Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * nbtsort.c
4 : * Build a btree from sorted input by loading leaf pages sequentially.
5 : *
6 : * NOTES
7 : *
8 : * We use tuplesort.c to sort the given index tuples into order.
9 : * Then we scan the index tuples in order and build the btree pages
10 : * for each level. We load source tuples into leaf-level pages.
11 : * Whenever we fill a page at one level, we add a link to it to its
12 : * parent level (starting a new parent level if necessary). When
13 : * done, we write out each final page on each level, adding it to
14 : * its parent level. When we have only one page on a level, it must be
15 : * the root -- it can be attached to the btree metapage and we are done.
16 : *
17 : * It is not wise to pack the pages entirely full, since then *any*
18 : * insertion would cause a split (and not only of the leaf page; the need
19 : * for a split would cascade right up the tree). The steady-state load
20 : * factor for btrees is usually estimated at 70%. We choose to pack leaf
21 : * pages to the user-controllable fill factor (default 90%) while upper pages
22 : * are always packed to 70%. This gives us reasonable density (there aren't
23 : * many upper pages if the keys are reasonable-size) without risking a lot of
24 : * cascading splits during early insertions.
25 : *
26 : * We use the bulk smgr loading facility to bypass the buffer cache and
27 : * WAL-log the pages efficiently.
28 : *
29 : * This code isn't concerned about the FSM at all. The caller is responsible
30 : * for initializing that.
31 : *
32 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
33 : * Portions Copyright (c) 1994, Regents of the University of California
34 : *
35 : * IDENTIFICATION
36 : * src/backend/access/nbtree/nbtsort.c
37 : *
38 : *-------------------------------------------------------------------------
39 : */
40 :
41 : #include "postgres.h"
42 :
43 : #include "access/nbtree.h"
44 : #include "access/parallel.h"
45 : #include "access/relscan.h"
46 : #include "access/table.h"
47 : #include "access/tableam.h"
48 : #include "access/xact.h"
49 : #include "catalog/index.h"
50 : #include "commands/progress.h"
51 : #include "executor/instrument.h"
52 : #include "miscadmin.h"
53 : #include "pgstat.h"
54 : #include "storage/bulk_write.h"
55 : #include "tcop/tcopprot.h"
56 : #include "utils/rel.h"
57 : #include "utils/sortsupport.h"
58 : #include "utils/tuplesort.h"
59 :
60 :
61 : /* Magic numbers for parallel state sharing */
62 : #define PARALLEL_KEY_BTREE_SHARED UINT64CONST(0xA000000000000001)
63 : #define PARALLEL_KEY_TUPLESORT UINT64CONST(0xA000000000000002)
64 : #define PARALLEL_KEY_TUPLESORT_SPOOL2 UINT64CONST(0xA000000000000003)
65 : #define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xA000000000000004)
66 : #define PARALLEL_KEY_WAL_USAGE UINT64CONST(0xA000000000000005)
67 : #define PARALLEL_KEY_BUFFER_USAGE UINT64CONST(0xA000000000000006)
68 :
69 : /*
70 : * DISABLE_LEADER_PARTICIPATION disables the leader's participation in
71 : * parallel index builds. This may be useful as a debugging aid.
72 : #undef DISABLE_LEADER_PARTICIPATION
73 : */
74 :
75 : /*
76 : * Status record for spooling/sorting phase. (Note we may have two of
77 : * these due to the special requirements for uniqueness-checking with
78 : * dead tuples.)
79 : */
80 : typedef struct BTSpool
81 : {
82 : Tuplesortstate *sortstate; /* state data for tuplesort.c */
83 : Relation heap;
84 : Relation index;
85 : bool isunique;
86 : bool nulls_not_distinct;
87 : } BTSpool;
88 :
89 : /*
90 : * Status for index builds performed in parallel. This is allocated in a
91 : * dynamic shared memory segment. Note that there is a separate tuplesort TOC
92 : * entry, private to tuplesort.c but allocated by this module on its behalf.
93 : */
94 : typedef struct BTShared
95 : {
96 : /*
97 : * These fields are not modified during the sort. They primarily exist
98 : * for the benefit of worker processes that need to create BTSpool state
99 : * corresponding to that used by the leader.
100 : */
101 : Oid heaprelid;
102 : Oid indexrelid;
103 : bool isunique;
104 : bool nulls_not_distinct;
105 : bool isconcurrent;
106 : int scantuplesortstates;
107 :
108 : /* Query ID, for report in worker processes */
109 : int64 queryid;
110 :
111 : /*
112 : * workersdonecv is used to monitor the progress of workers. All parallel
113 : * participants must indicate that they are done before leader can use
114 : * mutable state that workers maintain during scan (and before leader can
115 : * proceed to tuplesort_performsort()).
116 : */
117 : ConditionVariable workersdonecv;
118 :
119 : /*
120 : * mutex protects all fields before heapdesc.
121 : *
122 : * These fields contain status information of interest to B-Tree index
123 : * builds that must work just the same when an index is built in parallel.
124 : */
125 : slock_t mutex;
126 :
127 : /*
128 : * Mutable state that is maintained by workers, and reported back to
129 : * leader at end of parallel scan.
130 : *
131 : * nparticipantsdone is number of worker processes finished.
132 : *
133 : * reltuples is the total number of input heap tuples.
134 : *
135 : * havedead indicates if RECENTLY_DEAD tuples were encountered during
136 : * build.
137 : *
138 : * indtuples is the total number of tuples that made it into the index.
139 : *
140 : * brokenhotchain indicates if any worker detected a broken HOT chain
141 : * during build.
142 : */
143 : int nparticipantsdone;
144 : double reltuples;
145 : bool havedead;
146 : double indtuples;
147 : bool brokenhotchain;
148 :
149 : /*
150 : * ParallelTableScanDescData data follows. Can't directly embed here, as
151 : * implementations of the parallel table scan desc interface might need
152 : * stronger alignment.
153 : */
154 : } BTShared;
155 :
156 : /*
157 : * Return pointer to a BTShared's parallel table scan.
158 : *
159 : * c.f. shm_toc_allocate as to why BUFFERALIGN is used, rather than just
160 : * MAXALIGN.
161 : */
162 : #define ParallelTableScanFromBTShared(shared) \
163 : (ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(BTShared)))
164 :
165 : /*
166 : * Status for leader in parallel index build.
167 : */
168 : typedef struct BTLeader
169 : {
170 : /* parallel context itself */
171 : ParallelContext *pcxt;
172 :
173 : /*
174 : * nparticipanttuplesorts is the exact number of worker processes
175 : * successfully launched, plus one leader process if it participates as a
176 : * worker (only DISABLE_LEADER_PARTICIPATION builds avoid leader
177 : * participating as a worker).
178 : */
179 : int nparticipanttuplesorts;
180 :
181 : /*
182 : * Leader process convenience pointers to shared state (leader avoids TOC
183 : * lookups).
184 : *
185 : * btshared is the shared state for entire build. sharedsort is the
186 : * shared, tuplesort-managed state passed to each process tuplesort.
187 : * sharedsort2 is the corresponding btspool2 shared state, used only when
188 : * building unique indexes. snapshot is the snapshot used by the scan iff
189 : * an MVCC snapshot is required.
190 : */
191 : BTShared *btshared;
192 : Sharedsort *sharedsort;
193 : Sharedsort *sharedsort2;
194 : Snapshot snapshot;
195 : WalUsage *walusage;
196 : BufferUsage *bufferusage;
197 : } BTLeader;
198 :
199 : /*
200 : * Working state for btbuild and its callback.
201 : *
202 : * When parallel CREATE INDEX is used, there is a BTBuildState for each
203 : * participant.
204 : */
205 : typedef struct BTBuildState
206 : {
207 : bool isunique;
208 : bool nulls_not_distinct;
209 : bool havedead;
210 : Relation heap;
211 : BTSpool *spool;
212 :
213 : /*
214 : * spool2 is needed only when the index is a unique index. Dead tuples are
215 : * put into spool2 instead of spool in order to avoid uniqueness check.
216 : */
217 : BTSpool *spool2;
218 : double indtuples;
219 :
220 : /*
221 : * btleader is only present when a parallel index build is performed, and
222 : * only in the leader process. (Actually, only the leader has a
223 : * BTBuildState. Workers have their own spool and spool2, though.)
224 : */
225 : BTLeader *btleader;
226 : } BTBuildState;
227 :
228 : /*
229 : * Status record for a btree page being built. We have one of these
230 : * for each active tree level.
231 : */
232 : typedef struct BTPageState
233 : {
234 : BulkWriteBuffer btps_buf; /* workspace for page building */
235 : BlockNumber btps_blkno; /* block # to write this page at */
236 : IndexTuple btps_lowkey; /* page's strict lower bound pivot tuple */
237 : OffsetNumber btps_lastoff; /* last item offset loaded */
238 : Size btps_lastextra; /* last item's extra posting list space */
239 : uint32 btps_level; /* tree level (0 = leaf) */
240 : Size btps_full; /* "full" if less than this much free space */
241 : struct BTPageState *btps_next; /* link to parent level, if any */
242 : } BTPageState;
243 :
244 : /*
245 : * Overall status record for index writing phase.
246 : */
247 : typedef struct BTWriteState
248 : {
249 : Relation heap;
250 : Relation index;
251 : BulkWriteState *bulkstate;
252 : BTScanInsert inskey; /* generic insertion scankey */
253 : BlockNumber btws_pages_alloced; /* # pages allocated */
254 : } BTWriteState;
255 :
256 :
257 : static double _bt_spools_heapscan(Relation heap, Relation index,
258 : BTBuildState *buildstate, IndexInfo *indexInfo);
259 : static void _bt_spooldestroy(BTSpool *btspool);
260 : static void _bt_spool(BTSpool *btspool, ItemPointer self,
261 : Datum *values, bool *isnull);
262 : static void _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2);
263 : static void _bt_build_callback(Relation index, ItemPointer tid, Datum *values,
264 : bool *isnull, bool tupleIsAlive, void *state);
265 : static BulkWriteBuffer _bt_blnewpage(BTWriteState *wstate, uint32 level);
266 : static BTPageState *_bt_pagestate(BTWriteState *wstate, uint32 level);
267 : static void _bt_slideleft(Page rightmostpage);
268 : static void _bt_sortaddtup(Page page, Size itemsize,
269 : IndexTuple itup, OffsetNumber itup_off,
270 : bool newfirstdataitem);
271 : static void _bt_buildadd(BTWriteState *wstate, BTPageState *state,
272 : IndexTuple itup, Size truncextra);
273 : static void _bt_sort_dedup_finish_pending(BTWriteState *wstate,
274 : BTPageState *state,
275 : BTDedupState dstate);
276 : static void _bt_uppershutdown(BTWriteState *wstate, BTPageState *state);
277 : static void _bt_load(BTWriteState *wstate,
278 : BTSpool *btspool, BTSpool *btspool2);
279 : static void _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent,
280 : int request);
281 : static void _bt_end_parallel(BTLeader *btleader);
282 : static Size _bt_parallel_estimate_shared(Relation heap, Snapshot snapshot);
283 : static double _bt_parallel_heapscan(BTBuildState *buildstate,
284 : bool *brokenhotchain);
285 : static void _bt_leader_participate_as_worker(BTBuildState *buildstate);
286 : static void _bt_parallel_scan_and_sort(BTSpool *btspool, BTSpool *btspool2,
287 : BTShared *btshared, Sharedsort *sharedsort,
288 : Sharedsort *sharedsort2, int sortmem,
289 : bool progress);
290 :
291 :
292 : /*
293 : * btbuild() -- build a new btree index.
294 : */
295 : IndexBuildResult *
296 50298 : btbuild(Relation heap, Relation index, IndexInfo *indexInfo)
297 : {
298 : IndexBuildResult *result;
299 : BTBuildState buildstate;
300 : double reltuples;
301 :
302 : #ifdef BTREE_BUILD_STATS
303 : if (log_btree_build_stats)
304 : ResetUsage();
305 : #endif /* BTREE_BUILD_STATS */
306 :
307 50298 : buildstate.isunique = indexInfo->ii_Unique;
308 50298 : buildstate.nulls_not_distinct = indexInfo->ii_NullsNotDistinct;
309 50298 : buildstate.havedead = false;
310 50298 : buildstate.heap = heap;
311 50298 : buildstate.spool = NULL;
312 50298 : buildstate.spool2 = NULL;
313 50298 : buildstate.indtuples = 0;
314 50298 : buildstate.btleader = NULL;
315 :
316 : /*
317 : * We expect to be called exactly once for any index relation. If that's
318 : * not the case, big trouble's what we have.
319 : */
320 50298 : if (RelationGetNumberOfBlocks(index) != 0)
321 0 : elog(ERROR, "index \"%s\" already contains data",
322 : RelationGetRelationName(index));
323 :
324 50298 : reltuples = _bt_spools_heapscan(heap, index, &buildstate, indexInfo);
325 :
326 : /*
327 : * Finish the build by (1) completing the sort of the spool file, (2)
328 : * inserting the sorted tuples into btree pages and (3) building the upper
329 : * levels. Finally, it may also be necessary to end use of parallelism.
330 : */
331 50286 : _bt_leafbuild(buildstate.spool, buildstate.spool2);
332 50202 : _bt_spooldestroy(buildstate.spool);
333 50202 : if (buildstate.spool2)
334 20 : _bt_spooldestroy(buildstate.spool2);
335 50202 : if (buildstate.btleader)
336 148 : _bt_end_parallel(buildstate.btleader);
337 :
338 50202 : result = (IndexBuildResult *) palloc(sizeof(IndexBuildResult));
339 :
340 50202 : result->heap_tuples = reltuples;
341 50202 : result->index_tuples = buildstate.indtuples;
342 :
343 : #ifdef BTREE_BUILD_STATS
344 : if (log_btree_build_stats)
345 : {
346 : ShowUsage("BTREE BUILD STATS");
347 : ResetUsage();
348 : }
349 : #endif /* BTREE_BUILD_STATS */
350 :
351 50202 : return result;
352 : }
353 :
354 : /*
355 : * Create and initialize one or two spool structures, and save them in caller's
356 : * buildstate argument. May also fill-in fields within indexInfo used by index
357 : * builds.
358 : *
359 : * Scans the heap, possibly in parallel, filling spools with IndexTuples. This
360 : * routine encapsulates all aspects of managing parallelism. Caller need only
361 : * call _bt_end_parallel() in parallel case after it is done with spool/spool2.
362 : *
363 : * Returns the total number of heap tuples scanned.
364 : */
365 : static double
366 50298 : _bt_spools_heapscan(Relation heap, Relation index, BTBuildState *buildstate,
367 : IndexInfo *indexInfo)
368 : {
369 50298 : BTSpool *btspool = (BTSpool *) palloc0(sizeof(BTSpool));
370 50298 : SortCoordinate coordinate = NULL;
371 50298 : double reltuples = 0;
372 :
373 : /*
374 : * We size the sort area as maintenance_work_mem rather than work_mem to
375 : * speed index creation. This should be OK since a single backend can't
376 : * run multiple index creations in parallel (see also: notes on
377 : * parallelism and maintenance_work_mem below).
378 : */
379 50298 : btspool->heap = heap;
380 50298 : btspool->index = index;
381 50298 : btspool->isunique = indexInfo->ii_Unique;
382 50298 : btspool->nulls_not_distinct = indexInfo->ii_NullsNotDistinct;
383 :
384 : /* Save as primary spool */
385 50298 : buildstate->spool = btspool;
386 :
387 : /* Report table scan phase started */
388 50298 : pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE,
389 : PROGRESS_BTREE_PHASE_INDEXBUILD_TABLESCAN);
390 :
391 : /* Attempt to launch parallel worker scan when required */
392 50298 : if (indexInfo->ii_ParallelWorkers > 0)
393 148 : _bt_begin_parallel(buildstate, indexInfo->ii_Concurrent,
394 : indexInfo->ii_ParallelWorkers);
395 :
396 : /*
397 : * If parallel build requested and at least one worker process was
398 : * successfully launched, set up coordination state
399 : */
400 50298 : if (buildstate->btleader)
401 : {
402 148 : coordinate = (SortCoordinate) palloc0(sizeof(SortCoordinateData));
403 148 : coordinate->isWorker = false;
404 148 : coordinate->nParticipants =
405 148 : buildstate->btleader->nparticipanttuplesorts;
406 148 : coordinate->sharedsort = buildstate->btleader->sharedsort;
407 : }
408 :
409 : /*
410 : * Begin serial/leader tuplesort.
411 : *
412 : * In cases where parallelism is involved, the leader receives the same
413 : * share of maintenance_work_mem as a serial sort (it is generally treated
414 : * in the same way as a serial sort once we return). Parallel worker
415 : * Tuplesortstates will have received only a fraction of
416 : * maintenance_work_mem, though.
417 : *
418 : * We rely on the lifetime of the Leader Tuplesortstate almost not
419 : * overlapping with any worker Tuplesortstate's lifetime. There may be
420 : * some small overlap, but that's okay because we rely on leader
421 : * Tuplesortstate only allocating a small, fixed amount of memory here.
422 : * When its tuplesort_performsort() is called (by our caller), and
423 : * significant amounts of memory are likely to be used, all workers must
424 : * have already freed almost all memory held by their Tuplesortstates
425 : * (they are about to go away completely, too). The overall effect is
426 : * that maintenance_work_mem always represents an absolute high watermark
427 : * on the amount of memory used by a CREATE INDEX operation, regardless of
428 : * the use of parallelism or any other factor.
429 : */
430 100596 : buildstate->spool->sortstate =
431 50298 : tuplesort_begin_index_btree(heap, index, buildstate->isunique,
432 50298 : buildstate->nulls_not_distinct,
433 : maintenance_work_mem, coordinate,
434 : TUPLESORT_NONE);
435 :
436 : /*
437 : * If building a unique index, put dead tuples in a second spool to keep
438 : * them out of the uniqueness check. We expect that the second spool (for
439 : * dead tuples) won't get very full, so we give it only work_mem.
440 : */
441 50298 : if (indexInfo->ii_Unique)
442 : {
443 41112 : BTSpool *btspool2 = (BTSpool *) palloc0(sizeof(BTSpool));
444 41112 : SortCoordinate coordinate2 = NULL;
445 :
446 : /* Initialize secondary spool */
447 41112 : btspool2->heap = heap;
448 41112 : btspool2->index = index;
449 41112 : btspool2->isunique = false;
450 : /* Save as secondary spool */
451 41112 : buildstate->spool2 = btspool2;
452 :
453 41112 : if (buildstate->btleader)
454 : {
455 : /*
456 : * Set up non-private state that is passed to
457 : * tuplesort_begin_index_btree() about the basic high level
458 : * coordination of a parallel sort.
459 : */
460 64 : coordinate2 = (SortCoordinate) palloc0(sizeof(SortCoordinateData));
461 64 : coordinate2->isWorker = false;
462 64 : coordinate2->nParticipants =
463 64 : buildstate->btleader->nparticipanttuplesorts;
464 64 : coordinate2->sharedsort = buildstate->btleader->sharedsort2;
465 : }
466 :
467 : /*
468 : * We expect that the second one (for dead tuples) won't get very
469 : * full, so we give it only work_mem
470 : */
471 41112 : buildstate->spool2->sortstate =
472 41112 : tuplesort_begin_index_btree(heap, index, false, false, work_mem,
473 : coordinate2, TUPLESORT_NONE);
474 : }
475 :
476 : /* Fill spool using either serial or parallel heap scan */
477 50298 : if (!buildstate->btleader)
478 50150 : reltuples = table_index_build_scan(heap, index, indexInfo, true, true,
479 : _bt_build_callback, buildstate,
480 : NULL);
481 : else
482 148 : reltuples = _bt_parallel_heapscan(buildstate,
483 : &indexInfo->ii_BrokenHotChain);
484 :
485 : /*
486 : * Set the progress target for the next phase. Reset the block number
487 : * values set by table_index_build_scan
488 : */
489 : {
490 50286 : const int progress_index[] = {
491 : PROGRESS_CREATEIDX_TUPLES_TOTAL,
492 : PROGRESS_SCAN_BLOCKS_TOTAL,
493 : PROGRESS_SCAN_BLOCKS_DONE
494 : };
495 50286 : const int64 progress_vals[] = {
496 50286 : buildstate->indtuples,
497 : 0, 0
498 : };
499 :
500 50286 : pgstat_progress_update_multi_param(3, progress_index, progress_vals);
501 : }
502 :
503 : /* okay, all heap tuples are spooled */
504 50286 : if (buildstate->spool2 && !buildstate->havedead)
505 : {
506 : /* spool2 turns out to be unnecessary */
507 41092 : _bt_spooldestroy(buildstate->spool2);
508 41092 : buildstate->spool2 = NULL;
509 : }
510 :
511 50286 : return reltuples;
512 : }
513 :
514 : /*
515 : * clean up a spool structure and its substructures.
516 : */
517 : static void
518 91314 : _bt_spooldestroy(BTSpool *btspool)
519 : {
520 91314 : tuplesort_end(btspool->sortstate);
521 91314 : pfree(btspool);
522 91314 : }
523 :
524 : /*
525 : * spool an index entry into the sort file.
526 : */
527 : static void
528 12776742 : _bt_spool(BTSpool *btspool, ItemPointer self, Datum *values, bool *isnull)
529 : {
530 12776742 : tuplesort_putindextuplevalues(btspool->sortstate, btspool->index,
531 : self, values, isnull);
532 12776742 : }
533 :
534 : /*
535 : * given a spool loaded by successive calls to _bt_spool,
536 : * create an entire btree.
537 : */
538 : static void
539 50286 : _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2)
540 : {
541 : BTWriteState wstate;
542 :
543 : #ifdef BTREE_BUILD_STATS
544 : if (log_btree_build_stats)
545 : {
546 : ShowUsage("BTREE BUILD (Spool) STATISTICS");
547 : ResetUsage();
548 : }
549 : #endif /* BTREE_BUILD_STATS */
550 :
551 : /* Execute the sort */
552 50286 : pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE,
553 : PROGRESS_BTREE_PHASE_PERFORMSORT_1);
554 50286 : tuplesort_performsort(btspool->sortstate);
555 50202 : if (btspool2)
556 : {
557 20 : pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE,
558 : PROGRESS_BTREE_PHASE_PERFORMSORT_2);
559 20 : tuplesort_performsort(btspool2->sortstate);
560 : }
561 :
562 50202 : wstate.heap = btspool->heap;
563 50202 : wstate.index = btspool->index;
564 50202 : wstate.inskey = _bt_mkscankey(wstate.index, NULL);
565 : /* _bt_mkscankey() won't set allequalimage without metapage */
566 50202 : wstate.inskey->allequalimage = _bt_allequalimage(wstate.index, true);
567 :
568 : /* reserve the metapage */
569 50202 : wstate.btws_pages_alloced = BTREE_METAPAGE + 1;
570 :
571 50202 : pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE,
572 : PROGRESS_BTREE_PHASE_LEAF_LOAD);
573 50202 : _bt_load(&wstate, btspool, btspool2);
574 50202 : }
575 :
576 : /*
577 : * Per-tuple callback for table_index_build_scan
578 : */
579 : static void
580 12776742 : _bt_build_callback(Relation index,
581 : ItemPointer tid,
582 : Datum *values,
583 : bool *isnull,
584 : bool tupleIsAlive,
585 : void *state)
586 : {
587 12776742 : BTBuildState *buildstate = (BTBuildState *) state;
588 :
589 : /*
590 : * insert the index tuple into the appropriate spool file for subsequent
591 : * processing
592 : */
593 12776742 : if (tupleIsAlive || buildstate->spool2 == NULL)
594 12776252 : _bt_spool(buildstate->spool, tid, values, isnull);
595 : else
596 : {
597 : /* dead tuples are put into spool2 */
598 490 : buildstate->havedead = true;
599 490 : _bt_spool(buildstate->spool2, tid, values, isnull);
600 : }
601 :
602 12776742 : buildstate->indtuples += 1;
603 12776742 : }
604 :
605 : /*
606 : * allocate workspace for a new, clean btree page, not linked to any siblings.
607 : */
608 : static BulkWriteBuffer
609 52962 : _bt_blnewpage(BTWriteState *wstate, uint32 level)
610 : {
611 : BulkWriteBuffer buf;
612 : Page page;
613 : BTPageOpaque opaque;
614 :
615 52962 : buf = smgr_bulk_get_buf(wstate->bulkstate);
616 52962 : page = (Page) buf;
617 :
618 : /* Zero the page and set up standard page header info */
619 52962 : _bt_pageinit(page, BLCKSZ);
620 :
621 : /* Initialize BT opaque state */
622 52962 : opaque = BTPageGetOpaque(page);
623 52962 : opaque->btpo_prev = opaque->btpo_next = P_NONE;
624 52962 : opaque->btpo_level = level;
625 52962 : opaque->btpo_flags = (level > 0) ? 0 : BTP_LEAF;
626 52962 : opaque->btpo_cycleid = 0;
627 :
628 : /* Make the P_HIKEY line pointer appear allocated */
629 52962 : ((PageHeader) page)->pd_lower += sizeof(ItemIdData);
630 :
631 52962 : return buf;
632 : }
633 :
634 : /*
635 : * emit a completed btree page, and release the working storage.
636 : */
637 : static void
638 103164 : _bt_blwritepage(BTWriteState *wstate, BulkWriteBuffer buf, BlockNumber blkno)
639 : {
640 103164 : smgr_bulk_write(wstate->bulkstate, blkno, buf, true);
641 : /* smgr_bulk_write took ownership of 'buf' */
642 103164 : }
643 :
644 : /*
645 : * allocate and initialize a new BTPageState. the returned structure
646 : * is suitable for immediate use by _bt_buildadd.
647 : */
648 : static BTPageState *
649 12862 : _bt_pagestate(BTWriteState *wstate, uint32 level)
650 : {
651 12862 : BTPageState *state = (BTPageState *) palloc0(sizeof(BTPageState));
652 :
653 : /* create initial page for level */
654 12862 : state->btps_buf = _bt_blnewpage(wstate, level);
655 :
656 : /* and assign it a page position */
657 12862 : state->btps_blkno = wstate->btws_pages_alloced++;
658 :
659 12862 : state->btps_lowkey = NULL;
660 : /* initialize lastoff so first item goes into P_FIRSTKEY */
661 12862 : state->btps_lastoff = P_HIKEY;
662 12862 : state->btps_lastextra = 0;
663 12862 : state->btps_level = level;
664 : /* set "full" threshold based on level. See notes at head of file. */
665 12862 : if (level > 0)
666 2692 : state->btps_full = (BLCKSZ * (100 - BTREE_NONLEAF_FILLFACTOR) / 100);
667 : else
668 10170 : state->btps_full = BTGetTargetPageFreeSpace(wstate->index);
669 :
670 : /* no parent level, yet */
671 12862 : state->btps_next = NULL;
672 :
673 12862 : return state;
674 : }
675 :
676 : /*
677 : * Slide the array of ItemIds from the page back one slot (from P_FIRSTKEY to
678 : * P_HIKEY, overwriting P_HIKEY).
679 : *
680 : * _bt_blnewpage() makes the P_HIKEY line pointer appear allocated, but the
681 : * rightmost page on its level is not supposed to get a high key. Now that
682 : * it's clear that this page is a rightmost page, remove the unneeded empty
683 : * P_HIKEY line pointer space.
684 : */
685 : static void
686 12862 : _bt_slideleft(Page rightmostpage)
687 : {
688 : OffsetNumber off;
689 : OffsetNumber maxoff;
690 : ItemId previi;
691 :
692 12862 : maxoff = PageGetMaxOffsetNumber(rightmostpage);
693 : Assert(maxoff >= P_FIRSTKEY);
694 12862 : previi = PageGetItemId(rightmostpage, P_HIKEY);
695 758084 : for (off = P_FIRSTKEY; off <= maxoff; off = OffsetNumberNext(off))
696 : {
697 745222 : ItemId thisii = PageGetItemId(rightmostpage, off);
698 :
699 745222 : *previi = *thisii;
700 745222 : previi = thisii;
701 : }
702 12862 : ((PageHeader) rightmostpage)->pd_lower -= sizeof(ItemIdData);
703 12862 : }
704 :
705 : /*
706 : * Add an item to a page being built.
707 : *
708 : * This is very similar to nbtinsert.c's _bt_pgaddtup(), but this variant
709 : * raises an error directly.
710 : *
711 : * Note that our nbtsort.c caller does not know yet if the page will be
712 : * rightmost. Offset P_FIRSTKEY is always assumed to be the first data key by
713 : * caller. Page that turns out to be the rightmost on its level is fixed by
714 : * calling _bt_slideleft().
715 : */
716 : static void
717 11493462 : _bt_sortaddtup(Page page,
718 : Size itemsize,
719 : IndexTuple itup,
720 : OffsetNumber itup_off,
721 : bool newfirstdataitem)
722 : {
723 : IndexTupleData trunctuple;
724 :
725 11493462 : if (newfirstdataitem)
726 : {
727 2816 : trunctuple = *itup;
728 2816 : trunctuple.t_info = sizeof(IndexTupleData);
729 2816 : BTreeTupleSetNAtts(&trunctuple, 0, false);
730 2816 : itup = &trunctuple;
731 2816 : itemsize = sizeof(IndexTupleData);
732 : }
733 :
734 11493462 : if (PageAddItem(page, (Item) itup, itemsize, itup_off,
735 : false, false) == InvalidOffsetNumber)
736 0 : elog(ERROR, "failed to add item to the index page");
737 11493462 : }
738 :
739 : /*----------
740 : * Add an item to a disk page from the sort output (or add a posting list
741 : * item formed from the sort output).
742 : *
743 : * We must be careful to observe the page layout conventions of nbtsearch.c:
744 : * - rightmost pages start data items at P_HIKEY instead of at P_FIRSTKEY.
745 : * - on non-leaf pages, the key portion of the first item need not be
746 : * stored, we should store only the link.
747 : *
748 : * A leaf page being built looks like:
749 : *
750 : * +----------------+---------------------------------+
751 : * | PageHeaderData | linp0 linp1 linp2 ... |
752 : * +-----------+----+---------------------------------+
753 : * | ... linpN | |
754 : * +-----------+--------------------------------------+
755 : * | ^ last |
756 : * | |
757 : * +-------------+------------------------------------+
758 : * | | itemN ... |
759 : * +-------------+------------------+-----------------+
760 : * | ... item3 item2 item1 | "special space" |
761 : * +--------------------------------+-----------------+
762 : *
763 : * Contrast this with the diagram in bufpage.h; note the mismatch
764 : * between linps and items. This is because we reserve linp0 as a
765 : * placeholder for the pointer to the "high key" item; when we have
766 : * filled up the page, we will set linp0 to point to itemN and clear
767 : * linpN. On the other hand, if we find this is the last (rightmost)
768 : * page, we leave the items alone and slide the linp array over. If
769 : * the high key is to be truncated, offset 1 is deleted, and we insert
770 : * the truncated high key at offset 1.
771 : *
772 : * 'last' pointer indicates the last offset added to the page.
773 : *
774 : * 'truncextra' is the size of the posting list in itup, if any. This
775 : * information is stashed for the next call here, when we may benefit
776 : * from considering the impact of truncating away the posting list on
777 : * the page before deciding to finish the page off. Posting lists are
778 : * often relatively large, so it is worth going to the trouble of
779 : * accounting for the saving from truncating away the posting list of
780 : * the tuple that becomes the high key (that may be the only way to
781 : * get close to target free space on the page). Note that this is
782 : * only used for the soft fillfactor-wise limit, not the critical hard
783 : * limit.
784 : *----------
785 : */
786 : static void
787 11453362 : _bt_buildadd(BTWriteState *wstate, BTPageState *state, IndexTuple itup,
788 : Size truncextra)
789 : {
790 : BulkWriteBuffer nbuf;
791 : Page npage;
792 : BlockNumber nblkno;
793 : OffsetNumber last_off;
794 : Size last_truncextra;
795 : Size pgspc;
796 : Size itupsz;
797 : bool isleaf;
798 :
799 : /*
800 : * This is a handy place to check for cancel interrupts during the btree
801 : * load phase of index creation.
802 : */
803 11453362 : CHECK_FOR_INTERRUPTS();
804 :
805 11453362 : nbuf = state->btps_buf;
806 11453362 : npage = (Page) nbuf;
807 11453362 : nblkno = state->btps_blkno;
808 11453362 : last_off = state->btps_lastoff;
809 11453362 : last_truncextra = state->btps_lastextra;
810 11453362 : state->btps_lastextra = truncextra;
811 :
812 11453362 : pgspc = PageGetFreeSpace(npage);
813 11453362 : itupsz = IndexTupleSize(itup);
814 11453362 : itupsz = MAXALIGN(itupsz);
815 : /* Leaf case has slightly different rules due to suffix truncation */
816 11453362 : isleaf = (state->btps_level == 0);
817 :
818 : /*
819 : * Check whether the new item can fit on a btree page on current level at
820 : * all.
821 : *
822 : * Every newly built index will treat heap TID as part of the keyspace,
823 : * which imposes the requirement that new high keys must occasionally have
824 : * a heap TID appended within _bt_truncate(). That may leave a new pivot
825 : * tuple one or two MAXALIGN() quantums larger than the original
826 : * firstright tuple it's derived from. v4 deals with the problem by
827 : * decreasing the limit on the size of tuples inserted on the leaf level
828 : * by the same small amount. Enforce the new v4+ limit on the leaf level,
829 : * and the old limit on internal levels, since pivot tuples may need to
830 : * make use of the reserved space. This should never fail on internal
831 : * pages.
832 : */
833 11453362 : if (unlikely(itupsz > BTMaxItemSize))
834 264 : _bt_check_third_page(wstate->index, wstate->heap, isleaf, npage,
835 : itup);
836 :
837 : /*
838 : * Check to see if current page will fit new item, with space left over to
839 : * append a heap TID during suffix truncation when page is a leaf page.
840 : *
841 : * It is guaranteed that we can fit at least 2 non-pivot tuples plus a
842 : * high key with heap TID when finishing off a leaf page, since we rely on
843 : * _bt_check_third_page() rejecting oversized non-pivot tuples. On
844 : * internal pages we can always fit 3 pivot tuples with larger internal
845 : * page tuple limit (includes page high key).
846 : *
847 : * Most of the time, a page is only "full" in the sense that the soft
848 : * fillfactor-wise limit has been exceeded. However, we must always leave
849 : * at least two items plus a high key on each page before starting a new
850 : * page. Disregard fillfactor and insert on "full" current page if we
851 : * don't have the minimum number of items yet. (Note that we deliberately
852 : * assume that suffix truncation neither enlarges nor shrinks new high key
853 : * when applying soft limit, except when last tuple has a posting list.)
854 : */
855 : Assert(last_truncextra == 0 || isleaf);
856 11453362 : if (pgspc < itupsz + (isleaf ? MAXALIGN(sizeof(ItemPointerData)) : 0) ||
857 11452206 : (pgspc + last_truncextra < state->btps_full && last_off > P_FIRSTKEY))
858 : {
859 : /*
860 : * Finish off the page and write it out.
861 : */
862 40100 : BulkWriteBuffer obuf = nbuf;
863 40100 : Page opage = npage;
864 40100 : BlockNumber oblkno = nblkno;
865 : ItemId ii;
866 : ItemId hii;
867 : IndexTuple oitup;
868 :
869 : /* Create new page of same level */
870 40100 : nbuf = _bt_blnewpage(wstate, state->btps_level);
871 40100 : npage = (Page) nbuf;
872 :
873 : /* and assign it a page position */
874 40100 : nblkno = wstate->btws_pages_alloced++;
875 :
876 : /*
877 : * We copy the last item on the page into the new page, and then
878 : * rearrange the old page so that the 'last item' becomes its high key
879 : * rather than a true data item. There had better be at least two
880 : * items on the page already, else the page would be empty of useful
881 : * data.
882 : */
883 : Assert(last_off > P_FIRSTKEY);
884 40100 : ii = PageGetItemId(opage, last_off);
885 40100 : oitup = (IndexTuple) PageGetItem(opage, ii);
886 40100 : _bt_sortaddtup(npage, ItemIdGetLength(ii), oitup, P_FIRSTKEY,
887 40100 : !isleaf);
888 :
889 : /*
890 : * Move 'last' into the high key position on opage. _bt_blnewpage()
891 : * allocated empty space for a line pointer when opage was first
892 : * created, so this is a matter of rearranging already-allocated space
893 : * on page, and initializing high key line pointer. (Actually, leaf
894 : * pages must also swap oitup with a truncated version of oitup, which
895 : * is sometimes larger than oitup, though never by more than the space
896 : * needed to append a heap TID.)
897 : */
898 40100 : hii = PageGetItemId(opage, P_HIKEY);
899 40100 : *hii = *ii;
900 40100 : ItemIdSetUnused(ii); /* redundant */
901 40100 : ((PageHeader) opage)->pd_lower -= sizeof(ItemIdData);
902 :
903 40100 : if (isleaf)
904 : {
905 : IndexTuple lastleft;
906 : IndexTuple truncated;
907 :
908 : /*
909 : * Truncate away any unneeded attributes from high key on leaf
910 : * level. This is only done at the leaf level because downlinks
911 : * in internal pages are either negative infinity items, or get
912 : * their contents from copying from one level down. See also:
913 : * _bt_split().
914 : *
915 : * We don't try to bias our choice of split point to make it more
916 : * likely that _bt_truncate() can truncate away more attributes,
917 : * whereas the split point used within _bt_split() is chosen much
918 : * more delicately. Even still, the lastleft and firstright
919 : * tuples passed to _bt_truncate() here are at least not fully
920 : * equal to each other when deduplication is used, unless there is
921 : * a large group of duplicates (also, unique index builds usually
922 : * have few or no spool2 duplicates). When the split point is
923 : * between two unequal tuples, _bt_truncate() will avoid including
924 : * a heap TID in the new high key, which is the most important
925 : * benefit of suffix truncation.
926 : *
927 : * Overwrite the old item with new truncated high key directly.
928 : * oitup is already located at the physical beginning of tuple
929 : * space, so this should directly reuse the existing tuple space.
930 : */
931 39976 : ii = PageGetItemId(opage, OffsetNumberPrev(last_off));
932 39976 : lastleft = (IndexTuple) PageGetItem(opage, ii);
933 :
934 : Assert(IndexTupleSize(oitup) > last_truncextra);
935 39976 : truncated = _bt_truncate(wstate->index, lastleft, oitup,
936 : wstate->inskey);
937 39976 : if (!PageIndexTupleOverwrite(opage, P_HIKEY, (Item) truncated,
938 : IndexTupleSize(truncated)))
939 0 : elog(ERROR, "failed to add high key to the index page");
940 39976 : pfree(truncated);
941 :
942 : /* oitup should continue to point to the page's high key */
943 39976 : hii = PageGetItemId(opage, P_HIKEY);
944 39976 : oitup = (IndexTuple) PageGetItem(opage, hii);
945 : }
946 :
947 : /*
948 : * Link the old page into its parent, using its low key. If we don't
949 : * have a parent, we have to create one; this adds a new btree level.
950 : */
951 40100 : if (state->btps_next == NULL)
952 2692 : state->btps_next = _bt_pagestate(wstate, state->btps_level + 1);
953 :
954 : Assert((BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) <=
955 : IndexRelationGetNumberOfKeyAttributes(wstate->index) &&
956 : BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) > 0) ||
957 : P_LEFTMOST(BTPageGetOpaque(opage)));
958 : Assert(BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) == 0 ||
959 : !P_LEFTMOST(BTPageGetOpaque(opage)));
960 40100 : BTreeTupleSetDownLink(state->btps_lowkey, oblkno);
961 40100 : _bt_buildadd(wstate, state->btps_next, state->btps_lowkey, 0);
962 40100 : pfree(state->btps_lowkey);
963 :
964 : /*
965 : * Save a copy of the high key from the old page. It is also the low
966 : * key for the new page.
967 : */
968 40100 : state->btps_lowkey = CopyIndexTuple(oitup);
969 :
970 : /*
971 : * Set the sibling links for both pages.
972 : */
973 : {
974 40100 : BTPageOpaque oopaque = BTPageGetOpaque(opage);
975 40100 : BTPageOpaque nopaque = BTPageGetOpaque(npage);
976 :
977 40100 : oopaque->btpo_next = nblkno;
978 40100 : nopaque->btpo_prev = oblkno;
979 40100 : nopaque->btpo_next = P_NONE; /* redundant */
980 : }
981 :
982 : /*
983 : * Write out the old page. _bt_blwritepage takes ownership of the
984 : * 'opage' buffer.
985 : */
986 40100 : _bt_blwritepage(wstate, obuf, oblkno);
987 :
988 : /*
989 : * Reset last_off to point to new page
990 : */
991 40100 : last_off = P_FIRSTKEY;
992 : }
993 :
994 : /*
995 : * By here, either original page is still the current page, or a new page
996 : * was created that became the current page. Either way, the current page
997 : * definitely has space for new item.
998 : *
999 : * If the new item is the first for its page, it must also be the first
1000 : * item on its entire level. On later same-level pages, a low key for a
1001 : * page will be copied from the prior page in the code above. Generate a
1002 : * minus infinity low key here instead.
1003 : */
1004 11453362 : if (last_off == P_HIKEY)
1005 : {
1006 : Assert(state->btps_lowkey == NULL);
1007 12862 : state->btps_lowkey = palloc0(sizeof(IndexTupleData));
1008 12862 : state->btps_lowkey->t_info = sizeof(IndexTupleData);
1009 12862 : BTreeTupleSetNAtts(state->btps_lowkey, 0, false);
1010 : }
1011 :
1012 : /*
1013 : * Add the new item into the current page.
1014 : */
1015 11453362 : last_off = OffsetNumberNext(last_off);
1016 11453362 : _bt_sortaddtup(npage, itupsz, itup, last_off,
1017 11453362 : !isleaf && last_off == P_FIRSTKEY);
1018 :
1019 11453362 : state->btps_buf = nbuf;
1020 11453362 : state->btps_blkno = nblkno;
1021 11453362 : state->btps_lastoff = last_off;
1022 11453362 : }
1023 :
1024 : /*
1025 : * Finalize pending posting list tuple, and add it to the index. Final tuple
1026 : * is based on saved base tuple, and saved list of heap TIDs.
1027 : *
1028 : * This is almost like _bt_dedup_finish_pending(), but it adds a new tuple
1029 : * using _bt_buildadd().
1030 : */
1031 : static void
1032 4740492 : _bt_sort_dedup_finish_pending(BTWriteState *wstate, BTPageState *state,
1033 : BTDedupState dstate)
1034 : {
1035 : Assert(dstate->nitems > 0);
1036 :
1037 4740492 : if (dstate->nitems == 1)
1038 4700334 : _bt_buildadd(wstate, state, dstate->base, 0);
1039 : else
1040 : {
1041 : IndexTuple postingtuple;
1042 : Size truncextra;
1043 :
1044 : /* form a tuple with a posting list */
1045 40158 : postingtuple = _bt_form_posting(dstate->base,
1046 : dstate->htids,
1047 : dstate->nhtids);
1048 : /* Calculate posting list overhead */
1049 40158 : truncextra = IndexTupleSize(postingtuple) -
1050 40158 : BTreeTupleGetPostingOffset(postingtuple);
1051 :
1052 40158 : _bt_buildadd(wstate, state, postingtuple, truncextra);
1053 40158 : pfree(postingtuple);
1054 : }
1055 :
1056 4740492 : dstate->nmaxitems = 0;
1057 4740492 : dstate->nhtids = 0;
1058 4740492 : dstate->nitems = 0;
1059 4740492 : dstate->phystupsize = 0;
1060 4740492 : }
1061 :
1062 : /*
1063 : * Finish writing out the completed btree.
1064 : */
1065 : static void
1066 50202 : _bt_uppershutdown(BTWriteState *wstate, BTPageState *state)
1067 : {
1068 : BTPageState *s;
1069 50202 : BlockNumber rootblkno = P_NONE;
1070 50202 : uint32 rootlevel = 0;
1071 : BulkWriteBuffer metabuf;
1072 :
1073 : /*
1074 : * Each iteration of this loop completes one more level of the tree.
1075 : */
1076 63064 : for (s = state; s != NULL; s = s->btps_next)
1077 : {
1078 : BlockNumber blkno;
1079 : BTPageOpaque opaque;
1080 :
1081 12862 : blkno = s->btps_blkno;
1082 12862 : opaque = BTPageGetOpaque((Page) s->btps_buf);
1083 :
1084 : /*
1085 : * We have to link the last page on this level to somewhere.
1086 : *
1087 : * If we're at the top, it's the root, so attach it to the metapage.
1088 : * Otherwise, add an entry for it to its parent using its low key.
1089 : * This may cause the last page of the parent level to split, but
1090 : * that's not a problem -- we haven't gotten to it yet.
1091 : */
1092 12862 : if (s->btps_next == NULL)
1093 : {
1094 10170 : opaque->btpo_flags |= BTP_ROOT;
1095 10170 : rootblkno = blkno;
1096 10170 : rootlevel = s->btps_level;
1097 : }
1098 : else
1099 : {
1100 : Assert((BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) <=
1101 : IndexRelationGetNumberOfKeyAttributes(wstate->index) &&
1102 : BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) > 0) ||
1103 : P_LEFTMOST(opaque));
1104 : Assert(BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) == 0 ||
1105 : !P_LEFTMOST(opaque));
1106 2692 : BTreeTupleSetDownLink(s->btps_lowkey, blkno);
1107 2692 : _bt_buildadd(wstate, s->btps_next, s->btps_lowkey, 0);
1108 2692 : pfree(s->btps_lowkey);
1109 2692 : s->btps_lowkey = NULL;
1110 : }
1111 :
1112 : /*
1113 : * This is the rightmost page, so the ItemId array needs to be slid
1114 : * back one slot. Then we can dump out the page.
1115 : */
1116 12862 : _bt_slideleft((Page) s->btps_buf);
1117 12862 : _bt_blwritepage(wstate, s->btps_buf, s->btps_blkno);
1118 12862 : s->btps_buf = NULL; /* writepage took ownership of the buffer */
1119 : }
1120 :
1121 : /*
1122 : * As the last step in the process, construct the metapage and make it
1123 : * point to the new root (unless we had no data at all, in which case it's
1124 : * set to point to "P_NONE"). This changes the index to the "valid" state
1125 : * by filling in a valid magic number in the metapage.
1126 : */
1127 50202 : metabuf = smgr_bulk_get_buf(wstate->bulkstate);
1128 50202 : _bt_initmetapage((Page) metabuf, rootblkno, rootlevel,
1129 50202 : wstate->inskey->allequalimage);
1130 50202 : _bt_blwritepage(wstate, metabuf, BTREE_METAPAGE);
1131 50202 : }
1132 :
1133 : /*
1134 : * Read tuples in correct sort order from tuplesort, and load them into
1135 : * btree leaves.
1136 : */
1137 : static void
1138 50202 : _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
1139 : {
1140 50202 : BTPageState *state = NULL;
1141 50202 : bool merge = (btspool2 != NULL);
1142 : IndexTuple itup,
1143 50202 : itup2 = NULL;
1144 : bool load1;
1145 50202 : TupleDesc tupdes = RelationGetDescr(wstate->index);
1146 : int i,
1147 50202 : keysz = IndexRelationGetNumberOfKeyAttributes(wstate->index);
1148 : SortSupport sortKeys;
1149 50202 : int64 tuples_done = 0;
1150 : bool deduplicate;
1151 :
1152 50202 : wstate->bulkstate = smgr_bulk_start_rel(wstate->index, MAIN_FORKNUM);
1153 :
1154 59120 : deduplicate = wstate->inskey->allequalimage && !btspool->isunique &&
1155 8918 : BTGetDeduplicateItems(wstate->index);
1156 :
1157 50202 : if (merge)
1158 : {
1159 : /*
1160 : * Another BTSpool for dead tuples exists. Now we have to merge
1161 : * btspool and btspool2.
1162 : */
1163 :
1164 : /* the preparation of merge */
1165 20 : itup = tuplesort_getindextuple(btspool->sortstate, true);
1166 20 : itup2 = tuplesort_getindextuple(btspool2->sortstate, true);
1167 :
1168 : /* Prepare SortSupport data for each column */
1169 20 : sortKeys = (SortSupport) palloc0(keysz * sizeof(SortSupportData));
1170 :
1171 42 : for (i = 0; i < keysz; i++)
1172 : {
1173 22 : SortSupport sortKey = sortKeys + i;
1174 22 : ScanKey scanKey = wstate->inskey->scankeys + i;
1175 : bool reverse;
1176 :
1177 22 : sortKey->ssup_cxt = CurrentMemoryContext;
1178 22 : sortKey->ssup_collation = scanKey->sk_collation;
1179 22 : sortKey->ssup_nulls_first =
1180 22 : (scanKey->sk_flags & SK_BT_NULLS_FIRST) != 0;
1181 22 : sortKey->ssup_attno = scanKey->sk_attno;
1182 : /* Abbreviation is not supported here */
1183 22 : sortKey->abbreviate = false;
1184 :
1185 : Assert(sortKey->ssup_attno != 0);
1186 :
1187 22 : reverse = (scanKey->sk_flags & SK_BT_DESC) != 0;
1188 :
1189 22 : PrepareSortSupportFromIndexRel(wstate->index, reverse, sortKey);
1190 : }
1191 :
1192 : for (;;)
1193 : {
1194 3306 : load1 = true; /* load BTSpool next ? */
1195 3306 : if (itup2 == NULL)
1196 : {
1197 148 : if (itup == NULL)
1198 20 : break;
1199 : }
1200 3158 : else if (itup != NULL)
1201 : {
1202 2964 : int32 compare = 0;
1203 :
1204 3200 : for (i = 1; i <= keysz; i++)
1205 : {
1206 : SortSupport entry;
1207 : Datum attrDatum1,
1208 : attrDatum2;
1209 : bool isNull1,
1210 : isNull2;
1211 :
1212 3016 : entry = sortKeys + i - 1;
1213 3016 : attrDatum1 = index_getattr(itup, i, tupdes, &isNull1);
1214 3016 : attrDatum2 = index_getattr(itup2, i, tupdes, &isNull2);
1215 :
1216 3016 : compare = ApplySortComparator(attrDatum1, isNull1,
1217 : attrDatum2, isNull2,
1218 : entry);
1219 3016 : if (compare > 0)
1220 : {
1221 276 : load1 = false;
1222 2780 : break;
1223 : }
1224 2740 : else if (compare < 0)
1225 2504 : break;
1226 : }
1227 :
1228 : /*
1229 : * If key values are equal, we sort on ItemPointer. This is
1230 : * required for btree indexes, since heap TID is treated as an
1231 : * implicit last key attribute in order to ensure that all
1232 : * keys in the index are physically unique.
1233 : */
1234 2964 : if (compare == 0)
1235 : {
1236 184 : compare = ItemPointerCompare(&itup->t_tid, &itup2->t_tid);
1237 : Assert(compare != 0);
1238 184 : if (compare > 0)
1239 20 : load1 = false;
1240 : }
1241 : }
1242 : else
1243 194 : load1 = false;
1244 :
1245 : /* When we see first tuple, create first index page */
1246 3286 : if (state == NULL)
1247 20 : state = _bt_pagestate(wstate, 0);
1248 :
1249 3286 : if (load1)
1250 : {
1251 2796 : _bt_buildadd(wstate, state, itup, 0);
1252 2796 : itup = tuplesort_getindextuple(btspool->sortstate, true);
1253 : }
1254 : else
1255 : {
1256 490 : _bt_buildadd(wstate, state, itup2, 0);
1257 490 : itup2 = tuplesort_getindextuple(btspool2->sortstate, true);
1258 : }
1259 :
1260 : /* Report progress */
1261 3286 : pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_DONE,
1262 : ++tuples_done);
1263 : }
1264 20 : pfree(sortKeys);
1265 : }
1266 50182 : else if (deduplicate)
1267 : {
1268 : /* merge is unnecessary, deduplicate into posting lists */
1269 : BTDedupState dstate;
1270 :
1271 8918 : dstate = (BTDedupState) palloc(sizeof(BTDedupStateData));
1272 8918 : dstate->deduplicate = true; /* unused */
1273 8918 : dstate->nmaxitems = 0; /* unused */
1274 8918 : dstate->maxpostingsize = 0; /* set later */
1275 : /* Metadata about base tuple of current pending posting list */
1276 8918 : dstate->base = NULL;
1277 8918 : dstate->baseoff = InvalidOffsetNumber; /* unused */
1278 8918 : dstate->basetupsize = 0;
1279 : /* Metadata about current pending posting list TIDs */
1280 8918 : dstate->htids = NULL;
1281 8918 : dstate->nhtids = 0;
1282 8918 : dstate->nitems = 0;
1283 8918 : dstate->phystupsize = 0; /* unused */
1284 8918 : dstate->nintervals = 0; /* unused */
1285 :
1286 6115204 : while ((itup = tuplesort_getindextuple(btspool->sortstate,
1287 6115204 : true)) != NULL)
1288 : {
1289 : /* When we see first tuple, create first index page */
1290 6106286 : if (state == NULL)
1291 : {
1292 2432 : state = _bt_pagestate(wstate, 0);
1293 :
1294 : /*
1295 : * Limit size of posting list tuples to 1/10 space we want to
1296 : * leave behind on the page, plus space for final item's line
1297 : * pointer. This is equal to the space that we'd like to
1298 : * leave behind on each leaf page when fillfactor is 90,
1299 : * allowing us to get close to fillfactor% space utilization
1300 : * when there happen to be a great many duplicates. (This
1301 : * makes higher leaf fillfactor settings ineffective when
1302 : * building indexes that have many duplicates, but packing
1303 : * leaf pages full with few very large tuples doesn't seem
1304 : * like a useful goal.)
1305 : */
1306 2432 : dstate->maxpostingsize = MAXALIGN_DOWN((BLCKSZ * 10 / 100)) -
1307 : sizeof(ItemIdData);
1308 : Assert(dstate->maxpostingsize <= BTMaxItemSize &&
1309 : dstate->maxpostingsize <= INDEX_SIZE_MASK);
1310 2432 : dstate->htids = palloc(dstate->maxpostingsize);
1311 :
1312 : /* start new pending posting list with itup copy */
1313 2432 : _bt_dedup_start_pending(dstate, CopyIndexTuple(itup),
1314 : InvalidOffsetNumber);
1315 : }
1316 6103854 : else if (_bt_keep_natts_fast(wstate->index, dstate->base,
1317 1373710 : itup) > keysz &&
1318 1373710 : _bt_dedup_save_htid(dstate, itup))
1319 : {
1320 : /*
1321 : * Tuple is equal to base tuple of pending posting list. Heap
1322 : * TID from itup has been saved in state.
1323 : */
1324 : }
1325 : else
1326 : {
1327 : /*
1328 : * Tuple is not equal to pending posting list tuple, or
1329 : * _bt_dedup_save_htid() opted to not merge current item into
1330 : * pending posting list.
1331 : */
1332 4738060 : _bt_sort_dedup_finish_pending(wstate, state, dstate);
1333 4738060 : pfree(dstate->base);
1334 :
1335 : /* start new pending posting list with itup copy */
1336 4738060 : _bt_dedup_start_pending(dstate, CopyIndexTuple(itup),
1337 : InvalidOffsetNumber);
1338 : }
1339 :
1340 : /* Report progress */
1341 6106286 : pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_DONE,
1342 : ++tuples_done);
1343 : }
1344 :
1345 8918 : if (state)
1346 : {
1347 : /*
1348 : * Handle the last item (there must be a last item when the
1349 : * tuplesort returned one or more tuples)
1350 : */
1351 2432 : _bt_sort_dedup_finish_pending(wstate, state, dstate);
1352 2432 : pfree(dstate->base);
1353 2432 : pfree(dstate->htids);
1354 : }
1355 :
1356 8918 : pfree(dstate);
1357 : }
1358 : else
1359 : {
1360 : /* merging and deduplication are both unnecessary */
1361 6708056 : while ((itup = tuplesort_getindextuple(btspool->sortstate,
1362 6708056 : true)) != NULL)
1363 : {
1364 : /* When we see first tuple, create first index page */
1365 6666792 : if (state == NULL)
1366 7718 : state = _bt_pagestate(wstate, 0);
1367 :
1368 6666792 : _bt_buildadd(wstate, state, itup, 0);
1369 :
1370 : /* Report progress */
1371 6666792 : pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_DONE,
1372 : ++tuples_done);
1373 : }
1374 : }
1375 :
1376 : /* Close down final pages and write the metapage */
1377 50202 : _bt_uppershutdown(wstate, state);
1378 50202 : smgr_bulk_finish(wstate->bulkstate);
1379 50202 : }
1380 :
1381 : /*
1382 : * Create parallel context, and launch workers for leader.
1383 : *
1384 : * buildstate argument should be initialized (with the exception of the
1385 : * tuplesort state in spools, which may later be created based on shared
1386 : * state initially set up here).
1387 : *
1388 : * isconcurrent indicates if operation is CREATE INDEX CONCURRENTLY.
1389 : *
1390 : * request is the target number of parallel worker processes to launch.
1391 : *
1392 : * Sets buildstate's BTLeader, which caller must use to shut down parallel
1393 : * mode by passing it to _bt_end_parallel() at the very end of its index
1394 : * build. If not even a single worker process can be launched, this is
1395 : * never set, and caller should proceed with a serial index build.
1396 : */
1397 : static void
1398 148 : _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent, int request)
1399 : {
1400 : ParallelContext *pcxt;
1401 : int scantuplesortstates;
1402 : Snapshot snapshot;
1403 : Size estbtshared;
1404 : Size estsort;
1405 : BTShared *btshared;
1406 : Sharedsort *sharedsort;
1407 : Sharedsort *sharedsort2;
1408 148 : BTSpool *btspool = buildstate->spool;
1409 148 : BTLeader *btleader = (BTLeader *) palloc0(sizeof(BTLeader));
1410 : WalUsage *walusage;
1411 : BufferUsage *bufferusage;
1412 148 : bool leaderparticipates = true;
1413 : int querylen;
1414 :
1415 : #ifdef DISABLE_LEADER_PARTICIPATION
1416 : leaderparticipates = false;
1417 : #endif
1418 :
1419 : /*
1420 : * Enter parallel mode, and create context for parallel build of btree
1421 : * index
1422 : */
1423 148 : EnterParallelMode();
1424 : Assert(request > 0);
1425 148 : pcxt = CreateParallelContext("postgres", "_bt_parallel_build_main",
1426 : request);
1427 :
1428 148 : scantuplesortstates = leaderparticipates ? request + 1 : request;
1429 :
1430 : /*
1431 : * Prepare for scan of the base relation. In a normal index build, we use
1432 : * SnapshotAny because we must retrieve all tuples and do our own time
1433 : * qual checks (because we have to index RECENTLY_DEAD tuples). In a
1434 : * concurrent build, we take a regular MVCC snapshot and index whatever's
1435 : * live according to that.
1436 : */
1437 148 : if (!isconcurrent)
1438 148 : snapshot = SnapshotAny;
1439 : else
1440 0 : snapshot = RegisterSnapshot(GetTransactionSnapshot());
1441 :
1442 : /*
1443 : * Estimate size for our own PARALLEL_KEY_BTREE_SHARED workspace, and
1444 : * PARALLEL_KEY_TUPLESORT tuplesort workspace
1445 : */
1446 148 : estbtshared = _bt_parallel_estimate_shared(btspool->heap, snapshot);
1447 148 : shm_toc_estimate_chunk(&pcxt->estimator, estbtshared);
1448 148 : estsort = tuplesort_estimate_shared(scantuplesortstates);
1449 148 : shm_toc_estimate_chunk(&pcxt->estimator, estsort);
1450 :
1451 : /*
1452 : * Unique case requires a second spool, and so we may have to account for
1453 : * another shared workspace for that -- PARALLEL_KEY_TUPLESORT_SPOOL2
1454 : */
1455 148 : if (!btspool->isunique)
1456 84 : shm_toc_estimate_keys(&pcxt->estimator, 2);
1457 : else
1458 : {
1459 64 : shm_toc_estimate_chunk(&pcxt->estimator, estsort);
1460 64 : shm_toc_estimate_keys(&pcxt->estimator, 3);
1461 : }
1462 :
1463 : /*
1464 : * Estimate space for WalUsage and BufferUsage -- PARALLEL_KEY_WAL_USAGE
1465 : * and PARALLEL_KEY_BUFFER_USAGE.
1466 : *
1467 : * If there are no extensions loaded that care, we could skip this. We
1468 : * have no way of knowing whether anyone's looking at pgWalUsage or
1469 : * pgBufferUsage, so do it unconditionally.
1470 : */
1471 148 : shm_toc_estimate_chunk(&pcxt->estimator,
1472 : mul_size(sizeof(WalUsage), pcxt->nworkers));
1473 148 : shm_toc_estimate_keys(&pcxt->estimator, 1);
1474 148 : shm_toc_estimate_chunk(&pcxt->estimator,
1475 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
1476 148 : shm_toc_estimate_keys(&pcxt->estimator, 1);
1477 :
1478 : /* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
1479 148 : if (debug_query_string)
1480 : {
1481 148 : querylen = strlen(debug_query_string);
1482 148 : shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
1483 148 : shm_toc_estimate_keys(&pcxt->estimator, 1);
1484 : }
1485 : else
1486 0 : querylen = 0; /* keep compiler quiet */
1487 :
1488 : /* Everyone's had a chance to ask for space, so now create the DSM */
1489 148 : InitializeParallelDSM(pcxt);
1490 :
1491 : /* If no DSM segment was available, back out (do serial build) */
1492 148 : if (pcxt->seg == NULL)
1493 : {
1494 0 : if (IsMVCCSnapshot(snapshot))
1495 0 : UnregisterSnapshot(snapshot);
1496 0 : DestroyParallelContext(pcxt);
1497 0 : ExitParallelMode();
1498 0 : return;
1499 : }
1500 :
1501 : /* Store shared build state, for which we reserved space */
1502 148 : btshared = (BTShared *) shm_toc_allocate(pcxt->toc, estbtshared);
1503 : /* Initialize immutable state */
1504 148 : btshared->heaprelid = RelationGetRelid(btspool->heap);
1505 148 : btshared->indexrelid = RelationGetRelid(btspool->index);
1506 148 : btshared->isunique = btspool->isunique;
1507 148 : btshared->nulls_not_distinct = btspool->nulls_not_distinct;
1508 148 : btshared->isconcurrent = isconcurrent;
1509 148 : btshared->scantuplesortstates = scantuplesortstates;
1510 148 : btshared->queryid = pgstat_get_my_query_id();
1511 148 : ConditionVariableInit(&btshared->workersdonecv);
1512 148 : SpinLockInit(&btshared->mutex);
1513 : /* Initialize mutable state */
1514 148 : btshared->nparticipantsdone = 0;
1515 148 : btshared->reltuples = 0.0;
1516 148 : btshared->havedead = false;
1517 148 : btshared->indtuples = 0.0;
1518 148 : btshared->brokenhotchain = false;
1519 148 : table_parallelscan_initialize(btspool->heap,
1520 : ParallelTableScanFromBTShared(btshared),
1521 : snapshot);
1522 :
1523 : /*
1524 : * Store shared tuplesort-private state, for which we reserved space.
1525 : * Then, initialize opaque state using tuplesort routine.
1526 : */
1527 148 : sharedsort = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
1528 148 : tuplesort_initialize_shared(sharedsort, scantuplesortstates,
1529 : pcxt->seg);
1530 :
1531 148 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_BTREE_SHARED, btshared);
1532 148 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLESORT, sharedsort);
1533 :
1534 : /* Unique case requires a second spool, and associated shared state */
1535 148 : if (!btspool->isunique)
1536 84 : sharedsort2 = NULL;
1537 : else
1538 : {
1539 : /*
1540 : * Store additional shared tuplesort-private state, for which we
1541 : * reserved space. Then, initialize opaque state using tuplesort
1542 : * routine.
1543 : */
1544 64 : sharedsort2 = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
1545 64 : tuplesort_initialize_shared(sharedsort2, scantuplesortstates,
1546 : pcxt->seg);
1547 :
1548 64 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLESORT_SPOOL2, sharedsort2);
1549 : }
1550 :
1551 : /* Store query string for workers */
1552 148 : if (debug_query_string)
1553 : {
1554 : char *sharedquery;
1555 :
1556 148 : sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
1557 148 : memcpy(sharedquery, debug_query_string, querylen + 1);
1558 148 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, sharedquery);
1559 : }
1560 :
1561 : /*
1562 : * Allocate space for each worker's WalUsage and BufferUsage; no need to
1563 : * initialize.
1564 : */
1565 148 : walusage = shm_toc_allocate(pcxt->toc,
1566 148 : mul_size(sizeof(WalUsage), pcxt->nworkers));
1567 148 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_WAL_USAGE, walusage);
1568 148 : bufferusage = shm_toc_allocate(pcxt->toc,
1569 148 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
1570 148 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_BUFFER_USAGE, bufferusage);
1571 :
1572 : /* Launch workers, saving status for leader/caller */
1573 148 : LaunchParallelWorkers(pcxt);
1574 148 : btleader->pcxt = pcxt;
1575 148 : btleader->nparticipanttuplesorts = pcxt->nworkers_launched;
1576 148 : if (leaderparticipates)
1577 148 : btleader->nparticipanttuplesorts++;
1578 148 : btleader->btshared = btshared;
1579 148 : btleader->sharedsort = sharedsort;
1580 148 : btleader->sharedsort2 = sharedsort2;
1581 148 : btleader->snapshot = snapshot;
1582 148 : btleader->walusage = walusage;
1583 148 : btleader->bufferusage = bufferusage;
1584 :
1585 : /* If no workers were successfully launched, back out (do serial build) */
1586 148 : if (pcxt->nworkers_launched == 0)
1587 : {
1588 0 : _bt_end_parallel(btleader);
1589 0 : return;
1590 : }
1591 :
1592 : /* Save leader state now that it's clear build will be parallel */
1593 148 : buildstate->btleader = btleader;
1594 :
1595 : /* Join heap scan ourselves */
1596 148 : if (leaderparticipates)
1597 148 : _bt_leader_participate_as_worker(buildstate);
1598 :
1599 : /*
1600 : * Caller needs to wait for all launched workers when we return. Make
1601 : * sure that the failure-to-start case will not hang forever.
1602 : */
1603 148 : WaitForParallelWorkersToAttach(pcxt);
1604 : }
1605 :
1606 : /*
1607 : * Shut down workers, destroy parallel context, and end parallel mode.
1608 : */
1609 : static void
1610 148 : _bt_end_parallel(BTLeader *btleader)
1611 : {
1612 : int i;
1613 :
1614 : /* Shutdown worker processes */
1615 148 : WaitForParallelWorkersToFinish(btleader->pcxt);
1616 :
1617 : /*
1618 : * Next, accumulate WAL usage. (This must wait for the workers to finish,
1619 : * or we might get incomplete data.)
1620 : */
1621 296 : for (i = 0; i < btleader->pcxt->nworkers_launched; i++)
1622 148 : InstrAccumParallelQuery(&btleader->bufferusage[i], &btleader->walusage[i]);
1623 :
1624 : /* Free last reference to MVCC snapshot, if one was used */
1625 148 : if (IsMVCCSnapshot(btleader->snapshot))
1626 0 : UnregisterSnapshot(btleader->snapshot);
1627 148 : DestroyParallelContext(btleader->pcxt);
1628 148 : ExitParallelMode();
1629 148 : }
1630 :
1631 : /*
1632 : * Returns size of shared memory required to store state for a parallel
1633 : * btree index build based on the snapshot its parallel scan will use.
1634 : */
1635 : static Size
1636 148 : _bt_parallel_estimate_shared(Relation heap, Snapshot snapshot)
1637 : {
1638 : /* c.f. shm_toc_allocate as to why BUFFERALIGN is used */
1639 148 : return add_size(BUFFERALIGN(sizeof(BTShared)),
1640 : table_parallelscan_estimate(heap, snapshot));
1641 : }
1642 :
1643 : /*
1644 : * Within leader, wait for end of heap scan.
1645 : *
1646 : * When called, parallel heap scan started by _bt_begin_parallel() will
1647 : * already be underway within worker processes (when leader participates
1648 : * as a worker, we should end up here just as workers are finishing).
1649 : *
1650 : * Fills in fields needed for ambuild statistics, and lets caller set
1651 : * field indicating that some worker encountered a broken HOT chain.
1652 : *
1653 : * Returns the total number of heap tuples scanned.
1654 : */
1655 : static double
1656 148 : _bt_parallel_heapscan(BTBuildState *buildstate, bool *brokenhotchain)
1657 : {
1658 148 : BTShared *btshared = buildstate->btleader->btshared;
1659 : int nparticipanttuplesorts;
1660 : double reltuples;
1661 :
1662 148 : nparticipanttuplesorts = buildstate->btleader->nparticipanttuplesorts;
1663 : for (;;)
1664 : {
1665 340 : SpinLockAcquire(&btshared->mutex);
1666 340 : if (btshared->nparticipantsdone == nparticipanttuplesorts)
1667 : {
1668 148 : buildstate->havedead = btshared->havedead;
1669 148 : buildstate->indtuples = btshared->indtuples;
1670 148 : *brokenhotchain = btshared->brokenhotchain;
1671 148 : reltuples = btshared->reltuples;
1672 148 : SpinLockRelease(&btshared->mutex);
1673 148 : break;
1674 : }
1675 192 : SpinLockRelease(&btshared->mutex);
1676 :
1677 192 : ConditionVariableSleep(&btshared->workersdonecv,
1678 : WAIT_EVENT_PARALLEL_CREATE_INDEX_SCAN);
1679 : }
1680 :
1681 148 : ConditionVariableCancelSleep();
1682 :
1683 148 : return reltuples;
1684 : }
1685 :
1686 : /*
1687 : * Within leader, participate as a parallel worker.
1688 : */
1689 : static void
1690 148 : _bt_leader_participate_as_worker(BTBuildState *buildstate)
1691 : {
1692 148 : BTLeader *btleader = buildstate->btleader;
1693 : BTSpool *leaderworker;
1694 : BTSpool *leaderworker2;
1695 : int sortmem;
1696 :
1697 : /* Allocate memory and initialize private spool */
1698 148 : leaderworker = (BTSpool *) palloc0(sizeof(BTSpool));
1699 148 : leaderworker->heap = buildstate->spool->heap;
1700 148 : leaderworker->index = buildstate->spool->index;
1701 148 : leaderworker->isunique = buildstate->spool->isunique;
1702 148 : leaderworker->nulls_not_distinct = buildstate->spool->nulls_not_distinct;
1703 :
1704 : /* Initialize second spool, if required */
1705 148 : if (!btleader->btshared->isunique)
1706 84 : leaderworker2 = NULL;
1707 : else
1708 : {
1709 : /* Allocate memory for worker's own private secondary spool */
1710 64 : leaderworker2 = (BTSpool *) palloc0(sizeof(BTSpool));
1711 :
1712 : /* Initialize worker's own secondary spool */
1713 64 : leaderworker2->heap = leaderworker->heap;
1714 64 : leaderworker2->index = leaderworker->index;
1715 64 : leaderworker2->isunique = false;
1716 : }
1717 :
1718 : /*
1719 : * Might as well use reliable figure when doling out maintenance_work_mem
1720 : * (when requested number of workers were not launched, this will be
1721 : * somewhat higher than it is for other workers).
1722 : */
1723 148 : sortmem = maintenance_work_mem / btleader->nparticipanttuplesorts;
1724 :
1725 : /* Perform work common to all participants */
1726 148 : _bt_parallel_scan_and_sort(leaderworker, leaderworker2, btleader->btshared,
1727 : btleader->sharedsort, btleader->sharedsort2,
1728 : sortmem, true);
1729 :
1730 : #ifdef BTREE_BUILD_STATS
1731 : if (log_btree_build_stats)
1732 : {
1733 : ShowUsage("BTREE BUILD (Leader Partial Spool) STATISTICS");
1734 : ResetUsage();
1735 : }
1736 : #endif /* BTREE_BUILD_STATS */
1737 148 : }
1738 :
1739 : /*
1740 : * Perform work within a launched parallel process.
1741 : */
1742 : void
1743 148 : _bt_parallel_build_main(dsm_segment *seg, shm_toc *toc)
1744 : {
1745 : char *sharedquery;
1746 : BTSpool *btspool;
1747 : BTSpool *btspool2;
1748 : BTShared *btshared;
1749 : Sharedsort *sharedsort;
1750 : Sharedsort *sharedsort2;
1751 : Relation heapRel;
1752 : Relation indexRel;
1753 : LOCKMODE heapLockmode;
1754 : LOCKMODE indexLockmode;
1755 : WalUsage *walusage;
1756 : BufferUsage *bufferusage;
1757 : int sortmem;
1758 :
1759 : #ifdef BTREE_BUILD_STATS
1760 : if (log_btree_build_stats)
1761 : ResetUsage();
1762 : #endif /* BTREE_BUILD_STATS */
1763 :
1764 : /*
1765 : * The only possible status flag that can be set to the parallel worker is
1766 : * PROC_IN_SAFE_IC.
1767 : */
1768 : Assert((MyProc->statusFlags == 0) ||
1769 : (MyProc->statusFlags == PROC_IN_SAFE_IC));
1770 :
1771 : /* Set debug_query_string for individual workers first */
1772 148 : sharedquery = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, true);
1773 148 : debug_query_string = sharedquery;
1774 :
1775 : /* Report the query string from leader */
1776 148 : pgstat_report_activity(STATE_RUNNING, debug_query_string);
1777 :
1778 : /* Look up nbtree shared state */
1779 148 : btshared = shm_toc_lookup(toc, PARALLEL_KEY_BTREE_SHARED, false);
1780 :
1781 : /* Open relations using lock modes known to be obtained by index.c */
1782 148 : if (!btshared->isconcurrent)
1783 : {
1784 148 : heapLockmode = ShareLock;
1785 148 : indexLockmode = AccessExclusiveLock;
1786 : }
1787 : else
1788 : {
1789 0 : heapLockmode = ShareUpdateExclusiveLock;
1790 0 : indexLockmode = RowExclusiveLock;
1791 : }
1792 :
1793 : /* Track query ID */
1794 148 : pgstat_report_query_id(btshared->queryid, false);
1795 :
1796 : /* Open relations within worker */
1797 148 : heapRel = table_open(btshared->heaprelid, heapLockmode);
1798 148 : indexRel = index_open(btshared->indexrelid, indexLockmode);
1799 :
1800 : /* Initialize worker's own spool */
1801 148 : btspool = (BTSpool *) palloc0(sizeof(BTSpool));
1802 148 : btspool->heap = heapRel;
1803 148 : btspool->index = indexRel;
1804 148 : btspool->isunique = btshared->isunique;
1805 148 : btspool->nulls_not_distinct = btshared->nulls_not_distinct;
1806 :
1807 : /* Look up shared state private to tuplesort.c */
1808 148 : sharedsort = shm_toc_lookup(toc, PARALLEL_KEY_TUPLESORT, false);
1809 148 : tuplesort_attach_shared(sharedsort, seg);
1810 148 : if (!btshared->isunique)
1811 : {
1812 84 : btspool2 = NULL;
1813 84 : sharedsort2 = NULL;
1814 : }
1815 : else
1816 : {
1817 : /* Allocate memory for worker's own private secondary spool */
1818 64 : btspool2 = (BTSpool *) palloc0(sizeof(BTSpool));
1819 :
1820 : /* Initialize worker's own secondary spool */
1821 64 : btspool2->heap = btspool->heap;
1822 64 : btspool2->index = btspool->index;
1823 64 : btspool2->isunique = false;
1824 : /* Look up shared state private to tuplesort.c */
1825 64 : sharedsort2 = shm_toc_lookup(toc, PARALLEL_KEY_TUPLESORT_SPOOL2, false);
1826 64 : tuplesort_attach_shared(sharedsort2, seg);
1827 : }
1828 :
1829 : /* Prepare to track buffer usage during parallel execution */
1830 148 : InstrStartParallelQuery();
1831 :
1832 : /* Perform sorting of spool, and possibly a spool2 */
1833 148 : sortmem = maintenance_work_mem / btshared->scantuplesortstates;
1834 148 : _bt_parallel_scan_and_sort(btspool, btspool2, btshared, sharedsort,
1835 : sharedsort2, sortmem, false);
1836 :
1837 : /* Report WAL/buffer usage during parallel execution */
1838 148 : bufferusage = shm_toc_lookup(toc, PARALLEL_KEY_BUFFER_USAGE, false);
1839 148 : walusage = shm_toc_lookup(toc, PARALLEL_KEY_WAL_USAGE, false);
1840 148 : InstrEndParallelQuery(&bufferusage[ParallelWorkerNumber],
1841 148 : &walusage[ParallelWorkerNumber]);
1842 :
1843 : #ifdef BTREE_BUILD_STATS
1844 : if (log_btree_build_stats)
1845 : {
1846 : ShowUsage("BTREE BUILD (Worker Partial Spool) STATISTICS");
1847 : ResetUsage();
1848 : }
1849 : #endif /* BTREE_BUILD_STATS */
1850 :
1851 148 : index_close(indexRel, indexLockmode);
1852 148 : table_close(heapRel, heapLockmode);
1853 148 : }
1854 :
1855 : /*
1856 : * Perform a worker's portion of a parallel sort.
1857 : *
1858 : * This generates a tuplesort for passed btspool, and a second tuplesort
1859 : * state if a second btspool is need (i.e. for unique index builds). All
1860 : * other spool fields should already be set when this is called.
1861 : *
1862 : * sortmem is the amount of working memory to use within each worker,
1863 : * expressed in KBs.
1864 : *
1865 : * When this returns, workers are done, and need only release resources.
1866 : */
1867 : static void
1868 296 : _bt_parallel_scan_and_sort(BTSpool *btspool, BTSpool *btspool2,
1869 : BTShared *btshared, Sharedsort *sharedsort,
1870 : Sharedsort *sharedsort2, int sortmem, bool progress)
1871 : {
1872 : SortCoordinate coordinate;
1873 : BTBuildState buildstate;
1874 : TableScanDesc scan;
1875 : double reltuples;
1876 : IndexInfo *indexInfo;
1877 :
1878 : /* Initialize local tuplesort coordination state */
1879 296 : coordinate = palloc0(sizeof(SortCoordinateData));
1880 296 : coordinate->isWorker = true;
1881 296 : coordinate->nParticipants = -1;
1882 296 : coordinate->sharedsort = sharedsort;
1883 :
1884 : /* Begin "partial" tuplesort */
1885 592 : btspool->sortstate = tuplesort_begin_index_btree(btspool->heap,
1886 : btspool->index,
1887 296 : btspool->isunique,
1888 296 : btspool->nulls_not_distinct,
1889 : sortmem, coordinate,
1890 : TUPLESORT_NONE);
1891 :
1892 : /*
1893 : * Just as with serial case, there may be a second spool. If so, a
1894 : * second, dedicated spool2 partial tuplesort is required.
1895 : */
1896 296 : if (btspool2)
1897 : {
1898 : SortCoordinate coordinate2;
1899 :
1900 : /*
1901 : * We expect that the second one (for dead tuples) won't get very
1902 : * full, so we give it only work_mem (unless sortmem is less for
1903 : * worker). Worker processes are generally permitted to allocate
1904 : * work_mem independently.
1905 : */
1906 128 : coordinate2 = palloc0(sizeof(SortCoordinateData));
1907 128 : coordinate2->isWorker = true;
1908 128 : coordinate2->nParticipants = -1;
1909 128 : coordinate2->sharedsort = sharedsort2;
1910 128 : btspool2->sortstate =
1911 128 : tuplesort_begin_index_btree(btspool->heap, btspool->index, false, false,
1912 : Min(sortmem, work_mem), coordinate2,
1913 : false);
1914 : }
1915 :
1916 : /* Fill in buildstate for _bt_build_callback() */
1917 296 : buildstate.isunique = btshared->isunique;
1918 296 : buildstate.nulls_not_distinct = btshared->nulls_not_distinct;
1919 296 : buildstate.havedead = false;
1920 296 : buildstate.heap = btspool->heap;
1921 296 : buildstate.spool = btspool;
1922 296 : buildstate.spool2 = btspool2;
1923 296 : buildstate.indtuples = 0;
1924 296 : buildstate.btleader = NULL;
1925 :
1926 : /* Join parallel scan */
1927 296 : indexInfo = BuildIndexInfo(btspool->index);
1928 296 : indexInfo->ii_Concurrent = btshared->isconcurrent;
1929 296 : scan = table_beginscan_parallel(btspool->heap,
1930 : ParallelTableScanFromBTShared(btshared));
1931 296 : reltuples = table_index_build_scan(btspool->heap, btspool->index, indexInfo,
1932 : true, progress, _bt_build_callback,
1933 : &buildstate, scan);
1934 :
1935 : /* Execute this worker's part of the sort */
1936 296 : if (progress)
1937 148 : pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE,
1938 : PROGRESS_BTREE_PHASE_PERFORMSORT_1);
1939 296 : tuplesort_performsort(btspool->sortstate);
1940 296 : if (btspool2)
1941 : {
1942 128 : if (progress)
1943 64 : pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE,
1944 : PROGRESS_BTREE_PHASE_PERFORMSORT_2);
1945 128 : tuplesort_performsort(btspool2->sortstate);
1946 : }
1947 :
1948 : /*
1949 : * Done. Record ambuild statistics, and whether we encountered a broken
1950 : * HOT chain.
1951 : */
1952 296 : SpinLockAcquire(&btshared->mutex);
1953 296 : btshared->nparticipantsdone++;
1954 296 : btshared->reltuples += reltuples;
1955 296 : if (buildstate.havedead)
1956 0 : btshared->havedead = true;
1957 296 : btshared->indtuples += buildstate.indtuples;
1958 296 : if (indexInfo->ii_BrokenHotChain)
1959 0 : btshared->brokenhotchain = true;
1960 296 : SpinLockRelease(&btshared->mutex);
1961 :
1962 : /* Notify leader */
1963 296 : ConditionVariableSignal(&btshared->workersdonecv);
1964 :
1965 : /* We can end tuplesorts immediately */
1966 296 : tuplesort_end(btspool->sortstate);
1967 296 : if (btspool2)
1968 128 : tuplesort_end(btspool2->sortstate);
1969 296 : }
|
