| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* checkpointer.c |
| 4 |
* |
| 5 |
* The checkpointer is new as of Postgres 9.2. It handles all checkpoints. |
| 6 |
* Checkpoints are automatically dispatched after a certain amount of time has |
| 7 |
* elapsed since the last one, and it can be signaled to perform requested |
| 8 |
* checkpoints as well. (The GUC parameter that mandates a checkpoint every |
| 9 |
* so many WAL segments is implemented by having backends signal when they |
| 10 |
* fill WAL segments; the checkpointer itself doesn't watch for the |
| 11 |
* condition.) |
| 12 |
* |
| 13 |
* The normal termination sequence is that checkpointer is instructed to |
| 14 |
* execute the shutdown checkpoint by SIGINT. After that checkpointer waits |
| 15 |
* to be terminated via SIGUSR2, which instructs the checkpointer to exit(0). |
| 16 |
* All backends must be stopped before SIGINT or SIGUSR2 is issued! |
| 17 |
* |
| 18 |
* Emergency termination is by SIGQUIT; like any backend, the checkpointer |
| 19 |
* will simply abort and exit on SIGQUIT. |
| 20 |
* |
| 21 |
* If the checkpointer exits unexpectedly, the postmaster treats that the same |
| 22 |
* as a backend crash: shared memory may be corrupted, so remaining backends |
| 23 |
* should be killed by SIGQUIT and then a recovery cycle started. (Even if |
| 24 |
* shared memory isn't corrupted, we have lost information about which |
| 25 |
* files need to be fsync'd for the next checkpoint, and so a system |
| 26 |
* restart needs to be forced.) |
| 27 |
* |
| 28 |
* |
| 29 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 30 |
* |
| 31 |
* |
| 32 |
* IDENTIFICATION |
| 33 |
* src/backend/postmaster/checkpointer.c |
| 34 |
* |
| 35 |
*------------------------------------------------------------------------- |
| 36 |
*/ |
| 37 |
#include "postgres.h" |
| 38 |
|
| 39 |
#include <sys/time.h> |
| 40 |
#include <time.h> |
| 41 |
|
| 42 |
#include "access/xlog.h" |
| 43 |
#include "access/xlog_internal.h" |
| 44 |
#include "access/xlogrecovery.h" |
| 45 |
#include "libpq/pqsignal.h" |
| 46 |
#include "miscadmin.h" |
| 47 |
#include "pgstat.h" |
| 48 |
#include "postmaster/auxprocess.h" |
| 49 |
#include "postmaster/bgwriter.h" |
| 50 |
#include "postmaster/interrupt.h" |
| 51 |
#include "replication/syncrep.h" |
| 52 |
#include "storage/aio_subsys.h" |
| 53 |
#include "storage/bufmgr.h" |
| 54 |
#include "storage/condition_variable.h" |
| 55 |
#include "storage/fd.h" |
| 56 |
#include "storage/ipc.h" |
| 57 |
#include "storage/lwlock.h" |
| 58 |
#include "storage/pmsignal.h" |
| 59 |
#include "storage/proc.h" |
| 60 |
#include "storage/procsignal.h" |
| 61 |
#include "storage/shmem.h" |
| 62 |
#include "storage/smgr.h" |
| 63 |
#include "storage/spin.h" |
| 64 |
#include "utils/guc.h" |
| 65 |
#include "utils/memutils.h" |
| 66 |
#include "utils/resowner.h" |
| 67 |
|
| 68 |
|
| 69 |
/*---------- |
| 70 |
* Shared memory area for communication between checkpointer and backends |
| 71 |
* |
| 72 |
* The ckpt counters allow backends to watch for completion of a checkpoint |
| 73 |
* request they send. Here's how it works: |
| 74 |
* * At start of a checkpoint, checkpointer reads (and clears) the request |
| 75 |
* flags and increments ckpt_started, while holding ckpt_lck. |
| 76 |
* * On completion of a checkpoint, checkpointer sets ckpt_done to |
| 77 |
* equal ckpt_started. |
| 78 |
* * On failure of a checkpoint, checkpointer increments ckpt_failed |
| 79 |
* and sets ckpt_done to equal ckpt_started. |
| 80 |
* |
| 81 |
* The algorithm for backends is: |
| 82 |
* 1. Record current values of ckpt_failed and ckpt_started, and |
| 83 |
* set request flags, while holding ckpt_lck. |
| 84 |
* 2. Send signal to request checkpoint. |
| 85 |
* 3. Sleep until ckpt_started changes. Now you know a checkpoint has |
| 86 |
* begun since you started this algorithm (although *not* that it was |
| 87 |
* specifically initiated by your signal), and that it is using your flags. |
| 88 |
* 4. Record new value of ckpt_started. |
| 89 |
* 5. Sleep until ckpt_done >= saved value of ckpt_started. (Use modulo |
| 90 |
* arithmetic here in case counters wrap around.) Now you know a |
| 91 |
* checkpoint has started and completed, but not whether it was |
| 92 |
* successful. |
| 93 |
* 6. If ckpt_failed is different from the originally saved value, |
| 94 |
* assume request failed; otherwise it was definitely successful. |
| 95 |
* |
| 96 |
* ckpt_flags holds the OR of the checkpoint request flags sent by all |
| 97 |
* requesting backends since the last checkpoint start. The flags are |
| 98 |
* chosen so that OR'ing is the correct way to combine multiple requests. |
| 99 |
* |
| 100 |
* The requests array holds fsync requests sent by backends and not yet |
| 101 |
* absorbed by the checkpointer. |
| 102 |
* |
| 103 |
* Unlike the checkpoint fields, requests related fields are protected by |
| 104 |
* CheckpointerCommLock. |
| 105 |
*---------- |
| 106 |
*/ |
| 107 |
typedef struct |
| 108 |
{ |
| 109 |
SyncRequestType type; /* request type */ |
| 110 |
FileTag ftag; /* file identifier */ |
| 111 |
} CheckpointerRequest; |
| 112 |
|
| 113 |
typedef struct |
| 114 |
{ |
| 115 |
pid_t checkpointer_pid; /* PID (0 if not started) */ |
| 116 |
|
| 117 |
slock_t ckpt_lck; /* protects all the ckpt_* fields */ |
| 118 |
|
| 119 |
int ckpt_started; /* advances when checkpoint starts */ |
| 120 |
int ckpt_done; /* advances when checkpoint done */ |
| 121 |
int ckpt_failed; /* advances when checkpoint fails */ |
| 122 |
|
| 123 |
int ckpt_flags; /* checkpoint flags, as defined in xlog.h */ |
| 124 |
|
| 125 |
ConditionVariable start_cv; /* signaled when ckpt_started advances */ |
| 126 |
ConditionVariable done_cv; /* signaled when ckpt_done advances */ |
| 127 |
|
| 128 |
int num_requests; /* current # of requests */ |
| 129 |
int max_requests; /* allocated array size */ |
| 130 |
CheckpointerRequest requests[FLEXIBLE_ARRAY_MEMBER]; |
| 131 |
} CheckpointerShmemStruct; |
| 132 |
|
| 133 |
static CheckpointerShmemStruct *CheckpointerShmem; |
| 134 |
|
| 135 |
/* interval for calling AbsorbSyncRequests in CheckpointWriteDelay */ |
| 136 |
#define WRITES_PER_ABSORB 1000 |
| 137 |
|
| 138 |
/* Max number of requests the checkpointer request queue can hold */ |
| 139 |
#define MAX_CHECKPOINT_REQUESTS 10000000 |
| 140 |
|
| 141 |
/* |
| 142 |
* GUC parameters |
| 143 |
*/ |
| 144 |
int CheckPointTimeout = 300; |
| 145 |
int CheckPointWarning = 30; |
| 146 |
double CheckPointCompletionTarget = 0.9; |
| 147 |
|
| 148 |
/* |
| 149 |
* Private state |
| 150 |
*/ |
| 151 |
static bool ckpt_active = false; |
| 152 |
static volatile sig_atomic_t ShutdownXLOGPending = false; |
| 153 |
|
| 154 |
/* these values are valid when ckpt_active is true: */ |
| 155 |
static pg_time_t ckpt_start_time; |
| 156 |
static XLogRecPtr ckpt_start_recptr; |
| 157 |
static double ckpt_cached_elapsed; |
| 158 |
|
| 159 |
static pg_time_t last_checkpoint_time; |
| 160 |
static pg_time_t last_xlog_switch_time; |
| 161 |
|
| 162 |
/* Prototypes for private functions */ |
| 163 |
|
| 164 |
static void ProcessCheckpointerInterrupts(void); |
| 165 |
static void CheckArchiveTimeout(void); |
| 166 |
static bool IsCheckpointOnSchedule(double progress); |
| 167 |
static bool ImmediateCheckpointRequested(void); |
| 168 |
static bool CompactCheckpointerRequestQueue(void); |
| 169 |
static void UpdateSharedMemoryConfig(void); |
| 170 |
|
| 171 |
/* Signal handlers */ |
| 172 |
static void ReqShutdownXLOG(SIGNAL_ARGS); |
| 173 |
|
| 174 |
|
| 175 |
/* |
| 176 |
* Main entry point for checkpointer process |
| 177 |
* |
| 178 |
* This is invoked from AuxiliaryProcessMain, which has already created the |
| 179 |
* basic execution environment, but not enabled signals yet. |
| 180 |
*/ |
| 181 |
void |
| 182 |
CheckpointerMain(const void *startup_data, size_t startup_data_len) |
| 183 |
{ |
| 184 |
sigjmp_buf local_sigjmp_buf; |
| 185 |
MemoryContext checkpointer_context; |
| 186 |
|
| 187 |
Assert(startup_data_len == 0); |
| 188 |
|
| 189 |
MyBackendType = B_CHECKPOINTER; |
| 190 |
AuxiliaryProcessMainCommon(); |
| 191 |
|
| 192 |
CheckpointerShmem->checkpointer_pid = MyProcPid; |
| 193 |
|
| 194 |
/* |
| 195 |
* Properly accept or ignore signals the postmaster might send us |
| 196 |
* |
| 197 |
* Note: we deliberately ignore SIGTERM, because during a standard Unix |
| 198 |
* system shutdown cycle, init will SIGTERM all processes at once. We |
| 199 |
* want to wait for the backends to exit, whereupon the postmaster will |
| 200 |
* tell us it's okay to shut down (via SIGUSR2). |
| 201 |
*/ |
| 202 |
pqsignal(SIGHUP, SignalHandlerForConfigReload); |
| 203 |
pqsignal(SIGINT, ReqShutdownXLOG); |
| 204 |
pqsignal(SIGTERM, SIG_IGN); /* ignore SIGTERM */ |
| 205 |
/* SIGQUIT handler was already set up by InitPostmasterChild */ |
| 206 |
pqsignal(SIGALRM, SIG_IGN); |
| 207 |
pqsignal(SIGPIPE, SIG_IGN); |
| 208 |
pqsignal(SIGUSR1, procsignal_sigusr1_handler); |
| 209 |
pqsignal(SIGUSR2, SignalHandlerForShutdownRequest); |
| 210 |
|
| 211 |
/* |
| 212 |
* Reset some signals that are accepted by postmaster but not here |
| 213 |
*/ |
| 214 |
pqsignal(SIGCHLD, SIG_DFL); |
| 215 |
|
| 216 |
/* |
| 217 |
* Initialize so that first time-driven event happens at the correct time. |
| 218 |
*/ |
| 219 |
last_checkpoint_time = last_xlog_switch_time = (pg_time_t) time(NULL); |
| 220 |
|
| 221 |
/* |
| 222 |
* Write out stats after shutdown. This needs to be called by exactly one |
| 223 |
* process during a normal shutdown, and since checkpointer is shut down |
| 224 |
* very late... |
| 225 |
* |
| 226 |
* While e.g. walsenders are active after the shutdown checkpoint has been |
| 227 |
* written (and thus could produce more stats), checkpointer stays around |
| 228 |
* after the shutdown checkpoint has been written. postmaster will only |
| 229 |
* signal checkpointer to exit after all processes that could emit stats |
| 230 |
* have been shut down. |
| 231 |
*/ |
| 232 |
before_shmem_exit(pgstat_before_server_shutdown, 0); |
| 233 |
|
| 234 |
/* |
| 235 |
* Create a memory context that we will do all our work in. We do this so |
| 236 |
* that we can reset the context during error recovery and thereby avoid |
| 237 |
* possible memory leaks. Formerly this code just ran in |
| 238 |
* TopMemoryContext, but resetting that would be a really bad idea. |
| 239 |
*/ |
| 240 |
checkpointer_context = AllocSetContextCreate(TopMemoryContext, |
| 241 |
"Checkpointer", |
| 242 |
ALLOCSET_DEFAULT_SIZES); |
| 243 |
MemoryContextSwitchTo(checkpointer_context); |
| 244 |
|
| 245 |
/* |
| 246 |
* If an exception is encountered, processing resumes here. |
| 247 |
* |
| 248 |
* You might wonder why this isn't coded as an infinite loop around a |
| 249 |
* PG_TRY construct. The reason is that this is the bottom of the |
| 250 |
* exception stack, and so with PG_TRY there would be no exception handler |
| 251 |
* in force at all during the CATCH part. By leaving the outermost setjmp |
| 252 |
* always active, we have at least some chance of recovering from an error |
| 253 |
* during error recovery. (If we get into an infinite loop thereby, it |
| 254 |
* will soon be stopped by overflow of elog.c's internal state stack.) |
| 255 |
* |
| 256 |
* Note that we use sigsetjmp(..., 1), so that the prevailing signal mask |
| 257 |
* (to wit, BlockSig) will be restored when longjmp'ing to here. Thus, |
| 258 |
* signals other than SIGQUIT will be blocked until we complete error |
| 259 |
* recovery. It might seem that this policy makes the HOLD_INTERRUPTS() |
| 260 |
* call redundant, but it is not since InterruptPending might be set |
| 261 |
* already. |
| 262 |
*/ |
| 263 |
if (sigsetjmp(local_sigjmp_buf, 1) != 0) |
| 264 |
{ |
| 265 |
/* Since not using PG_TRY, must reset error stack by hand */ |
| 266 |
error_context_stack = NULL; |
| 267 |
|
| 268 |
/* Prevent interrupts while cleaning up */ |
| 269 |
HOLD_INTERRUPTS(); |
| 270 |
|
| 271 |
/* Report the error to the server log */ |
| 272 |
EmitErrorReport(); |
| 273 |
|
| 274 |
/* |
| 275 |
* These operations are really just a minimal subset of |
| 276 |
* AbortTransaction(). We don't have very many resources to worry |
| 277 |
* about in checkpointer, but we do have LWLocks, buffers, and temp |
| 278 |
* files. |
| 279 |
*/ |
| 280 |
LWLockReleaseAll(); |
| 281 |
ConditionVariableCancelSleep(); |
| 282 |
pgstat_report_wait_end(); |
| 283 |
pgaio_error_cleanup(); |
| 284 |
UnlockBuffers(); |
| 285 |
ReleaseAuxProcessResources(false); |
| 286 |
AtEOXact_Buffers(false); |
| 287 |
AtEOXact_SMgr(); |
| 288 |
AtEOXact_Files(false); |
| 289 |
AtEOXact_HashTables(false); |
| 290 |
|
| 291 |
/* Warn any waiting backends that the checkpoint failed. */ |
| 292 |
if (ckpt_active) |
| 293 |
{ |
| 294 |
SpinLockAcquire(&CheckpointerShmem->ckpt_lck); |
| 295 |
CheckpointerShmem->ckpt_failed++; |
| 296 |
CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started; |
| 297 |
SpinLockRelease(&CheckpointerShmem->ckpt_lck); |
| 298 |
|
| 299 |
ConditionVariableBroadcast(&CheckpointerShmem->done_cv); |
| 300 |
|
| 301 |
ckpt_active = false; |
| 302 |
} |
| 303 |
|
| 304 |
/* |
| 305 |
* Now return to normal top-level context and clear ErrorContext for |
| 306 |
* next time. |
| 307 |
*/ |
| 308 |
MemoryContextSwitchTo(checkpointer_context); |
| 309 |
FlushErrorState(); |
| 310 |
|
| 311 |
/* Flush any leaked data in the top-level context */ |
| 312 |
MemoryContextReset(checkpointer_context); |
| 313 |
|
| 314 |
/* Now we can allow interrupts again */ |
| 315 |
RESUME_INTERRUPTS(); |
| 316 |
|
| 317 |
/* |
| 318 |
* Sleep at least 1 second after any error. A write error is likely |
| 319 |
* to be repeated, and we don't want to be filling the error logs as |
| 320 |
* fast as we can. |
| 321 |
*/ |
| 322 |
pg_usleep(1000000L); |
| 323 |
} |
| 324 |
|
| 325 |
/* We can now handle ereport(ERROR) */ |
| 326 |
PG_exception_stack = &local_sigjmp_buf; |
| 327 |
|
| 328 |
/* |
| 329 |
* Unblock signals (they were blocked when the postmaster forked us) |
| 330 |
*/ |
| 331 |
sigprocmask(SIG_SETMASK, &UnBlockSig, NULL); |
| 332 |
|
| 333 |
/* |
| 334 |
* Ensure all shared memory values are set correctly for the config. Doing |
| 335 |
* this here ensures no race conditions from other concurrent updaters. |
| 336 |
*/ |
| 337 |
UpdateSharedMemoryConfig(); |
| 338 |
|
| 339 |
/* |
| 340 |
* Advertise our proc number that backends can use to wake us up while |
| 341 |
* we're sleeping. |
| 342 |
*/ |
| 343 |
ProcGlobal->checkpointerProc = MyProcNumber; |
| 344 |
|
| 345 |
/* |
| 346 |
* Loop until we've been asked to write the shutdown checkpoint or |
| 347 |
* terminate. |
| 348 |
*/ |
| 349 |
for (;;) |
| 350 |
{ |
| 351 |
bool do_checkpoint = false; |
| 352 |
int flags = 0; |
| 353 |
pg_time_t now; |
| 354 |
int elapsed_secs; |
| 355 |
int cur_timeout; |
| 356 |
bool chkpt_or_rstpt_requested = false; |
| 357 |
bool chkpt_or_rstpt_timed = false; |
| 358 |
|
| 359 |
/* Clear any already-pending wakeups */ |
| 360 |
ResetLatch(MyLatch); |
| 361 |
|
| 362 |
/* |
| 363 |
* Process any requests or signals received recently. |
| 364 |
*/ |
| 365 |
AbsorbSyncRequests(); |
| 366 |
|
| 367 |
ProcessCheckpointerInterrupts(); |
| 368 |
if (ShutdownXLOGPending || ShutdownRequestPending) |
| 369 |
break; |
| 370 |
|
| 371 |
/* |
| 372 |
* Detect a pending checkpoint request by checking whether the flags |
| 373 |
* word in shared memory is nonzero. We shouldn't need to acquire the |
| 374 |
* ckpt_lck for this. |
| 375 |
*/ |
| 376 |
if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags) |
| 377 |
{ |
| 378 |
do_checkpoint = true; |
| 379 |
chkpt_or_rstpt_requested = true; |
| 380 |
} |
| 381 |
|
| 382 |
/* |
| 383 |
* Force a checkpoint if too much time has elapsed since the last one. |
| 384 |
* Note that we count a timed checkpoint in stats only when this |
| 385 |
* occurs without an external request, but we set the CAUSE_TIME flag |
| 386 |
* bit even if there is also an external request. |
| 387 |
*/ |
| 388 |
now = (pg_time_t) time(NULL); |
| 389 |
elapsed_secs = now - last_checkpoint_time; |
| 390 |
if (elapsed_secs >= CheckPointTimeout) |
| 391 |
{ |
| 392 |
if (!do_checkpoint) |
| 393 |
chkpt_or_rstpt_timed = true; |
| 394 |
do_checkpoint = true; |
| 395 |
flags |= CHECKPOINT_CAUSE_TIME; |
| 396 |
} |
| 397 |
|
| 398 |
/* |
| 399 |
* Do a checkpoint if requested. |
| 400 |
*/ |
| 401 |
if (do_checkpoint) |
| 402 |
{ |
| 403 |
bool ckpt_performed = false; |
| 404 |
bool do_restartpoint; |
| 405 |
|
| 406 |
/* Check if we should perform a checkpoint or a restartpoint. */ |
| 407 |
do_restartpoint = RecoveryInProgress(); |
| 408 |
|
| 409 |
/* |
| 410 |
* Atomically fetch the request flags to figure out what kind of a |
| 411 |
* checkpoint we should perform, and increase the started-counter |
| 412 |
* to acknowledge that we've started a new checkpoint. |
| 413 |
*/ |
| 414 |
SpinLockAcquire(&CheckpointerShmem->ckpt_lck); |
| 415 |
flags |= CheckpointerShmem->ckpt_flags; |
| 416 |
CheckpointerShmem->ckpt_flags = 0; |
| 417 |
CheckpointerShmem->ckpt_started++; |
| 418 |
SpinLockRelease(&CheckpointerShmem->ckpt_lck); |
| 419 |
|
| 420 |
ConditionVariableBroadcast(&CheckpointerShmem->start_cv); |
| 421 |
|
| 422 |
/* |
| 423 |
* The end-of-recovery checkpoint is a real checkpoint that's |
| 424 |
* performed while we're still in recovery. |
| 425 |
*/ |
| 426 |
if (flags & CHECKPOINT_END_OF_RECOVERY) |
| 427 |
do_restartpoint = false; |
| 428 |
|
| 429 |
if (chkpt_or_rstpt_timed) |
| 430 |
{ |
| 431 |
chkpt_or_rstpt_timed = false; |
| 432 |
if (do_restartpoint) |
| 433 |
PendingCheckpointerStats.restartpoints_timed++; |
| 434 |
else |
| 435 |
PendingCheckpointerStats.num_timed++; |
| 436 |
} |
| 437 |
|
| 438 |
if (chkpt_or_rstpt_requested) |
| 439 |
{ |
| 440 |
chkpt_or_rstpt_requested = false; |
| 441 |
if (do_restartpoint) |
| 442 |
PendingCheckpointerStats.restartpoints_requested++; |
| 443 |
else |
| 444 |
PendingCheckpointerStats.num_requested++; |
| 445 |
} |
| 446 |
|
| 447 |
/* |
| 448 |
* We will warn if (a) too soon since last checkpoint (whatever |
| 449 |
* caused it) and (b) somebody set the CHECKPOINT_CAUSE_XLOG flag |
| 450 |
* since the last checkpoint start. Note in particular that this |
| 451 |
* implementation will not generate warnings caused by |
| 452 |
* CheckPointTimeout < CheckPointWarning. |
| 453 |
*/ |
| 454 |
if (!do_restartpoint && |
| 455 |
(flags & CHECKPOINT_CAUSE_XLOG) && |
| 456 |
elapsed_secs < CheckPointWarning) |
| 457 |
ereport(LOG, |
| 458 |
(errmsg_plural("checkpoints are occurring too frequently (%d second apart)", |
| 459 |
"checkpoints are occurring too frequently (%d seconds apart)", |
| 460 |
elapsed_secs, |
| 461 |
elapsed_secs), |
| 462 |
errhint("Consider increasing the configuration parameter \"%s\".", "max_wal_size"))); |
| 463 |
|
| 464 |
/* |
| 465 |
* Initialize checkpointer-private variables used during |
| 466 |
* checkpoint. |
| 467 |
*/ |
| 468 |
ckpt_active = true; |
| 469 |
if (do_restartpoint) |
| 470 |
ckpt_start_recptr = GetXLogReplayRecPtr(NULL); |
| 471 |
else |
| 472 |
ckpt_start_recptr = GetInsertRecPtr(); |
| 473 |
ckpt_start_time = now; |
| 474 |
ckpt_cached_elapsed = 0; |
| 475 |
|
| 476 |
/* |
| 477 |
* Do the checkpoint. |
| 478 |
*/ |
| 479 |
if (!do_restartpoint) |
| 480 |
ckpt_performed = CreateCheckPoint(flags); |
| 481 |
else |
| 482 |
ckpt_performed = CreateRestartPoint(flags); |
| 483 |
|
| 484 |
/* |
| 485 |
* After any checkpoint, free all smgr objects. Otherwise we |
| 486 |
* would never do so for dropped relations, as the checkpointer |
| 487 |
* does not process shared invalidation messages or call |
| 488 |
* AtEOXact_SMgr(). |
| 489 |
*/ |
| 490 |
smgrdestroyall(); |
| 491 |
|
| 492 |
/* |
| 493 |
* Indicate checkpoint completion to any waiting backends. |
| 494 |
*/ |
| 495 |
SpinLockAcquire(&CheckpointerShmem->ckpt_lck); |
| 496 |
CheckpointerShmem->ckpt_done = CheckpointerShmem->ckpt_started; |
| 497 |
SpinLockRelease(&CheckpointerShmem->ckpt_lck); |
| 498 |
|
| 499 |
ConditionVariableBroadcast(&CheckpointerShmem->done_cv); |
| 500 |
|
| 501 |
if (!do_restartpoint) |
| 502 |
{ |
| 503 |
/* |
| 504 |
* Note we record the checkpoint start time not end time as |
| 505 |
* last_checkpoint_time. This is so that time-driven |
| 506 |
* checkpoints happen at a predictable spacing. |
| 507 |
*/ |
| 508 |
last_checkpoint_time = now; |
| 509 |
|
| 510 |
if (ckpt_performed) |
| 511 |
PendingCheckpointerStats.num_performed++; |
| 512 |
} |
| 513 |
else |
| 514 |
{ |
| 515 |
if (ckpt_performed) |
| 516 |
{ |
| 517 |
/* |
| 518 |
* The same as for checkpoint. Please see the |
| 519 |
* corresponding comment. |
| 520 |
*/ |
| 521 |
last_checkpoint_time = now; |
| 522 |
|
| 523 |
PendingCheckpointerStats.restartpoints_performed++; |
| 524 |
} |
| 525 |
else |
| 526 |
{ |
| 527 |
/* |
| 528 |
* We were not able to perform the restartpoint |
| 529 |
* (checkpoints throw an ERROR in case of error). Most |
| 530 |
* likely because we have not received any new checkpoint |
| 531 |
* WAL records since the last restartpoint. Try again in |
| 532 |
* 15 s. |
| 533 |
*/ |
| 534 |
last_checkpoint_time = now - CheckPointTimeout + 15; |
| 535 |
} |
| 536 |
} |
| 537 |
|
| 538 |
ckpt_active = false; |
| 539 |
|
| 540 |
/* |
| 541 |
* We may have received an interrupt during the checkpoint and the |
| 542 |
* latch might have been reset (e.g. in CheckpointWriteDelay). |
| 543 |
*/ |
| 544 |
ProcessCheckpointerInterrupts(); |
| 545 |
if (ShutdownXLOGPending || ShutdownRequestPending) |
| 546 |
break; |
| 547 |
} |
| 548 |
|
| 549 |
/* Check for archive_timeout and switch xlog files if necessary. */ |
| 550 |
CheckArchiveTimeout(); |
| 551 |
|
| 552 |
/* Report pending statistics to the cumulative stats system */ |
| 553 |
pgstat_report_checkpointer(); |
| 554 |
pgstat_report_wal(true); |
| 555 |
|
| 556 |
/* |
| 557 |
* If any checkpoint flags have been set, redo the loop to handle the |
| 558 |
* checkpoint without sleeping. |
| 559 |
*/ |
| 560 |
if (((volatile CheckpointerShmemStruct *) CheckpointerShmem)->ckpt_flags) |
| 561 |
continue; |
| 562 |
|
| 563 |
/* |
| 564 |
* Sleep until we are signaled or it's time for another checkpoint or |
| 565 |
* xlog file switch. |
| 566 |
*/ |
| 567 |
now = (pg_time_t) time(NULL); |
| 568 |
elapsed_secs = now - last_checkpoint_time; |
| 569 |
if (elapsed_secs >= CheckPointTimeout) |
| 570 |
continue; /* no sleep for us ... */ |
| 571 |
cur_timeout = CheckPointTimeout - elapsed_secs; |
| 572 |
if (XLogArchiveTimeout > 0 && !RecoveryInProgress()) |
| 573 |
{ |
| 574 |
elapsed_secs = now - last_xlog_switch_time; |
| 575 |
if (elapsed_secs >= XLogArchiveTimeout) |
| 576 |
continue; /* no sleep for us ... */ |
| 577 |
cur_timeout = Min(cur_timeout, XLogArchiveTimeout - elapsed_secs); |
| 578 |
} |
| 579 |
|
| 580 |
(void) WaitLatch(MyLatch, |
| 581 |
WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH, |
| 582 |
cur_timeout * 1000L /* convert to ms */ , |
| 583 |
WAIT_EVENT_CHECKPOINTER_MAIN); |
| 584 |
} |
| 585 |
|
| 586 |
/* |
| 587 |
* From here on, elog(ERROR) should end with exit(1), not send control |
| 588 |
* back to the sigsetjmp block above. |
| 589 |
*/ |
| 590 |
ExitOnAnyError = true; |
| 591 |
|
| 592 |
if (ShutdownXLOGPending) |
| 593 |
{ |
| 594 |
/* |
| 595 |
* Close down the database. |
| 596 |
* |
| 597 |
* Since ShutdownXLOG() creates restartpoint or checkpoint, and |
| 598 |
* updates the statistics, increment the checkpoint request and flush |
| 599 |
* out pending statistic. |
| 600 |
*/ |
| 601 |
PendingCheckpointerStats.num_requested++; |
| 602 |
ShutdownXLOG(0, 0); |
| 603 |
pgstat_report_checkpointer(); |
| 604 |
pgstat_report_wal(true); |
| 605 |
|
| 606 |
/* |
| 607 |
* Tell postmaster that we're done. |
| 608 |
*/ |
| 609 |
SendPostmasterSignal(PMSIGNAL_XLOG_IS_SHUTDOWN); |
| 610 |
ShutdownXLOGPending = false; |
| 611 |
} |
| 612 |
|
| 613 |
/* |
| 614 |
* Wait until we're asked to shut down. By separating the writing of the |
| 615 |
* shutdown checkpoint from checkpointer exiting, checkpointer can perform |
| 616 |
* some should-be-as-late-as-possible work like writing out stats. |
| 617 |
*/ |
| 618 |
for (;;) |
| 619 |
{ |
| 620 |
/* Clear any already-pending wakeups */ |
| 621 |
ResetLatch(MyLatch); |
| 622 |
|
| 623 |
ProcessCheckpointerInterrupts(); |
| 624 |
|
| 625 |
if (ShutdownRequestPending) |
| 626 |
break; |
| 627 |
|
| 628 |
(void) WaitLatch(MyLatch, |
| 629 |
WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, |
| 630 |
0, |
| 631 |
WAIT_EVENT_CHECKPOINTER_SHUTDOWN); |
| 632 |
} |
| 633 |
|
| 634 |
/* Normal exit from the checkpointer is here */ |
| 635 |
proc_exit(0); /* done */ |
| 636 |
} |
| 637 |
|
| 638 |
/* |
| 639 |
* Process any new interrupts. |
| 640 |
*/ |
| 641 |
static void |
| 642 |
ProcessCheckpointerInterrupts(void) |
| 643 |
{ |
| 644 |
if (ProcSignalBarrierPending) |
| 645 |
ProcessProcSignalBarrier(); |
| 646 |
|
| 647 |
if (ConfigReloadPending) |
| 648 |
{ |
| 649 |
ConfigReloadPending = false; |
| 650 |
ProcessConfigFile(PGC_SIGHUP); |
| 651 |
|
| 652 |
/* |
| 653 |
* Checkpointer is the last process to shut down, so we ask it to hold |
| 654 |
* the keys for a range of other tasks required most of which have |
| 655 |
* nothing to do with checkpointing at all. |
| 656 |
* |
| 657 |
* For various reasons, some config values can change dynamically so |
| 658 |
* the primary copy of them is held in shared memory to make sure all |
| 659 |
* backends see the same value. We make Checkpointer responsible for |
| 660 |
* updating the shared memory copy if the parameter setting changes |
| 661 |
* because of SIGHUP. |
| 662 |
*/ |
| 663 |
UpdateSharedMemoryConfig(); |
| 664 |
} |
| 665 |
|
| 666 |
/* Perform logging of memory contexts of this process */ |
| 667 |
if (LogMemoryContextPending) |
| 668 |
ProcessLogMemoryContextInterrupt(); |
| 669 |
} |
| 670 |
|
| 671 |
/* |
| 672 |
* CheckArchiveTimeout -- check for archive_timeout and switch xlog files |
| 673 |
* |
| 674 |
* This will switch to a new WAL file and force an archive file write if |
| 675 |
* meaningful activity is recorded in the current WAL file. This includes most |
| 676 |
* writes, including just a single checkpoint record, but excludes WAL records |
| 677 |
* that were inserted with the XLOG_MARK_UNIMPORTANT flag being set (like |
| 678 |
* snapshots of running transactions). Such records, depending on |
| 679 |
* configuration, occur on regular intervals and don't contain important |
| 680 |
* information. This avoids generating archives with a few unimportant |
| 681 |
* records. |
| 682 |
*/ |
| 683 |
static void |
| 684 |
CheckArchiveTimeout(void) |
| 685 |
{ |
| 686 |
pg_time_t now; |
| 687 |
pg_time_t last_time; |
| 688 |
XLogRecPtr last_switch_lsn; |
| 689 |
|
| 690 |
if (XLogArchiveTimeout <= 0 || RecoveryInProgress()) |
| 691 |
return; |
| 692 |
|
| 693 |
now = (pg_time_t) time(NULL); |
| 694 |
|
| 695 |
/* First we do a quick check using possibly-stale local state. */ |
| 696 |
if ((int) (now - last_xlog_switch_time) < XLogArchiveTimeout) |
| 697 |
return; |
| 698 |
|
| 699 |
/* |
| 700 |
* Update local state ... note that last_xlog_switch_time is the last time |
| 701 |
* a switch was performed *or requested*. |
| 702 |
*/ |
| 703 |
last_time = GetLastSegSwitchData(&last_switch_lsn); |
| 704 |
|
| 705 |
last_xlog_switch_time = Max(last_xlog_switch_time, last_time); |
| 706 |
|
| 707 |
/* Now we can do the real checks */ |
| 708 |
if ((int) (now - last_xlog_switch_time) >= XLogArchiveTimeout) |
| 709 |
{ |
| 710 |
/* |
| 711 |
* Switch segment only when "important" WAL has been logged since the |
| 712 |
* last segment switch (last_switch_lsn points to end of segment |
| 713 |
* switch occurred in). |
| 714 |
*/ |
| 715 |
if (GetLastImportantRecPtr() > last_switch_lsn) |
| 716 |
{ |
| 717 |
XLogRecPtr switchpoint; |
| 718 |
|
| 719 |
/* mark switch as unimportant, avoids triggering checkpoints */ |
| 720 |
switchpoint = RequestXLogSwitch(true); |
| 721 |
|
| 722 |
/* |
| 723 |
* If the returned pointer points exactly to a segment boundary, |
| 724 |
* assume nothing happened. |
| 725 |
*/ |
| 726 |
if (XLogSegmentOffset(switchpoint, wal_segment_size) != 0) |
| 727 |
elog(DEBUG1, "write-ahead log switch forced (\"archive_timeout\"=%d)", |
| 728 |
XLogArchiveTimeout); |
| 729 |
} |
| 730 |
|
| 731 |
/* |
| 732 |
* Update state in any case, so we don't retry constantly when the |
| 733 |
* system is idle. |
| 734 |
*/ |
| 735 |
last_xlog_switch_time = now; |
| 736 |
} |
| 737 |
} |
| 738 |
|
| 739 |
/* |
| 740 |
* Returns true if an immediate checkpoint request is pending. (Note that |
| 741 |
* this does not check the *current* checkpoint's IMMEDIATE flag, but whether |
| 742 |
* there is one pending behind it.) |
| 743 |
*/ |
| 744 |
static bool |
| 745 |
ImmediateCheckpointRequested(void) |
| 746 |
{ |
| 747 |
volatile CheckpointerShmemStruct *cps = CheckpointerShmem; |
| 748 |
|
| 749 |
/* |
| 750 |
* We don't need to acquire the ckpt_lck in this case because we're only |
| 751 |
* looking at a single flag bit. |
| 752 |
*/ |
| 753 |
if (cps->ckpt_flags & CHECKPOINT_IMMEDIATE) |
| 754 |
return true; |
| 755 |
return false; |
| 756 |
} |
| 757 |
|
| 758 |
/* |
| 759 |
* CheckpointWriteDelay -- control rate of checkpoint |
| 760 |
* |
| 761 |
* This function is called after each page write performed by BufferSync(). |
| 762 |
* It is responsible for throttling BufferSync()'s write rate to hit |
| 763 |
* checkpoint_completion_target. |
| 764 |
* |
| 765 |
* The checkpoint request flags should be passed in; currently the only one |
| 766 |
* examined is CHECKPOINT_IMMEDIATE, which disables delays between writes. |
| 767 |
* |
| 768 |
* 'progress' is an estimate of how much of the work has been done, as a |
| 769 |
* fraction between 0.0 meaning none, and 1.0 meaning all done. |
| 770 |
*/ |
| 771 |
void |
| 772 |
CheckpointWriteDelay(int flags, double progress) |
| 773 |
{ |
| 774 |
static int absorb_counter = WRITES_PER_ABSORB; |
| 775 |
|
| 776 |
/* Do nothing if checkpoint is being executed by non-checkpointer process */ |
| 777 |
if (!AmCheckpointerProcess()) |
| 778 |
return; |
| 779 |
|
| 780 |
/* |
| 781 |
* Perform the usual duties and take a nap, unless we're behind schedule, |
| 782 |
* in which case we just try to catch up as quickly as possible. |
| 783 |
*/ |
| 784 |
if (!(flags & CHECKPOINT_IMMEDIATE) && |
| 785 |
!ShutdownXLOGPending && |
| 786 |
!ShutdownRequestPending && |
| 787 |
!ImmediateCheckpointRequested() && |
| 788 |
IsCheckpointOnSchedule(progress)) |
| 789 |
{ |
| 790 |
if (ConfigReloadPending) |
| 791 |
{ |
| 792 |
ConfigReloadPending = false; |
| 793 |
ProcessConfigFile(PGC_SIGHUP); |
| 794 |
/* update shmem copies of config variables */ |
| 795 |
UpdateSharedMemoryConfig(); |
| 796 |
} |
| 797 |
|
| 798 |
AbsorbSyncRequests(); |
| 799 |
absorb_counter = WRITES_PER_ABSORB; |
| 800 |
|
| 801 |
CheckArchiveTimeout(); |
| 802 |
|
| 803 |
/* Report interim statistics to the cumulative stats system */ |
| 804 |
pgstat_report_checkpointer(); |
| 805 |
|
| 806 |
/* |
| 807 |
* This sleep used to be connected to bgwriter_delay, typically 200ms. |
| 808 |
* That resulted in more frequent wakeups if not much work to do. |
| 809 |
* Checkpointer and bgwriter are no longer related so take the Big |
| 810 |
* Sleep. |
| 811 |
*/ |
| 812 |
WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH | WL_TIMEOUT, |
| 813 |
100, |
| 814 |
WAIT_EVENT_CHECKPOINT_WRITE_DELAY); |
| 815 |
ResetLatch(MyLatch); |
| 816 |
} |
| 817 |
else if (--absorb_counter <= 0) |
| 818 |
{ |
| 819 |
/* |
| 820 |
* Absorb pending fsync requests after each WRITES_PER_ABSORB write |
| 821 |
* operations even when we don't sleep, to prevent overflow of the |
| 822 |
* fsync request queue. |
| 823 |
*/ |
| 824 |
AbsorbSyncRequests(); |
| 825 |
absorb_counter = WRITES_PER_ABSORB; |
| 826 |
} |
| 827 |
|
| 828 |
/* Check for barrier events. */ |
| 829 |
if (ProcSignalBarrierPending) |
| 830 |
ProcessProcSignalBarrier(); |
| 831 |
} |
| 832 |
|
| 833 |
/* |
| 834 |
* IsCheckpointOnSchedule -- are we on schedule to finish this checkpoint |
| 835 |
* (or restartpoint) in time? |
| 836 |
* |
| 837 |
* Compares the current progress against the time/segments elapsed since last |
| 838 |
* checkpoint, and returns true if the progress we've made this far is greater |
| 839 |
* than the elapsed time/segments. |
| 840 |
*/ |
| 841 |
static bool |
| 842 |
IsCheckpointOnSchedule(double progress) |
| 843 |
{ |
| 844 |
XLogRecPtr recptr; |
| 845 |
struct timeval now; |
| 846 |
double elapsed_xlogs, |
| 847 |
elapsed_time; |
| 848 |
|
| 849 |
Assert(ckpt_active); |
| 850 |
|
| 851 |
/* Scale progress according to checkpoint_completion_target. */ |
| 852 |
progress *= CheckPointCompletionTarget; |
| 853 |
|
| 854 |
/* |
| 855 |
* Check against the cached value first. Only do the more expensive |
| 856 |
* calculations once we reach the target previously calculated. Since |
| 857 |
* neither time or WAL insert pointer moves backwards, a freshly |
| 858 |
* calculated value can only be greater than or equal to the cached value. |
| 859 |
*/ |
| 860 |
if (progress < ckpt_cached_elapsed) |
| 861 |
return false; |
| 862 |
|
| 863 |
/* |
| 864 |
* Check progress against WAL segments written and CheckPointSegments. |
| 865 |
* |
| 866 |
* We compare the current WAL insert location against the location |
| 867 |
* computed before calling CreateCheckPoint. The code in XLogInsert that |
| 868 |
* actually triggers a checkpoint when CheckPointSegments is exceeded |
| 869 |
* compares against RedoRecPtr, so this is not completely accurate. |
| 870 |
* However, it's good enough for our purposes, we're only calculating an |
| 871 |
* estimate anyway. |
| 872 |
* |
| 873 |
* During recovery, we compare last replayed WAL record's location with |
| 874 |
* the location computed before calling CreateRestartPoint. That maintains |
| 875 |
* the same pacing as we have during checkpoints in normal operation, but |
| 876 |
* we might exceed max_wal_size by a fair amount. That's because there can |
| 877 |
* be a large gap between a checkpoint's redo-pointer and the checkpoint |
| 878 |
* record itself, and we only start the restartpoint after we've seen the |
| 879 |
* checkpoint record. (The gap is typically up to CheckPointSegments * |
| 880 |
* checkpoint_completion_target where checkpoint_completion_target is the |
| 881 |
* value that was in effect when the WAL was generated). |
| 882 |
*/ |
| 883 |
if (RecoveryInProgress()) |
| 884 |
recptr = GetXLogReplayRecPtr(NULL); |
| 885 |
else |
| 886 |
recptr = GetInsertRecPtr(); |
| 887 |
elapsed_xlogs = (((double) (recptr - ckpt_start_recptr)) / |
| 888 |
wal_segment_size) / CheckPointSegments; |
| 889 |
|
| 890 |
if (progress < elapsed_xlogs) |
| 891 |
{ |
| 892 |
ckpt_cached_elapsed = elapsed_xlogs; |
| 893 |
return false; |
| 894 |
} |
| 895 |
|
| 896 |
/* |
| 897 |
* Check progress against time elapsed and checkpoint_timeout. |
| 898 |
*/ |
| 899 |
gettimeofday(&now, NULL); |
| 900 |
elapsed_time = ((double) ((pg_time_t) now.tv_sec - ckpt_start_time) + |
| 901 |
now.tv_usec / 1000000.0) / CheckPointTimeout; |
| 902 |
|
| 903 |
if (progress < elapsed_time) |
| 904 |
{ |
| 905 |
ckpt_cached_elapsed = elapsed_time; |
| 906 |
return false; |
| 907 |
} |
| 908 |
|
| 909 |
/* It looks like we're on schedule. */ |
| 910 |
return true; |
| 911 |
} |
| 912 |
|
| 913 |
|
| 914 |
/* -------------------------------- |
| 915 |
* signal handler routines |
| 916 |
* -------------------------------- |
| 917 |
*/ |
| 918 |
|
| 919 |
/* SIGINT: set flag to trigger writing of shutdown checkpoint */ |
| 920 |
static void |
| 921 |
ReqShutdownXLOG(SIGNAL_ARGS) |
| 922 |
{ |
| 923 |
ShutdownXLOGPending = true; |
| 924 |
SetLatch(MyLatch); |
| 925 |
} |
| 926 |
|
| 927 |
|
| 928 |
/* -------------------------------- |
| 929 |
* communication with backends |
| 930 |
* -------------------------------- |
| 931 |
*/ |
| 932 |
|
| 933 |
/* |
| 934 |
* CheckpointerShmemSize |
| 935 |
* Compute space needed for checkpointer-related shared memory |
| 936 |
*/ |
| 937 |
Size |
| 938 |
CheckpointerShmemSize(void) |
| 939 |
{ |
| 940 |
Size size; |
| 941 |
|
| 942 |
/* |
| 943 |
* The size of the requests[] array is arbitrarily set equal to NBuffers. |
| 944 |
* But there is a cap of MAX_CHECKPOINT_REQUESTS to prevent accumulating |
| 945 |
* too many checkpoint requests in the ring buffer. |
| 946 |
*/ |
| 947 |
size = offsetof(CheckpointerShmemStruct, requests); |
| 948 |
size = add_size(size, mul_size(Min(NBuffers, |
| 949 |
MAX_CHECKPOINT_REQUESTS), |
| 950 |
sizeof(CheckpointerRequest))); |
| 951 |
|
| 952 |
return size; |
| 953 |
} |
| 954 |
|
| 955 |
/* |
| 956 |
* CheckpointerShmemInit |
| 957 |
* Allocate and initialize checkpointer-related shared memory |
| 958 |
*/ |
| 959 |
void |
| 960 |
CheckpointerShmemInit(void) |
| 961 |
{ |
| 962 |
Size size = CheckpointerShmemSize(); |
| 963 |
bool found; |
| 964 |
|
| 965 |
CheckpointerShmem = (CheckpointerShmemStruct *) |
| 966 |
ShmemInitStruct("Checkpointer Data", |
| 967 |
size, |
| 968 |
&found); |
| 969 |
|
| 970 |
if (!found) |
| 971 |
{ |
| 972 |
/* |
| 973 |
* First time through, so initialize. Note that we zero the whole |
| 974 |
* requests array; this is so that CompactCheckpointerRequestQueue can |
| 975 |
* assume that any pad bytes in the request structs are zeroes. |
| 976 |
*/ |
| 977 |
MemSet(CheckpointerShmem, 0, size); |
| 978 |
SpinLockInit(&CheckpointerShmem->ckpt_lck); |
| 979 |
CheckpointerShmem->max_requests = Min(NBuffers, MAX_CHECKPOINT_REQUESTS); |
| 980 |
ConditionVariableInit(&CheckpointerShmem->start_cv); |
| 981 |
ConditionVariableInit(&CheckpointerShmem->done_cv); |
| 982 |
} |
| 983 |
} |
| 984 |
|
| 985 |
/* |
| 986 |
* RequestCheckpoint |
| 987 |
* Called in backend processes to request a checkpoint |
| 988 |
* |
| 989 |
* flags is a bitwise OR of the following: |
| 990 |
* CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown. |
| 991 |
* CHECKPOINT_END_OF_RECOVERY: checkpoint is for end of WAL recovery. |
| 992 |
* CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP, |
| 993 |
* ignoring checkpoint_completion_target parameter. |
| 994 |
* CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occurred |
| 995 |
* since the last one (implied by CHECKPOINT_IS_SHUTDOWN or |
| 996 |
* CHECKPOINT_END_OF_RECOVERY). |
| 997 |
* CHECKPOINT_WAIT: wait for completion before returning (otherwise, |
| 998 |
* just signal checkpointer to do it, and return). |
| 999 |
* CHECKPOINT_CAUSE_XLOG: checkpoint is requested due to xlog filling. |
| 1000 |
* (This affects logging, and in particular enables CheckPointWarning.) |
| 1001 |
*/ |
| 1002 |
void |
| 1003 |
RequestCheckpoint(int flags) |
| 1004 |
{ |
| 1005 |
int ntries; |
| 1006 |
int old_failed, |
| 1007 |
old_started; |
| 1008 |
|
| 1009 |
/* |
| 1010 |
* If in a standalone backend, just do it ourselves. |
| 1011 |
*/ |
| 1012 |
if (!IsPostmasterEnvironment) |
| 1013 |
{ |
| 1014 |
/* |
| 1015 |
* There's no point in doing slow checkpoints in a standalone backend, |
| 1016 |
* because there's no other backends the checkpoint could disrupt. |
| 1017 |
*/ |
| 1018 |
CreateCheckPoint(flags | CHECKPOINT_IMMEDIATE); |
| 1019 |
|
| 1020 |
/* Free all smgr objects, as CheckpointerMain() normally would. */ |
| 1021 |
smgrdestroyall(); |
| 1022 |
|
| 1023 |
return; |
| 1024 |
} |
| 1025 |
|
| 1026 |
/* |
| 1027 |
* Atomically set the request flags, and take a snapshot of the counters. |
| 1028 |
* When we see ckpt_started > old_started, we know the flags we set here |
| 1029 |
* have been seen by checkpointer. |
| 1030 |
* |
| 1031 |
* Note that we OR the flags with any existing flags, to avoid overriding |
| 1032 |
* a "stronger" request by another backend. The flag senses must be |
| 1033 |
* chosen to make this work! |
| 1034 |
*/ |
| 1035 |
SpinLockAcquire(&CheckpointerShmem->ckpt_lck); |
| 1036 |
|
| 1037 |
old_failed = CheckpointerShmem->ckpt_failed; |
| 1038 |
old_started = CheckpointerShmem->ckpt_started; |
| 1039 |
CheckpointerShmem->ckpt_flags |= (flags | CHECKPOINT_REQUESTED); |
| 1040 |
|
| 1041 |
SpinLockRelease(&CheckpointerShmem->ckpt_lck); |
| 1042 |
|
| 1043 |
/* |
| 1044 |
* Set checkpointer's latch to request checkpoint. It's possible that the |
| 1045 |
* checkpointer hasn't started yet, so we will retry a few times if |
| 1046 |
* needed. (Actually, more than a few times, since on slow or overloaded |
| 1047 |
* buildfarm machines, it's been observed that the checkpointer can take |
| 1048 |
* several seconds to start.) However, if not told to wait for the |
| 1049 |
* checkpoint to occur, we consider failure to set the latch to be |
| 1050 |
* nonfatal and merely LOG it. The checkpointer should see the request |
| 1051 |
* when it does start, with or without the SetLatch(). |
| 1052 |
*/ |
| 1053 |
#define MAX_SIGNAL_TRIES 600 /* max wait 60.0 sec */ |
| 1054 |
for (ntries = 0;; ntries++) |
| 1055 |
{ |
| 1056 |
volatile PROC_HDR *procglobal = ProcGlobal; |
| 1057 |
ProcNumber checkpointerProc = procglobal->checkpointerProc; |
| 1058 |
|
| 1059 |
if (checkpointerProc == INVALID_PROC_NUMBER) |
| 1060 |
{ |
| 1061 |
if (ntries >= MAX_SIGNAL_TRIES || !(flags & CHECKPOINT_WAIT)) |
| 1062 |
{ |
| 1063 |
elog((flags & CHECKPOINT_WAIT) ? ERROR : LOG, |
| 1064 |
"could not notify checkpoint: checkpointer is not running"); |
| 1065 |
break; |
| 1066 |
} |
| 1067 |
} |
| 1068 |
else |
| 1069 |
{ |
| 1070 |
SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch); |
| 1071 |
/* notified successfully */ |
| 1072 |
break; |
| 1073 |
} |
| 1074 |
|
| 1075 |
CHECK_FOR_INTERRUPTS(); |
| 1076 |
pg_usleep(100000L); /* wait 0.1 sec, then retry */ |
| 1077 |
} |
| 1078 |
|
| 1079 |
/* |
| 1080 |
* If requested, wait for completion. We detect completion according to |
| 1081 |
* the algorithm given above. |
| 1082 |
*/ |
| 1083 |
if (flags & CHECKPOINT_WAIT) |
| 1084 |
{ |
| 1085 |
int new_started, |
| 1086 |
new_failed; |
| 1087 |
|
| 1088 |
/* Wait for a new checkpoint to start. */ |
| 1089 |
ConditionVariablePrepareToSleep(&CheckpointerShmem->start_cv); |
| 1090 |
for (;;) |
| 1091 |
{ |
| 1092 |
SpinLockAcquire(&CheckpointerShmem->ckpt_lck); |
| 1093 |
new_started = CheckpointerShmem->ckpt_started; |
| 1094 |
SpinLockRelease(&CheckpointerShmem->ckpt_lck); |
| 1095 |
|
| 1096 |
if (new_started != old_started) |
| 1097 |
break; |
| 1098 |
|
| 1099 |
ConditionVariableSleep(&CheckpointerShmem->start_cv, |
| 1100 |
WAIT_EVENT_CHECKPOINT_START); |
| 1101 |
} |
| 1102 |
ConditionVariableCancelSleep(); |
| 1103 |
|
| 1104 |
/* |
| 1105 |
* We are waiting for ckpt_done >= new_started, in a modulo sense. |
| 1106 |
*/ |
| 1107 |
ConditionVariablePrepareToSleep(&CheckpointerShmem->done_cv); |
| 1108 |
for (;;) |
| 1109 |
{ |
| 1110 |
int new_done; |
| 1111 |
|
| 1112 |
SpinLockAcquire(&CheckpointerShmem->ckpt_lck); |
| 1113 |
new_done = CheckpointerShmem->ckpt_done; |
| 1114 |
new_failed = CheckpointerShmem->ckpt_failed; |
| 1115 |
SpinLockRelease(&CheckpointerShmem->ckpt_lck); |
| 1116 |
|
| 1117 |
if (new_done - new_started >= 0) |
| 1118 |
break; |
| 1119 |
|
| 1120 |
ConditionVariableSleep(&CheckpointerShmem->done_cv, |
| 1121 |
WAIT_EVENT_CHECKPOINT_DONE); |
| 1122 |
} |
| 1123 |
ConditionVariableCancelSleep(); |
| 1124 |
|
| 1125 |
if (new_failed != old_failed) |
| 1126 |
ereport(ERROR, |
| 1127 |
(errmsg("checkpoint request failed"), |
| 1128 |
errhint("Consult recent messages in the server log for details."))); |
| 1129 |
} |
| 1130 |
} |
| 1131 |
|
| 1132 |
/* |
| 1133 |
* ForwardSyncRequest |
| 1134 |
* Forward a file-fsync request from a backend to the checkpointer |
| 1135 |
* |
| 1136 |
* Whenever a backend is compelled to write directly to a relation |
| 1137 |
* (which should be seldom, if the background writer is getting its job done), |
| 1138 |
* the backend calls this routine to pass over knowledge that the relation |
| 1139 |
* is dirty and must be fsync'd before next checkpoint. We also use this |
| 1140 |
* opportunity to count such writes for statistical purposes. |
| 1141 |
* |
| 1142 |
* To avoid holding the lock for longer than necessary, we normally write |
| 1143 |
* to the requests[] queue without checking for duplicates. The checkpointer |
| 1144 |
* will have to eliminate dups internally anyway. However, if we discover |
| 1145 |
* that the queue is full, we make a pass over the entire queue to compact |
| 1146 |
* it. This is somewhat expensive, but the alternative is for the backend |
| 1147 |
* to perform its own fsync, which is far more expensive in practice. It |
| 1148 |
* is theoretically possible a backend fsync might still be necessary, if |
| 1149 |
* the queue is full and contains no duplicate entries. In that case, we |
| 1150 |
* let the backend know by returning false. |
| 1151 |
*/ |
| 1152 |
bool |
| 1153 |
ForwardSyncRequest(const FileTag *ftag, SyncRequestType type) |
| 1154 |
{ |
| 1155 |
CheckpointerRequest *request; |
| 1156 |
bool too_full; |
| 1157 |
|
| 1158 |
if (!IsUnderPostmaster) |
| 1159 |
return false; /* probably shouldn't even get here */ |
| 1160 |
|
| 1161 |
if (AmCheckpointerProcess()) |
| 1162 |
elog(ERROR, "ForwardSyncRequest must not be called in checkpointer"); |
| 1163 |
|
| 1164 |
LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE); |
| 1165 |
|
| 1166 |
/* |
| 1167 |
* If the checkpointer isn't running or the request queue is full, the |
| 1168 |
* backend will have to perform its own fsync request. But before forcing |
| 1169 |
* that to happen, we can try to compact the request queue. |
| 1170 |
*/ |
| 1171 |
if (CheckpointerShmem->checkpointer_pid == 0 || |
| 1172 |
(CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests && |
| 1173 |
!CompactCheckpointerRequestQueue())) |
| 1174 |
{ |
| 1175 |
LWLockRelease(CheckpointerCommLock); |
| 1176 |
return false; |
| 1177 |
} |
| 1178 |
|
| 1179 |
/* OK, insert request */ |
| 1180 |
request = &CheckpointerShmem->requests[CheckpointerShmem->num_requests++]; |
| 1181 |
request->ftag = *ftag; |
| 1182 |
request->type = type; |
| 1183 |
|
| 1184 |
/* If queue is more than half full, nudge the checkpointer to empty it */ |
| 1185 |
too_full = (CheckpointerShmem->num_requests >= |
| 1186 |
CheckpointerShmem->max_requests / 2); |
| 1187 |
|
| 1188 |
LWLockRelease(CheckpointerCommLock); |
| 1189 |
|
| 1190 |
/* ... but not till after we release the lock */ |
| 1191 |
if (too_full) |
| 1192 |
{ |
| 1193 |
volatile PROC_HDR *procglobal = ProcGlobal; |
| 1194 |
ProcNumber checkpointerProc = procglobal->checkpointerProc; |
| 1195 |
|
| 1196 |
if (checkpointerProc != INVALID_PROC_NUMBER) |
| 1197 |
SetLatch(&GetPGProcByNumber(checkpointerProc)->procLatch); |
| 1198 |
} |
| 1199 |
|
| 1200 |
return true; |
| 1201 |
} |
| 1202 |
|
| 1203 |
/* |
| 1204 |
* CompactCheckpointerRequestQueue |
| 1205 |
* Remove duplicates from the request queue to avoid backend fsyncs. |
| 1206 |
* Returns "true" if any entries were removed. |
| 1207 |
* |
| 1208 |
* Although a full fsync request queue is not common, it can lead to severe |
| 1209 |
* performance problems when it does happen. So far, this situation has |
| 1210 |
* only been observed to occur when the system is under heavy write load, |
| 1211 |
* and especially during the "sync" phase of a checkpoint. Without this |
| 1212 |
* logic, each backend begins doing an fsync for every block written, which |
| 1213 |
* gets very expensive and can slow down the whole system. |
| 1214 |
* |
| 1215 |
* Trying to do this every time the queue is full could lose if there |
| 1216 |
* aren't any removable entries. But that should be vanishingly rare in |
| 1217 |
* practice: there's one queue entry per shared buffer. |
| 1218 |
*/ |
| 1219 |
static bool |
| 1220 |
CompactCheckpointerRequestQueue(void) |
| 1221 |
{ |
| 1222 |
struct CheckpointerSlotMapping |
| 1223 |
{ |
| 1224 |
CheckpointerRequest request; |
| 1225 |
int slot; |
| 1226 |
}; |
| 1227 |
|
| 1228 |
int n, |
| 1229 |
preserve_count; |
| 1230 |
int num_skipped = 0; |
| 1231 |
HASHCTL ctl; |
| 1232 |
HTAB *htab; |
| 1233 |
bool *skip_slot; |
| 1234 |
|
| 1235 |
/* must hold CheckpointerCommLock in exclusive mode */ |
| 1236 |
Assert(LWLockHeldByMe(CheckpointerCommLock)); |
| 1237 |
|
| 1238 |
/* Avoid memory allocations in a critical section. */ |
| 1239 |
if (CritSectionCount > 0) |
| 1240 |
return false; |
| 1241 |
|
| 1242 |
/* Initialize skip_slot array */ |
| 1243 |
skip_slot = palloc0(sizeof(bool) * CheckpointerShmem->num_requests); |
| 1244 |
|
| 1245 |
/* Initialize temporary hash table */ |
| 1246 |
ctl.keysize = sizeof(CheckpointerRequest); |
| 1247 |
ctl.entrysize = sizeof(struct CheckpointerSlotMapping); |
| 1248 |
ctl.hcxt = CurrentMemoryContext; |
| 1249 |
|
| 1250 |
htab = hash_create("CompactCheckpointerRequestQueue", |
| 1251 |
CheckpointerShmem->num_requests, |
| 1252 |
&ctl, |
| 1253 |
HASH_ELEM | HASH_BLOBS | HASH_CONTEXT); |
| 1254 |
|
| 1255 |
/* |
| 1256 |
* The basic idea here is that a request can be skipped if it's followed |
| 1257 |
* by a later, identical request. It might seem more sensible to work |
| 1258 |
* backwards from the end of the queue and check whether a request is |
| 1259 |
* *preceded* by an earlier, identical request, in the hopes of doing less |
| 1260 |
* copying. But that might change the semantics, if there's an |
| 1261 |
* intervening SYNC_FORGET_REQUEST or SYNC_FILTER_REQUEST, so we do it |
| 1262 |
* this way. It would be possible to be even smarter if we made the code |
| 1263 |
* below understand the specific semantics of such requests (it could blow |
| 1264 |
* away preceding entries that would end up being canceled anyhow), but |
| 1265 |
* it's not clear that the extra complexity would buy us anything. |
| 1266 |
*/ |
| 1267 |
for (n = 0; n < CheckpointerShmem->num_requests; n++) |
| 1268 |
{ |
| 1269 |
CheckpointerRequest *request; |
| 1270 |
struct CheckpointerSlotMapping *slotmap; |
| 1271 |
bool found; |
| 1272 |
|
| 1273 |
/* |
| 1274 |
* We use the request struct directly as a hashtable key. This |
| 1275 |
* assumes that any padding bytes in the structs are consistently the |
| 1276 |
* same, which should be okay because we zeroed them in |
| 1277 |
* CheckpointerShmemInit. Note also that RelFileLocator had better |
| 1278 |
* contain no pad bytes. |
| 1279 |
*/ |
| 1280 |
request = &CheckpointerShmem->requests[n]; |
| 1281 |
slotmap = hash_search(htab, request, HASH_ENTER, &found); |
| 1282 |
if (found) |
| 1283 |
{ |
| 1284 |
/* Duplicate, so mark the previous occurrence as skippable */ |
| 1285 |
skip_slot[slotmap->slot] = true; |
| 1286 |
num_skipped++; |
| 1287 |
} |
| 1288 |
/* Remember slot containing latest occurrence of this request value */ |
| 1289 |
slotmap->slot = n; |
| 1290 |
} |
| 1291 |
|
| 1292 |
/* Done with the hash table. */ |
| 1293 |
hash_destroy(htab); |
| 1294 |
|
| 1295 |
/* If no duplicates, we're out of luck. */ |
| 1296 |
if (!num_skipped) |
| 1297 |
{ |
| 1298 |
pfree(skip_slot); |
| 1299 |
return false; |
| 1300 |
} |
| 1301 |
|
| 1302 |
/* We found some duplicates; remove them. */ |
| 1303 |
preserve_count = 0; |
| 1304 |
for (n = 0; n < CheckpointerShmem->num_requests; n++) |
| 1305 |
{ |
| 1306 |
if (skip_slot[n]) |
| 1307 |
continue; |
| 1308 |
CheckpointerShmem->requests[preserve_count++] = CheckpointerShmem->requests[n]; |
| 1309 |
} |
| 1310 |
ereport(DEBUG1, |
| 1311 |
(errmsg_internal("compacted fsync request queue from %d entries to %d entries", |
| 1312 |
CheckpointerShmem->num_requests, preserve_count))); |
| 1313 |
CheckpointerShmem->num_requests = preserve_count; |
| 1314 |
|
| 1315 |
/* Cleanup. */ |
| 1316 |
pfree(skip_slot); |
| 1317 |
return true; |
| 1318 |
} |
| 1319 |
|
| 1320 |
/* |
| 1321 |
* AbsorbSyncRequests |
| 1322 |
* Retrieve queued sync requests and pass them to sync mechanism. |
| 1323 |
* |
| 1324 |
* This is exported because it must be called during CreateCheckPoint; |
| 1325 |
* we have to be sure we have accepted all pending requests just before |
| 1326 |
* we start fsync'ing. Since CreateCheckPoint sometimes runs in |
| 1327 |
* non-checkpointer processes, do nothing if not checkpointer. |
| 1328 |
*/ |
| 1329 |
void |
| 1330 |
AbsorbSyncRequests(void) |
| 1331 |
{ |
| 1332 |
CheckpointerRequest *requests = NULL; |
| 1333 |
CheckpointerRequest *request; |
| 1334 |
int n; |
| 1335 |
|
| 1336 |
if (!AmCheckpointerProcess()) |
| 1337 |
return; |
| 1338 |
|
| 1339 |
LWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE); |
| 1340 |
|
| 1341 |
/* |
| 1342 |
* We try to avoid holding the lock for a long time by copying the request |
| 1343 |
* array, and processing the requests after releasing the lock. |
| 1344 |
* |
| 1345 |
* Once we have cleared the requests from shared memory, we have to PANIC |
| 1346 |
* if we then fail to absorb them (eg, because our hashtable runs out of |
| 1347 |
* memory). This is because the system cannot run safely if we are unable |
| 1348 |
* to fsync what we have been told to fsync. Fortunately, the hashtable |
| 1349 |
* is so small that the problem is quite unlikely to arise in practice. |
| 1350 |
*/ |
| 1351 |
n = CheckpointerShmem->num_requests; |
| 1352 |
if (n > 0) |
| 1353 |
{ |
| 1354 |
requests = (CheckpointerRequest *) palloc(n * sizeof(CheckpointerRequest)); |
| 1355 |
memcpy(requests, CheckpointerShmem->requests, n * sizeof(CheckpointerRequest)); |
| 1356 |
} |
| 1357 |
|
| 1358 |
START_CRIT_SECTION(); |
| 1359 |
|
| 1360 |
CheckpointerShmem->num_requests = 0; |
| 1361 |
|
| 1362 |
LWLockRelease(CheckpointerCommLock); |
| 1363 |
|
| 1364 |
for (request = requests; n > 0; request++, n--) |
| 1365 |
RememberSyncRequest(&request->ftag, request->type); |
| 1366 |
|
| 1367 |
END_CRIT_SECTION(); |
| 1368 |
|
| 1369 |
if (requests) |
| 1370 |
pfree(requests); |
| 1371 |
} |
| 1372 |
|
| 1373 |
/* |
| 1374 |
* Update any shared memory configurations based on config parameters |
| 1375 |
*/ |
| 1376 |
static void |
| 1377 |
UpdateSharedMemoryConfig(void) |
| 1378 |
{ |
| 1379 |
/* update global shmem state for sync rep */ |
| 1380 |
SyncRepUpdateSyncStandbysDefined(); |
| 1381 |
|
| 1382 |
/* |
| 1383 |
* If full_page_writes has been changed by SIGHUP, we update it in shared |
| 1384 |
* memory and write an XLOG_FPW_CHANGE record. |
| 1385 |
*/ |
| 1386 |
UpdateFullPageWrites(); |
| 1387 |
|
| 1388 |
elog(DEBUG2, "checkpointer updated shared memory configuration values"); |
| 1389 |
} |
| 1390 |
|
| 1391 |
/* |
| 1392 |
* FirstCallSinceLastCheckpoint allows a process to take an action once |
| 1393 |
* per checkpoint cycle by asynchronously checking for checkpoint completion. |
| 1394 |
*/ |
| 1395 |
bool |
| 1396 |
FirstCallSinceLastCheckpoint(void) |
| 1397 |
{ |
| 1398 |
static int ckpt_done = 0; |
| 1399 |
int new_done; |
| 1400 |
bool FirstCall = false; |
| 1401 |
|
| 1402 |
SpinLockAcquire(&CheckpointerShmem->ckpt_lck); |
| 1403 |
new_done = CheckpointerShmem->ckpt_done; |
| 1404 |
SpinLockRelease(&CheckpointerShmem->ckpt_lck); |
| 1405 |
|
| 1406 |
if (new_done != ckpt_done) |
| 1407 |
FirstCall = true; |
| 1408 |
|
| 1409 |
ckpt_done = new_done; |
| 1410 |
|
| 1411 |
return FirstCall; |
| 1412 |
} |