantics. Normal use cases
pair an acquire load with a release store, in which case if the load returns the
value stored, then all code executed by the loading CPU after that acquire load will
see the effects of all memory-reference instructions executed by the storing CPU
prior to that release store. Acquiring a lock provides similar memory-ordering
semantics, hence the “acquire” in “acquire load”. (See also “memory barrier” and
“release store”.)
Amdahl’s Law: If sufficient numbers of CPUs are used to run a job that has both a
sequential portion and a concurrent portion, performance and scalability will be
limited by the overhead of the sequential portion.
Associativity: The number of cache lines that can be held simultaneously in a given
cache, when all of these cache lines hash identically in that cache. A cache
that could hold four cache lines for each possible hash value would be termed a
“four-way set-associative” cache, while a cache that could hold only one cache
line for each possible hash value would be termed a “direct-mapped” cache. A
cache whose associativity was equal to its capacity would be termed a “fully
associative” cache. Fully associative caches have the advantage of eliminating
associativity misses, but, due to hardware limitations, fully associative caches
are normally quite limited in size. The associativity of the large caches found on
modern microprocessors typically range from two-way to eight-way.
Associativity Miss: A cache miss incurred because the corresponding CPU has recently
accessed more data hashing to a given set of the cache than will fit in that set. Fully
associative caches are not subject to associativity misses (or, equivalently, in fully
associative caches, associativity and capacity misses are identical).
Atomic: An operation is considered “atomic” if it is not possible to observe any
intermediate state. For example, on most CPUs, a store to a properly aligned
pointer is atomic, because other CPUs will see either the old value or the new
value, but are guaranteed not to see some mixed value containing some pieces of
the new and old values.
Atomic Read-Modify-Write Operation: An atomic operation that both reads and
writes memory is considered an atomic read-modify-write operation, or atomic
RMW operation for short. Although the value written usually depends on the value
read, atomic_xchg() is the exception that proves this rule.
Bounded Wait Free: A forward-progress guarantee in which every thread makes
progress within a specific finite period of time, the specific time being the bound.
881
v2023.06.11a
882 GLOSSARY
Bounded Population-Oblivious Wait Free: A forward-progress guarantee in which

every thread makes progress within a specific finite period of time, the specific
time being the bound, where this bound is independent of the number of threads.
Cache: In modern computer systems, CPUs have caches in which to hold frequently
used data. These caches can be thought of as hardware hash tables with very
simple hash functions, but in which each hash bucket (termed a “set” by hardware
types) can hold only a limited number of data items. The number of data items that
can be held by each of a cache’s hash buckets is termed the cache’s “associativity”.
These data items are normally called “cache lines”, which can be thought of a
fixed-length blocks of data that circulate among the CPUs and memory.
Cache Coherence: A property of most modern SMP machines where all CPUs will
observe a sequence of values for a given variable that is consistent with at least
one global order of values for that variable. Cache coherence also guarantees that
at the end of a group of stores to a given variable, all CPUs will agree on the final
value for that variable. Note that cache coherence applies only to the series of
values taken on by a single variable. In contrast, the memory consistency model
for a given machine describes the order in which loads and stores to groups of
variables will appear to occur. See Section 15.2.6 for more information.
Cache-Coherence Protocol: A communications protocol, normally implemented in
hardware, that enforces memory consistency and ordering, preventing different
CPUs from seeing inconsistent views of data held in their caches.
Cache Geometry: The size and associativity of a cache is termed its geometry. Each
cache may be thought of as a two-dimensional array, with rows of cache lines
(“sets”) that have the same hash value, and columns of cache lines (“ways”) in
which every cache line has a different hash value. The associativity of a given
cache is its number of columns (hence the name “way”—a two-way set-associative
cache has two “ways”), and the size of the cache is its number of rows multiplied
by its number of columns.
Cache Line: (1) The unit of data that circulates among the CPUs and memory, usually
a moderate power of two in size. Typical cache-line sizes range from 16 to 256
bytes.
(2) A physical location in a CPU cache capable of holding one cache-line unit of
data.
(3) A physical location in memory capable of holding one cache-line unit of data,
but that it also aligned on a cache-line boundary. For example, the address of the
first word of a cache line in memory will end in 0x00 on systems with 256-byte
cache lines.
Cache Miss: A cache miss occurs when data needed by the CPU is not in that CPU’s
cache. The data might be missing because of a number of reasons, including:
(1) This CPU has never accessed the data before (“startup” or “warmup” miss),
(2) This CPU has recently accessed more data than would fit in its cache, so that
some of the older data had to be removed (“capacity” miss), (3) This CPU has
recently accessed more data in a given set1 than that set could hold (“associativity”
1 In hardware-cache terminology, the word “set” is used in the same way that the word
“bucket” is used when discussing software caches.
v2023.06.11a
883
miss), (4) Some other CPU has written to the data (or some other data in the same
cache line) since this CPU has accessed it (“communication miss”), or (5) This
CPU attempted to write to a cache line that is currently read-only, possibly due to
that line being replicated in other CPUs’ caches.
Capacity Miss: A cache miss incurred because the corresponding CPU has recently
accessed more data than will fit into the cache.
CAS: Compare-and-swap operation, which is an atomic operation that takes a pointer,

and old value, and a new value. If the pointed-to value is equal to the old value, it
is atomically replaced with the new value. There is some variety in CAS API. One
variation returns the actual pointed-to value, so that the caller compares the CAS
return value to the specified old value, with equality indicating a successful CAS
operation. Another variation returns a boolean success indication, in which case a
pointer to the old value may be passed in, and if so, the old value is updated in the
CAS failure case.
Clash Free: A forward-progress guarantee in which, in the absence of contention, at

least one thread makes progress within a finite period of time.
Code Locking: A simple locking design in which a “global lock” is used to protect
a set of critical sections, so that access by a given thread to that set is granted
or denied based only on the set of threads currently occupying the set of critical
sections, not based on what data the thread intends to access. The scalability of
a code-locked program is limited by the code; increasing the size of the data set
will normally not increase scalability (in fact, will typically decrease scalability by
increasing “lock contention”). Contrast with “data locking”.
Combinatorial Explosion: Denotes the exponential increase in executions that formal-

verification tools must analyze as problem size increases.
Combinatorial Implosion: Denotes the exponential decrease in executions that formal-

verification tools must analyze when a given code fragment is partitioned.
Communication Miss: A cache miss incurred because some other CPU has written to
the cache line since the last time this CPU accessed it.
Concurrent: In this book, a synonym of parallel. Please see Appendix A.6 on page 638
for a discussion of the recent distinction between these two terms.
Critical Section: A section of code guarded by some synchronization mechanism, so

that its execution constrained by that primitive. For example, if a set of critical
sections are guarded by the same global lock, then only one of those critical
sections may be executing at a given time. If a thread is executing in one such
critical section, any other threads must wait until the first thread completes before
executing any of the critical sections in the set.
Data Locking: A scalable locking design in which each instance of a given data
structure has its own lock. If each thread is using a different instance of the
data structure, then all of the threads may be executing in the set of critical
sections simultaneously. Data locking has the advantage of automatically scaling
to increasing numbers of CPUs as the number of instances of data grows. Contrast
with “code locking”.
v2023.06.11a
884 GLOSSARY
Data Race: A race condition in which several CPUs or threads access a variable
concurrently, and in which at least one of those accesses is a store and at least one
of those accesses is a plain access. It is important to note that while the presence
of data races often indicates the presence of bugs, the absence of data races in no
way implies the absence of bugs. (See “Plain access” and “Race condition”.)
Deadlock: A failure mode in which each of several threads is unable to make progress
until some other thread makes progress. For example, if two threads acquire a pair
of locks in opposite orders, deadlock can result. More information is provided in
Section 7.1.1.
Deadlock Free: A forward-progress guarantee in which, in the absence of failures, at
least one thread makes progress within a finite period of time.
Direct-Mapped Cache: A cache with only one way, so that it may hold only one cache
line with a given hash value.
Efficiency: A measure of effectiveness normally expressed as a ratio of some metric
actually achieved to some maximum value. The maximum value might be a theo-
retical maximum, but in parallel programming is often based on the corresponding
measured single-threaded metric.
Embarrassingly Parallel: A problem or algorithm where adding threads does not
significantly increase the overall cost of the computation, resulting in linear
speedups as threads are added (assuming sufficient CPUs are available).
Energy Efficiency: Shorthand for “energy-efficient use” in which the goal is to carry
out a given computation with reduced energy consumption. Sublinear scalability
can be an obstacle to energy-efficient use of a multicore system.
Epoch-Based Reclamation (EBR): An RCU implementation style put forward by
Keir Fraser [Fra03, Fra04, FH07].
Existence Guarantee: An existence guarantee is provided by a synchronization
mechanism that prevents a given dynamically allocated object from being freed for
the duration of that guarantee. For example, RCU provides existence guarantees
for the duration of RCU read-side critical sections. A similar but strictly weaker
guarantee is provided by type-safe memory.
Exclusive Lock: An exclusive lock is a mutual-exclusion mechanism that permits only
one thread at a time into the set of critical sections guarded by that lock.
False Sharing: If two CPUs each frequently write to one of a pair of data items, but
the pair of data items are located in the same cache line, this cache line will be
repeatedly invalidated, “ping-ponging” back and forth between the two CPUs’
caches. This is a common cause of “cache thrashing”, also called “cacheline
bouncing” (the latter most commonly in the Linux community). False sharing can
dramatically reduce both performance and scalability.
Forward-Progress Guarantee: Algorithms or programs that guarantee that execution
will progress at some rate under specified conditions. Academic forward-progress
guarantees are grouped into a formal hierarchy shown in Section 14.2. A wide
variety of practical forward-progress guarantees are provided by real-time systems,
as discussed in Section 14.3.
v2023.06.11a
885
Fragmentation: A memory pool that has a large amount of unused memory, but not
laid out to permit satisfying a relatively small request is said to be fragmented.
External fragmentation occurs when the space is divided up into small fragments
lying between allocated blocks of memory, while internal fragmentation occurs
when specific requests or types of requests have been allotted more memory than
they actually requested.
Fully Associative Cache: A fully associative cache contains only one set, so that it can
hold any subset of memory that fits within its capacity.
Grace Period: A grace period is any contiguous time interval such that any RCU
read-side critical section that began before the start of that interval has completed
before the end of that same interval. Many RCU implementations define a grace
period to be a time interval during which each thread has passed through at least
one quiescent state. Since RCU read-side critical sections by definition cannot
contain quiescent states, these two definitions are almost always interchangeable.
Hardware Transactional Memory (HTM): A transactional-memory system based

on hardware instructions provided for this purpose, as discussed in Section 17.3.
(See “Transactional memory”.)
Hazard Pointer: A scalable counterpart to a reference counter in which an object’s

reference count is represented implicitly by a count of the number of special hazard
pointers referencing that object.
Heisenbug: A timing-sensitive bug that disappears from sight when you add print
statements or tracing in an attempt to track it down.
Hot Spot: Data structure that is very heavily used, resulting in high levels of contention
on the corresponding lock. One example of this situation would be a hash table
with a poorly chosen hash function.
Humiliatingly Parallel: A problem or algorithm where adding threads significantly

decreases the overall cost of the computation, resulting in large superlinear speedups
as threads are added (assuming sufficient CPUs are available).
Immutable: In this book, a synonym for read-mostly.
Invalidation: When a CPU wishes to write to a data item, it must first ensure that this
data item is not present in any other CPUs’ cache. If necessary, the item is removed
from the other CPUs’ caches via “invalidation” messages from the writing CPUs
to any CPUs having a copy in their caches.
IPI: Inter-processor interrupt, which is an interrupt sent from one CPU to another.
IPIs are used heavily in the Linux kernel, for example, within the scheduler to alert
CPUs that a high-priority process is now runnable.
IRQ: Interrupt request, often used as an abbreviation for “interrupt” within the Linux
kernel community, as in “irq handler”.
Latency: The wall-clock time required for a given operation to complete.
v2023.06.11a
886 GLOSSARY
Linearizable: A sequence of operations is “linearizable” if there is at least one global

ordering of the sequence that is consistent with the observations of all CPUs and/or
threads. Linearizability is much prized by many researchers, but less useful in
practice than one might expect [HKLP12].
Livelock: A failure mode in which each of several threads is able to execute, but in
which a repeating series of failed operations prevents any of the threads from
making any useful forward progress. For example, incorrect use of conditional
locking (for example, spin_trylock() in the Linux kernel) can result in livelock.
More information is provided in Section 7.1.2.
Lock: A software abstraction that can be used to guard critical sections, as such, an
example of a “mutual exclusion mechanism”. An “exclusive lock” permits only
one thread at a time into the set of critical sections guarded by that lock, while
a “reader-writer lock” permits any number of reading threads, or but one writing
thread, into the set of critical sections guarded by that lock. (Just to be clear, the
presence of a writer thread in any of a given reader-writer lock’s critical sections
will prevent any reader from entering any of that lock’s critical sections and vice
versa.)
Lock Contention: A lock is said to be suffering contention when it is being used so

heavily that there is often a CPU waiting on it. Reducing lock contention is often
a concern when designing parallel algorithms and when implementing parallel
programs.
Lock Free: A forward-progress guarantee in which at least one thread makes progress
within a finite period of time.
Marked Access: A source-code memory access that uses a special function or macro,
such as READ_ONCE(), WRITE_ONCE(), atomic_inc(), and so on, in order to
protect that access from compiler and/or hardware optimizations. In contrast, a
plain access simply mentions the name of the object being accessed, so that in the
following, line 2 is the plain-access equivalent of line 1:
1 WRITE_ONCE(a, READ_ONCE(b) + READ_ONCE(c));

2 a = b + c;
Memory: From the viewpoint of memory models, the main memory, caches, and store
buffers in which values might be stored. However, this term is often used to denote
the main memory itself, excluding caches and store buffers.
Memory Barrier: A compiler directive that might also include a special memory-
barrier instruction. The purpose of a memory barrier is to order memory-reference
instructions that executed before the memory barrier to precede those that will
execute following that memory barrier. (See also “read memory barrier” and
“write memory barrier”.)
Memory Consistency: A set of properties that impose constraints on the order in which
accesses to groups of variables appear to occur. Memory consistency models
range from sequential consistency, a very constraining model popular in academic
circles, through process consistency, release consistency, and weak consistency.
v2023.06.11a
887
MESI Protocol: The cache-coherence protocol featuring modified, exclusive, shared,

and invalid (MESI) states, so that this protocol is named after the states that
the cache lines in a given cache can take on. A modified line has been recently
written to by this CPU, and is the sole representative of the current value of the
corresponding memory location. An exclusive cache line has not been written to,
but this CPU has the right to write to it at any time, as the line is guaranteed not
to be replicated into any other CPU’s cache (though the corresponding location
in main memory is up to date). A shared cache line is (or might be) replicated in
some other CPUs’ cache, meaning that this CPU must interact with those other
CPUs before writing to this cache line. An invalid cache line contains no value,
instead representing “empty space” in the cache into which data from memory
might be loaded.
Moore’s Law: A 1965 empirical projection by Gordon Moore that transistor density
increases exponentially over time [Moo65].
Mutual-Exclusion Mechanism: A software abstraction that regulates threads’ access

to “critical sections” and corresponding data.
NMI: Non-maskable interrupt. As the name indicates, this is an extremely high-priority

interrupt that cannot be masked. These are used for hardware-specific purposes
such as profiling. The advantage of using NMIs for profiling is that it allows you
to profile code that runs with interrupts disabled.
Non-Blocking: A group of academic forward-progress guarantees that includes bounded

population-oblivious wait free, bounded wait free, wait free, lock free, obstruction
free, clash free, starvation free, and deadlock free. See Section 14.2 for more
information.
Non-Blocking Synchronization (NBS): The use of algorithms, mechanisms, or

techniques that provide non-blocking forward-progress guarantees. NBS is often
used in a more restrictive sense of providing one of the stronger forward-progress
guarantees, usually wait free or lock free, but sometimes also obstruction free.
(See “Non-blocking”.)
NUCA: Non-uniform cache architecture, where groups of CPUs share caches and/or
store buffers. CPUs in a group can therefore exchange cache lines with each other
much more quickly than they can with CPUs in other groups. Systems comprised
of CPUs with hardware threads will generally have a NUCA architecture.
NUMA: Non-uniform memory architecture, where memory is split into banks and
each such bank is “close” to a group of CPUs, the group being termed a “NUMA
node”. An example NUMA machine is Sequent’s NUMA-Q system, where each
group of four CPUs had a bank of memory nearby. The CPUs in a given group can
access their memory much more quickly than another group’s memory.
NUMA Node: A group of closely placed CPUs and associated memory within a larger
NUMA machines.
Obstruction Free: A forward-progress guarantee in which, in the absence of contention,

every thread makes progress within a finite period of time.
v2023.06.11a
888 GLOSSARY
Overhead: Operations that must be executed, but which do not contribute directly to
the work that must be accomplished. For example, lock acquisition and release
is normally considered to be overhead, and specifically to be synchronization
overhead.
Parallel: In this book, a synonym of concurrent. Please see Appendix A.6 on page 638
for a discussion of the recent distinction between these two terms.
Performance: Rate at which work is done, expressed as work per unit time. If this
work is fully serialized, then the performance will be the reciprocal of the mean
latency of the work items.
Pipelined CPU: A CPU with a pipeline, which is an internal flow of instructions

internal to the CPU that is in some way similar to an assembly line, with many
of the same advantages and disadvantages. In the 1960s through the early 1980s,
pipelined CPUs were the province of supercomputers, but started appearing in
microprocessors (such as the 80486) in the late 1980s.
Plain Access: A source-code memory access that simply mentions the name of the
object being accessed. (See “Marked access”.)
Process Consistency: A memory-consistency model in which each CPU’s stores appear

to occur in program order, but in which different CPUs might see accesses from
more than one CPU as occurring in different orders.
Program Order: The order in which a given thread’s instructions would be executed by
a now-mythical “in-order” CPU that completely executed each instruction before
proceeding to the next instruction. (The reason such CPUs are now the stuff of
ancient myths and legends is that they were extremely slow. These dinosaurs were
one of the many victims of Moore’s-Law-driven increases in CPU clock frequency.
Some claim that these beasts will roam the earth once again, others vehemently
disagree.)
Quiescent State: In RCU, a point in the code where there can be no references held to
RCU-protected data structures, which is normally any point outside of an RCU
read-side critical section. Any interval of time during which all threads pass
through at least one quiescent state each is termed a “grace period”.
Quiescent-State-Based Reclamation (QSBR): An RCU implementation style char-

acterized by explicit quiescent states. In QSBR implementations, read-side
markers (rcu_read_lock() and rcu_read_unlock() in the Linux kernel) are
no-ops [MS98a, SM95]. Hooks in other parts of the software (for example, the
Linux-kernel scheduler) provide the quiescent states.
Race Condition: Any situation where multiple CPUs or threads can interact, though
this term is often used in cases where such interaction is undesirable. (See “Data
race”.)
RCU-Protected Data: A block of dynamically allocated memory whose freeing will

be deferred such that an RCU grace period will elapse between the time that there
were no longer any RCU-reader-accessible pointers to that block and the time that
that block is freed. This ensures that no RCU readers will have access to that block
at the time that it is freed.
v2023.06.11a
889
RCU-Protected Pointer: A pointer to RCU-protected data. Such pointers must be

handled carefully, for example, any reader that intends to dereference an RCU-
protected pointer must use rcu_dereference() (or stronger) to load that pointer,
and any updater must use rcu_assign_pointer() (or stronger) to store to that
pointer. More information is provided in Section 15.3.2.
RCU Read-Side Critical Section: A section of code protected by RCU, for example,
beginning with rcu_read_lock() and ending with rcu_read_unlock(). (See
“Read-side critical section”.)
Read-Copy Update (RCU): A synchronization mechanism that can be thought of

as a replacement for reader-writer locking or reference counting. RCU provides
extremely low-overhead access for readers, while writers incur additional overhead
maintaining old versions for the benefit of pre-existing readers. Readers neither
block nor spin, and thus cannot participate in deadlocks, however, they also can
see stale data and can run concurrently with updates. RCU is thus best-suited for
read-mostly situations where stale data can either be tolerated (as in routing tables)
or avoided (as in the Linux kernel’s System V IPC implementation).
Read Memory Barrier: A memory barrier that is only guaranteed to affect the ordering
of load instructions, that is, reads from memory. (See also “memory barrier” and
“write memory barrier”.)
Read Mostly: Read-mostly data is (again, as the name implies) rarely updated. However,
it might be updated at any time.
Read Only: Read-only data is, as the name implies, never updated except by beginning-
of-time initialization. In this book, a synonym for immutable.
Read-Side Critical Section: A section of code guarded by read-acquisition of some

reader-writer synchronization mechanism. For example, if one set of critical
sections are guarded by read-acquisition of a given global reader-writer lock,
while a second set of critical section are guarded by write-acquisition of that
same reader-writer lock, then the first set of critical sections will be the read-side
critical sections for that lock. Any number of threads may concurrently execute the
read-side critical sections, but only if no thread is executing one of the write-side
critical sections. (See also “RCU read-side critical section”.)
Reader-Writer Lock: A reader-writer lock is a mutual-exclusion mechanism that

permits any number of reading threads, or but one writing thread, into the set
of critical sections guarded by that lock. Threads attempting to write must wait
until all pre-existing reading threads release the lock, and, similarly, if there is
a pre-existing writer, any threads attempting to write must wait for the writer to
release the lock. A key concern for reader-writer locks is “fairness”: Can an
unending stream of readers starve a writer or vice versa?
Real Time: A situation in which getting the correct result is not sufficient, but where
this result must also be obtained within a given amount of time.
Reference Count: A counter that tracks the number of users of a given object or
entity. Reference counters provide existence guarantees and are sometimes used to
implement garbage collectors.
v2023.06.11a
890 GLOSSARY
Release Store: A write to memory that has release semantics. Normal use cases pair
an acquire load with a release store, in which case if the load returns the value
stored, then all code executed by the loading CPU after that acquire load will see
the effects of all memory-reference instructions executed by the storing CPU prior
to that release store. Releasing a lock provides similar memory-ordering semantics,
hence the “release” in “release store”. (See also “acquire load” and “memory
barrier”.)
Scalability: A measure of how effectively a given system is able to utilize additional

resources. For parallel computing, the additional resources are usually additional
CPUs.
Sequence Lock: A reader-writer synchronization mechanism in which readers retry

their operations if a writer was present.
Sequential Consistency: A memory-consistency model where all memory references

appear to occur in an order consistent with a single global order, and where each
CPU’s memory references appear to all CPUs to occur in program order.
Software Transactional Memory (HTM): A transactional-memory system capable

running on computer systems without special hardware support. (See “Transac-
tional memory”.)
Starvation: A condition where at least one CPU or thread is unable to make progress
due to an unfortunate series of resource-allocation decisions, as discussed in
Section 7.1.2. For example, in a multisocket system, CPUs on one socket having
privileged access to the data structure implementing a given lock could prevent
CPUs on other sockets from ever acquiring that lock.
Starvation Free: A forward-progress guarantee in which, in the absence of failures,

every thread makes progress within a finite period of time.
Store Buffer: A small set of internal registers used by a given CPU to record pending
stores while the corresponding cache lines are making their way to that CPU. Also
called “store queue”.
Store Forwarding: An arrangement where a given CPU refers to its store buffer as
well as its cache so as to ensure that the software sees the memory operations
performed by this CPU as if they were carried out in program order.
Superscalar CPU: A scalar (non-vector) CPU capable of executing multiple instructions

concurrently. This is a step up from a pipelined CPU that executes multiple
instructions in an assembly-line fashion—in a superscalar CPU, each stage of the
pipeline would be capable of handling more than one instruction. For example, if
the conditions were exactly right, the Intel Pentium Pro CPU from the mid-1990s
could execute two (and sometimes three) instructions per clock cycle. Thus, a
200 MHz Pentium Pro CPU could “retire”, or complete the execution of, up to 400
million instructions per second.
Synchronization: Means for avoiding destructive interactions among CPUs or threads.

Synchronization mechanisms include atomic RMW operations, memory barriers,
locking, reference counting, hazard pointers, sequence locking, RCU, non-blocking
synchronization, and transactional memory.
v2023.06.11a
891
Teachable: A topic, concept, method, or mechanism that teachers believe that they
understand completely and are therefore comfortable teaching.
Throughput: A performance metric featuring work items completed per unit time.
Transactional Lock Elision (TLE): The use of transactional memory to emulate

locking. Synchronization is instead carried out by conflicting accesses to the data
to be protected by the lock. In some cases, this can increase performance because
TLE avoids contention on the lock word [PD11, Kle14, FIMR16, PMDY20].
Transactional Memory (TM): A synchronization mechanism that gathers groups

of memory accesses so as to execute them atomically from the viewpoint of
transactions on other CPUs or threads, discussed in Sections 17.2 and 17.3.
Type-Safe Memory: Type-safe memory [GC96] is provided by a synchronization

mechanism that prevents a given dynamically allocated object from changing to an
incompatible type. Note that the object might well be freed and then reallocated,
but the reallocated object is guaranteed to be of a compatible type. Within the
Linux kernel, type-safe memory is provided within RCU read-side critical sections
for memory allocated from slabs marked with the SLAB_TYPESAFE_BY_RCU flag.
The strictly stronger existence guarantee also prevents freeing of the protected
object.
Unbounded Transactional Memory (UTM): A transactional-memory system based

on hardware instructions provided for this purpose, but with special hardware or
software capabilities that allow a given transaction to have a very large memory
footprint. Such a system would at least partially avoid HTM’s transaction-size
limitations called out in Section 17.3.2.1. (See “Hardware transactional memory”.)
Unfairness: A condition where the progress of at least one CPU or thread is impeded by
an unfortunate series of resource-allocation decisions, as discussed in Section 7.1.2.
Extreme levels of unfairness are termed “starvation”.
Unteachable: A topic, concept, method, or mechanism that the teacher does not
understand well is therefore uncomfortable teaching.
Vector CPU: A CPU that can apply a single instruction to multiple items of data
concurrently. In the 1960s through the 1980s, only supercomputers had vector
capabilities, but the advent of MMX in x86 CPUs and VMX in PowerPC CPUs
brought vector processing to the masses.
Wait Free: A forward-progress guarantee in which every thread makes progress within
a finite period of time.
Write Memory Barrier: A memory barrier that is only guaranteed to affect the ordering
of store instructions, that is, writes to memory. (See also “memory barrier” and
“read memory barrier”.)
Write Miss: A cache miss incurred because the corresponding CPU attempted to write
to a cache line that is read-only, most likely due to its being replicated in other
CPUs’ caches.
Write Mostly: Write-mostly data is (yet again, as the name implies) frequently updated.
v2023.06.11a
892 GLOSSARY
Write-Side Critical Section: A section of code guarded by write-acquisition of some

reader-writer synchronization mechanism. For example, if one set of critical
sections are guarded by write-acquisition of a given global reader-writer lock,
while a second set of critical section are guarded by read-acquisition of that same
reader-writer lock, then the first set of critical sections will be the write-side critical
sections for that lock. Only one thread may execute in the write-side critical section
at a time, and even then only if there are no threads are executing concurrently in
any of the corresponding read-side critical sections.
v2023.06.11a
Bibliography
[AA14] Maya Arbel and Hagit Attiya. Concurrent updates with RCU: Search
tree as an example. In Proceedings of the 2014 ACM Symposium on
Principles of Distributed Computing, PODC ’14, page 196–205, Paris,
France, 2014. ACM.
[AAKL06] C. Scott Ananian, Krste Asanovic, Bradley C. Kuszmaul, and Charles E.
Leiserson. Unbounded transactional memory. IEEE Micro, pages 59–69,
January-February 2006.
[AB13] Samy Al Bahra. Nonblocking algorithms and scalable multicore pro-
gramming. Commun. ACM, 56(7):50–61, July 2013.
[ABD+ 97] Jennifer M. Anderson, Lance M. Berc, Jeffrey Dean, Sanjay Ghemawat,
Monika R. Henzinger, Shun-Tak A. Leung, Richard L. Sites, Mark T.
Vandevoorde, Carl A. Waldspurger, and William E. Weihl. Continuous
profiling: Where have all the cycles gone? In Proceedings of the 16th
ACM Symposium on Operating Systems Principles, pages 1–14, New
York, NY, October 1997.
[ACA+ 18] A. Aljuhni, C. E. Chow, A. Aljaedi, S. Yusuf, and F. Torres-Reyes.
Towards understanding application performance and system behavior
with the full dynticks feature. In 2018 IEEE 8th Annual Computing
and Communication Workshop and Conference (CCWC), pages 394–401,
2018.
[ACHS13] Dan Alistarh, Keren Censor-Hillel, and Nir Shavit. Are lock-
free concurrent algorithms practically wait-free?, December 2013.
ArXiv:1311.3200v2.
[ACMS03] Andrea Arcangeli, Mingming Cao, Paul E. McKenney, and Dipankar
Sarma. Using read-copy update techniques for System V IPC in the
Linux 2.5 kernel. In Proceedings of the 2003 USENIX Annual Technical
Conference (FREENIX Track), pages 297–310, San Antonio, Texas, USA,
June 2003. USENIX Association.
[Ada11] Andrew Adamatzky. Slime mould solves maze in one pass . . . assisted
by gradient of chemo-attractants, August 2011. arXiv:1108.4956.
[ADF+ 19] Jade Alglave, Will Deacon, Boqun Feng, David Howells, Daniel Lustig,
Luc Maranget, Paul E. McKenney, Andrea Parri, Nicholas Piggin, Alan
Stern, Akira Yokosawa, and Peter Zĳlstra. Who’s afraid of a big bad
optimizing compiler?, July 2019. Linux Weekly News.
893
v2023.06.11a
894 BIBLIOGRAPHY
[Adv02] Advanced Micro Devices. AMD x86-64 Architecture Programmer’s

Manual Volumes 1–5, 2002.
[AGH+ 11a] Hagit Attiya, Rachid Guerraoui, Danny Hendler, Petr Kuznetsov,
Maged M. Michael, and Martin Vechev. Laws of order: Expensive
synchronization in concurrent algorithms cannot be eliminated. In 38th
ACM SIGACT-SIGPLAN Symposium on Principles of Programming
Languages, pages 487–498, Austin, TX, USA, 2011. ACM.
[AGH+ 11b] Hagit Attiya, Rachid Guerraoui, Danny Hendler, Petr Kuznetsov,
Maged M. Michael, and Martin Vechev. Laws of order: Expensive
synchronization in concurrent algorithms cannot be eliminated. SIG-
PLAN Not., 46(1):487–498, January 2011.
[AH22] Krzysztof R. Apt and Tony Hoare, editors. Edsger Wybe Dĳkstra: His
Life,Work, and Legacy, volume 45. Association for Computing Machinery,
New York, NY, USA, 1 edition, 2022.
[AHM09] Hagit Attiya, Eshcar Hillel, and Alessia Milani. Inherent limitations
on disjoint-access parallel implementations of transactional memory.
In Proceedings of the twenty-first annual symposium on Parallelism in
algorithms and architectures, SPAA ’09, pages 69–78, Calgary, AB,
Canada, 2009. ACM.
[AHS+ 03] J. Appavoo, K. Hui, C. A. N. Soules, R. W. Wisniewski, D. M. Da Silva,
O. Krieger, M. A. Auslander, D. J. Edelsohn, B. Gamsa, G. R. Ganger,
P. McKenney, M. Ostrowski, B. Rosenburg, M. Stumm, and J. Xenidis.
Enabling autonomic behavior in systems software with hot swapping.
IBM Systems Journal, 42(1):60–76, January 2003.
[AKK+ 14] Dan Alistarh, Justin Kopinsky, Petr Kuznetsov, Srivatsan Ravi, and Nir
Shavit. Inherent limitations of hybrid transactional memory. CoRR,
abs/1405.5689, 2014.
[AKNT13] Jade Alglave, Daniel Kroening, Vincent Nimal, and Michael Tautschnig.
Software verification for weak memory via program transformation.
In Proceedings of the 22nd European conference on Programming
Languages and Systems, ESOP’13, pages 512–532, Rome, Italy, 2013.
Springer-Verlag.
[AKT13] Jade Alglave, Daniel Kroening, and Michael Tautschnig. Partial orders for
efficient Bounded Model Checking of concurrent software. In Computer
Aided Verification (CAV), volume 8044 of LNCS, pages 141–157. Springer,
2013.
[Ale79] Christopher Alexander. The Timeless Way of Building. Oxford University
Press, New York, 1979.
[Alg13] Jade Alglave. Weakness is a virtue. In (EC)2 2013: 6th International
Workshop on Exploiting Concurrency Efficiently and Correctly, page 3,
2013.
[AM15] Maya Arbel and Adam Morrison. Predicate RCU: An RCU for scalable
concurrent updates. SIGPLAN Not., 50(8):21–30, January 2015.
v2023.06.11a
BIBLIOGRAPHY 895
[Amd67] Gene Amdahl. Validity of the single processor approach to achieving

large-scale computing capabilities. In AFIPS Conference Proceedings,
AFIPS ’67 (Spring), pages 483–485, Atlantic City, New Jersey, 1967.
Association for Computing Machinery.
[AMD20] AMD. Professional compute products - GPUOpen, March 2020. https:
//gpuopen.com/professional-compute/.
[AMM+ 17a] Jade Alglave, Luc Maranget, Paul E. McKenney, Andrea Parri, and Alan
Stern. A formal kernel memory-ordering model (part 1), April 2017.
https://lwn.net/Articles/718628/.
[AMM+ 17b] Jade Alglave, Luc Maranget, Paul E. McKenney, Andrea Parri, and Alan
Stern. A formal kernel memory-ordering model (part 2), April 2017.
[AMM+ 18] Jade Alglave, Luc Maranget, Paul E. McKenney, Andrea Parri, and Alan
Stern. Frightening small children and disconcerting grown-ups: Concur-
rency in the Linux kernel. In Proceedings of the Twenty-Third Interna-
tional Conference on Architectural Support for Programming Languages
and Operating Systems, ASPLOS ’18, pages 405–418, Williamsburg, VA,
USA, 2018. ACM.
[AMP+ 11] Jade Alglave, Luc Maranget, Pankaj Pawan, Susmit Sarkar, Peter Sewell,
Derek Williams, and Francesco Zappa Nardelli. PPCMEM/ARMMEM:
A tool for exploring the POWER and ARM memory models, June
2011. https://www.cl.cam.ac.uk/~pes20/ppc-supplemental/
pldi105-sarkar.pdf.
[AMT14] Jade Alglave, Luc Maranget, and Michael Tautschnig. Herding cats:
Modelling, simulation, testing, and data-mining for weak memory. In
Proceedings of the 35th ACM SIGPLAN Conference on Programming
Language Design and Implementation, PLDI ’14, pages 40–40, Edinburgh,
United Kingdom, 2014. ACM.
[And90] T. E. Anderson. The performance of spin lock alternatives for shared-
memory multiprocessors. IEEE Transactions on Parallel and Distributed
Systems, 1(1):6–16, January 1990.
[And91] Gregory R. Andrews. Concurrent Programming, Principles, and Prac-
tices. Benjamin Cummins, 1991.
[And19] Jim Anderson. Software transactional memory for real-time systems,
August 2019. https://www.cs.unc.edu/~anderson/projects/
rtstm.html.
[ARM10] ARM Limited. ARM Architecture Reference Manual: ARMv7-A and
ARMv7-R Edition, 2010.
[ARM17] ARM Limited. ARM Architecture Reference Manual (ARMv8, for ARMv8-
A architecture profile), 2017.
[Ash15] Mike Ash. Concurrent memory deallocation in the objective-c runtime,
May 2015. mikeash.com: just this guy, you know?
v2023.06.11a
896 BIBLIOGRAPHY
[ATC+ 11] Ege Akpinar, Sasa Tomic, Adrian Cristal, Osman Unsal, and Mateo
Valero. A comprehensive study of conflict resolution policies in hardware
transactional memory. In TRANSACT 2011, New Orleans, LA, USA,
June 2011. ACM SIGPLAN.
[ATS09] Ali-Reza Adl-Tabatabai and Tatiana Shpeisman. Draft specification of
transactional language constructs for C++, August 2009. URL: https://
software.intel.com/sites/default/files/ee/47/21569 (may
need to append .pdf to view after download).
[Att10] Hagit Attiya. The inherent complexity of transactional memory and
what to do about it. In Proceedings of the 29th ACM SIGACT-SIGOPS
Symposium on Principles of Distributed Computing, PODC ’10, pages
1–5, Zurich, Switzerland, 2010. ACM.
[BA01] Jeff Bonwick and Jonathan Adams. Magazines and vmem: Extending the
slab allocator to many CPUs and arbitrary resources. In USENIX Annual
Technical Conference, General Track 2001, pages 15–33, 2001.
[Bah11a] Samy Al Bahra. ck_epoch: Support per-object destructors, Oc-
tober 2011. https://github.com/concurrencykit/ck/commit/
10ffb2e6f1737a30e2dcf3862d105ad45fcd60a4.
[Bah11b] Samy Al Bahra. ck_hp.c, February 2011. Hazard pointers: https:
//github.com/concurrencykit/ck/blob/master/src/ck_hp.c.
[Bah11c] Samy Al Bahra. ck_sequence.h, February 2011. Sequence lock-
ing: https://github.com/concurrencykit/ck/blob/master/
include/ck_sequence.h.
[Bas18] JF Bastien. P1152R0: Deprecating volatile, October 2018.
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/
2018/p1152r0.html.
[BBC+ 10] Al Bessey, Ken Block, Ben Chelf, Andy Chou, Bryan Fulton, Seth Hallem,
Charles Henri-Gros, Asya Kamsky, Scott McPeak, and Dawson Engler.
A few billion lines of code later: Using static analysis to find bugs in the
real world. Commun. ACM, 53(2):66–75, February 2010.
[BCR03] David F. Bacon, Perry Cheng, and V. T. Rajan. A real-time garbage
collector with low overhead and consistent utilization. SIGPLAN Not.,
38(1):285–298, 2003.
[BD13] Paolo Bonzini and Mike Day. RCU implementation for Qemu, Au-
gust 2013. https://lists.gnu.org/archive/html/qemu-devel/
2013-08/msg02055.html.
[BD14] Hans-J. Boehm and Brian Demsky. Outlawing ghosts: Avoiding out-
of-thin-air results. In Proceedings of the Workshop on Memory Systems
Performance and Correctness, MSPC ’14, pages 7:1–7:6, Edinburgh,
United Kingdom, 2014. ACM.
[Bec11] Pete Becker. Working draft, standard for programming language C++, Feb-
ruary 2011. http://www.open-std.org/jtc1/sc22/wg21/docs/
papers/2011/n3242.pdf.
v2023.06.11a
BIBLIOGRAPHY 897
[BG87] D. Bertsekas and R. Gallager. Data Networks. Prentice-Hall, Inc., 1987.
[BGHZ16] Oana Balmau, Rachid Guerraoui, Maurice Herlihy, and Igor Zablotchi.
Fast and robust memory reclamation for concurrent data structures. In
Proceedings of the 28th ACM Symposium on Parallelism in Algorithms
and Architectures, SPAA ’16, pages 349–359, Pacific Grove, California,
USA, 2016. ACM.
[BGOS18] Sam Blackshear, Nikos Gorogiannis, Peter W. O’Hearn, and Ilya Sergey.
Racerd: Compositional static race detection. Proc. ACM Program. Lang.,
2(OOPSLA), October 2018.
[BGV17] Hans-J. Boehm, Olivier Giroux, and Viktor Vafeiades. P0668r1: Revising
the C++ memory model, July 2017. http://www.open-std.org/
jtc1/sc22/wg21/docs/papers/2017/p0668r1.html.
[Bha14] Srivatsa S. Bhat. percpu_rwlock: Implement the core design of per-CPU

reader-writer locks, February 2014. https://patchwork.kernel.
org/patch/2157401/.
[BHG87] Philip A. Bernstein, Vassos Hadzilacos, and Nathan Goodman. Con-

currency Control and Recovery in Database Systems. Addison Wesley
Publishing Company, 1987.
[BHS07] Frank Buschmann, Kevlin Henney, and Douglas C. Schmidt. Pattern-

Oriented Software Architecture Volume 4: A Pattern Language for
Distributed Computing. Wiley, Chichester, West Sussex, England, 2007.
[Bir89] Andrew D. Birrell. An Introduction to Programming with Threads. Digital

Systems Research Center, January 1989.
[BJ12] Rex Black and Capers Jones. Economics of software quality: An interview
with Capers Jones, part 1 of 2 (podcast transcript), January 2012. https:
//www.informit.com/articles/article.aspx?p=1824791.
[BK85] Bob Beck and Bob Kasten. VLSI assist in building a multiprocessor
UNIX system. In USENIX Conference Proceedings, pages 255–275,
Portland, OR, June 1985. USENIX Association.
[BLM05] C. Blundell, E. C. Lewis, and M. Martin. Deconstructing transac-

tional semantics: The subtleties of atomicity. In Annual Workshop
on Duplicating, Deconstructing, and Debunking (WDDD), June 2005.
Available: http://acg.cis.upenn.edu/papers/wddd05_atomic_
semantics.pdf [Viewed February 28, 2021].
[BLM06] C. Blundell, E. C. Lewis, and M. Martin. Subtleties of transac-

tional memory and atomicity semantics. Computer Architecture Let-
ters, 5(2), 2006. Available: http://acg.cis.upenn.edu/papers/
cal06_atomic_semantics.pdf [Viewed February 28, 2021].
[BM18] JF Bastien and Paul E. McKenney. P0750r1: Consume, Febru-

ary 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/
papers/2018/p0750r1.html.
v2023.06.11a
898 BIBLIOGRAPHY
[BMMM05] Luke Browning, Thomas Mathews, Paul E. McKenney, and James Moody.
Apparatus, method, and computer program product for converting simple
locks in a multiprocessor system. US Patent 6,842,809, Assigned to
International Business Machines Corporation, Washington, DC, January
2005.
[BMN+ 15] Mark Batty, Kayvan Memarian, Kyndylan Nienhuis, Jean Pichon-
Pharabod, and Peter Sewell. The problem of programming language
concurrency semantics. In Jan Vitek, editor, Programming Languages
and Systems, volume 9032 of Lecture Notes in Computer Science, pages
283–307. Springer Berlin Heidelberg, 2015.
[BMP08] R. F. Berry, P. E. McKenney, and F. N. Parr. Responsive systems: An
introduction. IBM Systems Journal, 47(2):197–206, April 2008.
[Boe05] Hans-J. Boehm. Threads cannot be implemented as a library. SIGPLAN
Not., 40(6):261–268, June 2005.
[Boe09] Hans-J. Boehm. Transactional memory should be an implementation
technique, not a programming interface. In HOTPAR 2009, page 6, Berke-
ley, CA, USA, March 2009. Available: https://www.usenix.org/
event/hotpar09/tech/full_papers/boehm/boehm.pdf [Viewed
May 24, 2009].
[Boe20] Hans Boehm. “Undefined behavior” and the concurrency memory model,
August 2020. http://www.open-std.org/jtc1/sc22/wg21/docs/
papers/2020/p2215r0.pdf.
[Boh01] Kristoffer Bohmann. Response time still matters, July
2001. URL: http://www.bohmann.dk/articles/response_time_
still_matters.html [broken, November 2016].
[Bon13] Paolo Bonzini. seqlock: introduce read-write seqlock, Septem-
ber 2013. https://git.qemu.org/?p=qemu.git;a=commit;h=
ea753d81e8b085d679f13e4a6023e003e9854d51.
[Bon15] Paolo Bonzini. rcu: add rcu library, February 2015.
https://git.qemu.org/?p=qemu.git;a=commit;h=
7911747bd46123ef8d8eef2ee49422bb8a4b274f.
[Bon21a] Paolo Bonzini. An introduction to lockless algorithms, February 2021.
Available: https://lwn.net/Articles/844224/ [Viewed February
19, 2021].
[Bon21b] Paolo Bonzini. Lockless patterns: an introduction to compare-and-
swap, March 2021. Available: https://lwn.net/Articles/847973/
[Viewed March 13, 2021].
[Bon21c] Paolo Bonzini. Lockless patterns: full memory barriers, March 2021.
Available: https://lwn.net/Articles/847481/ [Viewed March 8,
2021].
[Bon21d] Paolo Bonzini. Lockless patterns: more read-modify-write opera-
tions, March 2021. Available: https://lwn.net/Articles/849237/
v2023.06.11a
BIBLIOGRAPHY 899
[Bon21e] Paolo Bonzini. Lockless patterns: relaxed access and partial memory
barriers, February 2021. Available: https://lwn.net/Articles/
846700/ [Viewed February 27, 2021].
[Bon21f] Paolo Bonzini. Lockless patterns: some final topics, March 2021.
Available: https://lwn.net/Articles/850202/ [Viewed March 19,
2021].
[Bor06] Richard Bornat. Dividing the sheep from the goats, Jan-
uary 2006. Seminar at School of Computing, Univ. of
Kent. Abstract is available at https://www.cs.kent.ac.uk/
seminar_archive/2005_06/abs_2006_01_24.html. Retracted in
July 2014: http://www.eis.mdx.ac.uk/staffpages/r_bornat/
papers/camel_hump_retraction.pdf.
[Bos10] Keith Bostic. Switch lockless programming style from
epoch to hazard references, January 2010. https:
//github.com/wiredtiger/wiredtiger/commit/
dddc21014fc494a956778360a14d96c762495e09.
[Bos23] Mara Bos. Rust Atomics and Locks. O’Reilly Media, Inc., Sebastopol,
CA, USA, 2023.
[BPP+ 16] Adam Belay, George Prekas, Mia Primorac, Ana Klimovic, Samuel
Grossman, Christos Kozyrakis, and Edouard Bugnion. The IX operating
system: Combining low latency, high throughput, and efficiency in
a protected dataplane. ACM Trans. Comput. Syst., 34(4):11:1–11:39,
December 2016.
[Bra07] Reg Braithwaite. Don’t overthink fizzbuzz, January 2007.
http://weblog.raganwald.com/2007/01/dont-overthink-
fizzbuzz.html.
[Bra11] Björn Brandenburg. Scheduling and Locking in Multiprocessor Real-
Time Operating Systems. PhD thesis, The University of North Carolina
at Chapel Hill, 2011. URL: https://www.cs.unc.edu/~anderson/
diss/bbbdiss.pdf.
[Bro15a] Neil Brown. Pathname lookup in Linux, June 2015. https://lwn.
net/Articles/649115/.
[Bro15b] Neil Brown. RCU-walk: faster pathname lookup in Linux, July 2015.
[Bro15c] Neil Brown. A walk among the symlinks, July 2015. https://lwn.
[BS75] Paul J. Brown and Ronald M. Smith. Shared data controlled by a plurality
of users, May 1975. US Patent 3,886,525, filed June 29, 1973.
[BS14] Mark Batty and Peter Sewell. The thin-air problem, February 2014.
https://www.cl.cam.ac.uk/~pes20/cpp/notes42.html.
[But97] David Butenhof. Programming with POSIX Threads. Addison-Wesley,
Boston, MA, USA, 1997.
v2023.06.11a
900 BIBLIOGRAPHY
[BW14] Silas Boyd-Wickizer. Optimizing Communications Bottlenecks in Multi-

processor Operating Systems Kernels. PhD thesis, Massachusetts Institute
of Technology, 2014. https://pdos.csail.mit.edu/papers/sbw-
phd-thesis.pdf.
[BWCM+ 10] Silas Boyd-Wickizer, Austin T. Clements, Yandong Mao, Aleksey Pesterev,
M. Frans Kaashoek, Robert Morris, and Nickolai Zeldovich. An analysis
of Linux scalability to many cores. In 9th USENIX Symposium on
Operating System Design and Implementation, pages 1–16, Vancouver,
BC, Canada, October 2010. USENIX.
[CAK+ 96] Crispin Cowan, Tito Autrey, Charles Krasic, Calton Pu, and Jonathan
Walpole. Fast concurrent dynamic linking for an adaptive operating
system. In International Conference on Configurable Distributed Systems
(ICCDS’96), pages 108–115, Annapolis, MD, May 1996.
[CBF13] UPC Consortium, Dan Bonachea, and Gary Funck. UPC language and
library specifications, version 1.3. Technical report, UPC Consortium,
November 2013.
[CBM+ 08] Calin Cascaval, Colin Blundell, Maged Michael, Harold W. Cain, Peng
Wu, Stefanie Chiras, and Siddhartha Chatterjee. Software transactional
memory: Why is it only a research toy? ACM Queue, September 2008.
[Chi22] A.A. Chien. Computer Architecture for Scientists. Cambridge University

Press, 2022.
[CHP71] P. J. Courtois, F. Heymans, and D. L. Parnas. Concurrent control with

“readers” and “writers”. Communications of the ACM, 14(10):667–668,
October 1971.
[CKL04] Edmund Clarke, Daniel Kroening, and Flavio Lerda. A tool for checking
ANSI-C programs. In Kurt Jensen and Andreas Podelski, editors, Tools
and Algorithms for the Construction and Analysis of Systems (TACAS
2004), volume 2988 of Lecture Notes in Computer Science, pages 168–176.
Springer, 2004.
[CKZ12] Austin Clements, Frans Kaashoek, and Nickolai Zeldovich. Scalable

address spaces using RCU balanced trees. In Architectural Support for
Programming Languages and Operating Systems (ASPLOS 2012), pages
199–210, London, UK, March 2012. ACM.
[CKZ+ 13] Austin T. Clements, M. Frans Kaashoek, Nickolai Zeldovich, Robert T.

Morris, and Eddie Kohler. The scalable commutativity rule: Designing
scalable software for multicore processors. In Proceedings of the Twenty-
Fourth ACM Symposium on Operating Systems Principles, SOSP ’13,
pages 1–17, Farminton, Pennsylvania, 2013. ACM.
[Cli09] Cliff Click. And now some hardware transactional memory comments...,
February 2009. URL: http://www.cliffc.org/blog/2009/02/25/
and-now-some-hardware-transactional-memory-comments/.
v2023.06.11a
BIBLIOGRAPHY 901
[CLRS01] T.H. Cormen, C.E. Leiserson, R.L. Rivest, and C. Stein. Introduction to
Algorithms, Second Edition. MIT electrical engineering and computer
science series. MIT Press, 2001.
[CnRR18] Armando Castañeda, Sergio Rajsbaum, and Michel Raynal. Unifying
concurrent objects and distributed tasks: Interval-linearizability. J. ACM,
65(6), November 2018.
[Com01] Compaq Computer Corporation. Shared memory, threads, interpro-
cess communication, August 2001. Zipped archive: wiz_2637.
txt in https://www.digiater.nl/openvms/freeware/v70/ask_
the_wizard/wizard.zip.
[Coo18] Byron Cook. Formal reasoning about the security of amazon web services.
In Hana Chockler and Georg Weissenbacher, editors, Computer Aided
Verification, pages 38–47, Cham, 2018. Springer International Publishing.
[Cor02] Compaq Computer Corporation. Alpha Architecture Reference Manual.
Digital Press, fourth edition, 2002.
[Cor03] Jonathan Corbet. Driver porting: mutual exclusion with seqlocks, Febru-
ary 2003. https://lwn.net/Articles/22818/.
[Cor04a] Jonathan Corbet. Approaches to realtime Linux, October 2004. URL:
[Cor04b] Jonathan Corbet. Finding kernel problems automatically, June 2004.
[Cor04c] Jonathan Corbet. Realtime preemption, part 2, October 2004. URL:
[Cor06a] Jonathan Corbet. The kernel lock validator, May 2006. Available:
https://lwn.net/Articles/185666/ [Viewed: March 26, 2010].
[Cor06b] Jonathan Corbet. Priority inheritance in the kernel, April 2006. Available:
https://lwn.net/Articles/178253/ [Viewed June 29, 2009].
[Cor10a] Jonathan Corbet. Dcache scalability and RCU-walk, December 2010.
Available: https://lwn.net/Articles/419811/ [Viewed May 29,
2017].
[Cor10b] Jonathan Corbet. sys_membarrier(), January 2010. https://lwn.net/
Articles/369567/.
[Cor11] Jonathan Corbet. How to ruin linus’s vacation, July 2011. Available:
https://lwn.net/Articles/452117/ [Viewed May 29, 2017].
[Cor12] Jonathan Corbet. ACCESS_ONCE(), August 2012. https://lwn.net/
Articles/508991/.
[Cor13] Jonathan Corbet. (Nearly) full tickless operation in 3.10, May 2013.
[Cor14a] Jonathan Corbet. ACCESS_ONCE() and compiler bugs, December 2014.
v2023.06.11a
902 BIBLIOGRAPHY
[Cor14b] Jonathan Corbet. MCS locks and qspinlocks, March 2014. https:
//lwn.net/Articles/590243/.
[Cor14c] Jonathan Corbet. Relativistic hash tables, part 1: Algorithms, September
2014. https://lwn.net/Articles/612021/.
[Cor14d] Jonathan Corbet. Relativistic hash tables, part 2: Implementation,
September 2014. https://lwn.net/Articles/612100/.
[Cor16a] Jonathan Corbet. Finding race conditions with KCSAN, June 2016.
[Cor16b] Jonathan Corbet. Time to move to C11 atomics?, June 2016. https:
[Cor18] Jonathan Corbet. membarrier(2), October 2018. https://man7.org/
linux/man-pages/man2/membarrier.2.html.
[Cra93] Travis Craig. Building FIFO and priority-queuing spin locks from atomic
swap. Technical Report 93-02-02, University of Washington, Seattle,
Washington, February 1993.
[CRKH05] Jonathan Corbet, Alessandro Rubini, and Greg Kroah-Hartman. Linux
Device Drivers. O’Reilly Media, Inc., third edition, 2005. URL: https:
//lwn.net/Kernel/LDD3/.
[CSG99] David E. Culler, Jaswinder Pal Singh, and Anoop Gupta. Parallel Com-
puter Architecture: a Hardware/Software Approach. Morgan Kaufman,
1999.
[cut17] crates.io user ticki. conc v0.5.0: Hazard-pointer-based concurrent memory
reclamation, August 2017. https://crates.io/crates/conc.
[Dat82] C. J. Date. An Introduction to Database Systems, volume 1. Addison-
Wesley Publishing Company, 1982.
[DBA09] Saeed Dehnadi, Richard Bornat, and Ray Adams. Meta-analysis of the
effect of consistency on success in early learning of programming. In
PPIG 2009, pages 1–13, University of Limerick, Ireland, June 2009.
Psychology of Programming Interest Group.
[DCW+ 11] Luke Dalessandro, Francois Carouge, Sean White, Yossi Lev, Mark
Moir, Michael L. Scott, and Michael F. Spear. Hybrid NOrec: A case
study in the effectiveness of best effort hardware transactional memory.
In Proceedings of the 16th International Conference on Architectural
Support for Programming Languages and Operating Systems (ASPLOS),
ASPLOS ’11, page 39–52, Newport Beach, CA, USA, 2011. ACM.
[Dea18] Will Deacon. [PATCH 00/10] kernel/locking: qspinlock improvements,
April 2018. https://lkml.kernel.org/r/1522947547-24081-1-
git-send-email-will.deacon@arm.com.
[Dea19] Will Deacon. Re: [PATCH 1/1] Fix: trace sched switch start/stop
racy updates, August 2019. https://lore.kernel.org/lkml/
20190821103200.kpufwtviqhpbuv2n@willie-the-truck/.
v2023.06.11a
BIBLIOGRAPHY 903
[Den15] Peter Denning. Perspectives on OS foundations. In SOSP History Day

2015, SOSP ’15, pages 3:1–3:46, Monterey, California, 2015. ACM.
[Dep06] Department of Computing and Information Systems, University of Mel-

bourne. CSIRAC, 2006. https://cis.unimelb.edu.au/about/
csirac/.
[Des09a] Mathieu Desnoyers. Low-Impact Operating System Trac-

ing. PhD thesis, Ecole Polytechnique de Montréal, December
2009. Available: https://lttng.org/files/thesis/desnoyers-
dissertation-2009-12-v27.pdf [Viewed February 27, 2021].
[Des09b] Mathieu Desnoyers. [RFC git tree] userspace RCU (urcu) for Linux,
February 2009. https://liburcu.org.
[DFGG11] Aleksandar Dragovejic, Pascal Felber, Vincent Gramoli, and Rachid

Guerraoui. Why STM can be more than a research toy. Communications
of the ACM, pages 70–77, April 2011.
[DFLO19] Dino Distefano, Manuel Fähndrich, Francesco Logozzo, and Peter W.

O’Hearn. Scaling static analyses at facebook. Commun. ACM,
62(8):62–70, July 2019.
[DHJ+ 07] Giuseppe DeCandia, Deniz Hastorun, Madan Jampani, Gunavardhan

Kakulapati, Avinash Lakshman, Alex Pilchin, Swaminathan Sivasubra-
manian, Peter Vosshall, and Werner Vogels. Dynamo: Amazon’s highly
available key-value store. SIGOPS Oper. Syst. Rev., 41(6):205–220,
October 2007.
[DHK12] Vĳay D’Silva, Leopold Haller, and Daniel Kroening. Satisfiability solvers
are static analyzers. In Static Analysis Symposium (SAS), volume 7460 of
LNCS, pages 317–333. Springer, 2012.
[DHL+ 08] Dave Dice, Maurice Herlihy, Doug Lea, Yossi Lev, Victor Luchangco,
Wayne Mesard, Mark Moir, Kevin Moore, and Dan Nussbaum. Appli-
cations of the adaptive transactional memory test platform. In 3rd ACM
SIGPLAN Workshop on Transactional Computing, pages 1–10, Salt Lake
City, UT, USA, February 2008.
[Dĳ65] E. W. Dĳkstra. Solution of a problem in concurrent programming control.

Communications of the ACM, 8(9):569, Sept 1965.
[Dĳ68] Edsger W. Dĳkstra. Letters to the editor: Go to statement considered

harmful. Commun. ACM, 11(3):147–148, March 1968.
[Dĳ71] Edsger W. Dĳkstra. Hierarchical ordering of sequential processes. Acta In-

formatica, 1(2):115–138, 1971. Available: https://www.cs.utexas.
edu/users/EWD/ewd03xx/EWD310.PDF [Viewed January 13, 2008].
[DKS89] Alan Demers, Srinivasan Keshav, and Scott Shenker. Analysis and
simulation of a fair queuing algorithm. SIGCOMM ’89, pages 1–12,
1989.
v2023.06.11a
904 BIBLIOGRAPHY
[DLM+ 10] Dave Dice, Yossi Lev, Virendra J. Marathe, Mark Moir, Dan Nussbaum,
and Marek Oleszewski. Simplifying concurrent algorithms by exploiting
hardware transactional memory. In Proceedings of the 22nd ACM
symposium on Parallelism in algorithms and architectures, SPAA ’10,
pages 325–334, Thira, Santorini, Greece, 2010. ACM.
[DLMN09] Dave Dice, Yossi Lev, Mark Moir, and Dan Nussbaum. Early experi-
ence with a commercial hardware transactional memory implementation.
In Fourteenth International Conference on Architectural Support for
Programming Languages and Operating Systems (ASPLOS ’09), pages
157–168, Washington, DC, USA, March 2009.
[DMD13] Mathieu Desnoyers, Paul E. McKenney, and Michel R. Dagenais. Multi-

core systems modeling for formal verification of parallel algorithms.
SIGOPS Oper. Syst. Rev., 47(2):51–65, July 2013.
[DMLP79] Richard A. De Millo, Richard J. Lipton, and Alan J. Perlis. Social processes
and proofs of theorems and programs. Commun. ACM, 22(5):271–280,
May 1979.
[DMS+ 12] Mathieu Desnoyers, Paul E. McKenney, Alan Stern, Michel R. Dagenais,
and Jonathan Walpole. User-level implementations of read-copy update.
IEEE Transactions on Parallel and Distributed Systems, 23:375–382,
2012.
[dO18a] Daniel Bristot de Oliveira. Deadline scheduler part 2 – details and usage,
January 2018. URL: https://lwn.net/Articles/743946/.
[dO18b] Daniel Bristot de Oliveira. Deadline scheduling part 1 – overview and

theory, January 2018. URL: https://lwn.net/Articles/743740/.
[dOCdO19] Daniel Bristot de Oliveira, Tommaso Cucinotta, and Rômulo Silva

de Oliveira. Modeling the behavior of threads in the PREEMPT_RT
Linux kernel using automata. SIGBED Rev., 16(3):63–68, November
2019.
[Don21] Jason Donenfeld. Introduce WireGuardNT, August 2021. Git

commit: https://git.zx2c4.com/wireguard-nt/commit/?id=
d64c53776d7f72751d7bd580ead9846139c8f12f.
[Dov90] Ken F. Dove. A high capacity TCP/IP in parallel STREAMS. In UKUUG

Conference Proceedings, London, June 1990.
[Dow20] Travis Downs. Gathering intel on Intel AVX-512 transitions, Janu-

ary 2020. https://travisdowns.github.io/blog/2020/01/17/
avxfreq1.html.
[Dre11] Ulrich Drepper. Futexes are tricky. Technical Report FAT2011, Red Hat,
Inc., Raleigh, NC, USA, November 2011.
[DSS06] Dave Dice, Ori Shalev, and Nir Shavit. Transactional locking II. In Proc.
International Symposium on Distributed Computing. Springer Verlag,
2006.
v2023.06.11a
BIBLIOGRAPHY 905
[Duf10a] Joe Duffy. A (brief) retrospective on transactional memory, Jan-

uary 2010. http://joeduffyblog.com/2010/01/03/a-brief-
retrospective-on-transactional-memory/.
[Duf10b] Joe Duffy. More thoughts on transactional memory, May
2010. http://joeduffyblog.com/2010/05/16/more-thoughts-
on-transactional-memory/.
[Dug10] Abhinav Duggal. Stopping data races using redflag. Master’s thesis,
Stony Brook University, 2010.
[Eas71] William B. Easton. Process synchronization without long-term interlock.
In Proceedings of the Third ACM Symposium on Operating Systems
Principles, SOSP ’71, pages 95–100, Palo Alto, California, USA, 1971.
[Edg13] Jake Edge. The future of realtime Linux, November 2013. URL:
[Edg14] Jake Edge. The future of the realtime patch set, October 2014. URL:
[Edg22] Jake Edge. Introducing io_uring_spawn, September 2022. https:
[EGCD03] T. A. El-Ghazawi, W. W. Carlson, and J. M. Draper. UPC language
specifications v1.1, May 2003. URL: http://upc.gwu.edu [broken,
February 27, 2021].
[EGMdB11] Stephane Eranian, Eric Gouriou, Tipp Moseley, and Willem de Bruĳn.
Linux kernel profiling with perf, June 2011. https://perf.wiki.
kernel.org/index.php/Tutorial.
[Ell80] Carla Schlatter Ellis. Concurrent search and insertion in AVL trees. IEEE
Transactions on Computers, C-29(9):811–817, September 1980.
[ELLM07] Faith Ellen, Yossi Lev, Victor Luchangco, and Mark Moir. SNZI: scalable
NonZero indicators. In Proceedings of the twenty-sixth annual ACM
symposium on Principles of distributed computing, PODC ’07, pages
13–22, Portland, Oregon, USA, 2007. ACM.
[EMV+ 20a] Marco Elver, Paul E. McKenney, Dmitry Vyukov, Andrey Konovalov,
Alexander Potapenko, Kostya Serebryany, Alan Stern, Andrea Parri, Akira
Yokosawa, Peter Zĳlstra, Will Deacon, Daniel Lustig, Boqun Feng, Joel
Fernandes, Jade Alglave, and Luc Maranget. Concurrency bugs should
fear the big bad data-race detector (part 1), April 2020. Linux Weekly
News.
[EMV+ 20b] Marco Elver, Paul E. McKenney, Dmitry Vyukov, Andrey Konovalov,
Alexander Potapenko, Kostya Serebryany, Alan Stern, Andrea Parri, Akira
Yokosawa, Peter Zĳlstra, Will Deacon, Daniel Lustig, Boqun Feng, Joel
Fernandes, Jade Alglave, and Luc Maranget. Concurrency bugs should
fear the big bad data-race detector (part 2), April 2020. Linux Weekly
News.
v2023.06.11a
906 BIBLIOGRAPHY
[Eng68] Douglas Engelbart. The demo, December 1968. URL: http:

//thedemo.org/.
[ENS05] Ryan Eccles, Blair Nonneck, and Deborah A. Stacey. Exploring parallel
programming knowledge in the novice. In HPCS ’05: Proceedings of the
19th International Symposium on High Performance Computing Systems
and Applications, pages 97–102, Guelph, Ontario, Canada, 2005. IEEE
Computer Society.
[Eri08] Christer Ericson. Aiding pathfinding with cellular automata, June 2008.
http://realtimecollisiondetection.net/blog/?p=57.
[ES90] Margaret A. Ellis and Bjarne Stroustrup. The Annotated C++ Reference
Manual. Addison Wesley, 1990.
[ES05] Ryan Eccles and Deborah A. Stacey. Understanding the parallel program-
mer. In HPCS ’05: Proceedings of the 19th International Symposium on
High Performance Computing Systems and Applications, pages 156–160,
Guelph, Ontario, Canada, 2005. IEEE Computer Society.
[ETH11] ETH Zurich. Parallel solver for a perfect maze, March

2011. URL: http://nativesystems.inf.ethz.ch/pub/
Main/WebHomeLecturesParallelProgrammingExercises/
pp2011hw04.pdf [broken, November 2016].
[Eva11] Jason Evans. Scalable memory allocation using jemalloc, January

2011. https://engineering.fb.com/2011/01/03/core-data/
scalable-memory-allocation-using-jemalloc/.
[Fel50] W. Feller. An Introduction to Probability Theory and its Applications.

John Wiley, 1950.
[Fen73] J. Fennel. Instruction selection in a two-program counter instruction

unit. Technical Report US Patent 3,728,692, Assigned to International
Business Machines Corp, Washington, DC, April 1973.
[Fen15] Boqun Feng. powerpc: Make value-returning atomics fully ordered,

November 2015. Git commit: https://git.kernel.org/linus/
49e9cf3f0c04.
[FH07] Keir Fraser and Tim Harris. Concurrent programming without locks.
ACM Trans. Comput. Syst., 25(2):1–61, 2007.
[FIMR16] Pascal Felber, Shady Issa, Alexander Matveev, and Paolo Romano. Hard-
ware read-write lock elision. In Proceedings of the Eleventh European
Conference on Computer Systems, EuroSys ’16, London, United Kingdom,
2016. Association for Computing Machinery.
[Fos10] Ron Fosner. Scalable multithreaded programming with tasks. MSDN Mag-
azine, 2010(11):60–69, November 2010. http://msdn.microsoft.
com/en-us/magazine/gg309176.aspx.
[FPB79] Jr. Frederick P. Brooks. The Mythical Man-Month. Addison-Wesley,

1979.
v2023.06.11a
BIBLIOGRAPHY 907
[Fra03] Keir Anthony Fraser. Practical Lock-Freedom. PhD thesis, King’s

College, University of Cambridge, 2003.
[Fra04] Keir Fraser. Practical lock-freedom. Technical Report UCAM-CL-TR-579,

University of Cambridge, Computer Laboratory, February 2004.
[FRK02] Hubertus Francke, Rusty Russell, and Matthew Kirkwood. Fuss, futexes
and furwocks: Fast userlevel locking in linux. In Ottawa Linux Sympo-
sium, pages 479–495, June 2002. Available: https://www.kernel.
org/doc/ols/2002/ols2002-pages-479-495.pdf [Viewed May
22, 2011].
[FSP+ 17] Shaked Flur, Susmit Sarkar, Christopher Pulte, Kyndylan Nienhuis, Luc
Maranget, Kathryn E. Gray, Ali Sezgin, Mark Batty, and Peter Sewell.
Mixed-size concurrency: ARM, POWER, C/C++11, and SC. SIGPLAN
Not., 52(1):429–442, January 2017.
[GAJM15] Alex Groce, Iftekhar Ahmed, Carlos Jensen, and Paul E. McKenney. How
verified is my code? falsification-driven verification (t). In Proceedings
of the 2015 30th IEEE/ACM International Conference on Automated
Software Engineering (ASE), ASE ’15, pages 737–748, Washington, DC,
USA, 2015. IEEE Computer Society.
[Gar90] Arun Garg. Parallel STREAMS: a multi-processor implementation. In

USENIX Conference Proceedings, pages 163–176, Berkeley CA, Febru-
ary 1990. USENIX Association. Available: https://archive.org/
details/1990-proceedings-winter-dc/page/163/mode/2up.
[Gar07] Bryan Gardiner. IDF: Gordon Moore predicts end of Moore’s law
(again), September 2007. Available: https://www.wired.com/2007/
09/idf-gordon-mo-1/ [Viewed: February 27, 2021].
[GC96] Michael Greenwald and David R. Cheriton. The synergy between non-
blocking synchronization and operating system structure. In Proceedings
of the Second Symposium on Operating Systems Design and Implementa-
tion, pages 123–136, Seattle, WA, October 1996. USENIX Association.
[GDZE10] Olga Golovanevsky, Alon Dayan, Ayal Zaks, and David Edelsohn. Trace-
based data layout optimizations for multi-core processors. In Proceedings
of the 5th International Conference on High Performance Embedded
Architectures and Compilers, HiPEAC’10, pages 81–95, Pisa, Italy, 2010.
Springer-Verlag.
[GG14] Vincent Gramoli and Rachid Guerraoui. Democratizing transactional

programming. Commun. ACM, 57(1):86–93, January 2014.
[GGK18] Christina Giannoula, Georgios Goumas, and Nectarios Koziris. Combin-

ing HTM with RCU to speed up graph coloring on multicore platforms. In
Rio Yokota, Michèle Weiland, David Keyes, and Carsten Trinitis, editors,
High Performance Computing, pages 350–369, Cham, 2018. Springer
International Publishing.
v2023.06.11a
908 BIBLIOGRAPHY
[GGL+ 19] Rachid Guerraoui, Hugo Guiroux, Renaud Lachaize, Vivien Quéma, and
Vasileios Trigonakis. Lock–unlock: Is that all? a pragmatic analysis of
locking in software systems. ACM Trans. Comput. Syst., 36(1):1:1–1:149,
March 2019.
[Gha95] Kourosh Gharachorloo. Memory consistency models for shared-memory
multiprocessors. Technical Report CSL-TR-95-685, Computer Sys-
tems Laboratory, Departments of Electrical Engineering and Computer
Science, Stanford University, Stanford, CA, December 1995. Avail-
able: https://www.hpl.hp.com/techreports/Compaq-DEC/WRL-
95-9.pdf [Viewed: October 11, 2004].
[GHH+ 14] Alex Groce, Klaus Havelund, Gerard J. Holzmann, Rajeev Joshi, and
Ru-Gang Xu. Establishing flight software reliability: testing, model
checking, constraint-solving, monitoring and learning. Ann. Math. Artif.
Intell., 70(4):315–349, 2014.
[GHJV95] Erich Gamma, Richard Helm, Ralph Johnson, and John Vlissides. Design
Patterns: Elements of Reusable Object-Oriented Software. Addison-
Wesley, 1995.
[GKAS99] Ben Gamsa, Orran Krieger, Jonathan Appavoo, and Michael Stumm.
Tornado: Maximizing locality and concurrency in a shared memory
multiprocessor operating system. In Proceedings of the 3rd Symposium
on Operating System Design and Implementation, pages 87–100, New
Orleans, LA, February 1999.
[GKP13] Justin Gottschlich, Rob Knauerhase, and Gilles Pokam. But how do we
really debug transactional memory? In 5th USENIX Workshop on Hot
Topics in Parallelism (HotPar 2013), San Jose, CA, USA, June 2013.
[GKPS95] Ben Gamsa, Orran Krieger, E. Parsons, and Michael Stumm. Performance
issues for multiprocessor operating systems, November 1995. Techni-
cal Report CSRI-339, Available: ftp://ftp.cs.toronto.edu/pub/
reports/csri/339/339.ps.
[Gla18] Stjepan Glavina. Merge remaining subcrates, November
2018. https://github.com/crossbeam-rs/crossbeam/commit/
d9b1e3429450a64b490f68c08bd191417e68f00c.
[Gle10] Thomas Gleixner. Realtime linux: academia v. reality, July 2010. URL:
[Gle12] Thomas Gleixner. Linux -rt kvm guest demo, December 2012. Personal
communication.
[GMTW08] D. Guniguntala, P. E. McKenney, J. Triplett, and J. Walpole. The
read-copy-update mechanism for supporting real-time applications on
shared-memory multiprocessor systems with Linux. IBM Systems Journal,
47(2):221–236, May 2008.
[Gol18a] David Goldblatt. Add the Seq module, a simple seqlock implementa-
tion, April 2018. https://github.com/jemalloc/jemalloc/tree/
06a8c40b36403e902748d3f2a14e6dd43488ae89.
v2023.06.11a
BIBLIOGRAPHY 909
[Gol18b] David Goldblatt. P1202: Asymmetric fences, October 2018.

http://www.open-std.org/jtc1/sc22/wg21/docs/papers/
2018/p1202r0.pdf.
[Gol19] David Goldblatt. There might not be an elegant OOTA fix, Oc-
tober 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/
[GPB+ 07] Brian Goetz, Tim Peierls, Joshua Bloch, Joseph Bowbeer, David Holmes,
and Doug Lea. Java: Concurrency in Practice. Addison Wesley, Upper
Saddle River, NJ, USA, 2007.
[Gra91] Jim Gray. The Benchmark Handbook for Database and Transaction
Processing Systems. Morgan Kaufmann, 1991.
[Gra02] Jim Gray. Super-servers: Commodity computer clus-

ters pose a software challenge, April 2002. Available:
http://research.microsoft.com/en-us/um/people/gray/
papers/superservers(4t_computers).doc [Viewed: June 23,
2004].
[Gre19] Brendan Gregg. BPF Performance Tools: Linux System and Application
Observability. Addison-Wesley Professional, 1st edition, 2019.
[Gri00] Scott Griffen. Internet pioneers: Doug englebart, May 2000. Available:
https://www.ibiblio.org/pioneers/englebart.html [Viewed
November 28, 2008].
[Gro01] The Open Group. Single UNIX specification, July 2001. http://www.
opengroup.org/onlinepubs/007908799/index.html.
[Gro07] Dan Grossman. The transactional memory / garbage collection anal-

ogy. In OOPSLA ’07: Proceedings of the 22nd annual ACM
SIGPLAN conference on Object oriented programming systems and
applications, pages 695–706, Montreal, Quebec, Canada, October
2007. ACM. Available: https://homes.cs.washington.edu/
~djg/papers/analogy_oopsla07.pdf [Viewed February 27, 2021].
[GRY12] Alexey Gotsman, Noam Rinetzky, and Hongseok Yang. Verify-

ing highly concurrent algorithms with grace (extended version),
July 2012. https://software.imdea.org/~gotsman/papers/
recycling-esop13-ext.pdf.
[GRY13] Alexey Gotsman, Noam Rinetzky, and Hongseok Yang. Verifying

concurrent memory reclamation algorithms with grace. In ESOP’13:
European Symposium on Programming, pages 249–269, Rome, Italy,
2013. Springer-Verlag.
[GT90] Gary Graunke and Shreekant Thakkar. Synchronization algorithms for

shared-memory multiprocessors. IEEE Computer, 23(6):60–69, June
1990.
v2023.06.11a
910 BIBLIOGRAPHY
[Gui18] Hugo Guiroux. Understanding the performance of mutual exclusion

algorithms on modern multicore machines. PhD thesis, Université
Grenoble Alpes, 2018. https://hugoguiroux.github.io/assets/
these.pdf.
[Gwy15] David Gwynne. introduce srp, which according to the manpage i wrote
is short for “shared reference pointers”., July 2015. https://github.
com/openbsd/src/blob/HEAD/sys/kern/kern_srp.c.
[GYW+ 19] Jinyu Gu, Qianqian Yu, Xiayang Wang, Zhaoguo Wang, Binyu Zang,
Haibing Guan, and Haibo Chen. Pisces: A scalable and efficient persistent
transactional memory. In Proceedings of the 2019 USENIX Conference on
Usenix Annual Technical Conference, USENIX ATC ’19, pages 913–928,
Renton, WA, USA, 2019. USENIX Association.
[Har01] Timothy L. Harris. A pragmatic implementation of non-blocking linked-
lists. In Proc. 15th Intl. Conf. on Distributed Computing, pages 300–314.
Springer-Verlag, 2001.
[Har16] "No Bugs" Hare. Infographics: Operation costs in CPU clock cycles,
September 2016. http://ithare.com/infographics-operation-
costs-in-cpu-clock-cycles/.
[Hay20] Timothy Hayes. A shift to concurrency, October 2020. https:
//community.arm.com/developer/research/b/articles/
posts/arms-transactional-memory-extension-support-.
[HCS+ 05] Lorin Hochstein, Jeff Carver, Forrest Shull, Sima Asgari, and Victor
Basili. Parallel programmer productivity: A case study of novice parallel
programmers. In SC ’05: Proceedings of the 2005 ACM/IEEE conference
on Supercomputing, page 35, Seattle, WA, USA, 2005. IEEE Computer
Society.
[Hei27] W. Heisenberg. Über den anschaulichen Inhalt der quantentheoretischen
Kinematik und Mechanik. Zeitschrift für Physik, 43(3-4):172–198, 1927.
English translation in “Quantum theory and measurement” by Wheeler
and Zurek.
[Her90] Maurice P. Herlihy. A methodology for implementing highly concurrent
data structures. In Proceedings of the 2nd ACM SIGPLAN Symposium
on Principles and Practice of Parallel Programming, pages 197–206,
Seattle, WA, USA, March 1990.
[Her91] Maurice Herlihy. Wait-free synchronization. ACM TOPLAS, 13(1):124–
149, January 1991.
[Her93] Maurice Herlihy. A methodology for implementing highly concurrent
data objects. ACM Transactions on Programming Languages and Systems,
15(5):745–770, November 1993.
[Her05] Maurice Herlihy. The transactional manifesto: software engineering
and non-blocking synchronization. In PLDI ’05: Proceedings of the
2005 ACM SIGPLAN conference on Programming language design and
implementation, pages 280–280, Chicago, IL, USA, 2005. ACM Press.
v2023.06.11a
BIBLIOGRAPHY 911
[Her11] Benjamin Herrenschmidt. powerpc: Fix atomic_xxx_return barrier

semantics, November 2011. Git commit: https://git.kernel.org/
linus/b97021f85517.
[HHK+ 13] A. Haas, T.A. Henzinger, C.M. Kirsch, M. Lippautz, H. Payer, A. Sezgin,
and A. Sokolova. Distributed queues in shared memory—multicore
performance and scalability through quantitative relaxation. In Proc.
International Conference on Computing Frontiers, Ischia, Italy, 2013.
ACM.
[HKLP12] Andreas Haas, Christoph M. Kirsch, Michael Lippautz, and Hannes
Payer. How FIFO is your concurrent FIFO queue? In Proceedings of
the Workshop on Relaxing Synchronization for Multicore and Manycore
Scalability, Tucson, AZ USA, October 2012.
[HL86] Frederick S. Hillier and Gerald J. Lieberman. Introduction to Operations
Research. Holden-Day, 1986.
[HLM02] Maurice Herlihy, Victor Luchangco, and Mark Moir. The repeat offender
problem: A mechanism for supporting dynamic-sized, lock-free data
structures. In Proceedings of 16th International Symposium on Distributed
Computing, pages 339–353, Toulouse, France, October 2002.
[HLM03] Maurice Herlihy, Victor Luchangco, and Mark Moir. Obstruction-free
synchronization: Double-ended queues as an example. In Proceedings
of the 23rd IEEE International Conference on Distributed Computing
Systems (ICDCS), pages 73–82, Providence, RI, May 2003. The Institute
of Electrical and Electronics Engineers, Inc.
[HM93] Maurice Herlihy and J. Eliot B. Moss. Transactional memory: Archi-
tectural support for lock-free data structures. In ISCA ’93: Proceeding
of the 20th Annual International Symposium on Computer Architecture,
pages 289–300, San Diego, CA, USA, May 1993.
[HMB06] Thomas E. Hart, Paul E. McKenney, and Angela Demke Brown. Mak-
ing lockless synchronization fast: Performance implications of memory
reclamation. In 20th IEEE International Parallel and Distributed Pro-
cessing Symposium, Rhodes, Greece, April 2006. Available: http://
www.rdrop.com/users/paulmck/RCU/hart_ipdps06.pdf [Viewed
April 28, 2008].
[HMBW07] Thomas E. Hart, Paul E. McKenney, Angela Demke Brown, and Jonathan
Walpole. Performance of memory reclamation for lockless synchroniza-
tion. J. Parallel Distrib. Comput., 67(12):1270–1285, 2007.
[HMDZ06] David Howells, Paul E. McKenney, Will Deacon, and Peter Zĳlstra.
Linux kernel memory barriers, March 2006. https://www.kernel.
org/doc/Documentation/memory-barriers.txt.
[Hoa74] C. A. R. Hoare. Monitors: An operating system structuring concept.
Communications of the ACM, 17(10):549–557, October 1974.
[Hol03] Gerard J. Holzmann. The Spin Model Checker: Primer and Reference
Manual. Addison-Wesley, Boston, MA, USA, 2003.
v2023.06.11a
912 BIBLIOGRAPHY
[Hor18] Jann Horn. Reading privileged memory with a side-channel, Janu-

ary 2018. https://googleprojectzero.blogspot.com/2018/01/
reading-privileged-memory-with-side.html.
[HOS89] James P. Hennessy, Damian L. Osisek, and Joseph W. Seigh II. Pas-
sive serialization in a multitasking environment. Technical Report US
Patent 4,809,168, Assigned to International Business Machines Corp,
Washington, DC, February 1989.
[How12] Phil Howard. Extending Relativistic Programming to Multiple Writers.

PhD thesis, Portland State University, 2012.
[HP95] John L. Hennessy and David A. Patterson. Computer Architecture: A

Quantitative Approach. Morgan Kaufman, 1995.

Quantitative Approach, Fifth Edition. Morgan Kaufman, 2011.

Quantitative Approach, Sixth Edition. Morgan Kaufman, 2017.
[Hra13] Adam Hraška. Read-copy-update for helenos. Master’s thesis, Charles

University in Prague, Faculty of Mathematics and Physics, Department
of Distributed and Dependable Systems, 2013.
[HS08] Maurice Herlihy and Nir Shavit. The Art of Multiprocessor Programming.
Morgan Kaufmann, Burlington, MA, USA, 2008.
[HSLS20] Maurice Herlihy, Nir Shavit, Victor Luchangco, and Michael Spear. The
Art of Multiprocessor Programming, 2nd Edition. Morgan Kaufmann,
Burlington, MA, USA, 2020.
[HW90] Maurice P. Herlihy and Jeannette M. Wing. Linearizability: a correctness

condition for concurrent objects. ACM Trans. Program. Lang. Syst.,
12(3):463–492, July 1990.
[HW92] Wilson C. Hsieh and William E. Weihl. Scalable reader-writer locks

for parallel systems. In Proceedings of the 6th International Parallel
Processing Symposium, pages 216–230, Beverly Hills, CA, USA, March
1992.
[HW11] Philip W. Howard and Jonathan Walpole. A relativistic enhancement

to software transactional memory. In Proceedings of the 3rd USENIX
conference on Hot topics in parallelism, HotPar’11, pages 1–6, Berkeley,
CA, 2011. USENIX Association.
[HW14] Philip W. Howard and Jonathan Walpole. Relativistic red-black trees.

Concurrency and Computation: Practice and Experience, 26(16):2684–
2712, November 2014.
[IBM94] IBM Microelectronics and Motorola. PowerPC Microprocessor Family:

The Programming Environments, 1994.
v2023.06.11a
BIBLIOGRAPHY 913
[Inm85] Jack Inman. Implementing loosely coupled functions on tightly coupled

engines. In USENIX Conference Proceedings, pages 277–298, Portland,
OR, June 1985. USENIX Association.
[Inm07] Bill Inmon. Time value of information, January 2007. URL: http:
//www.b-eye-network.com/view/3365 [broken, February 2021].
[Int92] International Standards Organization. Information Technology - Database

Language SQL. ISO, 1992. Available (Second informal review draft
of ISO/IEC 9075:1992): http://www.contrib.andrew.cmu.edu/
~shadow/sql/sql1992.txt [Viewed September 19, 2008].
[Int02a] Intel Corporation. Intel Itanium Architecture Software Developer’s

Manual Volume 2: System Architecture, 2002.
[Int02b] Intel Corporation. Intel Itanium Architecture Software Developer’s

Manual Volume 3: Instruction Set Reference, 2002.
[Int04a] Intel Corporation. IA-32 Intel Architecture Software Developer’s Manual

Volume 2B: Instruction Set Reference, N-Z, 2004.
[Int04b] Intel Corporation. IA-32 Intel Architecture Software Developer’s Manual

Volume 3: System Programming Guide, 2004.
[Int04c] International Business Machines Corporation. z/Architecture principles

of operation, May 2004. Available: http://publibz.boulder.ibm.
com/epubs/pdf/dz9zr003.pdf [Viewed: February 16, 2005].
[Int07] Intel Corporation. Intel 64 Architecture Memory Ordering White Paper,

2007.
[Int11] Intel Corporation. Intel 64 and IA-32 Architectures Software Developer’s

Manual, Volume 3A: System Programming Guide, Part 1, 2011. Avail-
able: http://www.intel.com/Assets/PDF/manual/253668.pdf
[Viewed: February 12, 2011].
[Int16] Intel Corporation. Intel 64 and IA-32 Architectures Software Developer’s

Manual, Volume 3A: System Programming Guide, Part 1, 2016.
[Int20a] Intel. Desktop 4th Generation Intel® Core™ Processor Family, Desktop
Intel® Pentium® Processor Family, and Desktop Intel® Celeron® Pro-
cessor Family, April 2020. http://www.intel.com/content/dam/
www/public/us/en/documents/specification-updates/4th-
gen-core-family-desktop-specification-update.pdf.
[Int20b] Intel Corporation. Intel Transactional Synchronization Extensions

(Intel TSX) Programming Considerations, 2021.1 edition, December
2020. In Intel C++ Compiler Classic Developer Guide and Ref-
erence, https://software.intel.com/content/dam/develop/
external/us/en/documents/cpp_compiler_classic.pdf, page
1506.
[Int20c] International Business Machines Corporation. Power ISA™ Version 3.1,

2020.
v2023.06.11a
914 BIBLIOGRAPHY
[Int21] Intel. Performance monitoring impact of Intel® Transactional

Synchronization Extension memory ordering issue, June 2021.
https://www.intel.com/content/dam/support/us/en/
documents/processors/Performance-Monitoring-Impact-
of-TSX-Memory-Ordering-Issue-604224.pdf.
[Jac88] Van Jacobson. Congestion avoidance and control. In SIGCOMM ’88,

pages 314–329, August 1988.
[Jac93] Van Jacobson. Avoid read-side locking via delayed free, September 1993.
private communication.
[Jac08] Daniel Jackson. MapReduce course, January 2008. Available: https:

//sites.google.com/site/mriap2008/ [Viewed January 3, 2013].
[JED] JEDEC. mega (M) (as a prefix to units of semiconductor storage capacity)
[online].
[Jef14] Alan Jeffrey. Jmm revision status, July 2014. https://mail.openjdk.

java.net/pipermail/jmm-dev/2014-July/000072.html.
[JJKD21] Ralf Jung, Jacques-Henri Jourdan, Robbert Krebbers, and Derek Dreyer.
Safe systems programming in Rust. Commun. ACM, 64(4):144–152,
March 2021.
[JLK16a] Yeongjin Jang, Sangho Lee, and Taesoo Kim. Breaking kernel
address space layout randomization (KASLR) with Intel TSX, July
2016. Black Hat USA 2018 https://www.blackhat.com/us-
16/briefings.html#breaking-kernel-address-space-
layout-randomization-kaslr-with-intel-tsx.
[JLK16b] Yeongjin Jang, Sangho Lee, and Taesoo Kim. Breaking kernel address
space layout randomization with Intel TSX. In Proceedings of the 2016
ACM SIGSAC Conference on Computer and Communications Security,
CCS ’16, pages 380–392, Vienna, Austria, 2016. ACM.
[JMRR02] Benedict Joseph Jackson, Paul E. McKenney, Ramakrishnan Rajamony,

and Ronald Lynn Rockhold. Scalable interruptible queue locks for shared-
memory multiprocessor. US Patent 6,473,819, Assigned to International
Business Machines Corporation, Washington, DC, October 2002.
[Joh77] Stephen Johnson. Lint, a C program checker, December 1977. Computer

Science Technical Report 65, Bell Laboratories.
[Joh95] Aju John. Dynamic vnodes – design and implementation. In USENIX

Winter 1995, pages 11–23, New Orleans, LA, January 1995. USENIX
Association. Available: https://www.usenix.org/publications/
library/proceedings/neworl/full_papers/john.a [Viewed Oc-
tober 1, 2010].
[Jon11] Dave Jones. Trinity: A system call fuzzer. In 13th Ottawa Linux
Symposium, Ottawa, Canada, June 2011. Project repository: https:
//github.com/kernelslacker/trinity.
v2023.06.11a
BIBLIOGRAPHY 915
[JSG12] Christian Jacobi, Timothy Slegel, and Dan Greiner. Transactional memory
architecture and implementation for IBM System z. In Proceedings of the
45th Annual IEEE/ACM International Symposium on Microarchitecture,
MICRO 45, pages 25–36, Vancouver B.C. Canada, December 2012. Pre-
sentation slides: https://www.microarch.org/micro45/talks-
posters/3-jacobi-presentation.pdf.
[Kaa15] Frans Kaashoek. Parallel computing and the os. In SOSP History Day,
October 2015.
[KCH+ 06] Sanjeev Kumar, Michael Chu, Christopher J. Hughes, Partha Kumar,
and Anthony Nguyen. Hybrid transactional memory. In Proceedings
of the ACM SIGPLAN 2006 Symposium on Principles and Practice of
Parallel Programming, New York, New York, United States, 2006. ACM
SIGPLAN.
[KDI20] Alex Kogan, Dave Dice, and Shady Issa. Scalable range locks for scalable
address spaces and beyond. In Proceedings of the Fifteenth European
Conference on Computer Systems, EuroSys ’20, Heraklion, Greece, 2020.
[Kel17] Michael J. Kelly. How might the manufacturability of the hardware at

device level impact on exascale computing?, 2017. Keynote speech at Mul-
ticore World 2017, URL: https://openparallel.com/multicore-
world-2017/program-2017/abstracts2017/.
[Ken20] Chris Kennelly. TCMalloc overview, February 2020. https://google.

github.io/tcmalloc/overview.html.
[KFC11] KFC. Memristor processor solves mazes, March 2011. URL:

https://www.technologyreview.com/2011/03/03/196572/
memristor-processor-solves-mazes/.
[Khi14] Maxim Khizhinsky. Memory management schemes, June 2014.

https://kukuruku.co/post/lock-free-data-structures-
the-inside-memory-management-schemes/.
[Khi15] Max Khiszinsky. Lock-free data structures. the inside. RCU,

February 2015. https://kukuruku.co/post/lock-free-data-
structures-the-inside-rcu/.
[Kis14] Jan Kiszka. Real-time virtualization - how crazy are we?

In Linux Plumbers Conference, Duesseldorf, Germany, October
2014. URL: https://blog.linuxplumbersconf.org/2014/ocw/
proposals/1935.
[Kiv13] Avi Kivity. rcu: add basic read-copy-update implementation, August

2013. https://github.com/cloudius-systems/osv/commit/
94b69794fb9e6c99d78ca9a58ddaee1c31256b43.
[Kiv14a] Avi Kivity. rcu hashtable, July 2014. https:

//github.com/cloudius-systems/osv/commit/
7fa2728e5d03b2174b4a39d94b21940d11926e90.
v2023.06.11a
916 BIBLIOGRAPHY
[Kiv14b] Avi Kivity. rcu: introduce an rcu list type, April

2014. https://github.com/cloudius-systems/osv/commit/
4e46586093aeaf339fef8e08d123a6f6b0abde5b.
[KL80] H. T. Kung and Philip L. Lehman. Concurrent manipulation of binary
search trees. ACM Transactions on Database Systems, 5(3):354–382,
September 1980.
[Kle14] Andi Kleen. Scaling existing lock-based applications with lock elision.
Commun. ACM, 57(3):52–56, March 2014.
[Kle17] Matt Klein. Envoy threading model, July 2017. https://blog.
envoyproxy.io/envoy-threading-model-a8d44b922310.
[KLP12] Christoph M. Kirsch, Michael Lippautz, and Hannes Payer. Fast and scal-
able k-FIFO queues. Technical Report 2012-04, University of Salzburg,
Salzburg, Austria, June 2012.
[KM13] Konstantin Khlebnikov and Paul E. McKenney. RCU: non-atomic
assignment to long/pointer variables in gcc, January 2013. https:
//lore.kernel.org/lkml/50F52FC8.4000701@openvz.org/.
[KMK+ 19] Jaeho Kim, Ajit Mathew, Sanidhya Kashyap, Madhava Krishnan Ra-
manathan, and Changwoo Min. Mv-rlu: Scaling read-log-update with
multi-versioning. In Proceedings of the Twenty-Fourth International
Conference on Architectural Support for Programming Languages and
Operating Systems, ASPLOS ’19, pages 779–792, Providence, RI, USA,
2019. ACM.
[Kni86] Tom Knight. An architecture for mostly functional languages. In Proceed-
ings of the 1986 ACM Conference on LISP and Functional Programming,
LFP ’86, pages 105–112, Cambridge, Massachusetts, USA, 1986. ACM.
[Kni08] John U. Knickerbocker. 3D chip technology. IBM Journal of Research and
Development, 52(6), November 2008. URL: http://www.research.
ibm.com/journal/rd52-6.html [Link to each article is broken as
of November 2016; Available via https://ieeexplore.ieee.org/
xpl/tocresult.jsp?isnumber=5388557].
[Knu73] Donald Knuth. The Art of Computer Programming. Addison-Wesley,
1973.
[Kra17] Vlad Krasnov. On the dangers of Intel’s frequency scaling, Novem-
ber 2017. https://blog.cloudflare.com/on-the-dangers-of-
intels-frequency-scaling/.
[KS08] Daniel Kroening and Ofer Strichman. Decision Procedures: An Algo-
rithmic Point of View. Springer Publishing Company, Incorporated, 1
edition, 2008.
[KS17a] Michalis Kokologiannakis and Konstantinos Sagonas. Stateless model
checking of the linux kernel’s hierarchical read-copy update (Tree
RCU). Technical report, National Technical University of Athens, Jan-
uary 2017. https://github.com/michalis-/rcu/blob/master/
rcupaper.pdf.
v2023.06.11a
BIBLIOGRAPHY 917
[KS17b] Michalis Kokologiannakis and Konstantinos Sagonas. Stateless model

checking of the Linux kernel’s hierarchical read-copy-update (Tree RCU).
In Proceedings of International SPIN Symposium on Model Checking of
Software, SPIN 2017, New York, NY, USA, July 2017. ACM.
[KS19] Michalis Kokologiannakis and Konstantinos Sagonas. Stateless model
checking of the linux kernel’s read—copy update (RCU). Int. J. Softw.
Tools Technol. Transf., 21(3):287–306, June 2019.
[KWS97] Leonidas Kontothanassis, Robert W. Wisniewski, and Michael L. Scott.
Scheduler-conscious synchronization. ACM Transactions on Computer
Systems, 15(1):3–40, February 1997.
[LA94] Beng-Hong Lim and Anant Agarwal. Reactive synchronization algorithms
for multiprocessors. In Proceedings of the sixth international conference
on Architectural support for programming languages and operating
systems, ASPLOS VI, pages 25–35, San Jose, California, USA, October
1994. ACM.
[Lam74] Leslie Lamport. A new solution of Dĳkstra’s concurrent programming
problem. Communications of the ACM, 17(8):453–455, August 1974.
[Lam77] Leslie Lamport. Concurrent reading and writing. Commun. ACM,
20(11):806–811, November 1977.
[Lam02] Leslie Lamport. Specifying Systems: The TLA+ Language and Tools for
Hardware and Software Engineers. Addison-Wesley Longman Publishing
Co., Inc., Boston, MA, USA, 2002.
[Lar21] Michael Larabel. Intel to disable TSX by default on more CPUs with
new microcode, June 2021. https://www.phoronix.com/scan.php?
page=news_item&px=Intel-TSX-Off-New-Microcode.
[LBD+ 04] James R. Larus, Thomas Ball, Manuvir Das, Robert DeLine, Manuel Fah-
ndrich, Jon Pincus, Sriram K. Rajamani, and Ramanathan Venkatapathy.
Righting software. IEEE Softw., 21(3):92–100, May 2004.
[Lea97] Doug Lea. Concurrent Programming in Java: Design Principles and
Patterns. Addison Wesley Longman, Reading, MA, USA, 1997.
[Lem18] Daniel Lemire. AVX-512: when and how to use these new instruc-
tions, September 2018. https://lemire.me/blog/2018/09/07/
avx-512-when-and-how-to-use-these-new-instructions/.
[LGW+ 15] H. Q. Le, G. L. Guthrie, D. E. Williams, M. M. Michael, B. G. Frey,
W. J. Starke, C. May, R. Odaira, and T. Nakaike. Transactional memory
support in the ibm power8 processor. IBM J. Res. Dev., 59(1):8:1–8:14,
January 2015.
[LHF05] Michael Lyons, Bill Hay, and Brad Frey. PowerPC storage model
and AIX programming, November 2005. http://www.ibm.com/
developerworks/systems/articles/powerpc.html.
[Lis88] Barbara Liskov. Distributed programming in Argus. Commun. ACM,
31(3):300–312, 1988.
v2023.06.11a
918 BIBLIOGRAPHY
[LLO09] Yossi Lev, Victor Luchangco, and Marek Olszewski. Scalable reader-
writer locks. In SPAA ’09: Proceedings of the twenty-first annual
symposium on Parallelism in algorithms and architectures, pages 101–
110, Calgary, AB, Canada, 2009. ACM.
[LLS13] Yujie Liu, Victor Luchangco, and Michael Spear. Mindicators: A

scalable approach to quiescence. In Proceedings of the 2013 IEEE 33rd
International Conference on Distributed Computing Systems, ICDCS ’13,
pages 206–215, Washington, DC, USA, 2013. IEEE Computer Society.
[LMKM16] Lihao Liang, Paul E. McKenney, Daniel Kroening, and Tom Melham.
Verification of the tree-based hierarchical read-copy update in the Linux
kernel. Technical report, Cornell University Library, October 2016.
https://arxiv.org/abs/1610.03052.
[LMKM18] Lihao Liang, Paul E. McKenney, Daniel Kroening, and Tom Melham.
Verification of tree-based hierarchical Read-Copy Update in the Linux
Kernel. In 2018 Design, Automation & Test in Europe Conference &
Exhibition (DATE), Dresden, Germany, March 2018.
[Loc02] Doug Locke. Priority inheritance: The real story, July 2002.
URL: http://www.linuxdevices.com/articles/AT5698775833.
html [broken, November 2016], page capture available at https://www.
math.unipd.it/%7Etullio/SCD/2007/Materiale/Locke.pdf.
[Lom77] D. B. Lomet. Process structuring, synchronization, and recovery using

atomic actions. SIGSOFT Softw. Eng. Notes, 2(2):128–137, 1977. URL:
http://portal.acm.org/citation.cfm?id=808319#.
[LR80] Butler W. Lampson and David D. Redell. Experience with processes and
monitors in Mesa. Communications of the ACM, 23(2):105–117, 1980.
[LS86] Vladimir Lanin and Dennis Shasha. A symmetric concurrent b-tree

algorithm. In ACM ’86: Proceedings of 1986 ACM Fall joint computer
conference, pages 380–389, Dallas, Texas, United States, 1986. IEEE
Computer Society Press.
[LS11] Yujie Liu and Michael Spear. Toxic transactions. In TRANSACT 2011,
San Jose, CA, USA, June 2011. ACM SIGPLAN.
[LSLK14] Carl Leonardsson, Kostis Sagonas, Truc Nguyen Lam, and Michalis
Kokologiannakis. Nidhugg, July 2014. https://github.com/
nidhugg/nidhugg.
[LVK+ 17] Ori Lahav, Viktor Vafeiadis, Jeehoon Kang, Chung-Kil Hur, and Derek
Dreyer. Repairing sequential consistency in C/C++11. SIGPLAN Not.,
52(6):618–632, June 2017.
[LZC14] Ran Liu, Heng Zhang, and Haibo Chen. Scalable read-mostly synchro-
nization using passive reader-writer locks. In 2014 USENIX Annual
Technical Conference (USENIX ATC 14), pages 219–230, Philadelphia,
PA, June 2014. USENIX Association.
v2023.06.11a
BIBLIOGRAPHY 919
[MAK+ 01] Paul E. McKenney, Jonathan Appavoo, Andi Kleen, Orran Krieger, Rusty
Russell, Dipankar Sarma, and Maneesh Soni. Read-copy update. In Ot-
tawa Linux Symposium, July 2001. URL: https://www.kernel.org/
doc/ols/2001/read-copy.pdf, http://www.rdrop.com/users/
paulmck/RCU/rclock_OLS.2001.05.01c.pdf.
[Mar17] Luc Maraget. Aarch64 model vs. hardware, May 2017.
http://pauillac.inria.fr/~maranget/cats7/model-
aarch64/specific.html.
[Mar18] Catalin Marinas. Queued spinlocks model, March 2018.
https://git.kernel.org/pub/scm/linux/kernel/git/
cmarinas/kernel-tla.git.
[Mas92] H. Massalin. Synthesis: An Efficient Implementation of Fundamental
Operating System Services. PhD thesis, Columbia University, New York,
NY, 1992.
[Mat17] Norm Matloff. Programming on Parallel Machines. University of
California, Davis, Davis, CA, USA, 2017.
[MB20] Paul E. McKenney and Hans Boehm. P2055R0: A relaxed guide to
memory_order_relaxed, January 2020. http://www.open-std.org/
jtc1/sc22/wg21/docs/papers/2020/p2055r0.pdf.
[MBM+ 06] Kevin E. Moore, Jayaram Bobba, Michelle J. Moravan, Mark D. Hill, and
David A. Wood. LogTM: Log-based transactional memory. In Proceed-
ings of the 12th Annual International Symposium on High Performance
Computer Architecture (HPCA-12), Austin, Texas, United States, 2006.
IEEE. Available: http://www.cs.wisc.edu/multifacet/papers/
hpca06_logtm.pdf [Viewed December 21, 2006].
[MBWW12] Paul E. McKenney, Silas Boyd-Wickizer, and Jonathan Walpole. RCU
usage in the linux kernel: One decade later, September 2012. Techni-
cal report paulmck.2012.09.17, http://rdrop.com/users/paulmck/
techreports/survey.2012.09.17a.pdf.
[McK90] Paul E. McKenney. Stochastic fairness queuing. In IEEE IN-
FOCOM’90 Proceedings, pages 733–740, San Francisco, June
1990. The Institute of Electrical and Electronics Engineers,
Inc. Revision available: http://www.rdrop.com/users/paulmck/
scalability/paper/sfq.2002.06.04.pdf [Viewed May 26, 2008].
[McK95] Paul E. McKenney. Differential profiling. In MASCOTS 1995, pages
237–241, Toronto, Canada, January 1995.
[McK96a] Paul E. McKenney. Pattern Languages of Program Design, vol-
ume 2, chapter 31: Selecting Locking Designs for Parallel Pro-
grams, pages 501–531. Addison-Wesley, June 1996. Avail-
able: http://www.rdrop.com/users/paulmck/scalability/
paper/mutexdesignpat.pdf [Viewed February 17, 2005].
[McK96b] Paul E. McKenney. Selecting locking primitives for parallel programs.
Communications of the ACM, 39(10):75–82, October 1996.
v2023.06.11a
920 BIBLIOGRAPHY
[McK99] Paul E. McKenney. Differential profiling. Software - Practice and

Experience, 29(3):219–234, 1999.
[McK01] Paul E. McKenney. RFC: patch to allow lock-free traversal of lists with
insertion, October 2001. Available: https://lore.kernel.org/
lkml/200110090155.f991tPt22329@eng4.beaverton.ibm.com/
[Viewed January 05, 2021].
[McK03] Paul E. McKenney. Using RCU in the Linux 2.5 kernel. Linux
Journal, 1(114):18–26, October 2003. Available: https://www.
linuxjournal.com/article/6993 [Viewed November 14, 2007].
[McK04] Paul E. McKenney. Exploiting Deferred Destruction: An Analysis of
Read-Copy-Update Techniques in Operating System Kernels. PhD thesis,
OGI School of Science and Engineering at Oregon Health and Sciences
University, 2004.
[McK05a] Paul E. McKenney. Memory ordering in modern micropro-
cessors, part I. Linux Journal, 1(136):52–57, August 2005.
Available: https://www.linuxjournal.com/article/8211
http://www.rdrop.com/users/paulmck/scalability/paper/
ordering.2007.09.19a.pdf [Viewed November 30, 2007].
[McK05b] Paul E. McKenney. Memory ordering in modern microproces-
sors, part II. Linux Journal, 1(137):78–82, September 2005.
Available: https://www.linuxjournal.com/article/8212
http://www.rdrop.com/users/paulmck/scalability/paper/
ordering.2007.09.19a.pdf [Viewed November 30, 2007].
[McK05c] Paul E. McKenney. A realtime preemption overview, August 2005. URL:
[McK06] Paul E. McKenney. Sleepable RCU, October 2006. Available: https:
//lwn.net/Articles/202847/ Revised: http://www.rdrop.com/
users/paulmck/RCU/srcu.2007.01.14a.pdf [Viewed August 21,
2006].
[McK07a] Paul E. McKenney. The design of preemptible read-copy-update, October
2007. Available: https://lwn.net/Articles/253651/ [Viewed
October 25, 2007].
[McK07b] Paul E. McKenney. Immunize rcu_dereference() against crazy com-
piler writers, October 2007. Git commit: https://git.kernel.org/
linus/97b430320ce7.
[McK07c] Paul E. McKenney. [PATCH] QRCU with lockless fastpath, February
2007. Available: https://lkml.org/lkml/2007/2/25/18 [Viewed
March 27, 2008].
[McK07d] Paul E. McKenney. Priority-boosting RCU read-side critical sections,
February 2007. https://lwn.net/Articles/220677/.
[McK07e] Paul E. McKenney. RCU and unloadable modules, January 2007. Avail-
able: https://lwn.net/Articles/217484/ [Viewed November 22,
2007].
v2023.06.11a
BIBLIOGRAPHY 921
[McK07f] Paul E. McKenney. Using Promela and Spin to verify parallel algorithms,
August 2007. Available: https://lwn.net/Articles/243851/
[Viewed September 8, 2007].
[McK08a] Paul E. McKenney. Efficient support of consistent cyclic search with read-
copy update (lapsed). Technical Report US Patent 7,426,511, Assigned
to International Business Machines Corp, Washington, DC, September
2008.
[McK08b] Paul E. McKenney. Hierarchical RCU, November 2008. https://lwn.

[McK08c] Paul E. McKenney. rcu: fix rcu_try_flip_waitack_needed() to

prevent grace-period stall, May 2008. Git commit: https://git.
kernel.org/linus/d7c0651390b6.
[McK08d] Paul E. McKenney. rcu: fix misplaced mb() in rcu_enter/exit_

nohz(), March 2008. Git commit: https://git.kernel.org/
linus/ae66be9b71b1.
[McK08e] Paul E. McKenney. RCU part 3: the RCU API, January 2008. Available:
https://lwn.net/Articles/264090/ [Viewed January 10, 2008].
[McK08f] Paul E. McKenney. "Tree RCU": scalable classic RCU implementation,

December 2008. Git commit: https://git.kernel.org/linus/
64db4cfff99c.
[McK08g] Paul E. McKenney. What is RCU? part 2: Usage, January 2008. Available:
https://lwn.net/Articles/263130/ [Viewed January 4, 2008].
[McK09a] Paul E. McKenney. Re: [PATCH fyi] RCU: the bloatwatch edition,
January 2009. Available: https://lkml.org/lkml/2009/1/14/449
[Viewed January 15, 2009].
[McK09b] Paul E. McKenney. Transactional memory everywhere?, September 2009.

https://paulmck.livejournal.com/13841.html.
[McK10] Paul E. McKenney. Efficient support of consistent cyclic search with read-
copy update (lapsed). Technical Report US Patent 7,814,082, Assigned to
International Business Machines Corp, Washington, DC, October 2010.
[McK11a] Paul E. McKenney. 3.0 and RCU: what went wrong, July 2011. https:
[McK11b] Paul E. McKenney. Concurrent code and expensive instructions, Janu-

ary 2011. Available: https://lwn.net/Articles/423994 [Viewed
January 28, 2011].
[McK11c] Paul E. McKenney. Transactional memory everywhere: Htm and cache

geometry, June 2011. https://paulmck.livejournal.com/tag/
transactional%20memory%20everywhere.
[McK11d] Paul E. McKenney. Validating memory barriers and atomic instructions,

December 2011. https://lwn.net/Articles/470681/.
v2023.06.11a
922 BIBLIOGRAPHY
[McK11e] Paul E. McKenney. Verifying parallel software: Can theory meet

practice?, January 2011. http://www.rdrop.com/users/paulmck/
scalability/paper/VericoTheoryPractice.2011.01.28a.pdf.
[McK12a] Paul E. McKenney. Beyond expert-only parallel programming? In

Proceedings of the 2012 ACM workshop on Relaxing synchronization for
multicore and manycore scalability, RACES ’12, pages 25–32, Tucson,
Arizona, USA, 2012. ACM.
[McK12b] Paul E. McKenney. Making RCU safe for battery-powered devices, Feb-
ruary 2012. Available: http://www.rdrop.com/users/paulmck/
RCU/RCUdynticks.2012.02.15b.pdf [Viewed March 1, 2012].
[McK12c] Paul E. McKenney. Retrofitted parallelism considered grossly sub-optimal.

In 4th USENIX Workshop on Hot Topics on Parallelism, page 7, Berkeley,
CA, USA, June 2012.
[McK12d] Paul E. McKenney. Signed overflow optimization hazards in the kernel,

August 2012. https://lwn.net/Articles/511259/.
[McK12e] Paul E. McKenney. Transactional memory everywhere: Hardware

transactional lock elision, May 2012. Available: https://paulmck.
livejournal.com/32267.html [Viewed January 28, 2021].
[McK13] Paul E. McKenney. Structured deferral: synchronization via procrastina-

tion. Commun. ACM, 56(7):40–49, July 2013.
[McK14a] Paul E. McKenney. C++ memory model meets high-update-rate data struc-
tures, September 2014. http://www2.rdrop.com/users/paulmck/
RCU/C++Updates.2014.09.11a.pdf.
[McK14b] Paul E. McKenney. Efficient support of consistent cyclic search with read-
copy update (lapsed). Technical Report US Patent 8,874,535, Assigned to
International Business Machines Corp, Washington, DC, October 2014.
[McK14c] Paul E. McKenney. Is Parallel Programming Hard, And, If So, What

Can You Do About It? (First Edition). kernel.org, Corvallis, OR,
USA, 2014. https://kernel.org/pub/linux/kernel/people/
paulmck/perfbook/perfbook-e1.html.
[McK14d] Paul E. McKenney. N4037: Non-transactional implementation of

atomic tree move, May 2014. http://www.open-std.org/jtc1/
sc22/wg21/docs/papers/2014/n4037.pdf.
[McK14e] Paul E. McKenney. Proper care and feeding of return values from
rcu_dereference(), February 2014. https://www.kernel.org/
doc/Documentation/RCU/rcu_dereference.txt.
[McK14f] Paul E. McKenney. The RCU API, 2014 edition, September 2014.
[McK14g] Paul E. McKenney. Recent read-mostly research, November 2014. https:

v2023.06.11a
BIBLIOGRAPHY 923
[McK15a] Paul E. McKenney. Formal verification and Linux-kernel concurrency.

In Compositional Verification Methods for Next-Generation Concur-
rency, Dagstuhl Seminar Proceedings, Dagstuhl, Germany, 2015. Schloss
Dagstuhl - Leibniz-Zentrum fuer Informatik, Germany.
[McK15b] Paul E. McKenney. High-performance and scalable updates: The Issaquah
challenge, January 2015. http://www2.rdrop.com/users/paulmck/
scalability/paper/Updates.2015.01.16b.LCA.pdf.
[McK15c] Paul E. McKenney. [PATCH tip/core/rcu 01/10] rcu: Make
rcu_nmi_enter() handle nesting, January 2015. https:
//lore.kernel.org/lkml/1420651257-553-1-git-send-
email-paulmck@linux.vnet.ibm.com/.
[McK15d] Paul E. McKenney. Practical experience with formal verification tools. In
Verified Trustworthy Software Systems Specialist Meeting. The Royal
Society, April 2015. http://www.rdrop.com/users/paulmck/
scalability/paper/Validation.2016.04.06e.SpecMtg.pdf.
[McK15e] Paul E. McKenney. RCU requirements part 2 — parallelism and software
engineering, August 2015. https://lwn.net/Articles/652677/.
[McK15f] Paul E. McKenney. RCU requirements part 3, August 2015. https:
[McK15g] Paul E. McKenney. Re: [patch tip/locking/core v4 1/6] powerpc:
atomic: Make *xchg and *cmpxchg a full barrier, October 2015.
Email thread: https://lore.kernel.org/lkml/20151014201916.
GB3910@linux.vnet.ibm.com/.
[McK15h] Paul E. McKenney. Requirements for RCU part 1: the fundamentals, July
2015. https://lwn.net/Articles/652156/.
[McK16a] Paul E. McKenney. Beyond the Issaquah challenge: High-performance
scalable complex updates, September 2016. http://www2.rdrop.com/
users/paulmck/RCU/Updates.2016.09.19i.CPPCON.pdf.
[McK16b] Paul E. McKenney. High-performance and scalable updates: The Issaquah
challenge, June 2016. http://www2.rdrop.com/users/paulmck/
RCU/Updates.2016.06.01e.ACM.pdf.
[McK17] Paul E. McKenney. Verification challenge 6: Linux-kernel Tree RCU,
June 2017. https://paulmck.livejournal.com/46993.html.
[McK19a] Paul E. McKenney. A critical RCU safety property is... Ease of use! In
Proceedings of the 12th ACM International Conference on Systems and
Storage, SYSTOR ’19, pages 132–143, Haifa, Israel, 2019. ACM.
[McK19b] Paul E. McKenney. The RCU API, 2019 edition, January 2019. https:
[McK19c] Paul E. McKenney. RCU’s first-ever CVE, and how i lived to tell
the tale, January 2019. linux.conf.au Slides: http://www.rdrop.
com/users/paulmck/RCU/cve.2019.01.23e.pdf Video: https:
//www.youtube.com/watch?v=hZX1aokdNiY.
v2023.06.11a
924 BIBLIOGRAPHY
[MCM02] Paul E. McKenney, Kevin A. Closson, and Raghupathi Malige. Lingering

locks with fairness control for multi-node computer systems. US Patent
6,480,918, Assigned to International Business Machines Corporation,
Washington, DC, November 2002.
[MCS91] John M. Mellor-Crummey and Michael L. Scott. Algorithms for scalable
synchronization on shared-memory multiprocessors. Transactions of
Computer Systems, 9(1):21–65, February 1991.
[MD92] Paul E. McKenney and Ken F. Dove. Efficient demultiplexing of incoming
tcp packets. In SIGCOMM ’92, Proceedings of the Conference on
Communications Architecture & Protocols, pages 269–279, Baltimore,
MD, August 1992. Association for Computing Machinery.
[MDJ13a] Paul E. McKenney, Mathieu Desnoyers, and Lai Jiangshan. The RCU-
protected list API, November 2013. https://lwn.net/Articles/
573441/.
[MDJ13b] Paul E. McKenney, Mathieu Desnoyers, and Lai Jiangshan. The
URCU hash table API, November 2013. https://lwn.net/Articles/
573432/.
[MDJ13c] Paul E. McKenney, Mathieu Desnoyers, and Lai Jiangshan. URCU-
protected hash tables, November 2013. https://lwn.net/Articles/
573431/.
[MDJ13d] Paul E. McKenney, Mathieu Desnoyers, and Lai Jiangshan. URCU-
protected queues and stacks, November 2013. https://lwn.net/
Articles/573433/.
[MDJ13e] Paul E. McKenney, Mathieu Desnoyers, and Lai Jiangshan. The URCU
stack/queue API, November 2013. https://lwn.net/Articles/
573434/.
[MDJ13f] Paul E. McKenney, Mathieu Desnoyers, and Lai Jiangshan. User-space
RCU, November 2013. https://lwn.net/Articles/573424/.
[MDJ13g] Paul E. McKenney, Mathieu Desnoyers, and Lai Jiangshan. The user-space
RCU API, November 2013. https://lwn.net/Articles/573439/.
[MDJ13h] Paul E. McKenney, Mathieu Desnoyers, and Lai Jiangshan. User-space
RCU: Atomic-operation and utility API, November 2013. https://
lwn.net/Articles/573435/.
[MDJT13a] Paul E. McKenney, Mathieu Desnoyers, Lai Jiangshan, and Josh Triplett.
The RCU-barrier menagerie, November 2013. https://lwn.net/
Articles/573497/.
[MDJT13b] Paul E. McKenney, Mathieu Desnoyers, Lai Jiangshan, and Josh Triplett.
User-space RCU: Memory-barrier menagerie, November 2013. https:
[MDR16] Paul E. McKenney, Will Deacon, and Luis R. Rodriguez. Semantics of
MMIO mapping attributes across architectures, August 2016. https:
v2023.06.11a
BIBLIOGRAPHY 925
[MDSS20] Hans Meuer, Jack Dongarra, Erich Strohmaier, and Horst Simon. Top 500:
The list, November 2020. Available: https://top500.org/lists/
[Men16] Alexis Menard. Move OneWriterSeqLock and SharedMemorySe-

qLockBuffer from content/ to device/base/synchronization, September
2016. https://source.chromium.org/chromium/chromium/src/
+/b39a3082846d5877a15e8b7e18d66cb142abe8af.
[Mer11] Rick Merritt. IBM plants transactional memory in CPU, August 2011. EE
Times https://www.eetimes.com/ibm-plants-transactional-
memory-in-cpu/.
[Met99] Panagiotis Takis Metaxas. Fast dithering on a data-parallel computer. In

Proceedings of the IASTED International Conference on Parallel and
Distributed Computing and Systems, pages 570–576, Cambridge, MA,
USA, 1999. IASTED.
[MG92] Paul E. McKenney and Gary Graunke. Efficient buffer allocation on

shared-memory multiprocessors. In IEEE Workshop on the Architecture
and Implementation of High Performance Communication Subsystems,
pages 194–199, Tucson, AZ, February 1992. The Institute of Electrical
and Electronics Engineers, Inc.
[MGM+ 09] Paul E. McKenney, Manish Gupta, Maged M. Michael, Phil Howard,
Joshua Triplett, and Jonathan Walpole. Is parallel programming hard,
and if so, why? Technical Report TR-09-02, Portland State University,
Portland, OR, USA, February 2009. URL: https://archives.pdx.
edu/ds/psu/10386 [Viewed February 13, 2021].
[MHS12] Milo M. K. Martin, Mark D. Hill, and Daniel J. Sorin. Why on-chip
coherence is here to stay. Communications of the ACM, 55(7):78–89,
July 2012.
[Mic02] Maged M. Michael. Safe memory reclamation for dynamic lock-free

objects using atomic reads and writes. In Proceedings of the 21st Annual
ACM Symposium on Principles of Distributed Computing, pages 21–30,
August 2002.
[Mic03] Maged M. Michael. Cas-based lock-free algorithm for shared deques. In

Harald Kosch, László Böszörményi, and Hermann Hellwagner, editors,
Euro-Par, volume 2790 of Lecture Notes in Computer Science, pages
651–660. Springer, 2003.
[Mic04a] Maged M. Michael. Hazard pointers: Safe memory reclamation for

lock-free objects. IEEE Transactions on Parallel and Distributed Systems,
15(6):491–504, June 2004.
[Mic04b] Maged M. Michael. Scalable lock-free dynamic memory allocation.

SIGPLAN Not., 39(6):35–46, 2004.
[Mic08] Microsoft. FlushProcessWriteBuffers function, 2008.

https://docs.microsoft.com/en-us/windows/desktop/
v2023.06.11a
926 BIBLIOGRAPHY
api/processthreadsapi/nf-processthreadsapi-
flushprocesswritebuffers.
[Mic18] Maged Michael. Rewrite from experimental, use of de-

terministic schedule, improvements, June 2018. Hazard
pointers: https://github.com/facebook/folly/commit/
d42832d2a529156275543c7fa7183e1321df605d.
[Mil06] David S. Miller. Re: [PATCH, RFC] RCU : OOM avoidance and lower
latency, January 2006. Available: https://lkml.org/lkml/2006/1/
7/22 [Viewed February 29, 2012].
[MJST16] Paul E. McKenney, Alan Jeffrey, Ali Sezgin, and Tony Tye. Out-of-thin-air
execution is vacuous, July 2016. http://www.open-std.org/jtc1/
sc22/wg21/docs/papers/2016/p0422r0.html.
[MK88] Marshall Kirk McKusick and Michael J. Karels. Design of a general

purpose memory allocator for the 4.3BSD UNIX kernel. In USENIX
Conference Proceedings, Berkeley CA, June 1988.
[MKM12] Yandong Mao, Eddie Kohler, and Robert Tappan Morris. Cache craftiness
for fast multicore key-value storage. In Proceedings of the 7th ACM
European Conference on Computer Systems, EuroSys ’12, pages 183–196,
Bern, Switzerland, 2012. ACM.
[ML82] Udi Manber and Richard E. Ladner. Concurrency control in a dynamic

search structure. Technical Report 82-01-01, Department of Computer
Science, University of Washington, Seattle, Washington, January 1982.
[ML84] Udi Manber and Richard E. Ladner. Concurrency control in a dynamic

search structure. ACM Transactions on Database Systems, 9(3):439–455,
September 1984.
[MLH94] Peter Magnusson, Anders Landin, and Erik Hagersten. Efficient software
synchronization on large cache coherent multiprocessors. Technical
Report T94:07, Swedish Institute of Computer Science, Kista, Sweden,
February 1994.
[MM00] Ingo Molnar and David S. Miller. brlock, March 2000. URL: http:
//kernel.nic.funet.fi/pub/linux/kernel/v2.3/patch-
html/patch-2.3.49/linux_include_linux_brlock.h.html.
[MMM+ 20] Paul E. McKenney, Maged Michael, Jens Maurer, Peter Sewell, Mar-
tin Uecker, Hans Boehm, Hubert Tong, Niall Douglas, Thomas
Rodgers, Will Deacon, Michael Wong, David Goldblatt, Kostya Sere-
bryany, and Anthony Williams. P1726R4: Pointer lifetime-end zap,
July 2020. http://www.open-std.org/jtc1/sc22/wg21/docs/
[MMS19] Paul E. McKenney, Maged Michael, and Peter Sewell. N2369: Pointer
lifetime-end zap, April 2019. http://www.open-std.org/jtc1/
sc22/wg14/www/docs/n2369.pdf.
v2023.06.11a
BIBLIOGRAPHY 927
[MMTW10] Paul E. McKenney, Maged M. Michael, Josh Triplett, and Jonathan

Walpole. Why the grass may not be greener on the other side: a comparison
of locking vs. transactional memory. ACM Operating Systems Review,
44(3), July 2010.
[MMW07] Paul E. McKenney, Maged Michael, and Jonathan Walpole. Why the
grass may not be greener on the other side: A comparison of locking
vs. transactional memory. In Programming Languages and Operating
Systems, pages 1–5, Stevenson, Washington, USA, October 2007. ACM
SIGOPS.
[Mol05] Ingo Molnar. Index of /pub/linux/kernel/projects/rt, February 2005. URL:

https://www.kernel.org/pub/linux/kernel/projects/rt/.
[Mol06] Ingo Molnar. Lightweight robust futexes, March 2006. Avail-

able: https://www.kernel.org/doc/Documentation/robust-
futexes.txt [Viewed February 14, 2021].
[Moo65] Gordon E. Moore. Cramming more components onto integrated circuits.

Electronics, 38(8):114–117, April 1965.
[Moo03] Gordon Moore. No exponential is forever–but we can delay forever. In

IBM Academy of Technology 2003 Annual Meeting, San Francisco, CA,
October 2003.
[Mor07] Richard Morris. Sir Tony Hoare: Geek of the week, Au-
gust 2007. https://www.red-gate.com/simple-talk/opinion/
geek-of-the-week/sir-tony-hoare-geek-of-the-week/.
[MOZ09] Nicholas Mc Guire, Peter Odhiambo Okech, and Qingguo Zhou. Analysis
of inherent randomness of the linux kernel. In Eleventh Real Time Linux
Workshop, Dresden, Germany, September 2009.
[MP15a] Paul E. McKenney and Aravinda Prasad. Recent read-mostly research in

2015, December 2015. https://lwn.net/Articles/667593/.
[MP15b] Paul E. McKenney and Aravinda Prasad. Some more details on read-log-
update, December 2015. https://lwn.net/Articles/667720/.
[MPA+ 06] Paul E. McKenney, Chris Purcell, Algae, Ben Schumin, Gaius Cornelius,
Qwertyus, Neil Conway, Sbw, Blainster, Canis Rufus, Zoicon5, Anome,
and Hal Eisen. Read-copy update, July 2006. https://en.wikipedia.
org/wiki/Read-copy-update.
[MPI08] MPI Forum. Message passing interface forum, September 2008. Available:
http://www.mpi-forum.org/ [Viewed September 9, 2008].
[MR08] Paul E. McKenney and Steven Rostedt. Integrating and validating

dynticks and preemptable RCU, April 2008. Available: https://lwn.
net/Articles/279077/ [Viewed April 24, 2008].
[MRP+ 17] Paul E. McKenney, Torvald Riegel, Jeff Preshing, Hans Boehm, Clark
Nelson, Olivier Giroux, Lawrence Crowl, JF Bastian, and Michael
v2023.06.11a
928 BIBLIOGRAPHY
Wong. Marking memory order consume dependency chains, Feb-

ruary 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/
[MS93] Paul E. McKenney and Jack Slingwine. Efficient kernel memory al-
location on shared-memory multiprocessors. In USENIX Conference
Proceedings, pages 295–306, Berkeley CA, February 1993. USENIX
Association. Available: http://www.rdrop.com/users/paulmck/
scalability/paper/mpalloc.pdf [Viewed January 30, 2005].
[MS95] Maged M. Michael and Michael L. Scott. Correction of a memory man-

agement method for lock-free data structures, December 1995. Technical
Report TR599.
[MS96] M.M Michael and M. L. Scott. Simple, fast, and practical non-blocking
and blocking concurrent queue algorithms. In Proc of the Fifteenth ACM
Symposium on Principles of Distributed Computing, pages 267–275, May
1996.
[MS98a] Paul E. McKenney and John D. Slingwine. Read-copy update: Using exe-
cution history to solve concurrency problems. In Parallel and Distributed
Computing and Systems, pages 509–518, Las Vegas, NV, October 1998.
[MS98b] Maged M. Michael and Michael L. Scott. Nonblocking algorithms and

preemption-safe locking on multiprogrammed shared memory multipro-
cessors. J. Parallel Distrib. Comput., 51(1):1–26, 1998.
[MS01] Paul E. McKenney and Dipankar Sarma. Read-copy update mutual exclu-
sion in Linux, February 2001. Available: http://lse.sourceforge.
net/locking/rcu/rcupdate_doc.html [Viewed October 18, 2004].
[MS08] MySQL AB and Sun Microsystems. MySQL Downloads, November 2008.

Available: http://dev.mysql.com/downloads/ [Viewed November
26, 2008].
[MS09] Paul E. McKenney and Raul Silvera. Example POWER im-

plementation for C/C++ memory model, February 2009. Avail-
able: http://www.rdrop.com/users/paulmck/scalability/
paper/N2745r.2009.02.27a.html [Viewed: April 5, 2009].
[MS12] Alexander Matveev and Nir Shavit. Towards a fully pessimistic STM
model. In TRANSACT 2012, San Jose, CA, USA, February 2012. ACM
SIGPLAN.
[MS14] Paul E. McKenney and Alan Stern. Axiomatic validation of memory

barriers and atomic instructions, August 2014. https://lwn.net/
Articles/608550/.
[MS18] Luc Maranget and Alan Stern. lock.cat, May 2018. https:
//github.com/torvalds/linux/blob/master/tools/memory-
model/lock.cat.
v2023.06.11a
BIBLIOGRAPHY 929
[MSA+ 02] Paul E. McKenney, Dipankar Sarma, Andrea Arcangeli, Andi Kleen,
Orran Krieger, and Rusty Russell. Read-copy update. In Ottawa Linux Sym-
posium, pages 338–367, June 2002. Available: https://www.kernel.
org/doc/ols/2002/ols2002-pages-338-367.pdf [Viewed Febru-
ary 14, 2021].
[MSFM15] Alexander Matveev, Nir Shavit, Pascal Felber, and Patrick Marlier. Read-
log-update: A lightweight synchronization mechanism for concurrent
programming. In Proceedings of the 25th Symposium on Operating
Systems Principles, SOSP ’15, pages 168–183, Monterey, California,
2015. ACM.
[MSK01] Paul E. McKenney, Jack Slingwine, and Phil Krueger. Experience with
an efficient parallel kernel memory allocator. Software – Practice and
Experience, 31(3):235–257, March 2001.
[MSM05] Timothy G. Mattson, Beverly A. Sanders, and Berna L. Massingill.

Patterns for Parallel Programming. Addison Wesley, Boston, MA, USA,
2005.
[MSS04] Paul E. McKenney, Dipankar Sarma, and Maneesh Soni. Scaling dcache
with RCU. Linux Journal, 1(118):38–46, January 2004.
[MSS12] Luc Maranget, Susmit Sarkar, and Peter Sewell. A tutorial introduction to
the ARM and POWER relaxed memory models, October 2012. https:
//www.cl.cam.ac.uk/~pes20/ppc-supplemental/test7.pdf.
[MT01] Jose F. Martinez and Josep Torrellas. Speculative locks for concurrent exe-
cution of critical sections in shared-memory multiprocessors. In Workshop
on Memory Performance Issues, International Symposium on Computer
Architecture, Gothenburg, Sweden, June 2001. Available: https://
iacoma.cs.uiuc.edu/iacoma-papers/wmpi_locks.pdf [Viewed
June 23, 2004].
[MT02] Jose F. Martinez and Josep Torrellas. Speculative synchronization:

Applying thread-level speculation to explicitly parallel applications. In
Proceedings of the 10th International Conference on Architectural Support
for Programming Languages and Operating Systems, pages 18–29, San
Jose, CA, October 2002.
[Mud01] Trevor Mudge. POWER: A first-class architectural design constraint.

IEEE Computer, 34(4):52–58, April 2001.
[Mus04] Museum Victoria Australia. CSIRAC: Australia’s first computer, 2004.

URL: http://museumvictoria.com.au/csirac/.
[MW05] Paul E. McKenney and Jonathan Walpole. RCU semantics: A first attempt,
January 2005. Available: http://www.rdrop.com/users/paulmck/
RCU/rcu-semantics.2005.01.30a.pdf [Viewed December 6, 2009].
[MW07] Paul E. McKenney and Jonathan Walpole. What is RCU, fundamentally?,

December 2007. Available: https://lwn.net/Articles/262464/
[Viewed December 27, 2007].
v2023.06.11a
930 BIBLIOGRAPHY
[MW11] Paul E. McKenney and Jonathan Walpole. Efficient support of consistent

cyclic search with read-copy update and parallel updates (lapsed). Tech-
nical Report US Patent 7,953,778, Assigned to International Business
Machines Corp, Washington, DC, May 2011.
[MWB+ 17] Paul E. McKenney, Michael Wong, Hans Boehm, Jens Maurer,
Jeffrey Yasskin, and JF Bastien. P0190R4: Proposal for new
memory_order_consume definition, July 2017. http://www.open-
std.org/jtc1/sc22/wg21/docs/papers/2017/p0190r4.pdf.
[MWPF18] Paul E. McKenney, Ulrich Weigand, Andrea Parri, and Boqun Feng.
Linux-kernel memory model, September 2018. http://www.open-
std.org/jtc1/sc22/wg21/docs/papers/2018/p0124r6.html.
[Mye79] Glenford J. Myers. The Art of Software Testing. Wiley, 1979.
[NA18] Catherine E. Nemitz and James H. Anderson. Work-in-progress: Lock-

based software transactional memory for real-time systems. In 2018 IEEE
Real-Time Systems Symposium, RTSS’18, pages 147–150, Nashville, TN,
USA, 2018. IEEE.
[Nag18] Honnappa Nagarahalli. rcu: add RCU library supporting QSBR

mechanism, May 2018. https://git.dpdk.org/dpdk/tree/lib/
librte_rcu.
[Nata] National Institure of Standards and Technology. SI Unit rules and style
conventions [online].
[Natb] National Institure of Standards and Technology. Typefaces for Symbols

in Scientific Manuscripts [online].
[Nat19] National Institure of Standards and Technology. The international system

of units (SI). Technical Report NIST Special Publication 330 2019
EDITION, U.S. Department of Commerce, Washington, D.C., 2019.
[Nes06a] Oleg Nesterov. Re: [patch] cpufreq: mark cpufreq_tsc() as

core_initcall_sync, November 2006. Available: https://lkml.
org/lkml/2006/11/19/69 [Viewed May 28, 2007].
[Nes06b] Oleg Nesterov. Re: [rfc, patch 1/2] qrcu: "quick" srcu implementation,
November 2006. Available: https://lkml.org/lkml/2006/11/29/
330 [Viewed November 26, 2008].
[NSHW20] Vĳay Nagarajan, Daniel J. Sorin, Mark D. Hill, and David A. Wood. A
Primer on Memory Consistency and Cache Coherence, Second Edition.
Synthesis Lectures on Computer Architecture. Morgan & Claypool, 2020.
[NVi17a] NVidia. Accelerated computing — training, January 2017. https:

//developer.nvidia.com/accelerated-computing-training.
[NVi17b] NVidia. Existing university courses, January 2017. https://

developer.nvidia.com/educators/existing-courses.
v2023.06.11a
BIBLIOGRAPHY 931
[NZ13] Oleg Nesterov and Peter Zĳlstra. rcu: Create rcu_sync infrastructure, Oc-
tober 2013. https://lore.kernel.org/lkml/20131002150518.
675931976@infradead.org/.
[O’H19] Peter W. O’Hearn. Incorrectness logic. Proc. ACM Program. Lang.,
4(POPL), December 2019.
[OHOC20] Robert O’Callahan, Kyle Huey, Devon O’Dell, and Terry Coatta. To catch
a failure: The record-and-replay approach to debugging: A discussion
with robert o’callahan, kyle huey, devon o’dell, and terry coatta. Queue,
18(1):61–79, February 2020.
[ON07] Robert Olsson and Stefan Nilsson. TRASH: A dynamic LC-trie and hash
data structure. In Workshop on High Performance Switching and Routing
(HPSR’07), May 2007.
[ONH+ 96] Kunle Olukotun, Basem A. Nayfeh, Lance Hammond, Ken Wilson, and
Kunyung Chang. The case for a single-chip multiprocessor. In ASPLOS
VII, Cambridge, MA, USA, October 1996.
[Ope97] Open Group. The single UNIX specification, version 2: Threads, 1997.
Available: http://www.opengroup.org/onlinepubs/007908799/
xsh/threads.html [Viewed September 19, 2008].
[ORY01] Peter W. O’Hearn, John C. Reynolds, and Hongseok Yang. Local
reasoning about programs that alter data structures. In Proceedings of
the 15th International Workshop on Computer Science Logic, CSL ’01,
page 1–19, Berlin, Heidelberg, 2001. Springer-Verlag.
[PAB+ 95] Calton Pu, Tito Autrey, Andrew Black, Charles Consel, Crispin Cowan,
Jon Inouye, Lakshmi Kethana, Jonathan Walpole, and Ke Zhang. Opti-
mistic incremental specialization: Streamlining a commercial operating
system. In 15th ACM Symposium on Operating Systems Principles
(SOSP’95), pages 314–321, Copper Mountain, CO, December 1995.
[Pat10] David Patterson. The trouble with multicore. IEEE Spectrum, 2010:28–32,
52–53, July 2010.
[PAT11] V Pankratius and A R Adl-Tabatabai. A study of transactional memory
vs. locks in practice. In Proceedings of the 23rd ACM symposium on
Parallelism in algorithms and architectures (2011), SPAA ’11, pages
43–52, San Jose, CA, USA, 2011. ACM.
[PBCE20] Elizabeth Patitsas, Jesse Berlin, Michelle Craig, and Steve Easterbrook.
Evidence that computer science grades are not bimodal. Commun. ACM,
63(1):91–98, January 2020.
[PD11] Martin Pohlack and Stephan Diestelhorst. From lightweight hardware
transactional memory to lightweight lock elision. In TRANSACT 2011,
San Jose, CA, USA, June 2011. ACM SIGPLAN.
[Pen18] Roman Penyaev. [PATCH v2 01/26] introduce list_next_or_null_rr_rcu(),
May 2018. https://lkml.kernel.org/r/20180518130413.
16997-2-roman.penyaev@profitbricks.com.
v2023.06.11a
932 BIBLIOGRAPHY
[Pet06] Jeremy Peters. From reuters, automatic trading linked to news events,
December 2006. URL: http://www.nytimes.com/2006/12/11/
technology/11reuters.html?ei=5088&en=e5e9416415a9eeb2&
ex=1323493200...
[Pig06] Nick Piggin. [patch 3/3] radix-tree: RCU lockless readside, June
March 25, 2008].
[Pik17] Fedor G. Pikus. Read, copy, update... Then what?, September 2017.
https://www.youtube.com/watch?v=rxQ5K9lo034.
[PMDY20] SeongJae Park, Paul E. McKenney, Laurent Dufour, and Heon Y. Yeom.
An htm-based update-side synchronization for rcu on numa systems. In
Proceedings of the Fifteenth European Conference on Computer Systems,
EuroSys ’20, Heraklion, Greece, 2020. Association for Computing
Machinery.
[Pod10] Andrej Podzimek. Read-copy-update for opensolaris. Master’s thesis,

Charles University in Prague, 2010.
[Pok16] Michael Pokorny. The deadlock empire, February 2016. https://

deadlockempire.github.io/.
[Pos08] PostgreSQL Global Development Group. PostgreSQL, November 2008.

Available: https://www.postgresql.org/ [Viewed November 26,
2008].
[Pug90] William Pugh. Concurrent maintenance of skip lists. Technical Report CS-
TR-2222.1, Institute of Advanced Computer Science Studies, Department
of Computer Science, University of Maryland, College Park, Maryland,
June 1990.
[Pug00] William Pugh. Reordering on an Alpha processor, 2000.

Available: https://www.cs.umd.edu/~pugh/java/memoryModel/
AlphaReordering.html [Viewed: June 23, 2004].
[Pul00] Geoffrey K. Pullum. How Dr. Seuss would prove the halting problem
undecidable. Mathematics Magazine, 73(4):319–320, 2000. http:
//www.lel.ed.ac.uk/~gpullum/loopsnoop.html.
[PW07] Donald E. Porter and Emmett Witchel. Lessons from large

transactional systems, December 2007. Personal communication
<20071214220521.GA5721@olive-green.cs.utexas.edu>.
[Ras14] Mindaugas Rasiukevicius. NPF—progress and perspective. In AsiaBSD-

Con, Tokyo, Japan, March 2014.
[Ras16] Mindaugas Rasiukevicius. Quiescent-state and epoch based reclamation,

July 2016. https://github.com/rmind/libqsbr.
[Ray99] Eric S. Raymond. The Cathedral and the Bazaar: Musings on Linux and
Open Source by an Accidental Revolutionary. O’Reilly, 1999.
v2023.06.11a
BIBLIOGRAPHY 933
[RC15] Pedro Ramalhete and Andreia Correia. Poor man’s URCU, Au-
gust 2015. https://github.com/pramalhe/ConcurrencyFreaks/
blob/master/papers/poormanurcu-2015.pdf.
[RD12] Ravi Rajwar and Martin Dixon. Intel transactional synchronization exten-
sions, September 2012. Intel Developer Forum (IDF) 2012 ARCS004.
[Reg10] John Regehr. A guide to undefined behavior in C and C++, part 1, July
2010. https://blog.regehr.org/archives/213.
[Rei07] James Reinders. Intel Threading Building Blocks. O’Reilly, Sebastopol,

CA, USA, 2007.
[RG01] Ravi Rajwar and James R. Goodman. Speculative lock elision: Enabling
highly concurrent multithreaded execution. In Proceedings of the 34th
Annual ACM/IEEE International Symposium on Microarchitecture, MI-
CRO 34, pages 294–305, Austin, TX, December 2001. The Institute of
Electrical and Electronics Engineers, Inc.
[RG02] Ravi Rajwar and James R. Goodman. Transactional lock-free execution of

lock-based programs. In Proceedings of the 10th International Conference
on Architectural Support for Programming Languages and Operating
Systems, pages 5–17, Austin, TX, October 2002.
[RH02] Zoran Radović and Erik Hagersten. Efficient synchronization for nonuni-
form communication architectures. In Proceedings of the 2002 ACM/IEEE
Conference on Supercomputing, pages 1–13, Baltimore, Maryland, USA,
November 2002. The Institute of Electrical and Electronics Engineers,
Inc.
[RH03] Zoran Radović and Erik Hagersten. Hierarchical backoff locks for
nonuniform communication architectures. In Proceedings of the Ninth
International Symposium on High Performance Computer Architecture
(HPCA-9), pages 241–252, Anaheim, California, USA, February 2003.
[RH18] Geoff Romer and Andrew Hunter. An RAII interface for deferred
reclamation, March 2018. http://www.open-std.org/jtc1/sc22/
wg21/docs/papers/2018/p0561r4.html.
[RHP+ 07] Chistopher J. Rossbach, Owen S. Hofmann, Donald E. Porter, Hany E.

Ramadan, Aditya Bhandari, and Emmett Witchel. TxLinux: Us-
ing and managing hardware transactional memory in an operating
system. In SOSP’07: Twenty-First ACM Symposium on Operat-
ing Systems Principles, Stevenson, WA, USA, October 2007. ACM
SIGOPS. Available: http://www.sosp2007.org/papers/sosp056-
rossbach.pdf [Viewed October 21, 2007].
[Rin13] Martin Rinard. Parallel synchronization-free approximate data structure

construction. In Proceedings of the 5th USENIX Conference on Hot
Topics in Parallelism, HotPar’13, page 6, San Jose, CA, 2013. USENIX
Association.
v2023.06.11a
934 BIBLIOGRAPHY
[RKM+ 10] Arun Raman, Hanjun Kim, Thomas R. Mason, Thomas B. Jablin, and
David I. August. Speculative parallelization using software multi-threaded
transactions. SIGARCH Comput. Archit. News, 38(1):65–76, 2010.
[RLPB18] Yuxin Ren, Guyue Liu, Gabriel Parmer, and Björn Brandenburg. Scalable
memory reclamation for multi-core, real-time systems. In Proceedings of
the 2018 IEEE Real-Time and Embedded Technology and Applications
Symposium (RTAS), page 12, Porto, Portugal, April 2018. IEEE.
[RMF19] Federico Reghenzani, Giuseppe Massari, and William Fornaciari. The
real-time Linux kernel: A survey on PREEMPT_RT. ACM Comput.
Surv., 52(1):18:1–18:36, February 2019.
[Ros06] Steven Rostedt. Lightweight PI-futexes, June 2006. Avail-
able: https://www.kernel.org/doc/html/latest/locking/pi-
futex.html [Viewed February 14, 2021].
[Ros10a] Steven Rostedt. tracing: Harry Potter and the Deathly Macros, December
2010. Available: https://lwn.net/Articles/418710/ [Viewed:
August 28, 2011].
[Ros10b] Steven Rostedt. Using the TRACE_EVENT() macro (part 1), March 2010.
Available: https://lwn.net/Articles/379903/ [Viewed: August
28, 2011].
[Ros10c] Steven Rostedt. Using the TRACE_EVENT() macro (part 2), March 2010.
28, 2011].
[Ros10d] Steven Rostedt. Using the TRACE_EVENT() macro (part 3), April 2010.
28, 2011].
[Ros11] Steven Rostedt. lockdep: How to read its cryptic output, September 2011.
http://www.linuxplumbersconf.org/2011/ocw/sessions/153.
[Roy17] Lance Roy. rcutorture: Add CBMC-based formal verification for SRCU,
January 2017. URL: https://www.spinics.net/lists/kernel/
msg2421833.html.
[RR20] Sergio Rajsbaum and Michel Raynal. Mastering concurrent computing
through sequential thinking. Commun. ACM, 63(1):78–87, January 2020.
[RSB+ 97] Rajeev Rastogi, S. Seshadri, Philip Bohannon, Dennis W. Leinbaugh,
Abraham Silberschatz, and S. Sudarshan. Logical and physical versioning
in main memory databases. In Proceedings of the 23rd International
Conference on Very Large Data Bases, VLDB ’97, pages 86–95, San
Francisco, CA, USA, August 1997. Morgan Kaufmann Publishers Inc.
[RTY+ 87] Richard Rashid, Avadis Tevanian, Michael Young, David Golub, Robert
Baron, David Black, William Bolosky, and Jonathan Chew. Machine-
independent virtual memory management for paged uniprocessor and
multiprocessor architectures. In 2nd Symposium on Architectural Support
for Programming Languages and Operating Systems, pages 31–39, Palo
Alto, CA, October 1987. Association for Computing Machinery.
v2023.06.11a
BIBLIOGRAPHY 935
[Rus00a] Rusty Russell. Re: modular net drivers, June 2000.

URL: http://oss.sgi.com/projects/netdev/archive/2000-
06/msg00250.html [broken, February 15, 2021].
[Rus00b] Rusty Russell. Re: modular net drivers, June 2000.

URL: http://oss.sgi.com/projects/netdev/archive/2000-
06/msg00254.html [broken, February 15, 2021].
[Rus03] Rusty Russell. Hanging out with smart people: or... things I learned
being a kernel monkey, July 2003. 2003 Ottawa Linux Symposium
Keynote https://ozlabs.org/~rusty/ols-2003-keynote/ols-
keynote-2003.html.
[Rut17] Mark Rutland. compiler.h: Remove ACCESS_ONCE(), November 2017.

Git commit: https://git.kernel.org/linus/b899a850431e.
[SAE+ 18] Caitlin Sadowski, Edward Aftandilian, Alex Eagle, Liam Miller-Cushon,
and Ciera Jaspan. Lessons from building static analysis tools at google.
Commun. ACM, 61(4):58–66, March 2018.
[SAH+ 03] Craig A. N. Soules, Jonathan Appavoo, Kevin Hui, Dilma Da Silva, Gre-
gory R. Ganger, Orran Krieger, Michael Stumm, Robert W. Wisniewski,
Marc Auslander, Michal Ostrowski, Bryan Rosenburg, and Jimi Xenidis.
System support for online reconfiguration. In Proceedings of the 2003
USENIX Annual Technical Conference, pages 141–154, San Antonio,
Texas, USA, June 2003. USENIX Association.
[SATG+ 09] Tatiana Shpeisman, Ali-Reza Adl-Tabatabai, Robert Geva, Yang Ni, and
Adam Welc. Towards transactional memory semantics for C++. In SPAA
’09: Proceedings of the twenty-first annual symposium on Parallelism in
algorithms and architectures, pages 49–58, Calgary, AB, Canada, 2009.
ACM.
[SBN+ 20] Dimitrios Siakavaras, Panagiotis Billis, Konstantinos Nikas, Georgios

Goumas, and Nectarios Koziris. Efficient concurrent range queries in
b+-trees using rcu-htm. In Proceedings of the 32nd ACM Symposium on
Parallelism in Algorithms and Architectures, SPAA ’20, page 571–573,
Virtual Event, USA, 2020. Association for Computing Machinery.
[SBV10] Martin Schoeberl, Florian Brandner, and Jan Vitek. RTTM: Real-time
transactional memory. In Proceedings of the 2010 ACM Symposium on
Applied Computing, pages 326–333, 01 2010.
[Sch35] E. Schrödinger. Die gegenwärtige Situation in der Quantenmechanik.

Naturwissenschaften, 23:807–812; 823–828; 844–849, November 1935.
[Sch94] Curt Schimmel. UNIX Systems for Modern Architectures: Symmetric

Multiprocessing and Caching for Kernel Programmers. Addison-Wesley
Longman Publishing Co., Inc., Boston, MA, USA, 1994.
[Sco06] Michael Scott. Programming Language Pragmatics. Morgan Kaufmann,

Burlington, MA, USA, 2006.
v2023.06.11a
936 BIBLIOGRAPHY
[Sco13] Michael L. Scott. Shared-Memory Synchronization. Morgan & Claypool,

San Rafael, CA, USA, 2013.
[Sco15] Michael Scott. Programming Language Pragmatics, 4th Edition. Morgan
Kaufmann, Burlington, MA, USA, 2015.
[Seq88] Sequent Computer Systems, Inc. Guide to Parallel Programming, 1988.
[Sew] Peter Sewell. Relaxed-memory concurrency. Available: https://www.
cl.cam.ac.uk/~pes20/weakmemory/ [Viewed: February 15, 2021].
[Sey12] Justin Seyster. Runtime Verification of Kernel-Level Concurrency Using
Compiler-Based Instrumentation. PhD thesis, Stony Brook University,
2012.
[SF95] Janice M. Stone and Robert P. Fitzgerald. Storage in the PowerPC. IEEE
Micro, 15(2):50–58, April 1995.
[Sha11] Nir Shavit. Data structures in the multicore age. Commun. ACM, 54(3):76–
84, March 2011.
[She06] Gautham R. Shenoy. [patch 4/5] lock_cpu_hotplug: Redesign - light-
weight implementation of lock_cpu_hotplug, October 2006. Avail-
able: https://lkml.org/lkml/2006/10/26/73 [Viewed January 26,
2009].
[SHW11] Daniel J. Sorin, Mark D. Hill, and David A. Wood. A Primer on Memory
Consistency and Cache Coherence. Synthesis Lectures on Computer
Architecture. Morgan & Claypool, 2011.
[Slo10] Lubos Slovak. First steps for utilizing userspace RCU library, July
2010. https://gitlab.labs.nic.cz/knot/knot-dns/commit/
f67acc0178ee9a781d7a63fb041b5d09eb5fb4a2.
[SM95] John D. Slingwine and Paul E. McKenney. Apparatus and method for
achieving reduced overhead mutual exclusion and maintaining coherency
in a multiprocessor system utilizing execution history and thread monitor-
ing. Technical Report US Patent 5,442,758, Assigned to International
Business Machines Corp, Washington, DC, August 1995.
[SM97] John D. Slingwine and Paul E. McKenney. Method for maintaining data
coherency using thread activity summaries in a multicomputer system.
Technical Report US Patent 5,608,893, Assigned to International Business
Machines Corp, Washington, DC, March 1997.
[SM98] John D. Slingwine and Paul E. McKenney. Apparatus and method for
achieving reduced overhead mutual exclusion and maintaining coherency
in a multiprocessor system utilizing execution history and thread monitor-
ing. Technical Report US Patent 5,727,209, Assigned to International
Business Machines Corp, Washington, DC, March 1998.
[SM04a] Dipankar Sarma and Paul E. McKenney. Issues with selected scalabil-
ity features of the 2.6 kernel. In Ottawa Linux Symposium, page 16,
July 2004. https://www.kernel.org/doc/ols/2004/ols2004v2-
pages-195-208.pdf.
v2023.06.11a
BIBLIOGRAPHY 937
[SM04b] Dipankar Sarma and Paul E. McKenney. Making RCU safe for deep
sub-millisecond response realtime applications. In Proceedings of the
2004 USENIX Annual Technical Conference (FREENIX Track), pages
182–191, Boston, MA, USA, June 2004. USENIX Association.
[SM13] Thomas Sewell and Toby Murray. Above and beyond: seL4 noninterfer-
ence and binary verification, May 2013. https://cps-vo.org/node/
7706.
[Smi19] Richard Smith. Working draft, standard for programming language
C++, January 2019. http://www.open-std.org/jtc1/sc22/wg21/
docs/papers/2019/n4800.pdf.
[SMS08] Michael Spear, Maged Michael, and Michael Scott. Inevitability mech-
anisms for software transactional memory. In 3rd ACM SIGPLAN
Workshop on Transactional Computing, Salt Lake City, Utah, February
2008. ACM. Available: http://www.cs.rochester.edu/u/scott/
papers/2008_TRANSACT_inevitability.pdf [Viewed January 10,
2009].
[SNGK17] Dimitrios Siakavaras, Konstantinos Nikas, Georgios Goumas, and Nec-
tarios Koziris. Combining HTM and RCU to implement highly efficient
balanced binary search trees. In 12th ACM SIGPLAN Workshop on
Transactional Computing, Austin, TX, USA, February 2017.
[SPA94] SPARC International. The SPARC Architecture Manual, 1994.
Available: https://sparc.org/wp-content/uploads/2014/01/
SPARCV9.pdf.gz.
[Spi77] Keith R. Spitz. Tell which is which and you’ll be rich, 1977. Inscription
on wall of dungeon.
[Spr01] Manfred Spraul. Re: RFC: patch to allow lock-free traversal of lists with
insertion, October 2001. URL: http://lkml.iu.edu/hypermail/
linux/kernel/0110.1/0410.html.
[Spr08] Manfred Spraul. [RFC, PATCH] state machine based rcu, August
December 8, 2008].
[SR84] Z. Segall and L. Rudolf. Dynamic decentralized cache schemes for
MIMD parallel processors. In 11th Annual International Symposium on
Computer Architecture, pages 340–347, June 1984.
[SRK+ 11] Justin Seyster, Prabakar Radhakrishnan, Samriti Katoch, Abhinav Duggal,
Scott D. Stoller, and Erez Zadok. Redflag: a framework for analysis
of kernel-level concurrency. In Proceedings of the 11th international
conference on Algorithms and architectures for parallel processing -
Volume Part I, ICA3PP’11, pages 66–79, Melbourne, Australia, 2011.
Springer-Verlag.
[SRL90] Lui Sha, Ragunathan Rajkumar, and John P. Lehoczky. Priority in-
heritance protocols: An approach to real-time synchronization. IEEE
Transactions on Computers, 39(9):1175–1185, 1990.
v2023.06.11a
938 BIBLIOGRAPHY
[SS94] Duane Szafron and Jonathan Schaeffer. Experimentally assessing the

usability of parallel programming systems. In IFIP WG10.3 Programming
Environments for Massively Parallel Distributed Systems, pages 19.1–19.7,
Monte Verita, Ascona, Switzerland, 1994.
[SS06] Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free extensible hash
tables. J. ACM, 53(3):379–405, May 2006.
[SSA+ 11] Susmit Sarkar, Peter Sewell, Jade Alglave, Luc Maranget, and Derek
Williams. POWER and ARM litmus tests, 2011. https://www.cl.
cam.ac.uk/~pes20/ppc-supplemental/test6.pdf.
[SSHT93] Janice S. Stone, Harold S. Stone, Philip Heidelberger, and John Turek.
Multiple reservations and the Oklahoma update. IEEE Parallel and
Distributed Technology Systems and Applications, 1(4):58–71, November
1993.
[SSRB00] Douglas C. Schmidt, Michael Stal, Hans Rohnert, and Frank Buschmann.
Pattern-Oriented Software Architecture Volume 2: Patterns for Concur-
rent and Networked Objects. Wiley, Chichester, West Sussex, England,
2000.
[SSVM02] S. Swaminathan, John Stultz, Jack Vogel, and Paul E. McKenney. Fairlocks
– a high performance fair locking scheme. In Proceedings of the 14th
IASTED International Conference on Parallel and Distributed Computing
and Systems, pages 246–251, Cambridge, MA, USA, November 2002.
[ST87] William E. Snaman and David W. Thiel. The VAX/VMS distributed lock
manager. Digital Technical Journal, 5:29–44, September 1987.
[ST95] Nir Shavit and Dan Touitou. Software transactional memory. In Proceed-
ings of the 14th Annual ACM Symposium on Principles of Distributed
Computing, pages 204–213, Ottawa, Ontario, Canada, August 1995.
[Ste92] W. Richard Stevens. Advanced Programming in the UNIX Environment.

Addison Wesley, 1992.
[Ste13] W. Richard Stevens. Advanced Programming in the UNIX Environment,

3rd Edition. Addison Wesley, 2013.
[Sut08] Herb Sutter. Effective concurrency, 2008. Series in Dr. Dobbs Journal.
[Sut13] Adrian Sutton. Concurrent programming with the Disruptor, January 2013.
Presentation at Linux.conf.au 2013, URL: https://www.youtube.
com/watch?v=ItpT_vmRHyI.
[SW95] Richard L. Sites and Richard T. Witek. Alpha AXP Architecture. Digital
Press, second edition, 1995.
[SWS16] Harshal Sheth, Aashish Welling, and Nihar Sheth. Read-copy up-
date in a garbage collected environment, 2016. MIT PRIMES pro-
gram: https://math.mit.edu/research/highschool/primes/
materials/2016/conf/10-1%20Sheth-Welling-Sheth.pdf.
v2023.06.11a
BIBLIOGRAPHY 939
[SZJ12] KC Sivaramakrishnan, Lukasz Ziarek, and Suresh Jagannathan. Eliminat-

ing read barriers through procrastination and cleanliness. In Proceedings
of the 2012 International Symposium on Memory Management, ISMM
’12, pages 49–60, Beĳing, China, 2012. ACM.
[Tal07] Nassim Nicholas Taleb. The Black Swan. Random House, 2007.
[TDV15] Joseph Tassarotti, Derek Dreyer, and Victor Vafeiadis. Verifying read-
copy-update in a logic for weak memory. In Proceedings of the 2015
Proceedings of the 36th annual ACM SIGPLAN conference on Program-
ming Language Design and Implementation, PLDI ’15, pages 110–120,
New York, NY, USA, June 2015. ACM.
[The08] The Open MPI Project. Open MPI, November 2008. Available: http:
//www.open-mpi.org/software/ [Viewed November 26, 2008].
[The11] The Valgrind Developers. Valgrind, November 2011. http://www.

valgrind.org/.
[The12a] The NetBSD Foundation. pserialize(9), October 2012. http://netbsd.

gw.com/cgi-bin/man-cgi?pserialize+9+NetBSD-current.
[The12b] The OProfile Developers. Oprofile, April 2012. http://oprofile.

sourceforge.net.
[TMW11] Josh Triplett, Paul E. McKenney, and Jonathan Walpole. Resizable,

scalable, concurrent hash tables via relativistic programming. In Proceed-
ings of the 2011 USENIX Annual Technical Conference, pages 145–158,
Portland, OR USA, June 2011. The USENIX Association.
[Tor01] Linus Torvalds. Re: [Lse-tech] Re: RFC: patch to allow lock-free traversal
of lists with insertion, October 2001. URL: https://lkml.org/lkml/
2001/10/13/105, https://lkml.org/lkml/2001/10/13/82.
[Tor02] Linus Torvalds. Linux 2.5.43, October 2002. Available: https://

lkml.org/lkml/2002/10/15/425 [Viewed March 30, 2008].
[Tor03] Linus Torvalds. Linux 2.6, August 2003. Available: https://kernel.

org/pub/linux/kernel/v2.6 [Viewed February 16, 2021].
[Tor08] Linus Torvalds. Move ACCESS_ONCE() to <linux/compiler.h>,

May 2008. Git commit: https://git.kernel.org/linus/
9c3cdc1f83a6.
[Tor19] Linus Torvalds. rcu: locking and unlocking need to always be at least
barriers, June 2019. Git commit: https://git.kernel.org/linus/
66be4e66a7f4.
[Tra01] Transaction Processing Performance Council. TPC, 2001. Available:

http://www.tpc.org/ [Viewed December 7, 2008].
[Tre86] R. K. Treiber. Systems programming: Coping with parallelism, April

1986. RJ 5118.
v2023.06.11a
940 BIBLIOGRAPHY
[Tri12] Josh Triplett. Relativistic Causal Ordering: A Memory Model for Scalable
Concurrent Data Structures. PhD thesis, Portland State University, 2012.
[Tri22] Josh Triplett. Spawning processes faster and easier with io_uring, Sep-
tember 2022. https://www.youtube.com/watch?v=_h-kV8AYYqM&
t=4074s.
[TS93] Hiroaki Takada and Ken Sakamura. A bounded spin lock algorithm with
preemption. Technical Report 93-02, University of Tokyo, Tokyo, Japan,
1993.
[TS95] H. Takada and K. Sakamura. Real-time scalability of nested spin locks. In
Proceedings of the 2nd International Workshop on Real-Time Computing
Systems and Applications, RTCSA ’95, pages 160–167, Tokyo, Japan,
1995. IEEE Computer Society.
[Tur37] Alan M. Turing. On computable numbers, with an application to the
entscheidungsproblem. In Proceedings of the London Mathematical
Society, volume 42 of 2, pages 230–265, 1937.
[TZK+ 13] Stephen Tu, Wenting Zheng, Eddie Kohler, Barbara Liskov, and Samuel
Madden. Speedy transactions in multicore in-memory databases. In
Proceedings of the Twenty-Fourth ACM Symposium on Operating Systems
Principles, SOSP ’13, pages 18–32, Farminton, Pennsylvania, 2013.
ACM.
[Ung11] David Ungar. Everything you know (about parallel programming) is
wrong!: A wild screed about the future. In Dynamic Languages Sympo-
sium 2011, Portland, OR, USA, October 2011. Invited talk presentation.
[Uni08a] University of California, Berkeley. BOINC: compute for science, October
2008. Available: http://boinc.berkeley.edu/ [Viewed January 31,
2008].
[Uni08b] University of California, Berkeley. SETI@HOME, December 2008.
Available: http://setiathome.berkeley.edu/ [Viewed January 31,
2008].
[Uni10] University of Maryland. Parallel maze solving, November 2010. URL:
http://www.cs.umd.edu/class/fall2010/cmsc433/p3/ [broken,
February 2021].
[Val95] John D. Valois. Lock-free linked lists using compare-and-swap. In
Proceedings of the Fourteenth Annual ACM Symposium on Principles
of Distributed Computing, PODC ’95, pages 214–222, Ottowa, Ontario,
Canada, 1995. ACM.
[Van18] Michal Vaner. ArcSwap, April 2018. https://crates.io/crates/
arc-swap.
[VBC+ 15] Viktor Vafeiadis, Thibaut Balabonski, Soham Chakraborty, Robin Moris-
set, and Francesco Zappa Nardelli. Common compiler optimisations are
invalid in the c11 memory model and what we can do about it. SIGPLAN
Not., 50(1):209–220, January 2015.
v2023.06.11a
BIBLIOGRAPHY 941
[VGS08] Haris Volos, Neelam Goyal, and Michael M. Swift. Pathological interac-
tion of locks with transactional memory. In 3rd ACM SIGPLAN Workshop
on Transactional Computing, Salt Lake City, Utah, USA, February 2008.
ACM. Available: http://www.cs.wisc.edu/multifacet/papers/
transact08_txlock.pdf [Viewed September 7, 2009].
[Vog09] Werner Vogels. Eventually consistent. Commun. ACM, 52:40–44, January

2009.
[Š11] Jaroslav Ševčík. Safe optimisations for shared-memory concurrent

programs. SIGPLAN Not., 46(6):306–316, June 2011.
[Was14] Scott Wasson. Errata prompts Intel to disable TSX in

Haswell, early Broadwell CPUs, August 2014. https:
//techreport.com/news/26911/errata-prompts-intel-
to-disable-tsx-in-haswell-early-broadwell-cpus/.
[Wav16] Wave Computing, Inc. MIPS®Architecture For Programmers

Volume II-A: The MIPS64®Instruction Set Reference Manual,
2016. URL: https://www.mips.com/downloads/the-mips64-
instruction-set-v6-06/.
[Wei63] J. Weizenbaum. Symmetric list processor. Commun. ACM, 6(9):524–536,

September 1963.
[Wei12] Frédéric Weisbecker. Interruption timer périodique, 2012. http:

//www.dailymotion.com/video/xtxtew_interruption-timer-
periodique-frederic-weisbecker-kernel-recipes-12_tech.
[Wei13] Stewart Weiss. Unix lecture notes, May 2013. Avail-

able: http://www.compsci.hunter.cuny.edu/~sweiss/course_
materials/unix_lecture_notes/ [Viewed April 8, 2014].
[Wei22a] Frédéric Weisbecker. CPU isolation—A practical example (part 5),

January 2022. URL: https://www.suse.com/c/cpu-isolation-
practical-example-part-5/.
[Wei22b] Frédéric Weisbecker. CPU isolation—Full dynticks internals (part 2),

February 2022. URL: https://www.suse.com/c/cpu-isolation-
full-dynticks-part2/.
[Wei22c] Frédéric Weisbecker. CPU isolation—Housekeeping and tradeoffs (part

4), May 2022. URL: https://www.suse.com/c/cpu-isolation-
housekeeping-and-tradeoffs-part-4/.
[Wei22d] Frédéric Weisbecker. CPU isolation—Introduction (part 1), Jan-

uary 2022. URL: https://www.suse.com/c/cpu-isolation-
introduction-part-1/.
[Wei22e] Frédéric Weisbecker. CPU isolation—Nohz_full (part 3), March

2022. URL: https://www.suse.com/c/cpu-isolation-nohz_
full-part-3/.
v2023.06.11a
942 BIBLIOGRAPHY
[Wei22f] Frédéric Weisbecker. CPU isolation—Nohz_full troubleshooting:

broken TSC/clocksource—by SUSE labs (part 6), November 2022.
URL: https://www.suse.com/c/cpu-isolation-nohz_full-
troubleshooting-tsc-clocksource-by-suse-labs-part-6/.
[Wik08] Wikipedia. Zilog Z80, 2008. Available: https://en.wikipedia.

org/wiki/Z80 [Viewed: December 7, 2008].
[Wik12] Wikipedia. Labyrinth, January 2012. https://en.wikipedia.org/

wiki/Labyrinth.
[Wil12] Anthony Williams. C++ Concurrency in Action: Practical Multithread-

ing. Manning, Shelter Island, NY, USA, 2012.
[Wil19] Anthony Williams. C++ Concurrency in Action, 2nd Edition. Manning,

Shelter Island, NY, USA, 2019.
[WKS94] Robert W. Wisniewski, Leonidas Kontothanassis, and Michael L. Scott.

Scalable spin locks for multiprogrammed systems. In 8th IEEE Int’l.
Parallel Processing Symposium, Cancun, Mexico, April 1994. The
Institute of Electrical and Electronics Engineers, Inc.
[Won19] William G. Wong. Vhs or betamax. . . ccix or cxl. . . so many choices,

March 2019. https://www.electronicdesign.com/industrial-
automation/article/21807721/vhs-or-betamaxccix-or-
cxlso-many-choices.
[WTS96] Cai-Dong Wang, Hiroaki Takada, and Ken Sakamura. Priority inheritance
spin locks for multiprocessor real-time systems. In Proceedings of the
2nd International Symposium on Parallel Architectures, Algorithms, and
Networks, ISPAN ’96, pages 70–76, Beĳing, China, 1996. IEEE Computer
Society.
[xen14] xenomai.org. Xenomai, December 2014. URL: http://xenomai.org/.
[Xu10] Herbert Xu. bridge: Add core IGMP snooping support, February
2010. Available: https://marc.info/?t=126719855400006&r=1&
w=2 [Viewed March 20, 2011].
[YHLR13] Richard M. Yoo, Christopher J. Hughes, Konrad Lai, and Ravi Rajwar. Per-
formance evaluation of Intel® Transactional Synchronization Extensions
for high-performance computing. In Proceedings of SC13: International
Conference for High Performance Computing, Networking, Storage and
Analysis, SC ’13, pages 19:1–19:11, Denver, Colorado, 2013. ACM.
[Yod04a] Victor Yodaiken. Against priority inheritance, September 2004. Available:

https://www.yodaiken.com/papers/inherit.pdf [Viewed May
26, 2007].
[Yod04b] Victor Yodaiken. Temporal inventory and real-time synchronization in

RTLinuxPro, September 2004. URL: https://www.yodaiken.com/
papers/sync.pdf.
v2023.06.11a
BIBLIOGRAPHY 943
[Zel11] Cyril Zeller. CUDA C/C++ basics: Supercomputing 2011 tutorial, No-
vember 2011. https://www.nvidia.com/docs/IO/116711/sc11-
cuda-c-basics.pdf.
[Zha89] Lixia Zhang. A New Architecture for Packet Switching Network Protocols.
PhD thesis, Massachusetts Institute of Technology, July 1989.
[Zĳ14] Peter Zĳlstra. Another go at speculative page faults, October 2014.
https://lkml.org/lkml/2014/10/20/620.
v2023.06.11a
944 BIBLIOGRAPHY
v2023.06.11a
If I have seen further it is by standing on the
shoulders of giants.
Isaac Newton, modernized
Credits
LATEX Advisor
Akira Yokosawa is this book’s LATEX advisor, which perhaps most notably includes the
care and feeding of the style guide laid out in Appendix D. This work includes table
layout, listings, fonts, rendering of math, acronyms, bibliography formatting, epigraphs,
hyperlinks, paper size. Akira also perfected the cross-referencing of quick quizzes,
allowing easy and exact navigation between quick quizzes and their answers. He also
added build options that permit quick quizzes to be hidden and to be gathered at the end
of each chapter, textbook style.
This role also includes the build system, which Akira has optimized and made
much more user-friendly. His enhancements have included automating response to
bibliography changes, automatically determining which source files are present, and
automatically generating listings (with automatically generated hyperlinked line-number
references) from the source files.
Reviewers
• Alan Stern (Chapter 15).
• Andy Whitcroft (Section 9.5.2, Section 9.5.3).
• Artem Bityutskiy (Chapter 15, Appendix C).
• Dave Keck (Appendix C).
• David S. Horner (Section 12.1.5).
• Gautham Shenoy (Section 9.5.2, Section 9.5.3).
• “jarkao2”, AKA LWN guest #41960 (Section 9.5.3).
• Jonathan Walpole (Section 9.5.3).
• Josh Triplett (Chapter 12).
• Michael Factor (Section 17.2).
• Mike Fulton (Section 9.5.2).
• Peter Zĳlstra (Section 9.5.4).
• Richard Woodruff (Appendix C).
945
v2023.06.11a
946 CREDITS
• Suparna Bhattacharya (Chapter 12).

• Vara Prasad (Section 12.1.5).
Reviewers whose feedback took the extremely welcome form of a patch are credited
in the git logs.
Machine Owners
Readers might have noticed some graphs showing scalability data out to several hundred
CPUs, courtesy of my current employer, with special thanks to Paul Saab, Yashar
Bayani, Joe Boyd, and Kyle McMartin.
From back in my time at IBM, a great debt of thanks goes to Martin Bligh, who
originated the Advanced Build and Test (ABAT) system at IBM’s Linux Technology
Center, as well as to Andy Whitcroft, Dustin Kirkland, and many others who extended
this system. Many thanks go also to a great number of machine owners: Andrew
Theurer, Andy Whitcroft, Anton Blanchard, Chris McDermott, Cody Schaefer, Darrick
Wong, David “Shaggy” Kleikamp, Jon M. Tollefson, Jose R. Santos, Marvin Heffler,
Nathan Lynch, Nishanth Aravamudan, Tim Pepper, and Tony Breeds.
Original Publications
1. Section 2.4 (“What Makes Parallel Programming Hard?”) on page 19 originally
appeared in a Portland State University Technical Report [MGM+ 09].
2. Section 4.3.4.1 (“Shared-Variable Shenanigans”) on page 63 originally appeared
in Linux Weekly News [ADF+ 19].
3. Section 6.5 (“Retrofitted Parallelism Considered Grossly Sub-Optimal”) on
page 147 originally appeared in 4th USENIX Workshop on Hot Topics on Parallel-
ism [McK12c].
4. Section 9.5.2 (“RCU Fundamentals”) on page 228 originally appeared in Linux
Weekly News [MW07].
5. Section 9.5.3 (“RCU Linux-Kernel API”) on page 238 originally appeared in Linux
Weekly News [McK08e].
6. Section 9.5.4 (“RCU Usage”) on page 251 originally appeared in Linux Weekly
News [McK08g].
7. Section 9.5.5 (“RCU Related Work”) on page 277 originally appeared in Linux
Weekly News [McK14g].
8. Section 9.5.5 (“RCU Related Work”) on page 277 originally appeared in Linux
Weekly News [MP15a].
9. Chapter 12 (“Formal Verification”) on page 357 originally appeared in Linux
Weekly News [McK07f, MR08, McK11d].
10. Section 12.3 (“Axiomatic Approaches”) on page 407 originally appeared in Linux
Weekly News [MS14].
v2023.06.11a
FIGURE CREDITS 947
11. Section 13.5.4 (“Correlated Fields”) on page 438 originally appeared in Oregon
Graduate Institute [McK04].
12. Chapter 15 (“Advanced Synchronization: Memory Ordering”) on page 485
originally appeared in the Linux kernel [HMDZ06].
originally appeared in Linux Weekly News [AMM+ 17a, AMM+ 17b].
originally appeared in ASPLOS ’18 [AMM+ 18].
15. Section 15.3.2 (“Address- and Data-Dependency Difficulties”) on page 524 origi-
nally appeared in the Linux kernel [McK14e].
16. Section 15.5 (“Memory-Barrier Instructions For Specific CPUs”) on page 550
originally appeared in Linux Journal [McK05a, McK05b].
Figure Credits
1. Figure 3.1 (p 26) by Melissa Broussard.
7. Figure 3.8 (p 31) by Melissa Broussard, remixed.
12. Figure 6.1 (p 114) by Kornilios Kourtis.
v2023.06.11a
948 CREDITS
33. Figure 14.15 (p 483) by Sarah McKenney.
34. Figure 14.16 (p 483) by Sarah McKenney.
35. Figure 15.1 (p 487) by Wikipedia user “I, Appaloosa” CC BY-SA 3.0, reformatted.
36. Figure 15.2 (p 487) by Wikipedia user “I, Appaloosa” CC BY-SA 3.0, reformatted.
38. Figure 15.7 (p 497) by Akira Yokosawa.
39. Figure 15.23 (p 556) by Melissa Brossard.
45. Figure 17.5 (p 578) by Melissa Broussard, remixed.
v2023.06.11a
OTHER SUPPORT 949

51. Figure A.1 (p 630) by Melissa Broussard.
52. Figure E.2 (p 750) by Kornilios Kourtis.
Figure 9.33 was adapted from Fedor Pikus’s “When to use RCU” slide [Pik17]. The
discussion of mechanical reference counters in Section 9.2 stemmed from a private
conversation with Dave Regan.
Other Support
We owe thanks to many CPU architects for patiently explaining the instruction- and
memory-reordering features of their CPUs, particularly Wayne Cardoza, Ed Silha,
Anton Blanchard, Tim Slegel, Juergen Probst, Ingo Adlung, Ravi Arimilli, Cathy May,
Derek Williams, H. Peter Anvin, Andy Glew, Leonid Yegoshin, Richard Grisenthwaite,
and Will Deacon. Wayne deserves special thanks for his patience in explaining Alpha’s
reordering of dependent loads, a lesson that Paul resisted quite strenuously!
The bibtex-generation service of the Association for Computing Machinery has saved
us a huge amount of time and effort compiling the bibliography, for which we are
grateful. Thanks are also due to Stamatis Karnouskos, who convinced me to drag
my antique bibliography database kicking and screaming into the 21st century. Any
technical work of this sort owes thanks to the many individuals and organizations that
keep Internet and the World Wide Web up and running, and this one is no exception.
Portions of this material are based upon work supported by the National Science
Foundation under Grant No. CNS-0719851.
v2023.06.11a
950 CREDITS
v2023.06.11a
Acronyms
CAS compare and swap, 35, 43, 57, 72, 404, 423, 605, TLE transactional lock elision, 605, 635, 891
717, 826, 883
TM transactional memory, 891
CBMC C bounded model checker, 280, 413, 617, 818
UTM unbounded transactional memory, 602, 891
EBR epoch-based reclamation, 4, 279, 286, 884
HTM hardware transactional memory, 599, 601, 856,

857, 860, 885
IPI inter-processor interrupt, 219, 554, 683, 885

IRQ interrupt request, 397, 471, 885
KCSAN kernel concurrency sanitizer, 331, 837
LKMM Linux kernel memory consistency model, 409,

536, 555, 837
NBS non-blocking synchronization, 125, 188, 277, 446,

582, 626, 630, 887
NMI non-maskable interrupt, 274, 377, 578, 887
NUCA non-uniform cache architecture, 680, 840, 887
NUMA non-uniform memory architecture, 174, 280, 295,
594, 840, 887
QSBR quiescent-state-based reclamation, 224, 254, 279,

280, 286, 299, 542, 888
RAII resource acquisition is initialization, 178

RCU read-copy update, 4, 219, 857, 884, 885, 888, 889
STM software transactional memory, 602, 857, 890
951
v2023.06.11a
952 Acronyms
v2023.06.11a
Index
Bold: Major reference.
Underline: Definition.
Acquire load, 72, 230, 503, 881 communication, 663, 883 Deadlock free, 446, 884
Ahmed, Iftekhar, 280 write, 663, 891 Desnoyers, Mathieu, 277, 279
Alglave, Jade, 376, 402, 408, 541, 546 Cache-coherence protocol, 664, 882 Dĳkstra, Edsger W., 2, 114
Amdahl’s Law, 9, 128, 154, 881 Cache-invalidation latency, see Latency, Dining philosophers problem, 114
Anti-Heisenbug, see Heisenbug, anti- cache-invalidation Direct-mapped cache, see Cache,
Arbel, Maya, 278, 279 Cache-miss latency, see Latency, direct-mapped
Ash, Mike, 279 cache-miss Dreyer, Derek, 280
Associativity, see Cache associativity Capacity miss, see Cache miss, capacity Dufour, Laurent, 278
Associativity miss, see Cache miss, Chen, Haibo, 279
associativity Chien, Andrew, 5 Efficiency, 12, 126, 136, 183, 638, 884
Atomic, 29, 43, 56–58, 72, 78, 87, 95, 881 Clash free, 446, 883 energy, 39, 347, 884
Atomic read-modify-write operation, 491, Clements, Austin, 278 Embarrassingly parallel, 17, 136, 147, 884
492, 666, 881 Code locking, see Locking, code Epoch-based reclamation (EBR), 279,
Attiya, Hagit, 278, 856 Combinatorial explosion, 883 286, 884
Combinatorial implosion, 883 Exclusive lock, see Lock, exclusive
Belay, Adam, 280 Communication miss, see Cache miss, Existence guarantee, 184, 259, 260, 282,
Bhat, Srivatsa, 279 communication 422, 755, 884
Bonzini, Paolo, 5 Compare and swap (CAS), 35, 43, 57,
Bornat, Richard, 4 404, 423, 605, 717, 826, 883 False sharing, 38, 121, 154, 301, 320, 736,
Bos, Mara, 5 Concurrent, 638, 883 765, 799, 884
Bounded population-oblivious wait free, Consistency Felber, Pascal, 280
see Wait free, bounded memory, 552, 886 Forward-progress guarantee, 192, 279,
population-oblivious process, 888 284, 446, 884
Bounded wait free, see Wait free, bounded sequential, 431, 617, 890 Fragmentation, 146, 885
Butenhof, David R., 5 weak, 557 Fraser, Keir, 279, 884
Corbet, Jonathan, 5 Full memory barrier, see Memory barrier,
C bounded model checker (CBMC), 280, Correia, Andreia, 279 full
413, 617, 818 Critical section, 30, 56, 127, 131, 132, Fully associative cache, see Cache, fully
Cache, 882 140, 174, 184, 883 associative
direct-mapped, 667, 884 RCU read-side, 222, 231, 889
fully associative, 602, 885 read-side, 176, 214, 889 Generality, 11, 14, 42, 126
Cache associativity, 602, 662, 881 write-side, 892 Giannoula, Christina, 280, 594
Cache coherence, 505, 555, 602, 882 Gotsman, Alexey, 280
Cache geometry, 662, 882 Data locking, see Locking, data Grace period, 223, 239, 284, 297, 331,
Cache line, 34, 80, 182, 320, 489, 508, Data race, 49, 63, 159, 332, 524, 884 376, 410, 428, 474, 539, 581, 642,
553, 600, 662, 882 Deacon, Will, 64 885
Cache miss, 882 Deadlock, 9, 21, 119, 161, 224, 309, 475, Grace-period latency, see Latency,
associativity, 663, 881 526, 571, 589, 604, 884 grace-period
capacity, 662, 883 Deadlock cycle, 642, 643 Groce, Alex, 280
953
v2023.06.11a
954 INDEX
Hardware transactional memory (HTM), Liu, Ran, 279 Non-maskable interrupt (NMI), 274, 377,
599, 601, 856, 857, 860, 885 Liu, Yujie, 279 578, 887
Harris, Timothy, 278, 279 Livelock, 9, 21, 161, 171, 362, 605, 762, Non-uniform cache architecture (NUCA),
Hawking, Stephen, 10 886 680, 840, 887
Hazard pointer, 207, 227, 236, 281, 295, Lock, 886 Non-uniform memory architecture
321, 429, 481, 583, 607, 755, 885 exclusive, 53, 175, 634, 884 (NUMA), 174, 280, 295, 594, 840,
Heisenberg, Weiner, 304, 341 reader-writer, 53, 176, 279, 889 887
Heisenbug, 341, 885 sequence, 890 NUMA node, 21, 814, 887
anti-, 341 Lock contention, 88, 109, 121, 127, 132,
Hennessy, John L., 5, 25 142, 172, 886 Obstruction free, 446, 887
Herlihy, Maurice P., 4 Lock free, 279, 446, 886 Overhead, 10, 33, 888
Hot spot, 135, 302, 885 Locking, 159 memory-barrier, 30
Howard, Phil, 277 code, 128, 131, 142, 883
Howlett, Liam, 278 data, 21, 128, 146, 883 Parallel, 638, 888
Hraska, Adam, 279 Luchangco, Victor, 4, 279 Park, SeongJae, 280, 594
Humiliatingly parallel, 152, 885 Patterson, David A., 5, 25
Hunter, Andrew, 280 Madden, Samuel, 278 Pawan, Pankaj, 402
Mao, Yandong, 278 Penyaev, Roman, 410
Immutable, 885 Maranget, Luc, 402 Performance, 11, 126, 638, 888
Inter-processor interrupt (IPI), 219, 554, Marked access, 886 Pikus, Fedor, 949
683, 885 Marlier, Patrick, 280 Pipelined CPU, 888
Interrupt request (IRQ), 397, 471, 885 Matloff, Norm, 5 Plain access, 63, 75, 229, 523, 888
Invalidation, 663, 674, 856, 885 Mattson, Timothy G., 5 Podzimek, Andrej, 279
Matveev, Alexander, 280 Process consistency, see Consistency,
Jensen, Carlos, 280 McKenney, Paul E., 280 process
Melham, Tom, 280 Productivity, 11, 13, 126, 478, 592
Kaashoek, Frans, 278 Memory, 886 Program order, 888
Kernel concurrency sanitizer (KCSAN), Memory barrier, 30, 57, 127, 174, 210, Promela, 358, 818
331, 837 283, 299, 364, 423, 490, 578, 635,
Kim, Jaeho, 280 643, 652, 661, 886 Quiescent state, 223, 398, 589, 656, 888
Knuth, Donald, 4, 277, 583 full, 219, 490, 534, 549, 551, 838 Quiescent-state-based reclamation
Kogan, Alex, 280 read, 528, 551, 678, 889 (QSBR), 224, 254, 280, 286, 299,
Kohler, Eddie, 278 write, 551, 678, 891 542, 888
Kokologiannakis, Michalis, 280 Memory consistency, see Consistency,
Kroah-Hartman, Greg, 5 memory Race condition, 9, 185, 342, 357, 358,
Kroening, Daniel, 280 Memory latency, see Latency, memory 437, 494, 645, 888
Kung, H. T., 4, 277 Memory-barrier latency, see Latency, Ramalhete, Pedro, 279
memory-barrier RCU read-side critical section, see
Latency, 28, 38, 461, 885 Memory-barrier overhead, see Overhead, Critical section, RCU read-side
cache-invalidation, 675 memory-barrier RCU-protected data, 779, 888
cache-miss, 39 MESI protocol, 664, 887 RCU-protected pointer, 222, 889
grace-period, 239, 653 Message latency, see Latency, message Read memory barrier, see Memory
memory, 579 Moore’s Law, 10, 12, 18, 25, 28, 39, 42, barrier, read
memory-barrier, 299 129, 575, 578, 579, 887 Read mostly, 889
message, 127 Morris, Robert, 278 Read only, 889
scheduling, 450 Morrison, Adam, 279 Read-copy update (RCU), 219, 857, 889
Lea, Doug, 5 Mutual-exclusion mechanism, 887 Read-side critical section, see Critical
Lehman, Philip L., 4, 277 section, read-side
Lespinasse, Michel, 278 Nardelli, Francesco Zappa, 402 Reader-writer lock, see Lock,
Liang, Lihao, 280 Nidhugg, 414, 617, 818 reader-writer
Linearizable, 278, 446, 797, 886 Non-blocking, 887 Real time, 889
Linux kernel memory consistency model Non-blocking synchronization (NBS), Reference count, 72, 77, 204, 272, 281,
(LKMM), 409, 536, 555, 837 125, 188, 277, 446, 582, 626, 630, 422, 439, 594, 641, 730, 889
Liskov, Barbara, 278 887 Regan, Dave, 949
v2023.06.11a
INDEX 955
Reinders, James, 5 Spear, Michael, 4, 279 Vector CPU, 891

Release store, 72, 529, 890 Spin, 358
Reliability, 478 Starvation, 114, 159, 171, 182, 218, 431, Wait free, 446, 891
Resource acquisition is initialization 473, 594, 644, 762, 890 bounded, 284, 446, 881
(RAII), 178 Starvation free, 446, 890 bounded population-oblivious, 446,
Rinetzky, Noam, 280 Stevens, W. Richard, 5 882
Romer, Geoff, 280 Store buffer, 890 Walpole, Jon, 277
Roy, Lance, 280 Store forwarding, 890 Weak consistency, see Consistency, weak
Rubini, Alessandro, 5 Superscalar CPU, 890 Weisbecker, Frédéric, 467
Sutter, Herb, 5 Weiss, Stewart, 5
Sagonas, Konstantinos, 280 Synchronization, 890 Weizenbaum, Joseph, 277
Sarkar, Susmit, 402 Wilcox, Matthew, 278
Scalability, 12, 638, 890 Tassarotti, Joseph, 280 Williams, Anthony, 5
Scheduling latency, see Latency, Teachable, 891 Williams, Derek, 402
scheduling Throughput, 891 Write memory barrier, see Memory
Schimmel, Curt, 5 Torvalds, Linus, 732, 792, 848 barrier, write
Schmidt, Douglas C., 4 Transactional lock elision (TLE), 605, Write miss, see Cache miss, write
Scott, Michael, 4 635, 891 Write mostly, 891
Sequence lock, see Lock, sequence Transactional memory (TM), 891 Write-side critical section, see Critical
Sequential consistency, see Consistency, Triplett, Josh, 277 section, write-side
sequential Tu, Stephen, 278
Type-safe memory, 241, 259, 423, 891 Xu, Herbert, 277, 304, 792
Sewell, Peter, 402
Shavit, Nir, 4, 280 Unbounded transactional memory (UTM),
Shenoy, Gautham, 279 602, 891 Yang, Hongseok, 280
Siakavaras, Dimitrios, 280, 594 Uncertainty principle, 341
Sivaramakrishnan, KC, 279 Unfairness, 159, 172, 182, 218, 891 Zeldovich, Nickolai, 278
Software transactional memory (STM), Unteachable, 891 Zhang, Heng, 279
602, 857, 890 Zheng, Wenting, 278
Sorin, Daniel, 5 Vafeiadis, Viktor, 280 Zĳlstra, Peter, 278
v2023.06.11a
956 INDEX
v2023.06.11a
API Index
(c): Cxx standard, (g): GCC extension, (k): Linux kernel,
(kh): Linux kernel historic, (pf): perfbook CodeSamples,
(px): POSIX, (ur): userspace RCU.
_Thread_local (c), 58, 73, 83 atomic_compare_exchange_ DECLARE_PER_THREAD() (pf), 73

__ATOMIC_ACQUIRE (g), 58 strong() (c), 57 DEFINE_PER_CPU() (k), 73
__ATOMIC_ACQ_REL (g), 58 atomic_compare_exchange_ DEFINE_PER_THREAD() (pf), 73
__ATOMIC_CONSUME (g), 58 weak() (c), 57 destroy_rcu_head() (k), 250
__ATOMIC_RELAXED (g), 58 atomic_dec() (k), 72 destroy_rcu_head_on_stack() (k),
__ATOMIC_RELEASE (g), 58 atomic_dec_and_test() (k), 72 250
__ATOMIC_SEQ_CST (g), 58 atomic_exchange() (c), 57
__atomic_load() (g), 58 atomic_fetch_add() (c), 57 exec() (px), 190
__atomic_load_n() (g), 57, 58 atomic_fetch_and() (c), 57 exit() (px), 46
__atomic_store() (g), 58 atomic_fetch_sub() (c), 57
__atomic_store_n() (g), 57, 58 atomic_fetch_xor() (c), 57 for_each_running_thread() (pf), 60
__atomic_thread_fence() (g), 58 atomic_inc() (k), 72 for_each_thread() (pf), 60, 82
__get_thread_var() (pf), 73, 81 atomic_inc_not_zero() (k), 72 fork() (px), 46, 61, 74, 190, 726, 729
__sync_add_and_fetch() (g), 56 atomic_load() (c), 57
__sync_and_and_fetch() (g), 56 atomic_load_explicit() (c), 58 get_nulls_value() (k), 245
__sync_bool_compare_and_ atomic_read() (k), 72
swap() (g), 57 atomic_set() (k), 72 hlist_del_rcu() (k), 251
__sync_fetch_and_add() (g), 56, 724, atomic_signal_fence() (c), 57 hlist_for_each_entry_rcu() (k),
726 atomic_store() (c), 57 251
__sync_fetch_and_and() (g), 56 atomic_sub() (k), 72
init_per_thread() (pf), 73
__sync_fetch_and_nand() (g), 56, 725 atomic_sub_and_test() (k), 72
init_rcu_head() (k), 250
__sync_fetch_and_or() (g), 56 atomic_t (k), 72, 95
init_rcu_head_on_stack() (k), 250
__sync_fetch_and_sub() (g), 56, 726 atomic_thread_fence() (c), 57
is_a_nulls() (k), 245
__sync_fetch_and_xor() (g), 56, 726 atomic_xchg() (k), 72
__sync_nand_and_fetch() (g), 56, 725
kfree() (k), 249
__sync_or_and_fetch() (g), 56 barrier() (k), 57, 67, 70 kill() (px), 46
__sync_sub_and_fetch() (g), 56
kmem_cache_create() (k), 241
__sync_synchronize() (g), 57 call_rcu() (k), 168, 222, 241 kthread_create() (k), 59
__sync_val_compare_and_swap() (g), call_rcu_tasks() (k), 242 kthread_should_stop() (k), 59
57 call_srcu() (k), 241 kthread_stop() (k), 59
__sync_xor_and_fetch() (g), 56 cds_list_add() (ur), 297
__thread (g), 55, 58, 73, 83, 724 cds_list_add_rcu() (ur), 297 list_add_rcu() (k), 251
cds_list_del_init() (ur), 297 list_for_each_entry_rcu() (k), 251
ACCESS_ONCE() (kh), 57, 726 cds_list_del_rcu() (ur), 297 list_replace_rcu() (k), 246, 249
atomic_add() (k), 72 cds_list_for_each_entry() (ur), 296 lockless_dereference() (kh), 502,
atomic_add_return() (k), 72 cds_list_for_each_entry_ 841
atomic_add_unless() (k), 72 rcu() (ur), 296
atomic_cmpxchg() (k), 72, 98 create_thread() (pf), 59, 60 NR_THREADS (pf), 60
957
v2023.06.11a
958 API INDEX
per_cpu() (k), 73 rcu_head (k), 250 smp_thread_id() (pf), 60, 728

per_thread() (pf), 73, 82 rcu_head_after_call_rcu() (k), 250 smp_wmb() (k), 67, 68
pthread_atfork() (px), 190 rcu_head_init() (k), 250 spin_lock() (k), 61, 62
pthread_cond_wait() (px), 168 rcu_init() (ur), 59 spin_lock_init() (k), 61
pthread_create() (px), 48 RCU_INIT_POINTER() (k), 244 spin_trylock() (k), 62, 169
pthread_exit() (px), 48, 49 rcu_is_watching() (k), 250 spin_unlock() (k), 62
pthread_getspecific() (px), 58 RCU_LOCKDEP_WARN() (k), 250 spinlock_t (k), 61
pthread_join() (px), 49 RCU_NONIDLE() (k), 250 srcu_barrier() (k), 242
pthread_key_create() (px), 58 rcu_pointer_handoff() (k), 244 srcu_read_lock() (k), 241
pthread_key_delete() (px), 58 RCU_POINTER_INITIALIZER() (k), 244 srcu_read_lock_held() (k), 250
pthread_kill() (px), 106 rcu_read_lock() (k), 222, 239 srcu_read_unlock() (k), 241
pthread_mutex_init() (px), 50 rcu_read_lock_bh() (k), 239 srcu_struct (k), 241
PTHREAD_MUTEX_INITIALIZER (px), 50 rcu_read_lock_bh_held() (k), 250 struct task_struct (k), 59
pthread_mutex_lock() (px), 50, 170 rcu_read_lock_held() (k), 250 synchronize_irq() (k), 784
pthread_mutex_t (px), 50, 53, 168 rcu_read_lock_sched() (k), 239 synchronize_net() (k), 239
pthread_mutex_unlock() (px), 50 rcu_read_lock_sched_held() (k), synchronize_rcu() (k), 222, 239
pthread_rwlock_init() (px), 53 250 synchronize_rcu_expedited() (k),
PTHREAD_RWLOCK_INITIALIZER (px), 53 rcu_read_unlock() (k), 222, 239 239
pthread_rwlock_rdlock() (px), 53 rcu_read_unlock_bh() (k), 239 synchronize_rcu_tasks() (k), 242
pthread_rwlock_t (px), 53 rcu_read_unlock_sched() (k), 239 synchronize_srcu() (k), 241
pthread_rwlock_unlock() (px), 53 rcu_register_thread() (ur), 59 synchronize_srcu_expedited() (k),
pthread_rwlock_wrlock() (px), 53 rcu_replace_pointer() (k), 243 241
pthread_setspecific() (px), 58 rcu_sleep_check() (k), 250
pthread_t (px), 59 this_cpu_ptr() (k), 73
rcu_unregister_thread() (ur), 59
thread_id_t (pf), 59, 60
rcu_access_pointer() (k), 244 READ_ONCE() (k), 50, 55, 57, 58, 65, 66,
rcu_assign_pointer() (k), 222, 243 69, 71, 72, 722, 724, 726
unlikely() (k), 67
rcu_barrier() (k), 241
rcu_barrier_tasks() (k), 242 schedule() (k), 242 vfork() (px), 74, 729
rcu_cpu_stall_reset() (k), 250 schedule_timeout_ volatile (c), 68–70, 75
rcu_dereference() (k), 222, 244 interruptible() (k), 59
rcu_dereference_check() (k), 244 sig_atomic_t (c), 66 wait() (px), 47, 49, 61, 74, 726
rcu_dereference_protected() (k), SLAB_TYPESAFE_BY_RCU (k), 241 wait_all_threads() (pf), 60
244 smp_init() (pf), 59 wait_thread() (pf), 60
rcu_dereference_raw() (k), 244 smp_load_acquire() (k), 72, 728 waitall() (px), 47
rcu_dereference_raw_notrace() (k), smp_mb() (k), 70 WRITE_ONCE() (k), 50, 57, 64, 66, 69–72,
244 smp_store_release() (k), 68, 72, 728 722, 726, 727
v2023.06.11a

Perfbook-1c 2023 06 11a

Uploaded by

Copyright:

Available Formats

Perfbook-1c 2023 06 11a

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Perfbook-1c 2023 06 11a

Uploaded by

Copyright:

Available Formats

Is Parallel Programming Hard, And, If So,

What Can You Do About It?

June 11, 2023

1 How To Use This Book 1

3 Hardware and its Habits 25

3.2.3 Hardware Optimizations . . . . . . . . . . . . . . . . . . . . 38

4 Tools of the Trade 45

5.5 Parallel Counting Discussion . . . . . . . . . . . . . . . . . . . . . . 108

6 Partitioning and Synchronization Design 113

7.5.2 Locking For Parallel Libraries: Just Another Tool . . . . . . 187

8 Data Ownership 195

9 Deferred Processing 201

10 Data Structures 289

10.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321

12 Formal Verification 357

13 Putting It All Together 421

14 Advanced Synchronization 445

15 Advanced Synchronization: Memory Ordering 485

15.2.5 Control Dependencies . . . . . . . . . . . . . . . . . . . . . 504

16 Ease of Use 569

17 Conflicting Visions of the Future 575

17.3.5 Potential Game Changers . . . . . . . . . . . . . . . . . . . 612

18 Looking Forward and Back 625

A Important Questions 629

B “Toy” RCU Implementations 641

C Why Memory Barriers? 661

C.3.3 Store Buffers and Memory Barriers . . . . . . . . . . . . . . 671

D Style Guide 685

E Answers to Quick Quizzes 705

API Index 957

How To Use This Book

Cat: Where are you going?

1.2 Quick Quizzes

1.3 Alternatives to This Book

1. If you prefer a more academic and rigorous treatment of parallel programming,

2. If you would like an academic treatment of parallel programming from a program-

3 Albeit an implementation that contains a reader-preemption bug noted by Richard

Listing 1.1: Creating an Up-To-Date PDF

1.4 Sample Source Code

Use the source, Luke!

find CodeSamples -name rcu_rcpls.c -print

1.5 Whose Book Is This?

If you become a teacher, by your pupils you’ll be

Listing 1.2: Generating an Updated PDF

Signed-off-by: My Name <myname@example.org>

Please see https://lkml.org/lkml/2007/1/15/219 for an example patch with

2.1 Historic Parallel Programming Difficulties

Not the power to remember, but its very opposite,

1. The historic high cost and relative rarity of parallel systems.

3. The paucity of publicly accessible parallel code.

4. The lack of a widely understood engineering discipline of parallel programming.

5. The high overhead of communication relative to that of processing, even in tightly

However, even though parallel programming might not be as hard as is commonly

It therefore makes sense to consider alternatives to parallel programming. However,

2.2 Parallel Programming Goals

Quick Quiz 2.6: What about just having fun?

Each of these goals is elaborated upon in the following sections.

1 Kudos to Michael Wong for naming the iron triangle.

CPU Clock Frequency / MIPS