Branch data Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * walwriter.c
4 : : *
5 : : * The WAL writer background process is new as of Postgres 8.3. It attempts
6 : : * to keep regular backends from having to write out (and fsync) WAL pages.
7 : : * Also, it guarantees that transaction commit records that weren't synced
8 : : * to disk immediately upon commit (ie, were "asynchronously committed")
9 : : * will reach disk within a knowable time --- which, as it happens, is at
10 : : * most three times the wal_writer_delay cycle time.
11 : : *
12 : : * Note that as with the bgwriter for shared buffers, regular backends are
13 : : * still empowered to issue WAL writes and fsyncs when the walwriter doesn't
14 : : * keep up. This means that the WALWriter is not an essential process and
15 : : * can shutdown quickly when requested.
16 : : *
17 : : * Because the walwriter's cycle is directly linked to the maximum delay
18 : : * before async-commit transactions are guaranteed committed, it's probably
19 : : * unwise to load additional functionality onto it. For instance, if you've
20 : : * got a yen to create xlog segments further in advance, that'd be better done
21 : : * in bgwriter than in walwriter.
22 : : *
23 : : * The walwriter is started by the postmaster as soon as the startup subprocess
24 : : * finishes. It remains alive until the postmaster commands it to terminate.
25 : : * Normal termination is by SIGTERM, which instructs the walwriter to exit(0).
26 : : * Emergency termination is by SIGQUIT; like any backend, the walwriter will
27 : : * simply abort and exit on SIGQUIT.
28 : : *
29 : : * If the walwriter exits unexpectedly, the postmaster treats that the same
30 : : * as a backend crash: shared memory may be corrupted, so remaining backends
31 : : * should be killed by SIGQUIT and then a recovery cycle started.
32 : : *
33 : : *
34 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
35 : : *
36 : : *
37 : : * IDENTIFICATION
38 : : * src/backend/postmaster/walwriter.c
39 : : *
40 : : *-------------------------------------------------------------------------
41 : : */
42 : : #include "postgres.h"
43 : :
44 : : #include <signal.h>
45 : : #include <unistd.h>
46 : :
47 : : #include "access/xlog.h"
48 : : #include "libpq/pqsignal.h"
49 : : #include "miscadmin.h"
50 : : #include "pgstat.h"
51 : : #include "postmaster/auxprocess.h"
52 : : #include "postmaster/interrupt.h"
53 : : #include "postmaster/walwriter.h"
54 : : #include "storage/aio_subsys.h"
55 : : #include "storage/bufmgr.h"
56 : : #include "storage/condition_variable.h"
57 : : #include "storage/fd.h"
58 : : #include "storage/lwlock.h"
59 : : #include "storage/proc.h"
60 : : #include "storage/procsignal.h"
61 : : #include "storage/smgr.h"
62 : : #include "utils/hsearch.h"
63 : : #include "utils/injection_point.h"
64 : : #include "utils/memutils.h"
65 : : #include "utils/resowner.h"
66 : : #include "utils/wait_event.h"
67 : :
68 : :
69 : : /*
70 : : * GUC parameters
71 : : */
72 : : int WalWriterDelay = 200;
73 : : int WalWriterFlushAfter = DEFAULT_WAL_WRITER_FLUSH_AFTER;
74 : :
75 : : /*
76 : : * Number of do-nothing loops before lengthening the delay time, and the
77 : : * multiplier to apply to WalWriterDelay when we do decide to hibernate.
78 : : * (Perhaps these need to be configurable?)
79 : : */
80 : : #define LOOPS_UNTIL_HIBERNATE 50
81 : : #define HIBERNATE_FACTOR 25
82 : :
83 : : /*
84 : : * Main entry point for walwriter process
85 : : *
86 : : * This is invoked from AuxiliaryProcessMain, which has already created the
87 : : * basic execution environment, but not enabled signals yet.
88 : : */
89 : : void
90 : 602 : WalWriterMain(const void *startup_data, size_t startup_data_len)
91 : : {
92 : : sigjmp_buf local_sigjmp_buf;
93 : : MemoryContext walwriter_context;
94 : : int left_till_hibernate;
95 : : bool hibernating;
96 : :
97 : : Assert(startup_data_len == 0);
98 : :
99 : 602 : AuxiliaryProcessMainCommon();
100 : :
101 : : /*
102 : : * Properly accept or ignore signals the postmaster might send us
103 : : */
104 : 602 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
105 : 602 : pqsignal(SIGINT, PG_SIG_IGN); /* no query to cancel */
106 : 602 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
107 : : /* SIGQUIT handler was already set up by InitPostmasterChild */
108 : 602 : pqsignal(SIGALRM, PG_SIG_IGN);
109 : 602 : pqsignal(SIGPIPE, PG_SIG_IGN);
110 : 602 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
111 : 602 : pqsignal(SIGUSR2, PG_SIG_IGN); /* not used */
112 : :
113 : : /*
114 : : * Reset some signals that are accepted by postmaster but not here
115 : : */
116 : 602 : pqsignal(SIGCHLD, PG_SIG_DFL);
117 : :
118 : : /*
119 : : * Create a memory context that we will do all our work in. We do this so
120 : : * that we can reset the context during error recovery and thereby avoid
121 : : * possible memory leaks. Formerly this code just ran in
122 : : * TopMemoryContext, but resetting that would be a really bad idea.
123 : : */
124 : 602 : walwriter_context = AllocSetContextCreate(TopMemoryContext,
125 : : "Wal Writer",
126 : : ALLOCSET_DEFAULT_SIZES);
127 : 602 : MemoryContextSwitchTo(walwriter_context);
128 : :
129 : : /*
130 : : * If an exception is encountered, processing resumes here.
131 : : *
132 : : * You might wonder why this isn't coded as an infinite loop around a
133 : : * PG_TRY construct. The reason is that this is the bottom of the
134 : : * exception stack, and so with PG_TRY there would be no exception handler
135 : : * in force at all during the CATCH part. By leaving the outermost setjmp
136 : : * always active, we have at least some chance of recovering from an error
137 : : * during error recovery. (If we get into an infinite loop thereby, it
138 : : * will soon be stopped by overflow of elog.c's internal state stack.)
139 : : *
140 : : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
141 : : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
142 : : * signals other than SIGQUIT will be blocked until we complete error
143 : : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
144 : : * call redundant, but it is not since InterruptPending might be set
145 : : * already.
146 : : */
147 [ - + ]: 602 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
148 : : {
149 : : /* Since not using PG_TRY, must reset error stack by hand */
150 : 0 : error_context_stack = NULL;
151 : :
152 : : /* Prevent interrupts while cleaning up */
153 : 0 : HOLD_INTERRUPTS();
154 : :
155 : : /* Report the error to the server log */
156 : 0 : EmitErrorReport();
157 : :
158 : : /*
159 : : * These operations are really just a minimal subset of
160 : : * AbortTransaction(). We don't have very many resources to worry
161 : : * about in walwriter, but we do have LWLocks, and perhaps buffers?
162 : : */
163 : 0 : LWLockReleaseAll();
164 : 0 : ConditionVariableCancelSleep();
165 : 0 : pgstat_report_wait_end();
166 : 0 : pgaio_error_cleanup();
167 : 0 : UnlockBuffers();
168 : 0 : ReleaseAuxProcessResources(false);
169 : 0 : AtEOXact_Buffers(false);
170 : 0 : AtEOXact_SMgr();
171 : 0 : AtEOXact_Files(false);
172 : 0 : AtEOXact_HashTables(false);
173 : :
174 : : /*
175 : : * Now return to normal top-level context and clear ErrorContext for
176 : : * next time.
177 : : */
178 : 0 : MemoryContextSwitchTo(walwriter_context);
179 : 0 : FlushErrorState();
180 : :
181 : : /* Flush any leaked data in the top-level context */
182 : 0 : MemoryContextReset(walwriter_context);
183 : :
184 : : /* Now we can allow interrupts again */
185 : 0 : 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 : 0 : pg_usleep(1000000L);
193 : : }
194 : :
195 : : /* We can now handle ereport(ERROR) */
196 : 602 : PG_exception_stack = &local_sigjmp_buf;
197 : :
198 : : /*
199 : : * Unblock signals (they were blocked when the postmaster forked us)
200 : : */
201 : 602 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
202 : :
203 : : /*
204 : : * Reset hibernation state after any error.
205 : : */
206 : 602 : left_till_hibernate = LOOPS_UNTIL_HIBERNATE;
207 : 602 : hibernating = false;
208 : 602 : SetWalWriterSleeping(false);
209 : :
210 : : /*
211 : : * Loop forever
212 : : */
213 : : for (;;)
214 : 15543 : {
215 : : long cur_timeout;
216 : :
217 : : /*
218 : : * Advertise whether we might hibernate in this cycle. We do this
219 : : * before resetting the latch to ensure that any async commits will
220 : : * see the flag set if they might possibly need to wake us up, and
221 : : * that we won't miss any signal they send us. (If we discover work
222 : : * to do in the last cycle before we would hibernate, the global flag
223 : : * will be set unnecessarily, but little harm is done.) But avoid
224 : : * touching the global flag if it doesn't need to change.
225 : : */
226 [ + + ]: 16145 : if (hibernating != (left_till_hibernate <= 1))
227 : : {
228 : 33 : hibernating = (left_till_hibernate <= 1);
229 : 33 : SetWalWriterSleeping(hibernating);
230 : : }
231 : :
232 : : /* Clear any already-pending wakeups */
233 : 16145 : ResetLatch(MyLatch);
234 : :
235 : : /* Process any signals received recently */
236 : 16145 : ProcessMainLoopInterrupts();
237 : :
238 : 15546 : INJECTION_POINT("walwriter-before-background-flush", NULL);
239 : :
240 : : /*
241 : : * Do what we're here for; then, if XLogBackgroundFlush() found useful
242 : : * work to do, reset hibernation counter.
243 : : */
244 [ + + ]: 15546 : if (XLogBackgroundFlush())
245 : 4939 : left_till_hibernate = LOOPS_UNTIL_HIBERNATE;
246 [ + + ]: 10607 : else if (left_till_hibernate > 0)
247 : 10579 : left_till_hibernate--;
248 : :
249 : : /* report pending statistics to the cumulative stats system */
250 : 15546 : pgstat_report_wal(false);
251 : :
252 : : /*
253 : : * Sleep until we are signaled or WalWriterDelay has elapsed. If we
254 : : * haven't done anything useful for quite some time, lengthen the
255 : : * sleep time so as to reduce the server's idle power consumption.
256 : : */
257 [ + + ]: 15546 : if (left_till_hibernate > 0)
258 : 15500 : cur_timeout = WalWriterDelay; /* in ms */
259 : : else
260 : 46 : cur_timeout = WalWriterDelay * HIBERNATE_FACTOR;
261 : :
262 : 15546 : (void) WaitLatch(MyLatch,
263 : : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
264 : : cur_timeout,
265 : : WAIT_EVENT_WAL_WRITER_MAIN);
266 : : }
267 : : }
|
