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) 1996-2024, 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/buf_internals.h"
42 : #include "storage/bufmgr.h"
43 : #include "storage/condition_variable.h"
44 : #include "storage/fd.h"
45 : #include "storage/lwlock.h"
46 : #include "storage/proc.h"
47 : #include "storage/procsignal.h"
48 : #include "storage/smgr.h"
49 : #include "storage/standby.h"
50 : #include "utils/memutils.h"
51 : #include "utils/resowner.h"
52 : #include "utils/timestamp.h"
53 :
54 : /*
55 : * GUC parameters
56 : */
57 : int BgWriterDelay = 200;
58 :
59 : /*
60 : * Multiplier to apply to BgWriterDelay when we decide to hibernate.
61 : * (Perhaps this needs to be configurable?)
62 : */
63 : #define HIBERNATE_FACTOR 50
64 :
65 : /*
66 : * Interval in which standby snapshots are logged into the WAL stream, in
67 : * milliseconds.
68 : */
69 : #define LOG_SNAPSHOT_INTERVAL_MS 15000
70 :
71 : /*
72 : * LSN and timestamp at which we last issued a LogStandbySnapshot(), to avoid
73 : * doing so too often or repeatedly if there has been no other write activity
74 : * in the system.
75 : */
76 : static TimestampTz last_snapshot_ts;
77 : static XLogRecPtr last_snapshot_lsn = InvalidXLogRecPtr;
78 :
79 :
80 : /*
81 : * Main entry point for bgwriter process
82 : *
83 : * This is invoked from AuxiliaryProcessMain, which has already created the
84 : * basic execution environment, but not enabled signals yet.
85 : */
86 : void
87 894 : BackgroundWriterMain(char *startup_data, size_t startup_data_len)
88 : {
89 : sigjmp_buf local_sigjmp_buf;
90 : MemoryContext bgwriter_context;
91 : bool prev_hibernate;
92 : WritebackContext wb_context;
93 :
94 : Assert(startup_data_len == 0);
95 :
96 894 : MyBackendType = B_BG_WRITER;
97 894 : AuxiliaryProcessMainCommon();
98 :
99 : /*
100 : * Properly accept or ignore signals that might be sent to us.
101 : */
102 894 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
103 894 : pqsignal(SIGINT, SIG_IGN);
104 894 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
105 : /* SIGQUIT handler was already set up by InitPostmasterChild */
106 894 : pqsignal(SIGALRM, SIG_IGN);
107 894 : pqsignal(SIGPIPE, SIG_IGN);
108 894 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
109 894 : pqsignal(SIGUSR2, SIG_IGN);
110 :
111 : /*
112 : * Reset some signals that are accepted by postmaster but not here
113 : */
114 894 : pqsignal(SIGCHLD, SIG_DFL);
115 :
116 : /*
117 : * We just started, assume there has been either a shutdown or
118 : * end-of-recovery snapshot.
119 : */
120 894 : 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 : */
128 894 : bgwriter_context = AllocSetContextCreate(TopMemoryContext,
129 : "Background Writer",
130 : ALLOCSET_DEFAULT_SIZES);
131 894 : MemoryContextSwitchTo(bgwriter_context);
132 :
133 894 : 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 : */
153 894 : 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 : UnlockBuffers();
172 0 : ReleaseAuxProcessResources(false);
173 0 : AtEOXact_Buffers(false);
174 0 : AtEOXact_SMgr();
175 0 : AtEOXact_Files(false);
176 0 : AtEOXact_HashTables(false);
177 :
178 : /*
179 : * Now return to normal top-level context and clear ErrorContext for
180 : * next time.
181 : */
182 0 : MemoryContextSwitchTo(bgwriter_context);
183 0 : FlushErrorState();
184 :
185 : /* Flush any leaked data in the top-level context */
186 0 : MemoryContextReset(bgwriter_context);
187 :
188 : /* re-initialize to avoid repeated errors causing problems */
189 0 : WritebackContextInit(&wb_context, &bgwriter_flush_after);
190 :
191 : /* Now we can allow interrupts again */
192 0 : RESUME_INTERRUPTS();
193 :
194 : /*
195 : * Sleep at least 1 second after any error. A write error is likely
196 : * to be repeated, and we don't want to be filling the error logs as
197 : * fast as we can.
198 : */
199 0 : pg_usleep(1000000L);
200 :
201 : /* Report wait end here, when there is no further possibility of wait */
202 0 : pgstat_report_wait_end();
203 : }
204 :
205 : /* We can now handle ereport(ERROR) */
206 894 : PG_exception_stack = &local_sigjmp_buf;
207 :
208 : /*
209 : * Unblock signals (they were blocked when the postmaster forked us)
210 : */
211 894 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
212 :
213 : /*
214 : * Reset hibernation state after any error.
215 : */
216 894 : prev_hibernate = false;
217 :
218 : /*
219 : * Loop forever
220 : */
221 : for (;;)
222 17420 : {
223 : bool can_hibernate;
224 : int rc;
225 :
226 : /* Clear any already-pending wakeups */
227 18314 : ResetLatch(MyLatch);
228 :
229 18314 : HandleMainLoopInterrupts();
230 :
231 : /*
232 : * Do one cycle of dirty-buffer writing.
233 : */
234 17426 : can_hibernate = BgBufferSync(&wb_context);
235 :
236 : /* Report pending statistics to the cumulative stats system */
237 17426 : pgstat_report_bgwriter();
238 17426 : pgstat_report_wal(true);
239 :
240 17426 : if (FirstCallSinceLastCheckpoint())
241 : {
242 : /*
243 : * After any checkpoint, free all smgr objects. Otherwise we
244 : * would never do so for dropped relations, as the bgwriter does
245 : * not process shared invalidation messages or call
246 : * AtEOXact_SMgr().
247 : */
248 968 : smgrdestroyall();
249 : }
250 :
251 : /*
252 : * Log a new xl_running_xacts every now and then so replication can
253 : * get into a consistent state faster (think of suboverflowed
254 : * snapshots) and clean up resources (locks, KnownXids*) more
255 : * frequently. The costs of this are relatively low, so doing it 4
256 : * times (LOG_SNAPSHOT_INTERVAL_MS) a minute seems fine.
257 : *
258 : * We assume the interval for writing xl_running_xacts is
259 : * significantly bigger than BgWriterDelay, so we don't complicate the
260 : * overall timeout handling but just assume we're going to get called
261 : * often enough even if hibernation mode is active. It's not that
262 : * important that LOG_SNAPSHOT_INTERVAL_MS is met strictly. To make
263 : * sure we're not waking the disk up unnecessarily on an idle system
264 : * we check whether there has been any WAL inserted since the last
265 : * time we've logged a running xacts.
266 : *
267 : * We do this logging in the bgwriter as it is the only process that
268 : * is run regularly and returns to its mainloop all the time. E.g.
269 : * Checkpointer, when active, is barely ever in its mainloop and thus
270 : * makes it hard to log regularly.
271 : */
272 17426 : if (XLogStandbyInfoActive() && !RecoveryInProgress())
273 : {
274 8184 : TimestampTz timeout = 0;
275 8184 : TimestampTz now = GetCurrentTimestamp();
276 :
277 8184 : timeout = TimestampTzPlusMilliseconds(last_snapshot_ts,
278 : LOG_SNAPSHOT_INTERVAL_MS);
279 :
280 : /*
281 : * Only log if enough time has passed and interesting records have
282 : * been inserted since the last snapshot. Have to compare with <=
283 : * instead of < because GetLastImportantRecPtr() points at the
284 : * start of a record, whereas last_snapshot_lsn points just past
285 : * the end of the record.
286 : */
287 8184 : if (now >= timeout &&
288 52 : last_snapshot_lsn <= GetLastImportantRecPtr())
289 : {
290 52 : last_snapshot_lsn = LogStandbySnapshot();
291 52 : last_snapshot_ts = now;
292 : }
293 : }
294 :
295 : /*
296 : * Sleep until we are signaled or BgWriterDelay has elapsed.
297 : *
298 : * Note: the feedback control loop in BgBufferSync() expects that we
299 : * will call it every BgWriterDelay msec. While it's not critical for
300 : * correctness that that be exact, the feedback loop might misbehave
301 : * if we stray too far from that. Hence, avoid loading this process
302 : * down with latch events that are likely to happen frequently during
303 : * normal operation.
304 : */
305 17426 : rc = WaitLatch(MyLatch,
306 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
307 : BgWriterDelay /* ms */ , WAIT_EVENT_BGWRITER_MAIN);
308 :
309 : /*
310 : * If no latch event and BgBufferSync says nothing's happening, extend
311 : * the sleep in "hibernation" mode, where we sleep for much longer
312 : * than bgwriter_delay says. Fewer wakeups save electricity. When a
313 : * backend starts using buffers again, it will wake us up by setting
314 : * our latch. Because the extra sleep will persist only as long as no
315 : * buffer allocations happen, this should not distort the behavior of
316 : * BgBufferSync's control loop too badly; essentially, it will think
317 : * that the system-wide idle interval didn't exist.
318 : *
319 : * There is a race condition here, in that a backend might allocate a
320 : * buffer between the time BgBufferSync saw the alloc count as zero
321 : * and the time we call StrategyNotifyBgWriter. While it's not
322 : * critical that we not hibernate anyway, we try to reduce the odds of
323 : * that by only hibernating when BgBufferSync says nothing's happening
324 : * for two consecutive cycles. Also, we mitigate any possible
325 : * consequences of a missed wakeup by not hibernating forever.
326 : */
327 17420 : if (rc == WL_TIMEOUT && can_hibernate && prev_hibernate)
328 : {
329 : /* Ask for notification at next buffer allocation */
330 374 : StrategyNotifyBgWriter(MyProcNumber);
331 : /* Sleep ... */
332 374 : (void) WaitLatch(MyLatch,
333 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
334 374 : BgWriterDelay * HIBERNATE_FACTOR,
335 : WAIT_EVENT_BGWRITER_HIBERNATE);
336 : /* Reset the notification request in case we timed out */
337 374 : StrategyNotifyBgWriter(-1);
338 : }
339 :
340 17420 : prev_hibernate = can_hibernate;
341 : }
342 : }
|
