1 /*-------------------------------------------------------------------------
2  *
3  * indexam.c
4  *	  general index access method routines
5  *
6  * Portions Copyright (c) 1996-2017, 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_parallelscan_estimate - estimate shared memory for parallel scan
24  *		index_parallelscan_initialize - initialize parallel scan
25  *		index_parallelrescan  - (re)start a parallel scan of an index
26  *		index_beginscan_parallel - join parallel index scan
27  *		index_getnext_tid	- get the next TID from a scan
28  *		index_fetch_heap		- get the scan's next heap tuple
29  *		index_getnext	- get the next heap tuple from a scan
30  *		index_getbitmap - get all tuples from a scan
31  *		index_bulk_delete	- bulk deletion of index tuples
32  *		index_vacuum_cleanup	- post-deletion cleanup of an index
33  *		index_can_return	- does index support index-only scans?
34  *		index_getprocid - get a support procedure OID
35  *		index_getprocinfo - get a support procedure's lookup info
36  *
37  * NOTES
38  *		This file contains the index_ routines which used
39  *		to be a scattered collection of stuff in access/genam.
40  *
41  *
42  * old comments
43  *		Scans are implemented as follows:
44  *
45  *		`0' represents an invalid item pointer.
46  *		`-' represents an unknown item pointer.
47  *		`X' represents a known item pointers.
48  *		`+' represents known or invalid item pointers.
49  *		`*' represents any item pointers.
50  *
51  *		State is represented by a triple of these symbols in the order of
52  *		previous, current, next.  Note that the case of reverse scans works
53  *		identically.
54  *
55  *				State	Result
56  *		(1)		+ + -	+ 0 0			(if the next item pointer is invalid)
57  *		(2)				+ X -			(otherwise)
58  *		(3)		* 0 0	* 0 0			(no change)
59  *		(4)		+ X 0	X 0 0			(shift)
60  *		(5)		* + X	+ X -			(shift, add unknown)
61  *
62  *		All other states cannot occur.
63  *
64  *		Note: It would be possible to cache the status of the previous and
65  *			  next item pointer using the flags.
66  *
67  *-------------------------------------------------------------------------
68  */
69 
70 #include "postgres.h"
71 
72 #include "access/amapi.h"
73 #include "access/relscan.h"
74 #include "access/transam.h"
75 #include "access/xlog.h"
76 #include "catalog/catalog.h"
77 #include "catalog/index.h"
78 #include "pgstat.h"
79 #include "storage/bufmgr.h"
80 #include "storage/lmgr.h"
81 #include "storage/predicate.h"
82 #include "utils/snapmgr.h"
83 #include "utils/tqual.h"
84 
85 
86 /* ----------------------------------------------------------------
87  *					macros used in index_ routines
88  *
89  * Note: the ReindexIsProcessingIndex() check in RELATION_CHECKS is there
90  * to check that we don't try to scan or do retail insertions into an index
91  * that is currently being rebuilt or pending rebuild.  This helps to catch
92  * things that don't work when reindexing system catalogs.  The assertion
93  * doesn't prevent the actual rebuild because we don't use RELATION_CHECKS
94  * when calling the index AM's ambuild routine, and there is no reason for
95  * ambuild to call its subsidiary routines through this file.
96  * ----------------------------------------------------------------
97  */
98 #define RELATION_CHECKS \
99 ( \
100 	AssertMacro(RelationIsValid(indexRelation)), \
101 	AssertMacro(PointerIsValid(indexRelation->rd_amroutine)), \
102 	AssertMacro(!ReindexIsProcessingIndex(RelationGetRelid(indexRelation))) \
103 )
104 
105 #define SCAN_CHECKS \
106 ( \
107 	AssertMacro(IndexScanIsValid(scan)), \
108 	AssertMacro(RelationIsValid(scan->indexRelation)), \
109 	AssertMacro(PointerIsValid(scan->indexRelation->rd_amroutine)) \
110 )
111 
112 #define CHECK_REL_PROCEDURE(pname) \
113 do { \
114 	if (indexRelation->rd_amroutine->pname == NULL) \
115 		elog(ERROR, "function %s is not defined for index %s", \
116 			 CppAsString(pname), RelationGetRelationName(indexRelation)); \
117 } while(0)
118 
119 #define CHECK_SCAN_PROCEDURE(pname) \
120 do { \
121 	if (scan->indexRelation->rd_amroutine->pname == NULL) \
122 		elog(ERROR, "function %s is not defined for index %s", \
123 			 CppAsString(pname), RelationGetRelationName(scan->indexRelation)); \
124 } while(0)
125 
126 static IndexScanDesc index_beginscan_internal(Relation indexRelation,
127 						 int nkeys, int norderbys, Snapshot snapshot,
128 						 ParallelIndexScanDesc pscan, bool temp_snap);
129 
130 
131 /* ----------------------------------------------------------------
132  *				   index_ interface functions
133  * ----------------------------------------------------------------
134  */
135 
136 /* ----------------
137  *		index_open - open an index relation by relation OID
138  *
139  *		If lockmode is not "NoLock", the specified kind of lock is
140  *		obtained on the index.  (Generally, NoLock should only be
141  *		used if the caller knows it has some appropriate lock on the
142  *		index already.)
143  *
144  *		An error is raised if the index does not exist.
145  *
146  *		This is a convenience routine adapted for indexscan use.
147  *		Some callers may prefer to use relation_open directly.
148  * ----------------
149  */
150 Relation
index_open(Oid relationId,LOCKMODE lockmode)151 index_open(Oid relationId, LOCKMODE lockmode)
152 {
153 	Relation	r;
154 
155 	r = relation_open(relationId, lockmode);
156 
157 	if (r->rd_rel->relkind != RELKIND_INDEX)
158 		ereport(ERROR,
159 				(errcode(ERRCODE_WRONG_OBJECT_TYPE),
160 				 errmsg("\"%s\" is not an index",
161 						RelationGetRelationName(r))));
162 
163 	return r;
164 }
165 
166 /* ----------------
167  *		index_close - close an index relation
168  *
169  *		If lockmode is not "NoLock", we then release the specified lock.
170  *
171  *		Note that it is often sensible to hold a lock beyond index_close;
172  *		in that case, the lock is released automatically at xact end.
173  * ----------------
174  */
175 void
index_close(Relation relation,LOCKMODE lockmode)176 index_close(Relation relation, LOCKMODE lockmode)
177 {
178 	LockRelId	relid = relation->rd_lockInfo.lockRelId;
179 
180 	Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
181 
182 	/* The relcache does the real work... */
183 	RelationClose(relation);
184 
185 	if (lockmode != NoLock)
186 		UnlockRelationId(&relid, lockmode);
187 }
188 
189 /* ----------------
190  *		index_insert - insert an index tuple into a relation
191  * ----------------
192  */
193 bool
index_insert(Relation indexRelation,Datum * values,bool * isnull,ItemPointer heap_t_ctid,Relation heapRelation,IndexUniqueCheck checkUnique,IndexInfo * indexInfo)194 index_insert(Relation indexRelation,
195 			 Datum *values,
196 			 bool *isnull,
197 			 ItemPointer heap_t_ctid,
198 			 Relation heapRelation,
199 			 IndexUniqueCheck checkUnique,
200 			 IndexInfo *indexInfo)
201 {
202 	RELATION_CHECKS;
203 	CHECK_REL_PROCEDURE(aminsert);
204 
205 	if (!(indexRelation->rd_amroutine->ampredlocks))
206 		CheckForSerializableConflictIn(indexRelation,
207 									   (HeapTuple) NULL,
208 									   InvalidBuffer);
209 
210 	return indexRelation->rd_amroutine->aminsert(indexRelation, values, isnull,
211 												 heap_t_ctid, heapRelation,
212 												 checkUnique, indexInfo);
213 }
214 
215 /*
216  * index_beginscan - start a scan of an index with amgettuple
217  *
218  * Caller must be holding suitable locks on the heap and the index.
219  */
220 IndexScanDesc
index_beginscan(Relation heapRelation,Relation indexRelation,Snapshot snapshot,int nkeys,int norderbys)221 index_beginscan(Relation heapRelation,
222 				Relation indexRelation,
223 				Snapshot snapshot,
224 				int nkeys, int norderbys)
225 {
226 	IndexScanDesc scan;
227 
228 	scan = index_beginscan_internal(indexRelation, nkeys, norderbys, snapshot, NULL, false);
229 
230 	/*
231 	 * Save additional parameters into the scandesc.  Everything else was set
232 	 * up by RelationGetIndexScan.
233 	 */
234 	scan->heapRelation = heapRelation;
235 	scan->xs_snapshot = snapshot;
236 
237 	return scan;
238 }
239 
240 /*
241  * index_beginscan_bitmap - start a scan of an index with amgetbitmap
242  *
243  * As above, caller had better be holding some lock on the parent heap
244  * relation, even though it's not explicitly mentioned here.
245  */
246 IndexScanDesc
index_beginscan_bitmap(Relation indexRelation,Snapshot snapshot,int nkeys)247 index_beginscan_bitmap(Relation indexRelation,
248 					   Snapshot snapshot,
249 					   int nkeys)
250 {
251 	IndexScanDesc scan;
252 
253 	scan = index_beginscan_internal(indexRelation, nkeys, 0, snapshot, NULL, false);
254 
255 	/*
256 	 * Save additional parameters into the scandesc.  Everything else was set
257 	 * up by RelationGetIndexScan.
258 	 */
259 	scan->xs_snapshot = snapshot;
260 
261 	return scan;
262 }
263 
264 /*
265  * index_beginscan_internal --- common code for index_beginscan variants
266  */
267 static IndexScanDesc
index_beginscan_internal(Relation indexRelation,int nkeys,int norderbys,Snapshot snapshot,ParallelIndexScanDesc pscan,bool temp_snap)268 index_beginscan_internal(Relation indexRelation,
269 						 int nkeys, int norderbys, Snapshot snapshot,
270 						 ParallelIndexScanDesc pscan, bool temp_snap)
271 {
272 	IndexScanDesc scan;
273 
274 	RELATION_CHECKS;
275 	CHECK_REL_PROCEDURE(ambeginscan);
276 
277 	if (!(indexRelation->rd_amroutine->ampredlocks))
278 		PredicateLockRelation(indexRelation, snapshot);
279 
280 	/*
281 	 * We hold a reference count to the relcache entry throughout the scan.
282 	 */
283 	RelationIncrementReferenceCount(indexRelation);
284 
285 	/*
286 	 * Tell the AM to open a scan.
287 	 */
288 	scan = indexRelation->rd_amroutine->ambeginscan(indexRelation, nkeys,
289 													norderbys);
290 	/* Initialize information for parallel scan. */
291 	scan->parallel_scan = pscan;
292 	scan->xs_temp_snap = temp_snap;
293 
294 	return scan;
295 }
296 
297 /* ----------------
298  *		index_rescan  - (re)start a scan of an index
299  *
300  * During a restart, the caller may specify a new set of scankeys and/or
301  * orderbykeys; but the number of keys cannot differ from what index_beginscan
302  * was told.  (Later we might relax that to "must not exceed", but currently
303  * the index AMs tend to assume that scan->numberOfKeys is what to believe.)
304  * To restart the scan without changing keys, pass NULL for the key arrays.
305  * (Of course, keys *must* be passed on the first call, unless
306  * scan->numberOfKeys is zero.)
307  * ----------------
308  */
309 void
index_rescan(IndexScanDesc scan,ScanKey keys,int nkeys,ScanKey orderbys,int norderbys)310 index_rescan(IndexScanDesc scan,
311 			 ScanKey keys, int nkeys,
312 			 ScanKey orderbys, int norderbys)
313 {
314 	SCAN_CHECKS;
315 	CHECK_SCAN_PROCEDURE(amrescan);
316 
317 	Assert(nkeys == scan->numberOfKeys);
318 	Assert(norderbys == scan->numberOfOrderBys);
319 
320 	/* Release any held pin on a heap page */
321 	if (BufferIsValid(scan->xs_cbuf))
322 	{
323 		ReleaseBuffer(scan->xs_cbuf);
324 		scan->xs_cbuf = InvalidBuffer;
325 	}
326 
327 	scan->xs_continue_hot = false;
328 
329 	scan->kill_prior_tuple = false; /* for safety */
330 
331 	scan->indexRelation->rd_amroutine->amrescan(scan, keys, nkeys,
332 												orderbys, norderbys);
333 }
334 
335 /* ----------------
336  *		index_endscan - end a scan
337  * ----------------
338  */
339 void
index_endscan(IndexScanDesc scan)340 index_endscan(IndexScanDesc scan)
341 {
342 	SCAN_CHECKS;
343 	CHECK_SCAN_PROCEDURE(amendscan);
344 
345 	/* Release any held pin on a heap page */
346 	if (BufferIsValid(scan->xs_cbuf))
347 	{
348 		ReleaseBuffer(scan->xs_cbuf);
349 		scan->xs_cbuf = InvalidBuffer;
350 	}
351 
352 	/* End the AM's scan */
353 	scan->indexRelation->rd_amroutine->amendscan(scan);
354 
355 	/* Release index refcount acquired by index_beginscan */
356 	RelationDecrementReferenceCount(scan->indexRelation);
357 
358 	if (scan->xs_temp_snap)
359 		UnregisterSnapshot(scan->xs_snapshot);
360 
361 	/* Release the scan data structure itself */
362 	IndexScanEnd(scan);
363 }
364 
365 /* ----------------
366  *		index_markpos  - mark a scan position
367  * ----------------
368  */
369 void
index_markpos(IndexScanDesc scan)370 index_markpos(IndexScanDesc scan)
371 {
372 	SCAN_CHECKS;
373 	CHECK_SCAN_PROCEDURE(ammarkpos);
374 
375 	scan->indexRelation->rd_amroutine->ammarkpos(scan);
376 }
377 
378 /* ----------------
379  *		index_restrpos	- restore a scan position
380  *
381  * NOTE: this only restores the internal scan state of the index AM.
382  * The current result tuple (scan->xs_ctup) doesn't change.  See comments
383  * for ExecRestrPos().
384  *
385  * NOTE: in the presence of HOT chains, mark/restore only works correctly
386  * if the scan's snapshot is MVCC-safe; that ensures that there's at most one
387  * returnable tuple in each HOT chain, and so restoring the prior state at the
388  * granularity of the index AM is sufficient.  Since the only current user
389  * of mark/restore functionality is nodeMergejoin.c, this effectively means
390  * that merge-join plans only work for MVCC snapshots.  This could be fixed
391  * if necessary, but for now it seems unimportant.
392  * ----------------
393  */
394 void
index_restrpos(IndexScanDesc scan)395 index_restrpos(IndexScanDesc scan)
396 {
397 	Assert(IsMVCCSnapshot(scan->xs_snapshot));
398 
399 	SCAN_CHECKS;
400 	CHECK_SCAN_PROCEDURE(amrestrpos);
401 
402 	scan->xs_continue_hot = false;
403 
404 	scan->kill_prior_tuple = false; /* for safety */
405 
406 	scan->indexRelation->rd_amroutine->amrestrpos(scan);
407 }
408 
409 /*
410  * index_parallelscan_estimate - estimate shared memory for parallel scan
411  *
412  * Currently, we don't pass any information to the AM-specific estimator,
413  * so it can probably only return a constant.  In the future, we might need
414  * to pass more information.
415  */
416 Size
index_parallelscan_estimate(Relation indexRelation,Snapshot snapshot)417 index_parallelscan_estimate(Relation indexRelation, Snapshot snapshot)
418 {
419 	Size		nbytes;
420 
421 	RELATION_CHECKS;
422 
423 	nbytes = offsetof(ParallelIndexScanDescData, ps_snapshot_data);
424 	nbytes = add_size(nbytes, EstimateSnapshotSpace(snapshot));
425 	nbytes = MAXALIGN(nbytes);
426 
427 	/*
428 	 * If amestimateparallelscan is not provided, assume there is no
429 	 * AM-specific data needed.  (It's hard to believe that could work, but
430 	 * it's easy enough to cater to it here.)
431 	 */
432 	if (indexRelation->rd_amroutine->amestimateparallelscan != NULL)
433 		nbytes = add_size(nbytes,
434 						  indexRelation->rd_amroutine->amestimateparallelscan());
435 
436 	return nbytes;
437 }
438 
439 /*
440  * index_parallelscan_initialize - initialize parallel scan
441  *
442  * We initialize both the ParallelIndexScanDesc proper and the AM-specific
443  * information which follows it.
444  *
445  * This function calls access method specific initialization routine to
446  * initialize am specific information.  Call this just once in the leader
447  * process; then, individual workers attach via index_beginscan_parallel.
448  */
449 void
index_parallelscan_initialize(Relation heapRelation,Relation indexRelation,Snapshot snapshot,ParallelIndexScanDesc target)450 index_parallelscan_initialize(Relation heapRelation, Relation indexRelation,
451 							  Snapshot snapshot, ParallelIndexScanDesc target)
452 {
453 	Size		offset;
454 
455 	RELATION_CHECKS;
456 
457 	offset = add_size(offsetof(ParallelIndexScanDescData, ps_snapshot_data),
458 					  EstimateSnapshotSpace(snapshot));
459 	offset = MAXALIGN(offset);
460 
461 	target->ps_relid = RelationGetRelid(heapRelation);
462 	target->ps_indexid = RelationGetRelid(indexRelation);
463 	target->ps_offset = offset;
464 	SerializeSnapshot(snapshot, target->ps_snapshot_data);
465 
466 	/* aminitparallelscan is optional; assume no-op if not provided by AM */
467 	if (indexRelation->rd_amroutine->aminitparallelscan != NULL)
468 	{
469 		void	   *amtarget;
470 
471 		amtarget = OffsetToPointer(target, offset);
472 		indexRelation->rd_amroutine->aminitparallelscan(amtarget);
473 	}
474 }
475 
476 /* ----------------
477  *		index_parallelrescan  - (re)start a parallel scan of an index
478  * ----------------
479  */
480 void
index_parallelrescan(IndexScanDesc scan)481 index_parallelrescan(IndexScanDesc scan)
482 {
483 	SCAN_CHECKS;
484 
485 	/* amparallelrescan is optional; assume no-op if not provided by AM */
486 	if (scan->indexRelation->rd_amroutine->amparallelrescan != NULL)
487 		scan->indexRelation->rd_amroutine->amparallelrescan(scan);
488 }
489 
490 /*
491  * index_beginscan_parallel - join parallel index scan
492  *
493  * Caller must be holding suitable locks on the heap and the index.
494  */
495 IndexScanDesc
index_beginscan_parallel(Relation heaprel,Relation indexrel,int nkeys,int norderbys,ParallelIndexScanDesc pscan)496 index_beginscan_parallel(Relation heaprel, Relation indexrel, int nkeys,
497 						 int norderbys, ParallelIndexScanDesc pscan)
498 {
499 	Snapshot	snapshot;
500 	IndexScanDesc scan;
501 
502 	Assert(RelationGetRelid(heaprel) == pscan->ps_relid);
503 	snapshot = RestoreSnapshot(pscan->ps_snapshot_data);
504 	RegisterSnapshot(snapshot);
505 	scan = index_beginscan_internal(indexrel, nkeys, norderbys, snapshot,
506 									pscan, true);
507 
508 	/*
509 	 * Save additional parameters into the scandesc.  Everything else was set
510 	 * up by index_beginscan_internal.
511 	 */
512 	scan->heapRelation = heaprel;
513 	scan->xs_snapshot = snapshot;
514 
515 	return scan;
516 }
517 
518 /* ----------------
519  * index_getnext_tid - get the next TID from a scan
520  *
521  * The result is the next TID satisfying the scan keys,
522  * or NULL if no more matching tuples exist.
523  * ----------------
524  */
525 ItemPointer
index_getnext_tid(IndexScanDesc scan,ScanDirection direction)526 index_getnext_tid(IndexScanDesc scan, ScanDirection direction)
527 {
528 	bool		found;
529 
530 	SCAN_CHECKS;
531 	CHECK_SCAN_PROCEDURE(amgettuple);
532 
533 	Assert(TransactionIdIsValid(RecentGlobalXmin));
534 
535 	/*
536 	 * The AM's amgettuple proc finds the next index entry matching the scan
537 	 * keys, and puts the TID into scan->xs_ctup.t_self.  It should also set
538 	 * scan->xs_recheck and possibly scan->xs_itup/scan->xs_hitup, though we
539 	 * pay no attention to those fields here.
540 	 */
541 	found = scan->indexRelation->rd_amroutine->amgettuple(scan, direction);
542 
543 	/* Reset kill flag immediately for safety */
544 	scan->kill_prior_tuple = false;
545 
546 	/* If we're out of index entries, we're done */
547 	if (!found)
548 	{
549 		/* ... but first, release any held pin on a heap page */
550 		if (BufferIsValid(scan->xs_cbuf))
551 		{
552 			ReleaseBuffer(scan->xs_cbuf);
553 			scan->xs_cbuf = InvalidBuffer;
554 		}
555 		return NULL;
556 	}
557 
558 	pgstat_count_index_tuples(scan->indexRelation, 1);
559 
560 	/* Return the TID of the tuple we found. */
561 	return &scan->xs_ctup.t_self;
562 }
563 
564 /* ----------------
565  *		index_fetch_heap - get the scan's next heap tuple
566  *
567  * The result is a visible heap tuple associated with the index TID most
568  * recently fetched by index_getnext_tid, or NULL if no more matching tuples
569  * exist.  (There can be more than one matching tuple because of HOT chains,
570  * although when using an MVCC snapshot it should be impossible for more than
571  * one such tuple to exist.)
572  *
573  * On success, the buffer containing the heap tup is pinned (the pin will be
574  * dropped in a future index_getnext_tid, index_fetch_heap or index_endscan
575  * call).
576  *
577  * Note: caller must check scan->xs_recheck, and perform rechecking of the
578  * scan keys if required.  We do not do that here because we don't have
579  * enough information to do it efficiently in the general case.
580  * ----------------
581  */
582 HeapTuple
index_fetch_heap(IndexScanDesc scan)583 index_fetch_heap(IndexScanDesc scan)
584 {
585 	ItemPointer tid = &scan->xs_ctup.t_self;
586 	bool		all_dead = false;
587 	bool		got_heap_tuple;
588 
589 	/* We can skip the buffer-switching logic if we're in mid-HOT chain. */
590 	if (!scan->xs_continue_hot)
591 	{
592 		/* Switch to correct buffer if we don't have it already */
593 		Buffer		prev_buf = scan->xs_cbuf;
594 
595 		scan->xs_cbuf = ReleaseAndReadBuffer(scan->xs_cbuf,
596 											 scan->heapRelation,
597 											 ItemPointerGetBlockNumber(tid));
598 
599 		/*
600 		 * Prune page, but only if we weren't already on this page
601 		 */
602 		if (prev_buf != scan->xs_cbuf)
603 			heap_page_prune_opt(scan->heapRelation, scan->xs_cbuf);
604 	}
605 
606 	/* Obtain share-lock on the buffer so we can examine visibility */
607 	LockBuffer(scan->xs_cbuf, BUFFER_LOCK_SHARE);
608 	got_heap_tuple = heap_hot_search_buffer(tid, scan->heapRelation,
609 											scan->xs_cbuf,
610 											scan->xs_snapshot,
611 											&scan->xs_ctup,
612 											&all_dead,
613 											!scan->xs_continue_hot);
614 	LockBuffer(scan->xs_cbuf, BUFFER_LOCK_UNLOCK);
615 
616 	if (got_heap_tuple)
617 	{
618 		/*
619 		 * Only in a non-MVCC snapshot can more than one member of the HOT
620 		 * chain be visible.
621 		 */
622 		scan->xs_continue_hot = !IsMVCCSnapshot(scan->xs_snapshot);
623 		pgstat_count_heap_fetch(scan->indexRelation);
624 		return &scan->xs_ctup;
625 	}
626 
627 	/* We've reached the end of the HOT chain. */
628 	scan->xs_continue_hot = false;
629 
630 	/*
631 	 * If we scanned a whole HOT chain and found only dead tuples, tell index
632 	 * AM to kill its entry for that TID (this will take effect in the next
633 	 * amgettuple call, in index_getnext_tid).  We do not do this when in
634 	 * recovery because it may violate MVCC to do so.  See comments in
635 	 * RelationGetIndexScan().
636 	 */
637 	if (!scan->xactStartedInRecovery)
638 		scan->kill_prior_tuple = all_dead;
639 
640 	return NULL;
641 }
642 
643 /* ----------------
644  *		index_getnext - get the next heap tuple from a scan
645  *
646  * The result is the next heap tuple satisfying the scan keys and the
647  * snapshot, or NULL if no more matching tuples exist.
648  *
649  * On success, the buffer containing the heap tup is pinned (the pin will be
650  * dropped in a future index_getnext_tid, index_fetch_heap or index_endscan
651  * call).
652  *
653  * Note: caller must check scan->xs_recheck, and perform rechecking of the
654  * scan keys if required.  We do not do that here because we don't have
655  * enough information to do it efficiently in the general case.
656  * ----------------
657  */
658 HeapTuple
index_getnext(IndexScanDesc scan,ScanDirection direction)659 index_getnext(IndexScanDesc scan, ScanDirection direction)
660 {
661 	HeapTuple	heapTuple;
662 	ItemPointer tid;
663 
664 	for (;;)
665 	{
666 		if (scan->xs_continue_hot)
667 		{
668 			/*
669 			 * We are resuming scan of a HOT chain after having returned an
670 			 * earlier member.  Must still hold pin on current heap page.
671 			 */
672 			Assert(BufferIsValid(scan->xs_cbuf));
673 			Assert(ItemPointerGetBlockNumber(&scan->xs_ctup.t_self) ==
674 				   BufferGetBlockNumber(scan->xs_cbuf));
675 		}
676 		else
677 		{
678 			/* Time to fetch the next TID from the index */
679 			tid = index_getnext_tid(scan, direction);
680 
681 			/* If we're out of index entries, we're done */
682 			if (tid == NULL)
683 				break;
684 		}
685 
686 		/*
687 		 * Fetch the next (or only) visible heap tuple for this index entry.
688 		 * If we don't find anything, loop around and grab the next TID from
689 		 * the index.
690 		 */
691 		heapTuple = index_fetch_heap(scan);
692 		if (heapTuple != NULL)
693 			return heapTuple;
694 	}
695 
696 	return NULL;				/* failure exit */
697 }
698 
699 /* ----------------
700  *		index_getbitmap - get all tuples at once from an index scan
701  *
702  * Adds the TIDs of all heap tuples satisfying the scan keys to a bitmap.
703  * Since there's no interlock between the index scan and the eventual heap
704  * access, this is only safe to use with MVCC-based snapshots: the heap
705  * item slot could have been replaced by a newer tuple by the time we get
706  * to it.
707  *
708  * Returns the number of matching tuples found.  (Note: this might be only
709  * approximate, so it should only be used for statistical purposes.)
710  * ----------------
711  */
712 int64
index_getbitmap(IndexScanDesc scan,TIDBitmap * bitmap)713 index_getbitmap(IndexScanDesc scan, TIDBitmap *bitmap)
714 {
715 	int64		ntids;
716 
717 	SCAN_CHECKS;
718 	CHECK_SCAN_PROCEDURE(amgetbitmap);
719 
720 	/* just make sure this is false... */
721 	scan->kill_prior_tuple = false;
722 
723 	/*
724 	 * have the am's getbitmap proc do all the work.
725 	 */
726 	ntids = scan->indexRelation->rd_amroutine->amgetbitmap(scan, bitmap);
727 
728 	pgstat_count_index_tuples(scan->indexRelation, ntids);
729 
730 	return ntids;
731 }
732 
733 /* ----------------
734  *		index_bulk_delete - do mass deletion of index entries
735  *
736  *		callback routine tells whether a given main-heap tuple is
737  *		to be deleted
738  *
739  *		return value is an optional palloc'd struct of statistics
740  * ----------------
741  */
742 IndexBulkDeleteResult *
index_bulk_delete(IndexVacuumInfo * info,IndexBulkDeleteResult * stats,IndexBulkDeleteCallback callback,void * callback_state)743 index_bulk_delete(IndexVacuumInfo *info,
744 				  IndexBulkDeleteResult *stats,
745 				  IndexBulkDeleteCallback callback,
746 				  void *callback_state)
747 {
748 	Relation	indexRelation = info->index;
749 
750 	RELATION_CHECKS;
751 	CHECK_REL_PROCEDURE(ambulkdelete);
752 
753 	return indexRelation->rd_amroutine->ambulkdelete(info, stats,
754 													 callback, callback_state);
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 *
index_vacuum_cleanup(IndexVacuumInfo * info,IndexBulkDeleteResult * stats)764 index_vacuum_cleanup(IndexVacuumInfo *info,
765 					 IndexBulkDeleteResult *stats)
766 {
767 	Relation	indexRelation = info->index;
768 
769 	RELATION_CHECKS;
770 	CHECK_REL_PROCEDURE(amvacuumcleanup);
771 
772 	return indexRelation->rd_amroutine->amvacuumcleanup(info, stats);
773 }
774 
775 /* ----------------
776  *		index_can_return
777  *
778  *		Does the index access method support index-only scans for the given
779  *		column?
780  * ----------------
781  */
782 bool
index_can_return(Relation indexRelation,int attno)783 index_can_return(Relation indexRelation, int attno)
784 {
785 	RELATION_CHECKS;
786 
787 	/* amcanreturn is optional; assume FALSE if not provided by AM */
788 	if (indexRelation->rd_amroutine->amcanreturn == NULL)
789 		return false;
790 
791 	return indexRelation->rd_amroutine->amcanreturn(indexRelation, attno);
792 }
793 
794 /* ----------------
795  *		index_getprocid
796  *
797  *		Index access methods typically require support routines that are
798  *		not directly the implementation of any WHERE-clause query operator
799  *		and so cannot be kept in pg_amop.  Instead, such routines are kept
800  *		in pg_amproc.  These registered procedure OIDs are assigned numbers
801  *		according to a convention established by the access method.
802  *		The general index code doesn't know anything about the routines
803  *		involved; it just builds an ordered list of them for
804  *		each attribute on which an index is defined.
805  *
806  *		As of Postgres 8.3, support routines within an operator family
807  *		are further subdivided by the "left type" and "right type" of the
808  *		query operator(s) that they support.  The "default" functions for a
809  *		particular indexed attribute are those with both types equal to
810  *		the index opclass' opcintype (note that this is subtly different
811  *		from the indexed attribute's own type: it may be a binary-compatible
812  *		type instead).  Only the default functions are stored in relcache
813  *		entries --- access methods can use the syscache to look up non-default
814  *		functions.
815  *
816  *		This routine returns the requested default procedure OID for a
817  *		particular indexed attribute.
818  * ----------------
819  */
820 RegProcedure
index_getprocid(Relation irel,AttrNumber attnum,uint16 procnum)821 index_getprocid(Relation irel,
822 				AttrNumber attnum,
823 				uint16 procnum)
824 {
825 	RegProcedure *loc;
826 	int			nproc;
827 	int			procindex;
828 
829 	nproc = irel->rd_amroutine->amsupport;
830 
831 	Assert(procnum > 0 && procnum <= (uint16) nproc);
832 
833 	procindex = (nproc * (attnum - 1)) + (procnum - 1);
834 
835 	loc = irel->rd_support;
836 
837 	Assert(loc != NULL);
838 
839 	return loc[procindex];
840 }
841 
842 /* ----------------
843  *		index_getprocinfo
844  *
845  *		This routine allows index AMs to keep fmgr lookup info for
846  *		support procs in the relcache.  As above, only the "default"
847  *		functions for any particular indexed attribute are cached.
848  *
849  * Note: the return value points into cached data that will be lost during
850  * any relcache rebuild!  Therefore, either use the callinfo right away,
851  * or save it only after having acquired some type of lock on the index rel.
852  * ----------------
853  */
854 FmgrInfo *
index_getprocinfo(Relation irel,AttrNumber attnum,uint16 procnum)855 index_getprocinfo(Relation irel,
856 				  AttrNumber attnum,
857 				  uint16 procnum)
858 {
859 	FmgrInfo   *locinfo;
860 	int			nproc;
861 	int			procindex;
862 
863 	nproc = irel->rd_amroutine->amsupport;
864 
865 	Assert(procnum > 0 && procnum <= (uint16) nproc);
866 
867 	procindex = (nproc * (attnum - 1)) + (procnum - 1);
868 
869 	locinfo = irel->rd_supportinfo;
870 
871 	Assert(locinfo != NULL);
872 
873 	locinfo += procindex;
874 
875 	/* Initialize the lookup info if first time through */
876 	if (locinfo->fn_oid == InvalidOid)
877 	{
878 		RegProcedure *loc = irel->rd_support;
879 		RegProcedure procId;
880 
881 		Assert(loc != NULL);
882 
883 		procId = loc[procindex];
884 
885 		/*
886 		 * Complain if function was not found during IndexSupportInitialize.
887 		 * This should not happen unless the system tables contain bogus
888 		 * entries for the index opclass.  (If an AM wants to allow a support
889 		 * function to be optional, it can use index_getprocid.)
890 		 */
891 		if (!RegProcedureIsValid(procId))
892 			elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
893 				 procnum, attnum, RelationGetRelationName(irel));
894 
895 		fmgr_info_cxt(procId, locinfo, irel->rd_indexcxt);
896 	}
897 
898 	return locinfo;
899 }
900