commit 65928b9 (HEAD) Author: Noah Misch AuthorDate: Sun Feb 17 23:26:45 2019 -0800 Commit: Noah Misch CommitDate: Sun Feb 17 23:26:45 2019 -0800 Prevent concurrent SimpleLruTruncate() for any given SLRU. The SimpleLruTruncate() header comment states the new coding rule. This closes race conditions that presented rare opportunities for data loss. Symptoms and recommendations to users mirror the previous commit[1]. Back-patch to 9.4 (all supported versions). Reviewed by TBD. Discussion: https://postgr.es/m/TBD [1] slru-truncate-modulo-v2.patch of https://postgr.es/m/20190217040913.GA1305059@rfd.leadboat.com diff --git a/doc/src/sgml/monitoring.sgml b/doc/src/sgml/monitoring.sgml index 5c1408b..b7ddc73 100644 --- a/doc/src/sgml/monitoring.sgml +++ b/doc/src/sgml/monitoring.sgml @@ -849,7 +849,7 @@ postgres 27093 0.0 0.0 30096 2752 ? Ss 11:34 0:00 postgres: ser - LWLock + LWLock ShmemIndexLock Waiting to find or allocate space in shared memory. @@ -1044,6 +1044,16 @@ postgres 27093 0.0 0.0 30096 2752 ? Ss 11:34 0:00 postgres: ser the oldest transaction id available to it. + WrapLimitsVacuumLock + Waiting to update limits on transaction id and multixact + consumption. + + + AsyncQueueTailLock + Waiting to update limit on notification message + storage. + + clog Waiting for I/O on a clog (transaction status) buffer. @@ -1128,7 +1138,7 @@ postgres 27093 0.0 0.0 30096 2752 ? Ss 11:34 0:00 postgres: ser counters during Parallel Hash plan execution. - Lock + Lock relation Waiting to acquire a lock on a relation. @@ -1137,6 +1147,12 @@ postgres 27093 0.0 0.0 30096 2752 ? Ss 11:34 0:00 postgres: ser Waiting to extend a relation. + frozen IDs + Waiting to + update pg_database.datfrozenxid + and pg_database.datminmxid. + + page Waiting to acquire a lock on page of a relation. diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c index 3623352..8c9d5d8 100644 --- a/src/backend/access/transam/slru.c +++ b/src/backend/access/transam/slru.c @@ -1163,6 +1163,14 @@ SimpleLruFlush(SlruCtl ctl, bool allow_redirtied) /* * Remove all segments before the one holding the passed page number + * + * All SLRUs prevent concurrent calls to this function, either with an LWLock + * or by calling it only as part of a checkpoint. Mutual exclusion must begin + * before computing cutoffPage. Mutual exclusion must end after any limit + * update that would permit other backends to write fresh data into the + * segment immediately preceding the one containing cutoffPage. Otherwise, + * when the SLRU is quite full, SimpleLruTruncate() might delete that segment + * after it has accrued freshly-written data. */ void SimpleLruTruncate(SlruCtl ctl, int cutoffPage) diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c index cbc6129..a480dd5 100644 --- a/src/backend/access/transam/subtrans.c +++ b/src/backend/access/transam/subtrans.c @@ -347,8 +347,8 @@ ExtendSUBTRANS(TransactionId newestXact) /* * Remove all SUBTRANS segments before the one holding the passed transaction ID * - * This is normally called during checkpoint, with oldestXact being the - * oldest TransactionXmin of any running transaction. + * oldestXact is the oldest TransactionXmin of any running transaction. This + * is called only during checkpoint. */ void TruncateSUBTRANS(TransactionId oldestXact) diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c index 5a7ee0d..66e888c 100644 --- a/src/backend/commands/async.c +++ b/src/backend/commands/async.c @@ -223,19 +223,22 @@ typedef struct QueueBackendStatus /* * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff) * - * The AsyncQueueControl structure is protected by the AsyncQueueLock. + * The AsyncQueueControl structure is protected by the AsyncQueueLock and + * AsyncQueueTailLock. * - * When holding the lock in SHARED mode, backends may only inspect their own - * entries as well as the head and tail pointers. Consequently we can allow a - * backend to update its own record while holding only SHARED lock (since no - * other backend will inspect it). + * When holding AsyncQueueLock in SHARED mode, backends may only inspect their + * own entries as well as the head and tail pointers. Consequently we can + * allow a backend to update its own record while holding only SHARED lock + * (since no other backend will inspect it). * - * When holding the lock in EXCLUSIVE mode, backends can inspect the entries - * of other backends and also change the head and tail pointers. + * When holding AsyncQueueLock in EXCLUSIVE mode, backends can inspect the + * entries of other backends and also change the head pointer. When holding + * both AsyncQueueLock and AsyncQueueTailLock in EXCLUSIVE mode, backends can + * change the tail pointer. * * AsyncCtlLock is used as the control lock for the pg_notify SLRU buffers. - * In order to avoid deadlocks, whenever we need both locks, we always first - * get AsyncQueueLock and then AsyncCtlLock. + * In order to avoid deadlocks, whenever we need multiple locks, we first get + * AsyncQueueTailLock, then AsyncQueueLock, and lastly AsyncCtlLock. * * Each backend uses the backend[] array entry with index equal to its * BackendId (which can range from 1 to MaxBackends). We rely on this to make @@ -2010,6 +2013,10 @@ asyncQueueAdvanceTail(void) int newtailpage; int boundary; + /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */ + LWLockAcquire(AsyncQueueTailLock, LW_EXCLUSIVE); + + /* Compute the new tail. */ LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE); min = QUEUE_HEAD; for (i = 1; i <= MaxBackends; i++) @@ -2018,7 +2025,6 @@ asyncQueueAdvanceTail(void) min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i)); } oldtailpage = QUEUE_POS_PAGE(QUEUE_TAIL); - QUEUE_TAIL = min; LWLockRelease(AsyncQueueLock); /* @@ -2038,6 +2044,17 @@ asyncQueueAdvanceTail(void) */ SimpleLruTruncate(AsyncCtl, newtailpage); } + + /* + * Advertise the new tail. This changes asyncQueueIsFull()'s verdict for + * the segment immediately prior to the new tail, allowing fresh data into + * that segment. + */ + LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE); + QUEUE_TAIL = min; + LWLockRelease(AsyncQueueLock); + + LWLockRelease(AsyncQueueTailLock); } /* diff --git a/src/backend/commands/vacuum.c b/src/backend/commands/vacuum.c index e91df21..326f48b 100644 --- a/src/backend/commands/vacuum.c +++ b/src/backend/commands/vacuum.c @@ -1225,6 +1225,14 @@ vac_update_datfrozenxid(void) bool dirty = false; /* + * Restrict this task to one backend per database. This avoids race + * conditions that would move datfrozenxid or datminmxid backward or call + * vac_truncate_clog() with a datfrozenxid preceding a datfrozenxid passed + * to an earlier vac_truncate_clog() call. + */ + LockDatabaseFrozenIds(ExclusiveLock); + + /* * Initialize the "min" calculation with GetOldestXmin, which is a * reasonable approximation to the minimum relfrozenxid for not-yet- * committed pg_class entries for new tables; see AddNewRelationTuple(). @@ -1388,6 +1396,9 @@ vac_truncate_clog(TransactionId frozenXID, bool bogus = false; bool frozenAlreadyWrapped = false; + /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */ + LWLockAcquire(WrapLimitsVacuumLock, LW_EXCLUSIVE); + /* init oldest datoids to sync with my frozenXID/minMulti values */ oldestxid_datoid = MyDatabaseId; minmulti_datoid = MyDatabaseId; @@ -1497,6 +1508,8 @@ vac_truncate_clog(TransactionId frozenXID, */ SetTransactionIdLimit(frozenXID, oldestxid_datoid); SetMultiXactIdLimit(minMulti, minmulti_datoid, false); + + LWLockRelease(WrapLimitsVacuumLock); } diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c index e688ba8..4873835 100644 --- a/src/backend/storage/lmgr/lmgr.c +++ b/src/backend/storage/lmgr/lmgr.c @@ -458,6 +458,21 @@ UnlockRelationForExtension(Relation relation, LOCKMODE lockmode) } /* + * LockDatabaseFrozenIds + * + * This allows one backend per database to execute vac_update_datfrozenxid(). + */ +void +LockDatabaseFrozenIds(LOCKMODE lockmode) +{ + LOCKTAG tag; + + SET_LOCKTAG_DATABASE_FROZEN_IDS(tag, MyDatabaseId); + + (void) LockAcquire(&tag, lockmode, false, false); +} + +/* * LockPage * * Obtain a page-level lock. This is currently used by some index access @@ -1060,6 +1075,11 @@ DescribeLockTag(StringInfo buf, const LOCKTAG *tag) tag->locktag_field2, tag->locktag_field1); break; + case LOCKTAG_DATABASE_FROZEN_IDS: + appendStringInfo(buf, + _("pg_database.datfrozenxid of database %u"), + tag->locktag_field1); + break; case LOCKTAG_PAGE: appendStringInfo(buf, _("page %u of relation %u of database %u"), diff --git a/src/backend/storage/lmgr/lwlocknames.txt b/src/backend/storage/lmgr/lwlocknames.txt index db47843..d517023 100644 --- a/src/backend/storage/lmgr/lwlocknames.txt +++ b/src/backend/storage/lmgr/lwlocknames.txt @@ -49,3 +49,5 @@ MultiXactTruncationLock 41 OldSnapshotTimeMapLock 42 LogicalRepWorkerLock 43 CLogTruncationLock 44 +WrapLimitsVacuumLock 45 +AsyncQueueTailLock 46 diff --git a/src/backend/utils/adt/lockfuncs.c b/src/backend/utils/adt/lockfuncs.c index ffd1970..cbf0cab 100644 --- a/src/backend/utils/adt/lockfuncs.c +++ b/src/backend/utils/adt/lockfuncs.c @@ -26,6 +26,7 @@ const char *const LockTagTypeNames[] = { "relation", "extend", + "frozen IDs", "page", "tuple", "transactionid", @@ -245,6 +246,17 @@ pg_lock_status(PG_FUNCTION_ARGS) nulls[8] = true; nulls[9] = true; break; + case LOCKTAG_DATABASE_FROZEN_IDS: + values[1] = ObjectIdGetDatum(instance->locktag.locktag_field1); + nulls[2] = true; + nulls[3] = true; + nulls[4] = true; + nulls[5] = true; + nulls[6] = true; + nulls[7] = true; + nulls[8] = true; + nulls[9] = true; + break; case LOCKTAG_PAGE: values[1] = ObjectIdGetDatum(instance->locktag.locktag_field1); values[2] = ObjectIdGetDatum(instance->locktag.locktag_field2); diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h index 3d705fa..f113f55 100644 --- a/src/include/storage/lmgr.h +++ b/src/include/storage/lmgr.h @@ -59,6 +59,9 @@ extern bool ConditionalLockRelationForExtension(Relation relation, LOCKMODE lockmode); extern int RelationExtensionLockWaiterCount(Relation relation); +/* Lock to recompute pg_database.datfrozenxid in the current database */ +extern void LockDatabaseFrozenIds(LOCKMODE lockmode); + /* Lock a page (currently only used within indexes) */ extern void LockPage(Relation relation, BlockNumber blkno, LOCKMODE lockmode); extern bool ConditionalLockPage(Relation relation, BlockNumber blkno, LOCKMODE lockmode); diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h index 16b927c..f3b6ef4 100644 --- a/src/include/storage/lock.h +++ b/src/include/storage/lock.h @@ -141,6 +141,8 @@ typedef enum LockTagType /* ID info for a relation is DB OID + REL OID; DB OID = 0 if shared */ LOCKTAG_RELATION_EXTEND, /* the right to extend a relation */ /* same ID info as RELATION */ + LOCKTAG_DATABASE_FROZEN_IDS, /* pg_database.datfrozenxid */ + /* ID info for frozen IDs is DB OID */ LOCKTAG_PAGE, /* one page of a relation */ /* ID info for a page is RELATION info + BlockNumber */ LOCKTAG_TUPLE, /* one physical tuple */ @@ -206,6 +208,14 @@ typedef struct LOCKTAG (locktag).locktag_type = LOCKTAG_RELATION_EXTEND, \ (locktag).locktag_lockmethodid = DEFAULT_LOCKMETHOD) +#define SET_LOCKTAG_DATABASE_FROZEN_IDS(locktag,dboid) \ + ((locktag).locktag_field1 = (dboid), \ + (locktag).locktag_field2 = 0, \ + (locktag).locktag_field3 = 0, \ + (locktag).locktag_field4 = 0, \ + (locktag).locktag_type = LOCKTAG_DATABASE_FROZEN_IDS, \ + (locktag).locktag_lockmethodid = DEFAULT_LOCKMETHOD) + #define SET_LOCKTAG_PAGE(locktag,dboid,reloid,blocknum) \ ((locktag).locktag_field1 = (dboid), \ (locktag).locktag_field2 = (reloid), \