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 /* Now go through the common routine */
118 vacuum(vacstmt->options, vacstmt->relation, InvalidOid, ¶ms,
119 vacstmt->va_cols, NULL, isTopLevel);
123 * Primary entry point for VACUUM and ANALYZE commands.
125 * options is a bitmask of VacuumOption flags, indicating what to do.
127 * relid, if not InvalidOid, indicate the relation to process; otherwise,
128 * the RangeVar is used. (The latter must always be passed, because it's
129 * used for error messages.)
131 * params contains a set of parameters that can be used to customize the
134 * va_cols is a list of columns to analyze, or NIL to process them all.
136 * bstrategy is normally given as NULL, but in autovacuum it can be passed
137 * in to use the same buffer strategy object across multiple vacuum() calls.
139 * isTopLevel should be passed down from ProcessUtility.
141 * It is the caller's responsibility that all parameters are allocated in a
142 * memory context that will not disappear at transaction commit.
145 vacuum(int options, RangeVar *relation, Oid relid, VacuumParams *params,
146 List *va_cols, BufferAccessStrategy bstrategy, bool isTopLevel)
148 const char *stmttype;
149 volatile bool in_outer_xact,
152 static bool in_vacuum = false;
154 Assert(params != NULL);
156 stmttype = (options & VACOPT_VACUUM) ? "VACUUM" : "ANALYZE";
159 * We cannot run VACUUM inside a user transaction block; if we were inside
160 * a transaction, then our commit- and start-transaction-command calls
161 * would not have the intended effect! There are numerous other subtle
162 * dependencies on this, too.
164 * ANALYZE (without VACUUM) can run either way.
166 if (options & VACOPT_VACUUM)
168 PreventTransactionChain(isTopLevel, stmttype);
169 in_outer_xact = false;
172 in_outer_xact = IsInTransactionChain(isTopLevel);
175 * Due to static variables vac_context, anl_context and vac_strategy,
176 * vacuum() is not reentrant. This matters when VACUUM FULL or ANALYZE
177 * calls a hostile index expression that itself calls ANALYZE.
180 elog(ERROR, "%s cannot be executed from VACUUM or ANALYZE", stmttype);
183 * Send info about dead objects to the statistics collector, unless we are
184 * in autovacuum --- autovacuum.c does this for itself.
186 if ((options & VACOPT_VACUUM) && !IsAutoVacuumWorkerProcess())
187 pgstat_vacuum_stat();
190 * Create special memory context for cross-transaction storage.
192 * Since it is a child of PortalContext, it will go away eventually even
193 * if we suffer an error; there's no need for special abort cleanup logic.
195 vac_context = AllocSetContextCreate(PortalContext,
197 ALLOCSET_DEFAULT_MINSIZE,
198 ALLOCSET_DEFAULT_INITSIZE,
199 ALLOCSET_DEFAULT_MAXSIZE);
202 * If caller didn't give us a buffer strategy object, make one in the
203 * cross-transaction memory context.
205 if (bstrategy == NULL)
207 MemoryContext old_context = MemoryContextSwitchTo(vac_context);
209 bstrategy = GetAccessStrategy(BAS_VACUUM);
210 MemoryContextSwitchTo(old_context);
212 vac_strategy = bstrategy;
215 * Build list of relations to process, unless caller gave us one. (If we
216 * build one, we put it in vac_context for safekeeping.)
218 relations = get_rel_oids(relid, relation);
221 * Decide whether we need to start/commit our own transactions.
223 * For VACUUM (with or without ANALYZE): always do so, so that we can
224 * release locks as soon as possible. (We could possibly use the outer
225 * transaction for a one-table VACUUM, but handling TOAST tables would be
228 * For ANALYZE (no VACUUM): if inside a transaction block, we cannot
229 * start/commit our own transactions. Also, there's no need to do so if
230 * only processing one relation. For multiple relations when not within a
231 * transaction block, and also in an autovacuum worker, use own
232 * transactions so we can release locks sooner.
234 if (options & VACOPT_VACUUM)
235 use_own_xacts = true;
238 Assert(options & VACOPT_ANALYZE);
239 if (IsAutoVacuumWorkerProcess())
240 use_own_xacts = true;
241 else if (in_outer_xact)
242 use_own_xacts = false;
243 else if (list_length(relations) > 1)
244 use_own_xacts = true;
246 use_own_xacts = false;
250 * vacuum_rel expects to be entered with no transaction active; it will
251 * start and commit its own transaction. But we are called by an SQL
252 * command, and so we are executing inside a transaction already. We
253 * commit the transaction started in PostgresMain() here, and start
254 * another one before exiting to match the commit waiting for us back in
259 Assert(!in_outer_xact);
261 /* ActiveSnapshot is not set by autovacuum */
262 if (ActiveSnapshotSet())
265 /* matches the StartTransaction in PostgresMain() */
266 CommitTransactionCommand();
269 /* Turn vacuum cost accounting on or off */
275 VacuumCostActive = (VacuumCostDelay > 0);
276 VacuumCostBalance = 0;
282 * Loop to process each selected relation.
284 foreach(cur, relations)
286 Oid relid = lfirst_oid(cur);
288 if (options & VACOPT_VACUUM)
290 if (!vacuum_rel(relid, relation, options, params))
294 if (options & VACOPT_ANALYZE)
297 * If using separate xacts, start one for analyze. Otherwise,
298 * we can use the outer transaction.
302 StartTransactionCommand();
303 /* functions in indexes may want a snapshot set */
304 PushActiveSnapshot(GetTransactionSnapshot());
307 analyze_rel(relid, relation, options,
308 va_cols, in_outer_xact, vac_strategy);
313 CommitTransactionCommand();
321 VacuumCostActive = false;
327 VacuumCostActive = false;
330 * Finish up processing.
334 /* here, we are not in a transaction */
337 * This matches the CommitTransaction waiting for us in
340 StartTransactionCommand();
343 if ((options & VACOPT_VACUUM) && !IsAutoVacuumWorkerProcess())
346 * Update pg_database.datfrozenxid, and truncate pg_clog if possible.
347 * (autovacuum.c does this for itself.)
349 vac_update_datfrozenxid();
353 * Clean up working storage --- note we must do this after
354 * StartTransactionCommand, else we might be trying to delete the active
357 MemoryContextDelete(vac_context);
362 * Build a list of Oids for each relation to be processed
364 * The list is built in vac_context so that it will survive across our
365 * per-relation transactions.
368 get_rel_oids(Oid relid, const RangeVar *vacrel)
370 List *oid_list = NIL;
371 MemoryContext oldcontext;
373 /* OID supplied by VACUUM's caller? */
374 if (OidIsValid(relid))
376 oldcontext = MemoryContextSwitchTo(vac_context);
377 oid_list = lappend_oid(oid_list, relid);
378 MemoryContextSwitchTo(oldcontext);
382 /* Process a specific relation */
386 * Since we don't take a lock here, the relation might be gone, or the
387 * RangeVar might no longer refer to the OID we look up here. In the
388 * former case, VACUUM will do nothing; in the latter case, it will
389 * process the OID we looked up here, rather than the new one. Neither
390 * is ideal, but there's little practical alternative, since we're
391 * going to commit this transaction and begin a new one between now
394 relid = RangeVarGetRelid(vacrel, NoLock, false);
396 /* Make a relation list entry for this guy */
397 oldcontext = MemoryContextSwitchTo(vac_context);
398 oid_list = lappend_oid(oid_list, relid);
399 MemoryContextSwitchTo(oldcontext);
404 * Process all plain relations and materialized views listed in
411 pgclass = heap_open(RelationRelationId, AccessShareLock);
413 scan = heap_beginscan_catalog(pgclass, 0, NULL);
415 while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
417 Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
419 if (classForm->relkind != RELKIND_RELATION &&
420 classForm->relkind != RELKIND_MATVIEW)
423 /* Make a relation list entry for this guy */
424 oldcontext = MemoryContextSwitchTo(vac_context);
425 oid_list = lappend_oid(oid_list, HeapTupleGetOid(tuple));
426 MemoryContextSwitchTo(oldcontext);
430 heap_close(pgclass, AccessShareLock);
437 * vacuum_set_xid_limits() -- compute oldest-Xmin and freeze cutoff points
439 * The output parameters are:
440 * - oldestXmin is the cutoff value used to distinguish whether tuples are
441 * DEAD or RECENTLY_DEAD (see HeapTupleSatisfiesVacuum).
442 * - freezeLimit is the Xid below which all Xids are replaced by
443 * FrozenTransactionId during vacuum.
444 * - xidFullScanLimit (computed from table_freeze_age parameter)
445 * represents a minimum Xid value; a table whose relfrozenxid is older than
446 * this will have a full-table vacuum applied to it, to freeze tuples across
447 * the whole table. Vacuuming a table younger than this value can use a
449 * - multiXactCutoff is the value below which all MultiXactIds are removed from
451 * - mxactFullScanLimit is a value against which a table's relminmxid value is
452 * compared to produce a full-table vacuum, as with xidFullScanLimit.
454 * xidFullScanLimit and mxactFullScanLimit can be passed as NULL if caller is
458 vacuum_set_xid_limits(Relation rel,
460 int freeze_table_age,
461 int multixact_freeze_min_age,
462 int multixact_freeze_table_age,
463 TransactionId *oldestXmin,
464 TransactionId *freezeLimit,
465 TransactionId *xidFullScanLimit,
466 MultiXactId *multiXactCutoff,
467 MultiXactId *mxactFullScanLimit)
472 TransactionId safeLimit;
473 MultiXactId mxactLimit;
474 MultiXactId safeMxactLimit;
477 * We can always ignore processes running lazy vacuum. This is because we
478 * use these values only for deciding which tuples we must keep in the
479 * tables. Since lazy vacuum doesn't write its XID anywhere, it's safe to
480 * ignore it. In theory it could be problematic to ignore lazy vacuums in
481 * a full vacuum, but keep in mind that only one vacuum process can be
482 * working on a particular table at any time, and that each vacuum is
483 * always an independent transaction.
485 *oldestXmin = GetOldestXmin(rel, true);
487 Assert(TransactionIdIsNormal(*oldestXmin));
490 * Determine the minimum freeze age to use: as specified by the caller, or
491 * vacuum_freeze_min_age, but in any case not more than half
492 * autovacuum_freeze_max_age, so that autovacuums to prevent XID
493 * wraparound won't occur too frequently.
495 freezemin = freeze_min_age;
497 freezemin = vacuum_freeze_min_age;
498 freezemin = Min(freezemin, autovacuum_freeze_max_age / 2);
499 Assert(freezemin >= 0);
502 * Compute the cutoff XID, being careful not to generate a "permanent" XID
504 limit = *oldestXmin - freezemin;
505 if (!TransactionIdIsNormal(limit))
506 limit = FirstNormalTransactionId;
509 * If oldestXmin is very far back (in practice, more than
510 * autovacuum_freeze_max_age / 2 XIDs old), complain and force a minimum
511 * freeze age of zero.
513 safeLimit = ReadNewTransactionId() - autovacuum_freeze_max_age;
514 if (!TransactionIdIsNormal(safeLimit))
515 safeLimit = FirstNormalTransactionId;
517 if (TransactionIdPrecedes(limit, safeLimit))
520 (errmsg("oldest xmin is far in the past"),
521 errhint("Close open transactions soon to avoid wraparound problems.")));
525 *freezeLimit = limit;
528 * Determine the minimum multixact freeze age to use: as specified by
529 * caller, or vacuum_multixact_freeze_min_age, but in any case not more
530 * than half autovacuum_multixact_freeze_max_age, so that autovacuums to
531 * prevent MultiXact wraparound won't occur too frequently.
533 mxid_freezemin = multixact_freeze_min_age;
534 if (mxid_freezemin < 0)
535 mxid_freezemin = vacuum_multixact_freeze_min_age;
536 mxid_freezemin = Min(mxid_freezemin,
537 autovacuum_multixact_freeze_max_age / 2);
538 Assert(mxid_freezemin >= 0);
540 /* compute the cutoff multi, being careful to generate a valid value */
541 mxactLimit = GetOldestMultiXactId() - mxid_freezemin;
542 if (mxactLimit < FirstMultiXactId)
543 mxactLimit = FirstMultiXactId;
546 ReadNextMultiXactId() - autovacuum_multixact_freeze_max_age;
547 if (safeMxactLimit < FirstMultiXactId)
548 safeMxactLimit = FirstMultiXactId;
550 if (MultiXactIdPrecedes(mxactLimit, safeMxactLimit))
553 (errmsg("oldest multixact is far in the past"),
554 errhint("Close open transactions with multixacts soon to avoid wraparound problems.")));
555 mxactLimit = safeMxactLimit;
558 *multiXactCutoff = mxactLimit;
560 if (xidFullScanLimit != NULL)
564 Assert(mxactFullScanLimit != NULL);
567 * Determine the table freeze age to use: as specified by the caller,
568 * or vacuum_freeze_table_age, but in any case not more than
569 * autovacuum_freeze_max_age * 0.95, so that if you have e.g nightly
570 * VACUUM schedule, the nightly VACUUM gets a chance to freeze tuples
571 * before anti-wraparound autovacuum is launched.
573 freezetable = freeze_table_age;
575 freezetable = vacuum_freeze_table_age;
576 freezetable = Min(freezetable, autovacuum_freeze_max_age * 0.95);
577 Assert(freezetable >= 0);
580 * Compute XID limit causing a full-table vacuum, being careful not to
581 * generate a "permanent" XID.
583 limit = ReadNewTransactionId() - freezetable;
584 if (!TransactionIdIsNormal(limit))
585 limit = FirstNormalTransactionId;
587 *xidFullScanLimit = limit;
590 * Similar to the above, determine the table freeze age to use for
591 * multixacts: as specified by the caller, or
592 * vacuum_multixact_freeze_table_age, but in any case not more than
593 * autovacuum_multixact_freeze_table_age * 0.95, so that if you have
594 * e.g. nightly VACUUM schedule, the nightly VACUUM gets a chance to
595 * freeze multixacts before anti-wraparound autovacuum is launched.
597 freezetable = multixact_freeze_table_age;
599 freezetable = vacuum_multixact_freeze_table_age;
600 freezetable = Min(freezetable,
601 autovacuum_multixact_freeze_max_age * 0.95);
602 Assert(freezetable >= 0);
605 * Compute MultiXact limit causing a full-table vacuum, being careful
606 * to generate a valid MultiXact value.
608 mxactLimit = ReadNextMultiXactId() - freezetable;
609 if (mxactLimit < FirstMultiXactId)
610 mxactLimit = FirstMultiXactId;
612 *mxactFullScanLimit = mxactLimit;
616 Assert(mxactFullScanLimit == NULL);
621 * vac_estimate_reltuples() -- estimate the new value for pg_class.reltuples
623 * If we scanned the whole relation then we should just use the count of
624 * live tuples seen; but if we did not, we should not trust the count
625 * unreservedly, especially not in VACUUM, which may have scanned a quite
626 * nonrandom subset of the table. When we have only partial information,
627 * we take the old value of pg_class.reltuples as a measurement of the
628 * tuple density in the unscanned pages.
630 * This routine is shared by VACUUM and ANALYZE.
633 vac_estimate_reltuples(Relation relation, bool is_analyze,
634 BlockNumber total_pages,
635 BlockNumber scanned_pages,
636 double scanned_tuples)
638 BlockNumber old_rel_pages = relation->rd_rel->relpages;
639 double old_rel_tuples = relation->rd_rel->reltuples;
643 double updated_density;
645 /* If we did scan the whole table, just use the count as-is */
646 if (scanned_pages >= total_pages)
647 return scanned_tuples;
650 * If scanned_pages is zero but total_pages isn't, keep the existing value
651 * of reltuples. (Note: callers should avoid updating the pg_class
652 * statistics in this situation, since no new information has been
655 if (scanned_pages == 0)
656 return old_rel_tuples;
659 * If old value of relpages is zero, old density is indeterminate; we
660 * can't do much except scale up scanned_tuples to match total_pages.
662 if (old_rel_pages == 0)
663 return floor((scanned_tuples / scanned_pages) * total_pages + 0.5);
666 * Okay, we've covered the corner cases. The normal calculation is to
667 * convert the old measurement to a density (tuples per page), then update
668 * the density using an exponential-moving-average approach, and finally
669 * compute reltuples as updated_density * total_pages.
671 * For ANALYZE, the moving average multiplier is just the fraction of the
672 * table's pages we scanned. This is equivalent to assuming that the
673 * tuple density in the unscanned pages didn't change. Of course, it
674 * probably did, if the new density measurement is different. But over
675 * repeated cycles, the value of reltuples will converge towards the
676 * correct value, if repeated measurements show the same new density.
678 * For VACUUM, the situation is a bit different: we have looked at a
679 * nonrandom sample of pages, but we know for certain that the pages we
680 * didn't look at are precisely the ones that haven't changed lately.
681 * Thus, there is a reasonable argument for doing exactly the same thing
682 * as for the ANALYZE case, that is use the old density measurement as the
683 * value for the unscanned pages.
685 * This logic could probably use further refinement.
687 old_density = old_rel_tuples / old_rel_pages;
688 new_density = scanned_tuples / scanned_pages;
689 multiplier = (double) scanned_pages / (double) total_pages;
690 updated_density = old_density + (new_density - old_density) * multiplier;
691 return floor(updated_density * total_pages + 0.5);
696 * vac_update_relstats() -- update statistics for one relation
698 * Update the whole-relation statistics that are kept in its pg_class
699 * row. There are additional stats that will be updated if we are
700 * doing ANALYZE, but we always update these stats. This routine works
701 * for both index and heap relation entries in pg_class.
703 * We violate transaction semantics here by overwriting the rel's
704 * existing pg_class tuple with the new values. This is reasonably
705 * safe as long as we're sure that the new values are correct whether or
706 * not this transaction commits. The reason for doing this is that if
707 * we updated these tuples in the usual way, vacuuming pg_class itself
708 * wouldn't work very well --- by the time we got done with a vacuum
709 * cycle, most of the tuples in pg_class would've been obsoleted. Of
710 * course, this only works for fixed-size not-null columns, but these are.
712 * Another reason for doing it this way is that when we are in a lazy
713 * VACUUM and have PROC_IN_VACUUM set, we mustn't do any regular updates.
714 * Somebody vacuuming pg_class might think they could delete a tuple
715 * marked with xmin = our xid.
717 * In addition to fundamentally nontransactional statistics such as
718 * relpages and relallvisible, we try to maintain certain lazily-updated
719 * DDL flags such as relhasindex, by clearing them if no longer correct.
720 * It's safe to do this in VACUUM, which can't run in parallel with
721 * CREATE INDEX/RULE/TRIGGER and can't be part of a transaction block.
722 * However, it's *not* safe to do it in an ANALYZE that's within an
723 * outer transaction, because for example the current transaction might
724 * have dropped the last index; then we'd think relhasindex should be
725 * cleared, but if the transaction later rolls back this would be wrong.
726 * So we refrain from updating the DDL flags if we're inside an outer
727 * transaction. This is OK since postponing the flag maintenance is
730 * This routine is shared by VACUUM and ANALYZE.
733 vac_update_relstats(Relation relation,
734 BlockNumber num_pages, double num_tuples,
735 BlockNumber num_all_visible_pages,
736 bool hasindex, TransactionId frozenxid,
737 MultiXactId minmulti,
740 Oid relid = RelationGetRelid(relation);
743 Form_pg_class pgcform;
746 rd = heap_open(RelationRelationId, RowExclusiveLock);
748 /* Fetch a copy of the tuple to scribble on */
749 ctup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
750 if (!HeapTupleIsValid(ctup))
751 elog(ERROR, "pg_class entry for relid %u vanished during vacuuming",
753 pgcform = (Form_pg_class) GETSTRUCT(ctup);
755 /* Apply statistical updates, if any, to copied tuple */
758 if (pgcform->relpages != (int32) num_pages)
760 pgcform->relpages = (int32) num_pages;
763 if (pgcform->reltuples != (float4) num_tuples)
765 pgcform->reltuples = (float4) num_tuples;
768 if (pgcform->relallvisible != (int32) num_all_visible_pages)
770 pgcform->relallvisible = (int32) num_all_visible_pages;
774 /* Apply DDL updates, but not inside an outer transaction (see above) */
779 * If we didn't find any indexes, reset relhasindex.
781 if (pgcform->relhasindex && !hasindex)
783 pgcform->relhasindex = false;
788 * If we have discovered that there are no indexes, then there's no
789 * primary key either. This could be done more thoroughly...
791 if (pgcform->relhaspkey && !hasindex)
793 pgcform->relhaspkey = false;
797 /* We also clear relhasrules and relhastriggers if needed */
798 if (pgcform->relhasrules && relation->rd_rules == NULL)
800 pgcform->relhasrules = false;
803 if (pgcform->relhastriggers && relation->trigdesc == NULL)
805 pgcform->relhastriggers = false;
811 * Update relfrozenxid, unless caller passed InvalidTransactionId
812 * indicating it has no new data.
814 * Ordinarily, we don't let relfrozenxid go backwards: if things are
815 * working correctly, the only way the new frozenxid could be older would
816 * be if a previous VACUUM was done with a tighter freeze_min_age, in
817 * which case we don't want to forget the work it already did. However,
818 * if the stored relfrozenxid is "in the future", then it must be corrupt
819 * and it seems best to overwrite it with the cutoff we used this time.
820 * This should match vac_update_datfrozenxid() concerning what we consider
821 * to be "in the future".
823 if (TransactionIdIsNormal(frozenxid) &&
824 pgcform->relfrozenxid != frozenxid &&
825 (TransactionIdPrecedes(pgcform->relfrozenxid, frozenxid) ||
826 TransactionIdPrecedes(ReadNewTransactionId(),
827 pgcform->relfrozenxid)))
829 pgcform->relfrozenxid = frozenxid;
833 /* Similarly for relminmxid */
834 if (MultiXactIdIsValid(minmulti) &&
835 pgcform->relminmxid != minmulti &&
836 (MultiXactIdPrecedes(pgcform->relminmxid, minmulti) ||
837 MultiXactIdPrecedes(ReadNextMultiXactId(), pgcform->relminmxid)))
839 pgcform->relminmxid = minmulti;
843 /* If anything changed, write out the tuple. */
845 heap_inplace_update(rd, ctup);
847 heap_close(rd, RowExclusiveLock);
852 * vac_update_datfrozenxid() -- update pg_database.datfrozenxid for our DB
854 * Update pg_database's datfrozenxid entry for our database to be the
855 * minimum of the pg_class.relfrozenxid values.
857 * Similarly, update our datminmxid to be the minimum of the
858 * pg_class.relminmxid values.
860 * If we are able to advance either pg_database value, also try to
861 * truncate pg_clog and pg_multixact.
863 * We violate transaction semantics here by overwriting the database's
864 * existing pg_database tuple with the new values. This is reasonably
865 * safe since the new values are correct whether or not this transaction
866 * commits. As with vac_update_relstats, this avoids leaving dead tuples
867 * behind after a VACUUM.
870 vac_update_datfrozenxid(void)
873 Form_pg_database dbform;
877 TransactionId newFrozenXid;
878 MultiXactId newMinMulti;
879 TransactionId lastSaneFrozenXid;
880 MultiXactId lastSaneMinMulti;
885 * Initialize the "min" calculation with GetOldestXmin, which is a
886 * reasonable approximation to the minimum relfrozenxid for not-yet-
887 * committed pg_class entries for new tables; see AddNewRelationTuple().
888 * So we cannot produce a wrong minimum by starting with this.
890 newFrozenXid = GetOldestXmin(NULL, true);
893 * Similarly, initialize the MultiXact "min" with the value that would be
894 * used on pg_class for new tables. See AddNewRelationTuple().
896 newMinMulti = GetOldestMultiXactId();
899 * Identify the latest relfrozenxid and relminmxid values that we could
900 * validly see during the scan. These are conservative values, but it's
901 * not really worth trying to be more exact.
903 lastSaneFrozenXid = ReadNewTransactionId();
904 lastSaneMinMulti = ReadNextMultiXactId();
907 * We must seqscan pg_class to find the minimum Xid, because there is no
908 * index that can help us here.
910 relation = heap_open(RelationRelationId, AccessShareLock);
912 scan = systable_beginscan(relation, InvalidOid, false,
915 while ((classTup = systable_getnext(scan)) != NULL)
917 Form_pg_class classForm = (Form_pg_class) GETSTRUCT(classTup);
920 * Only consider relations able to hold unfrozen XIDs (anything else
921 * should have InvalidTransactionId in relfrozenxid anyway.)
923 if (classForm->relkind != RELKIND_RELATION &&
924 classForm->relkind != RELKIND_MATVIEW &&
925 classForm->relkind != RELKIND_TOASTVALUE)
928 Assert(TransactionIdIsNormal(classForm->relfrozenxid));
929 Assert(MultiXactIdIsValid(classForm->relminmxid));
932 * If things are working properly, no relation should have a
933 * relfrozenxid or relminmxid that is "in the future". However, such
934 * cases have been known to arise due to bugs in pg_upgrade. If we
935 * see any entries that are "in the future", chicken out and don't do
936 * anything. This ensures we won't truncate clog before those
937 * relations have been scanned and cleaned up.
939 if (TransactionIdPrecedes(lastSaneFrozenXid, classForm->relfrozenxid) ||
940 MultiXactIdPrecedes(lastSaneMinMulti, classForm->relminmxid))
946 if (TransactionIdPrecedes(classForm->relfrozenxid, newFrozenXid))
947 newFrozenXid = classForm->relfrozenxid;
949 if (MultiXactIdPrecedes(classForm->relminmxid, newMinMulti))
950 newMinMulti = classForm->relminmxid;
953 /* we're done with pg_class */
954 systable_endscan(scan);
955 heap_close(relation, AccessShareLock);
957 /* chicken out if bogus data found */
961 Assert(TransactionIdIsNormal(newFrozenXid));
962 Assert(MultiXactIdIsValid(newMinMulti));
964 /* Now fetch the pg_database tuple we need to update. */
965 relation = heap_open(DatabaseRelationId, RowExclusiveLock);
967 /* Fetch a copy of the tuple to scribble on */
968 tuple = SearchSysCacheCopy1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
969 if (!HeapTupleIsValid(tuple))
970 elog(ERROR, "could not find tuple for database %u", MyDatabaseId);
971 dbform = (Form_pg_database) GETSTRUCT(tuple);
974 * As in vac_update_relstats(), we ordinarily don't want to let
975 * datfrozenxid go backward; but if it's "in the future" then it must be
976 * corrupt and it seems best to overwrite it.
978 if (dbform->datfrozenxid != newFrozenXid &&
979 (TransactionIdPrecedes(dbform->datfrozenxid, newFrozenXid) ||
980 TransactionIdPrecedes(lastSaneFrozenXid, dbform->datfrozenxid)))
982 dbform->datfrozenxid = newFrozenXid;
986 newFrozenXid = dbform->datfrozenxid;
988 /* Ditto for datminmxid */
989 if (dbform->datminmxid != newMinMulti &&
990 (MultiXactIdPrecedes(dbform->datminmxid, newMinMulti) ||
991 MultiXactIdPrecedes(lastSaneMinMulti, dbform->datminmxid)))
993 dbform->datminmxid = newMinMulti;
997 newMinMulti = dbform->datminmxid;
1000 heap_inplace_update(relation, tuple);
1002 heap_freetuple(tuple);
1003 heap_close(relation, RowExclusiveLock);
1006 * If we were able to advance datfrozenxid or datminmxid, see if we can
1007 * truncate pg_clog and/or pg_multixact. Also do it if the shared
1008 * XID-wrap-limit info is stale, since this action will update that too.
1010 if (dirty || ForceTransactionIdLimitUpdate())
1011 vac_truncate_clog(newFrozenXid, newMinMulti,
1012 lastSaneFrozenXid, lastSaneMinMulti);
1017 * vac_truncate_clog() -- attempt to truncate the commit log
1019 * Scan pg_database to determine the system-wide oldest datfrozenxid,
1020 * and use it to truncate the transaction commit log (pg_clog).
1021 * Also update the XID wrap limit info maintained by varsup.c.
1022 * Likewise for datminmxid.
1024 * The passed frozenXID and minMulti are the updated values for my own
1025 * pg_database entry. They're used to initialize the "min" calculations.
1026 * The caller also passes the "last sane" XID and MXID, since it has
1027 * those at hand already.
1029 * This routine is only invoked when we've managed to change our
1030 * DB's datfrozenxid/datminmxid values, or we found that the shared
1031 * XID-wrap-limit info is stale.
1034 vac_truncate_clog(TransactionId frozenXID,
1035 MultiXactId minMulti,
1036 TransactionId lastSaneFrozenXid,
1037 MultiXactId lastSaneMinMulti)
1039 TransactionId myXID = GetCurrentTransactionId();
1043 Oid oldestxid_datoid;
1044 Oid minmulti_datoid;
1046 bool frozenAlreadyWrapped = false;
1048 /* init oldest datoids to sync with my frozenXID/minMulti values */
1049 oldestxid_datoid = MyDatabaseId;
1050 minmulti_datoid = MyDatabaseId;
1053 * Scan pg_database to compute the minimum datfrozenxid/datminmxid
1055 * Note: we need not worry about a race condition with new entries being
1056 * inserted by CREATE DATABASE. Any such entry will have a copy of some
1057 * existing DB's datfrozenxid, and that source DB cannot be ours because
1058 * of the interlock against copying a DB containing an active backend.
1059 * Hence the new entry will not reduce the minimum. Also, if two VACUUMs
1060 * concurrently modify the datfrozenxid's of different databases, the
1061 * worst possible outcome is that pg_clog is not truncated as aggressively
1064 relation = heap_open(DatabaseRelationId, AccessShareLock);
1066 scan = heap_beginscan_catalog(relation, 0, NULL);
1068 while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
1070 Form_pg_database dbform = (Form_pg_database) GETSTRUCT(tuple);
1072 Assert(TransactionIdIsNormal(dbform->datfrozenxid));
1073 Assert(MultiXactIdIsValid(dbform->datminmxid));
1076 * If things are working properly, no database should have a
1077 * datfrozenxid or datminmxid that is "in the future". However, such
1078 * cases have been known to arise due to bugs in pg_upgrade. If we
1079 * see any entries that are "in the future", chicken out and don't do
1080 * anything. This ensures we won't truncate clog before those
1081 * databases have been scanned and cleaned up. (We will issue the
1082 * "already wrapped" warning if appropriate, though.)
1084 if (TransactionIdPrecedes(lastSaneFrozenXid, dbform->datfrozenxid) ||
1085 MultiXactIdPrecedes(lastSaneMinMulti, dbform->datminmxid))
1088 if (TransactionIdPrecedes(myXID, dbform->datfrozenxid))
1089 frozenAlreadyWrapped = true;
1090 else if (TransactionIdPrecedes(dbform->datfrozenxid, frozenXID))
1092 frozenXID = dbform->datfrozenxid;
1093 oldestxid_datoid = HeapTupleGetOid(tuple);
1096 if (MultiXactIdPrecedes(dbform->datminmxid, minMulti))
1098 minMulti = dbform->datminmxid;
1099 minmulti_datoid = HeapTupleGetOid(tuple);
1105 heap_close(relation, AccessShareLock);
1108 * Do not truncate CLOG if we seem to have suffered wraparound already;
1109 * the computed minimum XID might be bogus. This case should now be
1110 * impossible due to the defenses in GetNewTransactionId, but we keep the
1113 if (frozenAlreadyWrapped)
1116 (errmsg("some databases have not been vacuumed in over 2 billion transactions"),
1117 errdetail("You might have already suffered transaction-wraparound data loss.")));
1121 /* chicken out if data is bogus in any other way */
1126 * Truncate CLOG and CommitTs to the oldest computed value.
1127 * Note we don't truncate multixacts; that will be done by the next
1130 TruncateCLOG(frozenXID);
1131 TruncateCommitTs(frozenXID, true);
1134 * Update the wrap limit for GetNewTransactionId and creation of new
1135 * MultiXactIds. Note: these functions will also signal the postmaster
1136 * for an(other) autovac cycle if needed. XXX should we avoid possibly
1139 SetTransactionIdLimit(frozenXID, oldestxid_datoid);
1140 SetMultiXactIdLimit(minMulti, minmulti_datoid);
1141 AdvanceOldestCommitTs(frozenXID);
1146 * vacuum_rel() -- vacuum one heap relation
1148 * Doing one heap at a time incurs extra overhead, since we need to
1149 * check that the heap exists again just before we vacuum it. The
1150 * reason that we do this is so that vacuuming can be spread across
1151 * many small transactions. Otherwise, two-phase locking would require
1152 * us to lock the entire database during one pass of the vacuum cleaner.
1154 * At entry and exit, we are not inside a transaction.
1157 vacuum_rel(Oid relid, RangeVar *relation, int options, VacuumParams *params)
1164 int save_sec_context;
1167 Assert(params != NULL);
1169 /* Begin a transaction for vacuuming this relation */
1170 StartTransactionCommand();
1173 * Functions in indexes may want a snapshot set. Also, setting a snapshot
1174 * ensures that RecentGlobalXmin is kept truly recent.
1176 PushActiveSnapshot(GetTransactionSnapshot());
1178 if (!(options & VACOPT_FULL))
1181 * In lazy vacuum, we can set the PROC_IN_VACUUM flag, which lets
1182 * other concurrent VACUUMs know that they can ignore this one while
1183 * determining their OldestXmin. (The reason we don't set it during a
1184 * full VACUUM is exactly that we may have to run user-defined
1185 * functions for functional indexes, and we want to make sure that if
1186 * they use the snapshot set above, any tuples it requires can't get
1187 * removed from other tables. An index function that depends on the
1188 * contents of other tables is arguably broken, but we won't break it
1189 * here by violating transaction semantics.)
1191 * We also set the VACUUM_FOR_WRAPAROUND flag, which is passed down by
1192 * autovacuum; it's used to avoid canceling a vacuum that was invoked
1195 * Note: these flags remain set until CommitTransaction or
1196 * AbortTransaction. We don't want to clear them until we reset
1197 * MyPgXact->xid/xmin, else OldestXmin might appear to go backwards,
1198 * which is probably Not Good.
1200 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1201 MyPgXact->vacuumFlags |= PROC_IN_VACUUM;
1202 if (params->is_wraparound)
1203 MyPgXact->vacuumFlags |= PROC_VACUUM_FOR_WRAPAROUND;
1204 LWLockRelease(ProcArrayLock);
1208 * Check for user-requested abort. Note we want this to be inside a
1209 * transaction, so xact.c doesn't issue useless WARNING.
1211 CHECK_FOR_INTERRUPTS();
1214 * Determine the type of lock we want --- hard exclusive lock for a FULL
1215 * vacuum, but just ShareUpdateExclusiveLock for concurrent vacuum. Either
1216 * way, we can be sure that no other backend is vacuuming the same table.
1218 lmode = (options & VACOPT_FULL) ? AccessExclusiveLock : ShareUpdateExclusiveLock;
1221 * Open the relation and get the appropriate lock on it.
1223 * There's a race condition here: the rel may have gone away since the
1224 * last time we saw it. If so, we don't need to vacuum it.
1226 * If we've been asked not to wait for the relation lock, acquire it first
1227 * in non-blocking mode, before calling try_relation_open().
1229 if (!(options & VACOPT_NOWAIT))
1230 onerel = try_relation_open(relid, lmode);
1231 else if (ConditionalLockRelationOid(relid, lmode))
1232 onerel = try_relation_open(relid, NoLock);
1236 if (IsAutoVacuumWorkerProcess() && Log_autovacuum_min_duration >= 0)
1238 (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
1239 errmsg("skipping vacuum of \"%s\" --- lock not available",
1240 relation->relname)));
1245 PopActiveSnapshot();
1246 CommitTransactionCommand();
1251 * Check permissions.
1253 * We allow the user to vacuum a table if he is superuser, the table
1254 * owner, or the database owner (but in the latter case, only if it's not
1255 * a shared relation). pg_class_ownercheck includes the superuser case.
1257 * Note we choose to treat permissions failure as a WARNING and keep
1258 * trying to vacuum the rest of the DB --- is this appropriate?
1260 if (!(pg_class_ownercheck(RelationGetRelid(onerel), GetUserId()) ||
1261 (pg_database_ownercheck(MyDatabaseId, GetUserId()) && !onerel->rd_rel->relisshared)))
1263 if (onerel->rd_rel->relisshared)
1265 (errmsg("skipping \"%s\" --- only superuser can vacuum it",
1266 RelationGetRelationName(onerel))));
1267 else if (onerel->rd_rel->relnamespace == PG_CATALOG_NAMESPACE)
1269 (errmsg("skipping \"%s\" --- only superuser or database owner can vacuum it",
1270 RelationGetRelationName(onerel))));
1273 (errmsg("skipping \"%s\" --- only table or database owner can vacuum it",
1274 RelationGetRelationName(onerel))));
1275 relation_close(onerel, lmode);
1276 PopActiveSnapshot();
1277 CommitTransactionCommand();
1282 * Check that it's a vacuumable relation; we used to do this in
1283 * get_rel_oids() but seems safer to check after we've locked the
1286 if (onerel->rd_rel->relkind != RELKIND_RELATION &&
1287 onerel->rd_rel->relkind != RELKIND_MATVIEW &&
1288 onerel->rd_rel->relkind != RELKIND_TOASTVALUE)
1291 (errmsg("skipping \"%s\" --- cannot vacuum non-tables or special system tables",
1292 RelationGetRelationName(onerel))));
1293 relation_close(onerel, lmode);
1294 PopActiveSnapshot();
1295 CommitTransactionCommand();
1300 * Silently ignore tables that are temp tables of other backends ---
1301 * trying to vacuum these will lead to great unhappiness, since their
1302 * contents are probably not up-to-date on disk. (We don't throw a
1303 * warning here; it would just lead to chatter during a database-wide
1306 if (RELATION_IS_OTHER_TEMP(onerel))
1308 relation_close(onerel, lmode);
1309 PopActiveSnapshot();
1310 CommitTransactionCommand();
1315 * Get a session-level lock too. This will protect our access to the
1316 * relation across multiple transactions, so that we can vacuum the
1317 * relation's TOAST table (if any) secure in the knowledge that no one is
1318 * deleting the parent relation.
1320 * NOTE: this cannot block, even if someone else is waiting for access,
1321 * because the lock manager knows that both lock requests are from the
1324 onerelid = onerel->rd_lockInfo.lockRelId;
1325 LockRelationIdForSession(&onerelid, lmode);
1328 * Remember the relation's TOAST relation for later, if the caller asked
1329 * us to process it. In VACUUM FULL, though, the toast table is
1330 * automatically rebuilt by cluster_rel so we shouldn't recurse to it.
1332 if (!(options & VACOPT_SKIPTOAST) && !(options & VACOPT_FULL))
1333 toast_relid = onerel->rd_rel->reltoastrelid;
1335 toast_relid = InvalidOid;
1338 * Switch to the table owner's userid, so that any index functions are run
1339 * as that user. Also lock down security-restricted operations and
1340 * arrange to make GUC variable changes local to this command. (This is
1341 * unnecessary, but harmless, for lazy VACUUM.)
1343 GetUserIdAndSecContext(&save_userid, &save_sec_context);
1344 SetUserIdAndSecContext(onerel->rd_rel->relowner,
1345 save_sec_context | SECURITY_RESTRICTED_OPERATION);
1346 save_nestlevel = NewGUCNestLevel();
1349 * Do the actual work --- either FULL or "lazy" vacuum
1351 if (options & VACOPT_FULL)
1353 /* close relation before vacuuming, but hold lock until commit */
1354 relation_close(onerel, NoLock);
1357 /* VACUUM FULL is now a variant of CLUSTER; see cluster.c */
1358 cluster_rel(relid, InvalidOid, false,
1359 (options & VACOPT_VERBOSE) != 0);
1362 lazy_vacuum_rel(onerel, options, params, vac_strategy);
1364 /* Roll back any GUC changes executed by index functions */
1365 AtEOXact_GUC(false, save_nestlevel);
1367 /* Restore userid and security context */
1368 SetUserIdAndSecContext(save_userid, save_sec_context);
1370 /* all done with this class, but hold lock until commit */
1372 relation_close(onerel, NoLock);
1375 * Complete the transaction and free all temporary memory used.
1377 PopActiveSnapshot();
1378 CommitTransactionCommand();
1381 * If the relation has a secondary toast rel, vacuum that too while we
1382 * still hold the session lock on the master table. Note however that
1383 * "analyze" will not get done on the toast table. This is good, because
1384 * the toaster always uses hardcoded index access and statistics are
1385 * totally unimportant for toast relations.
1387 if (toast_relid != InvalidOid)
1388 vacuum_rel(toast_relid, relation, options, params);
1391 * Now release the session-level lock on the master table.
1393 UnlockRelationIdForSession(&onerelid, lmode);
1395 /* Report that we really did it. */
1401 * Open all the vacuumable indexes of the given relation, obtaining the
1402 * specified kind of lock on each. Return an array of Relation pointers for
1403 * the indexes into *Irel, and the number of indexes into *nindexes.
1405 * We consider an index vacuumable if it is marked insertable (IndexIsReady).
1406 * If it isn't, probably a CREATE INDEX CONCURRENTLY command failed early in
1407 * execution, and what we have is too corrupt to be processable. We will
1408 * vacuum even if the index isn't indisvalid; this is important because in a
1409 * unique index, uniqueness checks will be performed anyway and had better not
1410 * hit dangling index pointers.
1413 vac_open_indexes(Relation relation, LOCKMODE lockmode,
1414 int *nindexes, Relation **Irel)
1417 ListCell *indexoidscan;
1420 Assert(lockmode != NoLock);
1422 indexoidlist = RelationGetIndexList(relation);
1424 /* allocate enough memory for all indexes */
1425 i = list_length(indexoidlist);
1428 *Irel = (Relation *) palloc(i * sizeof(Relation));
1432 /* collect just the ready indexes */
1434 foreach(indexoidscan, indexoidlist)
1436 Oid indexoid = lfirst_oid(indexoidscan);
1439 indrel = index_open(indexoid, lockmode);
1440 if (IndexIsReady(indrel->rd_index))
1441 (*Irel)[i++] = indrel;
1443 index_close(indrel, lockmode);
1448 list_free(indexoidlist);
1452 * Release the resources acquired by vac_open_indexes. Optionally release
1453 * the locks (say NoLock to keep 'em).
1456 vac_close_indexes(int nindexes, Relation *Irel, LOCKMODE lockmode)
1463 Relation ind = Irel[nindexes];
1465 index_close(ind, lockmode);
1471 * vacuum_delay_point --- check for interrupts and cost-based delay.
1473 * This should be called in each major loop of VACUUM processing,
1474 * typically once per page processed.
1477 vacuum_delay_point(void)
1479 /* Always check for interrupts */
1480 CHECK_FOR_INTERRUPTS();
1482 /* Nap if appropriate */
1483 if (VacuumCostActive && !InterruptPending &&
1484 VacuumCostBalance >= VacuumCostLimit)
1488 msec = VacuumCostDelay * VacuumCostBalance / VacuumCostLimit;
1489 if (msec > VacuumCostDelay * 4)
1490 msec = VacuumCostDelay * 4;
1492 pg_usleep(msec * 1000L);
1494 VacuumCostBalance = 0;
1496 /* update balance values for workers */
1497 AutoVacuumUpdateDelay();
1499 /* Might have gotten an interrupt while sleeping */
1500 CHECK_FOR_INTERRUPTS();