1 | /*------------------------------------------------------------------------- |
2 | * |
3 | * bgwriter.c |
4 | * |
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. |
12 | * |
13 | * As of Postgres 9.2 the bgwriter no longer handles checkpoints. |
14 | * |
15 | * The bgwriter is started by the postmaster as soon as the startup subprocess |
16 | * finishes, or as soon as recovery begins if we are doing archive recovery. |
17 | * It remains alive until the postmaster commands it to terminate. |
18 | * Normal termination is by SIGTERM, which instructs the bgwriter to exit(0). |
19 | * Emergency termination is by SIGQUIT; like any backend, the bgwriter will |
20 | * simply abort and exit on SIGQUIT. |
21 | * |
22 | * If the bgwriter exits unexpectedly, the postmaster treats that the same |
23 | * as a backend crash: shared memory may be corrupted, so remaining backends |
24 | * should be killed by SIGQUIT and then a recovery cycle started. |
25 | * |
26 | * |
27 | * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group |
28 | * |
29 | * |
30 | * IDENTIFICATION |
31 | * src/backend/postmaster/bgwriter.c |
32 | * |
33 | *------------------------------------------------------------------------- |
34 | */ |
35 | #include "postgres.h" |
36 | |
37 | #include <signal.h> |
38 | #include <sys/time.h> |
39 | #include <unistd.h> |
40 | |
41 | #include "access/xlog.h" |
42 | #include "access/xlog_internal.h" |
43 | #include "libpq/pqsignal.h" |
44 | #include "miscadmin.h" |
45 | #include "pgstat.h" |
46 | #include "postmaster/bgwriter.h" |
47 | #include "storage/bufmgr.h" |
48 | #include "storage/buf_internals.h" |
49 | #include "storage/condition_variable.h" |
50 | #include "storage/fd.h" |
51 | #include "storage/ipc.h" |
52 | #include "storage/lwlock.h" |
53 | #include "storage/proc.h" |
54 | #include "storage/shmem.h" |
55 | #include "storage/smgr.h" |
56 | #include "storage/spin.h" |
57 | #include "storage/standby.h" |
58 | #include "utils/guc.h" |
59 | #include "utils/memutils.h" |
60 | #include "utils/resowner.h" |
61 | #include "utils/timestamp.h" |
62 | |
63 | |
64 | /* |
65 | * GUC parameters |
66 | */ |
67 | int BgWriterDelay = 200; |
68 | |
69 | /* |
70 | * Multiplier to apply to BgWriterDelay when we decide to hibernate. |
71 | * (Perhaps this needs to be configurable?) |
72 | */ |
73 | #define HIBERNATE_FACTOR 50 |
74 | |
75 | /* |
76 | * Interval in which standby snapshots are logged into the WAL stream, in |
77 | * milliseconds. |
78 | */ |
79 | #define LOG_SNAPSHOT_INTERVAL_MS 15000 |
80 | |
81 | /* |
82 | * LSN and timestamp at which we last issued a LogStandbySnapshot(), to avoid |
83 | * doing so too often or repeatedly if there has been no other write activity |
84 | * in the system. |
85 | */ |
86 | static TimestampTz last_snapshot_ts; |
87 | static XLogRecPtr last_snapshot_lsn = InvalidXLogRecPtr; |
88 | |
89 | /* |
90 | * Flags set by interrupt handlers for later service in the main loop. |
91 | */ |
92 | static volatile sig_atomic_t got_SIGHUP = false; |
93 | static volatile sig_atomic_t shutdown_requested = false; |
94 | |
95 | /* Signal handlers */ |
96 | |
97 | static void bg_quickdie(SIGNAL_ARGS); |
98 | static void BgSigHupHandler(SIGNAL_ARGS); |
99 | static void ReqShutdownHandler(SIGNAL_ARGS); |
100 | static void bgwriter_sigusr1_handler(SIGNAL_ARGS); |
101 | |
102 | |
103 | /* |
104 | * Main entry point for bgwriter process |
105 | * |
106 | * This is invoked from AuxiliaryProcessMain, which has already created the |
107 | * basic execution environment, but not enabled signals yet. |
108 | */ |
109 | void |
110 | BackgroundWriterMain(void) |
111 | { |
112 | sigjmp_buf local_sigjmp_buf; |
113 | MemoryContext bgwriter_context; |
114 | bool prev_hibernate; |
115 | WritebackContext wb_context; |
116 | |
117 | /* |
118 | * Properly accept or ignore signals the postmaster might send us. |
119 | * |
120 | * bgwriter doesn't participate in ProcSignal signalling, but a SIGUSR1 |
121 | * handler is still needed for latch wakeups. |
122 | */ |
123 | pqsignal(SIGHUP, BgSigHupHandler); /* set flag to read config file */ |
124 | pqsignal(SIGINT, SIG_IGN); |
125 | pqsignal(SIGTERM, ReqShutdownHandler); /* shutdown */ |
126 | pqsignal(SIGQUIT, bg_quickdie); /* hard crash time */ |
127 | pqsignal(SIGALRM, SIG_IGN); |
128 | pqsignal(SIGPIPE, SIG_IGN); |
129 | pqsignal(SIGUSR1, bgwriter_sigusr1_handler); |
130 | pqsignal(SIGUSR2, SIG_IGN); |
131 | |
132 | /* |
133 | * Reset some signals that are accepted by postmaster but not here |
134 | */ |
135 | pqsignal(SIGCHLD, SIG_DFL); |
136 | |
137 | /* We allow SIGQUIT (quickdie) at all times */ |
138 | sigdelset(&BlockSig, SIGQUIT); |
139 | |
140 | /* |
141 | * We just started, assume there has been either a shutdown or |
142 | * end-of-recovery snapshot. |
143 | */ |
144 | last_snapshot_ts = GetCurrentTimestamp(); |
145 | |
146 | /* |
147 | * Create a memory context that we will do all our work in. We do this so |
148 | * that we can reset the context during error recovery and thereby avoid |
149 | * possible memory leaks. Formerly this code just ran in |
150 | * TopMemoryContext, but resetting that would be a really bad idea. |
151 | */ |
152 | bgwriter_context = AllocSetContextCreate(TopMemoryContext, |
153 | "Background Writer" , |
154 | ALLOCSET_DEFAULT_SIZES); |
155 | MemoryContextSwitchTo(bgwriter_context); |
156 | |
157 | WritebackContextInit(&wb_context, &bgwriter_flush_after); |
158 | |
159 | /* |
160 | * If an exception is encountered, processing resumes here. |
161 | * |
162 | * See notes in postgres.c about the design of this coding. |
163 | */ |
164 | if (sigsetjmp(local_sigjmp_buf, 1) != 0) |
165 | { |
166 | /* Since not using PG_TRY, must reset error stack by hand */ |
167 | error_context_stack = NULL; |
168 | |
169 | /* Prevent interrupts while cleaning up */ |
170 | HOLD_INTERRUPTS(); |
171 | |
172 | /* Report the error to the server log */ |
173 | EmitErrorReport(); |
174 | |
175 | /* |
176 | * These operations are really just a minimal subset of |
177 | * AbortTransaction(). We don't have very many resources to worry |
178 | * about in bgwriter, but we do have LWLocks, buffers, and temp files. |
179 | */ |
180 | LWLockReleaseAll(); |
181 | ConditionVariableCancelSleep(); |
182 | AbortBufferIO(); |
183 | UnlockBuffers(); |
184 | ReleaseAuxProcessResources(false); |
185 | AtEOXact_Buffers(false); |
186 | AtEOXact_SMgr(); |
187 | AtEOXact_Files(false); |
188 | AtEOXact_HashTables(false); |
189 | |
190 | /* |
191 | * Now return to normal top-level context and clear ErrorContext for |
192 | * next time. |
193 | */ |
194 | MemoryContextSwitchTo(bgwriter_context); |
195 | FlushErrorState(); |
196 | |
197 | /* Flush any leaked data in the top-level context */ |
198 | MemoryContextResetAndDeleteChildren(bgwriter_context); |
199 | |
200 | /* re-initialize to avoid repeated errors causing problems */ |
201 | WritebackContextInit(&wb_context, &bgwriter_flush_after); |
202 | |
203 | /* Now we can allow interrupts again */ |
204 | RESUME_INTERRUPTS(); |
205 | |
206 | /* |
207 | * Sleep at least 1 second after any error. A write error is likely |
208 | * to be repeated, and we don't want to be filling the error logs as |
209 | * fast as we can. |
210 | */ |
211 | pg_usleep(1000000L); |
212 | |
213 | /* |
214 | * Close all open files after any error. This is helpful on Windows, |
215 | * where holding deleted files open causes various strange errors. |
216 | * It's not clear we need it elsewhere, but shouldn't hurt. |
217 | */ |
218 | smgrcloseall(); |
219 | |
220 | /* Report wait end here, when there is no further possibility of wait */ |
221 | pgstat_report_wait_end(); |
222 | } |
223 | |
224 | /* We can now handle ereport(ERROR) */ |
225 | PG_exception_stack = &local_sigjmp_buf; |
226 | |
227 | /* |
228 | * Unblock signals (they were blocked when the postmaster forked us) |
229 | */ |
230 | PG_SETMASK(&UnBlockSig); |
231 | |
232 | /* |
233 | * Reset hibernation state after any error. |
234 | */ |
235 | prev_hibernate = false; |
236 | |
237 | /* |
238 | * Loop forever |
239 | */ |
240 | for (;;) |
241 | { |
242 | bool can_hibernate; |
243 | int rc; |
244 | |
245 | /* Clear any already-pending wakeups */ |
246 | ResetLatch(MyLatch); |
247 | |
248 | if (got_SIGHUP) |
249 | { |
250 | got_SIGHUP = false; |
251 | ProcessConfigFile(PGC_SIGHUP); |
252 | } |
253 | if (shutdown_requested) |
254 | { |
255 | /* |
256 | * From here on, elog(ERROR) should end with exit(1), not send |
257 | * control back to the sigsetjmp block above |
258 | */ |
259 | ExitOnAnyError = true; |
260 | /* Normal exit from the bgwriter is here */ |
261 | proc_exit(0); /* done */ |
262 | } |
263 | |
264 | /* |
265 | * Do one cycle of dirty-buffer writing. |
266 | */ |
267 | can_hibernate = BgBufferSync(&wb_context); |
268 | |
269 | /* |
270 | * Send off activity statistics to the stats collector |
271 | */ |
272 | pgstat_send_bgwriter(); |
273 | |
274 | if (FirstCallSinceLastCheckpoint()) |
275 | { |
276 | /* |
277 | * After any checkpoint, close all smgr files. This is so we |
278 | * won't hang onto smgr references to deleted files indefinitely. |
279 | */ |
280 | smgrcloseall(); |
281 | } |
282 | |
283 | /* |
284 | * Log a new xl_running_xacts every now and then so replication can |
285 | * get into a consistent state faster (think of suboverflowed |
286 | * snapshots) and clean up resources (locks, KnownXids*) more |
287 | * frequently. The costs of this are relatively low, so doing it 4 |
288 | * times (LOG_SNAPSHOT_INTERVAL_MS) a minute seems fine. |
289 | * |
290 | * We assume the interval for writing xl_running_xacts is |
291 | * significantly bigger than BgWriterDelay, so we don't complicate the |
292 | * overall timeout handling but just assume we're going to get called |
293 | * often enough even if hibernation mode is active. It's not that |
294 | * important that log_snap_interval_ms is met strictly. To make sure |
295 | * we're not waking the disk up unnecessarily on an idle system we |
296 | * check whether there has been any WAL inserted since the last time |
297 | * we've logged a running xacts. |
298 | * |
299 | * We do this logging in the bgwriter as it is the only process that |
300 | * is run regularly and returns to its mainloop all the time. E.g. |
301 | * Checkpointer, when active, is barely ever in its mainloop and thus |
302 | * makes it hard to log regularly. |
303 | */ |
304 | if (XLogStandbyInfoActive() && !RecoveryInProgress()) |
305 | { |
306 | TimestampTz timeout = 0; |
307 | TimestampTz now = GetCurrentTimestamp(); |
308 | |
309 | timeout = TimestampTzPlusMilliseconds(last_snapshot_ts, |
310 | LOG_SNAPSHOT_INTERVAL_MS); |
311 | |
312 | /* |
313 | * Only log if enough time has passed and interesting records have |
314 | * been inserted since the last snapshot. Have to compare with <= |
315 | * instead of < because GetLastImportantRecPtr() points at the |
316 | * start of a record, whereas last_snapshot_lsn points just past |
317 | * the end of the record. |
318 | */ |
319 | if (now >= timeout && |
320 | last_snapshot_lsn <= GetLastImportantRecPtr()) |
321 | { |
322 | last_snapshot_lsn = LogStandbySnapshot(); |
323 | last_snapshot_ts = now; |
324 | } |
325 | } |
326 | |
327 | /* |
328 | * Sleep until we are signaled or BgWriterDelay has elapsed. |
329 | * |
330 | * Note: the feedback control loop in BgBufferSync() expects that we |
331 | * will call it every BgWriterDelay msec. While it's not critical for |
332 | * correctness that that be exact, the feedback loop might misbehave |
333 | * if we stray too far from that. Hence, avoid loading this process |
334 | * down with latch events that are likely to happen frequently during |
335 | * normal operation. |
336 | */ |
337 | rc = WaitLatch(MyLatch, |
338 | WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH, |
339 | BgWriterDelay /* ms */ , WAIT_EVENT_BGWRITER_MAIN); |
340 | |
341 | /* |
342 | * If no latch event and BgBufferSync says nothing's happening, extend |
343 | * the sleep in "hibernation" mode, where we sleep for much longer |
344 | * than bgwriter_delay says. Fewer wakeups save electricity. When a |
345 | * backend starts using buffers again, it will wake us up by setting |
346 | * our latch. Because the extra sleep will persist only as long as no |
347 | * buffer allocations happen, this should not distort the behavior of |
348 | * BgBufferSync's control loop too badly; essentially, it will think |
349 | * that the system-wide idle interval didn't exist. |
350 | * |
351 | * There is a race condition here, in that a backend might allocate a |
352 | * buffer between the time BgBufferSync saw the alloc count as zero |
353 | * and the time we call StrategyNotifyBgWriter. While it's not |
354 | * critical that we not hibernate anyway, we try to reduce the odds of |
355 | * that by only hibernating when BgBufferSync says nothing's happening |
356 | * for two consecutive cycles. Also, we mitigate any possible |
357 | * consequences of a missed wakeup by not hibernating forever. |
358 | */ |
359 | if (rc == WL_TIMEOUT && can_hibernate && prev_hibernate) |
360 | { |
361 | /* Ask for notification at next buffer allocation */ |
362 | StrategyNotifyBgWriter(MyProc->pgprocno); |
363 | /* Sleep ... */ |
364 | (void) WaitLatch(MyLatch, |
365 | WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH, |
366 | BgWriterDelay * HIBERNATE_FACTOR, |
367 | WAIT_EVENT_BGWRITER_HIBERNATE); |
368 | /* Reset the notification request in case we timed out */ |
369 | StrategyNotifyBgWriter(-1); |
370 | } |
371 | |
372 | prev_hibernate = can_hibernate; |
373 | } |
374 | } |
375 | |
376 | |
377 | /* -------------------------------- |
378 | * signal handler routines |
379 | * -------------------------------- |
380 | */ |
381 | |
382 | /* |
383 | * bg_quickdie() occurs when signalled SIGQUIT by the postmaster. |
384 | * |
385 | * Some backend has bought the farm, |
386 | * so we need to stop what we're doing and exit. |
387 | */ |
388 | static void |
389 | bg_quickdie(SIGNAL_ARGS) |
390 | { |
391 | /* |
392 | * We DO NOT want to run proc_exit() or atexit() callbacks -- we're here |
393 | * because shared memory may be corrupted, so we don't want to try to |
394 | * clean up our transaction. Just nail the windows shut and get out of |
395 | * town. The callbacks wouldn't be safe to run from a signal handler, |
396 | * anyway. |
397 | * |
398 | * Note we do _exit(2) not _exit(0). This is to force the postmaster into |
399 | * a system reset cycle if someone sends a manual SIGQUIT to a random |
400 | * backend. This is necessary precisely because we don't clean up our |
401 | * shared memory state. (The "dead man switch" mechanism in pmsignal.c |
402 | * should ensure the postmaster sees this as a crash, too, but no harm in |
403 | * being doubly sure.) |
404 | */ |
405 | _exit(2); |
406 | } |
407 | |
408 | /* SIGHUP: set flag to re-read config file at next convenient time */ |
409 | static void |
410 | BgSigHupHandler(SIGNAL_ARGS) |
411 | { |
412 | int save_errno = errno; |
413 | |
414 | got_SIGHUP = true; |
415 | SetLatch(MyLatch); |
416 | |
417 | errno = save_errno; |
418 | } |
419 | |
420 | /* SIGTERM: set flag to shutdown and exit */ |
421 | static void |
422 | ReqShutdownHandler(SIGNAL_ARGS) |
423 | { |
424 | int save_errno = errno; |
425 | |
426 | shutdown_requested = true; |
427 | SetLatch(MyLatch); |
428 | |
429 | errno = save_errno; |
430 | } |
431 | |
432 | /* SIGUSR1: used for latch wakeups */ |
433 | static void |
434 | bgwriter_sigusr1_handler(SIGNAL_ARGS) |
435 | { |
436 | int save_errno = errno; |
437 | |
438 | latch_sigusr1_handler(); |
439 | |
440 | errno = save_errno; |
441 | } |
442 | |