Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * heapam.c
4 : * heap access method code
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/access/heap/heapam.c
12 : *
13 : *
14 : * INTERFACE ROUTINES
15 : * heap_beginscan - begin relation scan
16 : * heap_rescan - restart a relation scan
17 : * heap_endscan - end relation scan
18 : * heap_getnext - retrieve next tuple in scan
19 : * heap_fetch - retrieve tuple with given tid
20 : * heap_insert - insert tuple into a relation
21 : * heap_multi_insert - insert multiple tuples into a relation
22 : * heap_delete - delete a tuple from a relation
23 : * heap_update - replace a tuple in a relation with another tuple
24 : *
25 : * NOTES
26 : * This file contains the heap_ routines which implement
27 : * the POSTGRES heap access method used for all POSTGRES
28 : * relations.
29 : *
30 : *-------------------------------------------------------------------------
31 : */
32 : #include "postgres.h"
33 :
34 : #include "access/heapam.h"
35 : #include "access/heaptoast.h"
36 : #include "access/hio.h"
37 : #include "access/multixact.h"
38 : #include "access/subtrans.h"
39 : #include "access/syncscan.h"
40 : #include "access/valid.h"
41 : #include "access/visibilitymap.h"
42 : #include "access/xloginsert.h"
43 : #include "catalog/pg_database.h"
44 : #include "catalog/pg_database_d.h"
45 : #include "commands/vacuum.h"
46 : #include "pgstat.h"
47 : #include "port/pg_bitutils.h"
48 : #include "storage/lmgr.h"
49 : #include "storage/predicate.h"
50 : #include "storage/procarray.h"
51 : #include "utils/datum.h"
52 : #include "utils/injection_point.h"
53 : #include "utils/inval.h"
54 : #include "utils/spccache.h"
55 : #include "utils/syscache.h"
56 :
57 :
58 : static HeapTuple heap_prepare_insert(Relation relation, HeapTuple tup,
59 : TransactionId xid, CommandId cid, int options);
60 : static XLogRecPtr log_heap_update(Relation reln, Buffer oldbuf,
61 : Buffer newbuf, HeapTuple oldtup,
62 : HeapTuple newtup, HeapTuple old_key_tuple,
63 : bool all_visible_cleared, bool new_all_visible_cleared);
64 : #ifdef USE_ASSERT_CHECKING
65 : static void check_lock_if_inplace_updateable_rel(Relation relation,
66 : ItemPointer otid,
67 : HeapTuple newtup);
68 : static void check_inplace_rel_lock(HeapTuple oldtup);
69 : #endif
70 : static Bitmapset *HeapDetermineColumnsInfo(Relation relation,
71 : Bitmapset *interesting_cols,
72 : Bitmapset *external_cols,
73 : HeapTuple oldtup, HeapTuple newtup,
74 : bool *has_external);
75 : static bool heap_acquire_tuplock(Relation relation, ItemPointer tid,
76 : LockTupleMode mode, LockWaitPolicy wait_policy,
77 : bool *have_tuple_lock);
78 : static inline BlockNumber heapgettup_advance_block(HeapScanDesc scan,
79 : BlockNumber block,
80 : ScanDirection dir);
81 : static pg_noinline BlockNumber heapgettup_initial_block(HeapScanDesc scan,
82 : ScanDirection dir);
83 : static void compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
84 : uint16 old_infomask2, TransactionId add_to_xmax,
85 : LockTupleMode mode, bool is_update,
86 : TransactionId *result_xmax, uint16 *result_infomask,
87 : uint16 *result_infomask2);
88 : static TM_Result heap_lock_updated_tuple(Relation rel, HeapTuple tuple,
89 : ItemPointer ctid, TransactionId xid,
90 : LockTupleMode mode);
91 : static void GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
92 : uint16 *new_infomask2);
93 : static TransactionId MultiXactIdGetUpdateXid(TransactionId xmax,
94 : uint16 t_infomask);
95 : static bool DoesMultiXactIdConflict(MultiXactId multi, uint16 infomask,
96 : LockTupleMode lockmode, bool *current_is_member);
97 : static void MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
98 : Relation rel, ItemPointer ctid, XLTW_Oper oper,
99 : int *remaining);
100 : static bool ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
101 : uint16 infomask, Relation rel, int *remaining,
102 : bool logLockFailure);
103 : static void index_delete_sort(TM_IndexDeleteOp *delstate);
104 : static int bottomup_sort_and_shrink(TM_IndexDeleteOp *delstate);
105 : static XLogRecPtr log_heap_new_cid(Relation relation, HeapTuple tup);
106 : static HeapTuple ExtractReplicaIdentity(Relation relation, HeapTuple tp, bool key_required,
107 : bool *copy);
108 :
109 :
110 : /*
111 : * Each tuple lock mode has a corresponding heavyweight lock, and one or two
112 : * corresponding MultiXactStatuses (one to merely lock tuples, another one to
113 : * update them). This table (and the macros below) helps us determine the
114 : * heavyweight lock mode and MultiXactStatus values to use for any particular
115 : * tuple lock strength.
116 : *
117 : * These interact with InplaceUpdateTupleLock, an alias for ExclusiveLock.
118 : *
119 : * Don't look at lockstatus/updstatus directly! Use get_mxact_status_for_lock
120 : * instead.
121 : */
122 : static const struct
123 : {
124 : LOCKMODE hwlock;
125 : int lockstatus;
126 : int updstatus;
127 : }
128 :
129 : tupleLockExtraInfo[MaxLockTupleMode + 1] =
130 : {
131 : { /* LockTupleKeyShare */
132 : AccessShareLock,
133 : MultiXactStatusForKeyShare,
134 : -1 /* KeyShare does not allow updating tuples */
135 : },
136 : { /* LockTupleShare */
137 : RowShareLock,
138 : MultiXactStatusForShare,
139 : -1 /* Share does not allow updating tuples */
140 : },
141 : { /* LockTupleNoKeyExclusive */
142 : ExclusiveLock,
143 : MultiXactStatusForNoKeyUpdate,
144 : MultiXactStatusNoKeyUpdate
145 : },
146 : { /* LockTupleExclusive */
147 : AccessExclusiveLock,
148 : MultiXactStatusForUpdate,
149 : MultiXactStatusUpdate
150 : }
151 : };
152 :
153 : /* Get the LOCKMODE for a given MultiXactStatus */
154 : #define LOCKMODE_from_mxstatus(status) \
155 : (tupleLockExtraInfo[TUPLOCK_from_mxstatus((status))].hwlock)
156 :
157 : /*
158 : * Acquire heavyweight locks on tuples, using a LockTupleMode strength value.
159 : * This is more readable than having every caller translate it to lock.h's
160 : * LOCKMODE.
161 : */
162 : #define LockTupleTuplock(rel, tup, mode) \
163 : LockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
164 : #define UnlockTupleTuplock(rel, tup, mode) \
165 : UnlockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
166 : #define ConditionalLockTupleTuplock(rel, tup, mode, log) \
167 : ConditionalLockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock, (log))
168 :
169 : #ifdef USE_PREFETCH
170 : /*
171 : * heap_index_delete_tuples and index_delete_prefetch_buffer use this
172 : * structure to coordinate prefetching activity
173 : */
174 : typedef struct
175 : {
176 : BlockNumber cur_hblkno;
177 : int next_item;
178 : int ndeltids;
179 : TM_IndexDelete *deltids;
180 : } IndexDeletePrefetchState;
181 : #endif
182 :
183 : /* heap_index_delete_tuples bottom-up index deletion costing constants */
184 : #define BOTTOMUP_MAX_NBLOCKS 6
185 : #define BOTTOMUP_TOLERANCE_NBLOCKS 3
186 :
187 : /*
188 : * heap_index_delete_tuples uses this when determining which heap blocks it
189 : * must visit to help its bottom-up index deletion caller
190 : */
191 : typedef struct IndexDeleteCounts
192 : {
193 : int16 npromisingtids; /* Number of "promising" TIDs in group */
194 : int16 ntids; /* Number of TIDs in group */
195 : int16 ifirsttid; /* Offset to group's first deltid */
196 : } IndexDeleteCounts;
197 :
198 : /*
199 : * This table maps tuple lock strength values for each particular
200 : * MultiXactStatus value.
201 : */
202 : static const int MultiXactStatusLock[MaxMultiXactStatus + 1] =
203 : {
204 : LockTupleKeyShare, /* ForKeyShare */
205 : LockTupleShare, /* ForShare */
206 : LockTupleNoKeyExclusive, /* ForNoKeyUpdate */
207 : LockTupleExclusive, /* ForUpdate */
208 : LockTupleNoKeyExclusive, /* NoKeyUpdate */
209 : LockTupleExclusive /* Update */
210 : };
211 :
212 : /* Get the LockTupleMode for a given MultiXactStatus */
213 : #define TUPLOCK_from_mxstatus(status) \
214 : (MultiXactStatusLock[(status)])
215 :
216 : /*
217 : * Check that we have a valid snapshot if we might need TOAST access.
218 : */
219 : static inline void
220 20830296 : AssertHasSnapshotForToast(Relation rel)
221 : {
222 : #ifdef USE_ASSERT_CHECKING
223 :
224 : /* bootstrap mode in particular breaks this rule */
225 : if (!IsNormalProcessingMode())
226 : return;
227 :
228 : /* if the relation doesn't have a TOAST table, we are good */
229 : if (!OidIsValid(rel->rd_rel->reltoastrelid))
230 : return;
231 :
232 : Assert(HaveRegisteredOrActiveSnapshot());
233 :
234 : #endif /* USE_ASSERT_CHECKING */
235 20830296 : }
236 :
237 : /* ----------------------------------------------------------------
238 : * heap support routines
239 : * ----------------------------------------------------------------
240 : */
241 :
242 : /*
243 : * Streaming read API callback for parallel sequential scans. Returns the next
244 : * block the caller wants from the read stream or InvalidBlockNumber when done.
245 : */
246 : static BlockNumber
247 201572 : heap_scan_stream_read_next_parallel(ReadStream *stream,
248 : void *callback_private_data,
249 : void *per_buffer_data)
250 : {
251 201572 : HeapScanDesc scan = (HeapScanDesc) callback_private_data;
252 :
253 : Assert(ScanDirectionIsForward(scan->rs_dir));
254 : Assert(scan->rs_base.rs_parallel);
255 :
256 201572 : if (unlikely(!scan->rs_inited))
257 : {
258 : /* parallel scan */
259 2990 : table_block_parallelscan_startblock_init(scan->rs_base.rs_rd,
260 2990 : scan->rs_parallelworkerdata,
261 2990 : (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel);
262 :
263 : /* may return InvalidBlockNumber if there are no more blocks */
264 5980 : scan->rs_prefetch_block = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
265 2990 : scan->rs_parallelworkerdata,
266 2990 : (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel);
267 2990 : scan->rs_inited = true;
268 : }
269 : else
270 : {
271 198582 : scan->rs_prefetch_block = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
272 198582 : scan->rs_parallelworkerdata, (ParallelBlockTableScanDesc)
273 198582 : scan->rs_base.rs_parallel);
274 : }
275 :
276 201572 : return scan->rs_prefetch_block;
277 : }
278 :
279 : /*
280 : * Streaming read API callback for serial sequential and TID range scans.
281 : * Returns the next block the caller wants from the read stream or
282 : * InvalidBlockNumber when done.
283 : */
284 : static BlockNumber
285 7799868 : heap_scan_stream_read_next_serial(ReadStream *stream,
286 : void *callback_private_data,
287 : void *per_buffer_data)
288 : {
289 7799868 : HeapScanDesc scan = (HeapScanDesc) callback_private_data;
290 :
291 7799868 : if (unlikely(!scan->rs_inited))
292 : {
293 2252600 : scan->rs_prefetch_block = heapgettup_initial_block(scan, scan->rs_dir);
294 2252600 : scan->rs_inited = true;
295 : }
296 : else
297 5547268 : scan->rs_prefetch_block = heapgettup_advance_block(scan,
298 : scan->rs_prefetch_block,
299 : scan->rs_dir);
300 :
301 7799868 : return scan->rs_prefetch_block;
302 : }
303 :
304 : /*
305 : * Read stream API callback for bitmap heap scans.
306 : * Returns the next block the caller wants from the read stream or
307 : * InvalidBlockNumber when done.
308 : */
309 : static BlockNumber
310 419328 : bitmapheap_stream_read_next(ReadStream *pgsr, void *private_data,
311 : void *per_buffer_data)
312 : {
313 419328 : TBMIterateResult *tbmres = per_buffer_data;
314 419328 : BitmapHeapScanDesc bscan = (BitmapHeapScanDesc) private_data;
315 419328 : HeapScanDesc hscan = (HeapScanDesc) bscan;
316 419328 : TableScanDesc sscan = &hscan->rs_base;
317 :
318 : for (;;)
319 : {
320 419328 : CHECK_FOR_INTERRUPTS();
321 :
322 : /* no more entries in the bitmap */
323 419328 : if (!tbm_iterate(&sscan->st.rs_tbmiterator, tbmres))
324 20162 : return InvalidBlockNumber;
325 :
326 : /*
327 : * Ignore any claimed entries past what we think is the end of the
328 : * relation. It may have been extended after the start of our scan (we
329 : * only hold an AccessShareLock, and it could be inserts from this
330 : * backend). We don't take this optimization in SERIALIZABLE
331 : * isolation though, as we need to examine all invisible tuples
332 : * reachable by the index.
333 : */
334 399166 : if (!IsolationIsSerializable() &&
335 398960 : tbmres->blockno >= hscan->rs_nblocks)
336 0 : continue;
337 :
338 399166 : return tbmres->blockno;
339 : }
340 :
341 : /* not reachable */
342 : Assert(false);
343 : }
344 :
345 : /* ----------------
346 : * initscan - scan code common to heap_beginscan and heap_rescan
347 : * ----------------
348 : */
349 : static void
350 2294858 : initscan(HeapScanDesc scan, ScanKey key, bool keep_startblock)
351 : {
352 2294858 : ParallelBlockTableScanDesc bpscan = NULL;
353 : bool allow_strat;
354 : bool allow_sync;
355 :
356 : /*
357 : * Determine the number of blocks we have to scan.
358 : *
359 : * It is sufficient to do this once at scan start, since any tuples added
360 : * while the scan is in progress will be invisible to my snapshot anyway.
361 : * (That is not true when using a non-MVCC snapshot. However, we couldn't
362 : * guarantee to return tuples added after scan start anyway, since they
363 : * might go into pages we already scanned. To guarantee consistent
364 : * results for a non-MVCC snapshot, the caller must hold some higher-level
365 : * lock that ensures the interesting tuple(s) won't change.)
366 : */
367 2294858 : if (scan->rs_base.rs_parallel != NULL)
368 : {
369 4016 : bpscan = (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
370 4016 : scan->rs_nblocks = bpscan->phs_nblocks;
371 : }
372 : else
373 2290842 : scan->rs_nblocks = RelationGetNumberOfBlocks(scan->rs_base.rs_rd);
374 :
375 : /*
376 : * If the table is large relative to NBuffers, use a bulk-read access
377 : * strategy and enable synchronized scanning (see syncscan.c). Although
378 : * the thresholds for these features could be different, we make them the
379 : * same so that there are only two behaviors to tune rather than four.
380 : * (However, some callers need to be able to disable one or both of these
381 : * behaviors, independently of the size of the table; also there is a GUC
382 : * variable that can disable synchronized scanning.)
383 : *
384 : * Note that table_block_parallelscan_initialize has a very similar test;
385 : * if you change this, consider changing that one, too.
386 : */
387 2294854 : if (!RelationUsesLocalBuffers(scan->rs_base.rs_rd) &&
388 2280592 : scan->rs_nblocks > NBuffers / 4)
389 : {
390 25496 : allow_strat = (scan->rs_base.rs_flags & SO_ALLOW_STRAT) != 0;
391 25496 : allow_sync = (scan->rs_base.rs_flags & SO_ALLOW_SYNC) != 0;
392 : }
393 : else
394 2269358 : allow_strat = allow_sync = false;
395 :
396 2294854 : if (allow_strat)
397 : {
398 : /* During a rescan, keep the previous strategy object. */
399 22956 : if (scan->rs_strategy == NULL)
400 22650 : scan->rs_strategy = GetAccessStrategy(BAS_BULKREAD);
401 : }
402 : else
403 : {
404 2271898 : if (scan->rs_strategy != NULL)
405 0 : FreeAccessStrategy(scan->rs_strategy);
406 2271898 : scan->rs_strategy = NULL;
407 : }
408 :
409 2294854 : if (scan->rs_base.rs_parallel != NULL)
410 : {
411 : /* For parallel scan, believe whatever ParallelTableScanDesc says. */
412 4016 : if (scan->rs_base.rs_parallel->phs_syncscan)
413 4 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
414 : else
415 4012 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
416 : }
417 2290838 : else if (keep_startblock)
418 : {
419 : /*
420 : * When rescanning, we want to keep the previous startblock setting,
421 : * so that rewinding a cursor doesn't generate surprising results.
422 : * Reset the active syncscan setting, though.
423 : */
424 1531798 : if (allow_sync && synchronize_seqscans)
425 40 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
426 : else
427 1531758 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
428 : }
429 759040 : else if (allow_sync && synchronize_seqscans)
430 : {
431 112 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
432 112 : scan->rs_startblock = ss_get_location(scan->rs_base.rs_rd, scan->rs_nblocks);
433 : }
434 : else
435 : {
436 758928 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
437 758928 : scan->rs_startblock = 0;
438 : }
439 :
440 2294854 : scan->rs_numblocks = InvalidBlockNumber;
441 2294854 : scan->rs_inited = false;
442 2294854 : scan->rs_ctup.t_data = NULL;
443 2294854 : ItemPointerSetInvalid(&scan->rs_ctup.t_self);
444 2294854 : scan->rs_cbuf = InvalidBuffer;
445 2294854 : scan->rs_cblock = InvalidBlockNumber;
446 2294854 : scan->rs_ntuples = 0;
447 2294854 : scan->rs_cindex = 0;
448 :
449 : /*
450 : * Initialize to ForwardScanDirection because it is most common and
451 : * because heap scans go forward before going backward (e.g. CURSORs).
452 : */
453 2294854 : scan->rs_dir = ForwardScanDirection;
454 2294854 : scan->rs_prefetch_block = InvalidBlockNumber;
455 :
456 : /* page-at-a-time fields are always invalid when not rs_inited */
457 :
458 : /*
459 : * copy the scan key, if appropriate
460 : */
461 2294854 : if (key != NULL && scan->rs_base.rs_nkeys > 0)
462 434204 : memcpy(scan->rs_base.rs_key, key, scan->rs_base.rs_nkeys * sizeof(ScanKeyData));
463 :
464 : /*
465 : * Currently, we only have a stats counter for sequential heap scans (but
466 : * e.g for bitmap scans the underlying bitmap index scans will be counted,
467 : * and for sample scans we update stats for tuple fetches).
468 : */
469 2294854 : if (scan->rs_base.rs_flags & SO_TYPE_SEQSCAN)
470 2254942 : pgstat_count_heap_scan(scan->rs_base.rs_rd);
471 2294854 : }
472 :
473 : /*
474 : * heap_setscanlimits - restrict range of a heapscan
475 : *
476 : * startBlk is the page to start at
477 : * numBlks is number of pages to scan (InvalidBlockNumber means "all")
478 : */
479 : void
480 5626 : heap_setscanlimits(TableScanDesc sscan, BlockNumber startBlk, BlockNumber numBlks)
481 : {
482 5626 : HeapScanDesc scan = (HeapScanDesc) sscan;
483 :
484 : Assert(!scan->rs_inited); /* else too late to change */
485 : /* else rs_startblock is significant */
486 : Assert(!(scan->rs_base.rs_flags & SO_ALLOW_SYNC));
487 :
488 : /* Check startBlk is valid (but allow case of zero blocks...) */
489 : Assert(startBlk == 0 || startBlk < scan->rs_nblocks);
490 :
491 5626 : scan->rs_startblock = startBlk;
492 5626 : scan->rs_numblocks = numBlks;
493 5626 : }
494 :
495 : /*
496 : * Per-tuple loop for heap_prepare_pagescan(). Pulled out so it can be called
497 : * multiple times, with constant arguments for all_visible,
498 : * check_serializable.
499 : */
500 : pg_attribute_always_inline
501 : static int
502 5509528 : page_collect_tuples(HeapScanDesc scan, Snapshot snapshot,
503 : Page page, Buffer buffer,
504 : BlockNumber block, int lines,
505 : bool all_visible, bool check_serializable)
506 : {
507 5509528 : int ntup = 0;
508 : OffsetNumber lineoff;
509 :
510 275805620 : for (lineoff = FirstOffsetNumber; lineoff <= lines; lineoff++)
511 : {
512 270296108 : ItemId lpp = PageGetItemId(page, lineoff);
513 : HeapTupleData loctup;
514 : bool valid;
515 :
516 270296108 : if (!ItemIdIsNormal(lpp))
517 52610568 : continue;
518 :
519 217685540 : loctup.t_data = (HeapTupleHeader) PageGetItem(page, lpp);
520 217685540 : loctup.t_len = ItemIdGetLength(lpp);
521 217685540 : loctup.t_tableOid = RelationGetRelid(scan->rs_base.rs_rd);
522 217685540 : ItemPointerSet(&(loctup.t_self), block, lineoff);
523 :
524 217685540 : if (all_visible)
525 88962158 : valid = true;
526 : else
527 128723382 : valid = HeapTupleSatisfiesVisibility(&loctup, snapshot, buffer);
528 :
529 217685540 : if (check_serializable)
530 2830 : HeapCheckForSerializableConflictOut(valid, scan->rs_base.rs_rd,
531 : &loctup, buffer, snapshot);
532 :
533 217685524 : if (valid)
534 : {
535 202701550 : scan->rs_vistuples[ntup] = lineoff;
536 202701550 : ntup++;
537 : }
538 : }
539 :
540 : Assert(ntup <= MaxHeapTuplesPerPage);
541 :
542 5509512 : return ntup;
543 : }
544 :
545 : /*
546 : * heap_prepare_pagescan - Prepare current scan page to be scanned in pagemode
547 : *
548 : * Preparation currently consists of 1. prune the scan's rs_cbuf page, and 2.
549 : * fill the rs_vistuples[] array with the OffsetNumbers of visible tuples.
550 : */
551 : void
552 5509528 : heap_prepare_pagescan(TableScanDesc sscan)
553 : {
554 5509528 : HeapScanDesc scan = (HeapScanDesc) sscan;
555 5509528 : Buffer buffer = scan->rs_cbuf;
556 5509528 : BlockNumber block = scan->rs_cblock;
557 : Snapshot snapshot;
558 : Page page;
559 : int lines;
560 : bool all_visible;
561 : bool check_serializable;
562 :
563 : Assert(BufferGetBlockNumber(buffer) == block);
564 :
565 : /* ensure we're not accidentally being used when not in pagemode */
566 : Assert(scan->rs_base.rs_flags & SO_ALLOW_PAGEMODE);
567 5509528 : snapshot = scan->rs_base.rs_snapshot;
568 :
569 : /*
570 : * Prune and repair fragmentation for the whole page, if possible.
571 : */
572 5509528 : heap_page_prune_opt(scan->rs_base.rs_rd, buffer);
573 :
574 : /*
575 : * We must hold share lock on the buffer content while examining tuple
576 : * visibility. Afterwards, however, the tuples we have found to be
577 : * visible are guaranteed good as long as we hold the buffer pin.
578 : */
579 5509528 : LockBuffer(buffer, BUFFER_LOCK_SHARE);
580 :
581 5509528 : page = BufferGetPage(buffer);
582 5509528 : lines = PageGetMaxOffsetNumber(page);
583 :
584 : /*
585 : * If the all-visible flag indicates that all tuples on the page are
586 : * visible to everyone, we can skip the per-tuple visibility tests.
587 : *
588 : * Note: In hot standby, a tuple that's already visible to all
589 : * transactions on the primary might still be invisible to a read-only
590 : * transaction in the standby. We partly handle this problem by tracking
591 : * the minimum xmin of visible tuples as the cut-off XID while marking a
592 : * page all-visible on the primary and WAL log that along with the
593 : * visibility map SET operation. In hot standby, we wait for (or abort)
594 : * all transactions that can potentially may not see one or more tuples on
595 : * the page. That's how index-only scans work fine in hot standby. A
596 : * crucial difference between index-only scans and heap scans is that the
597 : * index-only scan completely relies on the visibility map where as heap
598 : * scan looks at the page-level PD_ALL_VISIBLE flag. We are not sure if
599 : * the page-level flag can be trusted in the same way, because it might
600 : * get propagated somehow without being explicitly WAL-logged, e.g. via a
601 : * full page write. Until we can prove that beyond doubt, let's check each
602 : * tuple for visibility the hard way.
603 : */
604 5509528 : all_visible = PageIsAllVisible(page) && !snapshot->takenDuringRecovery;
605 : check_serializable =
606 5509528 : CheckForSerializableConflictOutNeeded(scan->rs_base.rs_rd, snapshot);
607 :
608 : /*
609 : * We call page_collect_tuples() with constant arguments, to get the
610 : * compiler to constant fold the constant arguments. Separate calls with
611 : * constant arguments, rather than variables, are needed on several
612 : * compilers to actually perform constant folding.
613 : */
614 5509528 : if (likely(all_visible))
615 : {
616 2125680 : if (likely(!check_serializable))
617 2125680 : scan->rs_ntuples = page_collect_tuples(scan, snapshot, page, buffer,
618 : block, lines, true, false);
619 : else
620 0 : scan->rs_ntuples = page_collect_tuples(scan, snapshot, page, buffer,
621 : block, lines, true, true);
622 : }
623 : else
624 : {
625 3383848 : if (likely(!check_serializable))
626 3382608 : scan->rs_ntuples = page_collect_tuples(scan, snapshot, page, buffer,
627 : block, lines, false, false);
628 : else
629 1240 : scan->rs_ntuples = page_collect_tuples(scan, snapshot, page, buffer,
630 : block, lines, false, true);
631 : }
632 :
633 5509512 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
634 5509512 : }
635 :
636 : /*
637 : * heap_fetch_next_buffer - read and pin the next block from MAIN_FORKNUM.
638 : *
639 : * Read the next block of the scan relation from the read stream and save it
640 : * in the scan descriptor. It is already pinned.
641 : */
642 : static inline void
643 7635320 : heap_fetch_next_buffer(HeapScanDesc scan, ScanDirection dir)
644 : {
645 : Assert(scan->rs_read_stream);
646 :
647 : /* release previous scan buffer, if any */
648 7635320 : if (BufferIsValid(scan->rs_cbuf))
649 : {
650 5379728 : ReleaseBuffer(scan->rs_cbuf);
651 5379728 : scan->rs_cbuf = InvalidBuffer;
652 : }
653 :
654 : /*
655 : * Be sure to check for interrupts at least once per page. Checks at
656 : * higher code levels won't be able to stop a seqscan that encounters many
657 : * pages' worth of consecutive dead tuples.
658 : */
659 7635320 : CHECK_FOR_INTERRUPTS();
660 :
661 : /*
662 : * If the scan direction is changing, reset the prefetch block to the
663 : * current block. Otherwise, we will incorrectly prefetch the blocks
664 : * between the prefetch block and the current block again before
665 : * prefetching blocks in the new, correct scan direction.
666 : */
667 7635316 : if (unlikely(scan->rs_dir != dir))
668 : {
669 154 : scan->rs_prefetch_block = scan->rs_cblock;
670 154 : read_stream_reset(scan->rs_read_stream);
671 : }
672 :
673 7635316 : scan->rs_dir = dir;
674 :
675 7635316 : scan->rs_cbuf = read_stream_next_buffer(scan->rs_read_stream, NULL);
676 7635266 : if (BufferIsValid(scan->rs_cbuf))
677 5703956 : scan->rs_cblock = BufferGetBlockNumber(scan->rs_cbuf);
678 7635266 : }
679 :
680 : /*
681 : * heapgettup_initial_block - return the first BlockNumber to scan
682 : *
683 : * Returns InvalidBlockNumber when there are no blocks to scan. This can
684 : * occur with empty tables and in parallel scans when parallel workers get all
685 : * of the pages before we can get a chance to get our first page.
686 : */
687 : static pg_noinline BlockNumber
688 2252600 : heapgettup_initial_block(HeapScanDesc scan, ScanDirection dir)
689 : {
690 : Assert(!scan->rs_inited);
691 : Assert(scan->rs_base.rs_parallel == NULL);
692 :
693 : /* When there are no pages to scan, return InvalidBlockNumber */
694 2252600 : if (scan->rs_nblocks == 0 || scan->rs_numblocks == 0)
695 1245594 : return InvalidBlockNumber;
696 :
697 1007006 : if (ScanDirectionIsForward(dir))
698 : {
699 1006942 : return scan->rs_startblock;
700 : }
701 : else
702 : {
703 : /*
704 : * Disable reporting to syncscan logic in a backwards scan; it's not
705 : * very likely anyone else is doing the same thing at the same time,
706 : * and much more likely that we'll just bollix things for forward
707 : * scanners.
708 : */
709 64 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
710 :
711 : /*
712 : * Start from last page of the scan. Ensure we take into account
713 : * rs_numblocks if it's been adjusted by heap_setscanlimits().
714 : */
715 64 : if (scan->rs_numblocks != InvalidBlockNumber)
716 6 : return (scan->rs_startblock + scan->rs_numblocks - 1) % scan->rs_nblocks;
717 :
718 58 : if (scan->rs_startblock > 0)
719 0 : return scan->rs_startblock - 1;
720 :
721 58 : return scan->rs_nblocks - 1;
722 : }
723 : }
724 :
725 :
726 : /*
727 : * heapgettup_start_page - helper function for heapgettup()
728 : *
729 : * Return the next page to scan based on the scan->rs_cbuf and set *linesleft
730 : * to the number of tuples on this page. Also set *lineoff to the first
731 : * offset to scan with forward scans getting the first offset and backward
732 : * getting the final offset on the page.
733 : */
734 : static Page
735 202982 : heapgettup_start_page(HeapScanDesc scan, ScanDirection dir, int *linesleft,
736 : OffsetNumber *lineoff)
737 : {
738 : Page page;
739 :
740 : Assert(scan->rs_inited);
741 : Assert(BufferIsValid(scan->rs_cbuf));
742 :
743 : /* Caller is responsible for ensuring buffer is locked if needed */
744 202982 : page = BufferGetPage(scan->rs_cbuf);
745 :
746 202982 : *linesleft = PageGetMaxOffsetNumber(page) - FirstOffsetNumber + 1;
747 :
748 202982 : if (ScanDirectionIsForward(dir))
749 202982 : *lineoff = FirstOffsetNumber;
750 : else
751 0 : *lineoff = (OffsetNumber) (*linesleft);
752 :
753 : /* lineoff now references the physically previous or next tid */
754 202982 : return page;
755 : }
756 :
757 :
758 : /*
759 : * heapgettup_continue_page - helper function for heapgettup()
760 : *
761 : * Return the next page to scan based on the scan->rs_cbuf and set *linesleft
762 : * to the number of tuples left to scan on this page. Also set *lineoff to
763 : * the next offset to scan according to the ScanDirection in 'dir'.
764 : */
765 : static inline Page
766 16470998 : heapgettup_continue_page(HeapScanDesc scan, ScanDirection dir, int *linesleft,
767 : OffsetNumber *lineoff)
768 : {
769 : Page page;
770 :
771 : Assert(scan->rs_inited);
772 : Assert(BufferIsValid(scan->rs_cbuf));
773 :
774 : /* Caller is responsible for ensuring buffer is locked if needed */
775 16470998 : page = BufferGetPage(scan->rs_cbuf);
776 :
777 16470998 : if (ScanDirectionIsForward(dir))
778 : {
779 16470998 : *lineoff = OffsetNumberNext(scan->rs_coffset);
780 16470998 : *linesleft = PageGetMaxOffsetNumber(page) - (*lineoff) + 1;
781 : }
782 : else
783 : {
784 : /*
785 : * The previous returned tuple may have been vacuumed since the
786 : * previous scan when we use a non-MVCC snapshot, so we must
787 : * re-establish the lineoff <= PageGetMaxOffsetNumber(page) invariant
788 : */
789 0 : *lineoff = Min(PageGetMaxOffsetNumber(page), OffsetNumberPrev(scan->rs_coffset));
790 0 : *linesleft = *lineoff;
791 : }
792 :
793 : /* lineoff now references the physically previous or next tid */
794 16470998 : return page;
795 : }
796 :
797 : /*
798 : * heapgettup_advance_block - helper for heap_fetch_next_buffer()
799 : *
800 : * Given the current block number, the scan direction, and various information
801 : * contained in the scan descriptor, calculate the BlockNumber to scan next
802 : * and return it. If there are no further blocks to scan, return
803 : * InvalidBlockNumber to indicate this fact to the caller.
804 : *
805 : * This should not be called to determine the initial block number -- only for
806 : * subsequent blocks.
807 : *
808 : * This also adjusts rs_numblocks when a limit has been imposed by
809 : * heap_setscanlimits().
810 : */
811 : static inline BlockNumber
812 5547268 : heapgettup_advance_block(HeapScanDesc scan, BlockNumber block, ScanDirection dir)
813 : {
814 : Assert(scan->rs_base.rs_parallel == NULL);
815 :
816 5547268 : if (likely(ScanDirectionIsForward(dir)))
817 : {
818 5547150 : block++;
819 :
820 : /* wrap back to the start of the heap */
821 5547150 : if (block >= scan->rs_nblocks)
822 806220 : block = 0;
823 :
824 : /*
825 : * Report our new scan position for synchronization purposes. We don't
826 : * do that when moving backwards, however. That would just mess up any
827 : * other forward-moving scanners.
828 : *
829 : * Note: we do this before checking for end of scan so that the final
830 : * state of the position hint is back at the start of the rel. That's
831 : * not strictly necessary, but otherwise when you run the same query
832 : * multiple times the starting position would shift a little bit
833 : * backwards on every invocation, which is confusing. We don't
834 : * guarantee any specific ordering in general, though.
835 : */
836 5547150 : if (scan->rs_base.rs_flags & SO_ALLOW_SYNC)
837 17060 : ss_report_location(scan->rs_base.rs_rd, block);
838 :
839 : /* we're done if we're back at where we started */
840 5547150 : if (block == scan->rs_startblock)
841 806138 : return InvalidBlockNumber;
842 :
843 : /* check if the limit imposed by heap_setscanlimits() is met */
844 4741012 : if (scan->rs_numblocks != InvalidBlockNumber)
845 : {
846 4932 : if (--scan->rs_numblocks == 0)
847 3056 : return InvalidBlockNumber;
848 : }
849 :
850 4737956 : return block;
851 : }
852 : else
853 : {
854 : /* we're done if the last block is the start position */
855 118 : if (block == scan->rs_startblock)
856 118 : return InvalidBlockNumber;
857 :
858 : /* check if the limit imposed by heap_setscanlimits() is met */
859 0 : if (scan->rs_numblocks != InvalidBlockNumber)
860 : {
861 0 : if (--scan->rs_numblocks == 0)
862 0 : return InvalidBlockNumber;
863 : }
864 :
865 : /* wrap to the end of the heap when the last page was page 0 */
866 0 : if (block == 0)
867 0 : block = scan->rs_nblocks;
868 :
869 0 : block--;
870 :
871 0 : return block;
872 : }
873 : }
874 :
875 : /* ----------------
876 : * heapgettup - fetch next heap tuple
877 : *
878 : * Initialize the scan if not already done; then advance to the next
879 : * tuple as indicated by "dir"; return the next tuple in scan->rs_ctup,
880 : * or set scan->rs_ctup.t_data = NULL if no more tuples.
881 : *
882 : * Note: the reason nkeys/key are passed separately, even though they are
883 : * kept in the scan descriptor, is that the caller may not want us to check
884 : * the scankeys.
885 : *
886 : * Note: when we fall off the end of the scan in either direction, we
887 : * reset rs_inited. This means that a further request with the same
888 : * scan direction will restart the scan, which is a bit odd, but a
889 : * request with the opposite scan direction will start a fresh scan
890 : * in the proper direction. The latter is required behavior for cursors,
891 : * while the former case is generally undefined behavior in Postgres
892 : * so we don't care too much.
893 : * ----------------
894 : */
895 : static void
896 16512756 : heapgettup(HeapScanDesc scan,
897 : ScanDirection dir,
898 : int nkeys,
899 : ScanKey key)
900 : {
901 16512756 : HeapTuple tuple = &(scan->rs_ctup);
902 : Page page;
903 : OffsetNumber lineoff;
904 : int linesleft;
905 :
906 16512756 : if (likely(scan->rs_inited))
907 : {
908 : /* continue from previously returned page/tuple */
909 16470998 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
910 16470998 : page = heapgettup_continue_page(scan, dir, &linesleft, &lineoff);
911 16470998 : goto continue_page;
912 : }
913 :
914 : /*
915 : * advance the scan until we find a qualifying tuple or run out of stuff
916 : * to scan
917 : */
918 : while (true)
919 : {
920 244440 : heap_fetch_next_buffer(scan, dir);
921 :
922 : /* did we run out of blocks to scan? */
923 244440 : if (!BufferIsValid(scan->rs_cbuf))
924 41458 : break;
925 :
926 : Assert(BufferGetBlockNumber(scan->rs_cbuf) == scan->rs_cblock);
927 :
928 202982 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
929 202982 : page = heapgettup_start_page(scan, dir, &linesleft, &lineoff);
930 16673980 : continue_page:
931 :
932 : /*
933 : * Only continue scanning the page while we have lines left.
934 : *
935 : * Note that this protects us from accessing line pointers past
936 : * PageGetMaxOffsetNumber(); both for forward scans when we resume the
937 : * table scan, and for when we start scanning a new page.
938 : */
939 16753852 : for (; linesleft > 0; linesleft--, lineoff += dir)
940 : {
941 : bool visible;
942 16551170 : ItemId lpp = PageGetItemId(page, lineoff);
943 :
944 16551170 : if (!ItemIdIsNormal(lpp))
945 69416 : continue;
946 :
947 16481754 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lpp);
948 16481754 : tuple->t_len = ItemIdGetLength(lpp);
949 16481754 : ItemPointerSet(&(tuple->t_self), scan->rs_cblock, lineoff);
950 :
951 16481754 : visible = HeapTupleSatisfiesVisibility(tuple,
952 : scan->rs_base.rs_snapshot,
953 : scan->rs_cbuf);
954 :
955 16481754 : HeapCheckForSerializableConflictOut(visible, scan->rs_base.rs_rd,
956 : tuple, scan->rs_cbuf,
957 : scan->rs_base.rs_snapshot);
958 :
959 : /* skip tuples not visible to this snapshot */
960 16481754 : if (!visible)
961 10456 : continue;
962 :
963 : /* skip any tuples that don't match the scan key */
964 16471298 : if (key != NULL &&
965 0 : !HeapKeyTest(tuple, RelationGetDescr(scan->rs_base.rs_rd),
966 : nkeys, key))
967 0 : continue;
968 :
969 16471298 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
970 16471298 : scan->rs_coffset = lineoff;
971 16471298 : return;
972 : }
973 :
974 : /*
975 : * if we get here, it means we've exhausted the items on this page and
976 : * it's time to move to the next.
977 : */
978 202682 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
979 : }
980 :
981 : /* end of scan */
982 41458 : if (BufferIsValid(scan->rs_cbuf))
983 0 : ReleaseBuffer(scan->rs_cbuf);
984 :
985 41458 : scan->rs_cbuf = InvalidBuffer;
986 41458 : scan->rs_cblock = InvalidBlockNumber;
987 41458 : scan->rs_prefetch_block = InvalidBlockNumber;
988 41458 : tuple->t_data = NULL;
989 41458 : scan->rs_inited = false;
990 : }
991 :
992 : /* ----------------
993 : * heapgettup_pagemode - fetch next heap tuple in page-at-a-time mode
994 : *
995 : * Same API as heapgettup, but used in page-at-a-time mode
996 : *
997 : * The internal logic is much the same as heapgettup's too, but there are some
998 : * differences: we do not take the buffer content lock (that only needs to
999 : * happen inside heap_prepare_pagescan), and we iterate through just the
1000 : * tuples listed in rs_vistuples[] rather than all tuples on the page. Notice
1001 : * that lineindex is 0-based, where the corresponding loop variable lineoff in
1002 : * heapgettup is 1-based.
1003 : * ----------------
1004 : */
1005 : static void
1006 102006968 : heapgettup_pagemode(HeapScanDesc scan,
1007 : ScanDirection dir,
1008 : int nkeys,
1009 : ScanKey key)
1010 : {
1011 102006968 : HeapTuple tuple = &(scan->rs_ctup);
1012 : Page page;
1013 : uint32 lineindex;
1014 : uint32 linesleft;
1015 :
1016 102006968 : if (likely(scan->rs_inited))
1017 : {
1018 : /* continue from previously returned page/tuple */
1019 99793134 : page = BufferGetPage(scan->rs_cbuf);
1020 :
1021 99793134 : lineindex = scan->rs_cindex + dir;
1022 99793134 : if (ScanDirectionIsForward(dir))
1023 99792476 : linesleft = scan->rs_ntuples - lineindex;
1024 : else
1025 658 : linesleft = scan->rs_cindex;
1026 : /* lineindex now references the next or previous visible tid */
1027 :
1028 99793134 : goto continue_page;
1029 : }
1030 :
1031 : /*
1032 : * advance the scan until we find a qualifying tuple or run out of stuff
1033 : * to scan
1034 : */
1035 : while (true)
1036 : {
1037 7390880 : heap_fetch_next_buffer(scan, dir);
1038 :
1039 : /* did we run out of blocks to scan? */
1040 7390826 : if (!BufferIsValid(scan->rs_cbuf))
1041 1889852 : break;
1042 :
1043 : Assert(BufferGetBlockNumber(scan->rs_cbuf) == scan->rs_cblock);
1044 :
1045 : /* prune the page and determine visible tuple offsets */
1046 5500974 : heap_prepare_pagescan((TableScanDesc) scan);
1047 5500958 : page = BufferGetPage(scan->rs_cbuf);
1048 5500958 : linesleft = scan->rs_ntuples;
1049 5500958 : lineindex = ScanDirectionIsForward(dir) ? 0 : linesleft - 1;
1050 :
1051 : /* block is the same for all tuples, set it once outside the loop */
1052 5500958 : ItemPointerSetBlockNumber(&tuple->t_self, scan->rs_cblock);
1053 :
1054 : /* lineindex now references the next or previous visible tid */
1055 105294092 : continue_page:
1056 :
1057 201577546 : for (; linesleft > 0; linesleft--, lineindex += dir)
1058 : {
1059 : ItemId lpp;
1060 : OffsetNumber lineoff;
1061 :
1062 : Assert(lineindex <= scan->rs_ntuples);
1063 196400500 : lineoff = scan->rs_vistuples[lineindex];
1064 196400500 : lpp = PageGetItemId(page, lineoff);
1065 : Assert(ItemIdIsNormal(lpp));
1066 :
1067 196400500 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lpp);
1068 196400500 : tuple->t_len = ItemIdGetLength(lpp);
1069 196400500 : ItemPointerSetOffsetNumber(&tuple->t_self, lineoff);
1070 :
1071 : /* skip any tuples that don't match the scan key */
1072 196400500 : if (key != NULL &&
1073 97064862 : !HeapKeyTest(tuple, RelationGetDescr(scan->rs_base.rs_rd),
1074 : nkeys, key))
1075 96283454 : continue;
1076 :
1077 100117046 : scan->rs_cindex = lineindex;
1078 100117046 : return;
1079 : }
1080 : }
1081 :
1082 : /* end of scan */
1083 1889852 : if (BufferIsValid(scan->rs_cbuf))
1084 0 : ReleaseBuffer(scan->rs_cbuf);
1085 1889852 : scan->rs_cbuf = InvalidBuffer;
1086 1889852 : scan->rs_cblock = InvalidBlockNumber;
1087 1889852 : scan->rs_prefetch_block = InvalidBlockNumber;
1088 1889852 : tuple->t_data = NULL;
1089 1889852 : scan->rs_inited = false;
1090 : }
1091 :
1092 :
1093 : /* ----------------------------------------------------------------
1094 : * heap access method interface
1095 : * ----------------------------------------------------------------
1096 : */
1097 :
1098 :
1099 : TableScanDesc
1100 762952 : heap_beginscan(Relation relation, Snapshot snapshot,
1101 : int nkeys, ScanKey key,
1102 : ParallelTableScanDesc parallel_scan,
1103 : uint32 flags)
1104 : {
1105 : HeapScanDesc scan;
1106 :
1107 : /*
1108 : * increment relation ref count while scanning relation
1109 : *
1110 : * This is just to make really sure the relcache entry won't go away while
1111 : * the scan has a pointer to it. Caller should be holding the rel open
1112 : * anyway, so this is redundant in all normal scenarios...
1113 : */
1114 762952 : RelationIncrementReferenceCount(relation);
1115 :
1116 : /*
1117 : * allocate and initialize scan descriptor
1118 : */
1119 762952 : if (flags & SO_TYPE_BITMAPSCAN)
1120 : {
1121 16036 : BitmapHeapScanDesc bscan = palloc(sizeof(BitmapHeapScanDescData));
1122 :
1123 : /*
1124 : * Bitmap Heap scans do not have any fields that a normal Heap Scan
1125 : * does not have, so no special initializations required here.
1126 : */
1127 16036 : scan = (HeapScanDesc) bscan;
1128 : }
1129 : else
1130 746916 : scan = (HeapScanDesc) palloc(sizeof(HeapScanDescData));
1131 :
1132 762952 : scan->rs_base.rs_rd = relation;
1133 762952 : scan->rs_base.rs_snapshot = snapshot;
1134 762952 : scan->rs_base.rs_nkeys = nkeys;
1135 762952 : scan->rs_base.rs_flags = flags;
1136 762952 : scan->rs_base.rs_parallel = parallel_scan;
1137 762952 : scan->rs_strategy = NULL; /* set in initscan */
1138 762952 : scan->rs_cbuf = InvalidBuffer;
1139 :
1140 : /*
1141 : * Disable page-at-a-time mode if it's not a MVCC-safe snapshot.
1142 : */
1143 762952 : if (!(snapshot && IsMVCCSnapshot(snapshot)))
1144 58614 : scan->rs_base.rs_flags &= ~SO_ALLOW_PAGEMODE;
1145 :
1146 : /*
1147 : * For seqscan and sample scans in a serializable transaction, acquire a
1148 : * predicate lock on the entire relation. This is required not only to
1149 : * lock all the matching tuples, but also to conflict with new insertions
1150 : * into the table. In an indexscan, we take page locks on the index pages
1151 : * covering the range specified in the scan qual, but in a heap scan there
1152 : * is nothing more fine-grained to lock. A bitmap scan is a different
1153 : * story, there we have already scanned the index and locked the index
1154 : * pages covering the predicate. But in that case we still have to lock
1155 : * any matching heap tuples. For sample scan we could optimize the locking
1156 : * to be at least page-level granularity, but we'd need to add per-tuple
1157 : * locking for that.
1158 : */
1159 762952 : if (scan->rs_base.rs_flags & (SO_TYPE_SEQSCAN | SO_TYPE_SAMPLESCAN))
1160 : {
1161 : /*
1162 : * Ensure a missing snapshot is noticed reliably, even if the
1163 : * isolation mode means predicate locking isn't performed (and
1164 : * therefore the snapshot isn't used here).
1165 : */
1166 : Assert(snapshot);
1167 727452 : PredicateLockRelation(relation, snapshot);
1168 : }
1169 :
1170 : /* we only need to set this up once */
1171 762952 : scan->rs_ctup.t_tableOid = RelationGetRelid(relation);
1172 :
1173 : /*
1174 : * Allocate memory to keep track of page allocation for parallel workers
1175 : * when doing a parallel scan.
1176 : */
1177 762952 : if (parallel_scan != NULL)
1178 3908 : scan->rs_parallelworkerdata = palloc(sizeof(ParallelBlockTableScanWorkerData));
1179 : else
1180 759044 : scan->rs_parallelworkerdata = NULL;
1181 :
1182 : /*
1183 : * we do this here instead of in initscan() because heap_rescan also calls
1184 : * initscan() and we don't want to allocate memory again
1185 : */
1186 762952 : if (nkeys > 0)
1187 434204 : scan->rs_base.rs_key = (ScanKey) palloc(sizeof(ScanKeyData) * nkeys);
1188 : else
1189 328748 : scan->rs_base.rs_key = NULL;
1190 :
1191 762952 : initscan(scan, key, false);
1192 :
1193 762948 : scan->rs_read_stream = NULL;
1194 :
1195 : /*
1196 : * Set up a read stream for sequential scans and TID range scans. This
1197 : * should be done after initscan() because initscan() allocates the
1198 : * BufferAccessStrategy object passed to the read stream API.
1199 : */
1200 762948 : if (scan->rs_base.rs_flags & SO_TYPE_SEQSCAN ||
1201 35646 : scan->rs_base.rs_flags & SO_TYPE_TIDRANGESCAN)
1202 729152 : {
1203 : ReadStreamBlockNumberCB cb;
1204 :
1205 729152 : if (scan->rs_base.rs_parallel)
1206 3908 : cb = heap_scan_stream_read_next_parallel;
1207 : else
1208 725244 : cb = heap_scan_stream_read_next_serial;
1209 :
1210 : /* ---
1211 : * It is safe to use batchmode as the only locks taken by `cb`
1212 : * are never taken while waiting for IO:
1213 : * - SyncScanLock is used in the non-parallel case
1214 : * - in the parallel case, only spinlocks and atomics are used
1215 : * ---
1216 : */
1217 729152 : scan->rs_read_stream = read_stream_begin_relation(READ_STREAM_SEQUENTIAL |
1218 : READ_STREAM_USE_BATCHING,
1219 : scan->rs_strategy,
1220 : scan->rs_base.rs_rd,
1221 : MAIN_FORKNUM,
1222 : cb,
1223 : scan,
1224 : 0);
1225 : }
1226 33796 : else if (scan->rs_base.rs_flags & SO_TYPE_BITMAPSCAN)
1227 : {
1228 16036 : scan->rs_read_stream = read_stream_begin_relation(READ_STREAM_DEFAULT |
1229 : READ_STREAM_USE_BATCHING,
1230 : scan->rs_strategy,
1231 : scan->rs_base.rs_rd,
1232 : MAIN_FORKNUM,
1233 : bitmapheap_stream_read_next,
1234 : scan,
1235 : sizeof(TBMIterateResult));
1236 : }
1237 :
1238 :
1239 762948 : return (TableScanDesc) scan;
1240 : }
1241 :
1242 : void
1243 1531906 : heap_rescan(TableScanDesc sscan, ScanKey key, bool set_params,
1244 : bool allow_strat, bool allow_sync, bool allow_pagemode)
1245 : {
1246 1531906 : HeapScanDesc scan = (HeapScanDesc) sscan;
1247 :
1248 1531906 : if (set_params)
1249 : {
1250 30 : if (allow_strat)
1251 30 : scan->rs_base.rs_flags |= SO_ALLOW_STRAT;
1252 : else
1253 0 : scan->rs_base.rs_flags &= ~SO_ALLOW_STRAT;
1254 :
1255 30 : if (allow_sync)
1256 12 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
1257 : else
1258 18 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
1259 :
1260 30 : if (allow_pagemode && scan->rs_base.rs_snapshot &&
1261 30 : IsMVCCSnapshot(scan->rs_base.rs_snapshot))
1262 30 : scan->rs_base.rs_flags |= SO_ALLOW_PAGEMODE;
1263 : else
1264 0 : scan->rs_base.rs_flags &= ~SO_ALLOW_PAGEMODE;
1265 : }
1266 :
1267 : /*
1268 : * unpin scan buffers
1269 : */
1270 1531906 : if (BufferIsValid(scan->rs_cbuf))
1271 : {
1272 2860 : ReleaseBuffer(scan->rs_cbuf);
1273 2860 : scan->rs_cbuf = InvalidBuffer;
1274 : }
1275 :
1276 : /*
1277 : * SO_TYPE_BITMAPSCAN would be cleaned up here, but it does not hold any
1278 : * additional data vs a normal HeapScan
1279 : */
1280 :
1281 : /*
1282 : * The read stream is reset on rescan. This must be done before
1283 : * initscan(), as some state referred to by read_stream_reset() is reset
1284 : * in initscan().
1285 : */
1286 1531906 : if (scan->rs_read_stream)
1287 1531870 : read_stream_reset(scan->rs_read_stream);
1288 :
1289 : /*
1290 : * reinitialize scan descriptor
1291 : */
1292 1531906 : initscan(scan, key, true);
1293 1531906 : }
1294 :
1295 : void
1296 758310 : heap_endscan(TableScanDesc sscan)
1297 : {
1298 758310 : HeapScanDesc scan = (HeapScanDesc) sscan;
1299 :
1300 : /* Note: no locking manipulations needed */
1301 :
1302 : /*
1303 : * unpin scan buffers
1304 : */
1305 758310 : if (BufferIsValid(scan->rs_cbuf))
1306 318158 : ReleaseBuffer(scan->rs_cbuf);
1307 :
1308 : /*
1309 : * Must free the read stream before freeing the BufferAccessStrategy.
1310 : */
1311 758310 : if (scan->rs_read_stream)
1312 740656 : read_stream_end(scan->rs_read_stream);
1313 :
1314 : /*
1315 : * decrement relation reference count and free scan descriptor storage
1316 : */
1317 758310 : RelationDecrementReferenceCount(scan->rs_base.rs_rd);
1318 :
1319 758310 : if (scan->rs_base.rs_key)
1320 434146 : pfree(scan->rs_base.rs_key);
1321 :
1322 758310 : if (scan->rs_strategy != NULL)
1323 22634 : FreeAccessStrategy(scan->rs_strategy);
1324 :
1325 758310 : if (scan->rs_parallelworkerdata != NULL)
1326 3908 : pfree(scan->rs_parallelworkerdata);
1327 :
1328 758310 : if (scan->rs_base.rs_flags & SO_TEMP_SNAPSHOT)
1329 74938 : UnregisterSnapshot(scan->rs_base.rs_snapshot);
1330 :
1331 758310 : pfree(scan);
1332 758310 : }
1333 :
1334 : HeapTuple
1335 20348786 : heap_getnext(TableScanDesc sscan, ScanDirection direction)
1336 : {
1337 20348786 : HeapScanDesc scan = (HeapScanDesc) sscan;
1338 :
1339 : /*
1340 : * This is still widely used directly, without going through table AM, so
1341 : * add a safety check. It's possible we should, at a later point,
1342 : * downgrade this to an assert. The reason for checking the AM routine,
1343 : * rather than the AM oid, is that this allows to write regression tests
1344 : * that create another AM reusing the heap handler.
1345 : */
1346 20348786 : if (unlikely(sscan->rs_rd->rd_tableam != GetHeapamTableAmRoutine()))
1347 0 : ereport(ERROR,
1348 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1349 : errmsg_internal("only heap AM is supported")));
1350 :
1351 : /*
1352 : * We don't expect direct calls to heap_getnext with valid CheckXidAlive
1353 : * for catalog or regular tables. See detailed comments in xact.c where
1354 : * these variables are declared. Normally we have such a check at tableam
1355 : * level API but this is called from many places so we need to ensure it
1356 : * here.
1357 : */
1358 20348786 : if (unlikely(TransactionIdIsValid(CheckXidAlive) && !bsysscan))
1359 0 : elog(ERROR, "unexpected heap_getnext call during logical decoding");
1360 :
1361 : /* Note: no locking manipulations needed */
1362 :
1363 20348786 : if (scan->rs_base.rs_flags & SO_ALLOW_PAGEMODE)
1364 4795376 : heapgettup_pagemode(scan, direction,
1365 4795376 : scan->rs_base.rs_nkeys, scan->rs_base.rs_key);
1366 : else
1367 15553410 : heapgettup(scan, direction,
1368 15553410 : scan->rs_base.rs_nkeys, scan->rs_base.rs_key);
1369 :
1370 20348786 : if (scan->rs_ctup.t_data == NULL)
1371 128128 : return NULL;
1372 :
1373 : /*
1374 : * if we get here it means we have a new current scan tuple, so point to
1375 : * the proper return buffer and return the tuple.
1376 : */
1377 :
1378 20220658 : pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1379 :
1380 20220658 : return &scan->rs_ctup;
1381 : }
1382 :
1383 : bool
1384 98162158 : heap_getnextslot(TableScanDesc sscan, ScanDirection direction, TupleTableSlot *slot)
1385 : {
1386 98162158 : HeapScanDesc scan = (HeapScanDesc) sscan;
1387 :
1388 : /* Note: no locking manipulations needed */
1389 :
1390 98162158 : if (sscan->rs_flags & SO_ALLOW_PAGEMODE)
1391 97202812 : heapgettup_pagemode(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1392 : else
1393 959346 : heapgettup(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1394 :
1395 98162104 : if (scan->rs_ctup.t_data == NULL)
1396 : {
1397 1803088 : ExecClearTuple(slot);
1398 1803088 : return false;
1399 : }
1400 :
1401 : /*
1402 : * if we get here it means we have a new current scan tuple, so point to
1403 : * the proper return buffer and return the tuple.
1404 : */
1405 :
1406 96359016 : pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1407 :
1408 96359016 : ExecStoreBufferHeapTuple(&scan->rs_ctup, slot,
1409 : scan->rs_cbuf);
1410 96359016 : return true;
1411 : }
1412 :
1413 : void
1414 1916 : heap_set_tidrange(TableScanDesc sscan, ItemPointer mintid,
1415 : ItemPointer maxtid)
1416 : {
1417 1916 : HeapScanDesc scan = (HeapScanDesc) sscan;
1418 : BlockNumber startBlk;
1419 : BlockNumber numBlks;
1420 : ItemPointerData highestItem;
1421 : ItemPointerData lowestItem;
1422 :
1423 : /*
1424 : * For relations without any pages, we can simply leave the TID range
1425 : * unset. There will be no tuples to scan, therefore no tuples outside
1426 : * the given TID range.
1427 : */
1428 1916 : if (scan->rs_nblocks == 0)
1429 48 : return;
1430 :
1431 : /*
1432 : * Set up some ItemPointers which point to the first and last possible
1433 : * tuples in the heap.
1434 : */
1435 1904 : ItemPointerSet(&highestItem, scan->rs_nblocks - 1, MaxOffsetNumber);
1436 1904 : ItemPointerSet(&lowestItem, 0, FirstOffsetNumber);
1437 :
1438 : /*
1439 : * If the given maximum TID is below the highest possible TID in the
1440 : * relation, then restrict the range to that, otherwise we scan to the end
1441 : * of the relation.
1442 : */
1443 1904 : if (ItemPointerCompare(maxtid, &highestItem) < 0)
1444 136 : ItemPointerCopy(maxtid, &highestItem);
1445 :
1446 : /*
1447 : * If the given minimum TID is above the lowest possible TID in the
1448 : * relation, then restrict the range to only scan for TIDs above that.
1449 : */
1450 1904 : if (ItemPointerCompare(mintid, &lowestItem) > 0)
1451 1762 : ItemPointerCopy(mintid, &lowestItem);
1452 :
1453 : /*
1454 : * Check for an empty range and protect from would be negative results
1455 : * from the numBlks calculation below.
1456 : */
1457 1904 : if (ItemPointerCompare(&highestItem, &lowestItem) < 0)
1458 : {
1459 : /* Set an empty range of blocks to scan */
1460 36 : heap_setscanlimits(sscan, 0, 0);
1461 36 : return;
1462 : }
1463 :
1464 : /*
1465 : * Calculate the first block and the number of blocks we must scan. We
1466 : * could be more aggressive here and perform some more validation to try
1467 : * and further narrow the scope of blocks to scan by checking if the
1468 : * lowestItem has an offset above MaxOffsetNumber. In this case, we could
1469 : * advance startBlk by one. Likewise, if highestItem has an offset of 0
1470 : * we could scan one fewer blocks. However, such an optimization does not
1471 : * seem worth troubling over, currently.
1472 : */
1473 1868 : startBlk = ItemPointerGetBlockNumberNoCheck(&lowestItem);
1474 :
1475 1868 : numBlks = ItemPointerGetBlockNumberNoCheck(&highestItem) -
1476 1868 : ItemPointerGetBlockNumberNoCheck(&lowestItem) + 1;
1477 :
1478 : /* Set the start block and number of blocks to scan */
1479 1868 : heap_setscanlimits(sscan, startBlk, numBlks);
1480 :
1481 : /* Finally, set the TID range in sscan */
1482 1868 : ItemPointerCopy(&lowestItem, &sscan->st.tidrange.rs_mintid);
1483 1868 : ItemPointerCopy(&highestItem, &sscan->st.tidrange.rs_maxtid);
1484 : }
1485 :
1486 : bool
1487 8594 : heap_getnextslot_tidrange(TableScanDesc sscan, ScanDirection direction,
1488 : TupleTableSlot *slot)
1489 : {
1490 8594 : HeapScanDesc scan = (HeapScanDesc) sscan;
1491 8594 : ItemPointer mintid = &sscan->st.tidrange.rs_mintid;
1492 8594 : ItemPointer maxtid = &sscan->st.tidrange.rs_maxtid;
1493 :
1494 : /* Note: no locking manipulations needed */
1495 : for (;;)
1496 : {
1497 8780 : if (sscan->rs_flags & SO_ALLOW_PAGEMODE)
1498 8780 : heapgettup_pagemode(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1499 : else
1500 0 : heapgettup(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1501 :
1502 8764 : if (scan->rs_ctup.t_data == NULL)
1503 : {
1504 94 : ExecClearTuple(slot);
1505 94 : return false;
1506 : }
1507 :
1508 : /*
1509 : * heap_set_tidrange will have used heap_setscanlimits to limit the
1510 : * range of pages we scan to only ones that can contain the TID range
1511 : * we're scanning for. Here we must filter out any tuples from these
1512 : * pages that are outside of that range.
1513 : */
1514 8670 : if (ItemPointerCompare(&scan->rs_ctup.t_self, mintid) < 0)
1515 : {
1516 186 : ExecClearTuple(slot);
1517 :
1518 : /*
1519 : * When scanning backwards, the TIDs will be in descending order.
1520 : * Future tuples in this direction will be lower still, so we can
1521 : * just return false to indicate there will be no more tuples.
1522 : */
1523 186 : if (ScanDirectionIsBackward(direction))
1524 0 : return false;
1525 :
1526 186 : continue;
1527 : }
1528 :
1529 : /*
1530 : * Likewise for the final page, we must filter out TIDs greater than
1531 : * maxtid.
1532 : */
1533 8484 : if (ItemPointerCompare(&scan->rs_ctup.t_self, maxtid) > 0)
1534 : {
1535 72 : ExecClearTuple(slot);
1536 :
1537 : /*
1538 : * When scanning forward, the TIDs will be in ascending order.
1539 : * Future tuples in this direction will be higher still, so we can
1540 : * just return false to indicate there will be no more tuples.
1541 : */
1542 72 : if (ScanDirectionIsForward(direction))
1543 72 : return false;
1544 0 : continue;
1545 : }
1546 :
1547 8412 : break;
1548 : }
1549 :
1550 : /*
1551 : * if we get here it means we have a new current scan tuple, so point to
1552 : * the proper return buffer and return the tuple.
1553 : */
1554 8412 : pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1555 :
1556 8412 : ExecStoreBufferHeapTuple(&scan->rs_ctup, slot, scan->rs_cbuf);
1557 8412 : return true;
1558 : }
1559 :
1560 : /*
1561 : * heap_fetch - retrieve tuple with given tid
1562 : *
1563 : * On entry, tuple->t_self is the TID to fetch. We pin the buffer holding
1564 : * the tuple, fill in the remaining fields of *tuple, and check the tuple
1565 : * against the specified snapshot.
1566 : *
1567 : * If successful (tuple found and passes snapshot time qual), then *userbuf
1568 : * is set to the buffer holding the tuple and true is returned. The caller
1569 : * must unpin the buffer when done with the tuple.
1570 : *
1571 : * If the tuple is not found (ie, item number references a deleted slot),
1572 : * then tuple->t_data is set to NULL, *userbuf is set to InvalidBuffer,
1573 : * and false is returned.
1574 : *
1575 : * If the tuple is found but fails the time qual check, then the behavior
1576 : * depends on the keep_buf parameter. If keep_buf is false, the results
1577 : * are the same as for the tuple-not-found case. If keep_buf is true,
1578 : * then tuple->t_data and *userbuf are returned as for the success case,
1579 : * and again the caller must unpin the buffer; but false is returned.
1580 : *
1581 : * heap_fetch does not follow HOT chains: only the exact TID requested will
1582 : * be fetched.
1583 : *
1584 : * It is somewhat inconsistent that we ereport() on invalid block number but
1585 : * return false on invalid item number. There are a couple of reasons though.
1586 : * One is that the caller can relatively easily check the block number for
1587 : * validity, but cannot check the item number without reading the page
1588 : * himself. Another is that when we are following a t_ctid link, we can be
1589 : * reasonably confident that the page number is valid (since VACUUM shouldn't
1590 : * truncate off the destination page without having killed the referencing
1591 : * tuple first), but the item number might well not be good.
1592 : */
1593 : bool
1594 355008 : heap_fetch(Relation relation,
1595 : Snapshot snapshot,
1596 : HeapTuple tuple,
1597 : Buffer *userbuf,
1598 : bool keep_buf)
1599 : {
1600 355008 : ItemPointer tid = &(tuple->t_self);
1601 : ItemId lp;
1602 : Buffer buffer;
1603 : Page page;
1604 : OffsetNumber offnum;
1605 : bool valid;
1606 :
1607 : /*
1608 : * Fetch and pin the appropriate page of the relation.
1609 : */
1610 355008 : buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
1611 :
1612 : /*
1613 : * Need share lock on buffer to examine tuple commit status.
1614 : */
1615 354992 : LockBuffer(buffer, BUFFER_LOCK_SHARE);
1616 354992 : page = BufferGetPage(buffer);
1617 :
1618 : /*
1619 : * We'd better check for out-of-range offnum in case of VACUUM since the
1620 : * TID was obtained.
1621 : */
1622 354992 : offnum = ItemPointerGetOffsetNumber(tid);
1623 354992 : if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1624 : {
1625 6 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1626 6 : ReleaseBuffer(buffer);
1627 6 : *userbuf = InvalidBuffer;
1628 6 : tuple->t_data = NULL;
1629 6 : return false;
1630 : }
1631 :
1632 : /*
1633 : * get the item line pointer corresponding to the requested tid
1634 : */
1635 354986 : lp = PageGetItemId(page, offnum);
1636 :
1637 : /*
1638 : * Must check for deleted tuple.
1639 : */
1640 354986 : if (!ItemIdIsNormal(lp))
1641 : {
1642 580 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1643 580 : ReleaseBuffer(buffer);
1644 580 : *userbuf = InvalidBuffer;
1645 580 : tuple->t_data = NULL;
1646 580 : return false;
1647 : }
1648 :
1649 : /*
1650 : * fill in *tuple fields
1651 : */
1652 354406 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
1653 354406 : tuple->t_len = ItemIdGetLength(lp);
1654 354406 : tuple->t_tableOid = RelationGetRelid(relation);
1655 :
1656 : /*
1657 : * check tuple visibility, then release lock
1658 : */
1659 354406 : valid = HeapTupleSatisfiesVisibility(tuple, snapshot, buffer);
1660 :
1661 354406 : if (valid)
1662 354312 : PredicateLockTID(relation, &(tuple->t_self), snapshot,
1663 354312 : HeapTupleHeaderGetXmin(tuple->t_data));
1664 :
1665 354406 : HeapCheckForSerializableConflictOut(valid, relation, tuple, buffer, snapshot);
1666 :
1667 354406 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1668 :
1669 354406 : if (valid)
1670 : {
1671 : /*
1672 : * All checks passed, so return the tuple as valid. Caller is now
1673 : * responsible for releasing the buffer.
1674 : */
1675 354312 : *userbuf = buffer;
1676 :
1677 354312 : return true;
1678 : }
1679 :
1680 : /* Tuple failed time qual, but maybe caller wants to see it anyway. */
1681 94 : if (keep_buf)
1682 58 : *userbuf = buffer;
1683 : else
1684 : {
1685 36 : ReleaseBuffer(buffer);
1686 36 : *userbuf = InvalidBuffer;
1687 36 : tuple->t_data = NULL;
1688 : }
1689 :
1690 94 : return false;
1691 : }
1692 :
1693 : /*
1694 : * heap_hot_search_buffer - search HOT chain for tuple satisfying snapshot
1695 : *
1696 : * On entry, *tid is the TID of a tuple (either a simple tuple, or the root
1697 : * of a HOT chain), and buffer is the buffer holding this tuple. We search
1698 : * for the first chain member satisfying the given snapshot. If one is
1699 : * found, we update *tid to reference that tuple's offset number, and
1700 : * return true. If no match, return false without modifying *tid.
1701 : *
1702 : * heapTuple is a caller-supplied buffer. When a match is found, we return
1703 : * the tuple here, in addition to updating *tid. If no match is found, the
1704 : * contents of this buffer on return are undefined.
1705 : *
1706 : * If all_dead is not NULL, we check non-visible tuples to see if they are
1707 : * globally dead; *all_dead is set true if all members of the HOT chain
1708 : * are vacuumable, false if not.
1709 : *
1710 : * Unlike heap_fetch, the caller must already have pin and (at least) share
1711 : * lock on the buffer; it is still pinned/locked at exit.
1712 : */
1713 : bool
1714 44882210 : heap_hot_search_buffer(ItemPointer tid, Relation relation, Buffer buffer,
1715 : Snapshot snapshot, HeapTuple heapTuple,
1716 : bool *all_dead, bool first_call)
1717 : {
1718 44882210 : Page page = BufferGetPage(buffer);
1719 44882210 : TransactionId prev_xmax = InvalidTransactionId;
1720 : BlockNumber blkno;
1721 : OffsetNumber offnum;
1722 : bool at_chain_start;
1723 : bool valid;
1724 : bool skip;
1725 44882210 : GlobalVisState *vistest = NULL;
1726 :
1727 : /* If this is not the first call, previous call returned a (live!) tuple */
1728 44882210 : if (all_dead)
1729 38366646 : *all_dead = first_call;
1730 :
1731 44882210 : blkno = ItemPointerGetBlockNumber(tid);
1732 44882210 : offnum = ItemPointerGetOffsetNumber(tid);
1733 44882210 : at_chain_start = first_call;
1734 44882210 : skip = !first_call;
1735 :
1736 : /* XXX: we should assert that a snapshot is pushed or registered */
1737 : Assert(TransactionIdIsValid(RecentXmin));
1738 : Assert(BufferGetBlockNumber(buffer) == blkno);
1739 :
1740 : /* Scan through possible multiple members of HOT-chain */
1741 : for (;;)
1742 3107238 : {
1743 : ItemId lp;
1744 :
1745 : /* check for bogus TID */
1746 47989448 : if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1747 : break;
1748 :
1749 47989448 : lp = PageGetItemId(page, offnum);
1750 :
1751 : /* check for unused, dead, or redirected items */
1752 47989448 : if (!ItemIdIsNormal(lp))
1753 : {
1754 : /* We should only see a redirect at start of chain */
1755 1684328 : if (ItemIdIsRedirected(lp) && at_chain_start)
1756 : {
1757 : /* Follow the redirect */
1758 1005812 : offnum = ItemIdGetRedirect(lp);
1759 1005812 : at_chain_start = false;
1760 1005812 : continue;
1761 : }
1762 : /* else must be end of chain */
1763 678516 : break;
1764 : }
1765 :
1766 : /*
1767 : * Update heapTuple to point to the element of the HOT chain we're
1768 : * currently investigating. Having t_self set correctly is important
1769 : * because the SSI checks and the *Satisfies routine for historical
1770 : * MVCC snapshots need the correct tid to decide about the visibility.
1771 : */
1772 46305120 : heapTuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
1773 46305120 : heapTuple->t_len = ItemIdGetLength(lp);
1774 46305120 : heapTuple->t_tableOid = RelationGetRelid(relation);
1775 46305120 : ItemPointerSet(&heapTuple->t_self, blkno, offnum);
1776 :
1777 : /*
1778 : * Shouldn't see a HEAP_ONLY tuple at chain start.
1779 : */
1780 46305120 : if (at_chain_start && HeapTupleIsHeapOnly(heapTuple))
1781 0 : break;
1782 :
1783 : /*
1784 : * The xmin should match the previous xmax value, else chain is
1785 : * broken.
1786 : */
1787 48406546 : if (TransactionIdIsValid(prev_xmax) &&
1788 2101426 : !TransactionIdEquals(prev_xmax,
1789 : HeapTupleHeaderGetXmin(heapTuple->t_data)))
1790 0 : break;
1791 :
1792 : /*
1793 : * When first_call is true (and thus, skip is initially false) we'll
1794 : * return the first tuple we find. But on later passes, heapTuple
1795 : * will initially be pointing to the tuple we returned last time.
1796 : * Returning it again would be incorrect (and would loop forever), so
1797 : * we skip it and return the next match we find.
1798 : */
1799 46305120 : if (!skip)
1800 : {
1801 : /* If it's visible per the snapshot, we must return it */
1802 46142392 : valid = HeapTupleSatisfiesVisibility(heapTuple, snapshot, buffer);
1803 46142392 : HeapCheckForSerializableConflictOut(valid, relation, heapTuple,
1804 : buffer, snapshot);
1805 :
1806 46142382 : if (valid)
1807 : {
1808 31782406 : ItemPointerSetOffsetNumber(tid, offnum);
1809 31782406 : PredicateLockTID(relation, &heapTuple->t_self, snapshot,
1810 31782406 : HeapTupleHeaderGetXmin(heapTuple->t_data));
1811 31782406 : if (all_dead)
1812 25806494 : *all_dead = false;
1813 31782406 : return true;
1814 : }
1815 : }
1816 14522704 : skip = false;
1817 :
1818 : /*
1819 : * If we can't see it, maybe no one else can either. At caller
1820 : * request, check whether all chain members are dead to all
1821 : * transactions.
1822 : *
1823 : * Note: if you change the criterion here for what is "dead", fix the
1824 : * planner's get_actual_variable_range() function to match.
1825 : */
1826 14522704 : if (all_dead && *all_dead)
1827 : {
1828 12886384 : if (!vistest)
1829 12641148 : vistest = GlobalVisTestFor(relation);
1830 :
1831 12886384 : if (!HeapTupleIsSurelyDead(heapTuple, vistest))
1832 12192500 : *all_dead = false;
1833 : }
1834 :
1835 : /*
1836 : * Check to see if HOT chain continues past this tuple; if so fetch
1837 : * the next offnum and loop around.
1838 : */
1839 14522704 : if (HeapTupleIsHotUpdated(heapTuple))
1840 : {
1841 : Assert(ItemPointerGetBlockNumber(&heapTuple->t_data->t_ctid) ==
1842 : blkno);
1843 2101426 : offnum = ItemPointerGetOffsetNumber(&heapTuple->t_data->t_ctid);
1844 2101426 : at_chain_start = false;
1845 2101426 : prev_xmax = HeapTupleHeaderGetUpdateXid(heapTuple->t_data);
1846 : }
1847 : else
1848 12421278 : break; /* end of chain */
1849 : }
1850 :
1851 13099794 : return false;
1852 : }
1853 :
1854 : /*
1855 : * heap_get_latest_tid - get the latest tid of a specified tuple
1856 : *
1857 : * Actually, this gets the latest version that is visible according to the
1858 : * scan's snapshot. Create a scan using SnapshotDirty to get the very latest,
1859 : * possibly uncommitted version.
1860 : *
1861 : * *tid is both an input and an output parameter: it is updated to
1862 : * show the latest version of the row. Note that it will not be changed
1863 : * if no version of the row passes the snapshot test.
1864 : */
1865 : void
1866 300 : heap_get_latest_tid(TableScanDesc sscan,
1867 : ItemPointer tid)
1868 : {
1869 300 : Relation relation = sscan->rs_rd;
1870 300 : Snapshot snapshot = sscan->rs_snapshot;
1871 : ItemPointerData ctid;
1872 : TransactionId priorXmax;
1873 :
1874 : /*
1875 : * table_tuple_get_latest_tid() verified that the passed in tid is valid.
1876 : * Assume that t_ctid links are valid however - there shouldn't be invalid
1877 : * ones in the table.
1878 : */
1879 : Assert(ItemPointerIsValid(tid));
1880 :
1881 : /*
1882 : * Loop to chase down t_ctid links. At top of loop, ctid is the tuple we
1883 : * need to examine, and *tid is the TID we will return if ctid turns out
1884 : * to be bogus.
1885 : *
1886 : * Note that we will loop until we reach the end of the t_ctid chain.
1887 : * Depending on the snapshot passed, there might be at most one visible
1888 : * version of the row, but we don't try to optimize for that.
1889 : */
1890 300 : ctid = *tid;
1891 300 : priorXmax = InvalidTransactionId; /* cannot check first XMIN */
1892 : for (;;)
1893 90 : {
1894 : Buffer buffer;
1895 : Page page;
1896 : OffsetNumber offnum;
1897 : ItemId lp;
1898 : HeapTupleData tp;
1899 : bool valid;
1900 :
1901 : /*
1902 : * Read, pin, and lock the page.
1903 : */
1904 390 : buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&ctid));
1905 390 : LockBuffer(buffer, BUFFER_LOCK_SHARE);
1906 390 : page = BufferGetPage(buffer);
1907 :
1908 : /*
1909 : * Check for bogus item number. This is not treated as an error
1910 : * condition because it can happen while following a t_ctid link. We
1911 : * just assume that the prior tid is OK and return it unchanged.
1912 : */
1913 390 : offnum = ItemPointerGetOffsetNumber(&ctid);
1914 390 : if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1915 : {
1916 0 : UnlockReleaseBuffer(buffer);
1917 0 : break;
1918 : }
1919 390 : lp = PageGetItemId(page, offnum);
1920 390 : if (!ItemIdIsNormal(lp))
1921 : {
1922 0 : UnlockReleaseBuffer(buffer);
1923 0 : break;
1924 : }
1925 :
1926 : /* OK to access the tuple */
1927 390 : tp.t_self = ctid;
1928 390 : tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
1929 390 : tp.t_len = ItemIdGetLength(lp);
1930 390 : tp.t_tableOid = RelationGetRelid(relation);
1931 :
1932 : /*
1933 : * After following a t_ctid link, we might arrive at an unrelated
1934 : * tuple. Check for XMIN match.
1935 : */
1936 480 : if (TransactionIdIsValid(priorXmax) &&
1937 90 : !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(tp.t_data)))
1938 : {
1939 0 : UnlockReleaseBuffer(buffer);
1940 0 : break;
1941 : }
1942 :
1943 : /*
1944 : * Check tuple visibility; if visible, set it as the new result
1945 : * candidate.
1946 : */
1947 390 : valid = HeapTupleSatisfiesVisibility(&tp, snapshot, buffer);
1948 390 : HeapCheckForSerializableConflictOut(valid, relation, &tp, buffer, snapshot);
1949 390 : if (valid)
1950 276 : *tid = ctid;
1951 :
1952 : /*
1953 : * If there's a valid t_ctid link, follow it, else we're done.
1954 : */
1955 552 : if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
1956 276 : HeapTupleHeaderIsOnlyLocked(tp.t_data) ||
1957 228 : HeapTupleHeaderIndicatesMovedPartitions(tp.t_data) ||
1958 114 : ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid))
1959 : {
1960 300 : UnlockReleaseBuffer(buffer);
1961 300 : break;
1962 : }
1963 :
1964 90 : ctid = tp.t_data->t_ctid;
1965 90 : priorXmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
1966 90 : UnlockReleaseBuffer(buffer);
1967 : } /* end of loop */
1968 300 : }
1969 :
1970 :
1971 : /*
1972 : * UpdateXmaxHintBits - update tuple hint bits after xmax transaction ends
1973 : *
1974 : * This is called after we have waited for the XMAX transaction to terminate.
1975 : * If the transaction aborted, we guarantee the XMAX_INVALID hint bit will
1976 : * be set on exit. If the transaction committed, we set the XMAX_COMMITTED
1977 : * hint bit if possible --- but beware that that may not yet be possible,
1978 : * if the transaction committed asynchronously.
1979 : *
1980 : * Note that if the transaction was a locker only, we set HEAP_XMAX_INVALID
1981 : * even if it commits.
1982 : *
1983 : * Hence callers should look only at XMAX_INVALID.
1984 : *
1985 : * Note this is not allowed for tuples whose xmax is a multixact.
1986 : */
1987 : static void
1988 354 : UpdateXmaxHintBits(HeapTupleHeader tuple, Buffer buffer, TransactionId xid)
1989 : {
1990 : Assert(TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple), xid));
1991 : Assert(!(tuple->t_infomask & HEAP_XMAX_IS_MULTI));
1992 :
1993 354 : if (!(tuple->t_infomask & (HEAP_XMAX_COMMITTED | HEAP_XMAX_INVALID)))
1994 : {
1995 642 : if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask) &&
1996 288 : TransactionIdDidCommit(xid))
1997 236 : HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_COMMITTED,
1998 : xid);
1999 : else
2000 118 : HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_INVALID,
2001 : InvalidTransactionId);
2002 : }
2003 354 : }
2004 :
2005 :
2006 : /*
2007 : * GetBulkInsertState - prepare status object for a bulk insert
2008 : */
2009 : BulkInsertState
2010 5792 : GetBulkInsertState(void)
2011 : {
2012 : BulkInsertState bistate;
2013 :
2014 5792 : bistate = (BulkInsertState) palloc(sizeof(BulkInsertStateData));
2015 5792 : bistate->strategy = GetAccessStrategy(BAS_BULKWRITE);
2016 5792 : bistate->current_buf = InvalidBuffer;
2017 5792 : bistate->next_free = InvalidBlockNumber;
2018 5792 : bistate->last_free = InvalidBlockNumber;
2019 5792 : bistate->already_extended_by = 0;
2020 5792 : return bistate;
2021 : }
2022 :
2023 : /*
2024 : * FreeBulkInsertState - clean up after finishing a bulk insert
2025 : */
2026 : void
2027 5508 : FreeBulkInsertState(BulkInsertState bistate)
2028 : {
2029 5508 : if (bistate->current_buf != InvalidBuffer)
2030 4360 : ReleaseBuffer(bistate->current_buf);
2031 5508 : FreeAccessStrategy(bistate->strategy);
2032 5508 : pfree(bistate);
2033 5508 : }
2034 :
2035 : /*
2036 : * ReleaseBulkInsertStatePin - release a buffer currently held in bistate
2037 : */
2038 : void
2039 161512 : ReleaseBulkInsertStatePin(BulkInsertState bistate)
2040 : {
2041 161512 : if (bistate->current_buf != InvalidBuffer)
2042 60042 : ReleaseBuffer(bistate->current_buf);
2043 161512 : bistate->current_buf = InvalidBuffer;
2044 :
2045 : /*
2046 : * Despite the name, we also reset bulk relation extension state.
2047 : * Otherwise we can end up erroring out due to looking for free space in
2048 : * ->next_free of one partition, even though ->next_free was set when
2049 : * extending another partition. It could obviously also be bad for
2050 : * efficiency to look at existing blocks at offsets from another
2051 : * partition, even if we don't error out.
2052 : */
2053 161512 : bistate->next_free = InvalidBlockNumber;
2054 161512 : bistate->last_free = InvalidBlockNumber;
2055 161512 : }
2056 :
2057 :
2058 : /*
2059 : * heap_insert - insert tuple into a heap
2060 : *
2061 : * The new tuple is stamped with current transaction ID and the specified
2062 : * command ID.
2063 : *
2064 : * See table_tuple_insert for comments about most of the input flags, except
2065 : * that this routine directly takes a tuple rather than a slot.
2066 : *
2067 : * There's corresponding HEAP_INSERT_ options to all the TABLE_INSERT_
2068 : * options, and there additionally is HEAP_INSERT_SPECULATIVE which is used to
2069 : * implement table_tuple_insert_speculative().
2070 : *
2071 : * On return the header fields of *tup are updated to match the stored tuple;
2072 : * in particular tup->t_self receives the actual TID where the tuple was
2073 : * stored. But note that any toasting of fields within the tuple data is NOT
2074 : * reflected into *tup.
2075 : */
2076 : void
2077 16554214 : heap_insert(Relation relation, HeapTuple tup, CommandId cid,
2078 : int options, BulkInsertState bistate)
2079 : {
2080 16554214 : TransactionId xid = GetCurrentTransactionId();
2081 : HeapTuple heaptup;
2082 : Buffer buffer;
2083 16554202 : Buffer vmbuffer = InvalidBuffer;
2084 16554202 : bool all_visible_cleared = false;
2085 :
2086 : /* Cheap, simplistic check that the tuple matches the rel's rowtype. */
2087 : Assert(HeapTupleHeaderGetNatts(tup->t_data) <=
2088 : RelationGetNumberOfAttributes(relation));
2089 :
2090 16554202 : AssertHasSnapshotForToast(relation);
2091 :
2092 : /*
2093 : * Fill in tuple header fields and toast the tuple if necessary.
2094 : *
2095 : * Note: below this point, heaptup is the data we actually intend to store
2096 : * into the relation; tup is the caller's original untoasted data.
2097 : */
2098 16554202 : heaptup = heap_prepare_insert(relation, tup, xid, cid, options);
2099 :
2100 : /*
2101 : * Find buffer to insert this tuple into. If the page is all visible,
2102 : * this will also pin the requisite visibility map page.
2103 : */
2104 16554202 : buffer = RelationGetBufferForTuple(relation, heaptup->t_len,
2105 : InvalidBuffer, options, bistate,
2106 : &vmbuffer, NULL,
2107 : 0);
2108 :
2109 : /*
2110 : * We're about to do the actual insert -- but check for conflict first, to
2111 : * avoid possibly having to roll back work we've just done.
2112 : *
2113 : * This is safe without a recheck as long as there is no possibility of
2114 : * another process scanning the page between this check and the insert
2115 : * being visible to the scan (i.e., an exclusive buffer content lock is
2116 : * continuously held from this point until the tuple insert is visible).
2117 : *
2118 : * For a heap insert, we only need to check for table-level SSI locks. Our
2119 : * new tuple can't possibly conflict with existing tuple locks, and heap
2120 : * page locks are only consolidated versions of tuple locks; they do not
2121 : * lock "gaps" as index page locks do. So we don't need to specify a
2122 : * buffer when making the call, which makes for a faster check.
2123 : */
2124 16554202 : CheckForSerializableConflictIn(relation, NULL, InvalidBlockNumber);
2125 :
2126 : /* NO EREPORT(ERROR) from here till changes are logged */
2127 16554178 : START_CRIT_SECTION();
2128 :
2129 16554178 : RelationPutHeapTuple(relation, buffer, heaptup,
2130 16554178 : (options & HEAP_INSERT_SPECULATIVE) != 0);
2131 :
2132 16554178 : if (PageIsAllVisible(BufferGetPage(buffer)))
2133 : {
2134 14380 : all_visible_cleared = true;
2135 14380 : PageClearAllVisible(BufferGetPage(buffer));
2136 14380 : visibilitymap_clear(relation,
2137 14380 : ItemPointerGetBlockNumber(&(heaptup->t_self)),
2138 : vmbuffer, VISIBILITYMAP_VALID_BITS);
2139 : }
2140 :
2141 : /*
2142 : * XXX Should we set PageSetPrunable on this page ?
2143 : *
2144 : * The inserting transaction may eventually abort thus making this tuple
2145 : * DEAD and hence available for pruning. Though we don't want to optimize
2146 : * for aborts, if no other tuple in this page is UPDATEd/DELETEd, the
2147 : * aborted tuple will never be pruned until next vacuum is triggered.
2148 : *
2149 : * If you do add PageSetPrunable here, add it in heap_xlog_insert too.
2150 : */
2151 :
2152 16554178 : MarkBufferDirty(buffer);
2153 :
2154 : /* XLOG stuff */
2155 16554178 : if (RelationNeedsWAL(relation))
2156 : {
2157 : xl_heap_insert xlrec;
2158 : xl_heap_header xlhdr;
2159 : XLogRecPtr recptr;
2160 14058328 : Page page = BufferGetPage(buffer);
2161 14058328 : uint8 info = XLOG_HEAP_INSERT;
2162 14058328 : int bufflags = 0;
2163 :
2164 : /*
2165 : * If this is a catalog, we need to transmit combo CIDs to properly
2166 : * decode, so log that as well.
2167 : */
2168 14058328 : if (RelationIsAccessibleInLogicalDecoding(relation))
2169 6560 : log_heap_new_cid(relation, heaptup);
2170 :
2171 : /*
2172 : * If this is the single and first tuple on page, we can reinit the
2173 : * page instead of restoring the whole thing. Set flag, and hide
2174 : * buffer references from XLogInsert.
2175 : */
2176 14239408 : if (ItemPointerGetOffsetNumber(&(heaptup->t_self)) == FirstOffsetNumber &&
2177 181080 : PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
2178 : {
2179 178890 : info |= XLOG_HEAP_INIT_PAGE;
2180 178890 : bufflags |= REGBUF_WILL_INIT;
2181 : }
2182 :
2183 14058328 : xlrec.offnum = ItemPointerGetOffsetNumber(&heaptup->t_self);
2184 14058328 : xlrec.flags = 0;
2185 14058328 : if (all_visible_cleared)
2186 14374 : xlrec.flags |= XLH_INSERT_ALL_VISIBLE_CLEARED;
2187 14058328 : if (options & HEAP_INSERT_SPECULATIVE)
2188 4110 : xlrec.flags |= XLH_INSERT_IS_SPECULATIVE;
2189 : Assert(ItemPointerGetBlockNumber(&heaptup->t_self) == BufferGetBlockNumber(buffer));
2190 :
2191 : /*
2192 : * For logical decoding, we need the tuple even if we're doing a full
2193 : * page write, so make sure it's included even if we take a full-page
2194 : * image. (XXX We could alternatively store a pointer into the FPW).
2195 : */
2196 14058328 : if (RelationIsLogicallyLogged(relation) &&
2197 499752 : !(options & HEAP_INSERT_NO_LOGICAL))
2198 : {
2199 499698 : xlrec.flags |= XLH_INSERT_CONTAINS_NEW_TUPLE;
2200 499698 : bufflags |= REGBUF_KEEP_DATA;
2201 :
2202 499698 : if (IsToastRelation(relation))
2203 3572 : xlrec.flags |= XLH_INSERT_ON_TOAST_RELATION;
2204 : }
2205 :
2206 14058328 : XLogBeginInsert();
2207 14058328 : XLogRegisterData(&xlrec, SizeOfHeapInsert);
2208 :
2209 14058328 : xlhdr.t_infomask2 = heaptup->t_data->t_infomask2;
2210 14058328 : xlhdr.t_infomask = heaptup->t_data->t_infomask;
2211 14058328 : xlhdr.t_hoff = heaptup->t_data->t_hoff;
2212 :
2213 : /*
2214 : * note we mark xlhdr as belonging to buffer; if XLogInsert decides to
2215 : * write the whole page to the xlog, we don't need to store
2216 : * xl_heap_header in the xlog.
2217 : */
2218 14058328 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2219 14058328 : XLogRegisterBufData(0, &xlhdr, SizeOfHeapHeader);
2220 : /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
2221 14058328 : XLogRegisterBufData(0,
2222 14058328 : (char *) heaptup->t_data + SizeofHeapTupleHeader,
2223 14058328 : heaptup->t_len - SizeofHeapTupleHeader);
2224 :
2225 : /* filtering by origin on a row level is much more efficient */
2226 14058328 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
2227 :
2228 14058328 : recptr = XLogInsert(RM_HEAP_ID, info);
2229 :
2230 14058328 : PageSetLSN(page, recptr);
2231 : }
2232 :
2233 16554178 : END_CRIT_SECTION();
2234 :
2235 16554178 : UnlockReleaseBuffer(buffer);
2236 16554178 : if (vmbuffer != InvalidBuffer)
2237 14992 : ReleaseBuffer(vmbuffer);
2238 :
2239 : /*
2240 : * If tuple is cachable, mark it for invalidation from the caches in case
2241 : * we abort. Note it is OK to do this after releasing the buffer, because
2242 : * the heaptup data structure is all in local memory, not in the shared
2243 : * buffer.
2244 : */
2245 16554178 : CacheInvalidateHeapTuple(relation, heaptup, NULL);
2246 :
2247 : /* Note: speculative insertions are counted too, even if aborted later */
2248 16554178 : pgstat_count_heap_insert(relation, 1);
2249 :
2250 : /*
2251 : * If heaptup is a private copy, release it. Don't forget to copy t_self
2252 : * back to the caller's image, too.
2253 : */
2254 16554178 : if (heaptup != tup)
2255 : {
2256 35986 : tup->t_self = heaptup->t_self;
2257 35986 : heap_freetuple(heaptup);
2258 : }
2259 16554178 : }
2260 :
2261 : /*
2262 : * Subroutine for heap_insert(). Prepares a tuple for insertion. This sets the
2263 : * tuple header fields and toasts the tuple if necessary. Returns a toasted
2264 : * version of the tuple if it was toasted, or the original tuple if not. Note
2265 : * that in any case, the header fields are also set in the original tuple.
2266 : */
2267 : static HeapTuple
2268 20332864 : heap_prepare_insert(Relation relation, HeapTuple tup, TransactionId xid,
2269 : CommandId cid, int options)
2270 : {
2271 : /*
2272 : * To allow parallel inserts, we need to ensure that they are safe to be
2273 : * performed in workers. We have the infrastructure to allow parallel
2274 : * inserts in general except for the cases where inserts generate a new
2275 : * CommandId (eg. inserts into a table having a foreign key column).
2276 : */
2277 20332864 : if (IsParallelWorker())
2278 0 : ereport(ERROR,
2279 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
2280 : errmsg("cannot insert tuples in a parallel worker")));
2281 :
2282 20332864 : tup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
2283 20332864 : tup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
2284 20332864 : tup->t_data->t_infomask |= HEAP_XMAX_INVALID;
2285 20332864 : HeapTupleHeaderSetXmin(tup->t_data, xid);
2286 20332864 : if (options & HEAP_INSERT_FROZEN)
2287 204264 : HeapTupleHeaderSetXminFrozen(tup->t_data);
2288 :
2289 20332864 : HeapTupleHeaderSetCmin(tup->t_data, cid);
2290 20332864 : HeapTupleHeaderSetXmax(tup->t_data, 0); /* for cleanliness */
2291 20332864 : tup->t_tableOid = RelationGetRelid(relation);
2292 :
2293 : /*
2294 : * If the new tuple is too big for storage or contains already toasted
2295 : * out-of-line attributes from some other relation, invoke the toaster.
2296 : */
2297 20332864 : if (relation->rd_rel->relkind != RELKIND_RELATION &&
2298 60608 : relation->rd_rel->relkind != RELKIND_MATVIEW)
2299 : {
2300 : /* toast table entries should never be recursively toasted */
2301 : Assert(!HeapTupleHasExternal(tup));
2302 60512 : return tup;
2303 : }
2304 20272352 : else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
2305 36094 : return heap_toast_insert_or_update(relation, tup, NULL, options);
2306 : else
2307 20236258 : return tup;
2308 : }
2309 :
2310 : /*
2311 : * Helper for heap_multi_insert() that computes the number of entire pages
2312 : * that inserting the remaining heaptuples requires. Used to determine how
2313 : * much the relation needs to be extended by.
2314 : */
2315 : static int
2316 728436 : heap_multi_insert_pages(HeapTuple *heaptuples, int done, int ntuples, Size saveFreeSpace)
2317 : {
2318 728436 : size_t page_avail = BLCKSZ - SizeOfPageHeaderData - saveFreeSpace;
2319 728436 : int npages = 1;
2320 :
2321 6322026 : for (int i = done; i < ntuples; i++)
2322 : {
2323 5593590 : size_t tup_sz = sizeof(ItemIdData) + MAXALIGN(heaptuples[i]->t_len);
2324 :
2325 5593590 : if (page_avail < tup_sz)
2326 : {
2327 43518 : npages++;
2328 43518 : page_avail = BLCKSZ - SizeOfPageHeaderData - saveFreeSpace;
2329 : }
2330 5593590 : page_avail -= tup_sz;
2331 : }
2332 :
2333 728436 : return npages;
2334 : }
2335 :
2336 : /*
2337 : * heap_multi_insert - insert multiple tuples into a heap
2338 : *
2339 : * This is like heap_insert(), but inserts multiple tuples in one operation.
2340 : * That's faster than calling heap_insert() in a loop, because when multiple
2341 : * tuples can be inserted on a single page, we can write just a single WAL
2342 : * record covering all of them, and only need to lock/unlock the page once.
2343 : *
2344 : * Note: this leaks memory into the current memory context. You can create a
2345 : * temporary context before calling this, if that's a problem.
2346 : */
2347 : void
2348 713988 : heap_multi_insert(Relation relation, TupleTableSlot **slots, int ntuples,
2349 : CommandId cid, int options, BulkInsertState bistate)
2350 : {
2351 713988 : TransactionId xid = GetCurrentTransactionId();
2352 : HeapTuple *heaptuples;
2353 : int i;
2354 : int ndone;
2355 : PGAlignedBlock scratch;
2356 : Page page;
2357 713988 : Buffer vmbuffer = InvalidBuffer;
2358 : bool needwal;
2359 : Size saveFreeSpace;
2360 713988 : bool need_tuple_data = RelationIsLogicallyLogged(relation);
2361 713988 : bool need_cids = RelationIsAccessibleInLogicalDecoding(relation);
2362 713988 : bool starting_with_empty_page = false;
2363 713988 : int npages = 0;
2364 713988 : int npages_used = 0;
2365 :
2366 : /* currently not needed (thus unsupported) for heap_multi_insert() */
2367 : Assert(!(options & HEAP_INSERT_NO_LOGICAL));
2368 :
2369 713988 : AssertHasSnapshotForToast(relation);
2370 :
2371 713988 : needwal = RelationNeedsWAL(relation);
2372 713988 : saveFreeSpace = RelationGetTargetPageFreeSpace(relation,
2373 : HEAP_DEFAULT_FILLFACTOR);
2374 :
2375 : /* Toast and set header data in all the slots */
2376 713988 : heaptuples = palloc(ntuples * sizeof(HeapTuple));
2377 4492650 : for (i = 0; i < ntuples; i++)
2378 : {
2379 : HeapTuple tuple;
2380 :
2381 3778662 : tuple = ExecFetchSlotHeapTuple(slots[i], true, NULL);
2382 3778662 : slots[i]->tts_tableOid = RelationGetRelid(relation);
2383 3778662 : tuple->t_tableOid = slots[i]->tts_tableOid;
2384 3778662 : heaptuples[i] = heap_prepare_insert(relation, tuple, xid, cid,
2385 : options);
2386 : }
2387 :
2388 : /*
2389 : * We're about to do the actual inserts -- but check for conflict first,
2390 : * to minimize the possibility of having to roll back work we've just
2391 : * done.
2392 : *
2393 : * A check here does not definitively prevent a serialization anomaly;
2394 : * that check MUST be done at least past the point of acquiring an
2395 : * exclusive buffer content lock on every buffer that will be affected,
2396 : * and MAY be done after all inserts are reflected in the buffers and
2397 : * those locks are released; otherwise there is a race condition. Since
2398 : * multiple buffers can be locked and unlocked in the loop below, and it
2399 : * would not be feasible to identify and lock all of those buffers before
2400 : * the loop, we must do a final check at the end.
2401 : *
2402 : * The check here could be omitted with no loss of correctness; it is
2403 : * present strictly as an optimization.
2404 : *
2405 : * For heap inserts, we only need to check for table-level SSI locks. Our
2406 : * new tuples can't possibly conflict with existing tuple locks, and heap
2407 : * page locks are only consolidated versions of tuple locks; they do not
2408 : * lock "gaps" as index page locks do. So we don't need to specify a
2409 : * buffer when making the call, which makes for a faster check.
2410 : */
2411 713988 : CheckForSerializableConflictIn(relation, NULL, InvalidBlockNumber);
2412 :
2413 713988 : ndone = 0;
2414 1465460 : while (ndone < ntuples)
2415 : {
2416 : Buffer buffer;
2417 751472 : bool all_visible_cleared = false;
2418 751472 : bool all_frozen_set = false;
2419 : int nthispage;
2420 :
2421 751472 : CHECK_FOR_INTERRUPTS();
2422 :
2423 : /*
2424 : * Compute number of pages needed to fit the to-be-inserted tuples in
2425 : * the worst case. This will be used to determine how much to extend
2426 : * the relation by in RelationGetBufferForTuple(), if needed. If we
2427 : * filled a prior page from scratch, we can just update our last
2428 : * computation, but if we started with a partially filled page,
2429 : * recompute from scratch, the number of potentially required pages
2430 : * can vary due to tuples needing to fit onto the page, page headers
2431 : * etc.
2432 : */
2433 751472 : if (ndone == 0 || !starting_with_empty_page)
2434 : {
2435 728436 : npages = heap_multi_insert_pages(heaptuples, ndone, ntuples,
2436 : saveFreeSpace);
2437 728436 : npages_used = 0;
2438 : }
2439 : else
2440 23036 : npages_used++;
2441 :
2442 : /*
2443 : * Find buffer where at least the next tuple will fit. If the page is
2444 : * all-visible, this will also pin the requisite visibility map page.
2445 : *
2446 : * Also pin visibility map page if COPY FREEZE inserts tuples into an
2447 : * empty page. See all_frozen_set below.
2448 : */
2449 751472 : buffer = RelationGetBufferForTuple(relation, heaptuples[ndone]->t_len,
2450 : InvalidBuffer, options, bistate,
2451 : &vmbuffer, NULL,
2452 : npages - npages_used);
2453 751472 : page = BufferGetPage(buffer);
2454 :
2455 751472 : starting_with_empty_page = PageGetMaxOffsetNumber(page) == 0;
2456 :
2457 751472 : if (starting_with_empty_page && (options & HEAP_INSERT_FROZEN))
2458 3322 : all_frozen_set = true;
2459 :
2460 : /* NO EREPORT(ERROR) from here till changes are logged */
2461 751472 : START_CRIT_SECTION();
2462 :
2463 : /*
2464 : * RelationGetBufferForTuple has ensured that the first tuple fits.
2465 : * Put that on the page, and then as many other tuples as fit.
2466 : */
2467 751472 : RelationPutHeapTuple(relation, buffer, heaptuples[ndone], false);
2468 :
2469 : /*
2470 : * For logical decoding we need combo CIDs to properly decode the
2471 : * catalog.
2472 : */
2473 751472 : if (needwal && need_cids)
2474 9752 : log_heap_new_cid(relation, heaptuples[ndone]);
2475 :
2476 3778662 : for (nthispage = 1; ndone + nthispage < ntuples; nthispage++)
2477 : {
2478 3064674 : HeapTuple heaptup = heaptuples[ndone + nthispage];
2479 :
2480 3064674 : if (PageGetHeapFreeSpace(page) < MAXALIGN(heaptup->t_len) + saveFreeSpace)
2481 37484 : break;
2482 :
2483 3027190 : RelationPutHeapTuple(relation, buffer, heaptup, false);
2484 :
2485 : /*
2486 : * For logical decoding we need combo CIDs to properly decode the
2487 : * catalog.
2488 : */
2489 3027190 : if (needwal && need_cids)
2490 9194 : log_heap_new_cid(relation, heaptup);
2491 : }
2492 :
2493 : /*
2494 : * If the page is all visible, need to clear that, unless we're only
2495 : * going to add further frozen rows to it.
2496 : *
2497 : * If we're only adding already frozen rows to a previously empty
2498 : * page, mark it as all-visible.
2499 : */
2500 751472 : if (PageIsAllVisible(page) && !(options & HEAP_INSERT_FROZEN))
2501 : {
2502 6954 : all_visible_cleared = true;
2503 6954 : PageClearAllVisible(page);
2504 6954 : visibilitymap_clear(relation,
2505 : BufferGetBlockNumber(buffer),
2506 : vmbuffer, VISIBILITYMAP_VALID_BITS);
2507 : }
2508 744518 : else if (all_frozen_set)
2509 3322 : PageSetAllVisible(page);
2510 :
2511 : /*
2512 : * XXX Should we set PageSetPrunable on this page ? See heap_insert()
2513 : */
2514 :
2515 751472 : MarkBufferDirty(buffer);
2516 :
2517 : /* XLOG stuff */
2518 751472 : if (needwal)
2519 : {
2520 : XLogRecPtr recptr;
2521 : xl_heap_multi_insert *xlrec;
2522 734724 : uint8 info = XLOG_HEAP2_MULTI_INSERT;
2523 : char *tupledata;
2524 : int totaldatalen;
2525 734724 : char *scratchptr = scratch.data;
2526 : bool init;
2527 734724 : int bufflags = 0;
2528 :
2529 : /*
2530 : * If the page was previously empty, we can reinit the page
2531 : * instead of restoring the whole thing.
2532 : */
2533 734724 : init = starting_with_empty_page;
2534 :
2535 : /* allocate xl_heap_multi_insert struct from the scratch area */
2536 734724 : xlrec = (xl_heap_multi_insert *) scratchptr;
2537 734724 : scratchptr += SizeOfHeapMultiInsert;
2538 :
2539 : /*
2540 : * Allocate offsets array. Unless we're reinitializing the page,
2541 : * in that case the tuples are stored in order starting at
2542 : * FirstOffsetNumber and we don't need to store the offsets
2543 : * explicitly.
2544 : */
2545 734724 : if (!init)
2546 708574 : scratchptr += nthispage * sizeof(OffsetNumber);
2547 :
2548 : /* the rest of the scratch space is used for tuple data */
2549 734724 : tupledata = scratchptr;
2550 :
2551 : /* check that the mutually exclusive flags are not both set */
2552 : Assert(!(all_visible_cleared && all_frozen_set));
2553 :
2554 734724 : xlrec->flags = 0;
2555 734724 : if (all_visible_cleared)
2556 6954 : xlrec->flags = XLH_INSERT_ALL_VISIBLE_CLEARED;
2557 734724 : if (all_frozen_set)
2558 26 : xlrec->flags = XLH_INSERT_ALL_FROZEN_SET;
2559 :
2560 734724 : xlrec->ntuples = nthispage;
2561 :
2562 : /*
2563 : * Write out an xl_multi_insert_tuple and the tuple data itself
2564 : * for each tuple.
2565 : */
2566 3265450 : for (i = 0; i < nthispage; i++)
2567 : {
2568 2530726 : HeapTuple heaptup = heaptuples[ndone + i];
2569 : xl_multi_insert_tuple *tuphdr;
2570 : int datalen;
2571 :
2572 2530726 : if (!init)
2573 1486696 : xlrec->offsets[i] = ItemPointerGetOffsetNumber(&heaptup->t_self);
2574 : /* xl_multi_insert_tuple needs two-byte alignment. */
2575 2530726 : tuphdr = (xl_multi_insert_tuple *) SHORTALIGN(scratchptr);
2576 2530726 : scratchptr = ((char *) tuphdr) + SizeOfMultiInsertTuple;
2577 :
2578 2530726 : tuphdr->t_infomask2 = heaptup->t_data->t_infomask2;
2579 2530726 : tuphdr->t_infomask = heaptup->t_data->t_infomask;
2580 2530726 : tuphdr->t_hoff = heaptup->t_data->t_hoff;
2581 :
2582 : /* write bitmap [+ padding] [+ oid] + data */
2583 2530726 : datalen = heaptup->t_len - SizeofHeapTupleHeader;
2584 2530726 : memcpy(scratchptr,
2585 2530726 : (char *) heaptup->t_data + SizeofHeapTupleHeader,
2586 : datalen);
2587 2530726 : tuphdr->datalen = datalen;
2588 2530726 : scratchptr += datalen;
2589 : }
2590 734724 : totaldatalen = scratchptr - tupledata;
2591 : Assert((scratchptr - scratch.data) < BLCKSZ);
2592 :
2593 734724 : if (need_tuple_data)
2594 144 : xlrec->flags |= XLH_INSERT_CONTAINS_NEW_TUPLE;
2595 :
2596 : /*
2597 : * Signal that this is the last xl_heap_multi_insert record
2598 : * emitted by this call to heap_multi_insert(). Needed for logical
2599 : * decoding so it knows when to cleanup temporary data.
2600 : */
2601 734724 : if (ndone + nthispage == ntuples)
2602 711470 : xlrec->flags |= XLH_INSERT_LAST_IN_MULTI;
2603 :
2604 734724 : if (init)
2605 : {
2606 26150 : info |= XLOG_HEAP_INIT_PAGE;
2607 26150 : bufflags |= REGBUF_WILL_INIT;
2608 : }
2609 :
2610 : /*
2611 : * If we're doing logical decoding, include the new tuple data
2612 : * even if we take a full-page image of the page.
2613 : */
2614 734724 : if (need_tuple_data)
2615 144 : bufflags |= REGBUF_KEEP_DATA;
2616 :
2617 734724 : XLogBeginInsert();
2618 734724 : XLogRegisterData(xlrec, tupledata - scratch.data);
2619 734724 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2620 :
2621 734724 : XLogRegisterBufData(0, tupledata, totaldatalen);
2622 :
2623 : /* filtering by origin on a row level is much more efficient */
2624 734724 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
2625 :
2626 734724 : recptr = XLogInsert(RM_HEAP2_ID, info);
2627 :
2628 734724 : PageSetLSN(page, recptr);
2629 : }
2630 :
2631 751472 : END_CRIT_SECTION();
2632 :
2633 : /*
2634 : * If we've frozen everything on the page, update the visibilitymap.
2635 : * We're already holding pin on the vmbuffer.
2636 : */
2637 751472 : if (all_frozen_set)
2638 : {
2639 : Assert(PageIsAllVisible(page));
2640 : Assert(visibilitymap_pin_ok(BufferGetBlockNumber(buffer), vmbuffer));
2641 :
2642 : /*
2643 : * It's fine to use InvalidTransactionId here - this is only used
2644 : * when HEAP_INSERT_FROZEN is specified, which intentionally
2645 : * violates visibility rules.
2646 : */
2647 3322 : visibilitymap_set(relation, BufferGetBlockNumber(buffer), buffer,
2648 : InvalidXLogRecPtr, vmbuffer,
2649 : InvalidTransactionId,
2650 : VISIBILITYMAP_ALL_VISIBLE | VISIBILITYMAP_ALL_FROZEN);
2651 : }
2652 :
2653 751472 : UnlockReleaseBuffer(buffer);
2654 751472 : ndone += nthispage;
2655 :
2656 : /*
2657 : * NB: Only release vmbuffer after inserting all tuples - it's fairly
2658 : * likely that we'll insert into subsequent heap pages that are likely
2659 : * to use the same vm page.
2660 : */
2661 : }
2662 :
2663 : /* We're done with inserting all tuples, so release the last vmbuffer. */
2664 713988 : if (vmbuffer != InvalidBuffer)
2665 7118 : ReleaseBuffer(vmbuffer);
2666 :
2667 : /*
2668 : * We're done with the actual inserts. Check for conflicts again, to
2669 : * ensure that all rw-conflicts in to these inserts are detected. Without
2670 : * this final check, a sequential scan of the heap may have locked the
2671 : * table after the "before" check, missing one opportunity to detect the
2672 : * conflict, and then scanned the table before the new tuples were there,
2673 : * missing the other chance to detect the conflict.
2674 : *
2675 : * For heap inserts, we only need to check for table-level SSI locks. Our
2676 : * new tuples can't possibly conflict with existing tuple locks, and heap
2677 : * page locks are only consolidated versions of tuple locks; they do not
2678 : * lock "gaps" as index page locks do. So we don't need to specify a
2679 : * buffer when making the call.
2680 : */
2681 713988 : CheckForSerializableConflictIn(relation, NULL, InvalidBlockNumber);
2682 :
2683 : /*
2684 : * If tuples are cachable, mark them for invalidation from the caches in
2685 : * case we abort. Note it is OK to do this after releasing the buffer,
2686 : * because the heaptuples data structure is all in local memory, not in
2687 : * the shared buffer.
2688 : */
2689 713988 : if (IsCatalogRelation(relation))
2690 : {
2691 2449744 : for (i = 0; i < ntuples; i++)
2692 1739902 : CacheInvalidateHeapTuple(relation, heaptuples[i], NULL);
2693 : }
2694 :
2695 : /* copy t_self fields back to the caller's slots */
2696 4492650 : for (i = 0; i < ntuples; i++)
2697 3778662 : slots[i]->tts_tid = heaptuples[i]->t_self;
2698 :
2699 713988 : pgstat_count_heap_insert(relation, ntuples);
2700 713988 : }
2701 :
2702 : /*
2703 : * simple_heap_insert - insert a tuple
2704 : *
2705 : * Currently, this routine differs from heap_insert only in supplying
2706 : * a default command ID and not allowing access to the speedup options.
2707 : *
2708 : * This should be used rather than using heap_insert directly in most places
2709 : * where we are modifying system catalogs.
2710 : */
2711 : void
2712 1767710 : simple_heap_insert(Relation relation, HeapTuple tup)
2713 : {
2714 1767710 : heap_insert(relation, tup, GetCurrentCommandId(true), 0, NULL);
2715 1767710 : }
2716 :
2717 : /*
2718 : * Given infomask/infomask2, compute the bits that must be saved in the
2719 : * "infobits" field of xl_heap_delete, xl_heap_update, xl_heap_lock,
2720 : * xl_heap_lock_updated WAL records.
2721 : *
2722 : * See fix_infomask_from_infobits.
2723 : */
2724 : static uint8
2725 3863480 : compute_infobits(uint16 infomask, uint16 infomask2)
2726 : {
2727 : return
2728 3863480 : ((infomask & HEAP_XMAX_IS_MULTI) != 0 ? XLHL_XMAX_IS_MULTI : 0) |
2729 3863480 : ((infomask & HEAP_XMAX_LOCK_ONLY) != 0 ? XLHL_XMAX_LOCK_ONLY : 0) |
2730 3863480 : ((infomask & HEAP_XMAX_EXCL_LOCK) != 0 ? XLHL_XMAX_EXCL_LOCK : 0) |
2731 : /* note we ignore HEAP_XMAX_SHR_LOCK here */
2732 7726960 : ((infomask & HEAP_XMAX_KEYSHR_LOCK) != 0 ? XLHL_XMAX_KEYSHR_LOCK : 0) |
2733 : ((infomask2 & HEAP_KEYS_UPDATED) != 0 ?
2734 3863480 : XLHL_KEYS_UPDATED : 0);
2735 : }
2736 :
2737 : /*
2738 : * Given two versions of the same t_infomask for a tuple, compare them and
2739 : * return whether the relevant status for a tuple Xmax has changed. This is
2740 : * used after a buffer lock has been released and reacquired: we want to ensure
2741 : * that the tuple state continues to be the same it was when we previously
2742 : * examined it.
2743 : *
2744 : * Note the Xmax field itself must be compared separately.
2745 : */
2746 : static inline bool
2747 10662 : xmax_infomask_changed(uint16 new_infomask, uint16 old_infomask)
2748 : {
2749 10662 : const uint16 interesting =
2750 : HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY | HEAP_LOCK_MASK;
2751 :
2752 10662 : if ((new_infomask & interesting) != (old_infomask & interesting))
2753 28 : return true;
2754 :
2755 10634 : return false;
2756 : }
2757 :
2758 : /*
2759 : * heap_delete - delete a tuple
2760 : *
2761 : * See table_tuple_delete() for an explanation of the parameters, except that
2762 : * this routine directly takes a tuple rather than a slot.
2763 : *
2764 : * In the failure cases, the routine fills *tmfd with the tuple's t_ctid,
2765 : * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax (the last
2766 : * only for TM_SelfModified, since we cannot obtain cmax from a combo CID
2767 : * generated by another transaction).
2768 : */
2769 : TM_Result
2770 2952762 : heap_delete(Relation relation, ItemPointer tid,
2771 : CommandId cid, Snapshot crosscheck, bool wait,
2772 : TM_FailureData *tmfd, bool changingPart)
2773 : {
2774 : TM_Result result;
2775 2952762 : TransactionId xid = GetCurrentTransactionId();
2776 : ItemId lp;
2777 : HeapTupleData tp;
2778 : Page page;
2779 : BlockNumber block;
2780 : Buffer buffer;
2781 2952762 : Buffer vmbuffer = InvalidBuffer;
2782 : TransactionId new_xmax;
2783 : uint16 new_infomask,
2784 : new_infomask2;
2785 2952762 : bool have_tuple_lock = false;
2786 : bool iscombo;
2787 2952762 : bool all_visible_cleared = false;
2788 2952762 : HeapTuple old_key_tuple = NULL; /* replica identity of the tuple */
2789 2952762 : bool old_key_copied = false;
2790 :
2791 : Assert(ItemPointerIsValid(tid));
2792 :
2793 2952762 : AssertHasSnapshotForToast(relation);
2794 :
2795 : /*
2796 : * Forbid this during a parallel operation, lest it allocate a combo CID.
2797 : * Other workers might need that combo CID for visibility checks, and we
2798 : * have no provision for broadcasting it to them.
2799 : */
2800 2952762 : if (IsInParallelMode())
2801 0 : ereport(ERROR,
2802 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
2803 : errmsg("cannot delete tuples during a parallel operation")));
2804 :
2805 2952762 : block = ItemPointerGetBlockNumber(tid);
2806 2952762 : buffer = ReadBuffer(relation, block);
2807 2952762 : page = BufferGetPage(buffer);
2808 :
2809 : /*
2810 : * Before locking the buffer, pin the visibility map page if it appears to
2811 : * be necessary. Since we haven't got the lock yet, someone else might be
2812 : * in the middle of changing this, so we'll need to recheck after we have
2813 : * the lock.
2814 : */
2815 2952762 : if (PageIsAllVisible(page))
2816 468 : visibilitymap_pin(relation, block, &vmbuffer);
2817 :
2818 2952762 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2819 :
2820 2952762 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
2821 : Assert(ItemIdIsNormal(lp));
2822 :
2823 2952762 : tp.t_tableOid = RelationGetRelid(relation);
2824 2952762 : tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
2825 2952762 : tp.t_len = ItemIdGetLength(lp);
2826 2952762 : tp.t_self = *tid;
2827 :
2828 2 : l1:
2829 :
2830 : /*
2831 : * If we didn't pin the visibility map page and the page has become all
2832 : * visible while we were busy locking the buffer, we'll have to unlock and
2833 : * re-lock, to avoid holding the buffer lock across an I/O. That's a bit
2834 : * unfortunate, but hopefully shouldn't happen often.
2835 : */
2836 2952764 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
2837 : {
2838 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2839 0 : visibilitymap_pin(relation, block, &vmbuffer);
2840 0 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2841 : }
2842 :
2843 2952764 : result = HeapTupleSatisfiesUpdate(&tp, cid, buffer);
2844 :
2845 2952764 : if (result == TM_Invisible)
2846 : {
2847 0 : UnlockReleaseBuffer(buffer);
2848 0 : ereport(ERROR,
2849 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
2850 : errmsg("attempted to delete invisible tuple")));
2851 : }
2852 2952764 : else if (result == TM_BeingModified && wait)
2853 : {
2854 : TransactionId xwait;
2855 : uint16 infomask;
2856 :
2857 : /* must copy state data before unlocking buffer */
2858 81080 : xwait = HeapTupleHeaderGetRawXmax(tp.t_data);
2859 81080 : infomask = tp.t_data->t_infomask;
2860 :
2861 : /*
2862 : * Sleep until concurrent transaction ends -- except when there's a
2863 : * single locker and it's our own transaction. Note we don't care
2864 : * which lock mode the locker has, because we need the strongest one.
2865 : *
2866 : * Before sleeping, we need to acquire tuple lock to establish our
2867 : * priority for the tuple (see heap_lock_tuple). LockTuple will
2868 : * release us when we are next-in-line for the tuple.
2869 : *
2870 : * If we are forced to "start over" below, we keep the tuple lock;
2871 : * this arranges that we stay at the head of the line while rechecking
2872 : * tuple state.
2873 : */
2874 81080 : if (infomask & HEAP_XMAX_IS_MULTI)
2875 : {
2876 16 : bool current_is_member = false;
2877 :
2878 16 : if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
2879 : LockTupleExclusive, ¤t_is_member))
2880 : {
2881 16 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2882 :
2883 : /*
2884 : * Acquire the lock, if necessary (but skip it when we're
2885 : * requesting a lock and already have one; avoids deadlock).
2886 : */
2887 16 : if (!current_is_member)
2888 12 : heap_acquire_tuplock(relation, &(tp.t_self), LockTupleExclusive,
2889 : LockWaitBlock, &have_tuple_lock);
2890 :
2891 : /* wait for multixact */
2892 16 : MultiXactIdWait((MultiXactId) xwait, MultiXactStatusUpdate, infomask,
2893 : relation, &(tp.t_self), XLTW_Delete,
2894 : NULL);
2895 16 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2896 :
2897 : /*
2898 : * If xwait had just locked the tuple then some other xact
2899 : * could update this tuple before we get to this point. Check
2900 : * for xmax change, and start over if so.
2901 : *
2902 : * We also must start over if we didn't pin the VM page, and
2903 : * the page has become all visible.
2904 : */
2905 32 : if ((vmbuffer == InvalidBuffer && PageIsAllVisible(page)) ||
2906 32 : xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2907 16 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2908 : xwait))
2909 0 : goto l1;
2910 : }
2911 :
2912 : /*
2913 : * You might think the multixact is necessarily done here, but not
2914 : * so: it could have surviving members, namely our own xact or
2915 : * other subxacts of this backend. It is legal for us to delete
2916 : * the tuple in either case, however (the latter case is
2917 : * essentially a situation of upgrading our former shared lock to
2918 : * exclusive). We don't bother changing the on-disk hint bits
2919 : * since we are about to overwrite the xmax altogether.
2920 : */
2921 : }
2922 81064 : else if (!TransactionIdIsCurrentTransactionId(xwait))
2923 : {
2924 : /*
2925 : * Wait for regular transaction to end; but first, acquire tuple
2926 : * lock.
2927 : */
2928 80 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2929 80 : heap_acquire_tuplock(relation, &(tp.t_self), LockTupleExclusive,
2930 : LockWaitBlock, &have_tuple_lock);
2931 80 : XactLockTableWait(xwait, relation, &(tp.t_self), XLTW_Delete);
2932 72 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2933 :
2934 : /*
2935 : * xwait is done, but if xwait had just locked the tuple then some
2936 : * other xact could update this tuple before we get to this point.
2937 : * Check for xmax change, and start over if so.
2938 : *
2939 : * We also must start over if we didn't pin the VM page, and the
2940 : * page has become all visible.
2941 : */
2942 144 : if ((vmbuffer == InvalidBuffer && PageIsAllVisible(page)) ||
2943 142 : xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2944 70 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2945 : xwait))
2946 2 : goto l1;
2947 :
2948 : /* Otherwise check if it committed or aborted */
2949 70 : UpdateXmaxHintBits(tp.t_data, buffer, xwait);
2950 : }
2951 :
2952 : /*
2953 : * We may overwrite if previous xmax aborted, or if it committed but
2954 : * only locked the tuple without updating it.
2955 : */
2956 162112 : if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
2957 81092 : HEAP_XMAX_IS_LOCKED_ONLY(tp.t_data->t_infomask) ||
2958 50 : HeapTupleHeaderIsOnlyLocked(tp.t_data))
2959 81028 : result = TM_Ok;
2960 42 : else if (!ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid))
2961 34 : result = TM_Updated;
2962 : else
2963 8 : result = TM_Deleted;
2964 : }
2965 :
2966 : /* sanity check the result HeapTupleSatisfiesUpdate() and the logic above */
2967 : if (result != TM_Ok)
2968 : {
2969 : Assert(result == TM_SelfModified ||
2970 : result == TM_Updated ||
2971 : result == TM_Deleted ||
2972 : result == TM_BeingModified);
2973 : Assert(!(tp.t_data->t_infomask & HEAP_XMAX_INVALID));
2974 : Assert(result != TM_Updated ||
2975 : !ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid));
2976 : }
2977 :
2978 2952754 : if (crosscheck != InvalidSnapshot && result == TM_Ok)
2979 : {
2980 : /* Perform additional check for transaction-snapshot mode RI updates */
2981 2 : if (!HeapTupleSatisfiesVisibility(&tp, crosscheck, buffer))
2982 2 : result = TM_Updated;
2983 : }
2984 :
2985 2952754 : if (result != TM_Ok)
2986 : {
2987 112 : tmfd->ctid = tp.t_data->t_ctid;
2988 112 : tmfd->xmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
2989 112 : if (result == TM_SelfModified)
2990 42 : tmfd->cmax = HeapTupleHeaderGetCmax(tp.t_data);
2991 : else
2992 70 : tmfd->cmax = InvalidCommandId;
2993 112 : UnlockReleaseBuffer(buffer);
2994 112 : if (have_tuple_lock)
2995 42 : UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
2996 112 : if (vmbuffer != InvalidBuffer)
2997 0 : ReleaseBuffer(vmbuffer);
2998 112 : return result;
2999 : }
3000 :
3001 : /*
3002 : * We're about to do the actual delete -- check for conflict first, to
3003 : * avoid possibly having to roll back work we've just done.
3004 : *
3005 : * This is safe without a recheck as long as there is no possibility of
3006 : * another process scanning the page between this check and the delete
3007 : * being visible to the scan (i.e., an exclusive buffer content lock is
3008 : * continuously held from this point until the tuple delete is visible).
3009 : */
3010 2952642 : CheckForSerializableConflictIn(relation, tid, BufferGetBlockNumber(buffer));
3011 :
3012 : /* replace cid with a combo CID if necessary */
3013 2952614 : HeapTupleHeaderAdjustCmax(tp.t_data, &cid, &iscombo);
3014 :
3015 : /*
3016 : * Compute replica identity tuple before entering the critical section so
3017 : * we don't PANIC upon a memory allocation failure.
3018 : */
3019 2952614 : old_key_tuple = ExtractReplicaIdentity(relation, &tp, true, &old_key_copied);
3020 :
3021 : /*
3022 : * If this is the first possibly-multixact-able operation in the current
3023 : * transaction, set my per-backend OldestMemberMXactId setting. We can be
3024 : * certain that the transaction will never become a member of any older
3025 : * MultiXactIds than that. (We have to do this even if we end up just
3026 : * using our own TransactionId below, since some other backend could
3027 : * incorporate our XID into a MultiXact immediately afterwards.)
3028 : */
3029 2952614 : MultiXactIdSetOldestMember();
3030 :
3031 2952614 : compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(tp.t_data),
3032 2952614 : tp.t_data->t_infomask, tp.t_data->t_infomask2,
3033 : xid, LockTupleExclusive, true,
3034 : &new_xmax, &new_infomask, &new_infomask2);
3035 :
3036 2952614 : START_CRIT_SECTION();
3037 :
3038 : /*
3039 : * If this transaction commits, the tuple will become DEAD sooner or
3040 : * later. Set flag that this page is a candidate for pruning once our xid
3041 : * falls below the OldestXmin horizon. If the transaction finally aborts,
3042 : * the subsequent page pruning will be a no-op and the hint will be
3043 : * cleared.
3044 : */
3045 2952614 : PageSetPrunable(page, xid);
3046 :
3047 2952614 : if (PageIsAllVisible(page))
3048 : {
3049 468 : all_visible_cleared = true;
3050 468 : PageClearAllVisible(page);
3051 468 : visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
3052 : vmbuffer, VISIBILITYMAP_VALID_BITS);
3053 : }
3054 :
3055 : /* store transaction information of xact deleting the tuple */
3056 2952614 : tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3057 2952614 : tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3058 2952614 : tp.t_data->t_infomask |= new_infomask;
3059 2952614 : tp.t_data->t_infomask2 |= new_infomask2;
3060 2952614 : HeapTupleHeaderClearHotUpdated(tp.t_data);
3061 2952614 : HeapTupleHeaderSetXmax(tp.t_data, new_xmax);
3062 2952614 : HeapTupleHeaderSetCmax(tp.t_data, cid, iscombo);
3063 : /* Make sure there is no forward chain link in t_ctid */
3064 2952614 : tp.t_data->t_ctid = tp.t_self;
3065 :
3066 : /* Signal that this is actually a move into another partition */
3067 2952614 : if (changingPart)
3068 964 : HeapTupleHeaderSetMovedPartitions(tp.t_data);
3069 :
3070 2952614 : MarkBufferDirty(buffer);
3071 :
3072 : /*
3073 : * XLOG stuff
3074 : *
3075 : * NB: heap_abort_speculative() uses the same xlog record and replay
3076 : * routines.
3077 : */
3078 2952614 : if (RelationNeedsWAL(relation))
3079 : {
3080 : xl_heap_delete xlrec;
3081 : xl_heap_header xlhdr;
3082 : XLogRecPtr recptr;
3083 :
3084 : /*
3085 : * For logical decode we need combo CIDs to properly decode the
3086 : * catalog
3087 : */
3088 2831396 : if (RelationIsAccessibleInLogicalDecoding(relation))
3089 12414 : log_heap_new_cid(relation, &tp);
3090 :
3091 2831396 : xlrec.flags = 0;
3092 2831396 : if (all_visible_cleared)
3093 468 : xlrec.flags |= XLH_DELETE_ALL_VISIBLE_CLEARED;
3094 2831396 : if (changingPart)
3095 964 : xlrec.flags |= XLH_DELETE_IS_PARTITION_MOVE;
3096 5662792 : xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
3097 2831396 : tp.t_data->t_infomask2);
3098 2831396 : xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
3099 2831396 : xlrec.xmax = new_xmax;
3100 :
3101 2831396 : if (old_key_tuple != NULL)
3102 : {
3103 94026 : if (relation->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
3104 256 : xlrec.flags |= XLH_DELETE_CONTAINS_OLD_TUPLE;
3105 : else
3106 93770 : xlrec.flags |= XLH_DELETE_CONTAINS_OLD_KEY;
3107 : }
3108 :
3109 2831396 : XLogBeginInsert();
3110 2831396 : XLogRegisterData(&xlrec, SizeOfHeapDelete);
3111 :
3112 2831396 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
3113 :
3114 : /*
3115 : * Log replica identity of the deleted tuple if there is one
3116 : */
3117 2831396 : if (old_key_tuple != NULL)
3118 : {
3119 94026 : xlhdr.t_infomask2 = old_key_tuple->t_data->t_infomask2;
3120 94026 : xlhdr.t_infomask = old_key_tuple->t_data->t_infomask;
3121 94026 : xlhdr.t_hoff = old_key_tuple->t_data->t_hoff;
3122 :
3123 94026 : XLogRegisterData(&xlhdr, SizeOfHeapHeader);
3124 94026 : XLogRegisterData((char *) old_key_tuple->t_data
3125 : + SizeofHeapTupleHeader,
3126 94026 : old_key_tuple->t_len
3127 : - SizeofHeapTupleHeader);
3128 : }
3129 :
3130 : /* filtering by origin on a row level is much more efficient */
3131 2831396 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
3132 :
3133 2831396 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
3134 :
3135 2831396 : PageSetLSN(page, recptr);
3136 : }
3137 :
3138 2952614 : END_CRIT_SECTION();
3139 :
3140 2952614 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3141 :
3142 2952614 : if (vmbuffer != InvalidBuffer)
3143 468 : ReleaseBuffer(vmbuffer);
3144 :
3145 : /*
3146 : * If the tuple has toasted out-of-line attributes, we need to delete
3147 : * those items too. We have to do this before releasing the buffer
3148 : * because we need to look at the contents of the tuple, but it's OK to
3149 : * release the content lock on the buffer first.
3150 : */
3151 2952614 : if (relation->rd_rel->relkind != RELKIND_RELATION &&
3152 4802 : relation->rd_rel->relkind != RELKIND_MATVIEW)
3153 : {
3154 : /* toast table entries should never be recursively toasted */
3155 : Assert(!HeapTupleHasExternal(&tp));
3156 : }
3157 2947832 : else if (HeapTupleHasExternal(&tp))
3158 586 : heap_toast_delete(relation, &tp, false);
3159 :
3160 : /*
3161 : * Mark tuple for invalidation from system caches at next command
3162 : * boundary. We have to do this before releasing the buffer because we
3163 : * need to look at the contents of the tuple.
3164 : */
3165 2952614 : CacheInvalidateHeapTuple(relation, &tp, NULL);
3166 :
3167 : /* Now we can release the buffer */
3168 2952614 : ReleaseBuffer(buffer);
3169 :
3170 : /*
3171 : * Release the lmgr tuple lock, if we had it.
3172 : */
3173 2952614 : if (have_tuple_lock)
3174 40 : UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
3175 :
3176 2952614 : pgstat_count_heap_delete(relation);
3177 :
3178 2952614 : if (old_key_tuple != NULL && old_key_copied)
3179 93772 : heap_freetuple(old_key_tuple);
3180 :
3181 2952614 : return TM_Ok;
3182 : }
3183 :
3184 : /*
3185 : * simple_heap_delete - delete a tuple
3186 : *
3187 : * This routine may be used to delete a tuple when concurrent updates of
3188 : * the target tuple are not expected (for example, because we have a lock
3189 : * on the relation associated with the tuple). Any failure is reported
3190 : * via ereport().
3191 : */
3192 : void
3193 1226274 : simple_heap_delete(Relation relation, ItemPointer tid)
3194 : {
3195 : TM_Result result;
3196 : TM_FailureData tmfd;
3197 :
3198 1226274 : result = heap_delete(relation, tid,
3199 : GetCurrentCommandId(true), InvalidSnapshot,
3200 : true /* wait for commit */ ,
3201 : &tmfd, false /* changingPart */ );
3202 1226274 : switch (result)
3203 : {
3204 0 : case TM_SelfModified:
3205 : /* Tuple was already updated in current command? */
3206 0 : elog(ERROR, "tuple already updated by self");
3207 : break;
3208 :
3209 1226274 : case TM_Ok:
3210 : /* done successfully */
3211 1226274 : break;
3212 :
3213 0 : case TM_Updated:
3214 0 : elog(ERROR, "tuple concurrently updated");
3215 : break;
3216 :
3217 0 : case TM_Deleted:
3218 0 : elog(ERROR, "tuple concurrently deleted");
3219 : break;
3220 :
3221 0 : default:
3222 0 : elog(ERROR, "unrecognized heap_delete status: %u", result);
3223 : break;
3224 : }
3225 1226274 : }
3226 :
3227 : /*
3228 : * heap_update - replace a tuple
3229 : *
3230 : * See table_tuple_update() for an explanation of the parameters, except that
3231 : * this routine directly takes a tuple rather than a slot.
3232 : *
3233 : * In the failure cases, the routine fills *tmfd with the tuple's t_ctid,
3234 : * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax (the last
3235 : * only for TM_SelfModified, since we cannot obtain cmax from a combo CID
3236 : * generated by another transaction).
3237 : */
3238 : TM_Result
3239 609344 : heap_update(Relation relation, ItemPointer otid, HeapTuple newtup,
3240 : CommandId cid, Snapshot crosscheck, bool wait,
3241 : TM_FailureData *tmfd, LockTupleMode *lockmode,
3242 : TU_UpdateIndexes *update_indexes)
3243 : {
3244 : TM_Result result;
3245 609344 : TransactionId xid = GetCurrentTransactionId();
3246 : Bitmapset *hot_attrs;
3247 : Bitmapset *sum_attrs;
3248 : Bitmapset *key_attrs;
3249 : Bitmapset *id_attrs;
3250 : Bitmapset *interesting_attrs;
3251 : Bitmapset *modified_attrs;
3252 : ItemId lp;
3253 : HeapTupleData oldtup;
3254 : HeapTuple heaptup;
3255 609344 : HeapTuple old_key_tuple = NULL;
3256 609344 : bool old_key_copied = false;
3257 : Page page;
3258 : BlockNumber block;
3259 : MultiXactStatus mxact_status;
3260 : Buffer buffer,
3261 : newbuf,
3262 609344 : vmbuffer = InvalidBuffer,
3263 609344 : vmbuffer_new = InvalidBuffer;
3264 : bool need_toast;
3265 : Size newtupsize,
3266 : pagefree;
3267 609344 : bool have_tuple_lock = false;
3268 : bool iscombo;
3269 609344 : bool use_hot_update = false;
3270 609344 : bool summarized_update = false;
3271 : bool key_intact;
3272 609344 : bool all_visible_cleared = false;
3273 609344 : bool all_visible_cleared_new = false;
3274 : bool checked_lockers;
3275 : bool locker_remains;
3276 609344 : bool id_has_external = false;
3277 : TransactionId xmax_new_tuple,
3278 : xmax_old_tuple;
3279 : uint16 infomask_old_tuple,
3280 : infomask2_old_tuple,
3281 : infomask_new_tuple,
3282 : infomask2_new_tuple;
3283 :
3284 : Assert(ItemPointerIsValid(otid));
3285 :
3286 : /* Cheap, simplistic check that the tuple matches the rel's rowtype. */
3287 : Assert(HeapTupleHeaderGetNatts(newtup->t_data) <=
3288 : RelationGetNumberOfAttributes(relation));
3289 :
3290 609344 : AssertHasSnapshotForToast(relation);
3291 :
3292 : /*
3293 : * Forbid this during a parallel operation, lest it allocate a combo CID.
3294 : * Other workers might need that combo CID for visibility checks, and we
3295 : * have no provision for broadcasting it to them.
3296 : */
3297 609344 : if (IsInParallelMode())
3298 0 : ereport(ERROR,
3299 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
3300 : errmsg("cannot update tuples during a parallel operation")));
3301 :
3302 : #ifdef USE_ASSERT_CHECKING
3303 : check_lock_if_inplace_updateable_rel(relation, otid, newtup);
3304 : #endif
3305 :
3306 : /*
3307 : * Fetch the list of attributes to be checked for various operations.
3308 : *
3309 : * For HOT considerations, this is wasted effort if we fail to update or
3310 : * have to put the new tuple on a different page. But we must compute the
3311 : * list before obtaining buffer lock --- in the worst case, if we are
3312 : * doing an update on one of the relevant system catalogs, we could
3313 : * deadlock if we try to fetch the list later. In any case, the relcache
3314 : * caches the data so this is usually pretty cheap.
3315 : *
3316 : * We also need columns used by the replica identity and columns that are
3317 : * considered the "key" of rows in the table.
3318 : *
3319 : * Note that we get copies of each bitmap, so we need not worry about
3320 : * relcache flush happening midway through.
3321 : */
3322 609344 : hot_attrs = RelationGetIndexAttrBitmap(relation,
3323 : INDEX_ATTR_BITMAP_HOT_BLOCKING);
3324 609344 : sum_attrs = RelationGetIndexAttrBitmap(relation,
3325 : INDEX_ATTR_BITMAP_SUMMARIZED);
3326 609344 : key_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_KEY);
3327 609344 : id_attrs = RelationGetIndexAttrBitmap(relation,
3328 : INDEX_ATTR_BITMAP_IDENTITY_KEY);
3329 609344 : interesting_attrs = NULL;
3330 609344 : interesting_attrs = bms_add_members(interesting_attrs, hot_attrs);
3331 609344 : interesting_attrs = bms_add_members(interesting_attrs, sum_attrs);
3332 609344 : interesting_attrs = bms_add_members(interesting_attrs, key_attrs);
3333 609344 : interesting_attrs = bms_add_members(interesting_attrs, id_attrs);
3334 :
3335 609344 : block = ItemPointerGetBlockNumber(otid);
3336 609344 : INJECTION_POINT("heap_update-before-pin", NULL);
3337 609344 : buffer = ReadBuffer(relation, block);
3338 609344 : page = BufferGetPage(buffer);
3339 :
3340 : /*
3341 : * Before locking the buffer, pin the visibility map page if it appears to
3342 : * be necessary. Since we haven't got the lock yet, someone else might be
3343 : * in the middle of changing this, so we'll need to recheck after we have
3344 : * the lock.
3345 : */
3346 609344 : if (PageIsAllVisible(page))
3347 2994 : visibilitymap_pin(relation, block, &vmbuffer);
3348 :
3349 609344 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3350 :
3351 609344 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(otid));
3352 :
3353 : /*
3354 : * Usually, a buffer pin and/or snapshot blocks pruning of otid, ensuring
3355 : * we see LP_NORMAL here. When the otid origin is a syscache, we may have
3356 : * neither a pin nor a snapshot. Hence, we may see other LP_ states, each
3357 : * of which indicates concurrent pruning.
3358 : *
3359 : * Failing with TM_Updated would be most accurate. However, unlike other
3360 : * TM_Updated scenarios, we don't know the successor ctid in LP_UNUSED and
3361 : * LP_DEAD cases. While the distinction between TM_Updated and TM_Deleted
3362 : * does matter to SQL statements UPDATE and MERGE, those SQL statements
3363 : * hold a snapshot that ensures LP_NORMAL. Hence, the choice between
3364 : * TM_Updated and TM_Deleted affects only the wording of error messages.
3365 : * Settle on TM_Deleted, for two reasons. First, it avoids complicating
3366 : * the specification of when tmfd->ctid is valid. Second, it creates
3367 : * error log evidence that we took this branch.
3368 : *
3369 : * Since it's possible to see LP_UNUSED at otid, it's also possible to see
3370 : * LP_NORMAL for a tuple that replaced LP_UNUSED. If it's a tuple for an
3371 : * unrelated row, we'll fail with "duplicate key value violates unique".
3372 : * XXX if otid is the live, newer version of the newtup row, we'll discard
3373 : * changes originating in versions of this catalog row after the version
3374 : * the caller got from syscache. See syscache-update-pruned.spec.
3375 : */
3376 609344 : if (!ItemIdIsNormal(lp))
3377 : {
3378 : Assert(RelationSupportsSysCache(RelationGetRelid(relation)));
3379 :
3380 2 : UnlockReleaseBuffer(buffer);
3381 : Assert(!have_tuple_lock);
3382 2 : if (vmbuffer != InvalidBuffer)
3383 2 : ReleaseBuffer(vmbuffer);
3384 2 : tmfd->ctid = *otid;
3385 2 : tmfd->xmax = InvalidTransactionId;
3386 2 : tmfd->cmax = InvalidCommandId;
3387 2 : *update_indexes = TU_None;
3388 :
3389 2 : bms_free(hot_attrs);
3390 2 : bms_free(sum_attrs);
3391 2 : bms_free(key_attrs);
3392 2 : bms_free(id_attrs);
3393 : /* modified_attrs not yet initialized */
3394 2 : bms_free(interesting_attrs);
3395 2 : return TM_Deleted;
3396 : }
3397 :
3398 : /*
3399 : * Fill in enough data in oldtup for HeapDetermineColumnsInfo to work
3400 : * properly.
3401 : */
3402 609342 : oldtup.t_tableOid = RelationGetRelid(relation);
3403 609342 : oldtup.t_data = (HeapTupleHeader) PageGetItem(page, lp);
3404 609342 : oldtup.t_len = ItemIdGetLength(lp);
3405 609342 : oldtup.t_self = *otid;
3406 :
3407 : /* the new tuple is ready, except for this: */
3408 609342 : newtup->t_tableOid = RelationGetRelid(relation);
3409 :
3410 : /*
3411 : * Determine columns modified by the update. Additionally, identify
3412 : * whether any of the unmodified replica identity key attributes in the
3413 : * old tuple is externally stored or not. This is required because for
3414 : * such attributes the flattened value won't be WAL logged as part of the
3415 : * new tuple so we must include it as part of the old_key_tuple. See
3416 : * ExtractReplicaIdentity.
3417 : */
3418 609342 : modified_attrs = HeapDetermineColumnsInfo(relation, interesting_attrs,
3419 : id_attrs, &oldtup,
3420 : newtup, &id_has_external);
3421 :
3422 : /*
3423 : * If we're not updating any "key" column, we can grab a weaker lock type.
3424 : * This allows for more concurrency when we are running simultaneously
3425 : * with foreign key checks.
3426 : *
3427 : * Note that if a column gets detoasted while executing the update, but
3428 : * the value ends up being the same, this test will fail and we will use
3429 : * the stronger lock. This is acceptable; the important case to optimize
3430 : * is updates that don't manipulate key columns, not those that
3431 : * serendipitously arrive at the same key values.
3432 : */
3433 609342 : if (!bms_overlap(modified_attrs, key_attrs))
3434 : {
3435 601050 : *lockmode = LockTupleNoKeyExclusive;
3436 601050 : mxact_status = MultiXactStatusNoKeyUpdate;
3437 601050 : key_intact = true;
3438 :
3439 : /*
3440 : * If this is the first possibly-multixact-able operation in the
3441 : * current transaction, set my per-backend OldestMemberMXactId
3442 : * setting. We can be certain that the transaction will never become a
3443 : * member of any older MultiXactIds than that. (We have to do this
3444 : * even if we end up just using our own TransactionId below, since
3445 : * some other backend could incorporate our XID into a MultiXact
3446 : * immediately afterwards.)
3447 : */
3448 601050 : MultiXactIdSetOldestMember();
3449 : }
3450 : else
3451 : {
3452 8292 : *lockmode = LockTupleExclusive;
3453 8292 : mxact_status = MultiXactStatusUpdate;
3454 8292 : key_intact = false;
3455 : }
3456 :
3457 : /*
3458 : * Note: beyond this point, use oldtup not otid to refer to old tuple.
3459 : * otid may very well point at newtup->t_self, which we will overwrite
3460 : * with the new tuple's location, so there's great risk of confusion if we
3461 : * use otid anymore.
3462 : */
3463 :
3464 609342 : l2:
3465 609344 : checked_lockers = false;
3466 609344 : locker_remains = false;
3467 609344 : result = HeapTupleSatisfiesUpdate(&oldtup, cid, buffer);
3468 :
3469 : /* see below about the "no wait" case */
3470 : Assert(result != TM_BeingModified || wait);
3471 :
3472 609344 : if (result == TM_Invisible)
3473 : {
3474 0 : UnlockReleaseBuffer(buffer);
3475 0 : ereport(ERROR,
3476 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
3477 : errmsg("attempted to update invisible tuple")));
3478 : }
3479 609344 : else if (result == TM_BeingModified && wait)
3480 : {
3481 : TransactionId xwait;
3482 : uint16 infomask;
3483 71862 : bool can_continue = false;
3484 :
3485 : /*
3486 : * XXX note that we don't consider the "no wait" case here. This
3487 : * isn't a problem currently because no caller uses that case, but it
3488 : * should be fixed if such a caller is introduced. It wasn't a
3489 : * problem previously because this code would always wait, but now
3490 : * that some tuple locks do not conflict with one of the lock modes we
3491 : * use, it is possible that this case is interesting to handle
3492 : * specially.
3493 : *
3494 : * This may cause failures with third-party code that calls
3495 : * heap_update directly.
3496 : */
3497 :
3498 : /* must copy state data before unlocking buffer */
3499 71862 : xwait = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3500 71862 : infomask = oldtup.t_data->t_infomask;
3501 :
3502 : /*
3503 : * Now we have to do something about the existing locker. If it's a
3504 : * multi, sleep on it; we might be awakened before it is completely
3505 : * gone (or even not sleep at all in some cases); we need to preserve
3506 : * it as locker, unless it is gone completely.
3507 : *
3508 : * If it's not a multi, we need to check for sleeping conditions
3509 : * before actually going to sleep. If the update doesn't conflict
3510 : * with the locks, we just continue without sleeping (but making sure
3511 : * it is preserved).
3512 : *
3513 : * Before sleeping, we need to acquire tuple lock to establish our
3514 : * priority for the tuple (see heap_lock_tuple). LockTuple will
3515 : * release us when we are next-in-line for the tuple. Note we must
3516 : * not acquire the tuple lock until we're sure we're going to sleep;
3517 : * otherwise we're open for race conditions with other transactions
3518 : * holding the tuple lock which sleep on us.
3519 : *
3520 : * If we are forced to "start over" below, we keep the tuple lock;
3521 : * this arranges that we stay at the head of the line while rechecking
3522 : * tuple state.
3523 : */
3524 71862 : if (infomask & HEAP_XMAX_IS_MULTI)
3525 : {
3526 : TransactionId update_xact;
3527 : int remain;
3528 120 : bool current_is_member = false;
3529 :
3530 120 : if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
3531 : *lockmode, ¤t_is_member))
3532 : {
3533 16 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3534 :
3535 : /*
3536 : * Acquire the lock, if necessary (but skip it when we're
3537 : * requesting a lock and already have one; avoids deadlock).
3538 : */
3539 16 : if (!current_is_member)
3540 0 : heap_acquire_tuplock(relation, &(oldtup.t_self), *lockmode,
3541 : LockWaitBlock, &have_tuple_lock);
3542 :
3543 : /* wait for multixact */
3544 16 : MultiXactIdWait((MultiXactId) xwait, mxact_status, infomask,
3545 : relation, &oldtup.t_self, XLTW_Update,
3546 : &remain);
3547 16 : checked_lockers = true;
3548 16 : locker_remains = remain != 0;
3549 16 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3550 :
3551 : /*
3552 : * If xwait had just locked the tuple then some other xact
3553 : * could update this tuple before we get to this point. Check
3554 : * for xmax change, and start over if so.
3555 : */
3556 16 : if (xmax_infomask_changed(oldtup.t_data->t_infomask,
3557 16 : infomask) ||
3558 16 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3559 : xwait))
3560 0 : goto l2;
3561 : }
3562 :
3563 : /*
3564 : * Note that the multixact may not be done by now. It could have
3565 : * surviving members; our own xact or other subxacts of this
3566 : * backend, and also any other concurrent transaction that locked
3567 : * the tuple with LockTupleKeyShare if we only got
3568 : * LockTupleNoKeyExclusive. If this is the case, we have to be
3569 : * careful to mark the updated tuple with the surviving members in
3570 : * Xmax.
3571 : *
3572 : * Note that there could have been another update in the
3573 : * MultiXact. In that case, we need to check whether it committed
3574 : * or aborted. If it aborted we are safe to update it again;
3575 : * otherwise there is an update conflict, and we have to return
3576 : * TableTuple{Deleted, Updated} below.
3577 : *
3578 : * In the LockTupleExclusive case, we still need to preserve the
3579 : * surviving members: those would include the tuple locks we had
3580 : * before this one, which are important to keep in case this
3581 : * subxact aborts.
3582 : */
3583 120 : if (!HEAP_XMAX_IS_LOCKED_ONLY(oldtup.t_data->t_infomask))
3584 16 : update_xact = HeapTupleGetUpdateXid(oldtup.t_data);
3585 : else
3586 104 : update_xact = InvalidTransactionId;
3587 :
3588 : /*
3589 : * There was no UPDATE in the MultiXact; or it aborted. No
3590 : * TransactionIdIsInProgress() call needed here, since we called
3591 : * MultiXactIdWait() above.
3592 : */
3593 136 : if (!TransactionIdIsValid(update_xact) ||
3594 16 : TransactionIdDidAbort(update_xact))
3595 106 : can_continue = true;
3596 : }
3597 71742 : else if (TransactionIdIsCurrentTransactionId(xwait))
3598 : {
3599 : /*
3600 : * The only locker is ourselves; we can avoid grabbing the tuple
3601 : * lock here, but must preserve our locking information.
3602 : */
3603 71554 : checked_lockers = true;
3604 71554 : locker_remains = true;
3605 71554 : can_continue = true;
3606 : }
3607 188 : else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) && key_intact)
3608 : {
3609 : /*
3610 : * If it's just a key-share locker, and we're not changing the key
3611 : * columns, we don't need to wait for it to end; but we need to
3612 : * preserve it as locker.
3613 : */
3614 58 : checked_lockers = true;
3615 58 : locker_remains = true;
3616 58 : can_continue = true;
3617 : }
3618 : else
3619 : {
3620 : /*
3621 : * Wait for regular transaction to end; but first, acquire tuple
3622 : * lock.
3623 : */
3624 130 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3625 130 : heap_acquire_tuplock(relation, &(oldtup.t_self), *lockmode,
3626 : LockWaitBlock, &have_tuple_lock);
3627 130 : XactLockTableWait(xwait, relation, &oldtup.t_self,
3628 : XLTW_Update);
3629 130 : checked_lockers = true;
3630 130 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3631 :
3632 : /*
3633 : * xwait is done, but if xwait had just locked the tuple then some
3634 : * other xact could update this tuple before we get to this point.
3635 : * Check for xmax change, and start over if so.
3636 : */
3637 258 : if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3638 128 : !TransactionIdEquals(xwait,
3639 : HeapTupleHeaderGetRawXmax(oldtup.t_data)))
3640 2 : goto l2;
3641 :
3642 : /* Otherwise check if it committed or aborted */
3643 128 : UpdateXmaxHintBits(oldtup.t_data, buffer, xwait);
3644 128 : if (oldtup.t_data->t_infomask & HEAP_XMAX_INVALID)
3645 30 : can_continue = true;
3646 : }
3647 :
3648 71860 : if (can_continue)
3649 71748 : result = TM_Ok;
3650 112 : else if (!ItemPointerEquals(&oldtup.t_self, &oldtup.t_data->t_ctid))
3651 102 : result = TM_Updated;
3652 : else
3653 10 : result = TM_Deleted;
3654 : }
3655 :
3656 : /* Sanity check the result HeapTupleSatisfiesUpdate() and the logic above */
3657 : if (result != TM_Ok)
3658 : {
3659 : Assert(result == TM_SelfModified ||
3660 : result == TM_Updated ||
3661 : result == TM_Deleted ||
3662 : result == TM_BeingModified);
3663 : Assert(!(oldtup.t_data->t_infomask & HEAP_XMAX_INVALID));
3664 : Assert(result != TM_Updated ||
3665 : !ItemPointerEquals(&oldtup.t_self, &oldtup.t_data->t_ctid));
3666 : }
3667 :
3668 609342 : if (crosscheck != InvalidSnapshot && result == TM_Ok)
3669 : {
3670 : /* Perform additional check for transaction-snapshot mode RI updates */
3671 2 : if (!HeapTupleSatisfiesVisibility(&oldtup, crosscheck, buffer))
3672 2 : result = TM_Updated;
3673 : }
3674 :
3675 609342 : if (result != TM_Ok)
3676 : {
3677 306 : tmfd->ctid = oldtup.t_data->t_ctid;
3678 306 : tmfd->xmax = HeapTupleHeaderGetUpdateXid(oldtup.t_data);
3679 306 : if (result == TM_SelfModified)
3680 104 : tmfd->cmax = HeapTupleHeaderGetCmax(oldtup.t_data);
3681 : else
3682 202 : tmfd->cmax = InvalidCommandId;
3683 306 : UnlockReleaseBuffer(buffer);
3684 306 : if (have_tuple_lock)
3685 98 : UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3686 306 : if (vmbuffer != InvalidBuffer)
3687 0 : ReleaseBuffer(vmbuffer);
3688 306 : *update_indexes = TU_None;
3689 :
3690 306 : bms_free(hot_attrs);
3691 306 : bms_free(sum_attrs);
3692 306 : bms_free(key_attrs);
3693 306 : bms_free(id_attrs);
3694 306 : bms_free(modified_attrs);
3695 306 : bms_free(interesting_attrs);
3696 306 : return result;
3697 : }
3698 :
3699 : /*
3700 : * If we didn't pin the visibility map page and the page has become all
3701 : * visible while we were busy locking the buffer, or during some
3702 : * subsequent window during which we had it unlocked, we'll have to unlock
3703 : * and re-lock, to avoid holding the buffer lock across an I/O. That's a
3704 : * bit unfortunate, especially since we'll now have to recheck whether the
3705 : * tuple has been locked or updated under us, but hopefully it won't
3706 : * happen very often.
3707 : */
3708 609036 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3709 : {
3710 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3711 0 : visibilitymap_pin(relation, block, &vmbuffer);
3712 0 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3713 0 : goto l2;
3714 : }
3715 :
3716 : /* Fill in transaction status data */
3717 :
3718 : /*
3719 : * If the tuple we're updating is locked, we need to preserve the locking
3720 : * info in the old tuple's Xmax. Prepare a new Xmax value for this.
3721 : */
3722 609036 : compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3723 609036 : oldtup.t_data->t_infomask,
3724 609036 : oldtup.t_data->t_infomask2,
3725 : xid, *lockmode, true,
3726 : &xmax_old_tuple, &infomask_old_tuple,
3727 : &infomask2_old_tuple);
3728 :
3729 : /*
3730 : * And also prepare an Xmax value for the new copy of the tuple. If there
3731 : * was no xmax previously, or there was one but all lockers are now gone,
3732 : * then use InvalidTransactionId; otherwise, get the xmax from the old
3733 : * tuple. (In rare cases that might also be InvalidTransactionId and yet
3734 : * not have the HEAP_XMAX_INVALID bit set; that's fine.)
3735 : */
3736 680754 : if ((oldtup.t_data->t_infomask & HEAP_XMAX_INVALID) ||
3737 143436 : HEAP_LOCKED_UPGRADED(oldtup.t_data->t_infomask) ||
3738 71614 : (checked_lockers && !locker_remains))
3739 537318 : xmax_new_tuple = InvalidTransactionId;
3740 : else
3741 71718 : xmax_new_tuple = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3742 :
3743 609036 : if (!TransactionIdIsValid(xmax_new_tuple))
3744 : {
3745 537318 : infomask_new_tuple = HEAP_XMAX_INVALID;
3746 537318 : infomask2_new_tuple = 0;
3747 : }
3748 : else
3749 : {
3750 : /*
3751 : * If we found a valid Xmax for the new tuple, then the infomask bits
3752 : * to use on the new tuple depend on what was there on the old one.
3753 : * Note that since we're doing an update, the only possibility is that
3754 : * the lockers had FOR KEY SHARE lock.
3755 : */
3756 71718 : if (oldtup.t_data->t_infomask & HEAP_XMAX_IS_MULTI)
3757 : {
3758 106 : GetMultiXactIdHintBits(xmax_new_tuple, &infomask_new_tuple,
3759 : &infomask2_new_tuple);
3760 : }
3761 : else
3762 : {
3763 71612 : infomask_new_tuple = HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_LOCK_ONLY;
3764 71612 : infomask2_new_tuple = 0;
3765 : }
3766 : }
3767 :
3768 : /*
3769 : * Prepare the new tuple with the appropriate initial values of Xmin and
3770 : * Xmax, as well as initial infomask bits as computed above.
3771 : */
3772 609036 : newtup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
3773 609036 : newtup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
3774 609036 : HeapTupleHeaderSetXmin(newtup->t_data, xid);
3775 609036 : HeapTupleHeaderSetCmin(newtup->t_data, cid);
3776 609036 : newtup->t_data->t_infomask |= HEAP_UPDATED | infomask_new_tuple;
3777 609036 : newtup->t_data->t_infomask2 |= infomask2_new_tuple;
3778 609036 : HeapTupleHeaderSetXmax(newtup->t_data, xmax_new_tuple);
3779 :
3780 : /*
3781 : * Replace cid with a combo CID if necessary. Note that we already put
3782 : * the plain cid into the new tuple.
3783 : */
3784 609036 : HeapTupleHeaderAdjustCmax(oldtup.t_data, &cid, &iscombo);
3785 :
3786 : /*
3787 : * If the toaster needs to be activated, OR if the new tuple will not fit
3788 : * on the same page as the old, then we need to release the content lock
3789 : * (but not the pin!) on the old tuple's buffer while we are off doing
3790 : * TOAST and/or table-file-extension work. We must mark the old tuple to
3791 : * show that it's locked, else other processes may try to update it
3792 : * themselves.
3793 : *
3794 : * We need to invoke the toaster if there are already any out-of-line
3795 : * toasted values present, or if the new tuple is over-threshold.
3796 : */
3797 609036 : if (relation->rd_rel->relkind != RELKIND_RELATION &&
3798 0 : relation->rd_rel->relkind != RELKIND_MATVIEW)
3799 : {
3800 : /* toast table entries should never be recursively toasted */
3801 : Assert(!HeapTupleHasExternal(&oldtup));
3802 : Assert(!HeapTupleHasExternal(newtup));
3803 0 : need_toast = false;
3804 : }
3805 : else
3806 1826392 : need_toast = (HeapTupleHasExternal(&oldtup) ||
3807 1217356 : HeapTupleHasExternal(newtup) ||
3808 608272 : newtup->t_len > TOAST_TUPLE_THRESHOLD);
3809 :
3810 609036 : pagefree = PageGetHeapFreeSpace(page);
3811 :
3812 609036 : newtupsize = MAXALIGN(newtup->t_len);
3813 :
3814 609036 : if (need_toast || newtupsize > pagefree)
3815 300312 : {
3816 : TransactionId xmax_lock_old_tuple;
3817 : uint16 infomask_lock_old_tuple,
3818 : infomask2_lock_old_tuple;
3819 300312 : bool cleared_all_frozen = false;
3820 :
3821 : /*
3822 : * To prevent concurrent sessions from updating the tuple, we have to
3823 : * temporarily mark it locked, while we release the page-level lock.
3824 : *
3825 : * To satisfy the rule that any xid potentially appearing in a buffer
3826 : * written out to disk, we unfortunately have to WAL log this
3827 : * temporary modification. We can reuse xl_heap_lock for this
3828 : * purpose. If we crash/error before following through with the
3829 : * actual update, xmax will be of an aborted transaction, allowing
3830 : * other sessions to proceed.
3831 : */
3832 :
3833 : /*
3834 : * Compute xmax / infomask appropriate for locking the tuple. This has
3835 : * to be done separately from the combo that's going to be used for
3836 : * updating, because the potentially created multixact would otherwise
3837 : * be wrong.
3838 : */
3839 300312 : compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3840 300312 : oldtup.t_data->t_infomask,
3841 300312 : oldtup.t_data->t_infomask2,
3842 : xid, *lockmode, false,
3843 : &xmax_lock_old_tuple, &infomask_lock_old_tuple,
3844 : &infomask2_lock_old_tuple);
3845 :
3846 : Assert(HEAP_XMAX_IS_LOCKED_ONLY(infomask_lock_old_tuple));
3847 :
3848 300312 : START_CRIT_SECTION();
3849 :
3850 : /* Clear obsolete visibility flags ... */
3851 300312 : oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3852 300312 : oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3853 300312 : HeapTupleClearHotUpdated(&oldtup);
3854 : /* ... and store info about transaction updating this tuple */
3855 : Assert(TransactionIdIsValid(xmax_lock_old_tuple));
3856 300312 : HeapTupleHeaderSetXmax(oldtup.t_data, xmax_lock_old_tuple);
3857 300312 : oldtup.t_data->t_infomask |= infomask_lock_old_tuple;
3858 300312 : oldtup.t_data->t_infomask2 |= infomask2_lock_old_tuple;
3859 300312 : HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
3860 :
3861 : /* temporarily make it look not-updated, but locked */
3862 300312 : oldtup.t_data->t_ctid = oldtup.t_self;
3863 :
3864 : /*
3865 : * Clear all-frozen bit on visibility map if needed. We could
3866 : * immediately reset ALL_VISIBLE, but given that the WAL logging
3867 : * overhead would be unchanged, that doesn't seem necessarily
3868 : * worthwhile.
3869 : */
3870 302218 : if (PageIsAllVisible(page) &&
3871 1906 : visibilitymap_clear(relation, block, vmbuffer,
3872 : VISIBILITYMAP_ALL_FROZEN))
3873 1564 : cleared_all_frozen = true;
3874 :
3875 300312 : MarkBufferDirty(buffer);
3876 :
3877 300312 : if (RelationNeedsWAL(relation))
3878 : {
3879 : xl_heap_lock xlrec;
3880 : XLogRecPtr recptr;
3881 :
3882 280050 : XLogBeginInsert();
3883 280050 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
3884 :
3885 280050 : xlrec.offnum = ItemPointerGetOffsetNumber(&oldtup.t_self);
3886 280050 : xlrec.xmax = xmax_lock_old_tuple;
3887 560100 : xlrec.infobits_set = compute_infobits(oldtup.t_data->t_infomask,
3888 280050 : oldtup.t_data->t_infomask2);
3889 280050 : xlrec.flags =
3890 280050 : cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
3891 280050 : XLogRegisterData(&xlrec, SizeOfHeapLock);
3892 280050 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK);
3893 280050 : PageSetLSN(page, recptr);
3894 : }
3895 :
3896 300312 : END_CRIT_SECTION();
3897 :
3898 300312 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3899 :
3900 : /*
3901 : * Let the toaster do its thing, if needed.
3902 : *
3903 : * Note: below this point, heaptup is the data we actually intend to
3904 : * store into the relation; newtup is the caller's original untoasted
3905 : * data.
3906 : */
3907 300312 : if (need_toast)
3908 : {
3909 : /* Note we always use WAL and FSM during updates */
3910 3214 : heaptup = heap_toast_insert_or_update(relation, newtup, &oldtup, 0);
3911 3214 : newtupsize = MAXALIGN(heaptup->t_len);
3912 : }
3913 : else
3914 297098 : heaptup = newtup;
3915 :
3916 : /*
3917 : * Now, do we need a new page for the tuple, or not? This is a bit
3918 : * tricky since someone else could have added tuples to the page while
3919 : * we weren't looking. We have to recheck the available space after
3920 : * reacquiring the buffer lock. But don't bother to do that if the
3921 : * former amount of free space is still not enough; it's unlikely
3922 : * there's more free now than before.
3923 : *
3924 : * What's more, if we need to get a new page, we will need to acquire
3925 : * buffer locks on both old and new pages. To avoid deadlock against
3926 : * some other backend trying to get the same two locks in the other
3927 : * order, we must be consistent about the order we get the locks in.
3928 : * We use the rule "lock the lower-numbered page of the relation
3929 : * first". To implement this, we must do RelationGetBufferForTuple
3930 : * while not holding the lock on the old page, and we must rely on it
3931 : * to get the locks on both pages in the correct order.
3932 : *
3933 : * Another consideration is that we need visibility map page pin(s) if
3934 : * we will have to clear the all-visible flag on either page. If we
3935 : * call RelationGetBufferForTuple, we rely on it to acquire any such
3936 : * pins; but if we don't, we have to handle that here. Hence we need
3937 : * a loop.
3938 : */
3939 : for (;;)
3940 : {
3941 300312 : if (newtupsize > pagefree)
3942 : {
3943 : /* It doesn't fit, must use RelationGetBufferForTuple. */
3944 299300 : newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3945 : buffer, 0, NULL,
3946 : &vmbuffer_new, &vmbuffer,
3947 : 0);
3948 : /* We're all done. */
3949 299300 : break;
3950 : }
3951 : /* Acquire VM page pin if needed and we don't have it. */
3952 1012 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3953 0 : visibilitymap_pin(relation, block, &vmbuffer);
3954 : /* Re-acquire the lock on the old tuple's page. */
3955 1012 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3956 : /* Re-check using the up-to-date free space */
3957 1012 : pagefree = PageGetHeapFreeSpace(page);
3958 1012 : if (newtupsize > pagefree ||
3959 1012 : (vmbuffer == InvalidBuffer && PageIsAllVisible(page)))
3960 : {
3961 : /*
3962 : * Rats, it doesn't fit anymore, or somebody just now set the
3963 : * all-visible flag. We must now unlock and loop to avoid
3964 : * deadlock. Fortunately, this path should seldom be taken.
3965 : */
3966 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3967 : }
3968 : else
3969 : {
3970 : /* We're all done. */
3971 1012 : newbuf = buffer;
3972 1012 : break;
3973 : }
3974 : }
3975 : }
3976 : else
3977 : {
3978 : /* No TOAST work needed, and it'll fit on same page */
3979 308724 : newbuf = buffer;
3980 308724 : heaptup = newtup;
3981 : }
3982 :
3983 : /*
3984 : * We're about to do the actual update -- check for conflict first, to
3985 : * avoid possibly having to roll back work we've just done.
3986 : *
3987 : * This is safe without a recheck as long as there is no possibility of
3988 : * another process scanning the pages between this check and the update
3989 : * being visible to the scan (i.e., exclusive buffer content lock(s) are
3990 : * continuously held from this point until the tuple update is visible).
3991 : *
3992 : * For the new tuple the only check needed is at the relation level, but
3993 : * since both tuples are in the same relation and the check for oldtup
3994 : * will include checking the relation level, there is no benefit to a
3995 : * separate check for the new tuple.
3996 : */
3997 609036 : CheckForSerializableConflictIn(relation, &oldtup.t_self,
3998 : BufferGetBlockNumber(buffer));
3999 :
4000 : /*
4001 : * At this point newbuf and buffer are both pinned and locked, and newbuf
4002 : * has enough space for the new tuple. If they are the same buffer, only
4003 : * one pin is held.
4004 : */
4005 :
4006 609012 : if (newbuf == buffer)
4007 : {
4008 : /*
4009 : * Since the new tuple is going into the same page, we might be able
4010 : * to do a HOT update. Check if any of the index columns have been
4011 : * changed.
4012 : */
4013 309712 : if (!bms_overlap(modified_attrs, hot_attrs))
4014 : {
4015 283750 : use_hot_update = true;
4016 :
4017 : /*
4018 : * If none of the columns that are used in hot-blocking indexes
4019 : * were updated, we can apply HOT, but we do still need to check
4020 : * if we need to update the summarizing indexes, and update those
4021 : * indexes if the columns were updated, or we may fail to detect
4022 : * e.g. value bound changes in BRIN minmax indexes.
4023 : */
4024 283750 : if (bms_overlap(modified_attrs, sum_attrs))
4025 3282 : summarized_update = true;
4026 : }
4027 : }
4028 : else
4029 : {
4030 : /* Set a hint that the old page could use prune/defrag */
4031 299300 : PageSetFull(page);
4032 : }
4033 :
4034 : /*
4035 : * Compute replica identity tuple before entering the critical section so
4036 : * we don't PANIC upon a memory allocation failure.
4037 : * ExtractReplicaIdentity() will return NULL if nothing needs to be
4038 : * logged. Pass old key required as true only if the replica identity key
4039 : * columns are modified or it has external data.
4040 : */
4041 609012 : old_key_tuple = ExtractReplicaIdentity(relation, &oldtup,
4042 609012 : bms_overlap(modified_attrs, id_attrs) ||
4043 : id_has_external,
4044 : &old_key_copied);
4045 :
4046 : /* NO EREPORT(ERROR) from here till changes are logged */
4047 609012 : START_CRIT_SECTION();
4048 :
4049 : /*
4050 : * If this transaction commits, the old tuple will become DEAD sooner or
4051 : * later. Set flag that this page is a candidate for pruning once our xid
4052 : * falls below the OldestXmin horizon. If the transaction finally aborts,
4053 : * the subsequent page pruning will be a no-op and the hint will be
4054 : * cleared.
4055 : *
4056 : * XXX Should we set hint on newbuf as well? If the transaction aborts,
4057 : * there would be a prunable tuple in the newbuf; but for now we choose
4058 : * not to optimize for aborts. Note that heap_xlog_update must be kept in
4059 : * sync if this decision changes.
4060 : */
4061 609012 : PageSetPrunable(page, xid);
4062 :
4063 609012 : if (use_hot_update)
4064 : {
4065 : /* Mark the old tuple as HOT-updated */
4066 283750 : HeapTupleSetHotUpdated(&oldtup);
4067 : /* And mark the new tuple as heap-only */
4068 283750 : HeapTupleSetHeapOnly(heaptup);
4069 : /* Mark the caller's copy too, in case different from heaptup */
4070 283750 : HeapTupleSetHeapOnly(newtup);
4071 : }
4072 : else
4073 : {
4074 : /* Make sure tuples are correctly marked as not-HOT */
4075 325262 : HeapTupleClearHotUpdated(&oldtup);
4076 325262 : HeapTupleClearHeapOnly(heaptup);
4077 325262 : HeapTupleClearHeapOnly(newtup);
4078 : }
4079 :
4080 609012 : RelationPutHeapTuple(relation, newbuf, heaptup, false); /* insert new tuple */
4081 :
4082 :
4083 : /* Clear obsolete visibility flags, possibly set by ourselves above... */
4084 609012 : oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
4085 609012 : oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
4086 : /* ... and store info about transaction updating this tuple */
4087 : Assert(TransactionIdIsValid(xmax_old_tuple));
4088 609012 : HeapTupleHeaderSetXmax(oldtup.t_data, xmax_old_tuple);
4089 609012 : oldtup.t_data->t_infomask |= infomask_old_tuple;
4090 609012 : oldtup.t_data->t_infomask2 |= infomask2_old_tuple;
4091 609012 : HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
4092 :
4093 : /* record address of new tuple in t_ctid of old one */
4094 609012 : oldtup.t_data->t_ctid = heaptup->t_self;
4095 :
4096 : /* clear PD_ALL_VISIBLE flags, reset all visibilitymap bits */
4097 609012 : if (PageIsAllVisible(BufferGetPage(buffer)))
4098 : {
4099 2992 : all_visible_cleared = true;
4100 2992 : PageClearAllVisible(BufferGetPage(buffer));
4101 2992 : visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
4102 : vmbuffer, VISIBILITYMAP_VALID_BITS);
4103 : }
4104 609012 : if (newbuf != buffer && PageIsAllVisible(BufferGetPage(newbuf)))
4105 : {
4106 1634 : all_visible_cleared_new = true;
4107 1634 : PageClearAllVisible(BufferGetPage(newbuf));
4108 1634 : visibilitymap_clear(relation, BufferGetBlockNumber(newbuf),
4109 : vmbuffer_new, VISIBILITYMAP_VALID_BITS);
4110 : }
4111 :
4112 609012 : if (newbuf != buffer)
4113 299300 : MarkBufferDirty(newbuf);
4114 609012 : MarkBufferDirty(buffer);
4115 :
4116 : /* XLOG stuff */
4117 609012 : if (RelationNeedsWAL(relation))
4118 : {
4119 : XLogRecPtr recptr;
4120 :
4121 : /*
4122 : * For logical decoding we need combo CIDs to properly decode the
4123 : * catalog.
4124 : */
4125 586288 : if (RelationIsAccessibleInLogicalDecoding(relation))
4126 : {
4127 5002 : log_heap_new_cid(relation, &oldtup);
4128 5002 : log_heap_new_cid(relation, heaptup);
4129 : }
4130 :
4131 586288 : recptr = log_heap_update(relation, buffer,
4132 : newbuf, &oldtup, heaptup,
4133 : old_key_tuple,
4134 : all_visible_cleared,
4135 : all_visible_cleared_new);
4136 586288 : if (newbuf != buffer)
4137 : {
4138 279050 : PageSetLSN(BufferGetPage(newbuf), recptr);
4139 : }
4140 586288 : PageSetLSN(BufferGetPage(buffer), recptr);
4141 : }
4142 :
4143 609012 : END_CRIT_SECTION();
4144 :
4145 609012 : if (newbuf != buffer)
4146 299300 : LockBuffer(newbuf, BUFFER_LOCK_UNLOCK);
4147 609012 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
4148 :
4149 : /*
4150 : * Mark old tuple for invalidation from system caches at next command
4151 : * boundary, and mark the new tuple for invalidation in case we abort. We
4152 : * have to do this before releasing the buffer because oldtup is in the
4153 : * buffer. (heaptup is all in local memory, but it's necessary to process
4154 : * both tuple versions in one call to inval.c so we can avoid redundant
4155 : * sinval messages.)
4156 : */
4157 609012 : CacheInvalidateHeapTuple(relation, &oldtup, heaptup);
4158 :
4159 : /* Now we can release the buffer(s) */
4160 609012 : if (newbuf != buffer)
4161 299300 : ReleaseBuffer(newbuf);
4162 609012 : ReleaseBuffer(buffer);
4163 609012 : if (BufferIsValid(vmbuffer_new))
4164 1634 : ReleaseBuffer(vmbuffer_new);
4165 609012 : if (BufferIsValid(vmbuffer))
4166 2992 : ReleaseBuffer(vmbuffer);
4167 :
4168 : /*
4169 : * Release the lmgr tuple lock, if we had it.
4170 : */
4171 609012 : if (have_tuple_lock)
4172 30 : UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
4173 :
4174 609012 : pgstat_count_heap_update(relation, use_hot_update, newbuf != buffer);
4175 :
4176 : /*
4177 : * If heaptup is a private copy, release it. Don't forget to copy t_self
4178 : * back to the caller's image, too.
4179 : */
4180 609012 : if (heaptup != newtup)
4181 : {
4182 3120 : newtup->t_self = heaptup->t_self;
4183 3120 : heap_freetuple(heaptup);
4184 : }
4185 :
4186 : /*
4187 : * If it is a HOT update, the update may still need to update summarized
4188 : * indexes, lest we fail to update those summaries and get incorrect
4189 : * results (for example, minmax bounds of the block may change with this
4190 : * update).
4191 : */
4192 609012 : if (use_hot_update)
4193 : {
4194 283750 : if (summarized_update)
4195 3282 : *update_indexes = TU_Summarizing;
4196 : else
4197 280468 : *update_indexes = TU_None;
4198 : }
4199 : else
4200 325262 : *update_indexes = TU_All;
4201 :
4202 609012 : if (old_key_tuple != NULL && old_key_copied)
4203 166 : heap_freetuple(old_key_tuple);
4204 :
4205 609012 : bms_free(hot_attrs);
4206 609012 : bms_free(sum_attrs);
4207 609012 : bms_free(key_attrs);
4208 609012 : bms_free(id_attrs);
4209 609012 : bms_free(modified_attrs);
4210 609012 : bms_free(interesting_attrs);
4211 :
4212 609012 : return TM_Ok;
4213 : }
4214 :
4215 : #ifdef USE_ASSERT_CHECKING
4216 : /*
4217 : * Confirm adequate lock held during heap_update(), per rules from
4218 : * README.tuplock section "Locking to write inplace-updated tables".
4219 : */
4220 : static void
4221 : check_lock_if_inplace_updateable_rel(Relation relation,
4222 : ItemPointer otid,
4223 : HeapTuple newtup)
4224 : {
4225 : /* LOCKTAG_TUPLE acceptable for any catalog */
4226 : switch (RelationGetRelid(relation))
4227 : {
4228 : case RelationRelationId:
4229 : case DatabaseRelationId:
4230 : {
4231 : LOCKTAG tuptag;
4232 :
4233 : SET_LOCKTAG_TUPLE(tuptag,
4234 : relation->rd_lockInfo.lockRelId.dbId,
4235 : relation->rd_lockInfo.lockRelId.relId,
4236 : ItemPointerGetBlockNumber(otid),
4237 : ItemPointerGetOffsetNumber(otid));
4238 : if (LockHeldByMe(&tuptag, InplaceUpdateTupleLock, false))
4239 : return;
4240 : }
4241 : break;
4242 : default:
4243 : Assert(!IsInplaceUpdateRelation(relation));
4244 : return;
4245 : }
4246 :
4247 : switch (RelationGetRelid(relation))
4248 : {
4249 : case RelationRelationId:
4250 : {
4251 : /* LOCKTAG_TUPLE or LOCKTAG_RELATION ok */
4252 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(newtup);
4253 : Oid relid = classForm->oid;
4254 : Oid dbid;
4255 : LOCKTAG tag;
4256 :
4257 : if (IsSharedRelation(relid))
4258 : dbid = InvalidOid;
4259 : else
4260 : dbid = MyDatabaseId;
4261 :
4262 : if (classForm->relkind == RELKIND_INDEX)
4263 : {
4264 : Relation irel = index_open(relid, AccessShareLock);
4265 :
4266 : SET_LOCKTAG_RELATION(tag, dbid, irel->rd_index->indrelid);
4267 : index_close(irel, AccessShareLock);
4268 : }
4269 : else
4270 : SET_LOCKTAG_RELATION(tag, dbid, relid);
4271 :
4272 : if (!LockHeldByMe(&tag, ShareUpdateExclusiveLock, false) &&
4273 : !LockHeldByMe(&tag, ShareRowExclusiveLock, true))
4274 : elog(WARNING,
4275 : "missing lock for relation \"%s\" (OID %u, relkind %c) @ TID (%u,%u)",
4276 : NameStr(classForm->relname),
4277 : relid,
4278 : classForm->relkind,
4279 : ItemPointerGetBlockNumber(otid),
4280 : ItemPointerGetOffsetNumber(otid));
4281 : }
4282 : break;
4283 : case DatabaseRelationId:
4284 : {
4285 : /* LOCKTAG_TUPLE required */
4286 : Form_pg_database dbForm = (Form_pg_database) GETSTRUCT(newtup);
4287 :
4288 : elog(WARNING,
4289 : "missing lock on database \"%s\" (OID %u) @ TID (%u,%u)",
4290 : NameStr(dbForm->datname),
4291 : dbForm->oid,
4292 : ItemPointerGetBlockNumber(otid),
4293 : ItemPointerGetOffsetNumber(otid));
4294 : }
4295 : break;
4296 : }
4297 : }
4298 :
4299 : /*
4300 : * Confirm adequate relation lock held, per rules from README.tuplock section
4301 : * "Locking to write inplace-updated tables".
4302 : */
4303 : static void
4304 : check_inplace_rel_lock(HeapTuple oldtup)
4305 : {
4306 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(oldtup);
4307 : Oid relid = classForm->oid;
4308 : Oid dbid;
4309 : LOCKTAG tag;
4310 :
4311 : if (IsSharedRelation(relid))
4312 : dbid = InvalidOid;
4313 : else
4314 : dbid = MyDatabaseId;
4315 :
4316 : if (classForm->relkind == RELKIND_INDEX)
4317 : {
4318 : Relation irel = index_open(relid, AccessShareLock);
4319 :
4320 : SET_LOCKTAG_RELATION(tag, dbid, irel->rd_index->indrelid);
4321 : index_close(irel, AccessShareLock);
4322 : }
4323 : else
4324 : SET_LOCKTAG_RELATION(tag, dbid, relid);
4325 :
4326 : if (!LockHeldByMe(&tag, ShareUpdateExclusiveLock, true))
4327 : elog(WARNING,
4328 : "missing lock for relation \"%s\" (OID %u, relkind %c) @ TID (%u,%u)",
4329 : NameStr(classForm->relname),
4330 : relid,
4331 : classForm->relkind,
4332 : ItemPointerGetBlockNumber(&oldtup->t_self),
4333 : ItemPointerGetOffsetNumber(&oldtup->t_self));
4334 : }
4335 : #endif
4336 :
4337 : /*
4338 : * Check if the specified attribute's values are the same. Subroutine for
4339 : * HeapDetermineColumnsInfo.
4340 : */
4341 : static bool
4342 1476330 : heap_attr_equals(TupleDesc tupdesc, int attrnum, Datum value1, Datum value2,
4343 : bool isnull1, bool isnull2)
4344 : {
4345 : /*
4346 : * If one value is NULL and other is not, then they are certainly not
4347 : * equal
4348 : */
4349 1476330 : if (isnull1 != isnull2)
4350 90 : return false;
4351 :
4352 : /*
4353 : * If both are NULL, they can be considered equal.
4354 : */
4355 1476240 : if (isnull1)
4356 9982 : return true;
4357 :
4358 : /*
4359 : * We do simple binary comparison of the two datums. This may be overly
4360 : * strict because there can be multiple binary representations for the
4361 : * same logical value. But we should be OK as long as there are no false
4362 : * positives. Using a type-specific equality operator is messy because
4363 : * there could be multiple notions of equality in different operator
4364 : * classes; furthermore, we cannot safely invoke user-defined functions
4365 : * while holding exclusive buffer lock.
4366 : */
4367 1466258 : if (attrnum <= 0)
4368 : {
4369 : /* The only allowed system columns are OIDs, so do this */
4370 0 : return (DatumGetObjectId(value1) == DatumGetObjectId(value2));
4371 : }
4372 : else
4373 : {
4374 : CompactAttribute *att;
4375 :
4376 : Assert(attrnum <= tupdesc->natts);
4377 1466258 : att = TupleDescCompactAttr(tupdesc, attrnum - 1);
4378 1466258 : return datumIsEqual(value1, value2, att->attbyval, att->attlen);
4379 : }
4380 : }
4381 :
4382 : /*
4383 : * Check which columns are being updated.
4384 : *
4385 : * Given an updated tuple, determine (and return into the output bitmapset),
4386 : * from those listed as interesting, the set of columns that changed.
4387 : *
4388 : * has_external indicates if any of the unmodified attributes (from those
4389 : * listed as interesting) of the old tuple is a member of external_cols and is
4390 : * stored externally.
4391 : */
4392 : static Bitmapset *
4393 609342 : HeapDetermineColumnsInfo(Relation relation,
4394 : Bitmapset *interesting_cols,
4395 : Bitmapset *external_cols,
4396 : HeapTuple oldtup, HeapTuple newtup,
4397 : bool *has_external)
4398 : {
4399 : int attidx;
4400 609342 : Bitmapset *modified = NULL;
4401 609342 : TupleDesc tupdesc = RelationGetDescr(relation);
4402 :
4403 609342 : attidx = -1;
4404 2085672 : while ((attidx = bms_next_member(interesting_cols, attidx)) >= 0)
4405 : {
4406 : /* attidx is zero-based, attrnum is the normal attribute number */
4407 1476330 : AttrNumber attrnum = attidx + FirstLowInvalidHeapAttributeNumber;
4408 : Datum value1,
4409 : value2;
4410 : bool isnull1,
4411 : isnull2;
4412 :
4413 : /*
4414 : * If it's a whole-tuple reference, say "not equal". It's not really
4415 : * worth supporting this case, since it could only succeed after a
4416 : * no-op update, which is hardly a case worth optimizing for.
4417 : */
4418 1476330 : if (attrnum == 0)
4419 : {
4420 0 : modified = bms_add_member(modified, attidx);
4421 1413388 : continue;
4422 : }
4423 :
4424 : /*
4425 : * Likewise, automatically say "not equal" for any system attribute
4426 : * other than tableOID; we cannot expect these to be consistent in a
4427 : * HOT chain, or even to be set correctly yet in the new tuple.
4428 : */
4429 1476330 : if (attrnum < 0)
4430 : {
4431 0 : if (attrnum != TableOidAttributeNumber)
4432 : {
4433 0 : modified = bms_add_member(modified, attidx);
4434 0 : continue;
4435 : }
4436 : }
4437 :
4438 : /*
4439 : * Extract the corresponding values. XXX this is pretty inefficient
4440 : * if there are many indexed columns. Should we do a single
4441 : * heap_deform_tuple call on each tuple, instead? But that doesn't
4442 : * work for system columns ...
4443 : */
4444 1476330 : value1 = heap_getattr(oldtup, attrnum, tupdesc, &isnull1);
4445 1476330 : value2 = heap_getattr(newtup, attrnum, tupdesc, &isnull2);
4446 :
4447 1476330 : if (!heap_attr_equals(tupdesc, attrnum, value1,
4448 : value2, isnull1, isnull2))
4449 : {
4450 55776 : modified = bms_add_member(modified, attidx);
4451 55776 : continue;
4452 : }
4453 :
4454 : /*
4455 : * No need to check attributes that can't be stored externally. Note
4456 : * that system attributes can't be stored externally.
4457 : */
4458 1420554 : if (attrnum < 0 || isnull1 ||
4459 1410572 : TupleDescCompactAttr(tupdesc, attrnum - 1)->attlen != -1)
4460 1357612 : continue;
4461 :
4462 : /*
4463 : * Check if the old tuple's attribute is stored externally and is a
4464 : * member of external_cols.
4465 : */
4466 62952 : if (VARATT_IS_EXTERNAL((struct varlena *) DatumGetPointer(value1)) &&
4467 10 : bms_is_member(attidx, external_cols))
4468 4 : *has_external = true;
4469 : }
4470 :
4471 609342 : return modified;
4472 : }
4473 :
4474 : /*
4475 : * simple_heap_update - replace a tuple
4476 : *
4477 : * This routine may be used to update a tuple when concurrent updates of
4478 : * the target tuple are not expected (for example, because we have a lock
4479 : * on the relation associated with the tuple). Any failure is reported
4480 : * via ereport().
4481 : */
4482 : void
4483 222760 : simple_heap_update(Relation relation, ItemPointer otid, HeapTuple tup,
4484 : TU_UpdateIndexes *update_indexes)
4485 : {
4486 : TM_Result result;
4487 : TM_FailureData tmfd;
4488 : LockTupleMode lockmode;
4489 :
4490 222760 : result = heap_update(relation, otid, tup,
4491 : GetCurrentCommandId(true), InvalidSnapshot,
4492 : true /* wait for commit */ ,
4493 : &tmfd, &lockmode, update_indexes);
4494 222760 : switch (result)
4495 : {
4496 0 : case TM_SelfModified:
4497 : /* Tuple was already updated in current command? */
4498 0 : elog(ERROR, "tuple already updated by self");
4499 : break;
4500 :
4501 222758 : case TM_Ok:
4502 : /* done successfully */
4503 222758 : break;
4504 :
4505 0 : case TM_Updated:
4506 0 : elog(ERROR, "tuple concurrently updated");
4507 : break;
4508 :
4509 2 : case TM_Deleted:
4510 2 : elog(ERROR, "tuple concurrently deleted");
4511 : break;
4512 :
4513 0 : default:
4514 0 : elog(ERROR, "unrecognized heap_update status: %u", result);
4515 : break;
4516 : }
4517 222758 : }
4518 :
4519 :
4520 : /*
4521 : * Return the MultiXactStatus corresponding to the given tuple lock mode.
4522 : */
4523 : static MultiXactStatus
4524 2390 : get_mxact_status_for_lock(LockTupleMode mode, bool is_update)
4525 : {
4526 : int retval;
4527 :
4528 2390 : if (is_update)
4529 192 : retval = tupleLockExtraInfo[mode].updstatus;
4530 : else
4531 2198 : retval = tupleLockExtraInfo[mode].lockstatus;
4532 :
4533 2390 : if (retval == -1)
4534 0 : elog(ERROR, "invalid lock tuple mode %d/%s", mode,
4535 : is_update ? "true" : "false");
4536 :
4537 2390 : return (MultiXactStatus) retval;
4538 : }
4539 :
4540 : /*
4541 : * heap_lock_tuple - lock a tuple in shared or exclusive mode
4542 : *
4543 : * Note that this acquires a buffer pin, which the caller must release.
4544 : *
4545 : * Input parameters:
4546 : * relation: relation containing tuple (caller must hold suitable lock)
4547 : * tid: TID of tuple to lock
4548 : * cid: current command ID (used for visibility test, and stored into
4549 : * tuple's cmax if lock is successful)
4550 : * mode: indicates if shared or exclusive tuple lock is desired
4551 : * wait_policy: what to do if tuple lock is not available
4552 : * follow_updates: if true, follow the update chain to also lock descendant
4553 : * tuples.
4554 : *
4555 : * Output parameters:
4556 : * *tuple: all fields filled in
4557 : * *buffer: set to buffer holding tuple (pinned but not locked at exit)
4558 : * *tmfd: filled in failure cases (see below)
4559 : *
4560 : * Function results are the same as the ones for table_tuple_lock().
4561 : *
4562 : * In the failure cases other than TM_Invisible, the routine fills
4563 : * *tmfd with the tuple's t_ctid, t_xmax (resolving a possible MultiXact,
4564 : * if necessary), and t_cmax (the last only for TM_SelfModified,
4565 : * since we cannot obtain cmax from a combo CID generated by another
4566 : * transaction).
4567 : * See comments for struct TM_FailureData for additional info.
4568 : *
4569 : * See README.tuplock for a thorough explanation of this mechanism.
4570 : */
4571 : TM_Result
4572 169758 : heap_lock_tuple(Relation relation, HeapTuple tuple,
4573 : CommandId cid, LockTupleMode mode, LockWaitPolicy wait_policy,
4574 : bool follow_updates,
4575 : Buffer *buffer, TM_FailureData *tmfd)
4576 : {
4577 : TM_Result result;
4578 169758 : ItemPointer tid = &(tuple->t_self);
4579 : ItemId lp;
4580 : Page page;
4581 169758 : Buffer vmbuffer = InvalidBuffer;
4582 : BlockNumber block;
4583 : TransactionId xid,
4584 : xmax;
4585 : uint16 old_infomask,
4586 : new_infomask,
4587 : new_infomask2;
4588 169758 : bool first_time = true;
4589 169758 : bool skip_tuple_lock = false;
4590 169758 : bool have_tuple_lock = false;
4591 169758 : bool cleared_all_frozen = false;
4592 :
4593 169758 : *buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
4594 169758 : block = ItemPointerGetBlockNumber(tid);
4595 :
4596 : /*
4597 : * Before locking the buffer, pin the visibility map page if it appears to
4598 : * be necessary. Since we haven't got the lock yet, someone else might be
4599 : * in the middle of changing this, so we'll need to recheck after we have
4600 : * the lock.
4601 : */
4602 169758 : if (PageIsAllVisible(BufferGetPage(*buffer)))
4603 3318 : visibilitymap_pin(relation, block, &vmbuffer);
4604 :
4605 169758 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4606 :
4607 169758 : page = BufferGetPage(*buffer);
4608 169758 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
4609 : Assert(ItemIdIsNormal(lp));
4610 :
4611 169758 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
4612 169758 : tuple->t_len = ItemIdGetLength(lp);
4613 169758 : tuple->t_tableOid = RelationGetRelid(relation);
4614 :
4615 28 : l3:
4616 169786 : result = HeapTupleSatisfiesUpdate(tuple, cid, *buffer);
4617 :
4618 169786 : if (result == TM_Invisible)
4619 : {
4620 : /*
4621 : * This is possible, but only when locking a tuple for ON CONFLICT
4622 : * UPDATE. We return this value here rather than throwing an error in
4623 : * order to give that case the opportunity to throw a more specific
4624 : * error.
4625 : */
4626 24 : result = TM_Invisible;
4627 24 : goto out_locked;
4628 : }
4629 169762 : else if (result == TM_BeingModified ||
4630 154072 : result == TM_Updated ||
4631 : result == TM_Deleted)
4632 : {
4633 : TransactionId xwait;
4634 : uint16 infomask;
4635 : uint16 infomask2;
4636 : bool require_sleep;
4637 : ItemPointerData t_ctid;
4638 :
4639 : /* must copy state data before unlocking buffer */
4640 15692 : xwait = HeapTupleHeaderGetRawXmax(tuple->t_data);
4641 15692 : infomask = tuple->t_data->t_infomask;
4642 15692 : infomask2 = tuple->t_data->t_infomask2;
4643 15692 : ItemPointerCopy(&tuple->t_data->t_ctid, &t_ctid);
4644 :
4645 15692 : LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4646 :
4647 : /*
4648 : * If any subtransaction of the current top transaction already holds
4649 : * a lock as strong as or stronger than what we're requesting, we
4650 : * effectively hold the desired lock already. We *must* succeed
4651 : * without trying to take the tuple lock, else we will deadlock
4652 : * against anyone wanting to acquire a stronger lock.
4653 : *
4654 : * Note we only do this the first time we loop on the HTSU result;
4655 : * there is no point in testing in subsequent passes, because
4656 : * evidently our own transaction cannot have acquired a new lock after
4657 : * the first time we checked.
4658 : */
4659 15692 : if (first_time)
4660 : {
4661 15674 : first_time = false;
4662 :
4663 15674 : if (infomask & HEAP_XMAX_IS_MULTI)
4664 : {
4665 : int i;
4666 : int nmembers;
4667 : MultiXactMember *members;
4668 :
4669 : /*
4670 : * We don't need to allow old multixacts here; if that had
4671 : * been the case, HeapTupleSatisfiesUpdate would have returned
4672 : * MayBeUpdated and we wouldn't be here.
4673 : */
4674 : nmembers =
4675 166 : GetMultiXactIdMembers(xwait, &members, false,
4676 166 : HEAP_XMAX_IS_LOCKED_ONLY(infomask));
4677 :
4678 498 : for (i = 0; i < nmembers; i++)
4679 : {
4680 : /* only consider members of our own transaction */
4681 360 : if (!TransactionIdIsCurrentTransactionId(members[i].xid))
4682 262 : continue;
4683 :
4684 98 : if (TUPLOCK_from_mxstatus(members[i].status) >= mode)
4685 : {
4686 28 : pfree(members);
4687 28 : result = TM_Ok;
4688 28 : goto out_unlocked;
4689 : }
4690 : else
4691 : {
4692 : /*
4693 : * Disable acquisition of the heavyweight tuple lock.
4694 : * Otherwise, when promoting a weaker lock, we might
4695 : * deadlock with another locker that has acquired the
4696 : * heavyweight tuple lock and is waiting for our
4697 : * transaction to finish.
4698 : *
4699 : * Note that in this case we still need to wait for
4700 : * the multixact if required, to avoid acquiring
4701 : * conflicting locks.
4702 : */
4703 70 : skip_tuple_lock = true;
4704 : }
4705 : }
4706 :
4707 138 : if (members)
4708 138 : pfree(members);
4709 : }
4710 15508 : else if (TransactionIdIsCurrentTransactionId(xwait))
4711 : {
4712 13066 : switch (mode)
4713 : {
4714 328 : case LockTupleKeyShare:
4715 : Assert(HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) ||
4716 : HEAP_XMAX_IS_SHR_LOCKED(infomask) ||
4717 : HEAP_XMAX_IS_EXCL_LOCKED(infomask));
4718 328 : result = TM_Ok;
4719 328 : goto out_unlocked;
4720 14 : case LockTupleShare:
4721 26 : if (HEAP_XMAX_IS_SHR_LOCKED(infomask) ||
4722 12 : HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4723 : {
4724 2 : result = TM_Ok;
4725 2 : goto out_unlocked;
4726 : }
4727 12 : break;
4728 122 : case LockTupleNoKeyExclusive:
4729 122 : if (HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4730 : {
4731 100 : result = TM_Ok;
4732 100 : goto out_unlocked;
4733 : }
4734 22 : break;
4735 12602 : case LockTupleExclusive:
4736 12602 : if (HEAP_XMAX_IS_EXCL_LOCKED(infomask) &&
4737 2522 : infomask2 & HEAP_KEYS_UPDATED)
4738 : {
4739 2480 : result = TM_Ok;
4740 2480 : goto out_unlocked;
4741 : }
4742 10122 : break;
4743 : }
4744 : }
4745 : }
4746 :
4747 : /*
4748 : * Initially assume that we will have to wait for the locking
4749 : * transaction(s) to finish. We check various cases below in which
4750 : * this can be turned off.
4751 : */
4752 12754 : require_sleep = true;
4753 12754 : if (mode == LockTupleKeyShare)
4754 : {
4755 : /*
4756 : * If we're requesting KeyShare, and there's no update present, we
4757 : * don't need to wait. Even if there is an update, we can still
4758 : * continue if the key hasn't been modified.
4759 : *
4760 : * However, if there are updates, we need to walk the update chain
4761 : * to mark future versions of the row as locked, too. That way,
4762 : * if somebody deletes that future version, we're protected
4763 : * against the key going away. This locking of future versions
4764 : * could block momentarily, if a concurrent transaction is
4765 : * deleting a key; or it could return a value to the effect that
4766 : * the transaction deleting the key has already committed. So we
4767 : * do this before re-locking the buffer; otherwise this would be
4768 : * prone to deadlocks.
4769 : *
4770 : * Note that the TID we're locking was grabbed before we unlocked
4771 : * the buffer. For it to change while we're not looking, the
4772 : * other properties we're testing for below after re-locking the
4773 : * buffer would also change, in which case we would restart this
4774 : * loop above.
4775 : */
4776 1150 : if (!(infomask2 & HEAP_KEYS_UPDATED))
4777 : {
4778 : bool updated;
4779 :
4780 1088 : updated = !HEAP_XMAX_IS_LOCKED_ONLY(infomask);
4781 :
4782 : /*
4783 : * If there are updates, follow the update chain; bail out if
4784 : * that cannot be done.
4785 : */
4786 1088 : if (follow_updates && updated)
4787 : {
4788 : TM_Result res;
4789 :
4790 100 : res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4791 : GetCurrentTransactionId(),
4792 : mode);
4793 100 : if (res != TM_Ok)
4794 : {
4795 12 : result = res;
4796 : /* recovery code expects to have buffer lock held */
4797 12 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4798 362 : goto failed;
4799 : }
4800 : }
4801 :
4802 1076 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4803 :
4804 : /*
4805 : * Make sure it's still an appropriate lock, else start over.
4806 : * Also, if it wasn't updated before we released the lock, but
4807 : * is updated now, we start over too; the reason is that we
4808 : * now need to follow the update chain to lock the new
4809 : * versions.
4810 : */
4811 1076 : if (!HeapTupleHeaderIsOnlyLocked(tuple->t_data) &&
4812 86 : ((tuple->t_data->t_infomask2 & HEAP_KEYS_UPDATED) ||
4813 86 : !updated))
4814 28 : goto l3;
4815 :
4816 : /* Things look okay, so we can skip sleeping */
4817 1076 : require_sleep = false;
4818 :
4819 : /*
4820 : * Note we allow Xmax to change here; other updaters/lockers
4821 : * could have modified it before we grabbed the buffer lock.
4822 : * However, this is not a problem, because with the recheck we
4823 : * just did we ensure that they still don't conflict with the
4824 : * lock we want.
4825 : */
4826 : }
4827 : }
4828 11604 : else if (mode == LockTupleShare)
4829 : {
4830 : /*
4831 : * If we're requesting Share, we can similarly avoid sleeping if
4832 : * there's no update and no exclusive lock present.
4833 : */
4834 882 : if (HEAP_XMAX_IS_LOCKED_ONLY(infomask) &&
4835 882 : !HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4836 : {
4837 870 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4838 :
4839 : /*
4840 : * Make sure it's still an appropriate lock, else start over.
4841 : * See above about allowing xmax to change.
4842 : */
4843 1740 : if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4844 870 : HEAP_XMAX_IS_EXCL_LOCKED(tuple->t_data->t_infomask))
4845 0 : goto l3;
4846 870 : require_sleep = false;
4847 : }
4848 : }
4849 10722 : else if (mode == LockTupleNoKeyExclusive)
4850 : {
4851 : /*
4852 : * If we're requesting NoKeyExclusive, we might also be able to
4853 : * avoid sleeping; just ensure that there no conflicting lock
4854 : * already acquired.
4855 : */
4856 312 : if (infomask & HEAP_XMAX_IS_MULTI)
4857 : {
4858 52 : if (!DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
4859 : mode, NULL))
4860 : {
4861 : /*
4862 : * No conflict, but if the xmax changed under us in the
4863 : * meantime, start over.
4864 : */
4865 26 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4866 52 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4867 26 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4868 : xwait))
4869 0 : goto l3;
4870 :
4871 : /* otherwise, we're good */
4872 26 : require_sleep = false;
4873 : }
4874 : }
4875 260 : else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask))
4876 : {
4877 34 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4878 :
4879 : /* if the xmax changed in the meantime, start over */
4880 68 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4881 34 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4882 : xwait))
4883 0 : goto l3;
4884 : /* otherwise, we're good */
4885 34 : require_sleep = false;
4886 : }
4887 : }
4888 :
4889 : /*
4890 : * As a check independent from those above, we can also avoid sleeping
4891 : * if the current transaction is the sole locker of the tuple. Note
4892 : * that the strength of the lock already held is irrelevant; this is
4893 : * not about recording the lock in Xmax (which will be done regardless
4894 : * of this optimization, below). Also, note that the cases where we
4895 : * hold a lock stronger than we are requesting are already handled
4896 : * above by not doing anything.
4897 : *
4898 : * Note we only deal with the non-multixact case here; MultiXactIdWait
4899 : * is well equipped to deal with this situation on its own.
4900 : */
4901 23396 : if (require_sleep && !(infomask & HEAP_XMAX_IS_MULTI) &&
4902 10654 : TransactionIdIsCurrentTransactionId(xwait))
4903 : {
4904 : /* ... but if the xmax changed in the meantime, start over */
4905 10122 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4906 20244 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4907 10122 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4908 : xwait))
4909 0 : goto l3;
4910 : Assert(HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask));
4911 10122 : require_sleep = false;
4912 : }
4913 :
4914 : /*
4915 : * Time to sleep on the other transaction/multixact, if necessary.
4916 : *
4917 : * If the other transaction is an update/delete that's already
4918 : * committed, then sleeping cannot possibly do any good: if we're
4919 : * required to sleep, get out to raise an error instead.
4920 : *
4921 : * By here, we either have already acquired the buffer exclusive lock,
4922 : * or we must wait for the locking transaction or multixact; so below
4923 : * we ensure that we grab buffer lock after the sleep.
4924 : */
4925 12742 : if (require_sleep && (result == TM_Updated || result == TM_Deleted))
4926 : {
4927 274 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4928 274 : goto failed;
4929 : }
4930 12468 : else if (require_sleep)
4931 : {
4932 : /*
4933 : * Acquire tuple lock to establish our priority for the tuple, or
4934 : * die trying. LockTuple will release us when we are next-in-line
4935 : * for the tuple. We must do this even if we are share-locking,
4936 : * but not if we already have a weaker lock on the tuple.
4937 : *
4938 : * If we are forced to "start over" below, we keep the tuple lock;
4939 : * this arranges that we stay at the head of the line while
4940 : * rechecking tuple state.
4941 : */
4942 340 : if (!skip_tuple_lock &&
4943 308 : !heap_acquire_tuplock(relation, tid, mode, wait_policy,
4944 : &have_tuple_lock))
4945 : {
4946 : /*
4947 : * This can only happen if wait_policy is Skip and the lock
4948 : * couldn't be obtained.
4949 : */
4950 2 : result = TM_WouldBlock;
4951 : /* recovery code expects to have buffer lock held */
4952 2 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4953 2 : goto failed;
4954 : }
4955 :
4956 336 : if (infomask & HEAP_XMAX_IS_MULTI)
4957 : {
4958 80 : MultiXactStatus status = get_mxact_status_for_lock(mode, false);
4959 :
4960 : /* We only ever lock tuples, never update them */
4961 80 : if (status >= MultiXactStatusNoKeyUpdate)
4962 0 : elog(ERROR, "invalid lock mode in heap_lock_tuple");
4963 :
4964 : /* wait for multixact to end, or die trying */
4965 80 : switch (wait_policy)
4966 : {
4967 72 : case LockWaitBlock:
4968 72 : MultiXactIdWait((MultiXactId) xwait, status, infomask,
4969 : relation, &tuple->t_self, XLTW_Lock, NULL);
4970 72 : break;
4971 4 : case LockWaitSkip:
4972 4 : if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4973 : status, infomask, relation,
4974 : NULL, false))
4975 : {
4976 4 : result = TM_WouldBlock;
4977 : /* recovery code expects to have buffer lock held */
4978 4 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4979 4 : goto failed;
4980 : }
4981 0 : break;
4982 4 : case LockWaitError:
4983 4 : if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4984 : status, infomask, relation,
4985 : NULL, log_lock_failures))
4986 4 : ereport(ERROR,
4987 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4988 : errmsg("could not obtain lock on row in relation \"%s\"",
4989 : RelationGetRelationName(relation))));
4990 :
4991 0 : break;
4992 : }
4993 :
4994 : /*
4995 : * Of course, the multixact might not be done here: if we're
4996 : * requesting a light lock mode, other transactions with light
4997 : * locks could still be alive, as well as locks owned by our
4998 : * own xact or other subxacts of this backend. We need to
4999 : * preserve the surviving MultiXact members. Note that it
5000 : * isn't absolutely necessary in the latter case, but doing so
5001 : * is simpler.
5002 : */
5003 : }
5004 : else
5005 : {
5006 : /* wait for regular transaction to end, or die trying */
5007 256 : switch (wait_policy)
5008 : {
5009 178 : case LockWaitBlock:
5010 178 : XactLockTableWait(xwait, relation, &tuple->t_self,
5011 : XLTW_Lock);
5012 178 : break;
5013 66 : case LockWaitSkip:
5014 66 : if (!ConditionalXactLockTableWait(xwait, false))
5015 : {
5016 66 : result = TM_WouldBlock;
5017 : /* recovery code expects to have buffer lock held */
5018 66 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5019 66 : goto failed;
5020 : }
5021 0 : break;
5022 12 : case LockWaitError:
5023 12 : if (!ConditionalXactLockTableWait(xwait, log_lock_failures))
5024 12 : ereport(ERROR,
5025 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
5026 : errmsg("could not obtain lock on row in relation \"%s\"",
5027 : RelationGetRelationName(relation))));
5028 0 : break;
5029 : }
5030 : }
5031 :
5032 : /* if there are updates, follow the update chain */
5033 250 : if (follow_updates && !HEAP_XMAX_IS_LOCKED_ONLY(infomask))
5034 : {
5035 : TM_Result res;
5036 :
5037 80 : res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
5038 : GetCurrentTransactionId(),
5039 : mode);
5040 80 : if (res != TM_Ok)
5041 : {
5042 4 : result = res;
5043 : /* recovery code expects to have buffer lock held */
5044 4 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5045 4 : goto failed;
5046 : }
5047 : }
5048 :
5049 246 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5050 :
5051 : /*
5052 : * xwait is done, but if xwait had just locked the tuple then some
5053 : * other xact could update this tuple before we get to this point.
5054 : * Check for xmax change, and start over if so.
5055 : */
5056 468 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
5057 222 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
5058 : xwait))
5059 28 : goto l3;
5060 :
5061 218 : if (!(infomask & HEAP_XMAX_IS_MULTI))
5062 : {
5063 : /*
5064 : * Otherwise check if it committed or aborted. Note we cannot
5065 : * be here if the tuple was only locked by somebody who didn't
5066 : * conflict with us; that would have been handled above. So
5067 : * that transaction must necessarily be gone by now. But
5068 : * don't check for this in the multixact case, because some
5069 : * locker transactions might still be running.
5070 : */
5071 156 : UpdateXmaxHintBits(tuple->t_data, *buffer, xwait);
5072 : }
5073 : }
5074 :
5075 : /* By here, we're certain that we hold buffer exclusive lock again */
5076 :
5077 : /*
5078 : * We may lock if previous xmax aborted, or if it committed but only
5079 : * locked the tuple without updating it; or if we didn't have to wait
5080 : * at all for whatever reason.
5081 : */
5082 12346 : if (!require_sleep ||
5083 376 : (tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
5084 286 : HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
5085 128 : HeapTupleHeaderIsOnlyLocked(tuple->t_data))
5086 12230 : result = TM_Ok;
5087 116 : else if (!ItemPointerEquals(&tuple->t_self, &tuple->t_data->t_ctid))
5088 92 : result = TM_Updated;
5089 : else
5090 24 : result = TM_Deleted;
5091 : }
5092 :
5093 154070 : failed:
5094 166778 : if (result != TM_Ok)
5095 : {
5096 : Assert(result == TM_SelfModified || result == TM_Updated ||
5097 : result == TM_Deleted || result == TM_WouldBlock);
5098 :
5099 : /*
5100 : * When locking a tuple under LockWaitSkip semantics and we fail with
5101 : * TM_WouldBlock above, it's possible for concurrent transactions to
5102 : * release the lock and set HEAP_XMAX_INVALID in the meantime. So
5103 : * this assert is slightly different from the equivalent one in
5104 : * heap_delete and heap_update.
5105 : */
5106 : Assert((result == TM_WouldBlock) ||
5107 : !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));
5108 : Assert(result != TM_Updated ||
5109 : !ItemPointerEquals(&tuple->t_self, &tuple->t_data->t_ctid));
5110 490 : tmfd->ctid = tuple->t_data->t_ctid;
5111 490 : tmfd->xmax = HeapTupleHeaderGetUpdateXid(tuple->t_data);
5112 490 : if (result == TM_SelfModified)
5113 12 : tmfd->cmax = HeapTupleHeaderGetCmax(tuple->t_data);
5114 : else
5115 478 : tmfd->cmax = InvalidCommandId;
5116 490 : goto out_locked;
5117 : }
5118 :
5119 : /*
5120 : * If we didn't pin the visibility map page and the page has become all
5121 : * visible while we were busy locking the buffer, or during some
5122 : * subsequent window during which we had it unlocked, we'll have to unlock
5123 : * and re-lock, to avoid holding the buffer lock across I/O. That's a bit
5124 : * unfortunate, especially since we'll now have to recheck whether the
5125 : * tuple has been locked or updated under us, but hopefully it won't
5126 : * happen very often.
5127 : */
5128 166288 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
5129 : {
5130 0 : LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
5131 0 : visibilitymap_pin(relation, block, &vmbuffer);
5132 0 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5133 0 : goto l3;
5134 : }
5135 :
5136 166288 : xmax = HeapTupleHeaderGetRawXmax(tuple->t_data);
5137 166288 : old_infomask = tuple->t_data->t_infomask;
5138 :
5139 : /*
5140 : * If this is the first possibly-multixact-able operation in the current
5141 : * transaction, set my per-backend OldestMemberMXactId setting. We can be
5142 : * certain that the transaction will never become a member of any older
5143 : * MultiXactIds than that. (We have to do this even if we end up just
5144 : * using our own TransactionId below, since some other backend could
5145 : * incorporate our XID into a MultiXact immediately afterwards.)
5146 : */
5147 166288 : MultiXactIdSetOldestMember();
5148 :
5149 : /*
5150 : * Compute the new xmax and infomask to store into the tuple. Note we do
5151 : * not modify the tuple just yet, because that would leave it in the wrong
5152 : * state if multixact.c elogs.
5153 : */
5154 166288 : compute_new_xmax_infomask(xmax, old_infomask, tuple->t_data->t_infomask2,
5155 : GetCurrentTransactionId(), mode, false,
5156 : &xid, &new_infomask, &new_infomask2);
5157 :
5158 166288 : START_CRIT_SECTION();
5159 :
5160 : /*
5161 : * Store transaction information of xact locking the tuple.
5162 : *
5163 : * Note: Cmax is meaningless in this context, so don't set it; this avoids
5164 : * possibly generating a useless combo CID. Moreover, if we're locking a
5165 : * previously updated tuple, it's important to preserve the Cmax.
5166 : *
5167 : * Also reset the HOT UPDATE bit, but only if there's no update; otherwise
5168 : * we would break the HOT chain.
5169 : */
5170 166288 : tuple->t_data->t_infomask &= ~HEAP_XMAX_BITS;
5171 166288 : tuple->t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5172 166288 : tuple->t_data->t_infomask |= new_infomask;
5173 166288 : tuple->t_data->t_infomask2 |= new_infomask2;
5174 166288 : if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
5175 166210 : HeapTupleHeaderClearHotUpdated(tuple->t_data);
5176 166288 : HeapTupleHeaderSetXmax(tuple->t_data, xid);
5177 :
5178 : /*
5179 : * Make sure there is no forward chain link in t_ctid. Note that in the
5180 : * cases where the tuple has been updated, we must not overwrite t_ctid,
5181 : * because it was set by the updater. Moreover, if the tuple has been
5182 : * updated, we need to follow the update chain to lock the new versions of
5183 : * the tuple as well.
5184 : */
5185 166288 : if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
5186 166210 : tuple->t_data->t_ctid = *tid;
5187 :
5188 : /* Clear only the all-frozen bit on visibility map if needed */
5189 169606 : if (PageIsAllVisible(page) &&
5190 3318 : visibilitymap_clear(relation, block, vmbuffer,
5191 : VISIBILITYMAP_ALL_FROZEN))
5192 30 : cleared_all_frozen = true;
5193 :
5194 :
5195 166288 : MarkBufferDirty(*buffer);
5196 :
5197 : /*
5198 : * XLOG stuff. You might think that we don't need an XLOG record because
5199 : * there is no state change worth restoring after a crash. You would be
5200 : * wrong however: we have just written either a TransactionId or a
5201 : * MultiXactId that may never have been seen on disk before, and we need
5202 : * to make sure that there are XLOG entries covering those ID numbers.
5203 : * Else the same IDs might be re-used after a crash, which would be
5204 : * disastrous if this page made it to disk before the crash. Essentially
5205 : * we have to enforce the WAL log-before-data rule even in this case.
5206 : * (Also, in a PITR log-shipping or 2PC environment, we have to have XLOG
5207 : * entries for everything anyway.)
5208 : */
5209 166288 : if (RelationNeedsWAL(relation))
5210 : {
5211 : xl_heap_lock xlrec;
5212 : XLogRecPtr recptr;
5213 :
5214 165600 : XLogBeginInsert();
5215 165600 : XLogRegisterBuffer(0, *buffer, REGBUF_STANDARD);
5216 :
5217 165600 : xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
5218 165600 : xlrec.xmax = xid;
5219 331200 : xlrec.infobits_set = compute_infobits(new_infomask,
5220 165600 : tuple->t_data->t_infomask2);
5221 165600 : xlrec.flags = cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
5222 165600 : XLogRegisterData(&xlrec, SizeOfHeapLock);
5223 :
5224 : /* we don't decode row locks atm, so no need to log the origin */
5225 :
5226 165600 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK);
5227 :
5228 165600 : PageSetLSN(page, recptr);
5229 : }
5230 :
5231 166288 : END_CRIT_SECTION();
5232 :
5233 166288 : result = TM_Ok;
5234 :
5235 166802 : out_locked:
5236 166802 : LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
5237 :
5238 169740 : out_unlocked:
5239 169740 : if (BufferIsValid(vmbuffer))
5240 3318 : ReleaseBuffer(vmbuffer);
5241 :
5242 : /*
5243 : * Don't update the visibility map here. Locking a tuple doesn't change
5244 : * visibility info.
5245 : */
5246 :
5247 : /*
5248 : * Now that we have successfully marked the tuple as locked, we can
5249 : * release the lmgr tuple lock, if we had it.
5250 : */
5251 169740 : if (have_tuple_lock)
5252 278 : UnlockTupleTuplock(relation, tid, mode);
5253 :
5254 169740 : return result;
5255 : }
5256 :
5257 : /*
5258 : * Acquire heavyweight lock on the given tuple, in preparation for acquiring
5259 : * its normal, Xmax-based tuple lock.
5260 : *
5261 : * have_tuple_lock is an input and output parameter: on input, it indicates
5262 : * whether the lock has previously been acquired (and this function does
5263 : * nothing in that case). If this function returns success, have_tuple_lock
5264 : * has been flipped to true.
5265 : *
5266 : * Returns false if it was unable to obtain the lock; this can only happen if
5267 : * wait_policy is Skip.
5268 : */
5269 : static bool
5270 530 : heap_acquire_tuplock(Relation relation, ItemPointer tid, LockTupleMode mode,
5271 : LockWaitPolicy wait_policy, bool *have_tuple_lock)
5272 : {
5273 530 : if (*have_tuple_lock)
5274 18 : return true;
5275 :
5276 512 : switch (wait_policy)
5277 : {
5278 430 : case LockWaitBlock:
5279 430 : LockTupleTuplock(relation, tid, mode);
5280 430 : break;
5281 :
5282 68 : case LockWaitSkip:
5283 68 : if (!ConditionalLockTupleTuplock(relation, tid, mode, false))
5284 2 : return false;
5285 66 : break;
5286 :
5287 14 : case LockWaitError:
5288 14 : if (!ConditionalLockTupleTuplock(relation, tid, mode, log_lock_failures))
5289 2 : ereport(ERROR,
5290 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
5291 : errmsg("could not obtain lock on row in relation \"%s\"",
5292 : RelationGetRelationName(relation))));
5293 12 : break;
5294 : }
5295 508 : *have_tuple_lock = true;
5296 :
5297 508 : return true;
5298 : }
5299 :
5300 : /*
5301 : * Given an original set of Xmax and infomask, and a transaction (identified by
5302 : * add_to_xmax) acquiring a new lock of some mode, compute the new Xmax and
5303 : * corresponding infomasks to use on the tuple.
5304 : *
5305 : * Note that this might have side effects such as creating a new MultiXactId.
5306 : *
5307 : * Most callers will have called HeapTupleSatisfiesUpdate before this function;
5308 : * that will have set the HEAP_XMAX_INVALID bit if the xmax was a MultiXactId
5309 : * but it was not running anymore. There is a race condition, which is that the
5310 : * MultiXactId may have finished since then, but that uncommon case is handled
5311 : * either here, or within MultiXactIdExpand.
5312 : *
5313 : * There is a similar race condition possible when the old xmax was a regular
5314 : * TransactionId. We test TransactionIdIsInProgress again just to narrow the
5315 : * window, but it's still possible to end up creating an unnecessary
5316 : * MultiXactId. Fortunately this is harmless.
5317 : */
5318 : static void
5319 4028376 : compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
5320 : uint16 old_infomask2, TransactionId add_to_xmax,
5321 : LockTupleMode mode, bool is_update,
5322 : TransactionId *result_xmax, uint16 *result_infomask,
5323 : uint16 *result_infomask2)
5324 : {
5325 : TransactionId new_xmax;
5326 : uint16 new_infomask,
5327 : new_infomask2;
5328 :
5329 : Assert(TransactionIdIsCurrentTransactionId(add_to_xmax));
5330 :
5331 207940 : l5:
5332 4236316 : new_infomask = 0;
5333 4236316 : new_infomask2 = 0;
5334 4236316 : if (old_infomask & HEAP_XMAX_INVALID)
5335 : {
5336 : /*
5337 : * No previous locker; we just insert our own TransactionId.
5338 : *
5339 : * Note that it's critical that this case be the first one checked,
5340 : * because there are several blocks below that come back to this one
5341 : * to implement certain optimizations; old_infomask might contain
5342 : * other dirty bits in those cases, but we don't really care.
5343 : */
5344 4026130 : if (is_update)
5345 : {
5346 3561458 : new_xmax = add_to_xmax;
5347 3561458 : if (mode == LockTupleExclusive)
5348 3024976 : new_infomask2 |= HEAP_KEYS_UPDATED;
5349 : }
5350 : else
5351 : {
5352 464672 : new_infomask |= HEAP_XMAX_LOCK_ONLY;
5353 464672 : switch (mode)
5354 : {
5355 5072 : case LockTupleKeyShare:
5356 5072 : new_xmax = add_to_xmax;
5357 5072 : new_infomask |= HEAP_XMAX_KEYSHR_LOCK;
5358 5072 : break;
5359 1444 : case LockTupleShare:
5360 1444 : new_xmax = add_to_xmax;
5361 1444 : new_infomask |= HEAP_XMAX_SHR_LOCK;
5362 1444 : break;
5363 266798 : case LockTupleNoKeyExclusive:
5364 266798 : new_xmax = add_to_xmax;
5365 266798 : new_infomask |= HEAP_XMAX_EXCL_LOCK;
5366 266798 : break;
5367 191358 : case LockTupleExclusive:
5368 191358 : new_xmax = add_to_xmax;
5369 191358 : new_infomask |= HEAP_XMAX_EXCL_LOCK;
5370 191358 : new_infomask2 |= HEAP_KEYS_UPDATED;
5371 191358 : break;
5372 0 : default:
5373 0 : new_xmax = InvalidTransactionId; /* silence compiler */
5374 0 : elog(ERROR, "invalid lock mode");
5375 : }
5376 : }
5377 : }
5378 210186 : else if (old_infomask & HEAP_XMAX_IS_MULTI)
5379 : {
5380 : MultiXactStatus new_status;
5381 :
5382 : /*
5383 : * Currently we don't allow XMAX_COMMITTED to be set for multis, so
5384 : * cross-check.
5385 : */
5386 : Assert(!(old_infomask & HEAP_XMAX_COMMITTED));
5387 :
5388 : /*
5389 : * A multixact together with LOCK_ONLY set but neither lock bit set
5390 : * (i.e. a pg_upgraded share locked tuple) cannot possibly be running
5391 : * anymore. This check is critical for databases upgraded by
5392 : * pg_upgrade; both MultiXactIdIsRunning and MultiXactIdExpand assume
5393 : * that such multis are never passed.
5394 : */
5395 238 : if (HEAP_LOCKED_UPGRADED(old_infomask))
5396 : {
5397 0 : old_infomask &= ~HEAP_XMAX_IS_MULTI;
5398 0 : old_infomask |= HEAP_XMAX_INVALID;
5399 0 : goto l5;
5400 : }
5401 :
5402 : /*
5403 : * If the XMAX is already a MultiXactId, then we need to expand it to
5404 : * include add_to_xmax; but if all the members were lockers and are
5405 : * all gone, we can do away with the IS_MULTI bit and just set
5406 : * add_to_xmax as the only locker/updater. If all lockers are gone
5407 : * and we have an updater that aborted, we can also do without a
5408 : * multi.
5409 : *
5410 : * The cost of doing GetMultiXactIdMembers would be paid by
5411 : * MultiXactIdExpand if we weren't to do this, so this check is not
5412 : * incurring extra work anyhow.
5413 : */
5414 238 : if (!MultiXactIdIsRunning(xmax, HEAP_XMAX_IS_LOCKED_ONLY(old_infomask)))
5415 : {
5416 46 : if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) ||
5417 16 : !TransactionIdDidCommit(MultiXactIdGetUpdateXid(xmax,
5418 : old_infomask)))
5419 : {
5420 : /*
5421 : * Reset these bits and restart; otherwise fall through to
5422 : * create a new multi below.
5423 : */
5424 46 : old_infomask &= ~HEAP_XMAX_IS_MULTI;
5425 46 : old_infomask |= HEAP_XMAX_INVALID;
5426 46 : goto l5;
5427 : }
5428 : }
5429 :
5430 192 : new_status = get_mxact_status_for_lock(mode, is_update);
5431 :
5432 192 : new_xmax = MultiXactIdExpand((MultiXactId) xmax, add_to_xmax,
5433 : new_status);
5434 192 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5435 : }
5436 209948 : else if (old_infomask & HEAP_XMAX_COMMITTED)
5437 : {
5438 : /*
5439 : * It's a committed update, so we need to preserve him as updater of
5440 : * the tuple.
5441 : */
5442 : MultiXactStatus status;
5443 : MultiXactStatus new_status;
5444 :
5445 26 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5446 0 : status = MultiXactStatusUpdate;
5447 : else
5448 26 : status = MultiXactStatusNoKeyUpdate;
5449 :
5450 26 : new_status = get_mxact_status_for_lock(mode, is_update);
5451 :
5452 : /*
5453 : * since it's not running, it's obviously impossible for the old
5454 : * updater to be identical to the current one, so we need not check
5455 : * for that case as we do in the block above.
5456 : */
5457 26 : new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
5458 26 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5459 : }
5460 209922 : else if (TransactionIdIsInProgress(xmax))
5461 : {
5462 : /*
5463 : * If the XMAX is a valid, in-progress TransactionId, then we need to
5464 : * create a new MultiXactId that includes both the old locker or
5465 : * updater and our own TransactionId.
5466 : */
5467 : MultiXactStatus new_status;
5468 : MultiXactStatus old_status;
5469 : LockTupleMode old_mode;
5470 :
5471 209904 : if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5472 : {
5473 209852 : if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5474 11230 : old_status = MultiXactStatusForKeyShare;
5475 198622 : else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5476 862 : old_status = MultiXactStatusForShare;
5477 197760 : else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5478 : {
5479 197760 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5480 185536 : old_status = MultiXactStatusForUpdate;
5481 : else
5482 12224 : old_status = MultiXactStatusForNoKeyUpdate;
5483 : }
5484 : else
5485 : {
5486 : /*
5487 : * LOCK_ONLY can be present alone only when a page has been
5488 : * upgraded by pg_upgrade. But in that case,
5489 : * TransactionIdIsInProgress() should have returned false. We
5490 : * assume it's no longer locked in this case.
5491 : */
5492 0 : elog(WARNING, "LOCK_ONLY found for Xid in progress %u", xmax);
5493 0 : old_infomask |= HEAP_XMAX_INVALID;
5494 0 : old_infomask &= ~HEAP_XMAX_LOCK_ONLY;
5495 0 : goto l5;
5496 : }
5497 : }
5498 : else
5499 : {
5500 : /* it's an update, but which kind? */
5501 52 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5502 0 : old_status = MultiXactStatusUpdate;
5503 : else
5504 52 : old_status = MultiXactStatusNoKeyUpdate;
5505 : }
5506 :
5507 209904 : old_mode = TUPLOCK_from_mxstatus(old_status);
5508 :
5509 : /*
5510 : * If the lock to be acquired is for the same TransactionId as the
5511 : * existing lock, there's an optimization possible: consider only the
5512 : * strongest of both locks as the only one present, and restart.
5513 : */
5514 209904 : if (xmax == add_to_xmax)
5515 : {
5516 : /*
5517 : * Note that it's not possible for the original tuple to be
5518 : * updated: we wouldn't be here because the tuple would have been
5519 : * invisible and we wouldn't try to update it. As a subtlety,
5520 : * this code can also run when traversing an update chain to lock
5521 : * future versions of a tuple. But we wouldn't be here either,
5522 : * because the add_to_xmax would be different from the original
5523 : * updater.
5524 : */
5525 : Assert(HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5526 :
5527 : /* acquire the strongest of both */
5528 207878 : if (mode < old_mode)
5529 104424 : mode = old_mode;
5530 : /* mustn't touch is_update */
5531 :
5532 207878 : old_infomask |= HEAP_XMAX_INVALID;
5533 207878 : goto l5;
5534 : }
5535 :
5536 : /* otherwise, just fall back to creating a new multixact */
5537 2026 : new_status = get_mxact_status_for_lock(mode, is_update);
5538 2026 : new_xmax = MultiXactIdCreate(xmax, old_status,
5539 : add_to_xmax, new_status);
5540 2026 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5541 : }
5542 28 : else if (!HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) &&
5543 10 : TransactionIdDidCommit(xmax))
5544 2 : {
5545 : /*
5546 : * It's a committed update, so we gotta preserve him as updater of the
5547 : * tuple.
5548 : */
5549 : MultiXactStatus status;
5550 : MultiXactStatus new_status;
5551 :
5552 2 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5553 0 : status = MultiXactStatusUpdate;
5554 : else
5555 2 : status = MultiXactStatusNoKeyUpdate;
5556 :
5557 2 : new_status = get_mxact_status_for_lock(mode, is_update);
5558 :
5559 : /*
5560 : * since it's not running, it's obviously impossible for the old
5561 : * updater to be identical to the current one, so we need not check
5562 : * for that case as we do in the block above.
5563 : */
5564 2 : new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
5565 2 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5566 : }
5567 : else
5568 : {
5569 : /*
5570 : * Can get here iff the locking/updating transaction was running when
5571 : * the infomask was extracted from the tuple, but finished before
5572 : * TransactionIdIsInProgress got to run. Deal with it as if there was
5573 : * no locker at all in the first place.
5574 : */
5575 16 : old_infomask |= HEAP_XMAX_INVALID;
5576 16 : goto l5;
5577 : }
5578 :
5579 4028376 : *result_infomask = new_infomask;
5580 4028376 : *result_infomask2 = new_infomask2;
5581 4028376 : *result_xmax = new_xmax;
5582 4028376 : }
5583 :
5584 : /*
5585 : * Subroutine for heap_lock_updated_tuple_rec.
5586 : *
5587 : * Given a hypothetical multixact status held by the transaction identified
5588 : * with the given xid, does the current transaction need to wait, fail, or can
5589 : * it continue if it wanted to acquire a lock of the given mode? "needwait"
5590 : * is set to true if waiting is necessary; if it can continue, then TM_Ok is
5591 : * returned. If the lock is already held by the current transaction, return
5592 : * TM_SelfModified. In case of a conflict with another transaction, a
5593 : * different HeapTupleSatisfiesUpdate return code is returned.
5594 : *
5595 : * The held status is said to be hypothetical because it might correspond to a
5596 : * lock held by a single Xid, i.e. not a real MultiXactId; we express it this
5597 : * way for simplicity of API.
5598 : */
5599 : static TM_Result
5600 64 : test_lockmode_for_conflict(MultiXactStatus status, TransactionId xid,
5601 : LockTupleMode mode, HeapTuple tup,
5602 : bool *needwait)
5603 : {
5604 : MultiXactStatus wantedstatus;
5605 :
5606 64 : *needwait = false;
5607 64 : wantedstatus = get_mxact_status_for_lock(mode, false);
5608 :
5609 : /*
5610 : * Note: we *must* check TransactionIdIsInProgress before
5611 : * TransactionIdDidAbort/Commit; see comment at top of heapam_visibility.c
5612 : * for an explanation.
5613 : */
5614 64 : if (TransactionIdIsCurrentTransactionId(xid))
5615 : {
5616 : /*
5617 : * The tuple has already been locked by our own transaction. This is
5618 : * very rare but can happen if multiple transactions are trying to
5619 : * lock an ancient version of the same tuple.
5620 : */
5621 0 : return TM_SelfModified;
5622 : }
5623 64 : else if (TransactionIdIsInProgress(xid))
5624 : {
5625 : /*
5626 : * If the locking transaction is running, what we do depends on
5627 : * whether the lock modes conflict: if they do, then we must wait for
5628 : * it to finish; otherwise we can fall through to lock this tuple
5629 : * version without waiting.
5630 : */
5631 32 : if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5632 32 : LOCKMODE_from_mxstatus(wantedstatus)))
5633 : {
5634 16 : *needwait = true;
5635 : }
5636 :
5637 : /*
5638 : * If we set needwait above, then this value doesn't matter;
5639 : * otherwise, this value signals to caller that it's okay to proceed.
5640 : */
5641 32 : return TM_Ok;
5642 : }
5643 32 : else if (TransactionIdDidAbort(xid))
5644 6 : return TM_Ok;
5645 26 : else if (TransactionIdDidCommit(xid))
5646 : {
5647 : /*
5648 : * The other transaction committed. If it was only a locker, then the
5649 : * lock is completely gone now and we can return success; but if it
5650 : * was an update, then what we do depends on whether the two lock
5651 : * modes conflict. If they conflict, then we must report error to
5652 : * caller. But if they don't, we can fall through to allow the current
5653 : * transaction to lock the tuple.
5654 : *
5655 : * Note: the reason we worry about ISUPDATE here is because as soon as
5656 : * a transaction ends, all its locks are gone and meaningless, and
5657 : * thus we can ignore them; whereas its updates persist. In the
5658 : * TransactionIdIsInProgress case, above, we don't need to check
5659 : * because we know the lock is still "alive" and thus a conflict needs
5660 : * always be checked.
5661 : */
5662 26 : if (!ISUPDATE_from_mxstatus(status))
5663 8 : return TM_Ok;
5664 :
5665 18 : if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5666 18 : LOCKMODE_from_mxstatus(wantedstatus)))
5667 : {
5668 : /* bummer */
5669 16 : if (!ItemPointerEquals(&tup->t_self, &tup->t_data->t_ctid))
5670 12 : return TM_Updated;
5671 : else
5672 4 : return TM_Deleted;
5673 : }
5674 :
5675 2 : return TM_Ok;
5676 : }
5677 :
5678 : /* Not in progress, not aborted, not committed -- must have crashed */
5679 0 : return TM_Ok;
5680 : }
5681 :
5682 :
5683 : /*
5684 : * Recursive part of heap_lock_updated_tuple
5685 : *
5686 : * Fetch the tuple pointed to by tid in rel, and mark it as locked by the given
5687 : * xid with the given mode; if this tuple is updated, recurse to lock the new
5688 : * version as well.
5689 : */
5690 : static TM_Result
5691 162 : heap_lock_updated_tuple_rec(Relation rel, ItemPointer tid, TransactionId xid,
5692 : LockTupleMode mode)
5693 : {
5694 : TM_Result result;
5695 : ItemPointerData tupid;
5696 : HeapTupleData mytup;
5697 : Buffer buf;
5698 : uint16 new_infomask,
5699 : new_infomask2,
5700 : old_infomask,
5701 : old_infomask2;
5702 : TransactionId xmax,
5703 : new_xmax;
5704 162 : TransactionId priorXmax = InvalidTransactionId;
5705 162 : bool cleared_all_frozen = false;
5706 : bool pinned_desired_page;
5707 162 : Buffer vmbuffer = InvalidBuffer;
5708 : BlockNumber block;
5709 :
5710 162 : ItemPointerCopy(tid, &tupid);
5711 :
5712 : for (;;)
5713 : {
5714 168 : new_infomask = 0;
5715 168 : new_xmax = InvalidTransactionId;
5716 168 : block = ItemPointerGetBlockNumber(&tupid);
5717 168 : ItemPointerCopy(&tupid, &(mytup.t_self));
5718 :
5719 168 : if (!heap_fetch(rel, SnapshotAny, &mytup, &buf, false))
5720 : {
5721 : /*
5722 : * if we fail to find the updated version of the tuple, it's
5723 : * because it was vacuumed/pruned away after its creator
5724 : * transaction aborted. So behave as if we got to the end of the
5725 : * chain, and there's no further tuple to lock: return success to
5726 : * caller.
5727 : */
5728 0 : result = TM_Ok;
5729 0 : goto out_unlocked;
5730 : }
5731 :
5732 168 : l4:
5733 184 : CHECK_FOR_INTERRUPTS();
5734 :
5735 : /*
5736 : * Before locking the buffer, pin the visibility map page if it
5737 : * appears to be necessary. Since we haven't got the lock yet,
5738 : * someone else might be in the middle of changing this, so we'll need
5739 : * to recheck after we have the lock.
5740 : */
5741 184 : if (PageIsAllVisible(BufferGetPage(buf)))
5742 : {
5743 0 : visibilitymap_pin(rel, block, &vmbuffer);
5744 0 : pinned_desired_page = true;
5745 : }
5746 : else
5747 184 : pinned_desired_page = false;
5748 :
5749 184 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5750 :
5751 : /*
5752 : * If we didn't pin the visibility map page and the page has become
5753 : * all visible while we were busy locking the buffer, we'll have to
5754 : * unlock and re-lock, to avoid holding the buffer lock across I/O.
5755 : * That's a bit unfortunate, but hopefully shouldn't happen often.
5756 : *
5757 : * Note: in some paths through this function, we will reach here
5758 : * holding a pin on a vm page that may or may not be the one matching
5759 : * this page. If this page isn't all-visible, we won't use the vm
5760 : * page, but we hold onto such a pin till the end of the function.
5761 : */
5762 184 : if (!pinned_desired_page && PageIsAllVisible(BufferGetPage(buf)))
5763 : {
5764 0 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5765 0 : visibilitymap_pin(rel, block, &vmbuffer);
5766 0 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5767 : }
5768 :
5769 : /*
5770 : * Check the tuple XMIN against prior XMAX, if any. If we reached the
5771 : * end of the chain, we're done, so return success.
5772 : */
5773 190 : if (TransactionIdIsValid(priorXmax) &&
5774 6 : !TransactionIdEquals(HeapTupleHeaderGetXmin(mytup.t_data),
5775 : priorXmax))
5776 : {
5777 0 : result = TM_Ok;
5778 0 : goto out_locked;
5779 : }
5780 :
5781 : /*
5782 : * Also check Xmin: if this tuple was created by an aborted
5783 : * (sub)transaction, then we already locked the last live one in the
5784 : * chain, thus we're done, so return success.
5785 : */
5786 184 : if (TransactionIdDidAbort(HeapTupleHeaderGetXmin(mytup.t_data)))
5787 : {
5788 26 : result = TM_Ok;
5789 26 : goto out_locked;
5790 : }
5791 :
5792 158 : old_infomask = mytup.t_data->t_infomask;
5793 158 : old_infomask2 = mytup.t_data->t_infomask2;
5794 158 : xmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5795 :
5796 : /*
5797 : * If this tuple version has been updated or locked by some concurrent
5798 : * transaction(s), what we do depends on whether our lock mode
5799 : * conflicts with what those other transactions hold, and also on the
5800 : * status of them.
5801 : */
5802 158 : if (!(old_infomask & HEAP_XMAX_INVALID))
5803 : {
5804 : TransactionId rawxmax;
5805 : bool needwait;
5806 :
5807 60 : rawxmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5808 60 : if (old_infomask & HEAP_XMAX_IS_MULTI)
5809 : {
5810 : int nmembers;
5811 : int i;
5812 : MultiXactMember *members;
5813 :
5814 : /*
5815 : * We don't need a test for pg_upgrade'd tuples: this is only
5816 : * applied to tuples after the first in an update chain. Said
5817 : * first tuple in the chain may well be locked-in-9.2-and-
5818 : * pg_upgraded, but that one was already locked by our caller,
5819 : * not us; and any subsequent ones cannot be because our
5820 : * caller must necessarily have obtained a snapshot later than
5821 : * the pg_upgrade itself.
5822 : */
5823 : Assert(!HEAP_LOCKED_UPGRADED(mytup.t_data->t_infomask));
5824 :
5825 2 : nmembers = GetMultiXactIdMembers(rawxmax, &members, false,
5826 2 : HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5827 8 : for (i = 0; i < nmembers; i++)
5828 : {
5829 6 : result = test_lockmode_for_conflict(members[i].status,
5830 6 : members[i].xid,
5831 : mode,
5832 : &mytup,
5833 : &needwait);
5834 :
5835 : /*
5836 : * If the tuple was already locked by ourselves in a
5837 : * previous iteration of this (say heap_lock_tuple was
5838 : * forced to restart the locking loop because of a change
5839 : * in xmax), then we hold the lock already on this tuple
5840 : * version and we don't need to do anything; and this is
5841 : * not an error condition either. We just need to skip
5842 : * this tuple and continue locking the next version in the
5843 : * update chain.
5844 : */
5845 6 : if (result == TM_SelfModified)
5846 : {
5847 0 : pfree(members);
5848 0 : goto next;
5849 : }
5850 :
5851 6 : if (needwait)
5852 : {
5853 0 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5854 0 : XactLockTableWait(members[i].xid, rel,
5855 : &mytup.t_self,
5856 : XLTW_LockUpdated);
5857 0 : pfree(members);
5858 0 : goto l4;
5859 : }
5860 6 : if (result != TM_Ok)
5861 : {
5862 0 : pfree(members);
5863 0 : goto out_locked;
5864 : }
5865 : }
5866 2 : if (members)
5867 2 : pfree(members);
5868 : }
5869 : else
5870 : {
5871 : MultiXactStatus status;
5872 :
5873 : /*
5874 : * For a non-multi Xmax, we first need to compute the
5875 : * corresponding MultiXactStatus by using the infomask bits.
5876 : */
5877 58 : if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5878 : {
5879 20 : if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5880 20 : status = MultiXactStatusForKeyShare;
5881 0 : else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5882 0 : status = MultiXactStatusForShare;
5883 0 : else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5884 : {
5885 0 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5886 0 : status = MultiXactStatusForUpdate;
5887 : else
5888 0 : status = MultiXactStatusForNoKeyUpdate;
5889 : }
5890 : else
5891 : {
5892 : /*
5893 : * LOCK_ONLY present alone (a pg_upgraded tuple marked
5894 : * as share-locked in the old cluster) shouldn't be
5895 : * seen in the middle of an update chain.
5896 : */
5897 0 : elog(ERROR, "invalid lock status in tuple");
5898 : }
5899 : }
5900 : else
5901 : {
5902 : /* it's an update, but which kind? */
5903 38 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5904 28 : status = MultiXactStatusUpdate;
5905 : else
5906 10 : status = MultiXactStatusNoKeyUpdate;
5907 : }
5908 :
5909 58 : result = test_lockmode_for_conflict(status, rawxmax, mode,
5910 : &mytup, &needwait);
5911 :
5912 : /*
5913 : * If the tuple was already locked by ourselves in a previous
5914 : * iteration of this (say heap_lock_tuple was forced to
5915 : * restart the locking loop because of a change in xmax), then
5916 : * we hold the lock already on this tuple version and we don't
5917 : * need to do anything; and this is not an error condition
5918 : * either. We just need to skip this tuple and continue
5919 : * locking the next version in the update chain.
5920 : */
5921 58 : if (result == TM_SelfModified)
5922 0 : goto next;
5923 :
5924 58 : if (needwait)
5925 : {
5926 16 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5927 16 : XactLockTableWait(rawxmax, rel, &mytup.t_self,
5928 : XLTW_LockUpdated);
5929 16 : goto l4;
5930 : }
5931 42 : if (result != TM_Ok)
5932 : {
5933 16 : goto out_locked;
5934 : }
5935 : }
5936 : }
5937 :
5938 : /* compute the new Xmax and infomask values for the tuple ... */
5939 126 : compute_new_xmax_infomask(xmax, old_infomask, mytup.t_data->t_infomask2,
5940 : xid, mode, false,
5941 : &new_xmax, &new_infomask, &new_infomask2);
5942 :
5943 126 : if (PageIsAllVisible(BufferGetPage(buf)) &&
5944 0 : visibilitymap_clear(rel, block, vmbuffer,
5945 : VISIBILITYMAP_ALL_FROZEN))
5946 0 : cleared_all_frozen = true;
5947 :
5948 126 : START_CRIT_SECTION();
5949 :
5950 : /* ... and set them */
5951 126 : HeapTupleHeaderSetXmax(mytup.t_data, new_xmax);
5952 126 : mytup.t_data->t_infomask &= ~HEAP_XMAX_BITS;
5953 126 : mytup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5954 126 : mytup.t_data->t_infomask |= new_infomask;
5955 126 : mytup.t_data->t_infomask2 |= new_infomask2;
5956 :
5957 126 : MarkBufferDirty(buf);
5958 :
5959 : /* XLOG stuff */
5960 126 : if (RelationNeedsWAL(rel))
5961 : {
5962 : xl_heap_lock_updated xlrec;
5963 : XLogRecPtr recptr;
5964 126 : Page page = BufferGetPage(buf);
5965 :
5966 126 : XLogBeginInsert();
5967 126 : XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
5968 :
5969 126 : xlrec.offnum = ItemPointerGetOffsetNumber(&mytup.t_self);
5970 126 : xlrec.xmax = new_xmax;
5971 126 : xlrec.infobits_set = compute_infobits(new_infomask, new_infomask2);
5972 126 : xlrec.flags =
5973 126 : cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
5974 :
5975 126 : XLogRegisterData(&xlrec, SizeOfHeapLockUpdated);
5976 :
5977 126 : recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_LOCK_UPDATED);
5978 :
5979 126 : PageSetLSN(page, recptr);
5980 : }
5981 :
5982 126 : END_CRIT_SECTION();
5983 :
5984 126 : next:
5985 : /* if we find the end of update chain, we're done. */
5986 252 : if (mytup.t_data->t_infomask & HEAP_XMAX_INVALID ||
5987 252 : HeapTupleHeaderIndicatesMovedPartitions(mytup.t_data) ||
5988 134 : ItemPointerEquals(&mytup.t_self, &mytup.t_data->t_ctid) ||
5989 8 : HeapTupleHeaderIsOnlyLocked(mytup.t_data))
5990 : {
5991 120 : result = TM_Ok;
5992 120 : goto out_locked;
5993 : }
5994 :
5995 : /* tail recursion */
5996 6 : priorXmax = HeapTupleHeaderGetUpdateXid(mytup.t_data);
5997 6 : ItemPointerCopy(&(mytup.t_data->t_ctid), &tupid);
5998 6 : UnlockReleaseBuffer(buf);
5999 : }
6000 :
6001 : result = TM_Ok;
6002 :
6003 162 : out_locked:
6004 162 : UnlockReleaseBuffer(buf);
6005 :
6006 162 : out_unlocked:
6007 162 : if (vmbuffer != InvalidBuffer)
6008 0 : ReleaseBuffer(vmbuffer);
6009 :
6010 162 : return result;
6011 : }
6012 :
6013 : /*
6014 : * heap_lock_updated_tuple
6015 : * Follow update chain when locking an updated tuple, acquiring locks (row
6016 : * marks) on the updated versions.
6017 : *
6018 : * The initial tuple is assumed to be already locked.
6019 : *
6020 : * This function doesn't check visibility, it just unconditionally marks the
6021 : * tuple(s) as locked. If any tuple in the updated chain is being deleted
6022 : * concurrently (or updated with the key being modified), sleep until the
6023 : * transaction doing it is finished.
6024 : *
6025 : * Note that we don't acquire heavyweight tuple locks on the tuples we walk
6026 : * when we have to wait for other transactions to release them, as opposed to
6027 : * what heap_lock_tuple does. The reason is that having more than one
6028 : * transaction walking the chain is probably uncommon enough that risk of
6029 : * starvation is not likely: one of the preconditions for being here is that
6030 : * the snapshot in use predates the update that created this tuple (because we
6031 : * started at an earlier version of the tuple), but at the same time such a
6032 : * transaction cannot be using repeatable read or serializable isolation
6033 : * levels, because that would lead to a serializability failure.
6034 : */
6035 : static TM_Result
6036 180 : heap_lock_updated_tuple(Relation rel, HeapTuple tuple, ItemPointer ctid,
6037 : TransactionId xid, LockTupleMode mode)
6038 : {
6039 : /*
6040 : * If the tuple has not been updated, or has moved into another partition
6041 : * (effectively a delete) stop here.
6042 : */
6043 180 : if (!HeapTupleHeaderIndicatesMovedPartitions(tuple->t_data) &&
6044 176 : !ItemPointerEquals(&tuple->t_self, ctid))
6045 : {
6046 : /*
6047 : * If this is the first possibly-multixact-able operation in the
6048 : * current transaction, set my per-backend OldestMemberMXactId
6049 : * setting. We can be certain that the transaction will never become a
6050 : * member of any older MultiXactIds than that. (We have to do this
6051 : * even if we end up just using our own TransactionId below, since
6052 : * some other backend could incorporate our XID into a MultiXact
6053 : * immediately afterwards.)
6054 : */
6055 162 : MultiXactIdSetOldestMember();
6056 :
6057 162 : return heap_lock_updated_tuple_rec(rel, ctid, xid, mode);
6058 : }
6059 :
6060 : /* nothing to lock */
6061 18 : return TM_Ok;
6062 : }
6063 :
6064 : /*
6065 : * heap_finish_speculative - mark speculative insertion as successful
6066 : *
6067 : * To successfully finish a speculative insertion we have to clear speculative
6068 : * token from tuple. To do so the t_ctid field, which will contain a
6069 : * speculative token value, is modified in place to point to the tuple itself,
6070 : * which is characteristic of a newly inserted ordinary tuple.
6071 : *
6072 : * NB: It is not ok to commit without either finishing or aborting a
6073 : * speculative insertion. We could treat speculative tuples of committed
6074 : * transactions implicitly as completed, but then we would have to be prepared
6075 : * to deal with speculative tokens on committed tuples. That wouldn't be
6076 : * difficult - no-one looks at the ctid field of a tuple with invalid xmax -
6077 : * but clearing the token at completion isn't very expensive either.
6078 : * An explicit confirmation WAL record also makes logical decoding simpler.
6079 : */
6080 : void
6081 4112 : heap_finish_speculative(Relation relation, ItemPointer tid)
6082 : {
6083 : Buffer buffer;
6084 : Page page;
6085 : OffsetNumber offnum;
6086 4112 : ItemId lp = NULL;
6087 : HeapTupleHeader htup;
6088 :
6089 4112 : buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
6090 4112 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6091 4112 : page = (Page) BufferGetPage(buffer);
6092 :
6093 4112 : offnum = ItemPointerGetOffsetNumber(tid);
6094 4112 : if (PageGetMaxOffsetNumber(page) >= offnum)
6095 4112 : lp = PageGetItemId(page, offnum);
6096 :
6097 4112 : if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
6098 0 : elog(ERROR, "invalid lp");
6099 :
6100 4112 : htup = (HeapTupleHeader) PageGetItem(page, lp);
6101 :
6102 : /* NO EREPORT(ERROR) from here till changes are logged */
6103 4112 : START_CRIT_SECTION();
6104 :
6105 : Assert(HeapTupleHeaderIsSpeculative(htup));
6106 :
6107 4112 : MarkBufferDirty(buffer);
6108 :
6109 : /*
6110 : * Replace the speculative insertion token with a real t_ctid, pointing to
6111 : * itself like it does on regular tuples.
6112 : */
6113 4112 : htup->t_ctid = *tid;
6114 :
6115 : /* XLOG stuff */
6116 4112 : if (RelationNeedsWAL(relation))
6117 : {
6118 : xl_heap_confirm xlrec;
6119 : XLogRecPtr recptr;
6120 :
6121 4094 : xlrec.offnum = ItemPointerGetOffsetNumber(tid);
6122 :
6123 4094 : XLogBeginInsert();
6124 :
6125 : /* We want the same filtering on this as on a plain insert */
6126 4094 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
6127 :
6128 4094 : XLogRegisterData(&xlrec, SizeOfHeapConfirm);
6129 4094 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
6130 :
6131 4094 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_CONFIRM);
6132 :
6133 4094 : PageSetLSN(page, recptr);
6134 : }
6135 :
6136 4112 : END_CRIT_SECTION();
6137 :
6138 4112 : UnlockReleaseBuffer(buffer);
6139 4112 : }
6140 :
6141 : /*
6142 : * heap_abort_speculative - kill a speculatively inserted tuple
6143 : *
6144 : * Marks a tuple that was speculatively inserted in the same command as dead,
6145 : * by setting its xmin as invalid. That makes it immediately appear as dead
6146 : * to all transactions, including our own. In particular, it makes
6147 : * HeapTupleSatisfiesDirty() regard the tuple as dead, so that another backend
6148 : * inserting a duplicate key value won't unnecessarily wait for our whole
6149 : * transaction to finish (it'll just wait for our speculative insertion to
6150 : * finish).
6151 : *
6152 : * Killing the tuple prevents "unprincipled deadlocks", which are deadlocks
6153 : * that arise due to a mutual dependency that is not user visible. By
6154 : * definition, unprincipled deadlocks cannot be prevented by the user
6155 : * reordering lock acquisition in client code, because the implementation level
6156 : * lock acquisitions are not under the user's direct control. If speculative
6157 : * inserters did not take this precaution, then under high concurrency they
6158 : * could deadlock with each other, which would not be acceptable.
6159 : *
6160 : * This is somewhat redundant with heap_delete, but we prefer to have a
6161 : * dedicated routine with stripped down requirements. Note that this is also
6162 : * used to delete the TOAST tuples created during speculative insertion.
6163 : *
6164 : * This routine does not affect logical decoding as it only looks at
6165 : * confirmation records.
6166 : */
6167 : void
6168 20 : heap_abort_speculative(Relation relation, ItemPointer tid)
6169 : {
6170 20 : TransactionId xid = GetCurrentTransactionId();
6171 : ItemId lp;
6172 : HeapTupleData tp;
6173 : Page page;
6174 : BlockNumber block;
6175 : Buffer buffer;
6176 :
6177 : Assert(ItemPointerIsValid(tid));
6178 :
6179 20 : block = ItemPointerGetBlockNumber(tid);
6180 20 : buffer = ReadBuffer(relation, block);
6181 20 : page = BufferGetPage(buffer);
6182 :
6183 20 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6184 :
6185 : /*
6186 : * Page can't be all visible, we just inserted into it, and are still
6187 : * running.
6188 : */
6189 : Assert(!PageIsAllVisible(page));
6190 :
6191 20 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
6192 : Assert(ItemIdIsNormal(lp));
6193 :
6194 20 : tp.t_tableOid = RelationGetRelid(relation);
6195 20 : tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
6196 20 : tp.t_len = ItemIdGetLength(lp);
6197 20 : tp.t_self = *tid;
6198 :
6199 : /*
6200 : * Sanity check that the tuple really is a speculatively inserted tuple,
6201 : * inserted by us.
6202 : */
6203 20 : if (tp.t_data->t_choice.t_heap.t_xmin != xid)
6204 0 : elog(ERROR, "attempted to kill a tuple inserted by another transaction");
6205 20 : if (!(IsToastRelation(relation) || HeapTupleHeaderIsSpeculative(tp.t_data)))
6206 0 : elog(ERROR, "attempted to kill a non-speculative tuple");
6207 : Assert(!HeapTupleHeaderIsHeapOnly(tp.t_data));
6208 :
6209 : /*
6210 : * No need to check for serializable conflicts here. There is never a
6211 : * need for a combo CID, either. No need to extract replica identity, or
6212 : * do anything special with infomask bits.
6213 : */
6214 :
6215 20 : START_CRIT_SECTION();
6216 :
6217 : /*
6218 : * The tuple will become DEAD immediately. Flag that this page is a
6219 : * candidate for pruning by setting xmin to TransactionXmin. While not
6220 : * immediately prunable, it is the oldest xid we can cheaply determine
6221 : * that's safe against wraparound / being older than the table's
6222 : * relfrozenxid. To defend against the unlikely case of a new relation
6223 : * having a newer relfrozenxid than our TransactionXmin, use relfrozenxid
6224 : * if so (vacuum can't subsequently move relfrozenxid to beyond
6225 : * TransactionXmin, so there's no race here).
6226 : */
6227 : Assert(TransactionIdIsValid(TransactionXmin));
6228 : {
6229 20 : TransactionId relfrozenxid = relation->rd_rel->relfrozenxid;
6230 : TransactionId prune_xid;
6231 :
6232 20 : if (TransactionIdPrecedes(TransactionXmin, relfrozenxid))
6233 0 : prune_xid = relfrozenxid;
6234 : else
6235 20 : prune_xid = TransactionXmin;
6236 20 : PageSetPrunable(page, prune_xid);
6237 : }
6238 :
6239 : /* store transaction information of xact deleting the tuple */
6240 20 : tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
6241 20 : tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
6242 :
6243 : /*
6244 : * Set the tuple header xmin to InvalidTransactionId. This makes the
6245 : * tuple immediately invisible everyone. (In particular, to any
6246 : * transactions waiting on the speculative token, woken up later.)
6247 : */
6248 20 : HeapTupleHeaderSetXmin(tp.t_data, InvalidTransactionId);
6249 :
6250 : /* Clear the speculative insertion token too */
6251 20 : tp.t_data->t_ctid = tp.t_self;
6252 :
6253 20 : MarkBufferDirty(buffer);
6254 :
6255 : /*
6256 : * XLOG stuff
6257 : *
6258 : * The WAL records generated here match heap_delete(). The same recovery
6259 : * routines are used.
6260 : */
6261 20 : if (RelationNeedsWAL(relation))
6262 : {
6263 : xl_heap_delete xlrec;
6264 : XLogRecPtr recptr;
6265 :
6266 20 : xlrec.flags = XLH_DELETE_IS_SUPER;
6267 40 : xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
6268 20 : tp.t_data->t_infomask2);
6269 20 : xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
6270 20 : xlrec.xmax = xid;
6271 :
6272 20 : XLogBeginInsert();
6273 20 : XLogRegisterData(&xlrec, SizeOfHeapDelete);
6274 20 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
6275 :
6276 : /* No replica identity & replication origin logged */
6277 :
6278 20 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
6279 :
6280 20 : PageSetLSN(page, recptr);
6281 : }
6282 :
6283 20 : END_CRIT_SECTION();
6284 :
6285 20 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6286 :
6287 20 : if (HeapTupleHasExternal(&tp))
6288 : {
6289 : Assert(!IsToastRelation(relation));
6290 2 : heap_toast_delete(relation, &tp, true);
6291 : }
6292 :
6293 : /*
6294 : * Never need to mark tuple for invalidation, since catalogs don't support
6295 : * speculative insertion
6296 : */
6297 :
6298 : /* Now we can release the buffer */
6299 20 : ReleaseBuffer(buffer);
6300 :
6301 : /* count deletion, as we counted the insertion too */
6302 20 : pgstat_count_heap_delete(relation);
6303 20 : }
6304 :
6305 : /*
6306 : * heap_inplace_lock - protect inplace update from concurrent heap_update()
6307 : *
6308 : * Evaluate whether the tuple's state is compatible with a no-key update.
6309 : * Current transaction rowmarks are fine, as is KEY SHARE from any
6310 : * transaction. If compatible, return true with the buffer exclusive-locked,
6311 : * and the caller must release that by calling
6312 : * heap_inplace_update_and_unlock(), calling heap_inplace_unlock(), or raising
6313 : * an error. Otherwise, call release_callback(arg), wait for blocking
6314 : * transactions to end, and return false.
6315 : *
6316 : * Since this is intended for system catalogs and SERIALIZABLE doesn't cover
6317 : * DDL, this doesn't guarantee any particular predicate locking.
6318 : *
6319 : * One could modify this to return true for tuples with delete in progress,
6320 : * All inplace updaters take a lock that conflicts with DROP. If explicit
6321 : * "DELETE FROM pg_class" is in progress, we'll wait for it like we would an
6322 : * update.
6323 : *
6324 : * Readers of inplace-updated fields expect changes to those fields are
6325 : * durable. For example, vac_truncate_clog() reads datfrozenxid from
6326 : * pg_database tuples via catalog snapshots. A future snapshot must not
6327 : * return a lower datfrozenxid for the same database OID (lower in the
6328 : * FullTransactionIdPrecedes() sense). We achieve that since no update of a
6329 : * tuple can start while we hold a lock on its buffer. In cases like
6330 : * BEGIN;GRANT;CREATE INDEX;COMMIT we're inplace-updating a tuple visible only
6331 : * to this transaction. ROLLBACK then is one case where it's okay to lose
6332 : * inplace updates. (Restoring relhasindex=false on ROLLBACK is fine, since
6333 : * any concurrent CREATE INDEX would have blocked, then inplace-updated the
6334 : * committed tuple.)
6335 : *
6336 : * In principle, we could avoid waiting by overwriting every tuple in the
6337 : * updated tuple chain. Reader expectations permit updating a tuple only if
6338 : * it's aborted, is the tail of the chain, or we already updated the tuple
6339 : * referenced in its t_ctid. Hence, we would need to overwrite the tuples in
6340 : * order from tail to head. That would imply either (a) mutating all tuples
6341 : * in one critical section or (b) accepting a chance of partial completion.
6342 : * Partial completion of a relfrozenxid update would have the weird
6343 : * consequence that the table's next VACUUM could see the table's relfrozenxid
6344 : * move forward between vacuum_get_cutoffs() and finishing.
6345 : */
6346 : bool
6347 368548 : heap_inplace_lock(Relation relation,
6348 : HeapTuple oldtup_ptr, Buffer buffer,
6349 : void (*release_callback) (void *), void *arg)
6350 : {
6351 368548 : HeapTupleData oldtup = *oldtup_ptr; /* minimize diff vs. heap_update() */
6352 : TM_Result result;
6353 : bool ret;
6354 :
6355 : #ifdef USE_ASSERT_CHECKING
6356 : if (RelationGetRelid(relation) == RelationRelationId)
6357 : check_inplace_rel_lock(oldtup_ptr);
6358 : #endif
6359 :
6360 : Assert(BufferIsValid(buffer));
6361 :
6362 : /*
6363 : * Construct shared cache inval if necessary. Because we pass a tuple
6364 : * version without our own inplace changes or inplace changes other
6365 : * sessions complete while we wait for locks, inplace update mustn't
6366 : * change catcache lookup keys. But we aren't bothering with index
6367 : * updates either, so that's true a fortiori. After LockBuffer(), it
6368 : * would be too late, because this might reach a
6369 : * CatalogCacheInitializeCache() that locks "buffer".
6370 : */
6371 368548 : CacheInvalidateHeapTupleInplace(relation, oldtup_ptr, NULL);
6372 :
6373 368548 : LockTuple(relation, &oldtup.t_self, InplaceUpdateTupleLock);
6374 368548 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6375 :
6376 : /*----------
6377 : * Interpret HeapTupleSatisfiesUpdate() like heap_update() does, except:
6378 : *
6379 : * - wait unconditionally
6380 : * - already locked tuple above, since inplace needs that unconditionally
6381 : * - don't recheck header after wait: simpler to defer to next iteration
6382 : * - don't try to continue even if the updater aborts: likewise
6383 : * - no crosscheck
6384 : */
6385 368548 : result = HeapTupleSatisfiesUpdate(&oldtup, GetCurrentCommandId(false),
6386 : buffer);
6387 :
6388 368548 : if (result == TM_Invisible)
6389 : {
6390 : /* no known way this can happen */
6391 0 : ereport(ERROR,
6392 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6393 : errmsg_internal("attempted to overwrite invisible tuple")));
6394 : }
6395 368548 : else if (result == TM_SelfModified)
6396 : {
6397 : /*
6398 : * CREATE INDEX might reach this if an expression is silly enough to
6399 : * call e.g. SELECT ... FROM pg_class FOR SHARE. C code of other SQL
6400 : * statements might get here after a heap_update() of the same row, in
6401 : * the absence of an intervening CommandCounterIncrement().
6402 : */
6403 0 : ereport(ERROR,
6404 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6405 : errmsg("tuple to be updated was already modified by an operation triggered by the current command")));
6406 : }
6407 368548 : else if (result == TM_BeingModified)
6408 : {
6409 : TransactionId xwait;
6410 : uint16 infomask;
6411 :
6412 204 : xwait = HeapTupleHeaderGetRawXmax(oldtup.t_data);
6413 204 : infomask = oldtup.t_data->t_infomask;
6414 :
6415 204 : if (infomask & HEAP_XMAX_IS_MULTI)
6416 : {
6417 10 : LockTupleMode lockmode = LockTupleNoKeyExclusive;
6418 10 : MultiXactStatus mxact_status = MultiXactStatusNoKeyUpdate;
6419 : int remain;
6420 :
6421 10 : if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
6422 : lockmode, NULL))
6423 : {
6424 4 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6425 4 : release_callback(arg);
6426 4 : ret = false;
6427 4 : MultiXactIdWait((MultiXactId) xwait, mxact_status, infomask,
6428 : relation, &oldtup.t_self, XLTW_Update,
6429 : &remain);
6430 : }
6431 : else
6432 6 : ret = true;
6433 : }
6434 194 : else if (TransactionIdIsCurrentTransactionId(xwait))
6435 2 : ret = true;
6436 192 : else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask))
6437 2 : ret = true;
6438 : else
6439 : {
6440 190 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6441 190 : release_callback(arg);
6442 190 : ret = false;
6443 190 : XactLockTableWait(xwait, relation, &oldtup.t_self,
6444 : XLTW_Update);
6445 : }
6446 : }
6447 : else
6448 : {
6449 368344 : ret = (result == TM_Ok);
6450 368344 : if (!ret)
6451 : {
6452 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6453 0 : release_callback(arg);
6454 : }
6455 : }
6456 :
6457 : /*
6458 : * GetCatalogSnapshot() relies on invalidation messages to know when to
6459 : * take a new snapshot. COMMIT of xwait is responsible for sending the
6460 : * invalidation. We're not acquiring heavyweight locks sufficient to
6461 : * block if not yet sent, so we must take a new snapshot to ensure a later
6462 : * attempt has a fair chance. While we don't need this if xwait aborted,
6463 : * don't bother optimizing that.
6464 : */
6465 368548 : if (!ret)
6466 : {
6467 194 : UnlockTuple(relation, &oldtup.t_self, InplaceUpdateTupleLock);
6468 194 : ForgetInplace_Inval();
6469 194 : InvalidateCatalogSnapshot();
6470 : }
6471 368548 : return ret;
6472 : }
6473 :
6474 : /*
6475 : * heap_inplace_update_and_unlock - core of systable_inplace_update_finish
6476 : *
6477 : * The tuple cannot change size, and therefore its header fields and null
6478 : * bitmap (if any) don't change either.
6479 : *
6480 : * Since we hold LOCKTAG_TUPLE, no updater has a local copy of this tuple.
6481 : */
6482 : void
6483 166490 : heap_inplace_update_and_unlock(Relation relation,
6484 : HeapTuple oldtup, HeapTuple tuple,
6485 : Buffer buffer)
6486 : {
6487 166490 : HeapTupleHeader htup = oldtup->t_data;
6488 : uint32 oldlen;
6489 : uint32 newlen;
6490 : char *dst;
6491 : char *src;
6492 166490 : int nmsgs = 0;
6493 166490 : SharedInvalidationMessage *invalMessages = NULL;
6494 166490 : bool RelcacheInitFileInval = false;
6495 :
6496 : Assert(ItemPointerEquals(&oldtup->t_self, &tuple->t_self));
6497 166490 : oldlen = oldtup->t_len - htup->t_hoff;
6498 166490 : newlen = tuple->t_len - tuple->t_data->t_hoff;
6499 166490 : if (oldlen != newlen || htup->t_hoff != tuple->t_data->t_hoff)
6500 0 : elog(ERROR, "wrong tuple length");
6501 :
6502 166490 : dst = (char *) htup + htup->t_hoff;
6503 166490 : src = (char *) tuple->t_data + tuple->t_data->t_hoff;
6504 :
6505 : /* Like RecordTransactionCommit(), log only if needed */
6506 166490 : if (XLogStandbyInfoActive())
6507 96710 : nmsgs = inplaceGetInvalidationMessages(&invalMessages,
6508 : &RelcacheInitFileInval);
6509 :
6510 : /*
6511 : * Unlink relcache init files as needed. If unlinking, acquire
6512 : * RelCacheInitLock until after associated invalidations. By doing this
6513 : * in advance, if we checkpoint and then crash between inplace
6514 : * XLogInsert() and inval, we don't rely on StartupXLOG() ->
6515 : * RelationCacheInitFileRemove(). That uses elevel==LOG, so replay would
6516 : * neglect to PANIC on EIO.
6517 : */
6518 166490 : PreInplace_Inval();
6519 :
6520 : /*----------
6521 : * NO EREPORT(ERROR) from here till changes are complete
6522 : *
6523 : * Our buffer lock won't stop a reader having already pinned and checked
6524 : * visibility for this tuple. Hence, we write WAL first, then mutate the
6525 : * buffer. Like in MarkBufferDirtyHint() or RecordTransactionCommit(),
6526 : * checkpoint delay makes that acceptable. With the usual order of
6527 : * changes, a crash after memcpy() and before XLogInsert() could allow
6528 : * datfrozenxid to overtake relfrozenxid:
6529 : *
6530 : * ["D" is a VACUUM (ONLY_DATABASE_STATS)]
6531 : * ["R" is a VACUUM tbl]
6532 : * D: vac_update_datfrozenxid() -> systable_beginscan(pg_class)
6533 : * D: systable_getnext() returns pg_class tuple of tbl
6534 : * R: memcpy() into pg_class tuple of tbl
6535 : * D: raise pg_database.datfrozenxid, XLogInsert(), finish
6536 : * [crash]
6537 : * [recovery restores datfrozenxid w/o relfrozenxid]
6538 : *
6539 : * Mimic MarkBufferDirtyHint() subroutine XLogSaveBufferForHint().
6540 : * Specifically, use DELAY_CHKPT_START, and copy the buffer to the stack.
6541 : * The stack copy facilitates a FPI of the post-mutation block before we
6542 : * accept other sessions seeing it. DELAY_CHKPT_START allows us to
6543 : * XLogInsert() before MarkBufferDirty(). Since XLogSaveBufferForHint()
6544 : * can operate under BUFFER_LOCK_SHARED, it can't avoid DELAY_CHKPT_START.
6545 : * This function, however, likely could avoid it with the following order
6546 : * of operations: MarkBufferDirty(), XLogInsert(), memcpy(). Opt to use
6547 : * DELAY_CHKPT_START here, too, as a way to have fewer distinct code
6548 : * patterns to analyze. Inplace update isn't so frequent that it should
6549 : * pursue the small optimization of skipping DELAY_CHKPT_START.
6550 : */
6551 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
6552 166490 : START_CRIT_SECTION();
6553 166490 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
6554 :
6555 : /* XLOG stuff */
6556 166490 : if (RelationNeedsWAL(relation))
6557 : {
6558 : xl_heap_inplace xlrec;
6559 : PGAlignedBlock copied_buffer;
6560 166474 : char *origdata = (char *) BufferGetBlock(buffer);
6561 166474 : Page page = BufferGetPage(buffer);
6562 166474 : uint16 lower = ((PageHeader) page)->pd_lower;
6563 166474 : uint16 upper = ((PageHeader) page)->pd_upper;
6564 : uintptr_t dst_offset_in_block;
6565 : RelFileLocator rlocator;
6566 : ForkNumber forkno;
6567 : BlockNumber blkno;
6568 : XLogRecPtr recptr;
6569 :
6570 166474 : xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
6571 166474 : xlrec.dbId = MyDatabaseId;
6572 166474 : xlrec.tsId = MyDatabaseTableSpace;
6573 166474 : xlrec.relcacheInitFileInval = RelcacheInitFileInval;
6574 166474 : xlrec.nmsgs = nmsgs;
6575 :
6576 166474 : XLogBeginInsert();
6577 166474 : XLogRegisterData(&xlrec, MinSizeOfHeapInplace);
6578 166474 : if (nmsgs != 0)
6579 67996 : XLogRegisterData(invalMessages,
6580 : nmsgs * sizeof(SharedInvalidationMessage));
6581 :
6582 : /* register block matching what buffer will look like after changes */
6583 166474 : memcpy(copied_buffer.data, origdata, lower);
6584 166474 : memcpy(copied_buffer.data + upper, origdata + upper, BLCKSZ - upper);
6585 166474 : dst_offset_in_block = dst - origdata;
6586 166474 : memcpy(copied_buffer.data + dst_offset_in_block, src, newlen);
6587 166474 : BufferGetTag(buffer, &rlocator, &forkno, &blkno);
6588 : Assert(forkno == MAIN_FORKNUM);
6589 166474 : XLogRegisterBlock(0, &rlocator, forkno, blkno, copied_buffer.data,
6590 : REGBUF_STANDARD);
6591 166474 : XLogRegisterBufData(0, src, newlen);
6592 :
6593 : /* inplace updates aren't decoded atm, don't log the origin */
6594 :
6595 166474 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_INPLACE);
6596 :
6597 166474 : PageSetLSN(page, recptr);
6598 : }
6599 :
6600 166490 : memcpy(dst, src, newlen);
6601 :
6602 166490 : MarkBufferDirty(buffer);
6603 :
6604 166490 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6605 :
6606 : /*
6607 : * Send invalidations to shared queue. SearchSysCacheLocked1() assumes we
6608 : * do this before UnlockTuple().
6609 : *
6610 : * If we're mutating a tuple visible only to this transaction, there's an
6611 : * equivalent transactional inval from the action that created the tuple,
6612 : * and this inval is superfluous.
6613 : */
6614 166490 : AtInplace_Inval();
6615 :
6616 166490 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
6617 166490 : END_CRIT_SECTION();
6618 166490 : UnlockTuple(relation, &tuple->t_self, InplaceUpdateTupleLock);
6619 :
6620 166490 : AcceptInvalidationMessages(); /* local processing of just-sent inval */
6621 :
6622 : /*
6623 : * Queue a transactional inval. The immediate invalidation we just sent
6624 : * is the only one known to be necessary. To reduce risk from the
6625 : * transition to immediate invalidation, continue sending a transactional
6626 : * invalidation like we've long done. Third-party code might rely on it.
6627 : */
6628 166490 : if (!IsBootstrapProcessingMode())
6629 137776 : CacheInvalidateHeapTuple(relation, tuple, NULL);
6630 166490 : }
6631 :
6632 : /*
6633 : * heap_inplace_unlock - reverse of heap_inplace_lock
6634 : */
6635 : void
6636 201864 : heap_inplace_unlock(Relation relation,
6637 : HeapTuple oldtup, Buffer buffer)
6638 : {
6639 201864 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6640 201864 : UnlockTuple(relation, &oldtup->t_self, InplaceUpdateTupleLock);
6641 201864 : ForgetInplace_Inval();
6642 201864 : }
6643 :
6644 : #define FRM_NOOP 0x0001
6645 : #define FRM_INVALIDATE_XMAX 0x0002
6646 : #define FRM_RETURN_IS_XID 0x0004
6647 : #define FRM_RETURN_IS_MULTI 0x0008
6648 : #define FRM_MARK_COMMITTED 0x0010
6649 :
6650 : /*
6651 : * FreezeMultiXactId
6652 : * Determine what to do during freezing when a tuple is marked by a
6653 : * MultiXactId.
6654 : *
6655 : * "flags" is an output value; it's used to tell caller what to do on return.
6656 : * "pagefrz" is an input/output value, used to manage page level freezing.
6657 : *
6658 : * Possible values that we can set in "flags":
6659 : * FRM_NOOP
6660 : * don't do anything -- keep existing Xmax
6661 : * FRM_INVALIDATE_XMAX
6662 : * mark Xmax as InvalidTransactionId and set XMAX_INVALID flag.
6663 : * FRM_RETURN_IS_XID
6664 : * The Xid return value is a single update Xid to set as xmax.
6665 : * FRM_MARK_COMMITTED
6666 : * Xmax can be marked as HEAP_XMAX_COMMITTED
6667 : * FRM_RETURN_IS_MULTI
6668 : * The return value is a new MultiXactId to set as new Xmax.
6669 : * (caller must obtain proper infomask bits using GetMultiXactIdHintBits)
6670 : *
6671 : * Caller delegates control of page freezing to us. In practice we always
6672 : * force freezing of caller's page unless FRM_NOOP processing is indicated.
6673 : * We help caller ensure that XIDs < FreezeLimit and MXIDs < MultiXactCutoff
6674 : * can never be left behind. We freely choose when and how to process each
6675 : * Multi, without ever violating the cutoff postconditions for freezing.
6676 : *
6677 : * It's useful to remove Multis on a proactive timeline (relative to freezing
6678 : * XIDs) to keep MultiXact member SLRU buffer misses to a minimum. It can also
6679 : * be cheaper in the short run, for us, since we too can avoid SLRU buffer
6680 : * misses through eager processing.
6681 : *
6682 : * NB: Creates a _new_ MultiXactId when FRM_RETURN_IS_MULTI is set, though only
6683 : * when FreezeLimit and/or MultiXactCutoff cutoffs leave us with no choice.
6684 : * This can usually be put off, which is usually enough to avoid it altogether.
6685 : * Allocating new multis during VACUUM should be avoided on general principle;
6686 : * only VACUUM can advance relminmxid, so allocating new Multis here comes with
6687 : * its own special risks.
6688 : *
6689 : * NB: Caller must maintain "no freeze" NewRelfrozenXid/NewRelminMxid trackers
6690 : * using heap_tuple_should_freeze when we haven't forced page-level freezing.
6691 : *
6692 : * NB: Caller should avoid needlessly calling heap_tuple_should_freeze when we
6693 : * have already forced page-level freezing, since that might incur the same
6694 : * SLRU buffer misses that we specifically intended to avoid by freezing.
6695 : */
6696 : static TransactionId
6697 12 : FreezeMultiXactId(MultiXactId multi, uint16 t_infomask,
6698 : const struct VacuumCutoffs *cutoffs, uint16 *flags,
6699 : HeapPageFreeze *pagefrz)
6700 : {
6701 : TransactionId newxmax;
6702 : MultiXactMember *members;
6703 : int nmembers;
6704 : bool need_replace;
6705 : int nnewmembers;
6706 : MultiXactMember *newmembers;
6707 : bool has_lockers;
6708 : TransactionId update_xid;
6709 : bool update_committed;
6710 : TransactionId FreezePageRelfrozenXid;
6711 :
6712 12 : *flags = 0;
6713 :
6714 : /* We should only be called in Multis */
6715 : Assert(t_infomask & HEAP_XMAX_IS_MULTI);
6716 :
6717 24 : if (!MultiXactIdIsValid(multi) ||
6718 12 : HEAP_LOCKED_UPGRADED(t_infomask))
6719 : {
6720 0 : *flags |= FRM_INVALIDATE_XMAX;
6721 0 : pagefrz->freeze_required = true;
6722 0 : return InvalidTransactionId;
6723 : }
6724 12 : else if (MultiXactIdPrecedes(multi, cutoffs->relminmxid))
6725 0 : ereport(ERROR,
6726 : (errcode(ERRCODE_DATA_CORRUPTED),
6727 : errmsg_internal("found multixact %u from before relminmxid %u",
6728 : multi, cutoffs->relminmxid)));
6729 12 : else if (MultiXactIdPrecedes(multi, cutoffs->OldestMxact))
6730 : {
6731 : TransactionId update_xact;
6732 :
6733 : /*
6734 : * This old multi cannot possibly have members still running, but
6735 : * verify just in case. If it was a locker only, it can be removed
6736 : * without any further consideration; but if it contained an update,
6737 : * we might need to preserve it.
6738 : */
6739 8 : if (MultiXactIdIsRunning(multi,
6740 8 : HEAP_XMAX_IS_LOCKED_ONLY(t_infomask)))
6741 0 : ereport(ERROR,
6742 : (errcode(ERRCODE_DATA_CORRUPTED),
6743 : errmsg_internal("multixact %u from before multi freeze cutoff %u found to be still running",
6744 : multi, cutoffs->OldestMxact)));
6745 :
6746 8 : if (HEAP_XMAX_IS_LOCKED_ONLY(t_infomask))
6747 : {
6748 8 : *flags |= FRM_INVALIDATE_XMAX;
6749 8 : pagefrz->freeze_required = true;
6750 8 : return InvalidTransactionId;
6751 : }
6752 :
6753 : /* replace multi with single XID for its updater? */
6754 0 : update_xact = MultiXactIdGetUpdateXid(multi, t_infomask);
6755 0 : if (TransactionIdPrecedes(update_xact, cutoffs->relfrozenxid))
6756 0 : ereport(ERROR,
6757 : (errcode(ERRCODE_DATA_CORRUPTED),
6758 : errmsg_internal("multixact %u contains update XID %u from before relfrozenxid %u",
6759 : multi, update_xact,
6760 : cutoffs->relfrozenxid)));
6761 0 : else if (TransactionIdPrecedes(update_xact, cutoffs->OldestXmin))
6762 : {
6763 : /*
6764 : * Updater XID has to have aborted (otherwise the tuple would have
6765 : * been pruned away instead, since updater XID is < OldestXmin).
6766 : * Just remove xmax.
6767 : */
6768 0 : if (TransactionIdDidCommit(update_xact))
6769 0 : ereport(ERROR,
6770 : (errcode(ERRCODE_DATA_CORRUPTED),
6771 : errmsg_internal("multixact %u contains committed update XID %u from before removable cutoff %u",
6772 : multi, update_xact,
6773 : cutoffs->OldestXmin)));
6774 0 : *flags |= FRM_INVALIDATE_XMAX;
6775 0 : pagefrz->freeze_required = true;
6776 0 : return InvalidTransactionId;
6777 : }
6778 :
6779 : /* Have to keep updater XID as new xmax */
6780 0 : *flags |= FRM_RETURN_IS_XID;
6781 0 : pagefrz->freeze_required = true;
6782 0 : return update_xact;
6783 : }
6784 :
6785 : /*
6786 : * Some member(s) of this Multi may be below FreezeLimit xid cutoff, so we
6787 : * need to walk the whole members array to figure out what to do, if
6788 : * anything.
6789 : */
6790 : nmembers =
6791 4 : GetMultiXactIdMembers(multi, &members, false,
6792 4 : HEAP_XMAX_IS_LOCKED_ONLY(t_infomask));
6793 4 : if (nmembers <= 0)
6794 : {
6795 : /* Nothing worth keeping */
6796 0 : *flags |= FRM_INVALIDATE_XMAX;
6797 0 : pagefrz->freeze_required = true;
6798 0 : return InvalidTransactionId;
6799 : }
6800 :
6801 : /*
6802 : * The FRM_NOOP case is the only case where we might need to ratchet back
6803 : * FreezePageRelfrozenXid or FreezePageRelminMxid. It is also the only
6804 : * case where our caller might ratchet back its NoFreezePageRelfrozenXid
6805 : * or NoFreezePageRelminMxid "no freeze" trackers to deal with a multi.
6806 : * FRM_NOOP handling should result in the NewRelfrozenXid/NewRelminMxid
6807 : * trackers managed by VACUUM being ratcheting back by xmax to the degree
6808 : * required to make it safe to leave xmax undisturbed, independent of
6809 : * whether or not page freezing is triggered somewhere else.
6810 : *
6811 : * Our policy is to force freezing in every case other than FRM_NOOP,
6812 : * which obviates the need to maintain either set of trackers, anywhere.
6813 : * Every other case will reliably execute a freeze plan for xmax that
6814 : * either replaces xmax with an XID/MXID >= OldestXmin/OldestMxact, or
6815 : * sets xmax to an InvalidTransactionId XID, rendering xmax fully frozen.
6816 : * (VACUUM's NewRelfrozenXid/NewRelminMxid trackers are initialized with
6817 : * OldestXmin/OldestMxact, so later values never need to be tracked here.)
6818 : */
6819 4 : need_replace = false;
6820 4 : FreezePageRelfrozenXid = pagefrz->FreezePageRelfrozenXid;
6821 8 : for (int i = 0; i < nmembers; i++)
6822 : {
6823 6 : TransactionId xid = members[i].xid;
6824 :
6825 : Assert(!TransactionIdPrecedes(xid, cutoffs->relfrozenxid));
6826 :
6827 6 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
6828 : {
6829 : /* Can't violate the FreezeLimit postcondition */
6830 2 : need_replace = true;
6831 2 : break;
6832 : }
6833 4 : if (TransactionIdPrecedes(xid, FreezePageRelfrozenXid))
6834 0 : FreezePageRelfrozenXid = xid;
6835 : }
6836 :
6837 : /* Can't violate the MultiXactCutoff postcondition, either */
6838 4 : if (!need_replace)
6839 2 : need_replace = MultiXactIdPrecedes(multi, cutoffs->MultiXactCutoff);
6840 :
6841 4 : if (!need_replace)
6842 : {
6843 : /*
6844 : * vacuumlazy.c might ratchet back NewRelminMxid, NewRelfrozenXid, or
6845 : * both together to make it safe to retain this particular multi after
6846 : * freezing its page
6847 : */
6848 2 : *flags |= FRM_NOOP;
6849 2 : pagefrz->FreezePageRelfrozenXid = FreezePageRelfrozenXid;
6850 2 : if (MultiXactIdPrecedes(multi, pagefrz->FreezePageRelminMxid))
6851 0 : pagefrz->FreezePageRelminMxid = multi;
6852 2 : pfree(members);
6853 2 : return multi;
6854 : }
6855 :
6856 : /*
6857 : * Do a more thorough second pass over the multi to figure out which
6858 : * member XIDs actually need to be kept. Checking the precise status of
6859 : * individual members might even show that we don't need to keep anything.
6860 : * That is quite possible even though the Multi must be >= OldestMxact,
6861 : * since our second pass only keeps member XIDs when it's truly necessary;
6862 : * even member XIDs >= OldestXmin often won't be kept by second pass.
6863 : */
6864 2 : nnewmembers = 0;
6865 2 : newmembers = palloc(sizeof(MultiXactMember) * nmembers);
6866 2 : has_lockers = false;
6867 2 : update_xid = InvalidTransactionId;
6868 2 : update_committed = false;
6869 :
6870 : /*
6871 : * Determine whether to keep each member xid, or to ignore it instead
6872 : */
6873 6 : for (int i = 0; i < nmembers; i++)
6874 : {
6875 4 : TransactionId xid = members[i].xid;
6876 4 : MultiXactStatus mstatus = members[i].status;
6877 :
6878 : Assert(!TransactionIdPrecedes(xid, cutoffs->relfrozenxid));
6879 :
6880 4 : if (!ISUPDATE_from_mxstatus(mstatus))
6881 : {
6882 : /*
6883 : * Locker XID (not updater XID). We only keep lockers that are
6884 : * still running.
6885 : */
6886 8 : if (TransactionIdIsCurrentTransactionId(xid) ||
6887 4 : TransactionIdIsInProgress(xid))
6888 : {
6889 2 : if (TransactionIdPrecedes(xid, cutoffs->OldestXmin))
6890 0 : ereport(ERROR,
6891 : (errcode(ERRCODE_DATA_CORRUPTED),
6892 : errmsg_internal("multixact %u contains running locker XID %u from before removable cutoff %u",
6893 : multi, xid,
6894 : cutoffs->OldestXmin)));
6895 2 : newmembers[nnewmembers++] = members[i];
6896 2 : has_lockers = true;
6897 : }
6898 :
6899 4 : continue;
6900 : }
6901 :
6902 : /*
6903 : * Updater XID (not locker XID). Should we keep it?
6904 : *
6905 : * Since the tuple wasn't totally removed when vacuum pruned, the
6906 : * update Xid cannot possibly be older than OldestXmin cutoff unless
6907 : * the updater XID aborted. If the updater transaction is known
6908 : * aborted or crashed then it's okay to ignore it, otherwise not.
6909 : *
6910 : * In any case the Multi should never contain two updaters, whatever
6911 : * their individual commit status. Check for that first, in passing.
6912 : */
6913 0 : if (TransactionIdIsValid(update_xid))
6914 0 : ereport(ERROR,
6915 : (errcode(ERRCODE_DATA_CORRUPTED),
6916 : errmsg_internal("multixact %u has two or more updating members",
6917 : multi),
6918 : errdetail_internal("First updater XID=%u second updater XID=%u.",
6919 : update_xid, xid)));
6920 :
6921 : /*
6922 : * As with all tuple visibility routines, it's critical to test
6923 : * TransactionIdIsInProgress before TransactionIdDidCommit, because of
6924 : * race conditions explained in detail in heapam_visibility.c.
6925 : */
6926 0 : if (TransactionIdIsCurrentTransactionId(xid) ||
6927 0 : TransactionIdIsInProgress(xid))
6928 0 : update_xid = xid;
6929 0 : else if (TransactionIdDidCommit(xid))
6930 : {
6931 : /*
6932 : * The transaction committed, so we can tell caller to set
6933 : * HEAP_XMAX_COMMITTED. (We can only do this because we know the
6934 : * transaction is not running.)
6935 : */
6936 0 : update_committed = true;
6937 0 : update_xid = xid;
6938 : }
6939 : else
6940 : {
6941 : /*
6942 : * Not in progress, not committed -- must be aborted or crashed;
6943 : * we can ignore it.
6944 : */
6945 0 : continue;
6946 : }
6947 :
6948 : /*
6949 : * We determined that updater must be kept -- add it to pending new
6950 : * members list
6951 : */
6952 0 : if (TransactionIdPrecedes(xid, cutoffs->OldestXmin))
6953 0 : ereport(ERROR,
6954 : (errcode(ERRCODE_DATA_CORRUPTED),
6955 : errmsg_internal("multixact %u contains committed update XID %u from before removable cutoff %u",
6956 : multi, xid, cutoffs->OldestXmin)));
6957 0 : newmembers[nnewmembers++] = members[i];
6958 : }
6959 :
6960 2 : pfree(members);
6961 :
6962 : /*
6963 : * Determine what to do with caller's multi based on information gathered
6964 : * during our second pass
6965 : */
6966 2 : if (nnewmembers == 0)
6967 : {
6968 : /* Nothing worth keeping */
6969 0 : *flags |= FRM_INVALIDATE_XMAX;
6970 0 : newxmax = InvalidTransactionId;
6971 : }
6972 2 : else if (TransactionIdIsValid(update_xid) && !has_lockers)
6973 : {
6974 : /*
6975 : * If there's a single member and it's an update, pass it back alone
6976 : * without creating a new Multi. (XXX we could do this when there's a
6977 : * single remaining locker, too, but that would complicate the API too
6978 : * much; moreover, the case with the single updater is more
6979 : * interesting, because those are longer-lived.)
6980 : */
6981 : Assert(nnewmembers == 1);
6982 0 : *flags |= FRM_RETURN_IS_XID;
6983 0 : if (update_committed)
6984 0 : *flags |= FRM_MARK_COMMITTED;
6985 0 : newxmax = update_xid;
6986 : }
6987 : else
6988 : {
6989 : /*
6990 : * Create a new multixact with the surviving members of the previous
6991 : * one, to set as new Xmax in the tuple
6992 : */
6993 2 : newxmax = MultiXactIdCreateFromMembers(nnewmembers, newmembers);
6994 2 : *flags |= FRM_RETURN_IS_MULTI;
6995 : }
6996 :
6997 2 : pfree(newmembers);
6998 :
6999 2 : pagefrz->freeze_required = true;
7000 2 : return newxmax;
7001 : }
7002 :
7003 : /*
7004 : * heap_prepare_freeze_tuple
7005 : *
7006 : * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
7007 : * are older than the OldestXmin and/or OldestMxact freeze cutoffs. If so,
7008 : * setup enough state (in the *frz output argument) to enable caller to
7009 : * process this tuple as part of freezing its page, and return true. Return
7010 : * false if nothing can be changed about the tuple right now.
7011 : *
7012 : * Also sets *totally_frozen to true if the tuple will be totally frozen once
7013 : * caller executes returned freeze plan (or if the tuple was already totally
7014 : * frozen by an earlier VACUUM). This indicates that there are no remaining
7015 : * XIDs or MultiXactIds that will need to be processed by a future VACUUM.
7016 : *
7017 : * VACUUM caller must assemble HeapTupleFreeze freeze plan entries for every
7018 : * tuple that we returned true for, and then execute freezing. Caller must
7019 : * initialize pagefrz fields for page as a whole before first call here for
7020 : * each heap page.
7021 : *
7022 : * VACUUM caller decides on whether or not to freeze the page as a whole.
7023 : * We'll often prepare freeze plans for a page that caller just discards.
7024 : * However, VACUUM doesn't always get to make a choice; it must freeze when
7025 : * pagefrz.freeze_required is set, to ensure that any XIDs < FreezeLimit (and
7026 : * MXIDs < MultiXactCutoff) can never be left behind. We help to make sure
7027 : * that VACUUM always follows that rule.
7028 : *
7029 : * We sometimes force freezing of xmax MultiXactId values long before it is
7030 : * strictly necessary to do so just to ensure the FreezeLimit postcondition.
7031 : * It's worth processing MultiXactIds proactively when it is cheap to do so,
7032 : * and it's convenient to make that happen by piggy-backing it on the "force
7033 : * freezing" mechanism. Conversely, we sometimes delay freezing MultiXactIds
7034 : * because it is expensive right now (though only when it's still possible to
7035 : * do so without violating the FreezeLimit/MultiXactCutoff postcondition).
7036 : *
7037 : * It is assumed that the caller has checked the tuple with
7038 : * HeapTupleSatisfiesVacuum() and determined that it is not HEAPTUPLE_DEAD
7039 : * (else we should be removing the tuple, not freezing it).
7040 : *
7041 : * NB: This function has side effects: it might allocate a new MultiXactId.
7042 : * It will be set as tuple's new xmax when our *frz output is processed within
7043 : * heap_execute_freeze_tuple later on. If the tuple is in a shared buffer
7044 : * then caller had better have an exclusive lock on it already.
7045 : */
7046 : bool
7047 35521072 : heap_prepare_freeze_tuple(HeapTupleHeader tuple,
7048 : const struct VacuumCutoffs *cutoffs,
7049 : HeapPageFreeze *pagefrz,
7050 : HeapTupleFreeze *frz, bool *totally_frozen)
7051 : {
7052 35521072 : bool xmin_already_frozen = false,
7053 35521072 : xmax_already_frozen = false;
7054 35521072 : bool freeze_xmin = false,
7055 35521072 : replace_xvac = false,
7056 35521072 : replace_xmax = false,
7057 35521072 : freeze_xmax = false;
7058 : TransactionId xid;
7059 :
7060 35521072 : frz->xmax = HeapTupleHeaderGetRawXmax(tuple);
7061 35521072 : frz->t_infomask2 = tuple->t_infomask2;
7062 35521072 : frz->t_infomask = tuple->t_infomask;
7063 35521072 : frz->frzflags = 0;
7064 35521072 : frz->checkflags = 0;
7065 :
7066 : /*
7067 : * Process xmin, while keeping track of whether it's already frozen, or
7068 : * will become frozen iff our freeze plan is executed by caller (could be
7069 : * neither).
7070 : */
7071 35521072 : xid = HeapTupleHeaderGetXmin(tuple);
7072 35521072 : if (!TransactionIdIsNormal(xid))
7073 29628928 : xmin_already_frozen = true;
7074 : else
7075 : {
7076 5892144 : if (TransactionIdPrecedes(xid, cutoffs->relfrozenxid))
7077 0 : ereport(ERROR,
7078 : (errcode(ERRCODE_DATA_CORRUPTED),
7079 : errmsg_internal("found xmin %u from before relfrozenxid %u",
7080 : xid, cutoffs->relfrozenxid)));
7081 :
7082 : /* Will set freeze_xmin flags in freeze plan below */
7083 5892144 : freeze_xmin = TransactionIdPrecedes(xid, cutoffs->OldestXmin);
7084 :
7085 : /* Verify that xmin committed if and when freeze plan is executed */
7086 5892144 : if (freeze_xmin)
7087 4613974 : frz->checkflags |= HEAP_FREEZE_CHECK_XMIN_COMMITTED;
7088 : }
7089 :
7090 : /*
7091 : * Old-style VACUUM FULL is gone, but we have to process xvac for as long
7092 : * as we support having MOVED_OFF/MOVED_IN tuples in the database
7093 : */
7094 35521072 : xid = HeapTupleHeaderGetXvac(tuple);
7095 35521072 : if (TransactionIdIsNormal(xid))
7096 : {
7097 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7098 : Assert(TransactionIdPrecedes(xid, cutoffs->OldestXmin));
7099 :
7100 : /*
7101 : * For Xvac, we always freeze proactively. This allows totally_frozen
7102 : * tracking to ignore xvac.
7103 : */
7104 0 : replace_xvac = pagefrz->freeze_required = true;
7105 :
7106 : /* Will set replace_xvac flags in freeze plan below */
7107 : }
7108 :
7109 : /* Now process xmax */
7110 35521072 : xid = frz->xmax;
7111 35521072 : if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
7112 : {
7113 : /* Raw xmax is a MultiXactId */
7114 : TransactionId newxmax;
7115 : uint16 flags;
7116 :
7117 : /*
7118 : * We will either remove xmax completely (in the "freeze_xmax" path),
7119 : * process xmax by replacing it (in the "replace_xmax" path), or
7120 : * perform no-op xmax processing. The only constraint is that the
7121 : * FreezeLimit/MultiXactCutoff postcondition must never be violated.
7122 : */
7123 12 : newxmax = FreezeMultiXactId(xid, tuple->t_infomask, cutoffs,
7124 : &flags, pagefrz);
7125 :
7126 12 : if (flags & FRM_NOOP)
7127 : {
7128 : /*
7129 : * xmax is a MultiXactId, and nothing about it changes for now.
7130 : * This is the only case where 'freeze_required' won't have been
7131 : * set for us by FreezeMultiXactId, as well as the only case where
7132 : * neither freeze_xmax nor replace_xmax are set (given a multi).
7133 : *
7134 : * This is a no-op, but the call to FreezeMultiXactId might have
7135 : * ratcheted back NewRelfrozenXid and/or NewRelminMxid trackers
7136 : * for us (the "freeze page" variants, specifically). That'll
7137 : * make it safe for our caller to freeze the page later on, while
7138 : * leaving this particular xmax undisturbed.
7139 : *
7140 : * FreezeMultiXactId is _not_ responsible for the "no freeze"
7141 : * NewRelfrozenXid/NewRelminMxid trackers, though -- that's our
7142 : * job. A call to heap_tuple_should_freeze for this same tuple
7143 : * will take place below if 'freeze_required' isn't set already.
7144 : * (This repeats work from FreezeMultiXactId, but allows "no
7145 : * freeze" tracker maintenance to happen in only one place.)
7146 : */
7147 : Assert(!MultiXactIdPrecedes(newxmax, cutoffs->MultiXactCutoff));
7148 : Assert(MultiXactIdIsValid(newxmax) && xid == newxmax);
7149 : }
7150 10 : else if (flags & FRM_RETURN_IS_XID)
7151 : {
7152 : /*
7153 : * xmax will become an updater Xid (original MultiXact's updater
7154 : * member Xid will be carried forward as a simple Xid in Xmax).
7155 : */
7156 : Assert(!TransactionIdPrecedes(newxmax, cutoffs->OldestXmin));
7157 :
7158 : /*
7159 : * NB -- some of these transformations are only valid because we
7160 : * know the return Xid is a tuple updater (i.e. not merely a
7161 : * locker.) Also note that the only reason we don't explicitly
7162 : * worry about HEAP_KEYS_UPDATED is because it lives in
7163 : * t_infomask2 rather than t_infomask.
7164 : */
7165 0 : frz->t_infomask &= ~HEAP_XMAX_BITS;
7166 0 : frz->xmax = newxmax;
7167 0 : if (flags & FRM_MARK_COMMITTED)
7168 0 : frz->t_infomask |= HEAP_XMAX_COMMITTED;
7169 0 : replace_xmax = true;
7170 : }
7171 10 : else if (flags & FRM_RETURN_IS_MULTI)
7172 : {
7173 : uint16 newbits;
7174 : uint16 newbits2;
7175 :
7176 : /*
7177 : * xmax is an old MultiXactId that we have to replace with a new
7178 : * MultiXactId, to carry forward two or more original member XIDs.
7179 : */
7180 : Assert(!MultiXactIdPrecedes(newxmax, cutoffs->OldestMxact));
7181 :
7182 : /*
7183 : * We can't use GetMultiXactIdHintBits directly on the new multi
7184 : * here; that routine initializes the masks to all zeroes, which
7185 : * would lose other bits we need. Doing it this way ensures all
7186 : * unrelated bits remain untouched.
7187 : */
7188 2 : frz->t_infomask &= ~HEAP_XMAX_BITS;
7189 2 : frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
7190 2 : GetMultiXactIdHintBits(newxmax, &newbits, &newbits2);
7191 2 : frz->t_infomask |= newbits;
7192 2 : frz->t_infomask2 |= newbits2;
7193 2 : frz->xmax = newxmax;
7194 2 : replace_xmax = true;
7195 : }
7196 : else
7197 : {
7198 : /*
7199 : * Freeze plan for tuple "freezes xmax" in the strictest sense:
7200 : * it'll leave nothing in xmax (neither an Xid nor a MultiXactId).
7201 : */
7202 : Assert(flags & FRM_INVALIDATE_XMAX);
7203 : Assert(!TransactionIdIsValid(newxmax));
7204 :
7205 : /* Will set freeze_xmax flags in freeze plan below */
7206 8 : freeze_xmax = true;
7207 : }
7208 :
7209 : /* MultiXactId processing forces freezing (barring FRM_NOOP case) */
7210 : Assert(pagefrz->freeze_required || (!freeze_xmax && !replace_xmax));
7211 : }
7212 35521060 : else if (TransactionIdIsNormal(xid))
7213 : {
7214 : /* Raw xmax is normal XID */
7215 8332762 : if (TransactionIdPrecedes(xid, cutoffs->relfrozenxid))
7216 0 : ereport(ERROR,
7217 : (errcode(ERRCODE_DATA_CORRUPTED),
7218 : errmsg_internal("found xmax %u from before relfrozenxid %u",
7219 : xid, cutoffs->relfrozenxid)));
7220 :
7221 : /* Will set freeze_xmax flags in freeze plan below */
7222 8332762 : freeze_xmax = TransactionIdPrecedes(xid, cutoffs->OldestXmin);
7223 :
7224 : /*
7225 : * Verify that xmax aborted if and when freeze plan is executed,
7226 : * provided it's from an update. (A lock-only xmax can be removed
7227 : * independent of this, since the lock is released at xact end.)
7228 : */
7229 8332762 : if (freeze_xmax && !HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask))
7230 2648 : frz->checkflags |= HEAP_FREEZE_CHECK_XMAX_ABORTED;
7231 : }
7232 27188298 : else if (!TransactionIdIsValid(xid))
7233 : {
7234 : /* Raw xmax is InvalidTransactionId XID */
7235 : Assert((tuple->t_infomask & HEAP_XMAX_IS_MULTI) == 0);
7236 27188298 : xmax_already_frozen = true;
7237 : }
7238 : else
7239 0 : ereport(ERROR,
7240 : (errcode(ERRCODE_DATA_CORRUPTED),
7241 : errmsg_internal("found raw xmax %u (infomask 0x%04x) not invalid and not multi",
7242 : xid, tuple->t_infomask)));
7243 :
7244 35521072 : if (freeze_xmin)
7245 : {
7246 : Assert(!xmin_already_frozen);
7247 :
7248 4613974 : frz->t_infomask |= HEAP_XMIN_FROZEN;
7249 : }
7250 35521072 : if (replace_xvac)
7251 : {
7252 : /*
7253 : * If a MOVED_OFF tuple is not dead, the xvac transaction must have
7254 : * failed; whereas a non-dead MOVED_IN tuple must mean the xvac
7255 : * transaction succeeded.
7256 : */
7257 : Assert(pagefrz->freeze_required);
7258 0 : if (tuple->t_infomask & HEAP_MOVED_OFF)
7259 0 : frz->frzflags |= XLH_INVALID_XVAC;
7260 : else
7261 0 : frz->frzflags |= XLH_FREEZE_XVAC;
7262 : }
7263 : if (replace_xmax)
7264 : {
7265 : Assert(!xmax_already_frozen && !freeze_xmax);
7266 : Assert(pagefrz->freeze_required);
7267 :
7268 : /* Already set replace_xmax flags in freeze plan earlier */
7269 : }
7270 35521072 : if (freeze_xmax)
7271 : {
7272 : Assert(!xmax_already_frozen && !replace_xmax);
7273 :
7274 4586 : frz->xmax = InvalidTransactionId;
7275 :
7276 : /*
7277 : * The tuple might be marked either XMAX_INVALID or XMAX_COMMITTED +
7278 : * LOCKED. Normalize to INVALID just to be sure no one gets confused.
7279 : * Also get rid of the HEAP_KEYS_UPDATED bit.
7280 : */
7281 4586 : frz->t_infomask &= ~HEAP_XMAX_BITS;
7282 4586 : frz->t_infomask |= HEAP_XMAX_INVALID;
7283 4586 : frz->t_infomask2 &= ~HEAP_HOT_UPDATED;
7284 4586 : frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
7285 : }
7286 :
7287 : /*
7288 : * Determine if this tuple is already totally frozen, or will become
7289 : * totally frozen (provided caller executes freeze plans for the page)
7290 : */
7291 69759388 : *totally_frozen = ((freeze_xmin || xmin_already_frozen) &&
7292 34238316 : (freeze_xmax || xmax_already_frozen));
7293 :
7294 35521072 : if (!pagefrz->freeze_required && !(xmin_already_frozen &&
7295 : xmax_already_frozen))
7296 : {
7297 : /*
7298 : * So far no previous tuple from the page made freezing mandatory.
7299 : * Does this tuple force caller to freeze the entire page?
7300 : */
7301 11257646 : pagefrz->freeze_required =
7302 11257646 : heap_tuple_should_freeze(tuple, cutoffs,
7303 : &pagefrz->NoFreezePageRelfrozenXid,
7304 : &pagefrz->NoFreezePageRelminMxid);
7305 : }
7306 :
7307 : /* Tell caller if this tuple has a usable freeze plan set in *frz */
7308 35521072 : return freeze_xmin || replace_xvac || replace_xmax || freeze_xmax;
7309 : }
7310 :
7311 : /*
7312 : * Perform xmin/xmax XID status sanity checks before actually executing freeze
7313 : * plans.
7314 : *
7315 : * heap_prepare_freeze_tuple doesn't perform these checks directly because
7316 : * pg_xact lookups are relatively expensive. They shouldn't be repeated by
7317 : * successive VACUUMs that each decide against freezing the same page.
7318 : */
7319 : void
7320 44854 : heap_pre_freeze_checks(Buffer buffer,
7321 : HeapTupleFreeze *tuples, int ntuples)
7322 : {
7323 44854 : Page page = BufferGetPage(buffer);
7324 :
7325 1932266 : for (int i = 0; i < ntuples; i++)
7326 : {
7327 1887412 : HeapTupleFreeze *frz = tuples + i;
7328 1887412 : ItemId itemid = PageGetItemId(page, frz->offset);
7329 : HeapTupleHeader htup;
7330 :
7331 1887412 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
7332 :
7333 : /* Deliberately avoid relying on tuple hint bits here */
7334 1887412 : if (frz->checkflags & HEAP_FREEZE_CHECK_XMIN_COMMITTED)
7335 : {
7336 1887410 : TransactionId xmin = HeapTupleHeaderGetRawXmin(htup);
7337 :
7338 : Assert(!HeapTupleHeaderXminFrozen(htup));
7339 1887410 : if (unlikely(!TransactionIdDidCommit(xmin)))
7340 0 : ereport(ERROR,
7341 : (errcode(ERRCODE_DATA_CORRUPTED),
7342 : errmsg_internal("uncommitted xmin %u needs to be frozen",
7343 : xmin)));
7344 : }
7345 :
7346 : /*
7347 : * TransactionIdDidAbort won't work reliably in the presence of XIDs
7348 : * left behind by transactions that were in progress during a crash,
7349 : * so we can only check that xmax didn't commit
7350 : */
7351 1887412 : if (frz->checkflags & HEAP_FREEZE_CHECK_XMAX_ABORTED)
7352 : {
7353 1156 : TransactionId xmax = HeapTupleHeaderGetRawXmax(htup);
7354 :
7355 : Assert(TransactionIdIsNormal(xmax));
7356 1156 : if (unlikely(TransactionIdDidCommit(xmax)))
7357 0 : ereport(ERROR,
7358 : (errcode(ERRCODE_DATA_CORRUPTED),
7359 : errmsg_internal("cannot freeze committed xmax %u",
7360 : xmax)));
7361 : }
7362 : }
7363 44854 : }
7364 :
7365 : /*
7366 : * Helper which executes freezing of one or more heap tuples on a page on
7367 : * behalf of caller. Caller passes an array of tuple plans from
7368 : * heap_prepare_freeze_tuple. Caller must set 'offset' in each plan for us.
7369 : * Must be called in a critical section that also marks the buffer dirty and,
7370 : * if needed, emits WAL.
7371 : */
7372 : void
7373 44854 : heap_freeze_prepared_tuples(Buffer buffer, HeapTupleFreeze *tuples, int ntuples)
7374 : {
7375 44854 : Page page = BufferGetPage(buffer);
7376 :
7377 1932266 : for (int i = 0; i < ntuples; i++)
7378 : {
7379 1887412 : HeapTupleFreeze *frz = tuples + i;
7380 1887412 : ItemId itemid = PageGetItemId(page, frz->offset);
7381 : HeapTupleHeader htup;
7382 :
7383 1887412 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
7384 1887412 : heap_execute_freeze_tuple(htup, frz);
7385 : }
7386 44854 : }
7387 :
7388 : /*
7389 : * heap_freeze_tuple
7390 : * Freeze tuple in place, without WAL logging.
7391 : *
7392 : * Useful for callers like CLUSTER that perform their own WAL logging.
7393 : */
7394 : bool
7395 715270 : heap_freeze_tuple(HeapTupleHeader tuple,
7396 : TransactionId relfrozenxid, TransactionId relminmxid,
7397 : TransactionId FreezeLimit, TransactionId MultiXactCutoff)
7398 : {
7399 : HeapTupleFreeze frz;
7400 : bool do_freeze;
7401 : bool totally_frozen;
7402 : struct VacuumCutoffs cutoffs;
7403 : HeapPageFreeze pagefrz;
7404 :
7405 715270 : cutoffs.relfrozenxid = relfrozenxid;
7406 715270 : cutoffs.relminmxid = relminmxid;
7407 715270 : cutoffs.OldestXmin = FreezeLimit;
7408 715270 : cutoffs.OldestMxact = MultiXactCutoff;
7409 715270 : cutoffs.FreezeLimit = FreezeLimit;
7410 715270 : cutoffs.MultiXactCutoff = MultiXactCutoff;
7411 :
7412 715270 : pagefrz.freeze_required = true;
7413 715270 : pagefrz.FreezePageRelfrozenXid = FreezeLimit;
7414 715270 : pagefrz.FreezePageRelminMxid = MultiXactCutoff;
7415 715270 : pagefrz.NoFreezePageRelfrozenXid = FreezeLimit;
7416 715270 : pagefrz.NoFreezePageRelminMxid = MultiXactCutoff;
7417 :
7418 715270 : do_freeze = heap_prepare_freeze_tuple(tuple, &cutoffs,
7419 : &pagefrz, &frz, &totally_frozen);
7420 :
7421 : /*
7422 : * Note that because this is not a WAL-logged operation, we don't need to
7423 : * fill in the offset in the freeze record.
7424 : */
7425 :
7426 715270 : if (do_freeze)
7427 529976 : heap_execute_freeze_tuple(tuple, &frz);
7428 715270 : return do_freeze;
7429 : }
7430 :
7431 : /*
7432 : * For a given MultiXactId, return the hint bits that should be set in the
7433 : * tuple's infomask.
7434 : *
7435 : * Normally this should be called for a multixact that was just created, and
7436 : * so is on our local cache, so the GetMembers call is fast.
7437 : */
7438 : static void
7439 2354 : GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
7440 : uint16 *new_infomask2)
7441 : {
7442 : int nmembers;
7443 : MultiXactMember *members;
7444 : int i;
7445 2354 : uint16 bits = HEAP_XMAX_IS_MULTI;
7446 2354 : uint16 bits2 = 0;
7447 2354 : bool has_update = false;
7448 2354 : LockTupleMode strongest = LockTupleKeyShare;
7449 :
7450 : /*
7451 : * We only use this in multis we just created, so they cannot be values
7452 : * pre-pg_upgrade.
7453 : */
7454 2354 : nmembers = GetMultiXactIdMembers(multi, &members, false, false);
7455 :
7456 7202 : for (i = 0; i < nmembers; i++)
7457 : {
7458 : LockTupleMode mode;
7459 :
7460 : /*
7461 : * Remember the strongest lock mode held by any member of the
7462 : * multixact.
7463 : */
7464 4848 : mode = TUPLOCK_from_mxstatus(members[i].status);
7465 4848 : if (mode > strongest)
7466 1318 : strongest = mode;
7467 :
7468 : /* See what other bits we need */
7469 4848 : switch (members[i].status)
7470 : {
7471 4466 : case MultiXactStatusForKeyShare:
7472 : case MultiXactStatusForShare:
7473 : case MultiXactStatusForNoKeyUpdate:
7474 4466 : break;
7475 :
7476 104 : case MultiXactStatusForUpdate:
7477 104 : bits2 |= HEAP_KEYS_UPDATED;
7478 104 : break;
7479 :
7480 258 : case MultiXactStatusNoKeyUpdate:
7481 258 : has_update = true;
7482 258 : break;
7483 :
7484 20 : case MultiXactStatusUpdate:
7485 20 : bits2 |= HEAP_KEYS_UPDATED;
7486 20 : has_update = true;
7487 20 : break;
7488 : }
7489 : }
7490 :
7491 2354 : if (strongest == LockTupleExclusive ||
7492 : strongest == LockTupleNoKeyExclusive)
7493 438 : bits |= HEAP_XMAX_EXCL_LOCK;
7494 1916 : else if (strongest == LockTupleShare)
7495 874 : bits |= HEAP_XMAX_SHR_LOCK;
7496 1042 : else if (strongest == LockTupleKeyShare)
7497 1042 : bits |= HEAP_XMAX_KEYSHR_LOCK;
7498 :
7499 2354 : if (!has_update)
7500 2076 : bits |= HEAP_XMAX_LOCK_ONLY;
7501 :
7502 2354 : if (nmembers > 0)
7503 2354 : pfree(members);
7504 :
7505 2354 : *new_infomask = bits;
7506 2354 : *new_infomask2 = bits2;
7507 2354 : }
7508 :
7509 : /*
7510 : * MultiXactIdGetUpdateXid
7511 : *
7512 : * Given a multixact Xmax and corresponding infomask, which does not have the
7513 : * HEAP_XMAX_LOCK_ONLY bit set, obtain and return the Xid of the updating
7514 : * transaction.
7515 : *
7516 : * Caller is expected to check the status of the updating transaction, if
7517 : * necessary.
7518 : */
7519 : static TransactionId
7520 922 : MultiXactIdGetUpdateXid(TransactionId xmax, uint16 t_infomask)
7521 : {
7522 922 : TransactionId update_xact = InvalidTransactionId;
7523 : MultiXactMember *members;
7524 : int nmembers;
7525 :
7526 : Assert(!(t_infomask & HEAP_XMAX_LOCK_ONLY));
7527 : Assert(t_infomask & HEAP_XMAX_IS_MULTI);
7528 :
7529 : /*
7530 : * Since we know the LOCK_ONLY bit is not set, this cannot be a multi from
7531 : * pre-pg_upgrade.
7532 : */
7533 922 : nmembers = GetMultiXactIdMembers(xmax, &members, false, false);
7534 :
7535 922 : if (nmembers > 0)
7536 : {
7537 : int i;
7538 :
7539 2374 : for (i = 0; i < nmembers; i++)
7540 : {
7541 : /* Ignore lockers */
7542 2374 : if (!ISUPDATE_from_mxstatus(members[i].status))
7543 1452 : continue;
7544 :
7545 : /* there can be at most one updater */
7546 : Assert(update_xact == InvalidTransactionId);
7547 922 : update_xact = members[i].xid;
7548 : #ifndef USE_ASSERT_CHECKING
7549 :
7550 : /*
7551 : * in an assert-enabled build, walk the whole array to ensure
7552 : * there's no other updater.
7553 : */
7554 922 : break;
7555 : #endif
7556 : }
7557 :
7558 922 : pfree(members);
7559 : }
7560 :
7561 922 : return update_xact;
7562 : }
7563 :
7564 : /*
7565 : * HeapTupleGetUpdateXid
7566 : * As above, but use a HeapTupleHeader
7567 : *
7568 : * See also HeapTupleHeaderGetUpdateXid, which can be used without previously
7569 : * checking the hint bits.
7570 : */
7571 : TransactionId
7572 906 : HeapTupleGetUpdateXid(const HeapTupleHeaderData *tup)
7573 : {
7574 906 : return MultiXactIdGetUpdateXid(HeapTupleHeaderGetRawXmax(tup),
7575 906 : tup->t_infomask);
7576 : }
7577 :
7578 : /*
7579 : * Does the given multixact conflict with the current transaction grabbing a
7580 : * tuple lock of the given strength?
7581 : *
7582 : * The passed infomask pairs up with the given multixact in the tuple header.
7583 : *
7584 : * If current_is_member is not NULL, it is set to 'true' if the current
7585 : * transaction is a member of the given multixact.
7586 : */
7587 : static bool
7588 198 : DoesMultiXactIdConflict(MultiXactId multi, uint16 infomask,
7589 : LockTupleMode lockmode, bool *current_is_member)
7590 : {
7591 : int nmembers;
7592 : MultiXactMember *members;
7593 198 : bool result = false;
7594 198 : LOCKMODE wanted = tupleLockExtraInfo[lockmode].hwlock;
7595 :
7596 198 : if (HEAP_LOCKED_UPGRADED(infomask))
7597 0 : return false;
7598 :
7599 198 : nmembers = GetMultiXactIdMembers(multi, &members, false,
7600 198 : HEAP_XMAX_IS_LOCKED_ONLY(infomask));
7601 198 : if (nmembers >= 0)
7602 : {
7603 : int i;
7604 :
7605 618 : for (i = 0; i < nmembers; i++)
7606 : {
7607 : TransactionId memxid;
7608 : LOCKMODE memlockmode;
7609 :
7610 434 : if (result && (current_is_member == NULL || *current_is_member))
7611 : break;
7612 :
7613 420 : memlockmode = LOCKMODE_from_mxstatus(members[i].status);
7614 :
7615 : /* ignore members from current xact (but track their presence) */
7616 420 : memxid = members[i].xid;
7617 420 : if (TransactionIdIsCurrentTransactionId(memxid))
7618 : {
7619 184 : if (current_is_member != NULL)
7620 156 : *current_is_member = true;
7621 184 : continue;
7622 : }
7623 236 : else if (result)
7624 16 : continue;
7625 :
7626 : /* ignore members that don't conflict with the lock we want */
7627 220 : if (!DoLockModesConflict(memlockmode, wanted))
7628 142 : continue;
7629 :
7630 78 : if (ISUPDATE_from_mxstatus(members[i].status))
7631 : {
7632 : /* ignore aborted updaters */
7633 34 : if (TransactionIdDidAbort(memxid))
7634 2 : continue;
7635 : }
7636 : else
7637 : {
7638 : /* ignore lockers-only that are no longer in progress */
7639 44 : if (!TransactionIdIsInProgress(memxid))
7640 14 : continue;
7641 : }
7642 :
7643 : /*
7644 : * Whatever remains are either live lockers that conflict with our
7645 : * wanted lock, and updaters that are not aborted. Those conflict
7646 : * with what we want. Set up to return true, but keep going to
7647 : * look for the current transaction among the multixact members,
7648 : * if needed.
7649 : */
7650 62 : result = true;
7651 : }
7652 198 : pfree(members);
7653 : }
7654 :
7655 198 : return result;
7656 : }
7657 :
7658 : /*
7659 : * Do_MultiXactIdWait
7660 : * Actual implementation for the two functions below.
7661 : *
7662 : * 'multi', 'status' and 'infomask' indicate what to sleep on (the status is
7663 : * needed to ensure we only sleep on conflicting members, and the infomask is
7664 : * used to optimize multixact access in case it's a lock-only multi); 'nowait'
7665 : * indicates whether to use conditional lock acquisition, to allow callers to
7666 : * fail if lock is unavailable. 'rel', 'ctid' and 'oper' are used to set up
7667 : * context information for error messages. 'remaining', if not NULL, receives
7668 : * the number of members that are still running, including any (non-aborted)
7669 : * subtransactions of our own transaction. 'logLockFailure' indicates whether
7670 : * to log details when a lock acquisition fails with 'nowait' enabled.
7671 : *
7672 : * We do this by sleeping on each member using XactLockTableWait. Any
7673 : * members that belong to the current backend are *not* waited for, however;
7674 : * this would not merely be useless but would lead to Assert failure inside
7675 : * XactLockTableWait. By the time this returns, it is certain that all
7676 : * transactions *of other backends* that were members of the MultiXactId
7677 : * that conflict with the requested status are dead (and no new ones can have
7678 : * been added, since it is not legal to add members to an existing
7679 : * MultiXactId).
7680 : *
7681 : * But by the time we finish sleeping, someone else may have changed the Xmax
7682 : * of the containing tuple, so the caller needs to iterate on us somehow.
7683 : *
7684 : * Note that in case we return false, the number of remaining members is
7685 : * not to be trusted.
7686 : */
7687 : static bool
7688 116 : Do_MultiXactIdWait(MultiXactId multi, MultiXactStatus status,
7689 : uint16 infomask, bool nowait,
7690 : Relation rel, ItemPointer ctid, XLTW_Oper oper,
7691 : int *remaining, bool logLockFailure)
7692 : {
7693 116 : bool result = true;
7694 : MultiXactMember *members;
7695 : int nmembers;
7696 116 : int remain = 0;
7697 :
7698 : /* for pre-pg_upgrade tuples, no need to sleep at all */
7699 116 : nmembers = HEAP_LOCKED_UPGRADED(infomask) ? -1 :
7700 116 : GetMultiXactIdMembers(multi, &members, false,
7701 116 : HEAP_XMAX_IS_LOCKED_ONLY(infomask));
7702 :
7703 116 : if (nmembers >= 0)
7704 : {
7705 : int i;
7706 :
7707 374 : for (i = 0; i < nmembers; i++)
7708 : {
7709 266 : TransactionId memxid = members[i].xid;
7710 266 : MultiXactStatus memstatus = members[i].status;
7711 :
7712 266 : if (TransactionIdIsCurrentTransactionId(memxid))
7713 : {
7714 48 : remain++;
7715 48 : continue;
7716 : }
7717 :
7718 218 : if (!DoLockModesConflict(LOCKMODE_from_mxstatus(memstatus),
7719 218 : LOCKMODE_from_mxstatus(status)))
7720 : {
7721 44 : if (remaining && TransactionIdIsInProgress(memxid))
7722 16 : remain++;
7723 44 : continue;
7724 : }
7725 :
7726 : /*
7727 : * This member conflicts with our multi, so we have to sleep (or
7728 : * return failure, if asked to avoid waiting.)
7729 : *
7730 : * Note that we don't set up an error context callback ourselves,
7731 : * but instead we pass the info down to XactLockTableWait. This
7732 : * might seem a bit wasteful because the context is set up and
7733 : * tore down for each member of the multixact, but in reality it
7734 : * should be barely noticeable, and it avoids duplicate code.
7735 : */
7736 174 : if (nowait)
7737 : {
7738 8 : result = ConditionalXactLockTableWait(memxid, logLockFailure);
7739 8 : if (!result)
7740 8 : break;
7741 : }
7742 : else
7743 166 : XactLockTableWait(memxid, rel, ctid, oper);
7744 : }
7745 :
7746 116 : pfree(members);
7747 : }
7748 :
7749 116 : if (remaining)
7750 20 : *remaining = remain;
7751 :
7752 116 : return result;
7753 : }
7754 :
7755 : /*
7756 : * MultiXactIdWait
7757 : * Sleep on a MultiXactId.
7758 : *
7759 : * By the time we finish sleeping, someone else may have changed the Xmax
7760 : * of the containing tuple, so the caller needs to iterate on us somehow.
7761 : *
7762 : * We return (in *remaining, if not NULL) the number of members that are still
7763 : * running, including any (non-aborted) subtransactions of our own transaction.
7764 : */
7765 : static void
7766 108 : MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
7767 : Relation rel, ItemPointer ctid, XLTW_Oper oper,
7768 : int *remaining)
7769 : {
7770 108 : (void) Do_MultiXactIdWait(multi, status, infomask, false,
7771 : rel, ctid, oper, remaining, false);
7772 108 : }
7773 :
7774 : /*
7775 : * ConditionalMultiXactIdWait
7776 : * As above, but only lock if we can get the lock without blocking.
7777 : *
7778 : * By the time we finish sleeping, someone else may have changed the Xmax
7779 : * of the containing tuple, so the caller needs to iterate on us somehow.
7780 : *
7781 : * If the multixact is now all gone, return true. Returns false if some
7782 : * transactions might still be running.
7783 : *
7784 : * We return (in *remaining, if not NULL) the number of members that are still
7785 : * running, including any (non-aborted) subtransactions of our own transaction.
7786 : */
7787 : static bool
7788 8 : ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
7789 : uint16 infomask, Relation rel, int *remaining,
7790 : bool logLockFailure)
7791 : {
7792 8 : return Do_MultiXactIdWait(multi, status, infomask, true,
7793 : rel, NULL, XLTW_None, remaining, logLockFailure);
7794 : }
7795 :
7796 : /*
7797 : * heap_tuple_needs_eventual_freeze
7798 : *
7799 : * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
7800 : * will eventually require freezing (if tuple isn't removed by pruning first).
7801 : */
7802 : bool
7803 301986 : heap_tuple_needs_eventual_freeze(HeapTupleHeader tuple)
7804 : {
7805 : TransactionId xid;
7806 :
7807 : /*
7808 : * If xmin is a normal transaction ID, this tuple is definitely not
7809 : * frozen.
7810 : */
7811 301986 : xid = HeapTupleHeaderGetXmin(tuple);
7812 301986 : if (TransactionIdIsNormal(xid))
7813 5734 : return true;
7814 :
7815 : /*
7816 : * If xmax is a valid xact or multixact, this tuple is also not frozen.
7817 : */
7818 296252 : if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
7819 : {
7820 : MultiXactId multi;
7821 :
7822 0 : multi = HeapTupleHeaderGetRawXmax(tuple);
7823 0 : if (MultiXactIdIsValid(multi))
7824 0 : return true;
7825 : }
7826 : else
7827 : {
7828 296252 : xid = HeapTupleHeaderGetRawXmax(tuple);
7829 296252 : if (TransactionIdIsNormal(xid))
7830 14 : return true;
7831 : }
7832 :
7833 296238 : if (tuple->t_infomask & HEAP_MOVED)
7834 : {
7835 0 : xid = HeapTupleHeaderGetXvac(tuple);
7836 0 : if (TransactionIdIsNormal(xid))
7837 0 : return true;
7838 : }
7839 :
7840 296238 : return false;
7841 : }
7842 :
7843 : /*
7844 : * heap_tuple_should_freeze
7845 : *
7846 : * Return value indicates if heap_prepare_freeze_tuple sibling function would
7847 : * (or should) force freezing of the heap page that contains caller's tuple.
7848 : * Tuple header XIDs/MXIDs < FreezeLimit/MultiXactCutoff trigger freezing.
7849 : * This includes (xmin, xmax, xvac) fields, as well as MultiXact member XIDs.
7850 : *
7851 : * The *NoFreezePageRelfrozenXid and *NoFreezePageRelminMxid input/output
7852 : * arguments help VACUUM track the oldest extant XID/MXID remaining in rel.
7853 : * Our working assumption is that caller won't decide to freeze this tuple.
7854 : * It's up to caller to only ratchet back its own top-level trackers after the
7855 : * point that it fully commits to not freezing the tuple/page in question.
7856 : */
7857 : bool
7858 11261962 : heap_tuple_should_freeze(HeapTupleHeader tuple,
7859 : const struct VacuumCutoffs *cutoffs,
7860 : TransactionId *NoFreezePageRelfrozenXid,
7861 : MultiXactId *NoFreezePageRelminMxid)
7862 : {
7863 : TransactionId xid;
7864 : MultiXactId multi;
7865 11261962 : bool freeze = false;
7866 :
7867 : /* First deal with xmin */
7868 11261962 : xid = HeapTupleHeaderGetXmin(tuple);
7869 11261962 : if (TransactionIdIsNormal(xid))
7870 : {
7871 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7872 3573600 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7873 44908 : *NoFreezePageRelfrozenXid = xid;
7874 3573600 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
7875 41038 : freeze = true;
7876 : }
7877 :
7878 : /* Now deal with xmax */
7879 11261962 : xid = InvalidTransactionId;
7880 11261962 : multi = InvalidMultiXactId;
7881 11261962 : if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
7882 4 : multi = HeapTupleHeaderGetRawXmax(tuple);
7883 : else
7884 11261958 : xid = HeapTupleHeaderGetRawXmax(tuple);
7885 :
7886 11261962 : if (TransactionIdIsNormal(xid))
7887 : {
7888 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7889 : /* xmax is a non-permanent XID */
7890 8200752 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7891 8 : *NoFreezePageRelfrozenXid = xid;
7892 8200752 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
7893 8 : freeze = true;
7894 : }
7895 3061210 : else if (!MultiXactIdIsValid(multi))
7896 : {
7897 : /* xmax is a permanent XID or invalid MultiXactId/XID */
7898 : }
7899 4 : else if (HEAP_LOCKED_UPGRADED(tuple->t_infomask))
7900 : {
7901 : /* xmax is a pg_upgrade'd MultiXact, which can't have updater XID */
7902 0 : if (MultiXactIdPrecedes(multi, *NoFreezePageRelminMxid))
7903 0 : *NoFreezePageRelminMxid = multi;
7904 : /* heap_prepare_freeze_tuple always freezes pg_upgrade'd xmax */
7905 0 : freeze = true;
7906 : }
7907 : else
7908 : {
7909 : /* xmax is a MultiXactId that may have an updater XID */
7910 : MultiXactMember *members;
7911 : int nmembers;
7912 :
7913 : Assert(MultiXactIdPrecedesOrEquals(cutoffs->relminmxid, multi));
7914 4 : if (MultiXactIdPrecedes(multi, *NoFreezePageRelminMxid))
7915 4 : *NoFreezePageRelminMxid = multi;
7916 4 : if (MultiXactIdPrecedes(multi, cutoffs->MultiXactCutoff))
7917 4 : freeze = true;
7918 :
7919 : /* need to check whether any member of the mxact is old */
7920 4 : nmembers = GetMultiXactIdMembers(multi, &members, false,
7921 4 : HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask));
7922 :
7923 10 : for (int i = 0; i < nmembers; i++)
7924 : {
7925 6 : xid = members[i].xid;
7926 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7927 6 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7928 0 : *NoFreezePageRelfrozenXid = xid;
7929 6 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
7930 0 : freeze = true;
7931 : }
7932 4 : if (nmembers > 0)
7933 2 : pfree(members);
7934 : }
7935 :
7936 11261962 : if (tuple->t_infomask & HEAP_MOVED)
7937 : {
7938 0 : xid = HeapTupleHeaderGetXvac(tuple);
7939 0 : if (TransactionIdIsNormal(xid))
7940 : {
7941 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7942 0 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7943 0 : *NoFreezePageRelfrozenXid = xid;
7944 : /* heap_prepare_freeze_tuple forces xvac freezing */
7945 0 : freeze = true;
7946 : }
7947 : }
7948 :
7949 11261962 : return freeze;
7950 : }
7951 :
7952 : /*
7953 : * Maintain snapshotConflictHorizon for caller by ratcheting forward its value
7954 : * using any committed XIDs contained in 'tuple', an obsolescent heap tuple
7955 : * that caller is in the process of physically removing, e.g. via HOT pruning
7956 : * or index deletion.
7957 : *
7958 : * Caller must initialize its value to InvalidTransactionId, which is
7959 : * generally interpreted as "definitely no need for a recovery conflict".
7960 : * Final value must reflect all heap tuples that caller will physically remove
7961 : * (or remove TID references to) via its ongoing pruning/deletion operation.
7962 : * ResolveRecoveryConflictWithSnapshot() is passed the final value (taken from
7963 : * caller's WAL record) by REDO routine when it replays caller's operation.
7964 : */
7965 : void
7966 2991154 : HeapTupleHeaderAdvanceConflictHorizon(HeapTupleHeader tuple,
7967 : TransactionId *snapshotConflictHorizon)
7968 : {
7969 2991154 : TransactionId xmin = HeapTupleHeaderGetXmin(tuple);
7970 2991154 : TransactionId xmax = HeapTupleHeaderGetUpdateXid(tuple);
7971 2991154 : TransactionId xvac = HeapTupleHeaderGetXvac(tuple);
7972 :
7973 2991154 : if (tuple->t_infomask & HEAP_MOVED)
7974 : {
7975 0 : if (TransactionIdPrecedes(*snapshotConflictHorizon, xvac))
7976 0 : *snapshotConflictHorizon = xvac;
7977 : }
7978 :
7979 : /*
7980 : * Ignore tuples inserted by an aborted transaction or if the tuple was
7981 : * updated/deleted by the inserting transaction.
7982 : *
7983 : * Look for a committed hint bit, or if no xmin bit is set, check clog.
7984 : */
7985 2991154 : if (HeapTupleHeaderXminCommitted(tuple) ||
7986 209162 : (!HeapTupleHeaderXminInvalid(tuple) && TransactionIdDidCommit(xmin)))
7987 : {
7988 5407208 : if (xmax != xmin &&
7989 2568694 : TransactionIdFollows(xmax, *snapshotConflictHorizon))
7990 193924 : *snapshotConflictHorizon = xmax;
7991 : }
7992 2991154 : }
7993 :
7994 : #ifdef USE_PREFETCH
7995 : /*
7996 : * Helper function for heap_index_delete_tuples. Issues prefetch requests for
7997 : * prefetch_count buffers. The prefetch_state keeps track of all the buffers
7998 : * we can prefetch, and which have already been prefetched; each call to this
7999 : * function picks up where the previous call left off.
8000 : *
8001 : * Note: we expect the deltids array to be sorted in an order that groups TIDs
8002 : * by heap block, with all TIDs for each block appearing together in exactly
8003 : * one group.
8004 : */
8005 : static void
8006 36822 : index_delete_prefetch_buffer(Relation rel,
8007 : IndexDeletePrefetchState *prefetch_state,
8008 : int prefetch_count)
8009 : {
8010 36822 : BlockNumber cur_hblkno = prefetch_state->cur_hblkno;
8011 36822 : int count = 0;
8012 : int i;
8013 36822 : int ndeltids = prefetch_state->ndeltids;
8014 36822 : TM_IndexDelete *deltids = prefetch_state->deltids;
8015 :
8016 36822 : for (i = prefetch_state->next_item;
8017 1286950 : i < ndeltids && count < prefetch_count;
8018 1250128 : i++)
8019 : {
8020 1250128 : ItemPointer htid = &deltids[i].tid;
8021 :
8022 2489304 : if (cur_hblkno == InvalidBlockNumber ||
8023 1239176 : ItemPointerGetBlockNumber(htid) != cur_hblkno)
8024 : {
8025 32838 : cur_hblkno = ItemPointerGetBlockNumber(htid);
8026 32838 : PrefetchBuffer(rel, MAIN_FORKNUM, cur_hblkno);
8027 32838 : count++;
8028 : }
8029 : }
8030 :
8031 : /*
8032 : * Save the prefetch position so that next time we can continue from that
8033 : * position.
8034 : */
8035 36822 : prefetch_state->next_item = i;
8036 36822 : prefetch_state->cur_hblkno = cur_hblkno;
8037 36822 : }
8038 : #endif
8039 :
8040 : /*
8041 : * Helper function for heap_index_delete_tuples. Checks for index corruption
8042 : * involving an invalid TID in index AM caller's index page.
8043 : *
8044 : * This is an ideal place for these checks. The index AM must hold a buffer
8045 : * lock on the index page containing the TIDs we examine here, so we don't
8046 : * have to worry about concurrent VACUUMs at all. We can be sure that the
8047 : * index is corrupt when htid points directly to an LP_UNUSED item or
8048 : * heap-only tuple, which is not the case during standard index scans.
8049 : */
8050 : static inline void
8051 1061266 : index_delete_check_htid(TM_IndexDeleteOp *delstate,
8052 : Page page, OffsetNumber maxoff,
8053 : ItemPointer htid, TM_IndexStatus *istatus)
8054 : {
8055 1061266 : OffsetNumber indexpagehoffnum = ItemPointerGetOffsetNumber(htid);
8056 : ItemId iid;
8057 :
8058 : Assert(OffsetNumberIsValid(istatus->idxoffnum));
8059 :
8060 1061266 : if (unlikely(indexpagehoffnum > maxoff))
8061 0 : ereport(ERROR,
8062 : (errcode(ERRCODE_INDEX_CORRUPTED),
8063 : errmsg_internal("heap tid from index tuple (%u,%u) points past end of heap page line pointer array at offset %u of block %u in index \"%s\"",
8064 : ItemPointerGetBlockNumber(htid),
8065 : indexpagehoffnum,
8066 : istatus->idxoffnum, delstate->iblknum,
8067 : RelationGetRelationName(delstate->irel))));
8068 :
8069 1061266 : iid = PageGetItemId(page, indexpagehoffnum);
8070 1061266 : if (unlikely(!ItemIdIsUsed(iid)))
8071 0 : ereport(ERROR,
8072 : (errcode(ERRCODE_INDEX_CORRUPTED),
8073 : errmsg_internal("heap tid from index tuple (%u,%u) points to unused heap page item at offset %u of block %u in index \"%s\"",
8074 : ItemPointerGetBlockNumber(htid),
8075 : indexpagehoffnum,
8076 : istatus->idxoffnum, delstate->iblknum,
8077 : RelationGetRelationName(delstate->irel))));
8078 :
8079 1061266 : if (ItemIdHasStorage(iid))
8080 : {
8081 : HeapTupleHeader htup;
8082 :
8083 : Assert(ItemIdIsNormal(iid));
8084 628596 : htup = (HeapTupleHeader) PageGetItem(page, iid);
8085 :
8086 628596 : if (unlikely(HeapTupleHeaderIsHeapOnly(htup)))
8087 0 : ereport(ERROR,
8088 : (errcode(ERRCODE_INDEX_CORRUPTED),
8089 : errmsg_internal("heap tid from index tuple (%u,%u) points to heap-only tuple at offset %u of block %u in index \"%s\"",
8090 : ItemPointerGetBlockNumber(htid),
8091 : indexpagehoffnum,
8092 : istatus->idxoffnum, delstate->iblknum,
8093 : RelationGetRelationName(delstate->irel))));
8094 : }
8095 1061266 : }
8096 :
8097 : /*
8098 : * heapam implementation of tableam's index_delete_tuples interface.
8099 : *
8100 : * This helper function is called by index AMs during index tuple deletion.
8101 : * See tableam header comments for an explanation of the interface implemented
8102 : * here and a general theory of operation. Note that each call here is either
8103 : * a simple index deletion call, or a bottom-up index deletion call.
8104 : *
8105 : * It's possible for this to generate a fair amount of I/O, since we may be
8106 : * deleting hundreds of tuples from a single index block. To amortize that
8107 : * cost to some degree, this uses prefetching and combines repeat accesses to
8108 : * the same heap block.
8109 : */
8110 : TransactionId
8111 10952 : heap_index_delete_tuples(Relation rel, TM_IndexDeleteOp *delstate)
8112 : {
8113 : /* Initial assumption is that earlier pruning took care of conflict */
8114 10952 : TransactionId snapshotConflictHorizon = InvalidTransactionId;
8115 10952 : BlockNumber blkno = InvalidBlockNumber;
8116 10952 : Buffer buf = InvalidBuffer;
8117 10952 : Page page = NULL;
8118 10952 : OffsetNumber maxoff = InvalidOffsetNumber;
8119 : TransactionId priorXmax;
8120 : #ifdef USE_PREFETCH
8121 : IndexDeletePrefetchState prefetch_state;
8122 : int prefetch_distance;
8123 : #endif
8124 : SnapshotData SnapshotNonVacuumable;
8125 10952 : int finalndeltids = 0,
8126 10952 : nblocksaccessed = 0;
8127 :
8128 : /* State that's only used in bottom-up index deletion case */
8129 10952 : int nblocksfavorable = 0;
8130 10952 : int curtargetfreespace = delstate->bottomupfreespace,
8131 10952 : lastfreespace = 0,
8132 10952 : actualfreespace = 0;
8133 10952 : bool bottomup_final_block = false;
8134 :
8135 10952 : InitNonVacuumableSnapshot(SnapshotNonVacuumable, GlobalVisTestFor(rel));
8136 :
8137 : /* Sort caller's deltids array by TID for further processing */
8138 10952 : index_delete_sort(delstate);
8139 :
8140 : /*
8141 : * Bottom-up case: resort deltids array in an order attuned to where the
8142 : * greatest number of promising TIDs are to be found, and determine how
8143 : * many blocks from the start of sorted array should be considered
8144 : * favorable. This will also shrink the deltids array in order to
8145 : * eliminate completely unfavorable blocks up front.
8146 : */
8147 10952 : if (delstate->bottomup)
8148 3752 : nblocksfavorable = bottomup_sort_and_shrink(delstate);
8149 :
8150 : #ifdef USE_PREFETCH
8151 : /* Initialize prefetch state. */
8152 10952 : prefetch_state.cur_hblkno = InvalidBlockNumber;
8153 10952 : prefetch_state.next_item = 0;
8154 10952 : prefetch_state.ndeltids = delstate->ndeltids;
8155 10952 : prefetch_state.deltids = delstate->deltids;
8156 :
8157 : /*
8158 : * Determine the prefetch distance that we will attempt to maintain.
8159 : *
8160 : * Since the caller holds a buffer lock somewhere in rel, we'd better make
8161 : * sure that isn't a catalog relation before we call code that does
8162 : * syscache lookups, to avoid risk of deadlock.
8163 : */
8164 10952 : if (IsCatalogRelation(rel))
8165 7764 : prefetch_distance = maintenance_io_concurrency;
8166 : else
8167 : prefetch_distance =
8168 3188 : get_tablespace_maintenance_io_concurrency(rel->rd_rel->reltablespace);
8169 :
8170 : /* Cap initial prefetch distance for bottom-up deletion caller */
8171 10952 : if (delstate->bottomup)
8172 : {
8173 : Assert(nblocksfavorable >= 1);
8174 : Assert(nblocksfavorable <= BOTTOMUP_MAX_NBLOCKS);
8175 3752 : prefetch_distance = Min(prefetch_distance, nblocksfavorable);
8176 : }
8177 :
8178 : /* Start prefetching. */
8179 10952 : index_delete_prefetch_buffer(rel, &prefetch_state, prefetch_distance);
8180 : #endif
8181 :
8182 : /* Iterate over deltids, determine which to delete, check their horizon */
8183 : Assert(delstate->ndeltids > 0);
8184 1072218 : for (int i = 0; i < delstate->ndeltids; i++)
8185 : {
8186 1065018 : TM_IndexDelete *ideltid = &delstate->deltids[i];
8187 1065018 : TM_IndexStatus *istatus = delstate->status + ideltid->id;
8188 1065018 : ItemPointer htid = &ideltid->tid;
8189 : OffsetNumber offnum;
8190 :
8191 : /*
8192 : * Read buffer, and perform required extra steps each time a new block
8193 : * is encountered. Avoid refetching if it's the same block as the one
8194 : * from the last htid.
8195 : */
8196 2119084 : if (blkno == InvalidBlockNumber ||
8197 1054066 : ItemPointerGetBlockNumber(htid) != blkno)
8198 : {
8199 : /*
8200 : * Consider giving up early for bottom-up index deletion caller
8201 : * first. (Only prefetch next-next block afterwards, when it
8202 : * becomes clear that we're at least going to access the next
8203 : * block in line.)
8204 : *
8205 : * Sometimes the first block frees so much space for bottom-up
8206 : * caller that the deletion process can end without accessing any
8207 : * more blocks. It is usually necessary to access 2 or 3 blocks
8208 : * per bottom-up deletion operation, though.
8209 : */
8210 29622 : if (delstate->bottomup)
8211 : {
8212 : /*
8213 : * We often allow caller to delete a few additional items
8214 : * whose entries we reached after the point that space target
8215 : * from caller was satisfied. The cost of accessing the page
8216 : * was already paid at that point, so it made sense to finish
8217 : * it off. When that happened, we finalize everything here
8218 : * (by finishing off the whole bottom-up deletion operation
8219 : * without needlessly paying the cost of accessing any more
8220 : * blocks).
8221 : */
8222 8218 : if (bottomup_final_block)
8223 304 : break;
8224 :
8225 : /*
8226 : * Give up when we didn't enable our caller to free any
8227 : * additional space as a result of processing the page that we
8228 : * just finished up with. This rule is the main way in which
8229 : * we keep the cost of bottom-up deletion under control.
8230 : */
8231 7914 : if (nblocksaccessed >= 1 && actualfreespace == lastfreespace)
8232 3448 : break;
8233 4466 : lastfreespace = actualfreespace; /* for next time */
8234 :
8235 : /*
8236 : * Deletion operation (which is bottom-up) will definitely
8237 : * access the next block in line. Prepare for that now.
8238 : *
8239 : * Decay target free space so that we don't hang on for too
8240 : * long with a marginal case. (Space target is only truly
8241 : * helpful when it allows us to recognize that we don't need
8242 : * to access more than 1 or 2 blocks to satisfy caller due to
8243 : * agreeable workload characteristics.)
8244 : *
8245 : * We are a bit more patient when we encounter contiguous
8246 : * blocks, though: these are treated as favorable blocks. The
8247 : * decay process is only applied when the next block in line
8248 : * is not a favorable/contiguous block. This is not an
8249 : * exception to the general rule; we still insist on finding
8250 : * at least one deletable item per block accessed. See
8251 : * bottomup_nblocksfavorable() for full details of the theory
8252 : * behind favorable blocks and heap block locality in general.
8253 : *
8254 : * Note: The first block in line is always treated as a
8255 : * favorable block, so the earliest possible point that the
8256 : * decay can be applied is just before we access the second
8257 : * block in line. The Assert() verifies this for us.
8258 : */
8259 : Assert(nblocksaccessed > 0 || nblocksfavorable > 0);
8260 4466 : if (nblocksfavorable > 0)
8261 4028 : nblocksfavorable--;
8262 : else
8263 438 : curtargetfreespace /= 2;
8264 : }
8265 :
8266 : /* release old buffer */
8267 25870 : if (BufferIsValid(buf))
8268 14918 : UnlockReleaseBuffer(buf);
8269 :
8270 25870 : blkno = ItemPointerGetBlockNumber(htid);
8271 25870 : buf = ReadBuffer(rel, blkno);
8272 25870 : nblocksaccessed++;
8273 : Assert(!delstate->bottomup ||
8274 : nblocksaccessed <= BOTTOMUP_MAX_NBLOCKS);
8275 :
8276 : #ifdef USE_PREFETCH
8277 :
8278 : /*
8279 : * To maintain the prefetch distance, prefetch one more page for
8280 : * each page we read.
8281 : */
8282 25870 : index_delete_prefetch_buffer(rel, &prefetch_state, 1);
8283 : #endif
8284 :
8285 25870 : LockBuffer(buf, BUFFER_LOCK_SHARE);
8286 :
8287 25870 : page = BufferGetPage(buf);
8288 25870 : maxoff = PageGetMaxOffsetNumber(page);
8289 : }
8290 :
8291 : /*
8292 : * In passing, detect index corruption involving an index page with a
8293 : * TID that points to a location in the heap that couldn't possibly be
8294 : * correct. We only do this with actual TIDs from caller's index page
8295 : * (not items reached by traversing through a HOT chain).
8296 : */
8297 1061266 : index_delete_check_htid(delstate, page, maxoff, htid, istatus);
8298 :
8299 1061266 : if (istatus->knowndeletable)
8300 : Assert(!delstate->bottomup && !istatus->promising);
8301 : else
8302 : {
8303 800242 : ItemPointerData tmp = *htid;
8304 : HeapTupleData heapTuple;
8305 :
8306 : /* Are any tuples from this HOT chain non-vacuumable? */
8307 800242 : if (heap_hot_search_buffer(&tmp, rel, buf, &SnapshotNonVacuumable,
8308 : &heapTuple, NULL, true))
8309 489294 : continue; /* can't delete entry */
8310 :
8311 : /* Caller will delete, since whole HOT chain is vacuumable */
8312 310948 : istatus->knowndeletable = true;
8313 :
8314 : /* Maintain index free space info for bottom-up deletion case */
8315 310948 : if (delstate->bottomup)
8316 : {
8317 : Assert(istatus->freespace > 0);
8318 12838 : actualfreespace += istatus->freespace;
8319 12838 : if (actualfreespace >= curtargetfreespace)
8320 4458 : bottomup_final_block = true;
8321 : }
8322 : }
8323 :
8324 : /*
8325 : * Maintain snapshotConflictHorizon value for deletion operation as a
8326 : * whole by advancing current value using heap tuple headers. This is
8327 : * loosely based on the logic for pruning a HOT chain.
8328 : */
8329 571972 : offnum = ItemPointerGetOffsetNumber(htid);
8330 571972 : priorXmax = InvalidTransactionId; /* cannot check first XMIN */
8331 : for (;;)
8332 41014 : {
8333 : ItemId lp;
8334 : HeapTupleHeader htup;
8335 :
8336 : /* Sanity check (pure paranoia) */
8337 612986 : if (offnum < FirstOffsetNumber)
8338 0 : break;
8339 :
8340 : /*
8341 : * An offset past the end of page's line pointer array is possible
8342 : * when the array was truncated
8343 : */
8344 612986 : if (offnum > maxoff)
8345 0 : break;
8346 :
8347 612986 : lp = PageGetItemId(page, offnum);
8348 612986 : if (ItemIdIsRedirected(lp))
8349 : {
8350 18170 : offnum = ItemIdGetRedirect(lp);
8351 18170 : continue;
8352 : }
8353 :
8354 : /*
8355 : * We'll often encounter LP_DEAD line pointers (especially with an
8356 : * entry marked knowndeletable by our caller up front). No heap
8357 : * tuple headers get examined for an htid that leads us to an
8358 : * LP_DEAD item. This is okay because the earlier pruning
8359 : * operation that made the line pointer LP_DEAD in the first place
8360 : * must have considered the original tuple header as part of
8361 : * generating its own snapshotConflictHorizon value.
8362 : *
8363 : * Relying on XLOG_HEAP2_PRUNE_VACUUM_SCAN records like this is
8364 : * the same strategy that index vacuuming uses in all cases. Index
8365 : * VACUUM WAL records don't even have a snapshotConflictHorizon
8366 : * field of their own for this reason.
8367 : */
8368 594816 : if (!ItemIdIsNormal(lp))
8369 380570 : break;
8370 :
8371 214246 : htup = (HeapTupleHeader) PageGetItem(page, lp);
8372 :
8373 : /*
8374 : * Check the tuple XMIN against prior XMAX, if any
8375 : */
8376 237090 : if (TransactionIdIsValid(priorXmax) &&
8377 22844 : !TransactionIdEquals(HeapTupleHeaderGetXmin(htup), priorXmax))
8378 0 : break;
8379 :
8380 214246 : HeapTupleHeaderAdvanceConflictHorizon(htup,
8381 : &snapshotConflictHorizon);
8382 :
8383 : /*
8384 : * If the tuple is not HOT-updated, then we are at the end of this
8385 : * HOT-chain. No need to visit later tuples from the same update
8386 : * chain (they get their own index entries) -- just move on to
8387 : * next htid from index AM caller.
8388 : */
8389 214246 : if (!HeapTupleHeaderIsHotUpdated(htup))
8390 191402 : break;
8391 :
8392 : /* Advance to next HOT chain member */
8393 : Assert(ItemPointerGetBlockNumber(&htup->t_ctid) == blkno);
8394 22844 : offnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
8395 22844 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
8396 : }
8397 :
8398 : /* Enable further/final shrinking of deltids for caller */
8399 571972 : finalndeltids = i + 1;
8400 : }
8401 :
8402 10952 : UnlockReleaseBuffer(buf);
8403 :
8404 : /*
8405 : * Shrink deltids array to exclude non-deletable entries at the end. This
8406 : * is not just a minor optimization. Final deltids array size might be
8407 : * zero for a bottom-up caller. Index AM is explicitly allowed to rely on
8408 : * ndeltids being zero in all cases with zero total deletable entries.
8409 : */
8410 : Assert(finalndeltids > 0 || delstate->bottomup);
8411 10952 : delstate->ndeltids = finalndeltids;
8412 :
8413 10952 : return snapshotConflictHorizon;
8414 : }
8415 :
8416 : /*
8417 : * Specialized inlineable comparison function for index_delete_sort()
8418 : */
8419 : static inline int
8420 24317898 : index_delete_sort_cmp(TM_IndexDelete *deltid1, TM_IndexDelete *deltid2)
8421 : {
8422 24317898 : ItemPointer tid1 = &deltid1->tid;
8423 24317898 : ItemPointer tid2 = &deltid2->tid;
8424 :
8425 : {
8426 24317898 : BlockNumber blk1 = ItemPointerGetBlockNumber(tid1);
8427 24317898 : BlockNumber blk2 = ItemPointerGetBlockNumber(tid2);
8428 :
8429 24317898 : if (blk1 != blk2)
8430 9823432 : return (blk1 < blk2) ? -1 : 1;
8431 : }
8432 : {
8433 14494466 : OffsetNumber pos1 = ItemPointerGetOffsetNumber(tid1);
8434 14494466 : OffsetNumber pos2 = ItemPointerGetOffsetNumber(tid2);
8435 :
8436 14494466 : if (pos1 != pos2)
8437 14494466 : return (pos1 < pos2) ? -1 : 1;
8438 : }
8439 :
8440 : Assert(false);
8441 :
8442 0 : return 0;
8443 : }
8444 :
8445 : /*
8446 : * Sort deltids array from delstate by TID. This prepares it for further
8447 : * processing by heap_index_delete_tuples().
8448 : *
8449 : * This operation becomes a noticeable consumer of CPU cycles with some
8450 : * workloads, so we go to the trouble of specialization/micro optimization.
8451 : * We use shellsort for this because it's easy to specialize, compiles to
8452 : * relatively few instructions, and is adaptive to presorted inputs/subsets
8453 : * (which are typical here).
8454 : */
8455 : static void
8456 10952 : index_delete_sort(TM_IndexDeleteOp *delstate)
8457 : {
8458 10952 : TM_IndexDelete *deltids = delstate->deltids;
8459 10952 : int ndeltids = delstate->ndeltids;
8460 :
8461 : /*
8462 : * Shellsort gap sequence (taken from Sedgewick-Incerpi paper).
8463 : *
8464 : * This implementation is fast with array sizes up to ~4500. This covers
8465 : * all supported BLCKSZ values.
8466 : */
8467 10952 : const int gaps[9] = {1968, 861, 336, 112, 48, 21, 7, 3, 1};
8468 :
8469 : /* Think carefully before changing anything here -- keep swaps cheap */
8470 : StaticAssertDecl(sizeof(TM_IndexDelete) <= 8,
8471 : "element size exceeds 8 bytes");
8472 :
8473 109520 : for (int g = 0; g < lengthof(gaps); g++)
8474 : {
8475 14568230 : for (int hi = gaps[g], i = hi; i < ndeltids; i++)
8476 : {
8477 14469662 : TM_IndexDelete d = deltids[i];
8478 14469662 : int j = i;
8479 :
8480 25008940 : while (j >= hi && index_delete_sort_cmp(&deltids[j - hi], &d) >= 0)
8481 : {
8482 10539278 : deltids[j] = deltids[j - hi];
8483 10539278 : j -= hi;
8484 : }
8485 14469662 : deltids[j] = d;
8486 : }
8487 : }
8488 10952 : }
8489 :
8490 : /*
8491 : * Returns how many blocks should be considered favorable/contiguous for a
8492 : * bottom-up index deletion pass. This is a number of heap blocks that starts
8493 : * from and includes the first block in line.
8494 : *
8495 : * There is always at least one favorable block during bottom-up index
8496 : * deletion. In the worst case (i.e. with totally random heap blocks) the
8497 : * first block in line (the only favorable block) can be thought of as a
8498 : * degenerate array of contiguous blocks that consists of a single block.
8499 : * heap_index_delete_tuples() will expect this.
8500 : *
8501 : * Caller passes blockgroups, a description of the final order that deltids
8502 : * will be sorted in for heap_index_delete_tuples() bottom-up index deletion
8503 : * processing. Note that deltids need not actually be sorted just yet (caller
8504 : * only passes deltids to us so that we can interpret blockgroups).
8505 : *
8506 : * You might guess that the existence of contiguous blocks cannot matter much,
8507 : * since in general the main factor that determines which blocks we visit is
8508 : * the number of promising TIDs, which is a fixed hint from the index AM.
8509 : * We're not really targeting the general case, though -- the actual goal is
8510 : * to adapt our behavior to a wide variety of naturally occurring conditions.
8511 : * The effects of most of the heuristics we apply are only noticeable in the
8512 : * aggregate, over time and across many _related_ bottom-up index deletion
8513 : * passes.
8514 : *
8515 : * Deeming certain blocks favorable allows heapam to recognize and adapt to
8516 : * workloads where heap blocks visited during bottom-up index deletion can be
8517 : * accessed contiguously, in the sense that each newly visited block is the
8518 : * neighbor of the block that bottom-up deletion just finished processing (or
8519 : * close enough to it). It will likely be cheaper to access more favorable
8520 : * blocks sooner rather than later (e.g. in this pass, not across a series of
8521 : * related bottom-up passes). Either way it is probably only a matter of time
8522 : * (or a matter of further correlated version churn) before all blocks that
8523 : * appear together as a single large batch of favorable blocks get accessed by
8524 : * _some_ bottom-up pass. Large batches of favorable blocks tend to either
8525 : * appear almost constantly or not even once (it all depends on per-index
8526 : * workload characteristics).
8527 : *
8528 : * Note that the blockgroups sort order applies a power-of-two bucketing
8529 : * scheme that creates opportunities for contiguous groups of blocks to get
8530 : * batched together, at least with workloads that are naturally amenable to
8531 : * being driven by heap block locality. This doesn't just enhance the spatial
8532 : * locality of bottom-up heap block processing in the obvious way. It also
8533 : * enables temporal locality of access, since sorting by heap block number
8534 : * naturally tends to make the bottom-up processing order deterministic.
8535 : *
8536 : * Consider the following example to get a sense of how temporal locality
8537 : * might matter: There is a heap relation with several indexes, each of which
8538 : * is low to medium cardinality. It is subject to constant non-HOT updates.
8539 : * The updates are skewed (in one part of the primary key, perhaps). None of
8540 : * the indexes are logically modified by the UPDATE statements (if they were
8541 : * then bottom-up index deletion would not be triggered in the first place).
8542 : * Naturally, each new round of index tuples (for each heap tuple that gets a
8543 : * heap_update() call) will have the same heap TID in each and every index.
8544 : * Since these indexes are low cardinality and never get logically modified,
8545 : * heapam processing during bottom-up deletion passes will access heap blocks
8546 : * in approximately sequential order. Temporal locality of access occurs due
8547 : * to bottom-up deletion passes behaving very similarly across each of the
8548 : * indexes at any given moment. This keeps the number of buffer misses needed
8549 : * to visit heap blocks to a minimum.
8550 : */
8551 : static int
8552 3752 : bottomup_nblocksfavorable(IndexDeleteCounts *blockgroups, int nblockgroups,
8553 : TM_IndexDelete *deltids)
8554 : {
8555 3752 : int64 lastblock = -1;
8556 3752 : int nblocksfavorable = 0;
8557 :
8558 : Assert(nblockgroups >= 1);
8559 : Assert(nblockgroups <= BOTTOMUP_MAX_NBLOCKS);
8560 :
8561 : /*
8562 : * We tolerate heap blocks that will be accessed only slightly out of
8563 : * physical order. Small blips occur when a pair of almost-contiguous
8564 : * blocks happen to fall into different buckets (perhaps due only to a
8565 : * small difference in npromisingtids that the bucketing scheme didn't
8566 : * quite manage to ignore). We effectively ignore these blips by applying
8567 : * a small tolerance. The precise tolerance we use is a little arbitrary,
8568 : * but it works well enough in practice.
8569 : */
8570 11202 : for (int b = 0; b < nblockgroups; b++)
8571 : {
8572 10772 : IndexDeleteCounts *group = blockgroups + b;
8573 10772 : TM_IndexDelete *firstdtid = deltids + group->ifirsttid;
8574 10772 : BlockNumber block = ItemPointerGetBlockNumber(&firstdtid->tid);
8575 :
8576 10772 : if (lastblock != -1 &&
8577 7020 : ((int64) block < lastblock - BOTTOMUP_TOLERANCE_NBLOCKS ||
8578 5764 : (int64) block > lastblock + BOTTOMUP_TOLERANCE_NBLOCKS))
8579 : break;
8580 :
8581 7450 : nblocksfavorable++;
8582 7450 : lastblock = block;
8583 : }
8584 :
8585 : /* Always indicate that there is at least 1 favorable block */
8586 : Assert(nblocksfavorable >= 1);
8587 :
8588 3752 : return nblocksfavorable;
8589 : }
8590 :
8591 : /*
8592 : * qsort comparison function for bottomup_sort_and_shrink()
8593 : */
8594 : static int
8595 381368 : bottomup_sort_and_shrink_cmp(const void *arg1, const void *arg2)
8596 : {
8597 381368 : const IndexDeleteCounts *group1 = (const IndexDeleteCounts *) arg1;
8598 381368 : const IndexDeleteCounts *group2 = (const IndexDeleteCounts *) arg2;
8599 :
8600 : /*
8601 : * Most significant field is npromisingtids (which we invert the order of
8602 : * so as to sort in desc order).
8603 : *
8604 : * Caller should have already normalized npromisingtids fields into
8605 : * power-of-two values (buckets).
8606 : */
8607 381368 : if (group1->npromisingtids > group2->npromisingtids)
8608 18118 : return -1;
8609 363250 : if (group1->npromisingtids < group2->npromisingtids)
8610 21484 : return 1;
8611 :
8612 : /*
8613 : * Tiebreak: desc ntids sort order.
8614 : *
8615 : * We cannot expect power-of-two values for ntids fields. We should
8616 : * behave as if they were already rounded up for us instead.
8617 : */
8618 341766 : if (group1->ntids != group2->ntids)
8619 : {
8620 237566 : uint32 ntids1 = pg_nextpower2_32((uint32) group1->ntids);
8621 237566 : uint32 ntids2 = pg_nextpower2_32((uint32) group2->ntids);
8622 :
8623 237566 : if (ntids1 > ntids2)
8624 38358 : return -1;
8625 199208 : if (ntids1 < ntids2)
8626 45346 : return 1;
8627 : }
8628 :
8629 : /*
8630 : * Tiebreak: asc offset-into-deltids-for-block (offset to first TID for
8631 : * block in deltids array) order.
8632 : *
8633 : * This is equivalent to sorting in ascending heap block number order
8634 : * (among otherwise equal subsets of the array). This approach allows us
8635 : * to avoid accessing the out-of-line TID. (We rely on the assumption
8636 : * that the deltids array was sorted in ascending heap TID order when
8637 : * these offsets to the first TID from each heap block group were formed.)
8638 : */
8639 258062 : if (group1->ifirsttid > group2->ifirsttid)
8640 130362 : return 1;
8641 127700 : if (group1->ifirsttid < group2->ifirsttid)
8642 127700 : return -1;
8643 :
8644 0 : pg_unreachable();
8645 :
8646 : return 0;
8647 : }
8648 :
8649 : /*
8650 : * heap_index_delete_tuples() helper function for bottom-up deletion callers.
8651 : *
8652 : * Sorts deltids array in the order needed for useful processing by bottom-up
8653 : * deletion. The array should already be sorted in TID order when we're
8654 : * called. The sort process groups heap TIDs from deltids into heap block
8655 : * groupings. Earlier/more-promising groups/blocks are usually those that are
8656 : * known to have the most "promising" TIDs.
8657 : *
8658 : * Sets new size of deltids array (ndeltids) in state. deltids will only have
8659 : * TIDs from the BOTTOMUP_MAX_NBLOCKS most promising heap blocks when we
8660 : * return. This often means that deltids will be shrunk to a small fraction
8661 : * of its original size (we eliminate many heap blocks from consideration for
8662 : * caller up front).
8663 : *
8664 : * Returns the number of "favorable" blocks. See bottomup_nblocksfavorable()
8665 : * for a definition and full details.
8666 : */
8667 : static int
8668 3752 : bottomup_sort_and_shrink(TM_IndexDeleteOp *delstate)
8669 : {
8670 : IndexDeleteCounts *blockgroups;
8671 : TM_IndexDelete *reordereddeltids;
8672 3752 : BlockNumber curblock = InvalidBlockNumber;
8673 3752 : int nblockgroups = 0;
8674 3752 : int ncopied = 0;
8675 3752 : int nblocksfavorable = 0;
8676 :
8677 : Assert(delstate->bottomup);
8678 : Assert(delstate->ndeltids > 0);
8679 :
8680 : /* Calculate per-heap-block count of TIDs */
8681 3752 : blockgroups = palloc(sizeof(IndexDeleteCounts) * delstate->ndeltids);
8682 1785946 : for (int i = 0; i < delstate->ndeltids; i++)
8683 : {
8684 1782194 : TM_IndexDelete *ideltid = &delstate->deltids[i];
8685 1782194 : TM_IndexStatus *istatus = delstate->status + ideltid->id;
8686 1782194 : ItemPointer htid = &ideltid->tid;
8687 1782194 : bool promising = istatus->promising;
8688 :
8689 1782194 : if (curblock != ItemPointerGetBlockNumber(htid))
8690 : {
8691 : /* New block group */
8692 73706 : nblockgroups++;
8693 :
8694 : Assert(curblock < ItemPointerGetBlockNumber(htid) ||
8695 : !BlockNumberIsValid(curblock));
8696 :
8697 73706 : curblock = ItemPointerGetBlockNumber(htid);
8698 73706 : blockgroups[nblockgroups - 1].ifirsttid = i;
8699 73706 : blockgroups[nblockgroups - 1].ntids = 1;
8700 73706 : blockgroups[nblockgroups - 1].npromisingtids = 0;
8701 : }
8702 : else
8703 : {
8704 1708488 : blockgroups[nblockgroups - 1].ntids++;
8705 : }
8706 :
8707 1782194 : if (promising)
8708 239646 : blockgroups[nblockgroups - 1].npromisingtids++;
8709 : }
8710 :
8711 : /*
8712 : * We're about ready to sort block groups to determine the optimal order
8713 : * for visiting heap blocks. But before we do, round the number of
8714 : * promising tuples for each block group up to the next power-of-two,
8715 : * unless it is very low (less than 4), in which case we round up to 4.
8716 : * npromisingtids is far too noisy to trust when choosing between a pair
8717 : * of block groups that both have very low values.
8718 : *
8719 : * This scheme divides heap blocks/block groups into buckets. Each bucket
8720 : * contains blocks that have _approximately_ the same number of promising
8721 : * TIDs as each other. The goal is to ignore relatively small differences
8722 : * in the total number of promising entries, so that the whole process can
8723 : * give a little weight to heapam factors (like heap block locality)
8724 : * instead. This isn't a trade-off, really -- we have nothing to lose. It
8725 : * would be foolish to interpret small differences in npromisingtids
8726 : * values as anything more than noise.
8727 : *
8728 : * We tiebreak on nhtids when sorting block group subsets that have the
8729 : * same npromisingtids, but this has the same issues as npromisingtids,
8730 : * and so nhtids is subject to the same power-of-two bucketing scheme. The
8731 : * only reason that we don't fix nhtids in the same way here too is that
8732 : * we'll need accurate nhtids values after the sort. We handle nhtids
8733 : * bucketization dynamically instead (in the sort comparator).
8734 : *
8735 : * See bottomup_nblocksfavorable() for a full explanation of when and how
8736 : * heap locality/favorable blocks can significantly influence when and how
8737 : * heap blocks are accessed.
8738 : */
8739 77458 : for (int b = 0; b < nblockgroups; b++)
8740 : {
8741 73706 : IndexDeleteCounts *group = blockgroups + b;
8742 :
8743 : /* Better off falling back on nhtids with low npromisingtids */
8744 73706 : if (group->npromisingtids <= 4)
8745 62564 : group->npromisingtids = 4;
8746 : else
8747 11142 : group->npromisingtids =
8748 11142 : pg_nextpower2_32((uint32) group->npromisingtids);
8749 : }
8750 :
8751 : /* Sort groups and rearrange caller's deltids array */
8752 3752 : qsort(blockgroups, nblockgroups, sizeof(IndexDeleteCounts),
8753 : bottomup_sort_and_shrink_cmp);
8754 3752 : reordereddeltids = palloc(delstate->ndeltids * sizeof(TM_IndexDelete));
8755 :
8756 3752 : nblockgroups = Min(BOTTOMUP_MAX_NBLOCKS, nblockgroups);
8757 : /* Determine number of favorable blocks at the start of final deltids */
8758 3752 : nblocksfavorable = bottomup_nblocksfavorable(blockgroups, nblockgroups,
8759 : delstate->deltids);
8760 :
8761 25044 : for (int b = 0; b < nblockgroups; b++)
8762 : {
8763 21292 : IndexDeleteCounts *group = blockgroups + b;
8764 21292 : TM_IndexDelete *firstdtid = delstate->deltids + group->ifirsttid;
8765 :
8766 21292 : memcpy(reordereddeltids + ncopied, firstdtid,
8767 21292 : sizeof(TM_IndexDelete) * group->ntids);
8768 21292 : ncopied += group->ntids;
8769 : }
8770 :
8771 : /* Copy final grouped and sorted TIDs back into start of caller's array */
8772 3752 : memcpy(delstate->deltids, reordereddeltids,
8773 : sizeof(TM_IndexDelete) * ncopied);
8774 3752 : delstate->ndeltids = ncopied;
8775 :
8776 3752 : pfree(reordereddeltids);
8777 3752 : pfree(blockgroups);
8778 :
8779 3752 : return nblocksfavorable;
8780 : }
8781 :
8782 : /*
8783 : * Perform XLogInsert for a heap-visible operation. 'block' is the block
8784 : * being marked all-visible, and vm_buffer is the buffer containing the
8785 : * corresponding visibility map block. Both should have already been modified
8786 : * and dirtied.
8787 : *
8788 : * snapshotConflictHorizon comes from the largest xmin on the page being
8789 : * marked all-visible. REDO routine uses it to generate recovery conflicts.
8790 : *
8791 : * If checksums or wal_log_hints are enabled, we may also generate a full-page
8792 : * image of heap_buffer. Otherwise, we optimize away the FPI (by specifying
8793 : * REGBUF_NO_IMAGE for the heap buffer), in which case the caller should *not*
8794 : * update the heap page's LSN.
8795 : */
8796 : XLogRecPtr
8797 91288 : log_heap_visible(Relation rel, Buffer heap_buffer, Buffer vm_buffer,
8798 : TransactionId snapshotConflictHorizon, uint8 vmflags)
8799 : {
8800 : xl_heap_visible xlrec;
8801 : XLogRecPtr recptr;
8802 : uint8 flags;
8803 :
8804 : Assert(BufferIsValid(heap_buffer));
8805 : Assert(BufferIsValid(vm_buffer));
8806 :
8807 91288 : xlrec.snapshotConflictHorizon = snapshotConflictHorizon;
8808 91288 : xlrec.flags = vmflags;
8809 91288 : if (RelationIsAccessibleInLogicalDecoding(rel))
8810 260 : xlrec.flags |= VISIBILITYMAP_XLOG_CATALOG_REL;
8811 91288 : XLogBeginInsert();
8812 91288 : XLogRegisterData(&xlrec, SizeOfHeapVisible);
8813 :
8814 91288 : XLogRegisterBuffer(0, vm_buffer, 0);
8815 :
8816 91288 : flags = REGBUF_STANDARD;
8817 91288 : if (!XLogHintBitIsNeeded())
8818 5890 : flags |= REGBUF_NO_IMAGE;
8819 91288 : XLogRegisterBuffer(1, heap_buffer, flags);
8820 :
8821 91288 : recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_VISIBLE);
8822 :
8823 91288 : return recptr;
8824 : }
8825 :
8826 : /*
8827 : * Perform XLogInsert for a heap-update operation. Caller must already
8828 : * have modified the buffer(s) and marked them dirty.
8829 : */
8830 : static XLogRecPtr
8831 586288 : log_heap_update(Relation reln, Buffer oldbuf,
8832 : Buffer newbuf, HeapTuple oldtup, HeapTuple newtup,
8833 : HeapTuple old_key_tuple,
8834 : bool all_visible_cleared, bool new_all_visible_cleared)
8835 : {
8836 : xl_heap_update xlrec;
8837 : xl_heap_header xlhdr;
8838 : xl_heap_header xlhdr_idx;
8839 : uint8 info;
8840 : uint16 prefix_suffix[2];
8841 586288 : uint16 prefixlen = 0,
8842 586288 : suffixlen = 0;
8843 : XLogRecPtr recptr;
8844 586288 : Page page = BufferGetPage(newbuf);
8845 586288 : bool need_tuple_data = RelationIsLogicallyLogged(reln);
8846 : bool init;
8847 : int bufflags;
8848 :
8849 : /* Caller should not call me on a non-WAL-logged relation */
8850 : Assert(RelationNeedsWAL(reln));
8851 :
8852 586288 : XLogBeginInsert();
8853 :
8854 586288 : if (HeapTupleIsHeapOnly(newtup))
8855 281392 : info = XLOG_HEAP_HOT_UPDATE;
8856 : else
8857 304896 : info = XLOG_HEAP_UPDATE;
8858 :
8859 : /*
8860 : * If the old and new tuple are on the same page, we only need to log the
8861 : * parts of the new tuple that were changed. That saves on the amount of
8862 : * WAL we need to write. Currently, we just count any unchanged bytes in
8863 : * the beginning and end of the tuple. That's quick to check, and
8864 : * perfectly covers the common case that only one field is updated.
8865 : *
8866 : * We could do this even if the old and new tuple are on different pages,
8867 : * but only if we don't make a full-page image of the old page, which is
8868 : * difficult to know in advance. Also, if the old tuple is corrupt for
8869 : * some reason, it would allow the corruption to propagate the new page,
8870 : * so it seems best to avoid. Under the general assumption that most
8871 : * updates tend to create the new tuple version on the same page, there
8872 : * isn't much to be gained by doing this across pages anyway.
8873 : *
8874 : * Skip this if we're taking a full-page image of the new page, as we
8875 : * don't include the new tuple in the WAL record in that case. Also
8876 : * disable if wal_level='logical', as logical decoding needs to be able to
8877 : * read the new tuple in whole from the WAL record alone.
8878 : */
8879 586288 : if (oldbuf == newbuf && !need_tuple_data &&
8880 283358 : !XLogCheckBufferNeedsBackup(newbuf))
8881 : {
8882 282066 : char *oldp = (char *) oldtup->t_data + oldtup->t_data->t_hoff;
8883 282066 : char *newp = (char *) newtup->t_data + newtup->t_data->t_hoff;
8884 282066 : int oldlen = oldtup->t_len - oldtup->t_data->t_hoff;
8885 282066 : int newlen = newtup->t_len - newtup->t_data->t_hoff;
8886 :
8887 : /* Check for common prefix between old and new tuple */
8888 23023990 : for (prefixlen = 0; prefixlen < Min(oldlen, newlen); prefixlen++)
8889 : {
8890 22974834 : if (newp[prefixlen] != oldp[prefixlen])
8891 232910 : break;
8892 : }
8893 :
8894 : /*
8895 : * Storing the length of the prefix takes 2 bytes, so we need to save
8896 : * at least 3 bytes or there's no point.
8897 : */
8898 282066 : if (prefixlen < 3)
8899 44134 : prefixlen = 0;
8900 :
8901 : /* Same for suffix */
8902 9240146 : for (suffixlen = 0; suffixlen < Min(oldlen, newlen) - prefixlen; suffixlen++)
8903 : {
8904 9190498 : if (newp[newlen - suffixlen - 1] != oldp[oldlen - suffixlen - 1])
8905 232418 : break;
8906 : }
8907 282066 : if (suffixlen < 3)
8908 68246 : suffixlen = 0;
8909 : }
8910 :
8911 : /* Prepare main WAL data chain */
8912 586288 : xlrec.flags = 0;
8913 586288 : if (all_visible_cleared)
8914 2990 : xlrec.flags |= XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED;
8915 586288 : if (new_all_visible_cleared)
8916 1634 : xlrec.flags |= XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED;
8917 586288 : if (prefixlen > 0)
8918 237932 : xlrec.flags |= XLH_UPDATE_PREFIX_FROM_OLD;
8919 586288 : if (suffixlen > 0)
8920 213820 : xlrec.flags |= XLH_UPDATE_SUFFIX_FROM_OLD;
8921 586288 : if (need_tuple_data)
8922 : {
8923 94040 : xlrec.flags |= XLH_UPDATE_CONTAINS_NEW_TUPLE;
8924 94040 : if (old_key_tuple)
8925 : {
8926 286 : if (reln->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
8927 126 : xlrec.flags |= XLH_UPDATE_CONTAINS_OLD_TUPLE;
8928 : else
8929 160 : xlrec.flags |= XLH_UPDATE_CONTAINS_OLD_KEY;
8930 : }
8931 : }
8932 :
8933 : /* If new tuple is the single and first tuple on page... */
8934 593348 : if (ItemPointerGetOffsetNumber(&(newtup->t_self)) == FirstOffsetNumber &&
8935 7060 : PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
8936 : {
8937 6622 : info |= XLOG_HEAP_INIT_PAGE;
8938 6622 : init = true;
8939 : }
8940 : else
8941 579666 : init = false;
8942 :
8943 : /* Prepare WAL data for the old page */
8944 586288 : xlrec.old_offnum = ItemPointerGetOffsetNumber(&oldtup->t_self);
8945 586288 : xlrec.old_xmax = HeapTupleHeaderGetRawXmax(oldtup->t_data);
8946 1172576 : xlrec.old_infobits_set = compute_infobits(oldtup->t_data->t_infomask,
8947 586288 : oldtup->t_data->t_infomask2);
8948 :
8949 : /* Prepare WAL data for the new page */
8950 586288 : xlrec.new_offnum = ItemPointerGetOffsetNumber(&newtup->t_self);
8951 586288 : xlrec.new_xmax = HeapTupleHeaderGetRawXmax(newtup->t_data);
8952 :
8953 586288 : bufflags = REGBUF_STANDARD;
8954 586288 : if (init)
8955 6622 : bufflags |= REGBUF_WILL_INIT;
8956 586288 : if (need_tuple_data)
8957 94040 : bufflags |= REGBUF_KEEP_DATA;
8958 :
8959 586288 : XLogRegisterBuffer(0, newbuf, bufflags);
8960 586288 : if (oldbuf != newbuf)
8961 279050 : XLogRegisterBuffer(1, oldbuf, REGBUF_STANDARD);
8962 :
8963 586288 : XLogRegisterData(&xlrec, SizeOfHeapUpdate);
8964 :
8965 : /*
8966 : * Prepare WAL data for the new tuple.
8967 : */
8968 586288 : if (prefixlen > 0 || suffixlen > 0)
8969 : {
8970 281154 : if (prefixlen > 0 && suffixlen > 0)
8971 : {
8972 170598 : prefix_suffix[0] = prefixlen;
8973 170598 : prefix_suffix[1] = suffixlen;
8974 170598 : XLogRegisterBufData(0, &prefix_suffix, sizeof(uint16) * 2);
8975 : }
8976 110556 : else if (prefixlen > 0)
8977 : {
8978 67334 : XLogRegisterBufData(0, &prefixlen, sizeof(uint16));
8979 : }
8980 : else
8981 : {
8982 43222 : XLogRegisterBufData(0, &suffixlen, sizeof(uint16));
8983 : }
8984 : }
8985 :
8986 586288 : xlhdr.t_infomask2 = newtup->t_data->t_infomask2;
8987 586288 : xlhdr.t_infomask = newtup->t_data->t_infomask;
8988 586288 : xlhdr.t_hoff = newtup->t_data->t_hoff;
8989 : Assert(SizeofHeapTupleHeader + prefixlen + suffixlen <= newtup->t_len);
8990 :
8991 : /*
8992 : * PG73FORMAT: write bitmap [+ padding] [+ oid] + data
8993 : *
8994 : * The 'data' doesn't include the common prefix or suffix.
8995 : */
8996 586288 : XLogRegisterBufData(0, &xlhdr, SizeOfHeapHeader);
8997 586288 : if (prefixlen == 0)
8998 : {
8999 348356 : XLogRegisterBufData(0,
9000 348356 : (char *) newtup->t_data + SizeofHeapTupleHeader,
9001 348356 : newtup->t_len - SizeofHeapTupleHeader - suffixlen);
9002 : }
9003 : else
9004 : {
9005 : /*
9006 : * Have to write the null bitmap and data after the common prefix as
9007 : * two separate rdata entries.
9008 : */
9009 : /* bitmap [+ padding] [+ oid] */
9010 237932 : if (newtup->t_data->t_hoff - SizeofHeapTupleHeader > 0)
9011 : {
9012 237932 : XLogRegisterBufData(0,
9013 237932 : (char *) newtup->t_data + SizeofHeapTupleHeader,
9014 237932 : newtup->t_data->t_hoff - SizeofHeapTupleHeader);
9015 : }
9016 :
9017 : /* data after common prefix */
9018 237932 : XLogRegisterBufData(0,
9019 237932 : (char *) newtup->t_data + newtup->t_data->t_hoff + prefixlen,
9020 237932 : newtup->t_len - newtup->t_data->t_hoff - prefixlen - suffixlen);
9021 : }
9022 :
9023 : /* We need to log a tuple identity */
9024 586288 : if (need_tuple_data && old_key_tuple)
9025 : {
9026 : /* don't really need this, but its more comfy to decode */
9027 286 : xlhdr_idx.t_infomask2 = old_key_tuple->t_data->t_infomask2;
9028 286 : xlhdr_idx.t_infomask = old_key_tuple->t_data->t_infomask;
9029 286 : xlhdr_idx.t_hoff = old_key_tuple->t_data->t_hoff;
9030 :
9031 286 : XLogRegisterData(&xlhdr_idx, SizeOfHeapHeader);
9032 :
9033 : /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
9034 286 : XLogRegisterData((char *) old_key_tuple->t_data + SizeofHeapTupleHeader,
9035 286 : old_key_tuple->t_len - SizeofHeapTupleHeader);
9036 : }
9037 :
9038 : /* filtering by origin on a row level is much more efficient */
9039 586288 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
9040 :
9041 586288 : recptr = XLogInsert(RM_HEAP_ID, info);
9042 :
9043 586288 : return recptr;
9044 : }
9045 :
9046 : /*
9047 : * Perform XLogInsert of an XLOG_HEAP2_NEW_CID record
9048 : *
9049 : * This is only used in wal_level >= WAL_LEVEL_LOGICAL, and only for catalog
9050 : * tuples.
9051 : */
9052 : static XLogRecPtr
9053 47924 : log_heap_new_cid(Relation relation, HeapTuple tup)
9054 : {
9055 : xl_heap_new_cid xlrec;
9056 :
9057 : XLogRecPtr recptr;
9058 47924 : HeapTupleHeader hdr = tup->t_data;
9059 :
9060 : Assert(ItemPointerIsValid(&tup->t_self));
9061 : Assert(tup->t_tableOid != InvalidOid);
9062 :
9063 47924 : xlrec.top_xid = GetTopTransactionId();
9064 47924 : xlrec.target_locator = relation->rd_locator;
9065 47924 : xlrec.target_tid = tup->t_self;
9066 :
9067 : /*
9068 : * If the tuple got inserted & deleted in the same TX we definitely have a
9069 : * combo CID, set cmin and cmax.
9070 : */
9071 47924 : if (hdr->t_infomask & HEAP_COMBOCID)
9072 : {
9073 : Assert(!(hdr->t_infomask & HEAP_XMAX_INVALID));
9074 : Assert(!HeapTupleHeaderXminInvalid(hdr));
9075 4038 : xlrec.cmin = HeapTupleHeaderGetCmin(hdr);
9076 4038 : xlrec.cmax = HeapTupleHeaderGetCmax(hdr);
9077 4038 : xlrec.combocid = HeapTupleHeaderGetRawCommandId(hdr);
9078 : }
9079 : /* No combo CID, so only cmin or cmax can be set by this TX */
9080 : else
9081 : {
9082 : /*
9083 : * Tuple inserted.
9084 : *
9085 : * We need to check for LOCK ONLY because multixacts might be
9086 : * transferred to the new tuple in case of FOR KEY SHARE updates in
9087 : * which case there will be an xmax, although the tuple just got
9088 : * inserted.
9089 : */
9090 57266 : if (hdr->t_infomask & HEAP_XMAX_INVALID ||
9091 13380 : HEAP_XMAX_IS_LOCKED_ONLY(hdr->t_infomask))
9092 : {
9093 30508 : xlrec.cmin = HeapTupleHeaderGetRawCommandId(hdr);
9094 30508 : xlrec.cmax = InvalidCommandId;
9095 : }
9096 : /* Tuple from a different tx updated or deleted. */
9097 : else
9098 : {
9099 13378 : xlrec.cmin = InvalidCommandId;
9100 13378 : xlrec.cmax = HeapTupleHeaderGetRawCommandId(hdr);
9101 : }
9102 43886 : xlrec.combocid = InvalidCommandId;
9103 : }
9104 :
9105 : /*
9106 : * Note that we don't need to register the buffer here, because this
9107 : * operation does not modify the page. The insert/update/delete that
9108 : * called us certainly did, but that's WAL-logged separately.
9109 : */
9110 47924 : XLogBeginInsert();
9111 47924 : XLogRegisterData(&xlrec, SizeOfHeapNewCid);
9112 :
9113 : /* will be looked at irrespective of origin */
9114 :
9115 47924 : recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_NEW_CID);
9116 :
9117 47924 : return recptr;
9118 : }
9119 :
9120 : /*
9121 : * Build a heap tuple representing the configured REPLICA IDENTITY to represent
9122 : * the old tuple in an UPDATE or DELETE.
9123 : *
9124 : * Returns NULL if there's no need to log an identity or if there's no suitable
9125 : * key defined.
9126 : *
9127 : * Pass key_required true if any replica identity columns changed value, or if
9128 : * any of them have any external data. Delete must always pass true.
9129 : *
9130 : * *copy is set to true if the returned tuple is a modified copy rather than
9131 : * the same tuple that was passed in.
9132 : */
9133 : static HeapTuple
9134 3561626 : ExtractReplicaIdentity(Relation relation, HeapTuple tp, bool key_required,
9135 : bool *copy)
9136 : {
9137 3561626 : TupleDesc desc = RelationGetDescr(relation);
9138 3561626 : char replident = relation->rd_rel->relreplident;
9139 : Bitmapset *idattrs;
9140 : HeapTuple key_tuple;
9141 : bool nulls[MaxHeapAttributeNumber];
9142 : Datum values[MaxHeapAttributeNumber];
9143 :
9144 3561626 : *copy = false;
9145 :
9146 3561626 : if (!RelationIsLogicallyLogged(relation))
9147 3361056 : return NULL;
9148 :
9149 200570 : if (replident == REPLICA_IDENTITY_NOTHING)
9150 462 : return NULL;
9151 :
9152 200108 : if (replident == REPLICA_IDENTITY_FULL)
9153 : {
9154 : /*
9155 : * When logging the entire old tuple, it very well could contain
9156 : * toasted columns. If so, force them to be inlined.
9157 : */
9158 382 : if (HeapTupleHasExternal(tp))
9159 : {
9160 8 : *copy = true;
9161 8 : tp = toast_flatten_tuple(tp, desc);
9162 : }
9163 382 : return tp;
9164 : }
9165 :
9166 : /* if the key isn't required and we're only logging the key, we're done */
9167 199726 : if (!key_required)
9168 93754 : return NULL;
9169 :
9170 : /* find out the replica identity columns */
9171 105972 : idattrs = RelationGetIndexAttrBitmap(relation,
9172 : INDEX_ATTR_BITMAP_IDENTITY_KEY);
9173 :
9174 : /*
9175 : * If there's no defined replica identity columns, treat as !key_required.
9176 : * (This case should not be reachable from heap_update, since that should
9177 : * calculate key_required accurately. But heap_delete just passes
9178 : * constant true for key_required, so we can hit this case in deletes.)
9179 : */
9180 105972 : if (bms_is_empty(idattrs))
9181 12042 : return NULL;
9182 :
9183 : /*
9184 : * Construct a new tuple containing only the replica identity columns,
9185 : * with nulls elsewhere. While we're at it, assert that the replica
9186 : * identity columns aren't null.
9187 : */
9188 93930 : heap_deform_tuple(tp, desc, values, nulls);
9189 :
9190 301776 : for (int i = 0; i < desc->natts; i++)
9191 : {
9192 207846 : if (bms_is_member(i + 1 - FirstLowInvalidHeapAttributeNumber,
9193 : idattrs))
9194 : Assert(!nulls[i]);
9195 : else
9196 113892 : nulls[i] = true;
9197 : }
9198 :
9199 93930 : key_tuple = heap_form_tuple(desc, values, nulls);
9200 93930 : *copy = true;
9201 :
9202 93930 : bms_free(idattrs);
9203 :
9204 : /*
9205 : * If the tuple, which by here only contains indexed columns, still has
9206 : * toasted columns, force them to be inlined. This is somewhat unlikely
9207 : * since there's limits on the size of indexed columns, so we don't
9208 : * duplicate toast_flatten_tuple()s functionality in the above loop over
9209 : * the indexed columns, even if it would be more efficient.
9210 : */
9211 93930 : if (HeapTupleHasExternal(key_tuple))
9212 : {
9213 8 : HeapTuple oldtup = key_tuple;
9214 :
9215 8 : key_tuple = toast_flatten_tuple(oldtup, desc);
9216 8 : heap_freetuple(oldtup);
9217 : }
9218 :
9219 93930 : return key_tuple;
9220 : }
9221 :
9222 : /*
9223 : * HeapCheckForSerializableConflictOut
9224 : * We are reading a tuple. If it's not visible, there may be a
9225 : * rw-conflict out with the inserter. Otherwise, if it is visible to us
9226 : * but has been deleted, there may be a rw-conflict out with the deleter.
9227 : *
9228 : * We will determine the top level xid of the writing transaction with which
9229 : * we may be in conflict, and ask CheckForSerializableConflictOut() to check
9230 : * for overlap with our own transaction.
9231 : *
9232 : * This function should be called just about anywhere in heapam.c where a
9233 : * tuple has been read. The caller must hold at least a shared lock on the
9234 : * buffer, because this function might set hint bits on the tuple. There is
9235 : * currently no known reason to call this function from an index AM.
9236 : */
9237 : void
9238 64207680 : HeapCheckForSerializableConflictOut(bool visible, Relation relation,
9239 : HeapTuple tuple, Buffer buffer,
9240 : Snapshot snapshot)
9241 : {
9242 : TransactionId xid;
9243 : HTSV_Result htsvResult;
9244 :
9245 64207680 : if (!CheckForSerializableConflictOutNeeded(relation, snapshot))
9246 64156990 : return;
9247 :
9248 : /*
9249 : * Check to see whether the tuple has been written to by a concurrent
9250 : * transaction, either to create it not visible to us, or to delete it
9251 : * while it is visible to us. The "visible" bool indicates whether the
9252 : * tuple is visible to us, while HeapTupleSatisfiesVacuum checks what else
9253 : * is going on with it.
9254 : *
9255 : * In the event of a concurrently inserted tuple that also happens to have
9256 : * been concurrently updated (by a separate transaction), the xmin of the
9257 : * tuple will be used -- not the updater's xid.
9258 : */
9259 50690 : htsvResult = HeapTupleSatisfiesVacuum(tuple, TransactionXmin, buffer);
9260 50690 : switch (htsvResult)
9261 : {
9262 49070 : case HEAPTUPLE_LIVE:
9263 49070 : if (visible)
9264 49044 : return;
9265 26 : xid = HeapTupleHeaderGetXmin(tuple->t_data);
9266 26 : break;
9267 704 : case HEAPTUPLE_RECENTLY_DEAD:
9268 : case HEAPTUPLE_DELETE_IN_PROGRESS:
9269 704 : if (visible)
9270 562 : xid = HeapTupleHeaderGetUpdateXid(tuple->t_data);
9271 : else
9272 142 : xid = HeapTupleHeaderGetXmin(tuple->t_data);
9273 :
9274 704 : if (TransactionIdPrecedes(xid, TransactionXmin))
9275 : {
9276 : /* This is like the HEAPTUPLE_DEAD case */
9277 : Assert(!visible);
9278 130 : return;
9279 : }
9280 574 : break;
9281 652 : case HEAPTUPLE_INSERT_IN_PROGRESS:
9282 652 : xid = HeapTupleHeaderGetXmin(tuple->t_data);
9283 652 : break;
9284 264 : case HEAPTUPLE_DEAD:
9285 : Assert(!visible);
9286 264 : return;
9287 0 : default:
9288 :
9289 : /*
9290 : * The only way to get to this default clause is if a new value is
9291 : * added to the enum type without adding it to this switch
9292 : * statement. That's a bug, so elog.
9293 : */
9294 0 : elog(ERROR, "unrecognized return value from HeapTupleSatisfiesVacuum: %u", htsvResult);
9295 :
9296 : /*
9297 : * In spite of having all enum values covered and calling elog on
9298 : * this default, some compilers think this is a code path which
9299 : * allows xid to be used below without initialization. Silence
9300 : * that warning.
9301 : */
9302 : xid = InvalidTransactionId;
9303 : }
9304 :
9305 : Assert(TransactionIdIsValid(xid));
9306 : Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
9307 :
9308 : /*
9309 : * Find top level xid. Bail out if xid is too early to be a conflict, or
9310 : * if it's our own xid.
9311 : */
9312 1252 : if (TransactionIdEquals(xid, GetTopTransactionIdIfAny()))
9313 124 : return;
9314 1128 : xid = SubTransGetTopmostTransaction(xid);
9315 1128 : if (TransactionIdPrecedes(xid, TransactionXmin))
9316 0 : return;
9317 :
9318 1128 : CheckForSerializableConflictOut(relation, xid, snapshot);
9319 : }
|