1 /*-------------------------------------------------------------------------
4 * The postgres vacuum cleaner.
6 * This file includes the "full" version of VACUUM, as well as control code
7 * used by all three of full VACUUM, lazy VACUUM, and ANALYZE. See
8 * vacuumlazy.c and analyze.c for the rest of the code for the latter two.
11 * Portions Copyright (c) 1996-2009, PostgreSQL Global Development Group
12 * Portions Copyright (c) 1994, Regents of the University of California
16 * $PostgreSQL: pgsql/src/backend/commands/vacuum.c,v 1.387 2009/03/31 22:12:48 tgl Exp $
18 *-------------------------------------------------------------------------
25 #include "access/clog.h"
26 #include "access/genam.h"
27 #include "access/heapam.h"
28 #include "access/transam.h"
29 #include "access/visibilitymap.h"
30 #include "access/xact.h"
31 #include "access/xlog.h"
32 #include "catalog/namespace.h"
33 #include "catalog/pg_database.h"
34 #include "catalog/pg_namespace.h"
35 #include "catalog/storage.h"
36 #include "commands/dbcommands.h"
37 #include "commands/vacuum.h"
38 #include "executor/executor.h"
39 #include "miscadmin.h"
41 #include "postmaster/autovacuum.h"
42 #include "storage/bufmgr.h"
43 #include "storage/freespace.h"
44 #include "storage/lmgr.h"
45 #include "storage/proc.h"
46 #include "storage/procarray.h"
47 #include "utils/acl.h"
48 #include "utils/builtins.h"
49 #include "utils/flatfiles.h"
50 #include "utils/fmgroids.h"
51 #include "utils/inval.h"
52 #include "utils/lsyscache.h"
53 #include "utils/memutils.h"
54 #include "utils/pg_rusage.h"
55 #include "utils/relcache.h"
56 #include "utils/snapmgr.h"
57 #include "utils/syscache.h"
58 #include "utils/tqual.h"
64 int vacuum_freeze_min_age;
65 int vacuum_freeze_table_age;
68 * VacPage structures keep track of each page on which we find useful
69 * amounts of free space.
71 typedef struct VacPageData
73 BlockNumber blkno; /* BlockNumber of this Page */
74 Size free; /* FreeSpace on this Page */
75 uint16 offsets_used; /* Number of OffNums used by vacuum */
76 uint16 offsets_free; /* Number of OffNums free or to be free */
77 OffsetNumber offsets[1]; /* Array of free OffNums */
80 typedef VacPageData *VacPage;
82 typedef struct VacPageListData
84 BlockNumber empty_end_pages; /* Number of "empty" end-pages */
85 int num_pages; /* Number of pages in pagedesc */
86 int num_allocated_pages; /* Number of allocated pages in
88 VacPage *pagedesc; /* Descriptions of pages */
91 typedef VacPageListData *VacPageList;
94 * The "vtlinks" array keeps information about each recently-updated tuple
95 * ("recent" meaning its XMAX is too new to let us recycle the tuple).
96 * We store the tuple's own TID as well as its t_ctid (its link to the next
97 * newer tuple version). Searching in this array allows us to follow update
98 * chains backwards from newer to older tuples. When we move a member of an
99 * update chain, we must move *all* the live members of the chain, so that we
100 * can maintain their t_ctid link relationships (we must not just overwrite
101 * t_ctid in an existing tuple).
103 * Note: because t_ctid links can be stale (this would only occur if a prior
104 * VACUUM crashed partway through), it is possible that new_tid points to an
105 * empty slot or unrelated tuple. We have to check the linkage as we follow
106 * it, just as is done in EvalPlanQual.
108 typedef struct VTupleLinkData
110 ItemPointerData new_tid; /* t_ctid of an updated tuple */
111 ItemPointerData this_tid; /* t_self of the tuple */
114 typedef VTupleLinkData *VTupleLink;
117 * We use an array of VTupleMoveData to plan a chain tuple move fully
120 typedef struct VTupleMoveData
122 ItemPointerData tid; /* tuple ID */
123 VacPage vacpage; /* where to move it to */
124 bool cleanVpd; /* clean vacpage before using? */
127 typedef VTupleMoveData *VTupleMove;
130 * VRelStats contains the data acquired by scan_heap for use later
132 typedef struct VRelStats
134 /* miscellaneous statistics */
135 BlockNumber rel_pages; /* pages in relation */
136 double rel_tuples; /* tuples that remain after vacuuming */
137 double rel_indexed_tuples; /* indexed tuples that remain */
138 Size min_tlen; /* min surviving tuple size */
139 Size max_tlen; /* max surviving tuple size */
141 /* vtlinks array for tuple chain following - sorted by new_tid */
146 /*----------------------------------------------------------------------
149 * As these variables always appear together, we put them into one struct
150 * and pull initialization and cleanup into separate routines.
151 * ExecContext is used by repair_frag() and move_xxx_tuple(). More
152 * accurately: It is *used* only in move_xxx_tuple(), but because this
153 * routine is called many times, we initialize the struct just once in
154 * repair_frag() and pass it on to move_xxx_tuple().
156 typedef struct ExecContextData
158 ResultRelInfo *resultRelInfo;
160 TupleTableSlot *slot;
163 typedef ExecContextData *ExecContext;
166 ExecContext_Init(ExecContext ec, Relation rel)
168 TupleDesc tupdesc = RelationGetDescr(rel);
171 * We need a ResultRelInfo and an EState so we can use the regular
172 * executor's index-entry-making machinery.
174 ec->estate = CreateExecutorState();
176 ec->resultRelInfo = makeNode(ResultRelInfo);
177 ec->resultRelInfo->ri_RangeTableIndex = 1; /* dummy */
178 ec->resultRelInfo->ri_RelationDesc = rel;
179 ec->resultRelInfo->ri_TrigDesc = NULL; /* we don't fire triggers */
181 ExecOpenIndices(ec->resultRelInfo);
183 ec->estate->es_result_relations = ec->resultRelInfo;
184 ec->estate->es_num_result_relations = 1;
185 ec->estate->es_result_relation_info = ec->resultRelInfo;
187 /* Set up a tuple slot too */
188 ec->slot = MakeSingleTupleTableSlot(tupdesc);
192 ExecContext_Finish(ExecContext ec)
194 ExecDropSingleTupleTableSlot(ec->slot);
195 ExecCloseIndices(ec->resultRelInfo);
196 FreeExecutorState(ec->estate);
200 * End of ExecContext Implementation
201 *----------------------------------------------------------------------
204 /* A few variables that don't seem worth passing around as parameters */
205 static MemoryContext vac_context = NULL;
207 static int elevel = -1;
209 static TransactionId OldestXmin;
210 static TransactionId FreezeLimit;
212 static BufferAccessStrategy vac_strategy;
215 /* non-export function prototypes */
216 static List *get_rel_oids(Oid relid, const RangeVar *vacrel,
217 const char *stmttype);
218 static void vac_truncate_clog(TransactionId frozenXID);
219 static void vacuum_rel(Oid relid, VacuumStmt *vacstmt, bool do_toast,
220 bool for_wraparound, bool *scanned_all);
221 static void full_vacuum_rel(Relation onerel, VacuumStmt *vacstmt);
222 static void scan_heap(VRelStats *vacrelstats, Relation onerel,
223 VacPageList vacuum_pages, VacPageList fraged_pages);
224 static void repair_frag(VRelStats *vacrelstats, Relation onerel,
225 VacPageList vacuum_pages, VacPageList fraged_pages,
226 int nindexes, Relation *Irel);
227 static void move_chain_tuple(Relation rel,
228 Buffer old_buf, Page old_page, HeapTuple old_tup,
229 Buffer dst_buf, Page dst_page, VacPage dst_vacpage,
230 ExecContext ec, ItemPointer ctid, bool cleanVpd);
231 static void move_plain_tuple(Relation rel,
232 Buffer old_buf, Page old_page, HeapTuple old_tup,
233 Buffer dst_buf, Page dst_page, VacPage dst_vacpage,
235 static void update_hint_bits(Relation rel, VacPageList fraged_pages,
236 int num_fraged_pages, BlockNumber last_move_dest_block,
238 static void vacuum_heap(VRelStats *vacrelstats, Relation onerel,
239 VacPageList vacpagelist);
240 static void vacuum_page(Relation onerel, Buffer buffer, VacPage vacpage);
241 static void vacuum_index(VacPageList vacpagelist, Relation indrel,
242 double num_tuples, int keep_tuples);
243 static void scan_index(Relation indrel, double num_tuples);
244 static bool tid_reaped(ItemPointer itemptr, void *state);
245 static void vac_update_fsm(Relation onerel, VacPageList fraged_pages,
246 BlockNumber rel_pages);
247 static VacPage copy_vac_page(VacPage vacpage);
248 static void vpage_insert(VacPageList vacpagelist, VacPage vpnew);
249 static void *vac_bsearch(const void *key, const void *base,
250 size_t nelem, size_t size,
251 int (*compar) (const void *, const void *));
252 static int vac_cmp_blk(const void *left, const void *right);
253 static int vac_cmp_offno(const void *left, const void *right);
254 static int vac_cmp_vtlinks(const void *left, const void *right);
255 static bool enough_space(VacPage vacpage, Size len);
256 static Size PageGetFreeSpaceWithFillFactor(Relation relation, Page page);
259 /****************************************************************************
261 * Code common to all flavors of VACUUM and ANALYZE *
263 ****************************************************************************
268 * Primary entry point for VACUUM and ANALYZE commands.
270 * relid is normally InvalidOid; if it is not, then it provides the relation
271 * OID to be processed, and vacstmt->relation is ignored. (The non-invalid
272 * case is currently only used by autovacuum.)
274 * do_toast is passed as FALSE by autovacuum, because it processes TOAST
277 * for_wraparound is used by autovacuum to let us know when it's forcing
278 * a vacuum for wraparound, which should not be auto-cancelled.
280 * bstrategy is normally given as NULL, but in autovacuum it can be passed
281 * in to use the same buffer strategy object across multiple vacuum() calls.
283 * isTopLevel should be passed down from ProcessUtility.
285 * It is the caller's responsibility that vacstmt and bstrategy
286 * (if given) be allocated in a memory context that won't disappear
287 * at transaction commit.
290 vacuum(VacuumStmt *vacstmt, Oid relid, bool do_toast,
291 BufferAccessStrategy bstrategy, bool for_wraparound, bool isTopLevel)
293 const char *stmttype = vacstmt->vacuum ? "VACUUM" : "ANALYZE";
294 volatile MemoryContext anl_context = NULL;
295 volatile bool all_rels,
300 if (vacstmt->verbose)
306 * We cannot run VACUUM inside a user transaction block; if we were inside
307 * a transaction, then our commit- and start-transaction-command calls
308 * would not have the intended effect! Furthermore, the forced commit that
309 * occurs before truncating the relation's file would have the effect of
310 * committing the rest of the user's transaction too, which would
311 * certainly not be the desired behavior. (This only applies to VACUUM
312 * FULL, though. We could in theory run lazy VACUUM inside a transaction
313 * block, but we choose to disallow that case because we'd rather commit
314 * as soon as possible after finishing the vacuum. This is mainly so that
315 * we can let go the AccessExclusiveLock that we may be holding.)
317 * ANALYZE (without VACUUM) can run either way.
321 PreventTransactionChain(isTopLevel, stmttype);
322 in_outer_xact = false;
325 in_outer_xact = IsInTransactionChain(isTopLevel);
328 * Send info about dead objects to the statistics collector, unless we are
329 * in autovacuum --- autovacuum.c does this for itself.
331 if (vacstmt->vacuum && !IsAutoVacuumWorkerProcess())
332 pgstat_vacuum_stat();
335 * Create special memory context for cross-transaction storage.
337 * Since it is a child of PortalContext, it will go away eventually even
338 * if we suffer an error; there's no need for special abort cleanup logic.
340 vac_context = AllocSetContextCreate(PortalContext,
342 ALLOCSET_DEFAULT_MINSIZE,
343 ALLOCSET_DEFAULT_INITSIZE,
344 ALLOCSET_DEFAULT_MAXSIZE);
347 * If caller didn't give us a buffer strategy object, make one in the
348 * cross-transaction memory context.
350 if (bstrategy == NULL)
352 MemoryContext old_context = MemoryContextSwitchTo(vac_context);
354 bstrategy = GetAccessStrategy(BAS_VACUUM);
355 MemoryContextSwitchTo(old_context);
357 vac_strategy = bstrategy;
359 /* Remember whether we are processing everything in the DB */
360 all_rels = (!OidIsValid(relid) && vacstmt->relation == NULL);
363 * Build list of relations to process, unless caller gave us one. (If we
364 * build one, we put it in vac_context for safekeeping.)
366 relations = get_rel_oids(relid, vacstmt->relation, stmttype);
369 * Decide whether we need to start/commit our own transactions.
371 * For VACUUM (with or without ANALYZE): always do so, so that we can
372 * release locks as soon as possible. (We could possibly use the outer
373 * transaction for a one-table VACUUM, but handling TOAST tables would be
376 * For ANALYZE (no VACUUM): if inside a transaction block, we cannot
377 * start/commit our own transactions. Also, there's no need to do so if
378 * only processing one relation. For multiple relations when not within a
379 * transaction block, and also in an autovacuum worker, use own
380 * transactions so we can release locks sooner.
383 use_own_xacts = true;
386 Assert(vacstmt->analyze);
387 if (IsAutoVacuumWorkerProcess())
388 use_own_xacts = true;
389 else if (in_outer_xact)
390 use_own_xacts = false;
391 else if (list_length(relations) > 1)
392 use_own_xacts = true;
394 use_own_xacts = false;
398 * If we are running ANALYZE without per-table transactions, we'll need a
399 * memory context with table lifetime.
402 anl_context = AllocSetContextCreate(PortalContext,
404 ALLOCSET_DEFAULT_MINSIZE,
405 ALLOCSET_DEFAULT_INITSIZE,
406 ALLOCSET_DEFAULT_MAXSIZE);
409 * vacuum_rel expects to be entered with no transaction active; it will
410 * start and commit its own transaction. But we are called by an SQL
411 * command, and so we are executing inside a transaction already. We
412 * commit the transaction started in PostgresMain() here, and start
413 * another one before exiting to match the commit waiting for us back in
418 /* ActiveSnapshot is not set by autovacuum */
419 if (ActiveSnapshotSet())
422 /* matches the StartTransaction in PostgresMain() */
423 CommitTransactionCommand();
426 /* Turn vacuum cost accounting on or off */
431 VacuumCostActive = (VacuumCostDelay > 0);
432 VacuumCostBalance = 0;
435 * Loop to process each selected relation.
437 foreach(cur, relations)
439 Oid relid = lfirst_oid(cur);
440 bool scanned_all = false;
443 vacuum_rel(relid, vacstmt, do_toast, for_wraparound,
446 if (vacstmt->analyze)
448 MemoryContext old_context = NULL;
451 * If using separate xacts, start one for analyze. Otherwise,
452 * we can use the outer transaction, but we still need to call
453 * analyze_rel in a memory context that will be cleaned up on
454 * return (else we leak memory while processing multiple
459 StartTransactionCommand();
460 /* functions in indexes may want a snapshot set */
461 PushActiveSnapshot(GetTransactionSnapshot());
464 old_context = MemoryContextSwitchTo(anl_context);
466 analyze_rel(relid, vacstmt, vac_strategy, !scanned_all);
471 CommitTransactionCommand();
475 MemoryContextSwitchTo(old_context);
476 MemoryContextResetAndDeleteChildren(anl_context);
483 /* Make sure cost accounting is turned off after error */
484 VacuumCostActive = false;
489 /* Turn off vacuum cost accounting */
490 VacuumCostActive = false;
493 * Finish up processing.
497 /* here, we are not in a transaction */
500 * This matches the CommitTransaction waiting for us in
503 StartTransactionCommand();
506 if (vacstmt->vacuum && !IsAutoVacuumWorkerProcess())
509 * Update pg_database.datfrozenxid, and truncate pg_clog if possible.
510 * (autovacuum.c does this for itself.)
512 vac_update_datfrozenxid();
516 * Clean up working storage --- note we must do this after
517 * StartTransactionCommand, else we might be trying to delete the active
520 MemoryContextDelete(vac_context);
524 MemoryContextDelete(anl_context);
528 * Build a list of Oids for each relation to be processed
530 * The list is built in vac_context so that it will survive across our
531 * per-relation transactions.
534 get_rel_oids(Oid relid, const RangeVar *vacrel, const char *stmttype)
536 List *oid_list = NIL;
537 MemoryContext oldcontext;
539 /* OID supplied by VACUUM's caller? */
540 if (OidIsValid(relid))
542 oldcontext = MemoryContextSwitchTo(vac_context);
543 oid_list = lappend_oid(oid_list, relid);
544 MemoryContextSwitchTo(oldcontext);
548 /* Process a specific relation */
551 relid = RangeVarGetRelid(vacrel, false);
553 /* Make a relation list entry for this guy */
554 oldcontext = MemoryContextSwitchTo(vac_context);
555 oid_list = lappend_oid(oid_list, relid);
556 MemoryContextSwitchTo(oldcontext);
560 /* Process all plain relations listed in pg_class */
567 Anum_pg_class_relkind,
568 BTEqualStrategyNumber, F_CHAREQ,
569 CharGetDatum(RELKIND_RELATION));
571 pgclass = heap_open(RelationRelationId, AccessShareLock);
573 scan = heap_beginscan(pgclass, SnapshotNow, 1, &key);
575 while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
577 /* Make a relation list entry for this guy */
578 oldcontext = MemoryContextSwitchTo(vac_context);
579 oid_list = lappend_oid(oid_list, HeapTupleGetOid(tuple));
580 MemoryContextSwitchTo(oldcontext);
584 heap_close(pgclass, AccessShareLock);
591 * vacuum_set_xid_limits() -- compute oldest-Xmin and freeze cutoff points
594 vacuum_set_xid_limits(int freeze_min_age,
595 int freeze_table_age,
597 TransactionId *oldestXmin,
598 TransactionId *freezeLimit,
599 TransactionId *freezeTableLimit)
603 TransactionId safeLimit;
606 * We can always ignore processes running lazy vacuum. This is because we
607 * use these values only for deciding which tuples we must keep in the
608 * tables. Since lazy vacuum doesn't write its XID anywhere, it's safe to
609 * ignore it. In theory it could be problematic to ignore lazy vacuums on
610 * a full vacuum, but keep in mind that only one vacuum process can be
611 * working on a particular table at any time, and that each vacuum is
612 * always an independent transaction.
614 *oldestXmin = GetOldestXmin(sharedRel, true);
616 Assert(TransactionIdIsNormal(*oldestXmin));
619 * Determine the minimum freeze age to use: as specified by the caller, or
620 * vacuum_freeze_min_age, but in any case not more than half
621 * autovacuum_freeze_max_age, so that autovacuums to prevent XID
622 * wraparound won't occur too frequently.
624 freezemin = freeze_min_age;
626 freezemin = vacuum_freeze_min_age;
627 freezemin = Min(freezemin, autovacuum_freeze_max_age / 2);
628 Assert(freezemin >= 0);
631 * Compute the cutoff XID, being careful not to generate a "permanent" XID
633 limit = *oldestXmin - freezemin;
634 if (!TransactionIdIsNormal(limit))
635 limit = FirstNormalTransactionId;
638 * If oldestXmin is very far back (in practice, more than
639 * autovacuum_freeze_max_age / 2 XIDs old), complain and force a minimum
640 * freeze age of zero.
642 safeLimit = ReadNewTransactionId() - autovacuum_freeze_max_age;
643 if (!TransactionIdIsNormal(safeLimit))
644 safeLimit = FirstNormalTransactionId;
646 if (TransactionIdPrecedes(limit, safeLimit))
649 (errmsg("oldest xmin is far in the past"),
650 errhint("Close open transactions soon to avoid wraparound problems.")));
654 *freezeLimit = limit;
656 if (freezeTableLimit != NULL)
661 * Determine the table freeze age to use: as specified by the caller,
662 * or vacuum_freeze_table_age, but in any case not more than
663 * autovacuum_freeze_max_age * 0.95, so that if you have e.g nightly
664 * VACUUM schedule, the nightly VACUUM gets a chance to freeze tuples
665 * before anti-wraparound autovacuum is launched.
667 freezetable = freeze_min_age;
669 freezetable = vacuum_freeze_table_age;
670 freezetable = Min(freezetable, autovacuum_freeze_max_age * 0.95);
671 Assert(freezetable >= 0);
674 * Compute the cutoff XID, being careful not to generate a
677 limit = ReadNewTransactionId() - freezetable;
678 if (!TransactionIdIsNormal(limit))
679 limit = FirstNormalTransactionId;
681 *freezeTableLimit = limit;
687 * vac_update_relstats() -- update statistics for one relation
689 * Update the whole-relation statistics that are kept in its pg_class
690 * row. There are additional stats that will be updated if we are
691 * doing ANALYZE, but we always update these stats. This routine works
692 * for both index and heap relation entries in pg_class.
694 * We violate transaction semantics here by overwriting the rel's
695 * existing pg_class tuple with the new values. This is reasonably
696 * safe since the new values are correct whether or not this transaction
697 * commits. The reason for this is that if we updated these tuples in
698 * the usual way, vacuuming pg_class itself wouldn't work very well ---
699 * by the time we got done with a vacuum cycle, most of the tuples in
700 * pg_class would've been obsoleted. Of course, this only works for
701 * fixed-size never-null columns, but these are.
703 * Note another assumption: that two VACUUMs/ANALYZEs on a table can't
704 * run in parallel, nor can VACUUM/ANALYZE run in parallel with a
705 * schema alteration such as adding an index, rule, or trigger. Otherwise
706 * our updates of relhasindex etc might overwrite uncommitted updates.
708 * Another reason for doing it this way is that when we are in a lazy
709 * VACUUM and have PROC_IN_VACUUM set, we mustn't do any updates ---
710 * somebody vacuuming pg_class might think they could delete a tuple
711 * marked with xmin = our xid.
713 * This routine is shared by full VACUUM, lazy VACUUM, and stand-alone
717 vac_update_relstats(Relation relation,
718 BlockNumber num_pages, double num_tuples,
719 bool hasindex, TransactionId frozenxid)
721 Oid relid = RelationGetRelid(relation);
724 Form_pg_class pgcform;
727 rd = heap_open(RelationRelationId, RowExclusiveLock);
729 /* Fetch a copy of the tuple to scribble on */
730 ctup = SearchSysCacheCopy(RELOID,
731 ObjectIdGetDatum(relid),
733 if (!HeapTupleIsValid(ctup))
734 elog(ERROR, "pg_class entry for relid %u vanished during vacuuming",
736 pgcform = (Form_pg_class) GETSTRUCT(ctup);
738 /* Apply required updates, if any, to copied tuple */
741 if (pgcform->relpages != (int32) num_pages)
743 pgcform->relpages = (int32) num_pages;
746 if (pgcform->reltuples != (float4) num_tuples)
748 pgcform->reltuples = (float4) num_tuples;
751 if (pgcform->relhasindex != hasindex)
753 pgcform->relhasindex = hasindex;
758 * If we have discovered that there are no indexes, then there's no
759 * primary key either. This could be done more thoroughly...
763 if (pgcform->relhaspkey)
765 pgcform->relhaspkey = false;
770 /* We also clear relhasrules and relhastriggers if needed */
771 if (pgcform->relhasrules && relation->rd_rules == NULL)
773 pgcform->relhasrules = false;
776 if (pgcform->relhastriggers && relation->trigdesc == NULL)
778 pgcform->relhastriggers = false;
783 * relfrozenxid should never go backward. Caller can pass
784 * InvalidTransactionId if it has no new data.
786 if (TransactionIdIsNormal(frozenxid) &&
787 TransactionIdPrecedes(pgcform->relfrozenxid, frozenxid))
789 pgcform->relfrozenxid = frozenxid;
793 /* If anything changed, write out the tuple. */
795 heap_inplace_update(rd, ctup);
797 heap_close(rd, RowExclusiveLock);
802 * vac_update_datfrozenxid() -- update pg_database.datfrozenxid for our DB
804 * Update pg_database's datfrozenxid entry for our database to be the
805 * minimum of the pg_class.relfrozenxid values. If we are able to
806 * advance pg_database.datfrozenxid, also try to truncate pg_clog.
808 * We violate transaction semantics here by overwriting the database's
809 * existing pg_database tuple with the new value. This is reasonably
810 * safe since the new value is correct whether or not this transaction
811 * commits. As with vac_update_relstats, this avoids leaving dead tuples
812 * behind after a VACUUM.
814 * This routine is shared by full and lazy VACUUM.
817 vac_update_datfrozenxid(void)
820 Form_pg_database dbform;
824 TransactionId newFrozenXid;
828 * Initialize the "min" calculation with GetOldestXmin, which is a
829 * reasonable approximation to the minimum relfrozenxid for not-yet-
830 * committed pg_class entries for new tables; see AddNewRelationTuple().
831 * Se we cannot produce a wrong minimum by starting with this.
833 newFrozenXid = GetOldestXmin(true, true);
836 * We must seqscan pg_class to find the minimum Xid, because there is no
837 * index that can help us here.
839 relation = heap_open(RelationRelationId, AccessShareLock);
841 scan = systable_beginscan(relation, InvalidOid, false,
842 SnapshotNow, 0, NULL);
844 while ((classTup = systable_getnext(scan)) != NULL)
846 Form_pg_class classForm = (Form_pg_class) GETSTRUCT(classTup);
849 * Only consider heap and TOAST tables (anything else should have
850 * InvalidTransactionId in relfrozenxid anyway.)
852 if (classForm->relkind != RELKIND_RELATION &&
853 classForm->relkind != RELKIND_TOASTVALUE)
856 Assert(TransactionIdIsNormal(classForm->relfrozenxid));
858 if (TransactionIdPrecedes(classForm->relfrozenxid, newFrozenXid))
859 newFrozenXid = classForm->relfrozenxid;
862 /* we're done with pg_class */
863 systable_endscan(scan);
864 heap_close(relation, AccessShareLock);
866 Assert(TransactionIdIsNormal(newFrozenXid));
868 /* Now fetch the pg_database tuple we need to update. */
869 relation = heap_open(DatabaseRelationId, RowExclusiveLock);
871 /* Fetch a copy of the tuple to scribble on */
872 tuple = SearchSysCacheCopy(DATABASEOID,
873 ObjectIdGetDatum(MyDatabaseId),
875 if (!HeapTupleIsValid(tuple))
876 elog(ERROR, "could not find tuple for database %u", MyDatabaseId);
877 dbform = (Form_pg_database) GETSTRUCT(tuple);
880 * Don't allow datfrozenxid to go backward (probably can't happen anyway);
881 * and detect the common case where it doesn't go forward either.
883 if (TransactionIdPrecedes(dbform->datfrozenxid, newFrozenXid))
885 dbform->datfrozenxid = newFrozenXid;
890 heap_inplace_update(relation, tuple);
892 heap_freetuple(tuple);
893 heap_close(relation, RowExclusiveLock);
896 * If we were able to advance datfrozenxid, mark the flat-file copy of
897 * pg_database for update at commit, and see if we can truncate pg_clog.
901 database_file_update_needed();
902 vac_truncate_clog(newFrozenXid);
908 * vac_truncate_clog() -- attempt to truncate the commit log
910 * Scan pg_database to determine the system-wide oldest datfrozenxid,
911 * and use it to truncate the transaction commit log (pg_clog).
912 * Also update the XID wrap limit info maintained by varsup.c.
914 * The passed XID is simply the one I just wrote into my pg_database
915 * entry. It's used to initialize the "min" calculation.
917 * This routine is shared by full and lazy VACUUM. Note that it's
918 * only invoked when we've managed to change our DB's datfrozenxid
922 vac_truncate_clog(TransactionId frozenXID)
924 TransactionId myXID = GetCurrentTransactionId();
928 NameData oldest_datname;
929 bool frozenAlreadyWrapped = false;
931 /* init oldest_datname to sync with my frozenXID */
932 namestrcpy(&oldest_datname, get_database_name(MyDatabaseId));
935 * Scan pg_database to compute the minimum datfrozenxid
937 * Note: we need not worry about a race condition with new entries being
938 * inserted by CREATE DATABASE. Any such entry will have a copy of some
939 * existing DB's datfrozenxid, and that source DB cannot be ours because
940 * of the interlock against copying a DB containing an active backend.
941 * Hence the new entry will not reduce the minimum. Also, if two VACUUMs
942 * concurrently modify the datfrozenxid's of different databases, the
943 * worst possible outcome is that pg_clog is not truncated as aggressively
946 relation = heap_open(DatabaseRelationId, AccessShareLock);
948 scan = heap_beginscan(relation, SnapshotNow, 0, NULL);
950 while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
952 Form_pg_database dbform = (Form_pg_database) GETSTRUCT(tuple);
954 Assert(TransactionIdIsNormal(dbform->datfrozenxid));
956 if (TransactionIdPrecedes(myXID, dbform->datfrozenxid))
957 frozenAlreadyWrapped = true;
958 else if (TransactionIdPrecedes(dbform->datfrozenxid, frozenXID))
960 frozenXID = dbform->datfrozenxid;
961 namecpy(&oldest_datname, &dbform->datname);
967 heap_close(relation, AccessShareLock);
970 * Do not truncate CLOG if we seem to have suffered wraparound already;
971 * the computed minimum XID might be bogus. This case should now be
972 * impossible due to the defenses in GetNewTransactionId, but we keep the
975 if (frozenAlreadyWrapped)
978 (errmsg("some databases have not been vacuumed in over 2 billion transactions"),
979 errdetail("You might have already suffered transaction-wraparound data loss.")));
983 /* Truncate CLOG to the oldest frozenxid */
984 TruncateCLOG(frozenXID);
987 * Update the wrap limit for GetNewTransactionId. Note: this function
988 * will also signal the postmaster for an(other) autovac cycle if needed.
990 SetTransactionIdLimit(frozenXID, &oldest_datname);
994 /****************************************************************************
996 * Code common to both flavors of VACUUM *
998 ****************************************************************************
1003 * vacuum_rel() -- vacuum one heap relation
1005 * Doing one heap at a time incurs extra overhead, since we need to
1006 * check that the heap exists again just before we vacuum it. The
1007 * reason that we do this is so that vacuuming can be spread across
1008 * many small transactions. Otherwise, two-phase locking would require
1009 * us to lock the entire database during one pass of the vacuum cleaner.
1011 * We'll return true in *scanned_all if the vacuum scanned all heap
1012 * pages, and updated pg_class.
1014 * At entry and exit, we are not inside a transaction.
1017 vacuum_rel(Oid relid, VacuumStmt *vacstmt, bool do_toast, bool for_wraparound,
1025 bool save_secdefcxt;
1028 *scanned_all = false;
1030 /* Begin a transaction for vacuuming this relation */
1031 StartTransactionCommand();
1034 * Functions in indexes may want a snapshot set. Also, setting
1035 * a snapshot ensures that RecentGlobalXmin is kept truly recent.
1037 PushActiveSnapshot(GetTransactionSnapshot());
1042 * In lazy vacuum, we can set the PROC_IN_VACUUM flag, which lets other
1043 * concurrent VACUUMs know that they can ignore this one while
1044 * determining their OldestXmin. (The reason we don't set it during a
1045 * full VACUUM is exactly that we may have to run user- defined
1046 * functions for functional indexes, and we want to make sure that if
1047 * they use the snapshot set above, any tuples it requires can't get
1048 * removed from other tables. An index function that depends on the
1049 * contents of other tables is arguably broken, but we won't break it
1050 * here by violating transaction semantics.)
1052 * We also set the VACUUM_FOR_WRAPAROUND flag, which is passed down
1053 * by autovacuum; it's used to avoid cancelling a vacuum that was
1054 * invoked in an emergency.
1056 * Note: these flags remain set until CommitTransaction or
1057 * AbortTransaction. We don't want to clear them until we reset
1058 * MyProc->xid/xmin, else OldestXmin might appear to go backwards,
1059 * which is probably Not Good.
1061 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1062 MyProc->vacuumFlags |= PROC_IN_VACUUM;
1064 MyProc->vacuumFlags |= PROC_VACUUM_FOR_WRAPAROUND;
1065 LWLockRelease(ProcArrayLock);
1069 * Check for user-requested abort. Note we want this to be inside a
1070 * transaction, so xact.c doesn't issue useless WARNING.
1072 CHECK_FOR_INTERRUPTS();
1075 * Determine the type of lock we want --- hard exclusive lock for a FULL
1076 * vacuum, but just ShareUpdateExclusiveLock for concurrent vacuum. Either
1077 * way, we can be sure that no other backend is vacuuming the same table.
1079 lmode = vacstmt->full ? AccessExclusiveLock : ShareUpdateExclusiveLock;
1082 * Open the relation and get the appropriate lock on it.
1084 * There's a race condition here: the rel may have gone away since the
1085 * last time we saw it. If so, we don't need to vacuum it.
1087 onerel = try_relation_open(relid, lmode);
1091 PopActiveSnapshot();
1092 CommitTransactionCommand();
1097 * Check permissions.
1099 * We allow the user to vacuum a table if he is superuser, the table
1100 * owner, or the database owner (but in the latter case, only if it's not
1101 * a shared relation). pg_class_ownercheck includes the superuser case.
1103 * Note we choose to treat permissions failure as a WARNING and keep
1104 * trying to vacuum the rest of the DB --- is this appropriate?
1106 if (!(pg_class_ownercheck(RelationGetRelid(onerel), GetUserId()) ||
1107 (pg_database_ownercheck(MyDatabaseId, GetUserId()) && !onerel->rd_rel->relisshared)))
1109 if (onerel->rd_rel->relisshared)
1111 (errmsg("skipping \"%s\" --- only superuser can vacuum it",
1112 RelationGetRelationName(onerel))));
1113 else if (onerel->rd_rel->relnamespace == PG_CATALOG_NAMESPACE)
1115 (errmsg("skipping \"%s\" --- only superuser or database owner can vacuum it",
1116 RelationGetRelationName(onerel))));
1119 (errmsg("skipping \"%s\" --- only table or database owner can vacuum it",
1120 RelationGetRelationName(onerel))));
1121 relation_close(onerel, lmode);
1122 PopActiveSnapshot();
1123 CommitTransactionCommand();
1128 * Check that it's a vacuumable table; we used to do this in get_rel_oids()
1129 * but seems safer to check after we've locked the relation.
1131 if (onerel->rd_rel->relkind != RELKIND_RELATION &&
1132 onerel->rd_rel->relkind != RELKIND_TOASTVALUE)
1135 (errmsg("skipping \"%s\" --- cannot vacuum indexes, views, or special system tables",
1136 RelationGetRelationName(onerel))));
1137 relation_close(onerel, lmode);
1138 PopActiveSnapshot();
1139 CommitTransactionCommand();
1144 * Silently ignore tables that are temp tables of other backends ---
1145 * trying to vacuum these will lead to great unhappiness, since their
1146 * contents are probably not up-to-date on disk. (We don't throw a
1147 * warning here; it would just lead to chatter during a database-wide
1150 if (RELATION_IS_OTHER_TEMP(onerel))
1152 relation_close(onerel, lmode);
1153 PopActiveSnapshot();
1154 CommitTransactionCommand();
1159 * Get a session-level lock too. This will protect our access to the
1160 * relation across multiple transactions, so that we can vacuum the
1161 * relation's TOAST table (if any) secure in the knowledge that no one is
1162 * deleting the parent relation.
1164 * NOTE: this cannot block, even if someone else is waiting for access,
1165 * because the lock manager knows that both lock requests are from the
1168 onerelid = onerel->rd_lockInfo.lockRelId;
1169 LockRelationIdForSession(&onerelid, lmode);
1172 * Remember the relation's TOAST relation for later, if the caller asked
1176 toast_relid = onerel->rd_rel->reltoastrelid;
1178 toast_relid = InvalidOid;
1181 * Switch to the table owner's userid, so that any index functions are
1182 * run as that user. (This is unnecessary, but harmless, for lazy
1185 GetUserIdAndContext(&save_userid, &save_secdefcxt);
1186 SetUserIdAndContext(onerel->rd_rel->relowner, true);
1189 * Do the actual work --- either FULL or "lazy" vacuum
1192 full_vacuum_rel(onerel, vacstmt);
1194 lazy_vacuum_rel(onerel, vacstmt, vac_strategy, scanned_all);
1196 /* Restore userid */
1197 SetUserIdAndContext(save_userid, save_secdefcxt);
1199 /* all done with this class, but hold lock until commit */
1200 relation_close(onerel, NoLock);
1203 * Complete the transaction and free all temporary memory used.
1205 PopActiveSnapshot();
1206 CommitTransactionCommand();
1209 * If the relation has a secondary toast rel, vacuum that too while we
1210 * still hold the session lock on the master table. Note however that
1211 * "analyze" will not get done on the toast table. This is good, because
1212 * the toaster always uses hardcoded index access and statistics are
1213 * totally unimportant for toast relations.
1215 if (toast_relid != InvalidOid)
1216 vacuum_rel(toast_relid, vacstmt, false, for_wraparound, NULL);
1219 * Now release the session-level lock on the master table.
1221 UnlockRelationIdForSession(&onerelid, lmode);
1225 /****************************************************************************
1227 * Code for VACUUM FULL (only) *
1229 ****************************************************************************
1234 * full_vacuum_rel() -- perform FULL VACUUM for one heap relation
1236 * This routine vacuums a single heap, cleans out its indexes, and
1237 * updates its num_pages and num_tuples statistics.
1239 * At entry, we have already established a transaction and opened
1240 * and locked the relation.
1243 full_vacuum_rel(Relation onerel, VacuumStmt *vacstmt)
1245 VacPageListData vacuum_pages; /* List of pages to vacuum and/or
1247 VacPageListData fraged_pages; /* List of pages with space enough for
1252 VRelStats *vacrelstats;
1254 vacuum_set_xid_limits(vacstmt->freeze_min_age, vacstmt->freeze_table_age,
1255 onerel->rd_rel->relisshared,
1256 &OldestXmin, &FreezeLimit, NULL);
1259 * Flush any previous async-commit transactions. This does not guarantee
1260 * that we will be able to set hint bits for tuples they inserted, but it
1261 * improves the probability, especially in simple sequential-commands
1262 * cases. See scan_heap() and repair_frag() for more about this.
1264 XLogAsyncCommitFlush();
1267 * Set up statistics-gathering machinery.
1269 vacrelstats = (VRelStats *) palloc(sizeof(VRelStats));
1270 vacrelstats->rel_pages = 0;
1271 vacrelstats->rel_tuples = 0;
1272 vacrelstats->rel_indexed_tuples = 0;
1273 vacrelstats->hasindex = false;
1276 vacuum_pages.num_pages = fraged_pages.num_pages = 0;
1277 scan_heap(vacrelstats, onerel, &vacuum_pages, &fraged_pages);
1279 /* Now open all indexes of the relation */
1280 vac_open_indexes(onerel, AccessExclusiveLock, &nindexes, &Irel);
1282 vacrelstats->hasindex = true;
1284 /* Clean/scan index relation(s) */
1287 if (vacuum_pages.num_pages > 0)
1289 for (i = 0; i < nindexes; i++)
1290 vacuum_index(&vacuum_pages, Irel[i],
1291 vacrelstats->rel_indexed_tuples, 0);
1295 /* just scan indexes to update statistic */
1296 for (i = 0; i < nindexes; i++)
1297 scan_index(Irel[i], vacrelstats->rel_indexed_tuples);
1301 if (fraged_pages.num_pages > 0)
1303 /* Try to shrink heap */
1304 repair_frag(vacrelstats, onerel, &vacuum_pages, &fraged_pages,
1306 vac_close_indexes(nindexes, Irel, NoLock);
1310 vac_close_indexes(nindexes, Irel, NoLock);
1311 if (vacuum_pages.num_pages > 0)
1313 /* Clean pages from vacuum_pages list */
1314 vacuum_heap(vacrelstats, onerel, &vacuum_pages);
1318 /* update thefree space map with final free space info, and vacuum it */
1319 vac_update_fsm(onerel, &fraged_pages, vacrelstats->rel_pages);
1320 FreeSpaceMapVacuum(onerel);
1322 /* update statistics in pg_class */
1323 vac_update_relstats(onerel,
1324 vacrelstats->rel_pages, vacrelstats->rel_tuples,
1325 vacrelstats->hasindex, FreezeLimit);
1327 /* report results to the stats collector, too */
1328 pgstat_report_vacuum(RelationGetRelid(onerel), onerel->rd_rel->relisshared,
1329 true, vacstmt->analyze, vacrelstats->rel_tuples);
1334 * scan_heap() -- scan an open heap relation
1336 * This routine sets commit status bits, constructs vacuum_pages (list
1337 * of pages we need to compact free space on and/or clean indexes of
1338 * deleted tuples), constructs fraged_pages (list of pages with free
1339 * space that tuples could be moved into), and calculates statistics
1340 * on the number of live tuples in the heap.
1343 scan_heap(VRelStats *vacrelstats, Relation onerel,
1344 VacPageList vacuum_pages, VacPageList fraged_pages)
1346 BlockNumber nblocks,
1350 BlockNumber empty_pages,
1359 Size min_tlen = MaxHeapTupleSize;
1361 bool do_shrinking = true;
1362 VTupleLink vtlinks = (VTupleLink) palloc(100 * sizeof(VTupleLinkData));
1363 int num_vtlinks = 0;
1364 int free_vtlinks = 100;
1367 pg_rusage_init(&ru0);
1369 relname = RelationGetRelationName(onerel);
1371 (errmsg("vacuuming \"%s.%s\"",
1372 get_namespace_name(RelationGetNamespace(onerel)),
1375 empty_pages = empty_end_pages = 0;
1376 num_tuples = num_indexed_tuples = tups_vacuumed = nkeep = nunused = 0;
1379 nblocks = RelationGetNumberOfBlocks(onerel);
1382 * We initially create each VacPage item in a maximal-sized workspace,
1383 * then copy the workspace into a just-large-enough copy.
1385 vacpage = (VacPage) palloc(sizeof(VacPageData) + MaxOffsetNumber * sizeof(OffsetNumber));
1387 for (blkno = 0; blkno < nblocks; blkno++)
1394 OffsetNumber offnum,
1397 OffsetNumber frozen[MaxOffsetNumber];
1400 vacuum_delay_point();
1402 buf = ReadBufferExtended(onerel, MAIN_FORKNUM, blkno, RBM_NORMAL,
1404 page = BufferGetPage(buf);
1407 * Since we are holding exclusive lock on the relation, no other
1408 * backend can be accessing the page; however it is possible that the
1409 * background writer will try to write the page if it's already marked
1410 * dirty. To ensure that invalid data doesn't get written to disk, we
1411 * must take exclusive buffer lock wherever we potentially modify
1412 * pages. In fact, we insist on cleanup lock so that we can safely
1413 * call heap_page_prune(). (This might be overkill, since the
1414 * bgwriter pays no attention to individual tuples, but on the other
1415 * hand it's unlikely that the bgwriter has this particular page
1416 * pinned at this instant. So violating the coding rule would buy us
1419 LockBufferForCleanup(buf);
1421 vacpage->blkno = blkno;
1422 vacpage->offsets_used = 0;
1423 vacpage->offsets_free = 0;
1425 if (PageIsNew(page))
1427 VacPage vacpagecopy;
1430 (errmsg("relation \"%s\" page %u is uninitialized --- fixing",
1432 PageInit(page, BufferGetPageSize(buf), 0);
1433 MarkBufferDirty(buf);
1434 vacpage->free = PageGetFreeSpaceWithFillFactor(onerel, page);
1435 free_space += vacpage->free;
1438 vacpagecopy = copy_vac_page(vacpage);
1439 vpage_insert(vacuum_pages, vacpagecopy);
1440 vpage_insert(fraged_pages, vacpagecopy);
1441 UnlockReleaseBuffer(buf);
1445 if (PageIsEmpty(page))
1447 VacPage vacpagecopy;
1449 vacpage->free = PageGetFreeSpaceWithFillFactor(onerel, page);
1450 free_space += vacpage->free;
1453 vacpagecopy = copy_vac_page(vacpage);
1454 vpage_insert(vacuum_pages, vacpagecopy);
1455 vpage_insert(fraged_pages, vacpagecopy);
1456 UnlockReleaseBuffer(buf);
1461 * Prune all HOT-update chains in this page.
1463 * We use the redirect_move option so that redirecting line pointers
1464 * get collapsed out; this allows us to not worry about them below.
1466 * We count tuples removed by the pruning step as removed by VACUUM.
1468 tups_vacuumed += heap_page_prune(onerel, buf, OldestXmin,
1472 * Now scan the page to collect vacuumable items and check for tuples
1473 * requiring freezing.
1477 maxoff = PageGetMaxOffsetNumber(page);
1478 for (offnum = FirstOffsetNumber;
1480 offnum = OffsetNumberNext(offnum))
1482 ItemId itemid = PageGetItemId(page, offnum);
1483 bool tupgone = false;
1484 HeapTupleData tuple;
1487 * Collect un-used items too - it's possible to have indexes
1488 * pointing here after crash. (That's an ancient comment and is
1489 * likely obsolete with WAL, but we might as well continue to
1490 * check for such problems.)
1492 if (!ItemIdIsUsed(itemid))
1494 vacpage->offsets[vacpage->offsets_free++] = offnum;
1500 * DEAD item pointers are to be vacuumed normally; but we don't
1501 * count them in tups_vacuumed, else we'd be double-counting (at
1502 * least in the common case where heap_page_prune() just freed up
1505 if (ItemIdIsDead(itemid))
1507 vacpage->offsets[vacpage->offsets_free++] = offnum;
1511 /* Shouldn't have any redirected items anymore */
1512 if (!ItemIdIsNormal(itemid))
1513 elog(ERROR, "relation \"%s\" TID %u/%u: unexpected redirect item",
1514 relname, blkno, offnum);
1516 tuple.t_data = (HeapTupleHeader) PageGetItem(page, itemid);
1517 tuple.t_len = ItemIdGetLength(itemid);
1518 ItemPointerSet(&(tuple.t_self), blkno, offnum);
1520 switch (HeapTupleSatisfiesVacuum(tuple.t_data, OldestXmin, buf))
1522 case HEAPTUPLE_LIVE:
1523 /* Tuple is good --- but let's do some validity checks */
1524 if (onerel->rd_rel->relhasoids &&
1525 !OidIsValid(HeapTupleGetOid(&tuple)))
1526 elog(WARNING, "relation \"%s\" TID %u/%u: OID is invalid",
1527 relname, blkno, offnum);
1530 * The shrinkage phase of VACUUM FULL requires that all
1531 * live tuples have XMIN_COMMITTED set --- see comments in
1532 * repair_frag()'s walk-along-page loop. Use of async
1533 * commit may prevent HeapTupleSatisfiesVacuum from
1534 * setting the bit for a recently committed tuple. Rather
1535 * than trying to handle this corner case, we just give up
1539 !(tuple.t_data->t_infomask & HEAP_XMIN_COMMITTED))
1542 (errmsg("relation \"%s\" TID %u/%u: XMIN_COMMITTED not set for transaction %u --- cannot shrink relation",
1543 relname, blkno, offnum,
1544 HeapTupleHeaderGetXmin(tuple.t_data))));
1545 do_shrinking = false;
1548 case HEAPTUPLE_DEAD:
1551 * Ordinarily, DEAD tuples would have been removed by
1552 * heap_page_prune(), but it's possible that the tuple
1553 * state changed since heap_page_prune() looked. In
1554 * particular an INSERT_IN_PROGRESS tuple could have
1555 * changed to DEAD if the inserter aborted. So this
1556 * cannot be considered an error condition, though it does
1557 * suggest that someone released a lock early.
1559 * If the tuple is HOT-updated then it must only be
1560 * removed by a prune operation; so we keep it as if it
1561 * were RECENTLY_DEAD, and abandon shrinking. (XXX is it
1562 * worth trying to make the shrinking code smart enough to
1563 * handle this? It's an unusual corner case.)
1565 * DEAD heap-only tuples can safely be removed if they
1566 * aren't themselves HOT-updated, although this is a bit
1567 * inefficient since we'll uselessly try to remove index
1570 if (HeapTupleIsHotUpdated(&tuple))
1575 (errmsg("relation \"%s\" TID %u/%u: dead HOT-updated tuple --- cannot shrink relation",
1576 relname, blkno, offnum)));
1577 do_shrinking = false;
1581 tupgone = true; /* we can delete the tuple */
1584 * We need not require XMIN_COMMITTED or
1585 * XMAX_COMMITTED to be set, since we will remove the
1586 * tuple without any further examination of its hint
1591 case HEAPTUPLE_RECENTLY_DEAD:
1594 * If tuple is recently deleted then we must not remove it
1600 * As with the LIVE case, shrinkage requires
1601 * XMIN_COMMITTED to be set.
1604 !(tuple.t_data->t_infomask & HEAP_XMIN_COMMITTED))
1607 (errmsg("relation \"%s\" TID %u/%u: XMIN_COMMITTED not set for transaction %u --- cannot shrink relation",
1608 relname, blkno, offnum,
1609 HeapTupleHeaderGetXmin(tuple.t_data))));
1610 do_shrinking = false;
1614 * If we do shrinking and this tuple is updated one then
1615 * remember it to construct updated tuple dependencies.
1618 !(ItemPointerEquals(&(tuple.t_self),
1619 &(tuple.t_data->t_ctid))))
1621 if (free_vtlinks == 0)
1623 free_vtlinks = 1000;
1624 vtlinks = (VTupleLink) repalloc(vtlinks,
1625 (free_vtlinks + num_vtlinks) *
1626 sizeof(VTupleLinkData));
1628 vtlinks[num_vtlinks].new_tid = tuple.t_data->t_ctid;
1629 vtlinks[num_vtlinks].this_tid = tuple.t_self;
1634 case HEAPTUPLE_INSERT_IN_PROGRESS:
1637 * This should not happen, since we hold exclusive lock on
1638 * the relation; shouldn't we raise an error? (Actually,
1639 * it can happen in system catalogs, since we tend to
1640 * release write lock before commit there.) As above, we
1641 * can't apply repair_frag() if the tuple state is
1646 (errmsg("relation \"%s\" TID %u/%u: InsertTransactionInProgress %u --- cannot shrink relation",
1647 relname, blkno, offnum,
1648 HeapTupleHeaderGetXmin(tuple.t_data))));
1649 do_shrinking = false;
1651 case HEAPTUPLE_DELETE_IN_PROGRESS:
1654 * This should not happen, since we hold exclusive lock on
1655 * the relation; shouldn't we raise an error? (Actually,
1656 * it can happen in system catalogs, since we tend to
1657 * release write lock before commit there.) As above, we
1658 * can't apply repair_frag() if the tuple state is
1663 (errmsg("relation \"%s\" TID %u/%u: DeleteTransactionInProgress %u --- cannot shrink relation",
1664 relname, blkno, offnum,
1665 HeapTupleHeaderGetXmax(tuple.t_data))));
1666 do_shrinking = false;
1669 elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
1678 * Here we are building a temporary copy of the page with dead
1679 * tuples removed. Below we will apply
1680 * PageRepairFragmentation to the copy, so that we can
1681 * determine how much space will be available after removal of
1682 * dead tuples. But note we are NOT changing the real page
1685 if (tempPage == NULL)
1689 pageSize = PageGetPageSize(page);
1690 tempPage = (Page) palloc(pageSize);
1691 memcpy(tempPage, page, pageSize);
1694 /* mark it unused on the temp page */
1695 lpp = PageGetItemId(tempPage, offnum);
1696 ItemIdSetUnused(lpp);
1698 vacpage->offsets[vacpage->offsets_free++] = offnum;
1704 if (!HeapTupleIsHeapOnly(&tuple))
1705 num_indexed_tuples += 1;
1707 if (tuple.t_len < min_tlen)
1708 min_tlen = tuple.t_len;
1709 if (tuple.t_len > max_tlen)
1710 max_tlen = tuple.t_len;
1713 * Each non-removable tuple must be checked to see if it needs
1716 if (heap_freeze_tuple(tuple.t_data, FreezeLimit,
1718 frozen[nfrozen++] = offnum;
1720 } /* scan along page */
1722 if (tempPage != NULL)
1724 /* Some tuples are removable; figure free space after removal */
1725 PageRepairFragmentation(tempPage);
1726 vacpage->free = PageGetFreeSpaceWithFillFactor(onerel, tempPage);
1732 /* Just use current available space */
1733 vacpage->free = PageGetFreeSpaceWithFillFactor(onerel, page);
1734 /* Need to reap the page if it has UNUSED or DEAD line pointers */
1735 do_reap = (vacpage->offsets_free > 0);
1738 free_space += vacpage->free;
1741 * Add the page to vacuum_pages if it requires reaping, and add it to
1742 * fraged_pages if it has a useful amount of free space. "Useful"
1743 * means enough for a minimal-sized tuple. But we don't know that
1744 * accurately near the start of the relation, so add pages
1745 * unconditionally if they have >= BLCKSZ/10 free space. Also
1746 * forcibly add pages with no live tuples, to avoid confusing the
1747 * empty_end_pages logic. (In the presence of unreasonably small
1748 * fillfactor, it seems possible that such pages might not pass
1749 * the free-space test, but they had better be in the list anyway.)
1751 do_frag = (vacpage->free >= min_tlen || vacpage->free >= BLCKSZ / 10 ||
1754 if (do_reap || do_frag)
1756 VacPage vacpagecopy = copy_vac_page(vacpage);
1759 vpage_insert(vacuum_pages, vacpagecopy);
1761 vpage_insert(fraged_pages, vacpagecopy);
1765 * Include the page in empty_end_pages if it will be empty after
1766 * vacuuming; this is to keep us from using it as a move destination.
1767 * Note that such pages are guaranteed to be in fraged_pages.
1775 empty_end_pages = 0;
1778 * If we froze any tuples, mark the buffer dirty, and write a WAL
1779 * record recording the changes. We must log the changes to be
1780 * crash-safe against future truncation of CLOG.
1784 MarkBufferDirty(buf);
1785 /* no XLOG for temp tables, though */
1786 if (!onerel->rd_istemp)
1790 recptr = log_heap_freeze(onerel, buf, FreezeLimit,
1792 PageSetLSN(page, recptr);
1793 PageSetTLI(page, ThisTimeLineID);
1797 UnlockReleaseBuffer(buf);
1802 /* save stats in the rel list for use later */
1803 vacrelstats->rel_tuples = num_tuples;
1804 vacrelstats->rel_indexed_tuples = num_indexed_tuples;
1805 vacrelstats->rel_pages = nblocks;
1806 if (num_tuples == 0)
1807 min_tlen = max_tlen = 0;
1808 vacrelstats->min_tlen = min_tlen;
1809 vacrelstats->max_tlen = max_tlen;
1811 vacuum_pages->empty_end_pages = empty_end_pages;
1812 fraged_pages->empty_end_pages = empty_end_pages;
1815 * Clear the fraged_pages list if we found we couldn't shrink. Else,
1816 * remove any "empty" end-pages from the list, and compute usable free
1817 * space = free space in remaining pages.
1823 Assert((BlockNumber) fraged_pages->num_pages >= empty_end_pages);
1824 fraged_pages->num_pages -= empty_end_pages;
1825 usable_free_space = 0;
1826 for (i = 0; i < fraged_pages->num_pages; i++)
1827 usable_free_space += fraged_pages->pagedesc[i]->free;
1831 fraged_pages->num_pages = 0;
1832 usable_free_space = 0;
1835 /* don't bother to save vtlinks if we will not call repair_frag */
1836 if (fraged_pages->num_pages > 0 && num_vtlinks > 0)
1838 qsort((char *) vtlinks, num_vtlinks, sizeof(VTupleLinkData),
1840 vacrelstats->vtlinks = vtlinks;
1841 vacrelstats->num_vtlinks = num_vtlinks;
1845 vacrelstats->vtlinks = NULL;
1846 vacrelstats->num_vtlinks = 0;
1851 (errmsg("\"%s\": found %.0f removable, %.0f nonremovable row versions in %u pages",
1852 RelationGetRelationName(onerel),
1853 tups_vacuumed, num_tuples, nblocks),
1854 errdetail("%.0f dead row versions cannot be removed yet.\n"
1855 "Nonremovable row versions range from %lu to %lu bytes long.\n"
1856 "There were %.0f unused item pointers.\n"
1857 "Total free space (including removable row versions) is %.0f bytes.\n"
1858 "%u pages are or will become empty, including %u at the end of the table.\n"
1859 "%u pages containing %.0f free bytes are potential move destinations.\n"
1862 (unsigned long) min_tlen, (unsigned long) max_tlen,
1865 empty_pages, empty_end_pages,
1866 fraged_pages->num_pages, usable_free_space,
1867 pg_rusage_show(&ru0))));
1872 * repair_frag() -- try to repair relation's fragmentation
1874 * This routine marks dead tuples as unused and tries re-use dead space
1875 * by moving tuples (and inserting indexes if needed). It constructs
1876 * Nvacpagelist list of free-ed pages (moved tuples) and clean indexes
1877 * for them after committing (in hack-manner - without losing locks
1878 * and freeing memory!) current transaction. It truncates relation
1879 * if some end-blocks are gone away.
1882 repair_frag(VRelStats *vacrelstats, Relation onerel,
1883 VacPageList vacuum_pages, VacPageList fraged_pages,
1884 int nindexes, Relation *Irel)
1886 TransactionId myXID = GetCurrentTransactionId();
1887 Buffer dst_buffer = InvalidBuffer;
1888 BlockNumber nblocks,
1890 BlockNumber last_move_dest_block = 0,
1892 Page dst_page = NULL;
1894 VacPageListData Nvacpagelist;
1895 VacPage dst_vacpage = NULL,
1903 int keep_tuples = 0;
1904 int keep_indexed_tuples = 0;
1907 pg_rusage_init(&ru0);
1909 ExecContext_Init(&ec, onerel);
1911 Nvacpagelist.num_pages = 0;
1912 num_fraged_pages = fraged_pages->num_pages;
1913 Assert((BlockNumber) vacuum_pages->num_pages >= vacuum_pages->empty_end_pages);
1914 vacuumed_pages = vacuum_pages->num_pages - vacuum_pages->empty_end_pages;
1915 if (vacuumed_pages > 0)
1917 /* get last reaped page from vacuum_pages */
1918 last_vacuum_page = vacuum_pages->pagedesc[vacuumed_pages - 1];
1919 last_vacuum_block = last_vacuum_page->blkno;
1923 last_vacuum_page = NULL;
1924 last_vacuum_block = InvalidBlockNumber;
1927 vacpage = (VacPage) palloc(sizeof(VacPageData) + MaxOffsetNumber * sizeof(OffsetNumber));
1928 vacpage->offsets_used = vacpage->offsets_free = 0;
1931 * Scan pages backwards from the last nonempty page, trying to move tuples
1932 * down to lower pages. Quit when we reach a page that we have moved any
1933 * tuples onto, or the first page if we haven't moved anything, or when we
1934 * find a page we cannot completely empty (this last condition is handled
1935 * by "break" statements within the loop).
1937 * NB: this code depends on the vacuum_pages and fraged_pages lists being
1938 * in order by blkno.
1940 nblocks = vacrelstats->rel_pages;
1941 for (blkno = nblocks - vacuum_pages->empty_end_pages - 1;
1942 blkno > last_move_dest_block;
1947 OffsetNumber offnum,
1952 vacuum_delay_point();
1955 * Forget fraged_pages pages at or after this one; they're no longer
1956 * useful as move targets, since we only want to move down. Note that
1957 * since we stop the outer loop at last_move_dest_block, pages removed
1958 * here cannot have had anything moved onto them already.
1960 * Also note that we don't change the stored fraged_pages list, only
1961 * our local variable num_fraged_pages; so the forgotten pages are
1962 * still available to be loaded into the free space map later.
1964 while (num_fraged_pages > 0 &&
1965 fraged_pages->pagedesc[num_fraged_pages - 1]->blkno >= blkno)
1967 Assert(fraged_pages->pagedesc[num_fraged_pages - 1]->offsets_used == 0);
1972 * Process this page of relation.
1974 buf = ReadBufferExtended(onerel, MAIN_FORKNUM, blkno, RBM_NORMAL,
1976 page = BufferGetPage(buf);
1978 vacpage->offsets_free = 0;
1980 isempty = PageIsEmpty(page);
1982 /* Is the page in the vacuum_pages list? */
1983 if (blkno == last_vacuum_block)
1985 if (last_vacuum_page->offsets_free > 0)
1987 /* there are dead tuples on this page - clean them */
1989 LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
1990 vacuum_page(onerel, buf, last_vacuum_page);
1991 LockBuffer(buf, BUFFER_LOCK_UNLOCK);
1996 if (vacuumed_pages > 0)
1998 /* get prev reaped page from vacuum_pages */
1999 last_vacuum_page = vacuum_pages->pagedesc[vacuumed_pages - 1];
2000 last_vacuum_block = last_vacuum_page->blkno;
2004 last_vacuum_page = NULL;
2005 last_vacuum_block = InvalidBlockNumber;
2016 chain_tuple_moved = false; /* no one chain-tuple was moved off
2018 vacpage->blkno = blkno;
2019 maxoff = PageGetMaxOffsetNumber(page);
2020 for (offnum = FirstOffsetNumber;
2022 offnum = OffsetNumberNext(offnum))
2025 HeapTupleData tuple;
2026 ItemId itemid = PageGetItemId(page, offnum);
2028 if (!ItemIdIsUsed(itemid))
2031 if (ItemIdIsDead(itemid))
2033 /* just remember it for vacuum_page() */
2034 vacpage->offsets[vacpage->offsets_free++] = offnum;
2038 /* Shouldn't have any redirected items now */
2039 Assert(ItemIdIsNormal(itemid));
2041 tuple.t_data = (HeapTupleHeader) PageGetItem(page, itemid);
2042 tuple_len = tuple.t_len = ItemIdGetLength(itemid);
2043 ItemPointerSet(&(tuple.t_self), blkno, offnum);
2046 * VACUUM FULL has an exclusive lock on the relation. So
2047 * normally no other transaction can have pending INSERTs or
2048 * DELETEs in this relation. A tuple is either:
2049 * (a) live (XMIN_COMMITTED)
2050 * (b) known dead (XMIN_INVALID, or XMAX_COMMITTED and xmax
2051 * is visible to all active transactions)
2052 * (c) inserted and deleted (XMIN_COMMITTED+XMAX_COMMITTED)
2053 * but at least one active transaction does not see the
2054 * deleting transaction (ie, it's RECENTLY_DEAD)
2055 * (d) moved by the currently running VACUUM
2056 * (e) inserted or deleted by a not yet committed transaction,
2057 * or by a transaction we couldn't set XMIN_COMMITTED for.
2058 * In case (e) we wouldn't be in repair_frag() at all, because
2059 * scan_heap() detects those cases and shuts off shrinking.
2060 * We can't see case (b) here either, because such tuples were
2061 * already removed by vacuum_page(). Cases (a) and (c) are
2062 * normal and will have XMIN_COMMITTED set. Case (d) is only
2063 * possible if a whole tuple chain has been moved while
2064 * processing this or a higher numbered block.
2067 if (!(tuple.t_data->t_infomask & HEAP_XMIN_COMMITTED))
2069 if (tuple.t_data->t_infomask & HEAP_MOVED_IN)
2070 elog(ERROR, "HEAP_MOVED_IN was not expected");
2071 if (!(tuple.t_data->t_infomask & HEAP_MOVED_OFF))
2072 elog(ERROR, "HEAP_MOVED_OFF was expected");
2075 * MOVED_OFF by another VACUUM would have caused the
2076 * visibility check to set XMIN_COMMITTED or XMIN_INVALID.
2078 if (HeapTupleHeaderGetXvac(tuple.t_data) != myXID)
2079 elog(ERROR, "invalid XVAC in tuple header");
2082 * If this (chain) tuple is moved by me already then I have to
2083 * check is it in vacpage or not - i.e. is it moved while
2084 * cleaning this page or some previous one.
2087 /* Can't we Assert(keep_tuples > 0) here? */
2088 if (keep_tuples == 0)
2090 if (chain_tuple_moved)
2092 /* some chains were moved while cleaning this page */
2093 Assert(vacpage->offsets_free > 0);
2094 for (i = 0; i < vacpage->offsets_free; i++)
2096 if (vacpage->offsets[i] == offnum)
2099 if (i >= vacpage->offsets_free) /* not found */
2101 vacpage->offsets[vacpage->offsets_free++] = offnum;
2104 * If this is not a heap-only tuple, there must be an
2105 * index entry for this item which will be removed in
2106 * the index cleanup. Decrement the
2107 * keep_indexed_tuples count to remember this.
2109 if (!HeapTupleHeaderIsHeapOnly(tuple.t_data))
2110 keep_indexed_tuples--;
2116 vacpage->offsets[vacpage->offsets_free++] = offnum;
2119 * If this is not a heap-only tuple, there must be an
2120 * index entry for this item which will be removed in the
2121 * index cleanup. Decrement the keep_indexed_tuples count
2124 if (!HeapTupleHeaderIsHeapOnly(tuple.t_data))
2125 keep_indexed_tuples--;
2132 * If this tuple is in a chain of tuples created in updates by
2133 * "recent" transactions then we have to move the whole chain of
2134 * tuples to other places, so that we can write new t_ctid links
2135 * that preserve the chain relationship.
2137 * This test is complicated. Read it as "if tuple is a recently
2138 * created updated version, OR if it is an obsoleted version". (In
2139 * the second half of the test, we needn't make any check on XMAX
2140 * --- it must be recently obsoleted, else scan_heap would have
2141 * deemed it removable.)
2143 * NOTE: this test is not 100% accurate: it is possible for a
2144 * tuple to be an updated one with recent xmin, and yet not match
2145 * any new_tid entry in the vtlinks list. Presumably there was
2146 * once a parent tuple with xmax matching the xmin, but it's
2147 * possible that that tuple has been removed --- for example, if
2148 * it had xmin = xmax and wasn't itself an updated version, then
2149 * HeapTupleSatisfiesVacuum would deem it removable as soon as the
2150 * xmin xact completes.
2152 * To be on the safe side, we abandon the repair_frag process if
2153 * we cannot find the parent tuple in vtlinks. This may be overly
2154 * conservative; AFAICS it would be safe to move the chain.
2156 * Also, because we distinguish DEAD and RECENTLY_DEAD tuples
2157 * using OldestXmin, which is a rather coarse test, it is quite
2158 * possible to have an update chain in which a tuple we think is
2159 * RECENTLY_DEAD links forward to one that is definitely DEAD. In
2160 * such a case the RECENTLY_DEAD tuple must actually be dead, but
2161 * it seems too complicated to try to make VACUUM remove it. We
2162 * treat each contiguous set of RECENTLY_DEAD tuples as a
2163 * separately movable chain, ignoring any intervening DEAD ones.
2165 if (((tuple.t_data->t_infomask & HEAP_UPDATED) &&
2166 !TransactionIdPrecedes(HeapTupleHeaderGetXmin(tuple.t_data),
2168 (!(tuple.t_data->t_infomask & (HEAP_XMAX_INVALID |
2170 !(ItemPointerEquals(&(tuple.t_self),
2171 &(tuple.t_data->t_ctid)))))
2174 bool freeCbuf = false;
2175 bool chain_move_failed = false;
2176 bool moved_target = false;
2177 ItemPointerData Ctid;
2178 HeapTupleData tp = tuple;
2179 Size tlen = tuple_len;
2183 VacPage to_vacpage = NULL;
2187 if (dst_buffer != InvalidBuffer)
2189 ReleaseBuffer(dst_buffer);
2190 dst_buffer = InvalidBuffer;
2193 /* Quick exit if we have no vtlinks to search in */
2194 if (vacrelstats->vtlinks == NULL)
2196 elog(DEBUG2, "parent item in update-chain not found --- cannot continue repair_frag");
2197 break; /* out of walk-along-page loop */
2201 * If this tuple is in the begin/middle of the chain then we
2202 * have to move to the end of chain. As with any t_ctid
2203 * chase, we have to verify that each new tuple is really the
2204 * descendant of the tuple we came from; however, here we need
2205 * even more than the normal amount of paranoia. If t_ctid
2206 * links forward to a tuple determined to be DEAD, then
2207 * depending on where that tuple is, it might already have
2208 * been removed, and perhaps even replaced by a MOVED_IN
2209 * tuple. We don't want to include any DEAD tuples in the
2210 * chain, so we have to recheck HeapTupleSatisfiesVacuum.
2212 while (!(tp.t_data->t_infomask & (HEAP_XMAX_INVALID |
2214 !(ItemPointerEquals(&(tp.t_self),
2215 &(tp.t_data->t_ctid))))
2217 ItemPointerData nextTid;
2218 TransactionId priorXmax;
2221 OffsetNumber nextOffnum;
2223 HeapTupleHeader nextTdata;
2224 HTSV_Result nextTstatus;
2226 nextTid = tp.t_data->t_ctid;
2227 priorXmax = HeapTupleHeaderGetXmax(tp.t_data);
2228 /* assume block# is OK (see heap_fetch comments) */
2229 nextBuf = ReadBufferExtended(onerel, MAIN_FORKNUM,
2230 ItemPointerGetBlockNumber(&nextTid),
2231 RBM_NORMAL, vac_strategy);
2232 nextPage = BufferGetPage(nextBuf);
2233 /* If bogus or unused slot, assume tp is end of chain */
2234 nextOffnum = ItemPointerGetOffsetNumber(&nextTid);
2235 if (nextOffnum < FirstOffsetNumber ||
2236 nextOffnum > PageGetMaxOffsetNumber(nextPage))
2238 ReleaseBuffer(nextBuf);
2241 nextItemid = PageGetItemId(nextPage, nextOffnum);
2242 if (!ItemIdIsNormal(nextItemid))
2244 ReleaseBuffer(nextBuf);
2247 /* if not matching XMIN, assume tp is end of chain */
2248 nextTdata = (HeapTupleHeader) PageGetItem(nextPage,
2250 if (!TransactionIdEquals(HeapTupleHeaderGetXmin(nextTdata),
2253 ReleaseBuffer(nextBuf);
2258 * Must check for DEAD or MOVED_IN tuple, too. This could
2259 * potentially update hint bits, so we'd better hold the
2260 * buffer content lock.
2262 LockBuffer(nextBuf, BUFFER_LOCK_SHARE);
2263 nextTstatus = HeapTupleSatisfiesVacuum(nextTdata,
2266 if (nextTstatus == HEAPTUPLE_DEAD ||
2267 nextTstatus == HEAPTUPLE_INSERT_IN_PROGRESS)
2269 UnlockReleaseBuffer(nextBuf);
2272 LockBuffer(nextBuf, BUFFER_LOCK_UNLOCK);
2273 /* if it's MOVED_OFF we shoulda moved this one with it */
2274 if (nextTstatus == HEAPTUPLE_DELETE_IN_PROGRESS)
2275 elog(ERROR, "updated tuple is already HEAP_MOVED_OFF");
2276 /* OK, switch our attention to the next tuple in chain */
2277 tp.t_data = nextTdata;
2278 tp.t_self = nextTid;
2279 tlen = tp.t_len = ItemIdGetLength(nextItemid);
2281 ReleaseBuffer(Cbuf);
2286 /* Set up workspace for planning the chain move */
2287 vtmove = (VTupleMove) palloc(100 * sizeof(VTupleMoveData));
2292 * Now, walk backwards up the chain (towards older tuples) and
2293 * check if all items in chain can be moved. We record all
2294 * the moves that need to be made in the vtmove array.
2301 HeapTupleHeader PTdata;
2302 VTupleLinkData vtld,
2305 /* Identify a target page to move this tuple to */
2306 if (to_vacpage == NULL ||
2307 !enough_space(to_vacpage, tlen))
2309 for (i = 0; i < num_fraged_pages; i++)
2311 if (enough_space(fraged_pages->pagedesc[i], tlen))
2315 if (i == num_fraged_pages)
2317 /* can't move item anywhere */
2318 chain_move_failed = true;
2319 break; /* out of check-all-items loop */
2322 to_vacpage = fraged_pages->pagedesc[to_item];
2324 to_vacpage->free -= MAXALIGN(tlen);
2325 if (to_vacpage->offsets_used >= to_vacpage->offsets_free)
2326 to_vacpage->free -= sizeof(ItemIdData);
2327 (to_vacpage->offsets_used)++;
2329 /* Add an entry to vtmove list */
2330 if (free_vtmove == 0)
2333 vtmove = (VTupleMove)
2335 (free_vtmove + num_vtmove) *
2336 sizeof(VTupleMoveData));
2338 vtmove[num_vtmove].tid = tp.t_self;
2339 vtmove[num_vtmove].vacpage = to_vacpage;
2340 if (to_vacpage->offsets_used == 1)
2341 vtmove[num_vtmove].cleanVpd = true;
2343 vtmove[num_vtmove].cleanVpd = false;
2347 /* Remember if we reached the original target tuple */
2348 if (ItemPointerGetBlockNumber(&tp.t_self) == blkno &&
2349 ItemPointerGetOffsetNumber(&tp.t_self) == offnum)
2350 moved_target = true;
2352 /* Done if at beginning of chain */
2353 if (!(tp.t_data->t_infomask & HEAP_UPDATED) ||
2354 TransactionIdPrecedes(HeapTupleHeaderGetXmin(tp.t_data),
2356 break; /* out of check-all-items loop */
2358 /* Move to tuple with prior row version */
2359 vtld.new_tid = tp.t_self;
2361 vac_bsearch((void *) &vtld,
2362 (void *) (vacrelstats->vtlinks),
2363 vacrelstats->num_vtlinks,
2364 sizeof(VTupleLinkData),
2368 /* see discussion above */
2369 elog(DEBUG2, "parent item in update-chain not found --- cannot continue repair_frag");
2370 chain_move_failed = true;
2371 break; /* out of check-all-items loop */
2373 tp.t_self = vtlp->this_tid;
2374 Pbuf = ReadBufferExtended(onerel, MAIN_FORKNUM,
2375 ItemPointerGetBlockNumber(&(tp.t_self)),
2376 RBM_NORMAL, vac_strategy);
2377 Ppage = BufferGetPage(Pbuf);
2378 Pitemid = PageGetItemId(Ppage,
2379 ItemPointerGetOffsetNumber(&(tp.t_self)));
2380 /* this can't happen since we saw tuple earlier: */
2381 if (!ItemIdIsNormal(Pitemid))
2382 elog(ERROR, "parent itemid marked as unused");
2383 PTdata = (HeapTupleHeader) PageGetItem(Ppage, Pitemid);
2385 /* ctid should not have changed since we saved it */
2386 Assert(ItemPointerEquals(&(vtld.new_tid),
2387 &(PTdata->t_ctid)));
2390 * Read above about cases when !ItemIdIsUsed(nextItemid)
2391 * (child item is removed)... Due to the fact that at the
2392 * moment we don't remove unuseful part of update-chain,
2393 * it's possible to get non-matching parent row here. Like
2394 * as in the case which caused this problem, we stop
2395 * shrinking here. I could try to find real parent row but
2396 * want not to do it because of real solution will be
2397 * implemented anyway, later, and we are too close to 6.5
2398 * release. - vadim 06/11/99
2400 if ((PTdata->t_infomask & HEAP_XMAX_IS_MULTI) ||
2401 !(TransactionIdEquals(HeapTupleHeaderGetXmax(PTdata),
2402 HeapTupleHeaderGetXmin(tp.t_data))))
2404 ReleaseBuffer(Pbuf);
2405 elog(DEBUG2, "too old parent tuple found --- cannot continue repair_frag");
2406 chain_move_failed = true;
2407 break; /* out of check-all-items loop */
2410 tlen = tp.t_len = ItemIdGetLength(Pitemid);
2412 ReleaseBuffer(Cbuf);
2415 } /* end of check-all-items loop */
2418 ReleaseBuffer(Cbuf);
2421 /* Double-check that we will move the current target tuple */
2422 if (!moved_target && !chain_move_failed)
2424 elog(DEBUG2, "failed to chain back to target --- cannot continue repair_frag");
2425 chain_move_failed = true;
2428 if (chain_move_failed)
2431 * Undo changes to offsets_used state. We don't bother
2432 * cleaning up the amount-free state, since we're not
2433 * going to do any further tuple motion.
2435 for (i = 0; i < num_vtmove; i++)
2437 Assert(vtmove[i].vacpage->offsets_used > 0);
2438 (vtmove[i].vacpage->offsets_used)--;
2441 break; /* out of walk-along-page loop */
2445 * Okay, move the whole tuple chain in reverse order.
2447 * Ctid tracks the new location of the previously-moved tuple.
2449 ItemPointerSetInvalid(&Ctid);
2450 for (ti = 0; ti < num_vtmove; ti++)
2452 VacPage destvacpage = vtmove[ti].vacpage;
2456 /* Get page to move from */
2457 tuple.t_self = vtmove[ti].tid;
2458 Cbuf = ReadBufferExtended(onerel, MAIN_FORKNUM,
2459 ItemPointerGetBlockNumber(&(tuple.t_self)),
2460 RBM_NORMAL, vac_strategy);
2462 /* Get page to move to */
2463 dst_buffer = ReadBufferExtended(onerel, MAIN_FORKNUM,
2465 RBM_NORMAL, vac_strategy);
2467 LockBuffer(dst_buffer, BUFFER_LOCK_EXCLUSIVE);
2468 if (dst_buffer != Cbuf)
2469 LockBuffer(Cbuf, BUFFER_LOCK_EXCLUSIVE);
2471 dst_page = BufferGetPage(dst_buffer);
2472 Cpage = BufferGetPage(Cbuf);
2474 Citemid = PageGetItemId(Cpage,
2475 ItemPointerGetOffsetNumber(&(tuple.t_self)));
2476 tuple.t_data = (HeapTupleHeader) PageGetItem(Cpage, Citemid);
2477 tuple_len = tuple.t_len = ItemIdGetLength(Citemid);
2479 move_chain_tuple(onerel, Cbuf, Cpage, &tuple,
2480 dst_buffer, dst_page, destvacpage,
2481 &ec, &Ctid, vtmove[ti].cleanVpd);
2484 * If the tuple we are moving is a heap-only tuple, this
2485 * move will generate an additional index entry, so
2486 * increment the rel_indexed_tuples count.
2488 if (HeapTupleHeaderIsHeapOnly(tuple.t_data))
2489 vacrelstats->rel_indexed_tuples++;
2492 if (destvacpage->blkno > last_move_dest_block)
2493 last_move_dest_block = destvacpage->blkno;
2496 * Remember that we moved tuple from the current page
2497 * (corresponding index tuple will be cleaned).
2500 vacpage->offsets[vacpage->offsets_free++] =
2501 ItemPointerGetOffsetNumber(&(tuple.t_self));
2505 * When we move tuple chains, we may need to move
2506 * tuples from a block that we haven't yet scanned in
2507 * the outer walk-along-the-relation loop. Note that
2508 * we can't be moving a tuple from a block that we
2509 * have already scanned because if such a tuple
2510 * exists, then we must have moved the chain along
2511 * with that tuple when we scanned that block. IOW the
2512 * test of (Cbuf != buf) guarantees that the tuple we
2513 * are looking at right now is in a block which is yet
2516 * We maintain two counters to correctly count the
2517 * moved-off tuples from blocks that are not yet
2518 * scanned (keep_tuples) and how many of them have
2519 * index pointers (keep_indexed_tuples). The main
2520 * reason to track the latter is to help verify that
2521 * indexes have the expected number of entries when
2522 * all the dust settles.
2524 if (!HeapTupleHeaderIsHeapOnly(tuple.t_data))
2525 keep_indexed_tuples++;
2529 ReleaseBuffer(dst_buffer);
2530 ReleaseBuffer(Cbuf);
2531 } /* end of move-the-tuple-chain loop */
2533 dst_buffer = InvalidBuffer;
2535 chain_tuple_moved = true;
2537 /* advance to next tuple in walk-along-page loop */
2539 } /* end of is-tuple-in-chain test */
2541 /* try to find new page for this tuple */
2542 if (dst_buffer == InvalidBuffer ||
2543 !enough_space(dst_vacpage, tuple_len))
2545 if (dst_buffer != InvalidBuffer)
2547 ReleaseBuffer(dst_buffer);
2548 dst_buffer = InvalidBuffer;
2550 for (i = 0; i < num_fraged_pages; i++)
2552 if (enough_space(fraged_pages->pagedesc[i], tuple_len))
2555 if (i == num_fraged_pages)
2556 break; /* can't move item anywhere */
2557 dst_vacpage = fraged_pages->pagedesc[i];
2558 dst_buffer = ReadBufferExtended(onerel, MAIN_FORKNUM,
2560 RBM_NORMAL, vac_strategy);
2561 LockBuffer(dst_buffer, BUFFER_LOCK_EXCLUSIVE);
2562 dst_page = BufferGetPage(dst_buffer);
2563 /* if this page was not used before - clean it */
2564 if (!PageIsEmpty(dst_page) && dst_vacpage->offsets_used == 0)
2565 vacuum_page(onerel, dst_buffer, dst_vacpage);
2568 LockBuffer(dst_buffer, BUFFER_LOCK_EXCLUSIVE);
2570 LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
2572 move_plain_tuple(onerel, buf, page, &tuple,
2573 dst_buffer, dst_page, dst_vacpage, &ec);
2576 * If the tuple we are moving is a heap-only tuple, this move will
2577 * generate an additional index entry, so increment the
2578 * rel_indexed_tuples count.
2580 if (HeapTupleHeaderIsHeapOnly(tuple.t_data))
2581 vacrelstats->rel_indexed_tuples++;
2584 if (dst_vacpage->blkno > last_move_dest_block)
2585 last_move_dest_block = dst_vacpage->blkno;
2588 * Remember that we moved tuple from the current page
2589 * (corresponding index tuple will be cleaned).
2591 vacpage->offsets[vacpage->offsets_free++] = offnum;
2592 } /* walk along page */
2595 * If we broke out of the walk-along-page loop early (ie, still have
2596 * offnum <= maxoff), then we failed to move some tuple off this page.
2597 * No point in shrinking any more, so clean up and exit the per-page
2600 if (offnum < maxoff && keep_tuples > 0)
2605 * Fix vacpage state for any unvisited tuples remaining on page
2607 for (off = OffsetNumberNext(offnum);
2609 off = OffsetNumberNext(off))
2611 ItemId itemid = PageGetItemId(page, off);
2612 HeapTupleHeader htup;
2614 if (!ItemIdIsUsed(itemid))
2616 /* Shouldn't be any DEAD or REDIRECT items anymore */
2617 Assert(ItemIdIsNormal(itemid));
2619 htup = (HeapTupleHeader) PageGetItem(page, itemid);
2620 if (htup->t_infomask & HEAP_XMIN_COMMITTED)
2624 * See comments in the walk-along-page loop above about why
2625 * only MOVED_OFF tuples should be found here.
2627 if (htup->t_infomask & HEAP_MOVED_IN)
2628 elog(ERROR, "HEAP_MOVED_IN was not expected");
2629 if (!(htup->t_infomask & HEAP_MOVED_OFF))
2630 elog(ERROR, "HEAP_MOVED_OFF was expected");
2631 if (HeapTupleHeaderGetXvac(htup) != myXID)
2632 elog(ERROR, "invalid XVAC in tuple header");
2634 if (chain_tuple_moved)
2636 /* some chains were moved while cleaning this page */
2637 Assert(vacpage->offsets_free > 0);
2638 for (i = 0; i < vacpage->offsets_free; i++)
2640 if (vacpage->offsets[i] == off)
2643 if (i >= vacpage->offsets_free) /* not found */
2645 vacpage->offsets[vacpage->offsets_free++] = off;
2646 Assert(keep_tuples > 0);
2649 * If this is not a heap-only tuple, there must be an
2650 * index entry for this item which will be removed in
2651 * the index cleanup. Decrement the
2652 * keep_indexed_tuples count to remember this.
2654 if (!HeapTupleHeaderIsHeapOnly(htup))
2655 keep_indexed_tuples--;
2661 vacpage->offsets[vacpage->offsets_free++] = off;
2662 Assert(keep_tuples > 0);
2663 if (!HeapTupleHeaderIsHeapOnly(htup))
2664 keep_indexed_tuples--;
2670 if (vacpage->offsets_free > 0) /* some tuples were moved */
2672 if (chain_tuple_moved) /* else - they are ordered */
2674 qsort((char *) (vacpage->offsets), vacpage->offsets_free,
2675 sizeof(OffsetNumber), vac_cmp_offno);
2677 vpage_insert(&Nvacpagelist, copy_vac_page(vacpage));
2682 if (offnum <= maxoff)
2683 break; /* had to quit early, see above note */
2685 } /* walk along relation */
2687 blkno++; /* new number of blocks */
2689 if (dst_buffer != InvalidBuffer)
2691 Assert(num_moved > 0);
2692 ReleaseBuffer(dst_buffer);
2698 * We have to commit our tuple movings before we truncate the
2699 * relation. Ideally we should do Commit/StartTransactionCommand
2700 * here, relying on the session-level table lock to protect our
2701 * exclusive access to the relation. However, that would require a
2702 * lot of extra code to close and re-open the relation, indexes, etc.
2703 * For now, a quick hack: record status of current transaction as
2704 * committed, and continue. We force the commit to be synchronous so
2705 * that it's down to disk before we truncate. (Note: tqual.c knows
2706 * that VACUUM FULL always uses sync commit, too.) The transaction
2707 * continues to be shown as running in the ProcArray.
2709 * XXX This desperately needs to be revisited. Any failure after this
2710 * point will result in a PANIC "cannot abort transaction nnn, it was
2711 * already committed"!
2714 (void) RecordTransactionCommit();
2718 * We are not going to move any more tuples across pages, but we still
2719 * need to apply vacuum_page to compact free space in the remaining pages
2720 * in vacuum_pages list. Note that some of these pages may also be in the
2721 * fraged_pages list, and may have had tuples moved onto them; if so, we
2722 * already did vacuum_page and needn't do it again.
2724 for (i = 0, curpage = vacuum_pages->pagedesc;
2728 vacuum_delay_point();
2730 Assert((*curpage)->blkno < blkno);
2731 if ((*curpage)->offsets_used == 0)
2736 /* this page was not used as a move target, so must clean it */
2737 buf = ReadBufferExtended(onerel, MAIN_FORKNUM, (*curpage)->blkno,
2738 RBM_NORMAL, vac_strategy);
2739 LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
2740 page = BufferGetPage(buf);
2741 if (!PageIsEmpty(page))
2742 vacuum_page(onerel, buf, *curpage);
2743 UnlockReleaseBuffer(buf);
2748 * Now scan all the pages that we moved tuples onto and update tuple
2749 * status bits. This is not really necessary, but will save time for
2750 * future transactions examining these tuples.
2752 update_hint_bits(onerel, fraged_pages, num_fraged_pages,
2753 last_move_dest_block, num_moved);
2756 * It'd be cleaner to make this report at the bottom of this routine, but
2757 * then the rusage would double-count the second pass of index vacuuming.
2758 * So do it here and ignore the relatively small amount of processing that
2762 (errmsg("\"%s\": moved %u row versions, truncated %u to %u pages",
2763 RelationGetRelationName(onerel),
2764 num_moved, nblocks, blkno),
2766 pg_rusage_show(&ru0))));
2769 * Reflect the motion of system tuples to catalog cache here.
2771 CommandCounterIncrement();
2773 if (Nvacpagelist.num_pages > 0)
2775 /* vacuum indexes again if needed */
2782 /* re-sort Nvacpagelist.pagedesc */
2783 for (vpleft = Nvacpagelist.pagedesc,
2784 vpright = Nvacpagelist.pagedesc + Nvacpagelist.num_pages - 1;
2785 vpleft < vpright; vpleft++, vpright--)
2793 * keep_tuples is the number of tuples that have been moved off a
2794 * page during chain moves but not been scanned over subsequently.
2795 * The tuple ids of these tuples are not recorded as free offsets
2796 * for any VacPage, so they will not be cleared from the indexes.
2797 * keep_indexed_tuples is the portion of these that are expected
2798 * to have index entries.
2800 Assert(keep_tuples >= 0);
2801 for (i = 0; i < nindexes; i++)
2802 vacuum_index(&Nvacpagelist, Irel[i],
2803 vacrelstats->rel_indexed_tuples,
2804 keep_indexed_tuples);
2808 * Clean moved-off tuples from last page in Nvacpagelist list.
2810 * We need only do this in this one page, because higher-numbered
2811 * pages are going to be truncated from the relation entirely. But see
2812 * comments for update_hint_bits().
2814 if (vacpage->blkno == (blkno - 1) &&
2815 vacpage->offsets_free > 0)
2819 OffsetNumber unused[MaxOffsetNumber];
2820 OffsetNumber offnum,
2825 buf = ReadBufferExtended(onerel, MAIN_FORKNUM, vacpage->blkno,
2826 RBM_NORMAL, vac_strategy);
2827 LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
2828 page = BufferGetPage(buf);
2829 maxoff = PageGetMaxOffsetNumber(page);
2830 for (offnum = FirstOffsetNumber;
2832 offnum = OffsetNumberNext(offnum))
2834 ItemId itemid = PageGetItemId(page, offnum);
2835 HeapTupleHeader htup;
2837 if (!ItemIdIsUsed(itemid))
2839 /* Shouldn't be any DEAD or REDIRECT items anymore */
2840 Assert(ItemIdIsNormal(itemid));
2842 htup = (HeapTupleHeader) PageGetItem(page, itemid);
2843 if (htup->t_infomask & HEAP_XMIN_COMMITTED)
2847 * See comments in the walk-along-page loop above about why
2848 * only MOVED_OFF tuples should be found here.
2850 if (htup->t_infomask & HEAP_MOVED_IN)
2851 elog(ERROR, "HEAP_MOVED_IN was not expected");
2852 if (!(htup->t_infomask & HEAP_MOVED_OFF))
2853 elog(ERROR, "HEAP_MOVED_OFF was expected");
2854 if (HeapTupleHeaderGetXvac(htup) != myXID)
2855 elog(ERROR, "invalid XVAC in tuple header");
2857 ItemIdSetUnused(itemid);
2860 unused[uncnt++] = offnum;
2862 Assert(vacpage->offsets_free == num_tuples);
2864 START_CRIT_SECTION();
2866 PageRepairFragmentation(page);
2868 MarkBufferDirty(buf);
2871 if (!onerel->rd_istemp)
2875 recptr = log_heap_clean(onerel, buf,
2879 PageSetLSN(page, recptr);
2880 PageSetTLI(page, ThisTimeLineID);
2885 UnlockReleaseBuffer(buf);
2888 /* now - free new list of reaped pages */
2889 curpage = Nvacpagelist.pagedesc;
2890 for (i = 0; i < Nvacpagelist.num_pages; i++, curpage++)
2892 pfree(Nvacpagelist.pagedesc);
2895 /* Truncate relation, if needed */
2896 if (blkno < nblocks)
2898 RelationTruncate(onerel, blkno);
2900 /* force relcache inval so all backends reset their rd_targblock */
2901 CacheInvalidateRelcache(onerel);
2903 vacrelstats->rel_pages = blkno; /* set new number of blocks */
2908 if (vacrelstats->vtlinks != NULL)
2909 pfree(vacrelstats->vtlinks);
2911 ExecContext_Finish(&ec);
2915 * move_chain_tuple() -- move one tuple that is part of a tuple chain
2917 * This routine moves old_tup from old_page to dst_page.
2918 * old_page and dst_page might be the same page.
2919 * On entry old_buf and dst_buf are locked exclusively, both locks (or
2920 * the single lock, if this is a intra-page-move) are released before
2923 * Yes, a routine with ten parameters is ugly, but it's still better
2924 * than having these 120 lines of code in repair_frag() which is
2925 * already too long and almost unreadable.
2928 move_chain_tuple(Relation rel,
2929 Buffer old_buf, Page old_page, HeapTuple old_tup,
2930 Buffer dst_buf, Page dst_page, VacPage dst_vacpage,
2931 ExecContext ec, ItemPointer ctid, bool cleanVpd)
2933 TransactionId myXID = GetCurrentTransactionId();
2934 HeapTupleData newtup;
2935 OffsetNumber newoff;
2937 Size tuple_len = old_tup->t_len;
2940 * make a modifiable copy of the source tuple.
2942 heap_copytuple_with_tuple(old_tup, &newtup);
2945 * register invalidation of source tuple in catcaches.
2947 CacheInvalidateHeapTuple(rel, old_tup);
2949 /* NO EREPORT(ERROR) TILL CHANGES ARE LOGGED */
2950 START_CRIT_SECTION();
2953 * mark the source tuple MOVED_OFF.
2955 old_tup->t_data->t_infomask &= ~(HEAP_XMIN_COMMITTED |
2958 old_tup->t_data->t_infomask |= HEAP_MOVED_OFF;
2959 HeapTupleHeaderSetXvac(old_tup->t_data, myXID);
2962 * If this page was not used before - clean it.
2964 * NOTE: a nasty bug used to lurk here. It is possible for the source and
2965 * destination pages to be the same (since this tuple-chain member can be
2966 * on a page lower than the one we're currently processing in the outer
2967 * loop). If that's true, then after vacuum_page() the source tuple will
2968 * have been moved, and tuple.t_data will be pointing at garbage.
2969 * Therefore we must do everything that uses old_tup->t_data BEFORE this
2972 * This path is different from the other callers of vacuum_page, because
2973 * we have already incremented the vacpage's offsets_used field to account
2974 * for the tuple(s) we expect to move onto the page. Therefore
2975 * vacuum_page's check for offsets_used == 0 is wrong. But since that's a
2976 * good debugging check for all other callers, we work around it here
2977 * rather than remove it.
2979 if (!PageIsEmpty(dst_page) && cleanVpd)
2981 int sv_offsets_used = dst_vacpage->offsets_used;
2983 dst_vacpage->offsets_used = 0;
2984 vacuum_page(rel, dst_buf, dst_vacpage);
2985 dst_vacpage->offsets_used = sv_offsets_used;
2989 * Update the state of the copied tuple, and store it on the destination
2990 * page. The copied tuple is never part of a HOT chain.
2992 newtup.t_data->t_infomask &= ~(HEAP_XMIN_COMMITTED |
2995 newtup.t_data->t_infomask |= HEAP_MOVED_IN;
2996 HeapTupleHeaderClearHotUpdated(newtup.t_data);
2997 HeapTupleHeaderClearHeapOnly(newtup.t_data);
2998 HeapTupleHeaderSetXvac(newtup.t_data, myXID);
2999 newoff = PageAddItem(dst_page, (Item) newtup.t_data, tuple_len,
3000 InvalidOffsetNumber, false, true);
3001 if (newoff == InvalidOffsetNumber)
3002 elog(PANIC, "failed to add item with len = %lu to page %u while moving tuple chain",
3003 (unsigned long) tuple_len, dst_vacpage->blkno);
3004 newitemid = PageGetItemId(dst_page, newoff);
3005 /* drop temporary copy, and point to the version on the dest page */
3006 pfree(newtup.t_data);
3007 newtup.t_data = (HeapTupleHeader) PageGetItem(dst_page, newitemid);
3009 ItemPointerSet(&(newtup.t_self), dst_vacpage->blkno, newoff);
3012 * Set new tuple's t_ctid pointing to itself if last tuple in chain, and
3013 * to next tuple in chain otherwise. (Since we move the chain in reverse
3014 * order, this is actually the previously processed tuple.)
3016 if (!ItemPointerIsValid(ctid))
3017 newtup.t_data->t_ctid = newtup.t_self;
3019 newtup.t_data->t_ctid = *ctid;
3020 *ctid = newtup.t_self;
3022 MarkBufferDirty(dst_buf);
3023 if (dst_buf != old_buf)
3024 MarkBufferDirty(old_buf);
3027 if (!rel->rd_istemp)
3029 XLogRecPtr recptr = log_heap_move(rel, old_buf, old_tup->t_self,
3032 if (old_buf != dst_buf)
3034 PageSetLSN(old_page, recptr);
3035 PageSetTLI(old_page, ThisTimeLineID);
3037 PageSetLSN(dst_page, recptr);
3038 PageSetTLI(dst_page, ThisTimeLineID);
3043 PageClearAllVisible(BufferGetPage(old_buf));
3044 if (dst_buf != old_buf)
3045 PageClearAllVisible(BufferGetPage(dst_buf));
3047 LockBuffer(dst_buf, BUFFER_LOCK_UNLOCK);
3048 if (dst_buf != old_buf)
3049 LockBuffer(old_buf, BUFFER_LOCK_UNLOCK);
3051 /* Clear the bits in the visibility map. */
3052 visibilitymap_clear(rel, BufferGetBlockNumber(old_buf));
3053 if (dst_buf != old_buf)
3054 visibilitymap_clear(rel, BufferGetBlockNumber(dst_buf));
3056 /* Create index entries for the moved tuple */
3057 if (ec->resultRelInfo->ri_NumIndices > 0)
3059 ExecStoreTuple(&newtup, ec->slot, InvalidBuffer, false);
3060 ExecInsertIndexTuples(ec->slot, &(newtup.t_self), ec->estate, true);
3061 ResetPerTupleExprContext(ec->estate);
3066 * move_plain_tuple() -- move one tuple that is not part of a chain
3068 * This routine moves old_tup from old_page to dst_page.
3069 * On entry old_buf and dst_buf are locked exclusively, both locks are
3070 * released before exit.
3072 * Yes, a routine with eight parameters is ugly, but it's still better
3073 * than having these 90 lines of code in repair_frag() which is already
3074 * too long and almost unreadable.
3077 move_plain_tuple(Relation rel,
3078 Buffer old_buf, Page old_page, HeapTuple old_tup,
3079 Buffer dst_buf, Page dst_page, VacPage dst_vacpage,
3082 TransactionId myXID = GetCurrentTransactionId();
3083 HeapTupleData newtup;
3084 OffsetNumber newoff;
3086 Size tuple_len = old_tup->t_len;
3089 heap_copytuple_with_tuple(old_tup, &newtup);
3092 * register invalidation of source tuple in catcaches.
3094 * (Note: we do not need to register the copied tuple, because we are not
3095 * changing the tuple contents and so there cannot be any need to flush
3096 * negative catcache entries.)
3098 CacheInvalidateHeapTuple(rel, old_tup);
3100 /* NO EREPORT(ERROR) TILL CHANGES ARE LOGGED */
3101 START_CRIT_SECTION();
3104 * Mark new tuple as MOVED_IN by me; also mark it not HOT.
3106 newtup.t_data->t_infomask &= ~(HEAP_XMIN_COMMITTED |
3109 newtup.t_data->t_infomask |= HEAP_MOVED_IN;
3110 HeapTupleHeaderClearHotUpdated(newtup.t_data);
3111 HeapTupleHeaderClearHeapOnly(newtup.t_data);
3112 HeapTupleHeaderSetXvac(newtup.t_data, myXID);
3114 /* add tuple to the page */
3115 newoff = PageAddItem(dst_page, (Item) newtup.t_data, tuple_len,
3116 InvalidOffsetNumber, false, true);
3117 if (newoff == InvalidOffsetNumber)
3118 elog(PANIC, "failed to add item with len = %lu to page %u (free space %lu, nusd %u, noff %u)",
3119 (unsigned long) tuple_len,
3120 dst_vacpage->blkno, (unsigned long) dst_vacpage->free,
3121 dst_vacpage->offsets_used, dst_vacpage->offsets_free);
3122 newitemid = PageGetItemId(dst_page, newoff);
3123 pfree(newtup.t_data);
3124 newtup.t_data = (HeapTupleHeader) PageGetItem(dst_page, newitemid);
3125 ItemPointerSet(&(newtup.t_data->t_ctid), dst_vacpage->blkno, newoff);
3126 newtup.t_self = newtup.t_data->t_ctid;
3129 * Mark old tuple as MOVED_OFF by me.
3131 old_tup->t_data->t_infomask &= ~(HEAP_XMIN_COMMITTED |
3134 old_tup->t_data->t_infomask |= HEAP_MOVED_OFF;
3135 HeapTupleHeaderSetXvac(old_tup->t_data, myXID);
3137 MarkBufferDirty(dst_buf);
3138 MarkBufferDirty(old_buf);
3141 if (!rel->rd_istemp)
3143 XLogRecPtr recptr = log_heap_move(rel, old_buf, old_tup->t_self,
3146 PageSetLSN(old_page, recptr);
3147 PageSetTLI(old_page, ThisTimeLineID);
3148 PageSetLSN(dst_page, recptr);
3149 PageSetTLI(dst_page, ThisTimeLineID);
3155 * Clear the visible-to-all hint bits on the page, and bits in the
3156 * visibility map. Normally we'd release the locks on the heap pages
3157 * before updating the visibility map, but doesn't really matter here
3158 * because we're holding an AccessExclusiveLock on the relation anyway.
3160 if (PageIsAllVisible(dst_page))
3162 PageClearAllVisible(dst_page);
3163 visibilitymap_clear(rel, BufferGetBlockNumber(dst_buf));
3165 if (PageIsAllVisible(old_page))
3167 PageClearAllVisible(old_page);
3168 visibilitymap_clear(rel, BufferGetBlockNumber(old_buf));
3171 dst_vacpage->free = PageGetFreeSpaceWithFillFactor(rel, dst_page);
3172 LockBuffer(dst_buf, BUFFER_LOCK_UNLOCK);
3173 LockBuffer(old_buf, BUFFER_LOCK_UNLOCK);
3175 dst_vacpage->offsets_used++;
3177 /* insert index' tuples if needed */
3178 if (ec->resultRelInfo->ri_NumIndices > 0)
3180 ExecStoreTuple(&newtup, ec->slot, InvalidBuffer, false);
3181 ExecInsertIndexTuples(ec->slot, &(newtup.t_self), ec->estate, true);
3182 ResetPerTupleExprContext(ec->estate);
3187 * update_hint_bits() -- update hint bits in destination pages
3189 * Scan all the pages that we moved tuples onto and update tuple status bits.
3190 * This is not really necessary, but it will save time for future transactions
3191 * examining these tuples.
3193 * This pass guarantees that all HEAP_MOVED_IN tuples are marked as
3194 * XMIN_COMMITTED, so that future tqual tests won't need to check their XVAC.
3196 * BUT NOTICE that this code fails to clear HEAP_MOVED_OFF tuples from
3197 * pages that were move source pages but not move dest pages. The bulk
3198 * of the move source pages will be physically truncated from the relation,
3199 * and the last page remaining in the rel will be fixed separately in
3200 * repair_frag(), so the only cases where a MOVED_OFF tuple won't get its
3201 * hint bits updated are tuples that are moved as part of a chain and were
3202 * on pages that were not either move destinations nor at the end of the rel.
3203 * To completely ensure that no MOVED_OFF tuples remain unmarked, we'd have
3204 * to remember and revisit those pages too.
3206 * One wonders whether it wouldn't be better to skip this work entirely,
3207 * and let the tuple status updates happen someplace that's not holding an
3208 * exclusive lock on the relation.
3211 update_hint_bits(Relation rel, VacPageList fraged_pages, int num_fraged_pages,
3212 BlockNumber last_move_dest_block, int num_moved)
3214 TransactionId myXID = GetCurrentTransactionId();
3215 int checked_moved = 0;
3219 for (i = 0, curpage = fraged_pages->pagedesc;
3220 i < num_fraged_pages;
3225 OffsetNumber max_offset;
3229 vacuum_delay_point();
3231 if ((*curpage)->blkno > last_move_dest_block)
3232 break; /* no need to scan any further */
3233 if ((*curpage)->offsets_used == 0)
3234 continue; /* this page was never used as a move dest */
3235 buf = ReadBufferExtended(rel, MAIN_FORKNUM, (*curpage)->blkno,
3236 RBM_NORMAL, vac_strategy);
3237 LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
3238 page = BufferGetPage(buf);
3239 max_offset = PageGetMaxOffsetNumber(page);
3240 for (off = FirstOffsetNumber;
3242 off = OffsetNumberNext(off))
3244 ItemId itemid = PageGetItemId(page, off);
3245 HeapTupleHeader htup;
3247 if (!ItemIdIsUsed(itemid))
3249 /* Shouldn't be any DEAD or REDIRECT items anymore */
3250 Assert(ItemIdIsNormal(itemid));
3252 htup = (HeapTupleHeader) PageGetItem(page, itemid);
3253 if (htup->t_infomask & HEAP_XMIN_COMMITTED)
3257 * Here we may see either MOVED_OFF or MOVED_IN tuples.
3259 if (!(htup->t_infomask & HEAP_MOVED))
3260 elog(ERROR, "HEAP_MOVED_OFF/HEAP_MOVED_IN was expected");
3261 if (HeapTupleHeaderGetXvac(htup) != myXID)
3262 elog(ERROR, "invalid XVAC in tuple header");
3264 if (htup->t_infomask & HEAP_MOVED_IN)
3266 htup->t_infomask |= HEAP_XMIN_COMMITTED;
3267 htup->t_infomask &= ~HEAP_MOVED;
3271 htup->t_infomask |= HEAP_XMIN_INVALID;
3273 MarkBufferDirty(buf);
3274 UnlockReleaseBuffer(buf);
3275 Assert((*curpage)->offsets_used == num_tuples);
3276 checked_moved += num_tuples;
3278 Assert(num_moved == checked_moved);
3282 * vacuum_heap() -- free dead tuples
3284 * This routine marks dead tuples as unused and truncates relation
3285 * if there are "empty" end-blocks.
3288 vacuum_heap(VRelStats *vacrelstats, Relation onerel, VacPageList vacuum_pages)
3292 BlockNumber relblocks;
3296 nblocks = vacuum_pages->num_pages;
3297 nblocks -= vacuum_pages->empty_end_pages; /* nothing to do with them */
3299 for (i = 0, vacpage = vacuum_pages->pagedesc; i < nblocks; i++, vacpage++)
3301 vacuum_delay_point();
3303 if ((*vacpage)->offsets_free > 0)
3305 buf = ReadBufferExtended(onerel, MAIN_FORKNUM, (*vacpage)->blkno,
3306 RBM_NORMAL, vac_strategy);
3307 LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
3308 vacuum_page(onerel, buf, *vacpage);
3309 UnlockReleaseBuffer(buf);
3313 /* Truncate relation if there are some empty end-pages */
3314 Assert(vacrelstats->rel_pages >= vacuum_pages->empty_end_pages);
3315 if (vacuum_pages->empty_end_pages > 0)
3317 relblocks = vacrelstats->rel_pages - vacuum_pages->empty_end_pages;
3319 (errmsg("\"%s\": truncated %u to %u pages",
3320 RelationGetRelationName(onerel),
3321 vacrelstats->rel_pages, relblocks)));
3322 RelationTruncate(onerel, relblocks);
3324 /* force relcache inval so all backends reset their rd_targblock */
3325 CacheInvalidateRelcache(onerel);
3327 vacrelstats->rel_pages = relblocks; /* set new number of blocks */
3332 * vacuum_page() -- free dead tuples on a page
3333 * and repair its fragmentation.
3335 * Caller must hold pin and lock on buffer.
3338 vacuum_page(Relation onerel, Buffer buffer, VacPage vacpage)
3340 Page page = BufferGetPage(buffer);
3343 /* There shouldn't be any tuples moved onto the page yet! */
3344 Assert(vacpage->offsets_used == 0);
3346 START_CRIT_SECTION();
3348 for (i = 0; i < vacpage->offsets_free; i++)
3350 ItemId itemid = PageGetItemId(page, vacpage->offsets[i]);
3352 ItemIdSetUnused(itemid);
3355 PageRepairFragmentation(page);
3357 MarkBufferDirty(buffer);
3360 if (!onerel->rd_istemp)
3364 recptr = log_heap_clean(onerel, buffer,
3366 vacpage->offsets, vacpage->offsets_free,
3368 PageSetLSN(page, recptr);
3369 PageSetTLI(page, ThisTimeLineID);
3376 * scan_index() -- scan one index relation to update pg_class statistics.
3378 * We use this when we have no deletions to do.
3381 scan_index(Relation indrel, double num_tuples)
3383 IndexBulkDeleteResult *stats;
3384 IndexVacuumInfo ivinfo;
3387 pg_rusage_init(&ru0);
3389 ivinfo.index = indrel;
3390 ivinfo.vacuum_full = true;
3391 ivinfo.analyze_only = false;
3392 ivinfo.message_level = elevel;
3393 ivinfo.num_heap_tuples = num_tuples;
3394 ivinfo.strategy = vac_strategy;
3396 stats = index_vacuum_cleanup(&ivinfo, NULL);
3401 /* now update statistics in pg_class */
3402 vac_update_relstats(indrel,
3403 stats->num_pages, stats->num_index_tuples,
3404 false, InvalidTransactionId);
3407 (errmsg("index \"%s\" now contains %.0f row versions in %u pages",
3408 RelationGetRelationName(indrel),
3409 stats->num_index_tuples,
3411 errdetail("%u index pages have been deleted, %u are currently reusable.\n"
3413 stats->pages_deleted, stats->pages_free,
3414 pg_rusage_show(&ru0))));
3417 * Check for tuple count mismatch. If the index is partial, then it's OK
3418 * for it to have fewer tuples than the heap; else we got trouble.
3420 if (stats->num_index_tuples != num_tuples)
3422 if (stats->num_index_tuples > num_tuples ||
3423 !vac_is_partial_index(indrel))
3425 (errmsg("index \"%s\" contains %.0f row versions, but table contains %.0f row versions",
3426 RelationGetRelationName(indrel),
3427 stats->num_index_tuples, num_tuples),
3428 errhint("Rebuild the index with REINDEX.")));
3435 * vacuum_index() -- vacuum one index relation.
3437 * Vpl is the VacPageList of the heap we're currently vacuuming.
3438 * It's locked. Indrel is an index relation on the vacuumed heap.
3440 * We don't bother to set locks on the index relation here, since
3441 * the parent table is exclusive-locked already.
3443 * Finally, we arrange to update the index relation's statistics in
3447 vacuum_index(VacPageList vacpagelist, Relation indrel,
3448 double num_tuples, int keep_tuples)
3450 IndexBulkDeleteResult *stats;
3451 IndexVacuumInfo ivinfo;
3454 pg_rusage_init(&ru0);
3456 ivinfo.index = indrel;
3457 ivinfo.vacuum_full = true;
3458 ivinfo.analyze_only = false;
3459 ivinfo.message_level = elevel;
3460 ivinfo.num_heap_tuples = num_tuples + keep_tuples;
3461 ivinfo.strategy = vac_strategy;
3463 /* Do bulk deletion */
3464 stats = index_bulk_delete(&ivinfo, NULL, tid_reaped, (void *) vacpagelist);
3466 /* Do post-VACUUM cleanup */
3467 stats = index_vacuum_cleanup(&ivinfo, stats);
3472 /* now update statistics in pg_class */
3473 vac_update_relstats(indrel,
3474 stats->num_pages, stats->num_index_tuples,
3475 false, InvalidTransactionId);
3478 (errmsg("index \"%s\" now contains %.0f row versions in %u pages",
3479 RelationGetRelationName(indrel),
3480 stats->num_index_tuples,
3482 errdetail("%.0f index row versions were removed.\n"
3483 "%u index pages have been deleted, %u are currently reusable.\n"
3485 stats->tuples_removed,
3486 stats->pages_deleted, stats->pages_free,
3487 pg_rusage_show(&ru0))));
3490 * Check for tuple count mismatch. If the index is partial, then it's OK
3491 * for it to have fewer tuples than the heap; else we got trouble.
3493 if (stats->num_index_tuples != num_tuples + keep_tuples)
3495 if (stats->num_index_tuples > num_tuples + keep_tuples ||
3496 !vac_is_partial_index(indrel))
3498 (errmsg("index \"%s\" contains %.0f row versions, but table contains %.0f row versions",
3499 RelationGetRelationName(indrel),
3500 stats->num_index_tuples, num_tuples + keep_tuples),
3501 errhint("Rebuild the index with REINDEX.")));
3508 * tid_reaped() -- is a particular tid reaped?
3510 * This has the right signature to be an IndexBulkDeleteCallback.
3512 * vacpagelist->VacPage_array is sorted in right order.
3515 tid_reaped(ItemPointer itemptr, void *state)
3517 VacPageList vacpagelist = (VacPageList) state;
3518 OffsetNumber ioffno;
3522 VacPageData vacpage;
3524 vacpage.blkno = ItemPointerGetBlockNumber(itemptr);
3525 ioffno = ItemPointerGetOffsetNumber(itemptr);
3528 vpp = (VacPage *) vac_bsearch((void *) &vp,
3529 (void *) (vacpagelist->pagedesc),
3530 vacpagelist->num_pages,
3537 /* ok - we are on a partially or fully reaped page */
3540 if (vp->offsets_free == 0)
3542 /* this is EmptyPage, so claim all tuples on it are reaped!!! */
3546 voff = (OffsetNumber *) vac_bsearch((void *) &ioffno,
3547 (void *) (vp->offsets),
3549 sizeof(OffsetNumber),
3560 * Update the Free Space Map with the info we now have about free space in
3564 vac_update_fsm(Relation onerel, VacPageList fraged_pages,
3565 BlockNumber rel_pages)
3567 int nPages = fraged_pages->num_pages;
3568 VacPage *pagedesc = fraged_pages->pagedesc;
3571 for (i = 0; i < nPages; i++)
3574 * fraged_pages may contain entries for pages that we later decided to
3575 * truncate from the relation; don't enter them into the free space
3578 if (pagedesc[i]->blkno >= rel_pages)
3581 RecordPageWithFreeSpace(onerel, pagedesc[i]->blkno, pagedesc[i]->free);
3586 /* Copy a VacPage structure */
3588 copy_vac_page(VacPage vacpage)
3592 /* allocate a VacPageData entry */
3593 newvacpage = (VacPage) palloc(sizeof(VacPageData) +
3594 vacpage->offsets_free * sizeof(OffsetNumber));
3597 if (vacpage->offsets_free > 0)
3598 memcpy(newvacpage->offsets, vacpage->offsets,
3599 vacpage->offsets_free * sizeof(OffsetNumber));
3600 newvacpage->blkno = vacpage->blkno;
3601 newvacpage->free = vacpage->free;
3602 newvacpage->offsets_used = vacpage->offsets_used;
3603 newvacpage->offsets_free = vacpage->offsets_free;
3609 * Add a VacPage pointer to a VacPageList.
3611 * As a side effect of the way that scan_heap works,
3612 * higher pages come after lower pages in the array
3613 * (and highest tid on a page is last).
3616 vpage_insert(VacPageList vacpagelist, VacPage vpnew)
3618 #define PG_NPAGEDESC 1024
3620 /* allocate a VacPage entry if needed */
3621 if (vacpagelist->num_pages == 0)
3623 vacpagelist->pagedesc = (VacPage *) palloc(PG_NPAGEDESC * sizeof(VacPage));
3624 vacpagelist->num_allocated_pages = PG_NPAGEDESC;
3626 else if (vacpagelist->num_pages >= vacpagelist->num_allocated_pages)
3628 vacpagelist->num_allocated_pages *= 2;
3629 vacpagelist->pagedesc = (VacPage *) repalloc(vacpagelist->pagedesc, vacpagelist->num_allocated_pages * sizeof(VacPage));
3631 vacpagelist->pagedesc[vacpagelist->num_pages] = vpnew;
3632 (vacpagelist->num_pages)++;
3636 * vac_bsearch: just like standard C library routine bsearch(),
3637 * except that we first test to see whether the target key is outside
3638 * the range of the table entries. This case is handled relatively slowly
3639 * by the normal binary search algorithm (ie, no faster than any other key)
3640 * but it occurs often enough in VACUUM to be worth optimizing.
3643 vac_bsearch(const void *key, const void *base,
3644 size_t nelem, size_t size,
3645 int (*compar) (const void *, const void *))
3652 res = compar(key, base);
3656 return (void *) base;
3659 last = (const void *) ((const char *) base + (nelem - 1) * size);
3660 res = compar(key, last);
3664 return (void *) last;
3667 return NULL; /* already checked 'em all */
3668 return bsearch(key, base, nelem, size, compar);
3672 * Comparator routines for use with qsort() and bsearch().
3675 vac_cmp_blk(const void *left, const void *right)
3680 lblk = (*((VacPage *) left))->blkno;
3681 rblk = (*((VacPage *) right))->blkno;
3691 vac_cmp_offno(const void *left, const void *right)
3693 if (*(OffsetNumber *) left < *(OffsetNumber *) right)
3695 if (*(OffsetNumber *) left == *(OffsetNumber *) right)
3701 vac_cmp_vtlinks(const void *left, const void *right)
3703 if (((VTupleLink) left)->new_tid.ip_blkid.bi_hi <
3704 ((VTupleLink) right)->new_tid.ip_blkid.bi_hi)
3706 if (((VTupleLink) left)->new_tid.ip_blkid.bi_hi >
3707 ((VTupleLink) right)->new_tid.ip_blkid.bi_hi)
3709 /* bi_hi-es are equal */
3710 if (((VTupleLink) left)->new_tid.ip_blkid.bi_lo <
3711 ((VTupleLink) right)->new_tid.ip_blkid.bi_lo)
3713 if (((VTupleLink) left)->new_tid.ip_blkid.bi_lo >
3714 ((VTupleLink) right)->new_tid.ip_blkid.bi_lo)
3716 /* bi_lo-es are equal */
3717 if (((VTupleLink) left)->new_tid.ip_posid <
3718 ((VTupleLink) right)->new_tid.ip_posid)
3720 if (((VTupleLink) left)->new_tid.ip_posid >
3721 ((VTupleLink) right)->new_tid.ip_posid)
3728 * Open all the indexes of the given relation, obtaining the specified kind
3729 * of lock on each. Return an array of Relation pointers for the indexes
3730 * into *Irel, and the number of indexes into *nindexes.
3733 vac_open_indexes(Relation relation, LOCKMODE lockmode,
3734 int *nindexes, Relation **Irel)
3737 ListCell *indexoidscan;
3740 Assert(lockmode != NoLock);
3742 indexoidlist = RelationGetIndexList(relation);
3744 *nindexes = list_length(indexoidlist);
3747 *Irel = (Relation *) palloc(*nindexes * sizeof(Relation));
3752 foreach(indexoidscan, indexoidlist)
3754 Oid indexoid = lfirst_oid(indexoidscan);
3756 (*Irel)[i++] = index_open(indexoid, lockmode);
3759 list_free(indexoidlist);
3763 * Release the resources acquired by vac_open_indexes. Optionally release
3764 * the locks (say NoLock to keep 'em).
3767 vac_close_indexes(int nindexes, Relation *Irel, LOCKMODE lockmode)
3774 Relation ind = Irel[nindexes];
3776 index_close(ind, lockmode);
3783 * Is an index partial (ie, could it contain fewer tuples than the heap?)
3786 vac_is_partial_index(Relation indrel)
3789 * If the index's AM doesn't support nulls, it's partial for our purposes
3791 if (!indrel->rd_am->amindexnulls)
3794 /* Otherwise, look to see if there's a partial-index predicate */
3795 if (!heap_attisnull(indrel->rd_indextuple, Anum_pg_index_indpred))
3803 enough_space(VacPage vacpage, Size len)
3805 len = MAXALIGN(len);
3807 if (len > vacpage->free)
3810 /* if there are free itemid(s) and len <= free_space... */
3811 if (vacpage->offsets_used < vacpage->offsets_free)
3814 /* noff_used >= noff_free and so we'll have to allocate new itemid */
3815 if (len + sizeof(ItemIdData) <= vacpage->free)
3822 PageGetFreeSpaceWithFillFactor(Relation relation, Page page)
3825 * It is correct to use PageGetExactFreeSpace() here, *not*
3826 * PageGetHeapFreeSpace(). This is because (a) we do our own, exact
3827 * accounting for whether line pointers must be added, and (b) we will
3828 * recycle any LP_DEAD line pointers before starting to add rows to a
3829 * page, but that may not have happened yet at the time this function is
3830 * applied to a page, which means PageGetHeapFreeSpace()'s protection
3831 * against too many line pointers on a page could fire incorrectly. We do
3832 * not need that protection here: since VACUUM FULL always recycles all
3833 * dead line pointers first, it'd be physically impossible to insert more
3834 * than MaxHeapTuplesPerPage tuples anyway.
3836 Size freespace = PageGetExactFreeSpace(page);
3839 targetfree = RelationGetTargetPageFreeSpace(relation,
3840 HEAP_DEFAULT_FILLFACTOR);
3841 if (freespace > targetfree)
3842 return freespace - targetfree;
3848 * vacuum_delay_point --- check for interrupts and cost-based delay.
3850 * This should be called in each major loop of VACUUM processing,
3851 * typically once per page processed.
3854 vacuum_delay_point(void)
3856 /* Always check for interrupts */
3857 CHECK_FOR_INTERRUPTS();
3859 /* Nap if appropriate */
3860 if (VacuumCostActive && !InterruptPending &&
3861 VacuumCostBalance >= VacuumCostLimit)
3865 msec = VacuumCostDelay * VacuumCostBalance / VacuumCostLimit;
3866 if (msec > VacuumCostDelay * 4)
3867 msec = VacuumCostDelay * 4;
3869 pg_usleep(msec * 1000L);
3871 VacuumCostBalance = 0;
3873 /* update balance values for workers */
3874 AutoVacuumUpdateDelay();
3876 /* Might have gotten an interrupt while sleeping */
3877 CHECK_FOR_INTERRUPTS();