Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * checkpointer.c
4 : *
5 : * The checkpointer is new as of Postgres 9.2. It handles all checkpoints.
6 : * Checkpoints are automatically dispatched after a certain amount of time has
7 : * elapsed since the last one, and it can be signaled to perform requested
8 : * checkpoints as well. (The GUC parameter that mandates a checkpoint every
9 : * so many WAL segments is implemented by having backends signal when they
10 : * fill WAL segments; the checkpointer itself doesn't watch for the
11 : * condition.)
12 : *
13 : * The normal termination sequence is that checkpointer is instructed to
14 : * execute the shutdown checkpoint by SIGINT. After that checkpointer waits
15 : * to be terminated via SIGUSR2, which instructs the checkpointer to exit(0).
16 : * All backends must be stopped before SIGINT or SIGUSR2 is issued!
17 : *
18 : * Emergency termination is by SIGQUIT; like any backend, the checkpointer
19 : * will simply abort and exit on SIGQUIT.
20 : *
21 : * If the checkpointer exits unexpectedly, the postmaster treats that the same
22 : * as a backend crash: shared memory may be corrupted, so remaining backends
23 : * should be killed by SIGQUIT and then a recovery cycle started. (Even if
24 : * shared memory isn't corrupted, we have lost information about which
25 : * files need to be fsync'd for the next checkpoint, and so a system
26 : * restart needs to be forced.)
27 : *
28 : *
29 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
30 : *
31 : *
32 : * IDENTIFICATION
33 : * src/backend/postmaster/checkpointer.c
34 : *
35 : *-------------------------------------------------------------------------
36 : */
37 : #include "postgres.h"
38 :
39 : #include <sys/time.h>
40 : #include <time.h>
41 :
42 : #include "access/xlog.h"
43 : #include "access/xlog_internal.h"
44 : #include "access/xlogrecovery.h"
45 : #include "catalog/pg_authid.h"
46 : #include "commands/defrem.h"
47 : #include "libpq/pqsignal.h"
48 : #include "miscadmin.h"
49 : #include "pgstat.h"
50 : #include "postmaster/auxprocess.h"
51 : #include "postmaster/bgwriter.h"
52 : #include "postmaster/interrupt.h"
53 : #include "replication/syncrep.h"
54 : #include "storage/aio_subsys.h"
55 : #include "storage/bufmgr.h"
56 : #include "storage/condition_variable.h"
57 : #include "storage/fd.h"
58 : #include "storage/ipc.h"
59 : #include "storage/lwlock.h"
60 : #include "storage/pmsignal.h"
61 : #include "storage/proc.h"
62 : #include "storage/procsignal.h"
63 : #include "storage/shmem.h"
64 : #include "storage/smgr.h"
65 : #include "storage/spin.h"
66 : #include "utils/acl.h"
67 : #include "utils/guc.h"
68 : #include "utils/memutils.h"
69 : #include "utils/resowner.h"
70 :
71 :
72 : /*----------
73 : * Shared memory area for communication between checkpointer and backends
74 : *
75 : * The ckpt counters allow backends to watch for completion of a checkpoint
76 : * request they send. Here's how it works:
77 : * * At start of a checkpoint, checkpointer reads (and clears) the request
78 : * flags and increments ckpt_started, while holding ckpt_lck.
79 : * * On completion of a checkpoint, checkpointer sets ckpt_done to
80 : * equal ckpt_started.
81 : * * On failure of a checkpoint, checkpointer increments ckpt_failed
82 : * and sets ckpt_done to equal ckpt_started.
83 : *
84 : * The algorithm for backends is:
85 : * 1. Record current values of ckpt_failed and ckpt_started, and
86 : * set request flags, while holding ckpt_lck.
87 : * 2. Send signal to request checkpoint.
88 : * 3. Sleep until ckpt_started changes. Now you know a checkpoint has
89 : * begun since you started this algorithm (although *not* that it was
90 : * specifically initiated by your signal), and that it is using your flags.
91 : * 4. Record new value of ckpt_started.
92 : * 5. Sleep until ckpt_done >= saved value of ckpt_started. (Use modulo
93 : * arithmetic here in case counters wrap around.) Now you know a
94 : * checkpoint has started and completed, but not whether it was
95 : * successful.
96 : * 6. If ckpt_failed is different from the originally saved value,
97 : * assume request failed; otherwise it was definitely successful.
98 : *
99 : * ckpt_flags holds the OR of the checkpoint request flags sent by all
100 : * requesting backends since the last checkpoint start. The flags are
101 : * chosen so that OR'ing is the correct way to combine multiple requests.
102 : *
103 : * The requests array holds fsync requests sent by backends and not yet
104 : * absorbed by the checkpointer.
105 : *
106 : * Unlike the checkpoint fields, requests related fields are protected by
107 : * CheckpointerCommLock.
108 : *----------
109 : */
110 : typedef struct
111 : {
112 : SyncRequestType type; /* request type */
113 : FileTag ftag; /* file identifier */
114 : } CheckpointerRequest;
115 :
116 : typedef struct
117 : {
118 : pid_t checkpointer_pid; /* PID (0 if not started) */
119 :
120 : slock_t ckpt_lck; /* protects all the ckpt_* fields */
121 :
122 : int ckpt_started; /* advances when checkpoint starts */
123 : int ckpt_done; /* advances when checkpoint done */
124 : int ckpt_failed; /* advances when checkpoint fails */
125 :
126 : int ckpt_flags; /* checkpoint flags, as defined in xlog.h */
127 :
128 : ConditionVariable start_cv; /* signaled when ckpt_started advances */
129 : ConditionVariable done_cv; /* signaled when ckpt_done advances */
130 :
131 : int num_requests; /* current # of requests */
132 : int max_requests; /* allocated array size */
133 :
134 : int head; /* Index of the first request in the ring
135 : * buffer */
136 : int tail; /* Index of the last request in the ring
137 : * buffer */
138 :
139 : /* The ring buffer of pending checkpointer requests */
140 : CheckpointerRequest requests[FLEXIBLE_ARRAY_MEMBER];
141 : } CheckpointerShmemStruct;
142 :
143 : static CheckpointerShmemStruct *CheckpointerShmem;
144 :
145 : /* interval for calling AbsorbSyncRequests in CheckpointWriteDelay */
146 : #define WRITES_PER_ABSORB 1000
147 :
148 : /* Maximum number of checkpointer requests to process in one batch */
149 : #define CKPT_REQ_BATCH_SIZE 10000
150 :
151 : /* Max number of requests the checkpointer request queue can hold */
152 : #define MAX_CHECKPOINT_REQUESTS 10000000
153 :
154 : /*
155 : * GUC parameters
156 : */
157 : int CheckPointTimeout = 300;
158 : int CheckPointWarning = 30;
159 : double CheckPointCompletionTarget = 0.9;
160 :
161 : /*
162 : * Private state
163 : */
164 : static bool ckpt_active = false;
165 : static volatile sig_atomic_t ShutdownXLOGPending = false;
166 :
167 : /* these values are valid when ckpt_active is true: */
168 : static pg_time_t ckpt_start_time;
169 : static XLogRecPtr ckpt_start_recptr;
170 : static double ckpt_cached_elapsed;
171 :
172 : static pg_time_t last_checkpoint_time;
173 : static pg_time_t last_xlog_switch_time;
174 :
175 : /* Prototypes for private functions */
176 :
177 : static void ProcessCheckpointerInterrupts(void);
178 : static void CheckArchiveTimeout(void);
179 : static bool IsCheckpointOnSchedule(double progress);
180 : static bool FastCheckpointRequested(void);
181 : static bool CompactCheckpointerRequestQueue(void);
182 : static void UpdateSharedMemoryConfig(void);
183 :
184 : /* Signal handlers */
185 : static void ReqShutdownXLOG(SIGNAL_ARGS);
186 :
187 :
188 : /*
189 : * Main entry point for checkpointer process
190 : *
191 : * This is invoked from AuxiliaryProcessMain, which has already created the
192 : * basic execution environment, but not enabled signals yet.
193 : */
194 : void
195 1142 : CheckpointerMain(const void *startup_data, size_t startup_data_len)
196 : {
197 : sigjmp_buf local_sigjmp_buf;
198 : MemoryContext checkpointer_context;
199 :
200 : Assert(startup_data_len == 0);
201 :
202 1142 : AuxiliaryProcessMainCommon();
203 :
204 1142 : CheckpointerShmem->checkpointer_pid = MyProcPid;
205 :
206 : /*
207 : * Properly accept or ignore signals the postmaster might send us
208 : *
209 : * Note: we deliberately ignore SIGTERM, because during a standard Unix
210 : * system shutdown cycle, init will SIGTERM all processes at once. We
211 : * want to wait for the backends to exit, whereupon the postmaster will
212 : * tell us it's okay to shut down (via SIGUSR2).
213 : */
214 1142 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
215 1142 : pqsignal(SIGINT, ReqShutdownXLOG);
216 1142 : pqsignal(SIGTERM, SIG_IGN); /* ignore SIGTERM */
217 : /* SIGQUIT handler was already set up by InitPostmasterChild */
218 1142 : pqsignal(SIGALRM, SIG_IGN);
219 1142 : pqsignal(SIGPIPE, SIG_IGN);
220 1142 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
221 1142 : pqsignal(SIGUSR2, SignalHandlerForShutdownRequest);
222 :
223 : /*
224 : * Reset some signals that are accepted by postmaster but not here
225 : */
226 1142 : pqsignal(SIGCHLD, SIG_DFL);
227 :
228 : /*
229 : * Initialize so that first time-driven event happens at the correct time.
230 : */
231 1142 : last_checkpoint_time = last_xlog_switch_time = (pg_time_t) time(NULL);
232 :
233 : /*
234 : * Write out stats after shutdown. This needs to be called by exactly one
235 : * process during a normal shutdown, and since checkpointer is shut down
236 : * very late...
237 : *
238 : * While e.g. walsenders are active after the shutdown checkpoint has been
239 : * written (and thus could produce more stats), checkpointer stays around
240 : * after the shutdown checkpoint has been written. postmaster will only
241 : * signal checkpointer to exit after all processes that could emit stats
242 : * have been shut down.
243 : */
244 1142 : before_shmem_exit(pgstat_before_server_shutdown, 0);
245 :
246 : /*
247 : * Create a memory context that we will do all our work in. We do this so
248 : * that we can reset the context during error recovery and thereby avoid
249 : * possible memory leaks. Formerly this code just ran in
250 : * TopMemoryContext, but resetting that would be a really bad idea.
251 : */
252 1142 : checkpointer_context = AllocSetContextCreate(TopMemoryContext,
253 : "Checkpointer",
254 : ALLOCSET_DEFAULT_SIZES);
255 1142 : MemoryContextSwitchTo(checkpointer_context);
256 :
257 : /*
258 : * If an exception is encountered, processing resumes here.
259 : *
260 : * You might wonder why this isn't coded as an infinite loop around a
261 : * PG_TRY construct. The reason is that this is the bottom of the
262 : * exception stack, and so with PG_TRY there would be no exception handler
263 : * in force at all during the CATCH part. By leaving the outermost setjmp
264 : * always active, we have at least some chance of recovering from an error
265 : * during error recovery. (If we get into an infinite loop thereby, it
266 : * will soon be stopped by overflow of elog.c's internal state stack.)
267 : *
268 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
269 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
270 : * signals other than SIGQUIT will be blocked until we complete error
271 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
272 : * call redundant, but it is not since InterruptPending might be set
273 : * already.
274 : */
275 1142 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
276 : {
277 : /* Since not using PG_TRY, must reset error stack by hand */
278 0 : error_context_stack = NULL;
279 :
280 : /* Prevent interrupts while cleaning up */
281 0 : HOLD_INTERRUPTS();
282 :
283 : /* Report the error to the server log */
284 0 : EmitErrorReport();
285 :
286 : /*
287 : * These operations are really just a minimal subset of
288 : * AbortTransaction(). We don't have very many resources to worry
289 : * about in checkpointer, but we do have LWLocks, buffers, and temp
290 : * files.
291 : */
292 0 : LWLockReleaseAll();
293 0 : ConditionVariableCancelSleep();
294 0 : pgstat_report_wait_end();
295 0 : pgaio_error_cleanup();
296 0 : UnlockBuffers();
297 0 : ReleaseAuxProcessResources(false);
298 0 : AtEOXact_Buffers(false);
299 0 : AtEOXact_SMgr();
300 0 : AtEOXact_Files(false);
301 0 : AtEOXact_HashTables(false);
302 :
303 : /* Warn any waiting backends that the checkpoint failed. */
304 0 : if (ckpt_active)
305 : {
306 0 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
307 0 : CheckpointerShmem->ckpt_failed++;
308 0 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
309 0 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
310 :
311 0 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
312 :
313 0 : ckpt_active = false;
314 : }
315 :
316 : /*
317 : * Now return to normal top-level context and clear ErrorContext for
318 : * next time.
319 : */
320 0 : MemoryContextSwitchTo(checkpointer_context);
321 0 : FlushErrorState();
322 :
323 : /* Flush any leaked data in the top-level context */
324 0 : MemoryContextReset(checkpointer_context);
325 :
326 : /* Now we can allow interrupts again */
327 0 : RESUME_INTERRUPTS();
328 :
329 : /*
330 : * Sleep at least 1 second after any error. A write error is likely
331 : * to be repeated, and we don't want to be filling the error logs as
332 : * fast as we can.
333 : */
334 0 : pg_usleep(1000000L);
335 : }
336 :
337 : /* We can now handle ereport(ERROR) */
338 1142 : PG_exception_stack = &local_sigjmp_buf;
339 :
340 : /*
341 : * Unblock signals (they were blocked when the postmaster forked us)
342 : */
343 1142 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
344 :
345 : /*
346 : * Ensure all shared memory values are set correctly for the config. Doing
347 : * this here ensures no race conditions from other concurrent updaters.
348 : */
349 1142 : UpdateSharedMemoryConfig();
350 :
351 : /*
352 : * Advertise our proc number that backends can use to wake us up while
353 : * we're sleeping.
354 : */
355 1142 : ProcGlobal->checkpointerProc = MyProcNumber;
356 :
357 : /*
358 : * Loop until we've been asked to write the shutdown checkpoint or
359 : * terminate.
360 : */
361 : for (;;)
362 8752 : {
363 9894 : bool do_checkpoint = false;
364 9894 : int flags = 0;
365 : pg_time_t now;
366 : int elapsed_secs;
367 : int cur_timeout;
368 9894 : bool chkpt_or_rstpt_requested = false;
369 9894 : bool chkpt_or_rstpt_timed = false;
370 :
371 : /* Clear any already-pending wakeups */
372 9894 : ResetLatch(MyLatch);
373 :
374 : /*
375 : * Process any requests or signals received recently.
376 : */
377 9894 : AbsorbSyncRequests();
378 :
379 9894 : ProcessCheckpointerInterrupts();
380 9894 : if (ShutdownXLOGPending || ShutdownRequestPending)
381 : break;
382 :
383 : /*
384 : * Detect a pending checkpoint request by checking whether the flags
385 : * word in shared memory is nonzero. We shouldn't need to acquire the
386 : * ckpt_lck for this.
387 : */
388 8778 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
389 : {
390 2610 : do_checkpoint = true;
391 2610 : chkpt_or_rstpt_requested = true;
392 : }
393 :
394 : /*
395 : * Force a checkpoint if too much time has elapsed since the last one.
396 : * Note that we count a timed checkpoint in stats only when this
397 : * occurs without an external request, but we set the CAUSE_TIME flag
398 : * bit even if there is also an external request.
399 : */
400 8778 : now = (pg_time_t) time(NULL);
401 8778 : elapsed_secs = now - last_checkpoint_time;
402 8778 : if (elapsed_secs >= CheckPointTimeout)
403 : {
404 2 : if (!do_checkpoint)
405 2 : chkpt_or_rstpt_timed = true;
406 2 : do_checkpoint = true;
407 2 : flags |= CHECKPOINT_CAUSE_TIME;
408 : }
409 :
410 : /*
411 : * Do a checkpoint if requested.
412 : */
413 8778 : if (do_checkpoint)
414 : {
415 2612 : bool ckpt_performed = false;
416 : bool do_restartpoint;
417 :
418 : /* Check if we should perform a checkpoint or a restartpoint. */
419 2612 : do_restartpoint = RecoveryInProgress();
420 :
421 : /*
422 : * Atomically fetch the request flags to figure out what kind of a
423 : * checkpoint we should perform, and increase the started-counter
424 : * to acknowledge that we've started a new checkpoint.
425 : */
426 2612 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
427 2612 : flags |= CheckpointerShmem->ckpt_flags;
428 2612 : CheckpointerShmem->ckpt_flags = 0;
429 2612 : CheckpointerShmem->ckpt_started++;
430 2612 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
431 :
432 2612 : ConditionVariableBroadcast(&CheckpointerShmem->start_cv);
433 :
434 : /*
435 : * The end-of-recovery checkpoint is a real checkpoint that's
436 : * performed while we're still in recovery.
437 : */
438 2612 : if (flags & CHECKPOINT_END_OF_RECOVERY)
439 40 : do_restartpoint = false;
440 :
441 2612 : if (chkpt_or_rstpt_timed)
442 : {
443 2 : chkpt_or_rstpt_timed = false;
444 2 : if (do_restartpoint)
445 0 : PendingCheckpointerStats.restartpoints_timed++;
446 : else
447 2 : PendingCheckpointerStats.num_timed++;
448 : }
449 :
450 2612 : if (chkpt_or_rstpt_requested)
451 : {
452 2610 : chkpt_or_rstpt_requested = false;
453 2610 : if (do_restartpoint)
454 1102 : PendingCheckpointerStats.restartpoints_requested++;
455 : else
456 1508 : PendingCheckpointerStats.num_requested++;
457 : }
458 :
459 : /*
460 : * We will warn if (a) too soon since last checkpoint (whatever
461 : * caused it) and (b) somebody set the CHECKPOINT_CAUSE_XLOG flag
462 : * since the last checkpoint start. Note in particular that this
463 : * implementation will not generate warnings caused by
464 : * CheckPointTimeout < CheckPointWarning.
465 : */
466 2612 : if (!do_restartpoint &&
467 1510 : (flags & CHECKPOINT_CAUSE_XLOG) &&
468 390 : elapsed_secs < CheckPointWarning)
469 390 : ereport(LOG,
470 : (errmsg_plural("checkpoints are occurring too frequently (%d second apart)",
471 : "checkpoints are occurring too frequently (%d seconds apart)",
472 : elapsed_secs,
473 : elapsed_secs),
474 : errhint("Consider increasing the configuration parameter \"%s\".", "max_wal_size")));
475 :
476 : /*
477 : * Initialize checkpointer-private variables used during
478 : * checkpoint.
479 : */
480 2612 : ckpt_active = true;
481 2612 : if (do_restartpoint)
482 1102 : ckpt_start_recptr = GetXLogReplayRecPtr(NULL);
483 : else
484 1510 : ckpt_start_recptr = GetInsertRecPtr();
485 2612 : ckpt_start_time = now;
486 2612 : ckpt_cached_elapsed = 0;
487 :
488 : /*
489 : * Do the checkpoint.
490 : */
491 2612 : if (!do_restartpoint)
492 1510 : ckpt_performed = CreateCheckPoint(flags);
493 : else
494 1102 : ckpt_performed = CreateRestartPoint(flags);
495 :
496 : /*
497 : * After any checkpoint, free all smgr objects. Otherwise we
498 : * would never do so for dropped relations, as the checkpointer
499 : * does not process shared invalidation messages or call
500 : * AtEOXact_SMgr().
501 : */
502 2612 : smgrdestroyall();
503 :
504 : /*
505 : * Indicate checkpoint completion to any waiting backends.
506 : */
507 2612 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
508 2612 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
509 2612 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
510 :
511 2612 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
512 :
513 2612 : if (!do_restartpoint)
514 : {
515 : /*
516 : * Note we record the checkpoint start time not end time as
517 : * last_checkpoint_time. This is so that time-driven
518 : * checkpoints happen at a predictable spacing.
519 : */
520 1510 : last_checkpoint_time = now;
521 :
522 1510 : if (ckpt_performed)
523 1508 : PendingCheckpointerStats.num_performed++;
524 : }
525 : else
526 : {
527 1102 : if (ckpt_performed)
528 : {
529 : /*
530 : * The same as for checkpoint. Please see the
531 : * corresponding comment.
532 : */
533 354 : last_checkpoint_time = now;
534 :
535 354 : PendingCheckpointerStats.restartpoints_performed++;
536 : }
537 : else
538 : {
539 : /*
540 : * We were not able to perform the restartpoint
541 : * (checkpoints throw an ERROR in case of error). Most
542 : * likely because we have not received any new checkpoint
543 : * WAL records since the last restartpoint. Try again in
544 : * 15 s.
545 : */
546 748 : last_checkpoint_time = now - CheckPointTimeout + 15;
547 : }
548 : }
549 :
550 2612 : ckpt_active = false;
551 :
552 : /*
553 : * We may have received an interrupt during the checkpoint and the
554 : * latch might have been reset (e.g. in CheckpointWriteDelay).
555 : */
556 2612 : ProcessCheckpointerInterrupts();
557 2612 : if (ShutdownXLOGPending || ShutdownRequestPending)
558 : break;
559 : }
560 :
561 : /*
562 : * Disable logical decoding if someone requested it. See comments atop
563 : * logicalctl.c.
564 : */
565 8764 : DisableLogicalDecodingIfNecessary();
566 :
567 : /* Check for archive_timeout and switch xlog files if necessary. */
568 8764 : CheckArchiveTimeout();
569 :
570 : /* Report pending statistics to the cumulative stats system */
571 8764 : pgstat_report_checkpointer();
572 8764 : pgstat_report_wal(true);
573 :
574 : /*
575 : * If any checkpoint flags have been set, redo the loop to handle the
576 : * checkpoint without sleeping.
577 : */
578 8764 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
579 444 : continue;
580 :
581 : /*
582 : * Sleep until we are signaled or it's time for another checkpoint or
583 : * xlog file switch.
584 : */
585 8320 : now = (pg_time_t) time(NULL);
586 8320 : elapsed_secs = now - last_checkpoint_time;
587 8320 : if (elapsed_secs >= CheckPointTimeout)
588 0 : continue; /* no sleep for us ... */
589 8320 : cur_timeout = CheckPointTimeout - elapsed_secs;
590 8320 : if (XLogArchiveTimeout > 0 && !RecoveryInProgress())
591 : {
592 0 : elapsed_secs = now - last_xlog_switch_time;
593 0 : if (elapsed_secs >= XLogArchiveTimeout)
594 0 : continue; /* no sleep for us ... */
595 0 : cur_timeout = Min(cur_timeout, XLogArchiveTimeout - elapsed_secs);
596 : }
597 :
598 8320 : (void) WaitLatch(MyLatch,
599 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
600 : cur_timeout * 1000L /* convert to ms */ ,
601 : WAIT_EVENT_CHECKPOINTER_MAIN);
602 : }
603 :
604 : /*
605 : * From here on, elog(ERROR) should end with exit(1), not send control
606 : * back to the sigsetjmp block above.
607 : */
608 1130 : ExitOnAnyError = true;
609 :
610 1130 : if (ShutdownXLOGPending)
611 : {
612 : /*
613 : * Close down the database.
614 : *
615 : * Since ShutdownXLOG() creates restartpoint or checkpoint, and
616 : * updates the statistics, increment the checkpoint request and flush
617 : * out pending statistic.
618 : */
619 1130 : PendingCheckpointerStats.num_requested++;
620 1130 : ShutdownXLOG(0, 0);
621 1130 : pgstat_report_checkpointer();
622 1130 : pgstat_report_wal(true);
623 :
624 : /*
625 : * Tell postmaster that we're done.
626 : */
627 1130 : SendPostmasterSignal(PMSIGNAL_XLOG_IS_SHUTDOWN);
628 1130 : ShutdownXLOGPending = false;
629 : }
630 :
631 : /*
632 : * Wait until we're asked to shut down. By separating the writing of the
633 : * shutdown checkpoint from checkpointer exiting, checkpointer can perform
634 : * some should-be-as-late-as-possible work like writing out stats.
635 : */
636 : for (;;)
637 : {
638 : /* Clear any already-pending wakeups */
639 2258 : ResetLatch(MyLatch);
640 :
641 2258 : ProcessCheckpointerInterrupts();
642 :
643 2258 : if (ShutdownRequestPending)
644 1130 : break;
645 :
646 1128 : (void) WaitLatch(MyLatch,
647 : WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
648 : 0,
649 : WAIT_EVENT_CHECKPOINTER_SHUTDOWN);
650 : }
651 :
652 : /* Normal exit from the checkpointer is here */
653 1130 : proc_exit(0); /* done */
654 : }
655 :
656 : /*
657 : * Process any new interrupts.
658 : */
659 : static void
660 14764 : ProcessCheckpointerInterrupts(void)
661 : {
662 14764 : if (ProcSignalBarrierPending)
663 586 : ProcessProcSignalBarrier();
664 :
665 14764 : if (ConfigReloadPending)
666 : {
667 134 : ConfigReloadPending = false;
668 134 : ProcessConfigFile(PGC_SIGHUP);
669 :
670 : /*
671 : * Checkpointer is the last process to shut down, so we ask it to hold
672 : * the keys for a range of other tasks required most of which have
673 : * nothing to do with checkpointing at all.
674 : *
675 : * For various reasons, some config values can change dynamically so
676 : * the primary copy of them is held in shared memory to make sure all
677 : * backends see the same value. We make Checkpointer responsible for
678 : * updating the shared memory copy if the parameter setting changes
679 : * because of SIGHUP.
680 : */
681 134 : UpdateSharedMemoryConfig();
682 : }
683 :
684 : /* Perform logging of memory contexts of this process */
685 14764 : if (LogMemoryContextPending)
686 2 : ProcessLogMemoryContextInterrupt();
687 14764 : }
688 :
689 : /*
690 : * CheckArchiveTimeout -- check for archive_timeout and switch xlog files
691 : *
692 : * This will switch to a new WAL file and force an archive file write if
693 : * meaningful activity is recorded in the current WAL file. This includes most
694 : * writes, including just a single checkpoint record, but excludes WAL records
695 : * that were inserted with the XLOG_MARK_UNIMPORTANT flag being set (like
696 : * snapshots of running transactions). Such records, depending on
697 : * configuration, occur on regular intervals and don't contain important
698 : * information. This avoids generating archives with a few unimportant
699 : * records.
700 : */
701 : static void
702 26128 : CheckArchiveTimeout(void)
703 : {
704 : pg_time_t now;
705 : pg_time_t last_time;
706 : XLogRecPtr last_switch_lsn;
707 :
708 26128 : if (XLogArchiveTimeout <= 0 || RecoveryInProgress())
709 26128 : return;
710 :
711 0 : now = (pg_time_t) time(NULL);
712 :
713 : /* First we do a quick check using possibly-stale local state. */
714 0 : if ((int) (now - last_xlog_switch_time) < XLogArchiveTimeout)
715 0 : return;
716 :
717 : /*
718 : * Update local state ... note that last_xlog_switch_time is the last time
719 : * a switch was performed *or requested*.
720 : */
721 0 : last_time = GetLastSegSwitchData(&last_switch_lsn);
722 :
723 0 : last_xlog_switch_time = Max(last_xlog_switch_time, last_time);
724 :
725 : /* Now we can do the real checks */
726 0 : if ((int) (now - last_xlog_switch_time) >= XLogArchiveTimeout)
727 : {
728 : /*
729 : * Switch segment only when "important" WAL has been logged since the
730 : * last segment switch (last_switch_lsn points to end of segment
731 : * switch occurred in).
732 : */
733 0 : if (GetLastImportantRecPtr() > last_switch_lsn)
734 : {
735 : XLogRecPtr switchpoint;
736 :
737 : /* mark switch as unimportant, avoids triggering checkpoints */
738 0 : switchpoint = RequestXLogSwitch(true);
739 :
740 : /*
741 : * If the returned pointer points exactly to a segment boundary,
742 : * assume nothing happened.
743 : */
744 0 : if (XLogSegmentOffset(switchpoint, wal_segment_size) != 0)
745 0 : elog(DEBUG1, "write-ahead log switch forced (\"archive_timeout\"=%d)",
746 : XLogArchiveTimeout);
747 : }
748 :
749 : /*
750 : * Update state in any case, so we don't retry constantly when the
751 : * system is idle.
752 : */
753 0 : last_xlog_switch_time = now;
754 : }
755 : }
756 :
757 : /*
758 : * Returns true if a fast checkpoint request is pending. (Note that this does
759 : * not check the *current* checkpoint's FAST flag, but whether there is one
760 : * pending behind it.)
761 : */
762 : static bool
763 93294 : FastCheckpointRequested(void)
764 : {
765 93294 : volatile CheckpointerShmemStruct *cps = CheckpointerShmem;
766 :
767 : /*
768 : * We don't need to acquire the ckpt_lck in this case because we're only
769 : * looking at a single flag bit.
770 : */
771 93294 : if (cps->ckpt_flags & CHECKPOINT_FAST)
772 14876 : return true;
773 78418 : return false;
774 : }
775 :
776 : /*
777 : * CheckpointWriteDelay -- control rate of checkpoint
778 : *
779 : * This function is called after each page write performed by BufferSync().
780 : * It is responsible for throttling BufferSync()'s write rate to hit
781 : * checkpoint_completion_target.
782 : *
783 : * The checkpoint request flags should be passed in; currently the only one
784 : * examined is CHECKPOINT_FAST, which disables delays between writes.
785 : *
786 : * 'progress' is an estimate of how much of the work has been done, as a
787 : * fraction between 0.0 meaning none, and 1.0 meaning all done.
788 : */
789 : void
790 589912 : CheckpointWriteDelay(int flags, double progress)
791 : {
792 : static int absorb_counter = WRITES_PER_ABSORB;
793 :
794 : /* Do nothing if checkpoint is being executed by non-checkpointer process */
795 589912 : if (!AmCheckpointerProcess())
796 103450 : return;
797 :
798 : /*
799 : * Perform the usual duties and take a nap, unless we're behind schedule,
800 : * in which case we just try to catch up as quickly as possible.
801 : */
802 486462 : if (!(flags & CHECKPOINT_FAST) &&
803 93670 : !ShutdownXLOGPending &&
804 93294 : !ShutdownRequestPending &&
805 171712 : !FastCheckpointRequested() &&
806 78418 : IsCheckpointOnSchedule(progress))
807 : {
808 17364 : if (ConfigReloadPending)
809 : {
810 0 : ConfigReloadPending = false;
811 0 : ProcessConfigFile(PGC_SIGHUP);
812 : /* update shmem copies of config variables */
813 0 : UpdateSharedMemoryConfig();
814 : }
815 :
816 17364 : AbsorbSyncRequests();
817 17364 : absorb_counter = WRITES_PER_ABSORB;
818 :
819 17364 : CheckArchiveTimeout();
820 :
821 : /* Report interim statistics to the cumulative stats system */
822 17364 : pgstat_report_checkpointer();
823 :
824 : /*
825 : * This sleep used to be connected to bgwriter_delay, typically 200ms.
826 : * That resulted in more frequent wakeups if not much work to do.
827 : * Checkpointer and bgwriter are no longer related so take the Big
828 : * Sleep.
829 : */
830 17364 : WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | WL_TIMEOUT,
831 : 100,
832 : WAIT_EVENT_CHECKPOINT_WRITE_DELAY);
833 17364 : ResetLatch(MyLatch);
834 : }
835 469098 : else if (--absorb_counter <= 0)
836 : {
837 : /*
838 : * Absorb pending fsync requests after each WRITES_PER_ABSORB write
839 : * operations even when we don't sleep, to prevent overflow of the
840 : * fsync request queue.
841 : */
842 204 : AbsorbSyncRequests();
843 204 : absorb_counter = WRITES_PER_ABSORB;
844 : }
845 :
846 : /* Check for barrier events. */
847 486462 : if (ProcSignalBarrierPending)
848 14 : ProcessProcSignalBarrier();
849 : }
850 :
851 : /*
852 : * IsCheckpointOnSchedule -- are we on schedule to finish this checkpoint
853 : * (or restartpoint) in time?
854 : *
855 : * Compares the current progress against the time/segments elapsed since last
856 : * checkpoint, and returns true if the progress we've made this far is greater
857 : * than the elapsed time/segments.
858 : */
859 : static bool
860 78418 : IsCheckpointOnSchedule(double progress)
861 : {
862 : XLogRecPtr recptr;
863 : struct timeval now;
864 : double elapsed_xlogs,
865 : elapsed_time;
866 :
867 : Assert(ckpt_active);
868 :
869 : /* Scale progress according to checkpoint_completion_target. */
870 78418 : progress *= CheckPointCompletionTarget;
871 :
872 : /*
873 : * Check against the cached value first. Only do the more expensive
874 : * calculations once we reach the target previously calculated. Since
875 : * neither time or WAL insert pointer moves backwards, a freshly
876 : * calculated value can only be greater than or equal to the cached value.
877 : */
878 78418 : if (progress < ckpt_cached_elapsed)
879 53468 : return false;
880 :
881 : /*
882 : * Check progress against WAL segments written and CheckPointSegments.
883 : *
884 : * We compare the current WAL insert location against the location
885 : * computed before calling CreateCheckPoint. The code in XLogInsert that
886 : * actually triggers a checkpoint when CheckPointSegments is exceeded
887 : * compares against RedoRecPtr, so this is not completely accurate.
888 : * However, it's good enough for our purposes, we're only calculating an
889 : * estimate anyway.
890 : *
891 : * During recovery, we compare last replayed WAL record's location with
892 : * the location computed before calling CreateRestartPoint. That maintains
893 : * the same pacing as we have during checkpoints in normal operation, but
894 : * we might exceed max_wal_size by a fair amount. That's because there can
895 : * be a large gap between a checkpoint's redo-pointer and the checkpoint
896 : * record itself, and we only start the restartpoint after we've seen the
897 : * checkpoint record. (The gap is typically up to CheckPointSegments *
898 : * checkpoint_completion_target where checkpoint_completion_target is the
899 : * value that was in effect when the WAL was generated).
900 : */
901 24950 : if (RecoveryInProgress())
902 12114 : recptr = GetXLogReplayRecPtr(NULL);
903 : else
904 12836 : recptr = GetInsertRecPtr();
905 24950 : elapsed_xlogs = (((double) (recptr - ckpt_start_recptr)) /
906 24950 : wal_segment_size) / CheckPointSegments;
907 :
908 24950 : if (progress < elapsed_xlogs)
909 : {
910 7586 : ckpt_cached_elapsed = elapsed_xlogs;
911 7586 : return false;
912 : }
913 :
914 : /*
915 : * Check progress against time elapsed and checkpoint_timeout.
916 : */
917 17364 : gettimeofday(&now, NULL);
918 17364 : elapsed_time = ((double) ((pg_time_t) now.tv_sec - ckpt_start_time) +
919 17364 : now.tv_usec / 1000000.0) / CheckPointTimeout;
920 :
921 17364 : if (progress < elapsed_time)
922 : {
923 0 : ckpt_cached_elapsed = elapsed_time;
924 0 : return false;
925 : }
926 :
927 : /* It looks like we're on schedule. */
928 17364 : return true;
929 : }
930 :
931 :
932 : /* --------------------------------
933 : * signal handler routines
934 : * --------------------------------
935 : */
936 :
937 : /* SIGINT: set flag to trigger writing of shutdown checkpoint */
938 : static void
939 1132 : ReqShutdownXLOG(SIGNAL_ARGS)
940 : {
941 1132 : ShutdownXLOGPending = true;
942 1132 : SetLatch(MyLatch);
943 1132 : }
944 :
945 :
946 : /* --------------------------------
947 : * communication with backends
948 : * --------------------------------
949 : */
950 :
951 : /*
952 : * CheckpointerShmemSize
953 : * Compute space needed for checkpointer-related shared memory
954 : */
955 : Size
956 6534 : CheckpointerShmemSize(void)
957 : {
958 : Size size;
959 :
960 : /*
961 : * The size of the requests[] array is arbitrarily set equal to NBuffers.
962 : * But there is a cap of MAX_CHECKPOINT_REQUESTS to prevent accumulating
963 : * too many checkpoint requests in the ring buffer.
964 : */
965 6534 : size = offsetof(CheckpointerShmemStruct, requests);
966 6534 : size = add_size(size, mul_size(Min(NBuffers,
967 : MAX_CHECKPOINT_REQUESTS),
968 : sizeof(CheckpointerRequest)));
969 :
970 6534 : return size;
971 : }
972 :
973 : /*
974 : * CheckpointerShmemInit
975 : * Allocate and initialize checkpointer-related shared memory
976 : */
977 : void
978 2280 : CheckpointerShmemInit(void)
979 : {
980 2280 : Size size = CheckpointerShmemSize();
981 : bool found;
982 :
983 2280 : CheckpointerShmem = (CheckpointerShmemStruct *)
984 2280 : ShmemInitStruct("Checkpointer Data",
985 : size,
986 : &found);
987 :
988 2280 : if (!found)
989 : {
990 : /*
991 : * First time through, so initialize. Note that we zero the whole
992 : * requests array; this is so that CompactCheckpointerRequestQueue can
993 : * assume that any pad bytes in the request structs are zeroes.
994 : */
995 2568 : MemSet(CheckpointerShmem, 0, size);
996 2280 : SpinLockInit(&CheckpointerShmem->ckpt_lck);
997 2280 : CheckpointerShmem->max_requests = Min(NBuffers, MAX_CHECKPOINT_REQUESTS);
998 2280 : CheckpointerShmem->head = CheckpointerShmem->tail = 0;
999 2280 : ConditionVariableInit(&CheckpointerShmem->start_cv);
1000 2280 : ConditionVariableInit(&CheckpointerShmem->done_cv);
1001 : }
1002 2280 : }
1003 :
1004 : /*
1005 : * ExecCheckpoint
1006 : * Primary entry point for manual CHECKPOINT commands
1007 : *
1008 : * This is mainly a wrapper for RequestCheckpoint().
1009 : */
1010 : void
1011 896 : ExecCheckpoint(ParseState *pstate, CheckPointStmt *stmt)
1012 : {
1013 896 : bool fast = true;
1014 896 : bool unlogged = false;
1015 :
1016 1798 : foreach_ptr(DefElem, opt, stmt->options)
1017 : {
1018 30 : if (strcmp(opt->defname, "mode") == 0)
1019 : {
1020 12 : char *mode = defGetString(opt);
1021 :
1022 12 : if (strcmp(mode, "spread") == 0)
1023 0 : fast = false;
1024 12 : else if (strcmp(mode, "fast") != 0)
1025 6 : ereport(ERROR,
1026 : (errcode(ERRCODE_SYNTAX_ERROR),
1027 : errmsg("unrecognized value for %s option \"%s\": \"%s\"",
1028 : "CHECKPOINT", "mode", mode),
1029 : parser_errposition(pstate, opt->location)));
1030 : }
1031 18 : else if (strcmp(opt->defname, "flush_unlogged") == 0)
1032 12 : unlogged = defGetBoolean(opt);
1033 : else
1034 6 : ereport(ERROR,
1035 : (errcode(ERRCODE_SYNTAX_ERROR),
1036 : errmsg("unrecognized %s option \"%s\"",
1037 : "CHECKPOINT", opt->defname),
1038 : parser_errposition(pstate, opt->location)));
1039 : }
1040 :
1041 884 : if (!has_privs_of_role(GetUserId(), ROLE_PG_CHECKPOINT))
1042 0 : ereport(ERROR,
1043 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
1044 : /* translator: %s is name of an SQL command (e.g., CHECKPOINT) */
1045 : errmsg("permission denied to execute %s command",
1046 : "CHECKPOINT"),
1047 : errdetail("Only roles with privileges of the \"%s\" role may execute this command.",
1048 : "pg_checkpoint")));
1049 :
1050 1768 : RequestCheckpoint(CHECKPOINT_WAIT |
1051 884 : (fast ? CHECKPOINT_FAST : 0) |
1052 884 : (unlogged ? CHECKPOINT_FLUSH_UNLOGGED : 0) |
1053 884 : (RecoveryInProgress() ? 0 : CHECKPOINT_FORCE));
1054 884 : }
1055 :
1056 : /*
1057 : * RequestCheckpoint
1058 : * Called in backend processes to request a checkpoint
1059 : *
1060 : * flags is a bitwise OR of the following:
1061 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
1062 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
1063 : * CHECKPOINT_FAST: finish the checkpoint ASAP,
1064 : * ignoring checkpoint_completion_target parameter.
1065 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
1066 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
1067 : * CHECKPOINT_END_OF_RECOVERY, and the CHECKPOINT command).
1068 : * CHECKPOINT_WAIT: wait for completion before returning (otherwise,
1069 : * just signal checkpointer to do it, and return).
1070 : * CHECKPOINT_CAUSE_XLOG: checkpoint is requested due to xlog filling.
1071 : * (This affects logging, and in particular enables CheckPointWarning.)
1072 : */
1073 : void
1074 5134 : RequestCheckpoint(int flags)
1075 : {
1076 : int ntries;
1077 : int old_failed,
1078 : old_started;
1079 :
1080 : /*
1081 : * If in a standalone backend, just do it ourselves.
1082 : */
1083 5134 : if (!IsPostmasterEnvironment)
1084 : {
1085 : /*
1086 : * There's no point in doing slow checkpoints in a standalone backend,
1087 : * because there's no other backends the checkpoint could disrupt.
1088 : */
1089 410 : CreateCheckPoint(flags | CHECKPOINT_FAST);
1090 :
1091 : /* Free all smgr objects, as CheckpointerMain() normally would. */
1092 410 : smgrdestroyall();
1093 :
1094 410 : return;
1095 : }
1096 :
1097 : /*
1098 : * Atomically set the request flags, and take a snapshot of the counters.
1099 : * When we see ckpt_started > old_started, we know the flags we set here
1100 : * have been seen by checkpointer.
1101 : *
1102 : * Note that we OR the flags with any existing flags, to avoid overriding
1103 : * a "stronger" request by another backend. The flag senses must be
1104 : * chosen to make this work!
1105 : */
1106 4724 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1107 :
1108 4724 : old_failed = CheckpointerShmem->ckpt_failed;
1109 4724 : old_started = CheckpointerShmem->ckpt_started;
1110 4724 : CheckpointerShmem->ckpt_flags |= (flags | CHECKPOINT_REQUESTED);
1111 :
1112 4724 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1113 :
1114 : /*
1115 : * Set checkpointer's latch to request checkpoint. It's possible that the
1116 : * checkpointer hasn't started yet, so we will retry a few times if
1117 : * needed. (Actually, more than a few times, since on slow or overloaded
1118 : * buildfarm machines, it's been observed that the checkpointer can take
1119 : * several seconds to start.) However, if not told to wait for the
1120 : * checkpoint to occur, we consider failure to set the latch to be
1121 : * nonfatal and merely LOG it. The checkpointer should see the request
1122 : * when it does start, with or without the SetLatch().
1123 : */
1124 : #define MAX_SIGNAL_TRIES 600 /* max wait 60.0 sec */
1125 4724 : for (ntries = 0;; ntries++)
1126 56 : {
1127 4780 : volatile PROC_HDR *procglobal = ProcGlobal;
1128 4780 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1129 :
1130 4780 : if (checkpointerProc == INVALID_PROC_NUMBER)
1131 : {
1132 66 : if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT))
1133 : {
1134 10 : elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG,
1135 : "could not notify checkpoint: checkpointer is not running");
1136 10 : break;
1137 : }
1138 : }
1139 : else
1140 : {
1141 4714 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1142 : /* notified successfully */
1143 4714 : break;
1144 : }
1145 :
1146 56 : CHECK_FOR_INTERRUPTS();
1147 56 : pg_usleep(100000L); /* wait 0.1 sec, then retry */
1148 : }
1149 :
1150 : /*
1151 : * If requested, wait for completion. We detect completion according to
1152 : * the algorithm given above.
1153 : */
1154 4724 : if (flags & CHECKPOINT_WAIT)
1155 : {
1156 : int new_started,
1157 : new_failed;
1158 :
1159 : /* Wait for a new checkpoint to start. */
1160 1650 : ConditionVariablePrepareToSleep(&CheckpointerShmem->start_cv);
1161 : for (;;)
1162 : {
1163 2922 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1164 2922 : new_started = CheckpointerShmem->ckpt_started;
1165 2922 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1166 :
1167 2922 : if (new_started != old_started)
1168 1650 : break;
1169 :
1170 1272 : ConditionVariableSleep(&CheckpointerShmem->start_cv,
1171 : WAIT_EVENT_CHECKPOINT_START);
1172 : }
1173 1650 : ConditionVariableCancelSleep();
1174 :
1175 : /*
1176 : * We are waiting for ckpt_done >= new_started, in a modulo sense.
1177 : */
1178 1650 : ConditionVariablePrepareToSleep(&CheckpointerShmem->done_cv);
1179 : for (;;)
1180 1244 : {
1181 : int new_done;
1182 :
1183 2894 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1184 2894 : new_done = CheckpointerShmem->ckpt_done;
1185 2894 : new_failed = CheckpointerShmem->ckpt_failed;
1186 2894 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1187 :
1188 2894 : if (new_done - new_started >= 0)
1189 1650 : break;
1190 :
1191 1244 : ConditionVariableSleep(&CheckpointerShmem->done_cv,
1192 : WAIT_EVENT_CHECKPOINT_DONE);
1193 : }
1194 1650 : ConditionVariableCancelSleep();
1195 :
1196 1650 : if (new_failed != old_failed)
1197 0 : ereport(ERROR,
1198 : (errmsg("checkpoint request failed"),
1199 : errhint("Consult recent messages in the server log for details.")));
1200 : }
1201 : }
1202 :
1203 : /*
1204 : * ForwardSyncRequest
1205 : * Forward a file-fsync request from a backend to the checkpointer
1206 : *
1207 : * Whenever a backend is compelled to write directly to a relation
1208 : * (which should be seldom, if the background writer is getting its job done),
1209 : * the backend calls this routine to pass over knowledge that the relation
1210 : * is dirty and must be fsync'd before next checkpoint. We also use this
1211 : * opportunity to count such writes for statistical purposes.
1212 : *
1213 : * To avoid holding the lock for longer than necessary, we normally write
1214 : * to the requests[] queue without checking for duplicates. The checkpointer
1215 : * will have to eliminate dups internally anyway. However, if we discover
1216 : * that the queue is full, we make a pass over the entire queue to compact
1217 : * it. This is somewhat expensive, but the alternative is for the backend
1218 : * to perform its own fsync, which is far more expensive in practice. It
1219 : * is theoretically possible a backend fsync might still be necessary, if
1220 : * the queue is full and contains no duplicate entries. In that case, we
1221 : * let the backend know by returning false.
1222 : */
1223 : bool
1224 2524840 : ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)
1225 : {
1226 : CheckpointerRequest *request;
1227 : bool too_full;
1228 : int insert_pos;
1229 :
1230 2524840 : if (!IsUnderPostmaster)
1231 0 : return false; /* probably shouldn't even get here */
1232 :
1233 2524840 : if (AmCheckpointerProcess())
1234 0 : elog(ERROR, "ForwardSyncRequest must not be called in checkpointer");
1235 :
1236 2524840 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1237 :
1238 : /*
1239 : * If the checkpointer isn't running or the request queue is full, the
1240 : * backend will have to perform its own fsync request. But before forcing
1241 : * that to happen, we can try to compact the request queue.
1242 : */
1243 2524840 : if (CheckpointerShmem->checkpointer_pid == 0 ||
1244 2521586 : (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&
1245 2470 : !CompactCheckpointerRequestQueue()))
1246 : {
1247 5068 : LWLockRelease(CheckpointerCommLock);
1248 5068 : return false;
1249 : }
1250 :
1251 : /* OK, insert request */
1252 2519772 : insert_pos = CheckpointerShmem->tail;
1253 2519772 : request = &CheckpointerShmem->requests[insert_pos];
1254 2519772 : request->ftag = *ftag;
1255 2519772 : request->type = type;
1256 :
1257 2519772 : CheckpointerShmem->tail = (CheckpointerShmem->tail + 1) % CheckpointerShmem->max_requests;
1258 2519772 : CheckpointerShmem->num_requests++;
1259 :
1260 : /* If queue is more than half full, nudge the checkpointer to empty it */
1261 2519772 : too_full = (CheckpointerShmem->num_requests >=
1262 2519772 : CheckpointerShmem->max_requests / 2);
1263 :
1264 2519772 : LWLockRelease(CheckpointerCommLock);
1265 :
1266 : /* ... but not till after we release the lock */
1267 2519772 : if (too_full)
1268 : {
1269 63886 : volatile PROC_HDR *procglobal = ProcGlobal;
1270 63886 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1271 :
1272 63886 : if (checkpointerProc != INVALID_PROC_NUMBER)
1273 63886 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1274 : }
1275 :
1276 2519772 : return true;
1277 : }
1278 :
1279 : /*
1280 : * CompactCheckpointerRequestQueue
1281 : * Remove duplicates from the request queue to avoid backend fsyncs.
1282 : * Returns "true" if any entries were removed.
1283 : *
1284 : * Although a full fsync request queue is not common, it can lead to severe
1285 : * performance problems when it does happen. So far, this situation has
1286 : * only been observed to occur when the system is under heavy write load,
1287 : * and especially during the "sync" phase of a checkpoint. Without this
1288 : * logic, each backend begins doing an fsync for every block written, which
1289 : * gets very expensive and can slow down the whole system.
1290 : *
1291 : * Trying to do this every time the queue is full could lose if there
1292 : * aren't any removable entries. But that should be vanishingly rare in
1293 : * practice: there's one queue entry per shared buffer.
1294 : */
1295 : static bool
1296 2470 : CompactCheckpointerRequestQueue(void)
1297 : {
1298 : struct CheckpointerSlotMapping
1299 : {
1300 : CheckpointerRequest request;
1301 : int ring_idx;
1302 : };
1303 :
1304 : int n;
1305 2470 : int num_skipped = 0;
1306 : int head;
1307 : int max_requests;
1308 : int num_requests;
1309 : int read_idx,
1310 : write_idx;
1311 : HASHCTL ctl;
1312 : HTAB *htab;
1313 : bool *skip_slot;
1314 :
1315 : /* must hold CheckpointerCommLock in exclusive mode */
1316 : Assert(LWLockHeldByMe(CheckpointerCommLock));
1317 :
1318 : /* Avoid memory allocations in a critical section. */
1319 2470 : if (CritSectionCount > 0)
1320 0 : return false;
1321 :
1322 2470 : max_requests = CheckpointerShmem->max_requests;
1323 2470 : num_requests = CheckpointerShmem->num_requests;
1324 :
1325 : /* Initialize skip_slot array */
1326 2470 : skip_slot = palloc0_array(bool, max_requests);
1327 :
1328 2470 : head = CheckpointerShmem->head;
1329 :
1330 : /* Initialize temporary hash table */
1331 2470 : ctl.keysize = sizeof(CheckpointerRequest);
1332 2470 : ctl.entrysize = sizeof(struct CheckpointerSlotMapping);
1333 2470 : ctl.hcxt = CurrentMemoryContext;
1334 :
1335 2470 : htab = hash_create("CompactCheckpointerRequestQueue",
1336 2470 : CheckpointerShmem->num_requests,
1337 : &ctl,
1338 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1339 :
1340 : /*
1341 : * The basic idea here is that a request can be skipped if it's followed
1342 : * by a later, identical request. It might seem more sensible to work
1343 : * backwards from the end of the queue and check whether a request is
1344 : * *preceded* by an earlier, identical request, in the hopes of doing less
1345 : * copying. But that might change the semantics, if there's an
1346 : * intervening SYNC_FORGET_REQUEST or SYNC_FILTER_REQUEST, so we do it
1347 : * this way. It would be possible to be even smarter if we made the code
1348 : * below understand the specific semantics of such requests (it could blow
1349 : * away preceding entries that would end up being canceled anyhow), but
1350 : * it's not clear that the extra complexity would buy us anything.
1351 : */
1352 2470 : read_idx = head;
1353 312134 : for (n = 0; n < num_requests; n++)
1354 : {
1355 : CheckpointerRequest *request;
1356 : struct CheckpointerSlotMapping *slotmap;
1357 : bool found;
1358 :
1359 : /*
1360 : * We use the request struct directly as a hashtable key. This
1361 : * assumes that any padding bytes in the structs are consistently the
1362 : * same, which should be okay because we zeroed them in
1363 : * CheckpointerShmemInit. Note also that RelFileLocator had better
1364 : * contain no pad bytes.
1365 : */
1366 309664 : request = &CheckpointerShmem->requests[read_idx];
1367 309664 : slotmap = hash_search(htab, request, HASH_ENTER, &found);
1368 309664 : if (found)
1369 : {
1370 : /* Duplicate, so mark the previous occurrence as skippable */
1371 25782 : skip_slot[slotmap->ring_idx] = true;
1372 25782 : num_skipped++;
1373 : }
1374 : /* Remember slot containing latest occurrence of this request value */
1375 309664 : slotmap->ring_idx = read_idx;
1376 :
1377 : /* Move to the next request in the ring buffer */
1378 309664 : read_idx = (read_idx + 1) % max_requests;
1379 : }
1380 :
1381 : /* Done with the hash table. */
1382 2470 : hash_destroy(htab);
1383 :
1384 : /* If no duplicates, we're out of luck. */
1385 2470 : if (!num_skipped)
1386 : {
1387 1814 : pfree(skip_slot);
1388 1814 : return false;
1389 : }
1390 :
1391 : /* We found some duplicates; remove them. */
1392 656 : read_idx = write_idx = head;
1393 78128 : for (n = 0; n < num_requests; n++)
1394 : {
1395 : /* If this slot is NOT skipped, keep it */
1396 77472 : if (!skip_slot[read_idx])
1397 : {
1398 : /* If the read and write positions are different, copy the request */
1399 51690 : if (write_idx != read_idx)
1400 29960 : CheckpointerShmem->requests[write_idx] =
1401 29960 : CheckpointerShmem->requests[read_idx];
1402 :
1403 : /* Advance the write position */
1404 51690 : write_idx = (write_idx + 1) % max_requests;
1405 : }
1406 :
1407 77472 : read_idx = (read_idx + 1) % max_requests;
1408 : }
1409 :
1410 : /*
1411 : * Update ring buffer state: head remains the same, tail moves, count
1412 : * decreases
1413 : */
1414 656 : CheckpointerShmem->tail = write_idx;
1415 656 : CheckpointerShmem->num_requests -= num_skipped;
1416 :
1417 656 : ereport(DEBUG1,
1418 : (errmsg_internal("compacted fsync request queue from %d entries to %d entries",
1419 : num_requests, CheckpointerShmem->num_requests)));
1420 :
1421 : /* Cleanup. */
1422 656 : pfree(skip_slot);
1423 656 : return true;
1424 : }
1425 :
1426 : /*
1427 : * AbsorbSyncRequests
1428 : * Retrieve queued sync requests and pass them to sync mechanism.
1429 : *
1430 : * This is exported because it must be called during CreateCheckPoint;
1431 : * we have to be sure we have accepted all pending requests just before
1432 : * we start fsync'ing. Since CreateCheckPoint sometimes runs in
1433 : * non-checkpointer processes, do nothing if not checkpointer.
1434 : */
1435 : void
1436 41064 : AbsorbSyncRequests(void)
1437 : {
1438 41064 : CheckpointerRequest *requests = NULL;
1439 : CheckpointerRequest *request;
1440 : int n,
1441 : i;
1442 : bool loop;
1443 :
1444 41064 : if (!AmCheckpointerProcess())
1445 1304 : return;
1446 :
1447 : do
1448 : {
1449 39760 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1450 :
1451 : /*---
1452 : * We try to avoid holding the lock for a long time by:
1453 : * 1. Copying the request array and processing the requests after
1454 : * releasing the lock;
1455 : * 2. Processing not the whole queue, but only batches of
1456 : * CKPT_REQ_BATCH_SIZE at once.
1457 : *
1458 : * Once we have cleared the requests from shared memory, we must
1459 : * PANIC if we then fail to absorb them (e.g., because our hashtable
1460 : * runs out of memory). This is because the system cannot run safely
1461 : * if we are unable to fsync what we have been told to fsync.
1462 : * Fortunately, the hashtable is so small that the problem is quite
1463 : * unlikely to arise in practice.
1464 : *
1465 : * Note: The maximum possible size of a ring buffer is
1466 : * MAX_CHECKPOINT_REQUESTS entries, which fit into a maximum palloc
1467 : * allocation size of 1Gb. Our maximum batch size,
1468 : * CKPT_REQ_BATCH_SIZE, is even smaller.
1469 : */
1470 39760 : n = Min(CheckpointerShmem->num_requests, CKPT_REQ_BATCH_SIZE);
1471 39760 : if (n > 0)
1472 : {
1473 19422 : if (!requests)
1474 19422 : requests = (CheckpointerRequest *) palloc(n * sizeof(CheckpointerRequest));
1475 :
1476 2337456 : for (i = 0; i < n; i++)
1477 : {
1478 2318034 : requests[i] = CheckpointerShmem->requests[CheckpointerShmem->head];
1479 2318034 : CheckpointerShmem->head = (CheckpointerShmem->head + 1) % CheckpointerShmem->max_requests;
1480 : }
1481 :
1482 19422 : CheckpointerShmem->num_requests -= n;
1483 :
1484 : }
1485 :
1486 39760 : START_CRIT_SECTION();
1487 :
1488 : /* Are there any requests in the queue? If so, keep going. */
1489 39760 : loop = CheckpointerShmem->num_requests != 0;
1490 :
1491 39760 : LWLockRelease(CheckpointerCommLock);
1492 :
1493 2357794 : for (request = requests; n > 0; request++, n--)
1494 2318034 : RememberSyncRequest(&request->ftag, request->type);
1495 :
1496 39760 : END_CRIT_SECTION();
1497 39760 : } while (loop);
1498 :
1499 39760 : if (requests)
1500 19422 : pfree(requests);
1501 : }
1502 :
1503 : /*
1504 : * Update any shared memory configurations based on config parameters
1505 : */
1506 : static void
1507 1276 : UpdateSharedMemoryConfig(void)
1508 : {
1509 : /* update global shmem state for sync rep */
1510 1276 : SyncRepUpdateSyncStandbysDefined();
1511 :
1512 : /*
1513 : * If full_page_writes has been changed by SIGHUP, we update it in shared
1514 : * memory and write an XLOG_FPW_CHANGE record.
1515 : */
1516 1276 : UpdateFullPageWrites();
1517 :
1518 1276 : elog(DEBUG2, "checkpointer updated shared memory configuration values");
1519 1276 : }
1520 :
1521 : /*
1522 : * FirstCallSinceLastCheckpoint allows a process to take an action once
1523 : * per checkpoint cycle by asynchronously checking for checkpoint completion.
1524 : */
1525 : bool
1526 29572 : FirstCallSinceLastCheckpoint(void)
1527 : {
1528 : static int ckpt_done = 0;
1529 : int new_done;
1530 29572 : bool FirstCall = false;
1531 :
1532 29572 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1533 29572 : new_done = CheckpointerShmem->ckpt_done;
1534 29572 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1535 :
1536 29572 : if (new_done != ckpt_done)
1537 1260 : FirstCall = true;
1538 :
1539 29572 : ckpt_done = new_done;
1540 :
1541 29572 : return FirstCall;
1542 : }
1543 :
1544 : /*
1545 : * Wake up the checkpointer process.
1546 : */
1547 : void
1548 1884 : WakeupCheckpointer(void)
1549 : {
1550 1884 : volatile PROC_HDR *procglobal = ProcGlobal;
1551 1884 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1552 :
1553 1884 : if (checkpointerProc != INVALID_PROC_NUMBER)
1554 1256 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1555 1884 : }
|
