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) 1996-2021, 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/xact.h"
53 #include "access/xlog.h"
54 #include "access/xloginsert.h"
55 #include "catalog/index.h"
56 #include "commands/progress.h"
57 #include "executor/instrument.h"
58 #include "miscadmin.h"
59 #include "pgstat.h"
60 #include "storage/smgr.h"
61 #include "tcop/tcopprot.h" /* pgrminclude ignore */
62 #include "utils/rel.h"
63 #include "utils/sortsupport.h"
64 #include "utils/tuplesort.h"
65 
66 
67 /* Magic numbers for parallel state sharing */
68 #define PARALLEL_KEY_BTREE_SHARED UINT64CONST(0xA000000000000001)
69 #define PARALLEL_KEY_TUPLESORT UINT64CONST(0xA000000000000002)
70 #define PARALLEL_KEY_TUPLESORT_SPOOL2 UINT64CONST(0xA000000000000003)
71 #define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xA000000000000004)
72 #define PARALLEL_KEY_WAL_USAGE UINT64CONST(0xA000000000000005)
73 #define PARALLEL_KEY_BUFFER_USAGE UINT64CONST(0xA000000000000006)
74 
75 /*
76  * DISABLE_LEADER_PARTICIPATION disables the leader's participation in
77  * parallel index builds. This may be useful as a debugging aid.
78 #undef DISABLE_LEADER_PARTICIPATION
79  */
80 
81 /*
82  * Status record for spooling/sorting phase. (Note we may have two of
83  * these due to the special requirements for uniqueness-checking with
84  * dead tuples.)
85  */
86 typedef struct BTSpool
87 {
88  Tuplesortstate *sortstate; /* state data for tuplesort.c */
91  bool isunique;
92 } BTSpool;
93 
94 /*
95  * Status for index builds performed in parallel. This is allocated in a
96  * dynamic shared memory segment. Note that there is a separate tuplesort TOC
97  * entry, private to tuplesort.c but allocated by this module on its behalf.
98  */
99 typedef struct BTShared
100 {
101  /*
102  * These fields are not modified during the sort. They primarily exist
103  * for the benefit of worker processes that need to create BTSpool state
104  * corresponding to that used by the leader.
105  */
108  bool isunique;
111 
112  /*
113  * workersdonecv is used to monitor the progress of workers. All parallel
114  * participants must indicate that they are done before leader can use
115  * mutable state that workers maintain during scan (and before leader can
116  * proceed to tuplesort_performsort()).
117  */
119 
120  /*
121  * mutex protects all fields before heapdesc.
122  *
123  * These fields contain status information of interest to B-Tree index
124  * builds that must work just the same when an index is built in parallel.
125  */
127 
128  /*
129  * Mutable state that is maintained by workers, and reported back to
130  * leader at end of parallel scan.
131  *
132  * nparticipantsdone is number of worker processes finished.
133  *
134  * reltuples is the total number of input heap tuples.
135  *
136  * havedead indicates if RECENTLY_DEAD tuples were encountered during
137  * build.
138  *
139  * indtuples is the total number of tuples that made it into the index.
140  *
141  * brokenhotchain indicates if any worker detected a broken HOT chain
142  * during build.
143  */
145  double reltuples;
146  bool havedead;
147  double indtuples;
149 
150  /*
151  * ParallelTableScanDescData data follows. Can't directly embed here, as
152  * implementations of the parallel table scan desc interface might need
153  * stronger alignment.
154  */
155 } BTShared;
156 
157 /*
158  * Return pointer to a BTShared's parallel table scan.
159  *
160  * c.f. shm_toc_allocate as to why BUFFERALIGN is used, rather than just
161  * MAXALIGN.
162  */
163 #define ParallelTableScanFromBTShared(shared) \
164  (ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(BTShared)))
165 
166 /*
167  * Status for leader in parallel index build.
168  */
169 typedef struct BTLeader
170 {
171  /* parallel context itself */
173 
174  /*
175  * nparticipanttuplesorts is the exact number of worker processes
176  * successfully launched, plus one leader process if it participates as a
177  * worker (only DISABLE_LEADER_PARTICIPATION builds avoid leader
178  * participating as a worker).
179  */
181 
182  /*
183  * Leader process convenience pointers to shared state (leader avoids TOC
184  * lookups).
185  *
186  * btshared is the shared state for entire build. sharedsort is the
187  * shared, tuplesort-managed state passed to each process tuplesort.
188  * sharedsort2 is the corresponding btspool2 shared state, used only when
189  * building unique indexes. snapshot is the snapshot used by the scan iff
190  * an MVCC snapshot is required.
191  */
198 } BTLeader;
199 
200 /*
201  * Working state for btbuild and its callback.
202  *
203  * When parallel CREATE INDEX is used, there is a BTBuildState for each
204  * participant.
205  */
206 typedef struct BTBuildState
207 {
208  bool isunique;
209  bool havedead;
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  */
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  */
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  Page btps_page; /* 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 {
251  BTScanInsert inskey; /* generic insertion scankey */
252  bool btws_use_wal; /* dump pages to WAL? */
253  BlockNumber btws_pages_alloced; /* # pages allocated */
254  BlockNumber btws_pages_written; /* # pages written out */
255  Page btws_zeropage; /* workspace for filling zeroes */
256 } BTWriteState;
257 
258 
260  BTBuildState *buildstate, IndexInfo *indexInfo);
261 static void _bt_spooldestroy(BTSpool *btspool);
262 static void _bt_spool(BTSpool *btspool, ItemPointer self,
263  Datum *values, bool *isnull);
264 static void _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2);
266  bool *isnull, bool tupleIsAlive, void *state);
267 static Page _bt_blnewpage(uint32 level);
268 static BTPageState *_bt_pagestate(BTWriteState *wstate, uint32 level);
269 static void _bt_slideleft(Page rightmostpage);
270 static void _bt_sortaddtup(Page page, Size itemsize,
271  IndexTuple itup, OffsetNumber itup_off,
272  bool newfirstdataitem);
273 static void _bt_buildadd(BTWriteState *wstate, BTPageState *state,
274  IndexTuple itup, Size truncextra);
275 static void _bt_sort_dedup_finish_pending(BTWriteState *wstate,
277  BTDedupState dstate);
278 static void _bt_uppershutdown(BTWriteState *wstate, BTPageState *state);
279 static void _bt_load(BTWriteState *wstate,
280  BTSpool *btspool, BTSpool *btspool2);
281 static void _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent,
282  int request);
283 static void _bt_end_parallel(BTLeader *btleader);
285 static double _bt_parallel_heapscan(BTBuildState *buildstate,
286  bool *brokenhotchain);
287 static void _bt_leader_participate_as_worker(BTBuildState *buildstate);
288 static void _bt_parallel_scan_and_sort(BTSpool *btspool, BTSpool *btspool2,
289  BTShared *btshared, Sharedsort *sharedsort,
290  Sharedsort *sharedsort2, int sortmem,
291  bool progress);
292 
293 
294 /*
295  * btbuild() -- build a new btree index.
296  */
299 {
300  IndexBuildResult *result;
301  BTBuildState buildstate;
302  double reltuples;
303 
304 #ifdef BTREE_BUILD_STATS
306  ResetUsage();
307 #endif /* BTREE_BUILD_STATS */
308 
309  buildstate.isunique = indexInfo->ii_Unique;
310  buildstate.havedead = false;
311  buildstate.heap = heap;
312  buildstate.spool = NULL;
313  buildstate.spool2 = NULL;
314  buildstate.indtuples = 0;
315  buildstate.btleader = NULL;
316 
317  /*
318  * We expect to be called exactly once for any index relation. If that's
319  * not the case, big trouble's what we have.
320  */
321  if (RelationGetNumberOfBlocks(index) != 0)
322  elog(ERROR, "index \"%s\" already contains data",
323  RelationGetRelationName(index));
324 
325  reltuples = _bt_spools_heapscan(heap, index, &buildstate, indexInfo);
326 
327  /*
328  * Finish the build by (1) completing the sort of the spool file, (2)
329  * inserting the sorted tuples into btree pages and (3) building the upper
330  * levels. Finally, it may also be necessary to end use of parallelism.
331  */
332  _bt_leafbuild(buildstate.spool, buildstate.spool2);
333  _bt_spooldestroy(buildstate.spool);
334  if (buildstate.spool2)
335  _bt_spooldestroy(buildstate.spool2);
336  if (buildstate.btleader)
337  _bt_end_parallel(buildstate.btleader);
338 
339  result = (IndexBuildResult *) palloc(sizeof(IndexBuildResult));
340 
341  result->heap_tuples = reltuples;
342  result->index_tuples = buildstate.indtuples;
343 
344 #ifdef BTREE_BUILD_STATS
346  {
347  ShowUsage("BTREE BUILD STATS");
348  ResetUsage();
349  }
350 #endif /* BTREE_BUILD_STATS */
351 
352  return result;
353 }
354 
355 /*
356  * Create and initialize one or two spool structures, and save them in caller's
357  * buildstate argument. May also fill-in fields within indexInfo used by index
358  * builds.
359  *
360  * Scans the heap, possibly in parallel, filling spools with IndexTuples. This
361  * routine encapsulates all aspects of managing parallelism. Caller need only
362  * call _bt_end_parallel() in parallel case after it is done with spool/spool2.
363  *
364  * Returns the total number of heap tuples scanned.
365  */
366 static double
368  IndexInfo *indexInfo)
369 {
370  BTSpool *btspool = (BTSpool *) palloc0(sizeof(BTSpool));
371  SortCoordinate coordinate = NULL;
372  double reltuples = 0;
373 
374  /*
375  * We size the sort area as maintenance_work_mem rather than work_mem to
376  * speed index creation. This should be OK since a single backend can't
377  * run multiple index creations in parallel (see also: notes on
378  * parallelism and maintenance_work_mem below).
379  */
380  btspool->heap = heap;
381  btspool->index = index;
382  btspool->isunique = indexInfo->ii_Unique;
383 
384  /* Save as primary spool */
385  buildstate->spool = btspool;
386 
387  /* Report table scan phase started */
390 
391  /* Attempt to launch parallel worker scan when required */
392  if (indexInfo->ii_ParallelWorkers > 0)
393  _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  if (buildstate->btleader)
401  {
402  coordinate = (SortCoordinate) palloc0(sizeof(SortCoordinateData));
403  coordinate->isWorker = false;
404  coordinate->nParticipants =
405  buildstate->btleader->nparticipanttuplesorts;
406  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  buildstate->spool->sortstate =
431  tuplesort_begin_index_btree(heap, index, buildstate->isunique,
432  maintenance_work_mem, coordinate,
433  false);
434 
435  /*
436  * If building a unique index, put dead tuples in a second spool to keep
437  * them out of the uniqueness check. We expect that the second spool (for
438  * dead tuples) won't get very full, so we give it only work_mem.
439  */
440  if (indexInfo->ii_Unique)
441  {
442  BTSpool *btspool2 = (BTSpool *) palloc0(sizeof(BTSpool));
443  SortCoordinate coordinate2 = NULL;
444 
445  /* Initialize secondary spool */
446  btspool2->heap = heap;
447  btspool2->index = index;
448  btspool2->isunique = false;
449  /* Save as secondary spool */
450  buildstate->spool2 = btspool2;
451 
452  if (buildstate->btleader)
453  {
454  /*
455  * Set up non-private state that is passed to
456  * tuplesort_begin_index_btree() about the basic high level
457  * coordination of a parallel sort.
458  */
459  coordinate2 = (SortCoordinate) palloc0(sizeof(SortCoordinateData));
460  coordinate2->isWorker = false;
461  coordinate2->nParticipants =
462  buildstate->btleader->nparticipanttuplesorts;
463  coordinate2->sharedsort = buildstate->btleader->sharedsort2;
464  }
465 
466  /*
467  * We expect that the second one (for dead tuples) won't get very
468  * full, so we give it only work_mem
469  */
470  buildstate->spool2->sortstate =
471  tuplesort_begin_index_btree(heap, index, false, work_mem,
472  coordinate2, false);
473  }
474 
475  /* Fill spool using either serial or parallel heap scan */
476  if (!buildstate->btleader)
477  reltuples = table_index_build_scan(heap, index, indexInfo, true, true,
478  _bt_build_callback, (void *) buildstate,
479  NULL);
480  else
481  reltuples = _bt_parallel_heapscan(buildstate,
482  &indexInfo->ii_BrokenHotChain);
483 
484  /*
485  * Set the progress target for the next phase. Reset the block number
486  * values set by table_index_build_scan
487  */
488  {
489  const int progress_index[] = {
493  };
494  const int64 progress_vals[] = {
495  buildstate->indtuples,
496  0, 0
497  };
498 
499  pgstat_progress_update_multi_param(3, progress_index, progress_vals);
500  }
501 
502  /* okay, all heap tuples are spooled */
503  if (buildstate->spool2 && !buildstate->havedead)
504  {
505  /* spool2 turns out to be unnecessary */
506  _bt_spooldestroy(buildstate->spool2);
507  buildstate->spool2 = NULL;
508  }
509 
510  return reltuples;
511 }
512 
513 /*
514  * clean up a spool structure and its substructures.
515  */
516 static void
518 {
519  tuplesort_end(btspool->sortstate);
520  pfree(btspool);
521 }
522 
523 /*
524  * spool an index entry into the sort file.
525  */
526 static void
527 _bt_spool(BTSpool *btspool, ItemPointer self, Datum *values, bool *isnull)
528 {
529  tuplesort_putindextuplevalues(btspool->sortstate, btspool->index,
530  self, values, isnull);
531 }
532 
533 /*
534  * given a spool loaded by successive calls to _bt_spool,
535  * create an entire btree.
536  */
537 static void
538 _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2)
539 {
540  BTWriteState wstate;
541 
542 #ifdef BTREE_BUILD_STATS
544  {
545  ShowUsage("BTREE BUILD (Spool) STATISTICS");
546  ResetUsage();
547  }
548 #endif /* BTREE_BUILD_STATS */
549 
550  /* Execute the sort */
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  /* XLOG stuff */
641  if (wstate->btws_use_wal)
642  {
643  /* We use the XLOG_FPI record type for this */
644  log_newpage(&wstate->index->rd_node, MAIN_FORKNUM, blkno, page, true);
645  }
646 
647  /*
648  * If we have to write pages nonsequentially, fill in the space with
649  * zeroes until we come back and overwrite. This is not logically
650  * necessary on standard Unix filesystems (unwritten space will read as
651  * zeroes anyway), but it should help to avoid fragmentation. The dummy
652  * pages aren't WAL-logged though.
653  */
654  while (blkno > wstate->btws_pages_written)
655  {
656  if (!wstate->btws_zeropage)
657  wstate->btws_zeropage = (Page) palloc0(BLCKSZ);
658  /* don't set checksum for all-zero page */
660  wstate->btws_pages_written++,
661  (char *) wstate->btws_zeropage,
662  true);
663  }
664 
665  PageSetChecksumInplace(page, blkno);
666 
667  /*
668  * Now write the page. There's no need for smgr to schedule an fsync for
669  * this write; we'll do it ourselves before ending the build.
670  */
671  if (blkno == wstate->btws_pages_written)
672  {
673  /* extending the file... */
674  smgrextend(RelationGetSmgr(wstate->index), MAIN_FORKNUM, blkno,
675  (char *) page, true);
676  wstate->btws_pages_written++;
677  }
678  else
679  {
680  /* overwriting a block we zero-filled before */
681  smgrwrite(RelationGetSmgr(wstate->index), MAIN_FORKNUM, blkno,
682  (char *) page, true);
683  }
684 
685  pfree(page);
686 }
687 
688 /*
689  * allocate and initialize a new BTPageState. the returned structure
690  * is suitable for immediate use by _bt_buildadd.
691  */
692 static BTPageState *
694 {
696 
697  /* create initial page for level */
698  state->btps_page = _bt_blnewpage(level);
699 
700  /* and assign it a page position */
701  state->btps_blkno = wstate->btws_pages_alloced++;
702 
703  state->btps_lowkey = NULL;
704  /* initialize lastoff so first item goes into P_FIRSTKEY */
705  state->btps_lastoff = P_HIKEY;
706  state->btps_lastextra = 0;
707  state->btps_level = level;
708  /* set "full" threshold based on level. See notes at head of file. */
709  if (level > 0)
710  state->btps_full = (BLCKSZ * (100 - BTREE_NONLEAF_FILLFACTOR) / 100);
711  else
712  state->btps_full = BTGetTargetPageFreeSpace(wstate->index);
713 
714  /* no parent level, yet */
715  state->btps_next = NULL;
716 
717  return state;
718 }
719 
720 /*
721  * Slide the array of ItemIds from the page back one slot (from P_FIRSTKEY to
722  * P_HIKEY, overwriting P_HIKEY).
723  *
724  * _bt_blnewpage() makes the P_HIKEY line pointer appear allocated, but the
725  * rightmost page on its level is not supposed to get a high key. Now that
726  * it's clear that this page is a rightmost page, remove the unneeded empty
727  * P_HIKEY line pointer space.
728  */
729 static void
730 _bt_slideleft(Page rightmostpage)
731 {
732  OffsetNumber off;
733  OffsetNumber maxoff;
734  ItemId previi;
735 
736  maxoff = PageGetMaxOffsetNumber(rightmostpage);
737  Assert(maxoff >= P_FIRSTKEY);
738  previi = PageGetItemId(rightmostpage, P_HIKEY);
739  for (off = P_FIRSTKEY; off <= maxoff; off = OffsetNumberNext(off))
740  {
741  ItemId thisii = PageGetItemId(rightmostpage, off);
742 
743  *previi = *thisii;
744  previi = thisii;
745  }
746  ((PageHeader) rightmostpage)->pd_lower -= sizeof(ItemIdData);
747 }
748 
749 /*
750  * Add an item to a page being built.
751  *
752  * This is very similar to nbtinsert.c's _bt_pgaddtup(), but this variant
753  * raises an error directly.
754  *
755  * Note that our nbtsort.c caller does not know yet if the page will be
756  * rightmost. Offset P_FIRSTKEY is always assumed to be the first data key by
757  * caller. Page that turns out to be the rightmost on its level is fixed by
758  * calling _bt_slideleft().
759  */
760 static void
762  Size itemsize,
763  IndexTuple itup,
764  OffsetNumber itup_off,
765  bool newfirstdataitem)
766 {
767  IndexTupleData trunctuple;
768 
769  if (newfirstdataitem)
770  {
771  trunctuple = *itup;
772  trunctuple.t_info = sizeof(IndexTupleData);
773  BTreeTupleSetNAtts(&trunctuple, 0, false);
774  itup = &trunctuple;
775  itemsize = sizeof(IndexTupleData);
776  }
777 
778  if (PageAddItem(page, (Item) itup, itemsize, itup_off,
779  false, false) == InvalidOffsetNumber)
780  elog(ERROR, "failed to add item to the index page");
781 }
782 
783 /*----------
784  * Add an item to a disk page from the sort output (or add a posting list
785  * item formed from the sort output).
786  *
787  * We must be careful to observe the page layout conventions of nbtsearch.c:
788  * - rightmost pages start data items at P_HIKEY instead of at P_FIRSTKEY.
789  * - on non-leaf pages, the key portion of the first item need not be
790  * stored, we should store only the link.
791  *
792  * A leaf page being built looks like:
793  *
794  * +----------------+---------------------------------+
795  * | PageHeaderData | linp0 linp1 linp2 ... |
796  * +-----------+----+---------------------------------+
797  * | ... linpN | |
798  * +-----------+--------------------------------------+
799  * | ^ last |
800  * | |
801  * +-------------+------------------------------------+
802  * | | itemN ... |
803  * +-------------+------------------+-----------------+
804  * | ... item3 item2 item1 | "special space" |
805  * +--------------------------------+-----------------+
806  *
807  * Contrast this with the diagram in bufpage.h; note the mismatch
808  * between linps and items. This is because we reserve linp0 as a
809  * placeholder for the pointer to the "high key" item; when we have
810  * filled up the page, we will set linp0 to point to itemN and clear
811  * linpN. On the other hand, if we find this is the last (rightmost)
812  * page, we leave the items alone and slide the linp array over. If
813  * the high key is to be truncated, offset 1 is deleted, and we insert
814  * the truncated high key at offset 1.
815  *
816  * 'last' pointer indicates the last offset added to the page.
817  *
818  * 'truncextra' is the size of the posting list in itup, if any. This
819  * information is stashed for the next call here, when we may benefit
820  * from considering the impact of truncating away the posting list on
821  * the page before deciding to finish the page off. Posting lists are
822  * often relatively large, so it is worth going to the trouble of
823  * accounting for the saving from truncating away the posting list of
824  * the tuple that becomes the high key (that may be the only way to
825  * get close to target free space on the page). Note that this is
826  * only used for the soft fillfactor-wise limit, not the critical hard
827  * limit.
828  *----------
829  */
830 static void
832  Size truncextra)
833 {
834  Page npage;
835  BlockNumber nblkno;
836  OffsetNumber last_off;
837  Size last_truncextra;
838  Size pgspc;
839  Size itupsz;
840  bool isleaf;
841 
842  /*
843  * This is a handy place to check for cancel interrupts during the btree
844  * load phase of index creation.
845  */
847 
848  npage = state->btps_page;
849  nblkno = state->btps_blkno;
850  last_off = state->btps_lastoff;
851  last_truncextra = state->btps_lastextra;
852  state->btps_lastextra = truncextra;
853 
854  pgspc = PageGetFreeSpace(npage);
855  itupsz = IndexTupleSize(itup);
856  itupsz = MAXALIGN(itupsz);
857  /* Leaf case has slightly different rules due to suffix truncation */
858  isleaf = (state->btps_level == 0);
859 
860  /*
861  * Check whether the new item can fit on a btree page on current level at
862  * all.
863  *
864  * Every newly built index will treat heap TID as part of the keyspace,
865  * which imposes the requirement that new high keys must occasionally have
866  * a heap TID appended within _bt_truncate(). That may leave a new pivot
867  * tuple one or two MAXALIGN() quantums larger than the original
868  * firstright tuple it's derived from. v4 deals with the problem by
869  * decreasing the limit on the size of tuples inserted on the leaf level
870  * by the same small amount. Enforce the new v4+ limit on the leaf level,
871  * and the old limit on internal levels, since pivot tuples may need to
872  * make use of the reserved space. This should never fail on internal
873  * pages.
874  */
875  if (unlikely(itupsz > BTMaxItemSize(npage)))
876  _bt_check_third_page(wstate->index, wstate->heap, isleaf, npage,
877  itup);
878 
879  /*
880  * Check to see if current page will fit new item, with space left over to
881  * append a heap TID during suffix truncation when page is a leaf page.
882  *
883  * It is guaranteed that we can fit at least 2 non-pivot tuples plus a
884  * high key with heap TID when finishing off a leaf page, since we rely on
885  * _bt_check_third_page() rejecting oversized non-pivot tuples. On
886  * internal pages we can always fit 3 pivot tuples with larger internal
887  * page tuple limit (includes page high key).
888  *
889  * Most of the time, a page is only "full" in the sense that the soft
890  * fillfactor-wise limit has been exceeded. However, we must always leave
891  * at least two items plus a high key on each page before starting a new
892  * page. Disregard fillfactor and insert on "full" current page if we
893  * don't have the minimum number of items yet. (Note that we deliberately
894  * assume that suffix truncation neither enlarges nor shrinks new high key
895  * when applying soft limit, except when last tuple has a posting list.)
896  */
897  Assert(last_truncextra == 0 || isleaf);
898  if (pgspc < itupsz + (isleaf ? MAXALIGN(sizeof(ItemPointerData)) : 0) ||
899  (pgspc + last_truncextra < state->btps_full && last_off > P_FIRSTKEY))
900  {
901  /*
902  * Finish off the page and write it out.
903  */
904  Page opage = npage;
905  BlockNumber oblkno = nblkno;
906  ItemId ii;
907  ItemId hii;
908  IndexTuple oitup;
909 
910  /* Create new page of same level */
911  npage = _bt_blnewpage(state->btps_level);
912 
913  /* and assign it a page position */
914  nblkno = wstate->btws_pages_alloced++;
915 
916  /*
917  * We copy the last item on the page into the new page, and then
918  * rearrange the old page so that the 'last item' becomes its high key
919  * rather than a true data item. There had better be at least two
920  * items on the page already, else the page would be empty of useful
921  * data.
922  */
923  Assert(last_off > P_FIRSTKEY);
924  ii = PageGetItemId(opage, last_off);
925  oitup = (IndexTuple) PageGetItem(opage, ii);
926  _bt_sortaddtup(npage, ItemIdGetLength(ii), oitup, P_FIRSTKEY,
927  !isleaf);
928 
929  /*
930  * Move 'last' into the high key position on opage. _bt_blnewpage()
931  * allocated empty space for a line pointer when opage was first
932  * created, so this is a matter of rearranging already-allocated space
933  * on page, and initializing high key line pointer. (Actually, leaf
934  * pages must also swap oitup with a truncated version of oitup, which
935  * is sometimes larger than oitup, though never by more than the space
936  * needed to append a heap TID.)
937  */
938  hii = PageGetItemId(opage, P_HIKEY);
939  *hii = *ii;
940  ItemIdSetUnused(ii); /* redundant */
941  ((PageHeader) opage)->pd_lower -= sizeof(ItemIdData);
942 
943  if (isleaf)
944  {
945  IndexTuple lastleft;
946  IndexTuple truncated;
947 
948  /*
949  * Truncate away any unneeded attributes from high key on leaf
950  * level. This is only done at the leaf level because downlinks
951  * in internal pages are either negative infinity items, or get
952  * their contents from copying from one level down. See also:
953  * _bt_split().
954  *
955  * We don't try to bias our choice of split point to make it more
956  * likely that _bt_truncate() can truncate away more attributes,
957  * whereas the split point used within _bt_split() is chosen much
958  * more delicately. Even still, the lastleft and firstright
959  * tuples passed to _bt_truncate() here are at least not fully
960  * equal to each other when deduplication is used, unless there is
961  * a large group of duplicates (also, unique index builds usually
962  * have few or no spool2 duplicates). When the split point is
963  * between two unequal tuples, _bt_truncate() will avoid including
964  * a heap TID in the new high key, which is the most important
965  * benefit of suffix truncation.
966  *
967  * Overwrite the old item with new truncated high key directly.
968  * oitup is already located at the physical beginning of tuple
969  * space, so this should directly reuse the existing tuple space.
970  */
971  ii = PageGetItemId(opage, OffsetNumberPrev(last_off));
972  lastleft = (IndexTuple) PageGetItem(opage, ii);
973 
974  Assert(IndexTupleSize(oitup) > last_truncextra);
975  truncated = _bt_truncate(wstate->index, lastleft, oitup,
976  wstate->inskey);
977  if (!PageIndexTupleOverwrite(opage, P_HIKEY, (Item) truncated,
978  IndexTupleSize(truncated)))
979  elog(ERROR, "failed to add high key to the index page");
980  pfree(truncated);
981 
982  /* oitup should continue to point to the page's high key */
983  hii = PageGetItemId(opage, P_HIKEY);
984  oitup = (IndexTuple) PageGetItem(opage, hii);
985  }
986 
987  /*
988  * Link the old page into its parent, using its low key. If we don't
989  * have a parent, we have to create one; this adds a new btree level.
990  */
991  if (state->btps_next == NULL)
992  state->btps_next = _bt_pagestate(wstate, state->btps_level + 1);
993 
994  Assert((BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) <=
996  BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) > 0) ||
998  Assert(BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) == 0 ||
1000  BTreeTupleSetDownLink(state->btps_lowkey, oblkno);
1001  _bt_buildadd(wstate, state->btps_next, state->btps_lowkey, 0);
1002  pfree(state->btps_lowkey);
1003 
1004  /*
1005  * Save a copy of the high key from the old page. It is also the low
1006  * key for the new page.
1007  */
1008  state->btps_lowkey = CopyIndexTuple(oitup);
1009 
1010  /*
1011  * Set the sibling links for both pages.
1012  */
1013  {
1014  BTPageOpaque oopaque = (BTPageOpaque) PageGetSpecialPointer(opage);
1015  BTPageOpaque nopaque = (BTPageOpaque) PageGetSpecialPointer(npage);
1016 
1017  oopaque->btpo_next = nblkno;
1018  nopaque->btpo_prev = oblkno;
1019  nopaque->btpo_next = P_NONE; /* redundant */
1020  }
1021 
1022  /*
1023  * Write out the old page. We never need to touch it again, so we can
1024  * free the opage workspace too.
1025  */
1026  _bt_blwritepage(wstate, opage, oblkno);
1027 
1028  /*
1029  * Reset last_off to point to new page
1030  */
1031  last_off = P_FIRSTKEY;
1032  }
1033 
1034  /*
1035  * By here, either original page is still the current page, or a new page
1036  * was created that became the current page. Either way, the current page
1037  * definitely has space for new item.
1038  *
1039  * If the new item is the first for its page, it must also be the first
1040  * item on its entire level. On later same-level pages, a low key for a
1041  * page will be copied from the prior page in the code above. Generate a
1042  * minus infinity low key here instead.
1043  */
1044  if (last_off == P_HIKEY)
1045  {
1046  Assert(state->btps_lowkey == NULL);
1047  state->btps_lowkey = palloc0(sizeof(IndexTupleData));
1048  state->btps_lowkey->t_info = sizeof(IndexTupleData);
1049  BTreeTupleSetNAtts(state->btps_lowkey, 0, false);
1050  }
1051 
1052  /*
1053  * Add the new item into the current page.
1054  */
1055  last_off = OffsetNumberNext(last_off);
1056  _bt_sortaddtup(npage, itupsz, itup, last_off,
1057  !isleaf && last_off == P_FIRSTKEY);
1058 
1059  state->btps_page = npage;
1060  state->btps_blkno = nblkno;
1061  state->btps_lastoff = last_off;
1062 }
1063 
1064 /*
1065  * Finalize pending posting list tuple, and add it to the index. Final tuple
1066  * is based on saved base tuple, and saved list of heap TIDs.
1067  *
1068  * This is almost like _bt_dedup_finish_pending(), but it adds a new tuple
1069  * using _bt_buildadd().
1070  */
1071 static void
1073  BTDedupState dstate)
1074 {
1075  Assert(dstate->nitems > 0);
1076 
1077  if (dstate->nitems == 1)
1078  _bt_buildadd(wstate, state, dstate->base, 0);
1079  else
1080  {
1081  IndexTuple postingtuple;
1082  Size truncextra;
1083 
1084  /* form a tuple with a posting list */
1085  postingtuple = _bt_form_posting(dstate->base,
1086  dstate->htids,
1087  dstate->nhtids);
1088  /* Calculate posting list overhead */
1089  truncextra = IndexTupleSize(postingtuple) -
1090  BTreeTupleGetPostingOffset(postingtuple);
1091 
1092  _bt_buildadd(wstate, state, postingtuple, truncextra);
1093  pfree(postingtuple);
1094  }
1095 
1096  dstate->nmaxitems = 0;
1097  dstate->nhtids = 0;
1098  dstate->nitems = 0;
1099  dstate->phystupsize = 0;
1100 }
1101 
1102 /*
1103  * Finish writing out the completed btree.
1104  */
1105 static void
1107 {
1108  BTPageState *s;
1109  BlockNumber rootblkno = P_NONE;
1110  uint32 rootlevel = 0;
1111  Page metapage;
1112 
1113  /*
1114  * Each iteration of this loop completes one more level of the tree.
1115  */
1116  for (s = state; s != NULL; s = s->btps_next)
1117  {
1118  BlockNumber blkno;
1119  BTPageOpaque opaque;
1120 
1121  blkno = s->btps_blkno;
1123 
1124  /*
1125  * We have to link the last page on this level to somewhere.
1126  *
1127  * If we're at the top, it's the root, so attach it to the metapage.
1128  * Otherwise, add an entry for it to its parent using its low key.
1129  * This may cause the last page of the parent level to split, but
1130  * that's not a problem -- we haven't gotten to it yet.
1131  */
1132  if (s->btps_next == NULL)
1133  {
1134  opaque->btpo_flags |= BTP_ROOT;
1135  rootblkno = blkno;
1136  rootlevel = s->btps_level;
1137  }
1138  else
1139  {
1140  Assert((BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) <=
1142  BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) > 0) ||
1143  P_LEFTMOST(opaque));
1144  Assert(BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) == 0 ||
1145  !P_LEFTMOST(opaque));
1147  _bt_buildadd(wstate, s->btps_next, s->btps_lowkey, 0);
1148  pfree(s->btps_lowkey);
1149  s->btps_lowkey = NULL;
1150  }
1151 
1152  /*
1153  * This is the rightmost page, so the ItemId array needs to be slid
1154  * back one slot. Then we can dump out the page.
1155  */
1157  _bt_blwritepage(wstate, s->btps_page, s->btps_blkno);
1158  s->btps_page = NULL; /* writepage freed the workspace */
1159  }
1160 
1161  /*
1162  * As the last step in the process, construct the metapage and make it
1163  * point to the new root (unless we had no data at all, in which case it's
1164  * set to point to "P_NONE"). This changes the index to the "valid" state
1165  * by filling in a valid magic number in the metapage.
1166  */
1167  metapage = (Page) palloc(BLCKSZ);
1168  _bt_initmetapage(metapage, rootblkno, rootlevel,
1169  wstate->inskey->allequalimage);
1170  _bt_blwritepage(wstate, metapage, BTREE_METAPAGE);
1171 }
1172 
1173 /*
1174  * Read tuples in correct sort order from tuplesort, and load them into
1175  * btree leaves.
1176  */
1177 static void
1178 _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
1179 {
1180  BTPageState *state = NULL;
1181  bool merge = (btspool2 != NULL);
1182  IndexTuple itup,
1183  itup2 = NULL;
1184  bool load1;
1185  TupleDesc tupdes = RelationGetDescr(wstate->index);
1186  int i,
1188  SortSupport sortKeys;
1189  int64 tuples_done = 0;
1190  bool deduplicate;
1191 
1192  deduplicate = wstate->inskey->allequalimage && !btspool->isunique &&
1193  BTGetDeduplicateItems(wstate->index);
1194 
1195  if (merge)
1196  {
1197  /*
1198  * Another BTSpool for dead tuples exists. Now we have to merge
1199  * btspool and btspool2.
1200  */
1201 
1202  /* the preparation of merge */
1203  itup = tuplesort_getindextuple(btspool->sortstate, true);
1204  itup2 = tuplesort_getindextuple(btspool2->sortstate, true);
1205 
1206  /* Prepare SortSupport data for each column */
1207  sortKeys = (SortSupport) palloc0(keysz * sizeof(SortSupportData));
1208 
1209  for (i = 0; i < keysz; i++)
1210  {
1211  SortSupport sortKey = sortKeys + i;
1212  ScanKey scanKey = wstate->inskey->scankeys + i;
1213  int16 strategy;
1214 
1215  sortKey->ssup_cxt = CurrentMemoryContext;
1216  sortKey->ssup_collation = scanKey->sk_collation;
1217  sortKey->ssup_nulls_first =
1218  (scanKey->sk_flags & SK_BT_NULLS_FIRST) != 0;
1219  sortKey->ssup_attno = scanKey->sk_attno;
1220  /* Abbreviation is not supported here */
1221  sortKey->abbreviate = false;
1222 
1223  AssertState(sortKey->ssup_attno != 0);
1224 
1225  strategy = (scanKey->sk_flags & SK_BT_DESC) != 0 ?
1227 
1228  PrepareSortSupportFromIndexRel(wstate->index, strategy, sortKey);
1229  }
1230 
1231  for (;;)
1232  {
1233  load1 = true; /* load BTSpool next ? */
1234  if (itup2 == NULL)
1235  {
1236  if (itup == NULL)
1237  break;
1238  }
1239  else if (itup != NULL)
1240  {
1241  int32 compare = 0;
1242 
1243  for (i = 1; i <= keysz; i++)
1244  {
1245  SortSupport entry;
1246  Datum attrDatum1,
1247  attrDatum2;
1248  bool isNull1,
1249  isNull2;
1250 
1251  entry = sortKeys + i - 1;
1252  attrDatum1 = index_getattr(itup, i, tupdes, &isNull1);
1253  attrDatum2 = index_getattr(itup2, i, tupdes, &isNull2);
1254 
1255  compare = ApplySortComparator(attrDatum1, isNull1,
1256  attrDatum2, isNull2,
1257  entry);
1258  if (compare > 0)
1259  {
1260  load1 = false;
1261  break;
1262  }
1263  else if (compare < 0)
1264  break;
1265  }
1266 
1267  /*
1268  * If key values are equal, we sort on ItemPointer. This is
1269  * required for btree indexes, since heap TID is treated as an
1270  * implicit last key attribute in order to ensure that all
1271  * keys in the index are physically unique.
1272  */
1273  if (compare == 0)
1274  {
1275  compare = ItemPointerCompare(&itup->t_tid, &itup2->t_tid);
1276  Assert(compare != 0);
1277  if (compare > 0)
1278  load1 = false;
1279  }
1280  }
1281  else
1282  load1 = false;
1283 
1284  /* When we see first tuple, create first index page */
1285  if (state == NULL)
1286  state = _bt_pagestate(wstate, 0);
1287 
1288  if (load1)
1289  {
1290  _bt_buildadd(wstate, state, itup, 0);
1291  itup = tuplesort_getindextuple(btspool->sortstate, true);
1292  }
1293  else
1294  {
1295  _bt_buildadd(wstate, state, itup2, 0);
1296  itup2 = tuplesort_getindextuple(btspool2->sortstate, true);
1297  }
1298 
1299  /* Report progress */
1301  ++tuples_done);
1302  }
1303  pfree(sortKeys);
1304  }
1305  else if (deduplicate)
1306  {
1307  /* merge is unnecessary, deduplicate into posting lists */
1308  BTDedupState dstate;
1309 
1310  dstate = (BTDedupState) palloc(sizeof(BTDedupStateData));
1311  dstate->deduplicate = true; /* unused */
1312  dstate->nmaxitems = 0; /* 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)
1429 }
1430 
1431 /*
1432  * Create parallel context, and launch workers for leader.
1433  *
1434  * buildstate argument should be initialized (with the exception of the
1435  * tuplesort state in spools, which may later be created based on shared
1436  * state initially set up here).
1437  *
1438  * isconcurrent indicates if operation is CREATE INDEX CONCURRENTLY.
1439  *
1440  * request is the target number of parallel worker processes to launch.
1441  *
1442  * Sets buildstate's BTLeader, which caller must use to shut down parallel
1443  * mode by passing it to _bt_end_parallel() at the very end of its index
1444  * build. If not even a single worker process can be launched, this is
1445  * never set, and caller should proceed with a serial index build.
1446  */
1447 static void
1448 _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent, int request)
1449 {
1450  ParallelContext *pcxt;
1451  int scantuplesortstates;
1452  Snapshot snapshot;
1453  Size estbtshared;
1454  Size estsort;
1455  BTShared *btshared;
1456  Sharedsort *sharedsort;
1457  Sharedsort *sharedsort2;
1458  BTSpool *btspool = buildstate->spool;
1459  BTLeader *btleader = (BTLeader *) palloc0(sizeof(BTLeader));
1460  WalUsage *walusage;
1461  BufferUsage *bufferusage;
1462  bool leaderparticipates = true;
1463  int querylen;
1464 
1465 #ifdef DISABLE_LEADER_PARTICIPATION
1466  leaderparticipates = false;
1467 #endif
1468 
1469  /*
1470  * Enter parallel mode, and create context for parallel build of btree
1471  * index
1472  */
1474  Assert(request > 0);
1475  pcxt = CreateParallelContext("postgres", "_bt_parallel_build_main",
1476  request);
1477 
1478  scantuplesortstates = leaderparticipates ? request + 1 : request;
1479 
1480  /*
1481  * Prepare for scan of the base relation. In a normal index build, we use
1482  * SnapshotAny because we must retrieve all tuples and do our own time
1483  * qual checks (because we have to index RECENTLY_DEAD tuples). In a
1484  * concurrent build, we take a regular MVCC snapshot and index whatever's
1485  * live according to that.
1486  */
1487  if (!isconcurrent)
1488  snapshot = SnapshotAny;
1489  else
1491 
1492  /*
1493  * Estimate size for our own PARALLEL_KEY_BTREE_SHARED workspace, and
1494  * PARALLEL_KEY_TUPLESORT tuplesort workspace
1495  */
1496  estbtshared = _bt_parallel_estimate_shared(btspool->heap, snapshot);
1497  shm_toc_estimate_chunk(&pcxt->estimator, estbtshared);
1498  estsort = tuplesort_estimate_shared(scantuplesortstates);
1499  shm_toc_estimate_chunk(&pcxt->estimator, estsort);
1500 
1501  /*
1502  * Unique case requires a second spool, and so we may have to account for
1503  * another shared workspace for that -- PARALLEL_KEY_TUPLESORT_SPOOL2
1504  */
1505  if (!btspool->isunique)
1506  shm_toc_estimate_keys(&pcxt->estimator, 2);
1507  else
1508  {
1509  shm_toc_estimate_chunk(&pcxt->estimator, estsort);
1510  shm_toc_estimate_keys(&pcxt->estimator, 3);
1511  }
1512 
1513  /*
1514  * Estimate space for WalUsage and BufferUsage -- PARALLEL_KEY_WAL_USAGE
1515  * and PARALLEL_KEY_BUFFER_USAGE.
1516  *
1517  * If there are no extensions loaded that care, we could skip this. We
1518  * have no way of knowing whether anyone's looking at pgWalUsage or
1519  * pgBufferUsage, so do it unconditionally.
1520  */
1522  mul_size(sizeof(WalUsage), pcxt->nworkers));
1523  shm_toc_estimate_keys(&pcxt->estimator, 1);
1525  mul_size(sizeof(BufferUsage), pcxt->nworkers));
1526  shm_toc_estimate_keys(&pcxt->estimator, 1);
1527 
1528  /* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
1529  if (debug_query_string)
1530  {
1531  querylen = strlen(debug_query_string);
1532  shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
1533  shm_toc_estimate_keys(&pcxt->estimator, 1);
1534  }
1535  else
1536  querylen = 0; /* keep compiler quiet */
1537 
1538  /* Everyone's had a chance to ask for space, so now create the DSM */
1539  InitializeParallelDSM(pcxt);
1540 
1541  /* If no DSM segment was available, back out (do serial build) */
1542  if (pcxt->seg == NULL)
1543  {
1544  if (IsMVCCSnapshot(snapshot))
1545  UnregisterSnapshot(snapshot);
1546  DestroyParallelContext(pcxt);
1547  ExitParallelMode();
1548  return;
1549  }
1550 
1551  /* Store shared build state, for which we reserved space */
1552  btshared = (BTShared *) shm_toc_allocate(pcxt->toc, estbtshared);
1553  /* Initialize immutable state */
1554  btshared->heaprelid = RelationGetRelid(btspool->heap);
1555  btshared->indexrelid = RelationGetRelid(btspool->index);
1556  btshared->isunique = btspool->isunique;
1557  btshared->isconcurrent = isconcurrent;
1558  btshared->scantuplesortstates = scantuplesortstates;
1560  SpinLockInit(&btshared->mutex);
1561  /* Initialize mutable state */
1562  btshared->nparticipantsdone = 0;
1563  btshared->reltuples = 0.0;
1564  btshared->havedead = false;
1565  btshared->indtuples = 0.0;
1566  btshared->brokenhotchain = false;
1569  snapshot);
1570 
1571  /*
1572  * Store shared tuplesort-private state, for which we reserved space.
1573  * Then, initialize opaque state using tuplesort routine.
1574  */
1575  sharedsort = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
1576  tuplesort_initialize_shared(sharedsort, scantuplesortstates,
1577  pcxt->seg);
1578 
1579  shm_toc_insert(pcxt->toc, PARALLEL_KEY_BTREE_SHARED, btshared);
1580  shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLESORT, sharedsort);
1581 
1582  /* Unique case requires a second spool, and associated shared state */
1583  if (!btspool->isunique)
1584  sharedsort2 = NULL;
1585  else
1586  {
1587  /*
1588  * Store additional shared tuplesort-private state, for which we
1589  * reserved space. Then, initialize opaque state using tuplesort
1590  * routine.
1591  */
1592  sharedsort2 = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
1593  tuplesort_initialize_shared(sharedsort2, scantuplesortstates,
1594  pcxt->seg);
1595 
1596  shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLESORT_SPOOL2, sharedsort2);
1597  }
1598 
1599  /* Store query string for workers */
1600  if (debug_query_string)
1601  {
1602  char *sharedquery;
1603 
1604  sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
1605  memcpy(sharedquery, debug_query_string, querylen + 1);
1606  shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, sharedquery);
1607  }
1608 
1609  /*
1610  * Allocate space for each worker's WalUsage and BufferUsage; no need to
1611  * initialize.
1612  */
1613  walusage = shm_toc_allocate(pcxt->toc,
1614  mul_size(sizeof(WalUsage), pcxt->nworkers));
1615  shm_toc_insert(pcxt->toc, PARALLEL_KEY_WAL_USAGE, walusage);
1616  bufferusage = shm_toc_allocate(pcxt->toc,
1617  mul_size(sizeof(BufferUsage), pcxt->nworkers));
1618  shm_toc_insert(pcxt->toc, PARALLEL_KEY_BUFFER_USAGE, bufferusage);
1619 
1620  /* Launch workers, saving status for leader/caller */
1621  LaunchParallelWorkers(pcxt);
1622  btleader->pcxt = pcxt;
1623  btleader->nparticipanttuplesorts = pcxt->nworkers_launched;
1624  if (leaderparticipates)
1625  btleader->nparticipanttuplesorts++;
1626  btleader->btshared = btshared;
1627  btleader->sharedsort = sharedsort;
1628  btleader->sharedsort2 = sharedsort2;
1629  btleader->snapshot = snapshot;
1630  btleader->walusage = walusage;
1631  btleader->bufferusage = bufferusage;
1632 
1633  /* If no workers were successfully launched, back out (do serial build) */
1634  if (pcxt->nworkers_launched == 0)
1635  {
1636  _bt_end_parallel(btleader);
1637  return;
1638  }
1639 
1640  /* Save leader state now that it's clear build will be parallel */
1641  buildstate->btleader = btleader;
1642 
1643  /* Join heap scan ourselves */
1644  if (leaderparticipates)
1646 
1647  /*
1648  * Caller needs to wait for all launched workers when we return. Make
1649  * sure that the failure-to-start case will not hang forever.
1650  */
1652 }
1653 
1654 /*
1655  * Shut down workers, destroy parallel context, and end parallel mode.
1656  */
1657 static void
1659 {
1660  int i;
1661 
1662  /* Shutdown worker processes */
1664 
1665  /*
1666  * Next, accumulate WAL usage. (This must wait for the workers to finish,
1667  * or we might get incomplete data.)
1668  */
1669  for (i = 0; i < btleader->pcxt->nworkers_launched; i++)
1670  InstrAccumParallelQuery(&btleader->bufferusage[i], &btleader->walusage[i]);
1671 
1672  /* Free last reference to MVCC snapshot, if one was used */
1673  if (IsMVCCSnapshot(btleader->snapshot))
1674  UnregisterSnapshot(btleader->snapshot);
1675  DestroyParallelContext(btleader->pcxt);
1676  ExitParallelMode();
1677 }
1678 
1679 /*
1680  * Returns size of shared memory required to store state for a parallel
1681  * btree index build based on the snapshot its parallel scan will use.
1682  */
1683 static Size
1685 {
1686  /* c.f. shm_toc_allocate as to why BUFFERALIGN is used */
1687  return add_size(BUFFERALIGN(sizeof(BTShared)),
1688  table_parallelscan_estimate(heap, snapshot));
1689 }
1690 
1691 /*
1692  * Within leader, wait for end of heap scan.
1693  *
1694  * When called, parallel heap scan started by _bt_begin_parallel() will
1695  * already be underway within worker processes (when leader participates
1696  * as a worker, we should end up here just as workers are finishing).
1697  *
1698  * Fills in fields needed for ambuild statistics, and lets caller set
1699  * field indicating that some worker encountered a broken HOT chain.
1700  *
1701  * Returns the total number of heap tuples scanned.
1702  */
1703 static double
1704 _bt_parallel_heapscan(BTBuildState *buildstate, bool *brokenhotchain)
1705 {
1706  BTShared *btshared = buildstate->btleader->btshared;
1707  int nparticipanttuplesorts;
1708  double reltuples;
1709 
1710  nparticipanttuplesorts = buildstate->btleader->nparticipanttuplesorts;
1711  for (;;)
1712  {
1713  SpinLockAcquire(&btshared->mutex);
1714  if (btshared->nparticipantsdone == nparticipanttuplesorts)
1715  {
1716  buildstate->havedead = btshared->havedead;
1717  buildstate->indtuples = btshared->indtuples;
1718  *brokenhotchain = btshared->brokenhotchain;
1719  reltuples = btshared->reltuples;
1720  SpinLockRelease(&btshared->mutex);
1721  break;
1722  }
1723  SpinLockRelease(&btshared->mutex);
1724 
1727  }
1728 
1730 
1731  return reltuples;
1732 }
1733 
1734 /*
1735  * Within leader, participate as a parallel worker.
1736  */
1737 static void
1739 {
1740  BTLeader *btleader = buildstate->btleader;
1741  BTSpool *leaderworker;
1742  BTSpool *leaderworker2;
1743  int sortmem;
1744 
1745  /* Allocate memory and initialize private spool */
1746  leaderworker = (BTSpool *) palloc0(sizeof(BTSpool));
1747  leaderworker->heap = buildstate->spool->heap;
1748  leaderworker->index = buildstate->spool->index;
1749  leaderworker->isunique = buildstate->spool->isunique;
1750 
1751  /* Initialize second spool, if required */
1752  if (!btleader->btshared->isunique)
1753  leaderworker2 = NULL;
1754  else
1755  {
1756  /* Allocate memory for worker's own private secondary spool */
1757  leaderworker2 = (BTSpool *) palloc0(sizeof(BTSpool));
1758 
1759  /* Initialize worker's own secondary spool */
1760  leaderworker2->heap = leaderworker->heap;
1761  leaderworker2->index = leaderworker->index;
1762  leaderworker2->isunique = false;
1763  }
1764 
1765  /*
1766  * Might as well use reliable figure when doling out maintenance_work_mem
1767  * (when requested number of workers were not launched, this will be
1768  * somewhat higher than it is for other workers).
1769  */
1770  sortmem = maintenance_work_mem / btleader->nparticipanttuplesorts;
1771 
1772  /* Perform work common to all participants */
1773  _bt_parallel_scan_and_sort(leaderworker, leaderworker2, btleader->btshared,
1774  btleader->sharedsort, btleader->sharedsort2,
1775  sortmem, true);
1776 
1777 #ifdef BTREE_BUILD_STATS
1779  {
1780  ShowUsage("BTREE BUILD (Leader Partial Spool) STATISTICS");
1781  ResetUsage();
1782  }
1783 #endif /* BTREE_BUILD_STATS */
1784 }
1785 
1786 /*
1787  * Perform work within a launched parallel process.
1788  */
1789 void
1791 {
1792  char *sharedquery;
1793  BTSpool *btspool;
1794  BTSpool *btspool2;
1795  BTShared *btshared;
1796  Sharedsort *sharedsort;
1797  Sharedsort *sharedsort2;
1798  Relation heapRel;
1799  Relation indexRel;
1800  LOCKMODE heapLockmode;
1801  LOCKMODE indexLockmode;
1802  WalUsage *walusage;
1803  BufferUsage *bufferusage;
1804  int sortmem;
1805 
1806 #ifdef BTREE_BUILD_STATS
1808  ResetUsage();
1809 #endif /* BTREE_BUILD_STATS */
1810 
1811  /* Set debug_query_string for individual workers first */
1812  sharedquery = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, true);
1813  debug_query_string = sharedquery;
1814 
1815  /* Report the query string from leader */
1817 
1818  /* Look up nbtree shared state */
1819  btshared = shm_toc_lookup(toc, PARALLEL_KEY_BTREE_SHARED, false);
1820 
1821  /* Open relations using lock modes known to be obtained by index.c */
1822  if (!btshared->isconcurrent)
1823  {
1824  heapLockmode = ShareLock;
1825  indexLockmode = AccessExclusiveLock;
1826  }
1827  else
1828  {
1829  heapLockmode = ShareUpdateExclusiveLock;
1830  indexLockmode = RowExclusiveLock;
1831  }
1832 
1833  /* Open relations within worker */
1834  heapRel = table_open(btshared->heaprelid, heapLockmode);
1835  indexRel = index_open(btshared->indexrelid, indexLockmode);
1836 
1837  /* Initialize worker's own spool */
1838  btspool = (BTSpool *) palloc0(sizeof(BTSpool));
1839  btspool->heap = heapRel;
1840  btspool->index = indexRel;
1841  btspool->isunique = btshared->isunique;
1842 
1843  /* Look up shared state private to tuplesort.c */
1844  sharedsort = shm_toc_lookup(toc, PARALLEL_KEY_TUPLESORT, false);
1845  tuplesort_attach_shared(sharedsort, seg);
1846  if (!btshared->isunique)
1847  {
1848  btspool2 = NULL;
1849  sharedsort2 = NULL;
1850  }
1851  else
1852  {
1853  /* Allocate memory for worker's own private secondary spool */
1854  btspool2 = (BTSpool *) palloc0(sizeof(BTSpool));
1855 
1856  /* Initialize worker's own secondary spool */
1857  btspool2->heap = btspool->heap;
1858  btspool2->index = btspool->index;
1859  btspool2->isunique = false;
1860  /* Look up shared state private to tuplesort.c */
1861  sharedsort2 = shm_toc_lookup(toc, PARALLEL_KEY_TUPLESORT_SPOOL2, false);
1862  tuplesort_attach_shared(sharedsort2, seg);
1863  }
1864 
1865  /* Prepare to track buffer usage during parallel execution */
1867 
1868  /* Perform sorting of spool, and possibly a spool2 */
1869  sortmem = maintenance_work_mem / btshared->scantuplesortstates;
1870  _bt_parallel_scan_and_sort(btspool, btspool2, btshared, sharedsort,
1871  sharedsort2, sortmem, false);
1872 
1873  /* Report WAL/buffer usage during parallel execution */
1874  bufferusage = shm_toc_lookup(toc, PARALLEL_KEY_BUFFER_USAGE, false);
1875  walusage = shm_toc_lookup(toc, PARALLEL_KEY_WAL_USAGE, false);
1877  &walusage[ParallelWorkerNumber]);
1878 
1879 #ifdef BTREE_BUILD_STATS
1881  {
1882  ShowUsage("BTREE BUILD (Worker Partial Spool) STATISTICS");
1883  ResetUsage();
1884  }
1885 #endif /* BTREE_BUILD_STATS */
1886 
1887  index_close(indexRel, indexLockmode);
1888  table_close(heapRel, heapLockmode);
1889 }
1890 
1891 /*
1892  * Perform a worker's portion of a parallel sort.
1893  *
1894  * This generates a tuplesort for passed btspool, and a second tuplesort
1895  * state if a second btspool is need (i.e. for unique index builds). All
1896  * other spool fields should already be set when this is called.
1897  *
1898  * sortmem is the amount of working memory to use within each worker,
1899  * expressed in KBs.
1900  *
1901  * When this returns, workers are done, and need only release resources.
1902  */
1903 static void
1905  BTShared *btshared, Sharedsort *sharedsort,
1906  Sharedsort *sharedsort2, int sortmem, bool progress)
1907 {
1908  SortCoordinate coordinate;
1909  BTBuildState buildstate;
1910  TableScanDesc scan;
1911  double reltuples;
1912  IndexInfo *indexInfo;
1913 
1914  /* Initialize local tuplesort coordination state */
1915  coordinate = palloc0(sizeof(SortCoordinateData));
1916  coordinate->isWorker = true;
1917  coordinate->nParticipants = -1;
1918  coordinate->sharedsort = sharedsort;
1919 
1920  /* Begin "partial" tuplesort */
1921  btspool->sortstate = tuplesort_begin_index_btree(btspool->heap,
1922  btspool->index,
1923  btspool->isunique,
1924  sortmem, coordinate,
1925  false);
1926 
1927  /*
1928  * Just as with serial case, there may be a second spool. If so, a
1929  * second, dedicated spool2 partial tuplesort is required.
1930  */
1931  if (btspool2)
1932  {
1933  SortCoordinate coordinate2;
1934 
1935  /*
1936  * We expect that the second one (for dead tuples) won't get very
1937  * full, so we give it only work_mem (unless sortmem is less for
1938  * worker). Worker processes are generally permitted to allocate
1939  * work_mem independently.
1940  */
1941  coordinate2 = palloc0(sizeof(SortCoordinateData));
1942  coordinate2->isWorker = true;
1943  coordinate2->nParticipants = -1;
1944  coordinate2->sharedsort = sharedsort2;
1945  btspool2->sortstate =
1946  tuplesort_begin_index_btree(btspool->heap, btspool->index, false,
1947  Min(sortmem, work_mem), coordinate2,
1948  false);
1949  }
1950 
1951  /* Fill in buildstate for _bt_build_callback() */
1952  buildstate.isunique = btshared->isunique;
1953  buildstate.havedead = false;
1954  buildstate.heap = btspool->heap;
1955  buildstate.spool = btspool;
1956  buildstate.spool2 = btspool2;
1957  buildstate.indtuples = 0;
1958  buildstate.btleader = NULL;
1959 
1960  /* Join parallel scan */
1961  indexInfo = BuildIndexInfo(btspool->index);
1962  indexInfo->ii_Concurrent = btshared->isconcurrent;
1963  scan = table_beginscan_parallel(btspool->heap,
1964  ParallelTableScanFromBTShared(btshared));
1965  reltuples = table_index_build_scan(btspool->heap, btspool->index, indexInfo,
1966  true, progress, _bt_build_callback,
1967  (void *) &buildstate, scan);
1968 
1969  /* Execute this worker's part of the sort */
1970  if (progress)
1973  tuplesort_performsort(btspool->sortstate);
1974  if (btspool2)
1975  {
1976  if (progress)
1979  tuplesort_performsort(btspool2->sortstate);
1980  }
1981 
1982  /*
1983  * Done. Record ambuild statistics, and whether we encountered a broken
1984  * HOT chain.
1985  */
1986  SpinLockAcquire(&btshared->mutex);
1987  btshared->nparticipantsdone++;
1988  btshared->reltuples += reltuples;
1989  if (buildstate.havedead)
1990  btshared->havedead = true;
1991  btshared->indtuples += buildstate.indtuples;
1992  if (indexInfo->ii_BrokenHotChain)
1993  btshared->brokenhotchain = true;
1994  SpinLockRelease(&btshared->mutex);
1995 
1996  /* Notify leader */
1998 
1999  /* We can end tuplesorts immediately */
2000  tuplesort_end(btspool->sortstate);
2001  if (btspool2)
2002  tuplesort_end(btspool2->sortstate);
2003 }
struct SortSupportData * SortSupport
Definition: sortsupport.h:58
IndexTuple tuplesort_getindextuple(Tuplesortstate *state, bool forward)
Definition: tuplesort.c:2465
static void _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent, int request)
Definition: nbtsort.c:1448
int32 ItemPointerCompare(ItemPointer arg1, ItemPointer arg2)
Definition: itemptr.c:52
signed short int16
Definition: c.h:428
bool ssup_nulls_first
Definition: sortsupport.h:75
int slock_t
Definition: s_lock.h:934
IndexTuple base
Definition: nbtree.h:871
#define BTP_ROOT
Definition: nbtree.h:75
IndexTuple _bt_form_posting(IndexTuple base, ItemPointer htids, int nhtids)
Definition: nbtdedup.c:859
Sharedsort * sharedsort2
Definition: nbtsort.c:194
void tuplesort_performsort(Tuplesortstate *state)
Definition: tuplesort.c:2040
BlockNumber btpo_next
Definition: nbtree.h:65
#define PROGRESS_BTREE_PHASE_INDEXBUILD_TABLESCAN
Definition: nbtree.h:1112
void table_close(Relation relation, LOCKMODE lockmode)
Definition: table.c:167
int scantuplesortstates
Definition: nbtsort.c:110
#define BTGreaterStrategyNumber
Definition: stratnum.h:33
struct BTWriteState BTWriteState
#define PARALLEL_KEY_TUPLESORT
Definition: nbtsort.c:69
Relation heap
Definition: nbtsort.c:210
#define AssertState(condition)
Definition: c.h:807
Page btws_zeropage
Definition: nbtsort.c:255
OffsetNumber baseoff
Definition: nbtree.h:872
ParallelContext * CreateParallelContext(const char *library_name, const char *function_name, int nworkers)
Definition: parallel.c:164
BTShared * btshared
Definition: nbtsort.c:192
int nparticipantsdone
Definition: nbtsort.c:144
static void _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
Definition: nbtsort.c:1178
Snapshot RegisterSnapshot(Snapshot snapshot)
Definition: snapmgr.c:810
#define RelationGetDescr(relation)
Definition: rel.h:503
int LOCKMODE
Definition: lockdefs.h:26
BTScanInsert _bt_mkscankey(Relation rel, IndexTuple itup)
Definition: nbtutils.c:90
#define BTGetTargetPageFreeSpace(relation)
Definition: nbtree.h:1099
Oid indexrelid
Definition: nbtsort.c:107
#define BTGetDeduplicateItems(relation)
Definition: nbtree.h:1101
#define PROGRESS_BTREE_PHASE_PERFORMSORT_1
Definition: nbtree.h:1113
static void _bt_sort_dedup_finish_pending(BTWriteState *wstate, BTPageState *state, BTDedupState dstate)
Definition: nbtsort.c:1072
void ShowUsage(const char *title)
Definition: postgres.c:4795
dsm_segment * seg
Definition: parallel.h:43
BlockNumber btws_pages_alloced
Definition: nbtsort.c:253
Sharedsort * sharedsort
Definition: nbtsort.c:193
shm_toc_estimator estimator
Definition: parallel.h:42
#define SpinLockInit(lock)
Definition: spin.h:60
#define BTP_LEAF
Definition: nbtree.h:74
ItemPointerData t_tid
Definition: itup.h:37
void _bt_parallel_build_main(dsm_segment *seg, shm_toc *toc)
Definition: nbtsort.c:1790
#define Min(x, y)
Definition: c.h:986
bool havedead
Definition: nbtsort.c:209
Snapshot snapshot
Definition: nbtsort.c:195
Pointer Item
Definition: item.h:17
#define P_NONE
Definition: nbtree.h:211
Sharedsort * sharedsort
Definition: tuplesort.h:55
bool isunique
Definition: nbtsort.c:108
#define INDEX_SIZE_MASK
Definition: itup.h:65
bool allequalimage
Definition: nbtree.h:785
#define PageAddItem(page, item, size, offsetNumber, overwrite, is_heap)
Definition: bufpage.h:416
static void _bt_slideleft(Page rightmostpage)
Definition: nbtsort.c:730
#define ParallelTableScanFromBTShared(shared)
Definition: nbtsort.c:163
bool isconcurrent
Definition: nbtsort.c:109
void _bt_initmetapage(Page page, BlockNumber rootbknum, uint32 level, bool allequalimage)
Definition: nbtpage.c:69
#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:423
static void _bt_end_parallel(BTLeader *btleader)
Definition: nbtsort.c:1658
Tuplesortstate * sortstate
Definition: nbtsort.c:88
IndexInfo * BuildIndexInfo(Relation index)
Definition: index.c:2378
static void _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2)
Definition: nbtsort.c:538
BlockNumber btps_blkno
Definition: nbtsort.c:235
void tuplesort_initialize_shared(Sharedsort *shared, int nWorkers, dsm_segment *seg)
Definition: tuplesort.c:4578
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:570
bool btws_use_wal
Definition: nbtsort.c:252
BTSpool * spool
Definition: nbtsort.c:211
void InstrEndParallelQuery(BufferUsage *bufusage, WalUsage *walusage)
Definition: instrument.c:208
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:250
#define PageGetMaxOffsetNumber(page)
Definition: bufpage.h:357
ParallelContext * pcxt
Definition: nbtsort.c:172
BTPageOpaqueData * BTPageOpaque
Definition: nbtree.h:71
Size PageGetFreeSpace(Page page)
Definition: bufpage.c:900
Relation heap
Definition: nbtsort.c:89
signed int int32
Definition: c.h:429
struct BTBuildState BTBuildState
void pgstat_progress_update_multi_param(int nparam, const int *index, const int64 *val)
uint16 OffsetNumber
Definition: off.h:24
Definition: type.h:89
static void BTreeTupleSetNAtts(IndexTuple itup, uint16 nkeyatts, bool heaptid)
Definition: nbtree.h:588
IndexTuple _bt_truncate(Relation rel, IndexTuple lastleft, IndexTuple firstright, BTScanInsert itup_key)
Definition: nbtutils.c:2200
void ResetUsage(void)
Definition: postgres.c:4788
void WaitForParallelWorkersToFinish(ParallelContext *pcxt)
Definition: parallel.c:751
#define SpinLockAcquire(lock)
Definition: spin.h:62
void ConditionVariableInit(ConditionVariable *cv)
ItemPointer htids
Definition: nbtree.h:876
void DestroyParallelContext(ParallelContext *pcxt)
Definition: parallel.c:905
Page btps_page
Definition: nbtsort.c:234
struct BTSpool BTSpool
#define ItemIdGetLength(itemId)
Definition: itemid.h:59
void pfree(void *pointer)
Definition: mcxt.c:1169
#define PARALLEL_KEY_BUFFER_USAGE
Definition: nbtsort.c:73
Size phystupsize
Definition: nbtree.h:879
void ConditionVariableCancelSleep(void)
#define BTREE_NONLEAF_FILLFACTOR
Definition: nbtree.h:200
static int compare(const void *arg1, const void *arg2)
Definition: geqo_pool.c:145
Size btps_full
Definition: nbtsort.c:240
#define ERROR
Definition: elog.h:46
static double _bt_parallel_heapscan(BTBuildState *buildstate, bool *brokenhotchain)
Definition: nbtsort.c:1704
IndexBuildResult * btbuild(Relation heap, Relation index, IndexInfo *indexInfo)
Definition: nbtsort.c:298
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:1745
bool isunique
Definition: nbtsort.c:91
MemoryContext ssup_cxt
Definition: sortsupport.h:66
BTCycleId btpo_cycleid
Definition: nbtree.h:68
void ConditionVariableSignal(ConditionVariable *cv)
#define PROGRESS_BTREE_PHASE_PERFORMSORT_2
Definition: nbtree.h:1114
void ExitParallelMode(void)
Definition: xact.c:992
Oid heaprelid
Definition: nbtsort.c:106
BufferUsage * bufferusage
Definition: nbtsort.c:197
bool PageIndexTupleOverwrite(Page page, OffsetNumber offnum, Item newtup, Size newsize)
Definition: bufpage.c:1398
BlockNumber btpo_prev
Definition: nbtree.h:64
IndexTuple CopyIndexTuple(IndexTuple source)
Definition: indextuple.c:528
uint32 btpo_level
Definition: nbtree.h:66
OffsetNumber btps_lastoff
Definition: nbtsort.c:237
static void _bt_sortaddtup(Page page, Size itemsize, IndexTuple itup, OffsetNumber itup_off, bool newfirstdataitem)
Definition: nbtsort.c:761
bool _bt_dedup_save_htid(BTDedupState state, IndexTuple itup)
Definition: nbtdedup.c:474
void smgrwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, char *buffer, bool skipFsync)
Definition: smgr.c:523
BTSpool * spool2
Definition: nbtsort.c:217
#define RowExclusiveLock
Definition: lockdefs.h:38
IndexTupleData * IndexTuple
Definition: itup.h:53
struct BTShared BTShared
#define PARALLEL_KEY_QUERY_TEXT
Definition: nbtsort.c:71
#define P_FIRSTKEY
Definition: nbtree.h:368
int ParallelWorkerNumber
Definition: parallel.c:112
double indtuples
Definition: nbtsort.c:218
bool isunique
Definition: nbtsort.c:208
#define RelationGetRelationName(relation)
Definition: rel.h:511
#define P_LEFTMOST(opaque)
Definition: nbtree.h:217
#define PARALLEL_KEY_TUPLESORT_SPOOL2
Definition: nbtsort.c:70
bool ii_BrokenHotChain
Definition: execnodes.h:175
unsigned int uint32
Definition: c.h:441
struct ItemIdData ItemIdData
MemoryContext CurrentMemoryContext
Definition: mcxt.c:42
static void BTreeTupleSetDownLink(IndexTuple pivot, BlockNumber blkno)
Definition: nbtree.h:555
#define IndexRelationGetNumberOfKeyAttributes(relation)
Definition: rel.h:496
int nworkers_launched
Definition: parallel.h:38
void PrepareSortSupportFromIndexRel(Relation indexRel, int16 strategy, SortSupport ssup)
Definition: sortsupport.c:162
bool brokenhotchain
Definition: nbtsort.c:148
void LaunchParallelWorkers(ParallelContext *pcxt)
Definition: parallel.c:528
void InstrAccumParallelQuery(BufferUsage *bufusage, WalUsage *walusage)
Definition: instrument.c:218
#define BTREE_METAPAGE
Definition: nbtree.h:146
void UnregisterSnapshot(Snapshot snapshot)
Definition: snapmgr.c:852
Size btps_lastextra
Definition: nbtsort.c:238
Relation index
Definition: nbtsort.c:90
Relation index
Definition: nbtsort.c:250
TableScanDesc table_beginscan_parallel(Relation relation, ParallelTableScanDesc parallel_scan)
Definition: tableam.c:175
const char * debug_query_string
Definition: postgres.c:89
void InitializeParallelDSM(ParallelContext *pcxt)
Definition: parallel.c:202
#define PageGetItemId(page, offsetNumber)
Definition: bufpage.h:235
#define SK_BT_NULLS_FIRST
Definition: nbtree.h:1083
int progress
Definition: pgbench.c:272
void InstrStartParallelQuery(void)
Definition: instrument.c:200
#define SpinLockRelease(lock)
Definition: spin.h:64
Size mul_size(Size s1, Size s2)
Definition: shmem.c:519
struct BTLeader BTLeader
BlockNumber btws_pages_written
Definition: nbtsort.c:254
void * palloc0(Size size)
Definition: mcxt.c:1093
struct SortCoordinateData * SortCoordinate
Definition: tuplesort.h:58
#define PROGRESS_SCAN_BLOCKS_DONE
Definition: progress.h:120
uint32 btps_level
Definition: nbtsort.c:239
uintptr_t Datum
Definition: postgres.h:411
struct IndexTupleData IndexTupleData
static BTPageState * _bt_pagestate(BTWriteState *wstate, uint32 level)
Definition: nbtsort.c:693
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:502
static double _bt_spools_heapscan(Relation heap, Relation index, BTBuildState *buildstate, IndexInfo *indexInfo)
Definition: nbtsort.c:367
bool log_btree_build_stats
Definition: guc.c:582
BTLeader * btleader
Definition: nbtsort.c:225
int work_mem
Definition: globals.c:124
double reltuples
Definition: nbtsort.c:145
AttrNumber ssup_attno
Definition: sortsupport.h:81
#define RelationGetNumberOfBlocks(reln)
Definition: bufmgr.h:213
double indtuples
Definition: nbtsort.c:147
static void _bt_parallel_scan_and_sort(BTSpool *btspool, BTSpool *btspool2, BTShared *btshared, Sharedsort *sharedsort, Sharedsort *sharedsort2, int sortmem, bool progress)
Definition: nbtsort.c:1904
#define IsMVCCSnapshot(snapshot)
Definition: snapmgr.h:96
struct BTPageState * btps_next
Definition: nbtsort.c:241
BTScanInsert inskey
Definition: nbtsort.c:251
#define InvalidOffsetNumber
Definition: off.h:26
IndexTuple btps_lowkey
Definition: nbtsort.c:236
void tuplesort_putindextuplevalues(Tuplesortstate *state, Relation rel, ItemPointer self, Datum *values, bool *isnull)
Definition: tuplesort.c:1727
slock_t mutex
Definition: nbtsort.c:126
int maintenance_work_mem
Definition: globals.c:126
static void _bt_buildadd(BTWriteState *wstate, BTPageState *state, IndexTuple itup, Size truncextra)
Definition: nbtsort.c:831
Size table_parallelscan_estimate(Relation rel, Snapshot snapshot)
Definition: tableam.c:140
WalUsage * walusage
Definition: nbtsort.c:196
bool havedead
Definition: nbtsort.c:146
RelFileNode rd_node
Definition: rel.h:56
bool ii_Unique
Definition: execnodes.h:172
void tuplesort_attach_shared(Sharedsort *shared, dsm_segment *seg)
Definition: tuplesort.c:4601
static uint32 BTreeTupleGetPostingOffset(IndexTuple posting)
Definition: nbtree.h:522
#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:804
void pgstat_progress_update_param(int index, int64 val)
#define SK_BT_DESC
Definition: nbtree.h:1082
Definition: regguts.h:317
int nparticipanttuplesorts
Definition: nbtsort.c:180
static SMgrRelation RelationGetSmgr(Relation rel)
Definition: rel.h:544
#define OffsetNumberNext(offsetNumber)
Definition: off.h:52
int ii_ParallelWorkers
Definition: execnodes.h:176
size_t Size
Definition: c.h:540
#define PageGetSpecialPointer(page)
Definition: bufpage.h:326
static Size _bt_parallel_estimate_shared(Relation heap, Snapshot snapshot)
Definition: nbtsort.c:1684
#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:1532
#define MAXALIGN(LEN)
Definition: c.h:757
#define index_getattr(tup, attnum, tupleDesc, isnull)
Definition: itup.h:100
void EnterParallelMode(void)
Definition: xact.c:979
#define PROGRESS_SCAN_BLOCKS_TOTAL
Definition: progress.h:119
Size basetupsize
Definition: nbtree.h:873
ScanKeyData scankeys[INDEX_MAX_KEYS]
Definition: nbtree.h:791
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:1738
Size tuplesort_estimate_shared(int nWorkers)
Definition: tuplesort.c:4557
#define RelationNeedsWAL(relation)
Definition: rel.h:601
void _bt_check_third_page(Relation rel, Relation heap, bool needheaptidspace, Page page, IndexTuple newtup)
Definition: nbtutils.c:2634
bool ii_Concurrent
Definition: execnodes.h:174
#define SnapshotAny
Definition: snapmgr.h:67
#define PROGRESS_BTREE_PHASE_LEAF_LOAD
Definition: nbtree.h:1115
static void _bt_spooldestroy(BTSpool *btspool)
Definition: nbtsort.c:517
Size maxpostingsize
Definition: nbtree.h:868
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:162
#define P_HIKEY
Definition: nbtree.h:367
static Datum values[MAXATTR]
Definition: bootstrap.c:166
struct BTPageState BTPageState
#define AccessExclusiveLock
Definition: lockdefs.h:45
#define PARALLEL_KEY_WAL_USAGE
Definition: nbtsort.c:72
int _bt_keep_natts_fast(Relation rel, IndexTuple lastleft, IndexTuple firstright)
Definition: nbtutils.c:2419
void shm_toc_insert(shm_toc *toc, uint64 key, void *address)
Definition: shm_toc.c:171
void * palloc(Size size)
Definition: mcxt.c:1062
void pgstat_report_activity(BackendState state, const char *cmd_str)
#define PROGRESS_CREATEIDX_TUPLES_DONE
Definition: progress.h:85
bool deduplicate
Definition: nbtree.h:866
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:1048
#define elog(elevel,...)
Definition: elog.h:232
#define ShareLock
Definition: lockdefs.h:41
int i
static void _bt_spool(BTSpool *btspool, ItemPointer self, Datum *values, bool *isnull)
Definition: nbtsort.c:527
void _bt_pageinit(Page page, Size size)
Definition: nbtpage.c:1142
#define BUFFERALIGN(LEN)
Definition: c.h:759
#define unlikely(x)
Definition: c.h:273
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:120
#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:1066
static void _bt_uppershutdown(BTWriteState *wstate, BTPageState *state)
Definition: nbtsort.c:1106
ConditionVariable workersdonecv
Definition: nbtsort.c:118
unsigned short t_info
Definition: itup.h:49
BTDedupStateData * BTDedupState
Definition: nbtree.h:891
Relation heap
Definition: nbtsort.c:249
#define ItemIdSetUnused(itemId)
Definition: itemid.h:128
#define PARALLEL_KEY_BTREE_SHARED
Definition: nbtsort.c:68
void table_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan, Snapshot snapshot)
Definition: tableam.c:155
void tuplesort_end(Tuplesortstate *state)
Definition: tuplesort.c:1464
#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:2692
static int ApplySortComparator(Datum datum1, bool isNull1, Datum datum2, bool isNull2, SortSupport ssup)
Definition: sortsupport.h:200
uint16 btpo_flags
Definition: nbtree.h:67
void WaitForParallelWorkersToAttach(ParallelContext *pcxt)
Definition: parallel.c:648
void smgrimmedsync(SMgrRelation reln, ForkNumber forknum)
Definition: smgr.c:660
#define RelationGetRelid(relation)
Definition: rel.h:477
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:769