Fix corruption when relation truncation fails.

macdice · macdice · commit 2280912165d6 · 2024-12-20T23:58:08.000+13:00
RelationTruncate() does three things, while holding an AccessExclusiveLock and preventing checkpoints: 1. Logs the truncation. 2. Drops buffers, even if they're dirty. 3. Truncates some number of files. Step 2 could previously be canceled if it had to wait for I/O, and step 3 could and still can fail in file APIs. All orderings of these operations have data corruption hazards if interrupted, so we can't give up until the whole operation is done. When dirty pages were discarded but the corresponding blocks were left on disk due to ERROR, old page versions could come back from disk, reviving deleted data (see pgsql-bugs #18146 and several like it). When primary and standby were allowed to disagree on relation size, standbys could panic (see pgsql-bugs #18426) or revive data unknown to visibility management on the primary (theorized). Changes: * WAL is now unconditionally flushed first * smgrtruncate() is now called in a critical section, preventing interrupts and causing PANIC on file API failure * smgrtruncate() has a new parameter for existing fork sizes, because it can't call smgrnblocks() itself inside a critical section The changes apply to RelationTruncate(), smgr_redo() and pg_truncate_visibility_map(). That last is also brought up to date with other evolutions of the truncation protocol. The VACUUM FileTruncate() failure mode had been discussed in older reports than the ones referenced below, with independent analysis from many people, but earlier theories on how to fix it were too complicated to back-patch. The more recently invented cancellation bug was diagnosed by Alexander Lakhin. Other corruption scenarios were spotted by me while iterating on this patch and earlier commit 75818b3. Back-patch to all supported releases. Reviewed-by: Michael Paquier <michael@paquier.xyz> Reviewed-by: Robert Haas <robertmhaas@gmail.com> Reported-by: rootcause000@gmail.com Reported-by: Alexander Lakhin <exclusion@gmail.com> Discussion: https://postgr.es/m/18146-04e908c662113ad5%40postgresql.org Discussion: https://postgr.es/m/18426-2d18da6586f152d6%40postgresql.org
diff --git a/contrib/pg_visibility/pg_visibility.c b/contrib/pg_visibility/pg_visibility.c
@@ -18,6 +18,7 @@
 #include "funcapi.h"
 #include "miscadmin.h"
 #include "storage/bufmgr.h"
+#include "storage/proc.h"
 #include "storage/procarray.h"
 #include "storage/smgr.h"
 #include "utils/rel.h"
@@ -385,6 +386,7 @@ pg_truncate_visibility_map(PG_FUNCTION_ARGS)
 	Relation	rel;
 	ForkNumber	fork;
 	BlockNumber block;
+	BlockNumber old_block;
 
 	rel = relation_open(relid, AccessExclusiveLock);
 
@@ -394,15 +396,24 @@ pg_truncate_visibility_map(PG_FUNCTION_ARGS)
 	/* Forcibly reset cached file size */
 	RelationGetSmgr(rel)->smgr_vm_nblocks = InvalidBlockNumber;
 
+	/* Compute new and old size before entering critical section. */
+	fork = VISIBILITYMAP_FORKNUM;
 	block = visibilitymap_prepare_truncate(rel, 0);
-	if (BlockNumberIsValid(block))
-	{
-		fork = VISIBILITYMAP_FORKNUM;
-		smgrtruncate(RelationGetSmgr(rel), &fork, 1, &block);
-	}
+	old_block = BlockNumberIsValid(block) ? smgrnblocks(RelationGetSmgr(rel), fork) : 0;
+
+	/*
+	 * WAL-logging, buffer dropping, file truncation must be atomic and all on
+	 * one side of a checkpoint.  See RelationTruncate() for discussion.
+	 */
+	Assert(!MyProc->delayChkpt);
+	MyProc->delayChkpt = true;
+	Assert(!MyProc->delayChkptEnd);
+	MyProc->delayChkptEnd = true;
+	START_CRIT_SECTION();
 
 	if (RelationNeedsWAL(rel))
 	{
+		XLogRecPtr	lsn;
 		xl_smgr_truncate xlrec;
 
 		xlrec.blkno = 0;
@@ -412,9 +423,18 @@ pg_truncate_visibility_map(PG_FUNCTION_ARGS)
 		XLogBeginInsert();
 		XLogRegisterData((char *) &xlrec, sizeof(xlrec));
 
-		XLogInsert(RM_SMGR_ID, XLOG_SMGR_TRUNCATE | XLR_SPECIAL_REL_UPDATE);
+		lsn = XLogInsert(RM_SMGR_ID,
+						 XLOG_SMGR_TRUNCATE | XLR_SPECIAL_REL_UPDATE);
+		XLogFlush(lsn);
 	}
 
+	if (BlockNumberIsValid(block))
+		smgrtruncate(RelationGetSmgr(rel), &fork, 1, &old_block, &block);
+
+	END_CRIT_SECTION();
+	MyProc->delayChkpt = false;
+	MyProc->delayChkptEnd = false;
+
 	/*
 	 * Release the lock right away, not at commit time.
 	 *
diff --git a/src/backend/catalog/storage.c b/src/backend/catalog/storage.c
@@ -280,6 +280,7 @@ RelationTruncate(Relation rel, BlockNumber nblocks)
 	bool		vm;
 	bool		need_fsm_vacuum = false;
 	ForkNumber	forks[MAX_FORKNUM];
+	BlockNumber old_blocks[MAX_FORKNUM];
 	BlockNumber blocks[MAX_FORKNUM];
 	int			nforks = 0;
 	SMgrRelation reln;
@@ -295,6 +296,7 @@ RelationTruncate(Relation rel, BlockNumber nblocks)
 
 	/* Prepare for truncation of MAIN fork of the relation */
 	forks[nforks] = MAIN_FORKNUM;
+	old_blocks[nforks] = smgrnblocks(reln, MAIN_FORKNUM);
 	blocks[nforks] = nblocks;
 	nforks++;
 
@@ -306,6 +308,7 @@ RelationTruncate(Relation rel, BlockNumber nblocks)
 		if (BlockNumberIsValid(blocks[nforks]))
 		{
 			forks[nforks] = FSM_FORKNUM;
+			old_blocks[nforks] = smgrnblocks(reln, FSM_FORKNUM);
 			nforks++;
 			need_fsm_vacuum = true;
 		}
@@ -319,6 +322,7 @@ RelationTruncate(Relation rel, BlockNumber nblocks)
 		if (BlockNumberIsValid(blocks[nforks]))
 		{
 			forks[nforks] = VISIBILITYMAP_FORKNUM;
+			old_blocks[nforks] = smgrnblocks(reln, VISIBILITYMAP_FORKNUM);
 			nforks++;
 		}
 	}
@@ -357,14 +361,20 @@ RelationTruncate(Relation rel, BlockNumber nblocks)
 	MyProc->delayChkptEnd = true;		/* DELAY_CHKPT_COMPLETE */
 
 	/*
-	 * We WAL-log the truncation before actually truncating, which means
-	 * trouble if the truncation fails. If we then crash, the WAL replay
-	 * likely isn't going to succeed in the truncation either, and cause a
-	 * PANIC. It's tempting to put a critical section here, but that cure
-	 * would be worse than the disease. It would turn a usually harmless
-	 * failure to truncate, that might spell trouble at WAL replay, into a
-	 * certain PANIC.
+	 * We WAL-log the truncation first and then truncate in a critical
+	 * section. Truncation drops buffers, even if dirty, and then truncates
+	 * disk files. All of that work needs to complete before the lock is
+	 * released, or else old versions of pages on disk that are missing recent
+	 * changes would become accessible again.  We'll try the whole operation
+	 * again in crash recovery if we panic, but even then we can't give up
+	 * because we don't want standbys' relation sizes to diverge and break
+	 * replay or visibility invariants downstream.  The critical section also
+	 * suppresses interrupts.
+	 *
+	 * (See also pg_visibilitymap.c if changing this code.)
 	 */
+	START_CRIT_SECTION();
+
 	if (RelationNeedsWAL(rel))
 	{
 		/*
@@ -388,18 +398,20 @@ RelationTruncate(Relation rel, BlockNumber nblocks)
 		 * hit the disk before the WAL record, and the truncation of the FSM
 		 * or visibility map. If we crashed during that window, we'd be left
 		 * with a truncated heap, but the FSM or visibility map would still
-		 * contain entries for the non-existent heap pages.
+		 * contain entries for the non-existent heap pages, and standbys would
+		 * also never replay the truncation.
 		 */
-		if (fsm || vm)
-			XLogFlush(lsn);
+		XLogFlush(lsn);
 	}
 
 	/*
 	 * This will first remove any buffers from the buffer pool that should no
 	 * longer exist after truncation is complete, and then truncate the
 	 * corresponding files on disk.
 	 */
-	smgrtruncate(RelationGetSmgr(rel), forks, nforks, blocks);
+	smgrtruncate(RelationGetSmgr(rel), forks, nforks, old_blocks, blocks);
+
+	END_CRIT_SECTION();
 
 	/* We've done all the critical work, so checkpoints are OK now. */
 	MyProc->delayChkpt = false;
@@ -984,6 +996,7 @@ smgr_redo(XLogReaderState *record)
 		Relation	rel;
 		ForkNumber	forks[MAX_FORKNUM];
 		BlockNumber blocks[MAX_FORKNUM];
+		BlockNumber old_blocks[MAX_FORKNUM];
 		int			nforks = 0;
 		bool		need_fsm_vacuum = false;
 
@@ -1018,6 +1031,7 @@ smgr_redo(XLogReaderState *record)
 		if ((xlrec->flags & SMGR_TRUNCATE_HEAP) != 0)
 		{
 			forks[nforks] = MAIN_FORKNUM;
+			old_blocks[nforks] = smgrnblocks(reln, MAIN_FORKNUM);
 			blocks[nforks] = xlrec->blkno;
 			nforks++;
 
@@ -1035,6 +1049,7 @@ smgr_redo(XLogReaderState *record)
 			if (BlockNumberIsValid(blocks[nforks]))
 			{
 				forks[nforks] = FSM_FORKNUM;
+				old_blocks[nforks] = smgrnblocks(reln, FSM_FORKNUM);
 				nforks++;
 				need_fsm_vacuum = true;
 			}
@@ -1046,13 +1061,18 @@ smgr_redo(XLogReaderState *record)
 			if (BlockNumberIsValid(blocks[nforks]))
 			{
 				forks[nforks] = VISIBILITYMAP_FORKNUM;
+				old_blocks[nforks] = smgrnblocks(reln, VISIBILITYMAP_FORKNUM);
 				nforks++;
 			}
 		}
 
 		/* Do the real work to truncate relation forks */
 		if (nforks > 0)
-			smgrtruncate(reln, forks, nforks, blocks);
+		{
+			START_CRIT_SECTION();
+			smgrtruncate(reln, forks, nforks, old_blocks, blocks);
+			END_CRIT_SECTION();
+		}
 
 		/*
 		 * Update upper-level FSM pages to account for the truncation. This is
diff --git a/src/backend/storage/smgr/md.c b/src/backend/storage/smgr/md.c
@@ -827,19 +827,21 @@ mdnblocks(SMgrRelation reln, ForkNumber forknum)
 
 /*
  *	mdtruncate() -- Truncate relation to specified number of blocks.
+ *
+ * Guaranteed not to allocate memory, so it can be used in a critical section.
+ * Caller must have called smgrnblocks() to obtain curnblk while holding a
+ * sufficient lock to prevent a change in relation size, and not used any smgr
+ * functions for this relation or handled interrupts in between.  This makes
+ * sure we have opened all active segments, so that truncate loop will get
+ * them all!
  */
 void
-mdtruncate(SMgrRelation reln, ForkNumber forknum, BlockNumber nblocks)
+mdtruncate(SMgrRelation reln, ForkNumber forknum,
+		   BlockNumber curnblk, BlockNumber nblocks)
 {
-	BlockNumber curnblk;
 	BlockNumber priorblocks;
 	int			curopensegs;
 
-	/*
-	 * NOTE: mdnblocks makes sure we have opened all active segments, so that
-	 * truncation loop will get them all!
-	 */
-	curnblk = mdnblocks(reln, forknum);
 	if (nblocks > curnblk)
 	{
 		/* Bogus request ... but no complaint if InRecovery */
@@ -1108,7 +1110,7 @@ _fdvec_resize(SMgrRelation reln,
 		reln->md_seg_fds[forknum] =
 			MemoryContextAlloc(MdCxt, sizeof(MdfdVec) * nseg);
 	}
-	else
+	else if (nseg > reln->md_num_open_segs[forknum])
 	{
 		/*
 		 * It doesn't seem worthwhile complicating the code to amortize
@@ -1120,6 +1122,16 @@ _fdvec_resize(SMgrRelation reln,
 			repalloc(reln->md_seg_fds[forknum],
 					 sizeof(MdfdVec) * nseg);
 	}
+	else
+	{
+		/*
+		 * We don't reallocate a smaller array, because we want mdtruncate()
+		 * to be able to promise that it won't allocate memory, so that it is
+		 * allowed in a critical section.  This means that a bit of space in
+		 * the array is now wasted, until the next time we add a segment and
+		 * reallocate.
+		 */
+	}
 
 	reln->md_num_open_segs[forknum] = nseg;
 }
diff --git a/src/backend/storage/smgr/smgr.c b/src/backend/storage/smgr/smgr.c
@@ -59,7 +59,7 @@ typedef struct f_smgr
 								   BlockNumber blocknum, BlockNumber nblocks);
 	BlockNumber (*smgr_nblocks) (SMgrRelation reln, ForkNumber forknum);
 	void		(*smgr_truncate) (SMgrRelation reln, ForkNumber forknum,
-								  BlockNumber nblocks);
+								  BlockNumber old_blocks, BlockNumber nblocks);
 	void		(*smgr_immedsync) (SMgrRelation reln, ForkNumber forknum);
 } f_smgr;
 
@@ -548,10 +548,15 @@ smgrnblocks(SMgrRelation reln, ForkNumber forknum)
  *
  * The caller must hold AccessExclusiveLock on the relation, to ensure that
  * other backends receive the smgr invalidation event that this function sends
- * before they access any forks of the relation again.
+ * before they access any forks of the relation again.  The current size of
+ * the forks should be provided in old_nblocks.  This function should normally
+ * be called in a critical section, but the current size must be checked
+ * outside the critical section, and no interrupts or smgr functions relating
+ * to this relation should be called in between.
  */
 void
-smgrtruncate(SMgrRelation reln, ForkNumber *forknum, int nforks, BlockNumber *nblocks)
+smgrtruncate(SMgrRelation reln, ForkNumber *forknum, int nforks,
+			 BlockNumber *old_nblocks, BlockNumber *nblocks)
 {
 	int			i;
 
@@ -576,7 +581,8 @@ smgrtruncate(SMgrRelation reln, ForkNumber *forknum, int nforks, BlockNumber *nb
 	/* Do the truncation */
 	for (i = 0; i < nforks; i++)
 	{
-		smgrsw[reln->smgr_which].smgr_truncate(reln, forknum[i], nblocks[i]);
+		smgrsw[reln->smgr_which].smgr_truncate(reln, forknum[i],
+											   old_nblocks[i], nblocks[i]);
 
 		/*
 		 * We might as well update the local smgr_fsm_nblocks and
diff --git a/src/include/storage/md.h b/src/include/storage/md.h
@@ -38,7 +38,7 @@ extern void mdwriteback(SMgrRelation reln, ForkNumber forknum,
 						BlockNumber blocknum, BlockNumber nblocks);
 extern BlockNumber mdnblocks(SMgrRelation reln, ForkNumber forknum);
 extern void mdtruncate(SMgrRelation reln, ForkNumber forknum,
-					   BlockNumber nblocks);
+					   BlockNumber old_blocks, BlockNumber nblocks);
 extern void mdimmedsync(SMgrRelation reln, ForkNumber forknum);
 
 extern void ForgetDatabaseSyncRequests(Oid dbid);
diff --git a/src/include/storage/smgr.h b/src/include/storage/smgr.h
@@ -101,8 +101,9 @@ extern void smgrwrite(SMgrRelation reln, ForkNumber forknum,
 extern void smgrwriteback(SMgrRelation reln, ForkNumber forknum,
 						  BlockNumber blocknum, BlockNumber nblocks);
 extern BlockNumber smgrnblocks(SMgrRelation reln, ForkNumber forknum);
-extern void smgrtruncate(SMgrRelation reln, ForkNumber *forknum,
-						 int nforks, BlockNumber *nblocks);
+extern void smgrtruncate(SMgrRelation reln, ForkNumber *forknum, int nforks,
+						 BlockNumber *old_nblocks,
+						 BlockNumber *nblocks);
 extern void smgrimmedsync(SMgrRelation reln, ForkNumber forknum);
 extern void AtEOXact_SMgr(void);
 
