1 /*-------------------------------------------------------------------------
5 * The background writer (bgwriter) is new as of Postgres 8.0. It attempts
6 * to keep regular backends from having to write out dirty shared buffers
7 * (which they would only do when needing to free a shared buffer to read in
8 * another page). In the best scenario all writes from shared buffers will
9 * be issued by the background writer process. However, regular backends are
10 * still empowered to issue writes if the bgwriter fails to maintain enough
11 * clean shared buffers.
13 * The bgwriter is also charged with handling all checkpoints. It will
14 * automatically dispatch a checkpoint after a certain amount of time has
15 * elapsed since the last one, and it can be signaled to perform requested
16 * checkpoints as well. (The GUC parameter that mandates a checkpoint every
17 * so many WAL segments is implemented by having backends signal the bgwriter
18 * when they fill WAL segments; the bgwriter itself doesn't watch for the
21 * The bgwriter is started by the postmaster as soon as the startup subprocess
22 * finishes. It remains alive until the postmaster commands it to terminate.
23 * Normal termination is by SIGUSR2, which instructs the bgwriter to execute
24 * a shutdown checkpoint and then exit(0). (All backends must be stopped
25 * before SIGUSR2 is issued!) Emergency termination is by SIGQUIT; like any
26 * backend, the bgwriter will simply abort and exit on SIGQUIT.
28 * If the bgwriter exits unexpectedly, the postmaster treats that the same
29 * as a backend crash: shared memory may be corrupted, so remaining backends
30 * should be killed by SIGQUIT and then a recovery cycle started. (Even if
31 * shared memory isn't corrupted, we have lost information about which
32 * files need to be fsync'd for the next checkpoint, and so a system
33 * restart needs to be forced.)
36 * Portions Copyright (c) 1996-2007, PostgreSQL Global Development Group
42 *-------------------------------------------------------------------------
51 #include "access/xlog_internal.h"
52 #include "libpq/pqsignal.h"
53 #include "miscadmin.h"
55 #include "postmaster/bgwriter.h"
56 #include "storage/fd.h"
57 #include "storage/freespace.h"
58 #include "storage/ipc.h"
59 #include "storage/lwlock.h"
60 #include "storage/pmsignal.h"
61 #include "storage/shmem.h"
62 #include "storage/smgr.h"
63 #include "storage/spin.h"
64 #include "tcop/tcopprot.h"
65 #include "utils/guc.h"
66 #include "utils/memutils.h"
67 #include "utils/resowner.h"
71 * Shared memory area for communication between bgwriter and backends
73 * The ckpt counters allow backends to watch for completion of a checkpoint
74 * request they send. Here's how it works:
75 * * At start of a checkpoint, bgwriter reads (and clears) the request flags
76 * and increments ckpt_started, while holding ckpt_lck.
77 * * On completion of a checkpoint, bgwriter sets ckpt_done to
79 * * On failure of a checkpoint, bgwriter increments ckpt_failed
80 * and sets ckpt_done to equal ckpt_started.
82 * The algorithm for backends is:
83 * 1. Record current values of ckpt_failed and ckpt_started, and
84 * set request flags, while holding ckpt_lck.
85 * 2. Send signal to request checkpoint.
86 * 3. Sleep until ckpt_started changes. Now you know a checkpoint has
87 * begun since you started this algorithm (although *not* that it was
88 * specifically initiated by your signal), and that it is using your flags.
89 * 4. Record new value of ckpt_started.
90 * 5. Sleep until ckpt_done >= saved value of ckpt_started. (Use modulo
91 * arithmetic here in case counters wrap around.) Now you know a
92 * checkpoint has started and completed, but not whether it was
94 * 6. If ckpt_failed is different from the originally saved value,
95 * assume request failed; otherwise it was definitely successful.
97 * ckpt_flags holds the OR of the checkpoint request flags sent by all
98 * requesting backends since the last checkpoint start. The flags are
99 * chosen so that OR'ing is the correct way to combine multiple requests.
101 * num_backend_writes is used to count the number of buffer writes performed
102 * by non-bgwriter processes. This counter should be wide enough that it
103 * can't overflow during a single bgwriter cycle.
105 * The requests array holds fsync requests sent by backends and not yet
106 * absorbed by the bgwriter.
108 * Unlike the checkpoint fields, num_backend_writes and the requests
109 * fields are protected by BgWriterCommLock.
115 BlockNumber segno
; /* see md.c for special values */
116 /* might add a real request-type field later; not needed yet */
121 pid_t bgwriter_pid
; /* PID of bgwriter (0 if not started) */
123 slock_t ckpt_lck
; /* protects all the ckpt_* fields */
125 int ckpt_started
; /* advances when checkpoint starts */
126 int ckpt_done
; /* advances when checkpoint done */
127 int ckpt_failed
; /* advances when checkpoint fails */
129 int ckpt_flags
; /* checkpoint flags, as defined in xlog.h */
131 uint32 num_backend_writes
; /* counts non-bgwriter buffer writes */
133 int num_requests
; /* current # of requests */
134 int max_requests
; /* allocated array size */
135 BgWriterRequest requests
[1]; /* VARIABLE LENGTH ARRAY */
136 } BgWriterShmemStruct
;
138 static BgWriterShmemStruct
*BgWriterShmem
;
140 /* interval for calling AbsorbFsyncRequests in CheckpointWriteDelay */
141 #define WRITES_PER_ABSORB 1000
146 int BgWriterDelay
= 200;
147 int CheckPointTimeout
= 300;
148 int CheckPointWarning
= 30;
149 double CheckPointCompletionTarget
= 0.5;
152 * Flags set by interrupt handlers for later service in the main loop.
154 static volatile sig_atomic_t got_SIGHUP
= false;
155 static volatile sig_atomic_t checkpoint_requested
= false;
156 static volatile sig_atomic_t shutdown_requested
= false;
161 static bool am_bg_writer
= false;
163 static bool ckpt_active
= false;
165 /* these values are valid when ckpt_active is true: */
166 static time_t ckpt_start_time
;
167 static XLogRecPtr ckpt_start_recptr
;
168 static double ckpt_cached_elapsed
;
170 static time_t last_checkpoint_time
;
171 static time_t last_xlog_switch_time
;
173 /* Prototypes for private functions */
175 static void CheckArchiveTimeout(void);
176 static void BgWriterNap(void);
177 static bool IsCheckpointOnSchedule(double progress
);
178 static bool ImmediateCheckpointRequested(void);
180 /* Signal handlers */
182 static void bg_quickdie(SIGNAL_ARGS
);
183 static void BgSigHupHandler(SIGNAL_ARGS
);
184 static void ReqCheckpointHandler(SIGNAL_ARGS
);
185 static void ReqShutdownHandler(SIGNAL_ARGS
);
189 * Main entry point for bgwriter process
191 * This is invoked from BootstrapMain, which has already created the basic
192 * execution environment, but not enabled signals yet.
195 BackgroundWriterMain(void)
197 sigjmp_buf local_sigjmp_buf
;
198 MemoryContext bgwriter_context
;
200 BgWriterShmem
->bgwriter_pid
= MyProcPid
;
204 * If possible, make this process a group leader, so that the postmaster
205 * can signal any child processes too. (bgwriter probably never has
206 * any child processes, but for consistency we make all postmaster
207 * child processes do this.)
211 elog(FATAL
, "setsid() failed: %m");
215 * Properly accept or ignore signals the postmaster might send us
217 * Note: we deliberately ignore SIGTERM, because during a standard Unix
218 * system shutdown cycle, init will SIGTERM all processes at once. We
219 * want to wait for the backends to exit, whereupon the postmaster will
220 * tell us it's okay to shut down (via SIGUSR2).
222 * SIGUSR1 is presently unused; keep it spare in case someday we want this
223 * process to participate in sinval messaging.
225 pqsignal(SIGHUP
, BgSigHupHandler
); /* set flag to read config file */
226 pqsignal(SIGINT
, ReqCheckpointHandler
); /* request checkpoint */
227 pqsignal(SIGTERM
, SIG_IGN
); /* ignore SIGTERM */
228 pqsignal(SIGQUIT
, bg_quickdie
); /* hard crash time */
229 pqsignal(SIGALRM
, SIG_IGN
);
230 pqsignal(SIGPIPE
, SIG_IGN
);
231 pqsignal(SIGUSR1
, SIG_IGN
); /* reserve for sinval */
232 pqsignal(SIGUSR2
, ReqShutdownHandler
); /* request shutdown */
235 * Reset some signals that are accepted by postmaster but not here
237 pqsignal(SIGCHLD
, SIG_DFL
);
238 pqsignal(SIGTTIN
, SIG_DFL
);
239 pqsignal(SIGTTOU
, SIG_DFL
);
240 pqsignal(SIGCONT
, SIG_DFL
);
241 pqsignal(SIGWINCH
, SIG_DFL
);
243 /* We allow SIGQUIT (quickdie) at all times */
244 #ifdef HAVE_SIGPROCMASK
245 sigdelset(&BlockSig
, SIGQUIT
);
247 BlockSig
&= ~(sigmask(SIGQUIT
));
251 * Initialize so that first time-driven event happens at the correct time.
253 last_checkpoint_time
= last_xlog_switch_time
= time(NULL
);
256 * Create a resource owner to keep track of our resources (currently only
259 CurrentResourceOwner
= ResourceOwnerCreate(NULL
, "Background Writer");
262 * Create a memory context that we will do all our work in. We do this so
263 * that we can reset the context during error recovery and thereby avoid
264 * possible memory leaks. Formerly this code just ran in
265 * TopMemoryContext, but resetting that would be a really bad idea.
267 bgwriter_context
= AllocSetContextCreate(TopMemoryContext
,
269 ALLOCSET_DEFAULT_MINSIZE
,
270 ALLOCSET_DEFAULT_INITSIZE
,
271 ALLOCSET_DEFAULT_MAXSIZE
);
272 MemoryContextSwitchTo(bgwriter_context
);
275 * If an exception is encountered, processing resumes here.
277 * See notes in postgres.c about the design of this coding.
279 if (sigsetjmp(local_sigjmp_buf
, 1) != 0)
281 /* Since not using PG_TRY, must reset error stack by hand */
282 error_context_stack
= NULL
;
284 /* Prevent interrupts while cleaning up */
287 /* Report the error to the server log */
291 * These operations are really just a minimal subset of
292 * AbortTransaction(). We don't have very many resources to worry
293 * about in bgwriter, but we do have LWLocks, buffers, and temp files.
298 /* buffer pins are released here: */
299 ResourceOwnerRelease(CurrentResourceOwner
,
300 RESOURCE_RELEASE_BEFORE_LOCKS
,
302 /* we needn't bother with the other ResourceOwnerRelease phases */
303 AtEOXact_Buffers(false);
305 AtEOXact_HashTables(false);
307 /* Warn any waiting backends that the checkpoint failed. */
310 /* use volatile pointer to prevent code rearrangement */
311 volatile BgWriterShmemStruct
*bgs
= BgWriterShmem
;
313 SpinLockAcquire(&bgs
->ckpt_lck
);
315 bgs
->ckpt_done
= bgs
->ckpt_started
;
316 SpinLockRelease(&bgs
->ckpt_lck
);
322 * Now return to normal top-level context and clear ErrorContext for
325 MemoryContextSwitchTo(bgwriter_context
);
328 /* Flush any leaked data in the top-level context */
329 MemoryContextResetAndDeleteChildren(bgwriter_context
);
331 /* Now we can allow interrupts again */
335 * Sleep at least 1 second after any error. A write error is likely
336 * to be repeated, and we don't want to be filling the error logs as
342 * Close all open files after any error. This is helpful on Windows,
343 * where holding deleted files open causes various strange errors.
344 * It's not clear we need it elsewhere, but shouldn't hurt.
349 /* We can now handle ereport(ERROR) */
350 PG_exception_stack
= &local_sigjmp_buf
;
353 * Unblock signals (they were blocked when the postmaster forked us)
355 PG_SETMASK(&UnBlockSig
);
362 bool do_checkpoint
= false;
368 * Emergency bailout if postmaster has died. This is to avoid the
369 * necessity for manual cleanup of all postmaster children.
371 if (!PostmasterIsAlive(true))
375 * Process any requests or signals received recently.
377 AbsorbFsyncRequests();
382 ProcessConfigFile(PGC_SIGHUP
);
384 if (checkpoint_requested
)
386 checkpoint_requested
= false;
387 do_checkpoint
= true;
388 BgWriterStats
.m_requested_checkpoints
++;
390 if (shutdown_requested
)
393 * From here on, elog(ERROR) should end with exit(1), not send
394 * control back to the sigsetjmp block above
396 ExitOnAnyError
= true;
397 /* Close down the database */
399 DumpFreeSpaceMap(0, 0);
400 /* Normal exit from the bgwriter is here */
401 proc_exit(0); /* done */
405 * Force a checkpoint if too much time has elapsed since the
406 * last one. Note that we count a timed checkpoint in stats only
407 * when this occurs without an external request, but we set the
408 * CAUSE_TIME flag bit even if there is also an external request.
411 elapsed_secs
= now
- last_checkpoint_time
;
412 if (elapsed_secs
>= CheckPointTimeout
)
415 BgWriterStats
.m_timed_checkpoints
++;
416 do_checkpoint
= true;
417 flags
|= CHECKPOINT_CAUSE_TIME
;
421 * Do a checkpoint if requested, otherwise do one cycle of
422 * dirty-buffer writing.
426 /* use volatile pointer to prevent code rearrangement */
427 volatile BgWriterShmemStruct
*bgs
= BgWriterShmem
;
430 * Atomically fetch the request flags to figure out what
431 * kind of a checkpoint we should perform, and increase the
432 * started-counter to acknowledge that we've started
435 SpinLockAcquire(&bgs
->ckpt_lck
);
436 flags
|= bgs
->ckpt_flags
;
439 SpinLockRelease(&bgs
->ckpt_lck
);
442 * We will warn if (a) too soon since last checkpoint (whatever
443 * caused it) and (b) somebody set the CHECKPOINT_CAUSE_XLOG flag
444 * since the last checkpoint start. Note in particular that this
445 * implementation will not generate warnings caused by
446 * CheckPointTimeout < CheckPointWarning.
448 if ((flags
& CHECKPOINT_CAUSE_XLOG
) &&
449 elapsed_secs
< CheckPointWarning
)
451 (errmsg("checkpoints are occurring too frequently (%d seconds apart)",
453 errhint("Consider increasing the configuration parameter \"checkpoint_segments\".")));
456 * Initialize bgwriter-private variables used during checkpoint.
459 ckpt_start_recptr
= GetInsertRecPtr();
460 ckpt_start_time
= now
;
461 ckpt_cached_elapsed
= 0;
466 CreateCheckPoint(flags
);
469 * After any checkpoint, close all smgr files. This is so we
470 * won't hang onto smgr references to deleted files indefinitely.
475 * Indicate checkpoint completion to any waiting backends.
477 SpinLockAcquire(&bgs
->ckpt_lck
);
478 bgs
->ckpt_done
= bgs
->ckpt_started
;
479 SpinLockRelease(&bgs
->ckpt_lck
);
484 * Note we record the checkpoint start time not end time as
485 * last_checkpoint_time. This is so that time-driven checkpoints
486 * happen at a predictable spacing.
488 last_checkpoint_time
= now
;
493 /* Check for archive_timeout and switch xlog files if necessary. */
494 CheckArchiveTimeout();
496 /* Nap for the configured time. */
502 * CheckArchiveTimeout -- check for archive_timeout and switch xlog files
506 CheckArchiveTimeout(void)
511 if (XLogArchiveTimeout
<= 0)
516 /* First we do a quick check using possibly-stale local state. */
517 if ((int) (now
- last_xlog_switch_time
) < XLogArchiveTimeout
)
521 * Update local state ... note that last_xlog_switch_time is the
522 * last time a switch was performed *or requested*.
524 last_time
= GetLastSegSwitchTime();
526 last_xlog_switch_time
= Max(last_xlog_switch_time
, last_time
);
528 /* Now we can do the real check */
529 if ((int) (now
- last_xlog_switch_time
) >= XLogArchiveTimeout
)
531 XLogRecPtr switchpoint
;
533 /* OK, it's time to switch */
534 switchpoint
= RequestXLogSwitch();
537 * If the returned pointer points exactly to a segment
538 * boundary, assume nothing happened.
540 if ((switchpoint
.xrecoff
% XLogSegSize
) != 0)
542 (errmsg("transaction log switch forced (archive_timeout=%d)",
543 XLogArchiveTimeout
)));
546 * Update state in any case, so we don't retry constantly when
547 * the system is idle.
549 last_xlog_switch_time
= now
;
554 * BgWriterNap -- Nap for the configured time or until a signal is received.
562 * Send off activity statistics to the stats collector
564 pgstat_send_bgwriter();
567 * Nap for the configured time, or sleep for 10 seconds if there is no
568 * bgwriter activity configured.
570 * On some platforms, signals won't interrupt the sleep. To ensure we
571 * respond reasonably promptly when someone signals us, break down the
572 * sleep into 1-second increments, and check for interrupts after each
575 * We absorb pending requests after each short sleep.
577 if (bgwriter_lru_maxpages
> 0 || ckpt_active
)
578 udelay
= BgWriterDelay
* 1000L;
579 else if (XLogArchiveTimeout
> 0)
580 udelay
= 1000000L; /* One second */
582 udelay
= 10000000L; /* Ten seconds */
584 while (udelay
> 999999L)
586 if (got_SIGHUP
|| shutdown_requested
||
587 (ckpt_active
? ImmediateCheckpointRequested() : checkpoint_requested
))
590 AbsorbFsyncRequests();
594 if (!(got_SIGHUP
|| shutdown_requested
||
595 (ckpt_active
? ImmediateCheckpointRequested() : checkpoint_requested
)))
600 * Returns true if an immediate checkpoint request is pending. (Note that
601 * this does not check the *current* checkpoint's IMMEDIATE flag, but whether
602 * there is one pending behind it.)
605 ImmediateCheckpointRequested(void)
607 if (checkpoint_requested
)
609 volatile BgWriterShmemStruct
*bgs
= BgWriterShmem
;
612 * We don't need to acquire the ckpt_lck in this case because we're
613 * only looking at a single flag bit.
615 if (bgs
->ckpt_flags
& CHECKPOINT_IMMEDIATE
)
622 * CheckpointWriteDelay -- yield control to bgwriter during a checkpoint
624 * This function is called after each page write performed by BufferSync().
625 * It is responsible for keeping the bgwriter's normal activities in
626 * progress during a long checkpoint, and for throttling BufferSync()'s
627 * write rate to hit checkpoint_completion_target.
629 * The checkpoint request flags should be passed in; currently the only one
630 * examined is CHECKPOINT_IMMEDIATE, which disables delays between writes.
632 * 'progress' is an estimate of how much of the work has been done, as a
633 * fraction between 0.0 meaning none, and 1.0 meaning all done.
636 CheckpointWriteDelay(int flags
, double progress
)
638 static int absorb_counter
= WRITES_PER_ABSORB
;
640 /* Do nothing if checkpoint is being executed by non-bgwriter process */
645 * Perform the usual bgwriter duties and take a nap, unless we're behind
646 * schedule, in which case we just try to catch up as quickly as possible.
648 if (!(flags
& CHECKPOINT_IMMEDIATE
) &&
649 !shutdown_requested
&&
650 !ImmediateCheckpointRequested() &&
651 IsCheckpointOnSchedule(progress
))
656 ProcessConfigFile(PGC_SIGHUP
);
659 AbsorbFsyncRequests();
660 absorb_counter
= WRITES_PER_ABSORB
;
663 CheckArchiveTimeout();
666 else if (--absorb_counter
<= 0)
669 * Absorb pending fsync requests after each WRITES_PER_ABSORB write
670 * operations even when we don't sleep, to prevent overflow of the
671 * fsync request queue.
673 AbsorbFsyncRequests();
674 absorb_counter
= WRITES_PER_ABSORB
;
679 * IsCheckpointOnSchedule -- are we on schedule to finish this checkpoint
682 * Compares the current progress against the time/segments elapsed since last
683 * checkpoint, and returns true if the progress we've made this far is greater
684 * than the elapsed time/segments.
687 IsCheckpointOnSchedule(double progress
)
691 double elapsed_xlogs
,
696 /* Scale progress according to checkpoint_completion_target. */
697 progress
*= CheckPointCompletionTarget
;
700 * Check against the cached value first. Only do the more expensive
701 * calculations once we reach the target previously calculated. Since
702 * neither time or WAL insert pointer moves backwards, a freshly
703 * calculated value can only be greater than or equal to the cached value.
705 if (progress
< ckpt_cached_elapsed
)
709 * Check progress against WAL segments written and checkpoint_segments.
711 * We compare the current WAL insert location against the location
712 * computed before calling CreateCheckPoint. The code in XLogInsert that
713 * actually triggers a checkpoint when checkpoint_segments is exceeded
714 * compares against RedoRecptr, so this is not completely accurate.
715 * However, it's good enough for our purposes, we're only calculating
716 * an estimate anyway.
718 recptr
= GetInsertRecPtr();
720 (((double) (int32
) (recptr
.xlogid
- ckpt_start_recptr
.xlogid
)) * XLogSegsPerFile
+
721 ((double) (int32
) (recptr
.xrecoff
- ckpt_start_recptr
.xrecoff
)) / XLogSegSize
) /
724 if (progress
< elapsed_xlogs
)
726 ckpt_cached_elapsed
= elapsed_xlogs
;
731 * Check progress against time elapsed and checkpoint_timeout.
733 gettimeofday(&now
, NULL
);
734 elapsed_time
= ((double) (now
.tv_sec
- ckpt_start_time
) +
735 now
.tv_usec
/ 1000000.0) / CheckPointTimeout
;
737 if (progress
< elapsed_time
)
739 ckpt_cached_elapsed
= elapsed_time
;
743 /* It looks like we're on schedule. */
748 /* --------------------------------
749 * signal handler routines
750 * --------------------------------
754 * bg_quickdie() occurs when signalled SIGQUIT by the postmaster.
756 * Some backend has bought the farm,
757 * so we need to stop what we're doing and exit.
760 bg_quickdie(SIGNAL_ARGS
)
762 PG_SETMASK(&BlockSig
);
765 * DO NOT proc_exit() -- we're here because shared memory may be
766 * corrupted, so we don't want to try to clean up our transaction. Just
767 * nail the windows shut and get out of town.
769 * Note we do exit(2) not exit(0). This is to force the postmaster into a
770 * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
771 * backend. This is necessary precisely because we don't clean up our
772 * shared memory state.
777 /* SIGHUP: set flag to re-read config file at next convenient time */
779 BgSigHupHandler(SIGNAL_ARGS
)
784 /* SIGINT: set flag to run a normal checkpoint right away */
786 ReqCheckpointHandler(SIGNAL_ARGS
)
788 checkpoint_requested
= true;
791 /* SIGUSR2: set flag to run a shutdown checkpoint and exit */
793 ReqShutdownHandler(SIGNAL_ARGS
)
795 shutdown_requested
= true;
799 /* --------------------------------
800 * communication with backends
801 * --------------------------------
806 * Compute space needed for bgwriter-related shared memory
809 BgWriterShmemSize(void)
814 * Currently, the size of the requests[] array is arbitrarily set equal to
815 * NBuffers. This may prove too large or small ...
817 size
= offsetof(BgWriterShmemStruct
, requests
);
818 size
= add_size(size
, mul_size(NBuffers
, sizeof(BgWriterRequest
)));
825 * Allocate and initialize bgwriter-related shared memory
828 BgWriterShmemInit(void)
832 BgWriterShmem
= (BgWriterShmemStruct
*)
833 ShmemInitStruct("Background Writer Data",
836 if (BgWriterShmem
== NULL
)
838 (errcode(ERRCODE_OUT_OF_MEMORY
),
839 errmsg("not enough shared memory for background writer")));
841 return; /* already initialized */
843 MemSet(BgWriterShmem
, 0, sizeof(BgWriterShmemStruct
));
844 SpinLockInit(&BgWriterShmem
->ckpt_lck
);
845 BgWriterShmem
->max_requests
= NBuffers
;
850 * Called in backend processes to request a checkpoint
852 * flags is a bitwise OR of the following:
853 * CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
854 * CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP,
855 * ignoring checkpoint_completion_target parameter.
856 * CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occured
857 * since the last one (implied by CHECKPOINT_IS_SHUTDOWN).
858 * CHECKPOINT_WAIT: wait for completion before returning (otherwise,
859 * just signal bgwriter to do it, and return).
860 * CHECKPOINT_CAUSE_XLOG: checkpoint is requested due to xlog filling.
861 * (This affects logging, and in particular enables CheckPointWarning.)
864 RequestCheckpoint(int flags
)
866 /* use volatile pointer to prevent code rearrangement */
867 volatile BgWriterShmemStruct
*bgs
= BgWriterShmem
;
868 int old_failed
, old_started
;
871 * If in a standalone backend, just do it ourselves.
873 if (!IsPostmasterEnvironment
)
876 * There's no point in doing slow checkpoints in a standalone
877 * backend, because there's no other backends the checkpoint could
880 CreateCheckPoint(flags
| CHECKPOINT_IMMEDIATE
);
883 * After any checkpoint, close all smgr files. This is so we won't
884 * hang onto smgr references to deleted files indefinitely.
892 * Atomically set the request flags, and take a snapshot of the counters.
893 * When we see ckpt_started > old_started, we know the flags we set here
894 * have been seen by bgwriter.
896 * Note that we OR the flags with any existing flags, to avoid overriding
897 * a "stronger" request by another backend. The flag senses must be
898 * chosen to make this work!
900 SpinLockAcquire(&bgs
->ckpt_lck
);
902 old_failed
= bgs
->ckpt_failed
;
903 old_started
= bgs
->ckpt_started
;
904 bgs
->ckpt_flags
|= flags
;
906 SpinLockRelease(&bgs
->ckpt_lck
);
909 * Send signal to request checkpoint. When not waiting, we
910 * consider failure to send the signal to be nonfatal.
912 if (BgWriterShmem
->bgwriter_pid
== 0)
913 elog((flags
& CHECKPOINT_WAIT
) ? ERROR
: LOG
,
914 "could not request checkpoint because bgwriter not running");
915 if (kill(BgWriterShmem
->bgwriter_pid
, SIGINT
) != 0)
916 elog((flags
& CHECKPOINT_WAIT
) ? ERROR
: LOG
,
917 "could not signal for checkpoint: %m");
920 * If requested, wait for completion. We detect completion according to
921 * the algorithm given above.
923 if (flags
& CHECKPOINT_WAIT
)
925 int new_started
, new_failed
;
927 /* Wait for a new checkpoint to start. */
930 SpinLockAcquire(&bgs
->ckpt_lck
);
931 new_started
= bgs
->ckpt_started
;
932 SpinLockRelease(&bgs
->ckpt_lck
);
934 if (new_started
!= old_started
)
937 CHECK_FOR_INTERRUPTS();
942 * We are waiting for ckpt_done >= new_started, in a modulo sense.
948 SpinLockAcquire(&bgs
->ckpt_lck
);
949 new_done
= bgs
->ckpt_done
;
950 new_failed
= bgs
->ckpt_failed
;
951 SpinLockRelease(&bgs
->ckpt_lck
);
953 if (new_done
- new_started
>= 0)
956 CHECK_FOR_INTERRUPTS();
960 if (new_failed
!= old_failed
)
962 (errmsg("checkpoint request failed"),
963 errhint("Consult recent messages in the server log for details.")));
968 * ForwardFsyncRequest
969 * Forward a file-fsync request from a backend to the bgwriter
971 * Whenever a backend is compelled to write directly to a relation
972 * (which should be seldom, if the bgwriter is getting its job done),
973 * the backend calls this routine to pass over knowledge that the relation
974 * is dirty and must be fsync'd before next checkpoint. We also use this
975 * opportunity to count such writes for statistical purposes.
977 * segno specifies which segment (not block!) of the relation needs to be
978 * fsync'd. (Since the valid range is much less than BlockNumber, we can
979 * use high values for special flags; that's all internal to md.c, which
982 * If we are unable to pass over the request (at present, this can happen
983 * if the shared memory queue is full), we return false. That forces
984 * the backend to do its own fsync. We hope that will be even more seldom.
986 * Note: we presently make no attempt to eliminate duplicate requests
987 * in the requests[] queue. The bgwriter will have to eliminate dups
988 * internally anyway, so we may as well avoid holding the lock longer
989 * than we have to here.
992 ForwardFsyncRequest(RelFileNode rnode
, BlockNumber segno
)
994 BgWriterRequest
*request
;
996 if (!IsUnderPostmaster
)
997 return false; /* probably shouldn't even get here */
999 Assert(!am_bg_writer
);
1001 LWLockAcquire(BgWriterCommLock
, LW_EXCLUSIVE
);
1003 /* we count non-bgwriter writes even when the request queue overflows */
1004 BgWriterShmem
->num_backend_writes
++;
1006 if (BgWriterShmem
->bgwriter_pid
== 0 ||
1007 BgWriterShmem
->num_requests
>= BgWriterShmem
->max_requests
)
1009 LWLockRelease(BgWriterCommLock
);
1012 request
= &BgWriterShmem
->requests
[BgWriterShmem
->num_requests
++];
1013 request
->rnode
= rnode
;
1014 request
->segno
= segno
;
1015 LWLockRelease(BgWriterCommLock
);
1020 * AbsorbFsyncRequests
1021 * Retrieve queued fsync requests and pass them to local smgr.
1023 * This is exported because it must be called during CreateCheckPoint;
1024 * we have to be sure we have accepted all pending requests just before
1025 * we start fsync'ing. Since CreateCheckPoint sometimes runs in
1026 * non-bgwriter processes, do nothing if not bgwriter.
1029 AbsorbFsyncRequests(void)
1031 BgWriterRequest
*requests
= NULL
;
1032 BgWriterRequest
*request
;
1039 * We have to PANIC if we fail to absorb all the pending requests (eg,
1040 * because our hashtable runs out of memory). This is because the system
1041 * cannot run safely if we are unable to fsync what we have been told to
1042 * fsync. Fortunately, the hashtable is so small that the problem is
1043 * quite unlikely to arise in practice.
1045 START_CRIT_SECTION();
1048 * We try to avoid holding the lock for a long time by copying the request
1051 LWLockAcquire(BgWriterCommLock
, LW_EXCLUSIVE
);
1053 /* Transfer write count into pending pgstats message */
1054 BgWriterStats
.m_buf_written_backend
+= BgWriterShmem
->num_backend_writes
;
1055 BgWriterShmem
->num_backend_writes
= 0;
1057 n
= BgWriterShmem
->num_requests
;
1060 requests
= (BgWriterRequest
*) palloc(n
* sizeof(BgWriterRequest
));
1061 memcpy(requests
, BgWriterShmem
->requests
, n
* sizeof(BgWriterRequest
));
1063 BgWriterShmem
->num_requests
= 0;
1065 LWLockRelease(BgWriterCommLock
);
1067 for (request
= requests
; n
> 0; request
++, n
--)
1068 RememberFsyncRequest(request
->rnode
, request
->segno
);