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) 1996-2026, PostgreSQL Global Development Group
55 : * Portions Copyright (c) 1994, Regents of the University of California
56 : *
57 : *
58 : * IDENTIFICATION
59 : * src/backend/postmaster/autovacuum.c
60 : *
61 : *-------------------------------------------------------------------------
62 : */
63 : #include "postgres.h"
64 :
65 : #include <signal.h>
66 : #include <sys/time.h>
67 : #include <unistd.h>
68 :
69 : #include "access/heapam.h"
70 : #include "access/htup_details.h"
71 : #include "access/multixact.h"
72 : #include "access/reloptions.h"
73 : #include "access/tableam.h"
74 : #include "access/transam.h"
75 : #include "access/xact.h"
76 : #include "catalog/dependency.h"
77 : #include "catalog/namespace.h"
78 : #include "catalog/pg_database.h"
79 : #include "catalog/pg_namespace.h"
80 : #include "commands/vacuum.h"
81 : #include "common/int.h"
82 : #include "lib/ilist.h"
83 : #include "libpq/pqsignal.h"
84 : #include "miscadmin.h"
85 : #include "nodes/makefuncs.h"
86 : #include "pgstat.h"
87 : #include "postmaster/autovacuum.h"
88 : #include "postmaster/interrupt.h"
89 : #include "postmaster/postmaster.h"
90 : #include "storage/aio_subsys.h"
91 : #include "storage/bufmgr.h"
92 : #include "storage/ipc.h"
93 : #include "storage/latch.h"
94 : #include "storage/lmgr.h"
95 : #include "storage/pmsignal.h"
96 : #include "storage/proc.h"
97 : #include "storage/procsignal.h"
98 : #include "storage/smgr.h"
99 : #include "tcop/tcopprot.h"
100 : #include "utils/fmgroids.h"
101 : #include "utils/fmgrprotos.h"
102 : #include "utils/guc_hooks.h"
103 : #include "utils/injection_point.h"
104 : #include "utils/lsyscache.h"
105 : #include "utils/memutils.h"
106 : #include "utils/ps_status.h"
107 : #include "utils/rel.h"
108 : #include "utils/snapmgr.h"
109 : #include "utils/syscache.h"
110 : #include "utils/timeout.h"
111 : #include "utils/timestamp.h"
112 :
113 :
114 : /*
115 : * GUC parameters
116 : */
117 : bool autovacuum_start_daemon = false;
118 : int autovacuum_worker_slots;
119 : int autovacuum_max_workers;
120 : int autovacuum_work_mem = -1;
121 : int autovacuum_naptime;
122 : int autovacuum_vac_thresh;
123 : int autovacuum_vac_max_thresh;
124 : double autovacuum_vac_scale;
125 : int autovacuum_vac_ins_thresh;
126 : double autovacuum_vac_ins_scale;
127 : int autovacuum_anl_thresh;
128 : double autovacuum_anl_scale;
129 : int autovacuum_freeze_max_age;
130 : int autovacuum_multixact_freeze_max_age;
131 :
132 : double autovacuum_vac_cost_delay;
133 : int autovacuum_vac_cost_limit;
134 :
135 : int Log_autovacuum_min_duration = 600000;
136 : int Log_autoanalyze_min_duration = 600000;
137 :
138 : /* the minimum allowed time between two awakenings of the launcher */
139 : #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */
140 : #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */
141 :
142 : /*
143 : * Variables to save the cost-related storage parameters for the current
144 : * relation being vacuumed by this autovacuum worker. Using these, we can
145 : * ensure we don't overwrite the values of vacuum_cost_delay and
146 : * vacuum_cost_limit after reloading the configuration file. They are
147 : * initialized to "invalid" values to indicate that no cost-related storage
148 : * parameters were specified and will be set in do_autovacuum() after checking
149 : * the storage parameters in table_recheck_autovac().
150 : */
151 : static double av_storage_param_cost_delay = -1;
152 : static int av_storage_param_cost_limit = -1;
153 :
154 : /* Flags set by signal handlers */
155 : static volatile sig_atomic_t got_SIGUSR2 = false;
156 :
157 : /* Comparison points for determining whether freeze_max_age is exceeded */
158 : static TransactionId recentXid;
159 : static MultiXactId recentMulti;
160 :
161 : /* Default freeze ages to use for autovacuum (varies by database) */
162 : static int default_freeze_min_age;
163 : static int default_freeze_table_age;
164 : static int default_multixact_freeze_min_age;
165 : static int default_multixact_freeze_table_age;
166 :
167 : /* Memory context for long-lived data */
168 : static MemoryContext AutovacMemCxt;
169 :
170 : /* struct to keep track of databases in launcher */
171 : typedef struct avl_dbase
172 : {
173 : Oid adl_datid; /* hash key -- must be first */
174 : TimestampTz adl_next_worker;
175 : int adl_score;
176 : dlist_node adl_node;
177 : } avl_dbase;
178 :
179 : /* struct to keep track of databases in worker */
180 : typedef struct avw_dbase
181 : {
182 : Oid adw_datid;
183 : char *adw_name;
184 : TransactionId adw_frozenxid;
185 : MultiXactId adw_minmulti;
186 : PgStat_StatDBEntry *adw_entry;
187 : } avw_dbase;
188 :
189 : /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */
190 : typedef struct av_relation
191 : {
192 : Oid ar_toastrelid; /* hash key - must be first */
193 : Oid ar_relid;
194 : bool ar_hasrelopts;
195 : AutoVacOpts ar_reloptions; /* copy of AutoVacOpts from the main table's
196 : * reloptions, or NULL if none */
197 : } av_relation;
198 :
199 : /* struct to keep track of tables to vacuum and/or analyze, after rechecking */
200 : typedef struct autovac_table
201 : {
202 : Oid at_relid;
203 : VacuumParams at_params;
204 : double at_storage_param_vac_cost_delay;
205 : int at_storage_param_vac_cost_limit;
206 : bool at_dobalance;
207 : bool at_sharedrel;
208 : char *at_relname;
209 : char *at_nspname;
210 : char *at_datname;
211 : } autovac_table;
212 :
213 : /*-------------
214 : * This struct holds information about a single worker's whereabouts. We keep
215 : * an array of these in shared memory, sized according to
216 : * autovacuum_worker_slots.
217 : *
218 : * wi_links entry into free list or running list
219 : * wi_dboid OID of the database this worker is supposed to work on
220 : * wi_tableoid OID of the table currently being vacuumed, if any
221 : * wi_sharedrel flag indicating whether table is marked relisshared
222 : * wi_proc pointer to PGPROC of the running worker, NULL if not started
223 : * wi_launchtime Time at which this worker was launched
224 : * wi_dobalance Whether this worker should be included in balance calculations
225 : *
226 : * All fields are protected by AutovacuumLock, except for wi_tableoid and
227 : * wi_sharedrel which are protected by AutovacuumScheduleLock (note these
228 : * two fields are read-only for everyone except that worker itself).
229 : *-------------
230 : */
231 : typedef struct WorkerInfoData
232 : {
233 : dlist_node wi_links;
234 : Oid wi_dboid;
235 : Oid wi_tableoid;
236 : PGPROC *wi_proc;
237 : TimestampTz wi_launchtime;
238 : pg_atomic_flag wi_dobalance;
239 : bool wi_sharedrel;
240 : } WorkerInfoData;
241 :
242 : typedef struct WorkerInfoData *WorkerInfo;
243 :
244 : /*
245 : * Possible signals received by the launcher from remote processes. These are
246 : * stored atomically in shared memory so that other processes can set them
247 : * without locking.
248 : */
249 : typedef enum
250 : {
251 : AutoVacForkFailed, /* failed trying to start a worker */
252 : AutoVacRebalance, /* rebalance the cost limits */
253 : } AutoVacuumSignal;
254 :
255 : #define AutoVacNumSignals (AutoVacRebalance + 1)
256 :
257 : /*
258 : * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This
259 : * list is mostly protected by AutovacuumLock, except that if an item is
260 : * marked 'active' other processes must not modify the work-identifying
261 : * members.
262 : */
263 : typedef struct AutoVacuumWorkItem
264 : {
265 : AutoVacuumWorkItemType avw_type;
266 : bool avw_used; /* below data is valid */
267 : bool avw_active; /* being processed */
268 : Oid avw_database;
269 : Oid avw_relation;
270 : BlockNumber avw_blockNumber;
271 : } AutoVacuumWorkItem;
272 :
273 : #define NUM_WORKITEMS 256
274 :
275 : /*-------------
276 : * The main autovacuum shmem struct. On shared memory we store this main
277 : * struct and the array of WorkerInfo structs. This struct keeps:
278 : *
279 : * av_signal set by other processes to indicate various conditions
280 : * av_launcherpid the PID of the autovacuum launcher
281 : * av_freeWorkers the WorkerInfo freelist
282 : * av_runningWorkers the WorkerInfo non-free queue
283 : * av_startingWorker pointer to WorkerInfo currently being started (cleared by
284 : * the worker itself as soon as it's up and running)
285 : * av_workItems work item array
286 : * av_nworkersForBalance the number of autovacuum workers to use when
287 : * calculating the per worker cost limit
288 : *
289 : * This struct is protected by AutovacuumLock, except for av_signal and parts
290 : * of the worker list (see above).
291 : *-------------
292 : */
293 : typedef struct
294 : {
295 : sig_atomic_t av_signal[AutoVacNumSignals];
296 : pid_t av_launcherpid;
297 : dclist_head av_freeWorkers;
298 : dlist_head av_runningWorkers;
299 : WorkerInfo av_startingWorker;
300 : AutoVacuumWorkItem av_workItems[NUM_WORKITEMS];
301 : pg_atomic_uint32 av_nworkersForBalance;
302 : } AutoVacuumShmemStruct;
303 :
304 : static AutoVacuumShmemStruct *AutoVacuumShmem;
305 :
306 : /*
307 : * the database list (of avl_dbase elements) in the launcher, and the context
308 : * that contains it
309 : */
310 : static dlist_head DatabaseList = DLIST_STATIC_INIT(DatabaseList);
311 : static MemoryContext DatabaseListCxt = NULL;
312 :
313 : /*
314 : * Dummy pointer to persuade Valgrind that we've not leaked the array of
315 : * avl_dbase structs. Make it global to ensure the compiler doesn't
316 : * optimize it away.
317 : */
318 : #ifdef USE_VALGRIND
319 : extern avl_dbase *avl_dbase_array;
320 : avl_dbase *avl_dbase_array;
321 : #endif
322 :
323 : /* Pointer to my own WorkerInfo, valid on each worker */
324 : static WorkerInfo MyWorkerInfo = NULL;
325 :
326 : static Oid do_start_worker(void);
327 : static void ProcessAutoVacLauncherInterrupts(void);
328 : pg_noreturn static void AutoVacLauncherShutdown(void);
329 : static void launcher_determine_sleep(bool canlaunch, bool recursing,
330 : struct timeval *nap);
331 : static void launch_worker(TimestampTz now);
332 : static List *get_database_list(void);
333 : static void rebuild_database_list(Oid newdb);
334 : static int db_comparator(const void *a, const void *b);
335 : static void autovac_recalculate_workers_for_balance(void);
336 :
337 : static void do_autovacuum(void);
338 : static void FreeWorkerInfo(int code, Datum arg);
339 :
340 : static autovac_table *table_recheck_autovac(Oid relid, HTAB *table_toast_map,
341 : TupleDesc pg_class_desc,
342 : int effective_multixact_freeze_max_age);
343 : static void recheck_relation_needs_vacanalyze(Oid relid, AutoVacOpts *avopts,
344 : Form_pg_class classForm,
345 : int effective_multixact_freeze_max_age,
346 : bool *dovacuum, bool *doanalyze, bool *wraparound);
347 : static void relation_needs_vacanalyze(Oid relid, AutoVacOpts *relopts,
348 : Form_pg_class classForm,
349 : PgStat_StatTabEntry *tabentry,
350 : int effective_multixact_freeze_max_age,
351 : bool *dovacuum, bool *doanalyze, bool *wraparound);
352 :
353 : static void autovacuum_do_vac_analyze(autovac_table *tab,
354 : BufferAccessStrategy bstrategy);
355 : static AutoVacOpts *extract_autovac_opts(HeapTuple tup,
356 : TupleDesc pg_class_desc);
357 : static void perform_work_item(AutoVacuumWorkItem *workitem);
358 : static void autovac_report_activity(autovac_table *tab);
359 : static void autovac_report_workitem(AutoVacuumWorkItem *workitem,
360 : const char *nspname, const char *relname);
361 : static void avl_sigusr2_handler(SIGNAL_ARGS);
362 : static bool av_worker_available(void);
363 : static void check_av_worker_gucs(void);
364 :
365 :
366 :
367 : /********************************************************************
368 : * AUTOVACUUM LAUNCHER CODE
369 : ********************************************************************/
370 :
371 : /*
372 : * Main entry point for the autovacuum launcher process.
373 : */
374 : void
375 842 : AutoVacLauncherMain(const void *startup_data, size_t startup_data_len)
376 : {
377 : sigjmp_buf local_sigjmp_buf;
378 :
379 : Assert(startup_data_len == 0);
380 :
381 : /* Release postmaster's working memory context */
382 842 : if (PostmasterContext)
383 : {
384 842 : MemoryContextDelete(PostmasterContext);
385 842 : PostmasterContext = NULL;
386 : }
387 :
388 842 : MyBackendType = B_AUTOVAC_LAUNCHER;
389 842 : init_ps_display(NULL);
390 :
391 842 : ereport(DEBUG1,
392 : (errmsg_internal("autovacuum launcher started")));
393 :
394 842 : if (PostAuthDelay)
395 0 : pg_usleep(PostAuthDelay * 1000000L);
396 :
397 : Assert(GetProcessingMode() == InitProcessing);
398 :
399 : /*
400 : * Set up signal handlers. We operate on databases much like a regular
401 : * backend, so we use the same signal handling. See equivalent code in
402 : * tcop/postgres.c.
403 : */
404 842 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
405 842 : pqsignal(SIGINT, StatementCancelHandler);
406 842 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
407 : /* SIGQUIT handler was already set up by InitPostmasterChild */
408 :
409 842 : InitializeTimeouts(); /* establishes SIGALRM handler */
410 :
411 842 : pqsignal(SIGPIPE, SIG_IGN);
412 842 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
413 842 : pqsignal(SIGUSR2, avl_sigusr2_handler);
414 842 : pqsignal(SIGFPE, FloatExceptionHandler);
415 842 : pqsignal(SIGCHLD, SIG_DFL);
416 :
417 : /*
418 : * Create a per-backend PGPROC struct in shared memory. We must do this
419 : * before we can use LWLocks or access any shared memory.
420 : */
421 842 : InitProcess();
422 :
423 : /* Early initialization */
424 842 : BaseInit();
425 :
426 842 : InitPostgres(NULL, InvalidOid, NULL, InvalidOid, 0, NULL);
427 :
428 842 : SetProcessingMode(NormalProcessing);
429 :
430 : /*
431 : * Create a memory context that we will do all our work in. We do this so
432 : * that we can reset the context during error recovery and thereby avoid
433 : * possible memory leaks.
434 : */
435 842 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
436 : "Autovacuum Launcher",
437 : ALLOCSET_DEFAULT_SIZES);
438 842 : MemoryContextSwitchTo(AutovacMemCxt);
439 :
440 : /*
441 : * If an exception is encountered, processing resumes here.
442 : *
443 : * This code is a stripped down version of PostgresMain error recovery.
444 : *
445 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
446 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
447 : * signals other than SIGQUIT will be blocked until we complete error
448 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
449 : * call redundant, but it is not since InterruptPending might be set
450 : * already.
451 : */
452 842 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
453 : {
454 : /* since not using PG_TRY, must reset error stack by hand */
455 0 : error_context_stack = NULL;
456 :
457 : /* Prevents interrupts while cleaning up */
458 0 : HOLD_INTERRUPTS();
459 :
460 : /* Forget any pending QueryCancel or timeout request */
461 0 : disable_all_timeouts(false);
462 0 : QueryCancelPending = false; /* second to avoid race condition */
463 :
464 : /* Report the error to the server log */
465 0 : EmitErrorReport();
466 :
467 : /* Abort the current transaction in order to recover */
468 0 : AbortCurrentTransaction();
469 :
470 : /*
471 : * Release any other resources, for the case where we were not in a
472 : * transaction.
473 : */
474 0 : LWLockReleaseAll();
475 0 : pgstat_report_wait_end();
476 0 : pgaio_error_cleanup();
477 0 : UnlockBuffers();
478 : /* this is probably dead code, but let's be safe: */
479 0 : if (AuxProcessResourceOwner)
480 0 : ReleaseAuxProcessResources(false);
481 0 : AtEOXact_Buffers(false);
482 0 : AtEOXact_SMgr();
483 0 : AtEOXact_Files(false);
484 0 : AtEOXact_HashTables(false);
485 :
486 : /*
487 : * Now return to normal top-level context and clear ErrorContext for
488 : * next time.
489 : */
490 0 : MemoryContextSwitchTo(AutovacMemCxt);
491 0 : FlushErrorState();
492 :
493 : /* Flush any leaked data in the top-level context */
494 0 : MemoryContextReset(AutovacMemCxt);
495 :
496 : /* don't leave dangling pointers to freed memory */
497 0 : DatabaseListCxt = NULL;
498 0 : dlist_init(&DatabaseList);
499 :
500 : /* Now we can allow interrupts again */
501 0 : RESUME_INTERRUPTS();
502 :
503 : /* if in shutdown mode, no need for anything further; just go away */
504 0 : if (ShutdownRequestPending)
505 0 : AutoVacLauncherShutdown();
506 :
507 : /*
508 : * Sleep at least 1 second after any error. We don't want to be
509 : * filling the error logs as fast as we can.
510 : */
511 0 : pg_usleep(1000000L);
512 : }
513 :
514 : /* We can now handle ereport(ERROR) */
515 842 : PG_exception_stack = &local_sigjmp_buf;
516 :
517 : /* must unblock signals before calling rebuild_database_list */
518 842 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
519 :
520 : /*
521 : * Set always-secure search path. Launcher doesn't connect to a database,
522 : * so this has no effect.
523 : */
524 842 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
525 :
526 : /*
527 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
528 : * in postgresql.conf. We don't really want such a dangerous option being
529 : * applied non-interactively.
530 : */
531 842 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
532 :
533 : /*
534 : * Force settable timeouts off to avoid letting these settings prevent
535 : * regular maintenance from being executed.
536 : */
537 842 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
538 842 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
539 842 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
540 842 : SetConfigOption("idle_in_transaction_session_timeout", "0",
541 : PGC_SUSET, PGC_S_OVERRIDE);
542 :
543 : /*
544 : * Force default_transaction_isolation to READ COMMITTED. We don't want
545 : * to pay the overhead of serializable mode, nor add any risk of causing
546 : * deadlocks or delaying other transactions.
547 : */
548 842 : SetConfigOption("default_transaction_isolation", "read committed",
549 : PGC_SUSET, PGC_S_OVERRIDE);
550 :
551 : /*
552 : * Even when system is configured to use a different fetch consistency,
553 : * for autovac we always want fresh stats.
554 : */
555 842 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
556 :
557 : /*
558 : * In emergency mode, just start a worker (unless shutdown was requested)
559 : * and go away.
560 : */
561 842 : if (!AutoVacuumingActive())
562 : {
563 0 : if (!ShutdownRequestPending)
564 0 : do_start_worker();
565 0 : proc_exit(0); /* done */
566 : }
567 :
568 842 : AutoVacuumShmem->av_launcherpid = MyProcPid;
569 :
570 : /*
571 : * Create the initial database list. The invariant we want this list to
572 : * keep is that it's ordered by decreasing next_worker. As soon as an
573 : * entry is updated to a higher time, it will be moved to the front (which
574 : * is correct because the only operation is to add autovacuum_naptime to
575 : * the entry, and time always increases).
576 : */
577 842 : rebuild_database_list(InvalidOid);
578 :
579 : /* loop until shutdown request */
580 10820 : while (!ShutdownRequestPending)
581 : {
582 : struct timeval nap;
583 10818 : TimestampTz current_time = 0;
584 : bool can_launch;
585 :
586 : /*
587 : * This loop is a bit different from the normal use of WaitLatch,
588 : * because we'd like to sleep before the first launch of a child
589 : * process. So it's WaitLatch, then ResetLatch, then check for
590 : * wakening conditions.
591 : */
592 :
593 10818 : launcher_determine_sleep(av_worker_available(), false, &nap);
594 :
595 : /*
596 : * Wait until naptime expires or we get some type of signal (all the
597 : * signal handlers will wake us by calling SetLatch).
598 : */
599 10818 : (void) WaitLatch(MyLatch,
600 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
601 10818 : (nap.tv_sec * 1000L) + (nap.tv_usec / 1000L),
602 : WAIT_EVENT_AUTOVACUUM_MAIN);
603 :
604 10812 : ResetLatch(MyLatch);
605 :
606 10812 : ProcessAutoVacLauncherInterrupts();
607 :
608 : /*
609 : * a worker finished, or postmaster signaled failure to start a worker
610 : */
611 9978 : if (got_SIGUSR2)
612 : {
613 5838 : got_SIGUSR2 = false;
614 :
615 : /* rebalance cost limits, if needed */
616 5838 : if (AutoVacuumShmem->av_signal[AutoVacRebalance])
617 : {
618 2770 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
619 2770 : AutoVacuumShmem->av_signal[AutoVacRebalance] = false;
620 2770 : autovac_recalculate_workers_for_balance();
621 2770 : LWLockRelease(AutovacuumLock);
622 : }
623 :
624 5838 : if (AutoVacuumShmem->av_signal[AutoVacForkFailed])
625 : {
626 : /*
627 : * If the postmaster failed to start a new worker, we sleep
628 : * for a little while and resend the signal. The new worker's
629 : * state is still in memory, so this is sufficient. After
630 : * that, we restart the main loop.
631 : *
632 : * XXX should we put a limit to the number of times we retry?
633 : * I don't think it makes much sense, because a future start
634 : * of a worker will continue to fail in the same way.
635 : */
636 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = false;
637 0 : pg_usleep(1000000L); /* 1s */
638 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
639 3148 : continue;
640 : }
641 : }
642 :
643 : /*
644 : * There are some conditions that we need to check before trying to
645 : * start a worker. First, we need to make sure that there is a worker
646 : * slot available. Second, we need to make sure that no other worker
647 : * failed while starting up.
648 : */
649 :
650 9978 : current_time = GetCurrentTimestamp();
651 9978 : LWLockAcquire(AutovacuumLock, LW_SHARED);
652 :
653 9978 : can_launch = av_worker_available();
654 :
655 9978 : if (AutoVacuumShmem->av_startingWorker != NULL)
656 : {
657 : int waittime;
658 90 : WorkerInfo worker = AutoVacuumShmem->av_startingWorker;
659 :
660 : /*
661 : * We can't launch another worker when another one is still
662 : * starting up (or failed while doing so), so just sleep for a bit
663 : * more; that worker will wake us up again as soon as it's ready.
664 : * We will only wait autovacuum_naptime seconds (up to a maximum
665 : * of 60 seconds) for this to happen however. Note that failure
666 : * to connect to a particular database is not a problem here,
667 : * because the worker removes itself from the startingWorker
668 : * pointer before trying to connect. Problems detected by the
669 : * postmaster (like fork() failure) are also reported and handled
670 : * differently. The only problems that may cause this code to
671 : * fire are errors in the earlier sections of AutoVacWorkerMain,
672 : * before the worker removes the WorkerInfo from the
673 : * startingWorker pointer.
674 : */
675 90 : waittime = Min(autovacuum_naptime, 60) * 1000;
676 90 : if (TimestampDifferenceExceeds(worker->wi_launchtime, current_time,
677 : waittime))
678 : {
679 0 : LWLockRelease(AutovacuumLock);
680 0 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
681 :
682 : /*
683 : * No other process can put a worker in starting mode, so if
684 : * startingWorker is still INVALID after exchanging our lock,
685 : * we assume it's the same one we saw above (so we don't
686 : * recheck the launch time).
687 : */
688 0 : if (AutoVacuumShmem->av_startingWorker != NULL)
689 : {
690 0 : worker = AutoVacuumShmem->av_startingWorker;
691 0 : worker->wi_dboid = InvalidOid;
692 0 : worker->wi_tableoid = InvalidOid;
693 0 : worker->wi_sharedrel = false;
694 0 : worker->wi_proc = NULL;
695 0 : worker->wi_launchtime = 0;
696 0 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
697 : &worker->wi_links);
698 0 : AutoVacuumShmem->av_startingWorker = NULL;
699 0 : ereport(WARNING,
700 : errmsg("autovacuum worker took too long to start; canceled"));
701 : }
702 : }
703 : else
704 90 : can_launch = false;
705 : }
706 9978 : LWLockRelease(AutovacuumLock); /* either shared or exclusive */
707 :
708 : /* if we can't do anything, just go back to sleep */
709 9978 : if (!can_launch)
710 3148 : continue;
711 :
712 : /* We're OK to start a new worker */
713 :
714 6830 : if (dlist_is_empty(&DatabaseList))
715 : {
716 : /*
717 : * Special case when the list is empty: start a worker right away.
718 : * This covers the initial case, when no database is in pgstats
719 : * (thus the list is empty). Note that the constraints in
720 : * launcher_determine_sleep keep us from starting workers too
721 : * quickly (at most once every autovacuum_naptime when the list is
722 : * empty).
723 : */
724 14 : launch_worker(current_time);
725 : }
726 : else
727 : {
728 : /*
729 : * because rebuild_database_list constructs a list with most
730 : * distant adl_next_worker first, we obtain our database from the
731 : * tail of the list.
732 : */
733 : avl_dbase *avdb;
734 :
735 6816 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
736 :
737 : /*
738 : * launch a worker if next_worker is right now or it is in the
739 : * past
740 : */
741 6816 : if (TimestampDifferenceExceeds(avdb->adl_next_worker,
742 : current_time, 0))
743 2928 : launch_worker(current_time);
744 : }
745 : }
746 :
747 2 : AutoVacLauncherShutdown();
748 : }
749 :
750 : /*
751 : * Process any new interrupts.
752 : */
753 : static void
754 10812 : ProcessAutoVacLauncherInterrupts(void)
755 : {
756 : /* the normal shutdown case */
757 10812 : if (ShutdownRequestPending)
758 834 : AutoVacLauncherShutdown();
759 :
760 9978 : if (ConfigReloadPending)
761 : {
762 76 : int autovacuum_max_workers_prev = autovacuum_max_workers;
763 :
764 76 : ConfigReloadPending = false;
765 76 : ProcessConfigFile(PGC_SIGHUP);
766 :
767 : /* shutdown requested in config file? */
768 76 : if (!AutoVacuumingActive())
769 0 : AutoVacLauncherShutdown();
770 :
771 : /*
772 : * If autovacuum_max_workers changed, emit a WARNING if
773 : * autovacuum_worker_slots < autovacuum_max_workers. If it didn't
774 : * change, skip this to avoid too many repeated log messages.
775 : */
776 76 : if (autovacuum_max_workers_prev != autovacuum_max_workers)
777 0 : check_av_worker_gucs();
778 :
779 : /* rebuild the list in case the naptime changed */
780 76 : rebuild_database_list(InvalidOid);
781 : }
782 :
783 : /* Process barrier events */
784 9978 : if (ProcSignalBarrierPending)
785 84 : ProcessProcSignalBarrier();
786 :
787 : /* Perform logging of memory contexts of this process */
788 9978 : if (LogMemoryContextPending)
789 0 : ProcessLogMemoryContextInterrupt();
790 :
791 : /* Process sinval catchup interrupts that happened while sleeping */
792 9978 : ProcessCatchupInterrupt();
793 9978 : }
794 :
795 : /*
796 : * Perform a normal exit from the autovac launcher.
797 : */
798 : static void
799 836 : AutoVacLauncherShutdown(void)
800 : {
801 836 : ereport(DEBUG1,
802 : (errmsg_internal("autovacuum launcher shutting down")));
803 836 : AutoVacuumShmem->av_launcherpid = 0;
804 :
805 836 : proc_exit(0); /* done */
806 : }
807 :
808 : /*
809 : * Determine the time to sleep, based on the database list.
810 : *
811 : * The "canlaunch" parameter indicates whether we can start a worker right now,
812 : * for example due to the workers being all busy. If this is false, we will
813 : * cause a long sleep, which will be interrupted when a worker exits.
814 : */
815 : static void
816 11004 : launcher_determine_sleep(bool canlaunch, bool recursing, struct timeval *nap)
817 : {
818 : /*
819 : * We sleep until the next scheduled vacuum. We trust that when the
820 : * database list was built, care was taken so that no entries have times
821 : * in the past; if the first entry has too close a next_worker value, or a
822 : * time in the past, we will sleep a small nominal time.
823 : */
824 11004 : if (!canlaunch)
825 : {
826 5608 : nap->tv_sec = autovacuum_naptime;
827 5608 : nap->tv_usec = 0;
828 : }
829 5396 : else if (!dlist_is_empty(&DatabaseList))
830 : {
831 5344 : TimestampTz current_time = GetCurrentTimestamp();
832 : TimestampTz next_wakeup;
833 : avl_dbase *avdb;
834 : long secs;
835 : int usecs;
836 :
837 5344 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
838 :
839 5344 : next_wakeup = avdb->adl_next_worker;
840 5344 : TimestampDifference(current_time, next_wakeup, &secs, &usecs);
841 :
842 5344 : nap->tv_sec = secs;
843 5344 : nap->tv_usec = usecs;
844 : }
845 : else
846 : {
847 : /* list is empty, sleep for whole autovacuum_naptime seconds */
848 52 : nap->tv_sec = autovacuum_naptime;
849 52 : nap->tv_usec = 0;
850 : }
851 :
852 : /*
853 : * If the result is exactly zero, it means a database had an entry with
854 : * time in the past. Rebuild the list so that the databases are evenly
855 : * distributed again, and recalculate the time to sleep. This can happen
856 : * if there are more tables needing vacuum than workers, and they all take
857 : * longer to vacuum than autovacuum_naptime.
858 : *
859 : * We only recurse once. rebuild_database_list should always return times
860 : * in the future, but it seems best not to trust too much on that.
861 : */
862 11004 : if (nap->tv_sec == 0 && nap->tv_usec == 0 && !recursing)
863 : {
864 186 : rebuild_database_list(InvalidOid);
865 186 : launcher_determine_sleep(canlaunch, true, nap);
866 186 : return;
867 : }
868 :
869 : /* The smallest time we'll allow the launcher to sleep. */
870 10818 : if (nap->tv_sec <= 0 && nap->tv_usec <= MIN_AUTOVAC_SLEEPTIME * 1000)
871 : {
872 436 : nap->tv_sec = 0;
873 436 : nap->tv_usec = MIN_AUTOVAC_SLEEPTIME * 1000;
874 : }
875 :
876 : /*
877 : * If the sleep time is too large, clamp it to an arbitrary maximum (plus
878 : * any fractional seconds, for simplicity). This avoids an essentially
879 : * infinite sleep in strange cases like the system clock going backwards a
880 : * few years.
881 : */
882 10818 : if (nap->tv_sec > MAX_AUTOVAC_SLEEPTIME)
883 20 : nap->tv_sec = MAX_AUTOVAC_SLEEPTIME;
884 : }
885 :
886 : /*
887 : * Build an updated DatabaseList. It must only contain databases that appear
888 : * in pgstats, and must be sorted by next_worker from highest to lowest,
889 : * distributed regularly across the next autovacuum_naptime interval.
890 : *
891 : * Receives the Oid of the database that made this list be generated (we call
892 : * this the "new" database, because when the database was already present on
893 : * the list, we expect that this function is not called at all). The
894 : * preexisting list, if any, will be used to preserve the order of the
895 : * databases in the autovacuum_naptime period. The new database is put at the
896 : * end of the interval. The actual values are not saved, which should not be
897 : * much of a problem.
898 : */
899 : static void
900 1134 : rebuild_database_list(Oid newdb)
901 : {
902 : List *dblist;
903 : ListCell *cell;
904 : MemoryContext newcxt;
905 : MemoryContext oldcxt;
906 : MemoryContext tmpcxt;
907 : HASHCTL hctl;
908 : int score;
909 : int nelems;
910 : HTAB *dbhash;
911 : dlist_iter iter;
912 :
913 1134 : newcxt = AllocSetContextCreate(AutovacMemCxt,
914 : "Autovacuum database list",
915 : ALLOCSET_DEFAULT_SIZES);
916 1134 : tmpcxt = AllocSetContextCreate(newcxt,
917 : "Autovacuum database list (tmp)",
918 : ALLOCSET_DEFAULT_SIZES);
919 1134 : oldcxt = MemoryContextSwitchTo(tmpcxt);
920 :
921 : /*
922 : * Implementing this is not as simple as it sounds, because we need to put
923 : * the new database at the end of the list; next the databases that were
924 : * already on the list, and finally (at the tail of the list) all the
925 : * other databases that are not on the existing list.
926 : *
927 : * To do this, we build an empty hash table of scored databases. We will
928 : * start with the lowest score (zero) for the new database, then
929 : * increasing scores for the databases in the existing list, in order, and
930 : * lastly increasing scores for all databases gotten via
931 : * get_database_list() that are not already on the hash.
932 : *
933 : * Then we will put all the hash elements into an array, sort the array by
934 : * score, and finally put the array elements into the new doubly linked
935 : * list.
936 : */
937 1134 : hctl.keysize = sizeof(Oid);
938 1134 : hctl.entrysize = sizeof(avl_dbase);
939 1134 : hctl.hcxt = tmpcxt;
940 1134 : dbhash = hash_create("autovacuum db hash", 20, &hctl, /* magic number here
941 : * FIXME */
942 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
943 :
944 : /* start by inserting the new database */
945 1134 : score = 0;
946 1134 : if (OidIsValid(newdb))
947 : {
948 : avl_dbase *db;
949 : PgStat_StatDBEntry *entry;
950 :
951 : /* only consider this database if it has a pgstat entry */
952 30 : entry = pgstat_fetch_stat_dbentry(newdb);
953 30 : if (entry != NULL)
954 : {
955 : /* we assume it isn't found because the hash was just created */
956 24 : db = hash_search(dbhash, &newdb, HASH_ENTER, NULL);
957 :
958 : /* hash_search already filled in the key */
959 24 : db->adl_score = score++;
960 : /* next_worker is filled in later */
961 : }
962 : }
963 :
964 : /* Now insert the databases from the existing list */
965 1858 : dlist_foreach(iter, &DatabaseList)
966 : {
967 724 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
968 : avl_dbase *db;
969 : bool found;
970 : PgStat_StatDBEntry *entry;
971 :
972 : /*
973 : * skip databases with no stat entries -- in particular, this gets rid
974 : * of dropped databases
975 : */
976 724 : entry = pgstat_fetch_stat_dbentry(avdb->adl_datid);
977 724 : if (entry == NULL)
978 0 : continue;
979 :
980 724 : db = hash_search(dbhash, &(avdb->adl_datid), HASH_ENTER, &found);
981 :
982 724 : if (!found)
983 : {
984 : /* hash_search already filled in the key */
985 724 : db->adl_score = score++;
986 : /* next_worker is filled in later */
987 : }
988 : }
989 :
990 : /* finally, insert all qualifying databases not previously inserted */
991 1134 : dblist = get_database_list();
992 4894 : foreach(cell, dblist)
993 : {
994 3760 : avw_dbase *avdb = lfirst(cell);
995 : avl_dbase *db;
996 : bool found;
997 : PgStat_StatDBEntry *entry;
998 :
999 : /* only consider databases with a pgstat entry */
1000 3760 : entry = pgstat_fetch_stat_dbentry(avdb->adw_datid);
1001 3760 : if (entry == NULL)
1002 1816 : continue;
1003 :
1004 1944 : db = hash_search(dbhash, &(avdb->adw_datid), HASH_ENTER, &found);
1005 : /* only update the score if the database was not already on the hash */
1006 1944 : if (!found)
1007 : {
1008 : /* hash_search already filled in the key */
1009 1196 : db->adl_score = score++;
1010 : /* next_worker is filled in later */
1011 : }
1012 : }
1013 1134 : nelems = score;
1014 :
1015 : /* from here on, the allocated memory belongs to the new list */
1016 1134 : MemoryContextSwitchTo(newcxt);
1017 1134 : dlist_init(&DatabaseList);
1018 :
1019 1134 : if (nelems > 0)
1020 : {
1021 : TimestampTz current_time;
1022 : int millis_increment;
1023 : avl_dbase *dbary;
1024 : avl_dbase *db;
1025 : HASH_SEQ_STATUS seq;
1026 : int i;
1027 :
1028 : /* put all the hash elements into an array */
1029 1096 : dbary = palloc(nelems * sizeof(avl_dbase));
1030 : /* keep Valgrind quiet */
1031 : #ifdef USE_VALGRIND
1032 : avl_dbase_array = dbary;
1033 : #endif
1034 :
1035 1096 : i = 0;
1036 1096 : hash_seq_init(&seq, dbhash);
1037 3040 : while ((db = hash_seq_search(&seq)) != NULL)
1038 1944 : memcpy(&(dbary[i++]), db, sizeof(avl_dbase));
1039 :
1040 : /* sort the array */
1041 1096 : qsort(dbary, nelems, sizeof(avl_dbase), db_comparator);
1042 :
1043 : /*
1044 : * Determine the time interval between databases in the schedule. If
1045 : * we see that the configured naptime would take us to sleep times
1046 : * lower than our min sleep time (which launcher_determine_sleep is
1047 : * coded not to allow), silently use a larger naptime (but don't touch
1048 : * the GUC variable).
1049 : */
1050 1096 : millis_increment = 1000.0 * autovacuum_naptime / nelems;
1051 1096 : if (millis_increment <= MIN_AUTOVAC_SLEEPTIME)
1052 0 : millis_increment = MIN_AUTOVAC_SLEEPTIME * 1.1;
1053 :
1054 1096 : current_time = GetCurrentTimestamp();
1055 :
1056 : /*
1057 : * move the elements from the array into the dlist, setting the
1058 : * next_worker while walking the array
1059 : */
1060 3040 : for (i = 0; i < nelems; i++)
1061 : {
1062 1944 : db = &(dbary[i]);
1063 :
1064 1944 : current_time = TimestampTzPlusMilliseconds(current_time,
1065 : millis_increment);
1066 1944 : db->adl_next_worker = current_time;
1067 :
1068 : /* later elements should go closer to the head of the list */
1069 1944 : dlist_push_head(&DatabaseList, &db->adl_node);
1070 : }
1071 : }
1072 :
1073 : /* all done, clean up memory */
1074 1134 : if (DatabaseListCxt != NULL)
1075 292 : MemoryContextDelete(DatabaseListCxt);
1076 1134 : MemoryContextDelete(tmpcxt);
1077 1134 : DatabaseListCxt = newcxt;
1078 1134 : MemoryContextSwitchTo(oldcxt);
1079 1134 : }
1080 :
1081 : /* qsort comparator for avl_dbase, using adl_score */
1082 : static int
1083 1216 : db_comparator(const void *a, const void *b)
1084 : {
1085 2432 : return pg_cmp_s32(((const avl_dbase *) a)->adl_score,
1086 1216 : ((const avl_dbase *) b)->adl_score);
1087 : }
1088 :
1089 : /*
1090 : * do_start_worker
1091 : *
1092 : * Bare-bones procedure for starting an autovacuum worker from the launcher.
1093 : * It determines what database to work on, sets up shared memory stuff and
1094 : * signals postmaster to start the worker. It fails gracefully if invoked when
1095 : * autovacuum_workers are already active.
1096 : *
1097 : * Return value is the OID of the database that the worker is going to process,
1098 : * or InvalidOid if no worker was actually started.
1099 : */
1100 : static Oid
1101 2942 : do_start_worker(void)
1102 : {
1103 : List *dblist;
1104 : ListCell *cell;
1105 : TransactionId xidForceLimit;
1106 : MultiXactId multiForceLimit;
1107 : bool for_xid_wrap;
1108 : bool for_multi_wrap;
1109 : avw_dbase *avdb;
1110 : TimestampTz current_time;
1111 2942 : bool skipit = false;
1112 2942 : Oid retval = InvalidOid;
1113 : MemoryContext tmpcxt,
1114 : oldcxt;
1115 :
1116 : /* return quickly when there are no free workers */
1117 2942 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1118 2942 : if (!av_worker_available())
1119 : {
1120 0 : LWLockRelease(AutovacuumLock);
1121 0 : return InvalidOid;
1122 : }
1123 2942 : LWLockRelease(AutovacuumLock);
1124 :
1125 : /*
1126 : * Create and switch to a temporary context to avoid leaking the memory
1127 : * allocated for the database list.
1128 : */
1129 2942 : tmpcxt = AllocSetContextCreate(CurrentMemoryContext,
1130 : "Autovacuum start worker (tmp)",
1131 : ALLOCSET_DEFAULT_SIZES);
1132 2942 : oldcxt = MemoryContextSwitchTo(tmpcxt);
1133 :
1134 : /* Get a list of databases */
1135 2942 : dblist = get_database_list();
1136 :
1137 : /*
1138 : * Determine the oldest datfrozenxid/relfrozenxid that we will allow to
1139 : * pass without forcing a vacuum. (This limit can be tightened for
1140 : * particular tables, but not loosened.)
1141 : */
1142 2942 : recentXid = ReadNextTransactionId();
1143 2942 : xidForceLimit = recentXid - autovacuum_freeze_max_age;
1144 : /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */
1145 : /* this can cause the limit to go backwards by 3, but that's OK */
1146 2942 : if (xidForceLimit < FirstNormalTransactionId)
1147 0 : xidForceLimit -= FirstNormalTransactionId;
1148 :
1149 : /* Also determine the oldest datminmxid we will consider. */
1150 2942 : recentMulti = ReadNextMultiXactId();
1151 2942 : multiForceLimit = recentMulti - MultiXactMemberFreezeThreshold();
1152 2942 : if (multiForceLimit < FirstMultiXactId)
1153 0 : multiForceLimit -= FirstMultiXactId;
1154 :
1155 : /*
1156 : * Choose a database to connect to. We pick the database that was least
1157 : * recently auto-vacuumed, or one that needs vacuuming to prevent Xid
1158 : * wraparound-related data loss. If any db at risk of Xid wraparound is
1159 : * found, we pick the one with oldest datfrozenxid, independently of
1160 : * autovacuum times; similarly we pick the one with the oldest datminmxid
1161 : * if any is in MultiXactId wraparound. Note that those in Xid wraparound
1162 : * danger are given more priority than those in multi wraparound danger.
1163 : *
1164 : * Note that a database with no stats entry is not considered, except for
1165 : * Xid wraparound purposes. The theory is that if no one has ever
1166 : * connected to it since the stats were last initialized, it doesn't need
1167 : * vacuuming.
1168 : *
1169 : * XXX This could be improved if we had more info about whether it needs
1170 : * vacuuming before connecting to it. Perhaps look through the pgstats
1171 : * data for the database's tables? One idea is to keep track of the
1172 : * number of new and dead tuples per database in pgstats. However it
1173 : * isn't clear how to construct a metric that measures that and not cause
1174 : * starvation for less busy databases.
1175 : */
1176 2942 : avdb = NULL;
1177 2942 : for_xid_wrap = false;
1178 2942 : for_multi_wrap = false;
1179 2942 : current_time = GetCurrentTimestamp();
1180 11878 : foreach(cell, dblist)
1181 : {
1182 8936 : avw_dbase *tmp = lfirst(cell);
1183 : dlist_iter iter;
1184 :
1185 : /* Check to see if this one is at risk of wraparound */
1186 8936 : if (TransactionIdPrecedes(tmp->adw_frozenxid, xidForceLimit))
1187 : {
1188 7832 : if (avdb == NULL ||
1189 3072 : TransactionIdPrecedes(tmp->adw_frozenxid,
1190 : avdb->adw_frozenxid))
1191 1802 : avdb = tmp;
1192 4760 : for_xid_wrap = true;
1193 6766 : continue;
1194 : }
1195 4176 : else if (for_xid_wrap)
1196 224 : continue; /* ignore not-at-risk DBs */
1197 3952 : else if (MultiXactIdPrecedes(tmp->adw_minmulti, multiForceLimit))
1198 : {
1199 0 : if (avdb == NULL ||
1200 0 : MultiXactIdPrecedes(tmp->adw_minmulti, avdb->adw_minmulti))
1201 0 : avdb = tmp;
1202 0 : for_multi_wrap = true;
1203 0 : continue;
1204 : }
1205 3952 : else if (for_multi_wrap)
1206 0 : continue; /* ignore not-at-risk DBs */
1207 :
1208 : /* Find pgstat entry if any */
1209 3952 : tmp->adw_entry = pgstat_fetch_stat_dbentry(tmp->adw_datid);
1210 :
1211 : /*
1212 : * Skip a database with no pgstat entry; it means it hasn't seen any
1213 : * activity.
1214 : */
1215 3952 : if (!tmp->adw_entry)
1216 164 : continue;
1217 :
1218 : /*
1219 : * Also, skip a database that appears on the database list as having
1220 : * been processed recently (less than autovacuum_naptime seconds ago).
1221 : * We do this so that we don't select a database which we just
1222 : * selected, but that pgstat hasn't gotten around to updating the last
1223 : * autovacuum time yet.
1224 : */
1225 3788 : skipit = false;
1226 :
1227 7756 : dlist_reverse_foreach(iter, &DatabaseList)
1228 : {
1229 7706 : avl_dbase *dbp = dlist_container(avl_dbase, adl_node, iter.cur);
1230 :
1231 7706 : if (dbp->adl_datid == tmp->adw_datid)
1232 : {
1233 : /*
1234 : * Skip this database if its next_worker value falls between
1235 : * the current time and the current time plus naptime.
1236 : */
1237 3738 : if (!TimestampDifferenceExceeds(dbp->adl_next_worker,
1238 1618 : current_time, 0) &&
1239 1618 : !TimestampDifferenceExceeds(current_time,
1240 : dbp->adl_next_worker,
1241 : autovacuum_naptime * 1000))
1242 1618 : skipit = true;
1243 :
1244 3738 : break;
1245 : }
1246 : }
1247 3788 : if (skipit)
1248 1618 : continue;
1249 :
1250 : /*
1251 : * Remember the db with oldest autovac time. (If we are here, both
1252 : * tmp->entry and db->entry must be non-null.)
1253 : */
1254 2170 : if (avdb == NULL ||
1255 930 : tmp->adw_entry->last_autovac_time < avdb->adw_entry->last_autovac_time)
1256 1576 : avdb = tmp;
1257 : }
1258 :
1259 : /* Found a database -- process it */
1260 2942 : if (avdb != NULL)
1261 : {
1262 : WorkerInfo worker;
1263 : dlist_node *wptr;
1264 :
1265 2928 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1266 :
1267 : /*
1268 : * Get a worker entry from the freelist. We checked above, so there
1269 : * really should be a free slot.
1270 : */
1271 2928 : wptr = dclist_pop_head_node(&AutoVacuumShmem->av_freeWorkers);
1272 :
1273 2928 : worker = dlist_container(WorkerInfoData, wi_links, wptr);
1274 2928 : worker->wi_dboid = avdb->adw_datid;
1275 2928 : worker->wi_proc = NULL;
1276 2928 : worker->wi_launchtime = GetCurrentTimestamp();
1277 :
1278 2928 : AutoVacuumShmem->av_startingWorker = worker;
1279 :
1280 2928 : LWLockRelease(AutovacuumLock);
1281 :
1282 2928 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
1283 :
1284 2928 : retval = avdb->adw_datid;
1285 : }
1286 14 : else if (skipit)
1287 : {
1288 : /*
1289 : * If we skipped all databases on the list, rebuild it, because it
1290 : * probably contains a dropped database.
1291 : */
1292 0 : rebuild_database_list(InvalidOid);
1293 : }
1294 :
1295 2942 : MemoryContextSwitchTo(oldcxt);
1296 2942 : MemoryContextDelete(tmpcxt);
1297 :
1298 2942 : return retval;
1299 : }
1300 :
1301 : /*
1302 : * launch_worker
1303 : *
1304 : * Wrapper for starting a worker from the launcher. Besides actually starting
1305 : * it, update the database list to reflect the next time that another one will
1306 : * need to be started on the selected database. The actual database choice is
1307 : * left to do_start_worker.
1308 : *
1309 : * This routine is also expected to insert an entry into the database list if
1310 : * the selected database was previously absent from the list.
1311 : */
1312 : static void
1313 2942 : launch_worker(TimestampTz now)
1314 : {
1315 : Oid dbid;
1316 : dlist_iter iter;
1317 :
1318 2942 : dbid = do_start_worker();
1319 2942 : if (OidIsValid(dbid))
1320 : {
1321 2928 : bool found = false;
1322 :
1323 : /*
1324 : * Walk the database list and update the corresponding entry. If the
1325 : * database is not on the list, we'll recreate the list.
1326 : */
1327 6198 : dlist_foreach(iter, &DatabaseList)
1328 : {
1329 6168 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1330 :
1331 6168 : if (avdb->adl_datid == dbid)
1332 : {
1333 2898 : found = true;
1334 :
1335 : /*
1336 : * add autovacuum_naptime seconds to the current time, and use
1337 : * that as the new "next_worker" field for this database.
1338 : */
1339 2898 : avdb->adl_next_worker =
1340 2898 : TimestampTzPlusMilliseconds(now, autovacuum_naptime * 1000);
1341 :
1342 2898 : dlist_move_head(&DatabaseList, iter.cur);
1343 2898 : break;
1344 : }
1345 : }
1346 :
1347 : /*
1348 : * If the database was not present in the database list, we rebuild
1349 : * the list. It's possible that the database does not get into the
1350 : * list anyway, for example if it's a database that doesn't have a
1351 : * pgstat entry, but this is not a problem because we don't want to
1352 : * schedule workers regularly into those in any case.
1353 : */
1354 2928 : if (!found)
1355 30 : rebuild_database_list(dbid);
1356 : }
1357 2942 : }
1358 :
1359 : /*
1360 : * Called from postmaster to signal a failure to fork a process to become
1361 : * worker. The postmaster should kill(SIGUSR2) the launcher shortly
1362 : * after calling this function.
1363 : */
1364 : void
1365 0 : AutoVacWorkerFailed(void)
1366 : {
1367 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = true;
1368 0 : }
1369 :
1370 : /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */
1371 : static void
1372 5848 : avl_sigusr2_handler(SIGNAL_ARGS)
1373 : {
1374 5848 : got_SIGUSR2 = true;
1375 5848 : SetLatch(MyLatch);
1376 5848 : }
1377 :
1378 :
1379 : /********************************************************************
1380 : * AUTOVACUUM WORKER CODE
1381 : ********************************************************************/
1382 :
1383 : /*
1384 : * Main entry point for autovacuum worker processes.
1385 : */
1386 : void
1387 2930 : AutoVacWorkerMain(const void *startup_data, size_t startup_data_len)
1388 : {
1389 : sigjmp_buf local_sigjmp_buf;
1390 : Oid dbid;
1391 :
1392 : Assert(startup_data_len == 0);
1393 :
1394 : /* Release postmaster's working memory context */
1395 2930 : if (PostmasterContext)
1396 : {
1397 2930 : MemoryContextDelete(PostmasterContext);
1398 2930 : PostmasterContext = NULL;
1399 : }
1400 :
1401 2930 : MyBackendType = B_AUTOVAC_WORKER;
1402 2930 : init_ps_display(NULL);
1403 :
1404 : Assert(GetProcessingMode() == InitProcessing);
1405 :
1406 : /*
1407 : * Set up signal handlers. We operate on databases much like a regular
1408 : * backend, so we use the same signal handling. See equivalent code in
1409 : * tcop/postgres.c.
1410 : */
1411 2930 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1412 :
1413 : /*
1414 : * SIGINT is used to signal canceling the current table's vacuum; SIGTERM
1415 : * means abort and exit cleanly, and SIGQUIT means abandon ship.
1416 : */
1417 2930 : pqsignal(SIGINT, StatementCancelHandler);
1418 2930 : pqsignal(SIGTERM, die);
1419 : /* SIGQUIT handler was already set up by InitPostmasterChild */
1420 :
1421 2930 : InitializeTimeouts(); /* establishes SIGALRM handler */
1422 :
1423 2930 : pqsignal(SIGPIPE, SIG_IGN);
1424 2930 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1425 2930 : pqsignal(SIGUSR2, SIG_IGN);
1426 2930 : pqsignal(SIGFPE, FloatExceptionHandler);
1427 2930 : pqsignal(SIGCHLD, SIG_DFL);
1428 :
1429 : /*
1430 : * Create a per-backend PGPROC struct in shared memory. We must do this
1431 : * before we can use LWLocks or access any shared memory.
1432 : */
1433 2930 : InitProcess();
1434 :
1435 : /* Early initialization */
1436 2930 : BaseInit();
1437 :
1438 : /*
1439 : * If an exception is encountered, processing resumes here.
1440 : *
1441 : * Unlike most auxiliary processes, we don't attempt to continue
1442 : * processing after an error; we just clean up and exit. The autovac
1443 : * launcher is responsible for spawning another worker later.
1444 : *
1445 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
1446 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
1447 : * signals other than SIGQUIT will be blocked until we exit. It might
1448 : * seem that this policy makes the HOLD_INTERRUPTS() call redundant, but
1449 : * it is not since InterruptPending might be set already.
1450 : */
1451 2930 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1452 : {
1453 : /* since not using PG_TRY, must reset error stack by hand */
1454 0 : error_context_stack = NULL;
1455 :
1456 : /* Prevents interrupts while cleaning up */
1457 0 : HOLD_INTERRUPTS();
1458 :
1459 : /* Report the error to the server log */
1460 0 : EmitErrorReport();
1461 :
1462 : /*
1463 : * We can now go away. Note that because we called InitProcess, a
1464 : * callback was registered to do ProcKill, which will clean up
1465 : * necessary state.
1466 : */
1467 0 : proc_exit(0);
1468 : }
1469 :
1470 : /* We can now handle ereport(ERROR) */
1471 2930 : PG_exception_stack = &local_sigjmp_buf;
1472 :
1473 2930 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
1474 :
1475 : /*
1476 : * Set always-secure search path, so malicious users can't redirect user
1477 : * code (e.g. pg_index.indexprs). (That code runs in a
1478 : * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not
1479 : * take control of the entire autovacuum worker in any case.)
1480 : */
1481 2930 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1482 :
1483 : /*
1484 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
1485 : * in postgresql.conf. We don't really want such a dangerous option being
1486 : * applied non-interactively.
1487 : */
1488 2930 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
1489 :
1490 : /*
1491 : * Force settable timeouts off to avoid letting these settings prevent
1492 : * regular maintenance from being executed.
1493 : */
1494 2930 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1495 2930 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1496 2930 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1497 2930 : SetConfigOption("idle_in_transaction_session_timeout", "0",
1498 : PGC_SUSET, PGC_S_OVERRIDE);
1499 :
1500 : /*
1501 : * Force default_transaction_isolation to READ COMMITTED. We don't want
1502 : * to pay the overhead of serializable mode, nor add any risk of causing
1503 : * deadlocks or delaying other transactions.
1504 : */
1505 2930 : SetConfigOption("default_transaction_isolation", "read committed",
1506 : PGC_SUSET, PGC_S_OVERRIDE);
1507 :
1508 : /*
1509 : * Force synchronous replication off to allow regular maintenance even if
1510 : * we are waiting for standbys to connect. This is important to ensure we
1511 : * aren't blocked from performing anti-wraparound tasks.
1512 : */
1513 2930 : if (synchronous_commit > SYNCHRONOUS_COMMIT_LOCAL_FLUSH)
1514 2930 : SetConfigOption("synchronous_commit", "local",
1515 : PGC_SUSET, PGC_S_OVERRIDE);
1516 :
1517 : /*
1518 : * Even when system is configured to use a different fetch consistency,
1519 : * for autovac we always want fresh stats.
1520 : */
1521 2930 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
1522 :
1523 : /*
1524 : * Get the info about the database we're going to work on.
1525 : */
1526 2930 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1527 :
1528 : /*
1529 : * beware of startingWorker being INVALID; this should normally not
1530 : * happen, but if a worker fails after forking and before this, the
1531 : * launcher might have decided to remove it from the queue and start
1532 : * again.
1533 : */
1534 2930 : if (AutoVacuumShmem->av_startingWorker != NULL)
1535 : {
1536 2930 : MyWorkerInfo = AutoVacuumShmem->av_startingWorker;
1537 2930 : dbid = MyWorkerInfo->wi_dboid;
1538 2930 : MyWorkerInfo->wi_proc = MyProc;
1539 :
1540 : /* insert into the running list */
1541 2930 : dlist_push_head(&AutoVacuumShmem->av_runningWorkers,
1542 2930 : &MyWorkerInfo->wi_links);
1543 :
1544 : /*
1545 : * remove from the "starting" pointer, so that the launcher can start
1546 : * a new worker if required
1547 : */
1548 2930 : AutoVacuumShmem->av_startingWorker = NULL;
1549 2930 : LWLockRelease(AutovacuumLock);
1550 :
1551 2930 : on_shmem_exit(FreeWorkerInfo, 0);
1552 :
1553 : /* wake up the launcher */
1554 2930 : if (AutoVacuumShmem->av_launcherpid != 0)
1555 2930 : kill(AutoVacuumShmem->av_launcherpid, SIGUSR2);
1556 : }
1557 : else
1558 : {
1559 : /* no worker entry for me, go away */
1560 0 : elog(WARNING, "autovacuum worker started without a worker entry");
1561 0 : dbid = InvalidOid;
1562 0 : LWLockRelease(AutovacuumLock);
1563 : }
1564 :
1565 2930 : if (OidIsValid(dbid))
1566 : {
1567 : char dbname[NAMEDATALEN];
1568 :
1569 : /*
1570 : * Report autovac startup to the cumulative stats system. We
1571 : * deliberately do this before InitPostgres, so that the
1572 : * last_autovac_time will get updated even if the connection attempt
1573 : * fails. This is to prevent autovac from getting "stuck" repeatedly
1574 : * selecting an unopenable database, rather than making any progress
1575 : * on stuff it can connect to.
1576 : */
1577 2930 : pgstat_report_autovac(dbid);
1578 :
1579 : /*
1580 : * Connect to the selected database, specifying no particular user,
1581 : * and ignoring datallowconn. Collect the database's name for
1582 : * display.
1583 : *
1584 : * Note: if we have selected a just-deleted database (due to using
1585 : * stale stats info), we'll fail and exit here.
1586 : */
1587 2930 : InitPostgres(NULL, dbid, NULL, InvalidOid,
1588 : INIT_PG_OVERRIDE_ALLOW_CONNS,
1589 : dbname);
1590 2930 : SetProcessingMode(NormalProcessing);
1591 2930 : set_ps_display(dbname);
1592 2930 : ereport(DEBUG1,
1593 : (errmsg_internal("autovacuum: processing database \"%s\"", dbname)));
1594 :
1595 2930 : if (PostAuthDelay)
1596 0 : pg_usleep(PostAuthDelay * 1000000L);
1597 :
1598 : /* And do an appropriate amount of work */
1599 2930 : recentXid = ReadNextTransactionId();
1600 2930 : recentMulti = ReadNextMultiXactId();
1601 2930 : do_autovacuum();
1602 : }
1603 :
1604 : /* All done, go away */
1605 2928 : proc_exit(0);
1606 : }
1607 :
1608 : /*
1609 : * Return a WorkerInfo to the free list
1610 : */
1611 : static void
1612 2930 : FreeWorkerInfo(int code, Datum arg)
1613 : {
1614 2930 : if (MyWorkerInfo != NULL)
1615 : {
1616 2930 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1617 :
1618 2930 : dlist_delete(&MyWorkerInfo->wi_links);
1619 2930 : MyWorkerInfo->wi_dboid = InvalidOid;
1620 2930 : MyWorkerInfo->wi_tableoid = InvalidOid;
1621 2930 : MyWorkerInfo->wi_sharedrel = false;
1622 2930 : MyWorkerInfo->wi_proc = NULL;
1623 2930 : MyWorkerInfo->wi_launchtime = 0;
1624 2930 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
1625 2930 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
1626 2930 : &MyWorkerInfo->wi_links);
1627 : /* not mine anymore */
1628 2930 : MyWorkerInfo = NULL;
1629 :
1630 : /*
1631 : * now that we're inactive, cause a rebalancing of the surviving
1632 : * workers
1633 : */
1634 2930 : AutoVacuumShmem->av_signal[AutoVacRebalance] = true;
1635 2930 : LWLockRelease(AutovacuumLock);
1636 : }
1637 2930 : }
1638 :
1639 : /*
1640 : * Update vacuum cost-based delay-related parameters for autovacuum workers and
1641 : * backends executing VACUUM or ANALYZE using the value of relevant GUCs and
1642 : * global state. This must be called during setup for vacuum and after every
1643 : * config reload to ensure up-to-date values.
1644 : */
1645 : void
1646 386094 : VacuumUpdateCosts(void)
1647 : {
1648 386094 : if (MyWorkerInfo)
1649 : {
1650 371352 : if (av_storage_param_cost_delay >= 0)
1651 0 : vacuum_cost_delay = av_storage_param_cost_delay;
1652 371352 : else if (autovacuum_vac_cost_delay >= 0)
1653 371352 : vacuum_cost_delay = autovacuum_vac_cost_delay;
1654 : else
1655 : /* fall back to VacuumCostDelay */
1656 0 : vacuum_cost_delay = VacuumCostDelay;
1657 :
1658 371352 : AutoVacuumUpdateCostLimit();
1659 : }
1660 : else
1661 : {
1662 : /* Must be explicit VACUUM or ANALYZE */
1663 14742 : vacuum_cost_delay = VacuumCostDelay;
1664 14742 : vacuum_cost_limit = VacuumCostLimit;
1665 : }
1666 :
1667 : /*
1668 : * If configuration changes are allowed to impact VacuumCostActive, make
1669 : * sure it is updated.
1670 : */
1671 386094 : if (VacuumFailsafeActive)
1672 : Assert(!VacuumCostActive);
1673 386094 : else if (vacuum_cost_delay > 0)
1674 371352 : VacuumCostActive = true;
1675 : else
1676 : {
1677 14742 : VacuumCostActive = false;
1678 14742 : VacuumCostBalance = 0;
1679 : }
1680 :
1681 : /*
1682 : * Since the cost logging requires a lock, avoid rendering the log message
1683 : * in case we are using a message level where the log wouldn't be emitted.
1684 : */
1685 386094 : if (MyWorkerInfo && message_level_is_interesting(DEBUG2))
1686 : {
1687 : Oid dboid,
1688 : tableoid;
1689 :
1690 : Assert(!LWLockHeldByMe(AutovacuumLock));
1691 :
1692 0 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1693 0 : dboid = MyWorkerInfo->wi_dboid;
1694 0 : tableoid = MyWorkerInfo->wi_tableoid;
1695 0 : LWLockRelease(AutovacuumLock);
1696 :
1697 0 : elog(DEBUG2,
1698 : "Autovacuum VacuumUpdateCosts(db=%u, rel=%u, dobalance=%s, cost_limit=%d, cost_delay=%g active=%s failsafe=%s)",
1699 : dboid, tableoid, pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance) ? "no" : "yes",
1700 : vacuum_cost_limit, vacuum_cost_delay,
1701 : vacuum_cost_delay > 0 ? "yes" : "no",
1702 : VacuumFailsafeActive ? "yes" : "no");
1703 : }
1704 386094 : }
1705 :
1706 : /*
1707 : * Update vacuum_cost_limit with the correct value for an autovacuum worker,
1708 : * given the value of other relevant cost limit parameters and the number of
1709 : * workers across which the limit must be balanced. Autovacuum workers must
1710 : * call this regularly in case av_nworkersForBalance has been updated by
1711 : * another worker or by the autovacuum launcher. They must also call it after a
1712 : * config reload.
1713 : */
1714 : void
1715 377554 : AutoVacuumUpdateCostLimit(void)
1716 : {
1717 377554 : if (!MyWorkerInfo)
1718 0 : return;
1719 :
1720 : /*
1721 : * note: in cost_limit, zero also means use value from elsewhere, because
1722 : * zero is not a valid value.
1723 : */
1724 :
1725 377554 : if (av_storage_param_cost_limit > 0)
1726 0 : vacuum_cost_limit = av_storage_param_cost_limit;
1727 : else
1728 : {
1729 : int nworkers_for_balance;
1730 :
1731 377554 : if (autovacuum_vac_cost_limit > 0)
1732 0 : vacuum_cost_limit = autovacuum_vac_cost_limit;
1733 : else
1734 377554 : vacuum_cost_limit = VacuumCostLimit;
1735 :
1736 : /* Only balance limit if no cost-related storage parameters specified */
1737 377554 : if (pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance))
1738 0 : return;
1739 :
1740 : Assert(vacuum_cost_limit > 0);
1741 :
1742 377554 : nworkers_for_balance = pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1743 :
1744 : /* There is at least 1 autovac worker (this worker) */
1745 377554 : if (nworkers_for_balance <= 0)
1746 0 : elog(ERROR, "nworkers_for_balance must be > 0");
1747 :
1748 377554 : vacuum_cost_limit = Max(vacuum_cost_limit / nworkers_for_balance, 1);
1749 : }
1750 : }
1751 :
1752 : /*
1753 : * autovac_recalculate_workers_for_balance
1754 : * Recalculate the number of workers to consider, given cost-related
1755 : * storage parameters and the current number of active workers.
1756 : *
1757 : * Caller must hold the AutovacuumLock in at least shared mode to access
1758 : * worker->wi_proc.
1759 : */
1760 : static void
1761 188446 : autovac_recalculate_workers_for_balance(void)
1762 : {
1763 : dlist_iter iter;
1764 : int orig_nworkers_for_balance;
1765 188446 : int nworkers_for_balance = 0;
1766 :
1767 : Assert(LWLockHeldByMe(AutovacuumLock));
1768 :
1769 188446 : orig_nworkers_for_balance =
1770 188446 : pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1771 :
1772 577502 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1773 : {
1774 389056 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1775 :
1776 778112 : if (worker->wi_proc == NULL ||
1777 389056 : pg_atomic_unlocked_test_flag(&worker->wi_dobalance))
1778 6442 : continue;
1779 :
1780 382614 : nworkers_for_balance++;
1781 : }
1782 :
1783 188446 : if (nworkers_for_balance != orig_nworkers_for_balance)
1784 3708 : pg_atomic_write_u32(&AutoVacuumShmem->av_nworkersForBalance,
1785 : nworkers_for_balance);
1786 188446 : }
1787 :
1788 : /*
1789 : * get_database_list
1790 : * Return a list of all databases found in pg_database.
1791 : *
1792 : * The list and associated data is allocated in the caller's memory context,
1793 : * which is in charge of ensuring that it's properly cleaned up afterwards.
1794 : *
1795 : * Note: this is the only function in which the autovacuum launcher uses a
1796 : * transaction. Although we aren't attached to any particular database and
1797 : * therefore can't access most catalogs, we do have enough infrastructure
1798 : * to do a seqscan on pg_database.
1799 : */
1800 : static List *
1801 4076 : get_database_list(void)
1802 : {
1803 4076 : List *dblist = NIL;
1804 : Relation rel;
1805 : TableScanDesc scan;
1806 : HeapTuple tup;
1807 : MemoryContext resultcxt;
1808 :
1809 : /* This is the context that we will allocate our output data in */
1810 4076 : resultcxt = CurrentMemoryContext;
1811 :
1812 : /*
1813 : * Start a transaction so we can access pg_database.
1814 : */
1815 4076 : StartTransactionCommand();
1816 :
1817 4076 : rel = table_open(DatabaseRelationId, AccessShareLock);
1818 4076 : scan = table_beginscan_catalog(rel, 0, NULL);
1819 :
1820 16776 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1821 : {
1822 12700 : Form_pg_database pgdatabase = (Form_pg_database) GETSTRUCT(tup);
1823 : avw_dbase *avdb;
1824 : MemoryContext oldcxt;
1825 :
1826 : /*
1827 : * If database has partially been dropped, we can't, nor need to,
1828 : * vacuum it.
1829 : */
1830 12700 : if (database_is_invalid_form(pgdatabase))
1831 : {
1832 4 : elog(DEBUG2,
1833 : "autovacuum: skipping invalid database \"%s\"",
1834 : NameStr(pgdatabase->datname));
1835 4 : continue;
1836 : }
1837 :
1838 : /*
1839 : * Allocate our results in the caller's context, not the
1840 : * transaction's. We do this inside the loop, and restore the original
1841 : * context at the end, so that leaky things like heap_getnext() are
1842 : * not called in a potentially long-lived context.
1843 : */
1844 12696 : oldcxt = MemoryContextSwitchTo(resultcxt);
1845 :
1846 12696 : avdb = palloc_object(avw_dbase);
1847 :
1848 12696 : avdb->adw_datid = pgdatabase->oid;
1849 12696 : avdb->adw_name = pstrdup(NameStr(pgdatabase->datname));
1850 12696 : avdb->adw_frozenxid = pgdatabase->datfrozenxid;
1851 12696 : avdb->adw_minmulti = pgdatabase->datminmxid;
1852 : /* this gets set later: */
1853 12696 : avdb->adw_entry = NULL;
1854 :
1855 12696 : dblist = lappend(dblist, avdb);
1856 12696 : MemoryContextSwitchTo(oldcxt);
1857 : }
1858 :
1859 4076 : table_endscan(scan);
1860 4076 : table_close(rel, AccessShareLock);
1861 :
1862 4076 : CommitTransactionCommand();
1863 :
1864 : /* Be sure to restore caller's memory context */
1865 4076 : MemoryContextSwitchTo(resultcxt);
1866 :
1867 4076 : return dblist;
1868 : }
1869 :
1870 : /*
1871 : * Process a database table-by-table
1872 : *
1873 : * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in
1874 : * order not to ignore shutdown commands for too long.
1875 : */
1876 : static void
1877 2930 : do_autovacuum(void)
1878 : {
1879 : Relation classRel;
1880 : HeapTuple tuple;
1881 : TableScanDesc relScan;
1882 : Form_pg_database dbForm;
1883 2930 : List *table_oids = NIL;
1884 2930 : List *orphan_oids = NIL;
1885 : HASHCTL ctl;
1886 : HTAB *table_toast_map;
1887 : ListCell *volatile cell;
1888 : BufferAccessStrategy bstrategy;
1889 : ScanKeyData key;
1890 : TupleDesc pg_class_desc;
1891 : int effective_multixact_freeze_max_age;
1892 2930 : bool did_vacuum = false;
1893 2930 : bool found_concurrent_worker = false;
1894 : int i;
1895 :
1896 : /*
1897 : * StartTransactionCommand and CommitTransactionCommand will automatically
1898 : * switch to other contexts. We need this one to keep the list of
1899 : * relations to vacuum/analyze across transactions.
1900 : */
1901 2930 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
1902 : "Autovacuum worker",
1903 : ALLOCSET_DEFAULT_SIZES);
1904 2930 : MemoryContextSwitchTo(AutovacMemCxt);
1905 :
1906 : /* Start a transaction so our commands have one to play into. */
1907 2930 : StartTransactionCommand();
1908 :
1909 : /*
1910 : * This injection point is put in a transaction block to work with a wait
1911 : * that uses a condition variable.
1912 : */
1913 2930 : INJECTION_POINT("autovacuum-worker-start", NULL);
1914 :
1915 : /*
1916 : * Compute the multixact age for which freezing is urgent. This is
1917 : * normally autovacuum_multixact_freeze_max_age, but may be less if
1918 : * multixact members are bloated.
1919 : */
1920 2928 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1921 :
1922 : /*
1923 : * Find the pg_database entry and select the default freeze ages. We use
1924 : * zero in template and nonconnectable databases, else the system-wide
1925 : * default.
1926 : */
1927 2928 : tuple = SearchSysCache1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
1928 2928 : if (!HeapTupleIsValid(tuple))
1929 0 : elog(ERROR, "cache lookup failed for database %u", MyDatabaseId);
1930 2928 : dbForm = (Form_pg_database) GETSTRUCT(tuple);
1931 :
1932 2928 : if (dbForm->datistemplate || !dbForm->datallowconn)
1933 : {
1934 990 : default_freeze_min_age = 0;
1935 990 : default_freeze_table_age = 0;
1936 990 : default_multixact_freeze_min_age = 0;
1937 990 : default_multixact_freeze_table_age = 0;
1938 : }
1939 : else
1940 : {
1941 1938 : default_freeze_min_age = vacuum_freeze_min_age;
1942 1938 : default_freeze_table_age = vacuum_freeze_table_age;
1943 1938 : default_multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1944 1938 : default_multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1945 : }
1946 :
1947 2928 : ReleaseSysCache(tuple);
1948 :
1949 : /* StartTransactionCommand changed elsewhere */
1950 2928 : MemoryContextSwitchTo(AutovacMemCxt);
1951 :
1952 2928 : classRel = table_open(RelationRelationId, AccessShareLock);
1953 :
1954 : /* create a copy so we can use it after closing pg_class */
1955 2928 : pg_class_desc = CreateTupleDescCopy(RelationGetDescr(classRel));
1956 :
1957 : /* create hash table for toast <-> main relid mapping */
1958 2928 : ctl.keysize = sizeof(Oid);
1959 2928 : ctl.entrysize = sizeof(av_relation);
1960 :
1961 2928 : table_toast_map = hash_create("TOAST to main relid map",
1962 : 100,
1963 : &ctl,
1964 : HASH_ELEM | HASH_BLOBS);
1965 :
1966 : /*
1967 : * Scan pg_class to determine which tables to vacuum.
1968 : *
1969 : * We do this in two passes: on the first one we collect the list of plain
1970 : * relations and materialized views, and on the second one we collect
1971 : * TOAST tables. The reason for doing the second pass is that during it we
1972 : * want to use the main relation's pg_class.reloptions entry if the TOAST
1973 : * table does not have any, and we cannot obtain it unless we know
1974 : * beforehand what's the main table OID.
1975 : *
1976 : * We need to check TOAST tables separately because in cases with short,
1977 : * wide tables there might be proportionally much more activity in the
1978 : * TOAST table than in its parent.
1979 : */
1980 2928 : relScan = table_beginscan_catalog(classRel, 0, NULL);
1981 :
1982 : /*
1983 : * On the first pass, we collect main tables to vacuum, and also the main
1984 : * table relid to TOAST relid mapping.
1985 : */
1986 1240394 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
1987 : {
1988 1237466 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
1989 : PgStat_StatTabEntry *tabentry;
1990 : AutoVacOpts *relopts;
1991 : Oid relid;
1992 : bool dovacuum;
1993 : bool doanalyze;
1994 : bool wraparound;
1995 :
1996 1237466 : if (classForm->relkind != RELKIND_RELATION &&
1997 1032848 : classForm->relkind != RELKIND_MATVIEW)
1998 1032786 : continue;
1999 :
2000 204692 : relid = classForm->oid;
2001 :
2002 : /*
2003 : * Check if it is a temp table (presumably, of some other backend's).
2004 : * We cannot safely process other backends' temp tables.
2005 : */
2006 204692 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2007 : {
2008 : /*
2009 : * We just ignore it if the owning backend is still active and
2010 : * using the temporary schema. Also, for safety, ignore it if the
2011 : * namespace doesn't exist or isn't a temp namespace after all.
2012 : */
2013 12 : if (checkTempNamespaceStatus(classForm->relnamespace) == TEMP_NAMESPACE_IDLE)
2014 : {
2015 : /*
2016 : * The table seems to be orphaned -- although it might be that
2017 : * the owning backend has already deleted it and exited; our
2018 : * pg_class scan snapshot is not necessarily up-to-date
2019 : * anymore, so we could be looking at a committed-dead entry.
2020 : * Remember it so we can try to delete it later.
2021 : */
2022 0 : orphan_oids = lappend_oid(orphan_oids, relid);
2023 : }
2024 12 : continue;
2025 : }
2026 :
2027 : /* Fetch reloptions and the pgstat entry for this table */
2028 204680 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2029 204680 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2030 : relid);
2031 :
2032 : /* Check if it needs vacuum or analyze */
2033 204680 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2034 : effective_multixact_freeze_max_age,
2035 : &dovacuum, &doanalyze, &wraparound);
2036 :
2037 : /* Relations that need work are added to table_oids */
2038 204680 : if (dovacuum || doanalyze)
2039 118726 : table_oids = lappend_oid(table_oids, relid);
2040 :
2041 : /*
2042 : * Remember TOAST associations for the second pass. Note: we must do
2043 : * this whether or not the table is going to be vacuumed, because we
2044 : * don't automatically vacuum toast tables along the parent table.
2045 : */
2046 204680 : if (OidIsValid(classForm->reltoastrelid))
2047 : {
2048 : av_relation *hentry;
2049 : bool found;
2050 :
2051 235564 : hentry = hash_search(table_toast_map,
2052 117782 : &classForm->reltoastrelid,
2053 : HASH_ENTER, &found);
2054 :
2055 117782 : if (!found)
2056 : {
2057 : /* hash_search already filled in the key */
2058 117782 : hentry->ar_relid = relid;
2059 117782 : hentry->ar_hasrelopts = false;
2060 117782 : if (relopts != NULL)
2061 : {
2062 2526 : hentry->ar_hasrelopts = true;
2063 2526 : memcpy(&hentry->ar_reloptions, relopts,
2064 : sizeof(AutoVacOpts));
2065 : }
2066 : }
2067 : }
2068 :
2069 : /* Release stuff to avoid per-relation leakage */
2070 204680 : if (relopts)
2071 2588 : pfree(relopts);
2072 204680 : if (tabentry)
2073 200702 : pfree(tabentry);
2074 : }
2075 :
2076 2928 : table_endscan(relScan);
2077 :
2078 : /* second pass: check TOAST tables */
2079 2928 : ScanKeyInit(&key,
2080 : Anum_pg_class_relkind,
2081 : BTEqualStrategyNumber, F_CHAREQ,
2082 : CharGetDatum(RELKIND_TOASTVALUE));
2083 :
2084 2928 : relScan = table_beginscan_catalog(classRel, 1, &key);
2085 120714 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2086 : {
2087 117786 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2088 : PgStat_StatTabEntry *tabentry;
2089 : Oid relid;
2090 : AutoVacOpts *relopts;
2091 117786 : bool free_relopts = false;
2092 : bool dovacuum;
2093 : bool doanalyze;
2094 : bool wraparound;
2095 :
2096 : /*
2097 : * We cannot safely process other backends' temp tables, so skip 'em.
2098 : */
2099 117786 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2100 4 : continue;
2101 :
2102 117782 : relid = classForm->oid;
2103 :
2104 : /*
2105 : * fetch reloptions -- if this toast table does not have them, try the
2106 : * main rel
2107 : */
2108 117782 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2109 117782 : if (relopts)
2110 0 : free_relopts = true;
2111 : else
2112 : {
2113 : av_relation *hentry;
2114 : bool found;
2115 :
2116 117782 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2117 117782 : if (found && hentry->ar_hasrelopts)
2118 2526 : relopts = &hentry->ar_reloptions;
2119 : }
2120 :
2121 : /* Fetch the pgstat entry for this table */
2122 117782 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2123 : relid);
2124 :
2125 117782 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2126 : effective_multixact_freeze_max_age,
2127 : &dovacuum, &doanalyze, &wraparound);
2128 :
2129 : /* ignore analyze for toast tables */
2130 117782 : if (dovacuum)
2131 68164 : table_oids = lappend_oid(table_oids, relid);
2132 :
2133 : /* Release stuff to avoid leakage */
2134 117782 : if (free_relopts)
2135 0 : pfree(relopts);
2136 117782 : if (tabentry)
2137 114926 : pfree(tabentry);
2138 : }
2139 :
2140 2928 : table_endscan(relScan);
2141 2928 : table_close(classRel, AccessShareLock);
2142 :
2143 : /*
2144 : * Recheck orphan temporary tables, and if they still seem orphaned, drop
2145 : * them. We'll eat a transaction per dropped table, which might seem
2146 : * excessive, but we should only need to do anything as a result of a
2147 : * previous backend crash, so this should not happen often enough to
2148 : * justify "optimizing". Using separate transactions ensures that we
2149 : * don't bloat the lock table if there are many temp tables to be dropped,
2150 : * and it ensures that we don't lose work if a deletion attempt fails.
2151 : */
2152 2928 : foreach(cell, orphan_oids)
2153 : {
2154 0 : Oid relid = lfirst_oid(cell);
2155 : Form_pg_class classForm;
2156 : ObjectAddress object;
2157 :
2158 : /*
2159 : * Check for user-requested abort.
2160 : */
2161 0 : CHECK_FOR_INTERRUPTS();
2162 :
2163 : /*
2164 : * Try to lock the table. If we can't get the lock immediately,
2165 : * somebody else is using (or dropping) the table, so it's not our
2166 : * concern anymore. Having the lock prevents race conditions below.
2167 : */
2168 0 : if (!ConditionalLockRelationOid(relid, AccessExclusiveLock))
2169 0 : continue;
2170 :
2171 : /*
2172 : * Re-fetch the pg_class tuple and re-check whether it still seems to
2173 : * be an orphaned temp table. If it's not there or no longer the same
2174 : * relation, ignore it.
2175 : */
2176 0 : tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2177 0 : if (!HeapTupleIsValid(tuple))
2178 : {
2179 : /* be sure to drop useless lock so we don't bloat lock table */
2180 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2181 0 : continue;
2182 : }
2183 0 : classForm = (Form_pg_class) GETSTRUCT(tuple);
2184 :
2185 : /*
2186 : * Make all the same tests made in the loop above. In event of OID
2187 : * counter wraparound, the pg_class entry we have now might be
2188 : * completely unrelated to the one we saw before.
2189 : */
2190 0 : if (!((classForm->relkind == RELKIND_RELATION ||
2191 0 : classForm->relkind == RELKIND_MATVIEW) &&
2192 0 : classForm->relpersistence == RELPERSISTENCE_TEMP))
2193 : {
2194 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2195 0 : continue;
2196 : }
2197 :
2198 0 : if (checkTempNamespaceStatus(classForm->relnamespace) != TEMP_NAMESPACE_IDLE)
2199 : {
2200 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2201 0 : continue;
2202 : }
2203 :
2204 : /*
2205 : * Try to lock the temp namespace, too. Even though we have lock on
2206 : * the table itself, there's a risk of deadlock against an incoming
2207 : * backend trying to clean out the temp namespace, in case this table
2208 : * has dependencies (such as sequences) that the backend's
2209 : * performDeletion call might visit in a different order. If we can
2210 : * get AccessShareLock on the namespace, that's sufficient to ensure
2211 : * we're not running concurrently with RemoveTempRelations. If we
2212 : * can't, back off and let RemoveTempRelations do its thing.
2213 : */
2214 0 : if (!ConditionalLockDatabaseObject(NamespaceRelationId,
2215 : classForm->relnamespace, 0,
2216 : AccessShareLock))
2217 : {
2218 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2219 0 : continue;
2220 : }
2221 :
2222 : /* OK, let's delete it */
2223 0 : ereport(LOG,
2224 : (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"",
2225 : get_database_name(MyDatabaseId),
2226 : get_namespace_name(classForm->relnamespace),
2227 : NameStr(classForm->relname))));
2228 :
2229 : /*
2230 : * Deletion might involve TOAST table access, so ensure we have a
2231 : * valid snapshot.
2232 : */
2233 0 : PushActiveSnapshot(GetTransactionSnapshot());
2234 :
2235 0 : object.classId = RelationRelationId;
2236 0 : object.objectId = relid;
2237 0 : object.objectSubId = 0;
2238 0 : performDeletion(&object, DROP_CASCADE,
2239 : PERFORM_DELETION_INTERNAL |
2240 : PERFORM_DELETION_QUIETLY |
2241 : PERFORM_DELETION_SKIP_EXTENSIONS);
2242 :
2243 : /*
2244 : * To commit the deletion, end current transaction and start a new
2245 : * one. Note this also releases the locks we took.
2246 : */
2247 0 : PopActiveSnapshot();
2248 0 : CommitTransactionCommand();
2249 0 : StartTransactionCommand();
2250 :
2251 : /* StartTransactionCommand changed current memory context */
2252 0 : MemoryContextSwitchTo(AutovacMemCxt);
2253 : }
2254 :
2255 : /*
2256 : * Optionally, create a buffer access strategy object for VACUUM to use.
2257 : * We use the same BufferAccessStrategy object for all tables VACUUMed by
2258 : * this worker to prevent autovacuum from blowing out shared buffers.
2259 : *
2260 : * VacuumBufferUsageLimit being set to 0 results in
2261 : * GetAccessStrategyWithSize returning NULL, effectively meaning we can
2262 : * use up to all of shared buffers.
2263 : *
2264 : * If we later enter failsafe mode on any of the tables being vacuumed, we
2265 : * will cease use of the BufferAccessStrategy only for that table.
2266 : *
2267 : * XXX should we consider adding code to adjust the size of this if
2268 : * VacuumBufferUsageLimit changes?
2269 : */
2270 2928 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, VacuumBufferUsageLimit);
2271 :
2272 : /*
2273 : * create a memory context to act as fake PortalContext, so that the
2274 : * contexts created in the vacuum code are cleaned up for each table.
2275 : */
2276 2928 : PortalContext = AllocSetContextCreate(AutovacMemCxt,
2277 : "Autovacuum Portal",
2278 : ALLOCSET_DEFAULT_SIZES);
2279 :
2280 : /*
2281 : * Perform operations on collected tables.
2282 : */
2283 189818 : foreach(cell, table_oids)
2284 : {
2285 186890 : Oid relid = lfirst_oid(cell);
2286 : HeapTuple classTup;
2287 : autovac_table *tab;
2288 : bool isshared;
2289 : bool skipit;
2290 : dlist_iter iter;
2291 :
2292 186890 : CHECK_FOR_INTERRUPTS();
2293 :
2294 : /*
2295 : * Check for config changes before processing each collected table.
2296 : */
2297 186890 : if (ConfigReloadPending)
2298 : {
2299 0 : ConfigReloadPending = false;
2300 0 : ProcessConfigFile(PGC_SIGHUP);
2301 :
2302 : /*
2303 : * You might be tempted to bail out if we see autovacuum is now
2304 : * disabled. Must resist that temptation -- this might be a
2305 : * for-wraparound emergency worker, in which case that would be
2306 : * entirely inappropriate.
2307 : */
2308 : }
2309 :
2310 : /*
2311 : * Find out whether the table is shared or not. (It's slightly
2312 : * annoying to fetch the syscache entry just for this, but in typical
2313 : * cases it adds little cost because table_recheck_autovac would
2314 : * refetch the entry anyway. We could buy that back by copying the
2315 : * tuple here and passing it to table_recheck_autovac, but that
2316 : * increases the odds of that function working with stale data.)
2317 : */
2318 186890 : classTup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
2319 186890 : if (!HeapTupleIsValid(classTup))
2320 1214 : continue; /* somebody deleted the rel, forget it */
2321 186890 : isshared = ((Form_pg_class) GETSTRUCT(classTup))->relisshared;
2322 186890 : ReleaseSysCache(classTup);
2323 :
2324 : /*
2325 : * Hold schedule lock from here until we've claimed the table. We
2326 : * also need the AutovacuumLock to walk the worker array, but that one
2327 : * can just be a shared lock.
2328 : */
2329 186890 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2330 186890 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2331 :
2332 : /*
2333 : * Check whether the table is being vacuumed concurrently by another
2334 : * worker.
2335 : */
2336 186890 : skipit = false;
2337 575514 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
2338 : {
2339 389772 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
2340 :
2341 : /* ignore myself */
2342 389772 : if (worker == MyWorkerInfo)
2343 186358 : continue;
2344 :
2345 : /* ignore workers in other databases (unless table is shared) */
2346 203414 : if (!worker->wi_sharedrel && worker->wi_dboid != MyDatabaseId)
2347 8 : continue;
2348 :
2349 203406 : if (worker->wi_tableoid == relid)
2350 : {
2351 1148 : skipit = true;
2352 1148 : found_concurrent_worker = true;
2353 1148 : break;
2354 : }
2355 : }
2356 186890 : LWLockRelease(AutovacuumLock);
2357 186890 : if (skipit)
2358 : {
2359 1148 : LWLockRelease(AutovacuumScheduleLock);
2360 1148 : continue;
2361 : }
2362 :
2363 : /*
2364 : * Store the table's OID in shared memory before releasing the
2365 : * schedule lock, so that other workers don't try to vacuum it
2366 : * concurrently. (We claim it here so as not to hold
2367 : * AutovacuumScheduleLock while rechecking the stats.)
2368 : */
2369 185742 : MyWorkerInfo->wi_tableoid = relid;
2370 185742 : MyWorkerInfo->wi_sharedrel = isshared;
2371 185742 : LWLockRelease(AutovacuumScheduleLock);
2372 :
2373 : /*
2374 : * Check whether pgstat data still says we need to vacuum this table.
2375 : * It could have changed if something else processed the table while
2376 : * we weren't looking. This doesn't entirely close the race condition,
2377 : * but it is very small.
2378 : */
2379 185742 : MemoryContextSwitchTo(AutovacMemCxt);
2380 185742 : tab = table_recheck_autovac(relid, table_toast_map, pg_class_desc,
2381 : effective_multixact_freeze_max_age);
2382 185742 : if (tab == NULL)
2383 : {
2384 : /* someone else vacuumed the table, or it went away */
2385 66 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2386 66 : MyWorkerInfo->wi_tableoid = InvalidOid;
2387 66 : MyWorkerInfo->wi_sharedrel = false;
2388 66 : LWLockRelease(AutovacuumScheduleLock);
2389 66 : continue;
2390 : }
2391 :
2392 : /*
2393 : * Save the cost-related storage parameter values in global variables
2394 : * for reference when updating vacuum_cost_delay and vacuum_cost_limit
2395 : * during vacuuming this table.
2396 : */
2397 185676 : av_storage_param_cost_delay = tab->at_storage_param_vac_cost_delay;
2398 185676 : av_storage_param_cost_limit = tab->at_storage_param_vac_cost_limit;
2399 :
2400 : /*
2401 : * We only expect this worker to ever set the flag, so don't bother
2402 : * checking the return value. We shouldn't have to retry.
2403 : */
2404 185676 : if (tab->at_dobalance)
2405 185676 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2406 : else
2407 0 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
2408 :
2409 185676 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2410 185676 : autovac_recalculate_workers_for_balance();
2411 185676 : LWLockRelease(AutovacuumLock);
2412 :
2413 : /*
2414 : * We wait until this point to update cost delay and cost limit
2415 : * values, even though we reloaded the configuration file above, so
2416 : * that we can take into account the cost-related storage parameters.
2417 : */
2418 185676 : VacuumUpdateCosts();
2419 :
2420 :
2421 : /* clean up memory before each iteration */
2422 185676 : MemoryContextReset(PortalContext);
2423 :
2424 : /*
2425 : * Save the relation name for a possible error message, to avoid a
2426 : * catalog lookup in case of an error. If any of these return NULL,
2427 : * then the relation has been dropped since last we checked; skip it.
2428 : * Note: they must live in a long-lived memory context because we call
2429 : * vacuum and analyze in different transactions.
2430 : */
2431 :
2432 185676 : tab->at_relname = get_rel_name(tab->at_relid);
2433 185676 : tab->at_nspname = get_namespace_name(get_rel_namespace(tab->at_relid));
2434 185676 : tab->at_datname = get_database_name(MyDatabaseId);
2435 185676 : if (!tab->at_relname || !tab->at_nspname || !tab->at_datname)
2436 0 : goto deleted;
2437 :
2438 : /*
2439 : * We will abort vacuuming the current table if something errors out,
2440 : * and continue with the next one in schedule; in particular, this
2441 : * happens if we are interrupted with SIGINT.
2442 : */
2443 185676 : PG_TRY();
2444 : {
2445 : /* Use PortalContext for any per-table allocations */
2446 185676 : MemoryContextSwitchTo(PortalContext);
2447 :
2448 : /* have at it */
2449 185676 : autovacuum_do_vac_analyze(tab, bstrategy);
2450 :
2451 : /*
2452 : * Clear a possible query-cancel signal, to avoid a late reaction
2453 : * to an automatically-sent signal because of vacuuming the
2454 : * current table (we're done with it, so it would make no sense to
2455 : * cancel at this point.)
2456 : */
2457 185676 : QueryCancelPending = false;
2458 : }
2459 0 : PG_CATCH();
2460 : {
2461 : /*
2462 : * Abort the transaction, start a new one, and proceed with the
2463 : * next table in our list.
2464 : */
2465 0 : HOLD_INTERRUPTS();
2466 0 : if (tab->at_params.options & VACOPT_VACUUM)
2467 0 : errcontext("automatic vacuum of table \"%s.%s.%s\"",
2468 : tab->at_datname, tab->at_nspname, tab->at_relname);
2469 : else
2470 0 : errcontext("automatic analyze of table \"%s.%s.%s\"",
2471 : tab->at_datname, tab->at_nspname, tab->at_relname);
2472 0 : EmitErrorReport();
2473 :
2474 : /* this resets ProcGlobal->statusFlags[i] too */
2475 0 : AbortOutOfAnyTransaction();
2476 0 : FlushErrorState();
2477 0 : MemoryContextReset(PortalContext);
2478 :
2479 : /* restart our transaction for the following operations */
2480 0 : StartTransactionCommand();
2481 0 : RESUME_INTERRUPTS();
2482 : }
2483 185676 : PG_END_TRY();
2484 :
2485 : /* Make sure we're back in AutovacMemCxt */
2486 185676 : MemoryContextSwitchTo(AutovacMemCxt);
2487 :
2488 185676 : did_vacuum = true;
2489 :
2490 : /* ProcGlobal->statusFlags[i] are reset at the next end of xact */
2491 :
2492 : /* be tidy */
2493 185676 : deleted:
2494 185676 : if (tab->at_datname != NULL)
2495 185676 : pfree(tab->at_datname);
2496 185676 : if (tab->at_nspname != NULL)
2497 185676 : pfree(tab->at_nspname);
2498 185676 : if (tab->at_relname != NULL)
2499 185676 : pfree(tab->at_relname);
2500 185676 : pfree(tab);
2501 :
2502 : /*
2503 : * Remove my info from shared memory. We set wi_dobalance on the
2504 : * assumption that we are more likely than not to vacuum a table with
2505 : * no cost-related storage parameters next, so we want to claim our
2506 : * share of I/O as soon as possible to avoid thrashing the global
2507 : * balance.
2508 : */
2509 185676 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2510 185676 : MyWorkerInfo->wi_tableoid = InvalidOid;
2511 185676 : MyWorkerInfo->wi_sharedrel = false;
2512 185676 : LWLockRelease(AutovacuumScheduleLock);
2513 185676 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2514 : }
2515 :
2516 2928 : list_free(table_oids);
2517 :
2518 : /*
2519 : * Perform additional work items, as requested by backends.
2520 : */
2521 2928 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2522 752496 : for (i = 0; i < NUM_WORKITEMS; i++)
2523 : {
2524 749568 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
2525 :
2526 749568 : if (!workitem->avw_used)
2527 749556 : continue;
2528 12 : if (workitem->avw_active)
2529 0 : continue;
2530 12 : if (workitem->avw_database != MyDatabaseId)
2531 0 : continue;
2532 :
2533 : /* claim this one, and release lock while performing it */
2534 12 : workitem->avw_active = true;
2535 12 : LWLockRelease(AutovacuumLock);
2536 :
2537 12 : PushActiveSnapshot(GetTransactionSnapshot());
2538 12 : perform_work_item(workitem);
2539 12 : if (ActiveSnapshotSet()) /* transaction could have aborted */
2540 12 : PopActiveSnapshot();
2541 :
2542 : /*
2543 : * Check for config changes before acquiring lock for further jobs.
2544 : */
2545 12 : CHECK_FOR_INTERRUPTS();
2546 12 : if (ConfigReloadPending)
2547 : {
2548 0 : ConfigReloadPending = false;
2549 0 : ProcessConfigFile(PGC_SIGHUP);
2550 0 : VacuumUpdateCosts();
2551 : }
2552 :
2553 12 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2554 :
2555 : /* and mark it done */
2556 12 : workitem->avw_active = false;
2557 12 : workitem->avw_used = false;
2558 : }
2559 2928 : LWLockRelease(AutovacuumLock);
2560 :
2561 : /*
2562 : * We leak table_toast_map here (among other things), but since we're
2563 : * going away soon, it's not a problem normally. But when using Valgrind,
2564 : * release some stuff to reduce complaints about leaked storage.
2565 : */
2566 : #ifdef USE_VALGRIND
2567 : hash_destroy(table_toast_map);
2568 : FreeTupleDesc(pg_class_desc);
2569 : if (bstrategy)
2570 : pfree(bstrategy);
2571 : #endif
2572 :
2573 : /* Run the rest in xact context, mainly to avoid Valgrind leak warnings */
2574 2928 : MemoryContextSwitchTo(TopTransactionContext);
2575 :
2576 : /*
2577 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We
2578 : * only need to do this once, not after each table.
2579 : *
2580 : * Even if we didn't vacuum anything, it may still be important to do
2581 : * this, because one indirect effect of vac_update_datfrozenxid() is to
2582 : * update TransamVariables->xidVacLimit. That might need to be done even
2583 : * if we haven't vacuumed anything, because relations with older
2584 : * relfrozenxid values or other databases with older datfrozenxid values
2585 : * might have been dropped, allowing xidVacLimit to advance.
2586 : *
2587 : * However, it's also important not to do this blindly in all cases,
2588 : * because when autovacuum=off this will restart the autovacuum launcher.
2589 : * If we're not careful, an infinite loop can result, where workers find
2590 : * no work to do and restart the launcher, which starts another worker in
2591 : * the same database that finds no work to do. To prevent that, we skip
2592 : * this if (1) we found no work to do and (2) we skipped at least one
2593 : * table due to concurrent autovacuum activity. In that case, the other
2594 : * worker has already done it, or will do so when it finishes.
2595 : */
2596 2928 : if (did_vacuum || !found_concurrent_worker)
2597 2928 : vac_update_datfrozenxid();
2598 :
2599 : /* Finally close out the last transaction. */
2600 2928 : CommitTransactionCommand();
2601 2928 : }
2602 :
2603 : /*
2604 : * Execute a previously registered work item.
2605 : */
2606 : static void
2607 12 : perform_work_item(AutoVacuumWorkItem *workitem)
2608 : {
2609 12 : char *cur_datname = NULL;
2610 12 : char *cur_nspname = NULL;
2611 12 : char *cur_relname = NULL;
2612 :
2613 : /*
2614 : * Note we do not store table info in MyWorkerInfo, since this is not
2615 : * vacuuming proper.
2616 : */
2617 :
2618 : /*
2619 : * Save the relation name for a possible error message, to avoid a catalog
2620 : * lookup in case of an error. If any of these return NULL, then the
2621 : * relation has been dropped since last we checked; skip it.
2622 : */
2623 : Assert(CurrentMemoryContext == AutovacMemCxt);
2624 :
2625 12 : cur_relname = get_rel_name(workitem->avw_relation);
2626 12 : cur_nspname = get_namespace_name(get_rel_namespace(workitem->avw_relation));
2627 12 : cur_datname = get_database_name(MyDatabaseId);
2628 12 : if (!cur_relname || !cur_nspname || !cur_datname)
2629 0 : goto deleted2;
2630 :
2631 12 : autovac_report_workitem(workitem, cur_nspname, cur_relname);
2632 :
2633 : /* clean up memory before each work item */
2634 12 : MemoryContextReset(PortalContext);
2635 :
2636 : /*
2637 : * We will abort the current work item if something errors out, and
2638 : * continue with the next one; in particular, this happens if we are
2639 : * interrupted with SIGINT. Note that this means that the work item list
2640 : * can be lossy.
2641 : */
2642 12 : PG_TRY();
2643 : {
2644 : /* Use PortalContext for any per-work-item allocations */
2645 12 : MemoryContextSwitchTo(PortalContext);
2646 :
2647 : /*
2648 : * Have at it. Functions called here are responsible for any required
2649 : * user switch and sandbox.
2650 : */
2651 12 : switch (workitem->avw_type)
2652 : {
2653 12 : case AVW_BRINSummarizeRange:
2654 12 : DirectFunctionCall2(brin_summarize_range,
2655 : ObjectIdGetDatum(workitem->avw_relation),
2656 : Int64GetDatum((int64) workitem->avw_blockNumber));
2657 12 : break;
2658 0 : default:
2659 0 : elog(WARNING, "unrecognized work item found: type %d",
2660 : workitem->avw_type);
2661 0 : break;
2662 : }
2663 :
2664 : /*
2665 : * Clear a possible query-cancel signal, to avoid a late reaction to
2666 : * an automatically-sent signal because of vacuuming the current table
2667 : * (we're done with it, so it would make no sense to cancel at this
2668 : * point.)
2669 : */
2670 12 : QueryCancelPending = false;
2671 : }
2672 0 : PG_CATCH();
2673 : {
2674 : /*
2675 : * Abort the transaction, start a new one, and proceed with the next
2676 : * table in our list.
2677 : */
2678 0 : HOLD_INTERRUPTS();
2679 0 : errcontext("processing work entry for relation \"%s.%s.%s\"",
2680 : cur_datname, cur_nspname, cur_relname);
2681 0 : EmitErrorReport();
2682 :
2683 : /* this resets ProcGlobal->statusFlags[i] too */
2684 0 : AbortOutOfAnyTransaction();
2685 0 : FlushErrorState();
2686 0 : MemoryContextReset(PortalContext);
2687 :
2688 : /* restart our transaction for the following operations */
2689 0 : StartTransactionCommand();
2690 0 : RESUME_INTERRUPTS();
2691 : }
2692 12 : PG_END_TRY();
2693 :
2694 : /* Make sure we're back in AutovacMemCxt */
2695 12 : MemoryContextSwitchTo(AutovacMemCxt);
2696 :
2697 : /* We intentionally do not set did_vacuum here */
2698 :
2699 : /* be tidy */
2700 12 : deleted2:
2701 12 : if (cur_datname)
2702 12 : pfree(cur_datname);
2703 12 : if (cur_nspname)
2704 12 : pfree(cur_nspname);
2705 12 : if (cur_relname)
2706 12 : pfree(cur_relname);
2707 12 : }
2708 :
2709 : /*
2710 : * extract_autovac_opts
2711 : *
2712 : * Given a relation's pg_class tuple, return a palloc'd copy of the
2713 : * AutoVacOpts portion of reloptions, if set; otherwise, return NULL.
2714 : *
2715 : * Note: callers do not have a relation lock on the table at this point,
2716 : * so the table could have been dropped, and its catalog rows gone, after
2717 : * we acquired the pg_class row. If pg_class had a TOAST table, this would
2718 : * be a risk; fortunately, it doesn't.
2719 : */
2720 : static AutoVacOpts *
2721 508204 : extract_autovac_opts(HeapTuple tup, TupleDesc pg_class_desc)
2722 : {
2723 : bytea *relopts;
2724 : AutoVacOpts *av;
2725 :
2726 : Assert(((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_RELATION ||
2727 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_MATVIEW ||
2728 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_TOASTVALUE);
2729 :
2730 508204 : relopts = extractRelOptions(tup, pg_class_desc, NULL);
2731 508204 : if (relopts == NULL)
2732 503632 : return NULL;
2733 :
2734 4572 : av = palloc_object(AutoVacOpts);
2735 4572 : memcpy(av, &(((StdRdOptions *) relopts)->autovacuum), sizeof(AutoVacOpts));
2736 4572 : pfree(relopts);
2737 :
2738 4572 : return av;
2739 : }
2740 :
2741 :
2742 : /*
2743 : * table_recheck_autovac
2744 : *
2745 : * Recheck whether a table still needs vacuum or analyze. Return value is a
2746 : * valid autovac_table pointer if it does, NULL otherwise.
2747 : *
2748 : * Note that the returned autovac_table does not have the name fields set.
2749 : */
2750 : static autovac_table *
2751 185742 : table_recheck_autovac(Oid relid, HTAB *table_toast_map,
2752 : TupleDesc pg_class_desc,
2753 : int effective_multixact_freeze_max_age)
2754 : {
2755 : Form_pg_class classForm;
2756 : HeapTuple classTup;
2757 : bool dovacuum;
2758 : bool doanalyze;
2759 185742 : autovac_table *tab = NULL;
2760 : bool wraparound;
2761 : AutoVacOpts *avopts;
2762 185742 : bool free_avopts = false;
2763 :
2764 : /* fetch the relation's relcache entry */
2765 185742 : classTup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2766 185742 : if (!HeapTupleIsValid(classTup))
2767 0 : return NULL;
2768 185742 : classForm = (Form_pg_class) GETSTRUCT(classTup);
2769 :
2770 : /*
2771 : * Get the applicable reloptions. If it is a TOAST table, try to get the
2772 : * main table reloptions if the toast table itself doesn't have.
2773 : */
2774 185742 : avopts = extract_autovac_opts(classTup, pg_class_desc);
2775 185742 : if (avopts)
2776 1984 : free_avopts = true;
2777 183758 : else if (classForm->relkind == RELKIND_TOASTVALUE &&
2778 : table_toast_map != NULL)
2779 : {
2780 : av_relation *hentry;
2781 : bool found;
2782 :
2783 67660 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2784 67660 : if (found && hentry->ar_hasrelopts)
2785 1970 : avopts = &hentry->ar_reloptions;
2786 : }
2787 :
2788 185742 : recheck_relation_needs_vacanalyze(relid, avopts, classForm,
2789 : effective_multixact_freeze_max_age,
2790 : &dovacuum, &doanalyze, &wraparound);
2791 :
2792 : /* OK, it needs something done */
2793 185742 : if (doanalyze || dovacuum)
2794 : {
2795 : int freeze_min_age;
2796 : int freeze_table_age;
2797 : int multixact_freeze_min_age;
2798 : int multixact_freeze_table_age;
2799 : int log_vacuum_min_duration;
2800 : int log_analyze_min_duration;
2801 :
2802 : /*
2803 : * Calculate the vacuum cost parameters and the freeze ages. If there
2804 : * are options set in pg_class.reloptions, use them; in the case of a
2805 : * toast table, try the main table too. Otherwise use the GUC
2806 : * defaults, autovacuum's own first and plain vacuum second.
2807 : */
2808 :
2809 : /* -1 in autovac setting means use log_autovacuum_min_duration */
2810 3954 : log_vacuum_min_duration = (avopts && avopts->log_vacuum_min_duration >= 0)
2811 : ? avopts->log_vacuum_min_duration
2812 189630 : : Log_autovacuum_min_duration;
2813 :
2814 : /* -1 in autovac setting means use log_autoanalyze_min_duration */
2815 3954 : log_analyze_min_duration = (avopts && avopts->log_analyze_min_duration >= 0)
2816 : ? avopts->log_analyze_min_duration
2817 189630 : : Log_autoanalyze_min_duration;
2818 :
2819 : /* these do not have autovacuum-specific settings */
2820 3954 : freeze_min_age = (avopts && avopts->freeze_min_age >= 0)
2821 : ? avopts->freeze_min_age
2822 189630 : : default_freeze_min_age;
2823 :
2824 3954 : freeze_table_age = (avopts && avopts->freeze_table_age >= 0)
2825 : ? avopts->freeze_table_age
2826 189630 : : default_freeze_table_age;
2827 :
2828 189630 : multixact_freeze_min_age = (avopts &&
2829 3954 : avopts->multixact_freeze_min_age >= 0)
2830 : ? avopts->multixact_freeze_min_age
2831 189630 : : default_multixact_freeze_min_age;
2832 :
2833 189630 : multixact_freeze_table_age = (avopts &&
2834 3954 : avopts->multixact_freeze_table_age >= 0)
2835 : ? avopts->multixact_freeze_table_age
2836 189630 : : default_multixact_freeze_table_age;
2837 :
2838 185676 : tab = palloc_object(autovac_table);
2839 185676 : tab->at_relid = relid;
2840 185676 : tab->at_sharedrel = classForm->relisshared;
2841 :
2842 : /*
2843 : * Select VACUUM options. Note we don't say VACOPT_PROCESS_TOAST, so
2844 : * that vacuum() skips toast relations. Also note we tell vacuum() to
2845 : * skip vac_update_datfrozenxid(); we'll do that separately.
2846 : */
2847 185676 : tab->at_params.options =
2848 185676 : (dovacuum ? (VACOPT_VACUUM |
2849 : VACOPT_PROCESS_MAIN |
2850 185676 : VACOPT_SKIP_DATABASE_STATS) : 0) |
2851 185676 : (doanalyze ? VACOPT_ANALYZE : 0) |
2852 185676 : (!wraparound ? VACOPT_SKIP_LOCKED : 0);
2853 :
2854 : /*
2855 : * index_cleanup and truncate are unspecified at first in autovacuum.
2856 : * They will be filled in with usable values using their reloptions
2857 : * (or reloption defaults) later.
2858 : */
2859 185676 : tab->at_params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
2860 185676 : tab->at_params.truncate = VACOPTVALUE_UNSPECIFIED;
2861 : /* As of now, we don't support parallel vacuum for autovacuum */
2862 185676 : tab->at_params.nworkers = -1;
2863 185676 : tab->at_params.freeze_min_age = freeze_min_age;
2864 185676 : tab->at_params.freeze_table_age = freeze_table_age;
2865 185676 : tab->at_params.multixact_freeze_min_age = multixact_freeze_min_age;
2866 185676 : tab->at_params.multixact_freeze_table_age = multixact_freeze_table_age;
2867 185676 : tab->at_params.is_wraparound = wraparound;
2868 185676 : tab->at_params.log_vacuum_min_duration = log_vacuum_min_duration;
2869 185676 : tab->at_params.log_analyze_min_duration = log_analyze_min_duration;
2870 185676 : tab->at_params.toast_parent = InvalidOid;
2871 :
2872 : /*
2873 : * Later, in vacuum_rel(), we check reloptions for any
2874 : * vacuum_max_eager_freeze_failure_rate override.
2875 : */
2876 185676 : tab->at_params.max_eager_freeze_failure_rate = vacuum_max_eager_freeze_failure_rate;
2877 185676 : tab->at_storage_param_vac_cost_limit = avopts ?
2878 185676 : avopts->vacuum_cost_limit : 0;
2879 185676 : tab->at_storage_param_vac_cost_delay = avopts ?
2880 185676 : avopts->vacuum_cost_delay : -1;
2881 185676 : tab->at_relname = NULL;
2882 185676 : tab->at_nspname = NULL;
2883 185676 : tab->at_datname = NULL;
2884 :
2885 : /*
2886 : * If any of the cost delay parameters has been set individually for
2887 : * this table, disable the balancing algorithm.
2888 : */
2889 185676 : tab->at_dobalance =
2890 189630 : !(avopts && (avopts->vacuum_cost_limit > 0 ||
2891 3954 : avopts->vacuum_cost_delay >= 0));
2892 : }
2893 :
2894 185742 : if (free_avopts)
2895 1984 : pfree(avopts);
2896 185742 : heap_freetuple(classTup);
2897 185742 : return tab;
2898 : }
2899 :
2900 : /*
2901 : * recheck_relation_needs_vacanalyze
2902 : *
2903 : * Subroutine for table_recheck_autovac.
2904 : *
2905 : * Fetch the pgstat of a relation and recheck whether a relation
2906 : * needs to be vacuumed or analyzed.
2907 : */
2908 : static void
2909 185742 : recheck_relation_needs_vacanalyze(Oid relid,
2910 : AutoVacOpts *avopts,
2911 : Form_pg_class classForm,
2912 : int effective_multixact_freeze_max_age,
2913 : bool *dovacuum,
2914 : bool *doanalyze,
2915 : bool *wraparound)
2916 : {
2917 : PgStat_StatTabEntry *tabentry;
2918 :
2919 : /* fetch the pgstat table entry */
2920 185742 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2921 : relid);
2922 :
2923 185742 : relation_needs_vacanalyze(relid, avopts, classForm, tabentry,
2924 : effective_multixact_freeze_max_age,
2925 : dovacuum, doanalyze, wraparound);
2926 :
2927 : /* Release tabentry to avoid leakage */
2928 185742 : if (tabentry)
2929 184816 : pfree(tabentry);
2930 :
2931 : /* ignore ANALYZE for toast tables */
2932 185742 : if (classForm->relkind == RELKIND_TOASTVALUE)
2933 67660 : *doanalyze = false;
2934 185742 : }
2935 :
2936 : /*
2937 : * relation_needs_vacanalyze
2938 : *
2939 : * Check whether a relation needs to be vacuumed or analyzed; return each into
2940 : * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is
2941 : * being forced because of Xid or multixact wraparound.
2942 : *
2943 : * relopts is a pointer to the AutoVacOpts options (either for itself in the
2944 : * case of a plain table, or for either itself or its parent table in the case
2945 : * of a TOAST table), NULL if none; tabentry is the pgstats entry, which can be
2946 : * NULL.
2947 : *
2948 : * A table needs to be vacuumed if the number of dead tuples exceeds a
2949 : * threshold. This threshold is calculated as
2950 : *
2951 : * threshold = vac_base_thresh + vac_scale_factor * reltuples
2952 : * if (threshold > vac_max_thresh)
2953 : * threshold = vac_max_thresh;
2954 : *
2955 : * For analyze, the analysis done is that the number of tuples inserted,
2956 : * deleted and updated since the last analyze exceeds a threshold calculated
2957 : * in the same fashion as above. Note that the cumulative stats system stores
2958 : * the number of tuples (both live and dead) that there were as of the last
2959 : * analyze. This is asymmetric to the VACUUM case.
2960 : *
2961 : * We also force vacuum if the table's relfrozenxid is more than freeze_max_age
2962 : * transactions back, and if its relminmxid is more than
2963 : * multixact_freeze_max_age multixacts back.
2964 : *
2965 : * A table whose autovacuum_enabled option is false is
2966 : * automatically skipped (unless we have to vacuum it due to freeze_max_age).
2967 : * Thus autovacuum can be disabled for specific tables. Also, when the cumulative
2968 : * stats system does not have data about a table, it will be skipped.
2969 : *
2970 : * A table whose vac_base_thresh value is < 0 takes the base value from the
2971 : * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor
2972 : * value < 0 is substituted with the value of
2973 : * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze.
2974 : */
2975 : static void
2976 508204 : relation_needs_vacanalyze(Oid relid,
2977 : AutoVacOpts *relopts,
2978 : Form_pg_class classForm,
2979 : PgStat_StatTabEntry *tabentry,
2980 : int effective_multixact_freeze_max_age,
2981 : /* output params below */
2982 : bool *dovacuum,
2983 : bool *doanalyze,
2984 : bool *wraparound)
2985 : {
2986 : bool force_vacuum;
2987 : bool av_enabled;
2988 :
2989 : /* constants from reloptions or GUC variables */
2990 : int vac_base_thresh,
2991 : vac_max_thresh,
2992 : vac_ins_base_thresh,
2993 : anl_base_thresh;
2994 : float4 vac_scale_factor,
2995 : vac_ins_scale_factor,
2996 : anl_scale_factor;
2997 :
2998 : /* thresholds calculated from above constants */
2999 : float4 vacthresh,
3000 : vacinsthresh,
3001 : anlthresh;
3002 :
3003 : /* number of vacuum (resp. analyze) tuples at this time */
3004 : float4 vactuples,
3005 : instuples,
3006 : anltuples;
3007 :
3008 : /* freeze parameters */
3009 : int freeze_max_age;
3010 : int multixact_freeze_max_age;
3011 : TransactionId xidForceLimit;
3012 : TransactionId relfrozenxid;
3013 : MultiXactId multiForceLimit;
3014 :
3015 : Assert(classForm != NULL);
3016 : Assert(OidIsValid(relid));
3017 :
3018 : /*
3019 : * Determine vacuum/analyze equation parameters. We have two possible
3020 : * sources: the passed reloptions (which could be a main table or a toast
3021 : * table), or the autovacuum GUC variables.
3022 : */
3023 :
3024 : /* -1 in autovac setting means use plain vacuum_scale_factor */
3025 9068 : vac_scale_factor = (relopts && relopts->vacuum_scale_factor >= 0)
3026 0 : ? relopts->vacuum_scale_factor
3027 517272 : : autovacuum_vac_scale;
3028 :
3029 9068 : vac_base_thresh = (relopts && relopts->vacuum_threshold >= 0)
3030 : ? relopts->vacuum_threshold
3031 517272 : : autovacuum_vac_thresh;
3032 :
3033 : /* -1 is used to disable max threshold */
3034 9068 : vac_max_thresh = (relopts && relopts->vacuum_max_threshold >= -1)
3035 : ? relopts->vacuum_max_threshold
3036 517272 : : autovacuum_vac_max_thresh;
3037 :
3038 9068 : vac_ins_scale_factor = (relopts && relopts->vacuum_ins_scale_factor >= 0)
3039 0 : ? relopts->vacuum_ins_scale_factor
3040 517272 : : autovacuum_vac_ins_scale;
3041 :
3042 : /* -1 is used to disable insert vacuums */
3043 9068 : vac_ins_base_thresh = (relopts && relopts->vacuum_ins_threshold >= -1)
3044 : ? relopts->vacuum_ins_threshold
3045 517272 : : autovacuum_vac_ins_thresh;
3046 :
3047 9068 : anl_scale_factor = (relopts && relopts->analyze_scale_factor >= 0)
3048 0 : ? relopts->analyze_scale_factor
3049 517272 : : autovacuum_anl_scale;
3050 :
3051 9068 : anl_base_thresh = (relopts && relopts->analyze_threshold >= 0)
3052 : ? relopts->analyze_threshold
3053 517272 : : autovacuum_anl_thresh;
3054 :
3055 9068 : freeze_max_age = (relopts && relopts->freeze_max_age >= 0)
3056 0 : ? Min(relopts->freeze_max_age, autovacuum_freeze_max_age)
3057 517272 : : autovacuum_freeze_max_age;
3058 :
3059 9068 : multixact_freeze_max_age = (relopts && relopts->multixact_freeze_max_age >= 0)
3060 0 : ? Min(relopts->multixact_freeze_max_age, effective_multixact_freeze_max_age)
3061 517272 : : effective_multixact_freeze_max_age;
3062 :
3063 508204 : av_enabled = (relopts ? relopts->enabled : true);
3064 :
3065 : /* Force vacuum if table is at risk of wraparound */
3066 508204 : xidForceLimit = recentXid - freeze_max_age;
3067 508204 : if (xidForceLimit < FirstNormalTransactionId)
3068 0 : xidForceLimit -= FirstNormalTransactionId;
3069 508204 : relfrozenxid = classForm->relfrozenxid;
3070 1016408 : force_vacuum = (TransactionIdIsNormal(relfrozenxid) &&
3071 508204 : TransactionIdPrecedes(relfrozenxid, xidForceLimit));
3072 508204 : if (!force_vacuum)
3073 : {
3074 137054 : MultiXactId relminmxid = classForm->relminmxid;
3075 :
3076 137054 : multiForceLimit = recentMulti - multixact_freeze_max_age;
3077 137054 : if (multiForceLimit < FirstMultiXactId)
3078 0 : multiForceLimit -= FirstMultiXactId;
3079 274108 : force_vacuum = MultiXactIdIsValid(relminmxid) &&
3080 137054 : MultiXactIdPrecedes(relminmxid, multiForceLimit);
3081 : }
3082 508204 : *wraparound = force_vacuum;
3083 :
3084 : /* User disabled it in pg_class.reloptions? (But ignore if at risk) */
3085 508204 : if (!av_enabled && !force_vacuum)
3086 : {
3087 1096 : *doanalyze = false;
3088 1096 : *dovacuum = false;
3089 1096 : return;
3090 : }
3091 :
3092 : /*
3093 : * If we found stats for the table, and autovacuum is currently enabled,
3094 : * make a threshold-based decision whether to vacuum and/or analyze. If
3095 : * autovacuum is currently disabled, we must be here for anti-wraparound
3096 : * vacuuming only, so don't vacuum (or analyze) anything that's not being
3097 : * forced.
3098 : */
3099 507108 : if (tabentry && AutoVacuumingActive())
3100 499348 : {
3101 499348 : float4 pcnt_unfrozen = 1;
3102 499348 : float4 reltuples = classForm->reltuples;
3103 499348 : int32 relpages = classForm->relpages;
3104 499348 : int32 relallfrozen = classForm->relallfrozen;
3105 :
3106 499348 : vactuples = tabentry->dead_tuples;
3107 499348 : instuples = tabentry->ins_since_vacuum;
3108 499348 : anltuples = tabentry->mod_since_analyze;
3109 :
3110 : /* If the table hasn't yet been vacuumed, take reltuples as zero */
3111 499348 : if (reltuples < 0)
3112 1950 : reltuples = 0;
3113 :
3114 : /*
3115 : * If we have data for relallfrozen, calculate the unfrozen percentage
3116 : * of the table to modify insert scale factor. This helps us decide
3117 : * whether or not to vacuum an insert-heavy table based on the number
3118 : * of inserts to the more "active" part of the table.
3119 : */
3120 499348 : if (relpages > 0 && relallfrozen > 0)
3121 : {
3122 : /*
3123 : * It could be the stats were updated manually and relallfrozen >
3124 : * relpages. Clamp relallfrozen to relpages to avoid nonsensical
3125 : * calculations.
3126 : */
3127 193660 : relallfrozen = Min(relallfrozen, relpages);
3128 193660 : pcnt_unfrozen = 1 - ((float4) relallfrozen / relpages);
3129 : }
3130 :
3131 499348 : vacthresh = (float4) vac_base_thresh + vac_scale_factor * reltuples;
3132 499348 : if (vac_max_thresh >= 0 && vacthresh > (float4) vac_max_thresh)
3133 0 : vacthresh = (float4) vac_max_thresh;
3134 :
3135 499348 : vacinsthresh = (float4) vac_ins_base_thresh +
3136 499348 : vac_ins_scale_factor * reltuples * pcnt_unfrozen;
3137 499348 : anlthresh = (float4) anl_base_thresh + anl_scale_factor * reltuples;
3138 :
3139 499348 : if (vac_ins_base_thresh >= 0)
3140 499348 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: %.0f (threshold %.0f), anl: %.0f (threshold %.0f)",
3141 : NameStr(classForm->relname),
3142 : vactuples, vacthresh, instuples, vacinsthresh, anltuples, anlthresh);
3143 : else
3144 0 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: (disabled), anl: %.0f (threshold %.0f)",
3145 : NameStr(classForm->relname),
3146 : vactuples, vacthresh, anltuples, anlthresh);
3147 :
3148 : /* Determine if this table needs vacuum or analyze. */
3149 628990 : *dovacuum = force_vacuum || (vactuples > vacthresh) ||
3150 129642 : (vac_ins_base_thresh >= 0 && instuples > vacinsthresh);
3151 499348 : *doanalyze = (anltuples > anlthresh);
3152 : }
3153 : else
3154 : {
3155 : /*
3156 : * Skip a table not found in stat hash, unless we have to force vacuum
3157 : * for anti-wrap purposes. If it's not acted upon, there's no need to
3158 : * vacuum it.
3159 : */
3160 7760 : *dovacuum = force_vacuum;
3161 7760 : *doanalyze = false;
3162 : }
3163 :
3164 : /* ANALYZE refuses to work with pg_statistic */
3165 507108 : if (relid == StatisticRelationId)
3166 4638 : *doanalyze = false;
3167 : }
3168 :
3169 : /*
3170 : * autovacuum_do_vac_analyze
3171 : * Vacuum and/or analyze the specified table
3172 : *
3173 : * We expect the caller to have switched into a memory context that won't
3174 : * disappear at transaction commit.
3175 : */
3176 : static void
3177 185676 : autovacuum_do_vac_analyze(autovac_table *tab, BufferAccessStrategy bstrategy)
3178 : {
3179 : RangeVar *rangevar;
3180 : VacuumRelation *rel;
3181 : List *rel_list;
3182 : MemoryContext vac_context;
3183 : MemoryContext old_context;
3184 :
3185 : /* Let pgstat know what we're doing */
3186 185676 : autovac_report_activity(tab);
3187 :
3188 : /* Create a context that vacuum() can use as cross-transaction storage */
3189 185676 : vac_context = AllocSetContextCreate(CurrentMemoryContext,
3190 : "Vacuum",
3191 : ALLOCSET_DEFAULT_SIZES);
3192 :
3193 : /* Set up one VacuumRelation target, identified by OID, for vacuum() */
3194 185676 : old_context = MemoryContextSwitchTo(vac_context);
3195 185676 : rangevar = makeRangeVar(tab->at_nspname, tab->at_relname, -1);
3196 185676 : rel = makeVacuumRelation(rangevar, tab->at_relid, NIL);
3197 185676 : rel_list = list_make1(rel);
3198 185676 : MemoryContextSwitchTo(old_context);
3199 :
3200 185676 : vacuum(rel_list, tab->at_params, bstrategy, vac_context, true);
3201 :
3202 185676 : MemoryContextDelete(vac_context);
3203 185676 : }
3204 :
3205 : /*
3206 : * autovac_report_activity
3207 : * Report to pgstat what autovacuum is doing
3208 : *
3209 : * We send a SQL string corresponding to what the user would see if the
3210 : * equivalent command was to be issued manually.
3211 : *
3212 : * Note we assume that we are going to report the next command as soon as we're
3213 : * done with the current one, and exit right after the last one, so we don't
3214 : * bother to report "<IDLE>" or some such.
3215 : */
3216 : static void
3217 185676 : autovac_report_activity(autovac_table *tab)
3218 : {
3219 : #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56)
3220 : char activity[MAX_AUTOVAC_ACTIV_LEN];
3221 : int len;
3222 :
3223 : /* Report the command and possible options */
3224 185676 : if (tab->at_params.options & VACOPT_VACUUM)
3225 185320 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3226 : "autovacuum: VACUUM%s",
3227 185320 : tab->at_params.options & VACOPT_ANALYZE ? " ANALYZE" : "");
3228 : else
3229 356 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3230 : "autovacuum: ANALYZE");
3231 :
3232 : /*
3233 : * Report the qualified name of the relation.
3234 : */
3235 185676 : len = strlen(activity);
3236 :
3237 185676 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3238 : " %s.%s%s", tab->at_nspname, tab->at_relname,
3239 185676 : tab->at_params.is_wraparound ? " (to prevent wraparound)" : "");
3240 :
3241 : /* Set statement_timestamp() to current time for pg_stat_activity */
3242 185676 : SetCurrentStatementStartTimestamp();
3243 :
3244 185676 : pgstat_report_activity(STATE_RUNNING, activity);
3245 185676 : }
3246 :
3247 : /*
3248 : * autovac_report_workitem
3249 : * Report to pgstat that autovacuum is processing a work item
3250 : */
3251 : static void
3252 12 : autovac_report_workitem(AutoVacuumWorkItem *workitem,
3253 : const char *nspname, const char *relname)
3254 : {
3255 : char activity[MAX_AUTOVAC_ACTIV_LEN + 12 + 2];
3256 : char blk[12 + 2];
3257 : int len;
3258 :
3259 12 : switch (workitem->avw_type)
3260 : {
3261 12 : case AVW_BRINSummarizeRange:
3262 12 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3263 : "autovacuum: BRIN summarize");
3264 12 : break;
3265 : }
3266 :
3267 : /*
3268 : * Report the qualified name of the relation, and the block number if any
3269 : */
3270 12 : len = strlen(activity);
3271 :
3272 12 : if (BlockNumberIsValid(workitem->avw_blockNumber))
3273 12 : snprintf(blk, sizeof(blk), " %u", workitem->avw_blockNumber);
3274 : else
3275 0 : blk[0] = '\0';
3276 :
3277 12 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3278 : " %s.%s%s", nspname, relname, blk);
3279 :
3280 : /* Set statement_timestamp() to current time for pg_stat_activity */
3281 12 : SetCurrentStatementStartTimestamp();
3282 :
3283 12 : pgstat_report_activity(STATE_RUNNING, activity);
3284 12 : }
3285 :
3286 : /*
3287 : * AutoVacuumingActive
3288 : * Check GUC vars and report whether the autovacuum process should be
3289 : * running.
3290 : */
3291 : bool
3292 578562 : AutoVacuumingActive(void)
3293 : {
3294 578562 : if (!autovacuum_start_daemon || !pgstat_track_counts)
3295 10210 : return false;
3296 568352 : return true;
3297 : }
3298 :
3299 : /*
3300 : * Request one work item to the next autovacuum run processing our database.
3301 : * Return false if the request can't be recorded.
3302 : */
3303 : bool
3304 12 : AutoVacuumRequestWork(AutoVacuumWorkItemType type, Oid relationId,
3305 : BlockNumber blkno)
3306 : {
3307 : int i;
3308 12 : bool result = false;
3309 :
3310 12 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
3311 :
3312 : /*
3313 : * Locate an unused work item and fill it with the given data.
3314 : */
3315 42 : for (i = 0; i < NUM_WORKITEMS; i++)
3316 : {
3317 42 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
3318 :
3319 42 : if (workitem->avw_used)
3320 30 : continue;
3321 :
3322 12 : workitem->avw_used = true;
3323 12 : workitem->avw_active = false;
3324 12 : workitem->avw_type = type;
3325 12 : workitem->avw_database = MyDatabaseId;
3326 12 : workitem->avw_relation = relationId;
3327 12 : workitem->avw_blockNumber = blkno;
3328 12 : result = true;
3329 :
3330 : /* done */
3331 12 : break;
3332 : }
3333 :
3334 12 : LWLockRelease(AutovacuumLock);
3335 :
3336 12 : return result;
3337 : }
3338 :
3339 : /*
3340 : * autovac_init
3341 : * This is called at postmaster initialization.
3342 : *
3343 : * All we do here is annoy the user if he got it wrong.
3344 : */
3345 : void
3346 1822 : autovac_init(void)
3347 : {
3348 1822 : if (!autovacuum_start_daemon)
3349 244 : return;
3350 1578 : else if (!pgstat_track_counts)
3351 0 : ereport(WARNING,
3352 : (errmsg("autovacuum not started because of misconfiguration"),
3353 : errhint("Enable the \"track_counts\" option.")));
3354 : else
3355 1578 : check_av_worker_gucs();
3356 : }
3357 :
3358 : /*
3359 : * AutoVacuumShmemSize
3360 : * Compute space needed for autovacuum-related shared memory
3361 : */
3362 : Size
3363 6528 : AutoVacuumShmemSize(void)
3364 : {
3365 : Size size;
3366 :
3367 : /*
3368 : * Need the fixed struct and the array of WorkerInfoData.
3369 : */
3370 6528 : size = sizeof(AutoVacuumShmemStruct);
3371 6528 : size = MAXALIGN(size);
3372 6528 : size = add_size(size, mul_size(autovacuum_worker_slots,
3373 : sizeof(WorkerInfoData)));
3374 6528 : return size;
3375 : }
3376 :
3377 : /*
3378 : * AutoVacuumShmemInit
3379 : * Allocate and initialize autovacuum-related shared memory
3380 : */
3381 : void
3382 2278 : AutoVacuumShmemInit(void)
3383 : {
3384 : bool found;
3385 :
3386 2278 : AutoVacuumShmem = (AutoVacuumShmemStruct *)
3387 2278 : ShmemInitStruct("AutoVacuum Data",
3388 : AutoVacuumShmemSize(),
3389 : &found);
3390 :
3391 2278 : if (!IsUnderPostmaster)
3392 : {
3393 : WorkerInfo worker;
3394 : int i;
3395 :
3396 : Assert(!found);
3397 :
3398 2278 : AutoVacuumShmem->av_launcherpid = 0;
3399 2278 : dclist_init(&AutoVacuumShmem->av_freeWorkers);
3400 2278 : dlist_init(&AutoVacuumShmem->av_runningWorkers);
3401 2278 : AutoVacuumShmem->av_startingWorker = NULL;
3402 2278 : memset(AutoVacuumShmem->av_workItems, 0,
3403 : sizeof(AutoVacuumWorkItem) * NUM_WORKITEMS);
3404 :
3405 2278 : worker = (WorkerInfo) ((char *) AutoVacuumShmem +
3406 : MAXALIGN(sizeof(AutoVacuumShmemStruct)));
3407 :
3408 : /* initialize the WorkerInfo free list */
3409 27182 : for (i = 0; i < autovacuum_worker_slots; i++)
3410 : {
3411 24904 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
3412 24904 : &worker[i].wi_links);
3413 24904 : pg_atomic_init_flag(&worker[i].wi_dobalance);
3414 : }
3415 :
3416 2278 : pg_atomic_init_u32(&AutoVacuumShmem->av_nworkersForBalance, 0);
3417 :
3418 : }
3419 : else
3420 : Assert(found);
3421 2278 : }
3422 :
3423 : /*
3424 : * GUC check_hook for autovacuum_work_mem
3425 : */
3426 : bool
3427 2356 : check_autovacuum_work_mem(int *newval, void **extra, GucSource source)
3428 : {
3429 : /*
3430 : * -1 indicates fallback.
3431 : *
3432 : * If we haven't yet changed the boot_val default of -1, just let it be.
3433 : * Autovacuum will look to maintenance_work_mem instead.
3434 : */
3435 2356 : if (*newval == -1)
3436 2352 : return true;
3437 :
3438 : /*
3439 : * We clamp manually-set values to at least 64kB. Since
3440 : * maintenance_work_mem is always set to at least this value, do the same
3441 : * here.
3442 : */
3443 4 : if (*newval < 64)
3444 4 : *newval = 64;
3445 :
3446 4 : return true;
3447 : }
3448 :
3449 : /*
3450 : * Returns whether there is a free autovacuum worker slot available.
3451 : */
3452 : static bool
3453 23738 : av_worker_available(void)
3454 : {
3455 : int free_slots;
3456 : int reserved_slots;
3457 :
3458 23738 : free_slots = dclist_count(&AutoVacuumShmem->av_freeWorkers);
3459 :
3460 23738 : reserved_slots = autovacuum_worker_slots - autovacuum_max_workers;
3461 23738 : reserved_slots = Max(0, reserved_slots);
3462 :
3463 23738 : return free_slots > reserved_slots;
3464 : }
3465 :
3466 : /*
3467 : * Emits a WARNING if autovacuum_worker_slots < autovacuum_max_workers.
3468 : */
3469 : static void
3470 1578 : check_av_worker_gucs(void)
3471 : {
3472 1578 : if (autovacuum_worker_slots < autovacuum_max_workers)
3473 0 : ereport(WARNING,
3474 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3475 : errmsg("\"autovacuum_max_workers\" (%d) should be less than or equal to \"autovacuum_worker_slots\" (%d)",
3476 : autovacuum_max_workers, autovacuum_worker_slots),
3477 : errdetail("The server will only start up to \"autovacuum_worker_slots\" (%d) autovacuum workers at a given time.",
3478 : autovacuum_worker_slots)));
3479 1578 : }
|
