Line data Source code
1 : /*-------------------------------------------------------------------------
2 : * slotsync.c
3 : * Functionality for synchronizing slots to a standby server from the
4 : * primary server.
5 : *
6 : * Copyright (c) 2024-2026, PostgreSQL Global Development Group
7 : *
8 : * IDENTIFICATION
9 : * src/backend/replication/logical/slotsync.c
10 : *
11 : * This file contains the code for slot synchronization on a physical standby
12 : * to fetch logical failover slots information from the primary server, create
13 : * the slots on the standby and synchronize them periodically.
14 : *
15 : * Slot synchronization can be performed either automatically by enabling slot
16 : * sync worker or manually by calling SQL function pg_sync_replication_slots().
17 : *
18 : * If the WAL corresponding to the remote's restart_lsn is not available on the
19 : * physical standby or the remote's catalog_xmin precedes the oldest xid for
20 : * which it is guaranteed that rows wouldn't have been removed then we cannot
21 : * create the local standby slot because that would mean moving the local slot
22 : * backward and decoding won't be possible via such a slot. In this case, the
23 : * slot will be marked as RS_TEMPORARY. Once the primary server catches up,
24 : * the slot will be marked as RS_PERSISTENT (which means sync-ready) after
25 : * which slot sync worker can perform the sync periodically or user can call
26 : * pg_sync_replication_slots() periodically to perform the syncs.
27 : *
28 : * If synchronized slots fail to build a consistent snapshot from the
29 : * restart_lsn before reaching confirmed_flush_lsn, they would become
30 : * unreliable after promotion due to potential data loss from changes
31 : * before reaching a consistent point. This can happen because the slots can
32 : * be synced at some random time and we may not reach the consistent point
33 : * at the same WAL location as the primary. So, we mark such slots as
34 : * RS_TEMPORARY. Once the decoding from corresponding LSNs can reach a
35 : * consistent point, they will be marked as RS_PERSISTENT.
36 : *
37 : * If the WAL prior to the remote slot's confirmed_flush_lsn has not been
38 : * flushed on the standby, the slot is marked as RS_TEMPORARY. Once the standby
39 : * catches up and flushes that WAL, the slot will be marked as RS_PERSISTENT.
40 : *
41 : * The slot sync worker waits for some time before the next synchronization,
42 : * with the duration varying based on whether any slots were updated during
43 : * the last cycle. Refer to the comments above wait_for_slot_activity() for
44 : * more details.
45 : *
46 : * If the SQL function pg_sync_replication_slots() is used to sync the slots,
47 : * and if the slots are not ready to be synced and are marked as RS_TEMPORARY
48 : * because of any of the reasons mentioned above, then the SQL function also
49 : * waits and retries until the slots are marked as RS_PERSISTENT (which means
50 : * sync-ready). Refer to the comments in SyncReplicationSlots() for more
51 : * details.
52 : *
53 : * Any standby synchronized slots will be dropped if they no longer need
54 : * to be synchronized. See comment atop drop_local_obsolete_slots() for more
55 : * details.
56 : *---------------------------------------------------------------------------
57 : */
58 :
59 : #include "postgres.h"
60 :
61 : #include <time.h>
62 :
63 : #include "access/xlog_internal.h"
64 : #include "access/xlogrecovery.h"
65 : #include "catalog/pg_database.h"
66 : #include "libpq/pqsignal.h"
67 : #include "pgstat.h"
68 : #include "postmaster/interrupt.h"
69 : #include "replication/logical.h"
70 : #include "replication/slotsync.h"
71 : #include "replication/snapbuild.h"
72 : #include "storage/ipc.h"
73 : #include "storage/lmgr.h"
74 : #include "storage/proc.h"
75 : #include "storage/procarray.h"
76 : #include "tcop/tcopprot.h"
77 : #include "utils/builtins.h"
78 : #include "utils/memutils.h"
79 : #include "utils/pg_lsn.h"
80 : #include "utils/ps_status.h"
81 : #include "utils/timeout.h"
82 : #include "utils/wait_event.h"
83 :
84 : /*
85 : * Struct for sharing information to control slot synchronization.
86 : *
87 : * The 'pid' is either the slot sync worker's pid or the backend's pid running
88 : * the SQL function pg_sync_replication_slots(). When the startup process sets
89 : * 'stopSignaled' during promotion, it uses this 'pid' to wake up the currently
90 : * synchronizing process so that the process can immediately stop its
91 : * synchronizing work on seeing 'stopSignaled' set.
92 : * Setting 'stopSignaled' is also used to handle the race condition when the
93 : * postmaster has not noticed the promotion yet and thus may end up restarting
94 : * the slot sync worker. If 'stopSignaled' is set, the worker will exit in such a
95 : * case. The SQL function pg_sync_replication_slots() will also error out if
96 : * this flag is set. Note that we don't need to reset this variable as after
97 : * promotion the slot sync worker won't be restarted because the pmState
98 : * changes to PM_RUN from PM_HOT_STANDBY and we don't support demoting
99 : * primary without restarting the server. See LaunchMissingBackgroundProcesses.
100 : *
101 : * The 'syncing' flag is needed to prevent concurrent slot syncs to avoid slot
102 : * overwrites.
103 : *
104 : * The 'last_start_time' is needed by postmaster to start the slot sync worker
105 : * once per SLOTSYNC_RESTART_INTERVAL_SEC. In cases where an immediate restart
106 : * is expected (e.g., slot sync GUCs change), slot sync worker will reset
107 : * last_start_time before exiting, so that postmaster can start the worker
108 : * without waiting for SLOTSYNC_RESTART_INTERVAL_SEC.
109 : */
110 : typedef struct SlotSyncCtxStruct
111 : {
112 : pid_t pid;
113 : bool stopSignaled;
114 : bool syncing;
115 : time_t last_start_time;
116 : slock_t mutex;
117 : } SlotSyncCtxStruct;
118 :
119 : static SlotSyncCtxStruct *SlotSyncCtx = NULL;
120 :
121 : /* GUC variable */
122 : bool sync_replication_slots = false;
123 :
124 : /*
125 : * The sleep time (ms) between slot-sync cycles varies dynamically
126 : * (within a MIN/MAX range) according to slot activity. See
127 : * wait_for_slot_activity() for details.
128 : */
129 : #define MIN_SLOTSYNC_WORKER_NAPTIME_MS 200
130 : #define MAX_SLOTSYNC_WORKER_NAPTIME_MS 30000 /* 30s */
131 :
132 : static long sleep_ms = MIN_SLOTSYNC_WORKER_NAPTIME_MS;
133 :
134 : /* The restart interval for slot sync work used by postmaster */
135 : #define SLOTSYNC_RESTART_INTERVAL_SEC 10
136 :
137 : /*
138 : * Flag to tell if we are syncing replication slots. Unlike the 'syncing' flag
139 : * in SlotSyncCtxStruct, this flag is true only if the current process is
140 : * performing slot synchronization.
141 : */
142 : static bool syncing_slots = false;
143 :
144 : /*
145 : * Structure to hold information fetched from the primary server about a logical
146 : * replication slot.
147 : */
148 : typedef struct RemoteSlot
149 : {
150 : char *name;
151 : char *plugin;
152 : char *database;
153 : bool two_phase;
154 : bool failover;
155 : XLogRecPtr restart_lsn;
156 : XLogRecPtr confirmed_lsn;
157 : XLogRecPtr two_phase_at;
158 : TransactionId catalog_xmin;
159 :
160 : /* RS_INVAL_NONE if valid, or the reason of invalidation */
161 : ReplicationSlotInvalidationCause invalidated;
162 : } RemoteSlot;
163 :
164 : static void slotsync_failure_callback(int code, Datum arg);
165 : static void update_synced_slots_inactive_since(void);
166 :
167 : /*
168 : * Update slot sync skip stats. This function requires the caller to acquire
169 : * the slot.
170 : */
171 : static void
172 50 : update_slotsync_skip_stats(SlotSyncSkipReason skip_reason)
173 : {
174 : ReplicationSlot *slot;
175 :
176 : Assert(MyReplicationSlot);
177 :
178 50 : slot = MyReplicationSlot;
179 :
180 : /*
181 : * Update the slot sync related stats in pg_stat_replication_slots when a
182 : * slot sync is skipped
183 : */
184 50 : if (skip_reason != SS_SKIP_NONE)
185 4 : pgstat_report_replslotsync(slot);
186 :
187 : /* Update the slot sync skip reason */
188 50 : if (slot->slotsync_skip_reason != skip_reason)
189 : {
190 3 : SpinLockAcquire(&slot->mutex);
191 3 : slot->slotsync_skip_reason = skip_reason;
192 3 : SpinLockRelease(&slot->mutex);
193 : }
194 50 : }
195 :
196 : /*
197 : * If necessary, update the local synced slot's metadata based on the data
198 : * from the remote slot.
199 : *
200 : * If no update was needed (the data of the remote slot is the same as the
201 : * local slot) return false, otherwise true.
202 : */
203 : static bool
204 50 : update_local_synced_slot(RemoteSlot *remote_slot, Oid remote_dbid)
205 : {
206 50 : ReplicationSlot *slot = MyReplicationSlot;
207 50 : bool updated_xmin_or_lsn = false;
208 50 : bool updated_config = false;
209 50 : SlotSyncSkipReason skip_reason = SS_SKIP_NONE;
210 50 : XLogRecPtr latestFlushPtr = GetStandbyFlushRecPtr(NULL);
211 :
212 : Assert(slot->data.invalidated == RS_INVAL_NONE);
213 :
214 : /*
215 : * Make sure that concerned WAL is received and flushed before syncing
216 : * slot to target lsn received from the primary server.
217 : */
218 50 : if (remote_slot->confirmed_lsn > latestFlushPtr)
219 : {
220 0 : update_slotsync_skip_stats(SS_SKIP_WAL_NOT_FLUSHED);
221 :
222 : /*
223 : * Can get here only if GUC 'synchronized_standby_slots' on the
224 : * primary server was not configured correctly.
225 : */
226 0 : ereport(LOG,
227 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
228 : errmsg("skipping slot synchronization because the received slot sync"
229 : " LSN %X/%08X for slot \"%s\" is ahead of the standby position %X/%08X",
230 : LSN_FORMAT_ARGS(remote_slot->confirmed_lsn),
231 : remote_slot->name,
232 : LSN_FORMAT_ARGS(latestFlushPtr)));
233 :
234 0 : return false;
235 : }
236 :
237 : /*
238 : * Don't overwrite if we already have a newer catalog_xmin and
239 : * restart_lsn.
240 : */
241 98 : if (remote_slot->restart_lsn < slot->data.restart_lsn ||
242 48 : TransactionIdPrecedes(remote_slot->catalog_xmin,
243 : slot->data.catalog_xmin))
244 : {
245 : /* Update slot sync skip stats */
246 4 : update_slotsync_skip_stats(SS_SKIP_WAL_OR_ROWS_REMOVED);
247 :
248 : /*
249 : * This can happen in following situations:
250 : *
251 : * If the slot is temporary, it means either the initial WAL location
252 : * reserved for the local slot is ahead of the remote slot's
253 : * restart_lsn or the initial xmin_horizon computed for the local slot
254 : * is ahead of the remote slot.
255 : *
256 : * If the slot is persistent, both restart_lsn and catalog_xmin of the
257 : * synced slot could still be ahead of the remote slot. Since we use
258 : * slot advance functionality to keep snapbuild/slot updated, it is
259 : * possible that the restart_lsn and catalog_xmin are advanced to a
260 : * later position than it has on the primary. This can happen when
261 : * slot advancing machinery finds running xacts record after reaching
262 : * the consistent state at a later point than the primary where it
263 : * serializes the snapshot and updates the restart_lsn.
264 : *
265 : * We LOG the message if the slot is temporary as it can help the user
266 : * to understand why the slot is not sync-ready. In the case of a
267 : * persistent slot, it would be a more common case and won't directly
268 : * impact the users, so we used DEBUG1 level to log the message.
269 : */
270 4 : ereport(slot->data.persistency == RS_TEMPORARY ? LOG : DEBUG1,
271 : errmsg("could not synchronize replication slot \"%s\"",
272 : remote_slot->name),
273 : errdetail("Synchronization could lead to data loss, because the remote slot needs WAL at LSN %X/%08X and catalog xmin %u, but the standby has LSN %X/%08X and catalog xmin %u.",
274 : LSN_FORMAT_ARGS(remote_slot->restart_lsn),
275 : remote_slot->catalog_xmin,
276 : LSN_FORMAT_ARGS(slot->data.restart_lsn),
277 : slot->data.catalog_xmin));
278 :
279 : /*
280 : * Skip updating the configuration. This is required to avoid syncing
281 : * two_phase_at without syncing confirmed_lsn. Otherwise, the prepared
282 : * transaction between old confirmed_lsn and two_phase_at will
283 : * unexpectedly get decoded and sent to the downstream after
284 : * promotion. See comments in ReorderBufferFinishPrepared.
285 : */
286 4 : return false;
287 : }
288 :
289 : /*
290 : * Attempt to sync LSNs and xmins only if remote slot is ahead of local
291 : * slot.
292 : */
293 46 : if (remote_slot->confirmed_lsn > slot->data.confirmed_flush ||
294 67 : remote_slot->restart_lsn > slot->data.restart_lsn ||
295 33 : TransactionIdFollows(remote_slot->catalog_xmin,
296 : slot->data.catalog_xmin))
297 : {
298 : /*
299 : * We can't directly copy the remote slot's LSN or xmin unless there
300 : * exists a consistent snapshot at that point. Otherwise, after
301 : * promotion, the slots may not reach a consistent point before the
302 : * confirmed_flush_lsn which can lead to a data loss. To avoid data
303 : * loss, we let slot machinery advance the slot which ensures that
304 : * snapbuilder/slot statuses are updated properly.
305 : */
306 13 : if (SnapBuildSnapshotExists(remote_slot->restart_lsn))
307 : {
308 : /*
309 : * Update the slot info directly if there is a serialized snapshot
310 : * at the restart_lsn, as the slot can quickly reach consistency
311 : * at restart_lsn by restoring the snapshot.
312 : */
313 3 : SpinLockAcquire(&slot->mutex);
314 3 : slot->data.restart_lsn = remote_slot->restart_lsn;
315 3 : slot->data.confirmed_flush = remote_slot->confirmed_lsn;
316 3 : slot->data.catalog_xmin = remote_slot->catalog_xmin;
317 3 : SpinLockRelease(&slot->mutex);
318 : }
319 : else
320 : {
321 : bool found_consistent_snapshot;
322 :
323 10 : LogicalSlotAdvanceAndCheckSnapState(remote_slot->confirmed_lsn,
324 : &found_consistent_snapshot);
325 :
326 : /* Sanity check */
327 10 : if (slot->data.confirmed_flush != remote_slot->confirmed_lsn)
328 0 : ereport(ERROR,
329 : errmsg_internal("synchronized confirmed_flush for slot \"%s\" differs from remote slot",
330 : remote_slot->name),
331 : errdetail_internal("Remote slot has LSN %X/%08X but local slot has LSN %X/%08X.",
332 : LSN_FORMAT_ARGS(remote_slot->confirmed_lsn),
333 : LSN_FORMAT_ARGS(slot->data.confirmed_flush)));
334 :
335 : /*
336 : * If we can't reach a consistent snapshot, the slot won't be
337 : * persisted. See update_and_persist_local_synced_slot().
338 : */
339 10 : if (!found_consistent_snapshot)
340 : {
341 : Assert(MyReplicationSlot->data.persistency == RS_TEMPORARY);
342 :
343 0 : ereport(LOG,
344 : errmsg("could not synchronize replication slot \"%s\"",
345 : remote_slot->name),
346 : errdetail("Synchronization could lead to data loss, because the standby could not build a consistent snapshot to decode WALs at LSN %X/%08X.",
347 : LSN_FORMAT_ARGS(slot->data.restart_lsn)));
348 :
349 0 : skip_reason = SS_SKIP_NO_CONSISTENT_SNAPSHOT;
350 : }
351 : }
352 :
353 13 : updated_xmin_or_lsn = true;
354 : }
355 :
356 : /* Update slot sync skip stats */
357 46 : update_slotsync_skip_stats(skip_reason);
358 :
359 46 : if (remote_dbid != slot->data.database ||
360 46 : remote_slot->two_phase != slot->data.two_phase ||
361 45 : remote_slot->failover != slot->data.failover ||
362 45 : strcmp(remote_slot->plugin, NameStr(slot->data.plugin)) != 0 ||
363 45 : remote_slot->two_phase_at != slot->data.two_phase_at)
364 : {
365 : NameData plugin_name;
366 :
367 : /* Avoid expensive operations while holding a spinlock. */
368 1 : namestrcpy(&plugin_name, remote_slot->plugin);
369 :
370 1 : SpinLockAcquire(&slot->mutex);
371 1 : slot->data.plugin = plugin_name;
372 1 : slot->data.database = remote_dbid;
373 1 : slot->data.two_phase = remote_slot->two_phase;
374 1 : slot->data.two_phase_at = remote_slot->two_phase_at;
375 1 : slot->data.failover = remote_slot->failover;
376 1 : SpinLockRelease(&slot->mutex);
377 :
378 1 : updated_config = true;
379 :
380 : /*
381 : * Ensure that there is no risk of sending prepared transactions
382 : * unexpectedly after the promotion.
383 : */
384 : Assert(slot->data.two_phase_at <= slot->data.confirmed_flush);
385 : }
386 :
387 : /*
388 : * We have to write the changed xmin to disk *before* we change the
389 : * in-memory value, otherwise after a crash we wouldn't know that some
390 : * catalog tuples might have been removed already.
391 : */
392 46 : if (updated_config || updated_xmin_or_lsn)
393 : {
394 14 : ReplicationSlotMarkDirty();
395 14 : ReplicationSlotSave();
396 : }
397 :
398 : /*
399 : * Now the new xmin is safely on disk, we can let the global value
400 : * advance. We do not take ProcArrayLock or similar since we only advance
401 : * xmin here and there's not much harm done by a concurrent computation
402 : * missing that.
403 : */
404 46 : if (updated_xmin_or_lsn)
405 : {
406 13 : SpinLockAcquire(&slot->mutex);
407 13 : slot->effective_catalog_xmin = remote_slot->catalog_xmin;
408 13 : SpinLockRelease(&slot->mutex);
409 :
410 13 : ReplicationSlotsComputeRequiredXmin(false);
411 13 : ReplicationSlotsComputeRequiredLSN();
412 : }
413 :
414 46 : return updated_config || updated_xmin_or_lsn;
415 : }
416 :
417 : /*
418 : * Get the list of local logical slots that are synchronized from the
419 : * primary server.
420 : */
421 : static List *
422 29 : get_local_synced_slots(void)
423 : {
424 29 : List *local_slots = NIL;
425 :
426 29 : LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
427 :
428 319 : for (int i = 0; i < max_replication_slots; i++)
429 : {
430 290 : ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
431 :
432 : /* Check if it is a synchronized slot */
433 290 : if (s->in_use && s->data.synced)
434 : {
435 : Assert(SlotIsLogical(s));
436 45 : local_slots = lappend(local_slots, s);
437 : }
438 : }
439 :
440 29 : LWLockRelease(ReplicationSlotControlLock);
441 :
442 29 : return local_slots;
443 : }
444 :
445 : /*
446 : * Helper function to check if local_slot is required to be retained.
447 : *
448 : * Return false either if local_slot does not exist in the remote_slots list
449 : * or is invalidated while the corresponding remote slot is still valid,
450 : * otherwise true.
451 : */
452 : static bool
453 45 : local_sync_slot_required(ReplicationSlot *local_slot, List *remote_slots)
454 : {
455 45 : bool remote_exists = false;
456 45 : bool locally_invalidated = false;
457 :
458 110 : foreach_ptr(RemoteSlot, remote_slot, remote_slots)
459 : {
460 64 : if (strcmp(remote_slot->name, NameStr(local_slot->data.name)) == 0)
461 : {
462 44 : remote_exists = true;
463 :
464 : /*
465 : * If remote slot is not invalidated but local slot is marked as
466 : * invalidated, then set locally_invalidated flag.
467 : */
468 44 : SpinLockAcquire(&local_slot->mutex);
469 44 : locally_invalidated =
470 88 : (remote_slot->invalidated == RS_INVAL_NONE) &&
471 44 : (local_slot->data.invalidated != RS_INVAL_NONE);
472 44 : SpinLockRelease(&local_slot->mutex);
473 :
474 44 : break;
475 : }
476 : }
477 :
478 45 : return (remote_exists && !locally_invalidated);
479 : }
480 :
481 : /*
482 : * Drop local obsolete slots.
483 : *
484 : * Drop the local slots that no longer need to be synced i.e. these either do
485 : * not exist on the primary or are no longer enabled for failover.
486 : *
487 : * Additionally, drop any slots that are valid on the primary but got
488 : * invalidated on the standby. This situation may occur due to the following
489 : * reasons:
490 : * - The 'max_slot_wal_keep_size' on the standby is insufficient to retain WAL
491 : * records from the restart_lsn of the slot.
492 : * - 'primary_slot_name' is temporarily reset to null and the physical slot is
493 : * removed.
494 : * These dropped slots will get recreated in next sync-cycle and it is okay to
495 : * drop and recreate such slots as long as these are not consumable on the
496 : * standby (which is the case currently).
497 : *
498 : * Note: Change of 'wal_level' on the primary server to a level lower than
499 : * logical may also result in slot invalidation and removal on the standby.
500 : * This is because such 'wal_level' change is only possible if the logical
501 : * slots are removed on the primary server, so it's expected to see the
502 : * slots being invalidated and removed on the standby too (and re-created
503 : * if they are re-created on the primary server).
504 : */
505 : static void
506 29 : drop_local_obsolete_slots(List *remote_slot_list)
507 : {
508 29 : List *local_slots = get_local_synced_slots();
509 :
510 103 : foreach_ptr(ReplicationSlot, local_slot, local_slots)
511 : {
512 : /* Drop the local slot if it is not required to be retained. */
513 45 : if (!local_sync_slot_required(local_slot, remote_slot_list))
514 : {
515 : bool synced_slot;
516 :
517 : /*
518 : * Use shared lock to prevent a conflict with
519 : * ReplicationSlotsDropDBSlots(), trying to drop the same slot
520 : * during a drop-database operation.
521 : */
522 2 : LockSharedObject(DatabaseRelationId, local_slot->data.database,
523 : 0, AccessShareLock);
524 :
525 : /*
526 : * In the small window between getting the slot to drop and
527 : * locking the database, there is a possibility of a parallel
528 : * database drop by the startup process and the creation of a new
529 : * slot by the user. This new user-created slot may end up using
530 : * the same shared memory as that of 'local_slot'. Thus check if
531 : * local_slot is still the synced one before performing actual
532 : * drop.
533 : */
534 2 : SpinLockAcquire(&local_slot->mutex);
535 2 : synced_slot = local_slot->in_use && local_slot->data.synced;
536 2 : SpinLockRelease(&local_slot->mutex);
537 :
538 2 : if (synced_slot)
539 : {
540 2 : ReplicationSlotAcquire(NameStr(local_slot->data.name), true, false);
541 2 : ReplicationSlotDropAcquired();
542 : }
543 :
544 2 : UnlockSharedObject(DatabaseRelationId, local_slot->data.database,
545 : 0, AccessShareLock);
546 :
547 2 : ereport(LOG,
548 : errmsg("dropped replication slot \"%s\" of database with OID %u",
549 : NameStr(local_slot->data.name),
550 : local_slot->data.database));
551 : }
552 : }
553 29 : }
554 :
555 : /*
556 : * Reserve WAL for the currently active local slot using the specified WAL
557 : * location (restart_lsn).
558 : *
559 : * If the given WAL location has been removed or is at risk of removal,
560 : * reserve WAL using the oldest segment that is non-removable.
561 : */
562 : static void
563 7 : reserve_wal_for_local_slot(XLogRecPtr restart_lsn)
564 : {
565 : XLogRecPtr slot_min_lsn;
566 : XLogRecPtr min_safe_lsn;
567 : XLogSegNo segno;
568 7 : ReplicationSlot *slot = MyReplicationSlot;
569 :
570 : Assert(slot != NULL);
571 : Assert(!XLogRecPtrIsValid(slot->data.restart_lsn));
572 :
573 : /*
574 : * Acquire an exclusive lock to prevent the checkpoint process from
575 : * concurrently calculating the minimum slot LSN (see
576 : * CheckPointReplicationSlots), ensuring that if WAL reservation occurs
577 : * first, the checkpoint must wait for the restart_lsn update before
578 : * calculating the minimum LSN.
579 : *
580 : * Note: Unlike ReplicationSlotReserveWal(), this lock does not protect a
581 : * newly synced slot from being invalidated if a concurrent checkpoint has
582 : * invoked CheckPointReplicationSlots() before the WAL reservation here.
583 : * This can happen because the initial restart_lsn received from the
584 : * remote server can precede the redo pointer. Therefore, when selecting
585 : * the initial restart_lsn, we consider using the redo pointer or the
586 : * minimum slot LSN (if those values are greater than the remote
587 : * restart_lsn) instead of relying solely on the remote value.
588 : */
589 7 : LWLockAcquire(ReplicationSlotAllocationLock, LW_EXCLUSIVE);
590 :
591 : /*
592 : * Determine the minimum non-removable LSN by comparing the redo pointer
593 : * with the minimum slot LSN.
594 : *
595 : * The minimum slot LSN is considered because the redo pointer advances at
596 : * every checkpoint, even when replication slots are present on the
597 : * standby. In such scenarios, the redo pointer can exceed the remote
598 : * restart_lsn, while WALs preceding the remote restart_lsn remain
599 : * protected by a local replication slot.
600 : */
601 7 : min_safe_lsn = GetRedoRecPtr();
602 7 : slot_min_lsn = XLogGetReplicationSlotMinimumLSN();
603 :
604 7 : if (XLogRecPtrIsValid(slot_min_lsn) && min_safe_lsn > slot_min_lsn)
605 0 : min_safe_lsn = slot_min_lsn;
606 :
607 : /*
608 : * If the minimum safe LSN is greater than the given restart_lsn, use it
609 : * as the initial restart_lsn for the newly synced slot. Otherwise, use
610 : * the given remote restart_lsn.
611 : */
612 7 : SpinLockAcquire(&slot->mutex);
613 7 : slot->data.restart_lsn = Max(restart_lsn, min_safe_lsn);
614 7 : SpinLockRelease(&slot->mutex);
615 :
616 7 : ReplicationSlotsComputeRequiredLSN();
617 :
618 7 : XLByteToSeg(slot->data.restart_lsn, segno, wal_segment_size);
619 7 : if (XLogGetLastRemovedSegno() >= segno)
620 0 : elog(ERROR, "WAL required by replication slot %s has been removed concurrently",
621 : NameStr(slot->data.name));
622 :
623 7 : LWLockRelease(ReplicationSlotAllocationLock);
624 7 : }
625 :
626 : /*
627 : * If the remote restart_lsn and catalog_xmin have caught up with the
628 : * local ones, then update the LSNs and persist the local synced slot for
629 : * future synchronization; otherwise, do nothing.
630 : *
631 : * *slot_persistence_pending is set to true if any of the slots fail to
632 : * persist.
633 : *
634 : * Return true if the slot is marked as RS_PERSISTENT (sync-ready), otherwise
635 : * false.
636 : */
637 : static bool
638 10 : update_and_persist_local_synced_slot(RemoteSlot *remote_slot, Oid remote_dbid,
639 : bool *slot_persistence_pending)
640 : {
641 10 : ReplicationSlot *slot = MyReplicationSlot;
642 :
643 : /* Slotsync skip stats are handled in function update_local_synced_slot() */
644 10 : (void) update_local_synced_slot(remote_slot, remote_dbid);
645 :
646 : /*
647 : * Check if the slot cannot be synchronized. Refer to the comment atop the
648 : * file for details on this check.
649 : */
650 10 : if (slot->slotsync_skip_reason != SS_SKIP_NONE)
651 : {
652 : /*
653 : * We reach this point when the remote slot didn't catch up to locally
654 : * reserved position, or it cannot reach the consistent point from the
655 : * restart_lsn, or the WAL prior to the remote confirmed flush LSN has
656 : * not been received and flushed.
657 : *
658 : * We do not drop the slot because the restart_lsn and confirmed_lsn
659 : * can be ahead of the current location when recreating the slot in
660 : * the next cycle. It may take more time to create such a slot or
661 : * reach the consistent point. Therefore, we keep this slot and
662 : * attempt the synchronization in the next cycle.
663 : *
664 : * We also update the slot_persistence_pending parameter, so the SQL
665 : * function can retry.
666 : */
667 4 : if (slot_persistence_pending)
668 2 : *slot_persistence_pending = true;
669 :
670 4 : return false;
671 : }
672 :
673 6 : ReplicationSlotPersist();
674 :
675 6 : ereport(LOG,
676 : errmsg("newly created replication slot \"%s\" is sync-ready now",
677 : remote_slot->name));
678 :
679 6 : return true;
680 : }
681 :
682 : /*
683 : * Synchronize a single slot to the given position.
684 : *
685 : * This creates a new slot if there is no existing one and updates the
686 : * metadata of the slot as per the data received from the primary server.
687 : *
688 : * The slot is created as a temporary slot and stays in the same state until the
689 : * remote_slot catches up with locally reserved position and local slot is
690 : * updated. The slot is then persisted and is considered as sync-ready for
691 : * periodic syncs.
692 : *
693 : * *slot_persistence_pending is set to true if any of the slots fail to
694 : * persist.
695 : *
696 : * Returns TRUE if the local slot is updated.
697 : */
698 : static bool
699 50 : synchronize_one_slot(RemoteSlot *remote_slot, Oid remote_dbid,
700 : bool *slot_persistence_pending)
701 : {
702 : ReplicationSlot *slot;
703 50 : bool slot_updated = false;
704 :
705 : /* Search for the named slot */
706 50 : if ((slot = SearchNamedReplicationSlot(remote_slot->name, true)))
707 : {
708 : bool synced;
709 :
710 43 : SpinLockAcquire(&slot->mutex);
711 43 : synced = slot->data.synced;
712 43 : SpinLockRelease(&slot->mutex);
713 :
714 : /* User-created slot with the same name exists, raise ERROR. */
715 43 : if (!synced)
716 0 : ereport(ERROR,
717 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
718 : errmsg("exiting from slot synchronization because same"
719 : " name slot \"%s\" already exists on the standby",
720 : remote_slot->name));
721 :
722 : /*
723 : * The slot has been synchronized before.
724 : *
725 : * It is important to acquire the slot here before checking
726 : * invalidation. If we don't acquire the slot first, there could be a
727 : * race condition that the local slot could be invalidated just after
728 : * checking the 'invalidated' flag here and we could end up
729 : * overwriting 'invalidated' flag to remote_slot's value. See
730 : * InvalidatePossiblyObsoleteSlot() where it invalidates slot directly
731 : * if the slot is not acquired by other processes.
732 : *
733 : * XXX: If it ever turns out that slot acquire/release is costly for
734 : * cases when none of the slot properties is changed then we can do a
735 : * pre-check to ensure that at least one of the slot properties is
736 : * changed before acquiring the slot.
737 : */
738 43 : ReplicationSlotAcquire(remote_slot->name, true, false);
739 :
740 : Assert(slot == MyReplicationSlot);
741 :
742 : /*
743 : * Copy the invalidation cause from remote only if local slot is not
744 : * invalidated locally, we don't want to overwrite existing one.
745 : */
746 43 : if (slot->data.invalidated == RS_INVAL_NONE &&
747 43 : remote_slot->invalidated != RS_INVAL_NONE)
748 : {
749 0 : SpinLockAcquire(&slot->mutex);
750 0 : slot->data.invalidated = remote_slot->invalidated;
751 0 : SpinLockRelease(&slot->mutex);
752 :
753 : /* Make sure the invalidated state persists across server restart */
754 0 : ReplicationSlotMarkDirty();
755 0 : ReplicationSlotSave();
756 :
757 0 : slot_updated = true;
758 : }
759 :
760 : /* Skip the sync of an invalidated slot */
761 43 : if (slot->data.invalidated != RS_INVAL_NONE)
762 : {
763 0 : update_slotsync_skip_stats(SS_SKIP_INVALID);
764 :
765 0 : ReplicationSlotRelease();
766 0 : return slot_updated;
767 : }
768 :
769 : /* Slot not ready yet, let's attempt to make it sync-ready now. */
770 43 : if (slot->data.persistency == RS_TEMPORARY)
771 : {
772 3 : slot_updated = update_and_persist_local_synced_slot(remote_slot,
773 : remote_dbid,
774 : slot_persistence_pending);
775 : }
776 :
777 : /* Slot ready for sync, so sync it. */
778 : else
779 : {
780 : /*
781 : * Sanity check: As long as the invalidations are handled
782 : * appropriately as above, this should never happen.
783 : *
784 : * We don't need to check restart_lsn here. See the comments in
785 : * update_local_synced_slot() for details.
786 : */
787 40 : if (remote_slot->confirmed_lsn < slot->data.confirmed_flush)
788 0 : ereport(ERROR,
789 : errmsg_internal("cannot synchronize local slot \"%s\"",
790 : remote_slot->name),
791 : errdetail_internal("Local slot's start streaming location LSN(%X/%08X) is ahead of remote slot's LSN(%X/%08X).",
792 : LSN_FORMAT_ARGS(slot->data.confirmed_flush),
793 : LSN_FORMAT_ARGS(remote_slot->confirmed_lsn)));
794 :
795 40 : slot_updated = update_local_synced_slot(remote_slot, remote_dbid);
796 : }
797 : }
798 : /* Otherwise create the slot first. */
799 : else
800 : {
801 : NameData plugin_name;
802 7 : TransactionId xmin_horizon = InvalidTransactionId;
803 :
804 : /* Skip creating the local slot if remote_slot is invalidated already */
805 7 : if (remote_slot->invalidated != RS_INVAL_NONE)
806 0 : return false;
807 :
808 : /*
809 : * We create temporary slots instead of ephemeral slots here because
810 : * we want the slots to survive after releasing them. This is done to
811 : * avoid dropping and re-creating the slots in each synchronization
812 : * cycle if the restart_lsn or catalog_xmin of the remote slot has not
813 : * caught up.
814 : */
815 7 : ReplicationSlotCreate(remote_slot->name, true, RS_TEMPORARY,
816 7 : remote_slot->two_phase,
817 7 : remote_slot->failover,
818 : true);
819 :
820 : /* For shorter lines. */
821 7 : slot = MyReplicationSlot;
822 :
823 : /* Avoid expensive operations while holding a spinlock. */
824 7 : namestrcpy(&plugin_name, remote_slot->plugin);
825 :
826 7 : SpinLockAcquire(&slot->mutex);
827 7 : slot->data.database = remote_dbid;
828 7 : slot->data.plugin = plugin_name;
829 7 : SpinLockRelease(&slot->mutex);
830 :
831 7 : reserve_wal_for_local_slot(remote_slot->restart_lsn);
832 :
833 7 : LWLockAcquire(ReplicationSlotControlLock, LW_EXCLUSIVE);
834 7 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
835 7 : xmin_horizon = GetOldestSafeDecodingTransactionId(true);
836 7 : SpinLockAcquire(&slot->mutex);
837 7 : slot->effective_catalog_xmin = xmin_horizon;
838 7 : slot->data.catalog_xmin = xmin_horizon;
839 7 : SpinLockRelease(&slot->mutex);
840 7 : ReplicationSlotsComputeRequiredXmin(true);
841 7 : LWLockRelease(ProcArrayLock);
842 7 : LWLockRelease(ReplicationSlotControlLock);
843 :
844 7 : update_and_persist_local_synced_slot(remote_slot, remote_dbid,
845 : slot_persistence_pending);
846 :
847 7 : slot_updated = true;
848 : }
849 :
850 50 : ReplicationSlotRelease();
851 :
852 50 : return slot_updated;
853 : }
854 :
855 : /*
856 : * Fetch remote slots.
857 : *
858 : * If slot_names is NIL, fetches all failover logical slots from the
859 : * primary server, otherwise fetches only the ones with names in slot_names.
860 : *
861 : * Returns a list of remote slot information structures, or NIL if none
862 : * are found.
863 : */
864 : static List *
865 30 : fetch_remote_slots(WalReceiverConn *wrconn, List *slot_names)
866 : {
867 : #define SLOTSYNC_COLUMN_COUNT 10
868 30 : Oid slotRow[SLOTSYNC_COLUMN_COUNT] = {TEXTOID, TEXTOID, LSNOID,
869 : LSNOID, XIDOID, BOOLOID, LSNOID, BOOLOID, TEXTOID, TEXTOID};
870 :
871 : WalRcvExecResult *res;
872 : TupleTableSlot *tupslot;
873 30 : List *remote_slot_list = NIL;
874 : StringInfoData query;
875 :
876 30 : initStringInfo(&query);
877 30 : appendStringInfoString(&query,
878 : "SELECT slot_name, plugin, confirmed_flush_lsn,"
879 : " restart_lsn, catalog_xmin, two_phase,"
880 : " two_phase_at, failover,"
881 : " database, invalidation_reason"
882 : " FROM pg_catalog.pg_replication_slots"
883 : " WHERE failover and NOT temporary");
884 :
885 30 : if (slot_names != NIL)
886 : {
887 2 : bool first_slot = true;
888 :
889 : /*
890 : * Construct the query to fetch only the specified slots
891 : */
892 2 : appendStringInfoString(&query, " AND slot_name IN (");
893 :
894 6 : foreach_ptr(char, slot_name, slot_names)
895 : {
896 2 : if (!first_slot)
897 0 : appendStringInfoString(&query, ", ");
898 :
899 2 : appendStringInfo(&query, "%s", quote_literal_cstr(slot_name));
900 2 : first_slot = false;
901 : }
902 2 : appendStringInfoChar(&query, ')');
903 : }
904 :
905 : /* Execute the query */
906 30 : res = walrcv_exec(wrconn, query.data, SLOTSYNC_COLUMN_COUNT, slotRow);
907 30 : pfree(query.data);
908 30 : if (res->status != WALRCV_OK_TUPLES)
909 1 : ereport(ERROR,
910 : errmsg("could not fetch failover logical slots info from the primary server: %s",
911 : res->err));
912 :
913 29 : tupslot = MakeSingleTupleTableSlot(res->tupledesc, &TTSOpsMinimalTuple);
914 79 : while (tuplestore_gettupleslot(res->tuplestore, true, false, tupslot))
915 : {
916 : bool isnull;
917 50 : RemoteSlot *remote_slot = palloc0_object(RemoteSlot);
918 : Datum d;
919 50 : int col = 0;
920 :
921 50 : remote_slot->name = TextDatumGetCString(slot_getattr(tupslot, ++col,
922 : &isnull));
923 : Assert(!isnull);
924 :
925 50 : remote_slot->plugin = TextDatumGetCString(slot_getattr(tupslot, ++col,
926 : &isnull));
927 : Assert(!isnull);
928 :
929 : /*
930 : * It is possible to get null values for LSN and Xmin if slot is
931 : * invalidated on the primary server, so handle accordingly.
932 : */
933 50 : d = slot_getattr(tupslot, ++col, &isnull);
934 50 : remote_slot->confirmed_lsn = isnull ? InvalidXLogRecPtr :
935 50 : DatumGetLSN(d);
936 :
937 50 : d = slot_getattr(tupslot, ++col, &isnull);
938 50 : remote_slot->restart_lsn = isnull ? InvalidXLogRecPtr : DatumGetLSN(d);
939 :
940 50 : d = slot_getattr(tupslot, ++col, &isnull);
941 50 : remote_slot->catalog_xmin = isnull ? InvalidTransactionId :
942 50 : DatumGetTransactionId(d);
943 :
944 50 : remote_slot->two_phase = DatumGetBool(slot_getattr(tupslot, ++col,
945 : &isnull));
946 : Assert(!isnull);
947 :
948 50 : d = slot_getattr(tupslot, ++col, &isnull);
949 50 : remote_slot->two_phase_at = isnull ? InvalidXLogRecPtr : DatumGetLSN(d);
950 :
951 50 : remote_slot->failover = DatumGetBool(slot_getattr(tupslot, ++col,
952 : &isnull));
953 : Assert(!isnull);
954 :
955 50 : remote_slot->database = TextDatumGetCString(slot_getattr(tupslot,
956 : ++col, &isnull));
957 : Assert(!isnull);
958 :
959 50 : d = slot_getattr(tupslot, ++col, &isnull);
960 50 : remote_slot->invalidated = isnull ? RS_INVAL_NONE :
961 0 : GetSlotInvalidationCause(TextDatumGetCString(d));
962 :
963 : /* Sanity check */
964 : Assert(col == SLOTSYNC_COLUMN_COUNT);
965 :
966 : /*
967 : * If restart_lsn, confirmed_lsn or catalog_xmin is invalid but the
968 : * slot is valid, that means we have fetched the remote_slot in its
969 : * RS_EPHEMERAL state. In such a case, don't sync it; we can always
970 : * sync it in the next sync cycle when the remote_slot is persisted
971 : * and has valid lsn(s) and xmin values.
972 : *
973 : * XXX: In future, if we plan to expose 'slot->data.persistency' in
974 : * pg_replication_slots view, then we can avoid fetching RS_EPHEMERAL
975 : * slots in the first place.
976 : */
977 50 : if ((!XLogRecPtrIsValid(remote_slot->restart_lsn) ||
978 50 : !XLogRecPtrIsValid(remote_slot->confirmed_lsn) ||
979 50 : !TransactionIdIsValid(remote_slot->catalog_xmin)) &&
980 0 : remote_slot->invalidated == RS_INVAL_NONE)
981 0 : pfree(remote_slot);
982 : else
983 : /* Create list of remote slots */
984 50 : remote_slot_list = lappend(remote_slot_list, remote_slot);
985 :
986 50 : ExecClearTuple(tupslot);
987 : }
988 :
989 29 : walrcv_clear_result(res);
990 :
991 29 : return remote_slot_list;
992 : }
993 :
994 : /*
995 : * Synchronize slots.
996 : *
997 : * This function takes a list of remote slots and synchronizes them locally. It
998 : * creates the slots if not present on the standby and updates existing ones.
999 : *
1000 : * If slot_persistence_pending is not NULL, it will be set to true if one or
1001 : * more slots could not be persisted. This allows callers such as
1002 : * SyncReplicationSlots() to retry those slots.
1003 : *
1004 : * Returns TRUE if any of the slots gets updated in this sync-cycle.
1005 : */
1006 : static bool
1007 29 : synchronize_slots(WalReceiverConn *wrconn, List *remote_slot_list,
1008 : bool *slot_persistence_pending)
1009 : {
1010 29 : bool some_slot_updated = false;
1011 :
1012 : /* Drop local slots that no longer need to be synced. */
1013 29 : drop_local_obsolete_slots(remote_slot_list);
1014 :
1015 : /* Now sync the slots locally */
1016 108 : foreach_ptr(RemoteSlot, remote_slot, remote_slot_list)
1017 : {
1018 50 : Oid remote_dbid = get_database_oid(remote_slot->database, false);
1019 :
1020 : /*
1021 : * Use shared lock to prevent a conflict with
1022 : * ReplicationSlotsDropDBSlots(), trying to drop the same slot during
1023 : * a drop-database operation.
1024 : */
1025 50 : LockSharedObject(DatabaseRelationId, remote_dbid, 0, AccessShareLock);
1026 :
1027 50 : some_slot_updated |= synchronize_one_slot(remote_slot, remote_dbid,
1028 : slot_persistence_pending);
1029 :
1030 50 : UnlockSharedObject(DatabaseRelationId, remote_dbid, 0, AccessShareLock);
1031 : }
1032 :
1033 29 : return some_slot_updated;
1034 : }
1035 :
1036 : /*
1037 : * Checks the remote server info.
1038 : *
1039 : * We ensure that the 'primary_slot_name' exists on the remote server and the
1040 : * remote server is not a standby node.
1041 : */
1042 : static void
1043 14 : validate_remote_info(WalReceiverConn *wrconn)
1044 : {
1045 : #define PRIMARY_INFO_OUTPUT_COL_COUNT 2
1046 : WalRcvExecResult *res;
1047 14 : Oid slotRow[PRIMARY_INFO_OUTPUT_COL_COUNT] = {BOOLOID, BOOLOID};
1048 : StringInfoData cmd;
1049 : bool isnull;
1050 : TupleTableSlot *tupslot;
1051 : bool remote_in_recovery;
1052 : bool primary_slot_valid;
1053 14 : bool started_tx = false;
1054 :
1055 14 : initStringInfo(&cmd);
1056 14 : appendStringInfo(&cmd,
1057 : "SELECT pg_is_in_recovery(), count(*) = 1"
1058 : " FROM pg_catalog.pg_replication_slots"
1059 : " WHERE slot_type='physical' AND slot_name=%s",
1060 : quote_literal_cstr(PrimarySlotName));
1061 :
1062 : /* The syscache access in walrcv_exec() needs a transaction env. */
1063 14 : if (!IsTransactionState())
1064 : {
1065 5 : StartTransactionCommand();
1066 5 : started_tx = true;
1067 : }
1068 :
1069 14 : res = walrcv_exec(wrconn, cmd.data, PRIMARY_INFO_OUTPUT_COL_COUNT, slotRow);
1070 14 : pfree(cmd.data);
1071 :
1072 14 : if (res->status != WALRCV_OK_TUPLES)
1073 0 : ereport(ERROR,
1074 : errmsg("could not fetch primary slot name \"%s\" info from the primary server: %s",
1075 : PrimarySlotName, res->err),
1076 : errhint("Check if \"primary_slot_name\" is configured correctly."));
1077 :
1078 14 : tupslot = MakeSingleTupleTableSlot(res->tupledesc, &TTSOpsMinimalTuple);
1079 14 : if (!tuplestore_gettupleslot(res->tuplestore, true, false, tupslot))
1080 0 : elog(ERROR,
1081 : "failed to fetch tuple for the primary server slot specified by \"primary_slot_name\"");
1082 :
1083 14 : remote_in_recovery = DatumGetBool(slot_getattr(tupslot, 1, &isnull));
1084 : Assert(!isnull);
1085 :
1086 : /*
1087 : * Slot sync is currently not supported on a cascading standby. This is
1088 : * because if we allow it, the primary server needs to wait for all the
1089 : * cascading standbys, otherwise, logical subscribers can still be ahead
1090 : * of one of the cascading standbys which we plan to promote. Thus, to
1091 : * avoid this additional complexity, we restrict it for the time being.
1092 : */
1093 14 : if (remote_in_recovery)
1094 1 : ereport(ERROR,
1095 : errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1096 : errmsg("cannot synchronize replication slots from a standby server"));
1097 :
1098 13 : primary_slot_valid = DatumGetBool(slot_getattr(tupslot, 2, &isnull));
1099 : Assert(!isnull);
1100 :
1101 13 : if (!primary_slot_valid)
1102 0 : ereport(ERROR,
1103 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1104 : /* translator: second %s is a GUC variable name */
1105 : errmsg("replication slot \"%s\" specified by \"%s\" does not exist on primary server",
1106 : PrimarySlotName, "primary_slot_name"));
1107 :
1108 13 : ExecClearTuple(tupslot);
1109 13 : walrcv_clear_result(res);
1110 :
1111 13 : if (started_tx)
1112 5 : CommitTransactionCommand();
1113 13 : }
1114 :
1115 : /*
1116 : * Checks if dbname is specified in 'primary_conninfo'.
1117 : *
1118 : * Error out if not specified otherwise return it.
1119 : */
1120 : char *
1121 15 : CheckAndGetDbnameFromConninfo(void)
1122 : {
1123 : char *dbname;
1124 :
1125 : /*
1126 : * The slot synchronization needs a database connection for walrcv_exec to
1127 : * work.
1128 : */
1129 15 : dbname = walrcv_get_dbname_from_conninfo(PrimaryConnInfo);
1130 15 : if (dbname == NULL)
1131 1 : ereport(ERROR,
1132 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1133 :
1134 : /*
1135 : * translator: first %s is a connection option; second %s is a GUC
1136 : * variable name
1137 : */
1138 : errmsg("replication slot synchronization requires \"%s\" to be specified in \"%s\"",
1139 : "dbname", "primary_conninfo"));
1140 14 : return dbname;
1141 : }
1142 :
1143 : /*
1144 : * Return true if all necessary GUCs for slot synchronization are set
1145 : * appropriately, otherwise, return false.
1146 : */
1147 : bool
1148 23 : ValidateSlotSyncParams(int elevel)
1149 : {
1150 : /*
1151 : * Logical slot sync/creation requires logical decoding to be enabled.
1152 : */
1153 23 : if (!IsLogicalDecodingEnabled())
1154 : {
1155 0 : ereport(elevel,
1156 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1157 : errmsg("replication slot synchronization requires \"effective_wal_level\" >= \"logical\" on the primary"),
1158 : errhint("To enable logical decoding on primary, set \"wal_level\" >= \"logical\" or create at least one logical slot when \"wal_level\" = \"replica\"."));
1159 :
1160 0 : return false;
1161 : }
1162 :
1163 : /*
1164 : * A physical replication slot(primary_slot_name) is required on the
1165 : * primary to ensure that the rows needed by the standby are not removed
1166 : * after restarting, so that the synchronized slot on the standby will not
1167 : * be invalidated.
1168 : */
1169 23 : if (PrimarySlotName == NULL || *PrimarySlotName == '\0')
1170 : {
1171 0 : ereport(elevel,
1172 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1173 : /* translator: %s is a GUC variable name */
1174 : errmsg("replication slot synchronization requires \"%s\" to be set", "primary_slot_name"));
1175 0 : return false;
1176 : }
1177 :
1178 : /*
1179 : * hot_standby_feedback must be enabled to cooperate with the physical
1180 : * replication slot, which allows informing the primary about the xmin and
1181 : * catalog_xmin values on the standby.
1182 : */
1183 23 : if (!hot_standby_feedback)
1184 : {
1185 1 : ereport(elevel,
1186 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1187 : /* translator: %s is a GUC variable name */
1188 : errmsg("replication slot synchronization requires \"%s\" to be enabled",
1189 : "hot_standby_feedback"));
1190 1 : return false;
1191 : }
1192 :
1193 : /*
1194 : * The primary_conninfo is required to make connection to primary for
1195 : * getting slots information.
1196 : */
1197 22 : if (PrimaryConnInfo == NULL || *PrimaryConnInfo == '\0')
1198 : {
1199 0 : ereport(elevel,
1200 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1201 : /* translator: %s is a GUC variable name */
1202 : errmsg("replication slot synchronization requires \"%s\" to be set",
1203 : "primary_conninfo"));
1204 0 : return false;
1205 : }
1206 :
1207 22 : return true;
1208 : }
1209 :
1210 : /*
1211 : * Re-read the config file for slot synchronization.
1212 : *
1213 : * Exit or throw error if relevant GUCs have changed depending on whether
1214 : * called from slot sync worker or from the SQL function pg_sync_replication_slots()
1215 : */
1216 : static void
1217 1 : slotsync_reread_config(void)
1218 : {
1219 1 : char *old_primary_conninfo = pstrdup(PrimaryConnInfo);
1220 1 : char *old_primary_slotname = pstrdup(PrimarySlotName);
1221 1 : bool old_sync_replication_slots = sync_replication_slots;
1222 1 : bool old_hot_standby_feedback = hot_standby_feedback;
1223 : bool conninfo_changed;
1224 : bool primary_slotname_changed;
1225 1 : bool is_slotsync_worker = AmLogicalSlotSyncWorkerProcess();
1226 1 : bool parameter_changed = false;
1227 :
1228 : if (is_slotsync_worker)
1229 : Assert(sync_replication_slots);
1230 :
1231 1 : ConfigReloadPending = false;
1232 1 : ProcessConfigFile(PGC_SIGHUP);
1233 :
1234 1 : conninfo_changed = strcmp(old_primary_conninfo, PrimaryConnInfo) != 0;
1235 1 : primary_slotname_changed = strcmp(old_primary_slotname, PrimarySlotName) != 0;
1236 1 : pfree(old_primary_conninfo);
1237 1 : pfree(old_primary_slotname);
1238 :
1239 1 : if (old_sync_replication_slots != sync_replication_slots)
1240 : {
1241 0 : if (is_slotsync_worker)
1242 : {
1243 0 : ereport(LOG,
1244 : /* translator: %s is a GUC variable name */
1245 : errmsg("replication slot synchronization worker will stop because \"%s\" is disabled",
1246 : "sync_replication_slots"));
1247 :
1248 0 : proc_exit(0);
1249 : }
1250 :
1251 0 : parameter_changed = true;
1252 : }
1253 : else
1254 : {
1255 1 : if (conninfo_changed ||
1256 1 : primary_slotname_changed ||
1257 1 : (old_hot_standby_feedback != hot_standby_feedback))
1258 : {
1259 :
1260 1 : if (is_slotsync_worker)
1261 : {
1262 1 : ereport(LOG,
1263 : errmsg("replication slot synchronization worker will restart because of a parameter change"));
1264 :
1265 : /*
1266 : * Reset the last-start time for this worker so that the
1267 : * postmaster can restart it without waiting for
1268 : * SLOTSYNC_RESTART_INTERVAL_SEC.
1269 : */
1270 1 : SlotSyncCtx->last_start_time = 0;
1271 :
1272 1 : proc_exit(0);
1273 : }
1274 :
1275 0 : parameter_changed = true;
1276 : }
1277 : }
1278 :
1279 : /*
1280 : * If we have reached here with a parameter change, we must be running in
1281 : * SQL function, emit error in such a case.
1282 : */
1283 0 : if (parameter_changed)
1284 : {
1285 : Assert(!is_slotsync_worker);
1286 0 : ereport(ERROR,
1287 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1288 : errmsg("replication slot synchronization will stop because of a parameter change"));
1289 : }
1290 :
1291 0 : }
1292 :
1293 : /*
1294 : * Interrupt handler for process performing slot synchronization.
1295 : */
1296 : static void
1297 43 : ProcessSlotSyncInterrupts(void)
1298 : {
1299 43 : CHECK_FOR_INTERRUPTS();
1300 :
1301 41 : if (SlotSyncCtx->stopSignaled)
1302 : {
1303 1 : if (AmLogicalSlotSyncWorkerProcess())
1304 : {
1305 1 : ereport(LOG,
1306 : errmsg("replication slot synchronization worker will stop because promotion is triggered"));
1307 :
1308 1 : proc_exit(0);
1309 : }
1310 : else
1311 : {
1312 : /*
1313 : * For the backend executing SQL function
1314 : * pg_sync_replication_slots().
1315 : */
1316 0 : ereport(ERROR,
1317 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1318 : errmsg("replication slot synchronization will stop because promotion is triggered"));
1319 : }
1320 : }
1321 :
1322 40 : if (ConfigReloadPending)
1323 1 : slotsync_reread_config();
1324 39 : }
1325 :
1326 : /*
1327 : * Connection cleanup function for slotsync worker.
1328 : *
1329 : * Called on slotsync worker exit.
1330 : */
1331 : static void
1332 5 : slotsync_worker_disconnect(int code, Datum arg)
1333 : {
1334 5 : WalReceiverConn *wrconn = (WalReceiverConn *) DatumGetPointer(arg);
1335 :
1336 5 : walrcv_disconnect(wrconn);
1337 5 : }
1338 :
1339 : /*
1340 : * Cleanup function for slotsync worker.
1341 : *
1342 : * Called on slotsync worker exit.
1343 : */
1344 : static void
1345 5 : slotsync_worker_onexit(int code, Datum arg)
1346 : {
1347 : /*
1348 : * We need to do slots cleanup here just like WalSndErrorCleanup() does.
1349 : *
1350 : * The startup process during promotion invokes ShutDownSlotSync() which
1351 : * waits for slot sync to finish and it does that by checking the
1352 : * 'syncing' flag. Thus the slot sync worker must be done with slots'
1353 : * release and cleanup to avoid any dangling temporary slots or active
1354 : * slots before it marks itself as finished syncing.
1355 : */
1356 :
1357 : /* Make sure active replication slots are released */
1358 5 : if (MyReplicationSlot != NULL)
1359 0 : ReplicationSlotRelease();
1360 :
1361 : /* Also cleanup the temporary slots. */
1362 5 : ReplicationSlotCleanup(false);
1363 :
1364 5 : SpinLockAcquire(&SlotSyncCtx->mutex);
1365 :
1366 5 : SlotSyncCtx->pid = InvalidPid;
1367 :
1368 : /*
1369 : * If syncing_slots is true, it indicates that the process errored out
1370 : * without resetting the flag. So, we need to clean up shared memory and
1371 : * reset the flag here.
1372 : */
1373 5 : if (syncing_slots)
1374 : {
1375 5 : SlotSyncCtx->syncing = false;
1376 5 : syncing_slots = false;
1377 : }
1378 :
1379 5 : SpinLockRelease(&SlotSyncCtx->mutex);
1380 5 : }
1381 :
1382 : /*
1383 : * Sleep for long enough that we believe it's likely that the slots on primary
1384 : * get updated.
1385 : *
1386 : * If there is no slot activity the wait time between sync-cycles will double
1387 : * (to a maximum of 30s). If there is some slot activity the wait time between
1388 : * sync-cycles is reset to the minimum (200ms).
1389 : */
1390 : static void
1391 21 : wait_for_slot_activity(bool some_slot_updated)
1392 : {
1393 : int rc;
1394 :
1395 21 : if (!some_slot_updated)
1396 : {
1397 : /*
1398 : * No slots were updated, so double the sleep time, but not beyond the
1399 : * maximum allowable value.
1400 : */
1401 12 : sleep_ms = Min(sleep_ms * 2, MAX_SLOTSYNC_WORKER_NAPTIME_MS);
1402 : }
1403 : else
1404 : {
1405 : /*
1406 : * Some slots were updated since the last sleep, so reset the sleep
1407 : * time.
1408 : */
1409 9 : sleep_ms = MIN_SLOTSYNC_WORKER_NAPTIME_MS;
1410 : }
1411 :
1412 21 : rc = WaitLatch(MyLatch,
1413 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
1414 : sleep_ms,
1415 : WAIT_EVENT_REPLICATION_SLOTSYNC_MAIN);
1416 :
1417 21 : if (rc & WL_LATCH_SET)
1418 3 : ResetLatch(MyLatch);
1419 21 : }
1420 :
1421 : /*
1422 : * Emit an error if a concurrent sync call is in progress.
1423 : * Otherwise, advertise that a sync is in progress.
1424 : */
1425 : static void
1426 14 : check_and_set_sync_info(pid_t sync_process_pid)
1427 : {
1428 14 : SpinLockAcquire(&SlotSyncCtx->mutex);
1429 :
1430 14 : if (SlotSyncCtx->syncing)
1431 : {
1432 0 : SpinLockRelease(&SlotSyncCtx->mutex);
1433 0 : ereport(ERROR,
1434 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1435 : errmsg("cannot synchronize replication slots concurrently"));
1436 : }
1437 :
1438 : /* The pid must not be already assigned in SlotSyncCtx */
1439 : Assert(SlotSyncCtx->pid == InvalidPid);
1440 :
1441 14 : SlotSyncCtx->syncing = true;
1442 :
1443 : /*
1444 : * Advertise the required PID so that the startup process can kill the
1445 : * slot sync process on promotion.
1446 : */
1447 14 : SlotSyncCtx->pid = sync_process_pid;
1448 :
1449 14 : SpinLockRelease(&SlotSyncCtx->mutex);
1450 :
1451 14 : syncing_slots = true;
1452 14 : }
1453 :
1454 : /*
1455 : * Reset syncing flag.
1456 : */
1457 : static void
1458 9 : reset_syncing_flag(void)
1459 : {
1460 9 : SpinLockAcquire(&SlotSyncCtx->mutex);
1461 9 : SlotSyncCtx->syncing = false;
1462 9 : SlotSyncCtx->pid = InvalidPid;
1463 9 : SpinLockRelease(&SlotSyncCtx->mutex);
1464 :
1465 9 : syncing_slots = false;
1466 9 : }
1467 :
1468 : /*
1469 : * The main loop of our worker process.
1470 : *
1471 : * It connects to the primary server, fetches logical failover slots
1472 : * information periodically in order to create and sync the slots.
1473 : *
1474 : * Note: If any changes are made here, check if the corresponding SQL
1475 : * function logic in SyncReplicationSlots() also needs to be changed.
1476 : */
1477 : void
1478 5 : ReplSlotSyncWorkerMain(const void *startup_data, size_t startup_data_len)
1479 : {
1480 5 : WalReceiverConn *wrconn = NULL;
1481 : char *dbname;
1482 : char *err;
1483 : sigjmp_buf local_sigjmp_buf;
1484 : StringInfoData app_name;
1485 :
1486 : Assert(startup_data_len == 0);
1487 :
1488 5 : init_ps_display(NULL);
1489 :
1490 : Assert(GetProcessingMode() == InitProcessing);
1491 :
1492 : /*
1493 : * Create a per-backend PGPROC struct in shared memory. We must do this
1494 : * before we access any shared memory.
1495 : */
1496 5 : InitProcess();
1497 :
1498 : /*
1499 : * Early initialization.
1500 : */
1501 5 : BaseInit();
1502 :
1503 : Assert(SlotSyncCtx != NULL);
1504 :
1505 : /*
1506 : * If an exception is encountered, processing resumes here.
1507 : *
1508 : * We just need to clean up, report the error, and go away.
1509 : *
1510 : * If we do not have this handling here, then since this worker process
1511 : * operates at the bottom of the exception stack, ERRORs turn into FATALs.
1512 : * Therefore, we create our own exception handler to catch ERRORs.
1513 : */
1514 5 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1515 : {
1516 : /* since not using PG_TRY, must reset error stack by hand */
1517 1 : error_context_stack = NULL;
1518 :
1519 : /* Prevents interrupts while cleaning up */
1520 1 : HOLD_INTERRUPTS();
1521 :
1522 : /* Report the error to the server log */
1523 1 : EmitErrorReport();
1524 :
1525 : /*
1526 : * We can now go away. Note that because we called InitProcess, a
1527 : * callback was registered to do ProcKill, which will clean up
1528 : * necessary state.
1529 : */
1530 1 : proc_exit(0);
1531 : }
1532 :
1533 : /* We can now handle ereport(ERROR) */
1534 5 : PG_exception_stack = &local_sigjmp_buf;
1535 :
1536 : /* Setup signal handling */
1537 5 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1538 5 : pqsignal(SIGINT, StatementCancelHandler);
1539 5 : pqsignal(SIGTERM, die);
1540 5 : pqsignal(SIGFPE, FloatExceptionHandler);
1541 5 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1542 5 : pqsignal(SIGUSR2, SIG_IGN);
1543 5 : pqsignal(SIGPIPE, SIG_IGN);
1544 5 : pqsignal(SIGCHLD, SIG_DFL);
1545 :
1546 5 : check_and_set_sync_info(MyProcPid);
1547 :
1548 5 : ereport(LOG, errmsg("slot sync worker started"));
1549 :
1550 : /* Register it as soon as SlotSyncCtx->pid is initialized. */
1551 5 : before_shmem_exit(slotsync_worker_onexit, (Datum) 0);
1552 :
1553 : /*
1554 : * Establishes SIGALRM handler and initialize timeout module. It is needed
1555 : * by InitPostgres to register different timeouts.
1556 : */
1557 5 : InitializeTimeouts();
1558 :
1559 : /* Load the libpq-specific functions */
1560 5 : load_file("libpqwalreceiver", false);
1561 :
1562 : /*
1563 : * Unblock signals (they were blocked when the postmaster forked us)
1564 : */
1565 5 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
1566 :
1567 : /*
1568 : * Set always-secure search path, so malicious users can't redirect user
1569 : * code (e.g. operators).
1570 : *
1571 : * It's not strictly necessary since we won't be scanning or writing to
1572 : * any user table locally, but it's good to retain it here for added
1573 : * precaution.
1574 : */
1575 5 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1576 :
1577 5 : dbname = CheckAndGetDbnameFromConninfo();
1578 :
1579 : /*
1580 : * Connect to the database specified by the user in primary_conninfo. We
1581 : * need a database connection for walrcv_exec to work which we use to
1582 : * fetch slot information from the remote node. See comments atop
1583 : * libpqrcv_exec.
1584 : *
1585 : * We do not specify a specific user here since the slot sync worker will
1586 : * operate as a superuser. This is safe because the slot sync worker does
1587 : * not interact with user tables, eliminating the risk of executing
1588 : * arbitrary code within triggers.
1589 : */
1590 5 : InitPostgres(dbname, InvalidOid, NULL, InvalidOid, 0, NULL);
1591 :
1592 5 : SetProcessingMode(NormalProcessing);
1593 :
1594 5 : initStringInfo(&app_name);
1595 5 : if (cluster_name[0])
1596 5 : appendStringInfo(&app_name, "%s_%s", cluster_name, "slotsync worker");
1597 : else
1598 0 : appendStringInfoString(&app_name, "slotsync worker");
1599 :
1600 : /*
1601 : * Establish the connection to the primary server for slot
1602 : * synchronization.
1603 : */
1604 5 : wrconn = walrcv_connect(PrimaryConnInfo, false, false, false,
1605 : app_name.data, &err);
1606 :
1607 5 : if (!wrconn)
1608 0 : ereport(ERROR,
1609 : errcode(ERRCODE_CONNECTION_FAILURE),
1610 : errmsg("synchronization worker \"%s\" could not connect to the primary server: %s",
1611 : app_name.data, err));
1612 :
1613 5 : pfree(app_name.data);
1614 :
1615 : /*
1616 : * Register the disconnection callback.
1617 : *
1618 : * XXX: This can be combined with previous cleanup registration of
1619 : * slotsync_worker_onexit() but that will need the connection to be made
1620 : * global and we want to avoid introducing global for this purpose.
1621 : */
1622 5 : before_shmem_exit(slotsync_worker_disconnect, PointerGetDatum(wrconn));
1623 :
1624 : /*
1625 : * Using the specified primary server connection, check that we are not a
1626 : * cascading standby and slot configured in 'primary_slot_name' exists on
1627 : * the primary server.
1628 : */
1629 5 : validate_remote_info(wrconn);
1630 :
1631 : /* Main loop to synchronize slots */
1632 : for (;;)
1633 19 : {
1634 24 : bool some_slot_updated = false;
1635 24 : bool started_tx = false;
1636 : List *remote_slots;
1637 :
1638 24 : ProcessSlotSyncInterrupts();
1639 :
1640 : /*
1641 : * The syscache access in fetch_remote_slots() needs a transaction
1642 : * env.
1643 : */
1644 20 : if (!IsTransactionState())
1645 : {
1646 20 : StartTransactionCommand();
1647 20 : started_tx = true;
1648 : }
1649 :
1650 20 : remote_slots = fetch_remote_slots(wrconn, NIL);
1651 19 : some_slot_updated = synchronize_slots(wrconn, remote_slots, NULL);
1652 19 : list_free_deep(remote_slots);
1653 :
1654 19 : if (started_tx)
1655 19 : CommitTransactionCommand();
1656 :
1657 19 : wait_for_slot_activity(some_slot_updated);
1658 : }
1659 :
1660 : /*
1661 : * The slot sync worker can't get here because it will only stop when it
1662 : * receives a stop request from the startup process, or when there is an
1663 : * error.
1664 : */
1665 : Assert(false);
1666 : }
1667 :
1668 : /*
1669 : * Update the inactive_since property for synced slots.
1670 : *
1671 : * Note that this function is currently called when we shutdown the slot
1672 : * sync machinery.
1673 : */
1674 : static void
1675 949 : update_synced_slots_inactive_since(void)
1676 : {
1677 949 : TimestampTz now = 0;
1678 :
1679 : /*
1680 : * We need to update inactive_since only when we are promoting standby to
1681 : * correctly interpret the inactive_since if the standby gets promoted
1682 : * without a restart. We don't want the slots to appear inactive for a
1683 : * long time after promotion if they haven't been synchronized recently.
1684 : * Whoever acquires the slot, i.e., makes the slot active, will reset it.
1685 : */
1686 949 : if (!StandbyMode)
1687 897 : return;
1688 :
1689 : /* The slot sync worker or the SQL function mustn't be running by now */
1690 : Assert((SlotSyncCtx->pid == InvalidPid) && !SlotSyncCtx->syncing);
1691 :
1692 52 : LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
1693 :
1694 560 : for (int i = 0; i < max_replication_slots; i++)
1695 : {
1696 508 : ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
1697 :
1698 : /* Check if it is a synchronized slot */
1699 508 : if (s->in_use && s->data.synced)
1700 : {
1701 : Assert(SlotIsLogical(s));
1702 :
1703 : /* The slot must not be acquired by any process */
1704 : Assert(s->active_proc == INVALID_PROC_NUMBER);
1705 :
1706 : /* Use the same inactive_since time for all the slots. */
1707 3 : if (now == 0)
1708 2 : now = GetCurrentTimestamp();
1709 :
1710 3 : ReplicationSlotSetInactiveSince(s, now, true);
1711 : }
1712 : }
1713 :
1714 52 : LWLockRelease(ReplicationSlotControlLock);
1715 : }
1716 :
1717 : /*
1718 : * Shut down slot synchronization.
1719 : *
1720 : * This function sets stopSignaled=true and wakes up the slot sync process
1721 : * (either worker or backend running the SQL function pg_sync_replication_slots())
1722 : * so that worker can exit or the SQL function pg_sync_replication_slots() can
1723 : * finish. It also waits till the slot sync worker has exited or
1724 : * pg_sync_replication_slots() has finished.
1725 : */
1726 : void
1727 949 : ShutDownSlotSync(void)
1728 : {
1729 : pid_t sync_process_pid;
1730 :
1731 949 : SpinLockAcquire(&SlotSyncCtx->mutex);
1732 :
1733 949 : SlotSyncCtx->stopSignaled = true;
1734 :
1735 : /*
1736 : * Return if neither the slot sync worker is running nor the function
1737 : * pg_sync_replication_slots() is executing.
1738 : */
1739 949 : if (!SlotSyncCtx->syncing)
1740 : {
1741 948 : SpinLockRelease(&SlotSyncCtx->mutex);
1742 948 : update_synced_slots_inactive_since();
1743 948 : return;
1744 : }
1745 :
1746 1 : sync_process_pid = SlotSyncCtx->pid;
1747 :
1748 1 : SpinLockRelease(&SlotSyncCtx->mutex);
1749 :
1750 : /*
1751 : * Signal process doing slotsync, if any. The process will stop upon
1752 : * detecting that the stopSignaled flag is set to true.
1753 : */
1754 1 : if (sync_process_pid != InvalidPid)
1755 1 : kill(sync_process_pid, SIGUSR1);
1756 :
1757 : /* Wait for slot sync to end */
1758 : for (;;)
1759 0 : {
1760 : int rc;
1761 :
1762 : /* Wait a bit, we don't expect to have to wait long */
1763 1 : rc = WaitLatch(MyLatch,
1764 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
1765 : 10L, WAIT_EVENT_REPLICATION_SLOTSYNC_SHUTDOWN);
1766 :
1767 1 : if (rc & WL_LATCH_SET)
1768 : {
1769 0 : ResetLatch(MyLatch);
1770 0 : CHECK_FOR_INTERRUPTS();
1771 : }
1772 :
1773 1 : SpinLockAcquire(&SlotSyncCtx->mutex);
1774 :
1775 : /* Ensure that no process is syncing the slots. */
1776 1 : if (!SlotSyncCtx->syncing)
1777 1 : break;
1778 :
1779 0 : SpinLockRelease(&SlotSyncCtx->mutex);
1780 : }
1781 :
1782 1 : SpinLockRelease(&SlotSyncCtx->mutex);
1783 :
1784 1 : update_synced_slots_inactive_since();
1785 : }
1786 :
1787 : /*
1788 : * SlotSyncWorkerCanRestart
1789 : *
1790 : * Return true, indicating worker is allowed to restart, if enough time has
1791 : * passed since it was last launched to reach SLOTSYNC_RESTART_INTERVAL_SEC.
1792 : * Otherwise return false.
1793 : *
1794 : * This is a safety valve to protect against continuous respawn attempts if the
1795 : * worker is dying immediately at launch. Note that since we will retry to
1796 : * launch the worker from the postmaster main loop, we will get another
1797 : * chance later.
1798 : */
1799 : bool
1800 12 : SlotSyncWorkerCanRestart(void)
1801 : {
1802 12 : time_t curtime = time(NULL);
1803 :
1804 : /*
1805 : * If first time through, or time somehow went backwards, always update
1806 : * last_start_time to match the current clock and allow worker start.
1807 : * Otherwise allow it only once enough time has elapsed.
1808 : */
1809 12 : if (SlotSyncCtx->last_start_time == 0 ||
1810 7 : curtime < SlotSyncCtx->last_start_time ||
1811 7 : curtime - SlotSyncCtx->last_start_time >= SLOTSYNC_RESTART_INTERVAL_SEC)
1812 : {
1813 5 : SlotSyncCtx->last_start_time = curtime;
1814 5 : return true;
1815 : }
1816 7 : return false;
1817 : }
1818 :
1819 : /*
1820 : * Is current process syncing replication slots?
1821 : *
1822 : * Could be either backend executing SQL function or slot sync worker.
1823 : */
1824 : bool
1825 26 : IsSyncingReplicationSlots(void)
1826 : {
1827 26 : return syncing_slots;
1828 : }
1829 :
1830 : /*
1831 : * Amount of shared memory required for slot synchronization.
1832 : */
1833 : Size
1834 3324 : SlotSyncShmemSize(void)
1835 : {
1836 3324 : return sizeof(SlotSyncCtxStruct);
1837 : }
1838 :
1839 : /*
1840 : * Allocate and initialize the shared memory of slot synchronization.
1841 : */
1842 : void
1843 1159 : SlotSyncShmemInit(void)
1844 : {
1845 1159 : Size size = SlotSyncShmemSize();
1846 : bool found;
1847 :
1848 1159 : SlotSyncCtx = (SlotSyncCtxStruct *)
1849 1159 : ShmemInitStruct("Slot Sync Data", size, &found);
1850 :
1851 1159 : if (!found)
1852 : {
1853 1159 : memset(SlotSyncCtx, 0, size);
1854 1159 : SlotSyncCtx->pid = InvalidPid;
1855 1159 : SpinLockInit(&SlotSyncCtx->mutex);
1856 : }
1857 1159 : }
1858 :
1859 : /*
1860 : * Error cleanup callback for slot sync SQL function.
1861 : */
1862 : static void
1863 1 : slotsync_failure_callback(int code, Datum arg)
1864 : {
1865 1 : WalReceiverConn *wrconn = (WalReceiverConn *) DatumGetPointer(arg);
1866 :
1867 : /*
1868 : * We need to do slots cleanup here just like WalSndErrorCleanup() does.
1869 : *
1870 : * The startup process during promotion invokes ShutDownSlotSync() which
1871 : * waits for slot sync to finish and it does that by checking the
1872 : * 'syncing' flag. Thus the SQL function must be done with slots' release
1873 : * and cleanup to avoid any dangling temporary slots or active slots
1874 : * before it marks itself as finished syncing.
1875 : */
1876 :
1877 : /* Make sure active replication slots are released */
1878 1 : if (MyReplicationSlot != NULL)
1879 0 : ReplicationSlotRelease();
1880 :
1881 : /* Also cleanup the synced temporary slots. */
1882 1 : ReplicationSlotCleanup(true);
1883 :
1884 : /*
1885 : * The set syncing_slots indicates that the process errored out without
1886 : * resetting the flag. So, we need to clean up shared memory and reset the
1887 : * flag here.
1888 : */
1889 1 : if (syncing_slots)
1890 1 : reset_syncing_flag();
1891 :
1892 1 : walrcv_disconnect(wrconn);
1893 1 : }
1894 :
1895 : /*
1896 : * Helper function to extract slot names from a list of remote slots
1897 : */
1898 : static List *
1899 1 : extract_slot_names(List *remote_slots)
1900 : {
1901 1 : List *slot_names = NIL;
1902 :
1903 3 : foreach_ptr(RemoteSlot, remote_slot, remote_slots)
1904 : {
1905 : char *slot_name;
1906 :
1907 1 : slot_name = pstrdup(remote_slot->name);
1908 1 : slot_names = lappend(slot_names, slot_name);
1909 : }
1910 :
1911 1 : return slot_names;
1912 : }
1913 :
1914 : /*
1915 : * Synchronize the failover enabled replication slots using the specified
1916 : * primary server connection.
1917 : *
1918 : * Repeatedly fetches and updates replication slot information from the
1919 : * primary until all slots are at least "sync ready".
1920 : *
1921 : * Exits early if promotion is triggered or certain critical
1922 : * configuration parameters have changed.
1923 : */
1924 : void
1925 9 : SyncReplicationSlots(WalReceiverConn *wrconn)
1926 : {
1927 9 : PG_ENSURE_ERROR_CLEANUP(slotsync_failure_callback, PointerGetDatum(wrconn));
1928 : {
1929 9 : List *remote_slots = NIL;
1930 9 : List *slot_names = NIL; /* List of slot names to track */
1931 :
1932 9 : check_and_set_sync_info(MyProcPid);
1933 :
1934 : /* Check for interrupts and config changes */
1935 9 : ProcessSlotSyncInterrupts();
1936 :
1937 9 : validate_remote_info(wrconn);
1938 :
1939 : /* Retry until all the slots are sync-ready */
1940 : for (;;)
1941 2 : {
1942 10 : bool slot_persistence_pending = false;
1943 10 : bool some_slot_updated = false;
1944 :
1945 : /* Check for interrupts and config changes */
1946 10 : ProcessSlotSyncInterrupts();
1947 :
1948 : /* We must be in a valid transaction state */
1949 : Assert(IsTransactionState());
1950 :
1951 : /*
1952 : * Fetch remote slot info for the given slot_names. If slot_names
1953 : * is NIL, fetch all failover-enabled slots. Note that we reuse
1954 : * slot_names from the first iteration; re-fetching all failover
1955 : * slots each time could cause an endless loop. Instead of
1956 : * reprocessing only the pending slots in each iteration, it's
1957 : * better to process all the slots received in the first
1958 : * iteration. This ensures that by the time we're done, all slots
1959 : * reflect the latest values.
1960 : */
1961 10 : remote_slots = fetch_remote_slots(wrconn, slot_names);
1962 :
1963 : /* Attempt to synchronize slots */
1964 10 : some_slot_updated = synchronize_slots(wrconn, remote_slots,
1965 : &slot_persistence_pending);
1966 :
1967 : /*
1968 : * If slot_persistence_pending is true, extract slot names for
1969 : * future iterations (only needed if we haven't done it yet)
1970 : */
1971 10 : if (slot_names == NIL && slot_persistence_pending)
1972 1 : slot_names = extract_slot_names(remote_slots);
1973 :
1974 : /* Free the current remote_slots list */
1975 10 : list_free_deep(remote_slots);
1976 :
1977 : /* Done if all slots are persisted i.e are sync-ready */
1978 10 : if (!slot_persistence_pending)
1979 8 : break;
1980 :
1981 : /* wait before retrying again */
1982 2 : wait_for_slot_activity(some_slot_updated);
1983 : }
1984 :
1985 8 : if (slot_names)
1986 1 : list_free_deep(slot_names);
1987 :
1988 : /* Cleanup the synced temporary slots */
1989 8 : ReplicationSlotCleanup(true);
1990 :
1991 : /* We are done with sync, so reset sync flag */
1992 8 : reset_syncing_flag();
1993 : }
1994 9 : PG_END_ENSURE_ERROR_CLEANUP(slotsync_failure_callback, PointerGetDatum(wrconn));
1995 8 : }
|
