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) 2024-2025, 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 : * The slot sync worker waits for some time before the next synchronization,
38 : * with the duration varying based on whether any slots were updated during
39 : * the last cycle. Refer to the comments above wait_for_slot_activity() for
40 : * more details.
41 : *
42 : * If the SQL function pg_sync_replication_slots() is used to sync the slots,
43 : * and if the slots are not ready to be synced and are marked as RS_TEMPORARY
44 : * because of any of the reasons mentioned above, then the SQL function also
45 : * waits and retries until the slots are marked as RS_PERSISTENT (which means
46 : * sync-ready). Refer to the comments in SyncReplicationSlots() for more
47 : * details.
48 : *
49 : * Any standby synchronized slots will be dropped if they no longer need
50 : * to be synchronized. See comment atop drop_local_obsolete_slots() for more
51 : * details.
52 : *---------------------------------------------------------------------------
53 : */
54 :
55 : #include "postgres.h"
56 :
57 : #include <time.h>
58 :
59 : #include "access/xlog_internal.h"
60 : #include "access/xlogrecovery.h"
61 : #include "catalog/pg_database.h"
62 : #include "libpq/pqsignal.h"
63 : #include "pgstat.h"
64 : #include "postmaster/interrupt.h"
65 : #include "replication/logical.h"
66 : #include "replication/slotsync.h"
67 : #include "replication/snapbuild.h"
68 : #include "storage/ipc.h"
69 : #include "storage/lmgr.h"
70 : #include "storage/proc.h"
71 : #include "storage/procarray.h"
72 : #include "tcop/tcopprot.h"
73 : #include "utils/builtins.h"
74 : #include "utils/memutils.h"
75 : #include "utils/pg_lsn.h"
76 : #include "utils/ps_status.h"
77 : #include "utils/timeout.h"
78 :
79 : /*
80 : * Struct for sharing information to control slot synchronization.
81 : *
82 : * The 'pid' is either the slot sync worker's pid or the backend's pid running
83 : * the SQL function pg_sync_replication_slots(). When the startup process sets
84 : * 'stopSignaled' during promotion, it uses this 'pid' to wake up the currently
85 : * synchronizing process so that the process can immediately stop its
86 : * synchronizing work on seeing 'stopSignaled' set.
87 : * Setting 'stopSignaled' is also used to handle the race condition when the
88 : * postmaster has not noticed the promotion yet and thus may end up restarting
89 : * the slot sync worker. If 'stopSignaled' is set, the worker will exit in such a
90 : * case. The SQL function pg_sync_replication_slots() will also error out if
91 : * this flag is set. Note that we don't need to reset this variable as after
92 : * promotion the slot sync worker won't be restarted because the pmState
93 : * changes to PM_RUN from PM_HOT_STANDBY and we don't support demoting
94 : * primary without restarting the server. See LaunchMissingBackgroundProcesses.
95 : *
96 : * The 'syncing' flag is needed to prevent concurrent slot syncs to avoid slot
97 : * overwrites.
98 : *
99 : * The 'last_start_time' is needed by postmaster to start the slot sync worker
100 : * once per SLOTSYNC_RESTART_INTERVAL_SEC. In cases where an immediate restart
101 : * is expected (e.g., slot sync GUCs change), slot sync worker will reset
102 : * last_start_time before exiting, so that postmaster can start the worker
103 : * without waiting for SLOTSYNC_RESTART_INTERVAL_SEC.
104 : */
105 : typedef struct SlotSyncCtxStruct
106 : {
107 : pid_t pid;
108 : bool stopSignaled;
109 : bool syncing;
110 : time_t last_start_time;
111 : slock_t mutex;
112 : } SlotSyncCtxStruct;
113 :
114 : static SlotSyncCtxStruct *SlotSyncCtx = NULL;
115 :
116 : /* GUC variable */
117 : bool sync_replication_slots = false;
118 :
119 : /*
120 : * The sleep time (ms) between slot-sync cycles varies dynamically
121 : * (within a MIN/MAX range) according to slot activity. See
122 : * wait_for_slot_activity() for details.
123 : */
124 : #define MIN_SLOTSYNC_WORKER_NAPTIME_MS 200
125 : #define MAX_SLOTSYNC_WORKER_NAPTIME_MS 30000 /* 30s */
126 :
127 : static long sleep_ms = MIN_SLOTSYNC_WORKER_NAPTIME_MS;
128 :
129 : /* The restart interval for slot sync work used by postmaster */
130 : #define SLOTSYNC_RESTART_INTERVAL_SEC 10
131 :
132 : /*
133 : * Flag to tell if we are syncing replication slots. Unlike the 'syncing' flag
134 : * in SlotSyncCtxStruct, this flag is true only if the current process is
135 : * performing slot synchronization.
136 : */
137 : static bool syncing_slots = false;
138 :
139 : /*
140 : * Structure to hold information fetched from the primary server about a logical
141 : * replication slot.
142 : */
143 : typedef struct RemoteSlot
144 : {
145 : char *name;
146 : char *plugin;
147 : char *database;
148 : bool two_phase;
149 : bool failover;
150 : XLogRecPtr restart_lsn;
151 : XLogRecPtr confirmed_lsn;
152 : XLogRecPtr two_phase_at;
153 : TransactionId catalog_xmin;
154 :
155 : /* RS_INVAL_NONE if valid, or the reason of invalidation */
156 : ReplicationSlotInvalidationCause invalidated;
157 : } RemoteSlot;
158 :
159 : static void slotsync_failure_callback(int code, Datum arg);
160 : static void update_synced_slots_inactive_since(void);
161 :
162 : /*
163 : * Update slot sync skip stats. This function requires the caller to acquire
164 : * the slot.
165 : */
166 : static void
167 88 : update_slotsync_skip_stats(SlotSyncSkipReason skip_reason)
168 : {
169 : ReplicationSlot *slot;
170 :
171 : Assert(MyReplicationSlot);
172 :
173 88 : slot = MyReplicationSlot;
174 :
175 : /*
176 : * Update the slot sync related stats in pg_stat_replication_slot when a
177 : * slot sync is skipped
178 : */
179 88 : if (skip_reason != SS_SKIP_NONE)
180 4 : pgstat_report_replslotsync(slot);
181 :
182 : /* Update the slot sync skip reason */
183 88 : if (slot->slotsync_skip_reason != skip_reason)
184 : {
185 4 : SpinLockAcquire(&slot->mutex);
186 4 : slot->slotsync_skip_reason = skip_reason;
187 4 : SpinLockRelease(&slot->mutex);
188 : }
189 88 : }
190 :
191 : /*
192 : * If necessary, update the local synced slot's metadata based on the data
193 : * from the remote slot.
194 : *
195 : * If no update was needed (the data of the remote slot is the same as the
196 : * local slot) return false, otherwise true.
197 : *
198 : * *found_consistent_snapshot will be true iff the remote slot's LSN or xmin is
199 : * modified, and decoding from the corresponding LSN's can reach a
200 : * consistent snapshot.
201 : *
202 : * *remote_slot_precedes will be true if the remote slot's LSN or xmin
203 : * precedes locally reserved position.
204 : */
205 : static bool
206 88 : update_local_synced_slot(RemoteSlot *remote_slot, Oid remote_dbid,
207 : bool *found_consistent_snapshot,
208 : bool *remote_slot_precedes)
209 : {
210 88 : ReplicationSlot *slot = MyReplicationSlot;
211 88 : bool updated_xmin_or_lsn = false;
212 88 : bool updated_config = false;
213 88 : SlotSyncSkipReason skip_reason = SS_SKIP_NONE;
214 :
215 : Assert(slot->data.invalidated == RS_INVAL_NONE);
216 :
217 88 : if (found_consistent_snapshot)
218 16 : *found_consistent_snapshot = false;
219 :
220 88 : if (remote_slot_precedes)
221 16 : *remote_slot_precedes = false;
222 :
223 : /*
224 : * Don't overwrite if we already have a newer catalog_xmin and
225 : * restart_lsn.
226 : */
227 176 : if (remote_slot->restart_lsn < slot->data.restart_lsn ||
228 88 : TransactionIdPrecedes(remote_slot->catalog_xmin,
229 : slot->data.catalog_xmin))
230 : {
231 : /* Update slot sync skip stats */
232 4 : update_slotsync_skip_stats(SS_SKIP_WAL_OR_ROWS_REMOVED);
233 :
234 : /*
235 : * This can happen in following situations:
236 : *
237 : * If the slot is temporary, it means either the initial WAL location
238 : * reserved for the local slot is ahead of the remote slot's
239 : * restart_lsn or the initial xmin_horizon computed for the local slot
240 : * is ahead of the remote slot.
241 : *
242 : * If the slot is persistent, both restart_lsn and catalog_xmin of the
243 : * synced slot could still be ahead of the remote slot. Since we use
244 : * slot advance functionality to keep snapbuild/slot updated, it is
245 : * possible that the restart_lsn and catalog_xmin are advanced to a
246 : * later position than it has on the primary. This can happen when
247 : * slot advancing machinery finds running xacts record after reaching
248 : * the consistent state at a later point than the primary where it
249 : * serializes the snapshot and updates the restart_lsn.
250 : *
251 : * We LOG the message if the slot is temporary as it can help the user
252 : * to understand why the slot is not sync-ready. In the case of a
253 : * persistent slot, it would be a more common case and won't directly
254 : * impact the users, so we used DEBUG1 level to log the message.
255 : */
256 4 : ereport(slot->data.persistency == RS_TEMPORARY ? LOG : DEBUG1,
257 : errmsg("could not synchronize replication slot \"%s\"",
258 : remote_slot->name),
259 : 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.",
260 : LSN_FORMAT_ARGS(remote_slot->restart_lsn),
261 : remote_slot->catalog_xmin,
262 : LSN_FORMAT_ARGS(slot->data.restart_lsn),
263 : slot->data.catalog_xmin));
264 :
265 4 : if (remote_slot_precedes)
266 4 : *remote_slot_precedes = true;
267 :
268 : /*
269 : * Skip updating the configuration. This is required to avoid syncing
270 : * two_phase_at without syncing confirmed_lsn. Otherwise, the prepared
271 : * transaction between old confirmed_lsn and two_phase_at will
272 : * unexpectedly get decoded and sent to the downstream after
273 : * promotion. See comments in ReorderBufferFinishPrepared.
274 : */
275 4 : return false;
276 : }
277 :
278 : /*
279 : * Attempt to sync LSNs and xmins only if remote slot is ahead of local
280 : * slot.
281 : */
282 84 : if (remote_slot->confirmed_lsn > slot->data.confirmed_flush ||
283 126 : remote_slot->restart_lsn > slot->data.restart_lsn ||
284 62 : TransactionIdFollows(remote_slot->catalog_xmin,
285 : slot->data.catalog_xmin))
286 : {
287 : /*
288 : * We can't directly copy the remote slot's LSN or xmin unless there
289 : * exists a consistent snapshot at that point. Otherwise, after
290 : * promotion, the slots may not reach a consistent point before the
291 : * confirmed_flush_lsn which can lead to a data loss. To avoid data
292 : * loss, we let slot machinery advance the slot which ensures that
293 : * snapbuilder/slot statuses are updated properly.
294 : */
295 22 : if (SnapBuildSnapshotExists(remote_slot->restart_lsn))
296 : {
297 : /*
298 : * Update the slot info directly if there is a serialized snapshot
299 : * at the restart_lsn, as the slot can quickly reach consistency
300 : * at restart_lsn by restoring the snapshot.
301 : */
302 2 : SpinLockAcquire(&slot->mutex);
303 2 : slot->data.restart_lsn = remote_slot->restart_lsn;
304 2 : slot->data.confirmed_flush = remote_slot->confirmed_lsn;
305 2 : slot->data.catalog_xmin = remote_slot->catalog_xmin;
306 2 : SpinLockRelease(&slot->mutex);
307 :
308 2 : if (found_consistent_snapshot)
309 0 : *found_consistent_snapshot = true;
310 : }
311 : else
312 : {
313 20 : LogicalSlotAdvanceAndCheckSnapState(remote_slot->confirmed_lsn,
314 : found_consistent_snapshot);
315 :
316 : /* Sanity check */
317 20 : if (slot->data.confirmed_flush != remote_slot->confirmed_lsn)
318 0 : ereport(ERROR,
319 : errmsg_internal("synchronized confirmed_flush for slot \"%s\" differs from remote slot",
320 : remote_slot->name),
321 : errdetail_internal("Remote slot has LSN %X/%08X but local slot has LSN %X/%08X.",
322 : LSN_FORMAT_ARGS(remote_slot->confirmed_lsn),
323 : LSN_FORMAT_ARGS(slot->data.confirmed_flush)));
324 :
325 : /*
326 : * If we can't reach a consistent snapshot, the slot won't be
327 : * persisted. See update_and_persist_local_synced_slot().
328 : */
329 20 : if (found_consistent_snapshot && !(*found_consistent_snapshot))
330 0 : skip_reason = SS_SKIP_NO_CONSISTENT_SNAPSHOT;
331 : }
332 :
333 22 : updated_xmin_or_lsn = true;
334 : }
335 :
336 : /* Update slot sync skip stats */
337 84 : update_slotsync_skip_stats(skip_reason);
338 :
339 84 : if (remote_dbid != slot->data.database ||
340 84 : remote_slot->two_phase != slot->data.two_phase ||
341 82 : remote_slot->failover != slot->data.failover ||
342 82 : strcmp(remote_slot->plugin, NameStr(slot->data.plugin)) != 0 ||
343 82 : remote_slot->two_phase_at != slot->data.two_phase_at)
344 : {
345 : NameData plugin_name;
346 :
347 : /* Avoid expensive operations while holding a spinlock. */
348 2 : namestrcpy(&plugin_name, remote_slot->plugin);
349 :
350 2 : SpinLockAcquire(&slot->mutex);
351 2 : slot->data.plugin = plugin_name;
352 2 : slot->data.database = remote_dbid;
353 2 : slot->data.two_phase = remote_slot->two_phase;
354 2 : slot->data.two_phase_at = remote_slot->two_phase_at;
355 2 : slot->data.failover = remote_slot->failover;
356 2 : SpinLockRelease(&slot->mutex);
357 :
358 2 : updated_config = true;
359 :
360 : /*
361 : * Ensure that there is no risk of sending prepared transactions
362 : * unexpectedly after the promotion.
363 : */
364 : Assert(slot->data.two_phase_at <= slot->data.confirmed_flush);
365 : }
366 :
367 : /*
368 : * We have to write the changed xmin to disk *before* we change the
369 : * in-memory value, otherwise after a crash we wouldn't know that some
370 : * catalog tuples might have been removed already.
371 : */
372 84 : if (updated_config || updated_xmin_or_lsn)
373 : {
374 24 : ReplicationSlotMarkDirty();
375 24 : ReplicationSlotSave();
376 : }
377 :
378 : /*
379 : * Now the new xmin is safely on disk, we can let the global value
380 : * advance. We do not take ProcArrayLock or similar since we only advance
381 : * xmin here and there's not much harm done by a concurrent computation
382 : * missing that.
383 : */
384 84 : if (updated_xmin_or_lsn)
385 : {
386 22 : SpinLockAcquire(&slot->mutex);
387 22 : slot->effective_catalog_xmin = remote_slot->catalog_xmin;
388 22 : SpinLockRelease(&slot->mutex);
389 :
390 22 : ReplicationSlotsComputeRequiredXmin(false);
391 22 : ReplicationSlotsComputeRequiredLSN();
392 : }
393 :
394 84 : return updated_config || updated_xmin_or_lsn;
395 : }
396 :
397 : /*
398 : * Get the list of local logical slots that are synchronized from the
399 : * primary server.
400 : */
401 : static List *
402 50 : get_local_synced_slots(void)
403 : {
404 50 : List *local_slots = NIL;
405 :
406 50 : LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
407 :
408 550 : for (int i = 0; i < max_replication_slots; i++)
409 : {
410 500 : ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
411 :
412 : /* Check if it is a synchronized slot */
413 500 : if (s->in_use && s->data.synced)
414 : {
415 : Assert(SlotIsLogical(s));
416 80 : local_slots = lappend(local_slots, s);
417 : }
418 : }
419 :
420 50 : LWLockRelease(ReplicationSlotControlLock);
421 :
422 50 : return local_slots;
423 : }
424 :
425 : /*
426 : * Helper function to check if local_slot is required to be retained.
427 : *
428 : * Return false either if local_slot does not exist in the remote_slots list
429 : * or is invalidated while the corresponding remote slot is still valid,
430 : * otherwise true.
431 : */
432 : static bool
433 80 : local_sync_slot_required(ReplicationSlot *local_slot, List *remote_slots)
434 : {
435 80 : bool remote_exists = false;
436 80 : bool locally_invalidated = false;
437 :
438 196 : foreach_ptr(RemoteSlot, remote_slot, remote_slots)
439 : {
440 114 : if (strcmp(remote_slot->name, NameStr(local_slot->data.name)) == 0)
441 : {
442 78 : remote_exists = true;
443 :
444 : /*
445 : * If remote slot is not invalidated but local slot is marked as
446 : * invalidated, then set locally_invalidated flag.
447 : */
448 78 : SpinLockAcquire(&local_slot->mutex);
449 78 : locally_invalidated =
450 156 : (remote_slot->invalidated == RS_INVAL_NONE) &&
451 78 : (local_slot->data.invalidated != RS_INVAL_NONE);
452 78 : SpinLockRelease(&local_slot->mutex);
453 :
454 78 : break;
455 : }
456 : }
457 :
458 80 : return (remote_exists && !locally_invalidated);
459 : }
460 :
461 : /*
462 : * Drop local obsolete slots.
463 : *
464 : * Drop the local slots that no longer need to be synced i.e. these either do
465 : * not exist on the primary or are no longer enabled for failover.
466 : *
467 : * Additionally, drop any slots that are valid on the primary but got
468 : * invalidated on the standby. This situation may occur due to the following
469 : * reasons:
470 : * - The 'max_slot_wal_keep_size' on the standby is insufficient to retain WAL
471 : * records from the restart_lsn of the slot.
472 : * - 'primary_slot_name' is temporarily reset to null and the physical slot is
473 : * removed.
474 : * These dropped slots will get recreated in next sync-cycle and it is okay to
475 : * drop and recreate such slots as long as these are not consumable on the
476 : * standby (which is the case currently).
477 : *
478 : * Note: Change of 'wal_level' on the primary server to a level lower than
479 : * logical may also result in slot invalidation and removal on the standby.
480 : * This is because such 'wal_level' change is only possible if the logical
481 : * slots are removed on the primary server, so it's expected to see the
482 : * slots being invalidated and removed on the standby too (and re-created
483 : * if they are re-created on the primary server).
484 : */
485 : static void
486 50 : drop_local_obsolete_slots(List *remote_slot_list)
487 : {
488 50 : List *local_slots = get_local_synced_slots();
489 :
490 180 : foreach_ptr(ReplicationSlot, local_slot, local_slots)
491 : {
492 : /* Drop the local slot if it is not required to be retained. */
493 80 : if (!local_sync_slot_required(local_slot, remote_slot_list))
494 : {
495 : bool synced_slot;
496 :
497 : /*
498 : * Use shared lock to prevent a conflict with
499 : * ReplicationSlotsDropDBSlots(), trying to drop the same slot
500 : * during a drop-database operation.
501 : */
502 4 : LockSharedObject(DatabaseRelationId, local_slot->data.database,
503 : 0, AccessShareLock);
504 :
505 : /*
506 : * In the small window between getting the slot to drop and
507 : * locking the database, there is a possibility of a parallel
508 : * database drop by the startup process and the creation of a new
509 : * slot by the user. This new user-created slot may end up using
510 : * the same shared memory as that of 'local_slot'. Thus check if
511 : * local_slot is still the synced one before performing actual
512 : * drop.
513 : */
514 4 : SpinLockAcquire(&local_slot->mutex);
515 4 : synced_slot = local_slot->in_use && local_slot->data.synced;
516 4 : SpinLockRelease(&local_slot->mutex);
517 :
518 4 : if (synced_slot)
519 : {
520 4 : ReplicationSlotAcquire(NameStr(local_slot->data.name), true, false);
521 4 : ReplicationSlotDropAcquired();
522 : }
523 :
524 4 : UnlockSharedObject(DatabaseRelationId, local_slot->data.database,
525 : 0, AccessShareLock);
526 :
527 4 : ereport(LOG,
528 : errmsg("dropped replication slot \"%s\" of database with OID %u",
529 : NameStr(local_slot->data.name),
530 : local_slot->data.database));
531 : }
532 : }
533 50 : }
534 :
535 : /*
536 : * Reserve WAL for the currently active local slot using the specified WAL
537 : * location (restart_lsn).
538 : *
539 : * If the given WAL location has been removed, reserve WAL using the oldest
540 : * existing WAL segment.
541 : */
542 : static void
543 12 : reserve_wal_for_local_slot(XLogRecPtr restart_lsn)
544 : {
545 : XLogSegNo oldest_segno;
546 : XLogSegNo segno;
547 12 : ReplicationSlot *slot = MyReplicationSlot;
548 :
549 : Assert(slot != NULL);
550 : Assert(!XLogRecPtrIsValid(slot->data.restart_lsn));
551 :
552 : while (true)
553 : {
554 12 : SpinLockAcquire(&slot->mutex);
555 12 : slot->data.restart_lsn = restart_lsn;
556 12 : SpinLockRelease(&slot->mutex);
557 :
558 : /* Prevent WAL removal as fast as possible */
559 12 : ReplicationSlotsComputeRequiredLSN();
560 :
561 12 : XLByteToSeg(slot->data.restart_lsn, segno, wal_segment_size);
562 :
563 : /*
564 : * Find the oldest existing WAL segment file.
565 : *
566 : * Normally, we can determine it by using the last removed segment
567 : * number. However, if no WAL segment files have been removed by a
568 : * checkpoint since startup, we need to search for the oldest segment
569 : * file from the current timeline existing in XLOGDIR.
570 : *
571 : * XXX: Currently, we are searching for the oldest segment in the
572 : * current timeline as there is less chance of the slot's restart_lsn
573 : * from being some prior timeline, and even if it happens, in the
574 : * worst case, we will wait to sync till the slot's restart_lsn moved
575 : * to the current timeline.
576 : */
577 12 : oldest_segno = XLogGetLastRemovedSegno() + 1;
578 :
579 12 : if (oldest_segno == 1)
580 : {
581 : TimeLineID cur_timeline;
582 :
583 8 : GetWalRcvFlushRecPtr(NULL, &cur_timeline);
584 8 : oldest_segno = XLogGetOldestSegno(cur_timeline);
585 : }
586 :
587 12 : elog(DEBUG1, "segno: " UINT64_FORMAT " of purposed restart_lsn for the synced slot, oldest_segno: " UINT64_FORMAT " available",
588 : segno, oldest_segno);
589 :
590 : /*
591 : * If all required WAL is still there, great, otherwise retry. The
592 : * slot should prevent further removal of WAL, unless there's a
593 : * concurrent ReplicationSlotsComputeRequiredLSN() after we've written
594 : * the new restart_lsn above, so normally we should never need to loop
595 : * more than twice.
596 : */
597 12 : if (segno >= oldest_segno)
598 12 : break;
599 :
600 : /* Retry using the location of the oldest wal segment */
601 0 : XLogSegNoOffsetToRecPtr(oldest_segno, 0, wal_segment_size, restart_lsn);
602 : }
603 12 : }
604 :
605 : /*
606 : * If the remote restart_lsn and catalog_xmin have caught up with the
607 : * local ones, then update the LSNs and persist the local synced slot for
608 : * future synchronization; otherwise, do nothing.
609 : *
610 : * *slot_persistence_pending is set to true if any of the slots fail to
611 : * persist.
612 : *
613 : * Return true if the slot is marked as RS_PERSISTENT (sync-ready), otherwise
614 : * false.
615 : */
616 : static bool
617 16 : update_and_persist_local_synced_slot(RemoteSlot *remote_slot, Oid remote_dbid,
618 : bool *slot_persistence_pending)
619 : {
620 16 : ReplicationSlot *slot = MyReplicationSlot;
621 16 : bool found_consistent_snapshot = false;
622 16 : bool remote_slot_precedes = false;
623 :
624 : /* Slotsync skip stats are handled in function update_local_synced_slot() */
625 16 : (void) update_local_synced_slot(remote_slot, remote_dbid,
626 : &found_consistent_snapshot,
627 : &remote_slot_precedes);
628 :
629 : /*
630 : * Check if the primary server has caught up. Refer to the comment atop
631 : * the file for details on this check.
632 : */
633 16 : if (remote_slot_precedes)
634 : {
635 : /*
636 : * The remote slot didn't catch up to locally reserved position.
637 : *
638 : * We do not drop the slot because the restart_lsn can be ahead of the
639 : * current location when recreating the slot in the next cycle. It may
640 : * take more time to create such a slot. Therefore, we keep this slot
641 : * and attempt the synchronization in the next cycle.
642 : *
643 : * We also update the slot_persistence_pending parameter, so the SQL
644 : * function can retry.
645 : */
646 4 : if (slot_persistence_pending)
647 4 : *slot_persistence_pending = true;
648 :
649 4 : return false;
650 : }
651 :
652 : /*
653 : * Don't persist the slot if it cannot reach the consistent point from the
654 : * restart_lsn. See comments atop this file.
655 : */
656 12 : if (!found_consistent_snapshot)
657 : {
658 0 : ereport(LOG,
659 : errmsg("could not synchronize replication slot \"%s\"", remote_slot->name),
660 : errdetail("Synchronization could lead to data loss, because the standby could not build a consistent snapshot to decode WALs at LSN %X/%08X.",
661 : LSN_FORMAT_ARGS(slot->data.restart_lsn)));
662 :
663 : /* Set this, so that SQL function can retry */
664 0 : if (slot_persistence_pending)
665 0 : *slot_persistence_pending = true;
666 :
667 0 : return false;
668 : }
669 :
670 12 : ReplicationSlotPersist();
671 :
672 12 : ereport(LOG,
673 : errmsg("newly created replication slot \"%s\" is sync-ready now",
674 : remote_slot->name));
675 :
676 12 : return true;
677 : }
678 :
679 : /*
680 : * Synchronize a single slot to the given position.
681 : *
682 : * This creates a new slot if there is no existing one and updates the
683 : * metadata of the slot as per the data received from the primary server.
684 : *
685 : * The slot is created as a temporary slot and stays in the same state until the
686 : * remote_slot catches up with locally reserved position and local slot is
687 : * updated. The slot is then persisted and is considered as sync-ready for
688 : * periodic syncs.
689 : *
690 : * *slot_persistence_pending is set to true if any of the slots fail to
691 : * persist.
692 : *
693 : * Returns TRUE if the local slot is updated.
694 : */
695 : static bool
696 88 : synchronize_one_slot(RemoteSlot *remote_slot, Oid remote_dbid,
697 : bool *slot_persistence_pending)
698 : {
699 : ReplicationSlot *slot;
700 88 : XLogRecPtr latestFlushPtr = GetStandbyFlushRecPtr(NULL);
701 88 : bool slot_updated = false;
702 :
703 : /* Search for the named slot */
704 88 : if ((slot = SearchNamedReplicationSlot(remote_slot->name, true)))
705 : {
706 : bool synced;
707 :
708 76 : SpinLockAcquire(&slot->mutex);
709 76 : synced = slot->data.synced;
710 76 : SpinLockRelease(&slot->mutex);
711 :
712 : /* User-created slot with the same name exists, raise ERROR. */
713 76 : if (!synced)
714 0 : ereport(ERROR,
715 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
716 : errmsg("exiting from slot synchronization because same"
717 : " name slot \"%s\" already exists on the standby",
718 : remote_slot->name));
719 :
720 : /*
721 : * The slot has been synchronized before.
722 : *
723 : * It is important to acquire the slot here before checking
724 : * invalidation. If we don't acquire the slot first, there could be a
725 : * race condition that the local slot could be invalidated just after
726 : * checking the 'invalidated' flag here and we could end up
727 : * overwriting 'invalidated' flag to remote_slot's value. See
728 : * InvalidatePossiblyObsoleteSlot() where it invalidates slot directly
729 : * if the slot is not acquired by other processes.
730 : *
731 : * XXX: If it ever turns out that slot acquire/release is costly for
732 : * cases when none of the slot properties is changed then we can do a
733 : * pre-check to ensure that at least one of the slot properties is
734 : * changed before acquiring the slot.
735 : */
736 76 : ReplicationSlotAcquire(remote_slot->name, true, false);
737 :
738 : Assert(slot == MyReplicationSlot);
739 :
740 : /*
741 : * Copy the invalidation cause from remote only if local slot is not
742 : * invalidated locally, we don't want to overwrite existing one.
743 : */
744 76 : if (slot->data.invalidated == RS_INVAL_NONE &&
745 76 : remote_slot->invalidated != RS_INVAL_NONE)
746 : {
747 0 : SpinLockAcquire(&slot->mutex);
748 0 : slot->data.invalidated = remote_slot->invalidated;
749 0 : SpinLockRelease(&slot->mutex);
750 :
751 : /* Make sure the invalidated state persists across server restart */
752 0 : ReplicationSlotMarkDirty();
753 0 : ReplicationSlotSave();
754 :
755 0 : slot_updated = true;
756 : }
757 :
758 : /* Skip the sync of an invalidated slot */
759 76 : if (slot->data.invalidated != RS_INVAL_NONE)
760 : {
761 0 : update_slotsync_skip_stats(SS_SKIP_INVALID);
762 :
763 0 : ReplicationSlotRelease();
764 0 : return slot_updated;
765 : }
766 :
767 : /*
768 : * Make sure that concerned WAL is received and flushed before syncing
769 : * slot to target lsn received from the primary server.
770 : *
771 : * Report statistics only after the slot has been acquired, ensuring
772 : * it cannot be dropped during the reporting process.
773 : */
774 76 : if (remote_slot->confirmed_lsn > latestFlushPtr)
775 : {
776 0 : update_slotsync_skip_stats(SS_SKIP_WAL_NOT_FLUSHED);
777 :
778 : /*
779 : * Can get here only if GUC 'synchronized_standby_slots' on the
780 : * primary server was not configured correctly.
781 : */
782 0 : ereport(AmLogicalSlotSyncWorkerProcess() ? LOG : ERROR,
783 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
784 : errmsg("skipping slot synchronization because the received slot sync"
785 : " LSN %X/%08X for slot \"%s\" is ahead of the standby position %X/%08X",
786 : LSN_FORMAT_ARGS(remote_slot->confirmed_lsn),
787 : remote_slot->name,
788 : LSN_FORMAT_ARGS(latestFlushPtr)));
789 :
790 0 : ReplicationSlotRelease();
791 :
792 0 : return slot_updated;
793 : }
794 :
795 : /* Slot not ready yet, let's attempt to make it sync-ready now. */
796 76 : if (slot->data.persistency == RS_TEMPORARY)
797 : {
798 4 : slot_updated = update_and_persist_local_synced_slot(remote_slot,
799 : remote_dbid,
800 : slot_persistence_pending);
801 : }
802 :
803 : /* Slot ready for sync, so sync it. */
804 : else
805 : {
806 : /*
807 : * Sanity check: As long as the invalidations are handled
808 : * appropriately as above, this should never happen.
809 : *
810 : * We don't need to check restart_lsn here. See the comments in
811 : * update_local_synced_slot() for details.
812 : */
813 72 : if (remote_slot->confirmed_lsn < slot->data.confirmed_flush)
814 0 : ereport(ERROR,
815 : errmsg_internal("cannot synchronize local slot \"%s\"",
816 : remote_slot->name),
817 : errdetail_internal("Local slot's start streaming location LSN(%X/%08X) is ahead of remote slot's LSN(%X/%08X).",
818 : LSN_FORMAT_ARGS(slot->data.confirmed_flush),
819 : LSN_FORMAT_ARGS(remote_slot->confirmed_lsn)));
820 :
821 72 : slot_updated = update_local_synced_slot(remote_slot, remote_dbid,
822 : NULL, NULL);
823 : }
824 : }
825 : /* Otherwise create the slot first. */
826 : else
827 : {
828 : NameData plugin_name;
829 12 : TransactionId xmin_horizon = InvalidTransactionId;
830 :
831 : /* Skip creating the local slot if remote_slot is invalidated already */
832 12 : if (remote_slot->invalidated != RS_INVAL_NONE)
833 0 : return false;
834 :
835 : /*
836 : * We create temporary slots instead of ephemeral slots here because
837 : * we want the slots to survive after releasing them. This is done to
838 : * avoid dropping and re-creating the slots in each synchronization
839 : * cycle if the restart_lsn or catalog_xmin of the remote slot has not
840 : * caught up.
841 : */
842 12 : ReplicationSlotCreate(remote_slot->name, true, RS_TEMPORARY,
843 12 : remote_slot->two_phase,
844 12 : remote_slot->failover,
845 : true);
846 :
847 : /* For shorter lines. */
848 12 : slot = MyReplicationSlot;
849 :
850 : /* Avoid expensive operations while holding a spinlock. */
851 12 : namestrcpy(&plugin_name, remote_slot->plugin);
852 :
853 12 : SpinLockAcquire(&slot->mutex);
854 12 : slot->data.database = remote_dbid;
855 12 : slot->data.plugin = plugin_name;
856 12 : SpinLockRelease(&slot->mutex);
857 :
858 12 : reserve_wal_for_local_slot(remote_slot->restart_lsn);
859 :
860 12 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
861 12 : xmin_horizon = GetOldestSafeDecodingTransactionId(true);
862 12 : SpinLockAcquire(&slot->mutex);
863 12 : slot->effective_catalog_xmin = xmin_horizon;
864 12 : slot->data.catalog_xmin = xmin_horizon;
865 12 : SpinLockRelease(&slot->mutex);
866 12 : ReplicationSlotsComputeRequiredXmin(true);
867 12 : LWLockRelease(ProcArrayLock);
868 :
869 : /*
870 : * Make sure that concerned WAL is received and flushed before syncing
871 : * slot to target lsn received from the primary server.
872 : *
873 : * Report statistics only after the slot has been acquired, ensuring
874 : * it cannot be dropped during the reporting process.
875 : */
876 12 : if (remote_slot->confirmed_lsn > latestFlushPtr)
877 : {
878 0 : update_slotsync_skip_stats(SS_SKIP_WAL_NOT_FLUSHED);
879 :
880 : /*
881 : * Can get here only if GUC 'synchronized_standby_slots' on the
882 : * primary server was not configured correctly.
883 : */
884 0 : ereport(AmLogicalSlotSyncWorkerProcess() ? LOG : ERROR,
885 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
886 : errmsg("skipping slot synchronization because the received slot sync"
887 : " LSN %X/%08X for slot \"%s\" is ahead of the standby position %X/%08X",
888 : LSN_FORMAT_ARGS(remote_slot->confirmed_lsn),
889 : remote_slot->name,
890 : LSN_FORMAT_ARGS(latestFlushPtr)));
891 :
892 0 : ReplicationSlotRelease();
893 :
894 0 : return false;
895 : }
896 :
897 12 : update_and_persist_local_synced_slot(remote_slot, remote_dbid,
898 : slot_persistence_pending);
899 :
900 12 : slot_updated = true;
901 : }
902 :
903 88 : ReplicationSlotRelease();
904 :
905 88 : return slot_updated;
906 : }
907 :
908 : /*
909 : * Fetch remote slots.
910 : *
911 : * If slot_names is NIL, fetches all failover logical slots from the
912 : * primary server, otherwise fetches only the ones with names in slot_names.
913 : *
914 : * Returns a list of remote slot information structures, or NIL if none
915 : * are found.
916 : */
917 : static List *
918 50 : fetch_remote_slots(WalReceiverConn *wrconn, List *slot_names)
919 : {
920 : #define SLOTSYNC_COLUMN_COUNT 10
921 50 : Oid slotRow[SLOTSYNC_COLUMN_COUNT] = {TEXTOID, TEXTOID, LSNOID,
922 : LSNOID, XIDOID, BOOLOID, LSNOID, BOOLOID, TEXTOID, TEXTOID};
923 :
924 : WalRcvExecResult *res;
925 : TupleTableSlot *tupslot;
926 50 : List *remote_slot_list = NIL;
927 : StringInfoData query;
928 :
929 50 : initStringInfo(&query);
930 50 : appendStringInfoString(&query,
931 : "SELECT slot_name, plugin, confirmed_flush_lsn,"
932 : " restart_lsn, catalog_xmin, two_phase,"
933 : " two_phase_at, failover,"
934 : " database, invalidation_reason"
935 : " FROM pg_catalog.pg_replication_slots"
936 : " WHERE failover and NOT temporary");
937 :
938 50 : if (slot_names != NIL)
939 : {
940 4 : bool first_slot = true;
941 :
942 : /*
943 : * Construct the query to fetch only the specified slots
944 : */
945 4 : appendStringInfoString(&query, " AND slot_name IN (");
946 :
947 12 : foreach_ptr(char, slot_name, slot_names)
948 : {
949 4 : if (!first_slot)
950 0 : appendStringInfoString(&query, ", ");
951 :
952 4 : appendStringInfo(&query, "%s", quote_literal_cstr(slot_name));
953 4 : first_slot = false;
954 : }
955 4 : appendStringInfoChar(&query, ')');
956 : }
957 :
958 : /* Execute the query */
959 50 : res = walrcv_exec(wrconn, query.data, SLOTSYNC_COLUMN_COUNT, slotRow);
960 50 : pfree(query.data);
961 50 : if (res->status != WALRCV_OK_TUPLES)
962 0 : ereport(ERROR,
963 : errmsg("could not fetch failover logical slots info from the primary server: %s",
964 : res->err));
965 :
966 50 : tupslot = MakeSingleTupleTableSlot(res->tupledesc, &TTSOpsMinimalTuple);
967 138 : while (tuplestore_gettupleslot(res->tuplestore, true, false, tupslot))
968 : {
969 : bool isnull;
970 88 : RemoteSlot *remote_slot = palloc0_object(RemoteSlot);
971 : Datum d;
972 88 : int col = 0;
973 :
974 88 : remote_slot->name = TextDatumGetCString(slot_getattr(tupslot, ++col,
975 : &isnull));
976 : Assert(!isnull);
977 :
978 88 : remote_slot->plugin = TextDatumGetCString(slot_getattr(tupslot, ++col,
979 : &isnull));
980 : Assert(!isnull);
981 :
982 : /*
983 : * It is possible to get null values for LSN and Xmin if slot is
984 : * invalidated on the primary server, so handle accordingly.
985 : */
986 88 : d = slot_getattr(tupslot, ++col, &isnull);
987 88 : remote_slot->confirmed_lsn = isnull ? InvalidXLogRecPtr :
988 88 : DatumGetLSN(d);
989 :
990 88 : d = slot_getattr(tupslot, ++col, &isnull);
991 88 : remote_slot->restart_lsn = isnull ? InvalidXLogRecPtr : DatumGetLSN(d);
992 :
993 88 : d = slot_getattr(tupslot, ++col, &isnull);
994 88 : remote_slot->catalog_xmin = isnull ? InvalidTransactionId :
995 88 : DatumGetTransactionId(d);
996 :
997 88 : remote_slot->two_phase = DatumGetBool(slot_getattr(tupslot, ++col,
998 : &isnull));
999 : Assert(!isnull);
1000 :
1001 88 : d = slot_getattr(tupslot, ++col, &isnull);
1002 88 : remote_slot->two_phase_at = isnull ? InvalidXLogRecPtr : DatumGetLSN(d);
1003 :
1004 88 : remote_slot->failover = DatumGetBool(slot_getattr(tupslot, ++col,
1005 : &isnull));
1006 : Assert(!isnull);
1007 :
1008 88 : remote_slot->database = TextDatumGetCString(slot_getattr(tupslot,
1009 : ++col, &isnull));
1010 : Assert(!isnull);
1011 :
1012 88 : d = slot_getattr(tupslot, ++col, &isnull);
1013 88 : remote_slot->invalidated = isnull ? RS_INVAL_NONE :
1014 0 : GetSlotInvalidationCause(TextDatumGetCString(d));
1015 :
1016 : /* Sanity check */
1017 : Assert(col == SLOTSYNC_COLUMN_COUNT);
1018 :
1019 : /*
1020 : * If restart_lsn, confirmed_lsn or catalog_xmin is invalid but the
1021 : * slot is valid, that means we have fetched the remote_slot in its
1022 : * RS_EPHEMERAL state. In such a case, don't sync it; we can always
1023 : * sync it in the next sync cycle when the remote_slot is persisted
1024 : * and has valid lsn(s) and xmin values.
1025 : *
1026 : * XXX: In future, if we plan to expose 'slot->data.persistency' in
1027 : * pg_replication_slots view, then we can avoid fetching RS_EPHEMERAL
1028 : * slots in the first place.
1029 : */
1030 88 : if ((!XLogRecPtrIsValid(remote_slot->restart_lsn) ||
1031 88 : !XLogRecPtrIsValid(remote_slot->confirmed_lsn) ||
1032 88 : !TransactionIdIsValid(remote_slot->catalog_xmin)) &&
1033 0 : remote_slot->invalidated == RS_INVAL_NONE)
1034 0 : pfree(remote_slot);
1035 : else
1036 : /* Create list of remote slots */
1037 88 : remote_slot_list = lappend(remote_slot_list, remote_slot);
1038 :
1039 88 : ExecClearTuple(tupslot);
1040 : }
1041 :
1042 50 : walrcv_clear_result(res);
1043 :
1044 50 : return remote_slot_list;
1045 : }
1046 :
1047 : /*
1048 : * Synchronize slots.
1049 : *
1050 : * This function takes a list of remote slots and synchronizes them locally. It
1051 : * creates the slots if not present on the standby and updates existing ones.
1052 : *
1053 : * If slot_persistence_pending is not NULL, it will be set to true if one or
1054 : * more slots could not be persisted. This allows callers such as
1055 : * SyncReplicationSlots() to retry those slots.
1056 : *
1057 : * Returns TRUE if any of the slots gets updated in this sync-cycle.
1058 : */
1059 : static bool
1060 50 : synchronize_slots(WalReceiverConn *wrconn, List *remote_slot_list,
1061 : bool *slot_persistence_pending)
1062 : {
1063 50 : bool some_slot_updated = false;
1064 :
1065 : /* Drop local slots that no longer need to be synced. */
1066 50 : drop_local_obsolete_slots(remote_slot_list);
1067 :
1068 : /* Now sync the slots locally */
1069 188 : foreach_ptr(RemoteSlot, remote_slot, remote_slot_list)
1070 : {
1071 88 : Oid remote_dbid = get_database_oid(remote_slot->database, false);
1072 :
1073 : /*
1074 : * Use shared lock to prevent a conflict with
1075 : * ReplicationSlotsDropDBSlots(), trying to drop the same slot during
1076 : * a drop-database operation.
1077 : */
1078 88 : LockSharedObject(DatabaseRelationId, remote_dbid, 0, AccessShareLock);
1079 :
1080 88 : some_slot_updated |= synchronize_one_slot(remote_slot, remote_dbid,
1081 : slot_persistence_pending);
1082 :
1083 88 : UnlockSharedObject(DatabaseRelationId, remote_dbid, 0, AccessShareLock);
1084 : }
1085 :
1086 50 : return some_slot_updated;
1087 : }
1088 :
1089 : /*
1090 : * Checks the remote server info.
1091 : *
1092 : * We ensure that the 'primary_slot_name' exists on the remote server and the
1093 : * remote server is not a standby node.
1094 : */
1095 : static void
1096 26 : validate_remote_info(WalReceiverConn *wrconn)
1097 : {
1098 : #define PRIMARY_INFO_OUTPUT_COL_COUNT 2
1099 : WalRcvExecResult *res;
1100 26 : Oid slotRow[PRIMARY_INFO_OUTPUT_COL_COUNT] = {BOOLOID, BOOLOID};
1101 : StringInfoData cmd;
1102 : bool isnull;
1103 : TupleTableSlot *tupslot;
1104 : bool remote_in_recovery;
1105 : bool primary_slot_valid;
1106 26 : bool started_tx = false;
1107 :
1108 26 : initStringInfo(&cmd);
1109 26 : appendStringInfo(&cmd,
1110 : "SELECT pg_is_in_recovery(), count(*) = 1"
1111 : " FROM pg_catalog.pg_replication_slots"
1112 : " WHERE slot_type='physical' AND slot_name=%s",
1113 : quote_literal_cstr(PrimarySlotName));
1114 :
1115 : /* The syscache access in walrcv_exec() needs a transaction env. */
1116 26 : if (!IsTransactionState())
1117 : {
1118 8 : StartTransactionCommand();
1119 8 : started_tx = true;
1120 : }
1121 :
1122 26 : res = walrcv_exec(wrconn, cmd.data, PRIMARY_INFO_OUTPUT_COL_COUNT, slotRow);
1123 26 : pfree(cmd.data);
1124 :
1125 26 : if (res->status != WALRCV_OK_TUPLES)
1126 0 : ereport(ERROR,
1127 : errmsg("could not fetch primary slot name \"%s\" info from the primary server: %s",
1128 : PrimarySlotName, res->err),
1129 : errhint("Check if \"primary_slot_name\" is configured correctly."));
1130 :
1131 26 : tupslot = MakeSingleTupleTableSlot(res->tupledesc, &TTSOpsMinimalTuple);
1132 26 : if (!tuplestore_gettupleslot(res->tuplestore, true, false, tupslot))
1133 0 : elog(ERROR,
1134 : "failed to fetch tuple for the primary server slot specified by \"primary_slot_name\"");
1135 :
1136 26 : remote_in_recovery = DatumGetBool(slot_getattr(tupslot, 1, &isnull));
1137 : Assert(!isnull);
1138 :
1139 : /*
1140 : * Slot sync is currently not supported on a cascading standby. This is
1141 : * because if we allow it, the primary server needs to wait for all the
1142 : * cascading standbys, otherwise, logical subscribers can still be ahead
1143 : * of one of the cascading standbys which we plan to promote. Thus, to
1144 : * avoid this additional complexity, we restrict it for the time being.
1145 : */
1146 26 : if (remote_in_recovery)
1147 2 : ereport(ERROR,
1148 : errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1149 : errmsg("cannot synchronize replication slots from a standby server"));
1150 :
1151 24 : primary_slot_valid = DatumGetBool(slot_getattr(tupslot, 2, &isnull));
1152 : Assert(!isnull);
1153 :
1154 24 : if (!primary_slot_valid)
1155 0 : ereport(ERROR,
1156 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1157 : /* translator: second %s is a GUC variable name */
1158 : errmsg("replication slot \"%s\" specified by \"%s\" does not exist on primary server",
1159 : PrimarySlotName, "primary_slot_name"));
1160 :
1161 24 : ExecClearTuple(tupslot);
1162 24 : walrcv_clear_result(res);
1163 :
1164 24 : if (started_tx)
1165 8 : CommitTransactionCommand();
1166 24 : }
1167 :
1168 : /*
1169 : * Checks if dbname is specified in 'primary_conninfo'.
1170 : *
1171 : * Error out if not specified otherwise return it.
1172 : */
1173 : char *
1174 28 : CheckAndGetDbnameFromConninfo(void)
1175 : {
1176 : char *dbname;
1177 :
1178 : /*
1179 : * The slot synchronization needs a database connection for walrcv_exec to
1180 : * work.
1181 : */
1182 28 : dbname = walrcv_get_dbname_from_conninfo(PrimaryConnInfo);
1183 28 : if (dbname == NULL)
1184 2 : ereport(ERROR,
1185 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1186 :
1187 : /*
1188 : * translator: first %s is a connection option; second %s is a GUC
1189 : * variable name
1190 : */
1191 : errmsg("replication slot synchronization requires \"%s\" to be specified in \"%s\"",
1192 : "dbname", "primary_conninfo"));
1193 26 : return dbname;
1194 : }
1195 :
1196 : /*
1197 : * Return true if all necessary GUCs for slot synchronization are set
1198 : * appropriately, otherwise, return false.
1199 : */
1200 : bool
1201 36 : ValidateSlotSyncParams(int elevel)
1202 : {
1203 : /*
1204 : * Logical slot sync/creation requires logical decoding to be enabled.
1205 : */
1206 36 : if (!IsLogicalDecodingEnabled())
1207 : {
1208 0 : ereport(elevel,
1209 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1210 : errmsg("replication slot synchronization requires \"effective_wal_level\" >= \"logical\" on the primary"),
1211 : errhint("To enable logical decoding on primary, set \"wal_level\" >= \"logical\" or create at least one logical slot when \"wal_level\" = \"replica\"."));
1212 :
1213 0 : return false;
1214 : }
1215 :
1216 : /*
1217 : * A physical replication slot(primary_slot_name) is required on the
1218 : * primary to ensure that the rows needed by the standby are not removed
1219 : * after restarting, so that the synchronized slot on the standby will not
1220 : * be invalidated.
1221 : */
1222 36 : if (PrimarySlotName == NULL || *PrimarySlotName == '\0')
1223 : {
1224 0 : ereport(elevel,
1225 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1226 : /* translator: %s is a GUC variable name */
1227 : errmsg("replication slot synchronization requires \"%s\" to be set", "primary_slot_name"));
1228 0 : return false;
1229 : }
1230 :
1231 : /*
1232 : * hot_standby_feedback must be enabled to cooperate with the physical
1233 : * replication slot, which allows informing the primary about the xmin and
1234 : * catalog_xmin values on the standby.
1235 : */
1236 36 : if (!hot_standby_feedback)
1237 : {
1238 2 : ereport(elevel,
1239 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1240 : /* translator: %s is a GUC variable name */
1241 : errmsg("replication slot synchronization requires \"%s\" to be enabled",
1242 : "hot_standby_feedback"));
1243 2 : return false;
1244 : }
1245 :
1246 : /*
1247 : * The primary_conninfo is required to make connection to primary for
1248 : * getting slots information.
1249 : */
1250 34 : if (PrimaryConnInfo == NULL || *PrimaryConnInfo == '\0')
1251 : {
1252 0 : ereport(elevel,
1253 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1254 : /* translator: %s is a GUC variable name */
1255 : errmsg("replication slot synchronization requires \"%s\" to be set",
1256 : "primary_conninfo"));
1257 0 : return false;
1258 : }
1259 :
1260 34 : return true;
1261 : }
1262 :
1263 : /*
1264 : * Re-read the config file for slot synchronization.
1265 : *
1266 : * Exit or throw error if relevant GUCs have changed depending on whether
1267 : * called from slot sync worker or from the SQL function pg_sync_replication_slots()
1268 : */
1269 : static void
1270 2 : slotsync_reread_config(void)
1271 : {
1272 2 : char *old_primary_conninfo = pstrdup(PrimaryConnInfo);
1273 2 : char *old_primary_slotname = pstrdup(PrimarySlotName);
1274 2 : bool old_sync_replication_slots = sync_replication_slots;
1275 2 : bool old_hot_standby_feedback = hot_standby_feedback;
1276 : bool conninfo_changed;
1277 : bool primary_slotname_changed;
1278 2 : bool is_slotsync_worker = AmLogicalSlotSyncWorkerProcess();
1279 2 : bool parameter_changed = false;
1280 :
1281 : if (is_slotsync_worker)
1282 : Assert(sync_replication_slots);
1283 :
1284 2 : ConfigReloadPending = false;
1285 2 : ProcessConfigFile(PGC_SIGHUP);
1286 :
1287 2 : conninfo_changed = strcmp(old_primary_conninfo, PrimaryConnInfo) != 0;
1288 2 : primary_slotname_changed = strcmp(old_primary_slotname, PrimarySlotName) != 0;
1289 2 : pfree(old_primary_conninfo);
1290 2 : pfree(old_primary_slotname);
1291 :
1292 2 : if (old_sync_replication_slots != sync_replication_slots)
1293 : {
1294 0 : if (is_slotsync_worker)
1295 : {
1296 0 : ereport(LOG,
1297 : /* translator: %s is a GUC variable name */
1298 : errmsg("replication slot synchronization worker will stop because \"%s\" is disabled",
1299 : "sync_replication_slots"));
1300 :
1301 0 : proc_exit(0);
1302 : }
1303 :
1304 0 : parameter_changed = true;
1305 : }
1306 : else
1307 : {
1308 2 : if (conninfo_changed ||
1309 2 : primary_slotname_changed ||
1310 2 : (old_hot_standby_feedback != hot_standby_feedback))
1311 : {
1312 :
1313 2 : if (is_slotsync_worker)
1314 : {
1315 2 : ereport(LOG,
1316 : errmsg("replication slot synchronization worker will restart because of a parameter change"));
1317 :
1318 : /*
1319 : * Reset the last-start time for this worker so that the
1320 : * postmaster can restart it without waiting for
1321 : * SLOTSYNC_RESTART_INTERVAL_SEC.
1322 : */
1323 2 : SlotSyncCtx->last_start_time = 0;
1324 :
1325 2 : proc_exit(0);
1326 : }
1327 :
1328 0 : parameter_changed = true;
1329 : }
1330 : }
1331 :
1332 : /*
1333 : * If we have reached here with a parameter change, we must be running in
1334 : * SQL function, emit error in such a case.
1335 : */
1336 0 : if (parameter_changed)
1337 : {
1338 : Assert(!is_slotsync_worker);
1339 0 : ereport(ERROR,
1340 : errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1341 : errmsg("replication slot synchronization will stop because of a parameter change"));
1342 : }
1343 :
1344 0 : }
1345 :
1346 : /*
1347 : * Interrupt handler for process performing slot synchronization.
1348 : */
1349 : static void
1350 76 : ProcessSlotSyncInterrupts(void)
1351 : {
1352 76 : CHECK_FOR_INTERRUPTS();
1353 :
1354 72 : if (SlotSyncCtx->stopSignaled)
1355 : {
1356 2 : if (AmLogicalSlotSyncWorkerProcess())
1357 : {
1358 2 : ereport(LOG,
1359 : errmsg("replication slot synchronization worker will stop because promotion is triggered"));
1360 :
1361 2 : proc_exit(0);
1362 : }
1363 : else
1364 : {
1365 : /*
1366 : * For the backend executing SQL function
1367 : * pg_sync_replication_slots().
1368 : */
1369 0 : ereport(ERROR,
1370 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1371 : errmsg("replication slot synchronization will stop because promotion is triggered"));
1372 : }
1373 : }
1374 :
1375 70 : if (ConfigReloadPending)
1376 2 : slotsync_reread_config();
1377 68 : }
1378 :
1379 : /*
1380 : * Connection cleanup function for slotsync worker.
1381 : *
1382 : * Called on slotsync worker exit.
1383 : */
1384 : static void
1385 8 : slotsync_worker_disconnect(int code, Datum arg)
1386 : {
1387 8 : WalReceiverConn *wrconn = (WalReceiverConn *) DatumGetPointer(arg);
1388 :
1389 8 : walrcv_disconnect(wrconn);
1390 8 : }
1391 :
1392 : /*
1393 : * Cleanup function for slotsync worker.
1394 : *
1395 : * Called on slotsync worker exit.
1396 : */
1397 : static void
1398 8 : slotsync_worker_onexit(int code, Datum arg)
1399 : {
1400 : /*
1401 : * We need to do slots cleanup here just like WalSndErrorCleanup() does.
1402 : *
1403 : * The startup process during promotion invokes ShutDownSlotSync() which
1404 : * waits for slot sync to finish and it does that by checking the
1405 : * 'syncing' flag. Thus the slot sync worker must be done with slots'
1406 : * release and cleanup to avoid any dangling temporary slots or active
1407 : * slots before it marks itself as finished syncing.
1408 : */
1409 :
1410 : /* Make sure active replication slots are released */
1411 8 : if (MyReplicationSlot != NULL)
1412 0 : ReplicationSlotRelease();
1413 :
1414 : /* Also cleanup the temporary slots. */
1415 8 : ReplicationSlotCleanup(false);
1416 :
1417 8 : SpinLockAcquire(&SlotSyncCtx->mutex);
1418 :
1419 8 : SlotSyncCtx->pid = InvalidPid;
1420 :
1421 : /*
1422 : * If syncing_slots is true, it indicates that the process errored out
1423 : * without resetting the flag. So, we need to clean up shared memory and
1424 : * reset the flag here.
1425 : */
1426 8 : if (syncing_slots)
1427 : {
1428 8 : SlotSyncCtx->syncing = false;
1429 8 : syncing_slots = false;
1430 : }
1431 :
1432 8 : SpinLockRelease(&SlotSyncCtx->mutex);
1433 8 : }
1434 :
1435 : /*
1436 : * Sleep for long enough that we believe it's likely that the slots on primary
1437 : * get updated.
1438 : *
1439 : * If there is no slot activity the wait time between sync-cycles will double
1440 : * (to a maximum of 30s). If there is some slot activity the wait time between
1441 : * sync-cycles is reset to the minimum (200ms).
1442 : */
1443 : static void
1444 34 : wait_for_slot_activity(bool some_slot_updated)
1445 : {
1446 : int rc;
1447 :
1448 34 : if (!some_slot_updated)
1449 : {
1450 : /*
1451 : * No slots were updated, so double the sleep time, but not beyond the
1452 : * maximum allowable value.
1453 : */
1454 22 : sleep_ms = Min(sleep_ms * 2, MAX_SLOTSYNC_WORKER_NAPTIME_MS);
1455 : }
1456 : else
1457 : {
1458 : /*
1459 : * Some slots were updated since the last sleep, so reset the sleep
1460 : * time.
1461 : */
1462 12 : sleep_ms = MIN_SLOTSYNC_WORKER_NAPTIME_MS;
1463 : }
1464 :
1465 34 : rc = WaitLatch(MyLatch,
1466 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
1467 : sleep_ms,
1468 : WAIT_EVENT_REPLICATION_SLOTSYNC_MAIN);
1469 :
1470 34 : if (rc & WL_LATCH_SET)
1471 10 : ResetLatch(MyLatch);
1472 34 : }
1473 :
1474 : /*
1475 : * Emit an error if a concurrent sync call is in progress.
1476 : * Otherwise, advertise that a sync is in progress.
1477 : */
1478 : static void
1479 26 : check_and_set_sync_info(pid_t sync_process_pid)
1480 : {
1481 26 : SpinLockAcquire(&SlotSyncCtx->mutex);
1482 :
1483 26 : if (SlotSyncCtx->syncing)
1484 : {
1485 0 : SpinLockRelease(&SlotSyncCtx->mutex);
1486 0 : ereport(ERROR,
1487 : errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1488 : errmsg("cannot synchronize replication slots concurrently"));
1489 : }
1490 :
1491 : /* The pid must not be already assigned in SlotSyncCtx */
1492 : Assert(SlotSyncCtx->pid == InvalidPid);
1493 :
1494 26 : SlotSyncCtx->syncing = true;
1495 :
1496 : /*
1497 : * Advertise the required PID so that the startup process can kill the
1498 : * slot sync process on promotion.
1499 : */
1500 26 : SlotSyncCtx->pid = sync_process_pid;
1501 :
1502 26 : SpinLockRelease(&SlotSyncCtx->mutex);
1503 :
1504 26 : syncing_slots = true;
1505 26 : }
1506 :
1507 : /*
1508 : * Reset syncing flag.
1509 : */
1510 : static void
1511 18 : reset_syncing_flag(void)
1512 : {
1513 18 : SpinLockAcquire(&SlotSyncCtx->mutex);
1514 18 : SlotSyncCtx->syncing = false;
1515 18 : SlotSyncCtx->pid = InvalidPid;
1516 18 : SpinLockRelease(&SlotSyncCtx->mutex);
1517 :
1518 18 : syncing_slots = false;
1519 18 : }
1520 :
1521 : /*
1522 : * The main loop of our worker process.
1523 : *
1524 : * It connects to the primary server, fetches logical failover slots
1525 : * information periodically in order to create and sync the slots.
1526 : *
1527 : * Note: If any changes are made here, check if the corresponding SQL
1528 : * function logic in SyncReplicationSlots() also needs to be changed.
1529 : */
1530 : void
1531 8 : ReplSlotSyncWorkerMain(const void *startup_data, size_t startup_data_len)
1532 : {
1533 8 : WalReceiverConn *wrconn = NULL;
1534 : char *dbname;
1535 : char *err;
1536 : sigjmp_buf local_sigjmp_buf;
1537 : StringInfoData app_name;
1538 :
1539 : Assert(startup_data_len == 0);
1540 :
1541 8 : MyBackendType = B_SLOTSYNC_WORKER;
1542 :
1543 8 : init_ps_display(NULL);
1544 :
1545 : Assert(GetProcessingMode() == InitProcessing);
1546 :
1547 : /*
1548 : * Create a per-backend PGPROC struct in shared memory. We must do this
1549 : * before we access any shared memory.
1550 : */
1551 8 : InitProcess();
1552 :
1553 : /*
1554 : * Early initialization.
1555 : */
1556 8 : BaseInit();
1557 :
1558 : Assert(SlotSyncCtx != NULL);
1559 :
1560 : /*
1561 : * If an exception is encountered, processing resumes here.
1562 : *
1563 : * We just need to clean up, report the error, and go away.
1564 : *
1565 : * If we do not have this handling here, then since this worker process
1566 : * operates at the bottom of the exception stack, ERRORs turn into FATALs.
1567 : * Therefore, we create our own exception handler to catch ERRORs.
1568 : */
1569 8 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1570 : {
1571 : /* since not using PG_TRY, must reset error stack by hand */
1572 0 : error_context_stack = NULL;
1573 :
1574 : /* Prevents interrupts while cleaning up */
1575 0 : HOLD_INTERRUPTS();
1576 :
1577 : /* Report the error to the server log */
1578 0 : EmitErrorReport();
1579 :
1580 : /*
1581 : * We can now go away. Note that because we called InitProcess, a
1582 : * callback was registered to do ProcKill, which will clean up
1583 : * necessary state.
1584 : */
1585 0 : proc_exit(0);
1586 : }
1587 :
1588 : /* We can now handle ereport(ERROR) */
1589 8 : PG_exception_stack = &local_sigjmp_buf;
1590 :
1591 : /* Setup signal handling */
1592 8 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1593 8 : pqsignal(SIGINT, StatementCancelHandler);
1594 8 : pqsignal(SIGTERM, die);
1595 8 : pqsignal(SIGFPE, FloatExceptionHandler);
1596 8 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1597 8 : pqsignal(SIGUSR2, SIG_IGN);
1598 8 : pqsignal(SIGPIPE, SIG_IGN);
1599 8 : pqsignal(SIGCHLD, SIG_DFL);
1600 :
1601 8 : check_and_set_sync_info(MyProcPid);
1602 :
1603 8 : ereport(LOG, errmsg("slot sync worker started"));
1604 :
1605 : /* Register it as soon as SlotSyncCtx->pid is initialized. */
1606 8 : before_shmem_exit(slotsync_worker_onexit, (Datum) 0);
1607 :
1608 : /*
1609 : * Establishes SIGALRM handler and initialize timeout module. It is needed
1610 : * by InitPostgres to register different timeouts.
1611 : */
1612 8 : InitializeTimeouts();
1613 :
1614 : /* Load the libpq-specific functions */
1615 8 : load_file("libpqwalreceiver", false);
1616 :
1617 : /*
1618 : * Unblock signals (they were blocked when the postmaster forked us)
1619 : */
1620 8 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
1621 :
1622 : /*
1623 : * Set always-secure search path, so malicious users can't redirect user
1624 : * code (e.g. operators).
1625 : *
1626 : * It's not strictly necessary since we won't be scanning or writing to
1627 : * any user table locally, but it's good to retain it here for added
1628 : * precaution.
1629 : */
1630 8 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1631 :
1632 8 : dbname = CheckAndGetDbnameFromConninfo();
1633 :
1634 : /*
1635 : * Connect to the database specified by the user in primary_conninfo. We
1636 : * need a database connection for walrcv_exec to work which we use to
1637 : * fetch slot information from the remote node. See comments atop
1638 : * libpqrcv_exec.
1639 : *
1640 : * We do not specify a specific user here since the slot sync worker will
1641 : * operate as a superuser. This is safe because the slot sync worker does
1642 : * not interact with user tables, eliminating the risk of executing
1643 : * arbitrary code within triggers.
1644 : */
1645 8 : InitPostgres(dbname, InvalidOid, NULL, InvalidOid, 0, NULL);
1646 :
1647 8 : SetProcessingMode(NormalProcessing);
1648 :
1649 8 : initStringInfo(&app_name);
1650 8 : if (cluster_name[0])
1651 8 : appendStringInfo(&app_name, "%s_%s", cluster_name, "slotsync worker");
1652 : else
1653 0 : appendStringInfoString(&app_name, "slotsync worker");
1654 :
1655 : /*
1656 : * Establish the connection to the primary server for slot
1657 : * synchronization.
1658 : */
1659 8 : wrconn = walrcv_connect(PrimaryConnInfo, false, false, false,
1660 : app_name.data, &err);
1661 :
1662 8 : if (!wrconn)
1663 0 : ereport(ERROR,
1664 : errcode(ERRCODE_CONNECTION_FAILURE),
1665 : errmsg("synchronization worker \"%s\" could not connect to the primary server: %s",
1666 : app_name.data, err));
1667 :
1668 8 : pfree(app_name.data);
1669 :
1670 : /*
1671 : * Register the disconnection callback.
1672 : *
1673 : * XXX: This can be combined with previous cleanup registration of
1674 : * slotsync_worker_onexit() but that will need the connection to be made
1675 : * global and we want to avoid introducing global for this purpose.
1676 : */
1677 8 : before_shmem_exit(slotsync_worker_disconnect, PointerGetDatum(wrconn));
1678 :
1679 : /*
1680 : * Using the specified primary server connection, check that we are not a
1681 : * cascading standby and slot configured in 'primary_slot_name' exists on
1682 : * the primary server.
1683 : */
1684 8 : validate_remote_info(wrconn);
1685 :
1686 : /* Main loop to synchronize slots */
1687 : for (;;)
1688 30 : {
1689 38 : bool some_slot_updated = false;
1690 38 : bool started_tx = false;
1691 : List *remote_slots;
1692 :
1693 38 : ProcessSlotSyncInterrupts();
1694 :
1695 : /*
1696 : * The syscache access in fetch_remote_slots() needs a transaction
1697 : * env.
1698 : */
1699 30 : if (!IsTransactionState())
1700 : {
1701 30 : StartTransactionCommand();
1702 30 : started_tx = true;
1703 : }
1704 :
1705 30 : remote_slots = fetch_remote_slots(wrconn, NIL);
1706 30 : some_slot_updated = synchronize_slots(wrconn, remote_slots, NULL);
1707 30 : list_free_deep(remote_slots);
1708 :
1709 30 : if (started_tx)
1710 30 : CommitTransactionCommand();
1711 :
1712 30 : wait_for_slot_activity(some_slot_updated);
1713 : }
1714 :
1715 : /*
1716 : * The slot sync worker can't get here because it will only stop when it
1717 : * receives a stop request from the startup process, or when there is an
1718 : * error.
1719 : */
1720 : Assert(false);
1721 : }
1722 :
1723 : /*
1724 : * Update the inactive_since property for synced slots.
1725 : *
1726 : * Note that this function is currently called when we shutdown the slot
1727 : * sync machinery.
1728 : */
1729 : static void
1730 1852 : update_synced_slots_inactive_since(void)
1731 : {
1732 1852 : TimestampTz now = 0;
1733 :
1734 : /*
1735 : * We need to update inactive_since only when we are promoting standby to
1736 : * correctly interpret the inactive_since if the standby gets promoted
1737 : * without a restart. We don't want the slots to appear inactive for a
1738 : * long time after promotion if they haven't been synchronized recently.
1739 : * Whoever acquires the slot, i.e., makes the slot active, will reset it.
1740 : */
1741 1852 : if (!StandbyMode)
1742 1748 : return;
1743 :
1744 : /* The slot sync worker or the SQL function mustn't be running by now */
1745 : Assert((SlotSyncCtx->pid == InvalidPid) && !SlotSyncCtx->syncing);
1746 :
1747 104 : LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
1748 :
1749 1120 : for (int i = 0; i < max_replication_slots; i++)
1750 : {
1751 1016 : ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
1752 :
1753 : /* Check if it is a synchronized slot */
1754 1016 : if (s->in_use && s->data.synced)
1755 : {
1756 : Assert(SlotIsLogical(s));
1757 :
1758 : /* The slot must not be acquired by any process */
1759 : Assert(s->active_pid == 0);
1760 :
1761 : /* Use the same inactive_since time for all the slots. */
1762 6 : if (now == 0)
1763 4 : now = GetCurrentTimestamp();
1764 :
1765 6 : ReplicationSlotSetInactiveSince(s, now, true);
1766 : }
1767 : }
1768 :
1769 104 : LWLockRelease(ReplicationSlotControlLock);
1770 : }
1771 :
1772 : /*
1773 : * Shut down slot synchronization.
1774 : *
1775 : * This function sets stopSignaled=true and wakes up the slot sync process
1776 : * (either worker or backend running the SQL function pg_sync_replication_slots())
1777 : * so that worker can exit or the SQL function pg_sync_replication_slots() can
1778 : * finish. It also waits till the slot sync worker has exited or
1779 : * pg_sync_replication_slots() has finished.
1780 : */
1781 : void
1782 1852 : ShutDownSlotSync(void)
1783 : {
1784 : pid_t sync_process_pid;
1785 :
1786 1852 : SpinLockAcquire(&SlotSyncCtx->mutex);
1787 :
1788 1852 : SlotSyncCtx->stopSignaled = true;
1789 :
1790 : /*
1791 : * Return if neither the slot sync worker is running nor the function
1792 : * pg_sync_replication_slots() is executing.
1793 : */
1794 1852 : if (!SlotSyncCtx->syncing)
1795 : {
1796 1850 : SpinLockRelease(&SlotSyncCtx->mutex);
1797 1850 : update_synced_slots_inactive_since();
1798 1850 : return;
1799 : }
1800 :
1801 2 : sync_process_pid = SlotSyncCtx->pid;
1802 :
1803 2 : SpinLockRelease(&SlotSyncCtx->mutex);
1804 :
1805 : /*
1806 : * Signal process doing slotsync, if any. The process will stop upon
1807 : * detecting that the stopSignaled flag is set to true.
1808 : */
1809 2 : if (sync_process_pid != InvalidPid)
1810 2 : kill(sync_process_pid, SIGUSR1);
1811 :
1812 : /* Wait for slot sync to end */
1813 : for (;;)
1814 0 : {
1815 : int rc;
1816 :
1817 : /* Wait a bit, we don't expect to have to wait long */
1818 2 : rc = WaitLatch(MyLatch,
1819 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
1820 : 10L, WAIT_EVENT_REPLICATION_SLOTSYNC_SHUTDOWN);
1821 :
1822 2 : if (rc & WL_LATCH_SET)
1823 : {
1824 0 : ResetLatch(MyLatch);
1825 0 : CHECK_FOR_INTERRUPTS();
1826 : }
1827 :
1828 2 : SpinLockAcquire(&SlotSyncCtx->mutex);
1829 :
1830 : /* Ensure that no process is syncing the slots. */
1831 2 : if (!SlotSyncCtx->syncing)
1832 2 : break;
1833 :
1834 0 : SpinLockRelease(&SlotSyncCtx->mutex);
1835 : }
1836 :
1837 2 : SpinLockRelease(&SlotSyncCtx->mutex);
1838 :
1839 2 : update_synced_slots_inactive_since();
1840 : }
1841 :
1842 : /*
1843 : * SlotSyncWorkerCanRestart
1844 : *
1845 : * Return true, indicating worker is allowed to restart, if enough time has
1846 : * passed since it was last launched to reach SLOTSYNC_RESTART_INTERVAL_SEC.
1847 : * Otherwise return false.
1848 : *
1849 : * This is a safety valve to protect against continuous respawn attempts if the
1850 : * worker is dying immediately at launch. Note that since we will retry to
1851 : * launch the worker from the postmaster main loop, we will get another
1852 : * chance later.
1853 : */
1854 : bool
1855 14 : SlotSyncWorkerCanRestart(void)
1856 : {
1857 14 : time_t curtime = time(NULL);
1858 :
1859 : /*
1860 : * If first time through, or time somehow went backwards, always update
1861 : * last_start_time to match the current clock and allow worker start.
1862 : * Otherwise allow it only once enough time has elapsed.
1863 : */
1864 14 : if (SlotSyncCtx->last_start_time == 0 ||
1865 6 : curtime < SlotSyncCtx->last_start_time ||
1866 6 : curtime - SlotSyncCtx->last_start_time >= SLOTSYNC_RESTART_INTERVAL_SEC)
1867 : {
1868 8 : SlotSyncCtx->last_start_time = curtime;
1869 8 : return true;
1870 : }
1871 6 : return false;
1872 : }
1873 :
1874 : /*
1875 : * Is current process syncing replication slots?
1876 : *
1877 : * Could be either backend executing SQL function or slot sync worker.
1878 : */
1879 : bool
1880 48 : IsSyncingReplicationSlots(void)
1881 : {
1882 48 : return syncing_slots;
1883 : }
1884 :
1885 : /*
1886 : * Amount of shared memory required for slot synchronization.
1887 : */
1888 : Size
1889 6492 : SlotSyncShmemSize(void)
1890 : {
1891 6492 : return sizeof(SlotSyncCtxStruct);
1892 : }
1893 :
1894 : /*
1895 : * Allocate and initialize the shared memory of slot synchronization.
1896 : */
1897 : void
1898 2266 : SlotSyncShmemInit(void)
1899 : {
1900 2266 : Size size = SlotSyncShmemSize();
1901 : bool found;
1902 :
1903 2266 : SlotSyncCtx = (SlotSyncCtxStruct *)
1904 2266 : ShmemInitStruct("Slot Sync Data", size, &found);
1905 :
1906 2266 : if (!found)
1907 : {
1908 2266 : memset(SlotSyncCtx, 0, size);
1909 2266 : SlotSyncCtx->pid = InvalidPid;
1910 2266 : SpinLockInit(&SlotSyncCtx->mutex);
1911 : }
1912 2266 : }
1913 :
1914 : /*
1915 : * Error cleanup callback for slot sync SQL function.
1916 : */
1917 : static void
1918 2 : slotsync_failure_callback(int code, Datum arg)
1919 : {
1920 2 : WalReceiverConn *wrconn = (WalReceiverConn *) DatumGetPointer(arg);
1921 :
1922 : /*
1923 : * We need to do slots cleanup here just like WalSndErrorCleanup() does.
1924 : *
1925 : * The startup process during promotion invokes ShutDownSlotSync() which
1926 : * waits for slot sync to finish and it does that by checking the
1927 : * 'syncing' flag. Thus the SQL function must be done with slots' release
1928 : * and cleanup to avoid any dangling temporary slots or active slots
1929 : * before it marks itself as finished syncing.
1930 : */
1931 :
1932 : /* Make sure active replication slots are released */
1933 2 : if (MyReplicationSlot != NULL)
1934 0 : ReplicationSlotRelease();
1935 :
1936 : /* Also cleanup the synced temporary slots. */
1937 2 : ReplicationSlotCleanup(true);
1938 :
1939 : /*
1940 : * The set syncing_slots indicates that the process errored out without
1941 : * resetting the flag. So, we need to clean up shared memory and reset the
1942 : * flag here.
1943 : */
1944 2 : if (syncing_slots)
1945 2 : reset_syncing_flag();
1946 :
1947 2 : walrcv_disconnect(wrconn);
1948 2 : }
1949 :
1950 : /*
1951 : * Helper function to extract slot names from a list of remote slots
1952 : */
1953 : static List *
1954 2 : extract_slot_names(List *remote_slots)
1955 : {
1956 2 : List *slot_names = NIL;
1957 :
1958 6 : foreach_ptr(RemoteSlot, remote_slot, remote_slots)
1959 : {
1960 : char *slot_name;
1961 :
1962 2 : slot_name = pstrdup(remote_slot->name);
1963 2 : slot_names = lappend(slot_names, slot_name);
1964 : }
1965 :
1966 2 : return slot_names;
1967 : }
1968 :
1969 : /*
1970 : * Synchronize the failover enabled replication slots using the specified
1971 : * primary server connection.
1972 : *
1973 : * Repeatedly fetches and updates replication slot information from the
1974 : * primary until all slots are at least "sync ready".
1975 : *
1976 : * Exits early if promotion is triggered or certain critical
1977 : * configuration parameters have changed.
1978 : */
1979 : void
1980 18 : SyncReplicationSlots(WalReceiverConn *wrconn)
1981 : {
1982 18 : PG_ENSURE_ERROR_CLEANUP(slotsync_failure_callback, PointerGetDatum(wrconn));
1983 : {
1984 18 : List *remote_slots = NIL;
1985 18 : List *slot_names = NIL; /* List of slot names to track */
1986 :
1987 18 : check_and_set_sync_info(MyProcPid);
1988 :
1989 : /* Check for interrupts and config changes */
1990 18 : ProcessSlotSyncInterrupts();
1991 :
1992 18 : validate_remote_info(wrconn);
1993 :
1994 : /* Retry until all the slots are sync-ready */
1995 : for (;;)
1996 4 : {
1997 20 : bool slot_persistence_pending = false;
1998 20 : bool some_slot_updated = false;
1999 :
2000 : /* Check for interrupts and config changes */
2001 20 : ProcessSlotSyncInterrupts();
2002 :
2003 : /* We must be in a valid transaction state */
2004 : Assert(IsTransactionState());
2005 :
2006 : /*
2007 : * Fetch remote slot info for the given slot_names. If slot_names
2008 : * is NIL, fetch all failover-enabled slots. Note that we reuse
2009 : * slot_names from the first iteration; re-fetching all failover
2010 : * slots each time could cause an endless loop. Instead of
2011 : * reprocessing only the pending slots in each iteration, it's
2012 : * better to process all the slots received in the first
2013 : * iteration. This ensures that by the time we're done, all slots
2014 : * reflect the latest values.
2015 : */
2016 20 : remote_slots = fetch_remote_slots(wrconn, slot_names);
2017 :
2018 : /* Attempt to synchronize slots */
2019 20 : some_slot_updated = synchronize_slots(wrconn, remote_slots,
2020 : &slot_persistence_pending);
2021 :
2022 : /*
2023 : * If slot_persistence_pending is true, extract slot names for
2024 : * future iterations (only needed if we haven't done it yet)
2025 : */
2026 20 : if (slot_names == NIL && slot_persistence_pending)
2027 2 : slot_names = extract_slot_names(remote_slots);
2028 :
2029 : /* Free the current remote_slots list */
2030 20 : list_free_deep(remote_slots);
2031 :
2032 : /* Done if all slots are persisted i.e are sync-ready */
2033 20 : if (!slot_persistence_pending)
2034 16 : break;
2035 :
2036 : /* wait before retrying again */
2037 4 : wait_for_slot_activity(some_slot_updated);
2038 : }
2039 :
2040 16 : if (slot_names)
2041 2 : list_free_deep(slot_names);
2042 :
2043 : /* Cleanup the synced temporary slots */
2044 16 : ReplicationSlotCleanup(true);
2045 :
2046 : /* We are done with sync, so reset sync flag */
2047 16 : reset_syncing_flag();
2048 : }
2049 18 : PG_END_ENSURE_ERROR_CLEANUP(slotsync_failure_callback, PointerGetDatum(wrconn));
2050 16 : }
|
