Skip to content

Commit e9c8747

Browse files
hlinnakahlinnaka
committed
Fix MVCC bug with prepared xact with subxacts on standby
We did not recover the subtransaction IDs of prepared transactions when starting a hot standby from a shutdown checkpoint. As a result, such subtransactions were considered as aborted, rather than in-progress. That would lead to hint bits being set incorrectly, and the subtransactions suddenly becoming visible to old snapshots when the prepared transaction was committed. To fix, update pg_subtrans with prepared transactions's subxids when starting hot standby from a shutdown checkpoint. The snapshots taken from that state need to be marked as "suboverflowed", so that we also check the pg_subtrans. Backport to all supported versions. Discussion: https://www.postgresql.org/message-id/6b852e98-2d49-4ca1-9e95-db419a2696e0@iki.fi
1 parent 071e19a commit e9c8747

File tree

7 files changed

+85
-18
lines changed

7 files changed

+85
-18
lines changed

‎src/backend/access/transam/twophase.c

Lines changed: 3 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -1997,9 +1997,8 @@ PrescanPreparedTransactions(TransactionId **xids_p, int *nxids_p)
19971997
* This is never called at the end of recovery - we use
19981998
* RecoverPreparedTransactions() at that point.
19991999
*
2000-
* The lack of calls to SubTransSetParent() calls here is by design;
2001-
* those calls are made by RecoverPreparedTransactions() at the end of recovery
2002-
* for those xacts that need this.
2000+
* This updates pg_subtrans, so that any subtransactions will be correctly
2001+
* seen as in-progress in snapshots taken during recovery.
20032002
*/
20042003
void
20052004
StandbyRecoverPreparedTransactions(void)
@@ -2019,7 +2018,7 @@ StandbyRecoverPreparedTransactions(void)
20192018

20202019
buf = ProcessTwoPhaseBuffer(xid,
20212020
gxact->prepare_start_lsn,
2022-
gxact->ondisk, false, false);
2021+
gxact->ondisk, true, false);
20232022
if (buf != NULL)
20242023
pfree(buf);
20252024
}

‎src/backend/access/transam/xlog.c

Lines changed: 8 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -7134,6 +7134,9 @@ StartupXLOG(void)
71347134
RunningTransactionsData running;
71357135
TransactionId latestCompletedXid;
71367136

7137+
/* Update pg_subtrans entries for any prepared transactions */
7138+
StandbyRecoverPreparedTransactions();
7139+
71377140
/*
71387141
* Construct a RunningTransactions snapshot representing a
71397142
* shut down server, with only prepared transactions still
@@ -7142,7 +7145,7 @@ StartupXLOG(void)
71427145
*/
71437146
running.xcnt = nxids;
71447147
running.subxcnt = 0;
7145-
running.subxid_overflow = false;
7148+
running.subxid_status = SUBXIDS_IN_SUBTRANS;
71467149
running.nextXid = XidFromFullTransactionId(checkPoint.nextFullXid);
71477150
running.oldestRunningXid = oldestActiveXID;
71487151
latestCompletedXid = XidFromFullTransactionId(checkPoint.nextFullXid);
@@ -7152,8 +7155,6 @@ StartupXLOG(void)
71527155
running.xids = xids;
71537156

71547157
ProcArrayApplyRecoveryInfo(&running);
7155-
7156-
StandbyRecoverPreparedTransactions();
71577158
}
71587159
}
71597160

@@ -10217,6 +10218,9 @@ xlog_redo(XLogReaderState *record)
1021710218

1021810219
oldestActiveXID = PrescanPreparedTransactions(&xids, &nxids);
1021910220

10221+
/* Update pg_subtrans entries for any prepared transactions */
10222+
StandbyRecoverPreparedTransactions();
10223+
1022010224
/*
1022110225
* Construct a RunningTransactions snapshot representing a shut
1022210226
* down server, with only prepared transactions still alive. We're
@@ -10225,7 +10229,7 @@ xlog_redo(XLogReaderState *record)
1022510229
*/
1022610230
running.xcnt = nxids;
1022710231
running.subxcnt = 0;
10228-
running.subxid_overflow = false;
10232+
running.subxid_status = SUBXIDS_IN_SUBTRANS;
1022910233
running.nextXid = XidFromFullTransactionId(checkPoint.nextFullXid);
1023010234
running.oldestRunningXid = oldestActiveXID;
1023110235
latestCompletedXid = XidFromFullTransactionId(checkPoint.nextFullXid);
@@ -10235,8 +10239,6 @@ xlog_redo(XLogReaderState *record)
1023510239
running.xids = xids;
1023610240

1023710241
ProcArrayApplyRecoveryInfo(&running);
10238-
10239-
StandbyRecoverPreparedTransactions();
1024010242
}
1024110243

1024210244
/* ControlFile->checkPointCopy always tracks the latest ckpt XID */

‎src/backend/storage/ipc/procarray.c

Lines changed: 15 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -734,7 +734,7 @@ ProcArrayApplyRecoveryInfo(RunningTransactions running)
734734
* If the snapshot isn't overflowed or if its empty we can reset our
735735
* pending state and use this snapshot instead.
736736
*/
737-
if (!running->subxid_overflow || running->xcnt == 0)
737+
if (running->subxid_status != SUBXIDS_MISSING || running->xcnt == 0)
738738
{
739739
/*
740740
* If we have already collected known assigned xids, we need to
@@ -886,7 +886,7 @@ ProcArrayApplyRecoveryInfo(RunningTransactions running)
886886
* missing, so conservatively assume the last one is latestObservedXid.
887887
* ----------
888888
*/
889-
if (running->subxid_overflow)
889+
if (running->subxid_status == SUBXIDS_MISSING)
890890
{
891891
standbyState = STANDBY_SNAPSHOT_PENDING;
892892

@@ -898,6 +898,18 @@ ProcArrayApplyRecoveryInfo(RunningTransactions running)
898898
standbyState = STANDBY_SNAPSHOT_READY;
899899

900900
standbySnapshotPendingXmin = InvalidTransactionId;
901+
902+
/*
903+
* If the 'xids' array didn't include all subtransactions, we have to
904+
* mark any snapshots taken as overflowed.
905+
*/
906+
if (running->subxid_status == SUBXIDS_IN_SUBTRANS)
907+
procArray->lastOverflowedXid = latestObservedXid;
908+
else
909+
{
910+
Assert(running->subxid_status == SUBXIDS_IN_ARRAY);
911+
procArray->lastOverflowedXid = InvalidTransactionId;
912+
}
901913
}
902914

903915
/*
@@ -2129,7 +2141,7 @@ GetRunningTransactionData(void)
21292141

21302142
CurrentRunningXacts->xcnt = count - subcount;
21312143
CurrentRunningXacts->subxcnt = subcount;
2132-
CurrentRunningXacts->subxid_overflow = suboverflowed;
2144+
CurrentRunningXacts->subxid_status = suboverflowed ? SUBXIDS_IN_SUBTRANS : SUBXIDS_IN_ARRAY;
21332145
CurrentRunningXacts->nextXid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
21342146
CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
21352147
CurrentRunningXacts->latestCompletedXid = latestCompletedXid;

‎src/backend/storage/ipc/standby.c

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -937,7 +937,7 @@ standby_redo(XLogReaderState *record)
937937

938938
running.xcnt = xlrec->xcnt;
939939
running.subxcnt = xlrec->subxcnt;
940-
running.subxid_overflow = xlrec->subxid_overflow;
940+
running.subxid_status = xlrec->subxid_overflow ? SUBXIDS_MISSING : SUBXIDS_IN_ARRAY;
941941
running.nextXid = xlrec->nextXid;
942942
running.latestCompletedXid = xlrec->latestCompletedXid;
943943
running.oldestRunningXid = xlrec->oldestRunningXid;
@@ -1093,7 +1093,7 @@ LogCurrentRunningXacts(RunningTransactions CurrRunningXacts)
10931093

10941094
xlrec.xcnt = CurrRunningXacts->xcnt;
10951095
xlrec.subxcnt = CurrRunningXacts->subxcnt;
1096-
xlrec.subxid_overflow = CurrRunningXacts->subxid_overflow;
1096+
xlrec.subxid_overflow = (CurrRunningXacts->subxid_status != SUBXIDS_IN_ARRAY);
10971097
xlrec.nextXid = CurrRunningXacts->nextXid;
10981098
xlrec.oldestRunningXid = CurrRunningXacts->oldestRunningXid;
10991099
xlrec.latestCompletedXid = CurrRunningXacts->latestCompletedXid;
@@ -1110,7 +1110,7 @@ LogCurrentRunningXacts(RunningTransactions CurrRunningXacts)
11101110

11111111
recptr = XLogInsert(RM_STANDBY_ID, XLOG_RUNNING_XACTS);
11121112

1113-
if (CurrRunningXacts->subxid_overflow)
1113+
if (xlrec.subxid_overflow)
11141114
elog(trace_recovery(DEBUG2),
11151115
"snapshot of %u running transactions overflowed (lsn %X/%X oldest xid %u latest complete %u next xid %u)",
11161116
CurrRunningXacts->xcnt,

‎src/include/storage/standby.h

Lines changed: 9 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -67,11 +67,19 @@ extern void StandbyReleaseOldLocks(TransactionId oldxid);
6767
* almost immediately see the data we need to begin executing queries.
6868
*/
6969

70+
typedef enum
71+
{
72+
SUBXIDS_IN_ARRAY, /* xids array includes all running subxids */
73+
SUBXIDS_MISSING, /* snapshot overflowed, subxids are missing */
74+
SUBXIDS_IN_SUBTRANS, /* subxids are not included in 'xids', but
75+
* pg_subtrans is fully up-to-date */
76+
} subxids_array_status;
77+
7078
typedef struct RunningTransactionsData
7179
{
7280
int xcnt; /* # of xact ids in xids[] */
7381
int subxcnt; /* # of subxact ids in xids[] */
74-
bool subxid_overflow; /* snapshot overflowed, subxids missing */
82+
subxids_array_status subxid_status;
7583
TransactionId nextXid; /* xid from ShmemVariableCache->nextFullXid */
7684
TransactionId oldestRunningXid; /* *not* oldestXmin */
7785
TransactionId latestCompletedXid; /* so we can set xmax */

‎src/test/recovery/t/009_twophase.pl

Lines changed: 46 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4,7 +4,7 @@
44

55
use PostgresNode;
66
use TestLib;
7-
use Test::More tests => 24;
7+
use Test::More tests => 27;
88

99
my $psql_out = '';
1010
my $psql_rc = '';
@@ -305,6 +305,51 @@ sub configure_and_reload
305305

306306
$cur_master->psql('postgres', "COMMIT PREPARED 'xact_009_12'");
307307

308+
###############################################################################
309+
# Check visibility of prepared transactions in standby after a restart while
310+
# primary is down.
311+
###############################################################################
312+
313+
$cur_master->psql(
314+
'postgres', "
315+
CREATE TABLE t_009_tbl_standby_mvcc (id int, msg text);
316+
BEGIN;
317+
INSERT INTO t_009_tbl_standby_mvcc VALUES (1, 'issued to ${cur_master_name}');
318+
SAVEPOINT s1;
319+
INSERT INTO t_009_tbl_standby_mvcc VALUES (2, 'issued to ${cur_master_name}');
320+
PREPARE TRANSACTION 'xact_009_standby_mvcc';
321+
");
322+
$cur_master->stop;
323+
$cur_standby->restart;
324+
325+
# Acquire a snapshot in standby, before we commit the prepared transaction
326+
my $standby_session = $cur_standby->background_psql('postgres', on_error_die => 1);
327+
$standby_session->query_safe("BEGIN ISOLATION LEVEL REPEATABLE READ");
328+
$psql_out = $standby_session->query_safe(
329+
"SELECT count(*) FROM t_009_tbl_standby_mvcc");
330+
is($psql_out, '0',
331+
"Prepared transaction not visible in standby before commit");
332+
333+
# Commit the transaction in primary
334+
$cur_master->start;
335+
$cur_master->psql('postgres', "
336+
SET synchronous_commit='remote_apply'; -- To ensure the standby is caught up
337+
COMMIT PREPARED 'xact_009_standby_mvcc'
338+
");
339+
340+
# Still not visible to the old snapshot
341+
$psql_out = $standby_session->query_safe(
342+
"SELECT count(*) FROM t_009_tbl_standby_mvcc");
343+
is($psql_out, '0',
344+
"Committed prepared transaction not visible to old snapshot in standby");
345+
346+
# Is visible to a new snapshot
347+
$standby_session->query_safe("COMMIT");
348+
$psql_out = $standby_session->query_safe(
349+
"SELECT count(*) FROM t_009_tbl_standby_mvcc");
350+
is($psql_out, '2',
351+
"Committed prepared transaction is visible to new snapshot in standby");
352+
308353
###############################################################################
309354
# Check for a lock conflict between prepared transaction with DDL inside and
310355
# replay of XLOG_STANDBY_LOCK wal record.

‎src/tools/pgindent/typedefs.list

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -3373,6 +3373,7 @@ string
33733373
substitute_actual_parameters_context
33743374
substitute_actual_srf_parameters_context
33753375
substitute_phv_relids_context
3376+
subxids_array_status
33763377
svtype
33773378
symbol
33783379
tablespaceinfo

0 commit comments

Comments
 (0)