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/memutils.h"
64 : #include "utils/resowner.h"
65 : #include "utils/wait_event.h"
66 :
67 :
68 : /*
69 : * GUC parameters
70 : */
71 : int WalWriterDelay = 200;
72 : int WalWriterFlushAfter = DEFAULT_WAL_WRITER_FLUSH_AFTER;
73 :
74 : /*
75 : * Number of do-nothing loops before lengthening the delay time, and the
76 : * multiplier to apply to WalWriterDelay when we do decide to hibernate.
77 : * (Perhaps these need to be configurable?)
78 : */
79 : #define LOOPS_UNTIL_HIBERNATE 50
80 : #define HIBERNATE_FACTOR 25
81 :
82 : /*
83 : * Main entry point for walwriter process
84 : *
85 : * This is invoked from AuxiliaryProcessMain, which has already created the
86 : * basic execution environment, but not enabled signals yet.
87 : */
88 : void
89 525 : WalWriterMain(const void *startup_data, size_t startup_data_len)
90 : {
91 : sigjmp_buf local_sigjmp_buf;
92 : MemoryContext walwriter_context;
93 : int left_till_hibernate;
94 : bool hibernating;
95 :
96 : Assert(startup_data_len == 0);
97 :
98 525 : AuxiliaryProcessMainCommon();
99 :
100 : /*
101 : * Properly accept or ignore signals the postmaster might send us
102 : */
103 525 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
104 525 : pqsignal(SIGINT, SIG_IGN); /* no query to cancel */
105 525 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
106 : /* SIGQUIT handler was already set up by InitPostmasterChild */
107 525 : pqsignal(SIGALRM, SIG_IGN);
108 525 : pqsignal(SIGPIPE, SIG_IGN);
109 525 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
110 525 : pqsignal(SIGUSR2, SIG_IGN); /* not used */
111 :
112 : /*
113 : * Reset some signals that are accepted by postmaster but not here
114 : */
115 525 : pqsignal(SIGCHLD, SIG_DFL);
116 :
117 : /*
118 : * Create a memory context that we will do all our work in. We do this so
119 : * that we can reset the context during error recovery and thereby avoid
120 : * possible memory leaks. Formerly this code just ran in
121 : * TopMemoryContext, but resetting that would be a really bad idea.
122 : */
123 525 : walwriter_context = AllocSetContextCreate(TopMemoryContext,
124 : "Wal Writer",
125 : ALLOCSET_DEFAULT_SIZES);
126 525 : MemoryContextSwitchTo(walwriter_context);
127 :
128 : /*
129 : * If an exception is encountered, processing resumes here.
130 : *
131 : * You might wonder why this isn't coded as an infinite loop around a
132 : * PG_TRY construct. The reason is that this is the bottom of the
133 : * exception stack, and so with PG_TRY there would be no exception handler
134 : * in force at all during the CATCH part. By leaving the outermost setjmp
135 : * always active, we have at least some chance of recovering from an error
136 : * during error recovery. (If we get into an infinite loop thereby, it
137 : * will soon be stopped by overflow of elog.c's internal state stack.)
138 : *
139 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
140 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
141 : * signals other than SIGQUIT will be blocked until we complete error
142 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
143 : * call redundant, but it is not since InterruptPending might be set
144 : * already.
145 : */
146 525 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
147 : {
148 : /* Since not using PG_TRY, must reset error stack by hand */
149 0 : error_context_stack = NULL;
150 :
151 : /* Prevent interrupts while cleaning up */
152 0 : HOLD_INTERRUPTS();
153 :
154 : /* Report the error to the server log */
155 0 : EmitErrorReport();
156 :
157 : /*
158 : * These operations are really just a minimal subset of
159 : * AbortTransaction(). We don't have very many resources to worry
160 : * about in walwriter, but we do have LWLocks, and perhaps buffers?
161 : */
162 0 : LWLockReleaseAll();
163 0 : ConditionVariableCancelSleep();
164 0 : pgstat_report_wait_end();
165 0 : pgaio_error_cleanup();
166 0 : UnlockBuffers();
167 0 : ReleaseAuxProcessResources(false);
168 0 : AtEOXact_Buffers(false);
169 0 : AtEOXact_SMgr();
170 0 : AtEOXact_Files(false);
171 0 : AtEOXact_HashTables(false);
172 :
173 : /*
174 : * Now return to normal top-level context and clear ErrorContext for
175 : * next time.
176 : */
177 0 : MemoryContextSwitchTo(walwriter_context);
178 0 : FlushErrorState();
179 :
180 : /* Flush any leaked data in the top-level context */
181 0 : MemoryContextReset(walwriter_context);
182 :
183 : /* Now we can allow interrupts again */
184 0 : RESUME_INTERRUPTS();
185 :
186 : /*
187 : * Sleep at least 1 second after any error. A write error is likely
188 : * to be repeated, and we don't want to be filling the error logs as
189 : * fast as we can.
190 : */
191 0 : pg_usleep(1000000L);
192 : }
193 :
194 : /* We can now handle ereport(ERROR) */
195 525 : PG_exception_stack = &local_sigjmp_buf;
196 :
197 : /*
198 : * Unblock signals (they were blocked when the postmaster forked us)
199 : */
200 525 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
201 :
202 : /*
203 : * Reset hibernation state after any error.
204 : */
205 525 : left_till_hibernate = LOOPS_UNTIL_HIBERNATE;
206 525 : hibernating = false;
207 525 : SetWalWriterSleeping(false);
208 :
209 : /*
210 : * Advertise our proc number that backends can use to wake us up while
211 : * we're sleeping.
212 : */
213 525 : ProcGlobal->walwriterProc = MyProcNumber;
214 :
215 : /*
216 : * Loop forever
217 : */
218 : for (;;)
219 14917 : {
220 : long cur_timeout;
221 :
222 : /*
223 : * Advertise whether we might hibernate in this cycle. We do this
224 : * before resetting the latch to ensure that any async commits will
225 : * see the flag set if they might possibly need to wake us up, and
226 : * that we won't miss any signal they send us. (If we discover work
227 : * to do in the last cycle before we would hibernate, the global flag
228 : * will be set unnecessarily, but little harm is done.) But avoid
229 : * touching the global flag if it doesn't need to change.
230 : */
231 15442 : if (hibernating != (left_till_hibernate <= 1))
232 : {
233 30 : hibernating = (left_till_hibernate <= 1);
234 30 : SetWalWriterSleeping(hibernating);
235 : }
236 :
237 : /* Clear any already-pending wakeups */
238 15442 : ResetLatch(MyLatch);
239 :
240 : /* Process any signals received recently */
241 15442 : ProcessMainLoopInterrupts();
242 :
243 : /*
244 : * Do what we're here for; then, if XLogBackgroundFlush() found useful
245 : * work to do, reset hibernation counter.
246 : */
247 14920 : if (XLogBackgroundFlush())
248 4285 : left_till_hibernate = LOOPS_UNTIL_HIBERNATE;
249 10635 : else if (left_till_hibernate > 0)
250 10603 : left_till_hibernate--;
251 :
252 : /* report pending statistics to the cumulative stats system */
253 14920 : pgstat_report_wal(false);
254 :
255 : /*
256 : * Sleep until we are signaled or WalWriterDelay has elapsed. If we
257 : * haven't done anything useful for quite some time, lengthen the
258 : * sleep time so as to reduce the server's idle power consumption.
259 : */
260 14920 : if (left_till_hibernate > 0)
261 14869 : cur_timeout = WalWriterDelay; /* in ms */
262 : else
263 51 : cur_timeout = WalWriterDelay * HIBERNATE_FACTOR;
264 :
265 14920 : (void) WaitLatch(MyLatch,
266 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
267 : cur_timeout,
268 : WAIT_EVENT_WAL_WRITER_MAIN);
269 : }
270 : }
|
