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 "postmaster/auxprocess.h"
51 : : #include "postmaster/bgwriter.h"
52 : : #include "postmaster/interrupt.h"
53 : : #include "replication/syncrep.h"
54 : : #include "storage/aio_subsys.h"
55 : : #include "storage/bufmgr.h"
56 : : #include "storage/condition_variable.h"
57 : : #include "storage/fd.h"
58 : : #include "storage/ipc.h"
59 : : #include "storage/lwlock.h"
60 : : #include "storage/pmsignal.h"
61 : : #include "storage/proc.h"
62 : : #include "storage/procsignal.h"
63 : : #include "storage/shmem.h"
64 : : #include "storage/smgr.h"
65 : : #include "storage/spin.h"
66 : : #include "utils/acl.h"
67 : : #include "utils/guc.h"
68 : : #include "utils/memutils.h"
69 : : #include "utils/resowner.h"
70 : :
71 : :
72 : : /*----------
73 : : * Shared memory area for communication between checkpointer and backends
74 : : *
75 : : * The ckpt counters allow backends to watch for completion of a checkpoint
76 : : * request they send. Here's how it works:
77 : : * * At start of a checkpoint, checkpointer reads (and clears) the request
78 : : * flags and increments ckpt_started, while holding ckpt_lck.
79 : : * * On completion of a checkpoint, checkpointer sets ckpt_done to
80 : : * equal ckpt_started.
81 : : * * On failure of a checkpoint, checkpointer increments ckpt_failed
82 : : * and sets ckpt_done to equal ckpt_started.
83 : : *
84 : : * The algorithm for backends is:
85 : : * 1. Record current values of ckpt_failed and ckpt_started, and
86 : : * set request flags, while holding ckpt_lck.
87 : : * 2. Send signal to request checkpoint.
88 : : * 3. Sleep until ckpt_started changes. Now you know a checkpoint has
89 : : * begun since you started this algorithm (although *not* that it was
90 : : * specifically initiated by your signal), and that it is using your flags.
91 : : * 4. Record new value of ckpt_started.
92 : : * 5. Sleep until ckpt_done >= saved value of ckpt_started. (Use modulo
93 : : * arithmetic here in case counters wrap around.) Now you know a
94 : : * checkpoint has started and completed, but not whether it was
95 : : * successful.
96 : : * 6. If ckpt_failed is different from the originally saved value,
97 : : * assume request failed; otherwise it was definitely successful.
98 : : *
99 : : * ckpt_flags holds the OR of the checkpoint request flags sent by all
100 : : * requesting backends since the last checkpoint start. The flags are
101 : : * chosen so that OR'ing is the correct way to combine multiple requests.
102 : : *
103 : : * The requests array holds fsync requests sent by backends and not yet
104 : : * absorbed by the checkpointer.
105 : : *
106 : : * Unlike the checkpoint fields, requests related fields are protected by
107 : : * CheckpointerCommLock.
108 : : *----------
109 : : */
110 : : typedef struct
111 : : {
112 : : SyncRequestType type; /* request type */
113 : : FileTag ftag; /* file identifier */
114 : : } CheckpointerRequest;
115 : :
116 : : typedef struct
117 : : {
118 : : pid_t checkpointer_pid; /* PID (0 if not started) */
119 : :
120 : : slock_t ckpt_lck; /* protects all the ckpt_* fields */
121 : :
122 : : int ckpt_started; /* advances when checkpoint starts */
123 : : int ckpt_done; /* advances when checkpoint done */
124 : : int ckpt_failed; /* advances when checkpoint fails */
125 : :
126 : : int ckpt_flags; /* checkpoint flags, as defined in xlog.h */
127 : :
128 : : ConditionVariable start_cv; /* signaled when ckpt_started advances */
129 : : ConditionVariable done_cv; /* signaled when ckpt_done advances */
130 : :
131 : : int num_requests; /* current # of requests */
132 : : int max_requests; /* allocated array size */
133 : :
134 : : int head; /* Index of the first request in the ring
135 : : * buffer */
136 : : int tail; /* Index of the last request in the ring
137 : : * buffer */
138 : :
139 : : /* The ring buffer of pending checkpointer requests */
140 : : CheckpointerRequest requests[FLEXIBLE_ARRAY_MEMBER];
141 : : } CheckpointerShmemStruct;
142 : :
143 : : static CheckpointerShmemStruct *CheckpointerShmem;
144 : :
145 : : /* interval for calling AbsorbSyncRequests in CheckpointWriteDelay */
146 : : #define WRITES_PER_ABSORB 1000
147 : :
148 : : /* Maximum number of checkpointer requests to process in one batch */
149 : : #define CKPT_REQ_BATCH_SIZE 10000
150 : :
151 : : /* Max number of requests the checkpointer request queue can hold */
152 : : #define MAX_CHECKPOINT_REQUESTS 10000000
153 : :
154 : : /*
155 : : * GUC parameters
156 : : */
157 : : int CheckPointTimeout = 300;
158 : : int CheckPointWarning = 30;
159 : : double CheckPointCompletionTarget = 0.9;
160 : :
161 : : /*
162 : : * Private state
163 : : */
164 : : static bool ckpt_active = false;
165 : : static volatile sig_atomic_t ShutdownXLOGPending = false;
166 : :
167 : : /* these values are valid when ckpt_active is true: */
168 : : static pg_time_t ckpt_start_time;
169 : : static XLogRecPtr ckpt_start_recptr;
170 : : static double ckpt_cached_elapsed;
171 : :
172 : : static pg_time_t last_checkpoint_time;
173 : : static pg_time_t last_xlog_switch_time;
174 : :
175 : : /* Prototypes for private functions */
176 : :
177 : : static void ProcessCheckpointerInterrupts(void);
178 : : static void CheckArchiveTimeout(void);
179 : : static bool IsCheckpointOnSchedule(double progress);
180 : : static bool FastCheckpointRequested(void);
181 : : static bool CompactCheckpointerRequestQueue(void);
182 : : static void UpdateSharedMemoryConfig(void);
183 : :
184 : : /* Signal handlers */
185 : : static void ReqShutdownXLOG(SIGNAL_ARGS);
186 : :
187 : :
188 : : /*
189 : : * Main entry point for checkpointer process
190 : : *
191 : : * This is invoked from AuxiliaryProcessMain, which has already created the
192 : : * basic execution environment, but not enabled signals yet.
193 : : */
194 : : void
195 : 1 : CheckpointerMain(const void *startup_data, size_t startup_data_len)
196 : : {
197 : 1 : sigjmp_buf local_sigjmp_buf;
198 : 1 : MemoryContext checkpointer_context;
199 : :
200 [ + - ]: 1 : Assert(startup_data_len == 0);
201 : :
202 : 1 : MyBackendType = B_CHECKPOINTER;
203 : 1 : AuxiliaryProcessMainCommon();
204 : :
205 : 1 : CheckpointerShmem->checkpointer_pid = MyProcPid;
206 : :
207 : : /*
208 : : * Properly accept or ignore signals the postmaster might send us
209 : : *
210 : : * Note: we deliberately ignore SIGTERM, because during a standard Unix
211 : : * system shutdown cycle, init will SIGTERM all processes at once. We
212 : : * want to wait for the backends to exit, whereupon the postmaster will
213 : : * tell us it's okay to shut down (via SIGUSR2).
214 : : */
215 : 1 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
216 : 1 : pqsignal(SIGINT, ReqShutdownXLOG);
217 : 1 : pqsignal(SIGTERM, SIG_IGN); /* ignore SIGTERM */
218 : : /* SIGQUIT handler was already set up by InitPostmasterChild */
219 : 1 : pqsignal(SIGALRM, SIG_IGN);
220 : 1 : pqsignal(SIGPIPE, SIG_IGN);
221 : 1 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
222 : 1 : pqsignal(SIGUSR2, SignalHandlerForShutdownRequest);
223 : :
224 : : /*
225 : : * Reset some signals that are accepted by postmaster but not here
226 : : */
227 : 1 : pqsignal(SIGCHLD, SIG_DFL);
228 : :
229 : : /*
230 : : * Initialize so that first time-driven event happens at the correct time.
231 : : */
232 : 1 : last_checkpoint_time = last_xlog_switch_time = (pg_time_t) time(NULL);
233 : :
234 : : /*
235 : : * Write out stats after shutdown. This needs to be called by exactly one
236 : : * process during a normal shutdown, and since checkpointer is shut down
237 : : * very late...
238 : : *
239 : : * While e.g. walsenders are active after the shutdown checkpoint has been
240 : : * written (and thus could produce more stats), checkpointer stays around
241 : : * after the shutdown checkpoint has been written. postmaster will only
242 : : * signal checkpointer to exit after all processes that could emit stats
243 : : * have been shut down.
244 : : */
245 : 1 : before_shmem_exit(pgstat_before_server_shutdown, 0);
246 : :
247 : : /*
248 : : * Create a memory context that we will do all our work in. We do this so
249 : : * that we can reset the context during error recovery and thereby avoid
250 : : * possible memory leaks. Formerly this code just ran in
251 : : * TopMemoryContext, but resetting that would be a really bad idea.
252 : : */
253 : 1 : checkpointer_context = AllocSetContextCreate(TopMemoryContext,
254 : : "Checkpointer",
255 : : ALLOCSET_DEFAULT_SIZES);
256 : 1 : MemoryContextSwitchTo(checkpointer_context);
257 : :
258 : : /*
259 : : * If an exception is encountered, processing resumes here.
260 : : *
261 : : * You might wonder why this isn't coded as an infinite loop around a
262 : : * PG_TRY construct. The reason is that this is the bottom of the
263 : : * exception stack, and so with PG_TRY there would be no exception handler
264 : : * in force at all during the CATCH part. By leaving the outermost setjmp
265 : : * always active, we have at least some chance of recovering from an error
266 : : * during error recovery. (If we get into an infinite loop thereby, it
267 : : * will soon be stopped by overflow of elog.c's internal state stack.)
268 : : *
269 : : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
270 : : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
271 : : * signals other than SIGQUIT will be blocked until we complete error
272 : : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
273 : : * call redundant, but it is not since InterruptPending might be set
274 : : * already.
275 : : */
276 [ + - ]: 1 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
277 : : {
278 : : /* Since not using PG_TRY, must reset error stack by hand */
279 : 0 : error_context_stack = NULL;
280 : :
281 : : /* Prevent interrupts while cleaning up */
282 : 0 : HOLD_INTERRUPTS();
283 : :
284 : : /* Report the error to the server log */
285 : 0 : EmitErrorReport();
286 : :
287 : : /*
288 : : * These operations are really just a minimal subset of
289 : : * AbortTransaction(). We don't have very many resources to worry
290 : : * about in checkpointer, but we do have LWLocks, buffers, and temp
291 : : * files.
292 : : */
293 : 0 : LWLockReleaseAll();
294 : 0 : ConditionVariableCancelSleep();
295 : 0 : pgstat_report_wait_end();
296 : 0 : pgaio_error_cleanup();
297 : 0 : UnlockBuffers();
298 : 0 : ReleaseAuxProcessResources(false);
299 : 0 : AtEOXact_Buffers(false);
300 : 0 : AtEOXact_SMgr();
301 : 0 : AtEOXact_Files(false);
302 : 0 : AtEOXact_HashTables(false);
303 : :
304 : : /* Warn any waiting backends that the checkpoint failed. */
305 [ # # ]: 0 : if (ckpt_active)
306 : : {
307 [ # # ]: 0 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
308 : 0 : CheckpointerShmem->ckpt_failed++;
309 : 0 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
310 : 0 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
311 : :
312 : 0 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
313 : :
314 : 0 : ckpt_active = false;
315 : 0 : }
316 : :
317 : : /*
318 : : * Now return to normal top-level context and clear ErrorContext for
319 : : * next time.
320 : : */
321 : 0 : MemoryContextSwitchTo(checkpointer_context);
322 : 0 : FlushErrorState();
323 : :
324 : : /* Flush any leaked data in the top-level context */
325 : 0 : MemoryContextReset(checkpointer_context);
326 : :
327 : : /* Now we can allow interrupts again */
328 [ # # ]: 0 : RESUME_INTERRUPTS();
329 : :
330 : : /*
331 : : * Sleep at least 1 second after any error. A write error is likely
332 : : * to be repeated, and we don't want to be filling the error logs as
333 : : * fast as we can.
334 : : */
335 : 0 : pg_usleep(1000000L);
336 : 0 : }
337 : :
338 : : /* We can now handle ereport(ERROR) */
339 : 1 : PG_exception_stack = &local_sigjmp_buf;
340 : :
341 : : /*
342 : : * Unblock signals (they were blocked when the postmaster forked us)
343 : : */
344 : 1 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
345 : :
346 : : /*
347 : : * Ensure all shared memory values are set correctly for the config. Doing
348 : : * this here ensures no race conditions from other concurrent updaters.
349 : : */
350 : 1 : UpdateSharedMemoryConfig();
351 : :
352 : : /*
353 : : * Advertise our proc number that backends can use to wake us up while
354 : : * we're sleeping.
355 : : */
356 : 1 : ProcGlobal->checkpointerProc = MyProcNumber;
357 : :
358 : : /*
359 : : * Loop until we've been asked to write the shutdown checkpoint or
360 : : * terminate.
361 : : */
362 : 4 : for (;;)
363 : : {
364 : 4 : bool do_checkpoint = false;
365 : 4 : int flags = 0;
366 : 4 : pg_time_t now;
367 : 4 : int elapsed_secs;
368 : 4 : int cur_timeout;
369 : 4 : bool chkpt_or_rstpt_requested = false;
370 : 4 : bool chkpt_or_rstpt_timed = false;
371 : :
372 : : /* Clear any already-pending wakeups */
373 : 4 : ResetLatch(MyLatch);
374 : :
375 : : /*
376 : : * Process any requests or signals received recently.
377 : : */
378 : 4 : AbsorbSyncRequests();
379 : :
380 : 4 : ProcessCheckpointerInterrupts();
381 [ + + - + ]: 4 : if (ShutdownXLOGPending || ShutdownRequestPending)
382 : 1 : break;
383 : :
384 : : /*
385 : : * Detect a pending checkpoint request by checking whether the flags
386 : : * word in shared memory is nonzero. We shouldn't need to acquire the
387 : : * ckpt_lck for this.
388 : : */
389 [ + - ]: 3 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
390 : : {
391 : 0 : do_checkpoint = true;
392 : 0 : chkpt_or_rstpt_requested = true;
393 : 0 : }
394 : :
395 : : /*
396 : : * Force a checkpoint if too much time has elapsed since the last one.
397 : : * Note that we count a timed checkpoint in stats only when this
398 : : * occurs without an external request, but we set the CAUSE_TIME flag
399 : : * bit even if there is also an external request.
400 : : */
401 : 3 : now = (pg_time_t) time(NULL);
402 : 3 : elapsed_secs = now - last_checkpoint_time;
403 [ + - ]: 3 : if (elapsed_secs >= CheckPointTimeout)
404 : : {
405 [ # # ]: 0 : if (!do_checkpoint)
406 : 0 : chkpt_or_rstpt_timed = true;
407 : 0 : do_checkpoint = true;
408 : 0 : flags |= CHECKPOINT_CAUSE_TIME;
409 : 0 : }
410 : :
411 : : /*
412 : : * Do a checkpoint if requested.
413 : : */
414 [ + - ]: 3 : if (do_checkpoint)
415 : : {
416 : 0 : bool ckpt_performed = false;
417 : 0 : bool do_restartpoint;
418 : :
419 : : /* Check if we should perform a checkpoint or a restartpoint. */
420 : 0 : do_restartpoint = RecoveryInProgress();
421 : :
422 : : /*
423 : : * Atomically fetch the request flags to figure out what kind of a
424 : : * checkpoint we should perform, and increase the started-counter
425 : : * to acknowledge that we've started a new checkpoint.
426 : : */
427 [ # # ]: 0 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
428 : 0 : flags |= CheckpointerShmem->ckpt_flags;
429 : 0 : CheckpointerShmem->ckpt_flags = 0;
430 : 0 : CheckpointerShmem->ckpt_started++;
431 : 0 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
432 : :
433 : 0 : ConditionVariableBroadcast(&CheckpointerShmem->start_cv);
434 : :
435 : : /*
436 : : * The end-of-recovery checkpoint is a real checkpoint that's
437 : : * performed while we're still in recovery.
438 : : */
439 [ # # ]: 0 : if (flags & CHECKPOINT_END_OF_RECOVERY)
440 : 0 : do_restartpoint = false;
441 : :
442 [ # # ]: 0 : if (chkpt_or_rstpt_timed)
443 : : {
444 : 0 : chkpt_or_rstpt_timed = false;
445 [ # # ]: 0 : if (do_restartpoint)
446 : 0 : PendingCheckpointerStats.restartpoints_timed++;
447 : : else
448 : 0 : PendingCheckpointerStats.num_timed++;
449 : 0 : }
450 : :
451 [ # # ]: 0 : if (chkpt_or_rstpt_requested)
452 : : {
453 : 0 : chkpt_or_rstpt_requested = false;
454 [ # # ]: 0 : if (do_restartpoint)
455 : 0 : PendingCheckpointerStats.restartpoints_requested++;
456 : : else
457 : 0 : PendingCheckpointerStats.num_requested++;
458 : 0 : }
459 : :
460 : : /*
461 : : * We will warn if (a) too soon since last checkpoint (whatever
462 : : * caused it) and (b) somebody set the CHECKPOINT_CAUSE_XLOG flag
463 : : * since the last checkpoint start. Note in particular that this
464 : : * implementation will not generate warnings caused by
465 : : * CheckPointTimeout < CheckPointWarning.
466 : : */
467 [ # # ]: 0 : if (!do_restartpoint &&
468 [ # # # # ]: 0 : (flags & CHECKPOINT_CAUSE_XLOG) &&
469 : 0 : elapsed_secs < CheckPointWarning)
470 [ # # # # ]: 0 : ereport(LOG,
471 : : (errmsg_plural("checkpoints are occurring too frequently (%d second apart)",
472 : : "checkpoints are occurring too frequently (%d seconds apart)",
473 : : elapsed_secs,
474 : : elapsed_secs),
475 : : errhint("Consider increasing the configuration parameter \"%s\".", "max_wal_size")));
476 : :
477 : : /*
478 : : * Initialize checkpointer-private variables used during
479 : : * checkpoint.
480 : : */
481 : 0 : ckpt_active = true;
482 [ # # ]: 0 : if (do_restartpoint)
483 : 0 : ckpt_start_recptr = GetXLogReplayRecPtr(NULL);
484 : : else
485 : 0 : ckpt_start_recptr = GetInsertRecPtr();
486 : 0 : ckpt_start_time = now;
487 : 0 : ckpt_cached_elapsed = 0;
488 : :
489 : : /*
490 : : * Do the checkpoint.
491 : : */
492 [ # # ]: 0 : if (!do_restartpoint)
493 : 0 : ckpt_performed = CreateCheckPoint(flags);
494 : : else
495 : 0 : ckpt_performed = CreateRestartPoint(flags);
496 : :
497 : : /*
498 : : * After any checkpoint, free all smgr objects. Otherwise we
499 : : * would never do so for dropped relations, as the checkpointer
500 : : * does not process shared invalidation messages or call
501 : : * AtEOXact_SMgr().
502 : : */
503 : 0 : smgrdestroyall();
504 : :
505 : : /*
506 : : * Indicate checkpoint completion to any waiting backends.
507 : : */
508 [ # # ]: 0 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
509 : 0 : CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started;
510 : 0 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
511 : :
512 : 0 : ConditionVariableBroadcast(&CheckpointerShmem->done_cv);
513 : :
514 [ # # ]: 0 : if (!do_restartpoint)
515 : : {
516 : : /*
517 : : * Note we record the checkpoint start time not end time as
518 : : * last_checkpoint_time. This is so that time-driven
519 : : * checkpoints happen at a predictable spacing.
520 : : */
521 : 0 : last_checkpoint_time = now;
522 : :
523 [ # # ]: 0 : if (ckpt_performed)
524 : 0 : PendingCheckpointerStats.num_performed++;
525 : 0 : }
526 : : else
527 : : {
528 [ # # ]: 0 : if (ckpt_performed)
529 : : {
530 : : /*
531 : : * The same as for checkpoint. Please see the
532 : : * corresponding comment.
533 : : */
534 : 0 : last_checkpoint_time = now;
535 : :
536 : 0 : PendingCheckpointerStats.restartpoints_performed++;
537 : 0 : }
538 : : else
539 : : {
540 : : /*
541 : : * We were not able to perform the restartpoint
542 : : * (checkpoints throw an ERROR in case of error). Most
543 : : * likely because we have not received any new checkpoint
544 : : * WAL records since the last restartpoint. Try again in
545 : : * 15 s.
546 : : */
547 : 0 : last_checkpoint_time = now - CheckPointTimeout + 15;
548 : : }
549 : : }
550 : :
551 : 0 : ckpt_active = false;
552 : :
553 : : /*
554 : : * We may have received an interrupt during the checkpoint and the
555 : : * latch might have been reset (e.g. in CheckpointWriteDelay).
556 : : */
557 : 0 : ProcessCheckpointerInterrupts();
558 [ # # # # ]: 0 : if (ShutdownXLOGPending || ShutdownRequestPending)
559 : 0 : break;
560 [ # # ]: 0 : }
561 : :
562 : : /*
563 : : * Disable logical decoding if someone requested it. See comments atop
564 : : * logicalctl.c.
565 : : */
566 : 3 : DisableLogicalDecodingIfNecessary();
567 : :
568 : : /* Check for archive_timeout and switch xlog files if necessary. */
569 : 3 : CheckArchiveTimeout();
570 : :
571 : : /* Report pending statistics to the cumulative stats system */
572 : 3 : pgstat_report_checkpointer();
573 : 3 : pgstat_report_wal(true);
574 : :
575 : : /*
576 : : * If any checkpoint flags have been set, redo the loop to handle the
577 : : * checkpoint without sleeping.
578 : : */
579 [ - + ]: 3 : if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags)
580 : 0 : continue;
581 : :
582 : : /*
583 : : * Sleep until we are signaled or it's time for another checkpoint or
584 : : * xlog file switch.
585 : : */
586 : 3 : now = (pg_time_t) time(NULL);
587 : 3 : elapsed_secs = now - last_checkpoint_time;
588 [ - + ]: 3 : if (elapsed_secs >= CheckPointTimeout)
589 : 0 : continue; /* no sleep for us ... */
590 : 3 : cur_timeout = CheckPointTimeout - elapsed_secs;
591 [ - + # # ]: 3 : if (XLogArchiveTimeout > 0 && !RecoveryInProgress())
592 : : {
593 : 0 : elapsed_secs = now - last_xlog_switch_time;
594 [ # # ]: 0 : if (elapsed_secs >= XLogArchiveTimeout)
595 : 0 : continue; /* no sleep for us ... */
596 [ # # ]: 0 : cur_timeout = Min(cur_timeout, XLogArchiveTimeout - elapsed_secs);
597 : 0 : }
598 : :
599 : 6 : (void) WaitLatch(MyLatch,
600 : : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
601 : 3 : cur_timeout * 1000L /* convert to ms */ ,
602 : : WAIT_EVENT_CHECKPOINTER_MAIN);
603 [ + - - + ]: 4 : }
604 : :
605 : : /*
606 : : * From here on, elog(ERROR) should end with exit(1), not send control
607 : : * back to the sigsetjmp block above.
608 : : */
609 : 1 : ExitOnAnyError = true;
610 : :
611 [ - + ]: 1 : if (ShutdownXLOGPending)
612 : : {
613 : : /*
614 : : * Close down the database.
615 : : *
616 : : * Since ShutdownXLOG() creates restartpoint or checkpoint, and
617 : : * updates the statistics, increment the checkpoint request and flush
618 : : * out pending statistic.
619 : : */
620 : 1 : PendingCheckpointerStats.num_requested++;
621 : 1 : ShutdownXLOG(0, 0);
622 : 1 : pgstat_report_checkpointer();
623 : 1 : pgstat_report_wal(true);
624 : :
625 : : /*
626 : : * Tell postmaster that we're done.
627 : : */
628 : 1 : SendPostmasterSignal(PMSIGNAL_XLOG_IS_SHUTDOWN);
629 : 1 : ShutdownXLOGPending = false;
630 : 1 : }
631 : :
632 : : /*
633 : : * Wait until we're asked to shut down. By separating the writing of the
634 : : * shutdown checkpoint from checkpointer exiting, checkpointer can perform
635 : : * some should-be-as-late-as-possible work like writing out stats.
636 : : */
637 : 2 : for (;;)
638 : : {
639 : : /* Clear any already-pending wakeups */
640 : 2 : ResetLatch(MyLatch);
641 : :
642 : 2 : ProcessCheckpointerInterrupts();
643 : :
644 [ + + ]: 2 : if (ShutdownRequestPending)
645 : 1 : break;
646 : :
647 : 1 : (void) WaitLatch(MyLatch,
648 : : WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
649 : : 0,
650 : : WAIT_EVENT_CHECKPOINTER_SHUTDOWN);
651 : : }
652 : :
653 : : /* Normal exit from the checkpointer is here */
654 : 1 : proc_exit(0); /* done */
655 : : }
656 : :
657 : : /*
658 : : * Process any new interrupts.
659 : : */
660 : : static void
661 : 6 : ProcessCheckpointerInterrupts(void)
662 : : {
663 [ + + ]: 6 : if (ProcSignalBarrierPending)
664 : 1 : ProcessProcSignalBarrier();
665 : :
666 [ + - ]: 6 : if (ConfigReloadPending)
667 : : {
668 : 0 : ConfigReloadPending = false;
669 : 0 : ProcessConfigFile(PGC_SIGHUP);
670 : :
671 : : /*
672 : : * Checkpointer is the last process to shut down, so we ask it to hold
673 : : * the keys for a range of other tasks required most of which have
674 : : * nothing to do with checkpointing at all.
675 : : *
676 : : * For various reasons, some config values can change dynamically so
677 : : * the primary copy of them is held in shared memory to make sure all
678 : : * backends see the same value. We make Checkpointer responsible for
679 : : * updating the shared memory copy if the parameter setting changes
680 : : * because of SIGHUP.
681 : : */
682 : 0 : UpdateSharedMemoryConfig();
683 : 0 : }
684 : :
685 : : /* Perform logging of memory contexts of this process */
686 [ + - ]: 6 : if (LogMemoryContextPending)
687 : 0 : ProcessLogMemoryContextInterrupt();
688 : 6 : }
689 : :
690 : : /*
691 : : * CheckArchiveTimeout -- check for archive_timeout and switch xlog files
692 : : *
693 : : * This will switch to a new WAL file and force an archive file write if
694 : : * meaningful activity is recorded in the current WAL file. This includes most
695 : : * writes, including just a single checkpoint record, but excludes WAL records
696 : : * that were inserted with the XLOG_MARK_UNIMPORTANT flag being set (like
697 : : * snapshots of running transactions). Such records, depending on
698 : : * configuration, occur on regular intervals and don't contain important
699 : : * information. This avoids generating archives with a few unimportant
700 : : * records.
701 : : */
702 : : static void
703 : 3 : CheckArchiveTimeout(void)
704 : : {
705 : 3 : pg_time_t now;
706 : 3 : pg_time_t last_time;
707 : 3 : XLogRecPtr last_switch_lsn;
708 : :
709 [ - + # # ]: 3 : if (XLogArchiveTimeout <= 0 || RecoveryInProgress())
710 : 3 : return;
711 : :
712 : 0 : now = (pg_time_t) time(NULL);
713 : :
714 : : /* First we do a quick check using possibly-stale local state. */
715 [ # # ]: 0 : if ((int) (now - last_xlog_switch_time) < XLogArchiveTimeout)
716 : 0 : return;
717 : :
718 : : /*
719 : : * Update local state ... note that last_xlog_switch_time is the last time
720 : : * a switch was performed *or requested*.
721 : : */
722 : 0 : last_time = GetLastSegSwitchData(&last_switch_lsn);
723 : :
724 [ # # ]: 0 : last_xlog_switch_time = Max(last_xlog_switch_time, last_time);
725 : :
726 : : /* Now we can do the real checks */
727 [ # # ]: 0 : if ((int) (now - last_xlog_switch_time) >= XLogArchiveTimeout)
728 : : {
729 : : /*
730 : : * Switch segment only when "important" WAL has been logged since the
731 : : * last segment switch (last_switch_lsn points to end of segment
732 : : * switch occurred in).
733 : : */
734 [ # # ]: 0 : if (GetLastImportantRecPtr() > last_switch_lsn)
735 : : {
736 : 0 : XLogRecPtr switchpoint;
737 : :
738 : : /* mark switch as unimportant, avoids triggering checkpoints */
739 : 0 : switchpoint = RequestXLogSwitch(true);
740 : :
741 : : /*
742 : : * If the returned pointer points exactly to a segment boundary,
743 : : * assume nothing happened.
744 : : */
745 [ # # ]: 0 : if (XLogSegmentOffset(switchpoint, wal_segment_size) != 0)
746 [ # # # # ]: 0 : elog(DEBUG1, "write-ahead log switch forced (\"archive_timeout\"=%d)",
747 : : XLogArchiveTimeout);
748 : 0 : }
749 : :
750 : : /*
751 : : * Update state in any case, so we don't retry constantly when the
752 : : * system is idle.
753 : : */
754 : 0 : last_xlog_switch_time = now;
755 : 0 : }
756 [ - + ]: 3 : }
757 : :
758 : : /*
759 : : * Returns true if a fast checkpoint request is pending. (Note that this does
760 : : * not check the *current* checkpoint's FAST flag, but whether there is one
761 : : * pending behind it.)
762 : : */
763 : : static bool
764 : 0 : FastCheckpointRequested(void)
765 : : {
766 : 0 : volatile CheckpointerShmemStruct *cps = CheckpointerShmem;
767 : :
768 : : /*
769 : : * We don't need to acquire the ckpt_lck in this case because we're only
770 : : * looking at a single flag bit.
771 : : */
772 [ # # ]: 0 : if (cps->ckpt_flags & CHECKPOINT_FAST)
773 : 0 : return true;
774 : 0 : return false;
775 : 0 : }
776 : :
777 : : /*
778 : : * CheckpointWriteDelay -- control rate of checkpoint
779 : : *
780 : : * This function is called after each page write performed by BufferSync().
781 : : * It is responsible for throttling BufferSync()'s write rate to hit
782 : : * checkpoint_completion_target.
783 : : *
784 : : * The checkpoint request flags should be passed in; currently the only one
785 : : * examined is CHECKPOINT_FAST, which disables delays between writes.
786 : : *
787 : : * 'progress' is an estimate of how much of the work has been done, as a
788 : : * fraction between 0.0 meaning none, and 1.0 meaning all done.
789 : : */
790 : : void
791 : 5230 : CheckpointWriteDelay(int flags, double progress)
792 : : {
793 : : static int absorb_counter = WRITES_PER_ABSORB;
794 : :
795 : : /* Do nothing if checkpoint is being executed by non-checkpointer process */
796 [ + + ]: 5230 : if (!AmCheckpointerProcess())
797 : 1052 : return;
798 : :
799 : : /*
800 : : * Perform the usual duties and take a nap, unless we're behind schedule,
801 : : * in which case we just try to catch up as quickly as possible.
802 : : */
803 [ - + ]: 4178 : if (!(flags & CHECKPOINT_FAST) &&
804 [ # # ]: 0 : !ShutdownXLOGPending &&
805 [ # # ]: 0 : !ShutdownRequestPending &&
806 [ # # # # ]: 0 : !FastCheckpointRequested() &&
807 : 0 : IsCheckpointOnSchedule(progress))
808 : : {
809 [ # # ]: 0 : if (ConfigReloadPending)
810 : : {
811 : 0 : ConfigReloadPending = false;
812 : 0 : ProcessConfigFile(PGC_SIGHUP);
813 : : /* update shmem copies of config variables */
814 : 0 : UpdateSharedMemoryConfig();
815 : 0 : }
816 : :
817 : 0 : AbsorbSyncRequests();
818 : 0 : absorb_counter = WRITES_PER_ABSORB;
819 : :
820 : 0 : CheckArchiveTimeout();
821 : :
822 : : /* Report interim statistics to the cumulative stats system */
823 : 0 : pgstat_report_checkpointer();
824 : :
825 : : /*
826 : : * This sleep used to be connected to bgwriter_delay, typically 200ms.
827 : : * That resulted in more frequent wakeups if not much work to do.
828 : : * Checkpointer and bgwriter are no longer related so take the Big
829 : : * Sleep.
830 : : */
831 : 0 : WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | WL_TIMEOUT,
832 : : 100,
833 : : WAIT_EVENT_CHECKPOINT_WRITE_DELAY);
834 : 0 : ResetLatch(MyLatch);
835 : 0 : }
836 [ + + ]: 4178 : else if (--absorb_counter <= 0)
837 : : {
838 : : /*
839 : : * Absorb pending fsync requests after each WRITES_PER_ABSORB write
840 : : * operations even when we don't sleep, to prevent overflow of the
841 : : * fsync request queue.
842 : : */
843 : 4 : AbsorbSyncRequests();
844 : 4 : absorb_counter = WRITES_PER_ABSORB;
845 : 4 : }
846 : :
847 : : /* Check for barrier events. */
848 [ + - ]: 4178 : if (ProcSignalBarrierPending)
849 : 0 : ProcessProcSignalBarrier();
850 : 5230 : }
851 : :
852 : : /*
853 : : * IsCheckpointOnSchedule -- are we on schedule to finish this checkpoint
854 : : * (or restartpoint) in time?
855 : : *
856 : : * Compares the current progress against the time/segments elapsed since last
857 : : * checkpoint, and returns true if the progress we've made this far is greater
858 : : * than the elapsed time/segments.
859 : : */
860 : : static bool
861 : 0 : IsCheckpointOnSchedule(double progress)
862 : : {
863 : 0 : XLogRecPtr recptr;
864 : 0 : struct timeval now;
865 : 0 : double elapsed_xlogs,
866 : : elapsed_time;
867 : :
868 [ # # ]: 0 : Assert(ckpt_active);
869 : :
870 : : /* Scale progress according to checkpoint_completion_target. */
871 : 0 : progress *= CheckPointCompletionTarget;
872 : :
873 : : /*
874 : : * Check against the cached value first. Only do the more expensive
875 : : * calculations once we reach the target previously calculated. Since
876 : : * neither time or WAL insert pointer moves backwards, a freshly
877 : : * calculated value can only be greater than or equal to the cached value.
878 : : */
879 [ # # ]: 0 : if (progress < ckpt_cached_elapsed)
880 : 0 : return false;
881 : :
882 : : /*
883 : : * Check progress against WAL segments written and CheckPointSegments.
884 : : *
885 : : * We compare the current WAL insert location against the location
886 : : * computed before calling CreateCheckPoint. The code in XLogInsert that
887 : : * actually triggers a checkpoint when CheckPointSegments is exceeded
888 : : * compares against RedoRecPtr, so this is not completely accurate.
889 : : * However, it's good enough for our purposes, we're only calculating an
890 : : * estimate anyway.
891 : : *
892 : : * During recovery, we compare last replayed WAL record's location with
893 : : * the location computed before calling CreateRestartPoint. That maintains
894 : : * the same pacing as we have during checkpoints in normal operation, but
895 : : * we might exceed max_wal_size by a fair amount. That's because there can
896 : : * be a large gap between a checkpoint's redo-pointer and the checkpoint
897 : : * record itself, and we only start the restartpoint after we've seen the
898 : : * checkpoint record. (The gap is typically up to CheckPointSegments *
899 : : * checkpoint_completion_target where checkpoint_completion_target is the
900 : : * value that was in effect when the WAL was generated).
901 : : */
902 [ # # ]: 0 : if (RecoveryInProgress())
903 : 0 : recptr = GetXLogReplayRecPtr(NULL);
904 : : else
905 : 0 : recptr = GetInsertRecPtr();
906 : 0 : elapsed_xlogs = (((double) (recptr - ckpt_start_recptr)) /
907 : 0 : wal_segment_size) / CheckPointSegments;
908 : :
909 [ # # ]: 0 : if (progress < elapsed_xlogs)
910 : : {
911 : 0 : ckpt_cached_elapsed = elapsed_xlogs;
912 : 0 : return false;
913 : : }
914 : :
915 : : /*
916 : : * Check progress against time elapsed and checkpoint_timeout.
917 : : */
918 : 0 : gettimeofday(&now, NULL);
919 : 0 : elapsed_time = ((double) ((pg_time_t) now.tv_sec - ckpt_start_time) +
920 : 0 : now.tv_usec / 1000000.0) / CheckPointTimeout;
921 : :
922 [ # # ]: 0 : if (progress < elapsed_time)
923 : : {
924 : 0 : ckpt_cached_elapsed = elapsed_time;
925 : 0 : return false;
926 : : }
927 : :
928 : : /* It looks like we're on schedule. */
929 : 0 : return true;
930 : 0 : }
931 : :
932 : :
933 : : /* --------------------------------
934 : : * signal handler routines
935 : : * --------------------------------
936 : : */
937 : :
938 : : /* SIGINT: set flag to trigger writing of shutdown checkpoint */
939 : : static void
940 : 1 : ReqShutdownXLOG(SIGNAL_ARGS)
941 : : {
942 : 1 : ShutdownXLOGPending = true;
943 : 1 : SetLatch(MyLatch);
944 : 1 : }
945 : :
946 : :
947 : : /* --------------------------------
948 : : * communication with backends
949 : : * --------------------------------
950 : : */
951 : :
952 : : /*
953 : : * CheckpointerShmemSize
954 : : * Compute space needed for checkpointer-related shared memory
955 : : */
956 : : Size
957 : 15 : CheckpointerShmemSize(void)
958 : : {
959 : 15 : Size size;
960 : :
961 : : /*
962 : : * The size of the requests[] array is arbitrarily set equal to NBuffers.
963 : : * But there is a cap of MAX_CHECKPOINT_REQUESTS to prevent accumulating
964 : : * too many checkpoint requests in the ring buffer.
965 : : */
966 : 15 : size = offsetof(CheckpointerShmemStruct, requests);
967 [ + - ]: 15 : size = add_size(size, mul_size(Min(NBuffers,
968 : : MAX_CHECKPOINT_REQUESTS),
969 : : sizeof(CheckpointerRequest)));
970 : :
971 : 30 : return size;
972 : 15 : }
973 : :
974 : : /*
975 : : * CheckpointerShmemInit
976 : : * Allocate and initialize checkpointer-related shared memory
977 : : */
978 : : void
979 : 6 : CheckpointerShmemInit(void)
980 : : {
981 : 6 : Size size = CheckpointerShmemSize();
982 : 6 : bool found;
983 : :
984 : 6 : CheckpointerShmem = (CheckpointerShmemStruct *)
985 : 6 : ShmemInitStruct("Checkpointer Data",
986 : 6 : size,
987 : : &found);
988 : :
989 [ - + ]: 6 : if (!found)
990 : : {
991 : : /*
992 : : * First time through, so initialize. Note that we zero the whole
993 : : * requests array; this is so that CompactCheckpointerRequestQueue can
994 : : * assume that any pad bytes in the request structs are zeroes.
995 : : */
996 [ + - + - : 6 : MemSet(CheckpointerShmem, 0, size);
+ - + - #
# ]
997 : 6 : SpinLockInit(&CheckpointerShmem->ckpt_lck);
998 [ + - ]: 6 : CheckpointerShmem->max_requests = Min(NBuffers, MAX_CHECKPOINT_REQUESTS);
999 : 6 : CheckpointerShmem->head = CheckpointerShmem->tail = 0;
1000 : 6 : ConditionVariableInit(&CheckpointerShmem->start_cv);
1001 : 6 : ConditionVariableInit(&CheckpointerShmem->done_cv);
1002 : 6 : }
1003 : 6 : }
1004 : :
1005 : : /*
1006 : : * ExecCheckpoint
1007 : : * Primary entry point for manual CHECKPOINT commands
1008 : : *
1009 : : * This is mainly a wrapper for RequestCheckpoint().
1010 : : */
1011 : : void
1012 : 9 : ExecCheckpoint(ParseState *pstate, CheckPointStmt *stmt)
1013 : : {
1014 : 9 : bool fast = true;
1015 : 9 : bool unlogged = false;
1016 : :
1017 [ + + + + : 19 : foreach_ptr(DefElem, opt, stmt->options)
+ + + + ]
1018 : : {
1019 [ + + ]: 5 : if (strcmp(opt->defname, "mode") == 0)
1020 : : {
1021 : 2 : char *mode = defGetString(opt);
1022 : :
1023 [ + - ]: 2 : if (strcmp(mode, "spread") == 0)
1024 : 0 : fast = false;
1025 [ + + ]: 2 : else if (strcmp(mode, "fast") != 0)
1026 [ + - + - ]: 1 : ereport(ERROR,
1027 : : (errcode(ERRCODE_SYNTAX_ERROR),
1028 : : errmsg("unrecognized value for %s option \"%s\": \"%s\"",
1029 : : "CHECKPOINT", "mode", mode),
1030 : : parser_errposition(pstate, opt->location)));
1031 : 1 : }
1032 [ + + ]: 3 : else if (strcmp(opt->defname, "flush_unlogged") == 0)
1033 : 2 : unlogged = defGetBoolean(opt);
1034 : : else
1035 [ + - + - ]: 1 : ereport(ERROR,
1036 : : (errcode(ERRCODE_SYNTAX_ERROR),
1037 : : errmsg("unrecognized %s option \"%s\"",
1038 : : "CHECKPOINT", opt->defname),
1039 : : parser_errposition(pstate, opt->location)));
1040 : 10 : }
1041 : :
1042 [ + - ]: 7 : if (!has_privs_of_role(GetUserId(), ROLE_PG_CHECKPOINT))
1043 [ # # # # ]: 0 : ereport(ERROR,
1044 : : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
1045 : : /* translator: %s is name of an SQL command (e.g., CHECKPOINT) */
1046 : : errmsg("permission denied to execute %s command",
1047 : : "CHECKPOINT"),
1048 : : errdetail("Only roles with privileges of the \"%s\" role may execute this command.",
1049 : : "pg_checkpoint")));
1050 : :
1051 : 14 : RequestCheckpoint(CHECKPOINT_WAIT |
1052 : 14 : (fast ? CHECKPOINT_FAST : 0) |
1053 : 14 : (unlogged ? CHECKPOINT_FLUSH_UNLOGGED : 0) |
1054 : 7 : (RecoveryInProgress() ? 0 : CHECKPOINT_FORCE));
1055 : 7 : }
1056 : :
1057 : : /*
1058 : : * RequestCheckpoint
1059 : : * Called in backend processes to request a checkpoint
1060 : : *
1061 : : * flags is a bitwise OR of the following:
1062 : : * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
1063 : : * CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery.
1064 : : * CHECKPOINT_FAST: finish the checkpoint ASAP,
1065 : : * ignoring checkpoint_completion_target parameter.
1066 : : * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred
1067 : : * since the last one (implied by CHECKPOINT_IS_SHUTDOWN or
1068 : : * CHECKPOINT_END_OF_RECOVERY, and the CHECKPOINT command).
1069 : : * CHECKPOINT_WAIT: wait for completion before returning (otherwise,
1070 : : * just signal checkpointer to do it, and return).
1071 : : * CHECKPOINT_CAUSE_XLOG: checkpoint is requested due to xlog filling.
1072 : : * (This affects logging, and in particular enables CheckPointWarning.)
1073 : : */
1074 : : void
1075 : 18 : RequestCheckpoint(int flags)
1076 : : {
1077 : 18 : int ntries;
1078 : 18 : int old_failed,
1079 : : old_started;
1080 : :
1081 : : /*
1082 : : * If in a standalone backend, just do it ourselves.
1083 : : */
1084 [ + + ]: 18 : if (!IsPostmasterEnvironment)
1085 : : {
1086 : : /*
1087 : : * There's no point in doing slow checkpoints in a standalone backend,
1088 : : * because there's no other backends the checkpoint could disrupt.
1089 : : */
1090 : 4 : CreateCheckPoint(flags | CHECKPOINT_FAST);
1091 : :
1092 : : /* Free all smgr objects, as CheckpointerMain() normally would. */
1093 : 4 : smgrdestroyall();
1094 : :
1095 : 4 : return;
1096 : : }
1097 : :
1098 : : /*
1099 : : * Atomically set the request flags, and take a snapshot of the counters.
1100 : : * When we see ckpt_started > old_started, we know the flags we set here
1101 : : * have been seen by checkpointer.
1102 : : *
1103 : : * Note that we OR the flags with any existing flags, to avoid overriding
1104 : : * a "stronger" request by another backend. The flag senses must be
1105 : : * chosen to make this work!
1106 : : */
1107 [ - + ]: 14 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1108 : :
1109 : 14 : old_failed = CheckpointerShmem->ckpt_failed;
1110 : 14 : old_started = CheckpointerShmem->ckpt_started;
1111 : 14 : CheckpointerShmem->ckpt_flags |= (flags | CHECKPOINT_REQUESTED);
1112 : :
1113 : 14 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1114 : :
1115 : : /*
1116 : : * Set checkpointer's latch to request checkpoint. It's possible that the
1117 : : * checkpointer hasn't started yet, so we will retry a few times if
1118 : : * needed. (Actually, more than a few times, since on slow or overloaded
1119 : : * buildfarm machines, it's been observed that the checkpointer can take
1120 : : * several seconds to start.) However, if not told to wait for the
1121 : : * checkpoint to occur, we consider failure to set the latch to be
1122 : : * nonfatal and merely LOG it. The checkpointer should see the request
1123 : : * when it does start, with or without the SetLatch().
1124 : : */
1125 : : #define MAX_SIGNAL_TRIES 600 /* max wait 60.0 sec */
1126 : 14 : for (ntries = 0;; ntries++)
1127 : : {
1128 : 14 : volatile PROC_HDR *procglobal = ProcGlobal;
1129 : 14 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1130 : :
1131 [ + - ]: 14 : if (checkpointerProc == INVALID_PROC_NUMBER)
1132 : : {
1133 [ # # # # ]: 0 : if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT))
1134 : : {
1135 [ # # # # : 0 : elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG,
# # # # #
# ]
1136 : : "could not notify checkpoint: checkpointer is not running");
1137 : 0 : break;
1138 : : }
1139 : 0 : }
1140 : : else
1141 : : {
1142 : 14 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1143 : : /* notified successfully */
1144 : 14 : break;
1145 : : }
1146 : :
1147 [ # # ]: 0 : CHECK_FOR_INTERRUPTS();
1148 : 0 : pg_usleep(100000L); /* wait 0.1 sec, then retry */
1149 [ - + ]: 14 : }
1150 : :
1151 : : /*
1152 : : * If requested, wait for completion. We detect completion according to
1153 : : * the algorithm given above.
1154 : : */
1155 [ - + ]: 14 : if (flags & CHECKPOINT_WAIT)
1156 : : {
1157 : 14 : int new_started,
1158 : : new_failed;
1159 : :
1160 : : /* Wait for a new checkpoint to start. */
1161 : 14 : ConditionVariablePrepareToSleep(&CheckpointerShmem->start_cv);
1162 : 28 : for (;;)
1163 : : {
1164 [ - + ]: 28 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1165 : 28 : new_started = CheckpointerShmem->ckpt_started;
1166 : 28 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1167 : :
1168 [ + + ]: 28 : if (new_started != old_started)
1169 : 14 : break;
1170 : :
1171 : 14 : ConditionVariableSleep(&CheckpointerShmem->start_cv,
1172 : : WAIT_EVENT_CHECKPOINT_START);
1173 : : }
1174 : 14 : ConditionVariableCancelSleep();
1175 : :
1176 : : /*
1177 : : * We are waiting for ckpt_done >= new_started, in a modulo sense.
1178 : : */
1179 : 14 : ConditionVariablePrepareToSleep(&CheckpointerShmem->done_cv);
1180 : 28 : for (;;)
1181 : : {
1182 : 28 : int new_done;
1183 : :
1184 [ - + ]: 28 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1185 : 28 : new_done = CheckpointerShmem->ckpt_done;
1186 : 28 : new_failed = CheckpointerShmem->ckpt_failed;
1187 : 28 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1188 : :
1189 [ + + ]: 28 : if (new_done - new_started >= 0)
1190 : 14 : break;
1191 : :
1192 : 14 : ConditionVariableSleep(&CheckpointerShmem->done_cv,
1193 : : WAIT_EVENT_CHECKPOINT_DONE);
1194 [ + + ]: 28 : }
1195 : 14 : ConditionVariableCancelSleep();
1196 : :
1197 [ + - ]: 14 : if (new_failed != old_failed)
1198 [ # # # # ]: 0 : ereport(ERROR,
1199 : : (errmsg("checkpoint request failed"),
1200 : : errhint("Consult recent messages in the server log for details.")));
1201 : 14 : }
1202 : 18 : }
1203 : :
1204 : : /*
1205 : : * ForwardSyncRequest
1206 : : * Forward a file-fsync request from a backend to the checkpointer
1207 : : *
1208 : : * Whenever a backend is compelled to write directly to a relation
1209 : : * (which should be seldom, if the background writer is getting its job done),
1210 : : * the backend calls this routine to pass over knowledge that the relation
1211 : : * is dirty and must be fsync'd before next checkpoint. We also use this
1212 : : * opportunity to count such writes for statistical purposes.
1213 : : *
1214 : : * To avoid holding the lock for longer than necessary, we normally write
1215 : : * to the requests[] queue without checking for duplicates. The checkpointer
1216 : : * will have to eliminate dups internally anyway. However, if we discover
1217 : : * that the queue is full, we make a pass over the entire queue to compact
1218 : : * it. This is somewhat expensive, but the alternative is for the backend
1219 : : * to perform its own fsync, which is far more expensive in practice. It
1220 : : * is theoretically possible a backend fsync might still be necessary, if
1221 : : * the queue is full and contains no duplicate entries. In that case, we
1222 : : * let the backend know by returning false.
1223 : : */
1224 : : bool
1225 : 71605 : ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)
1226 : : {
1227 : 71605 : CheckpointerRequest *request;
1228 : 71605 : bool too_full;
1229 : 71605 : int insert_pos;
1230 : :
1231 [ - + ]: 71605 : if (!IsUnderPostmaster)
1232 : 0 : return false; /* probably shouldn't even get here */
1233 : :
1234 [ + - ]: 71605 : if (AmCheckpointerProcess())
1235 [ # # # # ]: 0 : elog(ERROR, "ForwardSyncRequest must not be called in checkpointer");
1236 : :
1237 : 71605 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1238 : :
1239 : : /*
1240 : : * If the checkpointer isn't running or the request queue is full, the
1241 : : * backend will have to perform its own fsync request. But before forcing
1242 : : * that to happen, we can try to compact the request queue.
1243 : : */
1244 [ + - # # ]: 71605 : if (CheckpointerShmem->checkpointer_pid == 0 ||
1245 [ - + ]: 71605 : (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&
1246 : 0 : !CompactCheckpointerRequestQueue()))
1247 : : {
1248 : 0 : LWLockRelease(CheckpointerCommLock);
1249 : 0 : return false;
1250 : : }
1251 : :
1252 : : /* OK, insert request */
1253 : 71605 : insert_pos = CheckpointerShmem->tail;
1254 : 71605 : request = &CheckpointerShmem->requests[insert_pos];
1255 : 71605 : request->ftag = *ftag;
1256 : 71605 : request->type = type;
1257 : :
1258 : 71605 : CheckpointerShmem->tail = (CheckpointerShmem->tail + 1) % CheckpointerShmem->max_requests;
1259 : 71605 : CheckpointerShmem->num_requests++;
1260 : :
1261 : : /* If queue is more than half full, nudge the checkpointer to empty it */
1262 : 143210 : too_full = (CheckpointerShmem->num_requests >=
1263 : 71605 : CheckpointerShmem->max_requests / 2);
1264 : :
1265 : 71605 : LWLockRelease(CheckpointerCommLock);
1266 : :
1267 : : /* ... but not till after we release the lock */
1268 [ + + ]: 71605 : if (too_full)
1269 : : {
1270 : 30 : volatile PROC_HDR *procglobal = ProcGlobal;
1271 : 30 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1272 : :
1273 [ - + ]: 30 : if (checkpointerProc != INVALID_PROC_NUMBER)
1274 : 30 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1275 : 30 : }
1276 : :
1277 : 71605 : return true;
1278 : 71605 : }
1279 : :
1280 : : /*
1281 : : * CompactCheckpointerRequestQueue
1282 : : * Remove duplicates from the request queue to avoid backend fsyncs.
1283 : : * Returns "true" if any entries were removed.
1284 : : *
1285 : : * Although a full fsync request queue is not common, it can lead to severe
1286 : : * performance problems when it does happen. So far, this situation has
1287 : : * only been observed to occur when the system is under heavy write load,
1288 : : * and especially during the "sync" phase of a checkpoint. Without this
1289 : : * logic, each backend begins doing an fsync for every block written, which
1290 : : * gets very expensive and can slow down the whole system.
1291 : : *
1292 : : * Trying to do this every time the queue is full could lose if there
1293 : : * aren't any removable entries. But that should be vanishingly rare in
1294 : : * practice: there's one queue entry per shared buffer.
1295 : : */
1296 : : static bool
1297 : 0 : CompactCheckpointerRequestQueue(void)
1298 : : {
1299 : : struct CheckpointerSlotMapping
1300 : : {
1301 : : CheckpointerRequest request;
1302 : : int ring_idx;
1303 : : };
1304 : :
1305 : 0 : int n;
1306 : 0 : int num_skipped = 0;
1307 : 0 : int head;
1308 : 0 : int max_requests;
1309 : 0 : int num_requests;
1310 : 0 : int read_idx,
1311 : : write_idx;
1312 : 0 : HASHCTL ctl;
1313 : 0 : HTAB *htab;
1314 : 0 : bool *skip_slot;
1315 : :
1316 : : /* must hold CheckpointerCommLock in exclusive mode */
1317 [ # # ]: 0 : Assert(LWLockHeldByMe(CheckpointerCommLock));
1318 : :
1319 : : /* Avoid memory allocations in a critical section. */
1320 [ # # ]: 0 : if (CritSectionCount > 0)
1321 : 0 : return false;
1322 : :
1323 : 0 : max_requests = CheckpointerShmem->max_requests;
1324 : 0 : num_requests = CheckpointerShmem->num_requests;
1325 : :
1326 : : /* Initialize skip_slot array */
1327 : 0 : skip_slot = palloc0_array(bool, max_requests);
1328 : :
1329 : 0 : head = CheckpointerShmem->head;
1330 : :
1331 : : /* Initialize temporary hash table */
1332 : 0 : ctl.keysize = sizeof(CheckpointerRequest);
1333 : 0 : ctl.entrysize = sizeof(struct CheckpointerSlotMapping);
1334 : 0 : ctl.hcxt = CurrentMemoryContext;
1335 : :
1336 : 0 : htab = hash_create("CompactCheckpointerRequestQueue",
1337 : 0 : CheckpointerShmem->num_requests,
1338 : : &ctl,
1339 : : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1340 : :
1341 : : /*
1342 : : * The basic idea here is that a request can be skipped if it's followed
1343 : : * by a later, identical request. It might seem more sensible to work
1344 : : * backwards from the end of the queue and check whether a request is
1345 : : * *preceded* by an earlier, identical request, in the hopes of doing less
1346 : : * copying. But that might change the semantics, if there's an
1347 : : * intervening SYNC_FORGET_REQUEST or SYNC_FILTER_REQUEST, so we do it
1348 : : * this way. It would be possible to be even smarter if we made the code
1349 : : * below understand the specific semantics of such requests (it could blow
1350 : : * away preceding entries that would end up being canceled anyhow), but
1351 : : * it's not clear that the extra complexity would buy us anything.
1352 : : */
1353 : 0 : read_idx = head;
1354 [ # # ]: 0 : for (n = 0; n < num_requests; n++)
1355 : : {
1356 : 0 : CheckpointerRequest *request;
1357 : 0 : struct CheckpointerSlotMapping *slotmap;
1358 : 0 : bool found;
1359 : :
1360 : : /*
1361 : : * We use the request struct directly as a hashtable key. This
1362 : : * assumes that any padding bytes in the structs are consistently the
1363 : : * same, which should be okay because we zeroed them in
1364 : : * CheckpointerShmemInit. Note also that RelFileLocator had better
1365 : : * contain no pad bytes.
1366 : : */
1367 : 0 : request = &CheckpointerShmem->requests[read_idx];
1368 : 0 : slotmap = hash_search(htab, request, HASH_ENTER, &found);
1369 [ # # ]: 0 : if (found)
1370 : : {
1371 : : /* Duplicate, so mark the previous occurrence as skippable */
1372 : 0 : skip_slot[slotmap->ring_idx] = true;
1373 : 0 : num_skipped++;
1374 : 0 : }
1375 : : /* Remember slot containing latest occurrence of this request value */
1376 : 0 : slotmap->ring_idx = read_idx;
1377 : :
1378 : : /* Move to the next request in the ring buffer */
1379 : 0 : read_idx = (read_idx + 1) % max_requests;
1380 : 0 : }
1381 : :
1382 : : /* Done with the hash table. */
1383 : 0 : hash_destroy(htab);
1384 : :
1385 : : /* If no duplicates, we're out of luck. */
1386 [ # # ]: 0 : if (!num_skipped)
1387 : : {
1388 : 0 : pfree(skip_slot);
1389 : 0 : return false;
1390 : : }
1391 : :
1392 : : /* We found some duplicates; remove them. */
1393 : 0 : read_idx = write_idx = head;
1394 [ # # ]: 0 : for (n = 0; n < num_requests; n++)
1395 : : {
1396 : : /* If this slot is NOT skipped, keep it */
1397 [ # # ]: 0 : if (!skip_slot[read_idx])
1398 : : {
1399 : : /* If the read and write positions are different, copy the request */
1400 [ # # ]: 0 : if (write_idx != read_idx)
1401 : 0 : CheckpointerShmem->requests[write_idx] =
1402 : 0 : CheckpointerShmem->requests[read_idx];
1403 : :
1404 : : /* Advance the write position */
1405 : 0 : write_idx = (write_idx + 1) % max_requests;
1406 : 0 : }
1407 : :
1408 : 0 : read_idx = (read_idx + 1) % max_requests;
1409 : 0 : }
1410 : :
1411 : : /*
1412 : : * Update ring buffer state: head remains the same, tail moves, count
1413 : : * decreases
1414 : : */
1415 : 0 : CheckpointerShmem->tail = write_idx;
1416 : 0 : CheckpointerShmem->num_requests -= num_skipped;
1417 : :
1418 [ # # # # ]: 0 : ereport(DEBUG1,
1419 : : (errmsg_internal("compacted fsync request queue from %d entries to %d entries",
1420 : : num_requests, CheckpointerShmem->num_requests)));
1421 : :
1422 : : /* Cleanup. */
1423 : 0 : pfree(skip_slot);
1424 : 0 : return true;
1425 : 0 : }
1426 : :
1427 : : /*
1428 : : * AbsorbSyncRequests
1429 : : * Retrieve queued sync requests and pass them to sync mechanism.
1430 : : *
1431 : : * This is exported because it must be called during CreateCheckPoint;
1432 : : * we have to be sure we have accepted all pending requests just before
1433 : : * we start fsync'ing. Since CreateCheckPoint sometimes runs in
1434 : : * non-checkpointer processes, do nothing if not checkpointer.
1435 : : */
1436 : : void
1437 : 22 : AbsorbSyncRequests(void)
1438 : : {
1439 : 22 : CheckpointerRequest *requests = NULL;
1440 : 22 : CheckpointerRequest *request;
1441 : 22 : int n,
1442 : : i;
1443 : 22 : bool loop;
1444 : :
1445 [ + + ]: 22 : if (!AmCheckpointerProcess())
1446 : 12 : return;
1447 : :
1448 : 10 : do
1449 : : {
1450 : 10 : LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);
1451 : :
1452 : : /*---
1453 : : * We try to avoid holding the lock for a long time by:
1454 : : * 1. Copying the request array and processing the requests after
1455 : : * releasing the lock;
1456 : : * 2. Processing not the whole queue, but only batches of
1457 : : * CKPT_REQ_BATCH_SIZE at once.
1458 : : *
1459 : : * Once we have cleared the requests from shared memory, we must
1460 : : * PANIC if we then fail to absorb them (e.g., because our hashtable
1461 : : * runs out of memory). This is because the system cannot run safely
1462 : : * if we are unable to fsync what we have been told to fsync.
1463 : : * Fortunately, the hashtable is so small that the problem is quite
1464 : : * unlikely to arise in practice.
1465 : : *
1466 : : * Note: The maximum possible size of a ring buffer is
1467 : : * MAX_CHECKPOINT_REQUESTS entries, which fit into a maximum palloc
1468 : : * allocation size of 1Gb. Our maximum batch size,
1469 : : * CKPT_REQ_BATCH_SIZE, is even smaller.
1470 : : */
1471 [ + - ]: 10 : n = Min(CheckpointerShmem->num_requests, CKPT_REQ_BATCH_SIZE);
1472 [ + + ]: 10 : if (n > 0)
1473 : : {
1474 [ - + ]: 1 : if (!requests)
1475 : 1 : requests = (CheckpointerRequest *) palloc(n * sizeof(CheckpointerRequest));
1476 : :
1477 [ + + ]: 3626 : for (i = 0; i < n; i++)
1478 : : {
1479 : 3625 : requests[i] = CheckpointerShmem->requests[CheckpointerShmem->head];
1480 : 3625 : CheckpointerShmem->head = (CheckpointerShmem->head + 1) % CheckpointerShmem->max_requests;
1481 : 3625 : }
1482 : :
1483 : 1 : CheckpointerShmem->num_requests -= n;
1484 : :
1485 : 1 : }
1486 : :
1487 : 10 : START_CRIT_SECTION();
1488 : :
1489 : : /* Are there any requests in the queue? If so, keep going. */
1490 : 10 : loop = CheckpointerShmem->num_requests != 0;
1491 : :
1492 : 10 : LWLockRelease(CheckpointerCommLock);
1493 : :
1494 [ + + ]: 3635 : for (request = requests; n > 0; request++, n--)
1495 : 3625 : RememberSyncRequest(&request->ftag, request->type);
1496 : :
1497 [ + - ]: 10 : END_CRIT_SECTION();
1498 [ + - ]: 10 : } while (loop);
1499 : :
1500 [ + + ]: 10 : if (requests)
1501 : 1 : pfree(requests);
1502 [ - + ]: 22 : }
1503 : :
1504 : : /*
1505 : : * Update any shared memory configurations based on config parameters
1506 : : */
1507 : : static void
1508 : 1 : UpdateSharedMemoryConfig(void)
1509 : : {
1510 : : /* update global shmem state for sync rep */
1511 : 1 : SyncRepUpdateSyncStandbysDefined();
1512 : :
1513 : : /*
1514 : : * If full_page_writes has been changed by SIGHUP, we update it in shared
1515 : : * memory and write an XLOG_FPW_CHANGE record.
1516 : : */
1517 : 1 : UpdateFullPageWrites();
1518 : :
1519 [ - + - + ]: 1 : elog(DEBUG2, "checkpointer updated shared memory configuration values");
1520 : 1 : }
1521 : :
1522 : : /*
1523 : : * FirstCallSinceLastCheckpoint allows a process to take an action once
1524 : : * per checkpoint cycle by asynchronously checking for checkpoint completion.
1525 : : */
1526 : : bool
1527 : 24 : FirstCallSinceLastCheckpoint(void)
1528 : : {
1529 : : static int ckpt_done = 0;
1530 : 24 : int new_done;
1531 : 24 : bool FirstCall = false;
1532 : :
1533 [ - + ]: 24 : SpinLockAcquire(&CheckpointerShmem->ckpt_lck);
1534 : 24 : new_done = CheckpointerShmem->ckpt_done;
1535 : 24 : SpinLockRelease(&CheckpointerShmem->ckpt_lck);
1536 : :
1537 [ + - ]: 24 : if (new_done != ckpt_done)
1538 : 0 : FirstCall = true;
1539 : :
1540 : 24 : ckpt_done = new_done;
1541 : :
1542 : 48 : return FirstCall;
1543 : 24 : }
1544 : :
1545 : : /*
1546 : : * Wake up the checkpointer process.
1547 : : */
1548 : : void
1549 : 4 : WakeupCheckpointer(void)
1550 : : {
1551 : 4 : volatile PROC_HDR *procglobal = ProcGlobal;
1552 : 4 : ProcNumber checkpointerProc = procglobal->checkpointerProc;
1553 : :
1554 [ + + ]: 4 : if (checkpointerProc != INVALID_PROC_NUMBER)
1555 : 2 : SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch);
1556 : 4 : }
|
