Age Owner Branch data TLA Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * syncrep.c
4 : : *
5 : : * Synchronous replication is new as of PostgreSQL 9.1.
6 : : *
7 : : * If requested, transaction commits wait until their commit LSN are
8 : : * acknowledged by the synchronous standbys.
9 : : *
10 : : * This module contains the code for waiting and release of backends.
11 : : * All code in this module executes on the primary. The core streaming
12 : : * replication transport remains within WALreceiver/WALsender modules.
13 : : *
14 : : * The essence of this design is that it isolates all logic about
15 : : * waiting/releasing onto the primary. The primary defines which standbys
16 : : * it wishes to wait for. The standbys are completely unaware of the
17 : : * durability requirements of transactions on the primary, reducing the
18 : : * complexity of the code and streamlining both standby operations and
19 : : * network bandwidth because there is no requirement to ship
20 : : * per-transaction state information.
21 : : *
22 : : * Replication is either synchronous or not synchronous (async). If it is
23 : : * async, we just fastpath out of here. If it is sync, then we wait for
24 : : * the write, flush or apply location on the standby before releasing
25 : : * the waiting backend. Further complexity in that interaction is
26 : : * expected in later releases.
27 : : *
28 : : * The best performing way to manage the waiting backends is to have a
29 : : * single ordered queue of waiting backends, so that we can avoid
30 : : * searching the through all waiters each time we receive a reply.
31 : : *
32 : : * In 9.5 or before only a single standby could be considered as
33 : : * synchronous. In 9.6 we support a priority-based multiple synchronous
34 : : * standbys. In 10.0 a quorum-based multiple synchronous standbys is also
35 : : * supported. The number of synchronous standbys that transactions
36 : : * must wait for replies from is specified in synchronous_standby_names.
37 : : * This parameter also specifies a list of standby names and the method
38 : : * (FIRST and ANY) to choose synchronous standbys from the listed ones.
39 : : *
40 : : * The method FIRST specifies a priority-based synchronous replication
41 : : * and makes transaction commits wait until their WAL records are
42 : : * replicated to the requested number of synchronous standbys chosen based
43 : : * on their priorities. The standbys whose names appear earlier in the list
44 : : * are given higher priority and will be considered as synchronous.
45 : : * Other standby servers appearing later in this list represent potential
46 : : * synchronous standbys. If any of the current synchronous standbys
47 : : * disconnects for whatever reason, it will be replaced immediately with
48 : : * the next-highest-priority standby.
49 : : *
50 : : * The method ANY specifies a quorum-based synchronous replication
51 : : * and makes transaction commits wait until their WAL records are
52 : : * replicated to at least the requested number of synchronous standbys
53 : : * in the list. All the standbys appearing in the list are considered as
54 : : * candidates for quorum synchronous standbys.
55 : : *
56 : : * If neither FIRST nor ANY is specified, FIRST is used as the method.
57 : : * This is for backward compatibility with 9.6 or before where only a
58 : : * priority-based sync replication was supported.
59 : : *
60 : : * Before the standbys chosen from synchronous_standby_names can
61 : : * become the synchronous standbys they must have caught up with
62 : : * the primary; that may take some time. Once caught up,
63 : : * the standbys which are considered as synchronous at that moment
64 : : * will release waiters from the queue.
65 : : *
66 : : * Portions Copyright (c) 2010-2026, PostgreSQL Global Development Group
67 : : *
68 : : * IDENTIFICATION
69 : : * src/backend/replication/syncrep.c
70 : : *
71 : : *-------------------------------------------------------------------------
72 : : */
73 : : #include "postgres.h"
74 : :
75 : : #include <unistd.h>
76 : :
77 : : #include "access/xact.h"
78 : : #include "common/int.h"
79 : : #include "miscadmin.h"
80 : : #include "pgstat.h"
81 : : #include "replication/syncrep.h"
82 : : #include "replication/walsender.h"
83 : : #include "replication/walsender_private.h"
84 : : #include "storage/proc.h"
85 : : #include "tcop/tcopprot.h"
86 : : #include "utils/guc_hooks.h"
87 : : #include "utils/ps_status.h"
88 : : #include "utils/wait_event.h"
89 : :
90 : : /* User-settable parameters for sync rep */
91 : : char *SyncRepStandbyNames;
92 : :
93 : : #define SyncStandbysDefined() \
94 : : (SyncRepStandbyNames != NULL && SyncRepStandbyNames[0] != '\0')
95 : :
96 : : static bool announce_next_takeover = true;
97 : :
98 : : SyncRepConfigData *SyncRepConfig = NULL;
99 : : static int SyncRepWaitMode = SYNC_REP_NO_WAIT;
100 : :
101 : : static void SyncRepQueueInsert(int mode);
102 : : static void SyncRepCancelWait(void);
103 : : static int SyncRepWakeQueue(bool all, int mode);
104 : :
105 : : static bool SyncRepGetSyncRecPtr(XLogRecPtr *writePtr,
106 : : XLogRecPtr *flushPtr,
107 : : XLogRecPtr *applyPtr,
108 : : bool *am_sync);
109 : : static void SyncRepGetOldestSyncRecPtr(XLogRecPtr *writePtr,
110 : : XLogRecPtr *flushPtr,
111 : : XLogRecPtr *applyPtr,
112 : : SyncRepStandbyData *sync_standbys,
113 : : int num_standbys);
114 : : static void SyncRepGetNthLatestSyncRecPtr(XLogRecPtr *writePtr,
115 : : XLogRecPtr *flushPtr,
116 : : XLogRecPtr *applyPtr,
117 : : SyncRepStandbyData *sync_standbys,
118 : : int num_standbys,
119 : : uint8 nth);
120 : : static int SyncRepGetStandbyPriority(void);
121 : : static int standby_priority_comparator(const void *a, const void *b);
122 : : static int cmp_lsn(const void *a, const void *b);
123 : :
124 : : #ifdef USE_ASSERT_CHECKING
125 : : static bool SyncRepQueueIsOrderedByLSN(int mode);
126 : : #endif
127 : :
128 : : /*
129 : : * ===========================================================
130 : : * Synchronous Replication functions for normal user backends
131 : : * ===========================================================
132 : : */
133 : :
134 : : /*
135 : : * Wait for synchronous replication, if requested by user.
136 : : *
137 : : * Initially backends start in state SYNC_REP_NOT_WAITING and then
138 : : * change that state to SYNC_REP_WAITING before adding ourselves
139 : : * to the wait queue. During SyncRepWakeQueue() a WALSender changes
140 : : * the state to SYNC_REP_WAIT_COMPLETE once replication is confirmed.
141 : : * This backend then resets its state to SYNC_REP_NOT_WAITING.
142 : : *
143 : : * 'lsn' represents the LSN to wait for. 'commit' indicates whether this LSN
144 : : * represents a commit record. If it doesn't, then we wait only for the WAL
145 : : * to be flushed if synchronous_commit is set to the higher level of
146 : : * remote_apply, because only commit records provide apply feedback.
147 : : */
148 : : void
3638 rhaas@postgresql.org 149 :CBC 122880 : SyncRepWaitForLSN(XLogRecPtr lsn, bool commit)
150 : : {
151 : : int mode;
152 : :
153 : : /*
154 : : * This should be called while holding interrupts during a transaction
155 : : * commit to prevent the follow-up shared memory queue cleanups to be
156 : : * influenced by external interruptions.
157 : : */
2326 michael@paquier.xyz 158 [ - + ]: 122880 : Assert(InterruptHoldoffCount > 0);
159 : :
160 : : /*
161 : : * Fast exit if user has not requested sync replication, or there are no
162 : : * sync replication standby names defined.
163 : : *
164 : : * Since this routine gets called every commit time, it's important to
165 : : * exit quickly if sync replication is not requested.
166 : : *
167 : : * We check WalSndCtl->sync_standbys_status flag without the lock and exit
168 : : * immediately if SYNC_STANDBY_INIT is set (the checkpointer has
169 : : * initialized this data) but SYNC_STANDBY_DEFINED is missing (no sync
170 : : * replication requested).
171 : : *
172 : : * If SYNC_STANDBY_DEFINED is set, we need to check the status again later
173 : : * while holding the lock, to check the flag and operate the sync rep
174 : : * queue atomically. This is necessary to avoid the race condition
175 : : * described in SyncRepUpdateSyncStandbysDefined(). On the other hand, if
176 : : * SYNC_STANDBY_DEFINED is not set, the lock is not necessary because we
177 : : * don't touch the queue.
178 : : */
2020 fujii@postgresql.org 179 [ + + + + ]: 122880 : if (!SyncRepRequested() ||
338 michael@paquier.xyz 180 [ + + ]: 91881 : ((((volatile WalSndCtlData *) WalSndCtl)->sync_standbys_status) &
181 : : (SYNC_STANDBY_INIT | SYNC_STANDBY_DEFINED)) == SYNC_STANDBY_INIT)
2020 fujii@postgresql.org 182 : 90546 : return;
183 : :
184 : : /* Cap the level for anything other than commit to remote flush only. */
3638 rhaas@postgresql.org 185 [ + + ]: 32334 : if (commit)
186 : 32316 : mode = SyncRepWaitMode;
187 : : else
188 : 18 : mode = Min(SyncRepWaitMode, SYNC_REP_WAIT_FLUSH);
189 : :
1152 andres@anarazel.de 190 [ - + ]: 32334 : Assert(dlist_node_is_detached(&MyProc->syncRepLinks));
5477 rhaas@postgresql.org 191 [ - + ]: 32334 : Assert(WalSndCtl != NULL);
192 : :
193 : 32334 : LWLockAcquire(SyncRepLock, LW_EXCLUSIVE);
194 [ - + ]: 32334 : Assert(MyProc->syncRepState == SYNC_REP_NOT_WAITING);
195 : :
196 : : /*
197 : : * We don't wait for sync rep if SYNC_STANDBY_DEFINED is not set. See
198 : : * SyncRepUpdateSyncStandbysDefined().
199 : : *
200 : : * Also check that the standby hasn't already replied. Unlikely race
201 : : * condition but we'll be fetching that cache line anyway so it's likely
202 : : * to be a low cost check.
203 : : *
204 : : * If the sync standby data has not been initialized yet
205 : : * (SYNC_STANDBY_INIT is not set), fall back to a check based on the LSN,
206 : : * then do a direct GUC check.
207 : : */
338 michael@paquier.xyz 208 [ + + ]: 32334 : if (WalSndCtl->sync_standbys_status & SYNC_STANDBY_INIT)
209 : : {
210 [ + - ]: 40 : if ((WalSndCtl->sync_standbys_status & SYNC_STANDBY_DEFINED) == 0 ||
211 [ + + ]: 40 : lsn <= WalSndCtl->lsn[mode])
212 : : {
213 : 5 : LWLockRelease(SyncRepLock);
214 : 5 : return;
215 : : }
216 : : }
217 [ - + ]: 32294 : else if (lsn <= WalSndCtl->lsn[mode])
218 : : {
219 : : /*
220 : : * The LSN is older than what we need to wait for. The sync standby
221 : : * data has not been initialized yet, but we are OK to not wait
222 : : * because we know that there is no point in doing so based on the
223 : : * LSN.
224 : : */
338 michael@paquier.xyz 225 :UBC 0 : LWLockRelease(SyncRepLock);
226 : 0 : return;
227 : : }
338 michael@paquier.xyz 228 [ + - + - ]:CBC 32294 : else if (!SyncStandbysDefined())
229 : : {
230 : : /*
231 : : * If we are here, the sync standby data has not been initialized yet,
232 : : * and the LSN is newer than what need to wait for, so we have fallen
233 : : * back to the best thing we could do in this case: a check on
234 : : * SyncStandbysDefined() to see if the GUC is set or not.
235 : : *
236 : : * When the GUC has a value, we wait until the checkpointer updates
237 : : * the status data because we cannot be sure yet if we should wait or
238 : : * not. Here, the GUC has *no* value, we are sure that there is no
239 : : * point to wait; this matters for example when initializing a
240 : : * cluster, where we should never wait, and no sync standbys is the
241 : : * default behavior.
242 : : */
5477 rhaas@postgresql.org 243 : 32294 : LWLockRelease(SyncRepLock);
244 : 32294 : return;
245 : : }
246 : :
247 : : /*
248 : : * Set our waitLSN so WALSender will know when to wake us, and add
249 : : * ourselves to the queue.
250 : : */
3638 251 : 35 : MyProc->waitLSN = lsn;
5477 252 : 35 : MyProc->syncRepState = SYNC_REP_WAITING;
5158 simon@2ndQuadrant.co 253 : 35 : SyncRepQueueInsert(mode);
254 [ - + ]: 35 : Assert(SyncRepQueueIsOrderedByLSN(mode));
5477 rhaas@postgresql.org 255 : 35 : LWLockRelease(SyncRepLock);
256 : :
257 : : /* Alter ps display to show waiting for sync rep. */
258 [ + - ]: 35 : if (update_process_title)
259 : : {
260 : : char buffer[32];
261 : :
251 alvherre@kurilemu.de 262 :GNC 35 : sprintf(buffer, "waiting for %X/%08X", LSN_FORMAT_ARGS(lsn));
1119 drowley@postgresql.o 263 :CBC 35 : set_ps_display_suffix(buffer);
264 : : }
265 : :
266 : : /*
267 : : * Wait for specified LSN to be confirmed.
268 : : *
269 : : * Each proc has its own wait latch, so we perform a normal latch
270 : : * check/wait loop here.
271 : : */
272 : : for (;;)
5488 simon@2ndQuadrant.co 273 : 35 : {
274 : : int rc;
275 : :
276 : : /* Must reset the latch before testing state. */
4078 andres@anarazel.de 277 : 70 : ResetLatch(MyLatch);
278 : :
279 : : /*
280 : : * Acquiring the lock is not needed, the latch ensures proper
281 : : * barriers. If it looks like we're done, we must really be done,
282 : : * because once walsender changes the state to SYNC_REP_WAIT_COMPLETE,
283 : : * it will never update it again, so we can't be seeing a stale value
284 : : * in that case.
285 : : */
3502 simon@2ndQuadrant.co 286 [ + + ]: 70 : if (MyProc->syncRepState == SYNC_REP_WAIT_COMPLETE)
5477 rhaas@postgresql.org 287 : 35 : break;
288 : :
289 : : /*
290 : : * If a wait for synchronous replication is pending, we can neither
291 : : * acknowledge the commit nor raise ERROR or FATAL. The latter would
292 : : * lead the client to believe that the transaction aborted, which is
293 : : * not true: it's already committed locally. The former is no good
294 : : * either: the client has requested synchronous replication, and is
295 : : * entitled to assume that an acknowledged commit is also replicated,
296 : : * which might not be true. So in this case we issue a WARNING (which
297 : : * some clients may be able to interpret) and shut off further output.
298 : : * We do NOT reset ProcDiePending, so that the process will die after
299 : : * the commit is cleaned up.
300 : : */
301 [ - + ]: 35 : if (ProcDiePending)
302 : : {
5477 rhaas@postgresql.org 303 [ # # ]:UBC 0 : ereport(WARNING,
304 : : (errcode(ERRCODE_ADMIN_SHUTDOWN),
305 : : errmsg("canceling the wait for synchronous replication and terminating connection due to administrator command"),
306 : : errdetail("The transaction has already committed locally, but might not have been replicated to the standby.")));
307 : 0 : whereToSendOutput = DestNone;
308 : 0 : SyncRepCancelWait();
309 : 0 : break;
310 : : }
311 : :
312 : : /*
313 : : * It's unclear what to do if a query cancel interrupt arrives. We
314 : : * can't actually abort at this point, but ignoring the interrupt
315 : : * altogether is not helpful, so we just terminate the wait with a
316 : : * suitable warning.
317 : : */
5477 rhaas@postgresql.org 318 [ - + ]:CBC 35 : if (QueryCancelPending)
319 : : {
5477 rhaas@postgresql.org 320 :UBC 0 : QueryCancelPending = false;
321 [ # # ]: 0 : ereport(WARNING,
322 : : (errmsg("canceling wait for synchronous replication due to user request"),
323 : : errdetail("The transaction has already committed locally, but might not have been replicated to the standby.")));
324 : 0 : SyncRepCancelWait();
325 : 0 : break;
326 : : }
327 : :
328 : : /*
329 : : * Wait on latch. Any condition that should wake us up will set the
330 : : * latch, so no need for timeout.
331 : : */
2669 tmunro@postgresql.or 332 :CBC 35 : rc = WaitLatch(MyLatch, WL_LATCH_SET | WL_POSTMASTER_DEATH, -1,
333 : : WAIT_EVENT_SYNC_REP);
334 : :
335 : : /*
336 : : * If the postmaster dies, we'll probably never get an acknowledgment,
337 : : * because all the wal sender processes will exit. So just bail out.
338 : : */
339 [ - + ]: 35 : if (rc & WL_POSTMASTER_DEATH)
340 : : {
5477 rhaas@postgresql.org 341 :UBC 0 : ProcDiePending = true;
342 : 0 : whereToSendOutput = DestNone;
343 : 0 : SyncRepCancelWait();
344 : 0 : break;
345 : : }
346 : : }
347 : :
348 : : /*
349 : : * WalSender has checked our LSN and has removed us from queue. Clean up
350 : : * state and leave. It's OK to reset these shared memory fields without
351 : : * holding SyncRepLock, because any walsenders will ignore us anyway when
352 : : * we're not on the queue. We need a read barrier to make sure we see the
353 : : * changes to the queue link (this might be unnecessary without
354 : : * assertions, but better safe than sorry).
355 : : */
3168 heikki.linnakangas@i 356 :CBC 35 : pg_read_barrier();
1152 andres@anarazel.de 357 [ - + ]: 35 : Assert(dlist_node_is_detached(&MyProc->syncRepLinks));
5477 rhaas@postgresql.org 358 : 35 : MyProc->syncRepState = SYNC_REP_NOT_WAITING;
45 alvherre@kurilemu.de 359 :GNC 35 : MyProc->waitLSN = InvalidXLogRecPtr;
360 : :
361 : : /* reset ps display to remove the suffix */
1119 drowley@postgresql.o 362 [ + - ]:CBC 35 : if (update_process_title)
363 : 35 : set_ps_display_remove_suffix();
364 : : }
365 : :
366 : : /*
367 : : * Insert MyProc into the specified SyncRepQueue, maintaining sorted invariant.
368 : : *
369 : : * Usually we will go at tail of queue, though it's possible that we arrive
370 : : * here out of order, so start at tail and work back to insertion point.
371 : : */
372 : : static void
5164 simon@2ndQuadrant.co 373 : 35 : SyncRepQueueInsert(int mode)
374 : : {
375 : : dlist_head *queue;
376 : : dlist_iter iter;
377 : :
378 [ + - - + ]: 35 : Assert(mode >= 0 && mode < NUM_SYNC_REP_WAIT_MODE);
1152 andres@anarazel.de 379 : 35 : queue = &WalSndCtl->SyncRepQueue[mode];
380 : :
381 [ + - - + ]: 35 : dlist_reverse_foreach(iter, queue)
382 : : {
1152 andres@anarazel.de 383 :UBC 0 : PGPROC *proc = dlist_container(PGPROC, syncRepLinks, iter.cur);
384 : :
385 : : /*
386 : : * Stop at the queue element that we should insert after to ensure the
387 : : * queue is ordered by LSN.
388 : : */
4825 alvherre@alvh.no-ip. 389 [ # # ]: 0 : if (proc->waitLSN < MyProc->waitLSN)
390 : : {
1152 andres@anarazel.de 391 : 0 : dlist_insert_after(&proc->syncRepLinks, &MyProc->syncRepLinks);
392 : 0 : return;
393 : : }
394 : : }
395 : :
396 : : /*
397 : : * If we get here, the list was either empty, or this process needs to be
398 : : * at the head.
399 : : */
1152 andres@anarazel.de 400 :CBC 35 : dlist_push_head(queue, &MyProc->syncRepLinks);
401 : : }
402 : :
403 : : /*
404 : : * Acquire SyncRepLock and cancel any wait currently in progress.
405 : : */
406 : : static void
5477 rhaas@postgresql.org 407 :UBC 0 : SyncRepCancelWait(void)
408 : : {
409 : 0 : LWLockAcquire(SyncRepLock, LW_EXCLUSIVE);
1152 andres@anarazel.de 410 [ # # ]: 0 : if (!dlist_node_is_detached(&MyProc->syncRepLinks))
411 : 0 : dlist_delete_thoroughly(&MyProc->syncRepLinks);
5477 rhaas@postgresql.org 412 : 0 : MyProc->syncRepState = SYNC_REP_NOT_WAITING;
413 : 0 : LWLockRelease(SyncRepLock);
414 : 0 : }
415 : :
416 : : void
5331 tgl@sss.pgh.pa.us 417 :CBC 16962 : SyncRepCleanupAtProcExit(void)
418 : : {
419 : : /*
420 : : * First check if we are removed from the queue without the lock to not
421 : : * slow down backend exit.
422 : : */
1152 andres@anarazel.de 423 [ - + ]: 16962 : if (!dlist_node_is_detached(&MyProc->syncRepLinks))
424 : : {
5488 simon@2ndQuadrant.co 425 :UBC 0 : LWLockAcquire(SyncRepLock, LW_EXCLUSIVE);
426 : :
427 : : /* maybe we have just been removed, so recheck */
1152 andres@anarazel.de 428 [ # # ]: 0 : if (!dlist_node_is_detached(&MyProc->syncRepLinks))
429 : 0 : dlist_delete_thoroughly(&MyProc->syncRepLinks);
430 : :
5488 simon@2ndQuadrant.co 431 : 0 : LWLockRelease(SyncRepLock);
432 : : }
5488 simon@2ndQuadrant.co 433 :CBC 16962 : }
434 : :
435 : : /*
436 : : * ===========================================================
437 : : * Synchronous Replication functions for wal sender processes
438 : : * ===========================================================
439 : : */
440 : :
441 : : /*
442 : : * Take any action required to initialise sync rep state from config
443 : : * data. Called at WALSender startup and after each SIGHUP.
444 : : */
445 : : void
446 : 757 : SyncRepInitConfig(void)
447 : : {
448 : : int priority;
449 : :
450 : : /*
451 : : * Determine if we are a potential sync standby and remember the result
452 : : * for handling replies from standby.
453 : : */
454 : 757 : priority = SyncRepGetStandbyPriority();
455 [ + + ]: 757 : if (MyWalSnd->sync_standby_priority != priority)
456 : : {
2157 tgl@sss.pgh.pa.us 457 [ - + ]: 18 : SpinLockAcquire(&MyWalSnd->mutex);
5488 simon@2ndQuadrant.co 458 : 18 : MyWalSnd->sync_standby_priority = priority;
2157 tgl@sss.pgh.pa.us 459 : 18 : SpinLockRelease(&MyWalSnd->mutex);
460 : :
5488 simon@2ndQuadrant.co 461 [ - + ]: 18 : ereport(DEBUG1,
462 : : (errmsg_internal("standby \"%s\" now has synchronous standby priority %d",
463 : : application_name, priority)));
464 : : }
465 : 757 : }
466 : :
467 : : /*
468 : : * Update the LSNs on each queue based upon our latest state. This
469 : : * implements a simple policy of first-valid-sync-standby-releases-waiter.
470 : : *
471 : : * Other policies are possible, which would change what we do here and
472 : : * perhaps also which information we store as well.
473 : : */
474 : : void
475 : 100000 : SyncRepReleaseWaiters(void)
476 : : {
477 : 100000 : volatile WalSndCtlData *walsndctl = WalSndCtl;
478 : : XLogRecPtr writePtr;
479 : : XLogRecPtr flushPtr;
480 : : XLogRecPtr applyPtr;
481 : : bool got_recptr;
482 : : bool am_sync;
5164 483 : 100000 : int numwrite = 0;
484 : 100000 : int numflush = 0;
3638 rhaas@postgresql.org 485 : 100000 : int numapply = 0;
486 : :
487 : : /*
488 : : * If this WALSender is serving a standby that is not on the list of
489 : : * potential sync standbys then we have nothing to do. If we are still
490 : : * starting up, still running base backup or the current flush position is
491 : : * still invalid, then leave quickly also. Streaming or stopping WAL
492 : : * senders are allowed to release waiters.
493 : : */
5488 simon@2ndQuadrant.co 494 [ + + ]: 100000 : if (MyWalSnd->sync_standby_priority == 0 ||
2663 michael@paquier.xyz 495 [ + + ]: 279 : (MyWalSnd->state != WALSNDSTATE_STREAMING &&
496 [ + + ]: 157 : MyWalSnd->state != WALSNDSTATE_STOPPING) ||
129 alvherre@kurilemu.de 497 [ - + ]:GNC 265 : !XLogRecPtrIsValid(MyWalSnd->flush))
498 : : {
3630 fujii@postgresql.org 499 :CBC 99735 : announce_next_takeover = true;
5488 simon@2ndQuadrant.co 500 : 99739 : return;
501 : : }
502 : :
503 : : /*
504 : : * We're a potential sync standby. Release waiters if there are enough
505 : : * sync standbys and we are considered as sync.
506 : : */
507 : 265 : LWLockAcquire(SyncRepLock, LW_EXCLUSIVE);
508 : :
509 : : /*
510 : : * Check whether we are a sync standby or not, and calculate the synced
511 : : * positions among all sync standbys. (Note: although this step does not
512 : : * of itself require holding SyncRepLock, it seems like a good idea to do
513 : : * it after acquiring the lock. This ensures that the WAL pointers we use
514 : : * to release waiters are newer than any previous execution of this
515 : : * routine used.)
516 : : */
3373 fujii@postgresql.org 517 : 265 : got_recptr = SyncRepGetSyncRecPtr(&writePtr, &flushPtr, &applyPtr, &am_sync);
518 : :
519 : : /*
520 : : * If we are managing a sync standby, though we weren't prior to this,
521 : : * then announce we are now a sync standby.
522 : : */
3630 523 [ + + + + ]: 265 : if (announce_next_takeover && am_sync)
524 : : {
525 : 16 : announce_next_takeover = false;
526 : :
3373 527 [ + - ]: 16 : if (SyncRepConfig->syncrep_method == SYNC_REP_PRIORITY)
528 [ + - ]: 16 : ereport(LOG,
529 : : (errmsg("standby \"%s\" is now a synchronous standby with priority %d",
530 : : application_name, MyWalSnd->sync_standby_priority)));
531 : : else
3373 fujii@postgresql.org 532 [ # # ]:UBC 0 : ereport(LOG,
533 : : (errmsg("standby \"%s\" is now a candidate for quorum synchronous standby",
534 : : application_name)));
535 : : }
536 : :
537 : : /*
538 : : * If the number of sync standbys is less than requested or we aren't
539 : : * managing a sync standby then just leave.
540 : : */
3373 fujii@postgresql.org 541 [ + + - + ]:CBC 265 : if (!got_recptr || !am_sync)
542 : : {
5488 simon@2ndQuadrant.co 543 : 4 : LWLockRelease(SyncRepLock);
3630 fujii@postgresql.org 544 : 4 : announce_next_takeover = !am_sync;
5488 simon@2ndQuadrant.co 545 : 4 : return;
546 : : }
547 : :
548 : : /*
549 : : * Set the lsn first so that when we wake backends they will release up to
550 : : * this location.
551 : : */
3630 fujii@postgresql.org 552 [ + + ]: 261 : if (walsndctl->lsn[SYNC_REP_WAIT_WRITE] < writePtr)
553 : : {
554 : 52 : walsndctl->lsn[SYNC_REP_WAIT_WRITE] = writePtr;
5164 simon@2ndQuadrant.co 555 : 52 : numwrite = SyncRepWakeQueue(false, SYNC_REP_WAIT_WRITE);
556 : : }
3630 fujii@postgresql.org 557 [ + + ]: 261 : if (walsndctl->lsn[SYNC_REP_WAIT_FLUSH] < flushPtr)
558 : : {
559 : 57 : walsndctl->lsn[SYNC_REP_WAIT_FLUSH] = flushPtr;
5164 simon@2ndQuadrant.co 560 : 57 : numflush = SyncRepWakeQueue(false, SYNC_REP_WAIT_FLUSH);
561 : : }
3630 fujii@postgresql.org 562 [ + + ]: 261 : if (walsndctl->lsn[SYNC_REP_WAIT_APPLY] < applyPtr)
563 : : {
564 : 53 : walsndctl->lsn[SYNC_REP_WAIT_APPLY] = applyPtr;
3638 rhaas@postgresql.org 565 : 53 : numapply = SyncRepWakeQueue(false, SYNC_REP_WAIT_APPLY);
566 : : }
567 : :
5488 simon@2ndQuadrant.co 568 : 261 : LWLockRelease(SyncRepLock);
569 : :
251 alvherre@kurilemu.de 570 [ - + ]:GNC 261 : elog(DEBUG3, "released %d procs up to write %X/%08X, %d procs up to flush %X/%08X, %d procs up to apply %X/%08X",
571 : : numwrite, LSN_FORMAT_ARGS(writePtr),
572 : : numflush, LSN_FORMAT_ARGS(flushPtr),
573 : : numapply, LSN_FORMAT_ARGS(applyPtr));
574 : : }
575 : :
576 : : /*
577 : : * Calculate the synced Write, Flush and Apply positions among sync standbys.
578 : : *
579 : : * Return false if the number of sync standbys is less than
580 : : * synchronous_standby_names specifies. Otherwise return true and
581 : : * store the positions into *writePtr, *flushPtr and *applyPtr.
582 : : *
583 : : * On return, *am_sync is set to true if this walsender is connecting to
584 : : * sync standby. Otherwise it's set to false.
585 : : */
586 : : static bool
3373 fujii@postgresql.org 587 :CBC 265 : SyncRepGetSyncRecPtr(XLogRecPtr *writePtr, XLogRecPtr *flushPtr,
588 : : XLogRecPtr *applyPtr, bool *am_sync)
589 : : {
590 : : SyncRepStandbyData *sync_standbys;
591 : : int num_standbys;
592 : : int i;
593 : :
594 : : /* Initialize default results */
3630 595 : 265 : *writePtr = InvalidXLogRecPtr;
596 : 265 : *flushPtr = InvalidXLogRecPtr;
597 : 265 : *applyPtr = InvalidXLogRecPtr;
598 : 265 : *am_sync = false;
599 : :
600 : : /* Quick out if not even configured to be synchronous */
2157 tgl@sss.pgh.pa.us 601 [ - + ]: 265 : if (SyncRepConfig == NULL)
2157 tgl@sss.pgh.pa.us 602 :UBC 0 : return false;
603 : :
604 : : /* Get standbys that are considered as synchronous at this moment */
2157 tgl@sss.pgh.pa.us 605 :CBC 265 : num_standbys = SyncRepGetCandidateStandbys(&sync_standbys);
606 : :
607 : : /* Am I among the candidate sync standbys? */
608 [ + + ]: 271 : for (i = 0; i < num_standbys; i++)
609 : : {
610 [ + + ]: 268 : if (sync_standbys[i].is_me)
611 : : {
612 : 262 : *am_sync = true;
613 : 262 : break;
614 : : }
615 : : }
616 : :
617 : : /*
618 : : * Nothing more to do if we are not managing a sync standby or there are
619 : : * not enough synchronous standbys.
620 : : */
3609 621 [ + + ]: 265 : if (!(*am_sync) ||
2157 622 [ + + ]: 262 : num_standbys < SyncRepConfig->num_sync)
623 : : {
624 : 4 : pfree(sync_standbys);
3630 fujii@postgresql.org 625 : 4 : return false;
626 : : }
627 : :
628 : : /*
629 : : * In a priority-based sync replication, the synced positions are the
630 : : * oldest ones among sync standbys. In a quorum-based, they are the Nth
631 : : * latest ones.
632 : : *
633 : : * SyncRepGetNthLatestSyncRecPtr() also can calculate the oldest
634 : : * positions. But we use SyncRepGetOldestSyncRecPtr() for that calculation
635 : : * because it's a bit more efficient.
636 : : *
637 : : * XXX If the numbers of current and requested sync standbys are the same,
638 : : * we can use SyncRepGetOldestSyncRecPtr() to calculate the synced
639 : : * positions even in a quorum-based sync replication.
640 : : */
3373 641 [ + - ]: 261 : if (SyncRepConfig->syncrep_method == SYNC_REP_PRIORITY)
642 : : {
643 : 261 : SyncRepGetOldestSyncRecPtr(writePtr, flushPtr, applyPtr,
644 : : sync_standbys, num_standbys);
645 : : }
646 : : else
647 : : {
3373 fujii@postgresql.org 648 :UBC 0 : SyncRepGetNthLatestSyncRecPtr(writePtr, flushPtr, applyPtr,
649 : : sync_standbys, num_standbys,
2157 tgl@sss.pgh.pa.us 650 : 0 : SyncRepConfig->num_sync);
651 : : }
652 : :
2157 tgl@sss.pgh.pa.us 653 :CBC 261 : pfree(sync_standbys);
3373 fujii@postgresql.org 654 : 261 : return true;
655 : : }
656 : :
657 : : /*
658 : : * Calculate the oldest Write, Flush and Apply positions among sync standbys.
659 : : */
660 : : static void
2157 tgl@sss.pgh.pa.us 661 : 261 : SyncRepGetOldestSyncRecPtr(XLogRecPtr *writePtr,
662 : : XLogRecPtr *flushPtr,
663 : : XLogRecPtr *applyPtr,
664 : : SyncRepStandbyData *sync_standbys,
665 : : int num_standbys)
666 : : {
667 : : int i;
668 : :
669 : : /*
670 : : * Scan through all sync standbys and calculate the oldest Write, Flush
671 : : * and Apply positions. We assume *writePtr et al were initialized to
672 : : * InvalidXLogRecPtr.
673 : : */
674 [ + + ]: 524 : for (i = 0; i < num_standbys; i++)
675 : : {
676 : 263 : XLogRecPtr write = sync_standbys[i].write;
677 : 263 : XLogRecPtr flush = sync_standbys[i].flush;
678 : 263 : XLogRecPtr apply = sync_standbys[i].apply;
679 : :
129 alvherre@kurilemu.de 680 [ + + - + ]:GNC 263 : if (!XLogRecPtrIsValid(*writePtr) || *writePtr > write)
3630 fujii@postgresql.org 681 :CBC 261 : *writePtr = write;
129 alvherre@kurilemu.de 682 [ + + - + ]:GNC 263 : if (!XLogRecPtrIsValid(*flushPtr) || *flushPtr > flush)
3630 fujii@postgresql.org 683 :CBC 261 : *flushPtr = flush;
129 alvherre@kurilemu.de 684 [ + + - + ]:GNC 263 : if (!XLogRecPtrIsValid(*applyPtr) || *applyPtr > apply)
3630 fujii@postgresql.org 685 :CBC 261 : *applyPtr = apply;
686 : : }
3373 687 : 261 : }
688 : :
689 : : /*
690 : : * Calculate the Nth latest Write, Flush and Apply positions among sync
691 : : * standbys.
692 : : */
693 : : static void
2157 tgl@sss.pgh.pa.us 694 :UBC 0 : SyncRepGetNthLatestSyncRecPtr(XLogRecPtr *writePtr,
695 : : XLogRecPtr *flushPtr,
696 : : XLogRecPtr *applyPtr,
697 : : SyncRepStandbyData *sync_standbys,
698 : : int num_standbys,
699 : : uint8 nth)
700 : : {
701 : : XLogRecPtr *write_array;
702 : : XLogRecPtr *flush_array;
703 : : XLogRecPtr *apply_array;
704 : : int i;
705 : :
706 : : /* Should have enough candidates, or somebody messed up */
707 [ # # # # ]: 0 : Assert(nth > 0 && nth <= num_standbys);
708 : :
95 michael@paquier.xyz 709 :UNC 0 : write_array = palloc_array(XLogRecPtr, num_standbys);
710 : 0 : flush_array = palloc_array(XLogRecPtr, num_standbys);
711 : 0 : apply_array = palloc_array(XLogRecPtr, num_standbys);
712 : :
2157 tgl@sss.pgh.pa.us 713 [ # # ]:UBC 0 : for (i = 0; i < num_standbys; i++)
714 : : {
715 : 0 : write_array[i] = sync_standbys[i].write;
716 : 0 : flush_array[i] = sync_standbys[i].flush;
717 : 0 : apply_array[i] = sync_standbys[i].apply;
718 : : }
719 : :
720 : : /* Sort each array in descending order */
721 : 0 : qsort(write_array, num_standbys, sizeof(XLogRecPtr), cmp_lsn);
722 : 0 : qsort(flush_array, num_standbys, sizeof(XLogRecPtr), cmp_lsn);
723 : 0 : qsort(apply_array, num_standbys, sizeof(XLogRecPtr), cmp_lsn);
724 : :
725 : : /* Get Nth latest Write, Flush, Apply positions */
3373 fujii@postgresql.org 726 : 0 : *writePtr = write_array[nth - 1];
727 : 0 : *flushPtr = flush_array[nth - 1];
728 : 0 : *applyPtr = apply_array[nth - 1];
729 : :
730 : 0 : pfree(write_array);
731 : 0 : pfree(flush_array);
732 : 0 : pfree(apply_array);
733 : 0 : }
734 : :
735 : : /*
736 : : * Compare lsn in order to sort array in descending order.
737 : : */
738 : : static int
739 : 0 : cmp_lsn(const void *a, const void *b)
740 : : {
3224 bruce@momjian.us 741 : 0 : XLogRecPtr lsn1 = *((const XLogRecPtr *) a);
742 : 0 : XLogRecPtr lsn2 = *((const XLogRecPtr *) b);
743 : :
758 nathan@postgresql.or 744 : 0 : return pg_cmp_u64(lsn2, lsn1);
745 : : }
746 : :
747 : : /*
748 : : * Return data about walsenders that are candidates to be sync standbys.
749 : : *
750 : : * *standbys is set to a palloc'd array of structs of per-walsender data,
751 : : * and the number of valid entries (candidate sync senders) is returned.
752 : : * (This might be more or fewer than num_sync; caller must check.)
753 : : */
754 : : int
2157 tgl@sss.pgh.pa.us 755 :CBC 1002 : SyncRepGetCandidateStandbys(SyncRepStandbyData **standbys)
756 : : {
757 : : int i;
758 : : int n;
759 : :
760 : : /* Create result array */
95 michael@paquier.xyz 761 :GNC 1002 : *standbys = palloc_array(SyncRepStandbyData, max_wal_senders);
762 : :
763 : : /* Quick exit if sync replication is not requested */
3373 fujii@postgresql.org 764 [ + + ]:CBC 1002 : if (SyncRepConfig == NULL)
2157 tgl@sss.pgh.pa.us 765 : 720 : return 0;
766 : :
767 : : /* Collect raw data from shared memory */
768 : 282 : n = 0;
3373 fujii@postgresql.org 769 [ + + ]: 3102 : for (i = 0; i < max_wal_senders; i++)
770 : : {
771 : : volatile WalSnd *walsnd; /* Use volatile pointer to prevent code
772 : : * rearrangement */
773 : : SyncRepStandbyData *stby;
774 : : WalSndState state; /* not included in SyncRepStandbyData */
775 : :
776 : 2820 : walsnd = &WalSndCtl->walsnds[i];
2157 tgl@sss.pgh.pa.us 777 : 2820 : stby = *standbys + n;
778 : :
3180 alvherre@alvh.no-ip. 779 [ - + ]: 2820 : SpinLockAcquire(&walsnd->mutex);
2157 tgl@sss.pgh.pa.us 780 : 2820 : stby->pid = walsnd->pid;
3180 alvherre@alvh.no-ip. 781 : 2820 : state = walsnd->state;
2157 tgl@sss.pgh.pa.us 782 : 2820 : stby->write = walsnd->write;
783 : 2820 : stby->flush = walsnd->flush;
784 : 2820 : stby->apply = walsnd->apply;
785 : 2820 : stby->sync_standby_priority = walsnd->sync_standby_priority;
3180 alvherre@alvh.no-ip. 786 : 2820 : SpinLockRelease(&walsnd->mutex);
787 : :
788 : : /* Must be active */
2157 tgl@sss.pgh.pa.us 789 [ + + ]: 2820 : if (stby->pid == 0)
3373 fujii@postgresql.org 790 : 2490 : continue;
791 : :
792 : : /* Must be streaming or stopping */
2663 michael@paquier.xyz 793 [ + + - + ]: 330 : if (state != WALSNDSTATE_STREAMING &&
794 : : state != WALSNDSTATE_STOPPING)
3373 fujii@postgresql.org 795 :UBC 0 : continue;
796 : :
797 : : /* Must be synchronous */
2157 tgl@sss.pgh.pa.us 798 [ + + ]:CBC 330 : if (stby->sync_standby_priority == 0)
3373 fujii@postgresql.org 799 : 14 : continue;
800 : :
801 : : /* Must have a valid flush position */
129 alvherre@kurilemu.de 802 [ - + ]:GNC 316 : if (!XLogRecPtrIsValid(stby->flush))
3373 fujii@postgresql.org 803 :UBC 0 : continue;
804 : :
805 : : /* OK, it's a candidate */
2157 tgl@sss.pgh.pa.us 806 :CBC 316 : stby->walsnd_index = i;
807 : 316 : stby->is_me = (walsnd == MyWalSnd);
808 : 316 : n++;
809 : : }
810 : :
811 : : /*
812 : : * In quorum mode, we return all the candidates. In priority mode, if we
813 : : * have too many candidates then return only the num_sync ones of highest
814 : : * priority.
815 : : */
816 [ + + ]: 282 : if (SyncRepConfig->syncrep_method == SYNC_REP_PRIORITY &&
817 [ + + ]: 281 : n > SyncRepConfig->num_sync)
818 : : {
819 : : /* Sort by priority ... */
820 : 14 : qsort(*standbys, n, sizeof(SyncRepStandbyData),
821 : : standby_priority_comparator);
822 : : /* ... then report just the first num_sync ones */
823 : 14 : n = SyncRepConfig->num_sync;
824 : : }
825 : :
826 : 282 : return n;
827 : : }
828 : :
829 : : /*
830 : : * qsort comparator to sort SyncRepStandbyData entries by priority
831 : : */
832 : : static int
833 : 31 : standby_priority_comparator(const void *a, const void *b)
834 : : {
835 : 31 : const SyncRepStandbyData *sa = (const SyncRepStandbyData *) a;
836 : 31 : const SyncRepStandbyData *sb = (const SyncRepStandbyData *) b;
837 : :
838 : : /* First, sort by increasing priority value */
839 [ + + ]: 31 : if (sa->sync_standby_priority != sb->sync_standby_priority)
840 : 14 : return sa->sync_standby_priority - sb->sync_standby_priority;
841 : :
842 : : /*
843 : : * We might have equal priority values; arbitrarily break ties by position
844 : : * in the WalSnd array. (This is utterly bogus, since that is arrival
845 : : * order dependent, but there are regression tests that rely on it.)
846 : : */
847 : 17 : return sa->walsnd_index - sb->walsnd_index;
848 : : }
849 : :
850 : :
851 : : /*
852 : : * Check if we are in the list of sync standbys, and if so, determine
853 : : * priority sequence. Return priority if set, or zero to indicate that
854 : : * we are not a potential sync standby.
855 : : *
856 : : * Compare the parameter SyncRepStandbyNames against the application_name
857 : : * for this WALSender, or allow any name if we find a wildcard "*".
858 : : */
859 : : static int
5488 simon@2ndQuadrant.co 860 : 757 : SyncRepGetStandbyPriority(void)
861 : : {
862 : : const char *standby_name;
863 : : int priority;
864 : 757 : bool found = false;
865 : :
866 : : /*
867 : : * Since synchronous cascade replication is not allowed, we always set the
868 : : * priority of cascading walsender to zero.
869 : : */
5353 870 [ + + ]: 757 : if (am_cascading_walsender)
871 : 26 : return 0;
872 : :
3609 tgl@sss.pgh.pa.us 873 [ + - + + : 731 : if (!SyncStandbysDefined() || SyncRepConfig == NULL)
- + ]
5488 simon@2ndQuadrant.co 874 : 706 : return 0;
875 : :
3609 tgl@sss.pgh.pa.us 876 : 25 : standby_name = SyncRepConfig->member_names;
877 [ + + ]: 33 : for (priority = 1; priority <= SyncRepConfig->nmembers; priority++)
878 : : {
5488 simon@2ndQuadrant.co 879 [ + + ]: 32 : if (pg_strcasecmp(standby_name, application_name) == 0 ||
3609 tgl@sss.pgh.pa.us 880 [ + + ]: 18 : strcmp(standby_name, "*") == 0)
881 : : {
5488 simon@2ndQuadrant.co 882 : 24 : found = true;
883 : 24 : break;
884 : : }
3609 tgl@sss.pgh.pa.us 885 : 8 : standby_name += strlen(standby_name) + 1;
886 : : }
887 : :
3245 fujii@postgresql.org 888 [ + + ]: 25 : if (!found)
889 : 1 : return 0;
890 : :
891 : : /*
892 : : * In quorum-based sync replication, all the standbys in the list have the
893 : : * same priority, one.
894 : : */
895 [ + - ]: 24 : return (SyncRepConfig->syncrep_method == SYNC_REP_PRIORITY) ? priority : 1;
896 : : }
897 : :
898 : : /*
899 : : * Walk the specified queue from head. Set the state of any backends that
900 : : * need to be woken, remove them from the queue, and then wake them.
901 : : * Pass all = true to wake whole queue; otherwise, just wake up to
902 : : * the walsender's LSN.
903 : : *
904 : : * The caller must hold SyncRepLock in exclusive mode.
905 : : */
906 : : static int
5164 simon@2ndQuadrant.co 907 : 165 : SyncRepWakeQueue(bool all, int mode)
908 : : {
5488 909 : 165 : volatile WalSndCtlData *walsndctl = WalSndCtl;
5453 bruce@momjian.us 910 : 165 : int numprocs = 0;
911 : : dlist_mutable_iter iter;
912 : :
5164 simon@2ndQuadrant.co 913 [ + - - + ]: 165 : Assert(mode >= 0 && mode < NUM_SYNC_REP_WAIT_MODE);
2326 michael@paquier.xyz 914 [ - + ]: 165 : Assert(LWLockHeldByMeInMode(SyncRepLock, LW_EXCLUSIVE));
5164 simon@2ndQuadrant.co 915 [ - + ]: 165 : Assert(SyncRepQueueIsOrderedByLSN(mode));
916 : :
1152 andres@anarazel.de 917 [ + - + + ]: 190 : dlist_foreach_modify(iter, &WalSndCtl->SyncRepQueue[mode])
918 : : {
1031 tgl@sss.pgh.pa.us 919 : 28 : PGPROC *proc = dlist_container(PGPROC, syncRepLinks, iter.cur);
920 : :
921 : : /*
922 : : * Assume the queue is ordered by LSN
923 : : */
4825 alvherre@alvh.no-ip. 924 [ + - + + ]: 28 : if (!all && walsndctl->lsn[mode] < proc->waitLSN)
5488 simon@2ndQuadrant.co 925 : 3 : return numprocs;
926 : :
927 : : /*
928 : : * Remove from queue.
929 : : */
1152 andres@anarazel.de 930 : 25 : dlist_delete_thoroughly(&proc->syncRepLinks);
931 : :
932 : : /*
933 : : * SyncRepWaitForLSN() reads syncRepState without holding the lock, so
934 : : * make sure that it sees the queue link being removed before the
935 : : * syncRepState change.
936 : : */
3168 heikki.linnakangas@i 937 : 25 : pg_write_barrier();
938 : :
939 : : /*
940 : : * Set state to complete; see SyncRepWaitForLSN() for discussion of
941 : : * the various states.
942 : : */
1152 andres@anarazel.de 943 : 25 : proc->syncRepState = SYNC_REP_WAIT_COMPLETE;
944 : :
945 : : /*
946 : : * Wake only when we have set state and removed from queue.
947 : : */
948 : 25 : SetLatch(&(proc->procLatch));
949 : :
5488 simon@2ndQuadrant.co 950 : 25 : numprocs++;
951 : : }
952 : :
953 : 162 : return numprocs;
954 : : }
955 : :
956 : : /*
957 : : * The checkpointer calls this as needed to update the shared
958 : : * sync_standbys_status flag, so that backends don't remain permanently wedged
959 : : * if synchronous_standby_names is unset. It's safe to check the current value
960 : : * without the lock, because it's only ever updated by one process. But we
961 : : * must take the lock to change it.
962 : : */
963 : : void
5477 rhaas@postgresql.org 964 : 654 : SyncRepUpdateSyncStandbysDefined(void)
965 : : {
966 [ + - + + ]: 654 : bool sync_standbys_defined = SyncStandbysDefined();
967 : :
338 michael@paquier.xyz 968 : 654 : if (sync_standbys_defined !=
969 [ + + ]: 654 : ((WalSndCtl->sync_standbys_status & SYNC_STANDBY_DEFINED) != 0))
970 : : {
5477 rhaas@postgresql.org 971 : 14 : LWLockAcquire(SyncRepLock, LW_EXCLUSIVE);
972 : :
973 : : /*
974 : : * If synchronous_standby_names has been reset to empty, it's futile
975 : : * for backends to continue waiting. Since the user no longer wants
976 : : * synchronous replication, we'd better wake them up.
977 : : */
978 [ + + ]: 14 : if (!sync_standbys_defined)
979 : : {
980 : : int i;
981 : :
5164 simon@2ndQuadrant.co 982 [ + + ]: 4 : for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; i++)
983 : 3 : SyncRepWakeQueue(true, i);
984 : : }
985 : :
986 : : /*
987 : : * Only allow people to join the queue when there are synchronous
988 : : * standbys defined. Without this interlock, there's a race
989 : : * condition: we might wake up all the current waiters; then, some
990 : : * backend that hasn't yet reloaded its config might go to sleep on
991 : : * the queue (and never wake up). This prevents that.
992 : : */
338 michael@paquier.xyz 993 [ + + ]: 14 : WalSndCtl->sync_standbys_status = SYNC_STANDBY_INIT |
994 : : (sync_standbys_defined ? SYNC_STANDBY_DEFINED : 0);
995 : :
996 : 14 : LWLockRelease(SyncRepLock);
997 : : }
998 [ + + ]: 640 : else if ((WalSndCtl->sync_standbys_status & SYNC_STANDBY_INIT) == 0)
999 : : {
1000 : 569 : LWLockAcquire(SyncRepLock, LW_EXCLUSIVE);
1001 : :
1002 : : /*
1003 : : * Note that there is no need to wake up the queues here. We would
1004 : : * reach this path only if SyncStandbysDefined() returns false, or it
1005 : : * would mean that some backends are waiting with the GUC set. See
1006 : : * SyncRepWaitForLSN().
1007 : : */
1008 [ + - - + ]: 569 : Assert(!SyncStandbysDefined());
1009 : :
1010 : : /*
1011 : : * Even if there is no sync standby defined, let the readers of this
1012 : : * information know that the sync standby data has been initialized.
1013 : : * This can just be done once, hence the previous check on
1014 : : * SYNC_STANDBY_INIT to avoid useless work.
1015 : : */
1016 : 569 : WalSndCtl->sync_standbys_status |= SYNC_STANDBY_INIT;
1017 : :
5477 rhaas@postgresql.org 1018 : 569 : LWLockRelease(SyncRepLock);
1019 : : }
1020 : 654 : }
1021 : :
1022 : : #ifdef USE_ASSERT_CHECKING
1023 : : static bool
5164 simon@2ndQuadrant.co 1024 : 200 : SyncRepQueueIsOrderedByLSN(int mode)
1025 : : {
1026 : : XLogRecPtr lastLSN;
1027 : : dlist_iter iter;
1028 : :
1029 [ + - - + ]: 200 : Assert(mode >= 0 && mode < NUM_SYNC_REP_WAIT_MODE);
1030 : :
45 alvherre@kurilemu.de 1031 :GNC 200 : lastLSN = InvalidXLogRecPtr;
1032 : :
1152 andres@anarazel.de 1033 [ + - + + ]:CBC 263 : dlist_foreach(iter, &WalSndCtl->SyncRepQueue[mode])
1034 : : {
1035 : 63 : PGPROC *proc = dlist_container(PGPROC, syncRepLinks, iter.cur);
1036 : :
1037 : : /*
1038 : : * Check the queue is ordered by LSN and that multiple procs don't
1039 : : * have matching LSNs
1040 : : */
4825 alvherre@alvh.no-ip. 1041 [ - + ]: 63 : if (proc->waitLSN <= lastLSN)
5488 simon@2ndQuadrant.co 1042 :UBC 0 : return false;
1043 : :
5488 simon@2ndQuadrant.co 1044 :CBC 63 : lastLSN = proc->waitLSN;
1045 : : }
1046 : :
1047 : 200 : return true;
1048 : : }
1049 : : #endif
1050 : :
1051 : : /*
1052 : : * ===========================================================
1053 : : * Synchronous Replication functions executed by any process
1054 : : * ===========================================================
1055 : : */
1056 : :
1057 : : bool
5456 tgl@sss.pgh.pa.us 1058 : 1272 : check_synchronous_standby_names(char **newval, void **extra, GucSource source)
1059 : : {
3630 fujii@postgresql.org 1060 [ + - + + ]: 1272 : if (*newval != NULL && (*newval)[0] != '\0')
5488 simon@2ndQuadrant.co 1061 : 76 : {
1062 : : yyscan_t scanner;
1063 : : int parse_rc;
1064 : : SyncRepConfigData *pconf;
1065 : :
1066 : : /* Result of parsing is returned in one of these two variables */
415 peter@eisentraut.org 1067 : 76 : SyncRepConfigData *syncrep_parse_result = NULL;
1068 : 76 : char *syncrep_parse_error_msg = NULL;
1069 : :
1070 : : /* Parse the synchronous_standby_names string */
468 1071 : 76 : syncrep_scanner_init(*newval, &scanner);
415 1072 : 76 : parse_rc = syncrep_yyparse(&syncrep_parse_result, &syncrep_parse_error_msg, scanner);
468 1073 : 76 : syncrep_scanner_finish(scanner);
1074 : :
3609 tgl@sss.pgh.pa.us 1075 [ + - - + ]: 76 : if (parse_rc != 0 || syncrep_parse_result == NULL)
1076 : : {
3630 fujii@postgresql.org 1077 :UBC 0 : GUC_check_errcode(ERRCODE_SYNTAX_ERROR);
3609 tgl@sss.pgh.pa.us 1078 [ # # ]: 0 : if (syncrep_parse_error_msg)
1079 : 0 : GUC_check_errdetail("%s", syncrep_parse_error_msg);
1080 : : else
1081 : : /* translator: %s is a GUC name */
473 alvherre@alvh.no-ip. 1082 : 0 : GUC_check_errdetail("\"%s\" parser failed.",
1083 : : "synchronous_standby_names");
3630 fujii@postgresql.org 1084 : 0 : return false;
1085 : : }
1086 : :
3375 fujii@postgresql.org 1087 [ - + ]:CBC 76 : if (syncrep_parse_result->num_sync <= 0)
1088 : : {
3375 fujii@postgresql.org 1089 :UBC 0 : GUC_check_errmsg("number of synchronous standbys (%d) must be greater than zero",
1090 : 0 : syncrep_parse_result->num_sync);
1091 : 0 : return false;
1092 : : }
1093 : :
1094 : : /* GUC extra value must be guc_malloc'd, not palloc'd */
1095 : : pconf = (SyncRepConfigData *)
1248 tgl@sss.pgh.pa.us 1096 :CBC 76 : guc_malloc(LOG, syncrep_parse_result->config_size);
3609 1097 [ - + ]: 76 : if (pconf == NULL)
3609 tgl@sss.pgh.pa.us 1098 :UBC 0 : return false;
3609 tgl@sss.pgh.pa.us 1099 :CBC 76 : memcpy(pconf, syncrep_parse_result, syncrep_parse_result->config_size);
1100 : :
472 peter@eisentraut.org 1101 : 76 : *extra = pconf;
1102 : :
1103 : : /*
1104 : : * We need not explicitly clean up syncrep_parse_result. It, and any
1105 : : * other cruft generated during parsing, will be freed when the
1106 : : * current memory context is deleted. (This code is generally run in
1107 : : * a short-lived context used for config file processing, so that will
1108 : : * not be very long.)
1109 : : */
1110 : : }
1111 : : else
3609 tgl@sss.pgh.pa.us 1112 : 1196 : *extra = NULL;
1113 : :
5456 1114 : 1272 : return true;
1115 : : }
1116 : :
1117 : : void
3609 1118 : 1262 : assign_synchronous_standby_names(const char *newval, void *extra)
1119 : : {
1120 : 1262 : SyncRepConfig = (SyncRepConfigData *) extra;
1121 : 1262 : }
1122 : :
1123 : : void
5164 simon@2ndQuadrant.co 1124 : 1954 : assign_synchronous_commit(int newval, void *extra)
1125 : : {
1126 [ - + + + ]: 1954 : switch (newval)
1127 : : {
5164 simon@2ndQuadrant.co 1128 :UBC 0 : case SYNCHRONOUS_COMMIT_REMOTE_WRITE:
1129 : 0 : SyncRepWaitMode = SYNC_REP_WAIT_WRITE;
1130 : 0 : break;
5164 simon@2ndQuadrant.co 1131 :CBC 1323 : case SYNCHRONOUS_COMMIT_REMOTE_FLUSH:
1132 : 1323 : SyncRepWaitMode = SYNC_REP_WAIT_FLUSH;
1133 : 1323 : break;
3638 rhaas@postgresql.org 1134 : 2 : case SYNCHRONOUS_COMMIT_REMOTE_APPLY:
1135 : 2 : SyncRepWaitMode = SYNC_REP_WAIT_APPLY;
1136 : 2 : break;
5164 simon@2ndQuadrant.co 1137 : 629 : default:
1138 : 629 : SyncRepWaitMode = SYNC_REP_NO_WAIT;
1139 : 629 : break;
1140 : : }
1141 : 1954 : }
|
