Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * gist.c
4 : * interface routines for the postgres GiST index access method.
5 : *
6 : *
7 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
8 : * Portions Copyright (c) 1994, Regents of the University of California
9 : *
10 : * IDENTIFICATION
11 : * src/backend/access/gist/gist.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include "access/gist_private.h"
18 : #include "access/gistscan.h"
19 : #include "access/xloginsert.h"
20 : #include "catalog/pg_collation.h"
21 : #include "commands/vacuum.h"
22 : #include "miscadmin.h"
23 : #include "nodes/execnodes.h"
24 : #include "storage/predicate.h"
25 : #include "utils/fmgrprotos.h"
26 : #include "utils/index_selfuncs.h"
27 : #include "utils/memutils.h"
28 : #include "utils/rel.h"
29 :
30 : /* non-export function prototypes */
31 : static void gistfixsplit(GISTInsertState *state, GISTSTATE *giststate);
32 : static bool gistinserttuple(GISTInsertState *state, GISTInsertStack *stack,
33 : GISTSTATE *giststate, IndexTuple tuple, OffsetNumber oldoffnum);
34 : static bool gistinserttuples(GISTInsertState *state, GISTInsertStack *stack,
35 : GISTSTATE *giststate,
36 : IndexTuple *tuples, int ntup, OffsetNumber oldoffnum,
37 : Buffer leftchild, Buffer rightchild,
38 : bool unlockbuf, bool unlockleftchild);
39 : static void gistfinishsplit(GISTInsertState *state, GISTInsertStack *stack,
40 : GISTSTATE *giststate, List *splitinfo, bool unlockbuf);
41 : static void gistprunepage(Relation rel, Page page, Buffer buffer,
42 : Relation heapRel);
43 :
44 :
45 : #define ROTATEDIST(d) do { \
46 : SplitPageLayout *tmp = (SplitPageLayout *) palloc0(sizeof(SplitPageLayout)); \
47 : tmp->block.blkno = InvalidBlockNumber; \
48 : tmp->buffer = InvalidBuffer; \
49 : tmp->next = (d); \
50 : (d)=tmp; \
51 : } while(0)
52 :
53 :
54 : /*
55 : * GiST handler function: return IndexAmRoutine with access method parameters
56 : * and callbacks.
57 : */
58 : Datum
59 7592 : gisthandler(PG_FUNCTION_ARGS)
60 : {
61 7592 : IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
62 :
63 7592 : amroutine->amstrategies = 0;
64 7592 : amroutine->amsupport = GISTNProcs;
65 7592 : amroutine->amoptsprocnum = GIST_OPTIONS_PROC;
66 7592 : amroutine->amcanorder = false;
67 7592 : amroutine->amcanorderbyop = true;
68 7592 : amroutine->amcanbackward = false;
69 7592 : amroutine->amcanunique = false;
70 7592 : amroutine->amcanmulticol = true;
71 7592 : amroutine->amoptionalkey = true;
72 7592 : amroutine->amsearcharray = false;
73 7592 : amroutine->amsearchnulls = true;
74 7592 : amroutine->amstorage = true;
75 7592 : amroutine->amclusterable = true;
76 7592 : amroutine->ampredlocks = true;
77 7592 : amroutine->amcanparallel = false;
78 7592 : amroutine->amcanbuildparallel = false;
79 7592 : amroutine->amcaninclude = true;
80 7592 : amroutine->amusemaintenanceworkmem = false;
81 7592 : amroutine->amsummarizing = false;
82 7592 : amroutine->amparallelvacuumoptions =
83 : VACUUM_OPTION_PARALLEL_BULKDEL | VACUUM_OPTION_PARALLEL_COND_CLEANUP;
84 7592 : amroutine->amkeytype = InvalidOid;
85 :
86 7592 : amroutine->ambuild = gistbuild;
87 7592 : amroutine->ambuildempty = gistbuildempty;
88 7592 : amroutine->aminsert = gistinsert;
89 7592 : amroutine->aminsertcleanup = NULL;
90 7592 : amroutine->ambulkdelete = gistbulkdelete;
91 7592 : amroutine->amvacuumcleanup = gistvacuumcleanup;
92 7592 : amroutine->amcanreturn = gistcanreturn;
93 7592 : amroutine->amcostestimate = gistcostestimate;
94 7592 : amroutine->amgettreeheight = NULL;
95 7592 : amroutine->amoptions = gistoptions;
96 7592 : amroutine->amproperty = gistproperty;
97 7592 : amroutine->ambuildphasename = NULL;
98 7592 : amroutine->amvalidate = gistvalidate;
99 7592 : amroutine->amadjustmembers = gistadjustmembers;
100 7592 : amroutine->ambeginscan = gistbeginscan;
101 7592 : amroutine->amrescan = gistrescan;
102 7592 : amroutine->amgettuple = gistgettuple;
103 7592 : amroutine->amgetbitmap = gistgetbitmap;
104 7592 : amroutine->amendscan = gistendscan;
105 7592 : amroutine->ammarkpos = NULL;
106 7592 : amroutine->amrestrpos = NULL;
107 7592 : amroutine->amestimateparallelscan = NULL;
108 7592 : amroutine->aminitparallelscan = NULL;
109 7592 : amroutine->amparallelrescan = NULL;
110 :
111 7592 : PG_RETURN_POINTER(amroutine);
112 : }
113 :
114 : /*
115 : * Create and return a temporary memory context for use by GiST. We
116 : * _always_ invoke user-provided methods in a temporary memory
117 : * context, so that memory leaks in those functions cannot cause
118 : * problems. Also, we use some additional temporary contexts in the
119 : * GiST code itself, to avoid the need to do some awkward manual
120 : * memory management.
121 : */
122 : MemoryContext
123 10084 : createTempGistContext(void)
124 : {
125 10084 : return AllocSetContextCreate(CurrentMemoryContext,
126 : "GiST temporary context",
127 : ALLOCSET_DEFAULT_SIZES);
128 : }
129 :
130 : /*
131 : * gistbuildempty() -- build an empty gist index in the initialization fork
132 : */
133 : void
134 10 : gistbuildempty(Relation index)
135 : {
136 : Buffer buffer;
137 :
138 : /* Initialize the root page */
139 10 : buffer = ExtendBufferedRel(BMR_REL(index), INIT_FORKNUM, NULL,
140 : EB_SKIP_EXTENSION_LOCK | EB_LOCK_FIRST);
141 :
142 : /* Initialize and xlog buffer */
143 10 : START_CRIT_SECTION();
144 10 : GISTInitBuffer(buffer, F_LEAF);
145 10 : MarkBufferDirty(buffer);
146 10 : log_newpage_buffer(buffer, true);
147 10 : END_CRIT_SECTION();
148 :
149 : /* Unlock and release the buffer */
150 10 : UnlockReleaseBuffer(buffer);
151 10 : }
152 :
153 : /*
154 : * gistinsert -- wrapper for GiST tuple insertion.
155 : *
156 : * This is the public interface routine for tuple insertion in GiSTs.
157 : * It doesn't do any work; just locks the relation and passes the buck.
158 : */
159 : bool
160 304308 : gistinsert(Relation r, Datum *values, bool *isnull,
161 : ItemPointer ht_ctid, Relation heapRel,
162 : IndexUniqueCheck checkUnique,
163 : bool indexUnchanged,
164 : IndexInfo *indexInfo)
165 : {
166 304308 : GISTSTATE *giststate = (GISTSTATE *) indexInfo->ii_AmCache;
167 : IndexTuple itup;
168 : MemoryContext oldCxt;
169 :
170 : /* Initialize GISTSTATE cache if first call in this statement */
171 304308 : if (giststate == NULL)
172 : {
173 1926 : oldCxt = MemoryContextSwitchTo(indexInfo->ii_Context);
174 1926 : giststate = initGISTstate(r);
175 1926 : giststate->tempCxt = createTempGistContext();
176 1926 : indexInfo->ii_AmCache = (void *) giststate;
177 1926 : MemoryContextSwitchTo(oldCxt);
178 : }
179 :
180 304308 : oldCxt = MemoryContextSwitchTo(giststate->tempCxt);
181 :
182 304308 : itup = gistFormTuple(giststate, r, values, isnull, true);
183 304306 : itup->t_tid = *ht_ctid;
184 :
185 304306 : gistdoinsert(r, itup, 0, giststate, heapRel, false);
186 :
187 : /* cleanup */
188 304294 : MemoryContextSwitchTo(oldCxt);
189 304294 : MemoryContextReset(giststate->tempCxt);
190 :
191 304294 : return false;
192 : }
193 :
194 :
195 : /*
196 : * Place tuples from 'itup' to 'buffer'. If 'oldoffnum' is valid, the tuple
197 : * at that offset is atomically removed along with inserting the new tuples.
198 : * This is used to replace a tuple with a new one.
199 : *
200 : * If 'leftchildbuf' is valid, we're inserting the downlink for the page
201 : * to the right of 'leftchildbuf', or updating the downlink for 'leftchildbuf'.
202 : * F_FOLLOW_RIGHT flag on 'leftchildbuf' is cleared and NSN is set.
203 : *
204 : * If 'markfollowright' is true and the page is split, the left child is
205 : * marked with F_FOLLOW_RIGHT flag. That is the normal case. During buffered
206 : * index build, however, there is no concurrent access and the page splitting
207 : * is done in a slightly simpler fashion, and false is passed.
208 : *
209 : * If there is not enough room on the page, it is split. All the split
210 : * pages are kept pinned and locked and returned in *splitinfo, the caller
211 : * is responsible for inserting the downlinks for them. However, if
212 : * 'buffer' is the root page and it needs to be split, gistplacetopage()
213 : * performs the split as one atomic operation, and *splitinfo is set to NIL.
214 : * In that case, we continue to hold the root page locked, and the child
215 : * pages are released; note that new tuple(s) are *not* on the root page
216 : * but in one of the new child pages.
217 : *
218 : * If 'newblkno' is not NULL, returns the block number of page the first
219 : * new/updated tuple was inserted to. Usually it's the given page, but could
220 : * be its right sibling if the page was split.
221 : *
222 : * Returns 'true' if the page was split, 'false' otherwise.
223 : */
224 : bool
225 1724078 : gistplacetopage(Relation rel, Size freespace, GISTSTATE *giststate,
226 : Buffer buffer,
227 : IndexTuple *itup, int ntup, OffsetNumber oldoffnum,
228 : BlockNumber *newblkno,
229 : Buffer leftchildbuf,
230 : List **splitinfo,
231 : bool markfollowright,
232 : Relation heapRel,
233 : bool is_build)
234 : {
235 1724078 : BlockNumber blkno = BufferGetBlockNumber(buffer);
236 1724078 : Page page = BufferGetPage(buffer);
237 1724078 : bool is_leaf = (GistPageIsLeaf(page)) ? true : false;
238 : XLogRecPtr recptr;
239 : bool is_split;
240 :
241 : /*
242 : * Refuse to modify a page that's incompletely split. This should not
243 : * happen because we finish any incomplete splits while we walk down the
244 : * tree. However, it's remotely possible that another concurrent inserter
245 : * splits a parent page, and errors out before completing the split. We
246 : * will just throw an error in that case, and leave any split we had in
247 : * progress unfinished too. The next insert that comes along will clean up
248 : * the mess.
249 : */
250 1724078 : if (GistFollowRight(page))
251 0 : elog(ERROR, "concurrent GiST page split was incomplete");
252 :
253 : /* should never try to insert to a deleted page */
254 : Assert(!GistPageIsDeleted(page));
255 :
256 1724078 : *splitinfo = NIL;
257 :
258 : /*
259 : * if isupdate, remove old key: This node's key has been modified, either
260 : * because a child split occurred or because we needed to adjust our key
261 : * for an insert in a child node. Therefore, remove the old version of
262 : * this node's key.
263 : *
264 : * for WAL replay, in the non-split case we handle this by setting up a
265 : * one-element todelete array; in the split case, it's handled implicitly
266 : * because the tuple vector passed to gistSplit won't include this tuple.
267 : */
268 1724078 : is_split = gistnospace(page, itup, ntup, oldoffnum, freespace);
269 :
270 : /*
271 : * If leaf page is full, try at first to delete dead tuples. And then
272 : * check again.
273 : */
274 1724078 : if (is_split && GistPageIsLeaf(page) && GistPageHasGarbage(page))
275 : {
276 0 : gistprunepage(rel, page, buffer, heapRel);
277 0 : is_split = gistnospace(page, itup, ntup, oldoffnum, freespace);
278 : }
279 :
280 1724078 : if (is_split)
281 : {
282 : /* no space for insertion */
283 : IndexTuple *itvec;
284 : int tlen;
285 25304 : SplitPageLayout *dist = NULL,
286 : *ptr;
287 25304 : BlockNumber oldrlink = InvalidBlockNumber;
288 25304 : GistNSN oldnsn = 0;
289 : SplitPageLayout rootpg;
290 : bool is_rootsplit;
291 : int npage;
292 :
293 25304 : is_rootsplit = (blkno == GIST_ROOT_BLKNO);
294 :
295 : /*
296 : * Form index tuples vector to split. If we're replacing an old tuple,
297 : * remove the old version from the vector.
298 : */
299 25304 : itvec = gistextractpage(page, &tlen);
300 25304 : if (OffsetNumberIsValid(oldoffnum))
301 : {
302 : /* on inner page we should remove old tuple */
303 5214 : int pos = oldoffnum - FirstOffsetNumber;
304 :
305 5214 : tlen--;
306 5214 : if (pos != tlen)
307 3386 : memmove(itvec + pos, itvec + pos + 1, sizeof(IndexTuple) * (tlen - pos));
308 : }
309 25304 : itvec = gistjoinvector(itvec, &tlen, itup, ntup);
310 25304 : dist = gistSplit(rel, page, itvec, tlen, giststate);
311 :
312 : /*
313 : * Check that split didn't produce too many pages.
314 : */
315 25304 : npage = 0;
316 75986 : for (ptr = dist; ptr; ptr = ptr->next)
317 50682 : npage++;
318 : /* in a root split, we'll add one more page to the list below */
319 25304 : if (is_rootsplit)
320 408 : npage++;
321 25304 : if (npage > GIST_MAX_SPLIT_PAGES)
322 0 : elog(ERROR, "GiST page split into too many halves (%d, maximum %d)",
323 : npage, GIST_MAX_SPLIT_PAGES);
324 :
325 : /*
326 : * Set up pages to work with. Allocate new buffers for all but the
327 : * leftmost page. The original page becomes the new leftmost page, and
328 : * is just replaced with the new contents.
329 : *
330 : * For a root-split, allocate new buffers for all child pages, the
331 : * original page is overwritten with new root page containing
332 : * downlinks to the new child pages.
333 : */
334 25304 : ptr = dist;
335 25304 : if (!is_rootsplit)
336 : {
337 : /* save old rightlink and NSN */
338 24896 : oldrlink = GistPageGetOpaque(page)->rightlink;
339 24896 : oldnsn = GistPageGetNSN(page);
340 :
341 24896 : dist->buffer = buffer;
342 24896 : dist->block.blkno = BufferGetBlockNumber(buffer);
343 24896 : dist->page = PageGetTempPageCopySpecial(BufferGetPage(buffer));
344 :
345 : /* clean all flags except F_LEAF */
346 24896 : GistPageGetOpaque(dist->page)->flags = (is_leaf) ? F_LEAF : 0;
347 :
348 24896 : ptr = ptr->next;
349 : }
350 51090 : for (; ptr; ptr = ptr->next)
351 : {
352 : /* Allocate new page */
353 25786 : ptr->buffer = gistNewBuffer(rel, heapRel);
354 25786 : GISTInitBuffer(ptr->buffer, (is_leaf) ? F_LEAF : 0);
355 25786 : ptr->page = BufferGetPage(ptr->buffer);
356 25786 : ptr->block.blkno = BufferGetBlockNumber(ptr->buffer);
357 25786 : PredicateLockPageSplit(rel,
358 : BufferGetBlockNumber(buffer),
359 : BufferGetBlockNumber(ptr->buffer));
360 : }
361 :
362 : /*
363 : * Now that we know which blocks the new pages go to, set up downlink
364 : * tuples to point to them.
365 : */
366 75986 : for (ptr = dist; ptr; ptr = ptr->next)
367 : {
368 50682 : ItemPointerSetBlockNumber(&(ptr->itup->t_tid), ptr->block.blkno);
369 50682 : GistTupleSetValid(ptr->itup);
370 : }
371 :
372 : /*
373 : * If this is a root split, we construct the new root page with the
374 : * downlinks here directly, instead of requiring the caller to insert
375 : * them. Add the new root page to the list along with the child pages.
376 : */
377 25304 : if (is_rootsplit)
378 : {
379 : IndexTuple *downlinks;
380 408 : int ndownlinks = 0;
381 : int i;
382 :
383 408 : rootpg.buffer = buffer;
384 408 : rootpg.page = PageGetTempPageCopySpecial(BufferGetPage(rootpg.buffer));
385 408 : GistPageGetOpaque(rootpg.page)->flags = 0;
386 :
387 : /* Prepare a vector of all the downlinks */
388 1228 : for (ptr = dist; ptr; ptr = ptr->next)
389 820 : ndownlinks++;
390 408 : downlinks = palloc(sizeof(IndexTuple) * ndownlinks);
391 1228 : for (i = 0, ptr = dist; ptr; ptr = ptr->next)
392 820 : downlinks[i++] = ptr->itup;
393 :
394 408 : rootpg.block.blkno = GIST_ROOT_BLKNO;
395 408 : rootpg.block.num = ndownlinks;
396 408 : rootpg.list = gistfillitupvec(downlinks, ndownlinks,
397 : &(rootpg.lenlist));
398 408 : rootpg.itup = NULL;
399 :
400 408 : rootpg.next = dist;
401 408 : dist = &rootpg;
402 : }
403 : else
404 : {
405 : /* Prepare split-info to be returned to caller */
406 74758 : for (ptr = dist; ptr; ptr = ptr->next)
407 : {
408 49862 : GISTPageSplitInfo *si = palloc(sizeof(GISTPageSplitInfo));
409 :
410 49862 : si->buf = ptr->buffer;
411 49862 : si->downlink = ptr->itup;
412 49862 : *splitinfo = lappend(*splitinfo, si);
413 : }
414 : }
415 :
416 : /*
417 : * Fill all pages. All the pages are new, ie. freshly allocated empty
418 : * pages, or a temporary copy of the old page.
419 : */
420 76394 : for (ptr = dist; ptr; ptr = ptr->next)
421 : {
422 51090 : char *data = (char *) (ptr->list);
423 :
424 1874780 : for (int i = 0; i < ptr->block.num; i++)
425 : {
426 1823690 : IndexTuple thistup = (IndexTuple) data;
427 :
428 1823690 : if (PageAddItem(ptr->page, (Item) data, IndexTupleSize(thistup), i + FirstOffsetNumber, false, false) == InvalidOffsetNumber)
429 0 : elog(ERROR, "failed to add item to index page in \"%s\"", RelationGetRelationName(rel));
430 :
431 : /*
432 : * If this is the first inserted/updated tuple, let the caller
433 : * know which page it landed on.
434 : */
435 1823690 : if (newblkno && ItemPointerEquals(&thistup->t_tid, &(*itup)->t_tid))
436 774 : *newblkno = ptr->block.blkno;
437 :
438 1823690 : data += IndexTupleSize(thistup);
439 : }
440 :
441 : /* Set up rightlinks */
442 51090 : if (ptr->next && ptr->block.blkno != GIST_ROOT_BLKNO)
443 50756 : GistPageGetOpaque(ptr->page)->rightlink =
444 25378 : ptr->next->block.blkno;
445 : else
446 25712 : GistPageGetOpaque(ptr->page)->rightlink = oldrlink;
447 :
448 : /*
449 : * Mark the all but the right-most page with the follow-right
450 : * flag. It will be cleared as soon as the downlink is inserted
451 : * into the parent, but this ensures that if we error out before
452 : * that, the index is still consistent. (in buffering build mode,
453 : * any error will abort the index build anyway, so this is not
454 : * needed.)
455 : */
456 51090 : if (ptr->next && !is_rootsplit && markfollowright)
457 24198 : GistMarkFollowRight(ptr->page);
458 : else
459 26892 : GistClearFollowRight(ptr->page);
460 :
461 : /*
462 : * Copy the NSN of the original page to all pages. The
463 : * F_FOLLOW_RIGHT flags ensure that scans will follow the
464 : * rightlinks until the downlinks are inserted.
465 : */
466 51090 : GistPageSetNSN(ptr->page, oldnsn);
467 : }
468 :
469 : /*
470 : * gistXLogSplit() needs to WAL log a lot of pages, prepare WAL
471 : * insertion for that. NB: The number of pages and data segments
472 : * specified here must match the calculations in gistXLogSplit()!
473 : */
474 25304 : if (!is_build && RelationNeedsWAL(rel))
475 3460 : XLogEnsureRecordSpace(npage, 1 + npage * 2);
476 :
477 25304 : START_CRIT_SECTION();
478 :
479 : /*
480 : * Must mark buffers dirty before XLogInsert, even though we'll still
481 : * be changing their opaque fields below.
482 : */
483 76394 : for (ptr = dist; ptr; ptr = ptr->next)
484 51090 : MarkBufferDirty(ptr->buffer);
485 25304 : if (BufferIsValid(leftchildbuf))
486 5078 : MarkBufferDirty(leftchildbuf);
487 :
488 : /*
489 : * The first page in the chain was a temporary working copy meant to
490 : * replace the old page. Copy it over the old page.
491 : */
492 25304 : PageRestoreTempPage(dist->page, BufferGetPage(dist->buffer));
493 25304 : dist->page = BufferGetPage(dist->buffer);
494 :
495 : /*
496 : * Write the WAL record.
497 : *
498 : * If we're building a new index, however, we don't WAL-log changes
499 : * yet. The LSN-NSN interlock between parent and child requires that
500 : * LSNs never move backwards, so set the LSNs to a value that's
501 : * smaller than any real or fake unlogged LSN that might be generated
502 : * later. (There can't be any concurrent scans during index build, so
503 : * we don't need to be able to detect concurrent splits yet.)
504 : */
505 25304 : if (is_build)
506 21840 : recptr = GistBuildLSN;
507 : else
508 : {
509 3464 : if (RelationNeedsWAL(rel))
510 3460 : recptr = gistXLogSplit(is_leaf,
511 : dist, oldrlink, oldnsn, leftchildbuf,
512 : markfollowright);
513 : else
514 4 : recptr = gistGetFakeLSN(rel);
515 : }
516 :
517 76394 : for (ptr = dist; ptr; ptr = ptr->next)
518 51090 : PageSetLSN(ptr->page, recptr);
519 :
520 : /*
521 : * Return the new child buffers to the caller.
522 : *
523 : * If this was a root split, we've already inserted the downlink
524 : * pointers, in the form of a new root page. Therefore we can release
525 : * all the new buffers, and keep just the root page locked.
526 : */
527 25304 : if (is_rootsplit)
528 : {
529 1228 : for (ptr = dist->next; ptr; ptr = ptr->next)
530 820 : UnlockReleaseBuffer(ptr->buffer);
531 : }
532 : }
533 : else
534 : {
535 : /*
536 : * Enough space. We always get here if ntup==0.
537 : */
538 1698774 : START_CRIT_SECTION();
539 :
540 : /*
541 : * Delete old tuple if any, then insert new tuple(s) if any. If
542 : * possible, use the fast path of PageIndexTupleOverwrite.
543 : */
544 1698774 : if (OffsetNumberIsValid(oldoffnum))
545 : {
546 754014 : if (ntup == 1)
547 : {
548 : /* One-for-one replacement, so use PageIndexTupleOverwrite */
549 734214 : if (!PageIndexTupleOverwrite(page, oldoffnum, (Item) *itup,
550 734214 : IndexTupleSize(*itup)))
551 0 : elog(ERROR, "failed to add item to index page in \"%s\"",
552 : RelationGetRelationName(rel));
553 : }
554 : else
555 : {
556 : /* Delete old, then append new tuple(s) to page */
557 19800 : PageIndexTupleDelete(page, oldoffnum);
558 19800 : gistfillbuffer(page, itup, ntup, InvalidOffsetNumber);
559 : }
560 : }
561 : else
562 : {
563 : /* Just append new tuples at the end of the page */
564 944760 : gistfillbuffer(page, itup, ntup, InvalidOffsetNumber);
565 : }
566 :
567 1698774 : MarkBufferDirty(buffer);
568 :
569 1698774 : if (BufferIsValid(leftchildbuf))
570 19120 : MarkBufferDirty(leftchildbuf);
571 :
572 1698774 : if (is_build)
573 1204214 : recptr = GistBuildLSN;
574 : else
575 : {
576 494560 : if (RelationNeedsWAL(rel))
577 494482 : {
578 494482 : OffsetNumber ndeloffs = 0,
579 : deloffs[1];
580 :
581 494482 : if (OffsetNumberIsValid(oldoffnum))
582 : {
583 193708 : deloffs[0] = oldoffnum;
584 193708 : ndeloffs = 1;
585 : }
586 :
587 494482 : recptr = gistXLogUpdate(buffer,
588 : deloffs, ndeloffs, itup, ntup,
589 : leftchildbuf);
590 : }
591 : else
592 78 : recptr = gistGetFakeLSN(rel);
593 : }
594 1698774 : PageSetLSN(page, recptr);
595 :
596 1698774 : if (newblkno)
597 104370 : *newblkno = blkno;
598 : }
599 :
600 : /*
601 : * If we inserted the downlink for a child page, set NSN and clear
602 : * F_FOLLOW_RIGHT flag on the left child, so that concurrent scans know to
603 : * follow the rightlink if and only if they looked at the parent page
604 : * before we inserted the downlink.
605 : *
606 : * Note that we do this *after* writing the WAL record. That means that
607 : * the possible full page image in the WAL record does not include these
608 : * changes, and they must be replayed even if the page is restored from
609 : * the full page image. There's a chicken-and-egg problem: if we updated
610 : * the child pages first, we wouldn't know the recptr of the WAL record
611 : * we're about to write.
612 : */
613 1724078 : if (BufferIsValid(leftchildbuf))
614 : {
615 24198 : Page leftpg = BufferGetPage(leftchildbuf);
616 :
617 24198 : GistPageSetNSN(leftpg, recptr);
618 24198 : GistClearFollowRight(leftpg);
619 :
620 24198 : PageSetLSN(leftpg, recptr);
621 : }
622 :
623 1724078 : END_CRIT_SECTION();
624 :
625 1724078 : return is_split;
626 : }
627 :
628 : /*
629 : * Workhorse routine for doing insertion into a GiST index. Note that
630 : * this routine assumes it is invoked in a short-lived memory context,
631 : * so it does not bother releasing palloc'd allocations.
632 : */
633 : void
634 929362 : gistdoinsert(Relation r, IndexTuple itup, Size freespace,
635 : GISTSTATE *giststate, Relation heapRel, bool is_build)
636 : {
637 : ItemId iid;
638 : IndexTuple idxtuple;
639 : GISTInsertStack firststack;
640 : GISTInsertStack *stack;
641 : GISTInsertState state;
642 929362 : bool xlocked = false;
643 :
644 929362 : memset(&state, 0, sizeof(GISTInsertState));
645 929362 : state.freespace = freespace;
646 929362 : state.r = r;
647 929362 : state.heapRel = heapRel;
648 929362 : state.is_build = is_build;
649 :
650 : /* Start from the root */
651 929362 : firststack.blkno = GIST_ROOT_BLKNO;
652 929362 : firststack.lsn = 0;
653 929362 : firststack.retry_from_parent = false;
654 929362 : firststack.parent = NULL;
655 929362 : firststack.downlinkoffnum = InvalidOffsetNumber;
656 929362 : state.stack = stack = &firststack;
657 :
658 : /*
659 : * Walk down along the path of smallest penalty, updating the parent
660 : * pointers with the key we're inserting as we go. If we crash in the
661 : * middle, the tree is consistent, although the possible parent updates
662 : * were a waste.
663 : */
664 : for (;;)
665 : {
666 : /*
667 : * If we split an internal page while descending the tree, we have to
668 : * retry at the parent. (Normally, the LSN-NSN interlock below would
669 : * also catch this and cause us to retry. But LSNs are not updated
670 : * during index build.)
671 : */
672 2102794 : while (stack->retry_from_parent)
673 : {
674 190 : if (xlocked)
675 0 : LockBuffer(stack->buffer, GIST_UNLOCK);
676 190 : xlocked = false;
677 190 : ReleaseBuffer(stack->buffer);
678 190 : state.stack = stack = stack->parent;
679 : }
680 :
681 2102604 : if (XLogRecPtrIsInvalid(stack->lsn))
682 2102486 : stack->buffer = ReadBuffer(state.r, stack->blkno);
683 :
684 : /*
685 : * Be optimistic and grab shared lock first. Swap it for an exclusive
686 : * lock later if we need to update the page.
687 : */
688 2102604 : if (!xlocked)
689 : {
690 2102602 : LockBuffer(stack->buffer, GIST_SHARE);
691 2102602 : gistcheckpage(state.r, stack->buffer);
692 : }
693 :
694 2102604 : stack->page = (Page) BufferGetPage(stack->buffer);
695 2102604 : stack->lsn = xlocked ?
696 2102604 : PageGetLSN(stack->page) : BufferGetLSNAtomic(stack->buffer);
697 : Assert(!RelationNeedsWAL(state.r) || !XLogRecPtrIsInvalid(stack->lsn));
698 :
699 : /*
700 : * If this page was split but the downlink was never inserted to the
701 : * parent because the inserting backend crashed before doing that, fix
702 : * that now.
703 : */
704 2102604 : if (GistFollowRight(stack->page))
705 : {
706 0 : if (!xlocked)
707 : {
708 0 : LockBuffer(stack->buffer, GIST_UNLOCK);
709 0 : LockBuffer(stack->buffer, GIST_EXCLUSIVE);
710 0 : xlocked = true;
711 : /* someone might've completed the split when we unlocked */
712 0 : if (!GistFollowRight(stack->page))
713 0 : continue;
714 : }
715 0 : gistfixsplit(&state, giststate);
716 :
717 0 : UnlockReleaseBuffer(stack->buffer);
718 0 : xlocked = false;
719 0 : state.stack = stack = stack->parent;
720 0 : continue;
721 : }
722 :
723 3275760 : if ((stack->blkno != GIST_ROOT_BLKNO &&
724 1173156 : stack->parent->lsn < GistPageGetNSN(stack->page)) ||
725 2102604 : GistPageIsDeleted(stack->page))
726 : {
727 : /*
728 : * Concurrent split or page deletion detected. There's no
729 : * guarantee that the downlink for this page is consistent with
730 : * the tuple we're inserting anymore, so go back to parent and
731 : * rechoose the best child.
732 : */
733 0 : UnlockReleaseBuffer(stack->buffer);
734 0 : xlocked = false;
735 0 : state.stack = stack = stack->parent;
736 0 : continue;
737 : }
738 :
739 2102604 : if (!GistPageIsLeaf(stack->page))
740 : {
741 : /*
742 : * This is an internal page so continue to walk down the tree.
743 : * Find the child node that has the minimum insertion penalty.
744 : */
745 : BlockNumber childblkno;
746 : IndexTuple newtup;
747 : GISTInsertStack *item;
748 : OffsetNumber downlinkoffnum;
749 :
750 1173242 : downlinkoffnum = gistchoose(state.r, stack->page, itup, giststate);
751 1173242 : iid = PageGetItemId(stack->page, downlinkoffnum);
752 1173242 : idxtuple = (IndexTuple) PageGetItem(stack->page, iid);
753 1173242 : childblkno = ItemPointerGetBlockNumber(&(idxtuple->t_tid));
754 :
755 : /*
756 : * Check that it's not a leftover invalid tuple from pre-9.1
757 : */
758 1173242 : if (GistTupleIsInvalid(idxtuple))
759 0 : ereport(ERROR,
760 : (errmsg("index \"%s\" contains an inner tuple marked as invalid",
761 : RelationGetRelationName(r)),
762 : errdetail("This is caused by an incomplete page split at crash recovery before upgrading to PostgreSQL 9.1."),
763 : errhint("Please REINDEX it.")));
764 :
765 : /*
766 : * Check that the key representing the target child node is
767 : * consistent with the key we're inserting. Update it if it's not.
768 : */
769 1173242 : newtup = gistgetadjusted(state.r, idxtuple, itup, giststate);
770 1173242 : if (newtup)
771 : {
772 : /*
773 : * Swap shared lock for an exclusive one. Beware, the page may
774 : * change while we unlock/lock the page...
775 : */
776 665386 : if (!xlocked)
777 : {
778 665386 : LockBuffer(stack->buffer, GIST_UNLOCK);
779 665386 : LockBuffer(stack->buffer, GIST_EXCLUSIVE);
780 665386 : xlocked = true;
781 665386 : stack->page = (Page) BufferGetPage(stack->buffer);
782 :
783 665386 : if (PageGetLSN(stack->page) != stack->lsn)
784 : {
785 : /* the page was changed while we unlocked it, retry */
786 0 : continue;
787 : }
788 : }
789 :
790 : /*
791 : * Update the tuple.
792 : *
793 : * We still hold the lock after gistinserttuple(), but it
794 : * might have to split the page to make the updated tuple fit.
795 : * In that case the updated tuple might migrate to the other
796 : * half of the split, so we have to go back to the parent and
797 : * descend back to the half that's a better fit for the new
798 : * tuple.
799 : */
800 665386 : if (gistinserttuple(&state, stack, giststate, newtup,
801 : downlinkoffnum))
802 : {
803 : /*
804 : * If this was a root split, the root page continues to be
805 : * the parent and the updated tuple went to one of the
806 : * child pages, so we just need to retry from the root
807 : * page.
808 : */
809 118 : if (stack->blkno != GIST_ROOT_BLKNO)
810 : {
811 116 : UnlockReleaseBuffer(stack->buffer);
812 116 : xlocked = false;
813 116 : state.stack = stack = stack->parent;
814 : }
815 118 : continue;
816 : }
817 : }
818 1173124 : LockBuffer(stack->buffer, GIST_UNLOCK);
819 1173124 : xlocked = false;
820 :
821 : /* descend to the chosen child */
822 1173124 : item = (GISTInsertStack *) palloc0(sizeof(GISTInsertStack));
823 1173124 : item->blkno = childblkno;
824 1173124 : item->parent = stack;
825 1173124 : item->downlinkoffnum = downlinkoffnum;
826 1173124 : state.stack = stack = item;
827 : }
828 : else
829 : {
830 : /*
831 : * Leaf page. Insert the new key. We've already updated all the
832 : * parents on the way down, but we might have to split the page if
833 : * it doesn't fit. gistinserttuple() will take care of that.
834 : */
835 :
836 : /*
837 : * Swap shared lock for an exclusive one. Be careful, the page may
838 : * change while we unlock/lock the page...
839 : */
840 929362 : if (!xlocked)
841 : {
842 929362 : LockBuffer(stack->buffer, GIST_UNLOCK);
843 929362 : LockBuffer(stack->buffer, GIST_EXCLUSIVE);
844 929362 : xlocked = true;
845 929362 : stack->page = (Page) BufferGetPage(stack->buffer);
846 929362 : stack->lsn = PageGetLSN(stack->page);
847 :
848 929362 : if (stack->blkno == GIST_ROOT_BLKNO)
849 : {
850 : /*
851 : * the only page that can become inner instead of leaf is
852 : * the root page, so for root we should recheck it
853 : */
854 54398 : if (!GistPageIsLeaf(stack->page))
855 : {
856 : /*
857 : * very rare situation: during unlock/lock index with
858 : * number of pages = 1 was increased
859 : */
860 0 : LockBuffer(stack->buffer, GIST_UNLOCK);
861 0 : xlocked = false;
862 0 : continue;
863 : }
864 :
865 : /*
866 : * we don't need to check root split, because checking
867 : * leaf/inner is enough to recognize split for root
868 : */
869 : }
870 1749928 : else if ((GistFollowRight(stack->page) ||
871 874964 : stack->parent->lsn < GistPageGetNSN(stack->page)) ||
872 874964 : GistPageIsDeleted(stack->page))
873 : {
874 : /*
875 : * The page was split or deleted while we momentarily
876 : * unlocked the page. Go back to parent.
877 : */
878 0 : UnlockReleaseBuffer(stack->buffer);
879 0 : xlocked = false;
880 0 : state.stack = stack = stack->parent;
881 0 : continue;
882 : }
883 : }
884 :
885 : /* now state.stack->(page, buffer and blkno) points to leaf page */
886 :
887 929362 : gistinserttuple(&state, stack, giststate, itup,
888 : InvalidOffsetNumber);
889 929350 : LockBuffer(stack->buffer, GIST_UNLOCK);
890 :
891 : /* Release any pins we might still hold before exiting */
892 3031506 : for (; stack; stack = stack->parent)
893 2102156 : ReleaseBuffer(stack->buffer);
894 929350 : break;
895 : }
896 : }
897 929350 : }
898 :
899 : /*
900 : * Traverse the tree to find path from root page to specified "child" block.
901 : *
902 : * returns a new insertion stack, starting from the parent of "child", up
903 : * to the root. *downlinkoffnum is set to the offset of the downlink in the
904 : * direct parent of child.
905 : *
906 : * To prevent deadlocks, this should lock only one page at a time.
907 : */
908 : static GISTInsertStack *
909 0 : gistFindPath(Relation r, BlockNumber child, OffsetNumber *downlinkoffnum)
910 : {
911 : Page page;
912 : Buffer buffer;
913 : OffsetNumber i,
914 : maxoff;
915 : ItemId iid;
916 : IndexTuple idxtuple;
917 : List *fifo;
918 : GISTInsertStack *top,
919 : *ptr;
920 : BlockNumber blkno;
921 :
922 0 : top = (GISTInsertStack *) palloc0(sizeof(GISTInsertStack));
923 0 : top->blkno = GIST_ROOT_BLKNO;
924 0 : top->downlinkoffnum = InvalidOffsetNumber;
925 :
926 0 : fifo = list_make1(top);
927 0 : while (fifo != NIL)
928 : {
929 : /* Get next page to visit */
930 0 : top = linitial(fifo);
931 0 : fifo = list_delete_first(fifo);
932 :
933 0 : buffer = ReadBuffer(r, top->blkno);
934 0 : LockBuffer(buffer, GIST_SHARE);
935 0 : gistcheckpage(r, buffer);
936 0 : page = (Page) BufferGetPage(buffer);
937 :
938 0 : if (GistPageIsLeaf(page))
939 : {
940 : /*
941 : * Because we scan the index top-down, all the rest of the pages
942 : * in the queue must be leaf pages as well.
943 : */
944 0 : UnlockReleaseBuffer(buffer);
945 0 : break;
946 : }
947 :
948 : /* currently, internal pages are never deleted */
949 : Assert(!GistPageIsDeleted(page));
950 :
951 0 : top->lsn = BufferGetLSNAtomic(buffer);
952 :
953 : /*
954 : * If F_FOLLOW_RIGHT is set, the page to the right doesn't have a
955 : * downlink. This should not normally happen..
956 : */
957 0 : if (GistFollowRight(page))
958 0 : elog(ERROR, "concurrent GiST page split was incomplete");
959 :
960 0 : if (top->parent && top->parent->lsn < GistPageGetNSN(page) &&
961 0 : GistPageGetOpaque(page)->rightlink != InvalidBlockNumber /* sanity check */ )
962 : {
963 : /*
964 : * Page was split while we looked elsewhere. We didn't see the
965 : * downlink to the right page when we scanned the parent, so add
966 : * it to the queue now.
967 : *
968 : * Put the right page ahead of the queue, so that we visit it
969 : * next. That's important, because if this is the lowest internal
970 : * level, just above leaves, we might already have queued up some
971 : * leaf pages, and we assume that there can't be any non-leaf
972 : * pages behind leaf pages.
973 : */
974 0 : ptr = (GISTInsertStack *) palloc0(sizeof(GISTInsertStack));
975 0 : ptr->blkno = GistPageGetOpaque(page)->rightlink;
976 0 : ptr->downlinkoffnum = InvalidOffsetNumber;
977 0 : ptr->parent = top->parent;
978 :
979 0 : fifo = lcons(ptr, fifo);
980 : }
981 :
982 0 : maxoff = PageGetMaxOffsetNumber(page);
983 :
984 0 : for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
985 : {
986 0 : iid = PageGetItemId(page, i);
987 0 : idxtuple = (IndexTuple) PageGetItem(page, iid);
988 0 : blkno = ItemPointerGetBlockNumber(&(idxtuple->t_tid));
989 0 : if (blkno == child)
990 : {
991 : /* Found it! */
992 0 : UnlockReleaseBuffer(buffer);
993 0 : *downlinkoffnum = i;
994 0 : return top;
995 : }
996 : else
997 : {
998 : /* Append this child to the list of pages to visit later */
999 0 : ptr = (GISTInsertStack *) palloc0(sizeof(GISTInsertStack));
1000 0 : ptr->blkno = blkno;
1001 0 : ptr->downlinkoffnum = i;
1002 0 : ptr->parent = top;
1003 :
1004 0 : fifo = lappend(fifo, ptr);
1005 : }
1006 : }
1007 :
1008 0 : UnlockReleaseBuffer(buffer);
1009 : }
1010 :
1011 0 : elog(ERROR, "failed to re-find parent of a page in index \"%s\", block %u",
1012 : RelationGetRelationName(r), child);
1013 : return NULL; /* keep compiler quiet */
1014 : }
1015 :
1016 : /*
1017 : * Updates the stack so that child->parent is the correct parent of the
1018 : * child. child->parent must be exclusively locked on entry, and will
1019 : * remain so at exit, but it might not be the same page anymore.
1020 : */
1021 : static void
1022 24198 : gistFindCorrectParent(Relation r, GISTInsertStack *child, bool is_build)
1023 : {
1024 24198 : GISTInsertStack *parent = child->parent;
1025 : ItemId iid;
1026 : IndexTuple idxtuple;
1027 : OffsetNumber maxoff;
1028 : GISTInsertStack *ptr;
1029 :
1030 24198 : gistcheckpage(r, parent->buffer);
1031 24198 : parent->page = (Page) BufferGetPage(parent->buffer);
1032 24198 : maxoff = PageGetMaxOffsetNumber(parent->page);
1033 :
1034 : /* Check if the downlink is still where it was before */
1035 24198 : if (child->downlinkoffnum != InvalidOffsetNumber && child->downlinkoffnum <= maxoff)
1036 : {
1037 24198 : iid = PageGetItemId(parent->page, child->downlinkoffnum);
1038 24198 : idxtuple = (IndexTuple) PageGetItem(parent->page, iid);
1039 24198 : if (ItemPointerGetBlockNumber(&(idxtuple->t_tid)) == child->blkno)
1040 24198 : return; /* still there */
1041 : }
1042 :
1043 : /*
1044 : * The page has changed since we looked. During normal operation, every
1045 : * update of a page changes its LSN, so the LSN we memorized should have
1046 : * changed too. During index build, however, we don't WAL-log the changes
1047 : * until we have built the index, so the LSN doesn't change. There is no
1048 : * concurrent activity during index build, but we might have changed the
1049 : * parent ourselves.
1050 : */
1051 : Assert(parent->lsn != PageGetLSN(parent->page) || is_build);
1052 :
1053 : /*
1054 : * Scan the page to re-find the downlink. If the page was split, it might
1055 : * have moved to a different page, so follow the right links until we find
1056 : * it.
1057 : */
1058 : while (true)
1059 0 : {
1060 : OffsetNumber i;
1061 :
1062 0 : maxoff = PageGetMaxOffsetNumber(parent->page);
1063 0 : for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
1064 : {
1065 0 : iid = PageGetItemId(parent->page, i);
1066 0 : idxtuple = (IndexTuple) PageGetItem(parent->page, iid);
1067 0 : if (ItemPointerGetBlockNumber(&(idxtuple->t_tid)) == child->blkno)
1068 : {
1069 : /* yes!!, found */
1070 0 : child->downlinkoffnum = i;
1071 0 : return;
1072 : }
1073 : }
1074 :
1075 0 : parent->blkno = GistPageGetOpaque(parent->page)->rightlink;
1076 0 : parent->downlinkoffnum = InvalidOffsetNumber;
1077 0 : UnlockReleaseBuffer(parent->buffer);
1078 0 : if (parent->blkno == InvalidBlockNumber)
1079 : {
1080 : /*
1081 : * End of chain and still didn't find parent. It's a very-very
1082 : * rare situation when the root was split.
1083 : */
1084 0 : break;
1085 : }
1086 0 : parent->buffer = ReadBuffer(r, parent->blkno);
1087 0 : LockBuffer(parent->buffer, GIST_EXCLUSIVE);
1088 0 : gistcheckpage(r, parent->buffer);
1089 0 : parent->page = (Page) BufferGetPage(parent->buffer);
1090 : }
1091 :
1092 : /*
1093 : * awful!!, we need search tree to find parent ... , but before we should
1094 : * release all old parent
1095 : */
1096 :
1097 0 : ptr = child->parent->parent; /* child->parent already released above */
1098 0 : while (ptr)
1099 : {
1100 0 : ReleaseBuffer(ptr->buffer);
1101 0 : ptr = ptr->parent;
1102 : }
1103 :
1104 : /* ok, find new path */
1105 0 : ptr = parent = gistFindPath(r, child->blkno, &child->downlinkoffnum);
1106 :
1107 : /* read all buffers as expected by caller */
1108 : /* note we don't lock them or gistcheckpage them here! */
1109 0 : while (ptr)
1110 : {
1111 0 : ptr->buffer = ReadBuffer(r, ptr->blkno);
1112 0 : ptr->page = (Page) BufferGetPage(ptr->buffer);
1113 0 : ptr = ptr->parent;
1114 : }
1115 :
1116 : /* install new chain of parents to stack */
1117 0 : child->parent = parent;
1118 :
1119 : /* make recursive call to normal processing */
1120 0 : LockBuffer(child->parent->buffer, GIST_EXCLUSIVE);
1121 0 : gistFindCorrectParent(r, child, is_build);
1122 : }
1123 :
1124 : /*
1125 : * Form a downlink pointer for the page in 'buf'.
1126 : */
1127 : static IndexTuple
1128 0 : gistformdownlink(Relation rel, Buffer buf, GISTSTATE *giststate,
1129 : GISTInsertStack *stack, bool is_build)
1130 : {
1131 0 : Page page = BufferGetPage(buf);
1132 : OffsetNumber maxoff;
1133 : OffsetNumber offset;
1134 0 : IndexTuple downlink = NULL;
1135 :
1136 0 : maxoff = PageGetMaxOffsetNumber(page);
1137 0 : for (offset = FirstOffsetNumber; offset <= maxoff; offset = OffsetNumberNext(offset))
1138 : {
1139 : IndexTuple ituple = (IndexTuple)
1140 0 : PageGetItem(page, PageGetItemId(page, offset));
1141 :
1142 0 : if (downlink == NULL)
1143 0 : downlink = CopyIndexTuple(ituple);
1144 : else
1145 : {
1146 : IndexTuple newdownlink;
1147 :
1148 0 : newdownlink = gistgetadjusted(rel, downlink, ituple,
1149 : giststate);
1150 0 : if (newdownlink)
1151 0 : downlink = newdownlink;
1152 : }
1153 : }
1154 :
1155 : /*
1156 : * If the page is completely empty, we can't form a meaningful downlink
1157 : * for it. But we have to insert a downlink for the page. Any key will do,
1158 : * as long as its consistent with the downlink of parent page, so that we
1159 : * can legally insert it to the parent. A minimal one that matches as few
1160 : * scans as possible would be best, to keep scans from doing useless work,
1161 : * but we don't know how to construct that. So we just use the downlink of
1162 : * the original page that was split - that's as far from optimal as it can
1163 : * get but will do..
1164 : */
1165 0 : if (!downlink)
1166 : {
1167 : ItemId iid;
1168 :
1169 0 : LockBuffer(stack->parent->buffer, GIST_EXCLUSIVE);
1170 0 : gistFindCorrectParent(rel, stack, is_build);
1171 0 : iid = PageGetItemId(stack->parent->page, stack->downlinkoffnum);
1172 0 : downlink = (IndexTuple) PageGetItem(stack->parent->page, iid);
1173 0 : downlink = CopyIndexTuple(downlink);
1174 0 : LockBuffer(stack->parent->buffer, GIST_UNLOCK);
1175 : }
1176 :
1177 0 : ItemPointerSetBlockNumber(&(downlink->t_tid), BufferGetBlockNumber(buf));
1178 0 : GistTupleSetValid(downlink);
1179 :
1180 0 : return downlink;
1181 : }
1182 :
1183 :
1184 : /*
1185 : * Complete the incomplete split of state->stack->page.
1186 : */
1187 : static void
1188 0 : gistfixsplit(GISTInsertState *state, GISTSTATE *giststate)
1189 : {
1190 0 : GISTInsertStack *stack = state->stack;
1191 : Buffer buf;
1192 : Page page;
1193 0 : List *splitinfo = NIL;
1194 :
1195 0 : ereport(LOG,
1196 : (errmsg("fixing incomplete split in index \"%s\", block %u",
1197 : RelationGetRelationName(state->r), stack->blkno)));
1198 :
1199 : Assert(GistFollowRight(stack->page));
1200 : Assert(OffsetNumberIsValid(stack->downlinkoffnum));
1201 :
1202 0 : buf = stack->buffer;
1203 :
1204 : /*
1205 : * Read the chain of split pages, following the rightlinks. Construct a
1206 : * downlink tuple for each page.
1207 : */
1208 : for (;;)
1209 0 : {
1210 0 : GISTPageSplitInfo *si = palloc(sizeof(GISTPageSplitInfo));
1211 : IndexTuple downlink;
1212 :
1213 0 : page = BufferGetPage(buf);
1214 :
1215 : /* Form the new downlink tuples to insert to parent */
1216 0 : downlink = gistformdownlink(state->r, buf, giststate, stack, state->is_build);
1217 :
1218 0 : si->buf = buf;
1219 0 : si->downlink = downlink;
1220 :
1221 0 : splitinfo = lappend(splitinfo, si);
1222 :
1223 0 : if (GistFollowRight(page))
1224 : {
1225 : /* lock next page */
1226 0 : buf = ReadBuffer(state->r, GistPageGetOpaque(page)->rightlink);
1227 0 : LockBuffer(buf, GIST_EXCLUSIVE);
1228 : }
1229 : else
1230 0 : break;
1231 : }
1232 :
1233 : /* Insert the downlinks */
1234 0 : gistfinishsplit(state, stack, giststate, splitinfo, false);
1235 0 : }
1236 :
1237 : /*
1238 : * Insert or replace a tuple in stack->buffer. If 'oldoffnum' is valid, the
1239 : * tuple at 'oldoffnum' is replaced, otherwise the tuple is inserted as new.
1240 : * 'stack' represents the path from the root to the page being updated.
1241 : *
1242 : * The caller must hold an exclusive lock on stack->buffer. The lock is still
1243 : * held on return, but the page might not contain the inserted tuple if the
1244 : * page was split. The function returns true if the page was split, false
1245 : * otherwise.
1246 : */
1247 : static bool
1248 1594748 : gistinserttuple(GISTInsertState *state, GISTInsertStack *stack,
1249 : GISTSTATE *giststate, IndexTuple tuple, OffsetNumber oldoffnum)
1250 : {
1251 1594748 : return gistinserttuples(state, stack, giststate, &tuple, 1, oldoffnum,
1252 : InvalidBuffer, InvalidBuffer, false, false);
1253 : }
1254 :
1255 : /* ----------------
1256 : * An extended workhorse version of gistinserttuple(). This version allows
1257 : * inserting multiple tuples, or replacing a single tuple with multiple tuples.
1258 : * This is used to recursively update the downlinks in the parent when a page
1259 : * is split.
1260 : *
1261 : * If leftchild and rightchild are valid, we're inserting/replacing the
1262 : * downlink for rightchild, and leftchild is its left sibling. We clear the
1263 : * F_FOLLOW_RIGHT flag and update NSN on leftchild, atomically with the
1264 : * insertion of the downlink.
1265 : *
1266 : * To avoid holding locks for longer than necessary, when recursing up the
1267 : * tree to update the parents, the locking is a bit peculiar here. On entry,
1268 : * the caller must hold an exclusive lock on stack->buffer, as well as
1269 : * leftchild and rightchild if given. On return:
1270 : *
1271 : * - Lock on stack->buffer is released, if 'unlockbuf' is true. The page is
1272 : * always kept pinned, however.
1273 : * - Lock on 'leftchild' is released, if 'unlockleftchild' is true. The page
1274 : * is kept pinned.
1275 : * - Lock and pin on 'rightchild' are always released.
1276 : *
1277 : * Returns 'true' if the page had to be split. Note that if the page was
1278 : * split, the inserted/updated tuples might've been inserted to a right
1279 : * sibling of stack->buffer instead of stack->buffer itself.
1280 : */
1281 : static bool
1282 1618946 : gistinserttuples(GISTInsertState *state, GISTInsertStack *stack,
1283 : GISTSTATE *giststate,
1284 : IndexTuple *tuples, int ntup, OffsetNumber oldoffnum,
1285 : Buffer leftchild, Buffer rightchild,
1286 : bool unlockbuf, bool unlockleftchild)
1287 : {
1288 : List *splitinfo;
1289 : bool is_split;
1290 :
1291 : /*
1292 : * Check for any rw conflicts (in serializable isolation level) just
1293 : * before we intend to modify the page
1294 : */
1295 1618946 : CheckForSerializableConflictIn(state->r, NULL, BufferGetBlockNumber(stack->buffer));
1296 :
1297 : /* Insert the tuple(s) to the page, splitting the page if necessary */
1298 1618934 : is_split = gistplacetopage(state->r, state->freespace, giststate,
1299 : stack->buffer,
1300 : tuples, ntup,
1301 : oldoffnum, NULL,
1302 : leftchild,
1303 : &splitinfo,
1304 : true,
1305 : state->heapRel,
1306 1618934 : state->is_build);
1307 :
1308 : /*
1309 : * Before recursing up in case the page was split, release locks on the
1310 : * child pages. We don't need to keep them locked when updating the
1311 : * parent.
1312 : */
1313 1618934 : if (BufferIsValid(rightchild))
1314 24198 : UnlockReleaseBuffer(rightchild);
1315 1618934 : if (BufferIsValid(leftchild) && unlockleftchild)
1316 4966 : LockBuffer(leftchild, GIST_UNLOCK);
1317 :
1318 : /*
1319 : * If we had to split, insert/update the downlinks in the parent. If the
1320 : * caller requested us to release the lock on stack->buffer, tell
1321 : * gistfinishsplit() to do that as soon as it's safe to do so. If we
1322 : * didn't have to split, release it ourselves.
1323 : */
1324 1618934 : if (splitinfo)
1325 24128 : gistfinishsplit(state, stack, giststate, splitinfo, unlockbuf);
1326 1594806 : else if (unlockbuf)
1327 19162 : LockBuffer(stack->buffer, GIST_UNLOCK);
1328 :
1329 1618934 : return is_split;
1330 : }
1331 :
1332 : /*
1333 : * Finish an incomplete split by inserting/updating the downlinks in parent
1334 : * page. 'splitinfo' contains all the child pages involved in the split,
1335 : * from left-to-right.
1336 : *
1337 : * On entry, the caller must hold a lock on stack->buffer and all the child
1338 : * pages in 'splitinfo'. If 'unlockbuf' is true, the lock on stack->buffer is
1339 : * released on return. The child pages are always unlocked and unpinned.
1340 : */
1341 : static void
1342 24128 : gistfinishsplit(GISTInsertState *state, GISTInsertStack *stack,
1343 : GISTSTATE *giststate, List *splitinfo, bool unlockbuf)
1344 : {
1345 : GISTPageSplitInfo *right;
1346 : GISTPageSplitInfo *left;
1347 : IndexTuple tuples[2];
1348 :
1349 : /* A split always contains at least two halves */
1350 : Assert(list_length(splitinfo) >= 2);
1351 :
1352 : /*
1353 : * We need to insert downlinks for each new page, and update the downlink
1354 : * for the original (leftmost) page in the split. Begin at the rightmost
1355 : * page, inserting one downlink at a time until there's only two pages
1356 : * left. Finally insert the downlink for the last new page and update the
1357 : * downlink for the original page as one operation.
1358 : */
1359 24128 : LockBuffer(stack->parent->buffer, GIST_EXCLUSIVE);
1360 :
1361 : /*
1362 : * Insert downlinks for the siblings from right to left, until there are
1363 : * only two siblings left.
1364 : */
1365 24198 : for (int pos = list_length(splitinfo) - 1; pos > 1; pos--)
1366 : {
1367 70 : right = (GISTPageSplitInfo *) list_nth(splitinfo, pos);
1368 70 : left = (GISTPageSplitInfo *) list_nth(splitinfo, pos - 1);
1369 :
1370 70 : gistFindCorrectParent(state->r, stack, state->is_build);
1371 70 : if (gistinserttuples(state, stack->parent, giststate,
1372 : &right->downlink, 1,
1373 : InvalidOffsetNumber,
1374 : left->buf, right->buf, false, false))
1375 : {
1376 : /*
1377 : * If the parent page was split, the existing downlink might have
1378 : * moved.
1379 : */
1380 0 : stack->downlinkoffnum = InvalidOffsetNumber;
1381 : }
1382 : /* gistinserttuples() released the lock on right->buf. */
1383 : }
1384 :
1385 24128 : right = (GISTPageSplitInfo *) lsecond(splitinfo);
1386 24128 : left = (GISTPageSplitInfo *) linitial(splitinfo);
1387 :
1388 : /*
1389 : * Finally insert downlink for the remaining right page and update the
1390 : * downlink for the original page to not contain the tuples that were
1391 : * moved to the new pages.
1392 : */
1393 24128 : tuples[0] = left->downlink;
1394 24128 : tuples[1] = right->downlink;
1395 24128 : gistFindCorrectParent(state->r, stack, state->is_build);
1396 24128 : (void) gistinserttuples(state, stack->parent, giststate,
1397 : tuples, 2,
1398 24128 : stack->downlinkoffnum,
1399 : left->buf, right->buf,
1400 : true, /* Unlock parent */
1401 : unlockbuf /* Unlock stack->buffer if caller
1402 : * wants that */
1403 : );
1404 :
1405 : /*
1406 : * The downlink might have moved when we updated it. Even if the page
1407 : * wasn't split, because gistinserttuples() implements updating the old
1408 : * tuple by removing and re-inserting it!
1409 : */
1410 24128 : stack->downlinkoffnum = InvalidOffsetNumber;
1411 :
1412 : Assert(left->buf == stack->buffer);
1413 :
1414 : /*
1415 : * If we split the page because we had to adjust the downlink on an
1416 : * internal page, while descending the tree for inserting a new tuple,
1417 : * then this might no longer be the correct page for the new tuple. The
1418 : * downlink to this page might not cover the new tuple anymore, it might
1419 : * need to go to the newly-created right sibling instead. Tell the caller
1420 : * to walk back up the stack, to re-check at the parent which page to
1421 : * insert to.
1422 : *
1423 : * Normally, the LSN-NSN interlock during the tree descend would also
1424 : * detect that a concurrent split happened (by ourselves), and cause us to
1425 : * retry at the parent. But that mechanism doesn't work during index
1426 : * build, because we don't do WAL-logging, and don't update LSNs, during
1427 : * index build.
1428 : */
1429 24128 : stack->retry_from_parent = true;
1430 24128 : }
1431 :
1432 : /*
1433 : * gistSplit -- split a page in the tree and fill struct
1434 : * used for XLOG and real writes buffers. Function is recursive, ie
1435 : * it will split page until keys will fit in every page.
1436 : */
1437 : SplitPageLayout *
1438 26508 : gistSplit(Relation r,
1439 : Page page,
1440 : IndexTuple *itup, /* contains compressed entry */
1441 : int len,
1442 : GISTSTATE *giststate)
1443 : {
1444 : IndexTuple *lvectup,
1445 : *rvectup;
1446 : GistSplitVector v;
1447 : int i;
1448 26508 : SplitPageLayout *res = NULL;
1449 :
1450 : /* this should never recurse very deeply, but better safe than sorry */
1451 26508 : check_stack_depth();
1452 :
1453 : /* there's no point in splitting an empty page */
1454 : Assert(len > 0);
1455 :
1456 : /*
1457 : * If a single tuple doesn't fit on a page, no amount of splitting will
1458 : * help.
1459 : */
1460 26508 : if (len == 1)
1461 0 : ereport(ERROR,
1462 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1463 : errmsg("index row size %zu exceeds maximum %zu for index \"%s\"",
1464 : IndexTupleSize(itup[0]), GiSTPageSize,
1465 : RelationGetRelationName(r))));
1466 :
1467 26508 : memset(v.spl_lisnull, true,
1468 26508 : sizeof(bool) * giststate->nonLeafTupdesc->natts);
1469 26508 : memset(v.spl_risnull, true,
1470 26508 : sizeof(bool) * giststate->nonLeafTupdesc->natts);
1471 26508 : gistSplitByKey(r, page, itup, len, giststate, &v, 0);
1472 :
1473 : /* form left and right vector */
1474 26508 : lvectup = (IndexTuple *) palloc(sizeof(IndexTuple) * (len + 1));
1475 26508 : rvectup = (IndexTuple *) palloc(sizeof(IndexTuple) * (len + 1));
1476 :
1477 1096948 : for (i = 0; i < v.splitVector.spl_nleft; i++)
1478 1070440 : lvectup[i] = itup[v.splitVector.spl_left[i] - 1];
1479 :
1480 1227658 : for (i = 0; i < v.splitVector.spl_nright; i++)
1481 1201150 : rvectup[i] = itup[v.splitVector.spl_right[i] - 1];
1482 :
1483 : /* finalize splitting (may need another split) */
1484 26508 : if (!gistfitpage(rvectup, v.splitVector.spl_nright))
1485 : {
1486 592 : res = gistSplit(r, page, rvectup, v.splitVector.spl_nright, giststate);
1487 : }
1488 : else
1489 : {
1490 25916 : ROTATEDIST(res);
1491 25916 : res->block.num = v.splitVector.spl_nright;
1492 25916 : res->list = gistfillitupvec(rvectup, v.splitVector.spl_nright, &(res->lenlist));
1493 25916 : res->itup = gistFormTuple(giststate, r, v.spl_rattr, v.spl_risnull, false);
1494 : }
1495 :
1496 26508 : if (!gistfitpage(lvectup, v.splitVector.spl_nleft))
1497 : {
1498 : SplitPageLayout *resptr,
1499 : *subres;
1500 :
1501 334 : resptr = subres = gistSplit(r, page, lvectup, v.splitVector.spl_nleft, giststate);
1502 :
1503 : /* install on list's tail */
1504 830 : while (resptr->next)
1505 496 : resptr = resptr->next;
1506 :
1507 334 : resptr->next = res;
1508 334 : res = subres;
1509 : }
1510 : else
1511 : {
1512 26174 : ROTATEDIST(res);
1513 26174 : res->block.num = v.splitVector.spl_nleft;
1514 26174 : res->list = gistfillitupvec(lvectup, v.splitVector.spl_nleft, &(res->lenlist));
1515 26174 : res->itup = gistFormTuple(giststate, r, v.spl_lattr, v.spl_lisnull, false);
1516 : }
1517 :
1518 26508 : return res;
1519 : }
1520 :
1521 : /*
1522 : * Create a GISTSTATE and fill it with information about the index
1523 : */
1524 : GISTSTATE *
1525 9692 : initGISTstate(Relation index)
1526 : {
1527 : GISTSTATE *giststate;
1528 : MemoryContext scanCxt;
1529 : MemoryContext oldCxt;
1530 : int i;
1531 :
1532 : /* safety check to protect fixed-size arrays in GISTSTATE */
1533 9692 : if (index->rd_att->natts > INDEX_MAX_KEYS)
1534 0 : elog(ERROR, "numberOfAttributes %d > %d",
1535 : index->rd_att->natts, INDEX_MAX_KEYS);
1536 :
1537 : /* Create the memory context that will hold the GISTSTATE */
1538 9692 : scanCxt = AllocSetContextCreate(CurrentMemoryContext,
1539 : "GiST scan context",
1540 : ALLOCSET_DEFAULT_SIZES);
1541 9692 : oldCxt = MemoryContextSwitchTo(scanCxt);
1542 :
1543 : /* Create and fill in the GISTSTATE */
1544 9692 : giststate = (GISTSTATE *) palloc(sizeof(GISTSTATE));
1545 :
1546 9692 : giststate->scanCxt = scanCxt;
1547 9692 : giststate->tempCxt = scanCxt; /* caller must change this if needed */
1548 9692 : giststate->leafTupdesc = index->rd_att;
1549 :
1550 : /*
1551 : * The truncated tupdesc for non-leaf index tuples, which doesn't contain
1552 : * the INCLUDE attributes.
1553 : *
1554 : * It is used to form tuples during tuple adjustment and page split.
1555 : * B-tree creates shortened tuple descriptor for every truncated tuple,
1556 : * because it is doing this less often: it does not have to form truncated
1557 : * tuples during page split. Also, B-tree is not adjusting tuples on
1558 : * internal pages the way GiST does.
1559 : */
1560 9692 : giststate->nonLeafTupdesc = CreateTupleDescCopyConstr(index->rd_att);
1561 9692 : giststate->nonLeafTupdesc->natts =
1562 9692 : IndexRelationGetNumberOfKeyAttributes(index);
1563 :
1564 25834 : for (i = 0; i < IndexRelationGetNumberOfKeyAttributes(index); i++)
1565 : {
1566 16142 : fmgr_info_copy(&(giststate->consistentFn[i]),
1567 16142 : index_getprocinfo(index, i + 1, GIST_CONSISTENT_PROC),
1568 : scanCxt);
1569 16142 : fmgr_info_copy(&(giststate->unionFn[i]),
1570 16142 : index_getprocinfo(index, i + 1, GIST_UNION_PROC),
1571 : scanCxt);
1572 :
1573 : /* opclasses are not required to provide a Compress method */
1574 16142 : if (OidIsValid(index_getprocid(index, i + 1, GIST_COMPRESS_PROC)))
1575 5778 : fmgr_info_copy(&(giststate->compressFn[i]),
1576 5778 : index_getprocinfo(index, i + 1, GIST_COMPRESS_PROC),
1577 : scanCxt);
1578 : else
1579 10364 : giststate->compressFn[i].fn_oid = InvalidOid;
1580 :
1581 : /* opclasses are not required to provide a Decompress method */
1582 16142 : if (OidIsValid(index_getprocid(index, i + 1, GIST_DECOMPRESS_PROC)))
1583 1594 : fmgr_info_copy(&(giststate->decompressFn[i]),
1584 1594 : index_getprocinfo(index, i + 1, GIST_DECOMPRESS_PROC),
1585 : scanCxt);
1586 : else
1587 14548 : giststate->decompressFn[i].fn_oid = InvalidOid;
1588 :
1589 16142 : fmgr_info_copy(&(giststate->penaltyFn[i]),
1590 16142 : index_getprocinfo(index, i + 1, GIST_PENALTY_PROC),
1591 : scanCxt);
1592 16142 : fmgr_info_copy(&(giststate->picksplitFn[i]),
1593 16142 : index_getprocinfo(index, i + 1, GIST_PICKSPLIT_PROC),
1594 : scanCxt);
1595 16142 : fmgr_info_copy(&(giststate->equalFn[i]),
1596 16142 : index_getprocinfo(index, i + 1, GIST_EQUAL_PROC),
1597 : scanCxt);
1598 :
1599 : /* opclasses are not required to provide a Distance method */
1600 16142 : if (OidIsValid(index_getprocid(index, i + 1, GIST_DISTANCE_PROC)))
1601 1894 : fmgr_info_copy(&(giststate->distanceFn[i]),
1602 1894 : index_getprocinfo(index, i + 1, GIST_DISTANCE_PROC),
1603 : scanCxt);
1604 : else
1605 14248 : giststate->distanceFn[i].fn_oid = InvalidOid;
1606 :
1607 : /* opclasses are not required to provide a Fetch method */
1608 16142 : if (OidIsValid(index_getprocid(index, i + 1, GIST_FETCH_PROC)))
1609 1358 : fmgr_info_copy(&(giststate->fetchFn[i]),
1610 1358 : index_getprocinfo(index, i + 1, GIST_FETCH_PROC),
1611 : scanCxt);
1612 : else
1613 14784 : giststate->fetchFn[i].fn_oid = InvalidOid;
1614 :
1615 : /*
1616 : * If the index column has a specified collation, we should honor that
1617 : * while doing comparisons. However, we may have a collatable storage
1618 : * type for a noncollatable indexed data type. If there's no index
1619 : * collation then specify default collation in case the support
1620 : * functions need collation. This is harmless if the support
1621 : * functions don't care about collation, so we just do it
1622 : * unconditionally. (We could alternatively call get_typcollation,
1623 : * but that seems like expensive overkill --- there aren't going to be
1624 : * any cases where a GiST storage type has a nondefault collation.)
1625 : */
1626 16142 : if (OidIsValid(index->rd_indcollation[i]))
1627 194 : giststate->supportCollation[i] = index->rd_indcollation[i];
1628 : else
1629 15948 : giststate->supportCollation[i] = DEFAULT_COLLATION_OID;
1630 : }
1631 :
1632 : /* No opclass information for INCLUDE attributes */
1633 10194 : for (; i < index->rd_att->natts; i++)
1634 : {
1635 502 : giststate->consistentFn[i].fn_oid = InvalidOid;
1636 502 : giststate->unionFn[i].fn_oid = InvalidOid;
1637 502 : giststate->compressFn[i].fn_oid = InvalidOid;
1638 502 : giststate->decompressFn[i].fn_oid = InvalidOid;
1639 502 : giststate->penaltyFn[i].fn_oid = InvalidOid;
1640 502 : giststate->picksplitFn[i].fn_oid = InvalidOid;
1641 502 : giststate->equalFn[i].fn_oid = InvalidOid;
1642 502 : giststate->distanceFn[i].fn_oid = InvalidOid;
1643 502 : giststate->fetchFn[i].fn_oid = InvalidOid;
1644 502 : giststate->supportCollation[i] = InvalidOid;
1645 : }
1646 :
1647 9692 : MemoryContextSwitchTo(oldCxt);
1648 :
1649 9692 : return giststate;
1650 : }
1651 :
1652 : void
1653 7454 : freeGISTstate(GISTSTATE *giststate)
1654 : {
1655 : /* It's sufficient to delete the scanCxt */
1656 7454 : MemoryContextDelete(giststate->scanCxt);
1657 7454 : }
1658 :
1659 : /*
1660 : * gistprunepage() -- try to remove LP_DEAD items from the given page.
1661 : * Function assumes that buffer is exclusively locked.
1662 : */
1663 : static void
1664 0 : gistprunepage(Relation rel, Page page, Buffer buffer, Relation heapRel)
1665 : {
1666 : OffsetNumber deletable[MaxIndexTuplesPerPage];
1667 0 : int ndeletable = 0;
1668 : OffsetNumber offnum,
1669 : maxoff;
1670 :
1671 : Assert(GistPageIsLeaf(page));
1672 :
1673 : /*
1674 : * Scan over all items to see which ones need to be deleted according to
1675 : * LP_DEAD flags.
1676 : */
1677 0 : maxoff = PageGetMaxOffsetNumber(page);
1678 0 : for (offnum = FirstOffsetNumber;
1679 : offnum <= maxoff;
1680 0 : offnum = OffsetNumberNext(offnum))
1681 : {
1682 0 : ItemId itemId = PageGetItemId(page, offnum);
1683 :
1684 0 : if (ItemIdIsDead(itemId))
1685 0 : deletable[ndeletable++] = offnum;
1686 : }
1687 :
1688 0 : if (ndeletable > 0)
1689 : {
1690 0 : TransactionId snapshotConflictHorizon = InvalidTransactionId;
1691 :
1692 0 : if (XLogStandbyInfoActive() && RelationNeedsWAL(rel))
1693 : snapshotConflictHorizon =
1694 0 : index_compute_xid_horizon_for_tuples(rel, heapRel, buffer,
1695 : deletable, ndeletable);
1696 :
1697 0 : START_CRIT_SECTION();
1698 :
1699 0 : PageIndexMultiDelete(page, deletable, ndeletable);
1700 :
1701 : /*
1702 : * Mark the page as not containing any LP_DEAD items. This is not
1703 : * certainly true (there might be some that have recently been marked,
1704 : * but weren't included in our target-item list), but it will almost
1705 : * always be true and it doesn't seem worth an additional page scan to
1706 : * check it. Remember that F_HAS_GARBAGE is only a hint anyway.
1707 : */
1708 0 : GistClearPageHasGarbage(page);
1709 :
1710 0 : MarkBufferDirty(buffer);
1711 :
1712 : /* XLOG stuff */
1713 0 : if (RelationNeedsWAL(rel))
1714 0 : {
1715 : XLogRecPtr recptr;
1716 :
1717 0 : recptr = gistXLogDelete(buffer,
1718 : deletable, ndeletable,
1719 : snapshotConflictHorizon,
1720 : heapRel);
1721 :
1722 0 : PageSetLSN(page, recptr);
1723 : }
1724 : else
1725 0 : PageSetLSN(page, gistGetFakeLSN(rel));
1726 :
1727 0 : END_CRIT_SECTION();
1728 : }
1729 :
1730 : /*
1731 : * Note: if we didn't find any LP_DEAD items, then the page's
1732 : * F_HAS_GARBAGE hint bit is falsely set. We do not bother expending a
1733 : * separate write to clear it, however. We will clear it when we split
1734 : * the page.
1735 : */
1736 0 : }
|
