Line data Source code
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 72136 : build_replindex_scan_key(ScanKey skey, Relation rel, Relation idxrel,
59 : TupleTableSlot *searchslot)
60 : {
61 : int index_attoff;
62 72136 : int skey_attoff = 0;
63 : Datum indclassDatum;
64 : oidvector *opclass;
65 72136 : int2vector *indkey = &idxrel->rd_index->indkey;
66 :
67 72136 : indclassDatum = SysCacheGetAttrNotNull(INDEXRELID, idxrel->rd_indextuple,
68 : Anum_pg_index_indclass);
69 72136 : opclass = (oidvector *) DatumGetPointer(indclassDatum);
70 :
71 : /* Build scankey for every non-expression attribute in the index. */
72 144305 : for (index_attoff = 0; index_attoff < IndexRelationGetNumberOfKeyAttributes(idxrel);
73 72169 : index_attoff++)
74 : {
75 : Oid operator;
76 : Oid optype;
77 : Oid opfamily;
78 : RegProcedure regop;
79 72169 : int table_attno = indkey->values[index_attoff];
80 : StrategyNumber eq_strategy;
81 :
82 72169 : if (!AttributeNumberIsValid(table_attno))
83 : {
84 : /*
85 : * XXX: Currently, we don't support expressions in the scan key,
86 : * see code below.
87 : */
88 2 : continue;
89 : }
90 :
91 : /*
92 : * Load the operator info. We need this to get the equality operator
93 : * function for the scan key.
94 : */
95 72167 : optype = get_opclass_input_type(opclass->values[index_attoff]);
96 72167 : opfamily = get_opclass_family(opclass->values[index_attoff]);
97 72167 : eq_strategy = IndexAmTranslateCompareType(COMPARE_EQ, idxrel->rd_rel->relam, opfamily, false);
98 72167 : operator = get_opfamily_member(opfamily, optype,
99 : optype,
100 : eq_strategy);
101 :
102 72167 : if (!OidIsValid(operator))
103 0 : elog(ERROR, "missing operator %d(%u,%u) in opfamily %u",
104 : eq_strategy, optype, optype, opfamily);
105 :
106 72167 : regop = get_opcode(operator);
107 :
108 : /* Initialize the scankey. */
109 72167 : ScanKeyInit(&skey[skey_attoff],
110 72167 : index_attoff + 1,
111 : eq_strategy,
112 : regop,
113 72167 : searchslot->tts_values[table_attno - 1]);
114 :
115 72167 : skey[skey_attoff].sk_collation = idxrel->rd_indcollation[index_attoff];
116 :
117 : /* Check for null value. */
118 72167 : if (searchslot->tts_isnull[table_attno - 1])
119 1 : skey[skey_attoff].sk_flags |= (SK_ISNULL | SK_SEARCHNULL);
120 :
121 72167 : skey_attoff++;
122 : }
123 :
124 : /* There must always be at least one attribute for the index scan. */
125 : Assert(skey_attoff > 0);
126 :
127 72136 : return skey_attoff;
128 : }
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 72340 : should_refetch_tuple(TM_Result res, TM_FailureData *tmfd)
138 : {
139 72340 : bool refetch = false;
140 :
141 72340 : switch (res)
142 : {
143 72340 : case TM_Ok:
144 72340 : break;
145 0 : case TM_Updated:
146 : /* XXX: Improve handling here */
147 0 : if (ItemPointerIndicatesMovedPartitions(&tmfd->ctid))
148 0 : ereport(LOG,
149 : (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
150 : errmsg("tuple to be locked was already moved to another partition due to concurrent update, retrying")));
151 : else
152 0 : ereport(LOG,
153 : (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
154 : 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 : (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
161 : 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 : break;
167 0 : default:
168 0 : elog(ERROR, "unexpected table_tuple_lock status: %u", res);
169 : break;
170 : }
171 :
172 72340 : return refetch;
173 : }
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 72135 : RelationFindReplTupleByIndex(Relation rel, Oid idxoid,
183 : LockTupleMode lockmode,
184 : TupleTableSlot *searchslot,
185 : TupleTableSlot *outslot)
186 : {
187 : ScanKeyData skey[INDEX_MAX_KEYS];
188 : int skey_attoff;
189 : IndexScanDesc scan;
190 : SnapshotData snap;
191 : TransactionId xwait;
192 : Relation idxrel;
193 : bool found;
194 72135 : TypeCacheEntry **eq = NULL;
195 : bool isIdxSafeToSkipDuplicates;
196 :
197 : /* Open the index. */
198 72135 : idxrel = index_open(idxoid, RowExclusiveLock);
199 :
200 72135 : isIdxSafeToSkipDuplicates = (GetRelationIdentityOrPK(rel) == idxoid);
201 :
202 72135 : InitDirtySnapshot(snap);
203 :
204 : /* Build scan key. */
205 72135 : skey_attoff = build_replindex_scan_key(skey, rel, idxrel, searchslot);
206 :
207 : /* Start an index scan. */
208 72135 : scan = index_beginscan(rel, idxrel,
209 : &snap, NULL, skey_attoff, 0, SO_NONE);
210 :
211 0 : retry:
212 72135 : found = false;
213 :
214 72135 : index_rescan(scan, skey, skey_attoff, NULL, 0);
215 :
216 : /* Try to find the tuple */
217 72135 : while (index_getnext_slot(scan, ForwardScanDirection, outslot))
218 : {
219 : /*
220 : * Avoid expensive equality check if the index is primary key or
221 : * replica identity index.
222 : */
223 72100 : if (!isIdxSafeToSkipDuplicates)
224 : {
225 19 : if (eq == NULL)
226 19 : eq = palloc0_array(TypeCacheEntry *, outslot->tts_tupleDescriptor->natts);
227 :
228 19 : if (!tuples_equal(outslot, searchslot, eq, NULL))
229 0 : continue;
230 : }
231 :
232 72100 : ExecMaterializeSlot(outslot);
233 :
234 144200 : xwait = TransactionIdIsValid(snap.xmin) ?
235 72100 : snap.xmin : snap.xmax;
236 :
237 : /*
238 : * If the tuple is locked, wait for locking transaction to finish and
239 : * retry.
240 : */
241 72100 : if (TransactionIdIsValid(xwait))
242 : {
243 0 : XactLockTableWait(xwait, NULL, NULL, XLTW_None);
244 0 : goto retry;
245 : }
246 :
247 : /* Found our tuple and it's not locked */
248 72100 : found = true;
249 72100 : break;
250 : }
251 :
252 : /* Found tuple, try to lock it in the lockmode. */
253 72135 : if (found)
254 : {
255 : TM_FailureData tmfd;
256 : TM_Result res;
257 :
258 72100 : PushActiveSnapshot(GetLatestSnapshot());
259 :
260 72100 : res = table_tuple_lock(rel, &(outslot->tts_tid), GetActiveSnapshot(),
261 : outslot,
262 : GetCurrentCommandId(false),
263 : lockmode,
264 : LockWaitBlock,
265 : 0 /* don't follow updates */ ,
266 : &tmfd);
267 :
268 72100 : PopActiveSnapshot();
269 :
270 72100 : if (should_refetch_tuple(res, &tmfd))
271 0 : goto retry;
272 : }
273 :
274 72135 : index_endscan(scan);
275 :
276 : /* Don't release lock until commit. */
277 72135 : index_close(idxrel, NoLock);
278 :
279 72135 : return found;
280 : }
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 105337 : tuples_equal(TupleTableSlot *slot1, TupleTableSlot *slot2,
290 : TypeCacheEntry **eq, Bitmapset *columns)
291 : {
292 : int attrnum;
293 :
294 : Assert(slot1->tts_tupleDescriptor->natts ==
295 : slot2->tts_tupleDescriptor->natts);
296 :
297 105337 : slot_getallattrs(slot1);
298 105337 : slot_getallattrs(slot2);
299 :
300 : /* Check equality of the attributes. */
301 105555 : for (attrnum = 0; attrnum < slot1->tts_tupleDescriptor->natts; attrnum++)
302 : {
303 : Form_pg_attribute att;
304 : TypeCacheEntry *typentry;
305 :
306 105385 : att = TupleDescAttr(slot1->tts_tupleDescriptor, attrnum);
307 :
308 : /*
309 : * Ignore dropped and generated columns as the publisher doesn't send
310 : * those
311 : */
312 105385 : if (att->attisdropped || att->attgenerated)
313 1 : continue;
314 :
315 : /*
316 : * Ignore columns that are not listed for checking.
317 : */
318 105384 : if (columns &&
319 0 : !bms_is_member(att->attnum - FirstLowInvalidHeapAttributeNumber,
320 : 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 105384 : 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 105384 : if (slot1->tts_isnull[attrnum] || slot2->tts_isnull[attrnum])
334 1 : continue;
335 :
336 105383 : typentry = eq[attrnum];
337 105383 : if (typentry == NULL)
338 : {
339 216 : typentry = lookup_type_cache(att->atttypid,
340 : TYPECACHE_EQ_OPR_FINFO);
341 216 : if (!OidIsValid(typentry->eq_opr_finfo.fn_oid))
342 0 : ereport(ERROR,
343 : (errcode(ERRCODE_UNDEFINED_FUNCTION),
344 : errmsg("could not identify an equality operator for type %s",
345 : format_type_be(att->atttypid))));
346 216 : eq[attrnum] = typentry;
347 : }
348 :
349 105383 : if (!DatumGetBool(FunctionCall2Coll(&typentry->eq_opr_finfo,
350 : att->attcollation,
351 105383 : slot1->tts_values[attrnum],
352 105383 : slot2->tts_values[attrnum])))
353 105167 : return false;
354 : }
355 :
356 170 : return true;
357 : }
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
370 153 : RelationFindReplTupleSeq(Relation rel, LockTupleMode lockmode,
371 : TupleTableSlot *searchslot, TupleTableSlot *outslot)
372 : {
373 : TupleTableSlot *scanslot;
374 : TableScanDesc scan;
375 : SnapshotData snap;
376 : TypeCacheEntry **eq;
377 : TransactionId xwait;
378 : bool found;
379 153 : TupleDesc desc PG_USED_FOR_ASSERTS_ONLY = RelationGetDescr(rel);
380 :
381 : Assert(equalTupleDescs(desc, outslot->tts_tupleDescriptor));
382 :
383 153 : eq = palloc0_array(TypeCacheEntry *, outslot->tts_tupleDescriptor->natts);
384 :
385 : /* Start a heap scan. */
386 153 : InitDirtySnapshot(snap);
387 153 : scan = table_beginscan(rel, &snap, 0, NULL,
388 : SO_NONE);
389 153 : scanslot = table_slot_create(rel, NULL);
390 :
391 0 : retry:
392 153 : found = false;
393 :
394 153 : table_rescan(scan, NULL);
395 :
396 : /* Try to find the tuple */
397 105319 : while (table_scan_getnextslot(scan, ForwardScanDirection, scanslot))
398 : {
399 105315 : if (!tuples_equal(scanslot, searchslot, eq, NULL))
400 105166 : continue;
401 :
402 149 : found = true;
403 149 : ExecCopySlot(outslot, scanslot);
404 :
405 298 : xwait = TransactionIdIsValid(snap.xmin) ?
406 149 : snap.xmin : snap.xmax;
407 :
408 : /*
409 : * If the tuple is locked, wait for locking transaction to finish and
410 : * retry.
411 : */
412 149 : if (TransactionIdIsValid(xwait))
413 : {
414 0 : XactLockTableWait(xwait, NULL, NULL, XLTW_None);
415 0 : goto retry;
416 : }
417 :
418 : /* Found our tuple and it's not locked */
419 149 : break;
420 : }
421 :
422 : /* Found tuple, try to lock it in the lockmode. */
423 153 : if (found)
424 : {
425 : TM_FailureData tmfd;
426 : TM_Result res;
427 :
428 149 : PushActiveSnapshot(GetLatestSnapshot());
429 :
430 149 : res = table_tuple_lock(rel, &(outslot->tts_tid), GetActiveSnapshot(),
431 : outslot,
432 : GetCurrentCommandId(false),
433 : lockmode,
434 : LockWaitBlock,
435 : 0 /* don't follow updates */ ,
436 : &tmfd);
437 :
438 149 : PopActiveSnapshot();
439 :
440 149 : if (should_refetch_tuple(res, &tmfd))
441 0 : goto retry;
442 : }
443 :
444 153 : table_endscan(scan);
445 153 : ExecDropSingleTupleTableSlot(scanslot);
446 :
447 153 : return found;
448 : }
449 :
450 : /*
451 : * Build additional index information necessary for conflict detection.
452 : */
453 : static void
454 94 : BuildConflictIndexInfo(ResultRelInfo *resultRelInfo, Oid conflictindex)
455 : {
456 284 : for (int i = 0; i < resultRelInfo->ri_NumIndices; i++)
457 : {
458 190 : Relation indexRelation = resultRelInfo->ri_IndexRelationDescs[i];
459 190 : IndexInfo *indexRelationInfo = resultRelInfo->ri_IndexRelationInfo[i];
460 :
461 190 : if (conflictindex != RelationGetRelid(indexRelation))
462 96 : continue;
463 :
464 : /*
465 : * This Assert will fail if BuildSpeculativeIndexInfo() is called
466 : * twice for the given index.
467 : */
468 : Assert(indexRelationInfo->ii_UniqueOps == NULL);
469 :
470 94 : BuildSpeculativeIndexInfo(indexRelation, indexRelationInfo);
471 : }
472 94 : }
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 3 : update_most_recent_deletion_info(TupleTableSlot *scanslot,
482 : TransactionId oldestxmin,
483 : TransactionId *delete_xid,
484 : TimestampTz *delete_time,
485 : ReplOriginId *delete_origin)
486 : {
487 : BufferHeapTupleTableSlot *hslot;
488 : HeapTuple tuple;
489 : Buffer buf;
490 3 : bool recently_dead = false;
491 : TransactionId xmax;
492 : TimestampTz localts;
493 : ReplOriginId localorigin;
494 :
495 3 : hslot = (BufferHeapTupleTableSlot *) scanslot;
496 :
497 3 : tuple = ExecFetchSlotHeapTuple(scanslot, false, NULL);
498 3 : buf = hslot->buffer;
499 :
500 3 : 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 3 : if (HeapTupleSatisfiesVacuum(tuple, oldestxmin, buf) == HEAPTUPLE_RECENTLY_DEAD)
510 3 : recently_dead = true;
511 :
512 3 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
513 :
514 3 : if (!recently_dead)
515 0 : return;
516 :
517 3 : xmax = HeapTupleHeaderGetUpdateXid(tuple->t_data);
518 3 : if (!TransactionIdIsValid(xmax))
519 0 : return;
520 :
521 : /* Select the dead tuple with the most recent commit timestamp */
522 6 : if (TransactionIdGetCommitTsData(xmax, &localts, &localorigin) &&
523 3 : TimestampDifferenceExceeds(*delete_time, localts, 0))
524 : {
525 3 : *delete_xid = xmax;
526 3 : *delete_time = localts;
527 3 : *delete_origin = localorigin;
528 : }
529 : }
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 2 : RelationFindDeletedTupleInfoSeq(Relation rel, TupleTableSlot *searchslot,
565 : TransactionId oldestxmin,
566 : TransactionId *delete_xid,
567 : ReplOriginId *delete_origin,
568 : TimestampTz *delete_time)
569 : {
570 : TupleTableSlot *scanslot;
571 : TableScanDesc scan;
572 : TypeCacheEntry **eq;
573 : Bitmapset *indexbitmap;
574 2 : TupleDesc desc PG_USED_FOR_ASSERTS_ONLY = RelationGetDescr(rel);
575 :
576 : Assert(equalTupleDescs(desc, searchslot->tts_tupleDescriptor));
577 :
578 2 : *delete_xid = InvalidTransactionId;
579 2 : *delete_origin = InvalidReplOriginId;
580 2 : *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 2 : indexbitmap = RelationGetIndexAttrBitmap(rel,
592 : INDEX_ATTR_BITMAP_IDENTITY_KEY);
593 :
594 : /* fallback to PK if no replica identity */
595 2 : if (!indexbitmap)
596 2 : indexbitmap = RelationGetIndexAttrBitmap(rel,
597 : INDEX_ATTR_BITMAP_PRIMARY_KEY);
598 :
599 2 : 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 2 : scan = table_beginscan(rel, SnapshotAny, 0, NULL,
608 : SO_NONE);
609 2 : scanslot = table_slot_create(rel, NULL);
610 :
611 2 : table_rescan(scan, NULL);
612 :
613 : /* Try to find the tuple */
614 5 : while (table_scan_getnextslot(scan, ForwardScanDirection, scanslot))
615 : {
616 3 : if (!tuples_equal(scanslot, searchslot, eq, indexbitmap))
617 1 : continue;
618 :
619 2 : update_most_recent_deletion_info(scanslot, oldestxmin, delete_xid,
620 : delete_time, delete_origin);
621 : }
622 :
623 2 : table_endscan(scan);
624 2 : ExecDropSingleTupleTableSlot(scanslot);
625 :
626 2 : return *delete_time != 0;
627 : }
628 :
629 : /*
630 : * Similar to RelationFindDeletedTupleInfoSeq() but using index scan to locate
631 : * the deleted tuple.
632 : */
633 : bool
634 1 : RelationFindDeletedTupleInfoByIndex(Relation rel, Oid idxoid,
635 : TupleTableSlot *searchslot,
636 : TransactionId oldestxmin,
637 : TransactionId *delete_xid,
638 : ReplOriginId *delete_origin,
639 : TimestampTz *delete_time)
640 : {
641 : Relation idxrel;
642 : ScanKeyData skey[INDEX_MAX_KEYS];
643 : int skey_attoff;
644 : IndexScanDesc scan;
645 : TupleTableSlot *scanslot;
646 1 : TypeCacheEntry **eq = NULL;
647 : bool isIdxSafeToSkipDuplicates;
648 1 : TupleDesc desc PG_USED_FOR_ASSERTS_ONLY = RelationGetDescr(rel);
649 :
650 : Assert(equalTupleDescs(desc, searchslot->tts_tupleDescriptor));
651 : Assert(OidIsValid(idxoid));
652 :
653 1 : *delete_xid = InvalidTransactionId;
654 1 : *delete_time = 0;
655 1 : *delete_origin = InvalidReplOriginId;
656 :
657 1 : isIdxSafeToSkipDuplicates = (GetRelationIdentityOrPK(rel) == idxoid);
658 :
659 1 : scanslot = table_slot_create(rel, NULL);
660 :
661 1 : idxrel = index_open(idxoid, RowExclusiveLock);
662 :
663 : /* Build scan key. */
664 1 : 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 1 : scan = index_beginscan(rel, idxrel,
673 : SnapshotAny, NULL, skey_attoff, 0, SO_NONE);
674 :
675 1 : index_rescan(scan, skey, skey_attoff, NULL, 0);
676 :
677 : /* Try to find the tuple */
678 2 : while (index_getnext_slot(scan, ForwardScanDirection, scanslot))
679 : {
680 : /*
681 : * Avoid expensive equality check if the index is primary key or
682 : * replica identity index.
683 : */
684 1 : if (!isIdxSafeToSkipDuplicates)
685 : {
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 : }
692 :
693 1 : update_most_recent_deletion_info(scanslot, oldestxmin, delete_xid,
694 : delete_time, delete_origin);
695 : }
696 :
697 1 : index_endscan(scan);
698 :
699 1 : index_close(idxrel, NoLock);
700 :
701 1 : ExecDropSingleTupleTableSlot(scanslot);
702 :
703 1 : return *delete_time != 0;
704 : }
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 94 : FindConflictTuple(ResultRelInfo *resultRelInfo, EState *estate,
717 : Oid conflictindex, TupleTableSlot *slot,
718 : TupleTableSlot **conflictslot)
719 : {
720 94 : Relation rel = resultRelInfo->ri_RelationDesc;
721 : ItemPointerData conflictTid;
722 : TM_FailureData tmfd;
723 : TM_Result res;
724 :
725 94 : *conflictslot = NULL;
726 :
727 : /*
728 : * Build additional information required to check constraints violations.
729 : * See check_exclusion_or_unique_constraint().
730 : */
731 94 : BuildConflictIndexInfo(resultRelInfo, conflictindex);
732 :
733 94 : retry:
734 187 : if (ExecCheckIndexConstraints(resultRelInfo, slot, estate,
735 94 : &conflictTid, &slot->tts_tid,
736 : list_make1_oid(conflictindex)))
737 : {
738 2 : if (*conflictslot)
739 0 : ExecDropSingleTupleTableSlot(*conflictslot);
740 :
741 2 : *conflictslot = NULL;
742 2 : return false;
743 : }
744 :
745 91 : *conflictslot = table_slot_create(rel, NULL);
746 :
747 91 : PushActiveSnapshot(GetLatestSnapshot());
748 :
749 91 : res = table_tuple_lock(rel, &conflictTid, GetActiveSnapshot(),
750 : *conflictslot,
751 : GetCurrentCommandId(false),
752 : LockTupleShare,
753 : LockWaitBlock,
754 : 0 /* don't follow updates */ ,
755 : &tmfd);
756 :
757 91 : PopActiveSnapshot();
758 :
759 91 : if (should_refetch_tuple(res, &tmfd))
760 0 : goto retry;
761 :
762 91 : return true;
763 : }
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 50 : CheckAndReportConflict(ResultRelInfo *resultRelInfo, EState *estate,
771 : ConflictType type, List *recheckIndexes,
772 : TupleTableSlot *searchslot, TupleTableSlot *remoteslot)
773 : {
774 50 : List *conflicttuples = NIL;
775 : TupleTableSlot *conflictslot;
776 :
777 : /* Check all the unique indexes for conflicts */
778 192 : foreach_oid(uniqueidx, resultRelInfo->ri_onConflictArbiterIndexes)
779 : {
780 187 : if (list_member_oid(recheckIndexes, uniqueidx) &&
781 94 : FindConflictTuple(resultRelInfo, estate, uniqueidx, remoteslot,
782 : &conflictslot))
783 : {
784 91 : ConflictTupleInfo *conflicttuple = palloc0_object(ConflictTupleInfo);
785 :
786 91 : conflicttuple->slot = conflictslot;
787 91 : conflicttuple->indexoid = uniqueidx;
788 :
789 91 : GetTupleTransactionInfo(conflictslot, &conflicttuple->xmin,
790 : &conflicttuple->origin, &conflicttuple->ts);
791 :
792 91 : conflicttuples = lappend(conflicttuples, conflicttuple);
793 : }
794 : }
795 :
796 : /* Report the conflict, if found */
797 49 : if (conflicttuples)
798 47 : ReportApplyConflict(estate, resultRelInfo, ERROR,
799 47 : list_length(conflicttuples) > 1 ? CT_MULTIPLE_UNIQUE_CONFLICTS : type,
800 : searchslot, remoteslot, conflicttuples);
801 2 : }
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 96320 : ExecSimpleRelationInsert(ResultRelInfo *resultRelInfo,
811 : EState *estate, TupleTableSlot *slot)
812 : {
813 96320 : bool skip_tuple = false;
814 96320 : Relation rel = resultRelInfo->ri_RelationDesc;
815 :
816 : /* For now we support only tables. */
817 : Assert(rel->rd_rel->relkind == RELKIND_RELATION);
818 :
819 96320 : CheckCmdReplicaIdentity(rel, CMD_INSERT);
820 :
821 : /* BEFORE ROW INSERT Triggers */
822 96320 : if (resultRelInfo->ri_TrigDesc &&
823 20 : resultRelInfo->ri_TrigDesc->trig_insert_before_row)
824 : {
825 3 : if (!ExecBRInsertTriggers(estate, resultRelInfo, slot))
826 1 : skip_tuple = true; /* "do nothing" */
827 : }
828 :
829 96320 : if (!skip_tuple)
830 : {
831 96319 : List *recheckIndexes = NIL;
832 : List *conflictindexes;
833 96319 : bool conflict = false;
834 :
835 : /* Compute stored generated columns */
836 96319 : if (rel->rd_att->constr &&
837 65812 : rel->rd_att->constr->has_generated_stored)
838 4 : ExecComputeStoredGenerated(resultRelInfo, estate, slot,
839 : CMD_INSERT);
840 :
841 : /* Check the constraints of the tuple */
842 96319 : if (rel->rd_att->constr)
843 65812 : ExecConstraints(resultRelInfo, slot, estate);
844 96319 : if (rel->rd_rel->relispartition)
845 96 : ExecPartitionCheck(resultRelInfo, slot, estate, true);
846 :
847 : /* OK, store the tuple and create index entries for it */
848 96319 : simple_table_tuple_insert(resultRelInfo->ri_RelationDesc, slot);
849 :
850 96319 : conflictindexes = resultRelInfo->ri_onConflictArbiterIndexes;
851 :
852 96319 : if (resultRelInfo->ri_NumIndices > 0)
853 : {
854 : uint32 flags;
855 :
856 80887 : if (conflictindexes != NIL)
857 80883 : flags = EIIT_NO_DUPE_ERROR;
858 : else
859 4 : flags = 0;
860 80887 : recheckIndexes = ExecInsertIndexTuples(resultRelInfo,
861 : estate, flags,
862 : slot, conflictindexes,
863 : &conflict);
864 : }
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 96319 : if (conflict)
882 47 : CheckAndReportConflict(resultRelInfo, estate, CT_INSERT_EXISTS,
883 : recheckIndexes, NULL, slot);
884 :
885 : /* AFTER ROW INSERT Triggers */
886 96274 : ExecARInsertTriggers(estate, resultRelInfo, slot,
887 : 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 96274 : list_free(recheckIndexes);
896 : }
897 96275 : }
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 31930 : ExecSimpleRelationUpdate(ResultRelInfo *resultRelInfo,
907 : EState *estate, EPQState *epqstate,
908 : TupleTableSlot *searchslot, TupleTableSlot *slot)
909 : {
910 31930 : bool skip_tuple = false;
911 31930 : Relation rel = resultRelInfo->ri_RelationDesc;
912 31930 : ItemPointer tid = &(searchslot->tts_tid);
913 :
914 : /*
915 : * We support only non-system tables, with
916 : * check_publication_add_relation() accountable.
917 : */
918 : Assert(rel->rd_rel->relkind == RELKIND_RELATION);
919 : Assert(!IsCatalogRelation(rel));
920 :
921 31930 : CheckCmdReplicaIdentity(rel, CMD_UPDATE);
922 :
923 : /* BEFORE ROW UPDATE Triggers */
924 31930 : if (resultRelInfo->ri_TrigDesc &&
925 10 : resultRelInfo->ri_TrigDesc->trig_update_before_row)
926 : {
927 3 : if (!ExecBRUpdateTriggers(estate, epqstate, resultRelInfo,
928 : tid, NULL, slot, NULL, NULL, false))
929 2 : skip_tuple = true; /* "do nothing" */
930 : }
931 :
932 31930 : if (!skip_tuple)
933 : {
934 31928 : List *recheckIndexes = NIL;
935 : TU_UpdateIndexes update_indexes;
936 : List *conflictindexes;
937 31928 : bool conflict = false;
938 :
939 : /* Compute stored generated columns */
940 31928 : if (rel->rd_att->constr &&
941 31881 : rel->rd_att->constr->has_generated_stored)
942 2 : ExecComputeStoredGenerated(resultRelInfo, estate, slot,
943 : CMD_UPDATE);
944 :
945 : /* Check the constraints of the tuple */
946 31928 : if (rel->rd_att->constr)
947 31881 : ExecConstraints(resultRelInfo, slot, estate);
948 31928 : if (rel->rd_rel->relispartition)
949 12 : ExecPartitionCheck(resultRelInfo, slot, estate, true);
950 :
951 31928 : simple_table_tuple_update(rel, tid, slot, estate->es_snapshot,
952 : &update_indexes);
953 :
954 31928 : conflictindexes = resultRelInfo->ri_onConflictArbiterIndexes;
955 :
956 31928 : if (resultRelInfo->ri_NumIndices > 0 && (update_indexes != TU_None))
957 : {
958 20209 : uint32 flags = EIIT_IS_UPDATE;
959 :
960 20209 : if (conflictindexes != NIL)
961 20200 : flags |= EIIT_NO_DUPE_ERROR;
962 20209 : if (update_indexes == TU_Summarizing)
963 0 : flags |= EIIT_ONLY_SUMMARIZING;
964 20209 : recheckIndexes = ExecInsertIndexTuples(resultRelInfo,
965 : estate, flags,
966 : slot, conflictindexes,
967 : &conflict);
968 : }
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 31928 : if (conflict)
976 3 : CheckAndReportConflict(resultRelInfo, estate, CT_UPDATE_EXISTS,
977 : recheckIndexes, searchslot, slot);
978 :
979 : /* AFTER ROW UPDATE Triggers */
980 31925 : ExecARUpdateTriggers(estate, resultRelInfo,
981 : NULL, NULL,
982 : tid, NULL, slot,
983 : recheckIndexes, NULL, false);
984 :
985 31925 : list_free(recheckIndexes);
986 : }
987 31927 : }
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 40319 : ExecSimpleRelationDelete(ResultRelInfo *resultRelInfo,
997 : EState *estate, EPQState *epqstate,
998 : TupleTableSlot *searchslot)
999 : {
1000 40319 : bool skip_tuple = false;
1001 40319 : Relation rel = resultRelInfo->ri_RelationDesc;
1002 40319 : ItemPointer tid = &searchslot->tts_tid;
1003 :
1004 40319 : CheckCmdReplicaIdentity(rel, CMD_DELETE);
1005 :
1006 : /* BEFORE ROW DELETE Triggers */
1007 40319 : if (resultRelInfo->ri_TrigDesc &&
1008 10 : resultRelInfo->ri_TrigDesc->trig_delete_before_row)
1009 : {
1010 0 : skip_tuple = !ExecBRDeleteTriggers(estate, epqstate, resultRelInfo,
1011 0 : tid, NULL, NULL, NULL, NULL, false);
1012 : }
1013 :
1014 40319 : if (!skip_tuple)
1015 : {
1016 : /* OK, delete the tuple */
1017 40319 : simple_table_tuple_delete(rel, tid, estate->es_snapshot);
1018 :
1019 : /* AFTER ROW DELETE Triggers */
1020 40319 : ExecARDeleteTriggers(estate, resultRelInfo,
1021 : tid, NULL, NULL, false);
1022 : }
1023 40319 : }
1024 :
1025 : /*
1026 : * Check if command can be executed with current replica identity.
1027 : */
1028 : void
1029 254095 : CheckCmdReplicaIdentity(Relation rel, CmdType cmd)
1030 : {
1031 : 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 254095 : if (rel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
1038 238297 : return;
1039 :
1040 : /* We only need to do checks for UPDATE and DELETE. */
1041 250004 : if (cmd != CMD_UPDATE && cmd != CMD_DELETE)
1042 156838 : 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 93166 : RelationBuildPublicationDesc(rel, &pubdesc);
1068 93166 : if (cmd == CMD_UPDATE && !pubdesc.rf_valid_for_update)
1069 40 : ereport(ERROR,
1070 : (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1071 : errmsg("cannot update table \"%s\"",
1072 : RelationGetRelationName(rel)),
1073 : errdetail("Column used in the publication WHERE expression is not part of the replica identity.")));
1074 93126 : else if (cmd == CMD_UPDATE && !pubdesc.cols_valid_for_update)
1075 72 : ereport(ERROR,
1076 : (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1077 : errmsg("cannot update table \"%s\"",
1078 : RelationGetRelationName(rel)),
1079 : errdetail("Column list used by the publication does not cover the replica identity.")));
1080 93054 : else if (cmd == CMD_UPDATE && !pubdesc.gencols_valid_for_update)
1081 16 : ereport(ERROR,
1082 : (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1083 : errmsg("cannot update table \"%s\"",
1084 : RelationGetRelationName(rel)),
1085 : errdetail("Replica identity must not contain unpublished generated columns.")));
1086 93038 : else if (cmd == CMD_DELETE && !pubdesc.rf_valid_for_delete)
1087 0 : ereport(ERROR,
1088 : (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1089 : errmsg("cannot delete from table \"%s\"",
1090 : RelationGetRelationName(rel)),
1091 : errdetail("Column used in the publication WHERE expression is not part of the replica identity.")));
1092 93038 : else if (cmd == CMD_DELETE && !pubdesc.cols_valid_for_delete)
1093 0 : ereport(ERROR,
1094 : (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1095 : errmsg("cannot delete from table \"%s\"",
1096 : RelationGetRelationName(rel)),
1097 : errdetail("Column list used by the publication does not cover the replica identity.")));
1098 93038 : else if (cmd == CMD_DELETE && !pubdesc.gencols_valid_for_delete)
1099 0 : ereport(ERROR,
1100 : (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1101 : errmsg("cannot delete from table \"%s\"",
1102 : RelationGetRelationName(rel)),
1103 : errdetail("Replica identity must not contain unpublished generated columns.")));
1104 :
1105 : /* If relation has replica identity we are always good. */
1106 93038 : if (OidIsValid(RelationGetReplicaIndex(rel)))
1107 77120 : return;
1108 :
1109 : /* REPLICA IDENTITY FULL is also good for UPDATE/DELETE. */
1110 15918 : if (rel->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
1111 248 : 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 15670 : if (cmd == CMD_UPDATE && pubdesc.pubactions.pubupdate)
1119 81 : ereport(ERROR,
1120 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1121 : errmsg("cannot update table \"%s\" because it does not have a replica identity and publishes updates",
1122 : RelationGetRelationName(rel)),
1123 : errhint("To enable updating the table, set REPLICA IDENTITY using ALTER TABLE.")));
1124 15589 : else if (cmd == CMD_DELETE && pubdesc.pubactions.pubdelete)
1125 14 : ereport(ERROR,
1126 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1127 : errmsg("cannot delete from table \"%s\" because it does not have a replica identity and publishes deletes",
1128 : RelationGetRelationName(rel)),
1129 : errhint("To enable deleting from the table, set REPLICA IDENTITY using ALTER TABLE.")));
1130 : }
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 1049 : CheckSubscriptionRelkind(char localrelkind, char remoterelkind,
1141 : const char *nspname, const char *relname)
1142 : {
1143 1049 : if (localrelkind != RELKIND_RELATION &&
1144 17 : localrelkind != RELKIND_PARTITIONED_TABLE &&
1145 : localrelkind != RELKIND_SEQUENCE)
1146 0 : ereport(ERROR,
1147 : (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1148 : errmsg("cannot use relation \"%s.%s\" as logical replication target",
1149 : nspname, relname),
1150 : 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 1049 : if ((localrelkind == RELKIND_SEQUENCE && remoterelkind != RELKIND_SEQUENCE) ||
1158 1032 : (localrelkind != RELKIND_SEQUENCE && remoterelkind == RELKIND_SEQUENCE))
1159 0 : ereport(ERROR,
1160 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1161 : /* translator: 3rd and 4th %s are "sequence" or "table" */
1162 : errmsg("relation \"%s.%s\" type mismatch: source \"%s\", target \"%s\"",
1163 : nspname, relname,
1164 : remoterelkind == RELKIND_SEQUENCE ? "sequence" : "table",
1165 : localrelkind == RELKIND_SEQUENCE ? "sequence" : "table"));
1166 1049 : }
|
