Coverage Report

Created: 2026-08-13 07:12

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/src/postgres/src/backend/executor/execReplication.c
Line
Count
Source
1
/*-------------------------------------------------------------------------
2
 *
3
 * execReplication.c
4
 *    miscellaneous executor routines for logical replication
5
 *
6
 * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7
 * Portions Copyright (c) 1994, Regents of the University of California
8
 *
9
 * IDENTIFICATION
10
 *    src/backend/executor/execReplication.c
11
 *
12
 *-------------------------------------------------------------------------
13
 */
14
15
#include "postgres.h"
16
17
#include "access/amapi.h"
18
#include "access/commit_ts.h"
19
#include "access/genam.h"
20
#include "access/gist.h"
21
#include "access/relscan.h"
22
#include "access/tableam.h"
23
#include "access/transam.h"
24
#include "access/xact.h"
25
#include "access/heapam.h"
26
#include "catalog/pg_am_d.h"
27
#include "commands/trigger.h"
28
#include "executor/executor.h"
29
#include "executor/nodeModifyTable.h"
30
#include "replication/conflict.h"
31
#include "replication/logicalrelation.h"
32
#include "storage/lmgr.h"
33
#include "utils/builtins.h"
34
#include "utils/lsyscache.h"
35
#include "utils/rel.h"
36
#include "utils/snapmgr.h"
37
#include "utils/syscache.h"
38
#include "utils/typcache.h"
39
40
41
static bool tuples_equal(TupleTableSlot *slot1, TupleTableSlot *slot2,
42
             TypeCacheEntry **eq, Bitmapset *columns);
43
44
/*
45
 * Setup a ScanKey for a search in the relation 'rel' for a tuple 'key' that
46
 * is setup to match 'rel' (*NOT* idxrel!).
47
 *
48
 * Returns how many columns to use for the index scan.
49
 *
50
 * This is not a generic routine, idxrel must be PK, RI, or an index that can be
51
 * used for a REPLICA IDENTITY FULL table. See FindUsableIndexForReplicaIdentityFull()
52
 * for details.
53
 *
54
 * By definition, replication identity of a rel meets all limitations associated
55
 * with that. Note that any other index could also meet these limitations.
56
 */
57
static int
58
build_replindex_scan_key(ScanKey skey, Relation rel, Relation idxrel,
59
             TupleTableSlot *searchslot)
60
0
{
61
0
  int     index_attoff;
62
0
  int     skey_attoff = 0;
63
0
  Datum   indclassDatum;
64
0
  oidvector  *opclass;
65
0
  int2vector *indkey = &idxrel->rd_index->indkey;
66
67
0
  indclassDatum = SysCacheGetAttrNotNull(INDEXRELID, idxrel->rd_indextuple,
68
0
                       Anum_pg_index_indclass);
69
0
  opclass = (oidvector *) DatumGetPointer(indclassDatum);
70
71
  /* Build scankey for every non-expression attribute in the index. */
72
0
  for (index_attoff = 0; index_attoff < IndexRelationGetNumberOfKeyAttributes(idxrel);
73
0
     index_attoff++)
74
0
  {
75
0
    Oid     operator;
76
0
    Oid     optype;
77
0
    Oid     opfamily;
78
0
    RegProcedure regop;
79
0
    int     table_attno = indkey->values[index_attoff];
80
0
    StrategyNumber eq_strategy;
81
82
0
    if (!AttributeNumberIsValid(table_attno))
83
0
    {
84
      /*
85
       * XXX: Currently, we don't support expressions in the scan key,
86
       * see code below.
87
       */
88
0
      continue;
89
0
    }
90
91
    /*
92
     * Load the operator info.  We need this to get the equality operator
93
     * function for the scan key.
94
     */
95
0
    optype = get_opclass_input_type(opclass->values[index_attoff]);
96
0
    opfamily = get_opclass_family(opclass->values[index_attoff]);
97
0
    eq_strategy = IndexAmTranslateCompareType(COMPARE_EQ, idxrel->rd_rel->relam, opfamily, false);
98
0
    operator = get_opfamily_member(opfamily, optype,
99
0
                     optype,
100
0
                     eq_strategy);
101
102
0
    if (!OidIsValid(operator))
103
0
      elog(ERROR, "missing operator %d(%u,%u) in opfamily %u",
104
0
         eq_strategy, optype, optype, opfamily);
105
106
0
    regop = get_opcode(operator);
107
108
    /* Initialize the scankey. */
109
0
    ScanKeyInit(&skey[skey_attoff],
110
0
          index_attoff + 1,
111
0
          eq_strategy,
112
0
          regop,
113
0
          searchslot->tts_values[table_attno - 1]);
114
115
0
    skey[skey_attoff].sk_collation = idxrel->rd_indcollation[index_attoff];
116
117
    /* Check for null value. */
118
0
    if (searchslot->tts_isnull[table_attno - 1])
119
0
      skey[skey_attoff].sk_flags |= (SK_ISNULL | SK_SEARCHNULL);
120
121
0
    skey_attoff++;
122
0
  }
123
124
  /* There must always be at least one attribute for the index scan. */
125
0
  Assert(skey_attoff > 0);
126
127
0
  return skey_attoff;
128
0
}
129
130
131
/*
132
 * Helper function to check if it is necessary to re-fetch and lock the tuple
133
 * due to concurrent modifications. This function should be called after
134
 * invoking table_tuple_lock.
135
 */
136
static bool
137
should_refetch_tuple(TM_Result res, TM_FailureData *tmfd)
138
0
{
139
0
  bool    refetch = false;
140
141
0
  switch (res)
142
0
  {
143
0
    case TM_Ok:
144
0
      break;
145
0
    case TM_Updated:
146
      /* XXX: Improve handling here */
147
0
      if (ItemPointerIndicatesMovedPartitions(&tmfd->ctid))
148
0
        ereport(LOG,
149
0
            (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
150
0
             errmsg("tuple to be locked was already moved to another partition due to concurrent update, retrying")));
151
0
      else
152
0
        ereport(LOG,
153
0
            (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
154
0
             errmsg("concurrent update, retrying")));
155
0
      refetch = true;
156
0
      break;
157
0
    case TM_Deleted:
158
      /* XXX: Improve handling here */
159
0
      ereport(LOG,
160
0
          (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
161
0
           errmsg("concurrent delete, retrying")));
162
0
      refetch = true;
163
0
      break;
164
0
    case TM_Invisible:
165
0
      elog(ERROR, "attempted to lock invisible tuple");
166
0
      break;
167
0
    default:
168
0
      elog(ERROR, "unexpected table_tuple_lock status: %u", res);
169
0
      break;
170
0
  }
171
172
0
  return refetch;
173
0
}
174
175
/*
176
 * Search the relation 'rel' for tuple using the index.
177
 *
178
 * If a matching tuple is found, lock it with lockmode, fill the slot with its
179
 * contents, and return true.  Return false otherwise.
180
 */
181
bool
182
RelationFindReplTupleByIndex(Relation rel, Oid idxoid,
183
               LockTupleMode lockmode,
184
               TupleTableSlot *searchslot,
185
               TupleTableSlot *outslot)
186
0
{
187
0
  ScanKeyData skey[INDEX_MAX_KEYS];
188
0
  int     skey_attoff;
189
0
  IndexScanDesc scan;
190
0
  SnapshotData snap;
191
0
  TransactionId xwait;
192
0
  Relation  idxrel;
193
0
  bool    found;
194
0
  TypeCacheEntry **eq = NULL;
195
0
  bool    isIdxSafeToSkipDuplicates;
196
197
  /* Open the index. */
198
0
  idxrel = index_open(idxoid, RowExclusiveLock);
199
200
0
  isIdxSafeToSkipDuplicates = (GetRelationIdentityOrPK(rel) == idxoid);
201
202
0
  InitDirtySnapshot(snap);
203
204
  /* Build scan key. */
205
0
  skey_attoff = build_replindex_scan_key(skey, rel, idxrel, searchslot);
206
207
  /* Start an index scan. */
208
0
  scan = index_beginscan(rel, idxrel,
209
0
               &snap, NULL, skey_attoff, 0, SO_NONE);
210
211
0
retry:
212
0
  found = false;
213
214
0
  index_rescan(scan, skey, skey_attoff, NULL, 0);
215
216
  /* Try to find the tuple */
217
0
  while (index_getnext_slot(scan, ForwardScanDirection, outslot))
218
0
  {
219
    /*
220
     * Avoid expensive equality check if the index is primary key or
221
     * replica identity index.
222
     */
223
0
    if (!isIdxSafeToSkipDuplicates)
224
0
    {
225
0
      if (eq == NULL)
226
0
        eq = palloc0_array(TypeCacheEntry *, outslot->tts_tupleDescriptor->natts);
227
228
0
      if (!tuples_equal(outslot, searchslot, eq, NULL))
229
0
        continue;
230
0
    }
231
232
0
    ExecMaterializeSlot(outslot);
233
234
0
    xwait = TransactionIdIsValid(snap.xmin) ?
235
0
      snap.xmin : snap.xmax;
236
237
    /*
238
     * If the tuple is locked, wait for locking transaction to finish and
239
     * retry.
240
     */
241
0
    if (TransactionIdIsValid(xwait))
242
0
    {
243
0
      XactLockTableWait(xwait, NULL, NULL, XLTW_None);
244
0
      goto retry;
245
0
    }
246
247
    /* Found our tuple and it's not locked */
248
0
    found = true;
249
0
    break;
250
0
  }
251
252
  /* Found tuple, try to lock it in the lockmode. */
253
0
  if (found)
254
0
  {
255
0
    TM_FailureData tmfd;
256
0
    TM_Result res;
257
258
0
    PushActiveSnapshot(GetLatestSnapshot());
259
260
0
    res = table_tuple_lock(rel, &(outslot->tts_tid), GetActiveSnapshot(),
261
0
                 outslot,
262
0
                 GetCurrentCommandId(false),
263
0
                 lockmode,
264
0
                 LockWaitBlock,
265
0
                 0 /* don't follow updates */ ,
266
0
                 &tmfd);
267
268
0
    PopActiveSnapshot();
269
270
0
    if (should_refetch_tuple(res, &tmfd))
271
0
      goto retry;
272
0
  }
273
274
0
  index_endscan(scan);
275
276
  /* Don't release lock until commit. */
277
0
  index_close(idxrel, NoLock);
278
279
0
  return found;
280
0
}
281
282
/*
283
 * Compare the tuples in the slots by checking if they have equal values.
284
 *
285
 * If 'columns' is not null, only the columns specified within it will be
286
 * considered for the equality check, ignoring all other columns.
287
 */
288
static bool
289
tuples_equal(TupleTableSlot *slot1, TupleTableSlot *slot2,
290
       TypeCacheEntry **eq, Bitmapset *columns)
291
0
{
292
0
  int     attrnum;
293
294
0
  Assert(slot1->tts_tupleDescriptor->natts ==
295
0
       slot2->tts_tupleDescriptor->natts);
296
297
0
  slot_getallattrs(slot1);
298
0
  slot_getallattrs(slot2);
299
300
  /* Check equality of the attributes. */
301
0
  for (attrnum = 0; attrnum < slot1->tts_tupleDescriptor->natts; attrnum++)
302
0
  {
303
0
    Form_pg_attribute att;
304
0
    TypeCacheEntry *typentry;
305
306
0
    att = TupleDescAttr(slot1->tts_tupleDescriptor, attrnum);
307
308
    /*
309
     * Ignore dropped and generated columns as the publisher doesn't send
310
     * those
311
     */
312
0
    if (att->attisdropped || att->attgenerated)
313
0
      continue;
314
315
    /*
316
     * Ignore columns that are not listed for checking.
317
     */
318
0
    if (columns &&
319
0
      !bms_is_member(att->attnum - FirstLowInvalidHeapAttributeNumber,
320
0
               columns))
321
0
      continue;
322
323
    /*
324
     * If one value is NULL and other is not, then they are certainly not
325
     * equal
326
     */
327
0
    if (slot1->tts_isnull[attrnum] != slot2->tts_isnull[attrnum])
328
0
      return false;
329
330
    /*
331
     * If both are NULL, they can be considered equal.
332
     */
333
0
    if (slot1->tts_isnull[attrnum] || slot2->tts_isnull[attrnum])
334
0
      continue;
335
336
0
    typentry = eq[attrnum];
337
0
    if (typentry == NULL)
338
0
    {
339
0
      typentry = lookup_type_cache(att->atttypid,
340
0
                     TYPECACHE_EQ_OPR_FINFO);
341
0
      if (!OidIsValid(typentry->eq_opr_finfo.fn_oid))
342
0
        ereport(ERROR,
343
0
            (errcode(ERRCODE_UNDEFINED_FUNCTION),
344
0
             errmsg("could not identify an equality operator for type %s",
345
0
                format_type_be(att->atttypid))));
346
0
      eq[attrnum] = typentry;
347
0
    }
348
349
0
    if (!DatumGetBool(FunctionCall2Coll(&typentry->eq_opr_finfo,
350
0
                      att->attcollation,
351
0
                      slot1->tts_values[attrnum],
352
0
                      slot2->tts_values[attrnum])))
353
0
      return false;
354
0
  }
355
356
0
  return true;
357
0
}
358
359
/*
360
 * Search the relation 'rel' for tuple using the sequential scan.
361
 *
362
 * If a matching tuple is found, lock it with lockmode, fill the slot with its
363
 * contents, and return true.  Return false otherwise.
364
 *
365
 * Note that this stops on the first matching tuple.
366
 *
367
 * This can obviously be quite slow on tables that have more than few rows.
368
 */
369
bool
370
RelationFindReplTupleSeq(Relation rel, LockTupleMode lockmode,
371
             TupleTableSlot *searchslot, TupleTableSlot *outslot)
372
0
{
373
0
  TupleTableSlot *scanslot;
374
0
  TableScanDesc scan;
375
0
  SnapshotData snap;
376
0
  TypeCacheEntry **eq;
377
0
  TransactionId xwait;
378
0
  bool    found;
379
0
  TupleDesc desc PG_USED_FOR_ASSERTS_ONLY = RelationGetDescr(rel);
380
381
0
  Assert(equalTupleDescs(desc, outslot->tts_tupleDescriptor));
382
383
0
  eq = palloc0_array(TypeCacheEntry *, outslot->tts_tupleDescriptor->natts);
384
385
  /* Start a heap scan. */
386
0
  InitDirtySnapshot(snap);
387
0
  scan = table_beginscan(rel, &snap, 0, NULL,
388
0
               SO_NONE);
389
0
  scanslot = table_slot_create(rel, NULL);
390
391
0
retry:
392
0
  found = false;
393
394
0
  table_rescan(scan, NULL);
395
396
  /* Try to find the tuple */
397
0
  while (table_scan_getnextslot(scan, ForwardScanDirection, scanslot))
398
0
  {
399
0
    if (!tuples_equal(scanslot, searchslot, eq, NULL))
400
0
      continue;
401
402
0
    found = true;
403
0
    ExecCopySlot(outslot, scanslot);
404
405
0
    xwait = TransactionIdIsValid(snap.xmin) ?
406
0
      snap.xmin : snap.xmax;
407
408
    /*
409
     * If the tuple is locked, wait for locking transaction to finish and
410
     * retry.
411
     */
412
0
    if (TransactionIdIsValid(xwait))
413
0
    {
414
0
      XactLockTableWait(xwait, NULL, NULL, XLTW_None);
415
0
      goto retry;
416
0
    }
417
418
    /* Found our tuple and it's not locked */
419
0
    break;
420
0
  }
421
422
  /* Found tuple, try to lock it in the lockmode. */
423
0
  if (found)
424
0
  {
425
0
    TM_FailureData tmfd;
426
0
    TM_Result res;
427
428
0
    PushActiveSnapshot(GetLatestSnapshot());
429
430
0
    res = table_tuple_lock(rel, &(outslot->tts_tid), GetActiveSnapshot(),
431
0
                 outslot,
432
0
                 GetCurrentCommandId(false),
433
0
                 lockmode,
434
0
                 LockWaitBlock,
435
0
                 0 /* don't follow updates */ ,
436
0
                 &tmfd);
437
438
0
    PopActiveSnapshot();
439
440
0
    if (should_refetch_tuple(res, &tmfd))
441
0
      goto retry;
442
0
  }
443
444
0
  table_endscan(scan);
445
0
  ExecDropSingleTupleTableSlot(scanslot);
446
447
0
  return found;
448
0
}
449
450
/*
451
 * Build additional index information necessary for conflict detection.
452
 */
453
static void
454
BuildConflictIndexInfo(ResultRelInfo *resultRelInfo, Oid conflictindex)
455
0
{
456
0
  for (int i = 0; i < resultRelInfo->ri_NumIndices; i++)
457
0
  {
458
0
    Relation  indexRelation = resultRelInfo->ri_IndexRelationDescs[i];
459
0
    IndexInfo  *indexRelationInfo = resultRelInfo->ri_IndexRelationInfo[i];
460
461
0
    if (conflictindex != RelationGetRelid(indexRelation))
462
0
      continue;
463
464
    /*
465
     * This Assert will fail if BuildSpeculativeIndexInfo() is called
466
     * twice for the given index.
467
     */
468
0
    Assert(indexRelationInfo->ii_UniqueOps == NULL);
469
470
0
    BuildSpeculativeIndexInfo(indexRelation, indexRelationInfo);
471
0
  }
472
0
}
473
474
/*
475
 * If the tuple is recently dead and was deleted by a transaction with a newer
476
 * commit timestamp than previously recorded, update the associated transaction
477
 * ID, commit time, and origin. This helps ensure that conflict detection uses
478
 * the most recent and relevant deletion metadata.
479
 */
480
static void
481
update_most_recent_deletion_info(TupleTableSlot *scanslot,
482
                 TransactionId oldestxmin,
483
                 TransactionId *delete_xid,
484
                 TimestampTz *delete_time,
485
                 ReplOriginId *delete_origin)
486
0
{
487
0
  BufferHeapTupleTableSlot *hslot;
488
0
  HeapTuple tuple;
489
0
  Buffer    buf;
490
0
  bool    recently_dead = false;
491
0
  TransactionId xmax;
492
0
  TimestampTz localts;
493
0
  ReplOriginId localorigin;
494
495
0
  hslot = (BufferHeapTupleTableSlot *) scanslot;
496
497
0
  tuple = ExecFetchSlotHeapTuple(scanslot, false, NULL);
498
0
  buf = hslot->buffer;
499
500
0
  LockBuffer(buf, BUFFER_LOCK_SHARE);
501
502
  /*
503
   * We do not consider HEAPTUPLE_DEAD status because it indicates either
504
   * tuples whose inserting transaction was aborted (meaning there is no
505
   * commit timestamp or origin), or tuples deleted by a transaction older
506
   * than oldestxmin, making it safe to ignore them during conflict
507
   * detection (See comments atop worker.c for details).
508
   */
509
0
  if (HeapTupleSatisfiesVacuum(tuple, oldestxmin, buf) == HEAPTUPLE_RECENTLY_DEAD)
510
0
    recently_dead = true;
511
512
0
  LockBuffer(buf, BUFFER_LOCK_UNLOCK);
513
514
0
  if (!recently_dead)
515
0
    return;
516
517
0
  xmax = HeapTupleHeaderGetUpdateXid(tuple->t_data);
518
0
  if (!TransactionIdIsValid(xmax))
519
0
    return;
520
521
  /* Select the dead tuple with the most recent commit timestamp */
522
0
  if (TransactionIdGetCommitTsData(xmax, &localts, &localorigin) &&
523
0
    TimestampDifferenceExceeds(*delete_time, localts, 0))
524
0
  {
525
0
    *delete_xid = xmax;
526
0
    *delete_time = localts;
527
0
    *delete_origin = localorigin;
528
0
  }
529
0
}
530
531
/*
532
 * Searches the relation 'rel' for the most recently deleted tuple that matches
533
 * the values in 'searchslot' and is not yet removable by VACUUM. The function
534
 * returns the transaction ID, origin, and commit timestamp of the transaction
535
 * that deleted this tuple.
536
 *
537
 * 'oldestxmin' acts as a cutoff transaction ID. Tuples deleted by transactions
538
 * with IDs >= 'oldestxmin' are considered recently dead and are eligible for
539
 * conflict detection.
540
 *
541
 * Instead of stopping at the first match, we scan all matching dead tuples to
542
 * identify most recent deletion. This is crucial because only the latest
543
 * deletion is relevant for resolving conflicts.
544
 *
545
 * For example, consider a scenario on the subscriber where a row is deleted,
546
 * re-inserted, and then deleted again only on the subscriber:
547
 *
548
 *   - (pk, 1) - deleted at 9:00,
549
 *   - (pk, 1) - deleted at 9:02,
550
 *
551
 * Now, a remote update arrives: (pk, 1) -> (pk, 2), timestamped at 9:01.
552
 *
553
 * If we mistakenly return the older deletion (9:00), the system may wrongly
554
 * apply the remote update using a last-update-wins strategy. Instead, we must
555
 * recognize the more recent deletion at 9:02 and skip the update. See
556
 * comments atop worker.c for details. Note, as of now, conflict resolution
557
 * is not implemented. Consequently, the system may incorrectly report the
558
 * older tuple as the conflicted one, leading to misleading results.
559
 *
560
 * The commit timestamp of the deleting transaction is used to determine which
561
 * tuple was deleted most recently.
562
 */
563
bool
564
RelationFindDeletedTupleInfoSeq(Relation rel, TupleTableSlot *searchslot,
565
                TransactionId oldestxmin,
566
                TransactionId *delete_xid,
567
                ReplOriginId *delete_origin,
568
                TimestampTz *delete_time)
569
0
{
570
0
  TupleTableSlot *scanslot;
571
0
  TableScanDesc scan;
572
0
  TypeCacheEntry **eq;
573
0
  Bitmapset  *indexbitmap;
574
0
  TupleDesc desc PG_USED_FOR_ASSERTS_ONLY = RelationGetDescr(rel);
575
576
0
  Assert(equalTupleDescs(desc, searchslot->tts_tupleDescriptor));
577
578
0
  *delete_xid = InvalidTransactionId;
579
0
  *delete_origin = InvalidReplOriginId;
580
0
  *delete_time = 0;
581
582
  /*
583
   * If the relation has a replica identity key or a primary key that is
584
   * unusable for locating deleted tuples (see
585
   * IsIndexUsableForFindingDeletedTuple), a full table scan becomes
586
   * necessary. In such cases, comparing the entire tuple is not required,
587
   * since the remote tuple might not include all column values. Instead,
588
   * the indexed columns alone are sufficient to identify the target tuple
589
   * (see logicalrep_rel_mark_updatable).
590
   */
591
0
  indexbitmap = RelationGetIndexAttrBitmap(rel,
592
0
                       INDEX_ATTR_BITMAP_IDENTITY_KEY);
593
594
  /* fallback to PK if no replica identity */
595
0
  if (!indexbitmap)
596
0
    indexbitmap = RelationGetIndexAttrBitmap(rel,
597
0
                         INDEX_ATTR_BITMAP_PRIMARY_KEY);
598
599
0
  eq = palloc0_array(TypeCacheEntry *, searchslot->tts_tupleDescriptor->natts);
600
601
  /*
602
   * Start a heap scan using SnapshotAny to identify dead tuples that are
603
   * not visible under a standard MVCC snapshot. Tuples from transactions
604
   * not yet committed or those just committed prior to the scan are
605
   * excluded in update_most_recent_deletion_info().
606
   */
607
0
  scan = table_beginscan(rel, SnapshotAny, 0, NULL,
608
0
               SO_NONE);
609
0
  scanslot = table_slot_create(rel, NULL);
610
611
0
  table_rescan(scan, NULL);
612
613
  /* Try to find the tuple */
614
0
  while (table_scan_getnextslot(scan, ForwardScanDirection, scanslot))
615
0
  {
616
0
    if (!tuples_equal(scanslot, searchslot, eq, indexbitmap))
617
0
      continue;
618
619
0
    update_most_recent_deletion_info(scanslot, oldestxmin, delete_xid,
620
0
                     delete_time, delete_origin);
621
0
  }
622
623
0
  table_endscan(scan);
624
0
  ExecDropSingleTupleTableSlot(scanslot);
625
626
0
  return *delete_time != 0;
627
0
}
628
629
/*
630
 * Similar to RelationFindDeletedTupleInfoSeq() but using index scan to locate
631
 * the deleted tuple.
632
 */
633
bool
634
RelationFindDeletedTupleInfoByIndex(Relation rel, Oid idxoid,
635
                  TupleTableSlot *searchslot,
636
                  TransactionId oldestxmin,
637
                  TransactionId *delete_xid,
638
                  ReplOriginId *delete_origin,
639
                  TimestampTz *delete_time)
640
0
{
641
0
  Relation  idxrel;
642
0
  ScanKeyData skey[INDEX_MAX_KEYS];
643
0
  int     skey_attoff;
644
0
  IndexScanDesc scan;
645
0
  TupleTableSlot *scanslot;
646
0
  TypeCacheEntry **eq = NULL;
647
0
  bool    isIdxSafeToSkipDuplicates;
648
0
  TupleDesc desc PG_USED_FOR_ASSERTS_ONLY = RelationGetDescr(rel);
649
650
0
  Assert(equalTupleDescs(desc, searchslot->tts_tupleDescriptor));
651
0
  Assert(OidIsValid(idxoid));
652
653
0
  *delete_xid = InvalidTransactionId;
654
0
  *delete_time = 0;
655
0
  *delete_origin = InvalidReplOriginId;
656
657
0
  isIdxSafeToSkipDuplicates = (GetRelationIdentityOrPK(rel) == idxoid);
658
659
0
  scanslot = table_slot_create(rel, NULL);
660
661
0
  idxrel = index_open(idxoid, RowExclusiveLock);
662
663
  /* Build scan key. */
664
0
  skey_attoff = build_replindex_scan_key(skey, rel, idxrel, searchslot);
665
666
  /*
667
   * Start an index scan using SnapshotAny to identify dead tuples that are
668
   * not visible under a standard MVCC snapshot. Tuples from transactions
669
   * not yet committed or those just committed prior to the scan are
670
   * excluded in update_most_recent_deletion_info().
671
   */
672
0
  scan = index_beginscan(rel, idxrel,
673
0
               SnapshotAny, NULL, skey_attoff, 0, SO_NONE);
674
675
0
  index_rescan(scan, skey, skey_attoff, NULL, 0);
676
677
  /* Try to find the tuple */
678
0
  while (index_getnext_slot(scan, ForwardScanDirection, scanslot))
679
0
  {
680
    /*
681
     * Avoid expensive equality check if the index is primary key or
682
     * replica identity index.
683
     */
684
0
    if (!isIdxSafeToSkipDuplicates)
685
0
    {
686
0
      if (eq == NULL)
687
0
        eq = palloc0_array(TypeCacheEntry *, scanslot->tts_tupleDescriptor->natts);
688
689
0
      if (!tuples_equal(scanslot, searchslot, eq, NULL))
690
0
        continue;
691
0
    }
692
693
0
    update_most_recent_deletion_info(scanslot, oldestxmin, delete_xid,
694
0
                     delete_time, delete_origin);
695
0
  }
696
697
0
  index_endscan(scan);
698
699
0
  index_close(idxrel, NoLock);
700
701
0
  ExecDropSingleTupleTableSlot(scanslot);
702
703
0
  return *delete_time != 0;
704
0
}
705
706
/*
707
 * Find the tuple that violates the passed unique index (conflictindex).
708
 *
709
 * If the conflicting tuple is found return true, otherwise false.
710
 *
711
 * We lock the tuple to avoid getting it deleted before the caller can fetch
712
 * the required information. Note that if the tuple is deleted before a lock
713
 * is acquired, we will retry to find the conflicting tuple again.
714
 */
715
static bool
716
FindConflictTuple(ResultRelInfo *resultRelInfo, EState *estate,
717
          Oid conflictindex, TupleTableSlot *slot,
718
          TupleTableSlot **conflictslot)
719
0
{
720
0
  Relation  rel = resultRelInfo->ri_RelationDesc;
721
0
  ItemPointerData conflictTid;
722
0
  TM_FailureData tmfd;
723
0
  TM_Result res;
724
725
0
  *conflictslot = NULL;
726
727
  /*
728
   * Build additional information required to check constraints violations.
729
   * See check_exclusion_or_unique_constraint().
730
   */
731
0
  BuildConflictIndexInfo(resultRelInfo, conflictindex);
732
733
0
retry:
734
0
  if (ExecCheckIndexConstraints(resultRelInfo, slot, estate,
735
0
                  &conflictTid, &slot->tts_tid,
736
0
                  list_make1_oid(conflictindex)))
737
0
  {
738
0
    if (*conflictslot)
739
0
      ExecDropSingleTupleTableSlot(*conflictslot);
740
741
0
    *conflictslot = NULL;
742
0
    return false;
743
0
  }
744
745
0
  *conflictslot = table_slot_create(rel, NULL);
746
747
0
  PushActiveSnapshot(GetLatestSnapshot());
748
749
0
  res = table_tuple_lock(rel, &conflictTid, GetActiveSnapshot(),
750
0
               *conflictslot,
751
0
               GetCurrentCommandId(false),
752
0
               LockTupleShare,
753
0
               LockWaitBlock,
754
0
               0 /* don't follow updates */ ,
755
0
               &tmfd);
756
757
0
  PopActiveSnapshot();
758
759
0
  if (should_refetch_tuple(res, &tmfd))
760
0
    goto retry;
761
762
0
  return true;
763
0
}
764
765
/*
766
 * Check all the unique indexes in 'recheckIndexes' for conflict with the
767
 * tuple in 'remoteslot' and report if found.
768
 */
769
static void
770
CheckAndReportConflict(ResultRelInfo *resultRelInfo, EState *estate,
771
             ConflictType type, List *recheckIndexes,
772
             TupleTableSlot *searchslot, TupleTableSlot *remoteslot)
773
0
{
774
0
  List     *conflicttuples = NIL;
775
0
  TupleTableSlot *conflictslot;
776
777
  /* Check all the unique indexes for conflicts */
778
0
  foreach_oid(uniqueidx, resultRelInfo->ri_onConflictArbiterIndexes)
779
0
  {
780
0
    if (list_member_oid(recheckIndexes, uniqueidx) &&
781
0
      FindConflictTuple(resultRelInfo, estate, uniqueidx, remoteslot,
782
0
                &conflictslot))
783
0
    {
784
0
      ConflictTupleInfo *conflicttuple = palloc0_object(ConflictTupleInfo);
785
786
0
      conflicttuple->slot = conflictslot;
787
0
      conflicttuple->indexoid = uniqueidx;
788
789
0
      GetTupleTransactionInfo(conflictslot, &conflicttuple->xmin,
790
0
                  &conflicttuple->origin, &conflicttuple->ts);
791
792
0
      conflicttuples = lappend(conflicttuples, conflicttuple);
793
0
    }
794
0
  }
795
796
  /* Report the conflict, if found */
797
0
  if (conflicttuples)
798
0
    ReportApplyConflict(estate, resultRelInfo, ERROR,
799
0
              list_length(conflicttuples) > 1 ? CT_MULTIPLE_UNIQUE_CONFLICTS : type,
800
0
              searchslot, remoteslot, conflicttuples);
801
0
}
802
803
/*
804
 * Insert tuple represented in the slot to the relation, update the indexes,
805
 * and execute any constraints and per-row triggers.
806
 *
807
 * Caller is responsible for opening the indexes.
808
 */
809
void
810
ExecSimpleRelationInsert(ResultRelInfo *resultRelInfo,
811
             EState *estate, TupleTableSlot *slot)
812
0
{
813
0
  bool    skip_tuple = false;
814
0
  Relation  rel = resultRelInfo->ri_RelationDesc;
815
816
  /* For now we support only tables. */
817
0
  Assert(rel->rd_rel->relkind == RELKIND_RELATION);
818
819
0
  CheckCmdReplicaIdentity(rel, CMD_INSERT);
820
821
  /* BEFORE ROW INSERT Triggers */
822
0
  if (resultRelInfo->ri_TrigDesc &&
823
0
    resultRelInfo->ri_TrigDesc->trig_insert_before_row)
824
0
  {
825
0
    if (!ExecBRInsertTriggers(estate, resultRelInfo, slot))
826
0
      skip_tuple = true; /* "do nothing" */
827
0
  }
828
829
0
  if (!skip_tuple)
830
0
  {
831
0
    List     *recheckIndexes = NIL;
832
0
    List     *conflictindexes;
833
0
    bool    conflict = false;
834
835
    /* Compute stored generated columns */
836
0
    if (rel->rd_att->constr &&
837
0
      rel->rd_att->constr->has_generated_stored)
838
0
      ExecComputeStoredGenerated(resultRelInfo, estate, slot,
839
0
                     CMD_INSERT);
840
841
    /* Check the constraints of the tuple */
842
0
    if (rel->rd_att->constr)
843
0
      ExecConstraints(resultRelInfo, slot, estate);
844
0
    if (rel->rd_rel->relispartition)
845
0
      ExecPartitionCheck(resultRelInfo, slot, estate, true);
846
847
    /* OK, store the tuple and create index entries for it */
848
0
    simple_table_tuple_insert(resultRelInfo->ri_RelationDesc, slot);
849
850
0
    conflictindexes = resultRelInfo->ri_onConflictArbiterIndexes;
851
852
0
    if (resultRelInfo->ri_NumIndices > 0)
853
0
    {
854
0
      uint32    flags;
855
856
0
      if (conflictindexes != NIL)
857
0
        flags = EIIT_NO_DUPE_ERROR;
858
0
      else
859
0
        flags = 0;
860
0
      recheckIndexes = ExecInsertIndexTuples(resultRelInfo,
861
0
                           estate, flags,
862
0
                           slot, conflictindexes,
863
0
                           &conflict);
864
0
    }
865
866
    /*
867
     * Checks the conflict indexes to fetch the conflicting local row and
868
     * reports the conflict. We perform this check here, instead of
869
     * performing an additional index scan before the actual insertion and
870
     * reporting the conflict if any conflicting rows are found. This is
871
     * to avoid the overhead of executing the extra scan for each INSERT
872
     * operation, even when no conflict arises, which could introduce
873
     * significant overhead to replication, particularly in cases where
874
     * conflicts are rare.
875
     *
876
     * XXX OTOH, this could lead to clean-up effort for dead tuples added
877
     * in heap and index in case of conflicts. But as conflicts shouldn't
878
     * be a frequent thing so we preferred to save the performance
879
     * overhead of extra scan before each insertion.
880
     */
881
0
    if (conflict)
882
0
      CheckAndReportConflict(resultRelInfo, estate, CT_INSERT_EXISTS,
883
0
                   recheckIndexes, NULL, slot);
884
885
    /* AFTER ROW INSERT Triggers */
886
0
    ExecARInsertTriggers(estate, resultRelInfo, slot,
887
0
               recheckIndexes, NULL);
888
889
    /*
890
     * XXX we should in theory pass a TransitionCaptureState object to the
891
     * above to capture transition tuples, but after statement triggers
892
     * don't actually get fired by replication yet anyway
893
     */
894
895
0
    list_free(recheckIndexes);
896
0
  }
897
0
}
898
899
/*
900
 * Find the searchslot tuple and update it with data in the slot,
901
 * update the indexes, and execute any constraints and per-row triggers.
902
 *
903
 * Caller is responsible for opening the indexes.
904
 */
905
void
906
ExecSimpleRelationUpdate(ResultRelInfo *resultRelInfo,
907
             EState *estate, EPQState *epqstate,
908
             TupleTableSlot *searchslot, TupleTableSlot *slot)
909
0
{
910
0
  bool    skip_tuple = false;
911
0
  Relation  rel = resultRelInfo->ri_RelationDesc;
912
0
  ItemPointer tid = &(searchslot->tts_tid);
913
914
  /*
915
   * We support only non-system tables, with
916
   * check_publication_add_relation() accountable.
917
   */
918
0
  Assert(rel->rd_rel->relkind == RELKIND_RELATION);
919
0
  Assert(!IsCatalogRelation(rel));
920
921
0
  CheckCmdReplicaIdentity(rel, CMD_UPDATE);
922
923
  /* BEFORE ROW UPDATE Triggers */
924
0
  if (resultRelInfo->ri_TrigDesc &&
925
0
    resultRelInfo->ri_TrigDesc->trig_update_before_row)
926
0
  {
927
0
    if (!ExecBRUpdateTriggers(estate, epqstate, resultRelInfo,
928
0
                  tid, NULL, slot, NULL, NULL, false))
929
0
      skip_tuple = true; /* "do nothing" */
930
0
  }
931
932
0
  if (!skip_tuple)
933
0
  {
934
0
    List     *recheckIndexes = NIL;
935
0
    TU_UpdateIndexes update_indexes;
936
0
    List     *conflictindexes;
937
0
    bool    conflict = false;
938
939
    /* Compute stored generated columns */
940
0
    if (rel->rd_att->constr &&
941
0
      rel->rd_att->constr->has_generated_stored)
942
0
      ExecComputeStoredGenerated(resultRelInfo, estate, slot,
943
0
                     CMD_UPDATE);
944
945
    /* Check the constraints of the tuple */
946
0
    if (rel->rd_att->constr)
947
0
      ExecConstraints(resultRelInfo, slot, estate);
948
0
    if (rel->rd_rel->relispartition)
949
0
      ExecPartitionCheck(resultRelInfo, slot, estate, true);
950
951
0
    simple_table_tuple_update(rel, tid, slot, estate->es_snapshot,
952
0
                  &update_indexes);
953
954
0
    conflictindexes = resultRelInfo->ri_onConflictArbiterIndexes;
955
956
0
    if (resultRelInfo->ri_NumIndices > 0 && (update_indexes != TU_None))
957
0
    {
958
0
      uint32    flags = EIIT_IS_UPDATE;
959
960
0
      if (conflictindexes != NIL)
961
0
        flags |= EIIT_NO_DUPE_ERROR;
962
0
      if (update_indexes == TU_Summarizing)
963
0
        flags |= EIIT_ONLY_SUMMARIZING;
964
0
      recheckIndexes = ExecInsertIndexTuples(resultRelInfo,
965
0
                           estate, flags,
966
0
                           slot, conflictindexes,
967
0
                           &conflict);
968
0
    }
969
970
    /*
971
     * Refer to the comments above the call to CheckAndReportConflict() in
972
     * ExecSimpleRelationInsert to understand why this check is done at
973
     * this point.
974
     */
975
0
    if (conflict)
976
0
      CheckAndReportConflict(resultRelInfo, estate, CT_UPDATE_EXISTS,
977
0
                   recheckIndexes, searchslot, slot);
978
979
    /* AFTER ROW UPDATE Triggers */
980
0
    ExecARUpdateTriggers(estate, resultRelInfo,
981
0
               NULL, NULL,
982
0
               tid, NULL, slot,
983
0
               recheckIndexes, NULL, false);
984
985
0
    list_free(recheckIndexes);
986
0
  }
987
0
}
988
989
/*
990
 * Find the searchslot tuple and delete it, and execute any constraints
991
 * and per-row triggers.
992
 *
993
 * Caller is responsible for opening the indexes.
994
 */
995
void
996
ExecSimpleRelationDelete(ResultRelInfo *resultRelInfo,
997
             EState *estate, EPQState *epqstate,
998
             TupleTableSlot *searchslot)
999
0
{
1000
0
  bool    skip_tuple = false;
1001
0
  Relation  rel = resultRelInfo->ri_RelationDesc;
1002
0
  ItemPointer tid = &searchslot->tts_tid;
1003
1004
0
  CheckCmdReplicaIdentity(rel, CMD_DELETE);
1005
1006
  /* BEFORE ROW DELETE Triggers */
1007
0
  if (resultRelInfo->ri_TrigDesc &&
1008
0
    resultRelInfo->ri_TrigDesc->trig_delete_before_row)
1009
0
  {
1010
0
    skip_tuple = !ExecBRDeleteTriggers(estate, epqstate, resultRelInfo,
1011
0
                       tid, NULL, NULL, NULL, NULL, false);
1012
0
  }
1013
1014
0
  if (!skip_tuple)
1015
0
  {
1016
    /* OK, delete the tuple */
1017
0
    simple_table_tuple_delete(rel, tid, estate->es_snapshot);
1018
1019
    /* AFTER ROW DELETE Triggers */
1020
0
    ExecARDeleteTriggers(estate, resultRelInfo,
1021
0
               tid, NULL, NULL, false);
1022
0
  }
1023
0
}
1024
1025
/*
1026
 * Check if command can be executed with current replica identity.
1027
 */
1028
void
1029
CheckCmdReplicaIdentity(Relation rel, CmdType cmd)
1030
0
{
1031
0
  PublicationDesc pubdesc;
1032
1033
  /*
1034
   * Skip checking the replica identity for partitioned tables, because the
1035
   * operations are actually performed on the leaf partitions.
1036
   */
1037
0
  if (rel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
1038
0
    return;
1039
1040
  /* We only need to do checks for UPDATE and DELETE. */
1041
0
  if (cmd != CMD_UPDATE && cmd != CMD_DELETE)
1042
0
    return;
1043
1044
  /*
1045
   * It is only safe to execute UPDATE/DELETE if the relation does not
1046
   * publish UPDATEs or DELETEs, or all the following conditions are
1047
   * satisfied:
1048
   *
1049
   * 1. All columns, referenced in the row filters from publications which
1050
   * the relation is in, are valid - i.e. when all referenced columns are
1051
   * part of REPLICA IDENTITY.
1052
   *
1053
   * 2. All columns, referenced in the column lists are valid - i.e. when
1054
   * all columns referenced in the REPLICA IDENTITY are covered by the
1055
   * column list.
1056
   *
1057
   * 3. All generated columns in REPLICA IDENTITY of the relation, are valid
1058
   * - i.e. when all these generated columns are published.
1059
   *
1060
   * XXX We could optimize it by first checking whether any of the
1061
   * publications have a row filter or column list for this relation, or if
1062
   * the relation contains a generated column. If none of these exist and
1063
   * the relation has replica identity then we can avoid building the
1064
   * descriptor but as this happens only one time it doesn't seem worth the
1065
   * additional complexity.
1066
   */
1067
0
  RelationBuildPublicationDesc(rel, &pubdesc);
1068
0
  if (cmd == CMD_UPDATE && !pubdesc.rf_valid_for_update)
1069
0
    ereport(ERROR,
1070
0
        (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1071
0
         errmsg("cannot update table \"%s\"",
1072
0
            RelationGetRelationName(rel)),
1073
0
         errdetail("Column used in the publication WHERE expression is not part of the replica identity.")));
1074
0
  else if (cmd == CMD_UPDATE && !pubdesc.cols_valid_for_update)
1075
0
    ereport(ERROR,
1076
0
        (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1077
0
         errmsg("cannot update table \"%s\"",
1078
0
            RelationGetRelationName(rel)),
1079
0
         errdetail("Column list used by the publication does not cover the replica identity.")));
1080
0
  else if (cmd == CMD_UPDATE && !pubdesc.gencols_valid_for_update)
1081
0
    ereport(ERROR,
1082
0
        (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1083
0
         errmsg("cannot update table \"%s\"",
1084
0
            RelationGetRelationName(rel)),
1085
0
         errdetail("Replica identity must not contain unpublished generated columns.")));
1086
0
  else if (cmd == CMD_DELETE && !pubdesc.rf_valid_for_delete)
1087
0
    ereport(ERROR,
1088
0
        (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1089
0
         errmsg("cannot delete from table \"%s\"",
1090
0
            RelationGetRelationName(rel)),
1091
0
         errdetail("Column used in the publication WHERE expression is not part of the replica identity.")));
1092
0
  else if (cmd == CMD_DELETE && !pubdesc.cols_valid_for_delete)
1093
0
    ereport(ERROR,
1094
0
        (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1095
0
         errmsg("cannot delete from table \"%s\"",
1096
0
            RelationGetRelationName(rel)),
1097
0
         errdetail("Column list used by the publication does not cover the replica identity.")));
1098
0
  else if (cmd == CMD_DELETE && !pubdesc.gencols_valid_for_delete)
1099
0
    ereport(ERROR,
1100
0
        (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1101
0
         errmsg("cannot delete from table \"%s\"",
1102
0
            RelationGetRelationName(rel)),
1103
0
         errdetail("Replica identity must not contain unpublished generated columns.")));
1104
1105
  /* If relation has replica identity we are always good. */
1106
0
  if (OidIsValid(RelationGetReplicaIndex(rel)))
1107
0
    return;
1108
1109
  /* REPLICA IDENTITY FULL is also good for UPDATE/DELETE. */
1110
0
  if (rel->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
1111
0
    return;
1112
1113
  /*
1114
   * This is UPDATE/DELETE and there is no replica identity.
1115
   *
1116
   * Check if the table publishes UPDATES or DELETES.
1117
   */
1118
0
  if (cmd == CMD_UPDATE && pubdesc.pubactions.pubupdate)
1119
0
    ereport(ERROR,
1120
0
        (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1121
0
         errmsg("cannot update table \"%s\" because it does not have a replica identity and publishes updates",
1122
0
            RelationGetRelationName(rel)),
1123
0
         errhint("To enable updating the table, set REPLICA IDENTITY using ALTER TABLE.")));
1124
0
  else if (cmd == CMD_DELETE && pubdesc.pubactions.pubdelete)
1125
0
    ereport(ERROR,
1126
0
        (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1127
0
         errmsg("cannot delete from table \"%s\" because it does not have a replica identity and publishes deletes",
1128
0
            RelationGetRelationName(rel)),
1129
0
         errhint("To enable deleting from the table, set REPLICA IDENTITY using ALTER TABLE.")));
1130
0
}
1131
1132
1133
/*
1134
 * Check if we support writing into specific relkind of local relation and check
1135
 * if it aligns with the relkind of the relation on the publisher.
1136
 *
1137
 * The nspname and relname are only needed for error reporting.
1138
 */
1139
void
1140
CheckSubscriptionRelkind(char localrelkind, char remoterelkind,
1141
             const char *nspname, const char *relname)
1142
0
{
1143
0
  if (localrelkind != RELKIND_RELATION &&
1144
0
    localrelkind != RELKIND_PARTITIONED_TABLE &&
1145
0
    localrelkind != RELKIND_SEQUENCE)
1146
0
    ereport(ERROR,
1147
0
        (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1148
0
         errmsg("cannot use relation \"%s.%s\" as logical replication target",
1149
0
            nspname, relname),
1150
0
         errdetail_relkind_not_supported(localrelkind)));
1151
1152
  /*
1153
   * Allow RELKIND_RELATION and RELKIND_PARTITIONED_TABLE to be treated
1154
   * interchangeably, but ensure that sequences (RELKIND_SEQUENCE) match
1155
   * exactly on both publisher and subscriber.
1156
   */
1157
0
  if ((localrelkind == RELKIND_SEQUENCE && remoterelkind != RELKIND_SEQUENCE) ||
1158
0
    (localrelkind != RELKIND_SEQUENCE && remoterelkind == RELKIND_SEQUENCE))
1159
0
    ereport(ERROR,
1160
0
        errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1161
    /* translator: 3rd and 4th %s are "sequence" or "table" */
1162
0
        errmsg("relation \"%s.%s\" type mismatch: source \"%s\", target \"%s\"",
1163
0
             nspname, relname,
1164
0
             remoterelkind == RELKIND_SEQUENCE ? "sequence" : "table",
1165
0
             localrelkind == RELKIND_SEQUENCE ? "sequence" : "table"));
1166
0
}