Line data Source code
1 : /*-------------------------------------------------------------------------
2 : * tablesync.c
3 : * PostgreSQL logical replication: initial table data synchronization
4 : *
5 : * Copyright (c) 2012-2025, PostgreSQL Global Development Group
6 : *
7 : * IDENTIFICATION
8 : * src/backend/replication/logical/tablesync.c
9 : *
10 : * NOTES
11 : * This file contains code for initial table data synchronization for
12 : * logical replication.
13 : *
14 : * The initial data synchronization is done separately for each table,
15 : * in a separate apply worker that only fetches the initial snapshot data
16 : * from the publisher and then synchronizes the position in the stream with
17 : * the leader apply worker.
18 : *
19 : * There are several reasons for doing the synchronization this way:
20 : * - It allows us to parallelize the initial data synchronization
21 : * which lowers the time needed for it to happen.
22 : * - The initial synchronization does not have to hold the xid and LSN
23 : * for the time it takes to copy data of all tables, causing less
24 : * bloat and lower disk consumption compared to doing the
25 : * synchronization in a single process for the whole database.
26 : * - It allows us to synchronize any tables added after the initial
27 : * synchronization has finished.
28 : *
29 : * The stream position synchronization works in multiple steps:
30 : * - Apply worker requests a tablesync worker to start, setting the new
31 : * table state to INIT.
32 : * - Tablesync worker starts; changes table state from INIT to DATASYNC while
33 : * copying.
34 : * - Tablesync worker does initial table copy; there is a FINISHEDCOPY (sync
35 : * worker specific) state to indicate when the copy phase has completed, so
36 : * if the worker crashes with this (non-memory) state then the copy will not
37 : * be re-attempted.
38 : * - Tablesync worker then sets table state to SYNCWAIT; waits for state change.
39 : * - Apply worker periodically checks for tables in SYNCWAIT state. When
40 : * any appear, it sets the table state to CATCHUP and starts loop-waiting
41 : * until either the table state is set to SYNCDONE or the sync worker
42 : * exits.
43 : * - After the sync worker has seen the state change to CATCHUP, it will
44 : * read the stream and apply changes (acting like an apply worker) until
45 : * it catches up to the specified stream position. Then it sets the
46 : * state to SYNCDONE. There might be zero changes applied between
47 : * CATCHUP and SYNCDONE, because the sync worker might be ahead of the
48 : * apply worker.
49 : * - Once the state is set to SYNCDONE, the apply will continue tracking
50 : * the table until it reaches the SYNCDONE stream position, at which
51 : * point it sets state to READY and stops tracking. Again, there might
52 : * be zero changes in between.
53 : *
54 : * So the state progression is always: INIT -> DATASYNC -> FINISHEDCOPY
55 : * -> SYNCWAIT -> CATCHUP -> SYNCDONE -> READY.
56 : *
57 : * The catalog pg_subscription_rel is used to keep information about
58 : * subscribed tables and their state. The catalog holds all states
59 : * except SYNCWAIT and CATCHUP which are only in shared memory.
60 : *
61 : * Example flows look like this:
62 : * - Apply is in front:
63 : * sync:8
64 : * -> set in catalog FINISHEDCOPY
65 : * -> set in memory SYNCWAIT
66 : * apply:10
67 : * -> set in memory CATCHUP
68 : * -> enter wait-loop
69 : * sync:10
70 : * -> set in catalog SYNCDONE
71 : * -> exit
72 : * apply:10
73 : * -> exit wait-loop
74 : * -> continue rep
75 : * apply:11
76 : * -> set in catalog READY
77 : *
78 : * - Sync is in front:
79 : * sync:10
80 : * -> set in catalog FINISHEDCOPY
81 : * -> set in memory SYNCWAIT
82 : * apply:8
83 : * -> set in memory CATCHUP
84 : * -> continue per-table filtering
85 : * sync:10
86 : * -> set in catalog SYNCDONE
87 : * -> exit
88 : * apply:10
89 : * -> set in catalog READY
90 : * -> stop per-table filtering
91 : * -> continue rep
92 : *-------------------------------------------------------------------------
93 : */
94 :
95 : #include "postgres.h"
96 :
97 : #include "access/table.h"
98 : #include "access/xact.h"
99 : #include "catalog/indexing.h"
100 : #include "catalog/pg_subscription_rel.h"
101 : #include "catalog/pg_type.h"
102 : #include "commands/copy.h"
103 : #include "miscadmin.h"
104 : #include "nodes/makefuncs.h"
105 : #include "parser/parse_relation.h"
106 : #include "pgstat.h"
107 : #include "replication/logicallauncher.h"
108 : #include "replication/logicalrelation.h"
109 : #include "replication/logicalworker.h"
110 : #include "replication/origin.h"
111 : #include "replication/slot.h"
112 : #include "replication/walreceiver.h"
113 : #include "replication/worker_internal.h"
114 : #include "storage/ipc.h"
115 : #include "storage/lmgr.h"
116 : #include "utils/acl.h"
117 : #include "utils/array.h"
118 : #include "utils/builtins.h"
119 : #include "utils/lsyscache.h"
120 : #include "utils/rls.h"
121 : #include "utils/snapmgr.h"
122 : #include "utils/syscache.h"
123 : #include "utils/usercontext.h"
124 :
125 : List *table_states_not_ready = NIL;
126 :
127 : static StringInfo copybuf = NULL;
128 :
129 : /*
130 : * Wait until the relation sync state is set in the catalog to the expected
131 : * one; return true when it happens.
132 : *
133 : * Returns false if the table sync worker or the table itself have
134 : * disappeared, or the table state has been reset.
135 : *
136 : * Currently, this is used in the apply worker when transitioning from
137 : * CATCHUP state to SYNCDONE.
138 : */
139 : static bool
140 364 : wait_for_table_state_change(Oid relid, char expected_state)
141 : {
142 : char state;
143 :
144 : for (;;)
145 442 : {
146 : LogicalRepWorker *worker;
147 : XLogRecPtr statelsn;
148 :
149 806 : CHECK_FOR_INTERRUPTS();
150 :
151 806 : InvalidateCatalogSnapshot();
152 806 : state = GetSubscriptionRelState(MyLogicalRepWorker->subid,
153 : relid, &statelsn);
154 :
155 806 : if (state == SUBREL_STATE_UNKNOWN)
156 0 : break;
157 :
158 806 : if (state == expected_state)
159 364 : return true;
160 :
161 : /* Check if the sync worker is still running and bail if not. */
162 442 : LWLockAcquire(LogicalRepWorkerLock, LW_SHARED);
163 442 : worker = logicalrep_worker_find(WORKERTYPE_TABLESYNC,
164 442 : MyLogicalRepWorker->subid, relid,
165 : false);
166 442 : LWLockRelease(LogicalRepWorkerLock);
167 442 : if (!worker)
168 0 : break;
169 :
170 442 : (void) WaitLatch(MyLatch,
171 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
172 : 1000L, WAIT_EVENT_LOGICAL_SYNC_STATE_CHANGE);
173 :
174 442 : ResetLatch(MyLatch);
175 : }
176 :
177 0 : return false;
178 : }
179 :
180 : /*
181 : * Wait until the apply worker changes the state of our synchronization
182 : * worker to the expected one.
183 : *
184 : * Used when transitioning from SYNCWAIT state to CATCHUP.
185 : *
186 : * Returns false if the apply worker has disappeared.
187 : */
188 : static bool
189 368 : wait_for_worker_state_change(char expected_state)
190 : {
191 : int rc;
192 :
193 : for (;;)
194 370 : {
195 : LogicalRepWorker *worker;
196 :
197 738 : CHECK_FOR_INTERRUPTS();
198 :
199 : /*
200 : * Done if already in correct state. (We assume this fetch is atomic
201 : * enough to not give a misleading answer if we do it with no lock.)
202 : */
203 738 : if (MyLogicalRepWorker->relstate == expected_state)
204 368 : return true;
205 :
206 : /*
207 : * Bail out if the apply worker has died, else signal it we're
208 : * waiting.
209 : */
210 370 : LWLockAcquire(LogicalRepWorkerLock, LW_SHARED);
211 370 : worker = logicalrep_worker_find(WORKERTYPE_APPLY,
212 370 : MyLogicalRepWorker->subid, InvalidOid,
213 : false);
214 370 : if (worker && worker->proc)
215 370 : logicalrep_worker_wakeup_ptr(worker);
216 370 : LWLockRelease(LogicalRepWorkerLock);
217 370 : if (!worker)
218 0 : break;
219 :
220 : /*
221 : * Wait. We expect to get a latch signal back from the apply worker,
222 : * but use a timeout in case it dies without sending one.
223 : */
224 370 : rc = WaitLatch(MyLatch,
225 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
226 : 1000L, WAIT_EVENT_LOGICAL_SYNC_STATE_CHANGE);
227 :
228 370 : if (rc & WL_LATCH_SET)
229 370 : ResetLatch(MyLatch);
230 : }
231 :
232 0 : return false;
233 : }
234 :
235 : /*
236 : * Handle table synchronization cooperation from the synchronization
237 : * worker.
238 : *
239 : * If the sync worker is in CATCHUP state and reached (or passed) the
240 : * predetermined synchronization point in the WAL stream, mark the table as
241 : * SYNCDONE and finish.
242 : */
243 : void
244 452 : ProcessSyncingTablesForSync(XLogRecPtr current_lsn)
245 : {
246 452 : SpinLockAcquire(&MyLogicalRepWorker->relmutex);
247 :
248 452 : if (MyLogicalRepWorker->relstate == SUBREL_STATE_CATCHUP &&
249 452 : current_lsn >= MyLogicalRepWorker->relstate_lsn)
250 : {
251 : TimeLineID tli;
252 368 : char syncslotname[NAMEDATALEN] = {0};
253 368 : char originname[NAMEDATALEN] = {0};
254 :
255 368 : MyLogicalRepWorker->relstate = SUBREL_STATE_SYNCDONE;
256 368 : MyLogicalRepWorker->relstate_lsn = current_lsn;
257 :
258 368 : SpinLockRelease(&MyLogicalRepWorker->relmutex);
259 :
260 : /*
261 : * UpdateSubscriptionRelState must be called within a transaction.
262 : */
263 368 : if (!IsTransactionState())
264 368 : StartTransactionCommand();
265 :
266 368 : UpdateSubscriptionRelState(MyLogicalRepWorker->subid,
267 368 : MyLogicalRepWorker->relid,
268 368 : MyLogicalRepWorker->relstate,
269 368 : MyLogicalRepWorker->relstate_lsn,
270 : false);
271 :
272 : /*
273 : * End streaming so that LogRepWorkerWalRcvConn can be used to drop
274 : * the slot.
275 : */
276 368 : walrcv_endstreaming(LogRepWorkerWalRcvConn, &tli);
277 :
278 : /*
279 : * Cleanup the tablesync slot.
280 : *
281 : * This has to be done after updating the state because otherwise if
282 : * there is an error while doing the database operations we won't be
283 : * able to rollback dropped slot.
284 : */
285 368 : ReplicationSlotNameForTablesync(MyLogicalRepWorker->subid,
286 368 : MyLogicalRepWorker->relid,
287 : syncslotname,
288 : sizeof(syncslotname));
289 :
290 : /*
291 : * It is important to give an error if we are unable to drop the slot,
292 : * otherwise, it won't be dropped till the corresponding subscription
293 : * is dropped. So passing missing_ok = false.
294 : */
295 368 : ReplicationSlotDropAtPubNode(LogRepWorkerWalRcvConn, syncslotname, false);
296 :
297 368 : CommitTransactionCommand();
298 368 : pgstat_report_stat(false);
299 :
300 : /*
301 : * Start a new transaction to clean up the tablesync origin tracking.
302 : * This transaction will be ended within the FinishSyncWorker(). Now,
303 : * even, if we fail to remove this here, the apply worker will ensure
304 : * to clean it up afterward.
305 : *
306 : * We need to do this after the table state is set to SYNCDONE.
307 : * Otherwise, if an error occurs while performing the database
308 : * operation, the worker will be restarted and the in-memory state of
309 : * replication progress (remote_lsn) won't be rolled-back which would
310 : * have been cleared before restart. So, the restarted worker will use
311 : * invalid replication progress state resulting in replay of
312 : * transactions that have already been applied.
313 : */
314 368 : StartTransactionCommand();
315 :
316 368 : ReplicationOriginNameForLogicalRep(MyLogicalRepWorker->subid,
317 368 : MyLogicalRepWorker->relid,
318 : originname,
319 : sizeof(originname));
320 :
321 : /*
322 : * Resetting the origin session removes the ownership of the slot.
323 : * This is needed to allow the origin to be dropped.
324 : */
325 368 : replorigin_session_reset();
326 368 : replorigin_session_origin = InvalidRepOriginId;
327 368 : replorigin_session_origin_lsn = InvalidXLogRecPtr;
328 368 : replorigin_session_origin_timestamp = 0;
329 :
330 : /*
331 : * Drop the tablesync's origin tracking if exists.
332 : *
333 : * There is a chance that the user is concurrently performing refresh
334 : * for the subscription where we remove the table state and its origin
335 : * or the apply worker would have removed this origin. So passing
336 : * missing_ok = true.
337 : */
338 368 : replorigin_drop_by_name(originname, true, false);
339 :
340 368 : FinishSyncWorker();
341 : }
342 : else
343 84 : SpinLockRelease(&MyLogicalRepWorker->relmutex);
344 84 : }
345 :
346 : /*
347 : * Handle table synchronization cooperation from the apply worker.
348 : *
349 : * Walk over all subscription tables that are individually tracked by the
350 : * apply process (currently, all that have state other than
351 : * SUBREL_STATE_READY) and manage synchronization for them.
352 : *
353 : * If there are tables that need synchronizing and are not being synchronized
354 : * yet, start sync workers for them (if there are free slots for sync
355 : * workers). To prevent starting the sync worker for the same relation at a
356 : * high frequency after a failure, we store its last start time with each sync
357 : * state info. We start the sync worker for the same relation after waiting
358 : * at least wal_retrieve_retry_interval.
359 : *
360 : * For tables that are being synchronized already, check if sync workers
361 : * either need action from the apply worker or have finished. This is the
362 : * SYNCWAIT to CATCHUP transition.
363 : *
364 : * If the synchronization position is reached (SYNCDONE), then the table can
365 : * be marked as READY and is no longer tracked.
366 : */
367 : void
368 21638 : ProcessSyncingTablesForApply(XLogRecPtr current_lsn)
369 : {
370 : struct tablesync_start_time_mapping
371 : {
372 : Oid relid;
373 : TimestampTz last_start_time;
374 : };
375 : static HTAB *last_start_times = NULL;
376 : ListCell *lc;
377 : bool started_tx;
378 21638 : bool should_exit = false;
379 21638 : Relation rel = NULL;
380 :
381 : Assert(!IsTransactionState());
382 :
383 : /* We need up-to-date sync state info for subscription tables here. */
384 21638 : FetchRelationStates(NULL, NULL, &started_tx);
385 :
386 : /*
387 : * Prepare a hash table for tracking last start times of workers, to avoid
388 : * immediate restarts. We don't need it if there are no tables that need
389 : * syncing.
390 : */
391 21638 : if (table_states_not_ready != NIL && !last_start_times)
392 242 : {
393 : HASHCTL ctl;
394 :
395 242 : ctl.keysize = sizeof(Oid);
396 242 : ctl.entrysize = sizeof(struct tablesync_start_time_mapping);
397 242 : last_start_times = hash_create("Logical replication table sync worker start times",
398 : 256, &ctl, HASH_ELEM | HASH_BLOBS);
399 : }
400 :
401 : /*
402 : * Clean up the hash table when we're done with all tables (just to
403 : * release the bit of memory).
404 : */
405 21396 : else if (table_states_not_ready == NIL && last_start_times)
406 : {
407 180 : hash_destroy(last_start_times);
408 180 : last_start_times = NULL;
409 : }
410 :
411 : /*
412 : * Process all tables that are being synchronized.
413 : */
414 24974 : foreach(lc, table_states_not_ready)
415 : {
416 3340 : SubscriptionRelState *rstate = (SubscriptionRelState *) lfirst(lc);
417 :
418 3340 : if (!started_tx)
419 : {
420 566 : StartTransactionCommand();
421 566 : started_tx = true;
422 : }
423 :
424 : Assert(get_rel_relkind(rstate->relid) != RELKIND_SEQUENCE);
425 :
426 3340 : if (rstate->state == SUBREL_STATE_SYNCDONE)
427 : {
428 : /*
429 : * Apply has caught up to the position where the table sync has
430 : * finished. Mark the table as ready so that the apply will just
431 : * continue to replicate it normally.
432 : */
433 362 : if (current_lsn >= rstate->lsn)
434 : {
435 : char originname[NAMEDATALEN];
436 :
437 360 : rstate->state = SUBREL_STATE_READY;
438 360 : rstate->lsn = current_lsn;
439 :
440 : /*
441 : * Remove the tablesync origin tracking if exists.
442 : *
443 : * There is a chance that the user is concurrently performing
444 : * refresh for the subscription where we remove the table
445 : * state and its origin or the tablesync worker would have
446 : * already removed this origin. We can't rely on tablesync
447 : * worker to remove the origin tracking as if there is any
448 : * error while dropping we won't restart it to drop the
449 : * origin. So passing missing_ok = true.
450 : *
451 : * Lock the subscription and origin in the same order as we
452 : * are doing during DDL commands to avoid deadlocks. See
453 : * AlterSubscription_refresh.
454 : */
455 360 : LockSharedObject(SubscriptionRelationId, MyLogicalRepWorker->subid,
456 : 0, AccessShareLock);
457 :
458 360 : if (!rel)
459 358 : rel = table_open(SubscriptionRelRelationId, RowExclusiveLock);
460 :
461 360 : ReplicationOriginNameForLogicalRep(MyLogicalRepWorker->subid,
462 : rstate->relid,
463 : originname,
464 : sizeof(originname));
465 360 : replorigin_drop_by_name(originname, true, false);
466 :
467 : /*
468 : * Update the state to READY only after the origin cleanup.
469 : */
470 360 : UpdateSubscriptionRelState(MyLogicalRepWorker->subid,
471 360 : rstate->relid, rstate->state,
472 : rstate->lsn, true);
473 : }
474 : }
475 : else
476 : {
477 : LogicalRepWorker *syncworker;
478 :
479 : /*
480 : * Look for a sync worker for this relation.
481 : */
482 2978 : LWLockAcquire(LogicalRepWorkerLock, LW_SHARED);
483 :
484 2978 : syncworker = logicalrep_worker_find(WORKERTYPE_TABLESYNC,
485 2978 : MyLogicalRepWorker->subid,
486 : rstate->relid, false);
487 :
488 2978 : if (syncworker)
489 : {
490 : /* Found one, update our copy of its state */
491 1366 : SpinLockAcquire(&syncworker->relmutex);
492 1366 : rstate->state = syncworker->relstate;
493 1366 : rstate->lsn = syncworker->relstate_lsn;
494 1366 : if (rstate->state == SUBREL_STATE_SYNCWAIT)
495 : {
496 : /*
497 : * Sync worker is waiting for apply. Tell sync worker it
498 : * can catchup now.
499 : */
500 364 : syncworker->relstate = SUBREL_STATE_CATCHUP;
501 364 : syncworker->relstate_lsn =
502 364 : Max(syncworker->relstate_lsn, current_lsn);
503 : }
504 1366 : SpinLockRelease(&syncworker->relmutex);
505 :
506 : /* If we told worker to catch up, wait for it. */
507 1366 : if (rstate->state == SUBREL_STATE_SYNCWAIT)
508 : {
509 : /* Signal the sync worker, as it may be waiting for us. */
510 364 : if (syncworker->proc)
511 364 : logicalrep_worker_wakeup_ptr(syncworker);
512 :
513 : /* Now safe to release the LWLock */
514 364 : LWLockRelease(LogicalRepWorkerLock);
515 :
516 364 : if (started_tx)
517 : {
518 : /*
519 : * We must commit the existing transaction to release
520 : * the existing locks before entering a busy loop.
521 : * This is required to avoid any undetected deadlocks
522 : * due to any existing lock as deadlock detector won't
523 : * be able to detect the waits on the latch.
524 : *
525 : * Also close any tables prior to the commit.
526 : */
527 364 : if (rel)
528 : {
529 56 : table_close(rel, NoLock);
530 56 : rel = NULL;
531 : }
532 364 : CommitTransactionCommand();
533 364 : pgstat_report_stat(false);
534 : }
535 :
536 : /*
537 : * Enter busy loop and wait for synchronization worker to
538 : * reach expected state (or die trying).
539 : */
540 364 : StartTransactionCommand();
541 364 : started_tx = true;
542 :
543 364 : wait_for_table_state_change(rstate->relid,
544 : SUBREL_STATE_SYNCDONE);
545 : }
546 : else
547 1002 : LWLockRelease(LogicalRepWorkerLock);
548 : }
549 : else
550 : {
551 : /*
552 : * If there is no sync worker for this table yet, count
553 : * running sync workers for this subscription, while we have
554 : * the lock.
555 : */
556 : int nsyncworkers =
557 1612 : logicalrep_sync_worker_count(MyLogicalRepWorker->subid);
558 : struct tablesync_start_time_mapping *hentry;
559 : bool found;
560 :
561 : /* Now safe to release the LWLock */
562 1612 : LWLockRelease(LogicalRepWorkerLock);
563 :
564 1612 : hentry = hash_search(last_start_times, &rstate->relid,
565 : HASH_ENTER, &found);
566 1612 : if (!found)
567 382 : hentry->last_start_time = 0;
568 :
569 1612 : launch_sync_worker(WORKERTYPE_TABLESYNC, nsyncworkers,
570 : rstate->relid, &hentry->last_start_time);
571 : }
572 : }
573 : }
574 :
575 : /* Close table if opened */
576 21634 : if (rel)
577 302 : table_close(rel, NoLock);
578 :
579 :
580 21634 : if (started_tx)
581 : {
582 : /*
583 : * Even when the two_phase mode is requested by the user, it remains
584 : * as 'pending' until all tablesyncs have reached READY state.
585 : *
586 : * When this happens, we restart the apply worker and (if the
587 : * conditions are still ok) then the two_phase tri-state will become
588 : * 'enabled' at that time.
589 : *
590 : * Note: If the subscription has no tables then leave the state as
591 : * PENDING, which allows ALTER SUBSCRIPTION ... REFRESH PUBLICATION to
592 : * work.
593 : */
594 2054 : if (MySubscription->twophasestate == LOGICALREP_TWOPHASE_STATE_PENDING)
595 : {
596 74 : CommandCounterIncrement(); /* make updates visible */
597 74 : if (AllTablesyncsReady())
598 : {
599 12 : ereport(LOG,
600 : (errmsg("logical replication apply worker for subscription \"%s\" will restart so that two_phase can be enabled",
601 : MySubscription->name)));
602 12 : should_exit = true;
603 : }
604 : }
605 :
606 2054 : CommitTransactionCommand();
607 2054 : pgstat_report_stat(true);
608 : }
609 :
610 21634 : if (should_exit)
611 : {
612 : /*
613 : * Reset the last-start time for this worker so that the launcher will
614 : * restart it without waiting for wal_retrieve_retry_interval.
615 : */
616 12 : ApplyLauncherForgetWorkerStartTime(MySubscription->oid);
617 :
618 12 : proc_exit(0);
619 : }
620 21622 : }
621 :
622 : /*
623 : * Create list of columns for COPY based on logical relation mapping.
624 : */
625 : static List *
626 388 : make_copy_attnamelist(LogicalRepRelMapEntry *rel)
627 : {
628 388 : List *attnamelist = NIL;
629 : int i;
630 :
631 1040 : for (i = 0; i < rel->remoterel.natts; i++)
632 : {
633 652 : attnamelist = lappend(attnamelist,
634 652 : makeString(rel->remoterel.attnames[i]));
635 : }
636 :
637 :
638 388 : return attnamelist;
639 : }
640 :
641 : /*
642 : * Data source callback for the COPY FROM, which reads from the remote
643 : * connection and passes the data back to our local COPY.
644 : */
645 : static int
646 27978 : copy_read_data(void *outbuf, int minread, int maxread)
647 : {
648 27978 : int bytesread = 0;
649 : int avail;
650 :
651 : /* If there are some leftover data from previous read, use it. */
652 27978 : avail = copybuf->len - copybuf->cursor;
653 27978 : if (avail)
654 : {
655 0 : if (avail > maxread)
656 0 : avail = maxread;
657 0 : memcpy(outbuf, ©buf->data[copybuf->cursor], avail);
658 0 : copybuf->cursor += avail;
659 0 : maxread -= avail;
660 0 : bytesread += avail;
661 : }
662 :
663 27992 : while (maxread > 0 && bytesread < minread)
664 : {
665 27992 : pgsocket fd = PGINVALID_SOCKET;
666 : int len;
667 27992 : char *buf = NULL;
668 :
669 : for (;;)
670 : {
671 : /* Try read the data. */
672 27992 : len = walrcv_receive(LogRepWorkerWalRcvConn, &buf, &fd);
673 :
674 27992 : CHECK_FOR_INTERRUPTS();
675 :
676 27992 : if (len == 0)
677 14 : break;
678 27978 : else if (len < 0)
679 27978 : return bytesread;
680 : else
681 : {
682 : /* Process the data */
683 27594 : copybuf->data = buf;
684 27594 : copybuf->len = len;
685 27594 : copybuf->cursor = 0;
686 :
687 27594 : avail = copybuf->len - copybuf->cursor;
688 27594 : if (avail > maxread)
689 0 : avail = maxread;
690 27594 : memcpy(outbuf, ©buf->data[copybuf->cursor], avail);
691 27594 : outbuf = (char *) outbuf + avail;
692 27594 : copybuf->cursor += avail;
693 27594 : maxread -= avail;
694 27594 : bytesread += avail;
695 : }
696 :
697 27594 : if (maxread <= 0 || bytesread >= minread)
698 27594 : return bytesread;
699 : }
700 :
701 : /*
702 : * Wait for more data or latch.
703 : */
704 14 : (void) WaitLatchOrSocket(MyLatch,
705 : WL_SOCKET_READABLE | WL_LATCH_SET |
706 : WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
707 : fd, 1000L, WAIT_EVENT_LOGICAL_SYNC_DATA);
708 :
709 14 : ResetLatch(MyLatch);
710 : }
711 :
712 0 : return bytesread;
713 : }
714 :
715 :
716 : /*
717 : * Get information about remote relation in similar fashion the RELATION
718 : * message provides during replication.
719 : *
720 : * This function also returns (a) the relation qualifications to be used in
721 : * the COPY command, and (b) whether the remote relation has published any
722 : * generated column.
723 : */
724 : static void
725 392 : fetch_remote_table_info(char *nspname, char *relname, LogicalRepRelation *lrel,
726 : List **qual, bool *gencol_published)
727 : {
728 : WalRcvExecResult *res;
729 : StringInfoData cmd;
730 : TupleTableSlot *slot;
731 392 : Oid tableRow[] = {OIDOID, CHAROID, CHAROID};
732 392 : Oid attrRow[] = {INT2OID, TEXTOID, OIDOID, BOOLOID, BOOLOID};
733 392 : Oid qualRow[] = {TEXTOID};
734 : bool isnull;
735 : int natt;
736 392 : StringInfo pub_names = NULL;
737 392 : Bitmapset *included_cols = NULL;
738 392 : int server_version = walrcv_server_version(LogRepWorkerWalRcvConn);
739 :
740 392 : lrel->nspname = nspname;
741 392 : lrel->relname = relname;
742 :
743 : /* First fetch Oid and replica identity. */
744 392 : initStringInfo(&cmd);
745 392 : appendStringInfo(&cmd, "SELECT c.oid, c.relreplident, c.relkind"
746 : " FROM pg_catalog.pg_class c"
747 : " INNER JOIN pg_catalog.pg_namespace n"
748 : " ON (c.relnamespace = n.oid)"
749 : " WHERE n.nspname = %s"
750 : " AND c.relname = %s",
751 : quote_literal_cstr(nspname),
752 : quote_literal_cstr(relname));
753 392 : res = walrcv_exec(LogRepWorkerWalRcvConn, cmd.data,
754 : lengthof(tableRow), tableRow);
755 :
756 392 : if (res->status != WALRCV_OK_TUPLES)
757 0 : ereport(ERROR,
758 : (errcode(ERRCODE_CONNECTION_FAILURE),
759 : errmsg("could not fetch table info for table \"%s.%s\" from publisher: %s",
760 : nspname, relname, res->err)));
761 :
762 392 : slot = MakeSingleTupleTableSlot(res->tupledesc, &TTSOpsMinimalTuple);
763 392 : if (!tuplestore_gettupleslot(res->tuplestore, true, false, slot))
764 0 : ereport(ERROR,
765 : (errcode(ERRCODE_UNDEFINED_OBJECT),
766 : errmsg("table \"%s.%s\" not found on publisher",
767 : nspname, relname)));
768 :
769 392 : lrel->remoteid = DatumGetObjectId(slot_getattr(slot, 1, &isnull));
770 : Assert(!isnull);
771 392 : lrel->replident = DatumGetChar(slot_getattr(slot, 2, &isnull));
772 : Assert(!isnull);
773 392 : lrel->relkind = DatumGetChar(slot_getattr(slot, 3, &isnull));
774 : Assert(!isnull);
775 :
776 392 : ExecDropSingleTupleTableSlot(slot);
777 392 : walrcv_clear_result(res);
778 :
779 :
780 : /*
781 : * Get column lists for each relation.
782 : *
783 : * We need to do this before fetching info about column names and types,
784 : * so that we can skip columns that should not be replicated.
785 : */
786 392 : if (server_version >= 150000)
787 : {
788 : WalRcvExecResult *pubres;
789 : TupleTableSlot *tslot;
790 392 : Oid attrsRow[] = {INT2VECTOROID};
791 :
792 : /* Build the pub_names comma-separated string. */
793 392 : pub_names = makeStringInfo();
794 392 : GetPublicationsStr(MySubscription->publications, pub_names, true);
795 :
796 : /*
797 : * Fetch info about column lists for the relation (from all the
798 : * publications).
799 : */
800 392 : resetStringInfo(&cmd);
801 392 : appendStringInfo(&cmd,
802 : "SELECT DISTINCT"
803 : " (CASE WHEN (array_length(gpt.attrs, 1) = c.relnatts)"
804 : " THEN NULL ELSE gpt.attrs END)"
805 : " FROM pg_publication p,"
806 : " LATERAL pg_get_publication_tables(p.pubname) gpt,"
807 : " pg_class c"
808 : " WHERE gpt.relid = %u AND c.oid = gpt.relid"
809 : " AND p.pubname IN ( %s )",
810 : lrel->remoteid,
811 : pub_names->data);
812 :
813 392 : pubres = walrcv_exec(LogRepWorkerWalRcvConn, cmd.data,
814 : lengthof(attrsRow), attrsRow);
815 :
816 392 : if (pubres->status != WALRCV_OK_TUPLES)
817 0 : ereport(ERROR,
818 : (errcode(ERRCODE_CONNECTION_FAILURE),
819 : errmsg("could not fetch column list info for table \"%s.%s\" from publisher: %s",
820 : nspname, relname, pubres->err)));
821 :
822 : /*
823 : * We don't support the case where the column list is different for
824 : * the same table when combining publications. See comments atop
825 : * fetch_relation_list. So there should be only one row returned.
826 : * Although we already checked this when creating the subscription, we
827 : * still need to check here in case the column list was changed after
828 : * creating the subscription and before the sync worker is started.
829 : */
830 392 : if (tuplestore_tuple_count(pubres->tuplestore) > 1)
831 0 : ereport(ERROR,
832 : errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
833 : errmsg("cannot use different column lists for table \"%s.%s\" in different publications",
834 : nspname, relname));
835 :
836 : /*
837 : * Get the column list and build a single bitmap with the attnums.
838 : *
839 : * If we find a NULL value, it means all the columns should be
840 : * replicated.
841 : */
842 392 : tslot = MakeSingleTupleTableSlot(pubres->tupledesc, &TTSOpsMinimalTuple);
843 392 : if (tuplestore_gettupleslot(pubres->tuplestore, true, false, tslot))
844 : {
845 392 : Datum cfval = slot_getattr(tslot, 1, &isnull);
846 :
847 392 : if (!isnull)
848 : {
849 : ArrayType *arr;
850 : int nelems;
851 : int16 *elems;
852 :
853 44 : arr = DatumGetArrayTypeP(cfval);
854 44 : nelems = ARR_DIMS(arr)[0];
855 44 : elems = (int16 *) ARR_DATA_PTR(arr);
856 :
857 118 : for (natt = 0; natt < nelems; natt++)
858 74 : included_cols = bms_add_member(included_cols, elems[natt]);
859 : }
860 :
861 392 : ExecClearTuple(tslot);
862 : }
863 392 : ExecDropSingleTupleTableSlot(tslot);
864 :
865 392 : walrcv_clear_result(pubres);
866 : }
867 :
868 : /*
869 : * Now fetch column names and types.
870 : */
871 392 : resetStringInfo(&cmd);
872 392 : appendStringInfoString(&cmd,
873 : "SELECT a.attnum,"
874 : " a.attname,"
875 : " a.atttypid,"
876 : " a.attnum = ANY(i.indkey)");
877 :
878 : /* Generated columns can be replicated since version 18. */
879 392 : if (server_version >= 180000)
880 392 : appendStringInfoString(&cmd, ", a.attgenerated != ''");
881 :
882 784 : appendStringInfo(&cmd,
883 : " FROM pg_catalog.pg_attribute a"
884 : " LEFT JOIN pg_catalog.pg_index i"
885 : " ON (i.indexrelid = pg_get_replica_identity_index(%u))"
886 : " WHERE a.attnum > 0::pg_catalog.int2"
887 : " AND NOT a.attisdropped %s"
888 : " AND a.attrelid = %u"
889 : " ORDER BY a.attnum",
890 : lrel->remoteid,
891 392 : (server_version >= 120000 && server_version < 180000 ?
892 : "AND a.attgenerated = ''" : ""),
893 : lrel->remoteid);
894 392 : res = walrcv_exec(LogRepWorkerWalRcvConn, cmd.data,
895 : server_version >= 180000 ? lengthof(attrRow) : lengthof(attrRow) - 1, attrRow);
896 :
897 392 : if (res->status != WALRCV_OK_TUPLES)
898 0 : ereport(ERROR,
899 : (errcode(ERRCODE_CONNECTION_FAILURE),
900 : errmsg("could not fetch table info for table \"%s.%s\" from publisher: %s",
901 : nspname, relname, res->err)));
902 :
903 : /* We don't know the number of rows coming, so allocate enough space. */
904 392 : lrel->attnames = palloc0(MaxTupleAttributeNumber * sizeof(char *));
905 392 : lrel->atttyps = palloc0(MaxTupleAttributeNumber * sizeof(Oid));
906 392 : lrel->attkeys = NULL;
907 :
908 : /*
909 : * Store the columns as a list of names. Ignore those that are not
910 : * present in the column list, if there is one.
911 : */
912 392 : natt = 0;
913 392 : slot = MakeSingleTupleTableSlot(res->tupledesc, &TTSOpsMinimalTuple);
914 1118 : while (tuplestore_gettupleslot(res->tuplestore, true, false, slot))
915 : {
916 : char *rel_colname;
917 : AttrNumber attnum;
918 :
919 726 : attnum = DatumGetInt16(slot_getattr(slot, 1, &isnull));
920 : Assert(!isnull);
921 :
922 : /* If the column is not in the column list, skip it. */
923 726 : if (included_cols != NULL && !bms_is_member(attnum, included_cols))
924 : {
925 62 : ExecClearTuple(slot);
926 62 : continue;
927 : }
928 :
929 664 : rel_colname = TextDatumGetCString(slot_getattr(slot, 2, &isnull));
930 : Assert(!isnull);
931 :
932 664 : lrel->attnames[natt] = rel_colname;
933 664 : lrel->atttyps[natt] = DatumGetObjectId(slot_getattr(slot, 3, &isnull));
934 : Assert(!isnull);
935 :
936 664 : if (DatumGetBool(slot_getattr(slot, 4, &isnull)))
937 216 : lrel->attkeys = bms_add_member(lrel->attkeys, natt);
938 :
939 : /* Remember if the remote table has published any generated column. */
940 664 : if (server_version >= 180000 && !(*gencol_published))
941 : {
942 664 : *gencol_published = DatumGetBool(slot_getattr(slot, 5, &isnull));
943 : Assert(!isnull);
944 : }
945 :
946 : /* Should never happen. */
947 664 : if (++natt >= MaxTupleAttributeNumber)
948 0 : elog(ERROR, "too many columns in remote table \"%s.%s\"",
949 : nspname, relname);
950 :
951 664 : ExecClearTuple(slot);
952 : }
953 392 : ExecDropSingleTupleTableSlot(slot);
954 :
955 392 : lrel->natts = natt;
956 :
957 392 : walrcv_clear_result(res);
958 :
959 : /*
960 : * Get relation's row filter expressions. DISTINCT avoids the same
961 : * expression of a table in multiple publications from being included
962 : * multiple times in the final expression.
963 : *
964 : * We need to copy the row even if it matches just one of the
965 : * publications, so we later combine all the quals with OR.
966 : *
967 : * For initial synchronization, row filtering can be ignored in following
968 : * cases:
969 : *
970 : * 1) one of the subscribed publications for the table hasn't specified
971 : * any row filter
972 : *
973 : * 2) one of the subscribed publications has puballtables set to true
974 : *
975 : * 3) one of the subscribed publications is declared as TABLES IN SCHEMA
976 : * that includes this relation
977 : */
978 392 : if (server_version >= 150000)
979 : {
980 : /* Reuse the already-built pub_names. */
981 : Assert(pub_names != NULL);
982 :
983 : /* Check for row filters. */
984 392 : resetStringInfo(&cmd);
985 392 : appendStringInfo(&cmd,
986 : "SELECT DISTINCT pg_get_expr(gpt.qual, gpt.relid)"
987 : " FROM pg_publication p,"
988 : " LATERAL pg_get_publication_tables(p.pubname) gpt"
989 : " WHERE gpt.relid = %u"
990 : " AND p.pubname IN ( %s )",
991 : lrel->remoteid,
992 : pub_names->data);
993 :
994 392 : res = walrcv_exec(LogRepWorkerWalRcvConn, cmd.data, 1, qualRow);
995 :
996 392 : if (res->status != WALRCV_OK_TUPLES)
997 0 : ereport(ERROR,
998 : (errmsg("could not fetch table WHERE clause info for table \"%s.%s\" from publisher: %s",
999 : nspname, relname, res->err)));
1000 :
1001 : /*
1002 : * Multiple row filter expressions for the same table will be combined
1003 : * by COPY using OR. If any of the filter expressions for this table
1004 : * are null, it means the whole table will be copied. In this case it
1005 : * is not necessary to construct a unified row filter expression at
1006 : * all.
1007 : */
1008 392 : slot = MakeSingleTupleTableSlot(res->tupledesc, &TTSOpsMinimalTuple);
1009 422 : while (tuplestore_gettupleslot(res->tuplestore, true, false, slot))
1010 : {
1011 400 : Datum rf = slot_getattr(slot, 1, &isnull);
1012 :
1013 400 : if (!isnull)
1014 30 : *qual = lappend(*qual, makeString(TextDatumGetCString(rf)));
1015 : else
1016 : {
1017 : /* Ignore filters and cleanup as necessary. */
1018 370 : if (*qual)
1019 : {
1020 6 : list_free_deep(*qual);
1021 6 : *qual = NIL;
1022 : }
1023 370 : break;
1024 : }
1025 :
1026 30 : ExecClearTuple(slot);
1027 : }
1028 392 : ExecDropSingleTupleTableSlot(slot);
1029 :
1030 392 : walrcv_clear_result(res);
1031 392 : destroyStringInfo(pub_names);
1032 : }
1033 :
1034 392 : pfree(cmd.data);
1035 392 : }
1036 :
1037 : /*
1038 : * Copy existing data of a table from publisher.
1039 : *
1040 : * Caller is responsible for locking the local relation.
1041 : */
1042 : static void
1043 392 : copy_table(Relation rel)
1044 : {
1045 : LogicalRepRelMapEntry *relmapentry;
1046 : LogicalRepRelation lrel;
1047 392 : List *qual = NIL;
1048 : WalRcvExecResult *res;
1049 : StringInfoData cmd;
1050 : CopyFromState cstate;
1051 : List *attnamelist;
1052 : ParseState *pstate;
1053 392 : List *options = NIL;
1054 392 : bool gencol_published = false;
1055 :
1056 : /* Get the publisher relation info. */
1057 392 : fetch_remote_table_info(get_namespace_name(RelationGetNamespace(rel)),
1058 392 : RelationGetRelationName(rel), &lrel, &qual,
1059 : &gencol_published);
1060 :
1061 : /* Put the relation into relmap. */
1062 392 : logicalrep_relmap_update(&lrel);
1063 :
1064 : /* Map the publisher relation to local one. */
1065 392 : relmapentry = logicalrep_rel_open(lrel.remoteid, NoLock);
1066 : Assert(rel == relmapentry->localrel);
1067 :
1068 : /* Start copy on the publisher. */
1069 388 : initStringInfo(&cmd);
1070 :
1071 : /* Regular table with no row filter or generated columns */
1072 388 : if (lrel.relkind == RELKIND_RELATION && qual == NIL && !gencol_published)
1073 : {
1074 332 : appendStringInfo(&cmd, "COPY %s",
1075 332 : quote_qualified_identifier(lrel.nspname, lrel.relname));
1076 :
1077 : /* If the table has columns, then specify the columns */
1078 332 : if (lrel.natts)
1079 : {
1080 330 : appendStringInfoString(&cmd, " (");
1081 :
1082 : /*
1083 : * XXX Do we need to list the columns in all cases? Maybe we're
1084 : * replicating all columns?
1085 : */
1086 898 : for (int i = 0; i < lrel.natts; i++)
1087 : {
1088 568 : if (i > 0)
1089 238 : appendStringInfoString(&cmd, ", ");
1090 :
1091 568 : appendStringInfoString(&cmd, quote_identifier(lrel.attnames[i]));
1092 : }
1093 :
1094 330 : appendStringInfoChar(&cmd, ')');
1095 : }
1096 :
1097 332 : appendStringInfoString(&cmd, " TO STDOUT");
1098 : }
1099 : else
1100 : {
1101 : /*
1102 : * For non-tables and tables with row filters, we need to do COPY
1103 : * (SELECT ...), but we can't just do SELECT * because we may need to
1104 : * copy only subset of columns including generated columns. For tables
1105 : * with any row filters, build a SELECT query with OR'ed row filters
1106 : * for COPY.
1107 : *
1108 : * We also need to use this same COPY (SELECT ...) syntax when
1109 : * generated columns are published, because copy of generated columns
1110 : * is not supported by the normal COPY.
1111 : */
1112 56 : appendStringInfoString(&cmd, "COPY (SELECT ");
1113 140 : for (int i = 0; i < lrel.natts; i++)
1114 : {
1115 84 : appendStringInfoString(&cmd, quote_identifier(lrel.attnames[i]));
1116 84 : if (i < lrel.natts - 1)
1117 28 : appendStringInfoString(&cmd, ", ");
1118 : }
1119 :
1120 56 : appendStringInfoString(&cmd, " FROM ");
1121 :
1122 : /*
1123 : * For regular tables, make sure we don't copy data from a child that
1124 : * inherits the named table as those will be copied separately.
1125 : */
1126 56 : if (lrel.relkind == RELKIND_RELATION)
1127 22 : appendStringInfoString(&cmd, "ONLY ");
1128 :
1129 56 : appendStringInfoString(&cmd, quote_qualified_identifier(lrel.nspname, lrel.relname));
1130 : /* list of OR'ed filters */
1131 56 : if (qual != NIL)
1132 : {
1133 : ListCell *lc;
1134 22 : char *q = strVal(linitial(qual));
1135 :
1136 22 : appendStringInfo(&cmd, " WHERE %s", q);
1137 24 : for_each_from(lc, qual, 1)
1138 : {
1139 2 : q = strVal(lfirst(lc));
1140 2 : appendStringInfo(&cmd, " OR %s", q);
1141 : }
1142 22 : list_free_deep(qual);
1143 : }
1144 :
1145 56 : appendStringInfoString(&cmd, ") TO STDOUT");
1146 : }
1147 :
1148 : /*
1149 : * Prior to v16, initial table synchronization will use text format even
1150 : * if the binary option is enabled for a subscription.
1151 : */
1152 388 : if (walrcv_server_version(LogRepWorkerWalRcvConn) >= 160000 &&
1153 388 : MySubscription->binary)
1154 : {
1155 10 : appendStringInfoString(&cmd, " WITH (FORMAT binary)");
1156 10 : options = list_make1(makeDefElem("format",
1157 : (Node *) makeString("binary"), -1));
1158 : }
1159 :
1160 388 : res = walrcv_exec(LogRepWorkerWalRcvConn, cmd.data, 0, NULL);
1161 388 : pfree(cmd.data);
1162 388 : if (res->status != WALRCV_OK_COPY_OUT)
1163 0 : ereport(ERROR,
1164 : (errcode(ERRCODE_CONNECTION_FAILURE),
1165 : errmsg("could not start initial contents copy for table \"%s.%s\": %s",
1166 : lrel.nspname, lrel.relname, res->err)));
1167 388 : walrcv_clear_result(res);
1168 :
1169 388 : copybuf = makeStringInfo();
1170 :
1171 388 : pstate = make_parsestate(NULL);
1172 388 : (void) addRangeTableEntryForRelation(pstate, rel, AccessShareLock,
1173 : NULL, false, false);
1174 :
1175 388 : attnamelist = make_copy_attnamelist(relmapentry);
1176 388 : cstate = BeginCopyFrom(pstate, rel, NULL, NULL, false, copy_read_data, attnamelist, options);
1177 :
1178 : /* Do the copy */
1179 386 : (void) CopyFrom(cstate);
1180 :
1181 368 : logicalrep_rel_close(relmapentry, NoLock);
1182 368 : }
1183 :
1184 : /*
1185 : * Determine the tablesync slot name.
1186 : *
1187 : * The name must not exceed NAMEDATALEN - 1 because of remote node constraints
1188 : * on slot name length. We append system_identifier to avoid slot_name
1189 : * collision with subscriptions in other clusters. With the current scheme
1190 : * pg_%u_sync_%u_UINT64_FORMAT (3 + 10 + 6 + 10 + 20 + '\0'), the maximum
1191 : * length of slot_name will be 50.
1192 : *
1193 : * The returned slot name is stored in the supplied buffer (syncslotname) with
1194 : * the given size.
1195 : *
1196 : * Note: We don't use the subscription slot name as part of tablesync slot name
1197 : * because we are responsible for cleaning up these slots and it could become
1198 : * impossible to recalculate what name to cleanup if the subscription slot name
1199 : * had changed.
1200 : */
1201 : void
1202 774 : ReplicationSlotNameForTablesync(Oid suboid, Oid relid,
1203 : char *syncslotname, Size szslot)
1204 : {
1205 774 : snprintf(syncslotname, szslot, "pg_%u_sync_%u_" UINT64_FORMAT, suboid,
1206 : relid, GetSystemIdentifier());
1207 774 : }
1208 :
1209 : /*
1210 : * Start syncing the table in the sync worker.
1211 : *
1212 : * If nothing needs to be done to sync the table, we exit the worker without
1213 : * any further action.
1214 : *
1215 : * The returned slot name is palloc'ed in current memory context.
1216 : */
1217 : static char *
1218 396 : LogicalRepSyncTableStart(XLogRecPtr *origin_startpos)
1219 : {
1220 : char *slotname;
1221 : char *err;
1222 : char relstate;
1223 : XLogRecPtr relstate_lsn;
1224 : Relation rel;
1225 : AclResult aclresult;
1226 : WalRcvExecResult *res;
1227 : char originname[NAMEDATALEN];
1228 : RepOriginId originid;
1229 : UserContext ucxt;
1230 : bool must_use_password;
1231 : bool run_as_owner;
1232 :
1233 : /* Check the state of the table synchronization. */
1234 396 : StartTransactionCommand();
1235 396 : relstate = GetSubscriptionRelState(MyLogicalRepWorker->subid,
1236 396 : MyLogicalRepWorker->relid,
1237 : &relstate_lsn);
1238 396 : CommitTransactionCommand();
1239 :
1240 : /* Is the use of a password mandatory? */
1241 784 : must_use_password = MySubscription->passwordrequired &&
1242 388 : !MySubscription->ownersuperuser;
1243 :
1244 396 : SpinLockAcquire(&MyLogicalRepWorker->relmutex);
1245 396 : MyLogicalRepWorker->relstate = relstate;
1246 396 : MyLogicalRepWorker->relstate_lsn = relstate_lsn;
1247 396 : SpinLockRelease(&MyLogicalRepWorker->relmutex);
1248 :
1249 : /*
1250 : * If synchronization is already done or no longer necessary, exit now
1251 : * that we've updated shared memory state.
1252 : */
1253 396 : switch (relstate)
1254 : {
1255 0 : case SUBREL_STATE_SYNCDONE:
1256 : case SUBREL_STATE_READY:
1257 : case SUBREL_STATE_UNKNOWN:
1258 0 : FinishSyncWorker(); /* doesn't return */
1259 : }
1260 :
1261 : /* Calculate the name of the tablesync slot. */
1262 396 : slotname = (char *) palloc(NAMEDATALEN);
1263 396 : ReplicationSlotNameForTablesync(MySubscription->oid,
1264 396 : MyLogicalRepWorker->relid,
1265 : slotname,
1266 : NAMEDATALEN);
1267 :
1268 : /*
1269 : * Here we use the slot name instead of the subscription name as the
1270 : * application_name, so that it is different from the leader apply worker,
1271 : * so that synchronous replication can distinguish them.
1272 : */
1273 396 : LogRepWorkerWalRcvConn =
1274 396 : walrcv_connect(MySubscription->conninfo, true, true,
1275 : must_use_password,
1276 : slotname, &err);
1277 396 : if (LogRepWorkerWalRcvConn == NULL)
1278 0 : ereport(ERROR,
1279 : (errcode(ERRCODE_CONNECTION_FAILURE),
1280 : errmsg("table synchronization worker for subscription \"%s\" could not connect to the publisher: %s",
1281 : MySubscription->name, err)));
1282 :
1283 : Assert(MyLogicalRepWorker->relstate == SUBREL_STATE_INIT ||
1284 : MyLogicalRepWorker->relstate == SUBREL_STATE_DATASYNC ||
1285 : MyLogicalRepWorker->relstate == SUBREL_STATE_FINISHEDCOPY);
1286 :
1287 : /* Assign the origin tracking record name. */
1288 396 : ReplicationOriginNameForLogicalRep(MySubscription->oid,
1289 396 : MyLogicalRepWorker->relid,
1290 : originname,
1291 : sizeof(originname));
1292 :
1293 396 : if (MyLogicalRepWorker->relstate == SUBREL_STATE_DATASYNC)
1294 : {
1295 : /*
1296 : * We have previously errored out before finishing the copy so the
1297 : * replication slot might exist. We want to remove the slot if it
1298 : * already exists and proceed.
1299 : *
1300 : * XXX We could also instead try to drop the slot, last time we failed
1301 : * but for that, we might need to clean up the copy state as it might
1302 : * be in the middle of fetching the rows. Also, if there is a network
1303 : * breakdown then it wouldn't have succeeded so trying it next time
1304 : * seems like a better bet.
1305 : */
1306 18 : ReplicationSlotDropAtPubNode(LogRepWorkerWalRcvConn, slotname, true);
1307 : }
1308 378 : else if (MyLogicalRepWorker->relstate == SUBREL_STATE_FINISHEDCOPY)
1309 : {
1310 : /*
1311 : * The COPY phase was previously done, but tablesync then crashed
1312 : * before it was able to finish normally.
1313 : */
1314 0 : StartTransactionCommand();
1315 :
1316 : /*
1317 : * The origin tracking name must already exist. It was created first
1318 : * time this tablesync was launched.
1319 : */
1320 0 : originid = replorigin_by_name(originname, false);
1321 0 : replorigin_session_setup(originid, 0);
1322 0 : replorigin_session_origin = originid;
1323 0 : *origin_startpos = replorigin_session_get_progress(false);
1324 :
1325 0 : CommitTransactionCommand();
1326 :
1327 0 : goto copy_table_done;
1328 : }
1329 :
1330 396 : SpinLockAcquire(&MyLogicalRepWorker->relmutex);
1331 396 : MyLogicalRepWorker->relstate = SUBREL_STATE_DATASYNC;
1332 396 : MyLogicalRepWorker->relstate_lsn = InvalidXLogRecPtr;
1333 396 : SpinLockRelease(&MyLogicalRepWorker->relmutex);
1334 :
1335 : /* Update the state and make it visible to others. */
1336 396 : StartTransactionCommand();
1337 396 : UpdateSubscriptionRelState(MyLogicalRepWorker->subid,
1338 396 : MyLogicalRepWorker->relid,
1339 396 : MyLogicalRepWorker->relstate,
1340 396 : MyLogicalRepWorker->relstate_lsn,
1341 : false);
1342 394 : CommitTransactionCommand();
1343 394 : pgstat_report_stat(true);
1344 :
1345 394 : StartTransactionCommand();
1346 :
1347 : /*
1348 : * Use a standard write lock here. It might be better to disallow access
1349 : * to the table while it's being synchronized. But we don't want to block
1350 : * the main apply process from working and it has to open the relation in
1351 : * RowExclusiveLock when remapping remote relation id to local one.
1352 : */
1353 394 : rel = table_open(MyLogicalRepWorker->relid, RowExclusiveLock);
1354 :
1355 : /*
1356 : * Start a transaction in the remote node in REPEATABLE READ mode. This
1357 : * ensures that both the replication slot we create (see below) and the
1358 : * COPY are consistent with each other.
1359 : */
1360 394 : res = walrcv_exec(LogRepWorkerWalRcvConn,
1361 : "BEGIN READ ONLY ISOLATION LEVEL REPEATABLE READ",
1362 : 0, NULL);
1363 394 : if (res->status != WALRCV_OK_COMMAND)
1364 0 : ereport(ERROR,
1365 : (errcode(ERRCODE_CONNECTION_FAILURE),
1366 : errmsg("table copy could not start transaction on publisher: %s",
1367 : res->err)));
1368 394 : walrcv_clear_result(res);
1369 :
1370 : /*
1371 : * Create a new permanent logical decoding slot. This slot will be used
1372 : * for the catchup phase after COPY is done, so tell it to use the
1373 : * snapshot to make the final data consistent.
1374 : */
1375 394 : walrcv_create_slot(LogRepWorkerWalRcvConn,
1376 : slotname, false /* permanent */ , false /* two_phase */ ,
1377 : MySubscription->failover,
1378 : CRS_USE_SNAPSHOT, origin_startpos);
1379 :
1380 : /*
1381 : * Setup replication origin tracking. The purpose of doing this before the
1382 : * copy is to avoid doing the copy again due to any error in setting up
1383 : * origin tracking.
1384 : */
1385 392 : originid = replorigin_by_name(originname, true);
1386 392 : if (!OidIsValid(originid))
1387 : {
1388 : /*
1389 : * Origin tracking does not exist, so create it now.
1390 : *
1391 : * Then advance to the LSN got from walrcv_create_slot. This is WAL
1392 : * logged for the purpose of recovery. Locks are to prevent the
1393 : * replication origin from vanishing while advancing.
1394 : */
1395 392 : originid = replorigin_create(originname);
1396 :
1397 392 : LockRelationOid(ReplicationOriginRelationId, RowExclusiveLock);
1398 392 : replorigin_advance(originid, *origin_startpos, InvalidXLogRecPtr,
1399 : true /* go backward */ , true /* WAL log */ );
1400 392 : UnlockRelationOid(ReplicationOriginRelationId, RowExclusiveLock);
1401 :
1402 392 : replorigin_session_setup(originid, 0);
1403 392 : replorigin_session_origin = originid;
1404 : }
1405 : else
1406 : {
1407 0 : ereport(ERROR,
1408 : (errcode(ERRCODE_DUPLICATE_OBJECT),
1409 : errmsg("replication origin \"%s\" already exists",
1410 : originname)));
1411 : }
1412 :
1413 : /*
1414 : * If the user did not opt to run as the owner of the subscription
1415 : * ('run_as_owner'), then copy the table as the owner of the table.
1416 : */
1417 392 : run_as_owner = MySubscription->runasowner;
1418 392 : if (!run_as_owner)
1419 390 : SwitchToUntrustedUser(rel->rd_rel->relowner, &ucxt);
1420 :
1421 : /*
1422 : * Check that our table sync worker has permission to insert into the
1423 : * target table.
1424 : */
1425 392 : aclresult = pg_class_aclcheck(RelationGetRelid(rel), GetUserId(),
1426 : ACL_INSERT);
1427 392 : if (aclresult != ACLCHECK_OK)
1428 0 : aclcheck_error(aclresult,
1429 0 : get_relkind_objtype(rel->rd_rel->relkind),
1430 0 : RelationGetRelationName(rel));
1431 :
1432 : /*
1433 : * COPY FROM does not honor RLS policies. That is not a problem for
1434 : * subscriptions owned by roles with BYPASSRLS privilege (or superuser,
1435 : * who has it implicitly), but other roles should not be able to
1436 : * circumvent RLS. Disallow logical replication into RLS enabled
1437 : * relations for such roles.
1438 : */
1439 392 : if (check_enable_rls(RelationGetRelid(rel), InvalidOid, false) == RLS_ENABLED)
1440 0 : ereport(ERROR,
1441 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1442 : errmsg("user \"%s\" cannot replicate into relation with row-level security enabled: \"%s\"",
1443 : GetUserNameFromId(GetUserId(), true),
1444 : RelationGetRelationName(rel))));
1445 :
1446 : /* Now do the initial data copy */
1447 392 : PushActiveSnapshot(GetTransactionSnapshot());
1448 392 : copy_table(rel);
1449 368 : PopActiveSnapshot();
1450 :
1451 368 : res = walrcv_exec(LogRepWorkerWalRcvConn, "COMMIT", 0, NULL);
1452 368 : if (res->status != WALRCV_OK_COMMAND)
1453 0 : ereport(ERROR,
1454 : (errcode(ERRCODE_CONNECTION_FAILURE),
1455 : errmsg("table copy could not finish transaction on publisher: %s",
1456 : res->err)));
1457 368 : walrcv_clear_result(res);
1458 :
1459 368 : if (!run_as_owner)
1460 366 : RestoreUserContext(&ucxt);
1461 :
1462 368 : table_close(rel, NoLock);
1463 :
1464 : /* Make the copy visible. */
1465 368 : CommandCounterIncrement();
1466 :
1467 : /*
1468 : * Update the persisted state to indicate the COPY phase is done; make it
1469 : * visible to others.
1470 : */
1471 368 : UpdateSubscriptionRelState(MyLogicalRepWorker->subid,
1472 368 : MyLogicalRepWorker->relid,
1473 : SUBREL_STATE_FINISHEDCOPY,
1474 368 : MyLogicalRepWorker->relstate_lsn,
1475 : false);
1476 :
1477 368 : CommitTransactionCommand();
1478 :
1479 368 : copy_table_done:
1480 :
1481 368 : elog(DEBUG1,
1482 : "LogicalRepSyncTableStart: '%s' origin_startpos lsn %X/%08X",
1483 : originname, LSN_FORMAT_ARGS(*origin_startpos));
1484 :
1485 : /*
1486 : * We are done with the initial data synchronization, update the state.
1487 : */
1488 368 : SpinLockAcquire(&MyLogicalRepWorker->relmutex);
1489 368 : MyLogicalRepWorker->relstate = SUBREL_STATE_SYNCWAIT;
1490 368 : MyLogicalRepWorker->relstate_lsn = *origin_startpos;
1491 368 : SpinLockRelease(&MyLogicalRepWorker->relmutex);
1492 :
1493 : /*
1494 : * Finally, wait until the leader apply worker tells us to catch up and
1495 : * then return to let LogicalRepApplyLoop do it.
1496 : */
1497 368 : wait_for_worker_state_change(SUBREL_STATE_CATCHUP);
1498 368 : return slotname;
1499 : }
1500 :
1501 : /*
1502 : * Execute the initial sync with error handling. Disable the subscription,
1503 : * if it's required.
1504 : *
1505 : * Allocate the slot name in long-lived context on return. Note that we don't
1506 : * handle FATAL errors which are probably because of system resource error and
1507 : * are not repeatable.
1508 : */
1509 : static void
1510 396 : start_table_sync(XLogRecPtr *origin_startpos, char **slotname)
1511 : {
1512 396 : char *sync_slotname = NULL;
1513 :
1514 : Assert(am_tablesync_worker());
1515 :
1516 396 : PG_TRY();
1517 : {
1518 : /* Call initial sync. */
1519 396 : sync_slotname = LogicalRepSyncTableStart(origin_startpos);
1520 : }
1521 24 : PG_CATCH();
1522 : {
1523 24 : if (MySubscription->disableonerr)
1524 2 : DisableSubscriptionAndExit();
1525 : else
1526 : {
1527 : /*
1528 : * Report the worker failed during table synchronization. Abort
1529 : * the current transaction so that the stats message is sent in an
1530 : * idle state.
1531 : */
1532 22 : AbortOutOfAnyTransaction();
1533 22 : pgstat_report_subscription_error(MySubscription->oid,
1534 : WORKERTYPE_TABLESYNC);
1535 :
1536 22 : PG_RE_THROW();
1537 : }
1538 : }
1539 368 : PG_END_TRY();
1540 :
1541 : /* allocate slot name in long-lived context */
1542 368 : *slotname = MemoryContextStrdup(ApplyContext, sync_slotname);
1543 368 : pfree(sync_slotname);
1544 368 : }
1545 :
1546 : /*
1547 : * Runs the tablesync worker.
1548 : *
1549 : * It starts syncing tables. After a successful sync, sets streaming options
1550 : * and starts streaming to catchup with apply worker.
1551 : */
1552 : static void
1553 396 : run_tablesync_worker()
1554 : {
1555 : char originname[NAMEDATALEN];
1556 396 : XLogRecPtr origin_startpos = InvalidXLogRecPtr;
1557 396 : char *slotname = NULL;
1558 : WalRcvStreamOptions options;
1559 :
1560 396 : start_table_sync(&origin_startpos, &slotname);
1561 :
1562 368 : ReplicationOriginNameForLogicalRep(MySubscription->oid,
1563 368 : MyLogicalRepWorker->relid,
1564 : originname,
1565 : sizeof(originname));
1566 :
1567 368 : set_apply_error_context_origin(originname);
1568 :
1569 368 : set_stream_options(&options, slotname, &origin_startpos);
1570 :
1571 368 : walrcv_startstreaming(LogRepWorkerWalRcvConn, &options);
1572 :
1573 : /* Apply the changes till we catchup with the apply worker. */
1574 368 : start_apply(origin_startpos);
1575 0 : }
1576 :
1577 : /* Logical Replication Tablesync worker entry point */
1578 : void
1579 398 : TableSyncWorkerMain(Datum main_arg)
1580 : {
1581 398 : int worker_slot = DatumGetInt32(main_arg);
1582 :
1583 398 : SetupApplyOrSyncWorker(worker_slot);
1584 :
1585 396 : run_tablesync_worker();
1586 :
1587 0 : FinishSyncWorker();
1588 : }
1589 :
1590 : /*
1591 : * If the subscription has no tables then return false.
1592 : *
1593 : * Otherwise, are all tablesyncs READY?
1594 : *
1595 : * Note: This function is not suitable to be called from outside of apply or
1596 : * tablesync workers because MySubscription needs to be already initialized.
1597 : */
1598 : bool
1599 390 : AllTablesyncsReady(void)
1600 : {
1601 : bool started_tx;
1602 : bool has_tables;
1603 :
1604 : /* We need up-to-date sync state info for subscription tables here. */
1605 390 : FetchRelationStates(&has_tables, NULL, &started_tx);
1606 :
1607 390 : if (started_tx)
1608 : {
1609 30 : CommitTransactionCommand();
1610 30 : pgstat_report_stat(true);
1611 : }
1612 :
1613 : /*
1614 : * Return false when there are no tables in subscription or not all tables
1615 : * are in ready state; true otherwise.
1616 : */
1617 390 : return has_tables && (table_states_not_ready == NIL);
1618 : }
1619 :
1620 : /*
1621 : * Return whether the subscription currently has any tables.
1622 : *
1623 : * Note: Unlike HasSubscriptionTables(), this function relies on cached
1624 : * information for subscription tables. Additionally, it should not be
1625 : * invoked outside of apply or tablesync workers, as MySubscription must be
1626 : * initialized first.
1627 : */
1628 : bool
1629 236 : HasSubscriptionTablesCached(void)
1630 : {
1631 : bool started_tx;
1632 : bool has_tables;
1633 :
1634 : /* We need up-to-date subscription tables info here */
1635 236 : FetchRelationStates(&has_tables, NULL, &started_tx);
1636 :
1637 236 : if (started_tx)
1638 : {
1639 0 : CommitTransactionCommand();
1640 0 : pgstat_report_stat(true);
1641 : }
1642 :
1643 236 : return has_tables;
1644 : }
1645 :
1646 : /*
1647 : * Update the two_phase state of the specified subscription in pg_subscription.
1648 : */
1649 : void
1650 18 : UpdateTwoPhaseState(Oid suboid, char new_state)
1651 : {
1652 : Relation rel;
1653 : HeapTuple tup;
1654 : bool nulls[Natts_pg_subscription];
1655 : bool replaces[Natts_pg_subscription];
1656 : Datum values[Natts_pg_subscription];
1657 :
1658 : Assert(new_state == LOGICALREP_TWOPHASE_STATE_DISABLED ||
1659 : new_state == LOGICALREP_TWOPHASE_STATE_PENDING ||
1660 : new_state == LOGICALREP_TWOPHASE_STATE_ENABLED);
1661 :
1662 18 : rel = table_open(SubscriptionRelationId, RowExclusiveLock);
1663 18 : tup = SearchSysCacheCopy1(SUBSCRIPTIONOID, ObjectIdGetDatum(suboid));
1664 18 : if (!HeapTupleIsValid(tup))
1665 0 : elog(ERROR,
1666 : "cache lookup failed for subscription oid %u",
1667 : suboid);
1668 :
1669 : /* Form a new tuple. */
1670 18 : memset(values, 0, sizeof(values));
1671 18 : memset(nulls, false, sizeof(nulls));
1672 18 : memset(replaces, false, sizeof(replaces));
1673 :
1674 : /* And update/set two_phase state */
1675 18 : values[Anum_pg_subscription_subtwophasestate - 1] = CharGetDatum(new_state);
1676 18 : replaces[Anum_pg_subscription_subtwophasestate - 1] = true;
1677 :
1678 18 : tup = heap_modify_tuple(tup, RelationGetDescr(rel),
1679 : values, nulls, replaces);
1680 18 : CatalogTupleUpdate(rel, &tup->t_self, tup);
1681 :
1682 18 : heap_freetuple(tup);
1683 18 : table_close(rel, RowExclusiveLock);
1684 18 : }
|
