Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * heapam.c
4 : * heap access method code
5 : *
6 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/access/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 "executor/instrument_node.h"
47 : #include "pgstat.h"
48 : #include "port/pg_bitutils.h"
49 : #include "storage/lmgr.h"
50 : #include "storage/predicate.h"
51 : #include "storage/proc.h"
52 : #include "storage/procarray.h"
53 : #include "utils/datum.h"
54 : #include "utils/injection_point.h"
55 : #include "utils/inval.h"
56 : #include "utils/spccache.h"
57 : #include "utils/syscache.h"
58 :
59 :
60 : static HeapTuple heap_prepare_insert(Relation relation, HeapTuple tup,
61 : TransactionId xid, CommandId cid, uint32 options);
62 : static XLogRecPtr log_heap_update(Relation reln, Buffer oldbuf,
63 : Buffer newbuf, HeapTuple oldtup,
64 : HeapTuple newtup, HeapTuple old_key_tuple,
65 : bool all_visible_cleared, bool new_all_visible_cleared,
66 : bool walLogical);
67 : #ifdef USE_ASSERT_CHECKING
68 : static void check_lock_if_inplace_updateable_rel(Relation relation,
69 : const ItemPointerData *otid,
70 : HeapTuple newtup);
71 : static void check_inplace_rel_lock(HeapTuple oldtup);
72 : #endif
73 : static Bitmapset *HeapDetermineColumnsInfo(Relation relation,
74 : Bitmapset *interesting_cols,
75 : Bitmapset *external_cols,
76 : HeapTuple oldtup, HeapTuple newtup,
77 : bool *has_external);
78 : static bool heap_acquire_tuplock(Relation relation, const ItemPointerData *tid,
79 : LockTupleMode mode, LockWaitPolicy wait_policy,
80 : bool *have_tuple_lock);
81 : static inline BlockNumber heapgettup_advance_block(HeapScanDesc scan,
82 : BlockNumber block,
83 : ScanDirection dir);
84 : static pg_noinline BlockNumber heapgettup_initial_block(HeapScanDesc scan,
85 : ScanDirection dir);
86 : static void compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
87 : uint16 old_infomask2, TransactionId add_to_xmax,
88 : LockTupleMode mode, bool is_update,
89 : TransactionId *result_xmax, uint16 *result_infomask,
90 : uint16 *result_infomask2);
91 : static TM_Result heap_lock_updated_tuple(Relation rel,
92 : uint16 prior_infomask,
93 : TransactionId prior_raw_xmax,
94 : const ItemPointerData *prior_ctid,
95 : TransactionId xid,
96 : LockTupleMode mode);
97 : static void GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
98 : uint16 *new_infomask2);
99 : static TransactionId MultiXactIdGetUpdateXid(TransactionId xmax,
100 : uint16 t_infomask);
101 : static bool DoesMultiXactIdConflict(MultiXactId multi, uint16 infomask,
102 : LockTupleMode lockmode, bool *current_is_member);
103 : static void MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
104 : Relation rel, const ItemPointerData *ctid, XLTW_Oper oper,
105 : int *remaining);
106 : static bool ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
107 : uint16 infomask, Relation rel, int *remaining,
108 : bool logLockFailure);
109 : static void index_delete_sort(TM_IndexDeleteOp *delstate);
110 : static int bottomup_sort_and_shrink(TM_IndexDeleteOp *delstate);
111 : static XLogRecPtr log_heap_new_cid(Relation relation, HeapTuple tup);
112 : static HeapTuple ExtractReplicaIdentity(Relation relation, HeapTuple tp, bool key_required,
113 : bool *copy);
114 :
115 :
116 : /*
117 : * This table lists the heavyweight lock mode that corresponds to each tuple
118 : * lock mode, as well as one or two corresponding MultiXactStatus values:
119 : * .lockstatus to merely lock tuples, and .updstatus to update them. The
120 : * latter is set to -1 if the corresponding tuple lock mode does not allow
121 : * updating tuples -- see get_mxact_status_for_lock().
122 : *
123 : * These interact with InplaceUpdateTupleLock, an alias for ExclusiveLock.
124 : *
125 : * Don't look at lockstatus/updstatus directly! Use get_mxact_status_for_lock
126 : * instead.
127 : */
128 : static const struct
129 : {
130 : LOCKMODE hwlock;
131 : int lockstatus;
132 : int updstatus;
133 : } tupleLockExtraInfo[] =
134 :
135 : {
136 : [LockTupleKeyShare] = {
137 : .hwlock = AccessShareLock,
138 : .lockstatus = MultiXactStatusForKeyShare,
139 : /* KeyShare does not allow updating tuples */
140 : .updstatus = -1
141 : },
142 : [LockTupleShare] = {
143 : .hwlock = RowShareLock,
144 : .lockstatus = MultiXactStatusForShare,
145 : /* Share does not allow updating tuples */
146 : .updstatus = -1
147 : },
148 : [LockTupleNoKeyExclusive] = {
149 : .hwlock = ExclusiveLock,
150 : .lockstatus = MultiXactStatusForNoKeyUpdate,
151 : .updstatus = MultiXactStatusNoKeyUpdate
152 : },
153 : [LockTupleExclusive] = {
154 : .hwlock = AccessExclusiveLock,
155 : .lockstatus = MultiXactStatusForUpdate,
156 : .updstatus = MultiXactStatusUpdate
157 : }
158 : };
159 :
160 : /* Get the LOCKMODE for a given MultiXactStatus */
161 : #define LOCKMODE_from_mxstatus(status) \
162 : (tupleLockExtraInfo[TUPLOCK_from_mxstatus((status))].hwlock)
163 :
164 : /*
165 : * Acquire heavyweight locks on tuples, using a LockTupleMode strength value.
166 : * This is more readable than having every caller translate it to lock.h's
167 : * LOCKMODE.
168 : */
169 : #define LockTupleTuplock(rel, tup, mode) \
170 : LockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
171 : #define UnlockTupleTuplock(rel, tup, mode) \
172 : UnlockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
173 : #define ConditionalLockTupleTuplock(rel, tup, mode, log) \
174 : ConditionalLockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock, (log))
175 :
176 : #ifdef USE_PREFETCH
177 : /*
178 : * heap_index_delete_tuples and index_delete_prefetch_buffer use this
179 : * structure to coordinate prefetching activity
180 : */
181 : typedef struct
182 : {
183 : BlockNumber cur_hblkno;
184 : int next_item;
185 : int ndeltids;
186 : TM_IndexDelete *deltids;
187 : } IndexDeletePrefetchState;
188 : #endif
189 :
190 : /* heap_index_delete_tuples bottom-up index deletion costing constants */
191 : #define BOTTOMUP_MAX_NBLOCKS 6
192 : #define BOTTOMUP_TOLERANCE_NBLOCKS 3
193 :
194 : /*
195 : * heap_index_delete_tuples uses this when determining which heap blocks it
196 : * must visit to help its bottom-up index deletion caller
197 : */
198 : typedef struct IndexDeleteCounts
199 : {
200 : int16 npromisingtids; /* Number of "promising" TIDs in group */
201 : int16 ntids; /* Number of TIDs in group */
202 : int16 ifirsttid; /* Offset to group's first deltid */
203 : } IndexDeleteCounts;
204 :
205 : /*
206 : * This table maps tuple lock strength values for each particular
207 : * MultiXactStatus value.
208 : */
209 : static const int MultiXactStatusLock[MaxMultiXactStatus + 1] =
210 : {
211 : LockTupleKeyShare, /* ForKeyShare */
212 : LockTupleShare, /* ForShare */
213 : LockTupleNoKeyExclusive, /* ForNoKeyUpdate */
214 : LockTupleExclusive, /* ForUpdate */
215 : LockTupleNoKeyExclusive, /* NoKeyUpdate */
216 : LockTupleExclusive /* Update */
217 : };
218 :
219 : /* Get the LockTupleMode for a given MultiXactStatus */
220 : #define TUPLOCK_from_mxstatus(status) \
221 : (MultiXactStatusLock[(status)])
222 :
223 : /*
224 : * Check that we have a valid snapshot if we might need TOAST access.
225 : */
226 : static inline void
227 16323188 : AssertHasSnapshotForToast(Relation rel)
228 : {
229 : #ifdef USE_ASSERT_CHECKING
230 :
231 : /* bootstrap mode in particular breaks this rule */
232 : if (!IsNormalProcessingMode())
233 : return;
234 :
235 : /* if the relation doesn't have a TOAST table, we are good */
236 : if (!OidIsValid(rel->rd_rel->reltoastrelid))
237 : return;
238 :
239 : Assert(HaveRegisteredOrActiveSnapshot());
240 :
241 : #endif /* USE_ASSERT_CHECKING */
242 16323188 : }
243 :
244 : /* ----------------------------------------------------------------
245 : * heap support routines
246 : * ----------------------------------------------------------------
247 : */
248 :
249 : /*
250 : * Streaming read API callback for parallel sequential scans. Returns the next
251 : * block the caller wants from the read stream or InvalidBlockNumber when done.
252 : */
253 : static BlockNumber
254 146260 : heap_scan_stream_read_next_parallel(ReadStream *stream,
255 : void *callback_private_data,
256 : void *per_buffer_data)
257 : {
258 146260 : HeapScanDesc scan = (HeapScanDesc) callback_private_data;
259 :
260 : Assert(ScanDirectionIsForward(scan->rs_dir));
261 : Assert(scan->rs_base.rs_parallel);
262 :
263 146260 : if (unlikely(!scan->rs_inited))
264 : {
265 : /* parallel scan */
266 2783 : table_block_parallelscan_startblock_init(scan->rs_base.rs_rd,
267 2783 : scan->rs_parallelworkerdata,
268 2783 : (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel,
269 : scan->rs_startblock,
270 : scan->rs_numblocks);
271 :
272 : /* may return InvalidBlockNumber if there are no more blocks */
273 5566 : scan->rs_prefetch_block = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
274 2783 : scan->rs_parallelworkerdata,
275 2783 : (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel);
276 2783 : scan->rs_inited = true;
277 : }
278 : else
279 : {
280 143477 : scan->rs_prefetch_block = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
281 143477 : scan->rs_parallelworkerdata, (ParallelBlockTableScanDesc)
282 143477 : scan->rs_base.rs_parallel);
283 : }
284 :
285 146260 : return scan->rs_prefetch_block;
286 : }
287 :
288 : /*
289 : * Streaming read API callback for serial sequential and TID range scans.
290 : * Returns the next block the caller wants from the read stream or
291 : * InvalidBlockNumber when done.
292 : */
293 : static BlockNumber
294 5061409 : heap_scan_stream_read_next_serial(ReadStream *stream,
295 : void *callback_private_data,
296 : void *per_buffer_data)
297 : {
298 5061409 : HeapScanDesc scan = (HeapScanDesc) callback_private_data;
299 :
300 5061409 : if (unlikely(!scan->rs_inited))
301 : {
302 1320561 : scan->rs_prefetch_block = heapgettup_initial_block(scan, scan->rs_dir);
303 1320561 : scan->rs_inited = true;
304 : }
305 : else
306 3740848 : scan->rs_prefetch_block = heapgettup_advance_block(scan,
307 : scan->rs_prefetch_block,
308 : scan->rs_dir);
309 :
310 5061409 : return scan->rs_prefetch_block;
311 : }
312 :
313 : /*
314 : * Read stream API callback for bitmap heap scans.
315 : * Returns the next block the caller wants from the read stream or
316 : * InvalidBlockNumber when done.
317 : */
318 : static BlockNumber
319 263837 : bitmapheap_stream_read_next(ReadStream *pgsr, void *private_data,
320 : void *per_buffer_data)
321 : {
322 263837 : TBMIterateResult *tbmres = per_buffer_data;
323 263837 : BitmapHeapScanDesc bscan = (BitmapHeapScanDesc) private_data;
324 263837 : HeapScanDesc hscan = (HeapScanDesc) bscan;
325 263837 : TableScanDesc sscan = &hscan->rs_base;
326 :
327 : for (;;)
328 : {
329 263837 : CHECK_FOR_INTERRUPTS();
330 :
331 : /* no more entries in the bitmap */
332 263837 : if (!tbm_iterate(&sscan->st.rs_tbmiterator, tbmres))
333 15968 : return InvalidBlockNumber;
334 :
335 : /*
336 : * Ignore any claimed entries past what we think is the end of the
337 : * relation. It may have been extended after the start of our scan (we
338 : * only hold an AccessShareLock, and it could be inserts from this
339 : * backend). We don't take this optimization in SERIALIZABLE
340 : * isolation though, as we need to examine all invisible tuples
341 : * reachable by the index.
342 : */
343 247869 : if (!IsolationIsSerializable() &&
344 247760 : tbmres->blockno >= hscan->rs_nblocks)
345 0 : continue;
346 :
347 247869 : return tbmres->blockno;
348 : }
349 :
350 : /* not reachable */
351 : Assert(false);
352 : }
353 :
354 : /* ----------------
355 : * initscan - scan code common to heap_beginscan and heap_rescan
356 : * ----------------
357 : */
358 : static void
359 1352856 : initscan(HeapScanDesc scan, ScanKey key, bool keep_startblock)
360 : {
361 1352856 : ParallelBlockTableScanDesc bpscan = NULL;
362 : bool allow_strat;
363 : bool allow_sync;
364 :
365 : /*
366 : * Determine the number of blocks we have to scan.
367 : *
368 : * It is sufficient to do this once at scan start, since any tuples added
369 : * while the scan is in progress will be invisible to my snapshot anyway.
370 : * (That is not true when using a non-MVCC snapshot. However, we couldn't
371 : * guarantee to return tuples added after scan start anyway, since they
372 : * might go into pages we already scanned. To guarantee consistent
373 : * results for a non-MVCC snapshot, the caller must hold some higher-level
374 : * lock that ensures the interesting tuple(s) won't change.)
375 : */
376 1352856 : if (scan->rs_base.rs_parallel != NULL)
377 : {
378 4678 : bpscan = (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
379 4678 : scan->rs_nblocks = bpscan->phs_nblocks;
380 : }
381 : else
382 1348178 : scan->rs_nblocks = RelationGetNumberOfBlocks(scan->rs_base.rs_rd);
383 :
384 : /*
385 : * If the table is large relative to NBuffers, use a bulk-read access
386 : * strategy and enable synchronized scanning (see syncscan.c). Although
387 : * the thresholds for these features could be different, we make them the
388 : * same so that there are only two behaviors to tune rather than four.
389 : * (However, some callers need to be able to disable one or both of these
390 : * behaviors, independently of the size of the table; also there is a GUC
391 : * variable that can disable synchronized scanning.)
392 : *
393 : * Note that table_block_parallelscan_initialize has a very similar test;
394 : * if you change this, consider changing that one, too.
395 : */
396 1352854 : if (!RelationUsesLocalBuffers(scan->rs_base.rs_rd) &&
397 1343045 : scan->rs_nblocks > NBuffers / 4)
398 : {
399 14862 : allow_strat = (scan->rs_base.rs_flags & SO_ALLOW_STRAT) != 0;
400 14862 : allow_sync = (scan->rs_base.rs_flags & SO_ALLOW_SYNC) != 0;
401 : }
402 : else
403 1337992 : allow_strat = allow_sync = false;
404 :
405 1352854 : if (allow_strat)
406 : {
407 : /* During a rescan, keep the previous strategy object. */
408 13507 : if (scan->rs_strategy == NULL)
409 13312 : scan->rs_strategy = GetAccessStrategy(BAS_BULKREAD);
410 : }
411 : else
412 : {
413 1339347 : if (scan->rs_strategy != NULL)
414 0 : FreeAccessStrategy(scan->rs_strategy);
415 1339347 : scan->rs_strategy = NULL;
416 : }
417 :
418 1352854 : if (scan->rs_base.rs_parallel != NULL)
419 : {
420 : /* For parallel scan, believe whatever ParallelTableScanDesc says. */
421 4678 : if (scan->rs_base.rs_parallel->phs_syncscan)
422 5 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
423 : else
424 4673 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
425 :
426 : /*
427 : * If not rescanning, initialize the startblock. Finding the actual
428 : * start location is done in table_block_parallelscan_startblock_init,
429 : * based on whether an alternative start location has been set with
430 : * heap_setscanlimits, or using the syncscan location, when syncscan
431 : * is enabled.
432 : */
433 4678 : if (!keep_startblock)
434 4526 : scan->rs_startblock = InvalidBlockNumber;
435 : }
436 : else
437 : {
438 1348176 : if (keep_startblock)
439 : {
440 : /*
441 : * When rescanning, we want to keep the previous startblock
442 : * setting, so that rewinding a cursor doesn't generate surprising
443 : * results. Reset the active syncscan setting, though.
444 : */
445 862854 : if (allow_sync && synchronize_seqscans)
446 50 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
447 : else
448 862804 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
449 : }
450 485322 : else if (allow_sync && synchronize_seqscans)
451 : {
452 78 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
453 78 : scan->rs_startblock = ss_get_location(scan->rs_base.rs_rd, scan->rs_nblocks);
454 : }
455 : else
456 : {
457 485244 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
458 485244 : scan->rs_startblock = 0;
459 : }
460 : }
461 :
462 1352854 : scan->rs_numblocks = InvalidBlockNumber;
463 1352854 : scan->rs_inited = false;
464 1352854 : scan->rs_ctup.t_data = NULL;
465 1352854 : ItemPointerSetInvalid(&scan->rs_ctup.t_self);
466 1352854 : scan->rs_cbuf = InvalidBuffer;
467 1352854 : scan->rs_cblock = InvalidBlockNumber;
468 1352854 : scan->rs_ntuples = 0;
469 1352854 : scan->rs_cindex = 0;
470 :
471 : /*
472 : * Initialize to ForwardScanDirection because it is most common and
473 : * because heap scans go forward before going backward (e.g. CURSORs).
474 : */
475 1352854 : scan->rs_dir = ForwardScanDirection;
476 1352854 : scan->rs_prefetch_block = InvalidBlockNumber;
477 :
478 : /* page-at-a-time fields are always invalid when not rs_inited */
479 :
480 : /*
481 : * copy the scan key, if appropriate
482 : */
483 1352854 : if (key != NULL && scan->rs_base.rs_nkeys > 0)
484 268742 : memcpy(scan->rs_base.rs_key, key, scan->rs_base.rs_nkeys * sizeof(ScanKeyData));
485 :
486 : /*
487 : * Currently, we only have a stats counter for sequential heap scans (but
488 : * e.g for bitmap scans the underlying bitmap index scans will be counted,
489 : * and for sample scans we update stats for tuple fetches).
490 : */
491 1352854 : if (scan->rs_base.rs_flags & SO_TYPE_SEQSCAN)
492 1323902 : pgstat_count_heap_scan(scan->rs_base.rs_rd);
493 1352854 : }
494 :
495 : /*
496 : * heap_setscanlimits - restrict range of a heapscan
497 : *
498 : * startBlk is the page to start at
499 : * numBlks is number of pages to scan (InvalidBlockNumber means "all")
500 : */
501 : void
502 3279 : heap_setscanlimits(TableScanDesc sscan, BlockNumber startBlk, BlockNumber numBlks)
503 : {
504 3279 : HeapScanDesc scan = (HeapScanDesc) sscan;
505 :
506 : Assert(!scan->rs_inited); /* else too late to change */
507 : /* else rs_startblock is significant */
508 : Assert(!(scan->rs_base.rs_flags & SO_ALLOW_SYNC));
509 :
510 : /* Check startBlk is valid (but allow case of zero blocks...) */
511 : Assert(startBlk == 0 || startBlk < scan->rs_nblocks);
512 :
513 3279 : scan->rs_startblock = startBlk;
514 3279 : scan->rs_numblocks = numBlks;
515 3279 : }
516 :
517 : /*
518 : * Per-tuple loop for heap_prepare_pagescan(). Pulled out so it can be called
519 : * multiple times, with constant arguments for all_visible,
520 : * check_serializable.
521 : */
522 : pg_attribute_always_inline
523 : static int
524 3753111 : page_collect_tuples(HeapScanDesc scan, Snapshot snapshot,
525 : Page page, Buffer buffer,
526 : BlockNumber block, int lines,
527 : bool all_visible, bool check_serializable)
528 : {
529 3753111 : Oid relid = RelationGetRelid(scan->rs_base.rs_rd);
530 3753111 : int ntup = 0;
531 3753111 : int nvis = 0;
532 : BatchMVCCState batchmvcc;
533 :
534 : /* page at a time should have been disabled otherwise */
535 : Assert(IsMVCCSnapshot(snapshot));
536 :
537 : /* first find all tuples on the page */
538 202231837 : for (OffsetNumber lineoff = FirstOffsetNumber; lineoff <= lines; lineoff++)
539 : {
540 198478726 : ItemId lpp = PageGetItemId(page, lineoff);
541 : HeapTuple tup;
542 :
543 198478726 : if (unlikely(!ItemIdIsNormal(lpp)))
544 52366959 : continue;
545 :
546 : /*
547 : * If the page is not all-visible or we need to check serializability,
548 : * maintain enough state to be able to refind the tuple efficiently,
549 : * without again first needing to fetch the item and then via that the
550 : * tuple.
551 : */
552 146111767 : if (!all_visible || check_serializable)
553 : {
554 83301592 : tup = &batchmvcc.tuples[ntup];
555 :
556 83301592 : tup->t_data = (HeapTupleHeader) PageGetItem(page, lpp);
557 83301592 : tup->t_len = ItemIdGetLength(lpp);
558 83301592 : tup->t_tableOid = relid;
559 83301592 : ItemPointerSet(&(tup->t_self), block, lineoff);
560 : }
561 :
562 : /*
563 : * If the page is all visible, these fields otherwise won't be
564 : * populated in loop below.
565 : */
566 146111767 : if (all_visible)
567 : {
568 62810175 : if (check_serializable)
569 : {
570 0 : batchmvcc.visible[ntup] = true;
571 : }
572 62810175 : scan->rs_vistuples[ntup] = lineoff;
573 : }
574 :
575 146111767 : ntup++;
576 : }
577 :
578 : Assert(ntup <= MaxHeapTuplesPerPage);
579 :
580 : /*
581 : * Unless the page is all visible, test visibility for all tuples one go.
582 : * That is considerably more efficient than calling
583 : * HeapTupleSatisfiesMVCC() one-by-one.
584 : */
585 3753111 : if (all_visible)
586 1334987 : nvis = ntup;
587 : else
588 2418124 : nvis = HeapTupleSatisfiesMVCCBatch(snapshot, buffer,
589 : ntup,
590 : &batchmvcc,
591 2418124 : scan->rs_vistuples);
592 :
593 : /*
594 : * So far we don't have batch API for testing serializabilty, so do so
595 : * one-by-one.
596 : */
597 3753111 : if (check_serializable)
598 : {
599 2087 : for (int i = 0; i < ntup; i++)
600 : {
601 1464 : HeapCheckForSerializableConflictOut(batchmvcc.visible[i],
602 : scan->rs_base.rs_rd,
603 : &batchmvcc.tuples[i],
604 : buffer, snapshot);
605 : }
606 : }
607 :
608 3753103 : return nvis;
609 : }
610 :
611 : /*
612 : * heap_prepare_pagescan - Prepare current scan page to be scanned in pagemode
613 : *
614 : * Preparation currently consists of 1. prune the scan's rs_cbuf page, and 2.
615 : * fill the rs_vistuples[] array with the OffsetNumbers of visible tuples.
616 : */
617 : void
618 3753111 : heap_prepare_pagescan(TableScanDesc sscan)
619 : {
620 3753111 : HeapScanDesc scan = (HeapScanDesc) sscan;
621 3753111 : Buffer buffer = scan->rs_cbuf;
622 3753111 : BlockNumber block = scan->rs_cblock;
623 : Snapshot snapshot;
624 : Page page;
625 : int lines;
626 : bool all_visible;
627 : bool check_serializable;
628 :
629 : Assert(BufferGetBlockNumber(buffer) == block);
630 :
631 : /* ensure we're not accidentally being used when not in pagemode */
632 : Assert(scan->rs_base.rs_flags & SO_ALLOW_PAGEMODE);
633 3753111 : snapshot = scan->rs_base.rs_snapshot;
634 :
635 : /*
636 : * Prune and repair fragmentation for the whole page, if possible.
637 : */
638 3753111 : heap_page_prune_opt(scan->rs_base.rs_rd, buffer, &scan->rs_vmbuffer,
639 3753111 : sscan->rs_flags & SO_HINT_REL_READ_ONLY);
640 :
641 : /*
642 : * We must hold share lock on the buffer content while examining tuple
643 : * visibility. Afterwards, however, the tuples we have found to be
644 : * visible are guaranteed good as long as we hold the buffer pin.
645 : */
646 3753111 : LockBuffer(buffer, BUFFER_LOCK_SHARE);
647 :
648 3753111 : page = BufferGetPage(buffer);
649 3753111 : lines = PageGetMaxOffsetNumber(page);
650 :
651 : /*
652 : * If the all-visible flag indicates that all tuples on the page are
653 : * visible to everyone, we can skip the per-tuple visibility tests.
654 : *
655 : * Note: In hot standby, a tuple that's already visible to all
656 : * transactions on the primary might still be invisible to a read-only
657 : * transaction in the standby. We partly handle this problem by tracking
658 : * the minimum xmin of visible tuples as the cut-off XID while marking a
659 : * page all-visible on the primary and WAL log that along with the
660 : * visibility map SET operation. In hot standby, we wait for (or abort)
661 : * all transactions that can potentially may not see one or more tuples on
662 : * the page. That's how index-only scans work fine in hot standby. A
663 : * crucial difference between index-only scans and heap scans is that the
664 : * index-only scan completely relies on the visibility map where as heap
665 : * scan looks at the page-level PD_ALL_VISIBLE flag. We are not sure if
666 : * the page-level flag can be trusted in the same way, because it might
667 : * get propagated somehow without being explicitly WAL-logged, e.g. via a
668 : * full page write. Until we can prove that beyond doubt, let's check each
669 : * tuple for visibility the hard way.
670 : */
671 3753111 : all_visible = PageIsAllVisible(page) && !snapshot->takenDuringRecovery;
672 : check_serializable =
673 3753111 : CheckForSerializableConflictOutNeeded(scan->rs_base.rs_rd, snapshot);
674 :
675 : /*
676 : * We call page_collect_tuples() with constant arguments, to get the
677 : * compiler to constant fold the constant arguments. Separate calls with
678 : * constant arguments, rather than variables, are needed on several
679 : * compilers to actually perform constant folding.
680 : */
681 3753111 : if (likely(all_visible))
682 : {
683 1334987 : if (likely(!check_serializable))
684 1334987 : scan->rs_ntuples = page_collect_tuples(scan, snapshot, page, buffer,
685 : block, lines, true, false);
686 : else
687 0 : scan->rs_ntuples = page_collect_tuples(scan, snapshot, page, buffer,
688 : block, lines, true, true);
689 : }
690 : else
691 : {
692 2418124 : if (likely(!check_serializable))
693 2417493 : scan->rs_ntuples = page_collect_tuples(scan, snapshot, page, buffer,
694 : block, lines, false, false);
695 : else
696 631 : scan->rs_ntuples = page_collect_tuples(scan, snapshot, page, buffer,
697 : block, lines, false, true);
698 : }
699 :
700 3753103 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
701 3753103 : }
702 :
703 : /*
704 : * heap_fetch_next_buffer - read and pin the next block from MAIN_FORKNUM.
705 : *
706 : * Read the next block of the scan relation from the read stream and save it
707 : * in the scan descriptor. It is already pinned.
708 : */
709 : static inline void
710 4989874 : heap_fetch_next_buffer(HeapScanDesc scan, ScanDirection dir)
711 : {
712 : Assert(scan->rs_read_stream);
713 :
714 : /* release previous scan buffer, if any */
715 4989874 : if (BufferIsValid(scan->rs_cbuf))
716 : {
717 3666530 : ReleaseBuffer(scan->rs_cbuf);
718 3666530 : scan->rs_cbuf = InvalidBuffer;
719 : }
720 :
721 : /*
722 : * Be sure to check for interrupts at least once per page. Checks at
723 : * higher code levels won't be able to stop a seqscan that encounters many
724 : * pages' worth of consecutive dead tuples.
725 : */
726 4989874 : CHECK_FOR_INTERRUPTS();
727 :
728 : /*
729 : * If the scan direction is changing, reset the prefetch block to the
730 : * current block. Otherwise, we will incorrectly prefetch the blocks
731 : * between the prefetch block and the current block again before
732 : * prefetching blocks in the new, correct scan direction.
733 : */
734 4989874 : if (unlikely(scan->rs_dir != dir))
735 : {
736 102 : scan->rs_prefetch_block = scan->rs_cblock;
737 102 : read_stream_reset(scan->rs_read_stream);
738 : }
739 :
740 4989874 : scan->rs_dir = dir;
741 :
742 4989874 : scan->rs_cbuf = read_stream_next_buffer(scan->rs_read_stream, NULL);
743 4989847 : if (BufferIsValid(scan->rs_cbuf))
744 3863573 : scan->rs_cblock = BufferGetBlockNumber(scan->rs_cbuf);
745 4989847 : }
746 :
747 : /*
748 : * heapgettup_initial_block - return the first BlockNumber to scan
749 : *
750 : * Returns InvalidBlockNumber when there are no blocks to scan. This can
751 : * occur with empty tables and in parallel scans when parallel workers get all
752 : * of the pages before we can get a chance to get our first page.
753 : */
754 : static pg_noinline BlockNumber
755 1320561 : heapgettup_initial_block(HeapScanDesc scan, ScanDirection dir)
756 : {
757 : Assert(!scan->rs_inited);
758 : Assert(scan->rs_base.rs_parallel == NULL);
759 :
760 : /* When there are no pages to scan, return InvalidBlockNumber */
761 1320561 : if (scan->rs_nblocks == 0 || scan->rs_numblocks == 0)
762 760754 : return InvalidBlockNumber;
763 :
764 559807 : if (ScanDirectionIsForward(dir))
765 : {
766 559765 : return scan->rs_startblock;
767 : }
768 : else
769 : {
770 : /*
771 : * Disable reporting to syncscan logic in a backwards scan; it's not
772 : * very likely anyone else is doing the same thing at the same time,
773 : * and much more likely that we'll just bollix things for forward
774 : * scanners.
775 : */
776 42 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
777 :
778 : /*
779 : * Start from last page of the scan. Ensure we take into account
780 : * rs_numblocks if it's been adjusted by heap_setscanlimits().
781 : */
782 42 : if (scan->rs_numblocks != InvalidBlockNumber)
783 4 : return (scan->rs_startblock + scan->rs_numblocks - 1) % scan->rs_nblocks;
784 :
785 38 : if (scan->rs_startblock > 0)
786 0 : return scan->rs_startblock - 1;
787 :
788 38 : return scan->rs_nblocks - 1;
789 : }
790 : }
791 :
792 :
793 : /*
794 : * heapgettup_start_page - helper function for heapgettup()
795 : *
796 : * Return the next page to scan based on the scan->rs_cbuf and set *linesleft
797 : * to the number of tuples on this page. Also set *lineoff to the first
798 : * offset to scan with forward scans getting the first offset and backward
799 : * getting the final offset on the page.
800 : */
801 : static Page
802 116146 : heapgettup_start_page(HeapScanDesc scan, ScanDirection dir, int *linesleft,
803 : OffsetNumber *lineoff)
804 : {
805 : Page page;
806 :
807 : Assert(scan->rs_inited);
808 : Assert(BufferIsValid(scan->rs_cbuf));
809 :
810 : /* Caller is responsible for ensuring buffer is locked if needed */
811 116146 : page = BufferGetPage(scan->rs_cbuf);
812 :
813 116146 : *linesleft = PageGetMaxOffsetNumber(page) - FirstOffsetNumber + 1;
814 :
815 116146 : if (ScanDirectionIsForward(dir))
816 116146 : *lineoff = FirstOffsetNumber;
817 : else
818 0 : *lineoff = (OffsetNumber) (*linesleft);
819 :
820 : /* lineoff now references the physically previous or next tid */
821 116146 : return page;
822 : }
823 :
824 :
825 : /*
826 : * heapgettup_continue_page - helper function for heapgettup()
827 : *
828 : * Return the next page to scan based on the scan->rs_cbuf and set *linesleft
829 : * to the number of tuples left to scan on this page. Also set *lineoff to
830 : * the next offset to scan according to the ScanDirection in 'dir'.
831 : */
832 : static inline Page
833 9258225 : heapgettup_continue_page(HeapScanDesc scan, ScanDirection dir, int *linesleft,
834 : OffsetNumber *lineoff)
835 : {
836 : Page page;
837 :
838 : Assert(scan->rs_inited);
839 : Assert(BufferIsValid(scan->rs_cbuf));
840 :
841 : /* Caller is responsible for ensuring buffer is locked if needed */
842 9258225 : page = BufferGetPage(scan->rs_cbuf);
843 :
844 9258225 : if (ScanDirectionIsForward(dir))
845 : {
846 9258225 : *lineoff = OffsetNumberNext(scan->rs_coffset);
847 9258225 : *linesleft = PageGetMaxOffsetNumber(page) - (*lineoff) + 1;
848 : }
849 : else
850 : {
851 : /*
852 : * The previous returned tuple may have been vacuumed since the
853 : * previous scan when we use a non-MVCC snapshot, so we must
854 : * re-establish the lineoff <= PageGetMaxOffsetNumber(page) invariant
855 : */
856 0 : *lineoff = Min(PageGetMaxOffsetNumber(page), OffsetNumberPrev(scan->rs_coffset));
857 0 : *linesleft = *lineoff;
858 : }
859 :
860 : /* lineoff now references the physically previous or next tid */
861 9258225 : return page;
862 : }
863 :
864 : /*
865 : * heapgettup_advance_block - helper for heap_fetch_next_buffer()
866 : *
867 : * Given the current block number, the scan direction, and various information
868 : * contained in the scan descriptor, calculate the BlockNumber to scan next
869 : * and return it. If there are no further blocks to scan, return
870 : * InvalidBlockNumber to indicate this fact to the caller.
871 : *
872 : * This should not be called to determine the initial block number -- only for
873 : * subsequent blocks.
874 : *
875 : * This also adjusts rs_numblocks when a limit has been imposed by
876 : * heap_setscanlimits().
877 : */
878 : static inline BlockNumber
879 3740848 : heapgettup_advance_block(HeapScanDesc scan, BlockNumber block, ScanDirection dir)
880 : {
881 : Assert(scan->rs_base.rs_parallel == NULL);
882 :
883 3740848 : if (likely(ScanDirectionIsForward(dir)))
884 : {
885 3740770 : block++;
886 :
887 : /* wrap back to the start of the heap */
888 3740770 : if (block >= scan->rs_nblocks)
889 438525 : block = 0;
890 :
891 : /*
892 : * Report our new scan position for synchronization purposes. We don't
893 : * do that when moving backwards, however. That would just mess up any
894 : * other forward-moving scanners.
895 : *
896 : * Note: we do this before checking for end of scan so that the final
897 : * state of the position hint is back at the start of the rel. That's
898 : * not strictly necessary, but otherwise when you run the same query
899 : * multiple times the starting position would shift a little bit
900 : * backwards on every invocation, which is confusing. We don't
901 : * guarantee any specific ordering in general, though.
902 : */
903 3740770 : if (scan->rs_base.rs_flags & SO_ALLOW_SYNC)
904 38364 : ss_report_location(scan->rs_base.rs_rd, block);
905 :
906 : /* we're done if we're back at where we started */
907 3740770 : if (block == scan->rs_startblock)
908 438475 : return InvalidBlockNumber;
909 :
910 : /* check if the limit imposed by heap_setscanlimits() is met */
911 3302295 : if (scan->rs_numblocks != InvalidBlockNumber)
912 : {
913 2818 : if (--scan->rs_numblocks == 0)
914 1590 : return InvalidBlockNumber;
915 : }
916 :
917 3300705 : return block;
918 : }
919 : else
920 : {
921 : /* we're done if the last block is the start position */
922 78 : if (block == scan->rs_startblock)
923 78 : return InvalidBlockNumber;
924 :
925 : /* check if the limit imposed by heap_setscanlimits() is met */
926 0 : if (scan->rs_numblocks != InvalidBlockNumber)
927 : {
928 0 : if (--scan->rs_numblocks == 0)
929 0 : return InvalidBlockNumber;
930 : }
931 :
932 : /* wrap to the end of the heap when the last page was page 0 */
933 0 : if (block == 0)
934 0 : block = scan->rs_nblocks;
935 :
936 0 : block--;
937 :
938 0 : return block;
939 : }
940 : }
941 :
942 : /* ----------------
943 : * heapgettup - fetch next heap tuple
944 : *
945 : * Initialize the scan if not already done; then advance to the next
946 : * tuple as indicated by "dir"; return the next tuple in scan->rs_ctup,
947 : * or set scan->rs_ctup.t_data = NULL if no more tuples.
948 : *
949 : * Note: the reason nkeys/key are passed separately, even though they are
950 : * kept in the scan descriptor, is that the caller may not want us to check
951 : * the scankeys.
952 : *
953 : * Note: when we fall off the end of the scan in either direction, we
954 : * reset rs_inited. This means that a further request with the same
955 : * scan direction will restart the scan, which is a bit odd, but a
956 : * request with the opposite scan direction will start a fresh scan
957 : * in the proper direction. The latter is required behavior for cursors,
958 : * while the former case is generally undefined behavior in Postgres
959 : * so we don't care too much.
960 : * ----------------
961 : */
962 : static void
963 9285289 : heapgettup(HeapScanDesc scan,
964 : ScanDirection dir,
965 : int nkeys,
966 : ScanKey key)
967 : {
968 9285289 : HeapTuple tuple = &(scan->rs_ctup);
969 : Page page;
970 : OffsetNumber lineoff;
971 : int linesleft;
972 :
973 9285289 : if (likely(scan->rs_inited))
974 : {
975 : /* continue from previously returned page/tuple */
976 9258225 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
977 9258225 : page = heapgettup_continue_page(scan, dir, &linesleft, &lineoff);
978 9258225 : goto continue_page;
979 : }
980 :
981 : /*
982 : * advance the scan until we find a qualifying tuple or run out of stuff
983 : * to scan
984 : */
985 : while (true)
986 : {
987 142328 : heap_fetch_next_buffer(scan, dir);
988 :
989 : /* did we run out of blocks to scan? */
990 142328 : if (!BufferIsValid(scan->rs_cbuf))
991 26182 : break;
992 :
993 : Assert(BufferGetBlockNumber(scan->rs_cbuf) == scan->rs_cblock);
994 :
995 116146 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
996 116146 : page = heapgettup_start_page(scan, dir, &linesleft, &lineoff);
997 9374371 : continue_page:
998 :
999 : /*
1000 : * Only continue scanning the page while we have lines left.
1001 : *
1002 : * Note that this protects us from accessing line pointers past
1003 : * PageGetMaxOffsetNumber(); both for forward scans when we resume the
1004 : * table scan, and for when we start scanning a new page.
1005 : */
1006 9441864 : for (; linesleft > 0; linesleft--, lineoff += dir)
1007 : {
1008 : bool visible;
1009 9326600 : ItemId lpp = PageGetItemId(page, lineoff);
1010 :
1011 9326600 : if (!ItemIdIsNormal(lpp))
1012 45754 : continue;
1013 :
1014 9280846 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lpp);
1015 9280846 : tuple->t_len = ItemIdGetLength(lpp);
1016 9280846 : ItemPointerSet(&(tuple->t_self), scan->rs_cblock, lineoff);
1017 :
1018 9280846 : visible = HeapTupleSatisfiesVisibility(tuple,
1019 : scan->rs_base.rs_snapshot,
1020 : scan->rs_cbuf);
1021 :
1022 9280846 : HeapCheckForSerializableConflictOut(visible, scan->rs_base.rs_rd,
1023 : tuple, scan->rs_cbuf,
1024 : scan->rs_base.rs_snapshot);
1025 :
1026 : /* skip tuples not visible to this snapshot */
1027 9280846 : if (!visible)
1028 7362 : continue;
1029 :
1030 : /* skip any tuples that don't match the scan key */
1031 9273484 : if (key != NULL &&
1032 15106 : !HeapKeyTest(tuple, RelationGetDescr(scan->rs_base.rs_rd),
1033 : nkeys, key))
1034 14377 : continue;
1035 :
1036 9259107 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
1037 9259107 : scan->rs_coffset = lineoff;
1038 9259107 : return;
1039 : }
1040 :
1041 : /*
1042 : * if we get here, it means we've exhausted the items on this page and
1043 : * it's time to move to the next.
1044 : */
1045 115264 : LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
1046 : }
1047 :
1048 : /* end of scan */
1049 26182 : if (BufferIsValid(scan->rs_cbuf))
1050 0 : ReleaseBuffer(scan->rs_cbuf);
1051 :
1052 26182 : scan->rs_cbuf = InvalidBuffer;
1053 26182 : scan->rs_cblock = InvalidBlockNumber;
1054 26182 : scan->rs_prefetch_block = InvalidBlockNumber;
1055 26182 : tuple->t_data = NULL;
1056 26182 : scan->rs_inited = false;
1057 : }
1058 :
1059 : /* ----------------
1060 : * heapgettup_pagemode - fetch next heap tuple in page-at-a-time mode
1061 : *
1062 : * Same API as heapgettup, but used in page-at-a-time mode
1063 : *
1064 : * The internal logic is much the same as heapgettup's too, but there are some
1065 : * differences: we do not take the buffer content lock (that only needs to
1066 : * happen inside heap_prepare_pagescan), and we iterate through just the
1067 : * tuples listed in rs_vistuples[] rather than all tuples on the page. Notice
1068 : * that lineindex is 0-based, where the corresponding loop variable lineoff in
1069 : * heapgettup is 1-based.
1070 : * ----------------
1071 : */
1072 : static void
1073 69984125 : heapgettup_pagemode(HeapScanDesc scan,
1074 : ScanDirection dir,
1075 : int nkeys,
1076 : ScanKey key)
1077 : {
1078 69984125 : HeapTuple tuple = &(scan->rs_ctup);
1079 : Page page;
1080 : uint32 lineindex;
1081 : uint32 linesleft;
1082 :
1083 69984125 : if (likely(scan->rs_inited))
1084 : {
1085 : /* continue from previously returned page/tuple */
1086 68687845 : page = BufferGetPage(scan->rs_cbuf);
1087 :
1088 68687845 : lineindex = scan->rs_cindex + dir;
1089 68687845 : if (ScanDirectionIsForward(dir))
1090 68687407 : linesleft = scan->rs_ntuples - lineindex;
1091 : else
1092 438 : linesleft = scan->rs_cindex;
1093 : /* lineindex now references the next or previous visible tid */
1094 :
1095 68687845 : goto continue_page;
1096 : }
1097 :
1098 : /*
1099 : * advance the scan until we find a qualifying tuple or run out of stuff
1100 : * to scan
1101 : */
1102 : while (true)
1103 : {
1104 4847546 : heap_fetch_next_buffer(scan, dir);
1105 :
1106 : /* did we run out of blocks to scan? */
1107 4847519 : if (!BufferIsValid(scan->rs_cbuf))
1108 1100092 : break;
1109 :
1110 : Assert(BufferGetBlockNumber(scan->rs_cbuf) == scan->rs_cblock);
1111 :
1112 : /* prune the page and determine visible tuple offsets */
1113 3747427 : heap_prepare_pagescan((TableScanDesc) scan);
1114 3747419 : page = BufferGetPage(scan->rs_cbuf);
1115 3747419 : linesleft = scan->rs_ntuples;
1116 3747419 : lineindex = ScanDirectionIsForward(dir) ? 0 : linesleft - 1;
1117 :
1118 : /* block is the same for all tuples, set it once outside the loop */
1119 3747419 : ItemPointerSetBlockNumber(&tuple->t_self, scan->rs_cblock);
1120 :
1121 : /* lineindex now references the next or previous visible tid */
1122 72435264 : continue_page:
1123 :
1124 134158396 : for (; linesleft > 0; linesleft--, lineindex += dir)
1125 : {
1126 : ItemId lpp;
1127 : OffsetNumber lineoff;
1128 :
1129 : Assert(lineindex < scan->rs_ntuples);
1130 130607130 : lineoff = scan->rs_vistuples[lineindex];
1131 130607130 : lpp = PageGetItemId(page, lineoff);
1132 : Assert(ItemIdIsNormal(lpp));
1133 :
1134 130607130 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lpp);
1135 130607130 : tuple->t_len = ItemIdGetLength(lpp);
1136 130607130 : ItemPointerSetOffsetNumber(&tuple->t_self, lineoff);
1137 :
1138 : /* skip any tuples that don't match the scan key */
1139 130607130 : if (key != NULL &&
1140 62165641 : !HeapKeyTest(tuple, RelationGetDescr(scan->rs_base.rs_rd),
1141 : nkeys, key))
1142 61723132 : continue;
1143 :
1144 68883998 : scan->rs_cindex = lineindex;
1145 68883998 : return;
1146 : }
1147 : }
1148 :
1149 : /* end of scan */
1150 1100092 : if (BufferIsValid(scan->rs_cbuf))
1151 0 : ReleaseBuffer(scan->rs_cbuf);
1152 1100092 : scan->rs_cbuf = InvalidBuffer;
1153 1100092 : scan->rs_cblock = InvalidBlockNumber;
1154 1100092 : scan->rs_prefetch_block = InvalidBlockNumber;
1155 1100092 : tuple->t_data = NULL;
1156 1100092 : scan->rs_inited = false;
1157 : }
1158 :
1159 :
1160 : /* ----------------------------------------------------------------
1161 : * heap access method interface
1162 : * ----------------------------------------------------------------
1163 : */
1164 :
1165 :
1166 : TableScanDesc
1167 489850 : heap_beginscan(Relation relation, Snapshot snapshot,
1168 : int nkeys, ScanKey key,
1169 : ParallelTableScanDesc parallel_scan,
1170 : uint32 flags)
1171 : {
1172 : HeapScanDesc scan;
1173 :
1174 : /*
1175 : * increment relation ref count while scanning relation
1176 : *
1177 : * This is just to make really sure the relcache entry won't go away while
1178 : * the scan has a pointer to it. Caller should be holding the rel open
1179 : * anyway, so this is redundant in all normal scenarios...
1180 : */
1181 489850 : RelationIncrementReferenceCount(relation);
1182 :
1183 : /*
1184 : * allocate and initialize scan descriptor
1185 : */
1186 489850 : if (flags & SO_TYPE_BITMAPSCAN)
1187 : {
1188 13663 : BitmapHeapScanDesc bscan = palloc_object(BitmapHeapScanDescData);
1189 :
1190 : /*
1191 : * Bitmap Heap scans do not have any fields that a normal Heap Scan
1192 : * does not have, so no special initializations required here.
1193 : */
1194 13663 : scan = (HeapScanDesc) bscan;
1195 : }
1196 : else
1197 476187 : scan = (HeapScanDesc) palloc_object(HeapScanDescData);
1198 :
1199 489850 : scan->rs_base.rs_rd = relation;
1200 489850 : scan->rs_base.rs_snapshot = snapshot;
1201 489850 : scan->rs_base.rs_nkeys = nkeys;
1202 489850 : scan->rs_base.rs_flags = flags;
1203 489850 : scan->rs_base.rs_parallel = parallel_scan;
1204 489850 : scan->rs_base.rs_instrument = NULL;
1205 489850 : scan->rs_strategy = NULL; /* set in initscan */
1206 489850 : scan->rs_cbuf = InvalidBuffer;
1207 :
1208 : /*
1209 : * Disable page-at-a-time mode if it's not a MVCC-safe snapshot.
1210 : */
1211 489850 : if (!(snapshot && IsMVCCSnapshot(snapshot)))
1212 37976 : scan->rs_base.rs_flags &= ~SO_ALLOW_PAGEMODE;
1213 :
1214 : /* Check that a historic snapshot is not used for non-catalog tables */
1215 489850 : if (snapshot &&
1216 478938 : IsHistoricMVCCSnapshot(snapshot) &&
1217 729 : !RelationIsAccessibleInLogicalDecoding(relation))
1218 : {
1219 0 : ereport(ERROR,
1220 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
1221 : errmsg("cannot query non-catalog table \"%s\" during logical decoding",
1222 : RelationGetRelationName(relation))));
1223 : }
1224 :
1225 : /*
1226 : * For seqscan and sample scans in a serializable transaction, acquire a
1227 : * predicate lock on the entire relation. This is required not only to
1228 : * lock all the matching tuples, but also to conflict with new insertions
1229 : * into the table. In an indexscan, we take page locks on the index pages
1230 : * covering the range specified in the scan qual, but in a heap scan there
1231 : * is nothing more fine-grained to lock. A bitmap scan is a different
1232 : * story, there we have already scanned the index and locked the index
1233 : * pages covering the predicate. But in that case we still have to lock
1234 : * any matching heap tuples. For sample scan we could optimize the locking
1235 : * to be at least page-level granularity, but we'd need to add per-tuple
1236 : * locking for that.
1237 : */
1238 489850 : if (scan->rs_base.rs_flags & (SO_TYPE_SEQSCAN | SO_TYPE_SAMPLESCAN))
1239 : {
1240 : /*
1241 : * Ensure a missing snapshot is noticed reliably, even if the
1242 : * isolation mode means predicate locking isn't performed (and
1243 : * therefore the snapshot isn't used here).
1244 : */
1245 : Assert(snapshot);
1246 463479 : PredicateLockRelation(relation, snapshot);
1247 : }
1248 :
1249 : /* we only need to set this up once */
1250 489850 : scan->rs_ctup.t_tableOid = RelationGetRelid(relation);
1251 :
1252 : /*
1253 : * Allocate memory to keep track of page allocation for parallel workers
1254 : * when doing a parallel scan.
1255 : */
1256 489850 : if (parallel_scan != NULL)
1257 4526 : scan->rs_parallelworkerdata = palloc_object(ParallelBlockTableScanWorkerData);
1258 : else
1259 485324 : scan->rs_parallelworkerdata = NULL;
1260 :
1261 : /*
1262 : * we do this here instead of in initscan() because heap_rescan also calls
1263 : * initscan() and we don't want to allocate memory again
1264 : */
1265 489850 : if (nkeys > 0)
1266 268742 : scan->rs_base.rs_key = palloc_array(ScanKeyData, nkeys);
1267 : else
1268 221108 : scan->rs_base.rs_key = NULL;
1269 :
1270 489850 : initscan(scan, key, false);
1271 :
1272 489848 : scan->rs_read_stream = NULL;
1273 :
1274 : /*
1275 : * Set up a read stream for sequential scans and TID range scans. This
1276 : * should be done after initscan() because initscan() allocates the
1277 : * BufferAccessStrategy object passed to the read stream API.
1278 : */
1279 489848 : if (scan->rs_base.rs_flags & SO_TYPE_SEQSCAN ||
1280 26465 : scan->rs_base.rs_flags & SO_TYPE_TIDRANGESCAN)
1281 464697 : {
1282 : ReadStreamBlockNumberCB cb;
1283 :
1284 464697 : if (scan->rs_base.rs_parallel)
1285 4526 : cb = heap_scan_stream_read_next_parallel;
1286 : else
1287 460171 : cb = heap_scan_stream_read_next_serial;
1288 :
1289 : /* ---
1290 : * It is safe to use batchmode as the only locks taken by `cb`
1291 : * are never taken while waiting for IO:
1292 : * - SyncScanLock is used in the non-parallel case
1293 : * - in the parallel case, only spinlocks and atomics are used
1294 : * ---
1295 : */
1296 464697 : scan->rs_read_stream = read_stream_begin_relation(READ_STREAM_SEQUENTIAL |
1297 : READ_STREAM_USE_BATCHING,
1298 : scan->rs_strategy,
1299 : scan->rs_base.rs_rd,
1300 : MAIN_FORKNUM,
1301 : cb,
1302 : scan,
1303 : 0);
1304 : }
1305 25151 : else if (scan->rs_base.rs_flags & SO_TYPE_BITMAPSCAN)
1306 : {
1307 13663 : scan->rs_read_stream = read_stream_begin_relation(READ_STREAM_DEFAULT |
1308 : READ_STREAM_USE_BATCHING,
1309 : scan->rs_strategy,
1310 : scan->rs_base.rs_rd,
1311 : MAIN_FORKNUM,
1312 : bitmapheap_stream_read_next,
1313 : scan,
1314 : sizeof(TBMIterateResult));
1315 : }
1316 :
1317 : /* enable read stream instrumentation */
1318 489848 : if ((flags & SO_SCAN_INSTRUMENT) && (scan->rs_read_stream != NULL))
1319 : {
1320 8 : scan->rs_base.rs_instrument = palloc0_object(TableScanInstrumentation);
1321 8 : read_stream_enable_stats(scan->rs_read_stream,
1322 8 : &scan->rs_base.rs_instrument->io);
1323 : }
1324 :
1325 489848 : scan->rs_vmbuffer = InvalidBuffer;
1326 :
1327 489848 : return (TableScanDesc) scan;
1328 : }
1329 :
1330 : void
1331 863006 : heap_rescan(TableScanDesc sscan, ScanKey key, bool set_params,
1332 : bool allow_strat, bool allow_sync, bool allow_pagemode)
1333 : {
1334 863006 : HeapScanDesc scan = (HeapScanDesc) sscan;
1335 :
1336 863006 : if (set_params)
1337 : {
1338 19 : if (allow_strat)
1339 19 : scan->rs_base.rs_flags |= SO_ALLOW_STRAT;
1340 : else
1341 0 : scan->rs_base.rs_flags &= ~SO_ALLOW_STRAT;
1342 :
1343 19 : if (allow_sync)
1344 8 : scan->rs_base.rs_flags |= SO_ALLOW_SYNC;
1345 : else
1346 11 : scan->rs_base.rs_flags &= ~SO_ALLOW_SYNC;
1347 :
1348 19 : if (allow_pagemode && scan->rs_base.rs_snapshot &&
1349 19 : IsMVCCSnapshot(scan->rs_base.rs_snapshot))
1350 19 : scan->rs_base.rs_flags |= SO_ALLOW_PAGEMODE;
1351 : else
1352 0 : scan->rs_base.rs_flags &= ~SO_ALLOW_PAGEMODE;
1353 : }
1354 :
1355 : /*
1356 : * unpin scan buffers
1357 : */
1358 863006 : if (BufferIsValid(scan->rs_cbuf))
1359 : {
1360 3023 : ReleaseBuffer(scan->rs_cbuf);
1361 3023 : scan->rs_cbuf = InvalidBuffer;
1362 : }
1363 :
1364 863006 : if (BufferIsValid(scan->rs_vmbuffer))
1365 : {
1366 18 : ReleaseBuffer(scan->rs_vmbuffer);
1367 18 : scan->rs_vmbuffer = InvalidBuffer;
1368 : }
1369 :
1370 : /*
1371 : * SO_TYPE_BITMAPSCAN would be cleaned up here, but it does not hold any
1372 : * additional data vs a normal HeapScan
1373 : */
1374 :
1375 : /*
1376 : * The read stream is reset on rescan. This must be done before
1377 : * initscan(), as some state referred to by read_stream_reset() is reset
1378 : * in initscan().
1379 : */
1380 863006 : if (scan->rs_read_stream)
1381 862983 : read_stream_reset(scan->rs_read_stream);
1382 :
1383 : /*
1384 : * reinitialize scan descriptor
1385 : */
1386 863006 : initscan(scan, key, true);
1387 863006 : }
1388 :
1389 : void
1390 486565 : heap_endscan(TableScanDesc sscan)
1391 : {
1392 486565 : HeapScanDesc scan = (HeapScanDesc) sscan;
1393 :
1394 : /* Note: no locking manipulations needed */
1395 :
1396 : /*
1397 : * unpin scan buffers
1398 : */
1399 486565 : if (BufferIsValid(scan->rs_cbuf))
1400 191784 : ReleaseBuffer(scan->rs_cbuf);
1401 :
1402 486565 : if (BufferIsValid(scan->rs_vmbuffer))
1403 2899 : ReleaseBuffer(scan->rs_vmbuffer);
1404 :
1405 : /*
1406 : * Must free the read stream before freeing the BufferAccessStrategy.
1407 : */
1408 486565 : if (scan->rs_read_stream)
1409 475144 : read_stream_end(scan->rs_read_stream);
1410 :
1411 : /*
1412 : * decrement relation reference count and free scan descriptor storage
1413 : */
1414 486565 : RelationDecrementReferenceCount(scan->rs_base.rs_rd);
1415 :
1416 486565 : if (scan->rs_base.rs_key)
1417 268707 : pfree(scan->rs_base.rs_key);
1418 :
1419 486565 : if (scan->rs_strategy != NULL)
1420 13303 : FreeAccessStrategy(scan->rs_strategy);
1421 :
1422 486565 : if (scan->rs_parallelworkerdata != NULL)
1423 4526 : pfree(scan->rs_parallelworkerdata);
1424 :
1425 486565 : if (scan->rs_base.rs_flags & SO_TEMP_SNAPSHOT)
1426 51110 : UnregisterSnapshot(scan->rs_base.rs_snapshot);
1427 :
1428 486565 : if (scan->rs_base.rs_instrument)
1429 8 : pfree(scan->rs_base.rs_instrument);
1430 :
1431 486565 : pfree(scan);
1432 486565 : }
1433 :
1434 : HeapTuple
1435 11797065 : heap_getnext(TableScanDesc sscan, ScanDirection direction)
1436 : {
1437 11797065 : HeapScanDesc scan = (HeapScanDesc) sscan;
1438 :
1439 : /*
1440 : * This is still widely used directly, without going through table AM, so
1441 : * add a safety check. It's possible we should, at a later point,
1442 : * downgrade this to an assert. The reason for checking the AM routine,
1443 : * rather than the AM oid, is that this allows to write regression tests
1444 : * that create another AM reusing the heap handler.
1445 : */
1446 11797065 : if (unlikely(sscan->rs_rd->rd_tableam != GetHeapamTableAmRoutine()))
1447 0 : ereport(ERROR,
1448 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1449 : errmsg_internal("only heap AM is supported")));
1450 :
1451 : /* Note: no locking manipulations needed */
1452 :
1453 11797065 : if (scan->rs_base.rs_flags & SO_ALLOW_PAGEMODE)
1454 3094772 : heapgettup_pagemode(scan, direction,
1455 3094772 : scan->rs_base.rs_nkeys, scan->rs_base.rs_key);
1456 : else
1457 8702293 : heapgettup(scan, direction,
1458 8702293 : scan->rs_base.rs_nkeys, scan->rs_base.rs_key);
1459 :
1460 11797065 : if (scan->rs_ctup.t_data == NULL)
1461 82509 : return NULL;
1462 :
1463 : /*
1464 : * if we get here it means we have a new current scan tuple, so point to
1465 : * the proper return buffer and return the tuple.
1466 : */
1467 :
1468 11714556 : pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1469 :
1470 11714556 : return &scan->rs_ctup;
1471 : }
1472 :
1473 : bool
1474 67465713 : heap_getnextslot(TableScanDesc sscan, ScanDirection direction, TupleTableSlot *slot)
1475 : {
1476 67465713 : HeapScanDesc scan = (HeapScanDesc) sscan;
1477 :
1478 : /* Note: no locking manipulations needed */
1479 :
1480 67465713 : if (sscan->rs_flags & SO_ALLOW_PAGEMODE)
1481 66882717 : heapgettup_pagemode(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1482 : else
1483 582996 : heapgettup(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1484 :
1485 67465688 : if (scan->rs_ctup.t_data == NULL)
1486 : {
1487 1043627 : ExecClearTuple(slot);
1488 1043627 : return false;
1489 : }
1490 :
1491 : /*
1492 : * if we get here it means we have a new current scan tuple, so point to
1493 : * the proper return buffer and return the tuple.
1494 : */
1495 :
1496 66422061 : pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1497 :
1498 66422061 : ExecStoreBufferHeapTuple(&scan->rs_ctup, slot,
1499 : scan->rs_cbuf);
1500 66422061 : return true;
1501 : }
1502 :
1503 : void
1504 1374 : heap_set_tidrange(TableScanDesc sscan, ItemPointer mintid,
1505 : ItemPointer maxtid)
1506 : {
1507 1374 : HeapScanDesc scan = (HeapScanDesc) sscan;
1508 : BlockNumber startBlk;
1509 : BlockNumber numBlks;
1510 : ItemPointerData highestItem;
1511 : ItemPointerData lowestItem;
1512 :
1513 : /*
1514 : * For relations without any pages, we can simply leave the TID range
1515 : * unset. There will be no tuples to scan, therefore no tuples outside
1516 : * the given TID range.
1517 : */
1518 1374 : if (scan->rs_nblocks == 0)
1519 32 : return;
1520 :
1521 : /*
1522 : * Set up some ItemPointers which point to the first and last possible
1523 : * tuples in the heap.
1524 : */
1525 1366 : ItemPointerSet(&highestItem, scan->rs_nblocks - 1, MaxOffsetNumber);
1526 1366 : ItemPointerSet(&lowestItem, 0, FirstOffsetNumber);
1527 :
1528 : /*
1529 : * If the given maximum TID is below the highest possible TID in the
1530 : * relation, then restrict the range to that, otherwise we scan to the end
1531 : * of the relation.
1532 : */
1533 1366 : if (ItemPointerCompare(maxtid, &highestItem) < 0)
1534 172 : ItemPointerCopy(maxtid, &highestItem);
1535 :
1536 : /*
1537 : * If the given minimum TID is above the lowest possible TID in the
1538 : * relation, then restrict the range to only scan for TIDs above that.
1539 : */
1540 1366 : if (ItemPointerCompare(mintid, &lowestItem) > 0)
1541 1210 : ItemPointerCopy(mintid, &lowestItem);
1542 :
1543 : /*
1544 : * Check for an empty range and protect from would be negative results
1545 : * from the numBlks calculation below.
1546 : */
1547 1366 : if (ItemPointerCompare(&highestItem, &lowestItem) < 0)
1548 : {
1549 : /* Set an empty range of blocks to scan */
1550 24 : heap_setscanlimits(sscan, 0, 0);
1551 24 : return;
1552 : }
1553 :
1554 : /*
1555 : * Calculate the first block and the number of blocks we must scan. We
1556 : * could be more aggressive here and perform some more validation to try
1557 : * and further narrow the scope of blocks to scan by checking if the
1558 : * lowestItem has an offset above MaxOffsetNumber. In this case, we could
1559 : * advance startBlk by one. Likewise, if highestItem has an offset of 0
1560 : * we could scan one fewer blocks. However, such an optimization does not
1561 : * seem worth troubling over, currently.
1562 : */
1563 1342 : startBlk = ItemPointerGetBlockNumberNoCheck(&lowestItem);
1564 :
1565 1342 : numBlks = ItemPointerGetBlockNumberNoCheck(&highestItem) -
1566 1342 : ItemPointerGetBlockNumberNoCheck(&lowestItem) + 1;
1567 :
1568 : /* Set the start block and number of blocks to scan */
1569 1342 : heap_setscanlimits(sscan, startBlk, numBlks);
1570 :
1571 : /* Finally, set the TID range in sscan */
1572 1342 : ItemPointerCopy(&lowestItem, &sscan->st.tidrange.rs_mintid);
1573 1342 : ItemPointerCopy(&highestItem, &sscan->st.tidrange.rs_maxtid);
1574 : }
1575 :
1576 : bool
1577 6512 : heap_getnextslot_tidrange(TableScanDesc sscan, ScanDirection direction,
1578 : TupleTableSlot *slot)
1579 : {
1580 6512 : HeapScanDesc scan = (HeapScanDesc) sscan;
1581 6512 : ItemPointer mintid = &sscan->st.tidrange.rs_mintid;
1582 6512 : ItemPointer maxtid = &sscan->st.tidrange.rs_maxtid;
1583 :
1584 : /* Note: no locking manipulations needed */
1585 : for (;;)
1586 : {
1587 6636 : if (sscan->rs_flags & SO_ALLOW_PAGEMODE)
1588 6636 : heapgettup_pagemode(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1589 : else
1590 0 : heapgettup(scan, direction, sscan->rs_nkeys, sscan->rs_key);
1591 :
1592 6626 : if (scan->rs_ctup.t_data == NULL)
1593 : {
1594 138 : ExecClearTuple(slot);
1595 138 : return false;
1596 : }
1597 :
1598 : /*
1599 : * heap_set_tidrange will have used heap_setscanlimits to limit the
1600 : * range of pages we scan to only ones that can contain the TID range
1601 : * we're scanning for. Here we must filter out any tuples from these
1602 : * pages that are outside of that range.
1603 : */
1604 6488 : if (ItemPointerCompare(&scan->rs_ctup.t_self, mintid) < 0)
1605 : {
1606 124 : ExecClearTuple(slot);
1607 :
1608 : /*
1609 : * When scanning backwards, the TIDs will be in descending order.
1610 : * Future tuples in this direction will be lower still, so we can
1611 : * just return false to indicate there will be no more tuples.
1612 : */
1613 124 : if (ScanDirectionIsBackward(direction))
1614 0 : return false;
1615 :
1616 124 : continue;
1617 : }
1618 :
1619 : /*
1620 : * Likewise for the final page, we must filter out TIDs greater than
1621 : * maxtid.
1622 : */
1623 6364 : if (ItemPointerCompare(&scan->rs_ctup.t_self, maxtid) > 0)
1624 : {
1625 74 : ExecClearTuple(slot);
1626 :
1627 : /*
1628 : * When scanning forward, the TIDs will be in ascending order.
1629 : * Future tuples in this direction will be higher still, so we can
1630 : * just return false to indicate there will be no more tuples.
1631 : */
1632 74 : if (ScanDirectionIsForward(direction))
1633 74 : return false;
1634 0 : continue;
1635 : }
1636 :
1637 6290 : break;
1638 : }
1639 :
1640 : /*
1641 : * if we get here it means we have a new current scan tuple, so point to
1642 : * the proper return buffer and return the tuple.
1643 : */
1644 6290 : pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1645 :
1646 6290 : ExecStoreBufferHeapTuple(&scan->rs_ctup, slot, scan->rs_cbuf);
1647 6290 : return true;
1648 : }
1649 :
1650 : /*
1651 : * heap_fetch - retrieve tuple with given tid
1652 : *
1653 : * On entry, tuple->t_self is the TID to fetch. We pin the buffer holding
1654 : * the tuple, fill in the remaining fields of *tuple, and check the tuple
1655 : * against the specified snapshot.
1656 : *
1657 : * If successful (tuple found and passes snapshot time qual), then *userbuf
1658 : * is set to the buffer holding the tuple and true is returned. The caller
1659 : * must unpin the buffer when done with the tuple.
1660 : *
1661 : * If the tuple is not found (ie, item number references a deleted slot),
1662 : * then tuple->t_data is set to NULL, *userbuf is set to InvalidBuffer,
1663 : * and false is returned.
1664 : *
1665 : * If the tuple is found but fails the time qual check, then the behavior
1666 : * depends on the keep_buf parameter. If keep_buf is false, the results
1667 : * are the same as for the tuple-not-found case. If keep_buf is true,
1668 : * then tuple->t_data and *userbuf are returned as for the success case,
1669 : * and again the caller must unpin the buffer; but false is returned.
1670 : *
1671 : * heap_fetch does not follow HOT chains: only the exact TID requested will
1672 : * be fetched.
1673 : *
1674 : * It is somewhat inconsistent that we ereport() on invalid block number but
1675 : * return false on invalid item number. There are a couple of reasons though.
1676 : * One is that the caller can relatively easily check the block number for
1677 : * validity, but cannot check the item number without reading the page
1678 : * himself. Another is that when we are following a t_ctid link, we can be
1679 : * reasonably confident that the page number is valid (since VACUUM shouldn't
1680 : * truncate off the destination page without having killed the referencing
1681 : * tuple first), but the item number might well not be good.
1682 : */
1683 : bool
1684 2851882 : heap_fetch(Relation relation,
1685 : Snapshot snapshot,
1686 : HeapTuple tuple,
1687 : Buffer *userbuf,
1688 : bool keep_buf)
1689 : {
1690 2851882 : ItemPointer tid = &(tuple->t_self);
1691 : ItemId lp;
1692 : Buffer buffer;
1693 : Page page;
1694 : OffsetNumber offnum;
1695 : bool valid;
1696 :
1697 : /*
1698 : * Fetch and pin the appropriate page of the relation.
1699 : */
1700 2851882 : buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
1701 :
1702 : /*
1703 : * Need share lock on buffer to examine tuple commit status.
1704 : */
1705 2851874 : LockBuffer(buffer, BUFFER_LOCK_SHARE);
1706 2851874 : page = BufferGetPage(buffer);
1707 :
1708 : /*
1709 : * We'd better check for out-of-range offnum in case of VACUUM since the
1710 : * TID was obtained.
1711 : */
1712 2851874 : offnum = ItemPointerGetOffsetNumber(tid);
1713 2851874 : if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1714 : {
1715 4 : UnlockReleaseBuffer(buffer);
1716 4 : *userbuf = InvalidBuffer;
1717 4 : tuple->t_data = NULL;
1718 4 : return false;
1719 : }
1720 :
1721 : /*
1722 : * get the item line pointer corresponding to the requested tid
1723 : */
1724 2851870 : lp = PageGetItemId(page, offnum);
1725 :
1726 : /*
1727 : * Must check for deleted tuple.
1728 : */
1729 2851870 : if (!ItemIdIsNormal(lp))
1730 : {
1731 281 : UnlockReleaseBuffer(buffer);
1732 281 : *userbuf = InvalidBuffer;
1733 281 : tuple->t_data = NULL;
1734 281 : return false;
1735 : }
1736 :
1737 : /*
1738 : * fill in *tuple fields
1739 : */
1740 2851589 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
1741 2851589 : tuple->t_len = ItemIdGetLength(lp);
1742 2851589 : tuple->t_tableOid = RelationGetRelid(relation);
1743 :
1744 : /*
1745 : * check tuple visibility, then release lock
1746 : */
1747 2851589 : valid = HeapTupleSatisfiesVisibility(tuple, snapshot, buffer);
1748 :
1749 2851589 : if (valid)
1750 2851532 : PredicateLockTID(relation, &(tuple->t_self), snapshot,
1751 2851532 : HeapTupleHeaderGetXmin(tuple->t_data));
1752 :
1753 2851589 : HeapCheckForSerializableConflictOut(valid, relation, tuple, buffer, snapshot);
1754 :
1755 2851589 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1756 :
1757 2851589 : if (valid)
1758 : {
1759 : /*
1760 : * All checks passed, so return the tuple as valid. Caller is now
1761 : * responsible for releasing the buffer.
1762 : */
1763 2851532 : *userbuf = buffer;
1764 :
1765 2851532 : return true;
1766 : }
1767 :
1768 : /* Tuple failed time qual, but maybe caller wants to see it anyway. */
1769 57 : if (keep_buf)
1770 35 : *userbuf = buffer;
1771 : else
1772 : {
1773 22 : ReleaseBuffer(buffer);
1774 22 : *userbuf = InvalidBuffer;
1775 22 : tuple->t_data = NULL;
1776 : }
1777 :
1778 57 : return false;
1779 : }
1780 :
1781 : /*
1782 : * heap_get_latest_tid - get the latest tid of a specified tuple
1783 : *
1784 : * Actually, this gets the latest version that is visible according to the
1785 : * scan's snapshot. Create a scan using SnapshotDirty to get the very latest,
1786 : * possibly uncommitted version.
1787 : *
1788 : * *tid is both an input and an output parameter: it is updated to
1789 : * show the latest version of the row. Note that it will not be changed
1790 : * if no version of the row passes the snapshot test.
1791 : */
1792 : void
1793 207 : heap_get_latest_tid(TableScanDesc sscan,
1794 : ItemPointer tid)
1795 : {
1796 207 : Relation relation = sscan->rs_rd;
1797 207 : Snapshot snapshot = sscan->rs_snapshot;
1798 : ItemPointerData ctid;
1799 : TransactionId priorXmax;
1800 :
1801 : /*
1802 : * table_tuple_get_latest_tid() verified that the passed in tid is valid.
1803 : * Assume that t_ctid links are valid however - there shouldn't be invalid
1804 : * ones in the table.
1805 : */
1806 : Assert(ItemPointerIsValid(tid));
1807 :
1808 : /*
1809 : * Loop to chase down t_ctid links. At top of loop, ctid is the tuple we
1810 : * need to examine, and *tid is the TID we will return if ctid turns out
1811 : * to be bogus.
1812 : *
1813 : * Note that we will loop until we reach the end of the t_ctid chain.
1814 : * Depending on the snapshot passed, there might be at most one visible
1815 : * version of the row, but we don't try to optimize for that.
1816 : */
1817 207 : ctid = *tid;
1818 207 : priorXmax = InvalidTransactionId; /* cannot check first XMIN */
1819 : for (;;)
1820 64 : {
1821 : Buffer buffer;
1822 : Page page;
1823 : OffsetNumber offnum;
1824 : ItemId lp;
1825 : HeapTupleData tp;
1826 : bool valid;
1827 :
1828 : /*
1829 : * Read, pin, and lock the page.
1830 : */
1831 271 : buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&ctid));
1832 271 : LockBuffer(buffer, BUFFER_LOCK_SHARE);
1833 271 : page = BufferGetPage(buffer);
1834 :
1835 : /*
1836 : * Check for bogus item number. This is not treated as an error
1837 : * condition because it can happen while following a t_ctid link. We
1838 : * just assume that the prior tid is OK and return it unchanged.
1839 : */
1840 271 : offnum = ItemPointerGetOffsetNumber(&ctid);
1841 271 : if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1842 : {
1843 0 : UnlockReleaseBuffer(buffer);
1844 0 : break;
1845 : }
1846 271 : lp = PageGetItemId(page, offnum);
1847 271 : if (!ItemIdIsNormal(lp))
1848 : {
1849 0 : UnlockReleaseBuffer(buffer);
1850 0 : break;
1851 : }
1852 :
1853 : /* OK to access the tuple */
1854 271 : tp.t_self = ctid;
1855 271 : tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
1856 271 : tp.t_len = ItemIdGetLength(lp);
1857 271 : tp.t_tableOid = RelationGetRelid(relation);
1858 :
1859 : /*
1860 : * After following a t_ctid link, we might arrive at an unrelated
1861 : * tuple. Check for XMIN match.
1862 : */
1863 335 : if (TransactionIdIsValid(priorXmax) &&
1864 64 : !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(tp.t_data)))
1865 : {
1866 0 : UnlockReleaseBuffer(buffer);
1867 0 : break;
1868 : }
1869 :
1870 : /*
1871 : * Check tuple visibility; if visible, set it as the new result
1872 : * candidate.
1873 : */
1874 271 : valid = HeapTupleSatisfiesVisibility(&tp, snapshot, buffer);
1875 271 : HeapCheckForSerializableConflictOut(valid, relation, &tp, buffer, snapshot);
1876 271 : if (valid)
1877 191 : *tid = ctid;
1878 :
1879 : /*
1880 : * If there's a valid t_ctid link, follow it, else we're done.
1881 : */
1882 391 : if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
1883 200 : HeapTupleHeaderIsOnlyLocked(tp.t_data) ||
1884 160 : HeapTupleHeaderIndicatesMovedPartitions(tp.t_data) ||
1885 80 : ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid))
1886 : {
1887 207 : UnlockReleaseBuffer(buffer);
1888 207 : break;
1889 : }
1890 :
1891 64 : ctid = tp.t_data->t_ctid;
1892 64 : priorXmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
1893 64 : UnlockReleaseBuffer(buffer);
1894 : } /* end of loop */
1895 207 : }
1896 :
1897 :
1898 : /*
1899 : * UpdateXmaxHintBits - update tuple hint bits after xmax transaction ends
1900 : *
1901 : * This is called after we have waited for the XMAX transaction to terminate.
1902 : * If the transaction aborted, we guarantee the XMAX_INVALID hint bit will
1903 : * be set on exit. If the transaction committed, we set the XMAX_COMMITTED
1904 : * hint bit if possible --- but beware that that may not yet be possible,
1905 : * if the transaction committed asynchronously.
1906 : *
1907 : * Note that if the transaction was a locker only, we set HEAP_XMAX_INVALID
1908 : * even if it commits.
1909 : *
1910 : * Hence callers should look only at XMAX_INVALID.
1911 : *
1912 : * Note this is not allowed for tuples whose xmax is a multixact.
1913 : */
1914 : static void
1915 279 : UpdateXmaxHintBits(HeapTupleHeader tuple, Buffer buffer, TransactionId xid)
1916 : {
1917 : Assert(TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple), xid));
1918 : Assert(!(tuple->t_infomask & HEAP_XMAX_IS_MULTI));
1919 :
1920 279 : if (!(tuple->t_infomask & (HEAP_XMAX_COMMITTED | HEAP_XMAX_INVALID)))
1921 : {
1922 506 : if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask) &&
1923 228 : TransactionIdDidCommit(xid))
1924 201 : HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_COMMITTED,
1925 : xid);
1926 : else
1927 77 : HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_INVALID,
1928 : InvalidTransactionId);
1929 : }
1930 279 : }
1931 :
1932 :
1933 : /*
1934 : * GetBulkInsertState - prepare status object for a bulk insert
1935 : */
1936 : BulkInsertState
1937 3540 : GetBulkInsertState(void)
1938 : {
1939 : BulkInsertState bistate;
1940 :
1941 3540 : bistate = (BulkInsertState) palloc_object(BulkInsertStateData);
1942 3540 : bistate->strategy = GetAccessStrategy(BAS_BULKWRITE);
1943 3540 : bistate->current_buf = InvalidBuffer;
1944 3540 : bistate->next_free = InvalidBlockNumber;
1945 3540 : bistate->last_free = InvalidBlockNumber;
1946 3540 : bistate->already_extended_by = 0;
1947 3540 : return bistate;
1948 : }
1949 :
1950 : /*
1951 : * FreeBulkInsertState - clean up after finishing a bulk insert
1952 : */
1953 : void
1954 3299 : FreeBulkInsertState(BulkInsertState bistate)
1955 : {
1956 3299 : if (bistate->current_buf != InvalidBuffer)
1957 2593 : ReleaseBuffer(bistate->current_buf);
1958 3299 : FreeAccessStrategy(bistate->strategy);
1959 3299 : pfree(bistate);
1960 3299 : }
1961 :
1962 : /*
1963 : * ReleaseBulkInsertStatePin - release a buffer currently held in bistate
1964 : */
1965 : void
1966 90779 : ReleaseBulkInsertStatePin(BulkInsertState bistate)
1967 : {
1968 90779 : if (bistate->current_buf != InvalidBuffer)
1969 40028 : ReleaseBuffer(bistate->current_buf);
1970 90779 : bistate->current_buf = InvalidBuffer;
1971 :
1972 : /*
1973 : * Despite the name, we also reset bulk relation extension state.
1974 : * Otherwise we can end up erroring out due to looking for free space in
1975 : * ->next_free of one partition, even though ->next_free was set when
1976 : * extending another partition. It could obviously also be bad for
1977 : * efficiency to look at existing blocks at offsets from another
1978 : * partition, even if we don't error out.
1979 : */
1980 90779 : bistate->next_free = InvalidBlockNumber;
1981 90779 : bistate->last_free = InvalidBlockNumber;
1982 90779 : }
1983 :
1984 :
1985 : /*
1986 : * heap_insert - insert tuple into a heap
1987 : *
1988 : * The new tuple is stamped with current transaction ID and the specified
1989 : * command ID.
1990 : *
1991 : * See table_tuple_insert for comments about most of the input flags, except
1992 : * that this routine directly takes a tuple rather than a slot.
1993 : *
1994 : * There's corresponding HEAP_INSERT_ options to all the TABLE_INSERT_
1995 : * options, and there additionally is HEAP_INSERT_SPECULATIVE which is used to
1996 : * implement table_tuple_insert_speculative().
1997 : *
1998 : * On return the header fields of *tup are updated to match the stored tuple;
1999 : * in particular tup->t_self receives the actual TID where the tuple was
2000 : * stored. But note that any toasting of fields within the tuple data is NOT
2001 : * reflected into *tup.
2002 : */
2003 : void
2004 11549319 : heap_insert(Relation relation, HeapTuple tup, CommandId cid,
2005 : uint32 options, BulkInsertState bistate)
2006 : {
2007 11549319 : TransactionId xid = GetCurrentTransactionId();
2008 : HeapTuple heaptup;
2009 : Buffer buffer;
2010 : Page page;
2011 11549312 : Buffer vmbuffer = InvalidBuffer;
2012 11549312 : bool all_visible_cleared = false;
2013 :
2014 : /* Cheap, simplistic check that the tuple matches the rel's rowtype. */
2015 : Assert(HeapTupleHeaderGetNatts(tup->t_data) <=
2016 : RelationGetNumberOfAttributes(relation));
2017 :
2018 11549312 : AssertHasSnapshotForToast(relation);
2019 :
2020 : /*
2021 : * Fill in tuple header fields and toast the tuple if necessary.
2022 : *
2023 : * Note: below this point, heaptup is the data we actually intend to store
2024 : * into the relation; tup is the caller's original untoasted data.
2025 : */
2026 11549312 : heaptup = heap_prepare_insert(relation, tup, xid, cid, options);
2027 :
2028 : /*
2029 : * Find buffer to insert this tuple into. If the page is all visible,
2030 : * this will also pin the requisite visibility map page.
2031 : */
2032 11549312 : buffer = RelationGetBufferForTuple(relation, heaptup->t_len,
2033 : InvalidBuffer, options, bistate,
2034 : &vmbuffer, NULL,
2035 : 0);
2036 :
2037 11549312 : page = BufferGetPage(buffer);
2038 :
2039 : /*
2040 : * We're about to do the actual insert -- but check for conflict first, to
2041 : * avoid possibly having to roll back work we've just done.
2042 : *
2043 : * This is safe without a recheck as long as there is no possibility of
2044 : * another process scanning the page between this check and the insert
2045 : * being visible to the scan (i.e., an exclusive buffer content lock is
2046 : * continuously held from this point until the tuple insert is visible).
2047 : *
2048 : * For a heap insert, we only need to check for table-level SSI locks. Our
2049 : * new tuple can't possibly conflict with existing tuple locks, and heap
2050 : * page locks are only consolidated versions of tuple locks; they do not
2051 : * lock "gaps" as index page locks do. So we don't need to specify a
2052 : * buffer when making the call, which makes for a faster check.
2053 : */
2054 11549312 : CheckForSerializableConflictIn(relation, NULL, InvalidBlockNumber);
2055 :
2056 : /* NO EREPORT(ERROR) from here till changes are logged */
2057 11549300 : START_CRIT_SECTION();
2058 :
2059 11549300 : RelationPutHeapTuple(relation, buffer, heaptup,
2060 11549300 : (options & HEAP_INSERT_SPECULATIVE) != 0);
2061 :
2062 11549300 : if (PageIsAllVisible(page))
2063 : {
2064 7346 : all_visible_cleared = true;
2065 7346 : PageClearAllVisible(page);
2066 7346 : visibilitymap_clear(relation,
2067 7346 : ItemPointerGetBlockNumber(&(heaptup->t_self)),
2068 : vmbuffer, VISIBILITYMAP_VALID_BITS);
2069 : }
2070 :
2071 : /*
2072 : * Set pd_prune_xid to trigger heap_page_prune_and_freeze() once the page
2073 : * is full so that we can set the page all-visible in the VM on the next
2074 : * page access.
2075 : *
2076 : * Setting pd_prune_xid is also handy if the inserting transaction
2077 : * eventually aborts making this tuple DEAD and hence available for
2078 : * pruning. If no other tuple in this page is UPDATEd/DELETEd, the aborted
2079 : * tuple would never otherwise be pruned until next vacuum is triggered.
2080 : *
2081 : * Don't set it if we are in bootstrap mode or we are inserting a frozen
2082 : * tuple, as there is no further pruning/freezing needed in those cases.
2083 : */
2084 11549300 : if (TransactionIdIsNormal(xid) && !(options & HEAP_INSERT_FROZEN))
2085 10896755 : PageSetPrunable(page, xid);
2086 :
2087 11549300 : MarkBufferDirty(buffer);
2088 :
2089 : /* XLOG stuff */
2090 11549300 : if (RelationNeedsWAL(relation))
2091 : {
2092 : xl_heap_insert xlrec;
2093 : xl_heap_header xlhdr;
2094 : XLogRecPtr recptr;
2095 10437215 : uint8 info = XLOG_HEAP_INSERT;
2096 10437215 : int bufflags = 0;
2097 :
2098 : /*
2099 : * If this is a catalog, we need to transmit combo CIDs to properly
2100 : * decode, so log that as well.
2101 : */
2102 10437215 : if (RelationIsAccessibleInLogicalDecoding(relation))
2103 5241 : log_heap_new_cid(relation, heaptup);
2104 :
2105 : /*
2106 : * If this is the single and first tuple on page, we can reinit the
2107 : * page instead of restoring the whole thing. Set flag, and hide
2108 : * buffer references from XLogInsert.
2109 : */
2110 10559698 : if (ItemPointerGetOffsetNumber(&(heaptup->t_self)) == FirstOffsetNumber &&
2111 122483 : PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
2112 : {
2113 121609 : info |= XLOG_HEAP_INIT_PAGE;
2114 121609 : bufflags |= REGBUF_WILL_INIT;
2115 : }
2116 :
2117 10437215 : xlrec.offnum = ItemPointerGetOffsetNumber(&heaptup->t_self);
2118 10437215 : xlrec.flags = 0;
2119 10437215 : if (all_visible_cleared)
2120 7342 : xlrec.flags |= XLH_INSERT_ALL_VISIBLE_CLEARED;
2121 10437215 : if (options & HEAP_INSERT_SPECULATIVE)
2122 2206 : xlrec.flags |= XLH_INSERT_IS_SPECULATIVE;
2123 : Assert(ItemPointerGetBlockNumber(&heaptup->t_self) == BufferGetBlockNumber(buffer));
2124 :
2125 : /*
2126 : * For logical decoding, we need the tuple even if we're doing a full
2127 : * page write, so make sure it's included even if we take a full-page
2128 : * image. (XXX We could alternatively store a pointer into the FPW).
2129 : */
2130 10437215 : if (RelationIsLogicallyLogged(relation) &&
2131 273506 : !(options & HEAP_INSERT_NO_LOGICAL))
2132 : {
2133 273394 : xlrec.flags |= XLH_INSERT_CONTAINS_NEW_TUPLE;
2134 273394 : bufflags |= REGBUF_KEEP_DATA;
2135 :
2136 273394 : if (IsToastRelation(relation))
2137 1927 : xlrec.flags |= XLH_INSERT_ON_TOAST_RELATION;
2138 : }
2139 :
2140 10437215 : XLogBeginInsert();
2141 10437215 : XLogRegisterData(&xlrec, SizeOfHeapInsert);
2142 :
2143 10437215 : xlhdr.t_infomask2 = heaptup->t_data->t_infomask2;
2144 10437215 : xlhdr.t_infomask = heaptup->t_data->t_infomask;
2145 10437215 : xlhdr.t_hoff = heaptup->t_data->t_hoff;
2146 :
2147 : /*
2148 : * note we mark xlhdr as belonging to buffer; if XLogInsert decides to
2149 : * write the whole page to the xlog, we don't need to store
2150 : * xl_heap_header in the xlog.
2151 : */
2152 10437215 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2153 10437215 : XLogRegisterBufData(0, &xlhdr, SizeOfHeapHeader);
2154 : /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
2155 10437215 : XLogRegisterBufData(0,
2156 10437215 : (char *) heaptup->t_data + SizeofHeapTupleHeader,
2157 10437215 : heaptup->t_len - SizeofHeapTupleHeader);
2158 :
2159 : /* filtering by origin on a row level is much more efficient */
2160 10437215 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
2161 :
2162 10437215 : recptr = XLogInsert(RM_HEAP_ID, info);
2163 :
2164 10437215 : PageSetLSN(page, recptr);
2165 : }
2166 :
2167 11549300 : END_CRIT_SECTION();
2168 :
2169 11549300 : UnlockReleaseBuffer(buffer);
2170 11549300 : if (vmbuffer != InvalidBuffer)
2171 7689 : ReleaseBuffer(vmbuffer);
2172 :
2173 : /*
2174 : * If tuple is cacheable, mark it for invalidation from the caches in case
2175 : * we abort. Note it is OK to do this after releasing the buffer, because
2176 : * the heaptup data structure is all in local memory, not in the shared
2177 : * buffer.
2178 : */
2179 11549300 : CacheInvalidateHeapTuple(relation, heaptup, NULL);
2180 :
2181 : /* Note: speculative insertions are counted too, even if aborted later */
2182 11549300 : pgstat_count_heap_insert(relation, 1);
2183 :
2184 : /*
2185 : * If heaptup is a private copy, release it. Don't forget to copy t_self
2186 : * back to the caller's image, too.
2187 : */
2188 11549300 : if (heaptup != tup)
2189 : {
2190 22661 : tup->t_self = heaptup->t_self;
2191 22661 : heap_freetuple(heaptup);
2192 : }
2193 11549300 : }
2194 :
2195 : /*
2196 : * Subroutine for heap_insert(). Prepares a tuple for insertion. This sets the
2197 : * tuple header fields and toasts the tuple if necessary. Returns a toasted
2198 : * version of the tuple if it was toasted, or the original tuple if not. Note
2199 : * that in any case, the header fields are also set in the original tuple.
2200 : */
2201 : static HeapTuple
2202 13382794 : heap_prepare_insert(Relation relation, HeapTuple tup, TransactionId xid,
2203 : CommandId cid, uint32 options)
2204 : {
2205 : /*
2206 : * To allow parallel inserts, we need to ensure that they are safe to be
2207 : * performed in workers. We have the infrastructure to allow parallel
2208 : * inserts in general except for the cases where inserts generate a new
2209 : * CommandId (eg. inserts into a table having a foreign key column).
2210 : */
2211 13382794 : if (IsParallelWorker())
2212 0 : ereport(ERROR,
2213 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
2214 : errmsg("cannot insert tuples in a parallel worker")));
2215 :
2216 13382794 : tup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
2217 13382794 : tup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
2218 13382794 : tup->t_data->t_infomask |= HEAP_XMAX_INVALID;
2219 13382794 : HeapTupleHeaderSetXmin(tup->t_data, xid);
2220 13382794 : if (options & HEAP_INSERT_FROZEN)
2221 102651 : HeapTupleHeaderSetXminFrozen(tup->t_data);
2222 :
2223 13382794 : HeapTupleHeaderSetCmin(tup->t_data, cid);
2224 13382794 : HeapTupleHeaderSetXmax(tup->t_data, 0); /* for cleanliness */
2225 13382794 : tup->t_tableOid = RelationGetRelid(relation);
2226 :
2227 : /*
2228 : * If the new tuple is too big for storage or contains already toasted
2229 : * out-of-line attributes from some other relation, invoke the toaster.
2230 : */
2231 13382794 : if (relation->rd_rel->relkind != RELKIND_RELATION &&
2232 36358 : relation->rd_rel->relkind != RELKIND_MATVIEW)
2233 : {
2234 : /* toast table entries should never be recursively toasted */
2235 : Assert(!HeapTupleHasExternal(tup));
2236 36297 : return tup;
2237 : }
2238 13346497 : else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
2239 22720 : return heap_toast_insert_or_update(relation, tup, NULL, options);
2240 : else
2241 13323777 : return tup;
2242 : }
2243 :
2244 : /*
2245 : * Helper for heap_multi_insert() that computes the number of entire pages
2246 : * that inserting the remaining heaptuples requires. Used to determine how
2247 : * much the relation needs to be extended by.
2248 : */
2249 : static int
2250 491828 : heap_multi_insert_pages(HeapTuple *heaptuples, int done, int ntuples, Size saveFreeSpace)
2251 : {
2252 491828 : size_t page_avail = BLCKSZ - SizeOfPageHeaderData - saveFreeSpace;
2253 491828 : int npages = 1;
2254 :
2255 3006123 : for (int i = done; i < ntuples; i++)
2256 : {
2257 2514295 : size_t tup_sz = sizeof(ItemIdData) + MAXALIGN(heaptuples[i]->t_len);
2258 :
2259 2514295 : if (page_avail < tup_sz)
2260 : {
2261 17563 : npages++;
2262 17563 : page_avail = BLCKSZ - SizeOfPageHeaderData - saveFreeSpace;
2263 : }
2264 2514295 : page_avail -= tup_sz;
2265 : }
2266 :
2267 491828 : return npages;
2268 : }
2269 :
2270 : /*
2271 : * heap_multi_insert - insert multiple tuples into a heap
2272 : *
2273 : * This is like heap_insert(), but inserts multiple tuples in one operation.
2274 : * That's faster than calling heap_insert() in a loop, because when multiple
2275 : * tuples can be inserted on a single page, we can write just a single WAL
2276 : * record covering all of them, and only need to lock/unlock the page once.
2277 : *
2278 : * Note: this leaks memory into the current memory context. You can create a
2279 : * temporary context before calling this, if that's a problem.
2280 : */
2281 : void
2282 483565 : heap_multi_insert(Relation relation, TupleTableSlot **slots, int ntuples,
2283 : CommandId cid, uint32 options, BulkInsertState bistate)
2284 : {
2285 483565 : TransactionId xid = GetCurrentTransactionId();
2286 : HeapTuple *heaptuples;
2287 : int i;
2288 : int ndone;
2289 : PGAlignedBlock scratch;
2290 : Page page;
2291 483565 : Buffer vmbuffer = InvalidBuffer;
2292 : bool needwal;
2293 : Size saveFreeSpace;
2294 483565 : bool need_tuple_data = RelationIsLogicallyLogged(relation);
2295 483565 : bool need_cids = RelationIsAccessibleInLogicalDecoding(relation);
2296 483565 : bool starting_with_empty_page = false;
2297 483565 : int npages = 0;
2298 483565 : int npages_used = 0;
2299 :
2300 : /* currently not needed (thus unsupported) for heap_multi_insert() */
2301 : Assert(!(options & HEAP_INSERT_NO_LOGICAL));
2302 :
2303 483565 : AssertHasSnapshotForToast(relation);
2304 :
2305 483565 : needwal = RelationNeedsWAL(relation);
2306 483565 : saveFreeSpace = RelationGetTargetPageFreeSpace(relation,
2307 : HEAP_DEFAULT_FILLFACTOR);
2308 :
2309 : /* Toast and set header data in all the slots */
2310 483565 : heaptuples = palloc(ntuples * sizeof(HeapTuple));
2311 2317047 : for (i = 0; i < ntuples; i++)
2312 : {
2313 : HeapTuple tuple;
2314 :
2315 1833482 : tuple = ExecFetchSlotHeapTuple(slots[i], true, NULL);
2316 1833482 : slots[i]->tts_tableOid = RelationGetRelid(relation);
2317 1833482 : tuple->t_tableOid = slots[i]->tts_tableOid;
2318 1833482 : heaptuples[i] = heap_prepare_insert(relation, tuple, xid, cid,
2319 : options);
2320 : }
2321 :
2322 : /*
2323 : * We're about to do the actual inserts -- but check for conflict first,
2324 : * to minimize the possibility of having to roll back work we've just
2325 : * done.
2326 : *
2327 : * A check here does not definitively prevent a serialization anomaly;
2328 : * that check MUST be done at least past the point of acquiring an
2329 : * exclusive buffer content lock on every buffer that will be affected,
2330 : * and MAY be done after all inserts are reflected in the buffers and
2331 : * those locks are released; otherwise there is a race condition. Since
2332 : * multiple buffers can be locked and unlocked in the loop below, and it
2333 : * would not be feasible to identify and lock all of those buffers before
2334 : * the loop, we must do a final check at the end.
2335 : *
2336 : * The check here could be omitted with no loss of correctness; it is
2337 : * present strictly as an optimization.
2338 : *
2339 : * For heap inserts, we only need to check for table-level SSI locks. Our
2340 : * new tuples can't possibly conflict with existing tuple locks, and heap
2341 : * page locks are only consolidated versions of tuple locks; they do not
2342 : * lock "gaps" as index page locks do. So we don't need to specify a
2343 : * buffer when making the call, which makes for a faster check.
2344 : */
2345 483565 : CheckForSerializableConflictIn(relation, NULL, InvalidBlockNumber);
2346 :
2347 483565 : ndone = 0;
2348 984672 : while (ndone < ntuples)
2349 : {
2350 : Buffer buffer;
2351 501107 : bool all_visible_cleared = false;
2352 501107 : bool all_frozen_set = false;
2353 : int nthispage;
2354 :
2355 501107 : CHECK_FOR_INTERRUPTS();
2356 :
2357 : /*
2358 : * Compute number of pages needed to fit the to-be-inserted tuples in
2359 : * the worst case. This will be used to determine how much to extend
2360 : * the relation by in RelationGetBufferForTuple(), if needed. If we
2361 : * filled a prior page from scratch, we can just update our last
2362 : * computation, but if we started with a partially filled page,
2363 : * recompute from scratch, the number of potentially required pages
2364 : * can vary due to tuples needing to fit onto the page, page headers
2365 : * etc.
2366 : */
2367 501107 : if (ndone == 0 || !starting_with_empty_page)
2368 : {
2369 491828 : npages = heap_multi_insert_pages(heaptuples, ndone, ntuples,
2370 : saveFreeSpace);
2371 491828 : npages_used = 0;
2372 : }
2373 : else
2374 9279 : npages_used++;
2375 :
2376 : /*
2377 : * Find buffer where at least the next tuple will fit. If the page is
2378 : * all-visible, this will also pin the requisite visibility map page.
2379 : *
2380 : * Also pin visibility map page if COPY FREEZE inserts tuples into an
2381 : * empty page. See all_frozen_set below.
2382 : */
2383 501107 : buffer = RelationGetBufferForTuple(relation, heaptuples[ndone]->t_len,
2384 : InvalidBuffer, options, bistate,
2385 : &vmbuffer, NULL,
2386 : npages - npages_used);
2387 501107 : page = BufferGetPage(buffer);
2388 :
2389 501107 : starting_with_empty_page = PageGetMaxOffsetNumber(page) == 0;
2390 :
2391 501107 : if (starting_with_empty_page && (options & HEAP_INSERT_FROZEN))
2392 : {
2393 1665 : all_frozen_set = true;
2394 : /* Lock the vmbuffer before entering the critical section */
2395 1665 : LockBuffer(vmbuffer, BUFFER_LOCK_EXCLUSIVE);
2396 : }
2397 :
2398 : /* NO EREPORT(ERROR) from here till changes are logged */
2399 501107 : START_CRIT_SECTION();
2400 :
2401 : /*
2402 : * RelationGetBufferForTuple has ensured that the first tuple fits.
2403 : * Put that on the page, and then as many other tuples as fit.
2404 : */
2405 501107 : RelationPutHeapTuple(relation, buffer, heaptuples[ndone], false);
2406 :
2407 : /*
2408 : * For logical decoding we need combo CIDs to properly decode the
2409 : * catalog.
2410 : */
2411 501107 : if (needwal && need_cids)
2412 7037 : log_heap_new_cid(relation, heaptuples[ndone]);
2413 :
2414 1833482 : for (nthispage = 1; ndone + nthispage < ntuples; nthispage++)
2415 : {
2416 1349917 : HeapTuple heaptup = heaptuples[ndone + nthispage];
2417 :
2418 1349917 : if (PageGetHeapFreeSpace(page) < MAXALIGN(heaptup->t_len) + saveFreeSpace)
2419 17542 : break;
2420 :
2421 1332375 : RelationPutHeapTuple(relation, buffer, heaptup, false);
2422 :
2423 : /*
2424 : * For logical decoding we need combo CIDs to properly decode the
2425 : * catalog.
2426 : */
2427 1332375 : if (needwal && need_cids)
2428 5054 : log_heap_new_cid(relation, heaptup);
2429 : }
2430 :
2431 : /*
2432 : * If the page is all visible, need to clear that, unless we're only
2433 : * going to add further frozen rows to it.
2434 : *
2435 : * If we're only adding already frozen rows to a previously empty
2436 : * page, mark it as all-frozen and update the visibility map. We're
2437 : * already holding a pin on the vmbuffer.
2438 : */
2439 501107 : if (PageIsAllVisible(page) && !(options & HEAP_INSERT_FROZEN))
2440 : {
2441 3501 : all_visible_cleared = true;
2442 3501 : PageClearAllVisible(page);
2443 3501 : visibilitymap_clear(relation,
2444 : BufferGetBlockNumber(buffer),
2445 : vmbuffer, VISIBILITYMAP_VALID_BITS);
2446 : }
2447 497606 : else if (all_frozen_set)
2448 : {
2449 1665 : PageSetAllVisible(page);
2450 1665 : PageClearPrunable(page);
2451 1665 : visibilitymap_set(BufferGetBlockNumber(buffer),
2452 : vmbuffer,
2453 : VISIBILITYMAP_ALL_VISIBLE |
2454 : VISIBILITYMAP_ALL_FROZEN,
2455 : relation->rd_locator);
2456 : }
2457 :
2458 : /*
2459 : * Set pd_prune_xid. See heap_insert() for more on why we do this when
2460 : * inserting tuples. This only makes sense if we aren't already
2461 : * setting the page frozen in the VM and we're not in bootstrap mode.
2462 : */
2463 501107 : if (!all_frozen_set && TransactionIdIsNormal(xid))
2464 476870 : PageSetPrunable(page, xid);
2465 :
2466 501107 : MarkBufferDirty(buffer);
2467 :
2468 : /* XLOG stuff */
2469 501107 : if (needwal)
2470 : {
2471 : XLogRecPtr recptr;
2472 : xl_heap_multi_insert *xlrec;
2473 497212 : uint8 info = XLOG_HEAP2_MULTI_INSERT;
2474 : char *tupledata;
2475 : int totaldatalen;
2476 497212 : char *scratchptr = scratch.data;
2477 : bool init;
2478 497212 : int bufflags = 0;
2479 :
2480 : /*
2481 : * If the page was previously empty, we can reinit the page
2482 : * instead of restoring the whole thing.
2483 : */
2484 497212 : init = starting_with_empty_page;
2485 :
2486 : /* allocate xl_heap_multi_insert struct from the scratch area */
2487 497212 : xlrec = (xl_heap_multi_insert *) scratchptr;
2488 497212 : scratchptr += SizeOfHeapMultiInsert;
2489 :
2490 : /*
2491 : * Allocate offsets array. Unless we're reinitializing the page,
2492 : * in that case the tuples are stored in order starting at
2493 : * FirstOffsetNumber and we don't need to store the offsets
2494 : * explicitly.
2495 : */
2496 497212 : if (!init)
2497 480448 : scratchptr += nthispage * sizeof(OffsetNumber);
2498 :
2499 : /* the rest of the scratch space is used for tuple data */
2500 497212 : tupledata = scratchptr;
2501 :
2502 : /* check that the mutually exclusive flags are not both set */
2503 : Assert(!(all_visible_cleared && all_frozen_set));
2504 :
2505 497212 : xlrec->flags = 0;
2506 497212 : if (all_visible_cleared)
2507 3501 : xlrec->flags = XLH_INSERT_ALL_VISIBLE_CLEARED;
2508 :
2509 : /*
2510 : * We don't have to worry about including a conflict xid in the
2511 : * WAL record, as HEAP_INSERT_FROZEN intentionally violates
2512 : * visibility rules.
2513 : */
2514 497212 : if (all_frozen_set)
2515 21 : xlrec->flags = XLH_INSERT_ALL_FROZEN_SET;
2516 :
2517 497212 : xlrec->ntuples = nthispage;
2518 :
2519 : /*
2520 : * Write out an xl_multi_insert_tuple and the tuple data itself
2521 : * for each tuple.
2522 : */
2523 2123523 : for (i = 0; i < nthispage; i++)
2524 : {
2525 1626311 : HeapTuple heaptup = heaptuples[ndone + i];
2526 : xl_multi_insert_tuple *tuphdr;
2527 : int datalen;
2528 :
2529 1626311 : if (!init)
2530 986921 : xlrec->offsets[i] = ItemPointerGetOffsetNumber(&heaptup->t_self);
2531 : /* xl_multi_insert_tuple needs two-byte alignment. */
2532 1626311 : tuphdr = (xl_multi_insert_tuple *) SHORTALIGN(scratchptr);
2533 1626311 : scratchptr = ((char *) tuphdr) + SizeOfMultiInsertTuple;
2534 :
2535 1626311 : tuphdr->t_infomask2 = heaptup->t_data->t_infomask2;
2536 1626311 : tuphdr->t_infomask = heaptup->t_data->t_infomask;
2537 1626311 : tuphdr->t_hoff = heaptup->t_data->t_hoff;
2538 :
2539 : /* write bitmap [+ padding] [+ oid] + data */
2540 1626311 : datalen = heaptup->t_len - SizeofHeapTupleHeader;
2541 1626311 : memcpy(scratchptr,
2542 1626311 : (char *) heaptup->t_data + SizeofHeapTupleHeader,
2543 : datalen);
2544 1626311 : tuphdr->datalen = datalen;
2545 1626311 : scratchptr += datalen;
2546 : }
2547 497212 : totaldatalen = scratchptr - tupledata;
2548 : Assert((scratchptr - scratch.data) < BLCKSZ);
2549 :
2550 497212 : if (need_tuple_data)
2551 56 : xlrec->flags |= XLH_INSERT_CONTAINS_NEW_TUPLE;
2552 :
2553 : /*
2554 : * Signal that this is the last xl_heap_multi_insert record
2555 : * emitted by this call to heap_multi_insert(). Needed for logical
2556 : * decoding so it knows when to cleanup temporary data.
2557 : */
2558 497212 : if (ndone + nthispage == ntuples)
2559 483115 : xlrec->flags |= XLH_INSERT_LAST_IN_MULTI;
2560 :
2561 497212 : if (init)
2562 : {
2563 16764 : info |= XLOG_HEAP_INIT_PAGE;
2564 16764 : bufflags |= REGBUF_WILL_INIT;
2565 : }
2566 :
2567 : /*
2568 : * If we're doing logical decoding, include the new tuple data
2569 : * even if we take a full-page image of the page.
2570 : */
2571 497212 : if (need_tuple_data)
2572 56 : bufflags |= REGBUF_KEEP_DATA;
2573 :
2574 497212 : XLogBeginInsert();
2575 497212 : XLogRegisterData(xlrec, tupledata - scratch.data);
2576 497212 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2577 497212 : if (all_frozen_set)
2578 21 : XLogRegisterBuffer(1, vmbuffer, 0);
2579 :
2580 497212 : XLogRegisterBufData(0, tupledata, totaldatalen);
2581 :
2582 : /* filtering by origin on a row level is much more efficient */
2583 497212 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
2584 :
2585 497212 : recptr = XLogInsert(RM_HEAP2_ID, info);
2586 :
2587 497212 : PageSetLSN(page, recptr);
2588 497212 : if (all_frozen_set)
2589 : {
2590 : Assert(BufferIsDirty(vmbuffer));
2591 21 : PageSetLSN(BufferGetPage(vmbuffer), recptr);
2592 : }
2593 : }
2594 :
2595 501107 : END_CRIT_SECTION();
2596 :
2597 501107 : if (all_frozen_set)
2598 1665 : LockBuffer(vmbuffer, BUFFER_LOCK_UNLOCK);
2599 :
2600 501107 : UnlockReleaseBuffer(buffer);
2601 501107 : ndone += nthispage;
2602 :
2603 : /*
2604 : * NB: Only release vmbuffer after inserting all tuples - it's fairly
2605 : * likely that we'll insert into subsequent heap pages that are likely
2606 : * to use the same vm page.
2607 : */
2608 : }
2609 :
2610 : /* We're done with inserting all tuples, so release the last vmbuffer. */
2611 483565 : if (vmbuffer != InvalidBuffer)
2612 3594 : ReleaseBuffer(vmbuffer);
2613 :
2614 : /*
2615 : * We're done with the actual inserts. Check for conflicts again, to
2616 : * ensure that all rw-conflicts in to these inserts are detected. Without
2617 : * this final check, a sequential scan of the heap may have locked the
2618 : * table after the "before" check, missing one opportunity to detect the
2619 : * conflict, and then scanned the table before the new tuples were there,
2620 : * missing the other chance to detect the conflict.
2621 : *
2622 : * For heap inserts, we only need to check for table-level SSI locks. Our
2623 : * new tuples can't possibly conflict with existing tuple locks, and heap
2624 : * page locks are only consolidated versions of tuple locks; they do not
2625 : * lock "gaps" as index page locks do. So we don't need to specify a
2626 : * buffer when making the call.
2627 : */
2628 483565 : CheckForSerializableConflictIn(relation, NULL, InvalidBlockNumber);
2629 :
2630 : /*
2631 : * If tuples are cacheable, mark them for invalidation from the caches in
2632 : * case we abort. Note it is OK to do this after releasing the buffer,
2633 : * because the heaptuples data structure is all in local memory, not in
2634 : * the shared buffer.
2635 : */
2636 483565 : if (IsCatalogRelation(relation))
2637 : {
2638 1600487 : for (i = 0; i < ntuples; i++)
2639 1118386 : CacheInvalidateHeapTuple(relation, heaptuples[i], NULL);
2640 : }
2641 :
2642 : /* copy t_self fields back to the caller's slots */
2643 2317047 : for (i = 0; i < ntuples; i++)
2644 1833482 : slots[i]->tts_tid = heaptuples[i]->t_self;
2645 :
2646 483565 : pgstat_count_heap_insert(relation, ntuples);
2647 483565 : }
2648 :
2649 : /*
2650 : * simple_heap_insert - insert a tuple
2651 : *
2652 : * Currently, this routine differs from heap_insert only in supplying
2653 : * a default command ID and not allowing access to the speedup options.
2654 : *
2655 : * This should be used rather than using heap_insert directly in most places
2656 : * where we are modifying system catalogs.
2657 : */
2658 : void
2659 1083987 : simple_heap_insert(Relation relation, HeapTuple tup)
2660 : {
2661 1083987 : heap_insert(relation, tup, GetCurrentCommandId(true), 0, NULL);
2662 1083987 : }
2663 :
2664 : /*
2665 : * Given infomask/infomask2, compute the bits that must be saved in the
2666 : * "infobits" field of xl_heap_delete, xl_heap_update, xl_heap_lock,
2667 : * xl_heap_lock_updated WAL records.
2668 : *
2669 : * See fix_infomask_from_infobits.
2670 : */
2671 : static uint8
2672 6553712 : compute_infobits(uint16 infomask, uint16 infomask2)
2673 : {
2674 : return
2675 6553712 : ((infomask & HEAP_XMAX_IS_MULTI) != 0 ? XLHL_XMAX_IS_MULTI : 0) |
2676 6553712 : ((infomask & HEAP_XMAX_LOCK_ONLY) != 0 ? XLHL_XMAX_LOCK_ONLY : 0) |
2677 6553712 : ((infomask & HEAP_XMAX_EXCL_LOCK) != 0 ? XLHL_XMAX_EXCL_LOCK : 0) |
2678 : /* note we ignore HEAP_XMAX_SHR_LOCK here */
2679 13107424 : ((infomask & HEAP_XMAX_KEYSHR_LOCK) != 0 ? XLHL_XMAX_KEYSHR_LOCK : 0) |
2680 : ((infomask2 & HEAP_KEYS_UPDATED) != 0 ?
2681 6553712 : XLHL_KEYS_UPDATED : 0);
2682 : }
2683 :
2684 : /*
2685 : * Given two versions of the same t_infomask for a tuple, compare them and
2686 : * return whether the relevant status for a tuple Xmax has changed. This is
2687 : * used after a buffer lock has been released and reacquired: we want to ensure
2688 : * that the tuple state continues to be the same it was when we previously
2689 : * examined it.
2690 : *
2691 : * Note the Xmax field itself must be compared separately.
2692 : */
2693 : static inline bool
2694 5448 : xmax_infomask_changed(uint16 new_infomask, uint16 old_infomask)
2695 : {
2696 5448 : const uint16 interesting =
2697 : HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY | HEAP_LOCK_MASK;
2698 :
2699 5448 : if ((new_infomask & interesting) != (old_infomask & interesting))
2700 18 : return true;
2701 :
2702 5430 : return false;
2703 : }
2704 :
2705 : /*
2706 : * heap_delete - delete a tuple
2707 : *
2708 : * See table_tuple_delete() for an explanation of the parameters, except that
2709 : * this routine directly takes a tuple rather than a slot.
2710 : *
2711 : * In the failure cases, the routine fills *tmfd with the tuple's t_ctid,
2712 : * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax (the last
2713 : * only for TM_SelfModified, since we cannot obtain cmax from a combo CID
2714 : * generated by another transaction).
2715 : */
2716 : TM_Result
2717 1895514 : heap_delete(Relation relation, const ItemPointerData *tid,
2718 : CommandId cid, uint32 options, Snapshot crosscheck,
2719 : bool wait, TM_FailureData *tmfd)
2720 : {
2721 : TM_Result result;
2722 1895514 : TransactionId xid = GetCurrentTransactionId();
2723 : ItemId lp;
2724 : HeapTupleData tp;
2725 : Page page;
2726 : BlockNumber block;
2727 : Buffer buffer;
2728 1895514 : Buffer vmbuffer = InvalidBuffer;
2729 : TransactionId new_xmax;
2730 : uint16 new_infomask,
2731 : new_infomask2;
2732 1895514 : bool changingPart = (options & TABLE_DELETE_CHANGING_PARTITION) != 0;
2733 1895514 : bool walLogical = (options & TABLE_DELETE_NO_LOGICAL) == 0;
2734 1895514 : bool have_tuple_lock = false;
2735 : bool iscombo;
2736 1895514 : bool all_visible_cleared = false;
2737 1895514 : HeapTuple old_key_tuple = NULL; /* replica identity of the tuple */
2738 1895514 : bool old_key_copied = false;
2739 :
2740 : Assert(ItemPointerIsValid(tid));
2741 :
2742 1895514 : AssertHasSnapshotForToast(relation);
2743 :
2744 : /*
2745 : * Forbid this during a parallel operation, lest it allocate a combo CID.
2746 : * Other workers might need that combo CID for visibility checks, and we
2747 : * have no provision for broadcasting it to them.
2748 : */
2749 1895514 : if (IsInParallelMode())
2750 0 : ereport(ERROR,
2751 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
2752 : errmsg("cannot delete tuples during a parallel operation")));
2753 :
2754 1895514 : block = ItemPointerGetBlockNumber(tid);
2755 1895514 : buffer = ReadBuffer(relation, block);
2756 1895514 : page = BufferGetPage(buffer);
2757 :
2758 : /*
2759 : * Before locking the buffer, pin the visibility map page if it appears to
2760 : * be necessary. Since we haven't got the lock yet, someone else might be
2761 : * in the middle of changing this, so we'll need to recheck after we have
2762 : * the lock.
2763 : */
2764 1895514 : if (PageIsAllVisible(page))
2765 1773 : visibilitymap_pin(relation, block, &vmbuffer);
2766 :
2767 1895514 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2768 :
2769 1895514 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
2770 : Assert(ItemIdIsNormal(lp));
2771 :
2772 1895514 : tp.t_tableOid = RelationGetRelid(relation);
2773 1895514 : tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
2774 1895514 : tp.t_len = ItemIdGetLength(lp);
2775 1895514 : tp.t_self = *tid;
2776 :
2777 1 : l1:
2778 :
2779 : /*
2780 : * If we didn't pin the visibility map page and the page has become all
2781 : * visible while we were busy locking the buffer, we'll have to unlock and
2782 : * re-lock, to avoid holding the buffer lock across an I/O. That's a bit
2783 : * unfortunate, but hopefully shouldn't happen often.
2784 : */
2785 1895515 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
2786 : {
2787 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2788 0 : visibilitymap_pin(relation, block, &vmbuffer);
2789 0 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2790 : }
2791 :
2792 1895515 : result = HeapTupleSatisfiesUpdate(&tp, cid, buffer);
2793 :
2794 1895515 : if (result == TM_Invisible)
2795 : {
2796 0 : UnlockReleaseBuffer(buffer);
2797 0 : ereport(ERROR,
2798 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
2799 : errmsg("attempted to delete invisible tuple")));
2800 : }
2801 1895515 : else if (result == TM_BeingModified && wait)
2802 : {
2803 : TransactionId xwait;
2804 : uint16 infomask;
2805 :
2806 : /* must copy state data before unlocking buffer */
2807 40668 : xwait = HeapTupleHeaderGetRawXmax(tp.t_data);
2808 40668 : infomask = tp.t_data->t_infomask;
2809 :
2810 : /*
2811 : * Sleep until concurrent transaction ends -- except when there's a
2812 : * single locker and it's our own transaction. Note we don't care
2813 : * which lock mode the locker has, because we need the strongest one.
2814 : *
2815 : * Before sleeping, we need to acquire tuple lock to establish our
2816 : * priority for the tuple (see heap_lock_tuple). LockTuple will
2817 : * release us when we are next-in-line for the tuple.
2818 : *
2819 : * If we are forced to "start over" below, we keep the tuple lock;
2820 : * this arranges that we stay at the head of the line while rechecking
2821 : * tuple state.
2822 : */
2823 40668 : if (infomask & HEAP_XMAX_IS_MULTI)
2824 : {
2825 8 : bool current_is_member = false;
2826 :
2827 8 : if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
2828 : LockTupleExclusive, ¤t_is_member))
2829 : {
2830 8 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2831 :
2832 : /*
2833 : * Acquire the lock, if necessary (but skip it when we're
2834 : * requesting a lock and already have one; avoids deadlock).
2835 : */
2836 8 : if (!current_is_member)
2837 6 : heap_acquire_tuplock(relation, &(tp.t_self), LockTupleExclusive,
2838 : LockWaitBlock, &have_tuple_lock);
2839 :
2840 : /* wait for multixact */
2841 8 : MultiXactIdWait((MultiXactId) xwait, MultiXactStatusUpdate, infomask,
2842 : relation, &(tp.t_self), XLTW_Delete,
2843 : NULL);
2844 8 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2845 :
2846 : /*
2847 : * If xwait had just locked the tuple then some other xact
2848 : * could update this tuple before we get to this point. Check
2849 : * for xmax change, and start over if so.
2850 : *
2851 : * We also must start over if we didn't pin the VM page, and
2852 : * the page has become all visible.
2853 : */
2854 16 : if ((vmbuffer == InvalidBuffer && PageIsAllVisible(page)) ||
2855 16 : xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2856 8 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2857 : xwait))
2858 0 : goto l1;
2859 : }
2860 :
2861 : /*
2862 : * You might think the multixact is necessarily done here, but not
2863 : * so: it could have surviving members, namely our own xact or
2864 : * other subxacts of this backend. It is legal for us to delete
2865 : * the tuple in either case, however (the latter case is
2866 : * essentially a situation of upgrading our former shared lock to
2867 : * exclusive). We don't bother changing the on-disk hint bits
2868 : * since we are about to overwrite the xmax altogether.
2869 : */
2870 : }
2871 40660 : else if (!TransactionIdIsCurrentTransactionId(xwait))
2872 : {
2873 : /*
2874 : * Wait for regular transaction to end; but first, acquire tuple
2875 : * lock.
2876 : */
2877 70 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2878 70 : heap_acquire_tuplock(relation, &(tp.t_self), LockTupleExclusive,
2879 : LockWaitBlock, &have_tuple_lock);
2880 70 : XactLockTableWait(xwait, relation, &(tp.t_self), XLTW_Delete);
2881 66 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2882 :
2883 : /*
2884 : * xwait is done, but if xwait had just locked the tuple then some
2885 : * other xact could update this tuple before we get to this point.
2886 : * Check for xmax change, and start over if so.
2887 : *
2888 : * We also must start over if we didn't pin the VM page, and the
2889 : * page has become all visible.
2890 : */
2891 132 : if ((vmbuffer == InvalidBuffer && PageIsAllVisible(page)) ||
2892 131 : xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2893 65 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2894 : xwait))
2895 1 : goto l1;
2896 :
2897 : /* Otherwise check if it committed or aborted */
2898 65 : UpdateXmaxHintBits(tp.t_data, buffer, xwait);
2899 : }
2900 :
2901 : /*
2902 : * We may overwrite if previous xmax aborted, or if it committed but
2903 : * only locked the tuple without updating it.
2904 : */
2905 81306 : if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
2906 40692 : HEAP_XMAX_IS_LOCKED_ONLY(tp.t_data->t_infomask) ||
2907 49 : HeapTupleHeaderIsOnlyLocked(tp.t_data))
2908 40618 : result = TM_Ok;
2909 45 : else if (!ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid))
2910 32 : result = TM_Updated;
2911 : else
2912 13 : result = TM_Deleted;
2913 : }
2914 :
2915 : /* sanity check the result HeapTupleSatisfiesUpdate() and the logic above */
2916 : if (result != TM_Ok)
2917 : {
2918 : Assert(result == TM_SelfModified ||
2919 : result == TM_Updated ||
2920 : result == TM_Deleted ||
2921 : result == TM_BeingModified);
2922 : Assert(!(tp.t_data->t_infomask & HEAP_XMAX_INVALID));
2923 : Assert(result != TM_Updated ||
2924 : !ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid));
2925 : }
2926 :
2927 1895510 : if (crosscheck != InvalidSnapshot && result == TM_Ok)
2928 : {
2929 : /* Perform additional check for transaction-snapshot mode RI updates */
2930 1 : if (!HeapTupleSatisfiesVisibility(&tp, crosscheck, buffer))
2931 1 : result = TM_Updated;
2932 : }
2933 :
2934 1895510 : if (result != TM_Ok)
2935 : {
2936 101 : tmfd->ctid = tp.t_data->t_ctid;
2937 101 : tmfd->xmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
2938 101 : if (result == TM_SelfModified)
2939 36 : tmfd->cmax = HeapTupleHeaderGetCmax(tp.t_data);
2940 : else
2941 65 : tmfd->cmax = InvalidCommandId;
2942 101 : UnlockReleaseBuffer(buffer);
2943 101 : if (have_tuple_lock)
2944 45 : UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
2945 101 : if (vmbuffer != InvalidBuffer)
2946 0 : ReleaseBuffer(vmbuffer);
2947 101 : return result;
2948 : }
2949 :
2950 : /*
2951 : * We're about to do the actual delete -- check for conflict first, to
2952 : * avoid possibly having to roll back work we've just done.
2953 : *
2954 : * This is safe without a recheck as long as there is no possibility of
2955 : * another process scanning the page between this check and the delete
2956 : * being visible to the scan (i.e., an exclusive buffer content lock is
2957 : * continuously held from this point until the tuple delete is visible).
2958 : */
2959 1895409 : CheckForSerializableConflictIn(relation, tid, BufferGetBlockNumber(buffer));
2960 :
2961 : /* replace cid with a combo CID if necessary */
2962 1895395 : HeapTupleHeaderAdjustCmax(tp.t_data, &cid, &iscombo);
2963 :
2964 : /*
2965 : * Compute replica identity tuple before entering the critical section so
2966 : * we don't PANIC upon a memory allocation failure.
2967 : */
2968 1895395 : old_key_tuple = walLogical ?
2969 1895395 : ExtractReplicaIdentity(relation, &tp, true, &old_key_copied) : NULL;
2970 :
2971 : /*
2972 : * If this is the first possibly-multixact-able operation in the current
2973 : * transaction, set my per-backend OldestMemberMXactId setting. We can be
2974 : * certain that the transaction will never become a member of any older
2975 : * MultiXactIds than that. (We have to do this even if we end up just
2976 : * using our own TransactionId below, since some other backend could
2977 : * incorporate our XID into a MultiXact immediately afterwards.)
2978 : */
2979 1895395 : MultiXactIdSetOldestMember();
2980 :
2981 1895395 : compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(tp.t_data),
2982 1895395 : tp.t_data->t_infomask, tp.t_data->t_infomask2,
2983 : xid, LockTupleExclusive, true,
2984 : &new_xmax, &new_infomask, &new_infomask2);
2985 :
2986 1895395 : START_CRIT_SECTION();
2987 :
2988 : /*
2989 : * If this transaction commits, the tuple will become DEAD sooner or
2990 : * later. Set flag that this page is a candidate for pruning once our xid
2991 : * falls below the OldestXmin horizon. If the transaction finally aborts,
2992 : * the subsequent page pruning will be a no-op and the hint will be
2993 : * cleared.
2994 : */
2995 1895395 : PageSetPrunable(page, xid);
2996 :
2997 1895395 : if (PageIsAllVisible(page))
2998 : {
2999 1773 : all_visible_cleared = true;
3000 1773 : PageClearAllVisible(page);
3001 1773 : visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
3002 : vmbuffer, VISIBILITYMAP_VALID_BITS);
3003 : }
3004 :
3005 : /* store transaction information of xact deleting the tuple */
3006 1895395 : tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3007 1895395 : tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3008 1895395 : tp.t_data->t_infomask |= new_infomask;
3009 1895395 : tp.t_data->t_infomask2 |= new_infomask2;
3010 1895395 : HeapTupleHeaderClearHotUpdated(tp.t_data);
3011 1895395 : HeapTupleHeaderSetXmax(tp.t_data, new_xmax);
3012 1895395 : HeapTupleHeaderSetCmax(tp.t_data, cid, iscombo);
3013 : /* Make sure there is no forward chain link in t_ctid */
3014 1895395 : tp.t_data->t_ctid = tp.t_self;
3015 :
3016 : /* Signal that this is actually a move into another partition */
3017 1895395 : if (changingPart)
3018 666 : HeapTupleHeaderSetMovedPartitions(tp.t_data);
3019 :
3020 1895395 : MarkBufferDirty(buffer);
3021 :
3022 : /*
3023 : * XLOG stuff
3024 : *
3025 : * NB: heap_abort_speculative() uses the same xlog record and replay
3026 : * routines.
3027 : */
3028 1895395 : if (RelationNeedsWAL(relation))
3029 : {
3030 : xl_heap_delete xlrec;
3031 : xl_heap_header xlhdr;
3032 : XLogRecPtr recptr;
3033 :
3034 : /*
3035 : * For logical decode we need combo CIDs to properly decode the
3036 : * catalog
3037 : */
3038 1812152 : if (RelationIsAccessibleInLogicalDecoding(relation))
3039 10308 : log_heap_new_cid(relation, &tp);
3040 :
3041 1812152 : xlrec.flags = 0;
3042 1812152 : if (all_visible_cleared)
3043 1773 : xlrec.flags |= XLH_DELETE_ALL_VISIBLE_CLEARED;
3044 1812152 : if (changingPart)
3045 666 : xlrec.flags |= XLH_DELETE_IS_PARTITION_MOVE;
3046 3624304 : xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
3047 1812152 : tp.t_data->t_infomask2);
3048 1812152 : xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
3049 1812152 : xlrec.xmax = new_xmax;
3050 :
3051 1812152 : if (old_key_tuple != NULL)
3052 : {
3053 47028 : if (relation->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
3054 135 : xlrec.flags |= XLH_DELETE_CONTAINS_OLD_TUPLE;
3055 : else
3056 46893 : xlrec.flags |= XLH_DELETE_CONTAINS_OLD_KEY;
3057 : }
3058 :
3059 : /*
3060 : * Mark the change as not-for-logical-decoding if caller requested so.
3061 : *
3062 : * (This is used for changes that affect relations not visible to
3063 : * other transactions, such as the transient table during concurrent
3064 : * repack.)
3065 : */
3066 1812152 : if (!walLogical)
3067 3 : xlrec.flags |= XLH_DELETE_NO_LOGICAL;
3068 :
3069 1812152 : XLogBeginInsert();
3070 1812152 : XLogRegisterData(&xlrec, SizeOfHeapDelete);
3071 :
3072 1812152 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
3073 :
3074 : /*
3075 : * Log replica identity of the deleted tuple if there is one
3076 : */
3077 1812152 : if (old_key_tuple != NULL)
3078 : {
3079 47028 : xlhdr.t_infomask2 = old_key_tuple->t_data->t_infomask2;
3080 47028 : xlhdr.t_infomask = old_key_tuple->t_data->t_infomask;
3081 47028 : xlhdr.t_hoff = old_key_tuple->t_data->t_hoff;
3082 :
3083 47028 : XLogRegisterData(&xlhdr, SizeOfHeapHeader);
3084 47028 : XLogRegisterData((char *) old_key_tuple->t_data
3085 : + SizeofHeapTupleHeader,
3086 47028 : old_key_tuple->t_len
3087 : - SizeofHeapTupleHeader);
3088 : }
3089 :
3090 : /* filtering by origin on a row level is much more efficient */
3091 1812152 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
3092 :
3093 1812152 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
3094 :
3095 1812152 : PageSetLSN(page, recptr);
3096 : }
3097 :
3098 1895395 : END_CRIT_SECTION();
3099 :
3100 1895395 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3101 :
3102 1895395 : if (vmbuffer != InvalidBuffer)
3103 1773 : ReleaseBuffer(vmbuffer);
3104 :
3105 : /*
3106 : * If the tuple has toasted out-of-line attributes, we need to delete
3107 : * those items too. We have to do this before releasing the buffer
3108 : * because we need to look at the contents of the tuple, but it's OK to
3109 : * release the content lock on the buffer first.
3110 : */
3111 1899175 : if (relation->rd_rel->relkind != RELKIND_RELATION &&
3112 3793 : relation->rd_rel->relkind != RELKIND_MATVIEW)
3113 : {
3114 : /* toast table entries should never be recursively toasted */
3115 : Assert(!HeapTupleHasExternal(&tp));
3116 : }
3117 1891615 : else if (HeapTupleHasExternal(&tp))
3118 534 : heap_toast_delete(relation, &tp, false);
3119 :
3120 : /*
3121 : * Mark tuple for invalidation from system caches at next command
3122 : * boundary. We have to do this before releasing the buffer because we
3123 : * need to look at the contents of the tuple.
3124 : */
3125 1895395 : CacheInvalidateHeapTuple(relation, &tp, NULL);
3126 :
3127 : /* Now we can release the buffer */
3128 1895395 : ReleaseBuffer(buffer);
3129 :
3130 : /*
3131 : * Release the lmgr tuple lock, if we had it.
3132 : */
3133 1895395 : if (have_tuple_lock)
3134 26 : UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
3135 :
3136 1895395 : pgstat_count_heap_delete(relation);
3137 :
3138 1895395 : if (old_key_tuple != NULL && old_key_copied)
3139 46894 : heap_freetuple(old_key_tuple);
3140 :
3141 1895395 : return TM_Ok;
3142 : }
3143 :
3144 : /*
3145 : * simple_heap_delete - delete a tuple
3146 : *
3147 : * This routine may be used to delete a tuple when concurrent updates of
3148 : * the target tuple are not expected (for example, because we have a lock
3149 : * on the relation associated with the tuple). Any failure is reported
3150 : * via ereport().
3151 : */
3152 : void
3153 823639 : simple_heap_delete(Relation relation, const ItemPointerData *tid)
3154 : {
3155 : TM_Result result;
3156 : TM_FailureData tmfd;
3157 :
3158 823639 : result = heap_delete(relation, tid,
3159 : GetCurrentCommandId(true),
3160 : 0,
3161 : InvalidSnapshot,
3162 : true /* wait for commit */ ,
3163 : &tmfd);
3164 823639 : switch (result)
3165 : {
3166 0 : case TM_SelfModified:
3167 : /* Tuple was already updated in current command? */
3168 0 : elog(ERROR, "tuple already updated by self");
3169 : break;
3170 :
3171 823639 : case TM_Ok:
3172 : /* done successfully */
3173 823639 : break;
3174 :
3175 0 : case TM_Updated:
3176 0 : elog(ERROR, "tuple concurrently updated");
3177 : break;
3178 :
3179 0 : case TM_Deleted:
3180 0 : elog(ERROR, "tuple concurrently deleted");
3181 : break;
3182 :
3183 0 : default:
3184 0 : elog(ERROR, "unrecognized heap_delete status: %u", result);
3185 : break;
3186 : }
3187 823639 : }
3188 :
3189 : /*
3190 : * heap_update - replace a tuple
3191 : *
3192 : * See table_tuple_update() for an explanation of the parameters, except that
3193 : * this routine directly takes a tuple rather than a slot.
3194 : *
3195 : * In the failure cases, the routine fills *tmfd with the tuple's t_ctid,
3196 : * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax (the last
3197 : * only for TM_SelfModified, since we cannot obtain cmax from a combo CID
3198 : * generated by another transaction).
3199 : */
3200 : TM_Result
3201 2394797 : heap_update(Relation relation, const ItemPointerData *otid, HeapTuple newtup,
3202 : CommandId cid, uint32 options pg_attribute_unused(), Snapshot crosscheck, bool wait,
3203 : TM_FailureData *tmfd, LockTupleMode *lockmode,
3204 : TU_UpdateIndexes *update_indexes)
3205 : {
3206 : TM_Result result;
3207 2394797 : TransactionId xid = GetCurrentTransactionId();
3208 : Bitmapset *hot_attrs;
3209 : Bitmapset *sum_attrs;
3210 : Bitmapset *key_attrs;
3211 : Bitmapset *id_attrs;
3212 : Bitmapset *interesting_attrs;
3213 : Bitmapset *modified_attrs;
3214 : ItemId lp;
3215 : HeapTupleData oldtup;
3216 : HeapTuple heaptup;
3217 2394797 : HeapTuple old_key_tuple = NULL;
3218 2394797 : bool old_key_copied = false;
3219 2394797 : bool walLogical = (options & TABLE_UPDATE_NO_LOGICAL) == 0;
3220 : Page page,
3221 : newpage;
3222 : BlockNumber block;
3223 : MultiXactStatus mxact_status;
3224 : Buffer buffer,
3225 : newbuf,
3226 2394797 : vmbuffer = InvalidBuffer,
3227 2394797 : vmbuffer_new = InvalidBuffer;
3228 : bool need_toast;
3229 : Size newtupsize,
3230 : pagefree;
3231 2394797 : bool have_tuple_lock = false;
3232 : bool iscombo;
3233 2394797 : bool use_hot_update = false;
3234 2394797 : bool summarized_update = false;
3235 : bool key_intact;
3236 2394797 : bool all_visible_cleared = false;
3237 2394797 : bool all_visible_cleared_new = false;
3238 : bool checked_lockers;
3239 : bool locker_remains;
3240 2394797 : bool id_has_external = false;
3241 : TransactionId xmax_new_tuple,
3242 : xmax_old_tuple;
3243 : uint16 infomask_old_tuple,
3244 : infomask2_old_tuple,
3245 : infomask_new_tuple,
3246 : infomask2_new_tuple;
3247 :
3248 : Assert(ItemPointerIsValid(otid));
3249 :
3250 : /* Cheap, simplistic check that the tuple matches the rel's rowtype. */
3251 : Assert(HeapTupleHeaderGetNatts(newtup->t_data) <=
3252 : RelationGetNumberOfAttributes(relation));
3253 :
3254 2394797 : AssertHasSnapshotForToast(relation);
3255 :
3256 : /*
3257 : * Forbid this during a parallel operation, lest it allocate a combo CID.
3258 : * Other workers might need that combo CID for visibility checks, and we
3259 : * have no provision for broadcasting it to them.
3260 : */
3261 2394797 : if (IsInParallelMode())
3262 0 : ereport(ERROR,
3263 : (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
3264 : errmsg("cannot update tuples during a parallel operation")));
3265 :
3266 : #ifdef USE_ASSERT_CHECKING
3267 : check_lock_if_inplace_updateable_rel(relation, otid, newtup);
3268 : #endif
3269 :
3270 : /*
3271 : * Fetch the list of attributes to be checked for various operations.
3272 : *
3273 : * For HOT considerations, this is wasted effort if we fail to update or
3274 : * have to put the new tuple on a different page. But we must compute the
3275 : * list before obtaining buffer lock --- in the worst case, if we are
3276 : * doing an update on one of the relevant system catalogs, we could
3277 : * deadlock if we try to fetch the list later. In any case, the relcache
3278 : * caches the data so this is usually pretty cheap.
3279 : *
3280 : * We also need columns used by the replica identity and columns that are
3281 : * considered the "key" of rows in the table.
3282 : *
3283 : * Note that we get copies of each bitmap, so we need not worry about
3284 : * relcache flush happening midway through.
3285 : */
3286 2394797 : hot_attrs = RelationGetIndexAttrBitmap(relation,
3287 : INDEX_ATTR_BITMAP_HOT_BLOCKING);
3288 2394797 : sum_attrs = RelationGetIndexAttrBitmap(relation,
3289 : INDEX_ATTR_BITMAP_SUMMARIZED);
3290 2394797 : key_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_KEY);
3291 2394797 : id_attrs = RelationGetIndexAttrBitmap(relation,
3292 : INDEX_ATTR_BITMAP_IDENTITY_KEY);
3293 2394797 : interesting_attrs = NULL;
3294 2394797 : interesting_attrs = bms_add_members(interesting_attrs, hot_attrs);
3295 2394797 : interesting_attrs = bms_add_members(interesting_attrs, sum_attrs);
3296 2394797 : interesting_attrs = bms_add_members(interesting_attrs, key_attrs);
3297 2394797 : interesting_attrs = bms_add_members(interesting_attrs, id_attrs);
3298 :
3299 2394797 : block = ItemPointerGetBlockNumber(otid);
3300 2394797 : INJECTION_POINT("heap_update-before-pin", NULL);
3301 2394797 : buffer = ReadBuffer(relation, block);
3302 2394797 : page = BufferGetPage(buffer);
3303 :
3304 : /*
3305 : * Before locking the buffer, pin the visibility map page if it appears to
3306 : * be necessary. Since we haven't got the lock yet, someone else might be
3307 : * in the middle of changing this, so we'll need to recheck after we have
3308 : * the lock.
3309 : */
3310 2394797 : if (PageIsAllVisible(page))
3311 2140 : visibilitymap_pin(relation, block, &vmbuffer);
3312 :
3313 2394797 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3314 :
3315 2394797 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(otid));
3316 :
3317 : /*
3318 : * Usually, a buffer pin and/or snapshot blocks pruning of otid, ensuring
3319 : * we see LP_NORMAL here. When the otid origin is a syscache, we may have
3320 : * neither a pin nor a snapshot. Hence, we may see other LP_ states, each
3321 : * of which indicates concurrent pruning.
3322 : *
3323 : * Failing with TM_Updated would be most accurate. However, unlike other
3324 : * TM_Updated scenarios, we don't know the successor ctid in LP_UNUSED and
3325 : * LP_DEAD cases. While the distinction between TM_Updated and TM_Deleted
3326 : * does matter to SQL statements UPDATE and MERGE, those SQL statements
3327 : * hold a snapshot that ensures LP_NORMAL. Hence, the choice between
3328 : * TM_Updated and TM_Deleted affects only the wording of error messages.
3329 : * Settle on TM_Deleted, for two reasons. First, it avoids complicating
3330 : * the specification of when tmfd->ctid is valid. Second, it creates
3331 : * error log evidence that we took this branch.
3332 : *
3333 : * Since it's possible to see LP_UNUSED at otid, it's also possible to see
3334 : * LP_NORMAL for a tuple that replaced LP_UNUSED. If it's a tuple for an
3335 : * unrelated row, we'll fail with "duplicate key value violates unique".
3336 : * XXX if otid is the live, newer version of the newtup row, we'll discard
3337 : * changes originating in versions of this catalog row after the version
3338 : * the caller got from syscache. See syscache-update-pruned.spec.
3339 : */
3340 2394797 : if (!ItemIdIsNormal(lp))
3341 : {
3342 : Assert(RelationSupportsSysCache(RelationGetRelid(relation)));
3343 :
3344 1 : UnlockReleaseBuffer(buffer);
3345 : Assert(!have_tuple_lock);
3346 1 : if (vmbuffer != InvalidBuffer)
3347 1 : ReleaseBuffer(vmbuffer);
3348 1 : tmfd->ctid = *otid;
3349 1 : tmfd->xmax = InvalidTransactionId;
3350 1 : tmfd->cmax = InvalidCommandId;
3351 1 : *update_indexes = TU_None;
3352 :
3353 1 : bms_free(hot_attrs);
3354 1 : bms_free(sum_attrs);
3355 1 : bms_free(key_attrs);
3356 1 : bms_free(id_attrs);
3357 : /* modified_attrs not yet initialized */
3358 1 : bms_free(interesting_attrs);
3359 1 : return TM_Deleted;
3360 : }
3361 :
3362 : /*
3363 : * Fill in enough data in oldtup for HeapDetermineColumnsInfo to work
3364 : * properly.
3365 : */
3366 2394796 : oldtup.t_tableOid = RelationGetRelid(relation);
3367 2394796 : oldtup.t_data = (HeapTupleHeader) PageGetItem(page, lp);
3368 2394796 : oldtup.t_len = ItemIdGetLength(lp);
3369 2394796 : oldtup.t_self = *otid;
3370 :
3371 : /* the new tuple is ready, except for this: */
3372 2394796 : newtup->t_tableOid = RelationGetRelid(relation);
3373 :
3374 : /*
3375 : * Determine columns modified by the update. Additionally, identify
3376 : * whether any of the unmodified replica identity key attributes in the
3377 : * old tuple is externally stored or not. This is required because for
3378 : * such attributes the flattened value won't be WAL logged as part of the
3379 : * new tuple so we must include it as part of the old_key_tuple. See
3380 : * ExtractReplicaIdentity.
3381 : */
3382 2394796 : modified_attrs = HeapDetermineColumnsInfo(relation, interesting_attrs,
3383 : id_attrs, &oldtup,
3384 : newtup, &id_has_external);
3385 :
3386 : /*
3387 : * If we're not updating any "key" column, we can grab a weaker lock type.
3388 : * This allows for more concurrency when we are running simultaneously
3389 : * with foreign key checks.
3390 : *
3391 : * Note that if a column gets detoasted while executing the update, but
3392 : * the value ends up being the same, this test will fail and we will use
3393 : * the stronger lock. This is acceptable; the important case to optimize
3394 : * is updates that don't manipulate key columns, not those that
3395 : * serendipitously arrive at the same key values.
3396 : */
3397 2394796 : if (!bms_overlap(modified_attrs, key_attrs))
3398 : {
3399 2388884 : *lockmode = LockTupleNoKeyExclusive;
3400 2388884 : mxact_status = MultiXactStatusNoKeyUpdate;
3401 2388884 : key_intact = true;
3402 :
3403 : /*
3404 : * If this is the first possibly-multixact-able operation in the
3405 : * current transaction, set my per-backend OldestMemberMXactId
3406 : * setting. We can be certain that the transaction will never become a
3407 : * member of any older MultiXactIds than that. (We have to do this
3408 : * even if we end up just using our own TransactionId below, since
3409 : * some other backend could incorporate our XID into a MultiXact
3410 : * immediately afterwards.)
3411 : */
3412 2388884 : MultiXactIdSetOldestMember();
3413 : }
3414 : else
3415 : {
3416 5912 : *lockmode = LockTupleExclusive;
3417 5912 : mxact_status = MultiXactStatusUpdate;
3418 5912 : key_intact = false;
3419 : }
3420 :
3421 : /*
3422 : * Note: beyond this point, use oldtup not otid to refer to old tuple.
3423 : * otid may very well point at newtup->t_self, which we will overwrite
3424 : * with the new tuple's location, so there's great risk of confusion if we
3425 : * use otid anymore.
3426 : */
3427 :
3428 1 : l2:
3429 2394797 : checked_lockers = false;
3430 2394797 : locker_remains = false;
3431 2394797 : result = HeapTupleSatisfiesUpdate(&oldtup, cid, buffer);
3432 :
3433 : /* see below about the "no wait" case */
3434 : Assert(result != TM_BeingModified || wait);
3435 :
3436 2394797 : if (result == TM_Invisible)
3437 : {
3438 0 : UnlockReleaseBuffer(buffer);
3439 0 : ereport(ERROR,
3440 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
3441 : errmsg("attempted to update invisible tuple")));
3442 : }
3443 2394797 : else if (result == TM_BeingModified && wait)
3444 : {
3445 : TransactionId xwait;
3446 : uint16 infomask;
3447 36540 : bool can_continue = false;
3448 :
3449 : /*
3450 : * XXX note that we don't consider the "no wait" case here. This
3451 : * isn't a problem currently because no caller uses that case, but it
3452 : * should be fixed if such a caller is introduced. It wasn't a
3453 : * problem previously because this code would always wait, but now
3454 : * that some tuple locks do not conflict with one of the lock modes we
3455 : * use, it is possible that this case is interesting to handle
3456 : * specially.
3457 : *
3458 : * This may cause failures with third-party code that calls
3459 : * heap_update directly.
3460 : */
3461 :
3462 : /* must copy state data before unlocking buffer */
3463 36540 : xwait = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3464 36540 : infomask = oldtup.t_data->t_infomask;
3465 :
3466 : /*
3467 : * Now we have to do something about the existing locker. If it's a
3468 : * multi, sleep on it; we might be awakened before it is completely
3469 : * gone (or even not sleep at all in some cases); we need to preserve
3470 : * it as locker, unless it is gone completely.
3471 : *
3472 : * If it's not a multi, we need to check for sleeping conditions
3473 : * before actually going to sleep. If the update doesn't conflict
3474 : * with the locks, we just continue without sleeping (but making sure
3475 : * it is preserved).
3476 : *
3477 : * Before sleeping, we need to acquire tuple lock to establish our
3478 : * priority for the tuple (see heap_lock_tuple). LockTuple will
3479 : * release us when we are next-in-line for the tuple. Note we must
3480 : * not acquire the tuple lock until we're sure we're going to sleep;
3481 : * otherwise we're open for race conditions with other transactions
3482 : * holding the tuple lock which sleep on us.
3483 : *
3484 : * If we are forced to "start over" below, we keep the tuple lock;
3485 : * this arranges that we stay at the head of the line while rechecking
3486 : * tuple state.
3487 : */
3488 36540 : if (infomask & HEAP_XMAX_IS_MULTI)
3489 : {
3490 : TransactionId update_xact;
3491 : int remain;
3492 179 : bool current_is_member = false;
3493 :
3494 179 : if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
3495 : *lockmode, ¤t_is_member))
3496 : {
3497 8 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3498 :
3499 : /*
3500 : * Acquire the lock, if necessary (but skip it when we're
3501 : * requesting a lock and already have one; avoids deadlock).
3502 : */
3503 8 : if (!current_is_member)
3504 0 : heap_acquire_tuplock(relation, &(oldtup.t_self), *lockmode,
3505 : LockWaitBlock, &have_tuple_lock);
3506 :
3507 : /* wait for multixact */
3508 8 : MultiXactIdWait((MultiXactId) xwait, mxact_status, infomask,
3509 : relation, &oldtup.t_self, XLTW_Update,
3510 : &remain);
3511 8 : checked_lockers = true;
3512 8 : locker_remains = remain != 0;
3513 8 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3514 :
3515 : /*
3516 : * If xwait had just locked the tuple then some other xact
3517 : * could update this tuple before we get to this point. Check
3518 : * for xmax change, and start over if so.
3519 : */
3520 8 : if (xmax_infomask_changed(oldtup.t_data->t_infomask,
3521 8 : infomask) ||
3522 8 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3523 : xwait))
3524 0 : goto l2;
3525 : }
3526 :
3527 : /*
3528 : * Note that the multixact may not be done by now. It could have
3529 : * surviving members; our own xact or other subxacts of this
3530 : * backend, and also any other concurrent transaction that locked
3531 : * the tuple with LockTupleKeyShare if we only got
3532 : * LockTupleNoKeyExclusive. If this is the case, we have to be
3533 : * careful to mark the updated tuple with the surviving members in
3534 : * Xmax.
3535 : *
3536 : * Note that there could have been another update in the
3537 : * MultiXact. In that case, we need to check whether it committed
3538 : * or aborted. If it aborted we are safe to update it again;
3539 : * otherwise there is an update conflict, and we have to return
3540 : * TableTuple{Deleted, Updated} below.
3541 : *
3542 : * In the LockTupleExclusive case, we still need to preserve the
3543 : * surviving members: those would include the tuple locks we had
3544 : * before this one, which are important to keep in case this
3545 : * subxact aborts.
3546 : */
3547 179 : if (!HEAP_XMAX_IS_LOCKED_ONLY(oldtup.t_data->t_infomask))
3548 8 : update_xact = HeapTupleGetUpdateXid(oldtup.t_data);
3549 : else
3550 171 : update_xact = InvalidTransactionId;
3551 :
3552 : /*
3553 : * There was no UPDATE in the MultiXact; or it aborted. No
3554 : * TransactionIdIsInProgress() call needed here, since we called
3555 : * MultiXactIdWait() above.
3556 : */
3557 187 : if (!TransactionIdIsValid(update_xact) ||
3558 8 : TransactionIdDidAbort(update_xact))
3559 172 : can_continue = true;
3560 : }
3561 36361 : else if (TransactionIdIsCurrentTransactionId(xwait))
3562 : {
3563 : /*
3564 : * The only locker is ourselves; we can avoid grabbing the tuple
3565 : * lock here, but must preserve our locking information.
3566 : */
3567 36233 : checked_lockers = true;
3568 36233 : locker_remains = true;
3569 36233 : can_continue = true;
3570 : }
3571 128 : else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) && key_intact)
3572 : {
3573 : /*
3574 : * If it's just a key-share locker, and we're not changing the key
3575 : * columns, we don't need to wait for it to end; but we need to
3576 : * preserve it as locker.
3577 : */
3578 29 : checked_lockers = true;
3579 29 : locker_remains = true;
3580 29 : can_continue = true;
3581 : }
3582 : else
3583 : {
3584 : /*
3585 : * Wait for regular transaction to end; but first, acquire tuple
3586 : * lock.
3587 : */
3588 99 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3589 99 : heap_acquire_tuplock(relation, &(oldtup.t_self), *lockmode,
3590 : LockWaitBlock, &have_tuple_lock);
3591 99 : XactLockTableWait(xwait, relation, &oldtup.t_self,
3592 : XLTW_Update);
3593 99 : checked_lockers = true;
3594 99 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3595 :
3596 : /*
3597 : * xwait is done, but if xwait had just locked the tuple then some
3598 : * other xact could update this tuple before we get to this point.
3599 : * Check for xmax change, and start over if so.
3600 : */
3601 197 : if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3602 98 : !TransactionIdEquals(xwait,
3603 : HeapTupleHeaderGetRawXmax(oldtup.t_data)))
3604 1 : goto l2;
3605 :
3606 : /* Otherwise check if it committed or aborted */
3607 98 : UpdateXmaxHintBits(oldtup.t_data, buffer, xwait);
3608 98 : if (oldtup.t_data->t_infomask & HEAP_XMAX_INVALID)
3609 22 : can_continue = true;
3610 : }
3611 :
3612 36539 : if (can_continue)
3613 36456 : result = TM_Ok;
3614 83 : else if (!ItemPointerEquals(&oldtup.t_self, &oldtup.t_data->t_ctid))
3615 69 : result = TM_Updated;
3616 : else
3617 14 : result = TM_Deleted;
3618 : }
3619 :
3620 : /* Sanity check the result HeapTupleSatisfiesUpdate() and the logic above */
3621 : if (result != TM_Ok)
3622 : {
3623 : Assert(result == TM_SelfModified ||
3624 : result == TM_Updated ||
3625 : result == TM_Deleted ||
3626 : result == TM_BeingModified);
3627 : Assert(!(oldtup.t_data->t_infomask & HEAP_XMAX_INVALID));
3628 : Assert(result != TM_Updated ||
3629 : !ItemPointerEquals(&oldtup.t_self, &oldtup.t_data->t_ctid));
3630 : }
3631 :
3632 2394796 : if (crosscheck != InvalidSnapshot && result == TM_Ok)
3633 : {
3634 : /* Perform additional check for transaction-snapshot mode RI updates */
3635 1 : if (!HeapTupleSatisfiesVisibility(&oldtup, crosscheck, buffer))
3636 1 : result = TM_Updated;
3637 : }
3638 :
3639 2394796 : if (result != TM_Ok)
3640 : {
3641 209 : tmfd->ctid = oldtup.t_data->t_ctid;
3642 209 : tmfd->xmax = HeapTupleHeaderGetUpdateXid(oldtup.t_data);
3643 209 : if (result == TM_SelfModified)
3644 73 : tmfd->cmax = HeapTupleHeaderGetCmax(oldtup.t_data);
3645 : else
3646 136 : tmfd->cmax = InvalidCommandId;
3647 209 : UnlockReleaseBuffer(buffer);
3648 209 : if (have_tuple_lock)
3649 76 : UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3650 209 : if (vmbuffer != InvalidBuffer)
3651 0 : ReleaseBuffer(vmbuffer);
3652 209 : *update_indexes = TU_None;
3653 :
3654 209 : bms_free(hot_attrs);
3655 209 : bms_free(sum_attrs);
3656 209 : bms_free(key_attrs);
3657 209 : bms_free(id_attrs);
3658 209 : bms_free(modified_attrs);
3659 209 : bms_free(interesting_attrs);
3660 209 : return result;
3661 : }
3662 :
3663 : /*
3664 : * If we didn't pin the visibility map page and the page has become all
3665 : * visible while we were busy locking the buffer, or during some
3666 : * subsequent window during which we had it unlocked, we'll have to unlock
3667 : * and re-lock, to avoid holding the buffer lock across an I/O. That's a
3668 : * bit unfortunate, especially since we'll now have to recheck whether the
3669 : * tuple has been locked or updated under us, but hopefully it won't
3670 : * happen very often.
3671 : */
3672 2394587 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3673 : {
3674 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3675 0 : visibilitymap_pin(relation, block, &vmbuffer);
3676 0 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3677 0 : goto l2;
3678 : }
3679 :
3680 : /* Fill in transaction status data */
3681 :
3682 : /*
3683 : * If the tuple we're updating is locked, we need to preserve the locking
3684 : * info in the old tuple's Xmax. Prepare a new Xmax value for this.
3685 : */
3686 2394587 : compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3687 2394587 : oldtup.t_data->t_infomask,
3688 2394587 : oldtup.t_data->t_infomask2,
3689 : xid, *lockmode, true,
3690 : &xmax_old_tuple, &infomask_old_tuple,
3691 : &infomask2_old_tuple);
3692 :
3693 : /*
3694 : * And also prepare an Xmax value for the new copy of the tuple. If there
3695 : * was no xmax previously, or there was one but all lockers are now gone,
3696 : * then use InvalidTransactionId; otherwise, get the xmax from the old
3697 : * tuple. (In rare cases that might also be InvalidTransactionId and yet
3698 : * not have the HEAP_XMAX_INVALID bit set; that's fine.)
3699 : */
3700 2431021 : if ((oldtup.t_data->t_infomask & HEAP_XMAX_INVALID) ||
3701 72868 : HEAP_LOCKED_UPGRADED(oldtup.t_data->t_infomask) ||
3702 36263 : (checked_lockers && !locker_remains))
3703 2358153 : xmax_new_tuple = InvalidTransactionId;
3704 : else
3705 36434 : xmax_new_tuple = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3706 :
3707 2394587 : if (!TransactionIdIsValid(xmax_new_tuple))
3708 : {
3709 2358153 : infomask_new_tuple = HEAP_XMAX_INVALID;
3710 2358153 : infomask2_new_tuple = 0;
3711 : }
3712 : else
3713 : {
3714 : /*
3715 : * If we found a valid Xmax for the new tuple, then the infomask bits
3716 : * to use on the new tuple depend on what was there on the old one.
3717 : * Note that since we're doing an update, the only possibility is that
3718 : * the lockers had FOR KEY SHARE lock.
3719 : */
3720 36434 : if (oldtup.t_data->t_infomask & HEAP_XMAX_IS_MULTI)
3721 : {
3722 172 : GetMultiXactIdHintBits(xmax_new_tuple, &infomask_new_tuple,
3723 : &infomask2_new_tuple);
3724 : }
3725 : else
3726 : {
3727 36262 : infomask_new_tuple = HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_LOCK_ONLY;
3728 36262 : infomask2_new_tuple = 0;
3729 : }
3730 : }
3731 :
3732 : /*
3733 : * Prepare the new tuple with the appropriate initial values of Xmin and
3734 : * Xmax, as well as initial infomask bits as computed above.
3735 : */
3736 2394587 : newtup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
3737 2394587 : newtup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
3738 2394587 : HeapTupleHeaderSetXmin(newtup->t_data, xid);
3739 2394587 : HeapTupleHeaderSetCmin(newtup->t_data, cid);
3740 2394587 : newtup->t_data->t_infomask |= HEAP_UPDATED | infomask_new_tuple;
3741 2394587 : newtup->t_data->t_infomask2 |= infomask2_new_tuple;
3742 2394587 : HeapTupleHeaderSetXmax(newtup->t_data, xmax_new_tuple);
3743 :
3744 : /*
3745 : * Replace cid with a combo CID if necessary. Note that we already put
3746 : * the plain cid into the new tuple.
3747 : */
3748 2394587 : HeapTupleHeaderAdjustCmax(oldtup.t_data, &cid, &iscombo);
3749 :
3750 : /*
3751 : * If the toaster needs to be activated, OR if the new tuple will not fit
3752 : * on the same page as the old, then we need to release the content lock
3753 : * (but not the pin!) on the old tuple's buffer while we are off doing
3754 : * TOAST and/or table-file-extension work. We must mark the old tuple to
3755 : * show that it's locked, else other processes may try to update it
3756 : * themselves.
3757 : *
3758 : * We need to invoke the toaster if there are already any out-of-line
3759 : * toasted values present, or if the new tuple is over-threshold.
3760 : */
3761 2394587 : if (relation->rd_rel->relkind != RELKIND_RELATION &&
3762 0 : relation->rd_rel->relkind != RELKIND_MATVIEW)
3763 : {
3764 : /* toast table entries should never be recursively toasted */
3765 : Assert(!HeapTupleHasExternal(&oldtup));
3766 : Assert(!HeapTupleHasExternal(newtup));
3767 0 : need_toast = false;
3768 : }
3769 : else
3770 7183244 : need_toast = (HeapTupleHasExternal(&oldtup) ||
3771 4788657 : HeapTupleHasExternal(newtup) ||
3772 2394038 : newtup->t_len > TOAST_TUPLE_THRESHOLD);
3773 :
3774 2394587 : pagefree = PageGetHeapFreeSpace(page);
3775 :
3776 2394587 : newtupsize = MAXALIGN(newtup->t_len);
3777 :
3778 2394587 : if (need_toast || newtupsize > pagefree)
3779 2207443 : {
3780 : TransactionId xmax_lock_old_tuple;
3781 : uint16 infomask_lock_old_tuple,
3782 : infomask2_lock_old_tuple;
3783 2207443 : bool cleared_all_frozen = false;
3784 :
3785 : /*
3786 : * To prevent concurrent sessions from updating the tuple, we have to
3787 : * temporarily mark it locked, while we release the page-level lock.
3788 : *
3789 : * To satisfy the rule that any xid potentially appearing in a buffer
3790 : * written out to disk, we unfortunately have to WAL log this
3791 : * temporary modification. We can reuse xl_heap_lock for this
3792 : * purpose. If we crash/error before following through with the
3793 : * actual update, xmax will be of an aborted transaction, allowing
3794 : * other sessions to proceed.
3795 : */
3796 :
3797 : /*
3798 : * Compute xmax / infomask appropriate for locking the tuple. This has
3799 : * to be done separately from the combo that's going to be used for
3800 : * updating, because the potentially created multixact would otherwise
3801 : * be wrong.
3802 : */
3803 2207443 : compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3804 2207443 : oldtup.t_data->t_infomask,
3805 2207443 : oldtup.t_data->t_infomask2,
3806 : xid, *lockmode, false,
3807 : &xmax_lock_old_tuple, &infomask_lock_old_tuple,
3808 : &infomask2_lock_old_tuple);
3809 :
3810 : Assert(HEAP_XMAX_IS_LOCKED_ONLY(infomask_lock_old_tuple));
3811 :
3812 2207443 : START_CRIT_SECTION();
3813 :
3814 : /* Clear obsolete visibility flags ... */
3815 2207443 : oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3816 2207443 : oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3817 2207443 : HeapTupleClearHotUpdated(&oldtup);
3818 : /* ... and store info about transaction updating this tuple */
3819 : Assert(TransactionIdIsValid(xmax_lock_old_tuple));
3820 2207443 : HeapTupleHeaderSetXmax(oldtup.t_data, xmax_lock_old_tuple);
3821 2207443 : oldtup.t_data->t_infomask |= infomask_lock_old_tuple;
3822 2207443 : oldtup.t_data->t_infomask2 |= infomask2_lock_old_tuple;
3823 2207443 : HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
3824 :
3825 : /* temporarily make it look not-updated, but locked */
3826 2207443 : oldtup.t_data->t_ctid = oldtup.t_self;
3827 :
3828 : /*
3829 : * Clear all-frozen bit on visibility map if needed. We could
3830 : * immediately reset ALL_VISIBLE, but given that the WAL logging
3831 : * overhead would be unchanged, that doesn't seem necessarily
3832 : * worthwhile.
3833 : */
3834 2208735 : if (PageIsAllVisible(page) &&
3835 1292 : visibilitymap_clear(relation, block, vmbuffer,
3836 : VISIBILITYMAP_ALL_FROZEN))
3837 842 : cleared_all_frozen = true;
3838 :
3839 2207443 : MarkBufferDirty(buffer);
3840 :
3841 2207443 : if (RelationNeedsWAL(relation))
3842 : {
3843 : xl_heap_lock xlrec;
3844 : XLogRecPtr recptr;
3845 :
3846 2197307 : XLogBeginInsert();
3847 2197307 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
3848 :
3849 2197307 : xlrec.offnum = ItemPointerGetOffsetNumber(&oldtup.t_self);
3850 2197307 : xlrec.xmax = xmax_lock_old_tuple;
3851 4394614 : xlrec.infobits_set = compute_infobits(oldtup.t_data->t_infomask,
3852 2197307 : oldtup.t_data->t_infomask2);
3853 2197307 : xlrec.flags =
3854 2197307 : cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
3855 2197307 : XLogRegisterData(&xlrec, SizeOfHeapLock);
3856 2197307 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK);
3857 2197307 : PageSetLSN(page, recptr);
3858 : }
3859 :
3860 2207443 : END_CRIT_SECTION();
3861 :
3862 2207443 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3863 :
3864 : /*
3865 : * Let the toaster do its thing, if needed.
3866 : *
3867 : * Note: below this point, heaptup is the data we actually intend to
3868 : * store into the relation; newtup is the caller's original untoasted
3869 : * data.
3870 : */
3871 2207443 : if (need_toast)
3872 : {
3873 : /* Note we always use WAL and FSM during updates */
3874 1914 : heaptup = heap_toast_insert_or_update(relation, newtup, &oldtup, 0);
3875 1914 : newtupsize = MAXALIGN(heaptup->t_len);
3876 : }
3877 : else
3878 2205529 : heaptup = newtup;
3879 :
3880 : /*
3881 : * Now, do we need a new page for the tuple, or not? This is a bit
3882 : * tricky since someone else could have added tuples to the page while
3883 : * we weren't looking. We have to recheck the available space after
3884 : * reacquiring the buffer lock. But don't bother to do that if the
3885 : * former amount of free space is still not enough; it's unlikely
3886 : * there's more free now than before.
3887 : *
3888 : * What's more, if we need to get a new page, we will need to acquire
3889 : * buffer locks on both old and new pages. To avoid deadlock against
3890 : * some other backend trying to get the same two locks in the other
3891 : * order, we must be consistent about the order we get the locks in.
3892 : * We use the rule "lock the lower-numbered page of the relation
3893 : * first". To implement this, we must do RelationGetBufferForTuple
3894 : * while not holding the lock on the old page, and we must rely on it
3895 : * to get the locks on both pages in the correct order.
3896 : *
3897 : * Another consideration is that we need visibility map page pin(s) if
3898 : * we will have to clear the all-visible flag on either page. If we
3899 : * call RelationGetBufferForTuple, we rely on it to acquire any such
3900 : * pins; but if we don't, we have to handle that here. Hence we need
3901 : * a loop.
3902 : */
3903 : for (;;)
3904 : {
3905 2207443 : if (newtupsize > pagefree)
3906 : {
3907 : /* It doesn't fit, must use RelationGetBufferForTuple. */
3908 2206712 : newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3909 : buffer, 0, NULL,
3910 : &vmbuffer_new, &vmbuffer,
3911 : 0);
3912 : /* We're all done. */
3913 2206712 : break;
3914 : }
3915 : /* Acquire VM page pin if needed and we don't have it. */
3916 731 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3917 0 : visibilitymap_pin(relation, block, &vmbuffer);
3918 : /* Re-acquire the lock on the old tuple's page. */
3919 731 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3920 : /* Re-check using the up-to-date free space */
3921 731 : pagefree = PageGetHeapFreeSpace(page);
3922 731 : if (newtupsize > pagefree ||
3923 731 : (vmbuffer == InvalidBuffer && PageIsAllVisible(page)))
3924 : {
3925 : /*
3926 : * Rats, it doesn't fit anymore, or somebody just now set the
3927 : * all-visible flag. We must now unlock and loop to avoid
3928 : * deadlock. Fortunately, this path should seldom be taken.
3929 : */
3930 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3931 : }
3932 : else
3933 : {
3934 : /* We're all done. */
3935 731 : newbuf = buffer;
3936 731 : break;
3937 : }
3938 : }
3939 : }
3940 : else
3941 : {
3942 : /* No TOAST work needed, and it'll fit on same page */
3943 187144 : newbuf = buffer;
3944 187144 : heaptup = newtup;
3945 : }
3946 :
3947 2394587 : newpage = BufferGetPage(newbuf);
3948 :
3949 : /*
3950 : * We're about to do the actual update -- check for conflict first, to
3951 : * avoid possibly having to roll back work we've just done.
3952 : *
3953 : * This is safe without a recheck as long as there is no possibility of
3954 : * another process scanning the pages between this check and the update
3955 : * being visible to the scan (i.e., exclusive buffer content lock(s) are
3956 : * continuously held from this point until the tuple update is visible).
3957 : *
3958 : * For the new tuple the only check needed is at the relation level, but
3959 : * since both tuples are in the same relation and the check for oldtup
3960 : * will include checking the relation level, there is no benefit to a
3961 : * separate check for the new tuple.
3962 : */
3963 2394587 : CheckForSerializableConflictIn(relation, &oldtup.t_self,
3964 : BufferGetBlockNumber(buffer));
3965 :
3966 : /*
3967 : * At this point newbuf and buffer are both pinned and locked, and newbuf
3968 : * has enough space for the new tuple. If they are the same buffer, only
3969 : * one pin is held.
3970 : */
3971 :
3972 2394575 : if (newbuf == buffer)
3973 : {
3974 : /*
3975 : * Since the new tuple is going into the same page, we might be able
3976 : * to do a HOT update. Check if any of the index columns have been
3977 : * changed.
3978 : */
3979 187863 : if (!bms_overlap(modified_attrs, hot_attrs))
3980 : {
3981 171178 : use_hot_update = true;
3982 :
3983 : /*
3984 : * If none of the columns that are used in hot-blocking indexes
3985 : * were updated, we can apply HOT, but we do still need to check
3986 : * if we need to update the summarizing indexes, and update those
3987 : * indexes if the columns were updated, or we may fail to detect
3988 : * e.g. value bound changes in BRIN minmax indexes.
3989 : */
3990 171178 : if (bms_overlap(modified_attrs, sum_attrs))
3991 2188 : summarized_update = true;
3992 : }
3993 : }
3994 : else
3995 : {
3996 : /* Set a hint that the old page could use prune/defrag */
3997 2206712 : PageSetFull(page);
3998 : }
3999 :
4000 : /*
4001 : * Compute replica identity tuple before entering the critical section so
4002 : * we don't PANIC upon a memory allocation failure.
4003 : * ExtractReplicaIdentity() will return NULL if nothing needs to be
4004 : * logged. Pass old key required as true only if the replica identity key
4005 : * columns are modified or it has external data.
4006 : */
4007 2394575 : old_key_tuple = ExtractReplicaIdentity(relation, &oldtup,
4008 2394575 : bms_overlap(modified_attrs, id_attrs) ||
4009 : id_has_external,
4010 2394575 : &old_key_copied);
4011 :
4012 : /* NO EREPORT(ERROR) from here till changes are logged */
4013 2394575 : START_CRIT_SECTION();
4014 :
4015 : /*
4016 : * If this transaction commits, the old tuple will become DEAD sooner or
4017 : * later. Set flag that this page is a candidate for pruning once our xid
4018 : * falls below the OldestXmin horizon. If the transaction finally aborts,
4019 : * the subsequent page pruning will be a no-op and the hint will be
4020 : * cleared.
4021 : *
4022 : * We set the new page prunable as well. See heap_insert() for more on why
4023 : * we do this when inserting tuples.
4024 : */
4025 2394575 : PageSetPrunable(page, xid);
4026 2394575 : if (newbuf != buffer)
4027 2206712 : PageSetPrunable(newpage, xid);
4028 :
4029 2394575 : if (use_hot_update)
4030 : {
4031 : /* Mark the old tuple as HOT-updated */
4032 171178 : HeapTupleSetHotUpdated(&oldtup);
4033 : /* And mark the new tuple as heap-only */
4034 171178 : HeapTupleSetHeapOnly(heaptup);
4035 : /* Mark the caller's copy too, in case different from heaptup */
4036 171178 : HeapTupleSetHeapOnly(newtup);
4037 : }
4038 : else
4039 : {
4040 : /* Make sure tuples are correctly marked as not-HOT */
4041 2223397 : HeapTupleClearHotUpdated(&oldtup);
4042 2223397 : HeapTupleClearHeapOnly(heaptup);
4043 2223397 : HeapTupleClearHeapOnly(newtup);
4044 : }
4045 :
4046 2394575 : RelationPutHeapTuple(relation, newbuf, heaptup, false); /* insert new tuple */
4047 :
4048 :
4049 : /* Clear obsolete visibility flags, possibly set by ourselves above... */
4050 2394575 : oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
4051 2394575 : oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
4052 : /* ... and store info about transaction updating this tuple */
4053 : Assert(TransactionIdIsValid(xmax_old_tuple));
4054 2394575 : HeapTupleHeaderSetXmax(oldtup.t_data, xmax_old_tuple);
4055 2394575 : oldtup.t_data->t_infomask |= infomask_old_tuple;
4056 2394575 : oldtup.t_data->t_infomask2 |= infomask2_old_tuple;
4057 2394575 : HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
4058 :
4059 : /* record address of new tuple in t_ctid of old one */
4060 2394575 : oldtup.t_data->t_ctid = heaptup->t_self;
4061 :
4062 : /* clear PD_ALL_VISIBLE flags, reset all visibilitymap bits */
4063 2394575 : if (PageIsAllVisible(page))
4064 : {
4065 2139 : all_visible_cleared = true;
4066 2139 : PageClearAllVisible(page);
4067 2139 : visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
4068 : vmbuffer, VISIBILITYMAP_VALID_BITS);
4069 : }
4070 2394575 : if (newbuf != buffer && PageIsAllVisible(newpage))
4071 : {
4072 774 : all_visible_cleared_new = true;
4073 774 : PageClearAllVisible(newpage);
4074 774 : visibilitymap_clear(relation, BufferGetBlockNumber(newbuf),
4075 : vmbuffer_new, VISIBILITYMAP_VALID_BITS);
4076 : }
4077 :
4078 2394575 : if (newbuf != buffer)
4079 2206712 : MarkBufferDirty(newbuf);
4080 2394575 : MarkBufferDirty(buffer);
4081 :
4082 : /* XLOG stuff */
4083 2394575 : if (RelationNeedsWAL(relation))
4084 : {
4085 : XLogRecPtr recptr;
4086 :
4087 : /*
4088 : * For logical decoding we need combo CIDs to properly decode the
4089 : * catalog.
4090 : */
4091 2382791 : if (RelationIsAccessibleInLogicalDecoding(relation))
4092 : {
4093 2689 : log_heap_new_cid(relation, &oldtup);
4094 2689 : log_heap_new_cid(relation, heaptup);
4095 : }
4096 :
4097 2382791 : recptr = log_heap_update(relation, buffer,
4098 : newbuf, &oldtup, heaptup,
4099 : old_key_tuple,
4100 : all_visible_cleared,
4101 : all_visible_cleared_new,
4102 : walLogical);
4103 2382791 : if (newbuf != buffer)
4104 : {
4105 2196584 : PageSetLSN(newpage, recptr);
4106 : }
4107 2382791 : PageSetLSN(page, recptr);
4108 : }
4109 :
4110 2394575 : END_CRIT_SECTION();
4111 :
4112 2394575 : if (newbuf != buffer)
4113 2206712 : LockBuffer(newbuf, BUFFER_LOCK_UNLOCK);
4114 2394575 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
4115 :
4116 : /*
4117 : * Mark old tuple for invalidation from system caches at next command
4118 : * boundary, and mark the new tuple for invalidation in case we abort. We
4119 : * have to do this before releasing the buffer because oldtup is in the
4120 : * buffer. (heaptup is all in local memory, but it's necessary to process
4121 : * both tuple versions in one call to inval.c so we can avoid redundant
4122 : * sinval messages.)
4123 : */
4124 2394575 : CacheInvalidateHeapTuple(relation, &oldtup, heaptup);
4125 :
4126 : /* Now we can release the buffer(s) */
4127 2394575 : if (newbuf != buffer)
4128 2206712 : ReleaseBuffer(newbuf);
4129 2394575 : ReleaseBuffer(buffer);
4130 2394575 : if (BufferIsValid(vmbuffer_new))
4131 775 : ReleaseBuffer(vmbuffer_new);
4132 2394575 : if (BufferIsValid(vmbuffer))
4133 2139 : ReleaseBuffer(vmbuffer);
4134 :
4135 : /*
4136 : * Release the lmgr tuple lock, if we had it.
4137 : */
4138 2394575 : if (have_tuple_lock)
4139 22 : UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
4140 :
4141 2394575 : pgstat_count_heap_update(relation, use_hot_update, newbuf != buffer);
4142 :
4143 : /*
4144 : * If heaptup is a private copy, release it. Don't forget to copy t_self
4145 : * back to the caller's image, too.
4146 : */
4147 2394575 : if (heaptup != newtup)
4148 : {
4149 1858 : newtup->t_self = heaptup->t_self;
4150 1858 : heap_freetuple(heaptup);
4151 : }
4152 :
4153 : /*
4154 : * If it is a HOT update, the update may still need to update summarized
4155 : * indexes, lest we fail to update those summaries and get incorrect
4156 : * results (for example, minmax bounds of the block may change with this
4157 : * update).
4158 : */
4159 2394575 : if (use_hot_update)
4160 : {
4161 171178 : if (summarized_update)
4162 2188 : *update_indexes = TU_Summarizing;
4163 : else
4164 168990 : *update_indexes = TU_None;
4165 : }
4166 : else
4167 2223397 : *update_indexes = TU_All;
4168 :
4169 2394575 : if (old_key_tuple != NULL && old_key_copied)
4170 100 : heap_freetuple(old_key_tuple);
4171 :
4172 2394575 : bms_free(hot_attrs);
4173 2394575 : bms_free(sum_attrs);
4174 2394575 : bms_free(key_attrs);
4175 2394575 : bms_free(id_attrs);
4176 2394575 : bms_free(modified_attrs);
4177 2394575 : bms_free(interesting_attrs);
4178 :
4179 2394575 : return TM_Ok;
4180 : }
4181 :
4182 : #ifdef USE_ASSERT_CHECKING
4183 : /*
4184 : * Confirm adequate lock held during heap_update(), per rules from
4185 : * README.tuplock section "Locking to write inplace-updated tables".
4186 : */
4187 : static void
4188 : check_lock_if_inplace_updateable_rel(Relation relation,
4189 : const ItemPointerData *otid,
4190 : HeapTuple newtup)
4191 : {
4192 : /* LOCKTAG_TUPLE acceptable for any catalog */
4193 : switch (RelationGetRelid(relation))
4194 : {
4195 : case RelationRelationId:
4196 : case DatabaseRelationId:
4197 : {
4198 : LOCKTAG tuptag;
4199 :
4200 : SET_LOCKTAG_TUPLE(tuptag,
4201 : relation->rd_lockInfo.lockRelId.dbId,
4202 : relation->rd_lockInfo.lockRelId.relId,
4203 : ItemPointerGetBlockNumber(otid),
4204 : ItemPointerGetOffsetNumber(otid));
4205 : if (LockHeldByMe(&tuptag, InplaceUpdateTupleLock, false))
4206 : return;
4207 : }
4208 : break;
4209 : default:
4210 : Assert(!IsInplaceUpdateRelation(relation));
4211 : return;
4212 : }
4213 :
4214 : switch (RelationGetRelid(relation))
4215 : {
4216 : case RelationRelationId:
4217 : {
4218 : /* LOCKTAG_TUPLE or LOCKTAG_RELATION ok */
4219 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(newtup);
4220 : Oid relid = classForm->oid;
4221 : Oid dbid;
4222 : LOCKTAG tag;
4223 :
4224 : if (IsSharedRelation(relid))
4225 : dbid = InvalidOid;
4226 : else
4227 : dbid = MyDatabaseId;
4228 :
4229 : if (classForm->relkind == RELKIND_INDEX)
4230 : {
4231 : Relation irel = index_open(relid, AccessShareLock);
4232 :
4233 : SET_LOCKTAG_RELATION(tag, dbid, irel->rd_index->indrelid);
4234 : index_close(irel, AccessShareLock);
4235 : }
4236 : else
4237 : SET_LOCKTAG_RELATION(tag, dbid, relid);
4238 :
4239 : if (!LockHeldByMe(&tag, ShareUpdateExclusiveLock, false) &&
4240 : !LockHeldByMe(&tag, ShareRowExclusiveLock, true))
4241 : elog(WARNING,
4242 : "missing lock for relation \"%s\" (OID %u, relkind %c) @ TID (%u,%u)",
4243 : NameStr(classForm->relname),
4244 : relid,
4245 : classForm->relkind,
4246 : ItemPointerGetBlockNumber(otid),
4247 : ItemPointerGetOffsetNumber(otid));
4248 : }
4249 : break;
4250 : case DatabaseRelationId:
4251 : {
4252 : /* LOCKTAG_TUPLE required */
4253 : Form_pg_database dbForm = (Form_pg_database) GETSTRUCT(newtup);
4254 :
4255 : elog(WARNING,
4256 : "missing lock on database \"%s\" (OID %u) @ TID (%u,%u)",
4257 : NameStr(dbForm->datname),
4258 : dbForm->oid,
4259 : ItemPointerGetBlockNumber(otid),
4260 : ItemPointerGetOffsetNumber(otid));
4261 : }
4262 : break;
4263 : }
4264 : }
4265 :
4266 : /*
4267 : * Confirm adequate relation lock held, per rules from README.tuplock section
4268 : * "Locking to write inplace-updated tables".
4269 : */
4270 : static void
4271 : check_inplace_rel_lock(HeapTuple oldtup)
4272 : {
4273 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(oldtup);
4274 : Oid relid = classForm->oid;
4275 : Oid dbid;
4276 : LOCKTAG tag;
4277 :
4278 : if (IsSharedRelation(relid))
4279 : dbid = InvalidOid;
4280 : else
4281 : dbid = MyDatabaseId;
4282 :
4283 : if (classForm->relkind == RELKIND_INDEX)
4284 : {
4285 : Relation irel = index_open(relid, AccessShareLock);
4286 :
4287 : SET_LOCKTAG_RELATION(tag, dbid, irel->rd_index->indrelid);
4288 : index_close(irel, AccessShareLock);
4289 : }
4290 : else
4291 : SET_LOCKTAG_RELATION(tag, dbid, relid);
4292 :
4293 : if (!LockHeldByMe(&tag, ShareUpdateExclusiveLock, true))
4294 : elog(WARNING,
4295 : "missing lock for relation \"%s\" (OID %u, relkind %c) @ TID (%u,%u)",
4296 : NameStr(classForm->relname),
4297 : relid,
4298 : classForm->relkind,
4299 : ItemPointerGetBlockNumber(&oldtup->t_self),
4300 : ItemPointerGetOffsetNumber(&oldtup->t_self));
4301 : }
4302 : #endif
4303 :
4304 : /*
4305 : * Check if the specified attribute's values are the same. Subroutine for
4306 : * HeapDetermineColumnsInfo.
4307 : */
4308 : static bool
4309 989585 : heap_attr_equals(TupleDesc tupdesc, int attrnum, Datum value1, Datum value2,
4310 : bool isnull1, bool isnull2)
4311 : {
4312 : /*
4313 : * If one value is NULL and other is not, then they are certainly not
4314 : * equal
4315 : */
4316 989585 : if (isnull1 != isnull2)
4317 60 : return false;
4318 :
4319 : /*
4320 : * If both are NULL, they can be considered equal.
4321 : */
4322 989525 : if (isnull1)
4323 6641 : return true;
4324 :
4325 : /*
4326 : * We do simple binary comparison of the two datums. This may be overly
4327 : * strict because there can be multiple binary representations for the
4328 : * same logical value. But we should be OK as long as there are no false
4329 : * positives. Using a type-specific equality operator is messy because
4330 : * there could be multiple notions of equality in different operator
4331 : * classes; furthermore, we cannot safely invoke user-defined functions
4332 : * while holding exclusive buffer lock.
4333 : */
4334 982884 : if (attrnum <= 0)
4335 : {
4336 : /* The only allowed system columns are OIDs, so do this */
4337 0 : return (DatumGetObjectId(value1) == DatumGetObjectId(value2));
4338 : }
4339 : else
4340 : {
4341 : CompactAttribute *att;
4342 :
4343 : Assert(attrnum <= tupdesc->natts);
4344 982884 : att = TupleDescCompactAttr(tupdesc, attrnum - 1);
4345 982884 : return datumIsEqual(value1, value2, att->attbyval, att->attlen);
4346 : }
4347 : }
4348 :
4349 : /*
4350 : * Check which columns are being updated.
4351 : *
4352 : * Given an updated tuple, determine (and return into the output bitmapset),
4353 : * from those listed as interesting, the set of columns that changed.
4354 : *
4355 : * has_external indicates if any of the unmodified attributes (from those
4356 : * listed as interesting) of the old tuple is a member of external_cols and is
4357 : * stored externally.
4358 : */
4359 : static Bitmapset *
4360 2394796 : HeapDetermineColumnsInfo(Relation relation,
4361 : Bitmapset *interesting_cols,
4362 : Bitmapset *external_cols,
4363 : HeapTuple oldtup, HeapTuple newtup,
4364 : bool *has_external)
4365 : {
4366 : int attidx;
4367 2394796 : Bitmapset *modified = NULL;
4368 2394796 : TupleDesc tupdesc = RelationGetDescr(relation);
4369 :
4370 2394796 : attidx = -1;
4371 3384381 : while ((attidx = bms_next_member(interesting_cols, attidx)) >= 0)
4372 : {
4373 : /* attidx is zero-based, attrnum is the normal attribute number */
4374 989585 : AttrNumber attrnum = attidx + FirstLowInvalidHeapAttributeNumber;
4375 : Datum value1,
4376 : value2;
4377 : bool isnull1,
4378 : isnull2;
4379 :
4380 : /*
4381 : * If it's a whole-tuple reference, say "not equal". It's not really
4382 : * worth supporting this case, since it could only succeed after a
4383 : * no-op update, which is hardly a case worth optimizing for.
4384 : */
4385 989585 : if (attrnum == 0)
4386 : {
4387 0 : modified = bms_add_member(modified, attidx);
4388 957201 : continue;
4389 : }
4390 :
4391 : /*
4392 : * Likewise, automatically say "not equal" for any system attribute
4393 : * other than tableOID; we cannot expect these to be consistent in a
4394 : * HOT chain, or even to be set correctly yet in the new tuple.
4395 : */
4396 989585 : if (attrnum < 0)
4397 : {
4398 0 : if (attrnum != TableOidAttributeNumber)
4399 : {
4400 0 : modified = bms_add_member(modified, attidx);
4401 0 : continue;
4402 : }
4403 : }
4404 :
4405 : /*
4406 : * Extract the corresponding values. XXX this is pretty inefficient
4407 : * if there are many indexed columns. Should we do a single
4408 : * heap_deform_tuple call on each tuple, instead? But that doesn't
4409 : * work for system columns ...
4410 : */
4411 989585 : value1 = heap_getattr(oldtup, attrnum, tupdesc, &isnull1);
4412 989585 : value2 = heap_getattr(newtup, attrnum, tupdesc, &isnull2);
4413 :
4414 989585 : if (!heap_attr_equals(tupdesc, attrnum, value1,
4415 : value2, isnull1, isnull2))
4416 : {
4417 54610 : modified = bms_add_member(modified, attidx);
4418 54610 : continue;
4419 : }
4420 :
4421 : /*
4422 : * No need to check attributes that can't be stored externally. Note
4423 : * that system attributes can't be stored externally.
4424 : */
4425 934975 : if (attrnum < 0 || isnull1 ||
4426 928334 : TupleDescCompactAttr(tupdesc, attrnum - 1)->attlen != -1)
4427 902591 : continue;
4428 :
4429 : /*
4430 : * Check if the old tuple's attribute is stored externally and is a
4431 : * member of external_cols.
4432 : */
4433 32389 : if (VARATT_IS_EXTERNAL((varlena *) DatumGetPointer(value1)) &&
4434 5 : bms_is_member(attidx, external_cols))
4435 2 : *has_external = true;
4436 : }
4437 :
4438 2394796 : return modified;
4439 : }
4440 :
4441 : /*
4442 : * simple_heap_update - replace a tuple
4443 : *
4444 : * This routine may be used to update a tuple when concurrent updates of
4445 : * the target tuple are not expected (for example, because we have a lock
4446 : * on the relation associated with the tuple). Any failure is reported
4447 : * via ereport().
4448 : */
4449 : void
4450 138399 : simple_heap_update(Relation relation, const ItemPointerData *otid, HeapTuple tup,
4451 : TU_UpdateIndexes *update_indexes)
4452 : {
4453 : TM_Result result;
4454 : TM_FailureData tmfd;
4455 : LockTupleMode lockmode;
4456 :
4457 138399 : result = heap_update(relation, otid, tup,
4458 : GetCurrentCommandId(true), 0,
4459 : InvalidSnapshot,
4460 : true /* wait for commit */ ,
4461 : &tmfd, &lockmode, update_indexes);
4462 138399 : switch (result)
4463 : {
4464 0 : case TM_SelfModified:
4465 : /* Tuple was already updated in current command? */
4466 0 : elog(ERROR, "tuple already updated by self");
4467 : break;
4468 :
4469 138398 : case TM_Ok:
4470 : /* done successfully */
4471 138398 : break;
4472 :
4473 0 : case TM_Updated:
4474 0 : elog(ERROR, "tuple concurrently updated");
4475 : break;
4476 :
4477 1 : case TM_Deleted:
4478 1 : elog(ERROR, "tuple concurrently deleted");
4479 : break;
4480 :
4481 0 : default:
4482 0 : elog(ERROR, "unrecognized heap_update status: %u", result);
4483 : break;
4484 : }
4485 138398 : }
4486 :
4487 :
4488 : /*
4489 : * Return the MultiXactStatus corresponding to the given tuple lock mode.
4490 : */
4491 : static MultiXactStatus
4492 115534 : get_mxact_status_for_lock(LockTupleMode mode, bool is_update)
4493 : {
4494 : int retval;
4495 :
4496 115534 : if (is_update)
4497 216 : retval = tupleLockExtraInfo[mode].updstatus;
4498 : else
4499 115318 : retval = tupleLockExtraInfo[mode].lockstatus;
4500 :
4501 115534 : if (retval == -1)
4502 0 : elog(ERROR, "invalid lock tuple mode %d/%s", mode,
4503 : is_update ? "true" : "false");
4504 :
4505 115534 : return (MultiXactStatus) retval;
4506 : }
4507 :
4508 : /*
4509 : * heap_lock_tuple - lock a tuple in shared or exclusive mode
4510 : *
4511 : * Note that this acquires a buffer pin, which the caller must release.
4512 : *
4513 : * Input parameters:
4514 : * relation: relation containing tuple (caller must hold suitable lock)
4515 : * cid: current command ID (used for visibility test, and stored into
4516 : * tuple's cmax if lock is successful)
4517 : * mode: indicates if shared or exclusive tuple lock is desired
4518 : * wait_policy: what to do if tuple lock is not available
4519 : * follow_updates: if true, follow the update chain to also lock descendant
4520 : * tuples.
4521 : *
4522 : * Output parameters:
4523 : * *tuple: all fields filled in
4524 : * *buffer: set to buffer holding tuple (pinned but not locked at exit)
4525 : * *tmfd: filled in failure cases (see below)
4526 : *
4527 : * Function results are the same as the ones for table_tuple_lock().
4528 : *
4529 : * In the failure cases other than TM_Invisible, the routine fills
4530 : * *tmfd with the tuple's t_ctid, t_xmax (resolving a possible MultiXact,
4531 : * if necessary), and t_cmax (the last only for TM_SelfModified,
4532 : * since we cannot obtain cmax from a combo CID generated by another
4533 : * transaction).
4534 : * See comments for struct TM_FailureData for additional info.
4535 : *
4536 : * See README.tuplock for a thorough explanation of this mechanism.
4537 : */
4538 : TM_Result
4539 570496 : heap_lock_tuple(Relation relation, HeapTuple tuple,
4540 : CommandId cid, LockTupleMode mode, LockWaitPolicy wait_policy,
4541 : bool follow_updates,
4542 : Buffer *buffer, TM_FailureData *tmfd)
4543 : {
4544 : TM_Result result;
4545 570496 : ItemPointer tid = &(tuple->t_self);
4546 : ItemId lp;
4547 : Page page;
4548 570496 : Buffer vmbuffer = InvalidBuffer;
4549 : BlockNumber block;
4550 : TransactionId xid,
4551 : xmax;
4552 : uint16 old_infomask,
4553 : new_infomask,
4554 : new_infomask2;
4555 570496 : bool first_time = true;
4556 570496 : bool skip_tuple_lock = false;
4557 570496 : bool have_tuple_lock = false;
4558 570496 : bool cleared_all_frozen = false;
4559 :
4560 570496 : *buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
4561 570496 : block = ItemPointerGetBlockNumber(tid);
4562 :
4563 : /*
4564 : * Before locking the buffer, pin the visibility map page if it appears to
4565 : * be necessary. Since we haven't got the lock yet, someone else might be
4566 : * in the middle of changing this, so we'll need to recheck after we have
4567 : * the lock.
4568 : */
4569 570496 : if (PageIsAllVisible(BufferGetPage(*buffer)))
4570 413520 : visibilitymap_pin(relation, block, &vmbuffer);
4571 :
4572 570496 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4573 :
4574 570496 : page = BufferGetPage(*buffer);
4575 570496 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
4576 : Assert(ItemIdIsNormal(lp));
4577 :
4578 570496 : tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
4579 570496 : tuple->t_len = ItemIdGetLength(lp);
4580 570496 : tuple->t_tableOid = RelationGetRelid(relation);
4581 :
4582 18 : l3:
4583 570514 : result = HeapTupleSatisfiesUpdate(tuple, cid, *buffer);
4584 :
4585 570514 : if (result == TM_Invisible)
4586 : {
4587 : /*
4588 : * This is possible, but only when locking a tuple for ON CONFLICT DO
4589 : * SELECT/UPDATE. We return this value here rather than throwing an
4590 : * error in order to give that case the opportunity to throw a more
4591 : * specific error.
4592 : */
4593 28 : result = TM_Invisible;
4594 28 : goto out_locked;
4595 : }
4596 570486 : else if (result == TM_BeingModified ||
4597 80208 : result == TM_Updated ||
4598 : result == TM_Deleted)
4599 : {
4600 : TransactionId xwait;
4601 : uint16 infomask;
4602 : uint16 infomask2;
4603 : bool require_sleep;
4604 : ItemPointerData t_ctid;
4605 :
4606 : /* must copy state data before unlocking buffer */
4607 490280 : xwait = HeapTupleHeaderGetRawXmax(tuple->t_data);
4608 490280 : infomask = tuple->t_data->t_infomask;
4609 490280 : infomask2 = tuple->t_data->t_infomask2;
4610 490280 : ItemPointerCopy(&tuple->t_data->t_ctid, &t_ctid);
4611 :
4612 490280 : LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4613 :
4614 : /*
4615 : * If any subtransaction of the current top transaction already holds
4616 : * a lock as strong as or stronger than what we're requesting, we
4617 : * effectively hold the desired lock already. We *must* succeed
4618 : * without trying to take the tuple lock, else we will deadlock
4619 : * against anyone wanting to acquire a stronger lock.
4620 : *
4621 : * Note we only do this the first time we loop on the HTSU result;
4622 : * there is no point in testing in subsequent passes, because
4623 : * evidently our own transaction cannot have acquired a new lock after
4624 : * the first time we checked.
4625 : */
4626 490280 : if (first_time)
4627 : {
4628 490268 : first_time = false;
4629 :
4630 490268 : if (infomask & HEAP_XMAX_IS_MULTI)
4631 : {
4632 : int i;
4633 : int nmembers;
4634 : MultiXactMember *members;
4635 :
4636 : /*
4637 : * We don't need to allow old multixacts here; if that had
4638 : * been the case, HeapTupleSatisfiesUpdate would have returned
4639 : * MayBeUpdated and we wouldn't be here.
4640 : */
4641 : nmembers =
4642 73308 : GetMultiXactIdMembers(xwait, &members, false,
4643 73308 : HEAP_XMAX_IS_LOCKED_ONLY(infomask));
4644 :
4645 1422697 : for (i = 0; i < nmembers; i++)
4646 : {
4647 : /* only consider members of our own transaction */
4648 1349403 : if (!TransactionIdIsCurrentTransactionId(members[i].xid))
4649 1349353 : continue;
4650 :
4651 50 : if (TUPLOCK_from_mxstatus(members[i].status) >= mode)
4652 : {
4653 14 : pfree(members);
4654 14 : result = TM_Ok;
4655 14 : goto out_unlocked;
4656 : }
4657 : else
4658 : {
4659 : /*
4660 : * Disable acquisition of the heavyweight tuple lock.
4661 : * Otherwise, when promoting a weaker lock, we might
4662 : * deadlock with another locker that has acquired the
4663 : * heavyweight tuple lock and is waiting for our
4664 : * transaction to finish.
4665 : *
4666 : * Note that in this case we still need to wait for
4667 : * the multixact if required, to avoid acquiring
4668 : * conflicting locks.
4669 : */
4670 36 : skip_tuple_lock = true;
4671 : }
4672 : }
4673 :
4674 73294 : if (members)
4675 73294 : pfree(members);
4676 : }
4677 416960 : else if (TransactionIdIsCurrentTransactionId(xwait))
4678 : {
4679 415536 : switch (mode)
4680 : {
4681 409097 : case LockTupleKeyShare:
4682 : Assert(HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) ||
4683 : HEAP_XMAX_IS_SHR_LOCKED(infomask) ||
4684 : HEAP_XMAX_IS_EXCL_LOCKED(infomask));
4685 409097 : result = TM_Ok;
4686 409097 : goto out_unlocked;
4687 36 : case LockTupleShare:
4688 42 : if (HEAP_XMAX_IS_SHR_LOCKED(infomask) ||
4689 6 : HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4690 : {
4691 30 : result = TM_Ok;
4692 30 : goto out_unlocked;
4693 : }
4694 6 : break;
4695 82 : case LockTupleNoKeyExclusive:
4696 82 : if (HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4697 : {
4698 70 : result = TM_Ok;
4699 70 : goto out_unlocked;
4700 : }
4701 12 : break;
4702 6321 : case LockTupleExclusive:
4703 6321 : if (HEAP_XMAX_IS_EXCL_LOCKED(infomask) &&
4704 1279 : infomask2 & HEAP_KEYS_UPDATED)
4705 : {
4706 1251 : result = TM_Ok;
4707 1251 : goto out_unlocked;
4708 : }
4709 5070 : break;
4710 : }
4711 : }
4712 : }
4713 :
4714 : /*
4715 : * Initially assume that we will have to wait for the locking
4716 : * transaction(s) to finish. We check various cases below in which
4717 : * this can be turned off.
4718 : */
4719 79818 : require_sleep = true;
4720 79818 : if (mode == LockTupleKeyShare)
4721 : {
4722 : /*
4723 : * If we're requesting KeyShare, and there's no update present, we
4724 : * don't need to wait. Even if there is an update, we can still
4725 : * continue if the key hasn't been modified.
4726 : *
4727 : * However, if there are updates, we need to walk the update chain
4728 : * to mark future versions of the row as locked, too. That way,
4729 : * if somebody deletes that future version, we're protected
4730 : * against the key going away. This locking of future versions
4731 : * could block momentarily, if a concurrent transaction is
4732 : * deleting a key; or it could return a value to the effect that
4733 : * the transaction deleting the key has already committed. So we
4734 : * do this before re-locking the buffer; otherwise this would be
4735 : * prone to deadlocks.
4736 : *
4737 : * Note that the TID we're locking was grabbed before we unlocked
4738 : * the buffer. For it to change while we're not looking, the
4739 : * other properties we're testing for below after re-locking the
4740 : * buffer would also change, in which case we would restart this
4741 : * loop above.
4742 : */
4743 73910 : if (!(infomask2 & HEAP_KEYS_UPDATED))
4744 : {
4745 : bool updated;
4746 :
4747 73863 : updated = !HEAP_XMAX_IS_LOCKED_ONLY(infomask);
4748 :
4749 : /*
4750 : * If there are updates, follow the update chain; bail out if
4751 : * that cannot be done.
4752 : */
4753 73863 : if (follow_updates && updated &&
4754 2171 : !ItemPointerEquals(&tuple->t_self, &t_ctid))
4755 : {
4756 : TM_Result res;
4757 :
4758 2171 : res = heap_lock_updated_tuple(relation,
4759 : infomask, xwait, &t_ctid,
4760 : GetCurrentTransactionId(),
4761 : mode);
4762 2171 : if (res != TM_Ok)
4763 : {
4764 6 : result = res;
4765 : /* recovery code expects to have buffer lock held */
4766 6 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4767 207 : goto failed;
4768 : }
4769 : }
4770 :
4771 73857 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4772 :
4773 : /*
4774 : * Make sure it's still an appropriate lock, else start over.
4775 : * Also, if it wasn't updated before we released the lock, but
4776 : * is updated now, we start over too; the reason is that we
4777 : * now need to follow the update chain to lock the new
4778 : * versions.
4779 : */
4780 73857 : if (!HeapTupleHeaderIsOnlyLocked(tuple->t_data) &&
4781 2153 : ((tuple->t_data->t_infomask2 & HEAP_KEYS_UPDATED) ||
4782 2153 : !updated))
4783 18 : goto l3;
4784 :
4785 : /* Things look okay, so we can skip sleeping */
4786 73857 : require_sleep = false;
4787 :
4788 : /*
4789 : * Note we allow Xmax to change here; other updaters/lockers
4790 : * could have modified it before we grabbed the buffer lock.
4791 : * However, this is not a problem, because with the recheck we
4792 : * just did we ensure that they still don't conflict with the
4793 : * lock we want.
4794 : */
4795 : }
4796 : }
4797 5908 : else if (mode == LockTupleShare)
4798 : {
4799 : /*
4800 : * If we're requesting Share, we can similarly avoid sleeping if
4801 : * there's no update and no exclusive lock present.
4802 : */
4803 480 : if (HEAP_XMAX_IS_LOCKED_ONLY(infomask) &&
4804 480 : !HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4805 : {
4806 474 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4807 :
4808 : /*
4809 : * Make sure it's still an appropriate lock, else start over.
4810 : * See above about allowing xmax to change.
4811 : */
4812 948 : if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4813 474 : HEAP_XMAX_IS_EXCL_LOCKED(tuple->t_data->t_infomask))
4814 0 : goto l3;
4815 474 : require_sleep = false;
4816 : }
4817 : }
4818 5428 : else if (mode == LockTupleNoKeyExclusive)
4819 : {
4820 : /*
4821 : * If we're requesting NoKeyExclusive, we might also be able to
4822 : * avoid sleeping; just ensure that there no conflicting lock
4823 : * already acquired.
4824 : */
4825 173 : if (infomask & HEAP_XMAX_IS_MULTI)
4826 : {
4827 26 : if (!DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
4828 : mode, NULL))
4829 : {
4830 : /*
4831 : * No conflict, but if the xmax changed under us in the
4832 : * meantime, start over.
4833 : */
4834 13 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4835 26 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4836 13 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4837 : xwait))
4838 0 : goto l3;
4839 :
4840 : /* otherwise, we're good */
4841 13 : require_sleep = false;
4842 : }
4843 : }
4844 147 : else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask))
4845 : {
4846 18 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4847 :
4848 : /* if the xmax changed in the meantime, start over */
4849 36 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4850 18 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4851 : xwait))
4852 0 : goto l3;
4853 : /* otherwise, we're good */
4854 18 : require_sleep = false;
4855 : }
4856 : }
4857 :
4858 : /*
4859 : * As a check independent from those above, we can also avoid sleeping
4860 : * if the current transaction is the sole locker of the tuple. Note
4861 : * that the strength of the lock already held is irrelevant; this is
4862 : * not about recording the lock in Xmax (which will be done regardless
4863 : * of this optimization, below). Also, note that the cases where we
4864 : * hold a lock stronger than we are requesting are already handled
4865 : * above by not doing anything.
4866 : *
4867 : * Note we only deal with the non-multixact case here; MultiXactIdWait
4868 : * is well equipped to deal with this situation on its own.
4869 : */
4870 85218 : if (require_sleep && !(infomask & HEAP_XMAX_IS_MULTI) &&
4871 5406 : TransactionIdIsCurrentTransactionId(xwait))
4872 : {
4873 : /* ... but if the xmax changed in the meantime, start over */
4874 5070 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4875 10140 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4876 5070 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4877 : xwait))
4878 0 : goto l3;
4879 : Assert(HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask));
4880 5070 : require_sleep = false;
4881 : }
4882 :
4883 : /*
4884 : * Time to sleep on the other transaction/multixact, if necessary.
4885 : *
4886 : * If the other transaction is an update/delete that's already
4887 : * committed, then sleeping cannot possibly do any good: if we're
4888 : * required to sleep, get out to raise an error instead.
4889 : *
4890 : * By here, we either have already acquired the buffer exclusive lock,
4891 : * or we must wait for the locking transaction or multixact; so below
4892 : * we ensure that we grab buffer lock after the sleep.
4893 : */
4894 79812 : if (require_sleep && (result == TM_Updated || result == TM_Deleted))
4895 : {
4896 162 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4897 162 : goto failed;
4898 : }
4899 79650 : else if (require_sleep)
4900 : {
4901 : /*
4902 : * Acquire tuple lock to establish our priority for the tuple, or
4903 : * die trying. LockTuple will release us when we are next-in-line
4904 : * for the tuple. We must do this even if we are share-locking,
4905 : * but not if we already have a weaker lock on the tuple.
4906 : *
4907 : * If we are forced to "start over" below, we keep the tuple lock;
4908 : * this arranges that we stay at the head of the line while
4909 : * rechecking tuple state.
4910 : */
4911 218 : if (!skip_tuple_lock &&
4912 201 : !heap_acquire_tuplock(relation, tid, mode, wait_policy,
4913 : &have_tuple_lock))
4914 : {
4915 : /*
4916 : * This can only happen if wait_policy is Skip and the lock
4917 : * couldn't be obtained.
4918 : */
4919 1 : result = TM_WouldBlock;
4920 : /* recovery code expects to have buffer lock held */
4921 1 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4922 1 : goto failed;
4923 : }
4924 :
4925 216 : if (infomask & HEAP_XMAX_IS_MULTI)
4926 : {
4927 43 : MultiXactStatus status = get_mxact_status_for_lock(mode, false);
4928 :
4929 : /* We only ever lock tuples, never update them */
4930 43 : if (status >= MultiXactStatusNoKeyUpdate)
4931 0 : elog(ERROR, "invalid lock mode in heap_lock_tuple");
4932 :
4933 : /* wait for multixact to end, or die trying */
4934 43 : switch (wait_policy)
4935 : {
4936 37 : case LockWaitBlock:
4937 37 : MultiXactIdWait((MultiXactId) xwait, status, infomask,
4938 37 : relation, &tuple->t_self, XLTW_Lock, NULL);
4939 37 : break;
4940 2 : case LockWaitSkip:
4941 2 : if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4942 : status, infomask, relation,
4943 : NULL, false))
4944 : {
4945 2 : result = TM_WouldBlock;
4946 : /* recovery code expects to have buffer lock held */
4947 2 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4948 2 : goto failed;
4949 : }
4950 0 : break;
4951 4 : case LockWaitError:
4952 4 : if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4953 : status, infomask, relation,
4954 : NULL, log_lock_failures))
4955 4 : ereport(ERROR,
4956 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4957 : errmsg("could not obtain lock on row in relation \"%s\"",
4958 : RelationGetRelationName(relation))));
4959 :
4960 0 : break;
4961 : }
4962 :
4963 : /*
4964 : * Of course, the multixact might not be done here: if we're
4965 : * requesting a light lock mode, other transactions with light
4966 : * locks could still be alive, as well as locks owned by our
4967 : * own xact or other subxacts of this backend. We need to
4968 : * preserve the surviving MultiXact members. Note that it
4969 : * isn't absolutely necessary in the latter case, but doing so
4970 : * is simpler.
4971 : */
4972 : }
4973 : else
4974 : {
4975 : /* wait for regular transaction to end, or die trying */
4976 173 : switch (wait_policy)
4977 : {
4978 132 : case LockWaitBlock:
4979 132 : XactLockTableWait(xwait, relation, &tuple->t_self,
4980 : XLTW_Lock);
4981 132 : break;
4982 33 : case LockWaitSkip:
4983 33 : if (!ConditionalXactLockTableWait(xwait, false))
4984 : {
4985 33 : result = TM_WouldBlock;
4986 : /* recovery code expects to have buffer lock held */
4987 33 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4988 33 : goto failed;
4989 : }
4990 0 : break;
4991 8 : case LockWaitError:
4992 8 : if (!ConditionalXactLockTableWait(xwait, log_lock_failures))
4993 8 : ereport(ERROR,
4994 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4995 : errmsg("could not obtain lock on row in relation \"%s\"",
4996 : RelationGetRelationName(relation))));
4997 0 : break;
4998 : }
4999 : }
5000 :
5001 : /* if there are updates, follow the update chain */
5002 169 : if (follow_updates && !HEAP_XMAX_IS_LOCKED_ONLY(infomask) &&
5003 75 : !ItemPointerEquals(&tuple->t_self, &t_ctid))
5004 : {
5005 : TM_Result res;
5006 :
5007 55 : res = heap_lock_updated_tuple(relation,
5008 : infomask, xwait, &t_ctid,
5009 : GetCurrentTransactionId(),
5010 : mode);
5011 55 : if (res != TM_Ok)
5012 : {
5013 3 : result = res;
5014 : /* recovery code expects to have buffer lock held */
5015 3 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5016 3 : goto failed;
5017 : }
5018 : }
5019 :
5020 166 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5021 :
5022 : /*
5023 : * xwait is done, but if xwait had just locked the tuple then some
5024 : * other xact could update this tuple before we get to this point.
5025 : * Check for xmax change, and start over if so.
5026 : */
5027 316 : if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
5028 150 : !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
5029 : xwait))
5030 18 : goto l3;
5031 :
5032 148 : if (!(infomask & HEAP_XMAX_IS_MULTI))
5033 : {
5034 : /*
5035 : * Otherwise check if it committed or aborted. Note we cannot
5036 : * be here if the tuple was only locked by somebody who didn't
5037 : * conflict with us; that would have been handled above. So
5038 : * that transaction must necessarily be gone by now. But
5039 : * don't check for this in the multixact case, because some
5040 : * locker transactions might still be running.
5041 : */
5042 116 : UpdateXmaxHintBits(tuple->t_data, *buffer, xwait);
5043 : }
5044 : }
5045 :
5046 : /* By here, we're certain that we hold buffer exclusive lock again */
5047 :
5048 : /*
5049 : * We may lock if previous xmax aborted, or if it committed but only
5050 : * locked the tuple without updating it; or if we didn't have to wait
5051 : * at all for whatever reason.
5052 : */
5053 79580 : if (!require_sleep ||
5054 260 : (tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
5055 209 : HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
5056 97 : HeapTupleHeaderIsOnlyLocked(tuple->t_data))
5057 79490 : result = TM_Ok;
5058 90 : else if (!ItemPointerEquals(&tuple->t_self, &tuple->t_data->t_ctid))
5059 65 : result = TM_Updated;
5060 : else
5061 25 : result = TM_Deleted;
5062 : }
5063 :
5064 80206 : failed:
5065 159993 : if (result != TM_Ok)
5066 : {
5067 : Assert(result == TM_SelfModified || result == TM_Updated ||
5068 : result == TM_Deleted || result == TM_WouldBlock);
5069 :
5070 : /*
5071 : * When locking a tuple under LockWaitSkip semantics and we fail with
5072 : * TM_WouldBlock above, it's possible for concurrent transactions to
5073 : * release the lock and set HEAP_XMAX_INVALID in the meantime. So
5074 : * this assert is slightly different from the equivalent one in
5075 : * heap_delete and heap_update.
5076 : */
5077 : Assert((result == TM_WouldBlock) ||
5078 : !(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));
5079 : Assert(result != TM_Updated ||
5080 : !ItemPointerEquals(&tuple->t_self, &tuple->t_data->t_ctid));
5081 305 : tmfd->ctid = tuple->t_data->t_ctid;
5082 305 : tmfd->xmax = HeapTupleHeaderGetUpdateXid(tuple->t_data);
5083 305 : if (result == TM_SelfModified)
5084 8 : tmfd->cmax = HeapTupleHeaderGetCmax(tuple->t_data);
5085 : else
5086 297 : tmfd->cmax = InvalidCommandId;
5087 305 : goto out_locked;
5088 : }
5089 :
5090 : /*
5091 : * If we didn't pin the visibility map page and the page has become all
5092 : * visible while we were busy locking the buffer, or during some
5093 : * subsequent window during which we had it unlocked, we'll have to unlock
5094 : * and re-lock, to avoid holding the buffer lock across I/O. That's a bit
5095 : * unfortunate, especially since we'll now have to recheck whether the
5096 : * tuple has been locked or updated under us, but hopefully it won't
5097 : * happen very often.
5098 : */
5099 159688 : if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
5100 : {
5101 0 : LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
5102 0 : visibilitymap_pin(relation, block, &vmbuffer);
5103 0 : LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
5104 0 : goto l3;
5105 : }
5106 :
5107 159688 : xmax = HeapTupleHeaderGetRawXmax(tuple->t_data);
5108 159688 : old_infomask = tuple->t_data->t_infomask;
5109 :
5110 : /*
5111 : * If this is the first possibly-multixact-able operation in the current
5112 : * transaction, set my per-backend OldestMemberMXactId setting. We can be
5113 : * certain that the transaction will never become a member of any older
5114 : * MultiXactIds than that. (We have to do this even if we end up just
5115 : * using our own TransactionId below, since some other backend could
5116 : * incorporate our XID into a MultiXact immediately afterwards.)
5117 : */
5118 159688 : MultiXactIdSetOldestMember();
5119 :
5120 : /*
5121 : * Compute the new xmax and infomask to store into the tuple. Note we do
5122 : * not modify the tuple just yet, because that would leave it in the wrong
5123 : * state if multixact.c elogs.
5124 : */
5125 159688 : compute_new_xmax_infomask(xmax, old_infomask, tuple->t_data->t_infomask2,
5126 : GetCurrentTransactionId(), mode, false,
5127 : &xid, &new_infomask, &new_infomask2);
5128 :
5129 159688 : START_CRIT_SECTION();
5130 :
5131 : /*
5132 : * Store transaction information of xact locking the tuple.
5133 : *
5134 : * Note: Cmax is meaningless in this context, so don't set it; this avoids
5135 : * possibly generating a useless combo CID. Moreover, if we're locking a
5136 : * previously updated tuple, it's important to preserve the Cmax.
5137 : *
5138 : * Also reset the HOT UPDATE bit, but only if there's no update; otherwise
5139 : * we would break the HOT chain.
5140 : */
5141 159688 : tuple->t_data->t_infomask &= ~HEAP_XMAX_BITS;
5142 159688 : tuple->t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5143 159688 : tuple->t_data->t_infomask |= new_infomask;
5144 159688 : tuple->t_data->t_infomask2 |= new_infomask2;
5145 159688 : if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
5146 157539 : HeapTupleHeaderClearHotUpdated(tuple->t_data);
5147 159688 : HeapTupleHeaderSetXmax(tuple->t_data, xid);
5148 :
5149 : /*
5150 : * Make sure there is no forward chain link in t_ctid. Note that in the
5151 : * cases where the tuple has been updated, we must not overwrite t_ctid,
5152 : * because it was set by the updater. Moreover, if the tuple has been
5153 : * updated, we need to follow the update chain to lock the new versions of
5154 : * the tuple as well.
5155 : */
5156 159688 : if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
5157 157539 : tuple->t_data->t_ctid = *tid;
5158 :
5159 : /* Clear only the all-frozen bit on visibility map if needed */
5160 163311 : if (PageIsAllVisible(page) &&
5161 3623 : visibilitymap_clear(relation, block, vmbuffer,
5162 : VISIBILITYMAP_ALL_FROZEN))
5163 15 : cleared_all_frozen = true;
5164 :
5165 :
5166 159688 : MarkBufferDirty(*buffer);
5167 :
5168 : /*
5169 : * XLOG stuff. You might think that we don't need an XLOG record because
5170 : * there is no state change worth restoring after a crash. You would be
5171 : * wrong however: we have just written either a TransactionId or a
5172 : * MultiXactId that may never have been seen on disk before, and we need
5173 : * to make sure that there are XLOG entries covering those ID numbers.
5174 : * Else the same IDs might be re-used after a crash, which would be
5175 : * disastrous if this page made it to disk before the crash. Essentially
5176 : * we have to enforce the WAL log-before-data rule even in this case.
5177 : * (Also, in a PITR log-shipping or 2PC environment, we have to have XLOG
5178 : * entries for everything anyway.)
5179 : */
5180 159688 : if (RelationNeedsWAL(relation))
5181 : {
5182 : xl_heap_lock xlrec;
5183 : XLogRecPtr recptr;
5184 :
5185 159259 : XLogBeginInsert();
5186 159259 : XLogRegisterBuffer(0, *buffer, REGBUF_STANDARD);
5187 :
5188 159259 : xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
5189 159259 : xlrec.xmax = xid;
5190 318518 : xlrec.infobits_set = compute_infobits(new_infomask,
5191 159259 : tuple->t_data->t_infomask2);
5192 159259 : xlrec.flags = cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
5193 159259 : XLogRegisterData(&xlrec, SizeOfHeapLock);
5194 :
5195 : /* we don't decode row locks atm, so no need to log the origin */
5196 :
5197 159259 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK);
5198 :
5199 159259 : PageSetLSN(page, recptr);
5200 : }
5201 :
5202 159688 : END_CRIT_SECTION();
5203 :
5204 159688 : result = TM_Ok;
5205 :
5206 160021 : out_locked:
5207 160021 : LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
5208 :
5209 570483 : out_unlocked:
5210 570483 : if (BufferIsValid(vmbuffer))
5211 413520 : ReleaseBuffer(vmbuffer);
5212 :
5213 : /*
5214 : * Don't update the visibility map here. Locking a tuple doesn't change
5215 : * visibility info.
5216 : */
5217 :
5218 : /*
5219 : * Now that we have successfully marked the tuple as locked, we can
5220 : * release the lmgr tuple lock, if we had it.
5221 : */
5222 570483 : if (have_tuple_lock)
5223 182 : UnlockTupleTuplock(relation, tid, mode);
5224 :
5225 570483 : return result;
5226 : }
5227 :
5228 : /*
5229 : * Acquire heavyweight lock on the given tuple, in preparation for acquiring
5230 : * its normal, Xmax-based tuple lock.
5231 : *
5232 : * have_tuple_lock is an input and output parameter: on input, it indicates
5233 : * whether the lock has previously been acquired (and this function does
5234 : * nothing in that case). If this function returns success, have_tuple_lock
5235 : * has been flipped to true.
5236 : *
5237 : * Returns false if it was unable to obtain the lock; this can only happen if
5238 : * wait_policy is Skip.
5239 : */
5240 : static bool
5241 376 : heap_acquire_tuplock(Relation relation, const ItemPointerData *tid, LockTupleMode mode,
5242 : LockWaitPolicy wait_policy, bool *have_tuple_lock)
5243 : {
5244 376 : if (*have_tuple_lock)
5245 9 : return true;
5246 :
5247 367 : switch (wait_policy)
5248 : {
5249 322 : case LockWaitBlock:
5250 322 : LockTupleTuplock(relation, tid, mode);
5251 322 : break;
5252 :
5253 34 : case LockWaitSkip:
5254 34 : if (!ConditionalLockTupleTuplock(relation, tid, mode, false))
5255 1 : return false;
5256 33 : break;
5257 :
5258 11 : case LockWaitError:
5259 11 : if (!ConditionalLockTupleTuplock(relation, tid, mode, log_lock_failures))
5260 1 : ereport(ERROR,
5261 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
5262 : errmsg("could not obtain lock on row in relation \"%s\"",
5263 : RelationGetRelationName(relation))));
5264 10 : break;
5265 : }
5266 365 : *have_tuple_lock = true;
5267 :
5268 365 : return true;
5269 : }
5270 :
5271 : /*
5272 : * Given an original set of Xmax and infomask, and a transaction (identified by
5273 : * add_to_xmax) acquiring a new lock of some mode, compute the new Xmax and
5274 : * corresponding infomasks to use on the tuple.
5275 : *
5276 : * Note that this might have side effects such as creating a new MultiXactId.
5277 : *
5278 : * Most callers will have called HeapTupleSatisfiesUpdate before this function;
5279 : * that will have set the HEAP_XMAX_INVALID bit if the xmax was a MultiXactId
5280 : * but it was not running anymore. There is a race condition, which is that the
5281 : * MultiXactId may have finished since then, but that uncommon case is handled
5282 : * either here, or within MultiXactIdExpand.
5283 : *
5284 : * There is a similar race condition possible when the old xmax was a regular
5285 : * TransactionId. We test TransactionIdIsInProgress again just to narrow the
5286 : * window, but it's still possible to end up creating an unnecessary
5287 : * MultiXactId. Fortunately this is harmless.
5288 : */
5289 : static void
5290 6659304 : compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
5291 : uint16 old_infomask2, TransactionId add_to_xmax,
5292 : LockTupleMode mode, bool is_update,
5293 : TransactionId *result_xmax, uint16 *result_infomask,
5294 : uint16 *result_infomask2)
5295 : {
5296 : TransactionId new_xmax;
5297 : uint16 new_infomask,
5298 : new_infomask2;
5299 :
5300 : Assert(TransactionIdIsCurrentTransactionId(add_to_xmax));
5301 :
5302 104501 : l5:
5303 6763805 : new_infomask = 0;
5304 6763805 : new_infomask2 = 0;
5305 6763805 : if (old_infomask & HEAP_XMAX_INVALID)
5306 : {
5307 : /*
5308 : * No previous locker; we just insert our own TransactionId.
5309 : *
5310 : * Note that it's critical that this case be the first one checked,
5311 : * because there are several blocks below that come back to this one
5312 : * to implement certain optimizations; old_infomask might contain
5313 : * other dirty bits in those cases, but we don't really care.
5314 : */
5315 6582594 : if (is_update)
5316 : {
5317 4289766 : new_xmax = add_to_xmax;
5318 4289766 : if (mode == LockTupleExclusive)
5319 1933363 : new_infomask2 |= HEAP_KEYS_UPDATED;
5320 : }
5321 : else
5322 : {
5323 2292828 : new_infomask |= HEAP_XMAX_LOCK_ONLY;
5324 2292828 : switch (mode)
5325 : {
5326 4978 : case LockTupleKeyShare:
5327 4978 : new_xmax = add_to_xmax;
5328 4978 : new_infomask |= HEAP_XMAX_KEYSHR_LOCK;
5329 4978 : break;
5330 809 : case LockTupleShare:
5331 809 : new_xmax = add_to_xmax;
5332 809 : new_infomask |= HEAP_XMAX_SHR_LOCK;
5333 809 : break;
5334 2191084 : case LockTupleNoKeyExclusive:
5335 2191084 : new_xmax = add_to_xmax;
5336 2191084 : new_infomask |= HEAP_XMAX_EXCL_LOCK;
5337 2191084 : break;
5338 95957 : case LockTupleExclusive:
5339 95957 : new_xmax = add_to_xmax;
5340 95957 : new_infomask |= HEAP_XMAX_EXCL_LOCK;
5341 95957 : new_infomask2 |= HEAP_KEYS_UPDATED;
5342 95957 : break;
5343 0 : default:
5344 0 : new_xmax = InvalidTransactionId; /* silence compiler */
5345 0 : elog(ERROR, "invalid lock mode");
5346 : }
5347 : }
5348 : }
5349 181211 : else if (old_infomask & HEAP_XMAX_IS_MULTI)
5350 : {
5351 : MultiXactStatus new_status;
5352 :
5353 : /*
5354 : * Currently we don't allow XMAX_COMMITTED to be set for multis, so
5355 : * cross-check.
5356 : */
5357 : Assert(!(old_infomask & HEAP_XMAX_COMMITTED));
5358 :
5359 : /*
5360 : * A multixact together with LOCK_ONLY set but neither lock bit set
5361 : * (i.e. a pg_upgraded share locked tuple) cannot possibly be running
5362 : * anymore. This check is critical for databases upgraded by
5363 : * pg_upgrade; both MultiXactIdIsRunning and MultiXactIdExpand assume
5364 : * that such multis are never passed.
5365 : */
5366 75569 : if (HEAP_LOCKED_UPGRADED(old_infomask))
5367 : {
5368 0 : old_infomask &= ~HEAP_XMAX_IS_MULTI;
5369 0 : old_infomask |= HEAP_XMAX_INVALID;
5370 0 : goto l5;
5371 : }
5372 :
5373 : /*
5374 : * If the XMAX is already a MultiXactId, then we need to expand it to
5375 : * include add_to_xmax; but if all the members were lockers and are
5376 : * all gone, we can do away with the IS_MULTI bit and just set
5377 : * add_to_xmax as the only locker/updater. If all lockers are gone
5378 : * and we have an updater that aborted, we can also do without a
5379 : * multi.
5380 : *
5381 : * The cost of doing GetMultiXactIdMembers would be paid by
5382 : * MultiXactIdExpand if we weren't to do this, so this check is not
5383 : * incurring extra work anyhow.
5384 : */
5385 75569 : if (!MultiXactIdIsRunning(xmax, HEAP_XMAX_IS_LOCKED_ONLY(old_infomask)))
5386 : {
5387 23 : if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) ||
5388 8 : !TransactionIdDidCommit(MultiXactIdGetUpdateXid(xmax,
5389 : old_infomask)))
5390 : {
5391 : /*
5392 : * Reset these bits and restart; otherwise fall through to
5393 : * create a new multi below.
5394 : */
5395 23 : old_infomask &= ~HEAP_XMAX_IS_MULTI;
5396 23 : old_infomask |= HEAP_XMAX_INVALID;
5397 23 : goto l5;
5398 : }
5399 : }
5400 :
5401 75546 : new_status = get_mxact_status_for_lock(mode, is_update);
5402 :
5403 75546 : new_xmax = MultiXactIdExpand((MultiXactId) xmax, add_to_xmax,
5404 : new_status);
5405 75546 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5406 : }
5407 105642 : else if (old_infomask & HEAP_XMAX_COMMITTED)
5408 : {
5409 : /*
5410 : * It's a committed update, so we need to preserve him as updater of
5411 : * the tuple.
5412 : */
5413 : MultiXactStatus status;
5414 : MultiXactStatus new_status;
5415 :
5416 13 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5417 0 : status = MultiXactStatusUpdate;
5418 : else
5419 13 : status = MultiXactStatusNoKeyUpdate;
5420 :
5421 13 : new_status = get_mxact_status_for_lock(mode, is_update);
5422 :
5423 : /*
5424 : * since it's not running, it's obviously impossible for the old
5425 : * updater to be identical to the current one, so we need not check
5426 : * for that case as we do in the block above.
5427 : */
5428 13 : new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
5429 13 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5430 : }
5431 105629 : else if (TransactionIdIsInProgress(xmax))
5432 : {
5433 : /*
5434 : * If the XMAX is a valid, in-progress TransactionId, then we need to
5435 : * create a new MultiXactId that includes both the old locker or
5436 : * updater and our own TransactionId.
5437 : */
5438 : MultiXactStatus new_status;
5439 : MultiXactStatus old_status;
5440 : LockTupleMode old_mode;
5441 :
5442 105614 : if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5443 : {
5444 105586 : if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5445 5725 : old_status = MultiXactStatusForKeyShare;
5446 99861 : else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5447 471 : old_status = MultiXactStatusForShare;
5448 99390 : else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5449 : {
5450 99390 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5451 92885 : old_status = MultiXactStatusForUpdate;
5452 : else
5453 6505 : old_status = MultiXactStatusForNoKeyUpdate;
5454 : }
5455 : else
5456 : {
5457 : /*
5458 : * LOCK_ONLY can be present alone only when a page has been
5459 : * upgraded by pg_upgrade. But in that case,
5460 : * TransactionIdIsInProgress() should have returned false. We
5461 : * assume it's no longer locked in this case.
5462 : */
5463 0 : elog(WARNING, "LOCK_ONLY found for Xid in progress %u", xmax);
5464 0 : old_infomask |= HEAP_XMAX_INVALID;
5465 0 : old_infomask &= ~HEAP_XMAX_LOCK_ONLY;
5466 0 : goto l5;
5467 : }
5468 : }
5469 : else
5470 : {
5471 : /* it's an update, but which kind? */
5472 28 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5473 0 : old_status = MultiXactStatusUpdate;
5474 : else
5475 28 : old_status = MultiXactStatusNoKeyUpdate;
5476 : }
5477 :
5478 105614 : old_mode = TUPLOCK_from_mxstatus(old_status);
5479 :
5480 : /*
5481 : * If the lock to be acquired is for the same TransactionId as the
5482 : * existing lock, there's an optimization possible: consider only the
5483 : * strongest of both locks as the only one present, and restart.
5484 : */
5485 105614 : if (xmax == add_to_xmax)
5486 : {
5487 : /*
5488 : * Note that it's not possible for the original tuple to be
5489 : * updated: we wouldn't be here because the tuple would have been
5490 : * invisible and we wouldn't try to update it. As a subtlety,
5491 : * this code can also run when traversing an update chain to lock
5492 : * future versions of a tuple. But we wouldn't be here either,
5493 : * because the add_to_xmax would be different from the original
5494 : * updater.
5495 : */
5496 : Assert(HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5497 :
5498 : /* acquire the strongest of both */
5499 104464 : if (mode < old_mode)
5500 52191 : mode = old_mode;
5501 : /* mustn't touch is_update */
5502 :
5503 104464 : old_infomask |= HEAP_XMAX_INVALID;
5504 104464 : goto l5;
5505 : }
5506 :
5507 : /* otherwise, just fall back to creating a new multixact */
5508 1150 : new_status = get_mxact_status_for_lock(mode, is_update);
5509 1150 : new_xmax = MultiXactIdCreate(xmax, old_status,
5510 : add_to_xmax, new_status);
5511 1150 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5512 : }
5513 20 : else if (!HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) &&
5514 5 : TransactionIdDidCommit(xmax))
5515 1 : {
5516 : /*
5517 : * It's a committed update, so we gotta preserve him as updater of the
5518 : * tuple.
5519 : */
5520 : MultiXactStatus status;
5521 : MultiXactStatus new_status;
5522 :
5523 1 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5524 0 : status = MultiXactStatusUpdate;
5525 : else
5526 1 : status = MultiXactStatusNoKeyUpdate;
5527 :
5528 1 : new_status = get_mxact_status_for_lock(mode, is_update);
5529 :
5530 : /*
5531 : * since it's not running, it's obviously impossible for the old
5532 : * updater to be identical to the current one, so we need not check
5533 : * for that case as we do in the block above.
5534 : */
5535 1 : new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
5536 1 : GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5537 : }
5538 : else
5539 : {
5540 : /*
5541 : * Can get here iff the locking/updating transaction was running when
5542 : * the infomask was extracted from the tuple, but finished before
5543 : * TransactionIdIsInProgress got to run. Deal with it as if there was
5544 : * no locker at all in the first place.
5545 : */
5546 14 : old_infomask |= HEAP_XMAX_INVALID;
5547 14 : goto l5;
5548 : }
5549 :
5550 6659304 : *result_infomask = new_infomask;
5551 6659304 : *result_infomask2 = new_infomask2;
5552 6659304 : *result_xmax = new_xmax;
5553 6659304 : }
5554 :
5555 : /*
5556 : * Subroutine for heap_lock_updated_tuple_rec.
5557 : *
5558 : * Given a hypothetical multixact status held by the transaction identified
5559 : * with the given xid, does the current transaction need to wait, fail, or can
5560 : * it continue if it wanted to acquire a lock of the given mode? "needwait"
5561 : * is set to true if waiting is necessary; if it can continue, then TM_Ok is
5562 : * returned. If the lock is already held by the current transaction, return
5563 : * TM_SelfModified. In case of a conflict with another transaction, a
5564 : * different HeapTupleSatisfiesUpdate return code is returned.
5565 : *
5566 : * The held status is said to be hypothetical because it might correspond to a
5567 : * lock held by a single Xid, i.e. not a real MultiXactId; we express it this
5568 : * way for simplicity of API.
5569 : */
5570 : static TM_Result
5571 38781 : test_lockmode_for_conflict(MultiXactStatus status, TransactionId xid,
5572 : LockTupleMode mode, HeapTuple tup,
5573 : bool *needwait)
5574 : {
5575 : MultiXactStatus wantedstatus;
5576 :
5577 38781 : *needwait = false;
5578 38781 : wantedstatus = get_mxact_status_for_lock(mode, false);
5579 :
5580 : /*
5581 : * Note: we *must* check TransactionIdIsInProgress before
5582 : * TransactionIdDidAbort/Commit; see comment at top of heapam_visibility.c
5583 : * for an explanation.
5584 : */
5585 38781 : if (TransactionIdIsCurrentTransactionId(xid))
5586 : {
5587 : /*
5588 : * The tuple has already been locked by our own transaction. This is
5589 : * very rare but can happen if multiple transactions are trying to
5590 : * lock an ancient version of the same tuple.
5591 : */
5592 0 : return TM_SelfModified;
5593 : }
5594 38781 : else if (TransactionIdIsInProgress(xid))
5595 : {
5596 : /*
5597 : * If the locking transaction is running, what we do depends on
5598 : * whether the lock modes conflict: if they do, then we must wait for
5599 : * it to finish; otherwise we can fall through to lock this tuple
5600 : * version without waiting.
5601 : */
5602 36539 : if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5603 36539 : LOCKMODE_from_mxstatus(wantedstatus)))
5604 : {
5605 8 : *needwait = true;
5606 : }
5607 :
5608 : /*
5609 : * If we set needwait above, then this value doesn't matter;
5610 : * otherwise, this value signals to caller that it's okay to proceed.
5611 : */
5612 36539 : return TM_Ok;
5613 : }
5614 2242 : else if (TransactionIdDidAbort(xid))
5615 206 : return TM_Ok;
5616 2036 : else if (TransactionIdDidCommit(xid))
5617 : {
5618 : /*
5619 : * The other transaction committed. If it was only a locker, then the
5620 : * lock is completely gone now and we can return success; but if it
5621 : * was an update, then what we do depends on whether the two lock
5622 : * modes conflict. If they conflict, then we must report error to
5623 : * caller. But if they don't, we can fall through to allow the current
5624 : * transaction to lock the tuple.
5625 : *
5626 : * Note: the reason we worry about ISUPDATE here is because as soon as
5627 : * a transaction ends, all its locks are gone and meaningless, and
5628 : * thus we can ignore them; whereas its updates persist. In the
5629 : * TransactionIdIsInProgress case, above, we don't need to check
5630 : * because we know the lock is still "alive" and thus a conflict needs
5631 : * always be checked.
5632 : */
5633 2036 : if (!ISUPDATE_from_mxstatus(status))
5634 2026 : return TM_Ok;
5635 :
5636 10 : if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5637 10 : LOCKMODE_from_mxstatus(wantedstatus)))
5638 : {
5639 : /* bummer */
5640 9 : if (!ItemPointerEquals(&tup->t_self, &tup->t_data->t_ctid))
5641 7 : return TM_Updated;
5642 : else
5643 2 : return TM_Deleted;
5644 : }
5645 :
5646 1 : return TM_Ok;
5647 : }
5648 :
5649 : /* Not in progress, not aborted, not committed -- must have crashed */
5650 0 : return TM_Ok;
5651 : }
5652 :
5653 :
5654 : /*
5655 : * Recursive part of heap_lock_updated_tuple
5656 : *
5657 : * Fetch the tuple pointed to by tid in rel, and mark it as locked by the given
5658 : * xid with the given mode; if this tuple is updated, recurse to lock the new
5659 : * version as well.
5660 : */
5661 : static TM_Result
5662 2224 : heap_lock_updated_tuple_rec(Relation rel, TransactionId priorXmax,
5663 : const ItemPointerData *tid, TransactionId xid,
5664 : LockTupleMode mode)
5665 : {
5666 : TM_Result result;
5667 : ItemPointerData tupid;
5668 : HeapTupleData mytup;
5669 : Buffer buf;
5670 : uint16 new_infomask,
5671 : new_infomask2,
5672 : old_infomask,
5673 : old_infomask2;
5674 : TransactionId xmax,
5675 : new_xmax;
5676 2224 : bool cleared_all_frozen = false;
5677 : bool pinned_desired_page;
5678 2224 : Buffer vmbuffer = InvalidBuffer;
5679 : BlockNumber block;
5680 :
5681 2224 : ItemPointerCopy(tid, &tupid);
5682 :
5683 : for (;;)
5684 : {
5685 2227 : new_infomask = 0;
5686 2227 : new_xmax = InvalidTransactionId;
5687 2227 : block = ItemPointerGetBlockNumber(&tupid);
5688 2227 : ItemPointerCopy(&tupid, &(mytup.t_self));
5689 :
5690 2227 : if (!heap_fetch(rel, SnapshotAny, &mytup, &buf, false))
5691 : {
5692 : /*
5693 : * if we fail to find the updated version of the tuple, it's
5694 : * because it was vacuumed/pruned away after its creator
5695 : * transaction aborted. So behave as if we got to the end of the
5696 : * chain, and there's no further tuple to lock: return success to
5697 : * caller.
5698 : */
5699 0 : result = TM_Ok;
5700 0 : goto out_unlocked;
5701 : }
5702 :
5703 2227 : l4:
5704 2235 : CHECK_FOR_INTERRUPTS();
5705 :
5706 : /*
5707 : * Before locking the buffer, pin the visibility map page if it
5708 : * appears to be necessary. Since we haven't got the lock yet,
5709 : * someone else might be in the middle of changing this, so we'll need
5710 : * to recheck after we have the lock.
5711 : */
5712 2235 : if (PageIsAllVisible(BufferGetPage(buf)))
5713 : {
5714 0 : visibilitymap_pin(rel, block, &vmbuffer);
5715 0 : pinned_desired_page = true;
5716 : }
5717 : else
5718 2235 : pinned_desired_page = false;
5719 :
5720 2235 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5721 :
5722 : /*
5723 : * If we didn't pin the visibility map page and the page has become
5724 : * all visible while we were busy locking the buffer, we'll have to
5725 : * unlock and re-lock, to avoid holding the buffer lock across I/O.
5726 : * That's a bit unfortunate, but hopefully shouldn't happen often.
5727 : *
5728 : * Note: in some paths through this function, we will reach here
5729 : * holding a pin on a vm page that may or may not be the one matching
5730 : * this page. If this page isn't all-visible, we won't use the vm
5731 : * page, but we hold onto such a pin till the end of the function.
5732 : */
5733 2235 : if (!pinned_desired_page && PageIsAllVisible(BufferGetPage(buf)))
5734 : {
5735 0 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5736 0 : visibilitymap_pin(rel, block, &vmbuffer);
5737 0 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5738 : }
5739 :
5740 : /*
5741 : * Check the tuple XMIN against prior XMAX, if any. If we reached the
5742 : * end of the chain, we're done, so return success.
5743 : */
5744 4470 : if (TransactionIdIsValid(priorXmax) &&
5745 2235 : !TransactionIdEquals(HeapTupleHeaderGetXmin(mytup.t_data),
5746 : priorXmax))
5747 : {
5748 2 : result = TM_Ok;
5749 2 : goto out_locked;
5750 : }
5751 :
5752 : /*
5753 : * Also check Xmin: if this tuple was created by an aborted
5754 : * (sub)transaction, then we already locked the last live one in the
5755 : * chain, thus we're done, so return success.
5756 : */
5757 2233 : if (TransactionIdDidAbort(HeapTupleHeaderGetXmin(mytup.t_data)))
5758 : {
5759 25 : result = TM_Ok;
5760 25 : goto out_locked;
5761 : }
5762 :
5763 2208 : old_infomask = mytup.t_data->t_infomask;
5764 2208 : old_infomask2 = mytup.t_data->t_infomask2;
5765 2208 : xmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5766 :
5767 : /*
5768 : * If this tuple version has been updated or locked by some concurrent
5769 : * transaction(s), what we do depends on whether our lock mode
5770 : * conflicts with what those other transactions hold, and also on the
5771 : * status of them.
5772 : */
5773 2208 : if (!(old_infomask & HEAP_XMAX_INVALID))
5774 : {
5775 : TransactionId rawxmax;
5776 : bool needwait;
5777 :
5778 2145 : rawxmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5779 2145 : if (old_infomask & HEAP_XMAX_IS_MULTI)
5780 : {
5781 : int nmembers;
5782 : int i;
5783 : MultiXactMember *members;
5784 :
5785 : /*
5786 : * We don't need a test for pg_upgrade'd tuples: this is only
5787 : * applied to tuples after the first in an update chain. Said
5788 : * first tuple in the chain may well be locked-in-9.2-and-
5789 : * pg_upgraded, but that one was already locked by our caller,
5790 : * not us; and any subsequent ones cannot be because our
5791 : * caller must necessarily have obtained a snapshot later than
5792 : * the pg_upgrade itself.
5793 : */
5794 : Assert(!HEAP_LOCKED_UPGRADED(mytup.t_data->t_infomask));
5795 :
5796 2109 : nmembers = GetMultiXactIdMembers(rawxmax, &members, false,
5797 2109 : HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5798 40854 : for (i = 0; i < nmembers; i++)
5799 : {
5800 38745 : result = test_lockmode_for_conflict(members[i].status,
5801 38745 : members[i].xid,
5802 : mode,
5803 : &mytup,
5804 : &needwait);
5805 :
5806 : /*
5807 : * If the tuple was already locked by ourselves in a
5808 : * previous iteration of this (say heap_lock_tuple was
5809 : * forced to restart the locking loop because of a change
5810 : * in xmax), then we hold the lock already on this tuple
5811 : * version and we don't need to do anything; and this is
5812 : * not an error condition either. We just need to skip
5813 : * this tuple and continue locking the next version in the
5814 : * update chain.
5815 : */
5816 38745 : if (result == TM_SelfModified)
5817 : {
5818 0 : pfree(members);
5819 0 : goto next;
5820 : }
5821 :
5822 38745 : if (needwait)
5823 : {
5824 0 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5825 0 : XactLockTableWait(members[i].xid, rel,
5826 : &mytup.t_self,
5827 : XLTW_LockUpdated);
5828 0 : pfree(members);
5829 0 : goto l4;
5830 : }
5831 38745 : if (result != TM_Ok)
5832 : {
5833 0 : pfree(members);
5834 0 : goto out_locked;
5835 : }
5836 : }
5837 2109 : if (members)
5838 2109 : pfree(members);
5839 : }
5840 : else
5841 : {
5842 : MultiXactStatus status;
5843 :
5844 : /*
5845 : * For a non-multi Xmax, we first need to compute the
5846 : * corresponding MultiXactStatus by using the infomask bits.
5847 : */
5848 36 : if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5849 : {
5850 16 : if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5851 16 : status = MultiXactStatusForKeyShare;
5852 0 : else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5853 0 : status = MultiXactStatusForShare;
5854 0 : else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5855 : {
5856 0 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5857 0 : status = MultiXactStatusForUpdate;
5858 : else
5859 0 : status = MultiXactStatusForNoKeyUpdate;
5860 : }
5861 : else
5862 : {
5863 : /*
5864 : * LOCK_ONLY present alone (a pg_upgraded tuple marked
5865 : * as share-locked in the old cluster) shouldn't be
5866 : * seen in the middle of an update chain.
5867 : */
5868 0 : elog(ERROR, "invalid lock status in tuple");
5869 : }
5870 : }
5871 : else
5872 : {
5873 : /* it's an update, but which kind? */
5874 20 : if (old_infomask2 & HEAP_KEYS_UPDATED)
5875 15 : status = MultiXactStatusUpdate;
5876 : else
5877 5 : status = MultiXactStatusNoKeyUpdate;
5878 : }
5879 :
5880 36 : result = test_lockmode_for_conflict(status, rawxmax, mode,
5881 : &mytup, &needwait);
5882 :
5883 : /*
5884 : * If the tuple was already locked by ourselves in a previous
5885 : * iteration of this (say heap_lock_tuple was forced to
5886 : * restart the locking loop because of a change in xmax), then
5887 : * we hold the lock already on this tuple version and we don't
5888 : * need to do anything; and this is not an error condition
5889 : * either. We just need to skip this tuple and continue
5890 : * locking the next version in the update chain.
5891 : */
5892 36 : if (result == TM_SelfModified)
5893 0 : goto next;
5894 :
5895 36 : if (needwait)
5896 : {
5897 8 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5898 8 : XactLockTableWait(rawxmax, rel, &mytup.t_self,
5899 : XLTW_LockUpdated);
5900 8 : goto l4;
5901 : }
5902 28 : if (result != TM_Ok)
5903 : {
5904 9 : goto out_locked;
5905 : }
5906 : }
5907 : }
5908 :
5909 : /* compute the new Xmax and infomask values for the tuple ... */
5910 2191 : compute_new_xmax_infomask(xmax, old_infomask, mytup.t_data->t_infomask2,
5911 : xid, mode, false,
5912 : &new_xmax, &new_infomask, &new_infomask2);
5913 :
5914 2191 : if (PageIsAllVisible(BufferGetPage(buf)) &&
5915 0 : visibilitymap_clear(rel, block, vmbuffer,
5916 : VISIBILITYMAP_ALL_FROZEN))
5917 0 : cleared_all_frozen = true;
5918 :
5919 2191 : START_CRIT_SECTION();
5920 :
5921 : /* ... and set them */
5922 2191 : HeapTupleHeaderSetXmax(mytup.t_data, new_xmax);
5923 2191 : mytup.t_data->t_infomask &= ~HEAP_XMAX_BITS;
5924 2191 : mytup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5925 2191 : mytup.t_data->t_infomask |= new_infomask;
5926 2191 : mytup.t_data->t_infomask2 |= new_infomask2;
5927 :
5928 2191 : MarkBufferDirty(buf);
5929 :
5930 : /* XLOG stuff */
5931 2191 : if (RelationNeedsWAL(rel))
5932 : {
5933 : xl_heap_lock_updated xlrec;
5934 : XLogRecPtr recptr;
5935 2191 : Page page = BufferGetPage(buf);
5936 :
5937 2191 : XLogBeginInsert();
5938 2191 : XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
5939 :
5940 2191 : xlrec.offnum = ItemPointerGetOffsetNumber(&mytup.t_self);
5941 2191 : xlrec.xmax = new_xmax;
5942 2191 : xlrec.infobits_set = compute_infobits(new_infomask, new_infomask2);
5943 2191 : xlrec.flags =
5944 2191 : cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
5945 :
5946 2191 : XLogRegisterData(&xlrec, SizeOfHeapLockUpdated);
5947 :
5948 2191 : recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_LOCK_UPDATED);
5949 :
5950 2191 : PageSetLSN(page, recptr);
5951 : }
5952 :
5953 2191 : END_CRIT_SECTION();
5954 :
5955 2191 : next:
5956 : /* if we find the end of update chain, we're done. */
5957 4382 : if (mytup.t_data->t_infomask & HEAP_XMAX_INVALID ||
5958 4382 : HeapTupleHeaderIndicatesMovedPartitions(mytup.t_data) ||
5959 2195 : ItemPointerEquals(&mytup.t_self, &mytup.t_data->t_ctid) ||
5960 4 : HeapTupleHeaderIsOnlyLocked(mytup.t_data))
5961 : {
5962 2188 : result = TM_Ok;
5963 2188 : goto out_locked;
5964 : }
5965 :
5966 : /* tail recursion */
5967 3 : priorXmax = HeapTupleHeaderGetUpdateXid(mytup.t_data);
5968 3 : ItemPointerCopy(&(mytup.t_data->t_ctid), &tupid);
5969 3 : UnlockReleaseBuffer(buf);
5970 : }
5971 :
5972 : result = TM_Ok;
5973 :
5974 2224 : out_locked:
5975 2224 : UnlockReleaseBuffer(buf);
5976 :
5977 2224 : out_unlocked:
5978 2224 : if (vmbuffer != InvalidBuffer)
5979 0 : ReleaseBuffer(vmbuffer);
5980 :
5981 2224 : return result;
5982 : }
5983 :
5984 : /*
5985 : * heap_lock_updated_tuple
5986 : * Follow update chain when locking an updated tuple, acquiring locks (row
5987 : * marks) on the updated versions.
5988 : *
5989 : * 'prior_infomask', 'prior_raw_xmax' and 'prior_ctid' are the corresponding
5990 : * fields from the initial tuple. We will lock the tuples starting from the
5991 : * one that 'prior_ctid' points to. Note: This function does not lock the
5992 : * initial tuple itself.
5993 : *
5994 : * This function doesn't check visibility, it just unconditionally marks the
5995 : * tuple(s) as locked. If any tuple in the updated chain is being deleted
5996 : * concurrently (or updated with the key being modified), sleep until the
5997 : * transaction doing it is finished.
5998 : *
5999 : * Note that we don't acquire heavyweight tuple locks on the tuples we walk
6000 : * when we have to wait for other transactions to release them, as opposed to
6001 : * what heap_lock_tuple does. The reason is that having more than one
6002 : * transaction walking the chain is probably uncommon enough that risk of
6003 : * starvation is not likely: one of the preconditions for being here is that
6004 : * the snapshot in use predates the update that created this tuple (because we
6005 : * started at an earlier version of the tuple), but at the same time such a
6006 : * transaction cannot be using repeatable read or serializable isolation
6007 : * levels, because that would lead to a serializability failure.
6008 : */
6009 : static TM_Result
6010 2226 : heap_lock_updated_tuple(Relation rel,
6011 : uint16 prior_infomask,
6012 : TransactionId prior_raw_xmax,
6013 : const ItemPointerData *prior_ctid,
6014 : TransactionId xid, LockTupleMode mode)
6015 : {
6016 2226 : INJECTION_POINT("heap_lock_updated_tuple", NULL);
6017 :
6018 : /*
6019 : * If the tuple has moved into another partition (effectively a delete)
6020 : * stop here.
6021 : */
6022 2226 : if (!ItemPointerIndicatesMovedPartitions(prior_ctid))
6023 : {
6024 : TransactionId prior_xmax;
6025 :
6026 : /*
6027 : * If this is the first possibly-multixact-able operation in the
6028 : * current transaction, set my per-backend OldestMemberMXactId
6029 : * setting. We can be certain that the transaction will never become a
6030 : * member of any older MultiXactIds than that. (We have to do this
6031 : * even if we end up just using our own TransactionId below, since
6032 : * some other backend could incorporate our XID into a MultiXact
6033 : * immediately afterwards.)
6034 : */
6035 2224 : MultiXactIdSetOldestMember();
6036 :
6037 4448 : prior_xmax = (prior_infomask & HEAP_XMAX_IS_MULTI) ?
6038 2224 : MultiXactIdGetUpdateXid(prior_raw_xmax, prior_infomask) : prior_raw_xmax;
6039 2224 : return heap_lock_updated_tuple_rec(rel, prior_xmax, prior_ctid, xid, mode);
6040 : }
6041 :
6042 : /* nothing to lock */
6043 2 : return TM_Ok;
6044 : }
6045 :
6046 : /*
6047 : * heap_finish_speculative - mark speculative insertion as successful
6048 : *
6049 : * To successfully finish a speculative insertion we have to clear speculative
6050 : * token from tuple. To do so the t_ctid field, which will contain a
6051 : * speculative token value, is modified in place to point to the tuple itself,
6052 : * which is characteristic of a newly inserted ordinary tuple.
6053 : *
6054 : * NB: It is not ok to commit without either finishing or aborting a
6055 : * speculative insertion. We could treat speculative tuples of committed
6056 : * transactions implicitly as completed, but then we would have to be prepared
6057 : * to deal with speculative tokens on committed tuples. That wouldn't be
6058 : * difficult - no-one looks at the ctid field of a tuple with invalid xmax -
6059 : * but clearing the token at completion isn't very expensive either.
6060 : * An explicit confirmation WAL record also makes logical decoding simpler.
6061 : */
6062 : void
6063 2215 : heap_finish_speculative(Relation relation, const ItemPointerData *tid)
6064 : {
6065 : Buffer buffer;
6066 : Page page;
6067 : OffsetNumber offnum;
6068 : ItemId lp;
6069 : HeapTupleHeader htup;
6070 :
6071 2215 : buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
6072 2215 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6073 2215 : page = BufferGetPage(buffer);
6074 :
6075 2215 : offnum = ItemPointerGetOffsetNumber(tid);
6076 2215 : if (offnum < 1 || offnum > PageGetMaxOffsetNumber(page))
6077 0 : elog(ERROR, "offnum out of range");
6078 2215 : lp = PageGetItemId(page, offnum);
6079 2215 : if (!ItemIdIsNormal(lp))
6080 0 : elog(ERROR, "invalid lp");
6081 :
6082 2215 : htup = (HeapTupleHeader) PageGetItem(page, lp);
6083 :
6084 : /* NO EREPORT(ERROR) from here till changes are logged */
6085 2215 : START_CRIT_SECTION();
6086 :
6087 : Assert(HeapTupleHeaderIsSpeculative(htup));
6088 :
6089 2215 : MarkBufferDirty(buffer);
6090 :
6091 : /*
6092 : * Replace the speculative insertion token with a real t_ctid, pointing to
6093 : * itself like it does on regular tuples.
6094 : */
6095 2215 : htup->t_ctid = *tid;
6096 :
6097 : /* XLOG stuff */
6098 2215 : if (RelationNeedsWAL(relation))
6099 : {
6100 : xl_heap_confirm xlrec;
6101 : XLogRecPtr recptr;
6102 :
6103 2195 : xlrec.offnum = ItemPointerGetOffsetNumber(tid);
6104 :
6105 2195 : XLogBeginInsert();
6106 :
6107 : /* We want the same filtering on this as on a plain insert */
6108 2195 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
6109 :
6110 2195 : XLogRegisterData(&xlrec, SizeOfHeapConfirm);
6111 2195 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
6112 :
6113 2195 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_CONFIRM);
6114 :
6115 2195 : PageSetLSN(page, recptr);
6116 : }
6117 :
6118 2215 : END_CRIT_SECTION();
6119 :
6120 2215 : UnlockReleaseBuffer(buffer);
6121 2215 : }
6122 :
6123 : /*
6124 : * heap_abort_speculative - kill a speculatively inserted tuple
6125 : *
6126 : * Marks a tuple that was speculatively inserted in the same command as dead,
6127 : * by setting its xmin as invalid. That makes it immediately appear as dead
6128 : * to all transactions, including our own. In particular, it makes
6129 : * HeapTupleSatisfiesDirty() regard the tuple as dead, so that another backend
6130 : * inserting a duplicate key value won't unnecessarily wait for our whole
6131 : * transaction to finish (it'll just wait for our speculative insertion to
6132 : * finish).
6133 : *
6134 : * Killing the tuple prevents "unprincipled deadlocks", which are deadlocks
6135 : * that arise due to a mutual dependency that is not user visible. By
6136 : * definition, unprincipled deadlocks cannot be prevented by the user
6137 : * reordering lock acquisition in client code, because the implementation level
6138 : * lock acquisitions are not under the user's direct control. If speculative
6139 : * inserters did not take this precaution, then under high concurrency they
6140 : * could deadlock with each other, which would not be acceptable.
6141 : *
6142 : * This is somewhat redundant with heap_delete, but we prefer to have a
6143 : * dedicated routine with stripped down requirements. Note that this is also
6144 : * used to delete the TOAST tuples created during speculative insertion.
6145 : *
6146 : * This routine does not affect logical decoding as it only looks at
6147 : * confirmation records.
6148 : */
6149 : void
6150 16 : heap_abort_speculative(Relation relation, const ItemPointerData *tid)
6151 : {
6152 16 : TransactionId xid = GetCurrentTransactionId();
6153 : ItemId lp;
6154 : HeapTupleData tp;
6155 : Page page;
6156 : BlockNumber block;
6157 : Buffer buffer;
6158 :
6159 : Assert(ItemPointerIsValid(tid));
6160 :
6161 16 : block = ItemPointerGetBlockNumber(tid);
6162 16 : buffer = ReadBuffer(relation, block);
6163 16 : page = BufferGetPage(buffer);
6164 :
6165 16 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6166 :
6167 : /*
6168 : * Page can't be all visible, we just inserted into it, and are still
6169 : * running.
6170 : */
6171 : Assert(!PageIsAllVisible(page));
6172 :
6173 16 : lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
6174 : Assert(ItemIdIsNormal(lp));
6175 :
6176 16 : tp.t_tableOid = RelationGetRelid(relation);
6177 16 : tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
6178 16 : tp.t_len = ItemIdGetLength(lp);
6179 16 : tp.t_self = *tid;
6180 :
6181 : /*
6182 : * Sanity check that the tuple really is a speculatively inserted tuple,
6183 : * inserted by us.
6184 : */
6185 16 : if (tp.t_data->t_choice.t_heap.t_xmin != xid)
6186 0 : elog(ERROR, "attempted to kill a tuple inserted by another transaction");
6187 16 : if (!(IsToastRelation(relation) || HeapTupleHeaderIsSpeculative(tp.t_data)))
6188 0 : elog(ERROR, "attempted to kill a non-speculative tuple");
6189 : Assert(!HeapTupleHeaderIsHeapOnly(tp.t_data));
6190 :
6191 : /*
6192 : * No need to check for serializable conflicts here. There is never a
6193 : * need for a combo CID, either. No need to extract replica identity, or
6194 : * do anything special with infomask bits.
6195 : */
6196 :
6197 16 : START_CRIT_SECTION();
6198 :
6199 : /*
6200 : * The tuple will become DEAD immediately. Flag that this page is a
6201 : * candidate for pruning by setting xmin to TransactionXmin. While not
6202 : * immediately prunable, it is the oldest xid we can cheaply determine
6203 : * that's safe against wraparound / being older than the table's
6204 : * relfrozenxid. To defend against the unlikely case of a new relation
6205 : * having a newer relfrozenxid than our TransactionXmin, use relfrozenxid
6206 : * if so (vacuum can't subsequently move relfrozenxid to beyond
6207 : * TransactionXmin, so there's no race here).
6208 : */
6209 : Assert(TransactionIdIsValid(TransactionXmin));
6210 : {
6211 16 : TransactionId relfrozenxid = relation->rd_rel->relfrozenxid;
6212 : TransactionId prune_xid;
6213 :
6214 16 : if (TransactionIdPrecedes(TransactionXmin, relfrozenxid))
6215 0 : prune_xid = relfrozenxid;
6216 : else
6217 16 : prune_xid = TransactionXmin;
6218 16 : PageSetPrunable(page, prune_xid);
6219 : }
6220 :
6221 : /* store transaction information of xact deleting the tuple */
6222 16 : tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
6223 16 : tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
6224 :
6225 : /*
6226 : * Set the tuple header xmin to InvalidTransactionId. This makes the
6227 : * tuple immediately invisible everyone. (In particular, to any
6228 : * transactions waiting on the speculative token, woken up later.)
6229 : */
6230 16 : HeapTupleHeaderSetXmin(tp.t_data, InvalidTransactionId);
6231 :
6232 : /* Clear the speculative insertion token too */
6233 16 : tp.t_data->t_ctid = tp.t_self;
6234 :
6235 16 : MarkBufferDirty(buffer);
6236 :
6237 : /*
6238 : * XLOG stuff
6239 : *
6240 : * The WAL records generated here match heap_delete(). The same recovery
6241 : * routines are used.
6242 : */
6243 16 : if (RelationNeedsWAL(relation))
6244 : {
6245 : xl_heap_delete xlrec;
6246 : XLogRecPtr recptr;
6247 :
6248 12 : xlrec.flags = XLH_DELETE_IS_SUPER;
6249 24 : xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
6250 12 : tp.t_data->t_infomask2);
6251 12 : xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
6252 12 : xlrec.xmax = xid;
6253 :
6254 12 : XLogBeginInsert();
6255 12 : XLogRegisterData(&xlrec, SizeOfHeapDelete);
6256 12 : XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
6257 :
6258 : /* No replica identity & replication origin logged */
6259 :
6260 12 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
6261 :
6262 12 : PageSetLSN(page, recptr);
6263 : }
6264 :
6265 16 : END_CRIT_SECTION();
6266 :
6267 16 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6268 :
6269 16 : if (HeapTupleHasExternal(&tp))
6270 : {
6271 : Assert(!IsToastRelation(relation));
6272 1 : heap_toast_delete(relation, &tp, true);
6273 : }
6274 :
6275 : /*
6276 : * Never need to mark tuple for invalidation, since catalogs don't support
6277 : * speculative insertion
6278 : */
6279 :
6280 : /* Now we can release the buffer */
6281 16 : ReleaseBuffer(buffer);
6282 :
6283 : /* count deletion, as we counted the insertion too */
6284 16 : pgstat_count_heap_delete(relation);
6285 16 : }
6286 :
6287 : /*
6288 : * heap_inplace_lock - protect inplace update from concurrent heap_update()
6289 : *
6290 : * Evaluate whether the tuple's state is compatible with a no-key update.
6291 : * Current transaction rowmarks are fine, as is KEY SHARE from any
6292 : * transaction. If compatible, return true with the buffer exclusive-locked,
6293 : * and the caller must release that by calling
6294 : * heap_inplace_update_and_unlock(), calling heap_inplace_unlock(), or raising
6295 : * an error. Otherwise, call release_callback(arg), wait for blocking
6296 : * transactions to end, and return false.
6297 : *
6298 : * Since this is intended for system catalogs and SERIALIZABLE doesn't cover
6299 : * DDL, this doesn't guarantee any particular predicate locking.
6300 : *
6301 : * heap_delete() is a rarer source of blocking transactions (xwait). We'll
6302 : * wait for such a transaction just like for the normal heap_update() case.
6303 : * Normal concurrent DROP commands won't cause that, because all inplace
6304 : * updaters take some lock that conflicts with DROP. An explicit SQL "DELETE
6305 : * FROM pg_class" can cause it. By waiting, if the concurrent transaction
6306 : * executed both "DELETE FROM pg_class" and "INSERT INTO pg_class", our caller
6307 : * can find the successor tuple.
6308 : *
6309 : * Readers of inplace-updated fields expect changes to those fields are
6310 : * durable. For example, vac_truncate_clog() reads datfrozenxid from
6311 : * pg_database tuples via catalog snapshots. A future snapshot must not
6312 : * return a lower datfrozenxid for the same database OID (lower in the
6313 : * FullTransactionIdPrecedes() sense). We achieve that since no update of a
6314 : * tuple can start while we hold a lock on its buffer. In cases like
6315 : * BEGIN;GRANT;CREATE INDEX;COMMIT we're inplace-updating a tuple visible only
6316 : * to this transaction. ROLLBACK then is one case where it's okay to lose
6317 : * inplace updates. (Restoring relhasindex=false on ROLLBACK is fine, since
6318 : * any concurrent CREATE INDEX would have blocked, then inplace-updated the
6319 : * committed tuple.)
6320 : *
6321 : * In principle, we could avoid waiting by overwriting every tuple in the
6322 : * updated tuple chain. Reader expectations permit updating a tuple only if
6323 : * it's aborted, is the tail of the chain, or we already updated the tuple
6324 : * referenced in its t_ctid. Hence, we would need to overwrite the tuples in
6325 : * order from tail to head. That would imply either (a) mutating all tuples
6326 : * in one critical section or (b) accepting a chance of partial completion.
6327 : * Partial completion of a relfrozenxid update would have the weird
6328 : * consequence that the table's next VACUUM could see the table's relfrozenxid
6329 : * move forward between vacuum_get_cutoffs() and finishing.
6330 : */
6331 : bool
6332 228804 : heap_inplace_lock(Relation relation,
6333 : HeapTuple oldtup_ptr, Buffer buffer,
6334 : void (*release_callback) (void *), void *arg)
6335 : {
6336 228804 : HeapTupleData oldtup = *oldtup_ptr; /* minimize diff vs. heap_update() */
6337 : TM_Result result;
6338 : bool ret;
6339 :
6340 : #ifdef USE_ASSERT_CHECKING
6341 : if (RelationGetRelid(relation) == RelationRelationId)
6342 : check_inplace_rel_lock(oldtup_ptr);
6343 : #endif
6344 :
6345 : Assert(BufferIsValid(buffer));
6346 :
6347 : /*
6348 : * Register shared cache invals if necessary. Other sessions may finish
6349 : * inplace updates of this tuple between this step and LockTuple(). Since
6350 : * inplace updates don't change cache keys, that's harmless.
6351 : *
6352 : * While it's tempting to register invals only after confirming we can
6353 : * return true, the following obstacle precludes reordering steps that
6354 : * way. Registering invals might reach a CatalogCacheInitializeCache()
6355 : * that locks "buffer". That would hang indefinitely if running after our
6356 : * own LockBuffer(). Hence, we must register invals before LockBuffer().
6357 : */
6358 228804 : CacheInvalidateHeapTupleInplace(relation, oldtup_ptr);
6359 :
6360 228804 : LockTuple(relation, &oldtup.t_self, InplaceUpdateTupleLock);
6361 228804 : LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6362 :
6363 : /*----------
6364 : * Interpret HeapTupleSatisfiesUpdate() like heap_update() does, except:
6365 : *
6366 : * - wait unconditionally
6367 : * - already locked tuple above, since inplace needs that unconditionally
6368 : * - don't recheck header after wait: simpler to defer to next iteration
6369 : * - don't try to continue even if the updater aborts: likewise
6370 : * - no crosscheck
6371 : */
6372 228804 : result = HeapTupleSatisfiesUpdate(&oldtup, GetCurrentCommandId(false),
6373 : buffer);
6374 :
6375 228804 : if (result == TM_Invisible)
6376 : {
6377 : /* no known way this can happen */
6378 0 : ereport(ERROR,
6379 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6380 : errmsg_internal("attempted to overwrite invisible tuple")));
6381 : }
6382 228804 : else if (result == TM_SelfModified)
6383 : {
6384 : /*
6385 : * CREATE INDEX might reach this if an expression is silly enough to
6386 : * call e.g. SELECT ... FROM pg_class FOR SHARE. C code of other SQL
6387 : * statements might get here after a heap_update() of the same row, in
6388 : * the absence of an intervening CommandCounterIncrement().
6389 : */
6390 0 : ereport(ERROR,
6391 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6392 : errmsg("tuple to be updated was already modified by an operation triggered by the current command")));
6393 : }
6394 228804 : else if (result == TM_BeingModified)
6395 : {
6396 : TransactionId xwait;
6397 : uint16 infomask;
6398 :
6399 27 : xwait = HeapTupleHeaderGetRawXmax(oldtup.t_data);
6400 27 : infomask = oldtup.t_data->t_infomask;
6401 :
6402 27 : if (infomask & HEAP_XMAX_IS_MULTI)
6403 : {
6404 5 : LockTupleMode lockmode = LockTupleNoKeyExclusive;
6405 5 : MultiXactStatus mxact_status = MultiXactStatusNoKeyUpdate;
6406 : int remain;
6407 :
6408 5 : if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
6409 : lockmode, NULL))
6410 : {
6411 2 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6412 2 : release_callback(arg);
6413 2 : ret = false;
6414 2 : MultiXactIdWait((MultiXactId) xwait, mxact_status, infomask,
6415 : relation, &oldtup.t_self, XLTW_Update,
6416 : &remain);
6417 : }
6418 : else
6419 3 : ret = true;
6420 : }
6421 22 : else if (TransactionIdIsCurrentTransactionId(xwait))
6422 1 : ret = true;
6423 21 : else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask))
6424 1 : ret = true;
6425 : else
6426 : {
6427 20 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6428 20 : release_callback(arg);
6429 20 : ret = false;
6430 20 : XactLockTableWait(xwait, relation, &oldtup.t_self,
6431 : XLTW_Update);
6432 : }
6433 : }
6434 : else
6435 : {
6436 228777 : ret = (result == TM_Ok);
6437 228777 : if (!ret)
6438 : {
6439 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6440 0 : release_callback(arg);
6441 : }
6442 : }
6443 :
6444 : /*
6445 : * GetCatalogSnapshot() relies on invalidation messages to know when to
6446 : * take a new snapshot. COMMIT of xwait is responsible for sending the
6447 : * invalidation. We're not acquiring heavyweight locks sufficient to
6448 : * block if not yet sent, so we must take a new snapshot to ensure a later
6449 : * attempt has a fair chance. While we don't need this if xwait aborted,
6450 : * don't bother optimizing that.
6451 : */
6452 228804 : if (!ret)
6453 : {
6454 22 : UnlockTuple(relation, &oldtup.t_self, InplaceUpdateTupleLock);
6455 22 : ForgetInplace_Inval();
6456 22 : InvalidateCatalogSnapshot();
6457 : }
6458 228804 : return ret;
6459 : }
6460 :
6461 : /*
6462 : * heap_inplace_update_and_unlock - core of systable_inplace_update_finish
6463 : *
6464 : * The tuple cannot change size, and therefore its header fields and null
6465 : * bitmap (if any) don't change either.
6466 : *
6467 : * Since we hold LOCKTAG_TUPLE, no updater has a local copy of this tuple.
6468 : */
6469 : void
6470 100325 : heap_inplace_update_and_unlock(Relation relation,
6471 : HeapTuple oldtup, HeapTuple tuple,
6472 : Buffer buffer)
6473 : {
6474 100325 : HeapTupleHeader htup = oldtup->t_data;
6475 : uint32 oldlen;
6476 : uint32 newlen;
6477 : char *dst;
6478 : char *src;
6479 100325 : int nmsgs = 0;
6480 100325 : SharedInvalidationMessage *invalMessages = NULL;
6481 100325 : bool RelcacheInitFileInval = false;
6482 :
6483 : Assert(ItemPointerEquals(&oldtup->t_self, &tuple->t_self));
6484 100325 : oldlen = oldtup->t_len - htup->t_hoff;
6485 100325 : newlen = tuple->t_len - tuple->t_data->t_hoff;
6486 100325 : if (oldlen != newlen || htup->t_hoff != tuple->t_data->t_hoff)
6487 0 : elog(ERROR, "wrong tuple length");
6488 :
6489 100325 : dst = (char *) htup + htup->t_hoff;
6490 100325 : src = (char *) tuple->t_data + tuple->t_data->t_hoff;
6491 :
6492 : /* Like RecordTransactionCommit(), log only if needed */
6493 100325 : if (XLogStandbyInfoActive())
6494 72938 : nmsgs = inplaceGetInvalidationMessages(&invalMessages,
6495 : &RelcacheInitFileInval);
6496 :
6497 : /*
6498 : * Unlink relcache init files as needed. If unlinking, acquire
6499 : * RelCacheInitLock until after associated invalidations. By doing this
6500 : * in advance, if we checkpoint and then crash between inplace
6501 : * XLogInsert() and inval, we don't rely on StartupXLOG() ->
6502 : * RelationCacheInitFileRemove(). That uses elevel==LOG, so replay would
6503 : * neglect to PANIC on EIO.
6504 : */
6505 100325 : PreInplace_Inval();
6506 :
6507 : /*----------
6508 : * NO EREPORT(ERROR) from here till changes are complete
6509 : *
6510 : * Our exclusive buffer lock won't stop a reader having already pinned and
6511 : * checked visibility for this tuple. With the usual order of changes
6512 : * (i.e. updating the buffer contents before WAL logging), a reader could
6513 : * observe our not-yet-persistent update to relfrozenxid and update
6514 : * datfrozenxid based on that. A crash in that moment could allow
6515 : * datfrozenxid to overtake relfrozenxid:
6516 : *
6517 : * ["D" is a VACUUM (ONLY_DATABASE_STATS)]
6518 : * ["R" is a VACUUM tbl]
6519 : * D: vac_update_datfrozenxid() -> systable_beginscan(pg_class)
6520 : * D: systable_getnext() returns pg_class tuple of tbl
6521 : * R: memcpy() into pg_class tuple of tbl
6522 : * D: raise pg_database.datfrozenxid, XLogInsert(), finish
6523 : * [crash]
6524 : * [recovery restores datfrozenxid w/o relfrozenxid]
6525 : *
6526 : * We avoid that by using a temporary copy of the buffer to hide our
6527 : * change from other backends until the change has been WAL-logged. We
6528 : * apply our change to the temporary copy and WAL-log it, before modifying
6529 : * the real page. That way any action a reader of the in-place-updated
6530 : * value takes will be WAL logged after this change.
6531 : */
6532 100325 : START_CRIT_SECTION();
6533 :
6534 100325 : MarkBufferDirty(buffer);
6535 :
6536 : /* XLOG stuff */
6537 100325 : if (RelationNeedsWAL(relation))
6538 : {
6539 : xl_heap_inplace xlrec;
6540 : PGAlignedBlock copied_buffer;
6541 100321 : char *origdata = (char *) BufferGetBlock(buffer);
6542 100321 : Page page = BufferGetPage(buffer);
6543 100321 : uint16 lower = ((PageHeader) page)->pd_lower;
6544 100321 : uint16 upper = ((PageHeader) page)->pd_upper;
6545 : uintptr_t dst_offset_in_block;
6546 : RelFileLocator rlocator;
6547 : ForkNumber forkno;
6548 : BlockNumber blkno;
6549 : XLogRecPtr recptr;
6550 :
6551 100321 : xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
6552 100321 : xlrec.dbId = MyDatabaseId;
6553 100321 : xlrec.tsId = MyDatabaseTableSpace;
6554 100321 : xlrec.relcacheInitFileInval = RelcacheInitFileInval;
6555 100321 : xlrec.nmsgs = nmsgs;
6556 :
6557 100321 : XLogBeginInsert();
6558 100321 : XLogRegisterData(&xlrec, MinSizeOfHeapInplace);
6559 100321 : if (nmsgs != 0)
6560 54983 : XLogRegisterData(invalMessages,
6561 : nmsgs * sizeof(SharedInvalidationMessage));
6562 :
6563 : /* register block matching what buffer will look like after changes */
6564 100321 : memcpy(copied_buffer.data, origdata, lower);
6565 100321 : memcpy(copied_buffer.data + upper, origdata + upper, BLCKSZ - upper);
6566 100321 : dst_offset_in_block = dst - origdata;
6567 100321 : memcpy(copied_buffer.data + dst_offset_in_block, src, newlen);
6568 100321 : BufferGetTag(buffer, &rlocator, &forkno, &blkno);
6569 : Assert(forkno == MAIN_FORKNUM);
6570 100321 : XLogRegisterBlock(0, &rlocator, forkno, blkno, copied_buffer.data,
6571 : REGBUF_STANDARD);
6572 100321 : XLogRegisterBufData(0, src, newlen);
6573 :
6574 : /* inplace updates aren't decoded atm, don't log the origin */
6575 :
6576 100321 : recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_INPLACE);
6577 :
6578 100321 : PageSetLSN(page, recptr);
6579 : }
6580 :
6581 100325 : memcpy(dst, src, newlen);
6582 :
6583 100325 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6584 :
6585 : /*
6586 : * Send invalidations to shared queue. SearchSysCacheLocked1() assumes we
6587 : * do this before UnlockTuple().
6588 : */
6589 100325 : AtInplace_Inval();
6590 :
6591 100325 : END_CRIT_SECTION();
6592 100325 : UnlockTuple(relation, &tuple->t_self, InplaceUpdateTupleLock);
6593 :
6594 100325 : AcceptInvalidationMessages(); /* local processing of just-sent inval */
6595 :
6596 : /*
6597 : * Queue a transactional inval, for logical decoding and for third-party
6598 : * code that might have been relying on it since long before inplace
6599 : * update adopted immediate invalidation. See README.tuplock section
6600 : * "Reading inplace-updated columns" for logical decoding details.
6601 : */
6602 100325 : if (!IsBootstrapProcessingMode())
6603 82370 : CacheInvalidateHeapTuple(relation, tuple, NULL);
6604 100325 : }
6605 :
6606 : /*
6607 : * heap_inplace_unlock - reverse of heap_inplace_lock
6608 : */
6609 : void
6610 128457 : heap_inplace_unlock(Relation relation,
6611 : HeapTuple oldtup, Buffer buffer)
6612 : {
6613 128457 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
6614 128457 : UnlockTuple(relation, &oldtup->t_self, InplaceUpdateTupleLock);
6615 128457 : ForgetInplace_Inval();
6616 128457 : }
6617 :
6618 : #define FRM_NOOP 0x0001
6619 : #define FRM_INVALIDATE_XMAX 0x0002
6620 : #define FRM_RETURN_IS_XID 0x0004
6621 : #define FRM_RETURN_IS_MULTI 0x0008
6622 : #define FRM_MARK_COMMITTED 0x0010
6623 :
6624 : /*
6625 : * FreezeMultiXactId
6626 : * Determine what to do during freezing when a tuple is marked by a
6627 : * MultiXactId.
6628 : *
6629 : * "flags" is an output value; it's used to tell caller what to do on return.
6630 : * "pagefrz" is an input/output value, used to manage page level freezing.
6631 : *
6632 : * Possible values that we can set in "flags":
6633 : * FRM_NOOP
6634 : * don't do anything -- keep existing Xmax
6635 : * FRM_INVALIDATE_XMAX
6636 : * mark Xmax as InvalidTransactionId and set XMAX_INVALID flag.
6637 : * FRM_RETURN_IS_XID
6638 : * The Xid return value is a single update Xid to set as xmax.
6639 : * FRM_MARK_COMMITTED
6640 : * Xmax can be marked as HEAP_XMAX_COMMITTED
6641 : * FRM_RETURN_IS_MULTI
6642 : * The return value is a new MultiXactId to set as new Xmax.
6643 : * (caller must obtain proper infomask bits using GetMultiXactIdHintBits)
6644 : *
6645 : * Caller delegates control of page freezing to us. In practice we always
6646 : * force freezing of caller's page unless FRM_NOOP processing is indicated.
6647 : * We help caller ensure that XIDs < FreezeLimit and MXIDs < MultiXactCutoff
6648 : * can never be left behind. We freely choose when and how to process each
6649 : * Multi, without ever violating the cutoff postconditions for freezing.
6650 : *
6651 : * It's useful to remove Multis on a proactive timeline (relative to freezing
6652 : * XIDs) to keep MultiXact member SLRU buffer misses to a minimum. It can also
6653 : * be cheaper in the short run, for us, since we too can avoid SLRU buffer
6654 : * misses through eager processing.
6655 : *
6656 : * NB: Creates a _new_ MultiXactId when FRM_RETURN_IS_MULTI is set, though only
6657 : * when FreezeLimit and/or MultiXactCutoff cutoffs leave us with no choice.
6658 : * This can usually be put off, which is usually enough to avoid it altogether.
6659 : * Allocating new multis during VACUUM should be avoided on general principle;
6660 : * only VACUUM can advance relminmxid, so allocating new Multis here comes with
6661 : * its own special risks.
6662 : *
6663 : * NB: Caller must maintain "no freeze" NewRelfrozenXid/NewRelminMxid trackers
6664 : * using heap_tuple_should_freeze when we haven't forced page-level freezing.
6665 : *
6666 : * NB: Caller should avoid needlessly calling heap_tuple_should_freeze when we
6667 : * have already forced page-level freezing, since that might incur the same
6668 : * SLRU buffer misses that we specifically intended to avoid by freezing.
6669 : */
6670 : static TransactionId
6671 7 : FreezeMultiXactId(MultiXactId multi, uint16 t_infomask,
6672 : const struct VacuumCutoffs *cutoffs, uint16 *flags,
6673 : HeapPageFreeze *pagefrz)
6674 : {
6675 : TransactionId newxmax;
6676 : MultiXactMember *members;
6677 : int nmembers;
6678 : bool need_replace;
6679 : int nnewmembers;
6680 : MultiXactMember *newmembers;
6681 : bool has_lockers;
6682 : TransactionId update_xid;
6683 : bool update_committed;
6684 : TransactionId FreezePageRelfrozenXid;
6685 :
6686 7 : *flags = 0;
6687 :
6688 : /* We should only be called in Multis */
6689 : Assert(t_infomask & HEAP_XMAX_IS_MULTI);
6690 :
6691 14 : if (!MultiXactIdIsValid(multi) ||
6692 7 : HEAP_LOCKED_UPGRADED(t_infomask))
6693 : {
6694 0 : *flags |= FRM_INVALIDATE_XMAX;
6695 0 : pagefrz->freeze_required = true;
6696 0 : return InvalidTransactionId;
6697 : }
6698 7 : else if (MultiXactIdPrecedes(multi, cutoffs->relminmxid))
6699 0 : ereport(ERROR,
6700 : (errcode(ERRCODE_DATA_CORRUPTED),
6701 : errmsg_internal("found multixact %u from before relminmxid %u",
6702 : multi, cutoffs->relminmxid)));
6703 7 : else if (MultiXactIdPrecedes(multi, cutoffs->OldestMxact))
6704 : {
6705 : TransactionId update_xact;
6706 :
6707 : /*
6708 : * This old multi cannot possibly have members still running, but
6709 : * verify just in case. If it was a locker only, it can be removed
6710 : * without any further consideration; but if it contained an update,
6711 : * we might need to preserve it.
6712 : */
6713 5 : if (MultiXactIdIsRunning(multi,
6714 5 : HEAP_XMAX_IS_LOCKED_ONLY(t_infomask)))
6715 0 : ereport(ERROR,
6716 : (errcode(ERRCODE_DATA_CORRUPTED),
6717 : errmsg_internal("multixact %u from before multi freeze cutoff %u found to be still running",
6718 : multi, cutoffs->OldestMxact)));
6719 :
6720 5 : if (HEAP_XMAX_IS_LOCKED_ONLY(t_infomask))
6721 : {
6722 5 : *flags |= FRM_INVALIDATE_XMAX;
6723 5 : pagefrz->freeze_required = true;
6724 5 : return InvalidTransactionId;
6725 : }
6726 :
6727 : /* replace multi with single XID for its updater? */
6728 0 : update_xact = MultiXactIdGetUpdateXid(multi, t_infomask);
6729 0 : if (TransactionIdPrecedes(update_xact, cutoffs->relfrozenxid))
6730 0 : ereport(ERROR,
6731 : (errcode(ERRCODE_DATA_CORRUPTED),
6732 : errmsg_internal("multixact %u contains update XID %u from before relfrozenxid %u",
6733 : multi, update_xact,
6734 : cutoffs->relfrozenxid)));
6735 0 : else if (TransactionIdPrecedes(update_xact, cutoffs->OldestXmin))
6736 : {
6737 : /*
6738 : * Updater XID has to have aborted (otherwise the tuple would have
6739 : * been pruned away instead, since updater XID is < OldestXmin).
6740 : * Just remove xmax.
6741 : */
6742 0 : if (TransactionIdDidCommit(update_xact))
6743 0 : ereport(ERROR,
6744 : (errcode(ERRCODE_DATA_CORRUPTED),
6745 : errmsg_internal("multixact %u contains committed update XID %u from before removable cutoff %u",
6746 : multi, update_xact,
6747 : cutoffs->OldestXmin)));
6748 0 : *flags |= FRM_INVALIDATE_XMAX;
6749 0 : pagefrz->freeze_required = true;
6750 0 : return InvalidTransactionId;
6751 : }
6752 :
6753 : /* Have to keep updater XID as new xmax */
6754 0 : *flags |= FRM_RETURN_IS_XID;
6755 0 : pagefrz->freeze_required = true;
6756 0 : return update_xact;
6757 : }
6758 :
6759 : /*
6760 : * Some member(s) of this Multi may be below FreezeLimit xid cutoff, so we
6761 : * need to walk the whole members array to figure out what to do, if
6762 : * anything.
6763 : */
6764 : nmembers =
6765 2 : GetMultiXactIdMembers(multi, &members, false,
6766 2 : HEAP_XMAX_IS_LOCKED_ONLY(t_infomask));
6767 2 : if (nmembers <= 0)
6768 : {
6769 : /* Nothing worth keeping */
6770 0 : *flags |= FRM_INVALIDATE_XMAX;
6771 0 : pagefrz->freeze_required = true;
6772 0 : return InvalidTransactionId;
6773 : }
6774 :
6775 : /*
6776 : * The FRM_NOOP case is the only case where we might need to ratchet back
6777 : * FreezePageRelfrozenXid or FreezePageRelminMxid. It is also the only
6778 : * case where our caller might ratchet back its NoFreezePageRelfrozenXid
6779 : * or NoFreezePageRelminMxid "no freeze" trackers to deal with a multi.
6780 : * FRM_NOOP handling should result in the NewRelfrozenXid/NewRelminMxid
6781 : * trackers managed by VACUUM being ratcheting back by xmax to the degree
6782 : * required to make it safe to leave xmax undisturbed, independent of
6783 : * whether or not page freezing is triggered somewhere else.
6784 : *
6785 : * Our policy is to force freezing in every case other than FRM_NOOP,
6786 : * which obviates the need to maintain either set of trackers, anywhere.
6787 : * Every other case will reliably execute a freeze plan for xmax that
6788 : * either replaces xmax with an XID/MXID >= OldestXmin/OldestMxact, or
6789 : * sets xmax to an InvalidTransactionId XID, rendering xmax fully frozen.
6790 : * (VACUUM's NewRelfrozenXid/NewRelminMxid trackers are initialized with
6791 : * OldestXmin/OldestMxact, so later values never need to be tracked here.)
6792 : */
6793 2 : need_replace = false;
6794 2 : FreezePageRelfrozenXid = pagefrz->FreezePageRelfrozenXid;
6795 4 : for (int i = 0; i < nmembers; i++)
6796 : {
6797 3 : TransactionId xid = members[i].xid;
6798 :
6799 : Assert(!TransactionIdPrecedes(xid, cutoffs->relfrozenxid));
6800 :
6801 3 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
6802 : {
6803 : /* Can't violate the FreezeLimit postcondition */
6804 1 : need_replace = true;
6805 1 : break;
6806 : }
6807 2 : if (TransactionIdPrecedes(xid, FreezePageRelfrozenXid))
6808 0 : FreezePageRelfrozenXid = xid;
6809 : }
6810 :
6811 : /* Can't violate the MultiXactCutoff postcondition, either */
6812 2 : if (!need_replace)
6813 1 : need_replace = MultiXactIdPrecedes(multi, cutoffs->MultiXactCutoff);
6814 :
6815 2 : if (!need_replace)
6816 : {
6817 : /*
6818 : * vacuumlazy.c might ratchet back NewRelminMxid, NewRelfrozenXid, or
6819 : * both together to make it safe to retain this particular multi after
6820 : * freezing its page
6821 : */
6822 1 : *flags |= FRM_NOOP;
6823 1 : pagefrz->FreezePageRelfrozenXid = FreezePageRelfrozenXid;
6824 1 : if (MultiXactIdPrecedes(multi, pagefrz->FreezePageRelminMxid))
6825 0 : pagefrz->FreezePageRelminMxid = multi;
6826 1 : pfree(members);
6827 1 : return multi;
6828 : }
6829 :
6830 : /*
6831 : * Do a more thorough second pass over the multi to figure out which
6832 : * member XIDs actually need to be kept. Checking the precise status of
6833 : * individual members might even show that we don't need to keep anything.
6834 : * That is quite possible even though the Multi must be >= OldestMxact,
6835 : * since our second pass only keeps member XIDs when it's truly necessary;
6836 : * even member XIDs >= OldestXmin often won't be kept by second pass.
6837 : */
6838 1 : nnewmembers = 0;
6839 1 : newmembers = palloc_array(MultiXactMember, nmembers);
6840 1 : has_lockers = false;
6841 1 : update_xid = InvalidTransactionId;
6842 1 : update_committed = false;
6843 :
6844 : /*
6845 : * Determine whether to keep each member xid, or to ignore it instead
6846 : */
6847 3 : for (int i = 0; i < nmembers; i++)
6848 : {
6849 2 : TransactionId xid = members[i].xid;
6850 2 : MultiXactStatus mstatus = members[i].status;
6851 :
6852 : Assert(!TransactionIdPrecedes(xid, cutoffs->relfrozenxid));
6853 :
6854 2 : if (!ISUPDATE_from_mxstatus(mstatus))
6855 : {
6856 : /*
6857 : * Locker XID (not updater XID). We only keep lockers that are
6858 : * still running.
6859 : */
6860 4 : if (TransactionIdIsCurrentTransactionId(xid) ||
6861 2 : TransactionIdIsInProgress(xid))
6862 : {
6863 1 : if (TransactionIdPrecedes(xid, cutoffs->OldestXmin))
6864 0 : ereport(ERROR,
6865 : (errcode(ERRCODE_DATA_CORRUPTED),
6866 : errmsg_internal("multixact %u contains running locker XID %u from before removable cutoff %u",
6867 : multi, xid,
6868 : cutoffs->OldestXmin)));
6869 1 : newmembers[nnewmembers++] = members[i];
6870 1 : has_lockers = true;
6871 : }
6872 :
6873 2 : continue;
6874 : }
6875 :
6876 : /*
6877 : * Updater XID (not locker XID). Should we keep it?
6878 : *
6879 : * Since the tuple wasn't totally removed when vacuum pruned, the
6880 : * update Xid cannot possibly be older than OldestXmin cutoff unless
6881 : * the updater XID aborted. If the updater transaction is known
6882 : * aborted or crashed then it's okay to ignore it, otherwise not.
6883 : *
6884 : * In any case the Multi should never contain two updaters, whatever
6885 : * their individual commit status. Check for that first, in passing.
6886 : */
6887 0 : if (TransactionIdIsValid(update_xid))
6888 0 : ereport(ERROR,
6889 : (errcode(ERRCODE_DATA_CORRUPTED),
6890 : errmsg_internal("multixact %u has two or more updating members",
6891 : multi),
6892 : errdetail_internal("First updater XID=%u second updater XID=%u.",
6893 : update_xid, xid)));
6894 :
6895 : /*
6896 : * As with all tuple visibility routines, it's critical to test
6897 : * TransactionIdIsInProgress before TransactionIdDidCommit, because of
6898 : * race conditions explained in detail in heapam_visibility.c.
6899 : */
6900 0 : if (TransactionIdIsCurrentTransactionId(xid) ||
6901 0 : TransactionIdIsInProgress(xid))
6902 0 : update_xid = xid;
6903 0 : else if (TransactionIdDidCommit(xid))
6904 : {
6905 : /*
6906 : * The transaction committed, so we can tell caller to set
6907 : * HEAP_XMAX_COMMITTED. (We can only do this because we know the
6908 : * transaction is not running.)
6909 : */
6910 0 : update_committed = true;
6911 0 : update_xid = xid;
6912 : }
6913 : else
6914 : {
6915 : /*
6916 : * Not in progress, not committed -- must be aborted or crashed;
6917 : * we can ignore it.
6918 : */
6919 0 : continue;
6920 : }
6921 :
6922 : /*
6923 : * We determined that updater must be kept -- add it to pending new
6924 : * members list
6925 : */
6926 0 : if (TransactionIdPrecedes(xid, cutoffs->OldestXmin))
6927 0 : ereport(ERROR,
6928 : (errcode(ERRCODE_DATA_CORRUPTED),
6929 : errmsg_internal("multixact %u contains committed update XID %u from before removable cutoff %u",
6930 : multi, xid, cutoffs->OldestXmin)));
6931 0 : newmembers[nnewmembers++] = members[i];
6932 : }
6933 :
6934 1 : pfree(members);
6935 :
6936 : /*
6937 : * Determine what to do with caller's multi based on information gathered
6938 : * during our second pass
6939 : */
6940 1 : if (nnewmembers == 0)
6941 : {
6942 : /* Nothing worth keeping */
6943 0 : *flags |= FRM_INVALIDATE_XMAX;
6944 0 : newxmax = InvalidTransactionId;
6945 : }
6946 1 : else if (TransactionIdIsValid(update_xid) && !has_lockers)
6947 : {
6948 : /*
6949 : * If there's a single member and it's an update, pass it back alone
6950 : * without creating a new Multi. (XXX we could do this when there's a
6951 : * single remaining locker, too, but that would complicate the API too
6952 : * much; moreover, the case with the single updater is more
6953 : * interesting, because those are longer-lived.)
6954 : */
6955 : Assert(nnewmembers == 1);
6956 0 : *flags |= FRM_RETURN_IS_XID;
6957 0 : if (update_committed)
6958 0 : *flags |= FRM_MARK_COMMITTED;
6959 0 : newxmax = update_xid;
6960 : }
6961 : else
6962 : {
6963 : /*
6964 : * Create a new multixact with the surviving members of the previous
6965 : * one, to set as new Xmax in the tuple
6966 : */
6967 1 : newxmax = MultiXactIdCreateFromMembers(nnewmembers, newmembers);
6968 1 : *flags |= FRM_RETURN_IS_MULTI;
6969 : }
6970 :
6971 1 : pfree(newmembers);
6972 :
6973 1 : pagefrz->freeze_required = true;
6974 1 : return newxmax;
6975 : }
6976 :
6977 : /*
6978 : * heap_prepare_freeze_tuple
6979 : *
6980 : * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
6981 : * are older than the OldestXmin and/or OldestMxact freeze cutoffs. If so,
6982 : * setup enough state (in the *frz output argument) to enable caller to
6983 : * process this tuple as part of freezing its page, and return true. Return
6984 : * false if nothing can be changed about the tuple right now.
6985 : *
6986 : * FreezePageConflictXid is advanced only for xmin/xvac freezing, not for xmax
6987 : * changes. We only remove xmax state here when it is lock-only, or when the
6988 : * updater XID (including an updater member of a MultiXact) must be aborted;
6989 : * otherwise, the tuple would already be removable. Neither case affects
6990 : * visibility on a standby.
6991 : *
6992 : * Also sets *totally_frozen to true if the tuple will be totally frozen once
6993 : * caller executes returned freeze plan (or if the tuple was already totally
6994 : * frozen by an earlier VACUUM). This indicates that there are no remaining
6995 : * XIDs or MultiXactIds that will need to be processed by a future VACUUM.
6996 : *
6997 : * VACUUM caller must assemble HeapTupleFreeze freeze plan entries for every
6998 : * tuple that we returned true for, and then execute freezing. Caller must
6999 : * initialize pagefrz fields for page as a whole before first call here for
7000 : * each heap page.
7001 : *
7002 : * VACUUM caller decides on whether or not to freeze the page as a whole.
7003 : * We'll often prepare freeze plans for a page that caller just discards.
7004 : * However, VACUUM doesn't always get to make a choice; it must freeze when
7005 : * pagefrz.freeze_required is set, to ensure that any XIDs < FreezeLimit (and
7006 : * MXIDs < MultiXactCutoff) can never be left behind. We help to make sure
7007 : * that VACUUM always follows that rule.
7008 : *
7009 : * We sometimes force freezing of xmax MultiXactId values long before it is
7010 : * strictly necessary to do so just to ensure the FreezeLimit postcondition.
7011 : * It's worth processing MultiXactIds proactively when it is cheap to do so,
7012 : * and it's convenient to make that happen by piggy-backing it on the "force
7013 : * freezing" mechanism. Conversely, we sometimes delay freezing MultiXactIds
7014 : * because it is expensive right now (though only when it's still possible to
7015 : * do so without violating the FreezeLimit/MultiXactCutoff postcondition).
7016 : *
7017 : * It is assumed that the caller has checked the tuple with
7018 : * HeapTupleSatisfiesVacuum() and determined that it is not HEAPTUPLE_DEAD
7019 : * (else we should be removing the tuple, not freezing it).
7020 : *
7021 : * NB: This function has side effects: it might allocate a new MultiXactId.
7022 : * It will be set as tuple's new xmax when our *frz output is processed within
7023 : * heap_execute_freeze_tuple later on. If the tuple is in a shared buffer
7024 : * then caller had better have an exclusive lock on it already.
7025 : */
7026 : bool
7027 11926819 : heap_prepare_freeze_tuple(HeapTupleHeader tuple,
7028 : const struct VacuumCutoffs *cutoffs,
7029 : HeapPageFreeze *pagefrz,
7030 : HeapTupleFreeze *frz, bool *totally_frozen)
7031 : {
7032 11926819 : bool xmin_already_frozen = false,
7033 11926819 : xmax_already_frozen = false;
7034 11926819 : bool freeze_xmin = false,
7035 11926819 : replace_xvac = false,
7036 11926819 : replace_xmax = false,
7037 11926819 : freeze_xmax = false;
7038 : TransactionId xid;
7039 :
7040 11926819 : frz->xmax = HeapTupleHeaderGetRawXmax(tuple);
7041 11926819 : frz->t_infomask2 = tuple->t_infomask2;
7042 11926819 : frz->t_infomask = tuple->t_infomask;
7043 11926819 : frz->frzflags = 0;
7044 11926819 : frz->checkflags = 0;
7045 :
7046 : /*
7047 : * Process xmin, while keeping track of whether it's already frozen, or
7048 : * will become frozen iff our freeze plan is executed by caller (could be
7049 : * neither).
7050 : */
7051 11926819 : xid = HeapTupleHeaderGetXmin(tuple);
7052 11926819 : if (!TransactionIdIsNormal(xid))
7053 7982872 : xmin_already_frozen = true;
7054 : else
7055 : {
7056 3943947 : if (TransactionIdPrecedes(xid, cutoffs->relfrozenxid))
7057 0 : ereport(ERROR,
7058 : (errcode(ERRCODE_DATA_CORRUPTED),
7059 : errmsg_internal("found xmin %u from before relfrozenxid %u",
7060 : xid, cutoffs->relfrozenxid)));
7061 :
7062 : /* Will set freeze_xmin flags in freeze plan below */
7063 3943947 : freeze_xmin = TransactionIdPrecedes(xid, cutoffs->OldestXmin);
7064 :
7065 : /* Verify that xmin committed if and when freeze plan is executed */
7066 3943947 : if (freeze_xmin)
7067 : {
7068 3209783 : frz->checkflags |= HEAP_FREEZE_CHECK_XMIN_COMMITTED;
7069 3209783 : if (TransactionIdFollows(xid, pagefrz->FreezePageConflictXid))
7070 507788 : pagefrz->FreezePageConflictXid = xid;
7071 : }
7072 : }
7073 :
7074 : /*
7075 : * Old-style VACUUM FULL is gone, but we have to process xvac for as long
7076 : * as we support having MOVED_OFF/MOVED_IN tuples in the database
7077 : */
7078 11926819 : xid = HeapTupleHeaderGetXvac(tuple);
7079 11926819 : if (TransactionIdIsNormal(xid))
7080 : {
7081 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7082 : Assert(TransactionIdPrecedes(xid, cutoffs->OldestXmin));
7083 :
7084 : /*
7085 : * For Xvac, we always freeze proactively. This allows totally_frozen
7086 : * tracking to ignore xvac.
7087 : */
7088 0 : replace_xvac = pagefrz->freeze_required = true;
7089 :
7090 0 : if (TransactionIdFollows(xid, pagefrz->FreezePageConflictXid))
7091 0 : pagefrz->FreezePageConflictXid = xid;
7092 :
7093 : /* Will set replace_xvac flags in freeze plan below */
7094 : }
7095 :
7096 : /* Now process xmax */
7097 11926819 : xid = frz->xmax;
7098 11926819 : if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
7099 : {
7100 : /* Raw xmax is a MultiXactId */
7101 : TransactionId newxmax;
7102 : uint16 flags;
7103 :
7104 : /*
7105 : * We will either remove xmax completely (in the "freeze_xmax" path),
7106 : * process xmax by replacing it (in the "replace_xmax" path), or
7107 : * perform no-op xmax processing. The only constraint is that the
7108 : * FreezeLimit/MultiXactCutoff postcondition must never be violated.
7109 : */
7110 7 : newxmax = FreezeMultiXactId(xid, tuple->t_infomask, cutoffs,
7111 : &flags, pagefrz);
7112 :
7113 7 : if (flags & FRM_NOOP)
7114 : {
7115 : /*
7116 : * xmax is a MultiXactId, and nothing about it changes for now.
7117 : * This is the only case where 'freeze_required' won't have been
7118 : * set for us by FreezeMultiXactId, as well as the only case where
7119 : * neither freeze_xmax nor replace_xmax are set (given a multi).
7120 : *
7121 : * This is a no-op, but the call to FreezeMultiXactId might have
7122 : * ratcheted back NewRelfrozenXid and/or NewRelminMxid trackers
7123 : * for us (the "freeze page" variants, specifically). That'll
7124 : * make it safe for our caller to freeze the page later on, while
7125 : * leaving this particular xmax undisturbed.
7126 : *
7127 : * FreezeMultiXactId is _not_ responsible for the "no freeze"
7128 : * NewRelfrozenXid/NewRelminMxid trackers, though -- that's our
7129 : * job. A call to heap_tuple_should_freeze for this same tuple
7130 : * will take place below if 'freeze_required' isn't set already.
7131 : * (This repeats work from FreezeMultiXactId, but allows "no
7132 : * freeze" tracker maintenance to happen in only one place.)
7133 : */
7134 : Assert(!MultiXactIdPrecedes(newxmax, cutoffs->MultiXactCutoff));
7135 : Assert(MultiXactIdIsValid(newxmax) && xid == newxmax);
7136 : }
7137 6 : else if (flags & FRM_RETURN_IS_XID)
7138 : {
7139 : /*
7140 : * xmax will become an updater Xid (original MultiXact's updater
7141 : * member Xid will be carried forward as a simple Xid in Xmax).
7142 : */
7143 : Assert(!TransactionIdPrecedes(newxmax, cutoffs->OldestXmin));
7144 :
7145 : /*
7146 : * NB -- some of these transformations are only valid because we
7147 : * know the return Xid is a tuple updater (i.e. not merely a
7148 : * locker.) Also note that the only reason we don't explicitly
7149 : * worry about HEAP_KEYS_UPDATED is because it lives in
7150 : * t_infomask2 rather than t_infomask.
7151 : */
7152 0 : frz->t_infomask &= ~HEAP_XMAX_BITS;
7153 0 : frz->xmax = newxmax;
7154 0 : if (flags & FRM_MARK_COMMITTED)
7155 0 : frz->t_infomask |= HEAP_XMAX_COMMITTED;
7156 0 : replace_xmax = true;
7157 : }
7158 6 : else if (flags & FRM_RETURN_IS_MULTI)
7159 : {
7160 : uint16 newbits;
7161 : uint16 newbits2;
7162 :
7163 : /*
7164 : * xmax is an old MultiXactId that we have to replace with a new
7165 : * MultiXactId, to carry forward two or more original member XIDs.
7166 : */
7167 : Assert(!MultiXactIdPrecedes(newxmax, cutoffs->OldestMxact));
7168 :
7169 : /*
7170 : * We can't use GetMultiXactIdHintBits directly on the new multi
7171 : * here; that routine initializes the masks to all zeroes, which
7172 : * would lose other bits we need. Doing it this way ensures all
7173 : * unrelated bits remain untouched.
7174 : */
7175 1 : frz->t_infomask &= ~HEAP_XMAX_BITS;
7176 1 : frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
7177 1 : GetMultiXactIdHintBits(newxmax, &newbits, &newbits2);
7178 1 : frz->t_infomask |= newbits;
7179 1 : frz->t_infomask2 |= newbits2;
7180 1 : frz->xmax = newxmax;
7181 1 : replace_xmax = true;
7182 : }
7183 : else
7184 : {
7185 : /*
7186 : * Freeze plan for tuple "freezes xmax" in the strictest sense:
7187 : * it'll leave nothing in xmax (neither an Xid nor a MultiXactId).
7188 : */
7189 : Assert(flags & FRM_INVALIDATE_XMAX);
7190 : Assert(!TransactionIdIsValid(newxmax));
7191 :
7192 : /* Will set freeze_xmax flags in freeze plan below */
7193 5 : freeze_xmax = true;
7194 : }
7195 :
7196 : /* MultiXactId processing forces freezing (barring FRM_NOOP case) */
7197 : Assert(pagefrz->freeze_required || (!freeze_xmax && !replace_xmax));
7198 : }
7199 11926812 : else if (TransactionIdIsNormal(xid))
7200 : {
7201 : /* Raw xmax is normal XID */
7202 5480910 : if (TransactionIdPrecedes(xid, cutoffs->relfrozenxid))
7203 0 : ereport(ERROR,
7204 : (errcode(ERRCODE_DATA_CORRUPTED),
7205 : errmsg_internal("found xmax %u from before relfrozenxid %u",
7206 : xid, cutoffs->relfrozenxid)));
7207 :
7208 : /* Will set freeze_xmax flags in freeze plan below */
7209 5480910 : freeze_xmax = TransactionIdPrecedes(xid, cutoffs->OldestXmin);
7210 :
7211 : /*
7212 : * Verify that xmax aborted if and when freeze plan is executed,
7213 : * provided it's from an update. (A lock-only xmax can be removed
7214 : * independent of this, since the lock is released at xact end.)
7215 : */
7216 5480910 : if (freeze_xmax && !HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask))
7217 997 : frz->checkflags |= HEAP_FREEZE_CHECK_XMAX_ABORTED;
7218 : }
7219 6445902 : else if (!TransactionIdIsValid(xid))
7220 : {
7221 : /* Raw xmax is InvalidTransactionId XID */
7222 : Assert((tuple->t_infomask & HEAP_XMAX_IS_MULTI) == 0);
7223 6445902 : xmax_already_frozen = true;
7224 : }
7225 : else
7226 0 : ereport(ERROR,
7227 : (errcode(ERRCODE_DATA_CORRUPTED),
7228 : errmsg_internal("found raw xmax %u (infomask 0x%04x) not invalid and not multi",
7229 : xid, tuple->t_infomask)));
7230 :
7231 11926819 : if (freeze_xmin)
7232 : {
7233 : Assert(!xmin_already_frozen);
7234 :
7235 3209783 : frz->t_infomask |= HEAP_XMIN_FROZEN;
7236 : }
7237 11926819 : if (replace_xvac)
7238 : {
7239 : /*
7240 : * If a MOVED_OFF tuple is not dead, the xvac transaction must have
7241 : * failed; whereas a non-dead MOVED_IN tuple must mean the xvac
7242 : * transaction succeeded.
7243 : */
7244 : Assert(pagefrz->freeze_required);
7245 0 : if (tuple->t_infomask & HEAP_MOVED_OFF)
7246 0 : frz->frzflags |= XLH_INVALID_XVAC;
7247 : else
7248 0 : frz->frzflags |= XLH_FREEZE_XVAC;
7249 : }
7250 : if (replace_xmax)
7251 : {
7252 : Assert(!xmax_already_frozen && !freeze_xmax);
7253 : Assert(pagefrz->freeze_required);
7254 :
7255 : /* Already set replace_xmax flags in freeze plan earlier */
7256 : }
7257 11926819 : if (freeze_xmax)
7258 : {
7259 : Assert(!xmax_already_frozen && !replace_xmax);
7260 :
7261 1879 : frz->xmax = InvalidTransactionId;
7262 :
7263 : /*
7264 : * The tuple might be marked either XMAX_INVALID or XMAX_COMMITTED +
7265 : * LOCKED. Normalize to INVALID just to be sure no one gets confused.
7266 : * Also get rid of the HEAP_KEYS_UPDATED bit.
7267 : */
7268 1879 : frz->t_infomask &= ~HEAP_XMAX_BITS;
7269 1879 : frz->t_infomask |= HEAP_XMAX_INVALID;
7270 1879 : frz->t_infomask2 &= ~HEAP_HOT_UPDATED;
7271 1879 : frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
7272 : }
7273 :
7274 : /*
7275 : * Determine if this tuple is already totally frozen, or will become
7276 : * totally frozen (provided caller executes freeze plans for the page)
7277 : */
7278 23117595 : *totally_frozen = ((freeze_xmin || xmin_already_frozen) &&
7279 11190776 : (freeze_xmax || xmax_already_frozen));
7280 :
7281 11926819 : if (!pagefrz->freeze_required && !(xmin_already_frozen &&
7282 : xmax_already_frozen))
7283 : {
7284 : /*
7285 : * So far no previous tuple from the page made freezing mandatory.
7286 : * Does this tuple force caller to freeze the entire page?
7287 : */
7288 7593850 : pagefrz->freeze_required =
7289 7593850 : heap_tuple_should_freeze(tuple, cutoffs,
7290 : &pagefrz->NoFreezePageRelfrozenXid,
7291 : &pagefrz->NoFreezePageRelminMxid);
7292 : }
7293 :
7294 : /* Tell caller if this tuple has a usable freeze plan set in *frz */
7295 11926819 : return freeze_xmin || replace_xvac || replace_xmax || freeze_xmax;
7296 : }
7297 :
7298 : /*
7299 : * Perform xmin/xmax XID status sanity checks before actually executing freeze
7300 : * plans.
7301 : *
7302 : * heap_prepare_freeze_tuple doesn't perform these checks directly because
7303 : * pg_xact lookups are relatively expensive. They shouldn't be repeated by
7304 : * successive VACUUMs that each decide against freezing the same page.
7305 : */
7306 : void
7307 25741 : heap_pre_freeze_checks(Buffer buffer,
7308 : HeapTupleFreeze *tuples, int ntuples)
7309 : {
7310 25741 : Page page = BufferGetPage(buffer);
7311 :
7312 1226959 : for (int i = 0; i < ntuples; i++)
7313 : {
7314 1201218 : HeapTupleFreeze *frz = tuples + i;
7315 1201218 : ItemId itemid = PageGetItemId(page, frz->offset);
7316 : HeapTupleHeader htup;
7317 :
7318 1201218 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
7319 :
7320 : /* Deliberately avoid relying on tuple hint bits here */
7321 1201218 : if (frz->checkflags & HEAP_FREEZE_CHECK_XMIN_COMMITTED)
7322 : {
7323 1201217 : TransactionId xmin = HeapTupleHeaderGetRawXmin(htup);
7324 :
7325 : Assert(!HeapTupleHeaderXminFrozen(htup));
7326 1201217 : if (unlikely(!TransactionIdDidCommit(xmin)))
7327 0 : ereport(ERROR,
7328 : (errcode(ERRCODE_DATA_CORRUPTED),
7329 : errmsg_internal("uncommitted xmin %u needs to be frozen",
7330 : xmin)));
7331 : }
7332 :
7333 : /*
7334 : * TransactionIdDidAbort won't work reliably in the presence of XIDs
7335 : * left behind by transactions that were in progress during a crash,
7336 : * so we can only check that xmax didn't commit
7337 : */
7338 1201218 : if (frz->checkflags & HEAP_FREEZE_CHECK_XMAX_ABORTED)
7339 : {
7340 246 : TransactionId xmax = HeapTupleHeaderGetRawXmax(htup);
7341 :
7342 : Assert(TransactionIdIsNormal(xmax));
7343 246 : if (unlikely(TransactionIdDidCommit(xmax)))
7344 0 : ereport(ERROR,
7345 : (errcode(ERRCODE_DATA_CORRUPTED),
7346 : errmsg_internal("cannot freeze committed xmax %u",
7347 : xmax)));
7348 : }
7349 : }
7350 25741 : }
7351 :
7352 : /*
7353 : * Helper which executes freezing of one or more heap tuples on a page on
7354 : * behalf of caller. Caller passes an array of tuple plans from
7355 : * heap_prepare_freeze_tuple. Caller must set 'offset' in each plan for us.
7356 : * Must be called in a critical section that also marks the buffer dirty and,
7357 : * if needed, emits WAL.
7358 : */
7359 : void
7360 25741 : heap_freeze_prepared_tuples(Buffer buffer, HeapTupleFreeze *tuples, int ntuples)
7361 : {
7362 25741 : Page page = BufferGetPage(buffer);
7363 :
7364 1226959 : for (int i = 0; i < ntuples; i++)
7365 : {
7366 1201218 : HeapTupleFreeze *frz = tuples + i;
7367 1201218 : ItemId itemid = PageGetItemId(page, frz->offset);
7368 : HeapTupleHeader htup;
7369 :
7370 1201218 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
7371 1201218 : heap_execute_freeze_tuple(htup, frz);
7372 : }
7373 25741 : }
7374 :
7375 : /*
7376 : * heap_freeze_tuple
7377 : * Freeze tuple in place, without WAL logging.
7378 : *
7379 : * Useful for callers like CLUSTER that perform their own WAL logging.
7380 : */
7381 : bool
7382 456141 : heap_freeze_tuple(HeapTupleHeader tuple,
7383 : TransactionId relfrozenxid, TransactionId relminmxid,
7384 : TransactionId FreezeLimit, TransactionId MultiXactCutoff)
7385 : {
7386 : HeapTupleFreeze frz;
7387 : bool do_freeze;
7388 : bool totally_frozen;
7389 : struct VacuumCutoffs cutoffs;
7390 : HeapPageFreeze pagefrz;
7391 :
7392 456141 : cutoffs.relfrozenxid = relfrozenxid;
7393 456141 : cutoffs.relminmxid = relminmxid;
7394 456141 : cutoffs.OldestXmin = FreezeLimit;
7395 456141 : cutoffs.OldestMxact = MultiXactCutoff;
7396 456141 : cutoffs.FreezeLimit = FreezeLimit;
7397 456141 : cutoffs.MultiXactCutoff = MultiXactCutoff;
7398 :
7399 456141 : pagefrz.freeze_required = true;
7400 456141 : pagefrz.FreezePageRelfrozenXid = FreezeLimit;
7401 456141 : pagefrz.FreezePageRelminMxid = MultiXactCutoff;
7402 456141 : pagefrz.FreezePageConflictXid = InvalidTransactionId;
7403 456141 : pagefrz.NoFreezePageRelfrozenXid = FreezeLimit;
7404 456141 : pagefrz.NoFreezePageRelminMxid = MultiXactCutoff;
7405 :
7406 456141 : do_freeze = heap_prepare_freeze_tuple(tuple, &cutoffs,
7407 : &pagefrz, &frz, &totally_frozen);
7408 :
7409 : /*
7410 : * Note that because this is not a WAL-logged operation, we don't need to
7411 : * fill in the offset in the freeze record.
7412 : */
7413 :
7414 456141 : if (do_freeze)
7415 378170 : heap_execute_freeze_tuple(tuple, &frz);
7416 456141 : return do_freeze;
7417 : }
7418 :
7419 : /*
7420 : * For a given MultiXactId, return the hint bits that should be set in the
7421 : * tuple's infomask.
7422 : *
7423 : * Normally this should be called for a multixact that was just created, and
7424 : * so is on our local cache, so the GetMembers call is fast.
7425 : */
7426 : static void
7427 76883 : GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
7428 : uint16 *new_infomask2)
7429 : {
7430 : int nmembers;
7431 : MultiXactMember *members;
7432 : int i;
7433 76883 : uint16 bits = HEAP_XMAX_IS_MULTI;
7434 76883 : uint16 bits2 = 0;
7435 76883 : bool has_update = false;
7436 76883 : LockTupleMode strongest = LockTupleKeyShare;
7437 :
7438 : /*
7439 : * We only use this in multis we just created, so they cannot be values
7440 : * pre-pg_upgrade.
7441 : */
7442 76883 : nmembers = GetMultiXactIdMembers(multi, &members, false, false);
7443 :
7444 1472858 : for (i = 0; i < nmembers; i++)
7445 : {
7446 : LockTupleMode mode;
7447 :
7448 : /*
7449 : * Remember the strongest lock mode held by any member of the
7450 : * multixact.
7451 : */
7452 1395975 : mode = TUPLOCK_from_mxstatus(members[i].status);
7453 1395975 : if (mode > strongest)
7454 2929 : strongest = mode;
7455 :
7456 : /* See what other bits we need */
7457 1395975 : switch (members[i].status)
7458 : {
7459 1393553 : case MultiXactStatusForKeyShare:
7460 : case MultiXactStatusForShare:
7461 : case MultiXactStatusForNoKeyUpdate:
7462 1393553 : break;
7463 :
7464 53 : case MultiXactStatusForUpdate:
7465 53 : bits2 |= HEAP_KEYS_UPDATED;
7466 53 : break;
7467 :
7468 2359 : case MultiXactStatusNoKeyUpdate:
7469 2359 : has_update = true;
7470 2359 : break;
7471 :
7472 10 : case MultiXactStatusUpdate:
7473 10 : bits2 |= HEAP_KEYS_UPDATED;
7474 10 : has_update = true;
7475 10 : break;
7476 : }
7477 : }
7478 :
7479 76883 : if (strongest == LockTupleExclusive ||
7480 : strongest == LockTupleNoKeyExclusive)
7481 2450 : bits |= HEAP_XMAX_EXCL_LOCK;
7482 74433 : else if (strongest == LockTupleShare)
7483 476 : bits |= HEAP_XMAX_SHR_LOCK;
7484 73957 : else if (strongest == LockTupleKeyShare)
7485 73957 : bits |= HEAP_XMAX_KEYSHR_LOCK;
7486 :
7487 76883 : if (!has_update)
7488 74514 : bits |= HEAP_XMAX_LOCK_ONLY;
7489 :
7490 76883 : if (nmembers > 0)
7491 76883 : pfree(members);
7492 :
7493 76883 : *new_infomask = bits;
7494 76883 : *new_infomask2 = bits2;
7495 76883 : }
7496 :
7497 : /*
7498 : * MultiXactIdGetUpdateXid
7499 : *
7500 : * Given a multixact Xmax and corresponding infomask, which does not have the
7501 : * HEAP_XMAX_LOCK_ONLY bit set, obtain and return the Xid of the updating
7502 : * transaction.
7503 : *
7504 : * Caller is expected to check the status of the updating transaction, if
7505 : * necessary.
7506 : */
7507 : static TransactionId
7508 162051 : MultiXactIdGetUpdateXid(TransactionId xmax, uint16 t_infomask)
7509 : {
7510 162051 : TransactionId update_xact = InvalidTransactionId;
7511 : MultiXactMember *members;
7512 : int nmembers;
7513 :
7514 : Assert(!(t_infomask & HEAP_XMAX_LOCK_ONLY));
7515 : Assert(t_infomask & HEAP_XMAX_IS_MULTI);
7516 :
7517 : /*
7518 : * Since we know the LOCK_ONLY bit is not set, this cannot be a multi from
7519 : * pre-pg_upgrade.
7520 : */
7521 162051 : nmembers = GetMultiXactIdMembers(xmax, &members, false, false);
7522 :
7523 162051 : if (nmembers > 0)
7524 : {
7525 : int i;
7526 :
7527 245869 : for (i = 0; i < nmembers; i++)
7528 : {
7529 : /* Ignore lockers */
7530 245869 : if (!ISUPDATE_from_mxstatus(members[i].status))
7531 83818 : continue;
7532 :
7533 : /* there can be at most one updater */
7534 : Assert(update_xact == InvalidTransactionId);
7535 162051 : update_xact = members[i].xid;
7536 : #ifndef USE_ASSERT_CHECKING
7537 :
7538 : /*
7539 : * in an assert-enabled build, walk the whole array to ensure
7540 : * there's no other updater.
7541 : */
7542 162051 : break;
7543 : #endif
7544 : }
7545 :
7546 162051 : pfree(members);
7547 : }
7548 :
7549 162051 : return update_xact;
7550 : }
7551 :
7552 : /*
7553 : * HeapTupleGetUpdateXid
7554 : * As above, but use a HeapTupleHeader
7555 : *
7556 : * See also HeapTupleHeaderGetUpdateXid, which can be used without previously
7557 : * checking the hint bits.
7558 : */
7559 : TransactionId
7560 159912 : HeapTupleGetUpdateXid(const HeapTupleHeaderData *tup)
7561 : {
7562 159912 : return MultiXactIdGetUpdateXid(HeapTupleHeaderGetRawXmax(tup),
7563 159912 : tup->t_infomask);
7564 : }
7565 :
7566 : /*
7567 : * Does the given multixact conflict with the current transaction grabbing a
7568 : * tuple lock of the given strength?
7569 : *
7570 : * The passed infomask pairs up with the given multixact in the tuple header.
7571 : *
7572 : * If current_is_member is not NULL, it is set to 'true' if the current
7573 : * transaction is a member of the given multixact.
7574 : */
7575 : static bool
7576 218 : DoesMultiXactIdConflict(MultiXactId multi, uint16 infomask,
7577 : LockTupleMode lockmode, bool *current_is_member)
7578 : {
7579 : int nmembers;
7580 : MultiXactMember *members;
7581 218 : bool result = false;
7582 218 : LOCKMODE wanted = tupleLockExtraInfo[lockmode].hwlock;
7583 :
7584 218 : if (HEAP_LOCKED_UPGRADED(infomask))
7585 0 : return false;
7586 :
7587 218 : nmembers = GetMultiXactIdMembers(multi, &members, false,
7588 218 : HEAP_XMAX_IS_LOCKED_ONLY(infomask));
7589 218 : if (nmembers >= 0)
7590 : {
7591 : int i;
7592 :
7593 2682 : for (i = 0; i < nmembers; i++)
7594 : {
7595 : TransactionId memxid;
7596 : LOCKMODE memlockmode;
7597 :
7598 2471 : if (result && (current_is_member == NULL || *current_is_member))
7599 : break;
7600 :
7601 2464 : memlockmode = LOCKMODE_from_mxstatus(members[i].status);
7602 :
7603 : /* ignore members from current xact (but track their presence) */
7604 2464 : memxid = members[i].xid;
7605 2464 : if (TransactionIdIsCurrentTransactionId(memxid))
7606 : {
7607 92 : if (current_is_member != NULL)
7608 78 : *current_is_member = true;
7609 92 : continue;
7610 : }
7611 2372 : else if (result)
7612 8 : continue;
7613 :
7614 : /* ignore members that don't conflict with the lock we want */
7615 2364 : if (!DoLockModesConflict(memlockmode, wanted))
7616 2325 : continue;
7617 :
7618 39 : if (ISUPDATE_from_mxstatus(members[i].status))
7619 : {
7620 : /* ignore aborted updaters */
7621 17 : if (TransactionIdDidAbort(memxid))
7622 1 : continue;
7623 : }
7624 : else
7625 : {
7626 : /* ignore lockers-only that are no longer in progress */
7627 22 : if (!TransactionIdIsInProgress(memxid))
7628 7 : continue;
7629 : }
7630 :
7631 : /*
7632 : * Whatever remains are either live lockers that conflict with our
7633 : * wanted lock, and updaters that are not aborted. Those conflict
7634 : * with what we want. Set up to return true, but keep going to
7635 : * look for the current transaction among the multixact members,
7636 : * if needed.
7637 : */
7638 31 : result = true;
7639 : }
7640 218 : pfree(members);
7641 : }
7642 :
7643 218 : return result;
7644 : }
7645 :
7646 : /*
7647 : * Do_MultiXactIdWait
7648 : * Actual implementation for the two functions below.
7649 : *
7650 : * 'multi', 'status' and 'infomask' indicate what to sleep on (the status is
7651 : * needed to ensure we only sleep on conflicting members, and the infomask is
7652 : * used to optimize multixact access in case it's a lock-only multi); 'nowait'
7653 : * indicates whether to use conditional lock acquisition, to allow callers to
7654 : * fail if lock is unavailable. 'rel', 'ctid' and 'oper' are used to set up
7655 : * context information for error messages. 'remaining', if not NULL, receives
7656 : * the number of members that are still running, including any (non-aborted)
7657 : * subtransactions of our own transaction. 'logLockFailure' indicates whether
7658 : * to log details when a lock acquisition fails with 'nowait' enabled.
7659 : *
7660 : * We do this by sleeping on each member using XactLockTableWait. Any
7661 : * members that belong to the current backend are *not* waited for, however;
7662 : * this would not merely be useless but would lead to Assert failure inside
7663 : * XactLockTableWait. By the time this returns, it is certain that all
7664 : * transactions *of other backends* that were members of the MultiXactId
7665 : * that conflict with the requested status are dead (and no new ones can have
7666 : * been added, since it is not legal to add members to an existing
7667 : * MultiXactId).
7668 : *
7669 : * But by the time we finish sleeping, someone else may have changed the Xmax
7670 : * of the containing tuple, so the caller needs to iterate on us somehow.
7671 : *
7672 : * Note that in case we return false, the number of remaining members is
7673 : * not to be trusted.
7674 : */
7675 : static bool
7676 61 : Do_MultiXactIdWait(MultiXactId multi, MultiXactStatus status,
7677 : uint16 infomask, bool nowait,
7678 : Relation rel, const ItemPointerData *ctid, XLTW_Oper oper,
7679 : int *remaining, bool logLockFailure)
7680 : {
7681 61 : bool result = true;
7682 : MultiXactMember *members;
7683 : int nmembers;
7684 61 : int remain = 0;
7685 :
7686 : /* for pre-pg_upgrade tuples, no need to sleep at all */
7687 61 : nmembers = HEAP_LOCKED_UPGRADED(infomask) ? -1 :
7688 61 : GetMultiXactIdMembers(multi, &members, false,
7689 61 : HEAP_XMAX_IS_LOCKED_ONLY(infomask));
7690 :
7691 61 : if (nmembers >= 0)
7692 : {
7693 : int i;
7694 :
7695 192 : for (i = 0; i < nmembers; i++)
7696 : {
7697 137 : TransactionId memxid = members[i].xid;
7698 137 : MultiXactStatus memstatus = members[i].status;
7699 :
7700 137 : if (TransactionIdIsCurrentTransactionId(memxid))
7701 : {
7702 25 : remain++;
7703 25 : continue;
7704 : }
7705 :
7706 112 : if (!DoLockModesConflict(LOCKMODE_from_mxstatus(memstatus),
7707 112 : LOCKMODE_from_mxstatus(status)))
7708 : {
7709 22 : if (remaining && TransactionIdIsInProgress(memxid))
7710 8 : remain++;
7711 22 : continue;
7712 : }
7713 :
7714 : /*
7715 : * This member conflicts with our multi, so we have to sleep (or
7716 : * return failure, if asked to avoid waiting.)
7717 : *
7718 : * Note that we don't set up an error context callback ourselves,
7719 : * but instead we pass the info down to XactLockTableWait. This
7720 : * might seem a bit wasteful because the context is set up and
7721 : * tore down for each member of the multixact, but in reality it
7722 : * should be barely noticeable, and it avoids duplicate code.
7723 : */
7724 90 : if (nowait)
7725 : {
7726 6 : result = ConditionalXactLockTableWait(memxid, logLockFailure);
7727 6 : if (!result)
7728 6 : break;
7729 : }
7730 : else
7731 84 : XactLockTableWait(memxid, rel, ctid, oper);
7732 : }
7733 :
7734 61 : pfree(members);
7735 : }
7736 :
7737 61 : if (remaining)
7738 10 : *remaining = remain;
7739 :
7740 61 : return result;
7741 : }
7742 :
7743 : /*
7744 : * MultiXactIdWait
7745 : * Sleep on a MultiXactId.
7746 : *
7747 : * By the time we finish sleeping, someone else may have changed the Xmax
7748 : * of the containing tuple, so the caller needs to iterate on us somehow.
7749 : *
7750 : * We return (in *remaining, if not NULL) the number of members that are still
7751 : * running, including any (non-aborted) subtransactions of our own transaction.
7752 : */
7753 : static void
7754 55 : MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
7755 : Relation rel, const ItemPointerData *ctid, XLTW_Oper oper,
7756 : int *remaining)
7757 : {
7758 55 : (void) Do_MultiXactIdWait(multi, status, infomask, false,
7759 : rel, ctid, oper, remaining, false);
7760 55 : }
7761 :
7762 : /*
7763 : * ConditionalMultiXactIdWait
7764 : * As above, but only lock if we can get the lock without blocking.
7765 : *
7766 : * By the time we finish sleeping, someone else may have changed the Xmax
7767 : * of the containing tuple, so the caller needs to iterate on us somehow.
7768 : *
7769 : * If the multixact is now all gone, return true. Returns false if some
7770 : * transactions might still be running.
7771 : *
7772 : * We return (in *remaining, if not NULL) the number of members that are still
7773 : * running, including any (non-aborted) subtransactions of our own transaction.
7774 : */
7775 : static bool
7776 6 : ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
7777 : uint16 infomask, Relation rel, int *remaining,
7778 : bool logLockFailure)
7779 : {
7780 6 : return Do_MultiXactIdWait(multi, status, infomask, true,
7781 : rel, NULL, XLTW_None, remaining, logLockFailure);
7782 : }
7783 :
7784 : /*
7785 : * heap_tuple_needs_eventual_freeze
7786 : *
7787 : * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
7788 : * will eventually require freezing (if tuple isn't removed by pruning first).
7789 : */
7790 : bool
7791 114574 : heap_tuple_needs_eventual_freeze(HeapTupleHeader tuple)
7792 : {
7793 : TransactionId xid;
7794 :
7795 : /*
7796 : * If xmin is a normal transaction ID, this tuple is definitely not
7797 : * frozen.
7798 : */
7799 114574 : xid = HeapTupleHeaderGetXmin(tuple);
7800 114574 : if (TransactionIdIsNormal(xid))
7801 3008 : return true;
7802 :
7803 : /*
7804 : * If xmax is a valid xact or multixact, this tuple is also not frozen.
7805 : */
7806 111566 : if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
7807 : {
7808 : MultiXactId multi;
7809 :
7810 0 : multi = HeapTupleHeaderGetRawXmax(tuple);
7811 0 : if (MultiXactIdIsValid(multi))
7812 0 : return true;
7813 : }
7814 : else
7815 : {
7816 111566 : xid = HeapTupleHeaderGetRawXmax(tuple);
7817 111566 : if (TransactionIdIsNormal(xid))
7818 7 : return true;
7819 : }
7820 :
7821 111559 : if (tuple->t_infomask & HEAP_MOVED)
7822 : {
7823 0 : xid = HeapTupleHeaderGetXvac(tuple);
7824 0 : if (TransactionIdIsNormal(xid))
7825 0 : return true;
7826 : }
7827 :
7828 111559 : return false;
7829 : }
7830 :
7831 : /*
7832 : * heap_tuple_should_freeze
7833 : *
7834 : * Return value indicates if heap_prepare_freeze_tuple sibling function would
7835 : * (or should) force freezing of the heap page that contains caller's tuple.
7836 : * Tuple header XIDs/MXIDs < FreezeLimit/MultiXactCutoff trigger freezing.
7837 : * This includes (xmin, xmax, xvac) fields, as well as MultiXact member XIDs.
7838 : *
7839 : * The *NoFreezePageRelfrozenXid and *NoFreezePageRelminMxid input/output
7840 : * arguments help VACUUM track the oldest extant XID/MXID remaining in rel.
7841 : * Our working assumption is that caller won't decide to freeze this tuple.
7842 : * It's up to caller to only ratchet back its own top-level trackers after the
7843 : * point that it fully commits to not freezing the tuple/page in question.
7844 : */
7845 : bool
7846 7597247 : heap_tuple_should_freeze(HeapTupleHeader tuple,
7847 : const struct VacuumCutoffs *cutoffs,
7848 : TransactionId *NoFreezePageRelfrozenXid,
7849 : MultiXactId *NoFreezePageRelminMxid)
7850 : {
7851 : TransactionId xid;
7852 : MultiXactId multi;
7853 7597247 : bool freeze = false;
7854 :
7855 : /* First deal with xmin */
7856 7597247 : xid = HeapTupleHeaderGetXmin(tuple);
7857 7597247 : if (TransactionIdIsNormal(xid))
7858 : {
7859 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7860 2508207 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7861 25792 : *NoFreezePageRelfrozenXid = xid;
7862 2508207 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
7863 23652 : freeze = true;
7864 : }
7865 :
7866 : /* Now deal with xmax */
7867 7597247 : xid = InvalidTransactionId;
7868 7597247 : multi = InvalidMultiXactId;
7869 7597247 : if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
7870 2 : multi = HeapTupleHeaderGetRawXmax(tuple);
7871 : else
7872 7597245 : xid = HeapTupleHeaderGetRawXmax(tuple);
7873 :
7874 7597247 : if (TransactionIdIsNormal(xid))
7875 : {
7876 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7877 : /* xmax is a non-permanent XID */
7878 5415054 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7879 1 : *NoFreezePageRelfrozenXid = xid;
7880 5415054 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
7881 17 : freeze = true;
7882 : }
7883 2182193 : else if (!MultiXactIdIsValid(multi))
7884 : {
7885 : /* xmax is a permanent XID or invalid MultiXactId/XID */
7886 : }
7887 2 : else if (HEAP_LOCKED_UPGRADED(tuple->t_infomask))
7888 : {
7889 : /* xmax is a pg_upgrade'd MultiXact, which can't have updater XID */
7890 0 : if (MultiXactIdPrecedes(multi, *NoFreezePageRelminMxid))
7891 0 : *NoFreezePageRelminMxid = multi;
7892 : /* heap_prepare_freeze_tuple always freezes pg_upgrade'd xmax */
7893 0 : freeze = true;
7894 : }
7895 : else
7896 : {
7897 : /* xmax is a MultiXactId that may have an updater XID */
7898 : MultiXactMember *members;
7899 : int nmembers;
7900 :
7901 : Assert(MultiXactIdPrecedesOrEquals(cutoffs->relminmxid, multi));
7902 2 : if (MultiXactIdPrecedes(multi, *NoFreezePageRelminMxid))
7903 2 : *NoFreezePageRelminMxid = multi;
7904 2 : if (MultiXactIdPrecedes(multi, cutoffs->MultiXactCutoff))
7905 2 : freeze = true;
7906 :
7907 : /* need to check whether any member of the mxact is old */
7908 2 : nmembers = GetMultiXactIdMembers(multi, &members, false,
7909 2 : HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask));
7910 :
7911 5 : for (int i = 0; i < nmembers; i++)
7912 : {
7913 3 : xid = members[i].xid;
7914 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7915 3 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7916 0 : *NoFreezePageRelfrozenXid = xid;
7917 3 : if (TransactionIdPrecedes(xid, cutoffs->FreezeLimit))
7918 0 : freeze = true;
7919 : }
7920 2 : if (nmembers > 0)
7921 1 : pfree(members);
7922 : }
7923 :
7924 7597247 : if (tuple->t_infomask & HEAP_MOVED)
7925 : {
7926 0 : xid = HeapTupleHeaderGetXvac(tuple);
7927 0 : if (TransactionIdIsNormal(xid))
7928 : {
7929 : Assert(TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid, xid));
7930 0 : if (TransactionIdPrecedes(xid, *NoFreezePageRelfrozenXid))
7931 0 : *NoFreezePageRelfrozenXid = xid;
7932 : /* heap_prepare_freeze_tuple forces xvac freezing */
7933 0 : freeze = true;
7934 : }
7935 : }
7936 :
7937 7597247 : return freeze;
7938 : }
7939 :
7940 : /*
7941 : * Maintain snapshotConflictHorizon for caller by ratcheting forward its value
7942 : * using any committed XIDs contained in 'tuple', an obsolescent heap tuple
7943 : * that caller is in the process of physically removing, e.g. via HOT pruning
7944 : * or index deletion.
7945 : *
7946 : * Caller must initialize its value to InvalidTransactionId, which is
7947 : * generally interpreted as "definitely no need for a recovery conflict".
7948 : * Final value must reflect all heap tuples that caller will physically remove
7949 : * (or remove TID references to) via its ongoing pruning/deletion operation.
7950 : * ResolveRecoveryConflictWithSnapshot() is passed the final value (taken from
7951 : * caller's WAL record) by REDO routine when it replays caller's operation.
7952 : */
7953 : void
7954 3981609 : HeapTupleHeaderAdvanceConflictHorizon(HeapTupleHeader tuple,
7955 : TransactionId *snapshotConflictHorizon)
7956 : {
7957 3981609 : TransactionId xmin = HeapTupleHeaderGetXmin(tuple);
7958 3981609 : TransactionId xmax = HeapTupleHeaderGetUpdateXid(tuple);
7959 3981609 : TransactionId xvac = HeapTupleHeaderGetXvac(tuple);
7960 :
7961 3981609 : if (tuple->t_infomask & HEAP_MOVED)
7962 : {
7963 0 : if (TransactionIdPrecedes(*snapshotConflictHorizon, xvac))
7964 0 : *snapshotConflictHorizon = xvac;
7965 : }
7966 :
7967 : /*
7968 : * Ignore tuples inserted by an aborted transaction or if the tuple was
7969 : * updated/deleted by the inserting transaction.
7970 : *
7971 : * Look for a committed hint bit, or if no xmin bit is set, check clog.
7972 : */
7973 3981609 : if (HeapTupleHeaderXminCommitted(tuple) ||
7974 131349 : (!HeapTupleHeaderXminInvalid(tuple) && TransactionIdDidCommit(xmin)))
7975 : {
7976 7511759 : if (xmax != xmin &&
7977 3656255 : TransactionIdFollows(xmax, *snapshotConflictHorizon))
7978 130698 : *snapshotConflictHorizon = xmax;
7979 : }
7980 3981609 : }
7981 :
7982 : #ifdef USE_PREFETCH
7983 : /*
7984 : * Helper function for heap_index_delete_tuples. Issues prefetch requests for
7985 : * prefetch_count buffers. The prefetch_state keeps track of all the buffers
7986 : * we can prefetch, and which have already been prefetched; each call to this
7987 : * function picks up where the previous call left off.
7988 : *
7989 : * Note: we expect the deltids array to be sorted in an order that groups TIDs
7990 : * by heap block, with all TIDs for each block appearing together in exactly
7991 : * one group.
7992 : */
7993 : static void
7994 27535 : index_delete_prefetch_buffer(Relation rel,
7995 : IndexDeletePrefetchState *prefetch_state,
7996 : int prefetch_count)
7997 : {
7998 27535 : BlockNumber cur_hblkno = prefetch_state->cur_hblkno;
7999 27535 : int count = 0;
8000 : int i;
8001 27535 : int ndeltids = prefetch_state->ndeltids;
8002 27535 : TM_IndexDelete *deltids = prefetch_state->deltids;
8003 :
8004 27535 : for (i = prefetch_state->next_item;
8005 971354 : i < ndeltids && count < prefetch_count;
8006 943819 : i++)
8007 : {
8008 943819 : ItemPointer htid = &deltids[i].tid;
8009 :
8010 1879198 : if (cur_hblkno == InvalidBlockNumber ||
8011 935379 : ItemPointerGetBlockNumber(htid) != cur_hblkno)
8012 : {
8013 24748 : cur_hblkno = ItemPointerGetBlockNumber(htid);
8014 24748 : PrefetchBuffer(rel, MAIN_FORKNUM, cur_hblkno);
8015 24748 : count++;
8016 : }
8017 : }
8018 :
8019 : /*
8020 : * Save the prefetch position so that next time we can continue from that
8021 : * position.
8022 : */
8023 27535 : prefetch_state->next_item = i;
8024 27535 : prefetch_state->cur_hblkno = cur_hblkno;
8025 27535 : }
8026 : #endif
8027 :
8028 : /*
8029 : * Helper function for heap_index_delete_tuples. Checks for index corruption
8030 : * involving an invalid TID in index AM caller's index page.
8031 : *
8032 : * This is an ideal place for these checks. The index AM must hold a buffer
8033 : * lock on the index page containing the TIDs we examine here, so we don't
8034 : * have to worry about concurrent VACUUMs at all. We can be sure that the
8035 : * index is corrupt when htid points directly to an LP_UNUSED item or
8036 : * heap-only tuple, which is not the case during standard index scans.
8037 : */
8038 : static inline void
8039 783565 : index_delete_check_htid(TM_IndexDeleteOp *delstate,
8040 : Page page, OffsetNumber maxoff,
8041 : const ItemPointerData *htid, TM_IndexStatus *istatus)
8042 : {
8043 783565 : OffsetNumber indexpagehoffnum = ItemPointerGetOffsetNumber(htid);
8044 : ItemId iid;
8045 :
8046 : Assert(OffsetNumberIsValid(istatus->idxoffnum));
8047 :
8048 783565 : if (unlikely(indexpagehoffnum > maxoff))
8049 0 : ereport(ERROR,
8050 : (errcode(ERRCODE_INDEX_CORRUPTED),
8051 : 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\"",
8052 : ItemPointerGetBlockNumber(htid),
8053 : indexpagehoffnum,
8054 : istatus->idxoffnum, delstate->iblknum,
8055 : RelationGetRelationName(delstate->irel))));
8056 :
8057 783565 : iid = PageGetItemId(page, indexpagehoffnum);
8058 783565 : if (unlikely(!ItemIdIsUsed(iid)))
8059 0 : ereport(ERROR,
8060 : (errcode(ERRCODE_INDEX_CORRUPTED),
8061 : errmsg_internal("heap tid from index tuple (%u,%u) points to unused heap page item at offset %u of block %u in index \"%s\"",
8062 : ItemPointerGetBlockNumber(htid),
8063 : indexpagehoffnum,
8064 : istatus->idxoffnum, delstate->iblknum,
8065 : RelationGetRelationName(delstate->irel))));
8066 :
8067 783565 : if (ItemIdHasStorage(iid))
8068 : {
8069 : HeapTupleHeader htup;
8070 :
8071 : Assert(ItemIdIsNormal(iid));
8072 481889 : htup = (HeapTupleHeader) PageGetItem(page, iid);
8073 :
8074 481889 : if (unlikely(HeapTupleHeaderIsHeapOnly(htup)))
8075 0 : ereport(ERROR,
8076 : (errcode(ERRCODE_INDEX_CORRUPTED),
8077 : errmsg_internal("heap tid from index tuple (%u,%u) points to heap-only tuple at offset %u of block %u in index \"%s\"",
8078 : ItemPointerGetBlockNumber(htid),
8079 : indexpagehoffnum,
8080 : istatus->idxoffnum, delstate->iblknum,
8081 : RelationGetRelationName(delstate->irel))));
8082 : }
8083 783565 : }
8084 :
8085 : /*
8086 : * heapam implementation of tableam's index_delete_tuples interface.
8087 : *
8088 : * This helper function is called by index AMs during index tuple deletion.
8089 : * See tableam header comments for an explanation of the interface implemented
8090 : * here and a general theory of operation. Note that each call here is either
8091 : * a simple index deletion call, or a bottom-up index deletion call.
8092 : *
8093 : * It's possible for this to generate a fair amount of I/O, since we may be
8094 : * deleting hundreds of tuples from a single index block. To amortize that
8095 : * cost to some degree, this uses prefetching and combines repeat accesses to
8096 : * the same heap block.
8097 : */
8098 : TransactionId
8099 8440 : heap_index_delete_tuples(Relation rel, TM_IndexDeleteOp *delstate)
8100 : {
8101 : /* Initial assumption is that earlier pruning took care of conflict */
8102 8440 : TransactionId snapshotConflictHorizon = InvalidTransactionId;
8103 8440 : BlockNumber blkno = InvalidBlockNumber;
8104 8440 : Buffer buf = InvalidBuffer;
8105 8440 : Page page = NULL;
8106 8440 : OffsetNumber maxoff = InvalidOffsetNumber;
8107 : TransactionId priorXmax;
8108 : #ifdef USE_PREFETCH
8109 : IndexDeletePrefetchState prefetch_state;
8110 : int prefetch_distance;
8111 : #endif
8112 : SnapshotData SnapshotNonVacuumable;
8113 8440 : int finalndeltids = 0,
8114 8440 : nblocksaccessed = 0;
8115 :
8116 : /* State that's only used in bottom-up index deletion case */
8117 8440 : int nblocksfavorable = 0;
8118 8440 : int curtargetfreespace = delstate->bottomupfreespace,
8119 8440 : lastfreespace = 0,
8120 8440 : actualfreespace = 0;
8121 8440 : bool bottomup_final_block = false;
8122 :
8123 8440 : InitNonVacuumableSnapshot(SnapshotNonVacuumable, GlobalVisTestFor(rel));
8124 :
8125 : /* Sort caller's deltids array by TID for further processing */
8126 8440 : index_delete_sort(delstate);
8127 :
8128 : /*
8129 : * Bottom-up case: resort deltids array in an order attuned to where the
8130 : * greatest number of promising TIDs are to be found, and determine how
8131 : * many blocks from the start of sorted array should be considered
8132 : * favorable. This will also shrink the deltids array in order to
8133 : * eliminate completely unfavorable blocks up front.
8134 : */
8135 8440 : if (delstate->bottomup)
8136 2903 : nblocksfavorable = bottomup_sort_and_shrink(delstate);
8137 :
8138 : #ifdef USE_PREFETCH
8139 : /* Initialize prefetch state. */
8140 8440 : prefetch_state.cur_hblkno = InvalidBlockNumber;
8141 8440 : prefetch_state.next_item = 0;
8142 8440 : prefetch_state.ndeltids = delstate->ndeltids;
8143 8440 : prefetch_state.deltids = delstate->deltids;
8144 :
8145 : /*
8146 : * Determine the prefetch distance that we will attempt to maintain.
8147 : *
8148 : * Since the caller holds a buffer lock somewhere in rel, we'd better make
8149 : * sure that isn't a catalog relation before we call code that does
8150 : * syscache lookups, to avoid risk of deadlock.
8151 : */
8152 8440 : if (IsCatalogRelation(rel))
8153 6111 : prefetch_distance = maintenance_io_concurrency;
8154 : else
8155 : prefetch_distance =
8156 2329 : get_tablespace_maintenance_io_concurrency(rel->rd_rel->reltablespace);
8157 :
8158 : /* Cap initial prefetch distance for bottom-up deletion caller */
8159 8440 : if (delstate->bottomup)
8160 : {
8161 : Assert(nblocksfavorable >= 1);
8162 : Assert(nblocksfavorable <= BOTTOMUP_MAX_NBLOCKS);
8163 2903 : prefetch_distance = Min(prefetch_distance, nblocksfavorable);
8164 : }
8165 :
8166 : /* Start prefetching. */
8167 8440 : index_delete_prefetch_buffer(rel, &prefetch_state, prefetch_distance);
8168 : #endif
8169 :
8170 : /* Iterate over deltids, determine which to delete, check their horizon */
8171 : Assert(delstate->ndeltids > 0);
8172 792005 : for (int i = 0; i < delstate->ndeltids; i++)
8173 : {
8174 786468 : TM_IndexDelete *ideltid = &delstate->deltids[i];
8175 786468 : TM_IndexStatus *istatus = delstate->status + ideltid->id;
8176 786468 : ItemPointer htid = &ideltid->tid;
8177 : OffsetNumber offnum;
8178 :
8179 : /*
8180 : * Read buffer, and perform required extra steps each time a new block
8181 : * is encountered. Avoid refetching if it's the same block as the one
8182 : * from the last htid.
8183 : */
8184 1564496 : if (blkno == InvalidBlockNumber ||
8185 778028 : ItemPointerGetBlockNumber(htid) != blkno)
8186 : {
8187 : /*
8188 : * Consider giving up early for bottom-up index deletion caller
8189 : * first. (Only prefetch next-next block afterwards, when it
8190 : * becomes clear that we're at least going to access the next
8191 : * block in line.)
8192 : *
8193 : * Sometimes the first block frees so much space for bottom-up
8194 : * caller that the deletion process can end without accessing any
8195 : * more blocks. It is usually necessary to access 2 or 3 blocks
8196 : * per bottom-up deletion operation, though.
8197 : */
8198 21998 : if (delstate->bottomup)
8199 : {
8200 : /*
8201 : * We often allow caller to delete a few additional items
8202 : * whose entries we reached after the point that space target
8203 : * from caller was satisfied. The cost of accessing the page
8204 : * was already paid at that point, so it made sense to finish
8205 : * it off. When that happened, we finalize everything here
8206 : * (by finishing off the whole bottom-up deletion operation
8207 : * without needlessly paying the cost of accessing any more
8208 : * blocks).
8209 : */
8210 6203 : if (bottomup_final_block)
8211 142 : break;
8212 :
8213 : /*
8214 : * Give up when we didn't enable our caller to free any
8215 : * additional space as a result of processing the page that we
8216 : * just finished up with. This rule is the main way in which
8217 : * we keep the cost of bottom-up deletion under control.
8218 : */
8219 6061 : if (nblocksaccessed >= 1 && actualfreespace == lastfreespace)
8220 2761 : break;
8221 3300 : lastfreespace = actualfreespace; /* for next time */
8222 :
8223 : /*
8224 : * Deletion operation (which is bottom-up) will definitely
8225 : * access the next block in line. Prepare for that now.
8226 : *
8227 : * Decay target free space so that we don't hang on for too
8228 : * long with a marginal case. (Space target is only truly
8229 : * helpful when it allows us to recognize that we don't need
8230 : * to access more than 1 or 2 blocks to satisfy caller due to
8231 : * agreeable workload characteristics.)
8232 : *
8233 : * We are a bit more patient when we encounter contiguous
8234 : * blocks, though: these are treated as favorable blocks. The
8235 : * decay process is only applied when the next block in line
8236 : * is not a favorable/contiguous block. This is not an
8237 : * exception to the general rule; we still insist on finding
8238 : * at least one deletable item per block accessed. See
8239 : * bottomup_nblocksfavorable() for full details of the theory
8240 : * behind favorable blocks and heap block locality in general.
8241 : *
8242 : * Note: The first block in line is always treated as a
8243 : * favorable block, so the earliest possible point that the
8244 : * decay can be applied is just before we access the second
8245 : * block in line. The Assert() verifies this for us.
8246 : */
8247 : Assert(nblocksaccessed > 0 || nblocksfavorable > 0);
8248 3300 : if (nblocksfavorable > 0)
8249 3074 : nblocksfavorable--;
8250 : else
8251 226 : curtargetfreespace /= 2;
8252 : }
8253 :
8254 : /* release old buffer */
8255 19095 : if (BufferIsValid(buf))
8256 10655 : UnlockReleaseBuffer(buf);
8257 :
8258 19095 : blkno = ItemPointerGetBlockNumber(htid);
8259 19095 : buf = ReadBuffer(rel, blkno);
8260 19095 : nblocksaccessed++;
8261 : Assert(!delstate->bottomup ||
8262 : nblocksaccessed <= BOTTOMUP_MAX_NBLOCKS);
8263 :
8264 : #ifdef USE_PREFETCH
8265 :
8266 : /*
8267 : * To maintain the prefetch distance, prefetch one more page for
8268 : * each page we read.
8269 : */
8270 19095 : index_delete_prefetch_buffer(rel, &prefetch_state, 1);
8271 : #endif
8272 :
8273 19095 : LockBuffer(buf, BUFFER_LOCK_SHARE);
8274 :
8275 19095 : page = BufferGetPage(buf);
8276 19095 : maxoff = PageGetMaxOffsetNumber(page);
8277 : }
8278 :
8279 : /*
8280 : * In passing, detect index corruption involving an index page with a
8281 : * TID that points to a location in the heap that couldn't possibly be
8282 : * correct. We only do this with actual TIDs from caller's index page
8283 : * (not items reached by traversing through a HOT chain).
8284 : */
8285 783565 : index_delete_check_htid(delstate, page, maxoff, htid, istatus);
8286 :
8287 783565 : if (istatus->knowndeletable)
8288 : Assert(!delstate->bottomup && !istatus->promising);
8289 : else
8290 : {
8291 614914 : ItemPointerData tmp = *htid;
8292 : HeapTupleData heapTuple;
8293 :
8294 : /* Are any tuples from this HOT chain non-vacuumable? */
8295 614914 : if (heap_hot_search_buffer(&tmp, rel, buf, &SnapshotNonVacuumable,
8296 : &heapTuple, NULL, true))
8297 371842 : continue; /* can't delete entry */
8298 :
8299 : /* Caller will delete, since whole HOT chain is vacuumable */
8300 243072 : istatus->knowndeletable = true;
8301 :
8302 : /* Maintain index free space info for bottom-up deletion case */
8303 243072 : if (delstate->bottomup)
8304 : {
8305 : Assert(istatus->freespace > 0);
8306 10041 : actualfreespace += istatus->freespace;
8307 10041 : if (actualfreespace >= curtargetfreespace)
8308 2409 : bottomup_final_block = true;
8309 : }
8310 : }
8311 :
8312 : /*
8313 : * Maintain snapshotConflictHorizon value for deletion operation as a
8314 : * whole by advancing current value using heap tuple headers. This is
8315 : * loosely based on the logic for pruning a HOT chain.
8316 : */
8317 411723 : offnum = ItemPointerGetOffsetNumber(htid);
8318 411723 : priorXmax = InvalidTransactionId; /* cannot check first XMIN */
8319 : for (;;)
8320 23736 : {
8321 : ItemId lp;
8322 : HeapTupleHeader htup;
8323 :
8324 : /* Sanity check (pure paranoia) */
8325 435459 : if (offnum < FirstOffsetNumber)
8326 0 : break;
8327 :
8328 : /*
8329 : * An offset past the end of page's line pointer array is possible
8330 : * when the array was truncated
8331 : */
8332 435459 : if (offnum > maxoff)
8333 0 : break;
8334 :
8335 435459 : lp = PageGetItemId(page, offnum);
8336 435459 : if (ItemIdIsRedirected(lp))
8337 : {
8338 11519 : offnum = ItemIdGetRedirect(lp);
8339 11519 : continue;
8340 : }
8341 :
8342 : /*
8343 : * We'll often encounter LP_DEAD line pointers (especially with an
8344 : * entry marked knowndeletable by our caller up front). No heap
8345 : * tuple headers get examined for an htid that leads us to an
8346 : * LP_DEAD item. This is okay because the earlier pruning
8347 : * operation that made the line pointer LP_DEAD in the first place
8348 : * must have considered the original tuple header as part of
8349 : * generating its own snapshotConflictHorizon value.
8350 : *
8351 : * Relying on XLOG_HEAP2_PRUNE_VACUUM_SCAN records like this is
8352 : * the same strategy that index vacuuming uses in all cases. Index
8353 : * VACUUM WAL records don't even have a snapshotConflictHorizon
8354 : * field of their own for this reason.
8355 : */
8356 423940 : if (!ItemIdIsNormal(lp))
8357 274330 : break;
8358 :
8359 149610 : htup = (HeapTupleHeader) PageGetItem(page, lp);
8360 :
8361 : /*
8362 : * Check the tuple XMIN against prior XMAX, if any
8363 : */
8364 161827 : if (TransactionIdIsValid(priorXmax) &&
8365 12217 : !TransactionIdEquals(HeapTupleHeaderGetXmin(htup), priorXmax))
8366 0 : break;
8367 :
8368 149610 : HeapTupleHeaderAdvanceConflictHorizon(htup,
8369 : &snapshotConflictHorizon);
8370 :
8371 : /*
8372 : * If the tuple is not HOT-updated, then we are at the end of this
8373 : * HOT-chain. No need to visit later tuples from the same update
8374 : * chain (they get their own index entries) -- just move on to
8375 : * next htid from index AM caller.
8376 : */
8377 149610 : if (!HeapTupleHeaderIsHotUpdated(htup))
8378 137393 : break;
8379 :
8380 : /* Advance to next HOT chain member */
8381 : Assert(ItemPointerGetBlockNumber(&htup->t_ctid) == blkno);
8382 12217 : offnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
8383 12217 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
8384 : }
8385 :
8386 : /* Enable further/final shrinking of deltids for caller */
8387 411723 : finalndeltids = i + 1;
8388 : }
8389 :
8390 8440 : UnlockReleaseBuffer(buf);
8391 :
8392 : /*
8393 : * Shrink deltids array to exclude non-deletable entries at the end. This
8394 : * is not just a minor optimization. Final deltids array size might be
8395 : * zero for a bottom-up caller. Index AM is explicitly allowed to rely on
8396 : * ndeltids being zero in all cases with zero total deletable entries.
8397 : */
8398 : Assert(finalndeltids > 0 || delstate->bottomup);
8399 8440 : delstate->ndeltids = finalndeltids;
8400 :
8401 8440 : return snapshotConflictHorizon;
8402 : }
8403 :
8404 : /*
8405 : * Specialized inlineable comparison function for index_delete_sort()
8406 : */
8407 : static inline int
8408 18186204 : index_delete_sort_cmp(TM_IndexDelete *deltid1, TM_IndexDelete *deltid2)
8409 : {
8410 18186204 : ItemPointer tid1 = &deltid1->tid;
8411 18186204 : ItemPointer tid2 = &deltid2->tid;
8412 :
8413 : {
8414 18186204 : BlockNumber blk1 = ItemPointerGetBlockNumber(tid1);
8415 18186204 : BlockNumber blk2 = ItemPointerGetBlockNumber(tid2);
8416 :
8417 18186204 : if (blk1 != blk2)
8418 7214455 : return (blk1 < blk2) ? -1 : 1;
8419 : }
8420 : {
8421 10971749 : OffsetNumber pos1 = ItemPointerGetOffsetNumber(tid1);
8422 10971749 : OffsetNumber pos2 = ItemPointerGetOffsetNumber(tid2);
8423 :
8424 10971749 : if (pos1 != pos2)
8425 10971749 : return (pos1 < pos2) ? -1 : 1;
8426 : }
8427 :
8428 : Assert(false);
8429 :
8430 0 : return 0;
8431 : }
8432 :
8433 : /*
8434 : * Sort deltids array from delstate by TID. This prepares it for further
8435 : * processing by heap_index_delete_tuples().
8436 : *
8437 : * This operation becomes a noticeable consumer of CPU cycles with some
8438 : * workloads, so we go to the trouble of specialization/micro optimization.
8439 : * We use shellsort for this because it's easy to specialize, compiles to
8440 : * relatively few instructions, and is adaptive to presorted inputs/subsets
8441 : * (which are typical here).
8442 : */
8443 : static void
8444 8440 : index_delete_sort(TM_IndexDeleteOp *delstate)
8445 : {
8446 8440 : TM_IndexDelete *deltids = delstate->deltids;
8447 8440 : int ndeltids = delstate->ndeltids;
8448 :
8449 : /*
8450 : * Shellsort gap sequence (taken from Sedgewick-Incerpi paper).
8451 : *
8452 : * This implementation is fast with array sizes up to ~4500. This covers
8453 : * all supported BLCKSZ values.
8454 : */
8455 8440 : const int gaps[9] = {1968, 861, 336, 112, 48, 21, 7, 3, 1};
8456 :
8457 : /* Think carefully before changing anything here -- keep swaps cheap */
8458 : StaticAssertDecl(sizeof(TM_IndexDelete) <= 8,
8459 : "element size exceeds 8 bytes");
8460 :
8461 84400 : for (int g = 0; g < lengthof(gaps); g++)
8462 : {
8463 11015773 : for (int hi = gaps[g], i = hi; i < ndeltids; i++)
8464 : {
8465 10939813 : TM_IndexDelete d = deltids[i];
8466 10939813 : int j = i;
8467 :
8468 18713165 : while (j >= hi && index_delete_sort_cmp(&deltids[j - hi], &d) >= 0)
8469 : {
8470 7773352 : deltids[j] = deltids[j - hi];
8471 7773352 : j -= hi;
8472 : }
8473 10939813 : deltids[j] = d;
8474 : }
8475 : }
8476 8440 : }
8477 :
8478 : /*
8479 : * Returns how many blocks should be considered favorable/contiguous for a
8480 : * bottom-up index deletion pass. This is a number of heap blocks that starts
8481 : * from and includes the first block in line.
8482 : *
8483 : * There is always at least one favorable block during bottom-up index
8484 : * deletion. In the worst case (i.e. with totally random heap blocks) the
8485 : * first block in line (the only favorable block) can be thought of as a
8486 : * degenerate array of contiguous blocks that consists of a single block.
8487 : * heap_index_delete_tuples() will expect this.
8488 : *
8489 : * Caller passes blockgroups, a description of the final order that deltids
8490 : * will be sorted in for heap_index_delete_tuples() bottom-up index deletion
8491 : * processing. Note that deltids need not actually be sorted just yet (caller
8492 : * only passes deltids to us so that we can interpret blockgroups).
8493 : *
8494 : * You might guess that the existence of contiguous blocks cannot matter much,
8495 : * since in general the main factor that determines which blocks we visit is
8496 : * the number of promising TIDs, which is a fixed hint from the index AM.
8497 : * We're not really targeting the general case, though -- the actual goal is
8498 : * to adapt our behavior to a wide variety of naturally occurring conditions.
8499 : * The effects of most of the heuristics we apply are only noticeable in the
8500 : * aggregate, over time and across many _related_ bottom-up index deletion
8501 : * passes.
8502 : *
8503 : * Deeming certain blocks favorable allows heapam to recognize and adapt to
8504 : * workloads where heap blocks visited during bottom-up index deletion can be
8505 : * accessed contiguously, in the sense that each newly visited block is the
8506 : * neighbor of the block that bottom-up deletion just finished processing (or
8507 : * close enough to it). It will likely be cheaper to access more favorable
8508 : * blocks sooner rather than later (e.g. in this pass, not across a series of
8509 : * related bottom-up passes). Either way it is probably only a matter of time
8510 : * (or a matter of further correlated version churn) before all blocks that
8511 : * appear together as a single large batch of favorable blocks get accessed by
8512 : * _some_ bottom-up pass. Large batches of favorable blocks tend to either
8513 : * appear almost constantly or not even once (it all depends on per-index
8514 : * workload characteristics).
8515 : *
8516 : * Note that the blockgroups sort order applies a power-of-two bucketing
8517 : * scheme that creates opportunities for contiguous groups of blocks to get
8518 : * batched together, at least with workloads that are naturally amenable to
8519 : * being driven by heap block locality. This doesn't just enhance the spatial
8520 : * locality of bottom-up heap block processing in the obvious way. It also
8521 : * enables temporal locality of access, since sorting by heap block number
8522 : * naturally tends to make the bottom-up processing order deterministic.
8523 : *
8524 : * Consider the following example to get a sense of how temporal locality
8525 : * might matter: There is a heap relation with several indexes, each of which
8526 : * is low to medium cardinality. It is subject to constant non-HOT updates.
8527 : * The updates are skewed (in one part of the primary key, perhaps). None of
8528 : * the indexes are logically modified by the UPDATE statements (if they were
8529 : * then bottom-up index deletion would not be triggered in the first place).
8530 : * Naturally, each new round of index tuples (for each heap tuple that gets a
8531 : * heap_update() call) will have the same heap TID in each and every index.
8532 : * Since these indexes are low cardinality and never get logically modified,
8533 : * heapam processing during bottom-up deletion passes will access heap blocks
8534 : * in approximately sequential order. Temporal locality of access occurs due
8535 : * to bottom-up deletion passes behaving very similarly across each of the
8536 : * indexes at any given moment. This keeps the number of buffer misses needed
8537 : * to visit heap blocks to a minimum.
8538 : */
8539 : static int
8540 2903 : bottomup_nblocksfavorable(IndexDeleteCounts *blockgroups, int nblockgroups,
8541 : TM_IndexDelete *deltids)
8542 : {
8543 2903 : int64 lastblock = -1;
8544 2903 : int nblocksfavorable = 0;
8545 :
8546 : Assert(nblockgroups >= 1);
8547 : Assert(nblockgroups <= BOTTOMUP_MAX_NBLOCKS);
8548 :
8549 : /*
8550 : * We tolerate heap blocks that will be accessed only slightly out of
8551 : * physical order. Small blips occur when a pair of almost-contiguous
8552 : * blocks happen to fall into different buckets (perhaps due only to a
8553 : * small difference in npromisingtids that the bucketing scheme didn't
8554 : * quite manage to ignore). We effectively ignore these blips by applying
8555 : * a small tolerance. The precise tolerance we use is a little arbitrary,
8556 : * but it works well enough in practice.
8557 : */
8558 9029 : for (int b = 0; b < nblockgroups; b++)
8559 : {
8560 8625 : IndexDeleteCounts *group = blockgroups + b;
8561 8625 : TM_IndexDelete *firstdtid = deltids + group->ifirsttid;
8562 8625 : BlockNumber block = ItemPointerGetBlockNumber(&firstdtid->tid);
8563 :
8564 8625 : if (lastblock != -1 &&
8565 5722 : ((int64) block < lastblock - BOTTOMUP_TOLERANCE_NBLOCKS ||
8566 5104 : (int64) block > lastblock + BOTTOMUP_TOLERANCE_NBLOCKS))
8567 : break;
8568 :
8569 6126 : nblocksfavorable++;
8570 6126 : lastblock = block;
8571 : }
8572 :
8573 : /* Always indicate that there is at least 1 favorable block */
8574 : Assert(nblocksfavorable >= 1);
8575 :
8576 2903 : return nblocksfavorable;
8577 : }
8578 :
8579 : /*
8580 : * qsort comparison function for bottomup_sort_and_shrink()
8581 : */
8582 : static int
8583 258965 : bottomup_sort_and_shrink_cmp(const void *arg1, const void *arg2)
8584 : {
8585 258965 : const IndexDeleteCounts *group1 = (const IndexDeleteCounts *) arg1;
8586 258965 : const IndexDeleteCounts *group2 = (const IndexDeleteCounts *) arg2;
8587 :
8588 : /*
8589 : * Most significant field is npromisingtids (which we invert the order of
8590 : * so as to sort in desc order).
8591 : *
8592 : * Caller should have already normalized npromisingtids fields into
8593 : * power-of-two values (buckets).
8594 : */
8595 258965 : if (group1->npromisingtids > group2->npromisingtids)
8596 12145 : return -1;
8597 246820 : if (group1->npromisingtids < group2->npromisingtids)
8598 15674 : return 1;
8599 :
8600 : /*
8601 : * Tiebreak: desc ntids sort order.
8602 : *
8603 : * We cannot expect power-of-two values for ntids fields. We should
8604 : * behave as if they were already rounded up for us instead.
8605 : */
8606 231146 : if (group1->ntids != group2->ntids)
8607 : {
8608 165228 : uint32 ntids1 = pg_nextpower2_32((uint32) group1->ntids);
8609 165228 : uint32 ntids2 = pg_nextpower2_32((uint32) group2->ntids);
8610 :
8611 165228 : if (ntids1 > ntids2)
8612 24342 : return -1;
8613 140886 : if (ntids1 < ntids2)
8614 30176 : return 1;
8615 : }
8616 :
8617 : /*
8618 : * Tiebreak: asc offset-into-deltids-for-block (offset to first TID for
8619 : * block in deltids array) order.
8620 : *
8621 : * This is equivalent to sorting in ascending heap block number order
8622 : * (among otherwise equal subsets of the array). This approach allows us
8623 : * to avoid accessing the out-of-line TID. (We rely on the assumption
8624 : * that the deltids array was sorted in ascending heap TID order when
8625 : * these offsets to the first TID from each heap block group were formed.)
8626 : */
8627 176628 : if (group1->ifirsttid > group2->ifirsttid)
8628 85826 : return 1;
8629 90802 : if (group1->ifirsttid < group2->ifirsttid)
8630 90802 : return -1;
8631 :
8632 0 : pg_unreachable();
8633 :
8634 : return 0;
8635 : }
8636 :
8637 : /*
8638 : * heap_index_delete_tuples() helper function for bottom-up deletion callers.
8639 : *
8640 : * Sorts deltids array in the order needed for useful processing by bottom-up
8641 : * deletion. The array should already be sorted in TID order when we're
8642 : * called. The sort process groups heap TIDs from deltids into heap block
8643 : * groupings. Earlier/more-promising groups/blocks are usually those that are
8644 : * known to have the most "promising" TIDs.
8645 : *
8646 : * Sets new size of deltids array (ndeltids) in state. deltids will only have
8647 : * TIDs from the BOTTOMUP_MAX_NBLOCKS most promising heap blocks when we
8648 : * return. This often means that deltids will be shrunk to a small fraction
8649 : * of its original size (we eliminate many heap blocks from consideration for
8650 : * caller up front).
8651 : *
8652 : * Returns the number of "favorable" blocks. See bottomup_nblocksfavorable()
8653 : * for a definition and full details.
8654 : */
8655 : static int
8656 2903 : bottomup_sort_and_shrink(TM_IndexDeleteOp *delstate)
8657 : {
8658 : IndexDeleteCounts *blockgroups;
8659 : TM_IndexDelete *reordereddeltids;
8660 2903 : BlockNumber curblock = InvalidBlockNumber;
8661 2903 : int nblockgroups = 0;
8662 2903 : int ncopied = 0;
8663 2903 : int nblocksfavorable = 0;
8664 :
8665 : Assert(delstate->bottomup);
8666 : Assert(delstate->ndeltids > 0);
8667 :
8668 : /* Calculate per-heap-block count of TIDs */
8669 2903 : blockgroups = palloc_array(IndexDeleteCounts, delstate->ndeltids);
8670 1393376 : for (int i = 0; i < delstate->ndeltids; i++)
8671 : {
8672 1390473 : TM_IndexDelete *ideltid = &delstate->deltids[i];
8673 1390473 : TM_IndexStatus *istatus = delstate->status + ideltid->id;
8674 1390473 : ItemPointer htid = &ideltid->tid;
8675 1390473 : bool promising = istatus->promising;
8676 :
8677 1390473 : if (curblock != ItemPointerGetBlockNumber(htid))
8678 : {
8679 : /* New block group */
8680 51111 : nblockgroups++;
8681 :
8682 : Assert(curblock < ItemPointerGetBlockNumber(htid) ||
8683 : !BlockNumberIsValid(curblock));
8684 :
8685 51111 : curblock = ItemPointerGetBlockNumber(htid);
8686 51111 : blockgroups[nblockgroups - 1].ifirsttid = i;
8687 51111 : blockgroups[nblockgroups - 1].ntids = 1;
8688 51111 : blockgroups[nblockgroups - 1].npromisingtids = 0;
8689 : }
8690 : else
8691 : {
8692 1339362 : blockgroups[nblockgroups - 1].ntids++;
8693 : }
8694 :
8695 1390473 : if (promising)
8696 230354 : blockgroups[nblockgroups - 1].npromisingtids++;
8697 : }
8698 :
8699 : /*
8700 : * We're about ready to sort block groups to determine the optimal order
8701 : * for visiting heap blocks. But before we do, round the number of
8702 : * promising tuples for each block group up to the next power-of-two,
8703 : * unless it is very low (less than 4), in which case we round up to 4.
8704 : * npromisingtids is far too noisy to trust when choosing between a pair
8705 : * of block groups that both have very low values.
8706 : *
8707 : * This scheme divides heap blocks/block groups into buckets. Each bucket
8708 : * contains blocks that have _approximately_ the same number of promising
8709 : * TIDs as each other. The goal is to ignore relatively small differences
8710 : * in the total number of promising entries, so that the whole process can
8711 : * give a little weight to heapam factors (like heap block locality)
8712 : * instead. This isn't a trade-off, really -- we have nothing to lose. It
8713 : * would be foolish to interpret small differences in npromisingtids
8714 : * values as anything more than noise.
8715 : *
8716 : * We tiebreak on nhtids when sorting block group subsets that have the
8717 : * same npromisingtids, but this has the same issues as npromisingtids,
8718 : * and so nhtids is subject to the same power-of-two bucketing scheme. The
8719 : * only reason that we don't fix nhtids in the same way here too is that
8720 : * we'll need accurate nhtids values after the sort. We handle nhtids
8721 : * bucketization dynamically instead (in the sort comparator).
8722 : *
8723 : * See bottomup_nblocksfavorable() for a full explanation of when and how
8724 : * heap locality/favorable blocks can significantly influence when and how
8725 : * heap blocks are accessed.
8726 : */
8727 54014 : for (int b = 0; b < nblockgroups; b++)
8728 : {
8729 51111 : IndexDeleteCounts *group = blockgroups + b;
8730 :
8731 : /* Better off falling back on nhtids with low npromisingtids */
8732 51111 : if (group->npromisingtids <= 4)
8733 42870 : group->npromisingtids = 4;
8734 : else
8735 8241 : group->npromisingtids =
8736 8241 : pg_nextpower2_32((uint32) group->npromisingtids);
8737 : }
8738 :
8739 : /* Sort groups and rearrange caller's deltids array */
8740 2903 : qsort(blockgroups, nblockgroups, sizeof(IndexDeleteCounts),
8741 : bottomup_sort_and_shrink_cmp);
8742 2903 : reordereddeltids = palloc(delstate->ndeltids * sizeof(TM_IndexDelete));
8743 :
8744 2903 : nblockgroups = Min(BOTTOMUP_MAX_NBLOCKS, nblockgroups);
8745 : /* Determine number of favorable blocks at the start of final deltids */
8746 2903 : nblocksfavorable = bottomup_nblocksfavorable(blockgroups, nblockgroups,
8747 : delstate->deltids);
8748 :
8749 18746 : for (int b = 0; b < nblockgroups; b++)
8750 : {
8751 15843 : IndexDeleteCounts *group = blockgroups + b;
8752 15843 : TM_IndexDelete *firstdtid = delstate->deltids + group->ifirsttid;
8753 :
8754 15843 : memcpy(reordereddeltids + ncopied, firstdtid,
8755 15843 : sizeof(TM_IndexDelete) * group->ntids);
8756 15843 : ncopied += group->ntids;
8757 : }
8758 :
8759 : /* Copy final grouped and sorted TIDs back into start of caller's array */
8760 2903 : memcpy(delstate->deltids, reordereddeltids,
8761 : sizeof(TM_IndexDelete) * ncopied);
8762 2903 : delstate->ndeltids = ncopied;
8763 :
8764 2903 : pfree(reordereddeltids);
8765 2903 : pfree(blockgroups);
8766 :
8767 2903 : return nblocksfavorable;
8768 : }
8769 :
8770 : /*
8771 : * Perform XLogInsert for a heap-update operation. Caller must already
8772 : * have modified the buffer(s) and marked them dirty.
8773 : */
8774 : static XLogRecPtr
8775 2382791 : log_heap_update(Relation reln, Buffer oldbuf,
8776 : Buffer newbuf, HeapTuple oldtup, HeapTuple newtup,
8777 : HeapTuple old_key_tuple,
8778 : bool all_visible_cleared, bool new_all_visible_cleared,
8779 : bool walLogical)
8780 : {
8781 : xl_heap_update xlrec;
8782 : xl_heap_header xlhdr;
8783 : xl_heap_header xlhdr_idx;
8784 : uint8 info;
8785 : uint16 prefix_suffix[2];
8786 2382791 : uint16 prefixlen = 0,
8787 2382791 : suffixlen = 0;
8788 : XLogRecPtr recptr;
8789 2382791 : Page page = BufferGetPage(newbuf);
8790 2382791 : bool need_tuple_data = walLogical && RelationIsLogicallyLogged(reln);
8791 : bool init;
8792 : int bufflags;
8793 :
8794 : /* Caller should not call me on a non-WAL-logged relation */
8795 : Assert(RelationNeedsWAL(reln));
8796 :
8797 2382791 : XLogBeginInsert();
8798 :
8799 2382791 : if (HeapTupleIsHeapOnly(newtup))
8800 169598 : info = XLOG_HEAP_HOT_UPDATE;
8801 : else
8802 2213193 : info = XLOG_HEAP_UPDATE;
8803 :
8804 : /*
8805 : * If the old and new tuple are on the same page, we only need to log the
8806 : * parts of the new tuple that were changed. That saves on the amount of
8807 : * WAL we need to write. Currently, we just count any unchanged bytes in
8808 : * the beginning and end of the tuple. That's quick to check, and
8809 : * perfectly covers the common case that only one field is updated.
8810 : *
8811 : * We could do this even if the old and new tuple are on different pages,
8812 : * but only if we don't make a full-page image of the old page, which is
8813 : * difficult to know in advance. Also, if the old tuple is corrupt for
8814 : * some reason, it would allow the corruption to propagate the new page,
8815 : * so it seems best to avoid. Under the general assumption that most
8816 : * updates tend to create the new tuple version on the same page, there
8817 : * isn't much to be gained by doing this across pages anyway.
8818 : *
8819 : * Skip this if we're taking a full-page image of the new page, as we
8820 : * don't include the new tuple in the WAL record in that case. Also
8821 : * disable if effective_wal_level='logical', as logical decoding needs to
8822 : * be able to read the new tuple in whole from the WAL record alone.
8823 : */
8824 2382791 : if (oldbuf == newbuf && !need_tuple_data &&
8825 174236 : !XLogCheckBufferNeedsBackup(newbuf))
8826 : {
8827 173703 : char *oldp = (char *) oldtup->t_data + oldtup->t_data->t_hoff;
8828 173703 : char *newp = (char *) newtup->t_data + newtup->t_data->t_hoff;
8829 173703 : int oldlen = oldtup->t_len - oldtup->t_data->t_hoff;
8830 173703 : int newlen = newtup->t_len - newtup->t_data->t_hoff;
8831 :
8832 : /* Check for common prefix between old and new tuple */
8833 14553319 : for (prefixlen = 0; prefixlen < Min(oldlen, newlen); prefixlen++)
8834 : {
8835 14525455 : if (newp[prefixlen] != oldp[prefixlen])
8836 145839 : break;
8837 : }
8838 :
8839 : /*
8840 : * Storing the length of the prefix takes 2 bytes, so we need to save
8841 : * at least 3 bytes or there's no point.
8842 : */
8843 173703 : if (prefixlen < 3)
8844 24829 : prefixlen = 0;
8845 :
8846 : /* Same for suffix */
8847 6295653 : for (suffixlen = 0; suffixlen < Min(oldlen, newlen) - prefixlen; suffixlen++)
8848 : {
8849 6267455 : if (newp[newlen - suffixlen - 1] != oldp[oldlen - suffixlen - 1])
8850 145505 : break;
8851 : }
8852 173703 : if (suffixlen < 3)
8853 43409 : suffixlen = 0;
8854 : }
8855 :
8856 : /* Prepare main WAL data chain */
8857 2382791 : xlrec.flags = 0;
8858 2382791 : if (all_visible_cleared)
8859 2138 : xlrec.flags |= XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED;
8860 2382791 : if (new_all_visible_cleared)
8861 774 : xlrec.flags |= XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED;
8862 2382791 : if (prefixlen > 0)
8863 148874 : xlrec.flags |= XLH_UPDATE_PREFIX_FROM_OLD;
8864 2382791 : if (suffixlen > 0)
8865 130294 : xlrec.flags |= XLH_UPDATE_SUFFIX_FROM_OLD;
8866 2382791 : if (need_tuple_data)
8867 : {
8868 47053 : xlrec.flags |= XLH_UPDATE_CONTAINS_NEW_TUPLE;
8869 47053 : if (old_key_tuple)
8870 : {
8871 165 : if (reln->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
8872 68 : xlrec.flags |= XLH_UPDATE_CONTAINS_OLD_TUPLE;
8873 : else
8874 97 : xlrec.flags |= XLH_UPDATE_CONTAINS_OLD_KEY;
8875 : }
8876 : }
8877 :
8878 : /* If new tuple is the single and first tuple on page... */
8879 2395908 : if (ItemPointerGetOffsetNumber(&(newtup->t_self)) == FirstOffsetNumber &&
8880 13117 : PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
8881 : {
8882 12964 : info |= XLOG_HEAP_INIT_PAGE;
8883 12964 : init = true;
8884 : }
8885 : else
8886 2369827 : init = false;
8887 :
8888 : /* Prepare WAL data for the old page */
8889 2382791 : xlrec.old_offnum = ItemPointerGetOffsetNumber(&oldtup->t_self);
8890 2382791 : xlrec.old_xmax = HeapTupleHeaderGetRawXmax(oldtup->t_data);
8891 4765582 : xlrec.old_infobits_set = compute_infobits(oldtup->t_data->t_infomask,
8892 2382791 : oldtup->t_data->t_infomask2);
8893 :
8894 : /* Prepare WAL data for the new page */
8895 2382791 : xlrec.new_offnum = ItemPointerGetOffsetNumber(&newtup->t_self);
8896 2382791 : xlrec.new_xmax = HeapTupleHeaderGetRawXmax(newtup->t_data);
8897 :
8898 2382791 : bufflags = REGBUF_STANDARD;
8899 2382791 : if (init)
8900 12964 : bufflags |= REGBUF_WILL_INIT;
8901 2382791 : if (need_tuple_data)
8902 47053 : bufflags |= REGBUF_KEEP_DATA;
8903 :
8904 2382791 : XLogRegisterBuffer(0, newbuf, bufflags);
8905 2382791 : if (oldbuf != newbuf)
8906 2196584 : XLogRegisterBuffer(1, oldbuf, REGBUF_STANDARD);
8907 :
8908 2382791 : XLogRegisterData(&xlrec, SizeOfHeapUpdate);
8909 :
8910 : /*
8911 : * Prepare WAL data for the new tuple.
8912 : */
8913 2382791 : if (prefixlen > 0 || suffixlen > 0)
8914 : {
8915 171640 : if (prefixlen > 0 && suffixlen > 0)
8916 : {
8917 107528 : prefix_suffix[0] = prefixlen;
8918 107528 : prefix_suffix[1] = suffixlen;
8919 107528 : XLogRegisterBufData(0, &prefix_suffix, sizeof(uint16) * 2);
8920 : }
8921 64112 : else if (prefixlen > 0)
8922 : {
8923 41346 : XLogRegisterBufData(0, &prefixlen, sizeof(uint16));
8924 : }
8925 : else
8926 : {
8927 22766 : XLogRegisterBufData(0, &suffixlen, sizeof(uint16));
8928 : }
8929 : }
8930 :
8931 2382791 : xlhdr.t_infomask2 = newtup->t_data->t_infomask2;
8932 2382791 : xlhdr.t_infomask = newtup->t_data->t_infomask;
8933 2382791 : xlhdr.t_hoff = newtup->t_data->t_hoff;
8934 : Assert(SizeofHeapTupleHeader + prefixlen + suffixlen <= newtup->t_len);
8935 :
8936 : /*
8937 : * PG73FORMAT: write bitmap [+ padding] [+ oid] + data
8938 : *
8939 : * The 'data' doesn't include the common prefix or suffix.
8940 : */
8941 2382791 : XLogRegisterBufData(0, &xlhdr, SizeOfHeapHeader);
8942 2382791 : if (prefixlen == 0)
8943 : {
8944 2233917 : XLogRegisterBufData(0,
8945 2233917 : (char *) newtup->t_data + SizeofHeapTupleHeader,
8946 2233917 : newtup->t_len - SizeofHeapTupleHeader - suffixlen);
8947 : }
8948 : else
8949 : {
8950 : /*
8951 : * Have to write the null bitmap and data after the common prefix as
8952 : * two separate rdata entries.
8953 : */
8954 : /* bitmap [+ padding] [+ oid] */
8955 148874 : if (newtup->t_data->t_hoff - SizeofHeapTupleHeader > 0)
8956 : {
8957 148874 : XLogRegisterBufData(0,
8958 148874 : (char *) newtup->t_data + SizeofHeapTupleHeader,
8959 148874 : newtup->t_data->t_hoff - SizeofHeapTupleHeader);
8960 : }
8961 :
8962 : /* data after common prefix */
8963 148874 : XLogRegisterBufData(0,
8964 148874 : (char *) newtup->t_data + newtup->t_data->t_hoff + prefixlen,
8965 148874 : newtup->t_len - newtup->t_data->t_hoff - prefixlen - suffixlen);
8966 : }
8967 :
8968 : /* We need to log a tuple identity */
8969 2382791 : if (need_tuple_data && old_key_tuple)
8970 : {
8971 : /* don't really need this, but its more comfy to decode */
8972 165 : xlhdr_idx.t_infomask2 = old_key_tuple->t_data->t_infomask2;
8973 165 : xlhdr_idx.t_infomask = old_key_tuple->t_data->t_infomask;
8974 165 : xlhdr_idx.t_hoff = old_key_tuple->t_data->t_hoff;
8975 :
8976 165 : XLogRegisterData(&xlhdr_idx, SizeOfHeapHeader);
8977 :
8978 : /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
8979 165 : XLogRegisterData((char *) old_key_tuple->t_data + SizeofHeapTupleHeader,
8980 165 : old_key_tuple->t_len - SizeofHeapTupleHeader);
8981 : }
8982 :
8983 : /* filtering by origin on a row level is much more efficient */
8984 2382791 : XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
8985 :
8986 2382791 : recptr = XLogInsert(RM_HEAP_ID, info);
8987 :
8988 2382791 : return recptr;
8989 : }
8990 :
8991 : /*
8992 : * Perform XLogInsert of an XLOG_HEAP2_NEW_CID record
8993 : *
8994 : * This is only used when effective_wal_level is logical, and only for
8995 : * catalog tuples.
8996 : */
8997 : static XLogRecPtr
8998 33018 : log_heap_new_cid(Relation relation, HeapTuple tup)
8999 : {
9000 : xl_heap_new_cid xlrec;
9001 :
9002 : XLogRecPtr recptr;
9003 33018 : HeapTupleHeader hdr = tup->t_data;
9004 :
9005 : Assert(ItemPointerIsValid(&tup->t_self));
9006 : Assert(tup->t_tableOid != InvalidOid);
9007 :
9008 33018 : xlrec.top_xid = GetTopTransactionId();
9009 33018 : xlrec.target_locator = relation->rd_locator;
9010 33018 : xlrec.target_tid = tup->t_self;
9011 :
9012 : /*
9013 : * If the tuple got inserted & deleted in the same TX we definitely have a
9014 : * combo CID, set cmin and cmax.
9015 : */
9016 33018 : if (hdr->t_infomask & HEAP_COMBOCID)
9017 : {
9018 : Assert(!(hdr->t_infomask & HEAP_XMAX_INVALID));
9019 : Assert(!HeapTupleHeaderXminInvalid(hdr));
9020 2256 : xlrec.cmin = HeapTupleHeaderGetCmin(hdr);
9021 2256 : xlrec.cmax = HeapTupleHeaderGetCmax(hdr);
9022 2256 : xlrec.combocid = HeapTupleHeaderGetRawCommandId(hdr);
9023 : }
9024 : /* No combo CID, so only cmin or cmax can be set by this TX */
9025 : else
9026 : {
9027 : /*
9028 : * Tuple inserted.
9029 : *
9030 : * We need to check for LOCK ONLY because multixacts might be
9031 : * transferred to the new tuple in case of FOR KEY SHARE updates in
9032 : * which case there will be an xmax, although the tuple just got
9033 : * inserted.
9034 : */
9035 41504 : if (hdr->t_infomask & HEAP_XMAX_INVALID ||
9036 10742 : HEAP_XMAX_IS_LOCKED_ONLY(hdr->t_infomask))
9037 : {
9038 20021 : xlrec.cmin = HeapTupleHeaderGetRawCommandId(hdr);
9039 20021 : xlrec.cmax = InvalidCommandId;
9040 : }
9041 : /* Tuple from a different tx updated or deleted. */
9042 : else
9043 : {
9044 10741 : xlrec.cmin = InvalidCommandId;
9045 10741 : xlrec.cmax = HeapTupleHeaderGetRawCommandId(hdr);
9046 : }
9047 30762 : xlrec.combocid = InvalidCommandId;
9048 : }
9049 :
9050 : /*
9051 : * Note that we don't need to register the buffer here, because this
9052 : * operation does not modify the page. The insert/update/delete that
9053 : * called us certainly did, but that's WAL-logged separately.
9054 : */
9055 33018 : XLogBeginInsert();
9056 33018 : XLogRegisterData(&xlrec, SizeOfHeapNewCid);
9057 :
9058 : /* will be looked at irrespective of origin */
9059 :
9060 33018 : recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_NEW_CID);
9061 :
9062 33018 : return recptr;
9063 : }
9064 :
9065 : /*
9066 : * Build a heap tuple representing the configured REPLICA IDENTITY to represent
9067 : * the old tuple in an UPDATE or DELETE.
9068 : *
9069 : * Returns NULL if there's no need to log an identity or if there's no suitable
9070 : * key defined.
9071 : *
9072 : * Pass key_required true if any replica identity columns changed value, or if
9073 : * any of them have any external data. Delete must always pass true.
9074 : *
9075 : * *copy is set to true if the returned tuple is a modified copy rather than
9076 : * the same tuple that was passed in.
9077 : */
9078 : static HeapTuple
9079 4289967 : ExtractReplicaIdentity(Relation relation, HeapTuple tp, bool key_required,
9080 : bool *copy)
9081 : {
9082 4289967 : TupleDesc desc = RelationGetDescr(relation);
9083 4289967 : char replident = relation->rd_rel->relreplident;
9084 : Bitmapset *idattrs;
9085 : HeapTuple key_tuple;
9086 : bool nulls[MaxHeapAttributeNumber];
9087 : Datum values[MaxHeapAttributeNumber];
9088 :
9089 4289967 : *copy = false;
9090 :
9091 4289967 : if (!RelationIsLogicallyLogged(relation))
9092 4189573 : return NULL;
9093 :
9094 100394 : if (replident == REPLICA_IDENTITY_NOTHING)
9095 272 : return NULL;
9096 :
9097 100122 : if (replident == REPLICA_IDENTITY_FULL)
9098 : {
9099 : /*
9100 : * When logging the entire old tuple, it very well could contain
9101 : * toasted columns. If so, force them to be inlined.
9102 : */
9103 203 : if (HeapTupleHasExternal(tp))
9104 : {
9105 4 : *copy = true;
9106 4 : tp = toast_flatten_tuple(tp, desc);
9107 : }
9108 203 : return tp;
9109 : }
9110 :
9111 : /* if the key isn't required and we're only logging the key, we're done */
9112 99919 : if (!key_required)
9113 46908 : return NULL;
9114 :
9115 : /* find out the replica identity columns */
9116 53011 : idattrs = RelationGetIndexAttrBitmap(relation,
9117 : INDEX_ATTR_BITMAP_IDENTITY_KEY);
9118 :
9119 : /*
9120 : * If there's no defined replica identity columns, treat as !key_required.
9121 : * (This case should not be reachable from heap_update, since that should
9122 : * calculate key_required accurately. But heap_delete just passes
9123 : * constant true for key_required, so we can hit this case in deletes.)
9124 : */
9125 53011 : if (bms_is_empty(idattrs))
9126 6021 : return NULL;
9127 :
9128 : /*
9129 : * Construct a new tuple containing only the replica identity columns,
9130 : * with nulls elsewhere. While we're at it, assert that the replica
9131 : * identity columns aren't null.
9132 : */
9133 46990 : heap_deform_tuple(tp, desc, values, nulls);
9134 :
9135 150979 : for (int i = 0; i < desc->natts; i++)
9136 : {
9137 103989 : if (bms_is_member(i + 1 - FirstLowInvalidHeapAttributeNumber,
9138 : idattrs))
9139 : Assert(!nulls[i]);
9140 : else
9141 56979 : nulls[i] = true;
9142 : }
9143 :
9144 46990 : key_tuple = heap_form_tuple(desc, values, nulls);
9145 46990 : *copy = true;
9146 :
9147 46990 : bms_free(idattrs);
9148 :
9149 : /*
9150 : * If the tuple, which by here only contains indexed columns, still has
9151 : * toasted columns, force them to be inlined. This is somewhat unlikely
9152 : * since there's limits on the size of indexed columns, so we don't
9153 : * duplicate toast_flatten_tuple()s functionality in the above loop over
9154 : * the indexed columns, even if it would be more efficient.
9155 : */
9156 46990 : if (HeapTupleHasExternal(key_tuple))
9157 : {
9158 4 : HeapTuple oldtup = key_tuple;
9159 :
9160 4 : key_tuple = toast_flatten_tuple(oldtup, desc);
9161 4 : heap_freetuple(oldtup);
9162 : }
9163 :
9164 46990 : return key_tuple;
9165 : }
9166 :
9167 : /*
9168 : * HeapCheckForSerializableConflictOut
9169 : * We are reading a tuple. If it's not visible, there may be a
9170 : * rw-conflict out with the inserter. Otherwise, if it is visible to us
9171 : * but has been deleted, there may be a rw-conflict out with the deleter.
9172 : *
9173 : * We will determine the top level xid of the writing transaction with which
9174 : * we may be in conflict, and ask CheckForSerializableConflictOut() to check
9175 : * for overlap with our own transaction.
9176 : *
9177 : * This function should be called just about anywhere in heapam.c where a
9178 : * tuple has been read. The caller must hold at least a shared lock on the
9179 : * buffer, because this function might set hint bits on the tuple. There is
9180 : * currently no known reason to call this function from an index AM.
9181 : */
9182 : void
9183 42324171 : HeapCheckForSerializableConflictOut(bool visible, Relation relation,
9184 : HeapTuple tuple, Buffer buffer,
9185 : Snapshot snapshot)
9186 : {
9187 : TransactionId xid;
9188 : HTSV_Result htsvResult;
9189 :
9190 42324171 : if (!CheckForSerializableConflictOutNeeded(relation, snapshot))
9191 42316226 : return;
9192 :
9193 : /*
9194 : * Check to see whether the tuple has been written to by a concurrent
9195 : * transaction, either to create it not visible to us, or to delete it
9196 : * while it is visible to us. The "visible" bool indicates whether the
9197 : * tuple is visible to us, while HeapTupleSatisfiesVacuum checks what else
9198 : * is going on with it.
9199 : *
9200 : * In the event of a concurrently inserted tuple that also happens to have
9201 : * been concurrently updated (by a separate transaction), the xmin of the
9202 : * tuple will be used -- not the updater's xid.
9203 : */
9204 7945 : htsvResult = HeapTupleSatisfiesVacuum(tuple, TransactionXmin, buffer);
9205 7945 : switch (htsvResult)
9206 : {
9207 7137 : case HEAPTUPLE_LIVE:
9208 7137 : if (visible)
9209 7124 : return;
9210 13 : xid = HeapTupleHeaderGetXmin(tuple->t_data);
9211 13 : break;
9212 361 : case HEAPTUPLE_RECENTLY_DEAD:
9213 : case HEAPTUPLE_DELETE_IN_PROGRESS:
9214 361 : if (visible)
9215 286 : xid = HeapTupleHeaderGetUpdateXid(tuple->t_data);
9216 : else
9217 75 : xid = HeapTupleHeaderGetXmin(tuple->t_data);
9218 :
9219 361 : if (TransactionIdPrecedes(xid, TransactionXmin))
9220 : {
9221 : /* This is like the HEAPTUPLE_DEAD case */
9222 : Assert(!visible);
9223 69 : return;
9224 : }
9225 292 : break;
9226 327 : case HEAPTUPLE_INSERT_IN_PROGRESS:
9227 327 : xid = HeapTupleHeaderGetXmin(tuple->t_data);
9228 327 : break;
9229 120 : case HEAPTUPLE_DEAD:
9230 : Assert(!visible);
9231 120 : return;
9232 0 : default:
9233 :
9234 : /*
9235 : * The only way to get to this default clause is if a new value is
9236 : * added to the enum type without adding it to this switch
9237 : * statement. That's a bug, so elog.
9238 : */
9239 0 : elog(ERROR, "unrecognized return value from HeapTupleSatisfiesVacuum: %u", htsvResult);
9240 :
9241 : /*
9242 : * In spite of having all enum values covered and calling elog on
9243 : * this default, some compilers think this is a code path which
9244 : * allows xid to be used below without initialization. Silence
9245 : * that warning.
9246 : */
9247 : xid = InvalidTransactionId;
9248 : }
9249 :
9250 : Assert(TransactionIdIsValid(xid));
9251 : Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
9252 :
9253 : /*
9254 : * Find top level xid. Bail out if xid is too early to be a conflict, or
9255 : * if it's our own xid.
9256 : */
9257 632 : if (TransactionIdEquals(xid, GetTopTransactionIdIfAny()))
9258 64 : return;
9259 568 : xid = SubTransGetTopmostTransaction(xid);
9260 568 : if (TransactionIdPrecedes(xid, TransactionXmin))
9261 0 : return;
9262 :
9263 568 : CheckForSerializableConflictOut(relation, xid, snapshot);
9264 : }
|
