PostgreSQL Source Code  git master
tablesync.c
Go to the documentation of this file.
1 /*-------------------------------------------------------------------------
2  * tablesync.c
3  * PostgreSQL logical replication
4  *
5  * Copyright (c) 2012-2020, 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 main 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  * - Sync finishes copy and sets worker state as SYNCWAIT and waits for
31  * state to change in a loop.
32  * - Apply periodically checks tables that are synchronizing for SYNCWAIT.
33  * When the desired state appears, it will set the worker state to
34  * CATCHUP and starts loop-waiting until either the table state is set
35  * to SYNCDONE or the sync worker exits.
36  * - After the sync worker has seen the state change to CATCHUP, it will
37  * read the stream and apply changes (acting like an apply worker) until
38  * it catches up to the specified stream position. Then it sets the
39  * state to SYNCDONE. There might be zero changes applied between
40  * CATCHUP and SYNCDONE, because the sync worker might be ahead of the
41  * apply worker.
42  * - Once the state was set to SYNCDONE, the apply will continue tracking
43  * the table until it reaches the SYNCDONE stream position, at which
44  * point it sets state to READY and stops tracking. Again, there might
45  * be zero changes in between.
46  *
47  * So the state progression is always: INIT -> DATASYNC -> SYNCWAIT -> CATCHUP ->
48  * SYNCDONE -> READY.
49  *
50  * The catalog pg_subscription_rel is used to keep information about
51  * subscribed tables and their state. Some transient state during data
52  * synchronization is kept in shared memory. The states SYNCWAIT and
53  * CATCHUP only appear in memory.
54  *
55  * Example flows look like this:
56  * - Apply is in front:
57  * sync:8
58  * -> set in memory SYNCWAIT
59  * apply:10
60  * -> set in memory CATCHUP
61  * -> enter wait-loop
62  * sync:10
63  * -> set in catalog SYNCDONE
64  * -> exit
65  * apply:10
66  * -> exit wait-loop
67  * -> continue rep
68  * apply:11
69  * -> set in catalog READY
70  * - Sync in front:
71  * sync:10
72  * -> set in memory SYNCWAIT
73  * apply:8
74  * -> set in memory CATCHUP
75  * -> continue per-table filtering
76  * sync:10
77  * -> set in catalog SYNCDONE
78  * -> exit
79  * apply:10
80  * -> set in catalog READY
81  * -> stop per-table filtering
82  * -> continue rep
83  *-------------------------------------------------------------------------
84  */
85 
86 #include "postgres.h"
87 
88 #include "access/table.h"
89 #include "access/xact.h"
91 #include "catalog/pg_type.h"
92 #include "commands/copy.h"
93 #include "miscadmin.h"
94 #include "parser/parse_relation.h"
95 #include "pgstat.h"
100 #include "storage/ipc.h"
101 #include "utils/builtins.h"
102 #include "utils/lsyscache.h"
103 #include "utils/memutils.h"
104 #include "utils/snapmgr.h"
105 
106 static bool table_states_valid = false;
107 
109 
110 /*
111  * Exit routine for synchronization worker.
112  */
113 static void
115 finish_sync_worker(void)
116 {
117  /*
118  * Commit any outstanding transaction. This is the usual case, unless
119  * there was nothing to do for the table.
120  */
121  if (IsTransactionState())
122  {
124  pgstat_report_stat(false);
125  }
126 
127  /* And flush all writes. */
129 
131  ereport(LOG,
132  (errmsg("logical replication table synchronization worker for subscription \"%s\", table \"%s\" has finished",
136 
137  /* Find the main apply worker and signal it. */
139 
140  /* Stop gracefully */
141  proc_exit(0);
142 }
143 
144 /*
145  * Wait until the relation synchronization state is set in the catalog to the
146  * expected one.
147  *
148  * Used when transitioning from CATCHUP state to SYNCDONE.
149  *
150  * Returns false if the synchronization worker has disappeared or the table state
151  * has been reset.
152  */
153 static bool
154 wait_for_relation_state_change(Oid relid, char expected_state)
155 {
156  char state;
157 
158  for (;;)
159  {
160  LogicalRepWorker *worker;
161  XLogRecPtr statelsn;
162 
164 
165  /* XXX use cache invalidation here to improve performance? */
168  relid, &statelsn, true);
170 
171  if (state == SUBREL_STATE_UNKNOWN)
172  return false;
173 
174  if (state == expected_state)
175  return true;
176 
177  /* Check if the sync worker is still running and bail if not. */
178  LWLockAcquire(LogicalRepWorkerLock, LW_SHARED);
179 
180  /* Check if the opposite worker is still running and bail if not. */
182  am_tablesync_worker() ? InvalidOid : relid,
183  false);
184  LWLockRelease(LogicalRepWorkerLock);
185  if (!worker)
186  return false;
187 
188  (void) WaitLatch(MyLatch,
191 
193  }
194 
195  return false;
196 }
197 
198 /*
199  * Wait until the apply worker changes the state of our synchronization
200  * worker to the expected one.
201  *
202  * Used when transitioning from SYNCWAIT state to CATCHUP.
203  *
204  * Returns false if the apply worker has disappeared.
205  */
206 static bool
207 wait_for_worker_state_change(char expected_state)
208 {
209  int rc;
210 
211  for (;;)
212  {
213  LogicalRepWorker *worker;
214 
216 
217  /*
218  * Done if already in correct state. (We assume this fetch is atomic
219  * enough to not give a misleading answer if we do it with no lock.)
220  */
221  if (MyLogicalRepWorker->relstate == expected_state)
222  return true;
223 
224  /*
225  * Bail out if the apply worker has died, else signal it we're
226  * waiting.
227  */
228  LWLockAcquire(LogicalRepWorkerLock, LW_SHARED);
230  InvalidOid, false);
231  if (worker && worker->proc)
233  LWLockRelease(LogicalRepWorkerLock);
234  if (!worker)
235  break;
236 
237  /*
238  * Wait. We expect to get a latch signal back from the apply worker,
239  * but use a timeout in case it dies without sending one.
240  */
241  rc = WaitLatch(MyLatch,
244 
245  if (rc & WL_LATCH_SET)
247  }
248 
249  return false;
250 }
251 
252 /*
253  * Callback from syscache invalidation.
254  */
255 void
257 {
258  table_states_valid = false;
259 }
260 
261 /*
262  * Handle table synchronization cooperation from the synchronization
263  * worker.
264  *
265  * If the sync worker is in CATCHUP state and reached (or passed) the
266  * predetermined synchronization point in the WAL stream, mark the table as
267  * SYNCDONE and finish.
268  */
269 static void
271 {
273 
275 
276  if (MyLogicalRepWorker->relstate == SUBREL_STATE_CATCHUP &&
277  current_lsn >= MyLogicalRepWorker->relstate_lsn)
278  {
279  TimeLineID tli;
280 
281  MyLogicalRepWorker->relstate = SUBREL_STATE_SYNCDONE;
282  MyLogicalRepWorker->relstate_lsn = current_lsn;
283 
285 
290 
292  finish_sync_worker();
293  }
294  else
296 }
297 
298 /*
299  * Handle table synchronization cooperation from the apply worker.
300  *
301  * Walk over all subscription tables that are individually tracked by the
302  * apply process (currently, all that have state other than
303  * SUBREL_STATE_READY) and manage synchronization for them.
304  *
305  * If there are tables that need synchronizing and are not being synchronized
306  * yet, start sync workers for them (if there are free slots for sync
307  * workers). To prevent starting the sync worker for the same relation at a
308  * high frequency after a failure, we store its last start time with each sync
309  * state info. We start the sync worker for the same relation after waiting
310  * at least wal_retrieve_retry_interval.
311  *
312  * For tables that are being synchronized already, check if sync workers
313  * either need action from the apply worker or have finished. This is the
314  * SYNCWAIT to CATCHUP transition.
315  *
316  * If the synchronization position is reached (SYNCDONE), then the table can
317  * be marked as READY and is no longer tracked.
318  */
319 static void
321 {
322  struct tablesync_start_time_mapping
323  {
324  Oid relid;
325  TimestampTz last_start_time;
326  };
327  static List *table_states = NIL;
328  static HTAB *last_start_times = NULL;
329  ListCell *lc;
330  bool started_tx = false;
331 
333 
334  /* We need up-to-date sync state info for subscription tables here. */
335  if (!table_states_valid)
336  {
337  MemoryContext oldctx;
338  List *rstates;
339  ListCell *lc;
340  SubscriptionRelState *rstate;
341 
342  /* Clean the old list. */
343  list_free_deep(table_states);
344  table_states = NIL;
345 
347  started_tx = true;
348 
349  /* Fetch all non-ready tables. */
351 
352  /* Allocate the tracking info in a permanent memory context. */
354  foreach(lc, rstates)
355  {
356  rstate = palloc(sizeof(SubscriptionRelState));
357  memcpy(rstate, lfirst(lc), sizeof(SubscriptionRelState));
358  table_states = lappend(table_states, rstate);
359  }
360  MemoryContextSwitchTo(oldctx);
361 
362  table_states_valid = true;
363  }
364 
365  /*
366  * Prepare a hash table for tracking last start times of workers, to avoid
367  * immediate restarts. We don't need it if there are no tables that need
368  * syncing.
369  */
370  if (table_states && !last_start_times)
371  {
372  HASHCTL ctl;
373 
374  memset(&ctl, 0, sizeof(ctl));
375  ctl.keysize = sizeof(Oid);
376  ctl.entrysize = sizeof(struct tablesync_start_time_mapping);
377  last_start_times = hash_create("Logical replication table sync worker start times",
378  256, &ctl, HASH_ELEM | HASH_BLOBS);
379  }
380 
381  /*
382  * Clean up the hash table when we're done with all tables (just to
383  * release the bit of memory).
384  */
385  else if (!table_states && last_start_times)
386  {
387  hash_destroy(last_start_times);
388  last_start_times = NULL;
389  }
390 
391  /*
392  * Process all tables that are being synchronized.
393  */
394  foreach(lc, table_states)
395  {
397 
398  if (rstate->state == SUBREL_STATE_SYNCDONE)
399  {
400  /*
401  * Apply has caught up to the position where the table sync has
402  * finished. Mark the table as ready so that the apply will just
403  * continue to replicate it normally.
404  */
405  if (current_lsn >= rstate->lsn)
406  {
407  rstate->state = SUBREL_STATE_READY;
408  rstate->lsn = current_lsn;
409  if (!started_tx)
410  {
412  started_tx = true;
413  }
414 
416  rstate->relid, rstate->state,
417  rstate->lsn);
418  }
419  }
420  else
421  {
422  LogicalRepWorker *syncworker;
423 
424  /*
425  * Look for a sync worker for this relation.
426  */
427  LWLockAcquire(LogicalRepWorkerLock, LW_SHARED);
428 
430  rstate->relid, false);
431 
432  if (syncworker)
433  {
434  /* Found one, update our copy of its state */
435  SpinLockAcquire(&syncworker->relmutex);
436  rstate->state = syncworker->relstate;
437  rstate->lsn = syncworker->relstate_lsn;
438  if (rstate->state == SUBREL_STATE_SYNCWAIT)
439  {
440  /*
441  * Sync worker is waiting for apply. Tell sync worker it
442  * can catchup now.
443  */
444  syncworker->relstate = SUBREL_STATE_CATCHUP;
445  syncworker->relstate_lsn =
446  Max(syncworker->relstate_lsn, current_lsn);
447  }
448  SpinLockRelease(&syncworker->relmutex);
449 
450  /* If we told worker to catch up, wait for it. */
451  if (rstate->state == SUBREL_STATE_SYNCWAIT)
452  {
453  /* Signal the sync worker, as it may be waiting for us. */
454  if (syncworker->proc)
455  logicalrep_worker_wakeup_ptr(syncworker);
456 
457  /* Now safe to release the LWLock */
458  LWLockRelease(LogicalRepWorkerLock);
459 
460  /*
461  * Enter busy loop and wait for synchronization worker to
462  * reach expected state (or die trying).
463  */
464  if (!started_tx)
465  {
467  started_tx = true;
468  }
469 
471  SUBREL_STATE_SYNCDONE);
472  }
473  else
474  LWLockRelease(LogicalRepWorkerLock);
475  }
476  else
477  {
478  /*
479  * If there is no sync worker for this table yet, count
480  * running sync workers for this subscription, while we have
481  * the lock.
482  */
483  int nsyncworkers =
485 
486  /* Now safe to release the LWLock */
487  LWLockRelease(LogicalRepWorkerLock);
488 
489  /*
490  * If there are free sync worker slot(s), start a new sync
491  * worker for the table.
492  */
493  if (nsyncworkers < max_sync_workers_per_subscription)
494  {
496  struct tablesync_start_time_mapping *hentry;
497  bool found;
498 
499  hentry = hash_search(last_start_times, &rstate->relid,
500  HASH_ENTER, &found);
501 
502  if (!found ||
503  TimestampDifferenceExceeds(hentry->last_start_time, now,
505  {
510  rstate->relid);
511  hentry->last_start_time = now;
512  }
513  }
514  }
515  }
516  }
517 
518  if (started_tx)
519  {
521  pgstat_report_stat(false);
522  }
523 }
524 
525 /*
526  * Process possible state change(s) of tables that are being synchronized.
527  */
528 void
530 {
531  if (am_tablesync_worker())
532  process_syncing_tables_for_sync(current_lsn);
533  else
535 }
536 
537 /*
538  * Create list of columns for COPY based on logical relation mapping.
539  */
540 static List *
542 {
543  List *attnamelist = NIL;
544  int i;
545 
546  for (i = 0; i < rel->remoterel.natts; i++)
547  {
548  attnamelist = lappend(attnamelist,
549  makeString(rel->remoterel.attnames[i]));
550  }
551 
552 
553  return attnamelist;
554 }
555 
556 /*
557  * Data source callback for the COPY FROM, which reads from the remote
558  * connection and passes the data back to our local COPY.
559  */
560 static int
561 copy_read_data(void *outbuf, int minread, int maxread)
562 {
563  int bytesread = 0;
564  int avail;
565 
566  /* If there are some leftover data from previous read, use it. */
567  avail = copybuf->len - copybuf->cursor;
568  if (avail)
569  {
570  if (avail > maxread)
571  avail = maxread;
572  memcpy(outbuf, &copybuf->data[copybuf->cursor], avail);
573  copybuf->cursor += avail;
574  maxread -= avail;
575  bytesread += avail;
576  }
577 
578  while (maxread > 0 && bytesread < minread)
579  {
581  int len;
582  char *buf = NULL;
583 
584  for (;;)
585  {
586  /* Try read the data. */
587  len = walrcv_receive(wrconn, &buf, &fd);
588 
590 
591  if (len == 0)
592  break;
593  else if (len < 0)
594  return bytesread;
595  else
596  {
597  /* Process the data */
598  copybuf->data = buf;
599  copybuf->len = len;
600  copybuf->cursor = 0;
601 
602  avail = copybuf->len - copybuf->cursor;
603  if (avail > maxread)
604  avail = maxread;
605  memcpy(outbuf, &copybuf->data[copybuf->cursor], avail);
606  outbuf = (void *) ((char *) outbuf + avail);
607  copybuf->cursor += avail;
608  maxread -= avail;
609  bytesread += avail;
610  }
611 
612  if (maxread <= 0 || bytesread >= minread)
613  return bytesread;
614  }
615 
616  /*
617  * Wait for more data or latch.
618  */
619  (void) WaitLatchOrSocket(MyLatch,
622  fd, 1000L, WAIT_EVENT_LOGICAL_SYNC_DATA);
623 
625  }
626 
627  return bytesread;
628 }
629 
630 
631 /*
632  * Get information about remote relation in similar fashion the RELATION
633  * message provides during replication.
634  */
635 static void
636 fetch_remote_table_info(char *nspname, char *relname,
637  LogicalRepRelation *lrel)
638 {
639  WalRcvExecResult *res;
640  StringInfoData cmd;
641  TupleTableSlot *slot;
642  Oid tableRow[] = {OIDOID, CHAROID, CHAROID};
643  Oid attrRow[] = {TEXTOID, OIDOID, INT4OID, BOOLOID};
644  bool isnull;
645  int natt;
646 
647  lrel->nspname = nspname;
648  lrel->relname = relname;
649 
650  /* First fetch Oid and replica identity. */
651  initStringInfo(&cmd);
652  appendStringInfo(&cmd, "SELECT c.oid, c.relreplident, c.relkind"
653  " FROM pg_catalog.pg_class c"
654  " INNER JOIN pg_catalog.pg_namespace n"
655  " ON (c.relnamespace = n.oid)"
656  " WHERE n.nspname = %s"
657  " AND c.relname = %s",
658  quote_literal_cstr(nspname),
659  quote_literal_cstr(relname));
660  res = walrcv_exec(wrconn, cmd.data, lengthof(tableRow), tableRow);
661 
662  if (res->status != WALRCV_OK_TUPLES)
663  ereport(ERROR,
664  (errmsg("could not fetch table info for table \"%s.%s\" from publisher: %s",
665  nspname, relname, res->err)));
666 
668  if (!tuplestore_gettupleslot(res->tuplestore, true, false, slot))
669  ereport(ERROR,
670  (errmsg("table \"%s.%s\" not found on publisher",
671  nspname, relname)));
672 
673  lrel->remoteid = DatumGetObjectId(slot_getattr(slot, 1, &isnull));
674  Assert(!isnull);
675  lrel->replident = DatumGetChar(slot_getattr(slot, 2, &isnull));
676  Assert(!isnull);
677  lrel->relkind = DatumGetChar(slot_getattr(slot, 3, &isnull));
678  Assert(!isnull);
679 
681  walrcv_clear_result(res);
682 
683  /* Now fetch columns. */
684  resetStringInfo(&cmd);
685  appendStringInfo(&cmd,
686  "SELECT a.attname,"
687  " a.atttypid,"
688  " a.atttypmod,"
689  " a.attnum = ANY(i.indkey)"
690  " FROM pg_catalog.pg_attribute a"
691  " LEFT JOIN pg_catalog.pg_index i"
692  " ON (i.indexrelid = pg_get_replica_identity_index(%u))"
693  " WHERE a.attnum > 0::pg_catalog.int2"
694  " AND NOT a.attisdropped %s"
695  " AND a.attrelid = %u"
696  " ORDER BY a.attnum",
697  lrel->remoteid,
698  (walrcv_server_version(wrconn) >= 120000 ? "AND a.attgenerated = ''" : ""),
699  lrel->remoteid);
700  res = walrcv_exec(wrconn, cmd.data, lengthof(attrRow), attrRow);
701 
702  if (res->status != WALRCV_OK_TUPLES)
703  ereport(ERROR,
704  (errmsg("could not fetch table info for table \"%s.%s\": %s",
705  nspname, relname, res->err)));
706 
707  /* We don't know the number of rows coming, so allocate enough space. */
708  lrel->attnames = palloc0(MaxTupleAttributeNumber * sizeof(char *));
709  lrel->atttyps = palloc0(MaxTupleAttributeNumber * sizeof(Oid));
710  lrel->attkeys = NULL;
711 
712  natt = 0;
714  while (tuplestore_gettupleslot(res->tuplestore, true, false, slot))
715  {
716  lrel->attnames[natt] =
717  TextDatumGetCString(slot_getattr(slot, 1, &isnull));
718  Assert(!isnull);
719  lrel->atttyps[natt] = DatumGetObjectId(slot_getattr(slot, 2, &isnull));
720  Assert(!isnull);
721  if (DatumGetBool(slot_getattr(slot, 4, &isnull)))
722  lrel->attkeys = bms_add_member(lrel->attkeys, natt);
723 
724  /* Should never happen. */
725  if (++natt >= MaxTupleAttributeNumber)
726  elog(ERROR, "too many columns in remote table \"%s.%s\"",
727  nspname, relname);
728 
729  ExecClearTuple(slot);
730  }
732 
733  lrel->natts = natt;
734 
735  walrcv_clear_result(res);
736  pfree(cmd.data);
737 }
738 
739 /*
740  * Copy existing data of a table from publisher.
741  *
742  * Caller is responsible for locking the local relation.
743  */
744 static void
746 {
747  LogicalRepRelMapEntry *relmapentry;
748  LogicalRepRelation lrel;
749  WalRcvExecResult *res;
750  StringInfoData cmd;
751  CopyState cstate;
752  List *attnamelist;
753  ParseState *pstate;
754 
755  /* Get the publisher relation info. */
757  RelationGetRelationName(rel), &lrel);
758 
759  /* Put the relation into relmap. */
761 
762  /* Map the publisher relation to local one. */
763  relmapentry = logicalrep_rel_open(lrel.remoteid, NoLock);
764  Assert(rel == relmapentry->localrel);
765 
766  /* Start copy on the publisher. */
767  initStringInfo(&cmd);
768  if (lrel.relkind == RELKIND_RELATION)
769  appendStringInfo(&cmd, "COPY %s TO STDOUT",
771  else
772  {
773  /*
774  * For non-tables, we need to do COPY (SELECT ...), but we can't just
775  * do SELECT * because we need to not copy generated columns.
776  */
777  appendStringInfo(&cmd, "COPY (SELECT ");
778  for (int i = 0; i < lrel.natts; i++)
779  {
781  if (i < lrel.natts - 1)
782  appendStringInfoString(&cmd, ", ");
783  }
784  appendStringInfo(&cmd, " FROM %s) TO STDOUT",
786  }
787  res = walrcv_exec(wrconn, cmd.data, 0, NULL);
788  pfree(cmd.data);
789  if (res->status != WALRCV_OK_COPY_OUT)
790  ereport(ERROR,
791  (errmsg("could not start initial contents copy for table \"%s.%s\": %s",
792  lrel.nspname, lrel.relname, res->err)));
793  walrcv_clear_result(res);
794 
795  copybuf = makeStringInfo();
796 
797  pstate = make_parsestate(NULL);
799  NULL, false, false);
800 
801  attnamelist = make_copy_attnamelist(relmapentry);
802  cstate = BeginCopyFrom(pstate, rel, NULL, false, copy_read_data, attnamelist, NIL);
803 
804  /* Do the copy */
805  (void) CopyFrom(cstate);
806 
807  logicalrep_rel_close(relmapentry, NoLock);
808 }
809 
810 /*
811  * Start syncing the table in the sync worker.
812  *
813  * The returned slot name is palloc'ed in current memory context.
814  */
815 char *
817 {
818  char *slotname;
819  char *err;
820  char relstate;
821  XLogRecPtr relstate_lsn;
822 
823  /* Check the state of the table synchronization. */
827  &relstate_lsn, true);
829 
831  MyLogicalRepWorker->relstate = relstate;
832  MyLogicalRepWorker->relstate_lsn = relstate_lsn;
834 
835  /*
836  * To build a slot name for the sync work, we are limited to NAMEDATALEN -
837  * 1 characters. We cut the original slot name to NAMEDATALEN - 28 chars
838  * and append _%u_sync_%u (1 + 10 + 6 + 10 + '\0'). (It's actually the
839  * NAMEDATALEN on the remote that matters, but this scheme will also work
840  * reasonably if that is different.)
841  */
842  StaticAssertStmt(NAMEDATALEN >= 32, "NAMEDATALEN too small"); /* for sanity */
843  slotname = psprintf("%.*s_%u_sync_%u",
844  NAMEDATALEN - 28,
848 
849  /*
850  * Here we use the slot name instead of the subscription name as the
851  * application_name, so that it is different from the main apply worker,
852  * so that synchronous replication can distinguish them.
853  */
854  wrconn = walrcv_connect(MySubscription->conninfo, true, slotname, &err);
855  if (wrconn == NULL)
856  ereport(ERROR,
857  (errmsg("could not connect to the publisher: %s", err)));
858 
859  switch (MyLogicalRepWorker->relstate)
860  {
861  case SUBREL_STATE_INIT:
862  case SUBREL_STATE_DATASYNC:
863  {
864  Relation rel;
865  WalRcvExecResult *res;
866 
868  MyLogicalRepWorker->relstate = SUBREL_STATE_DATASYNC;
871 
872  /* Update the state and make it visible to others. */
879  pgstat_report_stat(false);
880 
881  /*
882  * We want to do the table data sync in a single transaction.
883  */
885 
886  /*
887  * Use a standard write lock here. It might be better to
888  * disallow access to the table while it's being synchronized.
889  * But we don't want to block the main apply process from
890  * working and it has to open the relation in RowExclusiveLock
891  * when remapping remote relation id to local one.
892  */
894 
895  /*
896  * Create a temporary slot for the sync process. We do this
897  * inside the transaction so that we can use the snapshot made
898  * by the slot to get existing data.
899  */
900  res = walrcv_exec(wrconn,
901  "BEGIN READ ONLY ISOLATION LEVEL "
902  "REPEATABLE READ", 0, NULL);
903  if (res->status != WALRCV_OK_COMMAND)
904  ereport(ERROR,
905  (errmsg("table copy could not start transaction on publisher"),
906  errdetail("The error was: %s", res->err)));
907  walrcv_clear_result(res);
908 
909  /*
910  * Create new temporary logical decoding slot.
911  *
912  * We'll use slot for data copy so make sure the snapshot is
913  * used for the transaction; that way the COPY will get data
914  * that is consistent with the lsn used by the slot to start
915  * decoding.
916  */
917  walrcv_create_slot(wrconn, slotname, true,
918  CRS_USE_SNAPSHOT, origin_startpos);
919 
921  copy_table(rel);
923 
924  res = walrcv_exec(wrconn, "COMMIT", 0, NULL);
925  if (res->status != WALRCV_OK_COMMAND)
926  ereport(ERROR,
927  (errmsg("table copy could not finish transaction on publisher"),
928  errdetail("The error was: %s", res->err)));
929  walrcv_clear_result(res);
930 
931  table_close(rel, NoLock);
932 
933  /* Make the copy visible. */
935 
936  /*
937  * We are done with the initial data synchronization, update
938  * the state.
939  */
941  MyLogicalRepWorker->relstate = SUBREL_STATE_SYNCWAIT;
942  MyLogicalRepWorker->relstate_lsn = *origin_startpos;
944 
945  /* Wait for main apply worker to tell us to catchup. */
946  wait_for_worker_state_change(SUBREL_STATE_CATCHUP);
947 
948  /*----------
949  * There are now two possible states here:
950  * a) Sync is behind the apply. If that's the case we need to
951  * catch up with it by consuming the logical replication
952  * stream up to the relstate_lsn. For that, we exit this
953  * function and continue in ApplyWorkerMain().
954  * b) Sync is caught up with the apply. So it can just set
955  * the state to SYNCDONE and finish.
956  *----------
957  */
958  if (*origin_startpos >= MyLogicalRepWorker->relstate_lsn)
959  {
960  /*
961  * Update the new state in catalog. No need to bother
962  * with the shmem state as we are exiting for good.
963  */
966  SUBREL_STATE_SYNCDONE,
967  *origin_startpos);
968  finish_sync_worker();
969  }
970  break;
971  }
972  case SUBREL_STATE_SYNCDONE:
973  case SUBREL_STATE_READY:
974  case SUBREL_STATE_UNKNOWN:
975 
976  /*
977  * Nothing to do here but finish. (UNKNOWN means the relation was
978  * removed from pg_subscription_rel before the sync worker could
979  * start.)
980  */
981  finish_sync_worker();
982  break;
983  default:
984  elog(ERROR, "unknown relation state \"%c\"",
986  }
987 
988  return slotname;
989 }
Subscription * MySubscription
Definition: worker.c:107
Value * makeString(char *str)
Definition: value.c:53
#define NIL
Definition: pg_list.h:65
void hash_destroy(HTAB *hashp)
Definition: dynahash.c:816
WalReceiverConn * wrconn
Definition: worker.c:105
#define InvalidXLogRecPtr
Definition: xlogdefs.h:28
#define walrcv_endstreaming(conn, next_tli)
Definition: walreceiver.h:288
void table_close(Relation relation, LOCKMODE lockmode)
Definition: table.c:133
char * quote_literal_cstr(const char *rawstr)
Definition: quote.c:102
uint32 TimeLineID
Definition: xlogdefs.h:52
#define HASH_ELEM
Definition: hsearch.h:87
#define WL_TIMEOUT
Definition: latch.h:127
const char * quote_identifier(const char *ident)
Definition: ruleutils.c:10737
static TupleTableSlot * ExecClearTuple(TupleTableSlot *slot)
Definition: tuptable.h:425
#define MaxTupleAttributeNumber
Definition: htup_details.h:33
void process_syncing_tables(XLogRecPtr current_lsn)
Definition: tablesync.c:529
TupleTableSlot * MakeSingleTupleTableSlot(TupleDesc tupdesc, const TupleTableSlotOps *tts_ops)
Definition: execTuples.c:1208
TimestampTz GetCurrentTimestamp(void)
Definition: timestamp.c:1583
void logicalrep_worker_wakeup(Oid subid, Oid relid)
Definition: launcher.c:594
int64 TimestampTz
Definition: timestamp.h:39
static void process_syncing_tables_for_apply(XLogRecPtr current_lsn)
Definition: tablesync.c:320
CopyState BeginCopyFrom(ParseState *pstate, Relation rel, const char *filename, bool is_program, copy_data_source_cb data_source_cb, List *attnamelist, List *options)
Definition: copy.c:3334
#define DatumGetObjectId(X)
Definition: postgres.h:500
void CommitTransactionCommand(void)
Definition: xact.c:2917
char * psprintf(const char *fmt,...)
Definition: psprintf.c:46
StringInfo makeStringInfo(void)
Definition: stringinfo.c:41
#define walrcv_receive(conn, buffer, wait_fd)
Definition: walreceiver.h:290
#define walrcv_server_version(conn)
Definition: walreceiver.h:282
static MemoryContext MemoryContextSwitchTo(MemoryContext context)
Definition: palloc.h:109
#define AccessShareLock
Definition: lockdefs.h:36
Size entrysize
Definition: hsearch.h:73
static void copy_table(Relation rel)
Definition: tablesync.c:745
void proc_exit(int code)
Definition: ipc.c:104
static bool wait_for_relation_state_change(Oid relid, char expected_state)
Definition: tablesync.c:154
uint64 CopyFrom(CopyState cstate)
Definition: copy.c:2653
void logicalrep_rel_close(LogicalRepRelMapEntry *rel, LOCKMODE lockmode)
Definition: relation.c:399
#define WL_SOCKET_READABLE
Definition: latch.h:125
void PopActiveSnapshot(void)
Definition: snapmgr.c:814
#define lengthof(array)
Definition: c.h:668
void * hash_search(HTAB *hashp, const void *keyPtr, HASHACTION action, bool *foundPtr)
Definition: dynahash.c:908
#define LOG
Definition: elog.h:26
NameData relname
Definition: pg_class.h:38
unsigned int Oid
Definition: postgres_ext.h:31
static bool table_states_valid
Definition: tablesync.c:106
char * LogicalRepSyncTableStart(XLogRecPtr *origin_startpos)
Definition: tablesync.c:816
#define walrcv_create_slot(conn, slotname, temporary, snapshot_action, lsn)
Definition: walreceiver.h:294
bool TimestampDifferenceExceeds(TimestampTz start_time, TimestampTz stop_time, int msec)
Definition: timestamp.c:1682
Snapshot GetTransactionSnapshot(void)
Definition: snapmgr.c:306
void list_free_deep(List *list)
Definition: list.c:1390
Bitmapset * attkeys
Definition: logicalproto.h:53
void XLogFlush(XLogRecPtr record)
Definition: xlog.c:2844
static int fd(const char *x, int i)
Definition: preproc-init.c:105
void ResetLatch(Latch *latch)
Definition: latch.c:540
int WaitLatch(Latch *latch, int wakeEvents, long timeout, uint32 wait_event_info)
Definition: latch.c:365
LogicalRepRelId remoteid
Definition: logicalproto.h:45
XLogRecPtr relstate_lsn
ParseNamespaceItem * addRangeTableEntryForRelation(ParseState *pstate, Relation rel, int lockmode, Alias *alias, bool inh, bool inFromCl)
ParseState * make_parsestate(ParseState *parentParseState)
Definition: parse_node.c:43
void LWLockRelease(LWLock *lock)
Definition: lwlock.c:1812
#define NAMEDATALEN
void logicalrep_worker_wakeup_ptr(LogicalRepWorker *worker)
Definition: launcher.c:614
#define StaticAssertStmt(condition, errmessage)
Definition: c.h:852
void logicalrep_relmap_update(LogicalRepRelation *remoterel)
Definition: relation.c:172
#define SpinLockAcquire(lock)
Definition: spin.h:62
Definition: dynahash.c:210
static void walrcv_clear_result(WalRcvExecResult *walres)
Definition: walreceiver.h:304
void pfree(void *pointer)
Definition: mcxt.c:1056
void appendStringInfo(StringInfo str, const char *fmt,...)
Definition: stringinfo.c:91
TupleDesc tupledesc
Definition: walreceiver.h:213
LogicalRepWorker * MyLogicalRepWorker
Definition: launcher.c:57
#define ERROR
Definition: elog.h:43
LogicalRepRelation remoterel
int max_sync_workers_per_subscription
Definition: launcher.c:55
static bool am_tablesync_worker(void)
void appendStringInfoString(StringInfo str, const char *s)
Definition: stringinfo.c:176
char * get_namespace_name(Oid nspid)
Definition: lsyscache.c:3155
#define NoLock
Definition: lockdefs.h:34
static char * buf
Definition: pg_test_fsync.c:67
void PushActiveSnapshot(Snapshot snap)
Definition: snapmgr.c:735
void ExecDropSingleTupleTableSlot(TupleTableSlot *slot)
Definition: execTuples.c:1224
#define RowExclusiveLock
Definition: lockdefs.h:38
int errdetail(const char *fmt,...)
Definition: elog.c:957
#define DatumGetBool(X)
Definition: postgres.h:393
#define RelationGetRelationName(relation)
Definition: rel.h:490
void resetStringInfo(StringInfo str)
Definition: stringinfo.c:75
int WaitLatchOrSocket(Latch *latch, int wakeEvents, pgsocket sock, long timeout, uint32 wait_event_info)
Definition: latch.c:390
unsigned int uint32
Definition: c.h:367
int pgsocket
Definition: port.h:31
LogicalRepRelMapEntry * logicalrep_rel_open(LogicalRepRelId remoteid, LOCKMODE lockmode)
Definition: relation.c:237
List * lappend(List *list, void *datum)
Definition: list.c:321
void initStringInfo(StringInfo str)
Definition: stringinfo.c:59
char * quote_qualified_identifier(const char *qualifier, const char *ident)
Definition: ruleutils.c:10821
int wal_retrieve_retry_interval
Definition: xlog.c:109
#define SpinLockRelease(lock)
Definition: spin.h:64
#define HASH_BLOBS
Definition: hsearch.h:88
#define TextDatumGetCString(d)
Definition: builtins.h:88
static void pg_attribute_noreturn()
Definition: tablesync.c:114
void * palloc0(Size size)
Definition: mcxt.c:980
HTAB * hash_create(const char *tabname, long nelem, HASHCTL *info, int flags)
Definition: dynahash.c:318
uintptr_t Datum
Definition: postgres.h:367
void CommandCounterIncrement(void)
Definition: xact.c:1006
#define PGINVALID_SOCKET
Definition: port.h:33
#define DatumGetChar(X)
Definition: postgres.h:409
void UpdateSubscriptionRelState(Oid subid, Oid relid, char state, XLogRecPtr sublsn)
Size keysize
Definition: hsearch.h:72
StringInfo copybuf
Definition: tablesync.c:108
static void fetch_remote_table_info(char *nspname, char *relname, LogicalRepRelation *lrel)
Definition: tablesync.c:636
int logicalrep_sync_worker_count(Oid subid)
Definition: launcher.c:720
#define InvalidOid
Definition: postgres_ext.h:36
static Datum slot_getattr(TupleTableSlot *slot, int attnum, bool *isnull)
Definition: tuptable.h:381
#define ereport(elevel,...)
Definition: elog.h:144
XLogRecPtr GetXLogWriteRecPtr(void)
Definition: xlog.c:11498
#define Max(x, y)
Definition: c.h:914
bool tuplestore_gettupleslot(Tuplestorestate *state, bool forward, bool copy, TupleTableSlot *slot)
Definition: tuplestore.c:1078
Tuplestorestate * tuplestore
Definition: walreceiver.h:212
void logicalrep_worker_launch(Oid dbid, Oid subid, const char *subname, Oid userid, Oid relid)
Definition: launcher.c:282
uint64 XLogRecPtr
Definition: xlogdefs.h:21
#define Assert(condition)
Definition: c.h:738
#define lfirst(lc)
Definition: pg_list.h:190
WalRcvExecStatus status
Definition: walreceiver.h:210
Definition: regguts.h:298
LogicalRepWorker * logicalrep_worker_find(Oid subid, Oid relid, bool only_running)
Definition: launcher.c:231
void StartTransactionCommand(void)
Definition: xact.c:2816
bool LWLockAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1208
bool IsTransactionState(void)
Definition: xact.c:356
Bitmapset * bms_add_member(Bitmapset *a, int x)
Definition: bitmapset.c:736
char GetSubscriptionRelState(Oid subid, Oid relid, XLogRecPtr *sublsn, bool missing_ok)
Snapshot GetLatestSnapshot(void)
Definition: snapmgr.c:381
static void process_syncing_tables_for_sync(XLogRecPtr current_lsn)
Definition: tablesync.c:270
void * palloc(Size size)
Definition: mcxt.c:949
int errmsg(const char *fmt,...)
Definition: elog.c:824
#define elog(elevel,...)
Definition: elog.h:214
int i
void * arg
struct Latch * MyLatch
Definition: globals.c:54
static int copy_read_data(void *outbuf, int minread, int maxread)
Definition: tablesync.c:561
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:99
Relation table_open(Oid relationId, LOCKMODE lockmode)
Definition: table.c:39
static List * make_copy_attnamelist(LogicalRepRelMapEntry *rel)
Definition: tablesync.c:541
Definition: pg_list.h:50
List * GetSubscriptionNotReadyRelations(Oid subid)
char * get_rel_name(Oid relid)
Definition: lsyscache.c:1791
#define WL_LATCH_SET
Definition: latch.h:124
static bool wait_for_worker_state_change(char expected_state)
Definition: tablesync.c:207
const TupleTableSlotOps TTSOpsMinimalTuple
Definition: execTuples.c:85
Datum now(PG_FUNCTION_ARGS)
Definition: timestamp.c:1547
#define WL_EXIT_ON_PM_DEATH
Definition: latch.h:129
#define walrcv_exec(conn, exec, nRetTypes, retTypes)
Definition: walreceiver.h:298
void invalidate_syncing_table_states(Datum arg, int cacheid, uint32 hashvalue)
Definition: tablesync.c:256
MemoryContext CacheMemoryContext
Definition: mcxt.c:47
#define RelationGetNamespace(relation)
Definition: rel.h:497
void pgstat_report_stat(bool force)
Definition: pgstat.c:839
#define walrcv_connect(conninfo, logical, appname, err)
Definition: walreceiver.h:272