Line data Source code
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 : * Normal termination is by SIGTERM, which instructs the bgwriter to exit(0).
16 : * Emergency termination is by SIGQUIT; like any backend, the bgwriter will
17 : * simply abort and exit on SIGQUIT.
18 : *
19 : * If the bgwriter exits unexpectedly, the postmaster treats that the same
20 : * as a backend crash: shared memory may be corrupted, so remaining backends
21 : * should be killed by SIGQUIT and then a recovery cycle started.
22 : *
23 : *
24 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
25 : *
26 : *
27 : * IDENTIFICATION
28 : * src/backend/postmaster/bgwriter.c
29 : *
30 : *-------------------------------------------------------------------------
31 : */
32 : #include "postgres.h"
33 :
34 : #include "access/xlog.h"
35 : #include "libpq/pqsignal.h"
36 : #include "miscadmin.h"
37 : #include "pgstat.h"
38 : #include "postmaster/auxprocess.h"
39 : #include "postmaster/bgwriter.h"
40 : #include "postmaster/interrupt.h"
41 : #include "storage/aio_subsys.h"
42 : #include "storage/buf_internals.h"
43 : #include "storage/bufmgr.h"
44 : #include "storage/condition_variable.h"
45 : #include "storage/fd.h"
46 : #include "storage/lwlock.h"
47 : #include "storage/proc.h"
48 : #include "storage/procsignal.h"
49 : #include "storage/smgr.h"
50 : #include "storage/standby.h"
51 : #include "utils/memutils.h"
52 : #include "utils/resowner.h"
53 : #include "utils/timestamp.h"
54 :
55 : /*
56 : * GUC parameters
57 : */
58 : int BgWriterDelay = 200;
59 :
60 : /*
61 : * Multiplier to apply to BgWriterDelay when we decide to hibernate.
62 : * (Perhaps this needs to be configurable?)
63 : */
64 : #define HIBERNATE_FACTOR 50
65 :
66 : /*
67 : * Interval in which standby snapshots are logged into the WAL stream, in
68 : * milliseconds.
69 : */
70 : #define LOG_SNAPSHOT_INTERVAL_MS 15000
71 :
72 : /*
73 : * LSN and timestamp at which we last issued a LogStandbySnapshot(), to avoid
74 : * doing so too often or repeatedly if there has been no other write activity
75 : * in the system.
76 : */
77 : static TimestampTz last_snapshot_ts;
78 : static XLogRecPtr last_snapshot_lsn = InvalidXLogRecPtr;
79 :
80 :
81 : /*
82 : * Main entry point for bgwriter process
83 : *
84 : * This is invoked from AuxiliaryProcessMain, which has already created the
85 : * basic execution environment, but not enabled signals yet.
86 : */
87 : void
88 1142 : BackgroundWriterMain(const void *startup_data, size_t startup_data_len)
89 : {
90 : sigjmp_buf local_sigjmp_buf;
91 : MemoryContext bgwriter_context;
92 : bool prev_hibernate;
93 : WritebackContext wb_context;
94 :
95 : Assert(startup_data_len == 0);
96 :
97 1142 : AuxiliaryProcessMainCommon();
98 :
99 : /*
100 : * Properly accept or ignore signals that might be sent to us.
101 : */
102 1142 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
103 1142 : pqsignal(SIGINT, SIG_IGN);
104 1142 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
105 : /* SIGQUIT handler was already set up by InitPostmasterChild */
106 1142 : pqsignal(SIGALRM, SIG_IGN);
107 1142 : pqsignal(SIGPIPE, SIG_IGN);
108 1142 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
109 1142 : pqsignal(SIGUSR2, SIG_IGN);
110 :
111 : /*
112 : * Reset some signals that are accepted by postmaster but not here
113 : */
114 1142 : pqsignal(SIGCHLD, SIG_DFL);
115 :
116 : /*
117 : * We just started, assume there has been either a shutdown or
118 : * end-of-recovery snapshot.
119 : */
120 1142 : last_snapshot_ts = GetCurrentTimestamp();
121 :
122 : /*
123 : * Create a memory context that we will do all our work in. We do this so
124 : * that we can reset the context during error recovery and thereby avoid
125 : * possible memory leaks. Formerly this code just ran in
126 : * TopMemoryContext, but resetting that would be a really bad idea.
127 : */
128 1142 : bgwriter_context = AllocSetContextCreate(TopMemoryContext,
129 : "Background Writer",
130 : ALLOCSET_DEFAULT_SIZES);
131 1142 : MemoryContextSwitchTo(bgwriter_context);
132 :
133 1142 : WritebackContextInit(&wb_context, &bgwriter_flush_after);
134 :
135 : /*
136 : * If an exception is encountered, processing resumes here.
137 : *
138 : * You might wonder why this isn't coded as an infinite loop around a
139 : * PG_TRY construct. The reason is that this is the bottom of the
140 : * exception stack, and so with PG_TRY there would be no exception handler
141 : * in force at all during the CATCH part. By leaving the outermost setjmp
142 : * always active, we have at least some chance of recovering from an error
143 : * during error recovery. (If we get into an infinite loop thereby, it
144 : * will soon be stopped by overflow of elog.c's internal state stack.)
145 : *
146 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
147 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
148 : * signals other than SIGQUIT will be blocked until we complete error
149 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
150 : * call redundant, but it is not since InterruptPending might be set
151 : * already.
152 : */
153 1142 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
154 : {
155 : /* Since not using PG_TRY, must reset error stack by hand */
156 0 : error_context_stack = NULL;
157 :
158 : /* Prevent interrupts while cleaning up */
159 0 : HOLD_INTERRUPTS();
160 :
161 : /* Report the error to the server log */
162 0 : EmitErrorReport();
163 :
164 : /*
165 : * These operations are really just a minimal subset of
166 : * AbortTransaction(). We don't have very many resources to worry
167 : * about in bgwriter, but we do have LWLocks, buffers, and temp files.
168 : */
169 0 : LWLockReleaseAll();
170 0 : ConditionVariableCancelSleep();
171 0 : pgaio_error_cleanup();
172 0 : UnlockBuffers();
173 0 : ReleaseAuxProcessResources(false);
174 0 : AtEOXact_Buffers(false);
175 0 : AtEOXact_SMgr();
176 0 : AtEOXact_Files(false);
177 0 : AtEOXact_HashTables(false);
178 :
179 : /*
180 : * Now return to normal top-level context and clear ErrorContext for
181 : * next time.
182 : */
183 0 : MemoryContextSwitchTo(bgwriter_context);
184 0 : FlushErrorState();
185 :
186 : /* Flush any leaked data in the top-level context */
187 0 : MemoryContextReset(bgwriter_context);
188 :
189 : /* re-initialize to avoid repeated errors causing problems */
190 0 : WritebackContextInit(&wb_context, &bgwriter_flush_after);
191 :
192 : /* Now we can allow interrupts again */
193 0 : RESUME_INTERRUPTS();
194 :
195 : /*
196 : * Sleep at least 1 second after any error. A write error is likely
197 : * to be repeated, and we don't want to be filling the error logs as
198 : * fast as we can.
199 : */
200 0 : pg_usleep(1000000L);
201 :
202 : /* Report wait end here, when there is no further possibility of wait */
203 0 : pgstat_report_wait_end();
204 : }
205 :
206 : /* We can now handle ereport(ERROR) */
207 1142 : PG_exception_stack = &local_sigjmp_buf;
208 :
209 : /*
210 : * Unblock signals (they were blocked when the postmaster forked us)
211 : */
212 1142 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
213 :
214 : /*
215 : * Reset hibernation state after any error.
216 : */
217 1142 : prev_hibernate = false;
218 :
219 : /*
220 : * Loop forever
221 : */
222 : for (;;)
223 28642 : {
224 : bool can_hibernate;
225 : int rc;
226 :
227 : /* Clear any already-pending wakeups */
228 29784 : ResetLatch(MyLatch);
229 :
230 29784 : ProcessMainLoopInterrupts();
231 :
232 : /*
233 : * Do one cycle of dirty-buffer writing.
234 : */
235 28654 : can_hibernate = BgBufferSync(&wb_context);
236 :
237 : /* Report pending statistics to the cumulative stats system */
238 28654 : pgstat_report_bgwriter();
239 28654 : pgstat_report_wal(true);
240 :
241 28654 : if (FirstCallSinceLastCheckpoint())
242 : {
243 : /*
244 : * After any checkpoint, free all smgr objects. Otherwise we
245 : * would never do so for dropped relations, as the bgwriter does
246 : * not process shared invalidation messages or call
247 : * AtEOXact_SMgr().
248 : */
249 1328 : smgrdestroyall();
250 : }
251 :
252 : /*
253 : * Log a new xl_running_xacts every now and then so replication can
254 : * get into a consistent state faster (think of suboverflowed
255 : * snapshots) and clean up resources (locks, KnownXids*) more
256 : * frequently. The costs of this are relatively low, so doing it 4
257 : * times (LOG_SNAPSHOT_INTERVAL_MS) a minute seems fine.
258 : *
259 : * We assume the interval for writing xl_running_xacts is
260 : * significantly bigger than BgWriterDelay, so we don't complicate the
261 : * overall timeout handling but just assume we're going to get called
262 : * often enough even if hibernation mode is active. It's not that
263 : * important that LOG_SNAPSHOT_INTERVAL_MS is met strictly. To make
264 : * sure we're not waking the disk up unnecessarily on an idle system
265 : * we check whether there has been any WAL inserted since the last
266 : * time we've logged a running xacts.
267 : *
268 : * We do this logging in the bgwriter as it is the only process that
269 : * is run regularly and returns to its mainloop all the time. E.g.
270 : * Checkpointer, when active, is barely ever in its mainloop and thus
271 : * makes it hard to log regularly.
272 : */
273 28654 : if (XLogStandbyInfoActive() && !RecoveryInProgress())
274 : {
275 13710 : TimestampTz timeout = 0;
276 13710 : TimestampTz now = GetCurrentTimestamp();
277 :
278 13710 : timeout = TimestampTzPlusMilliseconds(last_snapshot_ts,
279 : LOG_SNAPSHOT_INTERVAL_MS);
280 :
281 : /*
282 : * Only log if enough time has passed and interesting records have
283 : * been inserted since the last snapshot. Have to compare with <=
284 : * instead of < because GetLastImportantRecPtr() points at the
285 : * start of a record, whereas last_snapshot_lsn points just past
286 : * the end of the record.
287 : */
288 13710 : if (now >= timeout &&
289 114 : last_snapshot_lsn <= GetLastImportantRecPtr())
290 : {
291 114 : last_snapshot_lsn = LogStandbySnapshot();
292 114 : last_snapshot_ts = now;
293 : }
294 : }
295 :
296 : /*
297 : * Sleep until we are signaled or BgWriterDelay has elapsed.
298 : *
299 : * Note: the feedback control loop in BgBufferSync() expects that we
300 : * will call it every BgWriterDelay msec. While it's not critical for
301 : * correctness that that be exact, the feedback loop might misbehave
302 : * if we stray too far from that. Hence, avoid loading this process
303 : * down with latch events that are likely to happen frequently during
304 : * normal operation.
305 : */
306 28654 : rc = WaitLatch(MyLatch,
307 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
308 : BgWriterDelay /* ms */ , WAIT_EVENT_BGWRITER_MAIN);
309 :
310 : /*
311 : * If no latch event and BgBufferSync says nothing's happening, extend
312 : * the sleep in "hibernation" mode, where we sleep for much longer
313 : * than bgwriter_delay says. Fewer wakeups save electricity. When a
314 : * backend starts using buffers again, it will wake us up by setting
315 : * our latch. Because the extra sleep will persist only as long as no
316 : * buffer allocations happen, this should not distort the behavior of
317 : * BgBufferSync's control loop too badly; essentially, it will think
318 : * that the system-wide idle interval didn't exist.
319 : *
320 : * There is a race condition here, in that a backend might allocate a
321 : * buffer between the time BgBufferSync saw the alloc count as zero
322 : * and the time we call StrategyNotifyBgWriter. While it's not
323 : * critical that we not hibernate anyway, we try to reduce the odds of
324 : * that by only hibernating when BgBufferSync says nothing's happening
325 : * for two consecutive cycles. Also, we mitigate any possible
326 : * consequences of a missed wakeup by not hibernating forever.
327 : */
328 28642 : if (rc == WL_TIMEOUT && can_hibernate && prev_hibernate)
329 : {
330 : /* Ask for notification at next buffer allocation */
331 1498 : StrategyNotifyBgWriter(MyProcNumber);
332 : /* Sleep ... */
333 1498 : (void) WaitLatch(MyLatch,
334 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
335 1498 : BgWriterDelay * HIBERNATE_FACTOR,
336 : WAIT_EVENT_BGWRITER_HIBERNATE);
337 : /* Reset the notification request in case we timed out */
338 1498 : StrategyNotifyBgWriter(-1);
339 : }
340 :
341 28642 : prev_hibernate = can_hibernate;
342 : }
343 : }
|
