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) 1996-2020, 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 pqsignal(SIGQUIT, SignalHandlerForCrashExit);
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 /* We allow SIGQUIT (quickdie) at all times */
119 sigdelset(&BlockSig, SIGQUIT);
120
121 /*
122 * We just started, assume there has been either a shutdown or
123 * end-of-recovery snapshot.
124 */
125 last_snapshot_ts = GetCurrentTimestamp();
126
127 /*
128 * Create a memory context that we will do all our work in. We do this so
129 * that we can reset the context during error recovery and thereby avoid
130 * possible memory leaks. Formerly this code just ran in
131 * TopMemoryContext, but resetting that would be a really bad idea.
132 */
133 bgwriter_context = AllocSetContextCreate(TopMemoryContext,
134 "Background Writer",
135 ALLOCSET_DEFAULT_SIZES);
136 MemoryContextSwitchTo(bgwriter_context);
137
138 WritebackContextInit(&wb_context, &bgwriter_flush_after);
139
140 /*
141 * If an exception is encountered, processing resumes here.
142 *
143 * See notes in postgres.c about the design of this coding.
144 */
145 if (sigsetjmp(local_sigjmp_buf, 1) != 0)
146 {
147 /* Since not using PG_TRY, must reset error stack by hand */
148 error_context_stack = NULL;
149
150 /* Prevent interrupts while cleaning up */
151 HOLD_INTERRUPTS();
152
153 /* Report the error to the server log */
154 EmitErrorReport();
155
156 /*
157 * These operations are really just a minimal subset of
158 * AbortTransaction(). We don't have very many resources to worry
159 * about in bgwriter, but we do have LWLocks, buffers, and temp files.
160 */
161 LWLockReleaseAll();
162 ConditionVariableCancelSleep();
163 AbortBufferIO();
164 UnlockBuffers();
165 ReleaseAuxProcessResources(false);
166 AtEOXact_Buffers(false);
167 AtEOXact_SMgr();
168 AtEOXact_Files(false);
169 AtEOXact_HashTables(false);
170
171 /*
172 * Now return to normal top-level context and clear ErrorContext for
173 * next time.
174 */
175 MemoryContextSwitchTo(bgwriter_context);
176 FlushErrorState();
177
178 /* Flush any leaked data in the top-level context */
179 MemoryContextResetAndDeleteChildren(bgwriter_context);
180
181 /* re-initialize to avoid repeated errors causing problems */
182 WritebackContextInit(&wb_context, &bgwriter_flush_after);
183
184 /* Now we can allow interrupts again */
185 RESUME_INTERRUPTS();
186
187 /*
188 * Sleep at least 1 second after any error. A write error is likely
189 * to be repeated, and we don't want to be filling the error logs as
190 * fast as we can.
191 */
192 pg_usleep(1000000L);
193
194 /*
195 * Close all open files after any error. This is helpful on Windows,
196 * where holding deleted files open causes various strange errors.
197 * It's not clear we need it elsewhere, but shouldn't hurt.
198 */
199 smgrcloseall();
200
201 /* Report wait end here, when there is no further possibility of wait */
202 pgstat_report_wait_end();
203 }
204
205 /* We can now handle ereport(ERROR) */
206 PG_exception_stack = &local_sigjmp_buf;
207
208 /*
209 * Unblock signals (they were blocked when the postmaster forked us)
210 */
211 PG_SETMASK(&UnBlockSig);
212
213 /*
214 * Reset hibernation state after any error.
215 */
216 prev_hibernate = false;
217
218 /*
219 * Loop forever
220 */
221 for (;;)
222 {
223 bool can_hibernate;
224 int rc;
225
226 /* Clear any already-pending wakeups */
227 ResetLatch(MyLatch);
228
229 HandleMainLoopInterrupts();
230
231 /*
232 * Do one cycle of dirty-buffer writing.
233 */
234 can_hibernate = BgBufferSync(&wb_context);
235
236 /*
237 * Send off activity statistics to the stats collector
238 */
239 pgstat_send_bgwriter();
240
241 if (FirstCallSinceLastCheckpoint())
242 {
243 /*
244 * After any checkpoint, close all smgr files. This is so we
245 * won't hang onto smgr references to deleted files indefinitely.
246 */
247 smgrcloseall();
248 }
249
250 /*
251 * Log a new xl_running_xacts every now and then so replication can
252 * get into a consistent state faster (think of suboverflowed
253 * snapshots) and clean up resources (locks, KnownXids*) more
254 * frequently. The costs of this are relatively low, so doing it 4
255 * times (LOG_SNAPSHOT_INTERVAL_MS) a minute seems fine.
256 *
257 * We assume the interval for writing xl_running_xacts is
258 * significantly bigger than BgWriterDelay, so we don't complicate the
259 * overall timeout handling but just assume we're going to get called
260 * often enough even if hibernation mode is active. It's not that
261 * important that LOG_SNAPSHOT_INTERVAL_MS is met strictly. To make
262 * sure we're not waking the disk up unnecessarily on an idle system
263 * we check whether there has been any WAL inserted since the last
264 * time we've logged a running xacts.
265 *
266 * We do this logging in the bgwriter as it is the only process that
267 * is run regularly and returns to its mainloop all the time. E.g.
268 * Checkpointer, when active, is barely ever in its mainloop and thus
269 * makes it hard to log regularly.
270 */
271 if (XLogStandbyInfoActive() && !RecoveryInProgress())
272 {
273 TimestampTz timeout = 0;
274 TimestampTz now = GetCurrentTimestamp();
275
276 timeout = TimestampTzPlusMilliseconds(last_snapshot_ts,
277 LOG_SNAPSHOT_INTERVAL_MS);
278
279 /*
280 * Only log if enough time has passed and interesting records have
281 * been inserted since the last snapshot. Have to compare with <=
282 * instead of < because GetLastImportantRecPtr() points at the
283 * start of a record, whereas last_snapshot_lsn points just past
284 * the end of the record.
285 */
286 if (now >= timeout &&
287 last_snapshot_lsn <= GetLastImportantRecPtr())
288 {
289 last_snapshot_lsn = LogStandbySnapshot();
290 last_snapshot_ts = now;
291 }
292 }
293
294 /*
295 * Sleep until we are signaled or BgWriterDelay has elapsed.
296 *
297 * Note: the feedback control loop in BgBufferSync() expects that we
298 * will call it every BgWriterDelay msec. While it's not critical for
299 * correctness that that be exact, the feedback loop might misbehave
300 * if we stray too far from that. Hence, avoid loading this process
301 * down with latch events that are likely to happen frequently during
302 * normal operation.
303 */
304 rc = WaitLatch(MyLatch,
305 WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
306 BgWriterDelay /* ms */ , WAIT_EVENT_BGWRITER_MAIN);
307
308 /*
309 * If no latch event and BgBufferSync says nothing's happening, extend
310 * the sleep in "hibernation" mode, where we sleep for much longer
311 * than bgwriter_delay says. Fewer wakeups save electricity. When a
312 * backend starts using buffers again, it will wake us up by setting
313 * our latch. Because the extra sleep will persist only as long as no
314 * buffer allocations happen, this should not distort the behavior of
315 * BgBufferSync's control loop too badly; essentially, it will think
316 * that the system-wide idle interval didn't exist.
317 *
318 * There is a race condition here, in that a backend might allocate a
319 * buffer between the time BgBufferSync saw the alloc count as zero
320 * and the time we call StrategyNotifyBgWriter. While it's not
321 * critical that we not hibernate anyway, we try to reduce the odds of
322 * that by only hibernating when BgBufferSync says nothing's happening
323 * for two consecutive cycles. Also, we mitigate any possible
324 * consequences of a missed wakeup by not hibernating forever.
325 */
326 if (rc == WL_TIMEOUT && can_hibernate && prev_hibernate)
327 {
328 /* Ask for notification at next buffer allocation */
329 StrategyNotifyBgWriter(MyProc->pgprocno);
330 /* Sleep ... */
331 (void) WaitLatch(MyLatch,
332 WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
333 BgWriterDelay * HIBERNATE_FACTOR,
334 WAIT_EVENT_BGWRITER_HIBERNATE);
335 /* Reset the notification request in case we timed out */
336 StrategyNotifyBgWriter(-1);
337 }
338
339 prev_hibernate = can_hibernate;
340 }
341 }
342