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 : #include "utils/wait_event.h"
71 :
72 :
73 : /*----------
74 : * Shared memory area for communication between checkpointer and backends
75 : *
76 : * The ckpt counters allow backends to watch for completion of a checkpoint
77 : * request they send. Here's how it works:
78 : * * At start of a checkpoint, checkpointer reads (and clears) the request
79 : * flags and increments ckpt_started, while holding ckpt_lck.
80 : * * On completion of a checkpoint, checkpointer sets ckpt_done to
81 : * equal ckpt_started.
82 : * * On failure of a checkpoint, checkpointer increments ckpt_failed
83 : * and sets ckpt_done to equal ckpt_started.
84 : *
85 : * The algorithm for backends is:
86 : * 1. Record current values of ckpt_failed and ckpt_started, and
87 : * set request flags, while holding ckpt_lck.
88 : * 2. Send signal to request checkpoint.
89 : * 3. Sleep until ckpt_started changes. Now you know a checkpoint has
90 : * begun since you started this algorithm (although *not* that it was
91 : * specifically initiated by your signal), and that it is using your flags.
92 : * 4. Record new value of ckpt_started.
93 : * 5. Sleep until ckpt_done >= saved value of ckpt_started. (Use modulo
94 : * arithmetic here in case counters wrap around.) Now you know a
95 : * checkpoint has started and completed, but not whether it was
96 : * successful.
97 : * 6. If ckpt_failed is different from the originally saved value,
98 : * assume request failed; otherwise it was definitely successful.
99 : *
100 : * ckpt_flags holds the OR of the checkpoint request flags sent by all
101 : * requesting backends since the last checkpoint start. The flags are
102 : * chosen so that OR'ing is the correct way to combine multiple requests.
103 : *
104 : * The requests array holds fsync requests sent by backends and not yet
105 : * absorbed by the checkpointer.
106 : *
107 : * Unlike the checkpoint fields, requests related fields are protected by
108 : * CheckpointerCommLock.
109 : *----------
110 : */
111 : typedef struct
112 : {
113 : SyncRequestType type; /* request type */
114 : FileTag ftag; /* file identifier */
115 : } CheckpointerRequest;
116 :
117 : typedef struct
118 : {
119 : pid_t checkpointer_pid; /* PID (0 if not started) */
120 :
121 : slock_t ckpt_lck; /* protects all the ckpt_* fields */
122 :
123 : int ckpt_started; /* advances when checkpoint starts */
124 : int ckpt_done; /* advances when checkpoint done */
125 : int ckpt_failed; /* advances when checkpoint fails */
126 :
127 : int ckpt_flags; /* checkpoint flags, as defined in xlog.h */
128 :
129 : ConditionVariable start_cv; /* signaled when ckpt_started advances */
130 : ConditionVariable done_cv; /* signaled when ckpt_done advances */
131 :
132 : int num_requests; /* current # of requests */
133 : int max_requests; /* allocated array size */
134 :
135 : int head; /* Index of the first request in the ring
136 : * buffer */
137 : int tail; /* Index of the last request in the ring
138 : * buffer */
139 :
140 : /* The ring buffer of pending checkpointer requests */
141 : CheckpointerRequest requests[FLEXIBLE_ARRAY_MEMBER];
142 : } CheckpointerShmemStruct;
143 :
144 : static CheckpointerShmemStruct *CheckpointerShmem;
145 :
146 : /* interval for calling AbsorbSyncRequests in CheckpointWriteDelay */
147 : #define WRITES_PER_ABSORB 1000
148 :
149 : /* Maximum number of checkpointer requests to process in one batch */
150 : #define CKPT_REQ_BATCH_SIZE 10000
151 :
152 : /* Max number of requests the checkpointer request queue can hold */
153 : #define MAX_CHECKPOINT_REQUESTS 10000000
154 :
155 : /*
156 : * GUC parameters
157 : */
158 : int CheckPointTimeout = 300;
159 : int CheckPointWarning = 30;
160 : double CheckPointCompletionTarget = 0.9;
161 :
162 : /*
163 : * Private state
164 : */
165 : static bool ckpt_active = false;
166 : static volatile sig_atomic_t ShutdownXLOGPending = false;
167 :
168 : /* these values are valid when ckpt_active is true: */
169 : static pg_time_t ckpt_start_time;
170 : static XLogRecPtr ckpt_start_recptr;
171 : static double ckpt_cached_elapsed;
172 :
173 : static pg_time_t last_checkpoint_time;
174 : static pg_time_t last_xlog_switch_time;
175 :
176 : /* Prototypes for private functions */
177 :
178 : static void ProcessCheckpointerInterrupts(void);
179 : static void CheckArchiveTimeout(void);
180 : static bool IsCheckpointOnSchedule(double progress);
181 : static bool FastCheckpointRequested(void);
182 : static bool CompactCheckpointerRequestQueue(void);
183 : static void UpdateSharedMemoryConfig(void);
184 :
185 : /* Signal handlers */
186 : static void ReqShutdownXLOG(SIGNAL_ARGS);
187 :
188 :
189 : /*
190 : * Main entry point for checkpointer process
191 : *
192 : * This is invoked from AuxiliaryProcessMain, which has already created the
193 : * basic execution environment, but not enabled signals yet.
194 : */
195 : void
196 607 : CheckpointerMain(const void *startup_data, size_t startup_data_len)
197 : {
198 : sigjmp_buf local_sigjmp_buf;
199 : MemoryContext checkpointer_context;
200 :
201 : Assert(startup_data_len == 0);
202 :
203 607 : AuxiliaryProcessMainCommon();
204 :
205 607 : CheckpointerShmem->checkpointer_pid = MyProcPid;
206 :
207 : /*
208 : * Properly accept or ignore signals the postmaster might send us
209 : *
210 : * Note: we deliberately ignore SIGTERM, because during a standard Unix
211 : * system shutdown cycle, init will SIGTERM all processes at once. We
212 : * want to wait for the backends to exit, whereupon the postmaster will
213 : * tell us it's okay to shut down (via SIGUSR2).
214 : */
215 607 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
216 607 : pqsignal(SIGINT, ReqShutdownXLOG);
217 607 : pqsignal(SIGTERM, SIG_IGN); /* ignore SIGTERM */
218 : /* SIGQUIT handler was already set up by InitPostmasterChild */
219 607 : pqsignal(SIGALRM, SIG_IGN);
220 607 : pqsignal(SIGPIPE, SIG_IGN);
221 607 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
222 607 : pqsignal(SIGUSR2, SignalHandlerForShutdownRequest);
223 :
224 : /*
225 : * Reset some signals that are accepted by postmaster but not here
226 : */
227 607 : pqsignal(SIGCHLD, SIG_DFL);
228 :
229 : /*
230 : * Initialize so that first time-driven event happens at the correct time.
231 : */
232 607 : last_checkpoint_time = last_xlog_switch_time = (pg_time_t) time(NULL);
233 :
234 : /*
235 : * Write out stats after shutdown. This needs to be called by exactly one
236 : * process during a normal shutdown, and since checkpointer is shut down
237 : * very late...
238 : *
239 : * While e.g. walsenders are active after the shutdown checkpoint has been
240 : * written (and thus could produce more stats), checkpointer stays around
241 : * after the shutdown checkpoint has been written. postmaster will only
242 : * signal checkpointer to exit after all processes that could emit stats
243 : * have been shut down.
244 : */
245 607 : before_shmem_exit(pgstat_before_server_shutdown, 0);
246 :
247 : /*
248 : * Create a memory context that we will do all our work in. We do this so
249 : * that we can reset the context during error recovery and thereby avoid
250 : * possible memory leaks. Formerly this code just ran in
251 : * TopMemoryContext, but resetting that would be a really bad idea.
252 : */
253 607 : checkpointer_context = AllocSetContextCreate(TopMemoryContext,
254 : "Checkpointer",
255 : ALLOCSET_DEFAULT_SIZES);
256 607 : MemoryContextSwitchTo(checkpointer_context);
257 :
258 : /*
259 : * If an exception is encountered, processing resumes here.
260 : *
261 : * You might wonder why this isn't coded as an infinite loop around a
262 : * PG_TRY construct. The reason is that this is the bottom of the
263 : * exception stack, and so with PG_TRY there would be no exception handler
264 : * in force at all during the CATCH part. By leaving the outermost setjmp
265 : * always active, we have at least some chance of recovering from an error
266 : * during error recovery. (If we get into an infinite loop thereby, it
267 : * will soon be stopped by overflow of elog.c's internal state stack.)
268 : *
269 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
270 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
271 : * signals other than SIGQUIT will be blocked until we complete error
272 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
273 : * call redundant, but it is not since InterruptPending might be set
274 : * already.
275 : */
276 607 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
277 : {
278 : /* Since not using PG_TRY, must reset error stack by hand */
279 0 : error_context_stack = NULL;
280 :
281 : /* Prevent interrupts while cleaning up */
282 0 : HOLD_INTERRUPTS();
283 :
284 : /* Report the error to the server log */
285 0 : EmitErrorReport();
286 :
287 : /*
288 : * These operations are really just a minimal subset of
289 : * AbortTransaction(). We don't have very many resources to worry
290 : * about in checkpointer, but we do have LWLocks, buffers, and temp
291 : * files.
292 : */
293 0 : LWLockReleaseAll();
294 0 : ConditionVariableCancelSleep();
295 0 : pgstat_report_wait_end();
296 0 : pgaio_error_cleanup();
297 0 : UnlockBuffers();
298 0 : ReleaseAuxProcessResources(false);
299 0 : AtEOXact_Buffers(false);
300 0 : AtEOXact_SMgr();
301 0 : AtEOXact_Files(false);
302 0 : AtEOXact_HashTables(false);
303 :
304 : /* Warn any waiting backends that the checkpoint failed. */
305 0 : if (ckpt_active)
306 : {
307 0 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
308 0 : CheckpointerShmem->ckpt_failed++;
309 0 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
310 0 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
311 :
312 0 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
313 :
314 0 : ckpt_active = false;
315 : }
316 :
317 : /*
318 : * Now return to normal top-level context and clear ErrorContext for
319 : * next time.
320 : */
321 0 : MemoryContextSwitchTo(checkpointer_context);
322 0 : FlushErrorState();
323 :
324 : /* Flush any leaked data in the top-level context */
325 0 : MemoryContextReset(checkpointer_context);
326 :
327 : /* Now we can allow interrupts again */
328 0 : RESUME_INTERRUPTS();
329 :
330 : /*
331 : * Sleep at least 1 second after any error. A write error is likely
332 : * to be repeated, and we don't want to be filling the error logs as
333 : * fast as we can.
334 : */
335 0 : pg_usleep(1000000L);
336 : }
337 :
338 : /* We can now handle ereport(ERROR) */
339 607 : PG_exception_stack = &local_sigjmp_buf;
340 :
341 : /*
342 : * Unblock signals (they were blocked when the postmaster forked us)
343 : */
344 607 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
345 :
346 : /*
347 : * Ensure all shared memory values are set correctly for the config. Doing
348 : * this here ensures no race conditions from other concurrent updaters.
349 : */
350 607 : UpdateSharedMemoryConfig();
351 :
352 : /*
353 : * Advertise our proc number that backends can use to wake us up while
354 : * we're sleeping.
355 : */
356 607 : ProcGlobal->checkpointerProc = MyProcNumber;
357 :
358 : /*
359 : * Loop until we've been asked to write the shutdown checkpoint or
360 : * terminate.
361 : */
362 : for (;;)
363 4493 : {
364 5100 : bool do_checkpoint = false;
365 5100 : int flags = 0;
366 : pg_time_t now;
367 : int elapsed_secs;
368 : int cur_timeout;
369 5100 : bool chkpt_or_rstpt_requested = false;
370 5100 : bool chkpt_or_rstpt_timed = false;
371 :
372 : /* Clear any already-pending wakeups */
373 5100 : ResetLatch(MyLatch);
374 :
375 : /*
376 : * Process any requests or signals received recently.
377 : */
378 5100 : AbsorbSyncRequests();
379 :
380 5100 : ProcessCheckpointerInterrupts();
381 5100 : if (ShutdownXLOGPending || ShutdownRequestPending)
382 : break;
383 :
384 : /*
385 : * Detect a pending checkpoint request by checking whether the flags
386 : * word in shared memory is nonzero. We shouldn't need to acquire the
387 : * ckpt_lck for this.
388 : */
389 4507 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
390 : {
391 1316 : do_checkpoint = true;
392 1316 : chkpt_or_rstpt_requested = true;
393 : }
394 :
395 : /*
396 : * Force a checkpoint if too much time has elapsed since the last one.
397 : * Note that we count a timed checkpoint in stats only when this
398 : * occurs without an external request, but we set the CAUSE_TIME flag
399 : * bit even if there is also an external request.
400 : */
401 4507 : now = (pg_time_t) time(NULL);
402 4507 : elapsed_secs = now - last_checkpoint_time;
403 4507 : if (elapsed_secs >= CheckPointTimeout)
404 : {
405 1 : if (!do_checkpoint)
406 1 : chkpt_or_rstpt_timed = true;
407 1 : do_checkpoint = true;
408 1 : flags |= CHECKPOINT_CAUSE_TIME;
409 : }
410 :
411 : /*
412 : * Do a checkpoint if requested.
413 : */
414 4507 : if (do_checkpoint)
415 : {
416 1317 : bool ckpt_performed = false;
417 : bool do_restartpoint;
418 :
419 : /* Check if we should perform a checkpoint or a restartpoint. */
420 1317 : do_restartpoint = RecoveryInProgress();
421 :
422 : /*
423 : * Atomically fetch the request flags to figure out what kind of a
424 : * checkpoint we should perform, and increase the started-counter
425 : * to acknowledge that we've started a new checkpoint.
426 : */
427 1317 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
428 1317 : flags |= CheckpointerShmem->ckpt_flags;
429 1317 : CheckpointerShmem->ckpt_flags = 0;
430 1317 : CheckpointerShmem->ckpt_started++;
431 1317 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
432 :
433 1317 : ConditionVariableBroadcast(&CheckpointerShmem->start_cv);
434 :
435 : /*
436 : * The end-of-recovery checkpoint is a real checkpoint that's
437 : * performed while we're still in recovery.
438 : */
439 1317 : if (flags & CHECKPOINT_END_OF_RECOVERY)
440 21 : do_restartpoint = false;
441 :
442 1317 : if (chkpt_or_rstpt_timed)
443 : {
444 1 : chkpt_or_rstpt_timed = false;
445 1 : if (do_restartpoint)
446 0 : PendingCheckpointerStats.restartpoints_timed++;
447 : else
448 1 : PendingCheckpointerStats.num_timed++;
449 : }
450 :
451 1317 : if (chkpt_or_rstpt_requested)
452 : {
453 1316 : chkpt_or_rstpt_requested = false;
454 1316 : if (do_restartpoint)
455 552 : PendingCheckpointerStats.restartpoints_requested++;
456 : else
457 764 : PendingCheckpointerStats.num_requested++;
458 : }
459 :
460 : /*
461 : * We will warn if (a) too soon since last checkpoint (whatever
462 : * caused it) and (b) somebody set the CHECKPOINT_CAUSE_XLOG flag
463 : * since the last checkpoint start. Note in particular that this
464 : * implementation will not generate warnings caused by
465 : * CheckPointTimeout < CheckPointWarning.
466 : */
467 1317 : if (!do_restartpoint &&
468 765 : (flags & CHECKPOINT_CAUSE_XLOG) &&
469 200 : elapsed_secs < CheckPointWarning)
470 200 : ereport(LOG,
471 : (errmsg_plural("checkpoints are occurring too frequently (%d second apart)",
472 : "checkpoints are occurring too frequently (%d seconds apart)",
473 : elapsed_secs,
474 : elapsed_secs),
475 : errhint("Consider increasing the configuration parameter \"%s\".", "max_wal_size")));
476 :
477 : /*
478 : * Initialize checkpointer-private variables used during
479 : * checkpoint.
480 : */
481 1317 : ckpt_active = true;
482 1317 : if (do_restartpoint)
483 552 : ckpt_start_recptr = GetXLogReplayRecPtr(NULL);
484 : else
485 765 : ckpt_start_recptr = GetInsertRecPtr();
486 1317 : ckpt_start_time = now;
487 1317 : ckpt_cached_elapsed = 0;
488 :
489 : /*
490 : * Do the checkpoint.
491 : */
492 1317 : if (!do_restartpoint)
493 765 : ckpt_performed = CreateCheckPoint(flags);
494 : else
495 552 : ckpt_performed = CreateRestartPoint(flags);
496 :
497 : /*
498 : * After any checkpoint, free all smgr objects. Otherwise we
499 : * would never do so for dropped relations, as the checkpointer
500 : * does not process shared invalidation messages or call
501 : * AtEOXact_SMgr().
502 : */
503 1317 : smgrdestroyall();
504 :
505 : /*
506 : * Indicate checkpoint completion to any waiting backends.
507 : */
508 1317 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
509 1317 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
510 1317 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
511 :
512 1317 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
513 :
514 1317 : if (!do_restartpoint)
515 : {
516 : /*
517 : * Note we record the checkpoint start time not end time as
518 : * last_checkpoint_time. This is so that time-driven
519 : * checkpoints happen at a predictable spacing.
520 : */
521 765 : last_checkpoint_time = now;
522 :
523 765 : if (ckpt_performed)
524 762 : PendingCheckpointerStats.num_performed++;
525 : }
526 : else
527 : {
528 552 : if (ckpt_performed)
529 : {
530 : /*
531 : * The same as for checkpoint. Please see the
532 : * corresponding comment.
533 : */
534 180 : last_checkpoint_time = now;
535 :
536 180 : PendingCheckpointerStats.restartpoints_performed++;
537 : }
538 : else
539 : {
540 : /*
541 : * We were not able to perform the restartpoint
542 : * (checkpoints throw an ERROR in case of error). Most
543 : * likely because we have not received any new checkpoint
544 : * WAL records since the last restartpoint. Try again in
545 : * 15 s.
546 : */
547 372 : last_checkpoint_time = now - CheckPointTimeout + 15;
548 : }
549 : }
550 :
551 1317 : ckpt_active = false;
552 :
553 : /*
554 : * We may have received an interrupt during the checkpoint and the
555 : * latch might have been reset (e.g. in CheckpointWriteDelay).
556 : */
557 1317 : ProcessCheckpointerInterrupts();
558 1317 : if (ShutdownXLOGPending || ShutdownRequestPending)
559 : break;
560 : }
561 :
562 : /*
563 : * Disable logical decoding if someone requested it. See comments atop
564 : * logicalctl.c.
565 : */
566 4500 : DisableLogicalDecodingIfNecessary();
567 :
568 : /* Check for archive_timeout and switch xlog files if necessary. */
569 4500 : CheckArchiveTimeout();
570 :
571 : /* Report pending statistics to the cumulative stats system */
572 4500 : pgstat_report_checkpointer();
573 4500 : pgstat_report_wal(true);
574 :
575 : /*
576 : * If any checkpoint flags have been set, redo the loop to handle the
577 : * checkpoint without sleeping.
578 : */
579 4500 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
580 351 : continue;
581 :
582 : /*
583 : * Sleep until we are signaled or it's time for another checkpoint or
584 : * xlog file switch.
585 : */
586 4149 : now = (pg_time_t) time(NULL);
587 4149 : elapsed_secs = now - last_checkpoint_time;
588 4149 : if (elapsed_secs >= CheckPointTimeout)
589 0 : continue; /* no sleep for us ... */
590 4149 : cur_timeout = CheckPointTimeout - elapsed_secs;
591 4149 : if (XLogArchiveTimeout > 0 && !RecoveryInProgress())
592 : {
593 0 : elapsed_secs = now - last_xlog_switch_time;
594 0 : if (elapsed_secs >= XLogArchiveTimeout)
595 0 : continue; /* no sleep for us ... */
596 0 : cur_timeout = Min(cur_timeout, XLogArchiveTimeout - elapsed_secs);
597 : }
598 :
599 4149 : (void) WaitLatch(MyLatch,
600 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
601 : cur_timeout * 1000L /* convert to ms */ ,
602 : WAIT_EVENT_CHECKPOINTER_MAIN);
603 : }
604 :
605 : /*
606 : * From here on, elog(ERROR) should end with exit(1), not send control
607 : * back to the sigsetjmp block above.
608 : */
609 600 : ExitOnAnyError = true;
610 :
611 600 : if (ShutdownXLOGPending)
612 : {
613 : /*
614 : * Close down the database.
615 : *
616 : * Since ShutdownXLOG() creates restartpoint or checkpoint, and
617 : * updates the statistics, increment the checkpoint request and flush
618 : * out pending statistic.
619 : */
620 600 : PendingCheckpointerStats.num_requested++;
621 600 : ShutdownXLOG(0, 0);
622 600 : pgstat_report_checkpointer();
623 600 : pgstat_report_wal(true);
624 :
625 : /*
626 : * Tell postmaster that we're done.
627 : */
628 600 : SendPostmasterSignal(PMSIGNAL_XLOG_IS_SHUTDOWN);
629 600 : ShutdownXLOGPending = false;
630 : }
631 :
632 : /*
633 : * Wait until we're asked to shut down. By separating the writing of the
634 : * shutdown checkpoint from checkpointer exiting, checkpointer can perform
635 : * some should-be-as-late-as-possible work like writing out stats.
636 : */
637 : for (;;)
638 : {
639 : /* Clear any already-pending wakeups */
640 1225 : ResetLatch(MyLatch);
641 :
642 1225 : ProcessCheckpointerInterrupts();
643 :
644 1225 : if (ShutdownRequestPending)
645 600 : break;
646 :
647 625 : (void) WaitLatch(MyLatch,
648 : WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
649 : 0,
650 : WAIT_EVENT_CHECKPOINTER_SHUTDOWN);
651 : }
652 :
653 : /* Normal exit from the checkpointer is here */
654 600 : proc_exit(0); /* done */
655 : }
656 :
657 : /*
658 : * Process any new interrupts.
659 : */
660 : static void
661 7642 : ProcessCheckpointerInterrupts(void)
662 : {
663 7642 : if (ProcSignalBarrierPending)
664 307 : ProcessProcSignalBarrier();
665 :
666 7642 : if (ConfigReloadPending)
667 : {
668 78 : ConfigReloadPending = false;
669 78 : ProcessConfigFile(PGC_SIGHUP);
670 :
671 : /*
672 : * Checkpointer is the last process to shut down, so we ask it to hold
673 : * the keys for a range of other tasks required most of which have
674 : * nothing to do with checkpointing at all.
675 : *
676 : * For various reasons, some config values can change dynamically so
677 : * the primary copy of them is held in shared memory to make sure all
678 : * backends see the same value. We make Checkpointer responsible for
679 : * updating the shared memory copy if the parameter setting changes
680 : * because of SIGHUP.
681 : */
682 78 : UpdateSharedMemoryConfig();
683 : }
684 :
685 : /* Perform logging of memory contexts of this process */
686 7642 : if (LogMemoryContextPending)
687 1 : ProcessLogMemoryContextInterrupt();
688 7642 : }
689 :
690 : /*
691 : * CheckArchiveTimeout -- check for archive_timeout and switch xlog files
692 : *
693 : * This will switch to a new WAL file and force an archive file write if
694 : * meaningful activity is recorded in the current WAL file. This includes most
695 : * writes, including just a single checkpoint record, but excludes WAL records
696 : * that were inserted with the XLOG_MARK_UNIMPORTANT flag being set (like
697 : * snapshots of running transactions). Such records, depending on
698 : * configuration, occur on regular intervals and don't contain important
699 : * information. This avoids generating archives with a few unimportant
700 : * records.
701 : */
702 : static void
703 11998 : CheckArchiveTimeout(void)
704 : {
705 : pg_time_t now;
706 : pg_time_t last_time;
707 : XLogRecPtr last_switch_lsn;
708 :
709 11998 : if (XLogArchiveTimeout <= 0 || RecoveryInProgress())
710 11998 : return;
711 :
712 0 : now = (pg_time_t) time(NULL);
713 :
714 : /* First we do a quick check using possibly-stale local state. */
715 0 : if ((int) (now - last_xlog_switch_time) < XLogArchiveTimeout)
716 0 : return;
717 :
718 : /*
719 : * Update local state ... note that last_xlog_switch_time is the last time
720 : * a switch was performed *or requested*.
721 : */
722 0 : last_time = GetLastSegSwitchData(&last_switch_lsn);
723 :
724 0 : last_xlog_switch_time = Max(last_xlog_switch_time, last_time);
725 :
726 : /* Now we can do the real checks */
727 0 : if ((int) (now - last_xlog_switch_time) >= XLogArchiveTimeout)
728 : {
729 : /*
730 : * Switch segment only when "important" WAL has been logged since the
731 : * last segment switch (last_switch_lsn points to end of segment
732 : * switch occurred in).
733 : */
734 0 : if (GetLastImportantRecPtr() > last_switch_lsn)
735 : {
736 : XLogRecPtr switchpoint;
737 :
738 : /* mark switch as unimportant, avoids triggering checkpoints */
739 0 : switchpoint = RequestXLogSwitch(true);
740 :
741 : /*
742 : * If the returned pointer points exactly to a segment boundary,
743 : * assume nothing happened.
744 : */
745 0 : if (XLogSegmentOffset(switchpoint, wal_segment_size) != 0)
746 0 : elog(DEBUG1, "write-ahead log switch forced (\"archive_timeout\"=%d)",
747 : XLogArchiveTimeout);
748 : }
749 :
750 : /*
751 : * Update state in any case, so we don't retry constantly when the
752 : * system is idle.
753 : */
754 0 : last_xlog_switch_time = now;
755 : }
756 : }
757 :
758 : /*
759 : * Returns true if a fast checkpoint request is pending. (Note that this does
760 : * not check the *current* checkpoint's FAST flag, but whether there is one
761 : * pending behind it.)
762 : */
763 : static bool
764 48974 : FastCheckpointRequested(void)
765 : {
766 48974 : volatile CheckpointerShmemStruct *cps = CheckpointerShmem;
767 :
768 : /*
769 : * We don't need to acquire the ckpt_lck in this case because we're only
770 : * looking at a single flag bit.
771 : */
772 48974 : if (cps->ckpt_flags & CHECKPOINT_FAST)
773 6700 : return true;
774 42274 : return false;
775 : }
776 :
777 : /*
778 : * CheckpointWriteDelay -- control rate of checkpoint
779 : *
780 : * This function is called after each page write performed by BufferSync().
781 : * It is responsible for throttling BufferSync()'s write rate to hit
782 : * checkpoint_completion_target.
783 : *
784 : * The checkpoint request flags should be passed in; currently the only one
785 : * examined is CHECKPOINT_FAST, which disables delays between writes.
786 : *
787 : * 'progress' is an estimate of how much of the work has been done, as a
788 : * fraction between 0.0 meaning none, and 1.0 meaning all done.
789 : */
790 : void
791 322542 : CheckpointWriteDelay(int flags, double progress)
792 : {
793 : static int absorb_counter = WRITES_PER_ABSORB;
794 :
795 : /* Do nothing if checkpoint is being executed by non-checkpointer process */
796 322542 : if (!AmCheckpointerProcess())
797 51929 : return;
798 :
799 : /*
800 : * Perform the usual duties and take a nap, unless we're behind schedule,
801 : * in which case we just try to catch up as quickly as possible.
802 : */
803 270613 : if (!(flags & CHECKPOINT_FAST) &&
804 49240 : !ShutdownXLOGPending &&
805 48974 : !ShutdownRequestPending &&
806 91248 : !FastCheckpointRequested() &&
807 42274 : IsCheckpointOnSchedule(progress))
808 : {
809 7498 : if (ConfigReloadPending)
810 : {
811 0 : ConfigReloadPending = false;
812 0 : ProcessConfigFile(PGC_SIGHUP);
813 : /* update shmem copies of config variables */
814 0 : UpdateSharedMemoryConfig();
815 : }
816 :
817 7498 : AbsorbSyncRequests();
818 7498 : absorb_counter = WRITES_PER_ABSORB;
819 :
820 7498 : CheckArchiveTimeout();
821 :
822 : /* Report interim statistics to the cumulative stats system */
823 7498 : pgstat_report_checkpointer();
824 :
825 : /*
826 : * This sleep used to be connected to bgwriter_delay, typically 200ms.
827 : * That resulted in more frequent wakeups if not much work to do.
828 : * Checkpointer and bgwriter are no longer related so take the Big
829 : * Sleep.
830 : */
831 7498 : WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | WL_TIMEOUT,
832 : 100,
833 : WAIT_EVENT_CHECKPOINT_WRITE_DELAY);
834 7498 : ResetLatch(MyLatch);
835 : }
836 263115 : else if (--absorb_counter <= 0)
837 : {
838 : /*
839 : * Absorb pending fsync requests after each WRITES_PER_ABSORB write
840 : * operations even when we don't sleep, to prevent overflow of the
841 : * fsync request queue.
842 : */
843 134 : AbsorbSyncRequests();
844 134 : absorb_counter = WRITES_PER_ABSORB;
845 : }
846 :
847 : /* Check for barrier events. */
848 270613 : if (ProcSignalBarrierPending)
849 7 : ProcessProcSignalBarrier();
850 : }
851 :
852 : /*
853 : * IsCheckpointOnSchedule -- are we on schedule to finish this checkpoint
854 : * (or restartpoint) in time?
855 : *
856 : * Compares the current progress against the time/segments elapsed since last
857 : * checkpoint, and returns true if the progress we've made this far is greater
858 : * than the elapsed time/segments.
859 : */
860 : static bool
861 42274 : IsCheckpointOnSchedule(double progress)
862 : {
863 : XLogRecPtr recptr;
864 : struct timeval now;
865 : double elapsed_xlogs,
866 : elapsed_time;
867 :
868 : Assert(ckpt_active);
869 :
870 : /* Scale progress according to checkpoint_completion_target. */
871 42274 : progress *= CheckPointCompletionTarget;
872 :
873 : /*
874 : * Check against the cached value first. Only do the more expensive
875 : * calculations once we reach the target previously calculated. Since
876 : * neither time or WAL insert pointer moves backwards, a freshly
877 : * calculated value can only be greater than or equal to the cached value.
878 : */
879 42274 : if (progress < ckpt_cached_elapsed)
880 31107 : return false;
881 :
882 : /*
883 : * Check progress against WAL segments written and CheckPointSegments.
884 : *
885 : * We compare the current WAL insert location against the location
886 : * computed before calling CreateCheckPoint. The code in XLogInsert that
887 : * actually triggers a checkpoint when CheckPointSegments is exceeded
888 : * compares against RedoRecPtr, so this is not completely accurate.
889 : * However, it's good enough for our purposes, we're only calculating an
890 : * estimate anyway.
891 : *
892 : * During recovery, we compare last replayed WAL record's location with
893 : * the location computed before calling CreateRestartPoint. That maintains
894 : * the same pacing as we have during checkpoints in normal operation, but
895 : * we might exceed max_wal_size by a fair amount. That's because there can
896 : * be a large gap between a checkpoint's redo-pointer and the checkpoint
897 : * record itself, and we only start the restartpoint after we've seen the
898 : * checkpoint record. (The gap is typically up to CheckPointSegments *
899 : * checkpoint_completion_target where checkpoint_completion_target is the
900 : * value that was in effect when the WAL was generated).
901 : */
902 11167 : if (RecoveryInProgress())
903 5229 : recptr = GetXLogReplayRecPtr(NULL);
904 : else
905 5938 : recptr = GetInsertRecPtr();
906 11167 : elapsed_xlogs = (((double) (recptr - ckpt_start_recptr)) /
907 11167 : wal_segment_size) / CheckPointSegments;
908 :
909 11167 : if (progress < elapsed_xlogs)
910 : {
911 3666 : ckpt_cached_elapsed = elapsed_xlogs;
912 3666 : return false;
913 : }
914 :
915 : /*
916 : * Check progress against time elapsed and checkpoint_timeout.
917 : */
918 7501 : gettimeofday(&now, NULL);
919 7501 : elapsed_time = ((double) ((pg_time_t) now.tv_sec - ckpt_start_time) +
920 7501 : now.tv_usec / 1000000.0) / CheckPointTimeout;
921 :
922 7501 : if (progress < elapsed_time)
923 : {
924 3 : ckpt_cached_elapsed = elapsed_time;
925 3 : return false;
926 : }
927 :
928 : /* It looks like we're on schedule. */
929 7498 : return true;
930 : }
931 :
932 :
933 : /* --------------------------------
934 : * signal handler routines
935 : * --------------------------------
936 : */
937 :
938 : /* SIGINT: set flag to trigger writing of shutdown checkpoint */
939 : static void
940 661 : ReqShutdownXLOG(SIGNAL_ARGS)
941 : {
942 661 : ShutdownXLOGPending = true;
943 661 : SetLatch(MyLatch);
944 661 : }
945 :
946 :
947 : /* --------------------------------
948 : * communication with backends
949 : * --------------------------------
950 : */
951 :
952 : /*
953 : * CheckpointerShmemSize
954 : * Compute space needed for checkpointer-related shared memory
955 : */
956 : Size
957 3387 : CheckpointerShmemSize(void)
958 : {
959 : Size size;
960 :
961 : /*
962 : * The size of the requests[] array is arbitrarily set equal to NBuffers.
963 : * But there is a cap of MAX_CHECKPOINT_REQUESTS to prevent accumulating
964 : * too many checkpoint requests in the ring buffer.
965 : */
966 3387 : size = offsetof(CheckpointerShmemStruct, requests);
967 3387 : size = add_size(size, mul_size(Min(NBuffers,
968 : MAX_CHECKPOINT_REQUESTS),
969 : sizeof(CheckpointerRequest)));
970 :
971 3387 : return size;
972 : }
973 :
974 : /*
975 : * CheckpointerShmemInit
976 : * Allocate and initialize checkpointer-related shared memory
977 : */
978 : void
979 1180 : CheckpointerShmemInit(void)
980 : {
981 1180 : Size size = CheckpointerShmemSize();
982 : bool found;
983 :
984 1180 : CheckpointerShmem = (CheckpointerShmemStruct *)
985 1180 : ShmemInitStruct("Checkpointer Data",
986 : size,
987 : &found);
988 :
989 1180 : if (!found)
990 : {
991 : /*
992 : * First time through, so initialize. Note that we zero the whole
993 : * requests array; this is so that CompactCheckpointerRequestQueue can
994 : * assume that any pad bytes in the request structs are zeroes.
995 : */
996 1324 : MemSet(CheckpointerShmem, 0, size);
997 1180 : SpinLockInit(&CheckpointerShmem->ckpt_lck);
998 1180 : CheckpointerShmem->max_requests = Min(NBuffers, MAX_CHECKPOINT_REQUESTS);
999 1180 : CheckpointerShmem->head = CheckpointerShmem->tail = 0;
1000 1180 : ConditionVariableInit(&CheckpointerShmem->start_cv);
1001 1180 : ConditionVariableInit(&CheckpointerShmem->done_cv);
1002 : }
1003 1180 : }
1004 :
1005 : /*
1006 : * ExecCheckpoint
1007 : * Primary entry point for manual CHECKPOINT commands
1008 : *
1009 : * This is mainly a wrapper for RequestCheckpoint().
1010 : */
1011 : void
1012 463 : ExecCheckpoint(ParseState *pstate, CheckPointStmt *stmt)
1013 : {
1014 463 : bool fast = true;
1015 463 : bool unlogged = false;
1016 :
1017 930 : foreach_ptr(DefElem, opt, stmt->options)
1018 : {
1019 20 : if (strcmp(opt->defname, "mode") == 0)
1020 : {
1021 8 : char *mode = defGetString(opt);
1022 :
1023 8 : if (strcmp(mode, "spread") == 0)
1024 0 : fast = false;
1025 8 : else if (strcmp(mode, "fast") != 0)
1026 4 : ereport(ERROR,
1027 : (errcode(ERRCODE_SYNTAX_ERROR),
1028 : errmsg("unrecognized value for %s option \"%s\": \"%s\"",
1029 : "CHECKPOINT", "mode", mode),
1030 : parser_errposition(pstate, opt->location)));
1031 : }
1032 12 : else if (strcmp(opt->defname, "flush_unlogged") == 0)
1033 8 : unlogged = defGetBoolean(opt);
1034 : else
1035 4 : ereport(ERROR,
1036 : (errcode(ERRCODE_SYNTAX_ERROR),
1037 : errmsg("unrecognized %s option \"%s\"",
1038 : "CHECKPOINT", opt->defname),
1039 : parser_errposition(pstate, opt->location)));
1040 : }
1041 :
1042 455 : if (!has_privs_of_role(GetUserId(), ROLE_PG_CHECKPOINT))
1043 0 : ereport(ERROR,
1044 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
1045 : /* translator: %s is name of an SQL command (e.g., CHECKPOINT) */
1046 : errmsg("permission denied to execute %s command",
1047 : "CHECKPOINT"),
1048 : errdetail("Only roles with privileges of the \"%s\" role may execute this command.",
1049 : "pg_checkpoint")));
1050 :
1051 910 : RequestCheckpoint(CHECKPOINT_WAIT |
1052 455 : (fast ? CHECKPOINT_FAST : 0) |
1053 455 : (unlogged ? CHECKPOINT_FLUSH_UNLOGGED : 0) |
1054 455 : (RecoveryInProgress() ? 0 : CHECKPOINT_FORCE));
1055 455 : }
1056 :
1057 : /*
1058 : * RequestCheckpoint
1059 : * Called in backend processes to request a checkpoint
1060 : *
1061 : * flags is a bitwise OR of the following:
1062 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
1063 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
1064 : * CHECKPOINT_FAST: finish the checkpoint ASAP,
1065 : * ignoring checkpoint_completion_target parameter.
1066 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
1067 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
1068 : * CHECKPOINT_END_OF_RECOVERY, and the CHECKPOINT command).
1069 : * CHECKPOINT_WAIT: wait for completion before returning (otherwise,
1070 : * just signal checkpointer to do it, and return).
1071 : * CHECKPOINT_CAUSE_XLOG: checkpoint is requested due to xlog filling.
1072 : * (This affects logging, and in particular enables CheckPointWarning.)
1073 : */
1074 : void
1075 2674 : RequestCheckpoint(int flags)
1076 : {
1077 : int ntries;
1078 : int old_failed,
1079 : old_started;
1080 :
1081 : /*
1082 : * If in a standalone backend, just do it ourselves.
1083 : */
1084 2674 : if (!IsPostmasterEnvironment)
1085 : {
1086 : /*
1087 : * There's no point in doing slow checkpoints in a standalone backend,
1088 : * because there's no other backends the checkpoint could disrupt.
1089 : */
1090 205 : CreateCheckPoint(flags | CHECKPOINT_FAST);
1091 :
1092 : /* Free all smgr objects, as CheckpointerMain() normally would. */
1093 205 : smgrdestroyall();
1094 :
1095 205 : return;
1096 : }
1097 :
1098 : /*
1099 : * Atomically set the request flags, and take a snapshot of the counters.
1100 : * When we see ckpt_started > old_started, we know the flags we set here
1101 : * have been seen by checkpointer.
1102 : *
1103 : * Note that we OR the flags with any existing flags, to avoid overriding
1104 : * a "stronger" request by another backend. The flag senses must be
1105 : * chosen to make this work!
1106 : */
1107 2469 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1108 :
1109 2469 : old_failed = CheckpointerShmem->ckpt_failed;
1110 2469 : old_started = CheckpointerShmem->ckpt_started;
1111 2469 : CheckpointerShmem->ckpt_flags |= (flags | CHECKPOINT_REQUESTED);
1112 :
1113 2469 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1114 :
1115 : /*
1116 : * Set checkpointer's latch to request checkpoint. It's possible that the
1117 : * checkpointer hasn't started yet, so we will retry a few times if
1118 : * needed. (Actually, more than a few times, since on slow or overloaded
1119 : * buildfarm machines, it's been observed that the checkpointer can take
1120 : * several seconds to start.) However, if not told to wait for the
1121 : * checkpoint to occur, we consider failure to set the latch to be
1122 : * nonfatal and merely LOG it. The checkpointer should see the request
1123 : * when it does start, with or without the SetLatch().
1124 : */
1125 : #define MAX_SIGNAL_TRIES 600 /* max wait 60.0 sec */
1126 2469 : for (ntries = 0;; ntries++)
1127 19 : {
1128 2488 : volatile PROC_HDR *procglobal = ProcGlobal;
1129 2488 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1130 :
1131 2488 : if (checkpointerProc == INVALID_PROC_NUMBER)
1132 : {
1133 19 : if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT))
1134 : {
1135 0 : elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG,
1136 : "could not notify checkpoint: checkpointer is not running");
1137 0 : break;
1138 : }
1139 : }
1140 : else
1141 : {
1142 2469 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1143 : /* notified successfully */
1144 2469 : break;
1145 : }
1146 :
1147 19 : CHECK_FOR_INTERRUPTS();
1148 19 : pg_usleep(100000L); /* wait 0.1 sec, then retry */
1149 : }
1150 :
1151 : /*
1152 : * If requested, wait for completion. We detect completion according to
1153 : * the algorithm given above.
1154 : */
1155 2469 : if (flags & CHECKPOINT_WAIT)
1156 : {
1157 : int new_started,
1158 : new_failed;
1159 :
1160 : /* Wait for a new checkpoint to start. */
1161 855 : ConditionVariablePrepareToSleep(&CheckpointerShmem->start_cv);
1162 : for (;;)
1163 : {
1164 1579 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1165 1579 : new_started = CheckpointerShmem->ckpt_started;
1166 1579 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1167 :
1168 1579 : if (new_started != old_started)
1169 855 : break;
1170 :
1171 724 : ConditionVariableSleep(&CheckpointerShmem->start_cv,
1172 : WAIT_EVENT_CHECKPOINT_START);
1173 : }
1174 855 : ConditionVariableCancelSleep();
1175 :
1176 : /*
1177 : * We are waiting for ckpt_done >= new_started, in a modulo sense.
1178 : */
1179 855 : ConditionVariablePrepareToSleep(&CheckpointerShmem->done_cv);
1180 : for (;;)
1181 652 : {
1182 : int new_done;
1183 :
1184 1507 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1185 1507 : new_done = CheckpointerShmem->ckpt_done;
1186 1507 : new_failed = CheckpointerShmem->ckpt_failed;
1187 1507 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1188 :
1189 1507 : if (new_done - new_started >= 0)
1190 855 : break;
1191 :
1192 652 : ConditionVariableSleep(&CheckpointerShmem->done_cv,
1193 : WAIT_EVENT_CHECKPOINT_DONE);
1194 : }
1195 855 : ConditionVariableCancelSleep();
1196 :
1197 855 : if (new_failed != old_failed)
1198 0 : ereport(ERROR,
1199 : (errmsg("checkpoint request failed"),
1200 : errhint("Consult recent messages in the server log for details.")));
1201 : }
1202 : }
1203 :
1204 : /*
1205 : * ForwardSyncRequest
1206 : * Forward a file-fsync request from a backend to the checkpointer
1207 : *
1208 : * Whenever a backend is compelled to write directly to a relation
1209 : * (which should be seldom, if the background writer is getting its job done),
1210 : * the backend calls this routine to pass over knowledge that the relation
1211 : * is dirty and must be fsync'd before next checkpoint. We also use this
1212 : * opportunity to count such writes for statistical purposes.
1213 : *
1214 : * To avoid holding the lock for longer than necessary, we normally write
1215 : * to the requests[] queue without checking for duplicates. The checkpointer
1216 : * will have to eliminate dups internally anyway. However, if we discover
1217 : * that the queue is full, we make a pass over the entire queue to compact
1218 : * it. This is somewhat expensive, but the alternative is for the backend
1219 : * to perform its own fsync, which is far more expensive in practice. It
1220 : * is theoretically possible a backend fsync might still be necessary, if
1221 : * the queue is full and contains no duplicate entries. In that case, we
1222 : * let the backend know by returning false.
1223 : */
1224 : bool
1225 1357651 : ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)
1226 : {
1227 : CheckpointerRequest *request;
1228 : bool too_full;
1229 : int insert_pos;
1230 :
1231 1357651 : if (!IsUnderPostmaster)
1232 0 : return false; /* probably shouldn't even get here */
1233 :
1234 1357651 : if (AmCheckpointerProcess())
1235 0 : elog(ERROR, "ForwardSyncRequest must not be called in checkpointer");
1236 :
1237 1357651 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1238 :
1239 : /*
1240 : * If the checkpointer isn't running or the request queue is full, the
1241 : * backend will have to perform its own fsync request. But before forcing
1242 : * that to happen, we can try to compact the request queue.
1243 : */
1244 1357651 : if (CheckpointerShmem->checkpointer_pid == 0 ||
1245 1357109 : (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&
1246 269 : !CompactCheckpointerRequestQueue()))
1247 : {
1248 601 : LWLockRelease(CheckpointerCommLock);
1249 601 : return false;
1250 : }
1251 :
1252 : /* OK, insert request */
1253 1357050 : insert_pos = CheckpointerShmem->tail;
1254 1357050 : request = &CheckpointerShmem->requests[insert_pos];
1255 1357050 : request->ftag = *ftag;
1256 1357050 : request->type = type;
1257 :
1258 1357050 : CheckpointerShmem->tail = (CheckpointerShmem->tail + 1) % CheckpointerShmem->max_requests;
1259 1357050 : CheckpointerShmem->num_requests++;
1260 :
1261 : /* If queue is more than half full, nudge the checkpointer to empty it */
1262 1357050 : too_full = (CheckpointerShmem->num_requests >=
1263 1357050 : CheckpointerShmem->max_requests / 2);
1264 :
1265 1357050 : LWLockRelease(CheckpointerCommLock);
1266 :
1267 : /* ... but not till after we release the lock */
1268 1357050 : if (too_full)
1269 : {
1270 37830 : volatile PROC_HDR *procglobal = ProcGlobal;
1271 37830 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1272 :
1273 37830 : if (checkpointerProc != INVALID_PROC_NUMBER)
1274 37830 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1275 : }
1276 :
1277 1357050 : return true;
1278 : }
1279 :
1280 : /*
1281 : * CompactCheckpointerRequestQueue
1282 : * Remove duplicates from the request queue to avoid backend fsyncs.
1283 : * Returns "true" if any entries were removed.
1284 : *
1285 : * Although a full fsync request queue is not common, it can lead to severe
1286 : * performance problems when it does happen. So far, this situation has
1287 : * only been observed to occur when the system is under heavy write load,
1288 : * and especially during the "sync" phase of a checkpoint. Without this
1289 : * logic, each backend begins doing an fsync for every block written, which
1290 : * gets very expensive and can slow down the whole system.
1291 : *
1292 : * Trying to do this every time the queue is full could lose if there
1293 : * aren't any removable entries. But that should be vanishingly rare in
1294 : * practice: there's one queue entry per shared buffer.
1295 : */
1296 : static bool
1297 269 : CompactCheckpointerRequestQueue(void)
1298 : {
1299 : struct CheckpointerSlotMapping
1300 : {
1301 : CheckpointerRequest request;
1302 : int ring_idx;
1303 : };
1304 :
1305 : int n;
1306 269 : int num_skipped = 0;
1307 : int head;
1308 : int max_requests;
1309 : int num_requests;
1310 : int read_idx,
1311 : write_idx;
1312 : HASHCTL ctl;
1313 : HTAB *htab;
1314 : bool *skip_slot;
1315 :
1316 : /* must hold CheckpointerCommLock in exclusive mode */
1317 : Assert(LWLockHeldByMe(CheckpointerCommLock));
1318 :
1319 : /* Avoid memory allocations in a critical section. */
1320 269 : if (CritSectionCount > 0)
1321 0 : return false;
1322 :
1323 269 : max_requests = CheckpointerShmem->max_requests;
1324 269 : num_requests = CheckpointerShmem->num_requests;
1325 :
1326 : /* Initialize skip_slot array */
1327 269 : skip_slot = palloc0_array(bool, max_requests);
1328 :
1329 269 : head = CheckpointerShmem->head;
1330 :
1331 : /* Initialize temporary hash table */
1332 269 : ctl.keysize = sizeof(CheckpointerRequest);
1333 269 : ctl.entrysize = sizeof(struct CheckpointerSlotMapping);
1334 269 : ctl.hcxt = CurrentMemoryContext;
1335 :
1336 269 : htab = hash_create("CompactCheckpointerRequestQueue",
1337 269 : CheckpointerShmem->num_requests,
1338 : &ctl,
1339 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1340 :
1341 : /*
1342 : * The basic idea here is that a request can be skipped if it's followed
1343 : * by a later, identical request. It might seem more sensible to work
1344 : * backwards from the end of the queue and check whether a request is
1345 : * *preceded* by an earlier, identical request, in the hopes of doing less
1346 : * copying. But that might change the semantics, if there's an
1347 : * intervening SYNC_FORGET_REQUEST or SYNC_FILTER_REQUEST, so we do it
1348 : * this way. It would be possible to be even smarter if we made the code
1349 : * below understand the specific semantics of such requests (it could blow
1350 : * away preceding entries that would end up being canceled anyhow), but
1351 : * it's not clear that the extra complexity would buy us anything.
1352 : */
1353 269 : read_idx = head;
1354 23725 : for (n = 0; n < num_requests; n++)
1355 : {
1356 : CheckpointerRequest *request;
1357 : struct CheckpointerSlotMapping *slotmap;
1358 : bool found;
1359 :
1360 : /*
1361 : * We use the request struct directly as a hashtable key. This
1362 : * assumes that any padding bytes in the structs are consistently the
1363 : * same, which should be okay because we zeroed them in
1364 : * CheckpointerShmemInit. Note also that RelFileLocator had better
1365 : * contain no pad bytes.
1366 : */
1367 23456 : request = &CheckpointerShmem->requests[read_idx];
1368 23456 : slotmap = hash_search(htab, request, HASH_ENTER, &found);
1369 23456 : if (found)
1370 : {
1371 : /* Duplicate, so mark the previous occurrence as skippable */
1372 11222 : skip_slot[slotmap->ring_idx] = true;
1373 11222 : num_skipped++;
1374 : }
1375 : /* Remember slot containing latest occurrence of this request value */
1376 23456 : slotmap->ring_idx = read_idx;
1377 :
1378 : /* Move to the next request in the ring buffer */
1379 23456 : read_idx = (read_idx + 1) % max_requests;
1380 : }
1381 :
1382 : /* Done with the hash table. */
1383 269 : hash_destroy(htab);
1384 :
1385 : /* If no duplicates, we're out of luck. */
1386 269 : if (!num_skipped)
1387 : {
1388 59 : pfree(skip_slot);
1389 59 : return false;
1390 : }
1391 :
1392 : /* We found some duplicates; remove them. */
1393 210 : read_idx = write_idx = head;
1394 19474 : for (n = 0; n < num_requests; n++)
1395 : {
1396 : /* If this slot is NOT skipped, keep it */
1397 19264 : if (!skip_slot[read_idx])
1398 : {
1399 : /* If the read and write positions are different, copy the request */
1400 8042 : if (write_idx != read_idx)
1401 6869 : CheckpointerShmem->requests[write_idx] =
1402 6869 : CheckpointerShmem->requests[read_idx];
1403 :
1404 : /* Advance the write position */
1405 8042 : write_idx = (write_idx + 1) % max_requests;
1406 : }
1407 :
1408 19264 : read_idx = (read_idx + 1) % max_requests;
1409 : }
1410 :
1411 : /*
1412 : * Update ring buffer state: head remains the same, tail moves, count
1413 : * decreases
1414 : */
1415 210 : CheckpointerShmem->tail = write_idx;
1416 210 : CheckpointerShmem->num_requests -= num_skipped;
1417 :
1418 210 : ereport(DEBUG1,
1419 : (errmsg_internal("compacted fsync request queue from %d entries to %d entries",
1420 : num_requests, CheckpointerShmem->num_requests)));
1421 :
1422 : /* Cleanup. */
1423 210 : pfree(skip_slot);
1424 210 : return true;
1425 : }
1426 :
1427 : /*
1428 : * AbsorbSyncRequests
1429 : * Retrieve queued sync requests and pass them to sync mechanism.
1430 : *
1431 : * This is exported because it must be called during CreateCheckPoint;
1432 : * we have to be sure we have accepted all pending requests just before
1433 : * we start fsync'ing. Since CreateCheckPoint sometimes runs in
1434 : * non-checkpointer processes, do nothing if not checkpointer.
1435 : */
1436 : void
1437 19627 : AbsorbSyncRequests(void)
1438 : {
1439 19627 : CheckpointerRequest *requests = NULL;
1440 : CheckpointerRequest *request;
1441 : int n,
1442 : i;
1443 : bool loop;
1444 :
1445 19627 : if (!AmCheckpointerProcess())
1446 652 : return;
1447 :
1448 : do
1449 : {
1450 18976 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1451 :
1452 : /*---
1453 : * We try to avoid holding the lock for a long time by:
1454 : * 1. Copying the request array and processing the requests after
1455 : * releasing the lock;
1456 : * 2. Processing not the whole queue, but only batches of
1457 : * CKPT_REQ_BATCH_SIZE at once.
1458 : *
1459 : * Once we have cleared the requests from shared memory, we must
1460 : * PANIC if we then fail to absorb them (e.g., because our hashtable
1461 : * runs out of memory). This is because the system cannot run safely
1462 : * if we are unable to fsync what we have been told to fsync.
1463 : * Fortunately, the hashtable is so small that the problem is quite
1464 : * unlikely to arise in practice.
1465 : *
1466 : * Note: The maximum possible size of a ring buffer is
1467 : * MAX_CHECKPOINT_REQUESTS entries, which fit into a maximum palloc
1468 : * allocation size of 1Gb. Our maximum batch size,
1469 : * CKPT_REQ_BATCH_SIZE, is even smaller.
1470 : */
1471 18976 : n = Min(CheckpointerShmem->num_requests, CKPT_REQ_BATCH_SIZE);
1472 18976 : if (n > 0)
1473 : {
1474 9517 : if (!requests)
1475 9516 : requests = (CheckpointerRequest *) palloc(n * sizeof(CheckpointerRequest));
1476 :
1477 1195416 : for (i = 0; i < n; i++)
1478 : {
1479 1185899 : requests[i] = CheckpointerShmem->requests[CheckpointerShmem->head];
1480 1185899 : CheckpointerShmem->head = (CheckpointerShmem->head + 1) % CheckpointerShmem->max_requests;
1481 : }
1482 :
1483 9517 : CheckpointerShmem->num_requests -= n;
1484 :
1485 : }
1486 :
1487 18976 : START_CRIT_SECTION();
1488 :
1489 : /* Are there any requests in the queue? If so, keep going. */
1490 18976 : loop = CheckpointerShmem->num_requests != 0;
1491 :
1492 18976 : LWLockRelease(CheckpointerCommLock);
1493 :
1494 1204875 : for (request = requests; n > 0; request++, n--)
1495 1185899 : RememberSyncRequest(&request->ftag, request->type);
1496 :
1497 18976 : END_CRIT_SECTION();
1498 18976 : } while (loop);
1499 :
1500 18975 : if (requests)
1501 9516 : pfree(requests);
1502 : }
1503 :
1504 : /*
1505 : * Update any shared memory configurations based on config parameters
1506 : */
1507 : static void
1508 685 : UpdateSharedMemoryConfig(void)
1509 : {
1510 : /* update global shmem state for sync rep */
1511 685 : SyncRepUpdateSyncStandbysDefined();
1512 :
1513 : /*
1514 : * If full_page_writes has been changed by SIGHUP, we update it in shared
1515 : * memory and write an XLOG_FPW_CHANGE record.
1516 : */
1517 685 : UpdateFullPageWrites();
1518 :
1519 685 : elog(DEBUG2, "checkpointer updated shared memory configuration values");
1520 685 : }
1521 :
1522 : /*
1523 : * FirstCallSinceLastCheckpoint allows a process to take an action once
1524 : * per checkpoint cycle by asynchronously checking for checkpoint completion.
1525 : */
1526 : bool
1527 14645 : FirstCallSinceLastCheckpoint(void)
1528 : {
1529 : static int ckpt_done = 0;
1530 : int new_done;
1531 14645 : bool FirstCall = false;
1532 :
1533 14645 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1534 14645 : new_done = CheckpointerShmem->ckpt_done;
1535 14645 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1536 :
1537 14645 : if (new_done != ckpt_done)
1538 590 : FirstCall = true;
1539 :
1540 14645 : ckpt_done = new_done;
1541 :
1542 14645 : return FirstCall;
1543 : }
1544 :
1545 : /*
1546 : * Wake up the checkpointer process.
1547 : */
1548 : void
1549 981 : WakeupCheckpointer(void)
1550 : {
1551 981 : volatile PROC_HDR *procglobal = ProcGlobal;
1552 981 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1553 :
1554 981 : if (checkpointerProc != INVALID_PROC_NUMBER)
1555 665 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1556 981 : }
|
