Age Owner Branch data TLA Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * nbtutils.c
4 : : * Utility code for Postgres btree implementation.
5 : : *
6 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : : * Portions Copyright (c) 1994, Regents of the University of California
8 : : *
9 : : *
10 : : * IDENTIFICATION
11 : : * src/backend/access/nbtree/nbtutils.c
12 : : *
13 : : *-------------------------------------------------------------------------
14 : : */
15 : :
16 : : #include "postgres.h"
17 : :
18 : : #include <time.h>
19 : :
20 : : #include "access/nbtree.h"
21 : : #include "access/reloptions.h"
22 : : #include "access/relscan.h"
23 : : #include "commands/progress.h"
24 : : #include "common/int.h"
25 : : #include "lib/qunique.h"
26 : : #include "miscadmin.h"
27 : : #include "storage/lwlock.h"
28 : : #include "storage/subsystems.h"
29 : : #include "utils/datum.h"
30 : : #include "utils/lsyscache.h"
31 : : #include "utils/rel.h"
32 : :
33 : :
34 : : static int _bt_compare_int(const void *va, const void *vb);
35 : : static int _bt_keep_natts(Relation rel, IndexTuple lastleft,
36 : : IndexTuple firstright, BTScanInsert itup_key);
37 : :
38 : :
39 : : /*
40 : : * _bt_mkscankey
41 : : * Build an insertion scan key that contains comparison data from itup
42 : : * as well as comparator routines appropriate to the key datatypes.
43 : : *
44 : : * The result is intended for use with _bt_compare() and _bt_truncate().
45 : : * Callers that don't need to fill out the insertion scankey arguments
46 : : * (e.g. they use an ad-hoc comparison routine, or only need a scankey
47 : : * for _bt_truncate()) can pass a NULL index tuple. The scankey will
48 : : * be initialized as if an "all truncated" pivot tuple was passed
49 : : * instead.
50 : : *
51 : : * Note that we may occasionally have to share lock the metapage to
52 : : * determine whether or not the keys in the index are expected to be
53 : : * unique (i.e. if this is a "heapkeyspace" index). We assume a
54 : : * heapkeyspace index when caller passes a NULL tuple, allowing index
55 : : * build callers to avoid accessing the non-existent metapage. We
56 : : * also assume that the index is _not_ allequalimage when a NULL tuple
57 : : * is passed; CREATE INDEX callers call _bt_allequalimage() to set the
58 : : * field themselves.
59 : : */
60 : : BTScanInsert
1060 pg@bowt.ie 61 :GNC 8900057 : _bt_mkscankey(Relation rel, IndexTuple itup)
62 : : {
63 : : BTScanInsert key;
64 : : ScanKey skey;
65 : : TupleDesc itupdesc;
66 : : int indnkeyatts;
67 : : int16 *indoption;
68 : : int tupnatts;
69 : : int i;
70 : :
10108 bruce@momjian.us 71 : 8900057 : itupdesc = RelationGetDescr(rel);
2950 teodor@sigaev.ru 72 : 8900057 : indnkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
7056 tgl@sss.pgh.pa.us 73 : 8900057 : indoption = rel->rd_indoption;
2603 pg@bowt.ie 74 [ + + + + ]: 8900057 : tupnatts = itup ? BTreeTupleGetNAtts(itup, rel) : 0;
75 : :
76 [ - + ]: 8900057 : Assert(tupnatts <= IndexRelationGetNumberOfAttributes(rel));
77 : :
78 : : /*
79 : : * We'll execute search using scan key constructed on key columns.
80 : : * Truncated attributes and non-key attributes are omitted from the final
81 : : * scan key.
82 : : */
83 : 8900057 : key = palloc(offsetof(BTScanInsertData, scankeys) +
84 : 8900057 : sizeof(ScanKeyData) * indnkeyatts);
2260 85 [ + + ]: 8900057 : if (itup)
1060 86 : 8808795 : _bt_metaversion(rel, &key->heapkeyspace, &key->allequalimage);
87 : : else
88 : : {
89 : : /* Utility statement callers can set these fields themselves */
2260 90 : 91262 : key->heapkeyspace = true;
91 : 91262 : key->allequalimage = false;
92 : : }
2540 tgl@sss.pgh.pa.us 93 : 8900057 : key->anynullkeys = false; /* initial assumption */
879 pg@bowt.ie 94 : 8900057 : key->nextkey = false; /* usual case, required by btinsert */
95 : 8900057 : key->backward = false; /* usual case, required by btinsert */
2603 96 : 8900057 : key->keysz = Min(indnkeyatts, tupnatts);
97 [ + + ]: 8900057 : key->scantid = key->heapkeyspace && itup ?
98 [ + - ]: 17800114 : BTreeTupleGetHeapTID(itup) : NULL;
99 : 8900057 : skey = key->scankeys;
2950 teodor@sigaev.ru 100 [ + + ]: 23043119 : for (i = 0; i < indnkeyatts; i++)
101 : : {
102 : : FmgrInfo *procinfo;
103 : : Datum arg;
104 : : bool null;
105 : : int flags;
106 : :
107 : : /*
108 : : * We can use the cached (default) support procs since no cross-type
109 : : * comparison can be needed.
110 : : */
8977 tgl@sss.pgh.pa.us 111 : 14143062 : procinfo = index_getprocinfo(rel, i + 1, BTORDER_PROC);
112 : :
113 : : /*
114 : : * Key arguments built from truncated attributes (or when caller
115 : : * provides no tuple) are defensively represented as NULL values. They
116 : : * should never be used.
117 : : */
2603 pg@bowt.ie 118 [ + + ]: 14143062 : if (i < tupnatts)
119 : 13981584 : arg = index_getattr(itup, i + 1, itupdesc, &null);
120 : : else
121 : : {
122 : 161478 : arg = (Datum) 0;
123 : 161478 : null = true;
124 : : }
125 : 14143062 : flags = (null ? SK_ISNULL : 0) | (indoption[i] << SK_BT_INDOPTION_SHIFT);
8977 tgl@sss.pgh.pa.us 126 : 14143062 : ScanKeyEntryInitializeWithInfo(&skey[i],
127 : : flags,
128 : 14143062 : (AttrNumber) (i + 1),
129 : : InvalidStrategy,
130 : : InvalidOid,
5502 131 : 14143062 : rel->rd_indcollation[i],
132 : : procinfo,
133 : : arg);
134 : : /* Record if any key attribute is NULL (or truncated) */
2569 pg@bowt.ie 135 [ + + ]: 14143062 : if (null)
136 : 171847 : key->anynullkeys = true;
137 : : }
138 : :
139 : : /*
140 : : * In NULLS NOT DISTINCT mode, we pretend that there are no null keys, so
141 : : * that full uniqueness check is done.
142 : : */
1552 peter@eisentraut.org 143 [ + + ]: 8900057 : if (rel->rd_index->indnullsnotdistinct)
144 : 124 : key->anynullkeys = false;
145 : :
2603 pg@bowt.ie 146 : 8900057 : return key;
147 : : }
148 : :
149 : : /*
150 : : * qsort comparison function for int arrays
151 : : */
152 : : static int
146 153 : 365793 : _bt_compare_int(const void *va, const void *vb)
154 : : {
155 : 365793 : int a = *((const int *) va);
156 : 365793 : int b = *((const int *) vb);
157 : :
158 : 365793 : return pg_cmp_s32(a, b);
159 : : }
160 : :
161 : : /*
162 : : * _bt_killitems - set LP_DEAD state for items an indexscan caller has
163 : : * told us were killed
164 : : *
165 : : * scan->opaque, referenced locally through so, contains information about the
166 : : * current page and killed tuples thereon (generally, this should only be
167 : : * called if so->numKilled > 0).
168 : : *
169 : : * Caller should not have a lock on the so->currPos page, but must hold a
170 : : * buffer pin when !so->dropPin. When we return, it still won't be locked.
171 : : * It'll continue to hold whatever pins were held before calling here.
172 : : *
173 : : * We match items by heap TID before assuming they are the right ones to set
174 : : * LP_DEAD. If the scan is one that holds a buffer pin on the target page
175 : : * continuously from initially reading the items until applying this function
176 : : * (if it is a !so->dropPin scan), VACUUM cannot have deleted any items on the
177 : : * page, so the page's TIDs can't have been recycled by now. There's no risk
178 : : * that we'll confuse a new index tuple that happens to use a recycled TID
179 : : * with a now-removed tuple with the same TID (that used to be on this same
180 : : * page). We can't rely on that during scans that drop buffer pins eagerly
181 : : * (so->dropPin scans), though, so we must condition setting LP_DEAD bits on
182 : : * the page LSN having not changed since back when _bt_readpage saw the page.
183 : : * We totally give up on setting LP_DEAD bits when the page LSN changed.
184 : : *
185 : : * We give up much less often during !so->dropPin scans, but it still happens.
186 : : * We cope with cases where items have moved right due to insertions. If an
187 : : * item has moved off the current page due to a split, we'll fail to find it
188 : : * and just give up on it.
189 : : */
190 : : void
4059 kgrittn@postgresql.o 191 :CBC 105958 : _bt_killitems(IndexScanDesc scan)
192 : : {
333 pg@bowt.ie 193 : 105958 : Relation rel = scan->indexRelation;
7303 tgl@sss.pgh.pa.us 194 : 105958 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
195 : : Page page;
196 : : BTPageOpaque opaque;
197 : : OffsetNumber minoff;
198 : : OffsetNumber maxoff;
4059 kgrittn@postgresql.o 199 : 105958 : int numKilled = so->numKilled;
7303 tgl@sss.pgh.pa.us 200 : 105958 : bool killedsomething = false;
201 : : Buffer buf;
202 : :
333 pg@bowt.ie 203 [ - + ]: 105958 : Assert(numKilled > 0);
4059 kgrittn@postgresql.o 204 [ - + - - : 105958 : Assert(BTScanPosIsValid(so->currPos));
- + ]
333 pg@bowt.ie 205 [ - + ]: 105958 : Assert(scan->heapRelation != NULL); /* can't be a bitmap index scan */
206 : :
207 : : /* Always invalidate so->killedItems[] before leaving so->currPos */
4059 kgrittn@postgresql.o 208 : 105958 : so->numKilled = 0;
209 : :
210 : : /*
211 : : * We need to iterate through so->killedItems[] in leaf page order; the
212 : : * loop below expects this (when marking posting list tuples, at least).
213 : : * so->killedItems[] is now in whatever order the scan returned items in.
214 : : * Scrollable cursor scans might have even saved the same item/TID twice.
215 : : *
216 : : * Sort and unique-ify so->killedItems[] to deal with all this.
217 : : */
146 pg@bowt.ie 218 [ + + ]:GNC 105958 : if (numKilled > 1)
219 : : {
220 : 11276 : qsort(so->killedItems, numKilled, sizeof(int), _bt_compare_int);
221 : 11276 : numKilled = qunique(so->killedItems, numKilled, sizeof(int),
222 : : _bt_compare_int);
223 : : }
224 : :
333 pg@bowt.ie 225 [ + + ]:CBC 105958 : if (!so->dropPin)
226 : : {
227 : : /*
228 : : * We have held the pin on this page since we read the index tuples,
229 : : * so all we need to do is lock it. The pin will have prevented
230 : : * concurrent VACUUMs from recycling any of the TIDs on the page.
231 : : */
232 [ - + - - : 20298 : Assert(BTScanPosIsPinned(so->currPos));
- + ]
328 233 : 20298 : buf = so->currPos.buf;
234 : 20298 : _bt_lockbuf(rel, buf, BT_READ);
235 : : }
236 : : else
237 : : {
238 : : XLogRecPtr latestlsn;
239 : :
333 240 [ - + - - : 85660 : Assert(!BTScanPosIsPinned(so->currPos));
- + ]
241 : 85660 : buf = _bt_getbuf(rel, so->currPos.currPage, BT_READ);
242 : :
243 : 85660 : latestlsn = BufferGetLSNAtomic(buf);
244 [ - + ]: 85660 : Assert(so->currPos.lsn <= latestlsn);
245 [ + + ]: 85660 : if (so->currPos.lsn != latestlsn)
246 : : {
247 : : /* Modified, give up on hinting */
248 : 95 : _bt_relbuf(rel, buf);
4059 kgrittn@postgresql.o 249 : 95 : return;
250 : : }
251 : :
252 : : /* Unmodified, hinting is safe */
253 : : }
254 : :
328 pg@bowt.ie 255 : 105863 : page = BufferGetPage(buf);
1495 michael@paquier.xyz 256 : 105863 : opaque = BTPageGetOpaque(page);
7303 tgl@sss.pgh.pa.us 257 [ + + ]: 105863 : minoff = P_FIRSTDATAKEY(opaque);
258 : 105863 : maxoff = PageGetMaxOffsetNumber(page);
259 : :
260 : : /* Iterate through so->killedItems[] in leaf page order */
333 pg@bowt.ie 261 [ + + ]: 363872 : for (int i = 0; i < numKilled; i++)
262 : : {
7153 bruce@momjian.us 263 : 258012 : int itemIndex = so->killedItems[i];
264 : 258012 : BTScanPosItem *kitem = &so->currPos.items[itemIndex];
265 : 258012 : OffsetNumber offnum = kitem->indexOffset;
266 : :
7303 tgl@sss.pgh.pa.us 267 [ + - - + ]: 258012 : Assert(itemIndex >= so->currPos.firstItem &&
268 : : itemIndex <= so->currPos.lastItem);
146 pg@bowt.ie 269 [ + + - + ]:GNC 258012 : Assert(i == 0 ||
270 : : offnum >= so->currPos.items[so->killedItems[i - 1]].indexOffset);
271 : :
7303 tgl@sss.pgh.pa.us 272 [ - + ]:CBC 258012 : if (offnum < minoff)
7303 tgl@sss.pgh.pa.us 273 :UBC 0 : continue; /* pure paranoia */
7303 tgl@sss.pgh.pa.us 274 [ + + ]:CBC 6619074 : while (offnum <= maxoff)
275 : : {
276 : 6576497 : ItemId iid = PageGetItemId(page, offnum);
277 : 6576497 : IndexTuple ituple = (IndexTuple) PageGetItem(page, iid);
2260 pg@bowt.ie 278 : 6576497 : bool killtuple = false;
279 : :
280 [ + + ]: 6576497 : if (BTreeTupleIsPosting(ituple))
281 : : {
282 : 1093463 : int pi = i + 1;
283 : 1093463 : int nposting = BTreeTupleGetNPosting(ituple);
284 : : int j;
285 : :
286 : : /*
287 : : * Note that the page may have been modified in almost any way
288 : : * since we first read it (in the !so->dropPin case), so it's
289 : : * possible that this posting list tuple wasn't a posting list
290 : : * tuple when we first encountered its heap TIDs.
291 : : */
292 [ + + ]: 1132822 : for (j = 0; j < nposting; j++)
293 : : {
294 : 1131724 : ItemPointer item = BTreeTupleGetPostingN(ituple, j);
295 : :
296 [ + + ]: 1131724 : if (!ItemPointerEquals(item, &kitem->heapTid))
297 : 1092365 : break; /* out of posting list loop */
298 : :
299 : : /*
300 : : * kitem must have matching offnum when heap TIDs match,
301 : : * though only in the common case where the page can't
302 : : * have been concurrently modified
303 : : */
333 304 [ - + - - ]: 39359 : Assert(kitem->indexOffset == offnum || !so->dropPin);
305 : :
306 : : /*
307 : : * Read-ahead to later kitems here.
308 : : *
309 : : * We rely on the assumption that not advancing kitem here
310 : : * will prevent us from considering the posting list tuple
311 : : * fully dead by not matching its next heap TID in next
312 : : * loop iteration.
313 : : *
314 : : * If, on the other hand, this is the final heap TID in
315 : : * the posting list tuple, then tuple gets killed
316 : : * regardless (i.e. we handle the case where the last
317 : : * kitem is also the last heap TID in the last index tuple
318 : : * correctly -- posting tuple still gets killed).
319 : : */
2260 320 [ + + ]: 39359 : if (pi < numKilled)
321 : 14171 : kitem = &so->currPos.items[so->killedItems[pi++]];
322 : : }
323 : :
324 : : /*
325 : : * Don't bother advancing the outermost loop's int iterator to
326 : : * avoid processing killed items that relate to the same
327 : : * offnum/posting list tuple. This micro-optimization hardly
328 : : * seems worth it. (Further iterations of the outermost loop
329 : : * will fail to match on this same posting list's first heap
330 : : * TID instead, so we'll advance to the next offnum/index
331 : : * tuple pretty quickly.)
332 : : */
333 [ + + ]: 1093463 : if (j == nposting)
334 : 1098 : killtuple = true;
335 : : }
336 [ + + ]: 5483034 : else if (ItemPointerEquals(&ituple->t_tid, &kitem->heapTid))
337 : 214753 : killtuple = true;
338 : :
339 : : /*
340 : : * Mark index item as dead, if it isn't already. Since this
341 : : * happens while holding a buffer lock possibly in shared mode,
342 : : * it's possible that multiple processes attempt to do this
343 : : * simultaneously, leading to multiple full-page images being sent
344 : : * to WAL (if wal_log_hints or data checksums are enabled), which
345 : : * is undesirable.
346 : : */
2181 alvherre@alvh.no-ip. 347 [ + + + + ]: 6576497 : if (killtuple && !ItemIdIsDead(iid))
348 : : {
56 andres@anarazel.de 349 [ + + ]:GNC 215435 : if (!killedsomething)
350 : : {
351 : : /*
352 : : * Use the hint bit infrastructure to check if we can
353 : : * update the page while just holding a share lock. If we
354 : : * are not allowed, there's no point continuing.
355 : : */
356 [ + + ]: 79824 : if (!BufferBeginSetHintBits(buf))
357 : 3 : goto unlock_page;
358 : : }
359 : :
360 : : /* found the item/all posting list items */
6810 tgl@sss.pgh.pa.us 361 :CBC 215432 : ItemIdMarkDead(iid);
7303 362 : 215432 : killedsomething = true;
363 : 215432 : break; /* out of inner search loop */
364 : : }
365 : 6361062 : offnum = OffsetNumberNext(offnum);
366 : : }
367 : : }
368 : :
369 : : /*
370 : : * Since this can be redone later if needed, mark as dirty hint.
371 : : *
372 : : * Whenever we mark anything LP_DEAD, we also set the page's
373 : : * BTP_HAS_GARBAGE flag, which is likewise just a hint. (Note that we
374 : : * only rely on the page-level flag in !heapkeyspace indexes.)
375 : : */
376 [ + + ]: 105860 : if (killedsomething)
377 : : {
7224 378 : 79821 : opaque->btpo_flags |= BTP_HAS_GARBAGE;
56 andres@anarazel.de 379 :GNC 79821 : BufferFinishSetHintBits(buf, true, true);
380 : : }
381 : :
382 : 26039 : unlock_page:
328 pg@bowt.ie 383 [ + + ]:CBC 105863 : if (!so->dropPin)
384 : 20298 : _bt_unlockbuf(rel, buf);
385 : : else
386 : 85565 : _bt_relbuf(rel, buf);
387 : : }
388 : :
389 : :
390 : : /*
391 : : * The following routines manage a shared-memory area in which we track
392 : : * assignment of "vacuum cycle IDs" to currently-active btree vacuuming
393 : : * operations. There is a single counter which increments each time we
394 : : * start a vacuum to assign it a cycle ID. Since multiple vacuums could
395 : : * be active concurrently, we have to track the cycle ID for each active
396 : : * vacuum; this requires at most MaxBackends entries (usually far fewer).
397 : : * We assume at most one vacuum can be active for a given index.
398 : : *
399 : : * Access to the shared memory area is controlled by BtreeVacuumLock.
400 : : * In principle we could use a separate lmgr locktag for each index,
401 : : * but a single LWLock is much cheaper, and given the short time that
402 : : * the lock is ever held, the concurrency hit should be minimal.
403 : : */
404 : :
405 : : typedef struct BTOneVacInfo
406 : : {
407 : : LockRelId relid; /* global identifier of an index */
408 : : BTCycleId cycleid; /* cycle ID for its active VACUUM */
409 : : } BTOneVacInfo;
410 : :
411 : : typedef struct BTVacInfo
412 : : {
413 : : BTCycleId cycle_ctr; /* cycle ID most recently assigned */
414 : : int num_vacuums; /* number of currently active VACUUMs */
415 : : int max_vacuums; /* allocated length of vacuums[] array */
416 : : BTOneVacInfo vacuums[FLEXIBLE_ARRAY_MEMBER];
417 : : } BTVacInfo;
418 : :
419 : : static BTVacInfo *btvacinfo;
420 : :
421 : : static void BTreeShmemRequest(void *arg);
422 : : static void BTreeShmemInit(void *arg);
423 : :
424 : : const ShmemCallbacks BTreeShmemCallbacks = {
425 : : .request_fn = BTreeShmemRequest,
426 : : .init_fn = BTreeShmemInit,
427 : : };
428 : :
429 : : /*
430 : : * _bt_vacuum_cycleid --- get the active vacuum cycle ID for an index,
431 : : * or zero if there is no active VACUUM
432 : : *
433 : : * Note: for correct interlocking, the caller must already hold pin and
434 : : * exclusive lock on each buffer it will store the cycle ID into. This
435 : : * ensures that even if a VACUUM starts immediately afterwards, it cannot
436 : : * process those pages until the page split is complete.
437 : : */
438 : : BTCycleId
7302 tgl@sss.pgh.pa.us 439 : 16321 : _bt_vacuum_cycleid(Relation rel)
440 : : {
441 : 16321 : BTCycleId result = 0;
442 : : int i;
443 : :
444 : : /* Share lock is enough since this is a read-only operation */
445 : 16321 : LWLockAcquire(BtreeVacuumLock, LW_SHARED);
446 : :
447 [ + + ]: 16328 : for (i = 0; i < btvacinfo->num_vacuums; i++)
448 : : {
449 : 8 : BTOneVacInfo *vac = &btvacinfo->vacuums[i];
450 : :
451 [ + + ]: 8 : if (vac->relid.relId == rel->rd_lockInfo.lockRelId.relId &&
452 [ + - ]: 1 : vac->relid.dbId == rel->rd_lockInfo.lockRelId.dbId)
453 : : {
454 : 1 : result = vac->cycleid;
455 : 1 : break;
456 : : }
457 : : }
458 : :
459 : 16321 : LWLockRelease(BtreeVacuumLock);
460 : 16321 : return result;
461 : : }
462 : :
463 : : /*
464 : : * _bt_start_vacuum --- assign a cycle ID to a just-starting VACUUM operation
465 : : *
466 : : * Note: the caller must guarantee that it will eventually call
467 : : * _bt_end_vacuum, else we'll permanently leak an array slot. To ensure
468 : : * that this happens even in elog(FATAL) scenarios, the appropriate coding
469 : : * is not just a PG_TRY, but
470 : : * PG_ENSURE_ERROR_CLEANUP(_bt_end_vacuum_callback, PointerGetDatum(rel))
471 : : */
472 : : BTCycleId
473 : 2056 : _bt_start_vacuum(Relation rel)
474 : : {
475 : : BTCycleId result;
476 : : int i;
477 : : BTOneVacInfo *vac;
478 : :
479 : 2056 : LWLockAcquire(BtreeVacuumLock, LW_EXCLUSIVE);
480 : :
481 : : /*
482 : : * Assign the next cycle ID, being careful to avoid zero as well as the
483 : : * reserved high values.
484 : : */
6966 485 : 2056 : result = ++(btvacinfo->cycle_ctr);
486 [ + - - + ]: 2056 : if (result == 0 || result > MAX_BT_CYCLE_ID)
6966 tgl@sss.pgh.pa.us 487 :UBC 0 : result = btvacinfo->cycle_ctr = 1;
488 : :
489 : : /* Let's just make sure there's no entry already for this index */
7302 tgl@sss.pgh.pa.us 490 [ + + ]:CBC 2064 : for (i = 0; i < btvacinfo->num_vacuums; i++)
491 : : {
7302 tgl@sss.pgh.pa.us 492 :GBC 8 : vac = &btvacinfo->vacuums[i];
493 [ - + ]: 8 : if (vac->relid.relId == rel->rd_lockInfo.lockRelId.relId &&
7302 tgl@sss.pgh.pa.us 494 [ # # ]:UBC 0 : vac->relid.dbId == rel->rd_lockInfo.lockRelId.dbId)
495 : : {
496 : : /*
497 : : * Unlike most places in the backend, we have to explicitly
498 : : * release our LWLock before throwing an error. This is because
499 : : * we expect _bt_end_vacuum() to be called before transaction
500 : : * abort cleanup can run to release LWLocks.
501 : : */
6976 502 : 0 : LWLockRelease(BtreeVacuumLock);
7302 503 [ # # ]: 0 : elog(ERROR, "multiple active vacuums for index \"%s\"",
504 : : RelationGetRelationName(rel));
505 : : }
506 : : }
507 : :
508 : : /* OK, add an entry */
7302 tgl@sss.pgh.pa.us 509 [ - + ]:CBC 2056 : if (btvacinfo->num_vacuums >= btvacinfo->max_vacuums)
510 : : {
6976 tgl@sss.pgh.pa.us 511 :UBC 0 : LWLockRelease(BtreeVacuumLock);
7302 512 [ # # ]: 0 : elog(ERROR, "out of btvacinfo slots");
513 : : }
7302 tgl@sss.pgh.pa.us 514 :CBC 2056 : vac = &btvacinfo->vacuums[btvacinfo->num_vacuums];
515 : 2056 : vac->relid = rel->rd_lockInfo.lockRelId;
516 : 2056 : vac->cycleid = result;
517 : 2056 : btvacinfo->num_vacuums++;
518 : :
519 : 2056 : LWLockRelease(BtreeVacuumLock);
520 : 2056 : return result;
521 : : }
522 : :
523 : : /*
524 : : * _bt_end_vacuum --- mark a btree VACUUM operation as done
525 : : *
526 : : * Note: this is deliberately coded not to complain if no entry is found;
527 : : * this allows the caller to put PG_TRY around the start_vacuum operation.
528 : : */
529 : : void
530 : 2056 : _bt_end_vacuum(Relation rel)
531 : : {
532 : : int i;
533 : :
534 : 2056 : LWLockAcquire(BtreeVacuumLock, LW_EXCLUSIVE);
535 : :
536 : : /* Find the array entry */
537 [ + - ]: 2059 : for (i = 0; i < btvacinfo->num_vacuums; i++)
538 : : {
539 : 2059 : BTOneVacInfo *vac = &btvacinfo->vacuums[i];
540 : :
541 [ + + ]: 2059 : if (vac->relid.relId == rel->rd_lockInfo.lockRelId.relId &&
542 [ + - ]: 2056 : vac->relid.dbId == rel->rd_lockInfo.lockRelId.dbId)
543 : : {
544 : : /* Remove it by shifting down the last entry */
545 : 2056 : *vac = btvacinfo->vacuums[btvacinfo->num_vacuums - 1];
546 : 2056 : btvacinfo->num_vacuums--;
547 : 2056 : break;
548 : : }
549 : : }
550 : :
551 : 2056 : LWLockRelease(BtreeVacuumLock);
552 : 2056 : }
553 : :
554 : : /*
555 : : * _bt_end_vacuum wrapped as an on_shmem_exit callback function
556 : : */
557 : : void
6593 tgl@sss.pgh.pa.us 558 :GBC 2 : _bt_end_vacuum_callback(int code, Datum arg)
559 : : {
560 : 2 : _bt_end_vacuum((Relation) DatumGetPointer(arg));
561 : 2 : }
562 : :
563 : : /*
564 : : * BTreeShmemRequest --- register this module's shared memory
565 : : */
566 : : static void
29 heikki.linnakangas@i 567 :GNC 1244 : BTreeShmemRequest(void *arg)
568 : : {
569 : : Size size;
570 : :
4092 tgl@sss.pgh.pa.us 571 :CBC 1244 : size = offsetof(BTVacInfo, vacuums);
1484 rhaas@postgresql.org 572 : 1244 : size = add_size(size, mul_size(MaxBackends, sizeof(BTOneVacInfo)));
573 : :
29 heikki.linnakangas@i 574 :GNC 1244 : ShmemRequestStruct(.name = "BTree Vacuum State",
575 : : .size = size,
576 : : .ptr = (void **) &btvacinfo,
577 : : );
7302 tgl@sss.pgh.pa.us 578 :GIC 1244 : }
579 : :
580 : : /*
581 : : * BTreeShmemInit --- initialize this module's shared memory
582 : : */
583 : : static void
29 heikki.linnakangas@i 584 :GNC 1241 : BTreeShmemInit(void *arg)
585 : : {
586 : : /*
587 : : * It doesn't really matter what the cycle counter starts at, but having
588 : : * it always start the same doesn't seem good. Seed with low-order bits
589 : : * of time() instead.
590 : : */
591 : 1241 : btvacinfo->cycle_ctr = (BTCycleId) time(NULL);
592 : :
593 : 1241 : btvacinfo->num_vacuums = 0;
594 : 1241 : btvacinfo->max_vacuums = MaxBackends;
7302 tgl@sss.pgh.pa.us 595 :CBC 1241 : }
596 : :
597 : : bytea *
3761 598 : 204 : btoptions(Datum reloptions, bool validate)
599 : : {
600 : : static const relopt_parse_elt tab[] = {
601 : : {"fillfactor", RELOPT_TYPE_INT, offsetof(BTOptions, fillfactor)},
602 : : {"vacuum_cleanup_index_scale_factor", RELOPT_TYPE_REAL,
603 : : offsetof(BTOptions, vacuum_cleanup_index_scale_factor)},
604 : : {"deduplicate_items", RELOPT_TYPE_BOOL,
605 : : offsetof(BTOptions, deduplicate_items)}
606 : : };
607 : :
2353 michael@paquier.xyz 608 : 204 : return (bytea *) build_reloptions(reloptions, validate,
609 : : RELOPT_KIND_BTREE,
610 : : sizeof(BTOptions),
611 : : tab, lengthof(tab));
612 : : }
613 : :
614 : : /*
615 : : * btproperty() -- Check boolean properties of indexes.
616 : : *
617 : : * This is optional, but handling AMPROP_RETURNABLE here saves opening the rel
618 : : * to call btcanreturn.
619 : : */
620 : : bool
3552 tgl@sss.pgh.pa.us 621 : 504 : btproperty(Oid index_oid, int attno,
622 : : IndexAMProperty prop, const char *propname,
623 : : bool *res, bool *isnull)
624 : : {
625 [ + + ]: 504 : switch (prop)
626 : : {
627 : 28 : case AMPROP_RETURNABLE:
628 : : /* answer only for columns, not AM or whole index */
629 [ + + ]: 28 : if (attno == 0)
630 : 8 : return false;
631 : : /* otherwise, btree can always return data */
632 : 20 : *res = true;
633 : 20 : return true;
634 : :
635 : 476 : default:
636 : 476 : return false; /* punt to generic code */
637 : : }
638 : : }
639 : :
640 : : /*
641 : : * btbuildphasename() -- Return name of index build phase.
642 : : */
643 : : char *
2590 alvherre@alvh.no-ip. 644 :UBC 0 : btbuildphasename(int64 phasenum)
645 : : {
646 [ # # # # : 0 : switch (phasenum)
# # ]
647 : : {
648 : 0 : case PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE:
649 : 0 : return "initializing";
650 : 0 : case PROGRESS_BTREE_PHASE_INDEXBUILD_TABLESCAN:
651 : 0 : return "scanning table";
652 : 0 : case PROGRESS_BTREE_PHASE_PERFORMSORT_1:
653 : 0 : return "sorting live tuples";
654 : 0 : case PROGRESS_BTREE_PHASE_PERFORMSORT_2:
655 : 0 : return "sorting dead tuples";
656 : 0 : case PROGRESS_BTREE_PHASE_LEAF_LOAD:
657 : 0 : return "loading tuples in tree";
658 : 0 : default:
659 : 0 : return NULL;
660 : : }
661 : : }
662 : :
663 : : /*
664 : : * _bt_truncate() -- create tuple without unneeded suffix attributes.
665 : : *
666 : : * Returns truncated pivot index tuple allocated in caller's memory context,
667 : : * with key attributes copied from caller's firstright argument. If rel is
668 : : * an INCLUDE index, non-key attributes will definitely be truncated away,
669 : : * since they're not part of the key space. More aggressive suffix
670 : : * truncation can take place when it's clear that the returned tuple does not
671 : : * need one or more suffix key attributes. We only need to keep firstright
672 : : * attributes up to and including the first non-lastleft-equal attribute.
673 : : * Caller's insertion scankey is used to compare the tuples; the scankey's
674 : : * argument values are not considered here.
675 : : *
676 : : * Note that returned tuple's t_tid offset will hold the number of attributes
677 : : * present, so the original item pointer offset is not represented. Caller
678 : : * should only change truncated tuple's downlink. Note also that truncated
679 : : * key attributes are treated as containing "minus infinity" values by
680 : : * _bt_compare().
681 : : *
682 : : * In the worst case (when a heap TID must be appended to distinguish lastleft
683 : : * from firstright), the size of the returned tuple is the size of firstright
684 : : * plus the size of an additional MAXALIGN()'d item pointer. This guarantee
685 : : * is important, since callers need to stay under the 1/3 of a page
686 : : * restriction on tuple size. If this routine is ever taught to truncate
687 : : * within an attribute/datum, it will need to avoid returning an enlarged
688 : : * tuple to caller when truncation + TOAST compression ends up enlarging the
689 : : * final datum.
690 : : */
691 : : IndexTuple
2603 pg@bowt.ie 692 :CBC 40547 : _bt_truncate(Relation rel, IndexTuple lastleft, IndexTuple firstright,
693 : : BTScanInsert itup_key)
694 : : {
695 : 40547 : TupleDesc itupdesc = RelationGetDescr(rel);
696 : 40547 : int16 nkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
697 : : int keepnatts;
698 : : IndexTuple pivot;
699 : : IndexTuple tidpivot;
700 : : ItemPointer pivotheaptid;
701 : : Size newsize;
702 : :
703 : : /*
704 : : * We should only ever truncate non-pivot tuples from leaf pages. It's
705 : : * never okay to truncate when splitting an internal page.
706 : : */
2260 707 [ + - - + ]: 40547 : Assert(!BTreeTupleIsPivot(lastleft) && !BTreeTupleIsPivot(firstright));
708 : :
709 : : /* Determine how many attributes must be kept in truncated tuple */
2603 710 : 40547 : keepnatts = _bt_keep_natts(rel, lastleft, firstright, itup_key);
711 : :
712 : : #ifdef DEBUG_NO_TRUNCATE
713 : : /* Force truncation to be ineffective for testing purposes */
714 : : keepnatts = nkeyatts + 1;
715 : : #endif
716 : :
2227 717 : 40547 : pivot = index_truncate_tuple(itupdesc, firstright,
718 : : Min(keepnatts, nkeyatts));
719 : :
720 [ + + ]: 40547 : if (BTreeTupleIsPosting(pivot))
721 : : {
722 : : /*
723 : : * index_truncate_tuple() just returns a straight copy of firstright
724 : : * when it has no attributes to truncate. When that happens, we may
725 : : * need to truncate away a posting list here instead.
726 : : */
727 [ + + - + ]: 888 : Assert(keepnatts == nkeyatts || keepnatts == nkeyatts + 1);
728 [ - + ]: 888 : Assert(IndexRelationGetNumberOfAttributes(rel) == nkeyatts);
729 : 888 : pivot->t_info &= ~INDEX_SIZE_MASK;
730 : 888 : pivot->t_info |= MAXALIGN(BTreeTupleGetPostingOffset(firstright));
731 : : }
732 : :
733 : : /*
734 : : * If there is a distinguishing key attribute within pivot tuple, we're
735 : : * done
736 : : */
737 [ + + ]: 40547 : if (keepnatts <= nkeyatts)
738 : : {
2219 739 : 39733 : BTreeTupleSetNAtts(pivot, keepnatts, false);
2227 740 : 39733 : return pivot;
741 : : }
742 : :
743 : : /*
744 : : * We have to store a heap TID in the new pivot tuple, since no non-TID
745 : : * key attribute value in firstright distinguishes the right side of the
746 : : * split from the left side. nbtree conceptualizes this case as an
747 : : * inability to truncate away any key attributes, since heap TID is
748 : : * treated as just another key attribute (despite lacking a pg_attribute
749 : : * entry).
750 : : *
751 : : * Use enlarged space that holds a copy of pivot. We need the extra space
752 : : * to store a heap TID at the end (using the special pivot tuple
753 : : * representation). Note that the original pivot already has firstright's
754 : : * possible posting list/non-key attribute values removed at this point.
755 : : */
756 : 814 : newsize = MAXALIGN(IndexTupleSize(pivot)) + MAXALIGN(sizeof(ItemPointerData));
757 : 814 : tidpivot = palloc0(newsize);
758 : 814 : memcpy(tidpivot, pivot, MAXALIGN(IndexTupleSize(pivot)));
759 : : /* Cannot leak memory here */
760 : 814 : pfree(pivot);
761 : :
762 : : /*
763 : : * Store all of firstright's key attribute values plus a tiebreaker heap
764 : : * TID value in enlarged pivot tuple
765 : : */
766 : 814 : tidpivot->t_info &= ~INDEX_SIZE_MASK;
767 : 814 : tidpivot->t_info |= newsize;
2219 768 : 814 : BTreeTupleSetNAtts(tidpivot, nkeyatts, true);
2227 769 : 814 : pivotheaptid = BTreeTupleGetHeapTID(tidpivot);
770 : :
771 : : /*
772 : : * Lehman & Yao use lastleft as the leaf high key in all cases, but don't
773 : : * consider suffix truncation. It seems like a good idea to follow that
774 : : * example in cases where no truncation takes place -- use lastleft's heap
775 : : * TID. (This is also the closest value to negative infinity that's
776 : : * legally usable.)
777 : : */
2260 778 : 814 : ItemPointerCopy(BTreeTupleGetMaxHeapTID(lastleft), pivotheaptid);
779 : :
780 : : /*
781 : : * We're done. Assert() that heap TID invariants hold before returning.
782 : : *
783 : : * Lehman and Yao require that the downlink to the right page, which is to
784 : : * be inserted into the parent page in the second phase of a page split be
785 : : * a strict lower bound on items on the right page, and a non-strict upper
786 : : * bound for items on the left page. Assert that heap TIDs follow these
787 : : * invariants, since a heap TID value is apparently needed as a
788 : : * tiebreaker.
789 : : */
790 : : #ifndef DEBUG_NO_TRUNCATE
791 [ - + ]: 814 : Assert(ItemPointerCompare(BTreeTupleGetMaxHeapTID(lastleft),
792 : : BTreeTupleGetHeapTID(firstright)) < 0);
793 [ - + ]: 814 : Assert(ItemPointerCompare(pivotheaptid,
794 : : BTreeTupleGetHeapTID(lastleft)) >= 0);
795 [ - + ]: 814 : Assert(ItemPointerCompare(pivotheaptid,
796 : : BTreeTupleGetHeapTID(firstright)) < 0);
797 : : #else
798 : :
799 : : /*
800 : : * Those invariants aren't guaranteed to hold for lastleft + firstright
801 : : * heap TID attribute values when they're considered here only because
802 : : * DEBUG_NO_TRUNCATE is defined (a heap TID is probably not actually
803 : : * needed as a tiebreaker). DEBUG_NO_TRUNCATE must therefore use a heap
804 : : * TID value that always works as a strict lower bound for items to the
805 : : * right. In particular, it must avoid using firstright's leading key
806 : : * attribute values along with lastleft's heap TID value when lastleft's
807 : : * TID happens to be greater than firstright's TID.
808 : : */
809 : : ItemPointerCopy(BTreeTupleGetHeapTID(firstright), pivotheaptid);
810 : :
811 : : /*
812 : : * Pivot heap TID should never be fully equal to firstright. Note that
813 : : * the pivot heap TID will still end up equal to lastleft's heap TID when
814 : : * that's the only usable value.
815 : : */
816 : : ItemPointerSetOffsetNumber(pivotheaptid,
817 : : OffsetNumberPrev(ItemPointerGetOffsetNumber(pivotheaptid)));
818 : : Assert(ItemPointerCompare(pivotheaptid,
819 : : BTreeTupleGetHeapTID(firstright)) < 0);
820 : : #endif
821 : :
2227 822 : 814 : return tidpivot;
823 : : }
824 : :
825 : : /*
826 : : * _bt_keep_natts - how many key attributes to keep when truncating.
827 : : *
828 : : * Caller provides two tuples that enclose a split point. Caller's insertion
829 : : * scankey is used to compare the tuples; the scankey's argument values are
830 : : * not considered here.
831 : : *
832 : : * This can return a number of attributes that is one greater than the
833 : : * number of key attributes for the index relation. This indicates that the
834 : : * caller must use a heap TID as a unique-ifier in new pivot tuple.
835 : : */
836 : : static int
2603 837 : 40547 : _bt_keep_natts(Relation rel, IndexTuple lastleft, IndexTuple firstright,
838 : : BTScanInsert itup_key)
839 : : {
840 : 40547 : int nkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
841 : 40547 : TupleDesc itupdesc = RelationGetDescr(rel);
842 : : int keepnatts;
843 : : ScanKey scankey;
844 : :
845 : : /*
846 : : * _bt_compare() treats truncated key attributes as having the value minus
847 : : * infinity, which would break searches within !heapkeyspace indexes. We
848 : : * must still truncate away non-key attribute values, though.
849 : : */
850 [ - + ]: 40547 : if (!itup_key->heapkeyspace)
2603 pg@bowt.ie 851 :UBC 0 : return nkeyatts;
852 : :
2603 pg@bowt.ie 853 :CBC 40547 : scankey = itup_key->scankeys;
854 : 40547 : keepnatts = 1;
855 [ + + ]: 49233 : for (int attnum = 1; attnum <= nkeyatts; attnum++, scankey++)
856 : : {
857 : : Datum datum1,
858 : : datum2;
859 : : bool isNull1,
860 : : isNull2;
861 : :
862 : 48419 : datum1 = index_getattr(lastleft, attnum, itupdesc, &isNull1);
863 : 48419 : datum2 = index_getattr(firstright, attnum, itupdesc, &isNull2);
864 : :
865 [ - + ]: 48419 : if (isNull1 != isNull2)
866 : 39733 : break;
867 : :
868 [ + + + + ]: 96823 : if (!isNull1 &&
869 : 48404 : DatumGetInt32(FunctionCall2Coll(&scankey->sk_func,
870 : : scankey->sk_collation,
871 : : datum1,
872 : : datum2)) != 0)
873 : 39733 : break;
874 : :
875 : 8686 : keepnatts++;
876 : : }
877 : :
878 : : /*
879 : : * Assert that _bt_keep_natts_fast() agrees with us in passing. This is
880 : : * expected in an allequalimage index.
881 : : */
2260 882 [ + + - + ]: 40547 : Assert(!itup_key->allequalimage ||
883 : : keepnatts == _bt_keep_natts_fast(rel, lastleft, firstright));
884 : :
2603 885 : 40547 : return keepnatts;
886 : : }
887 : :
888 : : /*
889 : : * _bt_keep_natts_fast - fast bitwise variant of _bt_keep_natts.
890 : : *
891 : : * This is exported so that a candidate split point can have its effect on
892 : : * suffix truncation inexpensively evaluated ahead of time when finding a
893 : : * split location. A naive bitwise approach to datum comparisons is used to
894 : : * save cycles.
895 : : *
896 : : * The approach taken here usually provides the same answer as _bt_keep_natts
897 : : * will (for the same pair of tuples from a heapkeyspace index), since the
898 : : * majority of btree opclasses can never indicate that two datums are equal
899 : : * unless they're bitwise equal after detoasting. When an index only has
900 : : * "equal image" columns, routine is guaranteed to give the same result as
901 : : * _bt_keep_natts would.
902 : : *
903 : : * Callers can rely on the fact that attributes considered equal here are
904 : : * definitely also equal according to _bt_keep_natts, even when the index uses
905 : : * an opclass or collation that is not "allequalimage"/deduplication-safe.
906 : : * This weaker guarantee is good enough for nbtsplitloc.c caller, since false
907 : : * negatives generally only have the effect of making leaf page splits use a
908 : : * more balanced split point.
909 : : */
910 : : int
911 : 9840115 : _bt_keep_natts_fast(Relation rel, IndexTuple lastleft, IndexTuple firstright)
912 : : {
913 : 9840115 : TupleDesc itupdesc = RelationGetDescr(rel);
914 : 9840115 : int keysz = IndexRelationGetNumberOfKeyAttributes(rel);
915 : : int keepnatts;
916 : :
917 : 9840115 : keepnatts = 1;
918 [ + + ]: 16256746 : for (int attnum = 1; attnum <= keysz; attnum++)
919 : : {
920 : : Datum datum1,
921 : : datum2;
922 : : bool isNull1,
923 : : isNull2;
924 : : CompactAttribute *att;
925 : :
926 : 14497285 : datum1 = index_getattr(lastleft, attnum, itupdesc, &isNull1);
927 : 14497285 : datum2 = index_getattr(firstright, attnum, itupdesc, &isNull2);
501 drowley@postgresql.o 928 : 14497285 : att = TupleDescCompactAttr(itupdesc, attnum - 1);
929 : :
2603 pg@bowt.ie 930 [ + + ]: 14497285 : if (isNull1 != isNull2)
931 : 8080654 : break;
932 : :
933 [ + + ]: 14497149 : if (!isNull1 &&
2366 934 [ + + ]: 14472564 : !datum_image_eq(datum1, datum2, att->attbyval, att->attlen))
2603 935 : 8080518 : break;
936 : :
937 : 6416631 : keepnatts++;
938 : : }
939 : :
940 : 9840115 : return keepnatts;
941 : : }
942 : :
943 : : /*
944 : : * _bt_check_natts() -- Verify tuple has expected number of attributes.
945 : : *
946 : : * Returns value indicating if the expected number of attributes were found
947 : : * for a particular offset on page. This can be used as a general purpose
948 : : * sanity check.
949 : : *
950 : : * Testing a tuple directly with BTreeTupleGetNAtts() should generally be
951 : : * preferred to calling here. That's usually more convenient, and is always
952 : : * more explicit. Call here instead when offnum's tuple may be a negative
953 : : * infinity tuple that uses the pre-v11 on-disk representation, or when a low
954 : : * context check is appropriate. This routine is as strict as possible about
955 : : * what is expected on each version of btree.
956 : : */
957 : : bool
958 : 197635977 : _bt_check_natts(Relation rel, bool heapkeyspace, Page page, OffsetNumber offnum)
959 : : {
2931 tgl@sss.pgh.pa.us 960 : 197635977 : int16 natts = IndexRelationGetNumberOfAttributes(rel);
961 : 197635977 : int16 nkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
1495 michael@paquier.xyz 962 : 197635977 : BTPageOpaque opaque = BTPageGetOpaque(page);
963 : : IndexTuple itup;
964 : : int tupnatts;
965 : :
966 : : /*
967 : : * We cannot reliably test a deleted or half-dead page, since they have
968 : : * dummy high keys
969 : : */
2938 teodor@sigaev.ru 970 [ - + ]: 197635977 : if (P_IGNORE(opaque))
2938 teodor@sigaev.ru 971 :UBC 0 : return true;
972 : :
2938 teodor@sigaev.ru 973 [ + - - + ]:CBC 197635977 : Assert(offnum >= FirstOffsetNumber &&
974 : : offnum <= PageGetMaxOffsetNumber(page));
975 : :
976 : 197635977 : itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offnum));
2603 pg@bowt.ie 977 [ + + ]: 197635977 : tupnatts = BTreeTupleGetNAtts(itup, rel);
978 : :
979 : : /* !heapkeyspace indexes do not support deduplication */
2260 980 [ - + - - ]: 197635977 : if (!heapkeyspace && BTreeTupleIsPosting(itup))
2260 pg@bowt.ie 981 :UBC 0 : return false;
982 : :
983 : : /* Posting list tuples should never have "pivot heap TID" bit set */
2260 pg@bowt.ie 984 [ + + ]:CBC 197635977 : if (BTreeTupleIsPosting(itup) &&
985 [ - + ]: 2179968 : (ItemPointerGetOffsetNumberNoCheck(&itup->t_tid) &
986 : : BT_PIVOT_HEAP_TID_ATTR) != 0)
2260 pg@bowt.ie 987 :UBC 0 : return false;
988 : :
989 : : /* INCLUDE indexes do not support deduplication */
2260 pg@bowt.ie 990 [ + + - + ]:CBC 197635977 : if (natts != nkeyatts && BTreeTupleIsPosting(itup))
2260 pg@bowt.ie 991 :UBC 0 : return false;
992 : :
2938 teodor@sigaev.ru 993 [ + + ]:CBC 197635977 : if (P_ISLEAF(opaque))
994 : : {
995 [ + + + + ]: 140775864 : if (offnum >= P_FIRSTDATAKEY(opaque))
996 : : {
997 : : /*
998 : : * Non-pivot tuple should never be explicitly marked as a pivot
999 : : * tuple
1000 : : */
2260 pg@bowt.ie 1001 [ - + ]: 129724080 : if (BTreeTupleIsPivot(itup))
2603 pg@bowt.ie 1002 :UBC 0 : return false;
1003 : :
1004 : : /*
1005 : : * Leaf tuples that are not the page high key (non-pivot tuples)
1006 : : * should never be truncated. (Note that tupnatts must have been
1007 : : * inferred, even with a posting list tuple, because only pivot
1008 : : * tuples store tupnatts directly.)
1009 : : */
2603 pg@bowt.ie 1010 :CBC 129724080 : return tupnatts == natts;
1011 : : }
1012 : : else
1013 : : {
1014 : : /*
1015 : : * Rightmost page doesn't contain a page high key, so tuple was
1016 : : * checked above as ordinary leaf tuple
1017 : : */
2938 teodor@sigaev.ru 1018 [ - + ]: 11051784 : Assert(!P_RIGHTMOST(opaque));
1019 : :
1020 : : /*
1021 : : * !heapkeyspace high key tuple contains only key attributes. Note
1022 : : * that tupnatts will only have been explicitly represented in
1023 : : * !heapkeyspace indexes that happen to have non-key attributes.
1024 : : */
2603 pg@bowt.ie 1025 [ - + ]: 11051784 : if (!heapkeyspace)
2603 pg@bowt.ie 1026 :UBC 0 : return tupnatts == nkeyatts;
1027 : :
1028 : : /* Use generic heapkeyspace pivot tuple handling */
1029 : : }
1030 : : }
1031 : : else /* !P_ISLEAF(opaque) */
1032 : : {
2938 teodor@sigaev.ru 1033 [ + + + + ]:CBC 56860113 : if (offnum == P_FIRSTDATAKEY(opaque))
1034 : : {
1035 : : /*
1036 : : * The first tuple on any internal page (possibly the first after
1037 : : * its high key) is its negative infinity tuple. Negative
1038 : : * infinity tuples are always truncated to zero attributes. They
1039 : : * are a particular kind of pivot tuple.
1040 : : */
2603 pg@bowt.ie 1041 [ + - ]: 2320469 : if (heapkeyspace)
1042 : 2320469 : return tupnatts == 0;
1043 : :
1044 : : /*
1045 : : * The number of attributes won't be explicitly represented if the
1046 : : * negative infinity tuple was generated during a page split that
1047 : : * occurred with a version of Postgres before v11. There must be
1048 : : * a problem when there is an explicit representation that is
1049 : : * non-zero, or when there is no explicit representation and the
1050 : : * tuple is evidently not a pre-pg_upgrade tuple.
1051 : : *
1052 : : * Prior to v11, downlinks always had P_HIKEY as their offset.
1053 : : * Accept that as an alternative indication of a valid
1054 : : * !heapkeyspace negative infinity tuple.
1055 : : */
2603 pg@bowt.ie 1056 [ # # # # ]:UBC 0 : return tupnatts == 0 ||
2260 1057 : 0 : ItemPointerGetOffsetNumber(&(itup->t_tid)) == P_HIKEY;
1058 : : }
1059 : : else
1060 : : {
1061 : : /*
1062 : : * !heapkeyspace downlink tuple with separator key contains only
1063 : : * key attributes. Note that tupnatts will only have been
1064 : : * explicitly represented in !heapkeyspace indexes that happen to
1065 : : * have non-key attributes.
1066 : : */
2603 pg@bowt.ie 1067 [ - + ]:CBC 54539644 : if (!heapkeyspace)
2603 pg@bowt.ie 1068 :UBC 0 : return tupnatts == nkeyatts;
1069 : :
1070 : : /* Use generic heapkeyspace pivot tuple handling */
1071 : : }
1072 : : }
1073 : :
1074 : : /* Handle heapkeyspace pivot tuples (excluding minus infinity items) */
2603 pg@bowt.ie 1075 [ - + ]:CBC 65591428 : Assert(heapkeyspace);
1076 : :
1077 : : /*
1078 : : * Explicit representation of the number of attributes is mandatory with
1079 : : * heapkeyspace index pivot tuples, regardless of whether or not there are
1080 : : * non-key attributes.
1081 : : */
2260 1082 [ - + ]: 65591428 : if (!BTreeTupleIsPivot(itup))
2260 pg@bowt.ie 1083 :UBC 0 : return false;
1084 : :
1085 : : /* Pivot tuple should not use posting list representation (redundant) */
2260 pg@bowt.ie 1086 [ - + ]:CBC 65591428 : if (BTreeTupleIsPosting(itup))
2603 pg@bowt.ie 1087 :UBC 0 : return false;
1088 : :
1089 : : /*
1090 : : * Heap TID is a tiebreaker key attribute, so it cannot be untruncated
1091 : : * when any other key attribute is truncated
1092 : : */
2603 pg@bowt.ie 1093 [ + + - + ]:CBC 65591428 : if (BTreeTupleGetHeapTID(itup) != NULL && tupnatts != nkeyatts)
2603 pg@bowt.ie 1094 :UBC 0 : return false;
1095 : :
1096 : : /*
1097 : : * Pivot tuple must have at least one untruncated key attribute (minus
1098 : : * infinity pivot tuples are the only exception). Pivot tuples can never
1099 : : * represent that there is a value present for a key attribute that
1100 : : * exceeds pg_index.indnkeyatts for the index.
1101 : : */
2603 pg@bowt.ie 1102 [ + - + - ]:CBC 65591428 : return tupnatts > 0 && tupnatts <= nkeyatts;
1103 : : }
1104 : :
1105 : : /*
1106 : : *
1107 : : * _bt_check_third_page() -- check whether tuple fits on a btree page at all.
1108 : : *
1109 : : * We actually need to be able to fit three items on every page, so restrict
1110 : : * any one item to 1/3 the per-page available space. Note that itemsz should
1111 : : * not include the ItemId overhead.
1112 : : *
1113 : : * It might be useful to apply TOAST methods rather than throw an error here.
1114 : : * Using out of line storage would break assumptions made by suffix truncation
1115 : : * and by contrib/amcheck, though.
1116 : : */
1117 : : void
1118 : 176 : _bt_check_third_page(Relation rel, Relation heap, bool needheaptidspace,
1119 : : Page page, IndexTuple newtup)
1120 : : {
1121 : : Size itemsz;
1122 : : BTPageOpaque opaque;
1123 : :
1124 : 176 : itemsz = MAXALIGN(IndexTupleSize(newtup));
1125 : :
1126 : : /* Double check item size against limit */
420 1127 [ - + ]: 176 : if (itemsz <= BTMaxItemSize)
2603 pg@bowt.ie 1128 :UBC 0 : return;
1129 : :
1130 : : /*
1131 : : * Tuple is probably too large to fit on page, but it's possible that the
1132 : : * index uses version 2 or version 3, or that page is an internal page, in
1133 : : * which case a slightly higher limit applies.
1134 : : */
420 pg@bowt.ie 1135 [ + - + - ]:CBC 176 : if (!needheaptidspace && itemsz <= BTMaxItemSizeNoHeapTid)
2603 1136 : 176 : return;
1137 : :
1138 : : /*
1139 : : * Internal page insertions cannot fail here, because that would mean that
1140 : : * an earlier leaf level insertion that should have failed didn't
1141 : : */
1495 michael@paquier.xyz 1142 :UBC 0 : opaque = BTPageGetOpaque(page);
2603 pg@bowt.ie 1143 [ # # ]: 0 : if (!P_ISLEAF(opaque))
1144 [ # # ]: 0 : elog(ERROR, "cannot insert oversized tuple of size %zu on internal page of index \"%s\"",
1145 : : itemsz, RelationGetRelationName(rel));
1146 : :
1147 [ # # # # : 0 : ereport(ERROR,
# # ]
1148 : : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1149 : : errmsg("index row size %zu exceeds btree version %u maximum %zu for index \"%s\"",
1150 : : itemsz,
1151 : : needheaptidspace ? BTREE_VERSION : BTREE_NOVAC_VERSION,
1152 : : needheaptidspace ? BTMaxItemSize : BTMaxItemSizeNoHeapTid,
1153 : : RelationGetRelationName(rel)),
1154 : : errdetail("Index row references tuple (%u,%u) in relation \"%s\".",
1155 : : ItemPointerGetBlockNumber(BTreeTupleGetHeapTID(newtup)),
1156 : : ItemPointerGetOffsetNumber(BTreeTupleGetHeapTID(newtup)),
1157 : : RelationGetRelationName(heap)),
1158 : : errhint("Values larger than 1/3 of a buffer page cannot be indexed.\n"
1159 : : "Consider a function index of an MD5 hash of the value, "
1160 : : "or use full text indexing."),
1161 : : errtableconstraint(heap, RelationGetRelationName(rel))));
1162 : : }
1163 : :
1164 : : /*
1165 : : * Are all attributes in rel "equality is image equality" attributes?
1166 : : *
1167 : : * We use each attribute's BTEQUALIMAGE_PROC opclass procedure. If any
1168 : : * opclass either lacks a BTEQUALIMAGE_PROC procedure or returns false, we
1169 : : * return false; otherwise we return true.
1170 : : *
1171 : : * Returned boolean value is stored in index metapage during index builds.
1172 : : * Deduplication can only be used when we return true.
1173 : : */
1174 : : bool
2260 pg@bowt.ie 1175 :CBC 36699 : _bt_allequalimage(Relation rel, bool debugmessage)
1176 : : {
1177 : 36699 : bool allequalimage = true;
1178 : :
1179 : : /* INCLUDE indexes can never support deduplication */
1180 : 36699 : if (IndexRelationGetNumberOfAttributes(rel) !=
1181 [ + + ]: 36699 : IndexRelationGetNumberOfKeyAttributes(rel))
1182 : 166 : return false;
1183 : :
1184 [ + + ]: 95757 : for (int i = 0; i < IndexRelationGetNumberOfKeyAttributes(rel); i++)
1185 : : {
1186 : 59602 : Oid opfamily = rel->rd_opfamily[i];
1187 : 59602 : Oid opcintype = rel->rd_opcintype[i];
1188 : 59602 : Oid collation = rel->rd_indcollation[i];
1189 : : Oid equalimageproc;
1190 : :
1191 : 59602 : equalimageproc = get_opfamily_proc(opfamily, opcintype, opcintype,
1192 : : BTEQUALIMAGE_PROC);
1193 : :
1194 : : /*
1195 : : * If there is no BTEQUALIMAGE_PROC then deduplication is assumed to
1196 : : * be unsafe. Otherwise, actually call proc and see what it says.
1197 : : */
1198 [ + + ]: 59602 : if (!OidIsValid(equalimageproc) ||
1199 [ + + ]: 59253 : !DatumGetBool(OidFunctionCall1Coll(equalimageproc, collation,
1200 : : ObjectIdGetDatum(opcintype))))
1201 : : {
1202 : 378 : allequalimage = false;
1203 : 378 : break;
1204 : : }
1205 : : }
1206 : :
1207 [ + + ]: 36533 : if (debugmessage)
1208 : : {
1209 [ + + ]: 32201 : if (allequalimage)
1210 [ + + ]: 31823 : elog(DEBUG1, "index \"%s\" can safely use deduplication",
1211 : : RelationGetRelationName(rel));
1212 : : else
1213 [ - + ]: 378 : elog(DEBUG1, "index \"%s\" cannot use deduplication",
1214 : : RelationGetRelationName(rel));
1215 : : }
1216 : :
1217 : 36533 : return allequalimage;
1218 : : }
|
