Branch data Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * nbtpage.c
4 : : * BTree-specific page management code for the Postgres btree access
5 : : * method.
6 : : *
7 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
8 : : * Portions Copyright (c) 1994, Regents of the University of California
9 : : *
10 : : *
11 : : * IDENTIFICATION
12 : : * src/backend/access/nbtree/nbtpage.c
13 : : *
14 : : * NOTES
15 : : * Postgres btree pages look like ordinary relation pages. The opaque
16 : : * data at high addresses includes pointers to left and right siblings
17 : : * and flag data describing page state. The first page in a btree, page
18 : : * zero, is special -- it stores meta-information describing the tree.
19 : : * Pages one and higher store the actual tree data.
20 : : *
21 : : *-------------------------------------------------------------------------
22 : : */
23 : : #include "postgres.h"
24 : :
25 : : #include "access/nbtree.h"
26 : : #include "access/nbtxlog.h"
27 : : #include "access/tableam.h"
28 : : #include "access/transam.h"
29 : : #include "access/xlog.h"
30 : : #include "access/xloginsert.h"
31 : : #include "common/int.h"
32 : : #include "miscadmin.h"
33 : : #include "storage/indexfsm.h"
34 : : #include "storage/predicate.h"
35 : : #include "storage/procarray.h"
36 : : #include "utils/injection_point.h"
37 : : #include "utils/memdebug.h"
38 : : #include "utils/memutils.h"
39 : : #include "utils/snapmgr.h"
40 : :
41 : : static BTMetaPageData *_bt_getmeta(Relation rel, Buffer metabuf);
42 : : static void _bt_delitems_delete(Relation rel, Buffer buf,
43 : : TransactionId snapshotConflictHorizon,
44 : : bool isCatalogRel,
45 : : OffsetNumber *deletable, int ndeletable,
46 : : BTVacuumPosting *updatable, int nupdatable);
47 : : static char *_bt_delitems_update(BTVacuumPosting *updatable, int nupdatable,
48 : : OffsetNumber *updatedoffsets,
49 : : Size *updatedbuflen, bool needswal);
50 : : static bool _bt_mark_page_halfdead(Relation rel, Relation heaprel,
51 : : Buffer leafbuf, BTStack stack);
52 : : static bool _bt_unlink_halfdead_page(Relation rel, Buffer leafbuf,
53 : : BlockNumber scanblkno,
54 : : bool *rightsib_empty,
55 : : BTVacState *vstate);
56 : : static bool _bt_lock_subtree_parent(Relation rel, Relation heaprel,
57 : : BlockNumber child, BTStack stack,
58 : : Buffer *subtreeparent, OffsetNumber *poffset,
59 : : BlockNumber *topparent,
60 : : BlockNumber *topparentrightsib);
61 : : static void _bt_pendingfsm_add(BTVacState *vstate, BlockNumber target,
62 : : FullTransactionId safexid);
63 : :
64 : : /*
65 : : * _bt_initmetapage() -- Fill a page buffer with a correct metapage image
66 : : */
67 : : void
68 : 32936 : _bt_initmetapage(Page page, BlockNumber rootbknum, uint32 level,
69 : : bool allequalimage)
70 : : {
71 : : BTMetaPageData *metad;
72 : : BTPageOpaque metaopaque;
73 : :
74 : 32936 : _bt_pageinit(page, BLCKSZ);
75 : :
76 : 32936 : metad = BTPageGetMeta(page);
77 : 32936 : metad->btm_magic = BTREE_MAGIC;
78 : 32936 : metad->btm_version = BTREE_VERSION;
79 : 32936 : metad->btm_root = rootbknum;
80 : 32936 : metad->btm_level = level;
81 : 32936 : metad->btm_fastroot = rootbknum;
82 : 32936 : metad->btm_fastlevel = level;
83 : 32936 : metad->btm_last_cleanup_num_delpages = 0;
84 : 32936 : metad->btm_last_cleanup_num_heap_tuples = -1.0;
85 : 32936 : metad->btm_allequalimage = allequalimage;
86 : :
87 : 32936 : metaopaque = BTPageGetOpaque(page);
88 : 32936 : metaopaque->btpo_flags = BTP_META;
89 : :
90 : : /*
91 : : * Set pd_lower just past the end of the metadata. This is essential,
92 : : * because without doing so, metadata will be lost if xlog.c compresses
93 : : * the page.
94 : : */
95 : 32936 : ((PageHeader) page)->pd_lower =
96 : 32936 : ((char *) metad + sizeof(BTMetaPageData)) - (char *) page;
97 : 32936 : }
98 : :
99 : : /*
100 : : * _bt_upgrademetapage() -- Upgrade a meta-page from an old format to version
101 : : * 3, the last version that can be updated without broadly affecting
102 : : * on-disk compatibility. (A REINDEX is required to upgrade to v4.)
103 : : *
104 : : * This routine does purely in-memory image upgrade. Caller is
105 : : * responsible for locking, WAL-logging etc.
106 : : */
107 : : void
108 : 0 : _bt_upgrademetapage(Page page)
109 : : {
110 : : BTMetaPageData *metad;
111 : : BTPageOpaque metaopaque PG_USED_FOR_ASSERTS_ONLY;
112 : :
113 : 0 : metad = BTPageGetMeta(page);
114 : 0 : metaopaque = BTPageGetOpaque(page);
115 : :
116 : : /* It must be really a meta page of upgradable version */
117 : : Assert(metaopaque->btpo_flags & BTP_META);
118 : : Assert(metad->btm_version < BTREE_NOVAC_VERSION);
119 : : Assert(metad->btm_version >= BTREE_MIN_VERSION);
120 : :
121 : : /* Set version number and fill extra fields added into version 3 */
122 : 0 : metad->btm_version = BTREE_NOVAC_VERSION;
123 : 0 : metad->btm_last_cleanup_num_delpages = 0;
124 : 0 : metad->btm_last_cleanup_num_heap_tuples = -1.0;
125 : : /* Only a REINDEX can set this field */
126 : : Assert(!metad->btm_allequalimage);
127 : 0 : metad->btm_allequalimage = false;
128 : :
129 : : /* Adjust pd_lower (see _bt_initmetapage() for details) */
130 : 0 : ((PageHeader) page)->pd_lower =
131 : 0 : ((char *) metad + sizeof(BTMetaPageData)) - (char *) page;
132 : 0 : }
133 : :
134 : : /*
135 : : * Get metadata from share-locked buffer containing metapage, while performing
136 : : * standard sanity checks.
137 : : *
138 : : * Callers that cache data returned here in local cache should note that an
139 : : * on-the-fly upgrade using _bt_upgrademetapage() can change the version field
140 : : * and BTREE_NOVAC_VERSION specific fields without invalidating local cache.
141 : : */
142 : : static BTMetaPageData *
143 : 1143221 : _bt_getmeta(Relation rel, Buffer metabuf)
144 : : {
145 : : Page metapg;
146 : : BTPageOpaque metaopaque;
147 : : BTMetaPageData *metad;
148 : :
149 : 1143221 : metapg = BufferGetPage(metabuf);
150 : 1143221 : metaopaque = BTPageGetOpaque(metapg);
151 : 1143221 : metad = BTPageGetMeta(metapg);
152 : :
153 : : /* sanity-check the metapage */
154 [ + - ]: 1143221 : if (!P_ISMETA(metaopaque) ||
155 [ - + ]: 1143221 : metad->btm_magic != BTREE_MAGIC)
156 [ # # ]: 0 : ereport(ERROR,
157 : : (errcode(ERRCODE_INDEX_CORRUPTED),
158 : : errmsg("index \"%s\" is not a btree",
159 : : RelationGetRelationName(rel))));
160 : :
161 [ + - ]: 1143221 : if (metad->btm_version < BTREE_MIN_VERSION ||
162 [ - + ]: 1143221 : metad->btm_version > BTREE_VERSION)
163 [ # # ]: 0 : ereport(ERROR,
164 : : (errcode(ERRCODE_INDEX_CORRUPTED),
165 : : errmsg("version mismatch in index \"%s\": file version %d, "
166 : : "current version %d, minimal supported version %d",
167 : : RelationGetRelationName(rel),
168 : : metad->btm_version, BTREE_VERSION, BTREE_MIN_VERSION)));
169 : :
170 : 1143221 : return metad;
171 : : }
172 : :
173 : : /*
174 : : * _bt_vacuum_needs_cleanup() -- Checks if index needs cleanup
175 : : *
176 : : * Called by btvacuumcleanup when btbulkdelete was never called because no
177 : : * index tuples needed to be deleted.
178 : : */
179 : : bool
180 : 148317 : _bt_vacuum_needs_cleanup(Relation rel)
181 : : {
182 : : Buffer metabuf;
183 : : Page metapg;
184 : : BTMetaPageData *metad;
185 : : uint32 btm_version;
186 : : BlockNumber prev_num_delpages;
187 : :
188 : : /*
189 : : * Copy details from metapage to local variables quickly.
190 : : *
191 : : * Note that we deliberately avoid using cached version of metapage here.
192 : : */
193 : 148317 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_READ);
194 : 148317 : metapg = BufferGetPage(metabuf);
195 : 148317 : metad = BTPageGetMeta(metapg);
196 : 148317 : btm_version = metad->btm_version;
197 : :
198 [ - + ]: 148317 : if (btm_version < BTREE_NOVAC_VERSION)
199 : : {
200 : : /*
201 : : * Metapage needs to be dynamically upgraded to store fields that are
202 : : * only present when btm_version >= BTREE_NOVAC_VERSION
203 : : */
204 : 0 : _bt_relbuf(rel, metabuf);
205 : 0 : return true;
206 : : }
207 : :
208 : 148317 : prev_num_delpages = metad->btm_last_cleanup_num_delpages;
209 : 148317 : _bt_relbuf(rel, metabuf);
210 : :
211 : : /*
212 : : * Trigger cleanup in rare cases where prev_num_delpages exceeds 5% of the
213 : : * total size of the index. We can reasonably expect (though are not
214 : : * guaranteed) to be able to recycle this many pages if we decide to do a
215 : : * btvacuumscan call during the ongoing btvacuumcleanup. For further
216 : : * details see the nbtree/README section on placing deleted pages in the
217 : : * FSM.
218 : : */
219 [ + + ]: 148317 : if (prev_num_delpages > 0 &&
220 [ + - ]: 7 : prev_num_delpages > RelationGetNumberOfBlocks(rel) / 20)
221 : 7 : return true;
222 : :
223 : 148310 : return false;
224 : : }
225 : :
226 : : /*
227 : : * _bt_set_cleanup_info() -- Update metapage for btvacuumcleanup.
228 : : *
229 : : * Called at the end of btvacuumcleanup, when num_delpages value has been
230 : : * finalized.
231 : : */
232 : : void
233 : 1403 : _bt_set_cleanup_info(Relation rel, BlockNumber num_delpages)
234 : : {
235 : : Buffer metabuf;
236 : : Page metapg;
237 : : BTMetaPageData *metad;
238 : : XLogRecPtr recptr;
239 : :
240 : : /*
241 : : * On-disk compatibility note: The btm_last_cleanup_num_delpages metapage
242 : : * field started out as a TransactionId field called btm_oldest_btpo_xact.
243 : : * Both "versions" are just uint32 fields. It was convenient to repurpose
244 : : * the field when we began to use 64-bit XIDs in deleted pages.
245 : : *
246 : : * It's possible that a pg_upgrade'd database will contain an XID value in
247 : : * what is now recognized as the metapage's btm_last_cleanup_num_delpages
248 : : * field. _bt_vacuum_needs_cleanup() may even believe that this value
249 : : * indicates that there are lots of pages that it needs to recycle, when
250 : : * in reality there are only one or two. The worst that can happen is
251 : : * that there will be a call to btvacuumscan a little earlier, which will
252 : : * set btm_last_cleanup_num_delpages to a sane value when we're called.
253 : : *
254 : : * Note also that the metapage's btm_last_cleanup_num_heap_tuples field is
255 : : * no longer used as of PostgreSQL 14. We set it to -1.0 on rewrite, just
256 : : * to be consistent.
257 : : */
258 : 1403 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_READ);
259 : 1403 : metapg = BufferGetPage(metabuf);
260 : 1403 : metad = BTPageGetMeta(metapg);
261 : :
262 : : /* Don't miss chance to upgrade index/metapage when BTREE_MIN_VERSION */
263 [ + - ]: 1403 : if (metad->btm_version >= BTREE_NOVAC_VERSION &&
264 [ + + ]: 1403 : metad->btm_last_cleanup_num_delpages == num_delpages)
265 : : {
266 : : /* Usually means index continues to have num_delpages of 0 */
267 : 1319 : _bt_relbuf(rel, metabuf);
268 : 1319 : return;
269 : : }
270 : :
271 : : /* trade in our read lock for a write lock */
272 : 84 : _bt_unlockbuf(rel, metabuf);
273 : 84 : _bt_lockbuf(rel, metabuf, BT_WRITE);
274 : :
275 : 84 : START_CRIT_SECTION();
276 : :
277 : : /* upgrade meta-page if needed */
278 [ - + ]: 84 : if (metad->btm_version < BTREE_NOVAC_VERSION)
279 : 0 : _bt_upgrademetapage(metapg);
280 : :
281 : : /* update cleanup-related information */
282 : 84 : metad->btm_last_cleanup_num_delpages = num_delpages;
283 : 84 : metad->btm_last_cleanup_num_heap_tuples = -1.0;
284 : 84 : MarkBufferDirty(metabuf);
285 : :
286 : : /* write wal record if needed */
287 [ + - + + : 84 : if (RelationNeedsWAL(rel))
+ - + - ]
288 : 84 : {
289 : : xl_btree_metadata md;
290 : :
291 : 84 : XLogBeginInsert();
292 : 84 : XLogRegisterBuffer(0, metabuf, REGBUF_WILL_INIT | REGBUF_STANDARD);
293 : :
294 : : Assert(metad->btm_version >= BTREE_NOVAC_VERSION);
295 : 84 : md.version = metad->btm_version;
296 : 84 : md.root = metad->btm_root;
297 : 84 : md.level = metad->btm_level;
298 : 84 : md.fastroot = metad->btm_fastroot;
299 : 84 : md.fastlevel = metad->btm_fastlevel;
300 : 84 : md.last_cleanup_num_delpages = num_delpages;
301 : 84 : md.allequalimage = metad->btm_allequalimage;
302 : :
303 : 84 : XLogRegisterBufData(0, &md, sizeof(xl_btree_metadata));
304 : :
305 : 84 : recptr = XLogInsert(RM_BTREE_ID, XLOG_BTREE_META_CLEANUP);
306 : : }
307 : : else
308 : 0 : recptr = XLogGetFakeLSN(rel);
309 : :
310 : 84 : PageSetLSN(metapg, recptr);
311 : :
312 : 84 : END_CRIT_SECTION();
313 : :
314 : 84 : _bt_relbuf(rel, metabuf);
315 : : }
316 : :
317 : : /*
318 : : * _bt_getroot() -- Get the root page of the btree.
319 : : *
320 : : * Since the root page can move around the btree file, we have to read
321 : : * its location from the metadata page, and then read the root page
322 : : * itself. If no root page exists yet, we have to create one.
323 : : *
324 : : * The access type parameter (BT_READ or BT_WRITE) controls whether
325 : : * a new root page will be created or not. If access = BT_READ,
326 : : * and no root page exists, we just return InvalidBuffer. For
327 : : * BT_WRITE, we try to create the root page if it doesn't exist.
328 : : * NOTE that the returned root page will have only a read lock set
329 : : * on it even if access = BT_WRITE!
330 : : *
331 : : * If access = BT_WRITE, heaprel must be set; otherwise caller can just
332 : : * pass NULL. See _bt_allocbuf for an explanation.
333 : : *
334 : : * The returned page is not necessarily the true root --- it could be
335 : : * a "fast root" (a page that is alone in its level due to deletions).
336 : : * Also, if the root page is split while we are "in flight" to it,
337 : : * what we will return is the old root, which is now just the leftmost
338 : : * page on a probably-not-very-wide level. For most purposes this is
339 : : * as good as or better than the true root, so we do not bother to
340 : : * insist on finding the true root. We do, however, guarantee to
341 : : * return a live (not deleted or half-dead) page.
342 : : *
343 : : * On successful return, the root page is pinned and read-locked.
344 : : * The metadata page is not locked or pinned on exit.
345 : : */
346 : : Buffer
347 : 15850084 : _bt_getroot(Relation rel, Relation heaprel, int access)
348 : : {
349 : : Buffer metabuf;
350 : : Buffer rootbuf;
351 : : Page rootpage;
352 : : BTPageOpaque rootopaque;
353 : : BlockNumber rootblkno;
354 : : uint32 rootlevel;
355 : : BTMetaPageData *metad;
356 : : XLogRecPtr recptr;
357 : :
358 : : Assert(access == BT_READ || heaprel != NULL);
359 : :
360 : : /*
361 : : * Try to use previously-cached metapage data to find the root. This
362 : : * normally saves one buffer access per index search, which is a very
363 : : * helpful savings in bufmgr traffic and hence contention.
364 : : */
365 [ + + ]: 15850084 : if (rel->rd_amcache != NULL)
366 : : {
367 : 15501313 : metad = (BTMetaPageData *) rel->rd_amcache;
368 : : /* We shouldn't have cached it if any of these fail */
369 : : Assert(metad->btm_magic == BTREE_MAGIC);
370 : : Assert(metad->btm_version >= BTREE_MIN_VERSION);
371 : : Assert(metad->btm_version <= BTREE_VERSION);
372 : : Assert(!metad->btm_allequalimage ||
373 : : metad->btm_version > BTREE_NOVAC_VERSION);
374 : : Assert(metad->btm_root != P_NONE);
375 : :
376 : 15501313 : rootblkno = metad->btm_fastroot;
377 : : Assert(rootblkno != P_NONE);
378 : 15501313 : rootlevel = metad->btm_fastlevel;
379 : :
380 : 15501313 : rootbuf = _bt_getbuf(rel, rootblkno, BT_READ);
381 : 15501313 : rootpage = BufferGetPage(rootbuf);
382 : 15501313 : rootopaque = BTPageGetOpaque(rootpage);
383 : :
384 : : /*
385 : : * Since the cache might be stale, we check the page more carefully
386 : : * here than normal. We *must* check that it's not deleted. If it's
387 : : * not alone on its level, then we reject too --- this may be overly
388 : : * paranoid but better safe than sorry. Note we don't check P_ISROOT,
389 : : * because that's not set in a "fast root".
390 : : */
391 [ + - ]: 15501313 : if (!P_IGNORE(rootopaque) &&
392 [ + - ]: 15501313 : rootopaque->btpo_level == rootlevel &&
393 [ + - ]: 15501313 : P_LEFTMOST(rootopaque) &&
394 [ + + ]: 15501313 : P_RIGHTMOST(rootopaque))
395 : : {
396 : : /* OK, accept cached page as the root */
397 : 15500288 : return rootbuf;
398 : : }
399 : 1025 : _bt_relbuf(rel, rootbuf);
400 : : /* Cache is stale, throw it away */
401 [ + - ]: 1025 : if (rel->rd_amcache)
402 : 1025 : pfree(rel->rd_amcache);
403 : 1025 : rel->rd_amcache = NULL;
404 : : }
405 : :
406 : 349796 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_READ);
407 : 349796 : metad = _bt_getmeta(rel, metabuf);
408 : :
409 : : /* if no root page initialized yet, do it */
410 [ + + ]: 349796 : if (metad->btm_root == P_NONE)
411 : : {
412 : : Page metapg;
413 : :
414 : : /* If access = BT_READ, caller doesn't want us to create root yet */
415 [ + + ]: 348516 : if (access == BT_READ)
416 : : {
417 : 341009 : _bt_relbuf(rel, metabuf);
418 : 341009 : return InvalidBuffer;
419 : : }
420 : :
421 : : /* trade in our read lock for a write lock */
422 : 7507 : _bt_unlockbuf(rel, metabuf);
423 : 7507 : _bt_lockbuf(rel, metabuf, BT_WRITE);
424 : :
425 : : /*
426 : : * Race condition: if someone else initialized the metadata between
427 : : * the time we released the read lock and acquired the write lock, we
428 : : * must avoid doing it again.
429 : : */
430 [ - + ]: 7507 : if (metad->btm_root != P_NONE)
431 : : {
432 : : /*
433 : : * Metadata initialized by someone else. In order to guarantee no
434 : : * deadlocks, we have to release the metadata page and start all
435 : : * over again. (Is that really true? But it's hardly worth trying
436 : : * to optimize this case.)
437 : : */
438 : 0 : _bt_relbuf(rel, metabuf);
439 : 0 : return _bt_getroot(rel, heaprel, access);
440 : : }
441 : :
442 : : /*
443 : : * Get, initialize, write, and leave a lock of the appropriate type on
444 : : * the new root page. Since this is the first page in the tree, it's
445 : : * a leaf as well as the root.
446 : : */
447 : 7507 : rootbuf = _bt_allocbuf(rel, heaprel);
448 : 7507 : rootblkno = BufferGetBlockNumber(rootbuf);
449 : 7507 : rootpage = BufferGetPage(rootbuf);
450 : 7507 : rootopaque = BTPageGetOpaque(rootpage);
451 : 7507 : rootopaque->btpo_prev = rootopaque->btpo_next = P_NONE;
452 : 7507 : rootopaque->btpo_flags = (BTP_LEAF | BTP_ROOT);
453 : 7507 : rootopaque->btpo_level = 0;
454 : 7507 : rootopaque->btpo_cycleid = 0;
455 : : /* Get raw page pointer for metapage */
456 : 7507 : metapg = BufferGetPage(metabuf);
457 : :
458 : : /* NO ELOG(ERROR) till meta is updated */
459 : 7507 : START_CRIT_SECTION();
460 : :
461 : : /* upgrade metapage if needed */
462 [ - + ]: 7507 : if (metad->btm_version < BTREE_NOVAC_VERSION)
463 : 0 : _bt_upgrademetapage(metapg);
464 : :
465 : 7507 : metad->btm_root = rootblkno;
466 : 7507 : metad->btm_level = 0;
467 : 7507 : metad->btm_fastroot = rootblkno;
468 : 7507 : metad->btm_fastlevel = 0;
469 : 7507 : metad->btm_last_cleanup_num_delpages = 0;
470 : 7507 : metad->btm_last_cleanup_num_heap_tuples = -1.0;
471 : :
472 : 7507 : MarkBufferDirty(rootbuf);
473 : 7507 : MarkBufferDirty(metabuf);
474 : :
475 : : /* XLOG stuff */
476 [ + + + + : 7507 : if (RelationNeedsWAL(rel))
+ + + + ]
477 : 7196 : {
478 : : xl_btree_newroot xlrec;
479 : : xl_btree_metadata md;
480 : :
481 : 7196 : XLogBeginInsert();
482 : 7196 : XLogRegisterBuffer(0, rootbuf, REGBUF_WILL_INIT);
483 : 7196 : XLogRegisterBuffer(2, metabuf, REGBUF_WILL_INIT | REGBUF_STANDARD);
484 : :
485 : : Assert(metad->btm_version >= BTREE_NOVAC_VERSION);
486 : 7196 : md.version = metad->btm_version;
487 : 7196 : md.root = rootblkno;
488 : 7196 : md.level = 0;
489 : 7196 : md.fastroot = rootblkno;
490 : 7196 : md.fastlevel = 0;
491 : 7196 : md.last_cleanup_num_delpages = 0;
492 : 7196 : md.allequalimage = metad->btm_allequalimage;
493 : :
494 : 7196 : XLogRegisterBufData(2, &md, sizeof(xl_btree_metadata));
495 : :
496 : 7196 : xlrec.rootblk = rootblkno;
497 : 7196 : xlrec.level = 0;
498 : :
499 : 7196 : XLogRegisterData(&xlrec, SizeOfBtreeNewroot);
500 : :
501 : 7196 : recptr = XLogInsert(RM_BTREE_ID, XLOG_BTREE_NEWROOT);
502 : : }
503 : : else
504 : 311 : recptr = XLogGetFakeLSN(rel);
505 : :
506 : 7507 : PageSetLSN(rootpage, recptr);
507 : 7507 : PageSetLSN(metapg, recptr);
508 : :
509 : 7507 : END_CRIT_SECTION();
510 : :
511 : : /*
512 : : * swap root write lock for read lock. There is no danger of anyone
513 : : * else accessing the new root page while it's unlocked, since no one
514 : : * else knows where it is yet.
515 : : */
516 : 7507 : _bt_unlockbuf(rel, rootbuf);
517 : 7507 : _bt_lockbuf(rel, rootbuf, BT_READ);
518 : :
519 : : /* okay, metadata is correct, release lock on it without caching */
520 : 7507 : _bt_relbuf(rel, metabuf);
521 : : }
522 : : else
523 : : {
524 : 1280 : rootblkno = metad->btm_fastroot;
525 : : Assert(rootblkno != P_NONE);
526 : 1280 : rootlevel = metad->btm_fastlevel;
527 : :
528 : : /*
529 : : * Cache the metapage data for next time
530 : : */
531 : 1280 : rel->rd_amcache = MemoryContextAlloc(rel->rd_indexcxt,
532 : : sizeof(BTMetaPageData));
533 : 1280 : memcpy(rel->rd_amcache, metad, sizeof(BTMetaPageData));
534 : :
535 : : /*
536 : : * We are done with the metapage; arrange to release it via first
537 : : * _bt_relandgetbuf call
538 : : */
539 : 1280 : rootbuf = metabuf;
540 : :
541 : : for (;;)
542 : : {
543 : 1280 : rootbuf = _bt_relandgetbuf(rel, rootbuf, rootblkno, BT_READ);
544 : 1280 : rootpage = BufferGetPage(rootbuf);
545 : 1280 : rootopaque = BTPageGetOpaque(rootpage);
546 : :
547 [ + - ]: 1280 : if (!P_IGNORE(rootopaque))
548 : 1280 : break;
549 : :
550 : : /* it's dead, Jim. step right one page */
551 [ # # ]: 0 : if (P_RIGHTMOST(rootopaque))
552 [ # # ]: 0 : elog(ERROR, "no live root page found in index \"%s\"",
553 : : RelationGetRelationName(rel));
554 : 0 : rootblkno = rootopaque->btpo_next;
555 : : }
556 : :
557 [ - + ]: 1280 : if (rootopaque->btpo_level != rootlevel)
558 [ # # ]: 0 : elog(ERROR, "root page %u of index \"%s\" has level %u, expected %u",
559 : : rootblkno, RelationGetRelationName(rel),
560 : : rootopaque->btpo_level, rootlevel);
561 : : }
562 : :
563 : : /*
564 : : * By here, we have a pin and read lock on the root page, and no lock set
565 : : * on the metadata page. Return the root page's buffer.
566 : : */
567 : 8787 : return rootbuf;
568 : : }
569 : :
570 : : /*
571 : : * _bt_gettrueroot() -- Get the true root page of the btree.
572 : : *
573 : : * This is the same as the BT_READ case of _bt_getroot(), except
574 : : * we follow the true-root link not the fast-root link.
575 : : *
576 : : * By the time we acquire lock on the root page, it might have been split and
577 : : * not be the true root anymore. This is okay for the present uses of this
578 : : * routine; we only really need to be able to move up at least one tree level
579 : : * from whatever non-root page we were at. If we ever do need to lock the
580 : : * one true root page, we could loop here, re-reading the metapage on each
581 : : * failure. (Note that it wouldn't do to hold the lock on the metapage while
582 : : * moving to the root --- that'd deadlock against any concurrent root split.)
583 : : */
584 : : Buffer
585 : 16 : _bt_gettrueroot(Relation rel)
586 : : {
587 : : Buffer metabuf;
588 : : Buffer rootbuf;
589 : : Page rootpage;
590 : : BTPageOpaque rootopaque;
591 : : BlockNumber rootblkno;
592 : : uint32 rootlevel;
593 : : BTMetaPageData *metad;
594 : :
595 : : /*
596 : : * We don't try to use cached metapage data here, since (a) this path is
597 : : * not performance-critical, and (b) if we are here it suggests our cache
598 : : * is out-of-date anyway. In light of point (b), it's probably safest to
599 : : * actively flush any cached metapage info.
600 : : */
601 [ + - ]: 16 : if (rel->rd_amcache)
602 : 16 : pfree(rel->rd_amcache);
603 : 16 : rel->rd_amcache = NULL;
604 : :
605 : 16 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_READ);
606 : 16 : metad = _bt_getmeta(rel, metabuf);
607 : :
608 : : /* if no root page initialized yet, fail */
609 [ - + ]: 16 : if (metad->btm_root == P_NONE)
610 : : {
611 : 0 : _bt_relbuf(rel, metabuf);
612 : 0 : return InvalidBuffer;
613 : : }
614 : :
615 : 16 : rootblkno = metad->btm_root;
616 : 16 : rootlevel = metad->btm_level;
617 : :
618 : : /*
619 : : * We are done with the metapage; arrange to release it via first
620 : : * _bt_relandgetbuf call
621 : : */
622 : 16 : rootbuf = metabuf;
623 : :
624 : : for (;;)
625 : : {
626 : 16 : rootbuf = _bt_relandgetbuf(rel, rootbuf, rootblkno, BT_READ);
627 : 16 : rootpage = BufferGetPage(rootbuf);
628 : 16 : rootopaque = BTPageGetOpaque(rootpage);
629 : :
630 [ + - ]: 16 : if (!P_IGNORE(rootopaque))
631 : 16 : break;
632 : :
633 : : /* it's dead, Jim. step right one page */
634 [ # # ]: 0 : if (P_RIGHTMOST(rootopaque))
635 [ # # ]: 0 : elog(ERROR, "no live root page found in index \"%s\"",
636 : : RelationGetRelationName(rel));
637 : 0 : rootblkno = rootopaque->btpo_next;
638 : : }
639 : :
640 [ - + ]: 16 : if (rootopaque->btpo_level != rootlevel)
641 [ # # ]: 0 : elog(ERROR, "root page %u of index \"%s\" has level %u, expected %u",
642 : : rootblkno, RelationGetRelationName(rel),
643 : : rootopaque->btpo_level, rootlevel);
644 : :
645 : 16 : return rootbuf;
646 : : }
647 : :
648 : : /*
649 : : * _bt_getrootheight() -- Get the height of the btree search tree.
650 : : *
651 : : * We return the level (counting from zero) of the current fast root.
652 : : * This represents the number of tree levels we'd have to descend through
653 : : * to start any btree index search.
654 : : *
655 : : * This is used by the planner for cost-estimation purposes. Since it's
656 : : * only an estimate, slightly-stale data is fine, hence we don't worry
657 : : * about updating previously cached data.
658 : : */
659 : : int
660 : 3611092 : _bt_getrootheight(Relation rel)
661 : : {
662 : : BTMetaPageData *metad;
663 : :
664 [ + + ]: 3611092 : if (rel->rd_amcache == NULL)
665 : : {
666 : : Buffer metabuf;
667 : :
668 : 58212 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_READ);
669 : 58211 : metad = _bt_getmeta(rel, metabuf);
670 : :
671 : : /*
672 : : * If there's no root page yet, _bt_getroot() doesn't expect a cache
673 : : * to be made, so just stop here and report the index height is zero.
674 : : * (XXX perhaps _bt_getroot() should be changed to allow this case.)
675 : : */
676 [ + + ]: 58211 : if (metad->btm_root == P_NONE)
677 : : {
678 : 33670 : _bt_relbuf(rel, metabuf);
679 : 33670 : return 0;
680 : : }
681 : :
682 : : /*
683 : : * Cache the metapage data for next time
684 : : */
685 : 24541 : rel->rd_amcache = MemoryContextAlloc(rel->rd_indexcxt,
686 : : sizeof(BTMetaPageData));
687 : 24541 : memcpy(rel->rd_amcache, metad, sizeof(BTMetaPageData));
688 : 24541 : _bt_relbuf(rel, metabuf);
689 : : }
690 : :
691 : : /* Get cached page */
692 : 3577421 : metad = (BTMetaPageData *) rel->rd_amcache;
693 : : /* We shouldn't have cached it if any of these fail */
694 : : Assert(metad->btm_magic == BTREE_MAGIC);
695 : : Assert(metad->btm_version >= BTREE_MIN_VERSION);
696 : : Assert(metad->btm_version <= BTREE_VERSION);
697 : : Assert(!metad->btm_allequalimage ||
698 : : metad->btm_version > BTREE_NOVAC_VERSION);
699 : : Assert(metad->btm_fastroot != P_NONE);
700 : :
701 : 3577421 : return metad->btm_fastlevel;
702 : : }
703 : :
704 : : /*
705 : : * _bt_metaversion() -- Get version/status info from metapage.
706 : : *
707 : : * Sets caller's *heapkeyspace and *allequalimage arguments using data
708 : : * from the B-Tree metapage (could be locally-cached version). This
709 : : * information needs to be stashed in insertion scankey, so we provide a
710 : : * single function that fetches both at once.
711 : : *
712 : : * This is used to determine the rules that must be used to descend a
713 : : * btree. Version 4 indexes treat heap TID as a tiebreaker attribute.
714 : : * pg_upgrade'd version 3 indexes need extra steps to preserve reasonable
715 : : * performance when inserting a new BTScanInsert-wise duplicate tuple
716 : : * among many leaf pages already full of such duplicates.
717 : : *
718 : : * Also sets allequalimage field, which indicates whether or not it is
719 : : * safe to apply deduplication. We rely on the assumption that
720 : : * btm_allequalimage will be zero'ed on heapkeyspace indexes that were
721 : : * pg_upgrade'd from Postgres 12.
722 : : */
723 : : void
724 : 17889237 : _bt_metaversion(Relation rel, bool *heapkeyspace, bool *allequalimage)
725 : : {
726 : : BTMetaPageData *metad;
727 : :
728 [ + + ]: 17889237 : if (rel->rd_amcache == NULL)
729 : : {
730 : : Buffer metabuf;
731 : :
732 : 735198 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_READ);
733 : 735198 : metad = _bt_getmeta(rel, metabuf);
734 : :
735 : : /*
736 : : * If there's no root page yet, _bt_getroot() doesn't expect a cache
737 : : * to be made, so just stop here. (XXX perhaps _bt_getroot() should
738 : : * be changed to allow this case.)
739 : : */
740 [ + + ]: 735198 : if (metad->btm_root == P_NONE)
741 : : {
742 : 343857 : *heapkeyspace = metad->btm_version > BTREE_NOVAC_VERSION;
743 : 343857 : *allequalimage = metad->btm_allequalimage;
744 : :
745 : 343857 : _bt_relbuf(rel, metabuf);
746 : 343857 : return;
747 : : }
748 : :
749 : : /*
750 : : * Cache the metapage data for next time
751 : : *
752 : : * An on-the-fly version upgrade performed by _bt_upgrademetapage()
753 : : * can change the nbtree version for an index without invalidating any
754 : : * local cache. This is okay because it can only happen when moving
755 : : * from version 2 to version 3, both of which are !heapkeyspace
756 : : * versions.
757 : : */
758 : 391341 : rel->rd_amcache = MemoryContextAlloc(rel->rd_indexcxt,
759 : : sizeof(BTMetaPageData));
760 : 391341 : memcpy(rel->rd_amcache, metad, sizeof(BTMetaPageData));
761 : 391341 : _bt_relbuf(rel, metabuf);
762 : : }
763 : :
764 : : /* Get cached page */
765 : 17545380 : metad = (BTMetaPageData *) rel->rd_amcache;
766 : : /* We shouldn't have cached it if any of these fail */
767 : : Assert(metad->btm_magic == BTREE_MAGIC);
768 : : Assert(metad->btm_version >= BTREE_MIN_VERSION);
769 : : Assert(metad->btm_version <= BTREE_VERSION);
770 : : Assert(!metad->btm_allequalimage ||
771 : : metad->btm_version > BTREE_NOVAC_VERSION);
772 : : Assert(metad->btm_fastroot != P_NONE);
773 : :
774 : 17545380 : *heapkeyspace = metad->btm_version > BTREE_NOVAC_VERSION;
775 : 17545380 : *allequalimage = metad->btm_allequalimage;
776 : : }
777 : :
778 : : /*
779 : : * _bt_checkpage() -- Verify that a freshly-read page looks sane.
780 : : */
781 : : void
782 : 29867291 : _bt_checkpage(Relation rel, Buffer buf)
783 : : {
784 : 29867291 : Page page = BufferGetPage(buf);
785 : :
786 : : /*
787 : : * ReadBuffer verifies that every newly-read page passes
788 : : * PageHeaderIsValid, which means it either contains a reasonably sane
789 : : * page header or is all-zero. We have to defend against the all-zero
790 : : * case, however.
791 : : */
792 [ - + ]: 29867291 : if (PageIsNew(page))
793 [ # # ]: 0 : ereport(ERROR,
794 : : (errcode(ERRCODE_INDEX_CORRUPTED),
795 : : errmsg("index \"%s\" contains unexpected zero page at block %u",
796 : : RelationGetRelationName(rel),
797 : : BufferGetBlockNumber(buf)),
798 : : errhint("Please REINDEX it.")));
799 : :
800 : : /*
801 : : * Additionally check that the special area looks sane.
802 : : */
803 [ - + ]: 29867291 : if (PageGetSpecialSize(page) != MAXALIGN(sizeof(BTPageOpaqueData)))
804 [ # # ]: 0 : ereport(ERROR,
805 : : (errcode(ERRCODE_INDEX_CORRUPTED),
806 : : errmsg("index \"%s\" contains corrupted page at block %u",
807 : : RelationGetRelationName(rel),
808 : : BufferGetBlockNumber(buf)),
809 : : errhint("Please REINDEX it.")));
810 : 29867291 : }
811 : :
812 : : /*
813 : : * _bt_getbuf() -- Get an existing block in a buffer, for read or write.
814 : : *
815 : : * The general rule in nbtree is that it's never okay to access a
816 : : * page without holding both a buffer pin and a buffer lock on
817 : : * the page's buffer.
818 : : *
819 : : * When this routine returns, the appropriate lock is set on the
820 : : * requested buffer and its reference count has been incremented
821 : : * (ie, the buffer is "locked and pinned"). Also, we apply
822 : : * _bt_checkpage to sanity-check the page, and perform Valgrind
823 : : * client requests that help Valgrind detect unsafe page accesses.
824 : : *
825 : : * Note: raw LockBuffer() calls are disallowed in nbtree; all
826 : : * buffer lock requests need to go through wrapper functions such
827 : : * as _bt_lockbuf().
828 : : */
829 : : Buffer
830 : 16937563 : _bt_getbuf(Relation rel, BlockNumber blkno, int access)
831 : : {
832 : : Buffer buf;
833 : :
834 : : Assert(BlockNumberIsValid(blkno));
835 : :
836 : : /* Read an existing block of the relation */
837 : 16937563 : buf = ReadBuffer(rel, blkno);
838 : 16937562 : _bt_lockbuf(rel, buf, access);
839 : 16937562 : _bt_checkpage(rel, buf);
840 : :
841 : 16937562 : return buf;
842 : : }
843 : :
844 : : /*
845 : : * _bt_allocbuf() -- Allocate a new block/page.
846 : : *
847 : : * Returns a write-locked buffer containing an unallocated nbtree page.
848 : : *
849 : : * Callers are required to pass a valid heaprel. We need heaprel so that we
850 : : * can handle generating a snapshotConflictHorizon that makes reusing a page
851 : : * from the FSM safe for queries that may be running on standbys.
852 : : */
853 : : Buffer
854 : 24057 : _bt_allocbuf(Relation rel, Relation heaprel)
855 : : {
856 : : Buffer buf;
857 : : BlockNumber blkno;
858 : : Page page;
859 : :
860 : : Assert(heaprel != NULL);
861 : :
862 : : /*
863 : : * First see if the FSM knows of any free pages.
864 : : *
865 : : * We can't trust the FSM's report unreservedly; we have to check that the
866 : : * page is still free. (For example, an already-free page could have been
867 : : * re-used between the time the last VACUUM scanned it and the time the
868 : : * VACUUM made its FSM updates.)
869 : : *
870 : : * In fact, it's worse than that: we can't even assume that it's safe to
871 : : * take a lock on the reported page. If somebody else has a lock on it,
872 : : * or even worse our own caller does, we could deadlock. (The own-caller
873 : : * scenario is actually not improbable. Consider an index on a serial or
874 : : * timestamp column. Nearly all splits will be at the rightmost page, so
875 : : * it's entirely likely that _bt_split will call us while holding a lock
876 : : * on the page most recently acquired from FSM. A VACUUM running
877 : : * concurrently with the previous split could well have placed that page
878 : : * back in FSM.)
879 : : *
880 : : * To get around that, we ask for only a conditional lock on the reported
881 : : * page. If we fail, then someone else is using the page, and we may
882 : : * reasonably assume it's not free. (If we happen to be wrong, the worst
883 : : * consequence is the page will be lost to use till the next VACUUM, which
884 : : * is no big problem.)
885 : : */
886 : : for (;;)
887 : : {
888 : 24057 : blkno = GetFreeIndexPage(rel);
889 [ + + ]: 24057 : if (blkno == InvalidBlockNumber)
890 : 23975 : break;
891 : 82 : buf = ReadBuffer(rel, blkno);
892 [ + - ]: 82 : if (_bt_conditionallockbuf(rel, buf))
893 : : {
894 : 82 : page = BufferGetPage(buf);
895 : :
896 : : /*
897 : : * It's possible to find an all-zeroes page in an index. For
898 : : * example, a backend might successfully extend the relation one
899 : : * page and then crash before it is able to make a WAL entry for
900 : : * adding the page. If we find a zeroed page then reclaim it
901 : : * immediately.
902 : : */
903 [ - + ]: 82 : if (PageIsNew(page))
904 : : {
905 : : /* Okay to use page. Initialize and return it. */
906 : 0 : _bt_pageinit(page, BufferGetPageSize(buf));
907 : 0 : return buf;
908 : : }
909 : :
910 [ + - ]: 82 : if (BTPageIsRecyclable(page, heaprel))
911 : : {
912 : : /*
913 : : * If we are generating WAL for Hot Standby then create a WAL
914 : : * record that will allow us to conflict with queries running
915 : : * on standby, in case they have snapshots older than safexid
916 : : * value
917 : : */
918 [ + - - + : 82 : if (RelationNeedsWAL(rel) && XLogStandbyInfoActive())
- - - - +
- ]
919 : : {
920 : : xl_btree_reuse_page xlrec_reuse;
921 : :
922 : : /*
923 : : * Note that we don't register the buffer with the record,
924 : : * because this operation doesn't modify the page (that
925 : : * already happened, back when VACUUM deleted the page).
926 : : * This record only exists to provide a conflict point for
927 : : * Hot Standby. See record REDO routine comments.
928 : : */
929 : 82 : xlrec_reuse.locator = rel->rd_locator;
930 : 82 : xlrec_reuse.block = blkno;
931 : 82 : xlrec_reuse.snapshotConflictHorizon = BTPageGetDeleteXid(page);
932 : 82 : xlrec_reuse.isCatalogRel =
933 [ + - - + : 82 : RelationIsAccessibleInLogicalDecoding(heaprel);
- - - - -
- - - - -
- - - - -
- - - ]
934 : :
935 : 82 : XLogBeginInsert();
936 : 82 : XLogRegisterData(&xlrec_reuse, SizeOfBtreeReusePage);
937 : :
938 : 82 : XLogInsert(RM_BTREE_ID, XLOG_BTREE_REUSE_PAGE);
939 : : }
940 : :
941 : : /* Okay to use page. Re-initialize and return it. */
942 : 82 : _bt_pageinit(page, BufferGetPageSize(buf));
943 : 82 : return buf;
944 : : }
945 [ # # ]: 0 : elog(DEBUG2, "FSM returned nonrecyclable page");
946 : 0 : _bt_relbuf(rel, buf);
947 : : }
948 : : else
949 : : {
950 [ # # ]: 0 : elog(DEBUG2, "FSM returned nonlockable page");
951 : : /* couldn't get lock, so just drop pin */
952 : 0 : ReleaseBuffer(buf);
953 : : }
954 : : }
955 : :
956 : : /*
957 : : * Extend the relation by one page. Need to use RBM_ZERO_AND_LOCK or we
958 : : * risk a race condition against btvacuumscan --- see comments therein.
959 : : * This forces us to repeat the valgrind request that _bt_lockbuf()
960 : : * otherwise would make, as we can't use _bt_lockbuf() without introducing
961 : : * a race.
962 : : */
963 : 23975 : buf = ExtendBufferedRel(BMR_REL(rel), MAIN_FORKNUM, NULL, EB_LOCK_FIRST);
964 : 23975 : if (!RelationUsesLocalBuffers(rel))
965 : : VALGRIND_MAKE_MEM_DEFINED(BufferGetPage(buf), BLCKSZ);
966 : :
967 : : /* Initialize the new page before returning it */
968 : 23975 : page = BufferGetPage(buf);
969 : : Assert(PageIsNew(page));
970 : 23975 : _bt_pageinit(page, BufferGetPageSize(buf));
971 : :
972 : 23975 : return buf;
973 : : }
974 : :
975 : : /*
976 : : * _bt_relandgetbuf() -- release a locked buffer and get another one.
977 : : *
978 : : * This is equivalent to _bt_relbuf followed by _bt_getbuf. Also, if obuf is
979 : : * InvalidBuffer then it reduces to just _bt_getbuf; allowing this case
980 : : * simplifies some callers.
981 : : *
982 : : * The original motivation for using this was to avoid two entries to the
983 : : * bufmgr when one would do. However, now it's mainly just a notational
984 : : * convenience. The only case where it saves work over _bt_relbuf/_bt_getbuf
985 : : * is when the target page is the same one already in the buffer.
986 : : */
987 : : Buffer
988 : 12859769 : _bt_relandgetbuf(Relation rel, Buffer obuf, BlockNumber blkno, int access)
989 : : {
990 : : Buffer buf;
991 : :
992 : : Assert(BlockNumberIsValid(blkno));
993 [ + + ]: 12859769 : if (BufferIsValid(obuf))
994 : : {
995 [ - + ]: 12849233 : if (BufferGetBlockNumber(obuf) == blkno)
996 : : {
997 : : /* trade in old lock mode for new lock */
998 : 0 : _bt_unlockbuf(rel, obuf);
999 : 0 : buf = obuf;
1000 : : }
1001 : : else
1002 : : {
1003 : : /* release lock and pin at once, that's a bit more efficient */
1004 : 12849233 : _bt_relbuf(rel, obuf);
1005 : 12849233 : buf = ReadBuffer(rel, blkno);
1006 : : }
1007 : : }
1008 : : else
1009 : 10536 : buf = ReadBuffer(rel, blkno);
1010 : :
1011 : 12859769 : _bt_lockbuf(rel, buf, access);
1012 : 12859769 : _bt_checkpage(rel, buf);
1013 : :
1014 : 12859769 : return buf;
1015 : : }
1016 : :
1017 : : /*
1018 : : * _bt_relbuf() -- release a locked buffer.
1019 : : *
1020 : : * Lock and pin (refcount) are both dropped. This is a bit more efficient than
1021 : : * doing the two operations separately.
1022 : : */
1023 : : void
1024 : 27004123 : _bt_relbuf(Relation rel, Buffer buf)
1025 : : {
1026 : : /*
1027 : : * Buffer is pinned and locked, which means that it is expected to be
1028 : : * defined and addressable. Check that proactively.
1029 : : */
1030 : : VALGRIND_CHECK_MEM_IS_DEFINED(BufferGetPage(buf), BLCKSZ);
1031 : 27004123 : if (!RelationUsesLocalBuffers(rel))
1032 : : VALGRIND_MAKE_MEM_NOACCESS(BufferGetPage(buf), BLCKSZ);
1033 : :
1034 : 27004123 : UnlockReleaseBuffer(buf);
1035 : 27004123 : }
1036 : :
1037 : : /*
1038 : : * _bt_lockbuf() -- lock a pinned buffer.
1039 : : *
1040 : : * Lock is acquired without acquiring another pin. This is like a raw
1041 : : * LockBuffer() call, but performs extra steps needed by Valgrind.
1042 : : *
1043 : : * Note: Caller may need to call _bt_checkpage() with buf when pin on buf
1044 : : * wasn't originally acquired in _bt_getbuf() or _bt_relandgetbuf().
1045 : : */
1046 : : void
1047 : 30422725 : _bt_lockbuf(Relation rel, Buffer buf, int access)
1048 : : {
1049 : : /* LockBuffer() asserts that pin is held by this backend */
1050 : 30422725 : LockBuffer(buf, access);
1051 : :
1052 : : /*
1053 : : * It doesn't matter that _bt_unlockbuf() won't get called in the event of
1054 : : * an nbtree error (e.g. a unique violation error). That won't cause
1055 : : * Valgrind false positives.
1056 : : *
1057 : : * The nbtree client requests are superimposed on top of the bufmgr.c
1058 : : * buffer pin client requests. In the event of an nbtree error the buffer
1059 : : * will certainly get marked as defined when the backend once again
1060 : : * acquires its first pin on the buffer. (Of course, if the backend never
1061 : : * touches the buffer again then it doesn't matter that it remains
1062 : : * non-accessible to Valgrind.)
1063 : : *
1064 : : * Note: When an IndexTuple C pointer gets computed using an ItemId read
1065 : : * from a page while a lock was held, the C pointer becomes unsafe to
1066 : : * dereference forever as soon as the lock is released. Valgrind can only
1067 : : * detect cases where the pointer gets dereferenced with no _current_
1068 : : * lock/pin held, though.
1069 : : */
1070 : 30422725 : if (!RelationUsesLocalBuffers(rel))
1071 : : VALGRIND_MAKE_MEM_DEFINED(BufferGetPage(buf), BLCKSZ);
1072 : 30422725 : }
1073 : :
1074 : : /*
1075 : : * _bt_unlockbuf() -- unlock a pinned buffer.
1076 : : */
1077 : : void
1078 : 3475184 : _bt_unlockbuf(Relation rel, Buffer buf)
1079 : : {
1080 : : /*
1081 : : * Buffer is pinned and locked, which means that it is expected to be
1082 : : * defined and addressable. Check that proactively.
1083 : : */
1084 : : VALGRIND_CHECK_MEM_IS_DEFINED(BufferGetPage(buf), BLCKSZ);
1085 : :
1086 : : /* LockBuffer() asserts that pin is held by this backend */
1087 : 3475184 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
1088 : :
1089 : 3475184 : if (!RelationUsesLocalBuffers(rel))
1090 : : VALGRIND_MAKE_MEM_NOACCESS(BufferGetPage(buf), BLCKSZ);
1091 : 3475184 : }
1092 : :
1093 : : /*
1094 : : * _bt_conditionallockbuf() -- conditionally BT_WRITE lock pinned
1095 : : * buffer.
1096 : : *
1097 : : * Note: Caller may need to call _bt_checkpage() with buf when pin on buf
1098 : : * wasn't originally acquired in _bt_getbuf() or _bt_relandgetbuf().
1099 : : */
1100 : : bool
1101 : 33232 : _bt_conditionallockbuf(Relation rel, Buffer buf)
1102 : : {
1103 : : /* ConditionalLockBuffer() asserts that pin is held by this backend */
1104 [ + + ]: 33232 : if (!ConditionalLockBuffer(buf))
1105 : 618 : return false;
1106 : :
1107 : 32614 : if (!RelationUsesLocalBuffers(rel))
1108 : : VALGRIND_MAKE_MEM_DEFINED(BufferGetPage(buf), BLCKSZ);
1109 : :
1110 : 32614 : return true;
1111 : : }
1112 : :
1113 : : /*
1114 : : * _bt_upgradelockbufcleanup() -- upgrade lock to a full cleanup lock.
1115 : : */
1116 : : void
1117 : 14013 : _bt_upgradelockbufcleanup(Relation rel, Buffer buf)
1118 : : {
1119 : : /*
1120 : : * Buffer is pinned and locked, which means that it is expected to be
1121 : : * defined and addressable. Check that proactively.
1122 : : */
1123 : : VALGRIND_CHECK_MEM_IS_DEFINED(BufferGetPage(buf), BLCKSZ);
1124 : :
1125 : : /* LockBuffer() asserts that pin is held by this backend */
1126 : 14013 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
1127 : 14013 : LockBufferForCleanup(buf);
1128 : 14013 : }
1129 : :
1130 : : /*
1131 : : * _bt_pageinit() -- Initialize a new page.
1132 : : *
1133 : : * On return, the page header is initialized; data space is empty;
1134 : : * special space is zeroed out.
1135 : : */
1136 : : void
1137 : 109500 : _bt_pageinit(Page page, Size size)
1138 : : {
1139 : 109500 : PageInit(page, size, sizeof(BTPageOpaqueData));
1140 : 109500 : }
1141 : :
1142 : : /*
1143 : : * Delete item(s) from a btree leaf page during VACUUM.
1144 : : *
1145 : : * This routine assumes that the caller already has a full cleanup lock on
1146 : : * the buffer. Also, the given deletable and updatable arrays *must* be
1147 : : * sorted in ascending order.
1148 : : *
1149 : : * Routine deals with deleting TIDs when some (but not all) of the heap TIDs
1150 : : * in an existing posting list item are to be removed. This works by
1151 : : * updating/overwriting an existing item with caller's new version of the item
1152 : : * (a version that lacks the TIDs that are to be deleted).
1153 : : *
1154 : : * We record VACUUMs and b-tree deletes differently in WAL. Deletes must
1155 : : * generate their own snapshotConflictHorizon directly from the tableam,
1156 : : * whereas VACUUMs rely on the initial VACUUM table scan performing
1157 : : * WAL-logging that takes care of the issue for the table's indexes
1158 : : * indirectly. Also, we remove the VACUUM cycle ID from pages, which b-tree
1159 : : * deletes don't do.
1160 : : */
1161 : : void
1162 : 8949 : _bt_delitems_vacuum(Relation rel, Buffer buf,
1163 : : OffsetNumber *deletable, int ndeletable,
1164 : : BTVacuumPosting *updatable, int nupdatable)
1165 : : {
1166 : 8949 : Page page = BufferGetPage(buf);
1167 : : BTPageOpaque opaque;
1168 [ + + + + : 8949 : bool needswal = RelationNeedsWAL(rel);
+ - + - ]
1169 : 8949 : char *updatedbuf = NULL;
1170 : 8949 : Size updatedbuflen = 0;
1171 : : OffsetNumber updatedoffsets[MaxIndexTuplesPerPage];
1172 : : XLogRecPtr recptr;
1173 : :
1174 : : /* Shouldn't be called unless there's something to do */
1175 : : Assert(ndeletable > 0 || nupdatable > 0);
1176 : :
1177 : : /* Generate new version of posting lists without deleted TIDs */
1178 [ + + ]: 8949 : if (nupdatable > 0)
1179 : 1173 : updatedbuf = _bt_delitems_update(updatable, nupdatable,
1180 : : updatedoffsets, &updatedbuflen,
1181 : : needswal);
1182 : :
1183 : : /* No ereport(ERROR) until changes are logged */
1184 : 8949 : START_CRIT_SECTION();
1185 : :
1186 : : /*
1187 : : * Handle posting tuple updates.
1188 : : *
1189 : : * Deliberately do this before handling simple deletes. If we did it the
1190 : : * other way around (i.e. WAL record order -- simple deletes before
1191 : : * updates) then we'd have to make compensating changes to the 'updatable'
1192 : : * array of offset numbers.
1193 : : *
1194 : : * PageIndexTupleOverwrite() won't unset each item's LP_DEAD bit when it
1195 : : * happens to already be set. It's important that we not interfere with
1196 : : * any future simple index tuple deletion operations.
1197 : : */
1198 [ + + ]: 67875 : for (int i = 0; i < nupdatable; i++)
1199 : : {
1200 : 58926 : OffsetNumber updatedoffset = updatedoffsets[i];
1201 : : IndexTuple itup;
1202 : : Size itemsz;
1203 : :
1204 : 58926 : itup = updatable[i]->itup;
1205 : 58926 : itemsz = MAXALIGN(IndexTupleSize(itup));
1206 [ - + ]: 58926 : if (!PageIndexTupleOverwrite(page, updatedoffset, itup, itemsz))
1207 [ # # ]: 0 : elog(PANIC, "failed to update partially dead item in block %u of index \"%s\"",
1208 : : BufferGetBlockNumber(buf), RelationGetRelationName(rel));
1209 : : }
1210 : :
1211 : : /* Now handle simple deletes of entire tuples */
1212 [ + + ]: 8949 : if (ndeletable > 0)
1213 : 8630 : PageIndexMultiDelete(page, deletable, ndeletable);
1214 : :
1215 : : /*
1216 : : * We can clear the vacuum cycle ID since this page has certainly been
1217 : : * processed by the current vacuum scan.
1218 : : */
1219 : 8949 : opaque = BTPageGetOpaque(page);
1220 : 8949 : opaque->btpo_cycleid = 0;
1221 : :
1222 : : /*
1223 : : * Clear the BTP_HAS_GARBAGE page flag.
1224 : : *
1225 : : * This flag indicates the presence of LP_DEAD items on the page (though
1226 : : * not reliably). Note that we only rely on it with pg_upgrade'd
1227 : : * !heapkeyspace indexes. That's why clearing it here won't usually
1228 : : * interfere with simple index tuple deletion.
1229 : : */
1230 : 8949 : opaque->btpo_flags &= ~BTP_HAS_GARBAGE;
1231 : :
1232 : 8949 : MarkBufferDirty(buf);
1233 : :
1234 : : /* XLOG stuff */
1235 [ + + ]: 8949 : if (needswal)
1236 : : {
1237 : : xl_btree_vacuum xlrec_vacuum;
1238 : :
1239 : 8948 : xlrec_vacuum.ndeleted = ndeletable;
1240 : 8948 : xlrec_vacuum.nupdated = nupdatable;
1241 : :
1242 : 8948 : XLogBeginInsert();
1243 : 8948 : XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
1244 : 8948 : XLogRegisterData(&xlrec_vacuum, SizeOfBtreeVacuum);
1245 : :
1246 [ + + ]: 8948 : if (ndeletable > 0)
1247 : 8629 : XLogRegisterBufData(0, deletable,
1248 : : ndeletable * sizeof(OffsetNumber));
1249 : :
1250 [ + + ]: 8948 : if (nupdatable > 0)
1251 : : {
1252 : 1173 : XLogRegisterBufData(0, updatedoffsets,
1253 : : nupdatable * sizeof(OffsetNumber));
1254 : 1173 : XLogRegisterBufData(0, updatedbuf, updatedbuflen);
1255 : : }
1256 : :
1257 : 8948 : recptr = XLogInsert(RM_BTREE_ID, XLOG_BTREE_VACUUM);
1258 : : }
1259 : : else
1260 : 1 : recptr = XLogGetFakeLSN(rel);
1261 : :
1262 : 8949 : PageSetLSN(page, recptr);
1263 : :
1264 : 8949 : END_CRIT_SECTION();
1265 : :
1266 : : /* can't leak memory here */
1267 [ + + ]: 8949 : if (updatedbuf != NULL)
1268 : 1173 : pfree(updatedbuf);
1269 : : /* free tuples allocated within _bt_delitems_update() */
1270 [ + + ]: 67875 : for (int i = 0; i < nupdatable; i++)
1271 : 58926 : pfree(updatable[i]->itup);
1272 : 8949 : }
1273 : :
1274 : : /*
1275 : : * Delete item(s) from a btree leaf page during single-page cleanup.
1276 : : *
1277 : : * This routine assumes that the caller has pinned and write locked the
1278 : : * buffer. Also, the given deletable and updatable arrays *must* be sorted in
1279 : : * ascending order.
1280 : : *
1281 : : * Routine deals with deleting TIDs when some (but not all) of the heap TIDs
1282 : : * in an existing posting list item are to be removed. This works by
1283 : : * updating/overwriting an existing item with caller's new version of the item
1284 : : * (a version that lacks the TIDs that are to be deleted).
1285 : : *
1286 : : * This is nearly the same as _bt_delitems_vacuum as far as what it does to
1287 : : * the page, but it needs its own snapshotConflictHorizon and isCatalogRel
1288 : : * (from the tableam). This is used by the REDO routine to generate recovery
1289 : : * conflicts. The other difference is that only _bt_delitems_vacuum will
1290 : : * clear page's VACUUM cycle ID.
1291 : : */
1292 : : static void
1293 : 5962 : _bt_delitems_delete(Relation rel, Buffer buf,
1294 : : TransactionId snapshotConflictHorizon, bool isCatalogRel,
1295 : : OffsetNumber *deletable, int ndeletable,
1296 : : BTVacuumPosting *updatable, int nupdatable)
1297 : : {
1298 : 5962 : Page page = BufferGetPage(buf);
1299 : : BTPageOpaque opaque;
1300 [ + - + + : 5962 : bool needswal = RelationNeedsWAL(rel);
+ - + - ]
1301 : 5962 : char *updatedbuf = NULL;
1302 : 5962 : Size updatedbuflen = 0;
1303 : : OffsetNumber updatedoffsets[MaxIndexTuplesPerPage];
1304 : : XLogRecPtr recptr;
1305 : :
1306 : : /* Shouldn't be called unless there's something to do */
1307 : : Assert(ndeletable > 0 || nupdatable > 0);
1308 : :
1309 : : /* Generate new versions of posting lists without deleted TIDs */
1310 [ + + ]: 5962 : if (nupdatable > 0)
1311 : 579 : updatedbuf = _bt_delitems_update(updatable, nupdatable,
1312 : : updatedoffsets, &updatedbuflen,
1313 : : needswal);
1314 : :
1315 : : /* No ereport(ERROR) until changes are logged */
1316 : 5962 : START_CRIT_SECTION();
1317 : :
1318 : : /* Handle updates and deletes just like _bt_delitems_vacuum */
1319 [ + + ]: 12661 : for (int i = 0; i < nupdatable; i++)
1320 : : {
1321 : 6699 : OffsetNumber updatedoffset = updatedoffsets[i];
1322 : : IndexTuple itup;
1323 : : Size itemsz;
1324 : :
1325 : 6699 : itup = updatable[i]->itup;
1326 : 6699 : itemsz = MAXALIGN(IndexTupleSize(itup));
1327 [ - + ]: 6699 : if (!PageIndexTupleOverwrite(page, updatedoffset, itup, itemsz))
1328 [ # # ]: 0 : elog(PANIC, "failed to update partially dead item in block %u of index \"%s\"",
1329 : : BufferGetBlockNumber(buf), RelationGetRelationName(rel));
1330 : : }
1331 : :
1332 [ + + ]: 5962 : if (ndeletable > 0)
1333 : 5871 : PageIndexMultiDelete(page, deletable, ndeletable);
1334 : :
1335 : : /*
1336 : : * Unlike _bt_delitems_vacuum, we *must not* clear the vacuum cycle ID at
1337 : : * this point. The VACUUM command alone controls vacuum cycle IDs.
1338 : : */
1339 : 5962 : opaque = BTPageGetOpaque(page);
1340 : :
1341 : : /*
1342 : : * Clear the BTP_HAS_GARBAGE page flag.
1343 : : *
1344 : : * This flag indicates the presence of LP_DEAD items on the page (though
1345 : : * not reliably). Note that we only rely on it with pg_upgrade'd
1346 : : * !heapkeyspace indexes.
1347 : : */
1348 : 5962 : opaque->btpo_flags &= ~BTP_HAS_GARBAGE;
1349 : :
1350 : 5962 : MarkBufferDirty(buf);
1351 : :
1352 : : /* XLOG stuff */
1353 [ + - ]: 5962 : if (needswal)
1354 : : {
1355 : : xl_btree_delete xlrec_delete;
1356 : :
1357 : 5962 : xlrec_delete.snapshotConflictHorizon = snapshotConflictHorizon;
1358 : 5962 : xlrec_delete.ndeleted = ndeletable;
1359 : 5962 : xlrec_delete.nupdated = nupdatable;
1360 : 5962 : xlrec_delete.isCatalogRel = isCatalogRel;
1361 : :
1362 : 5962 : XLogBeginInsert();
1363 : 5962 : XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
1364 : 5962 : XLogRegisterData(&xlrec_delete, SizeOfBtreeDelete);
1365 : :
1366 [ + + ]: 5962 : if (ndeletable > 0)
1367 : 5871 : XLogRegisterBufData(0, deletable,
1368 : : ndeletable * sizeof(OffsetNumber));
1369 : :
1370 [ + + ]: 5962 : if (nupdatable > 0)
1371 : : {
1372 : 579 : XLogRegisterBufData(0, updatedoffsets,
1373 : : nupdatable * sizeof(OffsetNumber));
1374 : 579 : XLogRegisterBufData(0, updatedbuf, updatedbuflen);
1375 : : }
1376 : :
1377 : 5962 : recptr = XLogInsert(RM_BTREE_ID, XLOG_BTREE_DELETE);
1378 : : }
1379 : : else
1380 : 0 : recptr = XLogGetFakeLSN(rel);
1381 : :
1382 : 5962 : PageSetLSN(page, recptr);
1383 : :
1384 : 5962 : END_CRIT_SECTION();
1385 : :
1386 : : /* can't leak memory here */
1387 [ + + ]: 5962 : if (updatedbuf != NULL)
1388 : 579 : pfree(updatedbuf);
1389 : : /* free tuples allocated within _bt_delitems_update() */
1390 [ + + ]: 12661 : for (int i = 0; i < nupdatable; i++)
1391 : 6699 : pfree(updatable[i]->itup);
1392 : 5962 : }
1393 : :
1394 : : /*
1395 : : * Set up state needed to delete TIDs from posting list tuples via "updating"
1396 : : * the tuple. Performs steps common to both _bt_delitems_vacuum and
1397 : : * _bt_delitems_delete. These steps must take place before each function's
1398 : : * critical section begins.
1399 : : *
1400 : : * updatable and nupdatable are inputs, though note that we will use
1401 : : * _bt_update_posting() to replace the original itup with a pointer to a final
1402 : : * version in palloc()'d memory. Caller should free the tuples when its done.
1403 : : *
1404 : : * The first nupdatable entries from updatedoffsets are set to the page offset
1405 : : * number for posting list tuples that caller updates. This is mostly useful
1406 : : * because caller may need to WAL-log the page offsets (though we always do
1407 : : * this for caller out of convenience).
1408 : : *
1409 : : * Returns buffer consisting of an array of xl_btree_update structs that
1410 : : * describe the steps we perform here for caller (though only when needswal is
1411 : : * true). Also sets *updatedbuflen to the final size of the buffer. This
1412 : : * buffer is used by caller when WAL logging is required.
1413 : : */
1414 : : static char *
1415 : 1752 : _bt_delitems_update(BTVacuumPosting *updatable, int nupdatable,
1416 : : OffsetNumber *updatedoffsets, Size *updatedbuflen,
1417 : : bool needswal)
1418 : : {
1419 : 1752 : char *updatedbuf = NULL;
1420 : 1752 : Size buflen = 0;
1421 : :
1422 : : /* Shouldn't be called unless there's something to do */
1423 : : Assert(nupdatable > 0);
1424 : :
1425 [ + + ]: 67377 : for (int i = 0; i < nupdatable; i++)
1426 : : {
1427 : 65625 : BTVacuumPosting vacposting = updatable[i];
1428 : : Size itemsz;
1429 : :
1430 : : /* Replace work area IndexTuple with updated version */
1431 : 65625 : _bt_update_posting(vacposting);
1432 : :
1433 : : /* Keep track of size of xl_btree_update for updatedbuf in passing */
1434 : 65625 : itemsz = SizeOfBtreeUpdate + vacposting->ndeletedtids * sizeof(uint16);
1435 : 65625 : buflen += itemsz;
1436 : :
1437 : : /* Build updatedoffsets buffer in passing */
1438 : 65625 : updatedoffsets[i] = vacposting->updatedoffset;
1439 : : }
1440 : :
1441 : : /* XLOG stuff */
1442 [ + - ]: 1752 : if (needswal)
1443 : : {
1444 : 1752 : Size offset = 0;
1445 : :
1446 : : /* Allocate, set final size for caller */
1447 : 1752 : updatedbuf = palloc(buflen);
1448 : 1752 : *updatedbuflen = buflen;
1449 [ + + ]: 67377 : for (int i = 0; i < nupdatable; i++)
1450 : : {
1451 : 65625 : BTVacuumPosting vacposting = updatable[i];
1452 : : Size itemsz;
1453 : : xl_btree_update update;
1454 : :
1455 : 65625 : update.ndeletedtids = vacposting->ndeletedtids;
1456 : 65625 : memcpy(updatedbuf + offset, &update.ndeletedtids,
1457 : : SizeOfBtreeUpdate);
1458 : 65625 : offset += SizeOfBtreeUpdate;
1459 : :
1460 : 65625 : itemsz = update.ndeletedtids * sizeof(uint16);
1461 : 65625 : memcpy(updatedbuf + offset, vacposting->deletetids, itemsz);
1462 : 65625 : offset += itemsz;
1463 : : }
1464 : : }
1465 : :
1466 : 1752 : return updatedbuf;
1467 : : }
1468 : :
1469 : : /*
1470 : : * Comparator used by _bt_delitems_delete_check() to restore deltids array
1471 : : * back to its original leaf-page-wise sort order
1472 : : */
1473 : : static int
1474 : 3451845 : _bt_delitems_cmp(const void *a, const void *b)
1475 : : {
1476 : 3451845 : const TM_IndexDelete *indexdelete1 = a;
1477 : 3451845 : const TM_IndexDelete *indexdelete2 = b;
1478 : :
1479 : : Assert(indexdelete1->id != indexdelete2->id);
1480 : :
1481 : 3451845 : return pg_cmp_s16(indexdelete1->id, indexdelete2->id);
1482 : : }
1483 : :
1484 : : /*
1485 : : * Try to delete item(s) from a btree leaf page during single-page cleanup.
1486 : : *
1487 : : * nbtree interface to table_index_delete_tuples(). Deletes a subset of index
1488 : : * tuples from caller's deltids array: those whose TIDs are found safe to
1489 : : * delete by the tableam (or already marked LP_DEAD in index, and so already
1490 : : * known to be deletable by our simple index deletion caller). We physically
1491 : : * delete index tuples from buf leaf page last of all (for index tuples where
1492 : : * that is known to be safe following our table_index_delete_tuples() call).
1493 : : *
1494 : : * Simple index deletion caller only includes TIDs from index tuples marked
1495 : : * LP_DEAD, as well as extra TIDs it found on the same leaf page that can be
1496 : : * included without increasing the total number of distinct table blocks for
1497 : : * the deletion operation as a whole. This approach often allows us to delete
1498 : : * some extra index tuples that were practically free for tableam to check in
1499 : : * passing (when they actually turn out to be safe to delete). It probably
1500 : : * only makes sense for the tableam to go ahead with these extra checks when
1501 : : * it is block-oriented (otherwise the checks probably won't be practically
1502 : : * free, which we rely on). The tableam interface requires the tableam side
1503 : : * to handle the problem, though, so this is okay (we as an index AM are free
1504 : : * to make the simplifying assumption that all tableams must be block-based).
1505 : : *
1506 : : * Bottom-up index deletion caller provides all the TIDs from the leaf page,
1507 : : * without expecting that tableam will check most of them. The tableam has
1508 : : * considerable discretion around which entries/blocks it checks. Our role in
1509 : : * costing the bottom-up deletion operation is strictly advisory.
1510 : : *
1511 : : * Note: Caller must have added deltids entries (i.e. entries that go in
1512 : : * delstate's main array) in leaf-page-wise order: page offset number order,
1513 : : * TID order among entries taken from the same posting list tuple (tiebreak on
1514 : : * TID). This order is convenient to work with here.
1515 : : *
1516 : : * Note: We also rely on the id field of each deltids element "capturing" this
1517 : : * original leaf-page-wise order. That is, we expect to be able to get back
1518 : : * to the original leaf-page-wise order just by sorting deltids on the id
1519 : : * field (tableam will sort deltids for its own reasons, so we'll need to put
1520 : : * it back in leaf-page-wise order afterwards).
1521 : : */
1522 : : void
1523 : 8297 : _bt_delitems_delete_check(Relation rel, Buffer buf, Relation heapRel,
1524 : : TM_IndexDeleteOp *delstate)
1525 : : {
1526 : 8297 : Page page = BufferGetPage(buf);
1527 : : TransactionId snapshotConflictHorizon;
1528 : : bool isCatalogRel;
1529 : 8297 : OffsetNumber postingidxoffnum = InvalidOffsetNumber;
1530 : 8297 : int ndeletable = 0,
1531 : 8297 : nupdatable = 0;
1532 : : OffsetNumber deletable[MaxIndexTuplesPerPage];
1533 : : BTVacuumPosting updatable[MaxIndexTuplesPerPage];
1534 : :
1535 : : /* Use tableam interface to determine which tuples to delete first */
1536 : 8297 : snapshotConflictHorizon = table_index_delete_tuples(heapRel, delstate);
1537 [ + + - + : 8297 : isCatalogRel = RelationIsAccessibleInLogicalDecoding(heapRel);
+ - - + -
- - - + +
- + - - -
- - - ]
1538 : :
1539 : : /* Should not WAL-log snapshotConflictHorizon unless it's required */
1540 [ + + ]: 8297 : if (!XLogStandbyInfoActive())
1541 : 1379 : snapshotConflictHorizon = InvalidTransactionId;
1542 : :
1543 : : /*
1544 : : * Construct a leaf-page-wise description of what _bt_delitems_delete()
1545 : : * needs to do to physically delete index tuples from the page.
1546 : : *
1547 : : * Must sort deltids array to restore leaf-page-wise order (original order
1548 : : * before call to tableam). This is the order that the loop expects.
1549 : : *
1550 : : * Note that deltids array might be a lot smaller now. It might even have
1551 : : * no entries at all (with bottom-up deletion caller), in which case there
1552 : : * is nothing left to do.
1553 : : */
1554 : 8297 : qsort(delstate->deltids, delstate->ndeltids, sizeof(TM_IndexDelete),
1555 : : _bt_delitems_cmp);
1556 [ + + ]: 8297 : if (delstate->ndeltids == 0)
1557 : : {
1558 : : Assert(delstate->bottomup);
1559 : 2335 : return;
1560 : : }
1561 : :
1562 : : /* We definitely have to delete at least one index tuple (or one TID) */
1563 [ + + ]: 508762 : for (int i = 0; i < delstate->ndeltids; i++)
1564 : : {
1565 : 502800 : TM_IndexStatus *dstatus = delstate->status + delstate->deltids[i].id;
1566 : 502800 : OffsetNumber idxoffnum = dstatus->idxoffnum;
1567 : 502800 : ItemId itemid = PageGetItemId(page, idxoffnum);
1568 : 502800 : IndexTuple itup = (IndexTuple) PageGetItem(page, itemid);
1569 : : int nestedi,
1570 : : nitem;
1571 : : BTVacuumPosting vacposting;
1572 : :
1573 : : Assert(OffsetNumberIsValid(idxoffnum));
1574 : :
1575 [ + + ]: 502800 : if (idxoffnum == postingidxoffnum)
1576 : : {
1577 : : /*
1578 : : * This deltid entry is a TID from a posting list tuple that has
1579 : : * already been completely processed
1580 : : */
1581 : : Assert(BTreeTupleIsPosting(itup));
1582 : : Assert(ItemPointerCompare(BTreeTupleGetHeapTID(itup),
1583 : : &delstate->deltids[i].tid) < 0);
1584 : : Assert(ItemPointerCompare(BTreeTupleGetMaxHeapTID(itup),
1585 : : &delstate->deltids[i].tid) >= 0);
1586 : 19285 : continue;
1587 : : }
1588 : :
1589 [ + + ]: 483515 : if (!BTreeTupleIsPosting(itup))
1590 : : {
1591 : : /* Plain non-pivot tuple */
1592 : : Assert(ItemPointerEquals(&itup->t_tid, &delstate->deltids[i].tid));
1593 [ + + ]: 466957 : if (dstatus->knowndeletable)
1594 : 376583 : deletable[ndeletable++] = idxoffnum;
1595 : 466957 : continue;
1596 : : }
1597 : :
1598 : : /*
1599 : : * itup is a posting list tuple whose lowest deltids entry (which may
1600 : : * or may not be for the first TID from itup) is considered here now.
1601 : : * We should process all of the deltids entries for the posting list
1602 : : * together now, though (not just the lowest). Remember to skip over
1603 : : * later itup-related entries during later iterations of outermost
1604 : : * loop.
1605 : : */
1606 : 16558 : postingidxoffnum = idxoffnum; /* Remember work in outermost loop */
1607 : 16558 : nestedi = i; /* Initialize for first itup deltids entry */
1608 : 16558 : vacposting = NULL; /* Describes final action for itup */
1609 : 16558 : nitem = BTreeTupleGetNPosting(itup);
1610 [ + + ]: 76993 : for (int p = 0; p < nitem; p++)
1611 : : {
1612 : 60435 : ItemPointer ptid = BTreeTupleGetPostingN(itup, p);
1613 : 60435 : int ptidcmp = -1;
1614 : :
1615 : : /*
1616 : : * This nested loop reuses work across ptid TIDs taken from itup.
1617 : : * We take advantage of the fact that both itup's TIDs and deltids
1618 : : * entries (within a single itup/posting list grouping) must both
1619 : : * be in ascending TID order.
1620 : : */
1621 [ + + ]: 87007 : for (; nestedi < delstate->ndeltids; nestedi++)
1622 : : {
1623 : 83821 : TM_IndexDelete *tcdeltid = &delstate->deltids[nestedi];
1624 : 83821 : TM_IndexStatus *tdstatus = (delstate->status + tcdeltid->id);
1625 : :
1626 : : /* Stop once we get past all itup related deltids entries */
1627 : : Assert(tdstatus->idxoffnum >= idxoffnum);
1628 [ + + ]: 83821 : if (tdstatus->idxoffnum != idxoffnum)
1629 : 17341 : break;
1630 : :
1631 : : /* Skip past non-deletable itup related entries up front */
1632 [ + + ]: 66480 : if (!tdstatus->knowndeletable)
1633 : 6037 : continue;
1634 : :
1635 : : /* Entry is first partial ptid match (or an exact match)? */
1636 : 60443 : ptidcmp = ItemPointerCompare(&tcdeltid->tid, ptid);
1637 [ + + ]: 60443 : if (ptidcmp >= 0)
1638 : : {
1639 : : /* Greater than or equal (partial or exact) match... */
1640 : 39908 : break;
1641 : : }
1642 : : }
1643 : :
1644 : : /* ...exact ptid match to a deletable deltids entry? */
1645 [ + + ]: 60435 : if (ptidcmp != 0)
1646 : 30629 : continue;
1647 : :
1648 : : /* Exact match for deletable deltids entry -- ptid gets deleted */
1649 [ + + ]: 29806 : if (vacposting == NULL)
1650 : : {
1651 : 14625 : vacposting = palloc(offsetof(BTVacuumPostingData, deletetids) +
1652 : : nitem * sizeof(uint16));
1653 : 14625 : vacposting->itup = itup;
1654 : 14625 : vacposting->updatedoffset = idxoffnum;
1655 : 14625 : vacposting->ndeletedtids = 0;
1656 : : }
1657 : 29806 : vacposting->deletetids[vacposting->ndeletedtids++] = p;
1658 : : }
1659 : :
1660 : : /* Final decision on itup, a posting list tuple */
1661 : :
1662 [ + + ]: 16558 : if (vacposting == NULL)
1663 : : {
1664 : : /* No TIDs to delete from itup -- do nothing */
1665 : : }
1666 [ + + ]: 14625 : else if (vacposting->ndeletedtids == nitem)
1667 : : {
1668 : : /* Straight delete of itup (to delete all TIDs) */
1669 : 7926 : deletable[ndeletable++] = idxoffnum;
1670 : : /* Turns out we won't need granular information */
1671 : 7926 : pfree(vacposting);
1672 : : }
1673 : : else
1674 : : {
1675 : : /* Delete some (but not all) TIDs from itup */
1676 : : Assert(vacposting->ndeletedtids > 0 &&
1677 : : vacposting->ndeletedtids < nitem);
1678 : 6699 : updatable[nupdatable++] = vacposting;
1679 : : }
1680 : : }
1681 : :
1682 : : /* Physically delete tuples (or TIDs) using deletable (or updatable) */
1683 : 5962 : _bt_delitems_delete(rel, buf, snapshotConflictHorizon, isCatalogRel,
1684 : : deletable, ndeletable, updatable, nupdatable);
1685 : :
1686 : : /* be tidy */
1687 [ + + ]: 12661 : for (int i = 0; i < nupdatable; i++)
1688 : 6699 : pfree(updatable[i]);
1689 : : }
1690 : :
1691 : : /*
1692 : : * Check that leftsib page (the btpo_prev of target page) is not marked with
1693 : : * INCOMPLETE_SPLIT flag. Used during page deletion.
1694 : : *
1695 : : * Returning true indicates that page flag is set in leftsib (which is
1696 : : * definitely still the left sibling of target). When that happens, the
1697 : : * target doesn't have a downlink in parent, and the page deletion algorithm
1698 : : * isn't prepared to handle that. Deletion of the target page (or the whole
1699 : : * subtree that contains the target page) cannot take place.
1700 : : *
1701 : : * Caller should not have a lock on the target page itself, since pages on the
1702 : : * same level must always be locked left to right to avoid deadlocks.
1703 : : */
1704 : : static bool
1705 : 3719 : _bt_leftsib_splitflag(Relation rel, BlockNumber leftsib, BlockNumber target)
1706 : : {
1707 : : Buffer buf;
1708 : : Page page;
1709 : : BTPageOpaque opaque;
1710 : : bool result;
1711 : :
1712 : : /* Easy case: No left sibling */
1713 [ + + ]: 3719 : if (leftsib == P_NONE)
1714 : 2905 : return false;
1715 : :
1716 : 814 : buf = _bt_getbuf(rel, leftsib, BT_READ);
1717 : 814 : page = BufferGetPage(buf);
1718 : 814 : opaque = BTPageGetOpaque(page);
1719 : :
1720 : : /*
1721 : : * If the left sibling was concurrently split, so that its next-pointer
1722 : : * doesn't point to the current page anymore, the split that created
1723 : : * target must be completed. Caller can reasonably expect that there will
1724 : : * be a downlink to the target page that it can relocate using its stack.
1725 : : * (We don't allow splitting an incompletely split page again until the
1726 : : * previous split has been completed.)
1727 : : */
1728 [ + - - + ]: 814 : result = (opaque->btpo_next == target && P_INCOMPLETE_SPLIT(opaque));
1729 : 814 : _bt_relbuf(rel, buf);
1730 : :
1731 : 814 : return result;
1732 : : }
1733 : :
1734 : : /*
1735 : : * Check that leafrightsib page (the btpo_next of target leaf page) is not
1736 : : * marked with ISHALFDEAD flag. Used during page deletion.
1737 : : *
1738 : : * Returning true indicates that page flag is set in leafrightsib, so page
1739 : : * deletion cannot go ahead. Our caller is not prepared to deal with the case
1740 : : * where the parent page does not have a pivot tuples whose downlink points to
1741 : : * leafrightsib (due to an earlier interrupted VACUUM operation). It doesn't
1742 : : * seem worth going to the trouble of teaching our caller to deal with it.
1743 : : * The situation will be resolved after VACUUM finishes the deletion of the
1744 : : * half-dead page (when a future VACUUM operation reaches the target page
1745 : : * again).
1746 : : *
1747 : : * _bt_leftsib_splitflag() is called for both leaf pages and internal pages.
1748 : : * _bt_rightsib_halfdeadflag() is only called for leaf pages, though. This is
1749 : : * okay because of the restriction on deleting pages that are the rightmost
1750 : : * page of their parent (i.e. that such deletions can only take place when the
1751 : : * entire subtree must be deleted). The leaf level check made here will apply
1752 : : * to a right "cousin" leaf page rather than a simple right sibling leaf page
1753 : : * in cases where caller actually goes on to attempt deleting pages that are
1754 : : * above the leaf page. The right cousin leaf page is representative of the
1755 : : * left edge of the subtree to the right of the to-be-deleted subtree as a
1756 : : * whole, which is exactly the condition that our caller cares about.
1757 : : * (Besides, internal pages are never marked half-dead, so it isn't even
1758 : : * possible to _directly_ assess if an internal page is part of some other
1759 : : * to-be-deleted subtree.)
1760 : : */
1761 : : static bool
1762 : 3552 : _bt_rightsib_halfdeadflag(Relation rel, BlockNumber leafrightsib)
1763 : : {
1764 : : Buffer buf;
1765 : : Page page;
1766 : : BTPageOpaque opaque;
1767 : : bool result;
1768 : :
1769 : : Assert(leafrightsib != P_NONE);
1770 : :
1771 : 3552 : buf = _bt_getbuf(rel, leafrightsib, BT_READ);
1772 : 3552 : page = BufferGetPage(buf);
1773 : 3552 : opaque = BTPageGetOpaque(page);
1774 : :
1775 : : Assert(P_ISLEAF(opaque) && !P_ISDELETED(opaque));
1776 : 3552 : result = P_ISHALFDEAD(opaque);
1777 : 3552 : _bt_relbuf(rel, buf);
1778 : :
1779 : 3552 : return result;
1780 : : }
1781 : :
1782 : : /*
1783 : : * _bt_pagedel() -- Delete a leaf page from the b-tree, if legal to do so.
1784 : : *
1785 : : * This action unlinks the leaf page from the b-tree structure, removing all
1786 : : * pointers leading to it --- but not touching its own left and right links.
1787 : : * The page cannot be physically reclaimed right away, since other processes
1788 : : * may currently be trying to follow links leading to the page; they have to
1789 : : * be allowed to use its right-link to recover. See nbtree/README.
1790 : : *
1791 : : * On entry, the target buffer must be pinned and locked (either read or write
1792 : : * lock is OK). The page must be an empty leaf page, which may be half-dead
1793 : : * already (a half-dead page should only be passed to us when an earlier
1794 : : * VACUUM operation was interrupted, though). Note in particular that caller
1795 : : * should never pass a buffer containing an existing deleted page here. The
1796 : : * lock and pin on caller's buffer will be dropped before we return.
1797 : : *
1798 : : * Maintains bulk delete stats for caller, which are taken from vstate. We
1799 : : * need to cooperate closely with caller here so that whole VACUUM operation
1800 : : * reliably avoids any double counting of subsidiary-to-leafbuf pages that we
1801 : : * delete in passing. If such pages happen to be from a block number that is
1802 : : * ahead of the current scanblkno position, then caller is expected to count
1803 : : * them directly later on. It's simpler for us to understand caller's
1804 : : * requirements than it would be for caller to understand when or how a
1805 : : * deleted page became deleted after the fact.
1806 : : *
1807 : : * NOTE: this leaks memory. Rather than trying to clean up everything
1808 : : * carefully, it's better to run it in a temp context that can be reset
1809 : : * frequently.
1810 : : */
1811 : : void
1812 : 3674 : _bt_pagedel(Relation rel, Buffer leafbuf, BTVacState *vstate)
1813 : : {
1814 : : BlockNumber rightsib;
1815 : : bool rightsib_empty;
1816 : : Page page;
1817 : : BTPageOpaque opaque;
1818 : :
1819 : : /*
1820 : : * Save original leafbuf block number from caller. Only deleted blocks
1821 : : * that are <= scanblkno are added to bulk delete stat's pages_deleted
1822 : : * count.
1823 : : */
1824 : 3674 : BlockNumber scanblkno = BufferGetBlockNumber(leafbuf);
1825 : :
1826 : : /*
1827 : : * "stack" is a search stack leading (approximately) to the target page.
1828 : : * It is initially NULL, but when iterating, we keep it to avoid
1829 : : * duplicated search effort.
1830 : : *
1831 : : * Also, when "stack" is not NULL, we have already checked that the
1832 : : * current page is not the right half of an incomplete split, i.e. the
1833 : : * left sibling does not have its INCOMPLETE_SPLIT flag set, including
1834 : : * when the current target page is to the right of caller's initial page
1835 : : * (the scanblkno page).
1836 : : */
1837 : 3674 : BTStack stack = NULL;
1838 : :
1839 : : for (;;)
1840 : : {
1841 : 7231 : page = BufferGetPage(leafbuf);
1842 : 7231 : opaque = BTPageGetOpaque(page);
1843 : :
1844 : : /*
1845 : : * Internal pages are never deleted directly, only as part of deleting
1846 : : * the whole subtree all the way down to leaf level.
1847 : : *
1848 : : * Also check for deleted pages here. Caller never passes us a fully
1849 : : * deleted page. Only VACUUM can delete pages, so there can't have
1850 : : * been a concurrent deletion. Assume that we reached any deleted
1851 : : * page encountered here by following a sibling link, and that the
1852 : : * index is corrupt.
1853 : : */
1854 : : Assert(!P_ISDELETED(opaque));
1855 [ + - - + ]: 7231 : if (!P_ISLEAF(opaque) || P_ISDELETED(opaque))
1856 : : {
1857 : : /*
1858 : : * Pre-9.4 page deletion only marked internal pages as half-dead,
1859 : : * but now we only use that flag on leaf pages. The old algorithm
1860 : : * was never supposed to leave half-dead pages in the tree, it was
1861 : : * just a transient state, but it was nevertheless possible in
1862 : : * error scenarios. We don't know how to deal with them here. They
1863 : : * are harmless as far as searches are considered, but inserts
1864 : : * into the deleted keyspace could add out-of-order downlinks in
1865 : : * the upper levels. Log a notice, hopefully the admin will notice
1866 : : * and reindex.
1867 : : */
1868 [ # # ]: 0 : if (P_ISHALFDEAD(opaque))
1869 [ # # ]: 0 : ereport(LOG,
1870 : : (errcode(ERRCODE_INDEX_CORRUPTED),
1871 : : errmsg("index \"%s\" contains a half-dead internal page",
1872 : : RelationGetRelationName(rel)),
1873 : : errhint("This can be caused by an interrupted VACUUM in version 9.3 or older, before upgrade. Please REINDEX it.")));
1874 : :
1875 [ # # ]: 0 : if (P_ISDELETED(opaque))
1876 [ # # ]: 0 : ereport(LOG,
1877 : : (errcode(ERRCODE_INDEX_CORRUPTED),
1878 : : errmsg_internal("found deleted block %u while following right link from block %u in index \"%s\"",
1879 : : BufferGetBlockNumber(leafbuf),
1880 : : scanblkno,
1881 : : RelationGetRelationName(rel))));
1882 : :
1883 : 0 : _bt_relbuf(rel, leafbuf);
1884 : 137 : return;
1885 : : }
1886 : :
1887 : : /*
1888 : : * We can never delete rightmost pages nor root pages. While at it,
1889 : : * check that page is empty, since it's possible that the leafbuf page
1890 : : * was empty a moment ago, but has since had some inserts.
1891 : : *
1892 : : * To keep the algorithm simple, we also never delete an incompletely
1893 : : * split page (they should be rare enough that this doesn't make any
1894 : : * meaningful difference to disk usage):
1895 : : *
1896 : : * The INCOMPLETE_SPLIT flag on the page tells us if the page is the
1897 : : * left half of an incomplete split, but ensuring that it's not the
1898 : : * right half is more complicated. For that, we have to check that
1899 : : * the left sibling doesn't have its INCOMPLETE_SPLIT flag set using
1900 : : * _bt_leftsib_splitflag(). On the first iteration, we temporarily
1901 : : * release the lock on scanblkno/leafbuf, check the left sibling, and
1902 : : * construct a search stack to scanblkno. On subsequent iterations,
1903 : : * we know we stepped right from a page that passed these tests, so
1904 : : * it's OK.
1905 : : */
1906 [ + + + - ]: 7231 : if (P_RIGHTMOST(opaque) || P_ISROOT(opaque) ||
1907 [ - + + - ]: 7099 : P_FIRSTDATAKEY(opaque) <= PageGetMaxOffsetNumber(page) ||
1908 [ - + ]: 7099 : P_INCOMPLETE_SPLIT(opaque))
1909 : : {
1910 : : /* Should never fail to delete a half-dead page */
1911 : : Assert(!P_ISHALFDEAD(opaque));
1912 : :
1913 : 132 : _bt_relbuf(rel, leafbuf);
1914 : 132 : return;
1915 : : }
1916 : :
1917 : : /*
1918 : : * First, remove downlink pointing to the page (or a parent of the
1919 : : * page, if we are going to delete a taller subtree), and mark the
1920 : : * leafbuf page half-dead
1921 : : */
1922 [ + - ]: 7099 : if (!P_ISHALFDEAD(opaque))
1923 : : {
1924 : : /*
1925 : : * We need an approximate pointer to the page's parent page. We
1926 : : * use a variant of the standard search mechanism to search for
1927 : : * the page's high key; this will give us a link to either the
1928 : : * current parent or someplace to its left (if there are multiple
1929 : : * equal high keys, which is possible with !heapkeyspace indexes).
1930 : : *
1931 : : * Also check if this is the right-half of an incomplete split
1932 : : * (see comment above).
1933 : : */
1934 [ + + ]: 7099 : if (!stack)
1935 : 3547 : {
1936 : : BTScanInsert itup_key;
1937 : : ItemId itemid;
1938 : : IndexTuple targetkey;
1939 : : BlockNumber leftsib,
1940 : : leafblkno;
1941 : : Buffer sleafbuf;
1942 : :
1943 : 3547 : itemid = PageGetItemId(page, P_HIKEY);
1944 : 3547 : targetkey = CopyIndexTuple((IndexTuple) PageGetItem(page, itemid));
1945 : :
1946 : 3547 : leftsib = opaque->btpo_prev;
1947 : 3547 : leafblkno = BufferGetBlockNumber(leafbuf);
1948 : :
1949 : : /*
1950 : : * To avoid deadlocks, we'd better drop the leaf page lock
1951 : : * before going further.
1952 : : */
1953 : 3547 : _bt_unlockbuf(rel, leafbuf);
1954 : :
1955 : : /*
1956 : : * Check that the left sibling of leafbuf (if any) is not
1957 : : * marked with INCOMPLETE_SPLIT flag before proceeding
1958 : : */
1959 : : Assert(leafblkno == scanblkno);
1960 [ - + ]: 3547 : if (_bt_leftsib_splitflag(rel, leftsib, leafblkno))
1961 : : {
1962 : 0 : ReleaseBuffer(leafbuf);
1963 : 0 : return;
1964 : : }
1965 : :
1966 : : /*
1967 : : * We need an insertion scan key, so build one.
1968 : : *
1969 : : * _bt_search searches for the leaf page that contains any
1970 : : * matching non-pivot tuples, but we need it to "search" for
1971 : : * the high key pivot from the page that we're set to delete.
1972 : : * Compensate for the mismatch by having _bt_search locate the
1973 : : * last position < equal-to-untruncated-prefix non-pivots.
1974 : : */
1975 : 3547 : itup_key = _bt_mkscankey(rel, targetkey);
1976 : :
1977 : : /* Set up a BTLessStrategyNumber-like insertion scan key */
1978 : 3547 : itup_key->nextkey = false;
1979 : 3547 : itup_key->backward = true;
1980 : 3547 : stack = _bt_search(rel, NULL, itup_key, &sleafbuf, BT_READ, true);
1981 : : /* won't need a second lock or pin on leafbuf */
1982 : 3547 : _bt_relbuf(rel, sleafbuf);
1983 : :
1984 : : /*
1985 : : * Re-lock the leaf page, and start over to use our stack
1986 : : * within _bt_mark_page_halfdead. We must do it that way
1987 : : * because it's possible that leafbuf can no longer be
1988 : : * deleted. We need to recheck.
1989 : : *
1990 : : * Note: We can't simply hold on to the sleafbuf lock instead,
1991 : : * because it's barely possible that sleafbuf is not the same
1992 : : * page as leafbuf. This happens when leafbuf split after our
1993 : : * original lock was dropped, but before _bt_search finished
1994 : : * its descent. We rely on the assumption that we'll find
1995 : : * leafbuf isn't safe to delete anymore in this scenario.
1996 : : * (Page deletion can cope with the stack being to the left of
1997 : : * leafbuf, but not to the right of leafbuf.)
1998 : : */
1999 : 3547 : _bt_lockbuf(rel, leafbuf, BT_WRITE);
2000 : 3547 : continue;
2001 : : }
2002 : :
2003 : : /*
2004 : : * See if it's safe to delete the leaf page, and determine how
2005 : : * many parent/internal pages above the leaf level will be
2006 : : * deleted. If it's safe then _bt_mark_page_halfdead will also
2007 : : * perform the first phase of deletion, which includes marking the
2008 : : * leafbuf page half-dead.
2009 : : */
2010 : : Assert(P_ISLEAF(opaque) && !P_IGNORE(opaque));
2011 [ + + ]: 3552 : if (!_bt_mark_page_halfdead(rel, vstate->info->heaprel, leafbuf,
2012 : : stack))
2013 : : {
2014 : 5 : _bt_relbuf(rel, leafbuf);
2015 : 5 : return;
2016 : : }
2017 : : }
2018 : : else
2019 : : {
2020 : 0 : INJECTION_POINT("nbtree-finish-half-dead-page-vacuum", NULL);
2021 : : }
2022 : :
2023 : : /*
2024 : : * Then unlink it from its siblings. Each call to
2025 : : * _bt_unlink_halfdead_page unlinks the topmost page from the subtree,
2026 : : * making it shallower. Iterate until the leafbuf page is deleted.
2027 : : */
2028 : 3547 : rightsib_empty = false;
2029 : : Assert(P_ISLEAF(opaque) && P_ISHALFDEAD(opaque));
2030 [ + + ]: 7255 : while (P_ISHALFDEAD(opaque))
2031 : : {
2032 : : /* Check for interrupts in _bt_unlink_halfdead_page */
2033 [ - + ]: 3708 : if (!_bt_unlink_halfdead_page(rel, leafbuf, scanblkno,
2034 : : &rightsib_empty, vstate))
2035 : : {
2036 : : /*
2037 : : * _bt_unlink_halfdead_page should never fail, since we
2038 : : * established that deletion is generally safe in
2039 : : * _bt_mark_page_halfdead -- index must be corrupt.
2040 : : *
2041 : : * Note that _bt_unlink_halfdead_page already released the
2042 : : * lock and pin on leafbuf for us.
2043 : : */
2044 : : Assert(false);
2045 : 0 : return;
2046 : : }
2047 : : }
2048 : :
2049 : : Assert(P_ISLEAF(opaque) && P_ISDELETED(opaque));
2050 : :
2051 : 3547 : rightsib = opaque->btpo_next;
2052 : :
2053 : 3547 : _bt_relbuf(rel, leafbuf);
2054 : :
2055 : : /*
2056 : : * Check here, as calling loops will have locks held, preventing
2057 : : * interrupts from being processed.
2058 : : */
2059 [ - + ]: 3547 : CHECK_FOR_INTERRUPTS();
2060 : :
2061 : : /*
2062 : : * The page has now been deleted. If its right sibling is completely
2063 : : * empty, it's possible that the reason we haven't deleted it earlier
2064 : : * is that it was the rightmost child of the parent. Now that we
2065 : : * removed the downlink for this page, the right sibling might now be
2066 : : * the only child of the parent, and could be removed. It would be
2067 : : * picked up by the next vacuum anyway, but might as well try to
2068 : : * remove it now, so loop back to process the right sibling.
2069 : : *
2070 : : * Note: This relies on the assumption that _bt_getstackbuf() will be
2071 : : * able to reuse our original descent stack with a different child
2072 : : * block (provided that the child block is to the right of the
2073 : : * original leaf page reached by _bt_search()). It will even update
2074 : : * the descent stack each time we loop around, avoiding repeated work.
2075 : : */
2076 [ + + ]: 3547 : if (!rightsib_empty)
2077 : 3537 : break;
2078 : :
2079 : 10 : leafbuf = _bt_getbuf(rel, rightsib, BT_WRITE);
2080 : : }
2081 : : }
2082 : :
2083 : : /*
2084 : : * First stage of page deletion.
2085 : : *
2086 : : * Establish the height of the to-be-deleted subtree with leafbuf at its
2087 : : * lowest level, remove the downlink to the subtree, and mark leafbuf
2088 : : * half-dead. The final to-be-deleted subtree is usually just leafbuf itself,
2089 : : * but may include additional internal pages (at most one per level of the
2090 : : * tree below the root).
2091 : : *
2092 : : * Caller must pass a valid heaprel, since it's just about possible that our
2093 : : * call to _bt_lock_subtree_parent will need to allocate a new index page to
2094 : : * complete a page split. Every call to _bt_allocbuf needs to pass a heaprel.
2095 : : *
2096 : : * Returns 'false' if leafbuf is unsafe to delete, usually because leafbuf is
2097 : : * the rightmost child of its parent (and parent has more than one downlink).
2098 : : * Returns 'true' when the first stage of page deletion completed
2099 : : * successfully.
2100 : : */
2101 : : static bool
2102 : 3552 : _bt_mark_page_halfdead(Relation rel, Relation heaprel, Buffer leafbuf,
2103 : : BTStack stack)
2104 : : {
2105 : : BlockNumber leafblkno;
2106 : : BlockNumber leafrightsib;
2107 : : BlockNumber topparent;
2108 : : BlockNumber topparentrightsib;
2109 : : ItemId itemid;
2110 : : Page page;
2111 : : BTPageOpaque opaque;
2112 : : Buffer subtreeparent;
2113 : : OffsetNumber poffset;
2114 : : OffsetNumber nextoffset;
2115 : : IndexTuple itup;
2116 : : IndexTupleData trunctuple;
2117 : : XLogRecPtr recptr;
2118 : :
2119 : 3552 : page = BufferGetPage(leafbuf);
2120 : 3552 : opaque = BTPageGetOpaque(page);
2121 : :
2122 : : Assert(!P_RIGHTMOST(opaque) && !P_ISROOT(opaque) &&
2123 : : P_ISLEAF(opaque) && !P_IGNORE(opaque) &&
2124 : : P_FIRSTDATAKEY(opaque) > PageGetMaxOffsetNumber(page));
2125 : : Assert(heaprel != NULL);
2126 : :
2127 : : /*
2128 : : * Save info about the leaf page.
2129 : : */
2130 : 3552 : leafblkno = BufferGetBlockNumber(leafbuf);
2131 : 3552 : leafrightsib = opaque->btpo_next;
2132 : :
2133 : : /*
2134 : : * Before attempting to lock the parent page, check that the right sibling
2135 : : * is not in half-dead state. A half-dead right sibling would have no
2136 : : * downlink in the parent, which would be highly confusing later when we
2137 : : * delete the downlink. It would fail the "right sibling of target page
2138 : : * is also the next child in parent page" cross-check below.
2139 : : */
2140 [ - + ]: 3552 : if (_bt_rightsib_halfdeadflag(rel, leafrightsib))
2141 : : {
2142 [ # # ]: 0 : elog(DEBUG1, "could not delete page %u because its right sibling %u is half-dead",
2143 : : leafblkno, leafrightsib);
2144 : 0 : return false;
2145 : : }
2146 : :
2147 : : /*
2148 : : * We cannot delete a page that is the rightmost child of its immediate
2149 : : * parent, unless it is the only child --- in which case the parent has to
2150 : : * be deleted too, and the same condition applies recursively to it. We
2151 : : * have to check this condition all the way up before trying to delete,
2152 : : * and lock the parent of the root of the to-be-deleted subtree (the
2153 : : * "subtree parent"). _bt_lock_subtree_parent() locks the subtree parent
2154 : : * for us. We remove the downlink to the "top parent" page (subtree root
2155 : : * page) from the subtree parent page below.
2156 : : *
2157 : : * Initialize topparent to be leafbuf page now. The final to-be-deleted
2158 : : * subtree is often a degenerate one page subtree consisting only of the
2159 : : * leafbuf page. When that happens, the leafbuf page is the final subtree
2160 : : * root page/top parent page.
2161 : : */
2162 : 3552 : topparent = leafblkno;
2163 : 3552 : topparentrightsib = leafrightsib;
2164 [ + + ]: 3552 : if (!_bt_lock_subtree_parent(rel, heaprel, leafblkno, stack,
2165 : : &subtreeparent, &poffset,
2166 : : &topparent, &topparentrightsib))
2167 : 5 : return false;
2168 : :
2169 : 3547 : page = BufferGetPage(subtreeparent);
2170 : 3547 : opaque = BTPageGetOpaque(page);
2171 : :
2172 : : #ifdef USE_ASSERT_CHECKING
2173 : :
2174 : : /*
2175 : : * This is just an assertion because _bt_lock_subtree_parent should have
2176 : : * guaranteed tuple has the expected contents
2177 : : */
2178 : : itemid = PageGetItemId(page, poffset);
2179 : : itup = (IndexTuple) PageGetItem(page, itemid);
2180 : : Assert(BTreeTupleGetDownLink(itup) == topparent);
2181 : : #endif
2182 : :
2183 : 3547 : nextoffset = OffsetNumberNext(poffset);
2184 : 3547 : itemid = PageGetItemId(page, nextoffset);
2185 : 3547 : itup = (IndexTuple) PageGetItem(page, itemid);
2186 : :
2187 : : /*
2188 : : * Check that the parent-page index items we're about to delete/overwrite
2189 : : * in subtree parent page contain what we expect. This can fail if the
2190 : : * index has become corrupt for some reason. When that happens we back
2191 : : * out of deletion of the leafbuf subtree. (This is just like the case
2192 : : * where _bt_lock_subtree_parent() cannot "re-find" leafbuf's downlink.)
2193 : : */
2194 [ - + ]: 3547 : if (BTreeTupleGetDownLink(itup) != topparentrightsib)
2195 : : {
2196 [ # # ]: 0 : ereport(LOG,
2197 : : (errcode(ERRCODE_INDEX_CORRUPTED),
2198 : : errmsg_internal("right sibling %u of block %u is not next child %u of block %u in index \"%s\"",
2199 : : topparentrightsib, topparent,
2200 : : BTreeTupleGetDownLink(itup),
2201 : : BufferGetBlockNumber(subtreeparent),
2202 : : RelationGetRelationName(rel))));
2203 : :
2204 : 0 : _bt_relbuf(rel, subtreeparent);
2205 : : Assert(false);
2206 : 0 : return false;
2207 : : }
2208 : :
2209 : : /*
2210 : : * Any insert which would have gone on the leaf block will now go to its
2211 : : * right sibling. In other words, the key space moves right.
2212 : : */
2213 : 3547 : PredicateLockPageCombine(rel, leafblkno, leafrightsib);
2214 : :
2215 : : /* No ereport(ERROR) until changes are logged */
2216 : 3547 : START_CRIT_SECTION();
2217 : :
2218 : : /*
2219 : : * Update parent of subtree. We want to delete the downlink to the top
2220 : : * parent page/root of the subtree, and the *following* key. Easiest way
2221 : : * is to copy the right sibling's downlink over the downlink that points
2222 : : * to top parent page, and then delete the right sibling's original pivot
2223 : : * tuple.
2224 : : *
2225 : : * Lanin and Shasha make the key space move left when deleting a page,
2226 : : * whereas the key space moves right here. That's why we cannot simply
2227 : : * delete the pivot tuple with the downlink to the top parent page. See
2228 : : * nbtree/README.
2229 : : */
2230 : 3547 : page = BufferGetPage(subtreeparent);
2231 : 3547 : opaque = BTPageGetOpaque(page);
2232 : :
2233 : 3547 : itemid = PageGetItemId(page, poffset);
2234 : 3547 : itup = (IndexTuple) PageGetItem(page, itemid);
2235 : 3547 : BTreeTupleSetDownLink(itup, topparentrightsib);
2236 : :
2237 : 3547 : nextoffset = OffsetNumberNext(poffset);
2238 : 3547 : PageIndexTupleDelete(page, nextoffset);
2239 : :
2240 : : /*
2241 : : * Mark the leaf page as half-dead, and stamp it with a link to the top
2242 : : * parent page. When the leaf page is also the top parent page, the link
2243 : : * is set to InvalidBlockNumber.
2244 : : */
2245 : 3547 : page = BufferGetPage(leafbuf);
2246 : 3547 : opaque = BTPageGetOpaque(page);
2247 : 3547 : opaque->btpo_flags |= BTP_HALF_DEAD;
2248 : :
2249 : : Assert(PageGetMaxOffsetNumber(page) == P_HIKEY);
2250 [ - + - - : 3547 : MemSet(&trunctuple, 0, sizeof(IndexTupleData));
- - - - -
- ]
2251 : 3547 : trunctuple.t_info = sizeof(IndexTupleData);
2252 [ + + ]: 3547 : if (topparent != leafblkno)
2253 : 76 : BTreeTupleSetTopParent(&trunctuple, topparent);
2254 : : else
2255 : 3471 : BTreeTupleSetTopParent(&trunctuple, InvalidBlockNumber);
2256 : :
2257 [ - + ]: 3547 : if (!PageIndexTupleOverwrite(page, P_HIKEY, &trunctuple, IndexTupleSize(&trunctuple)))
2258 [ # # ]: 0 : elog(ERROR, "could not overwrite high key in half-dead page");
2259 : :
2260 : : /* Must mark buffers dirty before XLogInsert */
2261 : 3547 : MarkBufferDirty(subtreeparent);
2262 : 3547 : MarkBufferDirty(leafbuf);
2263 : :
2264 : : /* XLOG stuff */
2265 [ + - + + : 3547 : if (RelationNeedsWAL(rel))
+ - + - ]
2266 : 3547 : {
2267 : : xl_btree_mark_page_halfdead xlrec;
2268 : :
2269 : 3547 : xlrec.poffset = poffset;
2270 : 3547 : xlrec.leafblk = leafblkno;
2271 [ + + ]: 3547 : if (topparent != leafblkno)
2272 : 76 : xlrec.topparent = topparent;
2273 : : else
2274 : 3471 : xlrec.topparent = InvalidBlockNumber;
2275 : :
2276 : 3547 : XLogBeginInsert();
2277 : 3547 : XLogRegisterBuffer(0, leafbuf, REGBUF_WILL_INIT);
2278 : 3547 : XLogRegisterBuffer(1, subtreeparent, REGBUF_STANDARD);
2279 : :
2280 : 3547 : page = BufferGetPage(leafbuf);
2281 : 3547 : opaque = BTPageGetOpaque(page);
2282 : 3547 : xlrec.leftblk = opaque->btpo_prev;
2283 : 3547 : xlrec.rightblk = opaque->btpo_next;
2284 : :
2285 : 3547 : XLogRegisterData(&xlrec, SizeOfBtreeMarkPageHalfDead);
2286 : :
2287 : 3547 : recptr = XLogInsert(RM_BTREE_ID, XLOG_BTREE_MARK_PAGE_HALFDEAD);
2288 : : }
2289 : : else
2290 : 0 : recptr = XLogGetFakeLSN(rel);
2291 : :
2292 : 3547 : page = BufferGetPage(subtreeparent);
2293 : 3547 : PageSetLSN(page, recptr);
2294 : 3547 : page = BufferGetPage(leafbuf);
2295 : 3547 : PageSetLSN(page, recptr);
2296 : :
2297 : 3547 : END_CRIT_SECTION();
2298 : :
2299 : 3547 : _bt_relbuf(rel, subtreeparent);
2300 : 3547 : return true;
2301 : : }
2302 : :
2303 : : /*
2304 : : * Second stage of page deletion.
2305 : : *
2306 : : * Unlinks a single page (in the subtree undergoing deletion) from its
2307 : : * siblings. Also marks the page deleted.
2308 : : *
2309 : : * To get rid of the whole subtree, including the leaf page itself, call here
2310 : : * until the leaf page is deleted. The original "top parent" established in
2311 : : * the first stage of deletion is deleted in the first call here, while the
2312 : : * leaf page is deleted in the last call here. Note that the leaf page itself
2313 : : * is often the initial top parent page.
2314 : : *
2315 : : * Returns 'false' if the page could not be unlinked (shouldn't happen). If
2316 : : * the right sibling of the current target page is empty, *rightsib_empty is
2317 : : * set to true, allowing caller to delete the target's right sibling page in
2318 : : * passing. Note that *rightsib_empty is only actually used by caller when
2319 : : * target page is leafbuf, following last call here for leafbuf/the subtree
2320 : : * containing leafbuf. (We always set *rightsib_empty for caller, just to be
2321 : : * consistent.)
2322 : : *
2323 : : * Must hold pin and lock on leafbuf at entry (read or write doesn't matter).
2324 : : * On success exit, we'll be holding pin and write lock. On failure exit,
2325 : : * we'll release both pin and lock before returning (we define it that way
2326 : : * to avoid having to reacquire a lock we already released).
2327 : : */
2328 : : static bool
2329 : 3708 : _bt_unlink_halfdead_page(Relation rel, Buffer leafbuf, BlockNumber scanblkno,
2330 : : bool *rightsib_empty, BTVacState *vstate)
2331 : : {
2332 : 3708 : BlockNumber leafblkno = BufferGetBlockNumber(leafbuf);
2333 : 3708 : IndexBulkDeleteResult *stats = vstate->stats;
2334 : : BlockNumber leafleftsib;
2335 : : BlockNumber leafrightsib;
2336 : : BlockNumber target;
2337 : : BlockNumber leftsib;
2338 : : BlockNumber rightsib;
2339 : 3708 : Buffer lbuf = InvalidBuffer;
2340 : : Buffer buf;
2341 : : Buffer rbuf;
2342 : 3708 : Buffer metabuf = InvalidBuffer;
2343 : 3708 : Page metapg = NULL;
2344 : 3708 : BTMetaPageData *metad = NULL;
2345 : : ItemId itemid;
2346 : : Page page;
2347 : : BTPageOpaque opaque;
2348 : : FullTransactionId safexid;
2349 : : bool rightsib_is_rightmost;
2350 : : uint32 targetlevel;
2351 : : IndexTuple leafhikey;
2352 : : BlockNumber leaftopparent;
2353 : : XLogRecPtr recptr;
2354 : :
2355 : 3708 : page = BufferGetPage(leafbuf);
2356 : 3708 : opaque = BTPageGetOpaque(page);
2357 : :
2358 : : Assert(P_ISLEAF(opaque) && !P_ISDELETED(opaque) && P_ISHALFDEAD(opaque));
2359 : :
2360 : : /*
2361 : : * Remember some information about the leaf page.
2362 : : */
2363 : 3708 : itemid = PageGetItemId(page, P_HIKEY);
2364 : 3708 : leafhikey = (IndexTuple) PageGetItem(page, itemid);
2365 : 3708 : target = BTreeTupleGetTopParent(leafhikey);
2366 : 3708 : leafleftsib = opaque->btpo_prev;
2367 : 3708 : leafrightsib = opaque->btpo_next;
2368 : :
2369 : 3708 : _bt_unlockbuf(rel, leafbuf);
2370 : :
2371 : 3708 : INJECTION_POINT("nbtree-leave-page-half-dead", NULL);
2372 : :
2373 : : /*
2374 : : * Check here, as calling loops will have locks held, preventing
2375 : : * interrupts from being processed.
2376 : : */
2377 [ - + ]: 3708 : CHECK_FOR_INTERRUPTS();
2378 : :
2379 : : /* Unlink the current top parent of the subtree */
2380 [ + + ]: 3708 : if (!BlockNumberIsValid(target))
2381 : : {
2382 : : /* Target is leaf page (or leaf page is top parent, if you prefer) */
2383 : 3547 : target = leafblkno;
2384 : :
2385 : 3547 : buf = leafbuf;
2386 : 3547 : leftsib = leafleftsib;
2387 : 3547 : targetlevel = 0;
2388 : : }
2389 : : else
2390 : : {
2391 : : /* Target is the internal page taken from leaf's top parent link */
2392 : : Assert(target != leafblkno);
2393 : :
2394 : : /* Fetch the block number of the target's left sibling */
2395 : 161 : buf = _bt_getbuf(rel, target, BT_READ);
2396 : 161 : page = BufferGetPage(buf);
2397 : 161 : opaque = BTPageGetOpaque(page);
2398 : 161 : leftsib = opaque->btpo_prev;
2399 : 161 : targetlevel = opaque->btpo_level;
2400 : : Assert(targetlevel > 0);
2401 : :
2402 : : /*
2403 : : * To avoid deadlocks, we'd better drop the target page lock before
2404 : : * going further.
2405 : : */
2406 : 161 : _bt_unlockbuf(rel, buf);
2407 : : }
2408 : :
2409 : : /*
2410 : : * We have to lock the pages we need to modify in the standard order:
2411 : : * moving right, then up. Else we will deadlock against other writers.
2412 : : *
2413 : : * So, first lock the leaf page, if it's not the target. Then find and
2414 : : * write-lock the current left sibling of the target page. The sibling
2415 : : * that was current a moment ago could have split, so we may have to move
2416 : : * right.
2417 : : */
2418 [ + + ]: 3708 : if (target != leafblkno)
2419 : 161 : _bt_lockbuf(rel, leafbuf, BT_WRITE);
2420 [ + + ]: 3708 : if (leftsib != P_NONE)
2421 : : {
2422 : 798 : lbuf = _bt_getbuf(rel, leftsib, BT_WRITE);
2423 : 798 : page = BufferGetPage(lbuf);
2424 : 798 : opaque = BTPageGetOpaque(page);
2425 [ - + - + ]: 798 : while (P_ISDELETED(opaque) || opaque->btpo_next != target)
2426 : : {
2427 : 0 : bool leftsibvalid = true;
2428 : :
2429 : : /*
2430 : : * Before we follow the link from the page that was the left
2431 : : * sibling mere moments ago, validate its right link. This
2432 : : * reduces the opportunities for loop to fail to ever make any
2433 : : * progress in the presence of index corruption.
2434 : : *
2435 : : * Note: we rely on the assumption that there can only be one
2436 : : * vacuum process running at a time (against the same index).
2437 : : */
2438 [ # # # # ]: 0 : if (P_RIGHTMOST(opaque) || P_ISDELETED(opaque) ||
2439 [ # # ]: 0 : leftsib == opaque->btpo_next)
2440 : 0 : leftsibvalid = false;
2441 : :
2442 : 0 : leftsib = opaque->btpo_next;
2443 : 0 : _bt_relbuf(rel, lbuf);
2444 : :
2445 [ # # ]: 0 : if (!leftsibvalid)
2446 : : {
2447 : : /*
2448 : : * This is known to fail in the field; sibling link corruption
2449 : : * is relatively common. Press on with vacuuming rather than
2450 : : * just throwing an ERROR.
2451 : : */
2452 [ # # ]: 0 : ereport(LOG,
2453 : : (errcode(ERRCODE_INDEX_CORRUPTED),
2454 : : errmsg_internal("valid left sibling for deletion target could not be located: "
2455 : : "left sibling %u of target %u with leafblkno %u and scanblkno %u on level %u of index \"%s\"",
2456 : : leftsib, target, leafblkno, scanblkno,
2457 : : targetlevel, RelationGetRelationName(rel))));
2458 : :
2459 : : /* Must release all pins and locks on failure exit */
2460 : 0 : ReleaseBuffer(buf);
2461 [ # # ]: 0 : if (target != leafblkno)
2462 : 0 : _bt_relbuf(rel, leafbuf);
2463 : :
2464 : 0 : return false;
2465 : : }
2466 : :
2467 [ # # ]: 0 : CHECK_FOR_INTERRUPTS();
2468 : :
2469 : : /* step right one page */
2470 : 0 : lbuf = _bt_getbuf(rel, leftsib, BT_WRITE);
2471 : 0 : page = BufferGetPage(lbuf);
2472 : 0 : opaque = BTPageGetOpaque(page);
2473 : : }
2474 : : }
2475 : : else
2476 : 2910 : lbuf = InvalidBuffer;
2477 : :
2478 : : /* Next write-lock the target page itself */
2479 : 3708 : _bt_lockbuf(rel, buf, BT_WRITE);
2480 : 3708 : page = BufferGetPage(buf);
2481 : 3708 : opaque = BTPageGetOpaque(page);
2482 : :
2483 : : /*
2484 : : * Check page is still empty etc, else abandon deletion. This is just for
2485 : : * paranoia's sake; a half-dead page cannot resurrect because there can be
2486 : : * only one vacuum process running at a time.
2487 : : */
2488 [ + - + - : 3708 : if (P_RIGHTMOST(opaque) || P_ISROOT(opaque) || P_ISDELETED(opaque))
- + ]
2489 [ # # ]: 0 : elog(ERROR, "target page changed status unexpectedly in block %u of index \"%s\"",
2490 : : target, RelationGetRelationName(rel));
2491 : :
2492 [ - + ]: 3708 : if (opaque->btpo_prev != leftsib)
2493 [ # # ]: 0 : ereport(ERROR,
2494 : : (errcode(ERRCODE_INDEX_CORRUPTED),
2495 : : errmsg_internal("target page left link unexpectedly changed from %u to %u in block %u of index \"%s\"",
2496 : : leftsib, opaque->btpo_prev, target,
2497 : : RelationGetRelationName(rel))));
2498 : :
2499 [ + + ]: 3708 : if (target == leafblkno)
2500 : : {
2501 [ - + + - ]: 3547 : if (P_FIRSTDATAKEY(opaque) <= PageGetMaxOffsetNumber(page) ||
2502 [ + - - + ]: 3547 : !P_ISLEAF(opaque) || !P_ISHALFDEAD(opaque))
2503 [ # # ]: 0 : elog(ERROR, "target leaf page changed status unexpectedly in block %u of index \"%s\"",
2504 : : target, RelationGetRelationName(rel));
2505 : :
2506 : : /* Leaf page is also target page: don't set leaftopparent */
2507 : 3547 : leaftopparent = InvalidBlockNumber;
2508 : : }
2509 : : else
2510 : : {
2511 : : IndexTuple finaldataitem;
2512 : :
2513 [ - + + - ]: 161 : if (P_FIRSTDATAKEY(opaque) != PageGetMaxOffsetNumber(page) ||
2514 [ - + ]: 161 : P_ISLEAF(opaque))
2515 [ # # ]: 0 : elog(ERROR, "target internal page on level %u changed status unexpectedly in block %u of index \"%s\"",
2516 : : targetlevel, target, RelationGetRelationName(rel));
2517 : :
2518 : : /* Target is internal: set leaftopparent for next call here... */
2519 [ - + ]: 161 : itemid = PageGetItemId(page, P_FIRSTDATAKEY(opaque));
2520 : 161 : finaldataitem = (IndexTuple) PageGetItem(page, itemid);
2521 : 161 : leaftopparent = BTreeTupleGetDownLink(finaldataitem);
2522 : : /* ...except when it would be a redundant pointer-to-self */
2523 [ + + ]: 161 : if (leaftopparent == leafblkno)
2524 : 76 : leaftopparent = InvalidBlockNumber;
2525 : : }
2526 : :
2527 : : /* No leaftopparent for level 0 (leaf page) or level 1 target */
2528 : : Assert(!BlockNumberIsValid(leaftopparent) || targetlevel > 1);
2529 : :
2530 : : /*
2531 : : * And next write-lock the (current) right sibling.
2532 : : */
2533 : 3708 : rightsib = opaque->btpo_next;
2534 : 3708 : rbuf = _bt_getbuf(rel, rightsib, BT_WRITE);
2535 : 3708 : page = BufferGetPage(rbuf);
2536 : 3708 : opaque = BTPageGetOpaque(page);
2537 : :
2538 : : /*
2539 : : * Validate target's right sibling page. Its left link must point back to
2540 : : * the target page.
2541 : : */
2542 [ - + ]: 3708 : if (opaque->btpo_prev != target)
2543 : : {
2544 : : /*
2545 : : * This is known to fail in the field; sibling link corruption is
2546 : : * relatively common. Press on with vacuuming rather than just
2547 : : * throwing an ERROR (same approach used for left-sibling's-right-link
2548 : : * validation check a moment ago).
2549 : : */
2550 [ # # ]: 0 : ereport(LOG,
2551 : : (errcode(ERRCODE_INDEX_CORRUPTED),
2552 : : errmsg_internal("right sibling's left-link doesn't match: "
2553 : : "right sibling %u of target %u with leafblkno %u "
2554 : : "and scanblkno %u spuriously links to non-target %u "
2555 : : "on level %u of index \"%s\"",
2556 : : rightsib, target, leafblkno,
2557 : : scanblkno, opaque->btpo_prev,
2558 : : targetlevel, RelationGetRelationName(rel))));
2559 : :
2560 : : /* Must release all pins and locks on failure exit */
2561 [ # # ]: 0 : if (BufferIsValid(lbuf))
2562 : 0 : _bt_relbuf(rel, lbuf);
2563 : 0 : _bt_relbuf(rel, rbuf);
2564 : 0 : _bt_relbuf(rel, buf);
2565 [ # # ]: 0 : if (target != leafblkno)
2566 : 0 : _bt_relbuf(rel, leafbuf);
2567 : :
2568 : 0 : return false;
2569 : : }
2570 : :
2571 : 3708 : rightsib_is_rightmost = P_RIGHTMOST(opaque);
2572 [ + + ]: 3708 : *rightsib_empty = (P_FIRSTDATAKEY(opaque) > PageGetMaxOffsetNumber(page));
2573 : :
2574 : : /*
2575 : : * If we are deleting the next-to-last page on the target's level, then
2576 : : * the rightsib is a candidate to become the new fast root. (In theory, it
2577 : : * might be possible to push the fast root even further down, but the odds
2578 : : * of doing so are slim, and the locking considerations daunting.)
2579 : : *
2580 : : * We can safely acquire a lock on the metapage here --- see comments for
2581 : : * _bt_newlevel().
2582 : : */
2583 [ + + + + ]: 3708 : if (leftsib == P_NONE && rightsib_is_rightmost)
2584 : : {
2585 : 37 : page = BufferGetPage(rbuf);
2586 : 37 : opaque = BTPageGetOpaque(page);
2587 [ + - ]: 37 : if (P_RIGHTMOST(opaque))
2588 : : {
2589 : : /* rightsib will be the only one left on the level */
2590 : 37 : metabuf = _bt_getbuf(rel, BTREE_METAPAGE, BT_WRITE);
2591 : 37 : metapg = BufferGetPage(metabuf);
2592 : 37 : metad = BTPageGetMeta(metapg);
2593 : :
2594 : : /*
2595 : : * The expected case here is btm_fastlevel == targetlevel+1; if
2596 : : * the fastlevel is <= targetlevel, something is wrong, and we
2597 : : * choose to overwrite it to fix it.
2598 : : */
2599 [ - + ]: 37 : if (metad->btm_fastlevel > targetlevel + 1)
2600 : : {
2601 : : /* no update wanted */
2602 : 0 : _bt_relbuf(rel, metabuf);
2603 : 0 : metabuf = InvalidBuffer;
2604 : : }
2605 : : }
2606 : : }
2607 : :
2608 : : /*
2609 : : * Here we begin doing the deletion.
2610 : : */
2611 : :
2612 : : /* No ereport(ERROR) until changes are logged */
2613 : 3708 : START_CRIT_SECTION();
2614 : :
2615 : : /*
2616 : : * Update siblings' side-links. Note the target page's side-links will
2617 : : * continue to point to the siblings. Asserts here are just rechecking
2618 : : * things we already verified above.
2619 : : */
2620 [ + + ]: 3708 : if (BufferIsValid(lbuf))
2621 : : {
2622 : 798 : page = BufferGetPage(lbuf);
2623 : 798 : opaque = BTPageGetOpaque(page);
2624 : : Assert(opaque->btpo_next == target);
2625 : 798 : opaque->btpo_next = rightsib;
2626 : : }
2627 : 3708 : page = BufferGetPage(rbuf);
2628 : 3708 : opaque = BTPageGetOpaque(page);
2629 : : Assert(opaque->btpo_prev == target);
2630 : 3708 : opaque->btpo_prev = leftsib;
2631 : :
2632 : : /*
2633 : : * If we deleted a parent of the targeted leaf page, instead of the leaf
2634 : : * itself, update the leaf to point to the next remaining child in the
2635 : : * subtree.
2636 : : *
2637 : : * Note: We rely on the fact that a buffer pin on the leaf page has been
2638 : : * held since leafhikey was initialized. This is safe, though only
2639 : : * because the page was already half-dead at that point. The leaf page
2640 : : * cannot have been modified by any other backend during the period when
2641 : : * no lock was held.
2642 : : */
2643 [ + + ]: 3708 : if (target != leafblkno)
2644 : 161 : BTreeTupleSetTopParent(leafhikey, leaftopparent);
2645 : :
2646 : : /*
2647 : : * Mark the page itself deleted. It can be recycled when all current
2648 : : * transactions are gone. Storing GetTopTransactionId() would work, but
2649 : : * we're in VACUUM and would not otherwise have an XID. Having already
2650 : : * updated links to the target, ReadNextFullTransactionId() suffices as an
2651 : : * upper bound. Any scan having retained a now-stale link is advertising
2652 : : * in its PGPROC an xmin less than or equal to the value we read here. It
2653 : : * will continue to do so, holding back the xmin horizon, for the duration
2654 : : * of that scan.
2655 : : */
2656 : 3708 : page = BufferGetPage(buf);
2657 : 3708 : opaque = BTPageGetOpaque(page);
2658 : : Assert(P_ISHALFDEAD(opaque) || !P_ISLEAF(opaque));
2659 : :
2660 : : /*
2661 : : * Store upper bound XID that's used to determine when deleted page is no
2662 : : * longer needed as a tombstone
2663 : : */
2664 : 3708 : safexid = ReadNextFullTransactionId();
2665 : 3708 : BTPageSetDeleted(page, safexid);
2666 : 3708 : opaque->btpo_cycleid = 0;
2667 : :
2668 : : /* And update the metapage, if needed */
2669 [ + + ]: 3708 : if (BufferIsValid(metabuf))
2670 : : {
2671 : : /* upgrade metapage if needed */
2672 [ - + ]: 37 : if (metad->btm_version < BTREE_NOVAC_VERSION)
2673 : 0 : _bt_upgrademetapage(metapg);
2674 : 37 : metad->btm_fastroot = rightsib;
2675 : 37 : metad->btm_fastlevel = targetlevel;
2676 : 37 : MarkBufferDirty(metabuf);
2677 : : }
2678 : :
2679 : : /* Must mark buffers dirty before XLogInsert */
2680 : 3708 : MarkBufferDirty(rbuf);
2681 : 3708 : MarkBufferDirty(buf);
2682 [ + + ]: 3708 : if (BufferIsValid(lbuf))
2683 : 798 : MarkBufferDirty(lbuf);
2684 [ + + ]: 3708 : if (target != leafblkno)
2685 : 161 : MarkBufferDirty(leafbuf);
2686 : :
2687 : : /* XLOG stuff */
2688 [ + - + + : 3708 : if (RelationNeedsWAL(rel))
+ - + - ]
2689 : 3708 : {
2690 : : xl_btree_unlink_page xlrec;
2691 : : xl_btree_metadata xlmeta;
2692 : : uint8 xlinfo;
2693 : :
2694 : 3708 : XLogBeginInsert();
2695 : :
2696 : 3708 : XLogRegisterBuffer(0, buf, REGBUF_WILL_INIT);
2697 [ + + ]: 3708 : if (BufferIsValid(lbuf))
2698 : 798 : XLogRegisterBuffer(1, lbuf, REGBUF_STANDARD);
2699 : 3708 : XLogRegisterBuffer(2, rbuf, REGBUF_STANDARD);
2700 [ + + ]: 3708 : if (target != leafblkno)
2701 : 161 : XLogRegisterBuffer(3, leafbuf, REGBUF_WILL_INIT);
2702 : :
2703 : : /* information stored on the target/to-be-unlinked block */
2704 : 3708 : xlrec.leftsib = leftsib;
2705 : 3708 : xlrec.rightsib = rightsib;
2706 : 3708 : xlrec.level = targetlevel;
2707 : 3708 : xlrec.safexid = safexid;
2708 : :
2709 : : /* information needed to recreate the leaf block (if not the target) */
2710 : 3708 : xlrec.leafleftsib = leafleftsib;
2711 : 3708 : xlrec.leafrightsib = leafrightsib;
2712 : 3708 : xlrec.leaftopparent = leaftopparent;
2713 : :
2714 : 3708 : XLogRegisterData(&xlrec, SizeOfBtreeUnlinkPage);
2715 : :
2716 [ + + ]: 3708 : if (BufferIsValid(metabuf))
2717 : : {
2718 : 37 : XLogRegisterBuffer(4, metabuf, REGBUF_WILL_INIT | REGBUF_STANDARD);
2719 : :
2720 : : Assert(metad->btm_version >= BTREE_NOVAC_VERSION);
2721 : 37 : xlmeta.version = metad->btm_version;
2722 : 37 : xlmeta.root = metad->btm_root;
2723 : 37 : xlmeta.level = metad->btm_level;
2724 : 37 : xlmeta.fastroot = metad->btm_fastroot;
2725 : 37 : xlmeta.fastlevel = metad->btm_fastlevel;
2726 : 37 : xlmeta.last_cleanup_num_delpages = metad->btm_last_cleanup_num_delpages;
2727 : 37 : xlmeta.allequalimage = metad->btm_allequalimage;
2728 : :
2729 : 37 : XLogRegisterBufData(4, &xlmeta, sizeof(xl_btree_metadata));
2730 : 37 : xlinfo = XLOG_BTREE_UNLINK_PAGE_META;
2731 : : }
2732 : : else
2733 : 3671 : xlinfo = XLOG_BTREE_UNLINK_PAGE;
2734 : :
2735 : 3708 : recptr = XLogInsert(RM_BTREE_ID, xlinfo);
2736 : : }
2737 : : else
2738 : 0 : recptr = XLogGetFakeLSN(rel);
2739 : :
2740 [ + + ]: 3708 : if (BufferIsValid(metabuf))
2741 : 37 : PageSetLSN(metapg, recptr);
2742 : 3708 : page = BufferGetPage(rbuf);
2743 : 3708 : PageSetLSN(page, recptr);
2744 : 3708 : page = BufferGetPage(buf);
2745 : 3708 : PageSetLSN(page, recptr);
2746 [ + + ]: 3708 : if (BufferIsValid(lbuf))
2747 : : {
2748 : 798 : page = BufferGetPage(lbuf);
2749 : 798 : PageSetLSN(page, recptr);
2750 : : }
2751 [ + + ]: 3708 : if (target != leafblkno)
2752 : : {
2753 : 161 : page = BufferGetPage(leafbuf);
2754 : 161 : PageSetLSN(page, recptr);
2755 : : }
2756 : :
2757 : 3708 : END_CRIT_SECTION();
2758 : :
2759 : : /* release metapage */
2760 [ + + ]: 3708 : if (BufferIsValid(metabuf))
2761 : 37 : _bt_relbuf(rel, metabuf);
2762 : :
2763 : : /* release siblings */
2764 [ + + ]: 3708 : if (BufferIsValid(lbuf))
2765 : 798 : _bt_relbuf(rel, lbuf);
2766 : 3708 : _bt_relbuf(rel, rbuf);
2767 : :
2768 : : /* If the target is not leafbuf, we're done with it now -- release it */
2769 [ + + ]: 3708 : if (target != leafblkno)
2770 : 161 : _bt_relbuf(rel, buf);
2771 : :
2772 : : /*
2773 : : * Maintain pages_newly_deleted, which is simply the number of pages
2774 : : * deleted by the ongoing VACUUM operation.
2775 : : *
2776 : : * Maintain pages_deleted in a way that takes into account how
2777 : : * btvacuumpage() will count deleted pages that have yet to become
2778 : : * scanblkno -- only count page when it's not going to get that treatment
2779 : : * later on.
2780 : : */
2781 : 3708 : stats->pages_newly_deleted++;
2782 [ + + ]: 3708 : if (target <= scanblkno)
2783 : 3565 : stats->pages_deleted++;
2784 : :
2785 : : /*
2786 : : * Remember information about the target page (now a newly deleted page)
2787 : : * in dedicated vstate space for later. The page will be considered as a
2788 : : * candidate to place in the FSM at the end of the current btvacuumscan()
2789 : : * call.
2790 : : */
2791 : 3708 : _bt_pendingfsm_add(vstate, target, safexid);
2792 : :
2793 : : /* Success - hold on to lock on leafbuf (might also have been target) */
2794 : 3708 : return true;
2795 : : }
2796 : :
2797 : : /*
2798 : : * Establish how tall the to-be-deleted subtree will be during the first stage
2799 : : * of page deletion.
2800 : : *
2801 : : * Caller's child argument is the block number of the page caller wants to
2802 : : * delete (this is leafbuf's block number, except when we're called
2803 : : * recursively). stack is a search stack leading to it. Note that we will
2804 : : * update the stack entry(s) to reflect current downlink positions --- this is
2805 : : * similar to the corresponding point in page split handling.
2806 : : *
2807 : : * If "first stage" caller cannot go ahead with deleting _any_ pages, returns
2808 : : * false. Returns true on success, in which case caller can use certain
2809 : : * details established here to perform the first stage of deletion. This
2810 : : * function is the last point at which page deletion may be deemed unsafe
2811 : : * (barring index corruption, or unexpected concurrent page deletions).
2812 : : *
2813 : : * We write lock the parent of the root of the to-be-deleted subtree for
2814 : : * caller on success (i.e. we leave our lock on the *subtreeparent buffer for
2815 : : * caller). Caller will have to remove a downlink from *subtreeparent. We
2816 : : * also set a *subtreeparent offset number in *poffset, to indicate the
2817 : : * location of the pivot tuple that contains the relevant downlink.
2818 : : *
2819 : : * The root of the to-be-deleted subtree is called the "top parent". Note
2820 : : * that the leafbuf page is often the final "top parent" page (you can think
2821 : : * of the leafbuf page as a degenerate single page subtree when that happens).
2822 : : * Caller should initialize *topparent to the target leafbuf page block number
2823 : : * (while *topparentrightsib should be set to leafbuf's right sibling block
2824 : : * number). We will update *topparent (and *topparentrightsib) for caller
2825 : : * here, though only when it turns out that caller will delete at least one
2826 : : * internal page (i.e. only when caller needs to store a valid link to the top
2827 : : * parent block in the leafbuf page using BTreeTupleSetTopParent()).
2828 : : */
2829 : : static bool
2830 : 3724 : _bt_lock_subtree_parent(Relation rel, Relation heaprel, BlockNumber child,
2831 : : BTStack stack, Buffer *subtreeparent,
2832 : : OffsetNumber *poffset, BlockNumber *topparent,
2833 : : BlockNumber *topparentrightsib)
2834 : : {
2835 : : BlockNumber parent,
2836 : : leftsibparent;
2837 : : OffsetNumber parentoffset,
2838 : : maxoff;
2839 : : Buffer pbuf;
2840 : : Page page;
2841 : : BTPageOpaque opaque;
2842 : :
2843 : : /*
2844 : : * Locate the pivot tuple whose downlink points to "child". Write lock
2845 : : * the parent page itself.
2846 : : */
2847 : 3724 : pbuf = _bt_getstackbuf(rel, heaprel, stack, child);
2848 [ - + ]: 3724 : if (pbuf == InvalidBuffer)
2849 : : {
2850 : : /*
2851 : : * Failed to "re-find" a pivot tuple whose downlink matched our child
2852 : : * block number on the parent level -- the index must be corrupt.
2853 : : * Don't even try to delete the leafbuf subtree. Just report the
2854 : : * issue and press on with vacuuming the index.
2855 : : *
2856 : : * Note: _bt_getstackbuf() recovers from concurrent page splits that
2857 : : * take place on the parent level. Its approach is a near-exhaustive
2858 : : * linear search. This also gives it a surprisingly good chance of
2859 : : * recovering in the event of a buggy or inconsistent opclass. But we
2860 : : * don't rely on that here.
2861 : : */
2862 [ # # ]: 0 : ereport(LOG,
2863 : : (errcode(ERRCODE_INDEX_CORRUPTED),
2864 : : errmsg_internal("failed to re-find parent key in index \"%s\" for deletion target page %u",
2865 : : RelationGetRelationName(rel), child)));
2866 : : Assert(false);
2867 : 0 : return false;
2868 : : }
2869 : :
2870 : 3724 : parent = stack->bts_blkno;
2871 : 3724 : parentoffset = stack->bts_offset;
2872 : :
2873 : 3724 : page = BufferGetPage(pbuf);
2874 : 3724 : opaque = BTPageGetOpaque(page);
2875 : 3724 : maxoff = PageGetMaxOffsetNumber(page);
2876 : 3724 : leftsibparent = opaque->btpo_prev;
2877 : :
2878 : : /*
2879 : : * _bt_getstackbuf() completes page splits on returned parent buffer when
2880 : : * required.
2881 : : *
2882 : : * In general it's a bad idea for VACUUM to use up more disk space, which
2883 : : * is why page deletion does not finish incomplete page splits most of the
2884 : : * time. We allow this limited exception because the risk is much lower,
2885 : : * and the potential downside of not proceeding is much higher: A single
2886 : : * internal page with the INCOMPLETE_SPLIT flag set might otherwise
2887 : : * prevent us from deleting hundreds of empty leaf pages from one level
2888 : : * down.
2889 : : */
2890 : : Assert(!P_INCOMPLETE_SPLIT(opaque));
2891 : :
2892 [ + + ]: 3724 : if (parentoffset < maxoff)
2893 : : {
2894 : : /*
2895 : : * Child is not the rightmost child in parent, so it's safe to delete
2896 : : * the subtree whose root/topparent is child page
2897 : : */
2898 : 3547 : *subtreeparent = pbuf;
2899 : 3547 : *poffset = parentoffset;
2900 : 3547 : return true;
2901 : : }
2902 : :
2903 : : /*
2904 : : * Child is the rightmost child of parent.
2905 : : *
2906 : : * Since it's the rightmost child of parent, deleting the child (or
2907 : : * deleting the subtree whose root/topparent is the child page) is only
2908 : : * safe when it's also possible to delete the parent.
2909 : : */
2910 : : Assert(parentoffset == maxoff);
2911 [ - + + + : 177 : if (parentoffset != P_FIRSTDATAKEY(opaque) || P_RIGHTMOST(opaque))
- + ]
2912 : : {
2913 : : /*
2914 : : * Child isn't parent's only child, or parent is rightmost on its
2915 : : * entire level. Definitely cannot delete any pages.
2916 : : */
2917 : 5 : _bt_relbuf(rel, pbuf);
2918 : 5 : return false;
2919 : : }
2920 : :
2921 : : /*
2922 : : * Now make sure that the parent deletion is itself safe by examining the
2923 : : * child's grandparent page. Recurse, passing the parent page as the
2924 : : * child page (child's grandparent is the parent on the next level up). If
2925 : : * parent deletion is unsafe, then child deletion must also be unsafe (in
2926 : : * which case caller cannot delete any pages at all).
2927 : : */
2928 : 172 : *topparent = parent;
2929 : 172 : *topparentrightsib = opaque->btpo_next;
2930 : :
2931 : : /*
2932 : : * Release lock on parent before recursing.
2933 : : *
2934 : : * It's OK to release page locks on parent before recursive call locks
2935 : : * grandparent. An internal page can only acquire an entry if the child
2936 : : * is split, but that cannot happen as long as we still hold a lock on the
2937 : : * leafbuf page.
2938 : : */
2939 : 172 : _bt_relbuf(rel, pbuf);
2940 : :
2941 : : /*
2942 : : * Before recursing, check that the left sibling of parent (if any) is not
2943 : : * marked with INCOMPLETE_SPLIT flag first (must do so after we drop the
2944 : : * parent lock).
2945 : : *
2946 : : * Note: We deliberately avoid completing incomplete splits here.
2947 : : */
2948 [ - + ]: 172 : if (_bt_leftsib_splitflag(rel, leftsibparent, parent))
2949 : 0 : return false;
2950 : :
2951 : : /* Recurse to examine child page's grandparent page */
2952 : 172 : return _bt_lock_subtree_parent(rel, heaprel, parent, stack->bts_parent,
2953 : : subtreeparent, poffset,
2954 : : topparent, topparentrightsib);
2955 : : }
2956 : :
2957 : : /*
2958 : : * Initialize local memory state used by VACUUM for _bt_pendingfsm_finalize
2959 : : * optimization.
2960 : : *
2961 : : * Called at the start of a btvacuumscan(). Caller's cleanuponly argument
2962 : : * indicates if ongoing VACUUM has not (and will not) call btbulkdelete().
2963 : : *
2964 : : * We expect to allocate memory inside VACUUM's top-level memory context here.
2965 : : * The working buffer is subject to a limit based on work_mem. Our strategy
2966 : : * when the array can no longer grow within the bounds of that limit is to
2967 : : * stop saving additional newly deleted pages, while proceeding as usual with
2968 : : * the pages that we can fit.
2969 : : */
2970 : : void
2971 : 1777 : _bt_pendingfsm_init(Relation rel, BTVacState *vstate, bool cleanuponly)
2972 : : {
2973 : : Size maxbufsize;
2974 : :
2975 : : /*
2976 : : * Don't bother with optimization in cleanup-only case -- we don't expect
2977 : : * any newly deleted pages. Besides, cleanup-only calls to btvacuumscan()
2978 : : * can only take place because this optimization didn't work out during
2979 : : * the last VACUUM.
2980 : : */
2981 [ + + ]: 1777 : if (cleanuponly)
2982 : 7 : return;
2983 : :
2984 : : /*
2985 : : * Cap maximum size of array so that we always respect work_mem. Avoid
2986 : : * int overflow here.
2987 : : */
2988 : 1770 : vstate->bufsize = 256;
2989 : 1770 : maxbufsize = (work_mem * (Size) 1024) / sizeof(BTPendingFSM);
2990 : 1770 : maxbufsize = Min(maxbufsize, MaxAllocSize / sizeof(BTPendingFSM));
2991 : : /* BTVacState.maxbufsize has type int */
2992 : 1770 : maxbufsize = Min(maxbufsize, INT_MAX);
2993 : : /* Stay sane with small work_mem */
2994 : 1770 : maxbufsize = Max(maxbufsize, vstate->bufsize);
2995 : 1770 : vstate->maxbufsize = (int) maxbufsize;
2996 : :
2997 : : /* Allocate buffer, indicate that there are currently 0 pending pages */
2998 : 1770 : vstate->pendingpages = palloc_array(BTPendingFSM, vstate->bufsize);
2999 : 1770 : vstate->npendingpages = 0;
3000 : : }
3001 : :
3002 : : /*
3003 : : * Place any newly deleted pages (i.e. pages that _bt_pagedel() deleted during
3004 : : * the ongoing VACUUM operation) into the free space map -- though only when
3005 : : * it is actually safe to do so by now.
3006 : : *
3007 : : * Called at the end of a btvacuumscan(), just before free space map vacuuming
3008 : : * takes place.
3009 : : *
3010 : : * Frees memory allocated by _bt_pendingfsm_init(), if any.
3011 : : */
3012 : : void
3013 : 1776 : _bt_pendingfsm_finalize(Relation rel, BTVacState *vstate)
3014 : : {
3015 : 1776 : IndexBulkDeleteResult *stats = vstate->stats;
3016 : 1776 : Relation heaprel = vstate->info->heaprel;
3017 : :
3018 : : Assert(stats->pages_newly_deleted >= vstate->npendingpages);
3019 : : Assert(heaprel != NULL);
3020 : :
3021 [ + + ]: 1776 : if (vstate->npendingpages == 0)
3022 : : {
3023 : : /* Just free memory when nothing to do */
3024 [ + + ]: 1698 : if (vstate->pendingpages)
3025 : 1691 : pfree(vstate->pendingpages);
3026 : :
3027 : 1698 : return;
3028 : : }
3029 : :
3030 : : #ifdef DEBUG_BTREE_PENDING_FSM
3031 : :
3032 : : /*
3033 : : * Debugging aid: Sleep for 5 seconds to greatly increase the chances of
3034 : : * placing pending pages in the FSM. Note that the optimization will
3035 : : * never be effective without some other backend concurrently consuming an
3036 : : * XID.
3037 : : */
3038 : : pg_usleep(5000000L);
3039 : : #endif
3040 : :
3041 : : /*
3042 : : * Recompute VACUUM XID boundaries.
3043 : : *
3044 : : * We don't actually care about the oldest non-removable XID. Computing
3045 : : * the oldest such XID has a useful side-effect that we rely on: it
3046 : : * forcibly updates the XID horizon state for this backend. This step is
3047 : : * essential; GlobalVisCheckRemovableFullXid() will not reliably recognize
3048 : : * that it is now safe to recycle newly deleted pages without this step.
3049 : : */
3050 : 78 : GetOldestNonRemovableTransactionId(heaprel);
3051 : :
3052 [ + + ]: 109 : for (unsigned int i = 0; i < vstate->npendingpages; i++)
3053 : : {
3054 : 108 : BlockNumber target = vstate->pendingpages[i].target;
3055 : 108 : FullTransactionId safexid = vstate->pendingpages[i].safexid;
3056 : :
3057 : : /*
3058 : : * Do the equivalent of checking BTPageIsRecyclable(), but without
3059 : : * accessing the page again a second time.
3060 : : *
3061 : : * Give up on finding the first non-recyclable page -- all later pages
3062 : : * must be non-recyclable too, since _bt_pendingfsm_add() adds pages
3063 : : * to the array in safexid order.
3064 : : */
3065 [ + + ]: 108 : if (!GlobalVisCheckRemovableFullXid(heaprel, safexid))
3066 : 77 : break;
3067 : :
3068 : 31 : RecordFreeIndexPage(rel, target);
3069 : 31 : stats->pages_free++;
3070 : : }
3071 : :
3072 : 78 : pfree(vstate->pendingpages);
3073 : : }
3074 : :
3075 : : /*
3076 : : * Maintain array of pages that were deleted during current btvacuumscan()
3077 : : * call, for use in _bt_pendingfsm_finalize()
3078 : : */
3079 : : static void
3080 : 3708 : _bt_pendingfsm_add(BTVacState *vstate,
3081 : : BlockNumber target,
3082 : : FullTransactionId safexid)
3083 : : {
3084 : : Assert(vstate->npendingpages <= vstate->bufsize);
3085 : : Assert(vstate->bufsize <= vstate->maxbufsize);
3086 : :
3087 : : #ifdef USE_ASSERT_CHECKING
3088 : :
3089 : : /*
3090 : : * Verify an assumption made by _bt_pendingfsm_finalize(): pages from the
3091 : : * array will always be in safexid order (since that is the order that we
3092 : : * save them in here)
3093 : : */
3094 : : if (vstate->npendingpages > 0)
3095 : : {
3096 : : FullTransactionId lastsafexid =
3097 : : vstate->pendingpages[vstate->npendingpages - 1].safexid;
3098 : :
3099 : : Assert(FullTransactionIdFollowsOrEquals(safexid, lastsafexid));
3100 : : }
3101 : : #endif
3102 : :
3103 : : /*
3104 : : * If temp buffer reaches maxbufsize/work_mem capacity then we discard
3105 : : * information about this page.
3106 : : *
3107 : : * Note that this also covers the case where we opted to not use the
3108 : : * optimization in _bt_pendingfsm_init().
3109 : : */
3110 [ - + ]: 3708 : if (vstate->npendingpages == vstate->maxbufsize)
3111 : 0 : return;
3112 : :
3113 : : /* Consider enlarging buffer */
3114 [ + + ]: 3708 : if (vstate->npendingpages == vstate->bufsize)
3115 : : {
3116 : 5 : int newbufsize = vstate->bufsize * 2;
3117 : :
3118 : : /* Respect work_mem */
3119 [ - + ]: 5 : if (newbufsize > vstate->maxbufsize)
3120 : 0 : newbufsize = vstate->maxbufsize;
3121 : :
3122 : 5 : vstate->bufsize = newbufsize;
3123 : 5 : vstate->pendingpages =
3124 : 5 : repalloc(vstate->pendingpages,
3125 : 5 : sizeof(BTPendingFSM) * vstate->bufsize);
3126 : : }
3127 : :
3128 : : /* Save metadata for newly deleted page */
3129 : 3708 : vstate->pendingpages[vstate->npendingpages].target = target;
3130 : 3708 : vstate->pendingpages[vstate->npendingpages].safexid = safexid;
3131 : 3708 : vstate->npendingpages++;
3132 : : }
|
