PostgreSQL Source Code  git master
nbtsort.c
Go to the documentation of this file.
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  * Formerly the index pages being built were kept in shared buffers, but
27  * that is of no value (since other backends have no interest in them yet)
28  * and it created locking problems for CHECKPOINT, because the upper-level
29  * pages were held exclusive-locked for long periods. Now we just build
30  * the pages in local memory and smgrwrite or smgrextend them as we finish
31  * them. They will need to be re-read into shared buffers on first use after
32  * the build finishes.
33  *
34  * This code isn't concerned about the FSM at all. The caller is responsible
35  * for initializing that.
36  *
37  * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
38  * Portions Copyright (c) 1994, Regents of the University of California
39  *
40  * IDENTIFICATION
41  * src/backend/access/nbtree/nbtsort.c
42  *
43  *-------------------------------------------------------------------------
44  */
45 
46 #include "postgres.h"
47 
48 #include "access/nbtree.h"
49 #include "access/parallel.h"
50 #include "access/relscan.h"
51 #include "access/table.h"
52 #include "access/tableam.h"
53 #include "access/xact.h"
54 #include "access/xlog.h"
55 #include "access/xloginsert.h"
56 #include "catalog/index.h"
57 #include "commands/progress.h"
58 #include "executor/instrument.h"
59 #include "miscadmin.h"
60 #include "pgstat.h"
61 #include "storage/smgr.h"
62 #include "tcop/tcopprot.h" /* pgrminclude ignore */
63 #include "utils/rel.h"
64 #include "utils/sortsupport.h"
65 #include "utils/tuplesort.h"
66 
67 
68 /* Magic numbers for parallel state sharing */
69 #define PARALLEL_KEY_BTREE_SHARED UINT64CONST(0xA000000000000001)
70 #define PARALLEL_KEY_TUPLESORT UINT64CONST(0xA000000000000002)
71 #define PARALLEL_KEY_TUPLESORT_SPOOL2 UINT64CONST(0xA000000000000003)
72 #define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xA000000000000004)
73 #define PARALLEL_KEY_WAL_USAGE UINT64CONST(0xA000000000000005)
74 #define PARALLEL_KEY_BUFFER_USAGE UINT64CONST(0xA000000000000006)
75 
76 /*
77  * DISABLE_LEADER_PARTICIPATION disables the leader's participation in
78  * parallel index builds. This may be useful as a debugging aid.
79 #undef DISABLE_LEADER_PARTICIPATION
80  */
81 
82 /*
83  * Status record for spooling/sorting phase. (Note we may have two of
84  * these due to the special requirements for uniqueness-checking with
85  * dead tuples.)
86  */
87 typedef struct BTSpool
88 {
89  Tuplesortstate *sortstate; /* state data for tuplesort.c */
92  bool isunique;
93 } BTSpool;
94 
95 /*
96  * Status for index builds performed in parallel. This is allocated in a
97  * dynamic shared memory segment. Note that there is a separate tuplesort TOC
98  * entry, private to tuplesort.c but allocated by this module on its behalf.
99  */
100 typedef struct BTShared
101 {
102  /*
103  * These fields are not modified during the sort. They primarily exist
104  * for the benefit of worker processes that need to create BTSpool state
105  * corresponding to that used by the leader.
106  */
109  bool isunique;
112 
113  /*
114  * workersdonecv is used to monitor the progress of workers. All parallel
115  * participants must indicate that they are done before leader can use
116  * mutable state that workers maintain during scan (and before leader can
117  * proceed to tuplesort_performsort()).
118  */
120 
121  /*
122  * mutex protects all fields before heapdesc.
123  *
124  * These fields contain status information of interest to B-Tree index
125  * builds that must work just the same when an index is built in parallel.
126  */
128 
129  /*
130  * Mutable state that is maintained by workers, and reported back to
131  * leader at end of parallel scan.
132  *
133  * nparticipantsdone is number of worker processes finished.
134  *
135  * reltuples is the total number of input heap tuples.
136  *
137  * havedead indicates if RECENTLY_DEAD tuples were encountered during
138  * build.
139  *
140  * indtuples is the total number of tuples that made it into the index.
141  *
142  * brokenhotchain indicates if any worker detected a broken HOT chain
143  * during build.
144  */
146  double reltuples;
147  bool havedead;
148  double indtuples;
150 
151  /*
152  * ParallelTableScanDescData data follows. Can't directly embed here, as
153  * implementations of the parallel table scan desc interface might need
154  * stronger alignment.
155  */
156 } BTShared;
157 
158 /*
159  * Return pointer to a BTShared's parallel table scan.
160  *
161  * c.f. shm_toc_allocate as to why BUFFERALIGN is used, rather than just
162  * MAXALIGN.
163  */
164 #define ParallelTableScanFromBTShared(shared) \
165  (ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(BTShared)))
166 
167 /*
168  * Status for leader in parallel index build.
169  */
170 typedef struct BTLeader
171 {
172  /* parallel context itself */
174 
175  /*
176  * nparticipanttuplesorts is the exact number of worker processes
177  * successfully launched, plus one leader process if it participates as a
178  * worker (only DISABLE_LEADER_PARTICIPATION builds avoid leader
179  * participating as a worker).
180  */
182 
183  /*
184  * Leader process convenience pointers to shared state (leader avoids TOC
185  * lookups).
186  *
187  * btshared is the shared state for entire build. sharedsort is the
188  * shared, tuplesort-managed state passed to each process tuplesort.
189  * sharedsort2 is the corresponding btspool2 shared state, used only when
190  * building unique indexes. snapshot is the snapshot used by the scan iff
191  * an MVCC snapshot is required.
192  */
199 } BTLeader;
200 
201 /*
202  * Working state for btbuild and its callback.
203  *
204  * When parallel CREATE INDEX is used, there is a BTBuildState for each
205  * participant.
206  */
207 typedef struct BTBuildState
208 {
209  bool isunique;
210  bool havedead;
213 
214  /*
215  * spool2 is needed only when the index is a unique index. Dead tuples are
216  * put into spool2 instead of spool in order to avoid uniqueness check.
217  */
219  double indtuples;
220 
221  /*
222  * btleader is only present when a parallel index build is performed, and
223  * only in the leader process. (Actually, only the leader has a
224  * BTBuildState. Workers have their own spool and spool2, though.)
225  */
227 } BTBuildState;
228 
229 /*
230  * Status record for a btree page being built. We have one of these
231  * for each active tree level.
232  */
233 typedef struct BTPageState
234 {
235  Page btps_page; /* workspace for page building */
236  BlockNumber btps_blkno; /* block # to write this page at */
237  IndexTuple btps_lowkey; /* page's strict lower bound pivot tuple */
238  OffsetNumber btps_lastoff; /* last item offset loaded */
239  Size btps_lastextra; /* last item's extra posting list space */
240  uint32 btps_level; /* tree level (0 = leaf) */
241  Size btps_full; /* "full" if less than this much free space */
242  struct BTPageState *btps_next; /* link to parent level, if any */
243 } BTPageState;
244 
245 /*
246  * Overall status record for index writing phase.
247  */
248 typedef struct BTWriteState
249 {
252  BTScanInsert inskey; /* generic insertion scankey */
253  bool btws_use_wal; /* dump pages to WAL? */
254  BlockNumber btws_pages_alloced; /* # pages allocated */
255  BlockNumber btws_pages_written; /* # pages written out */
256  Page btws_zeropage; /* workspace for filling zeroes */
257 } BTWriteState;
258 
259 
261  BTBuildState *buildstate, IndexInfo *indexInfo);
262 static void _bt_spooldestroy(BTSpool *btspool);
263 static void _bt_spool(BTSpool *btspool, ItemPointer self,
264  Datum *values, bool *isnull);
265 static void _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2);
267  bool *isnull, bool tupleIsAlive, void *state);
268 static Page _bt_blnewpage(uint32 level);
269 static BTPageState *_bt_pagestate(BTWriteState *wstate, uint32 level);
270 static void _bt_slideleft(Page page);
271 static void _bt_sortaddtup(Page page, Size itemsize,
272  IndexTuple itup, OffsetNumber itup_off,
273  bool newfirstdataitem);
274 static void _bt_buildadd(BTWriteState *wstate, BTPageState *state,
275  IndexTuple itup, Size truncextra);
276 static void _bt_sort_dedup_finish_pending(BTWriteState *wstate,
278  BTDedupState dstate);
279 static void _bt_uppershutdown(BTWriteState *wstate, BTPageState *state);
280 static void _bt_load(BTWriteState *wstate,
281  BTSpool *btspool, BTSpool *btspool2);
282 static void _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent,
283  int request);
284 static void _bt_end_parallel(BTLeader *btleader);
286 static double _bt_parallel_heapscan(BTBuildState *buildstate,
287  bool *brokenhotchain);
288 static void _bt_leader_participate_as_worker(BTBuildState *buildstate);
289 static void _bt_parallel_scan_and_sort(BTSpool *btspool, BTSpool *btspool2,
290  BTShared *btshared, Sharedsort *sharedsort,
291  Sharedsort *sharedsort2, int sortmem,
292  bool progress);
293 
294 
295 /*
296  * btbuild() -- build a new btree index.
297  */
300 {
301  IndexBuildResult *result;
302  BTBuildState buildstate;
303  double reltuples;
304 
305 #ifdef BTREE_BUILD_STATS
307  ResetUsage();
308 #endif /* BTREE_BUILD_STATS */
309 
310  buildstate.isunique = indexInfo->ii_Unique;
311  buildstate.havedead = false;
312  buildstate.heap = heap;
313  buildstate.spool = NULL;
314  buildstate.spool2 = NULL;
315  buildstate.indtuples = 0;
316  buildstate.btleader = NULL;
317 
318  /*
319  * We expect to be called exactly once for any index relation. If that's
320  * not the case, big trouble's what we have.
321  */
322  if (RelationGetNumberOfBlocks(index) != 0)
323  elog(ERROR, "index \"%s\" already contains data",
324  RelationGetRelationName(index));
325 
326  reltuples = _bt_spools_heapscan(heap, index, &buildstate, indexInfo);
327 
328  /*
329  * Finish the build by (1) completing the sort of the spool file, (2)
330  * inserting the sorted tuples into btree pages and (3) building the upper
331  * levels. Finally, it may also be necessary to end use of parallelism.
332  */
333  _bt_leafbuild(buildstate.spool, buildstate.spool2);
334  _bt_spooldestroy(buildstate.spool);
335  if (buildstate.spool2)
336  _bt_spooldestroy(buildstate.spool2);
337  if (buildstate.btleader)
338  _bt_end_parallel(buildstate.btleader);
339 
340  result = (IndexBuildResult *) palloc(sizeof(IndexBuildResult));
341 
342  result->heap_tuples = reltuples;
343  result->index_tuples = buildstate.indtuples;
344 
345 #ifdef BTREE_BUILD_STATS
347  {
348  ShowUsage("BTREE BUILD STATS");
349  ResetUsage();
350  }
351 #endif /* BTREE_BUILD_STATS */
352 
353  return result;
354 }
355 
356 /*
357  * Create and initialize one or two spool structures, and save them in caller's
358  * buildstate argument. May also fill-in fields within indexInfo used by index
359  * builds.
360  *
361  * Scans the heap, possibly in parallel, filling spools with IndexTuples. This
362  * routine encapsulates all aspects of managing parallelism. Caller need only
363  * call _bt_end_parallel() in parallel case after it is done with spool/spool2.
364  *
365  * Returns the total number of heap tuples scanned.
366  */
367 static double
369  IndexInfo *indexInfo)
370 {
371  BTSpool *btspool = (BTSpool *) palloc0(sizeof(BTSpool));
372  SortCoordinate coordinate = NULL;
373  double reltuples = 0;
374 
375  /*
376  * We size the sort area as maintenance_work_mem rather than work_mem to
377  * speed index creation. This should be OK since a single backend can't
378  * run multiple index creations in parallel (see also: notes on
379  * parallelism and maintenance_work_mem below).
380  */
381  btspool->heap = heap;
382  btspool->index = index;
383  btspool->isunique = indexInfo->ii_Unique;
384 
385  /* Save as primary spool */
386  buildstate->spool = btspool;
387 
388  /* Report table scan phase started */
391 
392  /* Attempt to launch parallel worker scan when required */
393  if (indexInfo->ii_ParallelWorkers > 0)
394  _bt_begin_parallel(buildstate, indexInfo->ii_Concurrent,
395  indexInfo->ii_ParallelWorkers);
396 
397  /*
398  * If parallel build requested and at least one worker process was
399  * successfully launched, set up coordination state
400  */
401  if (buildstate->btleader)
402  {
403  coordinate = (SortCoordinate) palloc0(sizeof(SortCoordinateData));
404  coordinate->isWorker = false;
405  coordinate->nParticipants =
406  buildstate->btleader->nparticipanttuplesorts;
407  coordinate->sharedsort = buildstate->btleader->sharedsort;
408  }
409 
410  /*
411  * Begin serial/leader tuplesort.
412  *
413  * In cases where parallelism is involved, the leader receives the same
414  * share of maintenance_work_mem as a serial sort (it is generally treated
415  * in the same way as a serial sort once we return). Parallel worker
416  * Tuplesortstates will have received only a fraction of
417  * maintenance_work_mem, though.
418  *
419  * We rely on the lifetime of the Leader Tuplesortstate almost not
420  * overlapping with any worker Tuplesortstate's lifetime. There may be
421  * some small overlap, but that's okay because we rely on leader
422  * Tuplesortstate only allocating a small, fixed amount of memory here.
423  * When its tuplesort_performsort() is called (by our caller), and
424  * significant amounts of memory are likely to be used, all workers must
425  * have already freed almost all memory held by their Tuplesortstates
426  * (they are about to go away completely, too). The overall effect is
427  * that maintenance_work_mem always represents an absolute high watermark
428  * on the amount of memory used by a CREATE INDEX operation, regardless of
429  * the use of parallelism or any other factor.
430  */
431  buildstate->spool->sortstate =
432  tuplesort_begin_index_btree(heap, index, buildstate->isunique,
433  maintenance_work_mem, coordinate,
434  false);
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  if (indexInfo->ii_Unique)
442  {
443  BTSpool *btspool2 = (BTSpool *) palloc0(sizeof(BTSpool));
444  SortCoordinate coordinate2 = NULL;
445 
446  /* Initialize secondary spool */
447  btspool2->heap = heap;
448  btspool2->index = index;
449  btspool2->isunique = false;
450  /* Save as secondary spool */
451  buildstate->spool2 = btspool2;
452 
453  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  coordinate2 = (SortCoordinate) palloc0(sizeof(SortCoordinateData));
461  coordinate2->isWorker = false;
462  coordinate2->nParticipants =
463  buildstate->btleader->nparticipanttuplesorts;
464  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  buildstate->spool2->sortstate =
472  tuplesort_begin_index_btree(heap, index, false, work_mem,
473  coordinate2, false);
474  }
475 
476  /* Fill spool using either serial or parallel heap scan */
477  if (!buildstate->btleader)
478  reltuples = table_index_build_scan(heap, index, indexInfo, true, true,
479  _bt_build_callback, (void *) buildstate,
480  NULL);
481  else
482  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  const int index[] = {
494  };
495  const int64 val[] = {
496  buildstate->indtuples,
497  0, 0
498  };
499 
500  pgstat_progress_update_multi_param(3, index, val);
501  }
502 
503  /* okay, all heap tuples are spooled */
504  if (buildstate->spool2 && !buildstate->havedead)
505  {
506  /* spool2 turns out to be unnecessary */
507  _bt_spooldestroy(buildstate->spool2);
508  buildstate->spool2 = NULL;
509  }
510 
511  return reltuples;
512 }
513 
514 /*
515  * clean up a spool structure and its substructures.
516  */
517 static void
519 {
520  tuplesort_end(btspool->sortstate);
521  pfree(btspool);
522 }
523 
524 /*
525  * spool an index entry into the sort file.
526  */
527 static void
528 _bt_spool(BTSpool *btspool, ItemPointer self, Datum *values, bool *isnull)
529 {
530  tuplesort_putindextuplevalues(btspool->sortstate, btspool->index,
531  self, values, isnull);
532 }
533 
534 /*
535  * given a spool loaded by successive calls to _bt_spool,
536  * create an entire btree.
537  */
538 static void
539 _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2)
540 {
541  BTWriteState wstate;
542 
543 #ifdef BTREE_BUILD_STATS
545  {
546  ShowUsage("BTREE BUILD (Spool) STATISTICS");
547  ResetUsage();
548  }
549 #endif /* BTREE_BUILD_STATS */
550 
554  if (btspool2)
555  {
558  tuplesort_performsort(btspool2->sortstate);
559  }
560 
561  wstate.heap = btspool->heap;
562  wstate.index = btspool->index;
563  wstate.inskey = _bt_mkscankey(wstate.index, NULL);
564  /* _bt_mkscankey() won't set allequalimage without metapage */
565  wstate.inskey->allequalimage = _bt_allequalimage(wstate.index, true);
566  wstate.btws_use_wal = RelationNeedsWAL(wstate.index);
567 
568  /* reserve the metapage */
569  wstate.btws_pages_alloced = BTREE_METAPAGE + 1;
570  wstate.btws_pages_written = 0;
571  wstate.btws_zeropage = NULL; /* until needed */
572 
575  _bt_load(&wstate, btspool, btspool2);
576 }
577 
578 /*
579  * Per-tuple callback for table_index_build_scan
580  */
581 static void
583  ItemPointer tid,
584  Datum *values,
585  bool *isnull,
586  bool tupleIsAlive,
587  void *state)
588 {
589  BTBuildState *buildstate = (BTBuildState *) state;
590 
591  /*
592  * insert the index tuple into the appropriate spool file for subsequent
593  * processing
594  */
595  if (tupleIsAlive || buildstate->spool2 == NULL)
596  _bt_spool(buildstate->spool, tid, values, isnull);
597  else
598  {
599  /* dead tuples are put into spool2 */
600  buildstate->havedead = true;
601  _bt_spool(buildstate->spool2, tid, values, isnull);
602  }
603 
604  buildstate->indtuples += 1;
605 }
606 
607 /*
608  * allocate workspace for a new, clean btree page, not linked to any siblings.
609  */
610 static Page
612 {
613  Page page;
614  BTPageOpaque opaque;
615 
616  page = (Page) palloc(BLCKSZ);
617 
618  /* Zero the page and set up standard page header info */
619  _bt_pageinit(page, BLCKSZ);
620 
621  /* Initialize BT opaque state */
622  opaque = (BTPageOpaque) PageGetSpecialPointer(page);
623  opaque->btpo_prev = opaque->btpo_next = P_NONE;
624  opaque->btpo.level = level;
625  opaque->btpo_flags = (level > 0) ? 0 : BTP_LEAF;
626  opaque->btpo_cycleid = 0;
627 
628  /* Make the P_HIKEY line pointer appear allocated */
629  ((PageHeader) page)->pd_lower += sizeof(ItemIdData);
630 
631  return page;
632 }
633 
634 /*
635  * emit a completed btree page, and release the working storage.
636  */
637 static void
639 {
640  /* Ensure rd_smgr is open (could have been closed by relcache flush!) */
641  RelationOpenSmgr(wstate->index);
642 
643  /* XLOG stuff */
644  if (wstate->btws_use_wal)
645  {
646  /* We use the XLOG_FPI record type for this */
647  log_newpage(&wstate->index->rd_node, MAIN_FORKNUM, blkno, page, true);
648  }
649 
650  /*
651  * If we have to write pages nonsequentially, fill in the space with
652  * zeroes until we come back and overwrite. This is not logically
653  * necessary on standard Unix filesystems (unwritten space will read as
654  * zeroes anyway), but it should help to avoid fragmentation. The dummy
655  * pages aren't WAL-logged though.
656  */
657  while (blkno > wstate->btws_pages_written)
658  {
659  if (!wstate->btws_zeropage)
660  wstate->btws_zeropage = (Page) palloc0(BLCKSZ);
661  /* don't set checksum for all-zero page */
663  wstate->btws_pages_written++,
664  (char *) wstate->btws_zeropage,
665  true);
666  }
667 
668  PageSetChecksumInplace(page, blkno);
669 
670  /*
671  * Now write the page. There's no need for smgr to schedule an fsync for
672  * this write; we'll do it ourselves before ending the build.
673  */
674  if (blkno == wstate->btws_pages_written)
675  {
676  /* extending the file... */
677  smgrextend(wstate->index->rd_smgr, MAIN_FORKNUM, blkno,
678  (char *) page, true);
679  wstate->btws_pages_written++;
680  }
681  else
682  {
683  /* overwriting a block we zero-filled before */
684  smgrwrite(wstate->index->rd_smgr, MAIN_FORKNUM, blkno,
685  (char *) page, true);
686  }
687 
688  pfree(page);
689 }
690 
691 /*
692  * allocate and initialize a new BTPageState. the returned structure
693  * is suitable for immediate use by _bt_buildadd.
694  */
695 static BTPageState *
697 {
699 
700  /* create initial page for level */
701  state->btps_page = _bt_blnewpage(level);
702 
703  /* and assign it a page position */
704  state->btps_blkno = wstate->btws_pages_alloced++;
705 
706  state->btps_lowkey = NULL;
707  /* initialize lastoff so first item goes into P_FIRSTKEY */
708  state->btps_lastoff = P_HIKEY;
709  state->btps_lastextra = 0;
710  state->btps_level = level;
711  /* set "full" threshold based on level. See notes at head of file. */
712  if (level > 0)
713  state->btps_full = (BLCKSZ * (100 - BTREE_NONLEAF_FILLFACTOR) / 100);
714  else
715  state->btps_full = BTGetTargetPageFreeSpace(wstate->index);
716 
717  /* no parent level, yet */
718  state->btps_next = NULL;
719 
720  return state;
721 }
722 
723 /*
724  * slide an array of ItemIds back one slot (from P_FIRSTKEY to
725  * P_HIKEY, overwriting P_HIKEY). we need to do this when we discover
726  * that we have built an ItemId array in what has turned out to be a
727  * P_RIGHTMOST page.
728  */
729 static void
731 {
732  OffsetNumber off;
733  OffsetNumber maxoff;
734  ItemId previi;
735  ItemId thisii;
736 
737  if (!PageIsEmpty(page))
738  {
739  maxoff = PageGetMaxOffsetNumber(page);
740  previi = PageGetItemId(page, P_HIKEY);
741  for (off = P_FIRSTKEY; off <= maxoff; off = OffsetNumberNext(off))
742  {
743  thisii = PageGetItemId(page, off);
744  *previi = *thisii;
745  previi = thisii;
746  }
747  ((PageHeader) page)->pd_lower -= sizeof(ItemIdData);
748  }
749 }
750 
751 /*
752  * Add an item to a page being built.
753  *
754  * This is very similar to nbtinsert.c's _bt_pgaddtup(), but this variant
755  * raises an error directly.
756  *
757  * Note that our nbtsort.c caller does not know yet if the page will be
758  * rightmost. Offset P_FIRSTKEY is always assumed to be the first data key by
759  * caller. Page that turns out to be the rightmost on its level is fixed by
760  * calling _bt_slideleft().
761  */
762 static void
764  Size itemsize,
765  IndexTuple itup,
766  OffsetNumber itup_off,
767  bool newfirstdataitem)
768 {
769  IndexTupleData trunctuple;
770 
771  if (newfirstdataitem)
772  {
773  trunctuple = *itup;
774  trunctuple.t_info = sizeof(IndexTupleData);
775  BTreeTupleSetNAtts(&trunctuple, 0, false);
776  itup = &trunctuple;
777  itemsize = sizeof(IndexTupleData);
778  }
779 
780  if (PageAddItem(page, (Item) itup, itemsize, itup_off,
781  false, false) == InvalidOffsetNumber)
782  elog(ERROR, "failed to add item to the index page");
783 }
784 
785 /*----------
786  * Add an item to a disk page from the sort output (or add a posting list
787  * item formed from the sort output).
788  *
789  * We must be careful to observe the page layout conventions of nbtsearch.c:
790  * - rightmost pages start data items at P_HIKEY instead of at P_FIRSTKEY.
791  * - on non-leaf pages, the key portion of the first item need not be
792  * stored, we should store only the link.
793  *
794  * A leaf page being built looks like:
795  *
796  * +----------------+---------------------------------+
797  * | PageHeaderData | linp0 linp1 linp2 ... |
798  * +-----------+----+---------------------------------+
799  * | ... linpN | |
800  * +-----------+--------------------------------------+
801  * | ^ last |
802  * | |
803  * +-------------+------------------------------------+
804  * | | itemN ... |
805  * +-------------+------------------+-----------------+
806  * | ... item3 item2 item1 | "special space" |
807  * +--------------------------------+-----------------+
808  *
809  * Contrast this with the diagram in bufpage.h; note the mismatch
810  * between linps and items. This is because we reserve linp0 as a
811  * placeholder for the pointer to the "high key" item; when we have
812  * filled up the page, we will set linp0 to point to itemN and clear
813  * linpN. On the other hand, if we find this is the last (rightmost)
814  * page, we leave the items alone and slide the linp array over. If
815  * the high key is to be truncated, offset 1 is deleted, and we insert
816  * the truncated high key at offset 1.
817  *
818  * 'last' pointer indicates the last offset added to the page.
819  *
820  * 'truncextra' is the size of the posting list in itup, if any. This
821  * information is stashed for the next call here, when we may benefit
822  * from considering the impact of truncating away the posting list on
823  * the page before deciding to finish the page off. Posting lists are
824  * often relatively large, so it is worth going to the trouble of
825  * accounting for the saving from truncating away the posting list of
826  * the tuple that becomes the high key (that may be the only way to
827  * get close to target free space on the page). Note that this is
828  * only used for the soft fillfactor-wise limit, not the critical hard
829  * limit.
830  *----------
831  */
832 static void
834  Size truncextra)
835 {
836  Page npage;
837  BlockNumber nblkno;
838  OffsetNumber last_off;
839  Size last_truncextra;
840  Size pgspc;
841  Size itupsz;
842  bool isleaf;
843 
844  /*
845  * This is a handy place to check for cancel interrupts during the btree
846  * load phase of index creation.
847  */
849 
850  npage = state->btps_page;
851  nblkno = state->btps_blkno;
852  last_off = state->btps_lastoff;
853  last_truncextra = state->btps_lastextra;
854  state->btps_lastextra = truncextra;
855 
856  pgspc = PageGetFreeSpace(npage);
857  itupsz = IndexTupleSize(itup);
858  itupsz = MAXALIGN(itupsz);
859  /* Leaf case has slightly different rules due to suffix truncation */
860  isleaf = (state->btps_level == 0);
861 
862  /*
863  * Check whether the new item can fit on a btree page on current level at
864  * all.
865  *
866  * Every newly built index will treat heap TID as part of the keyspace,
867  * which imposes the requirement that new high keys must occasionally have
868  * a heap TID appended within _bt_truncate(). That may leave a new pivot
869  * tuple one or two MAXALIGN() quantums larger than the original
870  * firstright tuple it's derived from. v4 deals with the problem by
871  * decreasing the limit on the size of tuples inserted on the leaf level
872  * by the same small amount. Enforce the new v4+ limit on the leaf level,
873  * and the old limit on internal levels, since pivot tuples may need to
874  * make use of the reserved space. This should never fail on internal
875  * pages.
876  */
877  if (unlikely(itupsz > BTMaxItemSize(npage)))
878  _bt_check_third_page(wstate->index, wstate->heap, isleaf, npage,
879  itup);
880 
881  /*
882  * Check to see if current page will fit new item, with space left over to
883  * append a heap TID during suffix truncation when page is a leaf page.
884  *
885  * It is guaranteed that we can fit at least 2 non-pivot tuples plus a
886  * high key with heap TID when finishing off a leaf page, since we rely on
887  * _bt_check_third_page() rejecting oversized non-pivot tuples. On
888  * internal pages we can always fit 3 pivot tuples with larger internal
889  * page tuple limit (includes page high key).
890  *
891  * Most of the time, a page is only "full" in the sense that the soft
892  * fillfactor-wise limit has been exceeded. However, we must always leave
893  * at least two items plus a high key on each page before starting a new
894  * page. Disregard fillfactor and insert on "full" current page if we
895  * don't have the minimum number of items yet. (Note that we deliberately
896  * assume that suffix truncation neither enlarges nor shrinks new high key
897  * when applying soft limit, except when last tuple has a posting list.)
898  */
899  Assert(last_truncextra == 0 || isleaf);
900  if (pgspc < itupsz + (isleaf ? MAXALIGN(sizeof(ItemPointerData)) : 0) ||
901  (pgspc + last_truncextra < state->btps_full && last_off > P_FIRSTKEY))
902  {
903  /*
904  * Finish off the page and write it out.
905  */
906  Page opage = npage;
907  BlockNumber oblkno = nblkno;
908  ItemId ii;
909  ItemId hii;
910  IndexTuple oitup;
911 
912  /* Create new page of same level */
913  npage = _bt_blnewpage(state->btps_level);
914 
915  /* and assign it a page position */
916  nblkno = wstate->btws_pages_alloced++;
917 
918  /*
919  * We copy the last item on the page into the new page, and then
920  * rearrange the old page so that the 'last item' becomes its high key
921  * rather than a true data item. There had better be at least two
922  * items on the page already, else the page would be empty of useful
923  * data.
924  */
925  Assert(last_off > P_FIRSTKEY);
926  ii = PageGetItemId(opage, last_off);
927  oitup = (IndexTuple) PageGetItem(opage, ii);
928  _bt_sortaddtup(npage, ItemIdGetLength(ii), oitup, P_FIRSTKEY,
929  !isleaf);
930 
931  /*
932  * Move 'last' into the high key position on opage. _bt_blnewpage()
933  * allocated empty space for a line pointer when opage was first
934  * created, so this is a matter of rearranging already-allocated space
935  * on page, and initializing high key line pointer. (Actually, leaf
936  * pages must also swap oitup with a truncated version of oitup, which
937  * is sometimes larger than oitup, though never by more than the space
938  * needed to append a heap TID.)
939  */
940  hii = PageGetItemId(opage, P_HIKEY);
941  *hii = *ii;
942  ItemIdSetUnused(ii); /* redundant */
943  ((PageHeader) opage)->pd_lower -= sizeof(ItemIdData);
944 
945  if (isleaf)
946  {
947  IndexTuple lastleft;
948  IndexTuple truncated;
949 
950  /*
951  * Truncate away any unneeded attributes from high key on leaf
952  * level. This is only done at the leaf level because downlinks
953  * in internal pages are either negative infinity items, or get
954  * their contents from copying from one level down. See also:
955  * _bt_split().
956  *
957  * We don't try to bias our choice of split point to make it more
958  * likely that _bt_truncate() can truncate away more attributes,
959  * whereas the split point used within _bt_split() is chosen much
960  * more delicately. Even still, the lastleft and firstright
961  * tuples passed to _bt_truncate() here are at least not fully
962  * equal to each other when deduplication is used, unless there is
963  * a large group of duplicates (also, unique index builds usually
964  * have few or no spool2 duplicates). When the split point is
965  * between two unequal tuples, _bt_truncate() will avoid including
966  * a heap TID in the new high key, which is the most important
967  * benefit of suffix truncation.
968  *
969  * Overwrite the old item with new truncated high key directly.
970  * oitup is already located at the physical beginning of tuple
971  * space, so this should directly reuse the existing tuple space.
972  */
973  ii = PageGetItemId(opage, OffsetNumberPrev(last_off));
974  lastleft = (IndexTuple) PageGetItem(opage, ii);
975 
976  Assert(IndexTupleSize(oitup) > last_truncextra);
977  truncated = _bt_truncate(wstate->index, lastleft, oitup,
978  wstate->inskey);
979  if (!PageIndexTupleOverwrite(opage, P_HIKEY, (Item) truncated,
980  IndexTupleSize(truncated)))
981  elog(ERROR, "failed to add high key to the index page");
982  pfree(truncated);
983 
984  /* oitup should continue to point to the page's high key */
985  hii = PageGetItemId(opage, P_HIKEY);
986  oitup = (IndexTuple) PageGetItem(opage, hii);
987  }
988 
989  /*
990  * Link the old page into its parent, using its low key. If we don't
991  * have a parent, we have to create one; this adds a new btree level.
992  */
993  if (state->btps_next == NULL)
994  state->btps_next = _bt_pagestate(wstate, state->btps_level + 1);
995 
996  Assert((BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) <=
998  BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) > 0) ||
1000  Assert(BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) == 0 ||
1002  BTreeTupleSetDownLink(state->btps_lowkey, oblkno);
1003  _bt_buildadd(wstate, state->btps_next, state->btps_lowkey, 0);
1004  pfree(state->btps_lowkey);
1005 
1006  /*
1007  * Save a copy of the high key from the old page. It is also the low
1008  * key for the new page.
1009  */
1010  state->btps_lowkey = CopyIndexTuple(oitup);
1011 
1012  /*
1013  * Set the sibling links for both pages.
1014  */
1015  {
1016  BTPageOpaque oopaque = (BTPageOpaque) PageGetSpecialPointer(opage);
1017  BTPageOpaque nopaque = (BTPageOpaque) PageGetSpecialPointer(npage);
1018 
1019  oopaque->btpo_next = nblkno;
1020  nopaque->btpo_prev = oblkno;
1021  nopaque->btpo_next = P_NONE; /* redundant */
1022  }
1023 
1024  /*
1025  * Write out the old page. We never need to touch it again, so we can
1026  * free the opage workspace too.
1027  */
1028  _bt_blwritepage(wstate, opage, oblkno);
1029 
1030  /*
1031  * Reset last_off to point to new page
1032  */
1033  last_off = P_FIRSTKEY;
1034  }
1035 
1036  /*
1037  * By here, either original page is still the current page, or a new page
1038  * was created that became the current page. Either way, the current page
1039  * definitely has space for new item.
1040  *
1041  * If the new item is the first for its page, it must also be the first
1042  * item on its entire level. On later same-level pages, a low key for a
1043  * page will be copied from the prior page in the code above. Generate a
1044  * minus infinity low key here instead.
1045  */
1046  if (last_off == P_HIKEY)
1047  {
1048  Assert(state->btps_lowkey == NULL);
1049  state->btps_lowkey = palloc0(sizeof(IndexTupleData));
1050  state->btps_lowkey->t_info = sizeof(IndexTupleData);
1051  BTreeTupleSetNAtts(state->btps_lowkey, 0, false);
1052  }
1053 
1054  /*
1055  * Add the new item into the current page.
1056  */
1057  last_off = OffsetNumberNext(last_off);
1058  _bt_sortaddtup(npage, itupsz, itup, last_off,
1059  !isleaf && last_off == P_FIRSTKEY);
1060 
1061  state->btps_page = npage;
1062  state->btps_blkno = nblkno;
1063  state->btps_lastoff = last_off;
1064 }
1065 
1066 /*
1067  * Finalize pending posting list tuple, and add it to the index. Final tuple
1068  * is based on saved base tuple, and saved list of heap TIDs.
1069  *
1070  * This is almost like _bt_dedup_finish_pending(), but it adds a new tuple
1071  * using _bt_buildadd().
1072  */
1073 static void
1075  BTDedupState dstate)
1076 {
1077  Assert(dstate->nitems > 0);
1078 
1079  if (dstate->nitems == 1)
1080  _bt_buildadd(wstate, state, dstate->base, 0);
1081  else
1082  {
1083  IndexTuple postingtuple;
1084  Size truncextra;
1085 
1086  /* form a tuple with a posting list */
1087  postingtuple = _bt_form_posting(dstate->base,
1088  dstate->htids,
1089  dstate->nhtids);
1090  /* Calculate posting list overhead */
1091  truncextra = IndexTupleSize(postingtuple) -
1092  BTreeTupleGetPostingOffset(postingtuple);
1093 
1094  _bt_buildadd(wstate, state, postingtuple, truncextra);
1095  pfree(postingtuple);
1096  }
1097 
1098  dstate->nhtids = 0;
1099  dstate->nitems = 0;
1100  dstate->phystupsize = 0;
1101 }
1102 
1103 /*
1104  * Finish writing out the completed btree.
1105  */
1106 static void
1108 {
1109  BTPageState *s;
1110  BlockNumber rootblkno = P_NONE;
1111  uint32 rootlevel = 0;
1112  Page metapage;
1113 
1114  /*
1115  * Each iteration of this loop completes one more level of the tree.
1116  */
1117  for (s = state; s != NULL; s = s->btps_next)
1118  {
1119  BlockNumber blkno;
1120  BTPageOpaque opaque;
1121 
1122  blkno = s->btps_blkno;
1124 
1125  /*
1126  * We have to link the last page on this level to somewhere.
1127  *
1128  * If we're at the top, it's the root, so attach it to the metapage.
1129  * Otherwise, add an entry for it to its parent using its low key.
1130  * This may cause the last page of the parent level to split, but
1131  * that's not a problem -- we haven't gotten to it yet.
1132  */
1133  if (s->btps_next == NULL)
1134  {
1135  opaque->btpo_flags |= BTP_ROOT;
1136  rootblkno = blkno;
1137  rootlevel = s->btps_level;
1138  }
1139  else
1140  {
1141  Assert((BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) <=
1143  BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) > 0) ||
1144  P_LEFTMOST(opaque));
1145  Assert(BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) == 0 ||
1146  !P_LEFTMOST(opaque));
1148  _bt_buildadd(wstate, s->btps_next, s->btps_lowkey, 0);
1149  pfree(s->btps_lowkey);
1150  s->btps_lowkey = NULL;
1151  }
1152 
1153  /*
1154  * This is the rightmost page, so the ItemId array needs to be slid
1155  * back one slot. Then we can dump out the page.
1156  */
1158  _bt_blwritepage(wstate, s->btps_page, s->btps_blkno);
1159  s->btps_page = NULL; /* writepage freed the workspace */
1160  }
1161 
1162  /*
1163  * As the last step in the process, construct the metapage and make it
1164  * point to the new root (unless we had no data at all, in which case it's
1165  * set to point to "P_NONE"). This changes the index to the "valid" state
1166  * by filling in a valid magic number in the metapage.
1167  */
1168  metapage = (Page) palloc(BLCKSZ);
1169  _bt_initmetapage(metapage, rootblkno, rootlevel,
1170  wstate->inskey->allequalimage);
1171  _bt_blwritepage(wstate, metapage, BTREE_METAPAGE);
1172 }
1173 
1174 /*
1175  * Read tuples in correct sort order from tuplesort, and load them into
1176  * btree leaves.
1177  */
1178 static void
1179 _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
1180 {
1181  BTPageState *state = NULL;
1182  bool merge = (btspool2 != NULL);
1183  IndexTuple itup,
1184  itup2 = NULL;
1185  bool load1;
1186  TupleDesc tupdes = RelationGetDescr(wstate->index);
1187  int i,
1189  SortSupport sortKeys;
1190  int64 tuples_done = 0;
1191  bool deduplicate;
1192 
1193  deduplicate = wstate->inskey->allequalimage &&
1194  BTGetDeduplicateItems(wstate->index);
1195 
1196  if (merge)
1197  {
1198  /*
1199  * Another BTSpool for dead tuples exists. Now we have to merge
1200  * btspool and btspool2.
1201  */
1202 
1203  /* the preparation of merge */
1204  itup = tuplesort_getindextuple(btspool->sortstate, true);
1205  itup2 = tuplesort_getindextuple(btspool2->sortstate, true);
1206 
1207  /* Prepare SortSupport data for each column */
1208  sortKeys = (SortSupport) palloc0(keysz * sizeof(SortSupportData));
1209 
1210  for (i = 0; i < keysz; i++)
1211  {
1212  SortSupport sortKey = sortKeys + i;
1213  ScanKey scanKey = wstate->inskey->scankeys + i;
1214  int16 strategy;
1215 
1216  sortKey->ssup_cxt = CurrentMemoryContext;
1217  sortKey->ssup_collation = scanKey->sk_collation;
1218  sortKey->ssup_nulls_first =
1219  (scanKey->sk_flags & SK_BT_NULLS_FIRST) != 0;
1220  sortKey->ssup_attno = scanKey->sk_attno;
1221  /* Abbreviation is not supported here */
1222  sortKey->abbreviate = false;
1223 
1224  AssertState(sortKey->ssup_attno != 0);
1225 
1226  strategy = (scanKey->sk_flags & SK_BT_DESC) != 0 ?
1228 
1229  PrepareSortSupportFromIndexRel(wstate->index, strategy, sortKey);
1230  }
1231 
1232  for (;;)
1233  {
1234  load1 = true; /* load BTSpool next ? */
1235  if (itup2 == NULL)
1236  {
1237  if (itup == NULL)
1238  break;
1239  }
1240  else if (itup != NULL)
1241  {
1242  int32 compare = 0;
1243 
1244  for (i = 1; i <= keysz; i++)
1245  {
1246  SortSupport entry;
1247  Datum attrDatum1,
1248  attrDatum2;
1249  bool isNull1,
1250  isNull2;
1251 
1252  entry = sortKeys + i - 1;
1253  attrDatum1 = index_getattr(itup, i, tupdes, &isNull1);
1254  attrDatum2 = index_getattr(itup2, i, tupdes, &isNull2);
1255 
1256  compare = ApplySortComparator(attrDatum1, isNull1,
1257  attrDatum2, isNull2,
1258  entry);
1259  if (compare > 0)
1260  {
1261  load1 = false;
1262  break;
1263  }
1264  else if (compare < 0)
1265  break;
1266  }
1267 
1268  /*
1269  * If key values are equal, we sort on ItemPointer. This is
1270  * required for btree indexes, since heap TID is treated as an
1271  * implicit last key attribute in order to ensure that all
1272  * keys in the index are physically unique.
1273  */
1274  if (compare == 0)
1275  {
1276  compare = ItemPointerCompare(&itup->t_tid, &itup2->t_tid);
1277  Assert(compare != 0);
1278  if (compare > 0)
1279  load1 = false;
1280  }
1281  }
1282  else
1283  load1 = false;
1284 
1285  /* When we see first tuple, create first index page */
1286  if (state == NULL)
1287  state = _bt_pagestate(wstate, 0);
1288 
1289  if (load1)
1290  {
1291  _bt_buildadd(wstate, state, itup, 0);
1292  itup = tuplesort_getindextuple(btspool->sortstate, true);
1293  }
1294  else
1295  {
1296  _bt_buildadd(wstate, state, itup2, 0);
1297  itup2 = tuplesort_getindextuple(btspool2->sortstate, true);
1298  }
1299 
1300  /* Report progress */
1302  ++tuples_done);
1303  }
1304  pfree(sortKeys);
1305  }
1306  else if (deduplicate)
1307  {
1308  /* merge is unnecessary, deduplicate into posting lists */
1309  BTDedupState dstate;
1310 
1311  dstate = (BTDedupState) palloc(sizeof(BTDedupStateData));
1312  dstate->deduplicate = true; /* unused */
1313  dstate->maxpostingsize = 0; /* set later */
1314  /* Metadata about base tuple of current pending posting list */
1315  dstate->base = NULL;
1316  dstate->baseoff = InvalidOffsetNumber; /* unused */
1317  dstate->basetupsize = 0;
1318  /* Metadata about current pending posting list TIDs */
1319  dstate->htids = NULL;
1320  dstate->nhtids = 0;
1321  dstate->nitems = 0;
1322  dstate->phystupsize = 0; /* unused */
1323  dstate->nintervals = 0; /* unused */
1324 
1325  while ((itup = tuplesort_getindextuple(btspool->sortstate,
1326  true)) != NULL)
1327  {
1328  /* When we see first tuple, create first index page */
1329  if (state == NULL)
1330  {
1331  state = _bt_pagestate(wstate, 0);
1332 
1333  /*
1334  * Limit size of posting list tuples to 1/10 space we want to
1335  * leave behind on the page, plus space for final item's line
1336  * pointer. This is equal to the space that we'd like to
1337  * leave behind on each leaf page when fillfactor is 90,
1338  * allowing us to get close to fillfactor% space utilization
1339  * when there happen to be a great many duplicates. (This
1340  * makes higher leaf fillfactor settings ineffective when
1341  * building indexes that have many duplicates, but packing
1342  * leaf pages full with few very large tuples doesn't seem
1343  * like a useful goal.)
1344  */
1345  dstate->maxpostingsize = MAXALIGN_DOWN((BLCKSZ * 10 / 100)) -
1346  sizeof(ItemIdData);
1347  Assert(dstate->maxpostingsize <= BTMaxItemSize(state->btps_page) &&
1348  dstate->maxpostingsize <= INDEX_SIZE_MASK);
1349  dstate->htids = palloc(dstate->maxpostingsize);
1350 
1351  /* start new pending posting list with itup copy */
1354  }
1355  else if (_bt_keep_natts_fast(wstate->index, dstate->base,
1356  itup) > keysz &&
1357  _bt_dedup_save_htid(dstate, itup))
1358  {
1359  /*
1360  * Tuple is equal to base tuple of pending posting list. Heap
1361  * TID from itup has been saved in state.
1362  */
1363  }
1364  else
1365  {
1366  /*
1367  * Tuple is not equal to pending posting list tuple, or
1368  * _bt_dedup_save_htid() opted to not merge current item into
1369  * pending posting list.
1370  */
1371  _bt_sort_dedup_finish_pending(wstate, state, dstate);
1372  pfree(dstate->base);
1373 
1374  /* start new pending posting list with itup copy */
1377  }
1378 
1379  /* Report progress */
1381  ++tuples_done);
1382  }
1383 
1384  if (state)
1385  {
1386  /*
1387  * Handle the last item (there must be a last item when the
1388  * tuplesort returned one or more tuples)
1389  */
1390  _bt_sort_dedup_finish_pending(wstate, state, dstate);
1391  pfree(dstate->base);
1392  pfree(dstate->htids);
1393  }
1394 
1395  pfree(dstate);
1396  }
1397  else
1398  {
1399  /* merging and deduplication are both unnecessary */
1400  while ((itup = tuplesort_getindextuple(btspool->sortstate,
1401  true)) != NULL)
1402  {
1403  /* When we see first tuple, create first index page */
1404  if (state == NULL)
1405  state = _bt_pagestate(wstate, 0);
1406 
1407  _bt_buildadd(wstate, state, itup, 0);
1408 
1409  /* Report progress */
1411  ++tuples_done);
1412  }
1413  }
1414 
1415  /* Close down final pages and write the metapage */
1416  _bt_uppershutdown(wstate, state);
1417 
1418  /*
1419  * When we WAL-logged index pages, we must nonetheless fsync index files.
1420  * Since we're building outside shared buffers, a CHECKPOINT occurring
1421  * during the build has no way to flush the previously written data to
1422  * disk (indeed it won't know the index even exists). A crash later on
1423  * would replay WAL from the checkpoint, therefore it wouldn't replay our
1424  * earlier WAL entries. If we do not fsync those pages here, they might
1425  * still not be on disk when the crash occurs.
1426  */
1427  if (wstate->btws_use_wal)
1428  {
1429  RelationOpenSmgr(wstate->index);
1431  }
1432 }
1433 
1434 /*
1435  * Create parallel context, and launch workers for leader.
1436  *
1437  * buildstate argument should be initialized (with the exception of the
1438  * tuplesort state in spools, which may later be created based on shared
1439  * state initially set up here).
1440  *
1441  * isconcurrent indicates if operation is CREATE INDEX CONCURRENTLY.
1442  *
1443  * request is the target number of parallel worker processes to launch.
1444  *
1445  * Sets buildstate's BTLeader, which caller must use to shut down parallel
1446  * mode by passing it to _bt_end_parallel() at the very end of its index
1447  * build. If not even a single worker process can be launched, this is
1448  * never set, and caller should proceed with a serial index build.
1449  */
1450 static void
1451 _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent, int request)
1452 {
1453  ParallelContext *pcxt;
1454  int scantuplesortstates;
1455  Snapshot snapshot;
1456  Size estbtshared;
1457  Size estsort;
1458  BTShared *btshared;
1459  Sharedsort *sharedsort;
1460  Sharedsort *sharedsort2;
1461  BTSpool *btspool = buildstate->spool;
1462  BTLeader *btleader = (BTLeader *) palloc0(sizeof(BTLeader));
1463  WalUsage *walusage;
1464  BufferUsage *bufferusage;
1465  bool leaderparticipates = true;
1466  char *sharedquery;
1467  int querylen;
1468 
1469 #ifdef DISABLE_LEADER_PARTICIPATION
1470  leaderparticipates = false;
1471 #endif
1472 
1473  /*
1474  * Enter parallel mode, and create context for parallel build of btree
1475  * index
1476  */
1478  Assert(request > 0);
1479  pcxt = CreateParallelContext("postgres", "_bt_parallel_build_main",
1480  request);
1481 
1482  scantuplesortstates = leaderparticipates ? request + 1 : request;
1483 
1484  /*
1485  * Prepare for scan of the base relation. In a normal index build, we use
1486  * SnapshotAny because we must retrieve all tuples and do our own time
1487  * qual checks (because we have to index RECENTLY_DEAD tuples). In a
1488  * concurrent build, we take a regular MVCC snapshot and index whatever's
1489  * live according to that.
1490  */
1491  if (!isconcurrent)
1492  snapshot = SnapshotAny;
1493  else
1495 
1496  /*
1497  * Estimate size for our own PARALLEL_KEY_BTREE_SHARED workspace, and
1498  * PARALLEL_KEY_TUPLESORT tuplesort workspace
1499  */
1500  estbtshared = _bt_parallel_estimate_shared(btspool->heap, snapshot);
1501  shm_toc_estimate_chunk(&pcxt->estimator, estbtshared);
1502  estsort = tuplesort_estimate_shared(scantuplesortstates);
1503  shm_toc_estimate_chunk(&pcxt->estimator, estsort);
1504 
1505  /*
1506  * Unique case requires a second spool, and so we may have to account for
1507  * another shared workspace for that -- PARALLEL_KEY_TUPLESORT_SPOOL2
1508  */
1509  if (!btspool->isunique)
1510  shm_toc_estimate_keys(&pcxt->estimator, 2);
1511  else
1512  {
1513  shm_toc_estimate_chunk(&pcxt->estimator, estsort);
1514  shm_toc_estimate_keys(&pcxt->estimator, 3);
1515  }
1516 
1517  /*
1518  * Estimate space for WalUsage and BufferUsage -- PARALLEL_KEY_WAL_USAGE
1519  * and PARALLEL_KEY_BUFFER_USAGE.
1520  *
1521  * If there are no extensions loaded that care, we could skip this. We
1522  * have no way of knowing whether anyone's looking at pgWalUsage or
1523  * pgBufferUsage, so do it unconditionally.
1524  */
1526  mul_size(sizeof(WalUsage), pcxt->nworkers));
1527  shm_toc_estimate_keys(&pcxt->estimator, 1);
1529  mul_size(sizeof(BufferUsage), pcxt->nworkers));
1530  shm_toc_estimate_keys(&pcxt->estimator, 1);
1531 
1532  /* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
1533  querylen = strlen(debug_query_string);
1534  shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
1535  shm_toc_estimate_keys(&pcxt->estimator, 1);
1536 
1537  /* Everyone's had a chance to ask for space, so now create the DSM */
1538  InitializeParallelDSM(pcxt);
1539 
1540  /* If no DSM segment was available, back out (do serial build) */
1541  if (pcxt->seg == NULL)
1542  {
1543  if (IsMVCCSnapshot(snapshot))
1544  UnregisterSnapshot(snapshot);
1545  DestroyParallelContext(pcxt);
1546  ExitParallelMode();
1547  return;
1548  }
1549 
1550  /* Store shared build state, for which we reserved space */
1551  btshared = (BTShared *) shm_toc_allocate(pcxt->toc, estbtshared);
1552  /* Initialize immutable state */
1553  btshared->heaprelid = RelationGetRelid(btspool->heap);
1554  btshared->indexrelid = RelationGetRelid(btspool->index);
1555  btshared->isunique = btspool->isunique;
1556  btshared->isconcurrent = isconcurrent;
1557  btshared->scantuplesortstates = scantuplesortstates;
1559  SpinLockInit(&btshared->mutex);
1560  /* Initialize mutable state */
1561  btshared->nparticipantsdone = 0;
1562  btshared->reltuples = 0.0;
1563  btshared->havedead = false;
1564  btshared->indtuples = 0.0;
1565  btshared->brokenhotchain = false;
1568  snapshot);
1569 
1570  /*
1571  * Store shared tuplesort-private state, for which we reserved space.
1572  * Then, initialize opaque state using tuplesort routine.
1573  */
1574  sharedsort = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
1575  tuplesort_initialize_shared(sharedsort, scantuplesortstates,
1576  pcxt->seg);
1577 
1578  shm_toc_insert(pcxt->toc, PARALLEL_KEY_BTREE_SHARED, btshared);
1579  shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLESORT, sharedsort);
1580 
1581  /* Unique case requires a second spool, and associated shared state */
1582  if (!btspool->isunique)
1583  sharedsort2 = NULL;
1584  else
1585  {
1586  /*
1587  * Store additional shared tuplesort-private state, for which we
1588  * reserved space. Then, initialize opaque state using tuplesort
1589  * routine.
1590  */
1591  sharedsort2 = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
1592  tuplesort_initialize_shared(sharedsort2, scantuplesortstates,
1593  pcxt->seg);
1594 
1595  shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLESORT_SPOOL2, sharedsort2);
1596  }
1597 
1598  /* Store query string for workers */
1599  sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
1600  memcpy(sharedquery, debug_query_string, querylen + 1);
1601  shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, sharedquery);
1602 
1603  /*
1604  * Allocate space for each worker's WalUsage and BufferUsage; no need to
1605  * initialize.
1606  */
1607  walusage = shm_toc_allocate(pcxt->toc,
1608  mul_size(sizeof(WalUsage), pcxt->nworkers));
1609  shm_toc_insert(pcxt->toc, PARALLEL_KEY_WAL_USAGE, walusage);
1610  bufferusage = shm_toc_allocate(pcxt->toc,
1611  mul_size(sizeof(BufferUsage), pcxt->nworkers));
1612  shm_toc_insert(pcxt->toc, PARALLEL_KEY_BUFFER_USAGE, bufferusage);
1613 
1614  /* Launch workers, saving status for leader/caller */
1615  LaunchParallelWorkers(pcxt);
1616  btleader->pcxt = pcxt;
1617  btleader->nparticipanttuplesorts = pcxt->nworkers_launched;
1618  if (leaderparticipates)
1619  btleader->nparticipanttuplesorts++;
1620  btleader->btshared = btshared;
1621  btleader->sharedsort = sharedsort;
1622  btleader->sharedsort2 = sharedsort2;
1623  btleader->snapshot = snapshot;
1624  btleader->walusage = walusage;
1625  btleader->bufferusage = bufferusage;
1626 
1627  /* If no workers were successfully launched, back out (do serial build) */
1628  if (pcxt->nworkers_launched == 0)
1629  {
1630  _bt_end_parallel(btleader);
1631  return;
1632  }
1633 
1634  /* Save leader state now that it's clear build will be parallel */
1635  buildstate->btleader = btleader;
1636 
1637  /* Join heap scan ourselves */
1638  if (leaderparticipates)
1640 
1641  /*
1642  * Caller needs to wait for all launched workers when we return. Make
1643  * sure that the failure-to-start case will not hang forever.
1644  */
1646 }
1647 
1648 /*
1649  * Shut down workers, destroy parallel context, and end parallel mode.
1650  */
1651 static void
1653 {
1654  int i;
1655 
1656  /* Shutdown worker processes */
1658 
1659  /*
1660  * Next, accumulate WAL usage. (This must wait for the workers to finish,
1661  * or we might get incomplete data.)
1662  */
1663  for (i = 0; i < btleader->pcxt->nworkers_launched; i++)
1664  InstrAccumParallelQuery(&btleader->bufferusage[i], &btleader->walusage[i]);
1665 
1666  /* Free last reference to MVCC snapshot, if one was used */
1667  if (IsMVCCSnapshot(btleader->snapshot))
1668  UnregisterSnapshot(btleader->snapshot);
1669  DestroyParallelContext(btleader->pcxt);
1670  ExitParallelMode();
1671 }
1672 
1673 /*
1674  * Returns size of shared memory required to store state for a parallel
1675  * btree index build based on the snapshot its parallel scan will use.
1676  */
1677 static Size
1679 {
1680  /* c.f. shm_toc_allocate as to why BUFFERALIGN is used */
1681  return add_size(BUFFERALIGN(sizeof(BTShared)),
1682  table_parallelscan_estimate(heap, snapshot));
1683 }
1684 
1685 /*
1686  * Within leader, wait for end of heap scan.
1687  *
1688  * When called, parallel heap scan started by _bt_begin_parallel() will
1689  * already be underway within worker processes (when leader participates
1690  * as a worker, we should end up here just as workers are finishing).
1691  *
1692  * Fills in fields needed for ambuild statistics, and lets caller set
1693  * field indicating that some worker encountered a broken HOT chain.
1694  *
1695  * Returns the total number of heap tuples scanned.
1696  */
1697 static double
1698 _bt_parallel_heapscan(BTBuildState *buildstate, bool *brokenhotchain)
1699 {
1700  BTShared *btshared = buildstate->btleader->btshared;
1701  int nparticipanttuplesorts;
1702  double reltuples;
1703 
1704  nparticipanttuplesorts = buildstate->btleader->nparticipanttuplesorts;
1705  for (;;)
1706  {
1707  SpinLockAcquire(&btshared->mutex);
1708  if (btshared->nparticipantsdone == nparticipanttuplesorts)
1709  {
1710  buildstate->havedead = btshared->havedead;
1711  buildstate->indtuples = btshared->indtuples;
1712  *brokenhotchain = btshared->brokenhotchain;
1713  reltuples = btshared->reltuples;
1714  SpinLockRelease(&btshared->mutex);
1715  break;
1716  }
1717  SpinLockRelease(&btshared->mutex);
1718 
1721  }
1722 
1724 
1725  return reltuples;
1726 }
1727 
1728 /*
1729  * Within leader, participate as a parallel worker.
1730  */
1731 static void
1733 {
1734  BTLeader *btleader = buildstate->btleader;
1735  BTSpool *leaderworker;
1736  BTSpool *leaderworker2;
1737  int sortmem;
1738 
1739  /* Allocate memory and initialize private spool */
1740  leaderworker = (BTSpool *) palloc0(sizeof(BTSpool));
1741  leaderworker->heap = buildstate->spool->heap;
1742  leaderworker->index = buildstate->spool->index;
1743  leaderworker->isunique = buildstate->spool->isunique;
1744 
1745  /* Initialize second spool, if required */
1746  if (!btleader->btshared->isunique)
1747  leaderworker2 = NULL;
1748  else
1749  {
1750  /* Allocate memory for worker's own private secondary spool */
1751  leaderworker2 = (BTSpool *) palloc0(sizeof(BTSpool));
1752 
1753  /* Initialize worker's own secondary spool */
1754  leaderworker2->heap = leaderworker->heap;
1755  leaderworker2->index = leaderworker->index;
1756  leaderworker2->isunique = false;
1757  }
1758 
1759  /*
1760  * Might as well use reliable figure when doling out maintenance_work_mem
1761  * (when requested number of workers were not launched, this will be
1762  * somewhat higher than it is for other workers).
1763  */
1764  sortmem = maintenance_work_mem / btleader->nparticipanttuplesorts;
1765 
1766  /* Perform work common to all participants */
1767  _bt_parallel_scan_and_sort(leaderworker, leaderworker2, btleader->btshared,
1768  btleader->sharedsort, btleader->sharedsort2,
1769  sortmem, true);
1770 
1771 #ifdef BTREE_BUILD_STATS
1773  {
1774  ShowUsage("BTREE BUILD (Leader Partial Spool) STATISTICS");
1775  ResetUsage();
1776  }
1777 #endif /* BTREE_BUILD_STATS */
1778 }
1779 
1780 /*
1781  * Perform work within a launched parallel process.
1782  */
1783 void
1785 {
1786  char *sharedquery;
1787  BTSpool *btspool;
1788  BTSpool *btspool2;
1789  BTShared *btshared;
1790  Sharedsort *sharedsort;
1791  Sharedsort *sharedsort2;
1792  Relation heapRel;
1793  Relation indexRel;
1794  LOCKMODE heapLockmode;
1795  LOCKMODE indexLockmode;
1796  WalUsage *walusage;
1797  BufferUsage *bufferusage;
1798  int sortmem;
1799 
1800 #ifdef BTREE_BUILD_STATS
1802  ResetUsage();
1803 #endif /* BTREE_BUILD_STATS */
1804 
1805  /* Set debug_query_string for individual workers first */
1806  sharedquery = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, false);
1807  debug_query_string = sharedquery;
1808 
1809  /* Report the query string from leader */
1811 
1812  /* Look up nbtree shared state */
1813  btshared = shm_toc_lookup(toc, PARALLEL_KEY_BTREE_SHARED, false);
1814 
1815  /* Open relations using lock modes known to be obtained by index.c */
1816  if (!btshared->isconcurrent)
1817  {
1818  heapLockmode = ShareLock;
1819  indexLockmode = AccessExclusiveLock;
1820  }
1821  else
1822  {
1823  heapLockmode = ShareUpdateExclusiveLock;
1824  indexLockmode = RowExclusiveLock;
1825  }
1826 
1827  /* Open relations within worker */
1828  heapRel = table_open(btshared->heaprelid, heapLockmode);
1829  indexRel = index_open(btshared->indexrelid, indexLockmode);
1830 
1831  /* Initialize worker's own spool */
1832  btspool = (BTSpool *) palloc0(sizeof(BTSpool));
1833  btspool->heap = heapRel;
1834  btspool->index = indexRel;
1835  btspool->isunique = btshared->isunique;
1836 
1837  /* Look up shared state private to tuplesort.c */
1838  sharedsort = shm_toc_lookup(toc, PARALLEL_KEY_TUPLESORT, false);
1839  tuplesort_attach_shared(sharedsort, seg);
1840  if (!btshared->isunique)
1841  {
1842  btspool2 = NULL;
1843  sharedsort2 = NULL;
1844  }
1845  else
1846  {
1847  /* Allocate memory for worker's own private secondary spool */
1848  btspool2 = (BTSpool *) palloc0(sizeof(BTSpool));
1849 
1850  /* Initialize worker's own secondary spool */
1851  btspool2->heap = btspool->heap;
1852  btspool2->index = btspool->index;
1853  btspool2->isunique = false;
1854  /* Look up shared state private to tuplesort.c */
1855  sharedsort2 = shm_toc_lookup(toc, PARALLEL_KEY_TUPLESORT_SPOOL2, false);
1856  tuplesort_attach_shared(sharedsort2, seg);
1857  }
1858 
1859  /* Prepare to track buffer usage during parallel execution */
1861 
1862  /* Perform sorting of spool, and possibly a spool2 */
1863  sortmem = maintenance_work_mem / btshared->scantuplesortstates;
1864  _bt_parallel_scan_and_sort(btspool, btspool2, btshared, sharedsort,
1865  sharedsort2, sortmem, false);
1866 
1867  /* Report WAL/buffer usage during parallel execution */
1868  bufferusage = shm_toc_lookup(toc, PARALLEL_KEY_BUFFER_USAGE, false);
1869  walusage = shm_toc_lookup(toc, PARALLEL_KEY_WAL_USAGE, false);
1871  &walusage[ParallelWorkerNumber]);
1872 
1873 #ifdef BTREE_BUILD_STATS
1875  {
1876  ShowUsage("BTREE BUILD (Worker Partial Spool) STATISTICS");
1877  ResetUsage();
1878  }
1879 #endif /* BTREE_BUILD_STATS */
1880 
1881  index_close(indexRel, indexLockmode);
1882  table_close(heapRel, heapLockmode);
1883 }
1884 
1885 /*
1886  * Perform a worker's portion of a parallel sort.
1887  *
1888  * This generates a tuplesort for passed btspool, and a second tuplesort
1889  * state if a second btspool is need (i.e. for unique index builds). All
1890  * other spool fields should already be set when this is called.
1891  *
1892  * sortmem is the amount of working memory to use within each worker,
1893  * expressed in KBs.
1894  *
1895  * When this returns, workers are done, and need only release resources.
1896  */
1897 static void
1899  BTShared *btshared, Sharedsort *sharedsort,
1900  Sharedsort *sharedsort2, int sortmem, bool progress)
1901 {
1902  SortCoordinate coordinate;
1903  BTBuildState buildstate;
1904  TableScanDesc scan;
1905  double reltuples;
1906  IndexInfo *indexInfo;
1907 
1908  /* Initialize local tuplesort coordination state */
1909  coordinate = palloc0(sizeof(SortCoordinateData));
1910  coordinate->isWorker = true;
1911  coordinate->nParticipants = -1;
1912  coordinate->sharedsort = sharedsort;
1913 
1914  /* Begin "partial" tuplesort */
1915  btspool->sortstate = tuplesort_begin_index_btree(btspool->heap,
1916  btspool->index,
1917  btspool->isunique,
1918  sortmem, coordinate,
1919  false);
1920 
1921  /*
1922  * Just as with serial case, there may be a second spool. If so, a
1923  * second, dedicated spool2 partial tuplesort is required.
1924  */
1925  if (btspool2)
1926  {
1927  SortCoordinate coordinate2;
1928 
1929  /*
1930  * We expect that the second one (for dead tuples) won't get very
1931  * full, so we give it only work_mem (unless sortmem is less for
1932  * worker). Worker processes are generally permitted to allocate
1933  * work_mem independently.
1934  */
1935  coordinate2 = palloc0(sizeof(SortCoordinateData));
1936  coordinate2->isWorker = true;
1937  coordinate2->nParticipants = -1;
1938  coordinate2->sharedsort = sharedsort2;
1939  btspool2->sortstate =
1940  tuplesort_begin_index_btree(btspool->heap, btspool->index, false,
1941  Min(sortmem, work_mem), coordinate2,
1942  false);
1943  }
1944 
1945  /* Fill in buildstate for _bt_build_callback() */
1946  buildstate.isunique = btshared->isunique;
1947  buildstate.havedead = false;
1948  buildstate.heap = btspool->heap;
1949  buildstate.spool = btspool;
1950  buildstate.spool2 = btspool2;
1951  buildstate.indtuples = 0;
1952  buildstate.btleader = NULL;
1953 
1954  /* Join parallel scan */
1955  indexInfo = BuildIndexInfo(btspool->index);
1956  indexInfo->ii_Concurrent = btshared->isconcurrent;
1957  scan = table_beginscan_parallel(btspool->heap,
1958  ParallelTableScanFromBTShared(btshared));
1959  reltuples = table_index_build_scan(btspool->heap, btspool->index, indexInfo,
1960  true, progress, _bt_build_callback,
1961  (void *) &buildstate, scan);
1962 
1963  /*
1964  * Execute this worker's part of the sort.
1965  *
1966  * Unlike leader and serial cases, we cannot avoid calling
1967  * tuplesort_performsort() for spool2 if it ends up containing no dead
1968  * tuples (this is disallowed for workers by tuplesort).
1969  */
1970  tuplesort_performsort(btspool->sortstate);
1971  if (btspool2)
1972  tuplesort_performsort(btspool2->sortstate);
1973 
1974  /*
1975  * Done. Record ambuild statistics, and whether we encountered a broken
1976  * HOT chain.
1977  */
1978  SpinLockAcquire(&btshared->mutex);
1979  btshared->nparticipantsdone++;
1980  btshared->reltuples += reltuples;
1981  if (buildstate.havedead)
1982  btshared->havedead = true;
1983  btshared->indtuples += buildstate.indtuples;
1984  if (indexInfo->ii_BrokenHotChain)
1985  btshared->brokenhotchain = true;
1986  SpinLockRelease(&btshared->mutex);
1987 
1988  /* Notify leader */
1990 
1991  /* We can end tuplesorts immediately */
1992  tuplesort_end(btspool->sortstate);
1993  if (btspool2)
1994  tuplesort_end(btspool2->sortstate);
1995 }
struct SortSupportData * SortSupport
Definition: sortsupport.h:58
IndexTuple tuplesort_getindextuple(Tuplesortstate *state, bool forward)
Definition: tuplesort.c:2389
static void _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent, int request)
Definition: nbtsort.c:1451
int32 ItemPointerCompare(ItemPointer arg1, ItemPointer arg2)
Definition: itemptr.c:52
signed short int16
Definition: c.h:354
bool ssup_nulls_first
Definition: sortsupport.h:75
int slock_t
Definition: s_lock.h:934
IndexTuple base
Definition: nbtree.h:745
#define BTP_ROOT
Definition: nbtree.h:73
IndexTuple _bt_form_posting(IndexTuple base, ItemPointer htids, int nhtids)
Definition: nbtdedup.c:601
Sharedsort * sharedsort2
Definition: nbtsort.c:195
void tuplesort_performsort(Tuplesortstate *state)
Definition: tuplesort.c:1964
BlockNumber btpo_next
Definition: nbtree.h:59
#define PROGRESS_BTREE_PHASE_INDEXBUILD_TABLESCAN
Definition: nbtree.h:986
#define PageIsEmpty(page)
Definition: bufpage.h:222
void table_close(Relation relation, LOCKMODE lockmode)
Definition: table.c:133
int scantuplesortstates
Definition: nbtsort.c:111
#define BTGreaterStrategyNumber
Definition: stratnum.h:33
struct BTWriteState BTWriteState
#define PARALLEL_KEY_TUPLESORT
Definition: nbtsort.c:70
Relation heap
Definition: nbtsort.c:211
#define AssertState(condition)
Definition: c.h:741
Page btws_zeropage
Definition: nbtsort.c:256
OffsetNumber baseoff
Definition: nbtree.h:746
ParallelContext * CreateParallelContext(const char *library_name, const char *function_name, int nworkers)
Definition: parallel.c:164
BTShared * btshared
Definition: nbtsort.c:193
int nparticipantsdone
Definition: nbtsort.c:145
static void _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
Definition: nbtsort.c:1179
Snapshot RegisterSnapshot(Snapshot snapshot)
Definition: snapmgr.c:865
#define RelationGetDescr(relation)
Definition: rel.h:482
int LOCKMODE
Definition: lockdefs.h:26
BTScanInsert _bt_mkscankey(Relation rel, IndexTuple itup)
Definition: nbtutils.c:90
#define BTGetTargetPageFreeSpace(relation)
Definition: nbtree.h:973
Oid indexrelid
Definition: nbtsort.c:108
void pgstat_report_activity(BackendState state, const char *cmd_str)
Definition: pgstat.c:3136
#define BTGetDeduplicateItems(relation)
Definition: nbtree.h:975
#define PROGRESS_BTREE_PHASE_PERFORMSORT_1
Definition: nbtree.h:987
static void _bt_sort_dedup_finish_pending(BTWriteState *wstate, BTPageState *state, BTDedupState dstate)
Definition: nbtsort.c:1074
void ShowUsage(const char *title)
Definition: postgres.c:4618
dsm_segment * seg
Definition: parallel.h:43
BlockNumber btws_pages_alloced
Definition: nbtsort.c:254
struct SMgrRelationData * rd_smgr
Definition: rel.h:57
void pgstat_progress_update_param(int index, int64 val)
Definition: pgstat.c:3235
Sharedsort * sharedsort
Definition: nbtsort.c:194
shm_toc_estimator estimator
Definition: parallel.h:42
#define SpinLockInit(lock)
Definition: spin.h:60
#define BTP_LEAF
Definition: nbtree.h:72
ItemPointerData t_tid
Definition: itup.h:37
void _bt_parallel_build_main(dsm_segment *seg, shm_toc *toc)
Definition: nbtsort.c:1784
#define Min(x, y)
Definition: c.h:920
bool havedead
Definition: nbtsort.c:210
Snapshot snapshot
Definition: nbtsort.c:196
union BTPageOpaqueData::@45 btpo
Pointer Item
Definition: item.h:17
#define P_NONE
Definition: nbtree.h:206
Sharedsort * sharedsort
Definition: tuplesort.h:55
bool isunique
Definition: nbtsort.c:109
#define INDEX_SIZE_MASK
Definition: itup.h:65
bool allequalimage
Definition: nbtree.h:660
#define PageAddItem(page, item, size, offsetNumber, overwrite, is_heap)
Definition: bufpage.h:416
#define ParallelTableScanFromBTShared(shared)
Definition: nbtsort.c:164
bool isconcurrent
Definition: nbtsort.c:110
void _bt_initmetapage(Page page, BlockNumber rootbknum, uint32 level, bool allequalimage)
Definition: nbtpage.c:60
#define PROGRESS_CREATEIDX_TUPLES_TOTAL
Definition: progress.h:84
uint32 BlockNumber
Definition: block.h:31
void _bt_dedup_start_pending(BTDedupState state, IndexTuple base, OffsetNumber baseoff)
Definition: nbtdedup.c:326
static void _bt_end_parallel(BTLeader *btleader)
Definition: nbtsort.c:1652
Tuplesortstate * sortstate
Definition: nbtsort.c:89
IndexInfo * BuildIndexInfo(Relation index)
Definition: index.c:2311
static void _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2)
Definition: nbtsort.c:539
BlockNumber btps_blkno
Definition: nbtsort.c:236
void tuplesort_initialize_shared(Sharedsort *shared, int nWorkers, dsm_segment *seg)
Definition: tuplesort.c:4502
unsigned int Oid
Definition: postgres_ext.h:31
#define shm_toc_estimate_chunk(e, sz)
Definition: shm_toc.h:51
#define BTreeTupleGetNAtts(itup, rel)
Definition: nbtree.h:445
bool btws_use_wal
Definition: nbtsort.c:253
BTSpool * spool
Definition: nbtsort.c:212
void InstrEndParallelQuery(BufferUsage *bufusage, WalUsage *walusage)
Definition: instrument.c:189
static void _bt_blwritepage(BTWriteState *wstate, Page page, BlockNumber blkno)
Definition: nbtsort.c:638
static pairingheap_node * merge(pairingheap *heap, pairingheap_node *a, pairingheap_node *b)
Definition: pairingheap.c:79
Snapshot GetTransactionSnapshot(void)
Definition: snapmgr.c:306
#define PageGetMaxOffsetNumber(page)
Definition: bufpage.h:357
ParallelContext * pcxt
Definition: nbtsort.c:173
BTPageOpaqueData * BTPageOpaque
Definition: nbtree.h:69
Size PageGetFreeSpace(Page page)
Definition: bufpage.c:574
Relation heap
Definition: nbtsort.c:90
signed int int32
Definition: c.h:355
struct BTBuildState BTBuildState
uint16 OffsetNumber
Definition: off.h:24
Definition: type.h:89
static void BTreeTupleSetNAtts(IndexTuple itup, uint16 nkeyatts, bool heaptid)
Definition: nbtree.h:463
#define RelationOpenSmgr(relation)
Definition: rel.h:513
IndexTuple _bt_truncate(Relation rel, IndexTuple lastleft, IndexTuple firstright, BTScanInsert itup_key)
Definition: nbtutils.c:2199
void ResetUsage(void)
Definition: postgres.c:4611
void WaitForParallelWorkersToFinish(ParallelContext *pcxt)
Definition: parallel.c:750
#define SpinLockAcquire(lock)
Definition: spin.h:62
void ConditionVariableInit(ConditionVariable *cv)
ItemPointer htids
Definition: nbtree.h:750
void DestroyParallelContext(ParallelContext *pcxt)
Definition: parallel.c:904
Page btps_page
Definition: nbtsort.c:235
struct BTSpool BTSpool
#define ItemIdGetLength(itemId)
Definition: itemid.h:59
void pfree(void *pointer)
Definition: mcxt.c:1056
#define PARALLEL_KEY_BUFFER_USAGE
Definition: nbtsort.c:74
Size phystupsize
Definition: nbtree.h:753
void ConditionVariableCancelSleep(void)
#define BTREE_NONLEAF_FILLFACTOR
Definition: nbtree.h:195
static int compare(const void *arg1, const void *arg2)
Definition: geqo_pool.c:145
Size btps_full
Definition: nbtsort.c:241
#define ERROR
Definition: elog.h:43
static double _bt_parallel_heapscan(BTBuildState *buildstate, bool *brokenhotchain)
Definition: nbtsort.c:1698
IndexBuildResult * btbuild(Relation heap, Relation index, IndexInfo *indexInfo)
Definition: nbtsort.c:299
static double table_index_build_scan(Relation table_rel, Relation index_rel, struct IndexInfo *index_info, bool allow_sync, bool progress, IndexBuildCallback callback, void *callback_state, TableScanDesc scan)
Definition: tableam.h:1524
bool isunique
Definition: nbtsort.c:92
MemoryContext ssup_cxt
Definition: sortsupport.h:66
BTCycleId btpo_cycleid
Definition: nbtree.h:66
void ConditionVariableSignal(ConditionVariable *cv)
#define PROGRESS_BTREE_PHASE_PERFORMSORT_2
Definition: nbtree.h:988
void ExitParallelMode(void)
Definition: xact.c:977
Oid heaprelid
Definition: nbtsort.c:107
BufferUsage * bufferusage
Definition: nbtsort.c:198
bool PageIndexTupleOverwrite(Page page, OffsetNumber offnum, Item newtup, Size newsize)
Definition: bufpage.c:1060
BlockNumber btpo_prev
Definition: nbtree.h:58
IndexTuple CopyIndexTuple(IndexTuple source)
Definition: indextuple.c:510
OffsetNumber btps_lastoff
Definition: nbtsort.c:238
static void _bt_sortaddtup(Page page, Size itemsize, IndexTuple itup, OffsetNumber itup_off, bool newfirstdataitem)
Definition: nbtsort.c:763
bool _bt_dedup_save_htid(BTDedupState state, IndexTuple itup)
Definition: nbtdedup.c:377
void smgrwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, char *buffer, bool skipFsync)
Definition: smgr.c:513
BTSpool * spool2
Definition: nbtsort.c:218
#define RowExclusiveLock
Definition: lockdefs.h:38
IndexTupleData * IndexTuple
Definition: itup.h:53
struct BTShared BTShared
#define PARALLEL_KEY_QUERY_TEXT
Definition: nbtsort.c:72
#define P_FIRSTKEY
Definition: nbtree.h:243
int ParallelWorkerNumber
Definition: parallel.c:112
double indtuples
Definition: nbtsort.c:219
bool isunique
Definition: nbtsort.c:209
#define RelationGetRelationName(relation)
Definition: rel.h:490
#define P_LEFTMOST(opaque)
Definition: nbtree.h:212
#define PARALLEL_KEY_TUPLESORT_SPOOL2
Definition: nbtsort.c:71
bool ii_BrokenHotChain
Definition: execnodes.h:175
unsigned int uint32
Definition: c.h:367
struct ItemIdData ItemIdData
MemoryContext CurrentMemoryContext
Definition: mcxt.c:38
static void BTreeTupleSetDownLink(IndexTuple pivot, BlockNumber blkno)
Definition: nbtree.h:430
#define IndexRelationGetNumberOfKeyAttributes(relation)
Definition: rel.h:475
int nworkers_launched
Definition: parallel.h:38
void PrepareSortSupportFromIndexRel(Relation indexRel, int16 strategy, SortSupport ssup)
Definition: sortsupport.c:161
bool brokenhotchain
Definition: nbtsort.c:149
void LaunchParallelWorkers(ParallelContext *pcxt)
Definition: parallel.c:527
void InstrAccumParallelQuery(BufferUsage *bufusage, WalUsage *walusage)
Definition: instrument.c:199
#define BTREE_METAPAGE
Definition: nbtree.h:141
void UnregisterSnapshot(Snapshot snapshot)
Definition: snapmgr.c:907
Size btps_lastextra
Definition: nbtsort.c:239
Relation index
Definition: nbtsort.c:91
Relation index
Definition: nbtsort.c:251
TableScanDesc table_beginscan_parallel(Relation relation, ParallelTableScanDesc parallel_scan)
Definition: tableam.c:161
const char * debug_query_string
Definition: postgres.c:88
void InitializeParallelDSM(ParallelContext *pcxt)
Definition: parallel.c:202
#define PageGetItemId(page, offsetNumber)
Definition: bufpage.h:235
#define SK_BT_NULLS_FIRST
Definition: nbtree.h:956
int progress
Definition: pgbench.c:234
void InstrStartParallelQuery(void)
Definition: instrument.c:181
uint32 level
Definition: nbtree.h:62
#define SpinLockRelease(lock)
Definition: spin.h:64
Size mul_size(Size s1, Size s2)
Definition: shmem.c:515
struct BTLeader BTLeader
BlockNumber btws_pages_written
Definition: nbtsort.c:255
void * palloc0(Size size)
Definition: mcxt.c:980
struct SortCoordinateData * SortCoordinate
Definition: tuplesort.h:58
#define PROGRESS_SCAN_BLOCKS_DONE
Definition: progress.h:120
uint32 btps_level
Definition: nbtsort.c:240
uintptr_t Datum
Definition: postgres.h:367
struct IndexTupleData IndexTupleData
static BTPageState * _bt_pagestate(BTWriteState *wstate, uint32 level)
Definition: nbtsort.c:696
static void _bt_build_callback(Relation index, ItemPointer tid, Datum *values, bool *isnull, bool tupleIsAlive, void *state)
Definition: nbtsort.c:582
Size add_size(Size s1, Size s2)
Definition: shmem.c:498
static double _bt_spools_heapscan(Relation heap, Relation index, BTBuildState *buildstate, IndexInfo *indexInfo)
Definition: nbtsort.c:368
bool log_btree_build_stats
Definition: guc.c:529
BTLeader * btleader
Definition: nbtsort.c:226
int work_mem
Definition: globals.c:121
double reltuples
Definition: nbtsort.c:146
AttrNumber ssup_attno
Definition: sortsupport.h:81
#define RelationGetNumberOfBlocks(reln)
Definition: bufmgr.h:211
double indtuples
Definition: nbtsort.c:148
static void _bt_parallel_scan_and_sort(BTSpool *btspool, BTSpool *btspool2, BTShared *btshared, Sharedsort *sharedsort, Sharedsort *sharedsort2, int sortmem, bool progress)
Definition: nbtsort.c:1898
#define IsMVCCSnapshot(snapshot)
Definition: snapmgr.h:97
struct BTPageState * btps_next
Definition: nbtsort.c:242
BTScanInsert inskey
Definition: nbtsort.c:252
#define InvalidOffsetNumber
Definition: off.h:26
IndexTuple btps_lowkey
Definition: nbtsort.c:237
void tuplesort_putindextuplevalues(Tuplesortstate *state, Relation rel, ItemPointer self, Datum *values, bool *isnull)
Definition: tuplesort.c:1651
slock_t mutex
Definition: nbtsort.c:127
int maintenance_work_mem
Definition: globals.c:122
static void _bt_buildadd(BTWriteState *wstate, BTPageState *state, IndexTuple itup, Size truncextra)
Definition: nbtsort.c:833
Size table_parallelscan_estimate(Relation rel, Snapshot snapshot)
Definition: tableam.c:126
WalUsage * walusage
Definition: nbtsort.c:197
bool havedead
Definition: nbtsort.c:147
RelFileNode rd_node
Definition: rel.h:55
bool ii_Unique
Definition: execnodes.h:172
void tuplesort_attach_shared(Sharedsort *shared, dsm_segment *seg)
Definition: tuplesort.c:4525
static uint32 BTreeTupleGetPostingOffset(IndexTuple posting)
Definition: nbtree.h:397
#define ShareUpdateExclusiveLock
Definition: lockdefs.h:39
void ConditionVariableSleep(ConditionVariable *cv, uint32 wait_event_info)
PageHeaderData * PageHeader
Definition: bufpage.h:166
int sk_flags
Definition: skey.h:66
#define Assert(condition)
Definition: c.h:738
#define SK_BT_DESC
Definition: nbtree.h:955
Definition: regguts.h:298
int nparticipanttuplesorts
Definition: nbtsort.c:181
void pgstat_progress_update_multi_param(int nparam, const int *index, const int64 *val)
Definition: pgstat.c:3257
#define OffsetNumberNext(offsetNumber)
Definition: off.h:52
int ii_ParallelWorkers
Definition: execnodes.h:176
size_t Size
Definition: c.h:466
#define PageGetSpecialPointer(page)
Definition: bufpage.h:326
static Size _bt_parallel_estimate_shared(Relation heap, Snapshot snapshot)
Definition: nbtsort.c:1678
#define shm_toc_estimate_keys(e, cnt)
Definition: shm_toc.h:53
#define OffsetNumberPrev(offsetNumber)
Definition: off.h:54
void PageSetChecksumInplace(Page page, BlockNumber blkno)
Definition: bufpage.c:1194
#define MAXALIGN(LEN)
Definition: c.h:691
#define index_getattr(tup, attnum, tupleDesc, isnull)
Definition: itup.h:100
void EnterParallelMode(void)
Definition: xact.c:964
#define PROGRESS_SCAN_BLOCKS_TOTAL
Definition: progress.h:119
Size basetupsize
Definition: nbtree.h:747
ScanKeyData scankeys[INDEX_MAX_KEYS]
Definition: nbtree.h:666
void * shm_toc_allocate(shm_toc *toc, Size nbytes)
Definition: shm_toc.c:88
static void _bt_leader_participate_as_worker(BTBuildState *buildstate)
Definition: nbtsort.c:1732
Size tuplesort_estimate_shared(int nWorkers)
Definition: tuplesort.c:4481
#define RelationNeedsWAL(relation)
Definition: rel.h:562
void _bt_check_third_page(Relation rel, Relation heap, bool needheaptidspace, Page page, IndexTuple newtup)
Definition: nbtutils.c:2633
bool ii_Concurrent
Definition: execnodes.h:174
#define SnapshotAny
Definition: snapmgr.h:69
#define PROGRESS_BTREE_PHASE_LEAF_LOAD
Definition: nbtree.h:989
static void _bt_spooldestroy(BTSpool *btspool)
Definition: nbtsort.c:518
Size maxpostingsize
Definition: nbtree.h:742
void index_close(Relation relation, LOCKMODE lockmode)
Definition: indexam.c:158
void smgrextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, char *buffer, bool skipFsync)
Definition: smgr.c:462
#define BTMaxItemSize(page)
Definition: nbtree.h:157
#define P_HIKEY
Definition: nbtree.h:242
static Datum values[MAXATTR]
Definition: bootstrap.c:167
static void _bt_slideleft(Page page)
Definition: nbtsort.c:730
struct BTPageState BTPageState
#define AccessExclusiveLock
Definition: lockdefs.h:45
#define PARALLEL_KEY_WAL_USAGE
Definition: nbtsort.c:73
int _bt_keep_natts_fast(Relation rel, IndexTuple lastleft, IndexTuple firstright)
Definition: nbtutils.c:2418
void shm_toc_insert(shm_toc *toc, uint64 key, void *address)
Definition: shm_toc.c:171
void * palloc(Size size)
Definition: mcxt.c:949
#define PROGRESS_CREATEIDX_TUPLES_DONE
Definition: progress.h:85
bool deduplicate
Definition: nbtree.h:741
static Page _bt_blnewpage(uint32 level)
Definition: nbtsort.c:611
Oid sk_collation
Definition: skey.h:70
XLogRecPtr log_newpage(RelFileNode *rnode, ForkNumber forkNum, BlockNumber blkno, Page page, bool page_std)
Definition: xloginsert.c:977
#define elog(elevel,...)
Definition: elog.h:214
#define ShareLock
Definition: lockdefs.h:41
int i
static void _bt_spool(BTSpool *btspool, ItemPointer self, Datum *values, bool *isnull)
Definition: nbtsort.c:528
void _bt_pageinit(Page page, Size size)
Definition: nbtpage.c:959
#define BUFFERALIGN(LEN)
Definition: c.h:693
#define unlikely(x)
Definition: c.h:206
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:99
#define PROGRESS_CREATEIDX_SUBPHASE
Definition: progress.h:83
Tuplesortstate * tuplesort_begin_index_btree(Relation heapRel, Relation indexRel, bool enforceUnique, int workMem, SortCoordinate coordinate, bool randomAccess)
Definition: tuplesort.c:1047
static void _bt_uppershutdown(BTWriteState *wstate, BTPageState *state)
Definition: nbtsort.c:1107
ConditionVariable workersdonecv
Definition: nbtsort.c:119
unsigned short t_info
Definition: itup.h:49
BTDedupStateData * BTDedupState
Definition: nbtree.h:765
Relation heap
Definition: nbtsort.c:250
#define ItemIdSetUnused(itemId)
Definition: itemid.h:128
#define PARALLEL_KEY_BTREE_SHARED
Definition: nbtsort.c:69
void table_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan, Snapshot snapshot)
Definition: tableam.c:141
void tuplesort_end(Tuplesortstate *state)
Definition: tuplesort.c:1388
#define BTLessStrategyNumber
Definition: stratnum.h:29
Relation table_open(Oid relationId, LOCKMODE lockmode)
Definition: table.c:39
bool _bt_allequalimage(Relation rel, bool debugmessage)
Definition: nbtutils.c:2691
static int ApplySortComparator(Datum datum1, bool isNull1, Datum datum2, bool isNull2, SortSupport ssup)
Definition: sortsupport.h:200
uint16 btpo_flags
Definition: nbtree.h:65
void WaitForParallelWorkersToAttach(ParallelContext *pcxt)
Definition: parallel.c:647
void smgrimmedsync(SMgrRelation reln, ForkNumber forknum)
Definition: smgr.c:620
#define RelationGetRelid(relation)
Definition: rel.h:456
long val
Definition: informix.c:664
void * shm_toc_lookup(shm_toc *toc, uint64 key, bool noError)
Definition: shm_toc.c:232
Relation index_open(Oid relationId, LOCKMODE lockmode)
Definition: indexam.c:132
#define PageGetItem(page, itemId)
Definition: bufpage.h:340
AttrNumber sk_attno
Definition: skey.h:67
Pointer Page
Definition: bufpage.h:78
#define IndexTupleSize(itup)
Definition: itup.h:71
shm_toc * toc
Definition: parallel.h:45
double index_tuples
Definition: genam.h:33
double heap_tuples
Definition: genam.h:32
#define MAXALIGN_DOWN(LEN)
Definition: c.h:703