Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * autovacuum.c
4 : *
5 : * PostgreSQL Integrated Autovacuum Daemon
6 : *
7 : * The autovacuum system is structured in two different kinds of processes: the
8 : * autovacuum launcher and the autovacuum worker. The launcher is an
9 : * always-running process, started by the postmaster when the autovacuum GUC
10 : * parameter is set. The launcher schedules autovacuum workers to be started
11 : * when appropriate. The workers are the processes which execute the actual
12 : * vacuuming; they connect to a database as determined in the launcher, and
13 : * once connected they examine the catalogs to select the tables to vacuum.
14 : *
15 : * The autovacuum launcher cannot start the worker processes by itself,
16 : * because doing so would cause robustness issues (namely, failure to shut
17 : * them down on exceptional conditions, and also, since the launcher is
18 : * connected to shared memory and is thus subject to corruption there, it is
19 : * not as robust as the postmaster). So it leaves that task to the postmaster.
20 : *
21 : * There is an autovacuum shared memory area, where the launcher stores
22 : * information about the database it wants vacuumed. When it wants a new
23 : * worker to start, it sets a flag in shared memory and sends a signal to the
24 : * postmaster. Then postmaster knows nothing more than it must start a worker;
25 : * so it forks a new child, which turns into a worker. This new process
26 : * connects to shared memory, and there it can inspect the information that the
27 : * launcher has set up.
28 : *
29 : * If the fork() call fails in the postmaster, it sets a flag in the shared
30 : * memory area, and sends a signal to the launcher. The launcher, upon
31 : * noticing the flag, can try starting the worker again by resending the
32 : * signal. Note that the failure can only be transient (fork failure due to
33 : * high load, memory pressure, too many processes, etc); more permanent
34 : * problems, like failure to connect to a database, are detected later in the
35 : * worker and dealt with just by having the worker exit normally. The launcher
36 : * will launch a new worker again later, per schedule.
37 : *
38 : * When the worker is done vacuuming it sends SIGUSR2 to the launcher. The
39 : * launcher then wakes up and is able to launch another worker, if the schedule
40 : * is so tight that a new worker is needed immediately. At this time the
41 : * launcher can also balance the settings for the various remaining workers'
42 : * cost-based vacuum delay feature.
43 : *
44 : * Note that there can be more than one worker in a database concurrently.
45 : * They will store the table they are currently vacuuming in shared memory, so
46 : * that other workers avoid being blocked waiting for the vacuum lock for that
47 : * table. They will also fetch the last time the table was vacuumed from
48 : * pgstats just before vacuuming each table, to avoid vacuuming a table that
49 : * was just finished being vacuumed by another worker and thus is no longer
50 : * noted in shared memory. However, there is a small window (due to not yet
51 : * holding the relation lock) during which a worker may choose a table that was
52 : * already vacuumed; this is a bug in the current design.
53 : *
54 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
55 : * Portions Copyright (c) 1994, Regents of the University of California
56 : *
57 : *
58 : * IDENTIFICATION
59 : * src/backend/postmaster/autovacuum.c
60 : *
61 : *-------------------------------------------------------------------------
62 : */
63 : #include "postgres.h"
64 :
65 : #include <signal.h>
66 : #include <sys/time.h>
67 : #include <unistd.h>
68 :
69 : #include "access/heapam.h"
70 : #include "access/htup_details.h"
71 : #include "access/multixact.h"
72 : #include "access/reloptions.h"
73 : #include "access/tableam.h"
74 : #include "access/transam.h"
75 : #include "access/xact.h"
76 : #include "catalog/dependency.h"
77 : #include "catalog/namespace.h"
78 : #include "catalog/pg_database.h"
79 : #include "catalog/pg_namespace.h"
80 : #include "commands/dbcommands.h"
81 : #include "commands/vacuum.h"
82 : #include "common/int.h"
83 : #include "lib/ilist.h"
84 : #include "libpq/pqsignal.h"
85 : #include "miscadmin.h"
86 : #include "nodes/makefuncs.h"
87 : #include "pgstat.h"
88 : #include "postmaster/autovacuum.h"
89 : #include "postmaster/interrupt.h"
90 : #include "postmaster/postmaster.h"
91 : #include "storage/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 : double autovacuum_vac_scale;
124 : int autovacuum_vac_ins_thresh;
125 : double autovacuum_vac_ins_scale;
126 : int autovacuum_anl_thresh;
127 : double autovacuum_anl_scale;
128 : int autovacuum_freeze_max_age;
129 : int autovacuum_multixact_freeze_max_age;
130 :
131 : double autovacuum_vac_cost_delay;
132 : int autovacuum_vac_cost_limit;
133 :
134 : int Log_autovacuum_min_duration = 600000;
135 :
136 : /* the minimum allowed time between two awakenings of the launcher */
137 : #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */
138 : #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */
139 :
140 : /*
141 : * Variables to save the cost-related storage parameters for the current
142 : * relation being vacuumed by this autovacuum worker. Using these, we can
143 : * ensure we don't overwrite the values of vacuum_cost_delay and
144 : * vacuum_cost_limit after reloading the configuration file. They are
145 : * initialized to "invalid" values to indicate that no cost-related storage
146 : * parameters were specified and will be set in do_autovacuum() after checking
147 : * the storage parameters in table_recheck_autovac().
148 : */
149 : static double av_storage_param_cost_delay = -1;
150 : static int av_storage_param_cost_limit = -1;
151 :
152 : /* Flags set by signal handlers */
153 : static volatile sig_atomic_t got_SIGUSR2 = false;
154 :
155 : /* Comparison points for determining whether freeze_max_age is exceeded */
156 : static TransactionId recentXid;
157 : static MultiXactId recentMulti;
158 :
159 : /* Default freeze ages to use for autovacuum (varies by database) */
160 : static int default_freeze_min_age;
161 : static int default_freeze_table_age;
162 : static int default_multixact_freeze_min_age;
163 : static int default_multixact_freeze_table_age;
164 :
165 : /* Memory context for long-lived data */
166 : static MemoryContext AutovacMemCxt;
167 :
168 : /* struct to keep track of databases in launcher */
169 : typedef struct avl_dbase
170 : {
171 : Oid adl_datid; /* hash key -- must be first */
172 : TimestampTz adl_next_worker;
173 : int adl_score;
174 : dlist_node adl_node;
175 : } avl_dbase;
176 :
177 : /* struct to keep track of databases in worker */
178 : typedef struct avw_dbase
179 : {
180 : Oid adw_datid;
181 : char *adw_name;
182 : TransactionId adw_frozenxid;
183 : MultiXactId adw_minmulti;
184 : PgStat_StatDBEntry *adw_entry;
185 : } avw_dbase;
186 :
187 : /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */
188 : typedef struct av_relation
189 : {
190 : Oid ar_toastrelid; /* hash key - must be first */
191 : Oid ar_relid;
192 : bool ar_hasrelopts;
193 : AutoVacOpts ar_reloptions; /* copy of AutoVacOpts from the main table's
194 : * reloptions, or NULL if none */
195 : } av_relation;
196 :
197 : /* struct to keep track of tables to vacuum and/or analyze, after rechecking */
198 : typedef struct autovac_table
199 : {
200 : Oid at_relid;
201 : VacuumParams at_params;
202 : double at_storage_param_vac_cost_delay;
203 : int at_storage_param_vac_cost_limit;
204 : bool at_dobalance;
205 : bool at_sharedrel;
206 : char *at_relname;
207 : char *at_nspname;
208 : char *at_datname;
209 : } autovac_table;
210 :
211 : /*-------------
212 : * This struct holds information about a single worker's whereabouts. We keep
213 : * an array of these in shared memory, sized according to
214 : * autovacuum_worker_slots.
215 : *
216 : * wi_links entry into free list or running list
217 : * wi_dboid OID of the database this worker is supposed to work on
218 : * wi_tableoid OID of the table currently being vacuumed, if any
219 : * wi_sharedrel flag indicating whether table is marked relisshared
220 : * wi_proc pointer to PGPROC of the running worker, NULL if not started
221 : * wi_launchtime Time at which this worker was launched
222 : * wi_dobalance Whether this worker should be included in balance calculations
223 : *
224 : * All fields are protected by AutovacuumLock, except for wi_tableoid and
225 : * wi_sharedrel which are protected by AutovacuumScheduleLock (note these
226 : * two fields are read-only for everyone except that worker itself).
227 : *-------------
228 : */
229 : typedef struct WorkerInfoData
230 : {
231 : dlist_node wi_links;
232 : Oid wi_dboid;
233 : Oid wi_tableoid;
234 : PGPROC *wi_proc;
235 : TimestampTz wi_launchtime;
236 : pg_atomic_flag wi_dobalance;
237 : bool wi_sharedrel;
238 : } WorkerInfoData;
239 :
240 : typedef struct WorkerInfoData *WorkerInfo;
241 :
242 : /*
243 : * Possible signals received by the launcher from remote processes. These are
244 : * stored atomically in shared memory so that other processes can set them
245 : * without locking.
246 : */
247 : typedef enum
248 : {
249 : AutoVacForkFailed, /* failed trying to start a worker */
250 : AutoVacRebalance, /* rebalance the cost limits */
251 : } AutoVacuumSignal;
252 :
253 : #define AutoVacNumSignals (AutoVacRebalance + 1)
254 :
255 : /*
256 : * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This
257 : * list is mostly protected by AutovacuumLock, except that if an item is
258 : * marked 'active' other processes must not modify the work-identifying
259 : * members.
260 : */
261 : typedef struct AutoVacuumWorkItem
262 : {
263 : AutoVacuumWorkItemType avw_type;
264 : bool avw_used; /* below data is valid */
265 : bool avw_active; /* being processed */
266 : Oid avw_database;
267 : Oid avw_relation;
268 : BlockNumber avw_blockNumber;
269 : } AutoVacuumWorkItem;
270 :
271 : #define NUM_WORKITEMS 256
272 :
273 : /*-------------
274 : * The main autovacuum shmem struct. On shared memory we store this main
275 : * struct and the array of WorkerInfo structs. This struct keeps:
276 : *
277 : * av_signal set by other processes to indicate various conditions
278 : * av_launcherpid the PID of the autovacuum launcher
279 : * av_freeWorkers the WorkerInfo freelist
280 : * av_runningWorkers the WorkerInfo non-free queue
281 : * av_startingWorker pointer to WorkerInfo currently being started (cleared by
282 : * the worker itself as soon as it's up and running)
283 : * av_workItems work item array
284 : * av_nworkersForBalance the number of autovacuum workers to use when
285 : * calculating the per worker cost limit
286 : *
287 : * This struct is protected by AutovacuumLock, except for av_signal and parts
288 : * of the worker list (see above).
289 : *-------------
290 : */
291 : typedef struct
292 : {
293 : sig_atomic_t av_signal[AutoVacNumSignals];
294 : pid_t av_launcherpid;
295 : dclist_head av_freeWorkers;
296 : dlist_head av_runningWorkers;
297 : WorkerInfo av_startingWorker;
298 : AutoVacuumWorkItem av_workItems[NUM_WORKITEMS];
299 : pg_atomic_uint32 av_nworkersForBalance;
300 : } AutoVacuumShmemStruct;
301 :
302 : static AutoVacuumShmemStruct *AutoVacuumShmem;
303 :
304 : /*
305 : * the database list (of avl_dbase elements) in the launcher, and the context
306 : * that contains it
307 : */
308 : static dlist_head DatabaseList = DLIST_STATIC_INIT(DatabaseList);
309 : static MemoryContext DatabaseListCxt = NULL;
310 :
311 : /* Pointer to my own WorkerInfo, valid on each worker */
312 : static WorkerInfo MyWorkerInfo = NULL;
313 :
314 : /* PID of launcher, valid only in worker while shutting down */
315 : int AutovacuumLauncherPid = 0;
316 :
317 : static Oid do_start_worker(void);
318 : static void HandleAutoVacLauncherInterrupts(void);
319 : static void AutoVacLauncherShutdown(void) pg_attribute_noreturn();
320 : static void launcher_determine_sleep(bool canlaunch, bool recursing,
321 : struct timeval *nap);
322 : static void launch_worker(TimestampTz now);
323 : static List *get_database_list(void);
324 : static void rebuild_database_list(Oid newdb);
325 : static int db_comparator(const void *a, const void *b);
326 : static void autovac_recalculate_workers_for_balance(void);
327 :
328 : static void do_autovacuum(void);
329 : static void FreeWorkerInfo(int code, Datum arg);
330 :
331 : static autovac_table *table_recheck_autovac(Oid relid, HTAB *table_toast_map,
332 : TupleDesc pg_class_desc,
333 : int effective_multixact_freeze_max_age);
334 : static void recheck_relation_needs_vacanalyze(Oid relid, AutoVacOpts *avopts,
335 : Form_pg_class classForm,
336 : int effective_multixact_freeze_max_age,
337 : bool *dovacuum, bool *doanalyze, bool *wraparound);
338 : static void relation_needs_vacanalyze(Oid relid, AutoVacOpts *relopts,
339 : Form_pg_class classForm,
340 : PgStat_StatTabEntry *tabentry,
341 : int effective_multixact_freeze_max_age,
342 : bool *dovacuum, bool *doanalyze, bool *wraparound);
343 :
344 : static void autovacuum_do_vac_analyze(autovac_table *tab,
345 : BufferAccessStrategy bstrategy);
346 : static AutoVacOpts *extract_autovac_opts(HeapTuple tup,
347 : TupleDesc pg_class_desc);
348 : static void perform_work_item(AutoVacuumWorkItem *workitem);
349 : static void autovac_report_activity(autovac_table *tab);
350 : static void autovac_report_workitem(AutoVacuumWorkItem *workitem,
351 : const char *nspname, const char *relname);
352 : static void avl_sigusr2_handler(SIGNAL_ARGS);
353 : static bool av_worker_available(void);
354 : static void check_av_worker_gucs(void);
355 :
356 :
357 :
358 : /********************************************************************
359 : * AUTOVACUUM LAUNCHER CODE
360 : ********************************************************************/
361 :
362 : /*
363 : * Main entry point for the autovacuum launcher process.
364 : */
365 : void
366 700 : AutoVacLauncherMain(char *startup_data, size_t startup_data_len)
367 : {
368 : sigjmp_buf local_sigjmp_buf;
369 :
370 : Assert(startup_data_len == 0);
371 :
372 : /* Release postmaster's working memory context */
373 700 : if (PostmasterContext)
374 : {
375 700 : MemoryContextDelete(PostmasterContext);
376 700 : PostmasterContext = NULL;
377 : }
378 :
379 700 : MyBackendType = B_AUTOVAC_LAUNCHER;
380 700 : init_ps_display(NULL);
381 :
382 700 : ereport(DEBUG1,
383 : (errmsg_internal("autovacuum launcher started")));
384 :
385 700 : if (PostAuthDelay)
386 0 : pg_usleep(PostAuthDelay * 1000000L);
387 :
388 : Assert(GetProcessingMode() == InitProcessing);
389 :
390 : /*
391 : * Set up signal handlers. We operate on databases much like a regular
392 : * backend, so we use the same signal handling. See equivalent code in
393 : * tcop/postgres.c.
394 : */
395 700 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
396 700 : pqsignal(SIGINT, StatementCancelHandler);
397 700 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
398 : /* SIGQUIT handler was already set up by InitPostmasterChild */
399 :
400 700 : InitializeTimeouts(); /* establishes SIGALRM handler */
401 :
402 700 : pqsignal(SIGPIPE, SIG_IGN);
403 700 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
404 700 : pqsignal(SIGUSR2, avl_sigusr2_handler);
405 700 : pqsignal(SIGFPE, FloatExceptionHandler);
406 700 : pqsignal(SIGCHLD, SIG_DFL);
407 :
408 : /*
409 : * Create a per-backend PGPROC struct in shared memory. We must do this
410 : * before we can use LWLocks or access any shared memory.
411 : */
412 700 : InitProcess();
413 :
414 : /* Early initialization */
415 700 : BaseInit();
416 :
417 700 : InitPostgres(NULL, InvalidOid, NULL, InvalidOid, 0, NULL);
418 :
419 700 : SetProcessingMode(NormalProcessing);
420 :
421 : /*
422 : * Create a memory context that we will do all our work in. We do this so
423 : * that we can reset the context during error recovery and thereby avoid
424 : * possible memory leaks.
425 : */
426 700 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
427 : "Autovacuum Launcher",
428 : ALLOCSET_DEFAULT_SIZES);
429 700 : MemoryContextSwitchTo(AutovacMemCxt);
430 :
431 : /*
432 : * If an exception is encountered, processing resumes here.
433 : *
434 : * This code is a stripped down version of PostgresMain error recovery.
435 : *
436 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
437 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
438 : * signals other than SIGQUIT will be blocked until we complete error
439 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
440 : * call redundant, but it is not since InterruptPending might be set
441 : * already.
442 : */
443 700 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
444 : {
445 : /* since not using PG_TRY, must reset error stack by hand */
446 0 : error_context_stack = NULL;
447 :
448 : /* Prevents interrupts while cleaning up */
449 0 : HOLD_INTERRUPTS();
450 :
451 : /* Forget any pending QueryCancel or timeout request */
452 0 : disable_all_timeouts(false);
453 0 : QueryCancelPending = false; /* second to avoid race condition */
454 :
455 : /* Report the error to the server log */
456 0 : EmitErrorReport();
457 :
458 : /* Abort the current transaction in order to recover */
459 0 : AbortCurrentTransaction();
460 :
461 : /*
462 : * Release any other resources, for the case where we were not in a
463 : * transaction.
464 : */
465 0 : LWLockReleaseAll();
466 0 : pgstat_report_wait_end();
467 0 : UnlockBuffers();
468 : /* this is probably dead code, but let's be safe: */
469 0 : if (AuxProcessResourceOwner)
470 0 : ReleaseAuxProcessResources(false);
471 0 : AtEOXact_Buffers(false);
472 0 : AtEOXact_SMgr();
473 0 : AtEOXact_Files(false);
474 0 : AtEOXact_HashTables(false);
475 :
476 : /*
477 : * Now return to normal top-level context and clear ErrorContext for
478 : * next time.
479 : */
480 0 : MemoryContextSwitchTo(AutovacMemCxt);
481 0 : FlushErrorState();
482 :
483 : /* Flush any leaked data in the top-level context */
484 0 : MemoryContextReset(AutovacMemCxt);
485 :
486 : /* don't leave dangling pointers to freed memory */
487 0 : DatabaseListCxt = NULL;
488 0 : dlist_init(&DatabaseList);
489 :
490 : /* Now we can allow interrupts again */
491 0 : RESUME_INTERRUPTS();
492 :
493 : /* if in shutdown mode, no need for anything further; just go away */
494 0 : if (ShutdownRequestPending)
495 0 : AutoVacLauncherShutdown();
496 :
497 : /*
498 : * Sleep at least 1 second after any error. We don't want to be
499 : * filling the error logs as fast as we can.
500 : */
501 0 : pg_usleep(1000000L);
502 : }
503 :
504 : /* We can now handle ereport(ERROR) */
505 700 : PG_exception_stack = &local_sigjmp_buf;
506 :
507 : /* must unblock signals before calling rebuild_database_list */
508 700 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
509 :
510 : /*
511 : * Set always-secure search path. Launcher doesn't connect to a database,
512 : * so this has no effect.
513 : */
514 700 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
515 :
516 : /*
517 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
518 : * in postgresql.conf. We don't really want such a dangerous option being
519 : * applied non-interactively.
520 : */
521 700 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
522 :
523 : /*
524 : * Force settable timeouts off to avoid letting these settings prevent
525 : * regular maintenance from being executed.
526 : */
527 700 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
528 700 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
529 700 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
530 700 : SetConfigOption("idle_in_transaction_session_timeout", "0",
531 : PGC_SUSET, PGC_S_OVERRIDE);
532 :
533 : /*
534 : * Force default_transaction_isolation to READ COMMITTED. We don't want
535 : * to pay the overhead of serializable mode, nor add any risk of causing
536 : * deadlocks or delaying other transactions.
537 : */
538 700 : SetConfigOption("default_transaction_isolation", "read committed",
539 : PGC_SUSET, PGC_S_OVERRIDE);
540 :
541 : /*
542 : * Even when system is configured to use a different fetch consistency,
543 : * for autovac we always want fresh stats.
544 : */
545 700 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
546 :
547 : /*
548 : * In emergency mode, just start a worker (unless shutdown was requested)
549 : * and go away.
550 : */
551 700 : if (!AutoVacuumingActive())
552 : {
553 0 : if (!ShutdownRequestPending)
554 0 : do_start_worker();
555 0 : proc_exit(0); /* done */
556 : }
557 :
558 700 : AutoVacuumShmem->av_launcherpid = MyProcPid;
559 :
560 : /*
561 : * Create the initial database list. The invariant we want this list to
562 : * keep is that it's ordered by decreasing next_time. As soon as an entry
563 : * is updated to a higher time, it will be moved to the front (which is
564 : * correct because the only operation is to add autovacuum_naptime to the
565 : * entry, and time always increases).
566 : */
567 700 : rebuild_database_list(InvalidOid);
568 :
569 : /* loop until shutdown request */
570 10038 : while (!ShutdownRequestPending)
571 : {
572 : struct timeval nap;
573 10036 : TimestampTz current_time = 0;
574 : bool can_launch;
575 :
576 : /*
577 : * This loop is a bit different from the normal use of WaitLatch,
578 : * because we'd like to sleep before the first launch of a child
579 : * process. So it's WaitLatch, then ResetLatch, then check for
580 : * wakening conditions.
581 : */
582 :
583 10036 : launcher_determine_sleep(av_worker_available(), false, &nap);
584 :
585 : /*
586 : * Wait until naptime expires or we get some type of signal (all the
587 : * signal handlers will wake us by calling SetLatch).
588 : */
589 10036 : (void) WaitLatch(MyLatch,
590 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
591 10036 : (nap.tv_sec * 1000L) + (nap.tv_usec / 1000L),
592 : WAIT_EVENT_AUTOVACUUM_MAIN);
593 :
594 10030 : ResetLatch(MyLatch);
595 :
596 10030 : HandleAutoVacLauncherInterrupts();
597 :
598 : /*
599 : * a worker finished, or postmaster signaled failure to start a worker
600 : */
601 9338 : if (got_SIGUSR2)
602 : {
603 4528 : got_SIGUSR2 = false;
604 :
605 : /* rebalance cost limits, if needed */
606 4528 : if (AutoVacuumShmem->av_signal[AutoVacRebalance])
607 : {
608 2262 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
609 2262 : AutoVacuumShmem->av_signal[AutoVacRebalance] = false;
610 2262 : autovac_recalculate_workers_for_balance();
611 2262 : LWLockRelease(AutovacuumLock);
612 : }
613 :
614 4528 : if (AutoVacuumShmem->av_signal[AutoVacForkFailed])
615 : {
616 : /*
617 : * If the postmaster failed to start a new worker, we sleep
618 : * for a little while and resend the signal. The new worker's
619 : * state is still in memory, so this is sufficient. After
620 : * that, we restart the main loop.
621 : *
622 : * XXX should we put a limit to the number of times we retry?
623 : * I don't think it makes much sense, because a future start
624 : * of a worker will continue to fail in the same way.
625 : */
626 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = false;
627 0 : pg_usleep(1000000L); /* 1s */
628 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
629 1968 : continue;
630 : }
631 : }
632 :
633 : /*
634 : * There are some conditions that we need to check before trying to
635 : * start a worker. First, we need to make sure that there is a worker
636 : * slot available. Second, we need to make sure that no other worker
637 : * failed while starting up.
638 : */
639 :
640 9338 : current_time = GetCurrentTimestamp();
641 9338 : LWLockAcquire(AutovacuumLock, LW_SHARED);
642 :
643 9338 : can_launch = av_worker_available();
644 :
645 9338 : if (AutoVacuumShmem->av_startingWorker != NULL)
646 : {
647 : int waittime;
648 16 : WorkerInfo worker = AutoVacuumShmem->av_startingWorker;
649 :
650 : /*
651 : * We can't launch another worker when another one is still
652 : * starting up (or failed while doing so), so just sleep for a bit
653 : * more; that worker will wake us up again as soon as it's ready.
654 : * We will only wait autovacuum_naptime seconds (up to a maximum
655 : * of 60 seconds) for this to happen however. Note that failure
656 : * to connect to a particular database is not a problem here,
657 : * because the worker removes itself from the startingWorker
658 : * pointer before trying to connect. Problems detected by the
659 : * postmaster (like fork() failure) are also reported and handled
660 : * differently. The only problems that may cause this code to
661 : * fire are errors in the earlier sections of AutoVacWorkerMain,
662 : * before the worker removes the WorkerInfo from the
663 : * startingWorker pointer.
664 : */
665 16 : waittime = Min(autovacuum_naptime, 60) * 1000;
666 16 : if (TimestampDifferenceExceeds(worker->wi_launchtime, current_time,
667 : waittime))
668 : {
669 0 : LWLockRelease(AutovacuumLock);
670 0 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
671 :
672 : /*
673 : * No other process can put a worker in starting mode, so if
674 : * startingWorker is still INVALID after exchanging our lock,
675 : * we assume it's the same one we saw above (so we don't
676 : * recheck the launch time).
677 : */
678 0 : if (AutoVacuumShmem->av_startingWorker != NULL)
679 : {
680 0 : worker = AutoVacuumShmem->av_startingWorker;
681 0 : worker->wi_dboid = InvalidOid;
682 0 : worker->wi_tableoid = InvalidOid;
683 0 : worker->wi_sharedrel = false;
684 0 : worker->wi_proc = NULL;
685 0 : worker->wi_launchtime = 0;
686 0 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
687 : &worker->wi_links);
688 0 : AutoVacuumShmem->av_startingWorker = NULL;
689 0 : ereport(WARNING,
690 : errmsg("autovacuum worker took too long to start; canceled"));
691 : }
692 : }
693 : else
694 16 : can_launch = false;
695 : }
696 9338 : LWLockRelease(AutovacuumLock); /* either shared or exclusive */
697 :
698 : /* if we can't do anything, just go back to sleep */
699 9338 : if (!can_launch)
700 1968 : continue;
701 :
702 : /* We're OK to start a new worker */
703 :
704 7370 : if (dlist_is_empty(&DatabaseList))
705 : {
706 : /*
707 : * Special case when the list is empty: start a worker right away.
708 : * This covers the initial case, when no database is in pgstats
709 : * (thus the list is empty). Note that the constraints in
710 : * launcher_determine_sleep keep us from starting workers too
711 : * quickly (at most once every autovacuum_naptime when the list is
712 : * empty).
713 : */
714 66 : launch_worker(current_time);
715 : }
716 : else
717 : {
718 : /*
719 : * because rebuild_database_list constructs a list with most
720 : * distant adl_next_worker first, we obtain our database from the
721 : * tail of the list.
722 : */
723 : avl_dbase *avdb;
724 :
725 7304 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
726 :
727 : /*
728 : * launch a worker if next_worker is right now or it is in the
729 : * past
730 : */
731 7304 : if (TimestampDifferenceExceeds(avdb->adl_next_worker,
732 : current_time, 0))
733 2266 : launch_worker(current_time);
734 : }
735 : }
736 :
737 2 : AutoVacLauncherShutdown();
738 : }
739 :
740 : /*
741 : * Process any new interrupts.
742 : */
743 : static void
744 10030 : HandleAutoVacLauncherInterrupts(void)
745 : {
746 : /* the normal shutdown case */
747 10030 : if (ShutdownRequestPending)
748 692 : AutoVacLauncherShutdown();
749 :
750 9338 : if (ConfigReloadPending)
751 : {
752 60 : int autovacuum_max_workers_prev = autovacuum_max_workers;
753 :
754 60 : ConfigReloadPending = false;
755 60 : ProcessConfigFile(PGC_SIGHUP);
756 :
757 : /* shutdown requested in config file? */
758 60 : if (!AutoVacuumingActive())
759 0 : AutoVacLauncherShutdown();
760 :
761 : /*
762 : * If autovacuum_max_workers changed, emit a WARNING if
763 : * autovacuum_worker_slots < autovacuum_max_workers. If it didn't
764 : * change, skip this to avoid too many repeated log messages.
765 : */
766 60 : if (autovacuum_max_workers_prev != autovacuum_max_workers)
767 0 : check_av_worker_gucs();
768 :
769 : /* rebuild the list in case the naptime changed */
770 60 : rebuild_database_list(InvalidOid);
771 : }
772 :
773 : /* Process barrier events */
774 9338 : if (ProcSignalBarrierPending)
775 54 : ProcessProcSignalBarrier();
776 :
777 : /* Perform logging of memory contexts of this process */
778 9338 : if (LogMemoryContextPending)
779 0 : ProcessLogMemoryContextInterrupt();
780 :
781 : /* Process sinval catchup interrupts that happened while sleeping */
782 9338 : ProcessCatchupInterrupt();
783 9338 : }
784 :
785 : /*
786 : * Perform a normal exit from the autovac launcher.
787 : */
788 : static void
789 694 : AutoVacLauncherShutdown(void)
790 : {
791 694 : ereport(DEBUG1,
792 : (errmsg_internal("autovacuum launcher shutting down")));
793 694 : AutoVacuumShmem->av_launcherpid = 0;
794 :
795 694 : proc_exit(0); /* done */
796 : }
797 :
798 : /*
799 : * Determine the time to sleep, based on the database list.
800 : *
801 : * The "canlaunch" parameter indicates whether we can start a worker right now,
802 : * for example due to the workers being all busy. If this is false, we will
803 : * cause a long sleep, which will be interrupted when a worker exits.
804 : */
805 : static void
806 10192 : launcher_determine_sleep(bool canlaunch, bool recursing, struct timeval *nap)
807 : {
808 : /*
809 : * We sleep until the next scheduled vacuum. We trust that when the
810 : * database list was built, care was taken so that no entries have times
811 : * in the past; if the first entry has too close a next_worker value, or a
812 : * time in the past, we will sleep a small nominal time.
813 : */
814 10192 : if (!canlaunch)
815 : {
816 3880 : nap->tv_sec = autovacuum_naptime;
817 3880 : nap->tv_usec = 0;
818 : }
819 6312 : else if (!dlist_is_empty(&DatabaseList))
820 : {
821 6180 : TimestampTz current_time = GetCurrentTimestamp();
822 : TimestampTz next_wakeup;
823 : avl_dbase *avdb;
824 : long secs;
825 : int usecs;
826 :
827 6180 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
828 :
829 6180 : next_wakeup = avdb->adl_next_worker;
830 6180 : TimestampDifference(current_time, next_wakeup, &secs, &usecs);
831 :
832 6180 : nap->tv_sec = secs;
833 6180 : nap->tv_usec = usecs;
834 : }
835 : else
836 : {
837 : /* list is empty, sleep for whole autovacuum_naptime seconds */
838 132 : nap->tv_sec = autovacuum_naptime;
839 132 : nap->tv_usec = 0;
840 : }
841 :
842 : /*
843 : * If the result is exactly zero, it means a database had an entry with
844 : * time in the past. Rebuild the list so that the databases are evenly
845 : * distributed again, and recalculate the time to sleep. This can happen
846 : * if there are more tables needing vacuum than workers, and they all take
847 : * longer to vacuum than autovacuum_naptime.
848 : *
849 : * We only recurse once. rebuild_database_list should always return times
850 : * in the future, but it seems best not to trust too much on that.
851 : */
852 10192 : if (nap->tv_sec == 0 && nap->tv_usec == 0 && !recursing)
853 : {
854 156 : rebuild_database_list(InvalidOid);
855 156 : launcher_determine_sleep(canlaunch, true, nap);
856 156 : return;
857 : }
858 :
859 : /* The smallest time we'll allow the launcher to sleep. */
860 10036 : if (nap->tv_sec <= 0 && nap->tv_usec <= MIN_AUTOVAC_SLEEPTIME * 1000)
861 : {
862 992 : nap->tv_sec = 0;
863 992 : nap->tv_usec = MIN_AUTOVAC_SLEEPTIME * 1000;
864 : }
865 :
866 : /*
867 : * If the sleep time is too large, clamp it to an arbitrary maximum (plus
868 : * any fractional seconds, for simplicity). This avoids an essentially
869 : * infinite sleep in strange cases like the system clock going backwards a
870 : * few years.
871 : */
872 10036 : if (nap->tv_sec > MAX_AUTOVAC_SLEEPTIME)
873 22 : nap->tv_sec = MAX_AUTOVAC_SLEEPTIME;
874 : }
875 :
876 : /*
877 : * Build an updated DatabaseList. It must only contain databases that appear
878 : * in pgstats, and must be sorted by next_worker from highest to lowest,
879 : * distributed regularly across the next autovacuum_naptime interval.
880 : *
881 : * Receives the Oid of the database that made this list be generated (we call
882 : * this the "new" database, because when the database was already present on
883 : * the list, we expect that this function is not called at all). The
884 : * preexisting list, if any, will be used to preserve the order of the
885 : * databases in the autovacuum_naptime period. The new database is put at the
886 : * end of the interval. The actual values are not saved, which should not be
887 : * much of a problem.
888 : */
889 : static void
890 942 : rebuild_database_list(Oid newdb)
891 : {
892 : List *dblist;
893 : ListCell *cell;
894 : MemoryContext newcxt;
895 : MemoryContext oldcxt;
896 : MemoryContext tmpcxt;
897 : HASHCTL hctl;
898 : int score;
899 : int nelems;
900 : HTAB *dbhash;
901 : dlist_iter iter;
902 :
903 942 : newcxt = AllocSetContextCreate(AutovacMemCxt,
904 : "Autovacuum database list",
905 : ALLOCSET_DEFAULT_SIZES);
906 942 : tmpcxt = AllocSetContextCreate(newcxt,
907 : "Autovacuum database list (tmp)",
908 : ALLOCSET_DEFAULT_SIZES);
909 942 : oldcxt = MemoryContextSwitchTo(tmpcxt);
910 :
911 : /*
912 : * Implementing this is not as simple as it sounds, because we need to put
913 : * the new database at the end of the list; next the databases that were
914 : * already on the list, and finally (at the tail of the list) all the
915 : * other databases that are not on the existing list.
916 : *
917 : * To do this, we build an empty hash table of scored databases. We will
918 : * start with the lowest score (zero) for the new database, then
919 : * increasing scores for the databases in the existing list, in order, and
920 : * lastly increasing scores for all databases gotten via
921 : * get_database_list() that are not already on the hash.
922 : *
923 : * Then we will put all the hash elements into an array, sort the array by
924 : * score, and finally put the array elements into the new doubly linked
925 : * list.
926 : */
927 942 : hctl.keysize = sizeof(Oid);
928 942 : hctl.entrysize = sizeof(avl_dbase);
929 942 : hctl.hcxt = tmpcxt;
930 942 : dbhash = hash_create("autovacuum db hash", 20, &hctl, /* magic number here
931 : * FIXME */
932 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
933 :
934 : /* start by inserting the new database */
935 942 : score = 0;
936 942 : if (OidIsValid(newdb))
937 : {
938 : avl_dbase *db;
939 : PgStat_StatDBEntry *entry;
940 :
941 : /* only consider this database if it has a pgstat entry */
942 26 : entry = pgstat_fetch_stat_dbentry(newdb);
943 26 : if (entry != NULL)
944 : {
945 : /* we assume it isn't found because the hash was just created */
946 20 : db = hash_search(dbhash, &newdb, HASH_ENTER, NULL);
947 :
948 : /* hash_search already filled in the key */
949 20 : db->adl_score = score++;
950 : /* next_worker is filled in later */
951 : }
952 : }
953 :
954 : /* Now insert the databases from the existing list */
955 1532 : dlist_foreach(iter, &DatabaseList)
956 : {
957 590 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
958 : avl_dbase *db;
959 : bool found;
960 : PgStat_StatDBEntry *entry;
961 :
962 : /*
963 : * skip databases with no stat entries -- in particular, this gets rid
964 : * of dropped databases
965 : */
966 590 : entry = pgstat_fetch_stat_dbentry(avdb->adl_datid);
967 590 : if (entry == NULL)
968 0 : continue;
969 :
970 590 : db = hash_search(dbhash, &(avdb->adl_datid), HASH_ENTER, &found);
971 :
972 590 : if (!found)
973 : {
974 : /* hash_search already filled in the key */
975 590 : db->adl_score = score++;
976 : /* next_worker is filled in later */
977 : }
978 : }
979 :
980 : /* finally, insert all qualifying databases not previously inserted */
981 942 : dblist = get_database_list();
982 4060 : foreach(cell, dblist)
983 : {
984 3118 : avw_dbase *avdb = lfirst(cell);
985 : avl_dbase *db;
986 : bool found;
987 : PgStat_StatDBEntry *entry;
988 :
989 : /* only consider databases with a pgstat entry */
990 3118 : entry = pgstat_fetch_stat_dbentry(avdb->adw_datid);
991 3118 : if (entry == NULL)
992 1610 : continue;
993 :
994 1508 : db = hash_search(dbhash, &(avdb->adw_datid), HASH_ENTER, &found);
995 : /* only update the score if the database was not already on the hash */
996 1508 : if (!found)
997 : {
998 : /* hash_search already filled in the key */
999 898 : db->adl_score = score++;
1000 : /* next_worker is filled in later */
1001 : }
1002 : }
1003 942 : nelems = score;
1004 :
1005 : /* from here on, the allocated memory belongs to the new list */
1006 942 : MemoryContextSwitchTo(newcxt);
1007 942 : dlist_init(&DatabaseList);
1008 :
1009 942 : if (nelems > 0)
1010 : {
1011 : TimestampTz current_time;
1012 : int millis_increment;
1013 : avl_dbase *dbary;
1014 : avl_dbase *db;
1015 : HASH_SEQ_STATUS seq;
1016 : int i;
1017 :
1018 : /* put all the hash elements into an array */
1019 876 : dbary = palloc(nelems * sizeof(avl_dbase));
1020 :
1021 876 : i = 0;
1022 876 : hash_seq_init(&seq, dbhash);
1023 2384 : while ((db = hash_seq_search(&seq)) != NULL)
1024 1508 : memcpy(&(dbary[i++]), db, sizeof(avl_dbase));
1025 :
1026 : /* sort the array */
1027 876 : qsort(dbary, nelems, sizeof(avl_dbase), db_comparator);
1028 :
1029 : /*
1030 : * Determine the time interval between databases in the schedule. If
1031 : * we see that the configured naptime would take us to sleep times
1032 : * lower than our min sleep time (which launcher_determine_sleep is
1033 : * coded not to allow), silently use a larger naptime (but don't touch
1034 : * the GUC variable).
1035 : */
1036 876 : millis_increment = 1000.0 * autovacuum_naptime / nelems;
1037 876 : if (millis_increment <= MIN_AUTOVAC_SLEEPTIME)
1038 0 : millis_increment = MIN_AUTOVAC_SLEEPTIME * 1.1;
1039 :
1040 876 : current_time = GetCurrentTimestamp();
1041 :
1042 : /*
1043 : * move the elements from the array into the dlist, setting the
1044 : * next_worker while walking the array
1045 : */
1046 2384 : for (i = 0; i < nelems; i++)
1047 : {
1048 1508 : db = &(dbary[i]);
1049 :
1050 1508 : current_time = TimestampTzPlusMilliseconds(current_time,
1051 : millis_increment);
1052 1508 : db->adl_next_worker = current_time;
1053 :
1054 : /* later elements should go closer to the head of the list */
1055 1508 : dlist_push_head(&DatabaseList, &db->adl_node);
1056 : }
1057 : }
1058 :
1059 : /* all done, clean up memory */
1060 942 : if (DatabaseListCxt != NULL)
1061 242 : MemoryContextDelete(DatabaseListCxt);
1062 942 : MemoryContextDelete(tmpcxt);
1063 942 : DatabaseListCxt = newcxt;
1064 942 : MemoryContextSwitchTo(oldcxt);
1065 942 : }
1066 :
1067 : /* qsort comparator for avl_dbase, using adl_score */
1068 : static int
1069 912 : db_comparator(const void *a, const void *b)
1070 : {
1071 912 : return pg_cmp_s32(((const avl_dbase *) a)->adl_score,
1072 : ((const avl_dbase *) b)->adl_score);
1073 : }
1074 :
1075 : /*
1076 : * do_start_worker
1077 : *
1078 : * Bare-bones procedure for starting an autovacuum worker from the launcher.
1079 : * It determines what database to work on, sets up shared memory stuff and
1080 : * signals postmaster to start the worker. It fails gracefully if invoked when
1081 : * autovacuum_workers are already active.
1082 : *
1083 : * Return value is the OID of the database that the worker is going to process,
1084 : * or InvalidOid if no worker was actually started.
1085 : */
1086 : static Oid
1087 2332 : do_start_worker(void)
1088 : {
1089 : List *dblist;
1090 : ListCell *cell;
1091 : TransactionId xidForceLimit;
1092 : MultiXactId multiForceLimit;
1093 : bool for_xid_wrap;
1094 : bool for_multi_wrap;
1095 : avw_dbase *avdb;
1096 : TimestampTz current_time;
1097 2332 : bool skipit = false;
1098 2332 : Oid retval = InvalidOid;
1099 : MemoryContext tmpcxt,
1100 : oldcxt;
1101 :
1102 : /* return quickly when there are no free workers */
1103 2332 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1104 2332 : if (!av_worker_available())
1105 : {
1106 0 : LWLockRelease(AutovacuumLock);
1107 0 : return InvalidOid;
1108 : }
1109 2332 : LWLockRelease(AutovacuumLock);
1110 :
1111 : /*
1112 : * Create and switch to a temporary context to avoid leaking the memory
1113 : * allocated for the database list.
1114 : */
1115 2332 : tmpcxt = AllocSetContextCreate(CurrentMemoryContext,
1116 : "Autovacuum start worker (tmp)",
1117 : ALLOCSET_DEFAULT_SIZES);
1118 2332 : oldcxt = MemoryContextSwitchTo(tmpcxt);
1119 :
1120 : /* Get a list of databases */
1121 2332 : dblist = get_database_list();
1122 :
1123 : /*
1124 : * Determine the oldest datfrozenxid/relfrozenxid that we will allow to
1125 : * pass without forcing a vacuum. (This limit can be tightened for
1126 : * particular tables, but not loosened.)
1127 : */
1128 2332 : recentXid = ReadNextTransactionId();
1129 2332 : xidForceLimit = recentXid - autovacuum_freeze_max_age;
1130 : /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */
1131 : /* this can cause the limit to go backwards by 3, but that's OK */
1132 2332 : if (xidForceLimit < FirstNormalTransactionId)
1133 0 : xidForceLimit -= FirstNormalTransactionId;
1134 :
1135 : /* Also determine the oldest datminmxid we will consider. */
1136 2332 : recentMulti = ReadNextMultiXactId();
1137 2332 : multiForceLimit = recentMulti - MultiXactMemberFreezeThreshold();
1138 2332 : if (multiForceLimit < FirstMultiXactId)
1139 0 : multiForceLimit -= FirstMultiXactId;
1140 :
1141 : /*
1142 : * Choose a database to connect to. We pick the database that was least
1143 : * recently auto-vacuumed, or one that needs vacuuming to prevent Xid
1144 : * wraparound-related data loss. If any db at risk of Xid wraparound is
1145 : * found, we pick the one with oldest datfrozenxid, independently of
1146 : * autovacuum times; similarly we pick the one with the oldest datminmxid
1147 : * if any is in MultiXactId wraparound. Note that those in Xid wraparound
1148 : * danger are given more priority than those in multi wraparound danger.
1149 : *
1150 : * Note that a database with no stats entry is not considered, except for
1151 : * Xid wraparound purposes. The theory is that if no one has ever
1152 : * connected to it since the stats were last initialized, it doesn't need
1153 : * vacuuming.
1154 : *
1155 : * XXX This could be improved if we had more info about whether it needs
1156 : * vacuuming before connecting to it. Perhaps look through the pgstats
1157 : * data for the database's tables? One idea is to keep track of the
1158 : * number of new and dead tuples per database in pgstats. However it
1159 : * isn't clear how to construct a metric that measures that and not cause
1160 : * starvation for less busy databases.
1161 : */
1162 2332 : avdb = NULL;
1163 2332 : for_xid_wrap = false;
1164 2332 : for_multi_wrap = false;
1165 2332 : current_time = GetCurrentTimestamp();
1166 9386 : foreach(cell, dblist)
1167 : {
1168 7054 : avw_dbase *tmp = lfirst(cell);
1169 : dlist_iter iter;
1170 :
1171 : /* Check to see if this one is at risk of wraparound */
1172 7054 : if (TransactionIdPrecedes(tmp->adw_frozenxid, xidForceLimit))
1173 : {
1174 3278 : if (avdb == NULL ||
1175 1256 : TransactionIdPrecedes(tmp->adw_frozenxid,
1176 : avdb->adw_frozenxid))
1177 884 : avdb = tmp;
1178 2022 : for_xid_wrap = true;
1179 5300 : continue;
1180 : }
1181 5032 : else if (for_xid_wrap)
1182 214 : continue; /* ignore not-at-risk DBs */
1183 4818 : else if (MultiXactIdPrecedes(tmp->adw_minmulti, multiForceLimit))
1184 : {
1185 0 : if (avdb == NULL ||
1186 0 : MultiXactIdPrecedes(tmp->adw_minmulti, avdb->adw_minmulti))
1187 0 : avdb = tmp;
1188 0 : for_multi_wrap = true;
1189 0 : continue;
1190 : }
1191 4818 : else if (for_multi_wrap)
1192 0 : continue; /* ignore not-at-risk DBs */
1193 :
1194 : /* Find pgstat entry if any */
1195 4818 : tmp->adw_entry = pgstat_fetch_stat_dbentry(tmp->adw_datid);
1196 :
1197 : /*
1198 : * Skip a database with no pgstat entry; it means it hasn't seen any
1199 : * activity.
1200 : */
1201 4818 : if (!tmp->adw_entry)
1202 284 : continue;
1203 :
1204 : /*
1205 : * Also, skip a database that appears on the database list as having
1206 : * been processed recently (less than autovacuum_naptime seconds ago).
1207 : * We do this so that we don't select a database which we just
1208 : * selected, but that pgstat hasn't gotten around to updating the last
1209 : * autovacuum time yet.
1210 : */
1211 4534 : skipit = false;
1212 :
1213 9112 : dlist_reverse_foreach(iter, &DatabaseList)
1214 : {
1215 9088 : avl_dbase *dbp = dlist_container(avl_dbase, adl_node, iter.cur);
1216 :
1217 9088 : if (dbp->adl_datid == tmp->adw_datid)
1218 : {
1219 : /*
1220 : * Skip this database if its next_worker value falls between
1221 : * the current time and the current time plus naptime.
1222 : */
1223 4510 : if (!TimestampDifferenceExceeds(dbp->adl_next_worker,
1224 2780 : current_time, 0) &&
1225 2780 : !TimestampDifferenceExceeds(current_time,
1226 : dbp->adl_next_worker,
1227 : autovacuum_naptime * 1000))
1228 2780 : skipit = true;
1229 :
1230 4510 : break;
1231 : }
1232 : }
1233 4534 : if (skipit)
1234 2780 : continue;
1235 :
1236 : /*
1237 : * Remember the db with oldest autovac time. (If we are here, both
1238 : * tmp->entry and db->entry must be non-null.)
1239 : */
1240 1754 : if (avdb == NULL ||
1241 254 : tmp->adw_entry->last_autovac_time < avdb->adw_entry->last_autovac_time)
1242 1596 : avdb = tmp;
1243 : }
1244 :
1245 : /* Found a database -- process it */
1246 2332 : if (avdb != NULL)
1247 : {
1248 : WorkerInfo worker;
1249 : dlist_node *wptr;
1250 :
1251 2266 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1252 :
1253 : /*
1254 : * Get a worker entry from the freelist. We checked above, so there
1255 : * really should be a free slot.
1256 : */
1257 2266 : wptr = dclist_pop_head_node(&AutoVacuumShmem->av_freeWorkers);
1258 :
1259 2266 : worker = dlist_container(WorkerInfoData, wi_links, wptr);
1260 2266 : worker->wi_dboid = avdb->adw_datid;
1261 2266 : worker->wi_proc = NULL;
1262 2266 : worker->wi_launchtime = GetCurrentTimestamp();
1263 :
1264 2266 : AutoVacuumShmem->av_startingWorker = worker;
1265 :
1266 2266 : LWLockRelease(AutovacuumLock);
1267 :
1268 2266 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
1269 :
1270 2266 : retval = avdb->adw_datid;
1271 : }
1272 66 : else if (skipit)
1273 : {
1274 : /*
1275 : * If we skipped all databases on the list, rebuild it, because it
1276 : * probably contains a dropped database.
1277 : */
1278 0 : rebuild_database_list(InvalidOid);
1279 : }
1280 :
1281 2332 : MemoryContextSwitchTo(oldcxt);
1282 2332 : MemoryContextDelete(tmpcxt);
1283 :
1284 2332 : return retval;
1285 : }
1286 :
1287 : /*
1288 : * launch_worker
1289 : *
1290 : * Wrapper for starting a worker from the launcher. Besides actually starting
1291 : * it, update the database list to reflect the next time that another one will
1292 : * need to be started on the selected database. The actual database choice is
1293 : * left to do_start_worker.
1294 : *
1295 : * This routine is also expected to insert an entry into the database list if
1296 : * the selected database was previously absent from the list.
1297 : */
1298 : static void
1299 2332 : launch_worker(TimestampTz now)
1300 : {
1301 : Oid dbid;
1302 : dlist_iter iter;
1303 :
1304 2332 : dbid = do_start_worker();
1305 2332 : if (OidIsValid(dbid))
1306 : {
1307 2266 : bool found = false;
1308 :
1309 : /*
1310 : * Walk the database list and update the corresponding entry. If the
1311 : * database is not on the list, we'll recreate the list.
1312 : */
1313 5854 : dlist_foreach(iter, &DatabaseList)
1314 : {
1315 5828 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1316 :
1317 5828 : if (avdb->adl_datid == dbid)
1318 : {
1319 2240 : found = true;
1320 :
1321 : /*
1322 : * add autovacuum_naptime seconds to the current time, and use
1323 : * that as the new "next_worker" field for this database.
1324 : */
1325 2240 : avdb->adl_next_worker =
1326 2240 : TimestampTzPlusMilliseconds(now, autovacuum_naptime * 1000);
1327 :
1328 2240 : dlist_move_head(&DatabaseList, iter.cur);
1329 2240 : break;
1330 : }
1331 : }
1332 :
1333 : /*
1334 : * If the database was not present in the database list, we rebuild
1335 : * the list. It's possible that the database does not get into the
1336 : * list anyway, for example if it's a database that doesn't have a
1337 : * pgstat entry, but this is not a problem because we don't want to
1338 : * schedule workers regularly into those in any case.
1339 : */
1340 2266 : if (!found)
1341 26 : rebuild_database_list(dbid);
1342 : }
1343 2332 : }
1344 :
1345 : /*
1346 : * Called from postmaster to signal a failure to fork a process to become
1347 : * worker. The postmaster should kill(SIGUSR2) the launcher shortly
1348 : * after calling this function.
1349 : */
1350 : void
1351 0 : AutoVacWorkerFailed(void)
1352 : {
1353 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = true;
1354 0 : }
1355 :
1356 : /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */
1357 : static void
1358 4528 : avl_sigusr2_handler(SIGNAL_ARGS)
1359 : {
1360 4528 : got_SIGUSR2 = true;
1361 4528 : SetLatch(MyLatch);
1362 4528 : }
1363 :
1364 :
1365 : /********************************************************************
1366 : * AUTOVACUUM WORKER CODE
1367 : ********************************************************************/
1368 :
1369 : /*
1370 : * Main entry point for autovacuum worker processes.
1371 : */
1372 : void
1373 2268 : AutoVacWorkerMain(char *startup_data, size_t startup_data_len)
1374 : {
1375 : sigjmp_buf local_sigjmp_buf;
1376 : Oid dbid;
1377 :
1378 : Assert(startup_data_len == 0);
1379 :
1380 : /* Release postmaster's working memory context */
1381 2268 : if (PostmasterContext)
1382 : {
1383 2268 : MemoryContextDelete(PostmasterContext);
1384 2268 : PostmasterContext = NULL;
1385 : }
1386 :
1387 2268 : MyBackendType = B_AUTOVAC_WORKER;
1388 2268 : init_ps_display(NULL);
1389 :
1390 : Assert(GetProcessingMode() == InitProcessing);
1391 :
1392 : /*
1393 : * Set up signal handlers. We operate on databases much like a regular
1394 : * backend, so we use the same signal handling. See equivalent code in
1395 : * tcop/postgres.c.
1396 : */
1397 2268 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1398 :
1399 : /*
1400 : * SIGINT is used to signal canceling the current table's vacuum; SIGTERM
1401 : * means abort and exit cleanly, and SIGQUIT means abandon ship.
1402 : */
1403 2268 : pqsignal(SIGINT, StatementCancelHandler);
1404 2268 : pqsignal(SIGTERM, die);
1405 : /* SIGQUIT handler was already set up by InitPostmasterChild */
1406 :
1407 2268 : InitializeTimeouts(); /* establishes SIGALRM handler */
1408 :
1409 2268 : pqsignal(SIGPIPE, SIG_IGN);
1410 2268 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1411 2268 : pqsignal(SIGUSR2, SIG_IGN);
1412 2268 : pqsignal(SIGFPE, FloatExceptionHandler);
1413 2268 : pqsignal(SIGCHLD, SIG_DFL);
1414 :
1415 : /*
1416 : * Create a per-backend PGPROC struct in shared memory. We must do this
1417 : * before we can use LWLocks or access any shared memory.
1418 : */
1419 2268 : InitProcess();
1420 :
1421 : /* Early initialization */
1422 2268 : BaseInit();
1423 :
1424 : /*
1425 : * If an exception is encountered, processing resumes here.
1426 : *
1427 : * Unlike most auxiliary processes, we don't attempt to continue
1428 : * processing after an error; we just clean up and exit. The autovac
1429 : * launcher is responsible for spawning another worker later.
1430 : *
1431 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
1432 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
1433 : * signals other than SIGQUIT will be blocked until we exit. It might
1434 : * seem that this policy makes the HOLD_INTERRUPTS() call redundant, but
1435 : * it is not since InterruptPending might be set already.
1436 : */
1437 2268 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1438 : {
1439 : /* since not using PG_TRY, must reset error stack by hand */
1440 0 : error_context_stack = NULL;
1441 :
1442 : /* Prevents interrupts while cleaning up */
1443 0 : HOLD_INTERRUPTS();
1444 :
1445 : /* Report the error to the server log */
1446 0 : EmitErrorReport();
1447 :
1448 : /*
1449 : * We can now go away. Note that because we called InitProcess, a
1450 : * callback was registered to do ProcKill, which will clean up
1451 : * necessary state.
1452 : */
1453 0 : proc_exit(0);
1454 : }
1455 :
1456 : /* We can now handle ereport(ERROR) */
1457 2268 : PG_exception_stack = &local_sigjmp_buf;
1458 :
1459 2268 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
1460 :
1461 : /*
1462 : * Set always-secure search path, so malicious users can't redirect user
1463 : * code (e.g. pg_index.indexprs). (That code runs in a
1464 : * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not
1465 : * take control of the entire autovacuum worker in any case.)
1466 : */
1467 2268 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1468 :
1469 : /*
1470 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
1471 : * in postgresql.conf. We don't really want such a dangerous option being
1472 : * applied non-interactively.
1473 : */
1474 2268 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
1475 :
1476 : /*
1477 : * Force settable timeouts off to avoid letting these settings prevent
1478 : * regular maintenance from being executed.
1479 : */
1480 2268 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1481 2268 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1482 2268 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1483 2268 : SetConfigOption("idle_in_transaction_session_timeout", "0",
1484 : PGC_SUSET, PGC_S_OVERRIDE);
1485 :
1486 : /*
1487 : * Force default_transaction_isolation to READ COMMITTED. We don't want
1488 : * to pay the overhead of serializable mode, nor add any risk of causing
1489 : * deadlocks or delaying other transactions.
1490 : */
1491 2268 : SetConfigOption("default_transaction_isolation", "read committed",
1492 : PGC_SUSET, PGC_S_OVERRIDE);
1493 :
1494 : /*
1495 : * Force synchronous replication off to allow regular maintenance even if
1496 : * we are waiting for standbys to connect. This is important to ensure we
1497 : * aren't blocked from performing anti-wraparound tasks.
1498 : */
1499 2268 : if (synchronous_commit > SYNCHRONOUS_COMMIT_LOCAL_FLUSH)
1500 2268 : SetConfigOption("synchronous_commit", "local",
1501 : PGC_SUSET, PGC_S_OVERRIDE);
1502 :
1503 : /*
1504 : * Even when system is configured to use a different fetch consistency,
1505 : * for autovac we always want fresh stats.
1506 : */
1507 2268 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
1508 :
1509 : /*
1510 : * Get the info about the database we're going to work on.
1511 : */
1512 2268 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1513 :
1514 : /*
1515 : * beware of startingWorker being INVALID; this should normally not
1516 : * happen, but if a worker fails after forking and before this, the
1517 : * launcher might have decided to remove it from the queue and start
1518 : * again.
1519 : */
1520 2268 : if (AutoVacuumShmem->av_startingWorker != NULL)
1521 : {
1522 2268 : MyWorkerInfo = AutoVacuumShmem->av_startingWorker;
1523 2268 : dbid = MyWorkerInfo->wi_dboid;
1524 2268 : MyWorkerInfo->wi_proc = MyProc;
1525 :
1526 : /* insert into the running list */
1527 2268 : dlist_push_head(&AutoVacuumShmem->av_runningWorkers,
1528 2268 : &MyWorkerInfo->wi_links);
1529 :
1530 : /*
1531 : * remove from the "starting" pointer, so that the launcher can start
1532 : * a new worker if required
1533 : */
1534 2268 : AutoVacuumShmem->av_startingWorker = NULL;
1535 2268 : LWLockRelease(AutovacuumLock);
1536 :
1537 2268 : on_shmem_exit(FreeWorkerInfo, 0);
1538 :
1539 : /* wake up the launcher */
1540 2268 : if (AutoVacuumShmem->av_launcherpid != 0)
1541 2268 : kill(AutoVacuumShmem->av_launcherpid, SIGUSR2);
1542 : }
1543 : else
1544 : {
1545 : /* no worker entry for me, go away */
1546 0 : elog(WARNING, "autovacuum worker started without a worker entry");
1547 0 : dbid = InvalidOid;
1548 0 : LWLockRelease(AutovacuumLock);
1549 : }
1550 :
1551 2268 : if (OidIsValid(dbid))
1552 : {
1553 : char dbname[NAMEDATALEN];
1554 :
1555 : /*
1556 : * Report autovac startup to the cumulative stats system. We
1557 : * deliberately do this before InitPostgres, so that the
1558 : * last_autovac_time will get updated even if the connection attempt
1559 : * fails. This is to prevent autovac from getting "stuck" repeatedly
1560 : * selecting an unopenable database, rather than making any progress
1561 : * on stuff it can connect to.
1562 : */
1563 2268 : pgstat_report_autovac(dbid);
1564 :
1565 : /*
1566 : * Connect to the selected database, specifying no particular user,
1567 : * and ignoring datallowconn. Collect the database's name for
1568 : * display.
1569 : *
1570 : * Note: if we have selected a just-deleted database (due to using
1571 : * stale stats info), we'll fail and exit here.
1572 : */
1573 2268 : InitPostgres(NULL, dbid, NULL, InvalidOid,
1574 : INIT_PG_OVERRIDE_ALLOW_CONNS,
1575 : dbname);
1576 2268 : SetProcessingMode(NormalProcessing);
1577 2268 : set_ps_display(dbname);
1578 2268 : ereport(DEBUG1,
1579 : (errmsg_internal("autovacuum: processing database \"%s\"", dbname)));
1580 :
1581 2268 : if (PostAuthDelay)
1582 0 : pg_usleep(PostAuthDelay * 1000000L);
1583 :
1584 : /* And do an appropriate amount of work */
1585 2268 : recentXid = ReadNextTransactionId();
1586 2268 : recentMulti = ReadNextMultiXactId();
1587 2268 : do_autovacuum();
1588 : }
1589 :
1590 : /*
1591 : * The launcher will be notified of my death in ProcKill, *if* we managed
1592 : * to get a worker slot at all
1593 : */
1594 :
1595 : /* All done, go away */
1596 2264 : proc_exit(0);
1597 : }
1598 :
1599 : /*
1600 : * Return a WorkerInfo to the free list
1601 : */
1602 : static void
1603 2268 : FreeWorkerInfo(int code, Datum arg)
1604 : {
1605 2268 : if (MyWorkerInfo != NULL)
1606 : {
1607 2268 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1608 :
1609 : /*
1610 : * Wake the launcher up so that he can launch a new worker immediately
1611 : * if required. We only save the launcher's PID in local memory here;
1612 : * the actual signal will be sent when the PGPROC is recycled. Note
1613 : * that we always do this, so that the launcher can rebalance the cost
1614 : * limit setting of the remaining workers.
1615 : *
1616 : * We somewhat ignore the risk that the launcher changes its PID
1617 : * between us reading it and the actual kill; we expect ProcKill to be
1618 : * called shortly after us, and we assume that PIDs are not reused too
1619 : * quickly after a process exits.
1620 : */
1621 2268 : AutovacuumLauncherPid = AutoVacuumShmem->av_launcherpid;
1622 :
1623 2268 : dlist_delete(&MyWorkerInfo->wi_links);
1624 2268 : MyWorkerInfo->wi_dboid = InvalidOid;
1625 2268 : MyWorkerInfo->wi_tableoid = InvalidOid;
1626 2268 : MyWorkerInfo->wi_sharedrel = false;
1627 2268 : MyWorkerInfo->wi_proc = NULL;
1628 2268 : MyWorkerInfo->wi_launchtime = 0;
1629 2268 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
1630 2268 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
1631 2268 : &MyWorkerInfo->wi_links);
1632 : /* not mine anymore */
1633 2268 : MyWorkerInfo = NULL;
1634 :
1635 : /*
1636 : * now that we're inactive, cause a rebalancing of the surviving
1637 : * workers
1638 : */
1639 2268 : AutoVacuumShmem->av_signal[AutoVacRebalance] = true;
1640 2268 : LWLockRelease(AutovacuumLock);
1641 : }
1642 2268 : }
1643 :
1644 : /*
1645 : * Update vacuum cost-based delay-related parameters for autovacuum workers and
1646 : * backends executing VACUUM or ANALYZE using the value of relevant GUCs and
1647 : * global state. This must be called during setup for vacuum and after every
1648 : * config reload to ensure up-to-date values.
1649 : */
1650 : void
1651 198284 : VacuumUpdateCosts(void)
1652 : {
1653 198284 : if (MyWorkerInfo)
1654 : {
1655 187644 : if (av_storage_param_cost_delay >= 0)
1656 0 : vacuum_cost_delay = av_storage_param_cost_delay;
1657 187644 : else if (autovacuum_vac_cost_delay >= 0)
1658 187644 : vacuum_cost_delay = autovacuum_vac_cost_delay;
1659 : else
1660 : /* fall back to VacuumCostDelay */
1661 0 : vacuum_cost_delay = VacuumCostDelay;
1662 :
1663 187644 : AutoVacuumUpdateCostLimit();
1664 : }
1665 : else
1666 : {
1667 : /* Must be explicit VACUUM or ANALYZE */
1668 10640 : vacuum_cost_delay = VacuumCostDelay;
1669 10640 : vacuum_cost_limit = VacuumCostLimit;
1670 : }
1671 :
1672 : /*
1673 : * If configuration changes are allowed to impact VacuumCostActive, make
1674 : * sure it is updated.
1675 : */
1676 198284 : if (VacuumFailsafeActive)
1677 : Assert(!VacuumCostActive);
1678 198284 : else if (vacuum_cost_delay > 0)
1679 187644 : VacuumCostActive = true;
1680 : else
1681 : {
1682 10640 : VacuumCostActive = false;
1683 10640 : VacuumCostBalance = 0;
1684 : }
1685 :
1686 : /*
1687 : * Since the cost logging requires a lock, avoid rendering the log message
1688 : * in case we are using a message level where the log wouldn't be emitted.
1689 : */
1690 198284 : if (MyWorkerInfo && message_level_is_interesting(DEBUG2))
1691 : {
1692 : Oid dboid,
1693 : tableoid;
1694 :
1695 : Assert(!LWLockHeldByMe(AutovacuumLock));
1696 :
1697 0 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1698 0 : dboid = MyWorkerInfo->wi_dboid;
1699 0 : tableoid = MyWorkerInfo->wi_tableoid;
1700 0 : LWLockRelease(AutovacuumLock);
1701 :
1702 0 : elog(DEBUG2,
1703 : "Autovacuum VacuumUpdateCosts(db=%u, rel=%u, dobalance=%s, cost_limit=%d, cost_delay=%g active=%s failsafe=%s)",
1704 : dboid, tableoid, pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance) ? "no" : "yes",
1705 : vacuum_cost_limit, vacuum_cost_delay,
1706 : vacuum_cost_delay > 0 ? "yes" : "no",
1707 : VacuumFailsafeActive ? "yes" : "no");
1708 : }
1709 198284 : }
1710 :
1711 : /*
1712 : * Update vacuum_cost_limit with the correct value for an autovacuum worker,
1713 : * given the value of other relevant cost limit parameters and the number of
1714 : * workers across which the limit must be balanced. Autovacuum workers must
1715 : * call this regularly in case av_nworkersForBalance has been updated by
1716 : * another worker or by the autovacuum launcher. They must also call it after a
1717 : * config reload.
1718 : */
1719 : void
1720 191706 : AutoVacuumUpdateCostLimit(void)
1721 : {
1722 191706 : if (!MyWorkerInfo)
1723 0 : return;
1724 :
1725 : /*
1726 : * note: in cost_limit, zero also means use value from elsewhere, because
1727 : * zero is not a valid value.
1728 : */
1729 :
1730 191706 : if (av_storage_param_cost_limit > 0)
1731 0 : vacuum_cost_limit = av_storage_param_cost_limit;
1732 : else
1733 : {
1734 : int nworkers_for_balance;
1735 :
1736 191706 : if (autovacuum_vac_cost_limit > 0)
1737 0 : vacuum_cost_limit = autovacuum_vac_cost_limit;
1738 : else
1739 191706 : vacuum_cost_limit = VacuumCostLimit;
1740 :
1741 : /* Only balance limit if no cost-related storage parameters specified */
1742 191706 : if (pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance))
1743 0 : return;
1744 :
1745 : Assert(vacuum_cost_limit > 0);
1746 :
1747 191706 : nworkers_for_balance = pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1748 :
1749 : /* There is at least 1 autovac worker (this worker) */
1750 191706 : if (nworkers_for_balance <= 0)
1751 0 : elog(ERROR, "nworkers_for_balance must be > 0");
1752 :
1753 191706 : vacuum_cost_limit = Max(vacuum_cost_limit / nworkers_for_balance, 1);
1754 : }
1755 : }
1756 :
1757 : /*
1758 : * autovac_recalculate_workers_for_balance
1759 : * Recalculate the number of workers to consider, given cost-related
1760 : * storage parameters and the current number of active workers.
1761 : *
1762 : * Caller must hold the AutovacuumLock in at least shared mode to access
1763 : * worker->wi_proc.
1764 : */
1765 : static void
1766 96084 : autovac_recalculate_workers_for_balance(void)
1767 : {
1768 : dlist_iter iter;
1769 : int orig_nworkers_for_balance;
1770 96084 : int nworkers_for_balance = 0;
1771 :
1772 : Assert(LWLockHeldByMe(AutovacuumLock));
1773 :
1774 96084 : orig_nworkers_for_balance =
1775 96084 : pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1776 :
1777 193794 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1778 : {
1779 97710 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1780 :
1781 195420 : if (worker->wi_proc == NULL ||
1782 97710 : pg_atomic_unlocked_test_flag(&worker->wi_dobalance))
1783 264 : continue;
1784 :
1785 97446 : nworkers_for_balance++;
1786 : }
1787 :
1788 96084 : if (nworkers_for_balance != orig_nworkers_for_balance)
1789 1804 : pg_atomic_write_u32(&AutoVacuumShmem->av_nworkersForBalance,
1790 : nworkers_for_balance);
1791 96084 : }
1792 :
1793 : /*
1794 : * get_database_list
1795 : * Return a list of all databases found in pg_database.
1796 : *
1797 : * The list and associated data is allocated in the caller's memory context,
1798 : * which is in charge of ensuring that it's properly cleaned up afterwards.
1799 : *
1800 : * Note: this is the only function in which the autovacuum launcher uses a
1801 : * transaction. Although we aren't attached to any particular database and
1802 : * therefore can't access most catalogs, we do have enough infrastructure
1803 : * to do a seqscan on pg_database.
1804 : */
1805 : static List *
1806 3274 : get_database_list(void)
1807 : {
1808 3274 : List *dblist = NIL;
1809 : Relation rel;
1810 : TableScanDesc scan;
1811 : HeapTuple tup;
1812 : MemoryContext resultcxt;
1813 :
1814 : /* This is the context that we will allocate our output data in */
1815 3274 : resultcxt = CurrentMemoryContext;
1816 :
1817 : /*
1818 : * Start a transaction so we can access pg_database.
1819 : */
1820 3274 : StartTransactionCommand();
1821 :
1822 3274 : rel = table_open(DatabaseRelationId, AccessShareLock);
1823 3274 : scan = table_beginscan_catalog(rel, 0, NULL);
1824 :
1825 13448 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1826 : {
1827 10174 : Form_pg_database pgdatabase = (Form_pg_database) GETSTRUCT(tup);
1828 : avw_dbase *avdb;
1829 : MemoryContext oldcxt;
1830 :
1831 : /*
1832 : * If database has partially been dropped, we can't, nor need to,
1833 : * vacuum it.
1834 : */
1835 10174 : if (database_is_invalid_form(pgdatabase))
1836 : {
1837 2 : elog(DEBUG2,
1838 : "autovacuum: skipping invalid database \"%s\"",
1839 : NameStr(pgdatabase->datname));
1840 2 : continue;
1841 : }
1842 :
1843 : /*
1844 : * Allocate our results in the caller's context, not the
1845 : * transaction's. We do this inside the loop, and restore the original
1846 : * context at the end, so that leaky things like heap_getnext() are
1847 : * not called in a potentially long-lived context.
1848 : */
1849 10172 : oldcxt = MemoryContextSwitchTo(resultcxt);
1850 :
1851 10172 : avdb = (avw_dbase *) palloc(sizeof(avw_dbase));
1852 :
1853 10172 : avdb->adw_datid = pgdatabase->oid;
1854 10172 : avdb->adw_name = pstrdup(NameStr(pgdatabase->datname));
1855 10172 : avdb->adw_frozenxid = pgdatabase->datfrozenxid;
1856 10172 : avdb->adw_minmulti = pgdatabase->datminmxid;
1857 : /* this gets set later: */
1858 10172 : avdb->adw_entry = NULL;
1859 :
1860 10172 : dblist = lappend(dblist, avdb);
1861 10172 : MemoryContextSwitchTo(oldcxt);
1862 : }
1863 :
1864 3274 : table_endscan(scan);
1865 3274 : table_close(rel, AccessShareLock);
1866 :
1867 3274 : CommitTransactionCommand();
1868 :
1869 : /* Be sure to restore caller's memory context */
1870 3274 : MemoryContextSwitchTo(resultcxt);
1871 :
1872 3274 : return dblist;
1873 : }
1874 :
1875 : /*
1876 : * Process a database table-by-table
1877 : *
1878 : * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in
1879 : * order not to ignore shutdown commands for too long.
1880 : */
1881 : static void
1882 2268 : do_autovacuum(void)
1883 : {
1884 : Relation classRel;
1885 : HeapTuple tuple;
1886 : TableScanDesc relScan;
1887 : Form_pg_database dbForm;
1888 2268 : List *table_oids = NIL;
1889 2268 : List *orphan_oids = NIL;
1890 : HASHCTL ctl;
1891 : HTAB *table_toast_map;
1892 : ListCell *volatile cell;
1893 : BufferAccessStrategy bstrategy;
1894 : ScanKeyData key;
1895 : TupleDesc pg_class_desc;
1896 : int effective_multixact_freeze_max_age;
1897 2268 : bool did_vacuum = false;
1898 2268 : bool found_concurrent_worker = false;
1899 : int i;
1900 :
1901 : /*
1902 : * StartTransactionCommand and CommitTransactionCommand will automatically
1903 : * switch to other contexts. We need this one to keep the list of
1904 : * relations to vacuum/analyze across transactions.
1905 : */
1906 2268 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
1907 : "Autovacuum worker",
1908 : ALLOCSET_DEFAULT_SIZES);
1909 2268 : MemoryContextSwitchTo(AutovacMemCxt);
1910 :
1911 : /* Start a transaction so our commands have one to play into. */
1912 2268 : StartTransactionCommand();
1913 :
1914 : /*
1915 : * This injection point is put in a transaction block to work with a wait
1916 : * that uses a condition variable.
1917 : */
1918 2268 : INJECTION_POINT("autovacuum-worker-start");
1919 :
1920 : /*
1921 : * Compute the multixact age for which freezing is urgent. This is
1922 : * normally autovacuum_multixact_freeze_max_age, but may be less if we are
1923 : * short of multixact member space.
1924 : */
1925 2266 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1926 :
1927 : /*
1928 : * Find the pg_database entry and select the default freeze ages. We use
1929 : * zero in template and nonconnectable databases, else the system-wide
1930 : * default.
1931 : */
1932 2266 : tuple = SearchSysCache1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
1933 2266 : if (!HeapTupleIsValid(tuple))
1934 0 : elog(ERROR, "cache lookup failed for database %u", MyDatabaseId);
1935 2266 : dbForm = (Form_pg_database) GETSTRUCT(tuple);
1936 :
1937 2266 : if (dbForm->datistemplate || !dbForm->datallowconn)
1938 : {
1939 1148 : default_freeze_min_age = 0;
1940 1148 : default_freeze_table_age = 0;
1941 1148 : default_multixact_freeze_min_age = 0;
1942 1148 : default_multixact_freeze_table_age = 0;
1943 : }
1944 : else
1945 : {
1946 1118 : default_freeze_min_age = vacuum_freeze_min_age;
1947 1118 : default_freeze_table_age = vacuum_freeze_table_age;
1948 1118 : default_multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1949 1118 : default_multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1950 : }
1951 :
1952 2266 : ReleaseSysCache(tuple);
1953 :
1954 : /* StartTransactionCommand changed elsewhere */
1955 2266 : MemoryContextSwitchTo(AutovacMemCxt);
1956 :
1957 2266 : classRel = table_open(RelationRelationId, AccessShareLock);
1958 :
1959 : /* create a copy so we can use it after closing pg_class */
1960 2266 : pg_class_desc = CreateTupleDescCopy(RelationGetDescr(classRel));
1961 :
1962 : /* create hash table for toast <-> main relid mapping */
1963 2266 : ctl.keysize = sizeof(Oid);
1964 2266 : ctl.entrysize = sizeof(av_relation);
1965 :
1966 2266 : table_toast_map = hash_create("TOAST to main relid map",
1967 : 100,
1968 : &ctl,
1969 : HASH_ELEM | HASH_BLOBS);
1970 :
1971 : /*
1972 : * Scan pg_class to determine which tables to vacuum.
1973 : *
1974 : * We do this in two passes: on the first one we collect the list of plain
1975 : * relations and materialized views, and on the second one we collect
1976 : * TOAST tables. The reason for doing the second pass is that during it we
1977 : * want to use the main relation's pg_class.reloptions entry if the TOAST
1978 : * table does not have any, and we cannot obtain it unless we know
1979 : * beforehand what's the main table OID.
1980 : *
1981 : * We need to check TOAST tables separately because in cases with short,
1982 : * wide tables there might be proportionally much more activity in the
1983 : * TOAST table than in its parent.
1984 : */
1985 2266 : relScan = table_beginscan_catalog(classRel, 0, NULL);
1986 :
1987 : /*
1988 : * On the first pass, we collect main tables to vacuum, and also the main
1989 : * table relid to TOAST relid mapping.
1990 : */
1991 955180 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
1992 : {
1993 952914 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
1994 : PgStat_StatTabEntry *tabentry;
1995 : AutoVacOpts *relopts;
1996 : Oid relid;
1997 : bool dovacuum;
1998 : bool doanalyze;
1999 : bool wraparound;
2000 :
2001 952914 : if (classForm->relkind != RELKIND_RELATION &&
2002 795048 : classForm->relkind != RELKIND_MATVIEW)
2003 795032 : continue;
2004 :
2005 157896 : relid = classForm->oid;
2006 :
2007 : /*
2008 : * Check if it is a temp table (presumably, of some other backend's).
2009 : * We cannot safely process other backends' temp tables.
2010 : */
2011 157896 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2012 : {
2013 : /*
2014 : * We just ignore it if the owning backend is still active and
2015 : * using the temporary schema. Also, for safety, ignore it if the
2016 : * namespace doesn't exist or isn't a temp namespace after all.
2017 : */
2018 14 : if (checkTempNamespaceStatus(classForm->relnamespace) == TEMP_NAMESPACE_IDLE)
2019 : {
2020 : /*
2021 : * The table seems to be orphaned -- although it might be that
2022 : * the owning backend has already deleted it and exited; our
2023 : * pg_class scan snapshot is not necessarily up-to-date
2024 : * anymore, so we could be looking at a committed-dead entry.
2025 : * Remember it so we can try to delete it later.
2026 : */
2027 0 : orphan_oids = lappend_oid(orphan_oids, relid);
2028 : }
2029 14 : continue;
2030 : }
2031 :
2032 : /* Fetch reloptions and the pgstat entry for this table */
2033 157882 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2034 157882 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2035 : relid);
2036 :
2037 : /* Check if it needs vacuum or analyze */
2038 157882 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2039 : effective_multixact_freeze_max_age,
2040 : &dovacuum, &doanalyze, &wraparound);
2041 :
2042 : /* Relations that need work are added to table_oids */
2043 157882 : if (dovacuum || doanalyze)
2044 59096 : table_oids = lappend_oid(table_oids, relid);
2045 :
2046 : /*
2047 : * Remember TOAST associations for the second pass. Note: we must do
2048 : * this whether or not the table is going to be vacuumed, because we
2049 : * don't automatically vacuum toast tables along the parent table.
2050 : */
2051 157882 : if (OidIsValid(classForm->reltoastrelid))
2052 : {
2053 : av_relation *hentry;
2054 : bool found;
2055 :
2056 186584 : hentry = hash_search(table_toast_map,
2057 93292 : &classForm->reltoastrelid,
2058 : HASH_ENTER, &found);
2059 :
2060 93292 : if (!found)
2061 : {
2062 : /* hash_search already filled in the key */
2063 93292 : hentry->ar_relid = relid;
2064 93292 : hentry->ar_hasrelopts = false;
2065 93292 : if (relopts != NULL)
2066 : {
2067 1798 : hentry->ar_hasrelopts = true;
2068 1798 : memcpy(&hentry->ar_reloptions, relopts,
2069 : sizeof(AutoVacOpts));
2070 : }
2071 : }
2072 : }
2073 : }
2074 :
2075 2266 : table_endscan(relScan);
2076 :
2077 : /* second pass: check TOAST tables */
2078 2266 : ScanKeyInit(&key,
2079 : Anum_pg_class_relkind,
2080 : BTEqualStrategyNumber, F_CHAREQ,
2081 : CharGetDatum(RELKIND_TOASTVALUE));
2082 :
2083 2266 : relScan = table_beginscan_catalog(classRel, 1, &key);
2084 95562 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2085 : {
2086 93296 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2087 : PgStat_StatTabEntry *tabentry;
2088 : Oid relid;
2089 93296 : AutoVacOpts *relopts = NULL;
2090 : bool dovacuum;
2091 : bool doanalyze;
2092 : bool wraparound;
2093 :
2094 : /*
2095 : * We cannot safely process other backends' temp tables, so skip 'em.
2096 : */
2097 93296 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2098 4 : continue;
2099 :
2100 93292 : relid = classForm->oid;
2101 :
2102 : /*
2103 : * fetch reloptions -- if this toast table does not have them, try the
2104 : * main rel
2105 : */
2106 93292 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2107 93292 : if (relopts == NULL)
2108 : {
2109 : av_relation *hentry;
2110 : bool found;
2111 :
2112 93290 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2113 93290 : if (found && hentry->ar_hasrelopts)
2114 1796 : relopts = &hentry->ar_reloptions;
2115 : }
2116 :
2117 : /* Fetch the pgstat entry for this table */
2118 93292 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2119 : relid);
2120 :
2121 93292 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2122 : effective_multixact_freeze_max_age,
2123 : &dovacuum, &doanalyze, &wraparound);
2124 :
2125 : /* ignore analyze for toast tables */
2126 93292 : if (dovacuum)
2127 34816 : table_oids = lappend_oid(table_oids, relid);
2128 : }
2129 :
2130 2266 : table_endscan(relScan);
2131 2266 : table_close(classRel, AccessShareLock);
2132 :
2133 : /*
2134 : * Recheck orphan temporary tables, and if they still seem orphaned, drop
2135 : * them. We'll eat a transaction per dropped table, which might seem
2136 : * excessive, but we should only need to do anything as a result of a
2137 : * previous backend crash, so this should not happen often enough to
2138 : * justify "optimizing". Using separate transactions ensures that we
2139 : * don't bloat the lock table if there are many temp tables to be dropped,
2140 : * and it ensures that we don't lose work if a deletion attempt fails.
2141 : */
2142 2266 : foreach(cell, orphan_oids)
2143 : {
2144 0 : Oid relid = lfirst_oid(cell);
2145 : Form_pg_class classForm;
2146 : ObjectAddress object;
2147 :
2148 : /*
2149 : * Check for user-requested abort.
2150 : */
2151 0 : CHECK_FOR_INTERRUPTS();
2152 :
2153 : /*
2154 : * Try to lock the table. If we can't get the lock immediately,
2155 : * somebody else is using (or dropping) the table, so it's not our
2156 : * concern anymore. Having the lock prevents race conditions below.
2157 : */
2158 0 : if (!ConditionalLockRelationOid(relid, AccessExclusiveLock))
2159 0 : continue;
2160 :
2161 : /*
2162 : * Re-fetch the pg_class tuple and re-check whether it still seems to
2163 : * be an orphaned temp table. If it's not there or no longer the same
2164 : * relation, ignore it.
2165 : */
2166 0 : tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2167 0 : if (!HeapTupleIsValid(tuple))
2168 : {
2169 : /* be sure to drop useless lock so we don't bloat lock table */
2170 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2171 0 : continue;
2172 : }
2173 0 : classForm = (Form_pg_class) GETSTRUCT(tuple);
2174 :
2175 : /*
2176 : * Make all the same tests made in the loop above. In event of OID
2177 : * counter wraparound, the pg_class entry we have now might be
2178 : * completely unrelated to the one we saw before.
2179 : */
2180 0 : if (!((classForm->relkind == RELKIND_RELATION ||
2181 0 : classForm->relkind == RELKIND_MATVIEW) &&
2182 0 : classForm->relpersistence == RELPERSISTENCE_TEMP))
2183 : {
2184 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2185 0 : continue;
2186 : }
2187 :
2188 0 : if (checkTempNamespaceStatus(classForm->relnamespace) != TEMP_NAMESPACE_IDLE)
2189 : {
2190 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2191 0 : continue;
2192 : }
2193 :
2194 : /*
2195 : * Try to lock the temp namespace, too. Even though we have lock on
2196 : * the table itself, there's a risk of deadlock against an incoming
2197 : * backend trying to clean out the temp namespace, in case this table
2198 : * has dependencies (such as sequences) that the backend's
2199 : * performDeletion call might visit in a different order. If we can
2200 : * get AccessShareLock on the namespace, that's sufficient to ensure
2201 : * we're not running concurrently with RemoveTempRelations. If we
2202 : * can't, back off and let RemoveTempRelations do its thing.
2203 : */
2204 0 : if (!ConditionalLockDatabaseObject(NamespaceRelationId,
2205 : classForm->relnamespace, 0,
2206 : AccessShareLock))
2207 : {
2208 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2209 0 : continue;
2210 : }
2211 :
2212 : /* OK, let's delete it */
2213 0 : ereport(LOG,
2214 : (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"",
2215 : get_database_name(MyDatabaseId),
2216 : get_namespace_name(classForm->relnamespace),
2217 : NameStr(classForm->relname))));
2218 :
2219 0 : object.classId = RelationRelationId;
2220 0 : object.objectId = relid;
2221 0 : object.objectSubId = 0;
2222 0 : performDeletion(&object, DROP_CASCADE,
2223 : PERFORM_DELETION_INTERNAL |
2224 : PERFORM_DELETION_QUIETLY |
2225 : PERFORM_DELETION_SKIP_EXTENSIONS);
2226 :
2227 : /*
2228 : * To commit the deletion, end current transaction and start a new
2229 : * one. Note this also releases the locks we took.
2230 : */
2231 0 : CommitTransactionCommand();
2232 0 : StartTransactionCommand();
2233 :
2234 : /* StartTransactionCommand changed current memory context */
2235 0 : MemoryContextSwitchTo(AutovacMemCxt);
2236 : }
2237 :
2238 : /*
2239 : * Optionally, create a buffer access strategy object for VACUUM to use.
2240 : * We use the same BufferAccessStrategy object for all tables VACUUMed by
2241 : * this worker to prevent autovacuum from blowing out shared buffers.
2242 : *
2243 : * VacuumBufferUsageLimit being set to 0 results in
2244 : * GetAccessStrategyWithSize returning NULL, effectively meaning we can
2245 : * use up to all of shared buffers.
2246 : *
2247 : * If we later enter failsafe mode on any of the tables being vacuumed, we
2248 : * will cease use of the BufferAccessStrategy only for that table.
2249 : *
2250 : * XXX should we consider adding code to adjust the size of this if
2251 : * VacuumBufferUsageLimit changes?
2252 : */
2253 2266 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, VacuumBufferUsageLimit);
2254 :
2255 : /*
2256 : * create a memory context to act as fake PortalContext, so that the
2257 : * contexts created in the vacuum code are cleaned up for each table.
2258 : */
2259 2266 : PortalContext = AllocSetContextCreate(AutovacMemCxt,
2260 : "Autovacuum Portal",
2261 : ALLOCSET_DEFAULT_SIZES);
2262 :
2263 : /*
2264 : * Perform operations on collected tables.
2265 : */
2266 96086 : foreach(cell, table_oids)
2267 : {
2268 93822 : Oid relid = lfirst_oid(cell);
2269 : HeapTuple classTup;
2270 : autovac_table *tab;
2271 : bool isshared;
2272 : bool skipit;
2273 : dlist_iter iter;
2274 :
2275 93822 : CHECK_FOR_INTERRUPTS();
2276 :
2277 : /*
2278 : * Check for config changes before processing each collected table.
2279 : */
2280 93822 : if (ConfigReloadPending)
2281 : {
2282 0 : ConfigReloadPending = false;
2283 0 : ProcessConfigFile(PGC_SIGHUP);
2284 :
2285 : /*
2286 : * You might be tempted to bail out if we see autovacuum is now
2287 : * disabled. Must resist that temptation -- this might be a
2288 : * for-wraparound emergency worker, in which case that would be
2289 : * entirely inappropriate.
2290 : */
2291 : }
2292 :
2293 : /*
2294 : * Find out whether the table is shared or not. (It's slightly
2295 : * annoying to fetch the syscache entry just for this, but in typical
2296 : * cases it adds little cost because table_recheck_autovac would
2297 : * refetch the entry anyway. We could buy that back by copying the
2298 : * tuple here and passing it to table_recheck_autovac, but that
2299 : * increases the odds of that function working with stale data.)
2300 : */
2301 93822 : classTup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
2302 93822 : if (!HeapTupleIsValid(classTup))
2303 0 : continue; /* somebody deleted the rel, forget it */
2304 93822 : isshared = ((Form_pg_class) GETSTRUCT(classTup))->relisshared;
2305 93822 : ReleaseSysCache(classTup);
2306 :
2307 : /*
2308 : * Hold schedule lock from here until we've claimed the table. We
2309 : * also need the AutovacuumLock to walk the worker array, but that one
2310 : * can just be a shared lock.
2311 : */
2312 93822 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2313 93822 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2314 :
2315 : /*
2316 : * Check whether the table is being vacuumed concurrently by another
2317 : * worker.
2318 : */
2319 93822 : skipit = false;
2320 191436 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
2321 : {
2322 97614 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
2323 :
2324 : /* ignore myself */
2325 97614 : if (worker == MyWorkerInfo)
2326 93822 : continue;
2327 :
2328 : /* ignore workers in other databases (unless table is shared) */
2329 3792 : if (!worker->wi_sharedrel && worker->wi_dboid != MyDatabaseId)
2330 0 : continue;
2331 :
2332 3792 : if (worker->wi_tableoid == relid)
2333 : {
2334 0 : skipit = true;
2335 0 : found_concurrent_worker = true;
2336 0 : break;
2337 : }
2338 : }
2339 93822 : LWLockRelease(AutovacuumLock);
2340 93822 : if (skipit)
2341 : {
2342 0 : LWLockRelease(AutovacuumScheduleLock);
2343 0 : continue;
2344 : }
2345 :
2346 : /*
2347 : * Store the table's OID in shared memory before releasing the
2348 : * schedule lock, so that other workers don't try to vacuum it
2349 : * concurrently. (We claim it here so as not to hold
2350 : * AutovacuumScheduleLock while rechecking the stats.)
2351 : */
2352 93822 : MyWorkerInfo->wi_tableoid = relid;
2353 93822 : MyWorkerInfo->wi_sharedrel = isshared;
2354 93822 : LWLockRelease(AutovacuumScheduleLock);
2355 :
2356 : /*
2357 : * Check whether pgstat data still says we need to vacuum this table.
2358 : * It could have changed if something else processed the table while
2359 : * we weren't looking. This doesn't entirely close the race condition,
2360 : * but it is very small.
2361 : */
2362 93822 : MemoryContextSwitchTo(AutovacMemCxt);
2363 93822 : tab = table_recheck_autovac(relid, table_toast_map, pg_class_desc,
2364 : effective_multixact_freeze_max_age);
2365 93822 : if (tab == NULL)
2366 : {
2367 : /* someone else vacuumed the table, or it went away */
2368 0 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2369 0 : MyWorkerInfo->wi_tableoid = InvalidOid;
2370 0 : MyWorkerInfo->wi_sharedrel = false;
2371 0 : LWLockRelease(AutovacuumScheduleLock);
2372 0 : continue;
2373 : }
2374 :
2375 : /*
2376 : * Save the cost-related storage parameter values in global variables
2377 : * for reference when updating vacuum_cost_delay and vacuum_cost_limit
2378 : * during vacuuming this table.
2379 : */
2380 93822 : av_storage_param_cost_delay = tab->at_storage_param_vac_cost_delay;
2381 93822 : av_storage_param_cost_limit = tab->at_storage_param_vac_cost_limit;
2382 :
2383 : /*
2384 : * We only expect this worker to ever set the flag, so don't bother
2385 : * checking the return value. We shouldn't have to retry.
2386 : */
2387 93822 : if (tab->at_dobalance)
2388 93822 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2389 : else
2390 0 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
2391 :
2392 93822 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2393 93822 : autovac_recalculate_workers_for_balance();
2394 93822 : LWLockRelease(AutovacuumLock);
2395 :
2396 : /*
2397 : * We wait until this point to update cost delay and cost limit
2398 : * values, even though we reloaded the configuration file above, so
2399 : * that we can take into account the cost-related storage parameters.
2400 : */
2401 93822 : VacuumUpdateCosts();
2402 :
2403 :
2404 : /* clean up memory before each iteration */
2405 93822 : MemoryContextReset(PortalContext);
2406 :
2407 : /*
2408 : * Save the relation name for a possible error message, to avoid a
2409 : * catalog lookup in case of an error. If any of these return NULL,
2410 : * then the relation has been dropped since last we checked; skip it.
2411 : * Note: they must live in a long-lived memory context because we call
2412 : * vacuum and analyze in different transactions.
2413 : */
2414 :
2415 93822 : tab->at_relname = get_rel_name(tab->at_relid);
2416 93822 : tab->at_nspname = get_namespace_name(get_rel_namespace(tab->at_relid));
2417 93822 : tab->at_datname = get_database_name(MyDatabaseId);
2418 93822 : if (!tab->at_relname || !tab->at_nspname || !tab->at_datname)
2419 0 : goto deleted;
2420 :
2421 : /*
2422 : * We will abort vacuuming the current table if something errors out,
2423 : * and continue with the next one in schedule; in particular, this
2424 : * happens if we are interrupted with SIGINT.
2425 : */
2426 93822 : PG_TRY();
2427 : {
2428 : /* Use PortalContext for any per-table allocations */
2429 93822 : MemoryContextSwitchTo(PortalContext);
2430 :
2431 : /* have at it */
2432 93822 : autovacuum_do_vac_analyze(tab, bstrategy);
2433 :
2434 : /*
2435 : * Clear a possible query-cancel signal, to avoid a late reaction
2436 : * to an automatically-sent signal because of vacuuming the
2437 : * current table (we're done with it, so it would make no sense to
2438 : * cancel at this point.)
2439 : */
2440 93820 : QueryCancelPending = false;
2441 : }
2442 0 : PG_CATCH();
2443 : {
2444 : /*
2445 : * Abort the transaction, start a new one, and proceed with the
2446 : * next table in our list.
2447 : */
2448 0 : HOLD_INTERRUPTS();
2449 0 : if (tab->at_params.options & VACOPT_VACUUM)
2450 0 : errcontext("automatic vacuum of table \"%s.%s.%s\"",
2451 : tab->at_datname, tab->at_nspname, tab->at_relname);
2452 : else
2453 0 : errcontext("automatic analyze of table \"%s.%s.%s\"",
2454 : tab->at_datname, tab->at_nspname, tab->at_relname);
2455 0 : EmitErrorReport();
2456 :
2457 : /* this resets ProcGlobal->statusFlags[i] too */
2458 0 : AbortOutOfAnyTransaction();
2459 0 : FlushErrorState();
2460 0 : MemoryContextReset(PortalContext);
2461 :
2462 : /* restart our transaction for the following operations */
2463 0 : StartTransactionCommand();
2464 0 : RESUME_INTERRUPTS();
2465 : }
2466 93820 : PG_END_TRY();
2467 :
2468 : /* Make sure we're back in AutovacMemCxt */
2469 93820 : MemoryContextSwitchTo(AutovacMemCxt);
2470 :
2471 93820 : did_vacuum = true;
2472 :
2473 : /* ProcGlobal->statusFlags[i] are reset at the next end of xact */
2474 :
2475 : /* be tidy */
2476 93820 : deleted:
2477 93820 : if (tab->at_datname != NULL)
2478 93820 : pfree(tab->at_datname);
2479 93820 : if (tab->at_nspname != NULL)
2480 93820 : pfree(tab->at_nspname);
2481 93820 : if (tab->at_relname != NULL)
2482 93820 : pfree(tab->at_relname);
2483 93820 : pfree(tab);
2484 :
2485 : /*
2486 : * Remove my info from shared memory. We set wi_dobalance on the
2487 : * assumption that we are more likely than not to vacuum a table with
2488 : * no cost-related storage parameters next, so we want to claim our
2489 : * share of I/O as soon as possible to avoid thrashing the global
2490 : * balance.
2491 : */
2492 93820 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2493 93820 : MyWorkerInfo->wi_tableoid = InvalidOid;
2494 93820 : MyWorkerInfo->wi_sharedrel = false;
2495 93820 : LWLockRelease(AutovacuumScheduleLock);
2496 93820 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2497 : }
2498 :
2499 : /*
2500 : * Perform additional work items, as requested by backends.
2501 : */
2502 2264 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2503 581848 : for (i = 0; i < NUM_WORKITEMS; i++)
2504 : {
2505 579584 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
2506 :
2507 579584 : if (!workitem->avw_used)
2508 579578 : continue;
2509 6 : if (workitem->avw_active)
2510 0 : continue;
2511 6 : if (workitem->avw_database != MyDatabaseId)
2512 0 : continue;
2513 :
2514 : /* claim this one, and release lock while performing it */
2515 6 : workitem->avw_active = true;
2516 6 : LWLockRelease(AutovacuumLock);
2517 :
2518 6 : perform_work_item(workitem);
2519 :
2520 : /*
2521 : * Check for config changes before acquiring lock for further jobs.
2522 : */
2523 6 : CHECK_FOR_INTERRUPTS();
2524 6 : if (ConfigReloadPending)
2525 : {
2526 0 : ConfigReloadPending = false;
2527 0 : ProcessConfigFile(PGC_SIGHUP);
2528 0 : VacuumUpdateCosts();
2529 : }
2530 :
2531 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2532 :
2533 : /* and mark it done */
2534 6 : workitem->avw_active = false;
2535 6 : workitem->avw_used = false;
2536 : }
2537 2264 : LWLockRelease(AutovacuumLock);
2538 :
2539 : /*
2540 : * We leak table_toast_map here (among other things), but since we're
2541 : * going away soon, it's not a problem.
2542 : */
2543 :
2544 : /*
2545 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We
2546 : * only need to do this once, not after each table.
2547 : *
2548 : * Even if we didn't vacuum anything, it may still be important to do
2549 : * this, because one indirect effect of vac_update_datfrozenxid() is to
2550 : * update TransamVariables->xidVacLimit. That might need to be done even
2551 : * if we haven't vacuumed anything, because relations with older
2552 : * relfrozenxid values or other databases with older datfrozenxid values
2553 : * might have been dropped, allowing xidVacLimit to advance.
2554 : *
2555 : * However, it's also important not to do this blindly in all cases,
2556 : * because when autovacuum=off this will restart the autovacuum launcher.
2557 : * If we're not careful, an infinite loop can result, where workers find
2558 : * no work to do and restart the launcher, which starts another worker in
2559 : * the same database that finds no work to do. To prevent that, we skip
2560 : * this if (1) we found no work to do and (2) we skipped at least one
2561 : * table due to concurrent autovacuum activity. In that case, the other
2562 : * worker has already done it, or will do so when it finishes.
2563 : */
2564 2264 : if (did_vacuum || !found_concurrent_worker)
2565 2264 : vac_update_datfrozenxid();
2566 :
2567 : /* Finally close out the last transaction. */
2568 2264 : CommitTransactionCommand();
2569 2264 : }
2570 :
2571 : /*
2572 : * Execute a previously registered work item.
2573 : */
2574 : static void
2575 6 : perform_work_item(AutoVacuumWorkItem *workitem)
2576 : {
2577 6 : char *cur_datname = NULL;
2578 6 : char *cur_nspname = NULL;
2579 6 : char *cur_relname = NULL;
2580 :
2581 : /*
2582 : * Note we do not store table info in MyWorkerInfo, since this is not
2583 : * vacuuming proper.
2584 : */
2585 :
2586 : /*
2587 : * Save the relation name for a possible error message, to avoid a catalog
2588 : * lookup in case of an error. If any of these return NULL, then the
2589 : * relation has been dropped since last we checked; skip it.
2590 : */
2591 : Assert(CurrentMemoryContext == AutovacMemCxt);
2592 :
2593 6 : cur_relname = get_rel_name(workitem->avw_relation);
2594 6 : cur_nspname = get_namespace_name(get_rel_namespace(workitem->avw_relation));
2595 6 : cur_datname = get_database_name(MyDatabaseId);
2596 6 : if (!cur_relname || !cur_nspname || !cur_datname)
2597 0 : goto deleted2;
2598 :
2599 6 : autovac_report_workitem(workitem, cur_nspname, cur_relname);
2600 :
2601 : /* clean up memory before each work item */
2602 6 : MemoryContextReset(PortalContext);
2603 :
2604 : /*
2605 : * We will abort the current work item if something errors out, and
2606 : * continue with the next one; in particular, this happens if we are
2607 : * interrupted with SIGINT. Note that this means that the work item list
2608 : * can be lossy.
2609 : */
2610 6 : PG_TRY();
2611 : {
2612 : /* Use PortalContext for any per-work-item allocations */
2613 6 : MemoryContextSwitchTo(PortalContext);
2614 :
2615 : /*
2616 : * Have at it. Functions called here are responsible for any required
2617 : * user switch and sandbox.
2618 : */
2619 6 : switch (workitem->avw_type)
2620 : {
2621 6 : case AVW_BRINSummarizeRange:
2622 6 : DirectFunctionCall2(brin_summarize_range,
2623 : ObjectIdGetDatum(workitem->avw_relation),
2624 : Int64GetDatum((int64) workitem->avw_blockNumber));
2625 6 : break;
2626 0 : default:
2627 0 : elog(WARNING, "unrecognized work item found: type %d",
2628 : workitem->avw_type);
2629 0 : break;
2630 : }
2631 :
2632 : /*
2633 : * Clear a possible query-cancel signal, to avoid a late reaction to
2634 : * an automatically-sent signal because of vacuuming the current table
2635 : * (we're done with it, so it would make no sense to cancel at this
2636 : * point.)
2637 : */
2638 6 : QueryCancelPending = false;
2639 : }
2640 0 : PG_CATCH();
2641 : {
2642 : /*
2643 : * Abort the transaction, start a new one, and proceed with the next
2644 : * table in our list.
2645 : */
2646 0 : HOLD_INTERRUPTS();
2647 0 : errcontext("processing work entry for relation \"%s.%s.%s\"",
2648 : cur_datname, cur_nspname, cur_relname);
2649 0 : EmitErrorReport();
2650 :
2651 : /* this resets ProcGlobal->statusFlags[i] too */
2652 0 : AbortOutOfAnyTransaction();
2653 0 : FlushErrorState();
2654 0 : MemoryContextReset(PortalContext);
2655 :
2656 : /* restart our transaction for the following operations */
2657 0 : StartTransactionCommand();
2658 0 : RESUME_INTERRUPTS();
2659 : }
2660 6 : PG_END_TRY();
2661 :
2662 : /* Make sure we're back in AutovacMemCxt */
2663 6 : MemoryContextSwitchTo(AutovacMemCxt);
2664 :
2665 : /* We intentionally do not set did_vacuum here */
2666 :
2667 : /* be tidy */
2668 6 : deleted2:
2669 6 : if (cur_datname)
2670 6 : pfree(cur_datname);
2671 6 : if (cur_nspname)
2672 6 : pfree(cur_nspname);
2673 6 : if (cur_relname)
2674 6 : pfree(cur_relname);
2675 6 : }
2676 :
2677 : /*
2678 : * extract_autovac_opts
2679 : *
2680 : * Given a relation's pg_class tuple, return the AutoVacOpts portion of
2681 : * reloptions, if set; otherwise, return NULL.
2682 : *
2683 : * Note: callers do not have a relation lock on the table at this point,
2684 : * so the table could have been dropped, and its catalog rows gone, after
2685 : * we acquired the pg_class row. If pg_class had a TOAST table, this would
2686 : * be a risk; fortunately, it doesn't.
2687 : */
2688 : static AutoVacOpts *
2689 344996 : extract_autovac_opts(HeapTuple tup, TupleDesc pg_class_desc)
2690 : {
2691 : bytea *relopts;
2692 : AutoVacOpts *av;
2693 :
2694 : Assert(((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_RELATION ||
2695 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_MATVIEW ||
2696 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_TOASTVALUE);
2697 :
2698 344996 : relopts = extractRelOptions(tup, pg_class_desc, NULL);
2699 344996 : if (relopts == NULL)
2700 341890 : return NULL;
2701 :
2702 3106 : av = palloc(sizeof(AutoVacOpts));
2703 3106 : memcpy(av, &(((StdRdOptions *) relopts)->autovacuum), sizeof(AutoVacOpts));
2704 3106 : pfree(relopts);
2705 :
2706 3106 : return av;
2707 : }
2708 :
2709 :
2710 : /*
2711 : * table_recheck_autovac
2712 : *
2713 : * Recheck whether a table still needs vacuum or analyze. Return value is a
2714 : * valid autovac_table pointer if it does, NULL otherwise.
2715 : *
2716 : * Note that the returned autovac_table does not have the name fields set.
2717 : */
2718 : static autovac_table *
2719 93822 : table_recheck_autovac(Oid relid, HTAB *table_toast_map,
2720 : TupleDesc pg_class_desc,
2721 : int effective_multixact_freeze_max_age)
2722 : {
2723 : Form_pg_class classForm;
2724 : HeapTuple classTup;
2725 : bool dovacuum;
2726 : bool doanalyze;
2727 93822 : autovac_table *tab = NULL;
2728 : bool wraparound;
2729 : AutoVacOpts *avopts;
2730 :
2731 : /* fetch the relation's relcache entry */
2732 93822 : classTup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2733 93822 : if (!HeapTupleIsValid(classTup))
2734 0 : return NULL;
2735 93822 : classForm = (Form_pg_class) GETSTRUCT(classTup);
2736 :
2737 : /*
2738 : * Get the applicable reloptions. If it is a TOAST table, try to get the
2739 : * main table reloptions if the toast table itself doesn't have.
2740 : */
2741 93822 : avopts = extract_autovac_opts(classTup, pg_class_desc);
2742 93822 : if (classForm->relkind == RELKIND_TOASTVALUE &&
2743 34814 : avopts == NULL && table_toast_map != NULL)
2744 : {
2745 : av_relation *hentry;
2746 : bool found;
2747 :
2748 34814 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2749 34814 : if (found && hentry->ar_hasrelopts)
2750 1248 : avopts = &hentry->ar_reloptions;
2751 : }
2752 :
2753 93822 : recheck_relation_needs_vacanalyze(relid, avopts, classForm,
2754 : effective_multixact_freeze_max_age,
2755 : &dovacuum, &doanalyze, &wraparound);
2756 :
2757 : /* OK, it needs something done */
2758 93822 : if (doanalyze || dovacuum)
2759 : {
2760 : int freeze_min_age;
2761 : int freeze_table_age;
2762 : int multixact_freeze_min_age;
2763 : int multixact_freeze_table_age;
2764 : int log_min_duration;
2765 :
2766 : /*
2767 : * Calculate the vacuum cost parameters and the freeze ages. If there
2768 : * are options set in pg_class.reloptions, use them; in the case of a
2769 : * toast table, try the main table too. Otherwise use the GUC
2770 : * defaults, autovacuum's own first and plain vacuum second.
2771 : */
2772 :
2773 : /* -1 in autovac setting means use log_autovacuum_min_duration */
2774 2512 : log_min_duration = (avopts && avopts->log_min_duration >= 0)
2775 : ? avopts->log_min_duration
2776 96334 : : Log_autovacuum_min_duration;
2777 :
2778 : /* these do not have autovacuum-specific settings */
2779 2512 : freeze_min_age = (avopts && avopts->freeze_min_age >= 0)
2780 : ? avopts->freeze_min_age
2781 96334 : : default_freeze_min_age;
2782 :
2783 2512 : freeze_table_age = (avopts && avopts->freeze_table_age >= 0)
2784 : ? avopts->freeze_table_age
2785 96334 : : default_freeze_table_age;
2786 :
2787 96334 : multixact_freeze_min_age = (avopts &&
2788 2512 : avopts->multixact_freeze_min_age >= 0)
2789 : ? avopts->multixact_freeze_min_age
2790 96334 : : default_multixact_freeze_min_age;
2791 :
2792 96334 : multixact_freeze_table_age = (avopts &&
2793 2512 : avopts->multixact_freeze_table_age >= 0)
2794 : ? avopts->multixact_freeze_table_age
2795 96334 : : default_multixact_freeze_table_age;
2796 :
2797 93822 : tab = palloc(sizeof(autovac_table));
2798 93822 : tab->at_relid = relid;
2799 93822 : tab->at_sharedrel = classForm->relisshared;
2800 :
2801 : /*
2802 : * Select VACUUM options. Note we don't say VACOPT_PROCESS_TOAST, so
2803 : * that vacuum() skips toast relations. Also note we tell vacuum() to
2804 : * skip vac_update_datfrozenxid(); we'll do that separately.
2805 : */
2806 93822 : tab->at_params.options =
2807 93822 : (dovacuum ? (VACOPT_VACUUM |
2808 : VACOPT_PROCESS_MAIN |
2809 93822 : VACOPT_SKIP_DATABASE_STATS) : 0) |
2810 93822 : (doanalyze ? VACOPT_ANALYZE : 0) |
2811 93822 : (!wraparound ? VACOPT_SKIP_LOCKED : 0);
2812 :
2813 : /*
2814 : * index_cleanup and truncate are unspecified at first in autovacuum.
2815 : * They will be filled in with usable values using their reloptions
2816 : * (or reloption defaults) later.
2817 : */
2818 93822 : tab->at_params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
2819 93822 : tab->at_params.truncate = VACOPTVALUE_UNSPECIFIED;
2820 : /* As of now, we don't support parallel vacuum for autovacuum */
2821 93822 : tab->at_params.nworkers = -1;
2822 93822 : tab->at_params.freeze_min_age = freeze_min_age;
2823 93822 : tab->at_params.freeze_table_age = freeze_table_age;
2824 93822 : tab->at_params.multixact_freeze_min_age = multixact_freeze_min_age;
2825 93822 : tab->at_params.multixact_freeze_table_age = multixact_freeze_table_age;
2826 93822 : tab->at_params.is_wraparound = wraparound;
2827 93822 : tab->at_params.log_min_duration = log_min_duration;
2828 93822 : tab->at_params.toast_parent = InvalidOid;
2829 93822 : tab->at_storage_param_vac_cost_limit = avopts ?
2830 93822 : avopts->vacuum_cost_limit : 0;
2831 93822 : tab->at_storage_param_vac_cost_delay = avopts ?
2832 93822 : avopts->vacuum_cost_delay : -1;
2833 93822 : tab->at_relname = NULL;
2834 93822 : tab->at_nspname = NULL;
2835 93822 : tab->at_datname = NULL;
2836 :
2837 : /*
2838 : * If any of the cost delay parameters has been set individually for
2839 : * this table, disable the balancing algorithm.
2840 : */
2841 93822 : tab->at_dobalance =
2842 96334 : !(avopts && (avopts->vacuum_cost_limit > 0 ||
2843 2512 : avopts->vacuum_cost_delay >= 0));
2844 : }
2845 :
2846 93822 : heap_freetuple(classTup);
2847 93822 : return tab;
2848 : }
2849 :
2850 : /*
2851 : * recheck_relation_needs_vacanalyze
2852 : *
2853 : * Subroutine for table_recheck_autovac.
2854 : *
2855 : * Fetch the pgstat of a relation and recheck whether a relation
2856 : * needs to be vacuumed or analyzed.
2857 : */
2858 : static void
2859 93822 : recheck_relation_needs_vacanalyze(Oid relid,
2860 : AutoVacOpts *avopts,
2861 : Form_pg_class classForm,
2862 : int effective_multixact_freeze_max_age,
2863 : bool *dovacuum,
2864 : bool *doanalyze,
2865 : bool *wraparound)
2866 : {
2867 : PgStat_StatTabEntry *tabentry;
2868 :
2869 : /* fetch the pgstat table entry */
2870 93822 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2871 : relid);
2872 :
2873 93822 : relation_needs_vacanalyze(relid, avopts, classForm, tabentry,
2874 : effective_multixact_freeze_max_age,
2875 : dovacuum, doanalyze, wraparound);
2876 :
2877 : /* ignore ANALYZE for toast tables */
2878 93822 : if (classForm->relkind == RELKIND_TOASTVALUE)
2879 34814 : *doanalyze = false;
2880 93822 : }
2881 :
2882 : /*
2883 : * relation_needs_vacanalyze
2884 : *
2885 : * Check whether a relation needs to be vacuumed or analyzed; return each into
2886 : * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is
2887 : * being forced because of Xid or multixact wraparound.
2888 : *
2889 : * relopts is a pointer to the AutoVacOpts options (either for itself in the
2890 : * case of a plain table, or for either itself or its parent table in the case
2891 : * of a TOAST table), NULL if none; tabentry is the pgstats entry, which can be
2892 : * NULL.
2893 : *
2894 : * A table needs to be vacuumed if the number of dead tuples exceeds a
2895 : * threshold. This threshold is calculated as
2896 : *
2897 : * threshold = vac_base_thresh + vac_scale_factor * reltuples
2898 : *
2899 : * For analyze, the analysis done is that the number of tuples inserted,
2900 : * deleted and updated since the last analyze exceeds a threshold calculated
2901 : * in the same fashion as above. Note that the cumulative stats system stores
2902 : * the number of tuples (both live and dead) that there were as of the last
2903 : * analyze. This is asymmetric to the VACUUM case.
2904 : *
2905 : * We also force vacuum if the table's relfrozenxid is more than freeze_max_age
2906 : * transactions back, and if its relminmxid is more than
2907 : * multixact_freeze_max_age multixacts back.
2908 : *
2909 : * A table whose autovacuum_enabled option is false is
2910 : * automatically skipped (unless we have to vacuum it due to freeze_max_age).
2911 : * Thus autovacuum can be disabled for specific tables. Also, when the cumulative
2912 : * stats system does not have data about a table, it will be skipped.
2913 : *
2914 : * A table whose vac_base_thresh value is < 0 takes the base value from the
2915 : * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor
2916 : * value < 0 is substituted with the value of
2917 : * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze.
2918 : */
2919 : static void
2920 344996 : relation_needs_vacanalyze(Oid relid,
2921 : AutoVacOpts *relopts,
2922 : Form_pg_class classForm,
2923 : PgStat_StatTabEntry *tabentry,
2924 : int effective_multixact_freeze_max_age,
2925 : /* output params below */
2926 : bool *dovacuum,
2927 : bool *doanalyze,
2928 : bool *wraparound)
2929 : {
2930 : bool force_vacuum;
2931 : bool av_enabled;
2932 : float4 reltuples; /* pg_class.reltuples */
2933 :
2934 : /* constants from reloptions or GUC variables */
2935 : int vac_base_thresh,
2936 : vac_ins_base_thresh,
2937 : anl_base_thresh;
2938 : float4 vac_scale_factor,
2939 : vac_ins_scale_factor,
2940 : anl_scale_factor;
2941 :
2942 : /* thresholds calculated from above constants */
2943 : float4 vacthresh,
2944 : vacinsthresh,
2945 : anlthresh;
2946 :
2947 : /* number of vacuum (resp. analyze) tuples at this time */
2948 : float4 vactuples,
2949 : instuples,
2950 : anltuples;
2951 :
2952 : /* freeze parameters */
2953 : int freeze_max_age;
2954 : int multixact_freeze_max_age;
2955 : TransactionId xidForceLimit;
2956 : TransactionId relfrozenxid;
2957 : MultiXactId multiForceLimit;
2958 :
2959 : Assert(classForm != NULL);
2960 : Assert(OidIsValid(relid));
2961 :
2962 : /*
2963 : * Determine vacuum/analyze equation parameters. We have two possible
2964 : * sources: the passed reloptions (which could be a main table or a toast
2965 : * table), or the autovacuum GUC variables.
2966 : */
2967 :
2968 : /* -1 in autovac setting means use plain vacuum_scale_factor */
2969 6150 : vac_scale_factor = (relopts && relopts->vacuum_scale_factor >= 0)
2970 0 : ? relopts->vacuum_scale_factor
2971 351146 : : autovacuum_vac_scale;
2972 :
2973 6150 : vac_base_thresh = (relopts && relopts->vacuum_threshold >= 0)
2974 : ? relopts->vacuum_threshold
2975 351146 : : autovacuum_vac_thresh;
2976 :
2977 6150 : vac_ins_scale_factor = (relopts && relopts->vacuum_ins_scale_factor >= 0)
2978 0 : ? relopts->vacuum_ins_scale_factor
2979 351146 : : autovacuum_vac_ins_scale;
2980 :
2981 : /* -1 is used to disable insert vacuums */
2982 6150 : vac_ins_base_thresh = (relopts && relopts->vacuum_ins_threshold >= -1)
2983 : ? relopts->vacuum_ins_threshold
2984 351146 : : autovacuum_vac_ins_thresh;
2985 :
2986 6150 : anl_scale_factor = (relopts && relopts->analyze_scale_factor >= 0)
2987 0 : ? relopts->analyze_scale_factor
2988 351146 : : autovacuum_anl_scale;
2989 :
2990 6150 : anl_base_thresh = (relopts && relopts->analyze_threshold >= 0)
2991 : ? relopts->analyze_threshold
2992 351146 : : autovacuum_anl_thresh;
2993 :
2994 6150 : freeze_max_age = (relopts && relopts->freeze_max_age >= 0)
2995 0 : ? Min(relopts->freeze_max_age, autovacuum_freeze_max_age)
2996 351146 : : autovacuum_freeze_max_age;
2997 :
2998 6150 : multixact_freeze_max_age = (relopts && relopts->multixact_freeze_max_age >= 0)
2999 0 : ? Min(relopts->multixact_freeze_max_age, effective_multixact_freeze_max_age)
3000 351146 : : effective_multixact_freeze_max_age;
3001 :
3002 344996 : av_enabled = (relopts ? relopts->enabled : true);
3003 :
3004 : /* Force vacuum if table is at risk of wraparound */
3005 344996 : xidForceLimit = recentXid - freeze_max_age;
3006 344996 : if (xidForceLimit < FirstNormalTransactionId)
3007 0 : xidForceLimit -= FirstNormalTransactionId;
3008 344996 : relfrozenxid = classForm->relfrozenxid;
3009 689992 : force_vacuum = (TransactionIdIsNormal(relfrozenxid) &&
3010 344996 : TransactionIdPrecedes(relfrozenxid, xidForceLimit));
3011 344996 : if (!force_vacuum)
3012 : {
3013 158016 : MultiXactId relminmxid = classForm->relminmxid;
3014 :
3015 158016 : multiForceLimit = recentMulti - multixact_freeze_max_age;
3016 158016 : if (multiForceLimit < FirstMultiXactId)
3017 0 : multiForceLimit -= FirstMultiXactId;
3018 316032 : force_vacuum = MultiXactIdIsValid(relminmxid) &&
3019 158016 : MultiXactIdPrecedes(relminmxid, multiForceLimit);
3020 : }
3021 344996 : *wraparound = force_vacuum;
3022 :
3023 : /* User disabled it in pg_class.reloptions? (But ignore if at risk) */
3024 344996 : if (!av_enabled && !force_vacuum)
3025 : {
3026 1096 : *doanalyze = false;
3027 1096 : *dovacuum = false;
3028 1096 : return;
3029 : }
3030 :
3031 : /*
3032 : * If we found stats for the table, and autovacuum is currently enabled,
3033 : * make a threshold-based decision whether to vacuum and/or analyze. If
3034 : * autovacuum is currently disabled, we must be here for anti-wraparound
3035 : * vacuuming only, so don't vacuum (or analyze) anything that's not being
3036 : * forced.
3037 : */
3038 343900 : if (PointerIsValid(tabentry) && AutoVacuumingActive())
3039 : {
3040 339434 : reltuples = classForm->reltuples;
3041 339434 : vactuples = tabentry->dead_tuples;
3042 339434 : instuples = tabentry->ins_since_vacuum;
3043 339434 : anltuples = tabentry->mod_since_analyze;
3044 :
3045 : /* If the table hasn't yet been vacuumed, take reltuples as zero */
3046 339434 : if (reltuples < 0)
3047 1698 : reltuples = 0;
3048 :
3049 339434 : vacthresh = (float4) vac_base_thresh + vac_scale_factor * reltuples;
3050 339434 : vacinsthresh = (float4) vac_ins_base_thresh + vac_ins_scale_factor * reltuples;
3051 339434 : anlthresh = (float4) anl_base_thresh + anl_scale_factor * reltuples;
3052 :
3053 : /*
3054 : * Note that we don't need to take special consideration for stat
3055 : * reset, because if that happens, the last vacuum and analyze counts
3056 : * will be reset too.
3057 : */
3058 339434 : if (vac_ins_base_thresh >= 0)
3059 339434 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: %.0f (threshold %.0f), anl: %.0f (threshold %.0f)",
3060 : NameStr(classForm->relname),
3061 : vactuples, vacthresh, instuples, vacinsthresh, anltuples, anlthresh);
3062 : else
3063 0 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: (disabled), anl: %.0f (threshold %.0f)",
3064 : NameStr(classForm->relname),
3065 : vactuples, vacthresh, anltuples, anlthresh);
3066 :
3067 : /* Determine if this table needs vacuum or analyze. */
3068 493500 : *dovacuum = force_vacuum || (vactuples > vacthresh) ||
3069 154066 : (vac_ins_base_thresh >= 0 && instuples > vacinsthresh);
3070 339434 : *doanalyze = (anltuples > anlthresh);
3071 : }
3072 : else
3073 : {
3074 : /*
3075 : * Skip a table not found in stat hash, unless we have to force vacuum
3076 : * for anti-wrap purposes. If it's not acted upon, there's no need to
3077 : * vacuum it.
3078 : */
3079 4466 : *dovacuum = force_vacuum;
3080 4466 : *doanalyze = false;
3081 : }
3082 :
3083 : /* ANALYZE refuses to work with pg_statistic */
3084 343900 : if (relid == StatisticRelationId)
3085 3118 : *doanalyze = false;
3086 : }
3087 :
3088 : /*
3089 : * autovacuum_do_vac_analyze
3090 : * Vacuum and/or analyze the specified table
3091 : *
3092 : * We expect the caller to have switched into a memory context that won't
3093 : * disappear at transaction commit.
3094 : */
3095 : static void
3096 93822 : autovacuum_do_vac_analyze(autovac_table *tab, BufferAccessStrategy bstrategy)
3097 : {
3098 : RangeVar *rangevar;
3099 : VacuumRelation *rel;
3100 : List *rel_list;
3101 : MemoryContext vac_context;
3102 :
3103 : /* Let pgstat know what we're doing */
3104 93822 : autovac_report_activity(tab);
3105 :
3106 : /* Set up one VacuumRelation target, identified by OID, for vacuum() */
3107 93822 : rangevar = makeRangeVar(tab->at_nspname, tab->at_relname, -1);
3108 93822 : rel = makeVacuumRelation(rangevar, tab->at_relid, NIL);
3109 93822 : rel_list = list_make1(rel);
3110 :
3111 93822 : vac_context = AllocSetContextCreate(CurrentMemoryContext,
3112 : "Vacuum",
3113 : ALLOCSET_DEFAULT_SIZES);
3114 :
3115 93822 : vacuum(rel_list, &tab->at_params, bstrategy, vac_context, true);
3116 :
3117 93820 : MemoryContextDelete(vac_context);
3118 93820 : }
3119 :
3120 : /*
3121 : * autovac_report_activity
3122 : * Report to pgstat what autovacuum is doing
3123 : *
3124 : * We send a SQL string corresponding to what the user would see if the
3125 : * equivalent command was to be issued manually.
3126 : *
3127 : * Note we assume that we are going to report the next command as soon as we're
3128 : * done with the current one, and exit right after the last one, so we don't
3129 : * bother to report "<IDLE>" or some such.
3130 : */
3131 : static void
3132 93822 : autovac_report_activity(autovac_table *tab)
3133 : {
3134 : #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56)
3135 : char activity[MAX_AUTOVAC_ACTIV_LEN];
3136 : int len;
3137 :
3138 : /* Report the command and possible options */
3139 93822 : if (tab->at_params.options & VACOPT_VACUUM)
3140 93642 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3141 : "autovacuum: VACUUM%s",
3142 93642 : tab->at_params.options & VACOPT_ANALYZE ? " ANALYZE" : "");
3143 : else
3144 180 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3145 : "autovacuum: ANALYZE");
3146 :
3147 : /*
3148 : * Report the qualified name of the relation.
3149 : */
3150 93822 : len = strlen(activity);
3151 :
3152 93822 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3153 : " %s.%s%s", tab->at_nspname, tab->at_relname,
3154 93822 : tab->at_params.is_wraparound ? " (to prevent wraparound)" : "");
3155 :
3156 : /* Set statement_timestamp() to current time for pg_stat_activity */
3157 93822 : SetCurrentStatementStartTimestamp();
3158 :
3159 93822 : pgstat_report_activity(STATE_RUNNING, activity);
3160 93822 : }
3161 :
3162 : /*
3163 : * autovac_report_workitem
3164 : * Report to pgstat that autovacuum is processing a work item
3165 : */
3166 : static void
3167 6 : autovac_report_workitem(AutoVacuumWorkItem *workitem,
3168 : const char *nspname, const char *relname)
3169 : {
3170 : char activity[MAX_AUTOVAC_ACTIV_LEN + 12 + 2];
3171 : char blk[12 + 2];
3172 : int len;
3173 :
3174 6 : switch (workitem->avw_type)
3175 : {
3176 6 : case AVW_BRINSummarizeRange:
3177 6 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3178 : "autovacuum: BRIN summarize");
3179 6 : break;
3180 : }
3181 :
3182 : /*
3183 : * Report the qualified name of the relation, and the block number if any
3184 : */
3185 6 : len = strlen(activity);
3186 :
3187 6 : if (BlockNumberIsValid(workitem->avw_blockNumber))
3188 6 : snprintf(blk, sizeof(blk), " %u", workitem->avw_blockNumber);
3189 : else
3190 0 : blk[0] = '\0';
3191 :
3192 6 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3193 : " %s.%s%s", nspname, relname, blk);
3194 :
3195 : /* Set statement_timestamp() to current time for pg_stat_activity */
3196 6 : SetCurrentStatementStartTimestamp();
3197 :
3198 6 : pgstat_report_activity(STATE_RUNNING, activity);
3199 6 : }
3200 :
3201 : /*
3202 : * AutoVacuumingActive
3203 : * Check GUC vars and report whether the autovacuum process should be
3204 : * running.
3205 : */
3206 : bool
3207 353462 : AutoVacuumingActive(void)
3208 : {
3209 353462 : if (!autovacuum_start_daemon || !pgstat_track_counts)
3210 6286 : return false;
3211 347176 : return true;
3212 : }
3213 :
3214 : /*
3215 : * Request one work item to the next autovacuum run processing our database.
3216 : * Return false if the request can't be recorded.
3217 : */
3218 : bool
3219 6 : AutoVacuumRequestWork(AutoVacuumWorkItemType type, Oid relationId,
3220 : BlockNumber blkno)
3221 : {
3222 : int i;
3223 6 : bool result = false;
3224 :
3225 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
3226 :
3227 : /*
3228 : * Locate an unused work item and fill it with the given data.
3229 : */
3230 12 : for (i = 0; i < NUM_WORKITEMS; i++)
3231 : {
3232 12 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
3233 :
3234 12 : if (workitem->avw_used)
3235 6 : continue;
3236 :
3237 6 : workitem->avw_used = true;
3238 6 : workitem->avw_active = false;
3239 6 : workitem->avw_type = type;
3240 6 : workitem->avw_database = MyDatabaseId;
3241 6 : workitem->avw_relation = relationId;
3242 6 : workitem->avw_blockNumber = blkno;
3243 6 : result = true;
3244 :
3245 : /* done */
3246 6 : break;
3247 : }
3248 :
3249 6 : LWLockRelease(AutovacuumLock);
3250 :
3251 6 : return result;
3252 : }
3253 :
3254 : /*
3255 : * autovac_init
3256 : * This is called at postmaster initialization.
3257 : *
3258 : * All we do here is annoy the user if he got it wrong.
3259 : */
3260 : void
3261 1530 : autovac_init(void)
3262 : {
3263 1530 : if (!autovacuum_start_daemon)
3264 186 : return;
3265 1344 : else if (!pgstat_track_counts)
3266 0 : ereport(WARNING,
3267 : (errmsg("autovacuum not started because of misconfiguration"),
3268 : errhint("Enable the \"track_counts\" option.")));
3269 : else
3270 1344 : check_av_worker_gucs();
3271 : }
3272 :
3273 : /*
3274 : * AutoVacuumShmemSize
3275 : * Compute space needed for autovacuum-related shared memory
3276 : */
3277 : Size
3278 5484 : AutoVacuumShmemSize(void)
3279 : {
3280 : Size size;
3281 :
3282 : /*
3283 : * Need the fixed struct and the array of WorkerInfoData.
3284 : */
3285 5484 : size = sizeof(AutoVacuumShmemStruct);
3286 5484 : size = MAXALIGN(size);
3287 5484 : size = add_size(size, mul_size(autovacuum_worker_slots,
3288 : sizeof(WorkerInfoData)));
3289 5484 : return size;
3290 : }
3291 :
3292 : /*
3293 : * AutoVacuumShmemInit
3294 : * Allocate and initialize autovacuum-related shared memory
3295 : */
3296 : void
3297 1918 : AutoVacuumShmemInit(void)
3298 : {
3299 : bool found;
3300 :
3301 1918 : AutoVacuumShmem = (AutoVacuumShmemStruct *)
3302 1918 : ShmemInitStruct("AutoVacuum Data",
3303 : AutoVacuumShmemSize(),
3304 : &found);
3305 :
3306 1918 : if (!IsUnderPostmaster)
3307 : {
3308 : WorkerInfo worker;
3309 : int i;
3310 :
3311 : Assert(!found);
3312 :
3313 1918 : AutoVacuumShmem->av_launcherpid = 0;
3314 1918 : dclist_init(&AutoVacuumShmem->av_freeWorkers);
3315 1918 : dlist_init(&AutoVacuumShmem->av_runningWorkers);
3316 1918 : AutoVacuumShmem->av_startingWorker = NULL;
3317 1918 : memset(AutoVacuumShmem->av_workItems, 0,
3318 : sizeof(AutoVacuumWorkItem) * NUM_WORKITEMS);
3319 :
3320 1918 : worker = (WorkerInfo) ((char *) AutoVacuumShmem +
3321 : MAXALIGN(sizeof(AutoVacuumShmemStruct)));
3322 :
3323 : /* initialize the WorkerInfo free list */
3324 22856 : for (i = 0; i < autovacuum_worker_slots; i++)
3325 : {
3326 20938 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
3327 20938 : &worker[i].wi_links);
3328 20938 : pg_atomic_init_flag(&worker[i].wi_dobalance);
3329 : }
3330 :
3331 1918 : pg_atomic_init_u32(&AutoVacuumShmem->av_nworkersForBalance, 0);
3332 :
3333 : }
3334 : else
3335 : Assert(found);
3336 1918 : }
3337 :
3338 : /*
3339 : * GUC check_hook for autovacuum_work_mem
3340 : */
3341 : bool
3342 1986 : check_autovacuum_work_mem(int *newval, void **extra, GucSource source)
3343 : {
3344 : /*
3345 : * -1 indicates fallback.
3346 : *
3347 : * If we haven't yet changed the boot_val default of -1, just let it be.
3348 : * Autovacuum will look to maintenance_work_mem instead.
3349 : */
3350 1986 : if (*newval == -1)
3351 1982 : return true;
3352 :
3353 : /*
3354 : * We clamp manually-set values to at least 64kB. Since
3355 : * maintenance_work_mem is always set to at least this value, do the same
3356 : * here.
3357 : */
3358 4 : if (*newval < 64)
3359 4 : *newval = 64;
3360 :
3361 4 : return true;
3362 : }
3363 :
3364 : /*
3365 : * Returns whether there is a free autovacuum worker slot available.
3366 : */
3367 : static bool
3368 21706 : av_worker_available(void)
3369 : {
3370 : int free_slots;
3371 : int reserved_slots;
3372 :
3373 21706 : free_slots = dclist_count(&AutoVacuumShmem->av_freeWorkers);
3374 :
3375 21706 : reserved_slots = autovacuum_worker_slots - autovacuum_max_workers;
3376 21706 : reserved_slots = Max(0, reserved_slots);
3377 :
3378 21706 : return free_slots > reserved_slots;
3379 : }
3380 :
3381 : /*
3382 : * Emits a WARNING if autovacuum_worker_slots < autovacuum_max_workers.
3383 : */
3384 : static void
3385 1344 : check_av_worker_gucs(void)
3386 : {
3387 1344 : if (autovacuum_worker_slots < autovacuum_max_workers)
3388 0 : ereport(WARNING,
3389 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3390 : errmsg("\"autovacuum_max_workers\" (%d) should be less than or equal to \"autovacuum_worker_slots\" (%d)",
3391 : autovacuum_max_workers, autovacuum_worker_slots),
3392 : errdetail("The server will only start up to \"autovacuum_worker_slots\" (%d) autovacuum workers at a given time.",
3393 : autovacuum_worker_slots)));
3394 1344 : }
|