Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * standby.c
4 : * Misc functions used in Hot Standby mode.
5 : *
6 : * All functions for handling RM_STANDBY_ID, which relate to
7 : * AccessExclusiveLocks and starting snapshots for Hot Standby mode.
8 : * Plus conflict recovery processing.
9 : *
10 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
11 : * Portions Copyright (c) 1994, Regents of the University of California
12 : *
13 : * IDENTIFICATION
14 : * src/backend/storage/ipc/standby.c
15 : *
16 : *-------------------------------------------------------------------------
17 : */
18 : #include "postgres.h"
19 : #include "access/transam.h"
20 : #include "access/twophase.h"
21 : #include "access/xact.h"
22 : #include "access/xloginsert.h"
23 : #include "access/xlogrecovery.h"
24 : #include "access/xlogutils.h"
25 : #include "miscadmin.h"
26 : #include "pgstat.h"
27 : #include "replication/slot.h"
28 : #include "storage/bufmgr.h"
29 : #include "storage/proc.h"
30 : #include "storage/procarray.h"
31 : #include "storage/sinvaladt.h"
32 : #include "storage/standby.h"
33 : #include "utils/hsearch.h"
34 : #include "utils/injection_point.h"
35 : #include "utils/ps_status.h"
36 : #include "utils/timeout.h"
37 : #include "utils/timestamp.h"
38 : #include "utils/wait_event.h"
39 :
40 : /* User-settable GUC parameters */
41 : int max_standby_archive_delay = 30 * 1000;
42 : int max_standby_streaming_delay = 30 * 1000;
43 : bool log_recovery_conflict_waits = false;
44 :
45 : /*
46 : * Keep track of all the exclusive locks owned by original transactions.
47 : * For each known exclusive lock, there is a RecoveryLockEntry in the
48 : * RecoveryLockHash hash table. All RecoveryLockEntrys belonging to a
49 : * given XID are chained together so that we can find them easily.
50 : * For each original transaction that is known to have any such locks,
51 : * there is a RecoveryLockXidEntry in the RecoveryLockXidHash hash table,
52 : * which stores the head of the chain of its locks.
53 : */
54 : typedef struct RecoveryLockEntry
55 : {
56 : xl_standby_lock key; /* hash key: xid, dbOid, relOid */
57 : struct RecoveryLockEntry *next; /* chain link */
58 : } RecoveryLockEntry;
59 :
60 : typedef struct RecoveryLockXidEntry
61 : {
62 : TransactionId xid; /* hash key -- must be first */
63 : struct RecoveryLockEntry *head; /* chain head */
64 : } RecoveryLockXidEntry;
65 :
66 : static HTAB *RecoveryLockHash = NULL;
67 : static HTAB *RecoveryLockXidHash = NULL;
68 :
69 : /* Flags set by timeout handlers */
70 : static volatile sig_atomic_t got_standby_deadlock_timeout = false;
71 : static volatile sig_atomic_t got_standby_delay_timeout = false;
72 : static volatile sig_atomic_t got_standby_lock_timeout = false;
73 :
74 : static void ResolveRecoveryConflictWithVirtualXIDs(VirtualTransactionId *waitlist,
75 : RecoveryConflictReason reason,
76 : uint32 wait_event_info,
77 : bool report_waiting);
78 : static void SendRecoveryConflictWithBufferPin(RecoveryConflictReason reason);
79 : static XLogRecPtr LogCurrentRunningXacts(RunningTransactions CurrRunningXacts);
80 : static void LogAccessExclusiveLocks(int nlocks, xl_standby_lock *locks);
81 : static const char *get_recovery_conflict_desc(RecoveryConflictReason reason);
82 :
83 : /*
84 : * InitRecoveryTransactionEnvironment
85 : * Initialize tracking of our primary's in-progress transactions.
86 : *
87 : * We need to issue shared invalidations and hold locks. Holding locks
88 : * means others may want to wait on us, so we need to make a lock table
89 : * vxact entry like a real transaction. We could create and delete
90 : * lock table entries for each transaction but its simpler just to create
91 : * one permanent entry and leave it there all the time. Locks are then
92 : * acquired and released as needed. Yes, this means you can see the
93 : * Startup process in pg_locks once we have run this.
94 : */
95 : void
96 116 : InitRecoveryTransactionEnvironment(void)
97 : {
98 : VirtualTransactionId vxid;
99 : HASHCTL hash_ctl;
100 :
101 : Assert(RecoveryLockHash == NULL); /* don't run this twice */
102 :
103 : /*
104 : * Initialize the hash tables for tracking the locks held by each
105 : * transaction.
106 : */
107 116 : hash_ctl.keysize = sizeof(xl_standby_lock);
108 116 : hash_ctl.entrysize = sizeof(RecoveryLockEntry);
109 116 : RecoveryLockHash = hash_create("RecoveryLockHash",
110 : 64,
111 : &hash_ctl,
112 : HASH_ELEM | HASH_BLOBS);
113 116 : hash_ctl.keysize = sizeof(TransactionId);
114 116 : hash_ctl.entrysize = sizeof(RecoveryLockXidEntry);
115 116 : RecoveryLockXidHash = hash_create("RecoveryLockXidHash",
116 : 64,
117 : &hash_ctl,
118 : HASH_ELEM | HASH_BLOBS);
119 :
120 : /*
121 : * Initialize shared invalidation management for Startup process, being
122 : * careful to register ourselves as a sendOnly process so we don't need to
123 : * read messages, nor will we get signaled when the queue starts filling
124 : * up.
125 : */
126 116 : SharedInvalBackendInit(true);
127 :
128 : /*
129 : * Lock a virtual transaction id for Startup process.
130 : *
131 : * We need to do GetNextLocalTransactionId() because
132 : * SharedInvalBackendInit() leaves localTransactionId invalid and the lock
133 : * manager doesn't like that at all.
134 : *
135 : * Note that we don't need to run XactLockTableInsert() because nobody
136 : * needs to wait on xids. That sounds a little strange, but table locks
137 : * are held by vxids and row level locks are held by xids. All queries
138 : * hold AccessShareLocks so never block while we write or lock new rows.
139 : */
140 116 : MyProc->vxid.procNumber = MyProcNumber;
141 116 : vxid.procNumber = MyProcNumber;
142 116 : vxid.localTransactionId = GetNextLocalTransactionId();
143 116 : VirtualXactLockTableInsert(vxid);
144 :
145 116 : standbyState = STANDBY_INITIALIZED;
146 116 : }
147 :
148 : /*
149 : * ShutdownRecoveryTransactionEnvironment
150 : * Shut down transaction tracking
151 : *
152 : * Prepare to switch from hot standby mode to normal operation. Shut down
153 : * recovery-time transaction tracking.
154 : *
155 : * This must be called even in shutdown of startup process if transaction
156 : * tracking has been initialized. Otherwise some locks the tracked
157 : * transactions were holding will not be released and may interfere with
158 : * the processes still running (but will exit soon later) at the exit of
159 : * startup process.
160 : */
161 : void
162 171 : ShutdownRecoveryTransactionEnvironment(void)
163 : {
164 : /*
165 : * Do nothing if RecoveryLockHash is NULL because that means that
166 : * transaction tracking has not yet been initialized or has already been
167 : * shut down. This makes it safe to have possibly-redundant calls of this
168 : * function during process exit.
169 : */
170 171 : if (RecoveryLockHash == NULL)
171 55 : return;
172 :
173 : /* Mark all tracked in-progress transactions as finished. */
174 116 : ExpireAllKnownAssignedTransactionIds();
175 :
176 : /* Release all locks the tracked transactions were holding */
177 116 : StandbyReleaseAllLocks();
178 :
179 : /* Destroy the lock hash tables. */
180 116 : hash_destroy(RecoveryLockHash);
181 116 : hash_destroy(RecoveryLockXidHash);
182 116 : RecoveryLockHash = NULL;
183 116 : RecoveryLockXidHash = NULL;
184 :
185 : /* Cleanup our VirtualTransaction */
186 116 : VirtualXactLockTableCleanup();
187 : }
188 :
189 :
190 : /*
191 : * -----------------------------------------------------
192 : * Standby wait timers and backend cancel logic
193 : * -----------------------------------------------------
194 : */
195 :
196 : /*
197 : * Determine the cutoff time at which we want to start canceling conflicting
198 : * transactions. Returns zero (a time safely in the past) if we are willing
199 : * to wait forever.
200 : */
201 : static TimestampTz
202 25 : GetStandbyLimitTime(void)
203 : {
204 : TimestampTz rtime;
205 : bool fromStream;
206 :
207 : /*
208 : * The cutoff time is the last WAL data receipt time plus the appropriate
209 : * delay variable. Delay of -1 means wait forever.
210 : */
211 25 : GetXLogReceiptTime(&rtime, &fromStream);
212 25 : if (fromStream)
213 : {
214 25 : if (max_standby_streaming_delay < 0)
215 0 : return 0; /* wait forever */
216 25 : return TimestampTzPlusMilliseconds(rtime, max_standby_streaming_delay);
217 : }
218 : else
219 : {
220 0 : if (max_standby_archive_delay < 0)
221 0 : return 0; /* wait forever */
222 0 : return TimestampTzPlusMilliseconds(rtime, max_standby_archive_delay);
223 : }
224 : }
225 :
226 : #define STANDBY_INITIAL_WAIT_US 1000
227 : static int standbyWait_us = STANDBY_INITIAL_WAIT_US;
228 :
229 : /*
230 : * Standby wait logic for ResolveRecoveryConflictWithVirtualXIDs.
231 : * We wait here for a while then return. If we decide we can't wait any
232 : * more then we return true, if we can wait some more return false.
233 : */
234 : static bool
235 15 : WaitExceedsMaxStandbyDelay(uint32 wait_event_info)
236 : {
237 : TimestampTz ltime;
238 :
239 15 : CHECK_FOR_INTERRUPTS();
240 :
241 : /* Are we past the limit time? */
242 15 : ltime = GetStandbyLimitTime();
243 15 : if (ltime && GetCurrentTimestamp() >= ltime)
244 3 : return true;
245 :
246 : /*
247 : * Sleep a bit (this is essential to avoid busy-waiting).
248 : */
249 12 : pgstat_report_wait_start(wait_event_info);
250 12 : pg_usleep(standbyWait_us);
251 12 : pgstat_report_wait_end();
252 :
253 : /*
254 : * Progressively increase the sleep times, but not to more than 1s, since
255 : * pg_usleep isn't interruptible on some platforms.
256 : */
257 12 : standbyWait_us *= 2;
258 12 : if (standbyWait_us > 1000000)
259 0 : standbyWait_us = 1000000;
260 :
261 12 : return false;
262 : }
263 :
264 : /*
265 : * Log the recovery conflict.
266 : *
267 : * wait_start is the timestamp when the caller started to wait.
268 : * now is the timestamp when this function has been called.
269 : * wait_list is the list of virtual transaction ids assigned to
270 : * conflicting processes. still_waiting indicates whether
271 : * the startup process is still waiting for the recovery conflict
272 : * to be resolved or not.
273 : */
274 : void
275 10 : LogRecoveryConflict(RecoveryConflictReason reason, TimestampTz wait_start,
276 : TimestampTz now, VirtualTransactionId *wait_list,
277 : bool still_waiting)
278 : {
279 : long secs;
280 : int usecs;
281 : long msecs;
282 : StringInfoData buf;
283 10 : int nprocs = 0;
284 :
285 : /*
286 : * There must be no conflicting processes when the recovery conflict has
287 : * already been resolved.
288 : */
289 : Assert(still_waiting || wait_list == NULL);
290 :
291 10 : TimestampDifference(wait_start, now, &secs, &usecs);
292 10 : msecs = secs * 1000 + usecs / 1000;
293 10 : usecs = usecs % 1000;
294 :
295 10 : if (wait_list)
296 : {
297 : VirtualTransactionId *vxids;
298 :
299 : /* Construct a string of list of the conflicting processes */
300 3 : vxids = wait_list;
301 6 : while (VirtualTransactionIdIsValid(*vxids))
302 : {
303 3 : PGPROC *proc = ProcNumberGetProc(vxids->procNumber);
304 :
305 : /* proc can be NULL if the target backend is not active */
306 3 : if (proc)
307 : {
308 3 : if (nprocs == 0)
309 : {
310 3 : initStringInfo(&buf);
311 3 : appendStringInfo(&buf, "%d", proc->pid);
312 : }
313 : else
314 0 : appendStringInfo(&buf, ", %d", proc->pid);
315 :
316 3 : nprocs++;
317 : }
318 :
319 3 : vxids++;
320 : }
321 : }
322 :
323 : /*
324 : * If wait_list is specified, report the list of PIDs of active
325 : * conflicting backends in a detail message. Note that if all the backends
326 : * in the list are not active, no detail message is logged.
327 : */
328 10 : if (still_waiting)
329 : {
330 5 : ereport(LOG,
331 : errmsg("recovery still waiting after %ld.%03d ms: %s",
332 : msecs, usecs, get_recovery_conflict_desc(reason)),
333 : nprocs > 0 ? errdetail_log_plural("Conflicting process: %s.",
334 : "Conflicting processes: %s.",
335 : nprocs, buf.data) : 0);
336 : }
337 : else
338 : {
339 5 : ereport(LOG,
340 : errmsg("recovery finished waiting after %ld.%03d ms: %s",
341 : msecs, usecs, get_recovery_conflict_desc(reason)));
342 : }
343 :
344 10 : if (nprocs > 0)
345 3 : pfree(buf.data);
346 10 : }
347 :
348 : /*
349 : * This is the main executioner for any query backend that conflicts with
350 : * recovery processing. Judgement has already been passed on it within
351 : * a specific rmgr. Here we just issue the orders to the procs. The procs
352 : * then throw the required error as instructed.
353 : *
354 : * If report_waiting is true, "waiting" is reported in PS display and the
355 : * wait for recovery conflict is reported in the log, if necessary. If
356 : * the caller is responsible for reporting them, report_waiting should be
357 : * false. Otherwise, both the caller and this function report the same
358 : * thing unexpectedly.
359 : */
360 : static void
361 16428 : ResolveRecoveryConflictWithVirtualXIDs(VirtualTransactionId *waitlist,
362 : RecoveryConflictReason reason,
363 : uint32 wait_event_info,
364 : bool report_waiting)
365 : {
366 16428 : TimestampTz waitStart = 0;
367 16428 : bool waiting = false;
368 16428 : bool logged_recovery_conflict = false;
369 :
370 : /* Fast exit, to avoid a kernel call if there's no work to be done. */
371 16428 : if (!VirtualTransactionIdIsValid(*waitlist))
372 16425 : return;
373 :
374 : /* Set the wait start timestamp for reporting */
375 3 : if (report_waiting && (log_recovery_conflict_waits || update_process_title))
376 2 : waitStart = GetCurrentTimestamp();
377 :
378 6 : while (VirtualTransactionIdIsValid(*waitlist))
379 : {
380 : /* reset standbyWait_us for each xact we wait for */
381 3 : standbyWait_us = STANDBY_INITIAL_WAIT_US;
382 :
383 : /* wait until the virtual xid is gone */
384 18 : while (!VirtualXactLock(*waitlist, false))
385 : {
386 : /* Is it time to kill it? */
387 15 : if (WaitExceedsMaxStandbyDelay(wait_event_info))
388 : {
389 : bool signaled;
390 :
391 : /*
392 : * Now find out who to throw out of the balloon.
393 : */
394 : Assert(VirtualTransactionIdIsValid(*waitlist));
395 3 : signaled = SignalRecoveryConflictWithVirtualXID(*waitlist, reason);
396 :
397 : /*
398 : * Wait a little bit for it to die so that we avoid flooding
399 : * an unresponsive backend when system is heavily loaded.
400 : */
401 3 : if (signaled)
402 3 : pg_usleep(5000L);
403 : }
404 :
405 15 : if (waitStart != 0 && (!logged_recovery_conflict || !waiting))
406 : {
407 14 : TimestampTz now = 0;
408 : bool maybe_log_conflict;
409 : bool maybe_update_title;
410 :
411 14 : maybe_log_conflict = (log_recovery_conflict_waits && !logged_recovery_conflict);
412 14 : maybe_update_title = (update_process_title && !waiting);
413 :
414 : /* Get the current timestamp if not report yet */
415 14 : if (maybe_log_conflict || maybe_update_title)
416 14 : now = GetCurrentTimestamp();
417 :
418 : /*
419 : * Report via ps if we have been waiting for more than 500
420 : * msec (should that be configurable?)
421 : */
422 28 : if (maybe_update_title &&
423 14 : TimestampDifferenceExceeds(waitStart, now, 500))
424 : {
425 0 : set_ps_display_suffix("waiting");
426 0 : waiting = true;
427 : }
428 :
429 : /*
430 : * Emit the log message if the startup process is waiting
431 : * longer than deadlock_timeout for recovery conflict.
432 : */
433 22 : if (maybe_log_conflict &&
434 8 : TimestampDifferenceExceeds(waitStart, now, DeadlockTimeout))
435 : {
436 2 : LogRecoveryConflict(reason, waitStart, now, waitlist, true);
437 2 : logged_recovery_conflict = true;
438 : }
439 : }
440 : }
441 :
442 : /* The virtual transaction is gone now, wait for the next one */
443 3 : waitlist++;
444 : }
445 :
446 : /*
447 : * Emit the log message if recovery conflict was resolved but the startup
448 : * process waited longer than deadlock_timeout for it.
449 : */
450 3 : if (logged_recovery_conflict)
451 2 : LogRecoveryConflict(reason, waitStart, GetCurrentTimestamp(),
452 : NULL, false);
453 :
454 : /* reset ps display to remove the suffix if we added one */
455 3 : if (waiting)
456 0 : set_ps_display_remove_suffix();
457 :
458 : }
459 :
460 : /*
461 : * Generate whatever recovery conflicts are needed to eliminate snapshots that
462 : * might see XIDs <= snapshotConflictHorizon as still running.
463 : *
464 : * snapshotConflictHorizon cutoffs are our standard approach to generating
465 : * granular recovery conflicts. Note that InvalidTransactionId values are
466 : * interpreted as "definitely don't need any conflicts" here, which is a
467 : * general convention that WAL records can (and often do) depend on.
468 : */
469 : void
470 17127 : ResolveRecoveryConflictWithSnapshot(TransactionId snapshotConflictHorizon,
471 : bool isCatalogRel,
472 : RelFileLocator locator)
473 : {
474 : VirtualTransactionId *backends;
475 :
476 : /*
477 : * If we get passed InvalidTransactionId then we do nothing (no conflict).
478 : *
479 : * This can happen whenever the changes in the WAL record do not affect
480 : * visibility on a standby. For example: a record that only freezes an
481 : * xmax from a locker.
482 : *
483 : * It's also quite common with records generated during index deletion
484 : * (original execution of the deletion can reason that a recovery conflict
485 : * which is sufficient for the deletion operation must take place before
486 : * replay of the deletion record itself).
487 : */
488 17127 : if (!TransactionIdIsValid(snapshotConflictHorizon))
489 701 : return;
490 :
491 : Assert(TransactionIdIsNormal(snapshotConflictHorizon));
492 16426 : backends = GetConflictingVirtualXIDs(snapshotConflictHorizon,
493 : locator.dbOid);
494 16426 : ResolveRecoveryConflictWithVirtualXIDs(backends,
495 : RECOVERY_CONFLICT_SNAPSHOT,
496 : WAIT_EVENT_RECOVERY_CONFLICT_SNAPSHOT,
497 : true);
498 :
499 : /*
500 : * Note that WaitExceedsMaxStandbyDelay() is not taken into account here
501 : * (as opposed to ResolveRecoveryConflictWithVirtualXIDs() above). That
502 : * seems OK, given that this kind of conflict should not normally be
503 : * reached, e.g. due to using a physical replication slot.
504 : */
505 16426 : if (IsLogicalDecodingEnabled() && isCatalogRel)
506 16 : InvalidateObsoleteReplicationSlots(RS_INVAL_HORIZON, 0, locator.dbOid,
507 : snapshotConflictHorizon);
508 : }
509 :
510 : /*
511 : * Variant of ResolveRecoveryConflictWithSnapshot that works with
512 : * FullTransactionId values
513 : */
514 : void
515 46 : ResolveRecoveryConflictWithSnapshotFullXid(FullTransactionId snapshotConflictHorizon,
516 : bool isCatalogRel,
517 : RelFileLocator locator)
518 : {
519 : /*
520 : * ResolveRecoveryConflictWithSnapshot operates on 32-bit TransactionIds,
521 : * so truncate the logged FullTransactionId. If the logged value is very
522 : * old, so that XID wrap-around already happened on it, there can't be any
523 : * snapshots that still see it.
524 : */
525 46 : FullTransactionId nextXid = ReadNextFullTransactionId();
526 : uint64 diff;
527 :
528 46 : diff = U64FromFullTransactionId(nextXid) -
529 46 : U64FromFullTransactionId(snapshotConflictHorizon);
530 46 : if (diff < MaxTransactionId / 2)
531 : {
532 : TransactionId truncated;
533 :
534 46 : truncated = XidFromFullTransactionId(snapshotConflictHorizon);
535 46 : ResolveRecoveryConflictWithSnapshot(truncated,
536 : isCatalogRel,
537 : locator);
538 : }
539 46 : }
540 :
541 : void
542 1 : ResolveRecoveryConflictWithTablespace(Oid tsid)
543 : {
544 : VirtualTransactionId *temp_file_users;
545 :
546 : /*
547 : * Standby users may be currently using this tablespace for their
548 : * temporary files. We only care about current users because
549 : * temp_tablespace parameter will just ignore tablespaces that no longer
550 : * exist.
551 : *
552 : * Ask everybody to cancel their queries immediately so we can ensure no
553 : * temp files remain and we can remove the tablespace. Nuke the entire
554 : * site from orbit, it's the only way to be sure.
555 : *
556 : * XXX: We could work out the pids of active backends using this
557 : * tablespace by examining the temp filenames in the directory. We would
558 : * then convert the pids into VirtualXIDs before attempting to cancel
559 : * them.
560 : *
561 : * We don't wait for commit because drop tablespace is non-transactional.
562 : */
563 1 : temp_file_users = GetConflictingVirtualXIDs(InvalidTransactionId,
564 : InvalidOid);
565 1 : ResolveRecoveryConflictWithVirtualXIDs(temp_file_users,
566 : RECOVERY_CONFLICT_TABLESPACE,
567 : WAIT_EVENT_RECOVERY_CONFLICT_TABLESPACE,
568 : true);
569 1 : }
570 :
571 : void
572 16 : ResolveRecoveryConflictWithDatabase(Oid dbid)
573 : {
574 : /*
575 : * We don't do ResolveRecoveryConflictWithVirtualXIDs() here since that
576 : * only waits for transactions and completely idle sessions would block
577 : * us. This is rare enough that we do this as simply as possible: no wait,
578 : * just force them off immediately.
579 : *
580 : * No locking is required here because we already acquired
581 : * AccessExclusiveLock. Anybody trying to connect while we do this will
582 : * block during InitPostgres() and then disconnect when they see the
583 : * database has been removed.
584 : */
585 18 : while (CountDBBackends(dbid) > 0)
586 : {
587 2 : SignalRecoveryConflictWithDatabase(dbid, RECOVERY_CONFLICT_DATABASE);
588 :
589 : /*
590 : * Wait awhile for them to die so that we avoid flooding an
591 : * unresponsive backend when system is heavily loaded.
592 : */
593 2 : pg_usleep(10000);
594 : }
595 16 : }
596 :
597 : /*
598 : * ResolveRecoveryConflictWithLock is called from ProcSleep()
599 : * to resolve conflicts with other backends holding relation locks.
600 : *
601 : * The WaitLatch sleep normally done in ProcSleep()
602 : * (when not InHotStandby) is performed here, for code clarity.
603 : *
604 : * We either resolve conflicts immediately or set a timeout to wake us at
605 : * the limit of our patience.
606 : *
607 : * Resolve conflicts by canceling to all backends holding a conflicting
608 : * lock. As we are already queued to be granted the lock, no new lock
609 : * requests conflicting with ours will be granted in the meantime.
610 : *
611 : * We also must check for deadlocks involving the Startup process and
612 : * hot-standby backend processes. If deadlock_timeout is reached in
613 : * this function, all the backends holding the conflicting locks are
614 : * requested to check themselves for deadlocks.
615 : *
616 : * logging_conflict should be true if the recovery conflict has not been
617 : * logged yet even though logging is enabled. After deadlock_timeout is
618 : * reached and the request for deadlock check is sent, we wait again to
619 : * be signaled by the release of the lock if logging_conflict is false.
620 : * Otherwise we return without waiting again so that the caller can report
621 : * the recovery conflict. In this case, then, this function is called again
622 : * with logging_conflict=false (because the recovery conflict has already
623 : * been logged) and we will wait again for the lock to be released.
624 : */
625 : void
626 3 : ResolveRecoveryConflictWithLock(LOCKTAG locktag, bool logging_conflict)
627 : {
628 : TimestampTz ltime;
629 : TimestampTz now;
630 :
631 : Assert(InHotStandby);
632 :
633 3 : ltime = GetStandbyLimitTime();
634 3 : now = GetCurrentTimestamp();
635 :
636 : /*
637 : * Update waitStart if first time through after the startup process
638 : * started waiting for the lock. It should not be updated every time
639 : * ResolveRecoveryConflictWithLock() is called during the wait.
640 : *
641 : * Use the current time obtained for comparison with ltime as waitStart
642 : * (i.e., the time when this process started waiting for the lock). Since
643 : * getting the current time newly can cause overhead, we reuse the
644 : * already-obtained time to avoid that overhead.
645 : *
646 : * Note that waitStart is updated without holding the lock table's
647 : * partition lock, to avoid the overhead by additional lock acquisition.
648 : * This can cause "waitstart" in pg_locks to become NULL for a very short
649 : * period of time after the wait started even though "granted" is false.
650 : * This is OK in practice because we can assume that users are likely to
651 : * look at "waitstart" when waiting for the lock for a long time.
652 : */
653 3 : if (pg_atomic_read_u64(&MyProc->waitStart) == 0)
654 1 : pg_atomic_write_u64(&MyProc->waitStart, now);
655 :
656 3 : if (now >= ltime && ltime != 0)
657 1 : {
658 : /*
659 : * We're already behind, so clear a path as quickly as possible.
660 : */
661 : VirtualTransactionId *backends;
662 :
663 1 : backends = GetLockConflicts(&locktag, AccessExclusiveLock, NULL);
664 :
665 : /*
666 : * Prevent ResolveRecoveryConflictWithVirtualXIDs() from reporting
667 : * "waiting" in PS display by disabling its argument report_waiting
668 : * because the caller, WaitOnLock(), has already reported that.
669 : */
670 1 : ResolveRecoveryConflictWithVirtualXIDs(backends,
671 : RECOVERY_CONFLICT_LOCK,
672 1 : PG_WAIT_LOCK | locktag.locktag_type,
673 : false);
674 : }
675 : else
676 : {
677 : /*
678 : * Wait (or wait again) until ltime, and check for deadlocks as well
679 : * if we will be waiting longer than deadlock_timeout
680 : */
681 : EnableTimeoutParams timeouts[2];
682 2 : int cnt = 0;
683 :
684 2 : if (ltime != 0)
685 : {
686 2 : got_standby_lock_timeout = false;
687 2 : timeouts[cnt].id = STANDBY_LOCK_TIMEOUT;
688 2 : timeouts[cnt].type = TMPARAM_AT;
689 2 : timeouts[cnt].fin_time = ltime;
690 2 : cnt++;
691 : }
692 :
693 2 : got_standby_deadlock_timeout = false;
694 2 : timeouts[cnt].id = STANDBY_DEADLOCK_TIMEOUT;
695 2 : timeouts[cnt].type = TMPARAM_AFTER;
696 2 : timeouts[cnt].delay_ms = DeadlockTimeout;
697 2 : cnt++;
698 :
699 2 : enable_timeouts(timeouts, cnt);
700 : }
701 :
702 : /* Wait to be signaled by the release of the Relation Lock */
703 3 : ProcWaitForSignal(PG_WAIT_LOCK | locktag.locktag_type);
704 :
705 : /*
706 : * Exit if ltime is reached. Then all the backends holding conflicting
707 : * locks will be canceled in the next ResolveRecoveryConflictWithLock()
708 : * call.
709 : */
710 3 : if (got_standby_lock_timeout)
711 0 : goto cleanup;
712 :
713 3 : if (got_standby_deadlock_timeout)
714 : {
715 : VirtualTransactionId *backends;
716 :
717 2 : backends = GetLockConflicts(&locktag, AccessExclusiveLock, NULL);
718 :
719 : /* Quick exit if there's no work to be done */
720 2 : if (!VirtualTransactionIdIsValid(*backends))
721 0 : goto cleanup;
722 :
723 : /*
724 : * Send signals to all the backends holding the conflicting locks, to
725 : * ask them to check themselves for deadlocks.
726 : */
727 4 : while (VirtualTransactionIdIsValid(*backends))
728 : {
729 2 : (void) SignalRecoveryConflictWithVirtualXID(*backends,
730 : RECOVERY_CONFLICT_STARTUP_DEADLOCK);
731 2 : backends++;
732 : }
733 :
734 : /*
735 : * Exit if the recovery conflict has not been logged yet even though
736 : * logging is enabled, so that the caller can log that. Then
737 : * RecoveryConflictWithLock() is called again and we will wait again
738 : * for the lock to be released.
739 : */
740 2 : if (logging_conflict)
741 1 : goto cleanup;
742 :
743 : /*
744 : * Wait again here to be signaled by the release of the Relation Lock,
745 : * to prevent the subsequent RecoveryConflictWithLock() from causing
746 : * deadlock_timeout and sending a request for deadlocks check again.
747 : * Otherwise the request continues to be sent every deadlock_timeout
748 : * until the relation locks are released or ltime is reached.
749 : */
750 1 : got_standby_deadlock_timeout = false;
751 1 : ProcWaitForSignal(PG_WAIT_LOCK | locktag.locktag_type);
752 : }
753 :
754 1 : cleanup:
755 :
756 : /*
757 : * Clear any timeout requests established above. We assume here that the
758 : * Startup process doesn't have any other outstanding timeouts than those
759 : * used by this function. If that stops being true, we could cancel the
760 : * timeouts individually, but that'd be slower.
761 : */
762 3 : disable_all_timeouts(false);
763 3 : got_standby_lock_timeout = false;
764 3 : got_standby_deadlock_timeout = false;
765 3 : }
766 :
767 : /*
768 : * ResolveRecoveryConflictWithBufferPin is called from LockBufferForCleanup()
769 : * to resolve conflicts with other backends holding buffer pins.
770 : *
771 : * The ProcWaitForSignal() sleep normally done in LockBufferForCleanup()
772 : * (when not InHotStandby) is performed here, for code clarity.
773 : *
774 : * We either resolve conflicts immediately or set a timeout to wake us at
775 : * the limit of our patience.
776 : *
777 : * Resolve conflicts by sending a PROCSIG signal to all backends to check if
778 : * they hold one of the buffer pins that is blocking Startup process. If so,
779 : * those backends will take an appropriate error action, ERROR or FATAL.
780 : *
781 : * We also must check for deadlocks. Deadlocks occur because if queries
782 : * wait on a lock, that must be behind an AccessExclusiveLock, which can only
783 : * be cleared if the Startup process replays a transaction completion record.
784 : * If Startup process is also waiting then that is a deadlock. The deadlock
785 : * can occur if the query is waiting and then the Startup sleeps, or if
786 : * Startup is sleeping and the query waits on a lock. We protect against
787 : * only the former sequence here, the latter sequence is checked prior to
788 : * the query sleeping, in CheckRecoveryConflictDeadlock().
789 : *
790 : * Deadlocks are extremely rare, and relatively expensive to check for,
791 : * so we don't do a deadlock check right away ... only if we have had to wait
792 : * at least deadlock_timeout.
793 : */
794 : void
795 7 : ResolveRecoveryConflictWithBufferPin(void)
796 : {
797 : TimestampTz ltime;
798 :
799 : Assert(InHotStandby);
800 :
801 7 : ltime = GetStandbyLimitTime();
802 :
803 7 : if (GetCurrentTimestamp() >= ltime && ltime != 0)
804 : {
805 : /*
806 : * We're already behind, so clear a path as quickly as possible.
807 : */
808 2 : SendRecoveryConflictWithBufferPin(RECOVERY_CONFLICT_BUFFERPIN);
809 : }
810 : else
811 : {
812 : /*
813 : * Wake up at ltime, and check for deadlocks as well if we will be
814 : * waiting longer than deadlock_timeout
815 : */
816 : EnableTimeoutParams timeouts[2];
817 5 : int cnt = 0;
818 :
819 5 : if (ltime != 0)
820 : {
821 5 : timeouts[cnt].id = STANDBY_TIMEOUT;
822 5 : timeouts[cnt].type = TMPARAM_AT;
823 5 : timeouts[cnt].fin_time = ltime;
824 5 : cnt++;
825 : }
826 :
827 5 : got_standby_deadlock_timeout = false;
828 5 : timeouts[cnt].id = STANDBY_DEADLOCK_TIMEOUT;
829 5 : timeouts[cnt].type = TMPARAM_AFTER;
830 5 : timeouts[cnt].delay_ms = DeadlockTimeout;
831 5 : cnt++;
832 :
833 5 : enable_timeouts(timeouts, cnt);
834 : }
835 :
836 : /*
837 : * Wait to be signaled by UnpinBuffer() or for the wait to be interrupted
838 : * by one of the timeouts established above.
839 : *
840 : * We assume that only UnpinBuffer() and the timeout requests established
841 : * above can wake us up here. WakeupRecovery() called by walreceiver or
842 : * SIGHUP signal handler, etc cannot do that because it uses the different
843 : * latch from that ProcWaitForSignal() waits on.
844 : */
845 7 : ProcWaitForSignal(WAIT_EVENT_BUFFER_CLEANUP);
846 :
847 7 : if (got_standby_delay_timeout)
848 1 : SendRecoveryConflictWithBufferPin(RECOVERY_CONFLICT_BUFFERPIN);
849 6 : else if (got_standby_deadlock_timeout)
850 : {
851 : /*
852 : * Send out a request for hot-standby backends to check themselves for
853 : * deadlocks.
854 : *
855 : * XXX The subsequent ResolveRecoveryConflictWithBufferPin() will wait
856 : * to be signaled by UnpinBuffer() again and send a request for
857 : * deadlocks check if deadlock_timeout happens. This causes the
858 : * request to continue to be sent every deadlock_timeout until the
859 : * buffer is unpinned or ltime is reached. This would increase the
860 : * workload in the startup process and backends. In practice it may
861 : * not be so harmful because the period that the buffer is kept pinned
862 : * is basically no so long. But we should fix this?
863 : */
864 2 : SendRecoveryConflictWithBufferPin(RECOVERY_CONFLICT_BUFFERPIN_DEADLOCK);
865 : }
866 :
867 : /*
868 : * Clear any timeout requests established above. We assume here that the
869 : * Startup process doesn't have any other timeouts than what this function
870 : * uses. If that stops being true, we could cancel the timeouts
871 : * individually, but that'd be slower.
872 : */
873 7 : disable_all_timeouts(false);
874 7 : got_standby_delay_timeout = false;
875 7 : got_standby_deadlock_timeout = false;
876 7 : }
877 :
878 : static void
879 5 : SendRecoveryConflictWithBufferPin(RecoveryConflictReason reason)
880 : {
881 : Assert(reason == RECOVERY_CONFLICT_BUFFERPIN ||
882 : reason == RECOVERY_CONFLICT_BUFFERPIN_DEADLOCK);
883 :
884 : /*
885 : * We send signal to all backends to ask them if they are holding the
886 : * buffer pin which is delaying the Startup process. Most of them will be
887 : * innocent, but we let the SIGUSR1 handling in each backend decide their
888 : * own fate.
889 : */
890 5 : SignalRecoveryConflictWithDatabase(InvalidOid, reason);
891 5 : }
892 :
893 : /*
894 : * In Hot Standby perform early deadlock detection. We abort the lock
895 : * wait if we are about to sleep while holding the buffer pin that Startup
896 : * process is waiting for.
897 : *
898 : * Note: this code is pessimistic, because there is no way for it to
899 : * determine whether an actual deadlock condition is present: the lock we
900 : * need to wait for might be unrelated to any held by the Startup process.
901 : * Sooner or later, this mechanism should get ripped out in favor of somehow
902 : * accounting for buffer locks in DeadLockCheck(). However, errors here
903 : * seem to be very low-probability in practice, so for now it's not worth
904 : * the trouble.
905 : */
906 : void
907 1 : CheckRecoveryConflictDeadlock(void)
908 : {
909 : Assert(!InRecovery); /* do not call in Startup process */
910 :
911 1 : if (!HoldingBufferPinThatDelaysRecovery())
912 1 : return;
913 :
914 : /*
915 : * Error message should match ProcessInterrupts() but we avoid calling
916 : * that because we aren't handling an interrupt at this point. Note that
917 : * we only cancel the current transaction here, so if we are in a
918 : * subtransaction and the pin is held by a parent, then the Startup
919 : * process will continue to wait even though we have avoided deadlock.
920 : */
921 0 : ereport(ERROR,
922 : (errcode(ERRCODE_T_R_DEADLOCK_DETECTED),
923 : errmsg("canceling statement due to conflict with recovery"),
924 : errdetail("User transaction caused buffer deadlock with recovery.")));
925 : }
926 :
927 :
928 : /* --------------------------------
929 : * timeout handler routines
930 : * --------------------------------
931 : */
932 :
933 : /*
934 : * StandbyDeadLockHandler() will be called if STANDBY_DEADLOCK_TIMEOUT is
935 : * exceeded.
936 : */
937 : void
938 4 : StandbyDeadLockHandler(void)
939 : {
940 4 : got_standby_deadlock_timeout = true;
941 4 : }
942 :
943 : /*
944 : * StandbyTimeoutHandler() will be called if STANDBY_TIMEOUT is exceeded.
945 : */
946 : void
947 1 : StandbyTimeoutHandler(void)
948 : {
949 1 : got_standby_delay_timeout = true;
950 1 : }
951 :
952 : /*
953 : * StandbyLockTimeoutHandler() will be called if STANDBY_LOCK_TIMEOUT is exceeded.
954 : */
955 : void
956 1 : StandbyLockTimeoutHandler(void)
957 : {
958 1 : got_standby_lock_timeout = true;
959 1 : }
960 :
961 : /*
962 : * -----------------------------------------------------
963 : * Locking in Recovery Mode
964 : * -----------------------------------------------------
965 : *
966 : * All locks are held by the Startup process using a single virtual
967 : * transaction. This implementation is both simpler and in some senses,
968 : * more correct. The locks held mean "some original transaction held
969 : * this lock, so query access is not allowed at this time". So the Startup
970 : * process is the proxy by which the original locks are implemented.
971 : *
972 : * We only keep track of AccessExclusiveLocks, which are only ever held by
973 : * one transaction on one relation.
974 : *
975 : * We keep a table of known locks in the RecoveryLockHash hash table.
976 : * The point of that table is to let us efficiently de-duplicate locks,
977 : * which is important because checkpoints will re-report the same locks
978 : * already held. There is also a RecoveryLockXidHash table with one entry
979 : * per xid, which allows us to efficiently find all the locks held by a
980 : * given original transaction.
981 : *
982 : * We use session locks rather than normal locks so we don't need
983 : * ResourceOwners.
984 : */
985 :
986 :
987 : void
988 29267 : StandbyAcquireAccessExclusiveLock(TransactionId xid, Oid dbOid, Oid relOid)
989 : {
990 : RecoveryLockXidEntry *xidentry;
991 : RecoveryLockEntry *lockentry;
992 : xl_standby_lock key;
993 : LOCKTAG locktag;
994 : bool found;
995 :
996 : /* Already processed? */
997 58534 : if (!TransactionIdIsValid(xid) ||
998 58515 : TransactionIdDidCommit(xid) ||
999 29248 : TransactionIdDidAbort(xid))
1000 21 : return;
1001 :
1002 29246 : elog(DEBUG4, "adding recovery lock: db %u rel %u", dbOid, relOid);
1003 :
1004 : /* dbOid is InvalidOid when we are locking a shared relation. */
1005 : Assert(OidIsValid(relOid));
1006 :
1007 : /* Create a hash entry for this xid, if we don't have one already. */
1008 29246 : xidentry = hash_search(RecoveryLockXidHash, &xid, HASH_ENTER, &found);
1009 29246 : if (!found)
1010 : {
1011 : Assert(xidentry->xid == xid); /* dynahash should have set this */
1012 11632 : xidentry->head = NULL;
1013 : }
1014 :
1015 : /* Create a hash entry for this lock, unless we have one already. */
1016 29246 : key.xid = xid;
1017 29246 : key.dbOid = dbOid;
1018 29246 : key.relOid = relOid;
1019 29246 : lockentry = hash_search(RecoveryLockHash, &key, HASH_ENTER, &found);
1020 29246 : if (!found)
1021 : {
1022 : /* It's new, so link it into the XID's list ... */
1023 27431 : lockentry->next = xidentry->head;
1024 27431 : xidentry->head = lockentry;
1025 :
1026 : /* ... and acquire the lock locally. */
1027 27431 : SET_LOCKTAG_RELATION(locktag, dbOid, relOid);
1028 :
1029 27431 : (void) LockAcquire(&locktag, AccessExclusiveLock, true, false);
1030 : }
1031 : }
1032 :
1033 : /*
1034 : * Release all the locks associated with this RecoveryLockXidEntry.
1035 : */
1036 : static void
1037 11632 : StandbyReleaseXidEntryLocks(RecoveryLockXidEntry *xidentry)
1038 : {
1039 : RecoveryLockEntry *entry;
1040 : RecoveryLockEntry *next;
1041 :
1042 39063 : for (entry = xidentry->head; entry != NULL; entry = next)
1043 : {
1044 : LOCKTAG locktag;
1045 :
1046 27431 : elog(DEBUG4,
1047 : "releasing recovery lock: xid %u db %u rel %u",
1048 : entry->key.xid, entry->key.dbOid, entry->key.relOid);
1049 : /* Release the lock ... */
1050 27431 : SET_LOCKTAG_RELATION(locktag, entry->key.dbOid, entry->key.relOid);
1051 27431 : if (!LockRelease(&locktag, AccessExclusiveLock, true))
1052 : {
1053 0 : elog(LOG,
1054 : "RecoveryLockHash contains entry for lock no longer recorded by lock manager: xid %u database %u relation %u",
1055 : entry->key.xid, entry->key.dbOid, entry->key.relOid);
1056 : Assert(false);
1057 : }
1058 : /* ... and remove the per-lock hash entry */
1059 27431 : next = entry->next;
1060 27431 : hash_search(RecoveryLockHash, entry, HASH_REMOVE, NULL);
1061 : }
1062 :
1063 11632 : xidentry->head = NULL; /* just for paranoia */
1064 11632 : }
1065 :
1066 : /*
1067 : * Release locks for specific XID, or all locks if it's InvalidXid.
1068 : */
1069 : static void
1070 12311 : StandbyReleaseLocks(TransactionId xid)
1071 : {
1072 : RecoveryLockXidEntry *entry;
1073 :
1074 12311 : if (TransactionIdIsValid(xid))
1075 : {
1076 12311 : if ((entry = hash_search(RecoveryLockXidHash, &xid, HASH_FIND, NULL)))
1077 : {
1078 11632 : StandbyReleaseXidEntryLocks(entry);
1079 11632 : hash_search(RecoveryLockXidHash, entry, HASH_REMOVE, NULL);
1080 : }
1081 : }
1082 : else
1083 0 : StandbyReleaseAllLocks();
1084 12311 : }
1085 :
1086 : /*
1087 : * Release locks for a transaction tree, starting at xid down, from
1088 : * RecoveryLockXidHash.
1089 : *
1090 : * Called during WAL replay of COMMIT/ROLLBACK when in hot standby mode,
1091 : * to remove any AccessExclusiveLocks requested by a transaction.
1092 : */
1093 : void
1094 11811 : StandbyReleaseLockTree(TransactionId xid, int nsubxids, TransactionId *subxids)
1095 : {
1096 : int i;
1097 :
1098 11811 : StandbyReleaseLocks(xid);
1099 :
1100 12311 : for (i = 0; i < nsubxids; i++)
1101 500 : StandbyReleaseLocks(subxids[i]);
1102 11811 : }
1103 :
1104 : /*
1105 : * Called at end of recovery and when we see a shutdown checkpoint.
1106 : */
1107 : void
1108 116 : StandbyReleaseAllLocks(void)
1109 : {
1110 : HASH_SEQ_STATUS status;
1111 : RecoveryLockXidEntry *entry;
1112 :
1113 116 : elog(DEBUG2, "release all standby locks");
1114 :
1115 116 : hash_seq_init(&status, RecoveryLockXidHash);
1116 116 : while ((entry = hash_seq_search(&status)))
1117 : {
1118 0 : StandbyReleaseXidEntryLocks(entry);
1119 0 : hash_search(RecoveryLockXidHash, entry, HASH_REMOVE, NULL);
1120 : }
1121 116 : }
1122 :
1123 : /*
1124 : * StandbyReleaseOldLocks
1125 : * Release standby locks held by top-level XIDs that aren't running,
1126 : * as long as they're not prepared transactions.
1127 : *
1128 : * This is needed to prune the locks of crashed transactions, which didn't
1129 : * write an ABORT/COMMIT record.
1130 : */
1131 : void
1132 836 : StandbyReleaseOldLocks(TransactionId oldxid)
1133 : {
1134 : HASH_SEQ_STATUS status;
1135 : RecoveryLockXidEntry *entry;
1136 :
1137 836 : hash_seq_init(&status, RecoveryLockXidHash);
1138 1136 : while ((entry = hash_seq_search(&status)))
1139 : {
1140 : Assert(TransactionIdIsValid(entry->xid));
1141 :
1142 : /* Skip if prepared transaction. */
1143 300 : if (StandbyTransactionIdIsPrepared(entry->xid))
1144 0 : continue;
1145 :
1146 : /* Skip if >= oldxid. */
1147 300 : if (!TransactionIdPrecedes(entry->xid, oldxid))
1148 300 : continue;
1149 :
1150 : /* Remove all locks and hash table entry. */
1151 0 : StandbyReleaseXidEntryLocks(entry);
1152 0 : hash_search(RecoveryLockXidHash, entry, HASH_REMOVE, NULL);
1153 : }
1154 836 : }
1155 :
1156 : /*
1157 : * --------------------------------------------------------------------
1158 : * Recovery handling for Rmgr RM_STANDBY_ID
1159 : *
1160 : * These record types will only be created if XLogStandbyInfoActive()
1161 : * --------------------------------------------------------------------
1162 : */
1163 :
1164 : void
1165 29221 : standby_redo(XLogReaderState *record)
1166 : {
1167 29221 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
1168 :
1169 : /* Backup blocks are not used in standby records */
1170 : Assert(!XLogRecHasAnyBlockRefs(record));
1171 :
1172 : /* Do nothing if we're not in hot standby mode */
1173 29221 : if (standbyState == STANDBY_DISABLED)
1174 161 : return;
1175 :
1176 29060 : if (info == XLOG_STANDBY_LOCK)
1177 : {
1178 27603 : xl_standby_locks *xlrec = (xl_standby_locks *) XLogRecGetData(record);
1179 : int i;
1180 :
1181 56870 : for (i = 0; i < xlrec->nlocks; i++)
1182 29267 : StandbyAcquireAccessExclusiveLock(xlrec->locks[i].xid,
1183 : xlrec->locks[i].dbOid,
1184 : xlrec->locks[i].relOid);
1185 : }
1186 1457 : else if (info == XLOG_RUNNING_XACTS)
1187 : {
1188 771 : xl_running_xacts *xlrec = (xl_running_xacts *) XLogRecGetData(record);
1189 : RunningTransactionsData running;
1190 :
1191 : /*
1192 : * Records issued for specific database are not suitable for physical
1193 : * replication because that affects the whole cluster. In particular,
1194 : * the list of XID is probably incomplete here.
1195 : */
1196 771 : if (OidIsValid(xlrec->dbid))
1197 0 : return;
1198 :
1199 771 : running.xcnt = xlrec->xcnt;
1200 771 : running.subxcnt = xlrec->subxcnt;
1201 771 : running.subxid_status = xlrec->subxid_overflow ? SUBXIDS_MISSING : SUBXIDS_IN_ARRAY;
1202 771 : running.nextXid = xlrec->nextXid;
1203 771 : running.latestCompletedXid = xlrec->latestCompletedXid;
1204 771 : running.oldestRunningXid = xlrec->oldestRunningXid;
1205 771 : running.xids = xlrec->xids;
1206 :
1207 771 : ProcArrayApplyRecoveryInfo(&running);
1208 :
1209 : /*
1210 : * The startup process currently has no convenient way to schedule
1211 : * stats to be reported. XLOG_RUNNING_XACTS records issued at a
1212 : * regular cadence, making this a convenient location to report stats.
1213 : * While these records aren't generated with wal_level=minimal, stats
1214 : * also cannot be accessed during WAL replay.
1215 : */
1216 771 : pgstat_report_stat(true);
1217 : }
1218 686 : else if (info == XLOG_INVALIDATIONS)
1219 : {
1220 686 : xl_invalidations *xlrec = (xl_invalidations *) XLogRecGetData(record);
1221 :
1222 686 : ProcessCommittedInvalidationMessages(xlrec->msgs,
1223 : xlrec->nmsgs,
1224 686 : xlrec->relcacheInitFileInval,
1225 : xlrec->dbId,
1226 : xlrec->tsId);
1227 : }
1228 : else
1229 0 : elog(PANIC, "standby_redo: unknown op code %u", info);
1230 : }
1231 :
1232 : /*
1233 : * Log details of the current snapshot to WAL. This allows the snapshot state
1234 : * to be reconstructed on the standby and for logical decoding.
1235 : *
1236 : * This is used for Hot Standby as follows:
1237 : *
1238 : * We can move directly to STANDBY_SNAPSHOT_READY at startup if we
1239 : * start from a shutdown checkpoint because we know nothing was running
1240 : * at that time and our recovery snapshot is known empty. In the more
1241 : * typical case of an online checkpoint we need to jump through a few
1242 : * hoops to get a correct recovery snapshot and this requires a two or
1243 : * sometimes a three stage process.
1244 : *
1245 : * The initial snapshot must contain all running xids and all current
1246 : * AccessExclusiveLocks at a point in time on the standby. Assembling
1247 : * that information while the server is running requires many and
1248 : * various LWLocks, so we choose to derive that information piece by
1249 : * piece and then re-assemble that info on the standby. When that
1250 : * information is fully assembled we move to STANDBY_SNAPSHOT_READY.
1251 : *
1252 : * Since locking on the primary when we derive the information is not
1253 : * strict, we note that there is a time window between the derivation and
1254 : * writing to WAL of the derived information. That allows race conditions
1255 : * that we must resolve, since xids and locks may enter or leave the
1256 : * snapshot during that window. This creates the issue that an xid or
1257 : * lock may start *after* the snapshot has been derived yet *before* the
1258 : * snapshot is logged in the running xacts WAL record. We resolve this by
1259 : * starting to accumulate changes at a point just prior to when we derive
1260 : * the snapshot on the primary, then ignore duplicates when we later apply
1261 : * the snapshot from the running xacts record. This is implemented during
1262 : * CreateCheckPoint() where we use the logical checkpoint location as
1263 : * our starting point and then write the running xacts record immediately
1264 : * before writing the main checkpoint WAL record. Since we always start
1265 : * up from a checkpoint and are immediately at our starting point, we
1266 : * unconditionally move to STANDBY_INITIALIZED. After this point we
1267 : * must do 4 things:
1268 : * * move shared nextXid forwards as we see new xids
1269 : * * extend the clog and subtrans with each new xid
1270 : * * keep track of uncommitted known assigned xids
1271 : * * keep track of uncommitted AccessExclusiveLocks
1272 : *
1273 : * When we see a commit/abort we must remove known assigned xids and locks
1274 : * from the completing transaction. Attempted removals that cannot locate
1275 : * an entry are expected and must not cause an error when we are in state
1276 : * STANDBY_INITIALIZED. This is implemented in StandbyReleaseLocks() and
1277 : * KnownAssignedXidsRemove().
1278 : *
1279 : * Later, when we apply the running xact data we must be careful to ignore
1280 : * transactions already committed, since those commits raced ahead when
1281 : * making WAL entries.
1282 : *
1283 : * For logical decoding only the running xacts information is needed;
1284 : * there's no need to look at the locking information, but it's logged anyway,
1285 : * as there's no independent knob to just enable logical decoding. For
1286 : * details of how this is used, check snapbuild.c's introductory comment.
1287 : *
1288 : * If 'dbid' is valid, only gather transactions running in that
1289 : * database. snapbuild.c can use such running xacts information for faster
1290 : * startup, but it still needs normal (cluster-wide) during the actual
1291 : * decoding - see standby_decode() and SnapBuildProcessRunningXacts() for
1292 : * details. Other processes (e.g. checkpointer) issue the cluster-wide records
1293 : * whether logical decoding is active or not.
1294 : *
1295 : * Please be careful about using this argument for other purposes. In
1296 : * particular, physical replication *must* ignore the database-specific
1297 : * records, exactly because they do not cover the whole cluster - see
1298 : * standby_redo().
1299 : *
1300 : * Returns the RecPtr of the last inserted record.
1301 : */
1302 : XLogRecPtr
1303 1514 : LogStandbySnapshot(Oid dbid)
1304 : {
1305 : XLogRecPtr recptr;
1306 : RunningTransactions running;
1307 : xl_standby_lock *locks;
1308 : int nlocks;
1309 1514 : bool logical_decoding_enabled = IsLogicalDecodingEnabled();
1310 :
1311 : Assert(XLogStandbyInfoActive());
1312 :
1313 : #ifdef USE_INJECTION_POINTS
1314 1514 : if (IS_INJECTION_POINT_ATTACHED("skip-log-running-xacts"))
1315 : {
1316 : /*
1317 : * This record could move slot's xmin forward during decoding, leading
1318 : * to unpredictable results, so skip it when requested by the test.
1319 : */
1320 0 : return GetInsertRecPtr();
1321 : }
1322 : #endif
1323 :
1324 : /*
1325 : * Get details of any AccessExclusiveLocks being held at the moment.
1326 : */
1327 1514 : locks = GetRunningTransactionLocks(&nlocks);
1328 1514 : if (nlocks > 0)
1329 155 : LogAccessExclusiveLocks(nlocks, locks);
1330 1514 : pfree(locks);
1331 :
1332 : /*
1333 : * Log details of all in-progress transactions. This should be the last
1334 : * record we write, because standby will open up when it sees this.
1335 : */
1336 1514 : running = GetRunningTransactionData(dbid);
1337 :
1338 : /*
1339 : * GetRunningTransactionData() acquired ProcArrayLock, we must release it.
1340 : * For Hot Standby this can be done before inserting the WAL record
1341 : * because ProcArrayApplyRecoveryInfo() rechecks the commit status using
1342 : * the clog. For logical decoding, though, the lock can't be released
1343 : * early because the clog might be "in the future" from the POV of the
1344 : * historic snapshot. This would allow for situations where we're waiting
1345 : * for the end of a transaction listed in the xl_running_xacts record
1346 : * which, according to the WAL, has committed before the xl_running_xacts
1347 : * record. Fortunately this routine isn't executed frequently, and it's
1348 : * only a shared lock.
1349 : */
1350 1514 : if (!logical_decoding_enabled)
1351 952 : LWLockRelease(ProcArrayLock);
1352 :
1353 1514 : recptr = LogCurrentRunningXacts(running);
1354 :
1355 : /* Release lock if we kept it longer ... */
1356 1514 : if (logical_decoding_enabled)
1357 562 : LWLockRelease(ProcArrayLock);
1358 :
1359 : /* GetRunningTransactionData() acquired XidGenLock, we must release it */
1360 1514 : LWLockRelease(XidGenLock);
1361 :
1362 1514 : return recptr;
1363 : }
1364 :
1365 : /*
1366 : * Record an enhanced snapshot of running transactions into WAL.
1367 : *
1368 : * The definitions of RunningTransactionsData and xl_running_xacts are
1369 : * similar. We keep them separate because xl_running_xacts is a contiguous
1370 : * chunk of memory and never exists fully until it is assembled in WAL.
1371 : * The inserted records are marked as not being important for durability,
1372 : * to avoid triggering superfluous checkpoint / archiving activity.
1373 : */
1374 : static XLogRecPtr
1375 1514 : LogCurrentRunningXacts(RunningTransactions CurrRunningXacts)
1376 : {
1377 : xl_running_xacts xlrec;
1378 : XLogRecPtr recptr;
1379 :
1380 1514 : xlrec.dbid = CurrRunningXacts->dbid;
1381 1514 : xlrec.xcnt = CurrRunningXacts->xcnt;
1382 1514 : xlrec.subxcnt = CurrRunningXacts->subxcnt;
1383 1514 : xlrec.subxid_overflow = (CurrRunningXacts->subxid_status != SUBXIDS_IN_ARRAY);
1384 1514 : xlrec.nextXid = CurrRunningXacts->nextXid;
1385 1514 : xlrec.oldestRunningXid = CurrRunningXacts->oldestRunningXid;
1386 1514 : xlrec.latestCompletedXid = CurrRunningXacts->latestCompletedXid;
1387 :
1388 : /* Header */
1389 1514 : XLogBeginInsert();
1390 1514 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
1391 1514 : XLogRegisterData(&xlrec, MinSizeOfXactRunningXacts);
1392 :
1393 : /* array of TransactionIds */
1394 1514 : if (xlrec.xcnt > 0)
1395 497 : XLogRegisterData(CurrRunningXacts->xids,
1396 497 : (xlrec.xcnt + xlrec.subxcnt) * sizeof(TransactionId));
1397 :
1398 1514 : recptr = XLogInsert(RM_STANDBY_ID, XLOG_RUNNING_XACTS);
1399 :
1400 1514 : if (xlrec.subxid_overflow)
1401 2 : elog(DEBUG2,
1402 : "snapshot of %d running transactions overflowed (lsn %X/%08X oldest xid %u latest complete %u next xid %u)",
1403 : CurrRunningXacts->xcnt,
1404 : LSN_FORMAT_ARGS(recptr),
1405 : CurrRunningXacts->oldestRunningXid,
1406 : CurrRunningXacts->latestCompletedXid,
1407 : CurrRunningXacts->nextXid);
1408 : else
1409 1512 : elog(DEBUG2,
1410 : "snapshot of %d+%d running transaction ids (lsn %X/%08X oldest xid %u latest complete %u next xid %u)",
1411 : CurrRunningXacts->xcnt, CurrRunningXacts->subxcnt,
1412 : LSN_FORMAT_ARGS(recptr),
1413 : CurrRunningXacts->oldestRunningXid,
1414 : CurrRunningXacts->latestCompletedXid,
1415 : CurrRunningXacts->nextXid);
1416 :
1417 : /*
1418 : * Ensure running_xacts information is synced to disk not too far in the
1419 : * future. We don't want to stall anything though (i.e. use XLogFlush()),
1420 : * so we let the wal writer do it during normal operation.
1421 : * XLogSetAsyncXactLSN() conveniently will mark the LSN as to-be-synced
1422 : * and nudge the WALWriter into action if sleeping. Check
1423 : * XLogBackgroundFlush() for details why a record might not be flushed
1424 : * without it.
1425 : */
1426 1514 : XLogSetAsyncXactLSN(recptr);
1427 :
1428 1514 : return recptr;
1429 : }
1430 :
1431 : /*
1432 : * Wholesale logging of AccessExclusiveLocks. Other lock types need not be
1433 : * logged, as described in backend/storage/lmgr/README.
1434 : */
1435 : static void
1436 171842 : LogAccessExclusiveLocks(int nlocks, xl_standby_lock *locks)
1437 : {
1438 : xl_standby_locks xlrec;
1439 :
1440 171842 : xlrec.nlocks = nlocks;
1441 :
1442 171842 : XLogBeginInsert();
1443 171842 : XLogRegisterData(&xlrec, offsetof(xl_standby_locks, locks));
1444 171842 : XLogRegisterData(locks, nlocks * sizeof(xl_standby_lock));
1445 171842 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
1446 :
1447 171842 : (void) XLogInsert(RM_STANDBY_ID, XLOG_STANDBY_LOCK);
1448 171842 : }
1449 :
1450 : /*
1451 : * Individual logging of AccessExclusiveLocks for use during LockAcquire()
1452 : */
1453 : void
1454 171687 : LogAccessExclusiveLock(Oid dbOid, Oid relOid)
1455 : {
1456 : xl_standby_lock xlrec;
1457 :
1458 171687 : xlrec.xid = GetCurrentTransactionId();
1459 :
1460 171687 : xlrec.dbOid = dbOid;
1461 171687 : xlrec.relOid = relOid;
1462 :
1463 171687 : LogAccessExclusiveLocks(1, &xlrec);
1464 171687 : MyXactFlags |= XACT_FLAGS_ACQUIREDACCESSEXCLUSIVELOCK;
1465 171687 : }
1466 :
1467 : /*
1468 : * Prepare to log an AccessExclusiveLock, for use during LockAcquire()
1469 : */
1470 : void
1471 172004 : LogAccessExclusiveLockPrepare(void)
1472 : {
1473 : /*
1474 : * Ensure that a TransactionId has been assigned to this transaction, for
1475 : * two reasons, both related to lock release on the standby. First, we
1476 : * must assign an xid so that RecordTransactionCommit() and
1477 : * RecordTransactionAbort() do not optimise away the transaction
1478 : * completion record which recovery relies upon to release locks. It's a
1479 : * hack, but for a corner case not worth adding code for into the main
1480 : * commit path. Second, we must assign an xid before the lock is recorded
1481 : * in shared memory, otherwise a concurrently executing
1482 : * GetRunningTransactionLocks() might see a lock associated with an
1483 : * InvalidTransactionId which we later assert cannot happen.
1484 : */
1485 172004 : (void) GetCurrentTransactionId();
1486 172004 : }
1487 :
1488 : /*
1489 : * Emit WAL for invalidations. This currently is only used for commits without
1490 : * an xid but which contain invalidations.
1491 : */
1492 : void
1493 11765 : LogStandbyInvalidations(int nmsgs, SharedInvalidationMessage *msgs,
1494 : bool relcacheInitFileInval)
1495 : {
1496 : xl_invalidations xlrec;
1497 :
1498 : /* prepare record */
1499 11765 : memset(&xlrec, 0, sizeof(xlrec));
1500 11765 : xlrec.dbId = MyDatabaseId;
1501 11765 : xlrec.tsId = MyDatabaseTableSpace;
1502 11765 : xlrec.relcacheInitFileInval = relcacheInitFileInval;
1503 11765 : xlrec.nmsgs = nmsgs;
1504 :
1505 : /* perform insertion */
1506 11765 : XLogBeginInsert();
1507 11765 : XLogRegisterData(&xlrec, MinSizeOfInvalidations);
1508 11765 : XLogRegisterData(msgs,
1509 : nmsgs * sizeof(SharedInvalidationMessage));
1510 11765 : XLogInsert(RM_STANDBY_ID, XLOG_INVALIDATIONS);
1511 11765 : }
1512 :
1513 : /* Return the description of recovery conflict */
1514 : static const char *
1515 10 : get_recovery_conflict_desc(RecoveryConflictReason reason)
1516 : {
1517 10 : const char *reasonDesc = _("unknown reason");
1518 :
1519 10 : switch (reason)
1520 : {
1521 4 : case RECOVERY_CONFLICT_BUFFERPIN:
1522 4 : reasonDesc = _("recovery conflict on buffer pin");
1523 4 : break;
1524 2 : case RECOVERY_CONFLICT_LOCK:
1525 2 : reasonDesc = _("recovery conflict on lock");
1526 2 : break;
1527 2 : case RECOVERY_CONFLICT_TABLESPACE:
1528 2 : reasonDesc = _("recovery conflict on tablespace");
1529 2 : break;
1530 2 : case RECOVERY_CONFLICT_SNAPSHOT:
1531 2 : reasonDesc = _("recovery conflict on snapshot");
1532 2 : break;
1533 0 : case RECOVERY_CONFLICT_LOGICALSLOT:
1534 0 : reasonDesc = _("recovery conflict on replication slot");
1535 0 : break;
1536 0 : case RECOVERY_CONFLICT_STARTUP_DEADLOCK:
1537 0 : reasonDesc = _("recovery conflict on deadlock");
1538 0 : break;
1539 0 : case RECOVERY_CONFLICT_BUFFERPIN_DEADLOCK:
1540 0 : reasonDesc = _("recovery conflict on buffer deadlock");
1541 0 : break;
1542 0 : case RECOVERY_CONFLICT_DATABASE:
1543 0 : reasonDesc = _("recovery conflict on database");
1544 0 : break;
1545 : }
1546 :
1547 10 : return reasonDesc;
1548 : }
|
