1 /*-------------------------------------------------------------------------
4 * The postgres vacuum cleaner.
6 * This file now includes only control and dispatch code for VACUUM and
7 * ANALYZE commands. Regular VACUUM is implemented in vacuumlazy.c,
8 * ANALYZE in analyze.c, and VACUUM FULL is a variant of CLUSTER, handled
12 * Portions Copyright (c) 1996-2015, PostgreSQL Global Development Group
13 * Portions Copyright (c) 1994, Regents of the University of California
17 * src/backend/commands/vacuum.c
19 *-------------------------------------------------------------------------
25 #include "access/clog.h"
26 #include "access/commit_ts.h"
27 #include "access/genam.h"
28 #include "access/heapam.h"
29 #include "access/htup_details.h"
30 #include "access/multixact.h"
31 #include "access/transam.h"
32 #include "access/xact.h"
33 #include "catalog/namespace.h"
34 #include "catalog/pg_database.h"
35 #include "catalog/pg_namespace.h"
36 #include "commands/cluster.h"
37 #include "commands/vacuum.h"
38 #include "miscadmin.h"
40 #include "postmaster/autovacuum.h"
41 #include "storage/bufmgr.h"
42 #include "storage/lmgr.h"
43 #include "storage/proc.h"
44 #include "storage/procarray.h"
45 #include "utils/acl.h"
46 #include "utils/fmgroids.h"
47 #include "utils/guc.h"
48 #include "utils/memutils.h"
49 #include "utils/snapmgr.h"
50 #include "utils/syscache.h"
51 #include "utils/tqual.h"
57 int vacuum_freeze_min_age;
58 int vacuum_freeze_table_age;
59 int vacuum_multixact_freeze_min_age;
60 int vacuum_multixact_freeze_table_age;
63 /* A few variables that don't seem worth passing around as parameters */
64 static MemoryContext vac_context = NULL;
65 static BufferAccessStrategy vac_strategy;
68 /* non-export function prototypes */
69 static List *get_rel_oids(Oid relid, const RangeVar *vacrel);
70 static void vac_truncate_clog(TransactionId frozenXID,
72 TransactionId lastSaneFrozenXid,
73 MultiXactId lastSaneMinMulti);
74 static bool vacuum_rel(Oid relid, RangeVar *relation, int options,
75 VacuumParams *params);
78 * Primary entry point for manual VACUUM and ANALYZE commands
80 * This is mainly a preparation wrapper for the real operations that will
84 ExecVacuum(VacuumStmt *vacstmt, bool isTopLevel)
88 /* sanity checks on options */
89 Assert(vacstmt->options & (VACOPT_VACUUM | VACOPT_ANALYZE));
90 Assert((vacstmt->options & VACOPT_VACUUM) ||
91 !(vacstmt->options & (VACOPT_FULL | VACOPT_FREEZE)));
92 Assert((vacstmt->options & VACOPT_ANALYZE) || vacstmt->va_cols == NIL);
93 Assert(!(vacstmt->options & VACOPT_SKIPTOAST));
96 * All freeze ages are zero if the FREEZE option is given; otherwise pass
97 * them as -1 which means to use the default values.
99 if (vacstmt->options & VACOPT_FREEZE)
101 params.freeze_min_age = 0;
102 params.freeze_table_age = 0;
103 params.multixact_freeze_min_age = 0;
104 params.multixact_freeze_table_age = 0;
108 params.freeze_min_age = -1;
109 params.freeze_table_age = -1;
110 params.multixact_freeze_min_age = -1;
111 params.multixact_freeze_table_age = -1;
114 /* user-invoked vacuum is never "for wraparound" */
115 params.is_wraparound = false;
117 /* user-invoked vacuum never uses this parameter */
118 params.log_min_duration = -1;
120 /* Now go through the common routine */
121 vacuum(vacstmt->options, vacstmt->relation, InvalidOid, ¶ms,
122 vacstmt->va_cols, NULL, isTopLevel);
126 * Primary entry point for VACUUM and ANALYZE commands.
128 * options is a bitmask of VacuumOption flags, indicating what to do.
130 * relid, if not InvalidOid, indicate the relation to process; otherwise,
131 * the RangeVar is used. (The latter must always be passed, because it's
132 * used for error messages.)
134 * params contains a set of parameters that can be used to customize the
137 * va_cols is a list of columns to analyze, or NIL to process them all.
139 * bstrategy is normally given as NULL, but in autovacuum it can be passed
140 * in to use the same buffer strategy object across multiple vacuum() calls.
142 * isTopLevel should be passed down from ProcessUtility.
144 * It is the caller's responsibility that all parameters are allocated in a
145 * memory context that will not disappear at transaction commit.
148 vacuum(int options, RangeVar *relation, Oid relid, VacuumParams *params,
149 List *va_cols, BufferAccessStrategy bstrategy, bool isTopLevel)
151 const char *stmttype;
152 volatile bool in_outer_xact,
155 static bool in_vacuum = false;
157 Assert(params != NULL);
159 stmttype = (options & VACOPT_VACUUM) ? "VACUUM" : "ANALYZE";
162 * We cannot run VACUUM inside a user transaction block; if we were inside
163 * a transaction, then our commit- and start-transaction-command calls
164 * would not have the intended effect! There are numerous other subtle
165 * dependencies on this, too.
167 * ANALYZE (without VACUUM) can run either way.
169 if (options & VACOPT_VACUUM)
171 PreventTransactionChain(isTopLevel, stmttype);
172 in_outer_xact = false;
175 in_outer_xact = IsInTransactionChain(isTopLevel);
178 * Due to static variables vac_context, anl_context and vac_strategy,
179 * vacuum() is not reentrant. This matters when VACUUM FULL or ANALYZE
180 * calls a hostile index expression that itself calls ANALYZE.
183 elog(ERROR, "%s cannot be executed from VACUUM or ANALYZE", stmttype);
186 * Send info about dead objects to the statistics collector, unless we are
187 * in autovacuum --- autovacuum.c does this for itself.
189 if ((options & VACOPT_VACUUM) && !IsAutoVacuumWorkerProcess())
190 pgstat_vacuum_stat();
193 * Create special memory context for cross-transaction storage.
195 * Since it is a child of PortalContext, it will go away eventually even
196 * if we suffer an error; there's no need for special abort cleanup logic.
198 vac_context = AllocSetContextCreate(PortalContext,
200 ALLOCSET_DEFAULT_MINSIZE,
201 ALLOCSET_DEFAULT_INITSIZE,
202 ALLOCSET_DEFAULT_MAXSIZE);
205 * If caller didn't give us a buffer strategy object, make one in the
206 * cross-transaction memory context.
208 if (bstrategy == NULL)
210 MemoryContext old_context = MemoryContextSwitchTo(vac_context);
212 bstrategy = GetAccessStrategy(BAS_VACUUM);
213 MemoryContextSwitchTo(old_context);
215 vac_strategy = bstrategy;
218 * Build list of relations to process, unless caller gave us one. (If we
219 * build one, we put it in vac_context for safekeeping.)
221 relations = get_rel_oids(relid, relation);
224 * Decide whether we need to start/commit our own transactions.
226 * For VACUUM (with or without ANALYZE): always do so, so that we can
227 * release locks as soon as possible. (We could possibly use the outer
228 * transaction for a one-table VACUUM, but handling TOAST tables would be
231 * For ANALYZE (no VACUUM): if inside a transaction block, we cannot
232 * start/commit our own transactions. Also, there's no need to do so if
233 * only processing one relation. For multiple relations when not within a
234 * transaction block, and also in an autovacuum worker, use own
235 * transactions so we can release locks sooner.
237 if (options & VACOPT_VACUUM)
238 use_own_xacts = true;
241 Assert(options & VACOPT_ANALYZE);
242 if (IsAutoVacuumWorkerProcess())
243 use_own_xacts = true;
244 else if (in_outer_xact)
245 use_own_xacts = false;
246 else if (list_length(relations) > 1)
247 use_own_xacts = true;
249 use_own_xacts = false;
253 * vacuum_rel expects to be entered with no transaction active; it will
254 * start and commit its own transaction. But we are called by an SQL
255 * command, and so we are executing inside a transaction already. We
256 * commit the transaction started in PostgresMain() here, and start
257 * another one before exiting to match the commit waiting for us back in
262 Assert(!in_outer_xact);
264 /* ActiveSnapshot is not set by autovacuum */
265 if (ActiveSnapshotSet())
268 /* matches the StartTransaction in PostgresMain() */
269 CommitTransactionCommand();
272 /* Turn vacuum cost accounting on or off */
278 VacuumCostActive = (VacuumCostDelay > 0);
279 VacuumCostBalance = 0;
285 * Loop to process each selected relation.
287 foreach(cur, relations)
289 Oid relid = lfirst_oid(cur);
291 if (options & VACOPT_VACUUM)
293 if (!vacuum_rel(relid, relation, options, params))
297 if (options & VACOPT_ANALYZE)
300 * If using separate xacts, start one for analyze. Otherwise,
301 * we can use the outer transaction.
305 StartTransactionCommand();
306 /* functions in indexes may want a snapshot set */
307 PushActiveSnapshot(GetTransactionSnapshot());
310 analyze_rel(relid, relation, options, params,
311 va_cols, in_outer_xact, vac_strategy);
316 CommitTransactionCommand();
324 VacuumCostActive = false;
330 VacuumCostActive = false;
333 * Finish up processing.
337 /* here, we are not in a transaction */
340 * This matches the CommitTransaction waiting for us in
343 StartTransactionCommand();
346 if ((options & VACOPT_VACUUM) && !IsAutoVacuumWorkerProcess())
349 * Update pg_database.datfrozenxid, and truncate pg_clog if possible.
350 * (autovacuum.c does this for itself.)
352 vac_update_datfrozenxid();
356 * Clean up working storage --- note we must do this after
357 * StartTransactionCommand, else we might be trying to delete the active
360 MemoryContextDelete(vac_context);
365 * Build a list of Oids for each relation to be processed
367 * The list is built in vac_context so that it will survive across our
368 * per-relation transactions.
371 get_rel_oids(Oid relid, const RangeVar *vacrel)
373 List *oid_list = NIL;
374 MemoryContext oldcontext;
376 /* OID supplied by VACUUM's caller? */
377 if (OidIsValid(relid))
379 oldcontext = MemoryContextSwitchTo(vac_context);
380 oid_list = lappend_oid(oid_list, relid);
381 MemoryContextSwitchTo(oldcontext);
385 /* Process a specific relation */
389 * Since we don't take a lock here, the relation might be gone, or the
390 * RangeVar might no longer refer to the OID we look up here. In the
391 * former case, VACUUM will do nothing; in the latter case, it will
392 * process the OID we looked up here, rather than the new one. Neither
393 * is ideal, but there's little practical alternative, since we're
394 * going to commit this transaction and begin a new one between now
397 relid = RangeVarGetRelid(vacrel, NoLock, false);
399 /* Make a relation list entry for this guy */
400 oldcontext = MemoryContextSwitchTo(vac_context);
401 oid_list = lappend_oid(oid_list, relid);
402 MemoryContextSwitchTo(oldcontext);
407 * Process all plain relations and materialized views listed in
414 pgclass = heap_open(RelationRelationId, AccessShareLock);
416 scan = heap_beginscan_catalog(pgclass, 0, NULL);
418 while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
420 Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
422 if (classForm->relkind != RELKIND_RELATION &&
423 classForm->relkind != RELKIND_MATVIEW)
426 /* Make a relation list entry for this guy */
427 oldcontext = MemoryContextSwitchTo(vac_context);
428 oid_list = lappend_oid(oid_list, HeapTupleGetOid(tuple));
429 MemoryContextSwitchTo(oldcontext);
433 heap_close(pgclass, AccessShareLock);
440 * vacuum_set_xid_limits() -- compute oldest-Xmin and freeze cutoff points
442 * The output parameters are:
443 * - oldestXmin is the cutoff value used to distinguish whether tuples are
444 * DEAD or RECENTLY_DEAD (see HeapTupleSatisfiesVacuum).
445 * - freezeLimit is the Xid below which all Xids are replaced by
446 * FrozenTransactionId during vacuum.
447 * - xidFullScanLimit (computed from table_freeze_age parameter)
448 * represents a minimum Xid value; a table whose relfrozenxid is older than
449 * this will have a full-table vacuum applied to it, to freeze tuples across
450 * the whole table. Vacuuming a table younger than this value can use a
452 * - multiXactCutoff is the value below which all MultiXactIds are removed from
454 * - mxactFullScanLimit is a value against which a table's relminmxid value is
455 * compared to produce a full-table vacuum, as with xidFullScanLimit.
457 * xidFullScanLimit and mxactFullScanLimit can be passed as NULL if caller is
461 vacuum_set_xid_limits(Relation rel,
463 int freeze_table_age,
464 int multixact_freeze_min_age,
465 int multixact_freeze_table_age,
466 TransactionId *oldestXmin,
467 TransactionId *freezeLimit,
468 TransactionId *xidFullScanLimit,
469 MultiXactId *multiXactCutoff,
470 MultiXactId *mxactFullScanLimit)
475 TransactionId safeLimit;
476 MultiXactId mxactLimit;
477 MultiXactId safeMxactLimit;
480 * We can always ignore processes running lazy vacuum. This is because we
481 * use these values only for deciding which tuples we must keep in the
482 * tables. Since lazy vacuum doesn't write its XID anywhere, it's safe to
483 * ignore it. In theory it could be problematic to ignore lazy vacuums in
484 * a full vacuum, but keep in mind that only one vacuum process can be
485 * working on a particular table at any time, and that each vacuum is
486 * always an independent transaction.
488 *oldestXmin = GetOldestXmin(rel, true);
490 Assert(TransactionIdIsNormal(*oldestXmin));
493 * Determine the minimum freeze age to use: as specified by the caller, or
494 * vacuum_freeze_min_age, but in any case not more than half
495 * autovacuum_freeze_max_age, so that autovacuums to prevent XID
496 * wraparound won't occur too frequently.
498 freezemin = freeze_min_age;
500 freezemin = vacuum_freeze_min_age;
501 freezemin = Min(freezemin, autovacuum_freeze_max_age / 2);
502 Assert(freezemin >= 0);
505 * Compute the cutoff XID, being careful not to generate a "permanent" XID
507 limit = *oldestXmin - freezemin;
508 if (!TransactionIdIsNormal(limit))
509 limit = FirstNormalTransactionId;
512 * If oldestXmin is very far back (in practice, more than
513 * autovacuum_freeze_max_age / 2 XIDs old), complain and force a minimum
514 * freeze age of zero.
516 safeLimit = ReadNewTransactionId() - autovacuum_freeze_max_age;
517 if (!TransactionIdIsNormal(safeLimit))
518 safeLimit = FirstNormalTransactionId;
520 if (TransactionIdPrecedes(limit, safeLimit))
523 (errmsg("oldest xmin is far in the past"),
524 errhint("Close open transactions soon to avoid wraparound problems.")));
528 *freezeLimit = limit;
531 * Determine the minimum multixact freeze age to use: as specified by
532 * caller, or vacuum_multixact_freeze_min_age, but in any case not more
533 * than half autovacuum_multixact_freeze_max_age, so that autovacuums to
534 * prevent MultiXact wraparound won't occur too frequently.
536 mxid_freezemin = multixact_freeze_min_age;
537 if (mxid_freezemin < 0)
538 mxid_freezemin = vacuum_multixact_freeze_min_age;
539 mxid_freezemin = Min(mxid_freezemin,
540 autovacuum_multixact_freeze_max_age / 2);
541 Assert(mxid_freezemin >= 0);
543 /* compute the cutoff multi, being careful to generate a valid value */
544 mxactLimit = GetOldestMultiXactId() - mxid_freezemin;
545 if (mxactLimit < FirstMultiXactId)
546 mxactLimit = FirstMultiXactId;
549 ReadNextMultiXactId() - autovacuum_multixact_freeze_max_age;
550 if (safeMxactLimit < FirstMultiXactId)
551 safeMxactLimit = FirstMultiXactId;
553 if (MultiXactIdPrecedes(mxactLimit, safeMxactLimit))
556 (errmsg("oldest multixact is far in the past"),
557 errhint("Close open transactions with multixacts soon to avoid wraparound problems.")));
558 mxactLimit = safeMxactLimit;
561 *multiXactCutoff = mxactLimit;
563 if (xidFullScanLimit != NULL)
567 Assert(mxactFullScanLimit != NULL);
570 * Determine the table freeze age to use: as specified by the caller,
571 * or vacuum_freeze_table_age, but in any case not more than
572 * autovacuum_freeze_max_age * 0.95, so that if you have e.g nightly
573 * VACUUM schedule, the nightly VACUUM gets a chance to freeze tuples
574 * before anti-wraparound autovacuum is launched.
576 freezetable = freeze_table_age;
578 freezetable = vacuum_freeze_table_age;
579 freezetable = Min(freezetable, autovacuum_freeze_max_age * 0.95);
580 Assert(freezetable >= 0);
583 * Compute XID limit causing a full-table vacuum, being careful not to
584 * generate a "permanent" XID.
586 limit = ReadNewTransactionId() - freezetable;
587 if (!TransactionIdIsNormal(limit))
588 limit = FirstNormalTransactionId;
590 *xidFullScanLimit = limit;
593 * Similar to the above, determine the table freeze age to use for
594 * multixacts: as specified by the caller, or
595 * vacuum_multixact_freeze_table_age, but in any case not more than
596 * autovacuum_multixact_freeze_table_age * 0.95, so that if you have
597 * e.g. nightly VACUUM schedule, the nightly VACUUM gets a chance to
598 * freeze multixacts before anti-wraparound autovacuum is launched.
600 freezetable = multixact_freeze_table_age;
602 freezetable = vacuum_multixact_freeze_table_age;
603 freezetable = Min(freezetable,
604 autovacuum_multixact_freeze_max_age * 0.95);
605 Assert(freezetable >= 0);
608 * Compute MultiXact limit causing a full-table vacuum, being careful
609 * to generate a valid MultiXact value.
611 mxactLimit = ReadNextMultiXactId() - freezetable;
612 if (mxactLimit < FirstMultiXactId)
613 mxactLimit = FirstMultiXactId;
615 *mxactFullScanLimit = mxactLimit;
619 Assert(mxactFullScanLimit == NULL);
624 * vac_estimate_reltuples() -- estimate the new value for pg_class.reltuples
626 * If we scanned the whole relation then we should just use the count of
627 * live tuples seen; but if we did not, we should not trust the count
628 * unreservedly, especially not in VACUUM, which may have scanned a quite
629 * nonrandom subset of the table. When we have only partial information,
630 * we take the old value of pg_class.reltuples as a measurement of the
631 * tuple density in the unscanned pages.
633 * This routine is shared by VACUUM and ANALYZE.
636 vac_estimate_reltuples(Relation relation, bool is_analyze,
637 BlockNumber total_pages,
638 BlockNumber scanned_pages,
639 double scanned_tuples)
641 BlockNumber old_rel_pages = relation->rd_rel->relpages;
642 double old_rel_tuples = relation->rd_rel->reltuples;
646 double updated_density;
648 /* If we did scan the whole table, just use the count as-is */
649 if (scanned_pages >= total_pages)
650 return scanned_tuples;
653 * If scanned_pages is zero but total_pages isn't, keep the existing value
654 * of reltuples. (Note: callers should avoid updating the pg_class
655 * statistics in this situation, since no new information has been
658 if (scanned_pages == 0)
659 return old_rel_tuples;
662 * If old value of relpages is zero, old density is indeterminate; we
663 * can't do much except scale up scanned_tuples to match total_pages.
665 if (old_rel_pages == 0)
666 return floor((scanned_tuples / scanned_pages) * total_pages + 0.5);
669 * Okay, we've covered the corner cases. The normal calculation is to
670 * convert the old measurement to a density (tuples per page), then update
671 * the density using an exponential-moving-average approach, and finally
672 * compute reltuples as updated_density * total_pages.
674 * For ANALYZE, the moving average multiplier is just the fraction of the
675 * table's pages we scanned. This is equivalent to assuming that the
676 * tuple density in the unscanned pages didn't change. Of course, it
677 * probably did, if the new density measurement is different. But over
678 * repeated cycles, the value of reltuples will converge towards the
679 * correct value, if repeated measurements show the same new density.
681 * For VACUUM, the situation is a bit different: we have looked at a
682 * nonrandom sample of pages, but we know for certain that the pages we
683 * didn't look at are precisely the ones that haven't changed lately.
684 * Thus, there is a reasonable argument for doing exactly the same thing
685 * as for the ANALYZE case, that is use the old density measurement as the
686 * value for the unscanned pages.
688 * This logic could probably use further refinement.
690 old_density = old_rel_tuples / old_rel_pages;
691 new_density = scanned_tuples / scanned_pages;
692 multiplier = (double) scanned_pages / (double) total_pages;
693 updated_density = old_density + (new_density - old_density) * multiplier;
694 return floor(updated_density * total_pages + 0.5);
699 * vac_update_relstats() -- update statistics for one relation
701 * Update the whole-relation statistics that are kept in its pg_class
702 * row. There are additional stats that will be updated if we are
703 * doing ANALYZE, but we always update these stats. This routine works
704 * for both index and heap relation entries in pg_class.
706 * We violate transaction semantics here by overwriting the rel's
707 * existing pg_class tuple with the new values. This is reasonably
708 * safe as long as we're sure that the new values are correct whether or
709 * not this transaction commits. The reason for doing this is that if
710 * we updated these tuples in the usual way, vacuuming pg_class itself
711 * wouldn't work very well --- by the time we got done with a vacuum
712 * cycle, most of the tuples in pg_class would've been obsoleted. Of
713 * course, this only works for fixed-size not-null columns, but these are.
715 * Another reason for doing it this way is that when we are in a lazy
716 * VACUUM and have PROC_IN_VACUUM set, we mustn't do any regular updates.
717 * Somebody vacuuming pg_class might think they could delete a tuple
718 * marked with xmin = our xid.
720 * In addition to fundamentally nontransactional statistics such as
721 * relpages and relallvisible, we try to maintain certain lazily-updated
722 * DDL flags such as relhasindex, by clearing them if no longer correct.
723 * It's safe to do this in VACUUM, which can't run in parallel with
724 * CREATE INDEX/RULE/TRIGGER and can't be part of a transaction block.
725 * However, it's *not* safe to do it in an ANALYZE that's within an
726 * outer transaction, because for example the current transaction might
727 * have dropped the last index; then we'd think relhasindex should be
728 * cleared, but if the transaction later rolls back this would be wrong.
729 * So we refrain from updating the DDL flags if we're inside an outer
730 * transaction. This is OK since postponing the flag maintenance is
733 * This routine is shared by VACUUM and ANALYZE.
736 vac_update_relstats(Relation relation,
737 BlockNumber num_pages, double num_tuples,
738 BlockNumber num_all_visible_pages,
739 bool hasindex, TransactionId frozenxid,
740 MultiXactId minmulti,
743 Oid relid = RelationGetRelid(relation);
746 Form_pg_class pgcform;
749 rd = heap_open(RelationRelationId, RowExclusiveLock);
751 /* Fetch a copy of the tuple to scribble on */
752 ctup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
753 if (!HeapTupleIsValid(ctup))
754 elog(ERROR, "pg_class entry for relid %u vanished during vacuuming",
756 pgcform = (Form_pg_class) GETSTRUCT(ctup);
758 /* Apply statistical updates, if any, to copied tuple */
761 if (pgcform->relpages != (int32) num_pages)
763 pgcform->relpages = (int32) num_pages;
766 if (pgcform->reltuples != (float4) num_tuples)
768 pgcform->reltuples = (float4) num_tuples;
771 if (pgcform->relallvisible != (int32) num_all_visible_pages)
773 pgcform->relallvisible = (int32) num_all_visible_pages;
777 /* Apply DDL updates, but not inside an outer transaction (see above) */
782 * If we didn't find any indexes, reset relhasindex.
784 if (pgcform->relhasindex && !hasindex)
786 pgcform->relhasindex = false;
791 * If we have discovered that there are no indexes, then there's no
792 * primary key either. This could be done more thoroughly...
794 if (pgcform->relhaspkey && !hasindex)
796 pgcform->relhaspkey = false;
800 /* We also clear relhasrules and relhastriggers if needed */
801 if (pgcform->relhasrules && relation->rd_rules == NULL)
803 pgcform->relhasrules = false;
806 if (pgcform->relhastriggers && relation->trigdesc == NULL)
808 pgcform->relhastriggers = false;
814 * Update relfrozenxid, unless caller passed InvalidTransactionId
815 * indicating it has no new data.
817 * Ordinarily, we don't let relfrozenxid go backwards: if things are
818 * working correctly, the only way the new frozenxid could be older would
819 * be if a previous VACUUM was done with a tighter freeze_min_age, in
820 * which case we don't want to forget the work it already did. However,
821 * if the stored relfrozenxid is "in the future", then it must be corrupt
822 * and it seems best to overwrite it with the cutoff we used this time.
823 * This should match vac_update_datfrozenxid() concerning what we consider
824 * to be "in the future".
826 if (TransactionIdIsNormal(frozenxid) &&
827 pgcform->relfrozenxid != frozenxid &&
828 (TransactionIdPrecedes(pgcform->relfrozenxid, frozenxid) ||
829 TransactionIdPrecedes(ReadNewTransactionId(),
830 pgcform->relfrozenxid)))
832 pgcform->relfrozenxid = frozenxid;
836 /* Similarly for relminmxid */
837 if (MultiXactIdIsValid(minmulti) &&
838 pgcform->relminmxid != minmulti &&
839 (MultiXactIdPrecedes(pgcform->relminmxid, minmulti) ||
840 MultiXactIdPrecedes(ReadNextMultiXactId(), pgcform->relminmxid)))
842 pgcform->relminmxid = minmulti;
846 /* If anything changed, write out the tuple. */
848 heap_inplace_update(rd, ctup);
850 heap_close(rd, RowExclusiveLock);
855 * vac_update_datfrozenxid() -- update pg_database.datfrozenxid for our DB
857 * Update pg_database's datfrozenxid entry for our database to be the
858 * minimum of the pg_class.relfrozenxid values.
860 * Similarly, update our datminmxid to be the minimum of the
861 * pg_class.relminmxid values.
863 * If we are able to advance either pg_database value, also try to
864 * truncate pg_clog and pg_multixact.
866 * We violate transaction semantics here by overwriting the database's
867 * existing pg_database tuple with the new values. This is reasonably
868 * safe since the new values are correct whether or not this transaction
869 * commits. As with vac_update_relstats, this avoids leaving dead tuples
870 * behind after a VACUUM.
873 vac_update_datfrozenxid(void)
876 Form_pg_database dbform;
880 TransactionId newFrozenXid;
881 MultiXactId newMinMulti;
882 TransactionId lastSaneFrozenXid;
883 MultiXactId lastSaneMinMulti;
888 * Initialize the "min" calculation with GetOldestXmin, which is a
889 * reasonable approximation to the minimum relfrozenxid for not-yet-
890 * committed pg_class entries for new tables; see AddNewRelationTuple().
891 * So we cannot produce a wrong minimum by starting with this.
893 newFrozenXid = GetOldestXmin(NULL, true);
896 * Similarly, initialize the MultiXact "min" with the value that would be
897 * used on pg_class for new tables. See AddNewRelationTuple().
899 newMinMulti = GetOldestMultiXactId();
902 * Identify the latest relfrozenxid and relminmxid values that we could
903 * validly see during the scan. These are conservative values, but it's
904 * not really worth trying to be more exact.
906 lastSaneFrozenXid = ReadNewTransactionId();
907 lastSaneMinMulti = ReadNextMultiXactId();
910 * We must seqscan pg_class to find the minimum Xid, because there is no
911 * index that can help us here.
913 relation = heap_open(RelationRelationId, AccessShareLock);
915 scan = systable_beginscan(relation, InvalidOid, false,
918 while ((classTup = systable_getnext(scan)) != NULL)
920 Form_pg_class classForm = (Form_pg_class) GETSTRUCT(classTup);
923 * Only consider relations able to hold unfrozen XIDs (anything else
924 * should have InvalidTransactionId in relfrozenxid anyway.)
926 if (classForm->relkind != RELKIND_RELATION &&
927 classForm->relkind != RELKIND_MATVIEW &&
928 classForm->relkind != RELKIND_TOASTVALUE)
931 Assert(TransactionIdIsNormal(classForm->relfrozenxid));
932 Assert(MultiXactIdIsValid(classForm->relminmxid));
935 * If things are working properly, no relation should have a
936 * relfrozenxid or relminmxid that is "in the future". However, such
937 * cases have been known to arise due to bugs in pg_upgrade. If we
938 * see any entries that are "in the future", chicken out and don't do
939 * anything. This ensures we won't truncate clog before those
940 * relations have been scanned and cleaned up.
942 if (TransactionIdPrecedes(lastSaneFrozenXid, classForm->relfrozenxid) ||
943 MultiXactIdPrecedes(lastSaneMinMulti, classForm->relminmxid))
949 if (TransactionIdPrecedes(classForm->relfrozenxid, newFrozenXid))
950 newFrozenXid = classForm->relfrozenxid;
952 if (MultiXactIdPrecedes(classForm->relminmxid, newMinMulti))
953 newMinMulti = classForm->relminmxid;
956 /* we're done with pg_class */
957 systable_endscan(scan);
958 heap_close(relation, AccessShareLock);
960 /* chicken out if bogus data found */
964 Assert(TransactionIdIsNormal(newFrozenXid));
965 Assert(MultiXactIdIsValid(newMinMulti));
967 /* Now fetch the pg_database tuple we need to update. */
968 relation = heap_open(DatabaseRelationId, RowExclusiveLock);
970 /* Fetch a copy of the tuple to scribble on */
971 tuple = SearchSysCacheCopy1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
972 if (!HeapTupleIsValid(tuple))
973 elog(ERROR, "could not find tuple for database %u", MyDatabaseId);
974 dbform = (Form_pg_database) GETSTRUCT(tuple);
977 * As in vac_update_relstats(), we ordinarily don't want to let
978 * datfrozenxid go backward; but if it's "in the future" then it must be
979 * corrupt and it seems best to overwrite it.
981 if (dbform->datfrozenxid != newFrozenXid &&
982 (TransactionIdPrecedes(dbform->datfrozenxid, newFrozenXid) ||
983 TransactionIdPrecedes(lastSaneFrozenXid, dbform->datfrozenxid)))
985 dbform->datfrozenxid = newFrozenXid;
989 newFrozenXid = dbform->datfrozenxid;
991 /* Ditto for datminmxid */
992 if (dbform->datminmxid != newMinMulti &&
993 (MultiXactIdPrecedes(dbform->datminmxid, newMinMulti) ||
994 MultiXactIdPrecedes(lastSaneMinMulti, dbform->datminmxid)))
996 dbform->datminmxid = newMinMulti;
1000 newMinMulti = dbform->datminmxid;
1003 heap_inplace_update(relation, tuple);
1005 heap_freetuple(tuple);
1006 heap_close(relation, RowExclusiveLock);
1009 * If we were able to advance datfrozenxid or datminmxid, see if we can
1010 * truncate pg_clog and/or pg_multixact. Also do it if the shared
1011 * XID-wrap-limit info is stale, since this action will update that too.
1013 if (dirty || ForceTransactionIdLimitUpdate())
1014 vac_truncate_clog(newFrozenXid, newMinMulti,
1015 lastSaneFrozenXid, lastSaneMinMulti);
1020 * vac_truncate_clog() -- attempt to truncate the commit log
1022 * Scan pg_database to determine the system-wide oldest datfrozenxid,
1023 * and use it to truncate the transaction commit log (pg_clog).
1024 * Also update the XID wrap limit info maintained by varsup.c.
1025 * Likewise for datminmxid.
1027 * The passed frozenXID and minMulti are the updated values for my own
1028 * pg_database entry. They're used to initialize the "min" calculations.
1029 * The caller also passes the "last sane" XID and MXID, since it has
1030 * those at hand already.
1032 * This routine is only invoked when we've managed to change our
1033 * DB's datfrozenxid/datminmxid values, or we found that the shared
1034 * XID-wrap-limit info is stale.
1037 vac_truncate_clog(TransactionId frozenXID,
1038 MultiXactId minMulti,
1039 TransactionId lastSaneFrozenXid,
1040 MultiXactId lastSaneMinMulti)
1042 TransactionId myXID = GetCurrentTransactionId();
1046 Oid oldestxid_datoid;
1047 Oid minmulti_datoid;
1049 bool frozenAlreadyWrapped = false;
1051 /* init oldest datoids to sync with my frozenXID/minMulti values */
1052 oldestxid_datoid = MyDatabaseId;
1053 minmulti_datoid = MyDatabaseId;
1056 * Scan pg_database to compute the minimum datfrozenxid/datminmxid
1058 * Note: we need not worry about a race condition with new entries being
1059 * inserted by CREATE DATABASE. Any such entry will have a copy of some
1060 * existing DB's datfrozenxid, and that source DB cannot be ours because
1061 * of the interlock against copying a DB containing an active backend.
1062 * Hence the new entry will not reduce the minimum. Also, if two VACUUMs
1063 * concurrently modify the datfrozenxid's of different databases, the
1064 * worst possible outcome is that pg_clog is not truncated as aggressively
1067 relation = heap_open(DatabaseRelationId, AccessShareLock);
1069 scan = heap_beginscan_catalog(relation, 0, NULL);
1071 while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
1073 Form_pg_database dbform = (Form_pg_database) GETSTRUCT(tuple);
1075 Assert(TransactionIdIsNormal(dbform->datfrozenxid));
1076 Assert(MultiXactIdIsValid(dbform->datminmxid));
1079 * If things are working properly, no database should have a
1080 * datfrozenxid or datminmxid that is "in the future". However, such
1081 * cases have been known to arise due to bugs in pg_upgrade. If we
1082 * see any entries that are "in the future", chicken out and don't do
1083 * anything. This ensures we won't truncate clog before those
1084 * databases have been scanned and cleaned up. (We will issue the
1085 * "already wrapped" warning if appropriate, though.)
1087 if (TransactionIdPrecedes(lastSaneFrozenXid, dbform->datfrozenxid) ||
1088 MultiXactIdPrecedes(lastSaneMinMulti, dbform->datminmxid))
1091 if (TransactionIdPrecedes(myXID, dbform->datfrozenxid))
1092 frozenAlreadyWrapped = true;
1093 else if (TransactionIdPrecedes(dbform->datfrozenxid, frozenXID))
1095 frozenXID = dbform->datfrozenxid;
1096 oldestxid_datoid = HeapTupleGetOid(tuple);
1099 if (MultiXactIdPrecedes(dbform->datminmxid, minMulti))
1101 minMulti = dbform->datminmxid;
1102 minmulti_datoid = HeapTupleGetOid(tuple);
1108 heap_close(relation, AccessShareLock);
1111 * Do not truncate CLOG if we seem to have suffered wraparound already;
1112 * the computed minimum XID might be bogus. This case should now be
1113 * impossible due to the defenses in GetNewTransactionId, but we keep the
1116 if (frozenAlreadyWrapped)
1119 (errmsg("some databases have not been vacuumed in over 2 billion transactions"),
1120 errdetail("You might have already suffered transaction-wraparound data loss.")));
1124 /* chicken out if data is bogus in any other way */
1129 * Truncate CLOG and CommitTs to the oldest computed value.
1130 * Note we don't truncate multixacts; that will be done by the next
1133 TruncateCLOG(frozenXID);
1134 TruncateCommitTs(frozenXID, true);
1137 * Update the wrap limit for GetNewTransactionId and creation of new
1138 * MultiXactIds. Note: these functions will also signal the postmaster
1139 * for an(other) autovac cycle if needed. XXX should we avoid possibly
1142 SetTransactionIdLimit(frozenXID, oldestxid_datoid);
1143 SetMultiXactIdLimit(minMulti, minmulti_datoid);
1144 AdvanceOldestCommitTs(frozenXID);
1149 * vacuum_rel() -- vacuum one heap relation
1151 * Doing one heap at a time incurs extra overhead, since we need to
1152 * check that the heap exists again just before we vacuum it. The
1153 * reason that we do this is so that vacuuming can be spread across
1154 * many small transactions. Otherwise, two-phase locking would require
1155 * us to lock the entire database during one pass of the vacuum cleaner.
1157 * At entry and exit, we are not inside a transaction.
1160 vacuum_rel(Oid relid, RangeVar *relation, int options, VacuumParams *params)
1167 int save_sec_context;
1170 Assert(params != NULL);
1172 /* Begin a transaction for vacuuming this relation */
1173 StartTransactionCommand();
1176 * Functions in indexes may want a snapshot set. Also, setting a snapshot
1177 * ensures that RecentGlobalXmin is kept truly recent.
1179 PushActiveSnapshot(GetTransactionSnapshot());
1181 if (!(options & VACOPT_FULL))
1184 * In lazy vacuum, we can set the PROC_IN_VACUUM flag, which lets
1185 * other concurrent VACUUMs know that they can ignore this one while
1186 * determining their OldestXmin. (The reason we don't set it during a
1187 * full VACUUM is exactly that we may have to run user-defined
1188 * functions for functional indexes, and we want to make sure that if
1189 * they use the snapshot set above, any tuples it requires can't get
1190 * removed from other tables. An index function that depends on the
1191 * contents of other tables is arguably broken, but we won't break it
1192 * here by violating transaction semantics.)
1194 * We also set the VACUUM_FOR_WRAPAROUND flag, which is passed down by
1195 * autovacuum; it's used to avoid canceling a vacuum that was invoked
1198 * Note: these flags remain set until CommitTransaction or
1199 * AbortTransaction. We don't want to clear them until we reset
1200 * MyPgXact->xid/xmin, else OldestXmin might appear to go backwards,
1201 * which is probably Not Good.
1203 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1204 MyPgXact->vacuumFlags |= PROC_IN_VACUUM;
1205 if (params->is_wraparound)
1206 MyPgXact->vacuumFlags |= PROC_VACUUM_FOR_WRAPAROUND;
1207 LWLockRelease(ProcArrayLock);
1211 * Check for user-requested abort. Note we want this to be inside a
1212 * transaction, so xact.c doesn't issue useless WARNING.
1214 CHECK_FOR_INTERRUPTS();
1217 * Determine the type of lock we want --- hard exclusive lock for a FULL
1218 * vacuum, but just ShareUpdateExclusiveLock for concurrent vacuum. Either
1219 * way, we can be sure that no other backend is vacuuming the same table.
1221 lmode = (options & VACOPT_FULL) ? AccessExclusiveLock : ShareUpdateExclusiveLock;
1224 * Open the relation and get the appropriate lock on it.
1226 * There's a race condition here: the rel may have gone away since the
1227 * last time we saw it. If so, we don't need to vacuum it.
1229 * If we've been asked not to wait for the relation lock, acquire it first
1230 * in non-blocking mode, before calling try_relation_open().
1232 if (!(options & VACOPT_NOWAIT))
1233 onerel = try_relation_open(relid, lmode);
1234 else if (ConditionalLockRelationOid(relid, lmode))
1235 onerel = try_relation_open(relid, NoLock);
1239 if (IsAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
1241 (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
1242 errmsg("skipping vacuum of \"%s\" --- lock not available",
1243 relation->relname)));
1248 PopActiveSnapshot();
1249 CommitTransactionCommand();
1254 * Check permissions.
1256 * We allow the user to vacuum a table if he is superuser, the table
1257 * owner, or the database owner (but in the latter case, only if it's not
1258 * a shared relation). pg_class_ownercheck includes the superuser case.
1260 * Note we choose to treat permissions failure as a WARNING and keep
1261 * trying to vacuum the rest of the DB --- is this appropriate?
1263 if (!(pg_class_ownercheck(RelationGetRelid(onerel), GetUserId()) ||
1264 (pg_database_ownercheck(MyDatabaseId, GetUserId()) && !onerel->rd_rel->relisshared)))
1266 if (onerel->rd_rel->relisshared)
1268 (errmsg("skipping \"%s\" --- only superuser can vacuum it",
1269 RelationGetRelationName(onerel))));
1270 else if (onerel->rd_rel->relnamespace == PG_CATALOG_NAMESPACE)
1272 (errmsg("skipping \"%s\" --- only superuser or database owner can vacuum it",
1273 RelationGetRelationName(onerel))));
1276 (errmsg("skipping \"%s\" --- only table or database owner can vacuum it",
1277 RelationGetRelationName(onerel))));
1278 relation_close(onerel, lmode);
1279 PopActiveSnapshot();
1280 CommitTransactionCommand();
1285 * Check that it's a vacuumable relation; we used to do this in
1286 * get_rel_oids() but seems safer to check after we've locked the
1289 if (onerel->rd_rel->relkind != RELKIND_RELATION &&
1290 onerel->rd_rel->relkind != RELKIND_MATVIEW &&
1291 onerel->rd_rel->relkind != RELKIND_TOASTVALUE)
1294 (errmsg("skipping \"%s\" --- cannot vacuum non-tables or special system tables",
1295 RelationGetRelationName(onerel))));
1296 relation_close(onerel, lmode);
1297 PopActiveSnapshot();
1298 CommitTransactionCommand();
1303 * Silently ignore tables that are temp tables of other backends ---
1304 * trying to vacuum these will lead to great unhappiness, since their
1305 * contents are probably not up-to-date on disk. (We don't throw a
1306 * warning here; it would just lead to chatter during a database-wide
1309 if (RELATION_IS_OTHER_TEMP(onerel))
1311 relation_close(onerel, lmode);
1312 PopActiveSnapshot();
1313 CommitTransactionCommand();
1318 * Get a session-level lock too. This will protect our access to the
1319 * relation across multiple transactions, so that we can vacuum the
1320 * relation's TOAST table (if any) secure in the knowledge that no one is
1321 * deleting the parent relation.
1323 * NOTE: this cannot block, even if someone else is waiting for access,
1324 * because the lock manager knows that both lock requests are from the
1327 onerelid = onerel->rd_lockInfo.lockRelId;
1328 LockRelationIdForSession(&onerelid, lmode);
1331 * Remember the relation's TOAST relation for later, if the caller asked
1332 * us to process it. In VACUUM FULL, though, the toast table is
1333 * automatically rebuilt by cluster_rel so we shouldn't recurse to it.
1335 if (!(options & VACOPT_SKIPTOAST) && !(options & VACOPT_FULL))
1336 toast_relid = onerel->rd_rel->reltoastrelid;
1338 toast_relid = InvalidOid;
1341 * Switch to the table owner's userid, so that any index functions are run
1342 * as that user. Also lock down security-restricted operations and
1343 * arrange to make GUC variable changes local to this command. (This is
1344 * unnecessary, but harmless, for lazy VACUUM.)
1346 GetUserIdAndSecContext(&save_userid, &save_sec_context);
1347 SetUserIdAndSecContext(onerel->rd_rel->relowner,
1348 save_sec_context | SECURITY_RESTRICTED_OPERATION);
1349 save_nestlevel = NewGUCNestLevel();
1352 * Do the actual work --- either FULL or "lazy" vacuum
1354 if (options & VACOPT_FULL)
1356 /* close relation before vacuuming, but hold lock until commit */
1357 relation_close(onerel, NoLock);
1360 /* VACUUM FULL is now a variant of CLUSTER; see cluster.c */
1361 cluster_rel(relid, InvalidOid, false,
1362 (options & VACOPT_VERBOSE) != 0);
1365 lazy_vacuum_rel(onerel, options, params, vac_strategy);
1367 /* Roll back any GUC changes executed by index functions */
1368 AtEOXact_GUC(false, save_nestlevel);
1370 /* Restore userid and security context */
1371 SetUserIdAndSecContext(save_userid, save_sec_context);
1373 /* all done with this class, but hold lock until commit */
1375 relation_close(onerel, NoLock);
1378 * Complete the transaction and free all temporary memory used.
1380 PopActiveSnapshot();
1381 CommitTransactionCommand();
1384 * If the relation has a secondary toast rel, vacuum that too while we
1385 * still hold the session lock on the master table. Note however that
1386 * "analyze" will not get done on the toast table. This is good, because
1387 * the toaster always uses hardcoded index access and statistics are
1388 * totally unimportant for toast relations.
1390 if (toast_relid != InvalidOid)
1391 vacuum_rel(toast_relid, relation, options, params);
1394 * Now release the session-level lock on the master table.
1396 UnlockRelationIdForSession(&onerelid, lmode);
1398 /* Report that we really did it. */
1404 * Open all the vacuumable indexes of the given relation, obtaining the
1405 * specified kind of lock on each. Return an array of Relation pointers for
1406 * the indexes into *Irel, and the number of indexes into *nindexes.
1408 * We consider an index vacuumable if it is marked insertable (IndexIsReady).
1409 * If it isn't, probably a CREATE INDEX CONCURRENTLY command failed early in
1410 * execution, and what we have is too corrupt to be processable. We will
1411 * vacuum even if the index isn't indisvalid; this is important because in a
1412 * unique index, uniqueness checks will be performed anyway and had better not
1413 * hit dangling index pointers.
1416 vac_open_indexes(Relation relation, LOCKMODE lockmode,
1417 int *nindexes, Relation **Irel)
1420 ListCell *indexoidscan;
1423 Assert(lockmode != NoLock);
1425 indexoidlist = RelationGetIndexList(relation);
1427 /* allocate enough memory for all indexes */
1428 i = list_length(indexoidlist);
1431 *Irel = (Relation *) palloc(i * sizeof(Relation));
1435 /* collect just the ready indexes */
1437 foreach(indexoidscan, indexoidlist)
1439 Oid indexoid = lfirst_oid(indexoidscan);
1442 indrel = index_open(indexoid, lockmode);
1443 if (IndexIsReady(indrel->rd_index))
1444 (*Irel)[i++] = indrel;
1446 index_close(indrel, lockmode);
1451 list_free(indexoidlist);
1455 * Release the resources acquired by vac_open_indexes. Optionally release
1456 * the locks (say NoLock to keep 'em).
1459 vac_close_indexes(int nindexes, Relation *Irel, LOCKMODE lockmode)
1466 Relation ind = Irel[nindexes];
1468 index_close(ind, lockmode);
1474 * vacuum_delay_point --- check for interrupts and cost-based delay.
1476 * This should be called in each major loop of VACUUM processing,
1477 * typically once per page processed.
1480 vacuum_delay_point(void)
1482 /* Always check for interrupts */
1483 CHECK_FOR_INTERRUPTS();
1485 /* Nap if appropriate */
1486 if (VacuumCostActive && !InterruptPending &&
1487 VacuumCostBalance >= VacuumCostLimit)
1491 msec = VacuumCostDelay * VacuumCostBalance / VacuumCostLimit;
1492 if (msec > VacuumCostDelay * 4)
1493 msec = VacuumCostDelay * 4;
1495 pg_usleep(msec * 1000L);
1497 VacuumCostBalance = 0;
1499 /* update balance values for workers */
1500 AutoVacuumUpdateDelay();
1502 /* Might have gotten an interrupt while sleeping */
1503 CHECK_FOR_INTERRUPTS();