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) 1996-2008, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        $PostgreSQL: pgsql/src/backend/access/index/indexam.c,v 1.101 2008/01/01 19:45:46 momjian Exp $
  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_multi - start a scan of an index with amgetmulti
  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_getmulti  - get multiple 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/genam.h"
  66 #include "access/heapam.h"
  67 #include "access/transam.h"
  68 #include "pgstat.h"
  69 #include "utils/relcache.h"
  70
  71
  72 /* ----------------------------------------------------------------
  73  *                                      macros used in index_ routines
  74  * ----------------------------------------------------------------
  75  */
  76 #define RELATION_CHECKS \
  77 ( \
  78         AssertMacro(RelationIsValid(indexRelation)), \
  79         AssertMacro(PointerIsValid(indexRelation->rd_am)) \
  80 )
  81
  82 #define SCAN_CHECKS \
  83 ( \
  84         AssertMacro(IndexScanIsValid(scan)), \
  85         AssertMacro(RelationIsValid(scan->indexRelation)), \
  86         AssertMacro(PointerIsValid(scan->indexRelation->rd_am)) \
  87 )
  88
  89 #define GET_REL_PROCEDURE(pname) \
  90 do { \
  91         procedure = &indexRelation->rd_aminfo->pname; \
  92         if (!OidIsValid(procedure->fn_oid)) \
  93         { \
  94                 RegProcedure    procOid = indexRelation->rd_am->pname; \
  95                 if (!RegProcedureIsValid(procOid)) \
  96                         elog(ERROR, "invalid %s regproc", CppAsString(pname)); \
  97                 fmgr_info_cxt(procOid, procedure, indexRelation->rd_indexcxt); \
  98         } \
  99 } while(0)
 100
 101 #define GET_SCAN_PROCEDURE(pname) \
 102 do { \
 103         procedure = &scan->indexRelation->rd_aminfo->pname; \
 104         if (!OidIsValid(procedure->fn_oid)) \
 105         { \
 106                 RegProcedure    procOid = scan->indexRelation->rd_am->pname; \
 107                 if (!RegProcedureIsValid(procOid)) \
 108                         elog(ERROR, "invalid %s regproc", CppAsString(pname)); \
 109                 fmgr_info_cxt(procOid, procedure, scan->indexRelation->rd_indexcxt); \
 110         } \
 111 } while(0)
 112
 113 static IndexScanDesc index_beginscan_internal(Relation indexRelation,
 114                                                  int nkeys, ScanKey key);
 115
 116
 117 /* ----------------------------------------------------------------
 118  *                                 index_ interface functions
 119  * ----------------------------------------------------------------
 120  */
 121
 122 /* ----------------
 123  *              index_open - open an index relation by relation OID
 124  *
 125  *              If lockmode is not "NoLock", the specified kind of lock is
 126  *              obtained on the index.  (Generally, NoLock should only be
 127  *              used if the caller knows it has some appropriate lock on the
 128  *              index already.)
 129  *
 130  *              An error is raised if the index does not exist.
 131  *
 132  *              This is a convenience routine adapted for indexscan use.
 133  *              Some callers may prefer to use relation_open directly.
 134  * ----------------
 135  */
 136 Relation
 137 index_open(Oid relationId, LOCKMODE lockmode)
 138 {
 139         Relation        r;
 140
 141         r = relation_open(relationId, lockmode);
 142
 143         if (r->rd_rel->relkind != RELKIND_INDEX)
 144                 ereport(ERROR,
 145                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
 146                                  errmsg("\"%s\" is not an index",
 147                                                 RelationGetRelationName(r))));
 148
 149         return r;
 150 }
 151
 152 /* ----------------
 153  *              index_close - close an index relation
 154  *
 155  *              If lockmode is not "NoLock", we then release the specified lock.
 156  *
 157  *              Note that it is often sensible to hold a lock beyond index_close;
 158  *              in that case, the lock is released automatically at xact end.
 159  * ----------------
 160  */
 161 void
 162 index_close(Relation relation, LOCKMODE lockmode)
 163 {
 164         LockRelId       relid = relation->rd_lockInfo.lockRelId;
 165
 166         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
 167
 168         /* The relcache does the real work... */
 169         RelationClose(relation);
 170
 171         if (lockmode != NoLock)
 172                 UnlockRelationId(&relid, lockmode);
 173 }
 174
 175 /* ----------------
 176  *              index_insert - insert an index tuple into a relation
 177  * ----------------
 178  */
 179 bool
 180 index_insert(Relation indexRelation,
 181                          Datum *values,
 182                          bool *isnull,
 183                          ItemPointer heap_t_ctid,
 184                          Relation heapRelation,
 185                          bool check_uniqueness)
 186 {
 187         FmgrInfo   *procedure;
 188
 189         RELATION_CHECKS;
 190         GET_REL_PROCEDURE(aminsert);
 191
 192         /*
 193          * have the am's insert proc do all the work.
 194          */
 195         return DatumGetBool(FunctionCall6(procedure,
 196                                                                           PointerGetDatum(indexRelation),
 197                                                                           PointerGetDatum(values),
 198                                                                           PointerGetDatum(isnull),
 199                                                                           PointerGetDatum(heap_t_ctid),
 200                                                                           PointerGetDatum(heapRelation),
 201                                                                           BoolGetDatum(check_uniqueness)));
 202 }
 203
 204 /*
 205  * index_beginscan - start a scan of an index with amgettuple
 206  *
 207  * Note: heapRelation may be NULL if there is no intention of calling
 208  * index_getnext on this scan; index_getnext_indexitem will not use the
 209  * heapRelation link (nor the snapshot).  However, the caller had better
 210  * be holding some kind of lock on the heap relation in any case, to ensure
 211  * no one deletes it (or the index) out from under us.  Caller must also
 212  * be holding a lock on the index.
 213  */
 214 IndexScanDesc
 215 index_beginscan(Relation heapRelation,
 216                                 Relation indexRelation,
 217                                 Snapshot snapshot,
 218                                 int nkeys, ScanKey key)
 219 {
 220         IndexScanDesc scan;
 221
 222         scan = index_beginscan_internal(indexRelation, nkeys, key);
 223
 224         /*
 225          * Save additional parameters into the scandesc.  Everything else was set
 226          * up by RelationGetIndexScan.
 227          */
 228         scan->is_multiscan = false;
 229         scan->heapRelation = heapRelation;
 230         scan->xs_snapshot = snapshot;
 231
 232         return scan;
 233 }
 234
 235 /*
 236  * index_beginscan_multi - start a scan of an index with amgetmulti
 237  *
 238  * As above, caller had better be holding some lock on the parent heap
 239  * relation, even though it's not explicitly mentioned here.
 240  */
 241 IndexScanDesc
 242 index_beginscan_multi(Relation indexRelation,
 243                                           Snapshot snapshot,
 244                                           int nkeys, ScanKey key)
 245 {
 246         IndexScanDesc scan;
 247
 248         scan = index_beginscan_internal(indexRelation, nkeys, key);
 249
 250         /*
 251          * Save additional parameters into the scandesc.  Everything else was set
 252          * up by RelationGetIndexScan.
 253          */
 254         scan->is_multiscan = true;
 255         scan->xs_snapshot = snapshot;
 256
 257         return scan;
 258 }
 259
 260 /*
 261  * index_beginscan_internal --- common code for index_beginscan variants
 262  */
 263 static IndexScanDesc
 264 index_beginscan_internal(Relation indexRelation,
 265                                                  int nkeys, ScanKey key)
 266 {
 267         IndexScanDesc scan;
 268         FmgrInfo   *procedure;
 269
 270         RELATION_CHECKS;
 271         GET_REL_PROCEDURE(ambeginscan);
 272
 273         /*
 274          * We hold a reference count to the relcache entry throughout the scan.
 275          */
 276         RelationIncrementReferenceCount(indexRelation);
 277
 278         /*
 279          * Tell the AM to open a scan.
 280          */
 281         scan = (IndexScanDesc)
 282                 DatumGetPointer(FunctionCall3(procedure,
 283                                                                           PointerGetDatum(indexRelation),
 284                                                                           Int32GetDatum(nkeys),
 285                                                                           PointerGetDatum(key)));
 286
 287         return scan;
 288 }
 289
 290 /* ----------------
 291  *              index_rescan  - (re)start a scan of an index
 292  *
 293  * The caller may specify a new set of scankeys (but the number of keys
 294  * cannot change).      To restart the scan without changing keys, pass NULL
 295  * for the key array.
 296  *
 297  * Note that this is also called when first starting an indexscan;
 298  * see RelationGetIndexScan.  Keys *must* be passed in that case,
 299  * unless scan->numberOfKeys is zero.
 300  * ----------------
 301  */
 302 void
 303 index_rescan(IndexScanDesc scan, ScanKey key)
 304 {
 305         FmgrInfo   *procedure;
 306
 307         SCAN_CHECKS;
 308         GET_SCAN_PROCEDURE(amrescan);
 309
 310         /* Release any held pin on a heap page */
 311         if (BufferIsValid(scan->xs_cbuf))
 312         {
 313                 ReleaseBuffer(scan->xs_cbuf);
 314                 scan->xs_cbuf = InvalidBuffer;
 315         }
 316
 317         scan->xs_next_hot = InvalidOffsetNumber;
 318
 319         scan->kill_prior_tuple = false;         /* for safety */
 320
 321         FunctionCall2(procedure,
 322                                   PointerGetDatum(scan),
 323                                   PointerGetDatum(key));
 324 }
 325
 326 /* ----------------
 327  *              index_endscan - end a scan
 328  * ----------------
 329  */
 330 void
 331 index_endscan(IndexScanDesc scan)
 332 {
 333         FmgrInfo   *procedure;
 334
 335         SCAN_CHECKS;
 336         GET_SCAN_PROCEDURE(amendscan);
 337
 338         /* Release any held pin on a heap page */
 339         if (BufferIsValid(scan->xs_cbuf))
 340         {
 341                 ReleaseBuffer(scan->xs_cbuf);
 342                 scan->xs_cbuf = InvalidBuffer;
 343         }
 344
 345         /* End the AM's scan */
 346         FunctionCall1(procedure, PointerGetDatum(scan));
 347
 348         /* Release index refcount acquired by index_beginscan */
 349         RelationDecrementReferenceCount(scan->indexRelation);
 350
 351         /* Release the scan data structure itself */
 352         IndexScanEnd(scan);
 353 }
 354
 355 /* ----------------
 356  *              index_markpos  - mark a scan position
 357  * ----------------
 358  */
 359 void
 360 index_markpos(IndexScanDesc scan)
 361 {
 362         FmgrInfo   *procedure;
 363
 364         SCAN_CHECKS;
 365         GET_SCAN_PROCEDURE(ammarkpos);
 366
 367         FunctionCall1(procedure, PointerGetDatum(scan));
 368 }
 369
 370 /* ----------------
 371  *              index_restrpos  - restore a scan position
 372  *
 373  * NOTE: this only restores the internal scan state of the index AM.
 374  * The current result tuple (scan->xs_ctup) doesn't change.  See comments
 375  * for ExecRestrPos().
 376  *
 377  * NOTE: in the presence of HOT chains, mark/restore only works correctly
 378  * if the scan's snapshot is MVCC-safe; that ensures that there's at most one
 379  * returnable tuple in each HOT chain, and so restoring the prior state at the
 380  * granularity of the index AM is sufficient.  Since the only current user
 381  * of mark/restore functionality is nodeMergejoin.c, this effectively means
 382  * that merge-join plans only work for MVCC snapshots.  This could be fixed
 383  * if necessary, but for now it seems unimportant.
 384  * ----------------
 385  */
 386 void
 387 index_restrpos(IndexScanDesc scan)
 388 {
 389         FmgrInfo   *procedure;
 390
 391         Assert(IsMVCCSnapshot(scan->xs_snapshot));
 392
 393         SCAN_CHECKS;
 394         GET_SCAN_PROCEDURE(amrestrpos);
 395
 396         scan->xs_next_hot = InvalidOffsetNumber;
 397
 398         scan->kill_prior_tuple = false;         /* for safety */
 399
 400         FunctionCall1(procedure, PointerGetDatum(scan));
 401 }
 402
 403 /* ----------------
 404  *              index_getnext - get the next heap tuple from a scan
 405  *
 406  * The result is the next heap tuple satisfying the scan keys and the
 407  * snapshot, or NULL if no more matching tuples exist.  On success,
 408  * the buffer containing the heap tuple is pinned (the pin will be dropped
 409  * at the next index_getnext or index_endscan).
 410  * ----------------
 411  */
 412 HeapTuple
 413 index_getnext(IndexScanDesc scan, ScanDirection direction)
 414 {
 415         HeapTuple       heapTuple = &scan->xs_ctup;
 416         ItemPointer tid = &heapTuple->t_self;
 417         FmgrInfo   *procedure;
 418
 419         SCAN_CHECKS;
 420         GET_SCAN_PROCEDURE(amgettuple);
 421
 422         /*
 423          * We always reset xs_hot_dead; if we are here then either we are just
 424          * starting the scan, or we previously returned a visible tuple, and in
 425          * either case it's inappropriate to kill the prior index entry.
 426          */
 427         scan->xs_hot_dead = false;
 428
 429         for (;;)
 430         {
 431                 OffsetNumber offnum;
 432                 bool            at_chain_start;
 433                 Page            dp;
 434
 435                 if (scan->xs_next_hot != InvalidOffsetNumber)
 436                 {
 437                         /*
 438                          * We are resuming scan of a HOT chain after having returned an
 439                          * earlier member.      Must still hold pin on current heap page.
 440                          */
 441                         Assert(BufferIsValid(scan->xs_cbuf));
 442                         Assert(ItemPointerGetBlockNumber(tid) ==
 443                                    BufferGetBlockNumber(scan->xs_cbuf));
 444                         Assert(TransactionIdIsValid(scan->xs_prev_xmax));
 445                         offnum = scan->xs_next_hot;
 446                         at_chain_start = false;
 447                         scan->xs_next_hot = InvalidOffsetNumber;
 448                 }
 449                 else
 450                 {
 451                         bool            found;
 452                         Buffer          prev_buf;
 453
 454                         /*
 455                          * If we scanned a whole HOT chain and found only dead tuples,
 456                          * tell index AM to kill its entry for that TID.
 457                          */
 458                         scan->kill_prior_tuple = scan->xs_hot_dead;
 459
 460                         /*
 461                          * The AM's gettuple proc finds the next index entry matching the
 462                          * scan keys, and puts the TID in xs_ctup.t_self (ie, *tid).
 463                          */
 464                         found = DatumGetBool(FunctionCall2(procedure,
 465                                                                                            PointerGetDatum(scan),
 466                                                                                            Int32GetDatum(direction)));
 467
 468                         /* Reset kill flag immediately for safety */
 469                         scan->kill_prior_tuple = false;
 470
 471                         /* If we're out of index entries, break out of outer loop */
 472                         if (!found)
 473                                 break;
 474
 475                         pgstat_count_index_tuples(scan->indexRelation, 1);
 476
 477                         /* Switch to correct buffer if we don't have it already */
 478                         prev_buf = scan->xs_cbuf;
 479                         scan->xs_cbuf = ReleaseAndReadBuffer(scan->xs_cbuf,
 480                                                                                                  scan->heapRelation,
 481                                                                                          ItemPointerGetBlockNumber(tid));
 482
 483                         /*
 484                          * Prune page, but only if we weren't already on this page
 485                          */
 486                         if (prev_buf != scan->xs_cbuf)
 487                                 heap_page_prune_opt(scan->heapRelation, scan->xs_cbuf,
 488                                                                         RecentGlobalXmin);
 489
 490                         /* Prepare to scan HOT chain starting at index-referenced offnum */
 491                         offnum = ItemPointerGetOffsetNumber(tid);
 492                         at_chain_start = true;
 493
 494                         /* We don't know what the first tuple's xmin should be */
 495                         scan->xs_prev_xmax = InvalidTransactionId;
 496
 497                         /* Initialize flag to detect if all entries are dead */
 498                         scan->xs_hot_dead = true;
 499                 }
 500
 501                 /* Obtain share-lock on the buffer so we can examine visibility */
 502                 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_SHARE);
 503
 504                 dp = (Page) BufferGetPage(scan->xs_cbuf);
 505
 506                 /* Scan through possible multiple members of HOT-chain */
 507                 for (;;)
 508                 {
 509                         ItemId          lp;
 510                         ItemPointer ctid;
 511
 512                         /* check for bogus TID */
 513                         if (offnum < FirstOffsetNumber ||
 514                                 offnum > PageGetMaxOffsetNumber(dp))
 515                                 break;
 516
 517                         lp = PageGetItemId(dp, offnum);
 518
 519                         /* check for unused, dead, or redirected items */
 520                         if (!ItemIdIsNormal(lp))
 521                         {
 522                                 /* We should only see a redirect at start of chain */
 523                                 if (ItemIdIsRedirected(lp) && at_chain_start)
 524                                 {
 525                                         /* Follow the redirect */
 526                                         offnum = ItemIdGetRedirect(lp);
 527                                         at_chain_start = false;
 528                                         continue;
 529                                 }
 530                                 /* else must be end of chain */
 531                                 break;
 532                         }
 533
 534                         /*
 535                          * We must initialize all of *heapTuple (ie, scan->xs_ctup) since
 536                          * it is returned to the executor on success.
 537                          */
 538                         heapTuple->t_data = (HeapTupleHeader) PageGetItem(dp, lp);
 539                         heapTuple->t_len = ItemIdGetLength(lp);
 540                         ItemPointerSetOffsetNumber(tid, offnum);
 541                         heapTuple->t_tableOid = RelationGetRelid(scan->heapRelation);
 542                         ctid = &heapTuple->t_data->t_ctid;
 543
 544                         /*
 545                          * Shouldn't see a HEAP_ONLY tuple at chain start.  (This test
 546                          * should be unnecessary, since the chain root can't be removed
 547                          * while we have pin on the index entry, but let's make it
 548                          * anyway.)
 549                          */
 550                         if (at_chain_start && HeapTupleIsHeapOnly(heapTuple))
 551                                 break;
 552
 553                         /*
 554                          * The xmin should match the previous xmax value, else chain is
 555                          * broken.      (Note: this test is not optional because it protects
 556                          * us against the case where the prior chain member's xmax aborted
 557                          * since we looked at it.)
 558                          */
 559                         if (TransactionIdIsValid(scan->xs_prev_xmax) &&
 560                                 !TransactionIdEquals(scan->xs_prev_xmax,
 561                                                                   HeapTupleHeaderGetXmin(heapTuple->t_data)))
 562                                 break;
 563
 564                         /* If it's visible per the snapshot, we must return it */
 565                         if (HeapTupleSatisfiesVisibility(heapTuple, scan->xs_snapshot,
 566                                                                                          scan->xs_cbuf))
 567                         {
 568                                 /*
 569                                  * If the snapshot is MVCC, we know that it could accept at
 570                                  * most one member of the HOT chain, so we can skip examining
 571                                  * any more members.  Otherwise, check for continuation of the
 572                                  * HOT-chain, and set state for next time.
 573                                  */
 574                                 if (IsMVCCSnapshot(scan->xs_snapshot))
 575                                         scan->xs_next_hot = InvalidOffsetNumber;
 576                                 else if (HeapTupleIsHotUpdated(heapTuple))
 577                                 {
 578                                         Assert(ItemPointerGetBlockNumber(ctid) ==
 579                                                    ItemPointerGetBlockNumber(tid));
 580                                         scan->xs_next_hot = ItemPointerGetOffsetNumber(ctid);
 581                                         scan->xs_prev_xmax = HeapTupleHeaderGetXmax(heapTuple->t_data);
 582                                 }
 583                                 else
 584                                         scan->xs_next_hot = InvalidOffsetNumber;
 585
 586                                 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_UNLOCK);
 587
 588                                 pgstat_count_heap_fetch(scan->indexRelation);
 589
 590                                 return heapTuple;
 591                         }
 592
 593                         /*
 594                          * If we can't see it, maybe no one else can either.  Check to see
 595                          * if the tuple is dead to all transactions.  If we find that all
 596                          * the tuples in the HOT chain are dead, we'll signal the index AM
 597                          * to not return that TID on future indexscans.
 598                          */
 599                         if (scan->xs_hot_dead &&
 600                                 HeapTupleSatisfiesVacuum(heapTuple->t_data, RecentGlobalXmin,
 601                                                                                  scan->xs_cbuf) != HEAPTUPLE_DEAD)
 602                                 scan->xs_hot_dead = false;
 603
 604                         /*
 605                          * Check to see if HOT chain continues past this tuple; if so
 606                          * fetch the next offnum (we don't bother storing it into
 607                          * xs_next_hot, but must store xs_prev_xmax), and loop around.
 608                          */
 609                         if (HeapTupleIsHotUpdated(heapTuple))
 610                         {
 611                                 Assert(ItemPointerGetBlockNumber(ctid) ==
 612                                            ItemPointerGetBlockNumber(tid));
 613                                 offnum = ItemPointerGetOffsetNumber(ctid);
 614                                 at_chain_start = false;
 615                                 scan->xs_prev_xmax = HeapTupleHeaderGetXmax(heapTuple->t_data);
 616                         }
 617                         else
 618                                 break;                  /* end of chain */
 619                 }                                               /* loop over a single HOT chain */
 620
 621                 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_UNLOCK);
 622
 623                 /* Loop around to ask index AM for another TID */
 624                 scan->xs_next_hot = InvalidOffsetNumber;
 625         }
 626
 627         /* Release any held pin on a heap page */
 628         if (BufferIsValid(scan->xs_cbuf))
 629         {
 630                 ReleaseBuffer(scan->xs_cbuf);
 631                 scan->xs_cbuf = InvalidBuffer;
 632         }
 633
 634         return NULL;                            /* failure exit */
 635 }
 636
 637 /* ----------------
 638  *              index_getnext_indexitem - get the next index tuple from a scan
 639  *
 640  * Finds the next index tuple satisfying the scan keys.  Note that the
 641  * corresponding heap tuple is not accessed, and thus no time qual (snapshot)
 642  * check is done, other than the index AM's internal check for killed tuples
 643  * (which most callers of this routine will probably want to suppress by
 644  * setting scan->ignore_killed_tuples = false).
 645  *
 646  * On success (TRUE return), the heap TID of the found index entry is in
 647  * scan->xs_ctup.t_self.  scan->xs_cbuf is untouched.
 648  * ----------------
 649  */
 650 bool
 651 index_getnext_indexitem(IndexScanDesc scan,
 652                                                 ScanDirection direction)
 653 {
 654         FmgrInfo   *procedure;
 655         bool            found;
 656
 657         SCAN_CHECKS;
 658         GET_SCAN_PROCEDURE(amgettuple);
 659
 660         /* just make sure this is false... */
 661         scan->kill_prior_tuple = false;
 662
 663         /*
 664          * have the am's gettuple proc do all the work.
 665          */
 666         found = DatumGetBool(FunctionCall2(procedure,
 667                                                                            PointerGetDatum(scan),
 668                                                                            Int32GetDatum(direction)));
 669
 670         if (found)
 671                 pgstat_count_index_tuples(scan->indexRelation, 1);
 672
 673         return found;
 674 }
 675
 676 /* ----------------
 677  *              index_getmulti - get multiple tuples from an index scan
 678  *
 679  * Collects the TIDs of multiple heap tuples satisfying the scan keys.
 680  * Since there's no interlock between the index scan and the eventual heap
 681  * access, this is only safe to use with MVCC-based snapshots: the heap
 682  * item slot could have been replaced by a newer tuple by the time we get
 683  * to it.
 684  *
 685  * A TRUE result indicates more calls should occur; a FALSE result says the
 686  * scan is done.  *returned_tids could be zero or nonzero in either case.
 687  * ----------------
 688  */
 689 bool
 690 index_getmulti(IndexScanDesc scan,
 691                            ItemPointer tids, int32 max_tids,
 692                            int32 *returned_tids)
 693 {
 694         FmgrInfo   *procedure;
 695         bool            found;
 696
 697         SCAN_CHECKS;
 698         GET_SCAN_PROCEDURE(amgetmulti);
 699
 700         /* just make sure this is false... */
 701         scan->kill_prior_tuple = false;
 702
 703         /*
 704          * have the am's getmulti proc do all the work.
 705          */
 706         found = DatumGetBool(FunctionCall4(procedure,
 707                                                                            PointerGetDatum(scan),
 708                                                                            PointerGetDatum(tids),
 709                                                                            Int32GetDatum(max_tids),
 710                                                                            PointerGetDatum(returned_tids)));
 711
 712         pgstat_count_index_tuples(scan->indexRelation, *returned_tids);
 713
 714         return found;
 715 }
 716
 717 /* ----------------
 718  *              index_bulk_delete - do mass deletion of index entries
 719  *
 720  *              callback routine tells whether a given main-heap tuple is
 721  *              to be deleted
 722  *
 723  *              return value is an optional palloc'd struct of statistics
 724  * ----------------
 725  */
 726 IndexBulkDeleteResult *
 727 index_bulk_delete(IndexVacuumInfo *info,
 728                                   IndexBulkDeleteResult *stats,
 729                                   IndexBulkDeleteCallback callback,
 730                                   void *callback_state)
 731 {
 732         Relation        indexRelation = info->index;
 733         FmgrInfo   *procedure;
 734         IndexBulkDeleteResult *result;
 735
 736         RELATION_CHECKS;
 737         GET_REL_PROCEDURE(ambulkdelete);
 738
 739         result = (IndexBulkDeleteResult *)
 740                 DatumGetPointer(FunctionCall4(procedure,
 741                                                                           PointerGetDatum(info),
 742                                                                           PointerGetDatum(stats),
 743                                                                           PointerGetDatum((Pointer) callback),
 744                                                                           PointerGetDatum(callback_state)));
 745
 746         return result;
 747 }
 748
 749 /* ----------------
 750  *              index_vacuum_cleanup - do post-deletion cleanup of an index
 751  *
 752  *              return value is an optional palloc'd struct of statistics
 753  * ----------------
 754  */
 755 IndexBulkDeleteResult *
 756 index_vacuum_cleanup(IndexVacuumInfo *info,
 757                                          IndexBulkDeleteResult *stats)
 758 {
 759         Relation        indexRelation = info->index;
 760         FmgrInfo   *procedure;
 761         IndexBulkDeleteResult *result;
 762
 763         RELATION_CHECKS;
 764         GET_REL_PROCEDURE(amvacuumcleanup);
 765
 766         result = (IndexBulkDeleteResult *)
 767                 DatumGetPointer(FunctionCall2(procedure,
 768                                                                           PointerGetDatum(info),
 769                                                                           PointerGetDatum(stats)));
 770
 771         return result;
 772 }
 773
 774 /* ----------------
 775  *              index_getprocid
 776  *
 777  *              Index access methods typically require support routines that are
 778  *              not directly the implementation of any WHERE-clause query operator
 779  *              and so cannot be kept in pg_amop.  Instead, such routines are kept
 780  *              in pg_amproc.  These registered procedure OIDs are assigned numbers
 781  *              according to a convention established by the access method.
 782  *              The general index code doesn't know anything about the routines
 783  *              involved; it just builds an ordered list of them for
 784  *              each attribute on which an index is defined.
 785  *
 786  *              As of Postgres 8.3, support routines within an operator family
 787  *              are further subdivided by the "left type" and "right type" of the
 788  *              query operator(s) that they support.  The "default" functions for a
 789  *              particular indexed attribute are those with both types equal to
 790  *              the index opclass' opcintype (note that this is subtly different
 791  *              from the indexed attribute's own type: it may be a binary-compatible
 792  *              type instead).  Only the default functions are stored in relcache
 793  *              entries --- access methods can use the syscache to look up non-default
 794  *              functions.
 795  *
 796  *              This routine returns the requested default procedure OID for a
 797  *              particular indexed attribute.
 798  * ----------------
 799  */
 800 RegProcedure
 801 index_getprocid(Relation irel,
 802                                 AttrNumber attnum,
 803                                 uint16 procnum)
 804 {
 805         RegProcedure *loc;
 806         int                     nproc;
 807         int                     procindex;
 808
 809         nproc = irel->rd_am->amsupport;
 810
 811         Assert(procnum > 0 && procnum <= (uint16) nproc);
 812
 813         procindex = (nproc * (attnum - 1)) + (procnum - 1);
 814
 815         loc = irel->rd_support;
 816
 817         Assert(loc != NULL);
 818
 819         return loc[procindex];
 820 }
 821
 822 /* ----------------
 823  *              index_getprocinfo
 824  *
 825  *              This routine allows index AMs to keep fmgr lookup info for
 826  *              support procs in the relcache.  As above, only the "default"
 827  *              functions for any particular indexed attribute are cached.
 828  *
 829  * Note: the return value points into cached data that will be lost during
 830  * any relcache rebuild!  Therefore, either use the callinfo right away,
 831  * or save it only after having acquired some type of lock on the index rel.
 832  * ----------------
 833  */
 834 FmgrInfo *
 835 index_getprocinfo(Relation irel,
 836                                   AttrNumber attnum,
 837                                   uint16 procnum)
 838 {
 839         FmgrInfo   *locinfo;
 840         int                     nproc;
 841         int                     procindex;
 842
 843         nproc = irel->rd_am->amsupport;
 844
 845         Assert(procnum > 0 && procnum <= (uint16) nproc);
 846
 847         procindex = (nproc * (attnum - 1)) + (procnum - 1);
 848
 849         locinfo = irel->rd_supportinfo;
 850
 851         Assert(locinfo != NULL);
 852
 853         locinfo += procindex;
 854
 855         /* Initialize the lookup info if first time through */
 856         if (locinfo->fn_oid == InvalidOid)
 857         {
 858                 RegProcedure *loc = irel->rd_support;
 859                 RegProcedure procId;
 860
 861                 Assert(loc != NULL);
 862
 863                 procId = loc[procindex];
 864
 865                 /*
 866                  * Complain if function was not found during IndexSupportInitialize.
 867                  * This should not happen unless the system tables contain bogus
 868                  * entries for the index opclass.  (If an AM wants to allow a support
 869                  * function to be optional, it can use index_getprocid.)
 870                  */
 871                 if (!RegProcedureIsValid(procId))
 872                         elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
 873                                  procnum, attnum, RelationGetRelationName(irel));
 874
 875                 fmgr_info_cxt(procId, locinfo, irel->rd_indexcxt);
 876         }
 877
 878         return locinfo;
 879 }