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) 1996-2025, 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 "catalog/pg_namespace.h"
80 : #include "commands/dbcommands.h"
81 : #include "commands/vacuum.h"
82 : #include "common/int.h"
83 : #include "lib/ilist.h"
84 : #include "libpq/pqsignal.h"
85 : #include "miscadmin.h"
86 : #include "nodes/makefuncs.h"
87 : #include "pgstat.h"
88 : #include "postmaster/autovacuum.h"
89 : #include "postmaster/interrupt.h"
90 : #include "postmaster/postmaster.h"
91 : #include "storage/aio_subsys.h"
92 : #include "storage/bufmgr.h"
93 : #include "storage/ipc.h"
94 : #include "storage/latch.h"
95 : #include "storage/lmgr.h"
96 : #include "storage/pmsignal.h"
97 : #include "storage/proc.h"
98 : #include "storage/procsignal.h"
99 : #include "storage/smgr.h"
100 : #include "tcop/tcopprot.h"
101 : #include "utils/fmgroids.h"
102 : #include "utils/fmgrprotos.h"
103 : #include "utils/guc_hooks.h"
104 : #include "utils/injection_point.h"
105 : #include "utils/lsyscache.h"
106 : #include "utils/memutils.h"
107 : #include "utils/ps_status.h"
108 : #include "utils/rel.h"
109 : #include "utils/snapmgr.h"
110 : #include "utils/syscache.h"
111 : #include "utils/timeout.h"
112 : #include "utils/timestamp.h"
113 :
114 :
115 : /*
116 : * GUC parameters
117 : */
118 : bool autovacuum_start_daemon = false;
119 : int autovacuum_worker_slots;
120 : int autovacuum_max_workers;
121 : int autovacuum_work_mem = -1;
122 : int autovacuum_naptime;
123 : int autovacuum_vac_thresh;
124 : int autovacuum_vac_max_thresh;
125 : double autovacuum_vac_scale;
126 : int autovacuum_vac_ins_thresh;
127 : double autovacuum_vac_ins_scale;
128 : int autovacuum_anl_thresh;
129 : double autovacuum_anl_scale;
130 : int autovacuum_freeze_max_age;
131 : int autovacuum_multixact_freeze_max_age;
132 :
133 : double autovacuum_vac_cost_delay;
134 : int autovacuum_vac_cost_limit;
135 :
136 : int Log_autovacuum_min_duration = 600000;
137 :
138 : /* the minimum allowed time between two awakenings of the launcher */
139 : #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */
140 : #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */
141 :
142 : /*
143 : * Variables to save the cost-related storage parameters for the current
144 : * relation being vacuumed by this autovacuum worker. Using these, we can
145 : * ensure we don't overwrite the values of vacuum_cost_delay and
146 : * vacuum_cost_limit after reloading the configuration file. They are
147 : * initialized to "invalid" values to indicate that no cost-related storage
148 : * parameters were specified and will be set in do_autovacuum() after checking
149 : * the storage parameters in table_recheck_autovac().
150 : */
151 : static double av_storage_param_cost_delay = -1;
152 : static int av_storage_param_cost_limit = -1;
153 :
154 : /* Flags set by signal handlers */
155 : static volatile sig_atomic_t got_SIGUSR2 = false;
156 :
157 : /* Comparison points for determining whether freeze_max_age is exceeded */
158 : static TransactionId recentXid;
159 : static MultiXactId recentMulti;
160 :
161 : /* Default freeze ages to use for autovacuum (varies by database) */
162 : static int default_freeze_min_age;
163 : static int default_freeze_table_age;
164 : static int default_multixact_freeze_min_age;
165 : static int default_multixact_freeze_table_age;
166 :
167 : /* Memory context for long-lived data */
168 : static MemoryContext AutovacMemCxt;
169 :
170 : /* struct to keep track of databases in launcher */
171 : typedef struct avl_dbase
172 : {
173 : Oid adl_datid; /* hash key -- must be first */
174 : TimestampTz adl_next_worker;
175 : int adl_score;
176 : dlist_node adl_node;
177 : } avl_dbase;
178 :
179 : /* struct to keep track of databases in worker */
180 : typedef struct avw_dbase
181 : {
182 : Oid adw_datid;
183 : char *adw_name;
184 : TransactionId adw_frozenxid;
185 : MultiXactId adw_minmulti;
186 : PgStat_StatDBEntry *adw_entry;
187 : } avw_dbase;
188 :
189 : /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */
190 : typedef struct av_relation
191 : {
192 : Oid ar_toastrelid; /* hash key - must be first */
193 : Oid ar_relid;
194 : bool ar_hasrelopts;
195 : AutoVacOpts ar_reloptions; /* copy of AutoVacOpts from the main table's
196 : * reloptions, or NULL if none */
197 : } av_relation;
198 :
199 : /* struct to keep track of tables to vacuum and/or analyze, after rechecking */
200 : typedef struct autovac_table
201 : {
202 : Oid at_relid;
203 : VacuumParams at_params;
204 : double at_storage_param_vac_cost_delay;
205 : int at_storage_param_vac_cost_limit;
206 : bool at_dobalance;
207 : bool at_sharedrel;
208 : char *at_relname;
209 : char *at_nspname;
210 : char *at_datname;
211 : } autovac_table;
212 :
213 : /*-------------
214 : * This struct holds information about a single worker's whereabouts. We keep
215 : * an array of these in shared memory, sized according to
216 : * autovacuum_worker_slots.
217 : *
218 : * wi_links entry into free list or running list
219 : * wi_dboid OID of the database this worker is supposed to work on
220 : * wi_tableoid OID of the table currently being vacuumed, if any
221 : * wi_sharedrel flag indicating whether table is marked relisshared
222 : * wi_proc pointer to PGPROC of the running worker, NULL if not started
223 : * wi_launchtime Time at which this worker was launched
224 : * wi_dobalance Whether this worker should be included in balance calculations
225 : *
226 : * All fields are protected by AutovacuumLock, except for wi_tableoid and
227 : * wi_sharedrel which are protected by AutovacuumScheduleLock (note these
228 : * two fields are read-only for everyone except that worker itself).
229 : *-------------
230 : */
231 : typedef struct WorkerInfoData
232 : {
233 : dlist_node wi_links;
234 : Oid wi_dboid;
235 : Oid wi_tableoid;
236 : PGPROC *wi_proc;
237 : TimestampTz wi_launchtime;
238 : pg_atomic_flag wi_dobalance;
239 : bool wi_sharedrel;
240 : } WorkerInfoData;
241 :
242 : typedef struct WorkerInfoData *WorkerInfo;
243 :
244 : /*
245 : * Possible signals received by the launcher from remote processes. These are
246 : * stored atomically in shared memory so that other processes can set them
247 : * without locking.
248 : */
249 : typedef enum
250 : {
251 : AutoVacForkFailed, /* failed trying to start a worker */
252 : AutoVacRebalance, /* rebalance the cost limits */
253 : } AutoVacuumSignal;
254 :
255 : #define AutoVacNumSignals (AutoVacRebalance + 1)
256 :
257 : /*
258 : * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This
259 : * list is mostly protected by AutovacuumLock, except that if an item is
260 : * marked 'active' other processes must not modify the work-identifying
261 : * members.
262 : */
263 : typedef struct AutoVacuumWorkItem
264 : {
265 : AutoVacuumWorkItemType avw_type;
266 : bool avw_used; /* below data is valid */
267 : bool avw_active; /* being processed */
268 : Oid avw_database;
269 : Oid avw_relation;
270 : BlockNumber avw_blockNumber;
271 : } AutoVacuumWorkItem;
272 :
273 : #define NUM_WORKITEMS 256
274 :
275 : /*-------------
276 : * The main autovacuum shmem struct. On shared memory we store this main
277 : * struct and the array of WorkerInfo structs. This struct keeps:
278 : *
279 : * av_signal set by other processes to indicate various conditions
280 : * av_launcherpid the PID of the autovacuum launcher
281 : * av_freeWorkers the WorkerInfo freelist
282 : * av_runningWorkers the WorkerInfo non-free queue
283 : * av_startingWorker pointer to WorkerInfo currently being started (cleared by
284 : * the worker itself as soon as it's up and running)
285 : * av_workItems work item array
286 : * av_nworkersForBalance the number of autovacuum workers to use when
287 : * calculating the per worker cost limit
288 : *
289 : * This struct is protected by AutovacuumLock, except for av_signal and parts
290 : * of the worker list (see above).
291 : *-------------
292 : */
293 : typedef struct
294 : {
295 : sig_atomic_t av_signal[AutoVacNumSignals];
296 : pid_t av_launcherpid;
297 : dclist_head av_freeWorkers;
298 : dlist_head av_runningWorkers;
299 : WorkerInfo av_startingWorker;
300 : AutoVacuumWorkItem av_workItems[NUM_WORKITEMS];
301 : pg_atomic_uint32 av_nworkersForBalance;
302 : } AutoVacuumShmemStruct;
303 :
304 : static AutoVacuumShmemStruct *AutoVacuumShmem;
305 :
306 : /*
307 : * the database list (of avl_dbase elements) in the launcher, and the context
308 : * that contains it
309 : */
310 : static dlist_head DatabaseList = DLIST_STATIC_INIT(DatabaseList);
311 : static MemoryContext DatabaseListCxt = NULL;
312 :
313 : /* Pointer to my own WorkerInfo, valid on each worker */
314 : static WorkerInfo MyWorkerInfo = NULL;
315 :
316 : /* PID of launcher, valid only in worker while shutting down */
317 : int AutovacuumLauncherPid = 0;
318 :
319 : static Oid do_start_worker(void);
320 : static void ProcessAutoVacLauncherInterrupts(void);
321 : pg_noreturn static void AutoVacLauncherShutdown(void);
322 : static void launcher_determine_sleep(bool canlaunch, bool recursing,
323 : struct timeval *nap);
324 : static void launch_worker(TimestampTz now);
325 : static List *get_database_list(void);
326 : static void rebuild_database_list(Oid newdb);
327 : static int db_comparator(const void *a, const void *b);
328 : static void autovac_recalculate_workers_for_balance(void);
329 :
330 : static void do_autovacuum(void);
331 : static void FreeWorkerInfo(int code, Datum arg);
332 :
333 : static autovac_table *table_recheck_autovac(Oid relid, HTAB *table_toast_map,
334 : TupleDesc pg_class_desc,
335 : int effective_multixact_freeze_max_age);
336 : static void recheck_relation_needs_vacanalyze(Oid relid, AutoVacOpts *avopts,
337 : Form_pg_class classForm,
338 : int effective_multixact_freeze_max_age,
339 : bool *dovacuum, bool *doanalyze, bool *wraparound);
340 : static void relation_needs_vacanalyze(Oid relid, AutoVacOpts *relopts,
341 : Form_pg_class classForm,
342 : PgStat_StatTabEntry *tabentry,
343 : int effective_multixact_freeze_max_age,
344 : bool *dovacuum, bool *doanalyze, bool *wraparound);
345 :
346 : static void autovacuum_do_vac_analyze(autovac_table *tab,
347 : BufferAccessStrategy bstrategy);
348 : static AutoVacOpts *extract_autovac_opts(HeapTuple tup,
349 : TupleDesc pg_class_desc);
350 : static void perform_work_item(AutoVacuumWorkItem *workitem);
351 : static void autovac_report_activity(autovac_table *tab);
352 : static void autovac_report_workitem(AutoVacuumWorkItem *workitem,
353 : const char *nspname, const char *relname);
354 : static void avl_sigusr2_handler(SIGNAL_ARGS);
355 : static bool av_worker_available(void);
356 : static void check_av_worker_gucs(void);
357 :
358 :
359 :
360 : /********************************************************************
361 : * AUTOVACUUM LAUNCHER CODE
362 : ********************************************************************/
363 :
364 : /*
365 : * Main entry point for the autovacuum launcher process.
366 : */
367 : void
368 746 : AutoVacLauncherMain(const void *startup_data, size_t startup_data_len)
369 : {
370 : sigjmp_buf local_sigjmp_buf;
371 :
372 : Assert(startup_data_len == 0);
373 :
374 : /* Release postmaster's working memory context */
375 746 : if (PostmasterContext)
376 : {
377 746 : MemoryContextDelete(PostmasterContext);
378 746 : PostmasterContext = NULL;
379 : }
380 :
381 746 : MyBackendType = B_AUTOVAC_LAUNCHER;
382 746 : init_ps_display(NULL);
383 :
384 746 : ereport(DEBUG1,
385 : (errmsg_internal("autovacuum launcher started")));
386 :
387 746 : if (PostAuthDelay)
388 0 : pg_usleep(PostAuthDelay * 1000000L);
389 :
390 : Assert(GetProcessingMode() == InitProcessing);
391 :
392 : /*
393 : * Set up signal handlers. We operate on databases much like a regular
394 : * backend, so we use the same signal handling. See equivalent code in
395 : * tcop/postgres.c.
396 : */
397 746 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
398 746 : pqsignal(SIGINT, StatementCancelHandler);
399 746 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
400 : /* SIGQUIT handler was already set up by InitPostmasterChild */
401 :
402 746 : InitializeTimeouts(); /* establishes SIGALRM handler */
403 :
404 746 : pqsignal(SIGPIPE, SIG_IGN);
405 746 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
406 746 : pqsignal(SIGUSR2, avl_sigusr2_handler);
407 746 : pqsignal(SIGFPE, FloatExceptionHandler);
408 746 : pqsignal(SIGCHLD, SIG_DFL);
409 :
410 : /*
411 : * Create a per-backend PGPROC struct in shared memory. We must do this
412 : * before we can use LWLocks or access any shared memory.
413 : */
414 746 : InitProcess();
415 :
416 : /* Early initialization */
417 746 : BaseInit();
418 :
419 746 : InitPostgres(NULL, InvalidOid, NULL, InvalidOid, 0, NULL);
420 :
421 746 : SetProcessingMode(NormalProcessing);
422 :
423 : /*
424 : * Create a memory context that we will do all our work in. We do this so
425 : * that we can reset the context during error recovery and thereby avoid
426 : * possible memory leaks.
427 : */
428 746 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
429 : "Autovacuum Launcher",
430 : ALLOCSET_DEFAULT_SIZES);
431 746 : MemoryContextSwitchTo(AutovacMemCxt);
432 :
433 : /*
434 : * If an exception is encountered, processing resumes here.
435 : *
436 : * This code is a stripped down version of PostgresMain error recovery.
437 : *
438 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
439 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
440 : * signals other than SIGQUIT will be blocked until we complete error
441 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
442 : * call redundant, but it is not since InterruptPending might be set
443 : * already.
444 : */
445 746 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
446 : {
447 : /* since not using PG_TRY, must reset error stack by hand */
448 0 : error_context_stack = NULL;
449 :
450 : /* Prevents interrupts while cleaning up */
451 0 : HOLD_INTERRUPTS();
452 :
453 : /* Forget any pending QueryCancel or timeout request */
454 0 : disable_all_timeouts(false);
455 0 : QueryCancelPending = false; /* second to avoid race condition */
456 :
457 : /* Report the error to the server log */
458 0 : EmitErrorReport();
459 :
460 : /* Abort the current transaction in order to recover */
461 0 : AbortCurrentTransaction();
462 :
463 : /*
464 : * Release any other resources, for the case where we were not in a
465 : * transaction.
466 : */
467 0 : LWLockReleaseAll();
468 0 : pgstat_report_wait_end();
469 0 : pgaio_error_cleanup();
470 0 : UnlockBuffers();
471 : /* this is probably dead code, but let's be safe: */
472 0 : if (AuxProcessResourceOwner)
473 0 : ReleaseAuxProcessResources(false);
474 0 : AtEOXact_Buffers(false);
475 0 : AtEOXact_SMgr();
476 0 : AtEOXact_Files(false);
477 0 : AtEOXact_HashTables(false);
478 :
479 : /*
480 : * Now return to normal top-level context and clear ErrorContext for
481 : * next time.
482 : */
483 0 : MemoryContextSwitchTo(AutovacMemCxt);
484 0 : FlushErrorState();
485 :
486 : /* Flush any leaked data in the top-level context */
487 0 : MemoryContextReset(AutovacMemCxt);
488 :
489 : /* don't leave dangling pointers to freed memory */
490 0 : DatabaseListCxt = NULL;
491 0 : dlist_init(&DatabaseList);
492 :
493 : /* Now we can allow interrupts again */
494 0 : RESUME_INTERRUPTS();
495 :
496 : /* if in shutdown mode, no need for anything further; just go away */
497 0 : if (ShutdownRequestPending)
498 0 : AutoVacLauncherShutdown();
499 :
500 : /*
501 : * Sleep at least 1 second after any error. We don't want to be
502 : * filling the error logs as fast as we can.
503 : */
504 0 : pg_usleep(1000000L);
505 : }
506 :
507 : /* We can now handle ereport(ERROR) */
508 746 : PG_exception_stack = &local_sigjmp_buf;
509 :
510 : /* must unblock signals before calling rebuild_database_list */
511 746 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
512 :
513 : /*
514 : * Set always-secure search path. Launcher doesn't connect to a database,
515 : * so this has no effect.
516 : */
517 746 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
518 :
519 : /*
520 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
521 : * in postgresql.conf. We don't really want such a dangerous option being
522 : * applied non-interactively.
523 : */
524 746 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
525 :
526 : /*
527 : * Force settable timeouts off to avoid letting these settings prevent
528 : * regular maintenance from being executed.
529 : */
530 746 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
531 746 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
532 746 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
533 746 : SetConfigOption("idle_in_transaction_session_timeout", "0",
534 : PGC_SUSET, PGC_S_OVERRIDE);
535 :
536 : /*
537 : * Force default_transaction_isolation to READ COMMITTED. We don't want
538 : * to pay the overhead of serializable mode, nor add any risk of causing
539 : * deadlocks or delaying other transactions.
540 : */
541 746 : SetConfigOption("default_transaction_isolation", "read committed",
542 : PGC_SUSET, PGC_S_OVERRIDE);
543 :
544 : /*
545 : * Even when system is configured to use a different fetch consistency,
546 : * for autovac we always want fresh stats.
547 : */
548 746 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
549 :
550 : /*
551 : * In emergency mode, just start a worker (unless shutdown was requested)
552 : * and go away.
553 : */
554 746 : if (!AutoVacuumingActive())
555 : {
556 0 : if (!ShutdownRequestPending)
557 0 : do_start_worker();
558 0 : proc_exit(0); /* done */
559 : }
560 :
561 746 : AutoVacuumShmem->av_launcherpid = MyProcPid;
562 :
563 : /*
564 : * Create the initial database list. The invariant we want this list to
565 : * keep is that it's ordered by decreasing next_time. As soon as an entry
566 : * is updated to a higher time, it will be moved to the front (which is
567 : * correct because the only operation is to add autovacuum_naptime to the
568 : * entry, and time always increases).
569 : */
570 746 : rebuild_database_list(InvalidOid);
571 :
572 : /* loop until shutdown request */
573 9784 : while (!ShutdownRequestPending)
574 : {
575 : struct timeval nap;
576 9782 : TimestampTz current_time = 0;
577 : bool can_launch;
578 :
579 : /*
580 : * This loop is a bit different from the normal use of WaitLatch,
581 : * because we'd like to sleep before the first launch of a child
582 : * process. So it's WaitLatch, then ResetLatch, then check for
583 : * wakening conditions.
584 : */
585 :
586 9782 : launcher_determine_sleep(av_worker_available(), false, &nap);
587 :
588 : /*
589 : * Wait until naptime expires or we get some type of signal (all the
590 : * signal handlers will wake us by calling SetLatch).
591 : */
592 9782 : (void) WaitLatch(MyLatch,
593 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
594 9782 : (nap.tv_sec * 1000L) + (nap.tv_usec / 1000L),
595 : WAIT_EVENT_AUTOVACUUM_MAIN);
596 :
597 9778 : ResetLatch(MyLatch);
598 :
599 9778 : ProcessAutoVacLauncherInterrupts();
600 :
601 : /*
602 : * a worker finished, or postmaster signaled failure to start a worker
603 : */
604 9038 : if (got_SIGUSR2)
605 : {
606 4740 : got_SIGUSR2 = false;
607 :
608 : /* rebalance cost limits, if needed */
609 4740 : if (AutoVacuumShmem->av_signal[AutoVacRebalance])
610 : {
611 2368 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
612 2368 : AutoVacuumShmem->av_signal[AutoVacRebalance] = false;
613 2368 : autovac_recalculate_workers_for_balance();
614 2368 : LWLockRelease(AutovacuumLock);
615 : }
616 :
617 4740 : if (AutoVacuumShmem->av_signal[AutoVacForkFailed])
618 : {
619 : /*
620 : * If the postmaster failed to start a new worker, we sleep
621 : * for a little while and resend the signal. The new worker's
622 : * state is still in memory, so this is sufficient. After
623 : * that, we restart the main loop.
624 : *
625 : * XXX should we put a limit to the number of times we retry?
626 : * I don't think it makes much sense, because a future start
627 : * of a worker will continue to fail in the same way.
628 : */
629 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = false;
630 0 : pg_usleep(1000000L); /* 1s */
631 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
632 2056 : continue;
633 : }
634 : }
635 :
636 : /*
637 : * There are some conditions that we need to check before trying to
638 : * start a worker. First, we need to make sure that there is a worker
639 : * slot available. Second, we need to make sure that no other worker
640 : * failed while starting up.
641 : */
642 :
643 9038 : current_time = GetCurrentTimestamp();
644 9038 : LWLockAcquire(AutovacuumLock, LW_SHARED);
645 :
646 9038 : can_launch = av_worker_available();
647 :
648 9038 : if (AutoVacuumShmem->av_startingWorker != NULL)
649 : {
650 : int waittime;
651 50 : WorkerInfo worker = AutoVacuumShmem->av_startingWorker;
652 :
653 : /*
654 : * We can't launch another worker when another one is still
655 : * starting up (or failed while doing so), so just sleep for a bit
656 : * more; that worker will wake us up again as soon as it's ready.
657 : * We will only wait autovacuum_naptime seconds (up to a maximum
658 : * of 60 seconds) for this to happen however. Note that failure
659 : * to connect to a particular database is not a problem here,
660 : * because the worker removes itself from the startingWorker
661 : * pointer before trying to connect. Problems detected by the
662 : * postmaster (like fork() failure) are also reported and handled
663 : * differently. The only problems that may cause this code to
664 : * fire are errors in the earlier sections of AutoVacWorkerMain,
665 : * before the worker removes the WorkerInfo from the
666 : * startingWorker pointer.
667 : */
668 50 : waittime = Min(autovacuum_naptime, 60) * 1000;
669 50 : if (TimestampDifferenceExceeds(worker->wi_launchtime, current_time,
670 : waittime))
671 : {
672 0 : LWLockRelease(AutovacuumLock);
673 0 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
674 :
675 : /*
676 : * No other process can put a worker in starting mode, so if
677 : * startingWorker is still INVALID after exchanging our lock,
678 : * we assume it's the same one we saw above (so we don't
679 : * recheck the launch time).
680 : */
681 0 : if (AutoVacuumShmem->av_startingWorker != NULL)
682 : {
683 0 : worker = AutoVacuumShmem->av_startingWorker;
684 0 : worker->wi_dboid = InvalidOid;
685 0 : worker->wi_tableoid = InvalidOid;
686 0 : worker->wi_sharedrel = false;
687 0 : worker->wi_proc = NULL;
688 0 : worker->wi_launchtime = 0;
689 0 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
690 : &worker->wi_links);
691 0 : AutoVacuumShmem->av_startingWorker = NULL;
692 0 : ereport(WARNING,
693 : errmsg("autovacuum worker took too long to start; canceled"));
694 : }
695 : }
696 : else
697 50 : can_launch = false;
698 : }
699 9038 : LWLockRelease(AutovacuumLock); /* either shared or exclusive */
700 :
701 : /* if we can't do anything, just go back to sleep */
702 9038 : if (!can_launch)
703 2056 : continue;
704 :
705 : /* We're OK to start a new worker */
706 :
707 6982 : if (dlist_is_empty(&DatabaseList))
708 : {
709 : /*
710 : * Special case when the list is empty: start a worker right away.
711 : * This covers the initial case, when no database is in pgstats
712 : * (thus the list is empty). Note that the constraints in
713 : * launcher_determine_sleep keep us from starting workers too
714 : * quickly (at most once every autovacuum_naptime when the list is
715 : * empty).
716 : */
717 2 : launch_worker(current_time);
718 : }
719 : else
720 : {
721 : /*
722 : * because rebuild_database_list constructs a list with most
723 : * distant adl_next_worker first, we obtain our database from the
724 : * tail of the list.
725 : */
726 : avl_dbase *avdb;
727 :
728 6980 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
729 :
730 : /*
731 : * launch a worker if next_worker is right now or it is in the
732 : * past
733 : */
734 6980 : if (TimestampDifferenceExceeds(avdb->adl_next_worker,
735 : current_time, 0))
736 2372 : launch_worker(current_time);
737 : }
738 : }
739 :
740 2 : AutoVacLauncherShutdown();
741 : }
742 :
743 : /*
744 : * Process any new interrupts.
745 : */
746 : static void
747 9778 : ProcessAutoVacLauncherInterrupts(void)
748 : {
749 : /* the normal shutdown case */
750 9778 : if (ShutdownRequestPending)
751 740 : AutoVacLauncherShutdown();
752 :
753 9038 : if (ConfigReloadPending)
754 : {
755 72 : int autovacuum_max_workers_prev = autovacuum_max_workers;
756 :
757 72 : ConfigReloadPending = false;
758 72 : ProcessConfigFile(PGC_SIGHUP);
759 :
760 : /* shutdown requested in config file? */
761 72 : if (!AutoVacuumingActive())
762 0 : AutoVacLauncherShutdown();
763 :
764 : /*
765 : * If autovacuum_max_workers changed, emit a WARNING if
766 : * autovacuum_worker_slots < autovacuum_max_workers. If it didn't
767 : * change, skip this to avoid too many repeated log messages.
768 : */
769 72 : if (autovacuum_max_workers_prev != autovacuum_max_workers)
770 0 : check_av_worker_gucs();
771 :
772 : /* rebuild the list in case the naptime changed */
773 72 : rebuild_database_list(InvalidOid);
774 : }
775 :
776 : /* Process barrier events */
777 9038 : if (ProcSignalBarrierPending)
778 52 : ProcessProcSignalBarrier();
779 :
780 : /* Perform logging of memory contexts of this process */
781 9038 : if (LogMemoryContextPending)
782 0 : ProcessLogMemoryContextInterrupt();
783 :
784 : /* Publish memory contexts of this process */
785 9038 : if (PublishMemoryContextPending)
786 0 : ProcessGetMemoryContextInterrupt();
787 :
788 : /* Process sinval catchup interrupts that happened while sleeping */
789 9038 : ProcessCatchupInterrupt();
790 9038 : }
791 :
792 : /*
793 : * Perform a normal exit from the autovac launcher.
794 : */
795 : static void
796 742 : AutoVacLauncherShutdown(void)
797 : {
798 742 : ereport(DEBUG1,
799 : (errmsg_internal("autovacuum launcher shutting down")));
800 742 : AutoVacuumShmem->av_launcherpid = 0;
801 :
802 742 : proc_exit(0); /* done */
803 : }
804 :
805 : /*
806 : * Determine the time to sleep, based on the database list.
807 : *
808 : * The "canlaunch" parameter indicates whether we can start a worker right now,
809 : * for example due to the workers being all busy. If this is false, we will
810 : * cause a long sleep, which will be interrupted when a worker exits.
811 : */
812 : static void
813 9968 : launcher_determine_sleep(bool canlaunch, bool recursing, struct timeval *nap)
814 : {
815 : /*
816 : * We sleep until the next scheduled vacuum. We trust that when the
817 : * database list was built, care was taken so that no entries have times
818 : * in the past; if the first entry has too close a next_worker value, or a
819 : * time in the past, we will sleep a small nominal time.
820 : */
821 9968 : if (!canlaunch)
822 : {
823 3978 : nap->tv_sec = autovacuum_naptime;
824 3978 : nap->tv_usec = 0;
825 : }
826 5990 : else if (!dlist_is_empty(&DatabaseList))
827 : {
828 5944 : TimestampTz current_time = GetCurrentTimestamp();
829 : TimestampTz next_wakeup;
830 : avl_dbase *avdb;
831 : long secs;
832 : int usecs;
833 :
834 5944 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
835 :
836 5944 : next_wakeup = avdb->adl_next_worker;
837 5944 : TimestampDifference(current_time, next_wakeup, &secs, &usecs);
838 :
839 5944 : nap->tv_sec = secs;
840 5944 : nap->tv_usec = usecs;
841 : }
842 : else
843 : {
844 : /* list is empty, sleep for whole autovacuum_naptime seconds */
845 46 : nap->tv_sec = autovacuum_naptime;
846 46 : nap->tv_usec = 0;
847 : }
848 :
849 : /*
850 : * If the result is exactly zero, it means a database had an entry with
851 : * time in the past. Rebuild the list so that the databases are evenly
852 : * distributed again, and recalculate the time to sleep. This can happen
853 : * if there are more tables needing vacuum than workers, and they all take
854 : * longer to vacuum than autovacuum_naptime.
855 : *
856 : * We only recurse once. rebuild_database_list should always return times
857 : * in the future, but it seems best not to trust too much on that.
858 : */
859 9968 : if (nap->tv_sec == 0 && nap->tv_usec == 0 && !recursing)
860 : {
861 186 : rebuild_database_list(InvalidOid);
862 186 : launcher_determine_sleep(canlaunch, true, nap);
863 186 : return;
864 : }
865 :
866 : /* The smallest time we'll allow the launcher to sleep. */
867 9782 : if (nap->tv_sec <= 0 && nap->tv_usec <= MIN_AUTOVAC_SLEEPTIME * 1000)
868 : {
869 1024 : nap->tv_sec = 0;
870 1024 : nap->tv_usec = MIN_AUTOVAC_SLEEPTIME * 1000;
871 : }
872 :
873 : /*
874 : * If the sleep time is too large, clamp it to an arbitrary maximum (plus
875 : * any fractional seconds, for simplicity). This avoids an essentially
876 : * infinite sleep in strange cases like the system clock going backwards a
877 : * few years.
878 : */
879 9782 : if (nap->tv_sec > MAX_AUTOVAC_SLEEPTIME)
880 18 : nap->tv_sec = MAX_AUTOVAC_SLEEPTIME;
881 : }
882 :
883 : /*
884 : * Build an updated DatabaseList. It must only contain databases that appear
885 : * in pgstats, and must be sorted by next_worker from highest to lowest,
886 : * distributed regularly across the next autovacuum_naptime interval.
887 : *
888 : * Receives the Oid of the database that made this list be generated (we call
889 : * this the "new" database, because when the database was already present on
890 : * the list, we expect that this function is not called at all). The
891 : * preexisting list, if any, will be used to preserve the order of the
892 : * databases in the autovacuum_naptime period. The new database is put at the
893 : * end of the interval. The actual values are not saved, which should not be
894 : * much of a problem.
895 : */
896 : static void
897 1030 : rebuild_database_list(Oid newdb)
898 : {
899 : List *dblist;
900 : ListCell *cell;
901 : MemoryContext newcxt;
902 : MemoryContext oldcxt;
903 : MemoryContext tmpcxt;
904 : HASHCTL hctl;
905 : int score;
906 : int nelems;
907 : HTAB *dbhash;
908 : dlist_iter iter;
909 :
910 1030 : newcxt = AllocSetContextCreate(AutovacMemCxt,
911 : "Autovacuum database list",
912 : ALLOCSET_DEFAULT_SIZES);
913 1030 : tmpcxt = AllocSetContextCreate(newcxt,
914 : "Autovacuum database list (tmp)",
915 : ALLOCSET_DEFAULT_SIZES);
916 1030 : oldcxt = MemoryContextSwitchTo(tmpcxt);
917 :
918 : /*
919 : * Implementing this is not as simple as it sounds, because we need to put
920 : * the new database at the end of the list; next the databases that were
921 : * already on the list, and finally (at the tail of the list) all the
922 : * other databases that are not on the existing list.
923 : *
924 : * To do this, we build an empty hash table of scored databases. We will
925 : * start with the lowest score (zero) for the new database, then
926 : * increasing scores for the databases in the existing list, in order, and
927 : * lastly increasing scores for all databases gotten via
928 : * get_database_list() that are not already on the hash.
929 : *
930 : * Then we will put all the hash elements into an array, sort the array by
931 : * score, and finally put the array elements into the new doubly linked
932 : * list.
933 : */
934 1030 : hctl.keysize = sizeof(Oid);
935 1030 : hctl.entrysize = sizeof(avl_dbase);
936 1030 : hctl.hcxt = tmpcxt;
937 1030 : dbhash = hash_create("autovacuum db hash", 20, &hctl, /* magic number here
938 : * FIXME */
939 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
940 :
941 : /* start by inserting the new database */
942 1030 : score = 0;
943 1030 : if (OidIsValid(newdb))
944 : {
945 : avl_dbase *db;
946 : PgStat_StatDBEntry *entry;
947 :
948 : /* only consider this database if it has a pgstat entry */
949 26 : entry = pgstat_fetch_stat_dbentry(newdb);
950 26 : if (entry != NULL)
951 : {
952 : /* we assume it isn't found because the hash was just created */
953 20 : db = hash_search(dbhash, &newdb, HASH_ENTER, NULL);
954 :
955 : /* hash_search already filled in the key */
956 20 : db->adl_score = score++;
957 : /* next_worker is filled in later */
958 : }
959 : }
960 :
961 : /* Now insert the databases from the existing list */
962 1728 : dlist_foreach(iter, &DatabaseList)
963 : {
964 698 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
965 : avl_dbase *db;
966 : bool found;
967 : PgStat_StatDBEntry *entry;
968 :
969 : /*
970 : * skip databases with no stat entries -- in particular, this gets rid
971 : * of dropped databases
972 : */
973 698 : entry = pgstat_fetch_stat_dbentry(avdb->adl_datid);
974 698 : if (entry == NULL)
975 0 : continue;
976 :
977 698 : db = hash_search(dbhash, &(avdb->adl_datid), HASH_ENTER, &found);
978 :
979 698 : if (!found)
980 : {
981 : /* hash_search already filled in the key */
982 698 : db->adl_score = score++;
983 : /* next_worker is filled in later */
984 : }
985 : }
986 :
987 : /* finally, insert all qualifying databases not previously inserted */
988 1030 : dblist = get_database_list();
989 4424 : foreach(cell, dblist)
990 : {
991 3394 : avw_dbase *avdb = lfirst(cell);
992 : avl_dbase *db;
993 : bool found;
994 : PgStat_StatDBEntry *entry;
995 :
996 : /* only consider databases with a pgstat entry */
997 3394 : entry = pgstat_fetch_stat_dbentry(avdb->adw_datid);
998 3394 : if (entry == NULL)
999 1656 : continue;
1000 :
1001 1738 : db = hash_search(dbhash, &(avdb->adw_datid), HASH_ENTER, &found);
1002 : /* only update the score if the database was not already on the hash */
1003 1738 : if (!found)
1004 : {
1005 : /* hash_search already filled in the key */
1006 1020 : db->adl_score = score++;
1007 : /* next_worker is filled in later */
1008 : }
1009 : }
1010 1030 : nelems = score;
1011 :
1012 : /* from here on, the allocated memory belongs to the new list */
1013 1030 : MemoryContextSwitchTo(newcxt);
1014 1030 : dlist_init(&DatabaseList);
1015 :
1016 1030 : if (nelems > 0)
1017 : {
1018 : TimestampTz current_time;
1019 : int millis_increment;
1020 : avl_dbase *dbary;
1021 : avl_dbase *db;
1022 : HASH_SEQ_STATUS seq;
1023 : int i;
1024 :
1025 : /* put all the hash elements into an array */
1026 986 : dbary = palloc(nelems * sizeof(avl_dbase));
1027 :
1028 986 : i = 0;
1029 986 : hash_seq_init(&seq, dbhash);
1030 2724 : while ((db = hash_seq_search(&seq)) != NULL)
1031 1738 : memcpy(&(dbary[i++]), db, sizeof(avl_dbase));
1032 :
1033 : /* sort the array */
1034 986 : qsort(dbary, nelems, sizeof(avl_dbase), db_comparator);
1035 :
1036 : /*
1037 : * Determine the time interval between databases in the schedule. If
1038 : * we see that the configured naptime would take us to sleep times
1039 : * lower than our min sleep time (which launcher_determine_sleep is
1040 : * coded not to allow), silently use a larger naptime (but don't touch
1041 : * the GUC variable).
1042 : */
1043 986 : millis_increment = 1000.0 * autovacuum_naptime / nelems;
1044 986 : if (millis_increment <= MIN_AUTOVAC_SLEEPTIME)
1045 0 : millis_increment = MIN_AUTOVAC_SLEEPTIME * 1.1;
1046 :
1047 986 : current_time = GetCurrentTimestamp();
1048 :
1049 : /*
1050 : * move the elements from the array into the dlist, setting the
1051 : * next_worker while walking the array
1052 : */
1053 2724 : for (i = 0; i < nelems; i++)
1054 : {
1055 1738 : db = &(dbary[i]);
1056 :
1057 1738 : current_time = TimestampTzPlusMilliseconds(current_time,
1058 : millis_increment);
1059 1738 : db->adl_next_worker = current_time;
1060 :
1061 : /* later elements should go closer to the head of the list */
1062 1738 : dlist_push_head(&DatabaseList, &db->adl_node);
1063 : }
1064 : }
1065 :
1066 : /* all done, clean up memory */
1067 1030 : if (DatabaseListCxt != NULL)
1068 284 : MemoryContextDelete(DatabaseListCxt);
1069 1030 : MemoryContextDelete(tmpcxt);
1070 1030 : DatabaseListCxt = newcxt;
1071 1030 : MemoryContextSwitchTo(oldcxt);
1072 1030 : }
1073 :
1074 : /* qsort comparator for avl_dbase, using adl_score */
1075 : static int
1076 1076 : db_comparator(const void *a, const void *b)
1077 : {
1078 1076 : return pg_cmp_s32(((const avl_dbase *) a)->adl_score,
1079 : ((const avl_dbase *) b)->adl_score);
1080 : }
1081 :
1082 : /*
1083 : * do_start_worker
1084 : *
1085 : * Bare-bones procedure for starting an autovacuum worker from the launcher.
1086 : * It determines what database to work on, sets up shared memory stuff and
1087 : * signals postmaster to start the worker. It fails gracefully if invoked when
1088 : * autovacuum_workers are already active.
1089 : *
1090 : * Return value is the OID of the database that the worker is going to process,
1091 : * or InvalidOid if no worker was actually started.
1092 : */
1093 : static Oid
1094 2374 : do_start_worker(void)
1095 : {
1096 : List *dblist;
1097 : ListCell *cell;
1098 : TransactionId xidForceLimit;
1099 : MultiXactId multiForceLimit;
1100 : bool for_xid_wrap;
1101 : bool for_multi_wrap;
1102 : avw_dbase *avdb;
1103 : TimestampTz current_time;
1104 2374 : bool skipit = false;
1105 2374 : Oid retval = InvalidOid;
1106 : MemoryContext tmpcxt,
1107 : oldcxt;
1108 :
1109 : /* return quickly when there are no free workers */
1110 2374 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1111 2374 : if (!av_worker_available())
1112 : {
1113 0 : LWLockRelease(AutovacuumLock);
1114 0 : return InvalidOid;
1115 : }
1116 2374 : LWLockRelease(AutovacuumLock);
1117 :
1118 : /*
1119 : * Create and switch to a temporary context to avoid leaking the memory
1120 : * allocated for the database list.
1121 : */
1122 2374 : tmpcxt = AllocSetContextCreate(CurrentMemoryContext,
1123 : "Autovacuum start worker (tmp)",
1124 : ALLOCSET_DEFAULT_SIZES);
1125 2374 : oldcxt = MemoryContextSwitchTo(tmpcxt);
1126 :
1127 : /* Get a list of databases */
1128 2374 : dblist = get_database_list();
1129 :
1130 : /*
1131 : * Determine the oldest datfrozenxid/relfrozenxid that we will allow to
1132 : * pass without forcing a vacuum. (This limit can be tightened for
1133 : * particular tables, but not loosened.)
1134 : */
1135 2374 : recentXid = ReadNextTransactionId();
1136 2374 : xidForceLimit = recentXid - autovacuum_freeze_max_age;
1137 : /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */
1138 : /* this can cause the limit to go backwards by 3, but that's OK */
1139 2374 : if (xidForceLimit < FirstNormalTransactionId)
1140 0 : xidForceLimit -= FirstNormalTransactionId;
1141 :
1142 : /* Also determine the oldest datminmxid we will consider. */
1143 2374 : recentMulti = ReadNextMultiXactId();
1144 2374 : multiForceLimit = recentMulti - MultiXactMemberFreezeThreshold();
1145 2374 : if (multiForceLimit < FirstMultiXactId)
1146 0 : multiForceLimit -= FirstMultiXactId;
1147 :
1148 : /*
1149 : * Choose a database to connect to. We pick the database that was least
1150 : * recently auto-vacuumed, or one that needs vacuuming to prevent Xid
1151 : * wraparound-related data loss. If any db at risk of Xid wraparound is
1152 : * found, we pick the one with oldest datfrozenxid, independently of
1153 : * autovacuum times; similarly we pick the one with the oldest datminmxid
1154 : * if any is in MultiXactId wraparound. Note that those in Xid wraparound
1155 : * danger are given more priority than those in multi wraparound danger.
1156 : *
1157 : * Note that a database with no stats entry is not considered, except for
1158 : * Xid wraparound purposes. The theory is that if no one has ever
1159 : * connected to it since the stats were last initialized, it doesn't need
1160 : * vacuuming.
1161 : *
1162 : * XXX This could be improved if we had more info about whether it needs
1163 : * vacuuming before connecting to it. Perhaps look through the pgstats
1164 : * data for the database's tables? One idea is to keep track of the
1165 : * number of new and dead tuples per database in pgstats. However it
1166 : * isn't clear how to construct a metric that measures that and not cause
1167 : * starvation for less busy databases.
1168 : */
1169 2374 : avdb = NULL;
1170 2374 : for_xid_wrap = false;
1171 2374 : for_multi_wrap = false;
1172 2374 : current_time = GetCurrentTimestamp();
1173 9552 : foreach(cell, dblist)
1174 : {
1175 7178 : avw_dbase *tmp = lfirst(cell);
1176 : dlist_iter iter;
1177 :
1178 : /* Check to see if this one is at risk of wraparound */
1179 7178 : if (TransactionIdPrecedes(tmp->adw_frozenxid, xidForceLimit))
1180 : {
1181 3850 : if (avdb == NULL ||
1182 1478 : TransactionIdPrecedes(tmp->adw_frozenxid,
1183 : avdb->adw_frozenxid))
1184 982 : avdb = tmp;
1185 2372 : for_xid_wrap = true;
1186 5458 : continue;
1187 : }
1188 4806 : else if (for_xid_wrap)
1189 192 : continue; /* ignore not-at-risk DBs */
1190 4614 : else if (MultiXactIdPrecedes(tmp->adw_minmulti, multiForceLimit))
1191 : {
1192 0 : if (avdb == NULL ||
1193 0 : MultiXactIdPrecedes(tmp->adw_minmulti, avdb->adw_minmulti))
1194 0 : avdb = tmp;
1195 0 : for_multi_wrap = true;
1196 0 : continue;
1197 : }
1198 4614 : else if (for_multi_wrap)
1199 0 : continue; /* ignore not-at-risk DBs */
1200 :
1201 : /* Find pgstat entry if any */
1202 4614 : tmp->adw_entry = pgstat_fetch_stat_dbentry(tmp->adw_datid);
1203 :
1204 : /*
1205 : * Skip a database with no pgstat entry; it means it hasn't seen any
1206 : * activity.
1207 : */
1208 4614 : if (!tmp->adw_entry)
1209 64 : continue;
1210 :
1211 : /*
1212 : * Also, skip a database that appears on the database list as having
1213 : * been processed recently (less than autovacuum_naptime seconds ago).
1214 : * We do this so that we don't select a database which we just
1215 : * selected, but that pgstat hasn't gotten around to updating the last
1216 : * autovacuum time yet.
1217 : */
1218 4550 : skipit = false;
1219 :
1220 9220 : dlist_reverse_foreach(iter, &DatabaseList)
1221 : {
1222 9186 : avl_dbase *dbp = dlist_container(avl_dbase, adl_node, iter.cur);
1223 :
1224 9186 : if (dbp->adl_datid == tmp->adw_datid)
1225 : {
1226 : /*
1227 : * Skip this database if its next_worker value falls between
1228 : * the current time and the current time plus naptime.
1229 : */
1230 4516 : if (!TimestampDifferenceExceeds(dbp->adl_next_worker,
1231 2830 : current_time, 0) &&
1232 2830 : !TimestampDifferenceExceeds(current_time,
1233 : dbp->adl_next_worker,
1234 : autovacuum_naptime * 1000))
1235 2830 : skipit = true;
1236 :
1237 4516 : break;
1238 : }
1239 : }
1240 4550 : if (skipit)
1241 2830 : continue;
1242 :
1243 : /*
1244 : * Remember the db with oldest autovac time. (If we are here, both
1245 : * tmp->entry and db->entry must be non-null.)
1246 : */
1247 1720 : if (avdb == NULL ||
1248 242 : tmp->adw_entry->last_autovac_time < avdb->adw_entry->last_autovac_time)
1249 1532 : avdb = tmp;
1250 : }
1251 :
1252 : /* Found a database -- process it */
1253 2374 : if (avdb != NULL)
1254 : {
1255 : WorkerInfo worker;
1256 : dlist_node *wptr;
1257 :
1258 2372 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1259 :
1260 : /*
1261 : * Get a worker entry from the freelist. We checked above, so there
1262 : * really should be a free slot.
1263 : */
1264 2372 : wptr = dclist_pop_head_node(&AutoVacuumShmem->av_freeWorkers);
1265 :
1266 2372 : worker = dlist_container(WorkerInfoData, wi_links, wptr);
1267 2372 : worker->wi_dboid = avdb->adw_datid;
1268 2372 : worker->wi_proc = NULL;
1269 2372 : worker->wi_launchtime = GetCurrentTimestamp();
1270 :
1271 2372 : AutoVacuumShmem->av_startingWorker = worker;
1272 :
1273 2372 : LWLockRelease(AutovacuumLock);
1274 :
1275 2372 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
1276 :
1277 2372 : retval = avdb->adw_datid;
1278 : }
1279 2 : else if (skipit)
1280 : {
1281 : /*
1282 : * If we skipped all databases on the list, rebuild it, because it
1283 : * probably contains a dropped database.
1284 : */
1285 0 : rebuild_database_list(InvalidOid);
1286 : }
1287 :
1288 2374 : MemoryContextSwitchTo(oldcxt);
1289 2374 : MemoryContextDelete(tmpcxt);
1290 :
1291 2374 : return retval;
1292 : }
1293 :
1294 : /*
1295 : * launch_worker
1296 : *
1297 : * Wrapper for starting a worker from the launcher. Besides actually starting
1298 : * it, update the database list to reflect the next time that another one will
1299 : * need to be started on the selected database. The actual database choice is
1300 : * left to do_start_worker.
1301 : *
1302 : * This routine is also expected to insert an entry into the database list if
1303 : * the selected database was previously absent from the list.
1304 : */
1305 : static void
1306 2374 : launch_worker(TimestampTz now)
1307 : {
1308 : Oid dbid;
1309 : dlist_iter iter;
1310 :
1311 2374 : dbid = do_start_worker();
1312 2374 : if (OidIsValid(dbid))
1313 : {
1314 2372 : bool found = false;
1315 :
1316 : /*
1317 : * Walk the database list and update the corresponding entry. If the
1318 : * database is not on the list, we'll recreate the list.
1319 : */
1320 6026 : dlist_foreach(iter, &DatabaseList)
1321 : {
1322 6000 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1323 :
1324 6000 : if (avdb->adl_datid == dbid)
1325 : {
1326 2346 : found = true;
1327 :
1328 : /*
1329 : * add autovacuum_naptime seconds to the current time, and use
1330 : * that as the new "next_worker" field for this database.
1331 : */
1332 2346 : avdb->adl_next_worker =
1333 2346 : TimestampTzPlusMilliseconds(now, autovacuum_naptime * 1000);
1334 :
1335 2346 : dlist_move_head(&DatabaseList, iter.cur);
1336 2346 : break;
1337 : }
1338 : }
1339 :
1340 : /*
1341 : * If the database was not present in the database list, we rebuild
1342 : * the list. It's possible that the database does not get into the
1343 : * list anyway, for example if it's a database that doesn't have a
1344 : * pgstat entry, but this is not a problem because we don't want to
1345 : * schedule workers regularly into those in any case.
1346 : */
1347 2372 : if (!found)
1348 26 : rebuild_database_list(dbid);
1349 : }
1350 2374 : }
1351 :
1352 : /*
1353 : * Called from postmaster to signal a failure to fork a process to become
1354 : * worker. The postmaster should kill(SIGUSR2) the launcher shortly
1355 : * after calling this function.
1356 : */
1357 : void
1358 0 : AutoVacWorkerFailed(void)
1359 : {
1360 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = true;
1361 0 : }
1362 :
1363 : /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */
1364 : static void
1365 4740 : avl_sigusr2_handler(SIGNAL_ARGS)
1366 : {
1367 4740 : got_SIGUSR2 = true;
1368 4740 : SetLatch(MyLatch);
1369 4740 : }
1370 :
1371 :
1372 : /********************************************************************
1373 : * AUTOVACUUM WORKER CODE
1374 : ********************************************************************/
1375 :
1376 : /*
1377 : * Main entry point for autovacuum worker processes.
1378 : */
1379 : void
1380 2374 : AutoVacWorkerMain(const void *startup_data, size_t startup_data_len)
1381 : {
1382 : sigjmp_buf local_sigjmp_buf;
1383 : Oid dbid;
1384 :
1385 : Assert(startup_data_len == 0);
1386 :
1387 : /* Release postmaster's working memory context */
1388 2374 : if (PostmasterContext)
1389 : {
1390 2374 : MemoryContextDelete(PostmasterContext);
1391 2374 : PostmasterContext = NULL;
1392 : }
1393 :
1394 2374 : MyBackendType = B_AUTOVAC_WORKER;
1395 2374 : init_ps_display(NULL);
1396 :
1397 : Assert(GetProcessingMode() == InitProcessing);
1398 :
1399 : /*
1400 : * Set up signal handlers. We operate on databases much like a regular
1401 : * backend, so we use the same signal handling. See equivalent code in
1402 : * tcop/postgres.c.
1403 : */
1404 2374 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1405 :
1406 : /*
1407 : * SIGINT is used to signal canceling the current table's vacuum; SIGTERM
1408 : * means abort and exit cleanly, and SIGQUIT means abandon ship.
1409 : */
1410 2374 : pqsignal(SIGINT, StatementCancelHandler);
1411 2374 : pqsignal(SIGTERM, die);
1412 : /* SIGQUIT handler was already set up by InitPostmasterChild */
1413 :
1414 2374 : InitializeTimeouts(); /* establishes SIGALRM handler */
1415 :
1416 2374 : pqsignal(SIGPIPE, SIG_IGN);
1417 2374 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1418 2374 : pqsignal(SIGUSR2, SIG_IGN);
1419 2374 : pqsignal(SIGFPE, FloatExceptionHandler);
1420 2374 : pqsignal(SIGCHLD, SIG_DFL);
1421 :
1422 : /*
1423 : * Create a per-backend PGPROC struct in shared memory. We must do this
1424 : * before we can use LWLocks or access any shared memory.
1425 : */
1426 2374 : InitProcess();
1427 :
1428 : /* Early initialization */
1429 2374 : BaseInit();
1430 :
1431 : /*
1432 : * If an exception is encountered, processing resumes here.
1433 : *
1434 : * Unlike most auxiliary processes, we don't attempt to continue
1435 : * processing after an error; we just clean up and exit. The autovac
1436 : * launcher is responsible for spawning another worker later.
1437 : *
1438 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
1439 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
1440 : * signals other than SIGQUIT will be blocked until we exit. It might
1441 : * seem that this policy makes the HOLD_INTERRUPTS() call redundant, but
1442 : * it is not since InterruptPending might be set already.
1443 : */
1444 2374 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1445 : {
1446 : /* since not using PG_TRY, must reset error stack by hand */
1447 0 : error_context_stack = NULL;
1448 :
1449 : /* Prevents interrupts while cleaning up */
1450 0 : HOLD_INTERRUPTS();
1451 :
1452 : /* Report the error to the server log */
1453 0 : EmitErrorReport();
1454 :
1455 : /*
1456 : * We can now go away. Note that because we called InitProcess, a
1457 : * callback was registered to do ProcKill, which will clean up
1458 : * necessary state.
1459 : */
1460 0 : proc_exit(0);
1461 : }
1462 :
1463 : /* We can now handle ereport(ERROR) */
1464 2374 : PG_exception_stack = &local_sigjmp_buf;
1465 :
1466 2374 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
1467 :
1468 : /*
1469 : * Set always-secure search path, so malicious users can't redirect user
1470 : * code (e.g. pg_index.indexprs). (That code runs in a
1471 : * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not
1472 : * take control of the entire autovacuum worker in any case.)
1473 : */
1474 2374 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1475 :
1476 : /*
1477 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
1478 : * in postgresql.conf. We don't really want such a dangerous option being
1479 : * applied non-interactively.
1480 : */
1481 2374 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
1482 :
1483 : /*
1484 : * Force settable timeouts off to avoid letting these settings prevent
1485 : * regular maintenance from being executed.
1486 : */
1487 2374 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1488 2374 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1489 2374 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1490 2374 : SetConfigOption("idle_in_transaction_session_timeout", "0",
1491 : PGC_SUSET, PGC_S_OVERRIDE);
1492 :
1493 : /*
1494 : * Force default_transaction_isolation to READ COMMITTED. We don't want
1495 : * to pay the overhead of serializable mode, nor add any risk of causing
1496 : * deadlocks or delaying other transactions.
1497 : */
1498 2374 : SetConfigOption("default_transaction_isolation", "read committed",
1499 : PGC_SUSET, PGC_S_OVERRIDE);
1500 :
1501 : /*
1502 : * Force synchronous replication off to allow regular maintenance even if
1503 : * we are waiting for standbys to connect. This is important to ensure we
1504 : * aren't blocked from performing anti-wraparound tasks.
1505 : */
1506 2374 : if (synchronous_commit > SYNCHRONOUS_COMMIT_LOCAL_FLUSH)
1507 2374 : SetConfigOption("synchronous_commit", "local",
1508 : PGC_SUSET, PGC_S_OVERRIDE);
1509 :
1510 : /*
1511 : * Even when system is configured to use a different fetch consistency,
1512 : * for autovac we always want fresh stats.
1513 : */
1514 2374 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
1515 :
1516 : /*
1517 : * Get the info about the database we're going to work on.
1518 : */
1519 2374 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1520 :
1521 : /*
1522 : * beware of startingWorker being INVALID; this should normally not
1523 : * happen, but if a worker fails after forking and before this, the
1524 : * launcher might have decided to remove it from the queue and start
1525 : * again.
1526 : */
1527 2374 : if (AutoVacuumShmem->av_startingWorker != NULL)
1528 : {
1529 2374 : MyWorkerInfo = AutoVacuumShmem->av_startingWorker;
1530 2374 : dbid = MyWorkerInfo->wi_dboid;
1531 2374 : MyWorkerInfo->wi_proc = MyProc;
1532 :
1533 : /* insert into the running list */
1534 2374 : dlist_push_head(&AutoVacuumShmem->av_runningWorkers,
1535 2374 : &MyWorkerInfo->wi_links);
1536 :
1537 : /*
1538 : * remove from the "starting" pointer, so that the launcher can start
1539 : * a new worker if required
1540 : */
1541 2374 : AutoVacuumShmem->av_startingWorker = NULL;
1542 2374 : LWLockRelease(AutovacuumLock);
1543 :
1544 2374 : on_shmem_exit(FreeWorkerInfo, 0);
1545 :
1546 : /* wake up the launcher */
1547 2374 : if (AutoVacuumShmem->av_launcherpid != 0)
1548 2374 : kill(AutoVacuumShmem->av_launcherpid, SIGUSR2);
1549 : }
1550 : else
1551 : {
1552 : /* no worker entry for me, go away */
1553 0 : elog(WARNING, "autovacuum worker started without a worker entry");
1554 0 : dbid = InvalidOid;
1555 0 : LWLockRelease(AutovacuumLock);
1556 : }
1557 :
1558 2374 : if (OidIsValid(dbid))
1559 : {
1560 : char dbname[NAMEDATALEN];
1561 :
1562 : /*
1563 : * Report autovac startup to the cumulative stats system. We
1564 : * deliberately do this before InitPostgres, so that the
1565 : * last_autovac_time will get updated even if the connection attempt
1566 : * fails. This is to prevent autovac from getting "stuck" repeatedly
1567 : * selecting an unopenable database, rather than making any progress
1568 : * on stuff it can connect to.
1569 : */
1570 2374 : pgstat_report_autovac(dbid);
1571 :
1572 : /*
1573 : * Connect to the selected database, specifying no particular user,
1574 : * and ignoring datallowconn. Collect the database's name for
1575 : * display.
1576 : *
1577 : * Note: if we have selected a just-deleted database (due to using
1578 : * stale stats info), we'll fail and exit here.
1579 : */
1580 2374 : InitPostgres(NULL, dbid, NULL, InvalidOid,
1581 : INIT_PG_OVERRIDE_ALLOW_CONNS,
1582 : dbname);
1583 2374 : SetProcessingMode(NormalProcessing);
1584 2374 : set_ps_display(dbname);
1585 2374 : ereport(DEBUG1,
1586 : (errmsg_internal("autovacuum: processing database \"%s\"", dbname)));
1587 :
1588 2374 : if (PostAuthDelay)
1589 0 : pg_usleep(PostAuthDelay * 1000000L);
1590 :
1591 : /* And do an appropriate amount of work */
1592 2374 : recentXid = ReadNextTransactionId();
1593 2374 : recentMulti = ReadNextMultiXactId();
1594 2374 : do_autovacuum();
1595 : }
1596 :
1597 : /*
1598 : * The launcher will be notified of my death in ProcKill, *if* we managed
1599 : * to get a worker slot at all
1600 : */
1601 :
1602 : /* All done, go away */
1603 2372 : proc_exit(0);
1604 : }
1605 :
1606 : /*
1607 : * Return a WorkerInfo to the free list
1608 : */
1609 : static void
1610 2374 : FreeWorkerInfo(int code, Datum arg)
1611 : {
1612 2374 : if (MyWorkerInfo != NULL)
1613 : {
1614 2374 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1615 :
1616 : /*
1617 : * Wake the launcher up so that he can launch a new worker immediately
1618 : * if required. We only save the launcher's PID in local memory here;
1619 : * the actual signal will be sent when the PGPROC is recycled. Note
1620 : * that we always do this, so that the launcher can rebalance the cost
1621 : * limit setting of the remaining workers.
1622 : *
1623 : * We somewhat ignore the risk that the launcher changes its PID
1624 : * between us reading it and the actual kill; we expect ProcKill to be
1625 : * called shortly after us, and we assume that PIDs are not reused too
1626 : * quickly after a process exits.
1627 : */
1628 2374 : AutovacuumLauncherPid = AutoVacuumShmem->av_launcherpid;
1629 :
1630 2374 : dlist_delete(&MyWorkerInfo->wi_links);
1631 2374 : MyWorkerInfo->wi_dboid = InvalidOid;
1632 2374 : MyWorkerInfo->wi_tableoid = InvalidOid;
1633 2374 : MyWorkerInfo->wi_sharedrel = false;
1634 2374 : MyWorkerInfo->wi_proc = NULL;
1635 2374 : MyWorkerInfo->wi_launchtime = 0;
1636 2374 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
1637 2374 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
1638 2374 : &MyWorkerInfo->wi_links);
1639 : /* not mine anymore */
1640 2374 : MyWorkerInfo = NULL;
1641 :
1642 : /*
1643 : * now that we're inactive, cause a rebalancing of the surviving
1644 : * workers
1645 : */
1646 2374 : AutoVacuumShmem->av_signal[AutoVacRebalance] = true;
1647 2374 : LWLockRelease(AutovacuumLock);
1648 : }
1649 2374 : }
1650 :
1651 : /*
1652 : * Update vacuum cost-based delay-related parameters for autovacuum workers and
1653 : * backends executing VACUUM or ANALYZE using the value of relevant GUCs and
1654 : * global state. This must be called during setup for vacuum and after every
1655 : * config reload to ensure up-to-date values.
1656 : */
1657 : void
1658 221428 : VacuumUpdateCosts(void)
1659 : {
1660 221428 : if (MyWorkerInfo)
1661 : {
1662 207824 : if (av_storage_param_cost_delay >= 0)
1663 0 : vacuum_cost_delay = av_storage_param_cost_delay;
1664 207824 : else if (autovacuum_vac_cost_delay >= 0)
1665 207824 : vacuum_cost_delay = autovacuum_vac_cost_delay;
1666 : else
1667 : /* fall back to VacuumCostDelay */
1668 0 : vacuum_cost_delay = VacuumCostDelay;
1669 :
1670 207824 : AutoVacuumUpdateCostLimit();
1671 : }
1672 : else
1673 : {
1674 : /* Must be explicit VACUUM or ANALYZE */
1675 13604 : vacuum_cost_delay = VacuumCostDelay;
1676 13604 : vacuum_cost_limit = VacuumCostLimit;
1677 : }
1678 :
1679 : /*
1680 : * If configuration changes are allowed to impact VacuumCostActive, make
1681 : * sure it is updated.
1682 : */
1683 221428 : if (VacuumFailsafeActive)
1684 : Assert(!VacuumCostActive);
1685 221428 : else if (vacuum_cost_delay > 0)
1686 207824 : VacuumCostActive = true;
1687 : else
1688 : {
1689 13604 : VacuumCostActive = false;
1690 13604 : VacuumCostBalance = 0;
1691 : }
1692 :
1693 : /*
1694 : * Since the cost logging requires a lock, avoid rendering the log message
1695 : * in case we are using a message level where the log wouldn't be emitted.
1696 : */
1697 221428 : if (MyWorkerInfo && message_level_is_interesting(DEBUG2))
1698 : {
1699 : Oid dboid,
1700 : tableoid;
1701 :
1702 : Assert(!LWLockHeldByMe(AutovacuumLock));
1703 :
1704 0 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1705 0 : dboid = MyWorkerInfo->wi_dboid;
1706 0 : tableoid = MyWorkerInfo->wi_tableoid;
1707 0 : LWLockRelease(AutovacuumLock);
1708 :
1709 0 : elog(DEBUG2,
1710 : "Autovacuum VacuumUpdateCosts(db=%u, rel=%u, dobalance=%s, cost_limit=%d, cost_delay=%g active=%s failsafe=%s)",
1711 : dboid, tableoid, pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance) ? "no" : "yes",
1712 : vacuum_cost_limit, vacuum_cost_delay,
1713 : vacuum_cost_delay > 0 ? "yes" : "no",
1714 : VacuumFailsafeActive ? "yes" : "no");
1715 : }
1716 221428 : }
1717 :
1718 : /*
1719 : * Update vacuum_cost_limit with the correct value for an autovacuum worker,
1720 : * given the value of other relevant cost limit parameters and the number of
1721 : * workers across which the limit must be balanced. Autovacuum workers must
1722 : * call this regularly in case av_nworkersForBalance has been updated by
1723 : * another worker or by the autovacuum launcher. They must also call it after a
1724 : * config reload.
1725 : */
1726 : void
1727 211532 : AutoVacuumUpdateCostLimit(void)
1728 : {
1729 211532 : if (!MyWorkerInfo)
1730 0 : return;
1731 :
1732 : /*
1733 : * note: in cost_limit, zero also means use value from elsewhere, because
1734 : * zero is not a valid value.
1735 : */
1736 :
1737 211532 : if (av_storage_param_cost_limit > 0)
1738 0 : vacuum_cost_limit = av_storage_param_cost_limit;
1739 : else
1740 : {
1741 : int nworkers_for_balance;
1742 :
1743 211532 : if (autovacuum_vac_cost_limit > 0)
1744 0 : vacuum_cost_limit = autovacuum_vac_cost_limit;
1745 : else
1746 211532 : vacuum_cost_limit = VacuumCostLimit;
1747 :
1748 : /* Only balance limit if no cost-related storage parameters specified */
1749 211532 : if (pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance))
1750 0 : return;
1751 :
1752 : Assert(vacuum_cost_limit > 0);
1753 :
1754 211532 : nworkers_for_balance = pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1755 :
1756 : /* There is at least 1 autovac worker (this worker) */
1757 211532 : if (nworkers_for_balance <= 0)
1758 0 : elog(ERROR, "nworkers_for_balance must be > 0");
1759 :
1760 211532 : vacuum_cost_limit = Max(vacuum_cost_limit / nworkers_for_balance, 1);
1761 : }
1762 : }
1763 :
1764 : /*
1765 : * autovac_recalculate_workers_for_balance
1766 : * Recalculate the number of workers to consider, given cost-related
1767 : * storage parameters and the current number of active workers.
1768 : *
1769 : * Caller must hold the AutovacuumLock in at least shared mode to access
1770 : * worker->wi_proc.
1771 : */
1772 : static void
1773 106280 : autovac_recalculate_workers_for_balance(void)
1774 : {
1775 : dlist_iter iter;
1776 : int orig_nworkers_for_balance;
1777 106280 : int nworkers_for_balance = 0;
1778 :
1779 : Assert(LWLockHeldByMe(AutovacuumLock));
1780 :
1781 106280 : orig_nworkers_for_balance =
1782 106280 : pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1783 :
1784 228362 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1785 : {
1786 122082 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1787 :
1788 244164 : if (worker->wi_proc == NULL ||
1789 122082 : pg_atomic_unlocked_test_flag(&worker->wi_dobalance))
1790 990 : continue;
1791 :
1792 121092 : nworkers_for_balance++;
1793 : }
1794 :
1795 106280 : if (nworkers_for_balance != orig_nworkers_for_balance)
1796 2006 : pg_atomic_write_u32(&AutoVacuumShmem->av_nworkersForBalance,
1797 : nworkers_for_balance);
1798 106280 : }
1799 :
1800 : /*
1801 : * get_database_list
1802 : * Return a list of all databases found in pg_database.
1803 : *
1804 : * The list and associated data is allocated in the caller's memory context,
1805 : * which is in charge of ensuring that it's properly cleaned up afterwards.
1806 : *
1807 : * Note: this is the only function in which the autovacuum launcher uses a
1808 : * transaction. Although we aren't attached to any particular database and
1809 : * therefore can't access most catalogs, we do have enough infrastructure
1810 : * to do a seqscan on pg_database.
1811 : */
1812 : static List *
1813 3404 : get_database_list(void)
1814 : {
1815 3404 : List *dblist = NIL;
1816 : Relation rel;
1817 : TableScanDesc scan;
1818 : HeapTuple tup;
1819 : MemoryContext resultcxt;
1820 :
1821 : /* This is the context that we will allocate our output data in */
1822 3404 : resultcxt = CurrentMemoryContext;
1823 :
1824 : /*
1825 : * Start a transaction so we can access pg_database.
1826 : */
1827 3404 : StartTransactionCommand();
1828 :
1829 3404 : rel = table_open(DatabaseRelationId, AccessShareLock);
1830 3404 : scan = table_beginscan_catalog(rel, 0, NULL);
1831 :
1832 13976 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1833 : {
1834 10572 : Form_pg_database pgdatabase = (Form_pg_database) GETSTRUCT(tup);
1835 : avw_dbase *avdb;
1836 : MemoryContext oldcxt;
1837 :
1838 : /*
1839 : * If database has partially been dropped, we can't, nor need to,
1840 : * vacuum it.
1841 : */
1842 10572 : if (database_is_invalid_form(pgdatabase))
1843 : {
1844 0 : elog(DEBUG2,
1845 : "autovacuum: skipping invalid database \"%s\"",
1846 : NameStr(pgdatabase->datname));
1847 0 : continue;
1848 : }
1849 :
1850 : /*
1851 : * Allocate our results in the caller's context, not the
1852 : * transaction's. We do this inside the loop, and restore the original
1853 : * context at the end, so that leaky things like heap_getnext() are
1854 : * not called in a potentially long-lived context.
1855 : */
1856 10572 : oldcxt = MemoryContextSwitchTo(resultcxt);
1857 :
1858 10572 : avdb = (avw_dbase *) palloc(sizeof(avw_dbase));
1859 :
1860 10572 : avdb->adw_datid = pgdatabase->oid;
1861 10572 : avdb->adw_name = pstrdup(NameStr(pgdatabase->datname));
1862 10572 : avdb->adw_frozenxid = pgdatabase->datfrozenxid;
1863 10572 : avdb->adw_minmulti = pgdatabase->datminmxid;
1864 : /* this gets set later: */
1865 10572 : avdb->adw_entry = NULL;
1866 :
1867 10572 : dblist = lappend(dblist, avdb);
1868 10572 : MemoryContextSwitchTo(oldcxt);
1869 : }
1870 :
1871 3404 : table_endscan(scan);
1872 3404 : table_close(rel, AccessShareLock);
1873 :
1874 3404 : CommitTransactionCommand();
1875 :
1876 : /* Be sure to restore caller's memory context */
1877 3404 : MemoryContextSwitchTo(resultcxt);
1878 :
1879 3404 : return dblist;
1880 : }
1881 :
1882 : /*
1883 : * Process a database table-by-table
1884 : *
1885 : * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in
1886 : * order not to ignore shutdown commands for too long.
1887 : */
1888 : static void
1889 2374 : do_autovacuum(void)
1890 : {
1891 : Relation classRel;
1892 : HeapTuple tuple;
1893 : TableScanDesc relScan;
1894 : Form_pg_database dbForm;
1895 2374 : List *table_oids = NIL;
1896 2374 : List *orphan_oids = NIL;
1897 : HASHCTL ctl;
1898 : HTAB *table_toast_map;
1899 : ListCell *volatile cell;
1900 : BufferAccessStrategy bstrategy;
1901 : ScanKeyData key;
1902 : TupleDesc pg_class_desc;
1903 : int effective_multixact_freeze_max_age;
1904 2374 : bool did_vacuum = false;
1905 2374 : bool found_concurrent_worker = false;
1906 : int i;
1907 :
1908 : /*
1909 : * StartTransactionCommand and CommitTransactionCommand will automatically
1910 : * switch to other contexts. We need this one to keep the list of
1911 : * relations to vacuum/analyze across transactions.
1912 : */
1913 2374 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
1914 : "Autovacuum worker",
1915 : ALLOCSET_DEFAULT_SIZES);
1916 2374 : MemoryContextSwitchTo(AutovacMemCxt);
1917 :
1918 : /* Start a transaction so our commands have one to play into. */
1919 2374 : StartTransactionCommand();
1920 :
1921 : /*
1922 : * This injection point is put in a transaction block to work with a wait
1923 : * that uses a condition variable.
1924 : */
1925 2374 : INJECTION_POINT("autovacuum-worker-start");
1926 :
1927 : /*
1928 : * Compute the multixact age for which freezing is urgent. This is
1929 : * normally autovacuum_multixact_freeze_max_age, but may be less if we are
1930 : * short of multixact member space.
1931 : */
1932 2372 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1933 :
1934 : /*
1935 : * Find the pg_database entry and select the default freeze ages. We use
1936 : * zero in template and nonconnectable databases, else the system-wide
1937 : * default.
1938 : */
1939 2372 : tuple = SearchSysCache1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
1940 2372 : if (!HeapTupleIsValid(tuple))
1941 0 : elog(ERROR, "cache lookup failed for database %u", MyDatabaseId);
1942 2372 : dbForm = (Form_pg_database) GETSTRUCT(tuple);
1943 :
1944 2372 : if (dbForm->datistemplate || !dbForm->datallowconn)
1945 : {
1946 1148 : default_freeze_min_age = 0;
1947 1148 : default_freeze_table_age = 0;
1948 1148 : default_multixact_freeze_min_age = 0;
1949 1148 : default_multixact_freeze_table_age = 0;
1950 : }
1951 : else
1952 : {
1953 1224 : default_freeze_min_age = vacuum_freeze_min_age;
1954 1224 : default_freeze_table_age = vacuum_freeze_table_age;
1955 1224 : default_multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1956 1224 : default_multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1957 : }
1958 :
1959 2372 : ReleaseSysCache(tuple);
1960 :
1961 : /* StartTransactionCommand changed elsewhere */
1962 2372 : MemoryContextSwitchTo(AutovacMemCxt);
1963 :
1964 2372 : classRel = table_open(RelationRelationId, AccessShareLock);
1965 :
1966 : /* create a copy so we can use it after closing pg_class */
1967 2372 : pg_class_desc = CreateTupleDescCopy(RelationGetDescr(classRel));
1968 :
1969 : /* create hash table for toast <-> main relid mapping */
1970 2372 : ctl.keysize = sizeof(Oid);
1971 2372 : ctl.entrysize = sizeof(av_relation);
1972 :
1973 2372 : table_toast_map = hash_create("TOAST to main relid map",
1974 : 100,
1975 : &ctl,
1976 : HASH_ELEM | HASH_BLOBS);
1977 :
1978 : /*
1979 : * Scan pg_class to determine which tables to vacuum.
1980 : *
1981 : * We do this in two passes: on the first one we collect the list of plain
1982 : * relations and materialized views, and on the second one we collect
1983 : * TOAST tables. The reason for doing the second pass is that during it we
1984 : * want to use the main relation's pg_class.reloptions entry if the TOAST
1985 : * table does not have any, and we cannot obtain it unless we know
1986 : * beforehand what's the main table OID.
1987 : *
1988 : * We need to check TOAST tables separately because in cases with short,
1989 : * wide tables there might be proportionally much more activity in the
1990 : * TOAST table than in its parent.
1991 : */
1992 2372 : relScan = table_beginscan_catalog(classRel, 0, NULL);
1993 :
1994 : /*
1995 : * On the first pass, we collect main tables to vacuum, and also the main
1996 : * table relid to TOAST relid mapping.
1997 : */
1998 1000960 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
1999 : {
2000 998588 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2001 : PgStat_StatTabEntry *tabentry;
2002 : AutoVacOpts *relopts;
2003 : Oid relid;
2004 : bool dovacuum;
2005 : bool doanalyze;
2006 : bool wraparound;
2007 :
2008 998588 : if (classForm->relkind != RELKIND_RELATION &&
2009 834588 : classForm->relkind != RELKIND_MATVIEW)
2010 834592 : continue;
2011 :
2012 164014 : relid = classForm->oid;
2013 :
2014 : /*
2015 : * Check if it is a temp table (presumably, of some other backend's).
2016 : * We cannot safely process other backends' temp tables.
2017 : */
2018 164014 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2019 : {
2020 : /*
2021 : * We just ignore it if the owning backend is still active and
2022 : * using the temporary schema. Also, for safety, ignore it if the
2023 : * namespace doesn't exist or isn't a temp namespace after all.
2024 : */
2025 18 : if (checkTempNamespaceStatus(classForm->relnamespace) == TEMP_NAMESPACE_IDLE)
2026 : {
2027 : /*
2028 : * The table seems to be orphaned -- although it might be that
2029 : * the owning backend has already deleted it and exited; our
2030 : * pg_class scan snapshot is not necessarily up-to-date
2031 : * anymore, so we could be looking at a committed-dead entry.
2032 : * Remember it so we can try to delete it later.
2033 : */
2034 0 : orphan_oids = lappend_oid(orphan_oids, relid);
2035 : }
2036 18 : continue;
2037 : }
2038 :
2039 : /* Fetch reloptions and the pgstat entry for this table */
2040 163996 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2041 163996 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2042 : relid);
2043 :
2044 : /* Check if it needs vacuum or analyze */
2045 163996 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2046 : effective_multixact_freeze_max_age,
2047 : &dovacuum, &doanalyze, &wraparound);
2048 :
2049 : /* Relations that need work are added to table_oids */
2050 163996 : if (dovacuum || doanalyze)
2051 65258 : table_oids = lappend_oid(table_oids, relid);
2052 :
2053 : /*
2054 : * Remember TOAST associations for the second pass. Note: we must do
2055 : * this whether or not the table is going to be vacuumed, because we
2056 : * don't automatically vacuum toast tables along the parent table.
2057 : */
2058 163996 : if (OidIsValid(classForm->reltoastrelid))
2059 : {
2060 : av_relation *hentry;
2061 : bool found;
2062 :
2063 194068 : hentry = hash_search(table_toast_map,
2064 97034 : &classForm->reltoastrelid,
2065 : HASH_ENTER, &found);
2066 :
2067 97034 : if (!found)
2068 : {
2069 : /* hash_search already filled in the key */
2070 97034 : hentry->ar_relid = relid;
2071 97034 : hentry->ar_hasrelopts = false;
2072 97034 : if (relopts != NULL)
2073 : {
2074 1874 : hentry->ar_hasrelopts = true;
2075 1874 : memcpy(&hentry->ar_reloptions, relopts,
2076 : sizeof(AutoVacOpts));
2077 : }
2078 : }
2079 : }
2080 : }
2081 :
2082 2372 : table_endscan(relScan);
2083 :
2084 : /* second pass: check TOAST tables */
2085 2372 : ScanKeyInit(&key,
2086 : Anum_pg_class_relkind,
2087 : BTEqualStrategyNumber, F_CHAREQ,
2088 : CharGetDatum(RELKIND_TOASTVALUE));
2089 :
2090 2372 : relScan = table_beginscan_catalog(classRel, 1, &key);
2091 99410 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2092 : {
2093 97038 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2094 : PgStat_StatTabEntry *tabentry;
2095 : Oid relid;
2096 97038 : AutoVacOpts *relopts = NULL;
2097 : bool dovacuum;
2098 : bool doanalyze;
2099 : bool wraparound;
2100 :
2101 : /*
2102 : * We cannot safely process other backends' temp tables, so skip 'em.
2103 : */
2104 97038 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2105 4 : continue;
2106 :
2107 97034 : relid = classForm->oid;
2108 :
2109 : /*
2110 : * fetch reloptions -- if this toast table does not have them, try the
2111 : * main rel
2112 : */
2113 97034 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2114 97034 : if (relopts == NULL)
2115 : {
2116 : av_relation *hentry;
2117 : bool found;
2118 :
2119 97034 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2120 97034 : if (found && hentry->ar_hasrelopts)
2121 1874 : relopts = &hentry->ar_reloptions;
2122 : }
2123 :
2124 : /* Fetch the pgstat entry for this table */
2125 97034 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2126 : relid);
2127 :
2128 97034 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2129 : effective_multixact_freeze_max_age,
2130 : &dovacuum, &doanalyze, &wraparound);
2131 :
2132 : /* ignore analyze for toast tables */
2133 97034 : if (dovacuum)
2134 38702 : table_oids = lappend_oid(table_oids, relid);
2135 : }
2136 :
2137 2372 : table_endscan(relScan);
2138 2372 : table_close(classRel, AccessShareLock);
2139 :
2140 : /*
2141 : * Recheck orphan temporary tables, and if they still seem orphaned, drop
2142 : * them. We'll eat a transaction per dropped table, which might seem
2143 : * excessive, but we should only need to do anything as a result of a
2144 : * previous backend crash, so this should not happen often enough to
2145 : * justify "optimizing". Using separate transactions ensures that we
2146 : * don't bloat the lock table if there are many temp tables to be dropped,
2147 : * and it ensures that we don't lose work if a deletion attempt fails.
2148 : */
2149 2372 : foreach(cell, orphan_oids)
2150 : {
2151 0 : Oid relid = lfirst_oid(cell);
2152 : Form_pg_class classForm;
2153 : ObjectAddress object;
2154 :
2155 : /*
2156 : * Check for user-requested abort.
2157 : */
2158 0 : CHECK_FOR_INTERRUPTS();
2159 :
2160 : /*
2161 : * Try to lock the table. If we can't get the lock immediately,
2162 : * somebody else is using (or dropping) the table, so it's not our
2163 : * concern anymore. Having the lock prevents race conditions below.
2164 : */
2165 0 : if (!ConditionalLockRelationOid(relid, AccessExclusiveLock))
2166 0 : continue;
2167 :
2168 : /*
2169 : * Re-fetch the pg_class tuple and re-check whether it still seems to
2170 : * be an orphaned temp table. If it's not there or no longer the same
2171 : * relation, ignore it.
2172 : */
2173 0 : tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2174 0 : if (!HeapTupleIsValid(tuple))
2175 : {
2176 : /* be sure to drop useless lock so we don't bloat lock table */
2177 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2178 0 : continue;
2179 : }
2180 0 : classForm = (Form_pg_class) GETSTRUCT(tuple);
2181 :
2182 : /*
2183 : * Make all the same tests made in the loop above. In event of OID
2184 : * counter wraparound, the pg_class entry we have now might be
2185 : * completely unrelated to the one we saw before.
2186 : */
2187 0 : if (!((classForm->relkind == RELKIND_RELATION ||
2188 0 : classForm->relkind == RELKIND_MATVIEW) &&
2189 0 : classForm->relpersistence == RELPERSISTENCE_TEMP))
2190 : {
2191 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2192 0 : continue;
2193 : }
2194 :
2195 0 : if (checkTempNamespaceStatus(classForm->relnamespace) != TEMP_NAMESPACE_IDLE)
2196 : {
2197 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2198 0 : continue;
2199 : }
2200 :
2201 : /*
2202 : * Try to lock the temp namespace, too. Even though we have lock on
2203 : * the table itself, there's a risk of deadlock against an incoming
2204 : * backend trying to clean out the temp namespace, in case this table
2205 : * has dependencies (such as sequences) that the backend's
2206 : * performDeletion call might visit in a different order. If we can
2207 : * get AccessShareLock on the namespace, that's sufficient to ensure
2208 : * we're not running concurrently with RemoveTempRelations. If we
2209 : * can't, back off and let RemoveTempRelations do its thing.
2210 : */
2211 0 : if (!ConditionalLockDatabaseObject(NamespaceRelationId,
2212 : classForm->relnamespace, 0,
2213 : AccessShareLock))
2214 : {
2215 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2216 0 : continue;
2217 : }
2218 :
2219 : /* OK, let's delete it */
2220 0 : ereport(LOG,
2221 : (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"",
2222 : get_database_name(MyDatabaseId),
2223 : get_namespace_name(classForm->relnamespace),
2224 : NameStr(classForm->relname))));
2225 :
2226 0 : object.classId = RelationRelationId;
2227 0 : object.objectId = relid;
2228 0 : object.objectSubId = 0;
2229 0 : performDeletion(&object, DROP_CASCADE,
2230 : PERFORM_DELETION_INTERNAL |
2231 : PERFORM_DELETION_QUIETLY |
2232 : PERFORM_DELETION_SKIP_EXTENSIONS);
2233 :
2234 : /*
2235 : * To commit the deletion, end current transaction and start a new
2236 : * one. Note this also releases the locks we took.
2237 : */
2238 0 : CommitTransactionCommand();
2239 0 : StartTransactionCommand();
2240 :
2241 : /* StartTransactionCommand changed current memory context */
2242 0 : MemoryContextSwitchTo(AutovacMemCxt);
2243 : }
2244 :
2245 : /*
2246 : * Optionally, create a buffer access strategy object for VACUUM to use.
2247 : * We use the same BufferAccessStrategy object for all tables VACUUMed by
2248 : * this worker to prevent autovacuum from blowing out shared buffers.
2249 : *
2250 : * VacuumBufferUsageLimit being set to 0 results in
2251 : * GetAccessStrategyWithSize returning NULL, effectively meaning we can
2252 : * use up to all of shared buffers.
2253 : *
2254 : * If we later enter failsafe mode on any of the tables being vacuumed, we
2255 : * will cease use of the BufferAccessStrategy only for that table.
2256 : *
2257 : * XXX should we consider adding code to adjust the size of this if
2258 : * VacuumBufferUsageLimit changes?
2259 : */
2260 2372 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, VacuumBufferUsageLimit);
2261 :
2262 : /*
2263 : * create a memory context to act as fake PortalContext, so that the
2264 : * contexts created in the vacuum code are cleaned up for each table.
2265 : */
2266 2372 : PortalContext = AllocSetContextCreate(AutovacMemCxt,
2267 : "Autovacuum Portal",
2268 : ALLOCSET_DEFAULT_SIZES);
2269 :
2270 : /*
2271 : * Perform operations on collected tables.
2272 : */
2273 106332 : foreach(cell, table_oids)
2274 : {
2275 103960 : Oid relid = lfirst_oid(cell);
2276 : HeapTuple classTup;
2277 : autovac_table *tab;
2278 : bool isshared;
2279 : bool skipit;
2280 : dlist_iter iter;
2281 :
2282 103960 : CHECK_FOR_INTERRUPTS();
2283 :
2284 : /*
2285 : * Check for config changes before processing each collected table.
2286 : */
2287 103960 : if (ConfigReloadPending)
2288 : {
2289 0 : ConfigReloadPending = false;
2290 0 : ProcessConfigFile(PGC_SIGHUP);
2291 :
2292 : /*
2293 : * You might be tempted to bail out if we see autovacuum is now
2294 : * disabled. Must resist that temptation -- this might be a
2295 : * for-wraparound emergency worker, in which case that would be
2296 : * entirely inappropriate.
2297 : */
2298 : }
2299 :
2300 : /*
2301 : * Find out whether the table is shared or not. (It's slightly
2302 : * annoying to fetch the syscache entry just for this, but in typical
2303 : * cases it adds little cost because table_recheck_autovac would
2304 : * refetch the entry anyway. We could buy that back by copying the
2305 : * tuple here and passing it to table_recheck_autovac, but that
2306 : * increases the odds of that function working with stale data.)
2307 : */
2308 103960 : classTup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
2309 103960 : if (!HeapTupleIsValid(classTup))
2310 48 : continue; /* somebody deleted the rel, forget it */
2311 103960 : isshared = ((Form_pg_class) GETSTRUCT(classTup))->relisshared;
2312 103960 : ReleaseSysCache(classTup);
2313 :
2314 : /*
2315 : * Hold schedule lock from here until we've claimed the table. We
2316 : * also need the AutovacuumLock to walk the worker array, but that one
2317 : * can just be a shared lock.
2318 : */
2319 103960 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2320 103960 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2321 :
2322 : /*
2323 : * Check whether the table is being vacuumed concurrently by another
2324 : * worker.
2325 : */
2326 103960 : skipit = false;
2327 225834 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
2328 : {
2329 121922 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
2330 :
2331 : /* ignore myself */
2332 121922 : if (worker == MyWorkerInfo)
2333 103938 : continue;
2334 :
2335 : /* ignore workers in other databases (unless table is shared) */
2336 17984 : if (!worker->wi_sharedrel && worker->wi_dboid != MyDatabaseId)
2337 0 : continue;
2338 :
2339 17984 : if (worker->wi_tableoid == relid)
2340 : {
2341 48 : skipit = true;
2342 48 : found_concurrent_worker = true;
2343 48 : break;
2344 : }
2345 : }
2346 103960 : LWLockRelease(AutovacuumLock);
2347 103960 : if (skipit)
2348 : {
2349 48 : LWLockRelease(AutovacuumScheduleLock);
2350 48 : continue;
2351 : }
2352 :
2353 : /*
2354 : * Store the table's OID in shared memory before releasing the
2355 : * schedule lock, so that other workers don't try to vacuum it
2356 : * concurrently. (We claim it here so as not to hold
2357 : * AutovacuumScheduleLock while rechecking the stats.)
2358 : */
2359 103912 : MyWorkerInfo->wi_tableoid = relid;
2360 103912 : MyWorkerInfo->wi_sharedrel = isshared;
2361 103912 : LWLockRelease(AutovacuumScheduleLock);
2362 :
2363 : /*
2364 : * Check whether pgstat data still says we need to vacuum this table.
2365 : * It could have changed if something else processed the table while
2366 : * we weren't looking. This doesn't entirely close the race condition,
2367 : * but it is very small.
2368 : */
2369 103912 : MemoryContextSwitchTo(AutovacMemCxt);
2370 103912 : tab = table_recheck_autovac(relid, table_toast_map, pg_class_desc,
2371 : effective_multixact_freeze_max_age);
2372 103912 : if (tab == NULL)
2373 : {
2374 : /* someone else vacuumed the table, or it went away */
2375 0 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2376 0 : MyWorkerInfo->wi_tableoid = InvalidOid;
2377 0 : MyWorkerInfo->wi_sharedrel = false;
2378 0 : LWLockRelease(AutovacuumScheduleLock);
2379 0 : continue;
2380 : }
2381 :
2382 : /*
2383 : * Save the cost-related storage parameter values in global variables
2384 : * for reference when updating vacuum_cost_delay and vacuum_cost_limit
2385 : * during vacuuming this table.
2386 : */
2387 103912 : av_storage_param_cost_delay = tab->at_storage_param_vac_cost_delay;
2388 103912 : av_storage_param_cost_limit = tab->at_storage_param_vac_cost_limit;
2389 :
2390 : /*
2391 : * We only expect this worker to ever set the flag, so don't bother
2392 : * checking the return value. We shouldn't have to retry.
2393 : */
2394 103912 : if (tab->at_dobalance)
2395 103912 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2396 : else
2397 0 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
2398 :
2399 103912 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2400 103912 : autovac_recalculate_workers_for_balance();
2401 103912 : LWLockRelease(AutovacuumLock);
2402 :
2403 : /*
2404 : * We wait until this point to update cost delay and cost limit
2405 : * values, even though we reloaded the configuration file above, so
2406 : * that we can take into account the cost-related storage parameters.
2407 : */
2408 103912 : VacuumUpdateCosts();
2409 :
2410 :
2411 : /* clean up memory before each iteration */
2412 103912 : MemoryContextReset(PortalContext);
2413 :
2414 : /*
2415 : * Save the relation name for a possible error message, to avoid a
2416 : * catalog lookup in case of an error. If any of these return NULL,
2417 : * then the relation has been dropped since last we checked; skip it.
2418 : * Note: they must live in a long-lived memory context because we call
2419 : * vacuum and analyze in different transactions.
2420 : */
2421 :
2422 103912 : tab->at_relname = get_rel_name(tab->at_relid);
2423 103912 : tab->at_nspname = get_namespace_name(get_rel_namespace(tab->at_relid));
2424 103912 : tab->at_datname = get_database_name(MyDatabaseId);
2425 103912 : if (!tab->at_relname || !tab->at_nspname || !tab->at_datname)
2426 0 : goto deleted;
2427 :
2428 : /*
2429 : * We will abort vacuuming the current table if something errors out,
2430 : * and continue with the next one in schedule; in particular, this
2431 : * happens if we are interrupted with SIGINT.
2432 : */
2433 103912 : PG_TRY();
2434 : {
2435 : /* Use PortalContext for any per-table allocations */
2436 103912 : MemoryContextSwitchTo(PortalContext);
2437 :
2438 : /* have at it */
2439 103912 : autovacuum_do_vac_analyze(tab, bstrategy);
2440 :
2441 : /*
2442 : * Clear a possible query-cancel signal, to avoid a late reaction
2443 : * to an automatically-sent signal because of vacuuming the
2444 : * current table (we're done with it, so it would make no sense to
2445 : * cancel at this point.)
2446 : */
2447 103912 : QueryCancelPending = false;
2448 : }
2449 0 : PG_CATCH();
2450 : {
2451 : /*
2452 : * Abort the transaction, start a new one, and proceed with the
2453 : * next table in our list.
2454 : */
2455 0 : HOLD_INTERRUPTS();
2456 0 : if (tab->at_params.options & VACOPT_VACUUM)
2457 0 : errcontext("automatic vacuum of table \"%s.%s.%s\"",
2458 : tab->at_datname, tab->at_nspname, tab->at_relname);
2459 : else
2460 0 : errcontext("automatic analyze of table \"%s.%s.%s\"",
2461 : tab->at_datname, tab->at_nspname, tab->at_relname);
2462 0 : EmitErrorReport();
2463 :
2464 : /* this resets ProcGlobal->statusFlags[i] too */
2465 0 : AbortOutOfAnyTransaction();
2466 0 : FlushErrorState();
2467 0 : MemoryContextReset(PortalContext);
2468 :
2469 : /* restart our transaction for the following operations */
2470 0 : StartTransactionCommand();
2471 0 : RESUME_INTERRUPTS();
2472 : }
2473 103912 : PG_END_TRY();
2474 :
2475 : /* Make sure we're back in AutovacMemCxt */
2476 103912 : MemoryContextSwitchTo(AutovacMemCxt);
2477 :
2478 103912 : did_vacuum = true;
2479 :
2480 : /* ProcGlobal->statusFlags[i] are reset at the next end of xact */
2481 :
2482 : /* be tidy */
2483 103912 : deleted:
2484 103912 : if (tab->at_datname != NULL)
2485 103912 : pfree(tab->at_datname);
2486 103912 : if (tab->at_nspname != NULL)
2487 103912 : pfree(tab->at_nspname);
2488 103912 : if (tab->at_relname != NULL)
2489 103912 : pfree(tab->at_relname);
2490 103912 : pfree(tab);
2491 :
2492 : /*
2493 : * Remove my info from shared memory. We set wi_dobalance on the
2494 : * assumption that we are more likely than not to vacuum a table with
2495 : * no cost-related storage parameters next, so we want to claim our
2496 : * share of I/O as soon as possible to avoid thrashing the global
2497 : * balance.
2498 : */
2499 103912 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2500 103912 : MyWorkerInfo->wi_tableoid = InvalidOid;
2501 103912 : MyWorkerInfo->wi_sharedrel = false;
2502 103912 : LWLockRelease(AutovacuumScheduleLock);
2503 103912 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2504 : }
2505 :
2506 : /*
2507 : * Perform additional work items, as requested by backends.
2508 : */
2509 2372 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2510 609604 : for (i = 0; i < NUM_WORKITEMS; i++)
2511 : {
2512 607232 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
2513 :
2514 607232 : if (!workitem->avw_used)
2515 607226 : continue;
2516 6 : if (workitem->avw_active)
2517 0 : continue;
2518 6 : if (workitem->avw_database != MyDatabaseId)
2519 0 : continue;
2520 :
2521 : /* claim this one, and release lock while performing it */
2522 6 : workitem->avw_active = true;
2523 6 : LWLockRelease(AutovacuumLock);
2524 :
2525 6 : perform_work_item(workitem);
2526 :
2527 : /*
2528 : * Check for config changes before acquiring lock for further jobs.
2529 : */
2530 6 : CHECK_FOR_INTERRUPTS();
2531 6 : if (ConfigReloadPending)
2532 : {
2533 0 : ConfigReloadPending = false;
2534 0 : ProcessConfigFile(PGC_SIGHUP);
2535 0 : VacuumUpdateCosts();
2536 : }
2537 :
2538 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2539 :
2540 : /* and mark it done */
2541 6 : workitem->avw_active = false;
2542 6 : workitem->avw_used = false;
2543 : }
2544 2372 : LWLockRelease(AutovacuumLock);
2545 :
2546 : /*
2547 : * We leak table_toast_map here (among other things), but since we're
2548 : * going away soon, it's not a problem.
2549 : */
2550 :
2551 : /*
2552 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We
2553 : * only need to do this once, not after each table.
2554 : *
2555 : * Even if we didn't vacuum anything, it may still be important to do
2556 : * this, because one indirect effect of vac_update_datfrozenxid() is to
2557 : * update TransamVariables->xidVacLimit. That might need to be done even
2558 : * if we haven't vacuumed anything, because relations with older
2559 : * relfrozenxid values or other databases with older datfrozenxid values
2560 : * might have been dropped, allowing xidVacLimit to advance.
2561 : *
2562 : * However, it's also important not to do this blindly in all cases,
2563 : * because when autovacuum=off this will restart the autovacuum launcher.
2564 : * If we're not careful, an infinite loop can result, where workers find
2565 : * no work to do and restart the launcher, which starts another worker in
2566 : * the same database that finds no work to do. To prevent that, we skip
2567 : * this if (1) we found no work to do and (2) we skipped at least one
2568 : * table due to concurrent autovacuum activity. In that case, the other
2569 : * worker has already done it, or will do so when it finishes.
2570 : */
2571 2372 : if (did_vacuum || !found_concurrent_worker)
2572 2372 : vac_update_datfrozenxid();
2573 :
2574 : /* Finally close out the last transaction. */
2575 2372 : CommitTransactionCommand();
2576 2372 : }
2577 :
2578 : /*
2579 : * Execute a previously registered work item.
2580 : */
2581 : static void
2582 6 : perform_work_item(AutoVacuumWorkItem *workitem)
2583 : {
2584 6 : char *cur_datname = NULL;
2585 6 : char *cur_nspname = NULL;
2586 6 : char *cur_relname = NULL;
2587 :
2588 : /*
2589 : * Note we do not store table info in MyWorkerInfo, since this is not
2590 : * vacuuming proper.
2591 : */
2592 :
2593 : /*
2594 : * Save the relation name for a possible error message, to avoid a catalog
2595 : * lookup in case of an error. If any of these return NULL, then the
2596 : * relation has been dropped since last we checked; skip it.
2597 : */
2598 : Assert(CurrentMemoryContext == AutovacMemCxt);
2599 :
2600 6 : cur_relname = get_rel_name(workitem->avw_relation);
2601 6 : cur_nspname = get_namespace_name(get_rel_namespace(workitem->avw_relation));
2602 6 : cur_datname = get_database_name(MyDatabaseId);
2603 6 : if (!cur_relname || !cur_nspname || !cur_datname)
2604 0 : goto deleted2;
2605 :
2606 6 : autovac_report_workitem(workitem, cur_nspname, cur_relname);
2607 :
2608 : /* clean up memory before each work item */
2609 6 : MemoryContextReset(PortalContext);
2610 :
2611 : /*
2612 : * We will abort the current work item if something errors out, and
2613 : * continue with the next one; in particular, this happens if we are
2614 : * interrupted with SIGINT. Note that this means that the work item list
2615 : * can be lossy.
2616 : */
2617 6 : PG_TRY();
2618 : {
2619 : /* Use PortalContext for any per-work-item allocations */
2620 6 : MemoryContextSwitchTo(PortalContext);
2621 :
2622 : /*
2623 : * Have at it. Functions called here are responsible for any required
2624 : * user switch and sandbox.
2625 : */
2626 6 : switch (workitem->avw_type)
2627 : {
2628 6 : case AVW_BRINSummarizeRange:
2629 6 : DirectFunctionCall2(brin_summarize_range,
2630 : ObjectIdGetDatum(workitem->avw_relation),
2631 : Int64GetDatum((int64) workitem->avw_blockNumber));
2632 6 : break;
2633 0 : default:
2634 0 : elog(WARNING, "unrecognized work item found: type %d",
2635 : workitem->avw_type);
2636 0 : break;
2637 : }
2638 :
2639 : /*
2640 : * Clear a possible query-cancel signal, to avoid a late reaction to
2641 : * an automatically-sent signal because of vacuuming the current table
2642 : * (we're done with it, so it would make no sense to cancel at this
2643 : * point.)
2644 : */
2645 6 : QueryCancelPending = false;
2646 : }
2647 0 : PG_CATCH();
2648 : {
2649 : /*
2650 : * Abort the transaction, start a new one, and proceed with the next
2651 : * table in our list.
2652 : */
2653 0 : HOLD_INTERRUPTS();
2654 0 : errcontext("processing work entry for relation \"%s.%s.%s\"",
2655 : cur_datname, cur_nspname, cur_relname);
2656 0 : EmitErrorReport();
2657 :
2658 : /* this resets ProcGlobal->statusFlags[i] too */
2659 0 : AbortOutOfAnyTransaction();
2660 0 : FlushErrorState();
2661 0 : MemoryContextReset(PortalContext);
2662 :
2663 : /* restart our transaction for the following operations */
2664 0 : StartTransactionCommand();
2665 0 : RESUME_INTERRUPTS();
2666 : }
2667 6 : PG_END_TRY();
2668 :
2669 : /* Make sure we're back in AutovacMemCxt */
2670 6 : MemoryContextSwitchTo(AutovacMemCxt);
2671 :
2672 : /* We intentionally do not set did_vacuum here */
2673 :
2674 : /* be tidy */
2675 6 : deleted2:
2676 6 : if (cur_datname)
2677 6 : pfree(cur_datname);
2678 6 : if (cur_nspname)
2679 6 : pfree(cur_nspname);
2680 6 : if (cur_relname)
2681 6 : pfree(cur_relname);
2682 6 : }
2683 :
2684 : /*
2685 : * extract_autovac_opts
2686 : *
2687 : * Given a relation's pg_class tuple, return the AutoVacOpts portion of
2688 : * reloptions, if set; otherwise, return NULL.
2689 : *
2690 : * Note: callers do not have a relation lock on the table at this point,
2691 : * so the table could have been dropped, and its catalog rows gone, after
2692 : * we acquired the pg_class row. If pg_class had a TOAST table, this would
2693 : * be a risk; fortunately, it doesn't.
2694 : */
2695 : static AutoVacOpts *
2696 364942 : extract_autovac_opts(HeapTuple tup, TupleDesc pg_class_desc)
2697 : {
2698 : bytea *relopts;
2699 : AutoVacOpts *av;
2700 :
2701 : Assert(((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_RELATION ||
2702 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_MATVIEW ||
2703 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_TOASTVALUE);
2704 :
2705 364942 : relopts = extractRelOptions(tup, pg_class_desc, NULL);
2706 364942 : if (relopts == NULL)
2707 361702 : return NULL;
2708 :
2709 3240 : av = palloc(sizeof(AutoVacOpts));
2710 3240 : memcpy(av, &(((StdRdOptions *) relopts)->autovacuum), sizeof(AutoVacOpts));
2711 3240 : pfree(relopts);
2712 :
2713 3240 : return av;
2714 : }
2715 :
2716 :
2717 : /*
2718 : * table_recheck_autovac
2719 : *
2720 : * Recheck whether a table still needs vacuum or analyze. Return value is a
2721 : * valid autovac_table pointer if it does, NULL otherwise.
2722 : *
2723 : * Note that the returned autovac_table does not have the name fields set.
2724 : */
2725 : static autovac_table *
2726 103912 : table_recheck_autovac(Oid relid, HTAB *table_toast_map,
2727 : TupleDesc pg_class_desc,
2728 : int effective_multixact_freeze_max_age)
2729 : {
2730 : Form_pg_class classForm;
2731 : HeapTuple classTup;
2732 : bool dovacuum;
2733 : bool doanalyze;
2734 103912 : autovac_table *tab = NULL;
2735 : bool wraparound;
2736 : AutoVacOpts *avopts;
2737 :
2738 : /* fetch the relation's relcache entry */
2739 103912 : classTup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2740 103912 : if (!HeapTupleIsValid(classTup))
2741 0 : return NULL;
2742 103912 : classForm = (Form_pg_class) GETSTRUCT(classTup);
2743 :
2744 : /*
2745 : * Get the applicable reloptions. If it is a TOAST table, try to get the
2746 : * main table reloptions if the toast table itself doesn't have.
2747 : */
2748 103912 : avopts = extract_autovac_opts(classTup, pg_class_desc);
2749 103912 : if (classForm->relkind == RELKIND_TOASTVALUE &&
2750 38684 : avopts == NULL && table_toast_map != NULL)
2751 : {
2752 : av_relation *hentry;
2753 : bool found;
2754 :
2755 38684 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2756 38684 : if (found && hentry->ar_hasrelopts)
2757 1340 : avopts = &hentry->ar_reloptions;
2758 : }
2759 :
2760 103912 : recheck_relation_needs_vacanalyze(relid, avopts, classForm,
2761 : effective_multixact_freeze_max_age,
2762 : &dovacuum, &doanalyze, &wraparound);
2763 :
2764 : /* OK, it needs something done */
2765 103912 : if (doanalyze || dovacuum)
2766 : {
2767 : int freeze_min_age;
2768 : int freeze_table_age;
2769 : int multixact_freeze_min_age;
2770 : int multixact_freeze_table_age;
2771 : int log_min_duration;
2772 :
2773 : /*
2774 : * Calculate the vacuum cost parameters and the freeze ages. If there
2775 : * are options set in pg_class.reloptions, use them; in the case of a
2776 : * toast table, try the main table too. Otherwise use the GUC
2777 : * defaults, autovacuum's own first and plain vacuum second.
2778 : */
2779 :
2780 : /* -1 in autovac setting means use log_autovacuum_min_duration */
2781 2688 : log_min_duration = (avopts && avopts->log_min_duration >= 0)
2782 : ? avopts->log_min_duration
2783 106600 : : Log_autovacuum_min_duration;
2784 :
2785 : /* these do not have autovacuum-specific settings */
2786 2688 : freeze_min_age = (avopts && avopts->freeze_min_age >= 0)
2787 : ? avopts->freeze_min_age
2788 106600 : : default_freeze_min_age;
2789 :
2790 2688 : freeze_table_age = (avopts && avopts->freeze_table_age >= 0)
2791 : ? avopts->freeze_table_age
2792 106600 : : default_freeze_table_age;
2793 :
2794 106600 : multixact_freeze_min_age = (avopts &&
2795 2688 : avopts->multixact_freeze_min_age >= 0)
2796 : ? avopts->multixact_freeze_min_age
2797 106600 : : default_multixact_freeze_min_age;
2798 :
2799 106600 : multixact_freeze_table_age = (avopts &&
2800 2688 : avopts->multixact_freeze_table_age >= 0)
2801 : ? avopts->multixact_freeze_table_age
2802 106600 : : default_multixact_freeze_table_age;
2803 :
2804 103912 : tab = palloc(sizeof(autovac_table));
2805 103912 : tab->at_relid = relid;
2806 103912 : tab->at_sharedrel = classForm->relisshared;
2807 :
2808 : /*
2809 : * Select VACUUM options. Note we don't say VACOPT_PROCESS_TOAST, so
2810 : * that vacuum() skips toast relations. Also note we tell vacuum() to
2811 : * skip vac_update_datfrozenxid(); we'll do that separately.
2812 : */
2813 103912 : tab->at_params.options =
2814 103912 : (dovacuum ? (VACOPT_VACUUM |
2815 : VACOPT_PROCESS_MAIN |
2816 103912 : VACOPT_SKIP_DATABASE_STATS) : 0) |
2817 103912 : (doanalyze ? VACOPT_ANALYZE : 0) |
2818 103912 : (!wraparound ? VACOPT_SKIP_LOCKED : 0);
2819 :
2820 : /*
2821 : * index_cleanup and truncate are unspecified at first in autovacuum.
2822 : * They will be filled in with usable values using their reloptions
2823 : * (or reloption defaults) later.
2824 : */
2825 103912 : tab->at_params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
2826 103912 : tab->at_params.truncate = VACOPTVALUE_UNSPECIFIED;
2827 : /* As of now, we don't support parallel vacuum for autovacuum */
2828 103912 : tab->at_params.nworkers = -1;
2829 103912 : tab->at_params.freeze_min_age = freeze_min_age;
2830 103912 : tab->at_params.freeze_table_age = freeze_table_age;
2831 103912 : tab->at_params.multixact_freeze_min_age = multixact_freeze_min_age;
2832 103912 : tab->at_params.multixact_freeze_table_age = multixact_freeze_table_age;
2833 103912 : tab->at_params.is_wraparound = wraparound;
2834 103912 : tab->at_params.log_min_duration = log_min_duration;
2835 103912 : tab->at_params.toast_parent = InvalidOid;
2836 :
2837 : /*
2838 : * Later, in vacuum_rel(), we check reloptions for any
2839 : * vacuum_max_eager_freeze_failure_rate override.
2840 : */
2841 103912 : tab->at_params.max_eager_freeze_failure_rate = vacuum_max_eager_freeze_failure_rate;
2842 103912 : tab->at_storage_param_vac_cost_limit = avopts ?
2843 103912 : avopts->vacuum_cost_limit : 0;
2844 103912 : tab->at_storage_param_vac_cost_delay = avopts ?
2845 103912 : avopts->vacuum_cost_delay : -1;
2846 103912 : tab->at_relname = NULL;
2847 103912 : tab->at_nspname = NULL;
2848 103912 : tab->at_datname = NULL;
2849 :
2850 : /*
2851 : * If any of the cost delay parameters has been set individually for
2852 : * this table, disable the balancing algorithm.
2853 : */
2854 103912 : tab->at_dobalance =
2855 106600 : !(avopts && (avopts->vacuum_cost_limit > 0 ||
2856 2688 : avopts->vacuum_cost_delay >= 0));
2857 : }
2858 :
2859 103912 : heap_freetuple(classTup);
2860 103912 : return tab;
2861 : }
2862 :
2863 : /*
2864 : * recheck_relation_needs_vacanalyze
2865 : *
2866 : * Subroutine for table_recheck_autovac.
2867 : *
2868 : * Fetch the pgstat of a relation and recheck whether a relation
2869 : * needs to be vacuumed or analyzed.
2870 : */
2871 : static void
2872 103912 : recheck_relation_needs_vacanalyze(Oid relid,
2873 : AutoVacOpts *avopts,
2874 : Form_pg_class classForm,
2875 : int effective_multixact_freeze_max_age,
2876 : bool *dovacuum,
2877 : bool *doanalyze,
2878 : bool *wraparound)
2879 : {
2880 : PgStat_StatTabEntry *tabentry;
2881 :
2882 : /* fetch the pgstat table entry */
2883 103912 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2884 : relid);
2885 :
2886 103912 : relation_needs_vacanalyze(relid, avopts, classForm, tabentry,
2887 : effective_multixact_freeze_max_age,
2888 : dovacuum, doanalyze, wraparound);
2889 :
2890 : /* ignore ANALYZE for toast tables */
2891 103912 : if (classForm->relkind == RELKIND_TOASTVALUE)
2892 38684 : *doanalyze = false;
2893 103912 : }
2894 :
2895 : /*
2896 : * relation_needs_vacanalyze
2897 : *
2898 : * Check whether a relation needs to be vacuumed or analyzed; return each into
2899 : * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is
2900 : * being forced because of Xid or multixact wraparound.
2901 : *
2902 : * relopts is a pointer to the AutoVacOpts options (either for itself in the
2903 : * case of a plain table, or for either itself or its parent table in the case
2904 : * of a TOAST table), NULL if none; tabentry is the pgstats entry, which can be
2905 : * NULL.
2906 : *
2907 : * A table needs to be vacuumed if the number of dead tuples exceeds a
2908 : * threshold. This threshold is calculated as
2909 : *
2910 : * threshold = vac_base_thresh + vac_scale_factor * reltuples
2911 : * if (threshold > vac_max_thresh)
2912 : * threshold = vac_max_thresh;
2913 : *
2914 : * For analyze, the analysis done is that the number of tuples inserted,
2915 : * deleted and updated since the last analyze exceeds a threshold calculated
2916 : * in the same fashion as above. Note that the cumulative stats system stores
2917 : * the number of tuples (both live and dead) that there were as of the last
2918 : * analyze. This is asymmetric to the VACUUM case.
2919 : *
2920 : * We also force vacuum if the table's relfrozenxid is more than freeze_max_age
2921 : * transactions back, and if its relminmxid is more than
2922 : * multixact_freeze_max_age multixacts back.
2923 : *
2924 : * A table whose autovacuum_enabled option is false is
2925 : * automatically skipped (unless we have to vacuum it due to freeze_max_age).
2926 : * Thus autovacuum can be disabled for specific tables. Also, when the cumulative
2927 : * stats system does not have data about a table, it will be skipped.
2928 : *
2929 : * A table whose vac_base_thresh value is < 0 takes the base value from the
2930 : * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor
2931 : * value < 0 is substituted with the value of
2932 : * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze.
2933 : */
2934 : static void
2935 364942 : relation_needs_vacanalyze(Oid relid,
2936 : AutoVacOpts *relopts,
2937 : Form_pg_class classForm,
2938 : PgStat_StatTabEntry *tabentry,
2939 : int effective_multixact_freeze_max_age,
2940 : /* output params below */
2941 : bool *dovacuum,
2942 : bool *doanalyze,
2943 : bool *wraparound)
2944 : {
2945 : bool force_vacuum;
2946 : bool av_enabled;
2947 :
2948 : /* constants from reloptions or GUC variables */
2949 : int vac_base_thresh,
2950 : vac_max_thresh,
2951 : vac_ins_base_thresh,
2952 : anl_base_thresh;
2953 : float4 vac_scale_factor,
2954 : vac_ins_scale_factor,
2955 : anl_scale_factor;
2956 :
2957 : /* thresholds calculated from above constants */
2958 : float4 vacthresh,
2959 : vacinsthresh,
2960 : anlthresh;
2961 :
2962 : /* number of vacuum (resp. analyze) tuples at this time */
2963 : float4 vactuples,
2964 : instuples,
2965 : anltuples;
2966 :
2967 : /* freeze parameters */
2968 : int freeze_max_age;
2969 : int multixact_freeze_max_age;
2970 : TransactionId xidForceLimit;
2971 : TransactionId relfrozenxid;
2972 : MultiXactId multiForceLimit;
2973 :
2974 : Assert(classForm != NULL);
2975 : Assert(OidIsValid(relid));
2976 :
2977 : /*
2978 : * Determine vacuum/analyze equation parameters. We have two possible
2979 : * sources: the passed reloptions (which could be a main table or a toast
2980 : * table), or the autovacuum GUC variables.
2981 : */
2982 :
2983 : /* -1 in autovac setting means use plain vacuum_scale_factor */
2984 6454 : vac_scale_factor = (relopts && relopts->vacuum_scale_factor >= 0)
2985 0 : ? relopts->vacuum_scale_factor
2986 371396 : : autovacuum_vac_scale;
2987 :
2988 6454 : vac_base_thresh = (relopts && relopts->vacuum_threshold >= 0)
2989 : ? relopts->vacuum_threshold
2990 371396 : : autovacuum_vac_thresh;
2991 :
2992 : /* -1 is used to disable max threshold */
2993 6454 : vac_max_thresh = (relopts && relopts->vacuum_max_threshold >= -1)
2994 : ? relopts->vacuum_max_threshold
2995 371396 : : autovacuum_vac_max_thresh;
2996 :
2997 6454 : vac_ins_scale_factor = (relopts && relopts->vacuum_ins_scale_factor >= 0)
2998 0 : ? relopts->vacuum_ins_scale_factor
2999 371396 : : autovacuum_vac_ins_scale;
3000 :
3001 : /* -1 is used to disable insert vacuums */
3002 6454 : vac_ins_base_thresh = (relopts && relopts->vacuum_ins_threshold >= -1)
3003 : ? relopts->vacuum_ins_threshold
3004 371396 : : autovacuum_vac_ins_thresh;
3005 :
3006 6454 : anl_scale_factor = (relopts && relopts->analyze_scale_factor >= 0)
3007 0 : ? relopts->analyze_scale_factor
3008 371396 : : autovacuum_anl_scale;
3009 :
3010 6454 : anl_base_thresh = (relopts && relopts->analyze_threshold >= 0)
3011 : ? relopts->analyze_threshold
3012 371396 : : autovacuum_anl_thresh;
3013 :
3014 6454 : freeze_max_age = (relopts && relopts->freeze_max_age >= 0)
3015 0 : ? Min(relopts->freeze_max_age, autovacuum_freeze_max_age)
3016 371396 : : autovacuum_freeze_max_age;
3017 :
3018 6454 : multixact_freeze_max_age = (relopts && relopts->multixact_freeze_max_age >= 0)
3019 0 : ? Min(relopts->multixact_freeze_max_age, effective_multixact_freeze_max_age)
3020 371396 : : effective_multixact_freeze_max_age;
3021 :
3022 364942 : av_enabled = (relopts ? relopts->enabled : true);
3023 :
3024 : /* Force vacuum if table is at risk of wraparound */
3025 364942 : xidForceLimit = recentXid - freeze_max_age;
3026 364942 : if (xidForceLimit < FirstNormalTransactionId)
3027 0 : xidForceLimit -= FirstNormalTransactionId;
3028 364942 : relfrozenxid = classForm->relfrozenxid;
3029 729884 : force_vacuum = (TransactionIdIsNormal(relfrozenxid) &&
3030 364942 : TransactionIdPrecedes(relfrozenxid, xidForceLimit));
3031 364942 : if (!force_vacuum)
3032 : {
3033 157742 : MultiXactId relminmxid = classForm->relminmxid;
3034 :
3035 157742 : multiForceLimit = recentMulti - multixact_freeze_max_age;
3036 157742 : if (multiForceLimit < FirstMultiXactId)
3037 0 : multiForceLimit -= FirstMultiXactId;
3038 315484 : force_vacuum = MultiXactIdIsValid(relminmxid) &&
3039 157742 : MultiXactIdPrecedes(relminmxid, multiForceLimit);
3040 : }
3041 364942 : *wraparound = force_vacuum;
3042 :
3043 : /* User disabled it in pg_class.reloptions? (But ignore if at risk) */
3044 364942 : if (!av_enabled && !force_vacuum)
3045 : {
3046 1070 : *doanalyze = false;
3047 1070 : *dovacuum = false;
3048 1070 : return;
3049 : }
3050 :
3051 : /*
3052 : * If we found stats for the table, and autovacuum is currently enabled,
3053 : * make a threshold-based decision whether to vacuum and/or analyze. If
3054 : * autovacuum is currently disabled, we must be here for anti-wraparound
3055 : * vacuuming only, so don't vacuum (or analyze) anything that's not being
3056 : * forced.
3057 : */
3058 363872 : if (PointerIsValid(tabentry) && AutoVacuumingActive())
3059 358644 : {
3060 358644 : float4 pcnt_unfrozen = 1;
3061 358644 : float4 reltuples = classForm->reltuples;
3062 358644 : int32 relpages = classForm->relpages;
3063 358644 : int32 relallfrozen = classForm->relallfrozen;
3064 :
3065 358644 : vactuples = tabentry->dead_tuples;
3066 358644 : instuples = tabentry->ins_since_vacuum;
3067 358644 : anltuples = tabentry->mod_since_analyze;
3068 :
3069 : /* If the table hasn't yet been vacuumed, take reltuples as zero */
3070 358644 : if (reltuples < 0)
3071 618 : reltuples = 0;
3072 :
3073 : /*
3074 : * If we have data for relallfrozen, calculate the unfrozen percentage
3075 : * of the table to modify insert scale factor. This helps us decide
3076 : * whether or not to vacuum an insert-heavy table based on the number
3077 : * of inserts to the more "active" part of the table.
3078 : */
3079 358644 : if (relpages > 0 && relallfrozen > 0)
3080 : {
3081 : /*
3082 : * It could be the stats were updated manually and relallfrozen >
3083 : * relpages. Clamp relallfrozen to relpages to avoid nonsensical
3084 : * calculations.
3085 : */
3086 139164 : relallfrozen = Min(relallfrozen, relpages);
3087 139164 : pcnt_unfrozen = 1 - ((float4) relallfrozen / relpages);
3088 : }
3089 :
3090 358644 : vacthresh = (float4) vac_base_thresh + vac_scale_factor * reltuples;
3091 358644 : if (vac_max_thresh >= 0 && vacthresh > (float4) vac_max_thresh)
3092 0 : vacthresh = (float4) vac_max_thresh;
3093 :
3094 358644 : vacinsthresh = (float4) vac_ins_base_thresh +
3095 358644 : vac_ins_scale_factor * reltuples * pcnt_unfrozen;
3096 358644 : anlthresh = (float4) anl_base_thresh + anl_scale_factor * reltuples;
3097 :
3098 : /*
3099 : * Note that we don't need to take special consideration for stat
3100 : * reset, because if that happens, the last vacuum and analyze counts
3101 : * will be reset too.
3102 : */
3103 358644 : if (vac_ins_base_thresh >= 0)
3104 358644 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: %.0f (threshold %.0f), anl: %.0f (threshold %.0f)",
3105 : NameStr(classForm->relname),
3106 : vactuples, vacthresh, instuples, vacinsthresh, anltuples, anlthresh);
3107 : else
3108 0 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: (disabled), anl: %.0f (threshold %.0f)",
3109 : NameStr(classForm->relname),
3110 : vactuples, vacthresh, anltuples, anlthresh);
3111 :
3112 : /* Determine if this table needs vacuum or analyze. */
3113 511762 : *dovacuum = force_vacuum || (vactuples > vacthresh) ||
3114 153118 : (vac_ins_base_thresh >= 0 && instuples > vacinsthresh);
3115 358644 : *doanalyze = (anltuples > anlthresh);
3116 : }
3117 : else
3118 : {
3119 : /*
3120 : * Skip a table not found in stat hash, unless we have to force vacuum
3121 : * for anti-wrap purposes. If it's not acted upon, there's no need to
3122 : * vacuum it.
3123 : */
3124 5228 : *dovacuum = force_vacuum;
3125 5228 : *doanalyze = false;
3126 : }
3127 :
3128 : /* ANALYZE refuses to work with pg_statistic */
3129 363872 : if (relid == StatisticRelationId)
3130 3314 : *doanalyze = false;
3131 : }
3132 :
3133 : /*
3134 : * autovacuum_do_vac_analyze
3135 : * Vacuum and/or analyze the specified table
3136 : *
3137 : * We expect the caller to have switched into a memory context that won't
3138 : * disappear at transaction commit.
3139 : */
3140 : static void
3141 103912 : autovacuum_do_vac_analyze(autovac_table *tab, BufferAccessStrategy bstrategy)
3142 : {
3143 : RangeVar *rangevar;
3144 : VacuumRelation *rel;
3145 : List *rel_list;
3146 : MemoryContext vac_context;
3147 :
3148 : /* Let pgstat know what we're doing */
3149 103912 : autovac_report_activity(tab);
3150 :
3151 : /* Set up one VacuumRelation target, identified by OID, for vacuum() */
3152 103912 : rangevar = makeRangeVar(tab->at_nspname, tab->at_relname, -1);
3153 103912 : rel = makeVacuumRelation(rangevar, tab->at_relid, NIL);
3154 103912 : rel_list = list_make1(rel);
3155 :
3156 103912 : vac_context = AllocSetContextCreate(CurrentMemoryContext,
3157 : "Vacuum",
3158 : ALLOCSET_DEFAULT_SIZES);
3159 :
3160 103912 : vacuum(rel_list, &tab->at_params, bstrategy, vac_context, true);
3161 :
3162 103912 : MemoryContextDelete(vac_context);
3163 103912 : }
3164 :
3165 : /*
3166 : * autovac_report_activity
3167 : * Report to pgstat what autovacuum is doing
3168 : *
3169 : * We send a SQL string corresponding to what the user would see if the
3170 : * equivalent command was to be issued manually.
3171 : *
3172 : * Note we assume that we are going to report the next command as soon as we're
3173 : * done with the current one, and exit right after the last one, so we don't
3174 : * bother to report "<IDLE>" or some such.
3175 : */
3176 : static void
3177 103912 : autovac_report_activity(autovac_table *tab)
3178 : {
3179 : #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56)
3180 : char activity[MAX_AUTOVAC_ACTIV_LEN];
3181 : int len;
3182 :
3183 : /* Report the command and possible options */
3184 103912 : if (tab->at_params.options & VACOPT_VACUUM)
3185 103724 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3186 : "autovacuum: VACUUM%s",
3187 103724 : tab->at_params.options & VACOPT_ANALYZE ? " ANALYZE" : "");
3188 : else
3189 188 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3190 : "autovacuum: ANALYZE");
3191 :
3192 : /*
3193 : * Report the qualified name of the relation.
3194 : */
3195 103912 : len = strlen(activity);
3196 :
3197 103912 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3198 : " %s.%s%s", tab->at_nspname, tab->at_relname,
3199 103912 : tab->at_params.is_wraparound ? " (to prevent wraparound)" : "");
3200 :
3201 : /* Set statement_timestamp() to current time for pg_stat_activity */
3202 103912 : SetCurrentStatementStartTimestamp();
3203 :
3204 103912 : pgstat_report_activity(STATE_RUNNING, activity);
3205 103912 : }
3206 :
3207 : /*
3208 : * autovac_report_workitem
3209 : * Report to pgstat that autovacuum is processing a work item
3210 : */
3211 : static void
3212 6 : autovac_report_workitem(AutoVacuumWorkItem *workitem,
3213 : const char *nspname, const char *relname)
3214 : {
3215 : char activity[MAX_AUTOVAC_ACTIV_LEN + 12 + 2];
3216 : char blk[12 + 2];
3217 : int len;
3218 :
3219 6 : switch (workitem->avw_type)
3220 : {
3221 6 : case AVW_BRINSummarizeRange:
3222 6 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3223 : "autovacuum: BRIN summarize");
3224 6 : break;
3225 : }
3226 :
3227 : /*
3228 : * Report the qualified name of the relation, and the block number if any
3229 : */
3230 6 : len = strlen(activity);
3231 :
3232 6 : if (BlockNumberIsValid(workitem->avw_blockNumber))
3233 6 : snprintf(blk, sizeof(blk), " %u", workitem->avw_blockNumber);
3234 : else
3235 0 : blk[0] = '\0';
3236 :
3237 6 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3238 : " %s.%s%s", nspname, relname, blk);
3239 :
3240 : /* Set statement_timestamp() to current time for pg_stat_activity */
3241 6 : SetCurrentStatementStartTimestamp();
3242 :
3243 6 : pgstat_report_activity(STATE_RUNNING, activity);
3244 6 : }
3245 :
3246 : /*
3247 : * AutoVacuumingActive
3248 : * Check GUC vars and report whether the autovacuum process should be
3249 : * running.
3250 : */
3251 : bool
3252 432936 : AutoVacuumingActive(void)
3253 : {
3254 432936 : if (!autovacuum_start_daemon || !pgstat_track_counts)
3255 10710 : return false;
3256 422226 : return true;
3257 : }
3258 :
3259 : /*
3260 : * Request one work item to the next autovacuum run processing our database.
3261 : * Return false if the request can't be recorded.
3262 : */
3263 : bool
3264 6 : AutoVacuumRequestWork(AutoVacuumWorkItemType type, Oid relationId,
3265 : BlockNumber blkno)
3266 : {
3267 : int i;
3268 6 : bool result = false;
3269 :
3270 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
3271 :
3272 : /*
3273 : * Locate an unused work item and fill it with the given data.
3274 : */
3275 12 : for (i = 0; i < NUM_WORKITEMS; i++)
3276 : {
3277 12 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
3278 :
3279 12 : if (workitem->avw_used)
3280 6 : continue;
3281 :
3282 6 : workitem->avw_used = true;
3283 6 : workitem->avw_active = false;
3284 6 : workitem->avw_type = type;
3285 6 : workitem->avw_database = MyDatabaseId;
3286 6 : workitem->avw_relation = relationId;
3287 6 : workitem->avw_blockNumber = blkno;
3288 6 : result = true;
3289 :
3290 : /* done */
3291 6 : break;
3292 : }
3293 :
3294 6 : LWLockRelease(AutovacuumLock);
3295 :
3296 6 : return result;
3297 : }
3298 :
3299 : /*
3300 : * autovac_init
3301 : * This is called at postmaster initialization.
3302 : *
3303 : * All we do here is annoy the user if he got it wrong.
3304 : */
3305 : void
3306 1680 : autovac_init(void)
3307 : {
3308 1680 : if (!autovacuum_start_daemon)
3309 224 : return;
3310 1456 : else if (!pgstat_track_counts)
3311 0 : ereport(WARNING,
3312 : (errmsg("autovacuum not started because of misconfiguration"),
3313 : errhint("Enable the \"track_counts\" option.")));
3314 : else
3315 1456 : check_av_worker_gucs();
3316 : }
3317 :
3318 : /*
3319 : * AutoVacuumShmemSize
3320 : * Compute space needed for autovacuum-related shared memory
3321 : */
3322 : Size
3323 6006 : AutoVacuumShmemSize(void)
3324 : {
3325 : Size size;
3326 :
3327 : /*
3328 : * Need the fixed struct and the array of WorkerInfoData.
3329 : */
3330 6006 : size = sizeof(AutoVacuumShmemStruct);
3331 6006 : size = MAXALIGN(size);
3332 6006 : size = add_size(size, mul_size(autovacuum_worker_slots,
3333 : sizeof(WorkerInfoData)));
3334 6006 : return size;
3335 : }
3336 :
3337 : /*
3338 : * AutoVacuumShmemInit
3339 : * Allocate and initialize autovacuum-related shared memory
3340 : */
3341 : void
3342 2100 : AutoVacuumShmemInit(void)
3343 : {
3344 : bool found;
3345 :
3346 2100 : AutoVacuumShmem = (AutoVacuumShmemStruct *)
3347 2100 : ShmemInitStruct("AutoVacuum Data",
3348 : AutoVacuumShmemSize(),
3349 : &found);
3350 :
3351 2100 : if (!IsUnderPostmaster)
3352 : {
3353 : WorkerInfo worker;
3354 : int i;
3355 :
3356 : Assert(!found);
3357 :
3358 2100 : AutoVacuumShmem->av_launcherpid = 0;
3359 2100 : dclist_init(&AutoVacuumShmem->av_freeWorkers);
3360 2100 : dlist_init(&AutoVacuumShmem->av_runningWorkers);
3361 2100 : AutoVacuumShmem->av_startingWorker = NULL;
3362 2100 : memset(AutoVacuumShmem->av_workItems, 0,
3363 : sizeof(AutoVacuumWorkItem) * NUM_WORKITEMS);
3364 :
3365 2100 : worker = (WorkerInfo) ((char *) AutoVacuumShmem +
3366 : MAXALIGN(sizeof(AutoVacuumShmemStruct)));
3367 :
3368 : /* initialize the WorkerInfo free list */
3369 25612 : for (i = 0; i < autovacuum_worker_slots; i++)
3370 : {
3371 23512 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
3372 23512 : &worker[i].wi_links);
3373 23512 : pg_atomic_init_flag(&worker[i].wi_dobalance);
3374 : }
3375 :
3376 2100 : pg_atomic_init_u32(&AutoVacuumShmem->av_nworkersForBalance, 0);
3377 :
3378 : }
3379 : else
3380 : Assert(found);
3381 2100 : }
3382 :
3383 : /*
3384 : * GUC check_hook for autovacuum_work_mem
3385 : */
3386 : bool
3387 2172 : check_autovacuum_work_mem(int *newval, void **extra, GucSource source)
3388 : {
3389 : /*
3390 : * -1 indicates fallback.
3391 : *
3392 : * If we haven't yet changed the boot_val default of -1, just let it be.
3393 : * Autovacuum will look to maintenance_work_mem instead.
3394 : */
3395 2172 : if (*newval == -1)
3396 2168 : return true;
3397 :
3398 : /*
3399 : * We clamp manually-set values to at least 64kB. Since
3400 : * maintenance_work_mem is always set to at least this value, do the same
3401 : * here.
3402 : */
3403 4 : if (*newval < 64)
3404 4 : *newval = 64;
3405 :
3406 4 : return true;
3407 : }
3408 :
3409 : /*
3410 : * Returns whether there is a free autovacuum worker slot available.
3411 : */
3412 : static bool
3413 21194 : av_worker_available(void)
3414 : {
3415 : int free_slots;
3416 : int reserved_slots;
3417 :
3418 21194 : free_slots = dclist_count(&AutoVacuumShmem->av_freeWorkers);
3419 :
3420 21194 : reserved_slots = autovacuum_worker_slots - autovacuum_max_workers;
3421 21194 : reserved_slots = Max(0, reserved_slots);
3422 :
3423 21194 : return free_slots > reserved_slots;
3424 : }
3425 :
3426 : /*
3427 : * Emits a WARNING if autovacuum_worker_slots < autovacuum_max_workers.
3428 : */
3429 : static void
3430 1456 : check_av_worker_gucs(void)
3431 : {
3432 1456 : if (autovacuum_worker_slots < autovacuum_max_workers)
3433 0 : ereport(WARNING,
3434 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3435 : errmsg("\"autovacuum_max_workers\" (%d) should be less than or equal to \"autovacuum_worker_slots\" (%d)",
3436 : autovacuum_max_workers, autovacuum_worker_slots),
3437 : errdetail("The server will only start up to \"autovacuum_worker_slots\" (%d) autovacuum workers at a given time.",
3438 : autovacuum_worker_slots)));
3439 1456 : }
|
