1 /*-------------------------------------------------------------------------
4 * general index access method routines
6 * Portions Copyright (c) 1996-2008, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * $PostgreSQL: pgsql/src/backend/access/index/indexam.c,v 1.109 2008/06/19 00:46:03 alvherre Exp $
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, ScanKey key);
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 bool check_uniqueness)
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 BoolGetDatum(check_uniqueness)));
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, ScanKey key)
220 scan = index_beginscan_internal(indexRelation, nkeys, key);
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,
241 int nkeys, ScanKey key)
245 scan = index_beginscan_internal(indexRelation, nkeys, key);
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, ScanKey key)
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 PointerGetDatum(key)));
287 * index_rescan - (re)start a scan of an index
289 * The caller may specify a new set of scankeys (but the number of keys
290 * cannot change). To restart the scan without changing keys, pass NULL
293 * Note that this is also called when first starting an indexscan;
294 * see RelationGetIndexScan. Keys *must* be passed in that case,
295 * unless scan->numberOfKeys is zero.
299 index_rescan(IndexScanDesc scan, ScanKey key)
304 GET_SCAN_PROCEDURE(amrescan);
306 /* Release any held pin on a heap page */
307 if (BufferIsValid(scan->xs_cbuf))
309 ReleaseBuffer(scan->xs_cbuf);
310 scan->xs_cbuf = InvalidBuffer;
313 scan->xs_next_hot = InvalidOffsetNumber;
315 scan->kill_prior_tuple = false; /* for safety */
317 FunctionCall2(procedure,
318 PointerGetDatum(scan),
319 PointerGetDatum(key));
323 * index_endscan - end a scan
327 index_endscan(IndexScanDesc scan)
332 GET_SCAN_PROCEDURE(amendscan);
334 /* Release any held pin on a heap page */
335 if (BufferIsValid(scan->xs_cbuf))
337 ReleaseBuffer(scan->xs_cbuf);
338 scan->xs_cbuf = InvalidBuffer;
341 /* End the AM's scan */
342 FunctionCall1(procedure, PointerGetDatum(scan));
344 /* Release index refcount acquired by index_beginscan */
345 RelationDecrementReferenceCount(scan->indexRelation);
347 /* Release the scan data structure itself */
352 * index_markpos - mark a scan position
356 index_markpos(IndexScanDesc scan)
361 GET_SCAN_PROCEDURE(ammarkpos);
363 FunctionCall1(procedure, PointerGetDatum(scan));
367 * index_restrpos - restore a scan position
369 * NOTE: this only restores the internal scan state of the index AM.
370 * The current result tuple (scan->xs_ctup) doesn't change. See comments
371 * for ExecRestrPos().
373 * NOTE: in the presence of HOT chains, mark/restore only works correctly
374 * if the scan's snapshot is MVCC-safe; that ensures that there's at most one
375 * returnable tuple in each HOT chain, and so restoring the prior state at the
376 * granularity of the index AM is sufficient. Since the only current user
377 * of mark/restore functionality is nodeMergejoin.c, this effectively means
378 * that merge-join plans only work for MVCC snapshots. This could be fixed
379 * if necessary, but for now it seems unimportant.
383 index_restrpos(IndexScanDesc scan)
387 Assert(IsMVCCSnapshot(scan->xs_snapshot));
390 GET_SCAN_PROCEDURE(amrestrpos);
392 scan->xs_next_hot = InvalidOffsetNumber;
394 scan->kill_prior_tuple = false; /* for safety */
396 FunctionCall1(procedure, PointerGetDatum(scan));
400 * index_getnext - get the next heap tuple from a scan
402 * The result is the next heap tuple satisfying the scan keys and the
403 * snapshot, or NULL if no more matching tuples exist. On success,
404 * the buffer containing the heap tuple is pinned (the pin will be dropped
405 * at the next index_getnext or index_endscan).
407 * Note: caller must check scan->xs_recheck, and perform rechecking of the
408 * scan keys if required. We do not do that here because we don't have
409 * enough information to do it efficiently in the general case.
413 index_getnext(IndexScanDesc scan, ScanDirection direction)
415 HeapTuple heapTuple = &scan->xs_ctup;
416 ItemPointer tid = &heapTuple->t_self;
420 GET_SCAN_PROCEDURE(amgettuple);
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.
427 scan->xs_hot_dead = false;
435 if (scan->xs_next_hot != InvalidOffsetNumber)
438 * We are resuming scan of a HOT chain after having returned an
439 * earlier member. Must still hold pin on current heap page.
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;
455 * If we scanned a whole HOT chain and found only dead tuples,
456 * tell index AM to kill its entry for that TID.
458 scan->kill_prior_tuple = scan->xs_hot_dead;
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 * It should also set scan->xs_recheck, though we pay no
464 * attention to that here.
466 found = DatumGetBool(FunctionCall2(procedure,
467 PointerGetDatum(scan),
468 Int32GetDatum(direction)));
470 /* Reset kill flag immediately for safety */
471 scan->kill_prior_tuple = false;
473 /* If we're out of index entries, break out of outer loop */
477 pgstat_count_index_tuples(scan->indexRelation, 1);
479 /* Switch to correct buffer if we don't have it already */
480 prev_buf = scan->xs_cbuf;
481 scan->xs_cbuf = ReleaseAndReadBuffer(scan->xs_cbuf,
483 ItemPointerGetBlockNumber(tid));
486 * Prune page, but only if we weren't already on this page
488 if (prev_buf != scan->xs_cbuf)
489 heap_page_prune_opt(scan->heapRelation, scan->xs_cbuf,
492 /* Prepare to scan HOT chain starting at index-referenced offnum */
493 offnum = ItemPointerGetOffsetNumber(tid);
494 at_chain_start = true;
496 /* We don't know what the first tuple's xmin should be */
497 scan->xs_prev_xmax = InvalidTransactionId;
499 /* Initialize flag to detect if all entries are dead */
500 scan->xs_hot_dead = true;
503 /* Obtain share-lock on the buffer so we can examine visibility */
504 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_SHARE);
506 dp = (Page) BufferGetPage(scan->xs_cbuf);
508 /* Scan through possible multiple members of HOT-chain */
514 /* check for bogus TID */
515 if (offnum < FirstOffsetNumber ||
516 offnum > PageGetMaxOffsetNumber(dp))
519 lp = PageGetItemId(dp, offnum);
521 /* check for unused, dead, or redirected items */
522 if (!ItemIdIsNormal(lp))
524 /* We should only see a redirect at start of chain */
525 if (ItemIdIsRedirected(lp) && at_chain_start)
527 /* Follow the redirect */
528 offnum = ItemIdGetRedirect(lp);
529 at_chain_start = false;
532 /* else must be end of chain */
537 * We must initialize all of *heapTuple (ie, scan->xs_ctup) since
538 * it is returned to the executor on success.
540 heapTuple->t_data = (HeapTupleHeader) PageGetItem(dp, lp);
541 heapTuple->t_len = ItemIdGetLength(lp);
542 ItemPointerSetOffsetNumber(tid, offnum);
543 heapTuple->t_tableOid = RelationGetRelid(scan->heapRelation);
544 ctid = &heapTuple->t_data->t_ctid;
547 * Shouldn't see a HEAP_ONLY tuple at chain start. (This test
548 * should be unnecessary, since the chain root can't be removed
549 * while we have pin on the index entry, but let's make it
552 if (at_chain_start && HeapTupleIsHeapOnly(heapTuple))
556 * The xmin should match the previous xmax value, else chain is
557 * broken. (Note: this test is not optional because it protects
558 * us against the case where the prior chain member's xmax aborted
559 * since we looked at it.)
561 if (TransactionIdIsValid(scan->xs_prev_xmax) &&
562 !TransactionIdEquals(scan->xs_prev_xmax,
563 HeapTupleHeaderGetXmin(heapTuple->t_data)))
566 /* If it's visible per the snapshot, we must return it */
567 if (HeapTupleSatisfiesVisibility(heapTuple, scan->xs_snapshot,
571 * If the snapshot is MVCC, we know that it could accept at
572 * most one member of the HOT chain, so we can skip examining
573 * any more members. Otherwise, check for continuation of the
574 * HOT-chain, and set state for next time.
576 if (IsMVCCSnapshot(scan->xs_snapshot))
577 scan->xs_next_hot = InvalidOffsetNumber;
578 else if (HeapTupleIsHotUpdated(heapTuple))
580 Assert(ItemPointerGetBlockNumber(ctid) ==
581 ItemPointerGetBlockNumber(tid));
582 scan->xs_next_hot = ItemPointerGetOffsetNumber(ctid);
583 scan->xs_prev_xmax = HeapTupleHeaderGetXmax(heapTuple->t_data);
586 scan->xs_next_hot = InvalidOffsetNumber;
588 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_UNLOCK);
590 pgstat_count_heap_fetch(scan->indexRelation);
596 * If we can't see it, maybe no one else can either. Check to see
597 * if the tuple is dead to all transactions. If we find that all
598 * the tuples in the HOT chain are dead, we'll signal the index AM
599 * to not return that TID on future indexscans.
601 if (scan->xs_hot_dead &&
602 HeapTupleSatisfiesVacuum(heapTuple->t_data, RecentGlobalXmin,
603 scan->xs_cbuf) != HEAPTUPLE_DEAD)
604 scan->xs_hot_dead = false;
607 * Check to see if HOT chain continues past this tuple; if so
608 * fetch the next offnum (we don't bother storing it into
609 * xs_next_hot, but must store xs_prev_xmax), and loop around.
611 if (HeapTupleIsHotUpdated(heapTuple))
613 Assert(ItemPointerGetBlockNumber(ctid) ==
614 ItemPointerGetBlockNumber(tid));
615 offnum = ItemPointerGetOffsetNumber(ctid);
616 at_chain_start = false;
617 scan->xs_prev_xmax = HeapTupleHeaderGetXmax(heapTuple->t_data);
620 break; /* end of chain */
621 } /* loop over a single HOT chain */
623 LockBuffer(scan->xs_cbuf, BUFFER_LOCK_UNLOCK);
625 /* Loop around to ask index AM for another TID */
626 scan->xs_next_hot = InvalidOffsetNumber;
629 /* Release any held pin on a heap page */
630 if (BufferIsValid(scan->xs_cbuf))
632 ReleaseBuffer(scan->xs_cbuf);
633 scan->xs_cbuf = InvalidBuffer;
636 return NULL; /* failure exit */
640 * index_getbitmap - get all tuples at once from an index scan
642 * Adds the TIDs of all heap tuples satisfying the scan keys to a bitmap.
643 * Since there's no interlock between the index scan and the eventual heap
644 * access, this is only safe to use with MVCC-based snapshots: the heap
645 * item slot could have been replaced by a newer tuple by the time we get
648 * Returns the number of matching tuples found.
652 index_getbitmap(IndexScanDesc scan, TIDBitmap *bitmap)
658 GET_SCAN_PROCEDURE(amgetbitmap);
660 /* just make sure this is false... */
661 scan->kill_prior_tuple = false;
664 * have the am's getbitmap proc do all the work.
666 ntids = DatumGetInt64(FunctionCall2(procedure,
667 PointerGetDatum(scan),
668 PointerGetDatum(bitmap)));
670 pgstat_count_index_tuples(scan->indexRelation, ntids);
676 * index_bulk_delete - do mass deletion of index entries
678 * callback routine tells whether a given main-heap tuple is
681 * return value is an optional palloc'd struct of statistics
684 IndexBulkDeleteResult *
685 index_bulk_delete(IndexVacuumInfo *info,
686 IndexBulkDeleteResult *stats,
687 IndexBulkDeleteCallback callback,
688 void *callback_state)
690 Relation indexRelation = info->index;
692 IndexBulkDeleteResult *result;
695 GET_REL_PROCEDURE(ambulkdelete);
697 result = (IndexBulkDeleteResult *)
698 DatumGetPointer(FunctionCall4(procedure,
699 PointerGetDatum(info),
700 PointerGetDatum(stats),
701 PointerGetDatum((Pointer) callback),
702 PointerGetDatum(callback_state)));
708 * index_vacuum_cleanup - do post-deletion cleanup of an index
710 * return value is an optional palloc'd struct of statistics
713 IndexBulkDeleteResult *
714 index_vacuum_cleanup(IndexVacuumInfo *info,
715 IndexBulkDeleteResult *stats)
717 Relation indexRelation = info->index;
719 IndexBulkDeleteResult *result;
722 GET_REL_PROCEDURE(amvacuumcleanup);
724 result = (IndexBulkDeleteResult *)
725 DatumGetPointer(FunctionCall2(procedure,
726 PointerGetDatum(info),
727 PointerGetDatum(stats)));
735 * Index access methods typically require support routines that are
736 * not directly the implementation of any WHERE-clause query operator
737 * and so cannot be kept in pg_amop. Instead, such routines are kept
738 * in pg_amproc. These registered procedure OIDs are assigned numbers
739 * according to a convention established by the access method.
740 * The general index code doesn't know anything about the routines
741 * involved; it just builds an ordered list of them for
742 * each attribute on which an index is defined.
744 * As of Postgres 8.3, support routines within an operator family
745 * are further subdivided by the "left type" and "right type" of the
746 * query operator(s) that they support. The "default" functions for a
747 * particular indexed attribute are those with both types equal to
748 * the index opclass' opcintype (note that this is subtly different
749 * from the indexed attribute's own type: it may be a binary-compatible
750 * type instead). Only the default functions are stored in relcache
751 * entries --- access methods can use the syscache to look up non-default
754 * This routine returns the requested default procedure OID for a
755 * particular indexed attribute.
759 index_getprocid(Relation irel,
767 nproc = irel->rd_am->amsupport;
769 Assert(procnum > 0 && procnum <= (uint16) nproc);
771 procindex = (nproc * (attnum - 1)) + (procnum - 1);
773 loc = irel->rd_support;
777 return loc[procindex];
783 * This routine allows index AMs to keep fmgr lookup info for
784 * support procs in the relcache. As above, only the "default"
785 * functions for any particular indexed attribute are cached.
787 * Note: the return value points into cached data that will be lost during
788 * any relcache rebuild! Therefore, either use the callinfo right away,
789 * or save it only after having acquired some type of lock on the index rel.
793 index_getprocinfo(Relation irel,
801 nproc = irel->rd_am->amsupport;
803 Assert(procnum > 0 && procnum <= (uint16) nproc);
805 procindex = (nproc * (attnum - 1)) + (procnum - 1);
807 locinfo = irel->rd_supportinfo;
809 Assert(locinfo != NULL);
811 locinfo += procindex;
813 /* Initialize the lookup info if first time through */
814 if (locinfo->fn_oid == InvalidOid)
816 RegProcedure *loc = irel->rd_support;
821 procId = loc[procindex];
824 * Complain if function was not found during IndexSupportInitialize.
825 * This should not happen unless the system tables contain bogus
826 * entries for the index opclass. (If an AM wants to allow a support
827 * function to be optional, it can use index_getprocid.)
829 if (!RegProcedureIsValid(procId))
830 elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
831 procnum, attnum, RelationGetRelationName(irel));
833 fmgr_info_cxt(procId, locinfo, irel->rd_indexcxt);