granicus.if.org Git - postgresql/blob - src/backend/access/index/indexam.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * indexam.c
   4  *        general index access method routines
   5  *
   6  * Portions Copyright (c) 1996-2011, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        src/backend/access/index/indexam.c
  12  *
  13  * INTERFACE ROUTINES
  14  *              index_open              - open an index relation by relation OID
  15  *              index_close             - close an index relation
  16  *              index_beginscan - start a scan of an index with amgettuple
  17  *              index_beginscan_bitmap - start a scan of an index with amgetbitmap
  18  *              index_rescan    - restart a scan of an index
  19  *              index_endscan   - end a scan
  20  *              index_insert    - insert an index tuple into a relation
  21  *              index_markpos   - mark a scan position
  22  *              index_restrpos  - restore a scan position
  23  *              index_getnext   - get the next tuple from a scan
  24  *              index_getbitmap - get all tuples from a scan
  25  *              index_bulk_delete       - bulk deletion of index tuples
  26  *              index_vacuum_cleanup    - post-deletion cleanup of an index
  27  *              index_getprocid - get a support procedure OID
  28  *              index_getprocinfo - get a support procedure's lookup info
  29  *
  30  * NOTES
  31  *              This file contains the index_ routines which used
  32  *              to be a scattered collection of stuff in access/genam.
  33  *
  34  *
  35  * old comments
  36  *              Scans are implemented as follows:
  37  *
  38  *              `0' represents an invalid item pointer.
  39  *              `-' represents an unknown item pointer.
  40  *              `X' represents a known item pointers.
  41  *              `+' represents known or invalid item pointers.
  42  *              `*' represents any item pointers.
  43  *
  44  *              State is represented by a triple of these symbols in the order of
  45  *              previous, current, next.  Note that the case of reverse scans works
  46  *              identically.
  47  *
  48  *                              State   Result
  49  *              (1)             + + -   + 0 0                   (if the next item pointer is invalid)
  50  *              (2)                             + X -                   (otherwise)
  51  *              (3)             * 0 0   * 0 0                   (no change)
  52  *              (4)             + X 0   X 0 0                   (shift)
  53  *              (5)             * + X   + X -                   (shift, add unknown)
  54  *
  55  *              All other states cannot occur.
  56  *
  57  *              Note: It would be possible to cache the status of the previous and
  58  *                        next item pointer using the flags.
  59  *
  60  *-------------------------------------------------------------------------
  61  */
  62
  63 #include "postgres.h"
  64
  65 #include "access/relscan.h"
  66 #include "access/transam.h"
  67 #include "access/xact.h"
  68 #include "catalog/index.h"
  69 #include "pgstat.h"
  70 #include "storage/bufmgr.h"
  71 #include "storage/lmgr.h"
  72 #include "storage/predicate.h"
  73 #include "utils/relcache.h"
  74 #include "utils/snapmgr.h"
  75 #include "utils/tqual.h"
  76
  77
  78 /* ----------------------------------------------------------------
  79  *                                      macros used in index_ routines
  80  *
  81  * Note: the ReindexIsProcessingIndex() check in RELATION_CHECKS is there
  82  * to check that we don't try to scan or do retail insertions into an index
  83  * that is currently being rebuilt or pending rebuild.  This helps to catch
  84  * things that don't work when reindexing system catalogs.  The assertion
  85  * doesn't prevent the actual rebuild because we don't use RELATION_CHECKS
  86  * when calling the index AM's ambuild routine, and there is no reason for
  87  * ambuild to call its subsidiary routines through this file.
  88  * ----------------------------------------------------------------
  89  */
  90 #define RELATION_CHECKS \
  91 ( \
  92         AssertMacro(RelationIsValid(indexRelation)), \
  93         AssertMacro(PointerIsValid(indexRelation->rd_am)), \
  94         AssertMacro(!ReindexIsProcessingIndex(RelationGetRelid(indexRelation))) \
  95 )
  96
  97 #define SCAN_CHECKS \
  98 ( \
  99         AssertMacro(IndexScanIsValid(scan)), \
 100         AssertMacro(RelationIsValid(scan->indexRelation)), \
 101         AssertMacro(PointerIsValid(scan->indexRelation->rd_am)) \
 102 )
 103
 104 #define GET_REL_PROCEDURE(pname) \
 105 do { \
 106         procedure = &indexRelation->rd_aminfo->pname; \
 107         if (!OidIsValid(procedure->fn_oid)) \
 108         { \
 109                 RegProcedure    procOid = indexRelation->rd_am->pname; \
 110                 if (!RegProcedureIsValid(procOid)) \
 111                         elog(ERROR, "invalid %s regproc", CppAsString(pname)); \
 112                 fmgr_info_cxt(procOid, procedure, indexRelation->rd_indexcxt); \
 113         } \
 114 } while(0)
 115
 116 #define GET_SCAN_PROCEDURE(pname) \
 117 do { \
 118         procedure = &scan->indexRelation->rd_aminfo->pname; \
 119         if (!OidIsValid(procedure->fn_oid)) \
 120         { \
 121                 RegProcedure    procOid = scan->indexRelation->rd_am->pname; \
 122                 if (!RegProcedureIsValid(procOid)) \
 123                         elog(ERROR, "invalid %s regproc", CppAsString(pname)); \
 124                 fmgr_info_cxt(procOid, procedure, scan->indexRelation->rd_indexcxt); \
 125         } \
 126 } while(0)
 127
 128 static IndexScanDesc index_beginscan_internal(Relation indexRelation,
 129                                                  int nkeys, int norderbys);
 130
 131
 132 /* ----------------------------------------------------------------
 133  *                                 index_ interface functions
 134  * ----------------------------------------------------------------
 135  */
 136
 137 /* ----------------
 138  *              index_open - open an index relation by relation OID
 139  *
 140  *              If lockmode is not "NoLock", the specified kind of lock is
 141  *              obtained on the index.  (Generally, NoLock should only be
 142  *              used if the caller knows it has some appropriate lock on the
 143  *              index already.)
 144  *
 145  *              An error is raised if the index does not exist.
 146  *
 147  *              This is a convenience routine adapted for indexscan use.
 148  *              Some callers may prefer to use relation_open directly.
 149  * ----------------
 150  */
 151 Relation
 152 index_open(Oid relationId, LOCKMODE lockmode)
 153 {
 154         Relation        r;
 155
 156         r = relation_open(relationId, lockmode);
 157
 158         if (r->rd_rel->relkind != RELKIND_INDEX)
 159                 ereport(ERROR,
 160                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 161                                  errmsg("\"%s\" is not an index",
 162                                                 RelationGetRelationName(r))));
 163
 164         return r;
 165 }
 166
 167 /* ----------------
 168  *              index_close - close an index relation
 169  *
 170  *              If lockmode is not "NoLock", we then release the specified lock.
 171  *
 172  *              Note that it is often sensible to hold a lock beyond index_close;
 173  *              in that case, the lock is released automatically at xact end.
 174  * ----------------
 175  */
 176 void
 177 index_close(Relation relation, LOCKMODE lockmode)
 178 {
 179         LockRelId       relid = relation->rd_lockInfo.lockRelId;
 180
 181         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
 182
 183         /* The relcache does the real work... */
 184         RelationClose(relation);
 185
 186         if (lockmode != NoLock)
 187                 UnlockRelationId(&relid, lockmode);
 188 }
 189
 190 /* ----------------
 191  *              index_insert - insert an index tuple into a relation
 192  * ----------------
 193  */
 194 bool
 195 index_insert(Relation indexRelation,
 196                          Datum *values,
 197                          bool *isnull,
 198                          ItemPointer heap_t_ctid,
 199                          Relation heapRelation,
 200                          IndexUniqueCheck checkUnique)
 201 {
 202         FmgrInfo   *procedure;
 203
 204         RELATION_CHECKS;
 205         GET_REL_PROCEDURE(aminsert);
 206
 207         if (!(indexRelation->rd_am->ampredlocks))
 208                 CheckForSerializableConflictIn(indexRelation,
 209                                                                            (HeapTuple) NULL,
 210                                                                            InvalidBuffer);
 211
 212         /*
 213          * have the am's insert proc do all the work.
 214          */
 215         return DatumGetBool(FunctionCall6(procedure,
 216                                                                           PointerGetDatum(indexRelation),
 217                                                                           PointerGetDatum(values),
 218                                                                           PointerGetDatum(isnull),
 219                                                                           PointerGetDatum(heap_t_ctid),
 220                                                                           PointerGetDatum(heapRelation),
 221                                                                           Int32GetDatum((int32) checkUnique)));
 222 }
 223
 224 /*
 225  * index_beginscan - start a scan of an index with amgettuple
 226  *
 227  * Caller must be holding suitable locks on the heap and the index.
 228  */
 229 IndexScanDesc
 230 index_beginscan(Relation heapRelation,
 231                                 Relation indexRelation,
 232                                 Snapshot snapshot,
 233                                 int nkeys, int norderbys)
 234 {
 235         IndexScanDesc scan;
 236
 237         scan = index_beginscan_internal(indexRelation, nkeys, norderbys);
 238
 239         /*
 240          * Save additional parameters into the scandesc.  Everything else was set
 241          * up by RelationGetIndexScan.
 242          */
 243         scan->heapRelation = heapRelation;
 244         scan->xs_snapshot = snapshot;
 245
 246         return scan;
 247 }
 248
 249 /*
 250  * index_beginscan_bitmap - start a scan of an index with amgetbitmap
 251  *
 252  * As above, caller had better be holding some lock on the parent heap
 253  * relation, even though it's not explicitly mentioned here.
 254  */
 255 IndexScanDesc
 256 index_beginscan_bitmap(Relation indexRelation,
 257                                            Snapshot snapshot,
 258                                            int nkeys)
 259 {
 260         IndexScanDesc scan;
 261
 262         scan = index_beginscan_internal(indexRelation, nkeys, 0);
 263
 264         /*
 265          * Save additional parameters into the scandesc.  Everything else was set
 266          * up by RelationGetIndexScan.
 267          */
 268         scan->xs_snapshot = snapshot;
 269
 270         return scan;
 271 }
 272
 273 /*
 274  * index_beginscan_internal --- common code for index_beginscan variants
 275  */
 276 static IndexScanDesc
 277 index_beginscan_internal(Relation indexRelation,
 278                                                  int nkeys, int norderbys)
 279 {
 280         IndexScanDesc scan;
 281         FmgrInfo   *procedure;
 282
 283         RELATION_CHECKS;
 284         GET_REL_PROCEDURE(ambeginscan);
 285
 286         if (!(indexRelation->rd_am->ampredlocks))
 287                 PredicateLockRelation(indexRelation);
 288
 289         /*
 290          * We hold a reference count to the relcache entry throughout the scan.
 291          */
 292         RelationIncrementReferenceCount(indexRelation);
 293
 294         /*
 295          * Tell the AM to open a scan.
 296          */
 297         scan = (IndexScanDesc)
 298                 DatumGetPointer(FunctionCall3(procedure,
 299                                                                           PointerGetDatum(indexRelation),
 300                                                                           Int32GetDatum(nkeys),
 301                                                                           Int32GetDatum(norderbys)));
 302
 303         return scan;
 304 }
 305
 306 /* ----------------
 307  *              index_rescan  - (re)start a scan of an index
 308  *
 309  * During a restart, the caller may specify a new set of scankeys and/or
 310  * orderbykeys; but the number of keys cannot differ from what index_beginscan
 311  * was told.  (Later we might relax that to "must not exceed", but currently
 312  * the index AMs tend to assume that scan->numberOfKeys is what to believe.)
 313  * To restart the scan without changing keys, pass NULL for the key arrays.
 314  * (Of course, keys *must* be passed on the first call, unless
 315  * scan->numberOfKeys is zero.)
 316  * ----------------
 317  */
 318 void
 319 index_rescan(IndexScanDesc scan,
 320                          ScanKey keys, int nkeys,
 321                          ScanKey orderbys, int norderbys)
 322 {
 323         FmgrInfo   *procedure;
 324
 325         SCAN_CHECKS;
 326         GET_SCAN_PROCEDURE(amrescan);
 327
 328         Assert(nkeys == scan->numberOfKeys);
 329         Assert(norderbys == scan->numberOfOrderBys);
 330
 331         /* Release any held pin on a heap page */
 332         if (BufferIsValid(scan->xs_cbuf))
 333         {
 334                 ReleaseBuffer(scan->xs_cbuf);
 335                 scan->xs_cbuf = InvalidBuffer;
 336         }
 337
 338         scan->xs_next_hot = InvalidOffsetNumber;
 339
 340         scan->kill_prior_tuple = false;         /* for safety */
 341
 342         FunctionCall5(procedure,
 343                                   PointerGetDatum(scan),
 344                                   PointerGetDatum(keys),
 345                                   Int32GetDatum(nkeys),
 346                                   PointerGetDatum(orderbys),
 347                                   Int32GetDatum(norderbys));
 348 }
 349
 350 /* ----------------
 351  *              index_endscan - end a scan
 352  * ----------------
 353  */
 354 void
 355 index_endscan(IndexScanDesc scan)
 356 {
 357         FmgrInfo   *procedure;
 358
 359         SCAN_CHECKS;
 360         GET_SCAN_PROCEDURE(amendscan);
 361
 362         /* Release any held pin on a heap page */
 363         if (BufferIsValid(scan->xs_cbuf))
 364         {
 365                 ReleaseBuffer(scan->xs_cbuf);
 366                 scan->xs_cbuf = InvalidBuffer;
 367         }
 368
 369         /* End the AM's scan */
 370         FunctionCall1(procedure, PointerGetDatum(scan));
 371
 372         /* Release index refcount acquired by index_beginscan */
 373         RelationDecrementReferenceCount(scan->indexRelation);
 374
 375         /* Release the scan data structure itself */
 376         IndexScanEnd(scan);
 377 }
 378
 379 /* ----------------
 380  *              index_markpos  - mark a scan position
 381  * ----------------
 382  */
 383 void
 384 index_markpos(IndexScanDesc scan)
 385 {
 386         FmgrInfo   *procedure;
 387
 388         SCAN_CHECKS;
 389         GET_SCAN_PROCEDURE(ammarkpos);
 390
 391         FunctionCall1(procedure, PointerGetDatum(scan));
 392 }
 393
 394 /* ----------------
 395  *              index_restrpos  - restore a scan position
 396  *
 397  * NOTE: this only restores the internal scan state of the index AM.
 398  * The current result tuple (scan->xs_ctup) doesn't change.  See comments
 399  * for ExecRestrPos().
 400  *
 401  * NOTE: in the presence of HOT chains, mark/restore only works correctly
 402  * if the scan's snapshot is MVCC-safe; that ensures that there's at most one
 403  * returnable tuple in each HOT chain, and so restoring the prior state at the
 404  * granularity of the index AM is sufficient.  Since the only current user
 405  * of mark/restore functionality is nodeMergejoin.c, this effectively means
 406  * that merge-join plans only work for MVCC snapshots.  This could be fixed
 407  * if necessary, but for now it seems unimportant.
 408  * ----------------
 409  */
 410 void
 411 index_restrpos(IndexScanDesc scan)
 412 {
 413         FmgrInfo   *procedure;
 414
 415         Assert(IsMVCCSnapshot(scan->xs_snapshot));
 416
 417         SCAN_CHECKS;
 418         GET_SCAN_PROCEDURE(amrestrpos);
 419
 420         scan->xs_next_hot = InvalidOffsetNumber;
 421
 422         scan->kill_prior_tuple = false;         /* for safety */
 423
 424         FunctionCall1(procedure, PointerGetDatum(scan));
 425 }
 426
 427 /* ----------------
 428  *              index_getnext - get the next heap tuple from a scan
 429  *
 430  * The result is the next heap tuple satisfying the scan keys and the
 431  * snapshot, or NULL if no more matching tuples exist.  On success,
 432  * the buffer containing the heap tuple is pinned (the pin will be dropped
 433  * at the next index_getnext or index_endscan).
 434  *
 435  * Note: caller must check scan->xs_recheck, and perform rechecking of the
 436  * scan keys if required.  We do not do that here because we don't have
 437  * enough information to do it efficiently in the general case.
 438  * ----------------
 439  */
 440 HeapTuple
 441 index_getnext(IndexScanDesc scan, ScanDirection direction)
 442 {
 443         HeapTuple       heapTuple = &scan->xs_ctup;
 444         ItemPointer tid = &heapTuple->t_self;
 445         FmgrInfo   *procedure;
 446
 447         SCAN_CHECKS;
 448         GET_SCAN_PROCEDURE(amgettuple);
 449
 450         Assert(TransactionIdIsValid(RecentGlobalXmin));
 451
 452         /*
 453          * We always reset xs_hot_dead; if we are here then either we are just
 454          * starting the scan, or we previously returned a visible tuple, and in
 455          * either case it's inappropriate to kill the prior index entry.
 456          */
 457         scan->xs_hot_dead = false;
 458
 459         for (;;)
 460         {
 461                 OffsetNumber offnum;
 462                 bool            at_chain_start;
 463                 Page            dp;
 464
 465                 if (scan->xs_next_hot != InvalidOffsetNumber)
 466                 {
 467                         /*
 468                          * We are resuming scan of a HOT chain after having returned an
 469                          * earlier member.      Must still hold pin on current heap page.
 470                          */
 471                         Assert(BufferIsValid(scan->xs_cbuf));
 472                         Assert(ItemPointerGetBlockNumber(tid) ==
 473                                    BufferGetBlockNumber(scan->xs_cbuf));
 474                         Assert(TransactionIdIsValid(scan->xs_prev_xmax));
 475                         offnum = scan->xs_next_hot;
 476                         at_chain_start = false;
 477                         scan->xs_next_hot = InvalidOffsetNumber;
 478                 }
 479                 else
 480                 {
 481                         bool            found;
 482                         Buffer          prev_buf;
 483
 484                         /*
 485                          * If we scanned a whole HOT chain and found only dead tuples,
 486                          * tell index AM to kill its entry for that TID. We do not do this
 487                          * when in recovery because it may violate MVCC to do so. see
 488                          * comments in RelationGetIndexScan().
 489                          */
 490                         if (!scan->xactStartedInRecovery)
 491                                 scan->kill_prior_tuple = scan->xs_hot_dead;
 492
 493                         /*
 494                          * The AM's gettuple proc finds the next index entry matching the
 495                          * scan keys, and puts the TID in xs_ctup.t_self (ie, *tid). It
 496                          * should also set scan->xs_recheck, though we pay no attention to
 497                          * that here.
 498                          */
 499                         found = DatumGetBool(FunctionCall2(procedure,
 500                                                                                            PointerGetDatum(scan),
 501                                                                                            Int32GetDatum(direction)));
 502
 503                         /* Reset kill flag immediately for safety */
 504                         scan->kill_prior_tuple = false;
 505
 506                         /* If we're out of index entries, break out of outer loop */
 507                         if (!found)
 508                                 break;
 509
 510                         pgstat_count_index_tuples(scan->indexRelation, 1);
 511
 512                         /* Switch to correct buffer if we don't have it already */
 513                         prev_buf = scan->xs_cbuf;
 514                         scan->xs_cbuf = ReleaseAndReadBuffer(scan->xs_cbuf,
 515                                                                                                  scan->heapRelation,
 516                                                                                          ItemPointerGetBlockNumber(tid));
 517
 518                         /*
 519                          * Prune page, but only if we weren't already on this page
 520                          */
 521                         if (prev_buf != scan->xs_cbuf)
 522                                 heap_page_prune_opt(scan->heapRelation, scan->xs_cbuf,
 523                                                                         RecentGlobalXmin);
 524
 525                         /* Prepare to scan HOT chain starting at index-referenced offnum */
 526                         offnum = ItemPointerGetOffsetNumber(tid);
 527                         at_chain_start = true;
 528
 529                         /* We don't know what the first tuple's xmin should be */
 530                         scan->xs_prev_xmax = InvalidTransactionId;
 531
 532                         /* Initialize flag to detect if all entries are dead */
 533                         scan->xs_hot_dead = true;
 534                 }
 535
 536                 /* Obtain share-lock on the buffer so we can examine visibility */
 537                 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_SHARE);
 538
 539                 dp = (Page) BufferGetPage(scan->xs_cbuf);
 540
 541                 /* Scan through possible multiple members of HOT-chain */
 542                 for (;;)
 543                 {
 544                         ItemId          lp;
 545                         ItemPointer ctid;
 546                         bool            valid;
 547
 548                         /* check for bogus TID */
 549                         if (offnum < FirstOffsetNumber ||
 550                                 offnum > PageGetMaxOffsetNumber(dp))
 551                                 break;
 552
 553                         lp = PageGetItemId(dp, offnum);
 554
 555                         /* check for unused, dead, or redirected items */
 556                         if (!ItemIdIsNormal(lp))
 557                         {
 558                                 /* We should only see a redirect at start of chain */
 559                                 if (ItemIdIsRedirected(lp) && at_chain_start)
 560                                 {
 561                                         /* Follow the redirect */
 562                                         offnum = ItemIdGetRedirect(lp);
 563                                         at_chain_start = false;
 564                                         continue;
 565                                 }
 566                                 /* else must be end of chain */
 567                                 break;
 568                         }
 569
 570                         /*
 571                          * We must initialize all of *heapTuple (ie, scan->xs_ctup) since
 572                          * it is returned to the executor on success.
 573                          */
 574                         heapTuple->t_data = (HeapTupleHeader) PageGetItem(dp, lp);
 575                         heapTuple->t_len = ItemIdGetLength(lp);
 576                         ItemPointerSetOffsetNumber(tid, offnum);
 577                         heapTuple->t_tableOid = RelationGetRelid(scan->heapRelation);
 578                         ctid = &heapTuple->t_data->t_ctid;
 579
 580                         /*
 581                          * Shouldn't see a HEAP_ONLY tuple at chain start.  (This test
 582                          * should be unnecessary, since the chain root can't be removed
 583                          * while we have pin on the index entry, but let's make it
 584                          * anyway.)
 585                          */
 586                         if (at_chain_start && HeapTupleIsHeapOnly(heapTuple))
 587                                 break;
 588
 589                         /*
 590                          * The xmin should match the previous xmax value, else chain is
 591                          * broken.      (Note: this test is not optional because it protects
 592                          * us against the case where the prior chain member's xmax aborted
 593                          * since we looked at it.)
 594                          */
 595                         if (TransactionIdIsValid(scan->xs_prev_xmax) &&
 596                                 !TransactionIdEquals(scan->xs_prev_xmax,
 597                                                                   HeapTupleHeaderGetXmin(heapTuple->t_data)))
 598                                 break;
 599
 600                         /* If it's visible per the snapshot, we must return it */
 601                         valid = HeapTupleSatisfiesVisibility(heapTuple, scan->xs_snapshot,
 602                                                                                                  scan->xs_cbuf);
 603
 604                         CheckForSerializableConflictOut(valid, scan->heapRelation,
 605                                                                                         heapTuple, scan->xs_cbuf);
 606
 607                         if (valid)
 608                         {
 609                                 /*
 610                                  * If the snapshot is MVCC, we know that it could accept at
 611                                  * most one member of the HOT chain, so we can skip examining
 612                                  * any more members.  Otherwise, check for continuation of the
 613                                  * HOT-chain, and set state for next time.
 614                                  */
 615                                 if (IsMVCCSnapshot(scan->xs_snapshot))
 616                                         scan->xs_next_hot = InvalidOffsetNumber;
 617                                 else if (HeapTupleIsHotUpdated(heapTuple))
 618                                 {
 619                                         Assert(ItemPointerGetBlockNumber(ctid) ==
 620                                                    ItemPointerGetBlockNumber(tid));
 621                                         scan->xs_next_hot = ItemPointerGetOffsetNumber(ctid);
 622                                         scan->xs_prev_xmax = HeapTupleHeaderGetXmax(heapTuple->t_data);
 623                                 }
 624                                 else
 625                                         scan->xs_next_hot = InvalidOffsetNumber;
 626
 627                                 PredicateLockTuple(scan->heapRelation, heapTuple);
 628
 629                                 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_UNLOCK);
 630
 631                                 pgstat_count_heap_fetch(scan->indexRelation);
 632
 633                                 return heapTuple;
 634                         }
 635
 636                         /*
 637                          * If we can't see it, maybe no one else can either.  Check to see
 638                          * if the tuple is dead to all transactions.  If we find that all
 639                          * the tuples in the HOT chain are dead, we'll signal the index AM
 640                          * to not return that TID on future indexscans.
 641                          */
 642                         if (scan->xs_hot_dead &&
 643                                 HeapTupleSatisfiesVacuum(heapTuple->t_data, RecentGlobalXmin,
 644                                                                                  scan->xs_cbuf) != HEAPTUPLE_DEAD)
 645                                 scan->xs_hot_dead = false;
 646
 647                         /*
 648                          * Check to see if HOT chain continues past this tuple; if so
 649                          * fetch the next offnum (we don't bother storing it into
 650                          * xs_next_hot, but must store xs_prev_xmax), and loop around.
 651                          */
 652                         if (HeapTupleIsHotUpdated(heapTuple))
 653                         {
 654                                 Assert(ItemPointerGetBlockNumber(ctid) ==
 655                                            ItemPointerGetBlockNumber(tid));
 656                                 offnum = ItemPointerGetOffsetNumber(ctid);
 657                                 at_chain_start = false;
 658                                 scan->xs_prev_xmax = HeapTupleHeaderGetXmax(heapTuple->t_data);
 659                         }
 660                         else
 661                                 break;                  /* end of chain */
 662                 }                                               /* loop over a single HOT chain */
 663
 664                 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_UNLOCK);
 665
 666                 /* Loop around to ask index AM for another TID */
 667                 scan->xs_next_hot = InvalidOffsetNumber;
 668         }
 669
 670         /* Release any held pin on a heap page */
 671         if (BufferIsValid(scan->xs_cbuf))
 672         {
 673                 ReleaseBuffer(scan->xs_cbuf);
 674                 scan->xs_cbuf = InvalidBuffer;
 675         }
 676
 677         return NULL;                            /* failure exit */
 678 }
 679
 680 /* ----------------
 681  *              index_getbitmap - get all tuples at once from an index scan
 682  *
 683  * Adds the TIDs of all heap tuples satisfying the scan keys to a bitmap.
 684  * Since there's no interlock between the index scan and the eventual heap
 685  * access, this is only safe to use with MVCC-based snapshots: the heap
 686  * item slot could have been replaced by a newer tuple by the time we get
 687  * to it.
 688  *
 689  * Returns the number of matching tuples found.  (Note: this might be only
 690  * approximate, so it should only be used for statistical purposes.)
 691  * ----------------
 692  */
 693 int64
 694 index_getbitmap(IndexScanDesc scan, TIDBitmap *bitmap)
 695 {
 696         FmgrInfo   *procedure;
 697         int64           ntids;
 698         Datum           d;
 699
 700         SCAN_CHECKS;
 701         GET_SCAN_PROCEDURE(amgetbitmap);
 702
 703         /* just make sure this is false... */
 704         scan->kill_prior_tuple = false;
 705
 706         /*
 707          * have the am's getbitmap proc do all the work.
 708          */
 709         d = FunctionCall2(procedure,
 710                                           PointerGetDatum(scan),
 711                                           PointerGetDatum(bitmap));
 712
 713         ntids = DatumGetInt64(d);
 714
 715         /* If int8 is pass-by-ref, must free the result to avoid memory leak */
 716 #ifndef USE_FLOAT8_BYVAL
 717         pfree(DatumGetPointer(d));
 718 #endif
 719
 720         pgstat_count_index_tuples(scan->indexRelation, ntids);
 721
 722         return ntids;
 723 }
 724
 725 /* ----------------
 726  *              index_bulk_delete - do mass deletion of index entries
 727  *
 728  *              callback routine tells whether a given main-heap tuple is
 729  *              to be deleted
 730  *
 731  *              return value is an optional palloc'd struct of statistics
 732  * ----------------
 733  */
 734 IndexBulkDeleteResult *
 735 index_bulk_delete(IndexVacuumInfo *info,
 736                                   IndexBulkDeleteResult *stats,
 737                                   IndexBulkDeleteCallback callback,
 738                                   void *callback_state)
 739 {
 740         Relation        indexRelation = info->index;
 741         FmgrInfo   *procedure;
 742         IndexBulkDeleteResult *result;
 743
 744         RELATION_CHECKS;
 745         GET_REL_PROCEDURE(ambulkdelete);
 746
 747         result = (IndexBulkDeleteResult *)
 748                 DatumGetPointer(FunctionCall4(procedure,
 749                                                                           PointerGetDatum(info),
 750                                                                           PointerGetDatum(stats),
 751                                                                           PointerGetDatum((Pointer) callback),
 752                                                                           PointerGetDatum(callback_state)));
 753
 754         return result;
 755 }
 756
 757 /* ----------------
 758  *              index_vacuum_cleanup - do post-deletion cleanup of an index
 759  *
 760  *              return value is an optional palloc'd struct of statistics
 761  * ----------------
 762  */
 763 IndexBulkDeleteResult *
 764 index_vacuum_cleanup(IndexVacuumInfo *info,
 765                                          IndexBulkDeleteResult *stats)
 766 {
 767         Relation        indexRelation = info->index;
 768         FmgrInfo   *procedure;
 769         IndexBulkDeleteResult *result;
 770
 771         RELATION_CHECKS;
 772         GET_REL_PROCEDURE(amvacuumcleanup);
 773
 774         result = (IndexBulkDeleteResult *)
 775                 DatumGetPointer(FunctionCall2(procedure,
 776                                                                           PointerGetDatum(info),
 777                                                                           PointerGetDatum(stats)));
 778
 779         return result;
 780 }
 781
 782 /* ----------------
 783  *              index_getprocid
 784  *
 785  *              Index access methods typically require support routines that are
 786  *              not directly the implementation of any WHERE-clause query operator
 787  *              and so cannot be kept in pg_amop.  Instead, such routines are kept
 788  *              in pg_amproc.  These registered procedure OIDs are assigned numbers
 789  *              according to a convention established by the access method.
 790  *              The general index code doesn't know anything about the routines
 791  *              involved; it just builds an ordered list of them for
 792  *              each attribute on which an index is defined.
 793  *
 794  *              As of Postgres 8.3, support routines within an operator family
 795  *              are further subdivided by the "left type" and "right type" of the
 796  *              query operator(s) that they support.  The "default" functions for a
 797  *              particular indexed attribute are those with both types equal to
 798  *              the index opclass' opcintype (note that this is subtly different
 799  *              from the indexed attribute's own type: it may be a binary-compatible
 800  *              type instead).  Only the default functions are stored in relcache
 801  *              entries --- access methods can use the syscache to look up non-default
 802  *              functions.
 803  *
 804  *              This routine returns the requested default procedure OID for a
 805  *              particular indexed attribute.
 806  * ----------------
 807  */
 808 RegProcedure
 809 index_getprocid(Relation irel,
 810                                 AttrNumber attnum,
 811                                 uint16 procnum)
 812 {
 813         RegProcedure *loc;
 814         int                     nproc;
 815         int                     procindex;
 816
 817         nproc = irel->rd_am->amsupport;
 818
 819         Assert(procnum > 0 && procnum <= (uint16) nproc);
 820
 821         procindex = (nproc * (attnum - 1)) + (procnum - 1);
 822
 823         loc = irel->rd_support;
 824
 825         Assert(loc != NULL);
 826
 827         return loc[procindex];
 828 }
 829
 830 /* ----------------
 831  *              index_getprocinfo
 832  *
 833  *              This routine allows index AMs to keep fmgr lookup info for
 834  *              support procs in the relcache.  As above, only the "default"
 835  *              functions for any particular indexed attribute are cached.
 836  *
 837  * Note: the return value points into cached data that will be lost during
 838  * any relcache rebuild!  Therefore, either use the callinfo right away,
 839  * or save it only after having acquired some type of lock on the index rel.
 840  * ----------------
 841  */
 842 FmgrInfo *
 843 index_getprocinfo(Relation irel,
 844                                   AttrNumber attnum,
 845                                   uint16 procnum)
 846 {
 847         FmgrInfo   *locinfo;
 848         int                     nproc;
 849         int                     procindex;
 850
 851         nproc = irel->rd_am->amsupport;
 852
 853         Assert(procnum > 0 && procnum <= (uint16) nproc);
 854
 855         procindex = (nproc * (attnum - 1)) + (procnum - 1);
 856
 857         locinfo = irel->rd_supportinfo;
 858
 859         Assert(locinfo != NULL);
 860
 861         locinfo += procindex;
 862
 863         /* Initialize the lookup info if first time through */
 864         if (locinfo->fn_oid == InvalidOid)
 865         {
 866                 RegProcedure *loc = irel->rd_support;
 867                 RegProcedure procId;
 868
 869                 Assert(loc != NULL);
 870
 871                 procId = loc[procindex];
 872
 873                 /*
 874                  * Complain if function was not found during IndexSupportInitialize.
 875                  * This should not happen unless the system tables contain bogus
 876                  * entries for the index opclass.  (If an AM wants to allow a support
 877                  * function to be optional, it can use index_getprocid.)
 878                  */
 879                 if (!RegProcedureIsValid(procId))
 880                         elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
 881                                  procnum, attnum, RelationGetRelationName(irel));
 882
 883                 fmgr_info_cxt(procId, locinfo, irel->rd_indexcxt);
 884         }
 885
 886         return locinfo;
 887 }