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, IndexTuple 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 7222816 : _bt_doinsert(Relation rel, IndexTuple itup,
104 : IndexUniqueCheck checkUnique, bool indexUnchanged,
105 : Relation heapRel)
106 : {
107 7222816 : bool is_unique = false;
108 : BTInsertStateData insertstate;
109 : BTScanInsert itup_key;
110 : BTStack stack;
111 7222816 : bool checkingunique = (checkUnique != UNIQUE_CHECK_NO);
112 :
113 : /* we need an insertion scan key to do our search, so build one */
114 7222816 : itup_key = _bt_mkscankey(rel, itup);
115 :
116 7222816 : if (checkingunique)
117 : {
118 5177276 : if (!itup_key->anynullkeys)
119 : {
120 : /* No (heapkeyspace) scantid until uniqueness established */
121 5157102 : 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 7222816 : insertstate.itup = itup;
155 7222816 : insertstate.itemsz = MAXALIGN(IndexTupleSize(itup));
156 7222816 : insertstate.itup_key = itup_key;
157 7222816 : insertstate.bounds_valid = false;
158 7222816 : insertstate.buf = InvalidBuffer;
159 7222816 : insertstate.postingoff = 0;
160 :
161 7222840 : 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 7222840 : 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 7222840 : if (checkingunique)
207 : {
208 : TransactionId xwait;
209 : uint32 speculativeToken;
210 :
211 5157126 : xwait = _bt_check_unique(rel, &insertstate, heapRel, checkUnique,
212 : &is_unique, &speculativeToken);
213 :
214 5156618 : 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 5156594 : if (itup_key->heapkeyspace)
238 5156594 : itup_key->scantid = &itup->t_tid;
239 : }
240 :
241 7222308 : 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 7222254 : 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 7222248 : newitemoff = _bt_findinsertloc(rel, &insertstate, checkingunique,
260 : indexUnchanged, stack, heapRel);
261 7222248 : _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 7222302 : if (stack)
273 6267556 : _bt_freestack(stack);
274 7222302 : pfree(itup_key);
275 :
276 7222302 : 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 7222840 : _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 7222840 : if (RelationGetTargetBlock(rel) != InvalidBlockNumber)
325 : {
326 : /* Simulate a _bt_getbuf() call with conditional locking */
327 65972 : insertstate->buf = ReadBuffer(rel, RelationGetTargetBlock(rel));
328 65972 : if (_bt_conditionallockbuf(rel, insertstate->buf))
329 : {
330 : Page page;
331 : BTPageOpaque opaque;
332 :
333 65076 : _bt_checkpage(rel, insertstate->buf);
334 65076 : page = BufferGetPage(insertstate->buf);
335 65076 : 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 65076 : if (P_RIGHTMOST(opaque) &&
346 65024 : P_ISLEAF(opaque) &&
347 65024 : !P_IGNORE(opaque) &&
348 129658 : PageGetFreeSpace(page) > insertstate->itemsz &&
349 129268 : PageGetMaxOffsetNumber(page) >= P_HIKEY &&
350 64634 : _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 64592 : return NULL;
365 : }
366 :
367 : /* Page unsuitable for caller, drop lock and pin */
368 484 : _bt_relbuf(rel, insertstate->buf);
369 : }
370 : else
371 : {
372 : /* Lock unavailable, drop pin */
373 896 : ReleaseBuffer(insertstate->buf);
374 : }
375 :
376 : /* Forget block, since cache doesn't appear to be useful */
377 1380 : RelationSetTargetBlock(rel, InvalidBlockNumber);
378 : }
379 :
380 : /* Cannot use optimization -- descend tree, return proper descent stack */
381 7158248 : 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 5157126 : _bt_check_unique(Relation rel, BTInsertState insertstate, Relation heapRel,
410 : IndexUniqueCheck checkUnique, bool *is_unique,
411 : uint32 *speculativeToken)
412 : {
413 5157126 : IndexTuple itup = insertstate->itup;
414 5157126 : IndexTuple curitup = NULL;
415 5157126 : ItemId curitemid = NULL;
416 5157126 : BTScanInsert itup_key = insertstate->itup_key;
417 : SnapshotData SnapshotDirty;
418 : OffsetNumber offset;
419 : OffsetNumber maxoff;
420 : Page page;
421 : BTPageOpaque opaque;
422 5157126 : Buffer nbuf = InvalidBuffer;
423 5157126 : bool found = false;
424 5157126 : bool inposting = false;
425 5157126 : bool prevalldead = true;
426 5157126 : int curposti = 0;
427 :
428 : /* Assume unique until we find a duplicate */
429 5157126 : *is_unique = true;
430 :
431 5157126 : InitDirtySnapshot(SnapshotDirty);
432 :
433 5157126 : page = BufferGetPage(insertstate->buf);
434 5157126 : opaque = BTPageGetOpaque(page);
435 5157126 : 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 5157126 : 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 17127372 : 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 14253986 : 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 2053454 : 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 12200532 : if (!inposting)
505 7678642 : curitemid = PageGetItemId(page, offset);
506 12200532 : if (inposting || !ItemIdIsDead(curitemid))
507 : {
508 : ItemPointerData htid;
509 11664848 : bool all_dead = false;
510 :
511 11664848 : if (!inposting)
512 : {
513 : /* Plain tuple, or first TID in posting list tuple */
514 7142958 : if (_bt_compare(rel, itup_key, page, offset) != 0)
515 204156 : break; /* we're past all the equal tuples */
516 :
517 : /* Advanced curitup */
518 6938802 : curitup = (IndexTuple) PageGetItem(page, curitemid);
519 : Assert(!BTreeTupleIsPivot(curitup));
520 : }
521 :
522 : /* okay, we gotta fetch the heap tuple using htid ... */
523 11460692 : if (!BTreeTupleIsPosting(curitup))
524 : {
525 : /* ... htid is from simple non-pivot tuple */
526 : Assert(!inposting);
527 6891622 : htid = curitup->t_tid;
528 : }
529 4569070 : else if (!inposting)
530 : {
531 : /* ... htid is first TID in new posting list */
532 47180 : inposting = true;
533 47180 : prevalldead = true;
534 47180 : curposti = 0;
535 47180 : htid = *BTreeTupleGetPostingN(curitup, 0);
536 : }
537 : else
538 : {
539 : /* ... htid is second or subsequent TID in posting list */
540 : Assert(curposti > 0);
541 4521890 : 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 11460904 : if (checkUnique == UNIQUE_CHECK_EXISTING &&
550 212 : 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 11460638 : 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 : */
575 732 : if (checkUnique == UNIQUE_CHECK_PARTIAL)
576 : {
577 200 : if (nbuf != InvalidBuffer)
578 0 : _bt_relbuf(rel, nbuf);
579 200 : *is_unique = false;
580 224 : 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 1064 : xwait = (TransactionIdIsValid(SnapshotDirty.xmin)) ?
588 532 : SnapshotDirty.xmin : SnapshotDirty.xmax;
589 :
590 532 : 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 : */
618 508 : htid = itup->t_tid;
619 508 : 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 : */
640 508 : 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 : */
650 500 : if (nbuf != InvalidBuffer)
651 0 : _bt_relbuf(rel, nbuf);
652 500 : _bt_relbuf(rel, insertstate->buf);
653 500 : insertstate->buf = InvalidBuffer;
654 500 : insertstate->bounds_valid = false;
655 :
656 : {
657 : Datum values[INDEX_MAX_KEYS];
658 : bool isnull[INDEX_MAX_KEYS];
659 : char *key_desc;
660 :
661 500 : index_deform_tuple(itup, RelationGetDescr(rel),
662 : values, isnull);
663 :
664 500 : key_desc = BuildIndexValueDescription(rel, values,
665 : isnull);
666 :
667 500 : 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 11459906 : else if (all_dead && (!inposting ||
678 38114 : (prevalldead &&
679 38114 : 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 105360 : ItemIdMarkDead(curitemid);
687 105360 : 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 105360 : if (nbuf != InvalidBuffer)
694 4 : MarkBufferDirtyHint(nbuf, true);
695 : else
696 105356 : 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 11459960 : if (!all_dead && inposting)
704 4530856 : prevalldead = false;
705 : }
706 : }
707 :
708 14869030 : if (inposting && curposti < BTreeTupleGetNPosting(curitup) - 1)
709 : {
710 : /* Advance to next TID in same posting list */
711 4521890 : curposti++;
712 4521890 : continue;
713 : }
714 10347140 : else if (offset < maxoff)
715 : {
716 : /* Advance to next tuple */
717 7438312 : curposti = 0;
718 7438312 : inposting = false;
719 7438312 : offset = OffsetNumberNext(offset);
720 : }
721 : else
722 : {
723 : int highkeycmp;
724 :
725 : /* If scankey == hikey we gotta check the next page too */
726 2908828 : if (P_RIGHTMOST(opaque))
727 2760378 : break;
728 148450 : highkeycmp = _bt_compare(rel, itup_key, page, P_HIKEY);
729 : Assert(highkeycmp <= 0);
730 148450 : if (highkeycmp != 0)
731 138406 : 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 5156394 : 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 5156394 : if (nbuf != InvalidBuffer)
770 5760 : _bt_relbuf(rel, nbuf);
771 :
772 5156394 : 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 7222248 : _bt_findinsertloc(Relation rel,
817 : BTInsertState insertstate,
818 : bool checkingunique,
819 : bool indexUnchanged,
820 : BTStack stack,
821 : Relation heapRel)
822 : {
823 7222248 : BTScanInsert itup_key = insertstate->itup_key;
824 7222248 : Page page = BufferGetPage(insertstate->buf);
825 : BTPageOpaque opaque;
826 : OffsetNumber newitemoff;
827 :
828 7222248 : opaque = BTPageGetOpaque(page);
829 :
830 : /* Check 1/3 of a page restriction */
831 7222248 : 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 7222248 : if (itup_key->heapkeyspace)
842 : {
843 : /* Keep track of whether checkingunique duplicate seen */
844 7222248 : 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 7222248 : if (checkingunique)
861 : {
862 5156534 : if (insertstate->low < insertstate->stricthigh)
863 : {
864 : /* Encountered a duplicate in _bt_check_unique() */
865 : Assert(insertstate->bounds_valid);
866 443104 : 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 5166578 : if (insertstate->bounds_valid &&
882 10291536 : insertstate->low <= insertstate->stricthigh &&
883 5145768 : insertstate->stricthigh <= PageGetMaxOffsetNumber(page))
884 2232472 : break;
885 :
886 : /* Test '<=', not '!=', since scantid is set now */
887 3096906 : if (P_RIGHTMOST(opaque) ||
888 162800 : _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 7222248 : if (PageGetFreeSpace(page) < insertstate->itemsz)
905 50748 : _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 7222248 : newitemoff = _bt_binsrch_insert(rel, insertstate);
990 :
991 7222248 : 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 6 : _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 6 : insertstate->postingoff = 0;
1008 6 : newitemoff = _bt_binsrch_insert(rel, insertstate);
1009 : Assert(insertstate->postingoff == 0);
1010 : }
1011 :
1012 7222248 : 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 7243328 : _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 7243328 : IndexTuple oposting = NULL;
1125 7243328 : IndexTuple origitup = NULL;
1126 7243328 : IndexTuple nposting = NULL;
1127 :
1128 7243328 : page = BufferGetPage(buf);
1129 7243328 : opaque = BTPageGetOpaque(page);
1130 7243328 : isleaf = P_ISLEAF(opaque);
1131 7243328 : isroot = P_ISROOT(opaque);
1132 7243328 : isrightmost = P_RIGHTMOST(opaque);
1133 7243328 : 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 7243328 : if (postingoff != 0)
1161 : {
1162 22072 : 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 22072 : 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 22072 : 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 22072 : origitup = itup;
1196 22072 : itup = CopyIndexTuple(origitup);
1197 22072 : 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 22072 : 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 7243328 : 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 22446 : rbuf = _bt_split(rel, heaprel, itup_key, buf, cbuf, newitemoff, itemsz,
1219 : itup, origitup, nposting, postingoff);
1220 22446 : 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 22446 : _bt_insert_parent(rel, heaprel, buf, rbuf, stack, isroot, isonly);
1243 : }
1244 : else
1245 : {
1246 7220882 : Buffer metabuf = InvalidBuffer;
1247 7220882 : Page metapg = NULL;
1248 7220882 : 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 7220882 : if (unlikely(split_only_page))
1259 : {
1260 : Assert(!isleaf);
1261 : Assert(BufferIsValid(cbuf));
1262 :
1263 24 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_WRITE);
1264 24 : metapg = BufferGetPage(metabuf);
1265 24 : metad = BTPageGetMeta(metapg);
1266 :
1267 24 : 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 7220882 : START_CRIT_SECTION();
1277 :
1278 7220882 : if (postingoff != 0)
1279 22004 : memcpy(oposting, nposting, MAXALIGN(IndexTupleSize(nposting)));
1280 :
1281 7220882 : if (PageAddItem(page, (Item) itup, itemsz, newitemoff, false,
1282 : false) == InvalidOffsetNumber)
1283 0 : elog(PANIC, "failed to add new item to block %u in index \"%s\"",
1284 : BufferGetBlockNumber(buf), RelationGetRelationName(rel));
1285 :
1286 7220882 : MarkBufferDirty(buf);
1287 :
1288 7220882 : if (BufferIsValid(metabuf))
1289 : {
1290 : /* upgrade meta-page if needed */
1291 24 : if (metad->btm_version < BTREE_NOVAC_VERSION)
1292 0 : _bt_upgrademetapage(metapg);
1293 24 : metad->btm_fastroot = BufferGetBlockNumber(buf);
1294 24 : metad->btm_fastlevel = opaque->btpo_level;
1295 24 : MarkBufferDirty(metabuf);
1296 : }
1297 :
1298 : /*
1299 : * Clear INCOMPLETE_SPLIT flag on child if inserting the new item
1300 : * finishes a split
1301 : */
1302 7220882 : if (!isleaf)
1303 : {
1304 20822 : Page cpage = BufferGetPage(cbuf);
1305 20822 : BTPageOpaque cpageop = BTPageGetOpaque(cpage);
1306 :
1307 : Assert(P_INCOMPLETE_SPLIT(cpageop));
1308 20822 : cpageop->btpo_flags &= ~BTP_INCOMPLETE_SPLIT;
1309 20822 : MarkBufferDirty(cbuf);
1310 : }
1311 :
1312 : /* XLOG stuff */
1313 7220882 : if (RelationNeedsWAL(rel))
1314 : {
1315 : xl_btree_insert xlrec;
1316 : xl_btree_metadata xlmeta;
1317 : uint8 xlinfo;
1318 : XLogRecPtr recptr;
1319 : uint16 upostingoff;
1320 :
1321 6714856 : xlrec.offnum = newitemoff;
1322 :
1323 6714856 : XLogBeginInsert();
1324 6714856 : XLogRegisterData(&xlrec, SizeOfBtreeInsert);
1325 :
1326 6714856 : if (isleaf && postingoff == 0)
1327 : {
1328 : /* Simple leaf insert */
1329 6673284 : xlinfo = XLOG_BTREE_INSERT_LEAF;
1330 : }
1331 41572 : else if (postingoff != 0)
1332 : {
1333 : /*
1334 : * Leaf insert with posting list split. Must include
1335 : * postingoff field before newitem/orignewitem.
1336 : */
1337 : Assert(isleaf);
1338 22004 : xlinfo = XLOG_BTREE_INSERT_POST;
1339 : }
1340 : else
1341 : {
1342 : /* Internal page insert, which finishes a split on cbuf */
1343 19568 : xlinfo = XLOG_BTREE_INSERT_UPPER;
1344 19568 : XLogRegisterBuffer(1, cbuf, REGBUF_STANDARD);
1345 :
1346 19568 : if (BufferIsValid(metabuf))
1347 : {
1348 : /* Actually, it's an internal page insert + meta update */
1349 24 : xlinfo = XLOG_BTREE_INSERT_META;
1350 :
1351 : Assert(metad->btm_version >= BTREE_NOVAC_VERSION);
1352 24 : xlmeta.version = metad->btm_version;
1353 24 : xlmeta.root = metad->btm_root;
1354 24 : xlmeta.level = metad->btm_level;
1355 24 : xlmeta.fastroot = metad->btm_fastroot;
1356 24 : xlmeta.fastlevel = metad->btm_fastlevel;
1357 24 : xlmeta.last_cleanup_num_delpages = metad->btm_last_cleanup_num_delpages;
1358 24 : xlmeta.allequalimage = metad->btm_allequalimage;
1359 :
1360 24 : XLogRegisterBuffer(2, metabuf,
1361 : REGBUF_WILL_INIT | REGBUF_STANDARD);
1362 24 : XLogRegisterBufData(2, &xlmeta,
1363 : sizeof(xl_btree_metadata));
1364 : }
1365 : }
1366 :
1367 6714856 : XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
1368 6714856 : if (postingoff == 0)
1369 : {
1370 : /* Just log itup from caller */
1371 6692852 : XLogRegisterBufData(0, itup, IndexTupleSize(itup));
1372 : }
1373 : else
1374 : {
1375 : /*
1376 : * Insert with posting list split (XLOG_BTREE_INSERT_POST
1377 : * record) case.
1378 : *
1379 : * Log postingoff. Also log origitup, not itup. REDO routine
1380 : * must reconstruct final itup (as well as nposting) using
1381 : * _bt_swap_posting().
1382 : */
1383 22004 : upostingoff = postingoff;
1384 :
1385 22004 : XLogRegisterBufData(0, &upostingoff, sizeof(uint16));
1386 22004 : XLogRegisterBufData(0, origitup,
1387 22004 : IndexTupleSize(origitup));
1388 : }
1389 :
1390 6714856 : recptr = XLogInsert(RM_BTREE_ID, xlinfo);
1391 :
1392 6714856 : if (BufferIsValid(metabuf))
1393 24 : PageSetLSN(metapg, recptr);
1394 6714856 : if (!isleaf)
1395 19568 : PageSetLSN(BufferGetPage(cbuf), recptr);
1396 :
1397 6714856 : PageSetLSN(page, recptr);
1398 : }
1399 :
1400 7220882 : END_CRIT_SECTION();
1401 :
1402 : /* Release subsidiary buffers */
1403 7220882 : if (BufferIsValid(metabuf))
1404 24 : _bt_relbuf(rel, metabuf);
1405 7220882 : if (!isleaf)
1406 20822 : _bt_relbuf(rel, cbuf);
1407 :
1408 : /*
1409 : * Cache the block number if this is the rightmost leaf page. Cache
1410 : * may be used by a future inserter within _bt_search_insert().
1411 : */
1412 7220882 : blockcache = InvalidBlockNumber;
1413 7220882 : if (isrightmost && isleaf && !isroot)
1414 3994202 : blockcache = BufferGetBlockNumber(buf);
1415 :
1416 : /* Release buffer for insertion target block */
1417 7220882 : _bt_relbuf(rel, buf);
1418 :
1419 : /*
1420 : * If we decided to cache the insertion target block before releasing
1421 : * its buffer lock, then cache it now. Check the height of the tree
1422 : * first, though. We don't go for the optimization with small
1423 : * indexes. Defer final check to this point to ensure that we don't
1424 : * call _bt_getrootheight while holding a buffer lock.
1425 : */
1426 11215084 : if (BlockNumberIsValid(blockcache) &&
1427 3994202 : _bt_getrootheight(rel) >= BTREE_FASTPATH_MIN_LEVEL)
1428 65996 : RelationSetTargetBlock(rel, blockcache);
1429 : }
1430 :
1431 : /* be tidy */
1432 7243328 : if (postingoff != 0)
1433 : {
1434 : /* itup is actually a modified copy of caller's original */
1435 22072 : pfree(nposting);
1436 22072 : pfree(itup);
1437 : }
1438 7243328 : }
1439 :
1440 : /*
1441 : * _bt_split() -- split a page in the btree.
1442 : *
1443 : * On entry, buf is the page to split, and is pinned and write-locked.
1444 : * newitemoff etc. tell us about the new item that must be inserted
1445 : * along with the data from the original page.
1446 : *
1447 : * itup_key is used for suffix truncation on leaf pages (internal
1448 : * page callers pass NULL). When splitting a non-leaf page, 'cbuf'
1449 : * is the left-sibling of the page we're inserting the downlink for.
1450 : * This function will clear the INCOMPLETE_SPLIT flag on it, and
1451 : * release the buffer.
1452 : *
1453 : * orignewitem, nposting, and postingoff are needed when an insert of
1454 : * orignewitem results in both a posting list split and a page split.
1455 : * These extra posting list split details are used here in the same
1456 : * way as they are used in the more common case where a posting list
1457 : * split does not coincide with a page split. We need to deal with
1458 : * posting list splits directly in order to ensure that everything
1459 : * that follows from the insert of orignewitem is handled as a single
1460 : * atomic operation (though caller's insert of a new pivot/downlink
1461 : * into parent page will still be a separate operation). See
1462 : * nbtree/README for details on the design of posting list splits.
1463 : *
1464 : * Returns the new right sibling of buf, pinned and write-locked.
1465 : * The pin and lock on buf are maintained.
1466 : */
1467 : static Buffer
1468 22446 : _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
1469 : Buffer cbuf, OffsetNumber newitemoff, Size newitemsz, IndexTuple newitem,
1470 : IndexTuple orignewitem, IndexTuple nposting, uint16 postingoff)
1471 : {
1472 : Buffer rbuf;
1473 : Page origpage;
1474 : Page leftpage,
1475 : rightpage;
1476 : PGAlignedBlock leftpage_buf,
1477 : rightpage_buf;
1478 : BlockNumber origpagenumber,
1479 : rightpagenumber;
1480 : BTPageOpaque ropaque,
1481 : lopaque,
1482 : oopaque;
1483 22446 : Buffer sbuf = InvalidBuffer;
1484 22446 : Page spage = NULL;
1485 22446 : BTPageOpaque sopaque = NULL;
1486 : Size itemsz;
1487 : ItemId itemid;
1488 : IndexTuple firstright,
1489 : lefthighkey;
1490 : OffsetNumber firstrightoff;
1491 : OffsetNumber afterleftoff,
1492 : afterrightoff,
1493 : minusinfoff;
1494 : OffsetNumber origpagepostingoff;
1495 : OffsetNumber maxoff;
1496 : OffsetNumber i;
1497 : bool newitemonleft,
1498 : isleaf,
1499 : isrightmost;
1500 :
1501 : /*
1502 : * origpage is the original page to be split. leftpage is a temporary
1503 : * buffer that receives the left-sibling data, which will be copied back
1504 : * into origpage on success. rightpage is the new page that will receive
1505 : * the right-sibling data.
1506 : *
1507 : * leftpage is allocated after choosing a split point. rightpage's new
1508 : * buffer isn't acquired until after leftpage is initialized and has new
1509 : * high key, the last point where splitting the page may fail (barring
1510 : * corruption). Failing before acquiring new buffer won't have lasting
1511 : * consequences, since origpage won't have been modified and leftpage is
1512 : * only workspace.
1513 : */
1514 22446 : origpage = BufferGetPage(buf);
1515 22446 : oopaque = BTPageGetOpaque(origpage);
1516 22446 : isleaf = P_ISLEAF(oopaque);
1517 22446 : isrightmost = P_RIGHTMOST(oopaque);
1518 22446 : maxoff = PageGetMaxOffsetNumber(origpage);
1519 22446 : origpagenumber = BufferGetBlockNumber(buf);
1520 :
1521 : /*
1522 : * Choose a point to split origpage at.
1523 : *
1524 : * A split point can be thought of as a point _between_ two existing data
1525 : * items on origpage (the lastleft and firstright tuples), provided you
1526 : * pretend that the new item that didn't fit is already on origpage.
1527 : *
1528 : * Since origpage does not actually contain newitem, the representation of
1529 : * split points needs to work with two boundary cases: splits where
1530 : * newitem is lastleft, and splits where newitem is firstright.
1531 : * newitemonleft resolves the ambiguity that would otherwise exist when
1532 : * newitemoff == firstrightoff. In all other cases it's clear which side
1533 : * of the split every tuple goes on from context. newitemonleft is
1534 : * usually (but not always) redundant information.
1535 : *
1536 : * firstrightoff is supposed to be an origpage offset number, but it's
1537 : * possible that its value will be maxoff+1, which is "past the end" of
1538 : * origpage. This happens in the rare case where newitem goes after all
1539 : * existing items (i.e. newitemoff is maxoff+1) and we end up splitting
1540 : * origpage at the point that leaves newitem alone on new right page. Any
1541 : * "!newitemonleft && newitemoff == firstrightoff" split point makes
1542 : * newitem the firstright tuple, though, so this case isn't a special
1543 : * case.
1544 : */
1545 22446 : firstrightoff = _bt_findsplitloc(rel, origpage, newitemoff, newitemsz,
1546 : newitem, &newitemonleft);
1547 :
1548 : /* Use temporary buffer for leftpage */
1549 22446 : leftpage = leftpage_buf.data;
1550 22446 : _bt_pageinit(leftpage, BufferGetPageSize(buf));
1551 22446 : lopaque = BTPageGetOpaque(leftpage);
1552 :
1553 : /*
1554 : * leftpage won't be the root when we're done. Also, clear the SPLIT_END
1555 : * and HAS_GARBAGE flags.
1556 : */
1557 22446 : lopaque->btpo_flags = oopaque->btpo_flags;
1558 22446 : lopaque->btpo_flags &= ~(BTP_ROOT | BTP_SPLIT_END | BTP_HAS_GARBAGE);
1559 : /* set flag in leftpage indicating that rightpage has no downlink yet */
1560 22446 : lopaque->btpo_flags |= BTP_INCOMPLETE_SPLIT;
1561 22446 : lopaque->btpo_prev = oopaque->btpo_prev;
1562 : /* handle btpo_next after rightpage buffer acquired */
1563 22446 : lopaque->btpo_level = oopaque->btpo_level;
1564 : /* handle btpo_cycleid after rightpage buffer acquired */
1565 :
1566 : /*
1567 : * Copy the original page's LSN into leftpage, which will become the
1568 : * updated version of the page. We need this because XLogInsert will
1569 : * examine the LSN and possibly dump it in a page image.
1570 : */
1571 22446 : PageSetLSN(leftpage, PageGetLSN(origpage));
1572 :
1573 : /*
1574 : * Determine page offset number of existing overlapped-with-orignewitem
1575 : * posting list when it is necessary to perform a posting list split in
1576 : * passing. Note that newitem was already changed by caller (newitem no
1577 : * longer has the orignewitem TID).
1578 : *
1579 : * This page offset number (origpagepostingoff) will be used to pretend
1580 : * that the posting split has already taken place, even though the
1581 : * required modifications to origpage won't occur until we reach the
1582 : * critical section. The lastleft and firstright tuples of our page split
1583 : * point should, in effect, come from an imaginary version of origpage
1584 : * that has the nposting tuple instead of the original posting list tuple.
1585 : *
1586 : * Note: _bt_findsplitloc() should have compensated for coinciding posting
1587 : * list splits in just the same way, at least in theory. It doesn't
1588 : * bother with that, though. In practice it won't affect its choice of
1589 : * split point.
1590 : */
1591 22446 : origpagepostingoff = InvalidOffsetNumber;
1592 22446 : if (postingoff != 0)
1593 : {
1594 : Assert(isleaf);
1595 : Assert(ItemPointerCompare(&orignewitem->t_tid,
1596 : &newitem->t_tid) < 0);
1597 : Assert(BTreeTupleIsPosting(nposting));
1598 68 : origpagepostingoff = OffsetNumberPrev(newitemoff);
1599 : }
1600 :
1601 : /*
1602 : * The high key for the new left page is a possibly-truncated copy of
1603 : * firstright on the leaf level (it's "firstright itself" on internal
1604 : * pages; see !isleaf comments below). This may seem to be contrary to
1605 : * Lehman & Yao's approach of using a copy of lastleft as the new high key
1606 : * when splitting on the leaf level. It isn't, though.
1607 : *
1608 : * Suffix truncation will leave the left page's high key fully equal to
1609 : * lastleft when lastleft and firstright are equal prior to heap TID (that
1610 : * is, the tiebreaker TID value comes from lastleft). It isn't actually
1611 : * necessary for a new leaf high key to be a copy of lastleft for the L&Y
1612 : * "subtree" invariant to hold. It's sufficient to make sure that the new
1613 : * leaf high key is strictly less than firstright, and greater than or
1614 : * equal to (not necessarily equal to) lastleft. In other words, when
1615 : * suffix truncation isn't possible during a leaf page split, we take
1616 : * L&Y's exact approach to generating a new high key for the left page.
1617 : * (Actually, that is slightly inaccurate. We don't just use a copy of
1618 : * lastleft. A tuple with all the keys from firstright but the max heap
1619 : * TID from lastleft is used, to avoid introducing a special case.)
1620 : */
1621 22446 : if (!newitemonleft && newitemoff == firstrightoff)
1622 : {
1623 : /* incoming tuple becomes firstright */
1624 42 : itemsz = newitemsz;
1625 42 : firstright = newitem;
1626 : }
1627 : else
1628 : {
1629 : /* existing item at firstrightoff becomes firstright */
1630 22404 : itemid = PageGetItemId(origpage, firstrightoff);
1631 22404 : itemsz = ItemIdGetLength(itemid);
1632 22404 : firstright = (IndexTuple) PageGetItem(origpage, itemid);
1633 22404 : if (firstrightoff == origpagepostingoff)
1634 0 : firstright = nposting;
1635 : }
1636 :
1637 22446 : if (isleaf)
1638 : {
1639 : IndexTuple lastleft;
1640 :
1641 : /* Attempt suffix truncation for leaf page splits */
1642 22188 : if (newitemonleft && newitemoff == firstrightoff)
1643 : {
1644 : /* incoming tuple becomes lastleft */
1645 340 : lastleft = newitem;
1646 : }
1647 : else
1648 : {
1649 : OffsetNumber lastleftoff;
1650 :
1651 : /* existing item before firstrightoff becomes lastleft */
1652 21848 : lastleftoff = OffsetNumberPrev(firstrightoff);
1653 : Assert(lastleftoff >= P_FIRSTDATAKEY(oopaque));
1654 21848 : itemid = PageGetItemId(origpage, lastleftoff);
1655 21848 : lastleft = (IndexTuple) PageGetItem(origpage, itemid);
1656 21848 : if (lastleftoff == origpagepostingoff)
1657 6 : lastleft = nposting;
1658 : }
1659 :
1660 22188 : lefthighkey = _bt_truncate(rel, lastleft, firstright, itup_key);
1661 22188 : itemsz = IndexTupleSize(lefthighkey);
1662 : }
1663 : else
1664 : {
1665 : /*
1666 : * Don't perform suffix truncation on a copy of firstright to make
1667 : * left page high key for internal page splits. Must use firstright
1668 : * as new high key directly.
1669 : *
1670 : * Each distinct separator key value originates as a leaf level high
1671 : * key; all other separator keys/pivot tuples are copied from one
1672 : * level down. A separator key in a grandparent page must be
1673 : * identical to high key in rightmost parent page of the subtree to
1674 : * its left, which must itself be identical to high key in rightmost
1675 : * child page of that same subtree (this even applies to separator
1676 : * from grandparent's high key). There must always be an unbroken
1677 : * "seam" of identical separator keys that guide index scans at every
1678 : * level, starting from the grandparent. That's why suffix truncation
1679 : * is unsafe here.
1680 : *
1681 : * Internal page splits will truncate firstright into a "negative
1682 : * infinity" data item when it gets inserted on the new right page
1683 : * below, though. This happens during the call to _bt_pgaddtup() for
1684 : * the new first data item for right page. Do not confuse this
1685 : * mechanism with suffix truncation. It is just a convenient way of
1686 : * implementing page splits that split the internal page "inside"
1687 : * firstright. The lefthighkey separator key cannot appear a second
1688 : * time in the right page (only firstright's downlink goes in right
1689 : * page).
1690 : */
1691 258 : lefthighkey = firstright;
1692 : }
1693 :
1694 : /*
1695 : * Add new high key to leftpage
1696 : */
1697 22446 : afterleftoff = P_HIKEY;
1698 :
1699 : Assert(BTreeTupleGetNAtts(lefthighkey, rel) > 0);
1700 : Assert(BTreeTupleGetNAtts(lefthighkey, rel) <=
1701 : IndexRelationGetNumberOfKeyAttributes(rel));
1702 : Assert(itemsz == MAXALIGN(IndexTupleSize(lefthighkey)));
1703 22446 : if (PageAddItem(leftpage, (Item) lefthighkey, itemsz, afterleftoff, false,
1704 : false) == InvalidOffsetNumber)
1705 0 : elog(ERROR, "failed to add high key to the left sibling"
1706 : " while splitting block %u of index \"%s\"",
1707 : origpagenumber, RelationGetRelationName(rel));
1708 22446 : afterleftoff = OffsetNumberNext(afterleftoff);
1709 :
1710 : /*
1711 : * Acquire a new right page to split into, now that left page has a new
1712 : * high key.
1713 : *
1714 : * To not confuse future VACUUM operations, we zero the right page and
1715 : * work on an in-memory copy of it before writing WAL, then copy its
1716 : * contents back to the actual page once we start the critical section
1717 : * work. This simplifies the split work, so as there is no need to zero
1718 : * the right page before throwing an error.
1719 : */
1720 22446 : rbuf = _bt_allocbuf(rel, heaprel);
1721 22446 : rightpage = rightpage_buf.data;
1722 :
1723 : /*
1724 : * Copy the contents of the right page into its temporary location, and
1725 : * zero the original space.
1726 : */
1727 22446 : memcpy(rightpage, BufferGetPage(rbuf), BLCKSZ);
1728 22446 : memset(BufferGetPage(rbuf), 0, BLCKSZ);
1729 22446 : rightpagenumber = BufferGetBlockNumber(rbuf);
1730 : /* rightpage was initialized by _bt_allocbuf */
1731 22446 : ropaque = BTPageGetOpaque(rightpage);
1732 :
1733 : /*
1734 : * Finish off remaining leftpage special area fields. They cannot be set
1735 : * before both origpage (leftpage) and rightpage buffers are acquired and
1736 : * locked.
1737 : *
1738 : * btpo_cycleid is only used with leaf pages, though we set it here in all
1739 : * cases just to be consistent.
1740 : */
1741 22446 : lopaque->btpo_next = rightpagenumber;
1742 22446 : lopaque->btpo_cycleid = _bt_vacuum_cycleid(rel);
1743 :
1744 : /*
1745 : * rightpage won't be the root when we're done. Also, clear the SPLIT_END
1746 : * and HAS_GARBAGE flags.
1747 : */
1748 22446 : ropaque->btpo_flags = oopaque->btpo_flags;
1749 22446 : ropaque->btpo_flags &= ~(BTP_ROOT | BTP_SPLIT_END | BTP_HAS_GARBAGE);
1750 22446 : ropaque->btpo_prev = origpagenumber;
1751 22446 : ropaque->btpo_next = oopaque->btpo_next;
1752 22446 : ropaque->btpo_level = oopaque->btpo_level;
1753 22446 : ropaque->btpo_cycleid = lopaque->btpo_cycleid;
1754 :
1755 : /*
1756 : * Add new high key to rightpage where necessary.
1757 : *
1758 : * If the page we're splitting is not the rightmost page at its level in
1759 : * the tree, then the first entry on the page is the high key from
1760 : * origpage.
1761 : */
1762 22446 : afterrightoff = P_HIKEY;
1763 :
1764 22446 : if (!isrightmost)
1765 : {
1766 : IndexTuple righthighkey;
1767 :
1768 9764 : itemid = PageGetItemId(origpage, P_HIKEY);
1769 9764 : itemsz = ItemIdGetLength(itemid);
1770 9764 : righthighkey = (IndexTuple) PageGetItem(origpage, itemid);
1771 : Assert(BTreeTupleGetNAtts(righthighkey, rel) > 0);
1772 : Assert(BTreeTupleGetNAtts(righthighkey, rel) <=
1773 : IndexRelationGetNumberOfKeyAttributes(rel));
1774 9764 : if (PageAddItem(rightpage, (Item) righthighkey, itemsz, afterrightoff,
1775 : false, false) == InvalidOffsetNumber)
1776 : {
1777 0 : elog(ERROR, "failed to add high key to the right sibling"
1778 : " while splitting block %u of index \"%s\"",
1779 : origpagenumber, RelationGetRelationName(rel));
1780 : }
1781 9764 : afterrightoff = OffsetNumberNext(afterrightoff);
1782 : }
1783 :
1784 : /*
1785 : * Internal page splits truncate first data item on right page -- it
1786 : * becomes "minus infinity" item for the page. Set this up here.
1787 : */
1788 22446 : minusinfoff = InvalidOffsetNumber;
1789 22446 : if (!isleaf)
1790 258 : minusinfoff = afterrightoff;
1791 :
1792 : /*
1793 : * Now transfer all the data items (non-pivot tuples in isleaf case, or
1794 : * additional pivot tuples in !isleaf case) to the appropriate page.
1795 : *
1796 : * Note: we *must* insert at least the right page's items in item-number
1797 : * order, for the benefit of _bt_restore_page().
1798 : */
1799 6807350 : for (i = P_FIRSTDATAKEY(oopaque); i <= maxoff; i = OffsetNumberNext(i))
1800 : {
1801 : IndexTuple dataitem;
1802 :
1803 6784904 : itemid = PageGetItemId(origpage, i);
1804 6784904 : itemsz = ItemIdGetLength(itemid);
1805 6784904 : dataitem = (IndexTuple) PageGetItem(origpage, itemid);
1806 :
1807 : /* replace original item with nposting due to posting split? */
1808 6784904 : if (i == origpagepostingoff)
1809 : {
1810 : Assert(BTreeTupleIsPosting(dataitem));
1811 : Assert(itemsz == MAXALIGN(IndexTupleSize(nposting)));
1812 68 : dataitem = nposting;
1813 : }
1814 :
1815 : /* does new item belong before this one? */
1816 6784836 : else if (i == newitemoff)
1817 : {
1818 12712 : if (newitemonleft)
1819 : {
1820 : Assert(newitemoff <= firstrightoff);
1821 3326 : if (!_bt_pgaddtup(leftpage, newitemsz, newitem, afterleftoff,
1822 : false))
1823 : {
1824 0 : elog(ERROR, "failed to add new item to the left sibling"
1825 : " while splitting block %u of index \"%s\"",
1826 : origpagenumber, RelationGetRelationName(rel));
1827 : }
1828 3326 : afterleftoff = OffsetNumberNext(afterleftoff);
1829 : }
1830 : else
1831 : {
1832 : Assert(newitemoff >= firstrightoff);
1833 9386 : if (!_bt_pgaddtup(rightpage, newitemsz, newitem, afterrightoff,
1834 : afterrightoff == minusinfoff))
1835 : {
1836 0 : elog(ERROR, "failed to add new item to the right sibling"
1837 : " while splitting block %u of index \"%s\"",
1838 : origpagenumber, RelationGetRelationName(rel));
1839 : }
1840 9386 : afterrightoff = OffsetNumberNext(afterrightoff);
1841 : }
1842 : }
1843 :
1844 : /* decide which page to put it on */
1845 6784904 : if (i < firstrightoff)
1846 : {
1847 5139744 : if (!_bt_pgaddtup(leftpage, itemsz, dataitem, afterleftoff, false))
1848 : {
1849 0 : elog(ERROR, "failed to add old item to the left sibling"
1850 : " while splitting block %u of index \"%s\"",
1851 : origpagenumber, RelationGetRelationName(rel));
1852 : }
1853 5139744 : afterleftoff = OffsetNumberNext(afterleftoff);
1854 : }
1855 : else
1856 : {
1857 1645160 : if (!_bt_pgaddtup(rightpage, itemsz, dataitem, afterrightoff,
1858 : afterrightoff == minusinfoff))
1859 : {
1860 0 : elog(ERROR, "failed to add old item to the right sibling"
1861 : " while splitting block %u of index \"%s\"",
1862 : origpagenumber, RelationGetRelationName(rel));
1863 : }
1864 1645160 : afterrightoff = OffsetNumberNext(afterrightoff);
1865 : }
1866 : }
1867 :
1868 : /* Handle case where newitem goes at the end of rightpage */
1869 22446 : if (i <= newitemoff)
1870 : {
1871 : /*
1872 : * Can't have newitemonleft here; that would imply we were told to put
1873 : * *everything* on the left page, which cannot fit (if it could, we'd
1874 : * not be splitting the page).
1875 : */
1876 : Assert(!newitemonleft && newitemoff == maxoff + 1);
1877 9734 : if (!_bt_pgaddtup(rightpage, newitemsz, newitem, afterrightoff,
1878 : afterrightoff == minusinfoff))
1879 : {
1880 0 : elog(ERROR, "failed to add new item to the right sibling"
1881 : " while splitting block %u of index \"%s\"",
1882 : origpagenumber, RelationGetRelationName(rel));
1883 : }
1884 9734 : afterrightoff = OffsetNumberNext(afterrightoff);
1885 : }
1886 :
1887 : /*
1888 : * We have to grab the original right sibling (if any) and update its prev
1889 : * link. We are guaranteed that this is deadlock-free, since we couple
1890 : * the locks in the standard order: left to right.
1891 : */
1892 22446 : if (!isrightmost)
1893 : {
1894 9764 : sbuf = _bt_getbuf(rel, oopaque->btpo_next, BT_WRITE);
1895 9764 : spage = BufferGetPage(sbuf);
1896 9764 : sopaque = BTPageGetOpaque(spage);
1897 9764 : if (sopaque->btpo_prev != origpagenumber)
1898 : {
1899 0 : ereport(ERROR,
1900 : (errcode(ERRCODE_INDEX_CORRUPTED),
1901 : errmsg_internal("right sibling's left-link doesn't match: "
1902 : "block %u links to %u instead of expected %u in index \"%s\"",
1903 : oopaque->btpo_next, sopaque->btpo_prev, origpagenumber,
1904 : RelationGetRelationName(rel))));
1905 : }
1906 :
1907 : /*
1908 : * Check to see if we can set the SPLIT_END flag in the right-hand
1909 : * split page; this can save some I/O for vacuum since it need not
1910 : * proceed to the right sibling. We can set the flag if the right
1911 : * sibling has a different cycleid: that means it could not be part of
1912 : * a group of pages that were all split off from the same ancestor
1913 : * page. If you're confused, imagine that page A splits to A B and
1914 : * then again, yielding A C B, while vacuum is in progress. Tuples
1915 : * originally in A could now be in either B or C, hence vacuum must
1916 : * examine both pages. But if D, our right sibling, has a different
1917 : * cycleid then it could not contain any tuples that were in A when
1918 : * the vacuum started.
1919 : */
1920 9764 : if (sopaque->btpo_cycleid != ropaque->btpo_cycleid)
1921 0 : ropaque->btpo_flags |= BTP_SPLIT_END;
1922 : }
1923 :
1924 : /*
1925 : * Right sibling is locked, new siblings are prepared, but original page
1926 : * is not updated yet.
1927 : *
1928 : * NO EREPORT(ERROR) till right sibling is updated. We can get away with
1929 : * not starting the critical section till here because we haven't been
1930 : * scribbling on the original page yet; see comments above.
1931 : */
1932 22446 : START_CRIT_SECTION();
1933 :
1934 : /*
1935 : * By here, the original data page has been split into two new halves, and
1936 : * these are correct. The algorithm requires that the left page never
1937 : * move during a split, so we copy the new left page back on top of the
1938 : * original. We need to do this before writing the WAL record, so that
1939 : * XLogInsert can WAL log an image of the page if necessary.
1940 : */
1941 22446 : memcpy(origpage, leftpage, BLCKSZ);
1942 : /* leftpage, lopaque must not be used below here */
1943 :
1944 : /*
1945 : * Move the contents of the right page from its temporary location to the
1946 : * destination buffer, before writing the WAL record. Unlike the left
1947 : * page, the right page and its opaque area are still needed to complete
1948 : * the update of the page, so reinitialize them.
1949 : */
1950 22446 : rightpage = BufferGetPage(rbuf);
1951 22446 : memcpy(rightpage, rightpage_buf.data, BLCKSZ);
1952 22446 : ropaque = BTPageGetOpaque(rightpage);
1953 :
1954 22446 : MarkBufferDirty(buf);
1955 22446 : MarkBufferDirty(rbuf);
1956 :
1957 22446 : if (!isrightmost)
1958 : {
1959 9764 : sopaque->btpo_prev = rightpagenumber;
1960 9764 : MarkBufferDirty(sbuf);
1961 : }
1962 :
1963 : /*
1964 : * Clear INCOMPLETE_SPLIT flag on child if inserting the new item finishes
1965 : * a split
1966 : */
1967 22446 : if (!isleaf)
1968 : {
1969 258 : Page cpage = BufferGetPage(cbuf);
1970 258 : BTPageOpaque cpageop = BTPageGetOpaque(cpage);
1971 :
1972 258 : cpageop->btpo_flags &= ~BTP_INCOMPLETE_SPLIT;
1973 258 : MarkBufferDirty(cbuf);
1974 : }
1975 :
1976 : /* XLOG stuff */
1977 22446 : if (RelationNeedsWAL(rel))
1978 : {
1979 : xl_btree_split xlrec;
1980 : uint8 xlinfo;
1981 : XLogRecPtr recptr;
1982 :
1983 21156 : xlrec.level = ropaque->btpo_level;
1984 : /* See comments below on newitem, orignewitem, and posting lists */
1985 21156 : xlrec.firstrightoff = firstrightoff;
1986 21156 : xlrec.newitemoff = newitemoff;
1987 21156 : xlrec.postingoff = 0;
1988 21156 : if (postingoff != 0 && origpagepostingoff < firstrightoff)
1989 36 : xlrec.postingoff = postingoff;
1990 :
1991 21156 : XLogBeginInsert();
1992 21156 : XLogRegisterData(&xlrec, SizeOfBtreeSplit);
1993 :
1994 21156 : XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
1995 21156 : XLogRegisterBuffer(1, rbuf, REGBUF_WILL_INIT);
1996 : /* Log original right sibling, since we've changed its prev-pointer */
1997 21156 : if (!isrightmost)
1998 9752 : XLogRegisterBuffer(2, sbuf, REGBUF_STANDARD);
1999 21156 : if (!isleaf)
2000 258 : XLogRegisterBuffer(3, cbuf, REGBUF_STANDARD);
2001 :
2002 : /*
2003 : * Log the new item, if it was inserted on the left page. (If it was
2004 : * put on the right page, we don't need to explicitly WAL log it
2005 : * because it's included with all the other items on the right page.)
2006 : * Show the new item as belonging to the left page buffer, so that it
2007 : * is not stored if XLogInsert decides it needs a full-page image of
2008 : * the left page. We always store newitemoff in the record, though.
2009 : *
2010 : * The details are sometimes slightly different for page splits that
2011 : * coincide with a posting list split. If both the replacement
2012 : * posting list and newitem go on the right page, then we don't need
2013 : * to log anything extra, just like the simple !newitemonleft
2014 : * no-posting-split case (postingoff is set to zero in the WAL record,
2015 : * so recovery doesn't need to process a posting list split at all).
2016 : * Otherwise, we set postingoff and log orignewitem instead of
2017 : * newitem, despite having actually inserted newitem. REDO routine
2018 : * must reconstruct nposting and newitem using _bt_swap_posting().
2019 : *
2020 : * Note: It's possible that our page split point is the point that
2021 : * makes the posting list lastleft and newitem firstright. This is
2022 : * the only case where we log orignewitem/newitem despite newitem
2023 : * going on the right page. If XLogInsert decides that it can omit
2024 : * orignewitem due to logging a full-page image of the left page,
2025 : * everything still works out, since recovery only needs to log
2026 : * orignewitem for items on the left page (just like the regular
2027 : * newitem-logged case).
2028 : */
2029 21156 : if (newitemonleft && xlrec.postingoff == 0)
2030 3290 : XLogRegisterBufData(0, newitem, newitemsz);
2031 17866 : else if (xlrec.postingoff != 0)
2032 : {
2033 : Assert(isleaf);
2034 : Assert(newitemonleft || firstrightoff == newitemoff);
2035 : Assert(newitemsz == IndexTupleSize(orignewitem));
2036 36 : XLogRegisterBufData(0, orignewitem, newitemsz);
2037 : }
2038 :
2039 : /* Log the left page's new high key */
2040 21156 : if (!isleaf)
2041 : {
2042 : /* lefthighkey isn't local copy, get current pointer */
2043 258 : itemid = PageGetItemId(origpage, P_HIKEY);
2044 258 : lefthighkey = (IndexTuple) PageGetItem(origpage, itemid);
2045 : }
2046 21156 : XLogRegisterBufData(0, lefthighkey,
2047 21156 : MAXALIGN(IndexTupleSize(lefthighkey)));
2048 :
2049 : /*
2050 : * Log the contents of the right page in the format understood by
2051 : * _bt_restore_page(). The whole right page will be recreated.
2052 : *
2053 : * Direct access to page is not good but faster - we should implement
2054 : * some new func in page API. Note we only store the tuples
2055 : * themselves, knowing that they were inserted in item-number order
2056 : * and so the line pointers can be reconstructed. See comments for
2057 : * _bt_restore_page().
2058 : */
2059 21156 : XLogRegisterBufData(1,
2060 21156 : (char *) rightpage + ((PageHeader) rightpage)->pd_upper,
2061 21156 : ((PageHeader) rightpage)->pd_special - ((PageHeader) rightpage)->pd_upper);
2062 :
2063 21156 : xlinfo = newitemonleft ? XLOG_BTREE_SPLIT_L : XLOG_BTREE_SPLIT_R;
2064 21156 : recptr = XLogInsert(RM_BTREE_ID, xlinfo);
2065 :
2066 21156 : PageSetLSN(origpage, recptr);
2067 21156 : PageSetLSN(rightpage, recptr);
2068 21156 : if (!isrightmost)
2069 9752 : PageSetLSN(spage, recptr);
2070 21156 : if (!isleaf)
2071 258 : PageSetLSN(BufferGetPage(cbuf), recptr);
2072 : }
2073 :
2074 22446 : END_CRIT_SECTION();
2075 :
2076 : /* release the old right sibling */
2077 22446 : if (!isrightmost)
2078 9764 : _bt_relbuf(rel, sbuf);
2079 :
2080 : /* release the child */
2081 22446 : if (!isleaf)
2082 258 : _bt_relbuf(rel, cbuf);
2083 :
2084 : /* be tidy */
2085 22446 : if (isleaf)
2086 22188 : pfree(lefthighkey);
2087 :
2088 : /* split's done */
2089 22446 : return rbuf;
2090 : }
2091 :
2092 : /*
2093 : * _bt_insert_parent() -- Insert downlink into parent, completing split.
2094 : *
2095 : * On entry, buf and rbuf are the left and right split pages, which we
2096 : * still hold write locks on. Both locks will be released here. We
2097 : * release the rbuf lock once we have a write lock on the page that we
2098 : * intend to insert a downlink to rbuf on (i.e. buf's current parent page).
2099 : * The lock on buf is released at the same point as the lock on the parent
2100 : * page, since buf's INCOMPLETE_SPLIT flag must be cleared by the same
2101 : * atomic operation that completes the split by inserting a new downlink.
2102 : *
2103 : * stack - stack showing how we got here. Will be NULL when splitting true
2104 : * root, or during concurrent root split, where we can be inefficient
2105 : * isroot - we split the true root
2106 : * isonly - we split a page alone on its level (might have been fast root)
2107 : */
2108 : static void
2109 22446 : _bt_insert_parent(Relation rel,
2110 : Relation heaprel,
2111 : Buffer buf,
2112 : Buffer rbuf,
2113 : BTStack stack,
2114 : bool isroot,
2115 : bool isonly)
2116 : {
2117 : Assert(heaprel != NULL);
2118 :
2119 : /*
2120 : * Here we have to do something Lehman and Yao don't talk about: deal with
2121 : * a root split and construction of a new root. If our stack is empty
2122 : * then we have just split a node on what had been the root level when we
2123 : * descended the tree. If it was still the root then we perform a
2124 : * new-root construction. If it *wasn't* the root anymore, search to find
2125 : * the next higher level that someone constructed meanwhile, and find the
2126 : * right place to insert as for the normal case.
2127 : *
2128 : * If we have to search for the parent level, we do so by re-descending
2129 : * from the root. This is not super-efficient, but it's rare enough not
2130 : * to matter.
2131 : */
2132 22446 : if (isroot)
2133 : {
2134 : Buffer rootbuf;
2135 :
2136 : Assert(stack == NULL);
2137 : Assert(isonly);
2138 : /* create a new root node one level up and update the metapage */
2139 1366 : rootbuf = _bt_newlevel(rel, heaprel, buf, rbuf);
2140 : /* release the split buffers */
2141 1366 : _bt_relbuf(rel, rootbuf);
2142 1366 : _bt_relbuf(rel, rbuf);
2143 1366 : _bt_relbuf(rel, buf);
2144 : }
2145 : else
2146 : {
2147 21080 : BlockNumber bknum = BufferGetBlockNumber(buf);
2148 21080 : BlockNumber rbknum = BufferGetBlockNumber(rbuf);
2149 21080 : Page page = BufferGetPage(buf);
2150 : IndexTuple new_item;
2151 : BTStackData fakestack;
2152 : IndexTuple ritem;
2153 : Buffer pbuf;
2154 :
2155 21080 : if (stack == NULL)
2156 : {
2157 : BTPageOpaque opaque;
2158 :
2159 24 : elog(DEBUG2, "concurrent ROOT page split");
2160 24 : opaque = BTPageGetOpaque(page);
2161 :
2162 : /*
2163 : * We should never reach here when a leaf page split takes place
2164 : * despite the insert of newitem being able to apply the fastpath
2165 : * optimization. Make sure of that with an assertion.
2166 : *
2167 : * This is more of a performance issue than a correctness issue.
2168 : * The fastpath won't have a descent stack. Using a phony stack
2169 : * here works, but never rely on that. The fastpath should be
2170 : * rejected within _bt_search_insert() when the rightmost leaf
2171 : * page will split, since it's faster to go through _bt_search()
2172 : * and get a stack in the usual way.
2173 : */
2174 : Assert(!(P_ISLEAF(opaque) &&
2175 : BlockNumberIsValid(RelationGetTargetBlock(rel))));
2176 :
2177 : /* Find the leftmost page at the next level up */
2178 24 : pbuf = _bt_get_endpoint(rel, opaque->btpo_level + 1, false);
2179 : /* Set up a phony stack entry pointing there */
2180 24 : stack = &fakestack;
2181 24 : stack->bts_blkno = BufferGetBlockNumber(pbuf);
2182 24 : stack->bts_offset = InvalidOffsetNumber;
2183 24 : stack->bts_parent = NULL;
2184 24 : _bt_relbuf(rel, pbuf);
2185 : }
2186 :
2187 : /* get high key from left, a strict lower bound for new right page */
2188 21080 : ritem = (IndexTuple) PageGetItem(page,
2189 21080 : PageGetItemId(page, P_HIKEY));
2190 :
2191 : /* form an index tuple that points at the new right page */
2192 21080 : new_item = CopyIndexTuple(ritem);
2193 21080 : BTreeTupleSetDownLink(new_item, rbknum);
2194 :
2195 : /*
2196 : * Re-find and write lock the parent of buf.
2197 : *
2198 : * It's possible that the location of buf's downlink has changed since
2199 : * our initial _bt_search() descent. _bt_getstackbuf() will detect
2200 : * and recover from this, updating the stack, which ensures that the
2201 : * new downlink will be inserted at the correct offset. Even buf's
2202 : * parent may have changed.
2203 : */
2204 21080 : pbuf = _bt_getstackbuf(rel, heaprel, stack, bknum);
2205 :
2206 : /*
2207 : * Unlock the right child. The left child will be unlocked in
2208 : * _bt_insertonpg().
2209 : *
2210 : * Unlocking the right child must be delayed until here to ensure that
2211 : * no concurrent VACUUM operation can become confused. Page deletion
2212 : * cannot be allowed to fail to re-find a downlink for the rbuf page.
2213 : * (Actually, this is just a vestige of how things used to work. The
2214 : * page deletion code is expected to check for the INCOMPLETE_SPLIT
2215 : * flag on the left child. It won't attempt deletion of the right
2216 : * child until the split is complete. Despite all this, we opt to
2217 : * conservatively delay unlocking the right child until here.)
2218 : */
2219 21080 : _bt_relbuf(rel, rbuf);
2220 :
2221 21080 : if (pbuf == InvalidBuffer)
2222 0 : ereport(ERROR,
2223 : (errcode(ERRCODE_INDEX_CORRUPTED),
2224 : errmsg_internal("failed to re-find parent key in index \"%s\" for split pages %u/%u",
2225 : RelationGetRelationName(rel), bknum, rbknum)));
2226 :
2227 : /* Recursively insert into the parent */
2228 42160 : _bt_insertonpg(rel, heaprel, NULL, pbuf, buf, stack->bts_parent,
2229 21080 : new_item, MAXALIGN(IndexTupleSize(new_item)),
2230 21080 : stack->bts_offset + 1, 0, isonly);
2231 :
2232 : /* be tidy */
2233 21080 : pfree(new_item);
2234 : }
2235 22446 : }
2236 :
2237 : /*
2238 : * _bt_finish_split() -- Finish an incomplete split
2239 : *
2240 : * A crash or other failure can leave a split incomplete. The insertion
2241 : * routines won't allow to insert on a page that is incompletely split.
2242 : * Before inserting on such a page, call _bt_finish_split().
2243 : *
2244 : * On entry, 'lbuf' must be locked in write-mode. On exit, it is unlocked
2245 : * and unpinned.
2246 : *
2247 : * Caller must provide a valid heaprel, since finishing a page split requires
2248 : * allocating a new page if and when the parent page splits in turn.
2249 : */
2250 : void
2251 0 : _bt_finish_split(Relation rel, Relation heaprel, Buffer lbuf, BTStack stack)
2252 : {
2253 0 : Page lpage = BufferGetPage(lbuf);
2254 0 : BTPageOpaque lpageop = BTPageGetOpaque(lpage);
2255 : Buffer rbuf;
2256 : Page rpage;
2257 : BTPageOpaque rpageop;
2258 : bool wasroot;
2259 : bool wasonly;
2260 :
2261 : Assert(P_INCOMPLETE_SPLIT(lpageop));
2262 : Assert(heaprel != NULL);
2263 :
2264 : /* Lock right sibling, the one missing the downlink */
2265 0 : rbuf = _bt_getbuf(rel, lpageop->btpo_next, BT_WRITE);
2266 0 : rpage = BufferGetPage(rbuf);
2267 0 : rpageop = BTPageGetOpaque(rpage);
2268 :
2269 : /* Could this be a root split? */
2270 0 : if (!stack)
2271 : {
2272 : Buffer metabuf;
2273 : Page metapg;
2274 : BTMetaPageData *metad;
2275 :
2276 : /* acquire lock on the metapage */
2277 0 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_WRITE);
2278 0 : metapg = BufferGetPage(metabuf);
2279 0 : metad = BTPageGetMeta(metapg);
2280 :
2281 0 : wasroot = (metad->btm_root == BufferGetBlockNumber(lbuf));
2282 :
2283 0 : _bt_relbuf(rel, metabuf);
2284 : }
2285 : else
2286 0 : wasroot = false;
2287 :
2288 : /* Was this the only page on the level before split? */
2289 0 : wasonly = (P_LEFTMOST(lpageop) && P_RIGHTMOST(rpageop));
2290 :
2291 0 : elog(DEBUG1, "finishing incomplete split of %u/%u",
2292 : BufferGetBlockNumber(lbuf), BufferGetBlockNumber(rbuf));
2293 :
2294 0 : _bt_insert_parent(rel, heaprel, lbuf, rbuf, stack, wasroot, wasonly);
2295 0 : }
2296 :
2297 : /*
2298 : * _bt_getstackbuf() -- Walk back up the tree one step, and find the pivot
2299 : * tuple whose downlink points to child page.
2300 : *
2301 : * Caller passes child's block number, which is used to identify
2302 : * associated pivot tuple in parent page using a linear search that
2303 : * matches on pivot's downlink/block number. The expected location of
2304 : * the pivot tuple is taken from the stack one level above the child
2305 : * page. This is used as a starting point. Insertions into the
2306 : * parent level could cause the pivot tuple to move right; deletions
2307 : * could cause it to move left, but not left of the page we previously
2308 : * found it on.
2309 : *
2310 : * Caller can use its stack to relocate the pivot tuple/downlink for
2311 : * any same-level page to the right of the page found by its initial
2312 : * descent. This is necessary because of the possibility that caller
2313 : * moved right to recover from a concurrent page split. It's also
2314 : * convenient for certain callers to be able to step right when there
2315 : * wasn't a concurrent page split, while still using their original
2316 : * stack. For example, the checkingunique _bt_doinsert() case may
2317 : * have to step right when there are many physical duplicates, and its
2318 : * scantid forces an insertion to the right of the "first page the
2319 : * value could be on". (This is also relied on by all of our callers
2320 : * when dealing with !heapkeyspace indexes.)
2321 : *
2322 : * Returns write-locked parent page buffer, or InvalidBuffer if pivot
2323 : * tuple not found (should not happen). Adjusts bts_blkno &
2324 : * bts_offset if changed. Page split caller should insert its new
2325 : * pivot tuple for its new right sibling page on parent page, at the
2326 : * offset number bts_offset + 1.
2327 : */
2328 : Buffer
2329 26882 : _bt_getstackbuf(Relation rel, Relation heaprel, BTStack stack, BlockNumber child)
2330 : {
2331 : BlockNumber blkno;
2332 : OffsetNumber start;
2333 :
2334 26882 : blkno = stack->bts_blkno;
2335 26882 : start = stack->bts_offset;
2336 :
2337 : for (;;)
2338 14 : {
2339 : Buffer buf;
2340 : Page page;
2341 : BTPageOpaque opaque;
2342 :
2343 26896 : buf = _bt_getbuf(rel, blkno, BT_WRITE);
2344 26896 : page = BufferGetPage(buf);
2345 26896 : opaque = BTPageGetOpaque(page);
2346 :
2347 : Assert(heaprel != NULL);
2348 26896 : if (P_INCOMPLETE_SPLIT(opaque))
2349 : {
2350 0 : _bt_finish_split(rel, heaprel, buf, stack->bts_parent);
2351 0 : continue;
2352 : }
2353 :
2354 26896 : if (!P_IGNORE(opaque))
2355 : {
2356 : OffsetNumber offnum,
2357 : minoff,
2358 : maxoff;
2359 : ItemId itemid;
2360 : IndexTuple item;
2361 :
2362 26882 : minoff = P_FIRSTDATAKEY(opaque);
2363 26882 : maxoff = PageGetMaxOffsetNumber(page);
2364 :
2365 : /*
2366 : * start = InvalidOffsetNumber means "search the whole page". We
2367 : * need this test anyway due to possibility that page has a high
2368 : * key now when it didn't before.
2369 : */
2370 26882 : if (start < minoff)
2371 38 : start = minoff;
2372 :
2373 : /*
2374 : * Need this check too, to guard against possibility that page
2375 : * split since we visited it originally.
2376 : */
2377 26882 : if (start > maxoff)
2378 0 : start = OffsetNumberNext(maxoff);
2379 :
2380 : /*
2381 : * These loops will check every item on the page --- but in an
2382 : * order that's attuned to the probability of where it actually
2383 : * is. Scan to the right first, then to the left.
2384 : */
2385 26882 : for (offnum = start;
2386 26978 : offnum <= maxoff;
2387 96 : offnum = OffsetNumberNext(offnum))
2388 : {
2389 26978 : itemid = PageGetItemId(page, offnum);
2390 26978 : item = (IndexTuple) PageGetItem(page, itemid);
2391 :
2392 26978 : if (BTreeTupleGetDownLink(item) == child)
2393 : {
2394 : /* Return accurate pointer to where link is now */
2395 26882 : stack->bts_blkno = blkno;
2396 26882 : stack->bts_offset = offnum;
2397 26882 : return buf;
2398 : }
2399 : }
2400 :
2401 0 : for (offnum = OffsetNumberPrev(start);
2402 0 : offnum >= minoff;
2403 0 : offnum = OffsetNumberPrev(offnum))
2404 : {
2405 0 : itemid = PageGetItemId(page, offnum);
2406 0 : item = (IndexTuple) PageGetItem(page, itemid);
2407 :
2408 0 : if (BTreeTupleGetDownLink(item) == child)
2409 : {
2410 : /* Return accurate pointer to where link is now */
2411 0 : stack->bts_blkno = blkno;
2412 0 : stack->bts_offset = offnum;
2413 0 : return buf;
2414 : }
2415 : }
2416 : }
2417 :
2418 : /*
2419 : * The item we're looking for moved right at least one page.
2420 : *
2421 : * Lehman and Yao couple/chain locks when moving right here, which we
2422 : * can avoid. See nbtree/README.
2423 : */
2424 14 : if (P_RIGHTMOST(opaque))
2425 : {
2426 0 : _bt_relbuf(rel, buf);
2427 0 : return InvalidBuffer;
2428 : }
2429 14 : blkno = opaque->btpo_next;
2430 14 : start = InvalidOffsetNumber;
2431 14 : _bt_relbuf(rel, buf);
2432 : }
2433 : }
2434 :
2435 : /*
2436 : * _bt_newlevel() -- Create a new level above root page.
2437 : *
2438 : * We've just split the old root page and need to create a new one.
2439 : * In order to do this, we add a new root page to the file, then lock
2440 : * the metadata page and update it. This is guaranteed to be deadlock-
2441 : * free, because all readers release their locks on the metadata page
2442 : * before trying to lock the root, and all writers lock the root before
2443 : * trying to lock the metadata page. We have a write lock on the old
2444 : * root page, so we have not introduced any cycles into the waits-for
2445 : * graph.
2446 : *
2447 : * On entry, lbuf (the old root) and rbuf (its new peer) are write-
2448 : * locked. On exit, a new root page exists with entries for the
2449 : * two new children, metapage is updated and unlocked/unpinned.
2450 : * The new root buffer is returned to caller which has to unlock/unpin
2451 : * lbuf, rbuf & rootbuf.
2452 : */
2453 : static Buffer
2454 1366 : _bt_newlevel(Relation rel, Relation heaprel, Buffer lbuf, Buffer rbuf)
2455 : {
2456 : Buffer rootbuf;
2457 : Page lpage,
2458 : rootpage;
2459 : BlockNumber lbkno,
2460 : rbkno;
2461 : BlockNumber rootblknum;
2462 : BTPageOpaque rootopaque;
2463 : BTPageOpaque lopaque;
2464 : ItemId itemid;
2465 : IndexTuple item;
2466 : IndexTuple left_item;
2467 : Size left_item_sz;
2468 : IndexTuple right_item;
2469 : Size right_item_sz;
2470 : Buffer metabuf;
2471 : Page metapg;
2472 : BTMetaPageData *metad;
2473 :
2474 1366 : lbkno = BufferGetBlockNumber(lbuf);
2475 1366 : rbkno = BufferGetBlockNumber(rbuf);
2476 1366 : lpage = BufferGetPage(lbuf);
2477 1366 : lopaque = BTPageGetOpaque(lpage);
2478 :
2479 : /* get a new root page */
2480 1366 : rootbuf = _bt_allocbuf(rel, heaprel);
2481 1366 : rootpage = BufferGetPage(rootbuf);
2482 1366 : rootblknum = BufferGetBlockNumber(rootbuf);
2483 :
2484 : /* acquire lock on the metapage */
2485 1366 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_WRITE);
2486 1366 : metapg = BufferGetPage(metabuf);
2487 1366 : metad = BTPageGetMeta(metapg);
2488 :
2489 : /*
2490 : * Create downlink item for left page (old root). The key value used is
2491 : * "minus infinity", a sentinel value that's reliably less than any real
2492 : * key value that could appear in the left page.
2493 : */
2494 1366 : left_item_sz = sizeof(IndexTupleData);
2495 1366 : left_item = (IndexTuple) palloc(left_item_sz);
2496 1366 : left_item->t_info = left_item_sz;
2497 1366 : BTreeTupleSetDownLink(left_item, lbkno);
2498 1366 : BTreeTupleSetNAtts(left_item, 0, false);
2499 :
2500 : /*
2501 : * Create downlink item for right page. The key for it is obtained from
2502 : * the "high key" position in the left page.
2503 : */
2504 1366 : itemid = PageGetItemId(lpage, P_HIKEY);
2505 1366 : right_item_sz = ItemIdGetLength(itemid);
2506 1366 : item = (IndexTuple) PageGetItem(lpage, itemid);
2507 1366 : right_item = CopyIndexTuple(item);
2508 1366 : BTreeTupleSetDownLink(right_item, rbkno);
2509 :
2510 : /* NO EREPORT(ERROR) from here till newroot op is logged */
2511 1366 : START_CRIT_SECTION();
2512 :
2513 : /* upgrade metapage if needed */
2514 1366 : if (metad->btm_version < BTREE_NOVAC_VERSION)
2515 0 : _bt_upgrademetapage(metapg);
2516 :
2517 : /* set btree special data */
2518 1366 : rootopaque = BTPageGetOpaque(rootpage);
2519 1366 : rootopaque->btpo_prev = rootopaque->btpo_next = P_NONE;
2520 1366 : rootopaque->btpo_flags = BTP_ROOT;
2521 1366 : rootopaque->btpo_level =
2522 1366 : (BTPageGetOpaque(lpage))->btpo_level + 1;
2523 1366 : rootopaque->btpo_cycleid = 0;
2524 :
2525 : /* update metapage data */
2526 1366 : metad->btm_root = rootblknum;
2527 1366 : metad->btm_level = rootopaque->btpo_level;
2528 1366 : metad->btm_fastroot = rootblknum;
2529 1366 : metad->btm_fastlevel = rootopaque->btpo_level;
2530 :
2531 : /*
2532 : * Insert the left page pointer into the new root page. The root page is
2533 : * the rightmost page on its level so there is no "high key" in it; the
2534 : * two items will go into positions P_HIKEY and P_FIRSTKEY.
2535 : *
2536 : * Note: we *must* insert the two items in item-number order, for the
2537 : * benefit of _bt_restore_page().
2538 : */
2539 : Assert(BTreeTupleGetNAtts(left_item, rel) == 0);
2540 1366 : if (PageAddItem(rootpage, (Item) left_item, left_item_sz, P_HIKEY,
2541 : false, false) == InvalidOffsetNumber)
2542 0 : elog(PANIC, "failed to add leftkey to new root page"
2543 : " while splitting block %u of index \"%s\"",
2544 : BufferGetBlockNumber(lbuf), RelationGetRelationName(rel));
2545 :
2546 : /*
2547 : * insert the right page pointer into the new root page.
2548 : */
2549 : Assert(BTreeTupleGetNAtts(right_item, rel) > 0);
2550 : Assert(BTreeTupleGetNAtts(right_item, rel) <=
2551 : IndexRelationGetNumberOfKeyAttributes(rel));
2552 1366 : if (PageAddItem(rootpage, (Item) right_item, right_item_sz, P_FIRSTKEY,
2553 : false, false) == InvalidOffsetNumber)
2554 0 : elog(PANIC, "failed to add rightkey to new root page"
2555 : " while splitting block %u of index \"%s\"",
2556 : BufferGetBlockNumber(lbuf), RelationGetRelationName(rel));
2557 :
2558 : /* Clear the incomplete-split flag in the left child */
2559 : Assert(P_INCOMPLETE_SPLIT(lopaque));
2560 1366 : lopaque->btpo_flags &= ~BTP_INCOMPLETE_SPLIT;
2561 1366 : MarkBufferDirty(lbuf);
2562 :
2563 1366 : MarkBufferDirty(rootbuf);
2564 1366 : MarkBufferDirty(metabuf);
2565 :
2566 : /* XLOG stuff */
2567 1366 : if (RelationNeedsWAL(rel))
2568 : {
2569 : xl_btree_newroot xlrec;
2570 : XLogRecPtr recptr;
2571 : xl_btree_metadata md;
2572 :
2573 1330 : xlrec.rootblk = rootblknum;
2574 1330 : xlrec.level = metad->btm_level;
2575 :
2576 1330 : XLogBeginInsert();
2577 1330 : XLogRegisterData(&xlrec, SizeOfBtreeNewroot);
2578 :
2579 1330 : XLogRegisterBuffer(0, rootbuf, REGBUF_WILL_INIT);
2580 1330 : XLogRegisterBuffer(1, lbuf, REGBUF_STANDARD);
2581 1330 : XLogRegisterBuffer(2, metabuf, REGBUF_WILL_INIT | REGBUF_STANDARD);
2582 :
2583 : Assert(metad->btm_version >= BTREE_NOVAC_VERSION);
2584 1330 : md.version = metad->btm_version;
2585 1330 : md.root = rootblknum;
2586 1330 : md.level = metad->btm_level;
2587 1330 : md.fastroot = rootblknum;
2588 1330 : md.fastlevel = metad->btm_level;
2589 1330 : md.last_cleanup_num_delpages = metad->btm_last_cleanup_num_delpages;
2590 1330 : md.allequalimage = metad->btm_allequalimage;
2591 :
2592 1330 : XLogRegisterBufData(2, &md, sizeof(xl_btree_metadata));
2593 :
2594 : /*
2595 : * Direct access to page is not good but faster - we should implement
2596 : * some new func in page API.
2597 : */
2598 1330 : XLogRegisterBufData(0,
2599 1330 : (char *) rootpage + ((PageHeader) rootpage)->pd_upper,
2600 1330 : ((PageHeader) rootpage)->pd_special -
2601 1330 : ((PageHeader) rootpage)->pd_upper);
2602 :
2603 1330 : recptr = XLogInsert(RM_BTREE_ID, XLOG_BTREE_NEWROOT);
2604 :
2605 1330 : PageSetLSN(lpage, recptr);
2606 1330 : PageSetLSN(rootpage, recptr);
2607 1330 : PageSetLSN(metapg, recptr);
2608 : }
2609 :
2610 1366 : END_CRIT_SECTION();
2611 :
2612 : /* done with metapage */
2613 1366 : _bt_relbuf(rel, metabuf);
2614 :
2615 1366 : pfree(left_item);
2616 1366 : pfree(right_item);
2617 :
2618 1366 : return rootbuf;
2619 : }
2620 :
2621 : /*
2622 : * _bt_pgaddtup() -- add a data item to a particular page during split.
2623 : *
2624 : * The difference between this routine and a bare PageAddItem call is
2625 : * that this code can deal with the first data item on an internal btree
2626 : * page in passing. This data item (which is called "firstright" within
2627 : * _bt_split()) has a key that must be treated as minus infinity after
2628 : * the split. Therefore, we truncate away all attributes when caller
2629 : * specifies it's the first data item on page (downlink is not changed,
2630 : * though). This extra step is only needed for the right page of an
2631 : * internal page split. There is no need to do this for the first data
2632 : * item on the existing/left page, since that will already have been
2633 : * truncated during an earlier page split.
2634 : *
2635 : * See _bt_split() for a high level explanation of why we truncate here.
2636 : * Note that this routine has nothing to do with suffix truncation,
2637 : * despite using some of the same infrastructure.
2638 : */
2639 : static inline bool
2640 6807350 : _bt_pgaddtup(Page page,
2641 : Size itemsize,
2642 : IndexTuple itup,
2643 : OffsetNumber itup_off,
2644 : bool newfirstdataitem)
2645 : {
2646 : IndexTupleData trunctuple;
2647 :
2648 6807350 : if (newfirstdataitem)
2649 : {
2650 258 : trunctuple = *itup;
2651 258 : trunctuple.t_info = sizeof(IndexTupleData);
2652 258 : BTreeTupleSetNAtts(&trunctuple, 0, false);
2653 258 : itup = &trunctuple;
2654 258 : itemsize = sizeof(IndexTupleData);
2655 : }
2656 :
2657 6807350 : if (unlikely(PageAddItem(page, (Item) itup, itemsize, itup_off, false,
2658 : false) == InvalidOffsetNumber))
2659 0 : return false;
2660 :
2661 6807350 : return true;
2662 : }
2663 :
2664 : /*
2665 : * _bt_delete_or_dedup_one_page - Try to avoid a leaf page split.
2666 : *
2667 : * There are three operations performed here: simple index deletion, bottom-up
2668 : * index deletion, and deduplication. If all three operations fail to free
2669 : * enough space for the incoming item then caller will go on to split the
2670 : * page. We always consider simple deletion first. If that doesn't work out
2671 : * we consider alternatives. Callers that only want us to consider simple
2672 : * deletion (without any fallback) ask for that using the 'simpleonly'
2673 : * argument.
2674 : *
2675 : * We usually pick only one alternative "complex" operation when simple
2676 : * deletion alone won't prevent a page split. The 'checkingunique',
2677 : * 'uniquedup', and 'indexUnchanged' arguments are used for that.
2678 : *
2679 : * Note: We used to only delete LP_DEAD items when the BTP_HAS_GARBAGE page
2680 : * level flag was found set. The flag was useful back when there wasn't
2681 : * necessarily one single page for a duplicate tuple to go on (before heap TID
2682 : * became a part of the key space in version 4 indexes). But we don't
2683 : * actually look at the flag anymore (it's not a gating condition for our
2684 : * caller). That would cause us to miss tuples that are safe to delete,
2685 : * without getting any benefit in return. We know that the alternative is to
2686 : * split the page; scanning the line pointer array in passing won't have
2687 : * noticeable overhead. (We still maintain the BTP_HAS_GARBAGE flag despite
2688 : * all this because !heapkeyspace indexes must still do a "getting tired"
2689 : * linear search, and so are likely to get some benefit from using it as a
2690 : * gating condition.)
2691 : */
2692 : static void
2693 50754 : _bt_delete_or_dedup_one_page(Relation rel, Relation heapRel,
2694 : BTInsertState insertstate,
2695 : bool simpleonly, bool checkingunique,
2696 : bool uniquedup, bool indexUnchanged)
2697 : {
2698 : OffsetNumber deletable[MaxIndexTuplesPerPage];
2699 50754 : int ndeletable = 0;
2700 : OffsetNumber offnum,
2701 : minoff,
2702 : maxoff;
2703 50754 : Buffer buffer = insertstate->buf;
2704 50754 : BTScanInsert itup_key = insertstate->itup_key;
2705 50754 : Page page = BufferGetPage(buffer);
2706 50754 : BTPageOpaque opaque = BTPageGetOpaque(page);
2707 :
2708 : Assert(P_ISLEAF(opaque));
2709 : Assert(simpleonly || itup_key->heapkeyspace);
2710 : Assert(!simpleonly || (!checkingunique && !uniquedup && !indexUnchanged));
2711 :
2712 : /*
2713 : * Scan over all items to see which ones need to be deleted according to
2714 : * LP_DEAD flags. We'll usually manage to delete a few extra items that
2715 : * are not marked LP_DEAD in passing. Often the extra items that actually
2716 : * end up getting deleted are items that would have had their LP_DEAD bit
2717 : * set before long anyway (if we opted not to include them as extras).
2718 : */
2719 50754 : minoff = P_FIRSTDATAKEY(opaque);
2720 50754 : maxoff = PageGetMaxOffsetNumber(page);
2721 50754 : for (offnum = minoff;
2722 13582680 : offnum <= maxoff;
2723 13531926 : offnum = OffsetNumberNext(offnum))
2724 : {
2725 13531926 : ItemId itemId = PageGetItemId(page, offnum);
2726 :
2727 13531926 : if (ItemIdIsDead(itemId))
2728 251706 : deletable[ndeletable++] = offnum;
2729 : }
2730 :
2731 50754 : if (ndeletable > 0)
2732 : {
2733 7450 : _bt_simpledel_pass(rel, buffer, heapRel, deletable, ndeletable,
2734 : insertstate->itup, minoff, maxoff);
2735 7450 : insertstate->bounds_valid = false;
2736 :
2737 : /* Return when a page split has already been avoided */
2738 7450 : if (PageGetFreeSpace(page) >= insertstate->itemsz)
2739 23068 : return;
2740 :
2741 : /* Might as well assume duplicates (if checkingunique) */
2742 98 : uniquedup = true;
2743 : }
2744 :
2745 : /*
2746 : * We're done with simple deletion. Return early with callers that only
2747 : * call here so that simple deletion can be considered. This includes
2748 : * callers that explicitly ask for this and checkingunique callers that
2749 : * probably don't have any version churn duplicates on the page.
2750 : *
2751 : * Note: The page's BTP_HAS_GARBAGE hint flag may still be set when we
2752 : * return at this point (or when we go on the try either or both of our
2753 : * other strategies and they also fail). We do not bother expending a
2754 : * separate write to clear it, however. Caller will definitely clear it
2755 : * when it goes on to split the page (note also that the deduplication
2756 : * process will clear the flag in passing, just to keep things tidy).
2757 : */
2758 43402 : if (simpleonly || (checkingunique && !uniquedup))
2759 : {
2760 : Assert(!indexUnchanged);
2761 15354 : return;
2762 : }
2763 :
2764 : /* Assume bounds about to be invalidated (this is almost certain now) */
2765 28048 : insertstate->bounds_valid = false;
2766 :
2767 : /*
2768 : * Perform bottom-up index deletion pass when executor hint indicated that
2769 : * incoming item is logically unchanged, or for a unique index that is
2770 : * known to have physical duplicates for some other reason. (There is a
2771 : * large overlap between these two cases for a unique index. It's worth
2772 : * having both triggering conditions in order to apply the optimization in
2773 : * the event of successive related INSERT and DELETE statements.)
2774 : *
2775 : * We'll go on to do a deduplication pass when a bottom-up pass fails to
2776 : * delete an acceptable amount of free space (a significant fraction of
2777 : * the page, or space for the new item, whichever is greater).
2778 : *
2779 : * Note: Bottom-up index deletion uses the same equality/equivalence
2780 : * routines as deduplication internally. However, it does not merge
2781 : * together index tuples, so the same correctness considerations do not
2782 : * apply. We deliberately omit an index-is-allequalimage test here.
2783 : */
2784 31952 : if ((indexUnchanged || uniquedup) &&
2785 3904 : _bt_bottomupdel_pass(rel, buffer, heapRel, insertstate->itemsz))
2786 362 : return;
2787 :
2788 : /* Perform deduplication pass (when enabled and index-is-allequalimage) */
2789 27686 : if (BTGetDeduplicateItems(rel) && itup_key->allequalimage)
2790 27668 : _bt_dedup_pass(rel, buffer, insertstate->itup, insertstate->itemsz,
2791 27668 : (indexUnchanged || uniquedup));
2792 : }
2793 :
2794 : /*
2795 : * _bt_simpledel_pass - Simple index tuple deletion pass.
2796 : *
2797 : * We delete all LP_DEAD-set index tuples on a leaf page. The offset numbers
2798 : * of all such tuples are determined by caller (caller passes these to us as
2799 : * its 'deletable' argument).
2800 : *
2801 : * We might also delete extra index tuples that turn out to be safe to delete
2802 : * in passing (though they must be cheap to check in passing to begin with).
2803 : * There is no certainty that any extra tuples will be deleted, though. The
2804 : * high level goal of the approach we take is to get the most out of each call
2805 : * here (without noticeably increasing the per-call overhead compared to what
2806 : * we need to do just to be able to delete the page's LP_DEAD-marked index
2807 : * tuples).
2808 : *
2809 : * The number of extra index tuples that turn out to be deletable might
2810 : * greatly exceed the number of LP_DEAD-marked index tuples due to various
2811 : * locality related effects. For example, it's possible that the total number
2812 : * of table blocks (pointed to by all TIDs on the leaf page) is naturally
2813 : * quite low, in which case we might end up checking if it's possible to
2814 : * delete _most_ index tuples on the page (without the tableam needing to
2815 : * access additional table blocks). The tableam will sometimes stumble upon
2816 : * _many_ extra deletable index tuples in indexes where this pattern is
2817 : * common.
2818 : *
2819 : * See nbtree/README for further details on simple index tuple deletion.
2820 : */
2821 : static void
2822 7450 : _bt_simpledel_pass(Relation rel, Buffer buffer, Relation heapRel,
2823 : OffsetNumber *deletable, int ndeletable, IndexTuple newitem,
2824 : OffsetNumber minoff, OffsetNumber maxoff)
2825 : {
2826 7450 : Page page = BufferGetPage(buffer);
2827 : BlockNumber *deadblocks;
2828 : int ndeadblocks;
2829 : TM_IndexDeleteOp delstate;
2830 : OffsetNumber offnum;
2831 :
2832 : /* Get array of table blocks pointed to by LP_DEAD-set tuples */
2833 7450 : deadblocks = _bt_deadblocks(page, deletable, ndeletable, newitem,
2834 : &ndeadblocks);
2835 :
2836 : /* Initialize tableam state that describes index deletion operation */
2837 7450 : delstate.irel = rel;
2838 7450 : delstate.iblknum = BufferGetBlockNumber(buffer);
2839 7450 : delstate.bottomup = false;
2840 7450 : delstate.bottomupfreespace = 0;
2841 7450 : delstate.ndeltids = 0;
2842 7450 : delstate.deltids = palloc(MaxTIDsPerBTreePage * sizeof(TM_IndexDelete));
2843 7450 : delstate.status = palloc(MaxTIDsPerBTreePage * sizeof(TM_IndexStatus));
2844 :
2845 7450 : for (offnum = minoff;
2846 2119750 : offnum <= maxoff;
2847 2112300 : offnum = OffsetNumberNext(offnum))
2848 : {
2849 2112300 : ItemId itemid = PageGetItemId(page, offnum);
2850 2112300 : IndexTuple itup = (IndexTuple) PageGetItem(page, itemid);
2851 2112300 : TM_IndexDelete *odeltid = &delstate.deltids[delstate.ndeltids];
2852 2112300 : TM_IndexStatus *ostatus = &delstate.status[delstate.ndeltids];
2853 : BlockNumber tidblock;
2854 : void *match;
2855 :
2856 2112300 : if (!BTreeTupleIsPosting(itup))
2857 : {
2858 2019334 : tidblock = ItemPointerGetBlockNumber(&itup->t_tid);
2859 2019334 : match = bsearch(&tidblock, deadblocks, ndeadblocks,
2860 : sizeof(BlockNumber), _bt_blk_cmp);
2861 :
2862 2019334 : if (!match)
2863 : {
2864 : Assert(!ItemIdIsDead(itemid));
2865 1277164 : continue;
2866 : }
2867 :
2868 : /*
2869 : * TID's table block is among those pointed to by the TIDs from
2870 : * LP_DEAD-bit set tuples on page -- add TID to deltids
2871 : */
2872 742170 : odeltid->tid = itup->t_tid;
2873 742170 : odeltid->id = delstate.ndeltids;
2874 742170 : ostatus->idxoffnum = offnum;
2875 742170 : ostatus->knowndeletable = ItemIdIsDead(itemid);
2876 742170 : ostatus->promising = false; /* unused */
2877 742170 : ostatus->freespace = 0; /* unused */
2878 :
2879 742170 : delstate.ndeltids++;
2880 : }
2881 : else
2882 : {
2883 92966 : int nitem = BTreeTupleGetNPosting(itup);
2884 :
2885 474768 : for (int p = 0; p < nitem; p++)
2886 : {
2887 381802 : ItemPointer tid = BTreeTupleGetPostingN(itup, p);
2888 :
2889 381802 : tidblock = ItemPointerGetBlockNumber(tid);
2890 381802 : match = bsearch(&tidblock, deadblocks, ndeadblocks,
2891 : sizeof(BlockNumber), _bt_blk_cmp);
2892 :
2893 381802 : if (!match)
2894 : {
2895 : Assert(!ItemIdIsDead(itemid));
2896 337686 : continue;
2897 : }
2898 :
2899 : /*
2900 : * TID's table block is among those pointed to by the TIDs
2901 : * from LP_DEAD-bit set tuples on page -- add TID to deltids
2902 : */
2903 44116 : odeltid->tid = *tid;
2904 44116 : odeltid->id = delstate.ndeltids;
2905 44116 : ostatus->idxoffnum = offnum;
2906 44116 : ostatus->knowndeletable = ItemIdIsDead(itemid);
2907 44116 : ostatus->promising = false; /* unused */
2908 44116 : ostatus->freespace = 0; /* unused */
2909 :
2910 44116 : odeltid++;
2911 44116 : ostatus++;
2912 44116 : delstate.ndeltids++;
2913 : }
2914 : }
2915 : }
2916 :
2917 7450 : pfree(deadblocks);
2918 :
2919 : Assert(delstate.ndeltids >= ndeletable);
2920 :
2921 : /* Physically delete LP_DEAD tuples (plus any delete-safe extra TIDs) */
2922 7450 : _bt_delitems_delete_check(rel, buffer, heapRel, &delstate);
2923 :
2924 7450 : pfree(delstate.deltids);
2925 7450 : pfree(delstate.status);
2926 7450 : }
2927 :
2928 : /*
2929 : * _bt_deadblocks() -- Get LP_DEAD related table blocks.
2930 : *
2931 : * Builds sorted and unique-ified array of table block numbers from index
2932 : * tuple TIDs whose line pointers are marked LP_DEAD. Also adds the table
2933 : * block from incoming newitem just in case it isn't among the LP_DEAD-related
2934 : * table blocks.
2935 : *
2936 : * Always counting the newitem's table block as an LP_DEAD related block makes
2937 : * sense because the cost is consistently low; it is practically certain that
2938 : * the table block will not incur a buffer miss in tableam. On the other hand
2939 : * the benefit is often quite high. There is a decent chance that there will
2940 : * be some deletable items from this block, since in general most garbage
2941 : * tuples became garbage in the recent past (in many cases this won't be the
2942 : * first logical row that core code added to/modified in table block
2943 : * recently).
2944 : *
2945 : * Returns final array, and sets *nblocks to its final size for caller.
2946 : */
2947 : static BlockNumber *
2948 7450 : _bt_deadblocks(Page page, OffsetNumber *deletable, int ndeletable,
2949 : IndexTuple newitem, int *nblocks)
2950 : {
2951 : int spacentids,
2952 : ntids;
2953 : BlockNumber *tidblocks;
2954 :
2955 : /*
2956 : * Accumulate each TID's block in array whose initial size has space for
2957 : * one table block per LP_DEAD-set tuple (plus space for the newitem table
2958 : * block). Array will only need to grow when there are LP_DEAD-marked
2959 : * posting list tuples (which is not that common).
2960 : */
2961 7450 : spacentids = ndeletable + 1;
2962 7450 : ntids = 0;
2963 7450 : tidblocks = (BlockNumber *) palloc(sizeof(BlockNumber) * spacentids);
2964 :
2965 : /*
2966 : * First add the table block for the incoming newitem. This is the one
2967 : * case where simple deletion can visit a table block that doesn't have
2968 : * any known deletable items.
2969 : */
2970 : Assert(!BTreeTupleIsPosting(newitem) && !BTreeTupleIsPivot(newitem));
2971 7450 : tidblocks[ntids++] = ItemPointerGetBlockNumber(&newitem->t_tid);
2972 :
2973 259156 : for (int i = 0; i < ndeletable; i++)
2974 : {
2975 251706 : ItemId itemid = PageGetItemId(page, deletable[i]);
2976 251706 : IndexTuple itup = (IndexTuple) PageGetItem(page, itemid);
2977 :
2978 : Assert(ItemIdIsDead(itemid));
2979 :
2980 251706 : if (!BTreeTupleIsPosting(itup))
2981 : {
2982 243164 : if (ntids + 1 > spacentids)
2983 : {
2984 212 : spacentids *= 2;
2985 : tidblocks = (BlockNumber *)
2986 212 : repalloc(tidblocks, sizeof(BlockNumber) * spacentids);
2987 : }
2988 :
2989 243164 : tidblocks[ntids++] = ItemPointerGetBlockNumber(&itup->t_tid);
2990 : }
2991 : else
2992 : {
2993 8542 : int nposting = BTreeTupleGetNPosting(itup);
2994 :
2995 8542 : if (ntids + nposting > spacentids)
2996 : {
2997 180 : spacentids = Max(spacentids * 2, ntids + nposting);
2998 : tidblocks = (BlockNumber *)
2999 180 : repalloc(tidblocks, sizeof(BlockNumber) * spacentids);
3000 : }
3001 :
3002 27878 : for (int j = 0; j < nposting; j++)
3003 : {
3004 19336 : ItemPointer tid = BTreeTupleGetPostingN(itup, j);
3005 :
3006 19336 : tidblocks[ntids++] = ItemPointerGetBlockNumber(tid);
3007 : }
3008 : }
3009 : }
3010 :
3011 7450 : qsort(tidblocks, ntids, sizeof(BlockNumber), _bt_blk_cmp);
3012 7450 : *nblocks = qunique(tidblocks, ntids, sizeof(BlockNumber), _bt_blk_cmp);
3013 :
3014 7450 : return tidblocks;
3015 : }
3016 :
3017 : /*
3018 : * _bt_blk_cmp() -- qsort comparison function for _bt_simpledel_pass
3019 : */
3020 : static inline int
3021 5314858 : _bt_blk_cmp(const void *arg1, const void *arg2)
3022 : {
3023 5314858 : BlockNumber b1 = *((BlockNumber *) arg1);
3024 5314858 : BlockNumber b2 = *((BlockNumber *) arg2);
3025 :
3026 5314858 : return pg_cmp_u32(b1, b2);
3027 : }
|
