Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * autovacuum.c
4 : *
5 : * PostgreSQL Integrated Autovacuum Daemon
6 : *
7 : * The autovacuum system is structured in two different kinds of processes: the
8 : * autovacuum launcher and the autovacuum worker. The launcher is an
9 : * always-running process, started by the postmaster when the autovacuum GUC
10 : * parameter is set. The launcher schedules autovacuum workers to be started
11 : * when appropriate. The workers are the processes which execute the actual
12 : * vacuuming; they connect to a database as determined in the launcher, and
13 : * once connected they examine the catalogs to select the tables to vacuum.
14 : *
15 : * The autovacuum launcher cannot start the worker processes by itself,
16 : * because doing so would cause robustness issues (namely, failure to shut
17 : * them down on exceptional conditions, and also, since the launcher is
18 : * connected to shared memory and is thus subject to corruption there, it is
19 : * not as robust as the postmaster). So it leaves that task to the postmaster.
20 : *
21 : * There is an autovacuum shared memory area, where the launcher stores
22 : * information about the database it wants vacuumed. When it wants a new
23 : * worker to start, it sets a flag in shared memory and sends a signal to the
24 : * postmaster. Then postmaster knows nothing more than it must start a worker;
25 : * so it forks a new child, which turns into a worker. This new process
26 : * connects to shared memory, and there it can inspect the information that the
27 : * launcher has set up.
28 : *
29 : * If the fork() call fails in the postmaster, it sets a flag in the shared
30 : * memory area, and sends a signal to the launcher. The launcher, upon
31 : * noticing the flag, can try starting the worker again by resending the
32 : * signal. Note that the failure can only be transient (fork failure due to
33 : * high load, memory pressure, too many processes, etc); more permanent
34 : * problems, like failure to connect to a database, are detected later in the
35 : * worker and dealt with just by having the worker exit normally. The launcher
36 : * will launch a new worker again later, per schedule.
37 : *
38 : * When the worker is done vacuuming it sends SIGUSR2 to the launcher. The
39 : * launcher then wakes up and is able to launch another worker, if the schedule
40 : * is so tight that a new worker is needed immediately. At this time the
41 : * launcher can also balance the settings for the various remaining workers'
42 : * cost-based vacuum delay feature.
43 : *
44 : * Note that there can be more than one worker in a database concurrently.
45 : * They will store the table they are currently vacuuming in shared memory, so
46 : * that other workers avoid being blocked waiting for the vacuum lock for that
47 : * table. They will also reload the pgstats data just before vacuuming each
48 : * table, to avoid vacuuming a table that was just finished being vacuumed by
49 : * another worker and thus is no longer noted in shared memory. However,
50 : * there is a window (caused by pgstat delay) on which a worker may choose a
51 : * table that was already vacuumed; this is a bug in the current design.
52 : *
53 : * Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
54 : * Portions Copyright (c) 1994, Regents of the University of California
55 : *
56 : *
57 : * IDENTIFICATION
58 : * src/backend/postmaster/autovacuum.c
59 : *
60 : *-------------------------------------------------------------------------
61 : */
62 : #include "postgres.h"
63 :
64 : #include <signal.h>
65 : #include <sys/time.h>
66 : #include <unistd.h>
67 :
68 : #include "access/heapam.h"
69 : #include "access/htup_details.h"
70 : #include "access/multixact.h"
71 : #include "access/reloptions.h"
72 : #include "access/tableam.h"
73 : #include "access/transam.h"
74 : #include "access/xact.h"
75 : #include "catalog/dependency.h"
76 : #include "catalog/namespace.h"
77 : #include "catalog/pg_database.h"
78 : #include "commands/dbcommands.h"
79 : #include "commands/vacuum.h"
80 : #include "lib/ilist.h"
81 : #include "libpq/pqsignal.h"
82 : #include "miscadmin.h"
83 : #include "nodes/makefuncs.h"
84 : #include "pgstat.h"
85 : #include "postmaster/autovacuum.h"
86 : #include "postmaster/fork_process.h"
87 : #include "postmaster/interrupt.h"
88 : #include "postmaster/postmaster.h"
89 : #include "storage/bufmgr.h"
90 : #include "storage/ipc.h"
91 : #include "storage/latch.h"
92 : #include "storage/lmgr.h"
93 : #include "storage/pmsignal.h"
94 : #include "storage/proc.h"
95 : #include "storage/procsignal.h"
96 : #include "storage/sinvaladt.h"
97 : #include "storage/smgr.h"
98 : #include "tcop/tcopprot.h"
99 : #include "utils/fmgroids.h"
100 : #include "utils/fmgrprotos.h"
101 : #include "utils/lsyscache.h"
102 : #include "utils/memutils.h"
103 : #include "utils/ps_status.h"
104 : #include "utils/rel.h"
105 : #include "utils/snapmgr.h"
106 : #include "utils/syscache.h"
107 : #include "utils/timeout.h"
108 : #include "utils/timestamp.h"
109 :
110 :
111 : /*
112 : * GUC parameters
113 : */
114 : bool autovacuum_start_daemon = false;
115 : int autovacuum_max_workers;
116 : int autovacuum_work_mem = -1;
117 : int autovacuum_naptime;
118 : int autovacuum_vac_thresh;
119 : double autovacuum_vac_scale;
120 : int autovacuum_vac_ins_thresh;
121 : double autovacuum_vac_ins_scale;
122 : int autovacuum_anl_thresh;
123 : double autovacuum_anl_scale;
124 : int autovacuum_freeze_max_age;
125 : int autovacuum_multixact_freeze_max_age;
126 :
127 : double autovacuum_vac_cost_delay;
128 : int autovacuum_vac_cost_limit;
129 :
130 : int Log_autovacuum_min_duration = -1;
131 :
132 : /* how long to keep pgstat data in the launcher, in milliseconds */
133 : #define STATS_READ_DELAY 1000
134 :
135 : /* the minimum allowed time between two awakenings of the launcher */
136 : #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */
137 : #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */
138 :
139 : /* Flags to tell if we are in an autovacuum process */
140 : static bool am_autovacuum_launcher = false;
141 : static bool am_autovacuum_worker = false;
142 :
143 : /* Flags set by signal handlers */
144 : static volatile sig_atomic_t got_SIGUSR2 = false;
145 :
146 : /* Comparison points for determining whether freeze_max_age is exceeded */
147 : static TransactionId recentXid;
148 : static MultiXactId recentMulti;
149 :
150 : /* Default freeze ages to use for autovacuum (varies by database) */
151 : static int default_freeze_min_age;
152 : static int default_freeze_table_age;
153 : static int default_multixact_freeze_min_age;
154 : static int default_multixact_freeze_table_age;
155 :
156 : /* Memory context for long-lived data */
157 : static MemoryContext AutovacMemCxt;
158 :
159 : /* struct to keep track of databases in launcher */
160 : typedef struct avl_dbase
161 : {
162 : Oid adl_datid; /* hash key -- must be first */
163 : TimestampTz adl_next_worker;
164 : int adl_score;
165 : dlist_node adl_node;
166 : } avl_dbase;
167 :
168 : /* struct to keep track of databases in worker */
169 : typedef struct avw_dbase
170 : {
171 : Oid adw_datid;
172 : char *adw_name;
173 : TransactionId adw_frozenxid;
174 : MultiXactId adw_minmulti;
175 : PgStat_StatDBEntry *adw_entry;
176 : } avw_dbase;
177 :
178 : /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */
179 : typedef struct av_relation
180 : {
181 : Oid ar_toastrelid; /* hash key - must be first */
182 : Oid ar_relid;
183 : bool ar_hasrelopts;
184 : AutoVacOpts ar_reloptions; /* copy of AutoVacOpts from the main table's
185 : * reloptions, or NULL if none */
186 : } av_relation;
187 :
188 : /* struct to keep track of tables to vacuum and/or analyze, after rechecking */
189 : typedef struct autovac_table
190 : {
191 : Oid at_relid;
192 : VacuumParams at_params;
193 : double at_vacuum_cost_delay;
194 : int at_vacuum_cost_limit;
195 : bool at_dobalance;
196 : bool at_sharedrel;
197 : char *at_relname;
198 : char *at_nspname;
199 : char *at_datname;
200 : } autovac_table;
201 :
202 : /*-------------
203 : * This struct holds information about a single worker's whereabouts. We keep
204 : * an array of these in shared memory, sized according to
205 : * autovacuum_max_workers.
206 : *
207 : * wi_links entry into free list or running list
208 : * wi_dboid OID of the database this worker is supposed to work on
209 : * wi_tableoid OID of the table currently being vacuumed, if any
210 : * wi_sharedrel flag indicating whether table is marked relisshared
211 : * wi_proc pointer to PGPROC of the running worker, NULL if not started
212 : * wi_launchtime Time at which this worker was launched
213 : * wi_cost_* Vacuum cost-based delay parameters current in this worker
214 : *
215 : * All fields are protected by AutovacuumLock, except for wi_tableoid and
216 : * wi_sharedrel which are protected by AutovacuumScheduleLock (note these
217 : * two fields are read-only for everyone except that worker itself).
218 : *-------------
219 : */
220 : typedef struct WorkerInfoData
221 : {
222 : dlist_node wi_links;
223 : Oid wi_dboid;
224 : Oid wi_tableoid;
225 : PGPROC *wi_proc;
226 : TimestampTz wi_launchtime;
227 : bool wi_dobalance;
228 : bool wi_sharedrel;
229 : double wi_cost_delay;
230 : int wi_cost_limit;
231 : int wi_cost_limit_base;
232 : } WorkerInfoData;
233 :
234 : typedef struct WorkerInfoData *WorkerInfo;
235 :
236 : /*
237 : * Possible signals received by the launcher from remote processes. These are
238 : * stored atomically in shared memory so that other processes can set them
239 : * without locking.
240 : */
241 : typedef enum
242 : {
243 : AutoVacForkFailed, /* failed trying to start a worker */
244 : AutoVacRebalance, /* rebalance the cost limits */
245 : AutoVacNumSignals /* must be last */
246 : } AutoVacuumSignal;
247 :
248 : /*
249 : * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This
250 : * list is mostly protected by AutovacuumLock, except that if an item is
251 : * marked 'active' other processes must not modify the work-identifying
252 : * members.
253 : */
254 : typedef struct AutoVacuumWorkItem
255 : {
256 : AutoVacuumWorkItemType avw_type;
257 : bool avw_used; /* below data is valid */
258 : bool avw_active; /* being processed */
259 : Oid avw_database;
260 : Oid avw_relation;
261 : BlockNumber avw_blockNumber;
262 : } AutoVacuumWorkItem;
263 :
264 : #define NUM_WORKITEMS 256
265 :
266 : /*-------------
267 : * The main autovacuum shmem struct. On shared memory we store this main
268 : * struct and the array of WorkerInfo structs. This struct keeps:
269 : *
270 : * av_signal set by other processes to indicate various conditions
271 : * av_launcherpid the PID of the autovacuum launcher
272 : * av_freeWorkers the WorkerInfo freelist
273 : * av_runningWorkers the WorkerInfo non-free queue
274 : * av_startingWorker pointer to WorkerInfo currently being started (cleared by
275 : * the worker itself as soon as it's up and running)
276 : * av_workItems work item array
277 : *
278 : * This struct is protected by AutovacuumLock, except for av_signal and parts
279 : * of the worker list (see above).
280 : *-------------
281 : */
282 : typedef struct
283 : {
284 : sig_atomic_t av_signal[AutoVacNumSignals];
285 : pid_t av_launcherpid;
286 : dlist_head av_freeWorkers;
287 : dlist_head av_runningWorkers;
288 : WorkerInfo av_startingWorker;
289 : AutoVacuumWorkItem av_workItems[NUM_WORKITEMS];
290 : } AutoVacuumShmemStruct;
291 :
292 : static AutoVacuumShmemStruct *AutoVacuumShmem;
293 :
294 : /*
295 : * the database list (of avl_dbase elements) in the launcher, and the context
296 : * that contains it
297 : */
298 : static dlist_head DatabaseList = DLIST_STATIC_INIT(DatabaseList);
299 : static MemoryContext DatabaseListCxt = NULL;
300 :
301 : /* Pointer to my own WorkerInfo, valid on each worker */
302 : static WorkerInfo MyWorkerInfo = NULL;
303 :
304 : /* PID of launcher, valid only in worker while shutting down */
305 : int AutovacuumLauncherPid = 0;
306 :
307 : #ifdef EXEC_BACKEND
308 : static pid_t avlauncher_forkexec(void);
309 : static pid_t avworker_forkexec(void);
310 : #endif
311 : NON_EXEC_STATIC void AutoVacWorkerMain(int argc, char *argv[]) pg_attribute_noreturn();
312 : NON_EXEC_STATIC void AutoVacLauncherMain(int argc, char *argv[]) pg_attribute_noreturn();
313 :
314 : static Oid do_start_worker(void);
315 : static void HandleAutoVacLauncherInterrupts(void);
316 : static void AutoVacLauncherShutdown(void) pg_attribute_noreturn();
317 : static void launcher_determine_sleep(bool canlaunch, bool recursing,
318 : struct timeval *nap);
319 : static void launch_worker(TimestampTz now);
320 : static List *get_database_list(void);
321 : static void rebuild_database_list(Oid newdb);
322 : static int db_comparator(const void *a, const void *b);
323 : static void autovac_balance_cost(void);
324 :
325 : static void do_autovacuum(void);
326 : static void FreeWorkerInfo(int code, Datum arg);
327 :
328 : static autovac_table *table_recheck_autovac(Oid relid, HTAB *table_toast_map,
329 : TupleDesc pg_class_desc,
330 : int effective_multixact_freeze_max_age);
331 : static void recheck_relation_needs_vacanalyze(Oid relid, AutoVacOpts *avopts,
332 : Form_pg_class classForm,
333 : int effective_multixact_freeze_max_age,
334 : bool *dovacuum, bool *doanalyze, bool *wraparound);
335 : static void relation_needs_vacanalyze(Oid relid, AutoVacOpts *relopts,
336 : Form_pg_class classForm,
337 : PgStat_StatTabEntry *tabentry,
338 : int effective_multixact_freeze_max_age,
339 : bool *dovacuum, bool *doanalyze, bool *wraparound);
340 :
341 : static void autovacuum_do_vac_analyze(autovac_table *tab,
342 : BufferAccessStrategy bstrategy);
343 : static AutoVacOpts *extract_autovac_opts(HeapTuple tup,
344 : TupleDesc pg_class_desc);
345 : static PgStat_StatTabEntry *get_pgstat_tabentry_relid(Oid relid, bool isshared,
346 : PgStat_StatDBEntry *shared,
347 : PgStat_StatDBEntry *dbentry);
348 : static void perform_work_item(AutoVacuumWorkItem *workitem);
349 : static void autovac_report_activity(autovac_table *tab);
350 : static void autovac_report_workitem(AutoVacuumWorkItem *workitem,
351 : const char *nspname, const char *relname);
352 : static void avl_sigusr2_handler(SIGNAL_ARGS);
353 : static void autovac_refresh_stats(void);
354 :
355 :
356 :
357 : /********************************************************************
358 : * AUTOVACUUM LAUNCHER CODE
359 : ********************************************************************/
360 :
361 : #ifdef EXEC_BACKEND
362 : /*
363 : * forkexec routine for the autovacuum launcher process.
364 : *
365 : * Format up the arglist, then fork and exec.
366 : */
367 : static pid_t
368 : avlauncher_forkexec(void)
369 : {
370 : char *av[10];
371 : int ac = 0;
372 :
373 : av[ac++] = "postgres";
374 : av[ac++] = "--forkavlauncher";
375 : av[ac++] = NULL; /* filled in by postmaster_forkexec */
376 : av[ac] = NULL;
377 :
378 : Assert(ac < lengthof(av));
379 :
380 : return postmaster_forkexec(ac, av);
381 : }
382 :
383 : /*
384 : * We need this set from the outside, before InitProcess is called
385 : */
386 : void
387 : AutovacuumLauncherIAm(void)
388 : {
389 : am_autovacuum_launcher = true;
390 : }
391 : #endif
392 :
393 : /*
394 : * Main entry point for autovacuum launcher process, to be called from the
395 : * postmaster.
396 : */
397 : int
398 782 : StartAutoVacLauncher(void)
399 : {
400 : pid_t AutoVacPID;
401 :
402 : #ifdef EXEC_BACKEND
403 : switch ((AutoVacPID = avlauncher_forkexec()))
404 : #else
405 782 : switch ((AutoVacPID = fork_process()))
406 : #endif
407 : {
408 0 : case -1:
409 0 : ereport(LOG,
410 : (errmsg("could not fork autovacuum launcher process: %m")));
411 0 : return 0;
412 :
413 : #ifndef EXEC_BACKEND
414 478 : case 0:
415 : /* in postmaster child ... */
416 478 : InitPostmasterChild();
417 :
418 : /* Close the postmaster's sockets */
419 478 : ClosePostmasterPorts(false);
420 :
421 478 : AutoVacLauncherMain(0, NULL);
422 : break;
423 : #endif
424 782 : default:
425 782 : return (int) AutoVacPID;
426 : }
427 :
428 : /* shouldn't get here */
429 : return 0;
430 : }
431 :
432 : /*
433 : * Main loop for the autovacuum launcher process.
434 : */
435 : NON_EXEC_STATIC void
436 478 : AutoVacLauncherMain(int argc, char *argv[])
437 : {
438 : sigjmp_buf local_sigjmp_buf;
439 :
440 478 : am_autovacuum_launcher = true;
441 :
442 478 : MyBackendType = B_AUTOVAC_LAUNCHER;
443 478 : init_ps_display(NULL);
444 :
445 478 : ereport(DEBUG1,
446 : (errmsg_internal("autovacuum launcher started")));
447 :
448 478 : if (PostAuthDelay)
449 0 : pg_usleep(PostAuthDelay * 1000000L);
450 :
451 478 : SetProcessingMode(InitProcessing);
452 :
453 : /*
454 : * Set up signal handlers. We operate on databases much like a regular
455 : * backend, so we use the same signal handling. See equivalent code in
456 : * tcop/postgres.c.
457 : */
458 478 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
459 478 : pqsignal(SIGINT, StatementCancelHandler);
460 478 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
461 : /* SIGQUIT handler was already set up by InitPostmasterChild */
462 :
463 478 : InitializeTimeouts(); /* establishes SIGALRM handler */
464 :
465 478 : pqsignal(SIGPIPE, SIG_IGN);
466 478 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
467 478 : pqsignal(SIGUSR2, avl_sigusr2_handler);
468 478 : pqsignal(SIGFPE, FloatExceptionHandler);
469 478 : pqsignal(SIGCHLD, SIG_DFL);
470 :
471 : /*
472 : * Create a per-backend PGPROC struct in shared memory, except in the
473 : * EXEC_BACKEND case where this was done in SubPostmasterMain. We must do
474 : * this before we can use LWLocks (and in the EXEC_BACKEND case we already
475 : * had to do some stuff with LWLocks).
476 : */
477 : #ifndef EXEC_BACKEND
478 478 : InitProcess();
479 : #endif
480 :
481 : /* Early initialization */
482 478 : BaseInit();
483 :
484 478 : InitPostgres(NULL, InvalidOid, NULL, InvalidOid, NULL, false);
485 :
486 478 : SetProcessingMode(NormalProcessing);
487 :
488 : /*
489 : * Create a memory context that we will do all our work in. We do this so
490 : * that we can reset the context during error recovery and thereby avoid
491 : * possible memory leaks.
492 : */
493 478 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
494 : "Autovacuum Launcher",
495 : ALLOCSET_DEFAULT_SIZES);
496 478 : MemoryContextSwitchTo(AutovacMemCxt);
497 :
498 : /*
499 : * If an exception is encountered, processing resumes here.
500 : *
501 : * This code is a stripped down version of PostgresMain error recovery.
502 : *
503 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
504 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
505 : * signals other than SIGQUIT will be blocked until we complete error
506 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
507 : * call redundant, but it is not since InterruptPending might be set
508 : * already.
509 : */
510 478 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
511 : {
512 : /* since not using PG_TRY, must reset error stack by hand */
513 0 : error_context_stack = NULL;
514 :
515 : /* Prevents interrupts while cleaning up */
516 0 : HOLD_INTERRUPTS();
517 :
518 : /* Forget any pending QueryCancel or timeout request */
519 0 : disable_all_timeouts(false);
520 0 : QueryCancelPending = false; /* second to avoid race condition */
521 :
522 : /* Report the error to the server log */
523 0 : EmitErrorReport();
524 :
525 : /* Abort the current transaction in order to recover */
526 0 : AbortCurrentTransaction();
527 :
528 : /*
529 : * Release any other resources, for the case where we were not in a
530 : * transaction.
531 : */
532 0 : LWLockReleaseAll();
533 0 : pgstat_report_wait_end();
534 0 : AbortBufferIO();
535 0 : UnlockBuffers();
536 : /* this is probably dead code, but let's be safe: */
537 0 : if (AuxProcessResourceOwner)
538 0 : ReleaseAuxProcessResources(false);
539 0 : AtEOXact_Buffers(false);
540 0 : AtEOXact_SMgr();
541 0 : AtEOXact_Files(false);
542 0 : AtEOXact_HashTables(false);
543 :
544 : /*
545 : * Now return to normal top-level context and clear ErrorContext for
546 : * next time.
547 : */
548 0 : MemoryContextSwitchTo(AutovacMemCxt);
549 0 : FlushErrorState();
550 :
551 : /* Flush any leaked data in the top-level context */
552 0 : MemoryContextResetAndDeleteChildren(AutovacMemCxt);
553 :
554 : /* don't leave dangling pointers to freed memory */
555 0 : DatabaseListCxt = NULL;
556 0 : dlist_init(&DatabaseList);
557 :
558 : /*
559 : * Make sure pgstat also considers our stat data as gone. Note: we
560 : * mustn't use autovac_refresh_stats here.
561 : */
562 0 : pgstat_clear_snapshot();
563 :
564 : /* Now we can allow interrupts again */
565 0 : RESUME_INTERRUPTS();
566 :
567 : /* if in shutdown mode, no need for anything further; just go away */
568 0 : if (ShutdownRequestPending)
569 0 : AutoVacLauncherShutdown();
570 :
571 : /*
572 : * Sleep at least 1 second after any error. We don't want to be
573 : * filling the error logs as fast as we can.
574 : */
575 0 : pg_usleep(1000000L);
576 : }
577 :
578 : /* We can now handle ereport(ERROR) */
579 478 : PG_exception_stack = &local_sigjmp_buf;
580 :
581 : /* must unblock signals before calling rebuild_database_list */
582 478 : PG_SETMASK(&UnBlockSig);
583 :
584 : /*
585 : * Set always-secure search path. Launcher doesn't connect to a database,
586 : * so this has no effect.
587 : */
588 478 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
589 :
590 : /*
591 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
592 : * in postgresql.conf. We don't really want such a dangerous option being
593 : * applied non-interactively.
594 : */
595 478 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
596 :
597 : /*
598 : * Force settable timeouts off to avoid letting these settings prevent
599 : * regular maintenance from being executed.
600 : */
601 478 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
602 478 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
603 478 : SetConfigOption("idle_in_transaction_session_timeout", "0",
604 : PGC_SUSET, PGC_S_OVERRIDE);
605 :
606 : /*
607 : * Force default_transaction_isolation to READ COMMITTED. We don't want
608 : * to pay the overhead of serializable mode, nor add any risk of causing
609 : * deadlocks or delaying other transactions.
610 : */
611 478 : SetConfigOption("default_transaction_isolation", "read committed",
612 : PGC_SUSET, PGC_S_OVERRIDE);
613 :
614 : /*
615 : * In emergency mode, just start a worker (unless shutdown was requested)
616 : * and go away.
617 : */
618 478 : if (!AutoVacuumingActive())
619 : {
620 0 : if (!ShutdownRequestPending)
621 0 : do_start_worker();
622 0 : proc_exit(0); /* done */
623 : }
624 :
625 478 : AutoVacuumShmem->av_launcherpid = MyProcPid;
626 :
627 : /*
628 : * Create the initial database list. The invariant we want this list to
629 : * keep is that it's ordered by decreasing next_time. As soon as an entry
630 : * is updated to a higher time, it will be moved to the front (which is
631 : * correct because the only operation is to add autovacuum_naptime to the
632 : * entry, and time always increases).
633 : */
634 478 : rebuild_database_list(InvalidOid);
635 :
636 : /* loop until shutdown request */
637 2328 : while (!ShutdownRequestPending)
638 : {
639 : struct timeval nap;
640 2326 : TimestampTz current_time = 0;
641 : bool can_launch;
642 :
643 : /*
644 : * This loop is a bit different from the normal use of WaitLatch,
645 : * because we'd like to sleep before the first launch of a child
646 : * process. So it's WaitLatch, then ResetLatch, then check for
647 : * wakening conditions.
648 : */
649 :
650 2326 : launcher_determine_sleep(!dlist_is_empty(&AutoVacuumShmem->av_freeWorkers),
651 2326 : false, &nap);
652 :
653 : /*
654 : * Wait until naptime expires or we get some type of signal (all the
655 : * signal handlers will wake us by calling SetLatch).
656 : */
657 2326 : (void) WaitLatch(MyLatch,
658 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
659 2326 : (nap.tv_sec * 1000L) + (nap.tv_usec / 1000L),
660 : WAIT_EVENT_AUTOVACUUM_MAIN);
661 :
662 2320 : ResetLatch(MyLatch);
663 :
664 2320 : HandleAutoVacLauncherInterrupts();
665 :
666 : /*
667 : * a worker finished, or postmaster signaled failure to start a worker
668 : */
669 1850 : if (got_SIGUSR2)
670 : {
671 96 : got_SIGUSR2 = false;
672 :
673 : /* rebalance cost limits, if needed */
674 96 : if (AutoVacuumShmem->av_signal[AutoVacRebalance])
675 : {
676 48 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
677 48 : AutoVacuumShmem->av_signal[AutoVacRebalance] = false;
678 48 : autovac_balance_cost();
679 48 : LWLockRelease(AutovacuumLock);
680 : }
681 :
682 96 : if (AutoVacuumShmem->av_signal[AutoVacForkFailed])
683 : {
684 : /*
685 : * If the postmaster failed to start a new worker, we sleep
686 : * for a little while and resend the signal. The new worker's
687 : * state is still in memory, so this is sufficient. After
688 : * that, we restart the main loop.
689 : *
690 : * XXX should we put a limit to the number of times we retry?
691 : * I don't think it makes much sense, because a future start
692 : * of a worker will continue to fail in the same way.
693 : */
694 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = false;
695 0 : pg_usleep(1000000L); /* 1s */
696 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
697 0 : continue;
698 : }
699 : }
700 :
701 : /*
702 : * There are some conditions that we need to check before trying to
703 : * start a worker. First, we need to make sure that there is a worker
704 : * slot available. Second, we need to make sure that no other worker
705 : * failed while starting up.
706 : */
707 :
708 1850 : current_time = GetCurrentTimestamp();
709 1850 : LWLockAcquire(AutovacuumLock, LW_SHARED);
710 :
711 1850 : can_launch = !dlist_is_empty(&AutoVacuumShmem->av_freeWorkers);
712 :
713 1850 : if (AutoVacuumShmem->av_startingWorker != NULL)
714 : {
715 : int waittime;
716 0 : WorkerInfo worker = AutoVacuumShmem->av_startingWorker;
717 :
718 : /*
719 : * We can't launch another worker when another one is still
720 : * starting up (or failed while doing so), so just sleep for a bit
721 : * more; that worker will wake us up again as soon as it's ready.
722 : * We will only wait autovacuum_naptime seconds (up to a maximum
723 : * of 60 seconds) for this to happen however. Note that failure
724 : * to connect to a particular database is not a problem here,
725 : * because the worker removes itself from the startingWorker
726 : * pointer before trying to connect. Problems detected by the
727 : * postmaster (like fork() failure) are also reported and handled
728 : * differently. The only problems that may cause this code to
729 : * fire are errors in the earlier sections of AutoVacWorkerMain,
730 : * before the worker removes the WorkerInfo from the
731 : * startingWorker pointer.
732 : */
733 0 : waittime = Min(autovacuum_naptime, 60) * 1000;
734 0 : if (TimestampDifferenceExceeds(worker->wi_launchtime, current_time,
735 : waittime))
736 : {
737 0 : LWLockRelease(AutovacuumLock);
738 0 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
739 :
740 : /*
741 : * No other process can put a worker in starting mode, so if
742 : * startingWorker is still INVALID after exchanging our lock,
743 : * we assume it's the same one we saw above (so we don't
744 : * recheck the launch time).
745 : */
746 0 : if (AutoVacuumShmem->av_startingWorker != NULL)
747 : {
748 0 : worker = AutoVacuumShmem->av_startingWorker;
749 0 : worker->wi_dboid = InvalidOid;
750 0 : worker->wi_tableoid = InvalidOid;
751 0 : worker->wi_sharedrel = false;
752 0 : worker->wi_proc = NULL;
753 0 : worker->wi_launchtime = 0;
754 0 : dlist_push_head(&AutoVacuumShmem->av_freeWorkers,
755 : &worker->wi_links);
756 0 : AutoVacuumShmem->av_startingWorker = NULL;
757 0 : ereport(WARNING,
758 : errmsg("autovacuum worker took too long to start; canceled"));
759 : }
760 : }
761 : else
762 0 : can_launch = false;
763 : }
764 1850 : LWLockRelease(AutovacuumLock); /* either shared or exclusive */
765 :
766 : /* if we can't do anything, just go back to sleep */
767 1850 : if (!can_launch)
768 0 : continue;
769 :
770 : /* We're OK to start a new worker */
771 :
772 1850 : if (dlist_is_empty(&DatabaseList))
773 : {
774 : /*
775 : * Special case when the list is empty: start a worker right away.
776 : * This covers the initial case, when no database is in pgstats
777 : * (thus the list is empty). Note that the constraints in
778 : * launcher_determine_sleep keep us from starting workers too
779 : * quickly (at most once every autovacuum_naptime when the list is
780 : * empty).
781 : */
782 428 : launch_worker(current_time);
783 : }
784 : else
785 : {
786 : /*
787 : * because rebuild_database_list constructs a list with most
788 : * distant adl_next_worker first, we obtain our database from the
789 : * tail of the list.
790 : */
791 : avl_dbase *avdb;
792 :
793 1422 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
794 :
795 : /*
796 : * launch a worker if next_worker is right now or it is in the
797 : * past
798 : */
799 1422 : if (TimestampDifferenceExceeds(avdb->adl_next_worker,
800 : current_time, 0))
801 12 : launch_worker(current_time);
802 : }
803 : }
804 :
805 2 : AutoVacLauncherShutdown();
806 : }
807 :
808 : /*
809 : * Process any new interrupts.
810 : */
811 : static void
812 2320 : HandleAutoVacLauncherInterrupts(void)
813 : {
814 : /* the normal shutdown case */
815 2320 : if (ShutdownRequestPending)
816 470 : AutoVacLauncherShutdown();
817 :
818 1850 : if (ConfigReloadPending)
819 : {
820 34 : ConfigReloadPending = false;
821 34 : ProcessConfigFile(PGC_SIGHUP);
822 :
823 : /* shutdown requested in config file? */
824 34 : if (!AutoVacuumingActive())
825 0 : AutoVacLauncherShutdown();
826 :
827 : /* rebalance in case the default cost parameters changed */
828 34 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
829 34 : autovac_balance_cost();
830 34 : LWLockRelease(AutovacuumLock);
831 :
832 : /* rebuild the list in case the naptime changed */
833 34 : rebuild_database_list(InvalidOid);
834 : }
835 :
836 : /* Process barrier events */
837 1850 : if (ProcSignalBarrierPending)
838 0 : ProcessProcSignalBarrier();
839 :
840 : /* Perform logging of memory contexts of this process */
841 1850 : if (LogMemoryContextPending)
842 0 : ProcessLogMemoryContextInterrupt();
843 :
844 : /* Process sinval catchup interrupts that happened while sleeping */
845 1850 : ProcessCatchupInterrupt();
846 1850 : }
847 :
848 : /*
849 : * Perform a normal exit from the autovac launcher.
850 : */
851 : static void
852 472 : AutoVacLauncherShutdown(void)
853 : {
854 472 : ereport(DEBUG1,
855 : (errmsg_internal("autovacuum launcher shutting down")));
856 472 : AutoVacuumShmem->av_launcherpid = 0;
857 :
858 472 : proc_exit(0); /* done */
859 : }
860 :
861 : /*
862 : * Determine the time to sleep, based on the database list.
863 : *
864 : * The "canlaunch" parameter indicates whether we can start a worker right now,
865 : * for example due to the workers being all busy. If this is false, we will
866 : * cause a long sleep, which will be interrupted when a worker exits.
867 : */
868 : static void
869 2326 : launcher_determine_sleep(bool canlaunch, bool recursing, struct timeval *nap)
870 : {
871 : /*
872 : * We sleep until the next scheduled vacuum. We trust that when the
873 : * database list was built, care was taken so that no entries have times
874 : * in the past; if the first entry has too close a next_worker value, or a
875 : * time in the past, we will sleep a small nominal time.
876 : */
877 2326 : if (!canlaunch)
878 : {
879 0 : nap->tv_sec = autovacuum_naptime;
880 0 : nap->tv_usec = 0;
881 : }
882 2326 : else if (!dlist_is_empty(&DatabaseList))
883 : {
884 1564 : TimestampTz current_time = GetCurrentTimestamp();
885 : TimestampTz next_wakeup;
886 : avl_dbase *avdb;
887 : long secs;
888 : int usecs;
889 :
890 1564 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
891 :
892 1564 : next_wakeup = avdb->adl_next_worker;
893 1564 : TimestampDifference(current_time, next_wakeup, &secs, &usecs);
894 :
895 1564 : nap->tv_sec = secs;
896 1564 : nap->tv_usec = usecs;
897 : }
898 : else
899 : {
900 : /* list is empty, sleep for whole autovacuum_naptime seconds */
901 762 : nap->tv_sec = autovacuum_naptime;
902 762 : nap->tv_usec = 0;
903 : }
904 :
905 : /*
906 : * If the result is exactly zero, it means a database had an entry with
907 : * time in the past. Rebuild the list so that the databases are evenly
908 : * distributed again, and recalculate the time to sleep. This can happen
909 : * if there are more tables needing vacuum than workers, and they all take
910 : * longer to vacuum than autovacuum_naptime.
911 : *
912 : * We only recurse once. rebuild_database_list should always return times
913 : * in the future, but it seems best not to trust too much on that.
914 : */
915 2326 : if (nap->tv_sec == 0 && nap->tv_usec == 0 && !recursing)
916 : {
917 0 : rebuild_database_list(InvalidOid);
918 0 : launcher_determine_sleep(canlaunch, true, nap);
919 0 : return;
920 : }
921 :
922 : /* The smallest time we'll allow the launcher to sleep. */
923 2326 : if (nap->tv_sec <= 0 && nap->tv_usec <= MIN_AUTOVAC_SLEEPTIME * 1000)
924 : {
925 4 : nap->tv_sec = 0;
926 4 : nap->tv_usec = MIN_AUTOVAC_SLEEPTIME * 1000;
927 : }
928 :
929 : /*
930 : * If the sleep time is too large, clamp it to an arbitrary maximum (plus
931 : * any fractional seconds, for simplicity). This avoids an essentially
932 : * infinite sleep in strange cases like the system clock going backwards a
933 : * few years.
934 : */
935 2326 : if (nap->tv_sec > MAX_AUTOVAC_SLEEPTIME)
936 0 : nap->tv_sec = MAX_AUTOVAC_SLEEPTIME;
937 : }
938 :
939 : /*
940 : * Build an updated DatabaseList. It must only contain databases that appear
941 : * in pgstats, and must be sorted by next_worker from highest to lowest,
942 : * distributed regularly across the next autovacuum_naptime interval.
943 : *
944 : * Receives the Oid of the database that made this list be generated (we call
945 : * this the "new" database, because when the database was already present on
946 : * the list, we expect that this function is not called at all). The
947 : * preexisting list, if any, will be used to preserve the order of the
948 : * databases in the autovacuum_naptime period. The new database is put at the
949 : * end of the interval. The actual values are not saved, which should not be
950 : * much of a problem.
951 : */
952 : static void
953 550 : rebuild_database_list(Oid newdb)
954 : {
955 : List *dblist;
956 : ListCell *cell;
957 : MemoryContext newcxt;
958 : MemoryContext oldcxt;
959 : MemoryContext tmpcxt;
960 : HASHCTL hctl;
961 : int score;
962 : int nelems;
963 : HTAB *dbhash;
964 : dlist_iter iter;
965 :
966 : /* use fresh stats */
967 550 : autovac_refresh_stats();
968 :
969 550 : newcxt = AllocSetContextCreate(AutovacMemCxt,
970 : "Autovacuum database list",
971 : ALLOCSET_DEFAULT_SIZES);
972 550 : tmpcxt = AllocSetContextCreate(newcxt,
973 : "Autovacuum database list (tmp)",
974 : ALLOCSET_DEFAULT_SIZES);
975 550 : oldcxt = MemoryContextSwitchTo(tmpcxt);
976 :
977 : /*
978 : * Implementing this is not as simple as it sounds, because we need to put
979 : * the new database at the end of the list; next the databases that were
980 : * already on the list, and finally (at the tail of the list) all the
981 : * other databases that are not on the existing list.
982 : *
983 : * To do this, we build an empty hash table of scored databases. We will
984 : * start with the lowest score (zero) for the new database, then
985 : * increasing scores for the databases in the existing list, in order, and
986 : * lastly increasing scores for all databases gotten via
987 : * get_database_list() that are not already on the hash.
988 : *
989 : * Then we will put all the hash elements into an array, sort the array by
990 : * score, and finally put the array elements into the new doubly linked
991 : * list.
992 : */
993 550 : hctl.keysize = sizeof(Oid);
994 550 : hctl.entrysize = sizeof(avl_dbase);
995 550 : hctl.hcxt = tmpcxt;
996 550 : dbhash = hash_create("autovacuum db hash", 20, &hctl, /* magic number here FIXME */
997 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
998 :
999 : /* start by inserting the new database */
1000 550 : score = 0;
1001 550 : if (OidIsValid(newdb))
1002 : {
1003 : avl_dbase *db;
1004 : PgStat_StatDBEntry *entry;
1005 :
1006 : /* only consider this database if it has a pgstat entry */
1007 38 : entry = pgstat_fetch_stat_dbentry(newdb);
1008 38 : if (entry != NULL)
1009 : {
1010 : /* we assume it isn't found because the hash was just created */
1011 38 : db = hash_search(dbhash, &newdb, HASH_ENTER, NULL);
1012 :
1013 : /* hash_search already filled in the key */
1014 38 : db->adl_score = score++;
1015 : /* next_worker is filled in later */
1016 : }
1017 : }
1018 :
1019 : /* Now insert the databases from the existing list */
1020 570 : dlist_foreach(iter, &DatabaseList)
1021 : {
1022 20 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1023 : avl_dbase *db;
1024 : bool found;
1025 : PgStat_StatDBEntry *entry;
1026 :
1027 : /*
1028 : * skip databases with no stat entries -- in particular, this gets rid
1029 : * of dropped databases
1030 : */
1031 20 : entry = pgstat_fetch_stat_dbentry(avdb->adl_datid);
1032 20 : if (entry == NULL)
1033 0 : continue;
1034 :
1035 20 : db = hash_search(dbhash, &(avdb->adl_datid), HASH_ENTER, &found);
1036 :
1037 20 : if (!found)
1038 : {
1039 : /* hash_search already filled in the key */
1040 20 : db->adl_score = score++;
1041 : /* next_worker is filled in later */
1042 : }
1043 : }
1044 :
1045 : /* finally, insert all qualifying databases not previously inserted */
1046 550 : dblist = get_database_list();
1047 2408 : foreach(cell, dblist)
1048 : {
1049 1858 : avw_dbase *avdb = lfirst(cell);
1050 : avl_dbase *db;
1051 : bool found;
1052 : PgStat_StatDBEntry *entry;
1053 :
1054 : /* only consider databases with a pgstat entry */
1055 1858 : entry = pgstat_fetch_stat_dbentry(avdb->adw_datid);
1056 1858 : if (entry == NULL)
1057 1598 : continue;
1058 :
1059 260 : db = hash_search(dbhash, &(avdb->adw_datid), HASH_ENTER, &found);
1060 : /* only update the score if the database was not already on the hash */
1061 260 : if (!found)
1062 : {
1063 : /* hash_search already filled in the key */
1064 202 : db->adl_score = score++;
1065 : /* next_worker is filled in later */
1066 : }
1067 : }
1068 550 : nelems = score;
1069 :
1070 : /* from here on, the allocated memory belongs to the new list */
1071 550 : MemoryContextSwitchTo(newcxt);
1072 550 : dlist_init(&DatabaseList);
1073 :
1074 550 : if (nelems > 0)
1075 : {
1076 : TimestampTz current_time;
1077 : int millis_increment;
1078 : avl_dbase *dbary;
1079 : avl_dbase *db;
1080 : HASH_SEQ_STATUS seq;
1081 : int i;
1082 :
1083 : /* put all the hash elements into an array */
1084 172 : dbary = palloc(nelems * sizeof(avl_dbase));
1085 :
1086 172 : i = 0;
1087 172 : hash_seq_init(&seq, dbhash);
1088 432 : while ((db = hash_seq_search(&seq)) != NULL)
1089 260 : memcpy(&(dbary[i++]), db, sizeof(avl_dbase));
1090 :
1091 : /* sort the array */
1092 172 : qsort(dbary, nelems, sizeof(avl_dbase), db_comparator);
1093 :
1094 : /*
1095 : * Determine the time interval between databases in the schedule. If
1096 : * we see that the configured naptime would take us to sleep times
1097 : * lower than our min sleep time (which launcher_determine_sleep is
1098 : * coded not to allow), silently use a larger naptime (but don't touch
1099 : * the GUC variable).
1100 : */
1101 172 : millis_increment = 1000.0 * autovacuum_naptime / nelems;
1102 172 : if (millis_increment <= MIN_AUTOVAC_SLEEPTIME)
1103 0 : millis_increment = MIN_AUTOVAC_SLEEPTIME * 1.1;
1104 :
1105 172 : current_time = GetCurrentTimestamp();
1106 :
1107 : /*
1108 : * move the elements from the array into the dlist, setting the
1109 : * next_worker while walking the array
1110 : */
1111 432 : for (i = 0; i < nelems; i++)
1112 : {
1113 260 : avl_dbase *db = &(dbary[i]);
1114 :
1115 260 : current_time = TimestampTzPlusMilliseconds(current_time,
1116 : millis_increment);
1117 260 : db->adl_next_worker = current_time;
1118 :
1119 : /* later elements should go closer to the head of the list */
1120 260 : dlist_push_head(&DatabaseList, &db->adl_node);
1121 : }
1122 : }
1123 :
1124 : /* all done, clean up memory */
1125 550 : if (DatabaseListCxt != NULL)
1126 72 : MemoryContextDelete(DatabaseListCxt);
1127 550 : MemoryContextDelete(tmpcxt);
1128 550 : DatabaseListCxt = newcxt;
1129 550 : MemoryContextSwitchTo(oldcxt);
1130 550 : }
1131 :
1132 : /* qsort comparator for avl_dbase, using adl_score */
1133 : static int
1134 150 : db_comparator(const void *a, const void *b)
1135 : {
1136 150 : if (((const avl_dbase *) a)->adl_score == ((const avl_dbase *) b)->adl_score)
1137 0 : return 0;
1138 : else
1139 150 : return (((const avl_dbase *) a)->adl_score < ((const avl_dbase *) b)->adl_score) ? 1 : -1;
1140 : }
1141 :
1142 : /*
1143 : * do_start_worker
1144 : *
1145 : * Bare-bones procedure for starting an autovacuum worker from the launcher.
1146 : * It determines what database to work on, sets up shared memory stuff and
1147 : * signals postmaster to start the worker. It fails gracefully if invoked when
1148 : * autovacuum_workers are already active.
1149 : *
1150 : * Return value is the OID of the database that the worker is going to process,
1151 : * or InvalidOid if no worker was actually started.
1152 : */
1153 : static Oid
1154 440 : do_start_worker(void)
1155 : {
1156 : List *dblist;
1157 : ListCell *cell;
1158 : TransactionId xidForceLimit;
1159 : MultiXactId multiForceLimit;
1160 : bool for_xid_wrap;
1161 : bool for_multi_wrap;
1162 : avw_dbase *avdb;
1163 : TimestampTz current_time;
1164 440 : bool skipit = false;
1165 440 : Oid retval = InvalidOid;
1166 : MemoryContext tmpcxt,
1167 : oldcxt;
1168 :
1169 : /* return quickly when there are no free workers */
1170 440 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1171 440 : if (dlist_is_empty(&AutoVacuumShmem->av_freeWorkers))
1172 : {
1173 0 : LWLockRelease(AutovacuumLock);
1174 0 : return InvalidOid;
1175 : }
1176 440 : LWLockRelease(AutovacuumLock);
1177 :
1178 : /*
1179 : * Create and switch to a temporary context to avoid leaking the memory
1180 : * allocated for the database list.
1181 : */
1182 440 : tmpcxt = AllocSetContextCreate(CurrentMemoryContext,
1183 : "Autovacuum start worker (tmp)",
1184 : ALLOCSET_DEFAULT_SIZES);
1185 440 : oldcxt = MemoryContextSwitchTo(tmpcxt);
1186 :
1187 : /* use fresh stats */
1188 440 : autovac_refresh_stats();
1189 :
1190 : /* Get a list of databases */
1191 440 : dblist = get_database_list();
1192 :
1193 : /*
1194 : * Determine the oldest datfrozenxid/relfrozenxid that we will allow to
1195 : * pass without forcing a vacuum. (This limit can be tightened for
1196 : * particular tables, but not loosened.)
1197 : */
1198 440 : recentXid = ReadNextTransactionId();
1199 440 : xidForceLimit = recentXid - autovacuum_freeze_max_age;
1200 : /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */
1201 : /* this can cause the limit to go backwards by 3, but that's OK */
1202 440 : if (xidForceLimit < FirstNormalTransactionId)
1203 0 : xidForceLimit -= FirstNormalTransactionId;
1204 :
1205 : /* Also determine the oldest datminmxid we will consider. */
1206 440 : recentMulti = ReadNextMultiXactId();
1207 440 : multiForceLimit = recentMulti - MultiXactMemberFreezeThreshold();
1208 440 : if (multiForceLimit < FirstMultiXactId)
1209 0 : multiForceLimit -= FirstMultiXactId;
1210 :
1211 : /*
1212 : * Choose a database to connect to. We pick the database that was least
1213 : * recently auto-vacuumed, or one that needs vacuuming to prevent Xid
1214 : * wraparound-related data loss. If any db at risk of Xid wraparound is
1215 : * found, we pick the one with oldest datfrozenxid, independently of
1216 : * autovacuum times; similarly we pick the one with the oldest datminmxid
1217 : * if any is in MultiXactId wraparound. Note that those in Xid wraparound
1218 : * danger are given more priority than those in multi wraparound danger.
1219 : *
1220 : * Note that a database with no stats entry is not considered, except for
1221 : * Xid wraparound purposes. The theory is that if no one has ever
1222 : * connected to it since the stats were last initialized, it doesn't need
1223 : * vacuuming.
1224 : *
1225 : * XXX This could be improved if we had more info about whether it needs
1226 : * vacuuming before connecting to it. Perhaps look through the pgstats
1227 : * data for the database's tables? One idea is to keep track of the
1228 : * number of new and dead tuples per database in pgstats. However it
1229 : * isn't clear how to construct a metric that measures that and not cause
1230 : * starvation for less busy databases.
1231 : */
1232 440 : avdb = NULL;
1233 440 : for_xid_wrap = false;
1234 440 : for_multi_wrap = false;
1235 440 : current_time = GetCurrentTimestamp();
1236 1838 : foreach(cell, dblist)
1237 : {
1238 1398 : avw_dbase *tmp = lfirst(cell);
1239 : dlist_iter iter;
1240 :
1241 : /* Check to see if this one is at risk of wraparound */
1242 1398 : if (TransactionIdPrecedes(tmp->adw_frozenxid, xidForceLimit))
1243 : {
1244 0 : if (avdb == NULL ||
1245 0 : TransactionIdPrecedes(tmp->adw_frozenxid,
1246 : avdb->adw_frozenxid))
1247 0 : avdb = tmp;
1248 0 : for_xid_wrap = true;
1249 1312 : continue;
1250 : }
1251 1398 : else if (for_xid_wrap)
1252 0 : continue; /* ignore not-at-risk DBs */
1253 1398 : else if (MultiXactIdPrecedes(tmp->adw_minmulti, multiForceLimit))
1254 : {
1255 0 : if (avdb == NULL ||
1256 0 : MultiXactIdPrecedes(tmp->adw_minmulti, avdb->adw_minmulti))
1257 0 : avdb = tmp;
1258 0 : for_multi_wrap = true;
1259 0 : continue;
1260 : }
1261 1398 : else if (for_multi_wrap)
1262 0 : continue; /* ignore not-at-risk DBs */
1263 :
1264 : /* Find pgstat entry if any */
1265 1398 : tmp->adw_entry = pgstat_fetch_stat_dbentry(tmp->adw_datid);
1266 :
1267 : /*
1268 : * Skip a database with no pgstat entry; it means it hasn't seen any
1269 : * activity.
1270 : */
1271 1398 : if (!tmp->adw_entry)
1272 1300 : continue;
1273 :
1274 : /*
1275 : * Also, skip a database that appears on the database list as having
1276 : * been processed recently (less than autovacuum_naptime seconds ago).
1277 : * We do this so that we don't select a database which we just
1278 : * selected, but that pgstat hasn't gotten around to updating the last
1279 : * autovacuum time yet.
1280 : */
1281 98 : skipit = false;
1282 :
1283 110 : dlist_reverse_foreach(iter, &DatabaseList)
1284 : {
1285 36 : avl_dbase *dbp = dlist_container(avl_dbase, adl_node, iter.cur);
1286 :
1287 36 : if (dbp->adl_datid == tmp->adw_datid)
1288 : {
1289 : /*
1290 : * Skip this database if its next_worker value falls between
1291 : * the current time and the current time plus naptime.
1292 : */
1293 24 : if (!TimestampDifferenceExceeds(dbp->adl_next_worker,
1294 12 : current_time, 0) &&
1295 12 : !TimestampDifferenceExceeds(current_time,
1296 : dbp->adl_next_worker,
1297 : autovacuum_naptime * 1000))
1298 12 : skipit = true;
1299 :
1300 24 : break;
1301 : }
1302 : }
1303 98 : if (skipit)
1304 12 : continue;
1305 :
1306 : /*
1307 : * Remember the db with oldest autovac time. (If we are here, both
1308 : * tmp->entry and db->entry must be non-null.)
1309 : */
1310 86 : if (avdb == NULL ||
1311 36 : tmp->adw_entry->last_autovac_time < avdb->adw_entry->last_autovac_time)
1312 50 : avdb = tmp;
1313 : }
1314 :
1315 : /* Found a database -- process it */
1316 440 : if (avdb != NULL)
1317 : {
1318 : WorkerInfo worker;
1319 : dlist_node *wptr;
1320 :
1321 50 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1322 :
1323 : /*
1324 : * Get a worker entry from the freelist. We checked above, so there
1325 : * really should be a free slot.
1326 : */
1327 50 : wptr = dlist_pop_head_node(&AutoVacuumShmem->av_freeWorkers);
1328 :
1329 50 : worker = dlist_container(WorkerInfoData, wi_links, wptr);
1330 50 : worker->wi_dboid = avdb->adw_datid;
1331 50 : worker->wi_proc = NULL;
1332 50 : worker->wi_launchtime = GetCurrentTimestamp();
1333 :
1334 50 : AutoVacuumShmem->av_startingWorker = worker;
1335 :
1336 50 : LWLockRelease(AutovacuumLock);
1337 :
1338 50 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
1339 :
1340 50 : retval = avdb->adw_datid;
1341 : }
1342 390 : else if (skipit)
1343 : {
1344 : /*
1345 : * If we skipped all databases on the list, rebuild it, because it
1346 : * probably contains a dropped database.
1347 : */
1348 0 : rebuild_database_list(InvalidOid);
1349 : }
1350 :
1351 440 : MemoryContextSwitchTo(oldcxt);
1352 440 : MemoryContextDelete(tmpcxt);
1353 :
1354 440 : return retval;
1355 : }
1356 :
1357 : /*
1358 : * launch_worker
1359 : *
1360 : * Wrapper for starting a worker from the launcher. Besides actually starting
1361 : * it, update the database list to reflect the next time that another one will
1362 : * need to be started on the selected database. The actual database choice is
1363 : * left to do_start_worker.
1364 : *
1365 : * This routine is also expected to insert an entry into the database list if
1366 : * the selected database was previously absent from the list.
1367 : */
1368 : static void
1369 440 : launch_worker(TimestampTz now)
1370 : {
1371 : Oid dbid;
1372 : dlist_iter iter;
1373 :
1374 440 : dbid = do_start_worker();
1375 440 : if (OidIsValid(dbid))
1376 : {
1377 50 : bool found = false;
1378 :
1379 : /*
1380 : * Walk the database list and update the corresponding entry. If the
1381 : * database is not on the list, we'll recreate the list.
1382 : */
1383 62 : dlist_foreach(iter, &DatabaseList)
1384 : {
1385 24 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1386 :
1387 24 : if (avdb->adl_datid == dbid)
1388 : {
1389 12 : found = true;
1390 :
1391 : /*
1392 : * add autovacuum_naptime seconds to the current time, and use
1393 : * that as the new "next_worker" field for this database.
1394 : */
1395 12 : avdb->adl_next_worker =
1396 12 : TimestampTzPlusMilliseconds(now, autovacuum_naptime * 1000);
1397 :
1398 12 : dlist_move_head(&DatabaseList, iter.cur);
1399 12 : break;
1400 : }
1401 : }
1402 :
1403 : /*
1404 : * If the database was not present in the database list, we rebuild
1405 : * the list. It's possible that the database does not get into the
1406 : * list anyway, for example if it's a database that doesn't have a
1407 : * pgstat entry, but this is not a problem because we don't want to
1408 : * schedule workers regularly into those in any case.
1409 : */
1410 50 : if (!found)
1411 38 : rebuild_database_list(dbid);
1412 : }
1413 440 : }
1414 :
1415 : /*
1416 : * Called from postmaster to signal a failure to fork a process to become
1417 : * worker. The postmaster should kill(SIGUSR2) the launcher shortly
1418 : * after calling this function.
1419 : */
1420 : void
1421 0 : AutoVacWorkerFailed(void)
1422 : {
1423 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = true;
1424 0 : }
1425 :
1426 : /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */
1427 : static void
1428 96 : avl_sigusr2_handler(SIGNAL_ARGS)
1429 : {
1430 96 : int save_errno = errno;
1431 :
1432 96 : got_SIGUSR2 = true;
1433 96 : SetLatch(MyLatch);
1434 :
1435 96 : errno = save_errno;
1436 96 : }
1437 :
1438 :
1439 : /********************************************************************
1440 : * AUTOVACUUM WORKER CODE
1441 : ********************************************************************/
1442 :
1443 : #ifdef EXEC_BACKEND
1444 : /*
1445 : * forkexec routines for the autovacuum worker.
1446 : *
1447 : * Format up the arglist, then fork and exec.
1448 : */
1449 : static pid_t
1450 : avworker_forkexec(void)
1451 : {
1452 : char *av[10];
1453 : int ac = 0;
1454 :
1455 : av[ac++] = "postgres";
1456 : av[ac++] = "--forkavworker";
1457 : av[ac++] = NULL; /* filled in by postmaster_forkexec */
1458 : av[ac] = NULL;
1459 :
1460 : Assert(ac < lengthof(av));
1461 :
1462 : return postmaster_forkexec(ac, av);
1463 : }
1464 :
1465 : /*
1466 : * We need this set from the outside, before InitProcess is called
1467 : */
1468 : void
1469 : AutovacuumWorkerIAm(void)
1470 : {
1471 : am_autovacuum_worker = true;
1472 : }
1473 : #endif
1474 :
1475 : /*
1476 : * Main entry point for autovacuum worker process.
1477 : *
1478 : * This code is heavily based on pgarch.c, q.v.
1479 : */
1480 : int
1481 58 : StartAutoVacWorker(void)
1482 : {
1483 : pid_t worker_pid;
1484 :
1485 : #ifdef EXEC_BACKEND
1486 : switch ((worker_pid = avworker_forkexec()))
1487 : #else
1488 58 : switch ((worker_pid = fork_process()))
1489 : #endif
1490 : {
1491 0 : case -1:
1492 0 : ereport(LOG,
1493 : (errmsg("could not fork autovacuum worker process: %m")));
1494 0 : return 0;
1495 :
1496 : #ifndef EXEC_BACKEND
1497 58 : case 0:
1498 : /* in postmaster child ... */
1499 58 : InitPostmasterChild();
1500 :
1501 : /* Close the postmaster's sockets */
1502 58 : ClosePostmasterPorts(false);
1503 :
1504 58 : AutoVacWorkerMain(0, NULL);
1505 : break;
1506 : #endif
1507 58 : default:
1508 58 : return (int) worker_pid;
1509 : }
1510 :
1511 : /* shouldn't get here */
1512 : return 0;
1513 : }
1514 :
1515 : /*
1516 : * AutoVacWorkerMain
1517 : */
1518 : NON_EXEC_STATIC void
1519 58 : AutoVacWorkerMain(int argc, char *argv[])
1520 : {
1521 : sigjmp_buf local_sigjmp_buf;
1522 : Oid dbid;
1523 :
1524 58 : am_autovacuum_worker = true;
1525 :
1526 58 : MyBackendType = B_AUTOVAC_WORKER;
1527 58 : init_ps_display(NULL);
1528 :
1529 58 : SetProcessingMode(InitProcessing);
1530 :
1531 : /*
1532 : * Set up signal handlers. We operate on databases much like a regular
1533 : * backend, so we use the same signal handling. See equivalent code in
1534 : * tcop/postgres.c.
1535 : */
1536 58 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1537 :
1538 : /*
1539 : * SIGINT is used to signal canceling the current table's vacuum; SIGTERM
1540 : * means abort and exit cleanly, and SIGQUIT means abandon ship.
1541 : */
1542 58 : pqsignal(SIGINT, StatementCancelHandler);
1543 58 : pqsignal(SIGTERM, die);
1544 : /* SIGQUIT handler was already set up by InitPostmasterChild */
1545 :
1546 58 : InitializeTimeouts(); /* establishes SIGALRM handler */
1547 :
1548 58 : pqsignal(SIGPIPE, SIG_IGN);
1549 58 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1550 58 : pqsignal(SIGUSR2, SIG_IGN);
1551 58 : pqsignal(SIGFPE, FloatExceptionHandler);
1552 58 : pqsignal(SIGCHLD, SIG_DFL);
1553 :
1554 : /*
1555 : * Create a per-backend PGPROC struct in shared memory, except in the
1556 : * EXEC_BACKEND case where this was done in SubPostmasterMain. We must do
1557 : * this before we can use LWLocks (and in the EXEC_BACKEND case we already
1558 : * had to do some stuff with LWLocks).
1559 : */
1560 : #ifndef EXEC_BACKEND
1561 58 : InitProcess();
1562 : #endif
1563 :
1564 : /* Early initialization */
1565 58 : BaseInit();
1566 :
1567 : /*
1568 : * If an exception is encountered, processing resumes here.
1569 : *
1570 : * Unlike most auxiliary processes, we don't attempt to continue
1571 : * processing after an error; we just clean up and exit. The autovac
1572 : * launcher is responsible for spawning another worker later.
1573 : *
1574 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
1575 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
1576 : * signals other than SIGQUIT will be blocked until we exit. It might
1577 : * seem that this policy makes the HOLD_INTERRUPTS() call redundant, but
1578 : * it is not since InterruptPending might be set already.
1579 : */
1580 58 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1581 : {
1582 : /* since not using PG_TRY, must reset error stack by hand */
1583 0 : error_context_stack = NULL;
1584 :
1585 : /* Prevents interrupts while cleaning up */
1586 0 : HOLD_INTERRUPTS();
1587 :
1588 : /* Report the error to the server log */
1589 0 : EmitErrorReport();
1590 :
1591 : /*
1592 : * We can now go away. Note that because we called InitProcess, a
1593 : * callback was registered to do ProcKill, which will clean up
1594 : * necessary state.
1595 : */
1596 0 : proc_exit(0);
1597 : }
1598 :
1599 : /* We can now handle ereport(ERROR) */
1600 58 : PG_exception_stack = &local_sigjmp_buf;
1601 :
1602 58 : PG_SETMASK(&UnBlockSig);
1603 :
1604 : /*
1605 : * Set always-secure search path, so malicious users can't redirect user
1606 : * code (e.g. pg_index.indexprs). (That code runs in a
1607 : * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not
1608 : * take control of the entire autovacuum worker in any case.)
1609 : */
1610 58 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1611 :
1612 : /*
1613 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
1614 : * in postgresql.conf. We don't really want such a dangerous option being
1615 : * applied non-interactively.
1616 : */
1617 58 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
1618 :
1619 : /*
1620 : * Force settable timeouts off to avoid letting these settings prevent
1621 : * regular maintenance from being executed.
1622 : */
1623 58 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1624 58 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1625 58 : SetConfigOption("idle_in_transaction_session_timeout", "0",
1626 : PGC_SUSET, PGC_S_OVERRIDE);
1627 :
1628 : /*
1629 : * Force default_transaction_isolation to READ COMMITTED. We don't want
1630 : * to pay the overhead of serializable mode, nor add any risk of causing
1631 : * deadlocks or delaying other transactions.
1632 : */
1633 58 : SetConfigOption("default_transaction_isolation", "read committed",
1634 : PGC_SUSET, PGC_S_OVERRIDE);
1635 :
1636 : /*
1637 : * Force synchronous replication off to allow regular maintenance even if
1638 : * we are waiting for standbys to connect. This is important to ensure we
1639 : * aren't blocked from performing anti-wraparound tasks.
1640 : */
1641 58 : if (synchronous_commit > SYNCHRONOUS_COMMIT_LOCAL_FLUSH)
1642 58 : SetConfigOption("synchronous_commit", "local",
1643 : PGC_SUSET, PGC_S_OVERRIDE);
1644 :
1645 : /*
1646 : * Get the info about the database we're going to work on.
1647 : */
1648 58 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1649 :
1650 : /*
1651 : * beware of startingWorker being INVALID; this should normally not
1652 : * happen, but if a worker fails after forking and before this, the
1653 : * launcher might have decided to remove it from the queue and start
1654 : * again.
1655 : */
1656 58 : if (AutoVacuumShmem->av_startingWorker != NULL)
1657 : {
1658 58 : MyWorkerInfo = AutoVacuumShmem->av_startingWorker;
1659 58 : dbid = MyWorkerInfo->wi_dboid;
1660 58 : MyWorkerInfo->wi_proc = MyProc;
1661 :
1662 : /* insert into the running list */
1663 58 : dlist_push_head(&AutoVacuumShmem->av_runningWorkers,
1664 58 : &MyWorkerInfo->wi_links);
1665 :
1666 : /*
1667 : * remove from the "starting" pointer, so that the launcher can start
1668 : * a new worker if required
1669 : */
1670 58 : AutoVacuumShmem->av_startingWorker = NULL;
1671 58 : LWLockRelease(AutovacuumLock);
1672 :
1673 58 : on_shmem_exit(FreeWorkerInfo, 0);
1674 :
1675 : /* wake up the launcher */
1676 58 : if (AutoVacuumShmem->av_launcherpid != 0)
1677 56 : kill(AutoVacuumShmem->av_launcherpid, SIGUSR2);
1678 : }
1679 : else
1680 : {
1681 : /* no worker entry for me, go away */
1682 0 : elog(WARNING, "autovacuum worker started without a worker entry");
1683 0 : dbid = InvalidOid;
1684 0 : LWLockRelease(AutovacuumLock);
1685 : }
1686 :
1687 58 : if (OidIsValid(dbid))
1688 : {
1689 : char dbname[NAMEDATALEN];
1690 :
1691 : /*
1692 : * Report autovac startup to the stats collector. We deliberately do
1693 : * this before InitPostgres, so that the last_autovac_time will get
1694 : * updated even if the connection attempt fails. This is to prevent
1695 : * autovac from getting "stuck" repeatedly selecting an unopenable
1696 : * database, rather than making any progress on stuff it can connect
1697 : * to.
1698 : */
1699 58 : pgstat_report_autovac(dbid);
1700 :
1701 : /*
1702 : * Connect to the selected database
1703 : *
1704 : * Note: if we have selected a just-deleted database (due to using
1705 : * stale stats info), we'll fail and exit here.
1706 : */
1707 58 : InitPostgres(NULL, dbid, NULL, InvalidOid, dbname, false);
1708 56 : SetProcessingMode(NormalProcessing);
1709 56 : set_ps_display(dbname);
1710 56 : ereport(DEBUG1,
1711 : (errmsg_internal("autovacuum: processing database \"%s\"", dbname)));
1712 :
1713 56 : if (PostAuthDelay)
1714 0 : pg_usleep(PostAuthDelay * 1000000L);
1715 :
1716 : /* And do an appropriate amount of work */
1717 56 : recentXid = ReadNextTransactionId();
1718 56 : recentMulti = ReadNextMultiXactId();
1719 56 : do_autovacuum();
1720 : }
1721 :
1722 : /*
1723 : * The launcher will be notified of my death in ProcKill, *if* we managed
1724 : * to get a worker slot at all
1725 : */
1726 :
1727 : /* All done, go away */
1728 56 : proc_exit(0);
1729 : }
1730 :
1731 : /*
1732 : * Return a WorkerInfo to the free list
1733 : */
1734 : static void
1735 58 : FreeWorkerInfo(int code, Datum arg)
1736 : {
1737 58 : if (MyWorkerInfo != NULL)
1738 : {
1739 58 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1740 :
1741 : /*
1742 : * Wake the launcher up so that he can launch a new worker immediately
1743 : * if required. We only save the launcher's PID in local memory here;
1744 : * the actual signal will be sent when the PGPROC is recycled. Note
1745 : * that we always do this, so that the launcher can rebalance the cost
1746 : * limit setting of the remaining workers.
1747 : *
1748 : * We somewhat ignore the risk that the launcher changes its PID
1749 : * between us reading it and the actual kill; we expect ProcKill to be
1750 : * called shortly after us, and we assume that PIDs are not reused too
1751 : * quickly after a process exits.
1752 : */
1753 58 : AutovacuumLauncherPid = AutoVacuumShmem->av_launcherpid;
1754 :
1755 58 : dlist_delete(&MyWorkerInfo->wi_links);
1756 58 : MyWorkerInfo->wi_dboid = InvalidOid;
1757 58 : MyWorkerInfo->wi_tableoid = InvalidOid;
1758 58 : MyWorkerInfo->wi_sharedrel = false;
1759 58 : MyWorkerInfo->wi_proc = NULL;
1760 58 : MyWorkerInfo->wi_launchtime = 0;
1761 58 : MyWorkerInfo->wi_dobalance = false;
1762 58 : MyWorkerInfo->wi_cost_delay = 0;
1763 58 : MyWorkerInfo->wi_cost_limit = 0;
1764 58 : MyWorkerInfo->wi_cost_limit_base = 0;
1765 58 : dlist_push_head(&AutoVacuumShmem->av_freeWorkers,
1766 58 : &MyWorkerInfo->wi_links);
1767 : /* not mine anymore */
1768 58 : MyWorkerInfo = NULL;
1769 :
1770 : /*
1771 : * now that we're inactive, cause a rebalancing of the surviving
1772 : * workers
1773 : */
1774 58 : AutoVacuumShmem->av_signal[AutoVacRebalance] = true;
1775 58 : LWLockRelease(AutovacuumLock);
1776 : }
1777 58 : }
1778 :
1779 : /*
1780 : * Update the cost-based delay parameters, so that multiple workers consume
1781 : * each a fraction of the total available I/O.
1782 : */
1783 : void
1784 586 : AutoVacuumUpdateDelay(void)
1785 : {
1786 586 : if (MyWorkerInfo)
1787 : {
1788 586 : VacuumCostDelay = MyWorkerInfo->wi_cost_delay;
1789 586 : VacuumCostLimit = MyWorkerInfo->wi_cost_limit;
1790 : }
1791 586 : }
1792 :
1793 : /*
1794 : * autovac_balance_cost
1795 : * Recalculate the cost limit setting for each active worker.
1796 : *
1797 : * Caller must hold the AutovacuumLock in exclusive mode.
1798 : */
1799 : static void
1800 338 : autovac_balance_cost(void)
1801 : {
1802 : /*
1803 : * The idea here is that we ration out I/O equally. The amount of I/O
1804 : * that a worker can consume is determined by cost_limit/cost_delay, so we
1805 : * try to equalize those ratios rather than the raw limit settings.
1806 : *
1807 : * note: in cost_limit, zero also means use value from elsewhere, because
1808 : * zero is not a valid value.
1809 : */
1810 676 : int vac_cost_limit = (autovacuum_vac_cost_limit > 0 ?
1811 338 : autovacuum_vac_cost_limit : VacuumCostLimit);
1812 676 : double vac_cost_delay = (autovacuum_vac_cost_delay >= 0 ?
1813 338 : autovacuum_vac_cost_delay : VacuumCostDelay);
1814 : double cost_total;
1815 : double cost_avail;
1816 : dlist_iter iter;
1817 :
1818 : /* not set? nothing to do */
1819 338 : if (vac_cost_limit <= 0 || vac_cost_delay <= 0)
1820 82 : return;
1821 :
1822 : /* calculate the total base cost limit of participating active workers */
1823 338 : cost_total = 0.0;
1824 594 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1825 : {
1826 256 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1827 :
1828 256 : if (worker->wi_proc != NULL &&
1829 256 : worker->wi_dobalance &&
1830 256 : worker->wi_cost_limit_base > 0 && worker->wi_cost_delay > 0)
1831 256 : cost_total +=
1832 256 : (double) worker->wi_cost_limit_base / worker->wi_cost_delay;
1833 : }
1834 :
1835 : /* there are no cost limits -- nothing to do */
1836 338 : if (cost_total <= 0)
1837 82 : return;
1838 :
1839 : /*
1840 : * Adjust cost limit of each active worker to balance the total of cost
1841 : * limit to autovacuum_vacuum_cost_limit.
1842 : */
1843 256 : cost_avail = (double) vac_cost_limit / vac_cost_delay;
1844 512 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1845 : {
1846 256 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1847 :
1848 256 : if (worker->wi_proc != NULL &&
1849 256 : worker->wi_dobalance &&
1850 256 : worker->wi_cost_limit_base > 0 && worker->wi_cost_delay > 0)
1851 : {
1852 256 : int limit = (int)
1853 256 : (cost_avail * worker->wi_cost_limit_base / cost_total);
1854 :
1855 : /*
1856 : * We put a lower bound of 1 on the cost_limit, to avoid division-
1857 : * by-zero in the vacuum code. Also, in case of roundoff trouble
1858 : * in these calculations, let's be sure we don't ever set
1859 : * cost_limit to more than the base value.
1860 : */
1861 256 : worker->wi_cost_limit = Max(Min(limit,
1862 : worker->wi_cost_limit_base),
1863 : 1);
1864 : }
1865 :
1866 256 : if (worker->wi_proc != NULL)
1867 256 : elog(DEBUG2, "autovac_balance_cost(pid=%d db=%u, rel=%u, dobalance=%s cost_limit=%d, cost_limit_base=%d, cost_delay=%g)",
1868 : worker->wi_proc->pid, worker->wi_dboid, worker->wi_tableoid,
1869 : worker->wi_dobalance ? "yes" : "no",
1870 : worker->wi_cost_limit, worker->wi_cost_limit_base,
1871 : worker->wi_cost_delay);
1872 : }
1873 : }
1874 :
1875 : /*
1876 : * get_database_list
1877 : * Return a list of all databases found in pg_database.
1878 : *
1879 : * The list and associated data is allocated in the caller's memory context,
1880 : * which is in charge of ensuring that it's properly cleaned up afterwards.
1881 : *
1882 : * Note: this is the only function in which the autovacuum launcher uses a
1883 : * transaction. Although we aren't attached to any particular database and
1884 : * therefore can't access most catalogs, we do have enough infrastructure
1885 : * to do a seqscan on pg_database.
1886 : */
1887 : static List *
1888 990 : get_database_list(void)
1889 : {
1890 990 : List *dblist = NIL;
1891 : Relation rel;
1892 : TableScanDesc scan;
1893 : HeapTuple tup;
1894 : MemoryContext resultcxt;
1895 :
1896 : /* This is the context that we will allocate our output data in */
1897 990 : resultcxt = CurrentMemoryContext;
1898 :
1899 : /*
1900 : * Start a transaction so we can access pg_database, and get a snapshot.
1901 : * We don't have a use for the snapshot itself, but we're interested in
1902 : * the secondary effect that it sets RecentGlobalXmin. (This is critical
1903 : * for anything that reads heap pages, because HOT may decide to prune
1904 : * them even if the process doesn't attempt to modify any tuples.)
1905 : *
1906 : * FIXME: This comment is inaccurate / the code buggy. A snapshot that is
1907 : * not pushed/active does not reliably prevent HOT pruning (->xmin could
1908 : * e.g. be cleared when cache invalidations are processed).
1909 : */
1910 990 : StartTransactionCommand();
1911 990 : (void) GetTransactionSnapshot();
1912 :
1913 990 : rel = table_open(DatabaseRelationId, AccessShareLock);
1914 990 : scan = table_beginscan_catalog(rel, 0, NULL);
1915 :
1916 4246 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1917 : {
1918 3256 : Form_pg_database pgdatabase = (Form_pg_database) GETSTRUCT(tup);
1919 : avw_dbase *avdb;
1920 : MemoryContext oldcxt;
1921 :
1922 : /*
1923 : * Allocate our results in the caller's context, not the
1924 : * transaction's. We do this inside the loop, and restore the original
1925 : * context at the end, so that leaky things like heap_getnext() are
1926 : * not called in a potentially long-lived context.
1927 : */
1928 3256 : oldcxt = MemoryContextSwitchTo(resultcxt);
1929 :
1930 3256 : avdb = (avw_dbase *) palloc(sizeof(avw_dbase));
1931 :
1932 3256 : avdb->adw_datid = pgdatabase->oid;
1933 3256 : avdb->adw_name = pstrdup(NameStr(pgdatabase->datname));
1934 3256 : avdb->adw_frozenxid = pgdatabase->datfrozenxid;
1935 3256 : avdb->adw_minmulti = pgdatabase->datminmxid;
1936 : /* this gets set later: */
1937 3256 : avdb->adw_entry = NULL;
1938 :
1939 3256 : dblist = lappend(dblist, avdb);
1940 3256 : MemoryContextSwitchTo(oldcxt);
1941 : }
1942 :
1943 990 : table_endscan(scan);
1944 990 : table_close(rel, AccessShareLock);
1945 :
1946 990 : CommitTransactionCommand();
1947 :
1948 990 : return dblist;
1949 : }
1950 :
1951 : /*
1952 : * Process a database table-by-table
1953 : *
1954 : * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in
1955 : * order not to ignore shutdown commands for too long.
1956 : */
1957 : static void
1958 56 : do_autovacuum(void)
1959 : {
1960 : Relation classRel;
1961 : HeapTuple tuple;
1962 : TableScanDesc relScan;
1963 : Form_pg_database dbForm;
1964 56 : List *table_oids = NIL;
1965 56 : List *orphan_oids = NIL;
1966 : HASHCTL ctl;
1967 : HTAB *table_toast_map;
1968 : ListCell *volatile cell;
1969 : PgStat_StatDBEntry *shared;
1970 : PgStat_StatDBEntry *dbentry;
1971 : BufferAccessStrategy bstrategy;
1972 : ScanKeyData key;
1973 : TupleDesc pg_class_desc;
1974 : int effective_multixact_freeze_max_age;
1975 56 : bool did_vacuum = false;
1976 56 : bool found_concurrent_worker = false;
1977 : int i;
1978 :
1979 : /*
1980 : * StartTransactionCommand and CommitTransactionCommand will automatically
1981 : * switch to other contexts. We need this one to keep the list of
1982 : * relations to vacuum/analyze across transactions.
1983 : */
1984 56 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
1985 : "Autovacuum worker",
1986 : ALLOCSET_DEFAULT_SIZES);
1987 56 : MemoryContextSwitchTo(AutovacMemCxt);
1988 :
1989 : /*
1990 : * may be NULL if we couldn't find an entry (only happens if we are
1991 : * forcing a vacuum for anti-wrap purposes).
1992 : */
1993 56 : dbentry = pgstat_fetch_stat_dbentry(MyDatabaseId);
1994 :
1995 : /* Start a transaction so our commands have one to play into. */
1996 56 : StartTransactionCommand();
1997 :
1998 : /*
1999 : * Clean up any dead statistics collector entries for this DB. We always
2000 : * want to do this exactly once per DB-processing cycle, even if we find
2001 : * nothing worth vacuuming in the database.
2002 : */
2003 56 : pgstat_vacuum_stat();
2004 :
2005 : /*
2006 : * Compute the multixact age for which freezing is urgent. This is
2007 : * normally autovacuum_multixact_freeze_max_age, but may be less if we are
2008 : * short of multixact member space.
2009 : */
2010 56 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
2011 :
2012 : /*
2013 : * Find the pg_database entry and select the default freeze ages. We use
2014 : * zero in template and nonconnectable databases, else the system-wide
2015 : * default.
2016 : */
2017 56 : tuple = SearchSysCache1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
2018 56 : if (!HeapTupleIsValid(tuple))
2019 0 : elog(ERROR, "cache lookup failed for database %u", MyDatabaseId);
2020 56 : dbForm = (Form_pg_database) GETSTRUCT(tuple);
2021 :
2022 56 : if (dbForm->datistemplate || !dbForm->datallowconn)
2023 : {
2024 0 : default_freeze_min_age = 0;
2025 0 : default_freeze_table_age = 0;
2026 0 : default_multixact_freeze_min_age = 0;
2027 0 : default_multixact_freeze_table_age = 0;
2028 : }
2029 : else
2030 : {
2031 56 : default_freeze_min_age = vacuum_freeze_min_age;
2032 56 : default_freeze_table_age = vacuum_freeze_table_age;
2033 56 : default_multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
2034 56 : default_multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
2035 : }
2036 :
2037 56 : ReleaseSysCache(tuple);
2038 :
2039 : /* StartTransactionCommand changed elsewhere */
2040 56 : MemoryContextSwitchTo(AutovacMemCxt);
2041 :
2042 : /* The database hash where pgstat keeps shared relations */
2043 56 : shared = pgstat_fetch_stat_dbentry(InvalidOid);
2044 :
2045 56 : classRel = table_open(RelationRelationId, AccessShareLock);
2046 :
2047 : /* create a copy so we can use it after closing pg_class */
2048 56 : pg_class_desc = CreateTupleDescCopy(RelationGetDescr(classRel));
2049 :
2050 : /* create hash table for toast <-> main relid mapping */
2051 56 : ctl.keysize = sizeof(Oid);
2052 56 : ctl.entrysize = sizeof(av_relation);
2053 :
2054 56 : table_toast_map = hash_create("TOAST to main relid map",
2055 : 100,
2056 : &ctl,
2057 : HASH_ELEM | HASH_BLOBS);
2058 :
2059 : /*
2060 : * Scan pg_class to determine which tables to vacuum.
2061 : *
2062 : * We do this in two passes: on the first one we collect the list of plain
2063 : * relations and materialized views, and on the second one we collect
2064 : * TOAST tables. The reason for doing the second pass is that during it we
2065 : * want to use the main relation's pg_class.reloptions entry if the TOAST
2066 : * table does not have any, and we cannot obtain it unless we know
2067 : * beforehand what's the main table OID.
2068 : *
2069 : * We need to check TOAST tables separately because in cases with short,
2070 : * wide tables there might be proportionally much more activity in the
2071 : * TOAST table than in its parent.
2072 : */
2073 56 : relScan = table_beginscan_catalog(classRel, 0, NULL);
2074 :
2075 : /*
2076 : * On the first pass, we collect main tables to vacuum, and also the main
2077 : * table relid to TOAST relid mapping.
2078 : */
2079 24818 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2080 : {
2081 24762 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2082 : PgStat_StatTabEntry *tabentry;
2083 : AutoVacOpts *relopts;
2084 : Oid relid;
2085 : bool dovacuum;
2086 : bool doanalyze;
2087 : bool wraparound;
2088 :
2089 24762 : if (classForm->relkind != RELKIND_RELATION &&
2090 20096 : classForm->relkind != RELKIND_MATVIEW)
2091 20052 : continue;
2092 :
2093 4710 : relid = classForm->oid;
2094 :
2095 : /*
2096 : * Check if it is a temp table (presumably, of some other backend's).
2097 : * We cannot safely process other backends' temp tables.
2098 : */
2099 4710 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2100 : {
2101 : /*
2102 : * We just ignore it if the owning backend is still active and
2103 : * using the temporary schema. Also, for safety, ignore it if the
2104 : * namespace doesn't exist or isn't a temp namespace after all.
2105 : */
2106 0 : if (checkTempNamespaceStatus(classForm->relnamespace) == TEMP_NAMESPACE_IDLE)
2107 : {
2108 : /*
2109 : * The table seems to be orphaned -- although it might be that
2110 : * the owning backend has already deleted it and exited; our
2111 : * pg_class scan snapshot is not necessarily up-to-date
2112 : * anymore, so we could be looking at a committed-dead entry.
2113 : * Remember it so we can try to delete it later.
2114 : */
2115 0 : orphan_oids = lappend_oid(orphan_oids, relid);
2116 : }
2117 0 : continue;
2118 : }
2119 :
2120 : /* Fetch reloptions and the pgstat entry for this table */
2121 4710 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2122 4710 : tabentry = get_pgstat_tabentry_relid(relid, classForm->relisshared,
2123 : shared, dbentry);
2124 :
2125 : /* Check if it needs vacuum or analyze */
2126 4710 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2127 : effective_multixact_freeze_max_age,
2128 : &dovacuum, &doanalyze, &wraparound);
2129 :
2130 : /* Relations that need work are added to table_oids */
2131 4710 : if (dovacuum || doanalyze)
2132 252 : table_oids = lappend_oid(table_oids, relid);
2133 :
2134 : /*
2135 : * Remember TOAST associations for the second pass. Note: we must do
2136 : * this whether or not the table is going to be vacuumed, because we
2137 : * don't automatically vacuum toast tables along the parent table.
2138 : */
2139 4710 : if (OidIsValid(classForm->reltoastrelid))
2140 : {
2141 : av_relation *hentry;
2142 : bool found;
2143 :
2144 4996 : hentry = hash_search(table_toast_map,
2145 2498 : &classForm->reltoastrelid,
2146 : HASH_ENTER, &found);
2147 :
2148 2498 : if (!found)
2149 : {
2150 : /* hash_search already filled in the key */
2151 2498 : hentry->ar_relid = relid;
2152 2498 : hentry->ar_hasrelopts = false;
2153 2498 : if (relopts != NULL)
2154 : {
2155 14 : hentry->ar_hasrelopts = true;
2156 14 : memcpy(&hentry->ar_reloptions, relopts,
2157 : sizeof(AutoVacOpts));
2158 : }
2159 : }
2160 : }
2161 : }
2162 :
2163 56 : table_endscan(relScan);
2164 :
2165 : /* second pass: check TOAST tables */
2166 56 : ScanKeyInit(&key,
2167 : Anum_pg_class_relkind,
2168 : BTEqualStrategyNumber, F_CHAREQ,
2169 : CharGetDatum(RELKIND_TOASTVALUE));
2170 :
2171 56 : relScan = table_beginscan_catalog(classRel, 1, &key);
2172 2554 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2173 : {
2174 2498 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2175 : PgStat_StatTabEntry *tabentry;
2176 : Oid relid;
2177 2498 : AutoVacOpts *relopts = NULL;
2178 : bool dovacuum;
2179 : bool doanalyze;
2180 : bool wraparound;
2181 :
2182 : /*
2183 : * We cannot safely process other backends' temp tables, so skip 'em.
2184 : */
2185 2498 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2186 0 : continue;
2187 :
2188 2498 : relid = classForm->oid;
2189 :
2190 : /*
2191 : * fetch reloptions -- if this toast table does not have them, try the
2192 : * main rel
2193 : */
2194 2498 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2195 2498 : if (relopts == NULL)
2196 : {
2197 : av_relation *hentry;
2198 : bool found;
2199 :
2200 2498 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2201 2498 : if (found && hentry->ar_hasrelopts)
2202 14 : relopts = &hentry->ar_reloptions;
2203 : }
2204 :
2205 : /* Fetch the pgstat entry for this table */
2206 2498 : tabentry = get_pgstat_tabentry_relid(relid, classForm->relisshared,
2207 : shared, dbentry);
2208 :
2209 2498 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2210 : effective_multixact_freeze_max_age,
2211 : &dovacuum, &doanalyze, &wraparound);
2212 :
2213 : /* ignore analyze for toast tables */
2214 2498 : if (dovacuum)
2215 4 : table_oids = lappend_oid(table_oids, relid);
2216 : }
2217 :
2218 56 : table_endscan(relScan);
2219 56 : table_close(classRel, AccessShareLock);
2220 :
2221 : /*
2222 : * Recheck orphan temporary tables, and if they still seem orphaned, drop
2223 : * them. We'll eat a transaction per dropped table, which might seem
2224 : * excessive, but we should only need to do anything as a result of a
2225 : * previous backend crash, so this should not happen often enough to
2226 : * justify "optimizing". Using separate transactions ensures that we
2227 : * don't bloat the lock table if there are many temp tables to be dropped,
2228 : * and it ensures that we don't lose work if a deletion attempt fails.
2229 : */
2230 56 : foreach(cell, orphan_oids)
2231 : {
2232 0 : Oid relid = lfirst_oid(cell);
2233 : Form_pg_class classForm;
2234 : ObjectAddress object;
2235 :
2236 : /*
2237 : * Check for user-requested abort.
2238 : */
2239 0 : CHECK_FOR_INTERRUPTS();
2240 :
2241 : /*
2242 : * Try to lock the table. If we can't get the lock immediately,
2243 : * somebody else is using (or dropping) the table, so it's not our
2244 : * concern anymore. Having the lock prevents race conditions below.
2245 : */
2246 0 : if (!ConditionalLockRelationOid(relid, AccessExclusiveLock))
2247 0 : continue;
2248 :
2249 : /*
2250 : * Re-fetch the pg_class tuple and re-check whether it still seems to
2251 : * be an orphaned temp table. If it's not there or no longer the same
2252 : * relation, ignore it.
2253 : */
2254 0 : tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2255 0 : if (!HeapTupleIsValid(tuple))
2256 : {
2257 : /* be sure to drop useless lock so we don't bloat lock table */
2258 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2259 0 : continue;
2260 : }
2261 0 : classForm = (Form_pg_class) GETSTRUCT(tuple);
2262 :
2263 : /*
2264 : * Make all the same tests made in the loop above. In event of OID
2265 : * counter wraparound, the pg_class entry we have now might be
2266 : * completely unrelated to the one we saw before.
2267 : */
2268 0 : if (!((classForm->relkind == RELKIND_RELATION ||
2269 0 : classForm->relkind == RELKIND_MATVIEW) &&
2270 0 : classForm->relpersistence == RELPERSISTENCE_TEMP))
2271 : {
2272 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2273 0 : continue;
2274 : }
2275 :
2276 0 : if (checkTempNamespaceStatus(classForm->relnamespace) != TEMP_NAMESPACE_IDLE)
2277 : {
2278 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2279 0 : continue;
2280 : }
2281 :
2282 : /* OK, let's delete it */
2283 0 : ereport(LOG,
2284 : (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"",
2285 : get_database_name(MyDatabaseId),
2286 : get_namespace_name(classForm->relnamespace),
2287 : NameStr(classForm->relname))));
2288 :
2289 0 : object.classId = RelationRelationId;
2290 0 : object.objectId = relid;
2291 0 : object.objectSubId = 0;
2292 0 : performDeletion(&object, DROP_CASCADE,
2293 : PERFORM_DELETION_INTERNAL |
2294 : PERFORM_DELETION_QUIETLY |
2295 : PERFORM_DELETION_SKIP_EXTENSIONS);
2296 :
2297 : /*
2298 : * To commit the deletion, end current transaction and start a new
2299 : * one. Note this also releases the lock we took.
2300 : */
2301 0 : CommitTransactionCommand();
2302 0 : StartTransactionCommand();
2303 :
2304 : /* StartTransactionCommand changed current memory context */
2305 0 : MemoryContextSwitchTo(AutovacMemCxt);
2306 : }
2307 :
2308 : /*
2309 : * Create a buffer access strategy object for VACUUM to use. We want to
2310 : * use the same one across all the vacuum operations we perform, since the
2311 : * point is for VACUUM not to blow out the shared cache.
2312 : */
2313 56 : bstrategy = GetAccessStrategy(BAS_VACUUM);
2314 :
2315 : /*
2316 : * create a memory context to act as fake PortalContext, so that the
2317 : * contexts created in the vacuum code are cleaned up for each table.
2318 : */
2319 56 : PortalContext = AllocSetContextCreate(AutovacMemCxt,
2320 : "Autovacuum Portal",
2321 : ALLOCSET_DEFAULT_SIZES);
2322 :
2323 : /*
2324 : * Perform operations on collected tables.
2325 : */
2326 312 : foreach(cell, table_oids)
2327 : {
2328 256 : Oid relid = lfirst_oid(cell);
2329 : HeapTuple classTup;
2330 : autovac_table *tab;
2331 : bool isshared;
2332 : bool skipit;
2333 : double stdVacuumCostDelay;
2334 : int stdVacuumCostLimit;
2335 : dlist_iter iter;
2336 :
2337 256 : CHECK_FOR_INTERRUPTS();
2338 :
2339 : /*
2340 : * Check for config changes before processing each collected table.
2341 : */
2342 256 : if (ConfigReloadPending)
2343 : {
2344 0 : ConfigReloadPending = false;
2345 0 : ProcessConfigFile(PGC_SIGHUP);
2346 :
2347 : /*
2348 : * You might be tempted to bail out if we see autovacuum is now
2349 : * disabled. Must resist that temptation -- this might be a
2350 : * for-wraparound emergency worker, in which case that would be
2351 : * entirely inappropriate.
2352 : */
2353 : }
2354 :
2355 : /*
2356 : * Find out whether the table is shared or not. (It's slightly
2357 : * annoying to fetch the syscache entry just for this, but in typical
2358 : * cases it adds little cost because table_recheck_autovac would
2359 : * refetch the entry anyway. We could buy that back by copying the
2360 : * tuple here and passing it to table_recheck_autovac, but that
2361 : * increases the odds of that function working with stale data.)
2362 : */
2363 256 : classTup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
2364 256 : if (!HeapTupleIsValid(classTup))
2365 0 : continue; /* somebody deleted the rel, forget it */
2366 256 : isshared = ((Form_pg_class) GETSTRUCT(classTup))->relisshared;
2367 256 : ReleaseSysCache(classTup);
2368 :
2369 : /*
2370 : * Hold schedule lock from here until we've claimed the table. We
2371 : * also need the AutovacuumLock to walk the worker array, but that one
2372 : * can just be a shared lock.
2373 : */
2374 256 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2375 256 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2376 :
2377 : /*
2378 : * Check whether the table is being vacuumed concurrently by another
2379 : * worker.
2380 : */
2381 256 : skipit = false;
2382 512 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
2383 : {
2384 256 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
2385 :
2386 : /* ignore myself */
2387 256 : if (worker == MyWorkerInfo)
2388 256 : continue;
2389 :
2390 : /* ignore workers in other databases (unless table is shared) */
2391 0 : if (!worker->wi_sharedrel && worker->wi_dboid != MyDatabaseId)
2392 0 : continue;
2393 :
2394 0 : if (worker->wi_tableoid == relid)
2395 : {
2396 0 : skipit = true;
2397 0 : found_concurrent_worker = true;
2398 0 : break;
2399 : }
2400 : }
2401 256 : LWLockRelease(AutovacuumLock);
2402 256 : if (skipit)
2403 : {
2404 0 : LWLockRelease(AutovacuumScheduleLock);
2405 0 : continue;
2406 : }
2407 :
2408 : /*
2409 : * Store the table's OID in shared memory before releasing the
2410 : * schedule lock, so that other workers don't try to vacuum it
2411 : * concurrently. (We claim it here so as not to hold
2412 : * AutovacuumScheduleLock while rechecking the stats.)
2413 : */
2414 256 : MyWorkerInfo->wi_tableoid = relid;
2415 256 : MyWorkerInfo->wi_sharedrel = isshared;
2416 256 : LWLockRelease(AutovacuumScheduleLock);
2417 :
2418 : /*
2419 : * Check whether pgstat data still says we need to vacuum this table.
2420 : * It could have changed if something else processed the table while
2421 : * we weren't looking.
2422 : *
2423 : * Note: we have a special case in pgstat code to ensure that the
2424 : * stats we read are as up-to-date as possible, to avoid the problem
2425 : * that somebody just finished vacuuming this table. The window to
2426 : * the race condition is not closed but it is very small.
2427 : */
2428 256 : MemoryContextSwitchTo(AutovacMemCxt);
2429 256 : tab = table_recheck_autovac(relid, table_toast_map, pg_class_desc,
2430 : effective_multixact_freeze_max_age);
2431 256 : if (tab == NULL)
2432 : {
2433 : /* someone else vacuumed the table, or it went away */
2434 0 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2435 0 : MyWorkerInfo->wi_tableoid = InvalidOid;
2436 0 : MyWorkerInfo->wi_sharedrel = false;
2437 0 : LWLockRelease(AutovacuumScheduleLock);
2438 0 : continue;
2439 : }
2440 :
2441 : /*
2442 : * Remember the prevailing values of the vacuum cost GUCs. We have to
2443 : * restore these at the bottom of the loop, else we'll compute wrong
2444 : * values in the next iteration of autovac_balance_cost().
2445 : */
2446 256 : stdVacuumCostDelay = VacuumCostDelay;
2447 256 : stdVacuumCostLimit = VacuumCostLimit;
2448 :
2449 : /* Must hold AutovacuumLock while mucking with cost balance info */
2450 256 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2451 :
2452 : /* advertise my cost delay parameters for the balancing algorithm */
2453 256 : MyWorkerInfo->wi_dobalance = tab->at_dobalance;
2454 256 : MyWorkerInfo->wi_cost_delay = tab->at_vacuum_cost_delay;
2455 256 : MyWorkerInfo->wi_cost_limit = tab->at_vacuum_cost_limit;
2456 256 : MyWorkerInfo->wi_cost_limit_base = tab->at_vacuum_cost_limit;
2457 :
2458 : /* do a balance */
2459 256 : autovac_balance_cost();
2460 :
2461 : /* set the active cost parameters from the result of that */
2462 256 : AutoVacuumUpdateDelay();
2463 :
2464 : /* done */
2465 256 : LWLockRelease(AutovacuumLock);
2466 :
2467 : /* clean up memory before each iteration */
2468 256 : MemoryContextResetAndDeleteChildren(PortalContext);
2469 :
2470 : /*
2471 : * Save the relation name for a possible error message, to avoid a
2472 : * catalog lookup in case of an error. If any of these return NULL,
2473 : * then the relation has been dropped since last we checked; skip it.
2474 : * Note: they must live in a long-lived memory context because we call
2475 : * vacuum and analyze in different transactions.
2476 : */
2477 :
2478 256 : tab->at_relname = get_rel_name(tab->at_relid);
2479 256 : tab->at_nspname = get_namespace_name(get_rel_namespace(tab->at_relid));
2480 256 : tab->at_datname = get_database_name(MyDatabaseId);
2481 256 : if (!tab->at_relname || !tab->at_nspname || !tab->at_datname)
2482 0 : goto deleted;
2483 :
2484 : /*
2485 : * We will abort vacuuming the current table if something errors out,
2486 : * and continue with the next one in schedule; in particular, this
2487 : * happens if we are interrupted with SIGINT.
2488 : */
2489 256 : PG_TRY();
2490 : {
2491 : /* Use PortalContext for any per-table allocations */
2492 256 : MemoryContextSwitchTo(PortalContext);
2493 :
2494 : /* have at it */
2495 256 : autovacuum_do_vac_analyze(tab, bstrategy);
2496 :
2497 : /*
2498 : * Clear a possible query-cancel signal, to avoid a late reaction
2499 : * to an automatically-sent signal because of vacuuming the
2500 : * current table (we're done with it, so it would make no sense to
2501 : * cancel at this point.)
2502 : */
2503 256 : QueryCancelPending = false;
2504 : }
2505 0 : PG_CATCH();
2506 : {
2507 : /*
2508 : * Abort the transaction, start a new one, and proceed with the
2509 : * next table in our list.
2510 : */
2511 0 : HOLD_INTERRUPTS();
2512 0 : if (tab->at_params.options & VACOPT_VACUUM)
2513 0 : errcontext("automatic vacuum of table \"%s.%s.%s\"",
2514 : tab->at_datname, tab->at_nspname, tab->at_relname);
2515 : else
2516 0 : errcontext("automatic analyze of table \"%s.%s.%s\"",
2517 : tab->at_datname, tab->at_nspname, tab->at_relname);
2518 0 : EmitErrorReport();
2519 :
2520 : /* this resets ProcGlobal->statusFlags[i] too */
2521 0 : AbortOutOfAnyTransaction();
2522 0 : FlushErrorState();
2523 0 : MemoryContextResetAndDeleteChildren(PortalContext);
2524 :
2525 : /* restart our transaction for the following operations */
2526 0 : StartTransactionCommand();
2527 0 : RESUME_INTERRUPTS();
2528 : }
2529 256 : PG_END_TRY();
2530 :
2531 : /* Make sure we're back in AutovacMemCxt */
2532 256 : MemoryContextSwitchTo(AutovacMemCxt);
2533 :
2534 256 : did_vacuum = true;
2535 :
2536 : /* ProcGlobal->statusFlags[i] are reset at the next end of xact */
2537 :
2538 : /* be tidy */
2539 256 : deleted:
2540 256 : if (tab->at_datname != NULL)
2541 256 : pfree(tab->at_datname);
2542 256 : if (tab->at_nspname != NULL)
2543 256 : pfree(tab->at_nspname);
2544 256 : if (tab->at_relname != NULL)
2545 256 : pfree(tab->at_relname);
2546 256 : pfree(tab);
2547 :
2548 : /*
2549 : * Remove my info from shared memory. We could, but intentionally
2550 : * don't, clear wi_cost_limit and friends --- this is on the
2551 : * assumption that we probably have more to do with similar cost
2552 : * settings, so we don't want to give up our share of I/O for a very
2553 : * short interval and thereby thrash the global balance.
2554 : */
2555 256 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2556 256 : MyWorkerInfo->wi_tableoid = InvalidOid;
2557 256 : MyWorkerInfo->wi_sharedrel = false;
2558 256 : LWLockRelease(AutovacuumScheduleLock);
2559 :
2560 : /* restore vacuum cost GUCs for the next iteration */
2561 256 : VacuumCostDelay = stdVacuumCostDelay;
2562 256 : VacuumCostLimit = stdVacuumCostLimit;
2563 : }
2564 :
2565 : /*
2566 : * Perform additional work items, as requested by backends.
2567 : */
2568 56 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2569 14392 : for (i = 0; i < NUM_WORKITEMS; i++)
2570 : {
2571 14336 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
2572 :
2573 14336 : if (!workitem->avw_used)
2574 14330 : continue;
2575 6 : if (workitem->avw_active)
2576 0 : continue;
2577 6 : if (workitem->avw_database != MyDatabaseId)
2578 0 : continue;
2579 :
2580 : /* claim this one, and release lock while performing it */
2581 6 : workitem->avw_active = true;
2582 6 : LWLockRelease(AutovacuumLock);
2583 :
2584 6 : perform_work_item(workitem);
2585 :
2586 : /*
2587 : * Check for config changes before acquiring lock for further jobs.
2588 : */
2589 6 : CHECK_FOR_INTERRUPTS();
2590 6 : if (ConfigReloadPending)
2591 : {
2592 0 : ConfigReloadPending = false;
2593 0 : ProcessConfigFile(PGC_SIGHUP);
2594 : }
2595 :
2596 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2597 :
2598 : /* and mark it done */
2599 6 : workitem->avw_active = false;
2600 6 : workitem->avw_used = false;
2601 : }
2602 56 : LWLockRelease(AutovacuumLock);
2603 :
2604 : /*
2605 : * We leak table_toast_map here (among other things), but since we're
2606 : * going away soon, it's not a problem.
2607 : */
2608 :
2609 : /*
2610 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We
2611 : * only need to do this once, not after each table.
2612 : *
2613 : * Even if we didn't vacuum anything, it may still be important to do
2614 : * this, because one indirect effect of vac_update_datfrozenxid() is to
2615 : * update ShmemVariableCache->xidVacLimit. That might need to be done
2616 : * even if we haven't vacuumed anything, because relations with older
2617 : * relfrozenxid values or other databases with older datfrozenxid values
2618 : * might have been dropped, allowing xidVacLimit to advance.
2619 : *
2620 : * However, it's also important not to do this blindly in all cases,
2621 : * because when autovacuum=off this will restart the autovacuum launcher.
2622 : * If we're not careful, an infinite loop can result, where workers find
2623 : * no work to do and restart the launcher, which starts another worker in
2624 : * the same database that finds no work to do. To prevent that, we skip
2625 : * this if (1) we found no work to do and (2) we skipped at least one
2626 : * table due to concurrent autovacuum activity. In that case, the other
2627 : * worker has already done it, or will do so when it finishes.
2628 : */
2629 56 : if (did_vacuum || !found_concurrent_worker)
2630 56 : vac_update_datfrozenxid();
2631 :
2632 : /* Finally close out the last transaction. */
2633 56 : CommitTransactionCommand();
2634 56 : }
2635 :
2636 : /*
2637 : * Execute a previously registered work item.
2638 : */
2639 : static void
2640 6 : perform_work_item(AutoVacuumWorkItem *workitem)
2641 : {
2642 6 : char *cur_datname = NULL;
2643 6 : char *cur_nspname = NULL;
2644 6 : char *cur_relname = NULL;
2645 :
2646 : /*
2647 : * Note we do not store table info in MyWorkerInfo, since this is not
2648 : * vacuuming proper.
2649 : */
2650 :
2651 : /*
2652 : * Save the relation name for a possible error message, to avoid a catalog
2653 : * lookup in case of an error. If any of these return NULL, then the
2654 : * relation has been dropped since last we checked; skip it.
2655 : */
2656 : Assert(CurrentMemoryContext == AutovacMemCxt);
2657 :
2658 6 : cur_relname = get_rel_name(workitem->avw_relation);
2659 6 : cur_nspname = get_namespace_name(get_rel_namespace(workitem->avw_relation));
2660 6 : cur_datname = get_database_name(MyDatabaseId);
2661 6 : if (!cur_relname || !cur_nspname || !cur_datname)
2662 0 : goto deleted2;
2663 :
2664 6 : autovac_report_workitem(workitem, cur_nspname, cur_relname);
2665 :
2666 : /* clean up memory before each work item */
2667 6 : MemoryContextResetAndDeleteChildren(PortalContext);
2668 :
2669 : /*
2670 : * We will abort the current work item if something errors out, and
2671 : * continue with the next one; in particular, this happens if we are
2672 : * interrupted with SIGINT. Note that this means that the work item list
2673 : * can be lossy.
2674 : */
2675 6 : PG_TRY();
2676 : {
2677 : /* Use PortalContext for any per-work-item allocations */
2678 6 : MemoryContextSwitchTo(PortalContext);
2679 :
2680 : /* have at it */
2681 6 : switch (workitem->avw_type)
2682 : {
2683 6 : case AVW_BRINSummarizeRange:
2684 6 : DirectFunctionCall2(brin_summarize_range,
2685 : ObjectIdGetDatum(workitem->avw_relation),
2686 : Int64GetDatum((int64) workitem->avw_blockNumber));
2687 6 : break;
2688 0 : default:
2689 0 : elog(WARNING, "unrecognized work item found: type %d",
2690 : workitem->avw_type);
2691 0 : break;
2692 : }
2693 :
2694 : /*
2695 : * Clear a possible query-cancel signal, to avoid a late reaction to
2696 : * an automatically-sent signal because of vacuuming the current table
2697 : * (we're done with it, so it would make no sense to cancel at this
2698 : * point.)
2699 : */
2700 6 : QueryCancelPending = false;
2701 : }
2702 0 : PG_CATCH();
2703 : {
2704 : /*
2705 : * Abort the transaction, start a new one, and proceed with the next
2706 : * table in our list.
2707 : */
2708 0 : HOLD_INTERRUPTS();
2709 0 : errcontext("processing work entry for relation \"%s.%s.%s\"",
2710 : cur_datname, cur_nspname, cur_relname);
2711 0 : EmitErrorReport();
2712 :
2713 : /* this resets ProcGlobal->statusFlags[i] too */
2714 0 : AbortOutOfAnyTransaction();
2715 0 : FlushErrorState();
2716 0 : MemoryContextResetAndDeleteChildren(PortalContext);
2717 :
2718 : /* restart our transaction for the following operations */
2719 0 : StartTransactionCommand();
2720 0 : RESUME_INTERRUPTS();
2721 : }
2722 6 : PG_END_TRY();
2723 :
2724 : /* Make sure we're back in AutovacMemCxt */
2725 6 : MemoryContextSwitchTo(AutovacMemCxt);
2726 :
2727 : /* We intentionally do not set did_vacuum here */
2728 :
2729 : /* be tidy */
2730 6 : deleted2:
2731 6 : if (cur_datname)
2732 6 : pfree(cur_datname);
2733 6 : if (cur_nspname)
2734 6 : pfree(cur_nspname);
2735 6 : if (cur_relname)
2736 6 : pfree(cur_relname);
2737 6 : }
2738 :
2739 : /*
2740 : * extract_autovac_opts
2741 : *
2742 : * Given a relation's pg_class tuple, return the AutoVacOpts portion of
2743 : * reloptions, if set; otherwise, return NULL.
2744 : *
2745 : * Note: callers do not have a relation lock on the table at this point,
2746 : * so the table could have been dropped, and its catalog rows gone, after
2747 : * we acquired the pg_class row. If pg_class had a TOAST table, this would
2748 : * be a risk; fortunately, it doesn't.
2749 : */
2750 : static AutoVacOpts *
2751 7464 : extract_autovac_opts(HeapTuple tup, TupleDesc pg_class_desc)
2752 : {
2753 : bytea *relopts;
2754 : AutoVacOpts *av;
2755 :
2756 : Assert(((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_RELATION ||
2757 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_MATVIEW ||
2758 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_TOASTVALUE);
2759 :
2760 7464 : relopts = extractRelOptions(tup, pg_class_desc, NULL);
2761 7464 : if (relopts == NULL)
2762 7430 : return NULL;
2763 :
2764 34 : av = palloc(sizeof(AutoVacOpts));
2765 34 : memcpy(av, &(((StdRdOptions *) relopts)->autovacuum), sizeof(AutoVacOpts));
2766 34 : pfree(relopts);
2767 :
2768 34 : return av;
2769 : }
2770 :
2771 : /*
2772 : * get_pgstat_tabentry_relid
2773 : *
2774 : * Fetch the pgstat entry of a table, either local to a database or shared.
2775 : */
2776 : static PgStat_StatTabEntry *
2777 7464 : get_pgstat_tabentry_relid(Oid relid, bool isshared, PgStat_StatDBEntry *shared,
2778 : PgStat_StatDBEntry *dbentry)
2779 : {
2780 7464 : PgStat_StatTabEntry *tabentry = NULL;
2781 :
2782 7464 : if (isshared)
2783 : {
2784 1024 : if (PointerIsValid(shared))
2785 1024 : tabentry = hash_search(shared->tables, &relid,
2786 : HASH_FIND, NULL);
2787 : }
2788 6440 : else if (PointerIsValid(dbentry))
2789 6440 : tabentry = hash_search(dbentry->tables, &relid,
2790 : HASH_FIND, NULL);
2791 :
2792 7464 : return tabentry;
2793 : }
2794 :
2795 : /*
2796 : * table_recheck_autovac
2797 : *
2798 : * Recheck whether a table still needs vacuum or analyze. Return value is a
2799 : * valid autovac_table pointer if it does, NULL otherwise.
2800 : *
2801 : * Note that the returned autovac_table does not have the name fields set.
2802 : */
2803 : static autovac_table *
2804 256 : table_recheck_autovac(Oid relid, HTAB *table_toast_map,
2805 : TupleDesc pg_class_desc,
2806 : int effective_multixact_freeze_max_age)
2807 : {
2808 : Form_pg_class classForm;
2809 : HeapTuple classTup;
2810 : bool dovacuum;
2811 : bool doanalyze;
2812 256 : autovac_table *tab = NULL;
2813 : bool wraparound;
2814 : AutoVacOpts *avopts;
2815 : static bool reuse_stats = false;
2816 :
2817 : /* fetch the relation's relcache entry */
2818 256 : classTup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2819 256 : if (!HeapTupleIsValid(classTup))
2820 0 : return NULL;
2821 256 : classForm = (Form_pg_class) GETSTRUCT(classTup);
2822 :
2823 : /*
2824 : * Get the applicable reloptions. If it is a TOAST table, try to get the
2825 : * main table reloptions if the toast table itself doesn't have.
2826 : */
2827 256 : avopts = extract_autovac_opts(classTup, pg_class_desc);
2828 256 : if (classForm->relkind == RELKIND_TOASTVALUE &&
2829 4 : avopts == NULL && table_toast_map != NULL)
2830 : {
2831 : av_relation *hentry;
2832 : bool found;
2833 :
2834 4 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2835 4 : if (found && hentry->ar_hasrelopts)
2836 0 : avopts = &hentry->ar_reloptions;
2837 : }
2838 :
2839 : /*
2840 : * Reuse the stats to recheck whether a relation needs to be vacuumed or
2841 : * analyzed if it was reloaded before and has not been cleared yet. This
2842 : * is necessary to avoid frequent refresh of stats, especially when there
2843 : * are very large number of relations and the refresh can cause lots of
2844 : * overhead.
2845 : *
2846 : * If we determined that a relation needs to be vacuumed or analyzed,
2847 : * based on the old stats, we refresh stats and recheck the necessity
2848 : * again. Because a relation may have already been vacuumed or analyzed by
2849 : * someone since the last reload of stats.
2850 : */
2851 256 : if (reuse_stats)
2852 : {
2853 0 : recheck_relation_needs_vacanalyze(relid, avopts, classForm,
2854 : effective_multixact_freeze_max_age,
2855 : &dovacuum, &doanalyze, &wraparound);
2856 :
2857 : /* Quick exit if a relation doesn't need to be vacuumed or analyzed */
2858 0 : if (!doanalyze && !dovacuum)
2859 : {
2860 0 : heap_freetuple(classTup);
2861 0 : return NULL;
2862 : }
2863 : }
2864 :
2865 : /* Use fresh stats and recheck again */
2866 256 : autovac_refresh_stats();
2867 :
2868 256 : recheck_relation_needs_vacanalyze(relid, avopts, classForm,
2869 : effective_multixact_freeze_max_age,
2870 : &dovacuum, &doanalyze, &wraparound);
2871 :
2872 : /* OK, it needs something done */
2873 256 : if (doanalyze || dovacuum)
2874 256 : {
2875 : int freeze_min_age;
2876 : int freeze_table_age;
2877 : int multixact_freeze_min_age;
2878 : int multixact_freeze_table_age;
2879 : int vac_cost_limit;
2880 : double vac_cost_delay;
2881 : int log_min_duration;
2882 :
2883 : /*
2884 : * Calculate the vacuum cost parameters and the freeze ages. If there
2885 : * are options set in pg_class.reloptions, use them; in the case of a
2886 : * toast table, try the main table too. Otherwise use the GUC
2887 : * defaults, autovacuum's own first and plain vacuum second.
2888 : */
2889 :
2890 : /* -1 in autovac setting means use plain vacuum_cost_delay */
2891 2 : vac_cost_delay = (avopts && avopts->vacuum_cost_delay >= 0)
2892 : ? avopts->vacuum_cost_delay
2893 512 : : (autovacuum_vac_cost_delay >= 0)
2894 : ? autovacuum_vac_cost_delay
2895 256 : : VacuumCostDelay;
2896 :
2897 : /* 0 or -1 in autovac setting means use plain vacuum_cost_limit */
2898 2 : vac_cost_limit = (avopts && avopts->vacuum_cost_limit > 0)
2899 : ? avopts->vacuum_cost_limit
2900 512 : : (autovacuum_vac_cost_limit > 0)
2901 : ? autovacuum_vac_cost_limit
2902 256 : : VacuumCostLimit;
2903 :
2904 : /* -1 in autovac setting means use log_autovacuum_min_duration */
2905 2 : log_min_duration = (avopts && avopts->log_min_duration >= 0)
2906 : ? avopts->log_min_duration
2907 258 : : Log_autovacuum_min_duration;
2908 :
2909 : /* these do not have autovacuum-specific settings */
2910 2 : freeze_min_age = (avopts && avopts->freeze_min_age >= 0)
2911 : ? avopts->freeze_min_age
2912 258 : : default_freeze_min_age;
2913 :
2914 2 : freeze_table_age = (avopts && avopts->freeze_table_age >= 0)
2915 : ? avopts->freeze_table_age
2916 258 : : default_freeze_table_age;
2917 :
2918 258 : multixact_freeze_min_age = (avopts &&
2919 2 : avopts->multixact_freeze_min_age >= 0)
2920 : ? avopts->multixact_freeze_min_age
2921 258 : : default_multixact_freeze_min_age;
2922 :
2923 258 : multixact_freeze_table_age = (avopts &&
2924 2 : avopts->multixact_freeze_table_age >= 0)
2925 : ? avopts->multixact_freeze_table_age
2926 258 : : default_multixact_freeze_table_age;
2927 :
2928 256 : tab = palloc(sizeof(autovac_table));
2929 256 : tab->at_relid = relid;
2930 256 : tab->at_sharedrel = classForm->relisshared;
2931 :
2932 : /* Note that this skips toast relations */
2933 768 : tab->at_params.options = (dovacuum ? VACOPT_VACUUM : 0) |
2934 256 : (doanalyze ? VACOPT_ANALYZE : 0) |
2935 256 : (!wraparound ? VACOPT_SKIP_LOCKED : 0);
2936 :
2937 : /*
2938 : * index_cleanup and truncate are unspecified at first in autovacuum.
2939 : * They will be filled in with usable values using their reloptions
2940 : * (or reloption defaults) later.
2941 : */
2942 256 : tab->at_params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
2943 256 : tab->at_params.truncate = VACOPTVALUE_UNSPECIFIED;
2944 : /* As of now, we don't support parallel vacuum for autovacuum */
2945 256 : tab->at_params.nworkers = -1;
2946 256 : tab->at_params.freeze_min_age = freeze_min_age;
2947 256 : tab->at_params.freeze_table_age = freeze_table_age;
2948 256 : tab->at_params.multixact_freeze_min_age = multixact_freeze_min_age;
2949 256 : tab->at_params.multixact_freeze_table_age = multixact_freeze_table_age;
2950 256 : tab->at_params.is_wraparound = wraparound;
2951 256 : tab->at_params.log_min_duration = log_min_duration;
2952 256 : tab->at_vacuum_cost_limit = vac_cost_limit;
2953 256 : tab->at_vacuum_cost_delay = vac_cost_delay;
2954 256 : tab->at_relname = NULL;
2955 256 : tab->at_nspname = NULL;
2956 256 : tab->at_datname = NULL;
2957 :
2958 : /*
2959 : * If any of the cost delay parameters has been set individually for
2960 : * this table, disable the balancing algorithm.
2961 : */
2962 256 : tab->at_dobalance =
2963 258 : !(avopts && (avopts->vacuum_cost_limit > 0 ||
2964 2 : avopts->vacuum_cost_delay > 0));
2965 :
2966 : /*
2967 : * When we decide to do vacuum or analyze, the existing stats cannot
2968 : * be reused in the next cycle because it's cleared at the end of
2969 : * vacuum or analyze (by AtEOXact_PgStat()).
2970 : */
2971 256 : reuse_stats = false;
2972 : }
2973 : else
2974 : {
2975 : /*
2976 : * If neither vacuum nor analyze is necessary, the existing stats is
2977 : * not cleared and can be reused in the next cycle.
2978 : */
2979 0 : reuse_stats = true;
2980 : }
2981 :
2982 256 : heap_freetuple(classTup);
2983 256 : return tab;
2984 : }
2985 :
2986 : /*
2987 : * recheck_relation_needs_vacanalyze
2988 : *
2989 : * Subroutine for table_recheck_autovac.
2990 : *
2991 : * Fetch the pgstat of a relation and recheck whether a relation
2992 : * needs to be vacuumed or analyzed.
2993 : */
2994 : static void
2995 256 : recheck_relation_needs_vacanalyze(Oid relid,
2996 : AutoVacOpts *avopts,
2997 : Form_pg_class classForm,
2998 : int effective_multixact_freeze_max_age,
2999 : bool *dovacuum,
3000 : bool *doanalyze,
3001 : bool *wraparound)
3002 : {
3003 : PgStat_StatTabEntry *tabentry;
3004 256 : PgStat_StatDBEntry *shared = NULL;
3005 256 : PgStat_StatDBEntry *dbentry = NULL;
3006 :
3007 256 : if (classForm->relisshared)
3008 16 : shared = pgstat_fetch_stat_dbentry(InvalidOid);
3009 : else
3010 240 : dbentry = pgstat_fetch_stat_dbentry(MyDatabaseId);
3011 :
3012 : /* fetch the pgstat table entry */
3013 256 : tabentry = get_pgstat_tabentry_relid(relid, classForm->relisshared,
3014 : shared, dbentry);
3015 :
3016 256 : relation_needs_vacanalyze(relid, avopts, classForm, tabentry,
3017 : effective_multixact_freeze_max_age,
3018 : dovacuum, doanalyze, wraparound);
3019 :
3020 : /* ignore ANALYZE for toast tables */
3021 256 : if (classForm->relkind == RELKIND_TOASTVALUE)
3022 4 : *doanalyze = false;
3023 256 : }
3024 :
3025 : /*
3026 : * relation_needs_vacanalyze
3027 : *
3028 : * Check whether a relation needs to be vacuumed or analyzed; return each into
3029 : * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is
3030 : * being forced because of Xid or multixact wraparound.
3031 : *
3032 : * relopts is a pointer to the AutoVacOpts options (either for itself in the
3033 : * case of a plain table, or for either itself or its parent table in the case
3034 : * of a TOAST table), NULL if none; tabentry is the pgstats entry, which can be
3035 : * NULL.
3036 : *
3037 : * A table needs to be vacuumed if the number of dead tuples exceeds a
3038 : * threshold. This threshold is calculated as
3039 : *
3040 : * threshold = vac_base_thresh + vac_scale_factor * reltuples
3041 : *
3042 : * For analyze, the analysis done is that the number of tuples inserted,
3043 : * deleted and updated since the last analyze exceeds a threshold calculated
3044 : * in the same fashion as above. Note that the collector actually stores
3045 : * the number of tuples (both live and dead) that there were as of the last
3046 : * analyze. This is asymmetric to the VACUUM case.
3047 : *
3048 : * We also force vacuum if the table's relfrozenxid is more than freeze_max_age
3049 : * transactions back, and if its relminmxid is more than
3050 : * multixact_freeze_max_age multixacts back.
3051 : *
3052 : * A table whose autovacuum_enabled option is false is
3053 : * automatically skipped (unless we have to vacuum it due to freeze_max_age).
3054 : * Thus autovacuum can be disabled for specific tables. Also, when the stats
3055 : * collector does not have data about a table, it will be skipped.
3056 : *
3057 : * A table whose vac_base_thresh value is < 0 takes the base value from the
3058 : * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor
3059 : * value < 0 is substituted with the value of
3060 : * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze.
3061 : */
3062 : static void
3063 7464 : relation_needs_vacanalyze(Oid relid,
3064 : AutoVacOpts *relopts,
3065 : Form_pg_class classForm,
3066 : PgStat_StatTabEntry *tabentry,
3067 : int effective_multixact_freeze_max_age,
3068 : /* output params below */
3069 : bool *dovacuum,
3070 : bool *doanalyze,
3071 : bool *wraparound)
3072 : {
3073 : bool force_vacuum;
3074 : bool av_enabled;
3075 : float4 reltuples; /* pg_class.reltuples */
3076 :
3077 : /* constants from reloptions or GUC variables */
3078 : int vac_base_thresh,
3079 : vac_ins_base_thresh,
3080 : anl_base_thresh;
3081 : float4 vac_scale_factor,
3082 : vac_ins_scale_factor,
3083 : anl_scale_factor;
3084 :
3085 : /* thresholds calculated from above constants */
3086 : float4 vacthresh,
3087 : vacinsthresh,
3088 : anlthresh;
3089 :
3090 : /* number of vacuum (resp. analyze) tuples at this time */
3091 : float4 vactuples,
3092 : instuples,
3093 : anltuples;
3094 :
3095 : /* freeze parameters */
3096 : int freeze_max_age;
3097 : int multixact_freeze_max_age;
3098 : TransactionId xidForceLimit;
3099 : MultiXactId multiForceLimit;
3100 :
3101 : AssertArg(classForm != NULL);
3102 : AssertArg(OidIsValid(relid));
3103 :
3104 : /*
3105 : * Determine vacuum/analyze equation parameters. We have two possible
3106 : * sources: the passed reloptions (which could be a main table or a toast
3107 : * table), or the autovacuum GUC variables.
3108 : */
3109 :
3110 : /* -1 in autovac setting means use plain vacuum_scale_factor */
3111 48 : vac_scale_factor = (relopts && relopts->vacuum_scale_factor >= 0)
3112 0 : ? relopts->vacuum_scale_factor
3113 7512 : : autovacuum_vac_scale;
3114 :
3115 48 : vac_base_thresh = (relopts && relopts->vacuum_threshold >= 0)
3116 : ? relopts->vacuum_threshold
3117 7512 : : autovacuum_vac_thresh;
3118 :
3119 48 : vac_ins_scale_factor = (relopts && relopts->vacuum_ins_scale_factor >= 0)
3120 0 : ? relopts->vacuum_ins_scale_factor
3121 7512 : : autovacuum_vac_ins_scale;
3122 :
3123 : /* -1 is used to disable insert vacuums */
3124 48 : vac_ins_base_thresh = (relopts && relopts->vacuum_ins_threshold >= -1)
3125 : ? relopts->vacuum_ins_threshold
3126 7512 : : autovacuum_vac_ins_thresh;
3127 :
3128 48 : anl_scale_factor = (relopts && relopts->analyze_scale_factor >= 0)
3129 0 : ? relopts->analyze_scale_factor
3130 7512 : : autovacuum_anl_scale;
3131 :
3132 48 : anl_base_thresh = (relopts && relopts->analyze_threshold >= 0)
3133 : ? relopts->analyze_threshold
3134 7512 : : autovacuum_anl_thresh;
3135 :
3136 48 : freeze_max_age = (relopts && relopts->freeze_max_age >= 0)
3137 0 : ? Min(relopts->freeze_max_age, autovacuum_freeze_max_age)
3138 7512 : : autovacuum_freeze_max_age;
3139 :
3140 48 : multixact_freeze_max_age = (relopts && relopts->multixact_freeze_max_age >= 0)
3141 0 : ? Min(relopts->multixact_freeze_max_age, effective_multixact_freeze_max_age)
3142 7512 : : effective_multixact_freeze_max_age;
3143 :
3144 7464 : av_enabled = (relopts ? relopts->enabled : true);
3145 :
3146 : /* Force vacuum if table is at risk of wraparound */
3147 7464 : xidForceLimit = recentXid - freeze_max_age;
3148 7464 : if (xidForceLimit < FirstNormalTransactionId)
3149 0 : xidForceLimit -= FirstNormalTransactionId;
3150 14928 : force_vacuum = (TransactionIdIsNormal(classForm->relfrozenxid) &&
3151 7464 : TransactionIdPrecedes(classForm->relfrozenxid,
3152 : xidForceLimit));
3153 7464 : if (!force_vacuum)
3154 : {
3155 7464 : multiForceLimit = recentMulti - multixact_freeze_max_age;
3156 7464 : if (multiForceLimit < FirstMultiXactId)
3157 0 : multiForceLimit -= FirstMultiXactId;
3158 14928 : force_vacuum = MultiXactIdIsValid(classForm->relminmxid) &&
3159 7464 : MultiXactIdPrecedes(classForm->relminmxid, multiForceLimit);
3160 : }
3161 7464 : *wraparound = force_vacuum;
3162 :
3163 : /* User disabled it in pg_class.reloptions? (But ignore if at risk) */
3164 7464 : if (!av_enabled && !force_vacuum)
3165 : {
3166 36 : *doanalyze = false;
3167 36 : *dovacuum = false;
3168 36 : return;
3169 : }
3170 :
3171 : /*
3172 : * If we found the table in the stats hash, and autovacuum is currently
3173 : * enabled, make a threshold-based decision whether to vacuum and/or
3174 : * analyze. If autovacuum is currently disabled, we must be here for
3175 : * anti-wraparound vacuuming only, so don't vacuum (or analyze) anything
3176 : * that's not being forced.
3177 : */
3178 7428 : if (PointerIsValid(tabentry) && AutoVacuumingActive())
3179 : {
3180 3230 : reltuples = classForm->reltuples;
3181 3230 : vactuples = tabentry->n_dead_tuples;
3182 3230 : instuples = tabentry->inserts_since_vacuum;
3183 3230 : anltuples = tabentry->changes_since_analyze;
3184 :
3185 : /* If the table hasn't yet been vacuumed, take reltuples as zero */
3186 3230 : if (reltuples < 0)
3187 390 : reltuples = 0;
3188 :
3189 3230 : vacthresh = (float4) vac_base_thresh + vac_scale_factor * reltuples;
3190 3230 : vacinsthresh = (float4) vac_ins_base_thresh + vac_ins_scale_factor * reltuples;
3191 3230 : anlthresh = (float4) anl_base_thresh + anl_scale_factor * reltuples;
3192 :
3193 : /*
3194 : * Note that we don't need to take special consideration for stat
3195 : * reset, because if that happens, the last vacuum and analyze counts
3196 : * will be reset too.
3197 : */
3198 3230 : if (vac_ins_base_thresh >= 0)
3199 3230 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: %.0f (threshold %.0f), anl: %.0f (threshold %.0f)",
3200 : NameStr(classForm->relname),
3201 : vactuples, vacthresh, instuples, vacinsthresh, anltuples, anlthresh);
3202 : else
3203 0 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: (disabled), anl: %.0f (threshold %.0f)",
3204 : NameStr(classForm->relname),
3205 : vactuples, vacthresh, anltuples, anlthresh);
3206 :
3207 : /* Determine if this table needs vacuum or analyze. */
3208 6260 : *dovacuum = force_vacuum || (vactuples > vacthresh) ||
3209 3030 : (vac_ins_base_thresh >= 0 && instuples > vacinsthresh);
3210 3230 : *doanalyze = (anltuples > anlthresh);
3211 : }
3212 : else
3213 : {
3214 : /*
3215 : * Skip a table not found in stat hash, unless we have to force vacuum
3216 : * for anti-wrap purposes. If it's not acted upon, there's no need to
3217 : * vacuum it.
3218 : */
3219 4198 : *dovacuum = force_vacuum;
3220 4198 : *doanalyze = false;
3221 : }
3222 :
3223 : /* ANALYZE refuses to work with pg_statistic */
3224 7428 : if (relid == StatisticRelationId)
3225 60 : *doanalyze = false;
3226 : }
3227 :
3228 : /*
3229 : * autovacuum_do_vac_analyze
3230 : * Vacuum and/or analyze the specified table
3231 : */
3232 : static void
3233 256 : autovacuum_do_vac_analyze(autovac_table *tab, BufferAccessStrategy bstrategy)
3234 : {
3235 : RangeVar *rangevar;
3236 : VacuumRelation *rel;
3237 : List *rel_list;
3238 :
3239 : /* Let pgstat know what we're doing */
3240 256 : autovac_report_activity(tab);
3241 :
3242 : /* Set up one VacuumRelation target, identified by OID, for vacuum() */
3243 256 : rangevar = makeRangeVar(tab->at_nspname, tab->at_relname, -1);
3244 256 : rel = makeVacuumRelation(rangevar, tab->at_relid, NIL);
3245 256 : rel_list = list_make1(rel);
3246 :
3247 256 : vacuum(rel_list, &tab->at_params, bstrategy, true);
3248 256 : }
3249 :
3250 : /*
3251 : * autovac_report_activity
3252 : * Report to pgstat what autovacuum is doing
3253 : *
3254 : * We send a SQL string corresponding to what the user would see if the
3255 : * equivalent command was to be issued manually.
3256 : *
3257 : * Note we assume that we are going to report the next command as soon as we're
3258 : * done with the current one, and exit right after the last one, so we don't
3259 : * bother to report "<IDLE>" or some such.
3260 : */
3261 : static void
3262 256 : autovac_report_activity(autovac_table *tab)
3263 : {
3264 : #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56)
3265 : char activity[MAX_AUTOVAC_ACTIV_LEN];
3266 : int len;
3267 :
3268 : /* Report the command and possible options */
3269 256 : if (tab->at_params.options & VACOPT_VACUUM)
3270 116 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3271 : "autovacuum: VACUUM%s",
3272 116 : tab->at_params.options & VACOPT_ANALYZE ? " ANALYZE" : "");
3273 : else
3274 140 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3275 : "autovacuum: ANALYZE");
3276 :
3277 : /*
3278 : * Report the qualified name of the relation.
3279 : */
3280 256 : len = strlen(activity);
3281 :
3282 256 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3283 : " %s.%s%s", tab->at_nspname, tab->at_relname,
3284 256 : tab->at_params.is_wraparound ? " (to prevent wraparound)" : "");
3285 :
3286 : /* Set statement_timestamp() to current time for pg_stat_activity */
3287 256 : SetCurrentStatementStartTimestamp();
3288 :
3289 256 : pgstat_report_activity(STATE_RUNNING, activity);
3290 256 : }
3291 :
3292 : /*
3293 : * autovac_report_workitem
3294 : * Report to pgstat that autovacuum is processing a work item
3295 : */
3296 : static void
3297 6 : autovac_report_workitem(AutoVacuumWorkItem *workitem,
3298 : const char *nspname, const char *relname)
3299 : {
3300 : char activity[MAX_AUTOVAC_ACTIV_LEN + 12 + 2];
3301 : char blk[12 + 2];
3302 : int len;
3303 :
3304 6 : switch (workitem->avw_type)
3305 : {
3306 6 : case AVW_BRINSummarizeRange:
3307 6 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3308 : "autovacuum: BRIN summarize");
3309 6 : break;
3310 : }
3311 :
3312 : /*
3313 : * Report the qualified name of the relation, and the block number if any
3314 : */
3315 6 : len = strlen(activity);
3316 :
3317 6 : if (BlockNumberIsValid(workitem->avw_blockNumber))
3318 6 : snprintf(blk, sizeof(blk), " %u", workitem->avw_blockNumber);
3319 : else
3320 0 : blk[0] = '\0';
3321 :
3322 6 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3323 : " %s.%s%s", nspname, relname, blk);
3324 :
3325 : /* Set statement_timestamp() to current time for pg_stat_activity */
3326 6 : SetCurrentStatementStartTimestamp();
3327 :
3328 6 : pgstat_report_activity(STATE_RUNNING, activity);
3329 6 : }
3330 :
3331 : /*
3332 : * AutoVacuumingActive
3333 : * Check GUC vars and report whether the autovacuum process should be
3334 : * running.
3335 : */
3336 : bool
3337 10590 : AutoVacuumingActive(void)
3338 : {
3339 10590 : if (!autovacuum_start_daemon || !pgstat_track_counts)
3340 1788 : return false;
3341 8802 : return true;
3342 : }
3343 :
3344 : /*
3345 : * Request one work item to the next autovacuum run processing our database.
3346 : * Return false if the request can't be recorded.
3347 : */
3348 : bool
3349 6 : AutoVacuumRequestWork(AutoVacuumWorkItemType type, Oid relationId,
3350 : BlockNumber blkno)
3351 : {
3352 : int i;
3353 6 : bool result = false;
3354 :
3355 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
3356 :
3357 : /*
3358 : * Locate an unused work item and fill it with the given data.
3359 : */
3360 12 : for (i = 0; i < NUM_WORKITEMS; i++)
3361 : {
3362 12 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
3363 :
3364 12 : if (workitem->avw_used)
3365 6 : continue;
3366 :
3367 6 : workitem->avw_used = true;
3368 6 : workitem->avw_active = false;
3369 6 : workitem->avw_type = type;
3370 6 : workitem->avw_database = MyDatabaseId;
3371 6 : workitem->avw_relation = relationId;
3372 6 : workitem->avw_blockNumber = blkno;
3373 6 : result = true;
3374 :
3375 : /* done */
3376 6 : break;
3377 : }
3378 :
3379 6 : LWLockRelease(AutovacuumLock);
3380 :
3381 6 : return result;
3382 : }
3383 :
3384 : /*
3385 : * autovac_init
3386 : * This is called at postmaster initialization.
3387 : *
3388 : * All we do here is annoy the user if he got it wrong.
3389 : */
3390 : void
3391 940 : autovac_init(void)
3392 : {
3393 940 : if (autovacuum_start_daemon && !pgstat_track_counts)
3394 0 : ereport(WARNING,
3395 : (errmsg("autovacuum not started because of misconfiguration"),
3396 : errhint("Enable the \"track_counts\" option.")));
3397 940 : }
3398 :
3399 : /*
3400 : * IsAutoVacuum functions
3401 : * Return whether this is either a launcher autovacuum process or a worker
3402 : * process.
3403 : */
3404 : bool
3405 72116 : IsAutoVacuumLauncherProcess(void)
3406 : {
3407 72116 : return am_autovacuum_launcher;
3408 : }
3409 :
3410 : bool
3411 392546 : IsAutoVacuumWorkerProcess(void)
3412 : {
3413 392546 : return am_autovacuum_worker;
3414 : }
3415 :
3416 :
3417 : /*
3418 : * AutoVacuumShmemSize
3419 : * Compute space needed for autovacuum-related shared memory
3420 : */
3421 : Size
3422 6736 : AutoVacuumShmemSize(void)
3423 : {
3424 : Size size;
3425 :
3426 : /*
3427 : * Need the fixed struct and the array of WorkerInfoData.
3428 : */
3429 6736 : size = sizeof(AutoVacuumShmemStruct);
3430 6736 : size = MAXALIGN(size);
3431 6736 : size = add_size(size, mul_size(autovacuum_max_workers,
3432 : sizeof(WorkerInfoData)));
3433 6736 : return size;
3434 : }
3435 :
3436 : /*
3437 : * AutoVacuumShmemInit
3438 : * Allocate and initialize autovacuum-related shared memory
3439 : */
3440 : void
3441 2894 : AutoVacuumShmemInit(void)
3442 : {
3443 : bool found;
3444 :
3445 2894 : AutoVacuumShmem = (AutoVacuumShmemStruct *)
3446 2894 : ShmemInitStruct("AutoVacuum Data",
3447 : AutoVacuumShmemSize(),
3448 : &found);
3449 :
3450 2894 : if (!IsUnderPostmaster)
3451 : {
3452 : WorkerInfo worker;
3453 : int i;
3454 :
3455 : Assert(!found);
3456 :
3457 2894 : AutoVacuumShmem->av_launcherpid = 0;
3458 2894 : dlist_init(&AutoVacuumShmem->av_freeWorkers);
3459 2894 : dlist_init(&AutoVacuumShmem->av_runningWorkers);
3460 2894 : AutoVacuumShmem->av_startingWorker = NULL;
3461 2894 : memset(AutoVacuumShmem->av_workItems, 0,
3462 : sizeof(AutoVacuumWorkItem) * NUM_WORKITEMS);
3463 :
3464 2894 : worker = (WorkerInfo) ((char *) AutoVacuumShmem +
3465 : MAXALIGN(sizeof(AutoVacuumShmemStruct)));
3466 :
3467 : /* initialize the WorkerInfo free list */
3468 11576 : for (i = 0; i < autovacuum_max_workers; i++)
3469 8682 : dlist_push_head(&AutoVacuumShmem->av_freeWorkers,
3470 8682 : &worker[i].wi_links);
3471 : }
3472 : else
3473 : Assert(found);
3474 2894 : }
3475 :
3476 : /*
3477 : * autovac_refresh_stats
3478 : * Refresh pgstats data for an autovacuum process
3479 : *
3480 : * Cause the next pgstats read operation to obtain fresh data, but throttle
3481 : * such refreshing in the autovacuum launcher. This is mostly to avoid
3482 : * rereading the pgstats files too many times in quick succession when there
3483 : * are many databases.
3484 : *
3485 : * Note: we avoid throttling in the autovac worker, as it would be
3486 : * counterproductive in the recheck logic.
3487 : */
3488 : static void
3489 1246 : autovac_refresh_stats(void)
3490 : {
3491 1246 : if (IsAutoVacuumLauncherProcess())
3492 : {
3493 : static TimestampTz last_read = 0;
3494 : TimestampTz current_time;
3495 :
3496 990 : current_time = GetCurrentTimestamp();
3497 :
3498 990 : if (!TimestampDifferenceExceeds(last_read, current_time,
3499 : STATS_READ_DELAY))
3500 456 : return;
3501 :
3502 534 : last_read = current_time;
3503 : }
3504 :
3505 790 : pgstat_clear_snapshot();
3506 : }
|
