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 20926062 : 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 20926062 : }
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 201580 : heap_scan_stream_read_next_parallel(ReadStream *stream,
248 : void *callback_private_data,
249 : void *per_buffer_data)
250 : {
251 201580 : HeapScanDesc scan = (HeapScanDesc) callback_private_data;
252 :
253 : Assert(ScanDirectionIsForward(scan->rs_dir));
254 : Assert(scan->rs_base.rs_parallel);
255 :
256 201580 : if (unlikely(!scan->rs_inited))
257 : {
258 : /* parallel scan */
259 2998 : table_block_parallelscan_startblock_init(scan->rs_base.rs_rd,
260 2998 : scan->rs_parallelworkerdata,
261 2998 : (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel);
262 :
263 : /* may return InvalidBlockNumber if there are no more blocks */
264 5996 : scan->rs_prefetch_block = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
265 2998 : scan->rs_parallelworkerdata,
266 2998 : (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel);
267 2998 : 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 201580 : 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 7502362 : heap_scan_stream_read_next_serial(ReadStream *stream,
286 : void *callback_private_data,
287 : void *per_buffer_data)
288 : {
289 7502362 : HeapScanDesc scan = (HeapScanDesc) callback_private_data;
290 :
291 7502362 : if (unlikely(!scan->rs_inited))
292 : {
293 1913890 : scan->rs_prefetch_block = heapgettup_initial_block(scan, scan->rs_dir);
294 1913890 : scan->rs_inited = true;
295 : }
296 : else
297 5588472 : scan->rs_prefetch_block = heapgettup_advance_block(scan,
298 : scan->rs_prefetch_block,
299 : scan->rs_dir);
300 :
301 7502362 : 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 421086 : bitmapheap_stream_read_next(ReadStream *pgsr, void *private_data,
311 : void *per_buffer_data)
312 : {
313 421086 : TBMIterateResult *tbmres = per_buffer_data;
314 421086 : BitmapHeapScanDesc bscan = (BitmapHeapScanDesc) private_data;
315 421086 : HeapScanDesc hscan = (HeapScanDesc) bscan;
316 421086 : TableScanDesc sscan = &hscan->rs_base;
317 :
318 : for (;;)
319 : {
320 421086 : CHECK_FOR_INTERRUPTS();
321 :
322 : /* no more entries in the bitmap */
323 421086 : if (!tbm_iterate(&sscan->st.rs_tbmiterator, tbmres))
324 20396 : 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 400690 : if (!IsolationIsSerializable() &&
335 400484 : tbmres->blockno >= hscan->rs_nblocks)
336 0 : continue;
337 :
338 400690 : 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 1956462 : initscan(HeapScanDesc scan, ScanKey key, bool keep_startblock)
351 : {
352 1956462 : 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 1956462 : 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 1952446 : 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 1956458 : if (!RelationUsesLocalBuffers(scan->rs_base.rs_rd) &&
388 1942086 : scan->rs_nblocks > NBuffers / 4)
389 : {
390 27558 : allow_strat = (scan->rs_base.rs_flags & SO_ALLOW_STRAT) != 0;
391 27558 : allow_sync = (scan->rs_base.rs_flags & SO_ALLOW_SYNC) != 0;
392 : }
393 : else
394 1928900 : allow_strat = allow_sync = false;
395 :
396 1956458 : if (allow_strat)
397 : {
398 : /* During a rescan, keep the previous strategy object. */
399 24972 : if (scan->rs_strategy == NULL)
400 24568 : scan->rs_strategy = GetAccessStrategy(BAS_BULKREAD);
401 : }
402 : else
403 : {
404 1931486 : if (scan->rs_strategy != NULL)
405 0 : FreeAccessStrategy(scan->rs_strategy);
406 1931486 : scan->rs_strategy = NULL;
407 : }
408 :
409 1956458 : 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 1952442 : 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 1196266 : if (allow_sync && synchronize_seqscans)
425 100 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
426 : else
427 1196166 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
428 : }
429 756176 : else if (allow_sync && synchronize_seqscans)
430 : {
431 144 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
432 144 : scan->rs_startblock = ss_get_location(scan->rs_base.rs_rd, scan->rs_nblocks);
433 : }
434 : else
435 : {
436 756032 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
437 756032 : scan->rs_startblock = 0;
438 : }
439 :
440 1956458 : scan->rs_numblocks = InvalidBlockNumber;
441 1956458 : scan->rs_inited = false;
442 1956458 : scan->rs_ctup.t_data = NULL;
443 1956458 : ItemPointerSetInvalid(&scan->rs_ctup.t_self);
444 1956458 : scan->rs_cbuf = InvalidBuffer;
445 1956458 : scan->rs_cblock = InvalidBlockNumber;
446 1956458 : scan->rs_ntuples = 0;
447 1956458 : 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 1956458 : scan->rs_dir = ForwardScanDirection;
454 1956458 : 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 1956458 : if (key != NULL && scan->rs_base.rs_nkeys > 0)
462 437950 : 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 1956458 : if (scan->rs_base.rs_flags & SO_TYPE_SEQSCAN)
470 1916234 : pgstat_count_heap_scan(scan->rs_base.rs_rd);
471 1956458 : }
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 5618 : heap_setscanlimits(TableScanDesc sscan, BlockNumber startBlk, BlockNumber numBlks)
481 : {
482 5618 : 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 5618 : scan->rs_startblock = startBlk;
492 5618 : scan->rs_numblocks = numBlks;
493 5618 : }
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 5564534 : 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 5564534 : int ntup = 0;
508 : OffsetNumber lineoff;
509 :
510 280707538 : for (lineoff = FirstOffsetNumber; lineoff <= lines; lineoff++)
511 : {
512 275143020 : ItemId lpp = PageGetItemId(page, lineoff);
513 : HeapTupleData loctup;
514 : bool valid;
515 :
516 275143020 : if (!ItemIdIsNormal(lpp))
517 58339050 : continue;
518 :
519 216803970 : loctup.t_data = (HeapTupleHeader) PageGetItem(page, lpp);
520 216803970 : loctup.t_len = ItemIdGetLength(lpp);
521 216803970 : loctup.t_tableOid = RelationGetRelid(scan->rs_base.rs_rd);
522 216803970 : ItemPointerSet(&(loctup.t_self), block, lineoff);
523 :
524 216803970 : if (all_visible)
525 85727218 : valid = true;
526 : else
527 131076752 : valid = HeapTupleSatisfiesVisibility(&loctup, snapshot, buffer);
528 :
529 216803970 : if (check_serializable)
530 2810 : HeapCheckForSerializableConflictOut(valid, scan->rs_base.rs_rd,
531 : &loctup, buffer, snapshot);
532 :
533 216803954 : if (valid)
534 : {
535 199953100 : scan->rs_vistuples[ntup] = lineoff;
536 199953100 : ntup++;
537 : }
538 : }
539 :
540 : Assert(ntup <= MaxHeapTuplesPerPage);
541 :
542 5564518 : 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 5564534 : heap_prepare_pagescan(TableScanDesc sscan)
553 : {
554 5564534 : HeapScanDesc scan = (HeapScanDesc) sscan;
555 5564534 : Buffer buffer = scan->rs_cbuf;
556 5564534 : 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 5564534 : snapshot = scan->rs_base.rs_snapshot;
568 :
569 : /*
570 : * Prune and repair fragmentation for the whole page, if possible.
571 : */
572 5564534 : 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 5564534 : LockBuffer(buffer, BUFFER_LOCK_SHARE);
580 :
581 5564534 : page = BufferGetPage(buffer);
582 5564534 : 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 5564534 : all_visible = PageIsAllVisible(page) && !snapshot->takenDuringRecovery;
605 : check_serializable =
606 5564534 : 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 5564534 : if (likely(all_visible))
615 : {
616 1990932 : if (likely(!check_serializable))
617 1990932 : 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 3573602 : if (likely(!check_serializable))
626 3572364 : scan->rs_ntuples = page_collect_tuples(scan, snapshot, page, buffer,
627 : block, lines, false, false);
628 : else
629 1238 : scan->rs_ntuples = page_collect_tuples(scan, snapshot, page, buffer,
630 : block, lines, false, true);
631 : }
632 :
633 5564518 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
634 5564518 : }
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 7337236 : heap_fetch_next_buffer(HeapScanDesc scan, ScanDirection dir)
644 : {
645 : Assert(scan->rs_read_stream);
646 :
647 : /* release previous scan buffer, if any */
648 7337236 : if (BufferIsValid(scan->rs_cbuf))
649 : {
650 5420348 : ReleaseBuffer(scan->rs_cbuf);
651 5420348 : 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 7337236 : 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 7337236 : 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 7337236 : scan->rs_dir = dir;
674 :
675 7337236 : scan->rs_cbuf = read_stream_next_buffer(scan->rs_read_stream, NULL);
676 7337186 : if (BufferIsValid(scan->rs_cbuf))
677 5745702 : scan->rs_cblock = BufferGetBlockNumber(scan->rs_cbuf);
678 7337186 : }
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 1913890 : 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 1913890 : if (scan->rs_nblocks == 0 || scan->rs_numblocks == 0)
695 931834 : return InvalidBlockNumber;
696 :
697 982056 : if (ScanDirectionIsForward(dir))
698 : {
699 981992 : 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 189720 : 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 189720 : page = BufferGetPage(scan->rs_cbuf);
745 :
746 189720 : *linesleft = PageGetMaxOffsetNumber(page) - FirstOffsetNumber + 1;
747 :
748 189720 : if (ScanDirectionIsForward(dir))
749 189720 : *lineoff = FirstOffsetNumber;
750 : else
751 0 : *lineoff = (OffsetNumber) (*linesleft);
752 :
753 : /* lineoff now references the physically previous or next tid */
754 189720 : 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 15696020 : 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 15696020 : page = BufferGetPage(scan->rs_cbuf);
776 :
777 15696020 : if (ScanDirectionIsForward(dir))
778 : {
779 15696020 : *lineoff = OffsetNumberNext(scan->rs_coffset);
780 15696020 : *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 15696020 : 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 5588472 : heapgettup_advance_block(HeapScanDesc scan, BlockNumber block, ScanDirection dir)
813 : {
814 : Assert(scan->rs_base.rs_parallel == NULL);
815 :
816 5588472 : if (likely(ScanDirectionIsForward(dir)))
817 : {
818 5588354 : block++;
819 :
820 : /* wrap back to the start of the heap */
821 5588354 : if (block >= scan->rs_nblocks)
822 781456 : 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 5588354 : if (scan->rs_base.rs_flags & SO_ALLOW_SYNC)
837 22530 : ss_report_location(scan->rs_base.rs_rd, block);
838 :
839 : /* we're done if we're back at where we started */
840 5588354 : if (block == scan->rs_startblock)
841 781374 : return InvalidBlockNumber;
842 :
843 : /* check if the limit imposed by heap_setscanlimits() is met */
844 4806980 : if (scan->rs_numblocks != InvalidBlockNumber)
845 : {
846 4932 : if (--scan->rs_numblocks == 0)
847 3056 : return InvalidBlockNumber;
848 : }
849 :
850 4803924 : 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 15736478 : heapgettup(HeapScanDesc scan,
897 : ScanDirection dir,
898 : int nkeys,
899 : ScanKey key)
900 : {
901 15736478 : HeapTuple tuple = &(scan->rs_ctup);
902 : Page page;
903 : OffsetNumber lineoff;
904 : int linesleft;
905 :
906 15736478 : if (likely(scan->rs_inited))
907 : {
908 : /* continue from previously returned page/tuple */
909 15696020 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
910 15696020 : page = heapgettup_continue_page(scan, dir, &linesleft, &lineoff);
911 15696020 : 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 229878 : heap_fetch_next_buffer(scan, dir);
921 :
922 : /* did we run out of blocks to scan? */
923 229878 : if (!BufferIsValid(scan->rs_cbuf))
924 40158 : break;
925 :
926 : Assert(BufferGetBlockNumber(scan->rs_cbuf) == scan->rs_cblock);
927 :
928 189720 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
929 189720 : page = heapgettup_start_page(scan, dir, &linesleft, &lineoff);
930 15885740 : 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 15969660 : for (; linesleft > 0; linesleft--, lineoff += dir)
940 : {
941 : bool visible;
942 15780240 : ItemId lpp = PageGetItemId(page, lineoff);
943 :
944 15780240 : if (!ItemIdIsNormal(lpp))
945 73462 : continue;
946 :
947 15706778 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lpp);
948 15706778 : tuple->t_len = ItemIdGetLength(lpp);
949 15706778 : ItemPointerSet(&(tuple->t_self), scan->rs_cblock, lineoff);
950 :
951 15706778 : visible = HeapTupleSatisfiesVisibility(tuple,
952 : scan->rs_base.rs_snapshot,
953 : scan->rs_cbuf);
954 :
955 15706778 : 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 15706778 : if (!visible)
961 10458 : continue;
962 :
963 : /* skip any tuples that don't match the scan key */
964 15696320 : if (key != NULL &&
965 0 : !HeapKeyTest(tuple, RelationGetDescr(scan->rs_base.rs_rd),
966 : nkeys, key))
967 0 : continue;
968 :
969 15696320 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
970 15696320 : scan->rs_coffset = lineoff;
971 15696320 : 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 189420 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
979 : }
980 :
981 : /* end of scan */
982 40158 : if (BufferIsValid(scan->rs_cbuf))
983 0 : ReleaseBuffer(scan->rs_cbuf);
984 :
985 40158 : scan->rs_cbuf = InvalidBuffer;
986 40158 : scan->rs_cblock = InvalidBlockNumber;
987 40158 : scan->rs_prefetch_block = InvalidBlockNumber;
988 40158 : tuple->t_data = NULL;
989 40158 : 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 99258218 : heapgettup_pagemode(HeapScanDesc scan,
1007 : ScanDirection dir,
1008 : int nkeys,
1009 : ScanKey key)
1010 : {
1011 99258218 : HeapTuple tuple = &(scan->rs_ctup);
1012 : Page page;
1013 : uint32 lineindex;
1014 : uint32 linesleft;
1015 :
1016 99258218 : if (likely(scan->rs_inited))
1017 : {
1018 : /* continue from previously returned page/tuple */
1019 97381788 : page = BufferGetPage(scan->rs_cbuf);
1020 :
1021 97381788 : lineindex = scan->rs_cindex + dir;
1022 97381788 : if (ScanDirectionIsForward(dir))
1023 97381130 : 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 97381788 : 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 7107358 : heap_fetch_next_buffer(scan, dir);
1038 :
1039 : /* did we run out of blocks to scan? */
1040 7107308 : if (!BufferIsValid(scan->rs_cbuf))
1041 1551326 : break;
1042 :
1043 : Assert(BufferGetBlockNumber(scan->rs_cbuf) == scan->rs_cblock);
1044 :
1045 : /* prune the page and determine visible tuple offsets */
1046 5555982 : heap_prepare_pagescan((TableScanDesc) scan);
1047 5555966 : page = BufferGetPage(scan->rs_cbuf);
1048 5555966 : linesleft = scan->rs_ntuples;
1049 5555966 : lineindex = ScanDirectionIsForward(dir) ? 0 : linesleft - 1;
1050 :
1051 : /* block is the same for all tuples, set it once outside the loop */
1052 5555966 : ItemPointerSetBlockNumber(&tuple->t_self, scan->rs_cblock);
1053 :
1054 : /* lineindex now references the next or previous visible tid */
1055 102937754 : continue_page:
1056 :
1057 198764816 : for (; linesleft > 0; linesleft--, lineindex += dir)
1058 : {
1059 : ItemId lpp;
1060 : OffsetNumber lineoff;
1061 :
1062 : Assert(lineindex <= scan->rs_ntuples);
1063 193533888 : lineoff = scan->rs_vistuples[lineindex];
1064 193533888 : lpp = PageGetItemId(page, lineoff);
1065 : Assert(ItemIdIsNormal(lpp));
1066 :
1067 193533888 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lpp);
1068 193533888 : tuple->t_len = ItemIdGetLength(lpp);
1069 193533888 : ItemPointerSetOffsetNumber(&tuple->t_self, lineoff);
1070 :
1071 : /* skip any tuples that don't match the scan key */
1072 193533888 : if (key != NULL &&
1073 96657826 : !HeapKeyTest(tuple, RelationGetDescr(scan->rs_base.rs_rd),
1074 : nkeys, key))
1075 95827062 : continue;
1076 :
1077 97706826 : scan->rs_cindex = lineindex;
1078 97706826 : return;
1079 : }
1080 : }
1081 :
1082 : /* end of scan */
1083 1551326 : if (BufferIsValid(scan->rs_cbuf))
1084 0 : ReleaseBuffer(scan->rs_cbuf);
1085 1551326 : scan->rs_cbuf = InvalidBuffer;
1086 1551326 : scan->rs_cblock = InvalidBlockNumber;
1087 1551326 : scan->rs_prefetch_block = InvalidBlockNumber;
1088 1551326 : tuple->t_data = NULL;
1089 1551326 : scan->rs_inited = false;
1090 : }
1091 :
1092 :
1093 : /* ----------------------------------------------------------------
1094 : * heap access method interface
1095 : * ----------------------------------------------------------------
1096 : */
1097 :
1098 :
1099 : TableScanDesc
1100 760088 : 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 760088 : RelationIncrementReferenceCount(relation);
1115 :
1116 : /*
1117 : * allocate and initialize scan descriptor
1118 : */
1119 760088 : if (flags & SO_TYPE_BITMAPSCAN)
1120 : {
1121 16198 : 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 16198 : scan = (HeapScanDesc) bscan;
1128 : }
1129 : else
1130 743890 : scan = (HeapScanDesc) palloc(sizeof(HeapScanDescData));
1131 :
1132 760088 : scan->rs_base.rs_rd = relation;
1133 760088 : scan->rs_base.rs_snapshot = snapshot;
1134 760088 : scan->rs_base.rs_nkeys = nkeys;
1135 760088 : scan->rs_base.rs_flags = flags;
1136 760088 : scan->rs_base.rs_parallel = parallel_scan;
1137 760088 : scan->rs_strategy = NULL; /* set in initscan */
1138 760088 : scan->rs_cbuf = InvalidBuffer;
1139 :
1140 : /*
1141 : * Disable page-at-a-time mode if it's not a MVCC-safe snapshot.
1142 : */
1143 760088 : if (!(snapshot && IsMVCCSnapshot(snapshot)))
1144 57386 : scan->rs_base.rs_flags &= ~SO_ALLOW_PAGEMODE;
1145 :
1146 : /* Check that a historic snapshot is not used for non-catalog tables */
1147 760088 : if (snapshot &&
1148 743164 : IsHistoricMVCCSnapshot(snapshot) &&
1149 1300 : !RelationIsAccessibleInLogicalDecoding(relation))
1150 : {
1151 0 : ereport(ERROR,
1152 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
1153 : errmsg("cannot query non-catalog table \"%s\" during logical decoding",
1154 : RelationGetRelationName(relation))));
1155 : }
1156 :
1157 : /*
1158 : * For seqscan and sample scans in a serializable transaction, acquire a
1159 : * predicate lock on the entire relation. This is required not only to
1160 : * lock all the matching tuples, but also to conflict with new insertions
1161 : * into the table. In an indexscan, we take page locks on the index pages
1162 : * covering the range specified in the scan qual, but in a heap scan there
1163 : * is nothing more fine-grained to lock. A bitmap scan is a different
1164 : * story, there we have already scanned the index and locked the index
1165 : * pages covering the predicate. But in that case we still have to lock
1166 : * any matching heap tuples. For sample scan we could optimize the locking
1167 : * to be at least page-level granularity, but we'd need to add per-tuple
1168 : * locking for that.
1169 : */
1170 760088 : if (scan->rs_base.rs_flags & (SO_TYPE_SEQSCAN | SO_TYPE_SAMPLESCAN))
1171 : {
1172 : /*
1173 : * Ensure a missing snapshot is noticed reliably, even if the
1174 : * isolation mode means predicate locking isn't performed (and
1175 : * therefore the snapshot isn't used here).
1176 : */
1177 : Assert(snapshot);
1178 724346 : PredicateLockRelation(relation, snapshot);
1179 : }
1180 :
1181 : /* we only need to set this up once */
1182 760088 : scan->rs_ctup.t_tableOid = RelationGetRelid(relation);
1183 :
1184 : /*
1185 : * Allocate memory to keep track of page allocation for parallel workers
1186 : * when doing a parallel scan.
1187 : */
1188 760088 : if (parallel_scan != NULL)
1189 3908 : scan->rs_parallelworkerdata = palloc(sizeof(ParallelBlockTableScanWorkerData));
1190 : else
1191 756180 : scan->rs_parallelworkerdata = NULL;
1192 :
1193 : /*
1194 : * we do this here instead of in initscan() because heap_rescan also calls
1195 : * initscan() and we don't want to allocate memory again
1196 : */
1197 760088 : if (nkeys > 0)
1198 437950 : scan->rs_base.rs_key = (ScanKey) palloc(sizeof(ScanKeyData) * nkeys);
1199 : else
1200 322138 : scan->rs_base.rs_key = NULL;
1201 :
1202 760088 : initscan(scan, key, false);
1203 :
1204 760084 : scan->rs_read_stream = NULL;
1205 :
1206 : /*
1207 : * Set up a read stream for sequential scans and TID range scans. This
1208 : * should be done after initscan() because initscan() allocates the
1209 : * BufferAccessStrategy object passed to the read stream API.
1210 : */
1211 760084 : if (scan->rs_base.rs_flags & SO_TYPE_SEQSCAN ||
1212 35888 : scan->rs_base.rs_flags & SO_TYPE_TIDRANGESCAN)
1213 726046 : {
1214 : ReadStreamBlockNumberCB cb;
1215 :
1216 726046 : if (scan->rs_base.rs_parallel)
1217 3908 : cb = heap_scan_stream_read_next_parallel;
1218 : else
1219 722138 : cb = heap_scan_stream_read_next_serial;
1220 :
1221 : /* ---
1222 : * It is safe to use batchmode as the only locks taken by `cb`
1223 : * are never taken while waiting for IO:
1224 : * - SyncScanLock is used in the non-parallel case
1225 : * - in the parallel case, only spinlocks and atomics are used
1226 : * ---
1227 : */
1228 726046 : scan->rs_read_stream = read_stream_begin_relation(READ_STREAM_SEQUENTIAL |
1229 : READ_STREAM_USE_BATCHING,
1230 : scan->rs_strategy,
1231 : scan->rs_base.rs_rd,
1232 : MAIN_FORKNUM,
1233 : cb,
1234 : scan,
1235 : 0);
1236 : }
1237 34038 : else if (scan->rs_base.rs_flags & SO_TYPE_BITMAPSCAN)
1238 : {
1239 16198 : scan->rs_read_stream = read_stream_begin_relation(READ_STREAM_DEFAULT |
1240 : READ_STREAM_USE_BATCHING,
1241 : scan->rs_strategy,
1242 : scan->rs_base.rs_rd,
1243 : MAIN_FORKNUM,
1244 : bitmapheap_stream_read_next,
1245 : scan,
1246 : sizeof(TBMIterateResult));
1247 : }
1248 :
1249 :
1250 760084 : return (TableScanDesc) scan;
1251 : }
1252 :
1253 : void
1254 1196374 : heap_rescan(TableScanDesc sscan, ScanKey key, bool set_params,
1255 : bool allow_strat, bool allow_sync, bool allow_pagemode)
1256 : {
1257 1196374 : HeapScanDesc scan = (HeapScanDesc) sscan;
1258 :
1259 1196374 : if (set_params)
1260 : {
1261 30 : if (allow_strat)
1262 30 : scan->rs_base.rs_flags |= SO_ALLOW_STRAT;
1263 : else
1264 0 : scan->rs_base.rs_flags &= ~SO_ALLOW_STRAT;
1265 :
1266 30 : if (allow_sync)
1267 12 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
1268 : else
1269 18 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
1270 :
1271 30 : if (allow_pagemode && scan->rs_base.rs_snapshot &&
1272 30 : IsMVCCSnapshot(scan->rs_base.rs_snapshot))
1273 30 : scan->rs_base.rs_flags |= SO_ALLOW_PAGEMODE;
1274 : else
1275 0 : scan->rs_base.rs_flags &= ~SO_ALLOW_PAGEMODE;
1276 : }
1277 :
1278 : /*
1279 : * unpin scan buffers
1280 : */
1281 1196374 : if (BufferIsValid(scan->rs_cbuf))
1282 : {
1283 3038 : ReleaseBuffer(scan->rs_cbuf);
1284 3038 : scan->rs_cbuf = InvalidBuffer;
1285 : }
1286 :
1287 : /*
1288 : * SO_TYPE_BITMAPSCAN would be cleaned up here, but it does not hold any
1289 : * additional data vs a normal HeapScan
1290 : */
1291 :
1292 : /*
1293 : * The read stream is reset on rescan. This must be done before
1294 : * initscan(), as some state referred to by read_stream_reset() is reset
1295 : * in initscan().
1296 : */
1297 1196374 : if (scan->rs_read_stream)
1298 1196338 : read_stream_reset(scan->rs_read_stream);
1299 :
1300 : /*
1301 : * reinitialize scan descriptor
1302 : */
1303 1196374 : initscan(scan, key, true);
1304 1196374 : }
1305 :
1306 : void
1307 755440 : heap_endscan(TableScanDesc sscan)
1308 : {
1309 755440 : HeapScanDesc scan = (HeapScanDesc) sscan;
1310 :
1311 : /* Note: no locking manipulations needed */
1312 :
1313 : /*
1314 : * unpin scan buffers
1315 : */
1316 755440 : if (BufferIsValid(scan->rs_cbuf))
1317 319106 : ReleaseBuffer(scan->rs_cbuf);
1318 :
1319 : /*
1320 : * Must free the read stream before freeing the BufferAccessStrategy.
1321 : */
1322 755440 : if (scan->rs_read_stream)
1323 737706 : read_stream_end(scan->rs_read_stream);
1324 :
1325 : /*
1326 : * decrement relation reference count and free scan descriptor storage
1327 : */
1328 755440 : RelationDecrementReferenceCount(scan->rs_base.rs_rd);
1329 :
1330 755440 : if (scan->rs_base.rs_key)
1331 437896 : pfree(scan->rs_base.rs_key);
1332 :
1333 755440 : if (scan->rs_strategy != NULL)
1334 24550 : FreeAccessStrategy(scan->rs_strategy);
1335 :
1336 755440 : if (scan->rs_parallelworkerdata != NULL)
1337 3908 : pfree(scan->rs_parallelworkerdata);
1338 :
1339 755440 : if (scan->rs_base.rs_flags & SO_TEMP_SNAPSHOT)
1340 80114 : UnregisterSnapshot(scan->rs_base.rs_snapshot);
1341 :
1342 755440 : pfree(scan);
1343 755440 : }
1344 :
1345 : HeapTuple
1346 20182878 : heap_getnext(TableScanDesc sscan, ScanDirection direction)
1347 : {
1348 20182878 : HeapScanDesc scan = (HeapScanDesc) sscan;
1349 :
1350 : /*
1351 : * This is still widely used directly, without going through table AM, so
1352 : * add a safety check. It's possible we should, at a later point,
1353 : * downgrade this to an assert. The reason for checking the AM routine,
1354 : * rather than the AM oid, is that this allows to write regression tests
1355 : * that create another AM reusing the heap handler.
1356 : */
1357 20182878 : if (unlikely(sscan->rs_rd->rd_tableam != GetHeapamTableAmRoutine()))
1358 0 : ereport(ERROR,
1359 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1360 : errmsg_internal("only heap AM is supported")));
1361 :
1362 : /*
1363 : * We don't expect direct calls to heap_getnext with valid CheckXidAlive
1364 : * for catalog or regular tables. See detailed comments in xact.c where
1365 : * these variables are declared. Normally we have such a check at tableam
1366 : * level API but this is called from many places so we need to ensure it
1367 : * here.
1368 : */
1369 20182878 : if (unlikely(TransactionIdIsValid(CheckXidAlive) && !bsysscan))
1370 0 : elog(ERROR, "unexpected heap_getnext call during logical decoding");
1371 :
1372 : /* Note: no locking manipulations needed */
1373 :
1374 20182878 : if (scan->rs_base.rs_flags & SO_ALLOW_PAGEMODE)
1375 5421980 : heapgettup_pagemode(scan, direction,
1376 5421980 : scan->rs_base.rs_nkeys, scan->rs_base.rs_key);
1377 : else
1378 14760898 : heapgettup(scan, direction,
1379 14760898 : scan->rs_base.rs_nkeys, scan->rs_base.rs_key);
1380 :
1381 20182878 : if (scan->rs_ctup.t_data == NULL)
1382 132282 : return NULL;
1383 :
1384 : /*
1385 : * if we get here it means we have a new current scan tuple, so point to
1386 : * the proper return buffer and return the tuple.
1387 : */
1388 :
1389 20050596 : pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1390 :
1391 20050596 : return &scan->rs_ctup;
1392 : }
1393 :
1394 : bool
1395 94803038 : heap_getnextslot(TableScanDesc sscan, ScanDirection direction, TupleTableSlot *slot)
1396 : {
1397 94803038 : HeapScanDesc scan = (HeapScanDesc) sscan;
1398 :
1399 : /* Note: no locking manipulations needed */
1400 :
1401 94803038 : if (sscan->rs_flags & SO_ALLOW_PAGEMODE)
1402 93827458 : heapgettup_pagemode(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1403 : else
1404 975580 : heapgettup(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1405 :
1406 94802988 : if (scan->rs_ctup.t_data == NULL)
1407 : {
1408 1459108 : ExecClearTuple(slot);
1409 1459108 : return false;
1410 : }
1411 :
1412 : /*
1413 : * if we get here it means we have a new current scan tuple, so point to
1414 : * the proper return buffer and return the tuple.
1415 : */
1416 :
1417 93343880 : pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1418 :
1419 93343880 : ExecStoreBufferHeapTuple(&scan->rs_ctup, slot,
1420 : scan->rs_cbuf);
1421 93343880 : return true;
1422 : }
1423 :
1424 : void
1425 1916 : heap_set_tidrange(TableScanDesc sscan, ItemPointer mintid,
1426 : ItemPointer maxtid)
1427 : {
1428 1916 : HeapScanDesc scan = (HeapScanDesc) sscan;
1429 : BlockNumber startBlk;
1430 : BlockNumber numBlks;
1431 : ItemPointerData highestItem;
1432 : ItemPointerData lowestItem;
1433 :
1434 : /*
1435 : * For relations without any pages, we can simply leave the TID range
1436 : * unset. There will be no tuples to scan, therefore no tuples outside
1437 : * the given TID range.
1438 : */
1439 1916 : if (scan->rs_nblocks == 0)
1440 48 : return;
1441 :
1442 : /*
1443 : * Set up some ItemPointers which point to the first and last possible
1444 : * tuples in the heap.
1445 : */
1446 1904 : ItemPointerSet(&highestItem, scan->rs_nblocks - 1, MaxOffsetNumber);
1447 1904 : ItemPointerSet(&lowestItem, 0, FirstOffsetNumber);
1448 :
1449 : /*
1450 : * If the given maximum TID is below the highest possible TID in the
1451 : * relation, then restrict the range to that, otherwise we scan to the end
1452 : * of the relation.
1453 : */
1454 1904 : if (ItemPointerCompare(maxtid, &highestItem) < 0)
1455 136 : ItemPointerCopy(maxtid, &highestItem);
1456 :
1457 : /*
1458 : * If the given minimum TID is above the lowest possible TID in the
1459 : * relation, then restrict the range to only scan for TIDs above that.
1460 : */
1461 1904 : if (ItemPointerCompare(mintid, &lowestItem) > 0)
1462 1762 : ItemPointerCopy(mintid, &lowestItem);
1463 :
1464 : /*
1465 : * Check for an empty range and protect from would be negative results
1466 : * from the numBlks calculation below.
1467 : */
1468 1904 : if (ItemPointerCompare(&highestItem, &lowestItem) < 0)
1469 : {
1470 : /* Set an empty range of blocks to scan */
1471 36 : heap_setscanlimits(sscan, 0, 0);
1472 36 : return;
1473 : }
1474 :
1475 : /*
1476 : * Calculate the first block and the number of blocks we must scan. We
1477 : * could be more aggressive here and perform some more validation to try
1478 : * and further narrow the scope of blocks to scan by checking if the
1479 : * lowestItem has an offset above MaxOffsetNumber. In this case, we could
1480 : * advance startBlk by one. Likewise, if highestItem has an offset of 0
1481 : * we could scan one fewer blocks. However, such an optimization does not
1482 : * seem worth troubling over, currently.
1483 : */
1484 1868 : startBlk = ItemPointerGetBlockNumberNoCheck(&lowestItem);
1485 :
1486 1868 : numBlks = ItemPointerGetBlockNumberNoCheck(&highestItem) -
1487 1868 : ItemPointerGetBlockNumberNoCheck(&lowestItem) + 1;
1488 :
1489 : /* Set the start block and number of blocks to scan */
1490 1868 : heap_setscanlimits(sscan, startBlk, numBlks);
1491 :
1492 : /* Finally, set the TID range in sscan */
1493 1868 : ItemPointerCopy(&lowestItem, &sscan->st.tidrange.rs_mintid);
1494 1868 : ItemPointerCopy(&highestItem, &sscan->st.tidrange.rs_maxtid);
1495 : }
1496 :
1497 : bool
1498 8594 : heap_getnextslot_tidrange(TableScanDesc sscan, ScanDirection direction,
1499 : TupleTableSlot *slot)
1500 : {
1501 8594 : HeapScanDesc scan = (HeapScanDesc) sscan;
1502 8594 : ItemPointer mintid = &sscan->st.tidrange.rs_mintid;
1503 8594 : ItemPointer maxtid = &sscan->st.tidrange.rs_maxtid;
1504 :
1505 : /* Note: no locking manipulations needed */
1506 : for (;;)
1507 : {
1508 8780 : if (sscan->rs_flags & SO_ALLOW_PAGEMODE)
1509 8780 : heapgettup_pagemode(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1510 : else
1511 0 : heapgettup(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1512 :
1513 8764 : if (scan->rs_ctup.t_data == NULL)
1514 : {
1515 94 : ExecClearTuple(slot);
1516 94 : return false;
1517 : }
1518 :
1519 : /*
1520 : * heap_set_tidrange will have used heap_setscanlimits to limit the
1521 : * range of pages we scan to only ones that can contain the TID range
1522 : * we're scanning for. Here we must filter out any tuples from these
1523 : * pages that are outside of that range.
1524 : */
1525 8670 : if (ItemPointerCompare(&scan->rs_ctup.t_self, mintid) < 0)
1526 : {
1527 186 : ExecClearTuple(slot);
1528 :
1529 : /*
1530 : * When scanning backwards, the TIDs will be in descending order.
1531 : * Future tuples in this direction will be lower still, so we can
1532 : * just return false to indicate there will be no more tuples.
1533 : */
1534 186 : if (ScanDirectionIsBackward(direction))
1535 0 : return false;
1536 :
1537 186 : continue;
1538 : }
1539 :
1540 : /*
1541 : * Likewise for the final page, we must filter out TIDs greater than
1542 : * maxtid.
1543 : */
1544 8484 : if (ItemPointerCompare(&scan->rs_ctup.t_self, maxtid) > 0)
1545 : {
1546 72 : ExecClearTuple(slot);
1547 :
1548 : /*
1549 : * When scanning forward, the TIDs will be in ascending order.
1550 : * Future tuples in this direction will be higher still, so we can
1551 : * just return false to indicate there will be no more tuples.
1552 : */
1553 72 : if (ScanDirectionIsForward(direction))
1554 72 : return false;
1555 0 : continue;
1556 : }
1557 :
1558 8412 : break;
1559 : }
1560 :
1561 : /*
1562 : * if we get here it means we have a new current scan tuple, so point to
1563 : * the proper return buffer and return the tuple.
1564 : */
1565 8412 : pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1566 :
1567 8412 : ExecStoreBufferHeapTuple(&scan->rs_ctup, slot, scan->rs_cbuf);
1568 8412 : return true;
1569 : }
1570 :
1571 : /*
1572 : * heap_fetch - retrieve tuple with given tid
1573 : *
1574 : * On entry, tuple->t_self is the TID to fetch. We pin the buffer holding
1575 : * the tuple, fill in the remaining fields of *tuple, and check the tuple
1576 : * against the specified snapshot.
1577 : *
1578 : * If successful (tuple found and passes snapshot time qual), then *userbuf
1579 : * is set to the buffer holding the tuple and true is returned. The caller
1580 : * must unpin the buffer when done with the tuple.
1581 : *
1582 : * If the tuple is not found (ie, item number references a deleted slot),
1583 : * then tuple->t_data is set to NULL, *userbuf is set to InvalidBuffer,
1584 : * and false is returned.
1585 : *
1586 : * If the tuple is found but fails the time qual check, then the behavior
1587 : * depends on the keep_buf parameter. If keep_buf is false, the results
1588 : * are the same as for the tuple-not-found case. If keep_buf is true,
1589 : * then tuple->t_data and *userbuf are returned as for the success case,
1590 : * and again the caller must unpin the buffer; but false is returned.
1591 : *
1592 : * heap_fetch does not follow HOT chains: only the exact TID requested will
1593 : * be fetched.
1594 : *
1595 : * It is somewhat inconsistent that we ereport() on invalid block number but
1596 : * return false on invalid item number. There are a couple of reasons though.
1597 : * One is that the caller can relatively easily check the block number for
1598 : * validity, but cannot check the item number without reading the page
1599 : * himself. Another is that when we are following a t_ctid link, we can be
1600 : * reasonably confident that the page number is valid (since VACUUM shouldn't
1601 : * truncate off the destination page without having killed the referencing
1602 : * tuple first), but the item number might well not be good.
1603 : */
1604 : bool
1605 355320 : heap_fetch(Relation relation,
1606 : Snapshot snapshot,
1607 : HeapTuple tuple,
1608 : Buffer *userbuf,
1609 : bool keep_buf)
1610 : {
1611 355320 : ItemPointer tid = &(tuple->t_self);
1612 : ItemId lp;
1613 : Buffer buffer;
1614 : Page page;
1615 : OffsetNumber offnum;
1616 : bool valid;
1617 :
1618 : /*
1619 : * Fetch and pin the appropriate page of the relation.
1620 : */
1621 355320 : buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
1622 :
1623 : /*
1624 : * Need share lock on buffer to examine tuple commit status.
1625 : */
1626 355304 : LockBuffer(buffer, BUFFER_LOCK_SHARE);
1627 355304 : page = BufferGetPage(buffer);
1628 :
1629 : /*
1630 : * We'd better check for out-of-range offnum in case of VACUUM since the
1631 : * TID was obtained.
1632 : */
1633 355304 : offnum = ItemPointerGetOffsetNumber(tid);
1634 355304 : if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1635 : {
1636 6 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1637 6 : ReleaseBuffer(buffer);
1638 6 : *userbuf = InvalidBuffer;
1639 6 : tuple->t_data = NULL;
1640 6 : return false;
1641 : }
1642 :
1643 : /*
1644 : * get the item line pointer corresponding to the requested tid
1645 : */
1646 355298 : lp = PageGetItemId(page, offnum);
1647 :
1648 : /*
1649 : * Must check for deleted tuple.
1650 : */
1651 355298 : if (!ItemIdIsNormal(lp))
1652 : {
1653 690 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1654 690 : ReleaseBuffer(buffer);
1655 690 : *userbuf = InvalidBuffer;
1656 690 : tuple->t_data = NULL;
1657 690 : return false;
1658 : }
1659 :
1660 : /*
1661 : * fill in *tuple fields
1662 : */
1663 354608 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
1664 354608 : tuple->t_len = ItemIdGetLength(lp);
1665 354608 : tuple->t_tableOid = RelationGetRelid(relation);
1666 :
1667 : /*
1668 : * check tuple visibility, then release lock
1669 : */
1670 354608 : valid = HeapTupleSatisfiesVisibility(tuple, snapshot, buffer);
1671 :
1672 354608 : if (valid)
1673 354514 : PredicateLockTID(relation, &(tuple->t_self), snapshot,
1674 354514 : HeapTupleHeaderGetXmin(tuple->t_data));
1675 :
1676 354608 : HeapCheckForSerializableConflictOut(valid, relation, tuple, buffer, snapshot);
1677 :
1678 354608 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1679 :
1680 354608 : if (valid)
1681 : {
1682 : /*
1683 : * All checks passed, so return the tuple as valid. Caller is now
1684 : * responsible for releasing the buffer.
1685 : */
1686 354514 : *userbuf = buffer;
1687 :
1688 354514 : return true;
1689 : }
1690 :
1691 : /* Tuple failed time qual, but maybe caller wants to see it anyway. */
1692 94 : if (keep_buf)
1693 58 : *userbuf = buffer;
1694 : else
1695 : {
1696 36 : ReleaseBuffer(buffer);
1697 36 : *userbuf = InvalidBuffer;
1698 36 : tuple->t_data = NULL;
1699 : }
1700 :
1701 94 : return false;
1702 : }
1703 :
1704 : /*
1705 : * heap_hot_search_buffer - search HOT chain for tuple satisfying snapshot
1706 : *
1707 : * On entry, *tid is the TID of a tuple (either a simple tuple, or the root
1708 : * of a HOT chain), and buffer is the buffer holding this tuple. We search
1709 : * for the first chain member satisfying the given snapshot. If one is
1710 : * found, we update *tid to reference that tuple's offset number, and
1711 : * return true. If no match, return false without modifying *tid.
1712 : *
1713 : * heapTuple is a caller-supplied buffer. When a match is found, we return
1714 : * the tuple here, in addition to updating *tid. If no match is found, the
1715 : * contents of this buffer on return are undefined.
1716 : *
1717 : * If all_dead is not NULL, we check non-visible tuples to see if they are
1718 : * globally dead; *all_dead is set true if all members of the HOT chain
1719 : * are vacuumable, false if not.
1720 : *
1721 : * Unlike heap_fetch, the caller must already have pin and (at least) share
1722 : * lock on the buffer; it is still pinned/locked at exit.
1723 : */
1724 : bool
1725 44873364 : heap_hot_search_buffer(ItemPointer tid, Relation relation, Buffer buffer,
1726 : Snapshot snapshot, HeapTuple heapTuple,
1727 : bool *all_dead, bool first_call)
1728 : {
1729 44873364 : Page page = BufferGetPage(buffer);
1730 44873364 : TransactionId prev_xmax = InvalidTransactionId;
1731 : BlockNumber blkno;
1732 : OffsetNumber offnum;
1733 : bool at_chain_start;
1734 : bool valid;
1735 : bool skip;
1736 44873364 : GlobalVisState *vistest = NULL;
1737 :
1738 : /* If this is not the first call, previous call returned a (live!) tuple */
1739 44873364 : if (all_dead)
1740 38386900 : *all_dead = first_call;
1741 :
1742 44873364 : blkno = ItemPointerGetBlockNumber(tid);
1743 44873364 : offnum = ItemPointerGetOffsetNumber(tid);
1744 44873364 : at_chain_start = first_call;
1745 44873364 : skip = !first_call;
1746 :
1747 : /* XXX: we should assert that a snapshot is pushed or registered */
1748 : Assert(TransactionIdIsValid(RecentXmin));
1749 : Assert(BufferGetBlockNumber(buffer) == blkno);
1750 :
1751 : /* Scan through possible multiple members of HOT-chain */
1752 : for (;;)
1753 2910478 : {
1754 : ItemId lp;
1755 :
1756 : /* check for bogus TID */
1757 47783842 : if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1758 : break;
1759 :
1760 47783842 : lp = PageGetItemId(page, offnum);
1761 :
1762 : /* check for unused, dead, or redirected items */
1763 47783842 : if (!ItemIdIsNormal(lp))
1764 : {
1765 : /* We should only see a redirect at start of chain */
1766 1773530 : if (ItemIdIsRedirected(lp) && at_chain_start)
1767 : {
1768 : /* Follow the redirect */
1769 1012250 : offnum = ItemIdGetRedirect(lp);
1770 1012250 : at_chain_start = false;
1771 1012250 : continue;
1772 : }
1773 : /* else must be end of chain */
1774 761280 : break;
1775 : }
1776 :
1777 : /*
1778 : * Update heapTuple to point to the element of the HOT chain we're
1779 : * currently investigating. Having t_self set correctly is important
1780 : * because the SSI checks and the *Satisfies routine for historical
1781 : * MVCC snapshots need the correct tid to decide about the visibility.
1782 : */
1783 46010312 : heapTuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
1784 46010312 : heapTuple->t_len = ItemIdGetLength(lp);
1785 46010312 : heapTuple->t_tableOid = RelationGetRelid(relation);
1786 46010312 : ItemPointerSet(&heapTuple->t_self, blkno, offnum);
1787 :
1788 : /*
1789 : * Shouldn't see a HEAP_ONLY tuple at chain start.
1790 : */
1791 46010312 : if (at_chain_start && HeapTupleIsHeapOnly(heapTuple))
1792 0 : break;
1793 :
1794 : /*
1795 : * The xmin should match the previous xmax value, else chain is
1796 : * broken.
1797 : */
1798 47908540 : if (TransactionIdIsValid(prev_xmax) &&
1799 1898228 : !TransactionIdEquals(prev_xmax,
1800 : HeapTupleHeaderGetXmin(heapTuple->t_data)))
1801 0 : break;
1802 :
1803 : /*
1804 : * When first_call is true (and thus, skip is initially false) we'll
1805 : * return the first tuple we find. But on later passes, heapTuple
1806 : * will initially be pointing to the tuple we returned last time.
1807 : * Returning it again would be incorrect (and would loop forever), so
1808 : * we skip it and return the next match we find.
1809 : */
1810 46010312 : if (!skip)
1811 : {
1812 : /* If it's visible per the snapshot, we must return it */
1813 45847498 : valid = HeapTupleSatisfiesVisibility(heapTuple, snapshot, buffer);
1814 45847498 : HeapCheckForSerializableConflictOut(valid, relation, heapTuple,
1815 : buffer, snapshot);
1816 :
1817 45847488 : if (valid)
1818 : {
1819 31677526 : ItemPointerSetOffsetNumber(tid, offnum);
1820 31677526 : PredicateLockTID(relation, &heapTuple->t_self, snapshot,
1821 31677526 : HeapTupleHeaderGetXmin(heapTuple->t_data));
1822 31677526 : if (all_dead)
1823 25764152 : *all_dead = false;
1824 31677526 : return true;
1825 : }
1826 : }
1827 14332776 : skip = false;
1828 :
1829 : /*
1830 : * If we can't see it, maybe no one else can either. At caller
1831 : * request, check whether all chain members are dead to all
1832 : * transactions.
1833 : *
1834 : * Note: if you change the criterion here for what is "dead", fix the
1835 : * planner's get_actual_variable_range() function to match.
1836 : */
1837 14332776 : if (all_dead && *all_dead)
1838 : {
1839 12898178 : if (!vistest)
1840 12643720 : vistest = GlobalVisTestFor(relation);
1841 :
1842 12898178 : if (!HeapTupleIsSurelyDead(heapTuple, vistest))
1843 12182534 : *all_dead = false;
1844 : }
1845 :
1846 : /*
1847 : * Check to see if HOT chain continues past this tuple; if so fetch
1848 : * the next offnum and loop around.
1849 : */
1850 14332776 : if (HeapTupleIsHotUpdated(heapTuple))
1851 : {
1852 : Assert(ItemPointerGetBlockNumber(&heapTuple->t_data->t_ctid) ==
1853 : blkno);
1854 1898228 : offnum = ItemPointerGetOffsetNumber(&heapTuple->t_data->t_ctid);
1855 1898228 : at_chain_start = false;
1856 1898228 : prev_xmax = HeapTupleHeaderGetUpdateXid(heapTuple->t_data);
1857 : }
1858 : else
1859 12434548 : break; /* end of chain */
1860 : }
1861 :
1862 13195828 : return false;
1863 : }
1864 :
1865 : /*
1866 : * heap_get_latest_tid - get the latest tid of a specified tuple
1867 : *
1868 : * Actually, this gets the latest version that is visible according to the
1869 : * scan's snapshot. Create a scan using SnapshotDirty to get the very latest,
1870 : * possibly uncommitted version.
1871 : *
1872 : * *tid is both an input and an output parameter: it is updated to
1873 : * show the latest version of the row. Note that it will not be changed
1874 : * if no version of the row passes the snapshot test.
1875 : */
1876 : void
1877 300 : heap_get_latest_tid(TableScanDesc sscan,
1878 : ItemPointer tid)
1879 : {
1880 300 : Relation relation = sscan->rs_rd;
1881 300 : Snapshot snapshot = sscan->rs_snapshot;
1882 : ItemPointerData ctid;
1883 : TransactionId priorXmax;
1884 :
1885 : /*
1886 : * table_tuple_get_latest_tid() verified that the passed in tid is valid.
1887 : * Assume that t_ctid links are valid however - there shouldn't be invalid
1888 : * ones in the table.
1889 : */
1890 : Assert(ItemPointerIsValid(tid));
1891 :
1892 : /*
1893 : * Loop to chase down t_ctid links. At top of loop, ctid is the tuple we
1894 : * need to examine, and *tid is the TID we will return if ctid turns out
1895 : * to be bogus.
1896 : *
1897 : * Note that we will loop until we reach the end of the t_ctid chain.
1898 : * Depending on the snapshot passed, there might be at most one visible
1899 : * version of the row, but we don't try to optimize for that.
1900 : */
1901 300 : ctid = *tid;
1902 300 : priorXmax = InvalidTransactionId; /* cannot check first XMIN */
1903 : for (;;)
1904 90 : {
1905 : Buffer buffer;
1906 : Page page;
1907 : OffsetNumber offnum;
1908 : ItemId lp;
1909 : HeapTupleData tp;
1910 : bool valid;
1911 :
1912 : /*
1913 : * Read, pin, and lock the page.
1914 : */
1915 390 : buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&ctid));
1916 390 : LockBuffer(buffer, BUFFER_LOCK_SHARE);
1917 390 : page = BufferGetPage(buffer);
1918 :
1919 : /*
1920 : * Check for bogus item number. This is not treated as an error
1921 : * condition because it can happen while following a t_ctid link. We
1922 : * just assume that the prior tid is OK and return it unchanged.
1923 : */
1924 390 : offnum = ItemPointerGetOffsetNumber(&ctid);
1925 390 : if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1926 : {
1927 0 : UnlockReleaseBuffer(buffer);
1928 0 : break;
1929 : }
1930 390 : lp = PageGetItemId(page, offnum);
1931 390 : if (!ItemIdIsNormal(lp))
1932 : {
1933 0 : UnlockReleaseBuffer(buffer);
1934 0 : break;
1935 : }
1936 :
1937 : /* OK to access the tuple */
1938 390 : tp.t_self = ctid;
1939 390 : tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
1940 390 : tp.t_len = ItemIdGetLength(lp);
1941 390 : tp.t_tableOid = RelationGetRelid(relation);
1942 :
1943 : /*
1944 : * After following a t_ctid link, we might arrive at an unrelated
1945 : * tuple. Check for XMIN match.
1946 : */
1947 480 : if (TransactionIdIsValid(priorXmax) &&
1948 90 : !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(tp.t_data)))
1949 : {
1950 0 : UnlockReleaseBuffer(buffer);
1951 0 : break;
1952 : }
1953 :
1954 : /*
1955 : * Check tuple visibility; if visible, set it as the new result
1956 : * candidate.
1957 : */
1958 390 : valid = HeapTupleSatisfiesVisibility(&tp, snapshot, buffer);
1959 390 : HeapCheckForSerializableConflictOut(valid, relation, &tp, buffer, snapshot);
1960 390 : if (valid)
1961 276 : *tid = ctid;
1962 :
1963 : /*
1964 : * If there's a valid t_ctid link, follow it, else we're done.
1965 : */
1966 552 : if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
1967 276 : HeapTupleHeaderIsOnlyLocked(tp.t_data) ||
1968 228 : HeapTupleHeaderIndicatesMovedPartitions(tp.t_data) ||
1969 114 : ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid))
1970 : {
1971 300 : UnlockReleaseBuffer(buffer);
1972 300 : break;
1973 : }
1974 :
1975 90 : ctid = tp.t_data->t_ctid;
1976 90 : priorXmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
1977 90 : UnlockReleaseBuffer(buffer);
1978 : } /* end of loop */
1979 300 : }
1980 :
1981 :
1982 : /*
1983 : * UpdateXmaxHintBits - update tuple hint bits after xmax transaction ends
1984 : *
1985 : * This is called after we have waited for the XMAX transaction to terminate.
1986 : * If the transaction aborted, we guarantee the XMAX_INVALID hint bit will
1987 : * be set on exit. If the transaction committed, we set the XMAX_COMMITTED
1988 : * hint bit if possible --- but beware that that may not yet be possible,
1989 : * if the transaction committed asynchronously.
1990 : *
1991 : * Note that if the transaction was a locker only, we set HEAP_XMAX_INVALID
1992 : * even if it commits.
1993 : *
1994 : * Hence callers should look only at XMAX_INVALID.
1995 : *
1996 : * Note this is not allowed for tuples whose xmax is a multixact.
1997 : */
1998 : static void
1999 350 : UpdateXmaxHintBits(HeapTupleHeader tuple, Buffer buffer, TransactionId xid)
2000 : {
2001 : Assert(TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple), xid));
2002 : Assert(!(tuple->t_infomask & HEAP_XMAX_IS_MULTI));
2003 :
2004 350 : if (!(tuple->t_infomask & (HEAP_XMAX_COMMITTED | HEAP_XMAX_INVALID)))
2005 : {
2006 634 : if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask) &&
2007 284 : TransactionIdDidCommit(xid))
2008 232 : HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_COMMITTED,
2009 : xid);
2010 : else
2011 118 : HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_INVALID,
2012 : InvalidTransactionId);
2013 : }
2014 350 : }
2015 :
2016 :
2017 : /*
2018 : * GetBulkInsertState - prepare status object for a bulk insert
2019 : */
2020 : BulkInsertState
2021 4650 : GetBulkInsertState(void)
2022 : {
2023 : BulkInsertState bistate;
2024 :
2025 4650 : bistate = (BulkInsertState) palloc(sizeof(BulkInsertStateData));
2026 4650 : bistate->strategy = GetAccessStrategy(BAS_BULKWRITE);
2027 4650 : bistate->current_buf = InvalidBuffer;
2028 4650 : bistate->next_free = InvalidBlockNumber;
2029 4650 : bistate->last_free = InvalidBlockNumber;
2030 4650 : bistate->already_extended_by = 0;
2031 4650 : return bistate;
2032 : }
2033 :
2034 : /*
2035 : * FreeBulkInsertState - clean up after finishing a bulk insert
2036 : */
2037 : void
2038 4364 : FreeBulkInsertState(BulkInsertState bistate)
2039 : {
2040 4364 : if (bistate->current_buf != InvalidBuffer)
2041 3482 : ReleaseBuffer(bistate->current_buf);
2042 4364 : FreeAccessStrategy(bistate->strategy);
2043 4364 : pfree(bistate);
2044 4364 : }
2045 :
2046 : /*
2047 : * ReleaseBulkInsertStatePin - release a buffer currently held in bistate
2048 : */
2049 : void
2050 161516 : ReleaseBulkInsertStatePin(BulkInsertState bistate)
2051 : {
2052 161516 : if (bistate->current_buf != InvalidBuffer)
2053 60042 : ReleaseBuffer(bistate->current_buf);
2054 161516 : bistate->current_buf = InvalidBuffer;
2055 :
2056 : /*
2057 : * Despite the name, we also reset bulk relation extension state.
2058 : * Otherwise we can end up erroring out due to looking for free space in
2059 : * ->next_free of one partition, even though ->next_free was set when
2060 : * extending another partition. It could obviously also be bad for
2061 : * efficiency to look at existing blocks at offsets from another
2062 : * partition, even if we don't error out.
2063 : */
2064 161516 : bistate->next_free = InvalidBlockNumber;
2065 161516 : bistate->last_free = InvalidBlockNumber;
2066 161516 : }
2067 :
2068 :
2069 : /*
2070 : * heap_insert - insert tuple into a heap
2071 : *
2072 : * The new tuple is stamped with current transaction ID and the specified
2073 : * command ID.
2074 : *
2075 : * See table_tuple_insert for comments about most of the input flags, except
2076 : * that this routine directly takes a tuple rather than a slot.
2077 : *
2078 : * There's corresponding HEAP_INSERT_ options to all the TABLE_INSERT_
2079 : * options, and there additionally is HEAP_INSERT_SPECULATIVE which is used to
2080 : * implement table_tuple_insert_speculative().
2081 : *
2082 : * On return the header fields of *tup are updated to match the stored tuple;
2083 : * in particular tup->t_self receives the actual TID where the tuple was
2084 : * stored. But note that any toasting of fields within the tuple data is NOT
2085 : * reflected into *tup.
2086 : */
2087 : void
2088 16634354 : heap_insert(Relation relation, HeapTuple tup, CommandId cid,
2089 : int options, BulkInsertState bistate)
2090 : {
2091 16634354 : TransactionId xid = GetCurrentTransactionId();
2092 : HeapTuple heaptup;
2093 : Buffer buffer;
2094 16634342 : Buffer vmbuffer = InvalidBuffer;
2095 16634342 : bool all_visible_cleared = false;
2096 :
2097 : /* Cheap, simplistic check that the tuple matches the rel's rowtype. */
2098 : Assert(HeapTupleHeaderGetNatts(tup->t_data) <=
2099 : RelationGetNumberOfAttributes(relation));
2100 :
2101 16634342 : AssertHasSnapshotForToast(relation);
2102 :
2103 : /*
2104 : * Fill in tuple header fields and toast the tuple if necessary.
2105 : *
2106 : * Note: below this point, heaptup is the data we actually intend to store
2107 : * into the relation; tup is the caller's original untoasted data.
2108 : */
2109 16634342 : heaptup = heap_prepare_insert(relation, tup, xid, cid, options);
2110 :
2111 : /*
2112 : * Find buffer to insert this tuple into. If the page is all visible,
2113 : * this will also pin the requisite visibility map page.
2114 : */
2115 16634342 : buffer = RelationGetBufferForTuple(relation, heaptup->t_len,
2116 : InvalidBuffer, options, bistate,
2117 : &vmbuffer, NULL,
2118 : 0);
2119 :
2120 : /*
2121 : * We're about to do the actual insert -- but check for conflict first, to
2122 : * avoid possibly having to roll back work we've just done.
2123 : *
2124 : * This is safe without a recheck as long as there is no possibility of
2125 : * another process scanning the page between this check and the insert
2126 : * being visible to the scan (i.e., an exclusive buffer content lock is
2127 : * continuously held from this point until the tuple insert is visible).
2128 : *
2129 : * For a heap insert, we only need to check for table-level SSI locks. Our
2130 : * new tuple can't possibly conflict with existing tuple locks, and heap
2131 : * page locks are only consolidated versions of tuple locks; they do not
2132 : * lock "gaps" as index page locks do. So we don't need to specify a
2133 : * buffer when making the call, which makes for a faster check.
2134 : */
2135 16634342 : CheckForSerializableConflictIn(relation, NULL, InvalidBlockNumber);
2136 :
2137 : /* NO EREPORT(ERROR) from here till changes are logged */
2138 16634318 : START_CRIT_SECTION();
2139 :
2140 16634318 : RelationPutHeapTuple(relation, buffer, heaptup,
2141 16634318 : (options & HEAP_INSERT_SPECULATIVE) != 0);
2142 :
2143 16634318 : if (PageIsAllVisible(BufferGetPage(buffer)))
2144 : {
2145 14178 : all_visible_cleared = true;
2146 14178 : PageClearAllVisible(BufferGetPage(buffer));
2147 14178 : visibilitymap_clear(relation,
2148 14178 : ItemPointerGetBlockNumber(&(heaptup->t_self)),
2149 : vmbuffer, VISIBILITYMAP_VALID_BITS);
2150 : }
2151 :
2152 : /*
2153 : * XXX Should we set PageSetPrunable on this page ?
2154 : *
2155 : * The inserting transaction may eventually abort thus making this tuple
2156 : * DEAD and hence available for pruning. Though we don't want to optimize
2157 : * for aborts, if no other tuple in this page is UPDATEd/DELETEd, the
2158 : * aborted tuple will never be pruned until next vacuum is triggered.
2159 : *
2160 : * If you do add PageSetPrunable here, add it in heap_xlog_insert too.
2161 : */
2162 :
2163 16634318 : MarkBufferDirty(buffer);
2164 :
2165 : /* XLOG stuff */
2166 16634318 : if (RelationNeedsWAL(relation))
2167 : {
2168 : xl_heap_insert xlrec;
2169 : xl_heap_header xlhdr;
2170 : XLogRecPtr recptr;
2171 14132374 : Page page = BufferGetPage(buffer);
2172 14132374 : uint8 info = XLOG_HEAP_INSERT;
2173 14132374 : int bufflags = 0;
2174 :
2175 : /*
2176 : * If this is a catalog, we need to transmit combo CIDs to properly
2177 : * decode, so log that as well.
2178 : */
2179 14132374 : if (RelationIsAccessibleInLogicalDecoding(relation))
2180 6756 : log_heap_new_cid(relation, heaptup);
2181 :
2182 : /*
2183 : * If this is the single and first tuple on page, we can reinit the
2184 : * page instead of restoring the whole thing. Set flag, and hide
2185 : * buffer references from XLogInsert.
2186 : */
2187 14314528 : if (ItemPointerGetOffsetNumber(&(heaptup->t_self)) == FirstOffsetNumber &&
2188 182154 : PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
2189 : {
2190 180260 : info |= XLOG_HEAP_INIT_PAGE;
2191 180260 : bufflags |= REGBUF_WILL_INIT;
2192 : }
2193 :
2194 14132374 : xlrec.offnum = ItemPointerGetOffsetNumber(&heaptup->t_self);
2195 14132374 : xlrec.flags = 0;
2196 14132374 : if (all_visible_cleared)
2197 14172 : xlrec.flags |= XLH_INSERT_ALL_VISIBLE_CLEARED;
2198 14132374 : if (options & HEAP_INSERT_SPECULATIVE)
2199 4110 : xlrec.flags |= XLH_INSERT_IS_SPECULATIVE;
2200 : Assert(ItemPointerGetBlockNumber(&heaptup->t_self) == BufferGetBlockNumber(buffer));
2201 :
2202 : /*
2203 : * For logical decoding, we need the tuple even if we're doing a full
2204 : * page write, so make sure it's included even if we take a full-page
2205 : * image. (XXX We could alternatively store a pointer into the FPW).
2206 : */
2207 14132374 : if (RelationIsLogicallyLogged(relation) &&
2208 499816 : !(options & HEAP_INSERT_NO_LOGICAL))
2209 : {
2210 499762 : xlrec.flags |= XLH_INSERT_CONTAINS_NEW_TUPLE;
2211 499762 : bufflags |= REGBUF_KEEP_DATA;
2212 :
2213 499762 : if (IsToastRelation(relation))
2214 3572 : xlrec.flags |= XLH_INSERT_ON_TOAST_RELATION;
2215 : }
2216 :
2217 14132374 : XLogBeginInsert();
2218 14132374 : XLogRegisterData(&xlrec, SizeOfHeapInsert);
2219 :
2220 14132374 : xlhdr.t_infomask2 = heaptup->t_data->t_infomask2;
2221 14132374 : xlhdr.t_infomask = heaptup->t_data->t_infomask;
2222 14132374 : xlhdr.t_hoff = heaptup->t_data->t_hoff;
2223 :
2224 : /*
2225 : * note we mark xlhdr as belonging to buffer; if XLogInsert decides to
2226 : * write the whole page to the xlog, we don't need to store
2227 : * xl_heap_header in the xlog.
2228 : */
2229 14132374 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2230 14132374 : XLogRegisterBufData(0, &xlhdr, SizeOfHeapHeader);
2231 : /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
2232 14132374 : XLogRegisterBufData(0,
2233 14132374 : (char *) heaptup->t_data + SizeofHeapTupleHeader,
2234 14132374 : heaptup->t_len - SizeofHeapTupleHeader);
2235 :
2236 : /* filtering by origin on a row level is much more efficient */
2237 14132374 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
2238 :
2239 14132374 : recptr = XLogInsert(RM_HEAP_ID, info);
2240 :
2241 14132374 : PageSetLSN(page, recptr);
2242 : }
2243 :
2244 16634318 : END_CRIT_SECTION();
2245 :
2246 16634318 : UnlockReleaseBuffer(buffer);
2247 16634318 : if (vmbuffer != InvalidBuffer)
2248 14738 : ReleaseBuffer(vmbuffer);
2249 :
2250 : /*
2251 : * If tuple is cachable, mark it for invalidation from the caches in case
2252 : * we abort. Note it is OK to do this after releasing the buffer, because
2253 : * the heaptup data structure is all in local memory, not in the shared
2254 : * buffer.
2255 : */
2256 16634318 : CacheInvalidateHeapTuple(relation, heaptup, NULL);
2257 :
2258 : /* Note: speculative insertions are counted too, even if aborted later */
2259 16634318 : pgstat_count_heap_insert(relation, 1);
2260 :
2261 : /*
2262 : * If heaptup is a private copy, release it. Don't forget to copy t_self
2263 : * back to the caller's image, too.
2264 : */
2265 16634318 : if (heaptup != tup)
2266 : {
2267 36304 : tup->t_self = heaptup->t_self;
2268 36304 : heap_freetuple(heaptup);
2269 : }
2270 16634318 : }
2271 :
2272 : /*
2273 : * Subroutine for heap_insert(). Prepares a tuple for insertion. This sets the
2274 : * tuple header fields and toasts the tuple if necessary. Returns a toasted
2275 : * version of the tuple if it was toasted, or the original tuple if not. Note
2276 : * that in any case, the header fields are also set in the original tuple.
2277 : */
2278 : static HeapTuple
2279 19566522 : heap_prepare_insert(Relation relation, HeapTuple tup, TransactionId xid,
2280 : CommandId cid, int options)
2281 : {
2282 : /*
2283 : * To allow parallel inserts, we need to ensure that they are safe to be
2284 : * performed in workers. We have the infrastructure to allow parallel
2285 : * inserts in general except for the cases where inserts generate a new
2286 : * CommandId (eg. inserts into a table having a foreign key column).
2287 : */
2288 19566522 : if (IsParallelWorker())
2289 0 : ereport(ERROR,
2290 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
2291 : errmsg("cannot insert tuples in a parallel worker")));
2292 :
2293 19566522 : tup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
2294 19566522 : tup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
2295 19566522 : tup->t_data->t_infomask |= HEAP_XMAX_INVALID;
2296 19566522 : HeapTupleHeaderSetXmin(tup->t_data, xid);
2297 19566522 : if (options & HEAP_INSERT_FROZEN)
2298 204192 : HeapTupleHeaderSetXminFrozen(tup->t_data);
2299 :
2300 19566522 : HeapTupleHeaderSetCmin(tup->t_data, cid);
2301 19566522 : HeapTupleHeaderSetXmax(tup->t_data, 0); /* for cleanliness */
2302 19566522 : tup->t_tableOid = RelationGetRelid(relation);
2303 :
2304 : /*
2305 : * If the new tuple is too big for storage or contains already toasted
2306 : * out-of-line attributes from some other relation, invoke the toaster.
2307 : */
2308 19566522 : if (relation->rd_rel->relkind != RELKIND_RELATION &&
2309 61284 : relation->rd_rel->relkind != RELKIND_MATVIEW)
2310 : {
2311 : /* toast table entries should never be recursively toasted */
2312 : Assert(!HeapTupleHasExternal(tup));
2313 61188 : return tup;
2314 : }
2315 19505334 : else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
2316 36392 : return heap_toast_insert_or_update(relation, tup, NULL, options);
2317 : else
2318 19468942 : return tup;
2319 : }
2320 :
2321 : /*
2322 : * Helper for heap_multi_insert() that computes the number of entire pages
2323 : * that inserting the remaining heaptuples requires. Used to determine how
2324 : * much the relation needs to be extended by.
2325 : */
2326 : static int
2327 723296 : heap_multi_insert_pages(HeapTuple *heaptuples, int done, int ntuples, Size saveFreeSpace)
2328 : {
2329 723296 : size_t page_avail = BLCKSZ - SizeOfPageHeaderData - saveFreeSpace;
2330 723296 : int npages = 1;
2331 :
2332 4831830 : for (int i = done; i < ntuples; i++)
2333 : {
2334 4108534 : size_t tup_sz = sizeof(ItemIdData) + MAXALIGN(heaptuples[i]->t_len);
2335 :
2336 4108534 : if (page_avail < tup_sz)
2337 : {
2338 31064 : npages++;
2339 31064 : page_avail = BLCKSZ - SizeOfPageHeaderData - saveFreeSpace;
2340 : }
2341 4108534 : page_avail -= tup_sz;
2342 : }
2343 :
2344 723296 : return npages;
2345 : }
2346 :
2347 : /*
2348 : * heap_multi_insert - insert multiple tuples into a heap
2349 : *
2350 : * This is like heap_insert(), but inserts multiple tuples in one operation.
2351 : * That's faster than calling heap_insert() in a loop, because when multiple
2352 : * tuples can be inserted on a single page, we can write just a single WAL
2353 : * record covering all of them, and only need to lock/unlock the page once.
2354 : *
2355 : * Note: this leaks memory into the current memory context. You can create a
2356 : * temporary context before calling this, if that's a problem.
2357 : */
2358 : void
2359 710188 : heap_multi_insert(Relation relation, TupleTableSlot **slots, int ntuples,
2360 : CommandId cid, int options, BulkInsertState bistate)
2361 : {
2362 710188 : TransactionId xid = GetCurrentTransactionId();
2363 : HeapTuple *heaptuples;
2364 : int i;
2365 : int ndone;
2366 : PGAlignedBlock scratch;
2367 : Page page;
2368 710188 : Buffer vmbuffer = InvalidBuffer;
2369 : bool needwal;
2370 : Size saveFreeSpace;
2371 710188 : bool need_tuple_data = RelationIsLogicallyLogged(relation);
2372 710188 : bool need_cids = RelationIsAccessibleInLogicalDecoding(relation);
2373 710188 : bool starting_with_empty_page = false;
2374 710188 : int npages = 0;
2375 710188 : int npages_used = 0;
2376 :
2377 : /* currently not needed (thus unsupported) for heap_multi_insert() */
2378 : Assert(!(options & HEAP_INSERT_NO_LOGICAL));
2379 :
2380 710188 : AssertHasSnapshotForToast(relation);
2381 :
2382 710188 : needwal = RelationNeedsWAL(relation);
2383 710188 : saveFreeSpace = RelationGetTargetPageFreeSpace(relation,
2384 : HEAP_DEFAULT_FILLFACTOR);
2385 :
2386 : /* Toast and set header data in all the slots */
2387 710188 : heaptuples = palloc(ntuples * sizeof(HeapTuple));
2388 3642368 : for (i = 0; i < ntuples; i++)
2389 : {
2390 : HeapTuple tuple;
2391 :
2392 2932180 : tuple = ExecFetchSlotHeapTuple(slots[i], true, NULL);
2393 2932180 : slots[i]->tts_tableOid = RelationGetRelid(relation);
2394 2932180 : tuple->t_tableOid = slots[i]->tts_tableOid;
2395 2932180 : heaptuples[i] = heap_prepare_insert(relation, tuple, xid, cid,
2396 : options);
2397 : }
2398 :
2399 : /*
2400 : * We're about to do the actual inserts -- but check for conflict first,
2401 : * to minimize the possibility of having to roll back work we've just
2402 : * done.
2403 : *
2404 : * A check here does not definitively prevent a serialization anomaly;
2405 : * that check MUST be done at least past the point of acquiring an
2406 : * exclusive buffer content lock on every buffer that will be affected,
2407 : * and MAY be done after all inserts are reflected in the buffers and
2408 : * those locks are released; otherwise there is a race condition. Since
2409 : * multiple buffers can be locked and unlocked in the loop below, and it
2410 : * would not be feasible to identify and lock all of those buffers before
2411 : * the loop, we must do a final check at the end.
2412 : *
2413 : * The check here could be omitted with no loss of correctness; it is
2414 : * present strictly as an optimization.
2415 : *
2416 : * For heap inserts, we only need to check for table-level SSI locks. Our
2417 : * new tuples can't possibly conflict with existing tuple locks, and heap
2418 : * page locks are only consolidated versions of tuple locks; they do not
2419 : * lock "gaps" as index page locks do. So we don't need to specify a
2420 : * buffer when making the call, which makes for a faster check.
2421 : */
2422 710188 : CheckForSerializableConflictIn(relation, NULL, InvalidBlockNumber);
2423 :
2424 710188 : ndone = 0;
2425 1449786 : while (ndone < ntuples)
2426 : {
2427 : Buffer buffer;
2428 739598 : bool all_visible_cleared = false;
2429 739598 : bool all_frozen_set = false;
2430 : int nthispage;
2431 :
2432 739598 : CHECK_FOR_INTERRUPTS();
2433 :
2434 : /*
2435 : * Compute number of pages needed to fit the to-be-inserted tuples in
2436 : * the worst case. This will be used to determine how much to extend
2437 : * the relation by in RelationGetBufferForTuple(), if needed. If we
2438 : * filled a prior page from scratch, we can just update our last
2439 : * computation, but if we started with a partially filled page,
2440 : * recompute from scratch, the number of potentially required pages
2441 : * can vary due to tuples needing to fit onto the page, page headers
2442 : * etc.
2443 : */
2444 739598 : if (ndone == 0 || !starting_with_empty_page)
2445 : {
2446 723296 : npages = heap_multi_insert_pages(heaptuples, ndone, ntuples,
2447 : saveFreeSpace);
2448 723296 : npages_used = 0;
2449 : }
2450 : else
2451 16302 : npages_used++;
2452 :
2453 : /*
2454 : * Find buffer where at least the next tuple will fit. If the page is
2455 : * all-visible, this will also pin the requisite visibility map page.
2456 : *
2457 : * Also pin visibility map page if COPY FREEZE inserts tuples into an
2458 : * empty page. See all_frozen_set below.
2459 : */
2460 739598 : buffer = RelationGetBufferForTuple(relation, heaptuples[ndone]->t_len,
2461 : InvalidBuffer, options, bistate,
2462 : &vmbuffer, NULL,
2463 : npages - npages_used);
2464 739598 : page = BufferGetPage(buffer);
2465 :
2466 739598 : starting_with_empty_page = PageGetMaxOffsetNumber(page) == 0;
2467 :
2468 739598 : if (starting_with_empty_page && (options & HEAP_INSERT_FROZEN))
2469 3322 : all_frozen_set = true;
2470 :
2471 : /* NO EREPORT(ERROR) from here till changes are logged */
2472 739598 : START_CRIT_SECTION();
2473 :
2474 : /*
2475 : * RelationGetBufferForTuple has ensured that the first tuple fits.
2476 : * Put that on the page, and then as many other tuples as fit.
2477 : */
2478 739598 : RelationPutHeapTuple(relation, buffer, heaptuples[ndone], false);
2479 :
2480 : /*
2481 : * For logical decoding we need combo CIDs to properly decode the
2482 : * catalog.
2483 : */
2484 739598 : if (needwal && need_cids)
2485 9912 : log_heap_new_cid(relation, heaptuples[ndone]);
2486 :
2487 2932180 : for (nthispage = 1; ndone + nthispage < ntuples; nthispage++)
2488 : {
2489 2221992 : HeapTuple heaptup = heaptuples[ndone + nthispage];
2490 :
2491 2221992 : if (PageGetHeapFreeSpace(page) < MAXALIGN(heaptup->t_len) + saveFreeSpace)
2492 29410 : break;
2493 :
2494 2192582 : RelationPutHeapTuple(relation, buffer, heaptup, false);
2495 :
2496 : /*
2497 : * For logical decoding we need combo CIDs to properly decode the
2498 : * catalog.
2499 : */
2500 2192582 : if (needwal && need_cids)
2501 9354 : log_heap_new_cid(relation, heaptup);
2502 : }
2503 :
2504 : /*
2505 : * If the page is all visible, need to clear that, unless we're only
2506 : * going to add further frozen rows to it.
2507 : *
2508 : * If we're only adding already frozen rows to a previously empty
2509 : * page, mark it as all-visible.
2510 : */
2511 739598 : if (PageIsAllVisible(page) && !(options & HEAP_INSERT_FROZEN))
2512 : {
2513 5712 : all_visible_cleared = true;
2514 5712 : PageClearAllVisible(page);
2515 5712 : visibilitymap_clear(relation,
2516 : BufferGetBlockNumber(buffer),
2517 : vmbuffer, VISIBILITYMAP_VALID_BITS);
2518 : }
2519 733886 : else if (all_frozen_set)
2520 3322 : PageSetAllVisible(page);
2521 :
2522 : /*
2523 : * XXX Should we set PageSetPrunable on this page ? See heap_insert()
2524 : */
2525 :
2526 739598 : MarkBufferDirty(buffer);
2527 :
2528 : /* XLOG stuff */
2529 739598 : if (needwal)
2530 : {
2531 : XLogRecPtr recptr;
2532 : xl_heap_multi_insert *xlrec;
2533 731950 : uint8 info = XLOG_HEAP2_MULTI_INSERT;
2534 : char *tupledata;
2535 : int totaldatalen;
2536 731950 : char *scratchptr = scratch.data;
2537 : bool init;
2538 731950 : int bufflags = 0;
2539 :
2540 : /*
2541 : * If the page was previously empty, we can reinit the page
2542 : * instead of restoring the whole thing.
2543 : */
2544 731950 : init = starting_with_empty_page;
2545 :
2546 : /* allocate xl_heap_multi_insert struct from the scratch area */
2547 731950 : xlrec = (xl_heap_multi_insert *) scratchptr;
2548 731950 : scratchptr += SizeOfHeapMultiInsert;
2549 :
2550 : /*
2551 : * Allocate offsets array. Unless we're reinitializing the page,
2552 : * in that case the tuples are stored in order starting at
2553 : * FirstOffsetNumber and we don't need to store the offsets
2554 : * explicitly.
2555 : */
2556 731950 : if (!init)
2557 705818 : scratchptr += nthispage * sizeof(OffsetNumber);
2558 :
2559 : /* the rest of the scratch space is used for tuple data */
2560 731950 : tupledata = scratchptr;
2561 :
2562 : /* check that the mutually exclusive flags are not both set */
2563 : Assert(!(all_visible_cleared && all_frozen_set));
2564 :
2565 731950 : xlrec->flags = 0;
2566 731950 : if (all_visible_cleared)
2567 5712 : xlrec->flags = XLH_INSERT_ALL_VISIBLE_CLEARED;
2568 731950 : if (all_frozen_set)
2569 26 : xlrec->flags = XLH_INSERT_ALL_FROZEN_SET;
2570 :
2571 731950 : xlrec->ntuples = nthispage;
2572 :
2573 : /*
2574 : * Write out an xl_multi_insert_tuple and the tuple data itself
2575 : * for each tuple.
2576 : */
2577 3253316 : for (i = 0; i < nthispage; i++)
2578 : {
2579 2521366 : HeapTuple heaptup = heaptuples[ndone + i];
2580 : xl_multi_insert_tuple *tuphdr;
2581 : int datalen;
2582 :
2583 2521366 : if (!init)
2584 1481282 : xlrec->offsets[i] = ItemPointerGetOffsetNumber(&heaptup->t_self);
2585 : /* xl_multi_insert_tuple needs two-byte alignment. */
2586 2521366 : tuphdr = (xl_multi_insert_tuple *) SHORTALIGN(scratchptr);
2587 2521366 : scratchptr = ((char *) tuphdr) + SizeOfMultiInsertTuple;
2588 :
2589 2521366 : tuphdr->t_infomask2 = heaptup->t_data->t_infomask2;
2590 2521366 : tuphdr->t_infomask = heaptup->t_data->t_infomask;
2591 2521366 : tuphdr->t_hoff = heaptup->t_data->t_hoff;
2592 :
2593 : /* write bitmap [+ padding] [+ oid] + data */
2594 2521366 : datalen = heaptup->t_len - SizeofHeapTupleHeader;
2595 2521366 : memcpy(scratchptr,
2596 2521366 : (char *) heaptup->t_data + SizeofHeapTupleHeader,
2597 : datalen);
2598 2521366 : tuphdr->datalen = datalen;
2599 2521366 : scratchptr += datalen;
2600 : }
2601 731950 : totaldatalen = scratchptr - tupledata;
2602 : Assert((scratchptr - scratch.data) < BLCKSZ);
2603 :
2604 731950 : if (need_tuple_data)
2605 144 : xlrec->flags |= XLH_INSERT_CONTAINS_NEW_TUPLE;
2606 :
2607 : /*
2608 : * Signal that this is the last xl_heap_multi_insert record
2609 : * emitted by this call to heap_multi_insert(). Needed for logical
2610 : * decoding so it knows when to cleanup temporary data.
2611 : */
2612 731950 : if (ndone + nthispage == ntuples)
2613 709358 : xlrec->flags |= XLH_INSERT_LAST_IN_MULTI;
2614 :
2615 731950 : if (init)
2616 : {
2617 26132 : info |= XLOG_HEAP_INIT_PAGE;
2618 26132 : bufflags |= REGBUF_WILL_INIT;
2619 : }
2620 :
2621 : /*
2622 : * If we're doing logical decoding, include the new tuple data
2623 : * even if we take a full-page image of the page.
2624 : */
2625 731950 : if (need_tuple_data)
2626 144 : bufflags |= REGBUF_KEEP_DATA;
2627 :
2628 731950 : XLogBeginInsert();
2629 731950 : XLogRegisterData(xlrec, tupledata - scratch.data);
2630 731950 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2631 :
2632 731950 : XLogRegisterBufData(0, tupledata, totaldatalen);
2633 :
2634 : /* filtering by origin on a row level is much more efficient */
2635 731950 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
2636 :
2637 731950 : recptr = XLogInsert(RM_HEAP2_ID, info);
2638 :
2639 731950 : PageSetLSN(page, recptr);
2640 : }
2641 :
2642 739598 : END_CRIT_SECTION();
2643 :
2644 : /*
2645 : * If we've frozen everything on the page, update the visibilitymap.
2646 : * We're already holding pin on the vmbuffer.
2647 : */
2648 739598 : if (all_frozen_set)
2649 : {
2650 : Assert(PageIsAllVisible(page));
2651 : Assert(visibilitymap_pin_ok(BufferGetBlockNumber(buffer), vmbuffer));
2652 :
2653 : /*
2654 : * It's fine to use InvalidTransactionId here - this is only used
2655 : * when HEAP_INSERT_FROZEN is specified, which intentionally
2656 : * violates visibility rules.
2657 : */
2658 3322 : visibilitymap_set(relation, BufferGetBlockNumber(buffer), buffer,
2659 : InvalidXLogRecPtr, vmbuffer,
2660 : InvalidTransactionId,
2661 : VISIBILITYMAP_ALL_VISIBLE | VISIBILITYMAP_ALL_FROZEN);
2662 : }
2663 :
2664 739598 : UnlockReleaseBuffer(buffer);
2665 739598 : ndone += nthispage;
2666 :
2667 : /*
2668 : * NB: Only release vmbuffer after inserting all tuples - it's fairly
2669 : * likely that we'll insert into subsequent heap pages that are likely
2670 : * to use the same vm page.
2671 : */
2672 : }
2673 :
2674 : /* We're done with inserting all tuples, so release the last vmbuffer. */
2675 710188 : if (vmbuffer != InvalidBuffer)
2676 5924 : ReleaseBuffer(vmbuffer);
2677 :
2678 : /*
2679 : * We're done with the actual inserts. Check for conflicts again, to
2680 : * ensure that all rw-conflicts in to these inserts are detected. Without
2681 : * this final check, a sequential scan of the heap may have locked the
2682 : * table after the "before" check, missing one opportunity to detect the
2683 : * conflict, and then scanned the table before the new tuples were there,
2684 : * missing the other chance to detect the conflict.
2685 : *
2686 : * For heap inserts, we only need to check for table-level SSI locks. Our
2687 : * new tuples can't possibly conflict with existing tuple locks, and heap
2688 : * page locks are only consolidated versions of tuple locks; they do not
2689 : * lock "gaps" as index page locks do. So we don't need to specify a
2690 : * buffer when making the call.
2691 : */
2692 710188 : CheckForSerializableConflictIn(relation, NULL, InvalidBlockNumber);
2693 :
2694 : /*
2695 : * If tuples are cachable, mark them for invalidation from the caches in
2696 : * case we abort. Note it is OK to do this after releasing the buffer,
2697 : * because the heaptuples data structure is all in local memory, not in
2698 : * the shared buffer.
2699 : */
2700 710188 : if (IsCatalogRelation(relation))
2701 : {
2702 2436828 : for (i = 0; i < ntuples; i++)
2703 1729066 : CacheInvalidateHeapTuple(relation, heaptuples[i], NULL);
2704 : }
2705 :
2706 : /* copy t_self fields back to the caller's slots */
2707 3642368 : for (i = 0; i < ntuples; i++)
2708 2932180 : slots[i]->tts_tid = heaptuples[i]->t_self;
2709 :
2710 710188 : pgstat_count_heap_insert(relation, ntuples);
2711 710188 : }
2712 :
2713 : /*
2714 : * simple_heap_insert - insert a tuple
2715 : *
2716 : * Currently, this routine differs from heap_insert only in supplying
2717 : * a default command ID and not allowing access to the speedup options.
2718 : *
2719 : * This should be used rather than using heap_insert directly in most places
2720 : * where we are modifying system catalogs.
2721 : */
2722 : void
2723 1794260 : simple_heap_insert(Relation relation, HeapTuple tup)
2724 : {
2725 1794260 : heap_insert(relation, tup, GetCurrentCommandId(true), 0, NULL);
2726 1794260 : }
2727 :
2728 : /*
2729 : * Given infomask/infomask2, compute the bits that must be saved in the
2730 : * "infobits" field of xl_heap_delete, xl_heap_update, xl_heap_lock,
2731 : * xl_heap_lock_updated WAL records.
2732 : *
2733 : * See fix_infomask_from_infobits.
2734 : */
2735 : static uint8
2736 3876316 : compute_infobits(uint16 infomask, uint16 infomask2)
2737 : {
2738 : return
2739 3876316 : ((infomask & HEAP_XMAX_IS_MULTI) != 0 ? XLHL_XMAX_IS_MULTI : 0) |
2740 3876316 : ((infomask & HEAP_XMAX_LOCK_ONLY) != 0 ? XLHL_XMAX_LOCK_ONLY : 0) |
2741 3876316 : ((infomask & HEAP_XMAX_EXCL_LOCK) != 0 ? XLHL_XMAX_EXCL_LOCK : 0) |
2742 : /* note we ignore HEAP_XMAX_SHR_LOCK here */
2743 7752632 : ((infomask & HEAP_XMAX_KEYSHR_LOCK) != 0 ? XLHL_XMAX_KEYSHR_LOCK : 0) |
2744 : ((infomask2 & HEAP_KEYS_UPDATED) != 0 ?
2745 3876316 : XLHL_KEYS_UPDATED : 0);
2746 : }
2747 :
2748 : /*
2749 : * Given two versions of the same t_infomask for a tuple, compare them and
2750 : * return whether the relevant status for a tuple Xmax has changed. This is
2751 : * used after a buffer lock has been released and reacquired: we want to ensure
2752 : * that the tuple state continues to be the same it was when we previously
2753 : * examined it.
2754 : *
2755 : * Note the Xmax field itself must be compared separately.
2756 : */
2757 : static inline bool
2758 10658 : xmax_infomask_changed(uint16 new_infomask, uint16 old_infomask)
2759 : {
2760 10658 : const uint16 interesting =
2761 : HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY | HEAP_LOCK_MASK;
2762 :
2763 10658 : if ((new_infomask & interesting) != (old_infomask & interesting))
2764 28 : return true;
2765 :
2766 10630 : return false;
2767 : }
2768 :
2769 : /*
2770 : * heap_delete - delete a tuple
2771 : *
2772 : * See table_tuple_delete() for an explanation of the parameters, except that
2773 : * this routine directly takes a tuple rather than a slot.
2774 : *
2775 : * In the failure cases, the routine fills *tmfd with the tuple's t_ctid,
2776 : * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax (the last
2777 : * only for TM_SelfModified, since we cannot obtain cmax from a combo CID
2778 : * generated by another transaction).
2779 : */
2780 : TM_Result
2781 2974976 : heap_delete(Relation relation, ItemPointer tid,
2782 : CommandId cid, Snapshot crosscheck, bool wait,
2783 : TM_FailureData *tmfd, bool changingPart)
2784 : {
2785 : TM_Result result;
2786 2974976 : TransactionId xid = GetCurrentTransactionId();
2787 : ItemId lp;
2788 : HeapTupleData tp;
2789 : Page page;
2790 : BlockNumber block;
2791 : Buffer buffer;
2792 2974976 : Buffer vmbuffer = InvalidBuffer;
2793 : TransactionId new_xmax;
2794 : uint16 new_infomask,
2795 : new_infomask2;
2796 2974976 : bool have_tuple_lock = false;
2797 : bool iscombo;
2798 2974976 : bool all_visible_cleared = false;
2799 2974976 : HeapTuple old_key_tuple = NULL; /* replica identity of the tuple */
2800 2974976 : bool old_key_copied = false;
2801 :
2802 : Assert(ItemPointerIsValid(tid));
2803 :
2804 2974976 : AssertHasSnapshotForToast(relation);
2805 :
2806 : /*
2807 : * Forbid this during a parallel operation, lest it allocate a combo CID.
2808 : * Other workers might need that combo CID for visibility checks, and we
2809 : * have no provision for broadcasting it to them.
2810 : */
2811 2974976 : if (IsInParallelMode())
2812 0 : ereport(ERROR,
2813 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
2814 : errmsg("cannot delete tuples during a parallel operation")));
2815 :
2816 2974976 : block = ItemPointerGetBlockNumber(tid);
2817 2974976 : buffer = ReadBuffer(relation, block);
2818 2974976 : page = BufferGetPage(buffer);
2819 :
2820 : /*
2821 : * Before locking the buffer, pin the visibility map page if it appears to
2822 : * be necessary. Since we haven't got the lock yet, someone else might be
2823 : * in the middle of changing this, so we'll need to recheck after we have
2824 : * the lock.
2825 : */
2826 2974976 : if (PageIsAllVisible(page))
2827 362 : visibilitymap_pin(relation, block, &vmbuffer);
2828 :
2829 2974976 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2830 :
2831 2974976 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
2832 : Assert(ItemIdIsNormal(lp));
2833 :
2834 2974976 : tp.t_tableOid = RelationGetRelid(relation);
2835 2974976 : tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
2836 2974976 : tp.t_len = ItemIdGetLength(lp);
2837 2974976 : tp.t_self = *tid;
2838 :
2839 2 : l1:
2840 :
2841 : /*
2842 : * If we didn't pin the visibility map page and the page has become all
2843 : * visible while we were busy locking the buffer, we'll have to unlock and
2844 : * re-lock, to avoid holding the buffer lock across an I/O. That's a bit
2845 : * unfortunate, but hopefully shouldn't happen often.
2846 : */
2847 2974978 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
2848 : {
2849 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2850 0 : visibilitymap_pin(relation, block, &vmbuffer);
2851 0 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2852 : }
2853 :
2854 2974978 : result = HeapTupleSatisfiesUpdate(&tp, cid, buffer);
2855 :
2856 2974978 : if (result == TM_Invisible)
2857 : {
2858 0 : UnlockReleaseBuffer(buffer);
2859 0 : ereport(ERROR,
2860 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
2861 : errmsg("attempted to delete invisible tuple")));
2862 : }
2863 2974978 : else if (result == TM_BeingModified && wait)
2864 : {
2865 : TransactionId xwait;
2866 : uint16 infomask;
2867 :
2868 : /* must copy state data before unlocking buffer */
2869 81088 : xwait = HeapTupleHeaderGetRawXmax(tp.t_data);
2870 81088 : infomask = tp.t_data->t_infomask;
2871 :
2872 : /*
2873 : * Sleep until concurrent transaction ends -- except when there's a
2874 : * single locker and it's our own transaction. Note we don't care
2875 : * which lock mode the locker has, because we need the strongest one.
2876 : *
2877 : * Before sleeping, we need to acquire tuple lock to establish our
2878 : * priority for the tuple (see heap_lock_tuple). LockTuple will
2879 : * release us when we are next-in-line for the tuple.
2880 : *
2881 : * If we are forced to "start over" below, we keep the tuple lock;
2882 : * this arranges that we stay at the head of the line while rechecking
2883 : * tuple state.
2884 : */
2885 81088 : if (infomask & HEAP_XMAX_IS_MULTI)
2886 : {
2887 16 : bool current_is_member = false;
2888 :
2889 16 : if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
2890 : LockTupleExclusive, ¤t_is_member))
2891 : {
2892 16 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2893 :
2894 : /*
2895 : * Acquire the lock, if necessary (but skip it when we're
2896 : * requesting a lock and already have one; avoids deadlock).
2897 : */
2898 16 : if (!current_is_member)
2899 12 : heap_acquire_tuplock(relation, &(tp.t_self), LockTupleExclusive,
2900 : LockWaitBlock, &have_tuple_lock);
2901 :
2902 : /* wait for multixact */
2903 16 : MultiXactIdWait((MultiXactId) xwait, MultiXactStatusUpdate, infomask,
2904 : relation, &(tp.t_self), XLTW_Delete,
2905 : NULL);
2906 16 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2907 :
2908 : /*
2909 : * If xwait had just locked the tuple then some other xact
2910 : * could update this tuple before we get to this point. Check
2911 : * for xmax change, and start over if so.
2912 : *
2913 : * We also must start over if we didn't pin the VM page, and
2914 : * the page has become all visible.
2915 : */
2916 32 : if ((vmbuffer == InvalidBuffer && PageIsAllVisible(page)) ||
2917 32 : xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2918 16 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2919 : xwait))
2920 0 : goto l1;
2921 : }
2922 :
2923 : /*
2924 : * You might think the multixact is necessarily done here, but not
2925 : * so: it could have surviving members, namely our own xact or
2926 : * other subxacts of this backend. It is legal for us to delete
2927 : * the tuple in either case, however (the latter case is
2928 : * essentially a situation of upgrading our former shared lock to
2929 : * exclusive). We don't bother changing the on-disk hint bits
2930 : * since we are about to overwrite the xmax altogether.
2931 : */
2932 : }
2933 81072 : else if (!TransactionIdIsCurrentTransactionId(xwait))
2934 : {
2935 : /*
2936 : * Wait for regular transaction to end; but first, acquire tuple
2937 : * lock.
2938 : */
2939 80 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2940 80 : heap_acquire_tuplock(relation, &(tp.t_self), LockTupleExclusive,
2941 : LockWaitBlock, &have_tuple_lock);
2942 80 : XactLockTableWait(xwait, relation, &(tp.t_self), XLTW_Delete);
2943 72 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2944 :
2945 : /*
2946 : * xwait is done, but if xwait had just locked the tuple then some
2947 : * other xact could update this tuple before we get to this point.
2948 : * Check for xmax change, and start over if so.
2949 : *
2950 : * We also must start over if we didn't pin the VM page, and the
2951 : * page has become all visible.
2952 : */
2953 144 : if ((vmbuffer == InvalidBuffer && PageIsAllVisible(page)) ||
2954 142 : xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2955 70 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2956 : xwait))
2957 2 : goto l1;
2958 :
2959 : /* Otherwise check if it committed or aborted */
2960 70 : UpdateXmaxHintBits(tp.t_data, buffer, xwait);
2961 : }
2962 :
2963 : /*
2964 : * We may overwrite if previous xmax aborted, or if it committed but
2965 : * only locked the tuple without updating it.
2966 : */
2967 162128 : if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
2968 81100 : HEAP_XMAX_IS_LOCKED_ONLY(tp.t_data->t_infomask) ||
2969 50 : HeapTupleHeaderIsOnlyLocked(tp.t_data))
2970 81036 : result = TM_Ok;
2971 42 : else if (!ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid))
2972 34 : result = TM_Updated;
2973 : else
2974 8 : result = TM_Deleted;
2975 : }
2976 :
2977 : /* sanity check the result HeapTupleSatisfiesUpdate() and the logic above */
2978 : if (result != TM_Ok)
2979 : {
2980 : Assert(result == TM_SelfModified ||
2981 : result == TM_Updated ||
2982 : result == TM_Deleted ||
2983 : result == TM_BeingModified);
2984 : Assert(!(tp.t_data->t_infomask & HEAP_XMAX_INVALID));
2985 : Assert(result != TM_Updated ||
2986 : !ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid));
2987 : }
2988 :
2989 2974968 : if (crosscheck != InvalidSnapshot && result == TM_Ok)
2990 : {
2991 : /* Perform additional check for transaction-snapshot mode RI updates */
2992 2 : if (!HeapTupleSatisfiesVisibility(&tp, crosscheck, buffer))
2993 2 : result = TM_Updated;
2994 : }
2995 :
2996 2974968 : if (result != TM_Ok)
2997 : {
2998 112 : tmfd->ctid = tp.t_data->t_ctid;
2999 112 : tmfd->xmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
3000 112 : if (result == TM_SelfModified)
3001 42 : tmfd->cmax = HeapTupleHeaderGetCmax(tp.t_data);
3002 : else
3003 70 : tmfd->cmax = InvalidCommandId;
3004 112 : UnlockReleaseBuffer(buffer);
3005 112 : if (have_tuple_lock)
3006 42 : UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
3007 112 : if (vmbuffer != InvalidBuffer)
3008 0 : ReleaseBuffer(vmbuffer);
3009 112 : return result;
3010 : }
3011 :
3012 : /*
3013 : * We're about to do the actual delete -- check for conflict first, to
3014 : * avoid possibly having to roll back work we've just done.
3015 : *
3016 : * This is safe without a recheck as long as there is no possibility of
3017 : * another process scanning the page between this check and the delete
3018 : * being visible to the scan (i.e., an exclusive buffer content lock is
3019 : * continuously held from this point until the tuple delete is visible).
3020 : */
3021 2974856 : CheckForSerializableConflictIn(relation, tid, BufferGetBlockNumber(buffer));
3022 :
3023 : /* replace cid with a combo CID if necessary */
3024 2974828 : HeapTupleHeaderAdjustCmax(tp.t_data, &cid, &iscombo);
3025 :
3026 : /*
3027 : * Compute replica identity tuple before entering the critical section so
3028 : * we don't PANIC upon a memory allocation failure.
3029 : */
3030 2974828 : old_key_tuple = ExtractReplicaIdentity(relation, &tp, true, &old_key_copied);
3031 :
3032 : /*
3033 : * If this is the first possibly-multixact-able operation in the current
3034 : * transaction, set my per-backend OldestMemberMXactId setting. We can be
3035 : * certain that the transaction will never become a member of any older
3036 : * MultiXactIds than that. (We have to do this even if we end up just
3037 : * using our own TransactionId below, since some other backend could
3038 : * incorporate our XID into a MultiXact immediately afterwards.)
3039 : */
3040 2974828 : MultiXactIdSetOldestMember();
3041 :
3042 2974828 : compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(tp.t_data),
3043 2974828 : tp.t_data->t_infomask, tp.t_data->t_infomask2,
3044 : xid, LockTupleExclusive, true,
3045 : &new_xmax, &new_infomask, &new_infomask2);
3046 :
3047 2974828 : START_CRIT_SECTION();
3048 :
3049 : /*
3050 : * If this transaction commits, the tuple will become DEAD sooner or
3051 : * later. Set flag that this page is a candidate for pruning once our xid
3052 : * falls below the OldestXmin horizon. If the transaction finally aborts,
3053 : * the subsequent page pruning will be a no-op and the hint will be
3054 : * cleared.
3055 : */
3056 2974828 : PageSetPrunable(page, xid);
3057 :
3058 2974828 : if (PageIsAllVisible(page))
3059 : {
3060 362 : all_visible_cleared = true;
3061 362 : PageClearAllVisible(page);
3062 362 : visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
3063 : vmbuffer, VISIBILITYMAP_VALID_BITS);
3064 : }
3065 :
3066 : /* store transaction information of xact deleting the tuple */
3067 2974828 : tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3068 2974828 : tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3069 2974828 : tp.t_data->t_infomask |= new_infomask;
3070 2974828 : tp.t_data->t_infomask2 |= new_infomask2;
3071 2974828 : HeapTupleHeaderClearHotUpdated(tp.t_data);
3072 2974828 : HeapTupleHeaderSetXmax(tp.t_data, new_xmax);
3073 2974828 : HeapTupleHeaderSetCmax(tp.t_data, cid, iscombo);
3074 : /* Make sure there is no forward chain link in t_ctid */
3075 2974828 : tp.t_data->t_ctid = tp.t_self;
3076 :
3077 : /* Signal that this is actually a move into another partition */
3078 2974828 : if (changingPart)
3079 964 : HeapTupleHeaderSetMovedPartitions(tp.t_data);
3080 :
3081 2974828 : MarkBufferDirty(buffer);
3082 :
3083 : /*
3084 : * XLOG stuff
3085 : *
3086 : * NB: heap_abort_speculative() uses the same xlog record and replay
3087 : * routines.
3088 : */
3089 2974828 : if (RelationNeedsWAL(relation))
3090 : {
3091 : xl_heap_delete xlrec;
3092 : xl_heap_header xlhdr;
3093 : XLogRecPtr recptr;
3094 :
3095 : /*
3096 : * For logical decode we need combo CIDs to properly decode the
3097 : * catalog
3098 : */
3099 2849590 : if (RelationIsAccessibleInLogicalDecoding(relation))
3100 12484 : log_heap_new_cid(relation, &tp);
3101 :
3102 2849590 : xlrec.flags = 0;
3103 2849590 : if (all_visible_cleared)
3104 362 : xlrec.flags |= XLH_DELETE_ALL_VISIBLE_CLEARED;
3105 2849590 : if (changingPart)
3106 964 : xlrec.flags |= XLH_DELETE_IS_PARTITION_MOVE;
3107 5699180 : xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
3108 2849590 : tp.t_data->t_infomask2);
3109 2849590 : xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
3110 2849590 : xlrec.xmax = new_xmax;
3111 :
3112 2849590 : if (old_key_tuple != NULL)
3113 : {
3114 94036 : if (relation->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
3115 260 : xlrec.flags |= XLH_DELETE_CONTAINS_OLD_TUPLE;
3116 : else
3117 93776 : xlrec.flags |= XLH_DELETE_CONTAINS_OLD_KEY;
3118 : }
3119 :
3120 2849590 : XLogBeginInsert();
3121 2849590 : XLogRegisterData(&xlrec, SizeOfHeapDelete);
3122 :
3123 2849590 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
3124 :
3125 : /*
3126 : * Log replica identity of the deleted tuple if there is one
3127 : */
3128 2849590 : if (old_key_tuple != NULL)
3129 : {
3130 94036 : xlhdr.t_infomask2 = old_key_tuple->t_data->t_infomask2;
3131 94036 : xlhdr.t_infomask = old_key_tuple->t_data->t_infomask;
3132 94036 : xlhdr.t_hoff = old_key_tuple->t_data->t_hoff;
3133 :
3134 94036 : XLogRegisterData(&xlhdr, SizeOfHeapHeader);
3135 94036 : XLogRegisterData((char *) old_key_tuple->t_data
3136 : + SizeofHeapTupleHeader,
3137 94036 : old_key_tuple->t_len
3138 : - SizeofHeapTupleHeader);
3139 : }
3140 :
3141 : /* filtering by origin on a row level is much more efficient */
3142 2849590 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
3143 :
3144 2849590 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
3145 :
3146 2849590 : PageSetLSN(page, recptr);
3147 : }
3148 :
3149 2974828 : END_CRIT_SECTION();
3150 :
3151 2974828 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3152 :
3153 2974828 : if (vmbuffer != InvalidBuffer)
3154 362 : ReleaseBuffer(vmbuffer);
3155 :
3156 : /*
3157 : * If the tuple has toasted out-of-line attributes, we need to delete
3158 : * those items too. We have to do this before releasing the buffer
3159 : * because we need to look at the contents of the tuple, but it's OK to
3160 : * release the content lock on the buffer first.
3161 : */
3162 2974828 : if (relation->rd_rel->relkind != RELKIND_RELATION &&
3163 4806 : relation->rd_rel->relkind != RELKIND_MATVIEW)
3164 : {
3165 : /* toast table entries should never be recursively toasted */
3166 : Assert(!HeapTupleHasExternal(&tp));
3167 : }
3168 2970042 : else if (HeapTupleHasExternal(&tp))
3169 598 : heap_toast_delete(relation, &tp, false);
3170 :
3171 : /*
3172 : * Mark tuple for invalidation from system caches at next command
3173 : * boundary. We have to do this before releasing the buffer because we
3174 : * need to look at the contents of the tuple.
3175 : */
3176 2974828 : CacheInvalidateHeapTuple(relation, &tp, NULL);
3177 :
3178 : /* Now we can release the buffer */
3179 2974828 : ReleaseBuffer(buffer);
3180 :
3181 : /*
3182 : * Release the lmgr tuple lock, if we had it.
3183 : */
3184 2974828 : if (have_tuple_lock)
3185 40 : UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
3186 :
3187 2974828 : pgstat_count_heap_delete(relation);
3188 :
3189 2974828 : if (old_key_tuple != NULL && old_key_copied)
3190 93778 : heap_freetuple(old_key_tuple);
3191 :
3192 2974828 : return TM_Ok;
3193 : }
3194 :
3195 : /*
3196 : * simple_heap_delete - delete a tuple
3197 : *
3198 : * This routine may be used to delete a tuple when concurrent updates of
3199 : * the target tuple are not expected (for example, because we have a lock
3200 : * on the relation associated with the tuple). Any failure is reported
3201 : * via ereport().
3202 : */
3203 : void
3204 1243120 : simple_heap_delete(Relation relation, ItemPointer tid)
3205 : {
3206 : TM_Result result;
3207 : TM_FailureData tmfd;
3208 :
3209 1243120 : result = heap_delete(relation, tid,
3210 : GetCurrentCommandId(true), InvalidSnapshot,
3211 : true /* wait for commit */ ,
3212 : &tmfd, false /* changingPart */ );
3213 1243120 : switch (result)
3214 : {
3215 0 : case TM_SelfModified:
3216 : /* Tuple was already updated in current command? */
3217 0 : elog(ERROR, "tuple already updated by self");
3218 : break;
3219 :
3220 1243120 : case TM_Ok:
3221 : /* done successfully */
3222 1243120 : break;
3223 :
3224 0 : case TM_Updated:
3225 0 : elog(ERROR, "tuple concurrently updated");
3226 : break;
3227 :
3228 0 : case TM_Deleted:
3229 0 : elog(ERROR, "tuple concurrently deleted");
3230 : break;
3231 :
3232 0 : default:
3233 0 : elog(ERROR, "unrecognized heap_delete status: %u", result);
3234 : break;
3235 : }
3236 1243120 : }
3237 :
3238 : /*
3239 : * heap_update - replace a tuple
3240 : *
3241 : * See table_tuple_update() for an explanation of the parameters, except that
3242 : * this routine directly takes a tuple rather than a slot.
3243 : *
3244 : * In the failure cases, the routine fills *tmfd with the tuple's t_ctid,
3245 : * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax (the last
3246 : * only for TM_SelfModified, since we cannot obtain cmax from a combo CID
3247 : * generated by another transaction).
3248 : */
3249 : TM_Result
3250 606556 : heap_update(Relation relation, ItemPointer otid, HeapTuple newtup,
3251 : CommandId cid, Snapshot crosscheck, bool wait,
3252 : TM_FailureData *tmfd, LockTupleMode *lockmode,
3253 : TU_UpdateIndexes *update_indexes)
3254 : {
3255 : TM_Result result;
3256 606556 : TransactionId xid = GetCurrentTransactionId();
3257 : Bitmapset *hot_attrs;
3258 : Bitmapset *sum_attrs;
3259 : Bitmapset *key_attrs;
3260 : Bitmapset *id_attrs;
3261 : Bitmapset *interesting_attrs;
3262 : Bitmapset *modified_attrs;
3263 : ItemId lp;
3264 : HeapTupleData oldtup;
3265 : HeapTuple heaptup;
3266 606556 : HeapTuple old_key_tuple = NULL;
3267 606556 : bool old_key_copied = false;
3268 : Page page;
3269 : BlockNumber block;
3270 : MultiXactStatus mxact_status;
3271 : Buffer buffer,
3272 : newbuf,
3273 606556 : vmbuffer = InvalidBuffer,
3274 606556 : vmbuffer_new = InvalidBuffer;
3275 : bool need_toast;
3276 : Size newtupsize,
3277 : pagefree;
3278 606556 : bool have_tuple_lock = false;
3279 : bool iscombo;
3280 606556 : bool use_hot_update = false;
3281 606556 : bool summarized_update = false;
3282 : bool key_intact;
3283 606556 : bool all_visible_cleared = false;
3284 606556 : bool all_visible_cleared_new = false;
3285 : bool checked_lockers;
3286 : bool locker_remains;
3287 606556 : bool id_has_external = false;
3288 : TransactionId xmax_new_tuple,
3289 : xmax_old_tuple;
3290 : uint16 infomask_old_tuple,
3291 : infomask2_old_tuple,
3292 : infomask_new_tuple,
3293 : infomask2_new_tuple;
3294 :
3295 : Assert(ItemPointerIsValid(otid));
3296 :
3297 : /* Cheap, simplistic check that the tuple matches the rel's rowtype. */
3298 : Assert(HeapTupleHeaderGetNatts(newtup->t_data) <=
3299 : RelationGetNumberOfAttributes(relation));
3300 :
3301 606556 : AssertHasSnapshotForToast(relation);
3302 :
3303 : /*
3304 : * Forbid this during a parallel operation, lest it allocate a combo CID.
3305 : * Other workers might need that combo CID for visibility checks, and we
3306 : * have no provision for broadcasting it to them.
3307 : */
3308 606556 : if (IsInParallelMode())
3309 0 : ereport(ERROR,
3310 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
3311 : errmsg("cannot update tuples during a parallel operation")));
3312 :
3313 : #ifdef USE_ASSERT_CHECKING
3314 : check_lock_if_inplace_updateable_rel(relation, otid, newtup);
3315 : #endif
3316 :
3317 : /*
3318 : * Fetch the list of attributes to be checked for various operations.
3319 : *
3320 : * For HOT considerations, this is wasted effort if we fail to update or
3321 : * have to put the new tuple on a different page. But we must compute the
3322 : * list before obtaining buffer lock --- in the worst case, if we are
3323 : * doing an update on one of the relevant system catalogs, we could
3324 : * deadlock if we try to fetch the list later. In any case, the relcache
3325 : * caches the data so this is usually pretty cheap.
3326 : *
3327 : * We also need columns used by the replica identity and columns that are
3328 : * considered the "key" of rows in the table.
3329 : *
3330 : * Note that we get copies of each bitmap, so we need not worry about
3331 : * relcache flush happening midway through.
3332 : */
3333 606556 : hot_attrs = RelationGetIndexAttrBitmap(relation,
3334 : INDEX_ATTR_BITMAP_HOT_BLOCKING);
3335 606556 : sum_attrs = RelationGetIndexAttrBitmap(relation,
3336 : INDEX_ATTR_BITMAP_SUMMARIZED);
3337 606556 : key_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_KEY);
3338 606556 : id_attrs = RelationGetIndexAttrBitmap(relation,
3339 : INDEX_ATTR_BITMAP_IDENTITY_KEY);
3340 606556 : interesting_attrs = NULL;
3341 606556 : interesting_attrs = bms_add_members(interesting_attrs, hot_attrs);
3342 606556 : interesting_attrs = bms_add_members(interesting_attrs, sum_attrs);
3343 606556 : interesting_attrs = bms_add_members(interesting_attrs, key_attrs);
3344 606556 : interesting_attrs = bms_add_members(interesting_attrs, id_attrs);
3345 :
3346 606556 : block = ItemPointerGetBlockNumber(otid);
3347 606556 : INJECTION_POINT("heap_update-before-pin", NULL);
3348 606556 : buffer = ReadBuffer(relation, block);
3349 606556 : page = BufferGetPage(buffer);
3350 :
3351 : /*
3352 : * Before locking the buffer, pin the visibility map page if it appears to
3353 : * be necessary. Since we haven't got the lock yet, someone else might be
3354 : * in the middle of changing this, so we'll need to recheck after we have
3355 : * the lock.
3356 : */
3357 606556 : if (PageIsAllVisible(page))
3358 3054 : visibilitymap_pin(relation, block, &vmbuffer);
3359 :
3360 606556 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3361 :
3362 606556 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(otid));
3363 :
3364 : /*
3365 : * Usually, a buffer pin and/or snapshot blocks pruning of otid, ensuring
3366 : * we see LP_NORMAL here. When the otid origin is a syscache, we may have
3367 : * neither a pin nor a snapshot. Hence, we may see other LP_ states, each
3368 : * of which indicates concurrent pruning.
3369 : *
3370 : * Failing with TM_Updated would be most accurate. However, unlike other
3371 : * TM_Updated scenarios, we don't know the successor ctid in LP_UNUSED and
3372 : * LP_DEAD cases. While the distinction between TM_Updated and TM_Deleted
3373 : * does matter to SQL statements UPDATE and MERGE, those SQL statements
3374 : * hold a snapshot that ensures LP_NORMAL. Hence, the choice between
3375 : * TM_Updated and TM_Deleted affects only the wording of error messages.
3376 : * Settle on TM_Deleted, for two reasons. First, it avoids complicating
3377 : * the specification of when tmfd->ctid is valid. Second, it creates
3378 : * error log evidence that we took this branch.
3379 : *
3380 : * Since it's possible to see LP_UNUSED at otid, it's also possible to see
3381 : * LP_NORMAL for a tuple that replaced LP_UNUSED. If it's a tuple for an
3382 : * unrelated row, we'll fail with "duplicate key value violates unique".
3383 : * XXX if otid is the live, newer version of the newtup row, we'll discard
3384 : * changes originating in versions of this catalog row after the version
3385 : * the caller got from syscache. See syscache-update-pruned.spec.
3386 : */
3387 606556 : if (!ItemIdIsNormal(lp))
3388 : {
3389 : Assert(RelationSupportsSysCache(RelationGetRelid(relation)));
3390 :
3391 2 : UnlockReleaseBuffer(buffer);
3392 : Assert(!have_tuple_lock);
3393 2 : if (vmbuffer != InvalidBuffer)
3394 2 : ReleaseBuffer(vmbuffer);
3395 2 : tmfd->ctid = *otid;
3396 2 : tmfd->xmax = InvalidTransactionId;
3397 2 : tmfd->cmax = InvalidCommandId;
3398 2 : *update_indexes = TU_None;
3399 :
3400 2 : bms_free(hot_attrs);
3401 2 : bms_free(sum_attrs);
3402 2 : bms_free(key_attrs);
3403 2 : bms_free(id_attrs);
3404 : /* modified_attrs not yet initialized */
3405 2 : bms_free(interesting_attrs);
3406 2 : return TM_Deleted;
3407 : }
3408 :
3409 : /*
3410 : * Fill in enough data in oldtup for HeapDetermineColumnsInfo to work
3411 : * properly.
3412 : */
3413 606554 : oldtup.t_tableOid = RelationGetRelid(relation);
3414 606554 : oldtup.t_data = (HeapTupleHeader) PageGetItem(page, lp);
3415 606554 : oldtup.t_len = ItemIdGetLength(lp);
3416 606554 : oldtup.t_self = *otid;
3417 :
3418 : /* the new tuple is ready, except for this: */
3419 606554 : newtup->t_tableOid = RelationGetRelid(relation);
3420 :
3421 : /*
3422 : * Determine columns modified by the update. Additionally, identify
3423 : * whether any of the unmodified replica identity key attributes in the
3424 : * old tuple is externally stored or not. This is required because for
3425 : * such attributes the flattened value won't be WAL logged as part of the
3426 : * new tuple so we must include it as part of the old_key_tuple. See
3427 : * ExtractReplicaIdentity.
3428 : */
3429 606554 : modified_attrs = HeapDetermineColumnsInfo(relation, interesting_attrs,
3430 : id_attrs, &oldtup,
3431 : newtup, &id_has_external);
3432 :
3433 : /*
3434 : * If we're not updating any "key" column, we can grab a weaker lock type.
3435 : * This allows for more concurrency when we are running simultaneously
3436 : * with foreign key checks.
3437 : *
3438 : * Note that if a column gets detoasted while executing the update, but
3439 : * the value ends up being the same, this test will fail and we will use
3440 : * the stronger lock. This is acceptable; the important case to optimize
3441 : * is updates that don't manipulate key columns, not those that
3442 : * serendipitously arrive at the same key values.
3443 : */
3444 606554 : if (!bms_overlap(modified_attrs, key_attrs))
3445 : {
3446 598202 : *lockmode = LockTupleNoKeyExclusive;
3447 598202 : mxact_status = MultiXactStatusNoKeyUpdate;
3448 598202 : key_intact = true;
3449 :
3450 : /*
3451 : * If this is the first possibly-multixact-able operation in the
3452 : * current transaction, set my per-backend OldestMemberMXactId
3453 : * setting. We can be certain that the transaction will never become a
3454 : * member of any older MultiXactIds than that. (We have to do this
3455 : * even if we end up just using our own TransactionId below, since
3456 : * some other backend could incorporate our XID into a MultiXact
3457 : * immediately afterwards.)
3458 : */
3459 598202 : MultiXactIdSetOldestMember();
3460 : }
3461 : else
3462 : {
3463 8352 : *lockmode = LockTupleExclusive;
3464 8352 : mxact_status = MultiXactStatusUpdate;
3465 8352 : key_intact = false;
3466 : }
3467 :
3468 : /*
3469 : * Note: beyond this point, use oldtup not otid to refer to old tuple.
3470 : * otid may very well point at newtup->t_self, which we will overwrite
3471 : * with the new tuple's location, so there's great risk of confusion if we
3472 : * use otid anymore.
3473 : */
3474 :
3475 606554 : l2:
3476 606556 : checked_lockers = false;
3477 606556 : locker_remains = false;
3478 606556 : result = HeapTupleSatisfiesUpdate(&oldtup, cid, buffer);
3479 :
3480 : /* see below about the "no wait" case */
3481 : Assert(result != TM_BeingModified || wait);
3482 :
3483 606556 : if (result == TM_Invisible)
3484 : {
3485 0 : UnlockReleaseBuffer(buffer);
3486 0 : ereport(ERROR,
3487 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
3488 : errmsg("attempted to update invisible tuple")));
3489 : }
3490 606556 : else if (result == TM_BeingModified && wait)
3491 : {
3492 : TransactionId xwait;
3493 : uint16 infomask;
3494 71856 : bool can_continue = false;
3495 :
3496 : /*
3497 : * XXX note that we don't consider the "no wait" case here. This
3498 : * isn't a problem currently because no caller uses that case, but it
3499 : * should be fixed if such a caller is introduced. It wasn't a
3500 : * problem previously because this code would always wait, but now
3501 : * that some tuple locks do not conflict with one of the lock modes we
3502 : * use, it is possible that this case is interesting to handle
3503 : * specially.
3504 : *
3505 : * This may cause failures with third-party code that calls
3506 : * heap_update directly.
3507 : */
3508 :
3509 : /* must copy state data before unlocking buffer */
3510 71856 : xwait = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3511 71856 : infomask = oldtup.t_data->t_infomask;
3512 :
3513 : /*
3514 : * Now we have to do something about the existing locker. If it's a
3515 : * multi, sleep on it; we might be awakened before it is completely
3516 : * gone (or even not sleep at all in some cases); we need to preserve
3517 : * it as locker, unless it is gone completely.
3518 : *
3519 : * If it's not a multi, we need to check for sleeping conditions
3520 : * before actually going to sleep. If the update doesn't conflict
3521 : * with the locks, we just continue without sleeping (but making sure
3522 : * it is preserved).
3523 : *
3524 : * Before sleeping, we need to acquire tuple lock to establish our
3525 : * priority for the tuple (see heap_lock_tuple). LockTuple will
3526 : * release us when we are next-in-line for the tuple. Note we must
3527 : * not acquire the tuple lock until we're sure we're going to sleep;
3528 : * otherwise we're open for race conditions with other transactions
3529 : * holding the tuple lock which sleep on us.
3530 : *
3531 : * If we are forced to "start over" below, we keep the tuple lock;
3532 : * this arranges that we stay at the head of the line while rechecking
3533 : * tuple state.
3534 : */
3535 71856 : if (infomask & HEAP_XMAX_IS_MULTI)
3536 : {
3537 : TransactionId update_xact;
3538 : int remain;
3539 120 : bool current_is_member = false;
3540 :
3541 120 : if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
3542 : *lockmode, ¤t_is_member))
3543 : {
3544 16 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3545 :
3546 : /*
3547 : * Acquire the lock, if necessary (but skip it when we're
3548 : * requesting a lock and already have one; avoids deadlock).
3549 : */
3550 16 : if (!current_is_member)
3551 0 : heap_acquire_tuplock(relation, &(oldtup.t_self), *lockmode,
3552 : LockWaitBlock, &have_tuple_lock);
3553 :
3554 : /* wait for multixact */
3555 16 : MultiXactIdWait((MultiXactId) xwait, mxact_status, infomask,
3556 : relation, &oldtup.t_self, XLTW_Update,
3557 : &remain);
3558 16 : checked_lockers = true;
3559 16 : locker_remains = remain != 0;
3560 16 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3561 :
3562 : /*
3563 : * If xwait had just locked the tuple then some other xact
3564 : * could update this tuple before we get to this point. Check
3565 : * for xmax change, and start over if so.
3566 : */
3567 16 : if (xmax_infomask_changed(oldtup.t_data->t_infomask,
3568 16 : infomask) ||
3569 16 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3570 : xwait))
3571 0 : goto l2;
3572 : }
3573 :
3574 : /*
3575 : * Note that the multixact may not be done by now. It could have
3576 : * surviving members; our own xact or other subxacts of this
3577 : * backend, and also any other concurrent transaction that locked
3578 : * the tuple with LockTupleKeyShare if we only got
3579 : * LockTupleNoKeyExclusive. If this is the case, we have to be
3580 : * careful to mark the updated tuple with the surviving members in
3581 : * Xmax.
3582 : *
3583 : * Note that there could have been another update in the
3584 : * MultiXact. In that case, we need to check whether it committed
3585 : * or aborted. If it aborted we are safe to update it again;
3586 : * otherwise there is an update conflict, and we have to return
3587 : * TableTuple{Deleted, Updated} below.
3588 : *
3589 : * In the LockTupleExclusive case, we still need to preserve the
3590 : * surviving members: those would include the tuple locks we had
3591 : * before this one, which are important to keep in case this
3592 : * subxact aborts.
3593 : */
3594 120 : if (!HEAP_XMAX_IS_LOCKED_ONLY(oldtup.t_data->t_infomask))
3595 16 : update_xact = HeapTupleGetUpdateXid(oldtup.t_data);
3596 : else
3597 104 : update_xact = InvalidTransactionId;
3598 :
3599 : /*
3600 : * There was no UPDATE in the MultiXact; or it aborted. No
3601 : * TransactionIdIsInProgress() call needed here, since we called
3602 : * MultiXactIdWait() above.
3603 : */
3604 136 : if (!TransactionIdIsValid(update_xact) ||
3605 16 : TransactionIdDidAbort(update_xact))
3606 106 : can_continue = true;
3607 : }
3608 71736 : else if (TransactionIdIsCurrentTransactionId(xwait))
3609 : {
3610 : /*
3611 : * The only locker is ourselves; we can avoid grabbing the tuple
3612 : * lock here, but must preserve our locking information.
3613 : */
3614 71552 : checked_lockers = true;
3615 71552 : locker_remains = true;
3616 71552 : can_continue = true;
3617 : }
3618 184 : else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) && key_intact)
3619 : {
3620 : /*
3621 : * If it's just a key-share locker, and we're not changing the key
3622 : * columns, we don't need to wait for it to end; but we need to
3623 : * preserve it as locker.
3624 : */
3625 58 : checked_lockers = true;
3626 58 : locker_remains = true;
3627 58 : can_continue = true;
3628 : }
3629 : else
3630 : {
3631 : /*
3632 : * Wait for regular transaction to end; but first, acquire tuple
3633 : * lock.
3634 : */
3635 126 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3636 126 : heap_acquire_tuplock(relation, &(oldtup.t_self), *lockmode,
3637 : LockWaitBlock, &have_tuple_lock);
3638 126 : XactLockTableWait(xwait, relation, &oldtup.t_self,
3639 : XLTW_Update);
3640 126 : checked_lockers = true;
3641 126 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3642 :
3643 : /*
3644 : * xwait is done, but if xwait had just locked the tuple then some
3645 : * other xact could update this tuple before we get to this point.
3646 : * Check for xmax change, and start over if so.
3647 : */
3648 250 : if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3649 124 : !TransactionIdEquals(xwait,
3650 : HeapTupleHeaderGetRawXmax(oldtup.t_data)))
3651 2 : goto l2;
3652 :
3653 : /* Otherwise check if it committed or aborted */
3654 124 : UpdateXmaxHintBits(oldtup.t_data, buffer, xwait);
3655 124 : if (oldtup.t_data->t_infomask & HEAP_XMAX_INVALID)
3656 30 : can_continue = true;
3657 : }
3658 :
3659 71854 : if (can_continue)
3660 71746 : result = TM_Ok;
3661 108 : else if (!ItemPointerEquals(&oldtup.t_self, &oldtup.t_data->t_ctid))
3662 98 : result = TM_Updated;
3663 : else
3664 10 : result = TM_Deleted;
3665 : }
3666 :
3667 : /* Sanity check the result HeapTupleSatisfiesUpdate() and the logic above */
3668 : if (result != TM_Ok)
3669 : {
3670 : Assert(result == TM_SelfModified ||
3671 : result == TM_Updated ||
3672 : result == TM_Deleted ||
3673 : result == TM_BeingModified);
3674 : Assert(!(oldtup.t_data->t_infomask & HEAP_XMAX_INVALID));
3675 : Assert(result != TM_Updated ||
3676 : !ItemPointerEquals(&oldtup.t_self, &oldtup.t_data->t_ctid));
3677 : }
3678 :
3679 606554 : if (crosscheck != InvalidSnapshot && result == TM_Ok)
3680 : {
3681 : /* Perform additional check for transaction-snapshot mode RI updates */
3682 2 : if (!HeapTupleSatisfiesVisibility(&oldtup, crosscheck, buffer))
3683 2 : result = TM_Updated;
3684 : }
3685 :
3686 606554 : if (result != TM_Ok)
3687 : {
3688 302 : tmfd->ctid = oldtup.t_data->t_ctid;
3689 302 : tmfd->xmax = HeapTupleHeaderGetUpdateXid(oldtup.t_data);
3690 302 : if (result == TM_SelfModified)
3691 104 : tmfd->cmax = HeapTupleHeaderGetCmax(oldtup.t_data);
3692 : else
3693 198 : tmfd->cmax = InvalidCommandId;
3694 302 : UnlockReleaseBuffer(buffer);
3695 302 : if (have_tuple_lock)
3696 94 : UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3697 302 : if (vmbuffer != InvalidBuffer)
3698 0 : ReleaseBuffer(vmbuffer);
3699 302 : *update_indexes = TU_None;
3700 :
3701 302 : bms_free(hot_attrs);
3702 302 : bms_free(sum_attrs);
3703 302 : bms_free(key_attrs);
3704 302 : bms_free(id_attrs);
3705 302 : bms_free(modified_attrs);
3706 302 : bms_free(interesting_attrs);
3707 302 : return result;
3708 : }
3709 :
3710 : /*
3711 : * If we didn't pin the visibility map page and the page has become all
3712 : * visible while we were busy locking the buffer, or during some
3713 : * subsequent window during which we had it unlocked, we'll have to unlock
3714 : * and re-lock, to avoid holding the buffer lock across an I/O. That's a
3715 : * bit unfortunate, especially since we'll now have to recheck whether the
3716 : * tuple has been locked or updated under us, but hopefully it won't
3717 : * happen very often.
3718 : */
3719 606252 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3720 : {
3721 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3722 0 : visibilitymap_pin(relation, block, &vmbuffer);
3723 0 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3724 0 : goto l2;
3725 : }
3726 :
3727 : /* Fill in transaction status data */
3728 :
3729 : /*
3730 : * If the tuple we're updating is locked, we need to preserve the locking
3731 : * info in the old tuple's Xmax. Prepare a new Xmax value for this.
3732 : */
3733 606252 : compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3734 606252 : oldtup.t_data->t_infomask,
3735 606252 : oldtup.t_data->t_infomask2,
3736 : xid, *lockmode, true,
3737 : &xmax_old_tuple, &infomask_old_tuple,
3738 : &infomask2_old_tuple);
3739 :
3740 : /*
3741 : * And also prepare an Xmax value for the new copy of the tuple. If there
3742 : * was no xmax previously, or there was one but all lockers are now gone,
3743 : * then use InvalidTransactionId; otherwise, get the xmax from the old
3744 : * tuple. (In rare cases that might also be InvalidTransactionId and yet
3745 : * not have the HEAP_XMAX_INVALID bit set; that's fine.)
3746 : */
3747 677968 : if ((oldtup.t_data->t_infomask & HEAP_XMAX_INVALID) ||
3748 143432 : HEAP_LOCKED_UPGRADED(oldtup.t_data->t_infomask) ||
3749 71612 : (checked_lockers && !locker_remains))
3750 534536 : xmax_new_tuple = InvalidTransactionId;
3751 : else
3752 71716 : xmax_new_tuple = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3753 :
3754 606252 : if (!TransactionIdIsValid(xmax_new_tuple))
3755 : {
3756 534536 : infomask_new_tuple = HEAP_XMAX_INVALID;
3757 534536 : infomask2_new_tuple = 0;
3758 : }
3759 : else
3760 : {
3761 : /*
3762 : * If we found a valid Xmax for the new tuple, then the infomask bits
3763 : * to use on the new tuple depend on what was there on the old one.
3764 : * Note that since we're doing an update, the only possibility is that
3765 : * the lockers had FOR KEY SHARE lock.
3766 : */
3767 71716 : if (oldtup.t_data->t_infomask & HEAP_XMAX_IS_MULTI)
3768 : {
3769 106 : GetMultiXactIdHintBits(xmax_new_tuple, &infomask_new_tuple,
3770 : &infomask2_new_tuple);
3771 : }
3772 : else
3773 : {
3774 71610 : infomask_new_tuple = HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_LOCK_ONLY;
3775 71610 : infomask2_new_tuple = 0;
3776 : }
3777 : }
3778 :
3779 : /*
3780 : * Prepare the new tuple with the appropriate initial values of Xmin and
3781 : * Xmax, as well as initial infomask bits as computed above.
3782 : */
3783 606252 : newtup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
3784 606252 : newtup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
3785 606252 : HeapTupleHeaderSetXmin(newtup->t_data, xid);
3786 606252 : HeapTupleHeaderSetCmin(newtup->t_data, cid);
3787 606252 : newtup->t_data->t_infomask |= HEAP_UPDATED | infomask_new_tuple;
3788 606252 : newtup->t_data->t_infomask2 |= infomask2_new_tuple;
3789 606252 : HeapTupleHeaderSetXmax(newtup->t_data, xmax_new_tuple);
3790 :
3791 : /*
3792 : * Replace cid with a combo CID if necessary. Note that we already put
3793 : * the plain cid into the new tuple.
3794 : */
3795 606252 : HeapTupleHeaderAdjustCmax(oldtup.t_data, &cid, &iscombo);
3796 :
3797 : /*
3798 : * If the toaster needs to be activated, OR if the new tuple will not fit
3799 : * on the same page as the old, then we need to release the content lock
3800 : * (but not the pin!) on the old tuple's buffer while we are off doing
3801 : * TOAST and/or table-file-extension work. We must mark the old tuple to
3802 : * show that it's locked, else other processes may try to update it
3803 : * themselves.
3804 : *
3805 : * We need to invoke the toaster if there are already any out-of-line
3806 : * toasted values present, or if the new tuple is over-threshold.
3807 : */
3808 606252 : if (relation->rd_rel->relkind != RELKIND_RELATION &&
3809 0 : relation->rd_rel->relkind != RELKIND_MATVIEW)
3810 : {
3811 : /* toast table entries should never be recursively toasted */
3812 : Assert(!HeapTupleHasExternal(&oldtup));
3813 : Assert(!HeapTupleHasExternal(newtup));
3814 0 : need_toast = false;
3815 : }
3816 : else
3817 1818054 : need_toast = (HeapTupleHasExternal(&oldtup) ||
3818 1211802 : HeapTupleHasExternal(newtup) ||
3819 605502 : newtup->t_len > TOAST_TUPLE_THRESHOLD);
3820 :
3821 606252 : pagefree = PageGetHeapFreeSpace(page);
3822 :
3823 606252 : newtupsize = MAXALIGN(newtup->t_len);
3824 :
3825 606252 : if (need_toast || newtupsize > pagefree)
3826 297710 : {
3827 : TransactionId xmax_lock_old_tuple;
3828 : uint16 infomask_lock_old_tuple,
3829 : infomask2_lock_old_tuple;
3830 297710 : bool cleared_all_frozen = false;
3831 :
3832 : /*
3833 : * To prevent concurrent sessions from updating the tuple, we have to
3834 : * temporarily mark it locked, while we release the page-level lock.
3835 : *
3836 : * To satisfy the rule that any xid potentially appearing in a buffer
3837 : * written out to disk, we unfortunately have to WAL log this
3838 : * temporary modification. We can reuse xl_heap_lock for this
3839 : * purpose. If we crash/error before following through with the
3840 : * actual update, xmax will be of an aborted transaction, allowing
3841 : * other sessions to proceed.
3842 : */
3843 :
3844 : /*
3845 : * Compute xmax / infomask appropriate for locking the tuple. This has
3846 : * to be done separately from the combo that's going to be used for
3847 : * updating, because the potentially created multixact would otherwise
3848 : * be wrong.
3849 : */
3850 297710 : compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3851 297710 : oldtup.t_data->t_infomask,
3852 297710 : oldtup.t_data->t_infomask2,
3853 : xid, *lockmode, false,
3854 : &xmax_lock_old_tuple, &infomask_lock_old_tuple,
3855 : &infomask2_lock_old_tuple);
3856 :
3857 : Assert(HEAP_XMAX_IS_LOCKED_ONLY(infomask_lock_old_tuple));
3858 :
3859 297710 : START_CRIT_SECTION();
3860 :
3861 : /* Clear obsolete visibility flags ... */
3862 297710 : oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3863 297710 : oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3864 297710 : HeapTupleClearHotUpdated(&oldtup);
3865 : /* ... and store info about transaction updating this tuple */
3866 : Assert(TransactionIdIsValid(xmax_lock_old_tuple));
3867 297710 : HeapTupleHeaderSetXmax(oldtup.t_data, xmax_lock_old_tuple);
3868 297710 : oldtup.t_data->t_infomask |= infomask_lock_old_tuple;
3869 297710 : oldtup.t_data->t_infomask2 |= infomask2_lock_old_tuple;
3870 297710 : HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
3871 :
3872 : /* temporarily make it look not-updated, but locked */
3873 297710 : oldtup.t_data->t_ctid = oldtup.t_self;
3874 :
3875 : /*
3876 : * Clear all-frozen bit on visibility map if needed. We could
3877 : * immediately reset ALL_VISIBLE, but given that the WAL logging
3878 : * overhead would be unchanged, that doesn't seem necessarily
3879 : * worthwhile.
3880 : */
3881 299444 : if (PageIsAllVisible(page) &&
3882 1734 : visibilitymap_clear(relation, block, vmbuffer,
3883 : VISIBILITYMAP_ALL_FROZEN))
3884 1384 : cleared_all_frozen = true;
3885 :
3886 297710 : MarkBufferDirty(buffer);
3887 :
3888 297710 : if (RelationNeedsWAL(relation))
3889 : {
3890 : xl_heap_lock xlrec;
3891 : XLogRecPtr recptr;
3892 :
3893 277448 : XLogBeginInsert();
3894 277448 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
3895 :
3896 277448 : xlrec.offnum = ItemPointerGetOffsetNumber(&oldtup.t_self);
3897 277448 : xlrec.xmax = xmax_lock_old_tuple;
3898 554896 : xlrec.infobits_set = compute_infobits(oldtup.t_data->t_infomask,
3899 277448 : oldtup.t_data->t_infomask2);
3900 277448 : xlrec.flags =
3901 277448 : cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
3902 277448 : XLogRegisterData(&xlrec, SizeOfHeapLock);
3903 277448 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK);
3904 277448 : PageSetLSN(page, recptr);
3905 : }
3906 :
3907 297710 : END_CRIT_SECTION();
3908 :
3909 297710 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3910 :
3911 : /*
3912 : * Let the toaster do its thing, if needed.
3913 : *
3914 : * Note: below this point, heaptup is the data we actually intend to
3915 : * store into the relation; newtup is the caller's original untoasted
3916 : * data.
3917 : */
3918 297710 : if (need_toast)
3919 : {
3920 : /* Note we always use WAL and FSM during updates */
3921 3170 : heaptup = heap_toast_insert_or_update(relation, newtup, &oldtup, 0);
3922 3170 : newtupsize = MAXALIGN(heaptup->t_len);
3923 : }
3924 : else
3925 294540 : heaptup = newtup;
3926 :
3927 : /*
3928 : * Now, do we need a new page for the tuple, or not? This is a bit
3929 : * tricky since someone else could have added tuples to the page while
3930 : * we weren't looking. We have to recheck the available space after
3931 : * reacquiring the buffer lock. But don't bother to do that if the
3932 : * former amount of free space is still not enough; it's unlikely
3933 : * there's more free now than before.
3934 : *
3935 : * What's more, if we need to get a new page, we will need to acquire
3936 : * buffer locks on both old and new pages. To avoid deadlock against
3937 : * some other backend trying to get the same two locks in the other
3938 : * order, we must be consistent about the order we get the locks in.
3939 : * We use the rule "lock the lower-numbered page of the relation
3940 : * first". To implement this, we must do RelationGetBufferForTuple
3941 : * while not holding the lock on the old page, and we must rely on it
3942 : * to get the locks on both pages in the correct order.
3943 : *
3944 : * Another consideration is that we need visibility map page pin(s) if
3945 : * we will have to clear the all-visible flag on either page. If we
3946 : * call RelationGetBufferForTuple, we rely on it to acquire any such
3947 : * pins; but if we don't, we have to handle that here. Hence we need
3948 : * a loop.
3949 : */
3950 : for (;;)
3951 : {
3952 297710 : if (newtupsize > pagefree)
3953 : {
3954 : /* It doesn't fit, must use RelationGetBufferForTuple. */
3955 296648 : newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3956 : buffer, 0, NULL,
3957 : &vmbuffer_new, &vmbuffer,
3958 : 0);
3959 : /* We're all done. */
3960 296648 : break;
3961 : }
3962 : /* Acquire VM page pin if needed and we don't have it. */
3963 1062 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3964 0 : visibilitymap_pin(relation, block, &vmbuffer);
3965 : /* Re-acquire the lock on the old tuple's page. */
3966 1062 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3967 : /* Re-check using the up-to-date free space */
3968 1062 : pagefree = PageGetHeapFreeSpace(page);
3969 1062 : if (newtupsize > pagefree ||
3970 1062 : (vmbuffer == InvalidBuffer && PageIsAllVisible(page)))
3971 : {
3972 : /*
3973 : * Rats, it doesn't fit anymore, or somebody just now set the
3974 : * all-visible flag. We must now unlock and loop to avoid
3975 : * deadlock. Fortunately, this path should seldom be taken.
3976 : */
3977 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3978 : }
3979 : else
3980 : {
3981 : /* We're all done. */
3982 1062 : newbuf = buffer;
3983 1062 : break;
3984 : }
3985 : }
3986 : }
3987 : else
3988 : {
3989 : /* No TOAST work needed, and it'll fit on same page */
3990 308542 : newbuf = buffer;
3991 308542 : heaptup = newtup;
3992 : }
3993 :
3994 : /*
3995 : * We're about to do the actual update -- check for conflict first, to
3996 : * avoid possibly having to roll back work we've just done.
3997 : *
3998 : * This is safe without a recheck as long as there is no possibility of
3999 : * another process scanning the pages between this check and the update
4000 : * being visible to the scan (i.e., exclusive buffer content lock(s) are
4001 : * continuously held from this point until the tuple update is visible).
4002 : *
4003 : * For the new tuple the only check needed is at the relation level, but
4004 : * since both tuples are in the same relation and the check for oldtup
4005 : * will include checking the relation level, there is no benefit to a
4006 : * separate check for the new tuple.
4007 : */
4008 606252 : CheckForSerializableConflictIn(relation, &oldtup.t_self,
4009 : BufferGetBlockNumber(buffer));
4010 :
4011 : /*
4012 : * At this point newbuf and buffer are both pinned and locked, and newbuf
4013 : * has enough space for the new tuple. If they are the same buffer, only
4014 : * one pin is held.
4015 : */
4016 :
4017 606228 : if (newbuf == buffer)
4018 : {
4019 : /*
4020 : * Since the new tuple is going into the same page, we might be able
4021 : * to do a HOT update. Check if any of the index columns have been
4022 : * changed.
4023 : */
4024 309580 : if (!bms_overlap(modified_attrs, hot_attrs))
4025 : {
4026 285734 : use_hot_update = true;
4027 :
4028 : /*
4029 : * If none of the columns that are used in hot-blocking indexes
4030 : * were updated, we can apply HOT, but we do still need to check
4031 : * if we need to update the summarizing indexes, and update those
4032 : * indexes if the columns were updated, or we may fail to detect
4033 : * e.g. value bound changes in BRIN minmax indexes.
4034 : */
4035 285734 : if (bms_overlap(modified_attrs, sum_attrs))
4036 3282 : summarized_update = true;
4037 : }
4038 : }
4039 : else
4040 : {
4041 : /* Set a hint that the old page could use prune/defrag */
4042 296648 : PageSetFull(page);
4043 : }
4044 :
4045 : /*
4046 : * Compute replica identity tuple before entering the critical section so
4047 : * we don't PANIC upon a memory allocation failure.
4048 : * ExtractReplicaIdentity() will return NULL if nothing needs to be
4049 : * logged. Pass old key required as true only if the replica identity key
4050 : * columns are modified or it has external data.
4051 : */
4052 606228 : old_key_tuple = ExtractReplicaIdentity(relation, &oldtup,
4053 606228 : bms_overlap(modified_attrs, id_attrs) ||
4054 : id_has_external,
4055 : &old_key_copied);
4056 :
4057 : /* NO EREPORT(ERROR) from here till changes are logged */
4058 606228 : START_CRIT_SECTION();
4059 :
4060 : /*
4061 : * If this transaction commits, the old tuple will become DEAD sooner or
4062 : * later. Set flag that this page is a candidate for pruning once our xid
4063 : * falls below the OldestXmin horizon. If the transaction finally aborts,
4064 : * the subsequent page pruning will be a no-op and the hint will be
4065 : * cleared.
4066 : *
4067 : * XXX Should we set hint on newbuf as well? If the transaction aborts,
4068 : * there would be a prunable tuple in the newbuf; but for now we choose
4069 : * not to optimize for aborts. Note that heap_xlog_update must be kept in
4070 : * sync if this decision changes.
4071 : */
4072 606228 : PageSetPrunable(page, xid);
4073 :
4074 606228 : if (use_hot_update)
4075 : {
4076 : /* Mark the old tuple as HOT-updated */
4077 285734 : HeapTupleSetHotUpdated(&oldtup);
4078 : /* And mark the new tuple as heap-only */
4079 285734 : HeapTupleSetHeapOnly(heaptup);
4080 : /* Mark the caller's copy too, in case different from heaptup */
4081 285734 : HeapTupleSetHeapOnly(newtup);
4082 : }
4083 : else
4084 : {
4085 : /* Make sure tuples are correctly marked as not-HOT */
4086 320494 : HeapTupleClearHotUpdated(&oldtup);
4087 320494 : HeapTupleClearHeapOnly(heaptup);
4088 320494 : HeapTupleClearHeapOnly(newtup);
4089 : }
4090 :
4091 606228 : RelationPutHeapTuple(relation, newbuf, heaptup, false); /* insert new tuple */
4092 :
4093 :
4094 : /* Clear obsolete visibility flags, possibly set by ourselves above... */
4095 606228 : oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
4096 606228 : oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
4097 : /* ... and store info about transaction updating this tuple */
4098 : Assert(TransactionIdIsValid(xmax_old_tuple));
4099 606228 : HeapTupleHeaderSetXmax(oldtup.t_data, xmax_old_tuple);
4100 606228 : oldtup.t_data->t_infomask |= infomask_old_tuple;
4101 606228 : oldtup.t_data->t_infomask2 |= infomask2_old_tuple;
4102 606228 : HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
4103 :
4104 : /* record address of new tuple in t_ctid of old one */
4105 606228 : oldtup.t_data->t_ctid = heaptup->t_self;
4106 :
4107 : /* clear PD_ALL_VISIBLE flags, reset all visibilitymap bits */
4108 606228 : if (PageIsAllVisible(BufferGetPage(buffer)))
4109 : {
4110 3052 : all_visible_cleared = true;
4111 3052 : PageClearAllVisible(BufferGetPage(buffer));
4112 3052 : visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
4113 : vmbuffer, VISIBILITYMAP_VALID_BITS);
4114 : }
4115 606228 : if (newbuf != buffer && PageIsAllVisible(BufferGetPage(newbuf)))
4116 : {
4117 1536 : all_visible_cleared_new = true;
4118 1536 : PageClearAllVisible(BufferGetPage(newbuf));
4119 1536 : visibilitymap_clear(relation, BufferGetBlockNumber(newbuf),
4120 : vmbuffer_new, VISIBILITYMAP_VALID_BITS);
4121 : }
4122 :
4123 606228 : if (newbuf != buffer)
4124 296648 : MarkBufferDirty(newbuf);
4125 606228 : MarkBufferDirty(buffer);
4126 :
4127 : /* XLOG stuff */
4128 606228 : if (RelationNeedsWAL(relation))
4129 : {
4130 : XLogRecPtr recptr;
4131 :
4132 : /*
4133 : * For logical decoding we need combo CIDs to properly decode the
4134 : * catalog.
4135 : */
4136 583504 : if (RelationIsAccessibleInLogicalDecoding(relation))
4137 : {
4138 5086 : log_heap_new_cid(relation, &oldtup);
4139 5086 : log_heap_new_cid(relation, heaptup);
4140 : }
4141 :
4142 583504 : recptr = log_heap_update(relation, buffer,
4143 : newbuf, &oldtup, heaptup,
4144 : old_key_tuple,
4145 : all_visible_cleared,
4146 : all_visible_cleared_new);
4147 583504 : if (newbuf != buffer)
4148 : {
4149 276398 : PageSetLSN(BufferGetPage(newbuf), recptr);
4150 : }
4151 583504 : PageSetLSN(BufferGetPage(buffer), recptr);
4152 : }
4153 :
4154 606228 : END_CRIT_SECTION();
4155 :
4156 606228 : if (newbuf != buffer)
4157 296648 : LockBuffer(newbuf, BUFFER_LOCK_UNLOCK);
4158 606228 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
4159 :
4160 : /*
4161 : * Mark old tuple for invalidation from system caches at next command
4162 : * boundary, and mark the new tuple for invalidation in case we abort. We
4163 : * have to do this before releasing the buffer because oldtup is in the
4164 : * buffer. (heaptup is all in local memory, but it's necessary to process
4165 : * both tuple versions in one call to inval.c so we can avoid redundant
4166 : * sinval messages.)
4167 : */
4168 606228 : CacheInvalidateHeapTuple(relation, &oldtup, heaptup);
4169 :
4170 : /* Now we can release the buffer(s) */
4171 606228 : if (newbuf != buffer)
4172 296648 : ReleaseBuffer(newbuf);
4173 606228 : ReleaseBuffer(buffer);
4174 606228 : if (BufferIsValid(vmbuffer_new))
4175 1536 : ReleaseBuffer(vmbuffer_new);
4176 606228 : if (BufferIsValid(vmbuffer))
4177 3052 : ReleaseBuffer(vmbuffer);
4178 :
4179 : /*
4180 : * Release the lmgr tuple lock, if we had it.
4181 : */
4182 606228 : if (have_tuple_lock)
4183 30 : UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
4184 :
4185 606228 : pgstat_count_heap_update(relation, use_hot_update, newbuf != buffer);
4186 :
4187 : /*
4188 : * If heaptup is a private copy, release it. Don't forget to copy t_self
4189 : * back to the caller's image, too.
4190 : */
4191 606228 : if (heaptup != newtup)
4192 : {
4193 3074 : newtup->t_self = heaptup->t_self;
4194 3074 : heap_freetuple(heaptup);
4195 : }
4196 :
4197 : /*
4198 : * If it is a HOT update, the update may still need to update summarized
4199 : * indexes, lest we fail to update those summaries and get incorrect
4200 : * results (for example, minmax bounds of the block may change with this
4201 : * update).
4202 : */
4203 606228 : if (use_hot_update)
4204 : {
4205 285734 : if (summarized_update)
4206 3282 : *update_indexes = TU_Summarizing;
4207 : else
4208 282452 : *update_indexes = TU_None;
4209 : }
4210 : else
4211 320494 : *update_indexes = TU_All;
4212 :
4213 606228 : if (old_key_tuple != NULL && old_key_copied)
4214 168 : heap_freetuple(old_key_tuple);
4215 :
4216 606228 : bms_free(hot_attrs);
4217 606228 : bms_free(sum_attrs);
4218 606228 : bms_free(key_attrs);
4219 606228 : bms_free(id_attrs);
4220 606228 : bms_free(modified_attrs);
4221 606228 : bms_free(interesting_attrs);
4222 :
4223 606228 : return TM_Ok;
4224 : }
4225 :
4226 : #ifdef USE_ASSERT_CHECKING
4227 : /*
4228 : * Confirm adequate lock held during heap_update(), per rules from
4229 : * README.tuplock section "Locking to write inplace-updated tables".
4230 : */
4231 : static void
4232 : check_lock_if_inplace_updateable_rel(Relation relation,
4233 : ItemPointer otid,
4234 : HeapTuple newtup)
4235 : {
4236 : /* LOCKTAG_TUPLE acceptable for any catalog */
4237 : switch (RelationGetRelid(relation))
4238 : {
4239 : case RelationRelationId:
4240 : case DatabaseRelationId:
4241 : {
4242 : LOCKTAG tuptag;
4243 :
4244 : SET_LOCKTAG_TUPLE(tuptag,
4245 : relation->rd_lockInfo.lockRelId.dbId,
4246 : relation->rd_lockInfo.lockRelId.relId,
4247 : ItemPointerGetBlockNumber(otid),
4248 : ItemPointerGetOffsetNumber(otid));
4249 : if (LockHeldByMe(&tuptag, InplaceUpdateTupleLock, false))
4250 : return;
4251 : }
4252 : break;
4253 : default:
4254 : Assert(!IsInplaceUpdateRelation(relation));
4255 : return;
4256 : }
4257 :
4258 : switch (RelationGetRelid(relation))
4259 : {
4260 : case RelationRelationId:
4261 : {
4262 : /* LOCKTAG_TUPLE or LOCKTAG_RELATION ok */
4263 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(newtup);
4264 : Oid relid = classForm->oid;
4265 : Oid dbid;
4266 : LOCKTAG tag;
4267 :
4268 : if (IsSharedRelation(relid))
4269 : dbid = InvalidOid;
4270 : else
4271 : dbid = MyDatabaseId;
4272 :
4273 : if (classForm->relkind == RELKIND_INDEX)
4274 : {
4275 : Relation irel = index_open(relid, AccessShareLock);
4276 :
4277 : SET_LOCKTAG_RELATION(tag, dbid, irel->rd_index->indrelid);
4278 : index_close(irel, AccessShareLock);
4279 : }
4280 : else
4281 : SET_LOCKTAG_RELATION(tag, dbid, relid);
4282 :
4283 : if (!LockHeldByMe(&tag, ShareUpdateExclusiveLock, false) &&
4284 : !LockHeldByMe(&tag, ShareRowExclusiveLock, true))
4285 : elog(WARNING,
4286 : "missing lock for relation \"%s\" (OID %u, relkind %c) @ TID (%u,%u)",
4287 : NameStr(classForm->relname),
4288 : relid,
4289 : classForm->relkind,
4290 : ItemPointerGetBlockNumber(otid),
4291 : ItemPointerGetOffsetNumber(otid));
4292 : }
4293 : break;
4294 : case DatabaseRelationId:
4295 : {
4296 : /* LOCKTAG_TUPLE required */
4297 : Form_pg_database dbForm = (Form_pg_database) GETSTRUCT(newtup);
4298 :
4299 : elog(WARNING,
4300 : "missing lock on database \"%s\" (OID %u) @ TID (%u,%u)",
4301 : NameStr(dbForm->datname),
4302 : dbForm->oid,
4303 : ItemPointerGetBlockNumber(otid),
4304 : ItemPointerGetOffsetNumber(otid));
4305 : }
4306 : break;
4307 : }
4308 : }
4309 :
4310 : /*
4311 : * Confirm adequate relation lock held, per rules from README.tuplock section
4312 : * "Locking to write inplace-updated tables".
4313 : */
4314 : static void
4315 : check_inplace_rel_lock(HeapTuple oldtup)
4316 : {
4317 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(oldtup);
4318 : Oid relid = classForm->oid;
4319 : Oid dbid;
4320 : LOCKTAG tag;
4321 :
4322 : if (IsSharedRelation(relid))
4323 : dbid = InvalidOid;
4324 : else
4325 : dbid = MyDatabaseId;
4326 :
4327 : if (classForm->relkind == RELKIND_INDEX)
4328 : {
4329 : Relation irel = index_open(relid, AccessShareLock);
4330 :
4331 : SET_LOCKTAG_RELATION(tag, dbid, irel->rd_index->indrelid);
4332 : index_close(irel, AccessShareLock);
4333 : }
4334 : else
4335 : SET_LOCKTAG_RELATION(tag, dbid, relid);
4336 :
4337 : if (!LockHeldByMe(&tag, ShareUpdateExclusiveLock, true))
4338 : elog(WARNING,
4339 : "missing lock for relation \"%s\" (OID %u, relkind %c) @ TID (%u,%u)",
4340 : NameStr(classForm->relname),
4341 : relid,
4342 : classForm->relkind,
4343 : ItemPointerGetBlockNumber(&oldtup->t_self),
4344 : ItemPointerGetOffsetNumber(&oldtup->t_self));
4345 : }
4346 : #endif
4347 :
4348 : /*
4349 : * Check if the specified attribute's values are the same. Subroutine for
4350 : * HeapDetermineColumnsInfo.
4351 : */
4352 : static bool
4353 1464396 : heap_attr_equals(TupleDesc tupdesc, int attrnum, Datum value1, Datum value2,
4354 : bool isnull1, bool isnull2)
4355 : {
4356 : /*
4357 : * If one value is NULL and other is not, then they are certainly not
4358 : * equal
4359 : */
4360 1464396 : if (isnull1 != isnull2)
4361 90 : return false;
4362 :
4363 : /*
4364 : * If both are NULL, they can be considered equal.
4365 : */
4366 1464306 : if (isnull1)
4367 9982 : return true;
4368 :
4369 : /*
4370 : * We do simple binary comparison of the two datums. This may be overly
4371 : * strict because there can be multiple binary representations for the
4372 : * same logical value. But we should be OK as long as there are no false
4373 : * positives. Using a type-specific equality operator is messy because
4374 : * there could be multiple notions of equality in different operator
4375 : * classes; furthermore, we cannot safely invoke user-defined functions
4376 : * while holding exclusive buffer lock.
4377 : */
4378 1454324 : if (attrnum <= 0)
4379 : {
4380 : /* The only allowed system columns are OIDs, so do this */
4381 0 : return (DatumGetObjectId(value1) == DatumGetObjectId(value2));
4382 : }
4383 : else
4384 : {
4385 : CompactAttribute *att;
4386 :
4387 : Assert(attrnum <= tupdesc->natts);
4388 1454324 : att = TupleDescCompactAttr(tupdesc, attrnum - 1);
4389 1454324 : return datumIsEqual(value1, value2, att->attbyval, att->attlen);
4390 : }
4391 : }
4392 :
4393 : /*
4394 : * Check which columns are being updated.
4395 : *
4396 : * Given an updated tuple, determine (and return into the output bitmapset),
4397 : * from those listed as interesting, the set of columns that changed.
4398 : *
4399 : * has_external indicates if any of the unmodified attributes (from those
4400 : * listed as interesting) of the old tuple is a member of external_cols and is
4401 : * stored externally.
4402 : */
4403 : static Bitmapset *
4404 606554 : HeapDetermineColumnsInfo(Relation relation,
4405 : Bitmapset *interesting_cols,
4406 : Bitmapset *external_cols,
4407 : HeapTuple oldtup, HeapTuple newtup,
4408 : bool *has_external)
4409 : {
4410 : int attidx;
4411 606554 : Bitmapset *modified = NULL;
4412 606554 : TupleDesc tupdesc = RelationGetDescr(relation);
4413 :
4414 606554 : attidx = -1;
4415 2070950 : while ((attidx = bms_next_member(interesting_cols, attidx)) >= 0)
4416 : {
4417 : /* attidx is zero-based, attrnum is the normal attribute number */
4418 1464396 : AttrNumber attrnum = attidx + FirstLowInvalidHeapAttributeNumber;
4419 : Datum value1,
4420 : value2;
4421 : bool isnull1,
4422 : isnull2;
4423 :
4424 : /*
4425 : * If it's a whole-tuple reference, say "not equal". It's not really
4426 : * worth supporting this case, since it could only succeed after a
4427 : * no-op update, which is hardly a case worth optimizing for.
4428 : */
4429 1464396 : if (attrnum == 0)
4430 : {
4431 0 : modified = bms_add_member(modified, attidx);
4432 1400202 : continue;
4433 : }
4434 :
4435 : /*
4436 : * Likewise, automatically say "not equal" for any system attribute
4437 : * other than tableOID; we cannot expect these to be consistent in a
4438 : * HOT chain, or even to be set correctly yet in the new tuple.
4439 : */
4440 1464396 : if (attrnum < 0)
4441 : {
4442 0 : if (attrnum != TableOidAttributeNumber)
4443 : {
4444 0 : modified = bms_add_member(modified, attidx);
4445 0 : continue;
4446 : }
4447 : }
4448 :
4449 : /*
4450 : * Extract the corresponding values. XXX this is pretty inefficient
4451 : * if there are many indexed columns. Should we do a single
4452 : * heap_deform_tuple call on each tuple, instead? But that doesn't
4453 : * work for system columns ...
4454 : */
4455 1464396 : value1 = heap_getattr(oldtup, attrnum, tupdesc, &isnull1);
4456 1464396 : value2 = heap_getattr(newtup, attrnum, tupdesc, &isnull2);
4457 :
4458 1464396 : if (!heap_attr_equals(tupdesc, attrnum, value1,
4459 : value2, isnull1, isnull2))
4460 : {
4461 53628 : modified = bms_add_member(modified, attidx);
4462 53628 : continue;
4463 : }
4464 :
4465 : /*
4466 : * No need to check attributes that can't be stored externally. Note
4467 : * that system attributes can't be stored externally.
4468 : */
4469 1410768 : if (attrnum < 0 || isnull1 ||
4470 1400786 : TupleDescCompactAttr(tupdesc, attrnum - 1)->attlen != -1)
4471 1346574 : continue;
4472 :
4473 : /*
4474 : * Check if the old tuple's attribute is stored externally and is a
4475 : * member of external_cols.
4476 : */
4477 64204 : if (VARATT_IS_EXTERNAL((struct varlena *) DatumGetPointer(value1)) &&
4478 10 : bms_is_member(attidx, external_cols))
4479 4 : *has_external = true;
4480 : }
4481 :
4482 606554 : return modified;
4483 : }
4484 :
4485 : /*
4486 : * simple_heap_update - replace a tuple
4487 : *
4488 : * This routine may be used to update a tuple when concurrent updates of
4489 : * the target tuple are not expected (for example, because we have a lock
4490 : * on the relation associated with the tuple). Any failure is reported
4491 : * via ereport().
4492 : */
4493 : void
4494 219592 : simple_heap_update(Relation relation, ItemPointer otid, HeapTuple tup,
4495 : TU_UpdateIndexes *update_indexes)
4496 : {
4497 : TM_Result result;
4498 : TM_FailureData tmfd;
4499 : LockTupleMode lockmode;
4500 :
4501 219592 : result = heap_update(relation, otid, tup,
4502 : GetCurrentCommandId(true), InvalidSnapshot,
4503 : true /* wait for commit */ ,
4504 : &tmfd, &lockmode, update_indexes);
4505 219592 : switch (result)
4506 : {
4507 0 : case TM_SelfModified:
4508 : /* Tuple was already updated in current command? */
4509 0 : elog(ERROR, "tuple already updated by self");
4510 : break;
4511 :
4512 219590 : case TM_Ok:
4513 : /* done successfully */
4514 219590 : break;
4515 :
4516 0 : case TM_Updated:
4517 0 : elog(ERROR, "tuple concurrently updated");
4518 : break;
4519 :
4520 2 : case TM_Deleted:
4521 2 : elog(ERROR, "tuple concurrently deleted");
4522 : break;
4523 :
4524 0 : default:
4525 0 : elog(ERROR, "unrecognized heap_update status: %u", result);
4526 : break;
4527 : }
4528 219590 : }
4529 :
4530 :
4531 : /*
4532 : * Return the MultiXactStatus corresponding to the given tuple lock mode.
4533 : */
4534 : static MultiXactStatus
4535 2432 : get_mxact_status_for_lock(LockTupleMode mode, bool is_update)
4536 : {
4537 : int retval;
4538 :
4539 2432 : if (is_update)
4540 192 : retval = tupleLockExtraInfo[mode].updstatus;
4541 : else
4542 2240 : retval = tupleLockExtraInfo[mode].lockstatus;
4543 :
4544 2432 : if (retval == -1)
4545 0 : elog(ERROR, "invalid lock tuple mode %d/%s", mode,
4546 : is_update ? "true" : "false");
4547 :
4548 2432 : return (MultiXactStatus) retval;
4549 : }
4550 :
4551 : /*
4552 : * heap_lock_tuple - lock a tuple in shared or exclusive mode
4553 : *
4554 : * Note that this acquires a buffer pin, which the caller must release.
4555 : *
4556 : * Input parameters:
4557 : * relation: relation containing tuple (caller must hold suitable lock)
4558 : * tid: TID of tuple to lock
4559 : * cid: current command ID (used for visibility test, and stored into
4560 : * tuple's cmax if lock is successful)
4561 : * mode: indicates if shared or exclusive tuple lock is desired
4562 : * wait_policy: what to do if tuple lock is not available
4563 : * follow_updates: if true, follow the update chain to also lock descendant
4564 : * tuples.
4565 : *
4566 : * Output parameters:
4567 : * *tuple: all fields filled in
4568 : * *buffer: set to buffer holding tuple (pinned but not locked at exit)
4569 : * *tmfd: filled in failure cases (see below)
4570 : *
4571 : * Function results are the same as the ones for table_tuple_lock().
4572 : *
4573 : * In the failure cases other than TM_Invisible, the routine fills
4574 : * *tmfd with the tuple's t_ctid, t_xmax (resolving a possible MultiXact,
4575 : * if necessary), and t_cmax (the last only for TM_SelfModified,
4576 : * since we cannot obtain cmax from a combo CID generated by another
4577 : * transaction).
4578 : * See comments for struct TM_FailureData for additional info.
4579 : *
4580 : * See README.tuplock for a thorough explanation of this mechanism.
4581 : */
4582 : TM_Result
4583 169794 : heap_lock_tuple(Relation relation, HeapTuple tuple,
4584 : CommandId cid, LockTupleMode mode, LockWaitPolicy wait_policy,
4585 : bool follow_updates,
4586 : Buffer *buffer, TM_FailureData *tmfd)
4587 : {
4588 : TM_Result result;
4589 169794 : ItemPointer tid = &(tuple->t_self);
4590 : ItemId lp;
4591 : Page page;
4592 169794 : Buffer vmbuffer = InvalidBuffer;
4593 : BlockNumber block;
4594 : TransactionId xid,
4595 : xmax;
4596 : uint16 old_infomask,
4597 : new_infomask,
4598 : new_infomask2;
4599 169794 : bool first_time = true;
4600 169794 : bool skip_tuple_lock = false;
4601 169794 : bool have_tuple_lock = false;
4602 169794 : bool cleared_all_frozen = false;
4603 :
4604 169794 : *buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
4605 169794 : block = ItemPointerGetBlockNumber(tid);
4606 :
4607 : /*
4608 : * Before locking the buffer, pin the visibility map page if it appears to
4609 : * be necessary. Since we haven't got the lock yet, someone else might be
4610 : * in the middle of changing this, so we'll need to recheck after we have
4611 : * the lock.
4612 : */
4613 169794 : if (PageIsAllVisible(BufferGetPage(*buffer)))
4614 3324 : visibilitymap_pin(relation, block, &vmbuffer);
4615 :
4616 169794 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4617 :
4618 169794 : page = BufferGetPage(*buffer);
4619 169794 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
4620 : Assert(ItemIdIsNormal(lp));
4621 :
4622 169794 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
4623 169794 : tuple->t_len = ItemIdGetLength(lp);
4624 169794 : tuple->t_tableOid = RelationGetRelid(relation);
4625 :
4626 28 : l3:
4627 169822 : result = HeapTupleSatisfiesUpdate(tuple, cid, *buffer);
4628 :
4629 169822 : if (result == TM_Invisible)
4630 : {
4631 : /*
4632 : * This is possible, but only when locking a tuple for ON CONFLICT
4633 : * UPDATE. We return this value here rather than throwing an error in
4634 : * order to give that case the opportunity to throw a more specific
4635 : * error.
4636 : */
4637 24 : result = TM_Invisible;
4638 24 : goto out_locked;
4639 : }
4640 169798 : else if (result == TM_BeingModified ||
4641 154058 : result == TM_Updated ||
4642 : result == TM_Deleted)
4643 : {
4644 : TransactionId xwait;
4645 : uint16 infomask;
4646 : uint16 infomask2;
4647 : bool require_sleep;
4648 : ItemPointerData t_ctid;
4649 :
4650 : /* must copy state data before unlocking buffer */
4651 15742 : xwait = HeapTupleHeaderGetRawXmax(tuple->t_data);
4652 15742 : infomask = tuple->t_data->t_infomask;
4653 15742 : infomask2 = tuple->t_data->t_infomask2;
4654 15742 : ItemPointerCopy(&tuple->t_data->t_ctid, &t_ctid);
4655 :
4656 15742 : LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4657 :
4658 : /*
4659 : * If any subtransaction of the current top transaction already holds
4660 : * a lock as strong as or stronger than what we're requesting, we
4661 : * effectively hold the desired lock already. We *must* succeed
4662 : * without trying to take the tuple lock, else we will deadlock
4663 : * against anyone wanting to acquire a stronger lock.
4664 : *
4665 : * Note we only do this the first time we loop on the HTSU result;
4666 : * there is no point in testing in subsequent passes, because
4667 : * evidently our own transaction cannot have acquired a new lock after
4668 : * the first time we checked.
4669 : */
4670 15742 : if (first_time)
4671 : {
4672 15724 : first_time = false;
4673 :
4674 15724 : if (infomask & HEAP_XMAX_IS_MULTI)
4675 : {
4676 : int i;
4677 : int nmembers;
4678 : MultiXactMember *members;
4679 :
4680 : /*
4681 : * We don't need to allow old multixacts here; if that had
4682 : * been the case, HeapTupleSatisfiesUpdate would have returned
4683 : * MayBeUpdated and we wouldn't be here.
4684 : */
4685 : nmembers =
4686 182 : GetMultiXactIdMembers(xwait, &members, false,
4687 182 : HEAP_XMAX_IS_LOCKED_ONLY(infomask));
4688 :
4689 554 : for (i = 0; i < nmembers; i++)
4690 : {
4691 : /* only consider members of our own transaction */
4692 400 : if (!TransactionIdIsCurrentTransactionId(members[i].xid))
4693 302 : continue;
4694 :
4695 98 : if (TUPLOCK_from_mxstatus(members[i].status) >= mode)
4696 : {
4697 28 : pfree(members);
4698 28 : result = TM_Ok;
4699 28 : goto out_unlocked;
4700 : }
4701 : else
4702 : {
4703 : /*
4704 : * Disable acquisition of the heavyweight tuple lock.
4705 : * Otherwise, when promoting a weaker lock, we might
4706 : * deadlock with another locker that has acquired the
4707 : * heavyweight tuple lock and is waiting for our
4708 : * transaction to finish.
4709 : *
4710 : * Note that in this case we still need to wait for
4711 : * the multixact if required, to avoid acquiring
4712 : * conflicting locks.
4713 : */
4714 70 : skip_tuple_lock = true;
4715 : }
4716 : }
4717 :
4718 154 : if (members)
4719 154 : pfree(members);
4720 : }
4721 15542 : else if (TransactionIdIsCurrentTransactionId(xwait))
4722 : {
4723 13078 : switch (mode)
4724 : {
4725 324 : case LockTupleKeyShare:
4726 : Assert(HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) ||
4727 : HEAP_XMAX_IS_SHR_LOCKED(infomask) ||
4728 : HEAP_XMAX_IS_EXCL_LOCKED(infomask));
4729 324 : result = TM_Ok;
4730 324 : goto out_unlocked;
4731 30 : case LockTupleShare:
4732 42 : if (HEAP_XMAX_IS_SHR_LOCKED(infomask) ||
4733 12 : HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4734 : {
4735 18 : result = TM_Ok;
4736 18 : goto out_unlocked;
4737 : }
4738 12 : break;
4739 122 : case LockTupleNoKeyExclusive:
4740 122 : if (HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4741 : {
4742 100 : result = TM_Ok;
4743 100 : goto out_unlocked;
4744 : }
4745 22 : break;
4746 12602 : case LockTupleExclusive:
4747 12602 : if (HEAP_XMAX_IS_EXCL_LOCKED(infomask) &&
4748 2522 : infomask2 & HEAP_KEYS_UPDATED)
4749 : {
4750 2480 : result = TM_Ok;
4751 2480 : goto out_unlocked;
4752 : }
4753 10122 : break;
4754 : }
4755 : }
4756 : }
4757 :
4758 : /*
4759 : * Initially assume that we will have to wait for the locking
4760 : * transaction(s) to finish. We check various cases below in which
4761 : * this can be turned off.
4762 : */
4763 12792 : require_sleep = true;
4764 12792 : if (mode == LockTupleKeyShare)
4765 : {
4766 : /*
4767 : * If we're requesting KeyShare, and there's no update present, we
4768 : * don't need to wait. Even if there is an update, we can still
4769 : * continue if the key hasn't been modified.
4770 : *
4771 : * However, if there are updates, we need to walk the update chain
4772 : * to mark future versions of the row as locked, too. That way,
4773 : * if somebody deletes that future version, we're protected
4774 : * against the key going away. This locking of future versions
4775 : * could block momentarily, if a concurrent transaction is
4776 : * deleting a key; or it could return a value to the effect that
4777 : * the transaction deleting the key has already committed. So we
4778 : * do this before re-locking the buffer; otherwise this would be
4779 : * prone to deadlocks.
4780 : *
4781 : * Note that the TID we're locking was grabbed before we unlocked
4782 : * the buffer. For it to change while we're not looking, the
4783 : * other properties we're testing for below after re-locking the
4784 : * buffer would also change, in which case we would restart this
4785 : * loop above.
4786 : */
4787 1190 : if (!(infomask2 & HEAP_KEYS_UPDATED))
4788 : {
4789 : bool updated;
4790 :
4791 1128 : updated = !HEAP_XMAX_IS_LOCKED_ONLY(infomask);
4792 :
4793 : /*
4794 : * If there are updates, follow the update chain; bail out if
4795 : * that cannot be done.
4796 : */
4797 1128 : if (follow_updates && updated)
4798 : {
4799 : TM_Result res;
4800 :
4801 100 : res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4802 : GetCurrentTransactionId(),
4803 : mode);
4804 100 : if (res != TM_Ok)
4805 : {
4806 12 : result = res;
4807 : /* recovery code expects to have buffer lock held */
4808 12 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4809 358 : goto failed;
4810 : }
4811 : }
4812 :
4813 1116 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4814 :
4815 : /*
4816 : * Make sure it's still an appropriate lock, else start over.
4817 : * Also, if it wasn't updated before we released the lock, but
4818 : * is updated now, we start over too; the reason is that we
4819 : * now need to follow the update chain to lock the new
4820 : * versions.
4821 : */
4822 1116 : if (!HeapTupleHeaderIsOnlyLocked(tuple->t_data) &&
4823 86 : ((tuple->t_data->t_infomask2 & HEAP_KEYS_UPDATED) ||
4824 86 : !updated))
4825 28 : goto l3;
4826 :
4827 : /* Things look okay, so we can skip sleeping */
4828 1116 : require_sleep = false;
4829 :
4830 : /*
4831 : * Note we allow Xmax to change here; other updaters/lockers
4832 : * could have modified it before we grabbed the buffer lock.
4833 : * However, this is not a problem, because with the recheck we
4834 : * just did we ensure that they still don't conflict with the
4835 : * lock we want.
4836 : */
4837 : }
4838 : }
4839 11602 : else if (mode == LockTupleShare)
4840 : {
4841 : /*
4842 : * If we're requesting Share, we can similarly avoid sleeping if
4843 : * there's no update and no exclusive lock present.
4844 : */
4845 884 : if (HEAP_XMAX_IS_LOCKED_ONLY(infomask) &&
4846 884 : !HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4847 : {
4848 872 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4849 :
4850 : /*
4851 : * Make sure it's still an appropriate lock, else start over.
4852 : * See above about allowing xmax to change.
4853 : */
4854 1744 : if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4855 872 : HEAP_XMAX_IS_EXCL_LOCKED(tuple->t_data->t_infomask))
4856 0 : goto l3;
4857 872 : require_sleep = false;
4858 : }
4859 : }
4860 10718 : else if (mode == LockTupleNoKeyExclusive)
4861 : {
4862 : /*
4863 : * If we're requesting NoKeyExclusive, we might also be able to
4864 : * avoid sleeping; just ensure that there no conflicting lock
4865 : * already acquired.
4866 : */
4867 308 : if (infomask & HEAP_XMAX_IS_MULTI)
4868 : {
4869 52 : if (!DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
4870 : mode, NULL))
4871 : {
4872 : /*
4873 : * No conflict, but if the xmax changed under us in the
4874 : * meantime, start over.
4875 : */
4876 26 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4877 52 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4878 26 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4879 : xwait))
4880 0 : goto l3;
4881 :
4882 : /* otherwise, we're good */
4883 26 : require_sleep = false;
4884 : }
4885 : }
4886 256 : else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask))
4887 : {
4888 34 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4889 :
4890 : /* if the xmax changed in the meantime, start over */
4891 68 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4892 34 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4893 : xwait))
4894 0 : goto l3;
4895 : /* otherwise, we're good */
4896 34 : require_sleep = false;
4897 : }
4898 : }
4899 :
4900 : /*
4901 : * As a check independent from those above, we can also avoid sleeping
4902 : * if the current transaction is the sole locker of the tuple. Note
4903 : * that the strength of the lock already held is irrelevant; this is
4904 : * not about recording the lock in Xmax (which will be done regardless
4905 : * of this optimization, below). Also, note that the cases where we
4906 : * hold a lock stronger than we are requesting are already handled
4907 : * above by not doing anything.
4908 : *
4909 : * Note we only deal with the non-multixact case here; MultiXactIdWait
4910 : * is well equipped to deal with this situation on its own.
4911 : */
4912 23430 : if (require_sleep && !(infomask & HEAP_XMAX_IS_MULTI) &&
4913 10650 : TransactionIdIsCurrentTransactionId(xwait))
4914 : {
4915 : /* ... but if the xmax changed in the meantime, start over */
4916 10122 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4917 20244 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4918 10122 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4919 : xwait))
4920 0 : goto l3;
4921 : Assert(HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask));
4922 10122 : require_sleep = false;
4923 : }
4924 :
4925 : /*
4926 : * Time to sleep on the other transaction/multixact, if necessary.
4927 : *
4928 : * If the other transaction is an update/delete that's already
4929 : * committed, then sleeping cannot possibly do any good: if we're
4930 : * required to sleep, get out to raise an error instead.
4931 : *
4932 : * By here, we either have already acquired the buffer exclusive lock,
4933 : * or we must wait for the locking transaction or multixact; so below
4934 : * we ensure that we grab buffer lock after the sleep.
4935 : */
4936 12780 : if (require_sleep && (result == TM_Updated || result == TM_Deleted))
4937 : {
4938 270 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4939 270 : goto failed;
4940 : }
4941 12510 : else if (require_sleep)
4942 : {
4943 : /*
4944 : * Acquire tuple lock to establish our priority for the tuple, or
4945 : * die trying. LockTuple will release us when we are next-in-line
4946 : * for the tuple. We must do this even if we are share-locking,
4947 : * but not if we already have a weaker lock on the tuple.
4948 : *
4949 : * If we are forced to "start over" below, we keep the tuple lock;
4950 : * this arranges that we stay at the head of the line while
4951 : * rechecking tuple state.
4952 : */
4953 340 : if (!skip_tuple_lock &&
4954 308 : !heap_acquire_tuplock(relation, tid, mode, wait_policy,
4955 : &have_tuple_lock))
4956 : {
4957 : /*
4958 : * This can only happen if wait_policy is Skip and the lock
4959 : * couldn't be obtained.
4960 : */
4961 2 : result = TM_WouldBlock;
4962 : /* recovery code expects to have buffer lock held */
4963 2 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4964 2 : goto failed;
4965 : }
4966 :
4967 336 : if (infomask & HEAP_XMAX_IS_MULTI)
4968 : {
4969 80 : MultiXactStatus status = get_mxact_status_for_lock(mode, false);
4970 :
4971 : /* We only ever lock tuples, never update them */
4972 80 : if (status >= MultiXactStatusNoKeyUpdate)
4973 0 : elog(ERROR, "invalid lock mode in heap_lock_tuple");
4974 :
4975 : /* wait for multixact to end, or die trying */
4976 80 : switch (wait_policy)
4977 : {
4978 72 : case LockWaitBlock:
4979 72 : MultiXactIdWait((MultiXactId) xwait, status, infomask,
4980 : relation, &tuple->t_self, XLTW_Lock, NULL);
4981 72 : break;
4982 4 : case LockWaitSkip:
4983 4 : if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4984 : status, infomask, relation,
4985 : NULL, false))
4986 : {
4987 4 : result = TM_WouldBlock;
4988 : /* recovery code expects to have buffer lock held */
4989 4 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4990 4 : goto failed;
4991 : }
4992 0 : break;
4993 4 : case LockWaitError:
4994 4 : if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4995 : status, infomask, relation,
4996 : NULL, log_lock_failures))
4997 4 : ereport(ERROR,
4998 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4999 : errmsg("could not obtain lock on row in relation \"%s\"",
5000 : RelationGetRelationName(relation))));
5001 :
5002 0 : break;
5003 : }
5004 :
5005 : /*
5006 : * Of course, the multixact might not be done here: if we're
5007 : * requesting a light lock mode, other transactions with light
5008 : * locks could still be alive, as well as locks owned by our
5009 : * own xact or other subxacts of this backend. We need to
5010 : * preserve the surviving MultiXact members. Note that it
5011 : * isn't absolutely necessary in the latter case, but doing so
5012 : * is simpler.
5013 : */
5014 : }
5015 : else
5016 : {
5017 : /* wait for regular transaction to end, or die trying */
5018 256 : switch (wait_policy)
5019 : {
5020 178 : case LockWaitBlock:
5021 178 : XactLockTableWait(xwait, relation, &tuple->t_self,
5022 : XLTW_Lock);
5023 178 : break;
5024 66 : case LockWaitSkip:
5025 66 : if (!ConditionalXactLockTableWait(xwait, false))
5026 : {
5027 66 : result = TM_WouldBlock;
5028 : /* recovery code expects to have buffer lock held */
5029 66 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5030 66 : goto failed;
5031 : }
5032 0 : break;
5033 12 : case LockWaitError:
5034 12 : if (!ConditionalXactLockTableWait(xwait, log_lock_failures))
5035 12 : ereport(ERROR,
5036 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
5037 : errmsg("could not obtain lock on row in relation \"%s\"",
5038 : RelationGetRelationName(relation))));
5039 0 : break;
5040 : }
5041 : }
5042 :
5043 : /* if there are updates, follow the update chain */
5044 250 : if (follow_updates && !HEAP_XMAX_IS_LOCKED_ONLY(infomask))
5045 : {
5046 : TM_Result res;
5047 :
5048 80 : res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
5049 : GetCurrentTransactionId(),
5050 : mode);
5051 80 : if (res != TM_Ok)
5052 : {
5053 4 : result = res;
5054 : /* recovery code expects to have buffer lock held */
5055 4 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5056 4 : goto failed;
5057 : }
5058 : }
5059 :
5060 246 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5061 :
5062 : /*
5063 : * xwait is done, but if xwait had just locked the tuple then some
5064 : * other xact could update this tuple before we get to this point.
5065 : * Check for xmax change, and start over if so.
5066 : */
5067 468 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
5068 222 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
5069 : xwait))
5070 28 : goto l3;
5071 :
5072 218 : if (!(infomask & HEAP_XMAX_IS_MULTI))
5073 : {
5074 : /*
5075 : * Otherwise check if it committed or aborted. Note we cannot
5076 : * be here if the tuple was only locked by somebody who didn't
5077 : * conflict with us; that would have been handled above. So
5078 : * that transaction must necessarily be gone by now. But
5079 : * don't check for this in the multixact case, because some
5080 : * locker transactions might still be running.
5081 : */
5082 156 : UpdateXmaxHintBits(tuple->t_data, *buffer, xwait);
5083 : }
5084 : }
5085 :
5086 : /* By here, we're certain that we hold buffer exclusive lock again */
5087 :
5088 : /*
5089 : * We may lock if previous xmax aborted, or if it committed but only
5090 : * locked the tuple without updating it; or if we didn't have to wait
5091 : * at all for whatever reason.
5092 : */
5093 12388 : if (!require_sleep ||
5094 376 : (tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
5095 286 : HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
5096 128 : HeapTupleHeaderIsOnlyLocked(tuple->t_data))
5097 12272 : result = TM_Ok;
5098 116 : else if (!ItemPointerEquals(&tuple->t_self, &tuple->t_data->t_ctid))
5099 92 : result = TM_Updated;
5100 : else
5101 24 : result = TM_Deleted;
5102 : }
5103 :
5104 154056 : failed:
5105 166802 : if (result != TM_Ok)
5106 : {
5107 : Assert(result == TM_SelfModified || result == TM_Updated ||
5108 : result == TM_Deleted || result == TM_WouldBlock);
5109 :
5110 : /*
5111 : * When locking a tuple under LockWaitSkip semantics and we fail with
5112 : * TM_WouldBlock above, it's possible for concurrent transactions to
5113 : * release the lock and set HEAP_XMAX_INVALID in the meantime. So
5114 : * this assert is slightly different from the equivalent one in
5115 : * heap_delete and heap_update.
5116 : */
5117 : Assert((result == TM_WouldBlock) ||
5118 : !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));
5119 : Assert(result != TM_Updated ||
5120 : !ItemPointerEquals(&tuple->t_self, &tuple->t_data->t_ctid));
5121 486 : tmfd->ctid = tuple->t_data->t_ctid;
5122 486 : tmfd->xmax = HeapTupleHeaderGetUpdateXid(tuple->t_data);
5123 486 : if (result == TM_SelfModified)
5124 12 : tmfd->cmax = HeapTupleHeaderGetCmax(tuple->t_data);
5125 : else
5126 474 : tmfd->cmax = InvalidCommandId;
5127 486 : goto out_locked;
5128 : }
5129 :
5130 : /*
5131 : * If we didn't pin the visibility map page and the page has become all
5132 : * visible while we were busy locking the buffer, or during some
5133 : * subsequent window during which we had it unlocked, we'll have to unlock
5134 : * and re-lock, to avoid holding the buffer lock across I/O. That's a bit
5135 : * unfortunate, especially since we'll now have to recheck whether the
5136 : * tuple has been locked or updated under us, but hopefully it won't
5137 : * happen very often.
5138 : */
5139 166316 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
5140 : {
5141 0 : LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
5142 0 : visibilitymap_pin(relation, block, &vmbuffer);
5143 0 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5144 0 : goto l3;
5145 : }
5146 :
5147 166316 : xmax = HeapTupleHeaderGetRawXmax(tuple->t_data);
5148 166316 : old_infomask = tuple->t_data->t_infomask;
5149 :
5150 : /*
5151 : * If this is the first possibly-multixact-able operation in the current
5152 : * transaction, set my per-backend OldestMemberMXactId setting. We can be
5153 : * certain that the transaction will never become a member of any older
5154 : * MultiXactIds than that. (We have to do this even if we end up just
5155 : * using our own TransactionId below, since some other backend could
5156 : * incorporate our XID into a MultiXact immediately afterwards.)
5157 : */
5158 166316 : MultiXactIdSetOldestMember();
5159 :
5160 : /*
5161 : * Compute the new xmax and infomask to store into the tuple. Note we do
5162 : * not modify the tuple just yet, because that would leave it in the wrong
5163 : * state if multixact.c elogs.
5164 : */
5165 166316 : compute_new_xmax_infomask(xmax, old_infomask, tuple->t_data->t_infomask2,
5166 : GetCurrentTransactionId(), mode, false,
5167 : &xid, &new_infomask, &new_infomask2);
5168 :
5169 166316 : START_CRIT_SECTION();
5170 :
5171 : /*
5172 : * Store transaction information of xact locking the tuple.
5173 : *
5174 : * Note: Cmax is meaningless in this context, so don't set it; this avoids
5175 : * possibly generating a useless combo CID. Moreover, if we're locking a
5176 : * previously updated tuple, it's important to preserve the Cmax.
5177 : *
5178 : * Also reset the HOT UPDATE bit, but only if there's no update; otherwise
5179 : * we would break the HOT chain.
5180 : */
5181 166316 : tuple->t_data->t_infomask &= ~HEAP_XMAX_BITS;
5182 166316 : tuple->t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5183 166316 : tuple->t_data->t_infomask |= new_infomask;
5184 166316 : tuple->t_data->t_infomask2 |= new_infomask2;
5185 166316 : if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
5186 166238 : HeapTupleHeaderClearHotUpdated(tuple->t_data);
5187 166316 : HeapTupleHeaderSetXmax(tuple->t_data, xid);
5188 :
5189 : /*
5190 : * Make sure there is no forward chain link in t_ctid. Note that in the
5191 : * cases where the tuple has been updated, we must not overwrite t_ctid,
5192 : * because it was set by the updater. Moreover, if the tuple has been
5193 : * updated, we need to follow the update chain to lock the new versions of
5194 : * the tuple as well.
5195 : */
5196 166316 : if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
5197 166238 : tuple->t_data->t_ctid = *tid;
5198 :
5199 : /* Clear only the all-frozen bit on visibility map if needed */
5200 169640 : if (PageIsAllVisible(page) &&
5201 3324 : visibilitymap_clear(relation, block, vmbuffer,
5202 : VISIBILITYMAP_ALL_FROZEN))
5203 28 : cleared_all_frozen = true;
5204 :
5205 :
5206 166316 : MarkBufferDirty(*buffer);
5207 :
5208 : /*
5209 : * XLOG stuff. You might think that we don't need an XLOG record because
5210 : * there is no state change worth restoring after a crash. You would be
5211 : * wrong however: we have just written either a TransactionId or a
5212 : * MultiXactId that may never have been seen on disk before, and we need
5213 : * to make sure that there are XLOG entries covering those ID numbers.
5214 : * Else the same IDs might be re-used after a crash, which would be
5215 : * disastrous if this page made it to disk before the crash. Essentially
5216 : * we have to enforce the WAL log-before-data rule even in this case.
5217 : * (Also, in a PITR log-shipping or 2PC environment, we have to have XLOG
5218 : * entries for everything anyway.)
5219 : */
5220 166316 : if (RelationNeedsWAL(relation))
5221 : {
5222 : xl_heap_lock xlrec;
5223 : XLogRecPtr recptr;
5224 :
5225 165628 : XLogBeginInsert();
5226 165628 : XLogRegisterBuffer(0, *buffer, REGBUF_STANDARD);
5227 :
5228 165628 : xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
5229 165628 : xlrec.xmax = xid;
5230 331256 : xlrec.infobits_set = compute_infobits(new_infomask,
5231 165628 : tuple->t_data->t_infomask2);
5232 165628 : xlrec.flags = cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
5233 165628 : XLogRegisterData(&xlrec, SizeOfHeapLock);
5234 :
5235 : /* we don't decode row locks atm, so no need to log the origin */
5236 :
5237 165628 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK);
5238 :
5239 165628 : PageSetLSN(page, recptr);
5240 : }
5241 :
5242 166316 : END_CRIT_SECTION();
5243 :
5244 166316 : result = TM_Ok;
5245 :
5246 166826 : out_locked:
5247 166826 : LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
5248 :
5249 169776 : out_unlocked:
5250 169776 : if (BufferIsValid(vmbuffer))
5251 3324 : ReleaseBuffer(vmbuffer);
5252 :
5253 : /*
5254 : * Don't update the visibility map here. Locking a tuple doesn't change
5255 : * visibility info.
5256 : */
5257 :
5258 : /*
5259 : * Now that we have successfully marked the tuple as locked, we can
5260 : * release the lmgr tuple lock, if we had it.
5261 : */
5262 169776 : if (have_tuple_lock)
5263 278 : UnlockTupleTuplock(relation, tid, mode);
5264 :
5265 169776 : return result;
5266 : }
5267 :
5268 : /*
5269 : * Acquire heavyweight lock on the given tuple, in preparation for acquiring
5270 : * its normal, Xmax-based tuple lock.
5271 : *
5272 : * have_tuple_lock is an input and output parameter: on input, it indicates
5273 : * whether the lock has previously been acquired (and this function does
5274 : * nothing in that case). If this function returns success, have_tuple_lock
5275 : * has been flipped to true.
5276 : *
5277 : * Returns false if it was unable to obtain the lock; this can only happen if
5278 : * wait_policy is Skip.
5279 : */
5280 : static bool
5281 526 : heap_acquire_tuplock(Relation relation, ItemPointer tid, LockTupleMode mode,
5282 : LockWaitPolicy wait_policy, bool *have_tuple_lock)
5283 : {
5284 526 : if (*have_tuple_lock)
5285 18 : return true;
5286 :
5287 508 : switch (wait_policy)
5288 : {
5289 426 : case LockWaitBlock:
5290 426 : LockTupleTuplock(relation, tid, mode);
5291 426 : break;
5292 :
5293 68 : case LockWaitSkip:
5294 68 : if (!ConditionalLockTupleTuplock(relation, tid, mode, false))
5295 2 : return false;
5296 66 : break;
5297 :
5298 14 : case LockWaitError:
5299 14 : if (!ConditionalLockTupleTuplock(relation, tid, mode, log_lock_failures))
5300 2 : ereport(ERROR,
5301 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
5302 : errmsg("could not obtain lock on row in relation \"%s\"",
5303 : RelationGetRelationName(relation))));
5304 12 : break;
5305 : }
5306 504 : *have_tuple_lock = true;
5307 :
5308 504 : return true;
5309 : }
5310 :
5311 : /*
5312 : * Given an original set of Xmax and infomask, and a transaction (identified by
5313 : * add_to_xmax) acquiring a new lock of some mode, compute the new Xmax and
5314 : * corresponding infomasks to use on the tuple.
5315 : *
5316 : * Note that this might have side effects such as creating a new MultiXactId.
5317 : *
5318 : * Most callers will have called HeapTupleSatisfiesUpdate before this function;
5319 : * that will have set the HEAP_XMAX_INVALID bit if the xmax was a MultiXactId
5320 : * but it was not running anymore. There is a race condition, which is that the
5321 : * MultiXactId may have finished since then, but that uncommon case is handled
5322 : * either here, or within MultiXactIdExpand.
5323 : *
5324 : * There is a similar race condition possible when the old xmax was a regular
5325 : * TransactionId. We test TransactionIdIsInProgress again just to narrow the
5326 : * window, but it's still possible to end up creating an unnecessary
5327 : * MultiXactId. Fortunately this is harmless.
5328 : */
5329 : static void
5330 4045232 : compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
5331 : uint16 old_infomask2, TransactionId add_to_xmax,
5332 : LockTupleMode mode, bool is_update,
5333 : TransactionId *result_xmax, uint16 *result_infomask,
5334 : uint16 *result_infomask2)
5335 : {
5336 : TransactionId new_xmax;
5337 : uint16 new_infomask,
5338 : new_infomask2;
5339 :
5340 : Assert(TransactionIdIsCurrentTransactionId(add_to_xmax));
5341 :
5342 207786 : l5:
5343 4253018 : new_infomask = 0;
5344 4253018 : new_infomask2 = 0;
5345 4253018 : if (old_infomask & HEAP_XMAX_INVALID)
5346 : {
5347 : /*
5348 : * No previous locker; we just insert our own TransactionId.
5349 : *
5350 : * Note that it's critical that this case be the first one checked,
5351 : * because there are several blocks below that come back to this one
5352 : * to implement certain optimizations; old_infomask might contain
5353 : * other dirty bits in those cases, but we don't really care.
5354 : */
5355 4042944 : if (is_update)
5356 : {
5357 3580888 : new_xmax = add_to_xmax;
5358 3580888 : if (mode == LockTupleExclusive)
5359 3047254 : new_infomask2 |= HEAP_KEYS_UPDATED;
5360 : }
5361 : else
5362 : {
5363 462056 : new_infomask |= HEAP_XMAX_LOCK_ONLY;
5364 462056 : switch (mode)
5365 : {
5366 5034 : case LockTupleKeyShare:
5367 5034 : new_xmax = add_to_xmax;
5368 5034 : new_infomask |= HEAP_XMAX_KEYSHR_LOCK;
5369 5034 : break;
5370 1462 : case LockTupleShare:
5371 1462 : new_xmax = add_to_xmax;
5372 1462 : new_infomask |= HEAP_XMAX_SHR_LOCK;
5373 1462 : break;
5374 264342 : case LockTupleNoKeyExclusive:
5375 264342 : new_xmax = add_to_xmax;
5376 264342 : new_infomask |= HEAP_XMAX_EXCL_LOCK;
5377 264342 : break;
5378 191218 : case LockTupleExclusive:
5379 191218 : new_xmax = add_to_xmax;
5380 191218 : new_infomask |= HEAP_XMAX_EXCL_LOCK;
5381 191218 : new_infomask2 |= HEAP_KEYS_UPDATED;
5382 191218 : break;
5383 0 : default:
5384 0 : new_xmax = InvalidTransactionId; /* silence compiler */
5385 0 : elog(ERROR, "invalid lock mode");
5386 : }
5387 : }
5388 : }
5389 210074 : else if (old_infomask & HEAP_XMAX_IS_MULTI)
5390 : {
5391 : MultiXactStatus new_status;
5392 :
5393 : /*
5394 : * Currently we don't allow XMAX_COMMITTED to be set for multis, so
5395 : * cross-check.
5396 : */
5397 : Assert(!(old_infomask & HEAP_XMAX_COMMITTED));
5398 :
5399 : /*
5400 : * A multixact together with LOCK_ONLY set but neither lock bit set
5401 : * (i.e. a pg_upgraded share locked tuple) cannot possibly be running
5402 : * anymore. This check is critical for databases upgraded by
5403 : * pg_upgrade; both MultiXactIdIsRunning and MultiXactIdExpand assume
5404 : * that such multis are never passed.
5405 : */
5406 254 : if (HEAP_LOCKED_UPGRADED(old_infomask))
5407 : {
5408 0 : old_infomask &= ~HEAP_XMAX_IS_MULTI;
5409 0 : old_infomask |= HEAP_XMAX_INVALID;
5410 0 : goto l5;
5411 : }
5412 :
5413 : /*
5414 : * If the XMAX is already a MultiXactId, then we need to expand it to
5415 : * include add_to_xmax; but if all the members were lockers and are
5416 : * all gone, we can do away with the IS_MULTI bit and just set
5417 : * add_to_xmax as the only locker/updater. If all lockers are gone
5418 : * and we have an updater that aborted, we can also do without a
5419 : * multi.
5420 : *
5421 : * The cost of doing GetMultiXactIdMembers would be paid by
5422 : * MultiXactIdExpand if we weren't to do this, so this check is not
5423 : * incurring extra work anyhow.
5424 : */
5425 254 : if (!MultiXactIdIsRunning(xmax, HEAP_XMAX_IS_LOCKED_ONLY(old_infomask)))
5426 : {
5427 46 : if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) ||
5428 16 : !TransactionIdDidCommit(MultiXactIdGetUpdateXid(xmax,
5429 : old_infomask)))
5430 : {
5431 : /*
5432 : * Reset these bits and restart; otherwise fall through to
5433 : * create a new multi below.
5434 : */
5435 46 : old_infomask &= ~HEAP_XMAX_IS_MULTI;
5436 46 : old_infomask |= HEAP_XMAX_INVALID;
5437 46 : goto l5;
5438 : }
5439 : }
5440 :
5441 208 : new_status = get_mxact_status_for_lock(mode, is_update);
5442 :
5443 208 : new_xmax = MultiXactIdExpand((MultiXactId) xmax, add_to_xmax,
5444 : new_status);
5445 208 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5446 : }
5447 209820 : else if (old_infomask & HEAP_XMAX_COMMITTED)
5448 : {
5449 : /*
5450 : * It's a committed update, so we need to preserve him as updater of
5451 : * the tuple.
5452 : */
5453 : MultiXactStatus status;
5454 : MultiXactStatus new_status;
5455 :
5456 26 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5457 0 : status = MultiXactStatusUpdate;
5458 : else
5459 26 : status = MultiXactStatusNoKeyUpdate;
5460 :
5461 26 : new_status = get_mxact_status_for_lock(mode, is_update);
5462 :
5463 : /*
5464 : * since it's not running, it's obviously impossible for the old
5465 : * updater to be identical to the current one, so we need not check
5466 : * for that case as we do in the block above.
5467 : */
5468 26 : new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
5469 26 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5470 : }
5471 209794 : else if (TransactionIdIsInProgress(xmax))
5472 : {
5473 : /*
5474 : * If the XMAX is a valid, in-progress TransactionId, then we need to
5475 : * create a new MultiXactId that includes both the old locker or
5476 : * updater and our own TransactionId.
5477 : */
5478 : MultiXactStatus new_status;
5479 : MultiXactStatus old_status;
5480 : LockTupleMode old_mode;
5481 :
5482 209776 : if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5483 : {
5484 209724 : if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5485 11254 : old_status = MultiXactStatusForKeyShare;
5486 198470 : else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5487 864 : old_status = MultiXactStatusForShare;
5488 197606 : else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5489 : {
5490 197606 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5491 185338 : old_status = MultiXactStatusForUpdate;
5492 : else
5493 12268 : old_status = MultiXactStatusForNoKeyUpdate;
5494 : }
5495 : else
5496 : {
5497 : /*
5498 : * LOCK_ONLY can be present alone only when a page has been
5499 : * upgraded by pg_upgrade. But in that case,
5500 : * TransactionIdIsInProgress() should have returned false. We
5501 : * assume it's no longer locked in this case.
5502 : */
5503 0 : elog(WARNING, "LOCK_ONLY found for Xid in progress %u", xmax);
5504 0 : old_infomask |= HEAP_XMAX_INVALID;
5505 0 : old_infomask &= ~HEAP_XMAX_LOCK_ONLY;
5506 0 : goto l5;
5507 : }
5508 : }
5509 : else
5510 : {
5511 : /* it's an update, but which kind? */
5512 52 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5513 0 : old_status = MultiXactStatusUpdate;
5514 : else
5515 52 : old_status = MultiXactStatusNoKeyUpdate;
5516 : }
5517 :
5518 209776 : old_mode = TUPLOCK_from_mxstatus(old_status);
5519 :
5520 : /*
5521 : * If the lock to be acquired is for the same TransactionId as the
5522 : * existing lock, there's an optimization possible: consider only the
5523 : * strongest of both locks as the only one present, and restart.
5524 : */
5525 209776 : if (xmax == add_to_xmax)
5526 : {
5527 : /*
5528 : * Note that it's not possible for the original tuple to be
5529 : * updated: we wouldn't be here because the tuple would have been
5530 : * invisible and we wouldn't try to update it. As a subtlety,
5531 : * this code can also run when traversing an update chain to lock
5532 : * future versions of a tuple. But we wouldn't be here either,
5533 : * because the add_to_xmax would be different from the original
5534 : * updater.
5535 : */
5536 : Assert(HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5537 :
5538 : /* acquire the strongest of both */
5539 207724 : if (mode < old_mode)
5540 104220 : mode = old_mode;
5541 : /* mustn't touch is_update */
5542 :
5543 207724 : old_infomask |= HEAP_XMAX_INVALID;
5544 207724 : goto l5;
5545 : }
5546 :
5547 : /* otherwise, just fall back to creating a new multixact */
5548 2052 : new_status = get_mxact_status_for_lock(mode, is_update);
5549 2052 : new_xmax = MultiXactIdCreate(xmax, old_status,
5550 : add_to_xmax, new_status);
5551 2052 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5552 : }
5553 28 : else if (!HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) &&
5554 10 : TransactionIdDidCommit(xmax))
5555 2 : {
5556 : /*
5557 : * It's a committed update, so we gotta preserve him as updater of the
5558 : * tuple.
5559 : */
5560 : MultiXactStatus status;
5561 : MultiXactStatus new_status;
5562 :
5563 2 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5564 0 : status = MultiXactStatusUpdate;
5565 : else
5566 2 : status = MultiXactStatusNoKeyUpdate;
5567 :
5568 2 : new_status = get_mxact_status_for_lock(mode, is_update);
5569 :
5570 : /*
5571 : * since it's not running, it's obviously impossible for the old
5572 : * updater to be identical to the current one, so we need not check
5573 : * for that case as we do in the block above.
5574 : */
5575 2 : new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
5576 2 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5577 : }
5578 : else
5579 : {
5580 : /*
5581 : * Can get here iff the locking/updating transaction was running when
5582 : * the infomask was extracted from the tuple, but finished before
5583 : * TransactionIdIsInProgress got to run. Deal with it as if there was
5584 : * no locker at all in the first place.
5585 : */
5586 16 : old_infomask |= HEAP_XMAX_INVALID;
5587 16 : goto l5;
5588 : }
5589 :
5590 4045232 : *result_infomask = new_infomask;
5591 4045232 : *result_infomask2 = new_infomask2;
5592 4045232 : *result_xmax = new_xmax;
5593 4045232 : }
5594 :
5595 : /*
5596 : * Subroutine for heap_lock_updated_tuple_rec.
5597 : *
5598 : * Given a hypothetical multixact status held by the transaction identified
5599 : * with the given xid, does the current transaction need to wait, fail, or can
5600 : * it continue if it wanted to acquire a lock of the given mode? "needwait"
5601 : * is set to true if waiting is necessary; if it can continue, then TM_Ok is
5602 : * returned. If the lock is already held by the current transaction, return
5603 : * TM_SelfModified. In case of a conflict with another transaction, a
5604 : * different HeapTupleSatisfiesUpdate return code is returned.
5605 : *
5606 : * The held status is said to be hypothetical because it might correspond to a
5607 : * lock held by a single Xid, i.e. not a real MultiXactId; we express it this
5608 : * way for simplicity of API.
5609 : */
5610 : static TM_Result
5611 64 : test_lockmode_for_conflict(MultiXactStatus status, TransactionId xid,
5612 : LockTupleMode mode, HeapTuple tup,
5613 : bool *needwait)
5614 : {
5615 : MultiXactStatus wantedstatus;
5616 :
5617 64 : *needwait = false;
5618 64 : wantedstatus = get_mxact_status_for_lock(mode, false);
5619 :
5620 : /*
5621 : * Note: we *must* check TransactionIdIsInProgress before
5622 : * TransactionIdDidAbort/Commit; see comment at top of heapam_visibility.c
5623 : * for an explanation.
5624 : */
5625 64 : if (TransactionIdIsCurrentTransactionId(xid))
5626 : {
5627 : /*
5628 : * The tuple has already been locked by our own transaction. This is
5629 : * very rare but can happen if multiple transactions are trying to
5630 : * lock an ancient version of the same tuple.
5631 : */
5632 0 : return TM_SelfModified;
5633 : }
5634 64 : else if (TransactionIdIsInProgress(xid))
5635 : {
5636 : /*
5637 : * If the locking transaction is running, what we do depends on
5638 : * whether the lock modes conflict: if they do, then we must wait for
5639 : * it to finish; otherwise we can fall through to lock this tuple
5640 : * version without waiting.
5641 : */
5642 32 : if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5643 32 : LOCKMODE_from_mxstatus(wantedstatus)))
5644 : {
5645 16 : *needwait = true;
5646 : }
5647 :
5648 : /*
5649 : * If we set needwait above, then this value doesn't matter;
5650 : * otherwise, this value signals to caller that it's okay to proceed.
5651 : */
5652 32 : return TM_Ok;
5653 : }
5654 32 : else if (TransactionIdDidAbort(xid))
5655 6 : return TM_Ok;
5656 26 : else if (TransactionIdDidCommit(xid))
5657 : {
5658 : /*
5659 : * The other transaction committed. If it was only a locker, then the
5660 : * lock is completely gone now and we can return success; but if it
5661 : * was an update, then what we do depends on whether the two lock
5662 : * modes conflict. If they conflict, then we must report error to
5663 : * caller. But if they don't, we can fall through to allow the current
5664 : * transaction to lock the tuple.
5665 : *
5666 : * Note: the reason we worry about ISUPDATE here is because as soon as
5667 : * a transaction ends, all its locks are gone and meaningless, and
5668 : * thus we can ignore them; whereas its updates persist. In the
5669 : * TransactionIdIsInProgress case, above, we don't need to check
5670 : * because we know the lock is still "alive" and thus a conflict needs
5671 : * always be checked.
5672 : */
5673 26 : if (!ISUPDATE_from_mxstatus(status))
5674 8 : return TM_Ok;
5675 :
5676 18 : if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5677 18 : LOCKMODE_from_mxstatus(wantedstatus)))
5678 : {
5679 : /* bummer */
5680 16 : if (!ItemPointerEquals(&tup->t_self, &tup->t_data->t_ctid))
5681 12 : return TM_Updated;
5682 : else
5683 4 : return TM_Deleted;
5684 : }
5685 :
5686 2 : return TM_Ok;
5687 : }
5688 :
5689 : /* Not in progress, not aborted, not committed -- must have crashed */
5690 0 : return TM_Ok;
5691 : }
5692 :
5693 :
5694 : /*
5695 : * Recursive part of heap_lock_updated_tuple
5696 : *
5697 : * Fetch the tuple pointed to by tid in rel, and mark it as locked by the given
5698 : * xid with the given mode; if this tuple is updated, recurse to lock the new
5699 : * version as well.
5700 : */
5701 : static TM_Result
5702 162 : heap_lock_updated_tuple_rec(Relation rel, ItemPointer tid, TransactionId xid,
5703 : LockTupleMode mode)
5704 : {
5705 : TM_Result result;
5706 : ItemPointerData tupid;
5707 : HeapTupleData mytup;
5708 : Buffer buf;
5709 : uint16 new_infomask,
5710 : new_infomask2,
5711 : old_infomask,
5712 : old_infomask2;
5713 : TransactionId xmax,
5714 : new_xmax;
5715 162 : TransactionId priorXmax = InvalidTransactionId;
5716 162 : bool cleared_all_frozen = false;
5717 : bool pinned_desired_page;
5718 162 : Buffer vmbuffer = InvalidBuffer;
5719 : BlockNumber block;
5720 :
5721 162 : ItemPointerCopy(tid, &tupid);
5722 :
5723 : for (;;)
5724 : {
5725 168 : new_infomask = 0;
5726 168 : new_xmax = InvalidTransactionId;
5727 168 : block = ItemPointerGetBlockNumber(&tupid);
5728 168 : ItemPointerCopy(&tupid, &(mytup.t_self));
5729 :
5730 168 : if (!heap_fetch(rel, SnapshotAny, &mytup, &buf, false))
5731 : {
5732 : /*
5733 : * if we fail to find the updated version of the tuple, it's
5734 : * because it was vacuumed/pruned away after its creator
5735 : * transaction aborted. So behave as if we got to the end of the
5736 : * chain, and there's no further tuple to lock: return success to
5737 : * caller.
5738 : */
5739 0 : result = TM_Ok;
5740 0 : goto out_unlocked;
5741 : }
5742 :
5743 168 : l4:
5744 184 : CHECK_FOR_INTERRUPTS();
5745 :
5746 : /*
5747 : * Before locking the buffer, pin the visibility map page if it
5748 : * appears to be necessary. Since we haven't got the lock yet,
5749 : * someone else might be in the middle of changing this, so we'll need
5750 : * to recheck after we have the lock.
5751 : */
5752 184 : if (PageIsAllVisible(BufferGetPage(buf)))
5753 : {
5754 0 : visibilitymap_pin(rel, block, &vmbuffer);
5755 0 : pinned_desired_page = true;
5756 : }
5757 : else
5758 184 : pinned_desired_page = false;
5759 :
5760 184 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5761 :
5762 : /*
5763 : * If we didn't pin the visibility map page and the page has become
5764 : * all visible while we were busy locking the buffer, we'll have to
5765 : * unlock and re-lock, to avoid holding the buffer lock across I/O.
5766 : * That's a bit unfortunate, but hopefully shouldn't happen often.
5767 : *
5768 : * Note: in some paths through this function, we will reach here
5769 : * holding a pin on a vm page that may or may not be the one matching
5770 : * this page. If this page isn't all-visible, we won't use the vm
5771 : * page, but we hold onto such a pin till the end of the function.
5772 : */
5773 184 : if (!pinned_desired_page && PageIsAllVisible(BufferGetPage(buf)))
5774 : {
5775 0 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5776 0 : visibilitymap_pin(rel, block, &vmbuffer);
5777 0 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5778 : }
5779 :
5780 : /*
5781 : * Check the tuple XMIN against prior XMAX, if any. If we reached the
5782 : * end of the chain, we're done, so return success.
5783 : */
5784 190 : if (TransactionIdIsValid(priorXmax) &&
5785 6 : !TransactionIdEquals(HeapTupleHeaderGetXmin(mytup.t_data),
5786 : priorXmax))
5787 : {
5788 0 : result = TM_Ok;
5789 0 : goto out_locked;
5790 : }
5791 :
5792 : /*
5793 : * Also check Xmin: if this tuple was created by an aborted
5794 : * (sub)transaction, then we already locked the last live one in the
5795 : * chain, thus we're done, so return success.
5796 : */
5797 184 : if (TransactionIdDidAbort(HeapTupleHeaderGetXmin(mytup.t_data)))
5798 : {
5799 26 : result = TM_Ok;
5800 26 : goto out_locked;
5801 : }
5802 :
5803 158 : old_infomask = mytup.t_data->t_infomask;
5804 158 : old_infomask2 = mytup.t_data->t_infomask2;
5805 158 : xmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5806 :
5807 : /*
5808 : * If this tuple version has been updated or locked by some concurrent
5809 : * transaction(s), what we do depends on whether our lock mode
5810 : * conflicts with what those other transactions hold, and also on the
5811 : * status of them.
5812 : */
5813 158 : if (!(old_infomask & HEAP_XMAX_INVALID))
5814 : {
5815 : TransactionId rawxmax;
5816 : bool needwait;
5817 :
5818 60 : rawxmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5819 60 : if (old_infomask & HEAP_XMAX_IS_MULTI)
5820 : {
5821 : int nmembers;
5822 : int i;
5823 : MultiXactMember *members;
5824 :
5825 : /*
5826 : * We don't need a test for pg_upgrade'd tuples: this is only
5827 : * applied to tuples after the first in an update chain. Said
5828 : * first tuple in the chain may well be locked-in-9.2-and-
5829 : * pg_upgraded, but that one was already locked by our caller,
5830 : * not us; and any subsequent ones cannot be because our
5831 : * caller must necessarily have obtained a snapshot later than
5832 : * the pg_upgrade itself.
5833 : */
5834 : Assert(!HEAP_LOCKED_UPGRADED(mytup.t_data->t_infomask));
5835 :
5836 2 : nmembers = GetMultiXactIdMembers(rawxmax, &members, false,
5837 2 : HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5838 8 : for (i = 0; i < nmembers; i++)
5839 : {
5840 6 : result = test_lockmode_for_conflict(members[i].status,
5841 6 : members[i].xid,
5842 : mode,
5843 : &mytup,
5844 : &needwait);
5845 :
5846 : /*
5847 : * If the tuple was already locked by ourselves in a
5848 : * previous iteration of this (say heap_lock_tuple was
5849 : * forced to restart the locking loop because of a change
5850 : * in xmax), then we hold the lock already on this tuple
5851 : * version and we don't need to do anything; and this is
5852 : * not an error condition either. We just need to skip
5853 : * this tuple and continue locking the next version in the
5854 : * update chain.
5855 : */
5856 6 : if (result == TM_SelfModified)
5857 : {
5858 0 : pfree(members);
5859 0 : goto next;
5860 : }
5861 :
5862 6 : if (needwait)
5863 : {
5864 0 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5865 0 : XactLockTableWait(members[i].xid, rel,
5866 : &mytup.t_self,
5867 : XLTW_LockUpdated);
5868 0 : pfree(members);
5869 0 : goto l4;
5870 : }
5871 6 : if (result != TM_Ok)
5872 : {
5873 0 : pfree(members);
5874 0 : goto out_locked;
5875 : }
5876 : }
5877 2 : if (members)
5878 2 : pfree(members);
5879 : }
5880 : else
5881 : {
5882 : MultiXactStatus status;
5883 :
5884 : /*
5885 : * For a non-multi Xmax, we first need to compute the
5886 : * corresponding MultiXactStatus by using the infomask bits.
5887 : */
5888 58 : if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5889 : {
5890 20 : if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5891 20 : status = MultiXactStatusForKeyShare;
5892 0 : else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5893 0 : status = MultiXactStatusForShare;
5894 0 : else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5895 : {
5896 0 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5897 0 : status = MultiXactStatusForUpdate;
5898 : else
5899 0 : status = MultiXactStatusForNoKeyUpdate;
5900 : }
5901 : else
5902 : {
5903 : /*
5904 : * LOCK_ONLY present alone (a pg_upgraded tuple marked
5905 : * as share-locked in the old cluster) shouldn't be
5906 : * seen in the middle of an update chain.
5907 : */
5908 0 : elog(ERROR, "invalid lock status in tuple");
5909 : }
5910 : }
5911 : else
5912 : {
5913 : /* it's an update, but which kind? */
5914 38 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5915 28 : status = MultiXactStatusUpdate;
5916 : else
5917 10 : status = MultiXactStatusNoKeyUpdate;
5918 : }
5919 :
5920 58 : result = test_lockmode_for_conflict(status, rawxmax, mode,
5921 : &mytup, &needwait);
5922 :
5923 : /*
5924 : * If the tuple was already locked by ourselves in a previous
5925 : * iteration of this (say heap_lock_tuple was forced to
5926 : * restart the locking loop because of a change in xmax), then
5927 : * we hold the lock already on this tuple version and we don't
5928 : * need to do anything; and this is not an error condition
5929 : * either. We just need to skip this tuple and continue
5930 : * locking the next version in the update chain.
5931 : */
5932 58 : if (result == TM_SelfModified)
5933 0 : goto next;
5934 :
5935 58 : if (needwait)
5936 : {
5937 16 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5938 16 : XactLockTableWait(rawxmax, rel, &mytup.t_self,
5939 : XLTW_LockUpdated);
5940 16 : goto l4;
5941 : }
5942 42 : if (result != TM_Ok)
5943 : {
5944 16 : goto out_locked;
5945 : }
5946 : }
5947 : }
5948 :
5949 : /* compute the new Xmax and infomask values for the tuple ... */
5950 126 : compute_new_xmax_infomask(xmax, old_infomask, mytup.t_data->t_infomask2,
5951 : xid, mode, false,
5952 : &new_xmax, &new_infomask, &new_infomask2);
5953 :
5954 126 : if (PageIsAllVisible(BufferGetPage(buf)) &&
5955 0 : visibilitymap_clear(rel, block, vmbuffer,
5956 : VISIBILITYMAP_ALL_FROZEN))
5957 0 : cleared_all_frozen = true;
5958 :
5959 126 : START_CRIT_SECTION();
5960 :
5961 : /* ... and set them */
5962 126 : HeapTupleHeaderSetXmax(mytup.t_data, new_xmax);
5963 126 : mytup.t_data->t_infomask &= ~HEAP_XMAX_BITS;
5964 126 : mytup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5965 126 : mytup.t_data->t_infomask |= new_infomask;
5966 126 : mytup.t_data->t_infomask2 |= new_infomask2;
5967 :
5968 126 : MarkBufferDirty(buf);
5969 :
5970 : /* XLOG stuff */
5971 126 : if (RelationNeedsWAL(rel))
5972 : {
5973 : xl_heap_lock_updated xlrec;
5974 : XLogRecPtr recptr;
5975 126 : Page page = BufferGetPage(buf);
5976 :
5977 126 : XLogBeginInsert();
5978 126 : XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
5979 :
5980 126 : xlrec.offnum = ItemPointerGetOffsetNumber(&mytup.t_self);
5981 126 : xlrec.xmax = new_xmax;
5982 126 : xlrec.infobits_set = compute_infobits(new_infomask, new_infomask2);
5983 126 : xlrec.flags =
5984 126 : cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
5985 :
5986 126 : XLogRegisterData(&xlrec, SizeOfHeapLockUpdated);
5987 :
5988 126 : recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_LOCK_UPDATED);
5989 :
5990 126 : PageSetLSN(page, recptr);
5991 : }
5992 :
5993 126 : END_CRIT_SECTION();
5994 :
5995 126 : next:
5996 : /* if we find the end of update chain, we're done. */
5997 252 : if (mytup.t_data->t_infomask & HEAP_XMAX_INVALID ||
5998 252 : HeapTupleHeaderIndicatesMovedPartitions(mytup.t_data) ||
5999 134 : ItemPointerEquals(&mytup.t_self, &mytup.t_data->t_ctid) ||
6000 8 : HeapTupleHeaderIsOnlyLocked(mytup.t_data))
6001 : {
6002 120 : result = TM_Ok;
6003 120 : goto out_locked;
6004 : }
6005 :
6006 : /* tail recursion */
6007 6 : priorXmax = HeapTupleHeaderGetUpdateXid(mytup.t_data);
6008 6 : ItemPointerCopy(&(mytup.t_data->t_ctid), &tupid);
6009 6 : UnlockReleaseBuffer(buf);
6010 : }
6011 :
6012 : result = TM_Ok;
6013 :
6014 162 : out_locked:
6015 162 : UnlockReleaseBuffer(buf);
6016 :
6017 162 : out_unlocked:
6018 162 : if (vmbuffer != InvalidBuffer)
6019 0 : ReleaseBuffer(vmbuffer);
6020 :
6021 162 : return result;
6022 : }
6023 :
6024 : /*
6025 : * heap_lock_updated_tuple
6026 : * Follow update chain when locking an updated tuple, acquiring locks (row
6027 : * marks) on the updated versions.
6028 : *
6029 : * The initial tuple is assumed to be already locked.
6030 : *
6031 : * This function doesn't check visibility, it just unconditionally marks the
6032 : * tuple(s) as locked. If any tuple in the updated chain is being deleted
6033 : * concurrently (or updated with the key being modified), sleep until the
6034 : * transaction doing it is finished.
6035 : *
6036 : * Note that we don't acquire heavyweight tuple locks on the tuples we walk
6037 : * when we have to wait for other transactions to release them, as opposed to
6038 : * what heap_lock_tuple does. The reason is that having more than one
6039 : * transaction walking the chain is probably uncommon enough that risk of
6040 : * starvation is not likely: one of the preconditions for being here is that
6041 : * the snapshot in use predates the update that created this tuple (because we
6042 : * started at an earlier version of the tuple), but at the same time such a
6043 : * transaction cannot be using repeatable read or serializable isolation
6044 : * levels, because that would lead to a serializability failure.
6045 : */
6046 : static TM_Result
6047 180 : heap_lock_updated_tuple(Relation rel, HeapTuple tuple, ItemPointer ctid,
6048 : TransactionId xid, LockTupleMode mode)
6049 : {
6050 : /*
6051 : * If the tuple has not been updated, or has moved into another partition
6052 : * (effectively a delete) stop here.
6053 : */
6054 180 : if (!HeapTupleHeaderIndicatesMovedPartitions(tuple->t_data) &&
6055 176 : !ItemPointerEquals(&tuple->t_self, ctid))
6056 : {
6057 : /*
6058 : * If this is the first possibly-multixact-able operation in the
6059 : * current transaction, set my per-backend OldestMemberMXactId
6060 : * setting. We can be certain that the transaction will never become a
6061 : * member of any older MultiXactIds than that. (We have to do this
6062 : * even if we end up just using our own TransactionId below, since
6063 : * some other backend could incorporate our XID into a MultiXact
6064 : * immediately afterwards.)
6065 : */
6066 162 : MultiXactIdSetOldestMember();
6067 :
6068 162 : return heap_lock_updated_tuple_rec(rel, ctid, xid, mode);
6069 : }
6070 :
6071 : /* nothing to lock */
6072 18 : return TM_Ok;
6073 : }
6074 :
6075 : /*
6076 : * heap_finish_speculative - mark speculative insertion as successful
6077 : *
6078 : * To successfully finish a speculative insertion we have to clear speculative
6079 : * token from tuple. To do so the t_ctid field, which will contain a
6080 : * speculative token value, is modified in place to point to the tuple itself,
6081 : * which is characteristic of a newly inserted ordinary tuple.
6082 : *
6083 : * NB: It is not ok to commit without either finishing or aborting a
6084 : * speculative insertion. We could treat speculative tuples of committed
6085 : * transactions implicitly as completed, but then we would have to be prepared
6086 : * to deal with speculative tokens on committed tuples. That wouldn't be
6087 : * difficult - no-one looks at the ctid field of a tuple with invalid xmax -
6088 : * but clearing the token at completion isn't very expensive either.
6089 : * An explicit confirmation WAL record also makes logical decoding simpler.
6090 : */
6091 : void
6092 4112 : heap_finish_speculative(Relation relation, ItemPointer tid)
6093 : {
6094 : Buffer buffer;
6095 : Page page;
6096 : OffsetNumber offnum;
6097 4112 : ItemId lp = NULL;
6098 : HeapTupleHeader htup;
6099 :
6100 4112 : buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
6101 4112 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6102 4112 : page = BufferGetPage(buffer);
6103 :
6104 4112 : offnum = ItemPointerGetOffsetNumber(tid);
6105 4112 : if (PageGetMaxOffsetNumber(page) >= offnum)
6106 4112 : lp = PageGetItemId(page, offnum);
6107 :
6108 4112 : if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
6109 0 : elog(ERROR, "invalid lp");
6110 :
6111 4112 : htup = (HeapTupleHeader) PageGetItem(page, lp);
6112 :
6113 : /* NO EREPORT(ERROR) from here till changes are logged */
6114 4112 : START_CRIT_SECTION();
6115 :
6116 : Assert(HeapTupleHeaderIsSpeculative(htup));
6117 :
6118 4112 : MarkBufferDirty(buffer);
6119 :
6120 : /*
6121 : * Replace the speculative insertion token with a real t_ctid, pointing to
6122 : * itself like it does on regular tuples.
6123 : */
6124 4112 : htup->t_ctid = *tid;
6125 :
6126 : /* XLOG stuff */
6127 4112 : if (RelationNeedsWAL(relation))
6128 : {
6129 : xl_heap_confirm xlrec;
6130 : XLogRecPtr recptr;
6131 :
6132 4094 : xlrec.offnum = ItemPointerGetOffsetNumber(tid);
6133 :
6134 4094 : XLogBeginInsert();
6135 :
6136 : /* We want the same filtering on this as on a plain insert */
6137 4094 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
6138 :
6139 4094 : XLogRegisterData(&xlrec, SizeOfHeapConfirm);
6140 4094 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
6141 :
6142 4094 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_CONFIRM);
6143 :
6144 4094 : PageSetLSN(page, recptr);
6145 : }
6146 :
6147 4112 : END_CRIT_SECTION();
6148 :
6149 4112 : UnlockReleaseBuffer(buffer);
6150 4112 : }
6151 :
6152 : /*
6153 : * heap_abort_speculative - kill a speculatively inserted tuple
6154 : *
6155 : * Marks a tuple that was speculatively inserted in the same command as dead,
6156 : * by setting its xmin as invalid. That makes it immediately appear as dead
6157 : * to all transactions, including our own. In particular, it makes
6158 : * HeapTupleSatisfiesDirty() regard the tuple as dead, so that another backend
6159 : * inserting a duplicate key value won't unnecessarily wait for our whole
6160 : * transaction to finish (it'll just wait for our speculative insertion to
6161 : * finish).
6162 : *
6163 : * Killing the tuple prevents "unprincipled deadlocks", which are deadlocks
6164 : * that arise due to a mutual dependency that is not user visible. By
6165 : * definition, unprincipled deadlocks cannot be prevented by the user
6166 : * reordering lock acquisition in client code, because the implementation level
6167 : * lock acquisitions are not under the user's direct control. If speculative
6168 : * inserters did not take this precaution, then under high concurrency they
6169 : * could deadlock with each other, which would not be acceptable.
6170 : *
6171 : * This is somewhat redundant with heap_delete, but we prefer to have a
6172 : * dedicated routine with stripped down requirements. Note that this is also
6173 : * used to delete the TOAST tuples created during speculative insertion.
6174 : *
6175 : * This routine does not affect logical decoding as it only looks at
6176 : * confirmation records.
6177 : */
6178 : void
6179 20 : heap_abort_speculative(Relation relation, ItemPointer tid)
6180 : {
6181 20 : TransactionId xid = GetCurrentTransactionId();
6182 : ItemId lp;
6183 : HeapTupleData tp;
6184 : Page page;
6185 : BlockNumber block;
6186 : Buffer buffer;
6187 :
6188 : Assert(ItemPointerIsValid(tid));
6189 :
6190 20 : block = ItemPointerGetBlockNumber(tid);
6191 20 : buffer = ReadBuffer(relation, block);
6192 20 : page = BufferGetPage(buffer);
6193 :
6194 20 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6195 :
6196 : /*
6197 : * Page can't be all visible, we just inserted into it, and are still
6198 : * running.
6199 : */
6200 : Assert(!PageIsAllVisible(page));
6201 :
6202 20 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
6203 : Assert(ItemIdIsNormal(lp));
6204 :
6205 20 : tp.t_tableOid = RelationGetRelid(relation);
6206 20 : tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
6207 20 : tp.t_len = ItemIdGetLength(lp);
6208 20 : tp.t_self = *tid;
6209 :
6210 : /*
6211 : * Sanity check that the tuple really is a speculatively inserted tuple,
6212 : * inserted by us.
6213 : */
6214 20 : if (tp.t_data->t_choice.t_heap.t_xmin != xid)
6215 0 : elog(ERROR, "attempted to kill a tuple inserted by another transaction");
6216 20 : if (!(IsToastRelation(relation) || HeapTupleHeaderIsSpeculative(tp.t_data)))
6217 0 : elog(ERROR, "attempted to kill a non-speculative tuple");
6218 : Assert(!HeapTupleHeaderIsHeapOnly(tp.t_data));
6219 :
6220 : /*
6221 : * No need to check for serializable conflicts here. There is never a
6222 : * need for a combo CID, either. No need to extract replica identity, or
6223 : * do anything special with infomask bits.
6224 : */
6225 :
6226 20 : START_CRIT_SECTION();
6227 :
6228 : /*
6229 : * The tuple will become DEAD immediately. Flag that this page is a
6230 : * candidate for pruning by setting xmin to TransactionXmin. While not
6231 : * immediately prunable, it is the oldest xid we can cheaply determine
6232 : * that's safe against wraparound / being older than the table's
6233 : * relfrozenxid. To defend against the unlikely case of a new relation
6234 : * having a newer relfrozenxid than our TransactionXmin, use relfrozenxid
6235 : * if so (vacuum can't subsequently move relfrozenxid to beyond
6236 : * TransactionXmin, so there's no race here).
6237 : */
6238 : Assert(TransactionIdIsValid(TransactionXmin));
6239 : {
6240 20 : TransactionId relfrozenxid = relation->rd_rel->relfrozenxid;
6241 : TransactionId prune_xid;
6242 :
6243 20 : if (TransactionIdPrecedes(TransactionXmin, relfrozenxid))
6244 0 : prune_xid = relfrozenxid;
6245 : else
6246 20 : prune_xid = TransactionXmin;
6247 20 : PageSetPrunable(page, prune_xid);
6248 : }
6249 :
6250 : /* store transaction information of xact deleting the tuple */
6251 20 : tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
6252 20 : tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
6253 :
6254 : /*
6255 : * Set the tuple header xmin to InvalidTransactionId. This makes the
6256 : * tuple immediately invisible everyone. (In particular, to any
6257 : * transactions waiting on the speculative token, woken up later.)
6258 : */
6259 20 : HeapTupleHeaderSetXmin(tp.t_data, InvalidTransactionId);
6260 :
6261 : /* Clear the speculative insertion token too */
6262 20 : tp.t_data->t_ctid = tp.t_self;
6263 :
6264 20 : MarkBufferDirty(buffer);
6265 :
6266 : /*
6267 : * XLOG stuff
6268 : *
6269 : * The WAL records generated here match heap_delete(). The same recovery
6270 : * routines are used.
6271 : */
6272 20 : if (RelationNeedsWAL(relation))
6273 : {
6274 : xl_heap_delete xlrec;
6275 : XLogRecPtr recptr;
6276 :
6277 20 : xlrec.flags = XLH_DELETE_IS_SUPER;
6278 40 : xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
6279 20 : tp.t_data->t_infomask2);
6280 20 : xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
6281 20 : xlrec.xmax = xid;
6282 :
6283 20 : XLogBeginInsert();
6284 20 : XLogRegisterData(&xlrec, SizeOfHeapDelete);
6285 20 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
6286 :
6287 : /* No replica identity & replication origin logged */
6288 :
6289 20 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
6290 :
6291 20 : PageSetLSN(page, recptr);
6292 : }
6293 :
6294 20 : END_CRIT_SECTION();
6295 :
6296 20 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6297 :
6298 20 : if (HeapTupleHasExternal(&tp))
6299 : {
6300 : Assert(!IsToastRelation(relation));
6301 2 : heap_toast_delete(relation, &tp, true);
6302 : }
6303 :
6304 : /*
6305 : * Never need to mark tuple for invalidation, since catalogs don't support
6306 : * speculative insertion
6307 : */
6308 :
6309 : /* Now we can release the buffer */
6310 20 : ReleaseBuffer(buffer);
6311 :
6312 : /* count deletion, as we counted the insertion too */
6313 20 : pgstat_count_heap_delete(relation);
6314 20 : }
6315 :
6316 : /*
6317 : * heap_inplace_lock - protect inplace update from concurrent heap_update()
6318 : *
6319 : * Evaluate whether the tuple's state is compatible with a no-key update.
6320 : * Current transaction rowmarks are fine, as is KEY SHARE from any
6321 : * transaction. If compatible, return true with the buffer exclusive-locked,
6322 : * and the caller must release that by calling
6323 : * heap_inplace_update_and_unlock(), calling heap_inplace_unlock(), or raising
6324 : * an error. Otherwise, call release_callback(arg), wait for blocking
6325 : * transactions to end, and return false.
6326 : *
6327 : * Since this is intended for system catalogs and SERIALIZABLE doesn't cover
6328 : * DDL, this doesn't guarantee any particular predicate locking.
6329 : *
6330 : * One could modify this to return true for tuples with delete in progress,
6331 : * All inplace updaters take a lock that conflicts with DROP. If explicit
6332 : * "DELETE FROM pg_class" is in progress, we'll wait for it like we would an
6333 : * update.
6334 : *
6335 : * Readers of inplace-updated fields expect changes to those fields are
6336 : * durable. For example, vac_truncate_clog() reads datfrozenxid from
6337 : * pg_database tuples via catalog snapshots. A future snapshot must not
6338 : * return a lower datfrozenxid for the same database OID (lower in the
6339 : * FullTransactionIdPrecedes() sense). We achieve that since no update of a
6340 : * tuple can start while we hold a lock on its buffer. In cases like
6341 : * BEGIN;GRANT;CREATE INDEX;COMMIT we're inplace-updating a tuple visible only
6342 : * to this transaction. ROLLBACK then is one case where it's okay to lose
6343 : * inplace updates. (Restoring relhasindex=false on ROLLBACK is fine, since
6344 : * any concurrent CREATE INDEX would have blocked, then inplace-updated the
6345 : * committed tuple.)
6346 : *
6347 : * In principle, we could avoid waiting by overwriting every tuple in the
6348 : * updated tuple chain. Reader expectations permit updating a tuple only if
6349 : * it's aborted, is the tail of the chain, or we already updated the tuple
6350 : * referenced in its t_ctid. Hence, we would need to overwrite the tuples in
6351 : * order from tail to head. That would imply either (a) mutating all tuples
6352 : * in one critical section or (b) accepting a chance of partial completion.
6353 : * Partial completion of a relfrozenxid update would have the weird
6354 : * consequence that the table's next VACUUM could see the table's relfrozenxid
6355 : * move forward between vacuum_get_cutoffs() and finishing.
6356 : */
6357 : bool
6358 416242 : heap_inplace_lock(Relation relation,
6359 : HeapTuple oldtup_ptr, Buffer buffer,
6360 : void (*release_callback) (void *), void *arg)
6361 : {
6362 416242 : HeapTupleData oldtup = *oldtup_ptr; /* minimize diff vs. heap_update() */
6363 : TM_Result result;
6364 : bool ret;
6365 :
6366 : #ifdef USE_ASSERT_CHECKING
6367 : if (RelationGetRelid(relation) == RelationRelationId)
6368 : check_inplace_rel_lock(oldtup_ptr);
6369 : #endif
6370 :
6371 : Assert(BufferIsValid(buffer));
6372 :
6373 : /*
6374 : * Construct shared cache inval if necessary. Because we pass a tuple
6375 : * version without our own inplace changes or inplace changes other
6376 : * sessions complete while we wait for locks, inplace update mustn't
6377 : * change catcache lookup keys. But we aren't bothering with index
6378 : * updates either, so that's true a fortiori. After LockBuffer(), it
6379 : * would be too late, because this might reach a
6380 : * CatalogCacheInitializeCache() that locks "buffer".
6381 : */
6382 416242 : CacheInvalidateHeapTupleInplace(relation, oldtup_ptr, NULL);
6383 :
6384 416242 : LockTuple(relation, &oldtup.t_self, InplaceUpdateTupleLock);
6385 416242 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6386 :
6387 : /*----------
6388 : * Interpret HeapTupleSatisfiesUpdate() like heap_update() does, except:
6389 : *
6390 : * - wait unconditionally
6391 : * - already locked tuple above, since inplace needs that unconditionally
6392 : * - don't recheck header after wait: simpler to defer to next iteration
6393 : * - don't try to continue even if the updater aborts: likewise
6394 : * - no crosscheck
6395 : */
6396 416242 : result = HeapTupleSatisfiesUpdate(&oldtup, GetCurrentCommandId(false),
6397 : buffer);
6398 :
6399 416242 : if (result == TM_Invisible)
6400 : {
6401 : /* no known way this can happen */
6402 0 : ereport(ERROR,
6403 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6404 : errmsg_internal("attempted to overwrite invisible tuple")));
6405 : }
6406 416242 : else if (result == TM_SelfModified)
6407 : {
6408 : /*
6409 : * CREATE INDEX might reach this if an expression is silly enough to
6410 : * call e.g. SELECT ... FROM pg_class FOR SHARE. C code of other SQL
6411 : * statements might get here after a heap_update() of the same row, in
6412 : * the absence of an intervening CommandCounterIncrement().
6413 : */
6414 0 : ereport(ERROR,
6415 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6416 : errmsg("tuple to be updated was already modified by an operation triggered by the current command")));
6417 : }
6418 416242 : else if (result == TM_BeingModified)
6419 : {
6420 : TransactionId xwait;
6421 : uint16 infomask;
6422 :
6423 78 : xwait = HeapTupleHeaderGetRawXmax(oldtup.t_data);
6424 78 : infomask = oldtup.t_data->t_infomask;
6425 :
6426 78 : if (infomask & HEAP_XMAX_IS_MULTI)
6427 : {
6428 10 : LockTupleMode lockmode = LockTupleNoKeyExclusive;
6429 10 : MultiXactStatus mxact_status = MultiXactStatusNoKeyUpdate;
6430 : int remain;
6431 :
6432 10 : if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
6433 : lockmode, NULL))
6434 : {
6435 4 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6436 4 : release_callback(arg);
6437 4 : ret = false;
6438 4 : MultiXactIdWait((MultiXactId) xwait, mxact_status, infomask,
6439 : relation, &oldtup.t_self, XLTW_Update,
6440 : &remain);
6441 : }
6442 : else
6443 6 : ret = true;
6444 : }
6445 68 : else if (TransactionIdIsCurrentTransactionId(xwait))
6446 2 : ret = true;
6447 66 : else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask))
6448 2 : ret = true;
6449 : else
6450 : {
6451 64 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6452 64 : release_callback(arg);
6453 64 : ret = false;
6454 64 : XactLockTableWait(xwait, relation, &oldtup.t_self,
6455 : XLTW_Update);
6456 : }
6457 : }
6458 : else
6459 : {
6460 416164 : ret = (result == TM_Ok);
6461 416164 : if (!ret)
6462 : {
6463 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6464 0 : release_callback(arg);
6465 : }
6466 : }
6467 :
6468 : /*
6469 : * GetCatalogSnapshot() relies on invalidation messages to know when to
6470 : * take a new snapshot. COMMIT of xwait is responsible for sending the
6471 : * invalidation. We're not acquiring heavyweight locks sufficient to
6472 : * block if not yet sent, so we must take a new snapshot to ensure a later
6473 : * attempt has a fair chance. While we don't need this if xwait aborted,
6474 : * don't bother optimizing that.
6475 : */
6476 416242 : if (!ret)
6477 : {
6478 68 : UnlockTuple(relation, &oldtup.t_self, InplaceUpdateTupleLock);
6479 68 : ForgetInplace_Inval();
6480 68 : InvalidateCatalogSnapshot();
6481 : }
6482 416242 : return ret;
6483 : }
6484 :
6485 : /*
6486 : * heap_inplace_update_and_unlock - core of systable_inplace_update_finish
6487 : *
6488 : * The tuple cannot change size, and therefore its header fields and null
6489 : * bitmap (if any) don't change either.
6490 : *
6491 : * Since we hold LOCKTAG_TUPLE, no updater has a local copy of this tuple.
6492 : */
6493 : void
6494 161362 : heap_inplace_update_and_unlock(Relation relation,
6495 : HeapTuple oldtup, HeapTuple tuple,
6496 : Buffer buffer)
6497 : {
6498 161362 : HeapTupleHeader htup = oldtup->t_data;
6499 : uint32 oldlen;
6500 : uint32 newlen;
6501 : char *dst;
6502 : char *src;
6503 161362 : int nmsgs = 0;
6504 161362 : SharedInvalidationMessage *invalMessages = NULL;
6505 161362 : bool RelcacheInitFileInval = false;
6506 :
6507 : Assert(ItemPointerEquals(&oldtup->t_self, &tuple->t_self));
6508 161362 : oldlen = oldtup->t_len - htup->t_hoff;
6509 161362 : newlen = tuple->t_len - tuple->t_data->t_hoff;
6510 161362 : if (oldlen != newlen || htup->t_hoff != tuple->t_data->t_hoff)
6511 0 : elog(ERROR, "wrong tuple length");
6512 :
6513 161362 : dst = (char *) htup + htup->t_hoff;
6514 161362 : src = (char *) tuple->t_data + tuple->t_data->t_hoff;
6515 :
6516 : /* Like RecordTransactionCommit(), log only if needed */
6517 161362 : if (XLogStandbyInfoActive())
6518 98614 : nmsgs = inplaceGetInvalidationMessages(&invalMessages,
6519 : &RelcacheInitFileInval);
6520 :
6521 : /*
6522 : * Unlink relcache init files as needed. If unlinking, acquire
6523 : * RelCacheInitLock until after associated invalidations. By doing this
6524 : * in advance, if we checkpoint and then crash between inplace
6525 : * XLogInsert() and inval, we don't rely on StartupXLOG() ->
6526 : * RelationCacheInitFileRemove(). That uses elevel==LOG, so replay would
6527 : * neglect to PANIC on EIO.
6528 : */
6529 161362 : PreInplace_Inval();
6530 :
6531 : /*----------
6532 : * NO EREPORT(ERROR) from here till changes are complete
6533 : *
6534 : * Our buffer lock won't stop a reader having already pinned and checked
6535 : * visibility for this tuple. Hence, we write WAL first, then mutate the
6536 : * buffer. Like in MarkBufferDirtyHint() or RecordTransactionCommit(),
6537 : * checkpoint delay makes that acceptable. With the usual order of
6538 : * changes, a crash after memcpy() and before XLogInsert() could allow
6539 : * datfrozenxid to overtake relfrozenxid:
6540 : *
6541 : * ["D" is a VACUUM (ONLY_DATABASE_STATS)]
6542 : * ["R" is a VACUUM tbl]
6543 : * D: vac_update_datfrozenxid() -> systable_beginscan(pg_class)
6544 : * D: systable_getnext() returns pg_class tuple of tbl
6545 : * R: memcpy() into pg_class tuple of tbl
6546 : * D: raise pg_database.datfrozenxid, XLogInsert(), finish
6547 : * [crash]
6548 : * [recovery restores datfrozenxid w/o relfrozenxid]
6549 : *
6550 : * Mimic MarkBufferDirtyHint() subroutine XLogSaveBufferForHint().
6551 : * Specifically, use DELAY_CHKPT_START, and copy the buffer to the stack.
6552 : * The stack copy facilitates a FPI of the post-mutation block before we
6553 : * accept other sessions seeing it. DELAY_CHKPT_START allows us to
6554 : * XLogInsert() before MarkBufferDirty(). Since XLogSaveBufferForHint()
6555 : * can operate under BUFFER_LOCK_SHARED, it can't avoid DELAY_CHKPT_START.
6556 : * This function, however, likely could avoid it with the following order
6557 : * of operations: MarkBufferDirty(), XLogInsert(), memcpy(). Opt to use
6558 : * DELAY_CHKPT_START here, too, as a way to have fewer distinct code
6559 : * patterns to analyze. Inplace update isn't so frequent that it should
6560 : * pursue the small optimization of skipping DELAY_CHKPT_START.
6561 : */
6562 : Assert((MyProc->delayChkptFlags & DELAY_CHKPT_START) == 0);
6563 161362 : START_CRIT_SECTION();
6564 161362 : MyProc->delayChkptFlags |= DELAY_CHKPT_START;
6565 :
6566 : /* XLOG stuff */
6567 161362 : if (RelationNeedsWAL(relation))
6568 : {
6569 : xl_heap_inplace xlrec;
6570 : PGAlignedBlock copied_buffer;
6571 161346 : char *origdata = (char *) BufferGetBlock(buffer);
6572 161346 : Page page = BufferGetPage(buffer);
6573 161346 : uint16 lower = ((PageHeader) page)->pd_lower;
6574 161346 : uint16 upper = ((PageHeader) page)->pd_upper;
6575 : uintptr_t dst_offset_in_block;
6576 : RelFileLocator rlocator;
6577 : ForkNumber forkno;
6578 : BlockNumber blkno;
6579 : XLogRecPtr recptr;
6580 :
6581 161346 : xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
6582 161346 : xlrec.dbId = MyDatabaseId;
6583 161346 : xlrec.tsId = MyDatabaseTableSpace;
6584 161346 : xlrec.relcacheInitFileInval = RelcacheInitFileInval;
6585 161346 : xlrec.nmsgs = nmsgs;
6586 :
6587 161346 : XLogBeginInsert();
6588 161346 : XLogRegisterData(&xlrec, MinSizeOfHeapInplace);
6589 161346 : if (nmsgs != 0)
6590 69314 : XLogRegisterData(invalMessages,
6591 : nmsgs * sizeof(SharedInvalidationMessage));
6592 :
6593 : /* register block matching what buffer will look like after changes */
6594 161346 : memcpy(copied_buffer.data, origdata, lower);
6595 161346 : memcpy(copied_buffer.data + upper, origdata + upper, BLCKSZ - upper);
6596 161346 : dst_offset_in_block = dst - origdata;
6597 161346 : memcpy(copied_buffer.data + dst_offset_in_block, src, newlen);
6598 161346 : BufferGetTag(buffer, &rlocator, &forkno, &blkno);
6599 : Assert(forkno == MAIN_FORKNUM);
6600 161346 : XLogRegisterBlock(0, &rlocator, forkno, blkno, copied_buffer.data,
6601 : REGBUF_STANDARD);
6602 161346 : XLogRegisterBufData(0, src, newlen);
6603 :
6604 : /* inplace updates aren't decoded atm, don't log the origin */
6605 :
6606 161346 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_INPLACE);
6607 :
6608 161346 : PageSetLSN(page, recptr);
6609 : }
6610 :
6611 161362 : memcpy(dst, src, newlen);
6612 :
6613 161362 : MarkBufferDirty(buffer);
6614 :
6615 161362 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6616 :
6617 : /*
6618 : * Send invalidations to shared queue. SearchSysCacheLocked1() assumes we
6619 : * do this before UnlockTuple().
6620 : *
6621 : * If we're mutating a tuple visible only to this transaction, there's an
6622 : * equivalent transactional inval from the action that created the tuple,
6623 : * and this inval is superfluous.
6624 : */
6625 161362 : AtInplace_Inval();
6626 :
6627 161362 : MyProc->delayChkptFlags &= ~DELAY_CHKPT_START;
6628 161362 : END_CRIT_SECTION();
6629 161362 : UnlockTuple(relation, &tuple->t_self, InplaceUpdateTupleLock);
6630 :
6631 161362 : AcceptInvalidationMessages(); /* local processing of just-sent inval */
6632 :
6633 : /*
6634 : * Queue a transactional inval. The immediate invalidation we just sent
6635 : * is the only one known to be necessary. To reduce risk from the
6636 : * transition to immediate invalidation, continue sending a transactional
6637 : * invalidation like we've long done. Third-party code might rely on it.
6638 : */
6639 161362 : if (!IsBootstrapProcessingMode())
6640 132062 : CacheInvalidateHeapTuple(relation, tuple, NULL);
6641 161362 : }
6642 :
6643 : /*
6644 : * heap_inplace_unlock - reverse of heap_inplace_lock
6645 : */
6646 : void
6647 254812 : heap_inplace_unlock(Relation relation,
6648 : HeapTuple oldtup, Buffer buffer)
6649 : {
6650 254812 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6651 254812 : UnlockTuple(relation, &oldtup->t_self, InplaceUpdateTupleLock);
6652 254812 : ForgetInplace_Inval();
6653 254812 : }
6654 :
6655 : #define FRM_NOOP 0x0001
6656 : #define FRM_INVALIDATE_XMAX 0x0002
6657 : #define FRM_RETURN_IS_XID 0x0004
6658 : #define FRM_RETURN_IS_MULTI 0x0008
6659 : #define FRM_MARK_COMMITTED 0x0010
6660 :
6661 : /*
6662 : * FreezeMultiXactId
6663 : * Determine what to do during freezing when a tuple is marked by a
6664 : * MultiXactId.
6665 : *
6666 : * "flags" is an output value; it's used to tell caller what to do on return.
6667 : * "pagefrz" is an input/output value, used to manage page level freezing.
6668 : *
6669 : * Possible values that we can set in "flags":
6670 : * FRM_NOOP
6671 : * don't do anything -- keep existing Xmax
6672 : * FRM_INVALIDATE_XMAX
6673 : * mark Xmax as InvalidTransactionId and set XMAX_INVALID flag.
6674 : * FRM_RETURN_IS_XID
6675 : * The Xid return value is a single update Xid to set as xmax.
6676 : * FRM_MARK_COMMITTED
6677 : * Xmax can be marked as HEAP_XMAX_COMMITTED
6678 : * FRM_RETURN_IS_MULTI
6679 : * The return value is a new MultiXactId to set as new Xmax.
6680 : * (caller must obtain proper infomask bits using GetMultiXactIdHintBits)
6681 : *
6682 : * Caller delegates control of page freezing to us. In practice we always
6683 : * force freezing of caller's page unless FRM_NOOP processing is indicated.
6684 : * We help caller ensure that XIDs < FreezeLimit and MXIDs < MultiXactCutoff
6685 : * can never be left behind. We freely choose when and how to process each
6686 : * Multi, without ever violating the cutoff postconditions for freezing.
6687 : *
6688 : * It's useful to remove Multis on a proactive timeline (relative to freezing
6689 : * XIDs) to keep MultiXact member SLRU buffer misses to a minimum. It can also
6690 : * be cheaper in the short run, for us, since we too can avoid SLRU buffer
6691 : * misses through eager processing.
6692 : *
6693 : * NB: Creates a _new_ MultiXactId when FRM_RETURN_IS_MULTI is set, though only
6694 : * when FreezeLimit and/or MultiXactCutoff cutoffs leave us with no choice.
6695 : * This can usually be put off, which is usually enough to avoid it altogether.
6696 : * Allocating new multis during VACUUM should be avoided on general principle;
6697 : * only VACUUM can advance relminmxid, so allocating new Multis here comes with
6698 : * its own special risks.
6699 : *
6700 : * NB: Caller must maintain "no freeze" NewRelfrozenXid/NewRelminMxid trackers
6701 : * using heap_tuple_should_freeze when we haven't forced page-level freezing.
6702 : *
6703 : * NB: Caller should avoid needlessly calling heap_tuple_should_freeze when we
6704 : * have already forced page-level freezing, since that might incur the same
6705 : * SLRU buffer misses that we specifically intended to avoid by freezing.
6706 : */
6707 : static TransactionId
6708 14 : FreezeMultiXactId(MultiXactId multi, uint16 t_infomask,
6709 : const struct VacuumCutoffs *cutoffs, uint16 *flags,
6710 : HeapPageFreeze *pagefrz)
6711 : {
6712 : TransactionId newxmax;
6713 : MultiXactMember *members;
6714 : int nmembers;
6715 : bool need_replace;
6716 : int nnewmembers;
6717 : MultiXactMember *newmembers;
6718 : bool has_lockers;
6719 : TransactionId update_xid;
6720 : bool update_committed;
6721 : TransactionId FreezePageRelfrozenXid;
6722 :
6723 14 : *flags = 0;
6724 :
6725 : /* We should only be called in Multis */
6726 : Assert(t_infomask & HEAP_XMAX_IS_MULTI);
6727 :
6728 28 : if (!MultiXactIdIsValid(multi) ||
6729 14 : HEAP_LOCKED_UPGRADED(t_infomask))
6730 : {
6731 0 : *flags |= FRM_INVALIDATE_XMAX;
6732 0 : pagefrz->freeze_required = true;
6733 0 : return InvalidTransactionId;
6734 : }
6735 14 : else if (MultiXactIdPrecedes(multi, cutoffs->relminmxid))
6736 0 : ereport(ERROR,
6737 : (errcode(ERRCODE_DATA_CORRUPTED),
6738 : errmsg_internal("found multixact %u from before relminmxid %u",
6739 : multi, cutoffs->relminmxid)));
6740 14 : else if (MultiXactIdPrecedes(multi, cutoffs->OldestMxact))
6741 : {
6742 : TransactionId update_xact;
6743 :
6744 : /*
6745 : * This old multi cannot possibly have members still running, but
6746 : * verify just in case. If it was a locker only, it can be removed
6747 : * without any further consideration; but if it contained an update,
6748 : * we might need to preserve it.
6749 : */
6750 10 : if (MultiXactIdIsRunning(multi,
6751 10 : HEAP_XMAX_IS_LOCKED_ONLY(t_infomask)))
6752 0 : ereport(ERROR,
6753 : (errcode(ERRCODE_DATA_CORRUPTED),
6754 : errmsg_internal("multixact %u from before multi freeze cutoff %u found to be still running",
6755 : multi, cutoffs->OldestMxact)));
6756 :
6757 10 : if (HEAP_XMAX_IS_LOCKED_ONLY(t_infomask))
6758 : {
6759 10 : *flags |= FRM_INVALIDATE_XMAX;
6760 10 : pagefrz->freeze_required = true;
6761 10 : return InvalidTransactionId;
6762 : }
6763 :
6764 : /* replace multi with single XID for its updater? */
6765 0 : update_xact = MultiXactIdGetUpdateXid(multi, t_infomask);
6766 0 : if (TransactionIdPrecedes(update_xact, cutoffs->relfrozenxid))
6767 0 : ereport(ERROR,
6768 : (errcode(ERRCODE_DATA_CORRUPTED),
6769 : errmsg_internal("multixact %u contains update XID %u from before relfrozenxid %u",
6770 : multi, update_xact,
6771 : cutoffs->relfrozenxid)));
6772 0 : else if (TransactionIdPrecedes(update_xact, cutoffs->OldestXmin))
6773 : {
6774 : /*
6775 : * Updater XID has to have aborted (otherwise the tuple would have
6776 : * been pruned away instead, since updater XID is < OldestXmin).
6777 : * Just remove xmax.
6778 : */
6779 0 : if (TransactionIdDidCommit(update_xact))
6780 0 : ereport(ERROR,
6781 : (errcode(ERRCODE_DATA_CORRUPTED),
6782 : errmsg_internal("multixact %u contains committed update XID %u from before removable cutoff %u",
6783 : multi, update_xact,
6784 : cutoffs->OldestXmin)));
6785 0 : *flags |= FRM_INVALIDATE_XMAX;
6786 0 : pagefrz->freeze_required = true;
6787 0 : return InvalidTransactionId;
6788 : }
6789 :
6790 : /* Have to keep updater XID as new xmax */
6791 0 : *flags |= FRM_RETURN_IS_XID;
6792 0 : pagefrz->freeze_required = true;
6793 0 : return update_xact;
6794 : }
6795 :
6796 : /*
6797 : * Some member(s) of this Multi may be below FreezeLimit xid cutoff, so we
6798 : * need to walk the whole members array to figure out what to do, if
6799 : * anything.
6800 : */
6801 : nmembers =
6802 4 : GetMultiXactIdMembers(multi, &members, false,
6803 4 : HEAP_XMAX_IS_LOCKED_ONLY(t_infomask));
6804 4 : if (nmembers <= 0)
6805 : {
6806 : /* Nothing worth keeping */
6807 0 : *flags |= FRM_INVALIDATE_XMAX;
6808 0 : pagefrz->freeze_required = true;
6809 0 : return InvalidTransactionId;
6810 : }
6811 :
6812 : /*
6813 : * The FRM_NOOP case is the only case where we might need to ratchet back
6814 : * FreezePageRelfrozenXid or FreezePageRelminMxid. It is also the only
6815 : * case where our caller might ratchet back its NoFreezePageRelfrozenXid
6816 : * or NoFreezePageRelminMxid "no freeze" trackers to deal with a multi.
6817 : * FRM_NOOP handling should result in the NewRelfrozenXid/NewRelminMxid
6818 : * trackers managed by VACUUM being ratcheting back by xmax to the degree
6819 : * required to make it safe to leave xmax undisturbed, independent of
6820 : * whether or not page freezing is triggered somewhere else.
6821 : *
6822 : * Our policy is to force freezing in every case other than FRM_NOOP,
6823 : * which obviates the need to maintain either set of trackers, anywhere.
6824 : * Every other case will reliably execute a freeze plan for xmax that
6825 : * either replaces xmax with an XID/MXID >= OldestXmin/OldestMxact, or
6826 : * sets xmax to an InvalidTransactionId XID, rendering xmax fully frozen.
6827 : * (VACUUM's NewRelfrozenXid/NewRelminMxid trackers are initialized with
6828 : * OldestXmin/OldestMxact, so later values never need to be tracked here.)
6829 : */
6830 4 : need_replace = false;
6831 4 : FreezePageRelfrozenXid = pagefrz->FreezePageRelfrozenXid;
6832 8 : for (int i = 0; i < nmembers; i++)
6833 : {
6834 6 : TransactionId xid = members[i].xid;
6835 :
6836 : Assert(!TransactionIdPrecedes(xid, cutoffs->relfrozenxid));
6837 :
6838 6 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
6839 : {
6840 : /* Can't violate the FreezeLimit postcondition */
6841 2 : need_replace = true;
6842 2 : break;
6843 : }
6844 4 : if (TransactionIdPrecedes(xid, FreezePageRelfrozenXid))
6845 0 : FreezePageRelfrozenXid = xid;
6846 : }
6847 :
6848 : /* Can't violate the MultiXactCutoff postcondition, either */
6849 4 : if (!need_replace)
6850 2 : need_replace = MultiXactIdPrecedes(multi, cutoffs->MultiXactCutoff);
6851 :
6852 4 : if (!need_replace)
6853 : {
6854 : /*
6855 : * vacuumlazy.c might ratchet back NewRelminMxid, NewRelfrozenXid, or
6856 : * both together to make it safe to retain this particular multi after
6857 : * freezing its page
6858 : */
6859 2 : *flags |= FRM_NOOP;
6860 2 : pagefrz->FreezePageRelfrozenXid = FreezePageRelfrozenXid;
6861 2 : if (MultiXactIdPrecedes(multi, pagefrz->FreezePageRelminMxid))
6862 0 : pagefrz->FreezePageRelminMxid = multi;
6863 2 : pfree(members);
6864 2 : return multi;
6865 : }
6866 :
6867 : /*
6868 : * Do a more thorough second pass over the multi to figure out which
6869 : * member XIDs actually need to be kept. Checking the precise status of
6870 : * individual members might even show that we don't need to keep anything.
6871 : * That is quite possible even though the Multi must be >= OldestMxact,
6872 : * since our second pass only keeps member XIDs when it's truly necessary;
6873 : * even member XIDs >= OldestXmin often won't be kept by second pass.
6874 : */
6875 2 : nnewmembers = 0;
6876 2 : newmembers = palloc(sizeof(MultiXactMember) * nmembers);
6877 2 : has_lockers = false;
6878 2 : update_xid = InvalidTransactionId;
6879 2 : update_committed = false;
6880 :
6881 : /*
6882 : * Determine whether to keep each member xid, or to ignore it instead
6883 : */
6884 6 : for (int i = 0; i < nmembers; i++)
6885 : {
6886 4 : TransactionId xid = members[i].xid;
6887 4 : MultiXactStatus mstatus = members[i].status;
6888 :
6889 : Assert(!TransactionIdPrecedes(xid, cutoffs->relfrozenxid));
6890 :
6891 4 : if (!ISUPDATE_from_mxstatus(mstatus))
6892 : {
6893 : /*
6894 : * Locker XID (not updater XID). We only keep lockers that are
6895 : * still running.
6896 : */
6897 8 : if (TransactionIdIsCurrentTransactionId(xid) ||
6898 4 : TransactionIdIsInProgress(xid))
6899 : {
6900 2 : if (TransactionIdPrecedes(xid, cutoffs->OldestXmin))
6901 0 : ereport(ERROR,
6902 : (errcode(ERRCODE_DATA_CORRUPTED),
6903 : errmsg_internal("multixact %u contains running locker XID %u from before removable cutoff %u",
6904 : multi, xid,
6905 : cutoffs->OldestXmin)));
6906 2 : newmembers[nnewmembers++] = members[i];
6907 2 : has_lockers = true;
6908 : }
6909 :
6910 4 : continue;
6911 : }
6912 :
6913 : /*
6914 : * Updater XID (not locker XID). Should we keep it?
6915 : *
6916 : * Since the tuple wasn't totally removed when vacuum pruned, the
6917 : * update Xid cannot possibly be older than OldestXmin cutoff unless
6918 : * the updater XID aborted. If the updater transaction is known
6919 : * aborted or crashed then it's okay to ignore it, otherwise not.
6920 : *
6921 : * In any case the Multi should never contain two updaters, whatever
6922 : * their individual commit status. Check for that first, in passing.
6923 : */
6924 0 : if (TransactionIdIsValid(update_xid))
6925 0 : ereport(ERROR,
6926 : (errcode(ERRCODE_DATA_CORRUPTED),
6927 : errmsg_internal("multixact %u has two or more updating members",
6928 : multi),
6929 : errdetail_internal("First updater XID=%u second updater XID=%u.",
6930 : update_xid, xid)));
6931 :
6932 : /*
6933 : * As with all tuple visibility routines, it's critical to test
6934 : * TransactionIdIsInProgress before TransactionIdDidCommit, because of
6935 : * race conditions explained in detail in heapam_visibility.c.
6936 : */
6937 0 : if (TransactionIdIsCurrentTransactionId(xid) ||
6938 0 : TransactionIdIsInProgress(xid))
6939 0 : update_xid = xid;
6940 0 : else if (TransactionIdDidCommit(xid))
6941 : {
6942 : /*
6943 : * The transaction committed, so we can tell caller to set
6944 : * HEAP_XMAX_COMMITTED. (We can only do this because we know the
6945 : * transaction is not running.)
6946 : */
6947 0 : update_committed = true;
6948 0 : update_xid = xid;
6949 : }
6950 : else
6951 : {
6952 : /*
6953 : * Not in progress, not committed -- must be aborted or crashed;
6954 : * we can ignore it.
6955 : */
6956 0 : continue;
6957 : }
6958 :
6959 : /*
6960 : * We determined that updater must be kept -- add it to pending new
6961 : * members list
6962 : */
6963 0 : if (TransactionIdPrecedes(xid, cutoffs->OldestXmin))
6964 0 : ereport(ERROR,
6965 : (errcode(ERRCODE_DATA_CORRUPTED),
6966 : errmsg_internal("multixact %u contains committed update XID %u from before removable cutoff %u",
6967 : multi, xid, cutoffs->OldestXmin)));
6968 0 : newmembers[nnewmembers++] = members[i];
6969 : }
6970 :
6971 2 : pfree(members);
6972 :
6973 : /*
6974 : * Determine what to do with caller's multi based on information gathered
6975 : * during our second pass
6976 : */
6977 2 : if (nnewmembers == 0)
6978 : {
6979 : /* Nothing worth keeping */
6980 0 : *flags |= FRM_INVALIDATE_XMAX;
6981 0 : newxmax = InvalidTransactionId;
6982 : }
6983 2 : else if (TransactionIdIsValid(update_xid) && !has_lockers)
6984 : {
6985 : /*
6986 : * If there's a single member and it's an update, pass it back alone
6987 : * without creating a new Multi. (XXX we could do this when there's a
6988 : * single remaining locker, too, but that would complicate the API too
6989 : * much; moreover, the case with the single updater is more
6990 : * interesting, because those are longer-lived.)
6991 : */
6992 : Assert(nnewmembers == 1);
6993 0 : *flags |= FRM_RETURN_IS_XID;
6994 0 : if (update_committed)
6995 0 : *flags |= FRM_MARK_COMMITTED;
6996 0 : newxmax = update_xid;
6997 : }
6998 : else
6999 : {
7000 : /*
7001 : * Create a new multixact with the surviving members of the previous
7002 : * one, to set as new Xmax in the tuple
7003 : */
7004 2 : newxmax = MultiXactIdCreateFromMembers(nnewmembers, newmembers);
7005 2 : *flags |= FRM_RETURN_IS_MULTI;
7006 : }
7007 :
7008 2 : pfree(newmembers);
7009 :
7010 2 : pagefrz->freeze_required = true;
7011 2 : return newxmax;
7012 : }
7013 :
7014 : /*
7015 : * heap_prepare_freeze_tuple
7016 : *
7017 : * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
7018 : * are older than the OldestXmin and/or OldestMxact freeze cutoffs. If so,
7019 : * setup enough state (in the *frz output argument) to enable caller to
7020 : * process this tuple as part of freezing its page, and return true. Return
7021 : * false if nothing can be changed about the tuple right now.
7022 : *
7023 : * Also sets *totally_frozen to true if the tuple will be totally frozen once
7024 : * caller executes returned freeze plan (or if the tuple was already totally
7025 : * frozen by an earlier VACUUM). This indicates that there are no remaining
7026 : * XIDs or MultiXactIds that will need to be processed by a future VACUUM.
7027 : *
7028 : * VACUUM caller must assemble HeapTupleFreeze freeze plan entries for every
7029 : * tuple that we returned true for, and then execute freezing. Caller must
7030 : * initialize pagefrz fields for page as a whole before first call here for
7031 : * each heap page.
7032 : *
7033 : * VACUUM caller decides on whether or not to freeze the page as a whole.
7034 : * We'll often prepare freeze plans for a page that caller just discards.
7035 : * However, VACUUM doesn't always get to make a choice; it must freeze when
7036 : * pagefrz.freeze_required is set, to ensure that any XIDs < FreezeLimit (and
7037 : * MXIDs < MultiXactCutoff) can never be left behind. We help to make sure
7038 : * that VACUUM always follows that rule.
7039 : *
7040 : * We sometimes force freezing of xmax MultiXactId values long before it is
7041 : * strictly necessary to do so just to ensure the FreezeLimit postcondition.
7042 : * It's worth processing MultiXactIds proactively when it is cheap to do so,
7043 : * and it's convenient to make that happen by piggy-backing it on the "force
7044 : * freezing" mechanism. Conversely, we sometimes delay freezing MultiXactIds
7045 : * because it is expensive right now (though only when it's still possible to
7046 : * do so without violating the FreezeLimit/MultiXactCutoff postcondition).
7047 : *
7048 : * It is assumed that the caller has checked the tuple with
7049 : * HeapTupleSatisfiesVacuum() and determined that it is not HEAPTUPLE_DEAD
7050 : * (else we should be removing the tuple, not freezing it).
7051 : *
7052 : * NB: This function has side effects: it might allocate a new MultiXactId.
7053 : * It will be set as tuple's new xmax when our *frz output is processed within
7054 : * heap_execute_freeze_tuple later on. If the tuple is in a shared buffer
7055 : * then caller had better have an exclusive lock on it already.
7056 : */
7057 : bool
7058 52459928 : heap_prepare_freeze_tuple(HeapTupleHeader tuple,
7059 : const struct VacuumCutoffs *cutoffs,
7060 : HeapPageFreeze *pagefrz,
7061 : HeapTupleFreeze *frz, bool *totally_frozen)
7062 : {
7063 52459928 : bool xmin_already_frozen = false,
7064 52459928 : xmax_already_frozen = false;
7065 52459928 : bool freeze_xmin = false,
7066 52459928 : replace_xvac = false,
7067 52459928 : replace_xmax = false,
7068 52459928 : freeze_xmax = false;
7069 : TransactionId xid;
7070 :
7071 52459928 : frz->xmax = HeapTupleHeaderGetRawXmax(tuple);
7072 52459928 : frz->t_infomask2 = tuple->t_infomask2;
7073 52459928 : frz->t_infomask = tuple->t_infomask;
7074 52459928 : frz->frzflags = 0;
7075 52459928 : frz->checkflags = 0;
7076 :
7077 : /*
7078 : * Process xmin, while keeping track of whether it's already frozen, or
7079 : * will become frozen iff our freeze plan is executed by caller (could be
7080 : * neither).
7081 : */
7082 52459928 : xid = HeapTupleHeaderGetXmin(tuple);
7083 52459928 : if (!TransactionIdIsNormal(xid))
7084 46530426 : xmin_already_frozen = true;
7085 : else
7086 : {
7087 5929502 : if (TransactionIdPrecedes(xid, cutoffs->relfrozenxid))
7088 0 : ereport(ERROR,
7089 : (errcode(ERRCODE_DATA_CORRUPTED),
7090 : errmsg_internal("found xmin %u from before relfrozenxid %u",
7091 : xid, cutoffs->relfrozenxid)));
7092 :
7093 : /* Will set freeze_xmin flags in freeze plan below */
7094 5929502 : freeze_xmin = TransactionIdPrecedes(xid, cutoffs->OldestXmin);
7095 :
7096 : /* Verify that xmin committed if and when freeze plan is executed */
7097 5929502 : if (freeze_xmin)
7098 4657900 : frz->checkflags |= HEAP_FREEZE_CHECK_XMIN_COMMITTED;
7099 : }
7100 :
7101 : /*
7102 : * Old-style VACUUM FULL is gone, but we have to process xvac for as long
7103 : * as we support having MOVED_OFF/MOVED_IN tuples in the database
7104 : */
7105 52459928 : xid = HeapTupleHeaderGetXvac(tuple);
7106 52459928 : if (TransactionIdIsNormal(xid))
7107 : {
7108 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7109 : Assert(TransactionIdPrecedes(xid, cutoffs->OldestXmin));
7110 :
7111 : /*
7112 : * For Xvac, we always freeze proactively. This allows totally_frozen
7113 : * tracking to ignore xvac.
7114 : */
7115 0 : replace_xvac = pagefrz->freeze_required = true;
7116 :
7117 : /* Will set replace_xvac flags in freeze plan below */
7118 : }
7119 :
7120 : /* Now process xmax */
7121 52459928 : xid = frz->xmax;
7122 52459928 : if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
7123 : {
7124 : /* Raw xmax is a MultiXactId */
7125 : TransactionId newxmax;
7126 : uint16 flags;
7127 :
7128 : /*
7129 : * We will either remove xmax completely (in the "freeze_xmax" path),
7130 : * process xmax by replacing it (in the "replace_xmax" path), or
7131 : * perform no-op xmax processing. The only constraint is that the
7132 : * FreezeLimit/MultiXactCutoff postcondition must never be violated.
7133 : */
7134 14 : newxmax = FreezeMultiXactId(xid, tuple->t_infomask, cutoffs,
7135 : &flags, pagefrz);
7136 :
7137 14 : if (flags & FRM_NOOP)
7138 : {
7139 : /*
7140 : * xmax is a MultiXactId, and nothing about it changes for now.
7141 : * This is the only case where 'freeze_required' won't have been
7142 : * set for us by FreezeMultiXactId, as well as the only case where
7143 : * neither freeze_xmax nor replace_xmax are set (given a multi).
7144 : *
7145 : * This is a no-op, but the call to FreezeMultiXactId might have
7146 : * ratcheted back NewRelfrozenXid and/or NewRelminMxid trackers
7147 : * for us (the "freeze page" variants, specifically). That'll
7148 : * make it safe for our caller to freeze the page later on, while
7149 : * leaving this particular xmax undisturbed.
7150 : *
7151 : * FreezeMultiXactId is _not_ responsible for the "no freeze"
7152 : * NewRelfrozenXid/NewRelminMxid trackers, though -- that's our
7153 : * job. A call to heap_tuple_should_freeze for this same tuple
7154 : * will take place below if 'freeze_required' isn't set already.
7155 : * (This repeats work from FreezeMultiXactId, but allows "no
7156 : * freeze" tracker maintenance to happen in only one place.)
7157 : */
7158 : Assert(!MultiXactIdPrecedes(newxmax, cutoffs->MultiXactCutoff));
7159 : Assert(MultiXactIdIsValid(newxmax) && xid == newxmax);
7160 : }
7161 12 : else if (flags & FRM_RETURN_IS_XID)
7162 : {
7163 : /*
7164 : * xmax will become an updater Xid (original MultiXact's updater
7165 : * member Xid will be carried forward as a simple Xid in Xmax).
7166 : */
7167 : Assert(!TransactionIdPrecedes(newxmax, cutoffs->OldestXmin));
7168 :
7169 : /*
7170 : * NB -- some of these transformations are only valid because we
7171 : * know the return Xid is a tuple updater (i.e. not merely a
7172 : * locker.) Also note that the only reason we don't explicitly
7173 : * worry about HEAP_KEYS_UPDATED is because it lives in
7174 : * t_infomask2 rather than t_infomask.
7175 : */
7176 0 : frz->t_infomask &= ~HEAP_XMAX_BITS;
7177 0 : frz->xmax = newxmax;
7178 0 : if (flags & FRM_MARK_COMMITTED)
7179 0 : frz->t_infomask |= HEAP_XMAX_COMMITTED;
7180 0 : replace_xmax = true;
7181 : }
7182 12 : else if (flags & FRM_RETURN_IS_MULTI)
7183 : {
7184 : uint16 newbits;
7185 : uint16 newbits2;
7186 :
7187 : /*
7188 : * xmax is an old MultiXactId that we have to replace with a new
7189 : * MultiXactId, to carry forward two or more original member XIDs.
7190 : */
7191 : Assert(!MultiXactIdPrecedes(newxmax, cutoffs->OldestMxact));
7192 :
7193 : /*
7194 : * We can't use GetMultiXactIdHintBits directly on the new multi
7195 : * here; that routine initializes the masks to all zeroes, which
7196 : * would lose other bits we need. Doing it this way ensures all
7197 : * unrelated bits remain untouched.
7198 : */
7199 2 : frz->t_infomask &= ~HEAP_XMAX_BITS;
7200 2 : frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
7201 2 : GetMultiXactIdHintBits(newxmax, &newbits, &newbits2);
7202 2 : frz->t_infomask |= newbits;
7203 2 : frz->t_infomask2 |= newbits2;
7204 2 : frz->xmax = newxmax;
7205 2 : replace_xmax = true;
7206 : }
7207 : else
7208 : {
7209 : /*
7210 : * Freeze plan for tuple "freezes xmax" in the strictest sense:
7211 : * it'll leave nothing in xmax (neither an Xid nor a MultiXactId).
7212 : */
7213 : Assert(flags & FRM_INVALIDATE_XMAX);
7214 : Assert(!TransactionIdIsValid(newxmax));
7215 :
7216 : /* Will set freeze_xmax flags in freeze plan below */
7217 10 : freeze_xmax = true;
7218 : }
7219 :
7220 : /* MultiXactId processing forces freezing (barring FRM_NOOP case) */
7221 : Assert(pagefrz->freeze_required || (!freeze_xmax && !replace_xmax));
7222 : }
7223 52459914 : else if (TransactionIdIsNormal(xid))
7224 : {
7225 : /* Raw xmax is normal XID */
7226 17419258 : if (TransactionIdPrecedes(xid, cutoffs->relfrozenxid))
7227 0 : ereport(ERROR,
7228 : (errcode(ERRCODE_DATA_CORRUPTED),
7229 : errmsg_internal("found xmax %u from before relfrozenxid %u",
7230 : xid, cutoffs->relfrozenxid)));
7231 :
7232 : /* Will set freeze_xmax flags in freeze plan below */
7233 17419258 : freeze_xmax = TransactionIdPrecedes(xid, cutoffs->OldestXmin);
7234 :
7235 : /*
7236 : * Verify that xmax aborted if and when freeze plan is executed,
7237 : * provided it's from an update. (A lock-only xmax can be removed
7238 : * independent of this, since the lock is released at xact end.)
7239 : */
7240 17419258 : if (freeze_xmax && !HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask))
7241 1476 : frz->checkflags |= HEAP_FREEZE_CHECK_XMAX_ABORTED;
7242 : }
7243 35040656 : else if (!TransactionIdIsValid(xid))
7244 : {
7245 : /* Raw xmax is InvalidTransactionId XID */
7246 : Assert((tuple->t_infomask & HEAP_XMAX_IS_MULTI) == 0);
7247 35040656 : xmax_already_frozen = true;
7248 : }
7249 : else
7250 0 : ereport(ERROR,
7251 : (errcode(ERRCODE_DATA_CORRUPTED),
7252 : errmsg_internal("found raw xmax %u (infomask 0x%04x) not invalid and not multi",
7253 : xid, tuple->t_infomask)));
7254 :
7255 52459928 : if (freeze_xmin)
7256 : {
7257 : Assert(!xmin_already_frozen);
7258 :
7259 4657900 : frz->t_infomask |= HEAP_XMIN_FROZEN;
7260 : }
7261 52459928 : if (replace_xvac)
7262 : {
7263 : /*
7264 : * If a MOVED_OFF tuple is not dead, the xvac transaction must have
7265 : * failed; whereas a non-dead MOVED_IN tuple must mean the xvac
7266 : * transaction succeeded.
7267 : */
7268 : Assert(pagefrz->freeze_required);
7269 0 : if (tuple->t_infomask & HEAP_MOVED_OFF)
7270 0 : frz->frzflags |= XLH_INVALID_XVAC;
7271 : else
7272 0 : frz->frzflags |= XLH_FREEZE_XVAC;
7273 : }
7274 : if (replace_xmax)
7275 : {
7276 : Assert(!xmax_already_frozen && !freeze_xmax);
7277 : Assert(pagefrz->freeze_required);
7278 :
7279 : /* Already set replace_xmax flags in freeze plan earlier */
7280 : }
7281 52459928 : if (freeze_xmax)
7282 : {
7283 : Assert(!xmax_already_frozen && !replace_xmax);
7284 :
7285 3226 : frz->xmax = InvalidTransactionId;
7286 :
7287 : /*
7288 : * The tuple might be marked either XMAX_INVALID or XMAX_COMMITTED +
7289 : * LOCKED. Normalize to INVALID just to be sure no one gets confused.
7290 : * Also get rid of the HEAP_KEYS_UPDATED bit.
7291 : */
7292 3226 : frz->t_infomask &= ~HEAP_XMAX_BITS;
7293 3226 : frz->t_infomask |= HEAP_XMAX_INVALID;
7294 3226 : frz->t_infomask2 &= ~HEAP_HOT_UPDATED;
7295 3226 : frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
7296 : }
7297 :
7298 : /*
7299 : * Determine if this tuple is already totally frozen, or will become
7300 : * totally frozen (provided caller executes freeze plans for the page)
7301 : */
7302 103645028 : *totally_frozen = ((freeze_xmin || xmin_already_frozen) &&
7303 51185100 : (freeze_xmax || xmax_already_frozen));
7304 :
7305 52459928 : if (!pagefrz->freeze_required && !(xmin_already_frozen &&
7306 : xmax_already_frozen))
7307 : {
7308 : /*
7309 : * So far no previous tuple from the page made freezing mandatory.
7310 : * Does this tuple force caller to freeze the entire page?
7311 : */
7312 20183830 : pagefrz->freeze_required =
7313 20183830 : heap_tuple_should_freeze(tuple, cutoffs,
7314 : &pagefrz->NoFreezePageRelfrozenXid,
7315 : &pagefrz->NoFreezePageRelminMxid);
7316 : }
7317 :
7318 : /* Tell caller if this tuple has a usable freeze plan set in *frz */
7319 52459928 : return freeze_xmin || replace_xvac || replace_xmax || freeze_xmax;
7320 : }
7321 :
7322 : /*
7323 : * Perform xmin/xmax XID status sanity checks before actually executing freeze
7324 : * plans.
7325 : *
7326 : * heap_prepare_freeze_tuple doesn't perform these checks directly because
7327 : * pg_xact lookups are relatively expensive. They shouldn't be repeated by
7328 : * successive VACUUMs that each decide against freezing the same page.
7329 : */
7330 : void
7331 45052 : heap_pre_freeze_checks(Buffer buffer,
7332 : HeapTupleFreeze *tuples, int ntuples)
7333 : {
7334 45052 : Page page = BufferGetPage(buffer);
7335 :
7336 1938064 : for (int i = 0; i < ntuples; i++)
7337 : {
7338 1893012 : HeapTupleFreeze *frz = tuples + i;
7339 1893012 : ItemId itemid = PageGetItemId(page, frz->offset);
7340 : HeapTupleHeader htup;
7341 :
7342 1893012 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
7343 :
7344 : /* Deliberately avoid relying on tuple hint bits here */
7345 1893012 : if (frz->checkflags & HEAP_FREEZE_CHECK_XMIN_COMMITTED)
7346 : {
7347 1893010 : TransactionId xmin = HeapTupleHeaderGetRawXmin(htup);
7348 :
7349 : Assert(!HeapTupleHeaderXminFrozen(htup));
7350 1893010 : if (unlikely(!TransactionIdDidCommit(xmin)))
7351 0 : ereport(ERROR,
7352 : (errcode(ERRCODE_DATA_CORRUPTED),
7353 : errmsg_internal("uncommitted xmin %u needs to be frozen",
7354 : xmin)));
7355 : }
7356 :
7357 : /*
7358 : * TransactionIdDidAbort won't work reliably in the presence of XIDs
7359 : * left behind by transactions that were in progress during a crash,
7360 : * so we can only check that xmax didn't commit
7361 : */
7362 1893012 : if (frz->checkflags & HEAP_FREEZE_CHECK_XMAX_ABORTED)
7363 : {
7364 358 : TransactionId xmax = HeapTupleHeaderGetRawXmax(htup);
7365 :
7366 : Assert(TransactionIdIsNormal(xmax));
7367 358 : if (unlikely(TransactionIdDidCommit(xmax)))
7368 0 : ereport(ERROR,
7369 : (errcode(ERRCODE_DATA_CORRUPTED),
7370 : errmsg_internal("cannot freeze committed xmax %u",
7371 : xmax)));
7372 : }
7373 : }
7374 45052 : }
7375 :
7376 : /*
7377 : * Helper which executes freezing of one or more heap tuples on a page on
7378 : * behalf of caller. Caller passes an array of tuple plans from
7379 : * heap_prepare_freeze_tuple. Caller must set 'offset' in each plan for us.
7380 : * Must be called in a critical section that also marks the buffer dirty and,
7381 : * if needed, emits WAL.
7382 : */
7383 : void
7384 45052 : heap_freeze_prepared_tuples(Buffer buffer, HeapTupleFreeze *tuples, int ntuples)
7385 : {
7386 45052 : Page page = BufferGetPage(buffer);
7387 :
7388 1938064 : for (int i = 0; i < ntuples; i++)
7389 : {
7390 1893012 : HeapTupleFreeze *frz = tuples + i;
7391 1893012 : ItemId itemid = PageGetItemId(page, frz->offset);
7392 : HeapTupleHeader htup;
7393 :
7394 1893012 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
7395 1893012 : heap_execute_freeze_tuple(htup, frz);
7396 : }
7397 45052 : }
7398 :
7399 : /*
7400 : * heap_freeze_tuple
7401 : * Freeze tuple in place, without WAL logging.
7402 : *
7403 : * Useful for callers like CLUSTER that perform their own WAL logging.
7404 : */
7405 : bool
7406 731946 : heap_freeze_tuple(HeapTupleHeader tuple,
7407 : TransactionId relfrozenxid, TransactionId relminmxid,
7408 : TransactionId FreezeLimit, TransactionId MultiXactCutoff)
7409 : {
7410 : HeapTupleFreeze frz;
7411 : bool do_freeze;
7412 : bool totally_frozen;
7413 : struct VacuumCutoffs cutoffs;
7414 : HeapPageFreeze pagefrz;
7415 :
7416 731946 : cutoffs.relfrozenxid = relfrozenxid;
7417 731946 : cutoffs.relminmxid = relminmxid;
7418 731946 : cutoffs.OldestXmin = FreezeLimit;
7419 731946 : cutoffs.OldestMxact = MultiXactCutoff;
7420 731946 : cutoffs.FreezeLimit = FreezeLimit;
7421 731946 : cutoffs.MultiXactCutoff = MultiXactCutoff;
7422 :
7423 731946 : pagefrz.freeze_required = true;
7424 731946 : pagefrz.FreezePageRelfrozenXid = FreezeLimit;
7425 731946 : pagefrz.FreezePageRelminMxid = MultiXactCutoff;
7426 731946 : pagefrz.NoFreezePageRelfrozenXid = FreezeLimit;
7427 731946 : pagefrz.NoFreezePageRelminMxid = MultiXactCutoff;
7428 :
7429 731946 : do_freeze = heap_prepare_freeze_tuple(tuple, &cutoffs,
7430 : &pagefrz, &frz, &totally_frozen);
7431 :
7432 : /*
7433 : * Note that because this is not a WAL-logged operation, we don't need to
7434 : * fill in the offset in the freeze record.
7435 : */
7436 :
7437 731946 : if (do_freeze)
7438 501504 : heap_execute_freeze_tuple(tuple, &frz);
7439 731946 : return do_freeze;
7440 : }
7441 :
7442 : /*
7443 : * For a given MultiXactId, return the hint bits that should be set in the
7444 : * tuple's infomask.
7445 : *
7446 : * Normally this should be called for a multixact that was just created, and
7447 : * so is on our local cache, so the GetMembers call is fast.
7448 : */
7449 : static void
7450 2396 : GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
7451 : uint16 *new_infomask2)
7452 : {
7453 : int nmembers;
7454 : MultiXactMember *members;
7455 : int i;
7456 2396 : uint16 bits = HEAP_XMAX_IS_MULTI;
7457 2396 : uint16 bits2 = 0;
7458 2396 : bool has_update = false;
7459 2396 : LockTupleMode strongest = LockTupleKeyShare;
7460 :
7461 : /*
7462 : * We only use this in multis we just created, so they cannot be values
7463 : * pre-pg_upgrade.
7464 : */
7465 2396 : nmembers = GetMultiXactIdMembers(multi, &members, false, false);
7466 :
7467 7344 : for (i = 0; i < nmembers; i++)
7468 : {
7469 : LockTupleMode mode;
7470 :
7471 : /*
7472 : * Remember the strongest lock mode held by any member of the
7473 : * multixact.
7474 : */
7475 4948 : mode = TUPLOCK_from_mxstatus(members[i].status);
7476 4948 : if (mode > strongest)
7477 1320 : strongest = mode;
7478 :
7479 : /* See what other bits we need */
7480 4948 : switch (members[i].status)
7481 : {
7482 4566 : case MultiXactStatusForKeyShare:
7483 : case MultiXactStatusForShare:
7484 : case MultiXactStatusForNoKeyUpdate:
7485 4566 : break;
7486 :
7487 104 : case MultiXactStatusForUpdate:
7488 104 : bits2 |= HEAP_KEYS_UPDATED;
7489 104 : break;
7490 :
7491 258 : case MultiXactStatusNoKeyUpdate:
7492 258 : has_update = true;
7493 258 : break;
7494 :
7495 20 : case MultiXactStatusUpdate:
7496 20 : bits2 |= HEAP_KEYS_UPDATED;
7497 20 : has_update = true;
7498 20 : break;
7499 : }
7500 : }
7501 :
7502 2396 : if (strongest == LockTupleExclusive ||
7503 : strongest == LockTupleNoKeyExclusive)
7504 438 : bits |= HEAP_XMAX_EXCL_LOCK;
7505 1958 : else if (strongest == LockTupleShare)
7506 876 : bits |= HEAP_XMAX_SHR_LOCK;
7507 1082 : else if (strongest == LockTupleKeyShare)
7508 1082 : bits |= HEAP_XMAX_KEYSHR_LOCK;
7509 :
7510 2396 : if (!has_update)
7511 2118 : bits |= HEAP_XMAX_LOCK_ONLY;
7512 :
7513 2396 : if (nmembers > 0)
7514 2396 : pfree(members);
7515 :
7516 2396 : *new_infomask = bits;
7517 2396 : *new_infomask2 = bits2;
7518 2396 : }
7519 :
7520 : /*
7521 : * MultiXactIdGetUpdateXid
7522 : *
7523 : * Given a multixact Xmax and corresponding infomask, which does not have the
7524 : * HEAP_XMAX_LOCK_ONLY bit set, obtain and return the Xid of the updating
7525 : * transaction.
7526 : *
7527 : * Caller is expected to check the status of the updating transaction, if
7528 : * necessary.
7529 : */
7530 : static TransactionId
7531 980 : MultiXactIdGetUpdateXid(TransactionId xmax, uint16 t_infomask)
7532 : {
7533 980 : TransactionId update_xact = InvalidTransactionId;
7534 : MultiXactMember *members;
7535 : int nmembers;
7536 :
7537 : Assert(!(t_infomask & HEAP_XMAX_LOCK_ONLY));
7538 : Assert(t_infomask & HEAP_XMAX_IS_MULTI);
7539 :
7540 : /*
7541 : * Since we know the LOCK_ONLY bit is not set, this cannot be a multi from
7542 : * pre-pg_upgrade.
7543 : */
7544 980 : nmembers = GetMultiXactIdMembers(xmax, &members, false, false);
7545 :
7546 980 : if (nmembers > 0)
7547 : {
7548 : int i;
7549 :
7550 2548 : for (i = 0; i < nmembers; i++)
7551 : {
7552 : /* Ignore lockers */
7553 2548 : if (!ISUPDATE_from_mxstatus(members[i].status))
7554 1568 : continue;
7555 :
7556 : /* there can be at most one updater */
7557 : Assert(update_xact == InvalidTransactionId);
7558 980 : update_xact = members[i].xid;
7559 : #ifndef USE_ASSERT_CHECKING
7560 :
7561 : /*
7562 : * in an assert-enabled build, walk the whole array to ensure
7563 : * there's no other updater.
7564 : */
7565 980 : break;
7566 : #endif
7567 : }
7568 :
7569 980 : pfree(members);
7570 : }
7571 :
7572 980 : return update_xact;
7573 : }
7574 :
7575 : /*
7576 : * HeapTupleGetUpdateXid
7577 : * As above, but use a HeapTupleHeader
7578 : *
7579 : * See also HeapTupleHeaderGetUpdateXid, which can be used without previously
7580 : * checking the hint bits.
7581 : */
7582 : TransactionId
7583 964 : HeapTupleGetUpdateXid(const HeapTupleHeaderData *tup)
7584 : {
7585 964 : return MultiXactIdGetUpdateXid(HeapTupleHeaderGetRawXmax(tup),
7586 964 : tup->t_infomask);
7587 : }
7588 :
7589 : /*
7590 : * Does the given multixact conflict with the current transaction grabbing a
7591 : * tuple lock of the given strength?
7592 : *
7593 : * The passed infomask pairs up with the given multixact in the tuple header.
7594 : *
7595 : * If current_is_member is not NULL, it is set to 'true' if the current
7596 : * transaction is a member of the given multixact.
7597 : */
7598 : static bool
7599 198 : DoesMultiXactIdConflict(MultiXactId multi, uint16 infomask,
7600 : LockTupleMode lockmode, bool *current_is_member)
7601 : {
7602 : int nmembers;
7603 : MultiXactMember *members;
7604 198 : bool result = false;
7605 198 : LOCKMODE wanted = tupleLockExtraInfo[lockmode].hwlock;
7606 :
7607 198 : if (HEAP_LOCKED_UPGRADED(infomask))
7608 0 : return false;
7609 :
7610 198 : nmembers = GetMultiXactIdMembers(multi, &members, false,
7611 198 : HEAP_XMAX_IS_LOCKED_ONLY(infomask));
7612 198 : if (nmembers >= 0)
7613 : {
7614 : int i;
7615 :
7616 618 : for (i = 0; i < nmembers; i++)
7617 : {
7618 : TransactionId memxid;
7619 : LOCKMODE memlockmode;
7620 :
7621 434 : if (result && (current_is_member == NULL || *current_is_member))
7622 : break;
7623 :
7624 420 : memlockmode = LOCKMODE_from_mxstatus(members[i].status);
7625 :
7626 : /* ignore members from current xact (but track their presence) */
7627 420 : memxid = members[i].xid;
7628 420 : if (TransactionIdIsCurrentTransactionId(memxid))
7629 : {
7630 184 : if (current_is_member != NULL)
7631 156 : *current_is_member = true;
7632 184 : continue;
7633 : }
7634 236 : else if (result)
7635 16 : continue;
7636 :
7637 : /* ignore members that don't conflict with the lock we want */
7638 220 : if (!DoLockModesConflict(memlockmode, wanted))
7639 142 : continue;
7640 :
7641 78 : if (ISUPDATE_from_mxstatus(members[i].status))
7642 : {
7643 : /* ignore aborted updaters */
7644 34 : if (TransactionIdDidAbort(memxid))
7645 2 : continue;
7646 : }
7647 : else
7648 : {
7649 : /* ignore lockers-only that are no longer in progress */
7650 44 : if (!TransactionIdIsInProgress(memxid))
7651 14 : continue;
7652 : }
7653 :
7654 : /*
7655 : * Whatever remains are either live lockers that conflict with our
7656 : * wanted lock, and updaters that are not aborted. Those conflict
7657 : * with what we want. Set up to return true, but keep going to
7658 : * look for the current transaction among the multixact members,
7659 : * if needed.
7660 : */
7661 62 : result = true;
7662 : }
7663 198 : pfree(members);
7664 : }
7665 :
7666 198 : return result;
7667 : }
7668 :
7669 : /*
7670 : * Do_MultiXactIdWait
7671 : * Actual implementation for the two functions below.
7672 : *
7673 : * 'multi', 'status' and 'infomask' indicate what to sleep on (the status is
7674 : * needed to ensure we only sleep on conflicting members, and the infomask is
7675 : * used to optimize multixact access in case it's a lock-only multi); 'nowait'
7676 : * indicates whether to use conditional lock acquisition, to allow callers to
7677 : * fail if lock is unavailable. 'rel', 'ctid' and 'oper' are used to set up
7678 : * context information for error messages. 'remaining', if not NULL, receives
7679 : * the number of members that are still running, including any (non-aborted)
7680 : * subtransactions of our own transaction. 'logLockFailure' indicates whether
7681 : * to log details when a lock acquisition fails with 'nowait' enabled.
7682 : *
7683 : * We do this by sleeping on each member using XactLockTableWait. Any
7684 : * members that belong to the current backend are *not* waited for, however;
7685 : * this would not merely be useless but would lead to Assert failure inside
7686 : * XactLockTableWait. By the time this returns, it is certain that all
7687 : * transactions *of other backends* that were members of the MultiXactId
7688 : * that conflict with the requested status are dead (and no new ones can have
7689 : * been added, since it is not legal to add members to an existing
7690 : * MultiXactId).
7691 : *
7692 : * But by the time we finish sleeping, someone else may have changed the Xmax
7693 : * of the containing tuple, so the caller needs to iterate on us somehow.
7694 : *
7695 : * Note that in case we return false, the number of remaining members is
7696 : * not to be trusted.
7697 : */
7698 : static bool
7699 116 : Do_MultiXactIdWait(MultiXactId multi, MultiXactStatus status,
7700 : uint16 infomask, bool nowait,
7701 : Relation rel, ItemPointer ctid, XLTW_Oper oper,
7702 : int *remaining, bool logLockFailure)
7703 : {
7704 116 : bool result = true;
7705 : MultiXactMember *members;
7706 : int nmembers;
7707 116 : int remain = 0;
7708 :
7709 : /* for pre-pg_upgrade tuples, no need to sleep at all */
7710 116 : nmembers = HEAP_LOCKED_UPGRADED(infomask) ? -1 :
7711 116 : GetMultiXactIdMembers(multi, &members, false,
7712 116 : HEAP_XMAX_IS_LOCKED_ONLY(infomask));
7713 :
7714 116 : if (nmembers >= 0)
7715 : {
7716 : int i;
7717 :
7718 374 : for (i = 0; i < nmembers; i++)
7719 : {
7720 266 : TransactionId memxid = members[i].xid;
7721 266 : MultiXactStatus memstatus = members[i].status;
7722 :
7723 266 : if (TransactionIdIsCurrentTransactionId(memxid))
7724 : {
7725 48 : remain++;
7726 48 : continue;
7727 : }
7728 :
7729 218 : if (!DoLockModesConflict(LOCKMODE_from_mxstatus(memstatus),
7730 218 : LOCKMODE_from_mxstatus(status)))
7731 : {
7732 44 : if (remaining && TransactionIdIsInProgress(memxid))
7733 16 : remain++;
7734 44 : continue;
7735 : }
7736 :
7737 : /*
7738 : * This member conflicts with our multi, so we have to sleep (or
7739 : * return failure, if asked to avoid waiting.)
7740 : *
7741 : * Note that we don't set up an error context callback ourselves,
7742 : * but instead we pass the info down to XactLockTableWait. This
7743 : * might seem a bit wasteful because the context is set up and
7744 : * tore down for each member of the multixact, but in reality it
7745 : * should be barely noticeable, and it avoids duplicate code.
7746 : */
7747 174 : if (nowait)
7748 : {
7749 8 : result = ConditionalXactLockTableWait(memxid, logLockFailure);
7750 8 : if (!result)
7751 8 : break;
7752 : }
7753 : else
7754 166 : XactLockTableWait(memxid, rel, ctid, oper);
7755 : }
7756 :
7757 116 : pfree(members);
7758 : }
7759 :
7760 116 : if (remaining)
7761 20 : *remaining = remain;
7762 :
7763 116 : return result;
7764 : }
7765 :
7766 : /*
7767 : * MultiXactIdWait
7768 : * Sleep on a MultiXactId.
7769 : *
7770 : * By the time we finish sleeping, someone else may have changed the Xmax
7771 : * of the containing tuple, so the caller needs to iterate on us somehow.
7772 : *
7773 : * We return (in *remaining, if not NULL) the number of members that are still
7774 : * running, including any (non-aborted) subtransactions of our own transaction.
7775 : */
7776 : static void
7777 108 : MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
7778 : Relation rel, ItemPointer ctid, XLTW_Oper oper,
7779 : int *remaining)
7780 : {
7781 108 : (void) Do_MultiXactIdWait(multi, status, infomask, false,
7782 : rel, ctid, oper, remaining, false);
7783 108 : }
7784 :
7785 : /*
7786 : * ConditionalMultiXactIdWait
7787 : * As above, but only lock if we can get the lock without blocking.
7788 : *
7789 : * By the time we finish sleeping, someone else may have changed the Xmax
7790 : * of the containing tuple, so the caller needs to iterate on us somehow.
7791 : *
7792 : * If the multixact is now all gone, return true. Returns false if some
7793 : * transactions might still be running.
7794 : *
7795 : * We return (in *remaining, if not NULL) the number of members that are still
7796 : * running, including any (non-aborted) subtransactions of our own transaction.
7797 : */
7798 : static bool
7799 8 : ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
7800 : uint16 infomask, Relation rel, int *remaining,
7801 : bool logLockFailure)
7802 : {
7803 8 : return Do_MultiXactIdWait(multi, status, infomask, true,
7804 : rel, NULL, XLTW_None, remaining, logLockFailure);
7805 : }
7806 :
7807 : /*
7808 : * heap_tuple_needs_eventual_freeze
7809 : *
7810 : * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
7811 : * will eventually require freezing (if tuple isn't removed by pruning first).
7812 : */
7813 : bool
7814 259940 : heap_tuple_needs_eventual_freeze(HeapTupleHeader tuple)
7815 : {
7816 : TransactionId xid;
7817 :
7818 : /*
7819 : * If xmin is a normal transaction ID, this tuple is definitely not
7820 : * frozen.
7821 : */
7822 259940 : xid = HeapTupleHeaderGetXmin(tuple);
7823 259940 : if (TransactionIdIsNormal(xid))
7824 4734 : return true;
7825 :
7826 : /*
7827 : * If xmax is a valid xact or multixact, this tuple is also not frozen.
7828 : */
7829 255206 : if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
7830 : {
7831 : MultiXactId multi;
7832 :
7833 0 : multi = HeapTupleHeaderGetRawXmax(tuple);
7834 0 : if (MultiXactIdIsValid(multi))
7835 0 : return true;
7836 : }
7837 : else
7838 : {
7839 255206 : xid = HeapTupleHeaderGetRawXmax(tuple);
7840 255206 : if (TransactionIdIsNormal(xid))
7841 10 : return true;
7842 : }
7843 :
7844 255196 : if (tuple->t_infomask & HEAP_MOVED)
7845 : {
7846 0 : xid = HeapTupleHeaderGetXvac(tuple);
7847 0 : if (TransactionIdIsNormal(xid))
7848 0 : return true;
7849 : }
7850 :
7851 255196 : return false;
7852 : }
7853 :
7854 : /*
7855 : * heap_tuple_should_freeze
7856 : *
7857 : * Return value indicates if heap_prepare_freeze_tuple sibling function would
7858 : * (or should) force freezing of the heap page that contains caller's tuple.
7859 : * Tuple header XIDs/MXIDs < FreezeLimit/MultiXactCutoff trigger freezing.
7860 : * This includes (xmin, xmax, xvac) fields, as well as MultiXact member XIDs.
7861 : *
7862 : * The *NoFreezePageRelfrozenXid and *NoFreezePageRelminMxid input/output
7863 : * arguments help VACUUM track the oldest extant XID/MXID remaining in rel.
7864 : * Our working assumption is that caller won't decide to freeze this tuple.
7865 : * It's up to caller to only ratchet back its own top-level trackers after the
7866 : * point that it fully commits to not freezing the tuple/page in question.
7867 : */
7868 : bool
7869 20188922 : heap_tuple_should_freeze(HeapTupleHeader tuple,
7870 : const struct VacuumCutoffs *cutoffs,
7871 : TransactionId *NoFreezePageRelfrozenXid,
7872 : MultiXactId *NoFreezePageRelminMxid)
7873 : {
7874 : TransactionId xid;
7875 : MultiXactId multi;
7876 20188922 : bool freeze = false;
7877 :
7878 : /* First deal with xmin */
7879 20188922 : xid = HeapTupleHeaderGetXmin(tuple);
7880 20188922 : if (TransactionIdIsNormal(xid))
7881 : {
7882 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7883 3572390 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7884 45822 : *NoFreezePageRelfrozenXid = xid;
7885 3572390 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
7886 42084 : freeze = true;
7887 : }
7888 :
7889 : /* Now deal with xmax */
7890 20188922 : xid = InvalidTransactionId;
7891 20188922 : multi = InvalidMultiXactId;
7892 20188922 : if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
7893 4 : multi = HeapTupleHeaderGetRawXmax(tuple);
7894 : else
7895 20188918 : xid = HeapTupleHeaderGetRawXmax(tuple);
7896 :
7897 20188922 : if (TransactionIdIsNormal(xid))
7898 : {
7899 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7900 : /* xmax is a non-permanent XID */
7901 17270980 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7902 4 : *NoFreezePageRelfrozenXid = xid;
7903 17270980 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
7904 52 : freeze = true;
7905 : }
7906 2917942 : else if (!MultiXactIdIsValid(multi))
7907 : {
7908 : /* xmax is a permanent XID or invalid MultiXactId/XID */
7909 : }
7910 4 : else if (HEAP_LOCKED_UPGRADED(tuple->t_infomask))
7911 : {
7912 : /* xmax is a pg_upgrade'd MultiXact, which can't have updater XID */
7913 0 : if (MultiXactIdPrecedes(multi, *NoFreezePageRelminMxid))
7914 0 : *NoFreezePageRelminMxid = multi;
7915 : /* heap_prepare_freeze_tuple always freezes pg_upgrade'd xmax */
7916 0 : freeze = true;
7917 : }
7918 : else
7919 : {
7920 : /* xmax is a MultiXactId that may have an updater XID */
7921 : MultiXactMember *members;
7922 : int nmembers;
7923 :
7924 : Assert(MultiXactIdPrecedesOrEquals(cutoffs->relminmxid, multi));
7925 4 : if (MultiXactIdPrecedes(multi, *NoFreezePageRelminMxid))
7926 4 : *NoFreezePageRelminMxid = multi;
7927 4 : if (MultiXactIdPrecedes(multi, cutoffs->MultiXactCutoff))
7928 4 : freeze = true;
7929 :
7930 : /* need to check whether any member of the mxact is old */
7931 4 : nmembers = GetMultiXactIdMembers(multi, &members, false,
7932 4 : HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask));
7933 :
7934 10 : for (int i = 0; i < nmembers; i++)
7935 : {
7936 6 : xid = members[i].xid;
7937 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7938 6 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7939 0 : *NoFreezePageRelfrozenXid = xid;
7940 6 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
7941 0 : freeze = true;
7942 : }
7943 4 : if (nmembers > 0)
7944 2 : pfree(members);
7945 : }
7946 :
7947 20188922 : if (tuple->t_infomask & HEAP_MOVED)
7948 : {
7949 0 : xid = HeapTupleHeaderGetXvac(tuple);
7950 0 : if (TransactionIdIsNormal(xid))
7951 : {
7952 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7953 0 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7954 0 : *NoFreezePageRelfrozenXid = xid;
7955 : /* heap_prepare_freeze_tuple forces xvac freezing */
7956 0 : freeze = true;
7957 : }
7958 : }
7959 :
7960 20188922 : return freeze;
7961 : }
7962 :
7963 : /*
7964 : * Maintain snapshotConflictHorizon for caller by ratcheting forward its value
7965 : * using any committed XIDs contained in 'tuple', an obsolescent heap tuple
7966 : * that caller is in the process of physically removing, e.g. via HOT pruning
7967 : * or index deletion.
7968 : *
7969 : * Caller must initialize its value to InvalidTransactionId, which is
7970 : * generally interpreted as "definitely no need for a recovery conflict".
7971 : * Final value must reflect all heap tuples that caller will physically remove
7972 : * (or remove TID references to) via its ongoing pruning/deletion operation.
7973 : * ResolveRecoveryConflictWithSnapshot() is passed the final value (taken from
7974 : * caller's WAL record) by REDO routine when it replays caller's operation.
7975 : */
7976 : void
7977 3002676 : HeapTupleHeaderAdvanceConflictHorizon(HeapTupleHeader tuple,
7978 : TransactionId *snapshotConflictHorizon)
7979 : {
7980 3002676 : TransactionId xmin = HeapTupleHeaderGetXmin(tuple);
7981 3002676 : TransactionId xmax = HeapTupleHeaderGetUpdateXid(tuple);
7982 3002676 : TransactionId xvac = HeapTupleHeaderGetXvac(tuple);
7983 :
7984 3002676 : if (tuple->t_infomask & HEAP_MOVED)
7985 : {
7986 0 : if (TransactionIdPrecedes(*snapshotConflictHorizon, xvac))
7987 0 : *snapshotConflictHorizon = xvac;
7988 : }
7989 :
7990 : /*
7991 : * Ignore tuples inserted by an aborted transaction or if the tuple was
7992 : * updated/deleted by the inserting transaction.
7993 : *
7994 : * Look for a committed hint bit, or if no xmin bit is set, check clog.
7995 : */
7996 3002676 : if (HeapTupleHeaderXminCommitted(tuple) ||
7997 209776 : (!HeapTupleHeaderXminInvalid(tuple) && TransactionIdDidCommit(xmin)))
7998 : {
7999 5426788 : if (xmax != xmin &&
8000 2578646 : TransactionIdFollows(xmax, *snapshotConflictHorizon))
8001 190258 : *snapshotConflictHorizon = xmax;
8002 : }
8003 3002676 : }
8004 :
8005 : #ifdef USE_PREFETCH
8006 : /*
8007 : * Helper function for heap_index_delete_tuples. Issues prefetch requests for
8008 : * prefetch_count buffers. The prefetch_state keeps track of all the buffers
8009 : * we can prefetch, and which have already been prefetched; each call to this
8010 : * function picks up where the previous call left off.
8011 : *
8012 : * Note: we expect the deltids array to be sorted in an order that groups TIDs
8013 : * by heap block, with all TIDs for each block appearing together in exactly
8014 : * one group.
8015 : */
8016 : static void
8017 37408 : index_delete_prefetch_buffer(Relation rel,
8018 : IndexDeletePrefetchState *prefetch_state,
8019 : int prefetch_count)
8020 : {
8021 37408 : BlockNumber cur_hblkno = prefetch_state->cur_hblkno;
8022 37408 : int count = 0;
8023 : int i;
8024 37408 : int ndeltids = prefetch_state->ndeltids;
8025 37408 : TM_IndexDelete *deltids = prefetch_state->deltids;
8026 :
8027 37408 : for (i = prefetch_state->next_item;
8028 1299788 : i < ndeltids && count < prefetch_count;
8029 1262380 : i++)
8030 : {
8031 1262380 : ItemPointer htid = &deltids[i].tid;
8032 :
8033 2513518 : if (cur_hblkno == InvalidBlockNumber ||
8034 1251138 : ItemPointerGetBlockNumber(htid) != cur_hblkno)
8035 : {
8036 33846 : cur_hblkno = ItemPointerGetBlockNumber(htid);
8037 33846 : PrefetchBuffer(rel, MAIN_FORKNUM, cur_hblkno);
8038 33846 : count++;
8039 : }
8040 : }
8041 :
8042 : /*
8043 : * Save the prefetch position so that next time we can continue from that
8044 : * position.
8045 : */
8046 37408 : prefetch_state->next_item = i;
8047 37408 : prefetch_state->cur_hblkno = cur_hblkno;
8048 37408 : }
8049 : #endif
8050 :
8051 : /*
8052 : * Helper function for heap_index_delete_tuples. Checks for index corruption
8053 : * involving an invalid TID in index AM caller's index page.
8054 : *
8055 : * This is an ideal place for these checks. The index AM must hold a buffer
8056 : * lock on the index page containing the TIDs we examine here, so we don't
8057 : * have to worry about concurrent VACUUMs at all. We can be sure that the
8058 : * index is corrupt when htid points directly to an LP_UNUSED item or
8059 : * heap-only tuple, which is not the case during standard index scans.
8060 : */
8061 : static inline void
8062 1055998 : index_delete_check_htid(TM_IndexDeleteOp *delstate,
8063 : Page page, OffsetNumber maxoff,
8064 : ItemPointer htid, TM_IndexStatus *istatus)
8065 : {
8066 1055998 : OffsetNumber indexpagehoffnum = ItemPointerGetOffsetNumber(htid);
8067 : ItemId iid;
8068 :
8069 : Assert(OffsetNumberIsValid(istatus->idxoffnum));
8070 :
8071 1055998 : if (unlikely(indexpagehoffnum > maxoff))
8072 0 : ereport(ERROR,
8073 : (errcode(ERRCODE_INDEX_CORRUPTED),
8074 : 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\"",
8075 : ItemPointerGetBlockNumber(htid),
8076 : indexpagehoffnum,
8077 : istatus->idxoffnum, delstate->iblknum,
8078 : RelationGetRelationName(delstate->irel))));
8079 :
8080 1055998 : iid = PageGetItemId(page, indexpagehoffnum);
8081 1055998 : if (unlikely(!ItemIdIsUsed(iid)))
8082 0 : ereport(ERROR,
8083 : (errcode(ERRCODE_INDEX_CORRUPTED),
8084 : errmsg_internal("heap tid from index tuple (%u,%u) points to unused heap page item at offset %u of block %u in index \"%s\"",
8085 : ItemPointerGetBlockNumber(htid),
8086 : indexpagehoffnum,
8087 : istatus->idxoffnum, delstate->iblknum,
8088 : RelationGetRelationName(delstate->irel))));
8089 :
8090 1055998 : if (ItemIdHasStorage(iid))
8091 : {
8092 : HeapTupleHeader htup;
8093 :
8094 : Assert(ItemIdIsNormal(iid));
8095 630180 : htup = (HeapTupleHeader) PageGetItem(page, iid);
8096 :
8097 630180 : if (unlikely(HeapTupleHeaderIsHeapOnly(htup)))
8098 0 : ereport(ERROR,
8099 : (errcode(ERRCODE_INDEX_CORRUPTED),
8100 : errmsg_internal("heap tid from index tuple (%u,%u) points to heap-only tuple at offset %u of block %u in index \"%s\"",
8101 : ItemPointerGetBlockNumber(htid),
8102 : indexpagehoffnum,
8103 : istatus->idxoffnum, delstate->iblknum,
8104 : RelationGetRelationName(delstate->irel))));
8105 : }
8106 1055998 : }
8107 :
8108 : /*
8109 : * heapam implementation of tableam's index_delete_tuples interface.
8110 : *
8111 : * This helper function is called by index AMs during index tuple deletion.
8112 : * See tableam header comments for an explanation of the interface implemented
8113 : * here and a general theory of operation. Note that each call here is either
8114 : * a simple index deletion call, or a bottom-up index deletion call.
8115 : *
8116 : * It's possible for this to generate a fair amount of I/O, since we may be
8117 : * deleting hundreds of tuples from a single index block. To amortize that
8118 : * cost to some degree, this uses prefetching and combines repeat accesses to
8119 : * the same heap block.
8120 : */
8121 : TransactionId
8122 11242 : heap_index_delete_tuples(Relation rel, TM_IndexDeleteOp *delstate)
8123 : {
8124 : /* Initial assumption is that earlier pruning took care of conflict */
8125 11242 : TransactionId snapshotConflictHorizon = InvalidTransactionId;
8126 11242 : BlockNumber blkno = InvalidBlockNumber;
8127 11242 : Buffer buf = InvalidBuffer;
8128 11242 : Page page = NULL;
8129 11242 : OffsetNumber maxoff = InvalidOffsetNumber;
8130 : TransactionId priorXmax;
8131 : #ifdef USE_PREFETCH
8132 : IndexDeletePrefetchState prefetch_state;
8133 : int prefetch_distance;
8134 : #endif
8135 : SnapshotData SnapshotNonVacuumable;
8136 11242 : int finalndeltids = 0,
8137 11242 : nblocksaccessed = 0;
8138 :
8139 : /* State that's only used in bottom-up index deletion case */
8140 11242 : int nblocksfavorable = 0;
8141 11242 : int curtargetfreespace = delstate->bottomupfreespace,
8142 11242 : lastfreespace = 0,
8143 11242 : actualfreespace = 0;
8144 11242 : bool bottomup_final_block = false;
8145 :
8146 11242 : InitNonVacuumableSnapshot(SnapshotNonVacuumable, GlobalVisTestFor(rel));
8147 :
8148 : /* Sort caller's deltids array by TID for further processing */
8149 11242 : index_delete_sort(delstate);
8150 :
8151 : /*
8152 : * Bottom-up case: resort deltids array in an order attuned to where the
8153 : * greatest number of promising TIDs are to be found, and determine how
8154 : * many blocks from the start of sorted array should be considered
8155 : * favorable. This will also shrink the deltids array in order to
8156 : * eliminate completely unfavorable blocks up front.
8157 : */
8158 11242 : if (delstate->bottomup)
8159 3914 : nblocksfavorable = bottomup_sort_and_shrink(delstate);
8160 :
8161 : #ifdef USE_PREFETCH
8162 : /* Initialize prefetch state. */
8163 11242 : prefetch_state.cur_hblkno = InvalidBlockNumber;
8164 11242 : prefetch_state.next_item = 0;
8165 11242 : prefetch_state.ndeltids = delstate->ndeltids;
8166 11242 : prefetch_state.deltids = delstate->deltids;
8167 :
8168 : /*
8169 : * Determine the prefetch distance that we will attempt to maintain.
8170 : *
8171 : * Since the caller holds a buffer lock somewhere in rel, we'd better make
8172 : * sure that isn't a catalog relation before we call code that does
8173 : * syscache lookups, to avoid risk of deadlock.
8174 : */
8175 11242 : if (IsCatalogRelation(rel))
8176 8032 : prefetch_distance = maintenance_io_concurrency;
8177 : else
8178 : prefetch_distance =
8179 3210 : get_tablespace_maintenance_io_concurrency(rel->rd_rel->reltablespace);
8180 :
8181 : /* Cap initial prefetch distance for bottom-up deletion caller */
8182 11242 : if (delstate->bottomup)
8183 : {
8184 : Assert(nblocksfavorable >= 1);
8185 : Assert(nblocksfavorable <= BOTTOMUP_MAX_NBLOCKS);
8186 3914 : prefetch_distance = Min(prefetch_distance, nblocksfavorable);
8187 : }
8188 :
8189 : /* Start prefetching. */
8190 11242 : index_delete_prefetch_buffer(rel, &prefetch_state, prefetch_distance);
8191 : #endif
8192 :
8193 : /* Iterate over deltids, determine which to delete, check their horizon */
8194 : Assert(delstate->ndeltids > 0);
8195 1067240 : for (int i = 0; i < delstate->ndeltids; i++)
8196 : {
8197 1059912 : TM_IndexDelete *ideltid = &delstate->deltids[i];
8198 1059912 : TM_IndexStatus *istatus = delstate->status + ideltid->id;
8199 1059912 : ItemPointer htid = &ideltid->tid;
8200 : OffsetNumber offnum;
8201 :
8202 : /*
8203 : * Read buffer, and perform required extra steps each time a new block
8204 : * is encountered. Avoid refetching if it's the same block as the one
8205 : * from the last htid.
8206 : */
8207 2108582 : if (blkno == InvalidBlockNumber ||
8208 1048670 : ItemPointerGetBlockNumber(htid) != blkno)
8209 : {
8210 : /*
8211 : * Consider giving up early for bottom-up index deletion caller
8212 : * first. (Only prefetch next-next block afterwards, when it
8213 : * becomes clear that we're at least going to access the next
8214 : * block in line.)
8215 : *
8216 : * Sometimes the first block frees so much space for bottom-up
8217 : * caller that the deletion process can end without accessing any
8218 : * more blocks. It is usually necessary to access 2 or 3 blocks
8219 : * per bottom-up deletion operation, though.
8220 : */
8221 30080 : if (delstate->bottomup)
8222 : {
8223 : /*
8224 : * We often allow caller to delete a few additional items
8225 : * whose entries we reached after the point that space target
8226 : * from caller was satisfied. The cost of accessing the page
8227 : * was already paid at that point, so it made sense to finish
8228 : * it off. When that happened, we finalize everything here
8229 : * (by finishing off the whole bottom-up deletion operation
8230 : * without needlessly paying the cost of accessing any more
8231 : * blocks).
8232 : */
8233 8358 : if (bottomup_final_block)
8234 294 : break;
8235 :
8236 : /*
8237 : * Give up when we didn't enable our caller to free any
8238 : * additional space as a result of processing the page that we
8239 : * just finished up with. This rule is the main way in which
8240 : * we keep the cost of bottom-up deletion under control.
8241 : */
8242 8064 : if (nblocksaccessed >= 1 && actualfreespace == lastfreespace)
8243 3620 : break;
8244 4444 : lastfreespace = actualfreespace; /* for next time */
8245 :
8246 : /*
8247 : * Deletion operation (which is bottom-up) will definitely
8248 : * access the next block in line. Prepare for that now.
8249 : *
8250 : * Decay target free space so that we don't hang on for too
8251 : * long with a marginal case. (Space target is only truly
8252 : * helpful when it allows us to recognize that we don't need
8253 : * to access more than 1 or 2 blocks to satisfy caller due to
8254 : * agreeable workload characteristics.)
8255 : *
8256 : * We are a bit more patient when we encounter contiguous
8257 : * blocks, though: these are treated as favorable blocks. The
8258 : * decay process is only applied when the next block in line
8259 : * is not a favorable/contiguous block. This is not an
8260 : * exception to the general rule; we still insist on finding
8261 : * at least one deletable item per block accessed. See
8262 : * bottomup_nblocksfavorable() for full details of the theory
8263 : * behind favorable blocks and heap block locality in general.
8264 : *
8265 : * Note: The first block in line is always treated as a
8266 : * favorable block, so the earliest possible point that the
8267 : * decay can be applied is just before we access the second
8268 : * block in line. The Assert() verifies this for us.
8269 : */
8270 : Assert(nblocksaccessed > 0 || nblocksfavorable > 0);
8271 4444 : if (nblocksfavorable > 0)
8272 4242 : nblocksfavorable--;
8273 : else
8274 202 : curtargetfreespace /= 2;
8275 : }
8276 :
8277 : /* release old buffer */
8278 26166 : if (BufferIsValid(buf))
8279 14924 : UnlockReleaseBuffer(buf);
8280 :
8281 26166 : blkno = ItemPointerGetBlockNumber(htid);
8282 26166 : buf = ReadBuffer(rel, blkno);
8283 26166 : nblocksaccessed++;
8284 : Assert(!delstate->bottomup ||
8285 : nblocksaccessed <= BOTTOMUP_MAX_NBLOCKS);
8286 :
8287 : #ifdef USE_PREFETCH
8288 :
8289 : /*
8290 : * To maintain the prefetch distance, prefetch one more page for
8291 : * each page we read.
8292 : */
8293 26166 : index_delete_prefetch_buffer(rel, &prefetch_state, 1);
8294 : #endif
8295 :
8296 26166 : LockBuffer(buf, BUFFER_LOCK_SHARE);
8297 :
8298 26166 : page = BufferGetPage(buf);
8299 26166 : maxoff = PageGetMaxOffsetNumber(page);
8300 : }
8301 :
8302 : /*
8303 : * In passing, detect index corruption involving an index page with a
8304 : * TID that points to a location in the heap that couldn't possibly be
8305 : * correct. We only do this with actual TIDs from caller's index page
8306 : * (not items reached by traversing through a HOT chain).
8307 : */
8308 1055998 : index_delete_check_htid(delstate, page, maxoff, htid, istatus);
8309 :
8310 1055998 : if (istatus->knowndeletable)
8311 : Assert(!delstate->bottomup && !istatus->promising);
8312 : else
8313 : {
8314 793442 : ItemPointerData tmp = *htid;
8315 : HeapTupleData heapTuple;
8316 :
8317 : /* Are any tuples from this HOT chain non-vacuumable? */
8318 793442 : if (heap_hot_search_buffer(&tmp, rel, buf, &SnapshotNonVacuumable,
8319 : &heapTuple, NULL, true))
8320 475348 : continue; /* can't delete entry */
8321 :
8322 : /* Caller will delete, since whole HOT chain is vacuumable */
8323 318094 : istatus->knowndeletable = true;
8324 :
8325 : /* Maintain index free space info for bottom-up deletion case */
8326 318094 : if (delstate->bottomup)
8327 : {
8328 : Assert(istatus->freespace > 0);
8329 14598 : actualfreespace += istatus->freespace;
8330 14598 : if (actualfreespace >= curtargetfreespace)
8331 4850 : bottomup_final_block = true;
8332 : }
8333 : }
8334 :
8335 : /*
8336 : * Maintain snapshotConflictHorizon value for deletion operation as a
8337 : * whole by advancing current value using heap tuple headers. This is
8338 : * loosely based on the logic for pruning a HOT chain.
8339 : */
8340 580650 : offnum = ItemPointerGetOffsetNumber(htid);
8341 580650 : priorXmax = InvalidTransactionId; /* cannot check first XMIN */
8342 : for (;;)
8343 40446 : {
8344 : ItemId lp;
8345 : HeapTupleHeader htup;
8346 :
8347 : /* Sanity check (pure paranoia) */
8348 621096 : if (offnum < FirstOffsetNumber)
8349 0 : break;
8350 :
8351 : /*
8352 : * An offset past the end of page's line pointer array is possible
8353 : * when the array was truncated
8354 : */
8355 621096 : if (offnum > maxoff)
8356 0 : break;
8357 :
8358 621096 : lp = PageGetItemId(page, offnum);
8359 621096 : if (ItemIdIsRedirected(lp))
8360 : {
8361 18216 : offnum = ItemIdGetRedirect(lp);
8362 18216 : continue;
8363 : }
8364 :
8365 : /*
8366 : * We'll often encounter LP_DEAD line pointers (especially with an
8367 : * entry marked knowndeletable by our caller up front). No heap
8368 : * tuple headers get examined for an htid that leads us to an
8369 : * LP_DEAD item. This is okay because the earlier pruning
8370 : * operation that made the line pointer LP_DEAD in the first place
8371 : * must have considered the original tuple header as part of
8372 : * generating its own snapshotConflictHorizon value.
8373 : *
8374 : * Relying on XLOG_HEAP2_PRUNE_VACUUM_SCAN records like this is
8375 : * the same strategy that index vacuuming uses in all cases. Index
8376 : * VACUUM WAL records don't even have a snapshotConflictHorizon
8377 : * field of their own for this reason.
8378 : */
8379 602880 : if (!ItemIdIsNormal(lp))
8380 381734 : break;
8381 :
8382 221146 : htup = (HeapTupleHeader) PageGetItem(page, lp);
8383 :
8384 : /*
8385 : * Check the tuple XMIN against prior XMAX, if any
8386 : */
8387 243376 : if (TransactionIdIsValid(priorXmax) &&
8388 22230 : !TransactionIdEquals(HeapTupleHeaderGetXmin(htup), priorXmax))
8389 0 : break;
8390 :
8391 221146 : HeapTupleHeaderAdvanceConflictHorizon(htup,
8392 : &snapshotConflictHorizon);
8393 :
8394 : /*
8395 : * If the tuple is not HOT-updated, then we are at the end of this
8396 : * HOT-chain. No need to visit later tuples from the same update
8397 : * chain (they get their own index entries) -- just move on to
8398 : * next htid from index AM caller.
8399 : */
8400 221146 : if (!HeapTupleHeaderIsHotUpdated(htup))
8401 198916 : break;
8402 :
8403 : /* Advance to next HOT chain member */
8404 : Assert(ItemPointerGetBlockNumber(&htup->t_ctid) == blkno);
8405 22230 : offnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
8406 22230 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
8407 : }
8408 :
8409 : /* Enable further/final shrinking of deltids for caller */
8410 580650 : finalndeltids = i + 1;
8411 : }
8412 :
8413 11242 : UnlockReleaseBuffer(buf);
8414 :
8415 : /*
8416 : * Shrink deltids array to exclude non-deletable entries at the end. This
8417 : * is not just a minor optimization. Final deltids array size might be
8418 : * zero for a bottom-up caller. Index AM is explicitly allowed to rely on
8419 : * ndeltids being zero in all cases with zero total deletable entries.
8420 : */
8421 : Assert(finalndeltids > 0 || delstate->bottomup);
8422 11242 : delstate->ndeltids = finalndeltids;
8423 :
8424 11242 : return snapshotConflictHorizon;
8425 : }
8426 :
8427 : /*
8428 : * Specialized inlineable comparison function for index_delete_sort()
8429 : */
8430 : static inline int
8431 24784032 : index_delete_sort_cmp(TM_IndexDelete *deltid1, TM_IndexDelete *deltid2)
8432 : {
8433 24784032 : ItemPointer tid1 = &deltid1->tid;
8434 24784032 : ItemPointer tid2 = &deltid2->tid;
8435 :
8436 : {
8437 24784032 : BlockNumber blk1 = ItemPointerGetBlockNumber(tid1);
8438 24784032 : BlockNumber blk2 = ItemPointerGetBlockNumber(tid2);
8439 :
8440 24784032 : if (blk1 != blk2)
8441 10139394 : return (blk1 < blk2) ? -1 : 1;
8442 : }
8443 : {
8444 14644638 : OffsetNumber pos1 = ItemPointerGetOffsetNumber(tid1);
8445 14644638 : OffsetNumber pos2 = ItemPointerGetOffsetNumber(tid2);
8446 :
8447 14644638 : if (pos1 != pos2)
8448 14644638 : return (pos1 < pos2) ? -1 : 1;
8449 : }
8450 :
8451 : Assert(false);
8452 :
8453 0 : return 0;
8454 : }
8455 :
8456 : /*
8457 : * Sort deltids array from delstate by TID. This prepares it for further
8458 : * processing by heap_index_delete_tuples().
8459 : *
8460 : * This operation becomes a noticeable consumer of CPU cycles with some
8461 : * workloads, so we go to the trouble of specialization/micro optimization.
8462 : * We use shellsort for this because it's easy to specialize, compiles to
8463 : * relatively few instructions, and is adaptive to presorted inputs/subsets
8464 : * (which are typical here).
8465 : */
8466 : static void
8467 11242 : index_delete_sort(TM_IndexDeleteOp *delstate)
8468 : {
8469 11242 : TM_IndexDelete *deltids = delstate->deltids;
8470 11242 : int ndeltids = delstate->ndeltids;
8471 :
8472 : /*
8473 : * Shellsort gap sequence (taken from Sedgewick-Incerpi paper).
8474 : *
8475 : * This implementation is fast with array sizes up to ~4500. This covers
8476 : * all supported BLCKSZ values.
8477 : */
8478 11242 : const int gaps[9] = {1968, 861, 336, 112, 48, 21, 7, 3, 1};
8479 :
8480 : /* Think carefully before changing anything here -- keep swaps cheap */
8481 : StaticAssertDecl(sizeof(TM_IndexDelete) <= 8,
8482 : "element size exceeds 8 bytes");
8483 :
8484 112420 : for (int g = 0; g < lengthof(gaps); g++)
8485 : {
8486 14820382 : for (int hi = gaps[g], i = hi; i < ndeltids; i++)
8487 : {
8488 14719204 : TM_IndexDelete d = deltids[i];
8489 14719204 : int j = i;
8490 :
8491 25493500 : while (j >= hi && index_delete_sort_cmp(&deltids[j - hi], &d) >= 0)
8492 : {
8493 10774296 : deltids[j] = deltids[j - hi];
8494 10774296 : j -= hi;
8495 : }
8496 14719204 : deltids[j] = d;
8497 : }
8498 : }
8499 11242 : }
8500 :
8501 : /*
8502 : * Returns how many blocks should be considered favorable/contiguous for a
8503 : * bottom-up index deletion pass. This is a number of heap blocks that starts
8504 : * from and includes the first block in line.
8505 : *
8506 : * There is always at least one favorable block during bottom-up index
8507 : * deletion. In the worst case (i.e. with totally random heap blocks) the
8508 : * first block in line (the only favorable block) can be thought of as a
8509 : * degenerate array of contiguous blocks that consists of a single block.
8510 : * heap_index_delete_tuples() will expect this.
8511 : *
8512 : * Caller passes blockgroups, a description of the final order that deltids
8513 : * will be sorted in for heap_index_delete_tuples() bottom-up index deletion
8514 : * processing. Note that deltids need not actually be sorted just yet (caller
8515 : * only passes deltids to us so that we can interpret blockgroups).
8516 : *
8517 : * You might guess that the existence of contiguous blocks cannot matter much,
8518 : * since in general the main factor that determines which blocks we visit is
8519 : * the number of promising TIDs, which is a fixed hint from the index AM.
8520 : * We're not really targeting the general case, though -- the actual goal is
8521 : * to adapt our behavior to a wide variety of naturally occurring conditions.
8522 : * The effects of most of the heuristics we apply are only noticeable in the
8523 : * aggregate, over time and across many _related_ bottom-up index deletion
8524 : * passes.
8525 : *
8526 : * Deeming certain blocks favorable allows heapam to recognize and adapt to
8527 : * workloads where heap blocks visited during bottom-up index deletion can be
8528 : * accessed contiguously, in the sense that each newly visited block is the
8529 : * neighbor of the block that bottom-up deletion just finished processing (or
8530 : * close enough to it). It will likely be cheaper to access more favorable
8531 : * blocks sooner rather than later (e.g. in this pass, not across a series of
8532 : * related bottom-up passes). Either way it is probably only a matter of time
8533 : * (or a matter of further correlated version churn) before all blocks that
8534 : * appear together as a single large batch of favorable blocks get accessed by
8535 : * _some_ bottom-up pass. Large batches of favorable blocks tend to either
8536 : * appear almost constantly or not even once (it all depends on per-index
8537 : * workload characteristics).
8538 : *
8539 : * Note that the blockgroups sort order applies a power-of-two bucketing
8540 : * scheme that creates opportunities for contiguous groups of blocks to get
8541 : * batched together, at least with workloads that are naturally amenable to
8542 : * being driven by heap block locality. This doesn't just enhance the spatial
8543 : * locality of bottom-up heap block processing in the obvious way. It also
8544 : * enables temporal locality of access, since sorting by heap block number
8545 : * naturally tends to make the bottom-up processing order deterministic.
8546 : *
8547 : * Consider the following example to get a sense of how temporal locality
8548 : * might matter: There is a heap relation with several indexes, each of which
8549 : * is low to medium cardinality. It is subject to constant non-HOT updates.
8550 : * The updates are skewed (in one part of the primary key, perhaps). None of
8551 : * the indexes are logically modified by the UPDATE statements (if they were
8552 : * then bottom-up index deletion would not be triggered in the first place).
8553 : * Naturally, each new round of index tuples (for each heap tuple that gets a
8554 : * heap_update() call) will have the same heap TID in each and every index.
8555 : * Since these indexes are low cardinality and never get logically modified,
8556 : * heapam processing during bottom-up deletion passes will access heap blocks
8557 : * in approximately sequential order. Temporal locality of access occurs due
8558 : * to bottom-up deletion passes behaving very similarly across each of the
8559 : * indexes at any given moment. This keeps the number of buffer misses needed
8560 : * to visit heap blocks to a minimum.
8561 : */
8562 : static int
8563 3914 : bottomup_nblocksfavorable(IndexDeleteCounts *blockgroups, int nblockgroups,
8564 : TM_IndexDelete *deltids)
8565 : {
8566 3914 : int64 lastblock = -1;
8567 3914 : int nblocksfavorable = 0;
8568 :
8569 : Assert(nblockgroups >= 1);
8570 : Assert(nblockgroups <= BOTTOMUP_MAX_NBLOCKS);
8571 :
8572 : /*
8573 : * We tolerate heap blocks that will be accessed only slightly out of
8574 : * physical order. Small blips occur when a pair of almost-contiguous
8575 : * blocks happen to fall into different buckets (perhaps due only to a
8576 : * small difference in npromisingtids that the bucketing scheme didn't
8577 : * quite manage to ignore). We effectively ignore these blips by applying
8578 : * a small tolerance. The precise tolerance we use is a little arbitrary,
8579 : * but it works well enough in practice.
8580 : */
8581 12270 : for (int b = 0; b < nblockgroups; b++)
8582 : {
8583 11766 : IndexDeleteCounts *group = blockgroups + b;
8584 11766 : TM_IndexDelete *firstdtid = deltids + group->ifirsttid;
8585 11766 : BlockNumber block = ItemPointerGetBlockNumber(&firstdtid->tid);
8586 :
8587 11766 : if (lastblock != -1 &&
8588 7852 : ((int64) block < lastblock - BOTTOMUP_TOLERANCE_NBLOCKS ||
8589 6694 : (int64) block > lastblock + BOTTOMUP_TOLERANCE_NBLOCKS))
8590 : break;
8591 :
8592 8356 : nblocksfavorable++;
8593 8356 : lastblock = block;
8594 : }
8595 :
8596 : /* Always indicate that there is at least 1 favorable block */
8597 : Assert(nblocksfavorable >= 1);
8598 :
8599 3914 : return nblocksfavorable;
8600 : }
8601 :
8602 : /*
8603 : * qsort comparison function for bottomup_sort_and_shrink()
8604 : */
8605 : static int
8606 416224 : bottomup_sort_and_shrink_cmp(const void *arg1, const void *arg2)
8607 : {
8608 416224 : const IndexDeleteCounts *group1 = (const IndexDeleteCounts *) arg1;
8609 416224 : const IndexDeleteCounts *group2 = (const IndexDeleteCounts *) arg2;
8610 :
8611 : /*
8612 : * Most significant field is npromisingtids (which we invert the order of
8613 : * so as to sort in desc order).
8614 : *
8615 : * Caller should have already normalized npromisingtids fields into
8616 : * power-of-two values (buckets).
8617 : */
8618 416224 : if (group1->npromisingtids > group2->npromisingtids)
8619 19536 : return -1;
8620 396688 : if (group1->npromisingtids < group2->npromisingtids)
8621 21832 : return 1;
8622 :
8623 : /*
8624 : * Tiebreak: desc ntids sort order.
8625 : *
8626 : * We cannot expect power-of-two values for ntids fields. We should
8627 : * behave as if they were already rounded up for us instead.
8628 : */
8629 374856 : if (group1->ntids != group2->ntids)
8630 : {
8631 260502 : uint32 ntids1 = pg_nextpower2_32((uint32) group1->ntids);
8632 260502 : uint32 ntids2 = pg_nextpower2_32((uint32) group2->ntids);
8633 :
8634 260502 : if (ntids1 > ntids2)
8635 42264 : return -1;
8636 218238 : if (ntids1 < ntids2)
8637 54774 : return 1;
8638 : }
8639 :
8640 : /*
8641 : * Tiebreak: asc offset-into-deltids-for-block (offset to first TID for
8642 : * block in deltids array) order.
8643 : *
8644 : * This is equivalent to sorting in ascending heap block number order
8645 : * (among otherwise equal subsets of the array). This approach allows us
8646 : * to avoid accessing the out-of-line TID. (We rely on the assumption
8647 : * that the deltids array was sorted in ascending heap TID order when
8648 : * these offsets to the first TID from each heap block group were formed.)
8649 : */
8650 277818 : if (group1->ifirsttid > group2->ifirsttid)
8651 138230 : return 1;
8652 139588 : if (group1->ifirsttid < group2->ifirsttid)
8653 139588 : return -1;
8654 :
8655 0 : pg_unreachable();
8656 :
8657 : return 0;
8658 : }
8659 :
8660 : /*
8661 : * heap_index_delete_tuples() helper function for bottom-up deletion callers.
8662 : *
8663 : * Sorts deltids array in the order needed for useful processing by bottom-up
8664 : * deletion. The array should already be sorted in TID order when we're
8665 : * called. The sort process groups heap TIDs from deltids into heap block
8666 : * groupings. Earlier/more-promising groups/blocks are usually those that are
8667 : * known to have the most "promising" TIDs.
8668 : *
8669 : * Sets new size of deltids array (ndeltids) in state. deltids will only have
8670 : * TIDs from the BOTTOMUP_MAX_NBLOCKS most promising heap blocks when we
8671 : * return. This often means that deltids will be shrunk to a small fraction
8672 : * of its original size (we eliminate many heap blocks from consideration for
8673 : * caller up front).
8674 : *
8675 : * Returns the number of "favorable" blocks. See bottomup_nblocksfavorable()
8676 : * for a definition and full details.
8677 : */
8678 : static int
8679 3914 : bottomup_sort_and_shrink(TM_IndexDeleteOp *delstate)
8680 : {
8681 : IndexDeleteCounts *blockgroups;
8682 : TM_IndexDelete *reordereddeltids;
8683 3914 : BlockNumber curblock = InvalidBlockNumber;
8684 3914 : int nblockgroups = 0;
8685 3914 : int ncopied = 0;
8686 3914 : int nblocksfavorable = 0;
8687 :
8688 : Assert(delstate->bottomup);
8689 : Assert(delstate->ndeltids > 0);
8690 :
8691 : /* Calculate per-heap-block count of TIDs */
8692 3914 : blockgroups = palloc(sizeof(IndexDeleteCounts) * delstate->ndeltids);
8693 1835968 : for (int i = 0; i < delstate->ndeltids; i++)
8694 : {
8695 1832054 : TM_IndexDelete *ideltid = &delstate->deltids[i];
8696 1832054 : TM_IndexStatus *istatus = delstate->status + ideltid->id;
8697 1832054 : ItemPointer htid = &ideltid->tid;
8698 1832054 : bool promising = istatus->promising;
8699 :
8700 1832054 : if (curblock != ItemPointerGetBlockNumber(htid))
8701 : {
8702 : /* New block group */
8703 79712 : nblockgroups++;
8704 :
8705 : Assert(curblock < ItemPointerGetBlockNumber(htid) ||
8706 : !BlockNumberIsValid(curblock));
8707 :
8708 79712 : curblock = ItemPointerGetBlockNumber(htid);
8709 79712 : blockgroups[nblockgroups - 1].ifirsttid = i;
8710 79712 : blockgroups[nblockgroups - 1].ntids = 1;
8711 79712 : blockgroups[nblockgroups - 1].npromisingtids = 0;
8712 : }
8713 : else
8714 : {
8715 1752342 : blockgroups[nblockgroups - 1].ntids++;
8716 : }
8717 :
8718 1832054 : if (promising)
8719 236440 : blockgroups[nblockgroups - 1].npromisingtids++;
8720 : }
8721 :
8722 : /*
8723 : * We're about ready to sort block groups to determine the optimal order
8724 : * for visiting heap blocks. But before we do, round the number of
8725 : * promising tuples for each block group up to the next power-of-two,
8726 : * unless it is very low (less than 4), in which case we round up to 4.
8727 : * npromisingtids is far too noisy to trust when choosing between a pair
8728 : * of block groups that both have very low values.
8729 : *
8730 : * This scheme divides heap blocks/block groups into buckets. Each bucket
8731 : * contains blocks that have _approximately_ the same number of promising
8732 : * TIDs as each other. The goal is to ignore relatively small differences
8733 : * in the total number of promising entries, so that the whole process can
8734 : * give a little weight to heapam factors (like heap block locality)
8735 : * instead. This isn't a trade-off, really -- we have nothing to lose. It
8736 : * would be foolish to interpret small differences in npromisingtids
8737 : * values as anything more than noise.
8738 : *
8739 : * We tiebreak on nhtids when sorting block group subsets that have the
8740 : * same npromisingtids, but this has the same issues as npromisingtids,
8741 : * and so nhtids is subject to the same power-of-two bucketing scheme. The
8742 : * only reason that we don't fix nhtids in the same way here too is that
8743 : * we'll need accurate nhtids values after the sort. We handle nhtids
8744 : * bucketization dynamically instead (in the sort comparator).
8745 : *
8746 : * See bottomup_nblocksfavorable() for a full explanation of when and how
8747 : * heap locality/favorable blocks can significantly influence when and how
8748 : * heap blocks are accessed.
8749 : */
8750 83626 : for (int b = 0; b < nblockgroups; b++)
8751 : {
8752 79712 : IndexDeleteCounts *group = blockgroups + b;
8753 :
8754 : /* Better off falling back on nhtids with low npromisingtids */
8755 79712 : if (group->npromisingtids <= 4)
8756 68816 : group->npromisingtids = 4;
8757 : else
8758 10896 : group->npromisingtids =
8759 10896 : pg_nextpower2_32((uint32) group->npromisingtids);
8760 : }
8761 :
8762 : /* Sort groups and rearrange caller's deltids array */
8763 3914 : qsort(blockgroups, nblockgroups, sizeof(IndexDeleteCounts),
8764 : bottomup_sort_and_shrink_cmp);
8765 3914 : reordereddeltids = palloc(delstate->ndeltids * sizeof(TM_IndexDelete));
8766 :
8767 3914 : nblockgroups = Min(BOTTOMUP_MAX_NBLOCKS, nblockgroups);
8768 : /* Determine number of favorable blocks at the start of final deltids */
8769 3914 : nblocksfavorable = bottomup_nblocksfavorable(blockgroups, nblockgroups,
8770 : delstate->deltids);
8771 :
8772 26176 : for (int b = 0; b < nblockgroups; b++)
8773 : {
8774 22262 : IndexDeleteCounts *group = blockgroups + b;
8775 22262 : TM_IndexDelete *firstdtid = delstate->deltids + group->ifirsttid;
8776 :
8777 22262 : memcpy(reordereddeltids + ncopied, firstdtid,
8778 22262 : sizeof(TM_IndexDelete) * group->ntids);
8779 22262 : ncopied += group->ntids;
8780 : }
8781 :
8782 : /* Copy final grouped and sorted TIDs back into start of caller's array */
8783 3914 : memcpy(delstate->deltids, reordereddeltids,
8784 : sizeof(TM_IndexDelete) * ncopied);
8785 3914 : delstate->ndeltids = ncopied;
8786 :
8787 3914 : pfree(reordereddeltids);
8788 3914 : pfree(blockgroups);
8789 :
8790 3914 : return nblocksfavorable;
8791 : }
8792 :
8793 : /*
8794 : * Perform XLogInsert for a heap-visible operation. 'block' is the block
8795 : * being marked all-visible, and vm_buffer is the buffer containing the
8796 : * corresponding visibility map block. Both should have already been modified
8797 : * and dirtied.
8798 : *
8799 : * snapshotConflictHorizon comes from the largest xmin on the page being
8800 : * marked all-visible. REDO routine uses it to generate recovery conflicts.
8801 : *
8802 : * If checksums or wal_log_hints are enabled, we may also generate a full-page
8803 : * image of heap_buffer. Otherwise, we optimize away the FPI (by specifying
8804 : * REGBUF_NO_IMAGE for the heap buffer), in which case the caller should *not*
8805 : * update the heap page's LSN.
8806 : */
8807 : XLogRecPtr
8808 89254 : log_heap_visible(Relation rel, Buffer heap_buffer, Buffer vm_buffer,
8809 : TransactionId snapshotConflictHorizon, uint8 vmflags)
8810 : {
8811 : xl_heap_visible xlrec;
8812 : XLogRecPtr recptr;
8813 : uint8 flags;
8814 :
8815 : Assert(BufferIsValid(heap_buffer));
8816 : Assert(BufferIsValid(vm_buffer));
8817 :
8818 89254 : xlrec.snapshotConflictHorizon = snapshotConflictHorizon;
8819 89254 : xlrec.flags = vmflags;
8820 89254 : if (RelationIsAccessibleInLogicalDecoding(rel))
8821 250 : xlrec.flags |= VISIBILITYMAP_XLOG_CATALOG_REL;
8822 89254 : XLogBeginInsert();
8823 89254 : XLogRegisterData(&xlrec, SizeOfHeapVisible);
8824 :
8825 89254 : XLogRegisterBuffer(0, vm_buffer, 0);
8826 :
8827 89254 : flags = REGBUF_STANDARD;
8828 89254 : if (!XLogHintBitIsNeeded())
8829 6788 : flags |= REGBUF_NO_IMAGE;
8830 89254 : XLogRegisterBuffer(1, heap_buffer, flags);
8831 :
8832 89254 : recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_VISIBLE);
8833 :
8834 89254 : return recptr;
8835 : }
8836 :
8837 : /*
8838 : * Perform XLogInsert for a heap-update operation. Caller must already
8839 : * have modified the buffer(s) and marked them dirty.
8840 : */
8841 : static XLogRecPtr
8842 583504 : log_heap_update(Relation reln, Buffer oldbuf,
8843 : Buffer newbuf, HeapTuple oldtup, HeapTuple newtup,
8844 : HeapTuple old_key_tuple,
8845 : bool all_visible_cleared, bool new_all_visible_cleared)
8846 : {
8847 : xl_heap_update xlrec;
8848 : xl_heap_header xlhdr;
8849 : xl_heap_header xlhdr_idx;
8850 : uint8 info;
8851 : uint16 prefix_suffix[2];
8852 583504 : uint16 prefixlen = 0,
8853 583504 : suffixlen = 0;
8854 : XLogRecPtr recptr;
8855 583504 : Page page = BufferGetPage(newbuf);
8856 583504 : bool need_tuple_data = RelationIsLogicallyLogged(reln);
8857 : bool init;
8858 : int bufflags;
8859 :
8860 : /* Caller should not call me on a non-WAL-logged relation */
8861 : Assert(RelationNeedsWAL(reln));
8862 :
8863 583504 : XLogBeginInsert();
8864 :
8865 583504 : if (HeapTupleIsHeapOnly(newtup))
8866 283376 : info = XLOG_HEAP_HOT_UPDATE;
8867 : else
8868 300128 : info = XLOG_HEAP_UPDATE;
8869 :
8870 : /*
8871 : * If the old and new tuple are on the same page, we only need to log the
8872 : * parts of the new tuple that were changed. That saves on the amount of
8873 : * WAL we need to write. Currently, we just count any unchanged bytes in
8874 : * the beginning and end of the tuple. That's quick to check, and
8875 : * perfectly covers the common case that only one field is updated.
8876 : *
8877 : * We could do this even if the old and new tuple are on different pages,
8878 : * but only if we don't make a full-page image of the old page, which is
8879 : * difficult to know in advance. Also, if the old tuple is corrupt for
8880 : * some reason, it would allow the corruption to propagate the new page,
8881 : * so it seems best to avoid. Under the general assumption that most
8882 : * updates tend to create the new tuple version on the same page, there
8883 : * isn't much to be gained by doing this across pages anyway.
8884 : *
8885 : * Skip this if we're taking a full-page image of the new page, as we
8886 : * don't include the new tuple in the WAL record in that case. Also
8887 : * disable if wal_level='logical', as logical decoding needs to be able to
8888 : * read the new tuple in whole from the WAL record alone.
8889 : */
8890 583504 : if (oldbuf == newbuf && !need_tuple_data &&
8891 283216 : !XLogCheckBufferNeedsBackup(newbuf))
8892 : {
8893 282048 : char *oldp = (char *) oldtup->t_data + oldtup->t_data->t_hoff;
8894 282048 : char *newp = (char *) newtup->t_data + newtup->t_data->t_hoff;
8895 282048 : int oldlen = oldtup->t_len - oldtup->t_data->t_hoff;
8896 282048 : int newlen = newtup->t_len - newtup->t_data->t_hoff;
8897 :
8898 : /* Check for common prefix between old and new tuple */
8899 23318950 : for (prefixlen = 0; prefixlen < Min(oldlen, newlen); prefixlen++)
8900 : {
8901 23268302 : if (newp[prefixlen] != oldp[prefixlen])
8902 231400 : break;
8903 : }
8904 :
8905 : /*
8906 : * Storing the length of the prefix takes 2 bytes, so we need to save
8907 : * at least 3 bytes or there's no point.
8908 : */
8909 282048 : if (prefixlen < 3)
8910 44166 : prefixlen = 0;
8911 :
8912 : /* Same for suffix */
8913 9071456 : for (suffixlen = 0; suffixlen < Min(oldlen, newlen) - prefixlen; suffixlen++)
8914 : {
8915 9020320 : if (newp[newlen - suffixlen - 1] != oldp[oldlen - suffixlen - 1])
8916 230912 : break;
8917 : }
8918 282048 : if (suffixlen < 3)
8919 69906 : suffixlen = 0;
8920 : }
8921 :
8922 : /* Prepare main WAL data chain */
8923 583504 : xlrec.flags = 0;
8924 583504 : if (all_visible_cleared)
8925 3050 : xlrec.flags |= XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED;
8926 583504 : if (new_all_visible_cleared)
8927 1536 : xlrec.flags |= XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED;
8928 583504 : if (prefixlen > 0)
8929 237882 : xlrec.flags |= XLH_UPDATE_PREFIX_FROM_OLD;
8930 583504 : if (suffixlen > 0)
8931 212142 : xlrec.flags |= XLH_UPDATE_SUFFIX_FROM_OLD;
8932 583504 : if (need_tuple_data)
8933 : {
8934 94050 : xlrec.flags |= XLH_UPDATE_CONTAINS_NEW_TUPLE;
8935 94050 : if (old_key_tuple)
8936 : {
8937 290 : if (reln->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
8938 128 : xlrec.flags |= XLH_UPDATE_CONTAINS_OLD_TUPLE;
8939 : else
8940 162 : xlrec.flags |= XLH_UPDATE_CONTAINS_OLD_KEY;
8941 : }
8942 : }
8943 :
8944 : /* If new tuple is the single and first tuple on page... */
8945 590346 : if (ItemPointerGetOffsetNumber(&(newtup->t_self)) == FirstOffsetNumber &&
8946 6842 : PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
8947 : {
8948 6514 : info |= XLOG_HEAP_INIT_PAGE;
8949 6514 : init = true;
8950 : }
8951 : else
8952 576990 : init = false;
8953 :
8954 : /* Prepare WAL data for the old page */
8955 583504 : xlrec.old_offnum = ItemPointerGetOffsetNumber(&oldtup->t_self);
8956 583504 : xlrec.old_xmax = HeapTupleHeaderGetRawXmax(oldtup->t_data);
8957 1167008 : xlrec.old_infobits_set = compute_infobits(oldtup->t_data->t_infomask,
8958 583504 : oldtup->t_data->t_infomask2);
8959 :
8960 : /* Prepare WAL data for the new page */
8961 583504 : xlrec.new_offnum = ItemPointerGetOffsetNumber(&newtup->t_self);
8962 583504 : xlrec.new_xmax = HeapTupleHeaderGetRawXmax(newtup->t_data);
8963 :
8964 583504 : bufflags = REGBUF_STANDARD;
8965 583504 : if (init)
8966 6514 : bufflags |= REGBUF_WILL_INIT;
8967 583504 : if (need_tuple_data)
8968 94050 : bufflags |= REGBUF_KEEP_DATA;
8969 :
8970 583504 : XLogRegisterBuffer(0, newbuf, bufflags);
8971 583504 : if (oldbuf != newbuf)
8972 276398 : XLogRegisterBuffer(1, oldbuf, REGBUF_STANDARD);
8973 :
8974 583504 : XLogRegisterData(&xlrec, SizeOfHeapUpdate);
8975 :
8976 : /*
8977 : * Prepare WAL data for the new tuple.
8978 : */
8979 583504 : if (prefixlen > 0 || suffixlen > 0)
8980 : {
8981 281130 : if (prefixlen > 0 && suffixlen > 0)
8982 : {
8983 168894 : prefix_suffix[0] = prefixlen;
8984 168894 : prefix_suffix[1] = suffixlen;
8985 168894 : XLogRegisterBufData(0, &prefix_suffix, sizeof(uint16) * 2);
8986 : }
8987 112236 : else if (prefixlen > 0)
8988 : {
8989 68988 : XLogRegisterBufData(0, &prefixlen, sizeof(uint16));
8990 : }
8991 : else
8992 : {
8993 43248 : XLogRegisterBufData(0, &suffixlen, sizeof(uint16));
8994 : }
8995 : }
8996 :
8997 583504 : xlhdr.t_infomask2 = newtup->t_data->t_infomask2;
8998 583504 : xlhdr.t_infomask = newtup->t_data->t_infomask;
8999 583504 : xlhdr.t_hoff = newtup->t_data->t_hoff;
9000 : Assert(SizeofHeapTupleHeader + prefixlen + suffixlen <= newtup->t_len);
9001 :
9002 : /*
9003 : * PG73FORMAT: write bitmap [+ padding] [+ oid] + data
9004 : *
9005 : * The 'data' doesn't include the common prefix or suffix.
9006 : */
9007 583504 : XLogRegisterBufData(0, &xlhdr, SizeOfHeapHeader);
9008 583504 : if (prefixlen == 0)
9009 : {
9010 345622 : XLogRegisterBufData(0,
9011 345622 : (char *) newtup->t_data + SizeofHeapTupleHeader,
9012 345622 : newtup->t_len - SizeofHeapTupleHeader - suffixlen);
9013 : }
9014 : else
9015 : {
9016 : /*
9017 : * Have to write the null bitmap and data after the common prefix as
9018 : * two separate rdata entries.
9019 : */
9020 : /* bitmap [+ padding] [+ oid] */
9021 237882 : if (newtup->t_data->t_hoff - SizeofHeapTupleHeader > 0)
9022 : {
9023 237882 : XLogRegisterBufData(0,
9024 237882 : (char *) newtup->t_data + SizeofHeapTupleHeader,
9025 237882 : newtup->t_data->t_hoff - SizeofHeapTupleHeader);
9026 : }
9027 :
9028 : /* data after common prefix */
9029 237882 : XLogRegisterBufData(0,
9030 237882 : (char *) newtup->t_data + newtup->t_data->t_hoff + prefixlen,
9031 237882 : newtup->t_len - newtup->t_data->t_hoff - prefixlen - suffixlen);
9032 : }
9033 :
9034 : /* We need to log a tuple identity */
9035 583504 : if (need_tuple_data && old_key_tuple)
9036 : {
9037 : /* don't really need this, but its more comfy to decode */
9038 290 : xlhdr_idx.t_infomask2 = old_key_tuple->t_data->t_infomask2;
9039 290 : xlhdr_idx.t_infomask = old_key_tuple->t_data->t_infomask;
9040 290 : xlhdr_idx.t_hoff = old_key_tuple->t_data->t_hoff;
9041 :
9042 290 : XLogRegisterData(&xlhdr_idx, SizeOfHeapHeader);
9043 :
9044 : /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
9045 290 : XLogRegisterData((char *) old_key_tuple->t_data + SizeofHeapTupleHeader,
9046 290 : old_key_tuple->t_len - SizeofHeapTupleHeader);
9047 : }
9048 :
9049 : /* filtering by origin on a row level is much more efficient */
9050 583504 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
9051 :
9052 583504 : recptr = XLogInsert(RM_HEAP_ID, info);
9053 :
9054 583504 : return recptr;
9055 : }
9056 :
9057 : /*
9058 : * Perform XLogInsert of an XLOG_HEAP2_NEW_CID record
9059 : *
9060 : * This is only used in wal_level >= WAL_LEVEL_LOGICAL, and only for catalog
9061 : * tuples.
9062 : */
9063 : static XLogRecPtr
9064 48678 : log_heap_new_cid(Relation relation, HeapTuple tup)
9065 : {
9066 : xl_heap_new_cid xlrec;
9067 :
9068 : XLogRecPtr recptr;
9069 48678 : HeapTupleHeader hdr = tup->t_data;
9070 :
9071 : Assert(ItemPointerIsValid(&tup->t_self));
9072 : Assert(tup->t_tableOid != InvalidOid);
9073 :
9074 48678 : xlrec.top_xid = GetTopTransactionId();
9075 48678 : xlrec.target_locator = relation->rd_locator;
9076 48678 : xlrec.target_tid = tup->t_self;
9077 :
9078 : /*
9079 : * If the tuple got inserted & deleted in the same TX we definitely have a
9080 : * combo CID, set cmin and cmax.
9081 : */
9082 48678 : if (hdr->t_infomask & HEAP_COMBOCID)
9083 : {
9084 : Assert(!(hdr->t_infomask & HEAP_XMAX_INVALID));
9085 : Assert(!HeapTupleHeaderXminInvalid(hdr));
9086 4040 : xlrec.cmin = HeapTupleHeaderGetCmin(hdr);
9087 4040 : xlrec.cmax = HeapTupleHeaderGetCmax(hdr);
9088 4040 : xlrec.combocid = HeapTupleHeaderGetRawCommandId(hdr);
9089 : }
9090 : /* No combo CID, so only cmin or cmax can be set by this TX */
9091 : else
9092 : {
9093 : /*
9094 : * Tuple inserted.
9095 : *
9096 : * We need to check for LOCK ONLY because multixacts might be
9097 : * transferred to the new tuple in case of FOR KEY SHARE updates in
9098 : * which case there will be an xmax, although the tuple just got
9099 : * inserted.
9100 : */
9101 58170 : if (hdr->t_infomask & HEAP_XMAX_INVALID ||
9102 13532 : HEAP_XMAX_IS_LOCKED_ONLY(hdr->t_infomask))
9103 : {
9104 31108 : xlrec.cmin = HeapTupleHeaderGetRawCommandId(hdr);
9105 31108 : xlrec.cmax = InvalidCommandId;
9106 : }
9107 : /* Tuple from a different tx updated or deleted. */
9108 : else
9109 : {
9110 13530 : xlrec.cmin = InvalidCommandId;
9111 13530 : xlrec.cmax = HeapTupleHeaderGetRawCommandId(hdr);
9112 : }
9113 44638 : xlrec.combocid = InvalidCommandId;
9114 : }
9115 :
9116 : /*
9117 : * Note that we don't need to register the buffer here, because this
9118 : * operation does not modify the page. The insert/update/delete that
9119 : * called us certainly did, but that's WAL-logged separately.
9120 : */
9121 48678 : XLogBeginInsert();
9122 48678 : XLogRegisterData(&xlrec, SizeOfHeapNewCid);
9123 :
9124 : /* will be looked at irrespective of origin */
9125 :
9126 48678 : recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_NEW_CID);
9127 :
9128 48678 : return recptr;
9129 : }
9130 :
9131 : /*
9132 : * Build a heap tuple representing the configured REPLICA IDENTITY to represent
9133 : * the old tuple in an UPDATE or DELETE.
9134 : *
9135 : * Returns NULL if there's no need to log an identity or if there's no suitable
9136 : * key defined.
9137 : *
9138 : * Pass key_required true if any replica identity columns changed value, or if
9139 : * any of them have any external data. Delete must always pass true.
9140 : *
9141 : * *copy is set to true if the returned tuple is a modified copy rather than
9142 : * the same tuple that was passed in.
9143 : */
9144 : static HeapTuple
9145 3581056 : ExtractReplicaIdentity(Relation relation, HeapTuple tp, bool key_required,
9146 : bool *copy)
9147 : {
9148 3581056 : TupleDesc desc = RelationGetDescr(relation);
9149 3581056 : char replident = relation->rd_rel->relreplident;
9150 : Bitmapset *idattrs;
9151 : HeapTuple key_tuple;
9152 : bool nulls[MaxHeapAttributeNumber];
9153 : Datum values[MaxHeapAttributeNumber];
9154 :
9155 3581056 : *copy = false;
9156 :
9157 3581056 : if (!RelationIsLogicallyLogged(relation))
9158 3380466 : return NULL;
9159 :
9160 200590 : if (replident == REPLICA_IDENTITY_NOTHING)
9161 462 : return NULL;
9162 :
9163 200128 : if (replident == REPLICA_IDENTITY_FULL)
9164 : {
9165 : /*
9166 : * When logging the entire old tuple, it very well could contain
9167 : * toasted columns. If so, force them to be inlined.
9168 : */
9169 388 : if (HeapTupleHasExternal(tp))
9170 : {
9171 8 : *copy = true;
9172 8 : tp = toast_flatten_tuple(tp, desc);
9173 : }
9174 388 : return tp;
9175 : }
9176 :
9177 : /* if the key isn't required and we're only logging the key, we're done */
9178 199740 : if (!key_required)
9179 93760 : return NULL;
9180 :
9181 : /* find out the replica identity columns */
9182 105980 : idattrs = RelationGetIndexAttrBitmap(relation,
9183 : INDEX_ATTR_BITMAP_IDENTITY_KEY);
9184 :
9185 : /*
9186 : * If there's no defined replica identity columns, treat as !key_required.
9187 : * (This case should not be reachable from heap_update, since that should
9188 : * calculate key_required accurately. But heap_delete just passes
9189 : * constant true for key_required, so we can hit this case in deletes.)
9190 : */
9191 105980 : if (bms_is_empty(idattrs))
9192 12042 : return NULL;
9193 :
9194 : /*
9195 : * Construct a new tuple containing only the replica identity columns,
9196 : * with nulls elsewhere. While we're at it, assert that the replica
9197 : * identity columns aren't null.
9198 : */
9199 93938 : heap_deform_tuple(tp, desc, values, nulls);
9200 :
9201 301802 : for (int i = 0; i < desc->natts; i++)
9202 : {
9203 207864 : if (bms_is_member(i + 1 - FirstLowInvalidHeapAttributeNumber,
9204 : idattrs))
9205 : Assert(!nulls[i]);
9206 : else
9207 113902 : nulls[i] = true;
9208 : }
9209 :
9210 93938 : key_tuple = heap_form_tuple(desc, values, nulls);
9211 93938 : *copy = true;
9212 :
9213 93938 : bms_free(idattrs);
9214 :
9215 : /*
9216 : * If the tuple, which by here only contains indexed columns, still has
9217 : * toasted columns, force them to be inlined. This is somewhat unlikely
9218 : * since there's limits on the size of indexed columns, so we don't
9219 : * duplicate toast_flatten_tuple()s functionality in the above loop over
9220 : * the indexed columns, even if it would be more efficient.
9221 : */
9222 93938 : if (HeapTupleHasExternal(key_tuple))
9223 : {
9224 8 : HeapTuple oldtup = key_tuple;
9225 :
9226 8 : key_tuple = toast_flatten_tuple(oldtup, desc);
9227 8 : heap_freetuple(oldtup);
9228 : }
9229 :
9230 93938 : return key_tuple;
9231 : }
9232 :
9233 : /*
9234 : * HeapCheckForSerializableConflictOut
9235 : * We are reading a tuple. If it's not visible, there may be a
9236 : * rw-conflict out with the inserter. Otherwise, if it is visible to us
9237 : * but has been deleted, there may be a rw-conflict out with the deleter.
9238 : *
9239 : * We will determine the top level xid of the writing transaction with which
9240 : * we may be in conflict, and ask CheckForSerializableConflictOut() to check
9241 : * for overlap with our own transaction.
9242 : *
9243 : * This function should be called just about anywhere in heapam.c where a
9244 : * tuple has been read. The caller must hold at least a shared lock on the
9245 : * buffer, because this function might set hint bits on the tuple. There is
9246 : * currently no known reason to call this function from an index AM.
9247 : */
9248 : void
9249 63136180 : HeapCheckForSerializableConflictOut(bool visible, Relation relation,
9250 : HeapTuple tuple, Buffer buffer,
9251 : Snapshot snapshot)
9252 : {
9253 : TransactionId xid;
9254 : HTSV_Result htsvResult;
9255 :
9256 63136180 : if (!CheckForSerializableConflictOutNeeded(relation, snapshot))
9257 63085510 : return;
9258 :
9259 : /*
9260 : * Check to see whether the tuple has been written to by a concurrent
9261 : * transaction, either to create it not visible to us, or to delete it
9262 : * while it is visible to us. The "visible" bool indicates whether the
9263 : * tuple is visible to us, while HeapTupleSatisfiesVacuum checks what else
9264 : * is going on with it.
9265 : *
9266 : * In the event of a concurrently inserted tuple that also happens to have
9267 : * been concurrently updated (by a separate transaction), the xmin of the
9268 : * tuple will be used -- not the updater's xid.
9269 : */
9270 50670 : htsvResult = HeapTupleSatisfiesVacuum(tuple, TransactionXmin, buffer);
9271 50670 : switch (htsvResult)
9272 : {
9273 49066 : case HEAPTUPLE_LIVE:
9274 49066 : if (visible)
9275 49040 : return;
9276 26 : xid = HeapTupleHeaderGetXmin(tuple->t_data);
9277 26 : break;
9278 704 : case HEAPTUPLE_RECENTLY_DEAD:
9279 : case HEAPTUPLE_DELETE_IN_PROGRESS:
9280 704 : if (visible)
9281 562 : xid = HeapTupleHeaderGetUpdateXid(tuple->t_data);
9282 : else
9283 142 : xid = HeapTupleHeaderGetXmin(tuple->t_data);
9284 :
9285 704 : if (TransactionIdPrecedes(xid, TransactionXmin))
9286 : {
9287 : /* This is like the HEAPTUPLE_DEAD case */
9288 : Assert(!visible);
9289 124 : return;
9290 : }
9291 580 : break;
9292 652 : case HEAPTUPLE_INSERT_IN_PROGRESS:
9293 652 : xid = HeapTupleHeaderGetXmin(tuple->t_data);
9294 652 : break;
9295 248 : case HEAPTUPLE_DEAD:
9296 : Assert(!visible);
9297 248 : return;
9298 0 : default:
9299 :
9300 : /*
9301 : * The only way to get to this default clause is if a new value is
9302 : * added to the enum type without adding it to this switch
9303 : * statement. That's a bug, so elog.
9304 : */
9305 0 : elog(ERROR, "unrecognized return value from HeapTupleSatisfiesVacuum: %u", htsvResult);
9306 :
9307 : /*
9308 : * In spite of having all enum values covered and calling elog on
9309 : * this default, some compilers think this is a code path which
9310 : * allows xid to be used below without initialization. Silence
9311 : * that warning.
9312 : */
9313 : xid = InvalidTransactionId;
9314 : }
9315 :
9316 : Assert(TransactionIdIsValid(xid));
9317 : Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
9318 :
9319 : /*
9320 : * Find top level xid. Bail out if xid is too early to be a conflict, or
9321 : * if it's our own xid.
9322 : */
9323 1258 : if (TransactionIdEquals(xid, GetTopTransactionIdIfAny()))
9324 124 : return;
9325 1134 : xid = SubTransGetTopmostTransaction(xid);
9326 1134 : if (TransactionIdPrecedes(xid, TransactionXmin))
9327 0 : return;
9328 :
9329 1134 : CheckForSerializableConflictOut(relation, xid, snapshot);
9330 : }
|