Debian Codegraph @

src/backend/access/gin/ginfast.c

raw
1 /*-------------------------------------------------------------------------
2 *
3 * ginfast.c
4 * Fast insert routines for the Postgres inverted index access method.
5 * Pending entries are stored in linear list of pages. Later on
6 * (typically during VACUUM), ginInsertCleanup() will be invoked to
7 * transfer pending entries into the regular index structure. This
8 * wins because bulk insertion is much more efficient than retail.
9 *
10 * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
11 * Portions Copyright (c) 1994, Regents of the University of California
12 *
13 * IDENTIFICATION
14 * src/backend/access/gin/ginfast.c
15 *
16 *-------------------------------------------------------------------------
17 */
18
19 #include "postgres.h"
20
21 #include "access/gin_private.h"
22 #include "access/ginxlog.h"
23 #include "access/xlog.h"
24 #include "access/xloginsert.h"
25 #include "catalog/pg_am.h"
26 #include "commands/vacuum.h"
27 #include "miscadmin.h"
28 #include "port/pg_bitutils.h"
29 #include "postmaster/autovacuum.h"
30 #include "storage/indexfsm.h"
31 #include "storage/lmgr.h"
32 #include "storage/predicate.h"
33 #include "utils/acl.h"
34 #include "utils/fmgrprotos.h"
35 #include "utils/memutils.h"
36 #include "utils/rel.h"
37
38 /* GUC parameter */
39 int gin_pending_list_limit = 0;
40
41 #define GIN_PAGE_FREESIZE \
42 ( (Size) BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(GinPageOpaqueData)) )
43
44 typedef struct KeyArray
45 {
46 Datum *keys; /* expansible array */
47 GinNullCategory *categories; /* another expansible array */
48 int32 nvalues; /* current number of valid entries */
49 int32 maxvalues; /* allocated size of arrays */
50 } KeyArray;
51
52
53 /*
54 * Build a pending-list page from the given array of tuples, and write it out.
55 *
56 * Returns amount of free space left on the page.
57 */
58 static int32
59 writeListPage(Relation index, Buffer buffer,
60 IndexTuple *tuples, int32 ntuples, BlockNumber rightlink)
61 {
62 Page page = BufferGetPage(buffer);
63 int32 i,
64 freesize,
65 size = 0;
66 OffsetNumber l,
67 off;
68 PGAlignedBlock workspace;
69 char *ptr;
70
71 START_CRIT_SECTION();
72
73 GinInitBuffer(buffer, GIN_LIST);
74
75 off = FirstOffsetNumber;
76 ptr = workspace.data;
77
78 for (i = 0; i < ntuples; i++)
79 {
80 int this_size = IndexTupleSize(tuples[i]);
81
82 memcpy(ptr, tuples[i], this_size);
83 ptr += this_size;
84 size += this_size;
85
86 l = PageAddItem(page, (Item) tuples[i], this_size, off, false, false);
87
88 if (l == InvalidOffsetNumber)
89 elog(ERROR, "failed to add item to index page in \"%s\"",
90 RelationGetRelationName(index));
91
92 off++;
93 }
94
95 Assert(size <= BLCKSZ); /* else we overran workspace */
96
97 GinPageGetOpaque(page)->rightlink = rightlink;
98
99 /*
100 * tail page may contain only whole row(s) or final part of row placed on
101 * previous pages (a "row" here meaning all the index tuples generated for
102 * one heap tuple)
103 */
104 if (rightlink == InvalidBlockNumber)
105 {
106 GinPageSetFullRow(page);
107 GinPageGetOpaque(page)->maxoff = 1;
108 }
109 else
110 {
111 GinPageGetOpaque(page)->maxoff = 0;
112 }
113
114 MarkBufferDirty(buffer);
115
116 if (RelationNeedsWAL(index))
117 {
118 ginxlogInsertListPage data;
119 XLogRecPtr recptr;
120
121 data.rightlink = rightlink;
122 data.ntuples = ntuples;
123
124 XLogBeginInsert();
125 XLogRegisterData(&data, sizeof(ginxlogInsertListPage));
126
127 XLogRegisterBuffer(0, buffer, REGBUF_WILL_INIT);
128 XLogRegisterBufData(0, workspace.data, size);
129
130 recptr = XLogInsert(RM_GIN_ID, XLOG_GIN_INSERT_LISTPAGE);
131 PageSetLSN(page, recptr);
132 }
133
134 /* get free space before releasing buffer */
135 freesize = PageGetExactFreeSpace(page);
136
137 UnlockReleaseBuffer(buffer);
138
139 END_CRIT_SECTION();
140
141 return freesize;
142 }
143
144 static void
145 makeSublist(Relation index, IndexTuple *tuples, int32 ntuples,
146 GinMetaPageData *res)
147 {
148 Buffer curBuffer = InvalidBuffer;
149 Buffer prevBuffer = InvalidBuffer;
150 int i,
151 size = 0,
152 tupsize;
153 int startTuple = 0;
154
155 Assert(ntuples > 0);
156
157 /*
158 * Split tuples into pages
159 */
160 for (i = 0; i < ntuples; i++)
161 {
162 if (curBuffer == InvalidBuffer)
163 {
164 curBuffer = GinNewBuffer(index);
165
166 if (prevBuffer != InvalidBuffer)
167 {
168 res->nPendingPages++;
169 writeListPage(index, prevBuffer,
170 tuples + startTuple,
171 i - startTuple,
172 BufferGetBlockNumber(curBuffer));
173 }
174 else
175 {
176 res->head = BufferGetBlockNumber(curBuffer);
177 }
178
179 prevBuffer = curBuffer;
180 startTuple = i;
181 size = 0;
182 }
183
184 tupsize = MAXALIGN(IndexTupleSize(tuples[i])) + sizeof(ItemIdData);
185
186 if (size + tupsize > GinListPageSize)
187 {
188 /* won't fit, force a new page and reprocess */
189 i--;
190 curBuffer = InvalidBuffer;
191 }
192 else
193 {
194 size += tupsize;
195 }
196 }
197
198 /*
199 * Write last page
200 */
201 res->tail = BufferGetBlockNumber(curBuffer);
202 res->tailFreeSize = writeListPage(index, curBuffer,
203 tuples + startTuple,
204 ntuples - startTuple,
205 InvalidBlockNumber);
206 res->nPendingPages++;
207 /* that was only one heap tuple */
208 res->nPendingHeapTuples = 1;
209 }
210
211 /*
212 * Write the index tuples contained in *collector into the index's
213 * pending list.
214 *
215 * Function guarantees that all these tuples will be inserted consecutively,
216 * preserving order
217 */
218 void
219 ginHeapTupleFastInsert(GinState *ginstate, GinTupleCollector *collector)
220 {
221 Relation index = ginstate->index;
222 Buffer metabuffer;
223 Page metapage;
224 GinMetaPageData *metadata = NULL;
225 Buffer buffer = InvalidBuffer;
226 Page page = NULL;
227 ginxlogUpdateMeta data;
228 bool separateList = false;
229 bool needCleanup = false;
230 int cleanupSize;
231 bool needWal;
232
233 if (collector->ntuples == 0)
234 return;
235
236 needWal = RelationNeedsWAL(index);
237
238 data.locator = index->rd_locator;
239 data.ntuples = 0;
240 data.newRightlink = data.prevTail = InvalidBlockNumber;
241
242 metabuffer = ReadBuffer(index, GIN_METAPAGE_BLKNO);
243 metapage = BufferGetPage(metabuffer);
244
245 /*
246 * An insertion to the pending list could logically belong anywhere in the
247 * tree, so it conflicts with all serializable scans. All scans acquire a
248 * predicate lock on the metabuffer to represent that. Therefore we'll
249 * check for conflicts in, but not until we have the page locked and are
250 * ready to modify the page.
251 */
252
253 if (collector->sumsize + collector->ntuples * sizeof(ItemIdData) > GinListPageSize)
254 {
255 /*
256 * Total size is greater than one page => make sublist
257 */
258 separateList = true;
259 }
260 else
261 {
262 LockBuffer(metabuffer, GIN_EXCLUSIVE);
263 metadata = GinPageGetMeta(metapage);
264
265 if (metadata->head == InvalidBlockNumber ||
266 collector->sumsize + collector->ntuples * sizeof(ItemIdData) > metadata->tailFreeSize)
267 {
268 /*
269 * Pending list is empty or total size is greater than freespace
270 * on tail page => make sublist
271 *
272 * We unlock metabuffer to keep high concurrency
273 */
274 separateList = true;
275 LockBuffer(metabuffer, GIN_UNLOCK);
276 }
277 }
278
279 if (separateList)
280 {
281 /*
282 * We should make sublist separately and append it to the tail
283 */
284 GinMetaPageData sublist;
285
286 memset(&sublist, 0, sizeof(GinMetaPageData));
287 makeSublist(index, collector->tuples, collector->ntuples, &sublist);
288
289 /*
290 * metapage was unlocked, see above
291 */
292 LockBuffer(metabuffer, GIN_EXCLUSIVE);
293 metadata = GinPageGetMeta(metapage);
294
295 CheckForSerializableConflictIn(index, NULL, GIN_METAPAGE_BLKNO);
296
297 if (metadata->head == InvalidBlockNumber)
298 {
299 /*
300 * Main list is empty, so just insert sublist as main list
301 */
302 START_CRIT_SECTION();
303
304 metadata->head = sublist.head;
305 metadata->tail = sublist.tail;
306 metadata->tailFreeSize = sublist.tailFreeSize;
307
308 metadata->nPendingPages = sublist.nPendingPages;
309 metadata->nPendingHeapTuples = sublist.nPendingHeapTuples;
310
311 if (needWal)
312 XLogBeginInsert();
313 }
314 else
315 {
316 /*
317 * Merge lists
318 */
319 data.prevTail = metadata->tail;
320 data.newRightlink = sublist.head;
321
322 buffer = ReadBuffer(index, metadata->tail);
323 LockBuffer(buffer, GIN_EXCLUSIVE);
324 page = BufferGetPage(buffer);
325
326 Assert(GinPageGetOpaque(page)->rightlink == InvalidBlockNumber);
327
328 START_CRIT_SECTION();
329
330 GinPageGetOpaque(page)->rightlink = sublist.head;
331
332 MarkBufferDirty(buffer);
333
334 metadata->tail = sublist.tail;
335 metadata->tailFreeSize = sublist.tailFreeSize;
336
337 metadata->nPendingPages += sublist.nPendingPages;
338 metadata->nPendingHeapTuples += sublist.nPendingHeapTuples;
339
340 if (needWal)
341 {
342 XLogBeginInsert();
343 XLogRegisterBuffer(1, buffer, REGBUF_STANDARD);
344 }
345 }
346 }
347 else
348 {
349 /*
350 * Insert into tail page. Metapage is already locked
351 */
352 OffsetNumber l,
353 off;
354 int i,
355 tupsize;
356 char *ptr;
357 char *collectordata;
358
359 CheckForSerializableConflictIn(index, NULL, GIN_METAPAGE_BLKNO);
360
361 buffer = ReadBuffer(index, metadata->tail);
362 LockBuffer(buffer, GIN_EXCLUSIVE);
363 page = BufferGetPage(buffer);
364
365 off = (PageIsEmpty(page)) ? FirstOffsetNumber :
366 OffsetNumberNext(PageGetMaxOffsetNumber(page));
367
368 collectordata = ptr = (char *) palloc(collector->sumsize);
369
370 data.ntuples = collector->ntuples;
371
372 START_CRIT_SECTION();
373
374 if (needWal)
375 XLogBeginInsert();
376
377 /*
378 * Increase counter of heap tuples
379 */
380 Assert(GinPageGetOpaque(page)->maxoff <= metadata->nPendingHeapTuples);
381 GinPageGetOpaque(page)->maxoff++;
382 metadata->nPendingHeapTuples++;
383
384 for (i = 0; i < collector->ntuples; i++)
385 {
386 tupsize = IndexTupleSize(collector->tuples[i]);
387 l = PageAddItem(page, (Item) collector->tuples[i], tupsize, off, false, false);
388
389 if (l == InvalidOffsetNumber)
390 elog(ERROR, "failed to add item to index page in \"%s\"",
391 RelationGetRelationName(index));
392
393 memcpy(ptr, collector->tuples[i], tupsize);
394 ptr += tupsize;
395
396 off++;
397 }
398
399 Assert((ptr - collectordata) <= collector->sumsize);
400
401 MarkBufferDirty(buffer);
402
403 if (needWal)
404 {
405 XLogRegisterBuffer(1, buffer, REGBUF_STANDARD);
406 XLogRegisterBufData(1, collectordata, collector->sumsize);
407 }
408
409 metadata->tailFreeSize = PageGetExactFreeSpace(page);
410 }
411
412 /*
413 * Set pd_lower just past the end of the metadata. This is essential,
414 * because without doing so, metadata will be lost if xlog.c compresses
415 * the page. (We must do this here because pre-v11 versions of PG did not
416 * set the metapage's pd_lower correctly, so a pg_upgraded index might
417 * contain the wrong value.)
418 */
419 ((PageHeader) metapage)->pd_lower =
420 ((char *) metadata + sizeof(GinMetaPageData)) - (char *) metapage;
421
422 /*
423 * Write metabuffer, make xlog entry
424 */
425 MarkBufferDirty(metabuffer);
426
427 if (needWal)
428 {
429 XLogRecPtr recptr;
430
431 memcpy(&data.metadata, metadata, sizeof(GinMetaPageData));
432
433 XLogRegisterBuffer(0, metabuffer, REGBUF_WILL_INIT | REGBUF_STANDARD);
434 XLogRegisterData(&data, sizeof(ginxlogUpdateMeta));
435
436 recptr = XLogInsert(RM_GIN_ID, XLOG_GIN_UPDATE_META_PAGE);
437 PageSetLSN(metapage, recptr);
438
439 if (buffer != InvalidBuffer)
440 {
441 PageSetLSN(page, recptr);
442 }
443 }
444
445 if (buffer != InvalidBuffer)
446 UnlockReleaseBuffer(buffer);
447
448 /*
449 * Force pending list cleanup when it becomes too long. And,
450 * ginInsertCleanup could take significant amount of time, so we prefer to
451 * call it when it can do all the work in a single collection cycle. In
452 * non-vacuum mode, it shouldn't require maintenance_work_mem, so fire it
453 * while pending list is still small enough to fit into
454 * gin_pending_list_limit.
455 *
456 * ginInsertCleanup() should not be called inside our CRIT_SECTION.
457 */
458 cleanupSize = GinGetPendingListCleanupSize(index);
459 if (metadata->nPendingPages * GIN_PAGE_FREESIZE > cleanupSize * (Size) 1024)
460 needCleanup = true;
461
462 UnlockReleaseBuffer(metabuffer);
463
464 END_CRIT_SECTION();
465
466 /*
467 * Since it could contend with concurrent cleanup process we cleanup
468 * pending list not forcibly.
469 */
470 if (needCleanup)
471 ginInsertCleanup(ginstate, false, true, false, NULL);
472 }
473
474 /*
475 * Create temporary index tuples for a single indexable item (one index column
476 * for the heap tuple specified by ht_ctid), and append them to the array
477 * in *collector. They will subsequently be written out using
478 * ginHeapTupleFastInsert. Note that to guarantee consistent state, all
479 * temp tuples for a given heap tuple must be written in one call to
480 * ginHeapTupleFastInsert.
481 */
482 void
483 ginHeapTupleFastCollect(GinState *ginstate,
484 GinTupleCollector *collector,
485 OffsetNumber attnum, Datum value, bool isNull,
486 ItemPointer ht_ctid)
487 {
488 Datum *entries;
489 GinNullCategory *categories;
490 int32 i,
491 nentries;
492
493 /*
494 * Extract the key values that need to be inserted in the index
495 */
496 entries = ginExtractEntries(ginstate, attnum, value, isNull,
497 &nentries, &categories);
498
499 /*
500 * Protect against integer overflow in allocation calculations
501 */
502 if (nentries < 0 ||
503 collector->ntuples + nentries > MaxAllocSize / sizeof(IndexTuple))
504 elog(ERROR, "too many entries for GIN index");
505
506 /*
507 * Allocate/reallocate memory for storing collected tuples
508 */
509 if (collector->tuples == NULL)
510 {
511 /*
512 * Determine the number of elements to allocate in the tuples array
513 * initially. Make it a power of 2 to avoid wasting memory when
514 * resizing (since palloc likes powers of 2).
515 */
516 collector->lentuples = pg_nextpower2_32(Max(16, nentries));
517 collector->tuples = palloc_array(IndexTuple, collector->lentuples);
518 }
519 else if (collector->lentuples < collector->ntuples + nentries)
520 {
521 /*
522 * Advance lentuples to the next suitable power of 2. This won't
523 * overflow, though we could get to a value that exceeds
524 * MaxAllocSize/sizeof(IndexTuple), causing an error in repalloc.
525 */
526 collector->lentuples = pg_nextpower2_32(collector->ntuples + nentries);
527 collector->tuples = repalloc_array(collector->tuples,
528 IndexTuple, collector->lentuples);
529 }
530
531 /*
532 * Build an index tuple for each key value, and add to array. In pending
533 * tuples we just stick the heap TID into t_tid.
534 */
535 for (i = 0; i < nentries; i++)
536 {
537 IndexTuple itup;
538
539 itup = GinFormTuple(ginstate, attnum, entries[i], categories[i],
540 NULL, 0, 0, true);
541 itup->t_tid = *ht_ctid;
542 collector->tuples[collector->ntuples++] = itup;
543 collector->sumsize += IndexTupleSize(itup);
544 }
545 }
546
547 /*
548 * Deletes pending list pages up to (not including) newHead page.
549 * If newHead == InvalidBlockNumber then function drops the whole list.
550 *
551 * metapage is pinned and exclusive-locked throughout this function.
552 */
553 static void
554 shiftList(Relation index, Buffer metabuffer, BlockNumber newHead,
555 bool fill_fsm, IndexBulkDeleteResult *stats)
556 {
557 Page metapage;
558 GinMetaPageData *metadata;
559 BlockNumber blknoToDelete;
560
561 metapage = BufferGetPage(metabuffer);
562 metadata = GinPageGetMeta(metapage);
563 blknoToDelete = metadata->head;
564
565 do
566 {
567 Page page;
568 int i;
569 int64 nDeletedHeapTuples = 0;
570 ginxlogDeleteListPages data;
571 Buffer buffers[GIN_NDELETE_AT_ONCE];
572 BlockNumber freespace[GIN_NDELETE_AT_ONCE];
573
574 data.ndeleted = 0;
575 while (data.ndeleted < GIN_NDELETE_AT_ONCE && blknoToDelete != newHead)
576 {
577 freespace[data.ndeleted] = blknoToDelete;
578 buffers[data.ndeleted] = ReadBuffer(index, blknoToDelete);
579 LockBuffer(buffers[data.ndeleted], GIN_EXCLUSIVE);
580 page = BufferGetPage(buffers[data.ndeleted]);
581
582 data.ndeleted++;
583
584 Assert(!GinPageIsDeleted(page));
585
586 nDeletedHeapTuples += GinPageGetOpaque(page)->maxoff;
587 blknoToDelete = GinPageGetOpaque(page)->rightlink;
588 }
589
590 if (stats)
591 stats->pages_deleted += data.ndeleted;
592
593 /*
594 * This operation touches an unusually large number of pages, so
595 * prepare the XLogInsert machinery for that before entering the
596 * critical section.
597 */
598 if (RelationNeedsWAL(index))
599 XLogEnsureRecordSpace(data.ndeleted, 0);
600
601 START_CRIT_SECTION();
602
603 metadata->head = blknoToDelete;
604
605 Assert(metadata->nPendingPages >= data.ndeleted);
606 metadata->nPendingPages -= data.ndeleted;
607 Assert(metadata->nPendingHeapTuples >= nDeletedHeapTuples);
608 metadata->nPendingHeapTuples -= nDeletedHeapTuples;
609
610 if (blknoToDelete == InvalidBlockNumber)
611 {
612 metadata->tail = InvalidBlockNumber;
613 metadata->tailFreeSize = 0;
614 metadata->nPendingPages = 0;
615 metadata->nPendingHeapTuples = 0;
616 }
617
618 /*
619 * Set pd_lower just past the end of the metadata. This is essential,
620 * because without doing so, metadata will be lost if xlog.c
621 * compresses the page. (We must do this here because pre-v11
622 * versions of PG did not set the metapage's pd_lower correctly, so a
623 * pg_upgraded index might contain the wrong value.)
624 */
625 ((PageHeader) metapage)->pd_lower =
626 ((char *) metadata + sizeof(GinMetaPageData)) - (char *) metapage;
627
628 MarkBufferDirty(metabuffer);
629
630 for (i = 0; i < data.ndeleted; i++)
631 {
632 page = BufferGetPage(buffers[i]);
633 GinPageGetOpaque(page)->flags = GIN_DELETED;
634 MarkBufferDirty(buffers[i]);
635 }
636
637 if (RelationNeedsWAL(index))
638 {
639 XLogRecPtr recptr;
640
641 XLogBeginInsert();
642 XLogRegisterBuffer(0, metabuffer,
643 REGBUF_WILL_INIT | REGBUF_STANDARD);
644 for (i = 0; i < data.ndeleted; i++)
645 XLogRegisterBuffer(i + 1, buffers[i], REGBUF_WILL_INIT);
646
647 memcpy(&data.metadata, metadata, sizeof(GinMetaPageData));
648
649 XLogRegisterData(&data,
650 sizeof(ginxlogDeleteListPages));
651
652 recptr = XLogInsert(RM_GIN_ID, XLOG_GIN_DELETE_LISTPAGE);
653 PageSetLSN(metapage, recptr);
654
655 for (i = 0; i < data.ndeleted; i++)
656 {
657 page = BufferGetPage(buffers[i]);
658 PageSetLSN(page, recptr);
659 }
660 }
661
662 for (i = 0; i < data.ndeleted; i++)
663 UnlockReleaseBuffer(buffers[i]);
664
665 END_CRIT_SECTION();
666
667 for (i = 0; fill_fsm && i < data.ndeleted; i++)
668 RecordFreeIndexPage(index, freespace[i]);
669
670 } while (blknoToDelete != newHead);
671 }
672
673 /* Initialize empty KeyArray */
674 static void
675 initKeyArray(KeyArray *keys, int32 maxvalues)
676 {
677 keys->keys = palloc_array(Datum, maxvalues);
678 keys->categories = palloc_array(GinNullCategory, maxvalues);
679 keys->nvalues = 0;
680 keys->maxvalues = maxvalues;
681 }
682
683 /* Add datum to KeyArray, resizing if needed */
684 static void
685 addDatum(KeyArray *keys, Datum datum, GinNullCategory category)
686 {
687 if (keys->nvalues >= keys->maxvalues)
688 {
689 keys->maxvalues *= 2;
690 keys->keys = repalloc_array(keys->keys, Datum, keys->maxvalues);
691 keys->categories = repalloc_array(keys->categories, GinNullCategory, keys->maxvalues);
692 }
693
694 keys->keys[keys->nvalues] = datum;
695 keys->categories[keys->nvalues] = category;
696 keys->nvalues++;
697 }
698
699 /*
700 * Collect data from a pending-list page in preparation for insertion into
701 * the main index.
702 *
703 * Go through all tuples >= startoff on page and collect values in accum
704 *
705 * Note that ka is just workspace --- it does not carry any state across
706 * calls.
707 */
708 static void
709 processPendingPage(BuildAccumulator *accum, KeyArray *ka,
710 Page page, OffsetNumber startoff)
711 {
712 ItemPointerData heapptr;
713 OffsetNumber i,
714 maxoff;
715 OffsetNumber attrnum;
716
717 /* reset *ka to empty */
718 ka->nvalues = 0;
719
720 maxoff = PageGetMaxOffsetNumber(page);
721 Assert(maxoff >= FirstOffsetNumber);
722 ItemPointerSetInvalid(&heapptr);
723 attrnum = 0;
724
725 for (i = startoff; i <= maxoff; i = OffsetNumberNext(i))
726 {
727 IndexTuple itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, i));
728 OffsetNumber curattnum;
729 Datum curkey;
730 GinNullCategory curcategory;
731
732 /* Check for change of heap TID or attnum */
733 curattnum = gintuple_get_attrnum(accum->ginstate, itup);
734
735 if (!ItemPointerIsValid(&heapptr))
736 {
737 heapptr = itup->t_tid;
738 attrnum = curattnum;
739 }
740 else if (!(ItemPointerEquals(&heapptr, &itup->t_tid) &&
741 curattnum == attrnum))
742 {
743 /*
744 * ginInsertBAEntries can insert several datums per call, but only
745 * for one heap tuple and one column. So call it at a boundary,
746 * and reset ka.
747 */
748 ginInsertBAEntries(accum, &heapptr, attrnum,
749 ka->keys, ka->categories, ka->nvalues);
750 ka->nvalues = 0;
751 heapptr = itup->t_tid;
752 attrnum = curattnum;
753 }
754
755 /* Add key to KeyArray */
756 curkey = gintuple_get_key(accum->ginstate, itup, &curcategory);
757 addDatum(ka, curkey, curcategory);
758 }
759
760 /* Dump out all remaining keys */
761 ginInsertBAEntries(accum, &heapptr, attrnum,
762 ka->keys, ka->categories, ka->nvalues);
763 }
764
765 /*
766 * Move tuples from pending pages into regular GIN structure.
767 *
768 * On first glance it looks completely not crash-safe. But if we crash
769 * after posting entries to the main index and before removing them from the
770 * pending list, it's okay because when we redo the posting later on, nothing
771 * bad will happen.
772 *
773 * fill_fsm indicates that ginInsertCleanup should add deleted pages
774 * to FSM otherwise caller is responsible to put deleted pages into
775 * FSM.
776 *
777 * If stats isn't null, we count deleted pending pages into the counts.
778 */
779 void
780 ginInsertCleanup(GinState *ginstate, bool full_clean,
781 bool fill_fsm, bool forceCleanup,
782 IndexBulkDeleteResult *stats)
783 {
784 Relation index = ginstate->index;
785 Buffer metabuffer,
786 buffer;
787 Page metapage,
788 page;
789 GinMetaPageData *metadata;
790 MemoryContext opCtx,
791 oldCtx;
792 BuildAccumulator accum;
793 KeyArray datums;
794 BlockNumber blkno,
795 blknoFinish;
796 bool cleanupFinish = false;
797 bool fsm_vac = false;
798 int workMemory;
799
800 /*
801 * We would like to prevent concurrent cleanup process. For that we will
802 * lock metapage in exclusive mode using LockPage() call. Nobody other
803 * will use that lock for metapage, so we keep possibility of concurrent
804 * insertion into pending list
805 */
806
807 if (forceCleanup)
808 {
809 /*
810 * We are called from [auto]vacuum/analyze or gin_clean_pending_list()
811 * and we would like to wait concurrent cleanup to finish.
812 */
813 LockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock);
814 workMemory =
815 (AmAutoVacuumWorkerProcess() && autovacuum_work_mem != -1) ?
816 autovacuum_work_mem : maintenance_work_mem;
817 }
818 else
819 {
820 /*
821 * We are called from regular insert and if we see concurrent cleanup
822 * just exit in hope that concurrent process will clean up pending
823 * list.
824 */
825 if (!ConditionalLockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock))
826 return;
827 workMemory = work_mem;
828 }
829
830 metabuffer = ReadBuffer(index, GIN_METAPAGE_BLKNO);
831 LockBuffer(metabuffer, GIN_SHARE);
832 metapage = BufferGetPage(metabuffer);
833 metadata = GinPageGetMeta(metapage);
834
835 if (metadata->head == InvalidBlockNumber)
836 {
837 /* Nothing to do */
838 UnlockReleaseBuffer(metabuffer);
839 UnlockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock);
840 return;
841 }
842
843 /*
844 * Remember a tail page to prevent infinite cleanup if other backends add
845 * new tuples faster than we can cleanup.
846 */
847 blknoFinish = metadata->tail;
848
849 /*
850 * Read and lock head of pending list
851 */
852 blkno = metadata->head;
853 buffer = ReadBuffer(index, blkno);
854 LockBuffer(buffer, GIN_SHARE);
855 page = BufferGetPage(buffer);
856
857 LockBuffer(metabuffer, GIN_UNLOCK);
858
859 /*
860 * Initialize. All temporary space will be in opCtx
861 */
862 opCtx = AllocSetContextCreate(CurrentMemoryContext,
863 "GIN insert cleanup temporary context",
864 ALLOCSET_DEFAULT_SIZES);
865
866 oldCtx = MemoryContextSwitchTo(opCtx);
867
868 initKeyArray(&datums, 128);
869 ginInitBA(&accum);
870 accum.ginstate = ginstate;
871
872 /*
873 * At the top of this loop, we have pin and lock on the current page of
874 * the pending list. However, we'll release that before exiting the loop.
875 * Note we also have pin but not lock on the metapage.
876 */
877 for (;;)
878 {
879 Assert(!GinPageIsDeleted(page));
880
881 /*
882 * Are we walk through the page which as we remember was a tail when
883 * we start our cleanup? But if caller asks us to clean up whole
884 * pending list then ignore old tail, we will work until list becomes
885 * empty.
886 */
887 if (blkno == blknoFinish && full_clean == false)
888 cleanupFinish = true;
889
890 /*
891 * read page's datums into accum
892 */
893 processPendingPage(&accum, &datums, page, FirstOffsetNumber);
894
895 vacuum_delay_point(false);
896
897 /*
898 * Is it time to flush memory to disk? Flush if we are at the end of
899 * the pending list, or if we have a full row and memory is getting
900 * full.
901 */
902 if (GinPageGetOpaque(page)->rightlink == InvalidBlockNumber ||
903 (GinPageHasFullRow(page) &&
904 accum.allocatedMemory >= workMemory * (Size) 1024))
905 {
906 ItemPointerData *list;
907 uint32 nlist;
908 Datum key;
909 GinNullCategory category;
910 OffsetNumber maxoff,
911 attnum;
912
913 /*
914 * Unlock current page to increase performance. Changes of page
915 * will be checked later by comparing maxoff after completion of
916 * memory flush.
917 */
918 maxoff = PageGetMaxOffsetNumber(page);
919 LockBuffer(buffer, GIN_UNLOCK);
920
921 /*
922 * Moving collected data into regular structure can take
923 * significant amount of time - so, run it without locking pending
924 * list.
925 */
926 ginBeginBAScan(&accum);
927 while ((list = ginGetBAEntry(&accum,
928 &attnum, &key, &category, &nlist)) != NULL)
929 {
930 ginEntryInsert(ginstate, attnum, key, category,
931 list, nlist, NULL);
932 vacuum_delay_point(false);
933 }
934
935 /*
936 * Lock the whole list to remove pages
937 */
938 LockBuffer(metabuffer, GIN_EXCLUSIVE);
939 LockBuffer(buffer, GIN_SHARE);
940
941 Assert(!GinPageIsDeleted(page));
942
943 /*
944 * While we left the page unlocked, more stuff might have gotten
945 * added to it. If so, process those entries immediately. There
946 * shouldn't be very many, so we don't worry about the fact that
947 * we're doing this with exclusive lock. Insertion algorithm
948 * guarantees that inserted row(s) will not continue on next page.
949 * NOTE: intentionally no vacuum_delay_point in this loop.
950 */
951 if (PageGetMaxOffsetNumber(page) != maxoff)
952 {
953 ginInitBA(&accum);
954 processPendingPage(&accum, &datums, page, maxoff + 1);
955
956 ginBeginBAScan(&accum);
957 while ((list = ginGetBAEntry(&accum,
958 &attnum, &key, &category, &nlist)) != NULL)
959 ginEntryInsert(ginstate, attnum, key, category,
960 list, nlist, NULL);
961 }
962
963 /*
964 * Remember next page - it will become the new list head
965 */
966 blkno = GinPageGetOpaque(page)->rightlink;
967 UnlockReleaseBuffer(buffer); /* shiftList will do exclusive
968 * locking */
969
970 /*
971 * remove read pages from pending list, at this point all content
972 * of read pages is in regular structure
973 */
974 shiftList(index, metabuffer, blkno, fill_fsm, stats);
975
976 /* At this point, some pending pages have been freed up */
977 fsm_vac = true;
978
979 Assert(blkno == metadata->head);
980 LockBuffer(metabuffer, GIN_UNLOCK);
981
982 /*
983 * if we removed the whole pending list or we cleanup tail (which
984 * we remembered on start our cleanup process) then just exit
985 */
986 if (blkno == InvalidBlockNumber || cleanupFinish)
987 break;
988
989 /*
990 * release memory used so far and reinit state
991 */
992 MemoryContextReset(opCtx);
993 initKeyArray(&datums, datums.maxvalues);
994 ginInitBA(&accum);
995 }
996 else
997 {
998 blkno = GinPageGetOpaque(page)->rightlink;
999 UnlockReleaseBuffer(buffer);
1000 }
1001
1002 /*
1003 * Read next page in pending list
1004 */
1005 vacuum_delay_point(false);
1006 buffer = ReadBuffer(index, blkno);
1007 LockBuffer(buffer, GIN_SHARE);
1008 page = BufferGetPage(buffer);
1009 }
1010
1011 UnlockPage(index, GIN_METAPAGE_BLKNO, ExclusiveLock);
1012 ReleaseBuffer(metabuffer);
1013
1014 /*
1015 * As pending list pages can have a high churn rate, it is desirable to
1016 * recycle them immediately to the FreeSpaceMap when ordinary backends
1017 * clean the list.
1018 */
1019 if (fsm_vac && fill_fsm)
1020 IndexFreeSpaceMapVacuum(index);
1021
1022 /* Clean up temporary space */
1023 MemoryContextSwitchTo(oldCtx);
1024 MemoryContextDelete(opCtx);
1025 }
1026
1027 /*
1028 * SQL-callable function to clean the insert pending list
1029 */
1030 Datum
1031 gin_clean_pending_list(PG_FUNCTION_ARGS)
1032 {
1033 Oid indexoid = PG_GETARG_OID(0);
1034 Relation indexRel = index_open(indexoid, RowExclusiveLock);
1035 IndexBulkDeleteResult stats;
1036
1037 if (RecoveryInProgress())
1038 ereport(ERROR,
1039 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1040 errmsg("recovery is in progress"),
1041 errhint("GIN pending list cannot be cleaned up during recovery.")));
1042
1043 /* Must be a GIN index */
1044 if (indexRel->rd_rel->relkind != RELKIND_INDEX ||
1045 indexRel->rd_rel->relam != GIN_AM_OID)
1046 ereport(ERROR,
1047 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1048 errmsg("\"%s\" is not a GIN index",
1049 RelationGetRelationName(indexRel))));
1050
1051 /*
1052 * Reject attempts to read non-local temporary relations; we would be
1053 * likely to get wrong data since we have no visibility into the owning
1054 * session's local buffers.
1055 */
1056 if (RELATION_IS_OTHER_TEMP(indexRel))
1057 ereport(ERROR,
1058 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1059 errmsg("cannot access temporary indexes of other sessions")));
1060
1061 /* User must own the index (comparable to privileges needed for VACUUM) */
1062 if (!object_ownercheck(RelationRelationId, indexoid, GetUserId()))
1063 aclcheck_error(ACLCHECK_NOT_OWNER, OBJECT_INDEX,
1064 RelationGetRelationName(indexRel));
1065
1066 memset(&stats, 0, sizeof(stats));
1067
1068 /*
1069 * Can't assume anything about the content of an !indisready index. Make
1070 * those a no-op, not an error, so users can just run this function on all
1071 * indexes of the access method. Since an indisready&&!indisvalid index
1072 * is merely awaiting missed aminsert calls, we're capable of processing
1073 * it. Decline to do so, out of an abundance of caution.
1074 */
1075 if (indexRel->rd_index->indisvalid)
1076 {
1077 GinState ginstate;
1078
1079 initGinState(&ginstate, indexRel);
1080 ginInsertCleanup(&ginstate, true, true, true, &stats);
1081 }
1082 else
1083 ereport(DEBUG1,
1084 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1085 errmsg("index \"%s\" is not valid",
1086 RelationGetRelationName(indexRel))));
1087
1088 index_close(indexRel, RowExclusiveLock);
1089
1090 PG_RETURN_INT64((int64) stats.pages_deleted);
1091 }