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 "storage/subsystems.h"
77 : #include "tcop/tcopprot.h"
78 : #include "utils/builtins.h"
79 : #include "utils/memutils.h"
80 : #include "utils/pg_lsn.h"
81 : #include "utils/ps_status.h"
82 : #include "utils/timeout.h"
83 : #include "utils/wait_event.h"
84 :
85 : /*
86 : * Struct for sharing information to control slot synchronization.
87 : *
88 : * The 'pid' is either the slot sync worker's pid or the backend's pid running
89 : * the SQL function pg_sync_replication_slots(). On promotion, the startup
90 : * process sets 'stopSignaled' and uses this 'pid' to signal the synchronizing
91 : * process with PROCSIG_SLOTSYNC_MESSAGE and also to wake it up so that the
92 : * process can immediately stop its synchronizing work.
93 : * Setting 'stopSignaled' on the other hand is used to handle the race
94 : * condition when the postmaster has not noticed the promotion yet and thus may
95 : * end up restarting the slot sync worker. If 'stopSignaled' is set, the worker
96 : * will exit in such a case. The SQL function pg_sync_replication_slots() will
97 : * also error out if this flag is set. Note that we don't need to reset this
98 : * variable as after promotion the slot sync worker won't be restarted because
99 : * the pmState changes to PM_RUN from PM_HOT_STANDBY and we don't support
100 : * demoting primary without restarting the server.
101 : * See LaunchMissingBackgroundProcesses.
102 : *
103 : * The 'syncing' flag is needed to prevent concurrent slot syncs to avoid slot
104 : * overwrites.
105 : *
106 : * The 'last_start_time' is needed by postmaster to start the slot sync worker
107 : * once per SLOTSYNC_RESTART_INTERVAL_SEC. In cases where an immediate restart
108 : * is expected (e.g., slot sync GUCs change), slot sync worker will reset
109 : * last_start_time before exiting, so that postmaster can start the worker
110 : * without waiting for SLOTSYNC_RESTART_INTERVAL_SEC.
111 : */
112 : typedef struct SlotSyncCtxStruct
113 : {
114 : pid_t pid;
115 : bool stopSignaled;
116 : bool syncing;
117 : time_t last_start_time;
118 : slock_t mutex;
119 : } SlotSyncCtxStruct;
120 :
121 : static SlotSyncCtxStruct *SlotSyncCtx = NULL;
122 :
123 : static void SlotSyncShmemRequest(void *arg);
124 : static void SlotSyncShmemInit(void *arg);
125 :
126 : const ShmemCallbacks SlotSyncShmemCallbacks = {
127 : .request_fn = SlotSyncShmemRequest,
128 : .init_fn = SlotSyncShmemInit,
129 : };
130 :
131 : /* GUC variable */
132 : bool sync_replication_slots = false;
133 :
134 : /*
135 : * The sleep time (ms) between slot-sync cycles varies dynamically
136 : * (within a MIN/MAX range) according to slot activity. See
137 : * wait_for_slot_activity() for details.
138 : */
139 : #define MIN_SLOTSYNC_WORKER_NAPTIME_MS 200
140 : #define MAX_SLOTSYNC_WORKER_NAPTIME_MS 30000 /* 30s */
141 :
142 : static long sleep_ms = MIN_SLOTSYNC_WORKER_NAPTIME_MS;
143 :
144 : /* The restart interval for slot sync work used by postmaster */
145 : #define SLOTSYNC_RESTART_INTERVAL_SEC 10
146 :
147 : /*
148 : * Flag to tell if we are syncing replication slots. Unlike the 'syncing' flag
149 : * in SlotSyncCtxStruct, this flag is true only if the current process is
150 : * performing slot synchronization.
151 : */
152 : static bool syncing_slots = false;
153 :
154 : /*
155 : * Interrupt flag set when PROCSIG_SLOTSYNC_MESSAGE is received, asking the
156 : * slotsync worker or pg_sync_replication_slots() to stop because
157 : * standby promotion has been triggered.
158 : */
159 : volatile sig_atomic_t SlotSyncShutdownPending = false;
160 :
161 : /*
162 : * Structure to hold information fetched from the primary server about a logical
163 : * replication slot.
164 : */
165 : typedef struct RemoteSlot
166 : {
167 : char *name;
168 : char *plugin;
169 : char *database;
170 : bool two_phase;
171 : bool failover;
172 : XLogRecPtr restart_lsn;
173 : XLogRecPtr confirmed_lsn;
174 : XLogRecPtr two_phase_at;
175 : TransactionId catalog_xmin;
176 :
177 : /* RS_INVAL_NONE if valid, or the reason of invalidation */
178 : ReplicationSlotInvalidationCause invalidated;
179 : } RemoteSlot;
180 :
181 : static void slotsync_failure_callback(int code, Datum arg);
182 : static void update_synced_slots_inactive_since(void);
183 :
184 : /*
185 : * Update slot sync skip stats. This function requires the caller to acquire
186 : * the slot.
187 : */
188 : static void
189 63 : update_slotsync_skip_stats(SlotSyncSkipReason skip_reason)
190 : {
191 : ReplicationSlot *slot;
192 :
193 : Assert(MyReplicationSlot);
194 :
195 63 : slot = MyReplicationSlot;
196 :
197 : /*
198 : * Update the slot sync related stats in pg_stat_replication_slots when a
199 : * slot sync is skipped
200 : */
201 63 : if (skip_reason != SS_SKIP_NONE)
202 8 : pgstat_report_replslotsync(slot);
203 :
204 : /* Update the slot sync skip reason */
205 63 : if (slot->slotsync_skip_reason != skip_reason)
206 : {
207 5 : SpinLockAcquire(&slot->mutex);
208 5 : slot->slotsync_skip_reason = skip_reason;
209 5 : SpinLockRelease(&slot->mutex);
210 : }
211 63 : }
212 :
213 : /*
214 : * If necessary, update the local synced slot's metadata based on the data
215 : * from the remote slot.
216 : *
217 : * If no update was needed (the data of the remote slot is the same as the
218 : * local slot) return false, otherwise true.
219 : */
220 : static bool
221 63 : update_local_synced_slot(RemoteSlot *remote_slot, Oid remote_dbid)
222 : {
223 63 : ReplicationSlot *slot = MyReplicationSlot;
224 63 : bool updated_xmin_or_lsn = false;
225 63 : bool updated_config = false;
226 63 : SlotSyncSkipReason skip_reason = SS_SKIP_NONE;
227 63 : XLogRecPtr latestFlushPtr = GetStandbyFlushRecPtr(NULL);
228 :
229 : Assert(slot->data.invalidated == RS_INVAL_NONE);
230 :
231 : /*
232 : * Make sure that concerned WAL is received and flushed before syncing
233 : * slot to target lsn received from the primary server.
234 : */
235 63 : if (remote_slot->confirmed_lsn > latestFlushPtr)
236 : {
237 0 : update_slotsync_skip_stats(SS_SKIP_WAL_NOT_FLUSHED);
238 :
239 : /*
240 : * Can get here only if GUC 'synchronized_standby_slots' on the
241 : * primary server was not configured correctly.
242 : */
243 0 : ereport(LOG,
244 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
245 : errmsg("skipping slot synchronization because the received slot sync"
246 : " LSN %X/%08X for slot \"%s\" is ahead of the standby position %X/%08X",
247 : LSN_FORMAT_ARGS(remote_slot->confirmed_lsn),
248 : remote_slot->name,
249 : LSN_FORMAT_ARGS(latestFlushPtr)));
250 :
251 0 : return false;
252 : }
253 :
254 : /*
255 : * Don't overwrite if we already have a newer catalog_xmin and
256 : * restart_lsn.
257 : */
258 120 : if (remote_slot->restart_lsn < slot->data.restart_lsn ||
259 57 : TransactionIdPrecedes(remote_slot->catalog_xmin,
260 : slot->data.catalog_xmin))
261 : {
262 : /* Update slot sync skip stats */
263 8 : update_slotsync_skip_stats(SS_SKIP_WAL_OR_ROWS_REMOVED);
264 :
265 : /*
266 : * This can happen in following situations:
267 : *
268 : * If the slot is temporary, it means either the initial WAL location
269 : * reserved for the local slot is ahead of the remote slot's
270 : * restart_lsn or the initial xmin_horizon computed for the local slot
271 : * is ahead of the remote slot.
272 : *
273 : * If the slot is persistent, both restart_lsn and catalog_xmin of the
274 : * synced slot could still be ahead of the remote slot. Since we use
275 : * slot advance functionality to keep snapbuild/slot updated, it is
276 : * possible that the restart_lsn and catalog_xmin are advanced to a
277 : * later position than it has on the primary. This can happen when
278 : * slot advancing machinery finds running xacts record after reaching
279 : * the consistent state at a later point than the primary where it
280 : * serializes the snapshot and updates the restart_lsn.
281 : *
282 : * We LOG the message if the slot is temporary as it can help the user
283 : * to understand why the slot is not sync-ready. In the case of a
284 : * persistent slot, it would be a more common case and won't directly
285 : * impact the users, so we used DEBUG1 level to log the message.
286 : */
287 8 : ereport(slot->data.persistency == RS_TEMPORARY ? LOG : DEBUG1,
288 : errmsg("could not synchronize replication slot \"%s\"",
289 : remote_slot->name),
290 : 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.",
291 : LSN_FORMAT_ARGS(remote_slot->restart_lsn),
292 : remote_slot->catalog_xmin,
293 : LSN_FORMAT_ARGS(slot->data.restart_lsn),
294 : slot->data.catalog_xmin));
295 :
296 : /*
297 : * Skip updating the configuration. This is required to avoid syncing
298 : * two_phase_at without syncing confirmed_lsn. Otherwise, the prepared
299 : * transaction between old confirmed_lsn and two_phase_at will
300 : * unexpectedly get decoded and sent to the downstream after
301 : * promotion. See comments in ReorderBufferFinishPrepared.
302 : */
303 8 : return false;
304 : }
305 :
306 : /*
307 : * Attempt to sync LSNs and xmins only if remote slot is ahead of local
308 : * slot.
309 : */
310 55 : if (remote_slot->confirmed_lsn > slot->data.confirmed_flush ||
311 81 : remote_slot->restart_lsn > slot->data.restart_lsn ||
312 40 : TransactionIdFollows(remote_slot->catalog_xmin,
313 : slot->data.catalog_xmin))
314 : {
315 : /*
316 : * We can't directly copy the remote slot's LSN or xmin unless there
317 : * exists a consistent snapshot at that point. Otherwise, after
318 : * promotion, the slots may not reach a consistent point before the
319 : * confirmed_flush_lsn which can lead to a data loss. To avoid data
320 : * loss, we let slot machinery advance the slot which ensures that
321 : * snapbuilder/slot statuses are updated properly.
322 : */
323 15 : if (SnapBuildSnapshotExists(remote_slot->restart_lsn))
324 : {
325 : /*
326 : * Update the slot info directly if there is a serialized snapshot
327 : * at the restart_lsn, as the slot can quickly reach consistency
328 : * at restart_lsn by restoring the snapshot.
329 : */
330 4 : SpinLockAcquire(&slot->mutex);
331 4 : slot->data.restart_lsn = remote_slot->restart_lsn;
332 4 : slot->data.confirmed_flush = remote_slot->confirmed_lsn;
333 4 : slot->data.catalog_xmin = remote_slot->catalog_xmin;
334 4 : SpinLockRelease(&slot->mutex);
335 :
336 4 : updated_xmin_or_lsn = true;
337 : }
338 : else
339 : {
340 : bool found_consistent_snapshot;
341 11 : XLogRecPtr old_confirmed_lsn = slot->data.confirmed_flush;
342 11 : XLogRecPtr old_restart_lsn = slot->data.restart_lsn;
343 11 : XLogRecPtr old_catalog_xmin = slot->data.catalog_xmin;
344 :
345 11 : LogicalSlotAdvanceAndCheckSnapState(remote_slot->confirmed_lsn,
346 : &found_consistent_snapshot);
347 :
348 : /* Sanity check */
349 11 : if (slot->data.confirmed_flush != remote_slot->confirmed_lsn)
350 0 : ereport(ERROR,
351 : errmsg_internal("synchronized confirmed_flush for slot \"%s\" differs from remote slot",
352 : remote_slot->name),
353 : errdetail_internal("Remote slot has LSN %X/%08X but local slot has LSN %X/%08X.",
354 : LSN_FORMAT_ARGS(remote_slot->confirmed_lsn),
355 : LSN_FORMAT_ARGS(slot->data.confirmed_flush)));
356 :
357 : /*
358 : * If we can't reach a consistent snapshot, the slot won't be
359 : * persisted. See update_and_persist_local_synced_slot().
360 : */
361 11 : if (!found_consistent_snapshot)
362 : {
363 : Assert(MyReplicationSlot->data.persistency == RS_TEMPORARY);
364 :
365 0 : ereport(LOG,
366 : errmsg("could not synchronize replication slot \"%s\"",
367 : remote_slot->name),
368 : errdetail("Synchronization could lead to data loss, because the standby could not build a consistent snapshot to decode WALs at LSN %X/%08X.",
369 : LSN_FORMAT_ARGS(slot->data.restart_lsn)));
370 :
371 0 : skip_reason = SS_SKIP_NO_CONSISTENT_SNAPSHOT;
372 : }
373 :
374 : /*
375 : * It is possible that the slot's xmin or LSNs are not updated,
376 : * when the synced slot has reached consistent snapshot state or
377 : * cannot build one at all.
378 : */
379 11 : updated_xmin_or_lsn = (old_confirmed_lsn != slot->data.confirmed_flush ||
380 11 : old_restart_lsn != slot->data.restart_lsn ||
381 0 : old_catalog_xmin != slot->data.catalog_xmin);
382 : }
383 : }
384 :
385 : /* Update slot sync skip stats */
386 55 : update_slotsync_skip_stats(skip_reason);
387 :
388 55 : if (remote_dbid != slot->data.database ||
389 55 : remote_slot->two_phase != slot->data.two_phase ||
390 54 : remote_slot->failover != slot->data.failover ||
391 54 : strcmp(remote_slot->plugin, NameStr(slot->data.plugin)) != 0 ||
392 54 : remote_slot->two_phase_at != slot->data.two_phase_at)
393 : {
394 : NameData plugin_name;
395 :
396 : /* Avoid expensive operations while holding a spinlock. */
397 1 : namestrcpy(&plugin_name, remote_slot->plugin);
398 :
399 1 : SpinLockAcquire(&slot->mutex);
400 1 : slot->data.plugin = plugin_name;
401 1 : slot->data.database = remote_dbid;
402 1 : slot->data.two_phase = remote_slot->two_phase;
403 1 : slot->data.two_phase_at = remote_slot->two_phase_at;
404 1 : slot->data.failover = remote_slot->failover;
405 1 : SpinLockRelease(&slot->mutex);
406 :
407 1 : updated_config = true;
408 :
409 : /*
410 : * Ensure that there is no risk of sending prepared transactions
411 : * unexpectedly after the promotion.
412 : */
413 : Assert(slot->data.two_phase_at <= slot->data.confirmed_flush);
414 : }
415 :
416 : /*
417 : * We have to write the changed xmin to disk *before* we change the
418 : * in-memory value, otherwise after a crash we wouldn't know that some
419 : * catalog tuples might have been removed already.
420 : */
421 55 : if (updated_config || updated_xmin_or_lsn)
422 : {
423 16 : ReplicationSlotMarkDirty();
424 16 : ReplicationSlotSave();
425 : }
426 :
427 : /*
428 : * Now the new xmin is safely on disk, we can let the global value
429 : * advance. We do not take ProcArrayLock or similar since we only advance
430 : * xmin here and there's not much harm done by a concurrent computation
431 : * missing that.
432 : */
433 55 : if (updated_xmin_or_lsn)
434 : {
435 15 : SpinLockAcquire(&slot->mutex);
436 15 : slot->effective_catalog_xmin = remote_slot->catalog_xmin;
437 15 : SpinLockRelease(&slot->mutex);
438 :
439 15 : ReplicationSlotsComputeRequiredXmin(false);
440 15 : ReplicationSlotsComputeRequiredLSN();
441 : }
442 :
443 55 : return updated_config || updated_xmin_or_lsn;
444 : }
445 :
446 : /*
447 : * Get the list of local logical slots that are synchronized from the
448 : * primary server.
449 : */
450 : static List *
451 38 : get_local_synced_slots(void)
452 : {
453 38 : List *local_slots = NIL;
454 :
455 38 : LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
456 :
457 608 : for (int i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
458 : {
459 570 : ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
460 :
461 : /* Check if it is a synchronized slot */
462 570 : if (s->in_use && s->data.synced)
463 : {
464 : Assert(SlotIsLogical(s));
465 57 : local_slots = lappend(local_slots, s);
466 : }
467 : }
468 :
469 38 : LWLockRelease(ReplicationSlotControlLock);
470 :
471 38 : return local_slots;
472 : }
473 :
474 : /*
475 : * Helper function to check if local_slot is required to be retained.
476 : *
477 : * Return false either if local_slot does not exist in the remote_slots list
478 : * or is invalidated while the corresponding remote slot is still valid,
479 : * otherwise true.
480 : */
481 : static bool
482 57 : local_sync_slot_required(ReplicationSlot *local_slot, List *remote_slots)
483 : {
484 57 : bool remote_exists = false;
485 57 : bool locally_invalidated = false;
486 :
487 138 : foreach_ptr(RemoteSlot, remote_slot, remote_slots)
488 : {
489 80 : if (strcmp(remote_slot->name, NameStr(local_slot->data.name)) == 0)
490 : {
491 56 : remote_exists = true;
492 :
493 : /*
494 : * If remote slot is not invalidated but local slot is marked as
495 : * invalidated, then set locally_invalidated flag.
496 : */
497 56 : SpinLockAcquire(&local_slot->mutex);
498 56 : locally_invalidated =
499 112 : (remote_slot->invalidated == RS_INVAL_NONE) &&
500 56 : (local_slot->data.invalidated != RS_INVAL_NONE);
501 56 : SpinLockRelease(&local_slot->mutex);
502 :
503 56 : break;
504 : }
505 : }
506 :
507 57 : return (remote_exists && !locally_invalidated);
508 : }
509 :
510 : /*
511 : * Drop local obsolete slots.
512 : *
513 : * Drop the local slots that no longer need to be synced i.e. these either do
514 : * not exist on the primary or are no longer enabled for failover.
515 : *
516 : * Additionally, drop any slots that are valid on the primary but got
517 : * invalidated on the standby. This situation may occur due to the following
518 : * reasons:
519 : * - The 'max_slot_wal_keep_size' on the standby is insufficient to retain WAL
520 : * records from the restart_lsn of the slot.
521 : * - 'primary_slot_name' is temporarily reset to null and the physical slot is
522 : * removed.
523 : * These dropped slots will get recreated in next sync-cycle and it is okay to
524 : * drop and recreate such slots as long as these are not consumable on the
525 : * standby (which is the case currently).
526 : *
527 : * Note: Change of 'wal_level' on the primary server to a level lower than
528 : * logical may also result in slot invalidation and removal on the standby.
529 : * This is because such 'wal_level' change is only possible if the logical
530 : * slots are removed on the primary server, so it's expected to see the
531 : * slots being invalidated and removed on the standby too (and re-created
532 : * if they are re-created on the primary server).
533 : */
534 : static void
535 38 : drop_local_obsolete_slots(List *remote_slot_list)
536 : {
537 38 : List *local_slots = get_local_synced_slots();
538 :
539 133 : foreach_ptr(ReplicationSlot, local_slot, local_slots)
540 : {
541 : /* Drop the local slot if it is not required to be retained. */
542 57 : if (!local_sync_slot_required(local_slot, remote_slot_list))
543 : {
544 : bool synced_slot;
545 :
546 : /*
547 : * Use shared lock to prevent a conflict with
548 : * ReplicationSlotsDropDBSlots(), trying to drop the same slot
549 : * during a drop-database operation.
550 : */
551 2 : LockSharedObject(DatabaseRelationId, local_slot->data.database,
552 : 0, AccessShareLock);
553 :
554 : /*
555 : * In the small window between getting the slot to drop and
556 : * locking the database, there is a possibility of a parallel
557 : * database drop by the startup process and the creation of a new
558 : * slot by the user. This new user-created slot may end up using
559 : * the same shared memory as that of 'local_slot'. Thus check if
560 : * local_slot is still the synced one before performing actual
561 : * drop.
562 : */
563 2 : SpinLockAcquire(&local_slot->mutex);
564 2 : synced_slot = local_slot->in_use && local_slot->data.synced;
565 2 : SpinLockRelease(&local_slot->mutex);
566 :
567 2 : if (synced_slot)
568 : {
569 2 : ReplicationSlotAcquire(NameStr(local_slot->data.name), true, false);
570 2 : ReplicationSlotDropAcquired();
571 : }
572 :
573 2 : UnlockSharedObject(DatabaseRelationId, local_slot->data.database,
574 : 0, AccessShareLock);
575 :
576 2 : ereport(LOG,
577 : errmsg("dropped replication slot \"%s\" of database with OID %u",
578 : NameStr(local_slot->data.name),
579 : local_slot->data.database));
580 : }
581 : }
582 38 : }
583 :
584 : /*
585 : * Reserve WAL for the currently active local slot using the specified WAL
586 : * location (restart_lsn).
587 : *
588 : * If the given WAL location has been removed or is at risk of removal,
589 : * reserve WAL using the oldest segment that is non-removable.
590 : */
591 : static void
592 8 : reserve_wal_for_local_slot(XLogRecPtr restart_lsn)
593 : {
594 : XLogRecPtr slot_min_lsn;
595 : XLogRecPtr min_safe_lsn;
596 : XLogSegNo segno;
597 8 : ReplicationSlot *slot = MyReplicationSlot;
598 :
599 : Assert(slot != NULL);
600 : Assert(!XLogRecPtrIsValid(slot->data.restart_lsn));
601 :
602 : /*
603 : * Acquire an exclusive lock to prevent the checkpoint process from
604 : * concurrently calculating the minimum slot LSN (see
605 : * CheckPointReplicationSlots), ensuring that if WAL reservation occurs
606 : * first, the checkpoint must wait for the restart_lsn update before
607 : * calculating the minimum LSN.
608 : *
609 : * Note: Unlike ReplicationSlotReserveWal(), this lock does not protect a
610 : * newly synced slot from being invalidated if a concurrent checkpoint has
611 : * invoked CheckPointReplicationSlots() before the WAL reservation here.
612 : * This can happen because the initial restart_lsn received from the
613 : * remote server can precede the redo pointer. Therefore, when selecting
614 : * the initial restart_lsn, we consider using the redo pointer or the
615 : * minimum slot LSN (if those values are greater than the remote
616 : * restart_lsn) instead of relying solely on the remote value.
617 : */
618 8 : LWLockAcquire(ReplicationSlotAllocationLock, LW_EXCLUSIVE);
619 :
620 : /*
621 : * Determine the minimum non-removable LSN by comparing the redo pointer
622 : * with the minimum slot LSN.
623 : *
624 : * The minimum slot LSN is considered because the redo pointer advances at
625 : * every checkpoint, even when replication slots are present on the
626 : * standby. In such scenarios, the redo pointer can exceed the remote
627 : * restart_lsn, while WALs preceding the remote restart_lsn remain
628 : * protected by a local replication slot.
629 : */
630 8 : min_safe_lsn = GetRedoRecPtr();
631 8 : slot_min_lsn = XLogGetReplicationSlotMinimumLSN();
632 :
633 8 : if (XLogRecPtrIsValid(slot_min_lsn) && min_safe_lsn > slot_min_lsn)
634 0 : min_safe_lsn = slot_min_lsn;
635 :
636 : /*
637 : * If the minimum safe LSN is greater than the given restart_lsn, use it
638 : * as the initial restart_lsn for the newly synced slot. Otherwise, use
639 : * the given remote restart_lsn.
640 : */
641 8 : SpinLockAcquire(&slot->mutex);
642 8 : slot->data.restart_lsn = Max(restart_lsn, min_safe_lsn);
643 8 : SpinLockRelease(&slot->mutex);
644 :
645 8 : ReplicationSlotsComputeRequiredLSN();
646 :
647 8 : XLByteToSeg(slot->data.restart_lsn, segno, wal_segment_size);
648 8 : if (XLogGetLastRemovedSegno() >= segno)
649 0 : elog(ERROR, "WAL required by replication slot %s has been removed concurrently",
650 : NameStr(slot->data.name));
651 :
652 8 : LWLockRelease(ReplicationSlotAllocationLock);
653 8 : }
654 :
655 : /*
656 : * If the remote restart_lsn and catalog_xmin have caught up with the
657 : * local ones, then update the LSNs and persist the local synced slot for
658 : * future synchronization; otherwise, do nothing.
659 : *
660 : * *slot_persistence_pending is set to true if any of the slots fail to
661 : * persist.
662 : *
663 : * Return true if the slot is marked as RS_PERSISTENT (sync-ready), otherwise
664 : * false.
665 : */
666 : static bool
667 15 : update_and_persist_local_synced_slot(RemoteSlot *remote_slot, Oid remote_dbid,
668 : bool *slot_persistence_pending)
669 : {
670 15 : ReplicationSlot *slot = MyReplicationSlot;
671 :
672 : /* Slotsync skip stats are handled in function update_local_synced_slot() */
673 15 : (void) update_local_synced_slot(remote_slot, remote_dbid);
674 :
675 : /*
676 : * Check if the slot cannot be synchronized. Refer to the comment atop the
677 : * file for details on this check.
678 : */
679 15 : if (slot->slotsync_skip_reason != SS_SKIP_NONE)
680 : {
681 : /*
682 : * We reach this point when the remote slot didn't catch up to locally
683 : * reserved position, or it cannot reach the consistent point from the
684 : * restart_lsn, or the WAL prior to the remote confirmed flush LSN has
685 : * not been received and flushed.
686 : *
687 : * We do not drop the slot because the restart_lsn and confirmed_lsn
688 : * can be ahead of the current location when recreating the slot in
689 : * the next cycle. It may take more time to create such a slot or
690 : * reach the consistent point. Therefore, we keep this slot and
691 : * attempt the synchronization in the next cycle.
692 : *
693 : * We also update the slot_persistence_pending parameter, so the SQL
694 : * function can retry.
695 : */
696 8 : if (slot_persistence_pending)
697 2 : *slot_persistence_pending = true;
698 :
699 8 : return false;
700 : }
701 :
702 7 : ReplicationSlotPersist();
703 :
704 7 : ereport(LOG,
705 : errmsg("newly created replication slot \"%s\" is sync-ready now",
706 : remote_slot->name));
707 :
708 7 : return true;
709 : }
710 :
711 : /*
712 : * Synchronize a single slot to the given position.
713 : *
714 : * This creates a new slot if there is no existing one and updates the
715 : * metadata of the slot as per the data received from the primary server.
716 : *
717 : * The slot is created as a temporary slot and stays in the same state until the
718 : * remote_slot catches up with locally reserved position and local slot is
719 : * updated. The slot is then persisted and is considered as sync-ready for
720 : * periodic syncs.
721 : *
722 : * *slot_persistence_pending is set to true if any of the slots fail to
723 : * persist.
724 : *
725 : * Returns TRUE if the local slot is updated.
726 : */
727 : static bool
728 63 : synchronize_one_slot(RemoteSlot *remote_slot, Oid remote_dbid,
729 : bool *slot_persistence_pending)
730 : {
731 : ReplicationSlot *slot;
732 63 : bool slot_updated = false;
733 :
734 : /* Search for the named slot */
735 63 : if ((slot = SearchNamedReplicationSlot(remote_slot->name, true)))
736 : {
737 : bool synced;
738 :
739 55 : SpinLockAcquire(&slot->mutex);
740 55 : synced = slot->data.synced;
741 55 : SpinLockRelease(&slot->mutex);
742 :
743 : /* User-created slot with the same name exists, raise ERROR. */
744 55 : if (!synced)
745 0 : ereport(ERROR,
746 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
747 : errmsg("exiting from slot synchronization because same"
748 : " name slot \"%s\" already exists on the standby",
749 : remote_slot->name));
750 :
751 : /*
752 : * The slot has been synchronized before.
753 : *
754 : * It is important to acquire the slot here before checking
755 : * invalidation. If we don't acquire the slot first, there could be a
756 : * race condition that the local slot could be invalidated just after
757 : * checking the 'invalidated' flag here and we could end up
758 : * overwriting 'invalidated' flag to remote_slot's value. See
759 : * InvalidatePossiblyObsoleteSlot() where it invalidates slot directly
760 : * if the slot is not acquired by other processes.
761 : *
762 : * XXX: If it ever turns out that slot acquire/release is costly for
763 : * cases when none of the slot properties is changed then we can do a
764 : * pre-check to ensure that at least one of the slot properties is
765 : * changed before acquiring the slot.
766 : */
767 55 : ReplicationSlotAcquire(remote_slot->name, true, false);
768 :
769 : Assert(slot == MyReplicationSlot);
770 :
771 : /*
772 : * Copy the invalidation cause from remote only if local slot is not
773 : * invalidated locally, we don't want to overwrite existing one.
774 : */
775 55 : if (slot->data.invalidated == RS_INVAL_NONE &&
776 55 : remote_slot->invalidated != RS_INVAL_NONE)
777 : {
778 0 : SpinLockAcquire(&slot->mutex);
779 0 : slot->data.invalidated = remote_slot->invalidated;
780 0 : SpinLockRelease(&slot->mutex);
781 :
782 : /* Make sure the invalidated state persists across server restart */
783 0 : ReplicationSlotMarkDirty();
784 0 : ReplicationSlotSave();
785 :
786 0 : slot_updated = true;
787 : }
788 :
789 : /* Skip the sync of an invalidated slot */
790 55 : if (slot->data.invalidated != RS_INVAL_NONE)
791 : {
792 0 : update_slotsync_skip_stats(SS_SKIP_INVALID);
793 :
794 0 : ReplicationSlotRelease();
795 0 : return slot_updated;
796 : }
797 :
798 : /* Slot not ready yet, let's attempt to make it sync-ready now. */
799 55 : if (slot->data.persistency == RS_TEMPORARY)
800 : {
801 7 : slot_updated = update_and_persist_local_synced_slot(remote_slot,
802 : remote_dbid,
803 : slot_persistence_pending);
804 : }
805 :
806 : /* Slot ready for sync, so sync it. */
807 : else
808 : {
809 : /*
810 : * Sanity check: As long as the invalidations are handled
811 : * appropriately as above, this should never happen.
812 : *
813 : * We don't need to check restart_lsn here. See the comments in
814 : * update_local_synced_slot() for details.
815 : */
816 48 : if (remote_slot->confirmed_lsn < slot->data.confirmed_flush)
817 0 : ereport(ERROR,
818 : errmsg_internal("cannot synchronize local slot \"%s\"",
819 : remote_slot->name),
820 : errdetail_internal("Local slot's start streaming location LSN(%X/%08X) is ahead of remote slot's LSN(%X/%08X).",
821 : LSN_FORMAT_ARGS(slot->data.confirmed_flush),
822 : LSN_FORMAT_ARGS(remote_slot->confirmed_lsn)));
823 :
824 48 : slot_updated = update_local_synced_slot(remote_slot, remote_dbid);
825 : }
826 : }
827 : /* Otherwise create the slot first. */
828 : else
829 : {
830 : NameData plugin_name;
831 8 : TransactionId xmin_horizon = InvalidTransactionId;
832 :
833 : /* Skip creating the local slot if remote_slot is invalidated already */
834 8 : if (remote_slot->invalidated != RS_INVAL_NONE)
835 0 : return false;
836 :
837 : /*
838 : * We create temporary slots instead of ephemeral slots here because
839 : * we want the slots to survive after releasing them. This is done to
840 : * avoid dropping and re-creating the slots in each synchronization
841 : * cycle if the restart_lsn or catalog_xmin of the remote slot has not
842 : * caught up.
843 : */
844 8 : ReplicationSlotCreate(remote_slot->name, true, RS_TEMPORARY,
845 8 : remote_slot->two_phase,
846 : false,
847 8 : remote_slot->failover,
848 : true);
849 :
850 : /* For shorter lines. */
851 8 : slot = MyReplicationSlot;
852 :
853 : /* Avoid expensive operations while holding a spinlock. */
854 8 : namestrcpy(&plugin_name, remote_slot->plugin);
855 :
856 8 : SpinLockAcquire(&slot->mutex);
857 8 : slot->data.database = remote_dbid;
858 8 : slot->data.plugin = plugin_name;
859 8 : SpinLockRelease(&slot->mutex);
860 :
861 8 : reserve_wal_for_local_slot(remote_slot->restart_lsn);
862 :
863 8 : LWLockAcquire(ReplicationSlotControlLock, LW_EXCLUSIVE);
864 8 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
865 8 : xmin_horizon = GetOldestSafeDecodingTransactionId(true);
866 8 : SpinLockAcquire(&slot->mutex);
867 8 : slot->effective_catalog_xmin = xmin_horizon;
868 8 : slot->data.catalog_xmin = xmin_horizon;
869 8 : SpinLockRelease(&slot->mutex);
870 8 : ReplicationSlotsComputeRequiredXmin(true);
871 8 : LWLockRelease(ProcArrayLock);
872 8 : LWLockRelease(ReplicationSlotControlLock);
873 :
874 8 : update_and_persist_local_synced_slot(remote_slot, remote_dbid,
875 : slot_persistence_pending);
876 :
877 8 : slot_updated = true;
878 : }
879 :
880 63 : ReplicationSlotRelease();
881 :
882 63 : return slot_updated;
883 : }
884 :
885 : /*
886 : * Fetch remote slots.
887 : *
888 : * If slot_names is NIL, fetches all failover logical slots from the
889 : * primary server, otherwise fetches only the ones with names in slot_names.
890 : *
891 : * Returns a list of remote slot information structures, or NIL if none
892 : * are found.
893 : */
894 : static List *
895 40 : fetch_remote_slots(WalReceiverConn *wrconn, List *slot_names)
896 : {
897 : #define SLOTSYNC_COLUMN_COUNT 10
898 40 : Oid slotRow[SLOTSYNC_COLUMN_COUNT] = {TEXTOID, TEXTOID, LSNOID,
899 : LSNOID, XIDOID, BOOLOID, LSNOID, BOOLOID, TEXTOID, TEXTOID};
900 :
901 : WalRcvExecResult *res;
902 : TupleTableSlot *tupslot;
903 40 : List *remote_slot_list = NIL;
904 : StringInfoData query;
905 :
906 40 : initStringInfo(&query);
907 40 : appendStringInfoString(&query,
908 : "SELECT slot_name, plugin, confirmed_flush_lsn,"
909 : " restart_lsn, catalog_xmin, two_phase,"
910 : " two_phase_at, failover,"
911 : " database, invalidation_reason"
912 : " FROM pg_catalog.pg_replication_slots"
913 : " WHERE failover and NOT temporary");
914 :
915 40 : if (slot_names != NIL)
916 : {
917 2 : bool first_slot = true;
918 :
919 : /*
920 : * Construct the query to fetch only the specified slots
921 : */
922 2 : appendStringInfoString(&query, " AND slot_name IN (");
923 :
924 6 : foreach_ptr(char, slot_name, slot_names)
925 : {
926 2 : if (!first_slot)
927 0 : appendStringInfoString(&query, ", ");
928 :
929 2 : appendStringInfoString(&query, quote_literal_cstr(slot_name));
930 2 : first_slot = false;
931 : }
932 2 : appendStringInfoChar(&query, ')');
933 : }
934 :
935 : /* Execute the query */
936 40 : res = walrcv_exec(wrconn, query.data, SLOTSYNC_COLUMN_COUNT, slotRow);
937 40 : pfree(query.data);
938 40 : if (res->status != WALRCV_OK_TUPLES)
939 2 : ereport(ERROR,
940 : errmsg("could not fetch failover logical slots info from the primary server: %s",
941 : res->err));
942 :
943 38 : tupslot = MakeSingleTupleTableSlot(res->tupledesc, &TTSOpsMinimalTuple);
944 101 : while (tuplestore_gettupleslot(res->tuplestore, true, false, tupslot))
945 : {
946 : bool isnull;
947 63 : RemoteSlot *remote_slot = palloc0_object(RemoteSlot);
948 : Datum d;
949 63 : int col = 0;
950 :
951 63 : remote_slot->name = TextDatumGetCString(slot_getattr(tupslot, ++col,
952 : &isnull));
953 : Assert(!isnull);
954 :
955 63 : remote_slot->plugin = TextDatumGetCString(slot_getattr(tupslot, ++col,
956 : &isnull));
957 : Assert(!isnull);
958 :
959 : /*
960 : * It is possible to get null values for LSN and Xmin if slot is
961 : * invalidated on the primary server, so handle accordingly.
962 : */
963 63 : d = slot_getattr(tupslot, ++col, &isnull);
964 63 : remote_slot->confirmed_lsn = isnull ? InvalidXLogRecPtr :
965 63 : DatumGetLSN(d);
966 :
967 63 : d = slot_getattr(tupslot, ++col, &isnull);
968 63 : remote_slot->restart_lsn = isnull ? InvalidXLogRecPtr : DatumGetLSN(d);
969 :
970 63 : d = slot_getattr(tupslot, ++col, &isnull);
971 63 : remote_slot->catalog_xmin = isnull ? InvalidTransactionId :
972 63 : DatumGetTransactionId(d);
973 :
974 63 : remote_slot->two_phase = DatumGetBool(slot_getattr(tupslot, ++col,
975 : &isnull));
976 : Assert(!isnull);
977 :
978 63 : d = slot_getattr(tupslot, ++col, &isnull);
979 63 : remote_slot->two_phase_at = isnull ? InvalidXLogRecPtr : DatumGetLSN(d);
980 :
981 63 : remote_slot->failover = DatumGetBool(slot_getattr(tupslot, ++col,
982 : &isnull));
983 : Assert(!isnull);
984 :
985 63 : remote_slot->database = TextDatumGetCString(slot_getattr(tupslot,
986 : ++col, &isnull));
987 : Assert(!isnull);
988 :
989 63 : d = slot_getattr(tupslot, ++col, &isnull);
990 63 : remote_slot->invalidated = isnull ? RS_INVAL_NONE :
991 0 : GetSlotInvalidationCause(TextDatumGetCString(d));
992 :
993 : /* Sanity check */
994 : Assert(col == SLOTSYNC_COLUMN_COUNT);
995 :
996 : /*
997 : * If restart_lsn, confirmed_lsn or catalog_xmin is invalid but the
998 : * slot is valid, that means we have fetched the remote_slot in its
999 : * RS_EPHEMERAL state. In such a case, don't sync it; we can always
1000 : * sync it in the next sync cycle when the remote_slot is persisted
1001 : * and has valid lsn(s) and xmin values.
1002 : *
1003 : * XXX: In future, if we plan to expose 'slot->data.persistency' in
1004 : * pg_replication_slots view, then we can avoid fetching RS_EPHEMERAL
1005 : * slots in the first place.
1006 : */
1007 63 : if ((!XLogRecPtrIsValid(remote_slot->restart_lsn) ||
1008 63 : !XLogRecPtrIsValid(remote_slot->confirmed_lsn) ||
1009 63 : !TransactionIdIsValid(remote_slot->catalog_xmin)) &&
1010 0 : remote_slot->invalidated == RS_INVAL_NONE)
1011 0 : pfree(remote_slot);
1012 : else
1013 : /* Create list of remote slots */
1014 63 : remote_slot_list = lappend(remote_slot_list, remote_slot);
1015 :
1016 63 : ExecClearTuple(tupslot);
1017 : }
1018 :
1019 38 : walrcv_clear_result(res);
1020 :
1021 38 : return remote_slot_list;
1022 : }
1023 :
1024 : /*
1025 : * Synchronize slots.
1026 : *
1027 : * This function takes a list of remote slots and synchronizes them locally. It
1028 : * creates the slots if not present on the standby and updates existing ones.
1029 : *
1030 : * If slot_persistence_pending is not NULL, it will be set to true if one or
1031 : * more slots could not be persisted. This allows callers such as
1032 : * SyncReplicationSlots() to retry those slots.
1033 : *
1034 : * Returns TRUE if any of the slots gets updated in this sync-cycle.
1035 : */
1036 : static bool
1037 38 : synchronize_slots(WalReceiverConn *wrconn, List *remote_slot_list,
1038 : bool *slot_persistence_pending)
1039 : {
1040 38 : bool some_slot_updated = false;
1041 :
1042 : /* Drop local slots that no longer need to be synced. */
1043 38 : drop_local_obsolete_slots(remote_slot_list);
1044 :
1045 : /* Now sync the slots locally */
1046 139 : foreach_ptr(RemoteSlot, remote_slot, remote_slot_list)
1047 : {
1048 63 : Oid remote_dbid = get_database_oid(remote_slot->database, false);
1049 :
1050 : /*
1051 : * Use shared lock to prevent a conflict with
1052 : * ReplicationSlotsDropDBSlots(), trying to drop the same slot during
1053 : * a drop-database operation.
1054 : */
1055 63 : LockSharedObject(DatabaseRelationId, remote_dbid, 0, AccessShareLock);
1056 :
1057 63 : some_slot_updated |= synchronize_one_slot(remote_slot, remote_dbid,
1058 : slot_persistence_pending);
1059 :
1060 63 : UnlockSharedObject(DatabaseRelationId, remote_dbid, 0, AccessShareLock);
1061 : }
1062 :
1063 38 : return some_slot_updated;
1064 : }
1065 :
1066 : /*
1067 : * Checks the remote server info.
1068 : *
1069 : * We ensure that the 'primary_slot_name' exists on the remote server and the
1070 : * remote server is not a standby node.
1071 : */
1072 : static void
1073 15 : validate_remote_info(WalReceiverConn *wrconn)
1074 : {
1075 : #define PRIMARY_INFO_OUTPUT_COL_COUNT 2
1076 : WalRcvExecResult *res;
1077 15 : Oid slotRow[PRIMARY_INFO_OUTPUT_COL_COUNT] = {BOOLOID, BOOLOID};
1078 : StringInfoData cmd;
1079 : bool isnull;
1080 : TupleTableSlot *tupslot;
1081 : bool remote_in_recovery;
1082 : bool primary_slot_valid;
1083 15 : bool started_tx = false;
1084 :
1085 15 : initStringInfo(&cmd);
1086 15 : appendStringInfo(&cmd,
1087 : "SELECT pg_is_in_recovery(), count(*) = 1"
1088 : " FROM pg_catalog.pg_replication_slots"
1089 : " WHERE slot_type='physical' AND slot_name=%s",
1090 : quote_literal_cstr(PrimarySlotName));
1091 :
1092 : /* The syscache access in walrcv_exec() needs a transaction env. */
1093 15 : if (!IsTransactionState())
1094 : {
1095 6 : StartTransactionCommand();
1096 6 : started_tx = true;
1097 : }
1098 :
1099 15 : res = walrcv_exec(wrconn, cmd.data, PRIMARY_INFO_OUTPUT_COL_COUNT, slotRow);
1100 15 : pfree(cmd.data);
1101 :
1102 15 : if (res->status != WALRCV_OK_TUPLES)
1103 0 : ereport(ERROR,
1104 : errmsg("could not fetch primary slot name \"%s\" info from the primary server: %s",
1105 : PrimarySlotName, res->err),
1106 : errhint("Check if \"primary_slot_name\" is configured correctly."));
1107 :
1108 15 : tupslot = MakeSingleTupleTableSlot(res->tupledesc, &TTSOpsMinimalTuple);
1109 15 : if (!tuplestore_gettupleslot(res->tuplestore, true, false, tupslot))
1110 0 : elog(ERROR,
1111 : "failed to fetch tuple for the primary server slot specified by \"primary_slot_name\"");
1112 :
1113 15 : remote_in_recovery = DatumGetBool(slot_getattr(tupslot, 1, &isnull));
1114 : Assert(!isnull);
1115 :
1116 : /*
1117 : * Slot sync is currently not supported on a cascading standby. This is
1118 : * because if we allow it, the primary server needs to wait for all the
1119 : * cascading standbys, otherwise, logical subscribers can still be ahead
1120 : * of one of the cascading standbys which we plan to promote. Thus, to
1121 : * avoid this additional complexity, we restrict it for the time being.
1122 : */
1123 15 : if (remote_in_recovery)
1124 1 : ereport(ERROR,
1125 : errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1126 : errmsg("cannot synchronize replication slots from a standby server"));
1127 :
1128 14 : primary_slot_valid = DatumGetBool(slot_getattr(tupslot, 2, &isnull));
1129 : Assert(!isnull);
1130 :
1131 14 : if (!primary_slot_valid)
1132 0 : ereport(ERROR,
1133 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1134 : /* translator: second %s is a GUC variable name */
1135 : errmsg("replication slot \"%s\" specified by \"%s\" does not exist on primary server",
1136 : PrimarySlotName, "primary_slot_name"));
1137 :
1138 14 : ExecClearTuple(tupslot);
1139 14 : walrcv_clear_result(res);
1140 :
1141 14 : if (started_tx)
1142 6 : CommitTransactionCommand();
1143 14 : }
1144 :
1145 : /*
1146 : * Checks if dbname is specified in 'primary_conninfo'.
1147 : *
1148 : * Error out if not specified otherwise return it.
1149 : */
1150 : char *
1151 16 : CheckAndGetDbnameFromConninfo(void)
1152 : {
1153 : char *dbname;
1154 :
1155 : /*
1156 : * The slot synchronization needs a database connection for walrcv_exec to
1157 : * work.
1158 : */
1159 16 : dbname = walrcv_get_dbname_from_conninfo(PrimaryConnInfo);
1160 16 : if (dbname == NULL)
1161 1 : ereport(ERROR,
1162 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1163 :
1164 : /*
1165 : * translator: first %s is a connection option; second %s is a GUC
1166 : * variable name
1167 : */
1168 : errmsg("replication slot synchronization requires \"%s\" to be specified in \"%s\"",
1169 : "dbname", "primary_conninfo"));
1170 15 : return dbname;
1171 : }
1172 :
1173 : /*
1174 : * Return true if all necessary GUCs for slot synchronization are set
1175 : * appropriately, otherwise, return false.
1176 : */
1177 : bool
1178 28 : ValidateSlotSyncParams(int elevel)
1179 : {
1180 : /*
1181 : * Logical slot sync/creation requires logical decoding to be enabled.
1182 : */
1183 28 : if (!IsLogicalDecodingEnabled())
1184 : {
1185 0 : ereport(elevel,
1186 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1187 : errmsg("replication slot synchronization requires \"effective_wal_level\" >= \"logical\" on the primary"),
1188 : errhint("To enable logical decoding on primary, set \"wal_level\" >= \"logical\" or create at least one logical slot when \"wal_level\" = \"replica\"."));
1189 :
1190 0 : return false;
1191 : }
1192 :
1193 : /*
1194 : * A physical replication slot(primary_slot_name) is required on the
1195 : * primary to ensure that the rows needed by the standby are not removed
1196 : * after restarting, so that the synchronized slot on the standby will not
1197 : * be invalidated.
1198 : */
1199 28 : if (PrimarySlotName == NULL || *PrimarySlotName == '\0')
1200 : {
1201 0 : ereport(elevel,
1202 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1203 : /* translator: %s is a GUC variable name */
1204 : errmsg("replication slot synchronization requires \"%s\" to be set", "primary_slot_name"));
1205 0 : return false;
1206 : }
1207 :
1208 : /*
1209 : * hot_standby_feedback must be enabled to cooperate with the physical
1210 : * replication slot, which allows informing the primary about the xmin and
1211 : * catalog_xmin values on the standby.
1212 : */
1213 28 : if (!hot_standby_feedback)
1214 : {
1215 1 : ereport(elevel,
1216 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1217 : /* translator: %s is a GUC variable name */
1218 : errmsg("replication slot synchronization requires \"%s\" to be enabled",
1219 : "hot_standby_feedback"));
1220 1 : return false;
1221 : }
1222 :
1223 : /*
1224 : * The primary_conninfo is required to make connection to primary for
1225 : * getting slots information.
1226 : */
1227 27 : if (PrimaryConnInfo == NULL || *PrimaryConnInfo == '\0')
1228 : {
1229 0 : ereport(elevel,
1230 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1231 : /* translator: %s is a GUC variable name */
1232 : errmsg("replication slot synchronization requires \"%s\" to be set",
1233 : "primary_conninfo"));
1234 0 : return false;
1235 : }
1236 :
1237 27 : return true;
1238 : }
1239 :
1240 : /*
1241 : * Re-read the config file for slot synchronization.
1242 : *
1243 : * Exit or throw error if relevant GUCs have changed depending on whether
1244 : * called from slot sync worker or from the SQL function pg_sync_replication_slots()
1245 : */
1246 : static void
1247 1 : slotsync_reread_config(void)
1248 : {
1249 1 : char *old_primary_conninfo = pstrdup(PrimaryConnInfo);
1250 1 : char *old_primary_slotname = pstrdup(PrimarySlotName);
1251 1 : bool old_sync_replication_slots = sync_replication_slots;
1252 1 : bool old_hot_standby_feedback = hot_standby_feedback;
1253 : bool conninfo_changed;
1254 : bool primary_slotname_changed;
1255 1 : bool is_slotsync_worker = AmLogicalSlotSyncWorkerProcess();
1256 1 : bool parameter_changed = false;
1257 :
1258 : if (is_slotsync_worker)
1259 : Assert(sync_replication_slots);
1260 :
1261 1 : ConfigReloadPending = false;
1262 1 : ProcessConfigFile(PGC_SIGHUP);
1263 :
1264 1 : conninfo_changed = strcmp(old_primary_conninfo, PrimaryConnInfo) != 0;
1265 1 : primary_slotname_changed = strcmp(old_primary_slotname, PrimarySlotName) != 0;
1266 1 : pfree(old_primary_conninfo);
1267 1 : pfree(old_primary_slotname);
1268 :
1269 1 : if (old_sync_replication_slots != sync_replication_slots)
1270 : {
1271 0 : if (is_slotsync_worker)
1272 : {
1273 0 : ereport(LOG,
1274 : /* translator: %s is a GUC variable name */
1275 : errmsg("replication slot synchronization worker will stop because \"%s\" is disabled",
1276 : "sync_replication_slots"));
1277 :
1278 0 : proc_exit(0);
1279 : }
1280 :
1281 0 : parameter_changed = true;
1282 : }
1283 : else
1284 : {
1285 1 : if (conninfo_changed ||
1286 1 : primary_slotname_changed ||
1287 1 : (old_hot_standby_feedback != hot_standby_feedback))
1288 : {
1289 :
1290 1 : if (is_slotsync_worker)
1291 : {
1292 1 : ereport(LOG,
1293 : errmsg("replication slot synchronization worker will restart because of a parameter change"));
1294 :
1295 : /*
1296 : * Reset the last-start time for this worker so that the
1297 : * postmaster can restart it without waiting for
1298 : * SLOTSYNC_RESTART_INTERVAL_SEC.
1299 : */
1300 1 : SlotSyncCtx->last_start_time = 0;
1301 :
1302 1 : proc_exit(0);
1303 : }
1304 :
1305 0 : parameter_changed = true;
1306 : }
1307 : }
1308 :
1309 : /*
1310 : * If we have reached here with a parameter change, we must be running in
1311 : * SQL function, emit error in such a case.
1312 : */
1313 0 : if (parameter_changed)
1314 : {
1315 : Assert(!is_slotsync_worker);
1316 0 : ereport(ERROR,
1317 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1318 : errmsg("replication slot synchronization will stop because of a parameter change"));
1319 : }
1320 :
1321 0 : }
1322 :
1323 : /*
1324 : * Handle receipt of an interrupt indicating a slotsync shutdown message.
1325 : *
1326 : * This is called within the SIGUSR1 handler. All we do here is set a flag
1327 : * that will cause the next CHECK_FOR_INTERRUPTS() to invoke
1328 : * ProcessSlotSyncMessage().
1329 : */
1330 : void
1331 1 : HandleSlotSyncMessageInterrupt(void)
1332 : {
1333 1 : InterruptPending = true;
1334 1 : SlotSyncShutdownPending = true;
1335 : /* latch will be set by procsignal_sigusr1_handler */
1336 1 : }
1337 :
1338 : /*
1339 : * Handle a PROCSIG_SLOTSYNC_MESSAGE signal, called from ProcessInterrupts().
1340 : *
1341 : * If the current process is the slotsync background worker, log a message
1342 : * and exit cleanly. If it is a backend executing pg_sync_replication_slots(),
1343 : * raise an error, unless the sync has already finished, in which case there
1344 : * is no need to interrupt the caller.
1345 : */
1346 : void
1347 1 : ProcessSlotSyncMessage(void)
1348 : {
1349 1 : SlotSyncShutdownPending = false;
1350 :
1351 1 : if (AmLogicalSlotSyncWorkerProcess())
1352 : {
1353 1 : ereport(LOG,
1354 : errmsg("replication slot synchronization worker will stop because promotion is triggered"));
1355 1 : proc_exit(0);
1356 : }
1357 : else
1358 : {
1359 : /*
1360 : * If sync has already completed, there is no need to interrupt the
1361 : * caller with an error.
1362 : */
1363 0 : if (!IsSyncingReplicationSlots())
1364 0 : return;
1365 :
1366 0 : ereport(ERROR,
1367 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1368 : errmsg("replication slot synchronization will stop because promotion is triggered"));
1369 : }
1370 : }
1371 :
1372 : /*
1373 : * Connection cleanup function for slotsync worker.
1374 : *
1375 : * Called on slotsync worker exit.
1376 : */
1377 : static void
1378 6 : slotsync_worker_disconnect(int code, Datum arg)
1379 : {
1380 6 : WalReceiverConn *wrconn = (WalReceiverConn *) DatumGetPointer(arg);
1381 :
1382 6 : walrcv_disconnect(wrconn);
1383 6 : }
1384 :
1385 : /*
1386 : * Cleanup function for slotsync worker.
1387 : *
1388 : * Called on slotsync worker exit.
1389 : */
1390 : static void
1391 6 : slotsync_worker_onexit(int code, Datum arg)
1392 : {
1393 : /*
1394 : * We need to do slots cleanup here just like WalSndErrorCleanup() does.
1395 : *
1396 : * The startup process during promotion invokes ShutDownSlotSync() which
1397 : * waits for slot sync to finish and it does that by checking the
1398 : * 'syncing' flag. Thus the slot sync worker must be done with slots'
1399 : * release and cleanup to avoid any dangling temporary slots or active
1400 : * slots before it marks itself as finished syncing.
1401 : */
1402 :
1403 : /* Make sure active replication slots are released */
1404 6 : if (MyReplicationSlot != NULL)
1405 0 : ReplicationSlotRelease();
1406 :
1407 : /* Also cleanup the temporary slots. */
1408 6 : ReplicationSlotCleanup(false);
1409 :
1410 6 : SpinLockAcquire(&SlotSyncCtx->mutex);
1411 :
1412 6 : SlotSyncCtx->pid = InvalidPid;
1413 :
1414 : /*
1415 : * If syncing_slots is true, it indicates that the process errored out
1416 : * without resetting the flag. So, we need to clean up shared memory and
1417 : * reset the flag here.
1418 : */
1419 6 : if (syncing_slots)
1420 : {
1421 6 : SlotSyncCtx->syncing = false;
1422 6 : syncing_slots = false;
1423 : }
1424 :
1425 6 : SpinLockRelease(&SlotSyncCtx->mutex);
1426 6 : }
1427 :
1428 : /*
1429 : * Sleep for long enough that we believe it's likely that the slots on primary
1430 : * get updated.
1431 : *
1432 : * If there is no slot activity the wait time between sync-cycles will double
1433 : * (to a maximum of 30s). If there is some slot activity the wait time between
1434 : * sync-cycles is reset to the minimum (200ms).
1435 : */
1436 : static void
1437 30 : wait_for_slot_activity(bool some_slot_updated)
1438 : {
1439 : int rc;
1440 :
1441 30 : if (!some_slot_updated)
1442 : {
1443 : /*
1444 : * No slots were updated, so double the sleep time, but not beyond the
1445 : * maximum allowable value.
1446 : */
1447 18 : sleep_ms = Min(sleep_ms * 2, MAX_SLOTSYNC_WORKER_NAPTIME_MS);
1448 : }
1449 : else
1450 : {
1451 : /*
1452 : * Some slots were updated since the last sleep, so reset the sleep
1453 : * time.
1454 : */
1455 12 : sleep_ms = MIN_SLOTSYNC_WORKER_NAPTIME_MS;
1456 : }
1457 :
1458 30 : rc = WaitLatch(MyLatch,
1459 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
1460 : sleep_ms,
1461 : WAIT_EVENT_REPLICATION_SLOTSYNC_MAIN);
1462 :
1463 30 : if (rc & WL_LATCH_SET)
1464 4 : ResetLatch(MyLatch);
1465 30 : }
1466 :
1467 : /*
1468 : * Emit an error if a concurrent sync call is in progress.
1469 : * Otherwise, advertise that a sync is in progress.
1470 : */
1471 : static void
1472 15 : check_and_set_sync_info(pid_t sync_process_pid)
1473 : {
1474 15 : SpinLockAcquire(&SlotSyncCtx->mutex);
1475 :
1476 : /*
1477 : * Exit immediately if promotion has been triggered. This guards against
1478 : * a new worker (or a call to pg_sync_replication_slots()) that starts
1479 : * after the old worker was stopped by ShutDownSlotSync().
1480 : */
1481 15 : if (SlotSyncCtx->stopSignaled)
1482 : {
1483 0 : SpinLockRelease(&SlotSyncCtx->mutex);
1484 :
1485 0 : if (AmLogicalSlotSyncWorkerProcess())
1486 : {
1487 0 : ereport(DEBUG1,
1488 : errmsg("replication slot synchronization worker will not start because promotion was triggered"));
1489 :
1490 0 : proc_exit(0);
1491 : }
1492 : else
1493 : {
1494 : /*
1495 : * For the backend executing SQL function
1496 : * pg_sync_replication_slots().
1497 : */
1498 0 : ereport(ERROR,
1499 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1500 : errmsg("replication slot synchronization will not start because promotion was triggered"));
1501 : }
1502 : }
1503 :
1504 15 : if (SlotSyncCtx->syncing)
1505 : {
1506 0 : SpinLockRelease(&SlotSyncCtx->mutex);
1507 0 : ereport(ERROR,
1508 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1509 : errmsg("cannot synchronize replication slots concurrently"));
1510 : }
1511 :
1512 : /* The pid must not be already assigned in SlotSyncCtx */
1513 : Assert(SlotSyncCtx->pid == InvalidPid);
1514 :
1515 15 : SlotSyncCtx->syncing = true;
1516 :
1517 : /*
1518 : * Advertise the required PID so that the startup process can kill the
1519 : * slot sync process on promotion.
1520 : */
1521 15 : SlotSyncCtx->pid = sync_process_pid;
1522 :
1523 15 : SpinLockRelease(&SlotSyncCtx->mutex);
1524 :
1525 15 : syncing_slots = true;
1526 15 : }
1527 :
1528 : /*
1529 : * Reset syncing flag.
1530 : */
1531 : static void
1532 9 : reset_syncing_flag(void)
1533 : {
1534 9 : SpinLockAcquire(&SlotSyncCtx->mutex);
1535 9 : SlotSyncCtx->syncing = false;
1536 9 : SlotSyncCtx->pid = InvalidPid;
1537 9 : SpinLockRelease(&SlotSyncCtx->mutex);
1538 :
1539 9 : syncing_slots = false;
1540 9 : }
1541 :
1542 : /*
1543 : * The main loop of our worker process.
1544 : *
1545 : * It connects to the primary server, fetches logical failover slots
1546 : * information periodically in order to create and sync the slots.
1547 : *
1548 : * Note: If any changes are made here, check if the corresponding SQL
1549 : * function logic in SyncReplicationSlots() also needs to be changed.
1550 : */
1551 : void
1552 6 : ReplSlotSyncWorkerMain(const void *startup_data, size_t startup_data_len)
1553 : {
1554 6 : WalReceiverConn *wrconn = NULL;
1555 : char *dbname;
1556 : char *err;
1557 : sigjmp_buf local_sigjmp_buf;
1558 : StringInfoData app_name;
1559 :
1560 : Assert(startup_data_len == 0);
1561 :
1562 : /* Release postmaster's working memory context */
1563 6 : if (PostmasterContext)
1564 : {
1565 6 : MemoryContextDelete(PostmasterContext);
1566 6 : PostmasterContext = NULL;
1567 : }
1568 :
1569 6 : init_ps_display(NULL);
1570 :
1571 : Assert(GetProcessingMode() == InitProcessing);
1572 :
1573 : /*
1574 : * Create a per-backend PGPROC struct in shared memory. We must do this
1575 : * before we access any shared memory.
1576 : */
1577 6 : InitProcess();
1578 :
1579 : /*
1580 : * Early initialization.
1581 : */
1582 6 : BaseInit();
1583 :
1584 : Assert(SlotSyncCtx != NULL);
1585 :
1586 : /*
1587 : * If an exception is encountered, processing resumes here.
1588 : *
1589 : * We just need to clean up, report the error, and go away.
1590 : *
1591 : * If we do not have this handling here, then since this worker process
1592 : * operates at the bottom of the exception stack, ERRORs turn into FATALs.
1593 : * Therefore, we create our own exception handler to catch ERRORs.
1594 : */
1595 6 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1596 : {
1597 : /* since not using PG_TRY, must reset error stack by hand */
1598 2 : error_context_stack = NULL;
1599 :
1600 : /* Prevents interrupts while cleaning up */
1601 2 : HOLD_INTERRUPTS();
1602 :
1603 : /* Report the error to the server log */
1604 2 : EmitErrorReport();
1605 :
1606 : /*
1607 : * We can now go away. Note that because we called InitProcess, a
1608 : * callback was registered to do ProcKill, which will clean up
1609 : * necessary state.
1610 : */
1611 2 : proc_exit(0);
1612 : }
1613 :
1614 : /* We can now handle ereport(ERROR) */
1615 6 : PG_exception_stack = &local_sigjmp_buf;
1616 :
1617 : /* Setup signal handling */
1618 6 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1619 6 : pqsignal(SIGINT, StatementCancelHandler);
1620 6 : pqsignal(SIGTERM, die);
1621 6 : pqsignal(SIGFPE, FloatExceptionHandler);
1622 6 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1623 6 : pqsignal(SIGUSR2, PG_SIG_IGN);
1624 6 : pqsignal(SIGPIPE, PG_SIG_IGN);
1625 6 : pqsignal(SIGCHLD, PG_SIG_DFL);
1626 :
1627 6 : check_and_set_sync_info(MyProcPid);
1628 :
1629 6 : ereport(LOG, errmsg("slot sync worker started"));
1630 :
1631 : /* Register it as soon as SlotSyncCtx->pid is initialized. */
1632 6 : before_shmem_exit(slotsync_worker_onexit, (Datum) 0);
1633 :
1634 : /*
1635 : * Establishes SIGALRM handler and initialize timeout module. It is needed
1636 : * by InitPostgres to register different timeouts.
1637 : */
1638 6 : InitializeTimeouts();
1639 :
1640 : /* Load the libpq-specific functions */
1641 6 : load_file("libpqwalreceiver", false);
1642 :
1643 : /*
1644 : * Unblock signals (they were blocked when the postmaster forked us)
1645 : */
1646 6 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
1647 :
1648 : /*
1649 : * Set always-secure search path, so malicious users can't redirect user
1650 : * code (e.g. operators).
1651 : *
1652 : * It's not strictly necessary since we won't be scanning or writing to
1653 : * any user table locally, but it's good to retain it here for added
1654 : * precaution.
1655 : */
1656 6 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1657 :
1658 6 : dbname = CheckAndGetDbnameFromConninfo();
1659 :
1660 : /*
1661 : * Connect to the database specified by the user in primary_conninfo. We
1662 : * need a database connection for walrcv_exec to work which we use to
1663 : * fetch slot information from the remote node. See comments atop
1664 : * libpqrcv_exec.
1665 : *
1666 : * We do not specify a specific user here since the slot sync worker will
1667 : * operate as a superuser. This is safe because the slot sync worker does
1668 : * not interact with user tables, eliminating the risk of executing
1669 : * arbitrary code within triggers.
1670 : */
1671 6 : InitPostgres(dbname, InvalidOid, NULL, InvalidOid, 0, NULL);
1672 :
1673 6 : SetProcessingMode(NormalProcessing);
1674 :
1675 6 : initStringInfo(&app_name);
1676 6 : if (cluster_name[0])
1677 6 : appendStringInfo(&app_name, "%s_%s", cluster_name, "slotsync worker");
1678 : else
1679 0 : appendStringInfoString(&app_name, "slotsync worker");
1680 :
1681 : /*
1682 : * Establish the connection to the primary server for slot
1683 : * synchronization.
1684 : */
1685 6 : wrconn = walrcv_connect(PrimaryConnInfo, false, false, false,
1686 : app_name.data, &err);
1687 :
1688 6 : if (!wrconn)
1689 0 : ereport(ERROR,
1690 : errcode(ERRCODE_CONNECTION_FAILURE),
1691 : errmsg("synchronization worker \"%s\" could not connect to the primary server: %s",
1692 : app_name.data, err));
1693 :
1694 6 : pfree(app_name.data);
1695 :
1696 : /*
1697 : * Register the disconnection callback.
1698 : *
1699 : * XXX: This can be combined with previous cleanup registration of
1700 : * slotsync_worker_onexit() but that will need the connection to be made
1701 : * global and we want to avoid introducing global for this purpose.
1702 : */
1703 6 : before_shmem_exit(slotsync_worker_disconnect, PointerGetDatum(wrconn));
1704 :
1705 : /*
1706 : * Using the specified primary server connection, check that we are not a
1707 : * cascading standby and slot configured in 'primary_slot_name' exists on
1708 : * the primary server.
1709 : */
1710 6 : validate_remote_info(wrconn);
1711 :
1712 : /* Main loop to synchronize slots */
1713 : for (;;)
1714 28 : {
1715 34 : bool some_slot_updated = false;
1716 34 : bool started_tx = false;
1717 : List *remote_slots;
1718 :
1719 34 : CHECK_FOR_INTERRUPTS();
1720 :
1721 31 : if (ConfigReloadPending)
1722 1 : slotsync_reread_config();
1723 :
1724 : /*
1725 : * The syscache access in fetch_remote_slots() needs a transaction
1726 : * env.
1727 : */
1728 30 : if (!IsTransactionState())
1729 : {
1730 30 : StartTransactionCommand();
1731 30 : started_tx = true;
1732 : }
1733 :
1734 30 : remote_slots = fetch_remote_slots(wrconn, NIL);
1735 28 : some_slot_updated = synchronize_slots(wrconn, remote_slots, NULL);
1736 28 : list_free_deep(remote_slots);
1737 :
1738 28 : if (started_tx)
1739 28 : CommitTransactionCommand();
1740 :
1741 28 : wait_for_slot_activity(some_slot_updated);
1742 : }
1743 :
1744 : /*
1745 : * The slot sync worker can't get here because it will only stop when it
1746 : * receives a stop request from the startup process, or when there is an
1747 : * error.
1748 : */
1749 : Assert(false);
1750 : }
1751 :
1752 : /*
1753 : * Update the inactive_since property for synced slots.
1754 : *
1755 : * Note that this function is currently called when we shutdown the slot
1756 : * sync machinery.
1757 : */
1758 : static void
1759 1011 : update_synced_slots_inactive_since(void)
1760 : {
1761 1011 : TimestampTz now = 0;
1762 :
1763 : /*
1764 : * We need to update inactive_since only when we are promoting standby to
1765 : * correctly interpret the inactive_since if the standby gets promoted
1766 : * without a restart. We don't want the slots to appear inactive for a
1767 : * long time after promotion if they haven't been synchronized recently.
1768 : * Whoever acquires the slot, i.e., makes the slot active, will reset it.
1769 : */
1770 1011 : if (!StandbyMode)
1771 959 : return;
1772 :
1773 : /* The slot sync worker or the SQL function mustn't be running by now */
1774 : Assert((SlotSyncCtx->pid == InvalidPid) && !SlotSyncCtx->syncing);
1775 :
1776 52 : LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
1777 :
1778 820 : for (int i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
1779 : {
1780 768 : ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
1781 :
1782 : /* Check if it is a synchronized slot */
1783 768 : if (s->in_use && s->data.synced)
1784 : {
1785 : Assert(SlotIsLogical(s));
1786 :
1787 : /* The slot must not be acquired by any process */
1788 : Assert(s->active_proc == INVALID_PROC_NUMBER);
1789 :
1790 : /* Use the same inactive_since time for all the slots. */
1791 3 : if (now == 0)
1792 2 : now = GetCurrentTimestamp();
1793 :
1794 3 : ReplicationSlotSetInactiveSince(s, now, true);
1795 : }
1796 : }
1797 :
1798 52 : LWLockRelease(ReplicationSlotControlLock);
1799 : }
1800 :
1801 : /*
1802 : * Shut down slot synchronization.
1803 : *
1804 : * This function sets stopSignaled=true and wakes up the slot sync process
1805 : * (either worker or backend running the SQL function pg_sync_replication_slots())
1806 : * so that worker can exit or the SQL function pg_sync_replication_slots() can
1807 : * finish. It also waits till the slot sync worker has exited or
1808 : * pg_sync_replication_slots() has finished.
1809 : */
1810 : void
1811 1011 : ShutDownSlotSync(void)
1812 : {
1813 : pid_t sync_process_pid;
1814 :
1815 1011 : SpinLockAcquire(&SlotSyncCtx->mutex);
1816 :
1817 1011 : SlotSyncCtx->stopSignaled = true;
1818 :
1819 : /*
1820 : * Return if neither the slot sync worker is running nor the function
1821 : * pg_sync_replication_slots() is executing.
1822 : */
1823 1011 : if (!SlotSyncCtx->syncing)
1824 : {
1825 1010 : SpinLockRelease(&SlotSyncCtx->mutex);
1826 1010 : update_synced_slots_inactive_since();
1827 1010 : return;
1828 : }
1829 :
1830 1 : sync_process_pid = SlotSyncCtx->pid;
1831 :
1832 1 : SpinLockRelease(&SlotSyncCtx->mutex);
1833 :
1834 : /*
1835 : * Signal process doing slotsync, if any, asking it to stop.
1836 : */
1837 1 : if (sync_process_pid != InvalidPid)
1838 1 : SendProcSignal(sync_process_pid, PROCSIG_SLOTSYNC_MESSAGE,
1839 : INVALID_PROC_NUMBER);
1840 :
1841 : /* Wait for slot sync to end */
1842 : for (;;)
1843 0 : {
1844 : int rc;
1845 :
1846 : /* Wait a bit, we don't expect to have to wait long */
1847 1 : rc = WaitLatch(MyLatch,
1848 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
1849 : 10L, WAIT_EVENT_REPLICATION_SLOTSYNC_SHUTDOWN);
1850 :
1851 1 : if (rc & WL_LATCH_SET)
1852 : {
1853 0 : ResetLatch(MyLatch);
1854 0 : CHECK_FOR_INTERRUPTS();
1855 : }
1856 :
1857 1 : SpinLockAcquire(&SlotSyncCtx->mutex);
1858 :
1859 : /* Ensure that no process is syncing the slots. */
1860 1 : if (!SlotSyncCtx->syncing)
1861 1 : break;
1862 :
1863 0 : SpinLockRelease(&SlotSyncCtx->mutex);
1864 : }
1865 :
1866 1 : SpinLockRelease(&SlotSyncCtx->mutex);
1867 :
1868 1 : update_synced_slots_inactive_since();
1869 : }
1870 :
1871 : /*
1872 : * SlotSyncWorkerCanRestart
1873 : *
1874 : * Return true, indicating worker is allowed to restart, if enough time has
1875 : * passed since it was last launched to reach SLOTSYNC_RESTART_INTERVAL_SEC.
1876 : * Otherwise return false.
1877 : *
1878 : * This is a safety valve to protect against continuous respawn attempts if the
1879 : * worker is dying immediately at launch. Note that since we will retry to
1880 : * launch the worker from the postmaster main loop, we will get another
1881 : * chance later.
1882 : */
1883 : bool
1884 17 : SlotSyncWorkerCanRestart(void)
1885 : {
1886 17 : time_t curtime = time(NULL);
1887 :
1888 : /*
1889 : * If first time through, or time somehow went backwards, always update
1890 : * last_start_time to match the current clock and allow worker start.
1891 : * Otherwise allow it only once enough time has elapsed.
1892 : */
1893 17 : if (SlotSyncCtx->last_start_time == 0 ||
1894 11 : curtime < SlotSyncCtx->last_start_time ||
1895 11 : curtime - SlotSyncCtx->last_start_time >= SLOTSYNC_RESTART_INTERVAL_SEC)
1896 : {
1897 6 : SlotSyncCtx->last_start_time = curtime;
1898 6 : return true;
1899 : }
1900 11 : return false;
1901 : }
1902 :
1903 : /*
1904 : * Is current process syncing replication slots?
1905 : *
1906 : * Could be either backend executing SQL function or slot sync worker.
1907 : */
1908 : bool
1909 29 : IsSyncingReplicationSlots(void)
1910 : {
1911 29 : return syncing_slots;
1912 : }
1913 :
1914 : /*
1915 : * Register shared memory space needed for slot synchronization.
1916 : */
1917 : static void
1918 1238 : SlotSyncShmemRequest(void *arg)
1919 : {
1920 1238 : ShmemRequestStruct(.name = "Slot Sync Data",
1921 : .size = sizeof(SlotSyncCtxStruct),
1922 : .ptr = (void **) &SlotSyncCtx,
1923 : );
1924 1238 : }
1925 :
1926 : /*
1927 : * Initialize shared memory for slot synchronization.
1928 : */
1929 : static void
1930 1235 : SlotSyncShmemInit(void *arg)
1931 : {
1932 1235 : memset(SlotSyncCtx, 0, sizeof(SlotSyncCtxStruct));
1933 1235 : SlotSyncCtx->pid = InvalidPid;
1934 1235 : SpinLockInit(&SlotSyncCtx->mutex);
1935 1235 : }
1936 :
1937 : /*
1938 : * Error cleanup callback for slot sync SQL function.
1939 : */
1940 : static void
1941 1 : slotsync_failure_callback(int code, Datum arg)
1942 : {
1943 1 : WalReceiverConn *wrconn = (WalReceiverConn *) DatumGetPointer(arg);
1944 :
1945 : /*
1946 : * We need to do slots cleanup here just like WalSndErrorCleanup() does.
1947 : *
1948 : * The startup process during promotion invokes ShutDownSlotSync() which
1949 : * waits for slot sync to finish and it does that by checking the
1950 : * 'syncing' flag. Thus the SQL function must be done with slots' release
1951 : * and cleanup to avoid any dangling temporary slots or active slots
1952 : * before it marks itself as finished syncing.
1953 : */
1954 :
1955 : /* Make sure active replication slots are released */
1956 1 : if (MyReplicationSlot != NULL)
1957 0 : ReplicationSlotRelease();
1958 :
1959 : /* Also cleanup the synced temporary slots. */
1960 1 : ReplicationSlotCleanup(true);
1961 :
1962 : /*
1963 : * The set syncing_slots indicates that the process errored out without
1964 : * resetting the flag. So, we need to clean up shared memory and reset the
1965 : * flag here.
1966 : */
1967 1 : if (syncing_slots)
1968 1 : reset_syncing_flag();
1969 :
1970 1 : walrcv_disconnect(wrconn);
1971 1 : }
1972 :
1973 : /*
1974 : * Helper function to extract slot names from a list of remote slots
1975 : */
1976 : static List *
1977 1 : extract_slot_names(List *remote_slots)
1978 : {
1979 1 : List *slot_names = NIL;
1980 :
1981 3 : foreach_ptr(RemoteSlot, remote_slot, remote_slots)
1982 : {
1983 : char *slot_name;
1984 :
1985 1 : slot_name = pstrdup(remote_slot->name);
1986 1 : slot_names = lappend(slot_names, slot_name);
1987 : }
1988 :
1989 1 : return slot_names;
1990 : }
1991 :
1992 : /*
1993 : * Synchronize the failover enabled replication slots using the specified
1994 : * primary server connection.
1995 : *
1996 : * Repeatedly fetches and updates replication slot information from the
1997 : * primary until all slots are at least "sync ready".
1998 : *
1999 : * Exits early if promotion is triggered or certain critical
2000 : * configuration parameters have changed.
2001 : */
2002 : void
2003 9 : SyncReplicationSlots(WalReceiverConn *wrconn)
2004 : {
2005 9 : PG_ENSURE_ERROR_CLEANUP(slotsync_failure_callback, PointerGetDatum(wrconn));
2006 : {
2007 9 : List *remote_slots = NIL;
2008 9 : List *slot_names = NIL; /* List of slot names to track */
2009 :
2010 9 : check_and_set_sync_info(MyProcPid);
2011 :
2012 9 : validate_remote_info(wrconn);
2013 :
2014 : /* Retry until all the slots are sync-ready */
2015 : for (;;)
2016 2 : {
2017 10 : bool slot_persistence_pending = false;
2018 10 : bool some_slot_updated = false;
2019 :
2020 : /* Check for interrupts and config changes */
2021 10 : CHECK_FOR_INTERRUPTS();
2022 :
2023 10 : if (ConfigReloadPending)
2024 0 : slotsync_reread_config();
2025 :
2026 : /* We must be in a valid transaction state */
2027 : Assert(IsTransactionState());
2028 :
2029 : /*
2030 : * Fetch remote slot info for the given slot_names. If slot_names
2031 : * is NIL, fetch all failover-enabled slots. Note that we reuse
2032 : * slot_names from the first iteration; re-fetching all failover
2033 : * slots each time could cause an endless loop. Instead of
2034 : * reprocessing only the pending slots in each iteration, it's
2035 : * better to process all the slots received in the first
2036 : * iteration. This ensures that by the time we're done, all slots
2037 : * reflect the latest values.
2038 : */
2039 10 : remote_slots = fetch_remote_slots(wrconn, slot_names);
2040 :
2041 : /* Attempt to synchronize slots */
2042 10 : some_slot_updated = synchronize_slots(wrconn, remote_slots,
2043 : &slot_persistence_pending);
2044 :
2045 : /*
2046 : * If slot_persistence_pending is true, extract slot names for
2047 : * future iterations (only needed if we haven't done it yet)
2048 : */
2049 10 : if (slot_names == NIL && slot_persistence_pending)
2050 1 : slot_names = extract_slot_names(remote_slots);
2051 :
2052 : /* Free the current remote_slots list */
2053 10 : list_free_deep(remote_slots);
2054 :
2055 : /* Done if all slots are persisted i.e are sync-ready */
2056 10 : if (!slot_persistence_pending)
2057 8 : break;
2058 :
2059 : /* wait before retrying again */
2060 2 : wait_for_slot_activity(some_slot_updated);
2061 : }
2062 :
2063 8 : if (slot_names)
2064 1 : list_free_deep(slot_names);
2065 :
2066 : /* Cleanup the synced temporary slots */
2067 8 : ReplicationSlotCleanup(true);
2068 :
2069 : /* We are done with sync, so reset sync flag */
2070 8 : reset_syncing_flag();
2071 : }
2072 9 : PG_END_ENSURE_ERROR_CLEANUP(slotsync_failure_callback, PointerGetDatum(wrconn));
2073 8 : }
|
