Line data Source code
1 : /*
2 : * brin.c
3 : * Implementation of BRIN indexes for Postgres
4 : *
5 : * See src/backend/access/brin/README for details.
6 : *
7 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
8 : * Portions Copyright (c) 1994, Regents of the University of California
9 : *
10 : * IDENTIFICATION
11 : * src/backend/access/brin/brin.c
12 : *
13 : * TODO
14 : * * ScalarArrayOpExpr (amsearcharray -> SK_SEARCHARRAY)
15 : */
16 : #include "postgres.h"
17 :
18 : #include "access/brin.h"
19 : #include "access/brin_page.h"
20 : #include "access/brin_pageops.h"
21 : #include "access/brin_xlog.h"
22 : #include "access/relation.h"
23 : #include "access/reloptions.h"
24 : #include "access/relscan.h"
25 : #include "access/table.h"
26 : #include "access/tableam.h"
27 : #include "access/xloginsert.h"
28 : #include "catalog/index.h"
29 : #include "catalog/pg_am.h"
30 : #include "commands/vacuum.h"
31 : #include "miscadmin.h"
32 : #include "pgstat.h"
33 : #include "postmaster/autovacuum.h"
34 : #include "storage/bufmgr.h"
35 : #include "storage/freespace.h"
36 : #include "tcop/tcopprot.h"
37 : #include "utils/acl.h"
38 : #include "utils/datum.h"
39 : #include "utils/fmgrprotos.h"
40 : #include "utils/guc.h"
41 : #include "utils/index_selfuncs.h"
42 : #include "utils/memutils.h"
43 : #include "utils/rel.h"
44 : #include "utils/tuplesort.h"
45 :
46 : /* Magic numbers for parallel state sharing */
47 : #define PARALLEL_KEY_BRIN_SHARED UINT64CONST(0xB000000000000001)
48 : #define PARALLEL_KEY_TUPLESORT UINT64CONST(0xB000000000000002)
49 : #define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xB000000000000003)
50 : #define PARALLEL_KEY_WAL_USAGE UINT64CONST(0xB000000000000004)
51 : #define PARALLEL_KEY_BUFFER_USAGE UINT64CONST(0xB000000000000005)
52 :
53 : /*
54 : * Status for index builds performed in parallel. This is allocated in a
55 : * dynamic shared memory segment.
56 : */
57 : typedef struct BrinShared
58 : {
59 : /*
60 : * These fields are not modified during the build. They primarily exist
61 : * for the benefit of worker processes that need to create state
62 : * corresponding to that used by the leader.
63 : */
64 : Oid heaprelid;
65 : Oid indexrelid;
66 : bool isconcurrent;
67 : BlockNumber pagesPerRange;
68 : int scantuplesortstates;
69 :
70 : /* Query ID, for report in worker processes */
71 : int64 queryid;
72 :
73 : /*
74 : * workersdonecv is used to monitor the progress of workers. All parallel
75 : * participants must indicate that they are done before leader can use
76 : * results built by the workers (and before leader can write the data into
77 : * the index).
78 : */
79 : ConditionVariable workersdonecv;
80 :
81 : /*
82 : * mutex protects all fields before heapdesc.
83 : *
84 : * These fields contain status information of interest to BRIN index
85 : * builds that must work just the same when an index is built in parallel.
86 : */
87 : slock_t mutex;
88 :
89 : /*
90 : * Mutable state that is maintained by workers, and reported back to
91 : * leader at end of the scans.
92 : *
93 : * nparticipantsdone is number of worker processes finished.
94 : *
95 : * reltuples is the total number of input heap tuples.
96 : *
97 : * indtuples is the total number of tuples that made it into the index.
98 : */
99 : int nparticipantsdone;
100 : double reltuples;
101 : double indtuples;
102 :
103 : /*
104 : * ParallelTableScanDescData data follows. Can't directly embed here, as
105 : * implementations of the parallel table scan desc interface might need
106 : * stronger alignment.
107 : */
108 : } BrinShared;
109 :
110 : /*
111 : * Return pointer to a BrinShared's parallel table scan.
112 : *
113 : * c.f. shm_toc_allocate as to why BUFFERALIGN is used, rather than just
114 : * MAXALIGN.
115 : */
116 : #define ParallelTableScanFromBrinShared(shared) \
117 : (ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(BrinShared)))
118 :
119 : /*
120 : * Status for leader in parallel index build.
121 : */
122 : typedef struct BrinLeader
123 : {
124 : /* parallel context itself */
125 : ParallelContext *pcxt;
126 :
127 : /*
128 : * nparticipanttuplesorts is the exact number of worker processes
129 : * successfully launched, plus one leader process if it participates as a
130 : * worker (only DISABLE_LEADER_PARTICIPATION builds avoid leader
131 : * participating as a worker).
132 : */
133 : int nparticipanttuplesorts;
134 :
135 : /*
136 : * Leader process convenience pointers to shared state (leader avoids TOC
137 : * lookups).
138 : *
139 : * brinshared is the shared state for entire build. sharedsort is the
140 : * shared, tuplesort-managed state passed to each process tuplesort.
141 : * snapshot is the snapshot used by the scan iff an MVCC snapshot is
142 : * required.
143 : */
144 : BrinShared *brinshared;
145 : Sharedsort *sharedsort;
146 : Snapshot snapshot;
147 : WalUsage *walusage;
148 : BufferUsage *bufferusage;
149 : } BrinLeader;
150 :
151 : /*
152 : * We use a BrinBuildState during initial construction of a BRIN index.
153 : * The running state is kept in a BrinMemTuple.
154 : */
155 : typedef struct BrinBuildState
156 : {
157 : Relation bs_irel;
158 : double bs_numtuples;
159 : double bs_reltuples;
160 : Buffer bs_currentInsertBuf;
161 : BlockNumber bs_pagesPerRange;
162 : BlockNumber bs_currRangeStart;
163 : BlockNumber bs_maxRangeStart;
164 : BrinRevmap *bs_rmAccess;
165 : BrinDesc *bs_bdesc;
166 : BrinMemTuple *bs_dtuple;
167 :
168 : BrinTuple *bs_emptyTuple;
169 : Size bs_emptyTupleLen;
170 : MemoryContext bs_context;
171 :
172 : /*
173 : * bs_leader is only present when a parallel index build is performed, and
174 : * only in the leader process. (Actually, only the leader process has a
175 : * BrinBuildState.)
176 : */
177 : BrinLeader *bs_leader;
178 : int bs_worker_id;
179 :
180 : /*
181 : * The sortstate is used by workers (including the leader). It has to be
182 : * part of the build state, because that's the only thing passed to the
183 : * build callback etc.
184 : */
185 : Tuplesortstate *bs_sortstate;
186 : } BrinBuildState;
187 :
188 : /*
189 : * We use a BrinInsertState to capture running state spanning multiple
190 : * brininsert invocations, within the same command.
191 : */
192 : typedef struct BrinInsertState
193 : {
194 : BrinRevmap *bis_rmAccess;
195 : BrinDesc *bis_desc;
196 : BlockNumber bis_pages_per_range;
197 : } BrinInsertState;
198 :
199 : /*
200 : * Struct used as "opaque" during index scans
201 : */
202 : typedef struct BrinOpaque
203 : {
204 : BlockNumber bo_pagesPerRange;
205 : BrinRevmap *bo_rmAccess;
206 : BrinDesc *bo_bdesc;
207 : } BrinOpaque;
208 :
209 : #define BRIN_ALL_BLOCKRANGES InvalidBlockNumber
210 :
211 : static BrinBuildState *initialize_brin_buildstate(Relation idxRel,
212 : BrinRevmap *revmap,
213 : BlockNumber pagesPerRange,
214 : BlockNumber tablePages);
215 : static BrinInsertState *initialize_brin_insertstate(Relation idxRel, IndexInfo *indexInfo);
216 : static void terminate_brin_buildstate(BrinBuildState *state);
217 : static void brinsummarize(Relation index, Relation heapRel, BlockNumber pageRange,
218 : bool include_partial, double *numSummarized, double *numExisting);
219 : static void form_and_insert_tuple(BrinBuildState *state);
220 : static void form_and_spill_tuple(BrinBuildState *state);
221 : static void union_tuples(BrinDesc *bdesc, BrinMemTuple *a,
222 : BrinTuple *b);
223 : static void brin_vacuum_scan(Relation idxrel, BufferAccessStrategy strategy);
224 : static bool add_values_to_range(Relation idxRel, BrinDesc *bdesc,
225 : BrinMemTuple *dtup, const Datum *values, const bool *nulls);
226 : static bool check_null_keys(BrinValues *bval, ScanKey *nullkeys, int nnullkeys);
227 : static void brin_fill_empty_ranges(BrinBuildState *state,
228 : BlockNumber prevRange, BlockNumber nextRange);
229 :
230 : /* parallel index builds */
231 : static void _brin_begin_parallel(BrinBuildState *buildstate, Relation heap, Relation index,
232 : bool isconcurrent, int request);
233 : static void _brin_end_parallel(BrinLeader *brinleader, BrinBuildState *state);
234 : static Size _brin_parallel_estimate_shared(Relation heap, Snapshot snapshot);
235 : static double _brin_parallel_heapscan(BrinBuildState *state);
236 : static double _brin_parallel_merge(BrinBuildState *state);
237 : static void _brin_leader_participate_as_worker(BrinBuildState *buildstate,
238 : Relation heap, Relation index);
239 : static void _brin_parallel_scan_and_build(BrinBuildState *state,
240 : BrinShared *brinshared,
241 : Sharedsort *sharedsort,
242 : Relation heap, Relation index,
243 : int sortmem, bool progress);
244 :
245 : /*
246 : * BRIN handler function: return IndexAmRoutine with access method parameters
247 : * and callbacks.
248 : */
249 : Datum
250 4404 : brinhandler(PG_FUNCTION_ARGS)
251 : {
252 4404 : IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
253 :
254 4404 : amroutine->amstrategies = 0;
255 4404 : amroutine->amsupport = BRIN_LAST_OPTIONAL_PROCNUM;
256 4404 : amroutine->amoptsprocnum = BRIN_PROCNUM_OPTIONS;
257 4404 : amroutine->amcanorder = false;
258 4404 : amroutine->amcanorderbyop = false;
259 4404 : amroutine->amcanhash = false;
260 4404 : amroutine->amconsistentequality = false;
261 4404 : amroutine->amconsistentordering = false;
262 4404 : amroutine->amcanbackward = false;
263 4404 : amroutine->amcanunique = false;
264 4404 : amroutine->amcanmulticol = true;
265 4404 : amroutine->amoptionalkey = true;
266 4404 : amroutine->amsearcharray = false;
267 4404 : amroutine->amsearchnulls = true;
268 4404 : amroutine->amstorage = true;
269 4404 : amroutine->amclusterable = false;
270 4404 : amroutine->ampredlocks = false;
271 4404 : amroutine->amcanparallel = false;
272 4404 : amroutine->amcanbuildparallel = true;
273 4404 : amroutine->amcaninclude = false;
274 4404 : amroutine->amusemaintenanceworkmem = false;
275 4404 : amroutine->amsummarizing = true;
276 4404 : amroutine->amparallelvacuumoptions =
277 : VACUUM_OPTION_PARALLEL_CLEANUP;
278 4404 : amroutine->amkeytype = InvalidOid;
279 :
280 4404 : amroutine->ambuild = brinbuild;
281 4404 : amroutine->ambuildempty = brinbuildempty;
282 4404 : amroutine->aminsert = brininsert;
283 4404 : amroutine->aminsertcleanup = brininsertcleanup;
284 4404 : amroutine->ambulkdelete = brinbulkdelete;
285 4404 : amroutine->amvacuumcleanup = brinvacuumcleanup;
286 4404 : amroutine->amcanreturn = NULL;
287 4404 : amroutine->amcostestimate = brincostestimate;
288 4404 : amroutine->amgettreeheight = NULL;
289 4404 : amroutine->amoptions = brinoptions;
290 4404 : amroutine->amproperty = NULL;
291 4404 : amroutine->ambuildphasename = NULL;
292 4404 : amroutine->amvalidate = brinvalidate;
293 4404 : amroutine->amadjustmembers = NULL;
294 4404 : amroutine->ambeginscan = brinbeginscan;
295 4404 : amroutine->amrescan = brinrescan;
296 4404 : amroutine->amgettuple = NULL;
297 4404 : amroutine->amgetbitmap = bringetbitmap;
298 4404 : amroutine->amendscan = brinendscan;
299 4404 : amroutine->ammarkpos = NULL;
300 4404 : amroutine->amrestrpos = NULL;
301 4404 : amroutine->amestimateparallelscan = NULL;
302 4404 : amroutine->aminitparallelscan = NULL;
303 4404 : amroutine->amparallelrescan = NULL;
304 4404 : amroutine->amtranslatestrategy = NULL;
305 4404 : amroutine->amtranslatecmptype = NULL;
306 :
307 4404 : PG_RETURN_POINTER(amroutine);
308 : }
309 :
310 : /*
311 : * Initialize a BrinInsertState to maintain state to be used across multiple
312 : * tuple inserts, within the same command.
313 : */
314 : static BrinInsertState *
315 1122 : initialize_brin_insertstate(Relation idxRel, IndexInfo *indexInfo)
316 : {
317 : BrinInsertState *bistate;
318 : MemoryContext oldcxt;
319 :
320 1122 : oldcxt = MemoryContextSwitchTo(indexInfo->ii_Context);
321 1122 : bistate = palloc0(sizeof(BrinInsertState));
322 1122 : bistate->bis_desc = brin_build_desc(idxRel);
323 1122 : bistate->bis_rmAccess = brinRevmapInitialize(idxRel,
324 : &bistate->bis_pages_per_range);
325 1122 : indexInfo->ii_AmCache = bistate;
326 1122 : MemoryContextSwitchTo(oldcxt);
327 :
328 1122 : return bistate;
329 : }
330 :
331 : /*
332 : * A tuple in the heap is being inserted. To keep a brin index up to date,
333 : * we need to obtain the relevant index tuple and compare its stored values
334 : * with those of the new tuple. If the tuple values are not consistent with
335 : * the summary tuple, we need to update the index tuple.
336 : *
337 : * If autosummarization is enabled, check if we need to summarize the previous
338 : * page range.
339 : *
340 : * If the range is not currently summarized (i.e. the revmap returns NULL for
341 : * it), there's nothing to do for this tuple.
342 : */
343 : bool
344 126136 : brininsert(Relation idxRel, Datum *values, bool *nulls,
345 : ItemPointer heaptid, Relation heapRel,
346 : IndexUniqueCheck checkUnique,
347 : bool indexUnchanged,
348 : IndexInfo *indexInfo)
349 : {
350 : BlockNumber pagesPerRange;
351 : BlockNumber origHeapBlk;
352 : BlockNumber heapBlk;
353 126136 : BrinInsertState *bistate = (BrinInsertState *) indexInfo->ii_AmCache;
354 : BrinRevmap *revmap;
355 : BrinDesc *bdesc;
356 126136 : Buffer buf = InvalidBuffer;
357 126136 : MemoryContext tupcxt = NULL;
358 126136 : MemoryContext oldcxt = CurrentMemoryContext;
359 126136 : bool autosummarize = BrinGetAutoSummarize(idxRel);
360 :
361 : /*
362 : * If first time through in this statement, initialize the insert state
363 : * that we keep for all the inserts in the command.
364 : */
365 126136 : if (!bistate)
366 1122 : bistate = initialize_brin_insertstate(idxRel, indexInfo);
367 :
368 126136 : revmap = bistate->bis_rmAccess;
369 126136 : bdesc = bistate->bis_desc;
370 126136 : pagesPerRange = bistate->bis_pages_per_range;
371 :
372 : /*
373 : * origHeapBlk is the block number where the insertion occurred. heapBlk
374 : * is the first block in the corresponding page range.
375 : */
376 126136 : origHeapBlk = ItemPointerGetBlockNumber(heaptid);
377 126136 : heapBlk = (origHeapBlk / pagesPerRange) * pagesPerRange;
378 :
379 : for (;;)
380 0 : {
381 126136 : bool need_insert = false;
382 : OffsetNumber off;
383 : BrinTuple *brtup;
384 : BrinMemTuple *dtup;
385 :
386 126136 : CHECK_FOR_INTERRUPTS();
387 :
388 : /*
389 : * If auto-summarization is enabled and we just inserted the first
390 : * tuple into the first block of a new non-first page range, request a
391 : * summarization run of the previous range.
392 : */
393 126136 : if (autosummarize &&
394 290 : heapBlk > 0 &&
395 290 : heapBlk == origHeapBlk &&
396 290 : ItemPointerGetOffsetNumber(heaptid) == FirstOffsetNumber)
397 : {
398 16 : BlockNumber lastPageRange = heapBlk - 1;
399 : BrinTuple *lastPageTuple;
400 :
401 : lastPageTuple =
402 16 : brinGetTupleForHeapBlock(revmap, lastPageRange, &buf, &off,
403 : NULL, BUFFER_LOCK_SHARE);
404 16 : if (!lastPageTuple)
405 : {
406 : bool recorded;
407 :
408 12 : recorded = AutoVacuumRequestWork(AVW_BRINSummarizeRange,
409 : RelationGetRelid(idxRel),
410 : lastPageRange);
411 12 : if (!recorded)
412 0 : ereport(LOG,
413 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
414 : errmsg("request for BRIN range summarization for index \"%s\" page %u was not recorded",
415 : RelationGetRelationName(idxRel),
416 : lastPageRange)));
417 : }
418 : else
419 4 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
420 : }
421 :
422 126136 : brtup = brinGetTupleForHeapBlock(revmap, heapBlk, &buf, &off,
423 : NULL, BUFFER_LOCK_SHARE);
424 :
425 : /* if range is unsummarized, there's nothing to do */
426 126136 : if (!brtup)
427 78254 : break;
428 :
429 : /* First time through in this brininsert call? */
430 47882 : if (tupcxt == NULL)
431 : {
432 47882 : tupcxt = AllocSetContextCreate(CurrentMemoryContext,
433 : "brininsert cxt",
434 : ALLOCSET_DEFAULT_SIZES);
435 47882 : MemoryContextSwitchTo(tupcxt);
436 : }
437 :
438 47882 : dtup = brin_deform_tuple(bdesc, brtup, NULL);
439 :
440 47882 : need_insert = add_values_to_range(idxRel, bdesc, dtup, values, nulls);
441 :
442 47882 : if (!need_insert)
443 : {
444 : /*
445 : * The tuple is consistent with the new values, so there's nothing
446 : * to do.
447 : */
448 23954 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
449 : }
450 : else
451 : {
452 23928 : Page page = BufferGetPage(buf);
453 23928 : ItemId lp = PageGetItemId(page, off);
454 : Size origsz;
455 : BrinTuple *origtup;
456 : Size newsz;
457 : BrinTuple *newtup;
458 : bool samepage;
459 :
460 : /*
461 : * Make a copy of the old tuple, so that we can compare it after
462 : * re-acquiring the lock.
463 : */
464 23928 : origsz = ItemIdGetLength(lp);
465 23928 : origtup = brin_copy_tuple(brtup, origsz, NULL, NULL);
466 :
467 : /*
468 : * Before releasing the lock, check if we can attempt a same-page
469 : * update. Another process could insert a tuple concurrently in
470 : * the same page though, so downstream we must be prepared to cope
471 : * if this turns out to not be possible after all.
472 : */
473 23928 : newtup = brin_form_tuple(bdesc, heapBlk, dtup, &newsz);
474 23928 : samepage = brin_can_do_samepage_update(buf, origsz, newsz);
475 23928 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
476 :
477 : /*
478 : * Try to update the tuple. If this doesn't work for whatever
479 : * reason, we need to restart from the top; the revmap might be
480 : * pointing at a different tuple for this block now, so we need to
481 : * recompute to ensure both our new heap tuple and the other
482 : * inserter's are covered by the combined tuple. It might be that
483 : * we don't need to update at all.
484 : */
485 23928 : if (!brin_doupdate(idxRel, pagesPerRange, revmap, heapBlk,
486 : buf, off, origtup, origsz, newtup, newsz,
487 : samepage))
488 : {
489 : /* no luck; start over */
490 0 : MemoryContextReset(tupcxt);
491 0 : continue;
492 : }
493 : }
494 :
495 : /* success! */
496 47882 : break;
497 : }
498 :
499 126136 : if (BufferIsValid(buf))
500 47886 : ReleaseBuffer(buf);
501 126136 : MemoryContextSwitchTo(oldcxt);
502 126136 : if (tupcxt != NULL)
503 47882 : MemoryContextDelete(tupcxt);
504 :
505 126136 : return false;
506 : }
507 :
508 : /*
509 : * Callback to clean up the BrinInsertState once all tuple inserts are done.
510 : */
511 : void
512 1156 : brininsertcleanup(Relation index, IndexInfo *indexInfo)
513 : {
514 1156 : BrinInsertState *bistate = (BrinInsertState *) indexInfo->ii_AmCache;
515 :
516 : /* bail out if cache not initialized */
517 1156 : if (bistate == NULL)
518 34 : return;
519 :
520 : /* do this first to avoid dangling pointer if we fail partway through */
521 1122 : indexInfo->ii_AmCache = NULL;
522 :
523 : /*
524 : * Clean up the revmap. Note that the brinDesc has already been cleaned up
525 : * as part of its own memory context.
526 : */
527 1122 : brinRevmapTerminate(bistate->bis_rmAccess);
528 1122 : pfree(bistate);
529 : }
530 :
531 : /*
532 : * Initialize state for a BRIN index scan.
533 : *
534 : * We read the metapage here to determine the pages-per-range number that this
535 : * index was built with. Note that since this cannot be changed while we're
536 : * holding lock on index, it's not necessary to recompute it during brinrescan.
537 : */
538 : IndexScanDesc
539 2946 : brinbeginscan(Relation r, int nkeys, int norderbys)
540 : {
541 : IndexScanDesc scan;
542 : BrinOpaque *opaque;
543 :
544 2946 : scan = RelationGetIndexScan(r, nkeys, norderbys);
545 :
546 2946 : opaque = palloc_object(BrinOpaque);
547 2946 : opaque->bo_rmAccess = brinRevmapInitialize(r, &opaque->bo_pagesPerRange);
548 2946 : opaque->bo_bdesc = brin_build_desc(r);
549 2946 : scan->opaque = opaque;
550 :
551 2946 : return scan;
552 : }
553 :
554 : /*
555 : * Execute the index scan.
556 : *
557 : * This works by reading index TIDs from the revmap, and obtaining the index
558 : * tuples pointed to by them; the summary values in the index tuples are
559 : * compared to the scan keys. We return into the TID bitmap all the pages in
560 : * ranges corresponding to index tuples that match the scan keys.
561 : *
562 : * If a TID from the revmap is read as InvalidTID, we know that range is
563 : * unsummarized. Pages in those ranges need to be returned regardless of scan
564 : * keys.
565 : */
566 : int64
567 2946 : bringetbitmap(IndexScanDesc scan, TIDBitmap *tbm)
568 : {
569 2946 : Relation idxRel = scan->indexRelation;
570 2946 : Buffer buf = InvalidBuffer;
571 : BrinDesc *bdesc;
572 : Oid heapOid;
573 : Relation heapRel;
574 : BrinOpaque *opaque;
575 : BlockNumber nblocks;
576 2946 : int64 totalpages = 0;
577 : FmgrInfo *consistentFn;
578 : MemoryContext oldcxt;
579 : MemoryContext perRangeCxt;
580 : BrinMemTuple *dtup;
581 2946 : BrinTuple *btup = NULL;
582 2946 : Size btupsz = 0;
583 : ScanKey **keys,
584 : **nullkeys;
585 : int *nkeys,
586 : *nnullkeys;
587 : char *ptr;
588 : Size len;
589 : char *tmp PG_USED_FOR_ASSERTS_ONLY;
590 :
591 2946 : opaque = (BrinOpaque *) scan->opaque;
592 2946 : bdesc = opaque->bo_bdesc;
593 2946 : pgstat_count_index_scan(idxRel);
594 2946 : if (scan->instrument)
595 2946 : scan->instrument->nsearches++;
596 :
597 : /*
598 : * We need to know the size of the table so that we know how long to
599 : * iterate on the revmap.
600 : */
601 2946 : heapOid = IndexGetRelation(RelationGetRelid(idxRel), false);
602 2946 : heapRel = table_open(heapOid, AccessShareLock);
603 2946 : nblocks = RelationGetNumberOfBlocks(heapRel);
604 2946 : table_close(heapRel, AccessShareLock);
605 :
606 : /*
607 : * Make room for the consistent support procedures of indexed columns. We
608 : * don't look them up here; we do that lazily the first time we see a scan
609 : * key reference each of them. We rely on zeroing fn_oid to InvalidOid.
610 : */
611 2946 : consistentFn = palloc0_array(FmgrInfo, bdesc->bd_tupdesc->natts);
612 :
613 : /*
614 : * Make room for per-attribute lists of scan keys that we'll pass to the
615 : * consistent support procedure. We don't know which attributes have scan
616 : * keys, so we allocate space for all attributes. That may use more memory
617 : * but it's probably cheaper than determining which attributes are used.
618 : *
619 : * We keep null and regular keys separate, so that we can pass just the
620 : * regular keys to the consistent function easily.
621 : *
622 : * To reduce the allocation overhead, we allocate one big chunk and then
623 : * carve it into smaller arrays ourselves. All the pieces have exactly the
624 : * same lifetime, so that's OK.
625 : *
626 : * XXX The widest index can have 32 attributes, so the amount of wasted
627 : * memory is negligible. We could invent a more compact approach (with
628 : * just space for used attributes) but that would make the matching more
629 : * complex so it's not a good trade-off.
630 : */
631 2946 : len =
632 2946 : MAXALIGN(sizeof(ScanKey *) * bdesc->bd_tupdesc->natts) + /* regular keys */
633 2946 : MAXALIGN(sizeof(ScanKey) * scan->numberOfKeys) * bdesc->bd_tupdesc->natts +
634 2946 : MAXALIGN(sizeof(int) * bdesc->bd_tupdesc->natts) +
635 2946 : MAXALIGN(sizeof(ScanKey *) * bdesc->bd_tupdesc->natts) + /* NULL keys */
636 2946 : MAXALIGN(sizeof(ScanKey) * scan->numberOfKeys) * bdesc->bd_tupdesc->natts +
637 2946 : MAXALIGN(sizeof(int) * bdesc->bd_tupdesc->natts);
638 :
639 2946 : ptr = palloc(len);
640 2946 : tmp = ptr;
641 :
642 2946 : keys = (ScanKey **) ptr;
643 2946 : ptr += MAXALIGN(sizeof(ScanKey *) * bdesc->bd_tupdesc->natts);
644 :
645 2946 : nullkeys = (ScanKey **) ptr;
646 2946 : ptr += MAXALIGN(sizeof(ScanKey *) * bdesc->bd_tupdesc->natts);
647 :
648 2946 : nkeys = (int *) ptr;
649 2946 : ptr += MAXALIGN(sizeof(int) * bdesc->bd_tupdesc->natts);
650 :
651 2946 : nnullkeys = (int *) ptr;
652 2946 : ptr += MAXALIGN(sizeof(int) * bdesc->bd_tupdesc->natts);
653 :
654 69978 : for (int i = 0; i < bdesc->bd_tupdesc->natts; i++)
655 : {
656 67032 : keys[i] = (ScanKey *) ptr;
657 67032 : ptr += MAXALIGN(sizeof(ScanKey) * scan->numberOfKeys);
658 :
659 67032 : nullkeys[i] = (ScanKey *) ptr;
660 67032 : ptr += MAXALIGN(sizeof(ScanKey) * scan->numberOfKeys);
661 : }
662 :
663 : Assert(tmp + len == ptr);
664 :
665 : /* zero the number of keys */
666 2946 : memset(nkeys, 0, sizeof(int) * bdesc->bd_tupdesc->natts);
667 2946 : memset(nnullkeys, 0, sizeof(int) * bdesc->bd_tupdesc->natts);
668 :
669 : /* Preprocess the scan keys - split them into per-attribute arrays. */
670 5892 : for (int keyno = 0; keyno < scan->numberOfKeys; keyno++)
671 : {
672 2946 : ScanKey key = &scan->keyData[keyno];
673 2946 : AttrNumber keyattno = key->sk_attno;
674 :
675 : /*
676 : * The collation of the scan key must match the collation used in the
677 : * index column (but only if the search is not IS NULL/ IS NOT NULL).
678 : * Otherwise we shouldn't be using this index ...
679 : */
680 : Assert((key->sk_flags & SK_ISNULL) ||
681 : (key->sk_collation ==
682 : TupleDescAttr(bdesc->bd_tupdesc,
683 : keyattno - 1)->attcollation));
684 :
685 : /*
686 : * First time we see this index attribute, so init as needed.
687 : *
688 : * This is a bit of an overkill - we don't know how many scan keys are
689 : * there for this attribute, so we simply allocate the largest number
690 : * possible (as if all keys were for this attribute). This may waste a
691 : * bit of memory, but we only expect small number of scan keys in
692 : * general, so this should be negligible, and repeated repalloc calls
693 : * are not free either.
694 : */
695 2946 : if (consistentFn[keyattno - 1].fn_oid == InvalidOid)
696 : {
697 : FmgrInfo *tmp;
698 :
699 : /* First time we see this attribute, so no key/null keys. */
700 : Assert(nkeys[keyattno - 1] == 0);
701 : Assert(nnullkeys[keyattno - 1] == 0);
702 :
703 2946 : tmp = index_getprocinfo(idxRel, keyattno,
704 : BRIN_PROCNUM_CONSISTENT);
705 2946 : fmgr_info_copy(&consistentFn[keyattno - 1], tmp,
706 : CurrentMemoryContext);
707 : }
708 :
709 : /* Add key to the proper per-attribute array. */
710 2946 : if (key->sk_flags & SK_ISNULL)
711 : {
712 36 : nullkeys[keyattno - 1][nnullkeys[keyattno - 1]] = key;
713 36 : nnullkeys[keyattno - 1]++;
714 : }
715 : else
716 : {
717 2910 : keys[keyattno - 1][nkeys[keyattno - 1]] = key;
718 2910 : nkeys[keyattno - 1]++;
719 : }
720 : }
721 :
722 : /* allocate an initial in-memory tuple, out of the per-range memcxt */
723 2946 : dtup = brin_new_memtuple(bdesc);
724 :
725 : /*
726 : * Setup and use a per-range memory context, which is reset every time we
727 : * loop below. This avoids having to free the tuples within the loop.
728 : */
729 2946 : perRangeCxt = AllocSetContextCreate(CurrentMemoryContext,
730 : "bringetbitmap cxt",
731 : ALLOCSET_DEFAULT_SIZES);
732 2946 : oldcxt = MemoryContextSwitchTo(perRangeCxt);
733 :
734 : /*
735 : * Now scan the revmap. We start by querying for heap page 0,
736 : * incrementing by the number of pages per range; this gives us a full
737 : * view of the table. We make use of uint64 for heapBlk as a BlockNumber
738 : * could wrap for tables with close to 2^32 pages.
739 : */
740 194598 : for (uint64 heapBlk = 0; heapBlk < nblocks; heapBlk += opaque->bo_pagesPerRange)
741 : {
742 : bool addrange;
743 191652 : bool gottuple = false;
744 : BrinTuple *tup;
745 : OffsetNumber off;
746 : Size size;
747 :
748 191652 : CHECK_FOR_INTERRUPTS();
749 :
750 191652 : MemoryContextReset(perRangeCxt);
751 :
752 191652 : tup = brinGetTupleForHeapBlock(opaque->bo_rmAccess, (BlockNumber) heapBlk, &buf,
753 : &off, &size, BUFFER_LOCK_SHARE);
754 191652 : if (tup)
755 : {
756 189936 : gottuple = true;
757 189936 : btup = brin_copy_tuple(tup, size, btup, &btupsz);
758 189936 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
759 : }
760 :
761 : /*
762 : * For page ranges with no indexed tuple, we must return the whole
763 : * range; otherwise, compare it to the scan keys.
764 : */
765 191652 : if (!gottuple)
766 : {
767 1716 : addrange = true;
768 : }
769 : else
770 : {
771 189936 : dtup = brin_deform_tuple(bdesc, btup, dtup);
772 189936 : if (dtup->bt_placeholder)
773 : {
774 : /*
775 : * Placeholder tuples are always returned, regardless of the
776 : * values stored in them.
777 : */
778 0 : addrange = true;
779 : }
780 : else
781 : {
782 : int attno;
783 :
784 : /*
785 : * Compare scan keys with summary values stored for the range.
786 : * If scan keys are matched, the page range must be added to
787 : * the bitmap. We initially assume the range needs to be
788 : * added; in particular this serves the case where there are
789 : * no keys.
790 : */
791 189936 : addrange = true;
792 4704068 : for (attno = 1; attno <= bdesc->bd_tupdesc->natts; attno++)
793 : {
794 : BrinValues *bval;
795 : Datum add;
796 : Oid collation;
797 :
798 : /*
799 : * skip attributes without any scan keys (both regular and
800 : * IS [NOT] NULL)
801 : */
802 4567734 : if (nkeys[attno - 1] == 0 && nnullkeys[attno - 1] == 0)
803 4377798 : continue;
804 :
805 189936 : bval = &dtup->bt_columns[attno - 1];
806 :
807 : /*
808 : * If the BRIN tuple indicates that this range is empty,
809 : * we can skip it: there's nothing to match. We don't
810 : * need to examine the next columns.
811 : */
812 189936 : if (dtup->bt_empty_range)
813 : {
814 0 : addrange = false;
815 0 : break;
816 : }
817 :
818 : /*
819 : * First check if there are any IS [NOT] NULL scan keys,
820 : * and if we're violating them. In that case we can
821 : * terminate early, without invoking the support function.
822 : *
823 : * As there may be more keys, we can only determine
824 : * mismatch within this loop.
825 : */
826 189936 : if (bdesc->bd_info[attno - 1]->oi_regular_nulls &&
827 189936 : !check_null_keys(bval, nullkeys[attno - 1],
828 189936 : nnullkeys[attno - 1]))
829 : {
830 : /*
831 : * If any of the IS [NOT] NULL keys failed, the page
832 : * range as a whole can't pass. So terminate the loop.
833 : */
834 996 : addrange = false;
835 996 : break;
836 : }
837 :
838 : /*
839 : * So either there are no IS [NOT] NULL keys, or all
840 : * passed. If there are no regular scan keys, we're done -
841 : * the page range matches. If there are regular keys, but
842 : * the page range is marked as 'all nulls' it can't
843 : * possibly pass (we're assuming the operators are
844 : * strict).
845 : */
846 :
847 : /* No regular scan keys - page range as a whole passes. */
848 188940 : if (!nkeys[attno - 1])
849 1236 : continue;
850 :
851 : Assert((nkeys[attno - 1] > 0) &&
852 : (nkeys[attno - 1] <= scan->numberOfKeys));
853 :
854 : /* If it is all nulls, it cannot possibly be consistent. */
855 187704 : if (bval->bv_allnulls)
856 : {
857 378 : addrange = false;
858 378 : break;
859 : }
860 :
861 : /*
862 : * Collation from the first key (has to be the same for
863 : * all keys for the same attribute).
864 : */
865 187326 : collation = keys[attno - 1][0]->sk_collation;
866 :
867 : /*
868 : * Check whether the scan key is consistent with the page
869 : * range values; if so, have the pages in the range added
870 : * to the output bitmap.
871 : *
872 : * The opclass may or may not support processing of
873 : * multiple scan keys. We can determine that based on the
874 : * number of arguments - functions with extra parameter
875 : * (number of scan keys) do support this, otherwise we
876 : * have to simply pass the scan keys one by one.
877 : */
878 187326 : if (consistentFn[attno - 1].fn_nargs >= 4)
879 : {
880 : /* Check all keys at once */
881 39594 : add = FunctionCall4Coll(&consistentFn[attno - 1],
882 : collation,
883 : PointerGetDatum(bdesc),
884 : PointerGetDatum(bval),
885 39594 : PointerGetDatum(keys[attno - 1]),
886 39594 : Int32GetDatum(nkeys[attno - 1]));
887 39594 : addrange = DatumGetBool(add);
888 : }
889 : else
890 : {
891 : /*
892 : * Check keys one by one
893 : *
894 : * When there are multiple scan keys, failure to meet
895 : * the criteria for a single one of them is enough to
896 : * discard the range as a whole, so break out of the
897 : * loop as soon as a false return value is obtained.
898 : */
899 : int keyno;
900 :
901 258078 : for (keyno = 0; keyno < nkeys[attno - 1]; keyno++)
902 : {
903 147732 : add = FunctionCall3Coll(&consistentFn[attno - 1],
904 147732 : keys[attno - 1][keyno]->sk_collation,
905 : PointerGetDatum(bdesc),
906 : PointerGetDatum(bval),
907 147732 : PointerGetDatum(keys[attno - 1][keyno]));
908 147732 : addrange = DatumGetBool(add);
909 147732 : if (!addrange)
910 37386 : break;
911 : }
912 : }
913 :
914 : /*
915 : * If we found a scan key eliminating the range, no need
916 : * to check additional ones.
917 : */
918 187326 : if (!addrange)
919 52228 : break;
920 : }
921 : }
922 : }
923 :
924 : /* add the pages in the range to the output bitmap, if needed */
925 191652 : if (addrange)
926 : {
927 : uint64 pageno;
928 :
929 138050 : for (pageno = heapBlk;
930 286020 : pageno <= Min(nblocks, heapBlk + opaque->bo_pagesPerRange) - 1;
931 147970 : pageno++)
932 : {
933 147970 : MemoryContextSwitchTo(oldcxt);
934 147970 : tbm_add_page(tbm, pageno);
935 147970 : totalpages++;
936 147970 : MemoryContextSwitchTo(perRangeCxt);
937 : }
938 : }
939 : }
940 :
941 2946 : MemoryContextSwitchTo(oldcxt);
942 2946 : MemoryContextDelete(perRangeCxt);
943 :
944 2946 : if (buf != InvalidBuffer)
945 2946 : ReleaseBuffer(buf);
946 :
947 : /*
948 : * XXX We have an approximation of the number of *pages* that our scan
949 : * returns, but we don't have a precise idea of the number of heap tuples
950 : * involved.
951 : */
952 2946 : return totalpages * 10;
953 : }
954 :
955 : /*
956 : * Re-initialize state for a BRIN index scan
957 : */
958 : void
959 2946 : brinrescan(IndexScanDesc scan, ScanKey scankey, int nscankeys,
960 : ScanKey orderbys, int norderbys)
961 : {
962 : /*
963 : * Other index AMs preprocess the scan keys at this point, or sometime
964 : * early during the scan; this lets them optimize by removing redundant
965 : * keys, or doing early returns when they are impossible to satisfy; see
966 : * _bt_preprocess_keys for an example. Something like that could be added
967 : * here someday, too.
968 : */
969 :
970 2946 : if (scankey && scan->numberOfKeys > 0)
971 2946 : memcpy(scan->keyData, scankey, scan->numberOfKeys * sizeof(ScanKeyData));
972 2946 : }
973 :
974 : /*
975 : * Close down a BRIN index scan
976 : */
977 : void
978 2946 : brinendscan(IndexScanDesc scan)
979 : {
980 2946 : BrinOpaque *opaque = (BrinOpaque *) scan->opaque;
981 :
982 2946 : brinRevmapTerminate(opaque->bo_rmAccess);
983 2946 : brin_free_desc(opaque->bo_bdesc);
984 2946 : pfree(opaque);
985 2946 : }
986 :
987 : /*
988 : * Per-heap-tuple callback for table_index_build_scan.
989 : *
990 : * Note we don't worry about the page range at the end of the table here; it is
991 : * present in the build state struct after we're called the last time, but not
992 : * inserted into the index. Caller must ensure to do so, if appropriate.
993 : */
994 : static void
995 728452 : brinbuildCallback(Relation index,
996 : ItemPointer tid,
997 : Datum *values,
998 : bool *isnull,
999 : bool tupleIsAlive,
1000 : void *brstate)
1001 : {
1002 728452 : BrinBuildState *state = (BrinBuildState *) brstate;
1003 : BlockNumber thisblock;
1004 :
1005 728452 : thisblock = ItemPointerGetBlockNumber(tid);
1006 :
1007 : /*
1008 : * If we're in a block that belongs to a future range, summarize what
1009 : * we've got and start afresh. Note the scan might have skipped many
1010 : * pages, if they were devoid of live tuples; make sure to insert index
1011 : * tuples for those too.
1012 : */
1013 730748 : while (thisblock > state->bs_currRangeStart + state->bs_pagesPerRange - 1)
1014 : {
1015 :
1016 : BRIN_elog((DEBUG2,
1017 : "brinbuildCallback: completed a range: %u--%u",
1018 : state->bs_currRangeStart,
1019 : state->bs_currRangeStart + state->bs_pagesPerRange));
1020 :
1021 : /* create the index tuple and insert it */
1022 2296 : form_and_insert_tuple(state);
1023 :
1024 : /* set state to correspond to the next range */
1025 2296 : state->bs_currRangeStart += state->bs_pagesPerRange;
1026 :
1027 : /* re-initialize state for it */
1028 2296 : brin_memtuple_initialize(state->bs_dtuple, state->bs_bdesc);
1029 : }
1030 :
1031 : /* Accumulate the current tuple into the running state */
1032 728452 : (void) add_values_to_range(index, state->bs_bdesc, state->bs_dtuple,
1033 : values, isnull);
1034 728452 : }
1035 :
1036 : /*
1037 : * Per-heap-tuple callback for table_index_build_scan with parallelism.
1038 : *
1039 : * A version of the callback used by parallel index builds. The main difference
1040 : * is that instead of writing the BRIN tuples into the index, we write them
1041 : * into a shared tuplesort, and leave the insertion up to the leader (which may
1042 : * reorder them a bit etc.). The callback also does not generate empty ranges,
1043 : * those will be added by the leader when merging results from workers.
1044 : */
1045 : static void
1046 7962 : brinbuildCallbackParallel(Relation index,
1047 : ItemPointer tid,
1048 : Datum *values,
1049 : bool *isnull,
1050 : bool tupleIsAlive,
1051 : void *brstate)
1052 : {
1053 7962 : BrinBuildState *state = (BrinBuildState *) brstate;
1054 : BlockNumber thisblock;
1055 :
1056 7962 : thisblock = ItemPointerGetBlockNumber(tid);
1057 :
1058 : /*
1059 : * If we're in a block that belongs to a different range, summarize what
1060 : * we've got and start afresh. Note the scan might have skipped many
1061 : * pages, if they were devoid of live tuples; we do not create empty BRIN
1062 : * ranges here - the leader is responsible for filling them in.
1063 : *
1064 : * Unlike serial builds, parallel index builds allow synchronized seqscans
1065 : * (because that's what parallel scans do). This means the block may wrap
1066 : * around to the beginning of the relation, so the condition needs to
1067 : * check for both future and past ranges.
1068 : */
1069 7962 : if ((thisblock < state->bs_currRangeStart) ||
1070 7962 : (thisblock > state->bs_currRangeStart + state->bs_pagesPerRange - 1))
1071 : {
1072 :
1073 : BRIN_elog((DEBUG2,
1074 : "brinbuildCallbackParallel: completed a range: %u--%u",
1075 : state->bs_currRangeStart,
1076 : state->bs_currRangeStart + state->bs_pagesPerRange));
1077 :
1078 : /* create the index tuple and write it into the tuplesort */
1079 40 : form_and_spill_tuple(state);
1080 :
1081 : /*
1082 : * Set state to correspond to the next range (for this block).
1083 : *
1084 : * This skips ranges that are either empty (and so we don't get any
1085 : * tuples to summarize), or processed by other workers. We can't
1086 : * differentiate those cases here easily, so we leave it up to the
1087 : * leader to fill empty ranges where needed.
1088 : */
1089 : state->bs_currRangeStart
1090 40 : = state->bs_pagesPerRange * (thisblock / state->bs_pagesPerRange);
1091 :
1092 : /* re-initialize state for it */
1093 40 : brin_memtuple_initialize(state->bs_dtuple, state->bs_bdesc);
1094 : }
1095 :
1096 : /* Accumulate the current tuple into the running state */
1097 7962 : (void) add_values_to_range(index, state->bs_bdesc, state->bs_dtuple,
1098 : values, isnull);
1099 7962 : }
1100 :
1101 : /*
1102 : * brinbuild() -- build a new BRIN index.
1103 : */
1104 : IndexBuildResult *
1105 368 : brinbuild(Relation heap, Relation index, IndexInfo *indexInfo)
1106 : {
1107 : IndexBuildResult *result;
1108 : double reltuples;
1109 : double idxtuples;
1110 : BrinRevmap *revmap;
1111 : BrinBuildState *state;
1112 : Buffer meta;
1113 : BlockNumber pagesPerRange;
1114 :
1115 : /*
1116 : * We expect to be called exactly once for any index relation.
1117 : */
1118 368 : if (RelationGetNumberOfBlocks(index) != 0)
1119 0 : elog(ERROR, "index \"%s\" already contains data",
1120 : RelationGetRelationName(index));
1121 :
1122 : /*
1123 : * Critical section not required, because on error the creation of the
1124 : * whole relation will be rolled back.
1125 : */
1126 :
1127 368 : meta = ExtendBufferedRel(BMR_REL(index), MAIN_FORKNUM, NULL,
1128 : EB_LOCK_FIRST | EB_SKIP_EXTENSION_LOCK);
1129 : Assert(BufferGetBlockNumber(meta) == BRIN_METAPAGE_BLKNO);
1130 :
1131 368 : brin_metapage_init(BufferGetPage(meta), BrinGetPagesPerRange(index),
1132 : BRIN_CURRENT_VERSION);
1133 368 : MarkBufferDirty(meta);
1134 :
1135 368 : if (RelationNeedsWAL(index))
1136 : {
1137 : xl_brin_createidx xlrec;
1138 : XLogRecPtr recptr;
1139 : Page page;
1140 :
1141 180 : xlrec.version = BRIN_CURRENT_VERSION;
1142 180 : xlrec.pagesPerRange = BrinGetPagesPerRange(index);
1143 :
1144 180 : XLogBeginInsert();
1145 180 : XLogRegisterData(&xlrec, SizeOfBrinCreateIdx);
1146 180 : XLogRegisterBuffer(0, meta, REGBUF_WILL_INIT | REGBUF_STANDARD);
1147 :
1148 180 : recptr = XLogInsert(RM_BRIN_ID, XLOG_BRIN_CREATE_INDEX);
1149 :
1150 180 : page = BufferGetPage(meta);
1151 180 : PageSetLSN(page, recptr);
1152 : }
1153 :
1154 368 : UnlockReleaseBuffer(meta);
1155 :
1156 : /*
1157 : * Initialize our state, including the deformed tuple state.
1158 : */
1159 368 : revmap = brinRevmapInitialize(index, &pagesPerRange);
1160 368 : state = initialize_brin_buildstate(index, revmap, pagesPerRange,
1161 : RelationGetNumberOfBlocks(heap));
1162 :
1163 : /*
1164 : * Attempt to launch parallel worker scan when required
1165 : *
1166 : * XXX plan_create_index_workers makes the number of workers dependent on
1167 : * maintenance_work_mem, requiring 32MB for each worker. That makes sense
1168 : * for btree, but not for BRIN, which can do with much less memory. So
1169 : * maybe make that somehow less strict, optionally?
1170 : */
1171 368 : if (indexInfo->ii_ParallelWorkers > 0)
1172 10 : _brin_begin_parallel(state, heap, index, indexInfo->ii_Concurrent,
1173 : indexInfo->ii_ParallelWorkers);
1174 :
1175 : /*
1176 : * If parallel build requested and at least one worker process was
1177 : * successfully launched, set up coordination state, wait for workers to
1178 : * complete. Then read all tuples from the shared tuplesort and insert
1179 : * them into the index.
1180 : *
1181 : * In serial mode, simply scan the table and build the index one index
1182 : * tuple at a time.
1183 : */
1184 368 : if (state->bs_leader)
1185 : {
1186 : SortCoordinate coordinate;
1187 :
1188 8 : coordinate = (SortCoordinate) palloc0(sizeof(SortCoordinateData));
1189 8 : coordinate->isWorker = false;
1190 8 : coordinate->nParticipants =
1191 8 : state->bs_leader->nparticipanttuplesorts;
1192 8 : coordinate->sharedsort = state->bs_leader->sharedsort;
1193 :
1194 : /*
1195 : * Begin leader tuplesort.
1196 : *
1197 : * In cases where parallelism is involved, the leader receives the
1198 : * same share of maintenance_work_mem as a serial sort (it is
1199 : * generally treated in the same way as a serial sort once we return).
1200 : * Parallel worker Tuplesortstates will have received only a fraction
1201 : * of maintenance_work_mem, though.
1202 : *
1203 : * We rely on the lifetime of the Leader Tuplesortstate almost not
1204 : * overlapping with any worker Tuplesortstate's lifetime. There may
1205 : * be some small overlap, but that's okay because we rely on leader
1206 : * Tuplesortstate only allocating a small, fixed amount of memory
1207 : * here. When its tuplesort_performsort() is called (by our caller),
1208 : * and significant amounts of memory are likely to be used, all
1209 : * workers must have already freed almost all memory held by their
1210 : * Tuplesortstates (they are about to go away completely, too). The
1211 : * overall effect is that maintenance_work_mem always represents an
1212 : * absolute high watermark on the amount of memory used by a CREATE
1213 : * INDEX operation, regardless of the use of parallelism or any other
1214 : * factor.
1215 : */
1216 8 : state->bs_sortstate =
1217 8 : tuplesort_begin_index_brin(maintenance_work_mem, coordinate,
1218 : TUPLESORT_NONE);
1219 :
1220 : /* scan the relation and merge per-worker results */
1221 8 : reltuples = _brin_parallel_merge(state);
1222 :
1223 8 : _brin_end_parallel(state->bs_leader, state);
1224 : }
1225 : else /* no parallel index build */
1226 : {
1227 : /*
1228 : * Now scan the relation. No syncscan allowed here because we want
1229 : * the heap blocks in physical order (we want to produce the ranges
1230 : * starting from block 0, and the callback also relies on this to not
1231 : * generate summary for the same range twice).
1232 : */
1233 360 : reltuples = table_index_build_scan(heap, index, indexInfo, false, true,
1234 : brinbuildCallback, state, NULL);
1235 :
1236 : /*
1237 : * process the final batch
1238 : *
1239 : * XXX Note this does not update state->bs_currRangeStart, i.e. it
1240 : * stays set to the last range added to the index. This is OK, because
1241 : * that's what brin_fill_empty_ranges expects.
1242 : */
1243 360 : form_and_insert_tuple(state);
1244 :
1245 : /*
1246 : * Backfill the final ranges with empty data.
1247 : *
1248 : * This saves us from doing what amounts to full table scans when the
1249 : * index with a predicate like WHERE (nonnull_column IS NULL), or
1250 : * other very selective predicates.
1251 : */
1252 360 : brin_fill_empty_ranges(state,
1253 : state->bs_currRangeStart,
1254 : state->bs_maxRangeStart);
1255 : }
1256 :
1257 : /* release resources */
1258 368 : idxtuples = state->bs_numtuples;
1259 368 : brinRevmapTerminate(state->bs_rmAccess);
1260 368 : terminate_brin_buildstate(state);
1261 :
1262 : /*
1263 : * Return statistics
1264 : */
1265 368 : result = palloc_object(IndexBuildResult);
1266 :
1267 368 : result->heap_tuples = reltuples;
1268 368 : result->index_tuples = idxtuples;
1269 :
1270 368 : return result;
1271 : }
1272 :
1273 : void
1274 6 : brinbuildempty(Relation index)
1275 : {
1276 : Buffer metabuf;
1277 :
1278 : /* An empty BRIN index has a metapage only. */
1279 6 : metabuf = ExtendBufferedRel(BMR_REL(index), INIT_FORKNUM, NULL,
1280 : EB_LOCK_FIRST | EB_SKIP_EXTENSION_LOCK);
1281 :
1282 : /* Initialize and xlog metabuffer. */
1283 6 : START_CRIT_SECTION();
1284 6 : brin_metapage_init(BufferGetPage(metabuf), BrinGetPagesPerRange(index),
1285 : BRIN_CURRENT_VERSION);
1286 6 : MarkBufferDirty(metabuf);
1287 6 : log_newpage_buffer(metabuf, true);
1288 6 : END_CRIT_SECTION();
1289 :
1290 6 : UnlockReleaseBuffer(metabuf);
1291 6 : }
1292 :
1293 : /*
1294 : * brinbulkdelete
1295 : * Since there are no per-heap-tuple index tuples in BRIN indexes,
1296 : * there's not a lot we can do here.
1297 : *
1298 : * XXX we could mark item tuples as "dirty" (when a minimum or maximum heap
1299 : * tuple is deleted), meaning the need to re-run summarization on the affected
1300 : * range. Would need to add an extra flag in brintuples for that.
1301 : */
1302 : IndexBulkDeleteResult *
1303 26 : brinbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
1304 : IndexBulkDeleteCallback callback, void *callback_state)
1305 : {
1306 : /* allocate stats if first time through, else re-use existing struct */
1307 26 : if (stats == NULL)
1308 26 : stats = palloc0_object(IndexBulkDeleteResult);
1309 :
1310 26 : return stats;
1311 : }
1312 :
1313 : /*
1314 : * This routine is in charge of "vacuuming" a BRIN index: we just summarize
1315 : * ranges that are currently unsummarized.
1316 : */
1317 : IndexBulkDeleteResult *
1318 100 : brinvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats)
1319 : {
1320 : Relation heapRel;
1321 :
1322 : /* No-op in ANALYZE ONLY mode */
1323 100 : if (info->analyze_only)
1324 6 : return stats;
1325 :
1326 94 : if (!stats)
1327 74 : stats = palloc0_object(IndexBulkDeleteResult);
1328 94 : stats->num_pages = RelationGetNumberOfBlocks(info->index);
1329 : /* rest of stats is initialized by zeroing */
1330 :
1331 94 : heapRel = table_open(IndexGetRelation(RelationGetRelid(info->index), false),
1332 : AccessShareLock);
1333 :
1334 94 : brin_vacuum_scan(info->index, info->strategy);
1335 :
1336 94 : brinsummarize(info->index, heapRel, BRIN_ALL_BLOCKRANGES, false,
1337 : &stats->num_index_tuples, &stats->num_index_tuples);
1338 :
1339 94 : table_close(heapRel, AccessShareLock);
1340 :
1341 94 : return stats;
1342 : }
1343 :
1344 : /*
1345 : * reloptions processor for BRIN indexes
1346 : */
1347 : bytea *
1348 1160 : brinoptions(Datum reloptions, bool validate)
1349 : {
1350 : static const relopt_parse_elt tab[] = {
1351 : {"pages_per_range", RELOPT_TYPE_INT, offsetof(BrinOptions, pagesPerRange)},
1352 : {"autosummarize", RELOPT_TYPE_BOOL, offsetof(BrinOptions, autosummarize)}
1353 : };
1354 :
1355 1160 : return (bytea *) build_reloptions(reloptions, validate,
1356 : RELOPT_KIND_BRIN,
1357 : sizeof(BrinOptions),
1358 : tab, lengthof(tab));
1359 : }
1360 :
1361 : /*
1362 : * SQL-callable function to scan through an index and summarize all ranges
1363 : * that are not currently summarized.
1364 : */
1365 : Datum
1366 76 : brin_summarize_new_values(PG_FUNCTION_ARGS)
1367 : {
1368 76 : Datum relation = PG_GETARG_DATUM(0);
1369 :
1370 76 : return DirectFunctionCall2(brin_summarize_range,
1371 : relation,
1372 : Int64GetDatum((int64) BRIN_ALL_BLOCKRANGES));
1373 : }
1374 :
1375 : /*
1376 : * SQL-callable function to summarize the indicated page range, if not already
1377 : * summarized. If the second argument is BRIN_ALL_BLOCKRANGES, all
1378 : * unsummarized ranges are summarized.
1379 : */
1380 : Datum
1381 210 : brin_summarize_range(PG_FUNCTION_ARGS)
1382 : {
1383 210 : Oid indexoid = PG_GETARG_OID(0);
1384 210 : int64 heapBlk64 = PG_GETARG_INT64(1);
1385 : BlockNumber heapBlk;
1386 : Oid heapoid;
1387 : Relation indexRel;
1388 : Relation heapRel;
1389 : Oid save_userid;
1390 : int save_sec_context;
1391 : int save_nestlevel;
1392 210 : double numSummarized = 0;
1393 :
1394 210 : if (RecoveryInProgress())
1395 0 : ereport(ERROR,
1396 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1397 : errmsg("recovery is in progress"),
1398 : errhint("BRIN control functions cannot be executed during recovery.")));
1399 :
1400 210 : if (heapBlk64 > BRIN_ALL_BLOCKRANGES || heapBlk64 < 0)
1401 36 : ereport(ERROR,
1402 : (errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
1403 : errmsg("block number out of range: %" PRId64, heapBlk64)));
1404 174 : heapBlk = (BlockNumber) heapBlk64;
1405 :
1406 : /*
1407 : * We must lock table before index to avoid deadlocks. However, if the
1408 : * passed indexoid isn't an index then IndexGetRelation() will fail.
1409 : * Rather than emitting a not-very-helpful error message, postpone
1410 : * complaining, expecting that the is-it-an-index test below will fail.
1411 : */
1412 174 : heapoid = IndexGetRelation(indexoid, true);
1413 174 : if (OidIsValid(heapoid))
1414 : {
1415 156 : heapRel = table_open(heapoid, ShareUpdateExclusiveLock);
1416 :
1417 : /*
1418 : * Autovacuum calls us. For its benefit, switch to the table owner's
1419 : * userid, so that any index functions are run as that user. Also
1420 : * lock down security-restricted operations and arrange to make GUC
1421 : * variable changes local to this command. This is harmless, albeit
1422 : * unnecessary, when called from SQL, because we fail shortly if the
1423 : * user does not own the index.
1424 : */
1425 156 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
1426 156 : SetUserIdAndSecContext(heapRel->rd_rel->relowner,
1427 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
1428 156 : save_nestlevel = NewGUCNestLevel();
1429 156 : RestrictSearchPath();
1430 : }
1431 : else
1432 : {
1433 18 : heapRel = NULL;
1434 : /* Set these just to suppress "uninitialized variable" warnings */
1435 18 : save_userid = InvalidOid;
1436 18 : save_sec_context = -1;
1437 18 : save_nestlevel = -1;
1438 : }
1439 :
1440 174 : indexRel = index_open(indexoid, ShareUpdateExclusiveLock);
1441 :
1442 : /* Must be a BRIN index */
1443 156 : if (indexRel->rd_rel->relkind != RELKIND_INDEX ||
1444 156 : indexRel->rd_rel->relam != BRIN_AM_OID)
1445 18 : ereport(ERROR,
1446 : (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1447 : errmsg("\"%s\" is not a BRIN index",
1448 : RelationGetRelationName(indexRel))));
1449 :
1450 : /* User must own the index (comparable to privileges needed for VACUUM) */
1451 138 : if (heapRel != NULL && !object_ownercheck(RelationRelationId, indexoid, save_userid))
1452 0 : aclcheck_error(ACLCHECK_NOT_OWNER, OBJECT_INDEX,
1453 0 : RelationGetRelationName(indexRel));
1454 :
1455 : /*
1456 : * Since we did the IndexGetRelation call above without any lock, it's
1457 : * barely possible that a race against an index drop/recreation could have
1458 : * netted us the wrong table. Recheck.
1459 : */
1460 138 : if (heapRel == NULL || heapoid != IndexGetRelation(indexoid, false))
1461 0 : ereport(ERROR,
1462 : (errcode(ERRCODE_UNDEFINED_TABLE),
1463 : errmsg("could not open parent table of index \"%s\"",
1464 : RelationGetRelationName(indexRel))));
1465 :
1466 : /* see gin_clean_pending_list() */
1467 138 : if (indexRel->rd_index->indisvalid)
1468 138 : brinsummarize(indexRel, heapRel, heapBlk, true, &numSummarized, NULL);
1469 : else
1470 0 : ereport(DEBUG1,
1471 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1472 : errmsg("index \"%s\" is not valid",
1473 : RelationGetRelationName(indexRel))));
1474 :
1475 : /* Roll back any GUC changes executed by index functions */
1476 138 : AtEOXact_GUC(false, save_nestlevel);
1477 :
1478 : /* Restore userid and security context */
1479 138 : SetUserIdAndSecContext(save_userid, save_sec_context);
1480 :
1481 138 : relation_close(indexRel, ShareUpdateExclusiveLock);
1482 138 : relation_close(heapRel, ShareUpdateExclusiveLock);
1483 :
1484 138 : PG_RETURN_INT32((int32) numSummarized);
1485 : }
1486 :
1487 : /*
1488 : * SQL-callable interface to mark a range as no longer summarized
1489 : */
1490 : Datum
1491 104 : brin_desummarize_range(PG_FUNCTION_ARGS)
1492 : {
1493 104 : Oid indexoid = PG_GETARG_OID(0);
1494 104 : int64 heapBlk64 = PG_GETARG_INT64(1);
1495 : BlockNumber heapBlk;
1496 : Oid heapoid;
1497 : Relation heapRel;
1498 : Relation indexRel;
1499 : bool done;
1500 :
1501 104 : if (RecoveryInProgress())
1502 0 : ereport(ERROR,
1503 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1504 : errmsg("recovery is in progress"),
1505 : errhint("BRIN control functions cannot be executed during recovery.")));
1506 :
1507 104 : if (heapBlk64 > MaxBlockNumber || heapBlk64 < 0)
1508 18 : ereport(ERROR,
1509 : (errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
1510 : errmsg("block number out of range: %" PRId64,
1511 : heapBlk64)));
1512 86 : heapBlk = (BlockNumber) heapBlk64;
1513 :
1514 : /*
1515 : * We must lock table before index to avoid deadlocks. However, if the
1516 : * passed indexoid isn't an index then IndexGetRelation() will fail.
1517 : * Rather than emitting a not-very-helpful error message, postpone
1518 : * complaining, expecting that the is-it-an-index test below will fail.
1519 : *
1520 : * Unlike brin_summarize_range(), autovacuum never calls this. Hence, we
1521 : * don't switch userid.
1522 : */
1523 86 : heapoid = IndexGetRelation(indexoid, true);
1524 86 : if (OidIsValid(heapoid))
1525 86 : heapRel = table_open(heapoid, ShareUpdateExclusiveLock);
1526 : else
1527 0 : heapRel = NULL;
1528 :
1529 86 : indexRel = index_open(indexoid, ShareUpdateExclusiveLock);
1530 :
1531 : /* Must be a BRIN index */
1532 86 : if (indexRel->rd_rel->relkind != RELKIND_INDEX ||
1533 86 : indexRel->rd_rel->relam != BRIN_AM_OID)
1534 0 : ereport(ERROR,
1535 : (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1536 : errmsg("\"%s\" is not a BRIN index",
1537 : RelationGetRelationName(indexRel))));
1538 :
1539 : /* User must own the index (comparable to privileges needed for VACUUM) */
1540 86 : if (!object_ownercheck(RelationRelationId, indexoid, GetUserId()))
1541 0 : aclcheck_error(ACLCHECK_NOT_OWNER, OBJECT_INDEX,
1542 0 : RelationGetRelationName(indexRel));
1543 :
1544 : /*
1545 : * Since we did the IndexGetRelation call above without any lock, it's
1546 : * barely possible that a race against an index drop/recreation could have
1547 : * netted us the wrong table. Recheck.
1548 : */
1549 86 : if (heapRel == NULL || heapoid != IndexGetRelation(indexoid, false))
1550 0 : ereport(ERROR,
1551 : (errcode(ERRCODE_UNDEFINED_TABLE),
1552 : errmsg("could not open parent table of index \"%s\"",
1553 : RelationGetRelationName(indexRel))));
1554 :
1555 : /* see gin_clean_pending_list() */
1556 86 : if (indexRel->rd_index->indisvalid)
1557 : {
1558 : /* the revmap does the hard work */
1559 : do
1560 : {
1561 86 : done = brinRevmapDesummarizeRange(indexRel, heapBlk);
1562 : }
1563 86 : while (!done);
1564 : }
1565 : else
1566 0 : ereport(DEBUG1,
1567 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1568 : errmsg("index \"%s\" is not valid",
1569 : RelationGetRelationName(indexRel))));
1570 :
1571 86 : relation_close(indexRel, ShareUpdateExclusiveLock);
1572 86 : relation_close(heapRel, ShareUpdateExclusiveLock);
1573 :
1574 86 : PG_RETURN_VOID();
1575 : }
1576 :
1577 : /*
1578 : * Build a BrinDesc used to create or scan a BRIN index
1579 : */
1580 : BrinDesc *
1581 4582 : brin_build_desc(Relation rel)
1582 : {
1583 : BrinOpcInfo **opcinfo;
1584 : BrinDesc *bdesc;
1585 : TupleDesc tupdesc;
1586 4582 : int totalstored = 0;
1587 : int keyno;
1588 : long totalsize;
1589 : MemoryContext cxt;
1590 : MemoryContext oldcxt;
1591 :
1592 4582 : cxt = AllocSetContextCreate(CurrentMemoryContext,
1593 : "brin desc cxt",
1594 : ALLOCSET_SMALL_SIZES);
1595 4582 : oldcxt = MemoryContextSwitchTo(cxt);
1596 4582 : tupdesc = RelationGetDescr(rel);
1597 :
1598 : /*
1599 : * Obtain BrinOpcInfo for each indexed column. While at it, accumulate
1600 : * the number of columns stored, since the number is opclass-defined.
1601 : */
1602 4582 : opcinfo = palloc_array(BrinOpcInfo *, tupdesc->natts);
1603 76082 : for (keyno = 0; keyno < tupdesc->natts; keyno++)
1604 : {
1605 : FmgrInfo *opcInfoFn;
1606 71500 : Form_pg_attribute attr = TupleDescAttr(tupdesc, keyno);
1607 :
1608 71500 : opcInfoFn = index_getprocinfo(rel, keyno + 1, BRIN_PROCNUM_OPCINFO);
1609 :
1610 143000 : opcinfo[keyno] = (BrinOpcInfo *)
1611 71500 : DatumGetPointer(FunctionCall1(opcInfoFn, ObjectIdGetDatum(attr->atttypid)));
1612 71500 : totalstored += opcinfo[keyno]->oi_nstored;
1613 : }
1614 :
1615 : /* Allocate our result struct and fill it in */
1616 4582 : totalsize = offsetof(BrinDesc, bd_info) +
1617 4582 : sizeof(BrinOpcInfo *) * tupdesc->natts;
1618 :
1619 4582 : bdesc = palloc(totalsize);
1620 4582 : bdesc->bd_context = cxt;
1621 4582 : bdesc->bd_index = rel;
1622 4582 : bdesc->bd_tupdesc = tupdesc;
1623 4582 : bdesc->bd_disktdesc = NULL; /* generated lazily */
1624 4582 : bdesc->bd_totalstored = totalstored;
1625 :
1626 76082 : for (keyno = 0; keyno < tupdesc->natts; keyno++)
1627 71500 : bdesc->bd_info[keyno] = opcinfo[keyno];
1628 4582 : pfree(opcinfo);
1629 :
1630 4582 : MemoryContextSwitchTo(oldcxt);
1631 :
1632 4582 : return bdesc;
1633 : }
1634 :
1635 : void
1636 3446 : brin_free_desc(BrinDesc *bdesc)
1637 : {
1638 : /* make sure the tupdesc is still valid */
1639 : Assert(bdesc->bd_tupdesc->tdrefcount >= 1);
1640 : /* no need for retail pfree */
1641 3446 : MemoryContextDelete(bdesc->bd_context);
1642 3446 : }
1643 :
1644 : /*
1645 : * Fetch index's statistical data into *stats
1646 : */
1647 : void
1648 10730 : brinGetStats(Relation index, BrinStatsData *stats)
1649 : {
1650 : Buffer metabuffer;
1651 : Page metapage;
1652 : BrinMetaPageData *metadata;
1653 :
1654 10730 : metabuffer = ReadBuffer(index, BRIN_METAPAGE_BLKNO);
1655 10730 : LockBuffer(metabuffer, BUFFER_LOCK_SHARE);
1656 10730 : metapage = BufferGetPage(metabuffer);
1657 10730 : metadata = (BrinMetaPageData *) PageGetContents(metapage);
1658 :
1659 10730 : stats->pagesPerRange = metadata->pagesPerRange;
1660 10730 : stats->revmapNumPages = metadata->lastRevmapPage - 1;
1661 :
1662 10730 : UnlockReleaseBuffer(metabuffer);
1663 10730 : }
1664 :
1665 : /*
1666 : * Initialize a BrinBuildState appropriate to create tuples on the given index.
1667 : */
1668 : static BrinBuildState *
1669 468 : initialize_brin_buildstate(Relation idxRel, BrinRevmap *revmap,
1670 : BlockNumber pagesPerRange, BlockNumber tablePages)
1671 : {
1672 : BrinBuildState *state;
1673 468 : BlockNumber lastRange = 0;
1674 :
1675 468 : state = palloc_object(BrinBuildState);
1676 :
1677 468 : state->bs_irel = idxRel;
1678 468 : state->bs_numtuples = 0;
1679 468 : state->bs_reltuples = 0;
1680 468 : state->bs_currentInsertBuf = InvalidBuffer;
1681 468 : state->bs_pagesPerRange = pagesPerRange;
1682 468 : state->bs_currRangeStart = 0;
1683 468 : state->bs_rmAccess = revmap;
1684 468 : state->bs_bdesc = brin_build_desc(idxRel);
1685 468 : state->bs_dtuple = brin_new_memtuple(state->bs_bdesc);
1686 468 : state->bs_leader = NULL;
1687 468 : state->bs_worker_id = 0;
1688 468 : state->bs_sortstate = NULL;
1689 468 : state->bs_context = CurrentMemoryContext;
1690 468 : state->bs_emptyTuple = NULL;
1691 468 : state->bs_emptyTupleLen = 0;
1692 :
1693 : /* Remember the memory context to use for an empty tuple, if needed. */
1694 468 : state->bs_context = CurrentMemoryContext;
1695 468 : state->bs_emptyTuple = NULL;
1696 468 : state->bs_emptyTupleLen = 0;
1697 :
1698 : /*
1699 : * Calculate the start of the last page range. Page numbers are 0-based,
1700 : * so to calculate the index we need to subtract one. The integer division
1701 : * gives us the index of the page range.
1702 : */
1703 468 : if (tablePages > 0)
1704 344 : lastRange = ((tablePages - 1) / pagesPerRange) * pagesPerRange;
1705 :
1706 : /* Now calculate the start of the next range. */
1707 468 : state->bs_maxRangeStart = lastRange + state->bs_pagesPerRange;
1708 :
1709 468 : return state;
1710 : }
1711 :
1712 : /*
1713 : * Release resources associated with a BrinBuildState.
1714 : */
1715 : static void
1716 456 : terminate_brin_buildstate(BrinBuildState *state)
1717 : {
1718 : /*
1719 : * Release the last index buffer used. We might as well ensure that
1720 : * whatever free space remains in that page is available in FSM, too.
1721 : */
1722 456 : if (!BufferIsInvalid(state->bs_currentInsertBuf))
1723 : {
1724 : Page page;
1725 : Size freespace;
1726 : BlockNumber blk;
1727 :
1728 368 : page = BufferGetPage(state->bs_currentInsertBuf);
1729 368 : freespace = PageGetFreeSpace(page);
1730 368 : blk = BufferGetBlockNumber(state->bs_currentInsertBuf);
1731 368 : ReleaseBuffer(state->bs_currentInsertBuf);
1732 368 : RecordPageWithFreeSpace(state->bs_irel, blk, freespace);
1733 368 : FreeSpaceMapVacuumRange(state->bs_irel, blk, blk + 1);
1734 : }
1735 :
1736 456 : brin_free_desc(state->bs_bdesc);
1737 456 : pfree(state->bs_dtuple);
1738 456 : pfree(state);
1739 456 : }
1740 :
1741 : /*
1742 : * On the given BRIN index, summarize the heap page range that corresponds
1743 : * to the heap block number given.
1744 : *
1745 : * This routine can run in parallel with insertions into the heap. To avoid
1746 : * missing those values from the summary tuple, we first insert a placeholder
1747 : * index tuple into the index, then execute the heap scan; transactions
1748 : * concurrent with the scan update the placeholder tuple. After the scan, we
1749 : * union the placeholder tuple with the one computed by this routine. The
1750 : * update of the index value happens in a loop, so that if somebody updates
1751 : * the placeholder tuple after we read it, we detect the case and try again.
1752 : * This ensures that the concurrently inserted tuples are not lost.
1753 : *
1754 : * A further corner case is this routine being asked to summarize the partial
1755 : * range at the end of the table. heapNumBlocks is the (possibly outdated)
1756 : * table size; if we notice that the requested range lies beyond that size,
1757 : * we re-compute the table size after inserting the placeholder tuple, to
1758 : * avoid missing pages that were appended recently.
1759 : */
1760 : static void
1761 2944 : summarize_range(IndexInfo *indexInfo, BrinBuildState *state, Relation heapRel,
1762 : BlockNumber heapBlk, BlockNumber heapNumBlks)
1763 : {
1764 : Buffer phbuf;
1765 : BrinTuple *phtup;
1766 : Size phsz;
1767 : OffsetNumber offset;
1768 : BlockNumber scanNumBlks;
1769 :
1770 : /*
1771 : * Insert the placeholder tuple
1772 : */
1773 2944 : phbuf = InvalidBuffer;
1774 2944 : phtup = brin_form_placeholder_tuple(state->bs_bdesc, heapBlk, &phsz);
1775 2944 : offset = brin_doinsert(state->bs_irel, state->bs_pagesPerRange,
1776 : state->bs_rmAccess, &phbuf,
1777 : heapBlk, phtup, phsz);
1778 :
1779 : /*
1780 : * Compute range end. We hold ShareUpdateExclusive lock on table, so it
1781 : * cannot shrink concurrently (but it can grow).
1782 : */
1783 : Assert(heapBlk % state->bs_pagesPerRange == 0);
1784 2944 : if (heapBlk + state->bs_pagesPerRange > heapNumBlks)
1785 : {
1786 : /*
1787 : * If we're asked to scan what we believe to be the final range on the
1788 : * table (i.e. a range that might be partial) we need to recompute our
1789 : * idea of what the latest page is after inserting the placeholder
1790 : * tuple. Anyone that grows the table later will update the
1791 : * placeholder tuple, so it doesn't matter that we won't scan these
1792 : * pages ourselves. Careful: the table might have been extended
1793 : * beyond the current range, so clamp our result.
1794 : *
1795 : * Fortunately, this should occur infrequently.
1796 : */
1797 24 : scanNumBlks = Min(RelationGetNumberOfBlocks(heapRel) - heapBlk,
1798 : state->bs_pagesPerRange);
1799 : }
1800 : else
1801 : {
1802 : /* Easy case: range is known to be complete */
1803 2920 : scanNumBlks = state->bs_pagesPerRange;
1804 : }
1805 :
1806 : /*
1807 : * Execute the partial heap scan covering the heap blocks in the specified
1808 : * page range, summarizing the heap tuples in it. This scan stops just
1809 : * short of brinbuildCallback creating the new index entry.
1810 : *
1811 : * Note that it is critical we use the "any visible" mode of
1812 : * table_index_build_range_scan here: otherwise, we would miss tuples
1813 : * inserted by transactions that are still in progress, among other corner
1814 : * cases.
1815 : */
1816 2944 : state->bs_currRangeStart = heapBlk;
1817 2944 : table_index_build_range_scan(heapRel, state->bs_irel, indexInfo, false, true, false,
1818 : heapBlk, scanNumBlks,
1819 : brinbuildCallback, state, NULL);
1820 :
1821 : /*
1822 : * Now we update the values obtained by the scan with the placeholder
1823 : * tuple. We do this in a loop which only terminates if we're able to
1824 : * update the placeholder tuple successfully; if we are not, this means
1825 : * somebody else modified the placeholder tuple after we read it.
1826 : */
1827 : for (;;)
1828 0 : {
1829 : BrinTuple *newtup;
1830 : Size newsize;
1831 : bool didupdate;
1832 : bool samepage;
1833 :
1834 2944 : CHECK_FOR_INTERRUPTS();
1835 :
1836 : /*
1837 : * Update the summary tuple and try to update.
1838 : */
1839 2944 : newtup = brin_form_tuple(state->bs_bdesc,
1840 : heapBlk, state->bs_dtuple, &newsize);
1841 2944 : samepage = brin_can_do_samepage_update(phbuf, phsz, newsize);
1842 : didupdate =
1843 2944 : brin_doupdate(state->bs_irel, state->bs_pagesPerRange,
1844 : state->bs_rmAccess, heapBlk, phbuf, offset,
1845 : phtup, phsz, newtup, newsize, samepage);
1846 2944 : brin_free_tuple(phtup);
1847 2944 : brin_free_tuple(newtup);
1848 :
1849 : /* If the update succeeded, we're done. */
1850 2944 : if (didupdate)
1851 2944 : break;
1852 :
1853 : /*
1854 : * If the update didn't work, it might be because somebody updated the
1855 : * placeholder tuple concurrently. Extract the new version, union it
1856 : * with the values we have from the scan, and start over. (There are
1857 : * other reasons for the update to fail, but it's simple to treat them
1858 : * the same.)
1859 : */
1860 0 : phtup = brinGetTupleForHeapBlock(state->bs_rmAccess, heapBlk, &phbuf,
1861 : &offset, &phsz, BUFFER_LOCK_SHARE);
1862 : /* the placeholder tuple must exist */
1863 0 : if (phtup == NULL)
1864 0 : elog(ERROR, "missing placeholder tuple");
1865 0 : phtup = brin_copy_tuple(phtup, phsz, NULL, NULL);
1866 0 : LockBuffer(phbuf, BUFFER_LOCK_UNLOCK);
1867 :
1868 : /* merge it into the tuple from the heap scan */
1869 0 : union_tuples(state->bs_bdesc, state->bs_dtuple, phtup);
1870 : }
1871 :
1872 2944 : ReleaseBuffer(phbuf);
1873 2944 : }
1874 :
1875 : /*
1876 : * Summarize page ranges that are not already summarized. If pageRange is
1877 : * BRIN_ALL_BLOCKRANGES then the whole table is scanned; otherwise, only the
1878 : * page range containing the given heap page number is scanned.
1879 : * If include_partial is true, then the partial range at the end of the table
1880 : * is summarized, otherwise not.
1881 : *
1882 : * For each new index tuple inserted, *numSummarized (if not NULL) is
1883 : * incremented; for each existing tuple, *numExisting (if not NULL) is
1884 : * incremented.
1885 : */
1886 : static void
1887 232 : brinsummarize(Relation index, Relation heapRel, BlockNumber pageRange,
1888 : bool include_partial, double *numSummarized, double *numExisting)
1889 : {
1890 : BrinRevmap *revmap;
1891 232 : BrinBuildState *state = NULL;
1892 232 : IndexInfo *indexInfo = NULL;
1893 : BlockNumber heapNumBlocks;
1894 : BlockNumber pagesPerRange;
1895 : Buffer buf;
1896 : BlockNumber startBlk;
1897 :
1898 232 : revmap = brinRevmapInitialize(index, &pagesPerRange);
1899 :
1900 : /* determine range of pages to process */
1901 232 : heapNumBlocks = RelationGetNumberOfBlocks(heapRel);
1902 232 : if (pageRange == BRIN_ALL_BLOCKRANGES)
1903 152 : startBlk = 0;
1904 : else
1905 : {
1906 80 : startBlk = (pageRange / pagesPerRange) * pagesPerRange;
1907 80 : heapNumBlocks = Min(heapNumBlocks, startBlk + pagesPerRange);
1908 : }
1909 232 : if (startBlk > heapNumBlocks)
1910 : {
1911 : /* Nothing to do if start point is beyond end of table */
1912 0 : brinRevmapTerminate(revmap);
1913 0 : return;
1914 : }
1915 :
1916 : /*
1917 : * Scan the revmap to find unsummarized items.
1918 : */
1919 232 : buf = InvalidBuffer;
1920 19238 : for (; startBlk < heapNumBlocks; startBlk += pagesPerRange)
1921 : {
1922 : BrinTuple *tup;
1923 : OffsetNumber off;
1924 :
1925 : /*
1926 : * Unless requested to summarize even a partial range, go away now if
1927 : * we think the next range is partial. Caller would pass true when it
1928 : * is typically run once bulk data loading is done
1929 : * (brin_summarize_new_values), and false when it is typically the
1930 : * result of arbitrarily-scheduled maintenance command (vacuuming).
1931 : */
1932 19076 : if (!include_partial &&
1933 2328 : (startBlk + pagesPerRange > heapNumBlocks))
1934 70 : break;
1935 :
1936 19006 : CHECK_FOR_INTERRUPTS();
1937 :
1938 19006 : tup = brinGetTupleForHeapBlock(revmap, startBlk, &buf, &off, NULL,
1939 : BUFFER_LOCK_SHARE);
1940 19006 : if (tup == NULL)
1941 : {
1942 : /* no revmap entry for this heap range. Summarize it. */
1943 2944 : if (state == NULL)
1944 : {
1945 : /* first time through */
1946 : Assert(!indexInfo);
1947 88 : state = initialize_brin_buildstate(index, revmap,
1948 : pagesPerRange,
1949 : InvalidBlockNumber);
1950 88 : indexInfo = BuildIndexInfo(index);
1951 : }
1952 2944 : summarize_range(indexInfo, state, heapRel, startBlk, heapNumBlocks);
1953 :
1954 : /* and re-initialize state for the next range */
1955 2944 : brin_memtuple_initialize(state->bs_dtuple, state->bs_bdesc);
1956 :
1957 2944 : if (numSummarized)
1958 2944 : *numSummarized += 1.0;
1959 : }
1960 : else
1961 : {
1962 16062 : if (numExisting)
1963 2162 : *numExisting += 1.0;
1964 16062 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
1965 : }
1966 : }
1967 :
1968 232 : if (BufferIsValid(buf))
1969 160 : ReleaseBuffer(buf);
1970 :
1971 : /* free resources */
1972 232 : brinRevmapTerminate(revmap);
1973 232 : if (state)
1974 : {
1975 88 : terminate_brin_buildstate(state);
1976 88 : pfree(indexInfo);
1977 : }
1978 : }
1979 :
1980 : /*
1981 : * Given a deformed tuple in the build state, convert it into the on-disk
1982 : * format and insert it into the index, making the revmap point to it.
1983 : */
1984 : static void
1985 2656 : form_and_insert_tuple(BrinBuildState *state)
1986 : {
1987 : BrinTuple *tup;
1988 : Size size;
1989 :
1990 2656 : tup = brin_form_tuple(state->bs_bdesc, state->bs_currRangeStart,
1991 : state->bs_dtuple, &size);
1992 2656 : brin_doinsert(state->bs_irel, state->bs_pagesPerRange, state->bs_rmAccess,
1993 : &state->bs_currentInsertBuf, state->bs_currRangeStart,
1994 : tup, size);
1995 2656 : state->bs_numtuples++;
1996 :
1997 2656 : pfree(tup);
1998 2656 : }
1999 :
2000 : /*
2001 : * Given a deformed tuple in the build state, convert it into the on-disk
2002 : * format and write it to a (shared) tuplesort (the leader will insert it
2003 : * into the index later).
2004 : */
2005 : static void
2006 60 : form_and_spill_tuple(BrinBuildState *state)
2007 : {
2008 : BrinTuple *tup;
2009 : Size size;
2010 :
2011 : /* don't insert empty tuples in parallel build */
2012 60 : if (state->bs_dtuple->bt_empty_range)
2013 18 : return;
2014 :
2015 42 : tup = brin_form_tuple(state->bs_bdesc, state->bs_currRangeStart,
2016 : state->bs_dtuple, &size);
2017 :
2018 : /* write the BRIN tuple to the tuplesort */
2019 42 : tuplesort_putbrintuple(state->bs_sortstate, tup, size);
2020 :
2021 42 : state->bs_numtuples++;
2022 :
2023 42 : pfree(tup);
2024 : }
2025 :
2026 : /*
2027 : * Given two deformed tuples, adjust the first one so that it's consistent
2028 : * with the summary values in both.
2029 : */
2030 : static void
2031 2 : union_tuples(BrinDesc *bdesc, BrinMemTuple *a, BrinTuple *b)
2032 : {
2033 : int keyno;
2034 : BrinMemTuple *db;
2035 : MemoryContext cxt;
2036 : MemoryContext oldcxt;
2037 :
2038 : /* Use our own memory context to avoid retail pfree */
2039 2 : cxt = AllocSetContextCreate(CurrentMemoryContext,
2040 : "brin union",
2041 : ALLOCSET_DEFAULT_SIZES);
2042 2 : oldcxt = MemoryContextSwitchTo(cxt);
2043 2 : db = brin_deform_tuple(bdesc, b, NULL);
2044 2 : MemoryContextSwitchTo(oldcxt);
2045 :
2046 : /*
2047 : * Check if the ranges are empty.
2048 : *
2049 : * If at least one of them is empty, we don't need to call per-key union
2050 : * functions at all. If "b" is empty, we just use "a" as the result (it
2051 : * might be empty fine, but that's fine). If "a" is empty but "b" is not,
2052 : * we use "b" as the result (but we have to copy the data into "a" first).
2053 : *
2054 : * Only when both ranges are non-empty, we actually do the per-key merge.
2055 : */
2056 :
2057 : /* If "b" is empty - ignore it and just use "a" (even if it's empty etc.). */
2058 2 : if (db->bt_empty_range)
2059 : {
2060 : /* skip the per-key merge */
2061 0 : MemoryContextDelete(cxt);
2062 0 : return;
2063 : }
2064 :
2065 : /*
2066 : * Now we know "b" is not empty. If "a" is empty, then "b" is the result.
2067 : * But we need to copy the data from "b" to "a" first, because that's how
2068 : * we pass result out.
2069 : *
2070 : * We have to copy all the global/per-key flags etc. too.
2071 : */
2072 2 : if (a->bt_empty_range)
2073 : {
2074 0 : for (keyno = 0; keyno < bdesc->bd_tupdesc->natts; keyno++)
2075 : {
2076 : int i;
2077 0 : BrinValues *col_a = &a->bt_columns[keyno];
2078 0 : BrinValues *col_b = &db->bt_columns[keyno];
2079 0 : BrinOpcInfo *opcinfo = bdesc->bd_info[keyno];
2080 :
2081 0 : col_a->bv_allnulls = col_b->bv_allnulls;
2082 0 : col_a->bv_hasnulls = col_b->bv_hasnulls;
2083 :
2084 : /* If "b" has no data, we're done. */
2085 0 : if (col_b->bv_allnulls)
2086 0 : continue;
2087 :
2088 0 : for (i = 0; i < opcinfo->oi_nstored; i++)
2089 0 : col_a->bv_values[i] =
2090 0 : datumCopy(col_b->bv_values[i],
2091 0 : opcinfo->oi_typcache[i]->typbyval,
2092 0 : opcinfo->oi_typcache[i]->typlen);
2093 : }
2094 :
2095 : /* "a" started empty, but "b" was not empty, so remember that */
2096 0 : a->bt_empty_range = false;
2097 :
2098 : /* skip the per-key merge */
2099 0 : MemoryContextDelete(cxt);
2100 0 : return;
2101 : }
2102 :
2103 : /* Now we know neither range is empty. */
2104 10 : for (keyno = 0; keyno < bdesc->bd_tupdesc->natts; keyno++)
2105 : {
2106 : FmgrInfo *unionFn;
2107 8 : BrinValues *col_a = &a->bt_columns[keyno];
2108 8 : BrinValues *col_b = &db->bt_columns[keyno];
2109 8 : BrinOpcInfo *opcinfo = bdesc->bd_info[keyno];
2110 :
2111 8 : if (opcinfo->oi_regular_nulls)
2112 : {
2113 : /* Does the "b" summary represent any NULL values? */
2114 8 : bool b_has_nulls = (col_b->bv_hasnulls || col_b->bv_allnulls);
2115 :
2116 : /* Adjust "hasnulls". */
2117 8 : if (!col_a->bv_allnulls && b_has_nulls)
2118 0 : col_a->bv_hasnulls = true;
2119 :
2120 : /* If there are no values in B, there's nothing left to do. */
2121 8 : if (col_b->bv_allnulls)
2122 2 : continue;
2123 :
2124 : /*
2125 : * Adjust "allnulls". If A doesn't have values, just copy the
2126 : * values from B into A, and we're done. We cannot run the
2127 : * operators in this case, because values in A might contain
2128 : * garbage. Note we already established that B contains values.
2129 : *
2130 : * Also adjust "hasnulls" in order not to forget the summary
2131 : * represents NULL values. This is not redundant with the earlier
2132 : * update, because that only happens when allnulls=false.
2133 : */
2134 6 : if (col_a->bv_allnulls)
2135 0 : {
2136 : int i;
2137 :
2138 0 : col_a->bv_allnulls = false;
2139 0 : col_a->bv_hasnulls = true;
2140 :
2141 0 : for (i = 0; i < opcinfo->oi_nstored; i++)
2142 0 : col_a->bv_values[i] =
2143 0 : datumCopy(col_b->bv_values[i],
2144 0 : opcinfo->oi_typcache[i]->typbyval,
2145 0 : opcinfo->oi_typcache[i]->typlen);
2146 :
2147 0 : continue;
2148 : }
2149 : }
2150 :
2151 6 : unionFn = index_getprocinfo(bdesc->bd_index, keyno + 1,
2152 : BRIN_PROCNUM_UNION);
2153 6 : FunctionCall3Coll(unionFn,
2154 6 : bdesc->bd_index->rd_indcollation[keyno],
2155 : PointerGetDatum(bdesc),
2156 : PointerGetDatum(col_a),
2157 : PointerGetDatum(col_b));
2158 : }
2159 :
2160 2 : MemoryContextDelete(cxt);
2161 : }
2162 :
2163 : /*
2164 : * brin_vacuum_scan
2165 : * Do a complete scan of the index during VACUUM.
2166 : *
2167 : * This routine scans the complete index looking for uncataloged index pages,
2168 : * i.e. those that might have been lost due to a crash after index extension
2169 : * and such.
2170 : */
2171 : static void
2172 94 : brin_vacuum_scan(Relation idxrel, BufferAccessStrategy strategy)
2173 : {
2174 : BlockRangeReadStreamPrivate p;
2175 : ReadStream *stream;
2176 : Buffer buf;
2177 :
2178 94 : p.current_blocknum = 0;
2179 94 : p.last_exclusive = RelationGetNumberOfBlocks(idxrel);
2180 :
2181 : /*
2182 : * It is safe to use batchmode as block_range_read_stream_cb takes no
2183 : * locks.
2184 : */
2185 94 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE |
2186 : READ_STREAM_FULL |
2187 : READ_STREAM_USE_BATCHING,
2188 : strategy,
2189 : idxrel,
2190 : MAIN_FORKNUM,
2191 : block_range_read_stream_cb,
2192 : &p,
2193 : 0);
2194 :
2195 : /*
2196 : * Scan the index in physical order, and clean up any possible mess in
2197 : * each page.
2198 : */
2199 512 : while ((buf = read_stream_next_buffer(stream, NULL)) != InvalidBuffer)
2200 : {
2201 418 : CHECK_FOR_INTERRUPTS();
2202 :
2203 418 : brin_page_cleanup(idxrel, buf);
2204 :
2205 418 : ReleaseBuffer(buf);
2206 : }
2207 :
2208 94 : read_stream_end(stream);
2209 :
2210 : /*
2211 : * Update all upper pages in the index's FSM, as well. This ensures not
2212 : * only that we propagate leaf-page FSM updates made by brin_page_cleanup,
2213 : * but also that any pre-existing damage or out-of-dateness is repaired.
2214 : */
2215 94 : FreeSpaceMapVacuum(idxrel);
2216 94 : }
2217 :
2218 : static bool
2219 784296 : add_values_to_range(Relation idxRel, BrinDesc *bdesc, BrinMemTuple *dtup,
2220 : const Datum *values, const bool *nulls)
2221 : {
2222 : int keyno;
2223 :
2224 : /* If the range starts empty, we're certainly going to modify it. */
2225 784296 : bool modified = dtup->bt_empty_range;
2226 :
2227 : /*
2228 : * Compare the key values of the new tuple to the stored index values; our
2229 : * deformed tuple will get updated if the new tuple doesn't fit the
2230 : * original range (note this means we can't break out of the loop early).
2231 : * Make a note of whether this happens, so that we know to insert the
2232 : * modified tuple later.
2233 : */
2234 1848422 : for (keyno = 0; keyno < bdesc->bd_tupdesc->natts; keyno++)
2235 : {
2236 : Datum result;
2237 : BrinValues *bval;
2238 : FmgrInfo *addValue;
2239 : bool has_nulls;
2240 :
2241 1064126 : bval = &dtup->bt_columns[keyno];
2242 :
2243 : /*
2244 : * Does the range have actual NULL values? Either of the flags can be
2245 : * set, but we ignore the state before adding first row.
2246 : *
2247 : * We have to remember this, because we'll modify the flags and we
2248 : * need to know if the range started as empty.
2249 : */
2250 2091578 : has_nulls = ((!dtup->bt_empty_range) &&
2251 1027452 : (bval->bv_hasnulls || bval->bv_allnulls));
2252 :
2253 : /*
2254 : * If the value we're adding is NULL, handle it locally. Otherwise
2255 : * call the BRIN_PROCNUM_ADDVALUE procedure.
2256 : */
2257 1064126 : if (bdesc->bd_info[keyno]->oi_regular_nulls && nulls[keyno])
2258 : {
2259 : /*
2260 : * If the new value is null, we record that we saw it if it's the
2261 : * first one; otherwise, there's nothing to do.
2262 : */
2263 18694 : if (!bval->bv_hasnulls)
2264 : {
2265 3610 : bval->bv_hasnulls = true;
2266 3610 : modified = true;
2267 : }
2268 :
2269 18694 : continue;
2270 : }
2271 :
2272 1045432 : addValue = index_getprocinfo(idxRel, keyno + 1,
2273 : BRIN_PROCNUM_ADDVALUE);
2274 1045432 : result = FunctionCall4Coll(addValue,
2275 1045432 : idxRel->rd_indcollation[keyno],
2276 : PointerGetDatum(bdesc),
2277 : PointerGetDatum(bval),
2278 1045432 : values[keyno],
2279 1045432 : BoolGetDatum(nulls[keyno]));
2280 : /* if that returned true, we need to insert the updated tuple */
2281 1045432 : modified |= DatumGetBool(result);
2282 :
2283 : /*
2284 : * If the range was had actual NULL values (i.e. did not start empty),
2285 : * make sure we don't forget about the NULL values. Either the
2286 : * allnulls flag is still set to true, or (if the opclass cleared it)
2287 : * we need to set hasnulls=true.
2288 : *
2289 : * XXX This can only happen when the opclass modified the tuple, so
2290 : * the modified flag should be set.
2291 : */
2292 1045432 : if (has_nulls && !(bval->bv_hasnulls || bval->bv_allnulls))
2293 : {
2294 : Assert(modified);
2295 4 : bval->bv_hasnulls = true;
2296 : }
2297 : }
2298 :
2299 : /*
2300 : * After updating summaries for all the keys, mark it as not empty.
2301 : *
2302 : * If we're actually changing the flag value (i.e. tuple started as
2303 : * empty), we should have modified the tuple. So we should not see empty
2304 : * range that was not modified.
2305 : */
2306 : Assert(!dtup->bt_empty_range || modified);
2307 784296 : dtup->bt_empty_range = false;
2308 :
2309 784296 : return modified;
2310 : }
2311 :
2312 : static bool
2313 189936 : check_null_keys(BrinValues *bval, ScanKey *nullkeys, int nnullkeys)
2314 : {
2315 : int keyno;
2316 :
2317 : /*
2318 : * First check if there are any IS [NOT] NULL scan keys, and if we're
2319 : * violating them.
2320 : */
2321 191172 : for (keyno = 0; keyno < nnullkeys; keyno++)
2322 : {
2323 2232 : ScanKey key = nullkeys[keyno];
2324 :
2325 : Assert(key->sk_attno == bval->bv_attno);
2326 :
2327 : /* Handle only IS NULL/IS NOT NULL tests */
2328 2232 : if (!(key->sk_flags & SK_ISNULL))
2329 0 : continue;
2330 :
2331 2232 : if (key->sk_flags & SK_SEARCHNULL)
2332 : {
2333 : /* IS NULL scan key, but range has no NULLs */
2334 1116 : if (!bval->bv_allnulls && !bval->bv_hasnulls)
2335 978 : return false;
2336 : }
2337 1116 : else if (key->sk_flags & SK_SEARCHNOTNULL)
2338 : {
2339 : /*
2340 : * For IS NOT NULL, we can only skip ranges that are known to have
2341 : * only nulls.
2342 : */
2343 1116 : if (bval->bv_allnulls)
2344 18 : return false;
2345 : }
2346 : else
2347 : {
2348 : /*
2349 : * Neither IS NULL nor IS NOT NULL was used; assume all indexable
2350 : * operators are strict and thus return false with NULL value in
2351 : * the scan key.
2352 : */
2353 0 : return false;
2354 : }
2355 : }
2356 :
2357 188940 : return true;
2358 : }
2359 :
2360 : /*
2361 : * Create parallel context, and launch workers for leader.
2362 : *
2363 : * buildstate argument should be initialized (with the exception of the
2364 : * tuplesort states, which may later be created based on shared
2365 : * state initially set up here).
2366 : *
2367 : * isconcurrent indicates if operation is CREATE INDEX CONCURRENTLY.
2368 : *
2369 : * request is the target number of parallel worker processes to launch.
2370 : *
2371 : * Sets buildstate's BrinLeader, which caller must use to shut down parallel
2372 : * mode by passing it to _brin_end_parallel() at the very end of its index
2373 : * build. If not even a single worker process can be launched, this is
2374 : * never set, and caller should proceed with a serial index build.
2375 : */
2376 : static void
2377 10 : _brin_begin_parallel(BrinBuildState *buildstate, Relation heap, Relation index,
2378 : bool isconcurrent, int request)
2379 : {
2380 : ParallelContext *pcxt;
2381 : int scantuplesortstates;
2382 : Snapshot snapshot;
2383 : Size estbrinshared;
2384 : Size estsort;
2385 : BrinShared *brinshared;
2386 : Sharedsort *sharedsort;
2387 10 : BrinLeader *brinleader = (BrinLeader *) palloc0(sizeof(BrinLeader));
2388 : WalUsage *walusage;
2389 : BufferUsage *bufferusage;
2390 10 : bool leaderparticipates = true;
2391 : int querylen;
2392 :
2393 : #ifdef DISABLE_LEADER_PARTICIPATION
2394 : leaderparticipates = false;
2395 : #endif
2396 :
2397 : /*
2398 : * Enter parallel mode, and create context for parallel build of brin
2399 : * index
2400 : */
2401 10 : EnterParallelMode();
2402 : Assert(request > 0);
2403 10 : pcxt = CreateParallelContext("postgres", "_brin_parallel_build_main",
2404 : request);
2405 :
2406 10 : scantuplesortstates = leaderparticipates ? request + 1 : request;
2407 :
2408 : /*
2409 : * Prepare for scan of the base relation. In a normal index build, we use
2410 : * SnapshotAny because we must retrieve all tuples and do our own time
2411 : * qual checks (because we have to index RECENTLY_DEAD tuples). In a
2412 : * concurrent build, we take a regular MVCC snapshot and index whatever's
2413 : * live according to that.
2414 : */
2415 10 : if (!isconcurrent)
2416 10 : snapshot = SnapshotAny;
2417 : else
2418 0 : snapshot = RegisterSnapshot(GetTransactionSnapshot());
2419 :
2420 : /*
2421 : * Estimate size for our own PARALLEL_KEY_BRIN_SHARED workspace.
2422 : */
2423 10 : estbrinshared = _brin_parallel_estimate_shared(heap, snapshot);
2424 10 : shm_toc_estimate_chunk(&pcxt->estimator, estbrinshared);
2425 10 : estsort = tuplesort_estimate_shared(scantuplesortstates);
2426 10 : shm_toc_estimate_chunk(&pcxt->estimator, estsort);
2427 :
2428 10 : shm_toc_estimate_keys(&pcxt->estimator, 2);
2429 :
2430 : /*
2431 : * Estimate space for WalUsage and BufferUsage -- PARALLEL_KEY_WAL_USAGE
2432 : * and PARALLEL_KEY_BUFFER_USAGE.
2433 : *
2434 : * If there are no extensions loaded that care, we could skip this. We
2435 : * have no way of knowing whether anyone's looking at pgWalUsage or
2436 : * pgBufferUsage, so do it unconditionally.
2437 : */
2438 10 : shm_toc_estimate_chunk(&pcxt->estimator,
2439 : mul_size(sizeof(WalUsage), pcxt->nworkers));
2440 10 : shm_toc_estimate_keys(&pcxt->estimator, 1);
2441 10 : shm_toc_estimate_chunk(&pcxt->estimator,
2442 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
2443 10 : shm_toc_estimate_keys(&pcxt->estimator, 1);
2444 :
2445 : /* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
2446 10 : if (debug_query_string)
2447 : {
2448 10 : querylen = strlen(debug_query_string);
2449 10 : shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
2450 10 : shm_toc_estimate_keys(&pcxt->estimator, 1);
2451 : }
2452 : else
2453 0 : querylen = 0; /* keep compiler quiet */
2454 :
2455 : /* Everyone's had a chance to ask for space, so now create the DSM */
2456 10 : InitializeParallelDSM(pcxt);
2457 :
2458 : /* If no DSM segment was available, back out (do serial build) */
2459 10 : if (pcxt->seg == NULL)
2460 : {
2461 0 : if (IsMVCCSnapshot(snapshot))
2462 0 : UnregisterSnapshot(snapshot);
2463 0 : DestroyParallelContext(pcxt);
2464 0 : ExitParallelMode();
2465 0 : return;
2466 : }
2467 :
2468 : /* Store shared build state, for which we reserved space */
2469 10 : brinshared = (BrinShared *) shm_toc_allocate(pcxt->toc, estbrinshared);
2470 : /* Initialize immutable state */
2471 10 : brinshared->heaprelid = RelationGetRelid(heap);
2472 10 : brinshared->indexrelid = RelationGetRelid(index);
2473 10 : brinshared->isconcurrent = isconcurrent;
2474 10 : brinshared->scantuplesortstates = scantuplesortstates;
2475 10 : brinshared->pagesPerRange = buildstate->bs_pagesPerRange;
2476 10 : brinshared->queryid = pgstat_get_my_query_id();
2477 10 : ConditionVariableInit(&brinshared->workersdonecv);
2478 10 : SpinLockInit(&brinshared->mutex);
2479 :
2480 : /* Initialize mutable state */
2481 10 : brinshared->nparticipantsdone = 0;
2482 10 : brinshared->reltuples = 0.0;
2483 10 : brinshared->indtuples = 0.0;
2484 :
2485 10 : table_parallelscan_initialize(heap,
2486 : ParallelTableScanFromBrinShared(brinshared),
2487 : snapshot);
2488 :
2489 : /*
2490 : * Store shared tuplesort-private state, for which we reserved space.
2491 : * Then, initialize opaque state using tuplesort routine.
2492 : */
2493 10 : sharedsort = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
2494 10 : tuplesort_initialize_shared(sharedsort, scantuplesortstates,
2495 : pcxt->seg);
2496 :
2497 : /*
2498 : * Store shared tuplesort-private state, for which we reserved space.
2499 : * Then, initialize opaque state using tuplesort routine.
2500 : */
2501 10 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_BRIN_SHARED, brinshared);
2502 10 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLESORT, sharedsort);
2503 :
2504 : /* Store query string for workers */
2505 10 : if (debug_query_string)
2506 : {
2507 : char *sharedquery;
2508 :
2509 10 : sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
2510 10 : memcpy(sharedquery, debug_query_string, querylen + 1);
2511 10 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, sharedquery);
2512 : }
2513 :
2514 : /*
2515 : * Allocate space for each worker's WalUsage and BufferUsage; no need to
2516 : * initialize.
2517 : */
2518 10 : walusage = shm_toc_allocate(pcxt->toc,
2519 10 : mul_size(sizeof(WalUsage), pcxt->nworkers));
2520 10 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_WAL_USAGE, walusage);
2521 10 : bufferusage = shm_toc_allocate(pcxt->toc,
2522 10 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
2523 10 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_BUFFER_USAGE, bufferusage);
2524 :
2525 : /* Launch workers, saving status for leader/caller */
2526 10 : LaunchParallelWorkers(pcxt);
2527 10 : brinleader->pcxt = pcxt;
2528 10 : brinleader->nparticipanttuplesorts = pcxt->nworkers_launched;
2529 10 : if (leaderparticipates)
2530 10 : brinleader->nparticipanttuplesorts++;
2531 10 : brinleader->brinshared = brinshared;
2532 10 : brinleader->sharedsort = sharedsort;
2533 10 : brinleader->snapshot = snapshot;
2534 10 : brinleader->walusage = walusage;
2535 10 : brinleader->bufferusage = bufferusage;
2536 :
2537 : /* If no workers were successfully launched, back out (do serial build) */
2538 10 : if (pcxt->nworkers_launched == 0)
2539 : {
2540 2 : _brin_end_parallel(brinleader, NULL);
2541 2 : return;
2542 : }
2543 :
2544 : /* Save leader state now that it's clear build will be parallel */
2545 8 : buildstate->bs_leader = brinleader;
2546 :
2547 : /* Join heap scan ourselves */
2548 8 : if (leaderparticipates)
2549 8 : _brin_leader_participate_as_worker(buildstate, heap, index);
2550 :
2551 : /*
2552 : * Caller needs to wait for all launched workers when we return. Make
2553 : * sure that the failure-to-start case will not hang forever.
2554 : */
2555 8 : WaitForParallelWorkersToAttach(pcxt);
2556 : }
2557 :
2558 : /*
2559 : * Shut down workers, destroy parallel context, and end parallel mode.
2560 : */
2561 : static void
2562 10 : _brin_end_parallel(BrinLeader *brinleader, BrinBuildState *state)
2563 : {
2564 : int i;
2565 :
2566 : /* Shutdown worker processes */
2567 10 : WaitForParallelWorkersToFinish(brinleader->pcxt);
2568 :
2569 : /*
2570 : * Next, accumulate WAL usage. (This must wait for the workers to finish,
2571 : * or we might get incomplete data.)
2572 : */
2573 22 : for (i = 0; i < brinleader->pcxt->nworkers_launched; i++)
2574 12 : InstrAccumParallelQuery(&brinleader->bufferusage[i], &brinleader->walusage[i]);
2575 :
2576 : /* Free last reference to MVCC snapshot, if one was used */
2577 10 : if (IsMVCCSnapshot(brinleader->snapshot))
2578 0 : UnregisterSnapshot(brinleader->snapshot);
2579 10 : DestroyParallelContext(brinleader->pcxt);
2580 10 : ExitParallelMode();
2581 10 : }
2582 :
2583 : /*
2584 : * Within leader, wait for end of heap scan.
2585 : *
2586 : * When called, parallel heap scan started by _brin_begin_parallel() will
2587 : * already be underway within worker processes (when leader participates
2588 : * as a worker, we should end up here just as workers are finishing).
2589 : *
2590 : * Returns the total number of heap tuples scanned.
2591 : */
2592 : static double
2593 8 : _brin_parallel_heapscan(BrinBuildState *state)
2594 : {
2595 8 : BrinShared *brinshared = state->bs_leader->brinshared;
2596 : int nparticipanttuplesorts;
2597 :
2598 8 : nparticipanttuplesorts = state->bs_leader->nparticipanttuplesorts;
2599 : for (;;)
2600 : {
2601 26 : SpinLockAcquire(&brinshared->mutex);
2602 26 : if (brinshared->nparticipantsdone == nparticipanttuplesorts)
2603 : {
2604 : /* copy the data into leader state */
2605 8 : state->bs_reltuples = brinshared->reltuples;
2606 8 : state->bs_numtuples = brinshared->indtuples;
2607 :
2608 8 : SpinLockRelease(&brinshared->mutex);
2609 8 : break;
2610 : }
2611 18 : SpinLockRelease(&brinshared->mutex);
2612 :
2613 18 : ConditionVariableSleep(&brinshared->workersdonecv,
2614 : WAIT_EVENT_PARALLEL_CREATE_INDEX_SCAN);
2615 : }
2616 :
2617 8 : ConditionVariableCancelSleep();
2618 :
2619 8 : return state->bs_reltuples;
2620 : }
2621 :
2622 : /*
2623 : * Within leader, wait for end of heap scan and merge per-worker results.
2624 : *
2625 : * After waiting for all workers to finish, merge the per-worker results into
2626 : * the complete index. The results from each worker are sorted by block number
2627 : * (start of the page range). While combining the per-worker results we merge
2628 : * summaries for the same page range, and also fill-in empty summaries for
2629 : * ranges without any tuples.
2630 : *
2631 : * Returns the total number of heap tuples scanned.
2632 : */
2633 : static double
2634 8 : _brin_parallel_merge(BrinBuildState *state)
2635 : {
2636 : BrinTuple *btup;
2637 8 : BrinMemTuple *memtuple = NULL;
2638 : Size tuplen;
2639 8 : BlockNumber prevblkno = InvalidBlockNumber;
2640 : MemoryContext rangeCxt,
2641 : oldCxt;
2642 : double reltuples;
2643 :
2644 : /* wait for workers to scan table and produce partial results */
2645 8 : reltuples = _brin_parallel_heapscan(state);
2646 :
2647 : /* do the actual sort in the leader */
2648 8 : tuplesort_performsort(state->bs_sortstate);
2649 :
2650 : /*
2651 : * Initialize BrinMemTuple we'll use to union summaries from workers (in
2652 : * case they happened to produce parts of the same page range).
2653 : */
2654 8 : memtuple = brin_new_memtuple(state->bs_bdesc);
2655 :
2656 : /*
2657 : * Create a memory context we'll reset to combine results for a single
2658 : * page range (received from the workers). We don't expect huge number of
2659 : * overlaps under regular circumstances, because for large tables the
2660 : * chunk size is likely larger than the BRIN page range), but it can
2661 : * happen, and the union functions may do all kinds of stuff. So we better
2662 : * reset the context once in a while.
2663 : */
2664 8 : rangeCxt = AllocSetContextCreate(CurrentMemoryContext,
2665 : "brin union",
2666 : ALLOCSET_DEFAULT_SIZES);
2667 8 : oldCxt = MemoryContextSwitchTo(rangeCxt);
2668 :
2669 : /*
2670 : * Read the BRIN tuples from the shared tuplesort, sorted by block number.
2671 : * That probably gives us an index that is cheaper to scan, thanks to
2672 : * mostly getting data from the same index page as before.
2673 : */
2674 50 : while ((btup = tuplesort_getbrintuple(state->bs_sortstate, &tuplen, true)) != NULL)
2675 : {
2676 : /* Ranges should be multiples of pages_per_range for the index. */
2677 : Assert(btup->bt_blkno % state->bs_leader->brinshared->pagesPerRange == 0);
2678 :
2679 : /*
2680 : * Do we need to union summaries for the same page range?
2681 : *
2682 : * If this is the first brin tuple we read, then just deform it into
2683 : * the memtuple, and continue with the next one from tuplesort. We
2684 : * however may need to insert empty summaries into the index.
2685 : *
2686 : * If it's the same block as the last we saw, we simply union the brin
2687 : * tuple into it, and we're done - we don't even need to insert empty
2688 : * ranges, because that was done earlier when we saw the first brin
2689 : * tuple (for this range).
2690 : *
2691 : * Finally, if it's not the first brin tuple, and it's not the same
2692 : * page range, we need to do the insert and then deform the tuple into
2693 : * the memtuple. Then we'll insert empty ranges before the new brin
2694 : * tuple, if needed.
2695 : */
2696 42 : if (prevblkno == InvalidBlockNumber)
2697 : {
2698 : /* First brin tuples, just deform into memtuple. */
2699 2 : memtuple = brin_deform_tuple(state->bs_bdesc, btup, memtuple);
2700 :
2701 : /* continue to insert empty pages before thisblock */
2702 : }
2703 40 : else if (memtuple->bt_blkno == btup->bt_blkno)
2704 : {
2705 : /*
2706 : * Not the first brin tuple, but same page range as the previous
2707 : * one, so we can merge it into the memtuple.
2708 : */
2709 2 : union_tuples(state->bs_bdesc, memtuple, btup);
2710 2 : continue;
2711 : }
2712 : else
2713 : {
2714 : BrinTuple *tmp;
2715 : Size len;
2716 :
2717 : /*
2718 : * We got brin tuple for a different page range, so form a brin
2719 : * tuple from the memtuple, insert it, and re-init the memtuple
2720 : * from the new brin tuple.
2721 : */
2722 38 : tmp = brin_form_tuple(state->bs_bdesc, memtuple->bt_blkno,
2723 : memtuple, &len);
2724 :
2725 38 : brin_doinsert(state->bs_irel, state->bs_pagesPerRange, state->bs_rmAccess,
2726 : &state->bs_currentInsertBuf, tmp->bt_blkno, tmp, len);
2727 :
2728 : /*
2729 : * Reset the per-output-range context. This frees all the memory
2730 : * possibly allocated by the union functions, and also the BRIN
2731 : * tuple we just formed and inserted.
2732 : */
2733 38 : MemoryContextReset(rangeCxt);
2734 :
2735 38 : memtuple = brin_deform_tuple(state->bs_bdesc, btup, memtuple);
2736 :
2737 : /* continue to insert empty pages before thisblock */
2738 : }
2739 :
2740 : /* Fill empty ranges for all ranges missing in the tuplesort. */
2741 40 : brin_fill_empty_ranges(state, prevblkno, btup->bt_blkno);
2742 :
2743 40 : prevblkno = btup->bt_blkno;
2744 : }
2745 :
2746 8 : tuplesort_end(state->bs_sortstate);
2747 :
2748 : /* Fill the BRIN tuple for the last page range with data. */
2749 8 : if (prevblkno != InvalidBlockNumber)
2750 : {
2751 : BrinTuple *tmp;
2752 : Size len;
2753 :
2754 2 : tmp = brin_form_tuple(state->bs_bdesc, memtuple->bt_blkno,
2755 : memtuple, &len);
2756 :
2757 2 : brin_doinsert(state->bs_irel, state->bs_pagesPerRange, state->bs_rmAccess,
2758 : &state->bs_currentInsertBuf, tmp->bt_blkno, tmp, len);
2759 :
2760 2 : pfree(tmp);
2761 : }
2762 :
2763 : /* Fill empty ranges at the end, for all ranges missing in the tuplesort. */
2764 8 : brin_fill_empty_ranges(state, prevblkno, state->bs_maxRangeStart);
2765 :
2766 : /*
2767 : * Switch back to the original memory context, and destroy the one we
2768 : * created to isolate the union_tuple calls.
2769 : */
2770 8 : MemoryContextSwitchTo(oldCxt);
2771 8 : MemoryContextDelete(rangeCxt);
2772 :
2773 8 : return reltuples;
2774 : }
2775 :
2776 : /*
2777 : * Returns size of shared memory required to store state for a parallel
2778 : * brin index build based on the snapshot its parallel scan will use.
2779 : */
2780 : static Size
2781 10 : _brin_parallel_estimate_shared(Relation heap, Snapshot snapshot)
2782 : {
2783 : /* c.f. shm_toc_allocate as to why BUFFERALIGN is used */
2784 10 : return add_size(BUFFERALIGN(sizeof(BrinShared)),
2785 : table_parallelscan_estimate(heap, snapshot));
2786 : }
2787 :
2788 : /*
2789 : * Within leader, participate as a parallel worker.
2790 : */
2791 : static void
2792 8 : _brin_leader_participate_as_worker(BrinBuildState *buildstate, Relation heap, Relation index)
2793 : {
2794 8 : BrinLeader *brinleader = buildstate->bs_leader;
2795 : int sortmem;
2796 :
2797 : /*
2798 : * Might as well use reliable figure when doling out maintenance_work_mem
2799 : * (when requested number of workers were not launched, this will be
2800 : * somewhat higher than it is for other workers).
2801 : */
2802 8 : sortmem = maintenance_work_mem / brinleader->nparticipanttuplesorts;
2803 :
2804 : /* Perform work common to all participants */
2805 8 : _brin_parallel_scan_and_build(buildstate, brinleader->brinshared,
2806 : brinleader->sharedsort, heap, index, sortmem, true);
2807 8 : }
2808 :
2809 : /*
2810 : * Perform a worker's portion of a parallel sort.
2811 : *
2812 : * This generates a tuplesort for the worker portion of the table.
2813 : *
2814 : * sortmem is the amount of working memory to use within each worker,
2815 : * expressed in KBs.
2816 : *
2817 : * When this returns, workers are done, and need only release resources.
2818 : */
2819 : static void
2820 20 : _brin_parallel_scan_and_build(BrinBuildState *state,
2821 : BrinShared *brinshared, Sharedsort *sharedsort,
2822 : Relation heap, Relation index,
2823 : int sortmem, bool progress)
2824 : {
2825 : SortCoordinate coordinate;
2826 : TableScanDesc scan;
2827 : double reltuples;
2828 : IndexInfo *indexInfo;
2829 :
2830 : /* Initialize local tuplesort coordination state */
2831 20 : coordinate = palloc0(sizeof(SortCoordinateData));
2832 20 : coordinate->isWorker = true;
2833 20 : coordinate->nParticipants = -1;
2834 20 : coordinate->sharedsort = sharedsort;
2835 :
2836 : /* Begin "partial" tuplesort */
2837 20 : state->bs_sortstate = tuplesort_begin_index_brin(sortmem, coordinate,
2838 : TUPLESORT_NONE);
2839 :
2840 : /* Join parallel scan */
2841 20 : indexInfo = BuildIndexInfo(index);
2842 20 : indexInfo->ii_Concurrent = brinshared->isconcurrent;
2843 :
2844 20 : scan = table_beginscan_parallel(heap,
2845 : ParallelTableScanFromBrinShared(brinshared));
2846 :
2847 20 : reltuples = table_index_build_scan(heap, index, indexInfo, true, true,
2848 : brinbuildCallbackParallel, state, scan);
2849 :
2850 : /* insert the last item */
2851 20 : form_and_spill_tuple(state);
2852 :
2853 : /* sort the BRIN ranges built by this worker */
2854 20 : tuplesort_performsort(state->bs_sortstate);
2855 :
2856 20 : state->bs_reltuples += reltuples;
2857 :
2858 : /*
2859 : * Done. Record ambuild statistics.
2860 : */
2861 20 : SpinLockAcquire(&brinshared->mutex);
2862 20 : brinshared->nparticipantsdone++;
2863 20 : brinshared->reltuples += state->bs_reltuples;
2864 20 : brinshared->indtuples += state->bs_numtuples;
2865 20 : SpinLockRelease(&brinshared->mutex);
2866 :
2867 : /* Notify leader */
2868 20 : ConditionVariableSignal(&brinshared->workersdonecv);
2869 :
2870 20 : tuplesort_end(state->bs_sortstate);
2871 20 : }
2872 :
2873 : /*
2874 : * Perform work within a launched parallel process.
2875 : */
2876 : void
2877 12 : _brin_parallel_build_main(dsm_segment *seg, shm_toc *toc)
2878 : {
2879 : char *sharedquery;
2880 : BrinShared *brinshared;
2881 : Sharedsort *sharedsort;
2882 : BrinBuildState *buildstate;
2883 : Relation heapRel;
2884 : Relation indexRel;
2885 : LOCKMODE heapLockmode;
2886 : LOCKMODE indexLockmode;
2887 : WalUsage *walusage;
2888 : BufferUsage *bufferusage;
2889 : int sortmem;
2890 :
2891 : /*
2892 : * The only possible status flag that can be set to the parallel worker is
2893 : * PROC_IN_SAFE_IC.
2894 : */
2895 : Assert((MyProc->statusFlags == 0) ||
2896 : (MyProc->statusFlags == PROC_IN_SAFE_IC));
2897 :
2898 : /* Set debug_query_string for individual workers first */
2899 12 : sharedquery = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, true);
2900 12 : debug_query_string = sharedquery;
2901 :
2902 : /* Report the query string from leader */
2903 12 : pgstat_report_activity(STATE_RUNNING, debug_query_string);
2904 :
2905 : /* Look up brin shared state */
2906 12 : brinshared = shm_toc_lookup(toc, PARALLEL_KEY_BRIN_SHARED, false);
2907 :
2908 : /* Open relations using lock modes known to be obtained by index.c */
2909 12 : if (!brinshared->isconcurrent)
2910 : {
2911 12 : heapLockmode = ShareLock;
2912 12 : indexLockmode = AccessExclusiveLock;
2913 : }
2914 : else
2915 : {
2916 0 : heapLockmode = ShareUpdateExclusiveLock;
2917 0 : indexLockmode = RowExclusiveLock;
2918 : }
2919 :
2920 : /* Track query ID */
2921 12 : pgstat_report_query_id(brinshared->queryid, false);
2922 :
2923 : /* Open relations within worker */
2924 12 : heapRel = table_open(brinshared->heaprelid, heapLockmode);
2925 12 : indexRel = index_open(brinshared->indexrelid, indexLockmode);
2926 :
2927 12 : buildstate = initialize_brin_buildstate(indexRel, NULL,
2928 : brinshared->pagesPerRange,
2929 : InvalidBlockNumber);
2930 :
2931 : /* Look up shared state private to tuplesort.c */
2932 12 : sharedsort = shm_toc_lookup(toc, PARALLEL_KEY_TUPLESORT, false);
2933 12 : tuplesort_attach_shared(sharedsort, seg);
2934 :
2935 : /* Prepare to track buffer usage during parallel execution */
2936 12 : InstrStartParallelQuery();
2937 :
2938 : /*
2939 : * Might as well use reliable figure when doling out maintenance_work_mem
2940 : * (when requested number of workers were not launched, this will be
2941 : * somewhat higher than it is for other workers).
2942 : */
2943 12 : sortmem = maintenance_work_mem / brinshared->scantuplesortstates;
2944 :
2945 12 : _brin_parallel_scan_and_build(buildstate, brinshared, sharedsort,
2946 : heapRel, indexRel, sortmem, false);
2947 :
2948 : /* Report WAL/buffer usage during parallel execution */
2949 12 : bufferusage = shm_toc_lookup(toc, PARALLEL_KEY_BUFFER_USAGE, false);
2950 12 : walusage = shm_toc_lookup(toc, PARALLEL_KEY_WAL_USAGE, false);
2951 12 : InstrEndParallelQuery(&bufferusage[ParallelWorkerNumber],
2952 12 : &walusage[ParallelWorkerNumber]);
2953 :
2954 12 : index_close(indexRel, indexLockmode);
2955 12 : table_close(heapRel, heapLockmode);
2956 12 : }
2957 :
2958 : /*
2959 : * brin_build_empty_tuple
2960 : * Maybe initialize a BRIN tuple representing empty range.
2961 : *
2962 : * Returns a BRIN tuple representing an empty page range starting at the
2963 : * specified block number. The empty tuple is initialized only once, when it's
2964 : * needed for the first time, stored in the memory context bs_context to ensure
2965 : * proper life span, and reused on following calls. All empty tuples are
2966 : * exactly the same except for the bt_blkno field, which is set to the value
2967 : * in blkno parameter.
2968 : */
2969 : static void
2970 20 : brin_build_empty_tuple(BrinBuildState *state, BlockNumber blkno)
2971 : {
2972 : /* First time an empty tuple is requested? If yes, initialize it. */
2973 20 : if (state->bs_emptyTuple == NULL)
2974 : {
2975 : MemoryContext oldcxt;
2976 10 : BrinMemTuple *dtuple = brin_new_memtuple(state->bs_bdesc);
2977 :
2978 : /* Allocate the tuple in context for the whole index build. */
2979 10 : oldcxt = MemoryContextSwitchTo(state->bs_context);
2980 :
2981 10 : state->bs_emptyTuple = brin_form_tuple(state->bs_bdesc, blkno, dtuple,
2982 : &state->bs_emptyTupleLen);
2983 :
2984 10 : MemoryContextSwitchTo(oldcxt);
2985 : }
2986 : else
2987 : {
2988 : /* If we already have an empty tuple, just update the block. */
2989 10 : state->bs_emptyTuple->bt_blkno = blkno;
2990 : }
2991 20 : }
2992 :
2993 : /*
2994 : * brin_fill_empty_ranges
2995 : * Add BRIN index tuples representing empty page ranges.
2996 : *
2997 : * prevRange/nextRange determine for which page ranges to add empty summaries.
2998 : * Both boundaries are exclusive, i.e. only ranges starting at blkno for which
2999 : * (prevRange < blkno < nextRange) will be added to the index.
3000 : *
3001 : * If prevRange is InvalidBlockNumber, this means there was no previous page
3002 : * range (i.e. the first empty range to add is for blkno=0).
3003 : *
3004 : * The empty tuple is built only once, and then reused for all future calls.
3005 : */
3006 : static void
3007 408 : brin_fill_empty_ranges(BrinBuildState *state,
3008 : BlockNumber prevRange, BlockNumber nextRange)
3009 : {
3010 : BlockNumber blkno;
3011 :
3012 : /*
3013 : * If we already summarized some ranges, we need to start with the next
3014 : * one. Otherwise start from the first range of the table.
3015 : */
3016 408 : blkno = (prevRange == InvalidBlockNumber) ? 0 : (prevRange + state->bs_pagesPerRange);
3017 :
3018 : /* Generate empty ranges until we hit the next non-empty range. */
3019 428 : while (blkno < nextRange)
3020 : {
3021 : /* Did we already build the empty tuple? If not, do it now. */
3022 20 : brin_build_empty_tuple(state, blkno);
3023 :
3024 20 : brin_doinsert(state->bs_irel, state->bs_pagesPerRange, state->bs_rmAccess,
3025 : &state->bs_currentInsertBuf,
3026 20 : blkno, state->bs_emptyTuple, state->bs_emptyTupleLen);
3027 :
3028 : /* try next page range */
3029 20 : blkno += state->bs_pagesPerRange;
3030 : }
3031 408 : }
|
