1 /*-------------------------------------------------------------------------
4 * general index access method routines
6 * Portions Copyright (c) 1996-2010, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * src/backend/access/index/indexam.c
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
31 * This file contains the index_ routines which used
32 * to be a scattered collection of stuff in access/genam.
36 * Scans are implemented as follows:
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.
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
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)
55 * All other states cannot occur.
57 * Note: It would be possible to cache the status of the previous and
58 * next item pointer using the flags.
60 *-------------------------------------------------------------------------
65 #include "access/relscan.h"
66 #include "access/transam.h"
68 #include "storage/bufmgr.h"
69 #include "storage/lmgr.h"
70 #include "utils/relcache.h"
71 #include "utils/snapmgr.h"
72 #include "utils/tqual.h"
75 /* ----------------------------------------------------------------
76 * macros used in index_ routines
77 * ----------------------------------------------------------------
79 #define RELATION_CHECKS \
81 AssertMacro(RelationIsValid(indexRelation)), \
82 AssertMacro(PointerIsValid(indexRelation->rd_am)) \
87 AssertMacro(IndexScanIsValid(scan)), \
88 AssertMacro(RelationIsValid(scan->indexRelation)), \
89 AssertMacro(PointerIsValid(scan->indexRelation->rd_am)) \
92 #define GET_REL_PROCEDURE(pname) \
94 procedure = &indexRelation->rd_aminfo->pname; \
95 if (!OidIsValid(procedure->fn_oid)) \
97 RegProcedure procOid = indexRelation->rd_am->pname; \
98 if (!RegProcedureIsValid(procOid)) \
99 elog(ERROR, "invalid %s regproc", CppAsString(pname)); \
100 fmgr_info_cxt(procOid, procedure, indexRelation->rd_indexcxt); \
104 #define GET_SCAN_PROCEDURE(pname) \
106 procedure = &scan->indexRelation->rd_aminfo->pname; \
107 if (!OidIsValid(procedure->fn_oid)) \
109 RegProcedure procOid = scan->indexRelation->rd_am->pname; \
110 if (!RegProcedureIsValid(procOid)) \
111 elog(ERROR, "invalid %s regproc", CppAsString(pname)); \
112 fmgr_info_cxt(procOid, procedure, scan->indexRelation->rd_indexcxt); \
116 static IndexScanDesc index_beginscan_internal(Relation indexRelation,
117 int nkeys, int norderbys);
120 /* ----------------------------------------------------------------
121 * index_ interface functions
122 * ----------------------------------------------------------------
126 * index_open - open an index relation by relation OID
128 * If lockmode is not "NoLock", the specified kind of lock is
129 * obtained on the index. (Generally, NoLock should only be
130 * used if the caller knows it has some appropriate lock on the
133 * An error is raised if the index does not exist.
135 * This is a convenience routine adapted for indexscan use.
136 * Some callers may prefer to use relation_open directly.
140 index_open(Oid relationId, LOCKMODE lockmode)
144 r = relation_open(relationId, lockmode);
146 if (r->rd_rel->relkind != RELKIND_INDEX)
148 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
149 errmsg("\"%s\" is not an index",
150 RelationGetRelationName(r))));
156 * index_close - close an index relation
158 * If lockmode is not "NoLock", we then release the specified lock.
160 * Note that it is often sensible to hold a lock beyond index_close;
161 * in that case, the lock is released automatically at xact end.
165 index_close(Relation relation, LOCKMODE lockmode)
167 LockRelId relid = relation->rd_lockInfo.lockRelId;
169 Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
171 /* The relcache does the real work... */
172 RelationClose(relation);
174 if (lockmode != NoLock)
175 UnlockRelationId(&relid, lockmode);
179 * index_insert - insert an index tuple into a relation
183 index_insert(Relation indexRelation,
186 ItemPointer heap_t_ctid,
187 Relation heapRelation,
188 IndexUniqueCheck checkUnique)
193 GET_REL_PROCEDURE(aminsert);
196 * have the am's insert proc do all the work.
198 return DatumGetBool(FunctionCall6(procedure,
199 PointerGetDatum(indexRelation),
200 PointerGetDatum(values),
201 PointerGetDatum(isnull),
202 PointerGetDatum(heap_t_ctid),
203 PointerGetDatum(heapRelation),
204 Int32GetDatum((int32) checkUnique)));
208 * index_beginscan - start a scan of an index with amgettuple
210 * Caller must be holding suitable locks on the heap and the index.
213 index_beginscan(Relation heapRelation,
214 Relation indexRelation,
216 int nkeys, int norderbys)
220 scan = index_beginscan_internal(indexRelation, nkeys, norderbys);
223 * Save additional parameters into the scandesc. Everything else was set
224 * up by RelationGetIndexScan.
226 scan->heapRelation = heapRelation;
227 scan->xs_snapshot = snapshot;
233 * index_beginscan_bitmap - start a scan of an index with amgetbitmap
235 * As above, caller had better be holding some lock on the parent heap
236 * relation, even though it's not explicitly mentioned here.
239 index_beginscan_bitmap(Relation indexRelation,
245 scan = index_beginscan_internal(indexRelation, nkeys, 0);
248 * Save additional parameters into the scandesc. Everything else was set
249 * up by RelationGetIndexScan.
251 scan->xs_snapshot = snapshot;
257 * index_beginscan_internal --- common code for index_beginscan variants
260 index_beginscan_internal(Relation indexRelation,
261 int nkeys, int norderbys)
267 GET_REL_PROCEDURE(ambeginscan);
270 * We hold a reference count to the relcache entry throughout the scan.
272 RelationIncrementReferenceCount(indexRelation);
275 * Tell the AM to open a scan.
277 scan = (IndexScanDesc)
278 DatumGetPointer(FunctionCall3(procedure,
279 PointerGetDatum(indexRelation),
280 Int32GetDatum(nkeys),
281 Int32GetDatum(norderbys)));
287 * index_rescan - (re)start a scan of an index
289 * During a restart, the caller may specify a new set of scankeys and/or
290 * orderbykeys; but the number of keys cannot differ from what index_beginscan
291 * was told. (Later we might relax that to "must not exceed", but currently
292 * the index AMs tend to assume that scan->numberOfKeys is what to believe.)
293 * To restart the scan without changing keys, pass NULL for the key arrays.
294 * (Of course, keys *must* be passed on the first call, unless
295 * scan->numberOfKeys is zero.)
299 index_rescan(IndexScanDesc scan,
300 ScanKey keys, int nkeys,
301 ScanKey orderbys, int norderbys)
306 GET_SCAN_PROCEDURE(amrescan);
308 Assert(nkeys == scan->numberOfKeys);
309 Assert(norderbys == scan->numberOfOrderBys);
311 /* Release any held pin on a heap page */
312 if (BufferIsValid(scan->xs_cbuf))
314 ReleaseBuffer(scan->xs_cbuf);
315 scan->xs_cbuf = InvalidBuffer;
318 scan->xs_next_hot = InvalidOffsetNumber;
320 scan->kill_prior_tuple = false; /* for safety */
322 FunctionCall5(procedure,
323 PointerGetDatum(scan),
324 PointerGetDatum(keys),
325 Int32GetDatum(nkeys),
326 PointerGetDatum(orderbys),
327 Int32GetDatum(norderbys));
331 * index_endscan - end a scan
335 index_endscan(IndexScanDesc scan)
340 GET_SCAN_PROCEDURE(amendscan);
342 /* Release any held pin on a heap page */
343 if (BufferIsValid(scan->xs_cbuf))
345 ReleaseBuffer(scan->xs_cbuf);
346 scan->xs_cbuf = InvalidBuffer;
349 /* End the AM's scan */
350 FunctionCall1(procedure, PointerGetDatum(scan));
352 /* Release index refcount acquired by index_beginscan */
353 RelationDecrementReferenceCount(scan->indexRelation);
355 /* Release the scan data structure itself */
360 * index_markpos - mark a scan position
364 index_markpos(IndexScanDesc scan)
369 GET_SCAN_PROCEDURE(ammarkpos);
371 FunctionCall1(procedure, PointerGetDatum(scan));
375 * index_restrpos - restore a scan position
377 * NOTE: this only restores the internal scan state of the index AM.
378 * The current result tuple (scan->xs_ctup) doesn't change. See comments
379 * for ExecRestrPos().
381 * NOTE: in the presence of HOT chains, mark/restore only works correctly
382 * if the scan's snapshot is MVCC-safe; that ensures that there's at most one
383 * returnable tuple in each HOT chain, and so restoring the prior state at the
384 * granularity of the index AM is sufficient. Since the only current user
385 * of mark/restore functionality is nodeMergejoin.c, this effectively means
386 * that merge-join plans only work for MVCC snapshots. This could be fixed
387 * if necessary, but for now it seems unimportant.
391 index_restrpos(IndexScanDesc scan)
395 Assert(IsMVCCSnapshot(scan->xs_snapshot));
398 GET_SCAN_PROCEDURE(amrestrpos);
400 scan->xs_next_hot = InvalidOffsetNumber;
402 scan->kill_prior_tuple = false; /* for safety */
404 FunctionCall1(procedure, PointerGetDatum(scan));
408 * index_getnext - get the next heap tuple from a scan
410 * The result is the next heap tuple satisfying the scan keys and the
411 * snapshot, or NULL if no more matching tuples exist. On success,
412 * the buffer containing the heap tuple is pinned (the pin will be dropped
413 * at the next index_getnext or index_endscan).
415 * Note: caller must check scan->xs_recheck, and perform rechecking of the
416 * scan keys if required. We do not do that here because we don't have
417 * enough information to do it efficiently in the general case.
421 index_getnext(IndexScanDesc scan, ScanDirection direction)
423 HeapTuple heapTuple = &scan->xs_ctup;
424 ItemPointer tid = &heapTuple->t_self;
428 GET_SCAN_PROCEDURE(amgettuple);
430 Assert(TransactionIdIsValid(RecentGlobalXmin));
433 * We always reset xs_hot_dead; if we are here then either we are just
434 * starting the scan, or we previously returned a visible tuple, and in
435 * either case it's inappropriate to kill the prior index entry.
437 scan->xs_hot_dead = false;
445 if (scan->xs_next_hot != InvalidOffsetNumber)
448 * We are resuming scan of a HOT chain after having returned an
449 * earlier member. Must still hold pin on current heap page.
451 Assert(BufferIsValid(scan->xs_cbuf));
452 Assert(ItemPointerGetBlockNumber(tid) ==
453 BufferGetBlockNumber(scan->xs_cbuf));
454 Assert(TransactionIdIsValid(scan->xs_prev_xmax));
455 offnum = scan->xs_next_hot;
456 at_chain_start = false;
457 scan->xs_next_hot = InvalidOffsetNumber;
465 * If we scanned a whole HOT chain and found only dead tuples,
466 * tell index AM to kill its entry for that TID. We do not do this
467 * when in recovery because it may violate MVCC to do so. see
468 * comments in RelationGetIndexScan().
470 if (!scan->xactStartedInRecovery)
471 scan->kill_prior_tuple = scan->xs_hot_dead;
474 * The AM's gettuple proc finds the next index entry matching the
475 * scan keys, and puts the TID in xs_ctup.t_self (ie, *tid). It
476 * should also set scan->xs_recheck, though we pay no attention to
479 found = DatumGetBool(FunctionCall2(procedure,
480 PointerGetDatum(scan),
481 Int32GetDatum(direction)));
483 /* Reset kill flag immediately for safety */
484 scan->kill_prior_tuple = false;
486 /* If we're out of index entries, break out of outer loop */
490 pgstat_count_index_tuples(scan->indexRelation, 1);
492 /* Switch to correct buffer if we don't have it already */
493 prev_buf = scan->xs_cbuf;
494 scan->xs_cbuf = ReleaseAndReadBuffer(scan->xs_cbuf,
496 ItemPointerGetBlockNumber(tid));
499 * Prune page, but only if we weren't already on this page
501 if (prev_buf != scan->xs_cbuf)
502 heap_page_prune_opt(scan->heapRelation, scan->xs_cbuf,
505 /* Prepare to scan HOT chain starting at index-referenced offnum */
506 offnum = ItemPointerGetOffsetNumber(tid);
507 at_chain_start = true;
509 /* We don't know what the first tuple's xmin should be */
510 scan->xs_prev_xmax = InvalidTransactionId;
512 /* Initialize flag to detect if all entries are dead */
513 scan->xs_hot_dead = true;
516 /* Obtain share-lock on the buffer so we can examine visibility */
517 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_SHARE);
519 dp = (Page) BufferGetPage(scan->xs_cbuf);
521 /* Scan through possible multiple members of HOT-chain */
527 /* check for bogus TID */
528 if (offnum < FirstOffsetNumber ||
529 offnum > PageGetMaxOffsetNumber(dp))
532 lp = PageGetItemId(dp, offnum);
534 /* check for unused, dead, or redirected items */
535 if (!ItemIdIsNormal(lp))
537 /* We should only see a redirect at start of chain */
538 if (ItemIdIsRedirected(lp) && at_chain_start)
540 /* Follow the redirect */
541 offnum = ItemIdGetRedirect(lp);
542 at_chain_start = false;
545 /* else must be end of chain */
550 * We must initialize all of *heapTuple (ie, scan->xs_ctup) since
551 * it is returned to the executor on success.
553 heapTuple->t_data = (HeapTupleHeader) PageGetItem(dp, lp);
554 heapTuple->t_len = ItemIdGetLength(lp);
555 ItemPointerSetOffsetNumber(tid, offnum);
556 heapTuple->t_tableOid = RelationGetRelid(scan->heapRelation);
557 ctid = &heapTuple->t_data->t_ctid;
560 * Shouldn't see a HEAP_ONLY tuple at chain start. (This test
561 * should be unnecessary, since the chain root can't be removed
562 * while we have pin on the index entry, but let's make it
565 if (at_chain_start && HeapTupleIsHeapOnly(heapTuple))
569 * The xmin should match the previous xmax value, else chain is
570 * broken. (Note: this test is not optional because it protects
571 * us against the case where the prior chain member's xmax aborted
572 * since we looked at it.)
574 if (TransactionIdIsValid(scan->xs_prev_xmax) &&
575 !TransactionIdEquals(scan->xs_prev_xmax,
576 HeapTupleHeaderGetXmin(heapTuple->t_data)))
579 /* If it's visible per the snapshot, we must return it */
580 if (HeapTupleSatisfiesVisibility(heapTuple, scan->xs_snapshot,
584 * If the snapshot is MVCC, we know that it could accept at
585 * most one member of the HOT chain, so we can skip examining
586 * any more members. Otherwise, check for continuation of the
587 * HOT-chain, and set state for next time.
589 if (IsMVCCSnapshot(scan->xs_snapshot))
590 scan->xs_next_hot = InvalidOffsetNumber;
591 else if (HeapTupleIsHotUpdated(heapTuple))
593 Assert(ItemPointerGetBlockNumber(ctid) ==
594 ItemPointerGetBlockNumber(tid));
595 scan->xs_next_hot = ItemPointerGetOffsetNumber(ctid);
596 scan->xs_prev_xmax = HeapTupleHeaderGetXmax(heapTuple->t_data);
599 scan->xs_next_hot = InvalidOffsetNumber;
601 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_UNLOCK);
603 pgstat_count_heap_fetch(scan->indexRelation);
609 * If we can't see it, maybe no one else can either. Check to see
610 * if the tuple is dead to all transactions. If we find that all
611 * the tuples in the HOT chain are dead, we'll signal the index AM
612 * to not return that TID on future indexscans.
614 if (scan->xs_hot_dead &&
615 HeapTupleSatisfiesVacuum(heapTuple->t_data, RecentGlobalXmin,
616 scan->xs_cbuf) != HEAPTUPLE_DEAD)
617 scan->xs_hot_dead = false;
620 * Check to see if HOT chain continues past this tuple; if so
621 * fetch the next offnum (we don't bother storing it into
622 * xs_next_hot, but must store xs_prev_xmax), and loop around.
624 if (HeapTupleIsHotUpdated(heapTuple))
626 Assert(ItemPointerGetBlockNumber(ctid) ==
627 ItemPointerGetBlockNumber(tid));
628 offnum = ItemPointerGetOffsetNumber(ctid);
629 at_chain_start = false;
630 scan->xs_prev_xmax = HeapTupleHeaderGetXmax(heapTuple->t_data);
633 break; /* end of chain */
634 } /* loop over a single HOT chain */
636 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_UNLOCK);
638 /* Loop around to ask index AM for another TID */
639 scan->xs_next_hot = InvalidOffsetNumber;
642 /* Release any held pin on a heap page */
643 if (BufferIsValid(scan->xs_cbuf))
645 ReleaseBuffer(scan->xs_cbuf);
646 scan->xs_cbuf = InvalidBuffer;
649 return NULL; /* failure exit */
653 * index_getbitmap - get all tuples at once from an index scan
655 * Adds the TIDs of all heap tuples satisfying the scan keys to a bitmap.
656 * Since there's no interlock between the index scan and the eventual heap
657 * access, this is only safe to use with MVCC-based snapshots: the heap
658 * item slot could have been replaced by a newer tuple by the time we get
661 * Returns the number of matching tuples found. (Note: this might be only
662 * approximate, so it should only be used for statistical purposes.)
666 index_getbitmap(IndexScanDesc scan, TIDBitmap *bitmap)
673 GET_SCAN_PROCEDURE(amgetbitmap);
675 /* just make sure this is false... */
676 scan->kill_prior_tuple = false;
679 * have the am's getbitmap proc do all the work.
681 d = FunctionCall2(procedure,
682 PointerGetDatum(scan),
683 PointerGetDatum(bitmap));
685 ntids = DatumGetInt64(d);
687 /* If int8 is pass-by-ref, must free the result to avoid memory leak */
688 #ifndef USE_FLOAT8_BYVAL
689 pfree(DatumGetPointer(d));
692 pgstat_count_index_tuples(scan->indexRelation, ntids);
698 * index_bulk_delete - do mass deletion of index entries
700 * callback routine tells whether a given main-heap tuple is
703 * return value is an optional palloc'd struct of statistics
706 IndexBulkDeleteResult *
707 index_bulk_delete(IndexVacuumInfo *info,
708 IndexBulkDeleteResult *stats,
709 IndexBulkDeleteCallback callback,
710 void *callback_state)
712 Relation indexRelation = info->index;
714 IndexBulkDeleteResult *result;
717 GET_REL_PROCEDURE(ambulkdelete);
719 result = (IndexBulkDeleteResult *)
720 DatumGetPointer(FunctionCall4(procedure,
721 PointerGetDatum(info),
722 PointerGetDatum(stats),
723 PointerGetDatum((Pointer) callback),
724 PointerGetDatum(callback_state)));
730 * index_vacuum_cleanup - do post-deletion cleanup of an index
732 * return value is an optional palloc'd struct of statistics
735 IndexBulkDeleteResult *
736 index_vacuum_cleanup(IndexVacuumInfo *info,
737 IndexBulkDeleteResult *stats)
739 Relation indexRelation = info->index;
741 IndexBulkDeleteResult *result;
744 GET_REL_PROCEDURE(amvacuumcleanup);
746 result = (IndexBulkDeleteResult *)
747 DatumGetPointer(FunctionCall2(procedure,
748 PointerGetDatum(info),
749 PointerGetDatum(stats)));
757 * Index access methods typically require support routines that are
758 * not directly the implementation of any WHERE-clause query operator
759 * and so cannot be kept in pg_amop. Instead, such routines are kept
760 * in pg_amproc. These registered procedure OIDs are assigned numbers
761 * according to a convention established by the access method.
762 * The general index code doesn't know anything about the routines
763 * involved; it just builds an ordered list of them for
764 * each attribute on which an index is defined.
766 * As of Postgres 8.3, support routines within an operator family
767 * are further subdivided by the "left type" and "right type" of the
768 * query operator(s) that they support. The "default" functions for a
769 * particular indexed attribute are those with both types equal to
770 * the index opclass' opcintype (note that this is subtly different
771 * from the indexed attribute's own type: it may be a binary-compatible
772 * type instead). Only the default functions are stored in relcache
773 * entries --- access methods can use the syscache to look up non-default
776 * This routine returns the requested default procedure OID for a
777 * particular indexed attribute.
781 index_getprocid(Relation irel,
789 nproc = irel->rd_am->amsupport;
791 Assert(procnum > 0 && procnum <= (uint16) nproc);
793 procindex = (nproc * (attnum - 1)) + (procnum - 1);
795 loc = irel->rd_support;
799 return loc[procindex];
805 * This routine allows index AMs to keep fmgr lookup info for
806 * support procs in the relcache. As above, only the "default"
807 * functions for any particular indexed attribute are cached.
809 * Note: the return value points into cached data that will be lost during
810 * any relcache rebuild! Therefore, either use the callinfo right away,
811 * or save it only after having acquired some type of lock on the index rel.
815 index_getprocinfo(Relation irel,
823 nproc = irel->rd_am->amsupport;
825 Assert(procnum > 0 && procnum <= (uint16) nproc);
827 procindex = (nproc * (attnum - 1)) + (procnum - 1);
829 locinfo = irel->rd_supportinfo;
831 Assert(locinfo != NULL);
833 locinfo += procindex;
835 /* Initialize the lookup info if first time through */
836 if (locinfo->fn_oid == InvalidOid)
838 RegProcedure *loc = irel->rd_support;
843 procId = loc[procindex];
846 * Complain if function was not found during IndexSupportInitialize.
847 * This should not happen unless the system tables contain bogus
848 * entries for the index opclass. (If an AM wants to allow a support
849 * function to be optional, it can use index_getprocid.)
851 if (!RegProcedureIsValid(procId))
852 elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
853 procnum, attnum, RelationGetRelationName(irel));
855 fmgr_info_cxt(procId, locinfo, irel->rd_indexcxt);