Branch data Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * checkpointer.c
4 : : *
5 : : * The checkpointer is new as of Postgres 9.2. It handles all checkpoints.
6 : : * Checkpoints are automatically dispatched after a certain amount of time has
7 : : * elapsed since the last one, and it can be signaled to perform requested
8 : : * checkpoints as well. (The GUC parameter that mandates a checkpoint every
9 : : * so many WAL segments is implemented by having backends signal when they
10 : : * fill WAL segments; the checkpointer itself doesn't watch for the
11 : : * condition.)
12 : : *
13 : : * The normal termination sequence is that checkpointer is instructed to
14 : : * execute the shutdown checkpoint by SIGINT. After that checkpointer waits
15 : : * to be terminated via SIGUSR2, which instructs the checkpointer to exit(0).
16 : : * All backends must be stopped before SIGINT or SIGUSR2 is issued!
17 : : *
18 : : * Emergency termination is by SIGQUIT; like any backend, the checkpointer
19 : : * will simply abort and exit on SIGQUIT.
20 : : *
21 : : * If the checkpointer exits unexpectedly, the postmaster treats that the same
22 : : * as a backend crash: shared memory may be corrupted, so remaining backends
23 : : * should be killed by SIGQUIT and then a recovery cycle started. (Even if
24 : : * shared memory isn't corrupted, we have lost information about which
25 : : * files need to be fsync'd for the next checkpoint, and so a system
26 : : * restart needs to be forced.)
27 : : *
28 : : *
29 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
30 : : *
31 : : *
32 : : * IDENTIFICATION
33 : : * src/backend/postmaster/checkpointer.c
34 : : *
35 : : *-------------------------------------------------------------------------
36 : : */
37 : : #include "postgres.h"
38 : :
39 : : #include <sys/time.h>
40 : : #include <time.h>
41 : :
42 : : #include "access/xlog.h"
43 : : #include "access/xlog_internal.h"
44 : : #include "access/xlogrecovery.h"
45 : : #include "catalog/pg_authid.h"
46 : : #include "commands/defrem.h"
47 : : #include "libpq/pqsignal.h"
48 : : #include "miscadmin.h"
49 : : #include "pgstat.h"
50 : : #include "port/atomics.h"
51 : : #include "postmaster/auxprocess.h"
52 : : #include "postmaster/bgwriter.h"
53 : : #include "postmaster/interrupt.h"
54 : : #include "replication/syncrep.h"
55 : : #include "storage/aio_subsys.h"
56 : : #include "storage/bufmgr.h"
57 : : #include "storage/condition_variable.h"
58 : : #include "storage/fd.h"
59 : : #include "storage/ipc.h"
60 : : #include "storage/lwlock.h"
61 : : #include "storage/pmsignal.h"
62 : : #include "storage/proc.h"
63 : : #include "storage/procsignal.h"
64 : : #include "storage/shmem.h"
65 : : #include "storage/smgr.h"
66 : : #include "storage/spin.h"
67 : : #include "storage/subsystems.h"
68 : : #include "utils/acl.h"
69 : : #include "utils/guc.h"
70 : : #include "utils/memutils.h"
71 : : #include "utils/resowner.h"
72 : : #include "utils/wait_event.h"
73 : :
74 : :
75 : : /*----------
76 : : * Shared memory area for communication between checkpointer and backends
77 : : *
78 : : * The ckpt counters allow backends to watch for completion of a checkpoint
79 : : * request they send. Here's how it works:
80 : : * * At start of a checkpoint, checkpointer reads (and clears) the request
81 : : * flags and increments ckpt_started, while holding ckpt_lck.
82 : : * * On completion of a checkpoint, checkpointer sets ckpt_done to
83 : : * equal ckpt_started.
84 : : * * On failure of a checkpoint, checkpointer increments ckpt_failed
85 : : * and sets ckpt_done to equal ckpt_started.
86 : : *
87 : : * The algorithm for backends is:
88 : : * 1. Record current values of ckpt_failed and ckpt_started, and
89 : : * set request flags, while holding ckpt_lck.
90 : : * 2. Send signal to request checkpoint.
91 : : * 3. Sleep until ckpt_started changes. Now you know a checkpoint has
92 : : * begun since you started this algorithm (although *not* that it was
93 : : * specifically initiated by your signal), and that it is using your flags.
94 : : * 4. Record new value of ckpt_started.
95 : : * 5. Sleep until ckpt_done >= saved value of ckpt_started. (Use modulo
96 : : * arithmetic here in case counters wrap around.) Now you know a
97 : : * checkpoint has started and completed, but not whether it was
98 : : * successful.
99 : : * 6. If ckpt_failed is different from the originally saved value,
100 : : * assume request failed; otherwise it was definitely successful.
101 : : *
102 : : * ckpt_flags holds the OR of the checkpoint request flags sent by all
103 : : * requesting backends since the last checkpoint start. The flags are
104 : : * chosen so that OR'ing is the correct way to combine multiple requests.
105 : : *
106 : : * The requests array holds fsync requests sent by backends and not yet
107 : : * absorbed by the checkpointer.
108 : : *
109 : : * Unlike the checkpoint fields, requests related fields are protected by
110 : : * CheckpointerCommLock.
111 : : *----------
112 : : */
113 : : typedef struct
114 : : {
115 : : SyncRequestType type; /* request type */
116 : : FileTag ftag; /* file identifier */
117 : : } CheckpointerRequest;
118 : :
119 : : typedef struct
120 : : {
121 : : pid_t checkpointer_pid; /* PID (0 if not started) */
122 : :
123 : : slock_t ckpt_lck; /* protects all the ckpt_* fields */
124 : :
125 : : int ckpt_started; /* advances when checkpoint starts */
126 : : int ckpt_done; /* advances when checkpoint done */
127 : : int ckpt_failed; /* advances when checkpoint fails */
128 : :
129 : : int ckpt_flags; /* checkpoint flags, as defined in xlog.h */
130 : :
131 : : ConditionVariable start_cv; /* signaled when ckpt_started advances */
132 : : ConditionVariable done_cv; /* signaled when ckpt_done advances */
133 : :
134 : : int num_requests; /* current # of requests */
135 : : int max_requests; /* allocated array size */
136 : :
137 : : int head; /* Index of the first request in the ring
138 : : * buffer */
139 : : int tail; /* Index of the last request in the ring
140 : : * buffer */
141 : :
142 : : /* The ring buffer of pending checkpointer requests */
143 : : CheckpointerRequest requests[FLEXIBLE_ARRAY_MEMBER];
144 : : } CheckpointerShmemStruct;
145 : :
146 : : static CheckpointerShmemStruct *CheckpointerShmem;
147 : :
148 : : static void CheckpointerShmemRequest(void *arg);
149 : : static void CheckpointerShmemInit(void *arg);
150 : :
151 : : const ShmemCallbacks CheckpointerShmemCallbacks = {
152 : : .request_fn = CheckpointerShmemRequest,
153 : : .init_fn = CheckpointerShmemInit,
154 : : };
155 : :
156 : : /* interval for calling AbsorbSyncRequests in CheckpointWriteDelay */
157 : : #define WRITES_PER_ABSORB 1000
158 : :
159 : : /* Maximum number of checkpointer requests to process in one batch */
160 : : #define CKPT_REQ_BATCH_SIZE 10000
161 : :
162 : : /* Max number of requests the checkpointer request queue can hold */
163 : : #define MAX_CHECKPOINT_REQUESTS 10000000
164 : :
165 : : /*
166 : : * GUC parameters
167 : : */
168 : : int CheckPointTimeout = 300;
169 : : int CheckPointWarning = 30;
170 : : double CheckPointCompletionTarget = 0.9;
171 : :
172 : : /*
173 : : * Private state
174 : : */
175 : : static bool ckpt_active = false;
176 : : static volatile sig_atomic_t ShutdownXLOGPending = false;
177 : :
178 : : /* these values are valid when ckpt_active is true: */
179 : : static pg_time_t ckpt_start_time;
180 : : static XLogRecPtr ckpt_start_recptr;
181 : : static double ckpt_cached_elapsed;
182 : :
183 : : static pg_time_t last_checkpoint_time;
184 : : static pg_time_t last_xlog_switch_time;
185 : :
186 : : /* Prototypes for private functions */
187 : :
188 : : static void ProcessCheckpointerInterrupts(void);
189 : : static void CheckArchiveTimeout(void);
190 : : static bool IsCheckpointOnSchedule(double progress);
191 : : static bool FastCheckpointRequested(void);
192 : : static bool CompactCheckpointerRequestQueue(void);
193 : : static void UpdateSharedMemoryConfig(void);
194 : :
195 : : /* Signal handlers */
196 : : static void ReqShutdownXLOG(SIGNAL_ARGS);
197 : :
198 : :
199 : : /*
200 : : * Main entry point for checkpointer process
201 : : *
202 : : * This is invoked from AuxiliaryProcessMain, which has already created the
203 : : * basic execution environment, but not enabled signals yet.
204 : : */
205 : : void
206 : 642 : CheckpointerMain(const void *startup_data, size_t startup_data_len)
207 : : {
208 : : sigjmp_buf local_sigjmp_buf;
209 : : MemoryContext checkpointer_context;
210 : :
211 : : Assert(startup_data_len == 0);
212 : :
213 : 642 : AuxiliaryProcessMainCommon();
214 : :
215 : 642 : CheckpointerShmem->checkpointer_pid = MyProcPid;
216 : :
217 : : /*
218 : : * Properly accept or ignore signals the postmaster might send us
219 : : *
220 : : * Note: we deliberately ignore SIGTERM, because during a standard Unix
221 : : * system shutdown cycle, init will SIGTERM all processes at once. We
222 : : * want to wait for the backends to exit, whereupon the postmaster will
223 : : * tell us it's okay to shut down (via SIGUSR2).
224 : : */
225 : 642 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
226 : 642 : pqsignal(SIGINT, ReqShutdownXLOG);
227 : 642 : pqsignal(SIGTERM, PG_SIG_IGN); /* ignore SIGTERM */
228 : : /* SIGQUIT handler was already set up by InitPostmasterChild */
229 : 642 : pqsignal(SIGALRM, PG_SIG_IGN);
230 : 642 : pqsignal(SIGPIPE, PG_SIG_IGN);
231 : 642 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
232 : 642 : pqsignal(SIGUSR2, SignalHandlerForShutdownRequest);
233 : :
234 : : /*
235 : : * Reset some signals that are accepted by postmaster but not here
236 : : */
237 : 642 : pqsignal(SIGCHLD, PG_SIG_DFL);
238 : :
239 : : /*
240 : : * Initialize so that first time-driven event happens at the correct time.
241 : : */
242 : 642 : last_checkpoint_time = last_xlog_switch_time = (pg_time_t) time(NULL);
243 : :
244 : : /*
245 : : * Write out stats after shutdown. This needs to be called by exactly one
246 : : * process during a normal shutdown, and since checkpointer is shut down
247 : : * very late...
248 : : *
249 : : * While e.g. walsenders are active after the shutdown checkpoint has been
250 : : * written (and thus could produce more stats), checkpointer stays around
251 : : * after the shutdown checkpoint has been written. postmaster will only
252 : : * signal checkpointer to exit after all processes that could emit stats
253 : : * have been shut down.
254 : : */
255 : 642 : before_shmem_exit(pgstat_before_server_shutdown, 0);
256 : :
257 : : /*
258 : : * Create a memory context that we will do all our work in. We do this so
259 : : * that we can reset the context during error recovery and thereby avoid
260 : : * possible memory leaks. Formerly this code just ran in
261 : : * TopMemoryContext, but resetting that would be a really bad idea.
262 : : */
263 : 642 : checkpointer_context = AllocSetContextCreate(TopMemoryContext,
264 : : "Checkpointer",
265 : : ALLOCSET_DEFAULT_SIZES);
266 : 642 : MemoryContextSwitchTo(checkpointer_context);
267 : :
268 : : /*
269 : : * If an exception is encountered, processing resumes here.
270 : : *
271 : : * You might wonder why this isn't coded as an infinite loop around a
272 : : * PG_TRY construct. The reason is that this is the bottom of the
273 : : * exception stack, and so with PG_TRY there would be no exception handler
274 : : * in force at all during the CATCH part. By leaving the outermost setjmp
275 : : * always active, we have at least some chance of recovering from an error
276 : : * during error recovery. (If we get into an infinite loop thereby, it
277 : : * will soon be stopped by overflow of elog.c's internal state stack.)
278 : : *
279 : : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
280 : : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
281 : : * signals other than SIGQUIT will be blocked until we complete error
282 : : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
283 : : * call redundant, but it is not since InterruptPending might be set
284 : : * already.
285 : : */
286 [ - + ]: 642 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
287 : : {
288 : : /* Since not using PG_TRY, must reset error stack by hand */
289 : 0 : error_context_stack = NULL;
290 : :
291 : : /* Prevent interrupts while cleaning up */
292 : 0 : HOLD_INTERRUPTS();
293 : :
294 : : /* Report the error to the server log */
295 : 0 : EmitErrorReport();
296 : :
297 : : /*
298 : : * These operations are really just a minimal subset of
299 : : * AbortTransaction(). We don't have very many resources to worry
300 : : * about in checkpointer, but we do have LWLocks, buffers, and temp
301 : : * files.
302 : : */
303 : 0 : LWLockReleaseAll();
304 : 0 : ConditionVariableCancelSleep();
305 : 0 : pgstat_report_wait_end();
306 : 0 : pgaio_error_cleanup();
307 : 0 : UnlockBuffers();
308 : 0 : ReleaseAuxProcessResources(false);
309 : 0 : AtEOXact_Buffers(false);
310 : 0 : AtEOXact_SMgr();
311 : 0 : AtEOXact_Files(false);
312 : 0 : AtEOXact_HashTables(false);
313 : :
314 : : /* Warn any waiting backends that the checkpoint failed. */
315 [ # # ]: 0 : if (ckpt_active)
316 : : {
317 : 0 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
318 : 0 : CheckpointerShmem->ckpt_failed++;
319 : 0 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
320 : 0 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
321 : :
322 : 0 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
323 : :
324 : 0 : ckpt_active = false;
325 : : }
326 : :
327 : : /*
328 : : * Now return to normal top-level context and clear ErrorContext for
329 : : * next time.
330 : : */
331 : 0 : MemoryContextSwitchTo(checkpointer_context);
332 : 0 : FlushErrorState();
333 : :
334 : : /* Flush any leaked data in the top-level context */
335 : 0 : MemoryContextReset(checkpointer_context);
336 : :
337 : : /* Now we can allow interrupts again */
338 : 0 : RESUME_INTERRUPTS();
339 : :
340 : : /*
341 : : * Sleep at least 1 second after any error. A write error is likely
342 : : * to be repeated, and we don't want to be filling the error logs as
343 : : * fast as we can.
344 : : */
345 : 0 : pg_usleep(1000000L);
346 : : }
347 : :
348 : : /* We can now handle ereport(ERROR) */
349 : 642 : PG_exception_stack = &local_sigjmp_buf;
350 : :
351 : : /*
352 : : * Unblock signals (they were blocked when the postmaster forked us)
353 : : */
354 : 642 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
355 : :
356 : : /*
357 : : * Ensure all shared memory values are set correctly for the config. Doing
358 : : * this here ensures no race conditions from other concurrent updaters.
359 : : */
360 : 642 : UpdateSharedMemoryConfig();
361 : :
362 : : /*
363 : : * Loop until we've been asked to write the shutdown checkpoint or
364 : : * terminate.
365 : : */
366 : : for (;;)
367 : 5888 : {
368 : 6530 : bool do_checkpoint = false;
369 : 6530 : int flags = 0;
370 : : pg_time_t now;
371 : : int elapsed_secs;
372 : : int cur_timeout;
373 : 6530 : bool chkpt_or_rstpt_requested = false;
374 : 6530 : bool chkpt_or_rstpt_timed = false;
375 : :
376 : : /* Clear any already-pending wakeups */
377 : 6530 : ResetLatch(MyLatch);
378 : :
379 : : /*
380 : : * Process any requests or signals received recently.
381 : : */
382 : 6530 : AbsorbSyncRequests();
383 : :
384 : 6530 : ProcessCheckpointerInterrupts();
385 [ + + + - ]: 6530 : if (ShutdownXLOGPending || ShutdownRequestPending)
386 : : break;
387 : :
388 : : /*
389 : : * Detect a pending checkpoint request by checking whether the flags
390 : : * word in shared memory is nonzero. We shouldn't need to acquire the
391 : : * ckpt_lck for this.
392 : : */
393 [ + + ]: 5896 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
394 : : {
395 : 1361 : do_checkpoint = true;
396 : 1361 : chkpt_or_rstpt_requested = true;
397 : : }
398 : :
399 : : /*
400 : : * Force a checkpoint if too much time has elapsed since the last one.
401 : : * Note that we count a timed checkpoint in stats only when this
402 : : * occurs without an external request, but we set the CAUSE_TIME flag
403 : : * bit even if there is also an external request.
404 : : */
405 : 5896 : now = (pg_time_t) time(NULL);
406 : 5896 : elapsed_secs = now - last_checkpoint_time;
407 [ + + ]: 5896 : if (elapsed_secs >= CheckPointTimeout)
408 : : {
409 [ + - ]: 1 : if (!do_checkpoint)
410 : 1 : chkpt_or_rstpt_timed = true;
411 : 1 : do_checkpoint = true;
412 : 1 : flags |= CHECKPOINT_CAUSE_TIME;
413 : : }
414 : :
415 : : /*
416 : : * Do a checkpoint if requested.
417 : : */
418 [ + + ]: 5896 : if (do_checkpoint)
419 : : {
420 : 1362 : bool ckpt_performed = false;
421 : : bool do_restartpoint;
422 : :
423 : : /* Check if we should perform a checkpoint or a restartpoint. */
424 : 1362 : do_restartpoint = RecoveryInProgress();
425 : :
426 : : /*
427 : : * Atomically fetch the request flags to figure out what kind of a
428 : : * checkpoint we should perform, and increase the started-counter
429 : : * to acknowledge that we've started a new checkpoint.
430 : : */
431 : 1362 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
432 : 1362 : flags |= CheckpointerShmem->ckpt_flags;
433 : 1362 : CheckpointerShmem->ckpt_flags = 0;
434 : 1362 : CheckpointerShmem->ckpt_started++;
435 : 1362 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
436 : :
437 : 1362 : ConditionVariableBroadcast(&CheckpointerShmem->start_cv);
438 : :
439 : : /*
440 : : * The end-of-recovery checkpoint is a real checkpoint that's
441 : : * performed while we're still in recovery.
442 : : */
443 [ + + ]: 1362 : if (flags & CHECKPOINT_END_OF_RECOVERY)
444 : 22 : do_restartpoint = false;
445 : :
446 [ + + ]: 1362 : if (chkpt_or_rstpt_timed)
447 : : {
448 : 1 : chkpt_or_rstpt_timed = false;
449 [ - + ]: 1 : if (do_restartpoint)
450 : 0 : PendingCheckpointerStats.restartpoints_timed++;
451 : : else
452 : 1 : PendingCheckpointerStats.num_timed++;
453 : : }
454 : :
455 [ + + ]: 1362 : if (chkpt_or_rstpt_requested)
456 : : {
457 : 1361 : chkpt_or_rstpt_requested = false;
458 [ + + ]: 1361 : if (do_restartpoint)
459 : 557 : PendingCheckpointerStats.restartpoints_requested++;
460 : : else
461 : 804 : PendingCheckpointerStats.num_requested++;
462 : : }
463 : :
464 : : /*
465 : : * We will warn if (a) too soon since last checkpoint (whatever
466 : : * caused it) and (b) somebody set the CHECKPOINT_CAUSE_XLOG flag
467 : : * since the last checkpoint start. Note in particular that this
468 : : * implementation will not generate warnings caused by
469 : : * CheckPointTimeout < CheckPointWarning.
470 : : */
471 [ + + ]: 1362 : if (!do_restartpoint &&
472 [ + + ]: 805 : (flags & CHECKPOINT_CAUSE_XLOG) &&
473 [ + - ]: 211 : elapsed_secs < CheckPointWarning)
474 [ + - ]: 211 : ereport(LOG,
475 : : (errmsg_plural("checkpoints are occurring too frequently (%d second apart)",
476 : : "checkpoints are occurring too frequently (%d seconds apart)",
477 : : elapsed_secs,
478 : : elapsed_secs),
479 : : errhint("Consider increasing the configuration parameter \"%s\".", "max_wal_size")));
480 : :
481 : : /*
482 : : * Initialize checkpointer-private variables used during
483 : : * checkpoint.
484 : : */
485 : 1362 : ckpt_active = true;
486 [ + + ]: 1362 : if (do_restartpoint)
487 : 557 : ckpt_start_recptr = GetXLogReplayRecPtr(NULL);
488 : : else
489 : 805 : ckpt_start_recptr = GetInsertRecPtr();
490 : 1362 : ckpt_start_time = now;
491 : 1362 : ckpt_cached_elapsed = 0;
492 : :
493 : : /*
494 : : * Do the checkpoint.
495 : : */
496 [ + + ]: 1362 : if (!do_restartpoint)
497 : 805 : ckpt_performed = CreateCheckPoint(flags);
498 : : else
499 : 557 : ckpt_performed = CreateRestartPoint(flags);
500 : :
501 : : /*
502 : : * After any checkpoint, free all smgr objects. Otherwise we
503 : : * would never do so for dropped relations, as the checkpointer
504 : : * does not process shared invalidation messages or call
505 : : * AtEOXact_SMgr().
506 : : */
507 : 1362 : smgrdestroyall();
508 : :
509 : : /*
510 : : * Indicate checkpoint completion to any waiting backends.
511 : : */
512 : 1362 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
513 : 1362 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
514 : 1362 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
515 : :
516 : 1362 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
517 : :
518 [ + + ]: 1362 : if (!do_restartpoint)
519 : : {
520 : : /*
521 : : * Note we record the checkpoint start time not end time as
522 : : * last_checkpoint_time. This is so that time-driven
523 : : * checkpoints happen at a predictable spacing.
524 : : */
525 : 805 : last_checkpoint_time = now;
526 : :
527 [ + + ]: 805 : if (ckpt_performed)
528 : 800 : PendingCheckpointerStats.num_performed++;
529 : : }
530 : : else
531 : : {
532 [ + + ]: 557 : if (ckpt_performed)
533 : : {
534 : : /*
535 : : * The same as for checkpoint. Please see the
536 : : * corresponding comment.
537 : : */
538 : 188 : last_checkpoint_time = now;
539 : :
540 : 188 : PendingCheckpointerStats.restartpoints_performed++;
541 : : }
542 : : else
543 : : {
544 : : /*
545 : : * We were not able to perform the restartpoint
546 : : * (checkpoints throw an ERROR in case of error). Most
547 : : * likely because we have not received any new checkpoint
548 : : * WAL records since the last restartpoint. Try again in
549 : : * 15 s.
550 : : */
551 : 369 : last_checkpoint_time = now - CheckPointTimeout + 15;
552 : : }
553 : : }
554 : :
555 : 1362 : ckpt_active = false;
556 : :
557 : : /*
558 : : * We may have received an interrupt during the checkpoint and the
559 : : * latch might have been reset (e.g. in CheckpointWriteDelay).
560 : : */
561 : 1362 : ProcessCheckpointerInterrupts();
562 [ + + + - ]: 1362 : if (ShutdownXLOGPending || ShutdownRequestPending)
563 : : break;
564 : : }
565 : :
566 : : /*
567 : : * Disable logical decoding if someone requested it. See comments atop
568 : : * logicalctl.c.
569 : : */
570 : 5891 : DisableLogicalDecodingIfNecessary();
571 : :
572 : : /* Check for archive_timeout and switch xlog files if necessary. */
573 : 5891 : CheckArchiveTimeout();
574 : :
575 : : /* Report pending statistics to the cumulative stats system */
576 : 5891 : pgstat_report_checkpointer();
577 : 5891 : pgstat_report_wal(true);
578 : :
579 : : /*
580 : : * If any checkpoint flags have been set, redo the loop to handle the
581 : : * checkpoint without sleeping.
582 : : */
583 [ + + ]: 5891 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
584 : 283 : continue;
585 : :
586 : : /*
587 : : * Sleep until we are signaled or it's time for another checkpoint or
588 : : * xlog file switch.
589 : : */
590 : 5608 : now = (pg_time_t) time(NULL);
591 : 5608 : elapsed_secs = now - last_checkpoint_time;
592 [ - + ]: 5608 : if (elapsed_secs >= CheckPointTimeout)
593 : 0 : continue; /* no sleep for us ... */
594 : 5608 : cur_timeout = CheckPointTimeout - elapsed_secs;
595 [ - + - - ]: 5608 : if (XLogArchiveTimeout > 0 && !RecoveryInProgress())
596 : : {
597 : 0 : elapsed_secs = now - last_xlog_switch_time;
598 [ # # ]: 0 : if (elapsed_secs >= XLogArchiveTimeout)
599 : 0 : continue; /* no sleep for us ... */
600 : 0 : cur_timeout = Min(cur_timeout, XLogArchiveTimeout - elapsed_secs);
601 : : }
602 : :
603 : 5608 : (void) WaitLatch(MyLatch,
604 : : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
605 : : cur_timeout * 1000L /* convert to ms */ ,
606 : : WAIT_EVENT_CHECKPOINTER_MAIN);
607 : : }
608 : :
609 : : /*
610 : : * From here on, elog(ERROR) should end with exit(1), not send control
611 : : * back to the sigsetjmp block above.
612 : : */
613 : 639 : ExitOnAnyError = true;
614 : :
615 [ + - ]: 639 : if (ShutdownXLOGPending)
616 : : {
617 : : /*
618 : : * Close down the database.
619 : : *
620 : : * Since ShutdownXLOG() creates restartpoint or checkpoint, and
621 : : * updates the statistics, increment the checkpoint request and flush
622 : : * out pending statistic.
623 : : */
624 : 639 : PendingCheckpointerStats.num_requested++;
625 : 639 : ShutdownXLOG(0, 0);
626 : 639 : pgstat_report_checkpointer();
627 : 639 : pgstat_report_wal(true);
628 : :
629 : : /*
630 : : * Tell postmaster that we're done.
631 : : */
632 : 639 : SendPostmasterSignal(PMSIGNAL_XLOG_IS_SHUTDOWN);
633 : 639 : ShutdownXLOGPending = false;
634 : : }
635 : :
636 : : /*
637 : : * Wait until we're asked to shut down. By separating the writing of the
638 : : * shutdown checkpoint from checkpointer exiting, checkpointer can perform
639 : : * some should-be-as-late-as-possible work like writing out stats.
640 : : */
641 : : for (;;)
642 : : {
643 : : /* Clear any already-pending wakeups */
644 : 1307 : ResetLatch(MyLatch);
645 : :
646 : 1307 : ProcessCheckpointerInterrupts();
647 : :
648 [ + + ]: 1307 : if (ShutdownRequestPending)
649 : 639 : break;
650 : :
651 : 668 : (void) WaitLatch(MyLatch,
652 : : WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
653 : : 0,
654 : : WAIT_EVENT_CHECKPOINTER_SHUTDOWN);
655 : : }
656 : :
657 : : /* Normal exit from the checkpointer is here */
658 : 639 : proc_exit(0); /* done */
659 : : }
660 : :
661 : : /*
662 : : * Process any new interrupts.
663 : : */
664 : : static void
665 : 9199 : ProcessCheckpointerInterrupts(void)
666 : : {
667 [ + + ]: 9199 : if (ProcSignalBarrierPending)
668 : 535 : ProcessProcSignalBarrier();
669 : :
670 [ + + ]: 9199 : if (ConfigReloadPending)
671 : : {
672 : 82 : ConfigReloadPending = false;
673 : 82 : ProcessConfigFile(PGC_SIGHUP);
674 : :
675 : : /*
676 : : * Checkpointer is the last process to shut down, so we ask it to hold
677 : : * the keys for a range of other tasks required most of which have
678 : : * nothing to do with checkpointing at all.
679 : : *
680 : : * For various reasons, some config values can change dynamically so
681 : : * the primary copy of them is held in shared memory to make sure all
682 : : * backends see the same value. We make Checkpointer responsible for
683 : : * updating the shared memory copy if the parameter setting changes
684 : : * because of SIGHUP.
685 : : */
686 : 82 : UpdateSharedMemoryConfig();
687 : : }
688 : :
689 : : /* Perform logging of memory contexts of this process */
690 [ + + ]: 9199 : if (LogMemoryContextPending)
691 : 1 : ProcessLogMemoryContextInterrupt();
692 : 9199 : }
693 : :
694 : : /*
695 : : * CheckArchiveTimeout -- check for archive_timeout and switch xlog files
696 : : *
697 : : * This will switch to a new WAL file and force an archive file write if
698 : : * meaningful activity is recorded in the current WAL file. This includes most
699 : : * writes, including just a single checkpoint record, but excludes WAL records
700 : : * that were inserted with the XLOG_MARK_UNIMPORTANT flag being set (like
701 : : * snapshots of running transactions). Such records, depending on
702 : : * configuration, occur on regular intervals and don't contain important
703 : : * information. This avoids generating archives with a few unimportant
704 : : * records.
705 : : */
706 : : static void
707 : 13850 : CheckArchiveTimeout(void)
708 : : {
709 : : pg_time_t now;
710 : : pg_time_t last_time;
711 : : XLogRecPtr last_switch_lsn;
712 : :
713 [ - + - - ]: 13850 : if (XLogArchiveTimeout <= 0 || RecoveryInProgress())
714 : 13850 : return;
715 : :
716 : 0 : now = (pg_time_t) time(NULL);
717 : :
718 : : /* First we do a quick check using possibly-stale local state. */
719 [ # # ]: 0 : if ((int) (now - last_xlog_switch_time) < XLogArchiveTimeout)
720 : 0 : return;
721 : :
722 : : /*
723 : : * Update local state ... note that last_xlog_switch_time is the last time
724 : : * a switch was performed *or requested*.
725 : : */
726 : 0 : last_time = GetLastSegSwitchData(&last_switch_lsn);
727 : :
728 : 0 : last_xlog_switch_time = Max(last_xlog_switch_time, last_time);
729 : :
730 : : /* Now we can do the real checks */
731 [ # # ]: 0 : if ((int) (now - last_xlog_switch_time) >= XLogArchiveTimeout)
732 : : {
733 : : /*
734 : : * Switch segment only when "important" WAL has been logged since the
735 : : * last segment switch (last_switch_lsn points to end of segment
736 : : * switch occurred in).
737 : : */
738 [ # # ]: 0 : if (GetLastImportantRecPtr() > last_switch_lsn)
739 : : {
740 : : XLogRecPtr switchpoint;
741 : :
742 : : /* mark switch as unimportant, avoids triggering checkpoints */
743 : 0 : switchpoint = RequestXLogSwitch(true);
744 : :
745 : : /*
746 : : * If the returned pointer points exactly to a segment boundary,
747 : : * assume nothing happened.
748 : : */
749 [ # # ]: 0 : if (XLogSegmentOffset(switchpoint, wal_segment_size) != 0)
750 [ # # ]: 0 : elog(DEBUG1, "write-ahead log switch forced (\"archive_timeout\"=%d)",
751 : : XLogArchiveTimeout);
752 : : }
753 : :
754 : : /*
755 : : * Update state in any case, so we don't retry constantly when the
756 : : * system is idle.
757 : : */
758 : 0 : last_xlog_switch_time = now;
759 : : }
760 : : }
761 : :
762 : : /*
763 : : * Returns true if a fast checkpoint request is pending. (Note that this does
764 : : * not check the *current* checkpoint's FAST flag, but whether there is one
765 : : * pending behind it.)
766 : : */
767 : : static bool
768 : 50725 : FastCheckpointRequested(void)
769 : : {
770 : 50725 : volatile CheckpointerShmemStruct *cps = CheckpointerShmem;
771 : :
772 : : /*
773 : : * We don't need to acquire the ckpt_lck in this case because we're only
774 : : * looking at a single flag bit.
775 : : */
776 [ + + ]: 50725 : if (cps->ckpt_flags & CHECKPOINT_FAST)
777 : 7349 : return true;
778 : 43376 : return false;
779 : : }
780 : :
781 : : /*
782 : : * CheckpointWriteDelay -- control rate of checkpoint
783 : : *
784 : : * This function is called after each page write performed by BufferSync().
785 : : * It is responsible for throttling BufferSync()'s write rate to hit
786 : : * checkpoint_completion_target.
787 : : *
788 : : * The checkpoint request flags should be passed in; currently the only one
789 : : * examined is CHECKPOINT_FAST, which disables delays between writes.
790 : : *
791 : : * 'progress' is an estimate of how much of the work has been done, as a
792 : : * fraction between 0.0 meaning none, and 1.0 meaning all done.
793 : : */
794 : : void
795 : 338242 : CheckpointWriteDelay(int flags, double progress)
796 : : {
797 : : static int absorb_counter = WRITES_PER_ABSORB;
798 : :
799 : : /* Do nothing if checkpoint is being executed by non-checkpointer process */
800 [ + + ]: 338242 : if (!AmCheckpointerProcess())
801 : 58303 : return;
802 : :
803 : : /*
804 : : * Perform the usual duties and take a nap, unless we're behind schedule,
805 : : * in which case we just try to catch up as quickly as possible.
806 : : */
807 [ + + ]: 279939 : if (!(flags & CHECKPOINT_FAST) &&
808 [ + + ]: 50920 : !ShutdownXLOGPending &&
809 [ + - ]: 50725 : !ShutdownRequestPending &&
810 [ + + + + ]: 94101 : !FastCheckpointRequested() &&
811 : 43376 : IsCheckpointOnSchedule(progress))
812 : : {
813 [ - + ]: 7959 : if (ConfigReloadPending)
814 : : {
815 : 0 : ConfigReloadPending = false;
816 : 0 : ProcessConfigFile(PGC_SIGHUP);
817 : : /* update shmem copies of config variables */
818 : 0 : UpdateSharedMemoryConfig();
819 : : }
820 : :
821 : 7959 : AbsorbSyncRequests();
822 : 7959 : absorb_counter = WRITES_PER_ABSORB;
823 : :
824 : 7959 : CheckArchiveTimeout();
825 : :
826 : : /* Report interim statistics to the cumulative stats system */
827 : 7959 : pgstat_report_checkpointer();
828 : :
829 : : /*
830 : : * This sleep used to be connected to bgwriter_delay, typically 200ms.
831 : : * That resulted in more frequent wakeups if not much work to do.
832 : : * Checkpointer and bgwriter are no longer related so take the Big
833 : : * Sleep.
834 : : */
835 : 7959 : WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | WL_TIMEOUT,
836 : : 100,
837 : : WAIT_EVENT_CHECKPOINT_WRITE_DELAY);
838 : 7959 : ResetLatch(MyLatch);
839 : : }
840 [ + + ]: 271980 : else if (--absorb_counter <= 0)
841 : : {
842 : : /*
843 : : * Absorb pending fsync requests after each WRITES_PER_ABSORB write
844 : : * operations even when we don't sleep, to prevent overflow of the
845 : : * fsync request queue.
846 : : */
847 : 139 : AbsorbSyncRequests();
848 : 139 : absorb_counter = WRITES_PER_ABSORB;
849 : : }
850 : :
851 : : /* Check for barrier events. */
852 [ + + ]: 279939 : if (ProcSignalBarrierPending)
853 : 12 : ProcessProcSignalBarrier();
854 : : }
855 : :
856 : : /*
857 : : * IsCheckpointOnSchedule -- are we on schedule to finish this checkpoint
858 : : * (or restartpoint) in time?
859 : : *
860 : : * Compares the current progress against the time/segments elapsed since last
861 : : * checkpoint, and returns true if the progress we've made this far is greater
862 : : * than the elapsed time/segments.
863 : : */
864 : : static bool
865 : 43376 : IsCheckpointOnSchedule(double progress)
866 : : {
867 : : XLogRecPtr recptr;
868 : : struct timeval now;
869 : : double elapsed_xlogs,
870 : : elapsed_time;
871 : :
872 : : Assert(ckpt_active);
873 : :
874 : : /* Scale progress according to checkpoint_completion_target. */
875 : 43376 : progress *= CheckPointCompletionTarget;
876 : :
877 : : /*
878 : : * Check against the cached value first. Only do the more expensive
879 : : * calculations once we reach the target previously calculated. Since
880 : : * neither time or WAL insert pointer moves backwards, a freshly
881 : : * calculated value can only be greater than or equal to the cached value.
882 : : */
883 [ + + ]: 43376 : if (progress < ckpt_cached_elapsed)
884 : 31365 : return false;
885 : :
886 : : /*
887 : : * Check progress against WAL segments written and CheckPointSegments.
888 : : *
889 : : * We compare the current WAL insert location against the location
890 : : * computed before calling CreateCheckPoint. The code in XLogInsert that
891 : : * actually triggers a checkpoint when CheckPointSegments is exceeded
892 : : * compares against RedoRecPtr, so this is not completely accurate.
893 : : * However, it's good enough for our purposes, we're only calculating an
894 : : * estimate anyway.
895 : : *
896 : : * During recovery, we compare last replayed WAL record's location with
897 : : * the location computed before calling CreateRestartPoint. That maintains
898 : : * the same pacing as we have during checkpoints in normal operation, but
899 : : * we might exceed max_wal_size by a fair amount. That's because there can
900 : : * be a large gap between a checkpoint's redo-pointer and the checkpoint
901 : : * record itself, and we only start the restartpoint after we've seen the
902 : : * checkpoint record. (The gap is typically up to CheckPointSegments *
903 : : * checkpoint_completion_target where checkpoint_completion_target is the
904 : : * value that was in effect when the WAL was generated).
905 : : */
906 [ + + ]: 12011 : if (RecoveryInProgress())
907 : 5332 : recptr = GetXLogReplayRecPtr(NULL);
908 : : else
909 : 6679 : recptr = GetInsertRecPtr();
910 : 12011 : elapsed_xlogs = (((double) (recptr - ckpt_start_recptr)) /
911 : 12011 : wal_segment_size) / CheckPointSegments;
912 : :
913 [ + + ]: 12011 : if (progress < elapsed_xlogs)
914 : : {
915 : 4049 : ckpt_cached_elapsed = elapsed_xlogs;
916 : 4049 : return false;
917 : : }
918 : :
919 : : /*
920 : : * Check progress against time elapsed and checkpoint_timeout.
921 : : */
922 : 7962 : gettimeofday(&now, NULL);
923 : 7962 : elapsed_time = ((double) ((pg_time_t) now.tv_sec - ckpt_start_time) +
924 : 7962 : now.tv_usec / 1000000.0) / CheckPointTimeout;
925 : :
926 [ + + ]: 7962 : if (progress < elapsed_time)
927 : : {
928 : 3 : ckpt_cached_elapsed = elapsed_time;
929 : 3 : return false;
930 : : }
931 : :
932 : : /* It looks like we're on schedule. */
933 : 7959 : return true;
934 : : }
935 : :
936 : :
937 : : /* --------------------------------
938 : : * signal handler routines
939 : : * --------------------------------
940 : : */
941 : :
942 : : /* SIGINT: set flag to trigger writing of shutdown checkpoint */
943 : : static void
944 : 697 : ReqShutdownXLOG(SIGNAL_ARGS)
945 : : {
946 : 697 : ShutdownXLOGPending = true;
947 : 697 : SetLatch(MyLatch);
948 : 697 : }
949 : :
950 : :
951 : : /* --------------------------------
952 : : * communication with backends
953 : : * --------------------------------
954 : : */
955 : :
956 : : /*
957 : : * CheckpointerShmemRequest
958 : : * Register shared memory space needed for checkpointer
959 : : */
960 : : static void
961 : 1259 : CheckpointerShmemRequest(void *arg)
962 : : {
963 : : Size size;
964 : :
965 : : /*
966 : : * The size of the requests[] array is arbitrarily set equal to NBuffers.
967 : : * But there is a cap of MAX_CHECKPOINT_REQUESTS to prevent accumulating
968 : : * too many checkpoint requests in the ring buffer.
969 : : */
970 : 1259 : size = offsetof(CheckpointerShmemStruct, requests);
971 : 1259 : size = add_size(size, mul_size(Min(NBuffers,
972 : : MAX_CHECKPOINT_REQUESTS),
973 : : sizeof(CheckpointerRequest)));
974 : 1259 : ShmemRequestStruct(.name = "Checkpointer Data",
975 : : .size = size,
976 : : .ptr = (void **) &CheckpointerShmem,
977 : : );
978 : 1259 : }
979 : :
980 : : /*
981 : : * CheckpointerShmemInit
982 : : * Initialize checkpointer-related shared memory
983 : : */
984 : : static void
985 : 1256 : CheckpointerShmemInit(void *arg)
986 : : {
987 : 1256 : SpinLockInit(&CheckpointerShmem->ckpt_lck);
988 : 1256 : CheckpointerShmem->max_requests = Min(NBuffers, MAX_CHECKPOINT_REQUESTS);
989 : 1256 : CheckpointerShmem->head = CheckpointerShmem->tail = 0;
990 : 1256 : ConditionVariableInit(&CheckpointerShmem->start_cv);
991 : 1256 : ConditionVariableInit(&CheckpointerShmem->done_cv);
992 : 1256 : }
993 : :
994 : : /*
995 : : * ExecCheckpoint
996 : : * Primary entry point for manual CHECKPOINT commands
997 : : *
998 : : * This is mainly a wrapper for RequestCheckpoint().
999 : : */
1000 : : void
1001 : 464 : ExecCheckpoint(ParseState *pstate, CheckPointStmt *stmt)
1002 : : {
1003 : 464 : bool fast = true;
1004 : 464 : bool unlogged = false;
1005 : :
1006 [ + + + + : 932 : foreach_ptr(DefElem, opt, stmt->options)
+ + ]
1007 : : {
1008 [ + + ]: 20 : if (strcmp(opt->defname, "mode") == 0)
1009 : : {
1010 : 8 : char *mode = defGetString(opt);
1011 : :
1012 [ - + ]: 8 : if (strcmp(mode, "spread") == 0)
1013 : 0 : fast = false;
1014 [ + + ]: 8 : else if (strcmp(mode, "fast") != 0)
1015 [ + - ]: 4 : ereport(ERROR,
1016 : : (errcode(ERRCODE_SYNTAX_ERROR),
1017 : : errmsg("unrecognized value for %s option \"%s\": \"%s\"",
1018 : : "CHECKPOINT", "mode", mode),
1019 : : parser_errposition(pstate, opt->location)));
1020 : : }
1021 [ + + ]: 12 : else if (strcmp(opt->defname, "flush_unlogged") == 0)
1022 : 8 : unlogged = defGetBoolean(opt);
1023 : : else
1024 [ + - ]: 4 : ereport(ERROR,
1025 : : (errcode(ERRCODE_SYNTAX_ERROR),
1026 : : errmsg("unrecognized %s option \"%s\"",
1027 : : "CHECKPOINT", opt->defname),
1028 : : parser_errposition(pstate, opt->location)));
1029 : : }
1030 : :
1031 [ - + ]: 456 : if (!has_privs_of_role(GetUserId(), ROLE_PG_CHECKPOINT))
1032 [ # # ]: 0 : ereport(ERROR,
1033 : : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
1034 : : /* translator: %s is name of an SQL command (e.g., CHECKPOINT) */
1035 : : errmsg("permission denied to execute %s command",
1036 : : "CHECKPOINT"),
1037 : : errdetail("Only roles with privileges of the \"%s\" role may execute this command.",
1038 : : "pg_checkpoint")));
1039 : :
1040 [ + - ]: 912 : RequestCheckpoint(CHECKPOINT_WAIT |
1041 : 456 : (fast ? CHECKPOINT_FAST : 0) |
1042 [ + + ]: 456 : (unlogged ? CHECKPOINT_FLUSH_UNLOGGED : 0) |
1043 [ + + ]: 456 : (RecoveryInProgress() ? 0 : CHECKPOINT_FORCE));
1044 : 456 : }
1045 : :
1046 : : /*
1047 : : * RequestCheckpoint
1048 : : * Called in backend processes to request a checkpoint
1049 : : *
1050 : : * flags is a bitwise OR of the following:
1051 : : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
1052 : : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
1053 : : * CHECKPOINT_FAST: finish the checkpoint ASAP,
1054 : : * ignoring checkpoint_completion_target parameter.
1055 : : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
1056 : : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
1057 : : * CHECKPOINT_END_OF_RECOVERY, and the CHECKPOINT command).
1058 : : * CHECKPOINT_WAIT: wait for completion before returning (otherwise,
1059 : : * just signal checkpointer to do it, and return).
1060 : : * CHECKPOINT_CAUSE_XLOG: checkpoint is requested due to xlog filling.
1061 : : * (This affects logging, and in particular enables CheckPointWarning.)
1062 : : */
1063 : : void
1064 : 2745 : RequestCheckpoint(int flags)
1065 : : {
1066 : : int ntries;
1067 : : int old_failed,
1068 : : old_started;
1069 : :
1070 : : /*
1071 : : * If in a standalone backend, just do it ourselves.
1072 : : */
1073 [ + + ]: 2745 : if (!IsPostmasterEnvironment)
1074 : : {
1075 : : /*
1076 : : * There's no point in doing slow checkpoints in a standalone backend,
1077 : : * because there's no other backends the checkpoint could disrupt.
1078 : : */
1079 : 229 : CreateCheckPoint(flags | CHECKPOINT_FAST);
1080 : :
1081 : : /* Free all smgr objects, as CheckpointerMain() normally would. */
1082 : 229 : smgrdestroyall();
1083 : :
1084 : 229 : return;
1085 : : }
1086 : :
1087 : : /*
1088 : : * Atomically set the request flags, and take a snapshot of the counters.
1089 : : * When we see ckpt_started > old_started, we know the flags we set here
1090 : : * have been seen by checkpointer.
1091 : : *
1092 : : * Note that we OR the flags with any existing flags, to avoid overriding
1093 : : * a "stronger" request by another backend. The flag senses must be
1094 : : * chosen to make this work!
1095 : : */
1096 : 2516 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1097 : :
1098 : 2516 : old_failed = CheckpointerShmem->ckpt_failed;
1099 : 2516 : old_started = CheckpointerShmem->ckpt_started;
1100 : 2516 : CheckpointerShmem->ckpt_flags |= (flags | CHECKPOINT_REQUESTED);
1101 : :
1102 : 2516 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1103 : :
1104 : : /*
1105 : : * Set checkpointer's latch to request checkpoint. It's possible that the
1106 : : * checkpointer hasn't started yet, so we will retry a few times if
1107 : : * needed. (Actually, more than a few times, since on slow or overloaded
1108 : : * buildfarm machines, it's been observed that the checkpointer can take
1109 : : * several seconds to start.) However, if not told to wait for the
1110 : : * checkpoint to occur, we consider failure to set the latch to be
1111 : : * nonfatal and merely LOG it. The checkpointer should see the request
1112 : : * when it does start, with or without the SetLatch().
1113 : : */
1114 : : #define MAX_SIGNAL_TRIES 600 /* max wait 60.0 sec */
1115 : 2516 : for (ntries = 0;; ntries++)
1116 : 20 : {
1117 : 2536 : ProcNumber checkpointerProc = pg_atomic_read_u32(&ProcGlobal->checkpointerProc);
1118 : :
1119 [ + + ]: 2536 : if (checkpointerProc == INVALID_PROC_NUMBER)
1120 : : {
1121 [ + - + + ]: 21 : if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT))
1122 : : {
1123 [ - + + - ]: 1 : elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG,
1124 : : "could not notify checkpoint: checkpointer is not running");
1125 : 1 : break;
1126 : : }
1127 : : }
1128 : : else
1129 : : {
1130 : 2515 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1131 : : /* notified successfully */
1132 : 2515 : break;
1133 : : }
1134 : :
1135 [ - + ]: 20 : CHECK_FOR_INTERRUPTS();
1136 : 20 : pg_usleep(100000L); /* wait 0.1 sec, then retry */
1137 : : }
1138 : :
1139 : : /*
1140 : : * If requested, wait for completion. We detect completion according to
1141 : : * the algorithm given above.
1142 : : */
1143 [ + + ]: 2516 : if (flags & CHECKPOINT_WAIT)
1144 : : {
1145 : : int new_started,
1146 : : new_failed;
1147 : :
1148 : : /* Wait for a new checkpoint to start. */
1149 : 886 : ConditionVariablePrepareToSleep(&CheckpointerShmem->start_cv);
1150 : : for (;;)
1151 : : {
1152 : 1658 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1153 : 1658 : new_started = CheckpointerShmem->ckpt_started;
1154 : 1658 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1155 : :
1156 [ + + ]: 1658 : if (new_started != old_started)
1157 : 886 : break;
1158 : :
1159 : 772 : ConditionVariableSleep(&CheckpointerShmem->start_cv,
1160 : : WAIT_EVENT_CHECKPOINT_START);
1161 : : }
1162 : 886 : ConditionVariableCancelSleep();
1163 : :
1164 : : /*
1165 : : * We are waiting for ckpt_done >= new_started, in a modulo sense.
1166 : : */
1167 : 886 : ConditionVariablePrepareToSleep(&CheckpointerShmem->done_cv);
1168 : : for (;;)
1169 : 691 : {
1170 : : int new_done;
1171 : :
1172 : 1577 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1173 : 1577 : new_done = CheckpointerShmem->ckpt_done;
1174 : 1577 : new_failed = CheckpointerShmem->ckpt_failed;
1175 : 1577 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1176 : :
1177 [ + + ]: 1577 : if (new_done - new_started >= 0)
1178 : 886 : break;
1179 : :
1180 : 691 : ConditionVariableSleep(&CheckpointerShmem->done_cv,
1181 : : WAIT_EVENT_CHECKPOINT_DONE);
1182 : : }
1183 : 886 : ConditionVariableCancelSleep();
1184 : :
1185 [ - + ]: 886 : if (new_failed != old_failed)
1186 [ # # ]: 0 : ereport(ERROR,
1187 : : (errmsg("checkpoint request failed"),
1188 : : errhint("Consult recent messages in the server log for details.")));
1189 : : }
1190 : : }
1191 : :
1192 : : /*
1193 : : * ForwardSyncRequest
1194 : : * Forward a file-fsync request from a backend to the checkpointer
1195 : : *
1196 : : * Whenever a backend is compelled to write directly to a relation
1197 : : * (which should be seldom, if the background writer is getting its job done),
1198 : : * the backend calls this routine to pass over knowledge that the relation
1199 : : * is dirty and must be fsync'd before next checkpoint. We also use this
1200 : : * opportunity to count such writes for statistical purposes.
1201 : : *
1202 : : * To avoid holding the lock for longer than necessary, we normally write
1203 : : * to the requests[] queue without checking for duplicates. The checkpointer
1204 : : * will have to eliminate dups internally anyway. However, if we discover
1205 : : * that the queue is full, we make a pass over the entire queue to compact
1206 : : * it. This is somewhat expensive, but the alternative is for the backend
1207 : : * to perform its own fsync, which is far more expensive in practice. It
1208 : : * is theoretically possible a backend fsync might still be necessary, if
1209 : : * the queue is full and contains no duplicate entries. In that case, we
1210 : : * let the backend know by returning false.
1211 : : */
1212 : : bool
1213 : 1487078 : ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)
1214 : : {
1215 : : CheckpointerRequest *request;
1216 : : bool too_full;
1217 : : int insert_pos;
1218 : :
1219 [ - + ]: 1487078 : if (!IsUnderPostmaster)
1220 : 0 : return false; /* probably shouldn't even get here */
1221 : :
1222 [ - + ]: 1487078 : if (AmCheckpointerProcess())
1223 [ # # ]: 0 : elog(ERROR, "ForwardSyncRequest must not be called in checkpointer");
1224 : :
1225 : 1487078 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1226 : :
1227 : : /*
1228 : : * If the checkpointer isn't running or the request queue is full, the
1229 : : * backend will have to perform its own fsync request. But before forcing
1230 : : * that to happen, we can try to compact the request queue.
1231 : : */
1232 [ + + ]: 1487078 : if (CheckpointerShmem->checkpointer_pid == 0 ||
1233 [ + + ]: 1486747 : (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&
1234 [ + + ]: 249 : !CompactCheckpointerRequestQueue()))
1235 : : {
1236 : 396 : LWLockRelease(CheckpointerCommLock);
1237 : 396 : return false;
1238 : : }
1239 : :
1240 : : /* OK, insert request */
1241 : 1486682 : insert_pos = CheckpointerShmem->tail;
1242 : 1486682 : request = &CheckpointerShmem->requests[insert_pos];
1243 : 1486682 : request->ftag = *ftag;
1244 : 1486682 : request->type = type;
1245 : :
1246 : 1486682 : CheckpointerShmem->tail = (CheckpointerShmem->tail + 1) % CheckpointerShmem->max_requests;
1247 : 1486682 : CheckpointerShmem->num_requests++;
1248 : :
1249 : : /* If queue is more than half full, nudge the checkpointer to empty it */
1250 : 1486682 : too_full = (CheckpointerShmem->num_requests >=
1251 : 1486682 : CheckpointerShmem->max_requests / 2);
1252 : :
1253 : 1486682 : LWLockRelease(CheckpointerCommLock);
1254 : :
1255 : : /* ... but not till after we release the lock */
1256 [ + + ]: 1486682 : if (too_full)
1257 : : {
1258 : 38593 : ProcNumber checkpointerProc = pg_atomic_read_u32(&ProcGlobal->checkpointerProc);
1259 : :
1260 [ + - ]: 38593 : if (checkpointerProc != INVALID_PROC_NUMBER)
1261 : 38593 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1262 : : }
1263 : :
1264 : 1486682 : return true;
1265 : : }
1266 : :
1267 : : /*
1268 : : * CompactCheckpointerRequestQueue
1269 : : * Remove duplicates from the request queue to avoid backend fsyncs.
1270 : : * Returns "true" if any entries were removed.
1271 : : *
1272 : : * Although a full fsync request queue is not common, it can lead to severe
1273 : : * performance problems when it does happen. So far, this situation has
1274 : : * only been observed to occur when the system is under heavy write load,
1275 : : * and especially during the "sync" phase of a checkpoint. Without this
1276 : : * logic, each backend begins doing an fsync for every block written, which
1277 : : * gets very expensive and can slow down the whole system.
1278 : : *
1279 : : * Trying to do this every time the queue is full could lose if there
1280 : : * aren't any removable entries. But that should be vanishingly rare in
1281 : : * practice: there's one queue entry per shared buffer.
1282 : : */
1283 : : static bool
1284 : 249 : CompactCheckpointerRequestQueue(void)
1285 : : {
1286 : : struct CheckpointerSlotMapping
1287 : : {
1288 : : CheckpointerRequest request;
1289 : : int ring_idx;
1290 : : };
1291 : :
1292 : : int n;
1293 : 249 : int num_skipped = 0;
1294 : : int head;
1295 : : int max_requests;
1296 : : int num_requests;
1297 : : int read_idx,
1298 : : write_idx;
1299 : : HASHCTL ctl;
1300 : : HTAB *htab;
1301 : : bool *skip_slot;
1302 : :
1303 : : /* must hold CheckpointerCommLock in exclusive mode */
1304 : : Assert(LWLockHeldByMe(CheckpointerCommLock));
1305 : :
1306 : : /* Avoid memory allocations in a critical section. */
1307 [ - + ]: 249 : if (CritSectionCount > 0)
1308 : 0 : return false;
1309 : :
1310 : 249 : max_requests = CheckpointerShmem->max_requests;
1311 : 249 : num_requests = CheckpointerShmem->num_requests;
1312 : :
1313 : : /* Initialize skip_slot array */
1314 : 249 : skip_slot = palloc0_array(bool, max_requests);
1315 : :
1316 : 249 : head = CheckpointerShmem->head;
1317 : :
1318 : : /* Initialize temporary hash table */
1319 : 249 : ctl.keysize = sizeof(CheckpointerRequest);
1320 : 249 : ctl.entrysize = sizeof(struct CheckpointerSlotMapping);
1321 : 249 : ctl.hcxt = CurrentMemoryContext;
1322 : :
1323 : 249 : htab = hash_create("CompactCheckpointerRequestQueue",
1324 : 249 : CheckpointerShmem->num_requests,
1325 : : &ctl,
1326 : : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1327 : :
1328 : : /*
1329 : : * The basic idea here is that a request can be skipped if it's followed
1330 : : * by a later, identical request. It might seem more sensible to work
1331 : : * backwards from the end of the queue and check whether a request is
1332 : : * *preceded* by an earlier, identical request, in the hopes of doing less
1333 : : * copying. But that might change the semantics, if there's an
1334 : : * intervening SYNC_FORGET_REQUEST or SYNC_FILTER_REQUEST, so we do it
1335 : : * this way. It would be possible to be even smarter if we made the code
1336 : : * below understand the specific semantics of such requests (it could blow
1337 : : * away preceding entries that would end up being canceled anyhow), but
1338 : : * it's not clear that the extra complexity would buy us anything.
1339 : : */
1340 : 249 : read_idx = head;
1341 [ + + ]: 22153 : for (n = 0; n < num_requests; n++)
1342 : : {
1343 : : CheckpointerRequest *request;
1344 : : struct CheckpointerSlotMapping *slotmap;
1345 : : bool found;
1346 : :
1347 : : /*
1348 : : * We use the request struct directly as a hashtable key. This
1349 : : * assumes that any padding bytes in the structs are consistently the
1350 : : * same, which should be okay because we zeroed them in
1351 : : * CheckpointerShmemInit. Note also that RelFileLocator had better
1352 : : * contain no pad bytes.
1353 : : */
1354 : 21904 : request = &CheckpointerShmem->requests[read_idx];
1355 : 21904 : slotmap = hash_search(htab, request, HASH_ENTER, &found);
1356 [ + + ]: 21904 : if (found)
1357 : : {
1358 : : /* Duplicate, so mark the previous occurrence as skippable */
1359 : 12683 : skip_slot[slotmap->ring_idx] = true;
1360 : 12683 : num_skipped++;
1361 : : }
1362 : : /* Remember slot containing latest occurrence of this request value */
1363 : 21904 : slotmap->ring_idx = read_idx;
1364 : :
1365 : : /* Move to the next request in the ring buffer */
1366 : 21904 : read_idx = (read_idx + 1) % max_requests;
1367 : : }
1368 : :
1369 : : /* Done with the hash table. */
1370 : 249 : hash_destroy(htab);
1371 : :
1372 : : /* If no duplicates, we're out of luck. */
1373 [ + + ]: 249 : if (!num_skipped)
1374 : : {
1375 : 65 : pfree(skip_slot);
1376 : 65 : return false;
1377 : : }
1378 : :
1379 : : /* We found some duplicates; remove them. */
1380 : 184 : read_idx = write_idx = head;
1381 [ + + ]: 19256 : for (n = 0; n < num_requests; n++)
1382 : : {
1383 : : /* If this slot is NOT skipped, keep it */
1384 [ + + ]: 19072 : if (!skip_slot[read_idx])
1385 : : {
1386 : : /* If the read and write positions are different, copy the request */
1387 [ + + ]: 6389 : if (write_idx != read_idx)
1388 : 5377 : CheckpointerShmem->requests[write_idx] =
1389 : 5377 : CheckpointerShmem->requests[read_idx];
1390 : :
1391 : : /* Advance the write position */
1392 : 6389 : write_idx = (write_idx + 1) % max_requests;
1393 : : }
1394 : :
1395 : 19072 : read_idx = (read_idx + 1) % max_requests;
1396 : : }
1397 : :
1398 : : /*
1399 : : * Update ring buffer state: head remains the same, tail moves, count
1400 : : * decreases
1401 : : */
1402 : 184 : CheckpointerShmem->tail = write_idx;
1403 : 184 : CheckpointerShmem->num_requests -= num_skipped;
1404 : :
1405 [ + + ]: 184 : ereport(DEBUG1,
1406 : : (errmsg_internal("compacted fsync request queue from %d entries to %d entries",
1407 : : num_requests, CheckpointerShmem->num_requests)));
1408 : :
1409 : : /* Cleanup. */
1410 : 184 : pfree(skip_slot);
1411 : 184 : return true;
1412 : : }
1413 : :
1414 : : /*
1415 : : * AbsorbSyncRequests
1416 : : * Retrieve queued sync requests and pass them to sync mechanism.
1417 : : *
1418 : : * This is exported because it must be called during CreateCheckPoint;
1419 : : * we have to be sure we have accepted all pending requests just before
1420 : : * we start fsync'ing. Since CreateCheckPoint sometimes runs in
1421 : : * non-checkpointer processes, do nothing if not checkpointer.
1422 : : */
1423 : : void
1424 : 22001 : AbsorbSyncRequests(void)
1425 : : {
1426 : 22001 : CheckpointerRequest *requests = NULL;
1427 : : CheckpointerRequest *request;
1428 : : int n,
1429 : : i;
1430 : : bool loop;
1431 : :
1432 [ + + ]: 22001 : if (!AmCheckpointerProcess())
1433 : 724 : return;
1434 : :
1435 : : do
1436 : : {
1437 : 21277 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1438 : :
1439 : : /*---
1440 : : * We try to avoid holding the lock for a long time by:
1441 : : * 1. Copying the request array and processing the requests after
1442 : : * releasing the lock;
1443 : : * 2. Processing not the whole queue, but only batches of
1444 : : * CKPT_REQ_BATCH_SIZE at once.
1445 : : *
1446 : : * Once we have cleared the requests from shared memory, we must
1447 : : * PANIC if we then fail to absorb them (e.g., because our hashtable
1448 : : * runs out of memory). This is because the system cannot run safely
1449 : : * if we are unable to fsync what we have been told to fsync.
1450 : : * Fortunately, the hashtable is so small that the problem is quite
1451 : : * unlikely to arise in practice.
1452 : : *
1453 : : * Note: The maximum possible size of a ring buffer is
1454 : : * MAX_CHECKPOINT_REQUESTS entries, which fit into a maximum palloc
1455 : : * allocation size of 1Gb. Our maximum batch size,
1456 : : * CKPT_REQ_BATCH_SIZE, is even smaller.
1457 : : */
1458 : 21277 : n = Min(CheckpointerShmem->num_requests, CKPT_REQ_BATCH_SIZE);
1459 [ + + ]: 21277 : if (n > 0)
1460 : : {
1461 [ + - ]: 11093 : if (!requests)
1462 : 11093 : requests = (CheckpointerRequest *) palloc(n * sizeof(CheckpointerRequest));
1463 : :
1464 [ + + ]: 1322876 : for (i = 0; i < n; i++)
1465 : : {
1466 : 1311783 : requests[i] = CheckpointerShmem->requests[CheckpointerShmem->head];
1467 : 1311783 : CheckpointerShmem->head = (CheckpointerShmem->head + 1) % CheckpointerShmem->max_requests;
1468 : : }
1469 : :
1470 : 11093 : CheckpointerShmem->num_requests -= n;
1471 : :
1472 : : }
1473 : :
1474 : 21277 : START_CRIT_SECTION();
1475 : :
1476 : : /* Are there any requests in the queue? If so, keep going. */
1477 : 21277 : loop = CheckpointerShmem->num_requests != 0;
1478 : :
1479 : 21277 : LWLockRelease(CheckpointerCommLock);
1480 : :
1481 [ + + ]: 1333060 : for (request = requests; n > 0; request++, n--)
1482 : 1311783 : RememberSyncRequest(&request->ftag, request->type);
1483 : :
1484 : 21277 : END_CRIT_SECTION();
1485 [ - + ]: 21277 : } while (loop);
1486 : :
1487 [ + + ]: 21277 : if (requests)
1488 : 11093 : pfree(requests);
1489 : : }
1490 : :
1491 : : /*
1492 : : * Update any shared memory configurations based on config parameters
1493 : : */
1494 : : static void
1495 : 724 : UpdateSharedMemoryConfig(void)
1496 : : {
1497 : : /* update global shmem state for sync rep */
1498 : 724 : SyncRepUpdateSyncStandbysDefined();
1499 : :
1500 : : /*
1501 : : * If full_page_writes has been changed by SIGHUP, we update it in shared
1502 : : * memory and write an XLOG_FPW_CHANGE record.
1503 : : */
1504 : 724 : UpdateFullPageWrites();
1505 : :
1506 [ + + ]: 724 : elog(DEBUG2, "checkpointer updated shared memory configuration values");
1507 : 724 : }
1508 : :
1509 : : /*
1510 : : * FirstCallSinceLastCheckpoint allows a process to take an action once
1511 : : * per checkpoint cycle by asynchronously checking for checkpoint completion.
1512 : : */
1513 : : bool
1514 : 15193 : FirstCallSinceLastCheckpoint(void)
1515 : : {
1516 : : static int ckpt_done = 0;
1517 : : int new_done;
1518 : 15193 : bool FirstCall = false;
1519 : :
1520 : 15193 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1521 : 15193 : new_done = CheckpointerShmem->ckpt_done;
1522 : 15193 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1523 : :
1524 [ + + ]: 15193 : if (new_done != ckpt_done)
1525 : 622 : FirstCall = true;
1526 : :
1527 : 15193 : ckpt_done = new_done;
1528 : :
1529 : 15193 : return FirstCall;
1530 : : }
1531 : :
1532 : : /*
1533 : : * Wake up the checkpointer process.
1534 : : */
1535 : : void
1536 : 1041 : WakeupCheckpointer(void)
1537 : : {
1538 : 1041 : ProcNumber checkpointerProc = pg_atomic_read_u32(&ProcGlobal->checkpointerProc);
1539 : :
1540 [ + + ]: 1041 : if (checkpointerProc != INVALID_PROC_NUMBER)
1541 : 710 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1542 : 1041 : }
|
