Age Owner Branch data TLA 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
5930 simon@2ndQuadrant.co 96 :CBC 114 : InitRecoveryTransactionEnvironment(void)
97 : : {
98 : : VirtualTransactionId vxid;
99 : : HASHCTL hash_ctl;
100 : :
1256 tgl@sss.pgh.pa.us 101 [ - + ]: 114 : 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 : 114 : hash_ctl.keysize = sizeof(xl_standby_lock);
108 : 114 : hash_ctl.entrysize = sizeof(RecoveryLockEntry);
109 : 114 : RecoveryLockHash = hash_create("RecoveryLockHash",
110 : : 64,
111 : : &hash_ctl,
112 : : HASH_ELEM | HASH_BLOBS);
2819 tmunro@postgresql.or 113 : 114 : hash_ctl.keysize = sizeof(TransactionId);
1256 tgl@sss.pgh.pa.us 114 : 114 : hash_ctl.entrysize = sizeof(RecoveryLockXidEntry);
115 : 114 : 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 : : */
5930 simon@2ndQuadrant.co 126 : 114 : 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 : : */
742 heikki.linnakangas@i 140 : 114 : MyProc->vxid.procNumber = MyProcNumber;
141 : 114 : vxid.procNumber = MyProcNumber;
5930 simon@2ndQuadrant.co 142 : 114 : vxid.localTransactionId = GetNextLocalTransactionId();
143 : 114 : VirtualXactLockTableInsert(vxid);
144 : :
145 : 114 : standbyState = STANDBY_INITIALIZED;
146 : 114 : }
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 : 169 : 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 : : */
1256 tgl@sss.pgh.pa.us 170 [ + + ]: 169 : if (RecoveryLockHash == NULL)
1804 fujii@postgresql.org 171 : 55 : return;
172 : :
173 : : /* Mark all tracked in-progress transactions as finished. */
5930 simon@2ndQuadrant.co 174 : 114 : ExpireAllKnownAssignedTransactionIds();
175 : :
176 : : /* Release all locks the tracked transactions were holding */
177 : 114 : StandbyReleaseAllLocks();
178 : :
179 : : /* Destroy the lock hash tables. */
1256 tgl@sss.pgh.pa.us 180 : 114 : hash_destroy(RecoveryLockHash);
181 : 114 : hash_destroy(RecoveryLockXidHash);
182 : 114 : RecoveryLockHash = NULL;
183 : 114 : RecoveryLockXidHash = NULL;
184 : :
185 : : /* Cleanup our VirtualTransaction */
4854 simon@2ndQuadrant.co 186 : 114 : 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
5734 tgl@sss.pgh.pa.us 202 : 28 : 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 : 28 : GetXLogReceiptTime(&rtime, &fromStream);
212 [ + - ]: 28 : if (fromStream)
213 : : {
214 [ - + ]: 28 : if (max_standby_streaming_delay < 0)
5734 tgl@sss.pgh.pa.us 215 :UBC 0 : return 0; /* wait forever */
5734 tgl@sss.pgh.pa.us 216 :CBC 28 : return TimestampTzPlusMilliseconds(rtime, max_standby_streaming_delay);
217 : : }
218 : : else
219 : : {
5734 tgl@sss.pgh.pa.us 220 [ # # ]:UBC 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
2172 fujii@postgresql.org 235 :CBC 15 : WaitExceedsMaxStandbyDelay(uint32 wait_event_info)
236 : : {
237 : : TimestampTz ltime;
238 : :
3335 simon@2ndQuadrant.co 239 [ - + ]: 15 : CHECK_FOR_INTERRUPTS();
240 : :
241 : : /* Are we past the limit time? */
5734 tgl@sss.pgh.pa.us 242 : 15 : ltime = GetStandbyLimitTime();
243 [ + - + + ]: 15 : if (ltime && GetCurrentTimestamp() >= ltime)
5930 simon@2ndQuadrant.co 244 : 3 : return true;
245 : :
246 : : /*
247 : : * Sleep a bit (this is essential to avoid busy-waiting).
248 : : */
2172 fujii@postgresql.org 249 : 12 : pgstat_report_wait_start(wait_event_info);
5930 simon@2ndQuadrant.co 250 : 12 : pg_usleep(standbyWait_us);
2172 fujii@postgresql.org 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 : : */
5930 simon@2ndQuadrant.co 257 : 12 : standbyWait_us *= 2;
258 [ - + ]: 12 : if (standbyWait_us > 1000000)
5930 simon@2ndQuadrant.co 259 :UBC 0 : standbyWait_us = 1000000;
260 : :
5930 simon@2ndQuadrant.co 261 :CBC 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
33 heikki.linnakangas@i 275 :GNC 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;
1892 fujii@postgresql.org 283 :CBC 10 : int nprocs = 0;
284 : :
285 : : /*
286 : : * There must be no conflicting processes when the recovery conflict has
287 : : * already been resolved.
288 : : */
1887 289 [ + + - + ]: 10 : Assert(still_waiting || wait_list == NULL);
290 : :
1892 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 : : {
742 heikki.linnakangas@i 303 : 3 : PGPROC *proc = ProcNumberGetProc(vxids->procNumber);
304 : :
305 : : /* proc can be NULL if the target backend is not active */
1892 fujii@postgresql.org 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
1892 fujii@postgresql.org 314 :UBC 0 : appendStringInfo(&buf, ", %d", proc->pid);
315 : :
1892 fujii@postgresql.org 316 :CBC 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 : : */
1887 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 : :
1892 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
5930 simon@2ndQuadrant.co 361 : 13383 : ResolveRecoveryConflictWithVirtualXIDs(VirtualTransactionId *waitlist,
362 : : RecoveryConflictReason reason,
363 : : uint32 wait_event_info,
364 : : bool report_waiting)
365 : : {
2196 fujii@postgresql.org 366 : 13383 : TimestampTz waitStart = 0;
1119 drowley@postgresql.o 367 : 13383 : bool waiting = false;
1892 fujii@postgresql.org 368 : 13383 : bool logged_recovery_conflict = false;
369 : :
370 : : /* Fast exit, to avoid a kernel call if there's no work to be done. */
5567 rhaas@postgresql.org 371 [ + + ]: 13383 : if (!VirtualTransactionIdIsValid(*waitlist))
372 : 13380 : return;
373 : :
374 : : /* Set the wait start timestamp for reporting */
1892 fujii@postgresql.org 375 [ + + - + : 3 : if (report_waiting && (log_recovery_conflict_waits || update_process_title))
- - ]
2196 376 : 2 : waitStart = GetCurrentTimestamp();
377 : :
5567 rhaas@postgresql.org 378 [ + + ]: 6 : while (VirtualTransactionIdIsValid(*waitlist))
379 : : {
380 : : /* reset standbyWait_us for each xact we wait for */
5930 simon@2ndQuadrant.co 381 : 3 : standbyWait_us = STANDBY_INITIAL_WAIT_US;
382 : :
383 : : /* wait until the virtual xid is gone */
5337 rhaas@postgresql.org 384 [ + + ]: 18 : while (!VirtualXactLock(*waitlist, false))
385 : : {
386 : : /* Is it time to kill it? */
2172 fujii@postgresql.org 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 : : */
5930 simon@2ndQuadrant.co 394 [ - + ]: 3 : Assert(VirtualTransactionIdIsValid(*waitlist));
33 heikki.linnakangas@i 395 :GNC 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)
5796 tgl@sss.pgh.pa.us 402 :CBC 3 : pg_usleep(5000L);
403 : : }
404 : :
1119 drowley@postgresql.o 405 [ + + + + : 15 : if (waitStart != 0 && (!logged_recovery_conflict || !waiting))
+ - ]
406 : : {
1892 fujii@postgresql.org 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);
1119 drowley@postgresql.o 412 [ + - + - ]: 14 : maybe_update_title = (update_process_title && !waiting);
413 : :
414 : : /* Get the current timestamp if not report yet */
1892 fujii@postgresql.org 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 : : {
1119 drowley@postgresql.o 425 :UBC 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 : : */
1892 fujii@postgresql.org 433 [ + + + + ]:CBC 21 : if (maybe_log_conflict &&
434 : 7 : TimestampDifferenceExceeds(waitStart, now, DeadlockTimeout))
435 : : {
1887 436 : 2 : LogRecoveryConflict(reason, waitStart, now, waitlist, true);
1892 437 : 2 : logged_recovery_conflict = true;
438 : : }
439 : : }
440 : : }
441 : :
442 : : /* The virtual transaction is gone now, wait for the next one */
5930 simon@2ndQuadrant.co 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 : : */
1887 fujii@postgresql.org 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 */
1119 drowley@postgresql.o 455 [ - + ]: 3 : if (waiting)
1119 drowley@postgresql.o 456 :UBC 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
1214 pg@bowt.ie 470 :CBC 16044 : 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 when replaying already-applied WAL records after a
480 : : * standby crash or restart, or when replaying an XLOG_HEAP2_VISIBLE
481 : : * record that marks as frozen a page which was already all-visible. It's
482 : : * also quite common with records generated during index deletion
483 : : * (original execution of the deletion can reason that a recovery conflict
484 : : * which is sufficient for the deletion operation must take place before
485 : : * replay of the deletion record itself).
486 : : */
487 [ + + ]: 16044 : if (!TransactionIdIsValid(snapshotConflictHorizon))
5599 simon@2ndQuadrant.co 488 : 2663 : return;
489 : :
1168 pg@bowt.ie 490 [ - + ]: 13381 : Assert(TransactionIdIsNormal(snapshotConflictHorizon));
1214 491 : 13381 : backends = GetConflictingVirtualXIDs(snapshotConflictHorizon,
492 : : locator.dbOid);
5904 simon@2ndQuadrant.co 493 : 13381 : ResolveRecoveryConflictWithVirtualXIDs(backends,
494 : : RECOVERY_CONFLICT_SNAPSHOT,
495 : : WAIT_EVENT_RECOVERY_CONFLICT_SNAPSHOT,
496 : : true);
497 : :
498 : : /*
499 : : * Note that WaitExceedsMaxStandbyDelay() is not taken into account here
500 : : * (as opposed to ResolveRecoveryConflictWithVirtualXIDs() above). That
501 : : * seems OK, given that this kind of conflict should not normally be
502 : : * reached, e.g. due to using a physical replication slot.
503 : : */
82 msawada@postgresql.o 504 [ + + + + ]:GNC 13381 : if (IsLogicalDecodingEnabled() && isCatalogRel)
1073 andres@anarazel.de 505 :CBC 21 : InvalidateObsoleteReplicationSlots(RS_INVAL_HORIZON, 0, locator.dbOid,
506 : : snapshotConflictHorizon);
507 : : }
508 : :
509 : : /*
510 : : * Variant of ResolveRecoveryConflictWithSnapshot that works with
511 : : * FullTransactionId values
512 : : */
513 : : void
1214 pg@bowt.ie 514 : 68 : ResolveRecoveryConflictWithSnapshotFullXid(FullTransactionId snapshotConflictHorizon,
515 : : bool isCatalogRel,
516 : : RelFileLocator locator)
517 : : {
518 : : /*
519 : : * ResolveRecoveryConflictWithSnapshot operates on 32-bit TransactionIds,
520 : : * so truncate the logged FullTransactionId. If the logged value is very
521 : : * old, so that XID wrap-around already happened on it, there can't be any
522 : : * snapshots that still see it.
523 : : */
1845 524 : 68 : FullTransactionId nextXid = ReadNextFullTransactionId();
525 : : uint64 diff;
526 : :
527 : 68 : diff = U64FromFullTransactionId(nextXid) -
1214 528 : 68 : U64FromFullTransactionId(snapshotConflictHorizon);
1845 529 [ + - ]: 68 : if (diff < MaxTransactionId / 2)
530 : : {
531 : : TransactionId truncated;
532 : :
1214 533 : 68 : truncated = XidFromFullTransactionId(snapshotConflictHorizon);
1073 andres@anarazel.de 534 : 68 : ResolveRecoveryConflictWithSnapshot(truncated,
535 : : isCatalogRel,
536 : : locator);
537 : : }
1845 pg@bowt.ie 538 : 68 : }
539 : :
540 : : void
5904 simon@2ndQuadrant.co 541 : 1 : ResolveRecoveryConflictWithTablespace(Oid tsid)
542 : : {
543 : : VirtualTransactionId *temp_file_users;
544 : :
545 : : /*
546 : : * Standby users may be currently using this tablespace for their
547 : : * temporary files. We only care about current users because
548 : : * temp_tablespace parameter will just ignore tablespaces that no longer
549 : : * exist.
550 : : *
551 : : * Ask everybody to cancel their queries immediately so we can ensure no
552 : : * temp files remain and we can remove the tablespace. Nuke the entire
553 : : * site from orbit, it's the only way to be sure.
554 : : *
555 : : * XXX: We could work out the pids of active backends using this
556 : : * tablespace by examining the temp filenames in the directory. We would
557 : : * then convert the pids into VirtualXIDs before attempting to cancel
558 : : * them.
559 : : *
560 : : * We don't wait for commit because drop tablespace is non-transactional.
561 : : */
562 : 1 : temp_file_users = GetConflictingVirtualXIDs(InvalidTransactionId,
563 : : InvalidOid);
564 : 1 : ResolveRecoveryConflictWithVirtualXIDs(temp_file_users,
565 : : RECOVERY_CONFLICT_TABLESPACE,
566 : : WAIT_EVENT_RECOVERY_CONFLICT_TABLESPACE,
567 : : true);
568 : 1 : }
569 : :
570 : : void
571 : 14 : ResolveRecoveryConflictWithDatabase(Oid dbid)
572 : : {
573 : : /*
574 : : * We don't do ResolveRecoveryConflictWithVirtualXIDs() here since that
575 : : * only waits for transactions and completely idle sessions would block
576 : : * us. This is rare enough that we do this as simply as possible: no wait,
577 : : * just force them off immediately.
578 : : *
579 : : * No locking is required here because we already acquired
580 : : * AccessExclusiveLock. Anybody trying to connect while we do this will
581 : : * block during InitPostgres() and then disconnect when they see the
582 : : * database has been removed.
583 : : */
584 [ + + ]: 16 : while (CountDBBackends(dbid) > 0)
585 : : {
33 heikki.linnakangas@i 586 :GNC 2 : SignalRecoveryConflictWithDatabase(dbid, RECOVERY_CONFLICT_DATABASE);
587 : :
588 : : /*
589 : : * Wait awhile for them to die so that we avoid flooding an
590 : : * unresponsive backend when system is heavily loaded.
591 : : */
5904 simon@2ndQuadrant.co 592 :CBC 2 : pg_usleep(10000);
593 : : }
594 : 14 : }
595 : :
596 : : /*
597 : : * ResolveRecoveryConflictWithLock is called from ProcSleep()
598 : : * to resolve conflicts with other backends holding relation locks.
599 : : *
600 : : * The WaitLatch sleep normally done in ProcSleep()
601 : : * (when not InHotStandby) is performed here, for code clarity.
602 : : *
603 : : * We either resolve conflicts immediately or set a timeout to wake us at
604 : : * the limit of our patience.
605 : : *
606 : : * Resolve conflicts by canceling to all backends holding a conflicting
607 : : * lock. As we are already queued to be granted the lock, no new lock
608 : : * requests conflicting with ours will be granted in the meantime.
609 : : *
610 : : * We also must check for deadlocks involving the Startup process and
611 : : * hot-standby backend processes. If deadlock_timeout is reached in
612 : : * this function, all the backends holding the conflicting locks are
613 : : * requested to check themselves for deadlocks.
614 : : *
615 : : * logging_conflict should be true if the recovery conflict has not been
616 : : * logged yet even though logging is enabled. After deadlock_timeout is
617 : : * reached and the request for deadlock check is sent, we wait again to
618 : : * be signaled by the release of the lock if logging_conflict is false.
619 : : * Otherwise we return without waiting again so that the caller can report
620 : : * the recovery conflict. In this case, then, this function is called again
621 : : * with logging_conflict=false (because the recovery conflict has already
622 : : * been logged) and we will wait again for the lock to be released.
623 : : */
624 : : void
1892 fujii@postgresql.org 625 : 3 : ResolveRecoveryConflictWithLock(LOCKTAG locktag, bool logging_conflict)
626 : : {
627 : : TimestampTz ltime;
628 : : TimestampTz now;
629 : :
3657 simon@2ndQuadrant.co 630 [ - + ]: 3 : Assert(InHotStandby);
631 : :
632 : 3 : ltime = GetStandbyLimitTime();
1854 fujii@postgresql.org 633 : 3 : now = GetCurrentTimestamp();
634 : :
635 : : /*
636 : : * Update waitStart if first time through after the startup process
637 : : * started waiting for the lock. It should not be updated every time
638 : : * ResolveRecoveryConflictWithLock() is called during the wait.
639 : : *
640 : : * Use the current time obtained for comparison with ltime as waitStart
641 : : * (i.e., the time when this process started waiting for the lock). Since
642 : : * getting the current time newly can cause overhead, we reuse the
643 : : * already-obtained time to avoid that overhead.
644 : : *
645 : : * Note that waitStart is updated without holding the lock table's
646 : : * partition lock, to avoid the overhead by additional lock acquisition.
647 : : * This can cause "waitstart" in pg_locks to become NULL for a very short
648 : : * period of time after the wait started even though "granted" is false.
649 : : * This is OK in practice because we can assume that users are likely to
650 : : * look at "waitstart" when waiting for the lock for a long time.
651 : : */
652 [ + + ]: 3 : if (pg_atomic_read_u64(&MyProc->waitStart) == 0)
653 : 1 : pg_atomic_write_u64(&MyProc->waitStart, now);
654 : :
655 [ + + + - ]: 3 : if (now >= ltime && ltime != 0)
3657 simon@2ndQuadrant.co 656 : 1 : {
657 : : /*
658 : : * We're already behind, so clear a path as quickly as possible.
659 : : */
660 : : VirtualTransactionId *backends;
661 : :
2539 alvherre@alvh.no-ip. 662 : 1 : backends = GetLockConflicts(&locktag, AccessExclusiveLock, NULL);
663 : :
664 : : /*
665 : : * Prevent ResolveRecoveryConflictWithVirtualXIDs() from reporting
666 : : * "waiting" in PS display by disabling its argument report_waiting
667 : : * because the caller, WaitOnLock(), has already reported that.
668 : : */
5904 simon@2ndQuadrant.co 669 : 1 : ResolveRecoveryConflictWithVirtualXIDs(backends,
670 : : RECOVERY_CONFLICT_LOCK,
2172 fujii@postgresql.org 671 : 1 : PG_WAIT_LOCK | locktag.locktag_type,
672 : : false);
673 : : }
674 : : else
675 : : {
676 : : /*
677 : : * Wait (or wait again) until ltime, and check for deadlocks as well
678 : : * if we will be waiting longer than deadlock_timeout
679 : : */
680 : : EnableTimeoutParams timeouts[2];
1894 681 : 2 : int cnt = 0;
682 : :
683 [ + - ]: 2 : if (ltime != 0)
684 : : {
685 : 2 : got_standby_lock_timeout = false;
686 : 2 : timeouts[cnt].id = STANDBY_LOCK_TIMEOUT;
687 : 2 : timeouts[cnt].type = TMPARAM_AT;
688 : 2 : timeouts[cnt].fin_time = ltime;
689 : 2 : cnt++;
690 : : }
691 : :
692 : 2 : got_standby_deadlock_timeout = false;
693 : 2 : timeouts[cnt].id = STANDBY_DEADLOCK_TIMEOUT;
694 : 2 : timeouts[cnt].type = TMPARAM_AFTER;
695 : 2 : timeouts[cnt].delay_ms = DeadlockTimeout;
696 : 2 : cnt++;
697 : :
698 : 2 : enable_timeouts(timeouts, cnt);
699 : : }
700 : :
701 : : /* Wait to be signaled by the release of the Relation Lock */
3448 rhaas@postgresql.org 702 : 3 : ProcWaitForSignal(PG_WAIT_LOCK | locktag.locktag_type);
703 : :
704 : : /*
705 : : * Exit if ltime is reached. Then all the backends holding conflicting
706 : : * locks will be canceled in the next ResolveRecoveryConflictWithLock()
707 : : * call.
708 : : */
1894 fujii@postgresql.org 709 [ - + ]: 3 : if (got_standby_lock_timeout)
1894 fujii@postgresql.org 710 :UBC 0 : goto cleanup;
711 : :
1894 fujii@postgresql.org 712 [ + + ]:CBC 3 : if (got_standby_deadlock_timeout)
713 : : {
714 : : VirtualTransactionId *backends;
715 : :
716 : 2 : backends = GetLockConflicts(&locktag, AccessExclusiveLock, NULL);
717 : :
718 : : /* Quick exit if there's no work to be done */
719 [ - + ]: 2 : if (!VirtualTransactionIdIsValid(*backends))
1894 fujii@postgresql.org 720 :UBC 0 : goto cleanup;
721 : :
722 : : /*
723 : : * Send signals to all the backends holding the conflicting locks, to
724 : : * ask them to check themselves for deadlocks.
725 : : */
1894 fujii@postgresql.org 726 [ + + ]:CBC 4 : while (VirtualTransactionIdIsValid(*backends))
727 : : {
33 heikki.linnakangas@i 728 :GNC 2 : (void) SignalRecoveryConflictWithVirtualXID(*backends,
729 : : RECOVERY_CONFLICT_STARTUP_DEADLOCK);
1894 fujii@postgresql.org 730 :CBC 2 : backends++;
731 : : }
732 : :
733 : : /*
734 : : * Exit if the recovery conflict has not been logged yet even though
735 : : * logging is enabled, so that the caller can log that. Then
736 : : * RecoveryConflictWithLock() is called again and we will wait again
737 : : * for the lock to be released.
738 : : */
1892 739 [ + + ]: 2 : if (logging_conflict)
740 : 1 : goto cleanup;
741 : :
742 : : /*
743 : : * Wait again here to be signaled by the release of the Relation Lock,
744 : : * to prevent the subsequent RecoveryConflictWithLock() from causing
745 : : * deadlock_timeout and sending a request for deadlocks check again.
746 : : * Otherwise the request continues to be sent every deadlock_timeout
747 : : * until the relation locks are released or ltime is reached.
748 : : */
1894 749 : 1 : got_standby_deadlock_timeout = false;
750 : 1 : ProcWaitForSignal(PG_WAIT_LOCK | locktag.locktag_type);
751 : : }
752 : :
753 : 1 : cleanup:
754 : :
755 : : /*
756 : : * Clear any timeout requests established above. We assume here that the
757 : : * Startup process doesn't have any other outstanding timeouts than those
758 : : * used by this function. If that stops being true, we could cancel the
759 : : * timeouts individually, but that'd be slower.
760 : : */
3657 simon@2ndQuadrant.co 761 : 3 : disable_all_timeouts(false);
1894 fujii@postgresql.org 762 : 3 : got_standby_lock_timeout = false;
763 : 3 : got_standby_deadlock_timeout = false;
5904 simon@2ndQuadrant.co 764 : 3 : }
765 : :
766 : : /*
767 : : * ResolveRecoveryConflictWithBufferPin is called from LockBufferForCleanup()
768 : : * to resolve conflicts with other backends holding buffer pins.
769 : : *
770 : : * The ProcWaitForSignal() sleep normally done in LockBufferForCleanup()
771 : : * (when not InHotStandby) is performed here, for code clarity.
772 : : *
773 : : * We either resolve conflicts immediately or set a timeout to wake us at
774 : : * the limit of our patience.
775 : : *
776 : : * Resolve conflicts by sending a PROCSIG signal to all backends to check if
777 : : * they hold one of the buffer pins that is blocking Startup process. If so,
778 : : * those backends will take an appropriate error action, ERROR or FATAL.
779 : : *
780 : : * We also must check for deadlocks. Deadlocks occur because if queries
781 : : * wait on a lock, that must be behind an AccessExclusiveLock, which can only
782 : : * be cleared if the Startup process replays a transaction completion record.
783 : : * If Startup process is also waiting then that is a deadlock. The deadlock
784 : : * can occur if the query is waiting and then the Startup sleeps, or if
785 : : * Startup is sleeping and the query waits on a lock. We protect against
786 : : * only the former sequence here, the latter sequence is checked prior to
787 : : * the query sleeping, in CheckRecoveryConflictDeadlock().
788 : : *
789 : : * Deadlocks are extremely rare, and relatively expensive to check for,
790 : : * so we don't do a deadlock check right away ... only if we have had to wait
791 : : * at least deadlock_timeout.
792 : : */
793 : : void
5895 794 : 10 : ResolveRecoveryConflictWithBufferPin(void)
795 : : {
796 : : TimestampTz ltime;
797 : :
798 [ - + ]: 10 : Assert(InHotStandby);
799 : :
5734 tgl@sss.pgh.pa.us 800 : 10 : ltime = GetStandbyLimitTime();
801 : :
1894 fujii@postgresql.org 802 [ + + + - ]: 10 : if (GetCurrentTimestamp() >= ltime && ltime != 0)
803 : : {
804 : : /*
805 : : * We're already behind, so clear a path as quickly as possible.
806 : : */
33 heikki.linnakangas@i 807 :GNC 1 : SendRecoveryConflictWithBufferPin(RECOVERY_CONFLICT_BUFFERPIN);
808 : : }
809 : : else
810 : : {
811 : : /*
812 : : * Wake up at ltime, and check for deadlocks as well if we will be
813 : : * waiting longer than deadlock_timeout
814 : : */
815 : : EnableTimeoutParams timeouts[2];
1894 fujii@postgresql.org 816 :CBC 9 : int cnt = 0;
817 : :
818 [ + - ]: 9 : if (ltime != 0)
819 : : {
820 : 9 : timeouts[cnt].id = STANDBY_TIMEOUT;
821 : 9 : timeouts[cnt].type = TMPARAM_AT;
822 : 9 : timeouts[cnt].fin_time = ltime;
823 : 9 : cnt++;
824 : : }
825 : :
826 : 9 : got_standby_deadlock_timeout = false;
827 : 9 : timeouts[cnt].id = STANDBY_DEADLOCK_TIMEOUT;
828 : 9 : timeouts[cnt].type = TMPARAM_AFTER;
829 : 9 : timeouts[cnt].delay_ms = DeadlockTimeout;
830 : 9 : cnt++;
831 : :
832 : 9 : enable_timeouts(timeouts, cnt);
833 : : }
834 : :
835 : : /*
836 : : * Wait to be signaled by UnpinBuffer() or for the wait to be interrupted
837 : : * by one of the timeouts established above.
838 : : *
839 : : * We assume that only UnpinBuffer() and the timeout requests established
840 : : * above can wake us up here. WakeupRecovery() called by walreceiver or
841 : : * SIGHUP signal handler, etc cannot do that because it uses the different
842 : : * latch from that ProcWaitForSignal() waits on.
843 : : */
102 andres@anarazel.de 844 :GNC 10 : ProcWaitForSignal(WAIT_EVENT_BUFFER_CLEANUP);
845 : :
1413 andres@anarazel.de 846 [ + + ]:CBC 10 : if (got_standby_delay_timeout)
33 heikki.linnakangas@i 847 :GNC 1 : SendRecoveryConflictWithBufferPin(RECOVERY_CONFLICT_BUFFERPIN);
1413 andres@anarazel.de 848 [ + + ]:CBC 9 : else if (got_standby_deadlock_timeout)
849 : : {
850 : : /*
851 : : * Send out a request for hot-standby backends to check themselves for
852 : : * deadlocks.
853 : : *
854 : : * XXX The subsequent ResolveRecoveryConflictWithBufferPin() will wait
855 : : * to be signaled by UnpinBuffer() again and send a request for
856 : : * deadlocks check if deadlock_timeout happens. This causes the
857 : : * request to continue to be sent every deadlock_timeout until the
858 : : * buffer is unpinned or ltime is reached. This would increase the
859 : : * workload in the startup process and backends. In practice it may
860 : : * not be so harmful because the period that the buffer is kept pinned
861 : : * is basically no so long. But we should fix this?
862 : : */
33 heikki.linnakangas@i 863 :GNC 6 : SendRecoveryConflictWithBufferPin(RECOVERY_CONFLICT_BUFFERPIN_DEADLOCK);
864 : : }
865 : :
866 : : /*
867 : : * Clear any timeout requests established above. We assume here that the
868 : : * Startup process doesn't have any other timeouts than what this function
869 : : * uses. If that stops being true, we could cancel the timeouts
870 : : * individually, but that'd be slower.
871 : : */
4990 alvherre@alvh.no-ip. 872 :CBC 10 : disable_all_timeouts(false);
1413 andres@anarazel.de 873 : 10 : got_standby_delay_timeout = false;
1894 fujii@postgresql.org 874 : 10 : got_standby_deadlock_timeout = false;
5895 simon@2ndQuadrant.co 875 : 10 : }
876 : :
877 : : static void
33 heikki.linnakangas@i 878 :GNC 8 : SendRecoveryConflictWithBufferPin(RecoveryConflictReason reason)
879 : : {
880 [ + + - + ]: 8 : Assert(reason == RECOVERY_CONFLICT_BUFFERPIN ||
881 : : reason == RECOVERY_CONFLICT_BUFFERPIN_DEADLOCK);
882 : :
883 : : /*
884 : : * We send signal to all backends to ask them if they are holding the
885 : : * buffer pin which is delaying the Startup process. Most of them will be
886 : : * innocent, but we let the SIGUSR1 handling in each backend decide their
887 : : * own fate.
888 : : */
889 : 8 : SignalRecoveryConflictWithDatabase(InvalidOid, reason);
5895 simon@2ndQuadrant.co 890 :CBC 8 : }
891 : :
892 : : /*
893 : : * In Hot Standby perform early deadlock detection. We abort the lock
894 : : * wait if we are about to sleep while holding the buffer pin that Startup
895 : : * process is waiting for.
896 : : *
897 : : * Note: this code is pessimistic, because there is no way for it to
898 : : * determine whether an actual deadlock condition is present: the lock we
899 : : * need to wait for might be unrelated to any held by the Startup process.
900 : : * Sooner or later, this mechanism should get ripped out in favor of somehow
901 : : * accounting for buffer locks in DeadLockCheck(). However, errors here
902 : : * seem to be very low-probability in practice, so for now it's not worth
903 : : * the trouble.
904 : : */
905 : : void
5339 tgl@sss.pgh.pa.us 906 : 1 : CheckRecoveryConflictDeadlock(void)
907 : : {
908 [ - + ]: 1 : Assert(!InRecovery); /* do not call in Startup process */
909 : :
5887 simon@2ndQuadrant.co 910 [ + - ]: 1 : if (!HoldingBufferPinThatDelaysRecovery())
911 : 1 : return;
912 : :
913 : : /*
914 : : * Error message should match ProcessInterrupts() but we avoid calling
915 : : * that because we aren't handling an interrupt at this point. Note that
916 : : * we only cancel the current transaction here, so if we are in a
917 : : * subtransaction and the pin is held by a parent, then the Startup
918 : : * process will continue to wait even though we have avoided deadlock.
919 : : */
5887 simon@2ndQuadrant.co 920 [ # # ]:UBC 0 : ereport(ERROR,
921 : : (errcode(ERRCODE_T_R_DEADLOCK_DETECTED),
922 : : errmsg("canceling statement due to conflict with recovery"),
923 : : errdetail("User transaction caused buffer deadlock with recovery.")));
924 : : }
925 : :
926 : :
927 : : /* --------------------------------
928 : : * timeout handler routines
929 : : * --------------------------------
930 : : */
931 : :
932 : : /*
933 : : * StandbyDeadLockHandler() will be called if STANDBY_DEADLOCK_TIMEOUT is
934 : : * exceeded.
935 : : */
936 : : void
4990 alvherre@alvh.no-ip. 937 :CBC 8 : StandbyDeadLockHandler(void)
938 : : {
1894 fujii@postgresql.org 939 : 8 : got_standby_deadlock_timeout = true;
4990 alvherre@alvh.no-ip. 940 : 8 : }
941 : :
942 : : /*
943 : : * StandbyTimeoutHandler() will be called if STANDBY_TIMEOUT is exceeded.
944 : : */
945 : : void
946 : 1 : StandbyTimeoutHandler(void)
947 : : {
1413 andres@anarazel.de 948 : 1 : got_standby_delay_timeout = true;
4990 alvherre@alvh.no-ip. 949 : 1 : }
950 : :
951 : : /*
952 : : * StandbyLockTimeoutHandler() will be called if STANDBY_LOCK_TIMEOUT is exceeded.
953 : : */
954 : : void
3657 simon@2ndQuadrant.co 955 : 1 : StandbyLockTimeoutHandler(void)
956 : : {
1894 fujii@postgresql.org 957 : 1 : got_standby_lock_timeout = true;
3657 simon@2ndQuadrant.co 958 : 1 : }
959 : :
960 : : /*
961 : : * -----------------------------------------------------
962 : : * Locking in Recovery Mode
963 : : * -----------------------------------------------------
964 : : *
965 : : * All locks are held by the Startup process using a single virtual
966 : : * transaction. This implementation is both simpler and in some senses,
967 : : * more correct. The locks held mean "some original transaction held
968 : : * this lock, so query access is not allowed at this time". So the Startup
969 : : * process is the proxy by which the original locks are implemented.
970 : : *
971 : : * We only keep track of AccessExclusiveLocks, which are only ever held by
972 : : * one transaction on one relation.
973 : : *
974 : : * We keep a table of known locks in the RecoveryLockHash hash table.
975 : : * The point of that table is to let us efficiently de-duplicate locks,
976 : : * which is important because checkpoints will re-report the same locks
977 : : * already held. There is also a RecoveryLockXidHash table with one entry
978 : : * per xid, which allows us to efficiently find all the locks held by a
979 : : * given original transaction.
980 : : *
981 : : * We use session locks rather than normal locks so we don't need
982 : : * ResourceOwners.
983 : : */
984 : :
985 : :
986 : : void
5930 987 : 27777 : StandbyAcquireAccessExclusiveLock(TransactionId xid, Oid dbOid, Oid relOid)
988 : : {
989 : : RecoveryLockXidEntry *xidentry;
990 : : RecoveryLockEntry *lockentry;
991 : : xl_standby_lock key;
992 : : LOCKTAG locktag;
993 : : bool found;
994 : :
995 : : /* Already processed? */
5165 996 [ + - + + ]: 55554 : if (!TransactionIdIsValid(xid) ||
997 [ - + ]: 55527 : TransactionIdDidCommit(xid) ||
998 : 27750 : TransactionIdDidAbort(xid))
5930 999 : 27 : return;
1000 : :
825 michael@paquier.xyz 1001 [ - + ]: 27750 : elog(DEBUG4, "adding recovery lock: db %u rel %u", dbOid, relOid);
1002 : :
1003 : : /* dbOid is InvalidOid when we are locking a shared relation. */
5930 simon@2ndQuadrant.co 1004 [ - + ]: 27750 : Assert(OidIsValid(relOid));
1005 : :
1006 : : /* Create a hash entry for this xid, if we don't have one already. */
1256 tgl@sss.pgh.pa.us 1007 : 27750 : xidentry = hash_search(RecoveryLockXidHash, &xid, HASH_ENTER, &found);
2819 tmunro@postgresql.or 1008 [ + + ]: 27750 : if (!found)
1009 : : {
1256 tgl@sss.pgh.pa.us 1010 [ - + ]: 11180 : Assert(xidentry->xid == xid); /* dynahash should have set this */
1011 : 11180 : xidentry->head = NULL;
1012 : : }
1013 : :
1014 : : /* Create a hash entry for this lock, unless we have one already. */
1015 : 27750 : key.xid = xid;
1016 : 27750 : key.dbOid = dbOid;
1017 : 27750 : key.relOid = relOid;
1018 : 27750 : lockentry = hash_search(RecoveryLockHash, &key, HASH_ENTER, &found);
1019 [ + + ]: 27750 : if (!found)
1020 : : {
1021 : : /* It's new, so link it into the XID's list ... */
1022 : 26219 : lockentry->next = xidentry->head;
1023 : 26219 : xidentry->head = lockentry;
1024 : :
1025 : : /* ... and acquire the lock locally. */
1026 : 26219 : SET_LOCKTAG_RELATION(locktag, dbOid, relOid);
1027 : :
1028 : 26219 : (void) LockAcquire(&locktag, AccessExclusiveLock, true, false);
1029 : : }
1030 : : }
1031 : :
1032 : : /*
1033 : : * Release all the locks associated with this RecoveryLockXidEntry.
1034 : : */
1035 : : static void
1036 : 11180 : StandbyReleaseXidEntryLocks(RecoveryLockXidEntry *xidentry)
1037 : : {
1038 : : RecoveryLockEntry *entry;
1039 : : RecoveryLockEntry *next;
1040 : :
1041 [ + + ]: 37399 : for (entry = xidentry->head; entry != NULL; entry = next)
1042 : : {
1043 : : LOCKTAG locktag;
1044 : :
825 michael@paquier.xyz 1045 [ - + ]: 26219 : elog(DEBUG4,
1046 : : "releasing recovery lock: xid %u db %u rel %u",
1047 : : entry->key.xid, entry->key.dbOid, entry->key.relOid);
1048 : : /* Release the lock ... */
1256 tgl@sss.pgh.pa.us 1049 : 26219 : SET_LOCKTAG_RELATION(locktag, entry->key.dbOid, entry->key.relOid);
2819 tmunro@postgresql.or 1050 [ - + ]: 26219 : if (!LockRelease(&locktag, AccessExclusiveLock, true))
1051 : : {
2819 tmunro@postgresql.or 1052 [ # # ]:UBC 0 : elog(LOG,
1053 : : "RecoveryLockHash contains entry for lock no longer recorded by lock manager: xid %u database %u relation %u",
1054 : : entry->key.xid, entry->key.dbOid, entry->key.relOid);
1055 : 0 : Assert(false);
1056 : : }
1057 : : /* ... and remove the per-lock hash entry */
1256 tgl@sss.pgh.pa.us 1058 :CBC 26219 : next = entry->next;
1059 : 26219 : hash_search(RecoveryLockHash, entry, HASH_REMOVE, NULL);
1060 : : }
1061 : :
1062 : 11180 : xidentry->head = NULL; /* just for paranoia */
2819 tmunro@postgresql.or 1063 : 11180 : }
1064 : :
1065 : : /*
1066 : : * Release locks for specific XID, or all locks if it's InvalidXid.
1067 : : */
1068 : : static void
1069 : 11860 : StandbyReleaseLocks(TransactionId xid)
1070 : : {
1071 : : RecoveryLockXidEntry *entry;
1072 : :
1073 [ + - ]: 11860 : if (TransactionIdIsValid(xid))
1074 : : {
1256 tgl@sss.pgh.pa.us 1075 [ + + ]: 11860 : if ((entry = hash_search(RecoveryLockXidHash, &xid, HASH_FIND, NULL)))
1076 : : {
1077 : 11180 : StandbyReleaseXidEntryLocks(entry);
1078 : 11180 : hash_search(RecoveryLockXidHash, entry, HASH_REMOVE, NULL);
1079 : : }
1080 : : }
1081 : : else
2819 tmunro@postgresql.or 1082 :UBC 0 : StandbyReleaseAllLocks();
5930 simon@2ndQuadrant.co 1083 :CBC 11860 : }
1084 : :
1085 : : /*
1086 : : * Release locks for a transaction tree, starting at xid down, from
1087 : : * RecoveryLockXidHash.
1088 : : *
1089 : : * Called during WAL replay of COMMIT/ROLLBACK when in hot standby mode,
1090 : : * to remove any AccessExclusiveLocks requested by a transaction.
1091 : : */
1092 : : void
1093 : 11360 : StandbyReleaseLockTree(TransactionId xid, int nsubxids, TransactionId *subxids)
1094 : : {
1095 : : int i;
1096 : :
1097 : 11360 : StandbyReleaseLocks(xid);
1098 : :
1099 [ + + ]: 11860 : for (i = 0; i < nsubxids; i++)
1100 : 500 : StandbyReleaseLocks(subxids[i]);
1101 : 11360 : }
1102 : :
1103 : : /*
1104 : : * Called at end of recovery and when we see a shutdown checkpoint.
1105 : : */
1106 : : void
5165 1107 : 114 : StandbyReleaseAllLocks(void)
1108 : : {
1109 : : HASH_SEQ_STATUS status;
1110 : : RecoveryLockXidEntry *entry;
1111 : :
825 michael@paquier.xyz 1112 [ + + ]: 114 : elog(DEBUG2, "release all standby locks");
1113 : :
1256 tgl@sss.pgh.pa.us 1114 : 114 : hash_seq_init(&status, RecoveryLockXidHash);
2819 tmunro@postgresql.or 1115 [ - + ]: 114 : while ((entry = hash_seq_search(&status)))
1116 : : {
1256 tgl@sss.pgh.pa.us 1117 :UBC 0 : StandbyReleaseXidEntryLocks(entry);
1118 : 0 : hash_search(RecoveryLockXidHash, entry, HASH_REMOVE, NULL);
1119 : : }
5165 simon@2ndQuadrant.co 1120 :CBC 114 : }
1121 : :
1122 : : /*
1123 : : * StandbyReleaseOldLocks
1124 : : * Release standby locks held by top-level XIDs that aren't running,
1125 : : * as long as they're not prepared transactions.
1126 : : *
1127 : : * This is needed to prune the locks of crashed transactions, which didn't
1128 : : * write an ABORT/COMMIT record.
1129 : : */
1130 : : void
2829 1131 : 824 : StandbyReleaseOldLocks(TransactionId oldxid)
1132 : : {
1133 : : HASH_SEQ_STATUS status;
1134 : : RecoveryLockXidEntry *entry;
1135 : :
1256 tgl@sss.pgh.pa.us 1136 : 824 : hash_seq_init(&status, RecoveryLockXidHash);
2819 tmunro@postgresql.or 1137 [ + + ]: 1182 : while ((entry = hash_seq_search(&status)))
1138 : : {
1139 [ - + ]: 358 : Assert(TransactionIdIsValid(entry->xid));
1140 : :
1141 : : /* Skip if prepared transaction. */
1142 [ - + ]: 358 : if (StandbyTransactionIdIsPrepared(entry->xid))
2819 tmunro@postgresql.or 1143 :UBC 0 : continue;
1144 : :
1145 : : /* Skip if >= oldxid. */
2819 tmunro@postgresql.or 1146 [ + - ]:CBC 358 : if (!TransactionIdPrecedes(entry->xid, oldxid))
1147 : 358 : continue;
1148 : :
1149 : : /* Remove all locks and hash table entry. */
1256 tgl@sss.pgh.pa.us 1150 :UBC 0 : StandbyReleaseXidEntryLocks(entry);
1151 : 0 : hash_search(RecoveryLockXidHash, entry, HASH_REMOVE, NULL);
1152 : : }
5930 simon@2ndQuadrant.co 1153 :CBC 824 : }
1154 : :
1155 : : /*
1156 : : * --------------------------------------------------------------------
1157 : : * Recovery handling for Rmgr RM_STANDBY_ID
1158 : : *
1159 : : * These record types will only be created if XLogStandbyInfoActive()
1160 : : * --------------------------------------------------------------------
1161 : : */
1162 : :
1163 : : void
4133 heikki.linnakangas@i 1164 : 27943 : standby_redo(XLogReaderState *record)
1165 : : {
1166 : 27943 : uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
1167 : :
1168 : : /* Backup blocks are not used in standby records */
1169 [ - + ]: 27943 : Assert(!XLogRecHasAnyBlockRefs(record));
1170 : :
1171 : : /* Do nothing if we're not in hot standby mode */
5930 simon@2ndQuadrant.co 1172 [ + + ]: 27943 : if (standbyState == STANDBY_DISABLED)
1173 : 154 : return;
1174 : :
1175 [ + + ]: 27789 : if (info == XLOG_STANDBY_LOCK)
1176 : : {
1177 : 26396 : xl_standby_locks *xlrec = (xl_standby_locks *) XLogRecGetData(record);
1178 : : int i;
1179 : :
1180 [ + + ]: 54173 : for (i = 0; i < xlrec->nlocks; i++)
1181 : 27777 : StandbyAcquireAccessExclusiveLock(xlrec->locks[i].xid,
1182 : : xlrec->locks[i].dbOid,
1183 : : xlrec->locks[i].relOid);
1184 : : }
1185 [ + + ]: 1393 : else if (info == XLOG_RUNNING_XACTS)
1186 : : {
1187 : 759 : xl_running_xacts *xlrec = (xl_running_xacts *) XLogRecGetData(record);
1188 : : RunningTransactionsData running;
1189 : :
1190 : 759 : running.xcnt = xlrec->xcnt;
4851 1191 : 759 : running.subxcnt = xlrec->subxcnt;
626 heikki.linnakangas@i 1192 : 759 : running.subxid_status = xlrec->subxid_overflow ? SUBXIDS_MISSING : SUBXIDS_IN_ARRAY;
5930 simon@2ndQuadrant.co 1193 : 759 : running.nextXid = xlrec->nextXid;
5784 1194 : 759 : running.latestCompletedXid = xlrec->latestCompletedXid;
5930 1195 : 759 : running.oldestRunningXid = xlrec->oldestRunningXid;
1196 : 759 : running.xids = xlrec->xids;
1197 : :
1198 : 759 : ProcArrayApplyRecoveryInfo(&running);
1199 : :
1200 : : /*
1201 : : * The startup process currently has no convenient way to schedule
1202 : : * stats to be reported. XLOG_RUNNING_XACTS records issued at a
1203 : : * regular cadence, making this a convenient location to report stats.
1204 : : * While these records aren't generated with wal_level=minimal, stats
1205 : : * also cannot be accessed during WAL replay.
1206 : : */
1007 andres@anarazel.de 1207 : 759 : pgstat_report_stat(true);
1208 : : }
3613 1209 [ + - ]: 634 : else if (info == XLOG_INVALIDATIONS)
1210 : : {
1211 : 634 : xl_invalidations *xlrec = (xl_invalidations *) XLogRecGetData(record);
1212 : :
1213 : 634 : ProcessCommittedInvalidationMessages(xlrec->msgs,
1214 : : xlrec->nmsgs,
1215 : 634 : xlrec->relcacheInitFileInval,
1216 : : xlrec->dbId,
1217 : : xlrec->tsId);
1218 : : }
1219 : : else
5153 tgl@sss.pgh.pa.us 1220 [ # # ]:UBC 0 : elog(PANIC, "standby_redo: unknown op code %u", info);
1221 : : }
1222 : :
1223 : : /*
1224 : : * Log details of the current snapshot to WAL. This allows the snapshot state
1225 : : * to be reconstructed on the standby and for logical decoding.
1226 : : *
1227 : : * This is used for Hot Standby as follows:
1228 : : *
1229 : : * We can move directly to STANDBY_SNAPSHOT_READY at startup if we
1230 : : * start from a shutdown checkpoint because we know nothing was running
1231 : : * at that time and our recovery snapshot is known empty. In the more
1232 : : * typical case of an online checkpoint we need to jump through a few
1233 : : * hoops to get a correct recovery snapshot and this requires a two or
1234 : : * sometimes a three stage process.
1235 : : *
1236 : : * The initial snapshot must contain all running xids and all current
1237 : : * AccessExclusiveLocks at a point in time on the standby. Assembling
1238 : : * that information while the server is running requires many and
1239 : : * various LWLocks, so we choose to derive that information piece by
1240 : : * piece and then re-assemble that info on the standby. When that
1241 : : * information is fully assembled we move to STANDBY_SNAPSHOT_READY.
1242 : : *
1243 : : * Since locking on the primary when we derive the information is not
1244 : : * strict, we note that there is a time window between the derivation and
1245 : : * writing to WAL of the derived information. That allows race conditions
1246 : : * that we must resolve, since xids and locks may enter or leave the
1247 : : * snapshot during that window. This creates the issue that an xid or
1248 : : * lock may start *after* the snapshot has been derived yet *before* the
1249 : : * snapshot is logged in the running xacts WAL record. We resolve this by
1250 : : * starting to accumulate changes at a point just prior to when we derive
1251 : : * the snapshot on the primary, then ignore duplicates when we later apply
1252 : : * the snapshot from the running xacts record. This is implemented during
1253 : : * CreateCheckPoint() where we use the logical checkpoint location as
1254 : : * our starting point and then write the running xacts record immediately
1255 : : * before writing the main checkpoint WAL record. Since we always start
1256 : : * up from a checkpoint and are immediately at our starting point, we
1257 : : * unconditionally move to STANDBY_INITIALIZED. After this point we
1258 : : * must do 4 things:
1259 : : * * move shared nextXid forwards as we see new xids
1260 : : * * extend the clog and subtrans with each new xid
1261 : : * * keep track of uncommitted known assigned xids
1262 : : * * keep track of uncommitted AccessExclusiveLocks
1263 : : *
1264 : : * When we see a commit/abort we must remove known assigned xids and locks
1265 : : * from the completing transaction. Attempted removals that cannot locate
1266 : : * an entry are expected and must not cause an error when we are in state
1267 : : * STANDBY_INITIALIZED. This is implemented in StandbyReleaseLocks() and
1268 : : * KnownAssignedXidsRemove().
1269 : : *
1270 : : * Later, when we apply the running xact data we must be careful to ignore
1271 : : * transactions already committed, since those commits raced ahead when
1272 : : * making WAL entries.
1273 : : *
1274 : : * For logical decoding only the running xacts information is needed;
1275 : : * there's no need to look at the locking information, but it's logged anyway,
1276 : : * as there's no independent knob to just enable logical decoding. For
1277 : : * details of how this is used, check snapbuild.c's introductory comment.
1278 : : *
1279 : : *
1280 : : * Returns the RecPtr of the last inserted record.
1281 : : */
1282 : : XLogRecPtr
4851 tgl@sss.pgh.pa.us 1283 :CBC 1472 : LogStandbySnapshot(void)
1284 : : {
1285 : : XLogRecPtr recptr;
1286 : : RunningTransactions running;
1287 : : xl_standby_lock *locks;
1288 : : int nlocks;
82 msawada@postgresql.o 1289 :GNC 1472 : bool logical_decoding_enabled = IsLogicalDecodingEnabled();
1290 : :
5930 simon@2ndQuadrant.co 1291 [ - + ]:CBC 1472 : Assert(XLogStandbyInfoActive());
1292 : :
1293 : : #ifdef USE_INJECTION_POINTS
341 akapila@postgresql.o 1294 [ - + ]: 1472 : if (IS_INJECTION_POINT_ATTACHED("skip-log-running-xacts"))
1295 : : {
1296 : : /*
1297 : : * This record could move slot's xmin forward during decoding, leading
1298 : : * to unpredictable results, so skip it when requested by the test.
1299 : : */
341 akapila@postgresql.o 1300 :UBC 0 : return GetInsertRecPtr();
1301 : : }
1302 : : #endif
1303 : :
1304 : : /*
1305 : : * Get details of any AccessExclusiveLocks being held at the moment.
1306 : : */
5930 simon@2ndQuadrant.co 1307 :CBC 1472 : locks = GetRunningTransactionLocks(&nlocks);
1308 [ + + ]: 1472 : if (nlocks > 0)
1309 : 185 : LogAccessExclusiveLocks(nlocks, locks);
4667 tgl@sss.pgh.pa.us 1310 : 1472 : pfree(locks);
1311 : :
1312 : : /*
1313 : : * Log details of all in-progress transactions. This should be the last
1314 : : * record we write, because standby will open up when it sees this.
1315 : : */
5930 simon@2ndQuadrant.co 1316 : 1472 : running = GetRunningTransactionData();
1317 : :
1318 : : /*
1319 : : * GetRunningTransactionData() acquired ProcArrayLock, we must release it.
1320 : : * For Hot Standby this can be done before inserting the WAL record
1321 : : * because ProcArrayApplyRecoveryInfo() rechecks the commit status using
1322 : : * the clog. For logical decoding, though, the lock can't be released
1323 : : * early because the clog might be "in the future" from the POV of the
1324 : : * historic snapshot. This would allow for situations where we're waiting
1325 : : * for the end of a transaction listed in the xl_running_xacts record
1326 : : * which, according to the WAL, has committed before the xl_running_xacts
1327 : : * record. Fortunately this routine isn't executed frequently, and it's
1328 : : * only a shared lock.
1329 : : */
82 msawada@postgresql.o 1330 [ + + ]:GNC 1472 : if (!logical_decoding_enabled)
4395 rhaas@postgresql.org 1331 :CBC 926 : LWLockRelease(ProcArrayLock);
1332 : :
4442 1333 : 1472 : recptr = LogCurrentRunningXacts(running);
1334 : :
1335 : : /* Release lock if we kept it longer ... */
82 msawada@postgresql.o 1336 [ + + ]:GNC 1472 : if (logical_decoding_enabled)
4395 rhaas@postgresql.org 1337 :CBC 546 : LWLockRelease(ProcArrayLock);
1338 : :
1339 : : /* GetRunningTransactionData() acquired XidGenLock, we must release it */
5577 heikki.linnakangas@i 1340 : 1472 : LWLockRelease(XidGenLock);
1341 : :
4442 rhaas@postgresql.org 1342 : 1472 : return recptr;
1343 : : }
1344 : :
1345 : : /*
1346 : : * Record an enhanced snapshot of running transactions into WAL.
1347 : : *
1348 : : * The definitions of RunningTransactionsData and xl_running_xacts are
1349 : : * similar. We keep them separate because xl_running_xacts is a contiguous
1350 : : * chunk of memory and never exists fully until it is assembled in WAL.
1351 : : * The inserted records are marked as not being important for durability,
1352 : : * to avoid triggering superfluous checkpoint / archiving activity.
1353 : : */
1354 : : static XLogRecPtr
5930 simon@2ndQuadrant.co 1355 : 1472 : LogCurrentRunningXacts(RunningTransactions CurrRunningXacts)
1356 : : {
1357 : : xl_running_xacts xlrec;
1358 : : XLogRecPtr recptr;
1359 : :
1360 : 1472 : xlrec.xcnt = CurrRunningXacts->xcnt;
4851 1361 : 1472 : xlrec.subxcnt = CurrRunningXacts->subxcnt;
626 heikki.linnakangas@i 1362 : 1472 : xlrec.subxid_overflow = (CurrRunningXacts->subxid_status != SUBXIDS_IN_ARRAY);
5930 simon@2ndQuadrant.co 1363 : 1472 : xlrec.nextXid = CurrRunningXacts->nextXid;
1364 : 1472 : xlrec.oldestRunningXid = CurrRunningXacts->oldestRunningXid;
5784 1365 : 1472 : xlrec.latestCompletedXid = CurrRunningXacts->latestCompletedXid;
1366 : :
1367 : : /* Header */
4133 heikki.linnakangas@i 1368 : 1472 : XLogBeginInsert();
3370 andres@anarazel.de 1369 : 1472 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
397 peter@eisentraut.org 1370 : 1472 : XLogRegisterData(&xlrec, MinSizeOfXactRunningXacts);
1371 : :
1372 : : /* array of TransactionIds */
5930 simon@2ndQuadrant.co 1373 [ + + ]: 1472 : if (xlrec.xcnt > 0)
397 peter@eisentraut.org 1374 : 493 : XLogRegisterData(CurrRunningXacts->xids,
3189 tgl@sss.pgh.pa.us 1375 : 493 : (xlrec.xcnt + xlrec.subxcnt) * sizeof(TransactionId));
1376 : :
4133 heikki.linnakangas@i 1377 : 1472 : recptr = XLogInsert(RM_STANDBY_ID, XLOG_RUNNING_XACTS);
1378 : :
626 1379 [ + + ]: 1472 : if (xlrec.subxid_overflow)
825 michael@paquier.xyz 1380 [ - + ]: 2 : elog(DEBUG2,
1381 : : "snapshot of %d running transactions overflowed (lsn %X/%08X oldest xid %u latest complete %u next xid %u)",
1382 : : CurrRunningXacts->xcnt,
1383 : : LSN_FORMAT_ARGS(recptr),
1384 : : CurrRunningXacts->oldestRunningXid,
1385 : : CurrRunningXacts->latestCompletedXid,
1386 : : CurrRunningXacts->nextXid);
1387 : : else
1388 [ + + ]: 1470 : elog(DEBUG2,
1389 : : "snapshot of %d+%d running transaction ids (lsn %X/%08X oldest xid %u latest complete %u next xid %u)",
1390 : : CurrRunningXacts->xcnt, CurrRunningXacts->subxcnt,
1391 : : LSN_FORMAT_ARGS(recptr),
1392 : : CurrRunningXacts->oldestRunningXid,
1393 : : CurrRunningXacts->latestCompletedXid,
1394 : : CurrRunningXacts->nextXid);
1395 : :
1396 : : /*
1397 : : * Ensure running_xacts information is synced to disk not too far in the
1398 : : * future. We don't want to stall anything though (i.e. use XLogFlush()),
1399 : : * so we let the wal writer do it during normal operation.
1400 : : * XLogSetAsyncXactLSN() conveniently will mark the LSN as to-be-synced
1401 : : * and nudge the WALWriter into action if sleeping. Check
1402 : : * XLogBackgroundFlush() for details why a record might not be flushed
1403 : : * without it.
1404 : : */
4442 rhaas@postgresql.org 1405 : 1472 : XLogSetAsyncXactLSN(recptr);
1406 : :
1407 : 1472 : return recptr;
1408 : : }
1409 : :
1410 : : /*
1411 : : * Wholesale logging of AccessExclusiveLocks. Other lock types need not be
1412 : : * logged, as described in backend/storage/lmgr/README.
1413 : : */
1414 : : static void
5930 simon@2ndQuadrant.co 1415 : 133824 : LogAccessExclusiveLocks(int nlocks, xl_standby_lock *locks)
1416 : : {
1417 : : xl_standby_locks xlrec;
1418 : :
1419 : 133824 : xlrec.nlocks = nlocks;
1420 : :
4133 heikki.linnakangas@i 1421 : 133824 : XLogBeginInsert();
397 peter@eisentraut.org 1422 : 133824 : XLogRegisterData(&xlrec, offsetof(xl_standby_locks, locks));
1423 : 133824 : XLogRegisterData(locks, nlocks * sizeof(xl_standby_lock));
3370 andres@anarazel.de 1424 : 133824 : XLogSetRecordFlags(XLOG_MARK_UNIMPORTANT);
1425 : :
4133 heikki.linnakangas@i 1426 : 133824 : (void) XLogInsert(RM_STANDBY_ID, XLOG_STANDBY_LOCK);
5930 simon@2ndQuadrant.co 1427 : 133824 : }
1428 : :
1429 : : /*
1430 : : * Individual logging of AccessExclusiveLocks for use during LockAcquire()
1431 : : */
1432 : : void
1433 : 133639 : LogAccessExclusiveLock(Oid dbOid, Oid relOid)
1434 : : {
1435 : : xl_standby_lock xlrec;
1436 : :
3280 1437 : 133639 : xlrec.xid = GetCurrentTransactionId();
1438 : :
5930 1439 : 133639 : xlrec.dbOid = dbOid;
1440 : 133639 : xlrec.relOid = relOid;
1441 : :
1442 : 133639 : LogAccessExclusiveLocks(1, &xlrec);
3280 1443 : 133639 : MyXactFlags |= XACT_FLAGS_ACQUIREDACCESSEXCLUSIVELOCK;
5930 1444 : 133639 : }
1445 : :
1446 : : /*
1447 : : * Prepare to log an AccessExclusiveLock, for use during LockAcquire()
1448 : : */
1449 : : void
5585 1450 : 133854 : LogAccessExclusiveLockPrepare(void)
1451 : : {
1452 : : /*
1453 : : * Ensure that a TransactionId has been assigned to this transaction, for
1454 : : * two reasons, both related to lock release on the standby. First, we
1455 : : * must assign an xid so that RecordTransactionCommit() and
1456 : : * RecordTransactionAbort() do not optimise away the transaction
1457 : : * completion record which recovery relies upon to release locks. It's a
1458 : : * hack, but for a corner case not worth adding code for into the main
1459 : : * commit path. Second, we must assign an xid before the lock is recorded
1460 : : * in shared memory, otherwise a concurrently executing
1461 : : * GetRunningTransactionLocks() might see a lock associated with an
1462 : : * InvalidTransactionId which we later assert cannot happen.
1463 : : */
3280 1464 : 133854 : (void) GetCurrentTransactionId();
5585 1465 : 133854 : }
1466 : :
1467 : : /*
1468 : : * Emit WAL for invalidations. This currently is only used for commits without
1469 : : * an xid but which contain invalidations.
1470 : : */
1471 : : void
3613 andres@anarazel.de 1472 : 9701 : LogStandbyInvalidations(int nmsgs, SharedInvalidationMessage *msgs,
1473 : : bool relcacheInitFileInval)
1474 : : {
1475 : : xl_invalidations xlrec;
1476 : :
1477 : : /* prepare record */
1478 : 9701 : memset(&xlrec, 0, sizeof(xlrec));
1479 : 9701 : xlrec.dbId = MyDatabaseId;
1480 : 9701 : xlrec.tsId = MyDatabaseTableSpace;
1481 : 9701 : xlrec.relcacheInitFileInval = relcacheInitFileInval;
1482 : 9701 : xlrec.nmsgs = nmsgs;
1483 : :
1484 : : /* perform insertion */
1485 : 9701 : XLogBeginInsert();
397 peter@eisentraut.org 1486 : 9701 : XLogRegisterData(&xlrec, MinSizeOfInvalidations);
1487 : 9701 : XLogRegisterData(msgs,
1488 : : nmsgs * sizeof(SharedInvalidationMessage));
3613 andres@anarazel.de 1489 : 9701 : XLogInsert(RM_STANDBY_ID, XLOG_INVALIDATIONS);
1490 : 9701 : }
1491 : :
1492 : : /* Return the description of recovery conflict */
1493 : : static const char *
33 heikki.linnakangas@i 1494 :GNC 10 : get_recovery_conflict_desc(RecoveryConflictReason reason)
1495 : : {
1721 peter@eisentraut.org 1496 :CBC 10 : const char *reasonDesc = _("unknown reason");
1497 : :
1892 fujii@postgresql.org 1498 [ + + + + : 10 : switch (reason)
- - - -
- ]
1499 : : {
33 heikki.linnakangas@i 1500 :GNC 4 : case RECOVERY_CONFLICT_BUFFERPIN:
1721 peter@eisentraut.org 1501 :CBC 4 : reasonDesc = _("recovery conflict on buffer pin");
1892 fujii@postgresql.org 1502 : 4 : break;
33 heikki.linnakangas@i 1503 :GNC 2 : case RECOVERY_CONFLICT_LOCK:
1721 peter@eisentraut.org 1504 :CBC 2 : reasonDesc = _("recovery conflict on lock");
1892 fujii@postgresql.org 1505 : 2 : break;
33 heikki.linnakangas@i 1506 :GNC 2 : case RECOVERY_CONFLICT_TABLESPACE:
1721 peter@eisentraut.org 1507 :CBC 2 : reasonDesc = _("recovery conflict on tablespace");
1892 fujii@postgresql.org 1508 : 2 : break;
33 heikki.linnakangas@i 1509 :GNC 2 : case RECOVERY_CONFLICT_SNAPSHOT:
1721 peter@eisentraut.org 1510 :CBC 2 : reasonDesc = _("recovery conflict on snapshot");
1892 fujii@postgresql.org 1511 : 2 : break;
33 heikki.linnakangas@i 1512 :UNC 0 : case RECOVERY_CONFLICT_LOGICALSLOT:
1073 andres@anarazel.de 1513 :UBC 0 : reasonDesc = _("recovery conflict on replication slot");
1514 : 0 : break;
33 heikki.linnakangas@i 1515 :UNC 0 : case RECOVERY_CONFLICT_STARTUP_DEADLOCK:
1516 : 0 : reasonDesc = _("recovery conflict on deadlock");
1517 : 0 : break;
1518 : 0 : case RECOVERY_CONFLICT_BUFFERPIN_DEADLOCK:
1721 peter@eisentraut.org 1519 :UBC 0 : reasonDesc = _("recovery conflict on buffer deadlock");
1892 fujii@postgresql.org 1520 : 0 : break;
33 heikki.linnakangas@i 1521 :UNC 0 : case RECOVERY_CONFLICT_DATABASE:
1721 peter@eisentraut.org 1522 :UBC 0 : reasonDesc = _("recovery conflict on database");
1892 fujii@postgresql.org 1523 : 0 : break;
1524 : : }
1525 : :
1892 fujii@postgresql.org 1526 :CBC 10 : return reasonDesc;
1527 : : }
|
