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 checkpointer is started by the postmaster as soon as the startup
14 * subprocess finishes, or as soon as recovery begins if we are doing archive
15 * recovery. It remains alive until the postmaster commands it to terminate.
16 * Normal termination is by SIGUSR2, which instructs the checkpointer to
17 * execute a shutdown checkpoint and then exit(0). (All backends must be
18 * stopped before SIGUSR2 is issued!) Emergency termination is by SIGQUIT;
19 * like any backend, the checkpointer 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) 1996-2019, PostgreSQL Global Development Group
30 *
31 *
32 * IDENTIFICATION
33 * src/backend/postmaster/checkpointer.c
34 *
35 *-------------------------------------------------------------------------
36 */
37#include "postgres.h"
38
39#include <signal.h>
40#include <sys/time.h>
41#include <time.h>
42#include <unistd.h>
43
44#include "access/xlog.h"
45#include "access/xlog_internal.h"
46#include "libpq/pqsignal.h"
47#include "miscadmin.h"
48#include "pgstat.h"
49#include "postmaster/bgwriter.h"
50#include "replication/syncrep.h"
51#include "storage/bufmgr.h"
52#include "storage/condition_variable.h"
53#include "storage/fd.h"
54#include "storage/ipc.h"
55#include "storage/lwlock.h"
56#include "storage/proc.h"
57#include "storage/shmem.h"
58#include "storage/smgr.h"
59#include "storage/spin.h"
60#include "utils/guc.h"
61#include "utils/memutils.h"
62#include "utils/resowner.h"
63
64
65/*----------
66 * Shared memory area for communication between checkpointer and backends
67 *
68 * The ckpt counters allow backends to watch for completion of a checkpoint
69 * request they send. Here's how it works:
70 * * At start of a checkpoint, checkpointer reads (and clears) the request
71 * flags and increments ckpt_started, while holding ckpt_lck.
72 * * On completion of a checkpoint, checkpointer sets ckpt_done to
73 * equal ckpt_started.
74 * * On failure of a checkpoint, checkpointer increments ckpt_failed
75 * and sets ckpt_done to equal ckpt_started.
76 *
77 * The algorithm for backends is:
78 * 1. Record current values of ckpt_failed and ckpt_started, and
79 * set request flags, while holding ckpt_lck.
80 * 2. Send signal to request checkpoint.
81 * 3. Sleep until ckpt_started changes. Now you know a checkpoint has
82 * begun since you started this algorithm (although *not* that it was
83 * specifically initiated by your signal), and that it is using your flags.
84 * 4. Record new value of ckpt_started.
85 * 5. Sleep until ckpt_done >= saved value of ckpt_started. (Use modulo
86 * arithmetic here in case counters wrap around.) Now you know a
87 * checkpoint has started and completed, but not whether it was
88 * successful.
89 * 6. If ckpt_failed is different from the originally saved value,
90 * assume request failed; otherwise it was definitely successful.
91 *
92 * ckpt_flags holds the OR of the checkpoint request flags sent by all
93 * requesting backends since the last checkpoint start. The flags are
94 * chosen so that OR'ing is the correct way to combine multiple requests.
95 *
96 * num_backend_writes is used to count the number of buffer writes performed
97 * by user backend processes. This counter should be wide enough that it
98 * can't overflow during a single processing cycle. num_backend_fsync
99 * counts the subset of those writes that also had to do their own fsync,
100 * because the checkpointer failed to absorb their request.
101 *
102 * The requests array holds fsync requests sent by backends and not yet
103 * absorbed by the checkpointer.
104 *
105 * Unlike the checkpoint fields, num_backend_writes, num_backend_fsync, and
106 * the requests fields are protected by CheckpointerCommLock.
107 *----------
108 */
109typedef struct
110{
111 SyncRequestType type; /* request type */
112 FileTag ftag; /* file identifier */
113} CheckpointerRequest;
114
115typedef struct
116{
117 pid_t checkpointer_pid; /* PID (0 if not started) */
118
119 slock_t ckpt_lck; /* protects all the ckpt_* fields */
120
121 int ckpt_started; /* advances when checkpoint starts */
122 int ckpt_done; /* advances when checkpoint done */
123 int ckpt_failed; /* advances when checkpoint fails */
124
125 int ckpt_flags; /* checkpoint flags, as defined in xlog.h */
126
127 ConditionVariable start_cv; /* signaled when ckpt_started advances */
128 ConditionVariable done_cv; /* signaled when ckpt_done advances */
129
130 uint32 num_backend_writes; /* counts user backend buffer writes */
131 uint32 num_backend_fsync; /* counts user backend fsync calls */
132
133 int num_requests; /* current # of requests */
134 int max_requests; /* allocated array size */
135 CheckpointerRequest requests[FLEXIBLE_ARRAY_MEMBER];
136} CheckpointerShmemStruct;
137
138static CheckpointerShmemStruct *CheckpointerShmem;
139
140/* interval for calling AbsorbSyncRequests in CheckpointWriteDelay */
141#define WRITES_PER_ABSORB 1000
142
143/*
144 * GUC parameters
145 */
146int CheckPointTimeout = 300;
147int CheckPointWarning = 30;
148double CheckPointCompletionTarget = 0.5;
149
150/*
151 * Flags set by interrupt handlers for later service in the main loop.
152 */
153static volatile sig_atomic_t got_SIGHUP = false;
154static volatile sig_atomic_t shutdown_requested = false;
155
156/*
157 * Private state
158 */
159static bool ckpt_active = false;
160
161/* these values are valid when ckpt_active is true: */
162static pg_time_t ckpt_start_time;
163static XLogRecPtr ckpt_start_recptr;
164static double ckpt_cached_elapsed;
165
166static pg_time_t last_checkpoint_time;
167static pg_time_t last_xlog_switch_time;
168
169/* Prototypes for private functions */
170
171static void CheckArchiveTimeout(void);
172static bool IsCheckpointOnSchedule(double progress);
173static bool ImmediateCheckpointRequested(void);
174static bool CompactCheckpointerRequestQueue(void);
175static void UpdateSharedMemoryConfig(void);
176
177/* Signal handlers */
178
179static void chkpt_quickdie(SIGNAL_ARGS);
180static void ChkptSigHupHandler(SIGNAL_ARGS);
181static void ReqCheckpointHandler(SIGNAL_ARGS);
182static void chkpt_sigusr1_handler(SIGNAL_ARGS);
183static void ReqShutdownHandler(SIGNAL_ARGS);
184
185
186/*
187 * Main entry point for checkpointer process
188 *
189 * This is invoked from AuxiliaryProcessMain, which has already created the
190 * basic execution environment, but not enabled signals yet.
191 */
192void
193CheckpointerMain(void)
194{
195 sigjmp_buf local_sigjmp_buf;
196 MemoryContext checkpointer_context;
197
198 CheckpointerShmem->checkpointer_pid = MyProcPid;
199
200 /*
201 * Properly accept or ignore signals the postmaster might send us
202 *
203 * Note: we deliberately ignore SIGTERM, because during a standard Unix
204 * system shutdown cycle, init will SIGTERM all processes at once. We
205 * want to wait for the backends to exit, whereupon the postmaster will
206 * tell us it's okay to shut down (via SIGUSR2).
207 */
208 pqsignal(SIGHUP, ChkptSigHupHandler); /* set flag to read config file */
209 pqsignal(SIGINT, ReqCheckpointHandler); /* request checkpoint */
210 pqsignal(SIGTERM, SIG_IGN); /* ignore SIGTERM */
211 pqsignal(SIGQUIT, chkpt_quickdie); /* hard crash time */
212 pqsignal(SIGALRM, SIG_IGN);
213 pqsignal(SIGPIPE, SIG_IGN);
214 pqsignal(SIGUSR1, chkpt_sigusr1_handler);
215 pqsignal(SIGUSR2, ReqShutdownHandler); /* request shutdown */
216
217 /*
218 * Reset some signals that are accepted by postmaster but not here
219 */
220 pqsignal(SIGCHLD, SIG_DFL);
221
222 /* We allow SIGQUIT (quickdie) at all times */
223 sigdelset(&BlockSig, SIGQUIT);
224
225 /*
226 * Initialize so that first time-driven event happens at the correct time.
227 */
228 last_checkpoint_time = last_xlog_switch_time = (pg_time_t) time(NULL);
229
230 /*
231 * Create a memory context that we will do all our work in. We do this so
232 * that we can reset the context during error recovery and thereby avoid
233 * possible memory leaks. Formerly this code just ran in
234 * TopMemoryContext, but resetting that would be a really bad idea.
235 */
236 checkpointer_context = AllocSetContextCreate(TopMemoryContext,
237 "Checkpointer",
238 ALLOCSET_DEFAULT_SIZES);
239 MemoryContextSwitchTo(checkpointer_context);
240
241 /*
242 * If an exception is encountered, processing resumes here.
243 *
244 * See notes in postgres.c about the design of this coding.
245 */
246 if (sigsetjmp(local_sigjmp_buf, 1) != 0)
247 {
248 /* Since not using PG_TRY, must reset error stack by hand */
249 error_context_stack = NULL;
250
251 /* Prevent interrupts while cleaning up */
252 HOLD_INTERRUPTS();
253
254 /* Report the error to the server log */
255 EmitErrorReport();
256
257 /*
258 * These operations are really just a minimal subset of
259 * AbortTransaction(). We don't have very many resources to worry
260 * about in checkpointer, but we do have LWLocks, buffers, and temp
261 * files.
262 */
263 LWLockReleaseAll();
264 ConditionVariableCancelSleep();
265 pgstat_report_wait_end();
266 AbortBufferIO();
267 UnlockBuffers();
268 ReleaseAuxProcessResources(false);
269 AtEOXact_Buffers(false);
270 AtEOXact_SMgr();
271 AtEOXact_Files(false);
272 AtEOXact_HashTables(false);
273
274 /* Warn any waiting backends that the checkpoint failed. */
275 if (ckpt_active)
276 {
277 SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
278 CheckpointerShmem->ckpt_failed++;
279 CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
280 SpinLockRelease(&CheckpointerShmem->ckpt_lck);
281
282 ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
283
284 ckpt_active = false;
285 }
286
287 /*
288 * Now return to normal top-level context and clear ErrorContext for
289 * next time.
290 */
291 MemoryContextSwitchTo(checkpointer_context);
292 FlushErrorState();
293
294 /* Flush any leaked data in the top-level context */
295 MemoryContextResetAndDeleteChildren(checkpointer_context);
296
297 /* Now we can allow interrupts again */
298 RESUME_INTERRUPTS();
299
300 /*
301 * Sleep at least 1 second after any error. A write error is likely
302 * to be repeated, and we don't want to be filling the error logs as
303 * fast as we can.
304 */
305 pg_usleep(1000000L);
306
307 /*
308 * Close all open files after any error. This is helpful on Windows,
309 * where holding deleted files open causes various strange errors.
310 * It's not clear we need it elsewhere, but shouldn't hurt.
311 */
312 smgrcloseall();
313 }
314
315 /* We can now handle ereport(ERROR) */
316 PG_exception_stack = &local_sigjmp_buf;
317
318 /*
319 * Unblock signals (they were blocked when the postmaster forked us)
320 */
321 PG_SETMASK(&UnBlockSig);
322
323 /*
324 * Ensure all shared memory values are set correctly for the config. Doing
325 * this here ensures no race conditions from other concurrent updaters.
326 */
327 UpdateSharedMemoryConfig();
328
329 /*
330 * Advertise our latch that backends can use to wake us up while we're
331 * sleeping.
332 */
333 ProcGlobal->checkpointerLatch = &MyProc->procLatch;
334
335 /*
336 * Loop forever
337 */
338 for (;;)
339 {
340 bool do_checkpoint = false;
341 int flags = 0;
342 pg_time_t now;
343 int elapsed_secs;
344 int cur_timeout;
345
346 /* Clear any already-pending wakeups */
347 ResetLatch(MyLatch);
348
349 /*
350 * Process any requests or signals received recently.
351 */
352 AbsorbSyncRequests();
353
354 if (got_SIGHUP)
355 {
356 got_SIGHUP = false;
357 ProcessConfigFile(PGC_SIGHUP);
358
359 /*
360 * Checkpointer is the last process to shut down, so we ask it to
361 * hold the keys for a range of other tasks required most of which
362 * have nothing to do with checkpointing at all.
363 *
364 * For various reasons, some config values can change dynamically
365 * so the primary copy of them is held in shared memory to make
366 * sure all backends see the same value. We make Checkpointer
367 * responsible for updating the shared memory copy if the
368 * parameter setting changes because of SIGHUP.
369 */
370 UpdateSharedMemoryConfig();
371 }
372 if (shutdown_requested)
373 {
374 /*
375 * From here on, elog(ERROR) should end with exit(1), not send
376 * control back to the sigsetjmp block above
377 */
378 ExitOnAnyError = true;
379 /* Close down the database */
380 ShutdownXLOG(0, 0);
381 /* Normal exit from the checkpointer is here */
382 proc_exit(0); /* done */
383 }
384
385 /*
386 * Detect a pending checkpoint request by checking whether the flags
387 * word in shared memory is nonzero. We shouldn't need to acquire the
388 * ckpt_lck for this.
389 */
390 if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
391 {
392 do_checkpoint = true;
393 BgWriterStats.m_requested_checkpoints++;
394 }
395
396 /*
397 * Force a checkpoint if too much time has elapsed since the last one.
398 * Note that we count a timed checkpoint in stats only when this
399 * occurs without an external request, but we set the CAUSE_TIME flag
400 * bit even if there is also an external request.
401 */
402 now = (pg_time_t) time(NULL);
403 elapsed_secs = now - last_checkpoint_time;
404 if (elapsed_secs >= CheckPointTimeout)
405 {
406 if (!do_checkpoint)
407 BgWriterStats.m_timed_checkpoints++;
408 do_checkpoint = true;
409 flags |= CHECKPOINT_CAUSE_TIME;
410 }
411
412 /*
413 * Do a checkpoint if requested.
414 */
415 if (do_checkpoint)
416 {
417 bool ckpt_performed = false;
418 bool do_restartpoint;
419
420 /*
421 * Check if we should perform a checkpoint or a restartpoint. As a
422 * side-effect, RecoveryInProgress() initializes TimeLineID if
423 * it's not set yet.
424 */
425 do_restartpoint = RecoveryInProgress();
426
427 /*
428 * Atomically fetch the request flags to figure out what kind of a
429 * checkpoint we should perform, and increase the started-counter
430 * to acknowledge that we've started a new checkpoint.
431 */
432 SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
433 flags |= CheckpointerShmem->ckpt_flags;
434 CheckpointerShmem->ckpt_flags = 0;
435 CheckpointerShmem->ckpt_started++;
436 SpinLockRelease(&CheckpointerShmem->ckpt_lck);
437
438 ConditionVariableBroadcast(&CheckpointerShmem->start_cv);
439
440 /*
441 * The end-of-recovery checkpoint is a real checkpoint that's
442 * performed while we're still in recovery.
443 */
444 if (flags & CHECKPOINT_END_OF_RECOVERY)
445 do_restartpoint = false;
446
447 /*
448 * We will warn if (a) too soon since last checkpoint (whatever
449 * caused it) and (b) somebody set the CHECKPOINT_CAUSE_XLOG flag
450 * since the last checkpoint start. Note in particular that this
451 * implementation will not generate warnings caused by
452 * CheckPointTimeout < CheckPointWarning.
453 */
454 if (!do_restartpoint &&
455 (flags & CHECKPOINT_CAUSE_XLOG) &&
456 elapsed_secs < CheckPointWarning)
457 ereport(LOG,
458 (errmsg_plural("checkpoints are occurring too frequently (%d second apart)",
459 "checkpoints are occurring too frequently (%d seconds apart)",
460 elapsed_secs,
461 elapsed_secs),
462 errhint("Consider increasing the configuration parameter \"max_wal_size\".")));
463
464 /*
465 * Initialize checkpointer-private variables used during
466 * checkpoint.
467 */
468 ckpt_active = true;
469 if (do_restartpoint)
470 ckpt_start_recptr = GetXLogReplayRecPtr(NULL);
471 else
472 ckpt_start_recptr = GetInsertRecPtr();
473 ckpt_start_time = now;
474 ckpt_cached_elapsed = 0;
475
476 /*
477 * Do the checkpoint.
478 */
479 if (!do_restartpoint)
480 {
481 CreateCheckPoint(flags);
482 ckpt_performed = true;
483 }
484 else
485 ckpt_performed = CreateRestartPoint(flags);
486
487 /*
488 * After any checkpoint, close all smgr files. This is so we
489 * won't hang onto smgr references to deleted files indefinitely.
490 */
491 smgrcloseall();
492
493 /*
494 * Indicate checkpoint completion to any waiting backends.
495 */
496 SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
497 CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
498 SpinLockRelease(&CheckpointerShmem->ckpt_lck);
499
500 ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
501
502 if (ckpt_performed)
503 {
504 /*
505 * Note we record the checkpoint start time not end time as
506 * last_checkpoint_time. This is so that time-driven
507 * checkpoints happen at a predictable spacing.
508 */
509 last_checkpoint_time = now;
510 }
511 else
512 {
513 /*
514 * We were not able to perform the restartpoint (checkpoints
515 * throw an ERROR in case of error). Most likely because we
516 * have not received any new checkpoint WAL records since the
517 * last restartpoint. Try again in 15 s.
518 */
519 last_checkpoint_time = now - CheckPointTimeout + 15;
520 }
521
522 ckpt_active = false;
523 }
524
525 /* Check for archive_timeout and switch xlog files if necessary. */
526 CheckArchiveTimeout();
527
528 /*
529 * Send off activity statistics to the stats collector. (The reason
530 * why we re-use bgwriter-related code for this is that the bgwriter
531 * and checkpointer used to be just one process. It's probably not
532 * worth the trouble to split the stats support into two independent
533 * stats message types.)
534 */
535 pgstat_send_bgwriter();
536
537 /*
538 * Sleep until we are signaled or it's time for another checkpoint or
539 * xlog file switch.
540 */
541 now = (pg_time_t) time(NULL);
542 elapsed_secs = now - last_checkpoint_time;
543 if (elapsed_secs >= CheckPointTimeout)
544 continue; /* no sleep for us ... */
545 cur_timeout = CheckPointTimeout - elapsed_secs;
546 if (XLogArchiveTimeout > 0 && !RecoveryInProgress())
547 {
548 elapsed_secs = now - last_xlog_switch_time;
549 if (elapsed_secs >= XLogArchiveTimeout)
550 continue; /* no sleep for us ... */
551 cur_timeout = Min(cur_timeout, XLogArchiveTimeout - elapsed_secs);
552 }
553
554 (void) WaitLatch(MyLatch,
555 WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
556 cur_timeout * 1000L /* convert to ms */ ,
557 WAIT_EVENT_CHECKPOINTER_MAIN);
558 }
559}
560
561/*
562 * CheckArchiveTimeout -- check for archive_timeout and switch xlog files
563 *
564 * This will switch to a new WAL file and force an archive file write if
565 * meaningful activity is recorded in the current WAL file. This includes most
566 * writes, including just a single checkpoint record, but excludes WAL records
567 * that were inserted with the XLOG_MARK_UNIMPORTANT flag being set (like
568 * snapshots of running transactions). Such records, depending on
569 * configuration, occur on regular intervals and don't contain important
570 * information. This avoids generating archives with a few unimportant
571 * records.
572 */
573static void
574CheckArchiveTimeout(void)
575{
576 pg_time_t now;
577 pg_time_t last_time;
578 XLogRecPtr last_switch_lsn;
579
580 if (XLogArchiveTimeout <= 0 || RecoveryInProgress())
581 return;
582
583 now = (pg_time_t) time(NULL);
584
585 /* First we do a quick check using possibly-stale local state. */
586 if ((int) (now - last_xlog_switch_time) < XLogArchiveTimeout)
587 return;
588
589 /*
590 * Update local state ... note that last_xlog_switch_time is the last time
591 * a switch was performed *or requested*.
592 */
593 last_time = GetLastSegSwitchData(&last_switch_lsn);
594
595 last_xlog_switch_time = Max(last_xlog_switch_time, last_time);
596
597 /* Now we can do the real checks */
598 if ((int) (now - last_xlog_switch_time) >= XLogArchiveTimeout)
599 {
600 /*
601 * Switch segment only when "important" WAL has been logged since the
602 * last segment switch (last_switch_lsn points to end of segment
603 * switch occurred in).
604 */
605 if (GetLastImportantRecPtr() > last_switch_lsn)
606 {
607 XLogRecPtr switchpoint;
608
609 /* mark switch as unimportant, avoids triggering checkpoints */
610 switchpoint = RequestXLogSwitch(true);
611
612 /*
613 * If the returned pointer points exactly to a segment boundary,
614 * assume nothing happened.
615 */
616 if (XLogSegmentOffset(switchpoint, wal_segment_size) != 0)
617 elog(DEBUG1, "write-ahead log switch forced (archive_timeout=%d)",
618 XLogArchiveTimeout);
619 }
620
621 /*
622 * Update state in any case, so we don't retry constantly when the
623 * system is idle.
624 */
625 last_xlog_switch_time = now;
626 }
627}
628
629/*
630 * Returns true if an immediate checkpoint request is pending. (Note that
631 * this does not check the *current* checkpoint's IMMEDIATE flag, but whether
632 * there is one pending behind it.)
633 */
634static bool
635ImmediateCheckpointRequested(void)
636{
637 volatile CheckpointerShmemStruct *cps = CheckpointerShmem;
638
639 /*
640 * We don't need to acquire the ckpt_lck in this case because we're only
641 * looking at a single flag bit.
642 */
643 if (cps->ckpt_flags & CHECKPOINT_IMMEDIATE)
644 return true;
645 return false;
646}
647
648/*
649 * CheckpointWriteDelay -- control rate of checkpoint
650 *
651 * This function is called after each page write performed by BufferSync().
652 * It is responsible for throttling BufferSync()'s write rate to hit
653 * checkpoint_completion_target.
654 *
655 * The checkpoint request flags should be passed in; currently the only one
656 * examined is CHECKPOINT_IMMEDIATE, which disables delays between writes.
657 *
658 * 'progress' is an estimate of how much of the work has been done, as a
659 * fraction between 0.0 meaning none, and 1.0 meaning all done.
660 */
661void
662CheckpointWriteDelay(int flags, double progress)
663{
664 static int absorb_counter = WRITES_PER_ABSORB;
665
666 /* Do nothing if checkpoint is being executed by non-checkpointer process */
667 if (!AmCheckpointerProcess())
668 return;
669
670 /*
671 * Perform the usual duties and take a nap, unless we're behind schedule,
672 * in which case we just try to catch up as quickly as possible.
673 */
674 if (!(flags & CHECKPOINT_IMMEDIATE) &&
675 !shutdown_requested &&
676 !ImmediateCheckpointRequested() &&
677 IsCheckpointOnSchedule(progress))
678 {
679 if (got_SIGHUP)
680 {
681 got_SIGHUP = false;
682 ProcessConfigFile(PGC_SIGHUP);
683 /* update shmem copies of config variables */
684 UpdateSharedMemoryConfig();
685 }
686
687 AbsorbSyncRequests();
688 absorb_counter = WRITES_PER_ABSORB;
689
690 CheckArchiveTimeout();
691
692 /*
693 * Report interim activity statistics to the stats collector.
694 */
695 pgstat_send_bgwriter();
696
697 /*
698 * This sleep used to be connected to bgwriter_delay, typically 200ms.
699 * That resulted in more frequent wakeups if not much work to do.
700 * Checkpointer and bgwriter are no longer related so take the Big
701 * Sleep.
702 */
703 pg_usleep(100000L);
704 }
705 else if (--absorb_counter <= 0)
706 {
707 /*
708 * Absorb pending fsync requests after each WRITES_PER_ABSORB write
709 * operations even when we don't sleep, to prevent overflow of the
710 * fsync request queue.
711 */
712 AbsorbSyncRequests();
713 absorb_counter = WRITES_PER_ABSORB;
714 }
715}
716
717/*
718 * IsCheckpointOnSchedule -- are we on schedule to finish this checkpoint
719 * (or restartpoint) in time?
720 *
721 * Compares the current progress against the time/segments elapsed since last
722 * checkpoint, and returns true if the progress we've made this far is greater
723 * than the elapsed time/segments.
724 */
725static bool
726IsCheckpointOnSchedule(double progress)
727{
728 XLogRecPtr recptr;
729 struct timeval now;
730 double elapsed_xlogs,
731 elapsed_time;
732
733 Assert(ckpt_active);
734
735 /* Scale progress according to checkpoint_completion_target. */
736 progress *= CheckPointCompletionTarget;
737
738 /*
739 * Check against the cached value first. Only do the more expensive
740 * calculations once we reach the target previously calculated. Since
741 * neither time or WAL insert pointer moves backwards, a freshly
742 * calculated value can only be greater than or equal to the cached value.
743 */
744 if (progress < ckpt_cached_elapsed)
745 return false;
746
747 /*
748 * Check progress against WAL segments written and CheckPointSegments.
749 *
750 * We compare the current WAL insert location against the location
751 * computed before calling CreateCheckPoint. The code in XLogInsert that
752 * actually triggers a checkpoint when CheckPointSegments is exceeded
753 * compares against RedoRecptr, so this is not completely accurate.
754 * However, it's good enough for our purposes, we're only calculating an
755 * estimate anyway.
756 *
757 * During recovery, we compare last replayed WAL record's location with
758 * the location computed before calling CreateRestartPoint. That maintains
759 * the same pacing as we have during checkpoints in normal operation, but
760 * we might exceed max_wal_size by a fair amount. That's because there can
761 * be a large gap between a checkpoint's redo-pointer and the checkpoint
762 * record itself, and we only start the restartpoint after we've seen the
763 * checkpoint record. (The gap is typically up to CheckPointSegments *
764 * checkpoint_completion_target where checkpoint_completion_target is the
765 * value that was in effect when the WAL was generated).
766 */
767 if (RecoveryInProgress())
768 recptr = GetXLogReplayRecPtr(NULL);
769 else
770 recptr = GetInsertRecPtr();
771 elapsed_xlogs = (((double) (recptr - ckpt_start_recptr)) /
772 wal_segment_size) / CheckPointSegments;
773
774 if (progress < elapsed_xlogs)
775 {
776 ckpt_cached_elapsed = elapsed_xlogs;
777 return false;
778 }
779
780 /*
781 * Check progress against time elapsed and checkpoint_timeout.
782 */
783 gettimeofday(&now, NULL);
784 elapsed_time = ((double) ((pg_time_t) now.tv_sec - ckpt_start_time) +
785 now.tv_usec / 1000000.0) / CheckPointTimeout;
786
787 if (progress < elapsed_time)
788 {
789 ckpt_cached_elapsed = elapsed_time;
790 return false;
791 }
792
793 /* It looks like we're on schedule. */
794 return true;
795}
796
797
798/* --------------------------------
799 * signal handler routines
800 * --------------------------------
801 */
802
803/*
804 * chkpt_quickdie() occurs when signalled SIGQUIT by the postmaster.
805 *
806 * Some backend has bought the farm,
807 * so we need to stop what we're doing and exit.
808 */
809static void
810chkpt_quickdie(SIGNAL_ARGS)
811{
812 /*
813 * We DO NOT want to run proc_exit() or atexit() callbacks -- we're here
814 * because shared memory may be corrupted, so we don't want to try to
815 * clean up our transaction. Just nail the windows shut and get out of
816 * town. The callbacks wouldn't be safe to run from a signal handler,
817 * anyway.
818 *
819 * Note we do _exit(2) not _exit(0). This is to force the postmaster into
820 * a system reset cycle if someone sends a manual SIGQUIT to a random
821 * backend. This is necessary precisely because we don't clean up our
822 * shared memory state. (The "dead man switch" mechanism in pmsignal.c
823 * should ensure the postmaster sees this as a crash, too, but no harm in
824 * being doubly sure.)
825 */
826 _exit(2);
827}
828
829/* SIGHUP: set flag to re-read config file at next convenient time */
830static void
831ChkptSigHupHandler(SIGNAL_ARGS)
832{
833 int save_errno = errno;
834
835 got_SIGHUP = true;
836 SetLatch(MyLatch);
837
838 errno = save_errno;
839}
840
841/* SIGINT: set flag to run a normal checkpoint right away */
842static void
843ReqCheckpointHandler(SIGNAL_ARGS)
844{
845 int save_errno = errno;
846
847 /*
848 * The signalling process should have set ckpt_flags nonzero, so all we
849 * need do is ensure that our main loop gets kicked out of any wait.
850 */
851 SetLatch(MyLatch);
852
853 errno = save_errno;
854}
855
856/* SIGUSR1: used for latch wakeups */
857static void
858chkpt_sigusr1_handler(SIGNAL_ARGS)
859{
860 int save_errno = errno;
861
862 latch_sigusr1_handler();
863
864 errno = save_errno;
865}
866
867/* SIGUSR2: set flag to run a shutdown checkpoint and exit */
868static void
869ReqShutdownHandler(SIGNAL_ARGS)
870{
871 int save_errno = errno;
872
873 shutdown_requested = true;
874 SetLatch(MyLatch);
875
876 errno = save_errno;
877}
878
879
880/* --------------------------------
881 * communication with backends
882 * --------------------------------
883 */
884
885/*
886 * CheckpointerShmemSize
887 * Compute space needed for checkpointer-related shared memory
888 */
889Size
890CheckpointerShmemSize(void)
891{
892 Size size;
893
894 /*
895 * Currently, the size of the requests[] array is arbitrarily set equal to
896 * NBuffers. This may prove too large or small ...
897 */
898 size = offsetof(CheckpointerShmemStruct, requests);
899 size = add_size(size, mul_size(NBuffers, sizeof(CheckpointerRequest)));
900
901 return size;
902}
903
904/*
905 * CheckpointerShmemInit
906 * Allocate and initialize checkpointer-related shared memory
907 */
908void
909CheckpointerShmemInit(void)
910{
911 Size size = CheckpointerShmemSize();
912 bool found;
913
914 CheckpointerShmem = (CheckpointerShmemStruct *)
915 ShmemInitStruct("Checkpointer Data",
916 size,
917 &found);
918
919 if (!found)
920 {
921 /*
922 * First time through, so initialize. Note that we zero the whole
923 * requests array; this is so that CompactCheckpointerRequestQueue can
924 * assume that any pad bytes in the request structs are zeroes.
925 */
926 MemSet(CheckpointerShmem, 0, size);
927 SpinLockInit(&CheckpointerShmem->ckpt_lck);
928 CheckpointerShmem->max_requests = NBuffers;
929 ConditionVariableInit(&CheckpointerShmem->start_cv);
930 ConditionVariableInit(&CheckpointerShmem->done_cv);
931 }
932}
933
934/*
935 * RequestCheckpoint
936 * Called in backend processes to request a checkpoint
937 *
938 * flags is a bitwise OR of the following:
939 * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
940 * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
941 * CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP,
942 * ignoring checkpoint_completion_target parameter.
943 * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
944 * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
945 * CHECKPOINT_END_OF_RECOVERY).
946 * CHECKPOINT_WAIT: wait for completion before returning (otherwise,
947 * just signal checkpointer to do it, and return).
948 * CHECKPOINT_CAUSE_XLOG: checkpoint is requested due to xlog filling.
949 * (This affects logging, and in particular enables CheckPointWarning.)
950 */
951void
952RequestCheckpoint(int flags)
953{
954 int ntries;
955 int old_failed,
956 old_started;
957
958 /*
959 * If in a standalone backend, just do it ourselves.
960 */
961 if (!IsPostmasterEnvironment)
962 {
963 /*
964 * There's no point in doing slow checkpoints in a standalone backend,
965 * because there's no other backends the checkpoint could disrupt.
966 */
967 CreateCheckPoint(flags | CHECKPOINT_IMMEDIATE);
968
969 /*
970 * After any checkpoint, close all smgr files. This is so we won't
971 * hang onto smgr references to deleted files indefinitely.
972 */
973 smgrcloseall();
974
975 return;
976 }
977
978 /*
979 * Atomically set the request flags, and take a snapshot of the counters.
980 * When we see ckpt_started > old_started, we know the flags we set here
981 * have been seen by checkpointer.
982 *
983 * Note that we OR the flags with any existing flags, to avoid overriding
984 * a "stronger" request by another backend. The flag senses must be
985 * chosen to make this work!
986 */
987 SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
988
989 old_failed = CheckpointerShmem->ckpt_failed;
990 old_started = CheckpointerShmem->ckpt_started;
991 CheckpointerShmem->ckpt_flags |= (flags | CHECKPOINT_REQUESTED);
992
993 SpinLockRelease(&CheckpointerShmem->ckpt_lck);
994
995 /*
996 * Send signal to request checkpoint. It's possible that the checkpointer
997 * hasn't started yet, or is in process of restarting, so we will retry a
998 * few times if needed. (Actually, more than a few times, since on slow
999 * or overloaded buildfarm machines, it's been observed that the
1000 * checkpointer can take several seconds to start.) However, if not told
1001 * to wait for the checkpoint to occur, we consider failure to send the
1002 * signal to be nonfatal and merely LOG it. The checkpointer should see
1003 * the request when it does start, with or without getting a signal.
1004 */
1005#define MAX_SIGNAL_TRIES 600 /* max wait 60.0 sec */
1006 for (ntries = 0;; ntries++)
1007 {
1008 if (CheckpointerShmem->checkpointer_pid == 0)
1009 {
1010 if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT))
1011 {
1012 elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG,
1013 "could not signal for checkpoint: checkpointer is not running");
1014 break;
1015 }
1016 }
1017 else if (kill(CheckpointerShmem->checkpointer_pid, SIGINT) != 0)
1018 {
1019 if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT))
1020 {
1021 elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG,
1022 "could not signal for checkpoint: %m");
1023 break;
1024 }
1025 }
1026 else
1027 break; /* signal sent successfully */
1028
1029 CHECK_FOR_INTERRUPTS();
1030 pg_usleep(100000L); /* wait 0.1 sec, then retry */
1031 }
1032
1033 /*
1034 * If requested, wait for completion. We detect completion according to
1035 * the algorithm given above.
1036 */
1037 if (flags & CHECKPOINT_WAIT)
1038 {
1039 int new_started,
1040 new_failed;
1041
1042 /* Wait for a new checkpoint to start. */
1043 ConditionVariablePrepareToSleep(&CheckpointerShmem->start_cv);
1044 for (;;)
1045 {
1046 SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1047 new_started = CheckpointerShmem->ckpt_started;
1048 SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1049
1050 if (new_started != old_started)
1051 break;
1052
1053 ConditionVariableSleep(&CheckpointerShmem->start_cv,
1054 WAIT_EVENT_CHECKPOINT_START);
1055 }
1056 ConditionVariableCancelSleep();
1057
1058 /*
1059 * We are waiting for ckpt_done >= new_started, in a modulo sense.
1060 */
1061 ConditionVariablePrepareToSleep(&CheckpointerShmem->done_cv);
1062 for (;;)
1063 {
1064 int new_done;
1065
1066 SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1067 new_done = CheckpointerShmem->ckpt_done;
1068 new_failed = CheckpointerShmem->ckpt_failed;
1069 SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1070
1071 if (new_done - new_started >= 0)
1072 break;
1073
1074 ConditionVariableSleep(&CheckpointerShmem->done_cv,
1075 WAIT_EVENT_CHECKPOINT_DONE);
1076 }
1077 ConditionVariableCancelSleep();
1078
1079 if (new_failed != old_failed)
1080 ereport(ERROR,
1081 (errmsg("checkpoint request failed"),
1082 errhint("Consult recent messages in the server log for details.")));
1083 }
1084}
1085
1086/*
1087 * ForwardSyncRequest
1088 * Forward a file-fsync request from a backend to the checkpointer
1089 *
1090 * Whenever a backend is compelled to write directly to a relation
1091 * (which should be seldom, if the background writer is getting its job done),
1092 * the backend calls this routine to pass over knowledge that the relation
1093 * is dirty and must be fsync'd before next checkpoint. We also use this
1094 * opportunity to count such writes for statistical purposes.
1095 *
1096 * To avoid holding the lock for longer than necessary, we normally write
1097 * to the requests[] queue without checking for duplicates. The checkpointer
1098 * will have to eliminate dups internally anyway. However, if we discover
1099 * that the queue is full, we make a pass over the entire queue to compact
1100 * it. This is somewhat expensive, but the alternative is for the backend
1101 * to perform its own fsync, which is far more expensive in practice. It
1102 * is theoretically possible a backend fsync might still be necessary, if
1103 * the queue is full and contains no duplicate entries. In that case, we
1104 * let the backend know by returning false.
1105 */
1106bool
1107ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)
1108{
1109 CheckpointerRequest *request;
1110 bool too_full;
1111
1112 if (!IsUnderPostmaster)
1113 return false; /* probably shouldn't even get here */
1114
1115 if (AmCheckpointerProcess())
1116 elog(ERROR, "ForwardSyncRequest must not be called in checkpointer");
1117
1118 LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1119
1120 /* Count all backend writes regardless of if they fit in the queue */
1121 if (!AmBackgroundWriterProcess())
1122 CheckpointerShmem->num_backend_writes++;
1123
1124 /*
1125 * If the checkpointer isn't running or the request queue is full, the
1126 * backend will have to perform its own fsync request. But before forcing
1127 * that to happen, we can try to compact the request queue.
1128 */
1129 if (CheckpointerShmem->checkpointer_pid == 0 ||
1130 (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&
1131 !CompactCheckpointerRequestQueue()))
1132 {
1133 /*
1134 * Count the subset of writes where backends have to do their own
1135 * fsync
1136 */
1137 if (!AmBackgroundWriterProcess())
1138 CheckpointerShmem->num_backend_fsync++;
1139 LWLockRelease(CheckpointerCommLock);
1140 return false;
1141 }
1142
1143 /* OK, insert request */
1144 request = &CheckpointerShmem->requests[CheckpointerShmem->num_requests++];
1145 request->ftag = *ftag;
1146 request->type = type;
1147
1148 /* If queue is more than half full, nudge the checkpointer to empty it */
1149 too_full = (CheckpointerShmem->num_requests >=
1150 CheckpointerShmem->max_requests / 2);
1151
1152 LWLockRelease(CheckpointerCommLock);
1153
1154 /* ... but not till after we release the lock */
1155 if (too_full && ProcGlobal->checkpointerLatch)
1156 SetLatch(ProcGlobal->checkpointerLatch);
1157
1158 return true;
1159}
1160
1161/*
1162 * CompactCheckpointerRequestQueue
1163 * Remove duplicates from the request queue to avoid backend fsyncs.
1164 * Returns "true" if any entries were removed.
1165 *
1166 * Although a full fsync request queue is not common, it can lead to severe
1167 * performance problems when it does happen. So far, this situation has
1168 * only been observed to occur when the system is under heavy write load,
1169 * and especially during the "sync" phase of a checkpoint. Without this
1170 * logic, each backend begins doing an fsync for every block written, which
1171 * gets very expensive and can slow down the whole system.
1172 *
1173 * Trying to do this every time the queue is full could lose if there
1174 * aren't any removable entries. But that should be vanishingly rare in
1175 * practice: there's one queue entry per shared buffer.
1176 */
1177static bool
1178CompactCheckpointerRequestQueue(void)
1179{
1180 struct CheckpointerSlotMapping
1181 {
1182 CheckpointerRequest request;
1183 int slot;
1184 };
1185
1186 int n,
1187 preserve_count;
1188 int num_skipped = 0;
1189 HASHCTL ctl;
1190 HTAB *htab;
1191 bool *skip_slot;
1192
1193 /* must hold CheckpointerCommLock in exclusive mode */
1194 Assert(LWLockHeldByMe(CheckpointerCommLock));
1195
1196 /* Initialize skip_slot array */
1197 skip_slot = palloc0(sizeof(bool) * CheckpointerShmem->num_requests);
1198
1199 /* Initialize temporary hash table */
1200 MemSet(&ctl, 0, sizeof(ctl));
1201 ctl.keysize = sizeof(CheckpointerRequest);
1202 ctl.entrysize = sizeof(struct CheckpointerSlotMapping);
1203 ctl.hcxt = CurrentMemoryContext;
1204
1205 htab = hash_create("CompactCheckpointerRequestQueue",
1206 CheckpointerShmem->num_requests,
1207 &ctl,
1208 HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1209
1210 /*
1211 * The basic idea here is that a request can be skipped if it's followed
1212 * by a later, identical request. It might seem more sensible to work
1213 * backwards from the end of the queue and check whether a request is
1214 * *preceded* by an earlier, identical request, in the hopes of doing less
1215 * copying. But that might change the semantics, if there's an
1216 * intervening SYNC_FORGET_REQUEST or SYNC_FILTER_REQUEST, so we do it
1217 * this way. It would be possible to be even smarter if we made the code
1218 * below understand the specific semantics of such requests (it could blow
1219 * away preceding entries that would end up being canceled anyhow), but
1220 * it's not clear that the extra complexity would buy us anything.
1221 */
1222 for (n = 0; n < CheckpointerShmem->num_requests; n++)
1223 {
1224 CheckpointerRequest *request;
1225 struct CheckpointerSlotMapping *slotmap;
1226 bool found;
1227
1228 /*
1229 * We use the request struct directly as a hashtable key. This
1230 * assumes that any padding bytes in the structs are consistently the
1231 * same, which should be okay because we zeroed them in
1232 * CheckpointerShmemInit. Note also that RelFileNode had better
1233 * contain no pad bytes.
1234 */
1235 request = &CheckpointerShmem->requests[n];
1236 slotmap = hash_search(htab, request, HASH_ENTER, &found);
1237 if (found)
1238 {
1239 /* Duplicate, so mark the previous occurrence as skippable */
1240 skip_slot[slotmap->slot] = true;
1241 num_skipped++;
1242 }
1243 /* Remember slot containing latest occurrence of this request value */
1244 slotmap->slot = n;
1245 }
1246
1247 /* Done with the hash table. */
1248 hash_destroy(htab);
1249
1250 /* If no duplicates, we're out of luck. */
1251 if (!num_skipped)
1252 {
1253 pfree(skip_slot);
1254 return false;
1255 }
1256
1257 /* We found some duplicates; remove them. */
1258 preserve_count = 0;
1259 for (n = 0; n < CheckpointerShmem->num_requests; n++)
1260 {
1261 if (skip_slot[n])
1262 continue;
1263 CheckpointerShmem->requests[preserve_count++] = CheckpointerShmem->requests[n];
1264 }
1265 ereport(DEBUG1,
1266 (errmsg("compacted fsync request queue from %d entries to %d entries",
1267 CheckpointerShmem->num_requests, preserve_count)));
1268 CheckpointerShmem->num_requests = preserve_count;
1269
1270 /* Cleanup. */
1271 pfree(skip_slot);
1272 return true;
1273}
1274
1275/*
1276 * AbsorbSyncRequests
1277 * Retrieve queued sync requests and pass them to sync mechanism.
1278 *
1279 * This is exported because it must be called during CreateCheckPoint;
1280 * we have to be sure we have accepted all pending requests just before
1281 * we start fsync'ing. Since CreateCheckPoint sometimes runs in
1282 * non-checkpointer processes, do nothing if not checkpointer.
1283 */
1284void
1285AbsorbSyncRequests(void)
1286{
1287 CheckpointerRequest *requests = NULL;
1288 CheckpointerRequest *request;
1289 int n;
1290
1291 if (!AmCheckpointerProcess())
1292 return;
1293
1294 LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1295
1296 /* Transfer stats counts into pending pgstats message */
1297 BgWriterStats.m_buf_written_backend += CheckpointerShmem->num_backend_writes;
1298 BgWriterStats.m_buf_fsync_backend += CheckpointerShmem->num_backend_fsync;
1299
1300 CheckpointerShmem->num_backend_writes = 0;
1301 CheckpointerShmem->num_backend_fsync = 0;
1302
1303 /*
1304 * We try to avoid holding the lock for a long time by copying the request
1305 * array, and processing the requests after releasing the lock.
1306 *
1307 * Once we have cleared the requests from shared memory, we have to PANIC
1308 * if we then fail to absorb them (eg, because our hashtable runs out of
1309 * memory). This is because the system cannot run safely if we are unable
1310 * to fsync what we have been told to fsync. Fortunately, the hashtable
1311 * is so small that the problem is quite unlikely to arise in practice.
1312 */
1313 n = CheckpointerShmem->num_requests;
1314 if (n > 0)
1315 {
1316 requests = (CheckpointerRequest *) palloc(n * sizeof(CheckpointerRequest));
1317 memcpy(requests, CheckpointerShmem->requests, n * sizeof(CheckpointerRequest));
1318 }
1319
1320 START_CRIT_SECTION();
1321
1322 CheckpointerShmem->num_requests = 0;
1323
1324 LWLockRelease(CheckpointerCommLock);
1325
1326 for (request = requests; n > 0; request++, n--)
1327 RememberSyncRequest(&request->ftag, request->type);
1328
1329 END_CRIT_SECTION();
1330
1331 if (requests)
1332 pfree(requests);
1333}
1334
1335/*
1336 * Update any shared memory configurations based on config parameters
1337 */
1338static void
1339UpdateSharedMemoryConfig(void)
1340{
1341 /* update global shmem state for sync rep */
1342 SyncRepUpdateSyncStandbysDefined();
1343
1344 /*
1345 * If full_page_writes has been changed by SIGHUP, we update it in shared
1346 * memory and write an XLOG_FPW_CHANGE record.
1347 */
1348 UpdateFullPageWrites();
1349
1350 elog(DEBUG2, "checkpointer updated shared memory configuration values");
1351}
1352
1353/*
1354 * FirstCallSinceLastCheckpoint allows a process to take an action once
1355 * per checkpoint cycle by asynchronously checking for checkpoint completion.
1356 */
1357bool
1358FirstCallSinceLastCheckpoint(void)
1359{
1360 static int ckpt_done = 0;
1361 int new_done;
1362 bool FirstCall = false;
1363
1364 SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1365 new_done = CheckpointerShmem->ckpt_done;
1366 SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1367
1368 if (new_done != ckpt_done)
1369 FirstCall = true;
1370
1371 ckpt_done = new_done;
1372
1373 return FirstCall;
1374}
1375