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/vacuum.h"
81 : #include "common/int.h"
82 : #include "lib/ilist.h"
83 : #include "libpq/pqsignal.h"
84 : #include "miscadmin.h"
85 : #include "nodes/makefuncs.h"
86 : #include "pgstat.h"
87 : #include "postmaster/autovacuum.h"
88 : #include "postmaster/interrupt.h"
89 : #include "postmaster/postmaster.h"
90 : #include "storage/aio_subsys.h"
91 : #include "storage/bufmgr.h"
92 : #include "storage/ipc.h"
93 : #include "storage/latch.h"
94 : #include "storage/lmgr.h"
95 : #include "storage/pmsignal.h"
96 : #include "storage/proc.h"
97 : #include "storage/procsignal.h"
98 : #include "storage/smgr.h"
99 : #include "tcop/tcopprot.h"
100 : #include "utils/fmgroids.h"
101 : #include "utils/fmgrprotos.h"
102 : #include "utils/guc_hooks.h"
103 : #include "utils/injection_point.h"
104 : #include "utils/lsyscache.h"
105 : #include "utils/memutils.h"
106 : #include "utils/ps_status.h"
107 : #include "utils/rel.h"
108 : #include "utils/snapmgr.h"
109 : #include "utils/syscache.h"
110 : #include "utils/timeout.h"
111 : #include "utils/timestamp.h"
112 :
113 :
114 : /*
115 : * GUC parameters
116 : */
117 : bool autovacuum_start_daemon = false;
118 : int autovacuum_worker_slots;
119 : int autovacuum_max_workers;
120 : int autovacuum_work_mem = -1;
121 : int autovacuum_naptime;
122 : int autovacuum_vac_thresh;
123 : int autovacuum_vac_max_thresh;
124 : double autovacuum_vac_scale;
125 : int autovacuum_vac_ins_thresh;
126 : double autovacuum_vac_ins_scale;
127 : int autovacuum_anl_thresh;
128 : double autovacuum_anl_scale;
129 : int autovacuum_freeze_max_age;
130 : int autovacuum_multixact_freeze_max_age;
131 :
132 : double autovacuum_vac_cost_delay;
133 : int autovacuum_vac_cost_limit;
134 :
135 : int Log_autovacuum_min_duration = 600000;
136 :
137 : /* the minimum allowed time between two awakenings of the launcher */
138 : #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */
139 : #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */
140 :
141 : /*
142 : * Variables to save the cost-related storage parameters for the current
143 : * relation being vacuumed by this autovacuum worker. Using these, we can
144 : * ensure we don't overwrite the values of vacuum_cost_delay and
145 : * vacuum_cost_limit after reloading the configuration file. They are
146 : * initialized to "invalid" values to indicate that no cost-related storage
147 : * parameters were specified and will be set in do_autovacuum() after checking
148 : * the storage parameters in table_recheck_autovac().
149 : */
150 : static double av_storage_param_cost_delay = -1;
151 : static int av_storage_param_cost_limit = -1;
152 :
153 : /* Flags set by signal handlers */
154 : static volatile sig_atomic_t got_SIGUSR2 = false;
155 :
156 : /* Comparison points for determining whether freeze_max_age is exceeded */
157 : static TransactionId recentXid;
158 : static MultiXactId recentMulti;
159 :
160 : /* Default freeze ages to use for autovacuum (varies by database) */
161 : static int default_freeze_min_age;
162 : static int default_freeze_table_age;
163 : static int default_multixact_freeze_min_age;
164 : static int default_multixact_freeze_table_age;
165 :
166 : /* Memory context for long-lived data */
167 : static MemoryContext AutovacMemCxt;
168 :
169 : /* struct to keep track of databases in launcher */
170 : typedef struct avl_dbase
171 : {
172 : Oid adl_datid; /* hash key -- must be first */
173 : TimestampTz adl_next_worker;
174 : int adl_score;
175 : dlist_node adl_node;
176 : } avl_dbase;
177 :
178 : /* struct to keep track of databases in worker */
179 : typedef struct avw_dbase
180 : {
181 : Oid adw_datid;
182 : char *adw_name;
183 : TransactionId adw_frozenxid;
184 : MultiXactId adw_minmulti;
185 : PgStat_StatDBEntry *adw_entry;
186 : } avw_dbase;
187 :
188 : /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */
189 : typedef struct av_relation
190 : {
191 : Oid ar_toastrelid; /* hash key - must be first */
192 : Oid ar_relid;
193 : bool ar_hasrelopts;
194 : AutoVacOpts ar_reloptions; /* copy of AutoVacOpts from the main table's
195 : * reloptions, or NULL if none */
196 : } av_relation;
197 :
198 : /* struct to keep track of tables to vacuum and/or analyze, after rechecking */
199 : typedef struct autovac_table
200 : {
201 : Oid at_relid;
202 : VacuumParams at_params;
203 : double at_storage_param_vac_cost_delay;
204 : int at_storage_param_vac_cost_limit;
205 : bool at_dobalance;
206 : bool at_sharedrel;
207 : char *at_relname;
208 : char *at_nspname;
209 : char *at_datname;
210 : } autovac_table;
211 :
212 : /*-------------
213 : * This struct holds information about a single worker's whereabouts. We keep
214 : * an array of these in shared memory, sized according to
215 : * autovacuum_worker_slots.
216 : *
217 : * wi_links entry into free list or running list
218 : * wi_dboid OID of the database this worker is supposed to work on
219 : * wi_tableoid OID of the table currently being vacuumed, if any
220 : * wi_sharedrel flag indicating whether table is marked relisshared
221 : * wi_proc pointer to PGPROC of the running worker, NULL if not started
222 : * wi_launchtime Time at which this worker was launched
223 : * wi_dobalance Whether this worker should be included in balance calculations
224 : *
225 : * All fields are protected by AutovacuumLock, except for wi_tableoid and
226 : * wi_sharedrel which are protected by AutovacuumScheduleLock (note these
227 : * two fields are read-only for everyone except that worker itself).
228 : *-------------
229 : */
230 : typedef struct WorkerInfoData
231 : {
232 : dlist_node wi_links;
233 : Oid wi_dboid;
234 : Oid wi_tableoid;
235 : PGPROC *wi_proc;
236 : TimestampTz wi_launchtime;
237 : pg_atomic_flag wi_dobalance;
238 : bool wi_sharedrel;
239 : } WorkerInfoData;
240 :
241 : typedef struct WorkerInfoData *WorkerInfo;
242 :
243 : /*
244 : * Possible signals received by the launcher from remote processes. These are
245 : * stored atomically in shared memory so that other processes can set them
246 : * without locking.
247 : */
248 : typedef enum
249 : {
250 : AutoVacForkFailed, /* failed trying to start a worker */
251 : AutoVacRebalance, /* rebalance the cost limits */
252 : } AutoVacuumSignal;
253 :
254 : #define AutoVacNumSignals (AutoVacRebalance + 1)
255 :
256 : /*
257 : * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This
258 : * list is mostly protected by AutovacuumLock, except that if an item is
259 : * marked 'active' other processes must not modify the work-identifying
260 : * members.
261 : */
262 : typedef struct AutoVacuumWorkItem
263 : {
264 : AutoVacuumWorkItemType avw_type;
265 : bool avw_used; /* below data is valid */
266 : bool avw_active; /* being processed */
267 : Oid avw_database;
268 : Oid avw_relation;
269 : BlockNumber avw_blockNumber;
270 : } AutoVacuumWorkItem;
271 :
272 : #define NUM_WORKITEMS 256
273 :
274 : /*-------------
275 : * The main autovacuum shmem struct. On shared memory we store this main
276 : * struct and the array of WorkerInfo structs. This struct keeps:
277 : *
278 : * av_signal set by other processes to indicate various conditions
279 : * av_launcherpid the PID of the autovacuum launcher
280 : * av_freeWorkers the WorkerInfo freelist
281 : * av_runningWorkers the WorkerInfo non-free queue
282 : * av_startingWorker pointer to WorkerInfo currently being started (cleared by
283 : * the worker itself as soon as it's up and running)
284 : * av_workItems work item array
285 : * av_nworkersForBalance the number of autovacuum workers to use when
286 : * calculating the per worker cost limit
287 : *
288 : * This struct is protected by AutovacuumLock, except for av_signal and parts
289 : * of the worker list (see above).
290 : *-------------
291 : */
292 : typedef struct
293 : {
294 : sig_atomic_t av_signal[AutoVacNumSignals];
295 : pid_t av_launcherpid;
296 : dclist_head av_freeWorkers;
297 : dlist_head av_runningWorkers;
298 : WorkerInfo av_startingWorker;
299 : AutoVacuumWorkItem av_workItems[NUM_WORKITEMS];
300 : pg_atomic_uint32 av_nworkersForBalance;
301 : } AutoVacuumShmemStruct;
302 :
303 : static AutoVacuumShmemStruct *AutoVacuumShmem;
304 :
305 : /*
306 : * the database list (of avl_dbase elements) in the launcher, and the context
307 : * that contains it
308 : */
309 : static dlist_head DatabaseList = DLIST_STATIC_INIT(DatabaseList);
310 : static MemoryContext DatabaseListCxt = NULL;
311 :
312 : /*
313 : * Dummy pointer to persuade Valgrind that we've not leaked the array of
314 : * avl_dbase structs. Make it global to ensure the compiler doesn't
315 : * optimize it away.
316 : */
317 : #ifdef USE_VALGRIND
318 : extern avl_dbase *avl_dbase_array;
319 : avl_dbase *avl_dbase_array;
320 : #endif
321 :
322 : /* Pointer to my own WorkerInfo, valid on each worker */
323 : static WorkerInfo MyWorkerInfo = NULL;
324 :
325 : /* PID of launcher, valid only in worker while shutting down */
326 : int AutovacuumLauncherPid = 0;
327 :
328 : static Oid do_start_worker(void);
329 : static void ProcessAutoVacLauncherInterrupts(void);
330 : pg_noreturn static void AutoVacLauncherShutdown(void);
331 : static void launcher_determine_sleep(bool canlaunch, bool recursing,
332 : struct timeval *nap);
333 : static void launch_worker(TimestampTz now);
334 : static List *get_database_list(void);
335 : static void rebuild_database_list(Oid newdb);
336 : static int db_comparator(const void *a, const void *b);
337 : static void autovac_recalculate_workers_for_balance(void);
338 :
339 : static void do_autovacuum(void);
340 : static void FreeWorkerInfo(int code, Datum arg);
341 :
342 : static autovac_table *table_recheck_autovac(Oid relid, HTAB *table_toast_map,
343 : TupleDesc pg_class_desc,
344 : int effective_multixact_freeze_max_age);
345 : static void recheck_relation_needs_vacanalyze(Oid relid, AutoVacOpts *avopts,
346 : Form_pg_class classForm,
347 : int effective_multixact_freeze_max_age,
348 : bool *dovacuum, bool *doanalyze, bool *wraparound);
349 : static void relation_needs_vacanalyze(Oid relid, AutoVacOpts *relopts,
350 : Form_pg_class classForm,
351 : PgStat_StatTabEntry *tabentry,
352 : int effective_multixact_freeze_max_age,
353 : bool *dovacuum, bool *doanalyze, bool *wraparound);
354 :
355 : static void autovacuum_do_vac_analyze(autovac_table *tab,
356 : BufferAccessStrategy bstrategy);
357 : static AutoVacOpts *extract_autovac_opts(HeapTuple tup,
358 : TupleDesc pg_class_desc);
359 : static void perform_work_item(AutoVacuumWorkItem *workitem);
360 : static void autovac_report_activity(autovac_table *tab);
361 : static void autovac_report_workitem(AutoVacuumWorkItem *workitem,
362 : const char *nspname, const char *relname);
363 : static void avl_sigusr2_handler(SIGNAL_ARGS);
364 : static bool av_worker_available(void);
365 : static void check_av_worker_gucs(void);
366 :
367 :
368 :
369 : /********************************************************************
370 : * AUTOVACUUM LAUNCHER CODE
371 : ********************************************************************/
372 :
373 : /*
374 : * Main entry point for the autovacuum launcher process.
375 : */
376 : void
377 760 : AutoVacLauncherMain(const void *startup_data, size_t startup_data_len)
378 : {
379 : sigjmp_buf local_sigjmp_buf;
380 :
381 : Assert(startup_data_len == 0);
382 :
383 : /* Release postmaster's working memory context */
384 760 : if (PostmasterContext)
385 : {
386 760 : MemoryContextDelete(PostmasterContext);
387 760 : PostmasterContext = NULL;
388 : }
389 :
390 760 : MyBackendType = B_AUTOVAC_LAUNCHER;
391 760 : init_ps_display(NULL);
392 :
393 760 : ereport(DEBUG1,
394 : (errmsg_internal("autovacuum launcher started")));
395 :
396 760 : if (PostAuthDelay)
397 0 : pg_usleep(PostAuthDelay * 1000000L);
398 :
399 : Assert(GetProcessingMode() == InitProcessing);
400 :
401 : /*
402 : * Set up signal handlers. We operate on databases much like a regular
403 : * backend, so we use the same signal handling. See equivalent code in
404 : * tcop/postgres.c.
405 : */
406 760 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
407 760 : pqsignal(SIGINT, StatementCancelHandler);
408 760 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
409 : /* SIGQUIT handler was already set up by InitPostmasterChild */
410 :
411 760 : InitializeTimeouts(); /* establishes SIGALRM handler */
412 :
413 760 : pqsignal(SIGPIPE, SIG_IGN);
414 760 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
415 760 : pqsignal(SIGUSR2, avl_sigusr2_handler);
416 760 : pqsignal(SIGFPE, FloatExceptionHandler);
417 760 : pqsignal(SIGCHLD, SIG_DFL);
418 :
419 : /*
420 : * Create a per-backend PGPROC struct in shared memory. We must do this
421 : * before we can use LWLocks or access any shared memory.
422 : */
423 760 : InitProcess();
424 :
425 : /* Early initialization */
426 760 : BaseInit();
427 :
428 760 : InitPostgres(NULL, InvalidOid, NULL, InvalidOid, 0, NULL);
429 :
430 760 : SetProcessingMode(NormalProcessing);
431 :
432 : /*
433 : * Create a memory context that we will do all our work in. We do this so
434 : * that we can reset the context during error recovery and thereby avoid
435 : * possible memory leaks.
436 : */
437 760 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
438 : "Autovacuum Launcher",
439 : ALLOCSET_DEFAULT_SIZES);
440 760 : MemoryContextSwitchTo(AutovacMemCxt);
441 :
442 : /*
443 : * If an exception is encountered, processing resumes here.
444 : *
445 : * This code is a stripped down version of PostgresMain error recovery.
446 : *
447 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
448 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
449 : * signals other than SIGQUIT will be blocked until we complete error
450 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
451 : * call redundant, but it is not since InterruptPending might be set
452 : * already.
453 : */
454 760 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
455 : {
456 : /* since not using PG_TRY, must reset error stack by hand */
457 0 : error_context_stack = NULL;
458 :
459 : /* Prevents interrupts while cleaning up */
460 0 : HOLD_INTERRUPTS();
461 :
462 : /* Forget any pending QueryCancel or timeout request */
463 0 : disable_all_timeouts(false);
464 0 : QueryCancelPending = false; /* second to avoid race condition */
465 :
466 : /* Report the error to the server log */
467 0 : EmitErrorReport();
468 :
469 : /* Abort the current transaction in order to recover */
470 0 : AbortCurrentTransaction();
471 :
472 : /*
473 : * Release any other resources, for the case where we were not in a
474 : * transaction.
475 : */
476 0 : LWLockReleaseAll();
477 0 : pgstat_report_wait_end();
478 0 : pgaio_error_cleanup();
479 0 : UnlockBuffers();
480 : /* this is probably dead code, but let's be safe: */
481 0 : if (AuxProcessResourceOwner)
482 0 : ReleaseAuxProcessResources(false);
483 0 : AtEOXact_Buffers(false);
484 0 : AtEOXact_SMgr();
485 0 : AtEOXact_Files(false);
486 0 : AtEOXact_HashTables(false);
487 :
488 : /*
489 : * Now return to normal top-level context and clear ErrorContext for
490 : * next time.
491 : */
492 0 : MemoryContextSwitchTo(AutovacMemCxt);
493 0 : FlushErrorState();
494 :
495 : /* Flush any leaked data in the top-level context */
496 0 : MemoryContextReset(AutovacMemCxt);
497 :
498 : /* don't leave dangling pointers to freed memory */
499 0 : DatabaseListCxt = NULL;
500 0 : dlist_init(&DatabaseList);
501 :
502 : /* Now we can allow interrupts again */
503 0 : RESUME_INTERRUPTS();
504 :
505 : /* if in shutdown mode, no need for anything further; just go away */
506 0 : if (ShutdownRequestPending)
507 0 : AutoVacLauncherShutdown();
508 :
509 : /*
510 : * Sleep at least 1 second after any error. We don't want to be
511 : * filling the error logs as fast as we can.
512 : */
513 0 : pg_usleep(1000000L);
514 : }
515 :
516 : /* We can now handle ereport(ERROR) */
517 760 : PG_exception_stack = &local_sigjmp_buf;
518 :
519 : /* must unblock signals before calling rebuild_database_list */
520 760 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
521 :
522 : /*
523 : * Set always-secure search path. Launcher doesn't connect to a database,
524 : * so this has no effect.
525 : */
526 760 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
527 :
528 : /*
529 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
530 : * in postgresql.conf. We don't really want such a dangerous option being
531 : * applied non-interactively.
532 : */
533 760 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
534 :
535 : /*
536 : * Force settable timeouts off to avoid letting these settings prevent
537 : * regular maintenance from being executed.
538 : */
539 760 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
540 760 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
541 760 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
542 760 : SetConfigOption("idle_in_transaction_session_timeout", "0",
543 : PGC_SUSET, PGC_S_OVERRIDE);
544 :
545 : /*
546 : * Force default_transaction_isolation to READ COMMITTED. We don't want
547 : * to pay the overhead of serializable mode, nor add any risk of causing
548 : * deadlocks or delaying other transactions.
549 : */
550 760 : SetConfigOption("default_transaction_isolation", "read committed",
551 : PGC_SUSET, PGC_S_OVERRIDE);
552 :
553 : /*
554 : * Even when system is configured to use a different fetch consistency,
555 : * for autovac we always want fresh stats.
556 : */
557 760 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
558 :
559 : /*
560 : * In emergency mode, just start a worker (unless shutdown was requested)
561 : * and go away.
562 : */
563 760 : if (!AutoVacuumingActive())
564 : {
565 0 : if (!ShutdownRequestPending)
566 0 : do_start_worker();
567 0 : proc_exit(0); /* done */
568 : }
569 :
570 760 : AutoVacuumShmem->av_launcherpid = MyProcPid;
571 :
572 : /*
573 : * Create the initial database list. The invariant we want this list to
574 : * keep is that it's ordered by decreasing next_worker. As soon as an
575 : * entry is updated to a higher time, it will be moved to the front (which
576 : * is correct because the only operation is to add autovacuum_naptime to
577 : * the entry, and time always increases).
578 : */
579 760 : rebuild_database_list(InvalidOid);
580 :
581 : /* loop until shutdown request */
582 16618 : while (!ShutdownRequestPending)
583 : {
584 : struct timeval nap;
585 16612 : TimestampTz current_time = 0;
586 : bool can_launch;
587 :
588 : /*
589 : * This loop is a bit different from the normal use of WaitLatch,
590 : * because we'd like to sleep before the first launch of a child
591 : * process. So it's WaitLatch, then ResetLatch, then check for
592 : * wakening conditions.
593 : */
594 :
595 16612 : launcher_determine_sleep(av_worker_available(), false, &nap);
596 :
597 : /*
598 : * Wait until naptime expires or we get some type of signal (all the
599 : * signal handlers will wake us by calling SetLatch).
600 : */
601 16612 : (void) WaitLatch(MyLatch,
602 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
603 16612 : (nap.tv_sec * 1000L) + (nap.tv_usec / 1000L),
604 : WAIT_EVENT_AUTOVACUUM_MAIN);
605 :
606 16606 : ResetLatch(MyLatch);
607 :
608 16606 : ProcessAutoVacLauncherInterrupts();
609 :
610 : /*
611 : * a worker finished, or postmaster signaled failure to start a worker
612 : */
613 15858 : if (got_SIGUSR2)
614 : {
615 9446 : got_SIGUSR2 = false;
616 :
617 : /* rebalance cost limits, if needed */
618 9446 : if (AutoVacuumShmem->av_signal[AutoVacRebalance])
619 : {
620 4718 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
621 4718 : AutoVacuumShmem->av_signal[AutoVacRebalance] = false;
622 4718 : autovac_recalculate_workers_for_balance();
623 4718 : LWLockRelease(AutovacuumLock);
624 : }
625 :
626 9446 : if (AutoVacuumShmem->av_signal[AutoVacForkFailed])
627 : {
628 : /*
629 : * If the postmaster failed to start a new worker, we sleep
630 : * for a little while and resend the signal. The new worker's
631 : * state is still in memory, so this is sufficient. After
632 : * that, we restart the main loop.
633 : *
634 : * XXX should we put a limit to the number of times we retry?
635 : * I don't think it makes much sense, because a future start
636 : * of a worker will continue to fail in the same way.
637 : */
638 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = false;
639 0 : pg_usleep(1000000L); /* 1s */
640 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
641 4794 : continue;
642 : }
643 : }
644 :
645 : /*
646 : * There are some conditions that we need to check before trying to
647 : * start a worker. First, we need to make sure that there is a worker
648 : * slot available. Second, we need to make sure that no other worker
649 : * failed while starting up.
650 : */
651 :
652 15858 : current_time = GetCurrentTimestamp();
653 15858 : LWLockAcquire(AutovacuumLock, LW_SHARED);
654 :
655 15858 : can_launch = av_worker_available();
656 :
657 15858 : if (AutoVacuumShmem->av_startingWorker != NULL)
658 : {
659 : int waittime;
660 68 : WorkerInfo worker = AutoVacuumShmem->av_startingWorker;
661 :
662 : /*
663 : * We can't launch another worker when another one is still
664 : * starting up (or failed while doing so), so just sleep for a bit
665 : * more; that worker will wake us up again as soon as it's ready.
666 : * We will only wait autovacuum_naptime seconds (up to a maximum
667 : * of 60 seconds) for this to happen however. Note that failure
668 : * to connect to a particular database is not a problem here,
669 : * because the worker removes itself from the startingWorker
670 : * pointer before trying to connect. Problems detected by the
671 : * postmaster (like fork() failure) are also reported and handled
672 : * differently. The only problems that may cause this code to
673 : * fire are errors in the earlier sections of AutoVacWorkerMain,
674 : * before the worker removes the WorkerInfo from the
675 : * startingWorker pointer.
676 : */
677 68 : waittime = Min(autovacuum_naptime, 60) * 1000;
678 68 : if (TimestampDifferenceExceeds(worker->wi_launchtime, current_time,
679 : waittime))
680 : {
681 0 : LWLockRelease(AutovacuumLock);
682 0 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
683 :
684 : /*
685 : * No other process can put a worker in starting mode, so if
686 : * startingWorker is still INVALID after exchanging our lock,
687 : * we assume it's the same one we saw above (so we don't
688 : * recheck the launch time).
689 : */
690 0 : if (AutoVacuumShmem->av_startingWorker != NULL)
691 : {
692 0 : worker = AutoVacuumShmem->av_startingWorker;
693 0 : worker->wi_dboid = InvalidOid;
694 0 : worker->wi_tableoid = InvalidOid;
695 0 : worker->wi_sharedrel = false;
696 0 : worker->wi_proc = NULL;
697 0 : worker->wi_launchtime = 0;
698 0 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
699 : &worker->wi_links);
700 0 : AutoVacuumShmem->av_startingWorker = NULL;
701 0 : ereport(WARNING,
702 : errmsg("autovacuum worker took too long to start; canceled"));
703 : }
704 : }
705 : else
706 68 : can_launch = false;
707 : }
708 15858 : LWLockRelease(AutovacuumLock); /* either shared or exclusive */
709 :
710 : /* if we can't do anything, just go back to sleep */
711 15858 : if (!can_launch)
712 4794 : continue;
713 :
714 : /* We're OK to start a new worker */
715 :
716 11064 : if (dlist_is_empty(&DatabaseList))
717 : {
718 : /*
719 : * Special case when the list is empty: start a worker right away.
720 : * This covers the initial case, when no database is in pgstats
721 : * (thus the list is empty). Note that the constraints in
722 : * launcher_determine_sleep keep us from starting workers too
723 : * quickly (at most once every autovacuum_naptime when the list is
724 : * empty).
725 : */
726 10 : launch_worker(current_time);
727 : }
728 : else
729 : {
730 : /*
731 : * because rebuild_database_list constructs a list with most
732 : * distant adl_next_worker first, we obtain our database from the
733 : * tail of the list.
734 : */
735 : avl_dbase *avdb;
736 :
737 11054 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
738 :
739 : /*
740 : * launch a worker if next_worker is right now or it is in the
741 : * past
742 : */
743 11054 : if (TimestampDifferenceExceeds(avdb->adl_next_worker,
744 : current_time, 0))
745 4730 : launch_worker(current_time);
746 : }
747 : }
748 :
749 6 : AutoVacLauncherShutdown();
750 : }
751 :
752 : /*
753 : * Process any new interrupts.
754 : */
755 : static void
756 16606 : ProcessAutoVacLauncherInterrupts(void)
757 : {
758 : /* the normal shutdown case */
759 16606 : if (ShutdownRequestPending)
760 748 : AutoVacLauncherShutdown();
761 :
762 15858 : if (ConfigReloadPending)
763 : {
764 72 : int autovacuum_max_workers_prev = autovacuum_max_workers;
765 :
766 72 : ConfigReloadPending = false;
767 72 : ProcessConfigFile(PGC_SIGHUP);
768 :
769 : /* shutdown requested in config file? */
770 72 : if (!AutoVacuumingActive())
771 0 : AutoVacLauncherShutdown();
772 :
773 : /*
774 : * If autovacuum_max_workers changed, emit a WARNING if
775 : * autovacuum_worker_slots < autovacuum_max_workers. If it didn't
776 : * change, skip this to avoid too many repeated log messages.
777 : */
778 72 : if (autovacuum_max_workers_prev != autovacuum_max_workers)
779 0 : check_av_worker_gucs();
780 :
781 : /* rebuild the list in case the naptime changed */
782 72 : rebuild_database_list(InvalidOid);
783 : }
784 :
785 : /* Process barrier events */
786 15858 : if (ProcSignalBarrierPending)
787 54 : ProcessProcSignalBarrier();
788 :
789 : /* Perform logging of memory contexts of this process */
790 15858 : if (LogMemoryContextPending)
791 0 : ProcessLogMemoryContextInterrupt();
792 :
793 : /* Process sinval catchup interrupts that happened while sleeping */
794 15858 : ProcessCatchupInterrupt();
795 15858 : }
796 :
797 : /*
798 : * Perform a normal exit from the autovac launcher.
799 : */
800 : static void
801 754 : AutoVacLauncherShutdown(void)
802 : {
803 754 : ereport(DEBUG1,
804 : (errmsg_internal("autovacuum launcher shutting down")));
805 754 : AutoVacuumShmem->av_launcherpid = 0;
806 :
807 754 : proc_exit(0); /* done */
808 : }
809 :
810 : /*
811 : * Determine the time to sleep, based on the database list.
812 : *
813 : * The "canlaunch" parameter indicates whether we can start a worker right now,
814 : * for example due to the workers being all busy. If this is false, we will
815 : * cause a long sleep, which will be interrupted when a worker exits.
816 : */
817 : static void
818 16852 : launcher_determine_sleep(bool canlaunch, bool recursing, struct timeval *nap)
819 : {
820 : /*
821 : * We sleep until the next scheduled vacuum. We trust that when the
822 : * database list was built, care was taken so that no entries have times
823 : * in the past; if the first entry has too close a next_worker value, or a
824 : * time in the past, we will sleep a small nominal time.
825 : */
826 16852 : if (!canlaunch)
827 : {
828 8996 : nap->tv_sec = autovacuum_naptime;
829 8996 : nap->tv_usec = 0;
830 : }
831 7856 : else if (!dlist_is_empty(&DatabaseList))
832 : {
833 7808 : TimestampTz current_time = GetCurrentTimestamp();
834 : TimestampTz next_wakeup;
835 : avl_dbase *avdb;
836 : long secs;
837 : int usecs;
838 :
839 7808 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
840 :
841 7808 : next_wakeup = avdb->adl_next_worker;
842 7808 : TimestampDifference(current_time, next_wakeup, &secs, &usecs);
843 :
844 7808 : nap->tv_sec = secs;
845 7808 : nap->tv_usec = usecs;
846 : }
847 : else
848 : {
849 : /* list is empty, sleep for whole autovacuum_naptime seconds */
850 48 : nap->tv_sec = autovacuum_naptime;
851 48 : nap->tv_usec = 0;
852 : }
853 :
854 : /*
855 : * If the result is exactly zero, it means a database had an entry with
856 : * time in the past. Rebuild the list so that the databases are evenly
857 : * distributed again, and recalculate the time to sleep. This can happen
858 : * if there are more tables needing vacuum than workers, and they all take
859 : * longer to vacuum than autovacuum_naptime.
860 : *
861 : * We only recurse once. rebuild_database_list should always return times
862 : * in the future, but it seems best not to trust too much on that.
863 : */
864 16852 : if (nap->tv_sec == 0 && nap->tv_usec == 0 && !recursing)
865 : {
866 240 : rebuild_database_list(InvalidOid);
867 240 : launcher_determine_sleep(canlaunch, true, nap);
868 240 : return;
869 : }
870 :
871 : /* The smallest time we'll allow the launcher to sleep. */
872 16612 : if (nap->tv_sec <= 0 && nap->tv_usec <= MIN_AUTOVAC_SLEEPTIME * 1000)
873 : {
874 1662 : nap->tv_sec = 0;
875 1662 : nap->tv_usec = MIN_AUTOVAC_SLEEPTIME * 1000;
876 : }
877 :
878 : /*
879 : * If the sleep time is too large, clamp it to an arbitrary maximum (plus
880 : * any fractional seconds, for simplicity). This avoids an essentially
881 : * infinite sleep in strange cases like the system clock going backwards a
882 : * few years.
883 : */
884 16612 : if (nap->tv_sec > MAX_AUTOVAC_SLEEPTIME)
885 18 : nap->tv_sec = MAX_AUTOVAC_SLEEPTIME;
886 : }
887 :
888 : /*
889 : * Build an updated DatabaseList. It must only contain databases that appear
890 : * in pgstats, and must be sorted by next_worker from highest to lowest,
891 : * distributed regularly across the next autovacuum_naptime interval.
892 : *
893 : * Receives the Oid of the database that made this list be generated (we call
894 : * this the "new" database, because when the database was already present on
895 : * the list, we expect that this function is not called at all). The
896 : * preexisting list, if any, will be used to preserve the order of the
897 : * databases in the autovacuum_naptime period. The new database is put at the
898 : * end of the interval. The actual values are not saved, which should not be
899 : * much of a problem.
900 : */
901 : static void
902 1100 : rebuild_database_list(Oid newdb)
903 : {
904 : List *dblist;
905 : ListCell *cell;
906 : MemoryContext newcxt;
907 : MemoryContext oldcxt;
908 : MemoryContext tmpcxt;
909 : HASHCTL hctl;
910 : int score;
911 : int nelems;
912 : HTAB *dbhash;
913 : dlist_iter iter;
914 :
915 1100 : newcxt = AllocSetContextCreate(AutovacMemCxt,
916 : "Autovacuum database list",
917 : ALLOCSET_DEFAULT_SIZES);
918 1100 : tmpcxt = AllocSetContextCreate(newcxt,
919 : "Autovacuum database list (tmp)",
920 : ALLOCSET_DEFAULT_SIZES);
921 1100 : oldcxt = MemoryContextSwitchTo(tmpcxt);
922 :
923 : /*
924 : * Implementing this is not as simple as it sounds, because we need to put
925 : * the new database at the end of the list; next the databases that were
926 : * already on the list, and finally (at the tail of the list) all the
927 : * other databases that are not on the existing list.
928 : *
929 : * To do this, we build an empty hash table of scored databases. We will
930 : * start with the lowest score (zero) for the new database, then
931 : * increasing scores for the databases in the existing list, in order, and
932 : * lastly increasing scores for all databases gotten via
933 : * get_database_list() that are not already on the hash.
934 : *
935 : * Then we will put all the hash elements into an array, sort the array by
936 : * score, and finally put the array elements into the new doubly linked
937 : * list.
938 : */
939 1100 : hctl.keysize = sizeof(Oid);
940 1100 : hctl.entrysize = sizeof(avl_dbase);
941 1100 : hctl.hcxt = tmpcxt;
942 1100 : dbhash = hash_create("autovacuum db hash", 20, &hctl, /* magic number here
943 : * FIXME */
944 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
945 :
946 : /* start by inserting the new database */
947 1100 : score = 0;
948 1100 : if (OidIsValid(newdb))
949 : {
950 : avl_dbase *db;
951 : PgStat_StatDBEntry *entry;
952 :
953 : /* only consider this database if it has a pgstat entry */
954 28 : entry = pgstat_fetch_stat_dbentry(newdb);
955 28 : if (entry != NULL)
956 : {
957 : /* we assume it isn't found because the hash was just created */
958 22 : db = hash_search(dbhash, &newdb, HASH_ENTER, NULL);
959 :
960 : /* hash_search already filled in the key */
961 22 : db->adl_score = score++;
962 : /* next_worker is filled in later */
963 : }
964 : }
965 :
966 : /* Now insert the databases from the existing list */
967 1964 : dlist_foreach(iter, &DatabaseList)
968 : {
969 864 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
970 : avl_dbase *db;
971 : bool found;
972 : PgStat_StatDBEntry *entry;
973 :
974 : /*
975 : * skip databases with no stat entries -- in particular, this gets rid
976 : * of dropped databases
977 : */
978 864 : entry = pgstat_fetch_stat_dbentry(avdb->adl_datid);
979 864 : if (entry == NULL)
980 0 : continue;
981 :
982 864 : db = hash_search(dbhash, &(avdb->adl_datid), HASH_ENTER, &found);
983 :
984 864 : if (!found)
985 : {
986 : /* hash_search already filled in the key */
987 864 : db->adl_score = score++;
988 : /* next_worker is filled in later */
989 : }
990 : }
991 :
992 : /* finally, insert all qualifying databases not previously inserted */
993 1100 : dblist = get_database_list();
994 4720 : foreach(cell, dblist)
995 : {
996 3620 : avw_dbase *avdb = lfirst(cell);
997 : avl_dbase *db;
998 : bool found;
999 : PgStat_StatDBEntry *entry;
1000 :
1001 : /* only consider databases with a pgstat entry */
1002 3620 : entry = pgstat_fetch_stat_dbentry(avdb->adw_datid);
1003 3620 : if (entry == NULL)
1004 1650 : continue;
1005 :
1006 1970 : db = hash_search(dbhash, &(avdb->adw_datid), HASH_ENTER, &found);
1007 : /* only update the score if the database was not already on the hash */
1008 1970 : if (!found)
1009 : {
1010 : /* hash_search already filled in the key */
1011 1084 : db->adl_score = score++;
1012 : /* next_worker is filled in later */
1013 : }
1014 : }
1015 1100 : nelems = score;
1016 :
1017 : /* from here on, the allocated memory belongs to the new list */
1018 1100 : MemoryContextSwitchTo(newcxt);
1019 1100 : dlist_init(&DatabaseList);
1020 :
1021 1100 : if (nelems > 0)
1022 : {
1023 : TimestampTz current_time;
1024 : int millis_increment;
1025 : avl_dbase *dbary;
1026 : avl_dbase *db;
1027 : HASH_SEQ_STATUS seq;
1028 : int i;
1029 :
1030 : /* put all the hash elements into an array */
1031 1062 : dbary = palloc(nelems * sizeof(avl_dbase));
1032 : /* keep Valgrind quiet */
1033 : #ifdef USE_VALGRIND
1034 : avl_dbase_array = dbary;
1035 : #endif
1036 :
1037 1062 : i = 0;
1038 1062 : hash_seq_init(&seq, dbhash);
1039 3032 : while ((db = hash_seq_search(&seq)) != NULL)
1040 1970 : memcpy(&(dbary[i++]), db, sizeof(avl_dbase));
1041 :
1042 : /* sort the array */
1043 1062 : qsort(dbary, nelems, sizeof(avl_dbase), db_comparator);
1044 :
1045 : /*
1046 : * Determine the time interval between databases in the schedule. If
1047 : * we see that the configured naptime would take us to sleep times
1048 : * lower than our min sleep time (which launcher_determine_sleep is
1049 : * coded not to allow), silently use a larger naptime (but don't touch
1050 : * the GUC variable).
1051 : */
1052 1062 : millis_increment = 1000.0 * autovacuum_naptime / nelems;
1053 1062 : if (millis_increment <= MIN_AUTOVAC_SLEEPTIME)
1054 0 : millis_increment = MIN_AUTOVAC_SLEEPTIME * 1.1;
1055 :
1056 1062 : current_time = GetCurrentTimestamp();
1057 :
1058 : /*
1059 : * move the elements from the array into the dlist, setting the
1060 : * next_worker while walking the array
1061 : */
1062 3032 : for (i = 0; i < nelems; i++)
1063 : {
1064 1970 : db = &(dbary[i]);
1065 :
1066 1970 : current_time = TimestampTzPlusMilliseconds(current_time,
1067 : millis_increment);
1068 1970 : db->adl_next_worker = current_time;
1069 :
1070 : /* later elements should go closer to the head of the list */
1071 1970 : dlist_push_head(&DatabaseList, &db->adl_node);
1072 : }
1073 : }
1074 :
1075 : /* all done, clean up memory */
1076 1100 : if (DatabaseListCxt != NULL)
1077 340 : MemoryContextDelete(DatabaseListCxt);
1078 1100 : MemoryContextDelete(tmpcxt);
1079 1100 : DatabaseListCxt = newcxt;
1080 1100 : MemoryContextSwitchTo(oldcxt);
1081 1100 : }
1082 :
1083 : /* qsort comparator for avl_dbase, using adl_score */
1084 : static int
1085 1310 : db_comparator(const void *a, const void *b)
1086 : {
1087 2620 : return pg_cmp_s32(((const avl_dbase *) a)->adl_score,
1088 1310 : ((const avl_dbase *) b)->adl_score);
1089 : }
1090 :
1091 : /*
1092 : * do_start_worker
1093 : *
1094 : * Bare-bones procedure for starting an autovacuum worker from the launcher.
1095 : * It determines what database to work on, sets up shared memory stuff and
1096 : * signals postmaster to start the worker. It fails gracefully if invoked when
1097 : * autovacuum_workers are already active.
1098 : *
1099 : * Return value is the OID of the database that the worker is going to process,
1100 : * or InvalidOid if no worker was actually started.
1101 : */
1102 : static Oid
1103 4740 : do_start_worker(void)
1104 : {
1105 : List *dblist;
1106 : ListCell *cell;
1107 : TransactionId xidForceLimit;
1108 : MultiXactId multiForceLimit;
1109 : bool for_xid_wrap;
1110 : bool for_multi_wrap;
1111 : avw_dbase *avdb;
1112 : TimestampTz current_time;
1113 4740 : bool skipit = false;
1114 4740 : Oid retval = InvalidOid;
1115 : MemoryContext tmpcxt,
1116 : oldcxt;
1117 :
1118 : /* return quickly when there are no free workers */
1119 4740 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1120 4740 : if (!av_worker_available())
1121 : {
1122 0 : LWLockRelease(AutovacuumLock);
1123 0 : return InvalidOid;
1124 : }
1125 4740 : LWLockRelease(AutovacuumLock);
1126 :
1127 : /*
1128 : * Create and switch to a temporary context to avoid leaking the memory
1129 : * allocated for the database list.
1130 : */
1131 4740 : tmpcxt = AllocSetContextCreate(CurrentMemoryContext,
1132 : "Autovacuum start worker (tmp)",
1133 : ALLOCSET_DEFAULT_SIZES);
1134 4740 : oldcxt = MemoryContextSwitchTo(tmpcxt);
1135 :
1136 : /* Get a list of databases */
1137 4740 : dblist = get_database_list();
1138 :
1139 : /*
1140 : * Determine the oldest datfrozenxid/relfrozenxid that we will allow to
1141 : * pass without forcing a vacuum. (This limit can be tightened for
1142 : * particular tables, but not loosened.)
1143 : */
1144 4740 : recentXid = ReadNextTransactionId();
1145 4740 : xidForceLimit = recentXid - autovacuum_freeze_max_age;
1146 : /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */
1147 : /* this can cause the limit to go backwards by 3, but that's OK */
1148 4740 : if (xidForceLimit < FirstNormalTransactionId)
1149 0 : xidForceLimit -= FirstNormalTransactionId;
1150 :
1151 : /* Also determine the oldest datminmxid we will consider. */
1152 4740 : recentMulti = ReadNextMultiXactId();
1153 4740 : multiForceLimit = recentMulti - MultiXactMemberFreezeThreshold();
1154 4740 : if (multiForceLimit < FirstMultiXactId)
1155 0 : multiForceLimit -= FirstMultiXactId;
1156 :
1157 : /*
1158 : * Choose a database to connect to. We pick the database that was least
1159 : * recently auto-vacuumed, or one that needs vacuuming to prevent Xid
1160 : * wraparound-related data loss. If any db at risk of Xid wraparound is
1161 : * found, we pick the one with oldest datfrozenxid, independently of
1162 : * autovacuum times; similarly we pick the one with the oldest datminmxid
1163 : * if any is in MultiXactId wraparound. Note that those in Xid wraparound
1164 : * danger are given more priority than those in multi wraparound danger.
1165 : *
1166 : * Note that a database with no stats entry is not considered, except for
1167 : * Xid wraparound purposes. The theory is that if no one has ever
1168 : * connected to it since the stats were last initialized, it doesn't need
1169 : * vacuuming.
1170 : *
1171 : * XXX This could be improved if we had more info about whether it needs
1172 : * vacuuming before connecting to it. Perhaps look through the pgstats
1173 : * data for the database's tables? One idea is to keep track of the
1174 : * number of new and dead tuples per database in pgstats. However it
1175 : * isn't clear how to construct a metric that measures that and not cause
1176 : * starvation for less busy databases.
1177 : */
1178 4740 : avdb = NULL;
1179 4740 : for_xid_wrap = false;
1180 4740 : for_multi_wrap = false;
1181 4740 : current_time = GetCurrentTimestamp();
1182 19128 : foreach(cell, dblist)
1183 : {
1184 14388 : avw_dbase *tmp = lfirst(cell);
1185 : dlist_iter iter;
1186 :
1187 : /* Check to see if this one is at risk of wraparound */
1188 14388 : if (TransactionIdPrecedes(tmp->adw_frozenxid, xidForceLimit))
1189 : {
1190 11434 : if (avdb == NULL ||
1191 4548 : TransactionIdPrecedes(tmp->adw_frozenxid,
1192 : avdb->adw_frozenxid))
1193 2502 : avdb = tmp;
1194 6886 : for_xid_wrap = true;
1195 11678 : continue;
1196 : }
1197 7502 : else if (for_xid_wrap)
1198 120 : continue; /* ignore not-at-risk DBs */
1199 7382 : else if (MultiXactIdPrecedes(tmp->adw_minmulti, multiForceLimit))
1200 : {
1201 0 : if (avdb == NULL ||
1202 0 : MultiXactIdPrecedes(tmp->adw_minmulti, avdb->adw_minmulti))
1203 0 : avdb = tmp;
1204 0 : for_multi_wrap = true;
1205 0 : continue;
1206 : }
1207 7382 : else if (for_multi_wrap)
1208 0 : continue; /* ignore not-at-risk DBs */
1209 :
1210 : /* Find pgstat entry if any */
1211 7382 : tmp->adw_entry = pgstat_fetch_stat_dbentry(tmp->adw_datid);
1212 :
1213 : /*
1214 : * Skip a database with no pgstat entry; it means it hasn't seen any
1215 : * activity.
1216 : */
1217 7382 : if (!tmp->adw_entry)
1218 180 : continue;
1219 :
1220 : /*
1221 : * Also, skip a database that appears on the database list as having
1222 : * been processed recently (less than autovacuum_naptime seconds ago).
1223 : * We do this so that we don't select a database which we just
1224 : * selected, but that pgstat hasn't gotten around to updating the last
1225 : * autovacuum time yet.
1226 : */
1227 7202 : skipit = false;
1228 :
1229 14620 : dlist_reverse_foreach(iter, &DatabaseList)
1230 : {
1231 14574 : avl_dbase *dbp = dlist_container(avl_dbase, adl_node, iter.cur);
1232 :
1233 14574 : if (dbp->adl_datid == tmp->adw_datid)
1234 : {
1235 : /*
1236 : * Skip this database if its next_worker value falls between
1237 : * the current time and the current time plus naptime.
1238 : */
1239 7156 : if (!TimestampDifferenceExceeds(dbp->adl_next_worker,
1240 4492 : current_time, 0) &&
1241 4492 : !TimestampDifferenceExceeds(current_time,
1242 : dbp->adl_next_worker,
1243 : autovacuum_naptime * 1000))
1244 4492 : skipit = true;
1245 :
1246 7156 : break;
1247 : }
1248 : }
1249 7202 : if (skipit)
1250 4492 : continue;
1251 :
1252 : /*
1253 : * Remember the db with oldest autovac time. (If we are here, both
1254 : * tmp->entry and db->entry must be non-null.)
1255 : */
1256 2710 : if (avdb == NULL ||
1257 318 : tmp->adw_entry->last_autovac_time < avdb->adw_entry->last_autovac_time)
1258 2506 : avdb = tmp;
1259 : }
1260 :
1261 : /* Found a database -- process it */
1262 4740 : if (avdb != NULL)
1263 : {
1264 : WorkerInfo worker;
1265 : dlist_node *wptr;
1266 :
1267 4730 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1268 :
1269 : /*
1270 : * Get a worker entry from the freelist. We checked above, so there
1271 : * really should be a free slot.
1272 : */
1273 4730 : wptr = dclist_pop_head_node(&AutoVacuumShmem->av_freeWorkers);
1274 :
1275 4730 : worker = dlist_container(WorkerInfoData, wi_links, wptr);
1276 4730 : worker->wi_dboid = avdb->adw_datid;
1277 4730 : worker->wi_proc = NULL;
1278 4730 : worker->wi_launchtime = GetCurrentTimestamp();
1279 :
1280 4730 : AutoVacuumShmem->av_startingWorker = worker;
1281 :
1282 4730 : LWLockRelease(AutovacuumLock);
1283 :
1284 4730 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
1285 :
1286 4730 : retval = avdb->adw_datid;
1287 : }
1288 10 : else if (skipit)
1289 : {
1290 : /*
1291 : * If we skipped all databases on the list, rebuild it, because it
1292 : * probably contains a dropped database.
1293 : */
1294 0 : rebuild_database_list(InvalidOid);
1295 : }
1296 :
1297 4740 : MemoryContextSwitchTo(oldcxt);
1298 4740 : MemoryContextDelete(tmpcxt);
1299 :
1300 4740 : return retval;
1301 : }
1302 :
1303 : /*
1304 : * launch_worker
1305 : *
1306 : * Wrapper for starting a worker from the launcher. Besides actually starting
1307 : * it, update the database list to reflect the next time that another one will
1308 : * need to be started on the selected database. The actual database choice is
1309 : * left to do_start_worker.
1310 : *
1311 : * This routine is also expected to insert an entry into the database list if
1312 : * the selected database was previously absent from the list.
1313 : */
1314 : static void
1315 4740 : launch_worker(TimestampTz now)
1316 : {
1317 : Oid dbid;
1318 : dlist_iter iter;
1319 :
1320 4740 : dbid = do_start_worker();
1321 4740 : if (OidIsValid(dbid))
1322 : {
1323 4730 : bool found = false;
1324 :
1325 : /*
1326 : * Walk the database list and update the corresponding entry. If the
1327 : * database is not on the list, we'll recreate the list.
1328 : */
1329 10188 : dlist_foreach(iter, &DatabaseList)
1330 : {
1331 10160 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1332 :
1333 10160 : if (avdb->adl_datid == dbid)
1334 : {
1335 4702 : found = true;
1336 :
1337 : /*
1338 : * add autovacuum_naptime seconds to the current time, and use
1339 : * that as the new "next_worker" field for this database.
1340 : */
1341 4702 : avdb->adl_next_worker =
1342 4702 : TimestampTzPlusMilliseconds(now, autovacuum_naptime * 1000);
1343 :
1344 4702 : dlist_move_head(&DatabaseList, iter.cur);
1345 4702 : break;
1346 : }
1347 : }
1348 :
1349 : /*
1350 : * If the database was not present in the database list, we rebuild
1351 : * the list. It's possible that the database does not get into the
1352 : * list anyway, for example if it's a database that doesn't have a
1353 : * pgstat entry, but this is not a problem because we don't want to
1354 : * schedule workers regularly into those in any case.
1355 : */
1356 4730 : if (!found)
1357 28 : rebuild_database_list(dbid);
1358 : }
1359 4740 : }
1360 :
1361 : /*
1362 : * Called from postmaster to signal a failure to fork a process to become
1363 : * worker. The postmaster should kill(SIGUSR2) the launcher shortly
1364 : * after calling this function.
1365 : */
1366 : void
1367 0 : AutoVacWorkerFailed(void)
1368 : {
1369 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = true;
1370 0 : }
1371 :
1372 : /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */
1373 : static void
1374 9446 : avl_sigusr2_handler(SIGNAL_ARGS)
1375 : {
1376 9446 : got_SIGUSR2 = true;
1377 9446 : SetLatch(MyLatch);
1378 9446 : }
1379 :
1380 :
1381 : /********************************************************************
1382 : * AUTOVACUUM WORKER CODE
1383 : ********************************************************************/
1384 :
1385 : /*
1386 : * Main entry point for autovacuum worker processes.
1387 : */
1388 : void
1389 4732 : AutoVacWorkerMain(const void *startup_data, size_t startup_data_len)
1390 : {
1391 : sigjmp_buf local_sigjmp_buf;
1392 : Oid dbid;
1393 :
1394 : Assert(startup_data_len == 0);
1395 :
1396 : /* Release postmaster's working memory context */
1397 4732 : if (PostmasterContext)
1398 : {
1399 4732 : MemoryContextDelete(PostmasterContext);
1400 4732 : PostmasterContext = NULL;
1401 : }
1402 :
1403 4732 : MyBackendType = B_AUTOVAC_WORKER;
1404 4732 : init_ps_display(NULL);
1405 :
1406 : Assert(GetProcessingMode() == InitProcessing);
1407 :
1408 : /*
1409 : * Set up signal handlers. We operate on databases much like a regular
1410 : * backend, so we use the same signal handling. See equivalent code in
1411 : * tcop/postgres.c.
1412 : */
1413 4732 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1414 :
1415 : /*
1416 : * SIGINT is used to signal canceling the current table's vacuum; SIGTERM
1417 : * means abort and exit cleanly, and SIGQUIT means abandon ship.
1418 : */
1419 4732 : pqsignal(SIGINT, StatementCancelHandler);
1420 4732 : pqsignal(SIGTERM, die);
1421 : /* SIGQUIT handler was already set up by InitPostmasterChild */
1422 :
1423 4732 : InitializeTimeouts(); /* establishes SIGALRM handler */
1424 :
1425 4732 : pqsignal(SIGPIPE, SIG_IGN);
1426 4732 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1427 4732 : pqsignal(SIGUSR2, SIG_IGN);
1428 4732 : pqsignal(SIGFPE, FloatExceptionHandler);
1429 4732 : pqsignal(SIGCHLD, SIG_DFL);
1430 :
1431 : /*
1432 : * Create a per-backend PGPROC struct in shared memory. We must do this
1433 : * before we can use LWLocks or access any shared memory.
1434 : */
1435 4732 : InitProcess();
1436 :
1437 : /* Early initialization */
1438 4732 : BaseInit();
1439 :
1440 : /*
1441 : * If an exception is encountered, processing resumes here.
1442 : *
1443 : * Unlike most auxiliary processes, we don't attempt to continue
1444 : * processing after an error; we just clean up and exit. The autovac
1445 : * launcher is responsible for spawning another worker later.
1446 : *
1447 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
1448 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
1449 : * signals other than SIGQUIT will be blocked until we exit. It might
1450 : * seem that this policy makes the HOLD_INTERRUPTS() call redundant, but
1451 : * it is not since InterruptPending might be set already.
1452 : */
1453 4732 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1454 : {
1455 : /* since not using PG_TRY, must reset error stack by hand */
1456 0 : error_context_stack = NULL;
1457 :
1458 : /* Prevents interrupts while cleaning up */
1459 0 : HOLD_INTERRUPTS();
1460 :
1461 : /* Report the error to the server log */
1462 0 : EmitErrorReport();
1463 :
1464 : /*
1465 : * We can now go away. Note that because we called InitProcess, a
1466 : * callback was registered to do ProcKill, which will clean up
1467 : * necessary state.
1468 : */
1469 0 : proc_exit(0);
1470 : }
1471 :
1472 : /* We can now handle ereport(ERROR) */
1473 4732 : PG_exception_stack = &local_sigjmp_buf;
1474 :
1475 4732 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
1476 :
1477 : /*
1478 : * Set always-secure search path, so malicious users can't redirect user
1479 : * code (e.g. pg_index.indexprs). (That code runs in a
1480 : * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not
1481 : * take control of the entire autovacuum worker in any case.)
1482 : */
1483 4732 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1484 :
1485 : /*
1486 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
1487 : * in postgresql.conf. We don't really want such a dangerous option being
1488 : * applied non-interactively.
1489 : */
1490 4732 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
1491 :
1492 : /*
1493 : * Force settable timeouts off to avoid letting these settings prevent
1494 : * regular maintenance from being executed.
1495 : */
1496 4732 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1497 4732 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1498 4732 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1499 4732 : SetConfigOption("idle_in_transaction_session_timeout", "0",
1500 : PGC_SUSET, PGC_S_OVERRIDE);
1501 :
1502 : /*
1503 : * Force default_transaction_isolation to READ COMMITTED. We don't want
1504 : * to pay the overhead of serializable mode, nor add any risk of causing
1505 : * deadlocks or delaying other transactions.
1506 : */
1507 4732 : SetConfigOption("default_transaction_isolation", "read committed",
1508 : PGC_SUSET, PGC_S_OVERRIDE);
1509 :
1510 : /*
1511 : * Force synchronous replication off to allow regular maintenance even if
1512 : * we are waiting for standbys to connect. This is important to ensure we
1513 : * aren't blocked from performing anti-wraparound tasks.
1514 : */
1515 4732 : if (synchronous_commit > SYNCHRONOUS_COMMIT_LOCAL_FLUSH)
1516 4732 : SetConfigOption("synchronous_commit", "local",
1517 : PGC_SUSET, PGC_S_OVERRIDE);
1518 :
1519 : /*
1520 : * Even when system is configured to use a different fetch consistency,
1521 : * for autovac we always want fresh stats.
1522 : */
1523 4732 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
1524 :
1525 : /*
1526 : * Get the info about the database we're going to work on.
1527 : */
1528 4732 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1529 :
1530 : /*
1531 : * beware of startingWorker being INVALID; this should normally not
1532 : * happen, but if a worker fails after forking and before this, the
1533 : * launcher might have decided to remove it from the queue and start
1534 : * again.
1535 : */
1536 4732 : if (AutoVacuumShmem->av_startingWorker != NULL)
1537 : {
1538 4732 : MyWorkerInfo = AutoVacuumShmem->av_startingWorker;
1539 4732 : dbid = MyWorkerInfo->wi_dboid;
1540 4732 : MyWorkerInfo->wi_proc = MyProc;
1541 :
1542 : /* insert into the running list */
1543 4732 : dlist_push_head(&AutoVacuumShmem->av_runningWorkers,
1544 4732 : &MyWorkerInfo->wi_links);
1545 :
1546 : /*
1547 : * remove from the "starting" pointer, so that the launcher can start
1548 : * a new worker if required
1549 : */
1550 4732 : AutoVacuumShmem->av_startingWorker = NULL;
1551 4732 : LWLockRelease(AutovacuumLock);
1552 :
1553 4732 : on_shmem_exit(FreeWorkerInfo, 0);
1554 :
1555 : /* wake up the launcher */
1556 4732 : if (AutoVacuumShmem->av_launcherpid != 0)
1557 4732 : kill(AutoVacuumShmem->av_launcherpid, SIGUSR2);
1558 : }
1559 : else
1560 : {
1561 : /* no worker entry for me, go away */
1562 0 : elog(WARNING, "autovacuum worker started without a worker entry");
1563 0 : dbid = InvalidOid;
1564 0 : LWLockRelease(AutovacuumLock);
1565 : }
1566 :
1567 4732 : if (OidIsValid(dbid))
1568 : {
1569 : char dbname[NAMEDATALEN];
1570 :
1571 : /*
1572 : * Report autovac startup to the cumulative stats system. We
1573 : * deliberately do this before InitPostgres, so that the
1574 : * last_autovac_time will get updated even if the connection attempt
1575 : * fails. This is to prevent autovac from getting "stuck" repeatedly
1576 : * selecting an unopenable database, rather than making any progress
1577 : * on stuff it can connect to.
1578 : */
1579 4732 : pgstat_report_autovac(dbid);
1580 :
1581 : /*
1582 : * Connect to the selected database, specifying no particular user,
1583 : * and ignoring datallowconn. Collect the database's name for
1584 : * display.
1585 : *
1586 : * Note: if we have selected a just-deleted database (due to using
1587 : * stale stats info), we'll fail and exit here.
1588 : */
1589 4732 : InitPostgres(NULL, dbid, NULL, InvalidOid,
1590 : INIT_PG_OVERRIDE_ALLOW_CONNS,
1591 : dbname);
1592 4732 : SetProcessingMode(NormalProcessing);
1593 4732 : set_ps_display(dbname);
1594 4732 : ereport(DEBUG1,
1595 : (errmsg_internal("autovacuum: processing database \"%s\"", dbname)));
1596 :
1597 4732 : if (PostAuthDelay)
1598 0 : pg_usleep(PostAuthDelay * 1000000L);
1599 :
1600 : /* And do an appropriate amount of work */
1601 4732 : recentXid = ReadNextTransactionId();
1602 4732 : recentMulti = ReadNextMultiXactId();
1603 4732 : do_autovacuum();
1604 : }
1605 :
1606 : /*
1607 : * The launcher will be notified of my death in ProcKill, *if* we managed
1608 : * to get a worker slot at all
1609 : */
1610 :
1611 : /* All done, go away */
1612 4730 : proc_exit(0);
1613 : }
1614 :
1615 : /*
1616 : * Return a WorkerInfo to the free list
1617 : */
1618 : static void
1619 4732 : FreeWorkerInfo(int code, Datum arg)
1620 : {
1621 4732 : if (MyWorkerInfo != NULL)
1622 : {
1623 4732 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1624 :
1625 : /*
1626 : * Wake the launcher up so that he can launch a new worker immediately
1627 : * if required. We only save the launcher's PID in local memory here;
1628 : * the actual signal will be sent when the PGPROC is recycled. Note
1629 : * that we always do this, so that the launcher can rebalance the cost
1630 : * limit setting of the remaining workers.
1631 : *
1632 : * We somewhat ignore the risk that the launcher changes its PID
1633 : * between us reading it and the actual kill; we expect ProcKill to be
1634 : * called shortly after us, and we assume that PIDs are not reused too
1635 : * quickly after a process exits.
1636 : */
1637 4732 : AutovacuumLauncherPid = AutoVacuumShmem->av_launcherpid;
1638 :
1639 4732 : dlist_delete(&MyWorkerInfo->wi_links);
1640 4732 : MyWorkerInfo->wi_dboid = InvalidOid;
1641 4732 : MyWorkerInfo->wi_tableoid = InvalidOid;
1642 4732 : MyWorkerInfo->wi_sharedrel = false;
1643 4732 : MyWorkerInfo->wi_proc = NULL;
1644 4732 : MyWorkerInfo->wi_launchtime = 0;
1645 4732 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
1646 4732 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
1647 4732 : &MyWorkerInfo->wi_links);
1648 : /* not mine anymore */
1649 4732 : MyWorkerInfo = NULL;
1650 :
1651 : /*
1652 : * now that we're inactive, cause a rebalancing of the surviving
1653 : * workers
1654 : */
1655 4732 : AutoVacuumShmem->av_signal[AutoVacRebalance] = true;
1656 4732 : LWLockRelease(AutovacuumLock);
1657 : }
1658 4732 : }
1659 :
1660 : /*
1661 : * Update vacuum cost-based delay-related parameters for autovacuum workers and
1662 : * backends executing VACUUM or ANALYZE using the value of relevant GUCs and
1663 : * global state. This must be called during setup for vacuum and after every
1664 : * config reload to ensure up-to-date values.
1665 : */
1666 : void
1667 537944 : VacuumUpdateCosts(void)
1668 : {
1669 537944 : if (MyWorkerInfo)
1670 : {
1671 524280 : if (av_storage_param_cost_delay >= 0)
1672 0 : vacuum_cost_delay = av_storage_param_cost_delay;
1673 524280 : else if (autovacuum_vac_cost_delay >= 0)
1674 524280 : vacuum_cost_delay = autovacuum_vac_cost_delay;
1675 : else
1676 : /* fall back to VacuumCostDelay */
1677 0 : vacuum_cost_delay = VacuumCostDelay;
1678 :
1679 524280 : AutoVacuumUpdateCostLimit();
1680 : }
1681 : else
1682 : {
1683 : /* Must be explicit VACUUM or ANALYZE */
1684 13664 : vacuum_cost_delay = VacuumCostDelay;
1685 13664 : vacuum_cost_limit = VacuumCostLimit;
1686 : }
1687 :
1688 : /*
1689 : * If configuration changes are allowed to impact VacuumCostActive, make
1690 : * sure it is updated.
1691 : */
1692 537944 : if (VacuumFailsafeActive)
1693 : Assert(!VacuumCostActive);
1694 537944 : else if (vacuum_cost_delay > 0)
1695 524280 : VacuumCostActive = true;
1696 : else
1697 : {
1698 13664 : VacuumCostActive = false;
1699 13664 : VacuumCostBalance = 0;
1700 : }
1701 :
1702 : /*
1703 : * Since the cost logging requires a lock, avoid rendering the log message
1704 : * in case we are using a message level where the log wouldn't be emitted.
1705 : */
1706 537944 : if (MyWorkerInfo && message_level_is_interesting(DEBUG2))
1707 : {
1708 : Oid dboid,
1709 : tableoid;
1710 :
1711 : Assert(!LWLockHeldByMe(AutovacuumLock));
1712 :
1713 0 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1714 0 : dboid = MyWorkerInfo->wi_dboid;
1715 0 : tableoid = MyWorkerInfo->wi_tableoid;
1716 0 : LWLockRelease(AutovacuumLock);
1717 :
1718 0 : elog(DEBUG2,
1719 : "Autovacuum VacuumUpdateCosts(db=%u, rel=%u, dobalance=%s, cost_limit=%d, cost_delay=%g active=%s failsafe=%s)",
1720 : dboid, tableoid, pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance) ? "no" : "yes",
1721 : vacuum_cost_limit, vacuum_cost_delay,
1722 : vacuum_cost_delay > 0 ? "yes" : "no",
1723 : VacuumFailsafeActive ? "yes" : "no");
1724 : }
1725 537944 : }
1726 :
1727 : /*
1728 : * Update vacuum_cost_limit with the correct value for an autovacuum worker,
1729 : * given the value of other relevant cost limit parameters and the number of
1730 : * workers across which the limit must be balanced. Autovacuum workers must
1731 : * call this regularly in case av_nworkersForBalance has been updated by
1732 : * another worker or by the autovacuum launcher. They must also call it after a
1733 : * config reload.
1734 : */
1735 : void
1736 535824 : AutoVacuumUpdateCostLimit(void)
1737 : {
1738 535824 : if (!MyWorkerInfo)
1739 0 : return;
1740 :
1741 : /*
1742 : * note: in cost_limit, zero also means use value from elsewhere, because
1743 : * zero is not a valid value.
1744 : */
1745 :
1746 535824 : if (av_storage_param_cost_limit > 0)
1747 0 : vacuum_cost_limit = av_storage_param_cost_limit;
1748 : else
1749 : {
1750 : int nworkers_for_balance;
1751 :
1752 535824 : if (autovacuum_vac_cost_limit > 0)
1753 0 : vacuum_cost_limit = autovacuum_vac_cost_limit;
1754 : else
1755 535824 : vacuum_cost_limit = VacuumCostLimit;
1756 :
1757 : /* Only balance limit if no cost-related storage parameters specified */
1758 535824 : if (pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance))
1759 0 : return;
1760 :
1761 : Assert(vacuum_cost_limit > 0);
1762 :
1763 535824 : nworkers_for_balance = pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1764 :
1765 : /* There is at least 1 autovac worker (this worker) */
1766 535824 : if (nworkers_for_balance <= 0)
1767 0 : elog(ERROR, "nworkers_for_balance must be > 0");
1768 :
1769 535824 : vacuum_cost_limit = Max(vacuum_cost_limit / nworkers_for_balance, 1);
1770 : }
1771 : }
1772 :
1773 : /*
1774 : * autovac_recalculate_workers_for_balance
1775 : * Recalculate the number of workers to consider, given cost-related
1776 : * storage parameters and the current number of active workers.
1777 : *
1778 : * Caller must hold the AutovacuumLock in at least shared mode to access
1779 : * worker->wi_proc.
1780 : */
1781 : static void
1782 266858 : autovac_recalculate_workers_for_balance(void)
1783 : {
1784 : dlist_iter iter;
1785 : int orig_nworkers_for_balance;
1786 266858 : int nworkers_for_balance = 0;
1787 :
1788 : Assert(LWLockHeldByMe(AutovacuumLock));
1789 :
1790 266858 : orig_nworkers_for_balance =
1791 266858 : pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1792 :
1793 827904 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1794 : {
1795 561046 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1796 :
1797 1122092 : if (worker->wi_proc == NULL ||
1798 561046 : pg_atomic_unlocked_test_flag(&worker->wi_dobalance))
1799 12144 : continue;
1800 :
1801 548902 : nworkers_for_balance++;
1802 : }
1803 :
1804 266858 : if (nworkers_for_balance != orig_nworkers_for_balance)
1805 5110 : pg_atomic_write_u32(&AutoVacuumShmem->av_nworkersForBalance,
1806 : nworkers_for_balance);
1807 266858 : }
1808 :
1809 : /*
1810 : * get_database_list
1811 : * Return a list of all databases found in pg_database.
1812 : *
1813 : * The list and associated data is allocated in the caller's memory context,
1814 : * which is in charge of ensuring that it's properly cleaned up afterwards.
1815 : *
1816 : * Note: this is the only function in which the autovacuum launcher uses a
1817 : * transaction. Although we aren't attached to any particular database and
1818 : * therefore can't access most catalogs, we do have enough infrastructure
1819 : * to do a seqscan on pg_database.
1820 : */
1821 : static List *
1822 5840 : get_database_list(void)
1823 : {
1824 5840 : List *dblist = NIL;
1825 : Relation rel;
1826 : TableScanDesc scan;
1827 : HeapTuple tup;
1828 : MemoryContext resultcxt;
1829 :
1830 : /* This is the context that we will allocate our output data in */
1831 5840 : resultcxt = CurrentMemoryContext;
1832 :
1833 : /*
1834 : * Start a transaction so we can access pg_database.
1835 : */
1836 5840 : StartTransactionCommand();
1837 :
1838 5840 : rel = table_open(DatabaseRelationId, AccessShareLock);
1839 5840 : scan = table_beginscan_catalog(rel, 0, NULL);
1840 :
1841 23860 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1842 : {
1843 18020 : Form_pg_database pgdatabase = (Form_pg_database) GETSTRUCT(tup);
1844 : avw_dbase *avdb;
1845 : MemoryContext oldcxt;
1846 :
1847 : /*
1848 : * If database has partially been dropped, we can't, nor need to,
1849 : * vacuum it.
1850 : */
1851 18020 : if (database_is_invalid_form(pgdatabase))
1852 : {
1853 12 : elog(DEBUG2,
1854 : "autovacuum: skipping invalid database \"%s\"",
1855 : NameStr(pgdatabase->datname));
1856 12 : continue;
1857 : }
1858 :
1859 : /*
1860 : * Allocate our results in the caller's context, not the
1861 : * transaction's. We do this inside the loop, and restore the original
1862 : * context at the end, so that leaky things like heap_getnext() are
1863 : * not called in a potentially long-lived context.
1864 : */
1865 18008 : oldcxt = MemoryContextSwitchTo(resultcxt);
1866 :
1867 18008 : avdb = (avw_dbase *) palloc(sizeof(avw_dbase));
1868 :
1869 18008 : avdb->adw_datid = pgdatabase->oid;
1870 18008 : avdb->adw_name = pstrdup(NameStr(pgdatabase->datname));
1871 18008 : avdb->adw_frozenxid = pgdatabase->datfrozenxid;
1872 18008 : avdb->adw_minmulti = pgdatabase->datminmxid;
1873 : /* this gets set later: */
1874 18008 : avdb->adw_entry = NULL;
1875 :
1876 18008 : dblist = lappend(dblist, avdb);
1877 18008 : MemoryContextSwitchTo(oldcxt);
1878 : }
1879 :
1880 5840 : table_endscan(scan);
1881 5840 : table_close(rel, AccessShareLock);
1882 :
1883 5840 : CommitTransactionCommand();
1884 :
1885 : /* Be sure to restore caller's memory context */
1886 5840 : MemoryContextSwitchTo(resultcxt);
1887 :
1888 5840 : return dblist;
1889 : }
1890 :
1891 : /*
1892 : * Process a database table-by-table
1893 : *
1894 : * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in
1895 : * order not to ignore shutdown commands for too long.
1896 : */
1897 : static void
1898 4732 : do_autovacuum(void)
1899 : {
1900 : Relation classRel;
1901 : HeapTuple tuple;
1902 : TableScanDesc relScan;
1903 : Form_pg_database dbForm;
1904 4732 : List *table_oids = NIL;
1905 4732 : List *orphan_oids = NIL;
1906 : HASHCTL ctl;
1907 : HTAB *table_toast_map;
1908 : ListCell *volatile cell;
1909 : BufferAccessStrategy bstrategy;
1910 : ScanKeyData key;
1911 : TupleDesc pg_class_desc;
1912 : int effective_multixact_freeze_max_age;
1913 4732 : bool did_vacuum = false;
1914 4732 : bool found_concurrent_worker = false;
1915 : int i;
1916 :
1917 : /*
1918 : * StartTransactionCommand and CommitTransactionCommand will automatically
1919 : * switch to other contexts. We need this one to keep the list of
1920 : * relations to vacuum/analyze across transactions.
1921 : */
1922 4732 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
1923 : "Autovacuum worker",
1924 : ALLOCSET_DEFAULT_SIZES);
1925 4732 : MemoryContextSwitchTo(AutovacMemCxt);
1926 :
1927 : /* Start a transaction so our commands have one to play into. */
1928 4732 : StartTransactionCommand();
1929 :
1930 : /*
1931 : * This injection point is put in a transaction block to work with a wait
1932 : * that uses a condition variable.
1933 : */
1934 4732 : INJECTION_POINT("autovacuum-worker-start", NULL);
1935 :
1936 : /*
1937 : * Compute the multixact age for which freezing is urgent. This is
1938 : * normally autovacuum_multixact_freeze_max_age, but may be less if we are
1939 : * short of multixact member space.
1940 : */
1941 4730 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1942 :
1943 : /*
1944 : * Find the pg_database entry and select the default freeze ages. We use
1945 : * zero in template and nonconnectable databases, else the system-wide
1946 : * default.
1947 : */
1948 4730 : tuple = SearchSysCache1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
1949 4730 : if (!HeapTupleIsValid(tuple))
1950 0 : elog(ERROR, "cache lookup failed for database %u", MyDatabaseId);
1951 4730 : dbForm = (Form_pg_database) GETSTRUCT(tuple);
1952 :
1953 4730 : if (dbForm->datistemplate || !dbForm->datallowconn)
1954 : {
1955 1740 : default_freeze_min_age = 0;
1956 1740 : default_freeze_table_age = 0;
1957 1740 : default_multixact_freeze_min_age = 0;
1958 1740 : default_multixact_freeze_table_age = 0;
1959 : }
1960 : else
1961 : {
1962 2990 : default_freeze_min_age = vacuum_freeze_min_age;
1963 2990 : default_freeze_table_age = vacuum_freeze_table_age;
1964 2990 : default_multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1965 2990 : default_multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1966 : }
1967 :
1968 4730 : ReleaseSysCache(tuple);
1969 :
1970 : /* StartTransactionCommand changed elsewhere */
1971 4730 : MemoryContextSwitchTo(AutovacMemCxt);
1972 :
1973 4730 : classRel = table_open(RelationRelationId, AccessShareLock);
1974 :
1975 : /* create a copy so we can use it after closing pg_class */
1976 4730 : pg_class_desc = CreateTupleDescCopy(RelationGetDescr(classRel));
1977 :
1978 : /* create hash table for toast <-> main relid mapping */
1979 4730 : ctl.keysize = sizeof(Oid);
1980 4730 : ctl.entrysize = sizeof(av_relation);
1981 :
1982 4730 : table_toast_map = hash_create("TOAST to main relid map",
1983 : 100,
1984 : &ctl,
1985 : HASH_ELEM | HASH_BLOBS);
1986 :
1987 : /*
1988 : * Scan pg_class to determine which tables to vacuum.
1989 : *
1990 : * We do this in two passes: on the first one we collect the list of plain
1991 : * relations and materialized views, and on the second one we collect
1992 : * TOAST tables. The reason for doing the second pass is that during it we
1993 : * want to use the main relation's pg_class.reloptions entry if the TOAST
1994 : * table does not have any, and we cannot obtain it unless we know
1995 : * beforehand what's the main table OID.
1996 : *
1997 : * We need to check TOAST tables separately because in cases with short,
1998 : * wide tables there might be proportionally much more activity in the
1999 : * TOAST table than in its parent.
2000 : */
2001 4730 : relScan = table_beginscan_catalog(classRel, 0, NULL);
2002 :
2003 : /*
2004 : * On the first pass, we collect main tables to vacuum, and also the main
2005 : * table relid to TOAST relid mapping.
2006 : */
2007 2000074 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2008 : {
2009 1995344 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2010 : PgStat_StatTabEntry *tabentry;
2011 : AutoVacOpts *relopts;
2012 : Oid relid;
2013 : bool dovacuum;
2014 : bool doanalyze;
2015 : bool wraparound;
2016 :
2017 1995344 : if (classForm->relkind != RELKIND_RELATION &&
2018 1664842 : classForm->relkind != RELKIND_MATVIEW)
2019 1664746 : continue;
2020 :
2021 330616 : relid = classForm->oid;
2022 :
2023 : /*
2024 : * Check if it is a temp table (presumably, of some other backend's).
2025 : * We cannot safely process other backends' temp tables.
2026 : */
2027 330616 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2028 : {
2029 : /*
2030 : * We just ignore it if the owning backend is still active and
2031 : * using the temporary schema. Also, for safety, ignore it if the
2032 : * namespace doesn't exist or isn't a temp namespace after all.
2033 : */
2034 18 : if (checkTempNamespaceStatus(classForm->relnamespace) == TEMP_NAMESPACE_IDLE)
2035 : {
2036 : /*
2037 : * The table seems to be orphaned -- although it might be that
2038 : * the owning backend has already deleted it and exited; our
2039 : * pg_class scan snapshot is not necessarily up-to-date
2040 : * anymore, so we could be looking at a committed-dead entry.
2041 : * Remember it so we can try to delete it later.
2042 : */
2043 0 : orphan_oids = lappend_oid(orphan_oids, relid);
2044 : }
2045 18 : continue;
2046 : }
2047 :
2048 : /* Fetch reloptions and the pgstat entry for this table */
2049 330598 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2050 330598 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2051 : relid);
2052 :
2053 : /* Check if it needs vacuum or analyze */
2054 330598 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2055 : effective_multixact_freeze_max_age,
2056 : &dovacuum, &doanalyze, &wraparound);
2057 :
2058 : /* Relations that need work are added to table_oids */
2059 330598 : if (dovacuum || doanalyze)
2060 166860 : table_oids = lappend_oid(table_oids, relid);
2061 :
2062 : /*
2063 : * Remember TOAST associations for the second pass. Note: we must do
2064 : * this whether or not the table is going to be vacuumed, because we
2065 : * don't automatically vacuum toast tables along the parent table.
2066 : */
2067 330598 : if (OidIsValid(classForm->reltoastrelid))
2068 : {
2069 : av_relation *hentry;
2070 : bool found;
2071 :
2072 380872 : hentry = hash_search(table_toast_map,
2073 190436 : &classForm->reltoastrelid,
2074 : HASH_ENTER, &found);
2075 :
2076 190436 : if (!found)
2077 : {
2078 : /* hash_search already filled in the key */
2079 190436 : hentry->ar_relid = relid;
2080 190436 : hentry->ar_hasrelopts = false;
2081 190436 : if (relopts != NULL)
2082 : {
2083 4266 : hentry->ar_hasrelopts = true;
2084 4266 : memcpy(&hentry->ar_reloptions, relopts,
2085 : sizeof(AutoVacOpts));
2086 : }
2087 : }
2088 : }
2089 :
2090 : /* Release stuff to avoid per-relation leakage */
2091 330598 : if (relopts)
2092 4340 : pfree(relopts);
2093 330598 : if (tabentry)
2094 325632 : pfree(tabentry);
2095 : }
2096 :
2097 4730 : table_endscan(relScan);
2098 :
2099 : /* second pass: check TOAST tables */
2100 4730 : ScanKeyInit(&key,
2101 : Anum_pg_class_relkind,
2102 : BTEqualStrategyNumber, F_CHAREQ,
2103 : CharGetDatum(RELKIND_TOASTVALUE));
2104 :
2105 4730 : relScan = table_beginscan_catalog(classRel, 1, &key);
2106 195168 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2107 : {
2108 190438 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2109 : PgStat_StatTabEntry *tabentry;
2110 : Oid relid;
2111 : AutoVacOpts *relopts;
2112 190438 : bool free_relopts = false;
2113 : bool dovacuum;
2114 : bool doanalyze;
2115 : bool wraparound;
2116 :
2117 : /*
2118 : * We cannot safely process other backends' temp tables, so skip 'em.
2119 : */
2120 190438 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2121 2 : continue;
2122 :
2123 190436 : relid = classForm->oid;
2124 :
2125 : /*
2126 : * fetch reloptions -- if this toast table does not have them, try the
2127 : * main rel
2128 : */
2129 190436 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2130 190436 : if (relopts)
2131 2 : free_relopts = true;
2132 : else
2133 : {
2134 : av_relation *hentry;
2135 : bool found;
2136 :
2137 190434 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2138 190434 : if (found && hentry->ar_hasrelopts)
2139 4264 : relopts = &hentry->ar_reloptions;
2140 : }
2141 :
2142 : /* Fetch the pgstat entry for this table */
2143 190436 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2144 : relid);
2145 :
2146 190436 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2147 : effective_multixact_freeze_max_age,
2148 : &dovacuum, &doanalyze, &wraparound);
2149 :
2150 : /* ignore analyze for toast tables */
2151 190436 : if (dovacuum)
2152 96484 : table_oids = lappend_oid(table_oids, relid);
2153 :
2154 : /* Release stuff to avoid leakage */
2155 190436 : if (free_relopts)
2156 2 : pfree(relopts);
2157 190436 : if (tabentry)
2158 186856 : pfree(tabentry);
2159 : }
2160 :
2161 4730 : table_endscan(relScan);
2162 4730 : table_close(classRel, AccessShareLock);
2163 :
2164 : /*
2165 : * Recheck orphan temporary tables, and if they still seem orphaned, drop
2166 : * them. We'll eat a transaction per dropped table, which might seem
2167 : * excessive, but we should only need to do anything as a result of a
2168 : * previous backend crash, so this should not happen often enough to
2169 : * justify "optimizing". Using separate transactions ensures that we
2170 : * don't bloat the lock table if there are many temp tables to be dropped,
2171 : * and it ensures that we don't lose work if a deletion attempt fails.
2172 : */
2173 4730 : foreach(cell, orphan_oids)
2174 : {
2175 0 : Oid relid = lfirst_oid(cell);
2176 : Form_pg_class classForm;
2177 : ObjectAddress object;
2178 :
2179 : /*
2180 : * Check for user-requested abort.
2181 : */
2182 0 : CHECK_FOR_INTERRUPTS();
2183 :
2184 : /*
2185 : * Try to lock the table. If we can't get the lock immediately,
2186 : * somebody else is using (or dropping) the table, so it's not our
2187 : * concern anymore. Having the lock prevents race conditions below.
2188 : */
2189 0 : if (!ConditionalLockRelationOid(relid, AccessExclusiveLock))
2190 0 : continue;
2191 :
2192 : /*
2193 : * Re-fetch the pg_class tuple and re-check whether it still seems to
2194 : * be an orphaned temp table. If it's not there or no longer the same
2195 : * relation, ignore it.
2196 : */
2197 0 : tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2198 0 : if (!HeapTupleIsValid(tuple))
2199 : {
2200 : /* be sure to drop useless lock so we don't bloat lock table */
2201 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2202 0 : continue;
2203 : }
2204 0 : classForm = (Form_pg_class) GETSTRUCT(tuple);
2205 :
2206 : /*
2207 : * Make all the same tests made in the loop above. In event of OID
2208 : * counter wraparound, the pg_class entry we have now might be
2209 : * completely unrelated to the one we saw before.
2210 : */
2211 0 : if (!((classForm->relkind == RELKIND_RELATION ||
2212 0 : classForm->relkind == RELKIND_MATVIEW) &&
2213 0 : classForm->relpersistence == RELPERSISTENCE_TEMP))
2214 : {
2215 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2216 0 : continue;
2217 : }
2218 :
2219 0 : if (checkTempNamespaceStatus(classForm->relnamespace) != TEMP_NAMESPACE_IDLE)
2220 : {
2221 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2222 0 : continue;
2223 : }
2224 :
2225 : /*
2226 : * Try to lock the temp namespace, too. Even though we have lock on
2227 : * the table itself, there's a risk of deadlock against an incoming
2228 : * backend trying to clean out the temp namespace, in case this table
2229 : * has dependencies (such as sequences) that the backend's
2230 : * performDeletion call might visit in a different order. If we can
2231 : * get AccessShareLock on the namespace, that's sufficient to ensure
2232 : * we're not running concurrently with RemoveTempRelations. If we
2233 : * can't, back off and let RemoveTempRelations do its thing.
2234 : */
2235 0 : if (!ConditionalLockDatabaseObject(NamespaceRelationId,
2236 : classForm->relnamespace, 0,
2237 : AccessShareLock))
2238 : {
2239 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2240 0 : continue;
2241 : }
2242 :
2243 : /* OK, let's delete it */
2244 0 : ereport(LOG,
2245 : (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"",
2246 : get_database_name(MyDatabaseId),
2247 : get_namespace_name(classForm->relnamespace),
2248 : NameStr(classForm->relname))));
2249 :
2250 : /*
2251 : * Deletion might involve TOAST table access, so ensure we have a
2252 : * valid snapshot.
2253 : */
2254 0 : PushActiveSnapshot(GetTransactionSnapshot());
2255 :
2256 0 : object.classId = RelationRelationId;
2257 0 : object.objectId = relid;
2258 0 : object.objectSubId = 0;
2259 0 : performDeletion(&object, DROP_CASCADE,
2260 : PERFORM_DELETION_INTERNAL |
2261 : PERFORM_DELETION_QUIETLY |
2262 : PERFORM_DELETION_SKIP_EXTENSIONS);
2263 :
2264 : /*
2265 : * To commit the deletion, end current transaction and start a new
2266 : * one. Note this also releases the locks we took.
2267 : */
2268 0 : PopActiveSnapshot();
2269 0 : CommitTransactionCommand();
2270 0 : StartTransactionCommand();
2271 :
2272 : /* StartTransactionCommand changed current memory context */
2273 0 : MemoryContextSwitchTo(AutovacMemCxt);
2274 : }
2275 :
2276 : /*
2277 : * Optionally, create a buffer access strategy object for VACUUM to use.
2278 : * We use the same BufferAccessStrategy object for all tables VACUUMed by
2279 : * this worker to prevent autovacuum from blowing out shared buffers.
2280 : *
2281 : * VacuumBufferUsageLimit being set to 0 results in
2282 : * GetAccessStrategyWithSize returning NULL, effectively meaning we can
2283 : * use up to all of shared buffers.
2284 : *
2285 : * If we later enter failsafe mode on any of the tables being vacuumed, we
2286 : * will cease use of the BufferAccessStrategy only for that table.
2287 : *
2288 : * XXX should we consider adding code to adjust the size of this if
2289 : * VacuumBufferUsageLimit changes?
2290 : */
2291 4730 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, VacuumBufferUsageLimit);
2292 :
2293 : /*
2294 : * create a memory context to act as fake PortalContext, so that the
2295 : * contexts created in the vacuum code are cleaned up for each table.
2296 : */
2297 4730 : PortalContext = AllocSetContextCreate(AutovacMemCxt,
2298 : "Autovacuum Portal",
2299 : ALLOCSET_DEFAULT_SIZES);
2300 :
2301 : /*
2302 : * Perform operations on collected tables.
2303 : */
2304 268074 : foreach(cell, table_oids)
2305 : {
2306 263344 : Oid relid = lfirst_oid(cell);
2307 : HeapTuple classTup;
2308 : autovac_table *tab;
2309 : bool isshared;
2310 : bool skipit;
2311 : dlist_iter iter;
2312 :
2313 263344 : CHECK_FOR_INTERRUPTS();
2314 :
2315 : /*
2316 : * Check for config changes before processing each collected table.
2317 : */
2318 263344 : if (ConfigReloadPending)
2319 : {
2320 0 : ConfigReloadPending = false;
2321 0 : ProcessConfigFile(PGC_SIGHUP);
2322 :
2323 : /*
2324 : * You might be tempted to bail out if we see autovacuum is now
2325 : * disabled. Must resist that temptation -- this might be a
2326 : * for-wraparound emergency worker, in which case that would be
2327 : * entirely inappropriate.
2328 : */
2329 : }
2330 :
2331 : /*
2332 : * Find out whether the table is shared or not. (It's slightly
2333 : * annoying to fetch the syscache entry just for this, but in typical
2334 : * cases it adds little cost because table_recheck_autovac would
2335 : * refetch the entry anyway. We could buy that back by copying the
2336 : * tuple here and passing it to table_recheck_autovac, but that
2337 : * increases the odds of that function working with stale data.)
2338 : */
2339 263344 : classTup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
2340 263344 : if (!HeapTupleIsValid(classTup))
2341 1204 : continue; /* somebody deleted the rel, forget it */
2342 263344 : isshared = ((Form_pg_class) GETSTRUCT(classTup))->relisshared;
2343 263344 : ReleaseSysCache(classTup);
2344 :
2345 : /*
2346 : * Hold schedule lock from here until we've claimed the table. We
2347 : * also need the AutovacuumLock to walk the worker array, but that one
2348 : * can just be a shared lock.
2349 : */
2350 263344 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2351 263344 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2352 :
2353 : /*
2354 : * Check whether the table is being vacuumed concurrently by another
2355 : * worker.
2356 : */
2357 263344 : skipit = false;
2358 822784 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
2359 : {
2360 560644 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
2361 :
2362 : /* ignore myself */
2363 560644 : if (worker == MyWorkerInfo)
2364 262786 : continue;
2365 :
2366 : /* ignore workers in other databases (unless table is shared) */
2367 297858 : if (!worker->wi_sharedrel && worker->wi_dboid != MyDatabaseId)
2368 1344 : continue;
2369 :
2370 296514 : if (worker->wi_tableoid == relid)
2371 : {
2372 1204 : skipit = true;
2373 1204 : found_concurrent_worker = true;
2374 1204 : break;
2375 : }
2376 : }
2377 263344 : LWLockRelease(AutovacuumLock);
2378 263344 : if (skipit)
2379 : {
2380 1204 : LWLockRelease(AutovacuumScheduleLock);
2381 1204 : continue;
2382 : }
2383 :
2384 : /*
2385 : * Store the table's OID in shared memory before releasing the
2386 : * schedule lock, so that other workers don't try to vacuum it
2387 : * concurrently. (We claim it here so as not to hold
2388 : * AutovacuumScheduleLock while rechecking the stats.)
2389 : */
2390 262140 : MyWorkerInfo->wi_tableoid = relid;
2391 262140 : MyWorkerInfo->wi_sharedrel = isshared;
2392 262140 : LWLockRelease(AutovacuumScheduleLock);
2393 :
2394 : /*
2395 : * Check whether pgstat data still says we need to vacuum this table.
2396 : * It could have changed if something else processed the table while
2397 : * we weren't looking. This doesn't entirely close the race condition,
2398 : * but it is very small.
2399 : */
2400 262140 : MemoryContextSwitchTo(AutovacMemCxt);
2401 262140 : tab = table_recheck_autovac(relid, table_toast_map, pg_class_desc,
2402 : effective_multixact_freeze_max_age);
2403 262140 : if (tab == NULL)
2404 : {
2405 : /* someone else vacuumed the table, or it went away */
2406 0 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2407 0 : MyWorkerInfo->wi_tableoid = InvalidOid;
2408 0 : MyWorkerInfo->wi_sharedrel = false;
2409 0 : LWLockRelease(AutovacuumScheduleLock);
2410 0 : continue;
2411 : }
2412 :
2413 : /*
2414 : * Save the cost-related storage parameter values in global variables
2415 : * for reference when updating vacuum_cost_delay and vacuum_cost_limit
2416 : * during vacuuming this table.
2417 : */
2418 262140 : av_storage_param_cost_delay = tab->at_storage_param_vac_cost_delay;
2419 262140 : av_storage_param_cost_limit = tab->at_storage_param_vac_cost_limit;
2420 :
2421 : /*
2422 : * We only expect this worker to ever set the flag, so don't bother
2423 : * checking the return value. We shouldn't have to retry.
2424 : */
2425 262140 : if (tab->at_dobalance)
2426 262140 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2427 : else
2428 0 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
2429 :
2430 262140 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2431 262140 : autovac_recalculate_workers_for_balance();
2432 262140 : LWLockRelease(AutovacuumLock);
2433 :
2434 : /*
2435 : * We wait until this point to update cost delay and cost limit
2436 : * values, even though we reloaded the configuration file above, so
2437 : * that we can take into account the cost-related storage parameters.
2438 : */
2439 262140 : VacuumUpdateCosts();
2440 :
2441 :
2442 : /* clean up memory before each iteration */
2443 262140 : MemoryContextReset(PortalContext);
2444 :
2445 : /*
2446 : * Save the relation name for a possible error message, to avoid a
2447 : * catalog lookup in case of an error. If any of these return NULL,
2448 : * then the relation has been dropped since last we checked; skip it.
2449 : * Note: they must live in a long-lived memory context because we call
2450 : * vacuum and analyze in different transactions.
2451 : */
2452 :
2453 262140 : tab->at_relname = get_rel_name(tab->at_relid);
2454 262140 : tab->at_nspname = get_namespace_name(get_rel_namespace(tab->at_relid));
2455 262140 : tab->at_datname = get_database_name(MyDatabaseId);
2456 262140 : if (!tab->at_relname || !tab->at_nspname || !tab->at_datname)
2457 0 : goto deleted;
2458 :
2459 : /*
2460 : * We will abort vacuuming the current table if something errors out,
2461 : * and continue with the next one in schedule; in particular, this
2462 : * happens if we are interrupted with SIGINT.
2463 : */
2464 262140 : PG_TRY();
2465 : {
2466 : /* Use PortalContext for any per-table allocations */
2467 262140 : MemoryContextSwitchTo(PortalContext);
2468 :
2469 : /* have at it */
2470 262140 : autovacuum_do_vac_analyze(tab, bstrategy);
2471 :
2472 : /*
2473 : * Clear a possible query-cancel signal, to avoid a late reaction
2474 : * to an automatically-sent signal because of vacuuming the
2475 : * current table (we're done with it, so it would make no sense to
2476 : * cancel at this point.)
2477 : */
2478 262140 : QueryCancelPending = false;
2479 : }
2480 0 : PG_CATCH();
2481 : {
2482 : /*
2483 : * Abort the transaction, start a new one, and proceed with the
2484 : * next table in our list.
2485 : */
2486 0 : HOLD_INTERRUPTS();
2487 0 : if (tab->at_params.options & VACOPT_VACUUM)
2488 0 : errcontext("automatic vacuum of table \"%s.%s.%s\"",
2489 : tab->at_datname, tab->at_nspname, tab->at_relname);
2490 : else
2491 0 : errcontext("automatic analyze of table \"%s.%s.%s\"",
2492 : tab->at_datname, tab->at_nspname, tab->at_relname);
2493 0 : EmitErrorReport();
2494 :
2495 : /* this resets ProcGlobal->statusFlags[i] too */
2496 0 : AbortOutOfAnyTransaction();
2497 0 : FlushErrorState();
2498 0 : MemoryContextReset(PortalContext);
2499 :
2500 : /* restart our transaction for the following operations */
2501 0 : StartTransactionCommand();
2502 0 : RESUME_INTERRUPTS();
2503 : }
2504 262140 : PG_END_TRY();
2505 :
2506 : /* Make sure we're back in AutovacMemCxt */
2507 262140 : MemoryContextSwitchTo(AutovacMemCxt);
2508 :
2509 262140 : did_vacuum = true;
2510 :
2511 : /* ProcGlobal->statusFlags[i] are reset at the next end of xact */
2512 :
2513 : /* be tidy */
2514 262140 : deleted:
2515 262140 : if (tab->at_datname != NULL)
2516 262140 : pfree(tab->at_datname);
2517 262140 : if (tab->at_nspname != NULL)
2518 262140 : pfree(tab->at_nspname);
2519 262140 : if (tab->at_relname != NULL)
2520 262140 : pfree(tab->at_relname);
2521 262140 : pfree(tab);
2522 :
2523 : /*
2524 : * Remove my info from shared memory. We set wi_dobalance on the
2525 : * assumption that we are more likely than not to vacuum a table with
2526 : * no cost-related storage parameters next, so we want to claim our
2527 : * share of I/O as soon as possible to avoid thrashing the global
2528 : * balance.
2529 : */
2530 262140 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2531 262140 : MyWorkerInfo->wi_tableoid = InvalidOid;
2532 262140 : MyWorkerInfo->wi_sharedrel = false;
2533 262140 : LWLockRelease(AutovacuumScheduleLock);
2534 262140 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2535 : }
2536 :
2537 4730 : list_free(table_oids);
2538 :
2539 : /*
2540 : * Perform additional work items, as requested by backends.
2541 : */
2542 4730 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2543 1215610 : for (i = 0; i < NUM_WORKITEMS; i++)
2544 : {
2545 1210880 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
2546 :
2547 1210880 : if (!workitem->avw_used)
2548 1210874 : continue;
2549 6 : if (workitem->avw_active)
2550 0 : continue;
2551 6 : if (workitem->avw_database != MyDatabaseId)
2552 0 : continue;
2553 :
2554 : /* claim this one, and release lock while performing it */
2555 6 : workitem->avw_active = true;
2556 6 : LWLockRelease(AutovacuumLock);
2557 :
2558 6 : perform_work_item(workitem);
2559 :
2560 : /*
2561 : * Check for config changes before acquiring lock for further jobs.
2562 : */
2563 6 : CHECK_FOR_INTERRUPTS();
2564 6 : if (ConfigReloadPending)
2565 : {
2566 0 : ConfigReloadPending = false;
2567 0 : ProcessConfigFile(PGC_SIGHUP);
2568 0 : VacuumUpdateCosts();
2569 : }
2570 :
2571 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2572 :
2573 : /* and mark it done */
2574 6 : workitem->avw_active = false;
2575 6 : workitem->avw_used = false;
2576 : }
2577 4730 : LWLockRelease(AutovacuumLock);
2578 :
2579 : /*
2580 : * We leak table_toast_map here (among other things), but since we're
2581 : * going away soon, it's not a problem normally. But when using Valgrind,
2582 : * release some stuff to reduce complaints about leaked storage.
2583 : */
2584 : #ifdef USE_VALGRIND
2585 : hash_destroy(table_toast_map);
2586 : FreeTupleDesc(pg_class_desc);
2587 : if (bstrategy)
2588 : pfree(bstrategy);
2589 : #endif
2590 :
2591 : /* Run the rest in xact context, mainly to avoid Valgrind leak warnings */
2592 4730 : MemoryContextSwitchTo(TopTransactionContext);
2593 :
2594 : /*
2595 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We
2596 : * only need to do this once, not after each table.
2597 : *
2598 : * Even if we didn't vacuum anything, it may still be important to do
2599 : * this, because one indirect effect of vac_update_datfrozenxid() is to
2600 : * update TransamVariables->xidVacLimit. That might need to be done even
2601 : * if we haven't vacuumed anything, because relations with older
2602 : * relfrozenxid values or other databases with older datfrozenxid values
2603 : * might have been dropped, allowing xidVacLimit to advance.
2604 : *
2605 : * However, it's also important not to do this blindly in all cases,
2606 : * because when autovacuum=off this will restart the autovacuum launcher.
2607 : * If we're not careful, an infinite loop can result, where workers find
2608 : * no work to do and restart the launcher, which starts another worker in
2609 : * the same database that finds no work to do. To prevent that, we skip
2610 : * this if (1) we found no work to do and (2) we skipped at least one
2611 : * table due to concurrent autovacuum activity. In that case, the other
2612 : * worker has already done it, or will do so when it finishes.
2613 : */
2614 4730 : if (did_vacuum || !found_concurrent_worker)
2615 4730 : vac_update_datfrozenxid();
2616 :
2617 : /* Finally close out the last transaction. */
2618 4730 : CommitTransactionCommand();
2619 4730 : }
2620 :
2621 : /*
2622 : * Execute a previously registered work item.
2623 : */
2624 : static void
2625 6 : perform_work_item(AutoVacuumWorkItem *workitem)
2626 : {
2627 6 : char *cur_datname = NULL;
2628 6 : char *cur_nspname = NULL;
2629 6 : char *cur_relname = NULL;
2630 :
2631 : /*
2632 : * Note we do not store table info in MyWorkerInfo, since this is not
2633 : * vacuuming proper.
2634 : */
2635 :
2636 : /*
2637 : * Save the relation name for a possible error message, to avoid a catalog
2638 : * lookup in case of an error. If any of these return NULL, then the
2639 : * relation has been dropped since last we checked; skip it.
2640 : */
2641 : Assert(CurrentMemoryContext == AutovacMemCxt);
2642 :
2643 6 : cur_relname = get_rel_name(workitem->avw_relation);
2644 6 : cur_nspname = get_namespace_name(get_rel_namespace(workitem->avw_relation));
2645 6 : cur_datname = get_database_name(MyDatabaseId);
2646 6 : if (!cur_relname || !cur_nspname || !cur_datname)
2647 0 : goto deleted2;
2648 :
2649 6 : autovac_report_workitem(workitem, cur_nspname, cur_relname);
2650 :
2651 : /* clean up memory before each work item */
2652 6 : MemoryContextReset(PortalContext);
2653 :
2654 : /*
2655 : * We will abort the current work item if something errors out, and
2656 : * continue with the next one; in particular, this happens if we are
2657 : * interrupted with SIGINT. Note that this means that the work item list
2658 : * can be lossy.
2659 : */
2660 6 : PG_TRY();
2661 : {
2662 : /* Use PortalContext for any per-work-item allocations */
2663 6 : MemoryContextSwitchTo(PortalContext);
2664 :
2665 : /*
2666 : * Have at it. Functions called here are responsible for any required
2667 : * user switch and sandbox.
2668 : */
2669 6 : switch (workitem->avw_type)
2670 : {
2671 6 : case AVW_BRINSummarizeRange:
2672 6 : DirectFunctionCall2(brin_summarize_range,
2673 : ObjectIdGetDatum(workitem->avw_relation),
2674 : Int64GetDatum((int64) workitem->avw_blockNumber));
2675 6 : break;
2676 0 : default:
2677 0 : elog(WARNING, "unrecognized work item found: type %d",
2678 : workitem->avw_type);
2679 0 : break;
2680 : }
2681 :
2682 : /*
2683 : * Clear a possible query-cancel signal, to avoid a late reaction to
2684 : * an automatically-sent signal because of vacuuming the current table
2685 : * (we're done with it, so it would make no sense to cancel at this
2686 : * point.)
2687 : */
2688 6 : QueryCancelPending = false;
2689 : }
2690 0 : PG_CATCH();
2691 : {
2692 : /*
2693 : * Abort the transaction, start a new one, and proceed with the next
2694 : * table in our list.
2695 : */
2696 0 : HOLD_INTERRUPTS();
2697 0 : errcontext("processing work entry for relation \"%s.%s.%s\"",
2698 : cur_datname, cur_nspname, cur_relname);
2699 0 : EmitErrorReport();
2700 :
2701 : /* this resets ProcGlobal->statusFlags[i] too */
2702 0 : AbortOutOfAnyTransaction();
2703 0 : FlushErrorState();
2704 0 : MemoryContextReset(PortalContext);
2705 :
2706 : /* restart our transaction for the following operations */
2707 0 : StartTransactionCommand();
2708 0 : RESUME_INTERRUPTS();
2709 : }
2710 6 : PG_END_TRY();
2711 :
2712 : /* Make sure we're back in AutovacMemCxt */
2713 6 : MemoryContextSwitchTo(AutovacMemCxt);
2714 :
2715 : /* We intentionally do not set did_vacuum here */
2716 :
2717 : /* be tidy */
2718 6 : deleted2:
2719 6 : if (cur_datname)
2720 6 : pfree(cur_datname);
2721 6 : if (cur_nspname)
2722 6 : pfree(cur_nspname);
2723 6 : if (cur_relname)
2724 6 : pfree(cur_relname);
2725 6 : }
2726 :
2727 : /*
2728 : * extract_autovac_opts
2729 : *
2730 : * Given a relation's pg_class tuple, return a palloc'd copy of the
2731 : * AutoVacOpts portion of reloptions, if set; otherwise, return NULL.
2732 : *
2733 : * Note: callers do not have a relation lock on the table at this point,
2734 : * so the table could have been dropped, and its catalog rows gone, after
2735 : * we acquired the pg_class row. If pg_class had a TOAST table, this would
2736 : * be a risk; fortunately, it doesn't.
2737 : */
2738 : static AutoVacOpts *
2739 783174 : extract_autovac_opts(HeapTuple tup, TupleDesc pg_class_desc)
2740 : {
2741 : bytea *relopts;
2742 : AutoVacOpts *av;
2743 :
2744 : Assert(((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_RELATION ||
2745 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_MATVIEW ||
2746 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_TOASTVALUE);
2747 :
2748 783174 : relopts = extractRelOptions(tup, pg_class_desc, NULL);
2749 783174 : if (relopts == NULL)
2750 775448 : return NULL;
2751 :
2752 7726 : av = palloc(sizeof(AutoVacOpts));
2753 7726 : memcpy(av, &(((StdRdOptions *) relopts)->autovacuum), sizeof(AutoVacOpts));
2754 7726 : pfree(relopts);
2755 :
2756 7726 : return av;
2757 : }
2758 :
2759 :
2760 : /*
2761 : * table_recheck_autovac
2762 : *
2763 : * Recheck whether a table still needs vacuum or analyze. Return value is a
2764 : * valid autovac_table pointer if it does, NULL otherwise.
2765 : *
2766 : * Note that the returned autovac_table does not have the name fields set.
2767 : */
2768 : static autovac_table *
2769 262140 : table_recheck_autovac(Oid relid, HTAB *table_toast_map,
2770 : TupleDesc pg_class_desc,
2771 : int effective_multixact_freeze_max_age)
2772 : {
2773 : Form_pg_class classForm;
2774 : HeapTuple classTup;
2775 : bool dovacuum;
2776 : bool doanalyze;
2777 262140 : autovac_table *tab = NULL;
2778 : bool wraparound;
2779 : AutoVacOpts *avopts;
2780 262140 : bool free_avopts = false;
2781 :
2782 : /* fetch the relation's relcache entry */
2783 262140 : classTup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2784 262140 : if (!HeapTupleIsValid(classTup))
2785 0 : return NULL;
2786 262140 : classForm = (Form_pg_class) GETSTRUCT(classTup);
2787 :
2788 : /*
2789 : * Get the applicable reloptions. If it is a TOAST table, try to get the
2790 : * main table reloptions if the toast table itself doesn't have.
2791 : */
2792 262140 : avopts = extract_autovac_opts(classTup, pg_class_desc);
2793 262140 : if (avopts)
2794 3384 : free_avopts = true;
2795 258756 : else if (classForm->relkind == RELKIND_TOASTVALUE &&
2796 : table_toast_map != NULL)
2797 : {
2798 : av_relation *hentry;
2799 : bool found;
2800 :
2801 96070 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2802 96070 : if (found && hentry->ar_hasrelopts)
2803 3370 : avopts = &hentry->ar_reloptions;
2804 : }
2805 :
2806 262140 : recheck_relation_needs_vacanalyze(relid, avopts, classForm,
2807 : effective_multixact_freeze_max_age,
2808 : &dovacuum, &doanalyze, &wraparound);
2809 :
2810 : /* OK, it needs something done */
2811 262140 : if (doanalyze || dovacuum)
2812 : {
2813 : int freeze_min_age;
2814 : int freeze_table_age;
2815 : int multixact_freeze_min_age;
2816 : int multixact_freeze_table_age;
2817 : int log_min_duration;
2818 :
2819 : /*
2820 : * Calculate the vacuum cost parameters and the freeze ages. If there
2821 : * are options set in pg_class.reloptions, use them; in the case of a
2822 : * toast table, try the main table too. Otherwise use the GUC
2823 : * defaults, autovacuum's own first and plain vacuum second.
2824 : */
2825 :
2826 : /* -1 in autovac setting means use log_autovacuum_min_duration */
2827 6754 : log_min_duration = (avopts && avopts->log_min_duration >= 0)
2828 : ? avopts->log_min_duration
2829 268894 : : Log_autovacuum_min_duration;
2830 :
2831 : /* these do not have autovacuum-specific settings */
2832 6754 : freeze_min_age = (avopts && avopts->freeze_min_age >= 0)
2833 : ? avopts->freeze_min_age
2834 268894 : : default_freeze_min_age;
2835 :
2836 6754 : freeze_table_age = (avopts && avopts->freeze_table_age >= 0)
2837 : ? avopts->freeze_table_age
2838 268894 : : default_freeze_table_age;
2839 :
2840 268894 : multixact_freeze_min_age = (avopts &&
2841 6754 : avopts->multixact_freeze_min_age >= 0)
2842 : ? avopts->multixact_freeze_min_age
2843 268894 : : default_multixact_freeze_min_age;
2844 :
2845 268894 : multixact_freeze_table_age = (avopts &&
2846 6754 : avopts->multixact_freeze_table_age >= 0)
2847 : ? avopts->multixact_freeze_table_age
2848 268894 : : default_multixact_freeze_table_age;
2849 :
2850 262140 : tab = palloc(sizeof(autovac_table));
2851 262140 : tab->at_relid = relid;
2852 262140 : tab->at_sharedrel = classForm->relisshared;
2853 :
2854 : /*
2855 : * Select VACUUM options. Note we don't say VACOPT_PROCESS_TOAST, so
2856 : * that vacuum() skips toast relations. Also note we tell vacuum() to
2857 : * skip vac_update_datfrozenxid(); we'll do that separately.
2858 : */
2859 262140 : tab->at_params.options =
2860 262140 : (dovacuum ? (VACOPT_VACUUM |
2861 : VACOPT_PROCESS_MAIN |
2862 262140 : VACOPT_SKIP_DATABASE_STATS) : 0) |
2863 262140 : (doanalyze ? VACOPT_ANALYZE : 0) |
2864 262140 : (!wraparound ? VACOPT_SKIP_LOCKED : 0);
2865 :
2866 : /*
2867 : * index_cleanup and truncate are unspecified at first in autovacuum.
2868 : * They will be filled in with usable values using their reloptions
2869 : * (or reloption defaults) later.
2870 : */
2871 262140 : tab->at_params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
2872 262140 : tab->at_params.truncate = VACOPTVALUE_UNSPECIFIED;
2873 : /* As of now, we don't support parallel vacuum for autovacuum */
2874 262140 : tab->at_params.nworkers = -1;
2875 262140 : tab->at_params.freeze_min_age = freeze_min_age;
2876 262140 : tab->at_params.freeze_table_age = freeze_table_age;
2877 262140 : tab->at_params.multixact_freeze_min_age = multixact_freeze_min_age;
2878 262140 : tab->at_params.multixact_freeze_table_age = multixact_freeze_table_age;
2879 262140 : tab->at_params.is_wraparound = wraparound;
2880 262140 : tab->at_params.log_min_duration = log_min_duration;
2881 262140 : tab->at_params.toast_parent = InvalidOid;
2882 :
2883 : /*
2884 : * Later, in vacuum_rel(), we check reloptions for any
2885 : * vacuum_max_eager_freeze_failure_rate override.
2886 : */
2887 262140 : tab->at_params.max_eager_freeze_failure_rate = vacuum_max_eager_freeze_failure_rate;
2888 262140 : tab->at_storage_param_vac_cost_limit = avopts ?
2889 262140 : avopts->vacuum_cost_limit : 0;
2890 262140 : tab->at_storage_param_vac_cost_delay = avopts ?
2891 262140 : avopts->vacuum_cost_delay : -1;
2892 262140 : tab->at_relname = NULL;
2893 262140 : tab->at_nspname = NULL;
2894 262140 : tab->at_datname = NULL;
2895 :
2896 : /*
2897 : * If any of the cost delay parameters has been set individually for
2898 : * this table, disable the balancing algorithm.
2899 : */
2900 262140 : tab->at_dobalance =
2901 268894 : !(avopts && (avopts->vacuum_cost_limit > 0 ||
2902 6754 : avopts->vacuum_cost_delay >= 0));
2903 : }
2904 :
2905 262140 : if (free_avopts)
2906 3384 : pfree(avopts);
2907 262140 : heap_freetuple(classTup);
2908 262140 : return tab;
2909 : }
2910 :
2911 : /*
2912 : * recheck_relation_needs_vacanalyze
2913 : *
2914 : * Subroutine for table_recheck_autovac.
2915 : *
2916 : * Fetch the pgstat of a relation and recheck whether a relation
2917 : * needs to be vacuumed or analyzed.
2918 : */
2919 : static void
2920 262140 : recheck_relation_needs_vacanalyze(Oid relid,
2921 : AutoVacOpts *avopts,
2922 : Form_pg_class classForm,
2923 : int effective_multixact_freeze_max_age,
2924 : bool *dovacuum,
2925 : bool *doanalyze,
2926 : bool *wraparound)
2927 : {
2928 : PgStat_StatTabEntry *tabentry;
2929 :
2930 : /* fetch the pgstat table entry */
2931 262140 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2932 : relid);
2933 :
2934 262140 : relation_needs_vacanalyze(relid, avopts, classForm, tabentry,
2935 : effective_multixact_freeze_max_age,
2936 : dovacuum, doanalyze, wraparound);
2937 :
2938 : /* Release tabentry to avoid leakage */
2939 262140 : if (tabentry)
2940 261220 : pfree(tabentry);
2941 :
2942 : /* ignore ANALYZE for toast tables */
2943 262140 : if (classForm->relkind == RELKIND_TOASTVALUE)
2944 96070 : *doanalyze = false;
2945 262140 : }
2946 :
2947 : /*
2948 : * relation_needs_vacanalyze
2949 : *
2950 : * Check whether a relation needs to be vacuumed or analyzed; return each into
2951 : * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is
2952 : * being forced because of Xid or multixact wraparound.
2953 : *
2954 : * relopts is a pointer to the AutoVacOpts options (either for itself in the
2955 : * case of a plain table, or for either itself or its parent table in the case
2956 : * of a TOAST table), NULL if none; tabentry is the pgstats entry, which can be
2957 : * NULL.
2958 : *
2959 : * A table needs to be vacuumed if the number of dead tuples exceeds a
2960 : * threshold. This threshold is calculated as
2961 : *
2962 : * threshold = vac_base_thresh + vac_scale_factor * reltuples
2963 : * if (threshold > vac_max_thresh)
2964 : * threshold = vac_max_thresh;
2965 : *
2966 : * For analyze, the analysis done is that the number of tuples inserted,
2967 : * deleted and updated since the last analyze exceeds a threshold calculated
2968 : * in the same fashion as above. Note that the cumulative stats system stores
2969 : * the number of tuples (both live and dead) that there were as of the last
2970 : * analyze. This is asymmetric to the VACUUM case.
2971 : *
2972 : * We also force vacuum if the table's relfrozenxid is more than freeze_max_age
2973 : * transactions back, and if its relminmxid is more than
2974 : * multixact_freeze_max_age multixacts back.
2975 : *
2976 : * A table whose autovacuum_enabled option is false is
2977 : * automatically skipped (unless we have to vacuum it due to freeze_max_age).
2978 : * Thus autovacuum can be disabled for specific tables. Also, when the cumulative
2979 : * stats system does not have data about a table, it will be skipped.
2980 : *
2981 : * A table whose vac_base_thresh value is < 0 takes the base value from the
2982 : * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor
2983 : * value < 0 is substituted with the value of
2984 : * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze.
2985 : */
2986 : static void
2987 783174 : relation_needs_vacanalyze(Oid relid,
2988 : AutoVacOpts *relopts,
2989 : Form_pg_class classForm,
2990 : PgStat_StatTabEntry *tabentry,
2991 : int effective_multixact_freeze_max_age,
2992 : /* output params below */
2993 : bool *dovacuum,
2994 : bool *doanalyze,
2995 : bool *wraparound)
2996 : {
2997 : bool force_vacuum;
2998 : bool av_enabled;
2999 :
3000 : /* constants from reloptions or GUC variables */
3001 : int vac_base_thresh,
3002 : vac_max_thresh,
3003 : vac_ins_base_thresh,
3004 : anl_base_thresh;
3005 : float4 vac_scale_factor,
3006 : vac_ins_scale_factor,
3007 : anl_scale_factor;
3008 :
3009 : /* thresholds calculated from above constants */
3010 : float4 vacthresh,
3011 : vacinsthresh,
3012 : anlthresh;
3013 :
3014 : /* number of vacuum (resp. analyze) tuples at this time */
3015 : float4 vactuples,
3016 : instuples,
3017 : anltuples;
3018 :
3019 : /* freeze parameters */
3020 : int freeze_max_age;
3021 : int multixact_freeze_max_age;
3022 : TransactionId xidForceLimit;
3023 : TransactionId relfrozenxid;
3024 : MultiXactId multiForceLimit;
3025 :
3026 : Assert(classForm != NULL);
3027 : Assert(OidIsValid(relid));
3028 :
3029 : /*
3030 : * Determine vacuum/analyze equation parameters. We have two possible
3031 : * sources: the passed reloptions (which could be a main table or a toast
3032 : * table), or the autovacuum GUC variables.
3033 : */
3034 :
3035 : /* -1 in autovac setting means use plain vacuum_scale_factor */
3036 15360 : vac_scale_factor = (relopts && relopts->vacuum_scale_factor >= 0)
3037 0 : ? relopts->vacuum_scale_factor
3038 798534 : : autovacuum_vac_scale;
3039 :
3040 15360 : vac_base_thresh = (relopts && relopts->vacuum_threshold >= 0)
3041 : ? relopts->vacuum_threshold
3042 798534 : : autovacuum_vac_thresh;
3043 :
3044 : /* -1 is used to disable max threshold */
3045 15360 : vac_max_thresh = (relopts && relopts->vacuum_max_threshold >= -1)
3046 : ? relopts->vacuum_max_threshold
3047 798534 : : autovacuum_vac_max_thresh;
3048 :
3049 15360 : vac_ins_scale_factor = (relopts && relopts->vacuum_ins_scale_factor >= 0)
3050 0 : ? relopts->vacuum_ins_scale_factor
3051 798534 : : autovacuum_vac_ins_scale;
3052 :
3053 : /* -1 is used to disable insert vacuums */
3054 15360 : vac_ins_base_thresh = (relopts && relopts->vacuum_ins_threshold >= -1)
3055 : ? relopts->vacuum_ins_threshold
3056 798534 : : autovacuum_vac_ins_thresh;
3057 :
3058 15360 : anl_scale_factor = (relopts && relopts->analyze_scale_factor >= 0)
3059 0 : ? relopts->analyze_scale_factor
3060 798534 : : autovacuum_anl_scale;
3061 :
3062 15360 : anl_base_thresh = (relopts && relopts->analyze_threshold >= 0)
3063 : ? relopts->analyze_threshold
3064 798534 : : autovacuum_anl_thresh;
3065 :
3066 15360 : freeze_max_age = (relopts && relopts->freeze_max_age >= 0)
3067 0 : ? Min(relopts->freeze_max_age, autovacuum_freeze_max_age)
3068 798534 : : autovacuum_freeze_max_age;
3069 :
3070 15360 : multixact_freeze_max_age = (relopts && relopts->multixact_freeze_max_age >= 0)
3071 0 : ? Min(relopts->multixact_freeze_max_age, effective_multixact_freeze_max_age)
3072 798534 : : effective_multixact_freeze_max_age;
3073 :
3074 783174 : av_enabled = (relopts ? relopts->enabled : true);
3075 :
3076 : /* Force vacuum if table is at risk of wraparound */
3077 783174 : xidForceLimit = recentXid - freeze_max_age;
3078 783174 : if (xidForceLimit < FirstNormalTransactionId)
3079 0 : xidForceLimit -= FirstNormalTransactionId;
3080 783174 : relfrozenxid = classForm->relfrozenxid;
3081 1566348 : force_vacuum = (TransactionIdIsNormal(relfrozenxid) &&
3082 783174 : TransactionIdPrecedes(relfrozenxid, xidForceLimit));
3083 783174 : if (!force_vacuum)
3084 : {
3085 259122 : MultiXactId relminmxid = classForm->relminmxid;
3086 :
3087 259122 : multiForceLimit = recentMulti - multixact_freeze_max_age;
3088 259122 : if (multiForceLimit < FirstMultiXactId)
3089 0 : multiForceLimit -= FirstMultiXactId;
3090 518244 : force_vacuum = MultiXactIdIsValid(relminmxid) &&
3091 259122 : MultiXactIdPrecedes(relminmxid, multiForceLimit);
3092 : }
3093 783174 : *wraparound = force_vacuum;
3094 :
3095 : /* User disabled it in pg_class.reloptions? (But ignore if at risk) */
3096 783174 : if (!av_enabled && !force_vacuum)
3097 : {
3098 1788 : *doanalyze = false;
3099 1788 : *dovacuum = false;
3100 1788 : return;
3101 : }
3102 :
3103 : /*
3104 : * If we found stats for the table, and autovacuum is currently enabled,
3105 : * make a threshold-based decision whether to vacuum and/or analyze. If
3106 : * autovacuum is currently disabled, we must be here for anti-wraparound
3107 : * vacuuming only, so don't vacuum (or analyze) anything that's not being
3108 : * forced.
3109 : */
3110 781386 : if (PointerIsValid(tabentry) && AutoVacuumingActive())
3111 771920 : {
3112 771920 : float4 pcnt_unfrozen = 1;
3113 771920 : float4 reltuples = classForm->reltuples;
3114 771920 : int32 relpages = classForm->relpages;
3115 771920 : int32 relallfrozen = classForm->relallfrozen;
3116 :
3117 771920 : vactuples = tabentry->dead_tuples;
3118 771920 : instuples = tabentry->ins_since_vacuum;
3119 771920 : anltuples = tabentry->mod_since_analyze;
3120 :
3121 : /* If the table hasn't yet been vacuumed, take reltuples as zero */
3122 771920 : if (reltuples < 0)
3123 3000 : reltuples = 0;
3124 :
3125 : /*
3126 : * If we have data for relallfrozen, calculate the unfrozen percentage
3127 : * of the table to modify insert scale factor. This helps us decide
3128 : * whether or not to vacuum an insert-heavy table based on the number
3129 : * of inserts to the more "active" part of the table.
3130 : */
3131 771920 : if (relpages > 0 && relallfrozen > 0)
3132 : {
3133 : /*
3134 : * It could be the stats were updated manually and relallfrozen >
3135 : * relpages. Clamp relallfrozen to relpages to avoid nonsensical
3136 : * calculations.
3137 : */
3138 299658 : relallfrozen = Min(relallfrozen, relpages);
3139 299658 : pcnt_unfrozen = 1 - ((float4) relallfrozen / relpages);
3140 : }
3141 :
3142 771920 : vacthresh = (float4) vac_base_thresh + vac_scale_factor * reltuples;
3143 771920 : if (vac_max_thresh >= 0 && vacthresh > (float4) vac_max_thresh)
3144 0 : vacthresh = (float4) vac_max_thresh;
3145 :
3146 771920 : vacinsthresh = (float4) vac_ins_base_thresh +
3147 771920 : vac_ins_scale_factor * reltuples * pcnt_unfrozen;
3148 771920 : anlthresh = (float4) anl_base_thresh + anl_scale_factor * reltuples;
3149 :
3150 : /*
3151 : * Note that we don't need to take special consideration for stat
3152 : * reset, because if that happens, the last vacuum and analyze counts
3153 : * will be reset too.
3154 : */
3155 771920 : if (vac_ins_base_thresh >= 0)
3156 771920 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: %.0f (threshold %.0f), anl: %.0f (threshold %.0f)",
3157 : NameStr(classForm->relname),
3158 : vactuples, vacthresh, instuples, vacinsthresh, anltuples, anlthresh);
3159 : else
3160 0 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: (disabled), anl: %.0f (threshold %.0f)",
3161 : NameStr(classForm->relname),
3162 : vactuples, vacthresh, anltuples, anlthresh);
3163 :
3164 : /* Determine if this table needs vacuum or analyze. */
3165 1021550 : *dovacuum = force_vacuum || (vactuples > vacthresh) ||
3166 249630 : (vac_ins_base_thresh >= 0 && instuples > vacinsthresh);
3167 771920 : *doanalyze = (anltuples > anlthresh);
3168 : }
3169 : else
3170 : {
3171 : /*
3172 : * Skip a table not found in stat hash, unless we have to force vacuum
3173 : * for anti-wrap purposes. If it's not acted upon, there's no need to
3174 : * vacuum it.
3175 : */
3176 9466 : *dovacuum = force_vacuum;
3177 9466 : *doanalyze = false;
3178 : }
3179 :
3180 : /* ANALYZE refuses to work with pg_statistic */
3181 781386 : if (relid == StatisticRelationId)
3182 7132 : *doanalyze = false;
3183 : }
3184 :
3185 : /*
3186 : * autovacuum_do_vac_analyze
3187 : * Vacuum and/or analyze the specified table
3188 : *
3189 : * We expect the caller to have switched into a memory context that won't
3190 : * disappear at transaction commit.
3191 : */
3192 : static void
3193 262140 : autovacuum_do_vac_analyze(autovac_table *tab, BufferAccessStrategy bstrategy)
3194 : {
3195 : RangeVar *rangevar;
3196 : VacuumRelation *rel;
3197 : List *rel_list;
3198 : MemoryContext vac_context;
3199 : MemoryContext old_context;
3200 :
3201 : /* Let pgstat know what we're doing */
3202 262140 : autovac_report_activity(tab);
3203 :
3204 : /* Create a context that vacuum() can use as cross-transaction storage */
3205 262140 : vac_context = AllocSetContextCreate(CurrentMemoryContext,
3206 : "Vacuum",
3207 : ALLOCSET_DEFAULT_SIZES);
3208 :
3209 : /* Set up one VacuumRelation target, identified by OID, for vacuum() */
3210 262140 : old_context = MemoryContextSwitchTo(vac_context);
3211 262140 : rangevar = makeRangeVar(tab->at_nspname, tab->at_relname, -1);
3212 262140 : rel = makeVacuumRelation(rangevar, tab->at_relid, NIL);
3213 262140 : rel_list = list_make1(rel);
3214 262140 : MemoryContextSwitchTo(old_context);
3215 :
3216 262140 : vacuum(rel_list, tab->at_params, bstrategy, vac_context, true);
3217 :
3218 262140 : MemoryContextDelete(vac_context);
3219 262140 : }
3220 :
3221 : /*
3222 : * autovac_report_activity
3223 : * Report to pgstat what autovacuum is doing
3224 : *
3225 : * We send a SQL string corresponding to what the user would see if the
3226 : * equivalent command was to be issued manually.
3227 : *
3228 : * Note we assume that we are going to report the next command as soon as we're
3229 : * done with the current one, and exit right after the last one, so we don't
3230 : * bother to report "<IDLE>" or some such.
3231 : */
3232 : static void
3233 262140 : autovac_report_activity(autovac_table *tab)
3234 : {
3235 : #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56)
3236 : char activity[MAX_AUTOVAC_ACTIV_LEN];
3237 : int len;
3238 :
3239 : /* Report the command and possible options */
3240 262140 : if (tab->at_params.options & VACOPT_VACUUM)
3241 261784 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3242 : "autovacuum: VACUUM%s",
3243 261784 : tab->at_params.options & VACOPT_ANALYZE ? " ANALYZE" : "");
3244 : else
3245 356 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3246 : "autovacuum: ANALYZE");
3247 :
3248 : /*
3249 : * Report the qualified name of the relation.
3250 : */
3251 262140 : len = strlen(activity);
3252 :
3253 262140 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3254 : " %s.%s%s", tab->at_nspname, tab->at_relname,
3255 262140 : tab->at_params.is_wraparound ? " (to prevent wraparound)" : "");
3256 :
3257 : /* Set statement_timestamp() to current time for pg_stat_activity */
3258 262140 : SetCurrentStatementStartTimestamp();
3259 :
3260 262140 : pgstat_report_activity(STATE_RUNNING, activity);
3261 262140 : }
3262 :
3263 : /*
3264 : * autovac_report_workitem
3265 : * Report to pgstat that autovacuum is processing a work item
3266 : */
3267 : static void
3268 6 : autovac_report_workitem(AutoVacuumWorkItem *workitem,
3269 : const char *nspname, const char *relname)
3270 : {
3271 : char activity[MAX_AUTOVAC_ACTIV_LEN + 12 + 2];
3272 : char blk[12 + 2];
3273 : int len;
3274 :
3275 6 : switch (workitem->avw_type)
3276 : {
3277 6 : case AVW_BRINSummarizeRange:
3278 6 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3279 : "autovacuum: BRIN summarize");
3280 6 : break;
3281 : }
3282 :
3283 : /*
3284 : * Report the qualified name of the relation, and the block number if any
3285 : */
3286 6 : len = strlen(activity);
3287 :
3288 6 : if (BlockNumberIsValid(workitem->avw_blockNumber))
3289 6 : snprintf(blk, sizeof(blk), " %u", workitem->avw_blockNumber);
3290 : else
3291 0 : blk[0] = '\0';
3292 :
3293 6 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3294 : " %s.%s%s", nspname, relname, blk);
3295 :
3296 : /* Set statement_timestamp() to current time for pg_stat_activity */
3297 6 : SetCurrentStatementStartTimestamp();
3298 :
3299 6 : pgstat_report_activity(STATE_RUNNING, activity);
3300 6 : }
3301 :
3302 : /*
3303 : * AutoVacuumingActive
3304 : * Check GUC vars and report whether the autovacuum process should be
3305 : * running.
3306 : */
3307 : bool
3308 846450 : AutoVacuumingActive(void)
3309 : {
3310 846450 : if (!autovacuum_start_daemon || !pgstat_track_counts)
3311 9804 : return false;
3312 836646 : return true;
3313 : }
3314 :
3315 : /*
3316 : * Request one work item to the next autovacuum run processing our database.
3317 : * Return false if the request can't be recorded.
3318 : */
3319 : bool
3320 6 : AutoVacuumRequestWork(AutoVacuumWorkItemType type, Oid relationId,
3321 : BlockNumber blkno)
3322 : {
3323 : int i;
3324 6 : bool result = false;
3325 :
3326 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
3327 :
3328 : /*
3329 : * Locate an unused work item and fill it with the given data.
3330 : */
3331 12 : for (i = 0; i < NUM_WORKITEMS; i++)
3332 : {
3333 12 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
3334 :
3335 12 : if (workitem->avw_used)
3336 6 : continue;
3337 :
3338 6 : workitem->avw_used = true;
3339 6 : workitem->avw_active = false;
3340 6 : workitem->avw_type = type;
3341 6 : workitem->avw_database = MyDatabaseId;
3342 6 : workitem->avw_relation = relationId;
3343 6 : workitem->avw_blockNumber = blkno;
3344 6 : result = true;
3345 :
3346 : /* done */
3347 6 : break;
3348 : }
3349 :
3350 6 : LWLockRelease(AutovacuumLock);
3351 :
3352 6 : return result;
3353 : }
3354 :
3355 : /*
3356 : * autovac_init
3357 : * This is called at postmaster initialization.
3358 : *
3359 : * All we do here is annoy the user if he got it wrong.
3360 : */
3361 : void
3362 1704 : autovac_init(void)
3363 : {
3364 1704 : if (!autovacuum_start_daemon)
3365 242 : return;
3366 1462 : else if (!pgstat_track_counts)
3367 0 : ereport(WARNING,
3368 : (errmsg("autovacuum not started because of misconfiguration"),
3369 : errhint("Enable the \"track_counts\" option.")));
3370 : else
3371 1462 : check_av_worker_gucs();
3372 : }
3373 :
3374 : /*
3375 : * AutoVacuumShmemSize
3376 : * Compute space needed for autovacuum-related shared memory
3377 : */
3378 : Size
3379 6156 : AutoVacuumShmemSize(void)
3380 : {
3381 : Size size;
3382 :
3383 : /*
3384 : * Need the fixed struct and the array of WorkerInfoData.
3385 : */
3386 6156 : size = sizeof(AutoVacuumShmemStruct);
3387 6156 : size = MAXALIGN(size);
3388 6156 : size = add_size(size, mul_size(autovacuum_worker_slots,
3389 : sizeof(WorkerInfoData)));
3390 6156 : return size;
3391 : }
3392 :
3393 : /*
3394 : * AutoVacuumShmemInit
3395 : * Allocate and initialize autovacuum-related shared memory
3396 : */
3397 : void
3398 2152 : AutoVacuumShmemInit(void)
3399 : {
3400 : bool found;
3401 :
3402 2152 : AutoVacuumShmem = (AutoVacuumShmemStruct *)
3403 2152 : ShmemInitStruct("AutoVacuum Data",
3404 : AutoVacuumShmemSize(),
3405 : &found);
3406 :
3407 2152 : if (!IsUnderPostmaster)
3408 : {
3409 : WorkerInfo worker;
3410 : int i;
3411 :
3412 : Assert(!found);
3413 :
3414 2152 : AutoVacuumShmem->av_launcherpid = 0;
3415 2152 : dclist_init(&AutoVacuumShmem->av_freeWorkers);
3416 2152 : dlist_init(&AutoVacuumShmem->av_runningWorkers);
3417 2152 : AutoVacuumShmem->av_startingWorker = NULL;
3418 2152 : memset(AutoVacuumShmem->av_workItems, 0,
3419 : sizeof(AutoVacuumWorkItem) * NUM_WORKITEMS);
3420 :
3421 2152 : worker = (WorkerInfo) ((char *) AutoVacuumShmem +
3422 : MAXALIGN(sizeof(AutoVacuumShmemStruct)));
3423 :
3424 : /* initialize the WorkerInfo free list */
3425 25612 : for (i = 0; i < autovacuum_worker_slots; i++)
3426 : {
3427 23460 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
3428 23460 : &worker[i].wi_links);
3429 23460 : pg_atomic_init_flag(&worker[i].wi_dobalance);
3430 : }
3431 :
3432 2152 : pg_atomic_init_u32(&AutoVacuumShmem->av_nworkersForBalance, 0);
3433 :
3434 : }
3435 : else
3436 : Assert(found);
3437 2152 : }
3438 :
3439 : /*
3440 : * GUC check_hook for autovacuum_work_mem
3441 : */
3442 : bool
3443 2232 : check_autovacuum_work_mem(int *newval, void **extra, GucSource source)
3444 : {
3445 : /*
3446 : * -1 indicates fallback.
3447 : *
3448 : * If we haven't yet changed the boot_val default of -1, just let it be.
3449 : * Autovacuum will look to maintenance_work_mem instead.
3450 : */
3451 2232 : if (*newval == -1)
3452 2228 : return true;
3453 :
3454 : /*
3455 : * We clamp manually-set values to at least 64kB. Since
3456 : * maintenance_work_mem is always set to at least this value, do the same
3457 : * here.
3458 : */
3459 4 : if (*newval < 64)
3460 4 : *newval = 64;
3461 :
3462 4 : return true;
3463 : }
3464 :
3465 : /*
3466 : * Returns whether there is a free autovacuum worker slot available.
3467 : */
3468 : static bool
3469 37210 : av_worker_available(void)
3470 : {
3471 : int free_slots;
3472 : int reserved_slots;
3473 :
3474 37210 : free_slots = dclist_count(&AutoVacuumShmem->av_freeWorkers);
3475 :
3476 37210 : reserved_slots = autovacuum_worker_slots - autovacuum_max_workers;
3477 37210 : reserved_slots = Max(0, reserved_slots);
3478 :
3479 37210 : return free_slots > reserved_slots;
3480 : }
3481 :
3482 : /*
3483 : * Emits a WARNING if autovacuum_worker_slots < autovacuum_max_workers.
3484 : */
3485 : static void
3486 1462 : check_av_worker_gucs(void)
3487 : {
3488 1462 : if (autovacuum_worker_slots < autovacuum_max_workers)
3489 0 : ereport(WARNING,
3490 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3491 : errmsg("\"autovacuum_max_workers\" (%d) should be less than or equal to \"autovacuum_worker_slots\" (%d)",
3492 : autovacuum_max_workers, autovacuum_worker_slots),
3493 : errdetail("The server will only start up to \"autovacuum_worker_slots\" (%d) autovacuum workers at a given time.",
3494 : autovacuum_worker_slots)));
3495 1462 : }
|