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 "storage/subsystems.h"
67 : #include "utils/acl.h"
68 : #include "utils/guc.h"
69 : #include "utils/memutils.h"
70 : #include "utils/resowner.h"
71 : #include "utils/wait_event.h"
72 :
73 :
74 : /*----------
75 : * Shared memory area for communication between checkpointer and backends
76 : *
77 : * The ckpt counters allow backends to watch for completion of a checkpoint
78 : * request they send. Here's how it works:
79 : * * At start of a checkpoint, checkpointer reads (and clears) the request
80 : * flags and increments ckpt_started, while holding ckpt_lck.
81 : * * On completion of a checkpoint, checkpointer sets ckpt_done to
82 : * equal ckpt_started.
83 : * * On failure of a checkpoint, checkpointer increments ckpt_failed
84 : * and sets ckpt_done to equal ckpt_started.
85 : *
86 : * The algorithm for backends is:
87 : * 1. Record current values of ckpt_failed and ckpt_started, and
88 : * set request flags, while holding ckpt_lck.
89 : * 2. Send signal to request checkpoint.
90 : * 3. Sleep until ckpt_started changes. Now you know a checkpoint has
91 : * begun since you started this algorithm (although *not* that it was
92 : * specifically initiated by your signal), and that it is using your flags.
93 : * 4. Record new value of ckpt_started.
94 : * 5. Sleep until ckpt_done >= saved value of ckpt_started. (Use modulo
95 : * arithmetic here in case counters wrap around.) Now you know a
96 : * checkpoint has started and completed, but not whether it was
97 : * successful.
98 : * 6. If ckpt_failed is different from the originally saved value,
99 : * assume request failed; otherwise it was definitely successful.
100 : *
101 : * ckpt_flags holds the OR of the checkpoint request flags sent by all
102 : * requesting backends since the last checkpoint start. The flags are
103 : * chosen so that OR'ing is the correct way to combine multiple requests.
104 : *
105 : * The requests array holds fsync requests sent by backends and not yet
106 : * absorbed by the checkpointer.
107 : *
108 : * Unlike the checkpoint fields, requests related fields are protected by
109 : * CheckpointerCommLock.
110 : *----------
111 : */
112 : typedef struct
113 : {
114 : SyncRequestType type; /* request type */
115 : FileTag ftag; /* file identifier */
116 : } CheckpointerRequest;
117 :
118 : typedef struct
119 : {
120 : pid_t checkpointer_pid; /* PID (0 if not started) */
121 :
122 : slock_t ckpt_lck; /* protects all the ckpt_* fields */
123 :
124 : int ckpt_started; /* advances when checkpoint starts */
125 : int ckpt_done; /* advances when checkpoint done */
126 : int ckpt_failed; /* advances when checkpoint fails */
127 :
128 : int ckpt_flags; /* checkpoint flags, as defined in xlog.h */
129 :
130 : ConditionVariable start_cv; /* signaled when ckpt_started advances */
131 : ConditionVariable done_cv; /* signaled when ckpt_done advances */
132 :
133 : int num_requests; /* current # of requests */
134 : int max_requests; /* allocated array size */
135 :
136 : int head; /* Index of the first request in the ring
137 : * buffer */
138 : int tail; /* Index of the last request in the ring
139 : * buffer */
140 :
141 : /* The ring buffer of pending checkpointer requests */
142 : CheckpointerRequest requests[FLEXIBLE_ARRAY_MEMBER];
143 : } CheckpointerShmemStruct;
144 :
145 : static CheckpointerShmemStruct *CheckpointerShmem;
146 :
147 : static void CheckpointerShmemRequest(void *arg);
148 : static void CheckpointerShmemInit(void *arg);
149 :
150 : const ShmemCallbacks CheckpointerShmemCallbacks = {
151 : .request_fn = CheckpointerShmemRequest,
152 : .init_fn = CheckpointerShmemInit,
153 : };
154 :
155 : /* interval for calling AbsorbSyncRequests in CheckpointWriteDelay */
156 : #define WRITES_PER_ABSORB 1000
157 :
158 : /* Maximum number of checkpointer requests to process in one batch */
159 : #define CKPT_REQ_BATCH_SIZE 10000
160 :
161 : /* Max number of requests the checkpointer request queue can hold */
162 : #define MAX_CHECKPOINT_REQUESTS 10000000
163 :
164 : /*
165 : * GUC parameters
166 : */
167 : int CheckPointTimeout = 300;
168 : int CheckPointWarning = 30;
169 : double CheckPointCompletionTarget = 0.9;
170 :
171 : /*
172 : * Private state
173 : */
174 : static bool ckpt_active = false;
175 : static volatile sig_atomic_t ShutdownXLOGPending = false;
176 :
177 : /* these values are valid when ckpt_active is true: */
178 : static pg_time_t ckpt_start_time;
179 : static XLogRecPtr ckpt_start_recptr;
180 : static double ckpt_cached_elapsed;
181 :
182 : static pg_time_t last_checkpoint_time;
183 : static pg_time_t last_xlog_switch_time;
184 :
185 : /* Prototypes for private functions */
186 :
187 : static void ProcessCheckpointerInterrupts(void);
188 : static void CheckArchiveTimeout(void);
189 : static bool IsCheckpointOnSchedule(double progress);
190 : static bool FastCheckpointRequested(void);
191 : static bool CompactCheckpointerRequestQueue(void);
192 : static void UpdateSharedMemoryConfig(void);
193 :
194 : /* Signal handlers */
195 : static void ReqShutdownXLOG(SIGNAL_ARGS);
196 :
197 :
198 : /*
199 : * Main entry point for checkpointer process
200 : *
201 : * This is invoked from AuxiliaryProcessMain, which has already created the
202 : * basic execution environment, but not enabled signals yet.
203 : */
204 : void
205 633 : CheckpointerMain(const void *startup_data, size_t startup_data_len)
206 : {
207 : sigjmp_buf local_sigjmp_buf;
208 : MemoryContext checkpointer_context;
209 :
210 : Assert(startup_data_len == 0);
211 :
212 633 : AuxiliaryProcessMainCommon();
213 :
214 633 : CheckpointerShmem->checkpointer_pid = MyProcPid;
215 :
216 : /*
217 : * Properly accept or ignore signals the postmaster might send us
218 : *
219 : * Note: we deliberately ignore SIGTERM, because during a standard Unix
220 : * system shutdown cycle, init will SIGTERM all processes at once. We
221 : * want to wait for the backends to exit, whereupon the postmaster will
222 : * tell us it's okay to shut down (via SIGUSR2).
223 : */
224 633 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
225 633 : pqsignal(SIGINT, ReqShutdownXLOG);
226 633 : pqsignal(SIGTERM, SIG_IGN); /* ignore SIGTERM */
227 : /* SIGQUIT handler was already set up by InitPostmasterChild */
228 633 : pqsignal(SIGALRM, SIG_IGN);
229 633 : pqsignal(SIGPIPE, SIG_IGN);
230 633 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
231 633 : pqsignal(SIGUSR2, SignalHandlerForShutdownRequest);
232 :
233 : /*
234 : * Reset some signals that are accepted by postmaster but not here
235 : */
236 633 : pqsignal(SIGCHLD, SIG_DFL);
237 :
238 : /*
239 : * Initialize so that first time-driven event happens at the correct time.
240 : */
241 633 : last_checkpoint_time = last_xlog_switch_time = (pg_time_t) time(NULL);
242 :
243 : /*
244 : * Write out stats after shutdown. This needs to be called by exactly one
245 : * process during a normal shutdown, and since checkpointer is shut down
246 : * very late...
247 : *
248 : * While e.g. walsenders are active after the shutdown checkpoint has been
249 : * written (and thus could produce more stats), checkpointer stays around
250 : * after the shutdown checkpoint has been written. postmaster will only
251 : * signal checkpointer to exit after all processes that could emit stats
252 : * have been shut down.
253 : */
254 633 : before_shmem_exit(pgstat_before_server_shutdown, 0);
255 :
256 : /*
257 : * Create a memory context that we will do all our work in. We do this so
258 : * that we can reset the context during error recovery and thereby avoid
259 : * possible memory leaks. Formerly this code just ran in
260 : * TopMemoryContext, but resetting that would be a really bad idea.
261 : */
262 633 : checkpointer_context = AllocSetContextCreate(TopMemoryContext,
263 : "Checkpointer",
264 : ALLOCSET_DEFAULT_SIZES);
265 633 : MemoryContextSwitchTo(checkpointer_context);
266 :
267 : /*
268 : * If an exception is encountered, processing resumes here.
269 : *
270 : * You might wonder why this isn't coded as an infinite loop around a
271 : * PG_TRY construct. The reason is that this is the bottom of the
272 : * exception stack, and so with PG_TRY there would be no exception handler
273 : * in force at all during the CATCH part. By leaving the outermost setjmp
274 : * always active, we have at least some chance of recovering from an error
275 : * during error recovery. (If we get into an infinite loop thereby, it
276 : * will soon be stopped by overflow of elog.c's internal state stack.)
277 : *
278 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
279 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
280 : * signals other than SIGQUIT will be blocked until we complete error
281 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
282 : * call redundant, but it is not since InterruptPending might be set
283 : * already.
284 : */
285 633 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
286 : {
287 : /* Since not using PG_TRY, must reset error stack by hand */
288 0 : error_context_stack = NULL;
289 :
290 : /* Prevent interrupts while cleaning up */
291 0 : HOLD_INTERRUPTS();
292 :
293 : /* Report the error to the server log */
294 0 : EmitErrorReport();
295 :
296 : /*
297 : * These operations are really just a minimal subset of
298 : * AbortTransaction(). We don't have very many resources to worry
299 : * about in checkpointer, but we do have LWLocks, buffers, and temp
300 : * files.
301 : */
302 0 : LWLockReleaseAll();
303 0 : ConditionVariableCancelSleep();
304 0 : pgstat_report_wait_end();
305 0 : pgaio_error_cleanup();
306 0 : UnlockBuffers();
307 0 : ReleaseAuxProcessResources(false);
308 0 : AtEOXact_Buffers(false);
309 0 : AtEOXact_SMgr();
310 0 : AtEOXact_Files(false);
311 0 : AtEOXact_HashTables(false);
312 :
313 : /* Warn any waiting backends that the checkpoint failed. */
314 0 : if (ckpt_active)
315 : {
316 0 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
317 0 : CheckpointerShmem->ckpt_failed++;
318 0 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
319 0 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
320 :
321 0 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
322 :
323 0 : ckpt_active = false;
324 : }
325 :
326 : /*
327 : * Now return to normal top-level context and clear ErrorContext for
328 : * next time.
329 : */
330 0 : MemoryContextSwitchTo(checkpointer_context);
331 0 : FlushErrorState();
332 :
333 : /* Flush any leaked data in the top-level context */
334 0 : MemoryContextReset(checkpointer_context);
335 :
336 : /* Now we can allow interrupts again */
337 0 : RESUME_INTERRUPTS();
338 :
339 : /*
340 : * Sleep at least 1 second after any error. A write error is likely
341 : * to be repeated, and we don't want to be filling the error logs as
342 : * fast as we can.
343 : */
344 0 : pg_usleep(1000000L);
345 : }
346 :
347 : /* We can now handle ereport(ERROR) */
348 633 : PG_exception_stack = &local_sigjmp_buf;
349 :
350 : /*
351 : * Unblock signals (they were blocked when the postmaster forked us)
352 : */
353 633 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
354 :
355 : /*
356 : * Ensure all shared memory values are set correctly for the config. Doing
357 : * this here ensures no race conditions from other concurrent updaters.
358 : */
359 633 : UpdateSharedMemoryConfig();
360 :
361 : /*
362 : * Advertise our proc number that backends can use to wake us up while
363 : * we're sleeping.
364 : */
365 633 : ProcGlobal->checkpointerProc = MyProcNumber;
366 :
367 : /*
368 : * Loop until we've been asked to write the shutdown checkpoint or
369 : * terminate.
370 : */
371 : for (;;)
372 5473 : {
373 6106 : bool do_checkpoint = false;
374 6106 : int flags = 0;
375 : pg_time_t now;
376 : int elapsed_secs;
377 : int cur_timeout;
378 6106 : bool chkpt_or_rstpt_requested = false;
379 6106 : bool chkpt_or_rstpt_timed = false;
380 :
381 : /* Clear any already-pending wakeups */
382 6106 : ResetLatch(MyLatch);
383 :
384 : /*
385 : * Process any requests or signals received recently.
386 : */
387 6106 : AbsorbSyncRequests();
388 :
389 6106 : ProcessCheckpointerInterrupts();
390 6106 : if (ShutdownXLOGPending || ShutdownRequestPending)
391 : break;
392 :
393 : /*
394 : * Detect a pending checkpoint request by checking whether the flags
395 : * word in shared memory is nonzero. We shouldn't need to acquire the
396 : * ckpt_lck for this.
397 : */
398 5484 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
399 : {
400 1350 : do_checkpoint = true;
401 1350 : chkpt_or_rstpt_requested = true;
402 : }
403 :
404 : /*
405 : * Force a checkpoint if too much time has elapsed since the last one.
406 : * Note that we count a timed checkpoint in stats only when this
407 : * occurs without an external request, but we set the CAUSE_TIME flag
408 : * bit even if there is also an external request.
409 : */
410 5484 : now = (pg_time_t) time(NULL);
411 5484 : elapsed_secs = now - last_checkpoint_time;
412 5484 : if (elapsed_secs >= CheckPointTimeout)
413 : {
414 1 : if (!do_checkpoint)
415 1 : chkpt_or_rstpt_timed = true;
416 1 : do_checkpoint = true;
417 1 : flags |= CHECKPOINT_CAUSE_TIME;
418 : }
419 :
420 : /*
421 : * Do a checkpoint if requested.
422 : */
423 5484 : if (do_checkpoint)
424 : {
425 1351 : bool ckpt_performed = false;
426 : bool do_restartpoint;
427 :
428 : /* Check if we should perform a checkpoint or a restartpoint. */
429 1351 : do_restartpoint = RecoveryInProgress();
430 :
431 : /*
432 : * Atomically fetch the request flags to figure out what kind of a
433 : * checkpoint we should perform, and increase the started-counter
434 : * to acknowledge that we've started a new checkpoint.
435 : */
436 1351 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
437 1351 : flags |= CheckpointerShmem->ckpt_flags;
438 1351 : CheckpointerShmem->ckpt_flags = 0;
439 1351 : CheckpointerShmem->ckpt_started++;
440 1351 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
441 :
442 1351 : ConditionVariableBroadcast(&CheckpointerShmem->start_cv);
443 :
444 : /*
445 : * The end-of-recovery checkpoint is a real checkpoint that's
446 : * performed while we're still in recovery.
447 : */
448 1351 : if (flags & CHECKPOINT_END_OF_RECOVERY)
449 21 : do_restartpoint = false;
450 :
451 1351 : if (chkpt_or_rstpt_timed)
452 : {
453 1 : chkpt_or_rstpt_timed = false;
454 1 : if (do_restartpoint)
455 0 : PendingCheckpointerStats.restartpoints_timed++;
456 : else
457 1 : PendingCheckpointerStats.num_timed++;
458 : }
459 :
460 1351 : if (chkpt_or_rstpt_requested)
461 : {
462 1350 : chkpt_or_rstpt_requested = false;
463 1350 : if (do_restartpoint)
464 554 : PendingCheckpointerStats.restartpoints_requested++;
465 : else
466 796 : PendingCheckpointerStats.num_requested++;
467 : }
468 :
469 : /*
470 : * We will warn if (a) too soon since last checkpoint (whatever
471 : * caused it) and (b) somebody set the CHECKPOINT_CAUSE_XLOG flag
472 : * since the last checkpoint start. Note in particular that this
473 : * implementation will not generate warnings caused by
474 : * CheckPointTimeout < CheckPointWarning.
475 : */
476 1351 : if (!do_restartpoint &&
477 797 : (flags & CHECKPOINT_CAUSE_XLOG) &&
478 208 : elapsed_secs < CheckPointWarning)
479 208 : ereport(LOG,
480 : (errmsg_plural("checkpoints are occurring too frequently (%d second apart)",
481 : "checkpoints are occurring too frequently (%d seconds apart)",
482 : elapsed_secs,
483 : elapsed_secs),
484 : errhint("Consider increasing the configuration parameter \"%s\".", "max_wal_size")));
485 :
486 : /*
487 : * Initialize checkpointer-private variables used during
488 : * checkpoint.
489 : */
490 1351 : ckpt_active = true;
491 1351 : if (do_restartpoint)
492 554 : ckpt_start_recptr = GetXLogReplayRecPtr(NULL);
493 : else
494 797 : ckpt_start_recptr = GetInsertRecPtr();
495 1351 : ckpt_start_time = now;
496 1351 : ckpt_cached_elapsed = 0;
497 :
498 : /*
499 : * Do the checkpoint.
500 : */
501 1351 : if (!do_restartpoint)
502 797 : ckpt_performed = CreateCheckPoint(flags);
503 : else
504 554 : ckpt_performed = CreateRestartPoint(flags);
505 :
506 : /*
507 : * After any checkpoint, free all smgr objects. Otherwise we
508 : * would never do so for dropped relations, as the checkpointer
509 : * does not process shared invalidation messages or call
510 : * AtEOXact_SMgr().
511 : */
512 1351 : smgrdestroyall();
513 :
514 : /*
515 : * Indicate checkpoint completion to any waiting backends.
516 : */
517 1351 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
518 1351 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
519 1351 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
520 :
521 1351 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
522 :
523 1351 : if (!do_restartpoint)
524 : {
525 : /*
526 : * Note we record the checkpoint start time not end time as
527 : * last_checkpoint_time. This is so that time-driven
528 : * checkpoints happen at a predictable spacing.
529 : */
530 797 : last_checkpoint_time = now;
531 :
532 797 : if (ckpt_performed)
533 793 : PendingCheckpointerStats.num_performed++;
534 : }
535 : else
536 : {
537 554 : if (ckpt_performed)
538 : {
539 : /*
540 : * The same as for checkpoint. Please see the
541 : * corresponding comment.
542 : */
543 186 : last_checkpoint_time = now;
544 :
545 186 : PendingCheckpointerStats.restartpoints_performed++;
546 : }
547 : else
548 : {
549 : /*
550 : * We were not able to perform the restartpoint
551 : * (checkpoints throw an ERROR in case of error). Most
552 : * likely because we have not received any new checkpoint
553 : * WAL records since the last restartpoint. Try again in
554 : * 15 s.
555 : */
556 368 : last_checkpoint_time = now - CheckPointTimeout + 15;
557 : }
558 : }
559 :
560 1351 : ckpt_active = false;
561 :
562 : /*
563 : * We may have received an interrupt during the checkpoint and the
564 : * latch might have been reset (e.g. in CheckpointWriteDelay).
565 : */
566 1351 : ProcessCheckpointerInterrupts();
567 1351 : if (ShutdownXLOGPending || ShutdownRequestPending)
568 : break;
569 : }
570 :
571 : /*
572 : * Disable logical decoding if someone requested it. See comments atop
573 : * logicalctl.c.
574 : */
575 5480 : DisableLogicalDecodingIfNecessary();
576 :
577 : /* Check for archive_timeout and switch xlog files if necessary. */
578 5480 : CheckArchiveTimeout();
579 :
580 : /* Report pending statistics to the cumulative stats system */
581 5480 : pgstat_report_checkpointer();
582 5480 : pgstat_report_wal(true);
583 :
584 : /*
585 : * If any checkpoint flags have been set, redo the loop to handle the
586 : * checkpoint without sleeping.
587 : */
588 5480 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
589 297 : continue;
590 :
591 : /*
592 : * Sleep until we are signaled or it's time for another checkpoint or
593 : * xlog file switch.
594 : */
595 5183 : now = (pg_time_t) time(NULL);
596 5183 : elapsed_secs = now - last_checkpoint_time;
597 5183 : if (elapsed_secs >= CheckPointTimeout)
598 0 : continue; /* no sleep for us ... */
599 5183 : cur_timeout = CheckPointTimeout - elapsed_secs;
600 5183 : if (XLogArchiveTimeout > 0 && !RecoveryInProgress())
601 : {
602 0 : elapsed_secs = now - last_xlog_switch_time;
603 0 : if (elapsed_secs >= XLogArchiveTimeout)
604 0 : continue; /* no sleep for us ... */
605 0 : cur_timeout = Min(cur_timeout, XLogArchiveTimeout - elapsed_secs);
606 : }
607 :
608 5183 : (void) WaitLatch(MyLatch,
609 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
610 : cur_timeout * 1000L /* convert to ms */ ,
611 : WAIT_EVENT_CHECKPOINTER_MAIN);
612 : }
613 :
614 : /*
615 : * From here on, elog(ERROR) should end with exit(1), not send control
616 : * back to the sigsetjmp block above.
617 : */
618 626 : ExitOnAnyError = true;
619 :
620 626 : if (ShutdownXLOGPending)
621 : {
622 : /*
623 : * Close down the database.
624 : *
625 : * Since ShutdownXLOG() creates restartpoint or checkpoint, and
626 : * updates the statistics, increment the checkpoint request and flush
627 : * out pending statistic.
628 : */
629 626 : PendingCheckpointerStats.num_requested++;
630 626 : ShutdownXLOG(0, 0);
631 626 : pgstat_report_checkpointer();
632 626 : pgstat_report_wal(true);
633 :
634 : /*
635 : * Tell postmaster that we're done.
636 : */
637 626 : SendPostmasterSignal(PMSIGNAL_XLOG_IS_SHUTDOWN);
638 626 : ShutdownXLOGPending = false;
639 : }
640 :
641 : /*
642 : * Wait until we're asked to shut down. By separating the writing of the
643 : * shutdown checkpoint from checkpointer exiting, checkpointer can perform
644 : * some should-be-as-late-as-possible work like writing out stats.
645 : */
646 : for (;;)
647 : {
648 : /* Clear any already-pending wakeups */
649 1283 : ResetLatch(MyLatch);
650 :
651 1283 : ProcessCheckpointerInterrupts();
652 :
653 1283 : if (ShutdownRequestPending)
654 626 : break;
655 :
656 657 : (void) WaitLatch(MyLatch,
657 : WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
658 : 0,
659 : WAIT_EVENT_CHECKPOINTER_SHUTDOWN);
660 : }
661 :
662 : /* Normal exit from the checkpointer is here */
663 626 : proc_exit(0); /* done */
664 : }
665 :
666 : /*
667 : * Process any new interrupts.
668 : */
669 : static void
670 8740 : ProcessCheckpointerInterrupts(void)
671 : {
672 8740 : if (ProcSignalBarrierPending)
673 344 : ProcessProcSignalBarrier();
674 :
675 8740 : if (ConfigReloadPending)
676 : {
677 80 : ConfigReloadPending = false;
678 80 : ProcessConfigFile(PGC_SIGHUP);
679 :
680 : /*
681 : * Checkpointer is the last process to shut down, so we ask it to hold
682 : * the keys for a range of other tasks required most of which have
683 : * nothing to do with checkpointing at all.
684 : *
685 : * For various reasons, some config values can change dynamically so
686 : * the primary copy of them is held in shared memory to make sure all
687 : * backends see the same value. We make Checkpointer responsible for
688 : * updating the shared memory copy if the parameter setting changes
689 : * because of SIGHUP.
690 : */
691 80 : UpdateSharedMemoryConfig();
692 : }
693 :
694 : /* Perform logging of memory contexts of this process */
695 8740 : if (LogMemoryContextPending)
696 1 : ProcessLogMemoryContextInterrupt();
697 8740 : }
698 :
699 : /*
700 : * CheckArchiveTimeout -- check for archive_timeout and switch xlog files
701 : *
702 : * This will switch to a new WAL file and force an archive file write if
703 : * meaningful activity is recorded in the current WAL file. This includes most
704 : * writes, including just a single checkpoint record, but excludes WAL records
705 : * that were inserted with the XLOG_MARK_UNIMPORTANT flag being set (like
706 : * snapshots of running transactions). Such records, depending on
707 : * configuration, occur on regular intervals and don't contain important
708 : * information. This avoids generating archives with a few unimportant
709 : * records.
710 : */
711 : static void
712 13354 : CheckArchiveTimeout(void)
713 : {
714 : pg_time_t now;
715 : pg_time_t last_time;
716 : XLogRecPtr last_switch_lsn;
717 :
718 13354 : if (XLogArchiveTimeout <= 0 || RecoveryInProgress())
719 13354 : return;
720 :
721 0 : now = (pg_time_t) time(NULL);
722 :
723 : /* First we do a quick check using possibly-stale local state. */
724 0 : if ((int) (now - last_xlog_switch_time) < XLogArchiveTimeout)
725 0 : return;
726 :
727 : /*
728 : * Update local state ... note that last_xlog_switch_time is the last time
729 : * a switch was performed *or requested*.
730 : */
731 0 : last_time = GetLastSegSwitchData(&last_switch_lsn);
732 :
733 0 : last_xlog_switch_time = Max(last_xlog_switch_time, last_time);
734 :
735 : /* Now we can do the real checks */
736 0 : if ((int) (now - last_xlog_switch_time) >= XLogArchiveTimeout)
737 : {
738 : /*
739 : * Switch segment only when "important" WAL has been logged since the
740 : * last segment switch (last_switch_lsn points to end of segment
741 : * switch occurred in).
742 : */
743 0 : if (GetLastImportantRecPtr() > last_switch_lsn)
744 : {
745 : XLogRecPtr switchpoint;
746 :
747 : /* mark switch as unimportant, avoids triggering checkpoints */
748 0 : switchpoint = RequestXLogSwitch(true);
749 :
750 : /*
751 : * If the returned pointer points exactly to a segment boundary,
752 : * assume nothing happened.
753 : */
754 0 : if (XLogSegmentOffset(switchpoint, wal_segment_size) != 0)
755 0 : elog(DEBUG1, "write-ahead log switch forced (\"archive_timeout\"=%d)",
756 : XLogArchiveTimeout);
757 : }
758 :
759 : /*
760 : * Update state in any case, so we don't retry constantly when the
761 : * system is idle.
762 : */
763 0 : last_xlog_switch_time = now;
764 : }
765 : }
766 :
767 : /*
768 : * Returns true if a fast checkpoint request is pending. (Note that this does
769 : * not check the *current* checkpoint's FAST flag, but whether there is one
770 : * pending behind it.)
771 : */
772 : static bool
773 49805 : FastCheckpointRequested(void)
774 : {
775 49805 : volatile CheckpointerShmemStruct *cps = CheckpointerShmem;
776 :
777 : /*
778 : * We don't need to acquire the ckpt_lck in this case because we're only
779 : * looking at a single flag bit.
780 : */
781 49805 : if (cps->ckpt_flags & CHECKPOINT_FAST)
782 7146 : return true;
783 42659 : return false;
784 : }
785 :
786 : /*
787 : * CheckpointWriteDelay -- control rate of checkpoint
788 : *
789 : * This function is called after each page write performed by BufferSync().
790 : * It is responsible for throttling BufferSync()'s write rate to hit
791 : * checkpoint_completion_target.
792 : *
793 : * The checkpoint request flags should be passed in; currently the only one
794 : * examined is CHECKPOINT_FAST, which disables delays between writes.
795 : *
796 : * 'progress' is an estimate of how much of the work has been done, as a
797 : * fraction between 0.0 meaning none, and 1.0 meaning all done.
798 : */
799 : void
800 340638 : CheckpointWriteDelay(int flags, double progress)
801 : {
802 : static int absorb_counter = WRITES_PER_ABSORB;
803 :
804 : /* Do nothing if checkpoint is being executed by non-checkpointer process */
805 340638 : if (!AmCheckpointerProcess())
806 58193 : return;
807 :
808 : /*
809 : * Perform the usual duties and take a nap, unless we're behind schedule,
810 : * in which case we just try to catch up as quickly as possible.
811 : */
812 282445 : if (!(flags & CHECKPOINT_FAST) &&
813 49835 : !ShutdownXLOGPending &&
814 49805 : !ShutdownRequestPending &&
815 92464 : !FastCheckpointRequested() &&
816 42659 : IsCheckpointOnSchedule(progress))
817 : {
818 7874 : if (ConfigReloadPending)
819 : {
820 0 : ConfigReloadPending = false;
821 0 : ProcessConfigFile(PGC_SIGHUP);
822 : /* update shmem copies of config variables */
823 0 : UpdateSharedMemoryConfig();
824 : }
825 :
826 7874 : AbsorbSyncRequests();
827 7874 : absorb_counter = WRITES_PER_ABSORB;
828 :
829 7874 : CheckArchiveTimeout();
830 :
831 : /* Report interim statistics to the cumulative stats system */
832 7874 : pgstat_report_checkpointer();
833 :
834 : /*
835 : * This sleep used to be connected to bgwriter_delay, typically 200ms.
836 : * That resulted in more frequent wakeups if not much work to do.
837 : * Checkpointer and bgwriter are no longer related so take the Big
838 : * Sleep.
839 : */
840 7874 : WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | WL_TIMEOUT,
841 : 100,
842 : WAIT_EVENT_CHECKPOINT_WRITE_DELAY);
843 7874 : ResetLatch(MyLatch);
844 : }
845 274571 : else if (--absorb_counter <= 0)
846 : {
847 : /*
848 : * Absorb pending fsync requests after each WRITES_PER_ABSORB write
849 : * operations even when we don't sleep, to prevent overflow of the
850 : * fsync request queue.
851 : */
852 145 : AbsorbSyncRequests();
853 145 : absorb_counter = WRITES_PER_ABSORB;
854 : }
855 :
856 : /* Check for barrier events. */
857 282445 : if (ProcSignalBarrierPending)
858 14 : ProcessProcSignalBarrier();
859 : }
860 :
861 : /*
862 : * IsCheckpointOnSchedule -- are we on schedule to finish this checkpoint
863 : * (or restartpoint) in time?
864 : *
865 : * Compares the current progress against the time/segments elapsed since last
866 : * checkpoint, and returns true if the progress we've made this far is greater
867 : * than the elapsed time/segments.
868 : */
869 : static bool
870 42659 : IsCheckpointOnSchedule(double progress)
871 : {
872 : XLogRecPtr recptr;
873 : struct timeval now;
874 : double elapsed_xlogs,
875 : elapsed_time;
876 :
877 : Assert(ckpt_active);
878 :
879 : /* Scale progress according to checkpoint_completion_target. */
880 42659 : progress *= CheckPointCompletionTarget;
881 :
882 : /*
883 : * Check against the cached value first. Only do the more expensive
884 : * calculations once we reach the target previously calculated. Since
885 : * neither time or WAL insert pointer moves backwards, a freshly
886 : * calculated value can only be greater than or equal to the cached value.
887 : */
888 42659 : if (progress < ckpt_cached_elapsed)
889 30889 : return false;
890 :
891 : /*
892 : * Check progress against WAL segments written and CheckPointSegments.
893 : *
894 : * We compare the current WAL insert location against the location
895 : * computed before calling CreateCheckPoint. The code in XLogInsert that
896 : * actually triggers a checkpoint when CheckPointSegments is exceeded
897 : * compares against RedoRecPtr, so this is not completely accurate.
898 : * However, it's good enough for our purposes, we're only calculating an
899 : * estimate anyway.
900 : *
901 : * During recovery, we compare last replayed WAL record's location with
902 : * the location computed before calling CreateRestartPoint. That maintains
903 : * the same pacing as we have during checkpoints in normal operation, but
904 : * we might exceed max_wal_size by a fair amount. That's because there can
905 : * be a large gap between a checkpoint's redo-pointer and the checkpoint
906 : * record itself, and we only start the restartpoint after we've seen the
907 : * checkpoint record. (The gap is typically up to CheckPointSegments *
908 : * checkpoint_completion_target where checkpoint_completion_target is the
909 : * value that was in effect when the WAL was generated).
910 : */
911 11770 : if (RecoveryInProgress())
912 5363 : recptr = GetXLogReplayRecPtr(NULL);
913 : else
914 6407 : recptr = GetInsertRecPtr();
915 11770 : elapsed_xlogs = (((double) (recptr - ckpt_start_recptr)) /
916 11770 : wal_segment_size) / CheckPointSegments;
917 :
918 11770 : if (progress < elapsed_xlogs)
919 : {
920 3894 : ckpt_cached_elapsed = elapsed_xlogs;
921 3894 : return false;
922 : }
923 :
924 : /*
925 : * Check progress against time elapsed and checkpoint_timeout.
926 : */
927 7876 : gettimeofday(&now, NULL);
928 7876 : elapsed_time = ((double) ((pg_time_t) now.tv_sec - ckpt_start_time) +
929 7876 : now.tv_usec / 1000000.0) / CheckPointTimeout;
930 :
931 7876 : if (progress < elapsed_time)
932 : {
933 2 : ckpt_cached_elapsed = elapsed_time;
934 2 : return false;
935 : }
936 :
937 : /* It looks like we're on schedule. */
938 7874 : return true;
939 : }
940 :
941 :
942 : /* --------------------------------
943 : * signal handler routines
944 : * --------------------------------
945 : */
946 :
947 : /* SIGINT: set flag to trigger writing of shutdown checkpoint */
948 : static void
949 688 : ReqShutdownXLOG(SIGNAL_ARGS)
950 : {
951 688 : ShutdownXLOGPending = true;
952 688 : SetLatch(MyLatch);
953 688 : }
954 :
955 :
956 : /* --------------------------------
957 : * communication with backends
958 : * --------------------------------
959 : */
960 :
961 : /*
962 : * CheckpointerShmemRequest
963 : * Register shared memory space needed for checkpointer
964 : */
965 : static void
966 1234 : CheckpointerShmemRequest(void *arg)
967 : {
968 : Size size;
969 :
970 : /*
971 : * The size of the requests[] array is arbitrarily set equal to NBuffers.
972 : * But there is a cap of MAX_CHECKPOINT_REQUESTS to prevent accumulating
973 : * too many checkpoint requests in the ring buffer.
974 : */
975 1234 : size = offsetof(CheckpointerShmemStruct, requests);
976 1234 : size = add_size(size, mul_size(Min(NBuffers,
977 : MAX_CHECKPOINT_REQUESTS),
978 : sizeof(CheckpointerRequest)));
979 1234 : ShmemRequestStruct(.name = "Checkpointer Data",
980 : .size = size,
981 : .ptr = (void **) &CheckpointerShmem,
982 : );
983 1234 : }
984 :
985 : /*
986 : * CheckpointerShmemInit
987 : * Initialize checkpointer-related shared memory
988 : */
989 : static void
990 1231 : CheckpointerShmemInit(void *arg)
991 : {
992 1231 : SpinLockInit(&CheckpointerShmem->ckpt_lck);
993 1231 : CheckpointerShmem->max_requests = Min(NBuffers, MAX_CHECKPOINT_REQUESTS);
994 1231 : CheckpointerShmem->head = CheckpointerShmem->tail = 0;
995 1231 : ConditionVariableInit(&CheckpointerShmem->start_cv);
996 1231 : ConditionVariableInit(&CheckpointerShmem->done_cv);
997 1231 : }
998 :
999 : /*
1000 : * ExecCheckpoint
1001 : * Primary entry point for manual CHECKPOINT commands
1002 : *
1003 : * This is mainly a wrapper for RequestCheckpoint().
1004 : */
1005 : void
1006 465 : ExecCheckpoint(ParseState *pstate, CheckPointStmt *stmt)
1007 : {
1008 465 : bool fast = true;
1009 465 : bool unlogged = false;
1010 :
1011 934 : foreach_ptr(DefElem, opt, stmt->options)
1012 : {
1013 20 : if (strcmp(opt->defname, "mode") == 0)
1014 : {
1015 8 : char *mode = defGetString(opt);
1016 :
1017 8 : if (strcmp(mode, "spread") == 0)
1018 0 : fast = false;
1019 8 : else if (strcmp(mode, "fast") != 0)
1020 4 : ereport(ERROR,
1021 : (errcode(ERRCODE_SYNTAX_ERROR),
1022 : errmsg("unrecognized value for %s option \"%s\": \"%s\"",
1023 : "CHECKPOINT", "mode", mode),
1024 : parser_errposition(pstate, opt->location)));
1025 : }
1026 12 : else if (strcmp(opt->defname, "flush_unlogged") == 0)
1027 8 : unlogged = defGetBoolean(opt);
1028 : else
1029 4 : ereport(ERROR,
1030 : (errcode(ERRCODE_SYNTAX_ERROR),
1031 : errmsg("unrecognized %s option \"%s\"",
1032 : "CHECKPOINT", opt->defname),
1033 : parser_errposition(pstate, opt->location)));
1034 : }
1035 :
1036 457 : if (!has_privs_of_role(GetUserId(), ROLE_PG_CHECKPOINT))
1037 0 : ereport(ERROR,
1038 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
1039 : /* translator: %s is name of an SQL command (e.g., CHECKPOINT) */
1040 : errmsg("permission denied to execute %s command",
1041 : "CHECKPOINT"),
1042 : errdetail("Only roles with privileges of the \"%s\" role may execute this command.",
1043 : "pg_checkpoint")));
1044 :
1045 914 : RequestCheckpoint(CHECKPOINT_WAIT |
1046 457 : (fast ? CHECKPOINT_FAST : 0) |
1047 457 : (unlogged ? CHECKPOINT_FLUSH_UNLOGGED : 0) |
1048 457 : (RecoveryInProgress() ? 0 : CHECKPOINT_FORCE));
1049 457 : }
1050 :
1051 : /*
1052 : * RequestCheckpoint
1053 : * Called in backend processes to request a checkpoint
1054 : *
1055 : * flags is a bitwise OR of the following:
1056 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
1057 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
1058 : * CHECKPOINT_FAST: finish the checkpoint ASAP,
1059 : * ignoring checkpoint_completion_target parameter.
1060 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
1061 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
1062 : * CHECKPOINT_END_OF_RECOVERY, and the CHECKPOINT command).
1063 : * CHECKPOINT_WAIT: wait for completion before returning (otherwise,
1064 : * just signal checkpointer to do it, and return).
1065 : * CHECKPOINT_CAUSE_XLOG: checkpoint is requested due to xlog filling.
1066 : * (This affects logging, and in particular enables CheckPointWarning.)
1067 : */
1068 : void
1069 2718 : RequestCheckpoint(int flags)
1070 : {
1071 : int ntries;
1072 : int old_failed,
1073 : old_started;
1074 :
1075 : /*
1076 : * If in a standalone backend, just do it ourselves.
1077 : */
1078 2718 : if (!IsPostmasterEnvironment)
1079 : {
1080 : /*
1081 : * There's no point in doing slow checkpoints in a standalone backend,
1082 : * because there's no other backends the checkpoint could disrupt.
1083 : */
1084 229 : CreateCheckPoint(flags | CHECKPOINT_FAST);
1085 :
1086 : /* Free all smgr objects, as CheckpointerMain() normally would. */
1087 229 : smgrdestroyall();
1088 :
1089 229 : return;
1090 : }
1091 :
1092 : /*
1093 : * Atomically set the request flags, and take a snapshot of the counters.
1094 : * When we see ckpt_started > old_started, we know the flags we set here
1095 : * have been seen by checkpointer.
1096 : *
1097 : * Note that we OR the flags with any existing flags, to avoid overriding
1098 : * a "stronger" request by another backend. The flag senses must be
1099 : * chosen to make this work!
1100 : */
1101 2489 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1102 :
1103 2489 : old_failed = CheckpointerShmem->ckpt_failed;
1104 2489 : old_started = CheckpointerShmem->ckpt_started;
1105 2489 : CheckpointerShmem->ckpt_flags |= (flags | CHECKPOINT_REQUESTED);
1106 :
1107 2489 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1108 :
1109 : /*
1110 : * Set checkpointer's latch to request checkpoint. It's possible that the
1111 : * checkpointer hasn't started yet, so we will retry a few times if
1112 : * needed. (Actually, more than a few times, since on slow or overloaded
1113 : * buildfarm machines, it's been observed that the checkpointer can take
1114 : * several seconds to start.) However, if not told to wait for the
1115 : * checkpoint to occur, we consider failure to set the latch to be
1116 : * nonfatal and merely LOG it. The checkpointer should see the request
1117 : * when it does start, with or without the SetLatch().
1118 : */
1119 : #define MAX_SIGNAL_TRIES 600 /* max wait 60.0 sec */
1120 2489 : for (ntries = 0;; ntries++)
1121 26 : {
1122 2515 : volatile PROC_HDR *procglobal = ProcGlobal;
1123 2515 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1124 :
1125 2515 : if (checkpointerProc == INVALID_PROC_NUMBER)
1126 : {
1127 27 : if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT))
1128 : {
1129 1 : elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG,
1130 : "could not notify checkpoint: checkpointer is not running");
1131 1 : break;
1132 : }
1133 : }
1134 : else
1135 : {
1136 2488 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1137 : /* notified successfully */
1138 2488 : break;
1139 : }
1140 :
1141 26 : CHECK_FOR_INTERRUPTS();
1142 26 : pg_usleep(100000L); /* wait 0.1 sec, then retry */
1143 : }
1144 :
1145 : /*
1146 : * If requested, wait for completion. We detect completion according to
1147 : * the algorithm given above.
1148 : */
1149 2489 : if (flags & CHECKPOINT_WAIT)
1150 : {
1151 : int new_started,
1152 : new_failed;
1153 :
1154 : /* Wait for a new checkpoint to start. */
1155 879 : ConditionVariablePrepareToSleep(&CheckpointerShmem->start_cv);
1156 : for (;;)
1157 : {
1158 1658 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1159 1658 : new_started = CheckpointerShmem->ckpt_started;
1160 1658 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1161 :
1162 1658 : if (new_started != old_started)
1163 879 : break;
1164 :
1165 779 : ConditionVariableSleep(&CheckpointerShmem->start_cv,
1166 : WAIT_EVENT_CHECKPOINT_START);
1167 : }
1168 879 : ConditionVariableCancelSleep();
1169 :
1170 : /*
1171 : * We are waiting for ckpt_done >= new_started, in a modulo sense.
1172 : */
1173 879 : ConditionVariablePrepareToSleep(&CheckpointerShmem->done_cv);
1174 : for (;;)
1175 707 : {
1176 : int new_done;
1177 :
1178 1586 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1179 1586 : new_done = CheckpointerShmem->ckpt_done;
1180 1586 : new_failed = CheckpointerShmem->ckpt_failed;
1181 1586 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1182 :
1183 1586 : if (new_done - new_started >= 0)
1184 879 : break;
1185 :
1186 707 : ConditionVariableSleep(&CheckpointerShmem->done_cv,
1187 : WAIT_EVENT_CHECKPOINT_DONE);
1188 : }
1189 879 : ConditionVariableCancelSleep();
1190 :
1191 879 : if (new_failed != old_failed)
1192 0 : ereport(ERROR,
1193 : (errmsg("checkpoint request failed"),
1194 : errhint("Consult recent messages in the server log for details.")));
1195 : }
1196 : }
1197 :
1198 : /*
1199 : * ForwardSyncRequest
1200 : * Forward a file-fsync request from a backend to the checkpointer
1201 : *
1202 : * Whenever a backend is compelled to write directly to a relation
1203 : * (which should be seldom, if the background writer is getting its job done),
1204 : * the backend calls this routine to pass over knowledge that the relation
1205 : * is dirty and must be fsync'd before next checkpoint. We also use this
1206 : * opportunity to count such writes for statistical purposes.
1207 : *
1208 : * To avoid holding the lock for longer than necessary, we normally write
1209 : * to the requests[] queue without checking for duplicates. The checkpointer
1210 : * will have to eliminate dups internally anyway. However, if we discover
1211 : * that the queue is full, we make a pass over the entire queue to compact
1212 : * it. This is somewhat expensive, but the alternative is for the backend
1213 : * to perform its own fsync, which is far more expensive in practice. It
1214 : * is theoretically possible a backend fsync might still be necessary, if
1215 : * the queue is full and contains no duplicate entries. In that case, we
1216 : * let the backend know by returning false.
1217 : */
1218 : bool
1219 1476575 : ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)
1220 : {
1221 : CheckpointerRequest *request;
1222 : bool too_full;
1223 : int insert_pos;
1224 :
1225 1476575 : if (!IsUnderPostmaster)
1226 0 : return false; /* probably shouldn't even get here */
1227 :
1228 1476575 : if (AmCheckpointerProcess())
1229 0 : elog(ERROR, "ForwardSyncRequest must not be called in checkpointer");
1230 :
1231 1476575 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1232 :
1233 : /*
1234 : * If the checkpointer isn't running or the request queue is full, the
1235 : * backend will have to perform its own fsync request. But before forcing
1236 : * that to happen, we can try to compact the request queue.
1237 : */
1238 1476575 : if (CheckpointerShmem->checkpointer_pid == 0 ||
1239 1476155 : (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&
1240 251 : !CompactCheckpointerRequestQueue()))
1241 : {
1242 454 : LWLockRelease(CheckpointerCommLock);
1243 454 : return false;
1244 : }
1245 :
1246 : /* OK, insert request */
1247 1476121 : insert_pos = CheckpointerShmem->tail;
1248 1476121 : request = &CheckpointerShmem->requests[insert_pos];
1249 1476121 : request->ftag = *ftag;
1250 1476121 : request->type = type;
1251 :
1252 1476121 : CheckpointerShmem->tail = (CheckpointerShmem->tail + 1) % CheckpointerShmem->max_requests;
1253 1476121 : CheckpointerShmem->num_requests++;
1254 :
1255 : /* If queue is more than half full, nudge the checkpointer to empty it */
1256 1476121 : too_full = (CheckpointerShmem->num_requests >=
1257 1476121 : CheckpointerShmem->max_requests / 2);
1258 :
1259 1476121 : LWLockRelease(CheckpointerCommLock);
1260 :
1261 : /* ... but not till after we release the lock */
1262 1476121 : if (too_full)
1263 : {
1264 40734 : volatile PROC_HDR *procglobal = ProcGlobal;
1265 40734 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1266 :
1267 40734 : if (checkpointerProc != INVALID_PROC_NUMBER)
1268 40734 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1269 : }
1270 :
1271 1476121 : return true;
1272 : }
1273 :
1274 : /*
1275 : * CompactCheckpointerRequestQueue
1276 : * Remove duplicates from the request queue to avoid backend fsyncs.
1277 : * Returns "true" if any entries were removed.
1278 : *
1279 : * Although a full fsync request queue is not common, it can lead to severe
1280 : * performance problems when it does happen. So far, this situation has
1281 : * only been observed to occur when the system is under heavy write load,
1282 : * and especially during the "sync" phase of a checkpoint. Without this
1283 : * logic, each backend begins doing an fsync for every block written, which
1284 : * gets very expensive and can slow down the whole system.
1285 : *
1286 : * Trying to do this every time the queue is full could lose if there
1287 : * aren't any removable entries. But that should be vanishingly rare in
1288 : * practice: there's one queue entry per shared buffer.
1289 : */
1290 : static bool
1291 251 : CompactCheckpointerRequestQueue(void)
1292 : {
1293 : struct CheckpointerSlotMapping
1294 : {
1295 : CheckpointerRequest request;
1296 : int ring_idx;
1297 : };
1298 :
1299 : int n;
1300 251 : int num_skipped = 0;
1301 : int head;
1302 : int max_requests;
1303 : int num_requests;
1304 : int read_idx,
1305 : write_idx;
1306 : HASHCTL ctl;
1307 : HTAB *htab;
1308 : bool *skip_slot;
1309 :
1310 : /* must hold CheckpointerCommLock in exclusive mode */
1311 : Assert(LWLockHeldByMe(CheckpointerCommLock));
1312 :
1313 : /* Avoid memory allocations in a critical section. */
1314 251 : if (CritSectionCount > 0)
1315 0 : return false;
1316 :
1317 251 : max_requests = CheckpointerShmem->max_requests;
1318 251 : num_requests = CheckpointerShmem->num_requests;
1319 :
1320 : /* Initialize skip_slot array */
1321 251 : skip_slot = palloc0_array(bool, max_requests);
1322 :
1323 251 : head = CheckpointerShmem->head;
1324 :
1325 : /* Initialize temporary hash table */
1326 251 : ctl.keysize = sizeof(CheckpointerRequest);
1327 251 : ctl.entrysize = sizeof(struct CheckpointerSlotMapping);
1328 251 : ctl.hcxt = CurrentMemoryContext;
1329 :
1330 251 : htab = hash_create("CompactCheckpointerRequestQueue",
1331 251 : CheckpointerShmem->num_requests,
1332 : &ctl,
1333 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1334 :
1335 : /*
1336 : * The basic idea here is that a request can be skipped if it's followed
1337 : * by a later, identical request. It might seem more sensible to work
1338 : * backwards from the end of the queue and check whether a request is
1339 : * *preceded* by an earlier, identical request, in the hopes of doing less
1340 : * copying. But that might change the semantics, if there's an
1341 : * intervening SYNC_FORGET_REQUEST or SYNC_FILTER_REQUEST, so we do it
1342 : * this way. It would be possible to be even smarter if we made the code
1343 : * below understand the specific semantics of such requests (it could blow
1344 : * away preceding entries that would end up being canceled anyhow), but
1345 : * it's not clear that the extra complexity would buy us anything.
1346 : */
1347 251 : read_idx = head;
1348 25323 : for (n = 0; n < num_requests; n++)
1349 : {
1350 : CheckpointerRequest *request;
1351 : struct CheckpointerSlotMapping *slotmap;
1352 : bool found;
1353 :
1354 : /*
1355 : * We use the request struct directly as a hashtable key. This
1356 : * assumes that any padding bytes in the structs are consistently the
1357 : * same, which should be okay because we zeroed them in
1358 : * CheckpointerShmemInit. Note also that RelFileLocator had better
1359 : * contain no pad bytes.
1360 : */
1361 25072 : request = &CheckpointerShmem->requests[read_idx];
1362 25072 : slotmap = hash_search(htab, request, HASH_ENTER, &found);
1363 25072 : if (found)
1364 : {
1365 : /* Duplicate, so mark the previous occurrence as skippable */
1366 14789 : skip_slot[slotmap->ring_idx] = true;
1367 14789 : num_skipped++;
1368 : }
1369 : /* Remember slot containing latest occurrence of this request value */
1370 25072 : slotmap->ring_idx = read_idx;
1371 :
1372 : /* Move to the next request in the ring buffer */
1373 25072 : read_idx = (read_idx + 1) % max_requests;
1374 : }
1375 :
1376 : /* Done with the hash table. */
1377 251 : hash_destroy(htab);
1378 :
1379 : /* If no duplicates, we're out of luck. */
1380 251 : if (!num_skipped)
1381 : {
1382 34 : pfree(skip_slot);
1383 34 : return false;
1384 : }
1385 :
1386 : /* We found some duplicates; remove them. */
1387 217 : read_idx = write_idx = head;
1388 23065 : for (n = 0; n < num_requests; n++)
1389 : {
1390 : /* If this slot is NOT skipped, keep it */
1391 22848 : if (!skip_slot[read_idx])
1392 : {
1393 : /* If the read and write positions are different, copy the request */
1394 8059 : if (write_idx != read_idx)
1395 6726 : CheckpointerShmem->requests[write_idx] =
1396 6726 : CheckpointerShmem->requests[read_idx];
1397 :
1398 : /* Advance the write position */
1399 8059 : write_idx = (write_idx + 1) % max_requests;
1400 : }
1401 :
1402 22848 : read_idx = (read_idx + 1) % max_requests;
1403 : }
1404 :
1405 : /*
1406 : * Update ring buffer state: head remains the same, tail moves, count
1407 : * decreases
1408 : */
1409 217 : CheckpointerShmem->tail = write_idx;
1410 217 : CheckpointerShmem->num_requests -= num_skipped;
1411 :
1412 217 : ereport(DEBUG1,
1413 : (errmsg_internal("compacted fsync request queue from %d entries to %d entries",
1414 : num_requests, CheckpointerShmem->num_requests)));
1415 :
1416 : /* Cleanup. */
1417 217 : pfree(skip_slot);
1418 217 : return true;
1419 : }
1420 :
1421 : /*
1422 : * AbsorbSyncRequests
1423 : * Retrieve queued sync requests and pass them to sync mechanism.
1424 : *
1425 : * This is exported because it must be called during CreateCheckPoint;
1426 : * we have to be sure we have accepted all pending requests just before
1427 : * we start fsync'ing. Since CreateCheckPoint sometimes runs in
1428 : * non-checkpointer processes, do nothing if not checkpointer.
1429 : */
1430 : void
1431 21358 : AbsorbSyncRequests(void)
1432 : {
1433 21358 : CheckpointerRequest *requests = NULL;
1434 : CheckpointerRequest *request;
1435 : int n,
1436 : i;
1437 : bool loop;
1438 :
1439 21358 : if (!AmCheckpointerProcess())
1440 724 : return;
1441 :
1442 : do
1443 : {
1444 20634 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1445 :
1446 : /*---
1447 : * We try to avoid holding the lock for a long time by:
1448 : * 1. Copying the request array and processing the requests after
1449 : * releasing the lock;
1450 : * 2. Processing not the whole queue, but only batches of
1451 : * CKPT_REQ_BATCH_SIZE at once.
1452 : *
1453 : * Once we have cleared the requests from shared memory, we must
1454 : * PANIC if we then fail to absorb them (e.g., because our hashtable
1455 : * runs out of memory). This is because the system cannot run safely
1456 : * if we are unable to fsync what we have been told to fsync.
1457 : * Fortunately, the hashtable is so small that the problem is quite
1458 : * unlikely to arise in practice.
1459 : *
1460 : * Note: The maximum possible size of a ring buffer is
1461 : * MAX_CHECKPOINT_REQUESTS entries, which fit into a maximum palloc
1462 : * allocation size of 1Gb. Our maximum batch size,
1463 : * CKPT_REQ_BATCH_SIZE, is even smaller.
1464 : */
1465 20634 : n = Min(CheckpointerShmem->num_requests, CKPT_REQ_BATCH_SIZE);
1466 20634 : if (n > 0)
1467 : {
1468 11010 : if (!requests)
1469 11010 : requests = (CheckpointerRequest *) palloc(n * sizeof(CheckpointerRequest));
1470 :
1471 1315057 : for (i = 0; i < n; i++)
1472 : {
1473 1304047 : requests[i] = CheckpointerShmem->requests[CheckpointerShmem->head];
1474 1304047 : CheckpointerShmem->head = (CheckpointerShmem->head + 1) % CheckpointerShmem->max_requests;
1475 : }
1476 :
1477 11010 : CheckpointerShmem->num_requests -= n;
1478 :
1479 : }
1480 :
1481 20634 : START_CRIT_SECTION();
1482 :
1483 : /* Are there any requests in the queue? If so, keep going. */
1484 20634 : loop = CheckpointerShmem->num_requests != 0;
1485 :
1486 20634 : LWLockRelease(CheckpointerCommLock);
1487 :
1488 1324681 : for (request = requests; n > 0; request++, n--)
1489 1304047 : RememberSyncRequest(&request->ftag, request->type);
1490 :
1491 20634 : END_CRIT_SECTION();
1492 20634 : } while (loop);
1493 :
1494 20634 : if (requests)
1495 11010 : pfree(requests);
1496 : }
1497 :
1498 : /*
1499 : * Update any shared memory configurations based on config parameters
1500 : */
1501 : static void
1502 713 : UpdateSharedMemoryConfig(void)
1503 : {
1504 : /* update global shmem state for sync rep */
1505 713 : SyncRepUpdateSyncStandbysDefined();
1506 :
1507 : /*
1508 : * If full_page_writes has been changed by SIGHUP, we update it in shared
1509 : * memory and write an XLOG_FPW_CHANGE record.
1510 : */
1511 713 : UpdateFullPageWrites();
1512 :
1513 713 : elog(DEBUG2, "checkpointer updated shared memory configuration values");
1514 713 : }
1515 :
1516 : /*
1517 : * FirstCallSinceLastCheckpoint allows a process to take an action once
1518 : * per checkpoint cycle by asynchronously checking for checkpoint completion.
1519 : */
1520 : bool
1521 14504 : FirstCallSinceLastCheckpoint(void)
1522 : {
1523 : static int ckpt_done = 0;
1524 : int new_done;
1525 14504 : bool FirstCall = false;
1526 :
1527 14504 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1528 14504 : new_done = CheckpointerShmem->ckpt_done;
1529 14504 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1530 :
1531 14504 : if (new_done != ckpt_done)
1532 619 : FirstCall = true;
1533 :
1534 14504 : ckpt_done = new_done;
1535 :
1536 14504 : return FirstCall;
1537 : }
1538 :
1539 : /*
1540 : * Wake up the checkpointer process.
1541 : */
1542 : void
1543 1018 : WakeupCheckpointer(void)
1544 : {
1545 1018 : volatile PROC_HDR *procglobal = ProcGlobal;
1546 1018 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1547 :
1548 1018 : if (checkpointerProc != INVALID_PROC_NUMBER)
1549 680 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1550 1018 : }
|
