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) 1996-2021, PostgreSQL Global Development Group
28 *
29 *
30 * IDENTIFICATION
31 * src/backend/postmaster/bgwriter.c
32 *
33 *-------------------------------------------------------------------------
34 */
35 #include "postgres.h"
36
37 #include "access/xlog.h"
38 #include "access/xlog_internal.h"
39 #include "libpq/pqsignal.h"
40 #include "miscadmin.h"
41 #include "pgstat.h"
42 #include "postmaster/bgwriter.h"
43 #include "postmaster/interrupt.h"
44 #include "storage/buf_internals.h"
45 #include "storage/bufmgr.h"
46 #include "storage/condition_variable.h"
47 #include "storage/fd.h"
48 #include "storage/ipc.h"
49 #include "storage/lwlock.h"
50 #include "storage/proc.h"
51 #include "storage/procsignal.h"
52 #include "storage/shmem.h"
53 #include "storage/smgr.h"
54 #include "storage/spin.h"
55 #include "storage/standby.h"
56 #include "utils/guc.h"
57 #include "utils/memutils.h"
58 #include "utils/resowner.h"
59 #include "utils/timestamp.h"
60
61 /*
62 * GUC parameters
63 */
64 int BgWriterDelay = 200;
65
66 /*
67 * Multiplier to apply to BgWriterDelay when we decide to hibernate.
68 * (Perhaps this needs to be configurable?)
69 */
70 #define HIBERNATE_FACTOR 50
71
72 /*
73 * Interval in which standby snapshots are logged into the WAL stream, in
74 * milliseconds.
75 */
76 #define LOG_SNAPSHOT_INTERVAL_MS 15000
77
78 /*
79 * LSN and timestamp at which we last issued a LogStandbySnapshot(), to avoid
80 * doing so too often or repeatedly if there has been no other write activity
81 * in the system.
82 */
83 static TimestampTz last_snapshot_ts;
84 static XLogRecPtr last_snapshot_lsn = InvalidXLogRecPtr;
85
86
87 /*
88 * Main entry point for bgwriter process
89 *
90 * This is invoked from AuxiliaryProcessMain, which has already created the
91 * basic execution environment, but not enabled signals yet.
92 */
93 void
BackgroundWriterMain(void)94 BackgroundWriterMain(void)
95 {
96 sigjmp_buf local_sigjmp_buf;
97 MemoryContext bgwriter_context;
98 bool prev_hibernate;
99 WritebackContext wb_context;
100
101 /*
102 * Properly accept or ignore signals that might be sent to us.
103 */
104 pqsignal(SIGHUP, SignalHandlerForConfigReload);
105 pqsignal(SIGINT, SIG_IGN);
106 pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
107 /* SIGQUIT handler was already set up by InitPostmasterChild */
108 pqsignal(SIGALRM, SIG_IGN);
109 pqsignal(SIGPIPE, SIG_IGN);
110 pqsignal(SIGUSR1, procsignal_sigusr1_handler);
111 pqsignal(SIGUSR2, SIG_IGN);
112
113 /*
114 * Reset some signals that are accepted by postmaster but not here
115 */
116 pqsignal(SIGCHLD, SIG_DFL);
117
118 /*
119 * We just started, assume there has been either a shutdown or
120 * end-of-recovery snapshot.
121 */
122 last_snapshot_ts = GetCurrentTimestamp();
123
124 /*
125 * Create a memory context that we will do all our work in. We do this so
126 * that we can reset the context during error recovery and thereby avoid
127 * possible memory leaks. Formerly this code just ran in
128 * TopMemoryContext, but resetting that would be a really bad idea.
129 */
130 bgwriter_context = AllocSetContextCreate(TopMemoryContext,
131 "Background Writer",
132 ALLOCSET_DEFAULT_SIZES);
133 MemoryContextSwitchTo(bgwriter_context);
134
135 WritebackContextInit(&wb_context, &bgwriter_flush_after);
136
137 /*
138 * If an exception is encountered, processing resumes here.
139 *
140 * You might wonder why this isn't coded as an infinite loop around a
141 * PG_TRY construct. The reason is that this is the bottom of the
142 * exception stack, and so with PG_TRY there would be no exception handler
143 * in force at all during the CATCH part. By leaving the outermost setjmp
144 * always active, we have at least some chance of recovering from an error
145 * during error recovery. (If we get into an infinite loop thereby, it
146 * will soon be stopped by overflow of elog.c's internal state stack.)
147 *
148 * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
149 * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
150 * signals other than SIGQUIT will be blocked until we complete error
151 * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
152 * call redundant, but it is not since InterruptPending might be set
153 * already.
154 */
155 if (sigsetjmp(local_sigjmp_buf, 1) != 0)
156 {
157 /* Since not using PG_TRY, must reset error stack by hand */
158 error_context_stack = NULL;
159
160 /* Prevent interrupts while cleaning up */
161 HOLD_INTERRUPTS();
162
163 /* Report the error to the server log */
164 EmitErrorReport();
165
166 /*
167 * These operations are really just a minimal subset of
168 * AbortTransaction(). We don't have very many resources to worry
169 * about in bgwriter, but we do have LWLocks, buffers, and temp files.
170 */
171 LWLockReleaseAll();
172 ConditionVariableCancelSleep();
173 AbortBufferIO();
174 UnlockBuffers();
175 ReleaseAuxProcessResources(false);
176 AtEOXact_Buffers(false);
177 AtEOXact_SMgr();
178 AtEOXact_Files(false);
179 AtEOXact_HashTables(false);
180
181 /*
182 * Now return to normal top-level context and clear ErrorContext for
183 * next time.
184 */
185 MemoryContextSwitchTo(bgwriter_context);
186 FlushErrorState();
187
188 /* Flush any leaked data in the top-level context */
189 MemoryContextResetAndDeleteChildren(bgwriter_context);
190
191 /* re-initialize to avoid repeated errors causing problems */
192 WritebackContextInit(&wb_context, &bgwriter_flush_after);
193
194 /* Now we can allow interrupts again */
195 RESUME_INTERRUPTS();
196
197 /*
198 * Sleep at least 1 second after any error. A write error is likely
199 * to be repeated, and we don't want to be filling the error logs as
200 * fast as we can.
201 */
202 pg_usleep(1000000L);
203
204 /*
205 * Close all open files after any error. This is helpful on Windows,
206 * where holding deleted files open causes various strange errors.
207 * It's not clear we need it elsewhere, but shouldn't hurt.
208 */
209 smgrcloseall();
210
211 /* Report wait end here, when there is no further possibility of wait */
212 pgstat_report_wait_end();
213 }
214
215 /* We can now handle ereport(ERROR) */
216 PG_exception_stack = &local_sigjmp_buf;
217
218 /*
219 * Unblock signals (they were blocked when the postmaster forked us)
220 */
221 PG_SETMASK(&UnBlockSig);
222
223 /*
224 * Reset hibernation state after any error.
225 */
226 prev_hibernate = false;
227
228 /*
229 * Loop forever
230 */
231 for (;;)
232 {
233 bool can_hibernate;
234 int rc;
235
236 /* Clear any already-pending wakeups */
237 ResetLatch(MyLatch);
238
239 HandleMainLoopInterrupts();
240
241 /*
242 * Do one cycle of dirty-buffer writing.
243 */
244 can_hibernate = BgBufferSync(&wb_context);
245
246 /*
247 * Send off activity statistics to the stats collector
248 */
249 pgstat_send_bgwriter();
250
251 if (FirstCallSinceLastCheckpoint())
252 {
253 /*
254 * After any checkpoint, close all smgr files. This is so we
255 * won't hang onto smgr references to deleted files indefinitely.
256 */
257 smgrcloseall();
258 }
259
260 /*
261 * Log a new xl_running_xacts every now and then so replication can
262 * get into a consistent state faster (think of suboverflowed
263 * snapshots) and clean up resources (locks, KnownXids*) more
264 * frequently. The costs of this are relatively low, so doing it 4
265 * times (LOG_SNAPSHOT_INTERVAL_MS) a minute seems fine.
266 *
267 * We assume the interval for writing xl_running_xacts is
268 * significantly bigger than BgWriterDelay, so we don't complicate the
269 * overall timeout handling but just assume we're going to get called
270 * often enough even if hibernation mode is active. It's not that
271 * important that LOG_SNAPSHOT_INTERVAL_MS is met strictly. To make
272 * sure we're not waking the disk up unnecessarily on an idle system
273 * we check whether there has been any WAL inserted since the last
274 * time we've logged a running xacts.
275 *
276 * We do this logging in the bgwriter as it is the only process that
277 * is run regularly and returns to its mainloop all the time. E.g.
278 * Checkpointer, when active, is barely ever in its mainloop and thus
279 * makes it hard to log regularly.
280 */
281 if (XLogStandbyInfoActive() && !RecoveryInProgress())
282 {
283 TimestampTz timeout = 0;
284 TimestampTz now = GetCurrentTimestamp();
285
286 timeout = TimestampTzPlusMilliseconds(last_snapshot_ts,
287 LOG_SNAPSHOT_INTERVAL_MS);
288
289 /*
290 * Only log if enough time has passed and interesting records have
291 * been inserted since the last snapshot. Have to compare with <=
292 * instead of < because GetLastImportantRecPtr() points at the
293 * start of a record, whereas last_snapshot_lsn points just past
294 * the end of the record.
295 */
296 if (now >= timeout &&
297 last_snapshot_lsn <= GetLastImportantRecPtr())
298 {
299 last_snapshot_lsn = LogStandbySnapshot();
300 last_snapshot_ts = now;
301 }
302 }
303
304 /*
305 * Sleep until we are signaled or BgWriterDelay has elapsed.
306 *
307 * Note: the feedback control loop in BgBufferSync() expects that we
308 * will call it every BgWriterDelay msec. While it's not critical for
309 * correctness that that be exact, the feedback loop might misbehave
310 * if we stray too far from that. Hence, avoid loading this process
311 * down with latch events that are likely to happen frequently during
312 * normal operation.
313 */
314 rc = WaitLatch(MyLatch,
315 WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
316 BgWriterDelay /* ms */ , WAIT_EVENT_BGWRITER_MAIN);
317
318 /*
319 * If no latch event and BgBufferSync says nothing's happening, extend
320 * the sleep in "hibernation" mode, where we sleep for much longer
321 * than bgwriter_delay says. Fewer wakeups save electricity. When a
322 * backend starts using buffers again, it will wake us up by setting
323 * our latch. Because the extra sleep will persist only as long as no
324 * buffer allocations happen, this should not distort the behavior of
325 * BgBufferSync's control loop too badly; essentially, it will think
326 * that the system-wide idle interval didn't exist.
327 *
328 * There is a race condition here, in that a backend might allocate a
329 * buffer between the time BgBufferSync saw the alloc count as zero
330 * and the time we call StrategyNotifyBgWriter. While it's not
331 * critical that we not hibernate anyway, we try to reduce the odds of
332 * that by only hibernating when BgBufferSync says nothing's happening
333 * for two consecutive cycles. Also, we mitigate any possible
334 * consequences of a missed wakeup by not hibernating forever.
335 */
336 if (rc == WL_TIMEOUT && can_hibernate && prev_hibernate)
337 {
338 /* Ask for notification at next buffer allocation */
339 StrategyNotifyBgWriter(MyProc->pgprocno);
340 /* Sleep ... */
341 (void) WaitLatch(MyLatch,
342 WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
343 BgWriterDelay * HIBERNATE_FACTOR,
344 WAIT_EVENT_BGWRITER_HIBERNATE);
345 /* Reset the notification request in case we timed out */
346 StrategyNotifyBgWriter(-1);
347 }
348
349 prev_hibernate = can_hibernate;
350 }
351 }
352