1 /*-------------------------------------------------------------------------
4 * the postgres vacuum cleaner
6 * Copyright (c) 1994, Regents of the University of California
10 * $Header: /cvsroot/pgsql/src/backend/commands/vacuum.c,v 1.58 1998/01/15 19:42:40 pgsql Exp $
12 *-------------------------------------------------------------------------
14 #include <sys/types.h>
24 #include <utils/portal.h>
25 #include <access/genam.h>
26 #include <access/heapam.h>
27 #include <access/xact.h>
28 #include <storage/bufmgr.h>
29 #include <access/transam.h>
30 #include <catalog/pg_index.h>
31 #include <catalog/index.h>
32 #include <catalog/catname.h>
33 #include <catalog/catalog.h>
34 #include <catalog/pg_class.h>
35 #include <catalog/pg_proc.h>
36 #include <catalog/pg_statistic.h>
37 #include <catalog/pg_type.h>
38 #include <catalog/pg_operator.h>
39 #include <parser/parse_oper.h>
40 #include <storage/smgr.h>
41 #include <storage/lmgr.h>
42 #include <utils/inval.h>
43 #include <utils/mcxt.h>
44 #include <utils/inval.h>
45 #include <utils/syscache.h>
46 #include <utils/builtins.h>
47 #include <commands/vacuum.h>
48 #include <storage/bufpage.h>
49 #include "storage/shmem.h"
50 #ifndef HAVE_GETRUSAGE
51 #include <rusagestub.h>
54 #include <sys/resource.h>
57 /* #include <port-protos.h> */ /* Why? */
59 extern int BlowawayRelationBuffers(Relation rdesc, BlockNumber block);
61 bool VacuumRunning = false;
63 static Portal vc_portal;
65 static int MESSAGE_LEVEL; /* message level */
67 #define swapLong(a,b) {long tmp; tmp=a; a=b; b=tmp;}
68 #define swapInt(a,b) {int tmp; tmp=a; a=b; b=tmp;}
69 #define swapDatum(a,b) {Datum tmp; tmp=a; a=b; b=tmp;}
70 #define VacAttrStatsEqValid(stats) ( stats->f_cmpeq.fn_addr != NULL )
71 #define VacAttrStatsLtGtValid(stats) ( stats->f_cmplt.fn_addr != NULL && \
72 stats->f_cmpgt.fn_addr != NULL && \
73 RegProcedureIsValid(stats->outfunc) )
76 /* non-export function prototypes */
77 static void vc_init(void);
78 static void vc_shutdown(void);
79 static void vc_vacuum(NameData *VacRelP, bool analyze, List *va_cols);
80 static VRelList vc_getrels(NameData *VacRelP);
81 static void vc_vacone(Oid relid, bool analyze, List *va_cols);
82 static void vc_scanheap(VRelStats *vacrelstats, Relation onerel, VPageList Vvpl, VPageList Fvpl);
83 static void vc_rpfheap(VRelStats *vacrelstats, Relation onerel, VPageList Vvpl, VPageList Fvpl, int nindices, Relation *Irel);
84 static void vc_vacheap(VRelStats *vacrelstats, Relation onerel, VPageList vpl);
85 static void vc_vacpage(Page page, VPageDescr vpd);
86 static void vc_vaconeind(VPageList vpl, Relation indrel, int nhtups);
87 static void vc_scanoneind(Relation indrel, int nhtups);
88 static void vc_attrstats(Relation onerel, VRelStats *vacrelstats, HeapTuple htup);
89 static void vc_bucketcpy(AttributeTupleForm attr, Datum value, Datum *bucket, int16 *bucket_len);
90 static void vc_updstats(Oid relid, int npages, int ntups, bool hasindex, VRelStats *vacrelstats);
91 static void vc_delhilowstats(Oid relid, int attcnt, int *attnums);
92 static void vc_setpagelock(Relation rel, BlockNumber blkno);
93 static VPageDescr vc_tidreapped(ItemPointer itemptr, VPageList vpl);
94 static void vc_reappage(VPageList vpl, VPageDescr vpc);
95 static void vc_vpinsert(VPageList vpl, VPageDescr vpnew);
96 static void vc_free(VRelList vrl);
97 static void vc_getindices(Oid relid, int *nindices, Relation **Irel);
98 static void vc_clsindices(int nindices, Relation *Irel);
99 static void vc_mkindesc(Relation onerel, int nindices, Relation *Irel, IndDesc **Idesc);
100 static char *vc_find_eq(char *bot, int nelem, int size, char *elm, int (*compar) (char *, char *));
101 static int vc_cmp_blk(char *left, char *right);
102 static int vc_cmp_offno(char *left, char *right);
103 static bool vc_enough_space(VPageDescr vpd, Size len);
106 vacuum(char *vacrel, bool verbose, bool analyze, List *va_spec)
110 PortalVariableMemory pmem;
116 * Create a portal for safe memory across transctions. We need to
117 * palloc the name space for it because our hash function expects the
118 * name to be on a longword boundary. CreatePortal copies the name to
119 * safe storage for us.
121 pname = (char *) palloc(strlen(VACPNAME) + 1);
122 strcpy(pname, VACPNAME);
123 vc_portal = CreatePortal(pname);
127 MESSAGE_LEVEL = NOTICE;
129 MESSAGE_LEVEL = DEBUG;
131 /* vacrel gets de-allocated on transaction commit */
133 strcpy(VacRel.data, vacrel);
135 pmem = PortalGetVariableMemory(vc_portal);
136 old = MemoryContextSwitchTo((MemoryContext) pmem);
138 Assert(va_spec == NIL || analyze);
141 char *col = (char *) lfirst(le);
144 dest = (char *) palloc(strlen(col) + 1);
146 va_cols = lappend(va_cols, dest);
148 MemoryContextSwitchTo(old);
150 /* initialize vacuum cleaner */
153 /* vacuum the database */
155 vc_vacuum(&VacRel, analyze, va_cols);
157 vc_vacuum(NULL, analyze, NIL);
159 PortalDestroy(&vc_portal);
166 * vc_init(), vc_shutdown() -- start up and shut down the vacuum cleaner.
168 * We run exactly one vacuum cleaner at a time. We use the file system
169 * to guarantee an exclusive lock on vacuuming, since a single vacuum
170 * cleaner instantiation crosses transaction boundaries, and we'd lose
171 * postgres-style locks at the end of every transaction.
173 * The strangeness with committing and starting transactions in the
174 * init and shutdown routines is due to the fact that the vacuum cleaner
175 * is invoked via a sql command, and so is already executing inside
176 * a transaction. We need to leave ourselves in a predictable state
177 * on entry and exit to the vacuum cleaner. We commit the transaction
178 * started in PostgresMain() inside vc_init(), and start one in
179 * vc_shutdown() to match the commit waiting for us back in
187 if ((fd = open("pg_vlock", O_CREAT | O_EXCL, 0600)) < 0)
188 elog(ERROR, "can't create lock file -- another vacuum cleaner running?");
193 * By here, exclusive open on the lock file succeeded. If we abort
194 * for any reason during vacuuming, we need to remove the lock file.
195 * This global variable is checked in the transaction manager on xact
196 * abort, and the routine vc_abort() is called if necessary.
199 VacuumRunning = true;
201 /* matches the StartTransaction in PostgresMain() */
202 CommitTransactionCommand();
208 /* on entry, not in a transaction */
209 if (unlink("pg_vlock") < 0)
210 elog(ERROR, "vacuum: can't destroy lock file!");
212 /* okay, we're done */
213 VacuumRunning = false;
215 /* matches the CommitTransaction in PostgresMain() */
216 StartTransactionCommand();
223 /* on abort, remove the vacuum cleaner lock file */
226 VacuumRunning = false;
230 * vc_vacuum() -- vacuum the database.
232 * This routine builds a list of relations to vacuum, and then calls
233 * code that vacuums them one at a time. We are careful to vacuum each
234 * relation in a separate transaction in order to avoid holding too many
238 vc_vacuum(NameData *VacRelP, bool analyze, List *va_cols)
243 /* get list of relations */
244 vrl = vc_getrels(VacRelP);
246 if (analyze && VacRelP == NULL && vrl != NULL)
247 vc_delhilowstats(InvalidOid, 0, NULL);
249 /* vacuum each heap relation */
250 for (cur = vrl; cur != (VRelList) NULL; cur = cur->vrl_next)
251 vc_vacone(cur->vrl_relid, analyze, va_cols);
257 vc_getrels(NameData *VacRelP)
261 HeapScanDesc pgcscan;
264 PortalVariableMemory portalmem;
275 StartTransactionCommand();
279 ScanKeyEntryInitialize(&pgckey, 0x0, Anum_pg_class_relname,
280 NameEqualRegProcedure,
281 PointerGetDatum(VacRelP->data));
285 ScanKeyEntryInitialize(&pgckey, 0x0, Anum_pg_class_relkind,
286 CharacterEqualRegProcedure, CharGetDatum('r'));
289 portalmem = PortalGetVariableMemory(vc_portal);
290 vrl = cur = (VRelList) NULL;
292 pgclass = heap_openr(RelationRelationName);
293 pgcdesc = RelationGetTupleDescriptor(pgclass);
295 pgcscan = heap_beginscan(pgclass, false, false, 1, &pgckey);
297 while (HeapTupleIsValid(pgctup = heap_getnext(pgcscan, 0, &buf)))
302 d = heap_getattr(pgctup, buf, Anum_pg_class_relname, pgcdesc, &n);
306 * don't vacuum large objects for now - something breaks when we
309 if ((strlen(rname) >= 5) && rname[0] == 'x' &&
310 rname[1] == 'i' && rname[2] == 'n' &&
311 (rname[3] == 'v' || rname[3] == 'x') &&
312 rname[4] >= '0' && rname[4] <= '9')
314 elog(NOTICE, "Rel %s: can't vacuum LargeObjects now",
320 d = heap_getattr(pgctup, buf, Anum_pg_class_relkind, pgcdesc, &n);
322 rkind = DatumGetChar(d);
324 /* skip system relations */
328 elog(NOTICE, "Vacuum: can not process index and certain system tables");
332 /* get a relation list entry for this guy */
333 old = MemoryContextSwitchTo((MemoryContext) portalmem);
334 if (vrl == (VRelList) NULL)
336 vrl = cur = (VRelList) palloc(sizeof(VRelListData));
340 cur->vrl_next = (VRelList) palloc(sizeof(VRelListData));
343 MemoryContextSwitchTo(old);
345 cur->vrl_relid = pgctup->t_oid;
346 cur->vrl_next = (VRelList) NULL;
348 /* wei hates it if you forget to do this */
352 elog(NOTICE, "Vacuum: table not found");
355 heap_endscan(pgcscan);
358 CommitTransactionCommand();
364 * vc_vacone() -- vacuum one heap relation
366 * This routine vacuums a single heap, cleans out its indices, and
367 * updates its statistics npages and ntups statistics.
369 * Doing one heap at a time incurs extra overhead, since we need to
370 * check that the heap exists again just before we vacuum it. The
371 * reason that we do this is so that vacuuming can be spread across
372 * many small transactions. Otherwise, two-phase locking would require
373 * us to lock the entire database during one pass of the vacuum cleaner.
376 vc_vacone(Oid relid, bool analyze, List *va_cols)
383 HeapScanDesc pgcscan;
386 VPageListData Vvpl; /* List of pages to vacuum and/or clean
388 VPageListData Fvpl; /* List of pages with space enough for
394 VRelStats *vacrelstats;
396 StartTransactionCommand();
398 ScanKeyEntryInitialize(&pgckey, 0x0, ObjectIdAttributeNumber,
399 ObjectIdEqualRegProcedure,
400 ObjectIdGetDatum(relid));
402 pgclass = heap_openr(RelationRelationName);
403 pgcdesc = RelationGetTupleDescriptor(pgclass);
404 pgcscan = heap_beginscan(pgclass, false, false, 1, &pgckey);
407 * Race condition -- if the pg_class tuple has gone away since the
408 * last time we saw it, we don't need to vacuum it.
411 if (!HeapTupleIsValid(pgctup = heap_getnext(pgcscan, 0, &pgcbuf)))
413 heap_endscan(pgcscan);
415 CommitTransactionCommand();
419 /* now open the class and vacuum it */
420 onerel = heap_open(relid);
422 vacrelstats = (VRelStats *) palloc(sizeof(VRelStats));
423 vacrelstats->relid = relid;
424 vacrelstats->npages = vacrelstats->ntups = 0;
425 vacrelstats->hasindex = false;
426 if (analyze && !IsSystemRelationName((RelationGetRelationName(onerel))->data))
430 AttributeTupleForm *attr;
432 attr_cnt = onerel->rd_att->natts;
433 attr = onerel->rd_att->attrs;
440 if (length(va_cols) > attr_cnt)
441 elog(ERROR, "vacuum: too many attributes specified for relation %s",
442 (RelationGetRelationName(onerel))->data);
443 attnums = (int *) palloc(attr_cnt * sizeof(int));
446 char *col = (char *) lfirst(le);
448 for (i = 0; i < attr_cnt; i++)
450 if (namestrcmp(&(attr[i]->attname), col) == 0)
453 if (i < attr_cnt) /* found */
457 elog(ERROR, "vacuum: there is no attribute %s in %s",
458 col, (RelationGetRelationName(onerel))->data);
464 vacrelstats->vacattrstats =
465 (VacAttrStats *) palloc(attr_cnt * sizeof(VacAttrStats));
467 for (i = 0; i < attr_cnt; i++)
469 Operator func_operator;
470 OperatorTupleForm pgopform;
473 stats = &vacrelstats->vacattrstats[i];
474 stats->attr = palloc(ATTRIBUTE_TUPLE_SIZE);
475 memmove(stats->attr, attr[((attnums) ? attnums[i] : i)], ATTRIBUTE_TUPLE_SIZE);
476 stats->best = stats->guess1 = stats->guess2 = 0;
477 stats->max = stats->min = 0;
478 stats->best_len = stats->guess1_len = stats->guess2_len = 0;
479 stats->max_len = stats->min_len = 0;
480 stats->initialized = false;
481 stats->best_cnt = stats->guess1_cnt = stats->guess1_hits = stats->guess2_hits = 0;
482 stats->max_cnt = stats->min_cnt = stats->null_cnt = stats->nonnull_cnt = 0;
484 func_operator = oper("=", stats->attr->atttypid, stats->attr->atttypid, true);
485 if (func_operator != NULL)
487 pgopform = (OperatorTupleForm) GETSTRUCT(func_operator);
488 fmgr_info(pgopform->oprcode, &(stats->f_cmpeq));
491 stats->f_cmpeq.fn_addr = NULL;
493 func_operator = oper("<", stats->attr->atttypid, stats->attr->atttypid, true);
494 if (func_operator != NULL)
496 pgopform = (OperatorTupleForm) GETSTRUCT(func_operator);
497 fmgr_info(pgopform->oprcode, &(stats->f_cmplt));
500 stats->f_cmplt.fn_addr = NULL;
502 func_operator = oper(">", stats->attr->atttypid, stats->attr->atttypid, true);
503 if (func_operator != NULL)
505 pgopform = (OperatorTupleForm) GETSTRUCT(func_operator);
506 fmgr_info(pgopform->oprcode, &(stats->f_cmpgt));
509 stats->f_cmpgt.fn_addr = NULL;
511 pgttup = SearchSysCacheTuple(TYPOID,
512 ObjectIdGetDatum(stats->attr->atttypid),
514 if (HeapTupleIsValid(pgttup))
515 stats->outfunc = ((TypeTupleForm) GETSTRUCT(pgttup))->typoutput;
517 stats->outfunc = InvalidOid;
519 vacrelstats->va_natts = attr_cnt;
520 vc_delhilowstats(relid, ((attnums) ? attr_cnt : 0), attnums);
526 vacrelstats->va_natts = 0;
527 vacrelstats->vacattrstats = (VacAttrStats *) NULL;
530 /* we require the relation to be locked until the indices are cleaned */
531 RelationSetLockForWrite(onerel);
534 Vvpl.vpl_npages = Fvpl.vpl_npages = 0;
535 vc_scanheap(vacrelstats, onerel, &Vvpl, &Fvpl);
537 /* Now open indices */
538 Irel = (Relation *) NULL;
539 vc_getindices(vacrelstats->relid, &nindices, &Irel);
542 vacrelstats->hasindex = true;
544 vacrelstats->hasindex = false;
546 /* Clean/scan index relation(s) */
547 if (Irel != (Relation *) NULL)
549 if (Vvpl.vpl_npages > 0)
551 for (i = 0; i < nindices; i++)
552 vc_vaconeind(&Vvpl, Irel[i], vacrelstats->ntups);
555 /* just scan indices to update statistic */
557 for (i = 0; i < nindices; i++)
558 vc_scanoneind(Irel[i], vacrelstats->ntups);
562 if (Fvpl.vpl_npages > 0) /* Try to shrink heap */
563 vc_rpfheap(vacrelstats, onerel, &Vvpl, &Fvpl, nindices, Irel);
566 if (Irel != (Relation *) NULL)
567 vc_clsindices(nindices, Irel);
568 if (Vvpl.vpl_npages > 0)/* Clean pages from Vvpl list */
569 vc_vacheap(vacrelstats, onerel, &Vvpl);
572 /* ok - free Vvpl list of reapped pages */
573 if (Vvpl.vpl_npages > 0)
575 vpp = Vvpl.vpl_pgdesc;
576 for (i = 0; i < Vvpl.vpl_npages; i++, vpp++)
578 pfree(Vvpl.vpl_pgdesc);
579 if (Fvpl.vpl_npages > 0)
580 pfree(Fvpl.vpl_pgdesc);
583 /* all done with this class */
585 heap_endscan(pgcscan);
588 /* update statistics in pg_class */
589 vc_updstats(vacrelstats->relid, vacrelstats->npages, vacrelstats->ntups,
590 vacrelstats->hasindex, vacrelstats);
592 /* next command frees attribute stats */
594 CommitTransactionCommand();
598 * vc_scanheap() -- scan an open heap relation
600 * This routine sets commit times, constructs Vvpl list of
601 * empty/uninitialized pages and pages with dead tuples and
602 * ~LP_USED line pointers, constructs Fvpl list of pages
603 * appropriate for purposes of shrinking and maintains statistics
604 * on the number of live tuples in a heap.
607 vc_scanheap(VRelStats *vacrelstats, Relation onerel,
608 VPageList Vvpl, VPageList Fvpl)
637 Size min_tlen = MAXTUPLEN;
639 int32 i /* , attr_cnt */ ;
642 bool do_shrinking = true;
644 getrusage(RUSAGE_SELF, &ru0);
646 nvac = ntups = nunused = ncrash = nempg = nnepg = nchpg = nemend = 0;
649 relname = (RelationGetRelationName(onerel))->data;
651 nblocks = RelationGetNumberOfBlocks(onerel);
653 vpc = (VPageDescr) palloc(sizeof(VPageDescrData) + MaxOffsetNumber * sizeof(OffsetNumber));
656 for (blkno = 0; blkno < nblocks; blkno++)
658 buf = ReadBuffer(onerel, blkno);
659 page = BufferGetPage(buf);
660 vpc->vpd_blkno = blkno;
665 elog(NOTICE, "Rel %s: Uninitialized page %u - fixing",
667 PageInit(page, BufferGetPageSize(buf), 0);
668 vpc->vpd_free = ((PageHeader) page)->pd_upper - ((PageHeader) page)->pd_lower;
669 frsize += (vpc->vpd_free - sizeof(ItemIdData));
672 vc_reappage(Vvpl, vpc);
677 if (PageIsEmpty(page))
679 vpc->vpd_free = ((PageHeader) page)->pd_upper - ((PageHeader) page)->pd_lower;
680 frsize += (vpc->vpd_free - sizeof(ItemIdData));
683 vc_reappage(Vvpl, vpc);
690 maxoff = PageGetMaxOffsetNumber(page);
691 for (offnum = FirstOffsetNumber;
693 offnum = OffsetNumberNext(offnum))
695 itemid = PageGetItemId(page, offnum);
698 * Collect un-used items too - it's possible to have indices
699 * pointing here after crash.
701 if (!ItemIdIsUsed(itemid))
703 vpc->vpd_voff[vpc->vpd_noff++] = offnum;
708 htup = (HeapTuple) PageGetItem(page, itemid);
711 if (!(htup->t_infomask & HEAP_XMIN_COMMITTED))
713 if (htup->t_infomask & HEAP_XMIN_INVALID)
717 if (TransactionIdDidAbort(htup->t_xmin))
719 else if (TransactionIdDidCommit(htup->t_xmin))
721 htup->t_infomask |= HEAP_XMIN_COMMITTED;
724 else if (!TransactionIdIsInProgress(htup->t_xmin))
727 * Not Aborted, Not Committed, Not in Progress -
728 * so it's from crashed process. - vadim 11/26/96
735 elog(NOTICE, "Rel %s: TID %u/%u: InsertTransactionInProgress %u - can't shrink relation",
736 relname, blkno, offnum, htup->t_xmin);
737 do_shrinking = false;
743 * here we are concerned about tuples with xmin committed
744 * and xmax unknown or committed
746 if (htup->t_infomask & HEAP_XMIN_COMMITTED &&
747 !(htup->t_infomask & HEAP_XMAX_INVALID))
749 if (htup->t_infomask & HEAP_XMAX_COMMITTED)
751 else if (TransactionIdDidAbort(htup->t_xmax))
753 htup->t_infomask |= HEAP_XMAX_INVALID;
756 else if (TransactionIdDidCommit(htup->t_xmax))
758 else if (!TransactionIdIsInProgress(htup->t_xmax))
761 * Not Aborted, Not Committed, Not in Progress - so it
762 * from crashed process. - vadim 06/02/97
764 htup->t_infomask |= HEAP_XMAX_INVALID;;
769 elog(NOTICE, "Rel %s: TID %u/%u: DeleteTransactionInProgress %u - can't shrink relation",
770 relname, blkno, offnum, htup->t_xmax);
771 do_shrinking = false;
776 * It's possibly! But from where it comes ? And should we fix
777 * it ? - vadim 11/28/96
779 itemptr = &(htup->t_ctid);
780 if (!ItemPointerIsValid(itemptr) ||
781 BlockIdGetBlockNumber(&(itemptr->ip_blkid)) != blkno)
783 elog(NOTICE, "Rel %s: TID %u/%u: TID IN TUPLEHEADER %u/%u IS NOT THE SAME. TUPGONE %d.",
784 relname, blkno, offnum,
785 BlockIdGetBlockNumber(&(itemptr->ip_blkid)),
786 itemptr->ip_posid, tupgone);
792 if (htup->t_len != itemid->lp_len)
794 elog(NOTICE, "Rel %s: TID %u/%u: TUPLE_LEN IN PAGEHEADER %u IS NOT THE SAME AS IN TUPLEHEADER %u. TUPGONE %d.",
795 relname, blkno, offnum,
796 itemid->lp_len, htup->t_len, tupgone);
798 if (!OidIsValid(htup->t_oid))
800 elog(NOTICE, "Rel %s: TID %u/%u: OID IS INVALID. TUPGONE %d.",
801 relname, blkno, offnum, tupgone);
808 if (tempPage == (Page) NULL)
812 pageSize = PageGetPageSize(page);
813 tempPage = (Page) palloc(pageSize);
814 memmove(tempPage, page, pageSize);
817 lpp = &(((PageHeader) tempPage)->pd_linp[offnum - 1]);
820 lpp->lp_flags &= ~LP_USED;
822 vpc->vpd_voff[vpc->vpd_noff++] = offnum;
830 if (htup->t_len < min_tlen)
831 min_tlen = htup->t_len;
832 if (htup->t_len > max_tlen)
833 max_tlen = htup->t_len;
834 vc_attrstats(onerel, vacrelstats, htup);
846 if (tempPage != (Page) NULL)
847 { /* Some tuples are gone */
848 PageRepairFragmentation(tempPage);
849 vpc->vpd_free = ((PageHeader) tempPage)->pd_upper - ((PageHeader) tempPage)->pd_lower;
850 frsize += vpc->vpd_free;
851 vc_reappage(Vvpl, vpc);
853 tempPage = (Page) NULL;
855 else if (vpc->vpd_noff > 0)
856 { /* there are only ~LP_USED line pointers */
857 vpc->vpd_free = ((PageHeader) page)->pd_upper - ((PageHeader) page)->pd_lower;
858 frsize += vpc->vpd_free;
859 vc_reappage(Vvpl, vpc);
871 /* save stats in the rel list for use later */
872 vacrelstats->ntups = ntups;
873 vacrelstats->npages = nblocks;
874 /* vacrelstats->natts = attr_cnt;*/
876 min_tlen = max_tlen = 0;
877 vacrelstats->min_tlen = min_tlen;
878 vacrelstats->max_tlen = max_tlen;
880 Vvpl->vpl_nemend = nemend;
881 Fvpl->vpl_nemend = nemend;
884 * Try to make Fvpl keeping in mind that we can't use free space of
885 * "empty" end-pages and last page if it reapped.
887 if (do_shrinking && Vvpl->vpl_npages - nemend > 0)
889 int nusf; /* blocks usefull for re-using */
891 nusf = Vvpl->vpl_npages - nemend;
892 if ((Vvpl->vpl_pgdesc[nusf - 1])->vpd_blkno == nblocks - nemend - 1)
895 for (i = 0; i < nusf; i++)
897 vp = Vvpl->vpl_pgdesc[i];
898 if (vc_enough_space(vp, min_tlen))
900 vc_vpinsert(Fvpl, vp);
901 frsusf += vp->vpd_free;
906 getrusage(RUSAGE_SELF, &ru1);
908 elog(MESSAGE_LEVEL, "Rel %s: Pages %u: Changed %u, Reapped %u, Empty %u, New %u; \
909 Tup %u: Vac %u, Crash %u, UnUsed %u, MinLen %u, MaxLen %u; Re-using: Free/Avail. Space %u/%u; EndEmpty/Avail. Pages %u/%u. Elapsed %u/%u sec.",
911 nblocks, nchpg, Vvpl->vpl_npages, nempg, nnepg,
912 ntups, nvac, ncrash, nunused, min_tlen, max_tlen,
913 frsize, frsusf, nemend, Fvpl->vpl_npages,
914 ru1.ru_stime.tv_sec - ru0.ru_stime.tv_sec,
915 ru1.ru_utime.tv_sec - ru0.ru_utime.tv_sec);
921 * vc_rpfheap() -- try to repaire relation' fragmentation
923 * This routine marks dead tuples as unused and tries re-use dead space
924 * by moving tuples (and inserting indices if needed). It constructs
925 * Nvpl list of free-ed pages (moved tuples) and clean indices
926 * for them after committing (in hack-manner - without losing locks
927 * and freeing memory!) current transaction. It truncates relation
928 * if some end-blocks are gone away.
931 vc_rpfheap(VRelStats *vacrelstats, Relation onerel,
932 VPageList Vvpl, VPageList Fvpl, int nindices, Relation *Irel)
942 OffsetNumber offnum = 0,
950 TupleDesc tupdesc = NULL;
951 Datum *idatum = NULL;
953 InsertIndexResult iresult;
955 VPageDescr ToVpd = NULL,
977 getrusage(RUSAGE_SELF, &ru0);
979 myXID = GetCurrentTransactionId();
980 myCID = GetCurrentCommandId();
982 if (Irel != (Relation *) NULL) /* preparation for index' inserts */
984 vc_mkindesc(onerel, nindices, Irel, &Idesc);
985 tupdesc = RelationGetTupleDescriptor(onerel);
986 idatum = (Datum *) palloc(INDEX_MAX_KEYS * sizeof(*idatum));
987 inulls = (char *) palloc(INDEX_MAX_KEYS * sizeof(*inulls));
991 Fnpages = Fvpl->vpl_npages;
992 Fvplast = Fvpl->vpl_pgdesc[Fnpages - 1];
993 Fblklast = Fvplast->vpd_blkno;
994 Assert(Vvpl->vpl_npages > Vvpl->vpl_nemend);
995 Vnpages = Vvpl->vpl_npages - Vvpl->vpl_nemend;
996 Vvplast = Vvpl->vpl_pgdesc[Vnpages - 1];
997 Vblklast = Vvplast->vpd_blkno;
998 Assert(Vblklast >= Fblklast);
999 ToBuf = InvalidBuffer;
1002 vpc = (VPageDescr) palloc(sizeof(VPageDescrData) + MaxOffsetNumber * sizeof(OffsetNumber));
1003 vpc->vpd_nusd = vpc->vpd_noff = 0;
1005 nblocks = vacrelstats->npages;
1006 for (blkno = nblocks - Vvpl->vpl_nemend - 1;; blkno--)
1008 /* if it's reapped page and it was used by me - quit */
1009 if (blkno == Fblklast && Fvplast->vpd_nusd > 0)
1012 buf = ReadBuffer(onerel, blkno);
1013 page = BufferGetPage(buf);
1017 isempty = PageIsEmpty(page);
1020 if (blkno == Vblklast) /* it's reapped page */
1022 if (Vvplast->vpd_noff > 0) /* there are dead tuples */
1023 { /* on this page - clean */
1025 vc_vacpage(page, Vvplast);
1033 Assert(Vnpages > 0);
1034 /* get prev reapped page from Vvpl */
1035 Vvplast = Vvpl->vpl_pgdesc[Vnpages - 1];
1036 Vblklast = Vvplast->vpd_blkno;
1037 if (blkno == Fblklast) /* this page in Fvpl too */
1040 Assert(Fnpages > 0);
1041 Assert(Fvplast->vpd_nusd == 0);
1042 /* get prev reapped page from Fvpl */
1043 Fvplast = Fvpl->vpl_pgdesc[Fnpages - 1];
1044 Fblklast = Fvplast->vpd_blkno;
1046 Assert(Fblklast <= Vblklast);
1058 vpc->vpd_blkno = blkno;
1059 maxoff = PageGetMaxOffsetNumber(page);
1060 for (offnum = FirstOffsetNumber;
1062 offnum = OffsetNumberNext(offnum))
1064 itemid = PageGetItemId(page, offnum);
1066 if (!ItemIdIsUsed(itemid))
1069 htup = (HeapTuple) PageGetItem(page, itemid);
1072 /* try to find new page for this tuple */
1073 if (ToBuf == InvalidBuffer ||
1074 !vc_enough_space(ToVpd, tlen))
1076 if (ToBuf != InvalidBuffer)
1079 ToBuf = InvalidBuffer;
1082 * If no one tuple can't be added to this page -
1083 * remove page from Fvpl. - vadim 11/27/96
1085 if (!vc_enough_space(ToVpd, vacrelstats->min_tlen))
1087 if (ToVpd != Fvplast)
1089 Assert(Fnpages > ToVpI + 1);
1090 memmove(Fvpl->vpl_pgdesc + ToVpI,
1091 Fvpl->vpl_pgdesc + ToVpI + 1,
1092 sizeof(VPageDescr *) * (Fnpages - ToVpI - 1));
1094 Assert(Fnpages >= 1);
1098 /* get prev reapped page from Fvpl */
1099 Fvplast = Fvpl->vpl_pgdesc[Fnpages - 1];
1100 Fblklast = Fvplast->vpd_blkno;
1103 for (i = 0; i < Fnpages; i++)
1105 if (vc_enough_space(Fvpl->vpl_pgdesc[i], tlen))
1109 break; /* can't move item anywhere */
1111 ToVpd = Fvpl->vpl_pgdesc[ToVpI];
1112 ToBuf = ReadBuffer(onerel, ToVpd->vpd_blkno);
1113 ToPage = BufferGetPage(ToBuf);
1114 /* if this page was not used before - clean it */
1115 if (!PageIsEmpty(ToPage) && ToVpd->vpd_nusd == 0)
1116 vc_vacpage(ToPage, ToVpd);
1120 newtup = (HeapTuple) palloc(tlen);
1121 memmove((char *) newtup, (char *) htup, tlen);
1123 /* store transaction information */
1124 TransactionIdStore(myXID, &(newtup->t_xmin));
1125 newtup->t_cmin = myCID;
1126 StoreInvalidTransactionId(&(newtup->t_xmax));
1127 /* set xmin to unknown and xmax to invalid */
1128 newtup->t_infomask &= ~(HEAP_XACT_MASK);
1129 newtup->t_infomask |= HEAP_XMAX_INVALID;
1131 /* add tuple to the page */
1132 newoff = PageAddItem(ToPage, (Item) newtup, tlen,
1133 InvalidOffsetNumber, LP_USED);
1134 if (newoff == InvalidOffsetNumber)
1137 failed to add item with len = %u to page %u (free space %u, nusd %u, noff %u)",
1138 tlen, ToVpd->vpd_blkno, ToVpd->vpd_free,
1139 ToVpd->vpd_nusd, ToVpd->vpd_noff);
1141 newitemid = PageGetItemId(ToPage, newoff);
1143 newtup = (HeapTuple) PageGetItem(ToPage, newitemid);
1144 ItemPointerSet(&(newtup->t_ctid), ToVpd->vpd_blkno, newoff);
1146 /* now logically delete end-tuple */
1147 TransactionIdStore(myXID, &(htup->t_xmax));
1148 htup->t_cmax = myCID;
1149 /* set xmax to unknown */
1150 htup->t_infomask &= ~(HEAP_XMAX_INVALID | HEAP_XMAX_COMMITTED);
1154 ToVpd->vpd_free = ((PageHeader) ToPage)->pd_upper - ((PageHeader) ToPage)->pd_lower;
1155 vpc->vpd_voff[vpc->vpd_noff++] = offnum;
1157 /* insert index' tuples if needed */
1158 if (Irel != (Relation *) NULL)
1160 for (i = 0, idcur = Idesc; i < nindices; i++, idcur++)
1164 (AttrNumber *) &(idcur->tform->indkey[0]),
1171 iresult = index_insert(
1182 } /* walk along page */
1184 if (vpc->vpd_noff > 0) /* some tuples were moved */
1186 vc_reappage(&Nvpl, vpc);
1194 if (offnum <= maxoff)
1195 break; /* some item(s) left */
1197 } /* walk along relation */
1199 blkno++; /* new number of blocks */
1201 if (ToBuf != InvalidBuffer)
1211 * We have to commit our tuple' movings before we'll truncate
1212 * relation, but we shouldn't lose our locks. And so - quick hack:
1213 * flush buffers and record status of current transaction as
1214 * committed, and continue. - vadim 11/13/96
1216 FlushBufferPool(!TransactionFlushEnabled());
1217 TransactionIdCommit(myXID);
1218 FlushBufferPool(!TransactionFlushEnabled());
1222 * Clean uncleaned reapped pages from Vvpl list and set xmin committed
1223 * for inserted tuples
1226 for (i = 0, vpp = Vvpl->vpl_pgdesc; i < Vnpages; i++, vpp++)
1228 Assert((*vpp)->vpd_blkno < blkno);
1229 buf = ReadBuffer(onerel, (*vpp)->vpd_blkno);
1230 page = BufferGetPage(buf);
1231 if ((*vpp)->vpd_nusd == 0) /* this page was not used */
1235 * noff == 0 in empty pages only - such pages should be
1238 Assert((*vpp)->vpd_noff > 0);
1239 vc_vacpage(page, *vpp);
1242 /* this page was used */
1245 moff = PageGetMaxOffsetNumber(page);
1246 for (newoff = FirstOffsetNumber;
1248 newoff = OffsetNumberNext(newoff))
1250 itemid = PageGetItemId(page, newoff);
1251 if (!ItemIdIsUsed(itemid))
1253 htup = (HeapTuple) PageGetItem(page, itemid);
1254 if (TransactionIdEquals((TransactionId) htup->t_xmin, myXID))
1256 htup->t_infomask |= HEAP_XMIN_COMMITTED;
1260 Assert((*vpp)->vpd_nusd == ntups);
1265 Assert(nmoved == nchkmvd);
1267 getrusage(RUSAGE_SELF, &ru1);
1269 elog(MESSAGE_LEVEL, "Rel %s: Pages: %u --> %u; Tuple(s) moved: %u. \
1270 Elapsed %u/%u sec.",
1271 (RelationGetRelationName(onerel))->data,
1272 nblocks, blkno, nmoved,
1273 ru1.ru_stime.tv_sec - ru0.ru_stime.tv_sec,
1274 ru1.ru_utime.tv_sec - ru0.ru_utime.tv_sec);
1276 if (Nvpl.vpl_npages > 0)
1278 /* vacuum indices again if needed */
1279 if (Irel != (Relation *) NULL)
1285 /* re-sort Nvpl.vpl_pgdesc */
1286 for (vpleft = Nvpl.vpl_pgdesc,
1287 vpright = Nvpl.vpl_pgdesc + Nvpl.vpl_npages - 1;
1288 vpleft < vpright; vpleft++, vpright--)
1294 for (i = 0; i < nindices; i++)
1295 vc_vaconeind(&Nvpl, Irel[i], vacrelstats->ntups);
1299 * clean moved tuples from last page in Nvpl list if some tuples
1302 if (vpc->vpd_noff > 0 && offnum <= maxoff)
1304 Assert(vpc->vpd_blkno == blkno - 1);
1305 buf = ReadBuffer(onerel, vpc->vpd_blkno);
1306 page = BufferGetPage(buf);
1309 for (offnum = FirstOffsetNumber;
1311 offnum = OffsetNumberNext(offnum))
1313 itemid = PageGetItemId(page, offnum);
1314 if (!ItemIdIsUsed(itemid))
1316 htup = (HeapTuple) PageGetItem(page, itemid);
1317 Assert(TransactionIdEquals((TransactionId) htup->t_xmax, myXID));
1318 itemid->lp_flags &= ~LP_USED;
1321 Assert(vpc->vpd_noff == ntups);
1322 PageRepairFragmentation(page);
1326 /* now - free new list of reapped pages */
1327 vpp = Nvpl.vpl_pgdesc;
1328 for (i = 0; i < Nvpl.vpl_npages; i++, vpp++)
1330 pfree(Nvpl.vpl_pgdesc);
1333 /* truncate relation */
1334 if (blkno < nblocks)
1336 i = BlowawayRelationBuffers(onerel, blkno);
1338 elog (FATAL, "VACUUM (vc_rpfheap): BlowawayRelationBuffers returned %d", i);
1339 blkno = smgrtruncate(DEFAULT_SMGR, onerel, blkno);
1341 vacrelstats->npages = blkno; /* set new number of blocks */
1344 if (Irel != (Relation *) NULL) /* pfree index' allocations */
1349 vc_clsindices(nindices, Irel);
1357 * vc_vacheap() -- free dead tuples
1359 * This routine marks dead tuples as unused and truncates relation
1360 * if there are "empty" end-blocks.
1363 vc_vacheap(VRelStats *vacrelstats, Relation onerel, VPageList Vvpl)
1371 nblocks = Vvpl->vpl_npages;
1372 nblocks -= Vvpl->vpl_nemend; /* nothing to do with them */
1374 for (i = 0, vpp = Vvpl->vpl_pgdesc; i < nblocks; i++, vpp++)
1376 if ((*vpp)->vpd_noff > 0)
1378 buf = ReadBuffer(onerel, (*vpp)->vpd_blkno);
1379 page = BufferGetPage(buf);
1380 vc_vacpage(page, *vpp);
1385 /* truncate relation if there are some empty end-pages */
1386 if (Vvpl->vpl_nemend > 0)
1388 Assert(vacrelstats->npages >= Vvpl->vpl_nemend);
1389 nblocks = vacrelstats->npages - Vvpl->vpl_nemend;
1390 elog(MESSAGE_LEVEL, "Rel %s: Pages: %u --> %u.",
1391 (RelationGetRelationName(onerel))->data,
1392 vacrelstats->npages, nblocks);
1395 * we have to flush "empty" end-pages (if changed, but who knows
1396 * it) before truncation
1398 FlushBufferPool(!TransactionFlushEnabled());
1400 i = BlowawayRelationBuffers(onerel, nblocks);
1402 elog (FATAL, "VACUUM (vc_vacheap): BlowawayRelationBuffers returned %d", i);
1404 nblocks = smgrtruncate(DEFAULT_SMGR, onerel, nblocks);
1405 Assert(nblocks >= 0);
1406 vacrelstats->npages = nblocks; /* set new number of blocks */
1412 * vc_vacpage() -- free dead tuples on a page
1413 * and repaire its fragmentation.
1416 vc_vacpage(Page page, VPageDescr vpd)
1421 Assert(vpd->vpd_nusd == 0);
1422 for (i = 0; i < vpd->vpd_noff; i++)
1424 itemid = &(((PageHeader) page)->pd_linp[vpd->vpd_voff[i] - 1]);
1425 itemid->lp_flags &= ~LP_USED;
1427 PageRepairFragmentation(page);
1432 * _vc_scanoneind() -- scan one index relation to update statistic.
1436 vc_scanoneind(Relation indrel, int nhtups)
1438 RetrieveIndexResult res;
1439 IndexScanDesc iscan;
1445 getrusage(RUSAGE_SELF, &ru0);
1447 /* walk through the entire index */
1448 iscan = index_beginscan(indrel, false, 0, (ScanKey) NULL);
1451 while ((res = index_getnext(iscan, ForwardScanDirection))
1452 != (RetrieveIndexResult) NULL)
1458 index_endscan(iscan);
1460 /* now update statistics in pg_class */
1461 nipages = RelationGetNumberOfBlocks(indrel);
1462 vc_updstats(indrel->rd_id, nipages, nitups, false, NULL);
1464 getrusage(RUSAGE_SELF, &ru1);
1466 elog(MESSAGE_LEVEL, "Ind %s: Pages %u; Tuples %u. Elapsed %u/%u sec.",
1467 indrel->rd_rel->relname.data, nipages, nitups,
1468 ru1.ru_stime.tv_sec - ru0.ru_stime.tv_sec,
1469 ru1.ru_utime.tv_sec - ru0.ru_utime.tv_sec);
1471 if (nitups != nhtups)
1472 elog(NOTICE, "Ind %s: NUMBER OF INDEX' TUPLES (%u) IS NOT THE SAME AS HEAP' (%u)",
1473 indrel->rd_rel->relname.data, nitups, nhtups);
1475 } /* vc_scanoneind */
1478 * vc_vaconeind() -- vacuum one index relation.
1480 * Vpl is the VPageList of the heap we're currently vacuuming.
1481 * It's locked. Indrel is an index relation on the vacuumed heap.
1482 * We don't set locks on the index relation here, since the indexed
1483 * access methods support locking at different granularities.
1484 * We let them handle it.
1486 * Finally, we arrange to update the index relation's statistics in
1490 vc_vaconeind(VPageList vpl, Relation indrel, int nhtups)
1492 RetrieveIndexResult res;
1493 IndexScanDesc iscan;
1494 ItemPointer heapptr;
1502 getrusage(RUSAGE_SELF, &ru0);
1504 /* walk through the entire index */
1505 iscan = index_beginscan(indrel, false, 0, (ScanKey) NULL);
1509 while ((res = index_getnext(iscan, ForwardScanDirection))
1510 != (RetrieveIndexResult) NULL)
1512 heapptr = &res->heap_iptr;
1514 if ((vp = vc_tidreapped(heapptr, vpl)) != (VPageDescr) NULL)
1517 elog(DEBUG, "<%x,%x> -> <%x,%x>",
1518 ItemPointerGetBlockNumber(&(res->index_iptr)),
1519 ItemPointerGetOffsetNumber(&(res->index_iptr)),
1520 ItemPointerGetBlockNumber(&(res->heap_iptr)),
1521 ItemPointerGetOffsetNumber(&(res->heap_iptr)));
1523 if (vp->vpd_noff == 0)
1524 { /* this is EmptyPage !!! */
1525 elog(NOTICE, "Ind %s: pointer to EmptyPage (blk %u off %u) - fixing",
1526 indrel->rd_rel->relname.data,
1527 vp->vpd_blkno, ItemPointerGetOffsetNumber(heapptr));
1530 index_delete(indrel, &res->index_iptr);
1541 index_endscan(iscan);
1543 /* now update statistics in pg_class */
1544 nipages = RelationGetNumberOfBlocks(indrel);
1545 vc_updstats(indrel->rd_id, nipages, nitups, false, NULL);
1547 getrusage(RUSAGE_SELF, &ru1);
1549 elog(MESSAGE_LEVEL, "Ind %s: Pages %u; Tuples %u: Deleted %u. Elapsed %u/%u sec.",
1550 indrel->rd_rel->relname.data, nipages, nitups, nvac,
1551 ru1.ru_stime.tv_sec - ru0.ru_stime.tv_sec,
1552 ru1.ru_utime.tv_sec - ru0.ru_utime.tv_sec);
1554 if (nitups != nhtups)
1555 elog(NOTICE, "Ind %s: NUMBER OF INDEX' TUPLES (%u) IS NOT THE SAME AS HEAP' (%u)",
1556 indrel->rd_rel->relname.data, nitups, nhtups);
1558 } /* vc_vaconeind */
1561 * vc_tidreapped() -- is a particular tid reapped?
1563 * vpl->VPageDescr_array is sorted in right order.
1566 vc_tidreapped(ItemPointer itemptr, VPageList vpl)
1568 OffsetNumber ioffno;
1574 vpd.vpd_blkno = ItemPointerGetBlockNumber(itemptr);
1575 ioffno = ItemPointerGetOffsetNumber(itemptr);
1578 vpp = (VPageDescr *) vc_find_eq((char *) (vpl->vpl_pgdesc),
1579 vpl->vpl_npages, sizeof(VPageDescr), (char *) &vp,
1582 if (vpp == (VPageDescr *) NULL)
1583 return ((VPageDescr) NULL);
1586 /* ok - we are on true page */
1588 if (vp->vpd_noff == 0)
1589 { /* this is EmptyPage !!! */
1593 voff = (OffsetNumber *) vc_find_eq((char *) (vp->vpd_voff),
1594 vp->vpd_noff, sizeof(OffsetNumber), (char *) &ioffno,
1597 if (voff == (OffsetNumber *) NULL)
1598 return ((VPageDescr) NULL);
1602 } /* vc_tidreapped */
1605 * vc_attrstats() -- compute column statistics used by the optimzer
1607 * We compute the column min, max, null and non-null counts.
1608 * Plus we attempt to find the count of the value that occurs most
1609 * frequently in each column
1610 * These figures are used to compute the selectivity of the column
1612 * We use a three-bucked cache to get the most frequent item
1613 * The 'guess' buckets count hits. A cache miss causes guess1
1614 * to get the most hit 'guess' item in the most recent cycle, and
1615 * the new item goes into guess2. Whenever the total count of hits
1616 * of a 'guess' entry is larger than 'best', 'guess' becomes 'best'.
1618 * This method works perfectly for columns with unique values, and columns
1619 * with only two unique values, plus nulls.
1621 * It becomes less perfect as the number of unique values increases and
1622 * their distribution in the table becomes more random.
1626 vc_attrstats(Relation onerel, VRelStats *vacrelstats, HeapTuple htup)
1629 attr_cnt = vacrelstats->va_natts;
1630 VacAttrStats *vacattrstats = vacrelstats->vacattrstats;
1631 TupleDesc tupDesc = onerel->rd_att;
1635 for (i = 0; i < attr_cnt; i++)
1637 VacAttrStats *stats = &vacattrstats[i];
1638 bool value_hit = true;
1640 value = heap_getattr(htup, InvalidBuffer,
1641 stats->attr->attnum, tupDesc, &isnull);
1643 if (!VacAttrStatsEqValid(stats))
1650 stats->nonnull_cnt++;
1651 if (stats->initialized == false)
1653 vc_bucketcpy(stats->attr, value, &stats->best, &stats->best_len);
1654 /* best_cnt gets incremented later */
1655 vc_bucketcpy(stats->attr, value, &stats->guess1, &stats->guess1_len);
1656 stats->guess1_cnt = stats->guess1_hits = 1;
1657 vc_bucketcpy(stats->attr, value, &stats->guess2, &stats->guess2_len);
1658 stats->guess2_hits = 1;
1659 if (VacAttrStatsLtGtValid(stats))
1661 vc_bucketcpy(stats->attr, value, &stats->max, &stats->max_len);
1662 vc_bucketcpy(stats->attr, value, &stats->min, &stats->min_len);
1664 stats->initialized = true;
1666 if (VacAttrStatsLtGtValid(stats))
1668 if ((*fmgr_faddr(&stats->f_cmplt)) (value, stats->min))
1670 vc_bucketcpy(stats->attr, value, &stats->min, &stats->min_len);
1673 if ((*fmgr_faddr(&stats->f_cmpgt)) (value, stats->max))
1675 vc_bucketcpy(stats->attr, value, &stats->max, &stats->max_len);
1678 if ((*fmgr_faddr(&stats->f_cmpeq)) (value, stats->min))
1680 else if ((*fmgr_faddr(&stats->f_cmpeq)) (value, stats->max))
1683 if ((*fmgr_faddr(&stats->f_cmpeq)) (value, stats->best))
1685 else if ((*fmgr_faddr(&stats->f_cmpeq)) (value, stats->guess1))
1687 stats->guess1_cnt++;
1688 stats->guess1_hits++;
1690 else if ((*fmgr_faddr(&stats->f_cmpeq)) (value, stats->guess2))
1691 stats->guess2_hits++;
1695 if (stats->guess2_hits > stats->guess1_hits)
1697 swapDatum(stats->guess1, stats->guess2);
1698 swapInt(stats->guess1_len, stats->guess2_len);
1699 stats->guess1_cnt = stats->guess2_hits;
1700 swapLong(stats->guess1_hits, stats->guess2_hits);
1702 if (stats->guess1_cnt > stats->best_cnt)
1704 swapDatum(stats->best, stats->guess1);
1705 swapInt(stats->best_len, stats->guess1_len);
1706 swapLong(stats->best_cnt, stats->guess1_cnt);
1707 stats->guess1_hits = 1;
1708 stats->guess2_hits = 1;
1712 vc_bucketcpy(stats->attr, value, &stats->guess2, &stats->guess2_len);
1713 stats->guess1_hits = 1;
1714 stats->guess2_hits = 1;
1722 * vc_bucketcpy() -- update pg_class statistics for one relation
1726 vc_bucketcpy(AttributeTupleForm attr, Datum value, Datum *bucket, int16 *bucket_len)
1728 if (attr->attbyval && attr->attlen != -1)
1732 int len = (attr->attlen != -1 ? attr->attlen : VARSIZE(value));
1734 if (len > *bucket_len)
1736 if (*bucket_len != 0)
1737 pfree(DatumGetPointer(*bucket));
1738 *bucket = PointerGetDatum(palloc(len));
1741 memmove(DatumGetPointer(*bucket), DatumGetPointer(value), len);
1746 * vc_updstats() -- update pg_class statistics for one relation
1748 * This routine works for both index and heap relation entries in
1749 * pg_class. We violate no-overwrite semantics here by storing new
1750 * values for ntups, npages, and hasindex directly in the pg_class
1751 * tuple that's already on the page. The reason for this is that if
1752 * we updated these tuples in the usual way, then every tuple in pg_class
1753 * would be replaced every day. This would make planning and executing
1754 * historical queries very expensive.
1757 vc_updstats(Oid relid, int npages, int ntups, bool hasindex, VRelStats *vacrelstats)
1762 HeapScanDesc rsdesc,
1770 Form_pg_class pgcform;
1773 AttributeTupleForm attp;
1776 * update number of tuples and number of pages in pg_class
1778 ScanKeyEntryInitialize(&rskey, 0x0, ObjectIdAttributeNumber,
1779 ObjectIdEqualRegProcedure,
1780 ObjectIdGetDatum(relid));
1782 rd = heap_openr(RelationRelationName);
1783 rsdesc = heap_beginscan(rd, false, false, 1, &rskey);
1785 if (!HeapTupleIsValid(rtup = heap_getnext(rsdesc, 0, &rbuf)))
1786 elog(ERROR, "pg_class entry for relid %d vanished during vacuuming",
1789 /* overwrite the existing statistics in the tuple */
1790 vc_setpagelock(rd, BufferGetBlockNumber(rbuf));
1791 pgcform = (Form_pg_class) GETSTRUCT(rtup);
1792 pgcform->reltuples = ntups;
1793 pgcform->relpages = npages;
1794 pgcform->relhasindex = hasindex;
1796 if (vacrelstats != NULL && vacrelstats->va_natts > 0)
1798 VacAttrStats *vacattrstats = vacrelstats->vacattrstats;
1799 int natts = vacrelstats->va_natts;
1801 ad = heap_openr(AttributeRelationName);
1802 sd = heap_openr(StatisticRelationName);
1803 ScanKeyEntryInitialize(&askey, 0, Anum_pg_attribute_attrelid,
1806 asdesc = heap_beginscan(ad, false, false, 1, &askey);
1808 while (HeapTupleIsValid(atup = heap_getnext(asdesc, 0, &abuf)))
1811 float32data selratio; /* average ratio of rows selected
1812 * for a random constant */
1813 VacAttrStats *stats;
1814 Datum values[Natts_pg_statistic];
1815 char nulls[Natts_pg_statistic];
1817 attp = (AttributeTupleForm) GETSTRUCT(atup);
1818 if (attp->attnum <= 0) /* skip system attributes for now, */
1819 /* they are unique anyway */
1822 for (i = 0; i < natts; i++)
1824 if (attp->attnum == vacattrstats[i].attr->attnum)
1829 stats = &(vacattrstats[i]);
1831 /* overwrite the existing statistics in the tuple */
1832 if (VacAttrStatsEqValid(stats))
1835 vc_setpagelock(ad, BufferGetBlockNumber(abuf));
1837 if (stats->nonnull_cnt + stats->null_cnt == 0 ||
1838 (stats->null_cnt <= 1 && stats->best_cnt == 1))
1840 else if (VacAttrStatsLtGtValid(stats) && stats->min_cnt + stats->max_cnt == stats->nonnull_cnt)
1842 double min_cnt_d = stats->min_cnt,
1843 max_cnt_d = stats->max_cnt,
1844 null_cnt_d = stats->null_cnt,
1845 nonnullcnt_d = stats->nonnull_cnt; /* prevent overflow */
1847 selratio = (min_cnt_d * min_cnt_d + max_cnt_d * max_cnt_d + null_cnt_d * null_cnt_d) /
1848 (nonnullcnt_d + null_cnt_d) / (nonnullcnt_d + null_cnt_d);
1852 double most = (double) (stats->best_cnt > stats->null_cnt ? stats->best_cnt : stats->null_cnt);
1853 double total = ((double) stats->nonnull_cnt) + ((double) stats->null_cnt);
1856 * we assume count of other values are 20% of best
1859 selratio = (most * most + 0.20 * most * (total - most)) / total / total;
1863 attp->attdisbursion = selratio;
1864 WriteNoReleaseBuffer(abuf);
1866 /* DO PG_STATISTIC INSERTS */
1869 * doing system relations, especially pg_statistic is a
1872 if (VacAttrStatsLtGtValid(stats) && stats->initialized /* &&
1873 * !IsSystemRelationName(
1875 pgcform->relname.data) */ )
1877 FmgrInfo out_function;
1880 for (i = 0; i < Natts_pg_statistic; ++i)
1884 * initialize values[]
1888 values[i++] = (Datum) relid; /* 1 */
1889 values[i++] = (Datum) attp->attnum; /* 2 */
1890 values[i++] = (Datum) InvalidOid; /* 3 */
1891 fmgr_info(stats->outfunc, &out_function);
1892 out_string = (*fmgr_faddr(&out_function)) (stats->min, stats->attr->atttypid);
1893 values[i++] = (Datum) fmgr(TextInRegProcedure, out_string);
1895 out_string = (char *) (*fmgr_faddr(&out_function)) (stats->max, stats->attr->atttypid);
1896 values[i++] = (Datum) fmgr(TextInRegProcedure, out_string);
1901 stup = heap_formtuple(sdesc, values, nulls);
1904 * insert the tuple in the relation and get the tuple's oid.
1907 heap_insert(sd, stup);
1908 pfree(DatumGetPointer(values[3]));
1909 pfree(DatumGetPointer(values[4]));
1914 heap_endscan(asdesc);
1919 /* XXX -- after write, should invalidate relcache in other backends */
1920 WriteNoReleaseBuffer(rbuf); /* heap_endscan release scan' buffers ? */
1923 * invalidating system relations confuses the function cache of
1924 * pg_operator and pg_opclass
1926 if (!IsSystemRelationName(pgcform->relname.data))
1927 RelationInvalidateHeapTuple(rd, rtup);
1929 /* that's all, folks */
1930 heap_endscan(rsdesc);
1935 * vc_delhilowstats() -- delete pg_statistics rows
1939 vc_delhilowstats(Oid relid, int attcnt, int *attnums)
1941 Relation pgstatistic;
1942 HeapScanDesc pgsscan;
1946 pgstatistic = heap_openr(StatisticRelationName);
1948 if (relid != InvalidOid)
1950 ScanKeyEntryInitialize(&pgskey, 0x0, Anum_pg_statistic_starelid,
1951 ObjectIdEqualRegProcedure,
1952 ObjectIdGetDatum(relid));
1953 pgsscan = heap_beginscan(pgstatistic, false, false, 1, &pgskey);
1956 pgsscan = heap_beginscan(pgstatistic, false, false, 0, NULL);
1958 while (HeapTupleIsValid(pgstup = heap_getnext(pgsscan, 0, NULL)))
1962 Form_pg_statistic pgs = (Form_pg_statistic) GETSTRUCT(pgstup);
1965 for (i = 0; i < attcnt; i++)
1967 if (pgs->staattnum == attnums[i] + 1)
1971 continue; /* don't delete it */
1973 heap_delete(pgstatistic, &pgstup->t_ctid);
1976 heap_endscan(pgsscan);
1977 heap_close(pgstatistic);
1981 vc_setpagelock(Relation rel, BlockNumber blkno)
1983 ItemPointerData itm;
1985 ItemPointerSet(&itm, blkno, 1);
1987 RelationSetLockForWritePage(rel, &itm);
1991 * vc_reappage() -- save a page on the array of reapped pages.
1993 * As a side effect of the way that the vacuuming loop for a given
1994 * relation works, higher pages come after lower pages in the array
1995 * (and highest tid on a page is last).
1998 vc_reappage(VPageList vpl, VPageDescr vpc)
2002 /* allocate a VPageDescrData entry */
2003 newvpd = (VPageDescr) palloc(sizeof(VPageDescrData) + vpc->vpd_noff * sizeof(OffsetNumber));
2006 if (vpc->vpd_noff > 0)
2007 memmove(newvpd->vpd_voff, vpc->vpd_voff, vpc->vpd_noff * sizeof(OffsetNumber));
2008 newvpd->vpd_blkno = vpc->vpd_blkno;
2009 newvpd->vpd_free = vpc->vpd_free;
2010 newvpd->vpd_nusd = vpc->vpd_nusd;
2011 newvpd->vpd_noff = vpc->vpd_noff;
2013 /* insert this page into vpl list */
2014 vc_vpinsert(vpl, newvpd);
2019 vc_vpinsert(VPageList vpl, VPageDescr vpnew)
2022 /* allocate a VPageDescr entry if needed */
2023 if (vpl->vpl_npages == 0)
2024 vpl->vpl_pgdesc = (VPageDescr *) palloc(100 * sizeof(VPageDescr));
2025 else if (vpl->vpl_npages % 100 == 0)
2026 vpl->vpl_pgdesc = (VPageDescr *) repalloc(vpl->vpl_pgdesc, (vpl->vpl_npages + 100) * sizeof(VPageDescr));
2027 vpl->vpl_pgdesc[vpl->vpl_npages] = vpnew;
2028 (vpl->vpl_npages)++;
2033 vc_free(VRelList vrl)
2037 PortalVariableMemory pmem;
2039 pmem = PortalGetVariableMemory(vc_portal);
2040 old = MemoryContextSwitchTo((MemoryContext) pmem);
2042 while (vrl != (VRelList) NULL)
2045 /* free rel list entry */
2047 vrl = vrl->vrl_next;
2051 MemoryContextSwitchTo(old);
2055 vc_find_eq(char *bot, int nelem, int size, char *elm, int (*compar) (char *, char *))
2058 int last = nelem - 1;
2059 int celm = nelem / 2;
2063 last_move = first_move = true;
2066 if (first_move == true)
2068 res = compar(bot, elm);
2075 if (last_move == true)
2077 res = compar(elm, bot + last * size);
2081 return (bot + last * size);
2084 res = compar(elm, bot + celm * size);
2086 return (bot + celm * size);
2100 last = last - celm - 1;
2101 bot = bot + (celm + 1) * size;
2102 celm = (last + 1) / 2;
2109 vc_cmp_blk(char *left, char *right)
2114 lblk = (*((VPageDescr *) left))->vpd_blkno;
2115 rblk = (*((VPageDescr *) right))->vpd_blkno;
2126 vc_cmp_offno(char *left, char *right)
2129 if (*(OffsetNumber *) left < *(OffsetNumber *) right)
2131 if (*(OffsetNumber *) left == *(OffsetNumber *) right)
2135 } /* vc_cmp_offno */
2139 vc_getindices(Oid relid, int *nindices, Relation **Irel)
2145 HeapScanDesc pgiscan;
2155 ioid = (Oid *) palloc(10 * sizeof(Oid));
2157 /* prepare a heap scan on the pg_index relation */
2158 pgindex = heap_openr(IndexRelationName);
2159 pgidesc = RelationGetTupleDescriptor(pgindex);
2161 ScanKeyEntryInitialize(&pgikey, 0x0, Anum_pg_index_indrelid,
2162 ObjectIdEqualRegProcedure,
2163 ObjectIdGetDatum(relid));
2165 pgiscan = heap_beginscan(pgindex, false, false, 1, &pgikey);
2167 while (HeapTupleIsValid(pgitup = heap_getnext(pgiscan, 0, NULL)))
2169 d = heap_getattr(pgitup, InvalidBuffer, Anum_pg_index_indexrelid,
2173 ioid = (Oid *) repalloc(ioid, (i + 10) * sizeof(Oid));
2174 ioid[i - 1] = DatumGetObjectId(d);
2177 heap_endscan(pgiscan);
2178 heap_close(pgindex);
2181 { /* No one index found */
2186 if (Irel != (Relation **) NULL)
2187 *Irel = (Relation *) palloc(i * sizeof(Relation));
2191 irel = index_open(ioid[--i]);
2192 if (irel != (Relation) NULL)
2194 if (Irel != (Relation **) NULL)
2201 elog(NOTICE, "CAN't OPEN INDEX %u - SKIP IT", ioid[i]);
2206 if (Irel != (Relation **) NULL && *nindices == 0)
2209 *Irel = (Relation *) NULL;
2212 } /* vc_getindices */
2216 vc_clsindices(int nindices, Relation *Irel)
2219 if (Irel == (Relation *) NULL)
2224 index_close(Irel[nindices]);
2228 } /* vc_clsindices */
2232 vc_mkindesc(Relation onerel, int nindices, Relation *Irel, IndDesc **Idesc)
2235 HeapTuple pgIndexTup;
2236 AttrNumber *attnumP;
2240 *Idesc = (IndDesc *) palloc(nindices * sizeof(IndDesc));
2242 for (i = 0, idcur = *Idesc; i < nindices; i++, idcur++)
2245 SearchSysCacheTuple(INDEXRELID,
2246 ObjectIdGetDatum(Irel[i]->rd_id),
2249 idcur->tform = (IndexTupleForm) GETSTRUCT(pgIndexTup);
2250 for (attnumP = &(idcur->tform->indkey[0]), natts = 0;
2251 *attnumP != InvalidAttrNumber && natts != INDEX_MAX_KEYS;
2252 attnumP++, natts++);
2253 if (idcur->tform->indproc != InvalidOid)
2255 idcur->finfoP = &(idcur->finfo);
2256 FIgetnArgs(idcur->finfoP) = natts;
2258 FIgetProcOid(idcur->finfoP) = idcur->tform->indproc;
2259 *(FIgetname(idcur->finfoP)) = '\0';
2262 idcur->finfoP = (FuncIndexInfo *) NULL;
2264 idcur->natts = natts;
2271 vc_enough_space(VPageDescr vpd, Size len)
2274 len = DOUBLEALIGN(len);
2276 if (len > vpd->vpd_free)
2279 if (vpd->vpd_nusd < vpd->vpd_noff) /* there are free itemid(s) */
2280 return (true); /* and len <= free_space */
2282 /* ok. noff_usd >= noff_free and so we'll have to allocate new itemid */
2283 if (len <= vpd->vpd_free - sizeof(ItemIdData))
2288 } /* vc_enough_space */