Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * nbtinsert.c
4 : * Item insertion in Lehman and Yao btrees for Postgres.
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/access/nbtree/nbtinsert.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 :
16 : #include "postgres.h"
17 :
18 : #include "access/nbtree.h"
19 : #include "access/nbtxlog.h"
20 : #include "access/tableam.h"
21 : #include "access/transam.h"
22 : #include "access/xloginsert.h"
23 : #include "common/int.h"
24 : #include "common/pg_prng.h"
25 : #include "lib/qunique.h"
26 : #include "miscadmin.h"
27 : #include "storage/lmgr.h"
28 : #include "storage/predicate.h"
29 :
30 : /* Minimum tree height for application of fastpath optimization */
31 : #define BTREE_FASTPATH_MIN_LEVEL 2
32 :
33 :
34 : static BTStack _bt_search_insert(Relation rel, Relation heaprel,
35 : BTInsertState insertstate);
36 : static TransactionId _bt_check_unique(Relation rel, BTInsertState insertstate,
37 : Relation heapRel,
38 : IndexUniqueCheck checkUnique, bool *is_unique,
39 : uint32 *speculativeToken);
40 : static OffsetNumber _bt_findinsertloc(Relation rel,
41 : BTInsertState insertstate,
42 : bool checkingunique,
43 : bool indexUnchanged,
44 : BTStack stack,
45 : Relation heapRel);
46 : static void _bt_stepright(Relation rel, Relation heaprel,
47 : BTInsertState insertstate, BTStack stack);
48 : static void _bt_insertonpg(Relation rel, Relation heaprel, BTScanInsert itup_key,
49 : Buffer buf,
50 : Buffer cbuf,
51 : BTStack stack,
52 : IndexTuple itup,
53 : Size itemsz,
54 : OffsetNumber newitemoff,
55 : int postingoff,
56 : bool split_only_page);
57 : static Buffer _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key,
58 : Buffer buf, Buffer cbuf, OffsetNumber newitemoff,
59 : Size newitemsz, IndexTuple newitem, IndexTuple orignewitem,
60 : IndexTuple nposting, uint16 postingoff);
61 : static void _bt_insert_parent(Relation rel, Relation heaprel, Buffer buf,
62 : Buffer rbuf, BTStack stack, bool isroot, bool isonly);
63 : static Buffer _bt_newlevel(Relation rel, Relation heaprel, Buffer lbuf, Buffer rbuf);
64 : static inline bool _bt_pgaddtup(Page page, Size itemsize, const IndexTupleData *itup,
65 : OffsetNumber itup_off, bool newfirstdataitem);
66 : static void _bt_delete_or_dedup_one_page(Relation rel, Relation heapRel,
67 : BTInsertState insertstate,
68 : bool simpleonly, bool checkingunique,
69 : bool uniquedup, bool indexUnchanged);
70 : static void _bt_simpledel_pass(Relation rel, Buffer buffer, Relation heapRel,
71 : OffsetNumber *deletable, int ndeletable,
72 : IndexTuple newitem, OffsetNumber minoff,
73 : OffsetNumber maxoff);
74 : static BlockNumber *_bt_deadblocks(Page page, OffsetNumber *deletable,
75 : int ndeletable, IndexTuple newitem,
76 : int *nblocks);
77 : static inline int _bt_blk_cmp(const void *arg1, const void *arg2);
78 :
79 : /*
80 : * _bt_doinsert() -- Handle insertion of a single index tuple in the tree.
81 : *
82 : * This routine is called by the public interface routine, btinsert.
83 : * By here, itup is filled in, including the TID.
84 : *
85 : * If checkUnique is UNIQUE_CHECK_NO or UNIQUE_CHECK_PARTIAL, this
86 : * will allow duplicates. Otherwise (UNIQUE_CHECK_YES or
87 : * UNIQUE_CHECK_EXISTING) it will throw error for a duplicate.
88 : * For UNIQUE_CHECK_EXISTING we merely run the duplicate check, and
89 : * don't actually insert.
90 : *
91 : * indexUnchanged executor hint indicates if itup is from an
92 : * UPDATE that didn't logically change the indexed value, but
93 : * must nevertheless have a new entry to point to a successor
94 : * version.
95 : *
96 : * The result value is only significant for UNIQUE_CHECK_PARTIAL:
97 : * it must be true if the entry is known unique, else false.
98 : * (In the current implementation we'll also return true after a
99 : * successful UNIQUE_CHECK_YES or UNIQUE_CHECK_EXISTING call, but
100 : * that's just a coding artifact.)
101 : */
102 : bool
103 7300360 : _bt_doinsert(Relation rel, IndexTuple itup,
104 : IndexUniqueCheck checkUnique, bool indexUnchanged,
105 : Relation heapRel)
106 : {
107 7300360 : bool is_unique = false;
108 : BTInsertStateData insertstate;
109 : BTScanInsert itup_key;
110 : BTStack stack;
111 7300360 : bool checkingunique = (checkUnique != UNIQUE_CHECK_NO);
112 :
113 : /* we need an insertion scan key to do our search, so build one */
114 7300360 : itup_key = _bt_mkscankey(rel, itup);
115 :
116 7300360 : if (checkingunique)
117 : {
118 5242892 : if (!itup_key->anynullkeys)
119 : {
120 : /* No (heapkeyspace) scantid until uniqueness established */
121 5222718 : itup_key->scantid = NULL;
122 : }
123 : else
124 : {
125 : /*
126 : * Scan key for new tuple contains NULL key values. Bypass
127 : * checkingunique steps. They are unnecessary because core code
128 : * considers NULL unequal to every value, including NULL.
129 : *
130 : * This optimization avoids O(N^2) behavior within the
131 : * _bt_findinsertloc() heapkeyspace path when a unique index has a
132 : * large number of "duplicates" with NULL key values.
133 : */
134 20174 : checkingunique = false;
135 : /* Tuple is unique in the sense that core code cares about */
136 : Assert(checkUnique != UNIQUE_CHECK_EXISTING);
137 20174 : is_unique = true;
138 : }
139 : }
140 :
141 : /*
142 : * Fill in the BTInsertState working area, to track the current page and
143 : * position within the page to insert on.
144 : *
145 : * Note that itemsz is passed down to lower level code that deals with
146 : * inserting the item. It must be MAXALIGN()'d. This ensures that space
147 : * accounting code consistently considers the alignment overhead that we
148 : * expect PageAddItem() will add later. (Actually, index_form_tuple() is
149 : * already conservative about alignment, but we don't rely on that from
150 : * this distance. Besides, preserving the "true" tuple size in index
151 : * tuple headers for the benefit of nbtsplitloc.c might happen someday.
152 : * Note that heapam does not MAXALIGN() each heap tuple's lp_len field.)
153 : */
154 7300360 : insertstate.itup = itup;
155 7300360 : insertstate.itemsz = MAXALIGN(IndexTupleSize(itup));
156 7300360 : insertstate.itup_key = itup_key;
157 7300360 : insertstate.bounds_valid = false;
158 7300360 : insertstate.buf = InvalidBuffer;
159 7300360 : insertstate.postingoff = 0;
160 :
161 7300384 : search:
162 :
163 : /*
164 : * Find and lock the leaf page that the tuple should be added to by
165 : * searching from the root page. insertstate.buf will hold a buffer that
166 : * is locked in exclusive mode afterwards.
167 : */
168 7300384 : stack = _bt_search_insert(rel, heapRel, &insertstate);
169 :
170 : /*
171 : * checkingunique inserts are not allowed to go ahead when two tuples with
172 : * equal key attribute values would be visible to new MVCC snapshots once
173 : * the xact commits. Check for conflicts in the locked page/buffer (if
174 : * needed) here.
175 : *
176 : * It might be necessary to check a page to the right in _bt_check_unique,
177 : * though that should be very rare. In practice the first page the value
178 : * could be on (with scantid omitted) is almost always also the only page
179 : * that a matching tuple might be found on. This is due to the behavior
180 : * of _bt_findsplitloc with duplicate tuples -- a group of duplicates can
181 : * only be allowed to cross a page boundary when there is no candidate
182 : * leaf page split point that avoids it. Also, _bt_check_unique can use
183 : * the leaf page high key to determine that there will be no duplicates on
184 : * the right sibling without actually visiting it (it uses the high key in
185 : * cases where the new item happens to belong at the far right of the leaf
186 : * page).
187 : *
188 : * NOTE: obviously, _bt_check_unique can only detect keys that are already
189 : * in the index; so it cannot defend against concurrent insertions of the
190 : * same key. We protect against that by means of holding a write lock on
191 : * the first page the value could be on, with omitted/-inf value for the
192 : * implicit heap TID tiebreaker attribute. Any other would-be inserter of
193 : * the same key must acquire a write lock on the same page, so only one
194 : * would-be inserter can be making the check at one time. Furthermore,
195 : * once we are past the check we hold write locks continuously until we
196 : * have performed our insertion, so no later inserter can fail to see our
197 : * insertion. (This requires some care in _bt_findinsertloc.)
198 : *
199 : * If we must wait for another xact, we release the lock while waiting,
200 : * and then must perform a new search.
201 : *
202 : * For a partial uniqueness check, we don't wait for the other xact. Just
203 : * let the tuple in and return false for possibly non-unique, or true for
204 : * definitely unique.
205 : */
206 7300384 : if (checkingunique)
207 : {
208 : TransactionId xwait;
209 : uint32 speculativeToken;
210 :
211 5222742 : xwait = _bt_check_unique(rel, &insertstate, heapRel, checkUnique,
212 : &is_unique, &speculativeToken);
213 :
214 5222214 : if (unlikely(TransactionIdIsValid(xwait)))
215 : {
216 : /* Have to wait for the other guy ... */
217 24 : _bt_relbuf(rel, insertstate.buf);
218 24 : insertstate.buf = InvalidBuffer;
219 :
220 : /*
221 : * If it's a speculative insertion, wait for it to finish (ie. to
222 : * go ahead with the insertion, or kill the tuple). Otherwise
223 : * wait for the transaction to finish as usual.
224 : */
225 24 : if (speculativeToken)
226 0 : SpeculativeInsertionWait(xwait, speculativeToken);
227 : else
228 24 : XactLockTableWait(xwait, rel, &itup->t_tid, XLTW_InsertIndex);
229 :
230 : /* start over... */
231 24 : if (stack)
232 0 : _bt_freestack(stack);
233 24 : goto search;
234 : }
235 :
236 : /* Uniqueness is established -- restore heap tid as scantid */
237 5222190 : if (itup_key->heapkeyspace)
238 5222190 : itup_key->scantid = &itup->t_tid;
239 : }
240 :
241 7299832 : if (checkUnique != UNIQUE_CHECK_EXISTING)
242 : {
243 : OffsetNumber newitemoff;
244 :
245 : /*
246 : * The only conflict predicate locking cares about for indexes is when
247 : * an index tuple insert conflicts with an existing lock. We don't
248 : * know the actual page we're going to insert on for sure just yet in
249 : * checkingunique and !heapkeyspace cases, but it's okay to use the
250 : * first page the value could be on (with scantid omitted) instead.
251 : */
252 7299778 : CheckForSerializableConflictIn(rel, NULL, BufferGetBlockNumber(insertstate.buf));
253 :
254 : /*
255 : * Do the insertion. Note that insertstate contains cached binary
256 : * search bounds established within _bt_check_unique when insertion is
257 : * checkingunique.
258 : */
259 7299772 : newitemoff = _bt_findinsertloc(rel, &insertstate, checkingunique,
260 : indexUnchanged, stack, heapRel);
261 7299772 : _bt_insertonpg(rel, heapRel, itup_key, insertstate.buf, InvalidBuffer,
262 : stack, itup, insertstate.itemsz, newitemoff,
263 : insertstate.postingoff, false);
264 : }
265 : else
266 : {
267 : /* just release the buffer */
268 54 : _bt_relbuf(rel, insertstate.buf);
269 : }
270 :
271 : /* be tidy */
272 7299826 : if (stack)
273 6337424 : _bt_freestack(stack);
274 7299826 : pfree(itup_key);
275 :
276 7299826 : return is_unique;
277 : }
278 :
279 : /*
280 : * _bt_search_insert() -- _bt_search() wrapper for inserts
281 : *
282 : * Search the tree for a particular scankey, or more precisely for the first
283 : * leaf page it could be on. Try to make use of the fastpath optimization's
284 : * rightmost leaf page cache before actually searching the tree from the root
285 : * page, though.
286 : *
287 : * Return value is a stack of parent-page pointers (though see notes about
288 : * fastpath optimization and page splits below). insertstate->buf is set to
289 : * the address of the leaf-page buffer, which is write-locked and pinned in
290 : * all cases (if necessary by creating a new empty root page for caller).
291 : *
292 : * The fastpath optimization avoids most of the work of searching the tree
293 : * repeatedly when a single backend inserts successive new tuples on the
294 : * rightmost leaf page of an index. A backend cache of the rightmost leaf
295 : * page is maintained within _bt_insertonpg(), and used here. The cache is
296 : * invalidated here when an insert of a non-pivot tuple must take place on a
297 : * non-rightmost leaf page.
298 : *
299 : * The optimization helps with indexes on an auto-incremented field. It also
300 : * helps with indexes on datetime columns, as well as indexes with lots of
301 : * NULL values. (NULLs usually get inserted in the rightmost page for single
302 : * column indexes, since they usually get treated as coming after everything
303 : * else in the key space. Individual NULL tuples will generally be placed on
304 : * the rightmost leaf page due to the influence of the heap TID column.)
305 : *
306 : * Note that we avoid applying the optimization when there is insufficient
307 : * space on the rightmost page to fit caller's new item. This is necessary
308 : * because we'll need to return a real descent stack when a page split is
309 : * expected (actually, caller can cope with a leaf page split that uses a NULL
310 : * stack, but that's very slow and so must be avoided). Note also that the
311 : * fastpath optimization acquires the lock on the page conditionally as a way
312 : * of reducing extra contention when there are concurrent insertions into the
313 : * rightmost page (we give up if we'd have to wait for the lock). We assume
314 : * that it isn't useful to apply the optimization when there is contention,
315 : * since each per-backend cache won't stay valid for long.
316 : */
317 : static BTStack
318 7300384 : _bt_search_insert(Relation rel, Relation heaprel, BTInsertState insertstate)
319 : {
320 : Assert(insertstate->buf == InvalidBuffer);
321 : Assert(!insertstate->bounds_valid);
322 : Assert(insertstate->postingoff == 0);
323 :
324 7300384 : if (RelationGetTargetBlock(rel) != InvalidBlockNumber)
325 : {
326 : /* Simulate a _bt_getbuf() call with conditional locking */
327 69352 : insertstate->buf = ReadBuffer(rel, RelationGetTargetBlock(rel));
328 69352 : if (_bt_conditionallockbuf(rel, insertstate->buf))
329 : {
330 : Page page;
331 : BTPageOpaque opaque;
332 :
333 68932 : _bt_checkpage(rel, insertstate->buf);
334 68932 : page = BufferGetPage(insertstate->buf);
335 68932 : opaque = BTPageGetOpaque(page);
336 :
337 : /*
338 : * Check if the page is still the rightmost leaf page and has
339 : * enough free space to accommodate the new tuple. Also check
340 : * that the insertion scan key is strictly greater than the first
341 : * non-pivot tuple on the page. (Note that we expect itup_key's
342 : * scantid to be unset when our caller is a checkingunique
343 : * inserter.)
344 : */
345 68932 : if (P_RIGHTMOST(opaque) &&
346 68850 : P_ISLEAF(opaque) &&
347 68850 : !P_IGNORE(opaque) &&
348 137292 : PageGetFreeSpace(page) > insertstate->itemsz &&
349 136884 : PageGetMaxOffsetNumber(page) >= P_HIKEY &&
350 68442 : _bt_compare(rel, insertstate->itup_key, page, P_HIKEY) > 0)
351 : {
352 : /*
353 : * Caller can use the fastpath optimization because cached
354 : * block is still rightmost leaf page, which can fit caller's
355 : * new tuple without splitting. Keep block in local cache for
356 : * next insert, and have caller use NULL stack.
357 : *
358 : * Note that _bt_insert_parent() has an assertion that catches
359 : * leaf page splits that somehow follow from a fastpath insert
360 : * (it should only be passed a NULL stack when it must deal
361 : * with a concurrent root page split, and never because a NULL
362 : * stack was returned here).
363 : */
364 68396 : return NULL;
365 : }
366 :
367 : /* Page unsuitable for caller, drop lock and pin */
368 536 : _bt_relbuf(rel, insertstate->buf);
369 : }
370 : else
371 : {
372 : /* Lock unavailable, drop pin */
373 420 : ReleaseBuffer(insertstate->buf);
374 : }
375 :
376 : /* Forget block, since cache doesn't appear to be useful */
377 956 : RelationSetTargetBlock(rel, InvalidBlockNumber);
378 : }
379 :
380 : /* Cannot use optimization -- descend tree, return proper descent stack */
381 7231988 : return _bt_search(rel, heaprel, insertstate->itup_key, &insertstate->buf,
382 : BT_WRITE);
383 : }
384 :
385 : /*
386 : * _bt_check_unique() -- Check for violation of unique index constraint
387 : *
388 : * Returns InvalidTransactionId if there is no conflict, else an xact ID
389 : * we must wait for to see if it commits a conflicting tuple. If an actual
390 : * conflict is detected, no return --- just ereport(). If an xact ID is
391 : * returned, and the conflicting tuple still has a speculative insertion in
392 : * progress, *speculativeToken is set to non-zero, and the caller can wait for
393 : * the verdict on the insertion using SpeculativeInsertionWait().
394 : *
395 : * However, if checkUnique == UNIQUE_CHECK_PARTIAL, we always return
396 : * InvalidTransactionId because we don't want to wait. In this case we
397 : * set *is_unique to false if there is a potential conflict, and the
398 : * core code must redo the uniqueness check later.
399 : *
400 : * As a side-effect, sets state in insertstate that can later be used by
401 : * _bt_findinsertloc() to reuse most of the binary search work we do
402 : * here.
403 : *
404 : * This code treats NULLs as equal, unlike the default semantics for unique
405 : * indexes. So do not call here when there are NULL values in scan key and
406 : * the index uses the default NULLS DISTINCT mode.
407 : */
408 : static TransactionId
409 5222742 : _bt_check_unique(Relation rel, BTInsertState insertstate, Relation heapRel,
410 : IndexUniqueCheck checkUnique, bool *is_unique,
411 : uint32 *speculativeToken)
412 : {
413 5222742 : IndexTuple itup = insertstate->itup;
414 5222742 : IndexTuple curitup = NULL;
415 5222742 : ItemId curitemid = NULL;
416 5222742 : BTScanInsert itup_key = insertstate->itup_key;
417 : SnapshotData SnapshotDirty;
418 : OffsetNumber offset;
419 : OffsetNumber maxoff;
420 : Page page;
421 : BTPageOpaque opaque;
422 5222742 : Buffer nbuf = InvalidBuffer;
423 5222742 : bool found = false;
424 5222742 : bool inposting = false;
425 5222742 : bool prevalldead = true;
426 5222742 : int curposti = 0;
427 :
428 : /* Assume unique until we find a duplicate */
429 5222742 : *is_unique = true;
430 :
431 5222742 : InitDirtySnapshot(SnapshotDirty);
432 :
433 5222742 : page = BufferGetPage(insertstate->buf);
434 5222742 : opaque = BTPageGetOpaque(page);
435 5222742 : maxoff = PageGetMaxOffsetNumber(page);
436 :
437 : /*
438 : * Find the first tuple with the same key.
439 : *
440 : * This also saves the binary search bounds in insertstate. We use them
441 : * in the fastpath below, but also in the _bt_findinsertloc() call later.
442 : */
443 : Assert(!insertstate->bounds_valid);
444 5222742 : offset = _bt_binsrch_insert(rel, insertstate);
445 :
446 : /*
447 : * Scan over all equal tuples, looking for live conflicts.
448 : */
449 : Assert(!insertstate->bounds_valid || insertstate->low == offset);
450 : Assert(!itup_key->anynullkeys);
451 : Assert(itup_key->scantid == NULL);
452 : for (;;)
453 : {
454 : /*
455 : * Each iteration of the loop processes one heap TID, not one index
456 : * tuple. Current offset number for page isn't usually advanced on
457 : * iterations that process heap TIDs from posting list tuples.
458 : *
459 : * "inposting" state is set when _inside_ a posting list --- not when
460 : * we're at the start (or end) of a posting list. We advance curposti
461 : * at the end of the iteration when inside a posting list tuple. In
462 : * general, every loop iteration either advances the page offset or
463 : * advances curposti --- an iteration that handles the rightmost/max
464 : * heap TID in a posting list finally advances the page offset (and
465 : * unsets "inposting").
466 : *
467 : * Make sure the offset points to an actual index tuple before trying
468 : * to examine it...
469 : */
470 17221376 : if (offset <= maxoff)
471 : {
472 : /*
473 : * Fastpath: In most cases, we can use cached search bounds to
474 : * limit our consideration to items that are definitely
475 : * duplicates. This fastpath doesn't apply when the original page
476 : * is empty, or when initial offset is past the end of the
477 : * original page, which may indicate that we need to examine a
478 : * second or subsequent page.
479 : *
480 : * Note that this optimization allows us to avoid calling
481 : * _bt_compare() directly when there are no duplicates, as long as
482 : * the offset where the key will go is not at the end of the page.
483 : */
484 14295606 : if (nbuf == InvalidBuffer && offset == insertstate->stricthigh)
485 : {
486 : Assert(insertstate->bounds_valid);
487 : Assert(insertstate->low >= P_FIRSTDATAKEY(opaque));
488 : Assert(insertstate->low <= insertstate->stricthigh);
489 : Assert(_bt_compare(rel, itup_key, page, offset) < 0);
490 2064076 : break;
491 : }
492 :
493 : /*
494 : * We can skip items that are already marked killed.
495 : *
496 : * In the presence of heavy update activity an index may contain
497 : * many killed items with the same key; running _bt_compare() on
498 : * each killed item gets expensive. Just advance over killed
499 : * items as quickly as we can. We only apply _bt_compare() when
500 : * we get to a non-killed item. We could reuse the bounds to
501 : * avoid _bt_compare() calls for known equal tuples, but it
502 : * doesn't seem worth it.
503 : */
504 12231530 : if (!inposting)
505 7714164 : curitemid = PageGetItemId(page, offset);
506 12231530 : if (inposting || !ItemIdIsDead(curitemid))
507 : {
508 : ItemPointerData htid;
509 11672396 : bool all_dead = false;
510 :
511 11672396 : if (!inposting)
512 : {
513 : /* Plain tuple, or first TID in posting list tuple */
514 7155030 : if (_bt_compare(rel, itup_key, page, offset) != 0)
515 206942 : break; /* we're past all the equal tuples */
516 :
517 : /* Advanced curitup */
518 6948088 : curitup = (IndexTuple) PageGetItem(page, curitemid);
519 : Assert(!BTreeTupleIsPivot(curitup));
520 : }
521 :
522 : /* okay, we gotta fetch the heap tuple using htid ... */
523 11465454 : if (!BTreeTupleIsPosting(curitup))
524 : {
525 : /* ... htid is from simple non-pivot tuple */
526 : Assert(!inposting);
527 6901524 : htid = curitup->t_tid;
528 : }
529 4563930 : else if (!inposting)
530 : {
531 : /* ... htid is first TID in new posting list */
532 46564 : inposting = true;
533 46564 : prevalldead = true;
534 46564 : curposti = 0;
535 46564 : htid = *BTreeTupleGetPostingN(curitup, 0);
536 : }
537 : else
538 : {
539 : /* ... htid is second or subsequent TID in posting list */
540 : Assert(curposti > 0);
541 4517366 : htid = *BTreeTupleGetPostingN(curitup, curposti);
542 : }
543 :
544 : /*
545 : * If we are doing a recheck, we expect to find the tuple we
546 : * are rechecking. It's not a duplicate, but we have to keep
547 : * scanning.
548 : */
549 11465690 : if (checkUnique == UNIQUE_CHECK_EXISTING &&
550 236 : ItemPointerCompare(&htid, &itup->t_tid) == 0)
551 : {
552 54 : found = true;
553 : }
554 :
555 : /*
556 : * Check if there's any table tuples for this index entry
557 : * satisfying SnapshotDirty. This is necessary because for AMs
558 : * with optimizations like heap's HOT, we have just a single
559 : * index entry for the entire chain.
560 : */
561 11465400 : else if (table_index_fetch_tuple_check(heapRel, &htid,
562 : &SnapshotDirty,
563 : &all_dead))
564 : {
565 : TransactionId xwait;
566 :
567 : /*
568 : * It is a duplicate. If we are only doing a partial
569 : * check, then don't bother checking if the tuple is being
570 : * updated in another transaction. Just return the fact
571 : * that it is a potential conflict and leave the full
572 : * check till later. Don't invalidate binary search
573 : * bounds.
574 : */
575 748 : if (checkUnique == UNIQUE_CHECK_PARTIAL)
576 : {
577 196 : if (nbuf != InvalidBuffer)
578 0 : _bt_relbuf(rel, nbuf);
579 196 : *is_unique = false;
580 220 : return InvalidTransactionId;
581 : }
582 :
583 : /*
584 : * If this tuple is being updated by other transaction
585 : * then we have to wait for its commit/abort.
586 : */
587 1104 : xwait = (TransactionIdIsValid(SnapshotDirty.xmin)) ?
588 552 : SnapshotDirty.xmin : SnapshotDirty.xmax;
589 :
590 552 : if (TransactionIdIsValid(xwait))
591 : {
592 24 : if (nbuf != InvalidBuffer)
593 0 : _bt_relbuf(rel, nbuf);
594 : /* Tell _bt_doinsert to wait... */
595 24 : *speculativeToken = SnapshotDirty.speculativeToken;
596 : /* Caller releases lock on buf immediately */
597 24 : insertstate->bounds_valid = false;
598 24 : return xwait;
599 : }
600 :
601 : /*
602 : * Otherwise we have a definite conflict. But before
603 : * complaining, look to see if the tuple we want to insert
604 : * is itself now committed dead --- if so, don't complain.
605 : * This is a waste of time in normal scenarios but we must
606 : * do it to support CREATE INDEX CONCURRENTLY.
607 : *
608 : * We must follow HOT-chains here because during
609 : * concurrent index build, we insert the root TID though
610 : * the actual tuple may be somewhere in the HOT-chain.
611 : * While following the chain we might not stop at the
612 : * exact tuple which triggered the insert, but that's OK
613 : * because if we find a live tuple anywhere in this chain,
614 : * we have a unique key conflict. The other live tuple is
615 : * not part of this chain because it had a different index
616 : * entry.
617 : */
618 528 : htid = itup->t_tid;
619 528 : if (table_index_fetch_tuple_check(heapRel, &htid,
620 : SnapshotSelf, NULL))
621 : {
622 : /* Normal case --- it's still live */
623 : }
624 : else
625 : {
626 : /*
627 : * It's been deleted, so no error, and no need to
628 : * continue searching
629 : */
630 0 : break;
631 : }
632 :
633 : /*
634 : * Check for a conflict-in as we would if we were going to
635 : * write to this page. We aren't actually going to write,
636 : * but we want a chance to report SSI conflicts that would
637 : * otherwise be masked by this unique constraint
638 : * violation.
639 : */
640 528 : CheckForSerializableConflictIn(rel, NULL, BufferGetBlockNumber(insertstate->buf));
641 :
642 : /*
643 : * This is a definite conflict. Break the tuple down into
644 : * datums and report the error. But first, make sure we
645 : * release the buffer locks we're holding ---
646 : * BuildIndexValueDescription could make catalog accesses,
647 : * which in the worst case might touch this same index and
648 : * cause deadlocks.
649 : */
650 520 : if (nbuf != InvalidBuffer)
651 0 : _bt_relbuf(rel, nbuf);
652 520 : _bt_relbuf(rel, insertstate->buf);
653 520 : insertstate->buf = InvalidBuffer;
654 520 : insertstate->bounds_valid = false;
655 :
656 : {
657 : Datum values[INDEX_MAX_KEYS];
658 : bool isnull[INDEX_MAX_KEYS];
659 : char *key_desc;
660 :
661 520 : index_deform_tuple(itup, RelationGetDescr(rel),
662 : values, isnull);
663 :
664 520 : key_desc = BuildIndexValueDescription(rel, values,
665 : isnull);
666 :
667 520 : ereport(ERROR,
668 : (errcode(ERRCODE_UNIQUE_VIOLATION),
669 : errmsg("duplicate key value violates unique constraint \"%s\"",
670 : RelationGetRelationName(rel)),
671 : key_desc ? errdetail("Key %s already exists.",
672 : key_desc) : 0,
673 : errtableconstraint(heapRel,
674 : RelationGetRelationName(rel))));
675 : }
676 : }
677 11464652 : else if (all_dead && (!inposting ||
678 37950 : (prevalldead &&
679 37950 : curposti == BTreeTupleGetNPosting(curitup) - 1)))
680 : {
681 : /*
682 : * The conflicting tuple (or all HOT chains pointed to by
683 : * all posting list TIDs) is dead to everyone, so mark the
684 : * index entry killed.
685 : */
686 106262 : ItemIdMarkDead(curitemid);
687 106262 : opaque->btpo_flags |= BTP_HAS_GARBAGE;
688 :
689 : /*
690 : * Mark buffer with a dirty hint, since state is not
691 : * crucial. Be sure to mark the proper buffer dirty.
692 : */
693 106262 : if (nbuf != InvalidBuffer)
694 4 : MarkBufferDirtyHint(nbuf, true);
695 : else
696 106258 : MarkBufferDirtyHint(insertstate->buf, true);
697 : }
698 :
699 : /*
700 : * Remember if posting list tuple has even a single HOT chain
701 : * whose members are not all dead
702 : */
703 11464706 : if (!all_dead && inposting)
704 4525910 : prevalldead = false;
705 : }
706 : }
707 :
708 14949610 : if (inposting && curposti < BTreeTupleGetNPosting(curitup) - 1)
709 : {
710 : /* Advance to next TID in same posting list */
711 4517366 : curposti++;
712 4517366 : continue;
713 : }
714 10432244 : else if (offset < maxoff)
715 : {
716 : /* Advance to next tuple */
717 7471224 : curposti = 0;
718 7471224 : inposting = false;
719 7471224 : offset = OffsetNumberNext(offset);
720 : }
721 : else
722 : {
723 : int highkeycmp;
724 :
725 : /* If scankey == hikey we gotta check the next page too */
726 2961020 : if (P_RIGHTMOST(opaque))
727 2818402 : break;
728 142618 : highkeycmp = _bt_compare(rel, itup_key, page, P_HIKEY);
729 : Assert(highkeycmp <= 0);
730 142618 : if (highkeycmp != 0)
731 132574 : break;
732 : /* Advance to next non-dead page --- there must be one */
733 : for (;;)
734 0 : {
735 10044 : BlockNumber nblkno = opaque->btpo_next;
736 :
737 10044 : nbuf = _bt_relandgetbuf(rel, nbuf, nblkno, BT_READ);
738 10044 : page = BufferGetPage(nbuf);
739 10044 : opaque = BTPageGetOpaque(page);
740 10044 : if (!P_IGNORE(opaque))
741 10044 : break;
742 0 : if (P_RIGHTMOST(opaque))
743 0 : elog(ERROR, "fell off the end of index \"%s\"",
744 : RelationGetRelationName(rel));
745 : }
746 : /* Will also advance to next tuple */
747 10044 : curposti = 0;
748 10044 : inposting = false;
749 10044 : maxoff = PageGetMaxOffsetNumber(page);
750 10044 : offset = P_FIRSTDATAKEY(opaque);
751 : /* Don't invalidate binary search bounds */
752 : }
753 : }
754 :
755 : /*
756 : * If we are doing a recheck then we should have found the tuple we are
757 : * checking. Otherwise there's something very wrong --- probably, the
758 : * index is on a non-immutable expression.
759 : */
760 5221994 : if (checkUnique == UNIQUE_CHECK_EXISTING && !found)
761 0 : ereport(ERROR,
762 : (errcode(ERRCODE_INTERNAL_ERROR),
763 : errmsg("failed to re-find tuple within index \"%s\"",
764 : RelationGetRelationName(rel)),
765 : errhint("This may be because of a non-immutable index expression."),
766 : errtableconstraint(heapRel,
767 : RelationGetRelationName(rel))));
768 :
769 5221994 : if (nbuf != InvalidBuffer)
770 5760 : _bt_relbuf(rel, nbuf);
771 :
772 5221994 : return InvalidTransactionId;
773 : }
774 :
775 :
776 : /*
777 : * _bt_findinsertloc() -- Finds an insert location for a tuple
778 : *
779 : * On entry, insertstate buffer contains the page the new tuple belongs
780 : * on. It is exclusive-locked and pinned by the caller.
781 : *
782 : * If 'checkingunique' is true, the buffer on entry is the first page
783 : * that contains duplicates of the new key. If there are duplicates on
784 : * multiple pages, the correct insertion position might be some page to
785 : * the right, rather than the first page. In that case, this function
786 : * moves right to the correct target page.
787 : *
788 : * (In a !heapkeyspace index, there can be multiple pages with the same
789 : * high key, where the new tuple could legitimately be placed on. In
790 : * that case, the caller passes the first page containing duplicates,
791 : * just like when checkingunique=true. If that page doesn't have enough
792 : * room for the new tuple, this function moves right, trying to find a
793 : * legal page that does.)
794 : *
795 : * If 'indexUnchanged' is true, this is for an UPDATE that didn't
796 : * logically change the indexed value, but must nevertheless have a new
797 : * entry to point to a successor version. This hint from the executor
798 : * will influence our behavior when the page might have to be split and
799 : * we must consider our options. Bottom-up index deletion can avoid
800 : * pathological version-driven page splits, but we only want to go to the
801 : * trouble of trying it when we already have moderate confidence that
802 : * it's appropriate. The hint should not significantly affect our
803 : * behavior over time unless practically all inserts on to the leaf page
804 : * get the hint.
805 : *
806 : * On exit, insertstate buffer contains the chosen insertion page, and
807 : * the offset within that page is returned. If _bt_findinsertloc needed
808 : * to move right, the lock and pin on the original page are released, and
809 : * the new buffer is exclusively locked and pinned instead.
810 : *
811 : * If insertstate contains cached binary search bounds, we will take
812 : * advantage of them. This avoids repeating comparisons that we made in
813 : * _bt_check_unique() already.
814 : */
815 : static OffsetNumber
816 7299772 : _bt_findinsertloc(Relation rel,
817 : BTInsertState insertstate,
818 : bool checkingunique,
819 : bool indexUnchanged,
820 : BTStack stack,
821 : Relation heapRel)
822 : {
823 7299772 : BTScanInsert itup_key = insertstate->itup_key;
824 7299772 : Page page = BufferGetPage(insertstate->buf);
825 : BTPageOpaque opaque;
826 : OffsetNumber newitemoff;
827 :
828 7299772 : opaque = BTPageGetOpaque(page);
829 :
830 : /* Check 1/3 of a page restriction */
831 7299772 : if (unlikely(insertstate->itemsz > BTMaxItemSize))
832 0 : _bt_check_third_page(rel, heapRel, itup_key->heapkeyspace, page,
833 : insertstate->itup);
834 :
835 : Assert(P_ISLEAF(opaque) && !P_INCOMPLETE_SPLIT(opaque));
836 : Assert(!insertstate->bounds_valid || checkingunique);
837 : Assert(!itup_key->heapkeyspace || itup_key->scantid != NULL);
838 : Assert(itup_key->heapkeyspace || itup_key->scantid == NULL);
839 : Assert(!itup_key->allequalimage || itup_key->heapkeyspace);
840 :
841 7299772 : if (itup_key->heapkeyspace)
842 : {
843 : /* Keep track of whether checkingunique duplicate seen */
844 7299772 : bool uniquedup = indexUnchanged;
845 :
846 : /*
847 : * If we're inserting into a unique index, we may have to walk right
848 : * through leaf pages to find the one leaf page that we must insert on
849 : * to.
850 : *
851 : * This is needed for checkingunique callers because a scantid was not
852 : * used when we called _bt_search(). scantid can only be set after
853 : * _bt_check_unique() has checked for duplicates. The buffer
854 : * initially stored in insertstate->buf has the page where the first
855 : * duplicate key might be found, which isn't always the page that new
856 : * tuple belongs on. The heap TID attribute for new tuple (scantid)
857 : * could force us to insert on a sibling page, though that should be
858 : * very rare in practice.
859 : */
860 7299772 : if (checkingunique)
861 : {
862 5222130 : if (insertstate->low < insertstate->stricthigh)
863 : {
864 : /* Encountered a duplicate in _bt_check_unique() */
865 : Assert(insertstate->bounds_valid);
866 446686 : uniquedup = true;
867 : }
868 :
869 : for (;;)
870 : {
871 : /*
872 : * Does the new tuple belong on this page?
873 : *
874 : * The earlier _bt_check_unique() call may well have
875 : * established a strict upper bound on the offset for the new
876 : * item. If it's not the last item of the page (i.e. if there
877 : * is at least one tuple on the page that goes after the tuple
878 : * we're inserting) then we know that the tuple belongs on
879 : * this page. We can skip the high key check.
880 : */
881 5232174 : if (insertstate->bounds_valid &&
882 10422512 : insertstate->low <= insertstate->stricthigh &&
883 5211256 : insertstate->stricthigh <= PageGetMaxOffsetNumber(page))
884 2244744 : break;
885 :
886 : /* Test '<=', not '!=', since scantid is set now */
887 3145208 : if (P_RIGHTMOST(opaque) ||
888 157778 : _bt_compare(rel, itup_key, page, P_HIKEY) <= 0)
889 : break;
890 :
891 10044 : _bt_stepright(rel, heapRel, insertstate, stack);
892 : /* Update local state after stepping right */
893 10044 : page = BufferGetPage(insertstate->buf);
894 10044 : opaque = BTPageGetOpaque(page);
895 : /* Assume duplicates (if checkingunique) */
896 10044 : uniquedup = true;
897 : }
898 : }
899 :
900 : /*
901 : * If the target page cannot fit newitem, try to avoid splitting the
902 : * page on insert by performing deletion or deduplication now
903 : */
904 7299772 : if (PageGetFreeSpace(page) < insertstate->itemsz)
905 51288 : _bt_delete_or_dedup_one_page(rel, heapRel, insertstate, false,
906 : checkingunique, uniquedup,
907 : indexUnchanged);
908 : }
909 : else
910 : {
911 : /*----------
912 : * This is a !heapkeyspace (version 2 or 3) index. The current page
913 : * is the first page that we could insert the new tuple to, but there
914 : * may be other pages to the right that we could opt to use instead.
915 : *
916 : * If the new key is equal to one or more existing keys, we can
917 : * legitimately place it anywhere in the series of equal keys. In
918 : * fact, if the new key is equal to the page's "high key" we can place
919 : * it on the next page. If it is equal to the high key, and there's
920 : * not room to insert the new tuple on the current page without
921 : * splitting, then we move right hoping to find more free space and
922 : * avoid a split.
923 : *
924 : * Keep scanning right until we
925 : * (a) find a page with enough free space,
926 : * (b) reach the last page where the tuple can legally go, or
927 : * (c) get tired of searching.
928 : * (c) is not flippant; it is important because if there are many
929 : * pages' worth of equal keys, it's better to split one of the early
930 : * pages than to scan all the way to the end of the run of equal keys
931 : * on every insert. We implement "get tired" as a random choice,
932 : * since stopping after scanning a fixed number of pages wouldn't work
933 : * well (we'd never reach the right-hand side of previously split
934 : * pages). The probability of moving right is set at 0.99, which may
935 : * seem too high to change the behavior much, but it does an excellent
936 : * job of preventing O(N^2) behavior with many equal keys.
937 : *----------
938 : */
939 0 : while (PageGetFreeSpace(page) < insertstate->itemsz)
940 : {
941 : /*
942 : * Before considering moving right, see if we can obtain enough
943 : * space by erasing LP_DEAD items
944 : */
945 0 : if (P_HAS_GARBAGE(opaque))
946 : {
947 : /* Perform simple deletion */
948 0 : _bt_delete_or_dedup_one_page(rel, heapRel, insertstate, true,
949 : false, false, false);
950 :
951 0 : if (PageGetFreeSpace(page) >= insertstate->itemsz)
952 0 : break; /* OK, now we have enough space */
953 : }
954 :
955 : /*
956 : * Nope, so check conditions (b) and (c) enumerated above
957 : *
958 : * The earlier _bt_check_unique() call may well have established a
959 : * strict upper bound on the offset for the new item. If it's not
960 : * the last item of the page (i.e. if there is at least one tuple
961 : * on the page that's greater than the tuple we're inserting to)
962 : * then we know that the tuple belongs on this page. We can skip
963 : * the high key check.
964 : */
965 0 : if (insertstate->bounds_valid &&
966 0 : insertstate->low <= insertstate->stricthigh &&
967 0 : insertstate->stricthigh <= PageGetMaxOffsetNumber(page))
968 0 : break;
969 :
970 0 : if (P_RIGHTMOST(opaque) ||
971 0 : _bt_compare(rel, itup_key, page, P_HIKEY) != 0 ||
972 0 : pg_prng_uint32(&pg_global_prng_state) <= (PG_UINT32_MAX / 100))
973 : break;
974 :
975 0 : _bt_stepright(rel, heapRel, insertstate, stack);
976 : /* Update local state after stepping right */
977 0 : page = BufferGetPage(insertstate->buf);
978 0 : opaque = BTPageGetOpaque(page);
979 : }
980 : }
981 :
982 : /*
983 : * We should now be on the correct page. Find the offset within the page
984 : * for the new tuple. (Possibly reusing earlier search bounds.)
985 : */
986 : Assert(P_RIGHTMOST(opaque) ||
987 : _bt_compare(rel, itup_key, page, P_HIKEY) <= 0);
988 :
989 7299772 : newitemoff = _bt_binsrch_insert(rel, insertstate);
990 :
991 7299772 : if (insertstate->postingoff == -1)
992 : {
993 : /*
994 : * There is an overlapping posting list tuple with its LP_DEAD bit
995 : * set. We don't want to unnecessarily unset its LP_DEAD bit while
996 : * performing a posting list split, so perform simple index tuple
997 : * deletion early.
998 : */
999 2 : _bt_delete_or_dedup_one_page(rel, heapRel, insertstate, true,
1000 : false, false, false);
1001 :
1002 : /*
1003 : * Do new binary search. New insert location cannot overlap with any
1004 : * posting list now.
1005 : */
1006 : Assert(!insertstate->bounds_valid);
1007 2 : insertstate->postingoff = 0;
1008 2 : newitemoff = _bt_binsrch_insert(rel, insertstate);
1009 : Assert(insertstate->postingoff == 0);
1010 : }
1011 :
1012 7299772 : return newitemoff;
1013 : }
1014 :
1015 : /*
1016 : * Step right to next non-dead page, during insertion.
1017 : *
1018 : * This is a bit more complicated than moving right in a search. We must
1019 : * write-lock the target page before releasing write lock on current page;
1020 : * else someone else's _bt_check_unique scan could fail to see our insertion.
1021 : * Write locks on intermediate dead pages won't do because we don't know when
1022 : * they will get de-linked from the tree.
1023 : *
1024 : * This is more aggressive than it needs to be for non-unique !heapkeyspace
1025 : * indexes.
1026 : */
1027 : static void
1028 10044 : _bt_stepright(Relation rel, Relation heaprel, BTInsertState insertstate,
1029 : BTStack stack)
1030 : {
1031 : Page page;
1032 : BTPageOpaque opaque;
1033 : Buffer rbuf;
1034 : BlockNumber rblkno;
1035 :
1036 : Assert(heaprel != NULL);
1037 10044 : page = BufferGetPage(insertstate->buf);
1038 10044 : opaque = BTPageGetOpaque(page);
1039 :
1040 10044 : rbuf = InvalidBuffer;
1041 10044 : rblkno = opaque->btpo_next;
1042 : for (;;)
1043 : {
1044 10044 : rbuf = _bt_relandgetbuf(rel, rbuf, rblkno, BT_WRITE);
1045 10044 : page = BufferGetPage(rbuf);
1046 10044 : opaque = BTPageGetOpaque(page);
1047 :
1048 : /*
1049 : * If this page was incompletely split, finish the split now. We do
1050 : * this while holding a lock on the left sibling, which is not good
1051 : * because finishing the split could be a fairly lengthy operation.
1052 : * But this should happen very seldom.
1053 : */
1054 10044 : if (P_INCOMPLETE_SPLIT(opaque))
1055 : {
1056 0 : _bt_finish_split(rel, heaprel, rbuf, stack);
1057 0 : rbuf = InvalidBuffer;
1058 0 : continue;
1059 : }
1060 :
1061 10044 : if (!P_IGNORE(opaque))
1062 10044 : break;
1063 0 : if (P_RIGHTMOST(opaque))
1064 0 : elog(ERROR, "fell off the end of index \"%s\"",
1065 : RelationGetRelationName(rel));
1066 :
1067 0 : rblkno = opaque->btpo_next;
1068 : }
1069 : /* rbuf locked; unlock buf, update state for caller */
1070 10044 : _bt_relbuf(rel, insertstate->buf);
1071 10044 : insertstate->buf = rbuf;
1072 10044 : insertstate->bounds_valid = false;
1073 10044 : }
1074 :
1075 : /*----------
1076 : * _bt_insertonpg() -- Insert a tuple on a particular page in the index.
1077 : *
1078 : * This recursive procedure does the following things:
1079 : *
1080 : * + if postingoff != 0, splits existing posting list tuple
1081 : * (since it overlaps with new 'itup' tuple).
1082 : * + if necessary, splits the target page, using 'itup_key' for
1083 : * suffix truncation on leaf pages (caller passes NULL for
1084 : * non-leaf pages).
1085 : * + inserts the new tuple (might be split from posting list).
1086 : * + if the page was split, pops the parent stack, and finds the
1087 : * right place to insert the new child pointer (by walking
1088 : * right using information stored in the parent stack).
1089 : * + invokes itself with the appropriate tuple for the right
1090 : * child page on the parent.
1091 : * + updates the metapage if a true root or fast root is split.
1092 : *
1093 : * On entry, we must have the correct buffer in which to do the
1094 : * insertion, and the buffer must be pinned and write-locked. On return,
1095 : * we will have dropped both the pin and the lock on the buffer.
1096 : *
1097 : * This routine only performs retail tuple insertions. 'itup' should
1098 : * always be either a non-highkey leaf item, or a downlink (new high
1099 : * key items are created indirectly, when a page is split). When
1100 : * inserting to a non-leaf page, 'cbuf' is the left-sibling of the page
1101 : * we're inserting the downlink for. This function will clear the
1102 : * INCOMPLETE_SPLIT flag on it, and release the buffer.
1103 : *----------
1104 : */
1105 : static void
1106 7320988 : _bt_insertonpg(Relation rel,
1107 : Relation heaprel,
1108 : BTScanInsert itup_key,
1109 : Buffer buf,
1110 : Buffer cbuf,
1111 : BTStack stack,
1112 : IndexTuple itup,
1113 : Size itemsz,
1114 : OffsetNumber newitemoff,
1115 : int postingoff,
1116 : bool split_only_page)
1117 : {
1118 : Page page;
1119 : BTPageOpaque opaque;
1120 : bool isleaf,
1121 : isroot,
1122 : isrightmost,
1123 : isonly;
1124 7320988 : IndexTuple oposting = NULL;
1125 7320988 : IndexTuple origitup = NULL;
1126 7320988 : IndexTuple nposting = NULL;
1127 :
1128 7320988 : page = BufferGetPage(buf);
1129 7320988 : opaque = BTPageGetOpaque(page);
1130 7320988 : isleaf = P_ISLEAF(opaque);
1131 7320988 : isroot = P_ISROOT(opaque);
1132 7320988 : isrightmost = P_RIGHTMOST(opaque);
1133 7320988 : isonly = P_LEFTMOST(opaque) && P_RIGHTMOST(opaque);
1134 :
1135 : /* child buffer must be given iff inserting on an internal page */
1136 : Assert(isleaf == !BufferIsValid(cbuf));
1137 : /* tuple must have appropriate number of attributes */
1138 : Assert(!isleaf ||
1139 : BTreeTupleGetNAtts(itup, rel) ==
1140 : IndexRelationGetNumberOfAttributes(rel));
1141 : Assert(isleaf ||
1142 : BTreeTupleGetNAtts(itup, rel) <=
1143 : IndexRelationGetNumberOfKeyAttributes(rel));
1144 : Assert(!BTreeTupleIsPosting(itup));
1145 : Assert(MAXALIGN(IndexTupleSize(itup)) == itemsz);
1146 : /* Caller must always finish incomplete split for us */
1147 : Assert(!P_INCOMPLETE_SPLIT(opaque));
1148 :
1149 : /*
1150 : * Every internal page should have exactly one negative infinity item at
1151 : * all times. Only _bt_split() and _bt_newlevel() should add items that
1152 : * become negative infinity items through truncation, since they're the
1153 : * only routines that allocate new internal pages.
1154 : */
1155 : Assert(isleaf || newitemoff > P_FIRSTDATAKEY(opaque));
1156 :
1157 : /*
1158 : * Do we need to split an existing posting list item?
1159 : */
1160 7320988 : if (postingoff != 0)
1161 : {
1162 22442 : ItemId itemid = PageGetItemId(page, newitemoff);
1163 :
1164 : /*
1165 : * The new tuple is a duplicate with a heap TID that falls inside the
1166 : * range of an existing posting list tuple on a leaf page. Prepare to
1167 : * split an existing posting list. Overwriting the posting list with
1168 : * its post-split version is treated as an extra step in either the
1169 : * insert or page split critical section.
1170 : */
1171 : Assert(isleaf && itup_key->heapkeyspace && itup_key->allequalimage);
1172 22442 : oposting = (IndexTuple) PageGetItem(page, itemid);
1173 :
1174 : /*
1175 : * postingoff value comes from earlier call to _bt_binsrch_posting().
1176 : * Its binary search might think that a plain tuple must be a posting
1177 : * list tuple that needs to be split. This can happen with corruption
1178 : * involving an existing plain tuple that is a duplicate of the new
1179 : * item, up to and including its table TID. Check for that here in
1180 : * passing.
1181 : *
1182 : * Also verify that our caller has made sure that the existing posting
1183 : * list tuple does not have its LP_DEAD bit set.
1184 : */
1185 22442 : if (!BTreeTupleIsPosting(oposting) || ItemIdIsDead(itemid))
1186 0 : ereport(ERROR,
1187 : (errcode(ERRCODE_INDEX_CORRUPTED),
1188 : errmsg_internal("table tid from new index tuple (%u,%u) overlaps with invalid duplicate tuple at offset %u of block %u in index \"%s\"",
1189 : ItemPointerGetBlockNumber(&itup->t_tid),
1190 : ItemPointerGetOffsetNumber(&itup->t_tid),
1191 : newitemoff, BufferGetBlockNumber(buf),
1192 : RelationGetRelationName(rel))));
1193 :
1194 : /* use a mutable copy of itup as our itup from here on */
1195 22442 : origitup = itup;
1196 22442 : itup = CopyIndexTuple(origitup);
1197 22442 : nposting = _bt_swap_posting(itup, oposting, postingoff);
1198 : /* itup now contains rightmost/max TID from oposting */
1199 :
1200 : /* Alter offset so that newitem goes after posting list */
1201 22442 : newitemoff = OffsetNumberNext(newitemoff);
1202 : }
1203 :
1204 : /*
1205 : * Do we need to split the page to fit the item on it?
1206 : *
1207 : * Note: PageGetFreeSpace() subtracts sizeof(ItemIdData) from its result,
1208 : * so this comparison is correct even though we appear to be accounting
1209 : * only for the item and not for its line pointer.
1210 : */
1211 7320988 : if (PageGetFreeSpace(page) < itemsz)
1212 : {
1213 : Buffer rbuf;
1214 :
1215 : Assert(!split_only_page);
1216 :
1217 : /* split the buffer into left and right halves */
1218 22580 : rbuf = _bt_split(rel, heaprel, itup_key, buf, cbuf, newitemoff, itemsz,
1219 : itup, origitup, nposting, postingoff);
1220 22580 : PredicateLockPageSplit(rel,
1221 : BufferGetBlockNumber(buf),
1222 : BufferGetBlockNumber(rbuf));
1223 :
1224 : /*----------
1225 : * By here,
1226 : *
1227 : * + our target page has been split;
1228 : * + the original tuple has been inserted;
1229 : * + we have write locks on both the old (left half)
1230 : * and new (right half) buffers, after the split; and
1231 : * + we know the key we want to insert into the parent
1232 : * (it's the "high key" on the left child page).
1233 : *
1234 : * We're ready to do the parent insertion. We need to hold onto the
1235 : * locks for the child pages until we locate the parent, but we can
1236 : * at least release the lock on the right child before doing the
1237 : * actual insertion. The lock on the left child will be released
1238 : * last of all by parent insertion, where it is the 'cbuf' of parent
1239 : * page.
1240 : *----------
1241 : */
1242 22580 : _bt_insert_parent(rel, heaprel, buf, rbuf, stack, isroot, isonly);
1243 : }
1244 : else
1245 : {
1246 7298408 : Buffer metabuf = InvalidBuffer;
1247 7298408 : Page metapg = NULL;
1248 7298408 : BTMetaPageData *metad = NULL;
1249 : BlockNumber blockcache;
1250 :
1251 : /*
1252 : * If we are doing this insert because we split a page that was the
1253 : * only one on its tree level, but was not the root, it may have been
1254 : * the "fast root". We need to ensure that the fast root link points
1255 : * at or above the current page. We can safely acquire a lock on the
1256 : * metapage here --- see comments for _bt_newlevel().
1257 : */
1258 7298408 : if (unlikely(split_only_page))
1259 : {
1260 : Assert(!isleaf);
1261 : Assert(BufferIsValid(cbuf));
1262 :
1263 22 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_WRITE);
1264 22 : metapg = BufferGetPage(metabuf);
1265 22 : metad = BTPageGetMeta(metapg);
1266 :
1267 22 : if (metad->btm_fastlevel >= opaque->btpo_level)
1268 : {
1269 : /* no update wanted */
1270 0 : _bt_relbuf(rel, metabuf);
1271 0 : metabuf = InvalidBuffer;
1272 : }
1273 : }
1274 :
1275 : /* Do the update. No ereport(ERROR) until changes are logged */
1276 7298408 : START_CRIT_SECTION();
1277 :
1278 7298408 : if (postingoff != 0)
1279 22374 : memcpy(oposting, nposting, MAXALIGN(IndexTupleSize(nposting)));
1280 :
1281 7298408 : if (PageAddItem(page, itup, itemsz, newitemoff, false, false) == InvalidOffsetNumber)
1282 0 : elog(PANIC, "failed to add new item to block %u in index \"%s\"",
1283 : BufferGetBlockNumber(buf), RelationGetRelationName(rel));
1284 :
1285 7298408 : MarkBufferDirty(buf);
1286 :
1287 7298408 : if (BufferIsValid(metabuf))
1288 : {
1289 : /* upgrade meta-page if needed */
1290 22 : if (metad->btm_version < BTREE_NOVAC_VERSION)
1291 0 : _bt_upgrademetapage(metapg);
1292 22 : metad->btm_fastroot = BufferGetBlockNumber(buf);
1293 22 : metad->btm_fastlevel = opaque->btpo_level;
1294 22 : MarkBufferDirty(metabuf);
1295 : }
1296 :
1297 : /*
1298 : * Clear INCOMPLETE_SPLIT flag on child if inserting the new item
1299 : * finishes a split
1300 : */
1301 7298408 : if (!isleaf)
1302 : {
1303 21014 : Page cpage = BufferGetPage(cbuf);
1304 21014 : BTPageOpaque cpageop = BTPageGetOpaque(cpage);
1305 :
1306 : Assert(P_INCOMPLETE_SPLIT(cpageop));
1307 21014 : cpageop->btpo_flags &= ~BTP_INCOMPLETE_SPLIT;
1308 21014 : MarkBufferDirty(cbuf);
1309 : }
1310 :
1311 : /* XLOG stuff */
1312 7298408 : if (RelationNeedsWAL(rel))
1313 : {
1314 : xl_btree_insert xlrec;
1315 : xl_btree_metadata xlmeta;
1316 : uint8 xlinfo;
1317 : XLogRecPtr recptr;
1318 : uint16 upostingoff;
1319 :
1320 6792376 : xlrec.offnum = newitemoff;
1321 :
1322 6792376 : XLogBeginInsert();
1323 6792376 : XLogRegisterData(&xlrec, SizeOfBtreeInsert);
1324 :
1325 6792376 : if (isleaf && postingoff == 0)
1326 : {
1327 : /* Simple leaf insert */
1328 6750242 : xlinfo = XLOG_BTREE_INSERT_LEAF;
1329 : }
1330 42134 : else if (postingoff != 0)
1331 : {
1332 : /*
1333 : * Leaf insert with posting list split. Must include
1334 : * postingoff field before newitem/orignewitem.
1335 : */
1336 : Assert(isleaf);
1337 22374 : xlinfo = XLOG_BTREE_INSERT_POST;
1338 : }
1339 : else
1340 : {
1341 : /* Internal page insert, which finishes a split on cbuf */
1342 19760 : xlinfo = XLOG_BTREE_INSERT_UPPER;
1343 19760 : XLogRegisterBuffer(1, cbuf, REGBUF_STANDARD);
1344 :
1345 19760 : if (BufferIsValid(metabuf))
1346 : {
1347 : /* Actually, it's an internal page insert + meta update */
1348 22 : xlinfo = XLOG_BTREE_INSERT_META;
1349 :
1350 : Assert(metad->btm_version >= BTREE_NOVAC_VERSION);
1351 22 : xlmeta.version = metad->btm_version;
1352 22 : xlmeta.root = metad->btm_root;
1353 22 : xlmeta.level = metad->btm_level;
1354 22 : xlmeta.fastroot = metad->btm_fastroot;
1355 22 : xlmeta.fastlevel = metad->btm_fastlevel;
1356 22 : xlmeta.last_cleanup_num_delpages = metad->btm_last_cleanup_num_delpages;
1357 22 : xlmeta.allequalimage = metad->btm_allequalimage;
1358 :
1359 22 : XLogRegisterBuffer(2, metabuf,
1360 : REGBUF_WILL_INIT | REGBUF_STANDARD);
1361 22 : XLogRegisterBufData(2, &xlmeta,
1362 : sizeof(xl_btree_metadata));
1363 : }
1364 : }
1365 :
1366 6792376 : XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
1367 6792376 : if (postingoff == 0)
1368 : {
1369 : /* Just log itup from caller */
1370 6770002 : XLogRegisterBufData(0, itup, IndexTupleSize(itup));
1371 : }
1372 : else
1373 : {
1374 : /*
1375 : * Insert with posting list split (XLOG_BTREE_INSERT_POST
1376 : * record) case.
1377 : *
1378 : * Log postingoff. Also log origitup, not itup. REDO routine
1379 : * must reconstruct final itup (as well as nposting) using
1380 : * _bt_swap_posting().
1381 : */
1382 22374 : upostingoff = postingoff;
1383 :
1384 22374 : XLogRegisterBufData(0, &upostingoff, sizeof(uint16));
1385 22374 : XLogRegisterBufData(0, origitup,
1386 22374 : IndexTupleSize(origitup));
1387 : }
1388 :
1389 6792376 : recptr = XLogInsert(RM_BTREE_ID, xlinfo);
1390 :
1391 6792376 : if (BufferIsValid(metabuf))
1392 22 : PageSetLSN(metapg, recptr);
1393 6792376 : if (!isleaf)
1394 19760 : PageSetLSN(BufferGetPage(cbuf), recptr);
1395 :
1396 6792376 : PageSetLSN(page, recptr);
1397 : }
1398 :
1399 7298408 : END_CRIT_SECTION();
1400 :
1401 : /* Release subsidiary buffers */
1402 7298408 : if (BufferIsValid(metabuf))
1403 22 : _bt_relbuf(rel, metabuf);
1404 7298408 : if (!isleaf)
1405 21014 : _bt_relbuf(rel, cbuf);
1406 :
1407 : /*
1408 : * Cache the block number if this is the rightmost leaf page. Cache
1409 : * may be used by a future inserter within _bt_search_insert().
1410 : */
1411 7298408 : blockcache = InvalidBlockNumber;
1412 7298408 : if (isrightmost && isleaf && !isroot)
1413 4069628 : blockcache = BufferGetBlockNumber(buf);
1414 :
1415 : /* Release buffer for insertion target block */
1416 7298408 : _bt_relbuf(rel, buf);
1417 :
1418 : /*
1419 : * If we decided to cache the insertion target block before releasing
1420 : * its buffer lock, then cache it now. Check the height of the tree
1421 : * first, though. We don't go for the optimization with small
1422 : * indexes. Defer final check to this point to ensure that we don't
1423 : * call _bt_getrootheight while holding a buffer lock.
1424 : */
1425 11368036 : if (BlockNumberIsValid(blockcache) &&
1426 4069628 : _bt_getrootheight(rel) >= BTREE_FASTPATH_MIN_LEVEL)
1427 69380 : RelationSetTargetBlock(rel, blockcache);
1428 : }
1429 :
1430 : /* be tidy */
1431 7320988 : if (postingoff != 0)
1432 : {
1433 : /* itup is actually a modified copy of caller's original */
1434 22442 : pfree(nposting);
1435 22442 : pfree(itup);
1436 : }
1437 7320988 : }
1438 :
1439 : /*
1440 : * _bt_split() -- split a page in the btree.
1441 : *
1442 : * On entry, buf is the page to split, and is pinned and write-locked.
1443 : * newitemoff etc. tell us about the new item that must be inserted
1444 : * along with the data from the original page.
1445 : *
1446 : * itup_key is used for suffix truncation on leaf pages (internal
1447 : * page callers pass NULL). When splitting a non-leaf page, 'cbuf'
1448 : * is the left-sibling of the page we're inserting the downlink for.
1449 : * This function will clear the INCOMPLETE_SPLIT flag on it, and
1450 : * release the buffer.
1451 : *
1452 : * orignewitem, nposting, and postingoff are needed when an insert of
1453 : * orignewitem results in both a posting list split and a page split.
1454 : * These extra posting list split details are used here in the same
1455 : * way as they are used in the more common case where a posting list
1456 : * split does not coincide with a page split. We need to deal with
1457 : * posting list splits directly in order to ensure that everything
1458 : * that follows from the insert of orignewitem is handled as a single
1459 : * atomic operation (though caller's insert of a new pivot/downlink
1460 : * into parent page will still be a separate operation). See
1461 : * nbtree/README for details on the design of posting list splits.
1462 : *
1463 : * Returns the new right sibling of buf, pinned and write-locked.
1464 : * The pin and lock on buf are maintained.
1465 : */
1466 : static Buffer
1467 22580 : _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
1468 : Buffer cbuf, OffsetNumber newitemoff, Size newitemsz, IndexTuple newitem,
1469 : IndexTuple orignewitem, IndexTuple nposting, uint16 postingoff)
1470 : {
1471 : Buffer rbuf;
1472 : Page origpage;
1473 : Page leftpage,
1474 : rightpage;
1475 : PGAlignedBlock leftpage_buf,
1476 : rightpage_buf;
1477 : BlockNumber origpagenumber,
1478 : rightpagenumber;
1479 : BTPageOpaque ropaque,
1480 : lopaque,
1481 : oopaque;
1482 22580 : Buffer sbuf = InvalidBuffer;
1483 22580 : Page spage = NULL;
1484 22580 : BTPageOpaque sopaque = NULL;
1485 : Size itemsz;
1486 : ItemId itemid;
1487 : IndexTuple firstright,
1488 : lefthighkey;
1489 : OffsetNumber firstrightoff;
1490 : OffsetNumber afterleftoff,
1491 : afterrightoff,
1492 : minusinfoff;
1493 : OffsetNumber origpagepostingoff;
1494 : OffsetNumber maxoff;
1495 : OffsetNumber i;
1496 : bool newitemonleft,
1497 : isleaf,
1498 : isrightmost;
1499 :
1500 : /*
1501 : * origpage is the original page to be split. leftpage is a temporary
1502 : * buffer that receives the left-sibling data, which will be copied back
1503 : * into origpage on success. rightpage is the new page that will receive
1504 : * the right-sibling data.
1505 : *
1506 : * leftpage is allocated after choosing a split point. rightpage's new
1507 : * buffer isn't acquired until after leftpage is initialized and has new
1508 : * high key, the last point where splitting the page may fail (barring
1509 : * corruption). Failing before acquiring new buffer won't have lasting
1510 : * consequences, since origpage won't have been modified and leftpage is
1511 : * only workspace.
1512 : */
1513 22580 : origpage = BufferGetPage(buf);
1514 22580 : oopaque = BTPageGetOpaque(origpage);
1515 22580 : isleaf = P_ISLEAF(oopaque);
1516 22580 : isrightmost = P_RIGHTMOST(oopaque);
1517 22580 : maxoff = PageGetMaxOffsetNumber(origpage);
1518 22580 : origpagenumber = BufferGetBlockNumber(buf);
1519 :
1520 : /*
1521 : * Choose a point to split origpage at.
1522 : *
1523 : * A split point can be thought of as a point _between_ two existing data
1524 : * items on origpage (the lastleft and firstright tuples), provided you
1525 : * pretend that the new item that didn't fit is already on origpage.
1526 : *
1527 : * Since origpage does not actually contain newitem, the representation of
1528 : * split points needs to work with two boundary cases: splits where
1529 : * newitem is lastleft, and splits where newitem is firstright.
1530 : * newitemonleft resolves the ambiguity that would otherwise exist when
1531 : * newitemoff == firstrightoff. In all other cases it's clear which side
1532 : * of the split every tuple goes on from context. newitemonleft is
1533 : * usually (but not always) redundant information.
1534 : *
1535 : * firstrightoff is supposed to be an origpage offset number, but it's
1536 : * possible that its value will be maxoff+1, which is "past the end" of
1537 : * origpage. This happens in the rare case where newitem goes after all
1538 : * existing items (i.e. newitemoff is maxoff+1) and we end up splitting
1539 : * origpage at the point that leaves newitem alone on new right page. Any
1540 : * "!newitemonleft && newitemoff == firstrightoff" split point makes
1541 : * newitem the firstright tuple, though, so this case isn't a special
1542 : * case.
1543 : */
1544 22580 : firstrightoff = _bt_findsplitloc(rel, origpage, newitemoff, newitemsz,
1545 : newitem, &newitemonleft);
1546 :
1547 : /* Use temporary buffer for leftpage */
1548 22580 : leftpage = leftpage_buf.data;
1549 22580 : _bt_pageinit(leftpage, BufferGetPageSize(buf));
1550 22580 : lopaque = BTPageGetOpaque(leftpage);
1551 :
1552 : /*
1553 : * leftpage won't be the root when we're done. Also, clear the SPLIT_END
1554 : * and HAS_GARBAGE flags.
1555 : */
1556 22580 : lopaque->btpo_flags = oopaque->btpo_flags;
1557 22580 : lopaque->btpo_flags &= ~(BTP_ROOT | BTP_SPLIT_END | BTP_HAS_GARBAGE);
1558 : /* set flag in leftpage indicating that rightpage has no downlink yet */
1559 22580 : lopaque->btpo_flags |= BTP_INCOMPLETE_SPLIT;
1560 22580 : lopaque->btpo_prev = oopaque->btpo_prev;
1561 : /* handle btpo_next after rightpage buffer acquired */
1562 22580 : lopaque->btpo_level = oopaque->btpo_level;
1563 : /* handle btpo_cycleid after rightpage buffer acquired */
1564 :
1565 : /*
1566 : * Copy the original page's LSN into leftpage, which will become the
1567 : * updated version of the page. We need this because XLogInsert will
1568 : * examine the LSN and possibly dump it in a page image.
1569 : */
1570 22580 : PageSetLSN(leftpage, PageGetLSN(origpage));
1571 :
1572 : /*
1573 : * Determine page offset number of existing overlapped-with-orignewitem
1574 : * posting list when it is necessary to perform a posting list split in
1575 : * passing. Note that newitem was already changed by caller (newitem no
1576 : * longer has the orignewitem TID).
1577 : *
1578 : * This page offset number (origpagepostingoff) will be used to pretend
1579 : * that the posting split has already taken place, even though the
1580 : * required modifications to origpage won't occur until we reach the
1581 : * critical section. The lastleft and firstright tuples of our page split
1582 : * point should, in effect, come from an imaginary version of origpage
1583 : * that has the nposting tuple instead of the original posting list tuple.
1584 : *
1585 : * Note: _bt_findsplitloc() should have compensated for coinciding posting
1586 : * list splits in just the same way, at least in theory. It doesn't
1587 : * bother with that, though. In practice it won't affect its choice of
1588 : * split point.
1589 : */
1590 22580 : origpagepostingoff = InvalidOffsetNumber;
1591 22580 : if (postingoff != 0)
1592 : {
1593 : Assert(isleaf);
1594 : Assert(ItemPointerCompare(&orignewitem->t_tid,
1595 : &newitem->t_tid) < 0);
1596 : Assert(BTreeTupleIsPosting(nposting));
1597 68 : origpagepostingoff = OffsetNumberPrev(newitemoff);
1598 : }
1599 :
1600 : /*
1601 : * The high key for the new left page is a possibly-truncated copy of
1602 : * firstright on the leaf level (it's "firstright itself" on internal
1603 : * pages; see !isleaf comments below). This may seem to be contrary to
1604 : * Lehman & Yao's approach of using a copy of lastleft as the new high key
1605 : * when splitting on the leaf level. It isn't, though.
1606 : *
1607 : * Suffix truncation will leave the left page's high key fully equal to
1608 : * lastleft when lastleft and firstright are equal prior to heap TID (that
1609 : * is, the tiebreaker TID value comes from lastleft). It isn't actually
1610 : * necessary for a new leaf high key to be a copy of lastleft for the L&Y
1611 : * "subtree" invariant to hold. It's sufficient to make sure that the new
1612 : * leaf high key is strictly less than firstright, and greater than or
1613 : * equal to (not necessarily equal to) lastleft. In other words, when
1614 : * suffix truncation isn't possible during a leaf page split, we take
1615 : * L&Y's exact approach to generating a new high key for the left page.
1616 : * (Actually, that is slightly inaccurate. We don't just use a copy of
1617 : * lastleft. A tuple with all the keys from firstright but the max heap
1618 : * TID from lastleft is used, to avoid introducing a special case.)
1619 : */
1620 22580 : if (!newitemonleft && newitemoff == firstrightoff)
1621 : {
1622 : /* incoming tuple becomes firstright */
1623 30 : itemsz = newitemsz;
1624 30 : firstright = newitem;
1625 : }
1626 : else
1627 : {
1628 : /* existing item at firstrightoff becomes firstright */
1629 22550 : itemid = PageGetItemId(origpage, firstrightoff);
1630 22550 : itemsz = ItemIdGetLength(itemid);
1631 22550 : firstright = (IndexTuple) PageGetItem(origpage, itemid);
1632 22550 : if (firstrightoff == origpagepostingoff)
1633 0 : firstright = nposting;
1634 : }
1635 :
1636 22580 : if (isleaf)
1637 : {
1638 : IndexTuple lastleft;
1639 :
1640 : /* Attempt suffix truncation for leaf page splits */
1641 22378 : if (newitemonleft && newitemoff == firstrightoff)
1642 : {
1643 : /* incoming tuple becomes lastleft */
1644 426 : lastleft = newitem;
1645 : }
1646 : else
1647 : {
1648 : OffsetNumber lastleftoff;
1649 :
1650 : /* existing item before firstrightoff becomes lastleft */
1651 21952 : lastleftoff = OffsetNumberPrev(firstrightoff);
1652 : Assert(lastleftoff >= P_FIRSTDATAKEY(oopaque));
1653 21952 : itemid = PageGetItemId(origpage, lastleftoff);
1654 21952 : lastleft = (IndexTuple) PageGetItem(origpage, itemid);
1655 21952 : if (lastleftoff == origpagepostingoff)
1656 6 : lastleft = nposting;
1657 : }
1658 :
1659 22378 : lefthighkey = _bt_truncate(rel, lastleft, firstright, itup_key);
1660 22378 : itemsz = IndexTupleSize(lefthighkey);
1661 : }
1662 : else
1663 : {
1664 : /*
1665 : * Don't perform suffix truncation on a copy of firstright to make
1666 : * left page high key for internal page splits. Must use firstright
1667 : * as new high key directly.
1668 : *
1669 : * Each distinct separator key value originates as a leaf level high
1670 : * key; all other separator keys/pivot tuples are copied from one
1671 : * level down. A separator key in a grandparent page must be
1672 : * identical to high key in rightmost parent page of the subtree to
1673 : * its left, which must itself be identical to high key in rightmost
1674 : * child page of that same subtree (this even applies to separator
1675 : * from grandparent's high key). There must always be an unbroken
1676 : * "seam" of identical separator keys that guide index scans at every
1677 : * level, starting from the grandparent. That's why suffix truncation
1678 : * is unsafe here.
1679 : *
1680 : * Internal page splits will truncate firstright into a "negative
1681 : * infinity" data item when it gets inserted on the new right page
1682 : * below, though. This happens during the call to _bt_pgaddtup() for
1683 : * the new first data item for right page. Do not confuse this
1684 : * mechanism with suffix truncation. It is just a convenient way of
1685 : * implementing page splits that split the internal page "inside"
1686 : * firstright. The lefthighkey separator key cannot appear a second
1687 : * time in the right page (only firstright's downlink goes in right
1688 : * page).
1689 : */
1690 202 : lefthighkey = firstright;
1691 : }
1692 :
1693 : /*
1694 : * Add new high key to leftpage
1695 : */
1696 22580 : afterleftoff = P_HIKEY;
1697 :
1698 : Assert(BTreeTupleGetNAtts(lefthighkey, rel) > 0);
1699 : Assert(BTreeTupleGetNAtts(lefthighkey, rel) <=
1700 : IndexRelationGetNumberOfKeyAttributes(rel));
1701 : Assert(itemsz == MAXALIGN(IndexTupleSize(lefthighkey)));
1702 22580 : if (PageAddItem(leftpage, lefthighkey, itemsz, afterleftoff, false, false) == InvalidOffsetNumber)
1703 0 : elog(ERROR, "failed to add high key to the left sibling"
1704 : " while splitting block %u of index \"%s\"",
1705 : origpagenumber, RelationGetRelationName(rel));
1706 22580 : afterleftoff = OffsetNumberNext(afterleftoff);
1707 :
1708 : /*
1709 : * Acquire a new right page to split into, now that left page has a new
1710 : * high key.
1711 : *
1712 : * To not confuse future VACUUM operations, we zero the right page and
1713 : * work on an in-memory copy of it before writing WAL, then copy its
1714 : * contents back to the actual page once we start the critical section
1715 : * work. This simplifies the split work, so as there is no need to zero
1716 : * the right page before throwing an error.
1717 : */
1718 22580 : rbuf = _bt_allocbuf(rel, heaprel);
1719 22580 : rightpage = rightpage_buf.data;
1720 :
1721 : /*
1722 : * Copy the contents of the right page into its temporary location, and
1723 : * zero the original space.
1724 : */
1725 22580 : memcpy(rightpage, BufferGetPage(rbuf), BLCKSZ);
1726 22580 : memset(BufferGetPage(rbuf), 0, BLCKSZ);
1727 22580 : rightpagenumber = BufferGetBlockNumber(rbuf);
1728 : /* rightpage was initialized by _bt_allocbuf */
1729 22580 : ropaque = BTPageGetOpaque(rightpage);
1730 :
1731 : /*
1732 : * Finish off remaining leftpage special area fields. They cannot be set
1733 : * before both origpage (leftpage) and rightpage buffers are acquired and
1734 : * locked.
1735 : *
1736 : * btpo_cycleid is only used with leaf pages, though we set it here in all
1737 : * cases just to be consistent.
1738 : */
1739 22580 : lopaque->btpo_next = rightpagenumber;
1740 22580 : lopaque->btpo_cycleid = _bt_vacuum_cycleid(rel);
1741 :
1742 : /*
1743 : * rightpage won't be the root when we're done. Also, clear the SPLIT_END
1744 : * and HAS_GARBAGE flags.
1745 : */
1746 22580 : ropaque->btpo_flags = oopaque->btpo_flags;
1747 22580 : ropaque->btpo_flags &= ~(BTP_ROOT | BTP_SPLIT_END | BTP_HAS_GARBAGE);
1748 22580 : ropaque->btpo_prev = origpagenumber;
1749 22580 : ropaque->btpo_next = oopaque->btpo_next;
1750 22580 : ropaque->btpo_level = oopaque->btpo_level;
1751 22580 : ropaque->btpo_cycleid = lopaque->btpo_cycleid;
1752 :
1753 : /*
1754 : * Add new high key to rightpage where necessary.
1755 : *
1756 : * If the page we're splitting is not the rightmost page at its level in
1757 : * the tree, then the first entry on the page is the high key from
1758 : * origpage.
1759 : */
1760 22580 : afterrightoff = P_HIKEY;
1761 :
1762 22580 : if (!isrightmost)
1763 : {
1764 : IndexTuple righthighkey;
1765 :
1766 9714 : itemid = PageGetItemId(origpage, P_HIKEY);
1767 9714 : itemsz = ItemIdGetLength(itemid);
1768 9714 : righthighkey = (IndexTuple) PageGetItem(origpage, itemid);
1769 : Assert(BTreeTupleGetNAtts(righthighkey, rel) > 0);
1770 : Assert(BTreeTupleGetNAtts(righthighkey, rel) <=
1771 : IndexRelationGetNumberOfKeyAttributes(rel));
1772 9714 : if (PageAddItem(rightpage, righthighkey, itemsz, afterrightoff, false, false) == InvalidOffsetNumber)
1773 : {
1774 0 : elog(ERROR, "failed to add high key to the right sibling"
1775 : " while splitting block %u of index \"%s\"",
1776 : origpagenumber, RelationGetRelationName(rel));
1777 : }
1778 9714 : afterrightoff = OffsetNumberNext(afterrightoff);
1779 : }
1780 :
1781 : /*
1782 : * Internal page splits truncate first data item on right page -- it
1783 : * becomes "minus infinity" item for the page. Set this up here.
1784 : */
1785 22580 : minusinfoff = InvalidOffsetNumber;
1786 22580 : if (!isleaf)
1787 202 : minusinfoff = afterrightoff;
1788 :
1789 : /*
1790 : * Now transfer all the data items (non-pivot tuples in isleaf case, or
1791 : * additional pivot tuples in !isleaf case) to the appropriate page.
1792 : *
1793 : * Note: we *must* insert at least the right page's items in item-number
1794 : * order, for the benefit of _bt_restore_page().
1795 : */
1796 6877204 : for (i = P_FIRSTDATAKEY(oopaque); i <= maxoff; i = OffsetNumberNext(i))
1797 : {
1798 : IndexTuple dataitem;
1799 :
1800 6854624 : itemid = PageGetItemId(origpage, i);
1801 6854624 : itemsz = ItemIdGetLength(itemid);
1802 6854624 : dataitem = (IndexTuple) PageGetItem(origpage, itemid);
1803 :
1804 : /* replace original item with nposting due to posting split? */
1805 6854624 : if (i == origpagepostingoff)
1806 : {
1807 : Assert(BTreeTupleIsPosting(dataitem));
1808 : Assert(itemsz == MAXALIGN(IndexTupleSize(nposting)));
1809 68 : dataitem = nposting;
1810 : }
1811 :
1812 : /* does new item belong before this one? */
1813 6854556 : else if (i == newitemoff)
1814 : {
1815 12982 : if (newitemonleft)
1816 : {
1817 : Assert(newitemoff <= firstrightoff);
1818 3378 : if (!_bt_pgaddtup(leftpage, newitemsz, newitem, afterleftoff,
1819 : false))
1820 : {
1821 0 : elog(ERROR, "failed to add new item to the left sibling"
1822 : " while splitting block %u of index \"%s\"",
1823 : origpagenumber, RelationGetRelationName(rel));
1824 : }
1825 3378 : afterleftoff = OffsetNumberNext(afterleftoff);
1826 : }
1827 : else
1828 : {
1829 : Assert(newitemoff >= firstrightoff);
1830 9604 : if (!_bt_pgaddtup(rightpage, newitemsz, newitem, afterrightoff,
1831 : afterrightoff == minusinfoff))
1832 : {
1833 0 : elog(ERROR, "failed to add new item to the right sibling"
1834 : " while splitting block %u of index \"%s\"",
1835 : origpagenumber, RelationGetRelationName(rel));
1836 : }
1837 9604 : afterrightoff = OffsetNumberNext(afterrightoff);
1838 : }
1839 : }
1840 :
1841 : /* decide which page to put it on */
1842 6854624 : if (i < firstrightoff)
1843 : {
1844 5206244 : if (!_bt_pgaddtup(leftpage, itemsz, dataitem, afterleftoff, false))
1845 : {
1846 0 : elog(ERROR, "failed to add old item to the left sibling"
1847 : " while splitting block %u of index \"%s\"",
1848 : origpagenumber, RelationGetRelationName(rel));
1849 : }
1850 5206244 : afterleftoff = OffsetNumberNext(afterleftoff);
1851 : }
1852 : else
1853 : {
1854 1648380 : if (!_bt_pgaddtup(rightpage, itemsz, dataitem, afterrightoff,
1855 : afterrightoff == minusinfoff))
1856 : {
1857 0 : elog(ERROR, "failed to add old item to the right sibling"
1858 : " while splitting block %u of index \"%s\"",
1859 : origpagenumber, RelationGetRelationName(rel));
1860 : }
1861 1648380 : afterrightoff = OffsetNumberNext(afterrightoff);
1862 : }
1863 : }
1864 :
1865 : /* Handle case where newitem goes at the end of rightpage */
1866 22580 : if (i <= newitemoff)
1867 : {
1868 : /*
1869 : * Can't have newitemonleft here; that would imply we were told to put
1870 : * *everything* on the left page, which cannot fit (if it could, we'd
1871 : * not be splitting the page).
1872 : */
1873 : Assert(!newitemonleft && newitemoff == maxoff + 1);
1874 9598 : if (!_bt_pgaddtup(rightpage, newitemsz, newitem, afterrightoff,
1875 : afterrightoff == minusinfoff))
1876 : {
1877 0 : elog(ERROR, "failed to add new item to the right sibling"
1878 : " while splitting block %u of index \"%s\"",
1879 : origpagenumber, RelationGetRelationName(rel));
1880 : }
1881 9598 : afterrightoff = OffsetNumberNext(afterrightoff);
1882 : }
1883 :
1884 : /*
1885 : * We have to grab the original right sibling (if any) and update its prev
1886 : * link. We are guaranteed that this is deadlock-free, since we couple
1887 : * the locks in the standard order: left to right.
1888 : */
1889 22580 : if (!isrightmost)
1890 : {
1891 9714 : sbuf = _bt_getbuf(rel, oopaque->btpo_next, BT_WRITE);
1892 9714 : spage = BufferGetPage(sbuf);
1893 9714 : sopaque = BTPageGetOpaque(spage);
1894 9714 : if (sopaque->btpo_prev != origpagenumber)
1895 : {
1896 0 : ereport(ERROR,
1897 : (errcode(ERRCODE_INDEX_CORRUPTED),
1898 : errmsg_internal("right sibling's left-link doesn't match: "
1899 : "block %u links to %u instead of expected %u in index \"%s\"",
1900 : oopaque->btpo_next, sopaque->btpo_prev, origpagenumber,
1901 : RelationGetRelationName(rel))));
1902 : }
1903 :
1904 : /*
1905 : * Check to see if we can set the SPLIT_END flag in the right-hand
1906 : * split page; this can save some I/O for vacuum since it need not
1907 : * proceed to the right sibling. We can set the flag if the right
1908 : * sibling has a different cycleid: that means it could not be part of
1909 : * a group of pages that were all split off from the same ancestor
1910 : * page. If you're confused, imagine that page A splits to A B and
1911 : * then again, yielding A C B, while vacuum is in progress. Tuples
1912 : * originally in A could now be in either B or C, hence vacuum must
1913 : * examine both pages. But if D, our right sibling, has a different
1914 : * cycleid then it could not contain any tuples that were in A when
1915 : * the vacuum started.
1916 : */
1917 9714 : if (sopaque->btpo_cycleid != ropaque->btpo_cycleid)
1918 0 : ropaque->btpo_flags |= BTP_SPLIT_END;
1919 : }
1920 :
1921 : /*
1922 : * Right sibling is locked, new siblings are prepared, but original page
1923 : * is not updated yet.
1924 : *
1925 : * NO EREPORT(ERROR) till right sibling is updated. We can get away with
1926 : * not starting the critical section till here because we haven't been
1927 : * scribbling on the original page yet; see comments above.
1928 : */
1929 22580 : START_CRIT_SECTION();
1930 :
1931 : /*
1932 : * By here, the original data page has been split into two new halves, and
1933 : * these are correct. The algorithm requires that the left page never
1934 : * move during a split, so we copy the new left page back on top of the
1935 : * original. We need to do this before writing the WAL record, so that
1936 : * XLogInsert can WAL log an image of the page if necessary.
1937 : */
1938 22580 : memcpy(origpage, leftpage, BLCKSZ);
1939 : /* leftpage, lopaque must not be used below here */
1940 :
1941 : /*
1942 : * Move the contents of the right page from its temporary location to the
1943 : * destination buffer, before writing the WAL record. Unlike the left
1944 : * page, the right page and its opaque area are still needed to complete
1945 : * the update of the page, so reinitialize them.
1946 : */
1947 22580 : rightpage = BufferGetPage(rbuf);
1948 22580 : memcpy(rightpage, rightpage_buf.data, BLCKSZ);
1949 22580 : ropaque = BTPageGetOpaque(rightpage);
1950 :
1951 22580 : MarkBufferDirty(buf);
1952 22580 : MarkBufferDirty(rbuf);
1953 :
1954 22580 : if (!isrightmost)
1955 : {
1956 9714 : sopaque->btpo_prev = rightpagenumber;
1957 9714 : MarkBufferDirty(sbuf);
1958 : }
1959 :
1960 : /*
1961 : * Clear INCOMPLETE_SPLIT flag on child if inserting the new item finishes
1962 : * a split
1963 : */
1964 22580 : if (!isleaf)
1965 : {
1966 202 : Page cpage = BufferGetPage(cbuf);
1967 202 : BTPageOpaque cpageop = BTPageGetOpaque(cpage);
1968 :
1969 202 : cpageop->btpo_flags &= ~BTP_INCOMPLETE_SPLIT;
1970 202 : MarkBufferDirty(cbuf);
1971 : }
1972 :
1973 : /* XLOG stuff */
1974 22580 : if (RelationNeedsWAL(rel))
1975 : {
1976 : xl_btree_split xlrec;
1977 : uint8 xlinfo;
1978 : XLogRecPtr recptr;
1979 :
1980 21290 : xlrec.level = ropaque->btpo_level;
1981 : /* See comments below on newitem, orignewitem, and posting lists */
1982 21290 : xlrec.firstrightoff = firstrightoff;
1983 21290 : xlrec.newitemoff = newitemoff;
1984 21290 : xlrec.postingoff = 0;
1985 21290 : if (postingoff != 0 && origpagepostingoff < firstrightoff)
1986 38 : xlrec.postingoff = postingoff;
1987 :
1988 21290 : XLogBeginInsert();
1989 21290 : XLogRegisterData(&xlrec, SizeOfBtreeSplit);
1990 :
1991 21290 : XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
1992 21290 : XLogRegisterBuffer(1, rbuf, REGBUF_WILL_INIT);
1993 : /* Log original right sibling, since we've changed its prev-pointer */
1994 21290 : if (!isrightmost)
1995 9702 : XLogRegisterBuffer(2, sbuf, REGBUF_STANDARD);
1996 21290 : if (!isleaf)
1997 202 : XLogRegisterBuffer(3, cbuf, REGBUF_STANDARD);
1998 :
1999 : /*
2000 : * Log the new item, if it was inserted on the left page. (If it was
2001 : * put on the right page, we don't need to explicitly WAL log it
2002 : * because it's included with all the other items on the right page.)
2003 : * Show the new item as belonging to the left page buffer, so that it
2004 : * is not stored if XLogInsert decides it needs a full-page image of
2005 : * the left page. We always store newitemoff in the record, though.
2006 : *
2007 : * The details are sometimes slightly different for page splits that
2008 : * coincide with a posting list split. If both the replacement
2009 : * posting list and newitem go on the right page, then we don't need
2010 : * to log anything extra, just like the simple !newitemonleft
2011 : * no-posting-split case (postingoff is set to zero in the WAL record,
2012 : * so recovery doesn't need to process a posting list split at all).
2013 : * Otherwise, we set postingoff and log orignewitem instead of
2014 : * newitem, despite having actually inserted newitem. REDO routine
2015 : * must reconstruct nposting and newitem using _bt_swap_posting().
2016 : *
2017 : * Note: It's possible that our page split point is the point that
2018 : * makes the posting list lastleft and newitem firstright. This is
2019 : * the only case where we log orignewitem/newitem despite newitem
2020 : * going on the right page. If XLogInsert decides that it can omit
2021 : * orignewitem due to logging a full-page image of the left page,
2022 : * everything still works out, since recovery only needs to log
2023 : * orignewitem for items on the left page (just like the regular
2024 : * newitem-logged case).
2025 : */
2026 21290 : if (newitemonleft && xlrec.postingoff == 0)
2027 3340 : XLogRegisterBufData(0, newitem, newitemsz);
2028 17950 : else if (xlrec.postingoff != 0)
2029 : {
2030 : Assert(isleaf);
2031 : Assert(newitemonleft || firstrightoff == newitemoff);
2032 : Assert(newitemsz == IndexTupleSize(orignewitem));
2033 38 : XLogRegisterBufData(0, orignewitem, newitemsz);
2034 : }
2035 :
2036 : /* Log the left page's new high key */
2037 21290 : if (!isleaf)
2038 : {
2039 : /* lefthighkey isn't local copy, get current pointer */
2040 202 : itemid = PageGetItemId(origpage, P_HIKEY);
2041 202 : lefthighkey = (IndexTuple) PageGetItem(origpage, itemid);
2042 : }
2043 21290 : XLogRegisterBufData(0, lefthighkey,
2044 21290 : MAXALIGN(IndexTupleSize(lefthighkey)));
2045 :
2046 : /*
2047 : * Log the contents of the right page in the format understood by
2048 : * _bt_restore_page(). The whole right page will be recreated.
2049 : *
2050 : * Direct access to page is not good but faster - we should implement
2051 : * some new func in page API. Note we only store the tuples
2052 : * themselves, knowing that they were inserted in item-number order
2053 : * and so the line pointers can be reconstructed. See comments for
2054 : * _bt_restore_page().
2055 : */
2056 21290 : XLogRegisterBufData(1,
2057 21290 : (char *) rightpage + ((PageHeader) rightpage)->pd_upper,
2058 21290 : ((PageHeader) rightpage)->pd_special - ((PageHeader) rightpage)->pd_upper);
2059 :
2060 21290 : xlinfo = newitemonleft ? XLOG_BTREE_SPLIT_L : XLOG_BTREE_SPLIT_R;
2061 21290 : recptr = XLogInsert(RM_BTREE_ID, xlinfo);
2062 :
2063 21290 : PageSetLSN(origpage, recptr);
2064 21290 : PageSetLSN(rightpage, recptr);
2065 21290 : if (!isrightmost)
2066 9702 : PageSetLSN(spage, recptr);
2067 21290 : if (!isleaf)
2068 202 : PageSetLSN(BufferGetPage(cbuf), recptr);
2069 : }
2070 :
2071 22580 : END_CRIT_SECTION();
2072 :
2073 : /* release the old right sibling */
2074 22580 : if (!isrightmost)
2075 9714 : _bt_relbuf(rel, sbuf);
2076 :
2077 : /* release the child */
2078 22580 : if (!isleaf)
2079 202 : _bt_relbuf(rel, cbuf);
2080 :
2081 : /* be tidy */
2082 22580 : if (isleaf)
2083 22378 : pfree(lefthighkey);
2084 :
2085 : /* split's done */
2086 22580 : return rbuf;
2087 : }
2088 :
2089 : /*
2090 : * _bt_insert_parent() -- Insert downlink into parent, completing split.
2091 : *
2092 : * On entry, buf and rbuf are the left and right split pages, which we
2093 : * still hold write locks on. Both locks will be released here. We
2094 : * release the rbuf lock once we have a write lock on the page that we
2095 : * intend to insert a downlink to rbuf on (i.e. buf's current parent page).
2096 : * The lock on buf is released at the same point as the lock on the parent
2097 : * page, since buf's INCOMPLETE_SPLIT flag must be cleared by the same
2098 : * atomic operation that completes the split by inserting a new downlink.
2099 : *
2100 : * stack - stack showing how we got here. Will be NULL when splitting true
2101 : * root, or during concurrent root split, where we can be inefficient
2102 : * isroot - we split the true root
2103 : * isonly - we split a page alone on its level (might have been fast root)
2104 : */
2105 : static void
2106 22580 : _bt_insert_parent(Relation rel,
2107 : Relation heaprel,
2108 : Buffer buf,
2109 : Buffer rbuf,
2110 : BTStack stack,
2111 : bool isroot,
2112 : bool isonly)
2113 : {
2114 : Assert(heaprel != NULL);
2115 :
2116 : /*
2117 : * Here we have to do something Lehman and Yao don't talk about: deal with
2118 : * a root split and construction of a new root. If our stack is empty
2119 : * then we have just split a node on what had been the root level when we
2120 : * descended the tree. If it was still the root then we perform a
2121 : * new-root construction. If it *wasn't* the root anymore, search to find
2122 : * the next higher level that someone constructed meanwhile, and find the
2123 : * right place to insert as for the normal case.
2124 : *
2125 : * If we have to search for the parent level, we do so by re-descending
2126 : * from the root. This is not super-efficient, but it's rare enough not
2127 : * to matter.
2128 : */
2129 22580 : if (isroot)
2130 : {
2131 : Buffer rootbuf;
2132 :
2133 : Assert(stack == NULL);
2134 : Assert(isonly);
2135 : /* create a new root node one level up and update the metapage */
2136 1364 : rootbuf = _bt_newlevel(rel, heaprel, buf, rbuf);
2137 : /* release the split buffers */
2138 1364 : _bt_relbuf(rel, rootbuf);
2139 1364 : _bt_relbuf(rel, rbuf);
2140 1364 : _bt_relbuf(rel, buf);
2141 : }
2142 : else
2143 : {
2144 21216 : BlockNumber bknum = BufferGetBlockNumber(buf);
2145 21216 : BlockNumber rbknum = BufferGetBlockNumber(rbuf);
2146 21216 : Page page = BufferGetPage(buf);
2147 : IndexTuple new_item;
2148 : BTStackData fakestack;
2149 : IndexTuple ritem;
2150 : Buffer pbuf;
2151 :
2152 21216 : if (stack == NULL)
2153 : {
2154 : BTPageOpaque opaque;
2155 :
2156 22 : elog(DEBUG2, "concurrent ROOT page split");
2157 22 : opaque = BTPageGetOpaque(page);
2158 :
2159 : /*
2160 : * We should never reach here when a leaf page split takes place
2161 : * despite the insert of newitem being able to apply the fastpath
2162 : * optimization. Make sure of that with an assertion.
2163 : *
2164 : * This is more of a performance issue than a correctness issue.
2165 : * The fastpath won't have a descent stack. Using a phony stack
2166 : * here works, but never rely on that. The fastpath should be
2167 : * rejected within _bt_search_insert() when the rightmost leaf
2168 : * page will split, since it's faster to go through _bt_search()
2169 : * and get a stack in the usual way.
2170 : */
2171 : Assert(!(P_ISLEAF(opaque) &&
2172 : BlockNumberIsValid(RelationGetTargetBlock(rel))));
2173 :
2174 : /* Find the leftmost page at the next level up */
2175 22 : pbuf = _bt_get_endpoint(rel, opaque->btpo_level + 1, false);
2176 : /* Set up a phony stack entry pointing there */
2177 22 : stack = &fakestack;
2178 22 : stack->bts_blkno = BufferGetBlockNumber(pbuf);
2179 22 : stack->bts_offset = InvalidOffsetNumber;
2180 22 : stack->bts_parent = NULL;
2181 22 : _bt_relbuf(rel, pbuf);
2182 : }
2183 :
2184 : /* get high key from left, a strict lower bound for new right page */
2185 21216 : ritem = (IndexTuple) PageGetItem(page,
2186 21216 : PageGetItemId(page, P_HIKEY));
2187 :
2188 : /* form an index tuple that points at the new right page */
2189 21216 : new_item = CopyIndexTuple(ritem);
2190 21216 : BTreeTupleSetDownLink(new_item, rbknum);
2191 :
2192 : /*
2193 : * Re-find and write lock the parent of buf.
2194 : *
2195 : * It's possible that the location of buf's downlink has changed since
2196 : * our initial _bt_search() descent. _bt_getstackbuf() will detect
2197 : * and recover from this, updating the stack, which ensures that the
2198 : * new downlink will be inserted at the correct offset. Even buf's
2199 : * parent may have changed.
2200 : */
2201 21216 : pbuf = _bt_getstackbuf(rel, heaprel, stack, bknum);
2202 :
2203 : /*
2204 : * Unlock the right child. The left child will be unlocked in
2205 : * _bt_insertonpg().
2206 : *
2207 : * Unlocking the right child must be delayed until here to ensure that
2208 : * no concurrent VACUUM operation can become confused. Page deletion
2209 : * cannot be allowed to fail to re-find a downlink for the rbuf page.
2210 : * (Actually, this is just a vestige of how things used to work. The
2211 : * page deletion code is expected to check for the INCOMPLETE_SPLIT
2212 : * flag on the left child. It won't attempt deletion of the right
2213 : * child until the split is complete. Despite all this, we opt to
2214 : * conservatively delay unlocking the right child until here.)
2215 : */
2216 21216 : _bt_relbuf(rel, rbuf);
2217 :
2218 21216 : if (pbuf == InvalidBuffer)
2219 0 : ereport(ERROR,
2220 : (errcode(ERRCODE_INDEX_CORRUPTED),
2221 : errmsg_internal("failed to re-find parent key in index \"%s\" for split pages %u/%u",
2222 : RelationGetRelationName(rel), bknum, rbknum)));
2223 :
2224 : /* Recursively insert into the parent */
2225 42432 : _bt_insertonpg(rel, heaprel, NULL, pbuf, buf, stack->bts_parent,
2226 21216 : new_item, MAXALIGN(IndexTupleSize(new_item)),
2227 21216 : stack->bts_offset + 1, 0, isonly);
2228 :
2229 : /* be tidy */
2230 21216 : pfree(new_item);
2231 : }
2232 22580 : }
2233 :
2234 : /*
2235 : * _bt_finish_split() -- Finish an incomplete split
2236 : *
2237 : * A crash or other failure can leave a split incomplete. The insertion
2238 : * routines won't allow to insert on a page that is incompletely split.
2239 : * Before inserting on such a page, call _bt_finish_split().
2240 : *
2241 : * On entry, 'lbuf' must be locked in write-mode. On exit, it is unlocked
2242 : * and unpinned.
2243 : *
2244 : * Caller must provide a valid heaprel, since finishing a page split requires
2245 : * allocating a new page if and when the parent page splits in turn.
2246 : */
2247 : void
2248 0 : _bt_finish_split(Relation rel, Relation heaprel, Buffer lbuf, BTStack stack)
2249 : {
2250 0 : Page lpage = BufferGetPage(lbuf);
2251 0 : BTPageOpaque lpageop = BTPageGetOpaque(lpage);
2252 : Buffer rbuf;
2253 : Page rpage;
2254 : BTPageOpaque rpageop;
2255 : bool wasroot;
2256 : bool wasonly;
2257 :
2258 : Assert(P_INCOMPLETE_SPLIT(lpageop));
2259 : Assert(heaprel != NULL);
2260 :
2261 : /* Lock right sibling, the one missing the downlink */
2262 0 : rbuf = _bt_getbuf(rel, lpageop->btpo_next, BT_WRITE);
2263 0 : rpage = BufferGetPage(rbuf);
2264 0 : rpageop = BTPageGetOpaque(rpage);
2265 :
2266 : /* Could this be a root split? */
2267 0 : if (!stack)
2268 : {
2269 : Buffer metabuf;
2270 : Page metapg;
2271 : BTMetaPageData *metad;
2272 :
2273 : /* acquire lock on the metapage */
2274 0 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_WRITE);
2275 0 : metapg = BufferGetPage(metabuf);
2276 0 : metad = BTPageGetMeta(metapg);
2277 :
2278 0 : wasroot = (metad->btm_root == BufferGetBlockNumber(lbuf));
2279 :
2280 0 : _bt_relbuf(rel, metabuf);
2281 : }
2282 : else
2283 0 : wasroot = false;
2284 :
2285 : /* Was this the only page on the level before split? */
2286 0 : wasonly = (P_LEFTMOST(lpageop) && P_RIGHTMOST(rpageop));
2287 :
2288 0 : elog(DEBUG1, "finishing incomplete split of %u/%u",
2289 : BufferGetBlockNumber(lbuf), BufferGetBlockNumber(rbuf));
2290 :
2291 0 : _bt_insert_parent(rel, heaprel, lbuf, rbuf, stack, wasroot, wasonly);
2292 0 : }
2293 :
2294 : /*
2295 : * _bt_getstackbuf() -- Walk back up the tree one step, and find the pivot
2296 : * tuple whose downlink points to child page.
2297 : *
2298 : * Caller passes child's block number, which is used to identify
2299 : * associated pivot tuple in parent page using a linear search that
2300 : * matches on pivot's downlink/block number. The expected location of
2301 : * the pivot tuple is taken from the stack one level above the child
2302 : * page. This is used as a starting point. Insertions into the
2303 : * parent level could cause the pivot tuple to move right; deletions
2304 : * could cause it to move left, but not left of the page we previously
2305 : * found it on.
2306 : *
2307 : * Caller can use its stack to relocate the pivot tuple/downlink for
2308 : * any same-level page to the right of the page found by its initial
2309 : * descent. This is necessary because of the possibility that caller
2310 : * moved right to recover from a concurrent page split. It's also
2311 : * convenient for certain callers to be able to step right when there
2312 : * wasn't a concurrent page split, while still using their original
2313 : * stack. For example, the checkingunique _bt_doinsert() case may
2314 : * have to step right when there are many physical duplicates, and its
2315 : * scantid forces an insertion to the right of the "first page the
2316 : * value could be on". (This is also relied on by all of our callers
2317 : * when dealing with !heapkeyspace indexes.)
2318 : *
2319 : * Returns write-locked parent page buffer, or InvalidBuffer if pivot
2320 : * tuple not found (should not happen). Adjusts bts_blkno &
2321 : * bts_offset if changed. Page split caller should insert its new
2322 : * pivot tuple for its new right sibling page on parent page, at the
2323 : * offset number bts_offset + 1.
2324 : */
2325 : Buffer
2326 27008 : _bt_getstackbuf(Relation rel, Relation heaprel, BTStack stack, BlockNumber child)
2327 : {
2328 : BlockNumber blkno;
2329 : OffsetNumber start;
2330 :
2331 27008 : blkno = stack->bts_blkno;
2332 27008 : start = stack->bts_offset;
2333 :
2334 : for (;;)
2335 0 : {
2336 : Buffer buf;
2337 : Page page;
2338 : BTPageOpaque opaque;
2339 :
2340 27008 : buf = _bt_getbuf(rel, blkno, BT_WRITE);
2341 27008 : page = BufferGetPage(buf);
2342 27008 : opaque = BTPageGetOpaque(page);
2343 :
2344 : Assert(heaprel != NULL);
2345 27008 : if (P_INCOMPLETE_SPLIT(opaque))
2346 : {
2347 0 : _bt_finish_split(rel, heaprel, buf, stack->bts_parent);
2348 0 : continue;
2349 : }
2350 :
2351 27008 : if (!P_IGNORE(opaque))
2352 : {
2353 : OffsetNumber offnum,
2354 : minoff,
2355 : maxoff;
2356 : ItemId itemid;
2357 : IndexTuple item;
2358 :
2359 27008 : minoff = P_FIRSTDATAKEY(opaque);
2360 27008 : maxoff = PageGetMaxOffsetNumber(page);
2361 :
2362 : /*
2363 : * start = InvalidOffsetNumber means "search the whole page". We
2364 : * need this test anyway due to possibility that page has a high
2365 : * key now when it didn't before.
2366 : */
2367 27008 : if (start < minoff)
2368 22 : start = minoff;
2369 :
2370 : /*
2371 : * Need this check too, to guard against possibility that page
2372 : * split since we visited it originally.
2373 : */
2374 27008 : if (start > maxoff)
2375 0 : start = OffsetNumberNext(maxoff);
2376 :
2377 : /*
2378 : * These loops will check every item on the page --- but in an
2379 : * order that's attuned to the probability of where it actually
2380 : * is. Scan to the right first, then to the left.
2381 : */
2382 27008 : for (offnum = start;
2383 27100 : offnum <= maxoff;
2384 92 : offnum = OffsetNumberNext(offnum))
2385 : {
2386 27100 : itemid = PageGetItemId(page, offnum);
2387 27100 : item = (IndexTuple) PageGetItem(page, itemid);
2388 :
2389 27100 : if (BTreeTupleGetDownLink(item) == child)
2390 : {
2391 : /* Return accurate pointer to where link is now */
2392 27008 : stack->bts_blkno = blkno;
2393 27008 : stack->bts_offset = offnum;
2394 27008 : return buf;
2395 : }
2396 : }
2397 :
2398 0 : for (offnum = OffsetNumberPrev(start);
2399 0 : offnum >= minoff;
2400 0 : offnum = OffsetNumberPrev(offnum))
2401 : {
2402 0 : itemid = PageGetItemId(page, offnum);
2403 0 : item = (IndexTuple) PageGetItem(page, itemid);
2404 :
2405 0 : if (BTreeTupleGetDownLink(item) == child)
2406 : {
2407 : /* Return accurate pointer to where link is now */
2408 0 : stack->bts_blkno = blkno;
2409 0 : stack->bts_offset = offnum;
2410 0 : return buf;
2411 : }
2412 : }
2413 : }
2414 :
2415 : /*
2416 : * The item we're looking for moved right at least one page.
2417 : *
2418 : * Lehman and Yao couple/chain locks when moving right here, which we
2419 : * can avoid. See nbtree/README.
2420 : */
2421 0 : if (P_RIGHTMOST(opaque))
2422 : {
2423 0 : _bt_relbuf(rel, buf);
2424 0 : return InvalidBuffer;
2425 : }
2426 0 : blkno = opaque->btpo_next;
2427 0 : start = InvalidOffsetNumber;
2428 0 : _bt_relbuf(rel, buf);
2429 : }
2430 : }
2431 :
2432 : /*
2433 : * _bt_newlevel() -- Create a new level above root page.
2434 : *
2435 : * We've just split the old root page and need to create a new one.
2436 : * In order to do this, we add a new root page to the file, then lock
2437 : * the metadata page and update it. This is guaranteed to be deadlock-
2438 : * free, because all readers release their locks on the metadata page
2439 : * before trying to lock the root, and all writers lock the root before
2440 : * trying to lock the metadata page. We have a write lock on the old
2441 : * root page, so we have not introduced any cycles into the waits-for
2442 : * graph.
2443 : *
2444 : * On entry, lbuf (the old root) and rbuf (its new peer) are write-
2445 : * locked. On exit, a new root page exists with entries for the
2446 : * two new children, metapage is updated and unlocked/unpinned.
2447 : * The new root buffer is returned to caller which has to unlock/unpin
2448 : * lbuf, rbuf & rootbuf.
2449 : */
2450 : static Buffer
2451 1364 : _bt_newlevel(Relation rel, Relation heaprel, Buffer lbuf, Buffer rbuf)
2452 : {
2453 : Buffer rootbuf;
2454 : Page lpage,
2455 : rootpage;
2456 : BlockNumber lbkno,
2457 : rbkno;
2458 : BlockNumber rootblknum;
2459 : BTPageOpaque rootopaque;
2460 : BTPageOpaque lopaque;
2461 : ItemId itemid;
2462 : IndexTuple item;
2463 : IndexTuple left_item;
2464 : Size left_item_sz;
2465 : IndexTuple right_item;
2466 : Size right_item_sz;
2467 : Buffer metabuf;
2468 : Page metapg;
2469 : BTMetaPageData *metad;
2470 :
2471 1364 : lbkno = BufferGetBlockNumber(lbuf);
2472 1364 : rbkno = BufferGetBlockNumber(rbuf);
2473 1364 : lpage = BufferGetPage(lbuf);
2474 1364 : lopaque = BTPageGetOpaque(lpage);
2475 :
2476 : /* get a new root page */
2477 1364 : rootbuf = _bt_allocbuf(rel, heaprel);
2478 1364 : rootpage = BufferGetPage(rootbuf);
2479 1364 : rootblknum = BufferGetBlockNumber(rootbuf);
2480 :
2481 : /* acquire lock on the metapage */
2482 1364 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_WRITE);
2483 1364 : metapg = BufferGetPage(metabuf);
2484 1364 : metad = BTPageGetMeta(metapg);
2485 :
2486 : /*
2487 : * Create downlink item for left page (old root). The key value used is
2488 : * "minus infinity", a sentinel value that's reliably less than any real
2489 : * key value that could appear in the left page.
2490 : */
2491 1364 : left_item_sz = sizeof(IndexTupleData);
2492 1364 : left_item = (IndexTuple) palloc(left_item_sz);
2493 1364 : left_item->t_info = left_item_sz;
2494 1364 : BTreeTupleSetDownLink(left_item, lbkno);
2495 1364 : BTreeTupleSetNAtts(left_item, 0, false);
2496 :
2497 : /*
2498 : * Create downlink item for right page. The key for it is obtained from
2499 : * the "high key" position in the left page.
2500 : */
2501 1364 : itemid = PageGetItemId(lpage, P_HIKEY);
2502 1364 : right_item_sz = ItemIdGetLength(itemid);
2503 1364 : item = (IndexTuple) PageGetItem(lpage, itemid);
2504 1364 : right_item = CopyIndexTuple(item);
2505 1364 : BTreeTupleSetDownLink(right_item, rbkno);
2506 :
2507 : /* NO EREPORT(ERROR) from here till newroot op is logged */
2508 1364 : START_CRIT_SECTION();
2509 :
2510 : /* upgrade metapage if needed */
2511 1364 : if (metad->btm_version < BTREE_NOVAC_VERSION)
2512 0 : _bt_upgrademetapage(metapg);
2513 :
2514 : /* set btree special data */
2515 1364 : rootopaque = BTPageGetOpaque(rootpage);
2516 1364 : rootopaque->btpo_prev = rootopaque->btpo_next = P_NONE;
2517 1364 : rootopaque->btpo_flags = BTP_ROOT;
2518 1364 : rootopaque->btpo_level =
2519 1364 : (BTPageGetOpaque(lpage))->btpo_level + 1;
2520 1364 : rootopaque->btpo_cycleid = 0;
2521 :
2522 : /* update metapage data */
2523 1364 : metad->btm_root = rootblknum;
2524 1364 : metad->btm_level = rootopaque->btpo_level;
2525 1364 : metad->btm_fastroot = rootblknum;
2526 1364 : metad->btm_fastlevel = rootopaque->btpo_level;
2527 :
2528 : /*
2529 : * Insert the left page pointer into the new root page. The root page is
2530 : * the rightmost page on its level so there is no "high key" in it; the
2531 : * two items will go into positions P_HIKEY and P_FIRSTKEY.
2532 : *
2533 : * Note: we *must* insert the two items in item-number order, for the
2534 : * benefit of _bt_restore_page().
2535 : */
2536 : Assert(BTreeTupleGetNAtts(left_item, rel) == 0);
2537 1364 : if (PageAddItem(rootpage, left_item, left_item_sz, P_HIKEY, false, false) == InvalidOffsetNumber)
2538 0 : elog(PANIC, "failed to add leftkey to new root page"
2539 : " while splitting block %u of index \"%s\"",
2540 : BufferGetBlockNumber(lbuf), RelationGetRelationName(rel));
2541 :
2542 : /*
2543 : * insert the right page pointer into the new root page.
2544 : */
2545 : Assert(BTreeTupleGetNAtts(right_item, rel) > 0);
2546 : Assert(BTreeTupleGetNAtts(right_item, rel) <=
2547 : IndexRelationGetNumberOfKeyAttributes(rel));
2548 1364 : if (PageAddItem(rootpage, right_item, right_item_sz, P_FIRSTKEY, false, false) == InvalidOffsetNumber)
2549 0 : elog(PANIC, "failed to add rightkey to new root page"
2550 : " while splitting block %u of index \"%s\"",
2551 : BufferGetBlockNumber(lbuf), RelationGetRelationName(rel));
2552 :
2553 : /* Clear the incomplete-split flag in the left child */
2554 : Assert(P_INCOMPLETE_SPLIT(lopaque));
2555 1364 : lopaque->btpo_flags &= ~BTP_INCOMPLETE_SPLIT;
2556 1364 : MarkBufferDirty(lbuf);
2557 :
2558 1364 : MarkBufferDirty(rootbuf);
2559 1364 : MarkBufferDirty(metabuf);
2560 :
2561 : /* XLOG stuff */
2562 1364 : if (RelationNeedsWAL(rel))
2563 : {
2564 : xl_btree_newroot xlrec;
2565 : XLogRecPtr recptr;
2566 : xl_btree_metadata md;
2567 :
2568 1328 : xlrec.rootblk = rootblknum;
2569 1328 : xlrec.level = metad->btm_level;
2570 :
2571 1328 : XLogBeginInsert();
2572 1328 : XLogRegisterData(&xlrec, SizeOfBtreeNewroot);
2573 :
2574 1328 : XLogRegisterBuffer(0, rootbuf, REGBUF_WILL_INIT);
2575 1328 : XLogRegisterBuffer(1, lbuf, REGBUF_STANDARD);
2576 1328 : XLogRegisterBuffer(2, metabuf, REGBUF_WILL_INIT | REGBUF_STANDARD);
2577 :
2578 : Assert(metad->btm_version >= BTREE_NOVAC_VERSION);
2579 1328 : md.version = metad->btm_version;
2580 1328 : md.root = rootblknum;
2581 1328 : md.level = metad->btm_level;
2582 1328 : md.fastroot = rootblknum;
2583 1328 : md.fastlevel = metad->btm_level;
2584 1328 : md.last_cleanup_num_delpages = metad->btm_last_cleanup_num_delpages;
2585 1328 : md.allequalimage = metad->btm_allequalimage;
2586 :
2587 1328 : XLogRegisterBufData(2, &md, sizeof(xl_btree_metadata));
2588 :
2589 : /*
2590 : * Direct access to page is not good but faster - we should implement
2591 : * some new func in page API.
2592 : */
2593 1328 : XLogRegisterBufData(0,
2594 1328 : (char *) rootpage + ((PageHeader) rootpage)->pd_upper,
2595 1328 : ((PageHeader) rootpage)->pd_special -
2596 1328 : ((PageHeader) rootpage)->pd_upper);
2597 :
2598 1328 : recptr = XLogInsert(RM_BTREE_ID, XLOG_BTREE_NEWROOT);
2599 :
2600 1328 : PageSetLSN(lpage, recptr);
2601 1328 : PageSetLSN(rootpage, recptr);
2602 1328 : PageSetLSN(metapg, recptr);
2603 : }
2604 :
2605 1364 : END_CRIT_SECTION();
2606 :
2607 : /* done with metapage */
2608 1364 : _bt_relbuf(rel, metabuf);
2609 :
2610 1364 : pfree(left_item);
2611 1364 : pfree(right_item);
2612 :
2613 1364 : return rootbuf;
2614 : }
2615 :
2616 : /*
2617 : * _bt_pgaddtup() -- add a data item to a particular page during split.
2618 : *
2619 : * The difference between this routine and a bare PageAddItem call is
2620 : * that this code can deal with the first data item on an internal btree
2621 : * page in passing. This data item (which is called "firstright" within
2622 : * _bt_split()) has a key that must be treated as minus infinity after
2623 : * the split. Therefore, we truncate away all attributes when caller
2624 : * specifies it's the first data item on page (downlink is not changed,
2625 : * though). This extra step is only needed for the right page of an
2626 : * internal page split. There is no need to do this for the first data
2627 : * item on the existing/left page, since that will already have been
2628 : * truncated during an earlier page split.
2629 : *
2630 : * See _bt_split() for a high level explanation of why we truncate here.
2631 : * Note that this routine has nothing to do with suffix truncation,
2632 : * despite using some of the same infrastructure.
2633 : */
2634 : static inline bool
2635 6877204 : _bt_pgaddtup(Page page,
2636 : Size itemsize,
2637 : const IndexTupleData *itup,
2638 : OffsetNumber itup_off,
2639 : bool newfirstdataitem)
2640 : {
2641 : IndexTupleData trunctuple;
2642 :
2643 6877204 : if (newfirstdataitem)
2644 : {
2645 202 : trunctuple = *itup;
2646 202 : trunctuple.t_info = sizeof(IndexTupleData);
2647 202 : BTreeTupleSetNAtts(&trunctuple, 0, false);
2648 202 : itup = &trunctuple;
2649 202 : itemsize = sizeof(IndexTupleData);
2650 : }
2651 :
2652 6877204 : if (unlikely(PageAddItem(page, itup, itemsize, itup_off, false, false) == InvalidOffsetNumber))
2653 0 : return false;
2654 :
2655 6877204 : return true;
2656 : }
2657 :
2658 : /*
2659 : * _bt_delete_or_dedup_one_page - Try to avoid a leaf page split.
2660 : *
2661 : * There are three operations performed here: simple index deletion, bottom-up
2662 : * index deletion, and deduplication. If all three operations fail to free
2663 : * enough space for the incoming item then caller will go on to split the
2664 : * page. We always consider simple deletion first. If that doesn't work out
2665 : * we consider alternatives. Callers that only want us to consider simple
2666 : * deletion (without any fallback) ask for that using the 'simpleonly'
2667 : * argument.
2668 : *
2669 : * We usually pick only one alternative "complex" operation when simple
2670 : * deletion alone won't prevent a page split. The 'checkingunique',
2671 : * 'uniquedup', and 'indexUnchanged' arguments are used for that.
2672 : *
2673 : * Note: We used to only delete LP_DEAD items when the BTP_HAS_GARBAGE page
2674 : * level flag was found set. The flag was useful back when there wasn't
2675 : * necessarily one single page for a duplicate tuple to go on (before heap TID
2676 : * became a part of the key space in version 4 indexes). But we don't
2677 : * actually look at the flag anymore (it's not a gating condition for our
2678 : * caller). That would cause us to miss tuples that are safe to delete,
2679 : * without getting any benefit in return. We know that the alternative is to
2680 : * split the page; scanning the line pointer array in passing won't have
2681 : * noticeable overhead. (We still maintain the BTP_HAS_GARBAGE flag despite
2682 : * all this because !heapkeyspace indexes must still do a "getting tired"
2683 : * linear search, and so are likely to get some benefit from using it as a
2684 : * gating condition.)
2685 : */
2686 : static void
2687 51290 : _bt_delete_or_dedup_one_page(Relation rel, Relation heapRel,
2688 : BTInsertState insertstate,
2689 : bool simpleonly, bool checkingunique,
2690 : bool uniquedup, bool indexUnchanged)
2691 : {
2692 : OffsetNumber deletable[MaxIndexTuplesPerPage];
2693 51290 : int ndeletable = 0;
2694 : OffsetNumber offnum,
2695 : minoff,
2696 : maxoff;
2697 51290 : Buffer buffer = insertstate->buf;
2698 51290 : BTScanInsert itup_key = insertstate->itup_key;
2699 51290 : Page page = BufferGetPage(buffer);
2700 51290 : BTPageOpaque opaque = BTPageGetOpaque(page);
2701 :
2702 : Assert(P_ISLEAF(opaque));
2703 : Assert(simpleonly || itup_key->heapkeyspace);
2704 : Assert(!simpleonly || (!checkingunique && !uniquedup && !indexUnchanged));
2705 :
2706 : /*
2707 : * Scan over all items to see which ones need to be deleted according to
2708 : * LP_DEAD flags. We'll usually manage to delete a few extra items that
2709 : * are not marked LP_DEAD in passing. Often the extra items that actually
2710 : * end up getting deleted are items that would have had their LP_DEAD bit
2711 : * set before long anyway (if we opted not to include them as extras).
2712 : */
2713 51290 : minoff = P_FIRSTDATAKEY(opaque);
2714 51290 : maxoff = PageGetMaxOffsetNumber(page);
2715 51290 : for (offnum = minoff;
2716 13741110 : offnum <= maxoff;
2717 13689820 : offnum = OffsetNumberNext(offnum))
2718 : {
2719 13689820 : ItemId itemId = PageGetItemId(page, offnum);
2720 :
2721 13689820 : if (ItemIdIsDead(itemId))
2722 255056 : deletable[ndeletable++] = offnum;
2723 : }
2724 :
2725 51290 : if (ndeletable > 0)
2726 : {
2727 7718 : _bt_simpledel_pass(rel, buffer, heapRel, deletable, ndeletable,
2728 : insertstate->itup, minoff, maxoff);
2729 7718 : insertstate->bounds_valid = false;
2730 :
2731 : /* Return when a page split has already been avoided */
2732 7718 : if (PageGetFreeSpace(page) >= insertstate->itemsz)
2733 23592 : return;
2734 :
2735 : /* Might as well assume duplicates (if checkingunique) */
2736 100 : uniquedup = true;
2737 : }
2738 :
2739 : /*
2740 : * We're done with simple deletion. Return early with callers that only
2741 : * call here so that simple deletion can be considered. This includes
2742 : * callers that explicitly ask for this and checkingunique callers that
2743 : * probably don't have any version churn duplicates on the page.
2744 : *
2745 : * Note: The page's BTP_HAS_GARBAGE hint flag may still be set when we
2746 : * return at this point (or when we go on the try either or both of our
2747 : * other strategies and they also fail). We do not bother expending a
2748 : * separate write to clear it, however. Caller will definitely clear it
2749 : * when it goes on to split the page (note also that the deduplication
2750 : * process will clear the flag in passing, just to keep things tidy).
2751 : */
2752 43672 : if (simpleonly || (checkingunique && !uniquedup))
2753 : {
2754 : Assert(!indexUnchanged);
2755 15528 : return;
2756 : }
2757 :
2758 : /* Assume bounds about to be invalidated (this is almost certain now) */
2759 28144 : insertstate->bounds_valid = false;
2760 :
2761 : /*
2762 : * Perform bottom-up index deletion pass when executor hint indicated that
2763 : * incoming item is logically unchanged, or for a unique index that is
2764 : * known to have physical duplicates for some other reason. (There is a
2765 : * large overlap between these two cases for a unique index. It's worth
2766 : * having both triggering conditions in order to apply the optimization in
2767 : * the event of successive related INSERT and DELETE statements.)
2768 : *
2769 : * We'll go on to do a deduplication pass when a bottom-up pass fails to
2770 : * delete an acceptable amount of free space (a significant fraction of
2771 : * the page, or space for the new item, whichever is greater).
2772 : *
2773 : * Note: Bottom-up index deletion uses the same equality/equivalence
2774 : * routines as deduplication internally. However, it does not merge
2775 : * together index tuples, so the same correctness considerations do not
2776 : * apply. We deliberately omit an index-is-allequalimage test here.
2777 : */
2778 32070 : if ((indexUnchanged || uniquedup) &&
2779 3926 : _bt_bottomupdel_pass(rel, buffer, heapRel, insertstate->itemsz))
2780 446 : return;
2781 :
2782 : /* Perform deduplication pass (when enabled and index-is-allequalimage) */
2783 27698 : if (BTGetDeduplicateItems(rel) && itup_key->allequalimage)
2784 27680 : _bt_dedup_pass(rel, buffer, insertstate->itup, insertstate->itemsz,
2785 27680 : (indexUnchanged || uniquedup));
2786 : }
2787 :
2788 : /*
2789 : * _bt_simpledel_pass - Simple index tuple deletion pass.
2790 : *
2791 : * We delete all LP_DEAD-set index tuples on a leaf page. The offset numbers
2792 : * of all such tuples are determined by caller (caller passes these to us as
2793 : * its 'deletable' argument).
2794 : *
2795 : * We might also delete extra index tuples that turn out to be safe to delete
2796 : * in passing (though they must be cheap to check in passing to begin with).
2797 : * There is no certainty that any extra tuples will be deleted, though. The
2798 : * high level goal of the approach we take is to get the most out of each call
2799 : * here (without noticeably increasing the per-call overhead compared to what
2800 : * we need to do just to be able to delete the page's LP_DEAD-marked index
2801 : * tuples).
2802 : *
2803 : * The number of extra index tuples that turn out to be deletable might
2804 : * greatly exceed the number of LP_DEAD-marked index tuples due to various
2805 : * locality related effects. For example, it's possible that the total number
2806 : * of table blocks (pointed to by all TIDs on the leaf page) is naturally
2807 : * quite low, in which case we might end up checking if it's possible to
2808 : * delete _most_ index tuples on the page (without the tableam needing to
2809 : * access additional table blocks). The tableam will sometimes stumble upon
2810 : * _many_ extra deletable index tuples in indexes where this pattern is
2811 : * common.
2812 : *
2813 : * See nbtree/README for further details on simple index tuple deletion.
2814 : */
2815 : static void
2816 7718 : _bt_simpledel_pass(Relation rel, Buffer buffer, Relation heapRel,
2817 : OffsetNumber *deletable, int ndeletable, IndexTuple newitem,
2818 : OffsetNumber minoff, OffsetNumber maxoff)
2819 : {
2820 7718 : Page page = BufferGetPage(buffer);
2821 : BlockNumber *deadblocks;
2822 : int ndeadblocks;
2823 : TM_IndexDeleteOp delstate;
2824 : OffsetNumber offnum;
2825 :
2826 : /* Get array of table blocks pointed to by LP_DEAD-set tuples */
2827 7718 : deadblocks = _bt_deadblocks(page, deletable, ndeletable, newitem,
2828 : &ndeadblocks);
2829 :
2830 : /* Initialize tableam state that describes index deletion operation */
2831 7718 : delstate.irel = rel;
2832 7718 : delstate.iblknum = BufferGetBlockNumber(buffer);
2833 7718 : delstate.bottomup = false;
2834 7718 : delstate.bottomupfreespace = 0;
2835 7718 : delstate.ndeltids = 0;
2836 7718 : delstate.deltids = palloc(MaxTIDsPerBTreePage * sizeof(TM_IndexDelete));
2837 7718 : delstate.status = palloc(MaxTIDsPerBTreePage * sizeof(TM_IndexStatus));
2838 :
2839 7718 : for (offnum = minoff;
2840 2187376 : offnum <= maxoff;
2841 2179658 : offnum = OffsetNumberNext(offnum))
2842 : {
2843 2179658 : ItemId itemid = PageGetItemId(page, offnum);
2844 2179658 : IndexTuple itup = (IndexTuple) PageGetItem(page, itemid);
2845 2179658 : TM_IndexDelete *odeltid = &delstate.deltids[delstate.ndeltids];
2846 2179658 : TM_IndexStatus *ostatus = &delstate.status[delstate.ndeltids];
2847 : BlockNumber tidblock;
2848 : void *match;
2849 :
2850 2179658 : if (!BTreeTupleIsPosting(itup))
2851 : {
2852 2081766 : tidblock = ItemPointerGetBlockNumber(&itup->t_tid);
2853 2081766 : match = bsearch(&tidblock, deadblocks, ndeadblocks,
2854 : sizeof(BlockNumber), _bt_blk_cmp);
2855 :
2856 2081766 : if (!match)
2857 : {
2858 : Assert(!ItemIdIsDead(itemid));
2859 1322702 : continue;
2860 : }
2861 :
2862 : /*
2863 : * TID's table block is among those pointed to by the TIDs from
2864 : * LP_DEAD-bit set tuples on page -- add TID to deltids
2865 : */
2866 759064 : odeltid->tid = itup->t_tid;
2867 759064 : odeltid->id = delstate.ndeltids;
2868 759064 : ostatus->idxoffnum = offnum;
2869 759064 : ostatus->knowndeletable = ItemIdIsDead(itemid);
2870 759064 : ostatus->promising = false; /* unused */
2871 759064 : ostatus->freespace = 0; /* unused */
2872 :
2873 759064 : delstate.ndeltids++;
2874 : }
2875 : else
2876 : {
2877 97892 : int nitem = BTreeTupleGetNPosting(itup);
2878 :
2879 470304 : for (int p = 0; p < nitem; p++)
2880 : {
2881 372412 : ItemPointer tid = BTreeTupleGetPostingN(itup, p);
2882 :
2883 372412 : tidblock = ItemPointerGetBlockNumber(tid);
2884 372412 : match = bsearch(&tidblock, deadblocks, ndeadblocks,
2885 : sizeof(BlockNumber), _bt_blk_cmp);
2886 :
2887 372412 : if (!match)
2888 : {
2889 : Assert(!ItemIdIsDead(itemid));
2890 328566 : continue;
2891 : }
2892 :
2893 : /*
2894 : * TID's table block is among those pointed to by the TIDs
2895 : * from LP_DEAD-bit set tuples on page -- add TID to deltids
2896 : */
2897 43846 : odeltid->tid = *tid;
2898 43846 : odeltid->id = delstate.ndeltids;
2899 43846 : ostatus->idxoffnum = offnum;
2900 43846 : ostatus->knowndeletable = ItemIdIsDead(itemid);
2901 43846 : ostatus->promising = false; /* unused */
2902 43846 : ostatus->freespace = 0; /* unused */
2903 :
2904 43846 : odeltid++;
2905 43846 : ostatus++;
2906 43846 : delstate.ndeltids++;
2907 : }
2908 : }
2909 : }
2910 :
2911 7718 : pfree(deadblocks);
2912 :
2913 : Assert(delstate.ndeltids >= ndeletable);
2914 :
2915 : /* Physically delete LP_DEAD tuples (plus any delete-safe extra TIDs) */
2916 7718 : _bt_delitems_delete_check(rel, buffer, heapRel, &delstate);
2917 :
2918 7718 : pfree(delstate.deltids);
2919 7718 : pfree(delstate.status);
2920 7718 : }
2921 :
2922 : /*
2923 : * _bt_deadblocks() -- Get LP_DEAD related table blocks.
2924 : *
2925 : * Builds sorted and unique-ified array of table block numbers from index
2926 : * tuple TIDs whose line pointers are marked LP_DEAD. Also adds the table
2927 : * block from incoming newitem just in case it isn't among the LP_DEAD-related
2928 : * table blocks.
2929 : *
2930 : * Always counting the newitem's table block as an LP_DEAD related block makes
2931 : * sense because the cost is consistently low; it is practically certain that
2932 : * the table block will not incur a buffer miss in tableam. On the other hand
2933 : * the benefit is often quite high. There is a decent chance that there will
2934 : * be some deletable items from this block, since in general most garbage
2935 : * tuples became garbage in the recent past (in many cases this won't be the
2936 : * first logical row that core code added to/modified in table block
2937 : * recently).
2938 : *
2939 : * Returns final array, and sets *nblocks to its final size for caller.
2940 : */
2941 : static BlockNumber *
2942 7718 : _bt_deadblocks(Page page, OffsetNumber *deletable, int ndeletable,
2943 : IndexTuple newitem, int *nblocks)
2944 : {
2945 : int spacentids,
2946 : ntids;
2947 : BlockNumber *tidblocks;
2948 :
2949 : /*
2950 : * Accumulate each TID's block in array whose initial size has space for
2951 : * one table block per LP_DEAD-set tuple (plus space for the newitem table
2952 : * block). Array will only need to grow when there are LP_DEAD-marked
2953 : * posting list tuples (which is not that common).
2954 : */
2955 7718 : spacentids = ndeletable + 1;
2956 7718 : ntids = 0;
2957 7718 : tidblocks = (BlockNumber *) palloc(sizeof(BlockNumber) * spacentids);
2958 :
2959 : /*
2960 : * First add the table block for the incoming newitem. This is the one
2961 : * case where simple deletion can visit a table block that doesn't have
2962 : * any known deletable items.
2963 : */
2964 : Assert(!BTreeTupleIsPosting(newitem) && !BTreeTupleIsPivot(newitem));
2965 7718 : tidblocks[ntids++] = ItemPointerGetBlockNumber(&newitem->t_tid);
2966 :
2967 262774 : for (int i = 0; i < ndeletable; i++)
2968 : {
2969 255056 : ItemId itemid = PageGetItemId(page, deletable[i]);
2970 255056 : IndexTuple itup = (IndexTuple) PageGetItem(page, itemid);
2971 :
2972 : Assert(ItemIdIsDead(itemid));
2973 :
2974 255056 : if (!BTreeTupleIsPosting(itup))
2975 : {
2976 246612 : if (ntids + 1 > spacentids)
2977 : {
2978 194 : spacentids *= 2;
2979 : tidblocks = (BlockNumber *)
2980 194 : repalloc(tidblocks, sizeof(BlockNumber) * spacentids);
2981 : }
2982 :
2983 246612 : tidblocks[ntids++] = ItemPointerGetBlockNumber(&itup->t_tid);
2984 : }
2985 : else
2986 : {
2987 8444 : int nposting = BTreeTupleGetNPosting(itup);
2988 :
2989 8444 : if (ntids + nposting > spacentids)
2990 : {
2991 160 : spacentids = Max(spacentids * 2, ntids + nposting);
2992 : tidblocks = (BlockNumber *)
2993 160 : repalloc(tidblocks, sizeof(BlockNumber) * spacentids);
2994 : }
2995 :
2996 27330 : for (int j = 0; j < nposting; j++)
2997 : {
2998 18886 : ItemPointer tid = BTreeTupleGetPostingN(itup, j);
2999 :
3000 18886 : tidblocks[ntids++] = ItemPointerGetBlockNumber(tid);
3001 : }
3002 : }
3003 : }
3004 :
3005 7718 : qsort(tidblocks, ntids, sizeof(BlockNumber), _bt_blk_cmp);
3006 7718 : *nblocks = qunique(tidblocks, ntids, sizeof(BlockNumber), _bt_blk_cmp);
3007 :
3008 7718 : return tidblocks;
3009 : }
3010 :
3011 : /*
3012 : * _bt_blk_cmp() -- qsort comparison function for _bt_simpledel_pass
3013 : */
3014 : static inline int
3015 5461526 : _bt_blk_cmp(const void *arg1, const void *arg2)
3016 : {
3017 5461526 : BlockNumber b1 = *((BlockNumber *) arg1);
3018 5461526 : BlockNumber b2 = *((BlockNumber *) arg2);
3019 :
3020 5461526 : return pg_cmp_u32(b1, b2);
3021 : }
|
