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  * Since the index will never be used unless it is completely built,
35  * from a crash-recovery point of view there is no need to WAL-log the
36  * steps of the build. After completing the index build, we can just sync
37  * the whole file to disk using smgrimmedsync() before exiting this module.
38  * This can be seen to be sufficient for crash recovery by considering that
39  * it's effectively equivalent to what would happen if a CHECKPOINT occurred
40  * just after the index build. However, it is clearly not sufficient if the
41  * DBA is using the WAL log for PITR or replication purposes, since another
42  * machine would not be able to reconstruct the index from WAL. Therefore,
43  * we log the completed index pages to WAL if and only if WAL archiving is
44  * active.
45  *
46  * This code isn't concerned about the FSM at all. The caller is responsible
47  * for initializing that.
48  *
49  * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
50  * Portions Copyright (c) 1994, Regents of the University of California
51  *
52  * IDENTIFICATION
53  * src/backend/access/nbtree/nbtsort.c
54  *
55  *-------------------------------------------------------------------------
56  */
57 
58 #include "postgres.h"
59 
60 #include "access/nbtree.h"
61 #include "access/parallel.h"
62 #include "access/relscan.h"
63 #include "access/table.h"
64 #include "access/tableam.h"
65 #include "access/xact.h"
66 #include "access/xlog.h"
67 #include "access/xloginsert.h"
68 #include "catalog/index.h"
69 #include "commands/progress.h"
70 #include "miscadmin.h"
71 #include "pgstat.h"
72 #include "storage/smgr.h"
73 #include "tcop/tcopprot.h" /* pgrminclude ignore */
74 #include "utils/rel.h"
75 #include "utils/sortsupport.h"
76 #include "utils/tuplesort.h"
77 
78 
79 /* Magic numbers for parallel state sharing */
80 #define PARALLEL_KEY_BTREE_SHARED UINT64CONST(0xA000000000000001)
81 #define PARALLEL_KEY_TUPLESORT UINT64CONST(0xA000000000000002)
82 #define PARALLEL_KEY_TUPLESORT_SPOOL2 UINT64CONST(0xA000000000000003)
83 #define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xA000000000000004)
84 
85 /*
86  * DISABLE_LEADER_PARTICIPATION disables the leader's participation in
87  * parallel index builds. This may be useful as a debugging aid.
88 #undef DISABLE_LEADER_PARTICIPATION
89  */
90 
91 /*
92  * Status record for spooling/sorting phase. (Note we may have two of
93  * these due to the special requirements for uniqueness-checking with
94  * dead tuples.)
95  */
96 typedef struct BTSpool
97 {
98  Tuplesortstate *sortstate; /* state data for tuplesort.c */
101  bool isunique;
102 } BTSpool;
103 
104 /*
105  * Status for index builds performed in parallel. This is allocated in a
106  * dynamic shared memory segment. Note that there is a separate tuplesort TOC
107  * entry, private to tuplesort.c but allocated by this module on its behalf.
108  */
109 typedef struct BTShared
110 {
111  /*
112  * These fields are not modified during the sort. They primarily exist
113  * for the benefit of worker processes that need to create BTSpool state
114  * corresponding to that used by the leader.
115  */
118  bool isunique;
121 
122  /*
123  * workersdonecv is used to monitor the progress of workers. All parallel
124  * participants must indicate that they are done before leader can use
125  * mutable state that workers maintain during scan (and before leader can
126  * proceed to tuplesort_performsort()).
127  */
129 
130  /*
131  * mutex protects all fields before heapdesc.
132  *
133  * These fields contain status information of interest to B-Tree index
134  * builds that must work just the same when an index is built in parallel.
135  */
137 
138  /*
139  * Mutable state that is maintained by workers, and reported back to
140  * leader at end of parallel scan.
141  *
142  * nparticipantsdone is number of worker processes finished.
143  *
144  * reltuples is the total number of input heap tuples.
145  *
146  * havedead indicates if RECENTLY_DEAD tuples were encountered during
147  * build.
148  *
149  * indtuples is the total number of tuples that made it into the index.
150  *
151  * brokenhotchain indicates if any worker detected a broken HOT chain
152  * during build.
153  */
155  double reltuples;
156  bool havedead;
157  double indtuples;
159 
160  /*
161  * ParallelTableScanDescData data follows. Can't directly embed here, as
162  * implementations of the parallel table scan desc interface might need
163  * stronger alignment.
164  */
165 } BTShared;
166 
167 /*
168  * Return pointer to a BTShared's parallel table scan.
169  *
170  * c.f. shm_toc_allocate as to why BUFFERALIGN is used, rather than just
171  * MAXALIGN.
172  */
173 #define ParallelTableScanFromBTShared(shared) \
174  (ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(BTShared)))
175 
176 /*
177  * Status for leader in parallel index build.
178  */
179 typedef struct BTLeader
180 {
181  /* parallel context itself */
183 
184  /*
185  * nparticipanttuplesorts is the exact number of worker processes
186  * successfully launched, plus one leader process if it participates as a
187  * worker (only DISABLE_LEADER_PARTICIPATION builds avoid leader
188  * participating as a worker).
189  */
191 
192  /*
193  * Leader process convenience pointers to shared state (leader avoids TOC
194  * lookups).
195  *
196  * btshared is the shared state for entire build. sharedsort is the
197  * shared, tuplesort-managed state passed to each process tuplesort.
198  * sharedsort2 is the corresponding btspool2 shared state, used only when
199  * building unique indexes. snapshot is the snapshot used by the scan iff
200  * an MVCC snapshot is required.
201  */
206 } BTLeader;
207 
208 /*
209  * Working state for btbuild and its callback.
210  *
211  * When parallel CREATE INDEX is used, there is a BTBuildState for each
212  * participant.
213  */
214 typedef struct BTBuildState
215 {
216  bool isunique;
217  bool havedead;
220 
221  /*
222  * spool2 is needed only when the index is a unique index. Dead tuples are
223  * put into spool2 instead of spool in order to avoid uniqueness check.
224  */
226  double indtuples;
227 
228  /*
229  * btleader is only present when a parallel index build is performed, and
230  * only in the leader process. (Actually, only the leader has a
231  * BTBuildState. Workers have their own spool and spool2, though.)
232  */
234 } BTBuildState;
235 
236 /*
237  * Status record for a btree page being built. We have one of these
238  * for each active tree level.
239  */
240 typedef struct BTPageState
241 {
242  Page btps_page; /* workspace for page building */
243  BlockNumber btps_blkno; /* block # to write this page at */
244  IndexTuple btps_lowkey; /* page's strict lower bound pivot tuple */
245  OffsetNumber btps_lastoff; /* last item offset loaded */
246  uint32 btps_level; /* tree level (0 = leaf) */
247  Size btps_full; /* "full" if less than this much free space */
248  struct BTPageState *btps_next; /* link to parent level, if any */
249 } BTPageState;
250 
251 /*
252  * Overall status record for index writing phase.
253  */
254 typedef struct BTWriteState
255 {
258  BTScanInsert inskey; /* generic insertion scankey */
259  bool btws_use_wal; /* dump pages to WAL? */
260  BlockNumber btws_pages_alloced; /* # pages allocated */
261  BlockNumber btws_pages_written; /* # pages written out */
262  Page btws_zeropage; /* workspace for filling zeroes */
263 } BTWriteState;
264 
265 
267  BTBuildState *buildstate, IndexInfo *indexInfo);
268 static void _bt_spooldestroy(BTSpool *btspool);
269 static void _bt_spool(BTSpool *btspool, ItemPointer self,
270  Datum *values, bool *isnull);
271 static void _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2);
273  bool *isnull, bool tupleIsAlive, void *state);
274 static Page _bt_blnewpage(uint32 level);
275 static BTPageState *_bt_pagestate(BTWriteState *wstate, uint32 level);
276 static void _bt_slideleft(Page page);
277 static void _bt_sortaddtup(Page page, Size itemsize,
278  IndexTuple itup, OffsetNumber itup_off);
279 static void _bt_buildadd(BTWriteState *wstate, BTPageState *state,
280  IndexTuple itup);
281 static void _bt_uppershutdown(BTWriteState *wstate, BTPageState *state);
282 static void _bt_load(BTWriteState *wstate,
283  BTSpool *btspool, BTSpool *btspool2);
284 static void _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent,
285  int request);
286 static void _bt_end_parallel(BTLeader *btleader);
288 static double _bt_parallel_heapscan(BTBuildState *buildstate,
289  bool *brokenhotchain);
290 static void _bt_leader_participate_as_worker(BTBuildState *buildstate);
291 static void _bt_parallel_scan_and_sort(BTSpool *btspool, BTSpool *btspool2,
292  BTShared *btshared, Sharedsort *sharedsort,
293  Sharedsort *sharedsort2, int sortmem,
294  bool progress);
295 
296 
297 /*
298  * btbuild() -- build a new btree index.
299  */
302 {
303  IndexBuildResult *result;
304  BTBuildState buildstate;
305  double reltuples;
306 
307 #ifdef BTREE_BUILD_STATS
309  ResetUsage();
310 #endif /* BTREE_BUILD_STATS */
311 
312  buildstate.isunique = indexInfo->ii_Unique;
313  buildstate.havedead = false;
314  buildstate.heap = heap;
315  buildstate.spool = NULL;
316  buildstate.spool2 = NULL;
317  buildstate.indtuples = 0;
318  buildstate.btleader = NULL;
319 
320  /*
321  * We expect to be called exactly once for any index relation. If that's
322  * not the case, big trouble's what we have.
323  */
324  if (RelationGetNumberOfBlocks(index) != 0)
325  elog(ERROR, "index \"%s\" already contains data",
326  RelationGetRelationName(index));
327 
328  reltuples = _bt_spools_heapscan(heap, index, &buildstate, indexInfo);
329 
330  /*
331  * Finish the build by (1) completing the sort of the spool file, (2)
332  * inserting the sorted tuples into btree pages and (3) building the upper
333  * levels. Finally, it may also be necessary to end use of parallelism.
334  */
335  _bt_leafbuild(buildstate.spool, buildstate.spool2);
336  _bt_spooldestroy(buildstate.spool);
337  if (buildstate.spool2)
338  _bt_spooldestroy(buildstate.spool2);
339  if (buildstate.btleader)
340  _bt_end_parallel(buildstate.btleader);
341 
342  result = (IndexBuildResult *) palloc(sizeof(IndexBuildResult));
343 
344  result->heap_tuples = reltuples;
345  result->index_tuples = buildstate.indtuples;
346 
347 #ifdef BTREE_BUILD_STATS
349  {
350  ShowUsage("BTREE BUILD STATS");
351  ResetUsage();
352  }
353 #endif /* BTREE_BUILD_STATS */
354 
355  return result;
356 }
357 
358 /*
359  * Create and initialize one or two spool structures, and save them in caller's
360  * buildstate argument. May also fill-in fields within indexInfo used by index
361  * builds.
362  *
363  * Scans the heap, possibly in parallel, filling spools with IndexTuples. This
364  * routine encapsulates all aspects of managing parallelism. Caller need only
365  * call _bt_end_parallel() in parallel case after it is done with spool/spool2.
366  *
367  * Returns the total number of heap tuples scanned.
368  */
369 static double
371  IndexInfo *indexInfo)
372 {
373  BTSpool *btspool = (BTSpool *) palloc0(sizeof(BTSpool));
374  SortCoordinate coordinate = NULL;
375  double reltuples = 0;
376 
377  /*
378  * We size the sort area as maintenance_work_mem rather than work_mem to
379  * speed index creation. This should be OK since a single backend can't
380  * run multiple index creations in parallel (see also: notes on
381  * parallelism and maintenance_work_mem below).
382  */
383  btspool->heap = heap;
384  btspool->index = index;
385  btspool->isunique = indexInfo->ii_Unique;
386 
387  /* Save as primary spool */
388  buildstate->spool = btspool;
389 
390  /* Report table scan phase started */
393 
394  /* Attempt to launch parallel worker scan when required */
395  if (indexInfo->ii_ParallelWorkers > 0)
396  _bt_begin_parallel(buildstate, indexInfo->ii_Concurrent,
397  indexInfo->ii_ParallelWorkers);
398 
399  /*
400  * If parallel build requested and at least one worker process was
401  * successfully launched, set up coordination state
402  */
403  if (buildstate->btleader)
404  {
405  coordinate = (SortCoordinate) palloc0(sizeof(SortCoordinateData));
406  coordinate->isWorker = false;
407  coordinate->nParticipants =
408  buildstate->btleader->nparticipanttuplesorts;
409  coordinate->sharedsort = buildstate->btleader->sharedsort;
410  }
411 
412  /*
413  * Begin serial/leader tuplesort.
414  *
415  * In cases where parallelism is involved, the leader receives the same
416  * share of maintenance_work_mem as a serial sort (it is generally treated
417  * in the same way as a serial sort once we return). Parallel worker
418  * Tuplesortstates will have received only a fraction of
419  * maintenance_work_mem, though.
420  *
421  * We rely on the lifetime of the Leader Tuplesortstate almost not
422  * overlapping with any worker Tuplesortstate's lifetime. There may be
423  * some small overlap, but that's okay because we rely on leader
424  * Tuplesortstate only allocating a small, fixed amount of memory here.
425  * When its tuplesort_performsort() is called (by our caller), and
426  * significant amounts of memory are likely to be used, all workers must
427  * have already freed almost all memory held by their Tuplesortstates
428  * (they are about to go away completely, too). The overall effect is
429  * that maintenance_work_mem always represents an absolute high watermark
430  * on the amount of memory used by a CREATE INDEX operation, regardless of
431  * the use of parallelism or any other factor.
432  */
433  buildstate->spool->sortstate =
434  tuplesort_begin_index_btree(heap, index, buildstate->isunique,
435  maintenance_work_mem, coordinate,
436  false);
437 
438  /*
439  * If building a unique index, put dead tuples in a second spool to keep
440  * them out of the uniqueness check. We expect that the second spool (for
441  * dead tuples) won't get very full, so we give it only work_mem.
442  */
443  if (indexInfo->ii_Unique)
444  {
445  BTSpool *btspool2 = (BTSpool *) palloc0(sizeof(BTSpool));
446  SortCoordinate coordinate2 = NULL;
447 
448  /* Initialize secondary spool */
449  btspool2->heap = heap;
450  btspool2->index = index;
451  btspool2->isunique = false;
452  /* Save as secondary spool */
453  buildstate->spool2 = btspool2;
454 
455  if (buildstate->btleader)
456  {
457  /*
458  * Set up non-private state that is passed to
459  * tuplesort_begin_index_btree() about the basic high level
460  * coordination of a parallel sort.
461  */
462  coordinate2 = (SortCoordinate) palloc0(sizeof(SortCoordinateData));
463  coordinate2->isWorker = false;
464  coordinate2->nParticipants =
465  buildstate->btleader->nparticipanttuplesorts;
466  coordinate2->sharedsort = buildstate->btleader->sharedsort2;
467  }
468 
469  /*
470  * We expect that the second one (for dead tuples) won't get very
471  * full, so we give it only work_mem
472  */
473  buildstate->spool2->sortstate =
474  tuplesort_begin_index_btree(heap, index, false, work_mem,
475  coordinate2, false);
476  }
477 
478  /* Fill spool using either serial or parallel heap scan */
479  if (!buildstate->btleader)
480  reltuples = table_index_build_scan(heap, index, indexInfo, true, true,
481  _bt_build_callback, (void *) buildstate,
482  NULL);
483  else
484  reltuples = _bt_parallel_heapscan(buildstate,
485  &indexInfo->ii_BrokenHotChain);
486 
487  /*
488  * Set the progress target for the next phase. Reset the block number
489  * values set by table_index_build_scan
490  */
491  {
492  const int index[] = {
496  };
497  const int64 val[] = {
498  buildstate->indtuples,
499  0, 0
500  };
501 
502  pgstat_progress_update_multi_param(3, index, val);
503  }
504 
505  /* okay, all heap tuples are spooled */
506  if (buildstate->spool2 && !buildstate->havedead)
507  {
508  /* spool2 turns out to be unnecessary */
509  _bt_spooldestroy(buildstate->spool2);
510  buildstate->spool2 = NULL;
511  }
512 
513  return reltuples;
514 }
515 
516 /*
517  * clean up a spool structure and its substructures.
518  */
519 static void
521 {
522  tuplesort_end(btspool->sortstate);
523  pfree(btspool);
524 }
525 
526 /*
527  * spool an index entry into the sort file.
528  */
529 static void
530 _bt_spool(BTSpool *btspool, ItemPointer self, Datum *values, bool *isnull)
531 {
532  tuplesort_putindextuplevalues(btspool->sortstate, btspool->index,
533  self, values, isnull);
534 }
535 
536 /*
537  * given a spool loaded by successive calls to _bt_spool,
538  * create an entire btree.
539  */
540 static void
541 _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2)
542 {
543  BTWriteState wstate;
544 
545 #ifdef BTREE_BUILD_STATS
547  {
548  ShowUsage("BTREE BUILD (Spool) STATISTICS");
549  ResetUsage();
550  }
551 #endif /* BTREE_BUILD_STATS */
552 
556  if (btspool2)
557  {
560  tuplesort_performsort(btspool2->sortstate);
561  }
562 
563  wstate.heap = btspool->heap;
564  wstate.index = btspool->index;
565  wstate.inskey = _bt_mkscankey(wstate.index, NULL);
566 
567  /*
568  * We need to log index creation in WAL iff WAL archiving/streaming is
569  * enabled UNLESS the index isn't WAL-logged anyway.
570  */
571  wstate.btws_use_wal = XLogIsNeeded() && RelationNeedsWAL(wstate.index);
572 
573  /* reserve the metapage */
574  wstate.btws_pages_alloced = BTREE_METAPAGE + 1;
575  wstate.btws_pages_written = 0;
576  wstate.btws_zeropage = NULL; /* until needed */
577 
580  _bt_load(&wstate, btspool, btspool2);
581 }
582 
583 /*
584  * Per-tuple callback for table_index_build_scan
585  */
586 static void
588  ItemPointer tid,
589  Datum *values,
590  bool *isnull,
591  bool tupleIsAlive,
592  void *state)
593 {
594  BTBuildState *buildstate = (BTBuildState *) state;
595 
596  /*
597  * insert the index tuple into the appropriate spool file for subsequent
598  * processing
599  */
600  if (tupleIsAlive || buildstate->spool2 == NULL)
601  _bt_spool(buildstate->spool, tid, values, isnull);
602  else
603  {
604  /* dead tuples are put into spool2 */
605  buildstate->havedead = true;
606  _bt_spool(buildstate->spool2, tid, values, isnull);
607  }
608 
609  buildstate->indtuples += 1;
610 }
611 
612 /*
613  * allocate workspace for a new, clean btree page, not linked to any siblings.
614  */
615 static Page
617 {
618  Page page;
619  BTPageOpaque opaque;
620 
621  page = (Page) palloc(BLCKSZ);
622 
623  /* Zero the page and set up standard page header info */
624  _bt_pageinit(page, BLCKSZ);
625 
626  /* Initialize BT opaque state */
627  opaque = (BTPageOpaque) PageGetSpecialPointer(page);
628  opaque->btpo_prev = opaque->btpo_next = P_NONE;
629  opaque->btpo.level = level;
630  opaque->btpo_flags = (level > 0) ? 0 : BTP_LEAF;
631  opaque->btpo_cycleid = 0;
632 
633  /* Make the P_HIKEY line pointer appear allocated */
634  ((PageHeader) page)->pd_lower += sizeof(ItemIdData);
635 
636  return page;
637 }
638 
639 /*
640  * emit a completed btree page, and release the working storage.
641  */
642 static void
644 {
645  /* Ensure rd_smgr is open (could have been closed by relcache flush!) */
646  RelationOpenSmgr(wstate->index);
647 
648  /* XLOG stuff */
649  if (wstate->btws_use_wal)
650  {
651  /* We use the XLOG_FPI record type for this */
652  log_newpage(&wstate->index->rd_node, MAIN_FORKNUM, blkno, page, true);
653  }
654 
655  /*
656  * If we have to write pages nonsequentially, fill in the space with
657  * zeroes until we come back and overwrite. This is not logically
658  * necessary on standard Unix filesystems (unwritten space will read as
659  * zeroes anyway), but it should help to avoid fragmentation. The dummy
660  * pages aren't WAL-logged though.
661  */
662  while (blkno > wstate->btws_pages_written)
663  {
664  if (!wstate->btws_zeropage)
665  wstate->btws_zeropage = (Page) palloc0(BLCKSZ);
666  /* don't set checksum for all-zero page */
668  wstate->btws_pages_written++,
669  (char *) wstate->btws_zeropage,
670  true);
671  }
672 
673  PageSetChecksumInplace(page, blkno);
674 
675  /*
676  * Now write the page. There's no need for smgr to schedule an fsync for
677  * this write; we'll do it ourselves before ending the build.
678  */
679  if (blkno == wstate->btws_pages_written)
680  {
681  /* extending the file... */
682  smgrextend(wstate->index->rd_smgr, MAIN_FORKNUM, blkno,
683  (char *) page, true);
684  wstate->btws_pages_written++;
685  }
686  else
687  {
688  /* overwriting a block we zero-filled before */
689  smgrwrite(wstate->index->rd_smgr, MAIN_FORKNUM, blkno,
690  (char *) page, true);
691  }
692 
693  pfree(page);
694 }
695 
696 /*
697  * allocate and initialize a new BTPageState. the returned structure
698  * is suitable for immediate use by _bt_buildadd.
699  */
700 static BTPageState *
702 {
704 
705  /* create initial page for level */
706  state->btps_page = _bt_blnewpage(level);
707 
708  /* and assign it a page position */
709  state->btps_blkno = wstate->btws_pages_alloced++;
710 
711  state->btps_lowkey = NULL;
712  /* initialize lastoff so first item goes into P_FIRSTKEY */
713  state->btps_lastoff = P_HIKEY;
714  state->btps_level = level;
715  /* set "full" threshold based on level. See notes at head of file. */
716  if (level > 0)
717  state->btps_full = (BLCKSZ * (100 - BTREE_NONLEAF_FILLFACTOR) / 100);
718  else
719  state->btps_full = BTGetTargetPageFreeSpace(wstate->index);
720 
721  /* no parent level, yet */
722  state->btps_next = NULL;
723 
724  return state;
725 }
726 
727 /*
728  * slide an array of ItemIds back one slot (from P_FIRSTKEY to
729  * P_HIKEY, overwriting P_HIKEY). we need to do this when we discover
730  * that we have built an ItemId array in what has turned out to be a
731  * P_RIGHTMOST page.
732  */
733 static void
735 {
736  OffsetNumber off;
737  OffsetNumber maxoff;
738  ItemId previi;
739  ItemId thisii;
740 
741  if (!PageIsEmpty(page))
742  {
743  maxoff = PageGetMaxOffsetNumber(page);
744  previi = PageGetItemId(page, P_HIKEY);
745  for (off = P_FIRSTKEY; off <= maxoff; off = OffsetNumberNext(off))
746  {
747  thisii = PageGetItemId(page, off);
748  *previi = *thisii;
749  previi = thisii;
750  }
751  ((PageHeader) page)->pd_lower -= sizeof(ItemIdData);
752  }
753 }
754 
755 /*
756  * Add an item to a page being built.
757  *
758  * The main difference between this routine and a bare PageAddItem call
759  * is that this code knows that the leftmost data item on a non-leaf btree
760  * page has a key that must be treated as minus infinity. Therefore, it
761  * truncates away all attributes.
762  *
763  * This is almost like nbtinsert.c's _bt_pgaddtup(), but we can't use
764  * that because it assumes that P_RIGHTMOST() will return the correct
765  * answer for the page. Here, we don't know yet if the page will be
766  * rightmost. Offset P_FIRSTKEY is always the first data key.
767  */
768 static void
770  Size itemsize,
771  IndexTuple itup,
772  OffsetNumber itup_off)
773 {
775  IndexTupleData trunctuple;
776 
777  if (!P_ISLEAF(opaque) && itup_off == P_FIRSTKEY)
778  {
779  trunctuple = *itup;
780  trunctuple.t_info = sizeof(IndexTupleData);
781  BTreeTupleSetNAtts(&trunctuple, 0);
782  itup = &trunctuple;
783  itemsize = sizeof(IndexTupleData);
784  }
785 
786  if (PageAddItem(page, (Item) itup, itemsize, itup_off,
787  false, false) == InvalidOffsetNumber)
788  elog(ERROR, "failed to add item to the index page");
789 }
790 
791 /*----------
792  * Add an item to a disk page from the sort output.
793  *
794  * We must be careful to observe the page layout conventions of nbtsearch.c:
795  * - rightmost pages start data items at P_HIKEY instead of at P_FIRSTKEY.
796  * - on non-leaf pages, the key portion of the first item need not be
797  * stored, we should store only the link.
798  *
799  * A leaf page being built looks like:
800  *
801  * +----------------+---------------------------------+
802  * | PageHeaderData | linp0 linp1 linp2 ... |
803  * +-----------+----+---------------------------------+
804  * | ... linpN | |
805  * +-----------+--------------------------------------+
806  * | ^ last |
807  * | |
808  * +-------------+------------------------------------+
809  * | | itemN ... |
810  * +-------------+------------------+-----------------+
811  * | ... item3 item2 item1 | "special space" |
812  * +--------------------------------+-----------------+
813  *
814  * Contrast this with the diagram in bufpage.h; note the mismatch
815  * between linps and items. This is because we reserve linp0 as a
816  * placeholder for the pointer to the "high key" item; when we have
817  * filled up the page, we will set linp0 to point to itemN and clear
818  * linpN. On the other hand, if we find this is the last (rightmost)
819  * page, we leave the items alone and slide the linp array over. If
820  * the high key is to be truncated, offset 1 is deleted, and we insert
821  * the truncated high key at offset 1.
822  *
823  * 'last' pointer indicates the last offset added to the page.
824  *----------
825  */
826 static void
828 {
829  Page npage;
830  BlockNumber nblkno;
831  OffsetNumber last_off;
832  Size pgspc;
833  Size itupsz;
834  bool isleaf;
835 
836  /*
837  * This is a handy place to check for cancel interrupts during the btree
838  * load phase of index creation.
839  */
841 
842  npage = state->btps_page;
843  nblkno = state->btps_blkno;
844  last_off = state->btps_lastoff;
845 
846  pgspc = PageGetFreeSpace(npage);
847  itupsz = IndexTupleSize(itup);
848  itupsz = MAXALIGN(itupsz);
849  /* Leaf case has slightly different rules due to suffix truncation */
850  isleaf = (state->btps_level == 0);
851 
852  /*
853  * Check whether the new item can fit on a btree page on current level at
854  * all.
855  *
856  * Every newly built index will treat heap TID as part of the keyspace,
857  * which imposes the requirement that new high keys must occasionally have
858  * a heap TID appended within _bt_truncate(). That may leave a new pivot
859  * tuple one or two MAXALIGN() quantums larger than the original first
860  * right tuple it's derived from. v4 deals with the problem by decreasing
861  * the limit on the size of tuples inserted on the leaf level by the same
862  * small amount. Enforce the new v4+ limit on the leaf level, and the old
863  * limit on internal levels, since pivot tuples may need to make use of
864  * the reserved space. This should never fail on internal pages.
865  */
866  if (unlikely(itupsz > BTMaxItemSize(npage)))
867  _bt_check_third_page(wstate->index, wstate->heap, isleaf, npage,
868  itup);
869 
870  /*
871  * Check to see if current page will fit new item, with space left over to
872  * append a heap TID during suffix truncation when page is a leaf page.
873  *
874  * It is guaranteed that we can fit at least 2 non-pivot tuples plus a
875  * high key with heap TID when finishing off a leaf page, since we rely on
876  * _bt_check_third_page() rejecting oversized non-pivot tuples. On
877  * internal pages we can always fit 3 pivot tuples with larger internal
878  * page tuple limit (includes page high key).
879  *
880  * Most of the time, a page is only "full" in the sense that the soft
881  * fillfactor-wise limit has been exceeded. However, we must always leave
882  * at least two items plus a high key on each page before starting a new
883  * page. Disregard fillfactor and insert on "full" current page if we
884  * don't have the minimum number of items yet. (Note that we deliberately
885  * assume that suffix truncation neither enlarges nor shrinks new high key
886  * when applying soft limit.)
887  */
888  if (pgspc < itupsz + (isleaf ? MAXALIGN(sizeof(ItemPointerData)) : 0) ||
889  (pgspc < state->btps_full && last_off > P_FIRSTKEY))
890  {
891  /*
892  * Finish off the page and write it out.
893  */
894  Page opage = npage;
895  BlockNumber oblkno = nblkno;
896  ItemId ii;
897  ItemId hii;
898  IndexTuple oitup;
899 
900  /* Create new page of same level */
901  npage = _bt_blnewpage(state->btps_level);
902 
903  /* and assign it a page position */
904  nblkno = wstate->btws_pages_alloced++;
905 
906  /*
907  * We copy the last item on the page into the new page, and then
908  * rearrange the old page so that the 'last item' becomes its high key
909  * rather than a true data item. There had better be at least two
910  * items on the page already, else the page would be empty of useful
911  * data.
912  */
913  Assert(last_off > P_FIRSTKEY);
914  ii = PageGetItemId(opage, last_off);
915  oitup = (IndexTuple) PageGetItem(opage, ii);
916  _bt_sortaddtup(npage, ItemIdGetLength(ii), oitup, P_FIRSTKEY);
917 
918  /*
919  * Move 'last' into the high key position on opage. _bt_blnewpage()
920  * allocated empty space for a line pointer when opage was first
921  * created, so this is a matter of rearranging already-allocated space
922  * on page, and initializing high key line pointer. (Actually, leaf
923  * pages must also swap oitup with a truncated version of oitup, which
924  * is sometimes larger than oitup, though never by more than the space
925  * needed to append a heap TID.)
926  */
927  hii = PageGetItemId(opage, P_HIKEY);
928  *hii = *ii;
929  ItemIdSetUnused(ii); /* redundant */
930  ((PageHeader) opage)->pd_lower -= sizeof(ItemIdData);
931 
932  if (isleaf)
933  {
934  IndexTuple lastleft;
935  IndexTuple truncated;
936 
937  /*
938  * Truncate away any unneeded attributes from high key on leaf
939  * level. This is only done at the leaf level because downlinks
940  * in internal pages are either negative infinity items, or get
941  * their contents from copying from one level down. See also:
942  * _bt_split().
943  *
944  * We don't try to bias our choice of split point to make it more
945  * likely that _bt_truncate() can truncate away more attributes,
946  * whereas the split point used within _bt_split() is chosen much
947  * more delicately. Suffix truncation is mostly useful because it
948  * improves space utilization for workloads with random
949  * insertions. It doesn't seem worthwhile to add logic for
950  * choosing a split point here for a benefit that is bound to be
951  * much smaller.
952  *
953  * Overwrite the old item with new truncated high key directly.
954  * oitup is already located at the physical beginning of tuple
955  * space, so this should directly reuse the existing tuple space.
956  */
957  ii = PageGetItemId(opage, OffsetNumberPrev(last_off));
958  lastleft = (IndexTuple) PageGetItem(opage, ii);
959 
960  truncated = _bt_truncate(wstate->index, lastleft, oitup,
961  wstate->inskey);
962  if (!PageIndexTupleOverwrite(opage, P_HIKEY, (Item) truncated,
963  IndexTupleSize(truncated)))
964  elog(ERROR, "failed to add high key to the index page");
965  pfree(truncated);
966 
967  /* oitup should continue to point to the page's high key */
968  hii = PageGetItemId(opage, P_HIKEY);
969  oitup = (IndexTuple) PageGetItem(opage, hii);
970  }
971 
972  /*
973  * Link the old page into its parent, using its low key. If we don't
974  * have a parent, we have to create one; this adds a new btree level.
975  */
976  if (state->btps_next == NULL)
977  state->btps_next = _bt_pagestate(wstate, state->btps_level + 1);
978 
979  Assert((BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) <=
981  BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) > 0) ||
983  Assert(BTreeTupleGetNAtts(state->btps_lowkey, wstate->index) == 0 ||
985  BTreeTupleSetDownLink(state->btps_lowkey, oblkno);
986  _bt_buildadd(wstate, state->btps_next, state->btps_lowkey);
987  pfree(state->btps_lowkey);
988 
989  /*
990  * Save a copy of the high key from the old page. It is also the low
991  * key for the new page.
992  */
993  state->btps_lowkey = CopyIndexTuple(oitup);
994 
995  /*
996  * Set the sibling links for both pages.
997  */
998  {
1000  BTPageOpaque nopaque = (BTPageOpaque) PageGetSpecialPointer(npage);
1001 
1002  oopaque->btpo_next = nblkno;
1003  nopaque->btpo_prev = oblkno;
1004  nopaque->btpo_next = P_NONE; /* redundant */
1005  }
1006 
1007  /*
1008  * Write out the old page. We never need to touch it again, so we can
1009  * free the opage workspace too.
1010  */
1011  _bt_blwritepage(wstate, opage, oblkno);
1012 
1013  /*
1014  * Reset last_off to point to new page
1015  */
1016  last_off = P_FIRSTKEY;
1017  }
1018 
1019  /*
1020  * By here, either original page is still the current page, or a new page
1021  * was created that became the current page. Either way, the current page
1022  * definitely has space for new item.
1023  *
1024  * If the new item is the first for its page, it must also be the first
1025  * item on its entire level. On later same-level pages, a low key for a
1026  * page will be copied from the prior page in the code above. Generate a
1027  * minus infinity low key here instead.
1028  */
1029  if (last_off == P_HIKEY)
1030  {
1031  Assert(state->btps_lowkey == NULL);
1032  state->btps_lowkey = palloc0(sizeof(IndexTupleData));
1033  state->btps_lowkey->t_info = sizeof(IndexTupleData);
1034  BTreeTupleSetNAtts(state->btps_lowkey, 0);
1035  }
1036 
1037  /*
1038  * Add the new item into the current page.
1039  */
1040  last_off = OffsetNumberNext(last_off);
1041  _bt_sortaddtup(npage, itupsz, itup, last_off);
1042 
1043  state->btps_page = npage;
1044  state->btps_blkno = nblkno;
1045  state->btps_lastoff = last_off;
1046 }
1047 
1048 /*
1049  * Finish writing out the completed btree.
1050  */
1051 static void
1053 {
1054  BTPageState *s;
1055  BlockNumber rootblkno = P_NONE;
1056  uint32 rootlevel = 0;
1057  Page metapage;
1058 
1059  /*
1060  * Each iteration of this loop completes one more level of the tree.
1061  */
1062  for (s = state; s != NULL; s = s->btps_next)
1063  {
1064  BlockNumber blkno;
1065  BTPageOpaque opaque;
1066 
1067  blkno = s->btps_blkno;
1069 
1070  /*
1071  * We have to link the last page on this level to somewhere.
1072  *
1073  * If we're at the top, it's the root, so attach it to the metapage.
1074  * Otherwise, add an entry for it to its parent using its low key.
1075  * This may cause the last page of the parent level to split, but
1076  * that's not a problem -- we haven't gotten to it yet.
1077  */
1078  if (s->btps_next == NULL)
1079  {
1080  opaque->btpo_flags |= BTP_ROOT;
1081  rootblkno = blkno;
1082  rootlevel = s->btps_level;
1083  }
1084  else
1085  {
1086  Assert((BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) <=
1088  BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) > 0) ||
1089  P_LEFTMOST(opaque));
1090  Assert(BTreeTupleGetNAtts(s->btps_lowkey, wstate->index) == 0 ||
1091  !P_LEFTMOST(opaque));
1093  _bt_buildadd(wstate, s->btps_next, s->btps_lowkey);
1094  pfree(s->btps_lowkey);
1095  s->btps_lowkey = NULL;
1096  }
1097 
1098  /*
1099  * This is the rightmost page, so the ItemId array needs to be slid
1100  * back one slot. Then we can dump out the page.
1101  */
1103  _bt_blwritepage(wstate, s->btps_page, s->btps_blkno);
1104  s->btps_page = NULL; /* writepage freed the workspace */
1105  }
1106 
1107  /*
1108  * As the last step in the process, construct the metapage and make it
1109  * point to the new root (unless we had no data at all, in which case it's
1110  * set to point to "P_NONE"). This changes the index to the "valid" state
1111  * by filling in a valid magic number in the metapage.
1112  */
1113  metapage = (Page) palloc(BLCKSZ);
1114  _bt_initmetapage(metapage, rootblkno, rootlevel);
1115  _bt_blwritepage(wstate, metapage, BTREE_METAPAGE);
1116 }
1117 
1118 /*
1119  * Read tuples in correct sort order from tuplesort, and load them into
1120  * btree leaves.
1121  */
1122 static void
1123 _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
1124 {
1125  BTPageState *state = NULL;
1126  bool merge = (btspool2 != NULL);
1127  IndexTuple itup,
1128  itup2 = NULL;
1129  bool load1;
1130  TupleDesc tupdes = RelationGetDescr(wstate->index);
1131  int i,
1133  SortSupport sortKeys;
1134  int64 tuples_done = 0;
1135 
1136  if (merge)
1137  {
1138  /*
1139  * Another BTSpool for dead tuples exists. Now we have to merge
1140  * btspool and btspool2.
1141  */
1142 
1143  /* the preparation of merge */
1144  itup = tuplesort_getindextuple(btspool->sortstate, true);
1145  itup2 = tuplesort_getindextuple(btspool2->sortstate, true);
1146 
1147  /* Prepare SortSupport data for each column */
1148  sortKeys = (SortSupport) palloc0(keysz * sizeof(SortSupportData));
1149 
1150  for (i = 0; i < keysz; i++)
1151  {
1152  SortSupport sortKey = sortKeys + i;
1153  ScanKey scanKey = wstate->inskey->scankeys + i;
1154  int16 strategy;
1155 
1156  sortKey->ssup_cxt = CurrentMemoryContext;
1157  sortKey->ssup_collation = scanKey->sk_collation;
1158  sortKey->ssup_nulls_first =
1159  (scanKey->sk_flags & SK_BT_NULLS_FIRST) != 0;
1160  sortKey->ssup_attno = scanKey->sk_attno;
1161  /* Abbreviation is not supported here */
1162  sortKey->abbreviate = false;
1163 
1164  AssertState(sortKey->ssup_attno != 0);
1165 
1166  strategy = (scanKey->sk_flags & SK_BT_DESC) != 0 ?
1168 
1169  PrepareSortSupportFromIndexRel(wstate->index, strategy, sortKey);
1170  }
1171 
1172  for (;;)
1173  {
1174  load1 = true; /* load BTSpool next ? */
1175  if (itup2 == NULL)
1176  {
1177  if (itup == NULL)
1178  break;
1179  }
1180  else if (itup != NULL)
1181  {
1182  int32 compare = 0;
1183 
1184  for (i = 1; i <= keysz; i++)
1185  {
1186  SortSupport entry;
1187  Datum attrDatum1,
1188  attrDatum2;
1189  bool isNull1,
1190  isNull2;
1191 
1192  entry = sortKeys + i - 1;
1193  attrDatum1 = index_getattr(itup, i, tupdes, &isNull1);
1194  attrDatum2 = index_getattr(itup2, i, tupdes, &isNull2);
1195 
1196  compare = ApplySortComparator(attrDatum1, isNull1,
1197  attrDatum2, isNull2,
1198  entry);
1199  if (compare > 0)
1200  {
1201  load1 = false;
1202  break;
1203  }
1204  else if (compare < 0)
1205  break;
1206  }
1207 
1208  /*
1209  * If key values are equal, we sort on ItemPointer. This is
1210  * required for btree indexes, since heap TID is treated as an
1211  * implicit last key attribute in order to ensure that all
1212  * keys in the index are physically unique.
1213  */
1214  if (compare == 0)
1215  {
1216  compare = ItemPointerCompare(&itup->t_tid, &itup2->t_tid);
1217  Assert(compare != 0);
1218  if (compare > 0)
1219  load1 = false;
1220  }
1221  }
1222  else
1223  load1 = false;
1224 
1225  /* When we see first tuple, create first index page */
1226  if (state == NULL)
1227  state = _bt_pagestate(wstate, 0);
1228 
1229  if (load1)
1230  {
1231  _bt_buildadd(wstate, state, itup);
1232  itup = tuplesort_getindextuple(btspool->sortstate, true);
1233  }
1234  else
1235  {
1236  _bt_buildadd(wstate, state, itup2);
1237  itup2 = tuplesort_getindextuple(btspool2->sortstate, true);
1238  }
1239 
1240  /* Report progress */
1242  ++tuples_done);
1243  }
1244  pfree(sortKeys);
1245  }
1246  else
1247  {
1248  /* merge is unnecessary */
1249  while ((itup = tuplesort_getindextuple(btspool->sortstate,
1250  true)) != NULL)
1251  {
1252  /* When we see first tuple, create first index page */
1253  if (state == NULL)
1254  state = _bt_pagestate(wstate, 0);
1255 
1256  _bt_buildadd(wstate, state, itup);
1257 
1258  /* Report progress */
1260  ++tuples_done);
1261  }
1262  }
1263 
1264  /* Close down final pages and write the metapage */
1265  _bt_uppershutdown(wstate, state);
1266 
1267  /*
1268  * If the index is WAL-logged, we must fsync it down to disk before it's
1269  * safe to commit the transaction. (For a non-WAL-logged index we don't
1270  * care since the index will be uninteresting after a crash anyway.)
1271  *
1272  * It's obvious that we must do this when not WAL-logging the build. It's
1273  * less obvious that we have to do it even if we did WAL-log the index
1274  * pages. The reason is that since we're building outside shared buffers,
1275  * a CHECKPOINT occurring during the build has no way to flush the
1276  * previously written data to disk (indeed it won't know the index even
1277  * exists). A crash later on would replay WAL from the checkpoint,
1278  * therefore it wouldn't replay our earlier WAL entries. If we do not
1279  * fsync those pages here, they might still not be on disk when the crash
1280  * occurs.
1281  */
1282  if (RelationNeedsWAL(wstate->index))
1283  {
1284  RelationOpenSmgr(wstate->index);
1286  }
1287 }
1288 
1289 /*
1290  * Create parallel context, and launch workers for leader.
1291  *
1292  * buildstate argument should be initialized (with the exception of the
1293  * tuplesort state in spools, which may later be created based on shared
1294  * state initially set up here).
1295  *
1296  * isconcurrent indicates if operation is CREATE INDEX CONCURRENTLY.
1297  *
1298  * request is the target number of parallel worker processes to launch.
1299  *
1300  * Sets buildstate's BTLeader, which caller must use to shut down parallel
1301  * mode by passing it to _bt_end_parallel() at the very end of its index
1302  * build. If not even a single worker process can be launched, this is
1303  * never set, and caller should proceed with a serial index build.
1304  */
1305 static void
1306 _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent, int request)
1307 {
1308  ParallelContext *pcxt;
1309  int scantuplesortstates;
1310  Snapshot snapshot;
1311  Size estbtshared;
1312  Size estsort;
1313  BTShared *btshared;
1314  Sharedsort *sharedsort;
1315  Sharedsort *sharedsort2;
1316  BTSpool *btspool = buildstate->spool;
1317  BTLeader *btleader = (BTLeader *) palloc0(sizeof(BTLeader));
1318  bool leaderparticipates = true;
1319  char *sharedquery;
1320  int querylen;
1321 
1322 #ifdef DISABLE_LEADER_PARTICIPATION
1323  leaderparticipates = false;
1324 #endif
1325 
1326  /*
1327  * Enter parallel mode, and create context for parallel build of btree
1328  * index
1329  */
1331  Assert(request > 0);
1332  pcxt = CreateParallelContext("postgres", "_bt_parallel_build_main",
1333  request);
1334 
1335  scantuplesortstates = leaderparticipates ? request + 1 : request;
1336 
1337  /*
1338  * Prepare for scan of the base relation. In a normal index build, we use
1339  * SnapshotAny because we must retrieve all tuples and do our own time
1340  * qual checks (because we have to index RECENTLY_DEAD tuples). In a
1341  * concurrent build, we take a regular MVCC snapshot and index whatever's
1342  * live according to that.
1343  */
1344  if (!isconcurrent)
1345  snapshot = SnapshotAny;
1346  else
1348 
1349  /*
1350  * Estimate size for our own PARALLEL_KEY_BTREE_SHARED workspace, and
1351  * PARALLEL_KEY_TUPLESORT tuplesort workspace
1352  */
1353  estbtshared = _bt_parallel_estimate_shared(btspool->heap, snapshot);
1354  shm_toc_estimate_chunk(&pcxt->estimator, estbtshared);
1355  estsort = tuplesort_estimate_shared(scantuplesortstates);
1356  shm_toc_estimate_chunk(&pcxt->estimator, estsort);
1357 
1358  /*
1359  * Unique case requires a second spool, and so we may have to account for
1360  * another shared workspace for that -- PARALLEL_KEY_TUPLESORT_SPOOL2
1361  */
1362  if (!btspool->isunique)
1363  shm_toc_estimate_keys(&pcxt->estimator, 2);
1364  else
1365  {
1366  shm_toc_estimate_chunk(&pcxt->estimator, estsort);
1367  shm_toc_estimate_keys(&pcxt->estimator, 3);
1368  }
1369 
1370  /* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
1371  querylen = strlen(debug_query_string);
1372  shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
1373  shm_toc_estimate_keys(&pcxt->estimator, 1);
1374 
1375  /* Everyone's had a chance to ask for space, so now create the DSM */
1376  InitializeParallelDSM(pcxt);
1377 
1378  /* If no DSM segment was available, back out (do serial build) */
1379  if (pcxt->seg == NULL)
1380  {
1381  if (IsMVCCSnapshot(snapshot))
1382  UnregisterSnapshot(snapshot);
1383  DestroyParallelContext(pcxt);
1384  ExitParallelMode();
1385  return;
1386  }
1387 
1388  /* Store shared build state, for which we reserved space */
1389  btshared = (BTShared *) shm_toc_allocate(pcxt->toc, estbtshared);
1390  /* Initialize immutable state */
1391  btshared->heaprelid = RelationGetRelid(btspool->heap);
1392  btshared->indexrelid = RelationGetRelid(btspool->index);
1393  btshared->isunique = btspool->isunique;
1394  btshared->isconcurrent = isconcurrent;
1395  btshared->scantuplesortstates = scantuplesortstates;
1397  SpinLockInit(&btshared->mutex);
1398  /* Initialize mutable state */
1399  btshared->nparticipantsdone = 0;
1400  btshared->reltuples = 0.0;
1401  btshared->havedead = false;
1402  btshared->indtuples = 0.0;
1403  btshared->brokenhotchain = false;
1406  snapshot);
1407 
1408  /*
1409  * Store shared tuplesort-private state, for which we reserved space.
1410  * Then, initialize opaque state using tuplesort routine.
1411  */
1412  sharedsort = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
1413  tuplesort_initialize_shared(sharedsort, scantuplesortstates,
1414  pcxt->seg);
1415 
1416  shm_toc_insert(pcxt->toc, PARALLEL_KEY_BTREE_SHARED, btshared);
1417  shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLESORT, sharedsort);
1418 
1419  /* Unique case requires a second spool, and associated shared state */
1420  if (!btspool->isunique)
1421  sharedsort2 = NULL;
1422  else
1423  {
1424  /*
1425  * Store additional shared tuplesort-private state, for which we
1426  * reserved space. Then, initialize opaque state using tuplesort
1427  * routine.
1428  */
1429  sharedsort2 = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
1430  tuplesort_initialize_shared(sharedsort2, scantuplesortstates,
1431  pcxt->seg);
1432 
1433  shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLESORT_SPOOL2, sharedsort2);
1434  }
1435 
1436  /* Store query string for workers */
1437  sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
1438  memcpy(sharedquery, debug_query_string, querylen + 1);
1439  shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, sharedquery);
1440 
1441  /* Launch workers, saving status for leader/caller */
1442  LaunchParallelWorkers(pcxt);
1443  btleader->pcxt = pcxt;
1444  btleader->nparticipanttuplesorts = pcxt->nworkers_launched;
1445  if (leaderparticipates)
1446  btleader->nparticipanttuplesorts++;
1447  btleader->btshared = btshared;
1448  btleader->sharedsort = sharedsort;
1449  btleader->sharedsort2 = sharedsort2;
1450  btleader->snapshot = snapshot;
1451 
1452  /* If no workers were successfully launched, back out (do serial build) */
1453  if (pcxt->nworkers_launched == 0)
1454  {
1455  _bt_end_parallel(btleader);
1456  return;
1457  }
1458 
1459  /* Save leader state now that it's clear build will be parallel */
1460  buildstate->btleader = btleader;
1461 
1462  /* Join heap scan ourselves */
1463  if (leaderparticipates)
1465 
1466  /*
1467  * Caller needs to wait for all launched workers when we return. Make
1468  * sure that the failure-to-start case will not hang forever.
1469  */
1471 }
1472 
1473 /*
1474  * Shut down workers, destroy parallel context, and end parallel mode.
1475  */
1476 static void
1478 {
1479  /* Shutdown worker processes */
1481  /* Free last reference to MVCC snapshot, if one was used */
1482  if (IsMVCCSnapshot(btleader->snapshot))
1483  UnregisterSnapshot(btleader->snapshot);
1484  DestroyParallelContext(btleader->pcxt);
1485  ExitParallelMode();
1486 }
1487 
1488 /*
1489  * Returns size of shared memory required to store state for a parallel
1490  * btree index build based on the snapshot its parallel scan will use.
1491  */
1492 static Size
1494 {
1495  /* c.f. shm_toc_allocate as to why BUFFERALIGN is used */
1496  return add_size(BUFFERALIGN(sizeof(BTShared)),
1497  table_parallelscan_estimate(heap, snapshot));
1498 }
1499 
1500 /*
1501  * Within leader, wait for end of heap scan.
1502  *
1503  * When called, parallel heap scan started by _bt_begin_parallel() will
1504  * already be underway within worker processes (when leader participates
1505  * as a worker, we should end up here just as workers are finishing).
1506  *
1507  * Fills in fields needed for ambuild statistics, and lets caller set
1508  * field indicating that some worker encountered a broken HOT chain.
1509  *
1510  * Returns the total number of heap tuples scanned.
1511  */
1512 static double
1513 _bt_parallel_heapscan(BTBuildState *buildstate, bool *brokenhotchain)
1514 {
1515  BTShared *btshared = buildstate->btleader->btshared;
1516  int nparticipanttuplesorts;
1517  double reltuples;
1518 
1519  nparticipanttuplesorts = buildstate->btleader->nparticipanttuplesorts;
1520  for (;;)
1521  {
1522  SpinLockAcquire(&btshared->mutex);
1523  if (btshared->nparticipantsdone == nparticipanttuplesorts)
1524  {
1525  buildstate->havedead = btshared->havedead;
1526  buildstate->indtuples = btshared->indtuples;
1527  *brokenhotchain = btshared->brokenhotchain;
1528  reltuples = btshared->reltuples;
1529  SpinLockRelease(&btshared->mutex);
1530  break;
1531  }
1532  SpinLockRelease(&btshared->mutex);
1533 
1536  }
1537 
1539 
1540  return reltuples;
1541 }
1542 
1543 /*
1544  * Within leader, participate as a parallel worker.
1545  */
1546 static void
1548 {
1549  BTLeader *btleader = buildstate->btleader;
1550  BTSpool *leaderworker;
1551  BTSpool *leaderworker2;
1552  int sortmem;
1553 
1554  /* Allocate memory and initialize private spool */
1555  leaderworker = (BTSpool *) palloc0(sizeof(BTSpool));
1556  leaderworker->heap = buildstate->spool->heap;
1557  leaderworker->index = buildstate->spool->index;
1558  leaderworker->isunique = buildstate->spool->isunique;
1559 
1560  /* Initialize second spool, if required */
1561  if (!btleader->btshared->isunique)
1562  leaderworker2 = NULL;
1563  else
1564  {
1565  /* Allocate memory for worker's own private secondary spool */
1566  leaderworker2 = (BTSpool *) palloc0(sizeof(BTSpool));
1567 
1568  /* Initialize worker's own secondary spool */
1569  leaderworker2->heap = leaderworker->heap;
1570  leaderworker2->index = leaderworker->index;
1571  leaderworker2->isunique = false;
1572  }
1573 
1574  /*
1575  * Might as well use reliable figure when doling out maintenance_work_mem
1576  * (when requested number of workers were not launched, this will be
1577  * somewhat higher than it is for other workers).
1578  */
1579  sortmem = maintenance_work_mem / btleader->nparticipanttuplesorts;
1580 
1581  /* Perform work common to all participants */
1582  _bt_parallel_scan_and_sort(leaderworker, leaderworker2, btleader->btshared,
1583  btleader->sharedsort, btleader->sharedsort2,
1584  sortmem, true);
1585 
1586 #ifdef BTREE_BUILD_STATS
1588  {
1589  ShowUsage("BTREE BUILD (Leader Partial Spool) STATISTICS");
1590  ResetUsage();
1591  }
1592 #endif /* BTREE_BUILD_STATS */
1593 }
1594 
1595 /*
1596  * Perform work within a launched parallel process.
1597  */
1598 void
1600 {
1601  char *sharedquery;
1602  BTSpool *btspool;
1603  BTSpool *btspool2;
1604  BTShared *btshared;
1605  Sharedsort *sharedsort;
1606  Sharedsort *sharedsort2;
1607  Relation heapRel;
1608  Relation indexRel;
1609  LOCKMODE heapLockmode;
1610  LOCKMODE indexLockmode;
1611  int sortmem;
1612 
1613 #ifdef BTREE_BUILD_STATS
1615  ResetUsage();
1616 #endif /* BTREE_BUILD_STATS */
1617 
1618  /* Set debug_query_string for individual workers first */
1619  sharedquery = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, false);
1620  debug_query_string = sharedquery;
1621 
1622  /* Report the query string from leader */
1624 
1625  /* Look up nbtree shared state */
1626  btshared = shm_toc_lookup(toc, PARALLEL_KEY_BTREE_SHARED, false);
1627 
1628  /* Open relations using lock modes known to be obtained by index.c */
1629  if (!btshared->isconcurrent)
1630  {
1631  heapLockmode = ShareLock;
1632  indexLockmode = AccessExclusiveLock;
1633  }
1634  else
1635  {
1636  heapLockmode = ShareUpdateExclusiveLock;
1637  indexLockmode = RowExclusiveLock;
1638  }
1639 
1640  /* Open relations within worker */
1641  heapRel = table_open(btshared->heaprelid, heapLockmode);
1642  indexRel = index_open(btshared->indexrelid, indexLockmode);
1643 
1644  /* Initialize worker's own spool */
1645  btspool = (BTSpool *) palloc0(sizeof(BTSpool));
1646  btspool->heap = heapRel;
1647  btspool->index = indexRel;
1648  btspool->isunique = btshared->isunique;
1649 
1650  /* Look up shared state private to tuplesort.c */
1651  sharedsort = shm_toc_lookup(toc, PARALLEL_KEY_TUPLESORT, false);
1652  tuplesort_attach_shared(sharedsort, seg);
1653  if (!btshared->isunique)
1654  {
1655  btspool2 = NULL;
1656  sharedsort2 = NULL;
1657  }
1658  else
1659  {
1660  /* Allocate memory for worker's own private secondary spool */
1661  btspool2 = (BTSpool *) palloc0(sizeof(BTSpool));
1662 
1663  /* Initialize worker's own secondary spool */
1664  btspool2->heap = btspool->heap;
1665  btspool2->index = btspool->index;
1666  btspool2->isunique = false;
1667  /* Look up shared state private to tuplesort.c */
1668  sharedsort2 = shm_toc_lookup(toc, PARALLEL_KEY_TUPLESORT_SPOOL2, false);
1669  tuplesort_attach_shared(sharedsort2, seg);
1670  }
1671 
1672  /* Perform sorting of spool, and possibly a spool2 */
1673  sortmem = maintenance_work_mem / btshared->scantuplesortstates;
1674  _bt_parallel_scan_and_sort(btspool, btspool2, btshared, sharedsort,
1675  sharedsort2, sortmem, false);
1676 
1677 #ifdef BTREE_BUILD_STATS
1679  {
1680  ShowUsage("BTREE BUILD (Worker Partial Spool) STATISTICS");
1681  ResetUsage();
1682  }
1683 #endif /* BTREE_BUILD_STATS */
1684 
1685  index_close(indexRel, indexLockmode);
1686  table_close(heapRel, heapLockmode);
1687 }
1688 
1689 /*
1690  * Perform a worker's portion of a parallel sort.
1691  *
1692  * This generates a tuplesort for passed btspool, and a second tuplesort
1693  * state if a second btspool is need (i.e. for unique index builds). All
1694  * other spool fields should already be set when this is called.
1695  *
1696  * sortmem is the amount of working memory to use within each worker,
1697  * expressed in KBs.
1698  *
1699  * When this returns, workers are done, and need only release resources.
1700  */
1701 static void
1703  BTShared *btshared, Sharedsort *sharedsort,
1704  Sharedsort *sharedsort2, int sortmem, bool progress)
1705 {
1706  SortCoordinate coordinate;
1707  BTBuildState buildstate;
1708  TableScanDesc scan;
1709  double reltuples;
1710  IndexInfo *indexInfo;
1711 
1712  /* Initialize local tuplesort coordination state */
1713  coordinate = palloc0(sizeof(SortCoordinateData));
1714  coordinate->isWorker = true;
1715  coordinate->nParticipants = -1;
1716  coordinate->sharedsort = sharedsort;
1717 
1718  /* Begin "partial" tuplesort */
1719  btspool->sortstate = tuplesort_begin_index_btree(btspool->heap,
1720  btspool->index,
1721  btspool->isunique,
1722  sortmem, coordinate,
1723  false);
1724 
1725  /*
1726  * Just as with serial case, there may be a second spool. If so, a
1727  * second, dedicated spool2 partial tuplesort is required.
1728  */
1729  if (btspool2)
1730  {
1731  SortCoordinate coordinate2;
1732 
1733  /*
1734  * We expect that the second one (for dead tuples) won't get very
1735  * full, so we give it only work_mem (unless sortmem is less for
1736  * worker). Worker processes are generally permitted to allocate
1737  * work_mem independently.
1738  */
1739  coordinate2 = palloc0(sizeof(SortCoordinateData));
1740  coordinate2->isWorker = true;
1741  coordinate2->nParticipants = -1;
1742  coordinate2->sharedsort = sharedsort2;
1743  btspool2->sortstate =
1744  tuplesort_begin_index_btree(btspool->heap, btspool->index, false,
1745  Min(sortmem, work_mem), coordinate2,
1746  false);
1747  }
1748 
1749  /* Fill in buildstate for _bt_build_callback() */
1750  buildstate.isunique = btshared->isunique;
1751  buildstate.havedead = false;
1752  buildstate.heap = btspool->heap;
1753  buildstate.spool = btspool;
1754  buildstate.spool2 = btspool2;
1755  buildstate.indtuples = 0;
1756  buildstate.btleader = NULL;
1757 
1758  /* Join parallel scan */
1759  indexInfo = BuildIndexInfo(btspool->index);
1760  indexInfo->ii_Concurrent = btshared->isconcurrent;
1761  scan = table_beginscan_parallel(btspool->heap,
1762  ParallelTableScanFromBTShared(btshared));
1763  reltuples = table_index_build_scan(btspool->heap, btspool->index, indexInfo,
1764  true, progress, _bt_build_callback,
1765  (void *) &buildstate, scan);
1766 
1767  /*
1768  * Execute this worker's part of the sort.
1769  *
1770  * Unlike leader and serial cases, we cannot avoid calling
1771  * tuplesort_performsort() for spool2 if it ends up containing no dead
1772  * tuples (this is disallowed for workers by tuplesort).
1773  */
1774  tuplesort_performsort(btspool->sortstate);
1775  if (btspool2)
1776  tuplesort_performsort(btspool2->sortstate);
1777 
1778  /*
1779  * Done. Record ambuild statistics, and whether we encountered a broken
1780  * HOT chain.
1781  */
1782  SpinLockAcquire(&btshared->mutex);
1783  btshared->nparticipantsdone++;
1784  btshared->reltuples += reltuples;
1785  if (buildstate.havedead)
1786  btshared->havedead = true;
1787  btshared->indtuples += buildstate.indtuples;
1788  if (indexInfo->ii_BrokenHotChain)
1789  btshared->brokenhotchain = true;
1790  SpinLockRelease(&btshared->mutex);
1791 
1792  /* Notify leader */
1794 
1795  /* We can end tuplesorts immediately */
1796  tuplesort_end(btspool->sortstate);
1797  if (btspool2)
1798  tuplesort_end(btspool2->sortstate);
1799 }
struct SortSupportData * SortSupport
Definition: sortsupport.h:58
IndexTuple tuplesort_getindextuple(Tuplesortstate *state, bool forward)
Definition: tuplesort.c:2216
static void _bt_begin_parallel(BTBuildState *buildstate, bool isconcurrent, int request)
Definition: nbtsort.c:1306
int32 ItemPointerCompare(ItemPointer arg1, ItemPointer arg2)
Definition: itemptr.c:52
signed short int16
Definition: c.h:344
bool ssup_nulls_first
Definition: sortsupport.h:75
int slock_t
Definition: s_lock.h:934
#define BTP_ROOT
Definition: nbtree.h:73
Sharedsort * sharedsort2
Definition: nbtsort.c:204
void tuplesort_performsort(Tuplesortstate *state)
Definition: tuplesort.c:1791
BlockNumber btpo_next
Definition: nbtree.h:59
#define PROGRESS_BTREE_PHASE_INDEXBUILD_TABLESCAN
Definition: nbtree.h:705
#define PageIsEmpty(page)
Definition: bufpage.h:222
void table_close(Relation relation, LOCKMODE lockmode)
Definition: table.c:133
int scantuplesortstates
Definition: nbtsort.c:120
#define BTGreaterStrategyNumber
Definition: stratnum.h:33
struct BTWriteState BTWriteState
#define PARALLEL_KEY_TUPLESORT
Definition: nbtsort.c:81
Relation heap
Definition: nbtsort.c:218
#define AssertState(condition)
Definition: c.h:731
Page btws_zeropage
Definition: nbtsort.c:262
ParallelContext * CreateParallelContext(const char *library_name, const char *function_name, int nworkers)
Definition: parallel.c:162
BTShared * btshared
Definition: nbtsort.c:202
int nparticipantsdone
Definition: nbtsort.c:154
static void _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
Definition: nbtsort.c:1123
Snapshot RegisterSnapshot(Snapshot snapshot)
Definition: snapmgr.c:865
#define RelationGetDescr(relation)
Definition: rel.h:454
int LOCKMODE
Definition: lockdefs.h:26
BTScanInsert _bt_mkscankey(Relation rel, IndexTuple itup)
Definition: nbtutils.c:86
#define BTGetTargetPageFreeSpace(relation)
Definition: nbtree.h:697
Oid indexrelid
Definition: nbtsort.c:117
void pgstat_report_activity(BackendState state, const char *cmd_str)
Definition: pgstat.c:3114
#define PROGRESS_BTREE_PHASE_PERFORMSORT_1
Definition: nbtree.h:706
void ShowUsage(const char *title)
Definition: postgres.c:4594
dsm_segment * seg
Definition: parallel.h:43
BlockNumber btws_pages_alloced
Definition: nbtsort.c:260
struct SMgrRelationData * rd_smgr
Definition: rel.h:57
void pgstat_progress_update_param(int index, int64 val)
Definition: pgstat.c:3213
Sharedsort * sharedsort
Definition: nbtsort.c:203
shm_toc_estimator estimator
Definition: parallel.h:42
#define SpinLockInit(lock)
Definition: spin.h:60
#define BTP_LEAF
Definition: nbtree.h:72
#define XLogIsNeeded()
Definition: xlog.h:181
ItemPointerData t_tid
Definition: itup.h:37
void _bt_parallel_build_main(dsm_segment *seg, shm_toc *toc)
Definition: nbtsort.c:1599
#define Min(x, y)
Definition: c.h:910
union BTPageOpaqueData::@46 btpo
bool havedead
Definition: nbtsort.c:217
Snapshot snapshot
Definition: nbtsort.c:205
Pointer Item
Definition: item.h:17
#define P_NONE
Definition: nbtree.h:182
Sharedsort * sharedsort
Definition: tuplesort.h:55
bool isunique
Definition: nbtsort.c:118
#define PageAddItem(page, item, size, offsetNumber, overwrite, is_heap)
Definition: bufpage.h:416
#define ParallelTableScanFromBTShared(shared)
Definition: nbtsort.c:173
bool isconcurrent
Definition: nbtsort.c:119
#define PROGRESS_CREATEIDX_TUPLES_TOTAL
Definition: progress.h:84
uint32 BlockNumber
Definition: block.h:31
static void _bt_end_parallel(BTLeader *btleader)
Definition: nbtsort.c:1477
Tuplesortstate * sortstate
Definition: nbtsort.c:98
IndexInfo * BuildIndexInfo(Relation index)
Definition: index.c:2293
#define BTreeTupleSetDownLink(itup, blkno)
Definition: nbtree.h:303
static void _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2)
Definition: nbtsort.c:541
BlockNumber btps_blkno
Definition: nbtsort.c:243
void tuplesort_initialize_shared(Sharedsort *shared, int nWorkers, dsm_segment *seg)
Definition: tuplesort.c:4332
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:327
bool btws_use_wal
Definition: nbtsort.c:259
BTSpool * spool
Definition: nbtsort.c:219
static void _bt_blwritepage(BTWriteState *wstate, Page page, BlockNumber blkno)
Definition: nbtsort.c:643
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
#define BTreeTupleSetNAtts(itup, n)
Definition: nbtree.h:336
ParallelContext * pcxt
Definition: nbtsort.c:182
BTPageOpaqueData * BTPageOpaque
Definition: nbtree.h:69
Size PageGetFreeSpace(Page page)
Definition: bufpage.c:574
Relation heap
Definition: nbtsort.c:99
signed int int32
Definition: c.h:345
struct BTBuildState BTBuildState
uint16 OffsetNumber
Definition: off.h:24
Definition: type.h:89
#define RelationOpenSmgr(relation)
Definition: rel.h:485
IndexTuple _bt_truncate(Relation rel, IndexTuple lastleft, IndexTuple firstright, BTScanInsert itup_key)
Definition: nbtutils.c:2109
void ResetUsage(void)
Definition: postgres.c:4587
void WaitForParallelWorkersToFinish(ParallelContext *pcxt)
Definition: parallel.c:738
#define SpinLockAcquire(lock)
Definition: spin.h:62
void ConditionVariableInit(ConditionVariable *cv)
void DestroyParallelContext(ParallelContext *pcxt)
Definition: parallel.c:892
Page btps_page
Definition: nbtsort.c:242
struct BTSpool BTSpool
#define ItemIdGetLength(itemId)
Definition: itemid.h:59
void pfree(void *pointer)
Definition: mcxt.c:1056
void ConditionVariableCancelSleep(void)
#define BTREE_NONLEAF_FILLFACTOR
Definition: nbtree.h:171
static int compare(const void *arg1, const void *arg2)
Definition: geqo_pool.c:145
Size btps_full
Definition: nbtsort.c:247
#define ERROR
Definition: elog.h:43
static double _bt_parallel_heapscan(BTBuildState *buildstate, bool *brokenhotchain)
Definition: nbtsort.c:1513
IndexBuildResult * btbuild(Relation heap, Relation index, IndexInfo *indexInfo)
Definition: nbtsort.c:301
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:1531
bool isunique
Definition: nbtsort.c:101
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:707
static void _bt_buildadd(BTWriteState *wstate, BTPageState *state, IndexTuple itup)
Definition: nbtsort.c:827
void ExitParallelMode(void)
Definition: xact.c:976
Oid heaprelid
Definition: nbtsort.c:116
void _bt_initmetapage(Page page, BlockNumber rootbknum, uint32 level)
Definition: nbtpage.c:50
bool PageIndexTupleOverwrite(Page page, OffsetNumber offnum, Item newtup, Size newsize)
Definition: bufpage.c:1058
BlockNumber btpo_prev
Definition: nbtree.h:58
IndexTuple CopyIndexTuple(IndexTuple source)
Definition: indextuple.c:509
OffsetNumber btps_lastoff
Definition: nbtsort.c:245
void smgrwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, char *buffer, bool skipFsync)
Definition: smgr.c:530
BTSpool * spool2
Definition: nbtsort.c:225
#define RowExclusiveLock
Definition: lockdefs.h:38
IndexTupleData * IndexTuple
Definition: itup.h:53
struct BTShared BTShared
#define PARALLEL_KEY_QUERY_TEXT
Definition: nbtsort.c:83
#define P_FIRSTKEY
Definition: nbtree.h:219
double indtuples
Definition: nbtsort.c:226
bool isunique
Definition: nbtsort.c:216
#define RelationGetRelationName(relation)
Definition: rel.h:462
#define P_LEFTMOST(opaque)
Definition: nbtree.h:188
#define PARALLEL_KEY_TUPLESORT_SPOOL2
Definition: nbtsort.c:82
bool ii_BrokenHotChain
Definition: execnodes.h:173
unsigned int uint32
Definition: c.h:357
struct ItemIdData ItemIdData
MemoryContext CurrentMemoryContext
Definition: mcxt.c:38
#define IndexRelationGetNumberOfKeyAttributes(relation)
Definition: rel.h:447
int nworkers_launched
Definition: parallel.h:38
void PrepareSortSupportFromIndexRel(Relation indexRel, int16 strategy, SortSupport ssup)
Definition: sortsupport.c:161
bool brokenhotchain
Definition: nbtsort.c:158
void LaunchParallelWorkers(ParallelContext *pcxt)
Definition: parallel.c:515
#define BTREE_METAPAGE
Definition: nbtree.h:132
void UnregisterSnapshot(Snapshot snapshot)
Definition: snapmgr.c:907
Relation index
Definition: nbtsort.c:100
Relation index
Definition: nbtsort.c:257
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:200
#define PageGetItemId(page, offsetNumber)
Definition: bufpage.h:235
#define SK_BT_NULLS_FIRST
Definition: nbtree.h:681
int progress
Definition: pgbench.c:234
uint32 level
Definition: nbtree.h:62
#define SpinLockRelease(lock)
Definition: spin.h:64
struct BTLeader BTLeader
BlockNumber btws_pages_written
Definition: nbtsort.c:261
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:246
uintptr_t Datum
Definition: postgres.h:367
struct IndexTupleData IndexTupleData
static BTPageState * _bt_pagestate(BTWriteState *wstate, uint32 level)
Definition: nbtsort.c:701
static void _bt_build_callback(Relation index, ItemPointer tid, Datum *values, bool *isnull, bool tupleIsAlive, void *state)
Definition: nbtsort.c:587
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:370
bool log_btree_build_stats
Definition: guc.c:528
BTLeader * btleader
Definition: nbtsort.c:233
int work_mem
Definition: globals.c:121
double reltuples
Definition: nbtsort.c:155
AttrNumber ssup_attno
Definition: sortsupport.h:81
#define RelationGetNumberOfBlocks(reln)
Definition: bufmgr.h:198
double indtuples
Definition: nbtsort.c:157
static void _bt_parallel_scan_and_sort(BTSpool *btspool, BTSpool *btspool2, BTShared *btshared, Sharedsort *sharedsort, Sharedsort *sharedsort2, int sortmem, bool progress)
Definition: nbtsort.c:1702
#define IsMVCCSnapshot(snapshot)
Definition: snapmgr.h:97
struct BTPageState * btps_next
Definition: nbtsort.c:248
BTScanInsert inskey
Definition: nbtsort.c:258
#define InvalidOffsetNumber
Definition: off.h:26
IndexTuple btps_lowkey
Definition: nbtsort.c:244
void tuplesort_putindextuplevalues(Tuplesortstate *state, Relation rel, ItemPointer self, Datum *values, bool *isnull)
Definition: tuplesort.c:1478
slock_t mutex
Definition: nbtsort.c:136
int maintenance_work_mem
Definition: globals.c:122
Size table_parallelscan_estimate(Relation rel, Snapshot snapshot)
Definition: tableam.c:126
bool havedead
Definition: nbtsort.c:156
RelFileNode rd_node
Definition: rel.h:55
bool ii_Unique
Definition: execnodes.h:170
void tuplesort_attach_shared(Sharedsort *shared, dsm_segment *seg)
Definition: tuplesort.c:4355
#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:728
#define SK_BT_DESC
Definition: nbtree.h:680
Definition: regguts.h:298
int nparticipanttuplesorts
Definition: nbtsort.c:190
void pgstat_progress_update_multi_param(int nparam, const int *index, const int64 *val)
Definition: pgstat.c:3235
#define OffsetNumberNext(offsetNumber)
Definition: off.h:52
int ii_ParallelWorkers
Definition: execnodes.h:174
size_t Size
Definition: c.h:456
#define PageGetSpecialPointer(page)
Definition: bufpage.h:326
static Size _bt_parallel_estimate_shared(Relation heap, Snapshot snapshot)
Definition: nbtsort.c:1493
#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:1191
#define MAXALIGN(LEN)
Definition: c.h:681
#define index_getattr(tup, attnum, tupleDesc, isnull)
Definition: itup.h:100
void EnterParallelMode(void)
Definition: xact.c:963
#define PROGRESS_SCAN_BLOCKS_TOTAL
Definition: progress.h:119
ScanKeyData scankeys[INDEX_MAX_KEYS]
Definition: nbtree.h:477
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:1547
Size tuplesort_estimate_shared(int nWorkers)
Definition: tuplesort.c:4311
#define RelationNeedsWAL(relation)
Definition: rel.h:530
void _bt_check_third_page(Relation rel, Relation heap, bool needheaptidspace, Page page, IndexTuple newtup)
Definition: nbtutils.c:2523
bool ii_Concurrent
Definition: execnodes.h:172
#define SnapshotAny
Definition: snapmgr.h:69
#define PROGRESS_BTREE_PHASE_LEAF_LOAD
Definition: nbtree.h:708
static void _bt_spooldestroy(BTSpool *btspool)
Definition: nbtsort.c:520
void index_close(Relation relation, LOCKMODE lockmode)
Definition: indexam.c:152
void smgrextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, char *buffer, bool skipFsync)
Definition: smgr.c:483
#define BTMaxItemSize(page)
Definition: nbtree.h:148
#define P_HIKEY
Definition: nbtree.h:218
static Datum values[MAXATTR]
Definition: bootstrap.c:167
static void _bt_slideleft(Page page)
Definition: nbtsort.c:734
struct BTPageState BTPageState
#define AccessExclusiveLock
Definition: lockdefs.h:45
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
static Page _bt_blnewpage(uint32 level)
Definition: nbtsort.c:616
Oid sk_collation
Definition: skey.h:70
XLogRecPtr log_newpage(RelFileNode *rnode, ForkNumber forkNum, BlockNumber blkno, Page page, bool page_std)
Definition: xloginsert.c:972
#define elog(elevel,...)
Definition: elog.h:228
#define ShareLock
Definition: lockdefs.h:41
int i
static void _bt_spool(BTSpool *btspool, ItemPointer self, Datum *values, bool *isnull)
Definition: nbtsort.c:530
void _bt_pageinit(Page page, Size size)
Definition: nbtpage.c:924
static void _bt_sortaddtup(Page page, Size itemsize, IndexTuple itup, OffsetNumber itup_off)
Definition: nbtsort.c:769
#define BUFFERALIGN(LEN)
Definition: c.h:683
#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:975
static void _bt_uppershutdown(BTWriteState *wstate, BTPageState *state)
Definition: nbtsort.c:1052
ConditionVariable workersdonecv
Definition: nbtsort.c:128
unsigned short t_info
Definition: itup.h:49
Relation heap
Definition: nbtsort.c:256
#define ItemIdSetUnused(itemId)
Definition: itemid.h:128
#define PARALLEL_KEY_BTREE_SHARED
Definition: nbtsort.c:80
void table_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan, Snapshot snapshot)
Definition: tableam.c:141
void tuplesort_end(Tuplesortstate *state)
Definition: tuplesort.c:1236
#define BTLessStrategyNumber
Definition: stratnum.h:29
Relation table_open(Oid relationId, LOCKMODE lockmode)
Definition: table.c:39
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:635
void smgrimmedsync(SMgrRelation reln, ForkNumber forknum)
Definition: smgr.c:637
#define RelationGetRelid(relation)
Definition: rel.h:428
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:126
#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 P_ISLEAF(opaque)
Definition: nbtree.h:190