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 :
66 :
67 : /*
68 : * GUC parameters
69 : */
70 : int WalWriterDelay = 200;
71 : int WalWriterFlushAfter = DEFAULT_WAL_WRITER_FLUSH_AFTER;
72 :
73 : /*
74 : * Number of do-nothing loops before lengthening the delay time, and the
75 : * multiplier to apply to WalWriterDelay when we do decide to hibernate.
76 : * (Perhaps these need to be configurable?)
77 : */
78 : #define LOOPS_UNTIL_HIBERNATE 50
79 : #define HIBERNATE_FACTOR 25
80 :
81 : /*
82 : * Main entry point for walwriter 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 1014 : WalWriterMain(const void *startup_data, size_t startup_data_len)
89 : {
90 : sigjmp_buf local_sigjmp_buf;
91 : MemoryContext walwriter_context;
92 : int left_till_hibernate;
93 : bool hibernating;
94 :
95 : Assert(startup_data_len == 0);
96 :
97 1014 : AuxiliaryProcessMainCommon();
98 :
99 : /*
100 : * Properly accept or ignore signals the postmaster might send us
101 : *
102 : * We have no particular use for SIGINT at the moment, but seems
103 : * reasonable to treat like SIGTERM.
104 : */
105 1014 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
106 1014 : pqsignal(SIGINT, SignalHandlerForShutdownRequest);
107 1014 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
108 : /* SIGQUIT handler was already set up by InitPostmasterChild */
109 1014 : pqsignal(SIGALRM, SIG_IGN);
110 1014 : pqsignal(SIGPIPE, SIG_IGN);
111 1014 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
112 1014 : pqsignal(SIGUSR2, SIG_IGN); /* not used */
113 :
114 : /*
115 : * Reset some signals that are accepted by postmaster but not here
116 : */
117 1014 : pqsignal(SIGCHLD, SIG_DFL);
118 :
119 : /*
120 : * Create a memory context that we will do all our work in. We do this so
121 : * that we can reset the context during error recovery and thereby avoid
122 : * possible memory leaks. Formerly this code just ran in
123 : * TopMemoryContext, but resetting that would be a really bad idea.
124 : */
125 1014 : walwriter_context = AllocSetContextCreate(TopMemoryContext,
126 : "Wal Writer",
127 : ALLOCSET_DEFAULT_SIZES);
128 1014 : MemoryContextSwitchTo(walwriter_context);
129 :
130 : /*
131 : * If an exception is encountered, processing resumes here.
132 : *
133 : * You might wonder why this isn't coded as an infinite loop around a
134 : * PG_TRY construct. The reason is that this is the bottom of the
135 : * exception stack, and so with PG_TRY there would be no exception handler
136 : * in force at all during the CATCH part. By leaving the outermost setjmp
137 : * always active, we have at least some chance of recovering from an error
138 : * during error recovery. (If we get into an infinite loop thereby, it
139 : * will soon be stopped by overflow of elog.c's internal state stack.)
140 : *
141 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
142 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
143 : * signals other than SIGQUIT will be blocked until we complete error
144 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
145 : * call redundant, but it is not since InterruptPending might be set
146 : * already.
147 : */
148 1014 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
149 : {
150 : /* Since not using PG_TRY, must reset error stack by hand */
151 0 : error_context_stack = NULL;
152 :
153 : /* Prevent interrupts while cleaning up */
154 0 : HOLD_INTERRUPTS();
155 :
156 : /* Report the error to the server log */
157 0 : EmitErrorReport();
158 :
159 : /*
160 : * These operations are really just a minimal subset of
161 : * AbortTransaction(). We don't have very many resources to worry
162 : * about in walwriter, but we do have LWLocks, and perhaps buffers?
163 : */
164 0 : LWLockReleaseAll();
165 0 : ConditionVariableCancelSleep();
166 0 : pgstat_report_wait_end();
167 0 : pgaio_error_cleanup();
168 0 : UnlockBuffers();
169 0 : ReleaseAuxProcessResources(false);
170 0 : AtEOXact_Buffers(false);
171 0 : AtEOXact_SMgr();
172 0 : AtEOXact_Files(false);
173 0 : AtEOXact_HashTables(false);
174 :
175 : /*
176 : * Now return to normal top-level context and clear ErrorContext for
177 : * next time.
178 : */
179 0 : MemoryContextSwitchTo(walwriter_context);
180 0 : FlushErrorState();
181 :
182 : /* Flush any leaked data in the top-level context */
183 0 : MemoryContextReset(walwriter_context);
184 :
185 : /* Now we can allow interrupts again */
186 0 : RESUME_INTERRUPTS();
187 :
188 : /*
189 : * Sleep at least 1 second after any error. A write error is likely
190 : * to be repeated, and we don't want to be filling the error logs as
191 : * fast as we can.
192 : */
193 0 : pg_usleep(1000000L);
194 : }
195 :
196 : /* We can now handle ereport(ERROR) */
197 1014 : PG_exception_stack = &local_sigjmp_buf;
198 :
199 : /*
200 : * Unblock signals (they were blocked when the postmaster forked us)
201 : */
202 1014 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
203 :
204 : /*
205 : * Reset hibernation state after any error.
206 : */
207 1014 : left_till_hibernate = LOOPS_UNTIL_HIBERNATE;
208 1014 : hibernating = false;
209 1014 : SetWalWriterSleeping(false);
210 :
211 : /*
212 : * Advertise our proc number that backends can use to wake us up while
213 : * we're sleeping.
214 : */
215 1014 : ProcGlobal->walwriterProc = MyProcNumber;
216 :
217 : /*
218 : * Loop forever
219 : */
220 : for (;;)
221 32210 : {
222 : long cur_timeout;
223 :
224 : /*
225 : * Advertise whether we might hibernate in this cycle. We do this
226 : * before resetting the latch to ensure that any async commits will
227 : * see the flag set if they might possibly need to wake us up, and
228 : * that we won't miss any signal they send us. (If we discover work
229 : * to do in the last cycle before we would hibernate, the global flag
230 : * will be set unnecessarily, but little harm is done.) But avoid
231 : * touching the global flag if it doesn't need to change.
232 : */
233 33224 : if (hibernating != (left_till_hibernate <= 1))
234 : {
235 38 : hibernating = (left_till_hibernate <= 1);
236 38 : SetWalWriterSleeping(hibernating);
237 : }
238 :
239 : /* Clear any already-pending wakeups */
240 33224 : ResetLatch(MyLatch);
241 :
242 : /* Process any signals received recently */
243 33224 : ProcessMainLoopInterrupts();
244 :
245 : /*
246 : * Do what we're here for; then, if XLogBackgroundFlush() found useful
247 : * work to do, reset hibernation counter.
248 : */
249 32216 : if (XLogBackgroundFlush())
250 10318 : left_till_hibernate = LOOPS_UNTIL_HIBERNATE;
251 21898 : else if (left_till_hibernate > 0)
252 21868 : left_till_hibernate--;
253 :
254 : /* report pending statistics to the cumulative stats system */
255 32216 : pgstat_report_wal(false);
256 :
257 : /*
258 : * Sleep until we are signaled or WalWriterDelay has elapsed. If we
259 : * haven't done anything useful for quite some time, lengthen the
260 : * sleep time so as to reduce the server's idle power consumption.
261 : */
262 32216 : if (left_till_hibernate > 0)
263 32160 : cur_timeout = WalWriterDelay; /* in ms */
264 : else
265 56 : cur_timeout = WalWriterDelay * HIBERNATE_FACTOR;
266 :
267 32216 : (void) WaitLatch(MyLatch,
268 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
269 : cur_timeout,
270 : WAIT_EVENT_WAL_WRITER_MAIN);
271 : }
272 : }
|
