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 fetch the last time the table was vacuumed from
48 : * pgstats just before vacuuming each table, to avoid vacuuming a table that
49 : * was just finished being vacuumed by another worker and thus is no longer
50 : * noted in shared memory. However, there is a small window (due to not yet
51 : * holding the relation lock) during which a worker may choose a table that was
52 : * already vacuumed; this is a bug in the current design.
53 : *
54 : * Portions Copyright (c) 1996-2022, PostgreSQL Global Development Group
55 : * Portions Copyright (c) 1994, Regents of the University of California
56 : *
57 : *
58 : * IDENTIFICATION
59 : * src/backend/postmaster/autovacuum.c
60 : *
61 : *-------------------------------------------------------------------------
62 : */
63 : #include "postgres.h"
64 :
65 : #include <signal.h>
66 : #include <sys/time.h>
67 : #include <unistd.h>
68 :
69 : #include "access/heapam.h"
70 : #include "access/htup_details.h"
71 : #include "access/multixact.h"
72 : #include "access/reloptions.h"
73 : #include "access/tableam.h"
74 : #include "access/transam.h"
75 : #include "access/xact.h"
76 : #include "catalog/dependency.h"
77 : #include "catalog/namespace.h"
78 : #include "catalog/pg_database.h"
79 : #include "commands/dbcommands.h"
80 : #include "commands/vacuum.h"
81 : #include "lib/ilist.h"
82 : #include "libpq/pqsignal.h"
83 : #include "miscadmin.h"
84 : #include "nodes/makefuncs.h"
85 : #include "pgstat.h"
86 : #include "postmaster/autovacuum.h"
87 : #include "postmaster/fork_process.h"
88 : #include "postmaster/interrupt.h"
89 : #include "postmaster/postmaster.h"
90 : #include "storage/bufmgr.h"
91 : #include "storage/ipc.h"
92 : #include "storage/latch.h"
93 : #include "storage/lmgr.h"
94 : #include "storage/pmsignal.h"
95 : #include "storage/proc.h"
96 : #include "storage/procsignal.h"
97 : #include "storage/sinvaladt.h"
98 : #include "storage/smgr.h"
99 : #include "tcop/tcopprot.h"
100 : #include "utils/fmgroids.h"
101 : #include "utils/fmgrprotos.h"
102 : #include "utils/lsyscache.h"
103 : #include "utils/memutils.h"
104 : #include "utils/ps_status.h"
105 : #include "utils/rel.h"
106 : #include "utils/snapmgr.h"
107 : #include "utils/syscache.h"
108 : #include "utils/timeout.h"
109 : #include "utils/timestamp.h"
110 :
111 :
112 : /*
113 : * GUC parameters
114 : */
115 : bool autovacuum_start_daemon = false;
116 : int autovacuum_max_workers;
117 : int autovacuum_work_mem = -1;
118 : int autovacuum_naptime;
119 : int autovacuum_vac_thresh;
120 : double autovacuum_vac_scale;
121 : int autovacuum_vac_ins_thresh;
122 : double autovacuum_vac_ins_scale;
123 : int autovacuum_anl_thresh;
124 : double autovacuum_anl_scale;
125 : int autovacuum_freeze_max_age;
126 : int autovacuum_multixact_freeze_max_age;
127 :
128 : double autovacuum_vac_cost_delay;
129 : int autovacuum_vac_cost_limit;
130 :
131 : int Log_autovacuum_min_duration = 600000;
132 :
133 : /* the minimum allowed time between two awakenings of the launcher */
134 : #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */
135 : #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */
136 :
137 : /* Flags to tell if we are in an autovacuum process */
138 : static bool am_autovacuum_launcher = false;
139 : static bool am_autovacuum_worker = false;
140 :
141 : /* Flags set by signal handlers */
142 : static volatile sig_atomic_t got_SIGUSR2 = false;
143 :
144 : /* Comparison points for determining whether freeze_max_age is exceeded */
145 : static TransactionId recentXid;
146 : static MultiXactId recentMulti;
147 :
148 : /* Default freeze ages to use for autovacuum (varies by database) */
149 : static int default_freeze_min_age;
150 : static int default_freeze_table_age;
151 : static int default_multixact_freeze_min_age;
152 : static int default_multixact_freeze_table_age;
153 :
154 : /* Memory context for long-lived data */
155 : static MemoryContext AutovacMemCxt;
156 :
157 : /* struct to keep track of databases in launcher */
158 : typedef struct avl_dbase
159 : {
160 : Oid adl_datid; /* hash key -- must be first */
161 : TimestampTz adl_next_worker;
162 : int adl_score;
163 : dlist_node adl_node;
164 : } avl_dbase;
165 :
166 : /* struct to keep track of databases in worker */
167 : typedef struct avw_dbase
168 : {
169 : Oid adw_datid;
170 : char *adw_name;
171 : TransactionId adw_frozenxid;
172 : MultiXactId adw_minmulti;
173 : PgStat_StatDBEntry *adw_entry;
174 : } avw_dbase;
175 :
176 : /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */
177 : typedef struct av_relation
178 : {
179 : Oid ar_toastrelid; /* hash key - must be first */
180 : Oid ar_relid;
181 : bool ar_hasrelopts;
182 : AutoVacOpts ar_reloptions; /* copy of AutoVacOpts from the main table's
183 : * reloptions, or NULL if none */
184 : } av_relation;
185 :
186 : /* struct to keep track of tables to vacuum and/or analyze, after rechecking */
187 : typedef struct autovac_table
188 : {
189 : Oid at_relid;
190 : VacuumParams at_params;
191 : double at_vacuum_cost_delay;
192 : int at_vacuum_cost_limit;
193 : bool at_dobalance;
194 : bool at_sharedrel;
195 : char *at_relname;
196 : char *at_nspname;
197 : char *at_datname;
198 : } autovac_table;
199 :
200 : /*-------------
201 : * This struct holds information about a single worker's whereabouts. We keep
202 : * an array of these in shared memory, sized according to
203 : * autovacuum_max_workers.
204 : *
205 : * wi_links entry into free list or running list
206 : * wi_dboid OID of the database this worker is supposed to work on
207 : * wi_tableoid OID of the table currently being vacuumed, if any
208 : * wi_sharedrel flag indicating whether table is marked relisshared
209 : * wi_proc pointer to PGPROC of the running worker, NULL if not started
210 : * wi_launchtime Time at which this worker was launched
211 : * wi_cost_* Vacuum cost-based delay parameters current in this worker
212 : *
213 : * All fields are protected by AutovacuumLock, except for wi_tableoid and
214 : * wi_sharedrel which are protected by AutovacuumScheduleLock (note these
215 : * two fields are read-only for everyone except that worker itself).
216 : *-------------
217 : */
218 : typedef struct WorkerInfoData
219 : {
220 : dlist_node wi_links;
221 : Oid wi_dboid;
222 : Oid wi_tableoid;
223 : PGPROC *wi_proc;
224 : TimestampTz wi_launchtime;
225 : bool wi_dobalance;
226 : bool wi_sharedrel;
227 : double wi_cost_delay;
228 : int wi_cost_limit;
229 : int wi_cost_limit_base;
230 : } WorkerInfoData;
231 :
232 : typedef struct WorkerInfoData *WorkerInfo;
233 :
234 : /*
235 : * Possible signals received by the launcher from remote processes. These are
236 : * stored atomically in shared memory so that other processes can set them
237 : * without locking.
238 : */
239 : typedef enum
240 : {
241 : AutoVacForkFailed, /* failed trying to start a worker */
242 : AutoVacRebalance, /* rebalance the cost limits */
243 : AutoVacNumSignals /* must be last */
244 : } AutoVacuumSignal;
245 :
246 : /*
247 : * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This
248 : * list is mostly protected by AutovacuumLock, except that if an item is
249 : * marked 'active' other processes must not modify the work-identifying
250 : * members.
251 : */
252 : typedef struct AutoVacuumWorkItem
253 : {
254 : AutoVacuumWorkItemType avw_type;
255 : bool avw_used; /* below data is valid */
256 : bool avw_active; /* being processed */
257 : Oid avw_database;
258 : Oid avw_relation;
259 : BlockNumber avw_blockNumber;
260 : } AutoVacuumWorkItem;
261 :
262 : #define NUM_WORKITEMS 256
263 :
264 : /*-------------
265 : * The main autovacuum shmem struct. On shared memory we store this main
266 : * struct and the array of WorkerInfo structs. This struct keeps:
267 : *
268 : * av_signal set by other processes to indicate various conditions
269 : * av_launcherpid the PID of the autovacuum launcher
270 : * av_freeWorkers the WorkerInfo freelist
271 : * av_runningWorkers the WorkerInfo non-free queue
272 : * av_startingWorker pointer to WorkerInfo currently being started (cleared by
273 : * the worker itself as soon as it's up and running)
274 : * av_workItems work item array
275 : *
276 : * This struct is protected by AutovacuumLock, except for av_signal and parts
277 : * of the worker list (see above).
278 : *-------------
279 : */
280 : typedef struct
281 : {
282 : sig_atomic_t av_signal[AutoVacNumSignals];
283 : pid_t av_launcherpid;
284 : dlist_head av_freeWorkers;
285 : dlist_head av_runningWorkers;
286 : WorkerInfo av_startingWorker;
287 : AutoVacuumWorkItem av_workItems[NUM_WORKITEMS];
288 : } AutoVacuumShmemStruct;
289 :
290 : static AutoVacuumShmemStruct *AutoVacuumShmem;
291 :
292 : /*
293 : * the database list (of avl_dbase elements) in the launcher, and the context
294 : * that contains it
295 : */
296 : static dlist_head DatabaseList = DLIST_STATIC_INIT(DatabaseList);
297 : static MemoryContext DatabaseListCxt = NULL;
298 :
299 : /* Pointer to my own WorkerInfo, valid on each worker */
300 : static WorkerInfo MyWorkerInfo = NULL;
301 :
302 : /* PID of launcher, valid only in worker while shutting down */
303 : int AutovacuumLauncherPid = 0;
304 :
305 : #ifdef EXEC_BACKEND
306 : static pid_t avlauncher_forkexec(void);
307 : static pid_t avworker_forkexec(void);
308 : #endif
309 : NON_EXEC_STATIC void AutoVacWorkerMain(int argc, char *argv[]) pg_attribute_noreturn();
310 : NON_EXEC_STATIC void AutoVacLauncherMain(int argc, char *argv[]) pg_attribute_noreturn();
311 :
312 : static Oid do_start_worker(void);
313 : static void HandleAutoVacLauncherInterrupts(void);
314 : static void AutoVacLauncherShutdown(void) pg_attribute_noreturn();
315 : static void launcher_determine_sleep(bool canlaunch, bool recursing,
316 : struct timeval *nap);
317 : static void launch_worker(TimestampTz now);
318 : static List *get_database_list(void);
319 : static void rebuild_database_list(Oid newdb);
320 : static int db_comparator(const void *a, const void *b);
321 : static void autovac_balance_cost(void);
322 :
323 : static void do_autovacuum(void);
324 : static void FreeWorkerInfo(int code, Datum arg);
325 :
326 : static autovac_table *table_recheck_autovac(Oid relid, HTAB *table_toast_map,
327 : TupleDesc pg_class_desc,
328 : int effective_multixact_freeze_max_age);
329 : static void recheck_relation_needs_vacanalyze(Oid relid, AutoVacOpts *avopts,
330 : Form_pg_class classForm,
331 : int effective_multixact_freeze_max_age,
332 : bool *dovacuum, bool *doanalyze, bool *wraparound);
333 : static void relation_needs_vacanalyze(Oid relid, AutoVacOpts *relopts,
334 : Form_pg_class classForm,
335 : PgStat_StatTabEntry *tabentry,
336 : int effective_multixact_freeze_max_age,
337 : bool *dovacuum, bool *doanalyze, bool *wraparound);
338 :
339 : static void autovacuum_do_vac_analyze(autovac_table *tab,
340 : BufferAccessStrategy bstrategy);
341 : static AutoVacOpts *extract_autovac_opts(HeapTuple tup,
342 : TupleDesc pg_class_desc);
343 : static void perform_work_item(AutoVacuumWorkItem *workitem);
344 : static void autovac_report_activity(autovac_table *tab);
345 : static void autovac_report_workitem(AutoVacuumWorkItem *workitem,
346 : const char *nspname, const char *relname);
347 : static void avl_sigusr2_handler(SIGNAL_ARGS);
348 :
349 :
350 :
351 : /********************************************************************
352 : * AUTOVACUUM LAUNCHER CODE
353 : ********************************************************************/
354 :
355 : #ifdef EXEC_BACKEND
356 : /*
357 : * forkexec routine for the autovacuum launcher process.
358 : *
359 : * Format up the arglist, then fork and exec.
360 : */
361 : static pid_t
362 : avlauncher_forkexec(void)
363 : {
364 : char *av[10];
365 : int ac = 0;
366 :
367 : av[ac++] = "postgres";
368 : av[ac++] = "--forkavlauncher";
369 : av[ac++] = NULL; /* filled in by postmaster_forkexec */
370 : av[ac] = NULL;
371 :
372 : Assert(ac < lengthof(av));
373 :
374 : return postmaster_forkexec(ac, av);
375 : }
376 :
377 : /*
378 : * We need this set from the outside, before InitProcess is called
379 : */
380 : void
381 : AutovacuumLauncherIAm(void)
382 : {
383 : am_autovacuum_launcher = true;
384 : }
385 : #endif
386 :
387 : /*
388 : * Main entry point for autovacuum launcher process, to be called from the
389 : * postmaster.
390 : */
391 : int
392 868 : StartAutoVacLauncher(void)
393 : {
394 : pid_t AutoVacPID;
395 :
396 : #ifdef EXEC_BACKEND
397 : switch ((AutoVacPID = avlauncher_forkexec()))
398 : #else
399 868 : switch ((AutoVacPID = fork_process()))
400 : #endif
401 : {
402 0 : case -1:
403 0 : ereport(LOG,
404 : (errmsg("could not fork autovacuum launcher process: %m")));
405 0 : return 0;
406 :
407 : #ifndef EXEC_BACKEND
408 538 : case 0:
409 : /* in postmaster child ... */
410 538 : InitPostmasterChild();
411 :
412 : /* Close the postmaster's sockets */
413 538 : ClosePostmasterPorts(false);
414 :
415 538 : AutoVacLauncherMain(0, NULL);
416 : break;
417 : #endif
418 868 : default:
419 868 : return (int) AutoVacPID;
420 : }
421 :
422 : /* shouldn't get here */
423 : return 0;
424 : }
425 :
426 : /*
427 : * Main loop for the autovacuum launcher process.
428 : */
429 : NON_EXEC_STATIC void
430 538 : AutoVacLauncherMain(int argc, char *argv[])
431 : {
432 : sigjmp_buf local_sigjmp_buf;
433 :
434 538 : am_autovacuum_launcher = true;
435 :
436 538 : MyBackendType = B_AUTOVAC_LAUNCHER;
437 538 : init_ps_display(NULL);
438 :
439 538 : ereport(DEBUG1,
440 : (errmsg_internal("autovacuum launcher started")));
441 :
442 538 : if (PostAuthDelay)
443 0 : pg_usleep(PostAuthDelay * 1000000L);
444 :
445 538 : SetProcessingMode(InitProcessing);
446 :
447 : /*
448 : * Set up signal handlers. We operate on databases much like a regular
449 : * backend, so we use the same signal handling. See equivalent code in
450 : * tcop/postgres.c.
451 : */
452 538 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
453 538 : pqsignal(SIGINT, StatementCancelHandler);
454 538 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
455 : /* SIGQUIT handler was already set up by InitPostmasterChild */
456 :
457 538 : InitializeTimeouts(); /* establishes SIGALRM handler */
458 :
459 538 : pqsignal(SIGPIPE, SIG_IGN);
460 538 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
461 538 : pqsignal(SIGUSR2, avl_sigusr2_handler);
462 538 : pqsignal(SIGFPE, FloatExceptionHandler);
463 538 : pqsignal(SIGCHLD, SIG_DFL);
464 :
465 : /*
466 : * Create a per-backend PGPROC struct in shared memory, except in the
467 : * EXEC_BACKEND case where this was done in SubPostmasterMain. We must do
468 : * this before we can use LWLocks (and in the EXEC_BACKEND case we already
469 : * had to do some stuff with LWLocks).
470 : */
471 : #ifndef EXEC_BACKEND
472 538 : InitProcess();
473 : #endif
474 :
475 : /* Early initialization */
476 538 : BaseInit();
477 :
478 538 : InitPostgres(NULL, InvalidOid, NULL, InvalidOid, NULL, false);
479 :
480 538 : SetProcessingMode(NormalProcessing);
481 :
482 : /*
483 : * Create a memory context that we will do all our work in. We do this so
484 : * that we can reset the context during error recovery and thereby avoid
485 : * possible memory leaks.
486 : */
487 538 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
488 : "Autovacuum Launcher",
489 : ALLOCSET_DEFAULT_SIZES);
490 538 : MemoryContextSwitchTo(AutovacMemCxt);
491 :
492 : /*
493 : * If an exception is encountered, processing resumes here.
494 : *
495 : * This code is a stripped down version of PostgresMain error recovery.
496 : *
497 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
498 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
499 : * signals other than SIGQUIT will be blocked until we complete error
500 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
501 : * call redundant, but it is not since InterruptPending might be set
502 : * already.
503 : */
504 538 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
505 : {
506 : /* since not using PG_TRY, must reset error stack by hand */
507 0 : error_context_stack = NULL;
508 :
509 : /* Prevents interrupts while cleaning up */
510 0 : HOLD_INTERRUPTS();
511 :
512 : /* Forget any pending QueryCancel or timeout request */
513 0 : disable_all_timeouts(false);
514 0 : QueryCancelPending = false; /* second to avoid race condition */
515 :
516 : /* Report the error to the server log */
517 0 : EmitErrorReport();
518 :
519 : /* Abort the current transaction in order to recover */
520 0 : AbortCurrentTransaction();
521 :
522 : /*
523 : * Release any other resources, for the case where we were not in a
524 : * transaction.
525 : */
526 0 : LWLockReleaseAll();
527 0 : pgstat_report_wait_end();
528 0 : AbortBufferIO();
529 0 : UnlockBuffers();
530 : /* this is probably dead code, but let's be safe: */
531 0 : if (AuxProcessResourceOwner)
532 0 : ReleaseAuxProcessResources(false);
533 0 : AtEOXact_Buffers(false);
534 0 : AtEOXact_SMgr();
535 0 : AtEOXact_Files(false);
536 0 : AtEOXact_HashTables(false);
537 :
538 : /*
539 : * Now return to normal top-level context and clear ErrorContext for
540 : * next time.
541 : */
542 0 : MemoryContextSwitchTo(AutovacMemCxt);
543 0 : FlushErrorState();
544 :
545 : /* Flush any leaked data in the top-level context */
546 0 : MemoryContextResetAndDeleteChildren(AutovacMemCxt);
547 :
548 : /* don't leave dangling pointers to freed memory */
549 0 : DatabaseListCxt = NULL;
550 0 : dlist_init(&DatabaseList);
551 :
552 : /* Now we can allow interrupts again */
553 0 : RESUME_INTERRUPTS();
554 :
555 : /* if in shutdown mode, no need for anything further; just go away */
556 0 : if (ShutdownRequestPending)
557 0 : AutoVacLauncherShutdown();
558 :
559 : /*
560 : * Sleep at least 1 second after any error. We don't want to be
561 : * filling the error logs as fast as we can.
562 : */
563 0 : pg_usleep(1000000L);
564 : }
565 :
566 : /* We can now handle ereport(ERROR) */
567 538 : PG_exception_stack = &local_sigjmp_buf;
568 :
569 : /* must unblock signals before calling rebuild_database_list */
570 538 : PG_SETMASK(&UnBlockSig);
571 :
572 : /*
573 : * Set always-secure search path. Launcher doesn't connect to a database,
574 : * so this has no effect.
575 : */
576 538 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
577 :
578 : /*
579 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
580 : * in postgresql.conf. We don't really want such a dangerous option being
581 : * applied non-interactively.
582 : */
583 538 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
584 :
585 : /*
586 : * Force settable timeouts off to avoid letting these settings prevent
587 : * regular maintenance from being executed.
588 : */
589 538 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
590 538 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
591 538 : SetConfigOption("idle_in_transaction_session_timeout", "0",
592 : PGC_SUSET, PGC_S_OVERRIDE);
593 :
594 : /*
595 : * Force default_transaction_isolation to READ COMMITTED. We don't want
596 : * to pay the overhead of serializable mode, nor add any risk of causing
597 : * deadlocks or delaying other transactions.
598 : */
599 538 : SetConfigOption("default_transaction_isolation", "read committed",
600 : PGC_SUSET, PGC_S_OVERRIDE);
601 :
602 : /*
603 : * Even when system is configured to use a different fetch consistency,
604 : * for autovac we always want fresh stats.
605 : */
606 538 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
607 :
608 : /*
609 : * In emergency mode, just start a worker (unless shutdown was requested)
610 : * and go away.
611 : */
612 538 : if (!AutoVacuumingActive())
613 : {
614 0 : if (!ShutdownRequestPending)
615 0 : do_start_worker();
616 0 : proc_exit(0); /* done */
617 : }
618 :
619 538 : AutoVacuumShmem->av_launcherpid = MyProcPid;
620 :
621 : /*
622 : * Create the initial database list. The invariant we want this list to
623 : * keep is that it's ordered by decreasing next_time. As soon as an entry
624 : * is updated to a higher time, it will be moved to the front (which is
625 : * correct because the only operation is to add autovacuum_naptime to the
626 : * entry, and time always increases).
627 : */
628 538 : rebuild_database_list(InvalidOid);
629 :
630 : /* loop until shutdown request */
631 2868 : while (!ShutdownRequestPending)
632 : {
633 : struct timeval nap;
634 2868 : TimestampTz current_time = 0;
635 : bool can_launch;
636 :
637 : /*
638 : * This loop is a bit different from the normal use of WaitLatch,
639 : * because we'd like to sleep before the first launch of a child
640 : * process. So it's WaitLatch, then ResetLatch, then check for
641 : * wakening conditions.
642 : */
643 :
644 2868 : launcher_determine_sleep(!dlist_is_empty(&AutoVacuumShmem->av_freeWorkers),
645 2868 : false, &nap);
646 :
647 : /*
648 : * Wait until naptime expires or we get some type of signal (all the
649 : * signal handlers will wake us by calling SetLatch).
650 : */
651 2868 : (void) WaitLatch(MyLatch,
652 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
653 2868 : (nap.tv_sec * 1000L) + (nap.tv_usec / 1000L),
654 : WAIT_EVENT_AUTOVACUUM_MAIN);
655 :
656 2862 : ResetLatch(MyLatch);
657 :
658 2862 : HandleAutoVacLauncherInterrupts();
659 :
660 : /*
661 : * a worker finished, or postmaster signaled failure to start a worker
662 : */
663 2330 : if (got_SIGUSR2)
664 : {
665 32 : got_SIGUSR2 = false;
666 :
667 : /* rebalance cost limits, if needed */
668 32 : if (AutoVacuumShmem->av_signal[AutoVacRebalance])
669 : {
670 16 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
671 16 : AutoVacuumShmem->av_signal[AutoVacRebalance] = false;
672 16 : autovac_balance_cost();
673 16 : LWLockRelease(AutovacuumLock);
674 : }
675 :
676 32 : if (AutoVacuumShmem->av_signal[AutoVacForkFailed])
677 : {
678 : /*
679 : * If the postmaster failed to start a new worker, we sleep
680 : * for a little while and resend the signal. The new worker's
681 : * state is still in memory, so this is sufficient. After
682 : * that, we restart the main loop.
683 : *
684 : * XXX should we put a limit to the number of times we retry?
685 : * I don't think it makes much sense, because a future start
686 : * of a worker will continue to fail in the same way.
687 : */
688 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = false;
689 0 : pg_usleep(1000000L); /* 1s */
690 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
691 0 : continue;
692 : }
693 : }
694 :
695 : /*
696 : * There are some conditions that we need to check before trying to
697 : * start a worker. First, we need to make sure that there is a worker
698 : * slot available. Second, we need to make sure that no other worker
699 : * failed while starting up.
700 : */
701 :
702 2330 : current_time = GetCurrentTimestamp();
703 2330 : LWLockAcquire(AutovacuumLock, LW_SHARED);
704 :
705 2330 : can_launch = !dlist_is_empty(&AutoVacuumShmem->av_freeWorkers);
706 :
707 2330 : if (AutoVacuumShmem->av_startingWorker != NULL)
708 : {
709 : int waittime;
710 0 : WorkerInfo worker = AutoVacuumShmem->av_startingWorker;
711 :
712 : /*
713 : * We can't launch another worker when another one is still
714 : * starting up (or failed while doing so), so just sleep for a bit
715 : * more; that worker will wake us up again as soon as it's ready.
716 : * We will only wait autovacuum_naptime seconds (up to a maximum
717 : * of 60 seconds) for this to happen however. Note that failure
718 : * to connect to a particular database is not a problem here,
719 : * because the worker removes itself from the startingWorker
720 : * pointer before trying to connect. Problems detected by the
721 : * postmaster (like fork() failure) are also reported and handled
722 : * differently. The only problems that may cause this code to
723 : * fire are errors in the earlier sections of AutoVacWorkerMain,
724 : * before the worker removes the WorkerInfo from the
725 : * startingWorker pointer.
726 : */
727 0 : waittime = Min(autovacuum_naptime, 60) * 1000;
728 0 : if (TimestampDifferenceExceeds(worker->wi_launchtime, current_time,
729 : waittime))
730 : {
731 0 : LWLockRelease(AutovacuumLock);
732 0 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
733 :
734 : /*
735 : * No other process can put a worker in starting mode, so if
736 : * startingWorker is still INVALID after exchanging our lock,
737 : * we assume it's the same one we saw above (so we don't
738 : * recheck the launch time).
739 : */
740 0 : if (AutoVacuumShmem->av_startingWorker != NULL)
741 : {
742 0 : worker = AutoVacuumShmem->av_startingWorker;
743 0 : worker->wi_dboid = InvalidOid;
744 0 : worker->wi_tableoid = InvalidOid;
745 0 : worker->wi_sharedrel = false;
746 0 : worker->wi_proc = NULL;
747 0 : worker->wi_launchtime = 0;
748 0 : dlist_push_head(&AutoVacuumShmem->av_freeWorkers,
749 : &worker->wi_links);
750 0 : AutoVacuumShmem->av_startingWorker = NULL;
751 0 : ereport(WARNING,
752 : errmsg("autovacuum worker took too long to start; canceled"));
753 : }
754 : }
755 : else
756 0 : can_launch = false;
757 : }
758 2330 : LWLockRelease(AutovacuumLock); /* either shared or exclusive */
759 :
760 : /* if we can't do anything, just go back to sleep */
761 2330 : if (!can_launch)
762 0 : continue;
763 :
764 : /* We're OK to start a new worker */
765 :
766 2330 : if (dlist_is_empty(&DatabaseList))
767 : {
768 : /*
769 : * Special case when the list is empty: start a worker right away.
770 : * This covers the initial case, when no database is in pgstats
771 : * (thus the list is empty). Note that the constraints in
772 : * launcher_determine_sleep keep us from starting workers too
773 : * quickly (at most once every autovacuum_naptime when the list is
774 : * empty).
775 : */
776 54 : launch_worker(current_time);
777 : }
778 : else
779 : {
780 : /*
781 : * because rebuild_database_list constructs a list with most
782 : * distant adl_next_worker first, we obtain our database from the
783 : * tail of the list.
784 : */
785 : avl_dbase *avdb;
786 :
787 2276 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
788 :
789 : /*
790 : * launch a worker if next_worker is right now or it is in the
791 : * past
792 : */
793 2276 : if (TimestampDifferenceExceeds(avdb->adl_next_worker,
794 : current_time, 0))
795 16 : launch_worker(current_time);
796 : }
797 : }
798 :
799 0 : AutoVacLauncherShutdown();
800 : }
801 :
802 : /*
803 : * Process any new interrupts.
804 : */
805 : static void
806 2862 : HandleAutoVacLauncherInterrupts(void)
807 : {
808 : /* the normal shutdown case */
809 2862 : if (ShutdownRequestPending)
810 532 : AutoVacLauncherShutdown();
811 :
812 2330 : if (ConfigReloadPending)
813 : {
814 38 : ConfigReloadPending = false;
815 38 : ProcessConfigFile(PGC_SIGHUP);
816 :
817 : /* shutdown requested in config file? */
818 38 : if (!AutoVacuumingActive())
819 0 : AutoVacLauncherShutdown();
820 :
821 : /* rebalance in case the default cost parameters changed */
822 38 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
823 38 : autovac_balance_cost();
824 38 : LWLockRelease(AutovacuumLock);
825 :
826 : /* rebuild the list in case the naptime changed */
827 38 : rebuild_database_list(InvalidOid);
828 : }
829 :
830 : /* Process barrier events */
831 2330 : if (ProcSignalBarrierPending)
832 36 : ProcessProcSignalBarrier();
833 :
834 : /* Perform logging of memory contexts of this process */
835 2330 : if (LogMemoryContextPending)
836 0 : ProcessLogMemoryContextInterrupt();
837 :
838 : /* Process sinval catchup interrupts that happened while sleeping */
839 2330 : ProcessCatchupInterrupt();
840 2330 : }
841 :
842 : /*
843 : * Perform a normal exit from the autovac launcher.
844 : */
845 : static void
846 532 : AutoVacLauncherShutdown(void)
847 : {
848 532 : ereport(DEBUG1,
849 : (errmsg_internal("autovacuum launcher shutting down")));
850 532 : AutoVacuumShmem->av_launcherpid = 0;
851 :
852 532 : proc_exit(0); /* done */
853 : }
854 :
855 : /*
856 : * Determine the time to sleep, based on the database list.
857 : *
858 : * The "canlaunch" parameter indicates whether we can start a worker right now,
859 : * for example due to the workers being all busy. If this is false, we will
860 : * cause a long sleep, which will be interrupted when a worker exits.
861 : */
862 : static void
863 2868 : launcher_determine_sleep(bool canlaunch, bool recursing, struct timeval *nap)
864 : {
865 : /*
866 : * We sleep until the next scheduled vacuum. We trust that when the
867 : * database list was built, care was taken so that no entries have times
868 : * in the past; if the first entry has too close a next_worker value, or a
869 : * time in the past, we will sleep a small nominal time.
870 : */
871 2868 : if (!canlaunch)
872 : {
873 0 : nap->tv_sec = autovacuum_naptime;
874 0 : nap->tv_usec = 0;
875 : }
876 2868 : else if (!dlist_is_empty(&DatabaseList))
877 : {
878 2760 : TimestampTz current_time = GetCurrentTimestamp();
879 : TimestampTz next_wakeup;
880 : avl_dbase *avdb;
881 : long secs;
882 : int usecs;
883 :
884 2760 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
885 :
886 2760 : next_wakeup = avdb->adl_next_worker;
887 2760 : TimestampDifference(current_time, next_wakeup, &secs, &usecs);
888 :
889 2760 : nap->tv_sec = secs;
890 2760 : nap->tv_usec = usecs;
891 : }
892 : else
893 : {
894 : /* list is empty, sleep for whole autovacuum_naptime seconds */
895 108 : nap->tv_sec = autovacuum_naptime;
896 108 : nap->tv_usec = 0;
897 : }
898 :
899 : /*
900 : * If the result is exactly zero, it means a database had an entry with
901 : * time in the past. Rebuild the list so that the databases are evenly
902 : * distributed again, and recalculate the time to sleep. This can happen
903 : * if there are more tables needing vacuum than workers, and they all take
904 : * longer to vacuum than autovacuum_naptime.
905 : *
906 : * We only recurse once. rebuild_database_list should always return times
907 : * in the future, but it seems best not to trust too much on that.
908 : */
909 2868 : if (nap->tv_sec == 0 && nap->tv_usec == 0 && !recursing)
910 : {
911 0 : rebuild_database_list(InvalidOid);
912 0 : launcher_determine_sleep(canlaunch, true, nap);
913 0 : return;
914 : }
915 :
916 : /* The smallest time we'll allow the launcher to sleep. */
917 2868 : if (nap->tv_sec <= 0 && nap->tv_usec <= MIN_AUTOVAC_SLEEPTIME * 1000)
918 : {
919 8 : nap->tv_sec = 0;
920 8 : nap->tv_usec = MIN_AUTOVAC_SLEEPTIME * 1000;
921 : }
922 :
923 : /*
924 : * If the sleep time is too large, clamp it to an arbitrary maximum (plus
925 : * any fractional seconds, for simplicity). This avoids an essentially
926 : * infinite sleep in strange cases like the system clock going backwards a
927 : * few years.
928 : */
929 2868 : if (nap->tv_sec > MAX_AUTOVAC_SLEEPTIME)
930 22 : nap->tv_sec = MAX_AUTOVAC_SLEEPTIME;
931 : }
932 :
933 : /*
934 : * Build an updated DatabaseList. It must only contain databases that appear
935 : * in pgstats, and must be sorted by next_worker from highest to lowest,
936 : * distributed regularly across the next autovacuum_naptime interval.
937 : *
938 : * Receives the Oid of the database that made this list be generated (we call
939 : * this the "new" database, because when the database was already present on
940 : * the list, we expect that this function is not called at all). The
941 : * preexisting list, if any, will be used to preserve the order of the
942 : * databases in the autovacuum_naptime period. The new database is put at the
943 : * end of the interval. The actual values are not saved, which should not be
944 : * much of a problem.
945 : */
946 : static void
947 584 : rebuild_database_list(Oid newdb)
948 : {
949 : List *dblist;
950 : ListCell *cell;
951 : MemoryContext newcxt;
952 : MemoryContext oldcxt;
953 : MemoryContext tmpcxt;
954 : HASHCTL hctl;
955 : int score;
956 : int nelems;
957 : HTAB *dbhash;
958 : dlist_iter iter;
959 :
960 584 : newcxt = AllocSetContextCreate(AutovacMemCxt,
961 : "Autovacuum database list",
962 : ALLOCSET_DEFAULT_SIZES);
963 584 : tmpcxt = AllocSetContextCreate(newcxt,
964 : "Autovacuum database list (tmp)",
965 : ALLOCSET_DEFAULT_SIZES);
966 584 : oldcxt = MemoryContextSwitchTo(tmpcxt);
967 :
968 : /*
969 : * Implementing this is not as simple as it sounds, because we need to put
970 : * the new database at the end of the list; next the databases that were
971 : * already on the list, and finally (at the tail of the list) all the
972 : * other databases that are not on the existing list.
973 : *
974 : * To do this, we build an empty hash table of scored databases. We will
975 : * start with the lowest score (zero) for the new database, then
976 : * increasing scores for the databases in the existing list, in order, and
977 : * lastly increasing scores for all databases gotten via
978 : * get_database_list() that are not already on the hash.
979 : *
980 : * Then we will put all the hash elements into an array, sort the array by
981 : * score, and finally put the array elements into the new doubly linked
982 : * list.
983 : */
984 584 : hctl.keysize = sizeof(Oid);
985 584 : hctl.entrysize = sizeof(avl_dbase);
986 584 : hctl.hcxt = tmpcxt;
987 584 : dbhash = hash_create("autovacuum db hash", 20, &hctl, /* magic number here
988 : * FIXME */
989 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
990 :
991 : /* start by inserting the new database */
992 584 : score = 0;
993 584 : if (OidIsValid(newdb))
994 : {
995 : avl_dbase *db;
996 : PgStat_StatDBEntry *entry;
997 :
998 : /* only consider this database if it has a pgstat entry */
999 8 : entry = pgstat_fetch_stat_dbentry(newdb);
1000 8 : if (entry != NULL)
1001 : {
1002 : /* we assume it isn't found because the hash was just created */
1003 8 : db = hash_search(dbhash, &newdb, HASH_ENTER, NULL);
1004 :
1005 : /* hash_search already filled in the key */
1006 8 : db->adl_score = score++;
1007 : /* next_worker is filled in later */
1008 : }
1009 : }
1010 :
1011 : /* Now insert the databases from the existing list */
1012 652 : dlist_foreach(iter, &DatabaseList)
1013 : {
1014 68 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1015 : avl_dbase *db;
1016 : bool found;
1017 : PgStat_StatDBEntry *entry;
1018 :
1019 : /*
1020 : * skip databases with no stat entries -- in particular, this gets rid
1021 : * of dropped databases
1022 : */
1023 68 : entry = pgstat_fetch_stat_dbentry(avdb->adl_datid);
1024 68 : if (entry == NULL)
1025 0 : continue;
1026 :
1027 68 : db = hash_search(dbhash, &(avdb->adl_datid), HASH_ENTER, &found);
1028 :
1029 68 : if (!found)
1030 : {
1031 : /* hash_search already filled in the key */
1032 68 : db->adl_score = score++;
1033 : /* next_worker is filled in later */
1034 : }
1035 : }
1036 :
1037 : /* finally, insert all qualifying databases not previously inserted */
1038 584 : dblist = get_database_list();
1039 2582 : foreach(cell, dblist)
1040 : {
1041 1998 : avw_dbase *avdb = lfirst(cell);
1042 : avl_dbase *db;
1043 : bool found;
1044 : PgStat_StatDBEntry *entry;
1045 :
1046 : /* only consider databases with a pgstat entry */
1047 1998 : entry = pgstat_fetch_stat_dbentry(avdb->adw_datid);
1048 1998 : if (entry == NULL)
1049 1274 : continue;
1050 :
1051 724 : db = hash_search(dbhash, &(avdb->adw_datid), HASH_ENTER, &found);
1052 : /* only update the score if the database was not already on the hash */
1053 724 : if (!found)
1054 : {
1055 : /* hash_search already filled in the key */
1056 648 : db->adl_score = score++;
1057 : /* next_worker is filled in later */
1058 : }
1059 : }
1060 584 : nelems = score;
1061 :
1062 : /* from here on, the allocated memory belongs to the new list */
1063 584 : MemoryContextSwitchTo(newcxt);
1064 584 : dlist_init(&DatabaseList);
1065 :
1066 584 : if (nelems > 0)
1067 : {
1068 : TimestampTz current_time;
1069 : int millis_increment;
1070 : avl_dbase *dbary;
1071 : avl_dbase *db;
1072 : HASH_SEQ_STATUS seq;
1073 : int i;
1074 :
1075 : /* put all the hash elements into an array */
1076 530 : dbary = palloc(nelems * sizeof(avl_dbase));
1077 :
1078 530 : i = 0;
1079 530 : hash_seq_init(&seq, dbhash);
1080 1254 : while ((db = hash_seq_search(&seq)) != NULL)
1081 724 : memcpy(&(dbary[i++]), db, sizeof(avl_dbase));
1082 :
1083 : /* sort the array */
1084 530 : qsort(dbary, nelems, sizeof(avl_dbase), db_comparator);
1085 :
1086 : /*
1087 : * Determine the time interval between databases in the schedule. If
1088 : * we see that the configured naptime would take us to sleep times
1089 : * lower than our min sleep time (which launcher_determine_sleep is
1090 : * coded not to allow), silently use a larger naptime (but don't touch
1091 : * the GUC variable).
1092 : */
1093 530 : millis_increment = 1000.0 * autovacuum_naptime / nelems;
1094 530 : if (millis_increment <= MIN_AUTOVAC_SLEEPTIME)
1095 0 : millis_increment = MIN_AUTOVAC_SLEEPTIME * 1.1;
1096 :
1097 530 : current_time = GetCurrentTimestamp();
1098 :
1099 : /*
1100 : * move the elements from the array into the dlist, setting the
1101 : * next_worker while walking the array
1102 : */
1103 1254 : for (i = 0; i < nelems; i++)
1104 : {
1105 724 : avl_dbase *db = &(dbary[i]);
1106 :
1107 724 : current_time = TimestampTzPlusMilliseconds(current_time,
1108 : millis_increment);
1109 724 : db->adl_next_worker = current_time;
1110 :
1111 : /* later elements should go closer to the head of the list */
1112 724 : dlist_push_head(&DatabaseList, &db->adl_node);
1113 : }
1114 : }
1115 :
1116 : /* all done, clean up memory */
1117 584 : if (DatabaseListCxt != NULL)
1118 46 : MemoryContextDelete(DatabaseListCxt);
1119 584 : MemoryContextDelete(tmpcxt);
1120 584 : DatabaseListCxt = newcxt;
1121 584 : MemoryContextSwitchTo(oldcxt);
1122 584 : }
1123 :
1124 : /* qsort comparator for avl_dbase, using adl_score */
1125 : static int
1126 252 : db_comparator(const void *a, const void *b)
1127 : {
1128 252 : if (((const avl_dbase *) a)->adl_score == ((const avl_dbase *) b)->adl_score)
1129 0 : return 0;
1130 : else
1131 252 : return (((const avl_dbase *) a)->adl_score < ((const avl_dbase *) b)->adl_score) ? 1 : -1;
1132 : }
1133 :
1134 : /*
1135 : * do_start_worker
1136 : *
1137 : * Bare-bones procedure for starting an autovacuum worker from the launcher.
1138 : * It determines what database to work on, sets up shared memory stuff and
1139 : * signals postmaster to start the worker. It fails gracefully if invoked when
1140 : * autovacuum_workers are already active.
1141 : *
1142 : * Return value is the OID of the database that the worker is going to process,
1143 : * or InvalidOid if no worker was actually started.
1144 : */
1145 : static Oid
1146 70 : do_start_worker(void)
1147 : {
1148 : List *dblist;
1149 : ListCell *cell;
1150 : TransactionId xidForceLimit;
1151 : MultiXactId multiForceLimit;
1152 : bool for_xid_wrap;
1153 : bool for_multi_wrap;
1154 : avw_dbase *avdb;
1155 : TimestampTz current_time;
1156 70 : bool skipit = false;
1157 70 : Oid retval = InvalidOid;
1158 : MemoryContext tmpcxt,
1159 : oldcxt;
1160 :
1161 : /* return quickly when there are no free workers */
1162 70 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1163 70 : if (dlist_is_empty(&AutoVacuumShmem->av_freeWorkers))
1164 : {
1165 0 : LWLockRelease(AutovacuumLock);
1166 0 : return InvalidOid;
1167 : }
1168 70 : LWLockRelease(AutovacuumLock);
1169 :
1170 : /*
1171 : * Create and switch to a temporary context to avoid leaking the memory
1172 : * allocated for the database list.
1173 : */
1174 70 : tmpcxt = AllocSetContextCreate(CurrentMemoryContext,
1175 : "Autovacuum start worker (tmp)",
1176 : ALLOCSET_DEFAULT_SIZES);
1177 70 : oldcxt = MemoryContextSwitchTo(tmpcxt);
1178 :
1179 : /* Get a list of databases */
1180 70 : dblist = get_database_list();
1181 :
1182 : /*
1183 : * Determine the oldest datfrozenxid/relfrozenxid that we will allow to
1184 : * pass without forcing a vacuum. (This limit can be tightened for
1185 : * particular tables, but not loosened.)
1186 : */
1187 70 : recentXid = ReadNextTransactionId();
1188 70 : xidForceLimit = recentXid - autovacuum_freeze_max_age;
1189 : /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */
1190 : /* this can cause the limit to go backwards by 3, but that's OK */
1191 70 : if (xidForceLimit < FirstNormalTransactionId)
1192 0 : xidForceLimit -= FirstNormalTransactionId;
1193 :
1194 : /* Also determine the oldest datminmxid we will consider. */
1195 70 : recentMulti = ReadNextMultiXactId();
1196 70 : multiForceLimit = recentMulti - MultiXactMemberFreezeThreshold();
1197 70 : if (multiForceLimit < FirstMultiXactId)
1198 0 : multiForceLimit -= FirstMultiXactId;
1199 :
1200 : /*
1201 : * Choose a database to connect to. We pick the database that was least
1202 : * recently auto-vacuumed, or one that needs vacuuming to prevent Xid
1203 : * wraparound-related data loss. If any db at risk of Xid wraparound is
1204 : * found, we pick the one with oldest datfrozenxid, independently of
1205 : * autovacuum times; similarly we pick the one with the oldest datminmxid
1206 : * if any is in MultiXactId wraparound. Note that those in Xid wraparound
1207 : * danger are given more priority than those in multi wraparound danger.
1208 : *
1209 : * Note that a database with no stats entry is not considered, except for
1210 : * Xid wraparound purposes. The theory is that if no one has ever
1211 : * connected to it since the stats were last initialized, it doesn't need
1212 : * vacuuming.
1213 : *
1214 : * XXX This could be improved if we had more info about whether it needs
1215 : * vacuuming before connecting to it. Perhaps look through the pgstats
1216 : * data for the database's tables? One idea is to keep track of the
1217 : * number of new and dead tuples per database in pgstats. However it
1218 : * isn't clear how to construct a metric that measures that and not cause
1219 : * starvation for less busy databases.
1220 : */
1221 70 : avdb = NULL;
1222 70 : for_xid_wrap = false;
1223 70 : for_multi_wrap = false;
1224 70 : current_time = GetCurrentTimestamp();
1225 324 : foreach(cell, dblist)
1226 : {
1227 254 : avw_dbase *tmp = lfirst(cell);
1228 : dlist_iter iter;
1229 :
1230 : /* Check to see if this one is at risk of wraparound */
1231 254 : if (TransactionIdPrecedes(tmp->adw_frozenxid, xidForceLimit))
1232 : {
1233 0 : if (avdb == NULL ||
1234 0 : TransactionIdPrecedes(tmp->adw_frozenxid,
1235 : avdb->adw_frozenxid))
1236 0 : avdb = tmp;
1237 0 : for_xid_wrap = true;
1238 224 : continue;
1239 : }
1240 254 : else if (for_xid_wrap)
1241 0 : continue; /* ignore not-at-risk DBs */
1242 254 : else if (MultiXactIdPrecedes(tmp->adw_minmulti, multiForceLimit))
1243 : {
1244 0 : if (avdb == NULL ||
1245 0 : MultiXactIdPrecedes(tmp->adw_minmulti, avdb->adw_minmulti))
1246 0 : avdb = tmp;
1247 0 : for_multi_wrap = true;
1248 0 : continue;
1249 : }
1250 254 : else if (for_multi_wrap)
1251 0 : continue; /* ignore not-at-risk DBs */
1252 :
1253 : /* Find pgstat entry if any */
1254 254 : tmp->adw_entry = pgstat_fetch_stat_dbentry(tmp->adw_datid);
1255 :
1256 : /*
1257 : * Skip a database with no pgstat entry; it means it hasn't seen any
1258 : * activity.
1259 : */
1260 254 : if (!tmp->adw_entry)
1261 208 : continue;
1262 :
1263 : /*
1264 : * Also, skip a database that appears on the database list as having
1265 : * been processed recently (less than autovacuum_naptime seconds ago).
1266 : * We do this so that we don't select a database which we just
1267 : * selected, but that pgstat hasn't gotten around to updating the last
1268 : * autovacuum time yet.
1269 : */
1270 46 : skipit = false;
1271 :
1272 84 : dlist_reverse_foreach(iter, &DatabaseList)
1273 : {
1274 70 : avl_dbase *dbp = dlist_container(avl_dbase, adl_node, iter.cur);
1275 :
1276 70 : if (dbp->adl_datid == tmp->adw_datid)
1277 : {
1278 : /*
1279 : * Skip this database if its next_worker value falls between
1280 : * the current time and the current time plus naptime.
1281 : */
1282 32 : if (!TimestampDifferenceExceeds(dbp->adl_next_worker,
1283 16 : current_time, 0) &&
1284 16 : !TimestampDifferenceExceeds(current_time,
1285 : dbp->adl_next_worker,
1286 : autovacuum_naptime * 1000))
1287 16 : skipit = true;
1288 :
1289 32 : break;
1290 : }
1291 : }
1292 46 : if (skipit)
1293 16 : continue;
1294 :
1295 : /*
1296 : * Remember the db with oldest autovac time. (If we are here, both
1297 : * tmp->entry and db->entry must be non-null.)
1298 : */
1299 30 : if (avdb == NULL ||
1300 14 : tmp->adw_entry->last_autovac_time < avdb->adw_entry->last_autovac_time)
1301 16 : avdb = tmp;
1302 : }
1303 :
1304 : /* Found a database -- process it */
1305 70 : if (avdb != NULL)
1306 : {
1307 : WorkerInfo worker;
1308 : dlist_node *wptr;
1309 :
1310 16 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1311 :
1312 : /*
1313 : * Get a worker entry from the freelist. We checked above, so there
1314 : * really should be a free slot.
1315 : */
1316 16 : wptr = dlist_pop_head_node(&AutoVacuumShmem->av_freeWorkers);
1317 :
1318 16 : worker = dlist_container(WorkerInfoData, wi_links, wptr);
1319 16 : worker->wi_dboid = avdb->adw_datid;
1320 16 : worker->wi_proc = NULL;
1321 16 : worker->wi_launchtime = GetCurrentTimestamp();
1322 :
1323 16 : AutoVacuumShmem->av_startingWorker = worker;
1324 :
1325 16 : LWLockRelease(AutovacuumLock);
1326 :
1327 16 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
1328 :
1329 16 : retval = avdb->adw_datid;
1330 : }
1331 54 : else if (skipit)
1332 : {
1333 : /*
1334 : * If we skipped all databases on the list, rebuild it, because it
1335 : * probably contains a dropped database.
1336 : */
1337 0 : rebuild_database_list(InvalidOid);
1338 : }
1339 :
1340 70 : MemoryContextSwitchTo(oldcxt);
1341 70 : MemoryContextDelete(tmpcxt);
1342 :
1343 70 : return retval;
1344 : }
1345 :
1346 : /*
1347 : * launch_worker
1348 : *
1349 : * Wrapper for starting a worker from the launcher. Besides actually starting
1350 : * it, update the database list to reflect the next time that another one will
1351 : * need to be started on the selected database. The actual database choice is
1352 : * left to do_start_worker.
1353 : *
1354 : * This routine is also expected to insert an entry into the database list if
1355 : * the selected database was previously absent from the list.
1356 : */
1357 : static void
1358 70 : launch_worker(TimestampTz now)
1359 : {
1360 : Oid dbid;
1361 : dlist_iter iter;
1362 :
1363 70 : dbid = do_start_worker();
1364 70 : if (OidIsValid(dbid))
1365 : {
1366 16 : bool found = false;
1367 :
1368 : /*
1369 : * Walk the database list and update the corresponding entry. If the
1370 : * database is not on the list, we'll recreate the list.
1371 : */
1372 40 : dlist_foreach(iter, &DatabaseList)
1373 : {
1374 32 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1375 :
1376 32 : if (avdb->adl_datid == dbid)
1377 : {
1378 8 : found = true;
1379 :
1380 : /*
1381 : * add autovacuum_naptime seconds to the current time, and use
1382 : * that as the new "next_worker" field for this database.
1383 : */
1384 8 : avdb->adl_next_worker =
1385 8 : TimestampTzPlusMilliseconds(now, autovacuum_naptime * 1000);
1386 :
1387 8 : dlist_move_head(&DatabaseList, iter.cur);
1388 8 : break;
1389 : }
1390 : }
1391 :
1392 : /*
1393 : * If the database was not present in the database list, we rebuild
1394 : * the list. It's possible that the database does not get into the
1395 : * list anyway, for example if it's a database that doesn't have a
1396 : * pgstat entry, but this is not a problem because we don't want to
1397 : * schedule workers regularly into those in any case.
1398 : */
1399 16 : if (!found)
1400 8 : rebuild_database_list(dbid);
1401 : }
1402 70 : }
1403 :
1404 : /*
1405 : * Called from postmaster to signal a failure to fork a process to become
1406 : * worker. The postmaster should kill(SIGUSR2) the launcher shortly
1407 : * after calling this function.
1408 : */
1409 : void
1410 0 : AutoVacWorkerFailed(void)
1411 : {
1412 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = true;
1413 0 : }
1414 :
1415 : /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */
1416 : static void
1417 32 : avl_sigusr2_handler(SIGNAL_ARGS)
1418 : {
1419 32 : int save_errno = errno;
1420 :
1421 32 : got_SIGUSR2 = true;
1422 32 : SetLatch(MyLatch);
1423 :
1424 32 : errno = save_errno;
1425 32 : }
1426 :
1427 :
1428 : /********************************************************************
1429 : * AUTOVACUUM WORKER CODE
1430 : ********************************************************************/
1431 :
1432 : #ifdef EXEC_BACKEND
1433 : /*
1434 : * forkexec routines for the autovacuum worker.
1435 : *
1436 : * Format up the arglist, then fork and exec.
1437 : */
1438 : static pid_t
1439 : avworker_forkexec(void)
1440 : {
1441 : char *av[10];
1442 : int ac = 0;
1443 :
1444 : av[ac++] = "postgres";
1445 : av[ac++] = "--forkavworker";
1446 : av[ac++] = NULL; /* filled in by postmaster_forkexec */
1447 : av[ac] = NULL;
1448 :
1449 : Assert(ac < lengthof(av));
1450 :
1451 : return postmaster_forkexec(ac, av);
1452 : }
1453 :
1454 : /*
1455 : * We need this set from the outside, before InitProcess is called
1456 : */
1457 : void
1458 : AutovacuumWorkerIAm(void)
1459 : {
1460 : am_autovacuum_worker = true;
1461 : }
1462 : #endif
1463 :
1464 : /*
1465 : * Main entry point for autovacuum worker process.
1466 : *
1467 : * This code is heavily based on pgarch.c, q.v.
1468 : */
1469 : int
1470 16 : StartAutoVacWorker(void)
1471 : {
1472 : pid_t worker_pid;
1473 :
1474 : #ifdef EXEC_BACKEND
1475 : switch ((worker_pid = avworker_forkexec()))
1476 : #else
1477 16 : switch ((worker_pid = fork_process()))
1478 : #endif
1479 : {
1480 0 : case -1:
1481 0 : ereport(LOG,
1482 : (errmsg("could not fork autovacuum worker process: %m")));
1483 0 : return 0;
1484 :
1485 : #ifndef EXEC_BACKEND
1486 16 : case 0:
1487 : /* in postmaster child ... */
1488 16 : InitPostmasterChild();
1489 :
1490 : /* Close the postmaster's sockets */
1491 16 : ClosePostmasterPorts(false);
1492 :
1493 16 : AutoVacWorkerMain(0, NULL);
1494 : break;
1495 : #endif
1496 16 : default:
1497 16 : return (int) worker_pid;
1498 : }
1499 :
1500 : /* shouldn't get here */
1501 : return 0;
1502 : }
1503 :
1504 : /*
1505 : * AutoVacWorkerMain
1506 : */
1507 : NON_EXEC_STATIC void
1508 16 : AutoVacWorkerMain(int argc, char *argv[])
1509 : {
1510 : sigjmp_buf local_sigjmp_buf;
1511 : Oid dbid;
1512 :
1513 16 : am_autovacuum_worker = true;
1514 :
1515 16 : MyBackendType = B_AUTOVAC_WORKER;
1516 16 : init_ps_display(NULL);
1517 :
1518 16 : SetProcessingMode(InitProcessing);
1519 :
1520 : /*
1521 : * Set up signal handlers. We operate on databases much like a regular
1522 : * backend, so we use the same signal handling. See equivalent code in
1523 : * tcop/postgres.c.
1524 : */
1525 16 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1526 :
1527 : /*
1528 : * SIGINT is used to signal canceling the current table's vacuum; SIGTERM
1529 : * means abort and exit cleanly, and SIGQUIT means abandon ship.
1530 : */
1531 16 : pqsignal(SIGINT, StatementCancelHandler);
1532 16 : pqsignal(SIGTERM, die);
1533 : /* SIGQUIT handler was already set up by InitPostmasterChild */
1534 :
1535 16 : InitializeTimeouts(); /* establishes SIGALRM handler */
1536 :
1537 16 : pqsignal(SIGPIPE, SIG_IGN);
1538 16 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1539 16 : pqsignal(SIGUSR2, SIG_IGN);
1540 16 : pqsignal(SIGFPE, FloatExceptionHandler);
1541 16 : pqsignal(SIGCHLD, SIG_DFL);
1542 :
1543 : /*
1544 : * Create a per-backend PGPROC struct in shared memory, except in the
1545 : * EXEC_BACKEND case where this was done in SubPostmasterMain. We must do
1546 : * this before we can use LWLocks (and in the EXEC_BACKEND case we already
1547 : * had to do some stuff with LWLocks).
1548 : */
1549 : #ifndef EXEC_BACKEND
1550 16 : InitProcess();
1551 : #endif
1552 :
1553 : /* Early initialization */
1554 16 : BaseInit();
1555 :
1556 : /*
1557 : * If an exception is encountered, processing resumes here.
1558 : *
1559 : * Unlike most auxiliary processes, we don't attempt to continue
1560 : * processing after an error; we just clean up and exit. The autovac
1561 : * launcher is responsible for spawning another worker later.
1562 : *
1563 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
1564 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
1565 : * signals other than SIGQUIT will be blocked until we exit. It might
1566 : * seem that this policy makes the HOLD_INTERRUPTS() call redundant, but
1567 : * it is not since InterruptPending might be set already.
1568 : */
1569 16 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1570 : {
1571 : /* since not using PG_TRY, must reset error stack by hand */
1572 0 : error_context_stack = NULL;
1573 :
1574 : /* Prevents interrupts while cleaning up */
1575 0 : HOLD_INTERRUPTS();
1576 :
1577 : /* Report the error to the server log */
1578 0 : EmitErrorReport();
1579 :
1580 : /*
1581 : * We can now go away. Note that because we called InitProcess, a
1582 : * callback was registered to do ProcKill, which will clean up
1583 : * necessary state.
1584 : */
1585 0 : proc_exit(0);
1586 : }
1587 :
1588 : /* We can now handle ereport(ERROR) */
1589 16 : PG_exception_stack = &local_sigjmp_buf;
1590 :
1591 16 : PG_SETMASK(&UnBlockSig);
1592 :
1593 : /*
1594 : * Set always-secure search path, so malicious users can't redirect user
1595 : * code (e.g. pg_index.indexprs). (That code runs in a
1596 : * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not
1597 : * take control of the entire autovacuum worker in any case.)
1598 : */
1599 16 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1600 :
1601 : /*
1602 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
1603 : * in postgresql.conf. We don't really want such a dangerous option being
1604 : * applied non-interactively.
1605 : */
1606 16 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
1607 :
1608 : /*
1609 : * Force settable timeouts off to avoid letting these settings prevent
1610 : * regular maintenance from being executed.
1611 : */
1612 16 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1613 16 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1614 16 : SetConfigOption("idle_in_transaction_session_timeout", "0",
1615 : PGC_SUSET, PGC_S_OVERRIDE);
1616 :
1617 : /*
1618 : * Force default_transaction_isolation to READ COMMITTED. We don't want
1619 : * to pay the overhead of serializable mode, nor add any risk of causing
1620 : * deadlocks or delaying other transactions.
1621 : */
1622 16 : SetConfigOption("default_transaction_isolation", "read committed",
1623 : PGC_SUSET, PGC_S_OVERRIDE);
1624 :
1625 : /*
1626 : * Force synchronous replication off to allow regular maintenance even if
1627 : * we are waiting for standbys to connect. This is important to ensure we
1628 : * aren't blocked from performing anti-wraparound tasks.
1629 : */
1630 16 : if (synchronous_commit > SYNCHRONOUS_COMMIT_LOCAL_FLUSH)
1631 16 : SetConfigOption("synchronous_commit", "local",
1632 : PGC_SUSET, PGC_S_OVERRIDE);
1633 :
1634 : /*
1635 : * Even when system is configured to use a different fetch consistency,
1636 : * for autovac we always want fresh stats.
1637 : */
1638 16 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
1639 :
1640 : /*
1641 : * Get the info about the database we're going to work on.
1642 : */
1643 16 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1644 :
1645 : /*
1646 : * beware of startingWorker being INVALID; this should normally not
1647 : * happen, but if a worker fails after forking and before this, the
1648 : * launcher might have decided to remove it from the queue and start
1649 : * again.
1650 : */
1651 16 : if (AutoVacuumShmem->av_startingWorker != NULL)
1652 : {
1653 16 : MyWorkerInfo = AutoVacuumShmem->av_startingWorker;
1654 16 : dbid = MyWorkerInfo->wi_dboid;
1655 16 : MyWorkerInfo->wi_proc = MyProc;
1656 :
1657 : /* insert into the running list */
1658 16 : dlist_push_head(&AutoVacuumShmem->av_runningWorkers,
1659 16 : &MyWorkerInfo->wi_links);
1660 :
1661 : /*
1662 : * remove from the "starting" pointer, so that the launcher can start
1663 : * a new worker if required
1664 : */
1665 16 : AutoVacuumShmem->av_startingWorker = NULL;
1666 16 : LWLockRelease(AutovacuumLock);
1667 :
1668 16 : on_shmem_exit(FreeWorkerInfo, 0);
1669 :
1670 : /* wake up the launcher */
1671 16 : if (AutoVacuumShmem->av_launcherpid != 0)
1672 16 : kill(AutoVacuumShmem->av_launcherpid, SIGUSR2);
1673 : }
1674 : else
1675 : {
1676 : /* no worker entry for me, go away */
1677 0 : elog(WARNING, "autovacuum worker started without a worker entry");
1678 0 : dbid = InvalidOid;
1679 0 : LWLockRelease(AutovacuumLock);
1680 : }
1681 :
1682 16 : if (OidIsValid(dbid))
1683 : {
1684 : char dbname[NAMEDATALEN];
1685 :
1686 : /*
1687 : * Report autovac startup to the cumulative stats system. We
1688 : * deliberately do this before InitPostgres, so that the
1689 : * last_autovac_time will get updated even if the connection attempt
1690 : * fails. This is to prevent autovac from getting "stuck" repeatedly
1691 : * selecting an unopenable database, rather than making any progress
1692 : * on stuff it can connect to.
1693 : */
1694 16 : pgstat_report_autovac(dbid);
1695 :
1696 : /*
1697 : * Connect to the selected database
1698 : *
1699 : * Note: if we have selected a just-deleted database (due to using
1700 : * stale stats info), we'll fail and exit here.
1701 : */
1702 16 : InitPostgres(NULL, dbid, NULL, InvalidOid, dbname, false);
1703 16 : SetProcessingMode(NormalProcessing);
1704 16 : set_ps_display(dbname);
1705 16 : ereport(DEBUG1,
1706 : (errmsg_internal("autovacuum: processing database \"%s\"", dbname)));
1707 :
1708 16 : if (PostAuthDelay)
1709 0 : pg_usleep(PostAuthDelay * 1000000L);
1710 :
1711 : /* And do an appropriate amount of work */
1712 16 : recentXid = ReadNextTransactionId();
1713 16 : recentMulti = ReadNextMultiXactId();
1714 16 : do_autovacuum();
1715 : }
1716 :
1717 : /*
1718 : * The launcher will be notified of my death in ProcKill, *if* we managed
1719 : * to get a worker slot at all
1720 : */
1721 :
1722 : /* All done, go away */
1723 16 : proc_exit(0);
1724 : }
1725 :
1726 : /*
1727 : * Return a WorkerInfo to the free list
1728 : */
1729 : static void
1730 16 : FreeWorkerInfo(int code, Datum arg)
1731 : {
1732 16 : if (MyWorkerInfo != NULL)
1733 : {
1734 16 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1735 :
1736 : /*
1737 : * Wake the launcher up so that he can launch a new worker immediately
1738 : * if required. We only save the launcher's PID in local memory here;
1739 : * the actual signal will be sent when the PGPROC is recycled. Note
1740 : * that we always do this, so that the launcher can rebalance the cost
1741 : * limit setting of the remaining workers.
1742 : *
1743 : * We somewhat ignore the risk that the launcher changes its PID
1744 : * between us reading it and the actual kill; we expect ProcKill to be
1745 : * called shortly after us, and we assume that PIDs are not reused too
1746 : * quickly after a process exits.
1747 : */
1748 16 : AutovacuumLauncherPid = AutoVacuumShmem->av_launcherpid;
1749 :
1750 16 : dlist_delete(&MyWorkerInfo->wi_links);
1751 16 : MyWorkerInfo->wi_dboid = InvalidOid;
1752 16 : MyWorkerInfo->wi_tableoid = InvalidOid;
1753 16 : MyWorkerInfo->wi_sharedrel = false;
1754 16 : MyWorkerInfo->wi_proc = NULL;
1755 16 : MyWorkerInfo->wi_launchtime = 0;
1756 16 : MyWorkerInfo->wi_dobalance = false;
1757 16 : MyWorkerInfo->wi_cost_delay = 0;
1758 16 : MyWorkerInfo->wi_cost_limit = 0;
1759 16 : MyWorkerInfo->wi_cost_limit_base = 0;
1760 16 : dlist_push_head(&AutoVacuumShmem->av_freeWorkers,
1761 16 : &MyWorkerInfo->wi_links);
1762 : /* not mine anymore */
1763 16 : MyWorkerInfo = NULL;
1764 :
1765 : /*
1766 : * now that we're inactive, cause a rebalancing of the surviving
1767 : * workers
1768 : */
1769 16 : AutoVacuumShmem->av_signal[AutoVacRebalance] = true;
1770 16 : LWLockRelease(AutovacuumLock);
1771 : }
1772 16 : }
1773 :
1774 : /*
1775 : * Update the cost-based delay parameters, so that multiple workers consume
1776 : * each a fraction of the total available I/O.
1777 : */
1778 : void
1779 1028 : AutoVacuumUpdateDelay(void)
1780 : {
1781 1028 : if (MyWorkerInfo)
1782 : {
1783 1028 : VacuumCostDelay = MyWorkerInfo->wi_cost_delay;
1784 1028 : VacuumCostLimit = MyWorkerInfo->wi_cost_limit;
1785 : }
1786 1028 : }
1787 :
1788 : /*
1789 : * autovac_balance_cost
1790 : * Recalculate the cost limit setting for each active worker.
1791 : *
1792 : * Caller must hold the AutovacuumLock in exclusive mode.
1793 : */
1794 : static void
1795 400 : autovac_balance_cost(void)
1796 : {
1797 : /*
1798 : * The idea here is that we ration out I/O equally. The amount of I/O
1799 : * that a worker can consume is determined by cost_limit/cost_delay, so we
1800 : * try to equalize those ratios rather than the raw limit settings.
1801 : *
1802 : * note: in cost_limit, zero also means use value from elsewhere, because
1803 : * zero is not a valid value.
1804 : */
1805 800 : int vac_cost_limit = (autovacuum_vac_cost_limit > 0 ?
1806 400 : autovacuum_vac_cost_limit : VacuumCostLimit);
1807 800 : double vac_cost_delay = (autovacuum_vac_cost_delay >= 0 ?
1808 400 : autovacuum_vac_cost_delay : VacuumCostDelay);
1809 : double cost_total;
1810 : double cost_avail;
1811 : dlist_iter iter;
1812 :
1813 : /* not set? nothing to do */
1814 400 : if (vac_cost_limit <= 0 || vac_cost_delay <= 0)
1815 54 : return;
1816 :
1817 : /* calculate the total base cost limit of participating active workers */
1818 400 : cost_total = 0.0;
1819 746 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1820 : {
1821 346 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1822 :
1823 346 : if (worker->wi_proc != NULL &&
1824 346 : worker->wi_dobalance &&
1825 346 : worker->wi_cost_limit_base > 0 && worker->wi_cost_delay > 0)
1826 346 : cost_total +=
1827 346 : (double) worker->wi_cost_limit_base / worker->wi_cost_delay;
1828 : }
1829 :
1830 : /* there are no cost limits -- nothing to do */
1831 400 : if (cost_total <= 0)
1832 54 : return;
1833 :
1834 : /*
1835 : * Adjust cost limit of each active worker to balance the total of cost
1836 : * limit to autovacuum_vacuum_cost_limit.
1837 : */
1838 346 : cost_avail = (double) vac_cost_limit / vac_cost_delay;
1839 692 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1840 : {
1841 346 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1842 :
1843 346 : if (worker->wi_proc != NULL &&
1844 346 : worker->wi_dobalance &&
1845 346 : worker->wi_cost_limit_base > 0 && worker->wi_cost_delay > 0)
1846 : {
1847 346 : int limit = (int)
1848 346 : (cost_avail * worker->wi_cost_limit_base / cost_total);
1849 :
1850 : /*
1851 : * We put a lower bound of 1 on the cost_limit, to avoid division-
1852 : * by-zero in the vacuum code. Also, in case of roundoff trouble
1853 : * in these calculations, let's be sure we don't ever set
1854 : * cost_limit to more than the base value.
1855 : */
1856 346 : worker->wi_cost_limit = Max(Min(limit,
1857 : worker->wi_cost_limit_base),
1858 : 1);
1859 : }
1860 :
1861 346 : if (worker->wi_proc != NULL)
1862 346 : elog(DEBUG2, "autovac_balance_cost(pid=%d db=%u, rel=%u, dobalance=%s cost_limit=%d, cost_limit_base=%d, cost_delay=%g)",
1863 : worker->wi_proc->pid, worker->wi_dboid, worker->wi_tableoid,
1864 : worker->wi_dobalance ? "yes" : "no",
1865 : worker->wi_cost_limit, worker->wi_cost_limit_base,
1866 : worker->wi_cost_delay);
1867 : }
1868 : }
1869 :
1870 : /*
1871 : * get_database_list
1872 : * Return a list of all databases found in pg_database.
1873 : *
1874 : * The list and associated data is allocated in the caller's memory context,
1875 : * which is in charge of ensuring that it's properly cleaned up afterwards.
1876 : *
1877 : * Note: this is the only function in which the autovacuum launcher uses a
1878 : * transaction. Although we aren't attached to any particular database and
1879 : * therefore can't access most catalogs, we do have enough infrastructure
1880 : * to do a seqscan on pg_database.
1881 : */
1882 : static List *
1883 654 : get_database_list(void)
1884 : {
1885 654 : List *dblist = NIL;
1886 : Relation rel;
1887 : TableScanDesc scan;
1888 : HeapTuple tup;
1889 : MemoryContext resultcxt;
1890 :
1891 : /* This is the context that we will allocate our output data in */
1892 654 : resultcxt = CurrentMemoryContext;
1893 :
1894 : /*
1895 : * Start a transaction so we can access pg_database, and get a snapshot.
1896 : * We don't have a use for the snapshot itself, but we're interested in
1897 : * the secondary effect that it sets RecentGlobalXmin. (This is critical
1898 : * for anything that reads heap pages, because HOT may decide to prune
1899 : * them even if the process doesn't attempt to modify any tuples.)
1900 : *
1901 : * FIXME: This comment is inaccurate / the code buggy. A snapshot that is
1902 : * not pushed/active does not reliably prevent HOT pruning (->xmin could
1903 : * e.g. be cleared when cache invalidations are processed).
1904 : */
1905 654 : StartTransactionCommand();
1906 654 : (void) GetTransactionSnapshot();
1907 :
1908 654 : rel = table_open(DatabaseRelationId, AccessShareLock);
1909 654 : scan = table_beginscan_catalog(rel, 0, NULL);
1910 :
1911 2906 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1912 : {
1913 2252 : Form_pg_database pgdatabase = (Form_pg_database) GETSTRUCT(tup);
1914 : avw_dbase *avdb;
1915 : MemoryContext oldcxt;
1916 :
1917 : /*
1918 : * Allocate our results in the caller's context, not the
1919 : * transaction's. We do this inside the loop, and restore the original
1920 : * context at the end, so that leaky things like heap_getnext() are
1921 : * not called in a potentially long-lived context.
1922 : */
1923 2252 : oldcxt = MemoryContextSwitchTo(resultcxt);
1924 :
1925 2252 : avdb = (avw_dbase *) palloc(sizeof(avw_dbase));
1926 :
1927 2252 : avdb->adw_datid = pgdatabase->oid;
1928 2252 : avdb->adw_name = pstrdup(NameStr(pgdatabase->datname));
1929 2252 : avdb->adw_frozenxid = pgdatabase->datfrozenxid;
1930 2252 : avdb->adw_minmulti = pgdatabase->datminmxid;
1931 : /* this gets set later: */
1932 2252 : avdb->adw_entry = NULL;
1933 :
1934 2252 : dblist = lappend(dblist, avdb);
1935 2252 : MemoryContextSwitchTo(oldcxt);
1936 : }
1937 :
1938 654 : table_endscan(scan);
1939 654 : table_close(rel, AccessShareLock);
1940 :
1941 654 : CommitTransactionCommand();
1942 :
1943 654 : return dblist;
1944 : }
1945 :
1946 : /*
1947 : * Process a database table-by-table
1948 : *
1949 : * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in
1950 : * order not to ignore shutdown commands for too long.
1951 : */
1952 : static void
1953 16 : do_autovacuum(void)
1954 : {
1955 : Relation classRel;
1956 : HeapTuple tuple;
1957 : TableScanDesc relScan;
1958 : Form_pg_database dbForm;
1959 16 : List *table_oids = NIL;
1960 16 : List *orphan_oids = NIL;
1961 : HASHCTL ctl;
1962 : HTAB *table_toast_map;
1963 : ListCell *volatile cell;
1964 : BufferAccessStrategy bstrategy;
1965 : ScanKeyData key;
1966 : TupleDesc pg_class_desc;
1967 : int effective_multixact_freeze_max_age;
1968 16 : bool did_vacuum = false;
1969 16 : bool found_concurrent_worker = false;
1970 : int i;
1971 :
1972 : /*
1973 : * StartTransactionCommand and CommitTransactionCommand will automatically
1974 : * switch to other contexts. We need this one to keep the list of
1975 : * relations to vacuum/analyze across transactions.
1976 : */
1977 16 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
1978 : "Autovacuum worker",
1979 : ALLOCSET_DEFAULT_SIZES);
1980 16 : MemoryContextSwitchTo(AutovacMemCxt);
1981 :
1982 : /* Start a transaction so our commands have one to play into. */
1983 16 : StartTransactionCommand();
1984 :
1985 : /*
1986 : * Compute the multixact age for which freezing is urgent. This is
1987 : * normally autovacuum_multixact_freeze_max_age, but may be less if we are
1988 : * short of multixact member space.
1989 : */
1990 16 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1991 :
1992 : /*
1993 : * Find the pg_database entry and select the default freeze ages. We use
1994 : * zero in template and nonconnectable databases, else the system-wide
1995 : * default.
1996 : */
1997 16 : tuple = SearchSysCache1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
1998 16 : if (!HeapTupleIsValid(tuple))
1999 0 : elog(ERROR, "cache lookup failed for database %u", MyDatabaseId);
2000 16 : dbForm = (Form_pg_database) GETSTRUCT(tuple);
2001 :
2002 16 : if (dbForm->datistemplate || !dbForm->datallowconn)
2003 : {
2004 4 : default_freeze_min_age = 0;
2005 4 : default_freeze_table_age = 0;
2006 4 : default_multixact_freeze_min_age = 0;
2007 4 : default_multixact_freeze_table_age = 0;
2008 : }
2009 : else
2010 : {
2011 12 : default_freeze_min_age = vacuum_freeze_min_age;
2012 12 : default_freeze_table_age = vacuum_freeze_table_age;
2013 12 : default_multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
2014 12 : default_multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
2015 : }
2016 :
2017 16 : ReleaseSysCache(tuple);
2018 :
2019 : /* StartTransactionCommand changed elsewhere */
2020 16 : MemoryContextSwitchTo(AutovacMemCxt);
2021 :
2022 16 : classRel = table_open(RelationRelationId, AccessShareLock);
2023 :
2024 : /* create a copy so we can use it after closing pg_class */
2025 16 : pg_class_desc = CreateTupleDescCopy(RelationGetDescr(classRel));
2026 :
2027 : /* create hash table for toast <-> main relid mapping */
2028 16 : ctl.keysize = sizeof(Oid);
2029 16 : ctl.entrysize = sizeof(av_relation);
2030 :
2031 16 : table_toast_map = hash_create("TOAST to main relid map",
2032 : 100,
2033 : &ctl,
2034 : HASH_ELEM | HASH_BLOBS);
2035 :
2036 : /*
2037 : * Scan pg_class to determine which tables to vacuum.
2038 : *
2039 : * We do this in two passes: on the first one we collect the list of plain
2040 : * relations and materialized views, and on the second one we collect
2041 : * TOAST tables. The reason for doing the second pass is that during it we
2042 : * want to use the main relation's pg_class.reloptions entry if the TOAST
2043 : * table does not have any, and we cannot obtain it unless we know
2044 : * beforehand what's the main table OID.
2045 : *
2046 : * We need to check TOAST tables separately because in cases with short,
2047 : * wide tables there might be proportionally much more activity in the
2048 : * TOAST table than in its parent.
2049 : */
2050 16 : relScan = table_beginscan_catalog(classRel, 0, NULL);
2051 :
2052 : /*
2053 : * On the first pass, we collect main tables to vacuum, and also the main
2054 : * table relid to TOAST relid mapping.
2055 : */
2056 9406 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2057 : {
2058 9390 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2059 : PgStat_StatTabEntry *tabentry;
2060 : AutoVacOpts *relopts;
2061 : Oid relid;
2062 : bool dovacuum;
2063 : bool doanalyze;
2064 : bool wraparound;
2065 :
2066 9390 : if (classForm->relkind != RELKIND_RELATION &&
2067 7216 : classForm->relkind != RELKIND_MATVIEW)
2068 7188 : continue;
2069 :
2070 2202 : relid = classForm->oid;
2071 :
2072 : /*
2073 : * Check if it is a temp table (presumably, of some other backend's).
2074 : * We cannot safely process other backends' temp tables.
2075 : */
2076 2202 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2077 : {
2078 : /*
2079 : * We just ignore it if the owning backend is still active and
2080 : * using the temporary schema. Also, for safety, ignore it if the
2081 : * namespace doesn't exist or isn't a temp namespace after all.
2082 : */
2083 0 : if (checkTempNamespaceStatus(classForm->relnamespace) == TEMP_NAMESPACE_IDLE)
2084 : {
2085 : /*
2086 : * The table seems to be orphaned -- although it might be that
2087 : * the owning backend has already deleted it and exited; our
2088 : * pg_class scan snapshot is not necessarily up-to-date
2089 : * anymore, so we could be looking at a committed-dead entry.
2090 : * Remember it so we can try to delete it later.
2091 : */
2092 0 : orphan_oids = lappend_oid(orphan_oids, relid);
2093 : }
2094 0 : continue;
2095 : }
2096 :
2097 : /* Fetch reloptions and the pgstat entry for this table */
2098 2202 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2099 2202 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2100 : relid);
2101 :
2102 : /* Check if it needs vacuum or analyze */
2103 2202 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2104 : effective_multixact_freeze_max_age,
2105 : &dovacuum, &doanalyze, &wraparound);
2106 :
2107 : /* Relations that need work are added to table_oids */
2108 2202 : if (dovacuum || doanalyze)
2109 342 : table_oids = lappend_oid(table_oids, relid);
2110 :
2111 : /*
2112 : * Remember TOAST associations for the second pass. Note: we must do
2113 : * this whether or not the table is going to be vacuumed, because we
2114 : * don't automatically vacuum toast tables along the parent table.
2115 : */
2116 2202 : if (OidIsValid(classForm->reltoastrelid))
2117 : {
2118 : av_relation *hentry;
2119 : bool found;
2120 :
2121 2256 : hentry = hash_search(table_toast_map,
2122 1128 : &classForm->reltoastrelid,
2123 : HASH_ENTER, &found);
2124 :
2125 1128 : if (!found)
2126 : {
2127 : /* hash_search already filled in the key */
2128 1128 : hentry->ar_relid = relid;
2129 1128 : hentry->ar_hasrelopts = false;
2130 1128 : if (relopts != NULL)
2131 : {
2132 36 : hentry->ar_hasrelopts = true;
2133 36 : memcpy(&hentry->ar_reloptions, relopts,
2134 : sizeof(AutoVacOpts));
2135 : }
2136 : }
2137 : }
2138 : }
2139 :
2140 16 : table_endscan(relScan);
2141 :
2142 : /* second pass: check TOAST tables */
2143 16 : ScanKeyInit(&key,
2144 : Anum_pg_class_relkind,
2145 : BTEqualStrategyNumber, F_CHAREQ,
2146 : CharGetDatum(RELKIND_TOASTVALUE));
2147 :
2148 16 : relScan = table_beginscan_catalog(classRel, 1, &key);
2149 1144 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2150 : {
2151 1128 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2152 : PgStat_StatTabEntry *tabentry;
2153 : Oid relid;
2154 1128 : AutoVacOpts *relopts = NULL;
2155 : bool dovacuum;
2156 : bool doanalyze;
2157 : bool wraparound;
2158 :
2159 : /*
2160 : * We cannot safely process other backends' temp tables, so skip 'em.
2161 : */
2162 1128 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2163 0 : continue;
2164 :
2165 1128 : relid = classForm->oid;
2166 :
2167 : /*
2168 : * fetch reloptions -- if this toast table does not have them, try the
2169 : * main rel
2170 : */
2171 1128 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2172 1128 : if (relopts == NULL)
2173 : {
2174 : av_relation *hentry;
2175 : bool found;
2176 :
2177 1128 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2178 1128 : if (found && hentry->ar_hasrelopts)
2179 36 : relopts = &hentry->ar_reloptions;
2180 : }
2181 :
2182 : /* Fetch the pgstat entry for this table */
2183 1128 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2184 : relid);
2185 :
2186 1128 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2187 : effective_multixact_freeze_max_age,
2188 : &dovacuum, &doanalyze, &wraparound);
2189 :
2190 : /* ignore analyze for toast tables */
2191 1128 : if (dovacuum)
2192 4 : table_oids = lappend_oid(table_oids, relid);
2193 : }
2194 :
2195 16 : table_endscan(relScan);
2196 16 : table_close(classRel, AccessShareLock);
2197 :
2198 : /*
2199 : * Recheck orphan temporary tables, and if they still seem orphaned, drop
2200 : * them. We'll eat a transaction per dropped table, which might seem
2201 : * excessive, but we should only need to do anything as a result of a
2202 : * previous backend crash, so this should not happen often enough to
2203 : * justify "optimizing". Using separate transactions ensures that we
2204 : * don't bloat the lock table if there are many temp tables to be dropped,
2205 : * and it ensures that we don't lose work if a deletion attempt fails.
2206 : */
2207 16 : foreach(cell, orphan_oids)
2208 : {
2209 0 : Oid relid = lfirst_oid(cell);
2210 : Form_pg_class classForm;
2211 : ObjectAddress object;
2212 :
2213 : /*
2214 : * Check for user-requested abort.
2215 : */
2216 0 : CHECK_FOR_INTERRUPTS();
2217 :
2218 : /*
2219 : * Try to lock the table. If we can't get the lock immediately,
2220 : * somebody else is using (or dropping) the table, so it's not our
2221 : * concern anymore. Having the lock prevents race conditions below.
2222 : */
2223 0 : if (!ConditionalLockRelationOid(relid, AccessExclusiveLock))
2224 0 : continue;
2225 :
2226 : /*
2227 : * Re-fetch the pg_class tuple and re-check whether it still seems to
2228 : * be an orphaned temp table. If it's not there or no longer the same
2229 : * relation, ignore it.
2230 : */
2231 0 : tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2232 0 : if (!HeapTupleIsValid(tuple))
2233 : {
2234 : /* be sure to drop useless lock so we don't bloat lock table */
2235 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2236 0 : continue;
2237 : }
2238 0 : classForm = (Form_pg_class) GETSTRUCT(tuple);
2239 :
2240 : /*
2241 : * Make all the same tests made in the loop above. In event of OID
2242 : * counter wraparound, the pg_class entry we have now might be
2243 : * completely unrelated to the one we saw before.
2244 : */
2245 0 : if (!((classForm->relkind == RELKIND_RELATION ||
2246 0 : classForm->relkind == RELKIND_MATVIEW) &&
2247 0 : classForm->relpersistence == RELPERSISTENCE_TEMP))
2248 : {
2249 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2250 0 : continue;
2251 : }
2252 :
2253 0 : if (checkTempNamespaceStatus(classForm->relnamespace) != TEMP_NAMESPACE_IDLE)
2254 : {
2255 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2256 0 : continue;
2257 : }
2258 :
2259 : /* OK, let's delete it */
2260 0 : ereport(LOG,
2261 : (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"",
2262 : get_database_name(MyDatabaseId),
2263 : get_namespace_name(classForm->relnamespace),
2264 : NameStr(classForm->relname))));
2265 :
2266 0 : object.classId = RelationRelationId;
2267 0 : object.objectId = relid;
2268 0 : object.objectSubId = 0;
2269 0 : performDeletion(&object, DROP_CASCADE,
2270 : PERFORM_DELETION_INTERNAL |
2271 : PERFORM_DELETION_QUIETLY |
2272 : PERFORM_DELETION_SKIP_EXTENSIONS);
2273 :
2274 : /*
2275 : * To commit the deletion, end current transaction and start a new
2276 : * one. Note this also releases the lock we took.
2277 : */
2278 0 : CommitTransactionCommand();
2279 0 : StartTransactionCommand();
2280 :
2281 : /* StartTransactionCommand changed current memory context */
2282 0 : MemoryContextSwitchTo(AutovacMemCxt);
2283 : }
2284 :
2285 : /*
2286 : * Create a buffer access strategy object for VACUUM to use. We want to
2287 : * use the same one across all the vacuum operations we perform, since the
2288 : * point is for VACUUM not to blow out the shared cache.
2289 : */
2290 16 : bstrategy = GetAccessStrategy(BAS_VACUUM);
2291 :
2292 : /*
2293 : * create a memory context to act as fake PortalContext, so that the
2294 : * contexts created in the vacuum code are cleaned up for each table.
2295 : */
2296 16 : PortalContext = AllocSetContextCreate(AutovacMemCxt,
2297 : "Autovacuum Portal",
2298 : ALLOCSET_DEFAULT_SIZES);
2299 :
2300 : /*
2301 : * Perform operations on collected tables.
2302 : */
2303 362 : foreach(cell, table_oids)
2304 : {
2305 346 : Oid relid = lfirst_oid(cell);
2306 : HeapTuple classTup;
2307 : autovac_table *tab;
2308 : bool isshared;
2309 : bool skipit;
2310 : double stdVacuumCostDelay;
2311 : int stdVacuumCostLimit;
2312 : dlist_iter iter;
2313 :
2314 346 : CHECK_FOR_INTERRUPTS();
2315 :
2316 : /*
2317 : * Check for config changes before processing each collected table.
2318 : */
2319 346 : if (ConfigReloadPending)
2320 : {
2321 0 : ConfigReloadPending = false;
2322 0 : ProcessConfigFile(PGC_SIGHUP);
2323 :
2324 : /*
2325 : * You might be tempted to bail out if we see autovacuum is now
2326 : * disabled. Must resist that temptation -- this might be a
2327 : * for-wraparound emergency worker, in which case that would be
2328 : * entirely inappropriate.
2329 : */
2330 : }
2331 :
2332 : /*
2333 : * Find out whether the table is shared or not. (It's slightly
2334 : * annoying to fetch the syscache entry just for this, but in typical
2335 : * cases it adds little cost because table_recheck_autovac would
2336 : * refetch the entry anyway. We could buy that back by copying the
2337 : * tuple here and passing it to table_recheck_autovac, but that
2338 : * increases the odds of that function working with stale data.)
2339 : */
2340 346 : classTup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
2341 346 : if (!HeapTupleIsValid(classTup))
2342 0 : continue; /* somebody deleted the rel, forget it */
2343 346 : isshared = ((Form_pg_class) GETSTRUCT(classTup))->relisshared;
2344 346 : ReleaseSysCache(classTup);
2345 :
2346 : /*
2347 : * Hold schedule lock from here until we've claimed the table. We
2348 : * also need the AutovacuumLock to walk the worker array, but that one
2349 : * can just be a shared lock.
2350 : */
2351 346 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2352 346 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2353 :
2354 : /*
2355 : * Check whether the table is being vacuumed concurrently by another
2356 : * worker.
2357 : */
2358 346 : skipit = false;
2359 692 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
2360 : {
2361 346 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
2362 :
2363 : /* ignore myself */
2364 346 : if (worker == MyWorkerInfo)
2365 346 : continue;
2366 :
2367 : /* ignore workers in other databases (unless table is shared) */
2368 0 : if (!worker->wi_sharedrel && worker->wi_dboid != MyDatabaseId)
2369 0 : continue;
2370 :
2371 0 : if (worker->wi_tableoid == relid)
2372 : {
2373 0 : skipit = true;
2374 0 : found_concurrent_worker = true;
2375 0 : break;
2376 : }
2377 : }
2378 346 : LWLockRelease(AutovacuumLock);
2379 346 : if (skipit)
2380 : {
2381 0 : LWLockRelease(AutovacuumScheduleLock);
2382 0 : continue;
2383 : }
2384 :
2385 : /*
2386 : * Store the table's OID in shared memory before releasing the
2387 : * schedule lock, so that other workers don't try to vacuum it
2388 : * concurrently. (We claim it here so as not to hold
2389 : * AutovacuumScheduleLock while rechecking the stats.)
2390 : */
2391 346 : MyWorkerInfo->wi_tableoid = relid;
2392 346 : MyWorkerInfo->wi_sharedrel = isshared;
2393 346 : LWLockRelease(AutovacuumScheduleLock);
2394 :
2395 : /*
2396 : * Check whether pgstat data still says we need to vacuum this table.
2397 : * It could have changed if something else processed the table while
2398 : * we weren't looking. This doesn't entirely close the race condition,
2399 : * but it is very small.
2400 : */
2401 346 : MemoryContextSwitchTo(AutovacMemCxt);
2402 346 : tab = table_recheck_autovac(relid, table_toast_map, pg_class_desc,
2403 : effective_multixact_freeze_max_age);
2404 346 : if (tab == NULL)
2405 : {
2406 : /* someone else vacuumed the table, or it went away */
2407 0 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2408 0 : MyWorkerInfo->wi_tableoid = InvalidOid;
2409 0 : MyWorkerInfo->wi_sharedrel = false;
2410 0 : LWLockRelease(AutovacuumScheduleLock);
2411 0 : continue;
2412 : }
2413 :
2414 : /*
2415 : * Remember the prevailing values of the vacuum cost GUCs. We have to
2416 : * restore these at the bottom of the loop, else we'll compute wrong
2417 : * values in the next iteration of autovac_balance_cost().
2418 : */
2419 346 : stdVacuumCostDelay = VacuumCostDelay;
2420 346 : stdVacuumCostLimit = VacuumCostLimit;
2421 :
2422 : /* Must hold AutovacuumLock while mucking with cost balance info */
2423 346 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2424 :
2425 : /* advertise my cost delay parameters for the balancing algorithm */
2426 346 : MyWorkerInfo->wi_dobalance = tab->at_dobalance;
2427 346 : MyWorkerInfo->wi_cost_delay = tab->at_vacuum_cost_delay;
2428 346 : MyWorkerInfo->wi_cost_limit = tab->at_vacuum_cost_limit;
2429 346 : MyWorkerInfo->wi_cost_limit_base = tab->at_vacuum_cost_limit;
2430 :
2431 : /* do a balance */
2432 346 : autovac_balance_cost();
2433 :
2434 : /* set the active cost parameters from the result of that */
2435 346 : AutoVacuumUpdateDelay();
2436 :
2437 : /* done */
2438 346 : LWLockRelease(AutovacuumLock);
2439 :
2440 : /* clean up memory before each iteration */
2441 346 : MemoryContextResetAndDeleteChildren(PortalContext);
2442 :
2443 : /*
2444 : * Save the relation name for a possible error message, to avoid a
2445 : * catalog lookup in case of an error. If any of these return NULL,
2446 : * then the relation has been dropped since last we checked; skip it.
2447 : * Note: they must live in a long-lived memory context because we call
2448 : * vacuum and analyze in different transactions.
2449 : */
2450 :
2451 346 : tab->at_relname = get_rel_name(tab->at_relid);
2452 346 : tab->at_nspname = get_namespace_name(get_rel_namespace(tab->at_relid));
2453 346 : tab->at_datname = get_database_name(MyDatabaseId);
2454 346 : if (!tab->at_relname || !tab->at_nspname || !tab->at_datname)
2455 0 : goto deleted;
2456 :
2457 : /*
2458 : * We will abort vacuuming the current table if something errors out,
2459 : * and continue with the next one in schedule; in particular, this
2460 : * happens if we are interrupted with SIGINT.
2461 : */
2462 346 : PG_TRY();
2463 : {
2464 : /* Use PortalContext for any per-table allocations */
2465 346 : MemoryContextSwitchTo(PortalContext);
2466 :
2467 : /* have at it */
2468 346 : autovacuum_do_vac_analyze(tab, bstrategy);
2469 :
2470 : /*
2471 : * Clear a possible query-cancel signal, to avoid a late reaction
2472 : * to an automatically-sent signal because of vacuuming the
2473 : * current table (we're done with it, so it would make no sense to
2474 : * cancel at this point.)
2475 : */
2476 346 : QueryCancelPending = false;
2477 : }
2478 0 : PG_CATCH();
2479 : {
2480 : /*
2481 : * Abort the transaction, start a new one, and proceed with the
2482 : * next table in our list.
2483 : */
2484 0 : HOLD_INTERRUPTS();
2485 0 : if (tab->at_params.options & VACOPT_VACUUM)
2486 0 : errcontext("automatic vacuum of table \"%s.%s.%s\"",
2487 : tab->at_datname, tab->at_nspname, tab->at_relname);
2488 : else
2489 0 : errcontext("automatic analyze of table \"%s.%s.%s\"",
2490 : tab->at_datname, tab->at_nspname, tab->at_relname);
2491 0 : EmitErrorReport();
2492 :
2493 : /* this resets ProcGlobal->statusFlags[i] too */
2494 0 : AbortOutOfAnyTransaction();
2495 0 : FlushErrorState();
2496 0 : MemoryContextResetAndDeleteChildren(PortalContext);
2497 :
2498 : /* restart our transaction for the following operations */
2499 0 : StartTransactionCommand();
2500 0 : RESUME_INTERRUPTS();
2501 : }
2502 346 : PG_END_TRY();
2503 :
2504 : /* Make sure we're back in AutovacMemCxt */
2505 346 : MemoryContextSwitchTo(AutovacMemCxt);
2506 :
2507 346 : did_vacuum = true;
2508 :
2509 : /* ProcGlobal->statusFlags[i] are reset at the next end of xact */
2510 :
2511 : /* be tidy */
2512 346 : deleted:
2513 346 : if (tab->at_datname != NULL)
2514 346 : pfree(tab->at_datname);
2515 346 : if (tab->at_nspname != NULL)
2516 346 : pfree(tab->at_nspname);
2517 346 : if (tab->at_relname != NULL)
2518 346 : pfree(tab->at_relname);
2519 346 : pfree(tab);
2520 :
2521 : /*
2522 : * Remove my info from shared memory. We could, but intentionally
2523 : * don't, clear wi_cost_limit and friends --- this is on the
2524 : * assumption that we probably have more to do with similar cost
2525 : * settings, so we don't want to give up our share of I/O for a very
2526 : * short interval and thereby thrash the global balance.
2527 : */
2528 346 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2529 346 : MyWorkerInfo->wi_tableoid = InvalidOid;
2530 346 : MyWorkerInfo->wi_sharedrel = false;
2531 346 : LWLockRelease(AutovacuumScheduleLock);
2532 :
2533 : /* restore vacuum cost GUCs for the next iteration */
2534 346 : VacuumCostDelay = stdVacuumCostDelay;
2535 346 : VacuumCostLimit = stdVacuumCostLimit;
2536 : }
2537 :
2538 : /*
2539 : * Perform additional work items, as requested by backends.
2540 : */
2541 16 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2542 4112 : for (i = 0; i < NUM_WORKITEMS; i++)
2543 : {
2544 4096 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
2545 :
2546 4096 : if (!workitem->avw_used)
2547 4090 : continue;
2548 6 : if (workitem->avw_active)
2549 0 : continue;
2550 6 : if (workitem->avw_database != MyDatabaseId)
2551 0 : continue;
2552 :
2553 : /* claim this one, and release lock while performing it */
2554 6 : workitem->avw_active = true;
2555 6 : LWLockRelease(AutovacuumLock);
2556 :
2557 6 : perform_work_item(workitem);
2558 :
2559 : /*
2560 : * Check for config changes before acquiring lock for further jobs.
2561 : */
2562 6 : CHECK_FOR_INTERRUPTS();
2563 6 : if (ConfigReloadPending)
2564 : {
2565 0 : ConfigReloadPending = false;
2566 0 : ProcessConfigFile(PGC_SIGHUP);
2567 : }
2568 :
2569 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2570 :
2571 : /* and mark it done */
2572 6 : workitem->avw_active = false;
2573 6 : workitem->avw_used = false;
2574 : }
2575 16 : LWLockRelease(AutovacuumLock);
2576 :
2577 : /*
2578 : * We leak table_toast_map here (among other things), but since we're
2579 : * going away soon, it's not a problem.
2580 : */
2581 :
2582 : /*
2583 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We
2584 : * only need to do this once, not after each table.
2585 : *
2586 : * Even if we didn't vacuum anything, it may still be important to do
2587 : * this, because one indirect effect of vac_update_datfrozenxid() is to
2588 : * update ShmemVariableCache->xidVacLimit. That might need to be done
2589 : * even if we haven't vacuumed anything, because relations with older
2590 : * relfrozenxid values or other databases with older datfrozenxid values
2591 : * might have been dropped, allowing xidVacLimit to advance.
2592 : *
2593 : * However, it's also important not to do this blindly in all cases,
2594 : * because when autovacuum=off this will restart the autovacuum launcher.
2595 : * If we're not careful, an infinite loop can result, where workers find
2596 : * no work to do and restart the launcher, which starts another worker in
2597 : * the same database that finds no work to do. To prevent that, we skip
2598 : * this if (1) we found no work to do and (2) we skipped at least one
2599 : * table due to concurrent autovacuum activity. In that case, the other
2600 : * worker has already done it, or will do so when it finishes.
2601 : */
2602 16 : if (did_vacuum || !found_concurrent_worker)
2603 16 : vac_update_datfrozenxid();
2604 :
2605 : /* Finally close out the last transaction. */
2606 16 : CommitTransactionCommand();
2607 16 : }
2608 :
2609 : /*
2610 : * Execute a previously registered work item.
2611 : */
2612 : static void
2613 6 : perform_work_item(AutoVacuumWorkItem *workitem)
2614 : {
2615 6 : char *cur_datname = NULL;
2616 6 : char *cur_nspname = NULL;
2617 6 : char *cur_relname = NULL;
2618 :
2619 : /*
2620 : * Note we do not store table info in MyWorkerInfo, since this is not
2621 : * vacuuming proper.
2622 : */
2623 :
2624 : /*
2625 : * Save the relation name for a possible error message, to avoid a catalog
2626 : * lookup in case of an error. If any of these return NULL, then the
2627 : * relation has been dropped since last we checked; skip it.
2628 : */
2629 : Assert(CurrentMemoryContext == AutovacMemCxt);
2630 :
2631 6 : cur_relname = get_rel_name(workitem->avw_relation);
2632 6 : cur_nspname = get_namespace_name(get_rel_namespace(workitem->avw_relation));
2633 6 : cur_datname = get_database_name(MyDatabaseId);
2634 6 : if (!cur_relname || !cur_nspname || !cur_datname)
2635 0 : goto deleted2;
2636 :
2637 6 : autovac_report_workitem(workitem, cur_nspname, cur_relname);
2638 :
2639 : /* clean up memory before each work item */
2640 6 : MemoryContextResetAndDeleteChildren(PortalContext);
2641 :
2642 : /*
2643 : * We will abort the current work item if something errors out, and
2644 : * continue with the next one; in particular, this happens if we are
2645 : * interrupted with SIGINT. Note that this means that the work item list
2646 : * can be lossy.
2647 : */
2648 6 : PG_TRY();
2649 : {
2650 : /* Use PortalContext for any per-work-item allocations */
2651 6 : MemoryContextSwitchTo(PortalContext);
2652 :
2653 : /* have at it */
2654 6 : switch (workitem->avw_type)
2655 : {
2656 6 : case AVW_BRINSummarizeRange:
2657 6 : DirectFunctionCall2(brin_summarize_range,
2658 : ObjectIdGetDatum(workitem->avw_relation),
2659 : Int64GetDatum((int64) workitem->avw_blockNumber));
2660 6 : break;
2661 0 : default:
2662 0 : elog(WARNING, "unrecognized work item found: type %d",
2663 : workitem->avw_type);
2664 0 : break;
2665 : }
2666 :
2667 : /*
2668 : * Clear a possible query-cancel signal, to avoid a late reaction to
2669 : * an automatically-sent signal because of vacuuming the current table
2670 : * (we're done with it, so it would make no sense to cancel at this
2671 : * point.)
2672 : */
2673 6 : QueryCancelPending = false;
2674 : }
2675 0 : PG_CATCH();
2676 : {
2677 : /*
2678 : * Abort the transaction, start a new one, and proceed with the next
2679 : * table in our list.
2680 : */
2681 0 : HOLD_INTERRUPTS();
2682 0 : errcontext("processing work entry for relation \"%s.%s.%s\"",
2683 : cur_datname, cur_nspname, cur_relname);
2684 0 : EmitErrorReport();
2685 :
2686 : /* this resets ProcGlobal->statusFlags[i] too */
2687 0 : AbortOutOfAnyTransaction();
2688 0 : FlushErrorState();
2689 0 : MemoryContextResetAndDeleteChildren(PortalContext);
2690 :
2691 : /* restart our transaction for the following operations */
2692 0 : StartTransactionCommand();
2693 0 : RESUME_INTERRUPTS();
2694 : }
2695 6 : PG_END_TRY();
2696 :
2697 : /* Make sure we're back in AutovacMemCxt */
2698 6 : MemoryContextSwitchTo(AutovacMemCxt);
2699 :
2700 : /* We intentionally do not set did_vacuum here */
2701 :
2702 : /* be tidy */
2703 6 : deleted2:
2704 6 : if (cur_datname)
2705 6 : pfree(cur_datname);
2706 6 : if (cur_nspname)
2707 6 : pfree(cur_nspname);
2708 6 : if (cur_relname)
2709 6 : pfree(cur_relname);
2710 6 : }
2711 :
2712 : /*
2713 : * extract_autovac_opts
2714 : *
2715 : * Given a relation's pg_class tuple, return the AutoVacOpts portion of
2716 : * reloptions, if set; otherwise, return NULL.
2717 : *
2718 : * Note: callers do not have a relation lock on the table at this point,
2719 : * so the table could have been dropped, and its catalog rows gone, after
2720 : * we acquired the pg_class row. If pg_class had a TOAST table, this would
2721 : * be a risk; fortunately, it doesn't.
2722 : */
2723 : static AutoVacOpts *
2724 3676 : extract_autovac_opts(HeapTuple tup, TupleDesc pg_class_desc)
2725 : {
2726 : bytea *relopts;
2727 : AutoVacOpts *av;
2728 :
2729 : Assert(((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_RELATION ||
2730 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_MATVIEW ||
2731 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_TOASTVALUE);
2732 :
2733 3676 : relopts = extractRelOptions(tup, pg_class_desc, NULL);
2734 3676 : if (relopts == NULL)
2735 3596 : return NULL;
2736 :
2737 80 : av = palloc(sizeof(AutoVacOpts));
2738 80 : memcpy(av, &(((StdRdOptions *) relopts)->autovacuum), sizeof(AutoVacOpts));
2739 80 : pfree(relopts);
2740 :
2741 80 : return av;
2742 : }
2743 :
2744 :
2745 : /*
2746 : * table_recheck_autovac
2747 : *
2748 : * Recheck whether a table still needs vacuum or analyze. Return value is a
2749 : * valid autovac_table pointer if it does, NULL otherwise.
2750 : *
2751 : * Note that the returned autovac_table does not have the name fields set.
2752 : */
2753 : static autovac_table *
2754 346 : table_recheck_autovac(Oid relid, HTAB *table_toast_map,
2755 : TupleDesc pg_class_desc,
2756 : int effective_multixact_freeze_max_age)
2757 : {
2758 : Form_pg_class classForm;
2759 : HeapTuple classTup;
2760 : bool dovacuum;
2761 : bool doanalyze;
2762 346 : autovac_table *tab = NULL;
2763 : bool wraparound;
2764 : AutoVacOpts *avopts;
2765 :
2766 : /* fetch the relation's relcache entry */
2767 346 : classTup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2768 346 : if (!HeapTupleIsValid(classTup))
2769 0 : return NULL;
2770 346 : classForm = (Form_pg_class) GETSTRUCT(classTup);
2771 :
2772 : /*
2773 : * Get the applicable reloptions. If it is a TOAST table, try to get the
2774 : * main table reloptions if the toast table itself doesn't have.
2775 : */
2776 346 : avopts = extract_autovac_opts(classTup, pg_class_desc);
2777 346 : if (classForm->relkind == RELKIND_TOASTVALUE &&
2778 4 : avopts == NULL && table_toast_map != NULL)
2779 : {
2780 : av_relation *hentry;
2781 : bool found;
2782 :
2783 4 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2784 4 : if (found && hentry->ar_hasrelopts)
2785 0 : avopts = &hentry->ar_reloptions;
2786 : }
2787 :
2788 346 : recheck_relation_needs_vacanalyze(relid, avopts, classForm,
2789 : effective_multixact_freeze_max_age,
2790 : &dovacuum, &doanalyze, &wraparound);
2791 :
2792 : /* OK, it needs something done */
2793 346 : if (doanalyze || dovacuum)
2794 : {
2795 : int freeze_min_age;
2796 : int freeze_table_age;
2797 : int multixact_freeze_min_age;
2798 : int multixact_freeze_table_age;
2799 : int vac_cost_limit;
2800 : double vac_cost_delay;
2801 : int log_min_duration;
2802 :
2803 : /*
2804 : * Calculate the vacuum cost parameters and the freeze ages. If there
2805 : * are options set in pg_class.reloptions, use them; in the case of a
2806 : * toast table, try the main table too. Otherwise use the GUC
2807 : * defaults, autovacuum's own first and plain vacuum second.
2808 : */
2809 :
2810 : /* -1 in autovac setting means use plain vacuum_cost_delay */
2811 10 : vac_cost_delay = (avopts && avopts->vacuum_cost_delay >= 0)
2812 : ? avopts->vacuum_cost_delay
2813 692 : : (autovacuum_vac_cost_delay >= 0)
2814 : ? autovacuum_vac_cost_delay
2815 346 : : VacuumCostDelay;
2816 :
2817 : /* 0 or -1 in autovac setting means use plain vacuum_cost_limit */
2818 10 : vac_cost_limit = (avopts && avopts->vacuum_cost_limit > 0)
2819 : ? avopts->vacuum_cost_limit
2820 692 : : (autovacuum_vac_cost_limit > 0)
2821 : ? autovacuum_vac_cost_limit
2822 346 : : VacuumCostLimit;
2823 :
2824 : /* -1 in autovac setting means use log_autovacuum_min_duration */
2825 10 : log_min_duration = (avopts && avopts->log_min_duration >= 0)
2826 : ? avopts->log_min_duration
2827 356 : : Log_autovacuum_min_duration;
2828 :
2829 : /* these do not have autovacuum-specific settings */
2830 10 : freeze_min_age = (avopts && avopts->freeze_min_age >= 0)
2831 : ? avopts->freeze_min_age
2832 356 : : default_freeze_min_age;
2833 :
2834 10 : freeze_table_age = (avopts && avopts->freeze_table_age >= 0)
2835 : ? avopts->freeze_table_age
2836 356 : : default_freeze_table_age;
2837 :
2838 356 : multixact_freeze_min_age = (avopts &&
2839 10 : avopts->multixact_freeze_min_age >= 0)
2840 : ? avopts->multixact_freeze_min_age
2841 356 : : default_multixact_freeze_min_age;
2842 :
2843 356 : multixact_freeze_table_age = (avopts &&
2844 10 : avopts->multixact_freeze_table_age >= 0)
2845 : ? avopts->multixact_freeze_table_age
2846 356 : : default_multixact_freeze_table_age;
2847 :
2848 346 : tab = palloc(sizeof(autovac_table));
2849 346 : tab->at_relid = relid;
2850 346 : tab->at_sharedrel = classForm->relisshared;
2851 :
2852 : /* Note that this skips toast relations */
2853 1038 : tab->at_params.options = (dovacuum ? VACOPT_VACUUM : 0) |
2854 346 : (doanalyze ? VACOPT_ANALYZE : 0) |
2855 346 : (!wraparound ? VACOPT_SKIP_LOCKED : 0);
2856 :
2857 : /*
2858 : * index_cleanup and truncate are unspecified at first in autovacuum.
2859 : * They will be filled in with usable values using their reloptions
2860 : * (or reloption defaults) later.
2861 : */
2862 346 : tab->at_params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
2863 346 : tab->at_params.truncate = VACOPTVALUE_UNSPECIFIED;
2864 : /* As of now, we don't support parallel vacuum for autovacuum */
2865 346 : tab->at_params.nworkers = -1;
2866 346 : tab->at_params.freeze_min_age = freeze_min_age;
2867 346 : tab->at_params.freeze_table_age = freeze_table_age;
2868 346 : tab->at_params.multixact_freeze_min_age = multixact_freeze_min_age;
2869 346 : tab->at_params.multixact_freeze_table_age = multixact_freeze_table_age;
2870 346 : tab->at_params.is_wraparound = wraparound;
2871 346 : tab->at_params.log_min_duration = log_min_duration;
2872 346 : tab->at_vacuum_cost_limit = vac_cost_limit;
2873 346 : tab->at_vacuum_cost_delay = vac_cost_delay;
2874 346 : tab->at_relname = NULL;
2875 346 : tab->at_nspname = NULL;
2876 346 : tab->at_datname = NULL;
2877 :
2878 : /*
2879 : * If any of the cost delay parameters has been set individually for
2880 : * this table, disable the balancing algorithm.
2881 : */
2882 346 : tab->at_dobalance =
2883 356 : !(avopts && (avopts->vacuum_cost_limit > 0 ||
2884 10 : avopts->vacuum_cost_delay > 0));
2885 : }
2886 :
2887 346 : heap_freetuple(classTup);
2888 346 : return tab;
2889 : }
2890 :
2891 : /*
2892 : * recheck_relation_needs_vacanalyze
2893 : *
2894 : * Subroutine for table_recheck_autovac.
2895 : *
2896 : * Fetch the pgstat of a relation and recheck whether a relation
2897 : * needs to be vacuumed or analyzed.
2898 : */
2899 : static void
2900 346 : recheck_relation_needs_vacanalyze(Oid relid,
2901 : AutoVacOpts *avopts,
2902 : Form_pg_class classForm,
2903 : int effective_multixact_freeze_max_age,
2904 : bool *dovacuum,
2905 : bool *doanalyze,
2906 : bool *wraparound)
2907 : {
2908 : PgStat_StatTabEntry *tabentry;
2909 :
2910 : /* fetch the pgstat table entry */
2911 346 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2912 : relid);
2913 :
2914 346 : relation_needs_vacanalyze(relid, avopts, classForm, tabentry,
2915 : effective_multixact_freeze_max_age,
2916 : dovacuum, doanalyze, wraparound);
2917 :
2918 : /* ignore ANALYZE for toast tables */
2919 346 : if (classForm->relkind == RELKIND_TOASTVALUE)
2920 4 : *doanalyze = false;
2921 346 : }
2922 :
2923 : /*
2924 : * relation_needs_vacanalyze
2925 : *
2926 : * Check whether a relation needs to be vacuumed or analyzed; return each into
2927 : * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is
2928 : * being forced because of Xid or multixact wraparound.
2929 : *
2930 : * relopts is a pointer to the AutoVacOpts options (either for itself in the
2931 : * case of a plain table, or for either itself or its parent table in the case
2932 : * of a TOAST table), NULL if none; tabentry is the pgstats entry, which can be
2933 : * NULL.
2934 : *
2935 : * A table needs to be vacuumed if the number of dead tuples exceeds a
2936 : * threshold. This threshold is calculated as
2937 : *
2938 : * threshold = vac_base_thresh + vac_scale_factor * reltuples
2939 : *
2940 : * For analyze, the analysis done is that the number of tuples inserted,
2941 : * deleted and updated since the last analyze exceeds a threshold calculated
2942 : * in the same fashion as above. Note that the cumulative stats system stores
2943 : * the number of tuples (both live and dead) that there were as of the last
2944 : * analyze. This is asymmetric to the VACUUM case.
2945 : *
2946 : * We also force vacuum if the table's relfrozenxid is more than freeze_max_age
2947 : * transactions back, and if its relminmxid is more than
2948 : * multixact_freeze_max_age multixacts back.
2949 : *
2950 : * A table whose autovacuum_enabled option is false is
2951 : * automatically skipped (unless we have to vacuum it due to freeze_max_age).
2952 : * Thus autovacuum can be disabled for specific tables. Also, when the cumulative
2953 : * stats system does not have data about a table, it will be skipped.
2954 : *
2955 : * A table whose vac_base_thresh value is < 0 takes the base value from the
2956 : * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor
2957 : * value < 0 is substituted with the value of
2958 : * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze.
2959 : */
2960 : static void
2961 3676 : relation_needs_vacanalyze(Oid relid,
2962 : AutoVacOpts *relopts,
2963 : Form_pg_class classForm,
2964 : PgStat_StatTabEntry *tabentry,
2965 : int effective_multixact_freeze_max_age,
2966 : /* output params below */
2967 : bool *dovacuum,
2968 : bool *doanalyze,
2969 : bool *wraparound)
2970 : {
2971 : bool force_vacuum;
2972 : bool av_enabled;
2973 : float4 reltuples; /* pg_class.reltuples */
2974 :
2975 : /* constants from reloptions or GUC variables */
2976 : int vac_base_thresh,
2977 : vac_ins_base_thresh,
2978 : anl_base_thresh;
2979 : float4 vac_scale_factor,
2980 : vac_ins_scale_factor,
2981 : anl_scale_factor;
2982 :
2983 : /* thresholds calculated from above constants */
2984 : float4 vacthresh,
2985 : vacinsthresh,
2986 : anlthresh;
2987 :
2988 : /* number of vacuum (resp. analyze) tuples at this time */
2989 : float4 vactuples,
2990 : instuples,
2991 : anltuples;
2992 :
2993 : /* freeze parameters */
2994 : int freeze_max_age;
2995 : int multixact_freeze_max_age;
2996 : TransactionId xidForceLimit;
2997 : MultiXactId multiForceLimit;
2998 :
2999 : AssertArg(classForm != NULL);
3000 : AssertArg(OidIsValid(relid));
3001 :
3002 : /*
3003 : * Determine vacuum/analyze equation parameters. We have two possible
3004 : * sources: the passed reloptions (which could be a main table or a toast
3005 : * table), or the autovacuum GUC variables.
3006 : */
3007 :
3008 : /* -1 in autovac setting means use plain vacuum_scale_factor */
3009 116 : vac_scale_factor = (relopts && relopts->vacuum_scale_factor >= 0)
3010 0 : ? relopts->vacuum_scale_factor
3011 3792 : : autovacuum_vac_scale;
3012 :
3013 116 : vac_base_thresh = (relopts && relopts->vacuum_threshold >= 0)
3014 : ? relopts->vacuum_threshold
3015 3792 : : autovacuum_vac_thresh;
3016 :
3017 116 : vac_ins_scale_factor = (relopts && relopts->vacuum_ins_scale_factor >= 0)
3018 0 : ? relopts->vacuum_ins_scale_factor
3019 3792 : : autovacuum_vac_ins_scale;
3020 :
3021 : /* -1 is used to disable insert vacuums */
3022 116 : vac_ins_base_thresh = (relopts && relopts->vacuum_ins_threshold >= -1)
3023 : ? relopts->vacuum_ins_threshold
3024 3792 : : autovacuum_vac_ins_thresh;
3025 :
3026 116 : anl_scale_factor = (relopts && relopts->analyze_scale_factor >= 0)
3027 0 : ? relopts->analyze_scale_factor
3028 3792 : : autovacuum_anl_scale;
3029 :
3030 116 : anl_base_thresh = (relopts && relopts->analyze_threshold >= 0)
3031 : ? relopts->analyze_threshold
3032 3792 : : autovacuum_anl_thresh;
3033 :
3034 116 : freeze_max_age = (relopts && relopts->freeze_max_age >= 0)
3035 0 : ? Min(relopts->freeze_max_age, autovacuum_freeze_max_age)
3036 3792 : : autovacuum_freeze_max_age;
3037 :
3038 116 : multixact_freeze_max_age = (relopts && relopts->multixact_freeze_max_age >= 0)
3039 0 : ? Min(relopts->multixact_freeze_max_age, effective_multixact_freeze_max_age)
3040 3792 : : effective_multixact_freeze_max_age;
3041 :
3042 3676 : av_enabled = (relopts ? relopts->enabled : true);
3043 :
3044 : /* Force vacuum if table is at risk of wraparound */
3045 3676 : xidForceLimit = recentXid - freeze_max_age;
3046 3676 : if (xidForceLimit < FirstNormalTransactionId)
3047 0 : xidForceLimit -= FirstNormalTransactionId;
3048 7352 : force_vacuum = (TransactionIdIsNormal(classForm->relfrozenxid) &&
3049 3676 : TransactionIdPrecedes(classForm->relfrozenxid,
3050 : xidForceLimit));
3051 3676 : if (!force_vacuum)
3052 : {
3053 3676 : multiForceLimit = recentMulti - multixact_freeze_max_age;
3054 3676 : if (multiForceLimit < FirstMultiXactId)
3055 0 : multiForceLimit -= FirstMultiXactId;
3056 7352 : force_vacuum = MultiXactIdIsValid(classForm->relminmxid) &&
3057 3676 : MultiXactIdPrecedes(classForm->relminmxid, multiForceLimit);
3058 : }
3059 3676 : *wraparound = force_vacuum;
3060 :
3061 : /* User disabled it in pg_class.reloptions? (But ignore if at risk) */
3062 3676 : if (!av_enabled && !force_vacuum)
3063 : {
3064 80 : *doanalyze = false;
3065 80 : *dovacuum = false;
3066 80 : return;
3067 : }
3068 :
3069 : /*
3070 : * If we found stats for the table, and autovacuum is currently enabled,
3071 : * make a threshold-based decision whether to vacuum and/or analyze. If
3072 : * autovacuum is currently disabled, we must be here for anti-wraparound
3073 : * vacuuming only, so don't vacuum (or analyze) anything that's not being
3074 : * forced.
3075 : */
3076 3596 : if (PointerIsValid(tabentry) && AutoVacuumingActive())
3077 : {
3078 2896 : reltuples = classForm->reltuples;
3079 2896 : vactuples = tabentry->n_dead_tuples;
3080 2896 : instuples = tabentry->inserts_since_vacuum;
3081 2896 : anltuples = tabentry->changes_since_analyze;
3082 :
3083 : /* If the table hasn't yet been vacuumed, take reltuples as zero */
3084 2896 : if (reltuples < 0)
3085 648 : reltuples = 0;
3086 :
3087 2896 : vacthresh = (float4) vac_base_thresh + vac_scale_factor * reltuples;
3088 2896 : vacinsthresh = (float4) vac_ins_base_thresh + vac_ins_scale_factor * reltuples;
3089 2896 : anlthresh = (float4) anl_base_thresh + anl_scale_factor * reltuples;
3090 :
3091 : /*
3092 : * Note that we don't need to take special consideration for stat
3093 : * reset, because if that happens, the last vacuum and analyze counts
3094 : * will be reset too.
3095 : */
3096 2896 : if (vac_ins_base_thresh >= 0)
3097 2896 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: %.0f (threshold %.0f), anl: %.0f (threshold %.0f)",
3098 : NameStr(classForm->relname),
3099 : vactuples, vacthresh, instuples, vacinsthresh, anltuples, anlthresh);
3100 : else
3101 0 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: (disabled), anl: %.0f (threshold %.0f)",
3102 : NameStr(classForm->relname),
3103 : vactuples, vacthresh, anltuples, anlthresh);
3104 :
3105 : /* Determine if this table needs vacuum or analyze. */
3106 5586 : *dovacuum = force_vacuum || (vactuples > vacthresh) ||
3107 2690 : (vac_ins_base_thresh >= 0 && instuples > vacinsthresh);
3108 2896 : *doanalyze = (anltuples > anlthresh);
3109 : }
3110 : else
3111 : {
3112 : /*
3113 : * Skip a table not found in stat hash, unless we have to force vacuum
3114 : * for anti-wrap purposes. If it's not acted upon, there's no need to
3115 : * vacuum it.
3116 : */
3117 700 : *dovacuum = force_vacuum;
3118 700 : *doanalyze = false;
3119 : }
3120 :
3121 : /* ANALYZE refuses to work with pg_statistic */
3122 3596 : if (relid == StatisticRelationId)
3123 20 : *doanalyze = false;
3124 : }
3125 :
3126 : /*
3127 : * autovacuum_do_vac_analyze
3128 : * Vacuum and/or analyze the specified table
3129 : */
3130 : static void
3131 346 : autovacuum_do_vac_analyze(autovac_table *tab, BufferAccessStrategy bstrategy)
3132 : {
3133 : RangeVar *rangevar;
3134 : VacuumRelation *rel;
3135 : List *rel_list;
3136 :
3137 : /* Let pgstat know what we're doing */
3138 346 : autovac_report_activity(tab);
3139 :
3140 : /* Set up one VacuumRelation target, identified by OID, for vacuum() */
3141 346 : rangevar = makeRangeVar(tab->at_nspname, tab->at_relname, -1);
3142 346 : rel = makeVacuumRelation(rangevar, tab->at_relid, NIL);
3143 346 : rel_list = list_make1(rel);
3144 :
3145 346 : vacuum(rel_list, &tab->at_params, bstrategy, true);
3146 346 : }
3147 :
3148 : /*
3149 : * autovac_report_activity
3150 : * Report to pgstat what autovacuum is doing
3151 : *
3152 : * We send a SQL string corresponding to what the user would see if the
3153 : * equivalent command was to be issued manually.
3154 : *
3155 : * Note we assume that we are going to report the next command as soon as we're
3156 : * done with the current one, and exit right after the last one, so we don't
3157 : * bother to report "<IDLE>" or some such.
3158 : */
3159 : static void
3160 346 : autovac_report_activity(autovac_table *tab)
3161 : {
3162 : #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56)
3163 : char activity[MAX_AUTOVAC_ACTIV_LEN];
3164 : int len;
3165 :
3166 : /* Report the command and possible options */
3167 346 : if (tab->at_params.options & VACOPT_VACUUM)
3168 158 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3169 : "autovacuum: VACUUM%s",
3170 158 : tab->at_params.options & VACOPT_ANALYZE ? " ANALYZE" : "");
3171 : else
3172 188 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3173 : "autovacuum: ANALYZE");
3174 :
3175 : /*
3176 : * Report the qualified name of the relation.
3177 : */
3178 346 : len = strlen(activity);
3179 :
3180 346 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3181 : " %s.%s%s", tab->at_nspname, tab->at_relname,
3182 346 : tab->at_params.is_wraparound ? " (to prevent wraparound)" : "");
3183 :
3184 : /* Set statement_timestamp() to current time for pg_stat_activity */
3185 346 : SetCurrentStatementStartTimestamp();
3186 :
3187 346 : pgstat_report_activity(STATE_RUNNING, activity);
3188 346 : }
3189 :
3190 : /*
3191 : * autovac_report_workitem
3192 : * Report to pgstat that autovacuum is processing a work item
3193 : */
3194 : static void
3195 6 : autovac_report_workitem(AutoVacuumWorkItem *workitem,
3196 : const char *nspname, const char *relname)
3197 : {
3198 : char activity[MAX_AUTOVAC_ACTIV_LEN + 12 + 2];
3199 : char blk[12 + 2];
3200 : int len;
3201 :
3202 6 : switch (workitem->avw_type)
3203 : {
3204 6 : case AVW_BRINSummarizeRange:
3205 6 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3206 : "autovacuum: BRIN summarize");
3207 6 : break;
3208 : }
3209 :
3210 : /*
3211 : * Report the qualified name of the relation, and the block number if any
3212 : */
3213 6 : len = strlen(activity);
3214 :
3215 6 : if (BlockNumberIsValid(workitem->avw_blockNumber))
3216 6 : snprintf(blk, sizeof(blk), " %u", workitem->avw_blockNumber);
3217 : else
3218 0 : blk[0] = '\0';
3219 :
3220 6 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3221 : " %s.%s%s", nspname, relname, blk);
3222 :
3223 : /* Set statement_timestamp() to current time for pg_stat_activity */
3224 6 : SetCurrentStatementStartTimestamp();
3225 :
3226 6 : pgstat_report_activity(STATE_RUNNING, activity);
3227 6 : }
3228 :
3229 : /*
3230 : * AutoVacuumingActive
3231 : * Check GUC vars and report whether the autovacuum process should be
3232 : * running.
3233 : */
3234 : bool
3235 10160 : AutoVacuumingActive(void)
3236 : {
3237 10160 : if (!autovacuum_start_daemon || !pgstat_track_counts)
3238 1760 : return false;
3239 8400 : return true;
3240 : }
3241 :
3242 : /*
3243 : * Request one work item to the next autovacuum run processing our database.
3244 : * Return false if the request can't be recorded.
3245 : */
3246 : bool
3247 6 : AutoVacuumRequestWork(AutoVacuumWorkItemType type, Oid relationId,
3248 : BlockNumber blkno)
3249 : {
3250 : int i;
3251 6 : bool result = false;
3252 :
3253 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
3254 :
3255 : /*
3256 : * Locate an unused work item and fill it with the given data.
3257 : */
3258 12 : for (i = 0; i < NUM_WORKITEMS; i++)
3259 : {
3260 12 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
3261 :
3262 12 : if (workitem->avw_used)
3263 6 : continue;
3264 :
3265 6 : workitem->avw_used = true;
3266 6 : workitem->avw_active = false;
3267 6 : workitem->avw_type = type;
3268 6 : workitem->avw_database = MyDatabaseId;
3269 6 : workitem->avw_relation = relationId;
3270 6 : workitem->avw_blockNumber = blkno;
3271 6 : result = true;
3272 :
3273 : /* done */
3274 6 : break;
3275 : }
3276 :
3277 6 : LWLockRelease(AutovacuumLock);
3278 :
3279 6 : return result;
3280 : }
3281 :
3282 : /*
3283 : * autovac_init
3284 : * This is called at postmaster initialization.
3285 : *
3286 : * All we do here is annoy the user if he got it wrong.
3287 : */
3288 : void
3289 1044 : autovac_init(void)
3290 : {
3291 1044 : if (autovacuum_start_daemon && !pgstat_track_counts)
3292 0 : ereport(WARNING,
3293 : (errmsg("autovacuum not started because of misconfiguration"),
3294 : errhint("Enable the \"track_counts\" option.")));
3295 1044 : }
3296 :
3297 : /*
3298 : * IsAutoVacuum functions
3299 : * Return whether this is either a launcher autovacuum process or a worker
3300 : * process.
3301 : */
3302 : bool
3303 80330 : IsAutoVacuumLauncherProcess(void)
3304 : {
3305 80330 : return am_autovacuum_launcher;
3306 : }
3307 :
3308 : bool
3309 411960 : IsAutoVacuumWorkerProcess(void)
3310 : {
3311 411960 : return am_autovacuum_worker;
3312 : }
3313 :
3314 :
3315 : /*
3316 : * AutoVacuumShmemSize
3317 : * Compute space needed for autovacuum-related shared memory
3318 : */
3319 : Size
3320 7448 : AutoVacuumShmemSize(void)
3321 : {
3322 : Size size;
3323 :
3324 : /*
3325 : * Need the fixed struct and the array of WorkerInfoData.
3326 : */
3327 7448 : size = sizeof(AutoVacuumShmemStruct);
3328 7448 : size = MAXALIGN(size);
3329 7448 : size = add_size(size, mul_size(autovacuum_max_workers,
3330 : sizeof(WorkerInfoData)));
3331 7448 : return size;
3332 : }
3333 :
3334 : /*
3335 : * AutoVacuumShmemInit
3336 : * Allocate and initialize autovacuum-related shared memory
3337 : */
3338 : void
3339 3198 : AutoVacuumShmemInit(void)
3340 : {
3341 : bool found;
3342 :
3343 3198 : AutoVacuumShmem = (AutoVacuumShmemStruct *)
3344 3198 : ShmemInitStruct("AutoVacuum Data",
3345 : AutoVacuumShmemSize(),
3346 : &found);
3347 :
3348 3198 : if (!IsUnderPostmaster)
3349 : {
3350 : WorkerInfo worker;
3351 : int i;
3352 :
3353 : Assert(!found);
3354 :
3355 3198 : AutoVacuumShmem->av_launcherpid = 0;
3356 3198 : dlist_init(&AutoVacuumShmem->av_freeWorkers);
3357 3198 : dlist_init(&AutoVacuumShmem->av_runningWorkers);
3358 3198 : AutoVacuumShmem->av_startingWorker = NULL;
3359 3198 : memset(AutoVacuumShmem->av_workItems, 0,
3360 : sizeof(AutoVacuumWorkItem) * NUM_WORKITEMS);
3361 :
3362 3198 : worker = (WorkerInfo) ((char *) AutoVacuumShmem +
3363 : MAXALIGN(sizeof(AutoVacuumShmemStruct)));
3364 :
3365 : /* initialize the WorkerInfo free list */
3366 12792 : for (i = 0; i < autovacuum_max_workers; i++)
3367 9594 : dlist_push_head(&AutoVacuumShmem->av_freeWorkers,
3368 9594 : &worker[i].wi_links);
3369 : }
3370 : else
3371 : Assert(found);
3372 3198 : }
|
