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 : * Normal termination is by SIGUSR2, which instructs the checkpointer to
14 : * execute a shutdown checkpoint and then exit(0). (All backends must be
15 : * stopped before SIGUSR2 is issued!) Emergency termination is by SIGQUIT;
16 : * like any backend, the checkpointer will simply abort and exit on SIGQUIT.
17 : *
18 : * If the checkpointer exits unexpectedly, the postmaster treats that the same
19 : * as a backend crash: shared memory may be corrupted, so remaining backends
20 : * should be killed by SIGQUIT and then a recovery cycle started. (Even if
21 : * shared memory isn't corrupted, we have lost information about which
22 : * files need to be fsync'd for the next checkpoint, and so a system
23 : * restart needs to be forced.)
24 : *
25 : *
26 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
27 : *
28 : *
29 : * IDENTIFICATION
30 : * src/backend/postmaster/checkpointer.c
31 : *
32 : *-------------------------------------------------------------------------
33 : */
34 : #include "postgres.h"
35 :
36 : #include <sys/time.h>
37 : #include <time.h>
38 :
39 : #include "access/xlog.h"
40 : #include "access/xlog_internal.h"
41 : #include "access/xlogrecovery.h"
42 : #include "libpq/pqsignal.h"
43 : #include "miscadmin.h"
44 : #include "pgstat.h"
45 : #include "postmaster/auxprocess.h"
46 : #include "postmaster/bgwriter.h"
47 : #include "postmaster/interrupt.h"
48 : #include "replication/syncrep.h"
49 : #include "storage/bufmgr.h"
50 : #include "storage/condition_variable.h"
51 : #include "storage/fd.h"
52 : #include "storage/ipc.h"
53 : #include "storage/lwlock.h"
54 : #include "storage/proc.h"
55 : #include "storage/procsignal.h"
56 : #include "storage/shmem.h"
57 : #include "storage/smgr.h"
58 : #include "storage/spin.h"
59 : #include "utils/guc.h"
60 : #include "utils/memutils.h"
61 : #include "utils/resowner.h"
62 :
63 :
64 : /*----------
65 : * Shared memory area for communication between checkpointer and backends
66 : *
67 : * The ckpt counters allow backends to watch for completion of a checkpoint
68 : * request they send. Here's how it works:
69 : * * At start of a checkpoint, checkpointer reads (and clears) the request
70 : * flags and increments ckpt_started, while holding ckpt_lck.
71 : * * On completion of a checkpoint, checkpointer sets ckpt_done to
72 : * equal ckpt_started.
73 : * * On failure of a checkpoint, checkpointer increments ckpt_failed
74 : * and sets ckpt_done to equal ckpt_started.
75 : *
76 : * The algorithm for backends is:
77 : * 1. Record current values of ckpt_failed and ckpt_started, and
78 : * set request flags, while holding ckpt_lck.
79 : * 2. Send signal to request checkpoint.
80 : * 3. Sleep until ckpt_started changes. Now you know a checkpoint has
81 : * begun since you started this algorithm (although *not* that it was
82 : * specifically initiated by your signal), and that it is using your flags.
83 : * 4. Record new value of ckpt_started.
84 : * 5. Sleep until ckpt_done >= saved value of ckpt_started. (Use modulo
85 : * arithmetic here in case counters wrap around.) Now you know a
86 : * checkpoint has started and completed, but not whether it was
87 : * successful.
88 : * 6. If ckpt_failed is different from the originally saved value,
89 : * assume request failed; otherwise it was definitely successful.
90 : *
91 : * ckpt_flags holds the OR of the checkpoint request flags sent by all
92 : * requesting backends since the last checkpoint start. The flags are
93 : * chosen so that OR'ing is the correct way to combine multiple requests.
94 : *
95 : * The requests array holds fsync requests sent by backends and not yet
96 : * absorbed by the checkpointer.
97 : *
98 : * Unlike the checkpoint fields, requests related fields are protected by
99 : * CheckpointerCommLock.
100 : *----------
101 : */
102 : typedef struct
103 : {
104 : SyncRequestType type; /* request type */
105 : FileTag ftag; /* file identifier */
106 : } CheckpointerRequest;
107 :
108 : typedef struct
109 : {
110 : pid_t checkpointer_pid; /* PID (0 if not started) */
111 :
112 : slock_t ckpt_lck; /* protects all the ckpt_* fields */
113 :
114 : int ckpt_started; /* advances when checkpoint starts */
115 : int ckpt_done; /* advances when checkpoint done */
116 : int ckpt_failed; /* advances when checkpoint fails */
117 :
118 : int ckpt_flags; /* checkpoint flags, as defined in xlog.h */
119 :
120 : ConditionVariable start_cv; /* signaled when ckpt_started advances */
121 : ConditionVariable done_cv; /* signaled when ckpt_done advances */
122 :
123 : int num_requests; /* current # of requests */
124 : int max_requests; /* allocated array size */
125 : CheckpointerRequest requests[FLEXIBLE_ARRAY_MEMBER];
126 : } CheckpointerShmemStruct;
127 :
128 : static CheckpointerShmemStruct *CheckpointerShmem;
129 :
130 : /* interval for calling AbsorbSyncRequests in CheckpointWriteDelay */
131 : #define WRITES_PER_ABSORB 1000
132 :
133 : /*
134 : * GUC parameters
135 : */
136 : int CheckPointTimeout = 300;
137 : int CheckPointWarning = 30;
138 : double CheckPointCompletionTarget = 0.9;
139 :
140 : /*
141 : * Private state
142 : */
143 : static bool ckpt_active = false;
144 :
145 : /* these values are valid when ckpt_active is true: */
146 : static pg_time_t ckpt_start_time;
147 : static XLogRecPtr ckpt_start_recptr;
148 : static double ckpt_cached_elapsed;
149 :
150 : static pg_time_t last_checkpoint_time;
151 : static pg_time_t last_xlog_switch_time;
152 :
153 : /* Prototypes for private functions */
154 :
155 : static void HandleCheckpointerInterrupts(void);
156 : static void CheckArchiveTimeout(void);
157 : static bool IsCheckpointOnSchedule(double progress);
158 : static bool ImmediateCheckpointRequested(void);
159 : static bool CompactCheckpointerRequestQueue(void);
160 : static void UpdateSharedMemoryConfig(void);
161 :
162 : /* Signal handlers */
163 : static void ReqCheckpointHandler(SIGNAL_ARGS);
164 :
165 :
166 : /*
167 : * Main entry point for checkpointer process
168 : *
169 : * This is invoked from AuxiliaryProcessMain, which has already created the
170 : * basic execution environment, but not enabled signals yet.
171 : */
172 : void
173 828 : CheckpointerMain(char *startup_data, size_t startup_data_len)
174 : {
175 : sigjmp_buf local_sigjmp_buf;
176 : MemoryContext checkpointer_context;
177 :
178 : Assert(startup_data_len == 0);
179 :
180 828 : MyBackendType = B_CHECKPOINTER;
181 828 : AuxiliaryProcessMainCommon();
182 :
183 828 : CheckpointerShmem->checkpointer_pid = MyProcPid;
184 :
185 : /*
186 : * Properly accept or ignore signals the postmaster might send us
187 : *
188 : * Note: we deliberately ignore SIGTERM, because during a standard Unix
189 : * system shutdown cycle, init will SIGTERM all processes at once. We
190 : * want to wait for the backends to exit, whereupon the postmaster will
191 : * tell us it's okay to shut down (via SIGUSR2).
192 : */
193 828 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
194 828 : pqsignal(SIGINT, ReqCheckpointHandler); /* request checkpoint */
195 828 : pqsignal(SIGTERM, SIG_IGN); /* ignore SIGTERM */
196 : /* SIGQUIT handler was already set up by InitPostmasterChild */
197 828 : pqsignal(SIGALRM, SIG_IGN);
198 828 : pqsignal(SIGPIPE, SIG_IGN);
199 828 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
200 828 : pqsignal(SIGUSR2, SignalHandlerForShutdownRequest);
201 :
202 : /*
203 : * Reset some signals that are accepted by postmaster but not here
204 : */
205 828 : pqsignal(SIGCHLD, SIG_DFL);
206 :
207 : /*
208 : * Initialize so that first time-driven event happens at the correct time.
209 : */
210 828 : last_checkpoint_time = last_xlog_switch_time = (pg_time_t) time(NULL);
211 :
212 : /*
213 : * Write out stats after shutdown. This needs to be called by exactly one
214 : * process during a normal shutdown, and since checkpointer is shut down
215 : * very late...
216 : *
217 : * Walsenders are shut down after the checkpointer, but currently don't
218 : * report stats. If that changes, we need a more complicated solution.
219 : */
220 828 : before_shmem_exit(pgstat_before_server_shutdown, 0);
221 :
222 : /*
223 : * Create a memory context that we will do all our work in. We do this so
224 : * that we can reset the context during error recovery and thereby avoid
225 : * possible memory leaks. Formerly this code just ran in
226 : * TopMemoryContext, but resetting that would be a really bad idea.
227 : */
228 828 : checkpointer_context = AllocSetContextCreate(TopMemoryContext,
229 : "Checkpointer",
230 : ALLOCSET_DEFAULT_SIZES);
231 828 : MemoryContextSwitchTo(checkpointer_context);
232 :
233 : /*
234 : * If an exception is encountered, processing resumes here.
235 : *
236 : * You might wonder why this isn't coded as an infinite loop around a
237 : * PG_TRY construct. The reason is that this is the bottom of the
238 : * exception stack, and so with PG_TRY there would be no exception handler
239 : * in force at all during the CATCH part. By leaving the outermost setjmp
240 : * always active, we have at least some chance of recovering from an error
241 : * during error recovery. (If we get into an infinite loop thereby, it
242 : * will soon be stopped by overflow of elog.c's internal state stack.)
243 : *
244 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
245 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
246 : * signals other than SIGQUIT will be blocked until we complete error
247 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
248 : * call redundant, but it is not since InterruptPending might be set
249 : * already.
250 : */
251 828 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
252 : {
253 : /* Since not using PG_TRY, must reset error stack by hand */
254 0 : error_context_stack = NULL;
255 :
256 : /* Prevent interrupts while cleaning up */
257 0 : HOLD_INTERRUPTS();
258 :
259 : /* Report the error to the server log */
260 0 : EmitErrorReport();
261 :
262 : /*
263 : * These operations are really just a minimal subset of
264 : * AbortTransaction(). We don't have very many resources to worry
265 : * about in checkpointer, but we do have LWLocks, buffers, and temp
266 : * files.
267 : */
268 0 : LWLockReleaseAll();
269 0 : ConditionVariableCancelSleep();
270 0 : pgstat_report_wait_end();
271 0 : UnlockBuffers();
272 0 : ReleaseAuxProcessResources(false);
273 0 : AtEOXact_Buffers(false);
274 0 : AtEOXact_SMgr();
275 0 : AtEOXact_Files(false);
276 0 : AtEOXact_HashTables(false);
277 :
278 : /* Warn any waiting backends that the checkpoint failed. */
279 0 : if (ckpt_active)
280 : {
281 0 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
282 0 : CheckpointerShmem->ckpt_failed++;
283 0 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
284 0 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
285 :
286 0 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
287 :
288 0 : ckpt_active = false;
289 : }
290 :
291 : /*
292 : * Now return to normal top-level context and clear ErrorContext for
293 : * next time.
294 : */
295 0 : MemoryContextSwitchTo(checkpointer_context);
296 0 : FlushErrorState();
297 :
298 : /* Flush any leaked data in the top-level context */
299 0 : MemoryContextReset(checkpointer_context);
300 :
301 : /* Now we can allow interrupts again */
302 0 : RESUME_INTERRUPTS();
303 :
304 : /*
305 : * Sleep at least 1 second after any error. A write error is likely
306 : * to be repeated, and we don't want to be filling the error logs as
307 : * fast as we can.
308 : */
309 0 : pg_usleep(1000000L);
310 : }
311 :
312 : /* We can now handle ereport(ERROR) */
313 828 : PG_exception_stack = &local_sigjmp_buf;
314 :
315 : /*
316 : * Unblock signals (they were blocked when the postmaster forked us)
317 : */
318 828 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
319 :
320 : /*
321 : * Ensure all shared memory values are set correctly for the config. Doing
322 : * this here ensures no race conditions from other concurrent updaters.
323 : */
324 828 : UpdateSharedMemoryConfig();
325 :
326 : /*
327 : * Advertise our latch that backends can use to wake us up while we're
328 : * sleeping.
329 : */
330 828 : ProcGlobal->checkpointerLatch = &MyProc->procLatch;
331 :
332 : /*
333 : * Loop forever
334 : */
335 : for (;;)
336 11050 : {
337 11878 : bool do_checkpoint = false;
338 11878 : int flags = 0;
339 : pg_time_t now;
340 : int elapsed_secs;
341 : int cur_timeout;
342 11878 : bool chkpt_or_rstpt_requested = false;
343 11878 : bool chkpt_or_rstpt_timed = false;
344 :
345 : /* Clear any already-pending wakeups */
346 11878 : ResetLatch(MyLatch);
347 :
348 : /*
349 : * Process any requests or signals received recently.
350 : */
351 11878 : AbsorbSyncRequests();
352 11878 : HandleCheckpointerInterrupts();
353 :
354 : /*
355 : * Detect a pending checkpoint request by checking whether the flags
356 : * word in shared memory is nonzero. We shouldn't need to acquire the
357 : * ckpt_lck for this.
358 : */
359 11070 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
360 : {
361 478 : do_checkpoint = true;
362 478 : chkpt_or_rstpt_requested = true;
363 : }
364 :
365 : /*
366 : * Force a checkpoint if too much time has elapsed since the last one.
367 : * Note that we count a timed checkpoint in stats only when this
368 : * occurs without an external request, but we set the CAUSE_TIME flag
369 : * bit even if there is also an external request.
370 : */
371 11070 : now = (pg_time_t) time(NULL);
372 11070 : elapsed_secs = now - last_checkpoint_time;
373 11070 : if (elapsed_secs >= CheckPointTimeout)
374 : {
375 0 : if (!do_checkpoint)
376 0 : chkpt_or_rstpt_timed = true;
377 0 : do_checkpoint = true;
378 0 : flags |= CHECKPOINT_CAUSE_TIME;
379 : }
380 :
381 : /*
382 : * Do a checkpoint if requested.
383 : */
384 11070 : if (do_checkpoint)
385 : {
386 478 : bool ckpt_performed = false;
387 : bool do_restartpoint;
388 :
389 : /* Check if we should perform a checkpoint or a restartpoint. */
390 478 : do_restartpoint = RecoveryInProgress();
391 :
392 : /*
393 : * Atomically fetch the request flags to figure out what kind of a
394 : * checkpoint we should perform, and increase the started-counter
395 : * to acknowledge that we've started a new checkpoint.
396 : */
397 478 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
398 478 : flags |= CheckpointerShmem->ckpt_flags;
399 478 : CheckpointerShmem->ckpt_flags = 0;
400 478 : CheckpointerShmem->ckpt_started++;
401 478 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
402 :
403 478 : ConditionVariableBroadcast(&CheckpointerShmem->start_cv);
404 :
405 : /*
406 : * The end-of-recovery checkpoint is a real checkpoint that's
407 : * performed while we're still in recovery.
408 : */
409 478 : if (flags & CHECKPOINT_END_OF_RECOVERY)
410 38 : do_restartpoint = false;
411 :
412 478 : if (chkpt_or_rstpt_timed)
413 : {
414 0 : chkpt_or_rstpt_timed = false;
415 0 : if (do_restartpoint)
416 0 : PendingCheckpointerStats.restartpoints_timed++;
417 : else
418 0 : PendingCheckpointerStats.num_timed++;
419 : }
420 :
421 478 : if (chkpt_or_rstpt_requested)
422 : {
423 478 : chkpt_or_rstpt_requested = false;
424 478 : if (do_restartpoint)
425 78 : PendingCheckpointerStats.restartpoints_requested++;
426 : else
427 400 : PendingCheckpointerStats.num_requested++;
428 : }
429 :
430 : /*
431 : * We will warn if (a) too soon since last checkpoint (whatever
432 : * caused it) and (b) somebody set the CHECKPOINT_CAUSE_XLOG flag
433 : * since the last checkpoint start. Note in particular that this
434 : * implementation will not generate warnings caused by
435 : * CheckPointTimeout < CheckPointWarning.
436 : */
437 478 : if (!do_restartpoint &&
438 400 : (flags & CHECKPOINT_CAUSE_XLOG) &&
439 56 : elapsed_secs < CheckPointWarning)
440 56 : ereport(LOG,
441 : (errmsg_plural("checkpoints are occurring too frequently (%d second apart)",
442 : "checkpoints are occurring too frequently (%d seconds apart)",
443 : elapsed_secs,
444 : elapsed_secs),
445 : errhint("Consider increasing the configuration parameter max_wal_size.")));
446 :
447 : /*
448 : * Initialize checkpointer-private variables used during
449 : * checkpoint.
450 : */
451 478 : ckpt_active = true;
452 478 : if (do_restartpoint)
453 78 : ckpt_start_recptr = GetXLogReplayRecPtr(NULL);
454 : else
455 400 : ckpt_start_recptr = GetInsertRecPtr();
456 478 : ckpt_start_time = now;
457 478 : ckpt_cached_elapsed = 0;
458 :
459 : /*
460 : * Do the checkpoint.
461 : */
462 478 : if (!do_restartpoint)
463 : {
464 400 : CreateCheckPoint(flags);
465 400 : ckpt_performed = true;
466 : }
467 : else
468 78 : ckpt_performed = CreateRestartPoint(flags);
469 :
470 : /*
471 : * After any checkpoint, free all smgr objects. Otherwise we
472 : * would never do so for dropped relations, as the checkpointer
473 : * does not process shared invalidation messages or call
474 : * AtEOXact_SMgr().
475 : */
476 478 : smgrdestroyall();
477 :
478 : /*
479 : * Indicate checkpoint completion to any waiting backends.
480 : */
481 478 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
482 478 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
483 478 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
484 :
485 478 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
486 :
487 478 : if (ckpt_performed)
488 : {
489 : /*
490 : * Note we record the checkpoint start time not end time as
491 : * last_checkpoint_time. This is so that time-driven
492 : * checkpoints happen at a predictable spacing.
493 : */
494 442 : last_checkpoint_time = now;
495 :
496 442 : if (do_restartpoint)
497 42 : PendingCheckpointerStats.restartpoints_performed++;
498 : }
499 : else
500 : {
501 : /*
502 : * We were not able to perform the restartpoint (checkpoints
503 : * throw an ERROR in case of error). Most likely because we
504 : * have not received any new checkpoint WAL records since the
505 : * last restartpoint. Try again in 15 s.
506 : */
507 36 : last_checkpoint_time = now - CheckPointTimeout + 15;
508 : }
509 :
510 478 : ckpt_active = false;
511 :
512 : /* We may have received an interrupt during the checkpoint. */
513 478 : HandleCheckpointerInterrupts();
514 : }
515 :
516 : /* Check for archive_timeout and switch xlog files if necessary. */
517 11056 : CheckArchiveTimeout();
518 :
519 : /* Report pending statistics to the cumulative stats system */
520 11056 : pgstat_report_checkpointer();
521 11056 : pgstat_report_wal(true);
522 :
523 : /*
524 : * If any checkpoint flags have been set, redo the loop to handle the
525 : * checkpoint without sleeping.
526 : */
527 11056 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
528 8 : continue;
529 :
530 : /*
531 : * Sleep until we are signaled or it's time for another checkpoint or
532 : * xlog file switch.
533 : */
534 11048 : now = (pg_time_t) time(NULL);
535 11048 : elapsed_secs = now - last_checkpoint_time;
536 11048 : if (elapsed_secs >= CheckPointTimeout)
537 0 : continue; /* no sleep for us ... */
538 11048 : cur_timeout = CheckPointTimeout - elapsed_secs;
539 11048 : if (XLogArchiveTimeout > 0 && !RecoveryInProgress())
540 : {
541 0 : elapsed_secs = now - last_xlog_switch_time;
542 0 : if (elapsed_secs >= XLogArchiveTimeout)
543 0 : continue; /* no sleep for us ... */
544 0 : cur_timeout = Min(cur_timeout, XLogArchiveTimeout - elapsed_secs);
545 : }
546 :
547 11048 : (void) WaitLatch(MyLatch,
548 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
549 : cur_timeout * 1000L /* convert to ms */ ,
550 : WAIT_EVENT_CHECKPOINTER_MAIN);
551 : }
552 : }
553 :
554 : /*
555 : * Process any new interrupts.
556 : */
557 : static void
558 12356 : HandleCheckpointerInterrupts(void)
559 : {
560 12356 : if (ProcSignalBarrierPending)
561 86 : ProcessProcSignalBarrier();
562 :
563 12356 : if (ConfigReloadPending)
564 : {
565 96 : ConfigReloadPending = false;
566 96 : ProcessConfigFile(PGC_SIGHUP);
567 :
568 : /*
569 : * Checkpointer is the last process to shut down, so we ask it to hold
570 : * the keys for a range of other tasks required most of which have
571 : * nothing to do with checkpointing at all.
572 : *
573 : * For various reasons, some config values can change dynamically so
574 : * the primary copy of them is held in shared memory to make sure all
575 : * backends see the same value. We make Checkpointer responsible for
576 : * updating the shared memory copy if the parameter setting changes
577 : * because of SIGHUP.
578 : */
579 96 : UpdateSharedMemoryConfig();
580 : }
581 12356 : if (ShutdownRequestPending)
582 : {
583 : /*
584 : * From here on, elog(ERROR) should end with exit(1), not send control
585 : * back to the sigsetjmp block above
586 : */
587 822 : ExitOnAnyError = true;
588 :
589 : /*
590 : * Close down the database.
591 : *
592 : * Since ShutdownXLOG() creates restartpoint or checkpoint, and
593 : * updates the statistics, increment the checkpoint request and flush
594 : * out pending statistic.
595 : */
596 822 : PendingCheckpointerStats.num_requested++;
597 822 : ShutdownXLOG(0, 0);
598 822 : pgstat_report_checkpointer();
599 822 : pgstat_report_wal(true);
600 :
601 : /* Normal exit from the checkpointer is here */
602 822 : proc_exit(0); /* done */
603 : }
604 :
605 : /* Perform logging of memory contexts of this process */
606 11534 : if (LogMemoryContextPending)
607 4 : ProcessLogMemoryContextInterrupt();
608 11534 : }
609 :
610 : /*
611 : * CheckArchiveTimeout -- check for archive_timeout and switch xlog files
612 : *
613 : * This will switch to a new WAL file and force an archive file write if
614 : * meaningful activity is recorded in the current WAL file. This includes most
615 : * writes, including just a single checkpoint record, but excludes WAL records
616 : * that were inserted with the XLOG_MARK_UNIMPORTANT flag being set (like
617 : * snapshots of running transactions). Such records, depending on
618 : * configuration, occur on regular intervals and don't contain important
619 : * information. This avoids generating archives with a few unimportant
620 : * records.
621 : */
622 : static void
623 14174 : CheckArchiveTimeout(void)
624 : {
625 : pg_time_t now;
626 : pg_time_t last_time;
627 : XLogRecPtr last_switch_lsn;
628 :
629 14174 : if (XLogArchiveTimeout <= 0 || RecoveryInProgress())
630 14174 : return;
631 :
632 0 : now = (pg_time_t) time(NULL);
633 :
634 : /* First we do a quick check using possibly-stale local state. */
635 0 : if ((int) (now - last_xlog_switch_time) < XLogArchiveTimeout)
636 0 : return;
637 :
638 : /*
639 : * Update local state ... note that last_xlog_switch_time is the last time
640 : * a switch was performed *or requested*.
641 : */
642 0 : last_time = GetLastSegSwitchData(&last_switch_lsn);
643 :
644 0 : last_xlog_switch_time = Max(last_xlog_switch_time, last_time);
645 :
646 : /* Now we can do the real checks */
647 0 : if ((int) (now - last_xlog_switch_time) >= XLogArchiveTimeout)
648 : {
649 : /*
650 : * Switch segment only when "important" WAL has been logged since the
651 : * last segment switch (last_switch_lsn points to end of segment
652 : * switch occurred in).
653 : */
654 0 : if (GetLastImportantRecPtr() > last_switch_lsn)
655 : {
656 : XLogRecPtr switchpoint;
657 :
658 : /* mark switch as unimportant, avoids triggering checkpoints */
659 0 : switchpoint = RequestXLogSwitch(true);
660 :
661 : /*
662 : * If the returned pointer points exactly to a segment boundary,
663 : * assume nothing happened.
664 : */
665 0 : if (XLogSegmentOffset(switchpoint, wal_segment_size) != 0)
666 0 : elog(DEBUG1, "write-ahead log switch forced (archive_timeout=%d)",
667 : XLogArchiveTimeout);
668 : }
669 :
670 : /*
671 : * Update state in any case, so we don't retry constantly when the
672 : * system is idle.
673 : */
674 0 : last_xlog_switch_time = now;
675 : }
676 : }
677 :
678 : /*
679 : * Returns true if an immediate checkpoint request is pending. (Note that
680 : * this does not check the *current* checkpoint's IMMEDIATE flag, but whether
681 : * there is one pending behind it.)
682 : */
683 : static bool
684 43990 : ImmediateCheckpointRequested(void)
685 : {
686 43990 : volatile CheckpointerShmemStruct *cps = CheckpointerShmem;
687 :
688 : /*
689 : * We don't need to acquire the ckpt_lck in this case because we're only
690 : * looking at a single flag bit.
691 : */
692 43990 : if (cps->ckpt_flags & CHECKPOINT_IMMEDIATE)
693 5742 : return true;
694 38248 : return false;
695 : }
696 :
697 : /*
698 : * CheckpointWriteDelay -- control rate of checkpoint
699 : *
700 : * This function is called after each page write performed by BufferSync().
701 : * It is responsible for throttling BufferSync()'s write rate to hit
702 : * checkpoint_completion_target.
703 : *
704 : * The checkpoint request flags should be passed in; currently the only one
705 : * examined is CHECKPOINT_IMMEDIATE, which disables delays between writes.
706 : *
707 : * 'progress' is an estimate of how much of the work has been done, as a
708 : * fraction between 0.0 meaning none, and 1.0 meaning all done.
709 : */
710 : void
711 430570 : CheckpointWriteDelay(int flags, double progress)
712 : {
713 : static int absorb_counter = WRITES_PER_ABSORB;
714 :
715 : /* Do nothing if checkpoint is being executed by non-checkpointer process */
716 430570 : if (!AmCheckpointerProcess())
717 78202 : return;
718 :
719 : /*
720 : * Perform the usual duties and take a nap, unless we're behind schedule,
721 : * in which case we just try to catch up as quickly as possible.
722 : */
723 352368 : if (!(flags & CHECKPOINT_IMMEDIATE) &&
724 48702 : !ShutdownRequestPending &&
725 82238 : !ImmediateCheckpointRequested() &&
726 38248 : IsCheckpointOnSchedule(progress))
727 : {
728 3118 : if (ConfigReloadPending)
729 : {
730 0 : ConfigReloadPending = false;
731 0 : ProcessConfigFile(PGC_SIGHUP);
732 : /* update shmem copies of config variables */
733 0 : UpdateSharedMemoryConfig();
734 : }
735 :
736 3118 : AbsorbSyncRequests();
737 3118 : absorb_counter = WRITES_PER_ABSORB;
738 :
739 3118 : CheckArchiveTimeout();
740 :
741 : /* Report interim statistics to the cumulative stats system */
742 3118 : pgstat_report_checkpointer();
743 :
744 : /*
745 : * This sleep used to be connected to bgwriter_delay, typically 200ms.
746 : * That resulted in more frequent wakeups if not much work to do.
747 : * Checkpointer and bgwriter are no longer related so take the Big
748 : * Sleep.
749 : */
750 3118 : WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | WL_TIMEOUT,
751 : 100,
752 : WAIT_EVENT_CHECKPOINT_WRITE_DELAY);
753 3118 : ResetLatch(MyLatch);
754 : }
755 349250 : else if (--absorb_counter <= 0)
756 : {
757 : /*
758 : * Absorb pending fsync requests after each WRITES_PER_ABSORB write
759 : * operations even when we don't sleep, to prevent overflow of the
760 : * fsync request queue.
761 : */
762 144 : AbsorbSyncRequests();
763 144 : absorb_counter = WRITES_PER_ABSORB;
764 : }
765 :
766 : /* Check for barrier events. */
767 352368 : if (ProcSignalBarrierPending)
768 4 : ProcessProcSignalBarrier();
769 : }
770 :
771 : /*
772 : * IsCheckpointOnSchedule -- are we on schedule to finish this checkpoint
773 : * (or restartpoint) in time?
774 : *
775 : * Compares the current progress against the time/segments elapsed since last
776 : * checkpoint, and returns true if the progress we've made this far is greater
777 : * than the elapsed time/segments.
778 : */
779 : static bool
780 38248 : IsCheckpointOnSchedule(double progress)
781 : {
782 : XLogRecPtr recptr;
783 : struct timeval now;
784 : double elapsed_xlogs,
785 : elapsed_time;
786 :
787 : Assert(ckpt_active);
788 :
789 : /* Scale progress according to checkpoint_completion_target. */
790 38248 : progress *= CheckPointCompletionTarget;
791 :
792 : /*
793 : * Check against the cached value first. Only do the more expensive
794 : * calculations once we reach the target previously calculated. Since
795 : * neither time or WAL insert pointer moves backwards, a freshly
796 : * calculated value can only be greater than or equal to the cached value.
797 : */
798 38248 : if (progress < ckpt_cached_elapsed)
799 34236 : return false;
800 :
801 : /*
802 : * Check progress against WAL segments written and CheckPointSegments.
803 : *
804 : * We compare the current WAL insert location against the location
805 : * computed before calling CreateCheckPoint. The code in XLogInsert that
806 : * actually triggers a checkpoint when CheckPointSegments is exceeded
807 : * compares against RedoRecPtr, so this is not completely accurate.
808 : * However, it's good enough for our purposes, we're only calculating an
809 : * estimate anyway.
810 : *
811 : * During recovery, we compare last replayed WAL record's location with
812 : * the location computed before calling CreateRestartPoint. That maintains
813 : * the same pacing as we have during checkpoints in normal operation, but
814 : * we might exceed max_wal_size by a fair amount. That's because there can
815 : * be a large gap between a checkpoint's redo-pointer and the checkpoint
816 : * record itself, and we only start the restartpoint after we've seen the
817 : * checkpoint record. (The gap is typically up to CheckPointSegments *
818 : * checkpoint_completion_target where checkpoint_completion_target is the
819 : * value that was in effect when the WAL was generated).
820 : */
821 4012 : if (RecoveryInProgress())
822 1250 : recptr = GetXLogReplayRecPtr(NULL);
823 : else
824 2762 : recptr = GetInsertRecPtr();
825 4012 : elapsed_xlogs = (((double) (recptr - ckpt_start_recptr)) /
826 4012 : wal_segment_size) / CheckPointSegments;
827 :
828 4012 : if (progress < elapsed_xlogs)
829 : {
830 890 : ckpt_cached_elapsed = elapsed_xlogs;
831 890 : return false;
832 : }
833 :
834 : /*
835 : * Check progress against time elapsed and checkpoint_timeout.
836 : */
837 3122 : gettimeofday(&now, NULL);
838 3122 : elapsed_time = ((double) ((pg_time_t) now.tv_sec - ckpt_start_time) +
839 3122 : now.tv_usec / 1000000.0) / CheckPointTimeout;
840 :
841 3122 : if (progress < elapsed_time)
842 : {
843 4 : ckpt_cached_elapsed = elapsed_time;
844 4 : return false;
845 : }
846 :
847 : /* It looks like we're on schedule. */
848 3118 : return true;
849 : }
850 :
851 :
852 : /* --------------------------------
853 : * signal handler routines
854 : * --------------------------------
855 : */
856 :
857 : /* SIGINT: set flag to run a normal checkpoint right away */
858 : static void
859 478 : ReqCheckpointHandler(SIGNAL_ARGS)
860 : {
861 : /*
862 : * The signaling process should have set ckpt_flags nonzero, so all we
863 : * need do is ensure that our main loop gets kicked out of any wait.
864 : */
865 478 : SetLatch(MyLatch);
866 478 : }
867 :
868 :
869 : /* --------------------------------
870 : * communication with backends
871 : * --------------------------------
872 : */
873 :
874 : /*
875 : * CheckpointerShmemSize
876 : * Compute space needed for checkpointer-related shared memory
877 : */
878 : Size
879 5066 : CheckpointerShmemSize(void)
880 : {
881 : Size size;
882 :
883 : /*
884 : * Currently, the size of the requests[] array is arbitrarily set equal to
885 : * NBuffers. This may prove too large or small ...
886 : */
887 5066 : size = offsetof(CheckpointerShmemStruct, requests);
888 5066 : size = add_size(size, mul_size(NBuffers, sizeof(CheckpointerRequest)));
889 :
890 5066 : return size;
891 : }
892 :
893 : /*
894 : * CheckpointerShmemInit
895 : * Allocate and initialize checkpointer-related shared memory
896 : */
897 : void
898 1768 : CheckpointerShmemInit(void)
899 : {
900 1768 : Size size = CheckpointerShmemSize();
901 : bool found;
902 :
903 1768 : CheckpointerShmem = (CheckpointerShmemStruct *)
904 1768 : ShmemInitStruct("Checkpointer Data",
905 : size,
906 : &found);
907 :
908 1768 : if (!found)
909 : {
910 : /*
911 : * First time through, so initialize. Note that we zero the whole
912 : * requests array; this is so that CompactCheckpointerRequestQueue can
913 : * assume that any pad bytes in the request structs are zeroes.
914 : */
915 2052 : MemSet(CheckpointerShmem, 0, size);
916 1768 : SpinLockInit(&CheckpointerShmem->ckpt_lck);
917 1768 : CheckpointerShmem->max_requests = NBuffers;
918 1768 : ConditionVariableInit(&CheckpointerShmem->start_cv);
919 1768 : ConditionVariableInit(&CheckpointerShmem->done_cv);
920 : }
921 1768 : }
922 :
923 : /*
924 : * RequestCheckpoint
925 : * Called in backend processes to request a checkpoint
926 : *
927 : * flags is a bitwise OR of the following:
928 : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
929 : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
930 : * CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP,
931 : * ignoring checkpoint_completion_target parameter.
932 : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
933 : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
934 : * CHECKPOINT_END_OF_RECOVERY).
935 : * CHECKPOINT_WAIT: wait for completion before returning (otherwise,
936 : * just signal checkpointer to do it, and return).
937 : * CHECKPOINT_CAUSE_XLOG: checkpoint is requested due to xlog filling.
938 : * (This affects logging, and in particular enables CheckPointWarning.)
939 : */
940 : void
941 1334 : RequestCheckpoint(int flags)
942 : {
943 : int ntries;
944 : int old_failed,
945 : old_started;
946 :
947 : /*
948 : * If in a standalone backend, just do it ourselves.
949 : */
950 1334 : if (!IsPostmasterEnvironment)
951 : {
952 : /*
953 : * There's no point in doing slow checkpoints in a standalone backend,
954 : * because there's no other backends the checkpoint could disrupt.
955 : */
956 322 : CreateCheckPoint(flags | CHECKPOINT_IMMEDIATE);
957 :
958 : /* Free all smgr objects, as CheckpointerMain() normally would. */
959 322 : smgrdestroyall();
960 :
961 322 : return;
962 : }
963 :
964 : /*
965 : * Atomically set the request flags, and take a snapshot of the counters.
966 : * When we see ckpt_started > old_started, we know the flags we set here
967 : * have been seen by checkpointer.
968 : *
969 : * Note that we OR the flags with any existing flags, to avoid overriding
970 : * a "stronger" request by another backend. The flag senses must be
971 : * chosen to make this work!
972 : */
973 1012 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
974 :
975 1012 : old_failed = CheckpointerShmem->ckpt_failed;
976 1012 : old_started = CheckpointerShmem->ckpt_started;
977 1012 : CheckpointerShmem->ckpt_flags |= (flags | CHECKPOINT_REQUESTED);
978 :
979 1012 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
980 :
981 : /*
982 : * Send signal to request checkpoint. It's possible that the checkpointer
983 : * hasn't started yet, or is in process of restarting, so we will retry a
984 : * few times if needed. (Actually, more than a few times, since on slow
985 : * or overloaded buildfarm machines, it's been observed that the
986 : * checkpointer can take several seconds to start.) However, if not told
987 : * to wait for the checkpoint to occur, we consider failure to send the
988 : * signal to be nonfatal and merely LOG it. The checkpointer should see
989 : * the request when it does start, with or without getting a signal.
990 : */
991 : #define MAX_SIGNAL_TRIES 600 /* max wait 60.0 sec */
992 1020 : for (ntries = 0;; ntries++)
993 : {
994 1020 : if (CheckpointerShmem->checkpointer_pid == 0)
995 : {
996 8 : if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT))
997 : {
998 0 : elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG,
999 : "could not signal for checkpoint: checkpointer is not running");
1000 0 : break;
1001 : }
1002 : }
1003 1012 : else if (kill(CheckpointerShmem->checkpointer_pid, SIGINT) != 0)
1004 : {
1005 0 : if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT))
1006 : {
1007 0 : elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG,
1008 : "could not signal for checkpoint: %m");
1009 0 : break;
1010 : }
1011 : }
1012 : else
1013 1012 : break; /* signal sent successfully */
1014 :
1015 8 : CHECK_FOR_INTERRUPTS();
1016 8 : pg_usleep(100000L); /* wait 0.1 sec, then retry */
1017 : }
1018 :
1019 : /*
1020 : * If requested, wait for completion. We detect completion according to
1021 : * the algorithm given above.
1022 : */
1023 1012 : if (flags & CHECKPOINT_WAIT)
1024 : {
1025 : int new_started,
1026 : new_failed;
1027 :
1028 : /* Wait for a new checkpoint to start. */
1029 820 : ConditionVariablePrepareToSleep(&CheckpointerShmem->start_cv);
1030 : for (;;)
1031 : {
1032 1510 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1033 1510 : new_started = CheckpointerShmem->ckpt_started;
1034 1510 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1035 :
1036 1510 : if (new_started != old_started)
1037 820 : break;
1038 :
1039 690 : ConditionVariableSleep(&CheckpointerShmem->start_cv,
1040 : WAIT_EVENT_CHECKPOINT_START);
1041 : }
1042 820 : ConditionVariableCancelSleep();
1043 :
1044 : /*
1045 : * We are waiting for ckpt_done >= new_started, in a modulo sense.
1046 : */
1047 820 : ConditionVariablePrepareToSleep(&CheckpointerShmem->done_cv);
1048 : for (;;)
1049 642 : {
1050 : int new_done;
1051 :
1052 1462 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1053 1462 : new_done = CheckpointerShmem->ckpt_done;
1054 1462 : new_failed = CheckpointerShmem->ckpt_failed;
1055 1462 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1056 :
1057 1462 : if (new_done - new_started >= 0)
1058 820 : break;
1059 :
1060 642 : ConditionVariableSleep(&CheckpointerShmem->done_cv,
1061 : WAIT_EVENT_CHECKPOINT_DONE);
1062 : }
1063 820 : ConditionVariableCancelSleep();
1064 :
1065 820 : if (new_failed != old_failed)
1066 0 : ereport(ERROR,
1067 : (errmsg("checkpoint request failed"),
1068 : errhint("Consult recent messages in the server log for details.")));
1069 : }
1070 : }
1071 :
1072 : /*
1073 : * ForwardSyncRequest
1074 : * Forward a file-fsync request from a backend to the checkpointer
1075 : *
1076 : * Whenever a backend is compelled to write directly to a relation
1077 : * (which should be seldom, if the background writer is getting its job done),
1078 : * the backend calls this routine to pass over knowledge that the relation
1079 : * is dirty and must be fsync'd before next checkpoint. We also use this
1080 : * opportunity to count such writes for statistical purposes.
1081 : *
1082 : * To avoid holding the lock for longer than necessary, we normally write
1083 : * to the requests[] queue without checking for duplicates. The checkpointer
1084 : * will have to eliminate dups internally anyway. However, if we discover
1085 : * that the queue is full, we make a pass over the entire queue to compact
1086 : * it. This is somewhat expensive, but the alternative is for the backend
1087 : * to perform its own fsync, which is far more expensive in practice. It
1088 : * is theoretically possible a backend fsync might still be necessary, if
1089 : * the queue is full and contains no duplicate entries. In that case, we
1090 : * let the backend know by returning false.
1091 : */
1092 : bool
1093 1374622 : ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)
1094 : {
1095 : CheckpointerRequest *request;
1096 : bool too_full;
1097 :
1098 1374622 : if (!IsUnderPostmaster)
1099 0 : return false; /* probably shouldn't even get here */
1100 :
1101 1374622 : if (AmCheckpointerProcess())
1102 0 : elog(ERROR, "ForwardSyncRequest must not be called in checkpointer");
1103 :
1104 1374622 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1105 :
1106 : /*
1107 : * If the checkpointer isn't running or the request queue is full, the
1108 : * backend will have to perform its own fsync request. But before forcing
1109 : * that to happen, we can try to compact the request queue.
1110 : */
1111 1374622 : if (CheckpointerShmem->checkpointer_pid == 0 ||
1112 1374568 : (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&
1113 98 : !CompactCheckpointerRequestQueue()))
1114 : {
1115 98 : LWLockRelease(CheckpointerCommLock);
1116 98 : return false;
1117 : }
1118 :
1119 : /* OK, insert request */
1120 1374524 : request = &CheckpointerShmem->requests[CheckpointerShmem->num_requests++];
1121 1374524 : request->ftag = *ftag;
1122 1374524 : request->type = type;
1123 :
1124 : /* If queue is more than half full, nudge the checkpointer to empty it */
1125 1374524 : too_full = (CheckpointerShmem->num_requests >=
1126 1374524 : CheckpointerShmem->max_requests / 2);
1127 :
1128 1374524 : LWLockRelease(CheckpointerCommLock);
1129 :
1130 : /* ... but not till after we release the lock */
1131 1374524 : if (too_full && ProcGlobal->checkpointerLatch)
1132 26790 : SetLatch(ProcGlobal->checkpointerLatch);
1133 :
1134 1374524 : return true;
1135 : }
1136 :
1137 : /*
1138 : * CompactCheckpointerRequestQueue
1139 : * Remove duplicates from the request queue to avoid backend fsyncs.
1140 : * Returns "true" if any entries were removed.
1141 : *
1142 : * Although a full fsync request queue is not common, it can lead to severe
1143 : * performance problems when it does happen. So far, this situation has
1144 : * only been observed to occur when the system is under heavy write load,
1145 : * and especially during the "sync" phase of a checkpoint. Without this
1146 : * logic, each backend begins doing an fsync for every block written, which
1147 : * gets very expensive and can slow down the whole system.
1148 : *
1149 : * Trying to do this every time the queue is full could lose if there
1150 : * aren't any removable entries. But that should be vanishingly rare in
1151 : * practice: there's one queue entry per shared buffer.
1152 : */
1153 : static bool
1154 98 : CompactCheckpointerRequestQueue(void)
1155 : {
1156 : struct CheckpointerSlotMapping
1157 : {
1158 : CheckpointerRequest request;
1159 : int slot;
1160 : };
1161 :
1162 : int n,
1163 : preserve_count;
1164 98 : int num_skipped = 0;
1165 : HASHCTL ctl;
1166 : HTAB *htab;
1167 : bool *skip_slot;
1168 :
1169 : /* must hold CheckpointerCommLock in exclusive mode */
1170 : Assert(LWLockHeldByMe(CheckpointerCommLock));
1171 :
1172 : /* Initialize skip_slot array */
1173 98 : skip_slot = palloc0(sizeof(bool) * CheckpointerShmem->num_requests);
1174 :
1175 : /* Initialize temporary hash table */
1176 98 : ctl.keysize = sizeof(CheckpointerRequest);
1177 98 : ctl.entrysize = sizeof(struct CheckpointerSlotMapping);
1178 98 : ctl.hcxt = CurrentMemoryContext;
1179 :
1180 98 : htab = hash_create("CompactCheckpointerRequestQueue",
1181 98 : CheckpointerShmem->num_requests,
1182 : &ctl,
1183 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1184 :
1185 : /*
1186 : * The basic idea here is that a request can be skipped if it's followed
1187 : * by a later, identical request. It might seem more sensible to work
1188 : * backwards from the end of the queue and check whether a request is
1189 : * *preceded* by an earlier, identical request, in the hopes of doing less
1190 : * copying. But that might change the semantics, if there's an
1191 : * intervening SYNC_FORGET_REQUEST or SYNC_FILTER_REQUEST, so we do it
1192 : * this way. It would be possible to be even smarter if we made the code
1193 : * below understand the specific semantics of such requests (it could blow
1194 : * away preceding entries that would end up being canceled anyhow), but
1195 : * it's not clear that the extra complexity would buy us anything.
1196 : */
1197 12642 : for (n = 0; n < CheckpointerShmem->num_requests; n++)
1198 : {
1199 : CheckpointerRequest *request;
1200 : struct CheckpointerSlotMapping *slotmap;
1201 : bool found;
1202 :
1203 : /*
1204 : * We use the request struct directly as a hashtable key. This
1205 : * assumes that any padding bytes in the structs are consistently the
1206 : * same, which should be okay because we zeroed them in
1207 : * CheckpointerShmemInit. Note also that RelFileLocator had better
1208 : * contain no pad bytes.
1209 : */
1210 12544 : request = &CheckpointerShmem->requests[n];
1211 12544 : slotmap = hash_search(htab, request, HASH_ENTER, &found);
1212 12544 : if (found)
1213 : {
1214 : /* Duplicate, so mark the previous occurrence as skippable */
1215 3218 : skip_slot[slotmap->slot] = true;
1216 3218 : num_skipped++;
1217 : }
1218 : /* Remember slot containing latest occurrence of this request value */
1219 12544 : slotmap->slot = n;
1220 : }
1221 :
1222 : /* Done with the hash table. */
1223 98 : hash_destroy(htab);
1224 :
1225 : /* If no duplicates, we're out of luck. */
1226 98 : if (!num_skipped)
1227 : {
1228 44 : pfree(skip_slot);
1229 44 : return false;
1230 : }
1231 :
1232 : /* We found some duplicates; remove them. */
1233 54 : preserve_count = 0;
1234 6966 : for (n = 0; n < CheckpointerShmem->num_requests; n++)
1235 : {
1236 6912 : if (skip_slot[n])
1237 3218 : continue;
1238 3694 : CheckpointerShmem->requests[preserve_count++] = CheckpointerShmem->requests[n];
1239 : }
1240 54 : ereport(DEBUG1,
1241 : (errmsg_internal("compacted fsync request queue from %d entries to %d entries",
1242 : CheckpointerShmem->num_requests, preserve_count)));
1243 54 : CheckpointerShmem->num_requests = preserve_count;
1244 :
1245 : /* Cleanup. */
1246 54 : pfree(skip_slot);
1247 54 : return true;
1248 : }
1249 :
1250 : /*
1251 : * AbsorbSyncRequests
1252 : * Retrieve queued sync requests and pass them to sync mechanism.
1253 : *
1254 : * This is exported because it must be called during CreateCheckPoint;
1255 : * we have to be sure we have accepted all pending requests just before
1256 : * we start fsync'ing. Since CreateCheckPoint sometimes runs in
1257 : * non-checkpointer processes, do nothing if not checkpointer.
1258 : */
1259 : void
1260 24542 : AbsorbSyncRequests(void)
1261 : {
1262 24542 : CheckpointerRequest *requests = NULL;
1263 : CheckpointerRequest *request;
1264 : int n;
1265 :
1266 24542 : if (!AmCheckpointerProcess())
1267 1000 : return;
1268 :
1269 23542 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1270 :
1271 : /*
1272 : * We try to avoid holding the lock for a long time by copying the request
1273 : * array, and processing the requests after releasing the lock.
1274 : *
1275 : * Once we have cleared the requests from shared memory, we have to PANIC
1276 : * if we then fail to absorb them (eg, because our hashtable runs out of
1277 : * memory). This is because the system cannot run safely if we are unable
1278 : * to fsync what we have been told to fsync. Fortunately, the hashtable
1279 : * is so small that the problem is quite unlikely to arise in practice.
1280 : */
1281 23542 : n = CheckpointerShmem->num_requests;
1282 23542 : if (n > 0)
1283 : {
1284 13256 : requests = (CheckpointerRequest *) palloc(n * sizeof(CheckpointerRequest));
1285 13256 : memcpy(requests, CheckpointerShmem->requests, n * sizeof(CheckpointerRequest));
1286 : }
1287 :
1288 23542 : START_CRIT_SECTION();
1289 :
1290 23542 : CheckpointerShmem->num_requests = 0;
1291 :
1292 23542 : LWLockRelease(CheckpointerCommLock);
1293 :
1294 1222606 : for (request = requests; n > 0; request++, n--)
1295 1199064 : RememberSyncRequest(&request->ftag, request->type);
1296 :
1297 23542 : END_CRIT_SECTION();
1298 :
1299 23542 : if (requests)
1300 13256 : pfree(requests);
1301 : }
1302 :
1303 : /*
1304 : * Update any shared memory configurations based on config parameters
1305 : */
1306 : static void
1307 924 : UpdateSharedMemoryConfig(void)
1308 : {
1309 : /* update global shmem state for sync rep */
1310 924 : SyncRepUpdateSyncStandbysDefined();
1311 :
1312 : /*
1313 : * If full_page_writes has been changed by SIGHUP, we update it in shared
1314 : * memory and write an XLOG_FPW_CHANGE record.
1315 : */
1316 924 : UpdateFullPageWrites();
1317 :
1318 924 : elog(DEBUG2, "checkpointer updated shared memory configuration values");
1319 924 : }
1320 :
1321 : /*
1322 : * FirstCallSinceLastCheckpoint allows a process to take an action once
1323 : * per checkpoint cycle by asynchronously checking for checkpoint completion.
1324 : */
1325 : bool
1326 13022 : FirstCallSinceLastCheckpoint(void)
1327 : {
1328 : static int ckpt_done = 0;
1329 : int new_done;
1330 13022 : bool FirstCall = false;
1331 :
1332 13022 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1333 13022 : new_done = CheckpointerShmem->ckpt_done;
1334 13022 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1335 :
1336 13022 : if (new_done != ckpt_done)
1337 324 : FirstCall = true;
1338 :
1339 13022 : ckpt_done = new_done;
1340 :
1341 13022 : return FirstCall;
1342 : }
|