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) 1996-2025, PostgreSQL Global Development Group
30 : *
31 : *
32 : * IDENTIFICATION
33 : * src/backend/postmaster/checkpointer.c
34 : *
35 : *-------------------------------------------------------------------------
36 : */
37 : #include "postgres.h"
38 :
39 : #include <sys/time.h>
40 : #include <time.h>
41 :
42 : #include "access/xlog.h"
43 : #include "access/xlog_internal.h"
44 : #include "access/xlogrecovery.h"
45 : #include "catalog/pg_authid.h"
46 : #include "commands/defrem.h"
47 : #include "libpq/pqsignal.h"
48 : #include "miscadmin.h"
49 : #include "pgstat.h"
50 : #include "postmaster/auxprocess.h"
51 : #include "postmaster/bgwriter.h"
52 : #include "postmaster/interrupt.h"
53 : #include "replication/syncrep.h"
54 : #include "storage/aio_subsys.h"
55 : #include "storage/bufmgr.h"
56 : #include "storage/condition_variable.h"
57 : #include "storage/fd.h"
58 : #include "storage/ipc.h"
59 : #include "storage/lwlock.h"
60 : #include "storage/pmsignal.h"
61 : #include "storage/proc.h"
62 : #include "storage/procsignal.h"
63 : #include "storage/shmem.h"
64 : #include "storage/smgr.h"
65 : #include "storage/spin.h"
66 : #include "utils/acl.h"
67 : #include "utils/guc.h"
68 : #include "utils/memutils.h"
69 : #include "utils/resowner.h"
70 :
71 :
72 : /*----------
73 : * Shared memory area for communication between checkpointer and backends
74 : *
75 : * The ckpt counters allow backends to watch for completion of a checkpoint
76 : * request they send. Here's how it works:
77 : * * At start of a checkpoint, checkpointer reads (and clears) the request
78 : * flags and increments ckpt_started, while holding ckpt_lck.
79 : * * On completion of a checkpoint, checkpointer sets ckpt_done to
80 : * equal ckpt_started.
81 : * * On failure of a checkpoint, checkpointer increments ckpt_failed
82 : * and sets ckpt_done to equal ckpt_started.
83 : *
84 : * The algorithm for backends is:
85 : * 1. Record current values of ckpt_failed and ckpt_started, and
86 : * set request flags, while holding ckpt_lck.
87 : * 2. Send signal to request checkpoint.
88 : * 3. Sleep until ckpt_started changes. Now you know a checkpoint has
89 : * begun since you started this algorithm (although *not* that it was
90 : * specifically initiated by your signal), and that it is using your flags.
91 : * 4. Record new value of ckpt_started.
92 : * 5. Sleep until ckpt_done >= saved value of ckpt_started. (Use modulo
93 : * arithmetic here in case counters wrap around.) Now you know a
94 : * checkpoint has started and completed, but not whether it was
95 : * successful.
96 : * 6. If ckpt_failed is different from the originally saved value,
97 : * assume request failed; otherwise it was definitely successful.
98 : *
99 : * ckpt_flags holds the OR of the checkpoint request flags sent by all
100 : * requesting backends since the last checkpoint start. The flags are
101 : * chosen so that OR'ing is the correct way to combine multiple requests.
102 : *
103 : * The requests array holds fsync requests sent by backends and not yet
104 : * absorbed by the checkpointer.
105 : *
106 : * Unlike the checkpoint fields, requests related fields are protected by
107 : * CheckpointerCommLock.
108 : *----------
109 : */
110 : typedef struct
111 : {
112 : SyncRequestType type; /* request type */
113 : FileTag ftag; /* file identifier */
114 : } CheckpointerRequest;
115 :
116 : typedef struct
117 : {
118 : pid_t checkpointer_pid; /* PID (0 if not started) */
119 :
120 : slock_t ckpt_lck; /* protects all the ckpt_* fields */
121 :
122 : int ckpt_started; /* advances when checkpoint starts */
123 : int ckpt_done; /* advances when checkpoint done */
124 : int ckpt_failed; /* advances when checkpoint fails */
125 :
126 : int ckpt_flags; /* checkpoint flags, as defined in xlog.h */
127 :
128 : ConditionVariable start_cv; /* signaled when ckpt_started advances */
129 : ConditionVariable done_cv; /* signaled when ckpt_done advances */
130 :
131 : int num_requests; /* current # of requests */
132 : int max_requests; /* allocated array size */
133 :
134 : int head; /* Index of the first request in the ring
135 : * buffer */
136 : int tail; /* Index of the last request in the ring
137 : * buffer */
138 :
139 : /* The ring buffer of pending checkpointer requests */
140 : CheckpointerRequest requests[FLEXIBLE_ARRAY_MEMBER];
141 : } CheckpointerShmemStruct;
142 :
143 : static CheckpointerShmemStruct *CheckpointerShmem;
144 :
145 : /* interval for calling AbsorbSyncRequests in CheckpointWriteDelay */
146 : #define WRITES_PER_ABSORB 1000
147 :
148 : /* Maximum number of checkpointer requests to process in one batch */
149 : #define CKPT_REQ_BATCH_SIZE 10000
150 :
151 : /* Max number of requests the checkpointer request queue can hold */
152 : #define MAX_CHECKPOINT_REQUESTS 10000000
153 :
154 : /*
155 : * GUC parameters
156 : */
157 : int CheckPointTimeout = 300;
158 : int CheckPointWarning = 30;
159 : double CheckPointCompletionTarget = 0.9;
160 :
161 : /*
162 : * Private state
163 : */
164 : static bool ckpt_active = false;
165 : static volatile sig_atomic_t ShutdownXLOGPending = false;
166 :
167 : /* these values are valid when ckpt_active is true: */
168 : static pg_time_t ckpt_start_time;
169 : static XLogRecPtr ckpt_start_recptr;
170 : static double ckpt_cached_elapsed;
171 :
172 : static pg_time_t last_checkpoint_time;
173 : static pg_time_t last_xlog_switch_time;
174 :
175 : /* Prototypes for private functions */
176 :
177 : static void ProcessCheckpointerInterrupts(void);
178 : static void CheckArchiveTimeout(void);
179 : static bool IsCheckpointOnSchedule(double progress);
180 : static bool FastCheckpointRequested(void);
181 : static bool CompactCheckpointerRequestQueue(void);
182 : static void UpdateSharedMemoryConfig(void);
183 :
184 : /* Signal handlers */
185 : static void ReqShutdownXLOG(SIGNAL_ARGS);
186 :
187 :
188 : /*
189 : * Main entry point for checkpointer process
190 : *
191 : * This is invoked from AuxiliaryProcessMain, which has already created the
192 : * basic execution environment, but not enabled signals yet.
193 : */
194 : void
195 1038 : CheckpointerMain(const void *startup_data, size_t startup_data_len)
196 : {
197 : sigjmp_buf local_sigjmp_buf;
198 : MemoryContext checkpointer_context;
199 :
200 : Assert(startup_data_len == 0);
201 :
202 1038 : MyBackendType = B_CHECKPOINTER;
203 1038 : AuxiliaryProcessMainCommon();
204 :
205 1038 : CheckpointerShmem->checkpointer_pid = MyProcPid;
206 :
207 : /*
208 : * Properly accept or ignore signals the postmaster might send us
209 : *
210 : * Note: we deliberately ignore SIGTERM, because during a standard Unix
211 : * system shutdown cycle, init will SIGTERM all processes at once. We
212 : * want to wait for the backends to exit, whereupon the postmaster will
213 : * tell us it's okay to shut down (via SIGUSR2).
214 : */
215 1038 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
216 1038 : pqsignal(SIGINT, ReqShutdownXLOG);
217 1038 : pqsignal(SIGTERM, SIG_IGN); /* ignore SIGTERM */
218 : /* SIGQUIT handler was already set up by InitPostmasterChild */
219 1038 : pqsignal(SIGALRM, SIG_IGN);
220 1038 : pqsignal(SIGPIPE, SIG_IGN);
221 1038 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
222 1038 : pqsignal(SIGUSR2, SignalHandlerForShutdownRequest);
223 :
224 : /*
225 : * Reset some signals that are accepted by postmaster but not here
226 : */
227 1038 : pqsignal(SIGCHLD, SIG_DFL);
228 :
229 : /*
230 : * Initialize so that first time-driven event happens at the correct time.
231 : */
232 1038 : last_checkpoint_time = last_xlog_switch_time = (pg_time_t) time(NULL);
233 :
234 : /*
235 : * Write out stats after shutdown. This needs to be called by exactly one
236 : * process during a normal shutdown, and since checkpointer is shut down
237 : * very late...
238 : *
239 : * While e.g. walsenders are active after the shutdown checkpoint has been
240 : * written (and thus could produce more stats), checkpointer stays around
241 : * after the shutdown checkpoint has been written. postmaster will only
242 : * signal checkpointer to exit after all processes that could emit stats
243 : * have been shut down.
244 : */
245 1038 : before_shmem_exit(pgstat_before_server_shutdown, 0);
246 :
247 : /*
248 : * Create a memory context that we will do all our work in. We do this so
249 : * that we can reset the context during error recovery and thereby avoid
250 : * possible memory leaks. Formerly this code just ran in
251 : * TopMemoryContext, but resetting that would be a really bad idea.
252 : */
253 1038 : checkpointer_context = AllocSetContextCreate(TopMemoryContext,
254 : "Checkpointer",
255 : ALLOCSET_DEFAULT_SIZES);
256 1038 : MemoryContextSwitchTo(checkpointer_context);
257 :
258 : /*
259 : * If an exception is encountered, processing resumes here.
260 : *
261 : * You might wonder why this isn't coded as an infinite loop around a
262 : * PG_TRY construct. The reason is that this is the bottom of the
263 : * exception stack, and so with PG_TRY there would be no exception handler
264 : * in force at all during the CATCH part. By leaving the outermost setjmp
265 : * always active, we have at least some chance of recovering from an error
266 : * during error recovery. (If we get into an infinite loop thereby, it
267 : * will soon be stopped by overflow of elog.c's internal state stack.)
268 : *
269 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
270 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
271 : * signals other than SIGQUIT will be blocked until we complete error
272 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
273 : * call redundant, but it is not since InterruptPending might be set
274 : * already.
275 : */
276 1038 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
277 : {
278 : /* Since not using PG_TRY, must reset error stack by hand */
279 0 : error_context_stack = NULL;
280 :
281 : /* Prevent interrupts while cleaning up */
282 0 : HOLD_INTERRUPTS();
283 :
284 : /* Report the error to the server log */
285 0 : EmitErrorReport();
286 :
287 : /*
288 : * These operations are really just a minimal subset of
289 : * AbortTransaction(). We don't have very many resources to worry
290 : * about in checkpointer, but we do have LWLocks, buffers, and temp
291 : * files.
292 : */
293 0 : LWLockReleaseAll();
294 0 : ConditionVariableCancelSleep();
295 0 : pgstat_report_wait_end();
296 0 : pgaio_error_cleanup();
297 0 : UnlockBuffers();
298 0 : ReleaseAuxProcessResources(false);
299 0 : AtEOXact_Buffers(false);
300 0 : AtEOXact_SMgr();
301 0 : AtEOXact_Files(false);
302 0 : AtEOXact_HashTables(false);
303 :
304 : /* Warn any waiting backends that the checkpoint failed. */
305 0 : if (ckpt_active)
306 : {
307 0 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
308 0 : CheckpointerShmem->ckpt_failed++;
309 0 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
310 0 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
311 :
312 0 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
313 :
314 0 : ckpt_active = false;
315 : }
316 :
317 : /*
318 : * Now return to normal top-level context and clear ErrorContext for
319 : * next time.
320 : */
321 0 : MemoryContextSwitchTo(checkpointer_context);
322 0 : FlushErrorState();
323 :
324 : /* Flush any leaked data in the top-level context */
325 0 : MemoryContextReset(checkpointer_context);
326 :
327 : /* Now we can allow interrupts again */
328 0 : RESUME_INTERRUPTS();
329 :
330 : /*
331 : * Sleep at least 1 second after any error. A write error is likely
332 : * to be repeated, and we don't want to be filling the error logs as
333 : * fast as we can.
334 : */
335 0 : pg_usleep(1000000L);
336 : }
337 :
338 : /* We can now handle ereport(ERROR) */
339 1038 : PG_exception_stack = &local_sigjmp_buf;
340 :
341 : /*
342 : * Unblock signals (they were blocked when the postmaster forked us)
343 : */
344 1038 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
345 :
346 : /*
347 : * Ensure all shared memory values are set correctly for the config. Doing
348 : * this here ensures no race conditions from other concurrent updaters.
349 : */
350 1038 : UpdateSharedMemoryConfig();
351 :
352 : /*
353 : * Advertise our proc number that backends can use to wake us up while
354 : * we're sleeping.
355 : */
356 1038 : ProcGlobal->checkpointerProc = MyProcNumber;
357 :
358 : /*
359 : * Loop until we've been asked to write the shutdown checkpoint or
360 : * terminate.
361 : */
362 : for (;;)
363 7262 : {
364 8300 : bool do_checkpoint = false;
365 8300 : int flags = 0;
366 : pg_time_t now;
367 : int elapsed_secs;
368 : int cur_timeout;
369 8300 : bool chkpt_or_rstpt_requested = false;
370 8300 : bool chkpt_or_rstpt_timed = false;
371 :
372 : /* Clear any already-pending wakeups */
373 8300 : ResetLatch(MyLatch);
374 :
375 : /*
376 : * Process any requests or signals received recently.
377 : */
378 8300 : AbsorbSyncRequests();
379 :
380 8300 : ProcessCheckpointerInterrupts();
381 8300 : if (ShutdownXLOGPending || ShutdownRequestPending)
382 : break;
383 :
384 : /*
385 : * Detect a pending checkpoint request by checking whether the flags
386 : * word in shared memory is nonzero. We shouldn't need to acquire the
387 : * ckpt_lck for this.
388 : */
389 7282 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
390 : {
391 2542 : do_checkpoint = true;
392 2542 : chkpt_or_rstpt_requested = true;
393 : }
394 :
395 : /*
396 : * Force a checkpoint if too much time has elapsed since the last one.
397 : * Note that we count a timed checkpoint in stats only when this
398 : * occurs without an external request, but we set the CAUSE_TIME flag
399 : * bit even if there is also an external request.
400 : */
401 7282 : now = (pg_time_t) time(NULL);
402 7282 : elapsed_secs = now - last_checkpoint_time;
403 7282 : if (elapsed_secs >= CheckPointTimeout)
404 : {
405 2 : if (!do_checkpoint)
406 2 : chkpt_or_rstpt_timed = true;
407 2 : do_checkpoint = true;
408 2 : flags |= CHECKPOINT_CAUSE_TIME;
409 : }
410 :
411 : /*
412 : * Do a checkpoint if requested.
413 : */
414 7282 : if (do_checkpoint)
415 : {
416 2544 : bool ckpt_performed = false;
417 : bool do_restartpoint;
418 :
419 : /* Check if we should perform a checkpoint or a restartpoint. */
420 2544 : do_restartpoint = RecoveryInProgress();
421 :
422 : /*
423 : * Atomically fetch the request flags to figure out what kind of a
424 : * checkpoint we should perform, and increase the started-counter
425 : * to acknowledge that we've started a new checkpoint.
426 : */
427 2544 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
428 2544 : flags |= CheckpointerShmem->ckpt_flags;
429 2544 : CheckpointerShmem->ckpt_flags = 0;
430 2544 : CheckpointerShmem->ckpt_started++;
431 2544 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
432 :
433 2544 : ConditionVariableBroadcast(&CheckpointerShmem->start_cv);
434 :
435 : /*
436 : * The end-of-recovery checkpoint is a real checkpoint that's
437 : * performed while we're still in recovery.
438 : */
439 2544 : if (flags & CHECKPOINT_END_OF_RECOVERY)
440 40 : do_restartpoint = false;
441 :
442 2544 : if (chkpt_or_rstpt_timed)
443 : {
444 2 : chkpt_or_rstpt_timed = false;
445 2 : if (do_restartpoint)
446 0 : PendingCheckpointerStats.restartpoints_timed++;
447 : else
448 2 : PendingCheckpointerStats.num_timed++;
449 : }
450 :
451 2544 : if (chkpt_or_rstpt_requested)
452 : {
453 2542 : chkpt_or_rstpt_requested = false;
454 2542 : if (do_restartpoint)
455 1086 : PendingCheckpointerStats.restartpoints_requested++;
456 : else
457 1456 : PendingCheckpointerStats.num_requested++;
458 : }
459 :
460 : /*
461 : * We will warn if (a) too soon since last checkpoint (whatever
462 : * caused it) and (b) somebody set the CHECKPOINT_CAUSE_XLOG flag
463 : * since the last checkpoint start. Note in particular that this
464 : * implementation will not generate warnings caused by
465 : * CheckPointTimeout < CheckPointWarning.
466 : */
467 2544 : if (!do_restartpoint &&
468 1458 : (flags & CHECKPOINT_CAUSE_XLOG) &&
469 386 : elapsed_secs < CheckPointWarning)
470 386 : ereport(LOG,
471 : (errmsg_plural("checkpoints are occurring too frequently (%d second apart)",
472 : "checkpoints are occurring too frequently (%d seconds apart)",
473 : elapsed_secs,
474 : elapsed_secs),
475 : errhint("Consider increasing the configuration parameter \"%s\".", "max_wal_size")));
476 :
477 : /*
478 : * Initialize checkpointer-private variables used during
479 : * checkpoint.
480 : */
481 2544 : ckpt_active = true;
482 2544 : if (do_restartpoint)
483 1086 : ckpt_start_recptr = GetXLogReplayRecPtr(NULL);
484 : else
485 1458 : ckpt_start_recptr = GetInsertRecPtr();
486 2544 : ckpt_start_time = now;
487 2544 : ckpt_cached_elapsed = 0;
488 :
489 : /*
490 : * Do the checkpoint.
491 : */
492 2544 : if (!do_restartpoint)
493 1458 : ckpt_performed = CreateCheckPoint(flags);
494 : else
495 1086 : ckpt_performed = CreateRestartPoint(flags);
496 :
497 : /*
498 : * After any checkpoint, free all smgr objects. Otherwise we
499 : * would never do so for dropped relations, as the checkpointer
500 : * does not process shared invalidation messages or call
501 : * AtEOXact_SMgr().
502 : */
503 2544 : smgrdestroyall();
504 :
505 : /*
506 : * Indicate checkpoint completion to any waiting backends.
507 : */
508 2544 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
509 2544 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
510 2544 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
511 :
512 2544 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
513 :
514 2544 : if (!do_restartpoint)
515 : {
516 : /*
517 : * Note we record the checkpoint start time not end time as
518 : * last_checkpoint_time. This is so that time-driven
519 : * checkpoints happen at a predictable spacing.
520 : */
521 1458 : last_checkpoint_time = now;
522 :
523 1458 : if (ckpt_performed)
524 1452 : PendingCheckpointerStats.num_performed++;
525 : }
526 : else
527 : {
528 1086 : if (ckpt_performed)
529 : {
530 : /*
531 : * The same as for checkpoint. Please see the
532 : * corresponding comment.
533 : */
534 346 : last_checkpoint_time = now;
535 :
536 346 : PendingCheckpointerStats.restartpoints_performed++;
537 : }
538 : else
539 : {
540 : /*
541 : * We were not able to perform the restartpoint
542 : * (checkpoints throw an ERROR in case of error). Most
543 : * likely because we have not received any new checkpoint
544 : * WAL records since the last restartpoint. Try again in
545 : * 15 s.
546 : */
547 740 : last_checkpoint_time = now - CheckPointTimeout + 15;
548 : }
549 : }
550 :
551 2544 : ckpt_active = false;
552 :
553 : /*
554 : * We may have received an interrupt during the checkpoint and the
555 : * latch might have been reset (e.g. in CheckpointWriteDelay).
556 : */
557 2544 : ProcessCheckpointerInterrupts();
558 2544 : if (ShutdownXLOGPending || ShutdownRequestPending)
559 : break;
560 : }
561 :
562 : /* Check for archive_timeout and switch xlog files if necessary. */
563 7270 : CheckArchiveTimeout();
564 :
565 : /* Report pending statistics to the cumulative stats system */
566 7270 : pgstat_report_checkpointer();
567 7270 : pgstat_report_wal(true);
568 :
569 : /*
570 : * If any checkpoint flags have been set, redo the loop to handle the
571 : * checkpoint without sleeping.
572 : */
573 7270 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
574 434 : continue;
575 :
576 : /*
577 : * Sleep until we are signaled or it's time for another checkpoint or
578 : * xlog file switch.
579 : */
580 6836 : now = (pg_time_t) time(NULL);
581 6836 : elapsed_secs = now - last_checkpoint_time;
582 6836 : if (elapsed_secs >= CheckPointTimeout)
583 0 : continue; /* no sleep for us ... */
584 6836 : cur_timeout = CheckPointTimeout - elapsed_secs;
585 6836 : if (XLogArchiveTimeout > 0 && !RecoveryInProgress())
586 : {
587 0 : elapsed_secs = now - last_xlog_switch_time;
588 0 : if (elapsed_secs >= XLogArchiveTimeout)
589 0 : continue; /* no sleep for us ... */
590 0 : cur_timeout = Min(cur_timeout, XLogArchiveTimeout - elapsed_secs);
591 : }
592 :
593 6836 : (void) WaitLatch(MyLatch,
594 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
595 : cur_timeout * 1000L /* convert to ms */ ,
596 : WAIT_EVENT_CHECKPOINTER_MAIN);
597 : }
598 :
599 : /*
600 : * From here on, elog(ERROR) should end with exit(1), not send control
601 : * back to the sigsetjmp block above.
602 : */
603 1030 : ExitOnAnyError = true;
604 :
605 1030 : if (ShutdownXLOGPending)
606 : {
607 : /*
608 : * Close down the database.
609 : *
610 : * Since ShutdownXLOG() creates restartpoint or checkpoint, and
611 : * updates the statistics, increment the checkpoint request and flush
612 : * out pending statistic.
613 : */
614 1030 : PendingCheckpointerStats.num_requested++;
615 1030 : ShutdownXLOG(0, 0);
616 1030 : pgstat_report_checkpointer();
617 1030 : pgstat_report_wal(true);
618 :
619 : /*
620 : * Tell postmaster that we're done.
621 : */
622 1030 : SendPostmasterSignal(PMSIGNAL_XLOG_IS_SHUTDOWN);
623 1030 : ShutdownXLOGPending = false;
624 : }
625 :
626 : /*
627 : * Wait until we're asked to shut down. By separating the writing of the
628 : * shutdown checkpoint from checkpointer exiting, checkpointer can perform
629 : * some should-be-as-late-as-possible work like writing out stats.
630 : */
631 : for (;;)
632 : {
633 : /* Clear any already-pending wakeups */
634 2060 : ResetLatch(MyLatch);
635 :
636 2060 : ProcessCheckpointerInterrupts();
637 :
638 2060 : if (ShutdownRequestPending)
639 1030 : break;
640 :
641 1030 : (void) WaitLatch(MyLatch,
642 : WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
643 : 0,
644 : WAIT_EVENT_CHECKPOINTER_SHUTDOWN);
645 : }
646 :
647 : /* Normal exit from the checkpointer is here */
648 1030 : proc_exit(0); /* done */
649 : }
650 :
651 : /*
652 : * Process any new interrupts.
653 : */
654 : static void
655 12904 : ProcessCheckpointerInterrupts(void)
656 : {
657 12904 : if (ProcSignalBarrierPending)
658 118 : ProcessProcSignalBarrier();
659 :
660 12904 : if (ConfigReloadPending)
661 : {
662 118 : ConfigReloadPending = false;
663 118 : ProcessConfigFile(PGC_SIGHUP);
664 :
665 : /*
666 : * Checkpointer is the last process to shut down, so we ask it to hold
667 : * the keys for a range of other tasks required most of which have
668 : * nothing to do with checkpointing at all.
669 : *
670 : * For various reasons, some config values can change dynamically so
671 : * the primary copy of them is held in shared memory to make sure all
672 : * backends see the same value. We make Checkpointer responsible for
673 : * updating the shared memory copy if the parameter setting changes
674 : * because of SIGHUP.
675 : */
676 118 : UpdateSharedMemoryConfig();
677 : }
678 :
679 : /* Perform logging of memory contexts of this process */
680 12904 : if (LogMemoryContextPending)
681 2 : ProcessLogMemoryContextInterrupt();
682 12904 : }
683 :
684 : /*
685 : * CheckArchiveTimeout -- check for archive_timeout and switch xlog files
686 : *
687 : * This will switch to a new WAL file and force an archive file write if
688 : * meaningful activity is recorded in the current WAL file. This includes most
689 : * writes, including just a single checkpoint record, but excludes WAL records
690 : * that were inserted with the XLOG_MARK_UNIMPORTANT flag being set (like
691 : * snapshots of running transactions). Such records, depending on
692 : * configuration, occur on regular intervals and don't contain important
693 : * information. This avoids generating archives with a few unimportant
694 : * records.
695 : */
696 : static void
697 23408 : CheckArchiveTimeout(void)
698 : {
699 : pg_time_t now;
700 : pg_time_t last_time;
701 : XLogRecPtr last_switch_lsn;
702 :
703 23408 : if (XLogArchiveTimeout <= 0 || RecoveryInProgress())
704 23408 : return;
705 :
706 0 : now = (pg_time_t) time(NULL);
707 :
708 : /* First we do a quick check using possibly-stale local state. */
709 0 : if ((int) (now - last_xlog_switch_time) < XLogArchiveTimeout)
710 0 : return;
711 :
712 : /*
713 : * Update local state ... note that last_xlog_switch_time is the last time
714 : * a switch was performed *or requested*.
715 : */
716 0 : last_time = GetLastSegSwitchData(&last_switch_lsn);
717 :
718 0 : last_xlog_switch_time = Max(last_xlog_switch_time, last_time);
719 :
720 : /* Now we can do the real checks */
721 0 : if ((int) (now - last_xlog_switch_time) >= XLogArchiveTimeout)
722 : {
723 : /*
724 : * Switch segment only when "important" WAL has been logged since the
725 : * last segment switch (last_switch_lsn points to end of segment
726 : * switch occurred in).
727 : */
728 0 : if (GetLastImportantRecPtr() > last_switch_lsn)
729 : {
730 : XLogRecPtr switchpoint;
731 :
732 : /* mark switch as unimportant, avoids triggering checkpoints */
733 0 : switchpoint = RequestXLogSwitch(true);
734 :
735 : /*
736 : * If the returned pointer points exactly to a segment boundary,
737 : * assume nothing happened.
738 : */
739 0 : if (XLogSegmentOffset(switchpoint, wal_segment_size) != 0)
740 0 : elog(DEBUG1, "write-ahead log switch forced (\"archive_timeout\"=%d)",
741 : XLogArchiveTimeout);
742 : }
743 :
744 : /*
745 : * Update state in any case, so we don't retry constantly when the
746 : * system is idle.
747 : */
748 0 : last_xlog_switch_time = now;
749 : }
750 : }
751 :
752 : /*
753 : * Returns true if a fast checkpoint request is pending. (Note that this does
754 : * not check the *current* checkpoint's FAST flag, but whether there is one
755 : * pending behind it.)
756 : */
757 : static bool
758 95286 : FastCheckpointRequested(void)
759 : {
760 95286 : volatile CheckpointerShmemStruct *cps = CheckpointerShmem;
761 :
762 : /*
763 : * We don't need to acquire the ckpt_lck in this case because we're only
764 : * looking at a single flag bit.
765 : */
766 95286 : if (cps->ckpt_flags & CHECKPOINT_FAST)
767 7496 : return true;
768 87790 : return false;
769 : }
770 :
771 : /*
772 : * CheckpointWriteDelay -- control rate of checkpoint
773 : *
774 : * This function is called after each page write performed by BufferSync().
775 : * It is responsible for throttling BufferSync()'s write rate to hit
776 : * checkpoint_completion_target.
777 : *
778 : * The checkpoint request flags should be passed in; currently the only one
779 : * examined is CHECKPOINT_FAST, which disables delays between writes.
780 : *
781 : * 'progress' is an estimate of how much of the work has been done, as a
782 : * fraction between 0.0 meaning none, and 1.0 meaning all done.
783 : */
784 : void
785 580380 : CheckpointWriteDelay(int flags, double progress)
786 : {
787 : static int absorb_counter = WRITES_PER_ABSORB;
788 :
789 : /* Do nothing if checkpoint is being executed by non-checkpointer process */
790 580380 : if (!AmCheckpointerProcess())
791 101094 : return;
792 :
793 : /*
794 : * Perform the usual duties and take a nap, unless we're behind schedule,
795 : * in which case we just try to catch up as quickly as possible.
796 : */
797 479286 : if (!(flags & CHECKPOINT_FAST) &&
798 95570 : !ShutdownXLOGPending &&
799 95286 : !ShutdownRequestPending &&
800 183076 : !FastCheckpointRequested() &&
801 87790 : IsCheckpointOnSchedule(progress))
802 : {
803 16138 : if (ConfigReloadPending)
804 : {
805 0 : ConfigReloadPending = false;
806 0 : ProcessConfigFile(PGC_SIGHUP);
807 : /* update shmem copies of config variables */
808 0 : UpdateSharedMemoryConfig();
809 : }
810 :
811 16138 : AbsorbSyncRequests();
812 16138 : absorb_counter = WRITES_PER_ABSORB;
813 :
814 16138 : CheckArchiveTimeout();
815 :
816 : /* Report interim statistics to the cumulative stats system */
817 16138 : pgstat_report_checkpointer();
818 :
819 : /*
820 : * This sleep used to be connected to bgwriter_delay, typically 200ms.
821 : * That resulted in more frequent wakeups if not much work to do.
822 : * Checkpointer and bgwriter are no longer related so take the Big
823 : * Sleep.
824 : */
825 16138 : WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | WL_TIMEOUT,
826 : 100,
827 : WAIT_EVENT_CHECKPOINT_WRITE_DELAY);
828 16138 : ResetLatch(MyLatch);
829 : }
830 463148 : else if (--absorb_counter <= 0)
831 : {
832 : /*
833 : * Absorb pending fsync requests after each WRITES_PER_ABSORB write
834 : * operations even when we don't sleep, to prevent overflow of the
835 : * fsync request queue.
836 : */
837 194 : AbsorbSyncRequests();
838 194 : absorb_counter = WRITES_PER_ABSORB;
839 : }
840 :
841 : /* Check for barrier events. */
842 479286 : if (ProcSignalBarrierPending)
843 10 : ProcessProcSignalBarrier();
844 : }
845 :
846 : /*
847 : * IsCheckpointOnSchedule -- are we on schedule to finish this checkpoint
848 : * (or restartpoint) in time?
849 : *
850 : * Compares the current progress against the time/segments elapsed since last
851 : * checkpoint, and returns true if the progress we've made this far is greater
852 : * than the elapsed time/segments.
853 : */
854 : static bool
855 87790 : IsCheckpointOnSchedule(double progress)
856 : {
857 : XLogRecPtr recptr;
858 : struct timeval now;
859 : double elapsed_xlogs,
860 : elapsed_time;
861 :
862 : Assert(ckpt_active);
863 :
864 : /* Scale progress according to checkpoint_completion_target. */
865 87790 : progress *= CheckPointCompletionTarget;
866 :
867 : /*
868 : * Check against the cached value first. Only do the more expensive
869 : * calculations once we reach the target previously calculated. Since
870 : * neither time or WAL insert pointer moves backwards, a freshly
871 : * calculated value can only be greater than or equal to the cached value.
872 : */
873 87790 : if (progress < ckpt_cached_elapsed)
874 64586 : return false;
875 :
876 : /*
877 : * Check progress against WAL segments written and CheckPointSegments.
878 : *
879 : * We compare the current WAL insert location against the location
880 : * computed before calling CreateCheckPoint. The code in XLogInsert that
881 : * actually triggers a checkpoint when CheckPointSegments is exceeded
882 : * compares against RedoRecPtr, so this is not completely accurate.
883 : * However, it's good enough for our purposes, we're only calculating an
884 : * estimate anyway.
885 : *
886 : * During recovery, we compare last replayed WAL record's location with
887 : * the location computed before calling CreateRestartPoint. That maintains
888 : * the same pacing as we have during checkpoints in normal operation, but
889 : * we might exceed max_wal_size by a fair amount. That's because there can
890 : * be a large gap between a checkpoint's redo-pointer and the checkpoint
891 : * record itself, and we only start the restartpoint after we've seen the
892 : * checkpoint record. (The gap is typically up to CheckPointSegments *
893 : * checkpoint_completion_target where checkpoint_completion_target is the
894 : * value that was in effect when the WAL was generated).
895 : */
896 23204 : if (RecoveryInProgress())
897 11216 : recptr = GetXLogReplayRecPtr(NULL);
898 : else
899 11988 : recptr = GetInsertRecPtr();
900 23204 : elapsed_xlogs = (((double) (recptr - ckpt_start_recptr)) /
901 23204 : wal_segment_size) / CheckPointSegments;
902 :
903 23204 : if (progress < elapsed_xlogs)
904 : {
905 7066 : ckpt_cached_elapsed = elapsed_xlogs;
906 7066 : return false;
907 : }
908 :
909 : /*
910 : * Check progress against time elapsed and checkpoint_timeout.
911 : */
912 16138 : gettimeofday(&now, NULL);
913 16138 : elapsed_time = ((double) ((pg_time_t) now.tv_sec - ckpt_start_time) +
914 16138 : now.tv_usec / 1000000.0) / CheckPointTimeout;
915 :
916 16138 : if (progress < elapsed_time)
917 : {
918 0 : ckpt_cached_elapsed = elapsed_time;
919 0 : return false;
920 : }
921 :
922 : /* It looks like we're on schedule. */
923 16138 : return true;
924 : }
925 :
926 :
927 : /* --------------------------------
928 : * signal handler routines
929 : * --------------------------------
930 : */
931 :
932 : /* SIGINT: set flag to trigger writing of shutdown checkpoint */
933 : static void
934 1032 : ReqShutdownXLOG(SIGNAL_ARGS)
935 : {
936 1032 : ShutdownXLOGPending = true;
937 1032 : SetLatch(MyLatch);
938 1032 : }
939 :
940 :
941 : /* --------------------------------
942 : * communication with backends
943 : * --------------------------------
944 : */
945 :
946 : /*
947 : * CheckpointerShmemSize
948 : * Compute space needed for checkpointer-related shared memory
949 : */
950 : Size
951 6078 : CheckpointerShmemSize(void)
952 : {
953 : Size size;
954 :
955 : /*
956 : * The size of the requests[] array is arbitrarily set equal to NBuffers.
957 : * But there is a cap of MAX_CHECKPOINT_REQUESTS to prevent accumulating
958 : * too many checkpoint requests in the ring buffer.
959 : */
960 6078 : size = offsetof(CheckpointerShmemStruct, requests);
961 6078 : size = add_size(size, mul_size(Min(NBuffers,
962 : MAX_CHECKPOINT_REQUESTS),
963 : sizeof(CheckpointerRequest)));
964 :
965 6078 : return size;
966 : }
967 :
968 : /*
969 : * CheckpointerShmemInit
970 : * Allocate and initialize checkpointer-related shared memory
971 : */
972 : void
973 2126 : CheckpointerShmemInit(void)
974 : {
975 2126 : Size size = CheckpointerShmemSize();
976 : bool found;
977 :
978 2126 : CheckpointerShmem = (CheckpointerShmemStruct *)
979 2126 : ShmemInitStruct("Checkpointer Data",
980 : size,
981 : &found);
982 :
983 2126 : if (!found)
984 : {
985 : /*
986 : * First time through, so initialize. Note that we zero the whole
987 : * requests array; this is so that CompactCheckpointerRequestQueue can
988 : * assume that any pad bytes in the request structs are zeroes.
989 : */
990 2414 : MemSet(CheckpointerShmem, 0, size);
991 2126 : SpinLockInit(&CheckpointerShmem->ckpt_lck);
992 2126 : CheckpointerShmem->max_requests = Min(NBuffers, MAX_CHECKPOINT_REQUESTS);
993 2126 : CheckpointerShmem->head = CheckpointerShmem->tail = 0;
994 2126 : ConditionVariableInit(&CheckpointerShmem->start_cv);
995 2126 : ConditionVariableInit(&CheckpointerShmem->done_cv);
996 : }
997 2126 : }
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 882 : ExecCheckpoint(ParseState *pstate, CheckPointStmt *stmt)
1007 : {
1008 882 : bool fast = true;
1009 882 : bool unlogged = false;
1010 :
1011 1770 : foreach_ptr(DefElem, opt, stmt->options)
1012 : {
1013 30 : if (strcmp(opt->defname, "mode") == 0)
1014 : {
1015 12 : char *mode = defGetString(opt);
1016 :
1017 12 : if (strcmp(mode, "spread") == 0)
1018 0 : fast = false;
1019 12 : else if (strcmp(mode, "fast") != 0)
1020 6 : ereport(ERROR,
1021 : (errcode(ERRCODE_SYNTAX_ERROR),
1022 : errmsg("unrecognized MODE option \"%s\"", mode),
1023 : parser_errposition(pstate, opt->location)));
1024 : }
1025 18 : else if (strcmp(opt->defname, "flush_unlogged") == 0)
1026 12 : unlogged = defGetBoolean(opt);
1027 : else
1028 6 : ereport(ERROR,
1029 : (errcode(ERRCODE_SYNTAX_ERROR),
1030 : errmsg("unrecognized CHECKPOINT option \"%s\"", opt->defname),
1031 : parser_errposition(pstate, opt->location)));
1032 : }
1033 :
1034 870 : if (!has_privs_of_role(GetUserId(), ROLE_PG_CHECKPOINT))
1035 0 : ereport(ERROR,
1036 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
1037 : /* translator: %s is name of an SQL command (e.g., CHECKPOINT) */
1038 : errmsg("permission denied to execute %s command",
1039 : "CHECKPOINT"),
1040 : errdetail("Only roles with privileges of the \"%s\" role may execute this command.",
1041 : "pg_checkpoint")));
1042 :
1043 1740 : RequestCheckpoint(CHECKPOINT_WAIT |
1044 870 : (fast ? CHECKPOINT_FAST : 0) |
1045 870 : (unlogged ? CHECKPOINT_FLUSH_UNLOGGED : 0) |
1046 870 : (RecoveryInProgress() ? 0 : CHECKPOINT_FORCE));
1047 870 : }
1048 :
1049 : /*
1050 : * RequestCheckpoint
1051 : * Called in backend processes to request a checkpoint
1052 : *
1053 : * flags is a bitwise OR of the following:
1054 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
1055 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
1056 : * CHECKPOINT_FAST: finish the checkpoint ASAP,
1057 : * ignoring checkpoint_completion_target parameter.
1058 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
1059 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
1060 : * CHECKPOINT_END_OF_RECOVERY, and the CHECKPOINT command).
1061 : * CHECKPOINT_WAIT: wait for completion before returning (otherwise,
1062 : * just signal checkpointer to do it, and return).
1063 : * CHECKPOINT_CAUSE_XLOG: checkpoint is requested due to xlog filling.
1064 : * (This affects logging, and in particular enables CheckPointWarning.)
1065 : */
1066 : void
1067 4952 : RequestCheckpoint(int flags)
1068 : {
1069 : int ntries;
1070 : int old_failed,
1071 : old_started;
1072 :
1073 : /*
1074 : * If in a standalone backend, just do it ourselves.
1075 : */
1076 4952 : if (!IsPostmasterEnvironment)
1077 : {
1078 : /*
1079 : * There's no point in doing slow checkpoints in a standalone backend,
1080 : * because there's no other backends the checkpoint could disrupt.
1081 : */
1082 402 : CreateCheckPoint(flags | CHECKPOINT_FAST);
1083 :
1084 : /* Free all smgr objects, as CheckpointerMain() normally would. */
1085 402 : smgrdestroyall();
1086 :
1087 402 : return;
1088 : }
1089 :
1090 : /*
1091 : * Atomically set the request flags, and take a snapshot of the counters.
1092 : * When we see ckpt_started > old_started, we know the flags we set here
1093 : * have been seen by checkpointer.
1094 : *
1095 : * Note that we OR the flags with any existing flags, to avoid overriding
1096 : * a "stronger" request by another backend. The flag senses must be
1097 : * chosen to make this work!
1098 : */
1099 4550 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1100 :
1101 4550 : old_failed = CheckpointerShmem->ckpt_failed;
1102 4550 : old_started = CheckpointerShmem->ckpt_started;
1103 4550 : CheckpointerShmem->ckpt_flags |= (flags | CHECKPOINT_REQUESTED);
1104 :
1105 4550 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1106 :
1107 : /*
1108 : * Set checkpointer's latch to request checkpoint. It's possible that the
1109 : * checkpointer hasn't started yet, so we will retry a few times if
1110 : * needed. (Actually, more than a few times, since on slow or overloaded
1111 : * buildfarm machines, it's been observed that the checkpointer can take
1112 : * several seconds to start.) However, if not told to wait for the
1113 : * checkpoint to occur, we consider failure to set the latch to be
1114 : * nonfatal and merely LOG it. The checkpointer should see the request
1115 : * when it does start, with or without the SetLatch().
1116 : */
1117 : #define MAX_SIGNAL_TRIES 600 /* max wait 60.0 sec */
1118 4550 : for (ntries = 0;; ntries++)
1119 38 : {
1120 4588 : volatile PROC_HDR *procglobal = ProcGlobal;
1121 4588 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1122 :
1123 4588 : if (checkpointerProc == INVALID_PROC_NUMBER)
1124 : {
1125 38 : if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT))
1126 : {
1127 0 : elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG,
1128 : "could not notify checkpoint: checkpointer is not running");
1129 0 : break;
1130 : }
1131 : }
1132 : else
1133 : {
1134 4550 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1135 : /* notified successfully */
1136 4550 : break;
1137 : }
1138 :
1139 38 : CHECK_FOR_INTERRUPTS();
1140 38 : pg_usleep(100000L); /* wait 0.1 sec, then retry */
1141 : }
1142 :
1143 : /*
1144 : * If requested, wait for completion. We detect completion according to
1145 : * the algorithm given above.
1146 : */
1147 4550 : if (flags & CHECKPOINT_WAIT)
1148 : {
1149 : int new_started,
1150 : new_failed;
1151 :
1152 : /* Wait for a new checkpoint to start. */
1153 1594 : ConditionVariablePrepareToSleep(&CheckpointerShmem->start_cv);
1154 : for (;;)
1155 : {
1156 2918 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1157 2918 : new_started = CheckpointerShmem->ckpt_started;
1158 2918 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1159 :
1160 2918 : if (new_started != old_started)
1161 1594 : break;
1162 :
1163 1324 : ConditionVariableSleep(&CheckpointerShmem->start_cv,
1164 : WAIT_EVENT_CHECKPOINT_START);
1165 : }
1166 1594 : ConditionVariableCancelSleep();
1167 :
1168 : /*
1169 : * We are waiting for ckpt_done >= new_started, in a modulo sense.
1170 : */
1171 1594 : ConditionVariablePrepareToSleep(&CheckpointerShmem->done_cv);
1172 : for (;;)
1173 1214 : {
1174 : int new_done;
1175 :
1176 2808 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1177 2808 : new_done = CheckpointerShmem->ckpt_done;
1178 2808 : new_failed = CheckpointerShmem->ckpt_failed;
1179 2808 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1180 :
1181 2808 : if (new_done - new_started >= 0)
1182 1594 : break;
1183 :
1184 1214 : ConditionVariableSleep(&CheckpointerShmem->done_cv,
1185 : WAIT_EVENT_CHECKPOINT_DONE);
1186 : }
1187 1594 : ConditionVariableCancelSleep();
1188 :
1189 1594 : if (new_failed != old_failed)
1190 0 : ereport(ERROR,
1191 : (errmsg("checkpoint request failed"),
1192 : errhint("Consult recent messages in the server log for details.")));
1193 : }
1194 : }
1195 :
1196 : /*
1197 : * ForwardSyncRequest
1198 : * Forward a file-fsync request from a backend to the checkpointer
1199 : *
1200 : * Whenever a backend is compelled to write directly to a relation
1201 : * (which should be seldom, if the background writer is getting its job done),
1202 : * the backend calls this routine to pass over knowledge that the relation
1203 : * is dirty and must be fsync'd before next checkpoint. We also use this
1204 : * opportunity to count such writes for statistical purposes.
1205 : *
1206 : * To avoid holding the lock for longer than necessary, we normally write
1207 : * to the requests[] queue without checking for duplicates. The checkpointer
1208 : * will have to eliminate dups internally anyway. However, if we discover
1209 : * that the queue is full, we make a pass over the entire queue to compact
1210 : * it. This is somewhat expensive, but the alternative is for the backend
1211 : * to perform its own fsync, which is far more expensive in practice. It
1212 : * is theoretically possible a backend fsync might still be necessary, if
1213 : * the queue is full and contains no duplicate entries. In that case, we
1214 : * let the backend know by returning false.
1215 : */
1216 : bool
1217 2432224 : ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)
1218 : {
1219 : CheckpointerRequest *request;
1220 : bool too_full;
1221 : int insert_pos;
1222 :
1223 2432224 : if (!IsUnderPostmaster)
1224 0 : return false; /* probably shouldn't even get here */
1225 :
1226 2432224 : if (AmCheckpointerProcess())
1227 0 : elog(ERROR, "ForwardSyncRequest must not be called in checkpointer");
1228 :
1229 2432224 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1230 :
1231 : /*
1232 : * If the checkpointer isn't running or the request queue is full, the
1233 : * backend will have to perform its own fsync request. But before forcing
1234 : * that to happen, we can try to compact the request queue.
1235 : */
1236 2432224 : if (CheckpointerShmem->checkpointer_pid == 0 ||
1237 2430722 : (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&
1238 2358 : !CompactCheckpointerRequestQueue()))
1239 : {
1240 3254 : LWLockRelease(CheckpointerCommLock);
1241 3254 : return false;
1242 : }
1243 :
1244 : /* OK, insert request */
1245 2428970 : insert_pos = CheckpointerShmem->tail;
1246 2428970 : request = &CheckpointerShmem->requests[insert_pos];
1247 2428970 : request->ftag = *ftag;
1248 2428970 : request->type = type;
1249 :
1250 2428970 : CheckpointerShmem->tail = (CheckpointerShmem->tail + 1) % CheckpointerShmem->max_requests;
1251 2428970 : CheckpointerShmem->num_requests++;
1252 :
1253 : /* If queue is more than half full, nudge the checkpointer to empty it */
1254 2428970 : too_full = (CheckpointerShmem->num_requests >=
1255 2428970 : CheckpointerShmem->max_requests / 2);
1256 :
1257 2428970 : LWLockRelease(CheckpointerCommLock);
1258 :
1259 : /* ... but not till after we release the lock */
1260 2428970 : if (too_full)
1261 : {
1262 63556 : volatile PROC_HDR *procglobal = ProcGlobal;
1263 63556 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1264 :
1265 63556 : if (checkpointerProc != INVALID_PROC_NUMBER)
1266 63556 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1267 : }
1268 :
1269 2428970 : return true;
1270 : }
1271 :
1272 : /*
1273 : * CompactCheckpointerRequestQueue
1274 : * Remove duplicates from the request queue to avoid backend fsyncs.
1275 : * Returns "true" if any entries were removed.
1276 : *
1277 : * Although a full fsync request queue is not common, it can lead to severe
1278 : * performance problems when it does happen. So far, this situation has
1279 : * only been observed to occur when the system is under heavy write load,
1280 : * and especially during the "sync" phase of a checkpoint. Without this
1281 : * logic, each backend begins doing an fsync for every block written, which
1282 : * gets very expensive and can slow down the whole system.
1283 : *
1284 : * Trying to do this every time the queue is full could lose if there
1285 : * aren't any removable entries. But that should be vanishingly rare in
1286 : * practice: there's one queue entry per shared buffer.
1287 : */
1288 : static bool
1289 2358 : CompactCheckpointerRequestQueue(void)
1290 : {
1291 : struct CheckpointerSlotMapping
1292 : {
1293 : CheckpointerRequest request;
1294 : int ring_idx;
1295 : };
1296 :
1297 : int n;
1298 2358 : int num_skipped = 0;
1299 : int head;
1300 : int max_requests;
1301 : int num_requests;
1302 : int read_idx,
1303 : write_idx;
1304 : HASHCTL ctl;
1305 : HTAB *htab;
1306 : bool *skip_slot;
1307 :
1308 : /* must hold CheckpointerCommLock in exclusive mode */
1309 : Assert(LWLockHeldByMe(CheckpointerCommLock));
1310 :
1311 : /* Avoid memory allocations in a critical section. */
1312 2358 : if (CritSectionCount > 0)
1313 6 : return false;
1314 :
1315 2352 : max_requests = CheckpointerShmem->max_requests;
1316 2352 : num_requests = CheckpointerShmem->num_requests;
1317 :
1318 : /* Initialize skip_slot array */
1319 2352 : skip_slot = palloc0(sizeof(bool) * max_requests);
1320 :
1321 2352 : head = CheckpointerShmem->head;
1322 :
1323 : /* Initialize temporary hash table */
1324 2352 : ctl.keysize = sizeof(CheckpointerRequest);
1325 2352 : ctl.entrysize = sizeof(struct CheckpointerSlotMapping);
1326 2352 : ctl.hcxt = CurrentMemoryContext;
1327 :
1328 2352 : htab = hash_create("CompactCheckpointerRequestQueue",
1329 2352 : CheckpointerShmem->num_requests,
1330 : &ctl,
1331 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1332 :
1333 : /*
1334 : * The basic idea here is that a request can be skipped if it's followed
1335 : * by a later, identical request. It might seem more sensible to work
1336 : * backwards from the end of the queue and check whether a request is
1337 : * *preceded* by an earlier, identical request, in the hopes of doing less
1338 : * copying. But that might change the semantics, if there's an
1339 : * intervening SYNC_FORGET_REQUEST or SYNC_FILTER_REQUEST, so we do it
1340 : * this way. It would be possible to be even smarter if we made the code
1341 : * below understand the specific semantics of such requests (it could blow
1342 : * away preceding entries that would end up being canceled anyhow), but
1343 : * it's not clear that the extra complexity would buy us anything.
1344 : */
1345 2352 : read_idx = head;
1346 302288 : for (n = 0; n < num_requests; n++)
1347 : {
1348 : CheckpointerRequest *request;
1349 : struct CheckpointerSlotMapping *slotmap;
1350 : bool found;
1351 :
1352 : /*
1353 : * We use the request struct directly as a hashtable key. This
1354 : * assumes that any padding bytes in the structs are consistently the
1355 : * same, which should be okay because we zeroed them in
1356 : * CheckpointerShmemInit. Note also that RelFileLocator had better
1357 : * contain no pad bytes.
1358 : */
1359 299936 : request = &CheckpointerShmem->requests[read_idx];
1360 299936 : slotmap = hash_search(htab, request, HASH_ENTER, &found);
1361 299936 : if (found)
1362 : {
1363 : /* Duplicate, so mark the previous occurrence as skippable */
1364 26022 : skip_slot[slotmap->ring_idx] = true;
1365 26022 : num_skipped++;
1366 : }
1367 : /* Remember slot containing latest occurrence of this request value */
1368 299936 : slotmap->ring_idx = read_idx;
1369 :
1370 : /* Move to the next request in the ring buffer */
1371 299936 : read_idx = (read_idx + 1) % max_requests;
1372 : }
1373 :
1374 : /* Done with the hash table. */
1375 2352 : hash_destroy(htab);
1376 :
1377 : /* If no duplicates, we're out of luck. */
1378 2352 : if (!num_skipped)
1379 : {
1380 1746 : pfree(skip_slot);
1381 1746 : return false;
1382 : }
1383 :
1384 : /* We found some duplicates; remove them. */
1385 606 : read_idx = write_idx = head;
1386 77054 : for (n = 0; n < num_requests; n++)
1387 : {
1388 : /* If this slot is NOT skipped, keep it */
1389 76448 : if (!skip_slot[read_idx])
1390 : {
1391 : /* If the read and write positions are different, copy the request */
1392 50426 : if (write_idx != read_idx)
1393 22116 : CheckpointerShmem->requests[write_idx] =
1394 22116 : CheckpointerShmem->requests[read_idx];
1395 :
1396 : /* Advance the write position */
1397 50426 : write_idx = (write_idx + 1) % max_requests;
1398 : }
1399 :
1400 76448 : read_idx = (read_idx + 1) % max_requests;
1401 : }
1402 :
1403 : /*
1404 : * Update ring buffer state: head remains the same, tail moves, count
1405 : * decreases
1406 : */
1407 606 : CheckpointerShmem->tail = write_idx;
1408 606 : CheckpointerShmem->num_requests -= num_skipped;
1409 :
1410 606 : ereport(DEBUG1,
1411 : (errmsg_internal("compacted fsync request queue from %d entries to %d entries",
1412 : num_requests, CheckpointerShmem->num_requests)));
1413 :
1414 : /* Cleanup. */
1415 606 : pfree(skip_slot);
1416 606 : return true;
1417 : }
1418 :
1419 : /*
1420 : * AbsorbSyncRequests
1421 : * Retrieve queued sync requests and pass them to sync mechanism.
1422 : *
1423 : * This is exported because it must be called during CreateCheckPoint;
1424 : * we have to be sure we have accepted all pending requests just before
1425 : * we start fsync'ing. Since CreateCheckPoint sometimes runs in
1426 : * non-checkpointer processes, do nothing if not checkpointer.
1427 : */
1428 : void
1429 37366 : AbsorbSyncRequests(void)
1430 : {
1431 37366 : CheckpointerRequest *requests = NULL;
1432 : CheckpointerRequest *request;
1433 : int n,
1434 : i;
1435 : bool loop;
1436 :
1437 37366 : if (!AmCheckpointerProcess())
1438 1240 : return;
1439 :
1440 : do
1441 : {
1442 36128 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1443 :
1444 : /*---
1445 : * We try to avoid holding the lock for a long time by:
1446 : * 1. Copying the request array and processing the requests after
1447 : * releasing the lock;
1448 : * 2. Processing not the whole queue, but only batches of
1449 : * CKPT_REQ_BATCH_SIZE at once.
1450 : *
1451 : * Once we have cleared the requests from shared memory, we must
1452 : * PANIC if we then fail to absorb them (e.g., because our hashtable
1453 : * runs out of memory). This is because the system cannot run safely
1454 : * if we are unable to fsync what we have been told to fsync.
1455 : * Fortunately, the hashtable is so small that the problem is quite
1456 : * unlikely to arise in practice.
1457 : *
1458 : * Note: The maximum possible size of a ring buffer is
1459 : * MAX_CHECKPOINT_REQUESTS entries, which fit into a maximum palloc
1460 : * allocation size of 1Gb. Our maximum batch size,
1461 : * CKPT_REQ_BATCH_SIZE, is even smaller.
1462 : */
1463 36128 : n = Min(CheckpointerShmem->num_requests, CKPT_REQ_BATCH_SIZE);
1464 36128 : if (n > 0)
1465 : {
1466 17886 : if (!requests)
1467 17884 : requests = (CheckpointerRequest *) palloc(n * sizeof(CheckpointerRequest));
1468 :
1469 2249506 : for (i = 0; i < n; i++)
1470 : {
1471 2231620 : requests[i] = CheckpointerShmem->requests[CheckpointerShmem->head];
1472 2231620 : CheckpointerShmem->head = (CheckpointerShmem->head + 1) % CheckpointerShmem->max_requests;
1473 : }
1474 :
1475 17886 : CheckpointerShmem->num_requests -= n;
1476 :
1477 : }
1478 :
1479 36128 : START_CRIT_SECTION();
1480 :
1481 : /* Are there any requests in the queue? If so, keep going. */
1482 36128 : loop = CheckpointerShmem->num_requests != 0;
1483 :
1484 36128 : LWLockRelease(CheckpointerCommLock);
1485 :
1486 2267748 : for (request = requests; n > 0; request++, n--)
1487 2231620 : RememberSyncRequest(&request->ftag, request->type);
1488 :
1489 36128 : END_CRIT_SECTION();
1490 36128 : } while (loop);
1491 :
1492 36126 : if (requests)
1493 17884 : pfree(requests);
1494 : }
1495 :
1496 : /*
1497 : * Update any shared memory configurations based on config parameters
1498 : */
1499 : static void
1500 1156 : UpdateSharedMemoryConfig(void)
1501 : {
1502 : /* update global shmem state for sync rep */
1503 1156 : SyncRepUpdateSyncStandbysDefined();
1504 :
1505 : /*
1506 : * If full_page_writes has been changed by SIGHUP, we update it in shared
1507 : * memory and write an XLOG_FPW_CHANGE record.
1508 : */
1509 1156 : UpdateFullPageWrites();
1510 :
1511 1156 : elog(DEBUG2, "checkpointer updated shared memory configuration values");
1512 1156 : }
1513 :
1514 : /*
1515 : * FirstCallSinceLastCheckpoint allows a process to take an action once
1516 : * per checkpoint cycle by asynchronously checking for checkpoint completion.
1517 : */
1518 : bool
1519 26710 : FirstCallSinceLastCheckpoint(void)
1520 : {
1521 : static int ckpt_done = 0;
1522 : int new_done;
1523 26710 : bool FirstCall = false;
1524 :
1525 26710 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1526 26710 : new_done = CheckpointerShmem->ckpt_done;
1527 26710 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1528 :
1529 26710 : if (new_done != ckpt_done)
1530 1206 : FirstCall = true;
1531 :
1532 26710 : ckpt_done = new_done;
1533 :
1534 26710 : return FirstCall;
1535 : }
|
