| 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) 1996-2025, 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/aio_subsys.h" |
| 42 |
#include "storage/buf_internals.h" |
| 43 |
#include "storage/bufmgr.h" |
| 44 |
#include "storage/condition_variable.h" |
| 45 |
#include "storage/fd.h" |
| 46 |
#include "storage/lwlock.h" |
| 47 |
#include "storage/proc.h" |
| 48 |
#include "storage/procsignal.h" |
| 49 |
#include "storage/smgr.h" |
| 50 |
#include "storage/standby.h" |
| 51 |
#include "utils/memutils.h" |
| 52 |
#include "utils/resowner.h" |
| 53 |
#include "utils/timestamp.h" |
| 54 |
|
| 55 |
/* |
| 56 |
* GUC parameters |
| 57 |
*/ |
| 58 |
int BgWriterDelay = 200; |
| 59 |
|
| 60 |
/* |
| 61 |
* Multiplier to apply to BgWriterDelay when we decide to hibernate. |
| 62 |
* (Perhaps this needs to be configurable?) |
| 63 |
*/ |
| 64 |
#define HIBERNATE_FACTOR 50 |
| 65 |
|
| 66 |
/* |
| 67 |
* Interval in which standby snapshots are logged into the WAL stream, in |
| 68 |
* milliseconds. |
| 69 |
*/ |
| 70 |
#define LOG_SNAPSHOT_INTERVAL_MS 15000 |
| 71 |
|
| 72 |
/* |
| 73 |
* LSN and timestamp at which we last issued a LogStandbySnapshot(), to avoid |
| 74 |
* doing so too often or repeatedly if there has been no other write activity |
| 75 |
* in the system. |
| 76 |
*/ |
| 77 |
static TimestampTz last_snapshot_ts; |
| 78 |
static XLogRecPtr last_snapshot_lsn = InvalidXLogRecPtr; |
| 79 |
|
| 80 |
|
| 81 |
/* |
| 82 |
* Main entry point for bgwriter 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 |
BackgroundWriterMain(const void *startup_data, size_t startup_data_len) |
| 89 |
{ |
| 90 |
sigjmp_buf local_sigjmp_buf; |
| 91 |
MemoryContext bgwriter_context; |
| 92 |
bool prev_hibernate; |
| 93 |
WritebackContext wb_context; |
| 94 |
|
| 95 |
Assert(startup_data_len == 0); |
| 96 |
|
| 97 |
MyBackendType = B_BG_WRITER; |
| 98 |
AuxiliaryProcessMainCommon(); |
| 99 |
|
| 100 |
/* |
| 101 |
* Properly accept or ignore signals that might be sent to us. |
| 102 |
*/ |
| 103 |
pqsignal(SIGHUP, SignalHandlerForConfigReload); |
| 104 |
pqsignal(SIGINT, SIG_IGN); |
| 105 |
pqsignal(SIGTERM, SignalHandlerForShutdownRequest); |
| 106 |
/* SIGQUIT handler was already set up by InitPostmasterChild */ |
| 107 |
pqsignal(SIGALRM, SIG_IGN); |
| 108 |
pqsignal(SIGPIPE, SIG_IGN); |
| 109 |
pqsignal(SIGUSR1, procsignal_sigusr1_handler); |
| 110 |
pqsignal(SIGUSR2, SIG_IGN); |
| 111 |
|
| 112 |
/* |
| 113 |
* Reset some signals that are accepted by postmaster but not here |
| 114 |
*/ |
| 115 |
pqsignal(SIGCHLD, SIG_DFL); |
| 116 |
|
| 117 |
/* |
| 118 |
* We just started, assume there has been either a shutdown or |
| 119 |
* end-of-recovery snapshot. |
| 120 |
*/ |
| 121 |
last_snapshot_ts = GetCurrentTimestamp(); |
| 122 |
|
| 123 |
/* |
| 124 |
* Create a memory context that we will do all our work in. We do this so |
| 125 |
* that we can reset the context during error recovery and thereby avoid |
| 126 |
* possible memory leaks. Formerly this code just ran in |
| 127 |
* TopMemoryContext, but resetting that would be a really bad idea. |
| 128 |
*/ |
| 129 |
bgwriter_context = AllocSetContextCreate(TopMemoryContext, |
| 130 |
"Background Writer", |
| 131 |
ALLOCSET_DEFAULT_SIZES); |
| 132 |
MemoryContextSwitchTo(bgwriter_context); |
| 133 |
|
| 134 |
WritebackContextInit(&wb_context, &bgwriter_flush_after); |
| 135 |
|
| 136 |
/* |
| 137 |
* If an exception is encountered, processing resumes here. |
| 138 |
* |
| 139 |
* You might wonder why this isn't coded as an infinite loop around a |
| 140 |
* PG_TRY construct. The reason is that this is the bottom of the |
| 141 |
* exception stack, and so with PG_TRY there would be no exception handler |
| 142 |
* in force at all during the CATCH part. By leaving the outermost setjmp |
| 143 |
* always active, we have at least some chance of recovering from an error |
| 144 |
* during error recovery. (If we get into an infinite loop thereby, it |
| 145 |
* will soon be stopped by overflow of elog.c's internal state stack.) |
| 146 |
* |
| 147 |
* Note that we use sigsetjmp(..., 1), so that the prevailing signal mask |
| 148 |
* (to wit, BlockSig) will be restored when longjmp'ing to here. Thus, |
| 149 |
* signals other than SIGQUIT will be blocked until we complete error |
| 150 |
* recovery. It might seem that this policy makes the HOLD_INTERRUPTS() |
| 151 |
* call redundant, but it is not since InterruptPending might be set |
| 152 |
* already. |
| 153 |
*/ |
| 154 |
if (sigsetjmp(local_sigjmp_buf, 1) != 0) |
| 155 |
{ |
| 156 |
/* Since not using PG_TRY, must reset error stack by hand */ |
| 157 |
error_context_stack = NULL; |
| 158 |
|
| 159 |
/* Prevent interrupts while cleaning up */ |
| 160 |
HOLD_INTERRUPTS(); |
| 161 |
|
| 162 |
/* Report the error to the server log */ |
| 163 |
EmitErrorReport(); |
| 164 |
|
| 165 |
/* |
| 166 |
* These operations are really just a minimal subset of |
| 167 |
* AbortTransaction(). We don't have very many resources to worry |
| 168 |
* about in bgwriter, but we do have LWLocks, buffers, and temp files. |
| 169 |
*/ |
| 170 |
LWLockReleaseAll(); |
| 171 |
ConditionVariableCancelSleep(); |
| 172 |
pgaio_error_cleanup(); |
| 173 |
UnlockBuffers(); |
| 174 |
ReleaseAuxProcessResources(false); |
| 175 |
AtEOXact_Buffers(false); |
| 176 |
AtEOXact_SMgr(); |
| 177 |
AtEOXact_Files(false); |
| 178 |
AtEOXact_HashTables(false); |
| 179 |
|
| 180 |
/* |
| 181 |
* Now return to normal top-level context and clear ErrorContext for |
| 182 |
* next time. |
| 183 |
*/ |
| 184 |
MemoryContextSwitchTo(bgwriter_context); |
| 185 |
FlushErrorState(); |
| 186 |
|
| 187 |
/* Flush any leaked data in the top-level context */ |
| 188 |
MemoryContextReset(bgwriter_context); |
| 189 |
|
| 190 |
/* re-initialize to avoid repeated errors causing problems */ |
| 191 |
WritebackContextInit(&wb_context, &bgwriter_flush_after); |
| 192 |
|
| 193 |
/* Now we can allow interrupts again */ |
| 194 |
RESUME_INTERRUPTS(); |
| 195 |
|
| 196 |
/* |
| 197 |
* Sleep at least 1 second after any error. A write error is likely |
| 198 |
* to be repeated, and we don't want to be filling the error logs as |
| 199 |
* fast as we can. |
| 200 |
*/ |
| 201 |
pg_usleep(1000000L); |
| 202 |
|
| 203 |
/* Report wait end here, when there is no further possibility of wait */ |
| 204 |
pgstat_report_wait_end(); |
| 205 |
} |
| 206 |
|
| 207 |
/* We can now handle ereport(ERROR) */ |
| 208 |
PG_exception_stack = &local_sigjmp_buf; |
| 209 |
|
| 210 |
/* |
| 211 |
* Unblock signals (they were blocked when the postmaster forked us) |
| 212 |
*/ |
| 213 |
sigprocmask(SIG_SETMASK, &UnBlockSig, NULL); |
| 214 |
|
| 215 |
/* |
| 216 |
* Reset hibernation state after any error. |
| 217 |
*/ |
| 218 |
prev_hibernate = false; |
| 219 |
|
| 220 |
/* |
| 221 |
* Loop forever |
| 222 |
*/ |
| 223 |
for (;;) |
| 224 |
{ |
| 225 |
bool can_hibernate; |
| 226 |
int rc; |
| 227 |
|
| 228 |
/* Clear any already-pending wakeups */ |
| 229 |
ResetLatch(MyLatch); |
| 230 |
|
| 231 |
ProcessMainLoopInterrupts(); |
| 232 |
|
| 233 |
/* |
| 234 |
* Do one cycle of dirty-buffer writing. |
| 235 |
*/ |
| 236 |
can_hibernate = BgBufferSync(&wb_context); |
| 237 |
|
| 238 |
/* Report pending statistics to the cumulative stats system */ |
| 239 |
pgstat_report_bgwriter(); |
| 240 |
pgstat_report_wal(true); |
| 241 |
|
| 242 |
if (FirstCallSinceLastCheckpoint()) |
| 243 |
{ |
| 244 |
/* |
| 245 |
* After any checkpoint, free all smgr objects. Otherwise we |
| 246 |
* would never do so for dropped relations, as the bgwriter does |
| 247 |
* not process shared invalidation messages or call |
| 248 |
* AtEOXact_SMgr(). |
| 249 |
*/ |
| 250 |
smgrdestroyall(); |
| 251 |
} |
| 252 |
|
| 253 |
/* |
| 254 |
* Log a new xl_running_xacts every now and then so replication can |
| 255 |
* get into a consistent state faster (think of suboverflowed |
| 256 |
* snapshots) and clean up resources (locks, KnownXids*) more |
| 257 |
* frequently. The costs of this are relatively low, so doing it 4 |
| 258 |
* times (LOG_SNAPSHOT_INTERVAL_MS) a minute seems fine. |
| 259 |
* |
| 260 |
* We assume the interval for writing xl_running_xacts is |
| 261 |
* significantly bigger than BgWriterDelay, so we don't complicate the |
| 262 |
* overall timeout handling but just assume we're going to get called |
| 263 |
* often enough even if hibernation mode is active. It's not that |
| 264 |
* important that LOG_SNAPSHOT_INTERVAL_MS is met strictly. To make |
| 265 |
* sure we're not waking the disk up unnecessarily on an idle system |
| 266 |
* we check whether there has been any WAL inserted since the last |
| 267 |
* time we've logged a running xacts. |
| 268 |
* |
| 269 |
* We do this logging in the bgwriter as it is the only process that |
| 270 |
* is run regularly and returns to its mainloop all the time. E.g. |
| 271 |
* Checkpointer, when active, is barely ever in its mainloop and thus |
| 272 |
* makes it hard to log regularly. |
| 273 |
*/ |
| 274 |
if (XLogStandbyInfoActive() && !RecoveryInProgress()) |
| 275 |
{ |
| 276 |
TimestampTz timeout = 0; |
| 277 |
TimestampTz now = GetCurrentTimestamp(); |
| 278 |
|
| 279 |
timeout = TimestampTzPlusMilliseconds(last_snapshot_ts, |
| 280 |
LOG_SNAPSHOT_INTERVAL_MS); |
| 281 |
|
| 282 |
/* |
| 283 |
* Only log if enough time has passed and interesting records have |
| 284 |
* been inserted since the last snapshot. Have to compare with <= |
| 285 |
* instead of < because GetLastImportantRecPtr() points at the |
| 286 |
* start of a record, whereas last_snapshot_lsn points just past |
| 287 |
* the end of the record. |
| 288 |
*/ |
| 289 |
if (now >= timeout && |
| 290 |
last_snapshot_lsn <= GetLastImportantRecPtr()) |
| 291 |
{ |
| 292 |
last_snapshot_lsn = LogStandbySnapshot(); |
| 293 |
last_snapshot_ts = now; |
| 294 |
} |
| 295 |
} |
| 296 |
|
| 297 |
/* |
| 298 |
* Sleep until we are signaled or BgWriterDelay has elapsed. |
| 299 |
* |
| 300 |
* Note: the feedback control loop in BgBufferSync() expects that we |
| 301 |
* will call it every BgWriterDelay msec. While it's not critical for |
| 302 |
* correctness that that be exact, the feedback loop might misbehave |
| 303 |
* if we stray too far from that. Hence, avoid loading this process |
| 304 |
* down with latch events that are likely to happen frequently during |
| 305 |
* normal operation. |
| 306 |
*/ |
| 307 |
rc = WaitLatch(MyLatch, |
| 308 |
WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH, |
| 309 |
BgWriterDelay /* ms */ , WAIT_EVENT_BGWRITER_MAIN); |
| 310 |
|
| 311 |
/* |
| 312 |
* If no latch event and BgBufferSync says nothing's happening, extend |
| 313 |
* the sleep in "hibernation" mode, where we sleep for much longer |
| 314 |
* than bgwriter_delay says. Fewer wakeups save electricity. When a |
| 315 |
* backend starts using buffers again, it will wake us up by setting |
| 316 |
* our latch. Because the extra sleep will persist only as long as no |
| 317 |
* buffer allocations happen, this should not distort the behavior of |
| 318 |
* BgBufferSync's control loop too badly; essentially, it will think |
| 319 |
* that the system-wide idle interval didn't exist. |
| 320 |
* |
| 321 |
* There is a race condition here, in that a backend might allocate a |
| 322 |
* buffer between the time BgBufferSync saw the alloc count as zero |
| 323 |
* and the time we call StrategyNotifyBgWriter. While it's not |
| 324 |
* critical that we not hibernate anyway, we try to reduce the odds of |
| 325 |
* that by only hibernating when BgBufferSync says nothing's happening |
| 326 |
* for two consecutive cycles. Also, we mitigate any possible |
| 327 |
* consequences of a missed wakeup by not hibernating forever. |
| 328 |
*/ |
| 329 |
if (rc == WL_TIMEOUT && can_hibernate && prev_hibernate) |
| 330 |
{ |
| 331 |
/* Ask for notification at next buffer allocation */ |
| 332 |
StrategyNotifyBgWriter(MyProcNumber); |
| 333 |
/* Sleep ... */ |
| 334 |
(void) WaitLatch(MyLatch, |
| 335 |
WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH, |
| 336 |
BgWriterDelay * HIBERNATE_FACTOR, |
| 337 |
WAIT_EVENT_BGWRITER_HIBERNATE); |
| 338 |
/* Reset the notification request in case we timed out */ |
| 339 |
StrategyNotifyBgWriter(-1); |
| 340 |
} |
| 341 |
|
| 342 |
prev_hibernate = can_hibernate; |
| 343 |
} |
| 344 |
} |