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 <math.h>
66 : #include <signal.h>
67 : #include <sys/time.h>
68 : #include <unistd.h>
69 :
70 : #include "access/heapam.h"
71 : #include "access/htup_details.h"
72 : #include "access/multixact.h"
73 : #include "access/reloptions.h"
74 : #include "access/tableam.h"
75 : #include "access/transam.h"
76 : #include "access/xact.h"
77 : #include "catalog/dependency.h"
78 : #include "catalog/namespace.h"
79 : #include "catalog/pg_database.h"
80 : #include "catalog/pg_namespace.h"
81 : #include "commands/vacuum.h"
82 : #include "common/int.h"
83 : #include "lib/ilist.h"
84 : #include "libpq/pqsignal.h"
85 : #include "miscadmin.h"
86 : #include "nodes/makefuncs.h"
87 : #include "pgstat.h"
88 : #include "postmaster/autovacuum.h"
89 : #include "postmaster/interrupt.h"
90 : #include "postmaster/postmaster.h"
91 : #include "storage/aio_subsys.h"
92 : #include "storage/bufmgr.h"
93 : #include "storage/ipc.h"
94 : #include "storage/fd.h"
95 : #include "storage/latch.h"
96 : #include "storage/lmgr.h"
97 : #include "storage/pmsignal.h"
98 : #include "storage/proc.h"
99 : #include "storage/procsignal.h"
100 : #include "storage/smgr.h"
101 : #include "tcop/tcopprot.h"
102 : #include "utils/fmgroids.h"
103 : #include "utils/fmgrprotos.h"
104 : #include "utils/guc_hooks.h"
105 : #include "utils/injection_point.h"
106 : #include "utils/lsyscache.h"
107 : #include "utils/memutils.h"
108 : #include "utils/ps_status.h"
109 : #include "utils/rel.h"
110 : #include "utils/snapmgr.h"
111 : #include "utils/syscache.h"
112 : #include "utils/timeout.h"
113 : #include "utils/timestamp.h"
114 : #include "utils/wait_event.h"
115 :
116 :
117 : /*
118 : * GUC parameters
119 : */
120 : bool autovacuum_start_daemon = false;
121 : int autovacuum_worker_slots;
122 : int autovacuum_max_workers;
123 : int autovacuum_work_mem = -1;
124 : int autovacuum_naptime;
125 : int autovacuum_vac_thresh;
126 : int autovacuum_vac_max_thresh;
127 : double autovacuum_vac_scale;
128 : int autovacuum_vac_ins_thresh;
129 : double autovacuum_vac_ins_scale;
130 : int autovacuum_anl_thresh;
131 : double autovacuum_anl_scale;
132 : int autovacuum_freeze_max_age;
133 : int autovacuum_multixact_freeze_max_age;
134 : double autovacuum_freeze_score_weight = 1.0;
135 : double autovacuum_multixact_freeze_score_weight = 1.0;
136 : double autovacuum_vacuum_score_weight = 1.0;
137 : double autovacuum_vacuum_insert_score_weight = 1.0;
138 : double autovacuum_analyze_score_weight = 1.0;
139 : double autovacuum_vac_cost_delay;
140 : int autovacuum_vac_cost_limit;
141 :
142 : int Log_autovacuum_min_duration = 600000;
143 : int Log_autoanalyze_min_duration = 600000;
144 :
145 : /* the minimum allowed time between two awakenings of the launcher */
146 : #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */
147 : #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */
148 :
149 : /*
150 : * Variables to save the cost-related storage parameters for the current
151 : * relation being vacuumed by this autovacuum worker. Using these, we can
152 : * ensure we don't overwrite the values of vacuum_cost_delay and
153 : * vacuum_cost_limit after reloading the configuration file. They are
154 : * initialized to "invalid" values to indicate that no cost-related storage
155 : * parameters were specified and will be set in do_autovacuum() after checking
156 : * the storage parameters in table_recheck_autovac().
157 : */
158 : static double av_storage_param_cost_delay = -1;
159 : static int av_storage_param_cost_limit = -1;
160 :
161 : /* Flags set by signal handlers */
162 : static volatile sig_atomic_t got_SIGUSR2 = false;
163 :
164 : /* Comparison points for determining whether freeze_max_age is exceeded */
165 : static TransactionId recentXid;
166 : static MultiXactId recentMulti;
167 :
168 : /* Default freeze ages to use for autovacuum (varies by database) */
169 : static int default_freeze_min_age;
170 : static int default_freeze_table_age;
171 : static int default_multixact_freeze_min_age;
172 : static int default_multixact_freeze_table_age;
173 :
174 : /* Memory context for long-lived data */
175 : static MemoryContext AutovacMemCxt;
176 :
177 : /* struct to keep track of databases in launcher */
178 : typedef struct avl_dbase
179 : {
180 : Oid adl_datid; /* hash key -- must be first */
181 : TimestampTz adl_next_worker;
182 : int adl_score;
183 : dlist_node adl_node;
184 : } avl_dbase;
185 :
186 : /* struct to keep track of databases in worker */
187 : typedef struct avw_dbase
188 : {
189 : Oid adw_datid;
190 : char *adw_name;
191 : TransactionId adw_frozenxid;
192 : MultiXactId adw_minmulti;
193 : PgStat_StatDBEntry *adw_entry;
194 : } avw_dbase;
195 :
196 : /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */
197 : typedef struct av_relation
198 : {
199 : Oid ar_toastrelid; /* hash key - must be first */
200 : Oid ar_relid;
201 : bool ar_hasrelopts;
202 : AutoVacOpts ar_reloptions; /* copy of AutoVacOpts from the main table's
203 : * reloptions, or NULL if none */
204 : } av_relation;
205 :
206 : /* struct to keep track of tables to vacuum and/or analyze, after rechecking */
207 : typedef struct autovac_table
208 : {
209 : Oid at_relid;
210 : VacuumParams at_params;
211 : double at_storage_param_vac_cost_delay;
212 : int at_storage_param_vac_cost_limit;
213 : bool at_dobalance;
214 : char *at_relname;
215 : char *at_nspname;
216 : char *at_datname;
217 : } autovac_table;
218 :
219 : /*-------------
220 : * This struct holds information about a single worker's whereabouts. We keep
221 : * an array of these in shared memory, sized according to
222 : * autovacuum_worker_slots.
223 : *
224 : * wi_links entry into free list or running list
225 : * wi_dboid OID of the database this worker is supposed to work on
226 : * wi_tableoid OID of the table currently being vacuumed, if any
227 : * wi_sharedrel flag indicating whether table is marked relisshared
228 : * wi_proc pointer to PGPROC of the running worker, NULL if not started
229 : * wi_launchtime Time at which this worker was launched
230 : * wi_dobalance Whether this worker should be included in balance calculations
231 : *
232 : * All fields are protected by AutovacuumLock, except for wi_tableoid and
233 : * wi_sharedrel which are protected by AutovacuumScheduleLock (note these
234 : * two fields are read-only for everyone except that worker itself).
235 : *-------------
236 : */
237 : typedef struct WorkerInfoData
238 : {
239 : dlist_node wi_links;
240 : Oid wi_dboid;
241 : Oid wi_tableoid;
242 : PGPROC *wi_proc;
243 : TimestampTz wi_launchtime;
244 : pg_atomic_flag wi_dobalance;
245 : bool wi_sharedrel;
246 : } WorkerInfoData;
247 :
248 : typedef struct WorkerInfoData *WorkerInfo;
249 :
250 : /*
251 : * Possible signals received by the launcher from remote processes. These are
252 : * stored atomically in shared memory so that other processes can set them
253 : * without locking.
254 : */
255 : typedef enum
256 : {
257 : AutoVacForkFailed, /* failed trying to start a worker */
258 : AutoVacRebalance, /* rebalance the cost limits */
259 : } AutoVacuumSignal;
260 :
261 : #define AutoVacNumSignals (AutoVacRebalance + 1)
262 :
263 : /*
264 : * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This
265 : * list is mostly protected by AutovacuumLock, except that if an item is
266 : * marked 'active' other processes must not modify the work-identifying
267 : * members.
268 : */
269 : typedef struct AutoVacuumWorkItem
270 : {
271 : AutoVacuumWorkItemType avw_type;
272 : bool avw_used; /* below data is valid */
273 : bool avw_active; /* being processed */
274 : Oid avw_database;
275 : Oid avw_relation;
276 : BlockNumber avw_blockNumber;
277 : } AutoVacuumWorkItem;
278 :
279 : #define NUM_WORKITEMS 256
280 :
281 : /*-------------
282 : * The main autovacuum shmem struct. On shared memory we store this main
283 : * struct and the array of WorkerInfo structs. This struct keeps:
284 : *
285 : * av_signal set by other processes to indicate various conditions
286 : * av_launcherpid the PID of the autovacuum launcher
287 : * av_freeWorkers the WorkerInfo freelist
288 : * av_runningWorkers the WorkerInfo non-free queue
289 : * av_startingWorker pointer to WorkerInfo currently being started (cleared by
290 : * the worker itself as soon as it's up and running)
291 : * av_workItems work item array
292 : * av_nworkersForBalance the number of autovacuum workers to use when
293 : * calculating the per worker cost limit
294 : *
295 : * This struct is protected by AutovacuumLock, except for av_signal and parts
296 : * of the worker list (see above).
297 : *-------------
298 : */
299 : typedef struct
300 : {
301 : sig_atomic_t av_signal[AutoVacNumSignals];
302 : pid_t av_launcherpid;
303 : dclist_head av_freeWorkers;
304 : dlist_head av_runningWorkers;
305 : WorkerInfo av_startingWorker;
306 : AutoVacuumWorkItem av_workItems[NUM_WORKITEMS];
307 : pg_atomic_uint32 av_nworkersForBalance;
308 : } AutoVacuumShmemStruct;
309 :
310 : static AutoVacuumShmemStruct *AutoVacuumShmem;
311 :
312 : /*
313 : * the database list (of avl_dbase elements) in the launcher, and the context
314 : * that contains it
315 : */
316 : static dlist_head DatabaseList = DLIST_STATIC_INIT(DatabaseList);
317 : static MemoryContext DatabaseListCxt = NULL;
318 :
319 : /*
320 : * This struct is used by relation_needs_vacanalyze() to return the table's
321 : * score (i.e., the maximum of the component scores) as well as the component
322 : * scores themselves.
323 : */
324 : typedef struct
325 : {
326 : double max; /* maximum of all values below */
327 : double xid; /* transaction ID component */
328 : double mxid; /* multixact ID component */
329 : double vac; /* vacuum component */
330 : double vac_ins; /* vacuum insert component */
331 : double anl; /* analyze component */
332 : } AutoVacuumScores;
333 :
334 : /*
335 : * This struct is used to track and sort the list of tables to process.
336 : */
337 : typedef struct
338 : {
339 : Oid oid;
340 : double score;
341 : } TableToProcess;
342 :
343 : /*
344 : * Dummy pointer to persuade Valgrind that we've not leaked the array of
345 : * avl_dbase structs. Make it global to ensure the compiler doesn't
346 : * optimize it away.
347 : */
348 : #ifdef USE_VALGRIND
349 : extern avl_dbase *avl_dbase_array;
350 : avl_dbase *avl_dbase_array;
351 : #endif
352 :
353 : /* Pointer to my own WorkerInfo, valid on each worker */
354 : static WorkerInfo MyWorkerInfo = NULL;
355 :
356 : static Oid do_start_worker(void);
357 : static void ProcessAutoVacLauncherInterrupts(void);
358 : pg_noreturn static void AutoVacLauncherShutdown(void);
359 : static void launcher_determine_sleep(bool canlaunch, bool recursing,
360 : struct timeval *nap);
361 : static void launch_worker(TimestampTz now);
362 : static List *get_database_list(void);
363 : static void rebuild_database_list(Oid newdb);
364 : static int db_comparator(const void *a, const void *b);
365 : static void autovac_recalculate_workers_for_balance(void);
366 :
367 : static void do_autovacuum(void);
368 : static void FreeWorkerInfo(int code, Datum arg);
369 :
370 : static autovac_table *table_recheck_autovac(Oid relid, HTAB *table_toast_map,
371 : TupleDesc pg_class_desc,
372 : int effective_multixact_freeze_max_age);
373 : static void recheck_relation_needs_vacanalyze(Oid relid, AutoVacOpts *avopts,
374 : Form_pg_class classForm,
375 : int effective_multixact_freeze_max_age,
376 : bool *dovacuum, bool *doanalyze, bool *wraparound);
377 : static void relation_needs_vacanalyze(Oid relid, AutoVacOpts *relopts,
378 : Form_pg_class classForm,
379 : PgStat_StatTabEntry *tabentry,
380 : int effective_multixact_freeze_max_age,
381 : bool *dovacuum, bool *doanalyze, bool *wraparound,
382 : AutoVacuumScores *scores);
383 :
384 : static void autovacuum_do_vac_analyze(autovac_table *tab,
385 : BufferAccessStrategy bstrategy);
386 : static AutoVacOpts *extract_autovac_opts(HeapTuple tup,
387 : TupleDesc pg_class_desc);
388 : static void perform_work_item(AutoVacuumWorkItem *workitem);
389 : static void autovac_report_activity(autovac_table *tab);
390 : static void autovac_report_workitem(AutoVacuumWorkItem *workitem,
391 : const char *nspname, const char *relname);
392 : static void avl_sigusr2_handler(SIGNAL_ARGS);
393 : static bool av_worker_available(void);
394 : static void check_av_worker_gucs(void);
395 :
396 :
397 :
398 : /********************************************************************
399 : * AUTOVACUUM LAUNCHER CODE
400 : ********************************************************************/
401 :
402 : /*
403 : * Main entry point for the autovacuum launcher process.
404 : */
405 : void
406 459 : AutoVacLauncherMain(const void *startup_data, size_t startup_data_len)
407 : {
408 : sigjmp_buf local_sigjmp_buf;
409 :
410 : Assert(startup_data_len == 0);
411 :
412 : /* Release postmaster's working memory context */
413 459 : if (PostmasterContext)
414 : {
415 459 : MemoryContextDelete(PostmasterContext);
416 459 : PostmasterContext = NULL;
417 : }
418 :
419 459 : init_ps_display(NULL);
420 :
421 459 : ereport(DEBUG1,
422 : (errmsg_internal("autovacuum launcher started")));
423 :
424 459 : if (PostAuthDelay)
425 0 : pg_usleep(PostAuthDelay * 1000000L);
426 :
427 : Assert(GetProcessingMode() == InitProcessing);
428 :
429 : /*
430 : * Set up signal handlers. We operate on databases much like a regular
431 : * backend, so we use the same signal handling. See equivalent code in
432 : * tcop/postgres.c.
433 : */
434 459 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
435 459 : pqsignal(SIGINT, StatementCancelHandler);
436 459 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
437 : /* SIGQUIT handler was already set up by InitPostmasterChild */
438 :
439 459 : InitializeTimeouts(); /* establishes SIGALRM handler */
440 :
441 459 : pqsignal(SIGPIPE, SIG_IGN);
442 459 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
443 459 : pqsignal(SIGUSR2, avl_sigusr2_handler);
444 459 : pqsignal(SIGFPE, FloatExceptionHandler);
445 459 : pqsignal(SIGCHLD, SIG_DFL);
446 :
447 : /*
448 : * Create a per-backend PGPROC struct in shared memory. We must do this
449 : * before we can use LWLocks or access any shared memory.
450 : */
451 459 : InitProcess();
452 :
453 : /* Early initialization */
454 459 : BaseInit();
455 :
456 459 : InitPostgres(NULL, InvalidOid, NULL, InvalidOid, 0, NULL);
457 :
458 459 : SetProcessingMode(NormalProcessing);
459 :
460 : /*
461 : * Create a memory context that we will do all our work in. We do this so
462 : * that we can reset the context during error recovery and thereby avoid
463 : * possible memory leaks.
464 : */
465 459 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
466 : "Autovacuum Launcher",
467 : ALLOCSET_DEFAULT_SIZES);
468 459 : MemoryContextSwitchTo(AutovacMemCxt);
469 :
470 : /*
471 : * If an exception is encountered, processing resumes here.
472 : *
473 : * This code is a stripped down version of PostgresMain error recovery.
474 : *
475 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
476 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
477 : * signals other than SIGQUIT will be blocked until we complete error
478 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
479 : * call redundant, but it is not since InterruptPending might be set
480 : * already.
481 : */
482 459 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
483 : {
484 : /* since not using PG_TRY, must reset error stack by hand */
485 0 : error_context_stack = NULL;
486 :
487 : /* Prevents interrupts while cleaning up */
488 0 : HOLD_INTERRUPTS();
489 :
490 : /* Forget any pending QueryCancel or timeout request */
491 0 : disable_all_timeouts(false);
492 0 : QueryCancelPending = false; /* second to avoid race condition */
493 :
494 : /* Report the error to the server log */
495 0 : EmitErrorReport();
496 :
497 : /* Abort the current transaction in order to recover */
498 0 : AbortCurrentTransaction();
499 :
500 : /*
501 : * Release any other resources, for the case where we were not in a
502 : * transaction.
503 : */
504 0 : LWLockReleaseAll();
505 0 : pgstat_report_wait_end();
506 0 : pgaio_error_cleanup();
507 0 : UnlockBuffers();
508 : /* this is probably dead code, but let's be safe: */
509 0 : if (AuxProcessResourceOwner)
510 0 : ReleaseAuxProcessResources(false);
511 0 : AtEOXact_Buffers(false);
512 0 : AtEOXact_SMgr();
513 0 : AtEOXact_Files(false);
514 0 : AtEOXact_HashTables(false);
515 :
516 : /*
517 : * Now return to normal top-level context and clear ErrorContext for
518 : * next time.
519 : */
520 0 : MemoryContextSwitchTo(AutovacMemCxt);
521 0 : FlushErrorState();
522 :
523 : /* Flush any leaked data in the top-level context */
524 0 : MemoryContextReset(AutovacMemCxt);
525 :
526 : /* don't leave dangling pointers to freed memory */
527 0 : DatabaseListCxt = NULL;
528 0 : dlist_init(&DatabaseList);
529 :
530 : /* Now we can allow interrupts again */
531 0 : RESUME_INTERRUPTS();
532 :
533 : /* if in shutdown mode, no need for anything further; just go away */
534 0 : if (ShutdownRequestPending)
535 0 : AutoVacLauncherShutdown();
536 :
537 : /*
538 : * Sleep at least 1 second after any error. We don't want to be
539 : * filling the error logs as fast as we can.
540 : */
541 0 : pg_usleep(1000000L);
542 : }
543 :
544 : /* We can now handle ereport(ERROR) */
545 459 : PG_exception_stack = &local_sigjmp_buf;
546 :
547 : /* must unblock signals before calling rebuild_database_list */
548 459 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
549 :
550 : /*
551 : * Set always-secure search path. Launcher doesn't connect to a database,
552 : * so this has no effect.
553 : */
554 459 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
555 :
556 : /*
557 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
558 : * in postgresql.conf. We don't really want such a dangerous option being
559 : * applied non-interactively.
560 : */
561 459 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
562 :
563 : /*
564 : * Force settable timeouts off to avoid letting these settings prevent
565 : * regular maintenance from being executed.
566 : */
567 459 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
568 459 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
569 459 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
570 459 : SetConfigOption("idle_in_transaction_session_timeout", "0",
571 : PGC_SUSET, PGC_S_OVERRIDE);
572 :
573 : /*
574 : * Force default_transaction_isolation to READ COMMITTED. We don't want
575 : * to pay the overhead of serializable mode, nor add any risk of causing
576 : * deadlocks or delaying other transactions.
577 : */
578 459 : SetConfigOption("default_transaction_isolation", "read committed",
579 : PGC_SUSET, PGC_S_OVERRIDE);
580 :
581 : /*
582 : * Even when system is configured to use a different fetch consistency,
583 : * for autovac we always want fresh stats.
584 : */
585 459 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
586 :
587 : /*
588 : * In emergency mode, just start a worker (unless shutdown was requested)
589 : * and go away.
590 : */
591 459 : if (!AutoVacuumingActive())
592 : {
593 0 : if (!ShutdownRequestPending)
594 0 : do_start_worker();
595 0 : proc_exit(0); /* done */
596 : }
597 :
598 459 : AutoVacuumShmem->av_launcherpid = MyProcPid;
599 :
600 : /*
601 : * Create the initial database list. The invariant we want this list to
602 : * keep is that it's ordered by decreasing next_worker. As soon as an
603 : * entry is updated to a higher time, it will be moved to the front (which
604 : * is correct because the only operation is to add autovacuum_naptime to
605 : * the entry, and time always increases).
606 : */
607 459 : rebuild_database_list(InvalidOid);
608 :
609 : /* loop until shutdown request */
610 5157 : while (!ShutdownRequestPending)
611 : {
612 : struct timeval nap;
613 5157 : TimestampTz current_time = 0;
614 : bool can_launch;
615 :
616 : /*
617 : * This loop is a bit different from the normal use of WaitLatch,
618 : * because we'd like to sleep before the first launch of a child
619 : * process. So it's WaitLatch, then ResetLatch, then check for
620 : * wakening conditions.
621 : */
622 :
623 5157 : launcher_determine_sleep(av_worker_available(), false, &nap);
624 :
625 : /*
626 : * Wait until naptime expires or we get some type of signal (all the
627 : * signal handlers will wake us by calling SetLatch).
628 : */
629 5157 : (void) WaitLatch(MyLatch,
630 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
631 5157 : (nap.tv_sec * 1000L) + (nap.tv_usec / 1000L),
632 : WAIT_EVENT_AUTOVACUUM_MAIN);
633 :
634 5154 : ResetLatch(MyLatch);
635 :
636 5154 : ProcessAutoVacLauncherInterrupts();
637 :
638 : /*
639 : * a worker finished, or postmaster signaled failure to start a worker
640 : */
641 4698 : if (got_SIGUSR2)
642 : {
643 2782 : got_SIGUSR2 = false;
644 :
645 : /* rebalance cost limits, if needed */
646 2782 : if (AutoVacuumShmem->av_signal[AutoVacRebalance])
647 : {
648 1361 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
649 1361 : AutoVacuumShmem->av_signal[AutoVacRebalance] = false;
650 1361 : autovac_recalculate_workers_for_balance();
651 1361 : LWLockRelease(AutovacuumLock);
652 : }
653 :
654 2782 : if (AutoVacuumShmem->av_signal[AutoVacForkFailed])
655 : {
656 : /*
657 : * If the postmaster failed to start a new worker, we sleep
658 : * for a little while and resend the signal. The new worker's
659 : * state is still in memory, so this is sufficient. After
660 : * that, we restart the main loop.
661 : *
662 : * XXX should we put a limit to the number of times we retry?
663 : * I don't think it makes much sense, because a future start
664 : * of a worker will continue to fail in the same way.
665 : */
666 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = false;
667 0 : pg_usleep(1000000L); /* 1s */
668 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
669 1503 : continue;
670 : }
671 : }
672 :
673 : /*
674 : * There are some conditions that we need to check before trying to
675 : * start a worker. First, we need to make sure that there is a worker
676 : * slot available. Second, we need to make sure that no other worker
677 : * failed while starting up.
678 : */
679 :
680 4698 : current_time = GetCurrentTimestamp();
681 4698 : LWLockAcquire(AutovacuumLock, LW_SHARED);
682 :
683 4698 : can_launch = av_worker_available();
684 :
685 4698 : if (AutoVacuumShmem->av_startingWorker != NULL)
686 : {
687 : int waittime;
688 13 : WorkerInfo worker = AutoVacuumShmem->av_startingWorker;
689 :
690 : /*
691 : * We can't launch another worker when another one is still
692 : * starting up (or failed while doing so), so just sleep for a bit
693 : * more; that worker will wake us up again as soon as it's ready.
694 : * We will only wait autovacuum_naptime seconds (up to a maximum
695 : * of 60 seconds) for this to happen however. Note that failure
696 : * to connect to a particular database is not a problem here,
697 : * because the worker removes itself from the startingWorker
698 : * pointer before trying to connect. Problems detected by the
699 : * postmaster (like fork() failure) are also reported and handled
700 : * differently. The only problems that may cause this code to
701 : * fire are errors in the earlier sections of AutoVacWorkerMain,
702 : * before the worker removes the WorkerInfo from the
703 : * startingWorker pointer.
704 : */
705 13 : waittime = Min(autovacuum_naptime, 60) * 1000;
706 13 : if (TimestampDifferenceExceeds(worker->wi_launchtime, current_time,
707 : waittime))
708 : {
709 0 : LWLockRelease(AutovacuumLock);
710 0 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
711 :
712 : /*
713 : * No other process can put a worker in starting mode, so if
714 : * startingWorker is still INVALID after exchanging our lock,
715 : * we assume it's the same one we saw above (so we don't
716 : * recheck the launch time).
717 : */
718 0 : if (AutoVacuumShmem->av_startingWorker != NULL)
719 : {
720 0 : worker = AutoVacuumShmem->av_startingWorker;
721 0 : worker->wi_dboid = InvalidOid;
722 0 : worker->wi_tableoid = InvalidOid;
723 0 : worker->wi_sharedrel = false;
724 0 : worker->wi_proc = NULL;
725 0 : worker->wi_launchtime = 0;
726 0 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
727 : &worker->wi_links);
728 0 : AutoVacuumShmem->av_startingWorker = NULL;
729 0 : ereport(WARNING,
730 : errmsg("autovacuum worker took too long to start; canceled"));
731 : }
732 : }
733 : else
734 13 : can_launch = false;
735 : }
736 4698 : LWLockRelease(AutovacuumLock); /* either shared or exclusive */
737 :
738 : /* if we can't do anything, just go back to sleep */
739 4698 : if (!can_launch)
740 1503 : continue;
741 :
742 : /* We're OK to start a new worker */
743 :
744 3195 : if (dlist_is_empty(&DatabaseList))
745 : {
746 : /*
747 : * Special case when the list is empty: start a worker right away.
748 : * This covers the initial case, when no database is in pgstats
749 : * (thus the list is empty). Note that the constraints in
750 : * launcher_determine_sleep keep us from starting workers too
751 : * quickly (at most once every autovacuum_naptime when the list is
752 : * empty).
753 : */
754 4 : launch_worker(current_time);
755 : }
756 : else
757 : {
758 : /*
759 : * because rebuild_database_list constructs a list with most
760 : * distant adl_next_worker first, we obtain our database from the
761 : * tail of the list.
762 : */
763 : avl_dbase *avdb;
764 :
765 3191 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
766 :
767 : /*
768 : * launch a worker if next_worker is right now or it is in the
769 : * past
770 : */
771 3191 : if (TimestampDifferenceExceeds(avdb->adl_next_worker,
772 : current_time, 0))
773 1394 : launch_worker(current_time);
774 : }
775 : }
776 :
777 0 : AutoVacLauncherShutdown();
778 : }
779 :
780 : /*
781 : * Process any new interrupts.
782 : */
783 : static void
784 5154 : ProcessAutoVacLauncherInterrupts(void)
785 : {
786 : /* the normal shutdown case */
787 5154 : if (ShutdownRequestPending)
788 455 : AutoVacLauncherShutdown();
789 :
790 4699 : if (ConfigReloadPending)
791 : {
792 50 : int autovacuum_max_workers_prev = autovacuum_max_workers;
793 :
794 50 : ConfigReloadPending = false;
795 50 : ProcessConfigFile(PGC_SIGHUP);
796 :
797 : /* shutdown requested in config file? */
798 50 : if (!AutoVacuumingActive())
799 1 : AutoVacLauncherShutdown();
800 :
801 : /*
802 : * If autovacuum_max_workers changed, emit a WARNING if
803 : * autovacuum_worker_slots < autovacuum_max_workers. If it didn't
804 : * change, skip this to avoid too many repeated log messages.
805 : */
806 49 : if (autovacuum_max_workers_prev != autovacuum_max_workers)
807 0 : check_av_worker_gucs();
808 :
809 : /* rebuild the list in case the naptime changed */
810 49 : rebuild_database_list(InvalidOid);
811 : }
812 :
813 : /* Process barrier events */
814 4698 : if (ProcSignalBarrierPending)
815 42 : ProcessProcSignalBarrier();
816 :
817 : /* Perform logging of memory contexts of this process */
818 4698 : if (LogMemoryContextPending)
819 0 : ProcessLogMemoryContextInterrupt();
820 :
821 : /* Process sinval catchup interrupts that happened while sleeping */
822 4698 : ProcessCatchupInterrupt();
823 4698 : }
824 :
825 : /*
826 : * Perform a normal exit from the autovac launcher.
827 : */
828 : static void
829 456 : AutoVacLauncherShutdown(void)
830 : {
831 456 : ereport(DEBUG1,
832 : (errmsg_internal("autovacuum launcher shutting down")));
833 456 : AutoVacuumShmem->av_launcherpid = 0;
834 :
835 456 : proc_exit(0); /* done */
836 : }
837 :
838 : /*
839 : * Determine the time to sleep, based on the database list.
840 : *
841 : * The "canlaunch" parameter indicates whether we can start a worker right now,
842 : * for example due to the workers being all busy. If this is false, we will
843 : * cause a long sleep, which will be interrupted when a worker exits.
844 : */
845 : static void
846 5219 : launcher_determine_sleep(bool canlaunch, bool recursing, struct timeval *nap)
847 : {
848 : /*
849 : * We sleep until the next scheduled vacuum. We trust that when the
850 : * database list was built, care was taken so that no entries have times
851 : * in the past; if the first entry has too close a next_worker value, or a
852 : * time in the past, we will sleep a small nominal time.
853 : */
854 5219 : if (!canlaunch)
855 : {
856 2757 : nap->tv_sec = autovacuum_naptime;
857 2757 : nap->tv_usec = 0;
858 : }
859 2462 : else if (!dlist_is_empty(&DatabaseList))
860 : {
861 2432 : TimestampTz current_time = GetCurrentTimestamp();
862 : TimestampTz next_wakeup;
863 : avl_dbase *avdb;
864 : long secs;
865 : int usecs;
866 :
867 2432 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
868 :
869 2432 : next_wakeup = avdb->adl_next_worker;
870 2432 : TimestampDifference(current_time, next_wakeup, &secs, &usecs);
871 :
872 2432 : nap->tv_sec = secs;
873 2432 : nap->tv_usec = usecs;
874 : }
875 : else
876 : {
877 : /* list is empty, sleep for whole autovacuum_naptime seconds */
878 30 : nap->tv_sec = autovacuum_naptime;
879 30 : nap->tv_usec = 0;
880 : }
881 :
882 : /*
883 : * If the result is exactly zero, it means a database had an entry with
884 : * time in the past. Rebuild the list so that the databases are evenly
885 : * distributed again, and recalculate the time to sleep. This can happen
886 : * if there are more tables needing vacuum than workers, and they all take
887 : * longer to vacuum than autovacuum_naptime.
888 : *
889 : * We only recurse once. rebuild_database_list should always return times
890 : * in the future, but it seems best not to trust too much on that.
891 : */
892 5219 : if (nap->tv_sec == 0 && nap->tv_usec == 0 && !recursing)
893 : {
894 62 : rebuild_database_list(InvalidOid);
895 62 : launcher_determine_sleep(canlaunch, true, nap);
896 62 : return;
897 : }
898 :
899 : /* The smallest time we'll allow the launcher to sleep. */
900 5157 : if (nap->tv_sec <= 0 && nap->tv_usec <= MIN_AUTOVAC_SLEEPTIME * 1000)
901 : {
902 184 : nap->tv_sec = 0;
903 184 : nap->tv_usec = MIN_AUTOVAC_SLEEPTIME * 1000;
904 : }
905 :
906 : /*
907 : * If the sleep time is too large, clamp it to an arbitrary maximum (plus
908 : * any fractional seconds, for simplicity). This avoids an essentially
909 : * infinite sleep in strange cases like the system clock going backwards a
910 : * few years.
911 : */
912 5157 : if (nap->tv_sec > MAX_AUTOVAC_SLEEPTIME)
913 9 : nap->tv_sec = MAX_AUTOVAC_SLEEPTIME;
914 : }
915 :
916 : /*
917 : * Build an updated DatabaseList. It must only contain databases that appear
918 : * in pgstats, and must be sorted by next_worker from highest to lowest,
919 : * distributed regularly across the next autovacuum_naptime interval.
920 : *
921 : * Receives the Oid of the database that made this list be generated (we call
922 : * this the "new" database, because when the database was already present on
923 : * the list, we expect that this function is not called at all). The
924 : * preexisting list, if any, will be used to preserve the order of the
925 : * databases in the autovacuum_naptime period. The new database is put at the
926 : * end of the interval. The actual values are not saved, which should not be
927 : * much of a problem.
928 : */
929 : static void
930 584 : rebuild_database_list(Oid newdb)
931 : {
932 : List *dblist;
933 : ListCell *cell;
934 : MemoryContext newcxt;
935 : MemoryContext oldcxt;
936 : MemoryContext tmpcxt;
937 : HASHCTL hctl;
938 : int score;
939 : int nelems;
940 : HTAB *dbhash;
941 : dlist_iter iter;
942 :
943 584 : newcxt = AllocSetContextCreate(AutovacMemCxt,
944 : "Autovacuum database list",
945 : ALLOCSET_DEFAULT_SIZES);
946 584 : tmpcxt = AllocSetContextCreate(newcxt,
947 : "Autovacuum database list (tmp)",
948 : ALLOCSET_DEFAULT_SIZES);
949 584 : oldcxt = MemoryContextSwitchTo(tmpcxt);
950 :
951 : /*
952 : * Implementing this is not as simple as it sounds, because we need to put
953 : * the new database at the end of the list; next the databases that were
954 : * already on the list, and finally (at the tail of the list) all the
955 : * other databases that are not on the existing list.
956 : *
957 : * To do this, we build an empty hash table of scored databases. We will
958 : * start with the lowest score (zero) for the new database, then
959 : * increasing scores for the databases in the existing list, in order, and
960 : * lastly increasing scores for all databases gotten via
961 : * get_database_list() that are not already on the hash.
962 : *
963 : * Then we will put all the hash elements into an array, sort the array by
964 : * score, and finally put the array elements into the new doubly linked
965 : * list.
966 : */
967 584 : hctl.keysize = sizeof(Oid);
968 584 : hctl.entrysize = sizeof(avl_dbase);
969 584 : hctl.hcxt = tmpcxt;
970 584 : dbhash = hash_create("autovacuum db hash", 20, &hctl, /* magic number here
971 : * FIXME */
972 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
973 :
974 : /* start by inserting the new database */
975 584 : score = 0;
976 584 : if (OidIsValid(newdb))
977 : {
978 : avl_dbase *db;
979 : PgStat_StatDBEntry *entry;
980 :
981 : /* only consider this database if it has a pgstat entry */
982 14 : entry = pgstat_fetch_stat_dbentry(newdb);
983 14 : if (entry != NULL)
984 : {
985 : /* we assume it isn't found because the hash was just created */
986 11 : db = hash_search(dbhash, &newdb, HASH_ENTER, NULL);
987 :
988 : /* hash_search already filled in the key */
989 11 : db->adl_score = score++;
990 : /* next_worker is filled in later */
991 : }
992 : }
993 :
994 : /* Now insert the databases from the existing list */
995 880 : dlist_foreach(iter, &DatabaseList)
996 : {
997 296 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
998 : avl_dbase *db;
999 : bool found;
1000 : PgStat_StatDBEntry *entry;
1001 :
1002 : /*
1003 : * skip databases with no stat entries -- in particular, this gets rid
1004 : * of dropped databases
1005 : */
1006 296 : entry = pgstat_fetch_stat_dbentry(avdb->adl_datid);
1007 296 : if (entry == NULL)
1008 0 : continue;
1009 :
1010 296 : db = hash_search(dbhash, &(avdb->adl_datid), HASH_ENTER, &found);
1011 :
1012 296 : if (!found)
1013 : {
1014 : /* hash_search already filled in the key */
1015 296 : db->adl_score = score++;
1016 : /* next_worker is filled in later */
1017 : }
1018 : }
1019 :
1020 : /* finally, insert all qualifying databases not previously inserted */
1021 584 : dblist = get_database_list();
1022 2646 : foreach(cell, dblist)
1023 : {
1024 2062 : avw_dbase *avdb = lfirst(cell);
1025 : avl_dbase *db;
1026 : bool found;
1027 : PgStat_StatDBEntry *entry;
1028 :
1029 : /* only consider databases with a pgstat entry */
1030 2062 : entry = pgstat_fetch_stat_dbentry(avdb->adw_datid);
1031 2062 : if (entry == NULL)
1032 1085 : continue;
1033 :
1034 977 : db = hash_search(dbhash, &(avdb->adw_datid), HASH_ENTER, &found);
1035 : /* only update the score if the database was not already on the hash */
1036 977 : if (!found)
1037 : {
1038 : /* hash_search already filled in the key */
1039 670 : db->adl_score = score++;
1040 : /* next_worker is filled in later */
1041 : }
1042 : }
1043 584 : nelems = score;
1044 :
1045 : /* from here on, the allocated memory belongs to the new list */
1046 584 : MemoryContextSwitchTo(newcxt);
1047 584 : dlist_init(&DatabaseList);
1048 :
1049 584 : if (nelems > 0)
1050 : {
1051 : TimestampTz current_time;
1052 : int millis_increment;
1053 : avl_dbase *dbary;
1054 : avl_dbase *db;
1055 : HASH_SEQ_STATUS seq;
1056 : int i;
1057 :
1058 : /* put all the hash elements into an array */
1059 558 : dbary = palloc(nelems * sizeof(avl_dbase));
1060 : /* keep Valgrind quiet */
1061 : #ifdef USE_VALGRIND
1062 : avl_dbase_array = dbary;
1063 : #endif
1064 :
1065 558 : i = 0;
1066 558 : hash_seq_init(&seq, dbhash);
1067 1535 : while ((db = hash_seq_search(&seq)) != NULL)
1068 977 : memcpy(&(dbary[i++]), db, sizeof(avl_dbase));
1069 :
1070 : /* sort the array */
1071 558 : qsort(dbary, nelems, sizeof(avl_dbase), db_comparator);
1072 :
1073 : /*
1074 : * Determine the time interval between databases in the schedule. If
1075 : * we see that the configured naptime would take us to sleep times
1076 : * lower than our min sleep time (which launcher_determine_sleep is
1077 : * coded not to allow), silently use a larger naptime (but don't touch
1078 : * the GUC variable).
1079 : */
1080 558 : millis_increment = 1000.0 * autovacuum_naptime / nelems;
1081 558 : if (millis_increment <= MIN_AUTOVAC_SLEEPTIME)
1082 0 : millis_increment = MIN_AUTOVAC_SLEEPTIME * 1.1;
1083 :
1084 558 : current_time = GetCurrentTimestamp();
1085 :
1086 : /*
1087 : * move the elements from the array into the dlist, setting the
1088 : * next_worker while walking the array
1089 : */
1090 1535 : for (i = 0; i < nelems; i++)
1091 : {
1092 977 : db = &(dbary[i]);
1093 :
1094 977 : current_time = TimestampTzPlusMilliseconds(current_time,
1095 : millis_increment);
1096 977 : db->adl_next_worker = current_time;
1097 :
1098 : /* later elements should go closer to the head of the list */
1099 977 : dlist_push_head(&DatabaseList, &db->adl_node);
1100 : }
1101 : }
1102 :
1103 : /* all done, clean up memory */
1104 584 : if (DatabaseListCxt != NULL)
1105 125 : MemoryContextDelete(DatabaseListCxt);
1106 584 : MemoryContextDelete(tmpcxt);
1107 584 : DatabaseListCxt = newcxt;
1108 584 : MemoryContextSwitchTo(oldcxt);
1109 584 : }
1110 :
1111 : /* qsort comparator for avl_dbase, using adl_score */
1112 : static int
1113 626 : db_comparator(const void *a, const void *b)
1114 : {
1115 1252 : return pg_cmp_s32(((const avl_dbase *) a)->adl_score,
1116 626 : ((const avl_dbase *) b)->adl_score);
1117 : }
1118 :
1119 : /*
1120 : * do_start_worker
1121 : *
1122 : * Bare-bones procedure for starting an autovacuum worker from the launcher.
1123 : * It determines what database to work on, sets up shared memory stuff and
1124 : * signals postmaster to start the worker. It fails gracefully if invoked when
1125 : * autovacuum_workers are already active.
1126 : *
1127 : * Return value is the OID of the database that the worker is going to process,
1128 : * or InvalidOid if no worker was actually started.
1129 : */
1130 : static Oid
1131 1398 : do_start_worker(void)
1132 : {
1133 : List *dblist;
1134 : ListCell *cell;
1135 : TransactionId xidForceLimit;
1136 : MultiXactId multiForceLimit;
1137 : bool for_xid_wrap;
1138 : bool for_multi_wrap;
1139 : avw_dbase *avdb;
1140 : TimestampTz current_time;
1141 1398 : bool skipit = false;
1142 1398 : Oid retval = InvalidOid;
1143 : MemoryContext tmpcxt,
1144 : oldcxt;
1145 :
1146 : /* return quickly when there are no free workers */
1147 1398 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1148 1398 : if (!av_worker_available())
1149 : {
1150 0 : LWLockRelease(AutovacuumLock);
1151 0 : return InvalidOid;
1152 : }
1153 1398 : LWLockRelease(AutovacuumLock);
1154 :
1155 : /*
1156 : * Create and switch to a temporary context to avoid leaking the memory
1157 : * allocated for the database list.
1158 : */
1159 1398 : tmpcxt = AllocSetContextCreate(CurrentMemoryContext,
1160 : "Autovacuum start worker (tmp)",
1161 : ALLOCSET_DEFAULT_SIZES);
1162 1398 : oldcxt = MemoryContextSwitchTo(tmpcxt);
1163 :
1164 : /* Get a list of databases */
1165 1398 : dblist = get_database_list();
1166 :
1167 : /*
1168 : * Determine the oldest datfrozenxid/relfrozenxid that we will allow to
1169 : * pass without forcing a vacuum. (This limit can be tightened for
1170 : * particular tables, but not loosened.)
1171 : */
1172 1398 : recentXid = ReadNextTransactionId();
1173 1398 : xidForceLimit = recentXid - autovacuum_freeze_max_age;
1174 : /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */
1175 : /* this can cause the limit to go backwards by 3, but that's OK */
1176 1398 : if (xidForceLimit < FirstNormalTransactionId)
1177 0 : xidForceLimit -= FirstNormalTransactionId;
1178 :
1179 : /* Also determine the oldest datminmxid we will consider. */
1180 1398 : recentMulti = ReadNextMultiXactId();
1181 1398 : multiForceLimit = recentMulti - MultiXactMemberFreezeThreshold();
1182 1398 : if (multiForceLimit < FirstMultiXactId)
1183 0 : multiForceLimit -= FirstMultiXactId;
1184 :
1185 : /*
1186 : * Choose a database to connect to. We pick the database that was least
1187 : * recently auto-vacuumed, or one that needs vacuuming to prevent Xid
1188 : * wraparound-related data loss. If any db at risk of Xid wraparound is
1189 : * found, we pick the one with oldest datfrozenxid, independently of
1190 : * autovacuum times; similarly we pick the one with the oldest datminmxid
1191 : * if any is in MultiXactId wraparound. Note that those in Xid wraparound
1192 : * danger are given more priority than those in multi wraparound danger.
1193 : *
1194 : * Note that a database with no stats entry is not considered, except for
1195 : * Xid wraparound purposes. The theory is that if no one has ever
1196 : * connected to it since the stats were last initialized, it doesn't need
1197 : * vacuuming.
1198 : *
1199 : * XXX This could be improved if we had more info about whether it needs
1200 : * vacuuming before connecting to it. Perhaps look through the pgstats
1201 : * data for the database's tables? One idea is to keep track of the
1202 : * number of new and dead tuples per database in pgstats. However it
1203 : * isn't clear how to construct a metric that measures that and not cause
1204 : * starvation for less busy databases.
1205 : */
1206 1398 : avdb = NULL;
1207 1398 : for_xid_wrap = false;
1208 1398 : for_multi_wrap = false;
1209 1398 : current_time = GetCurrentTimestamp();
1210 5634 : foreach(cell, dblist)
1211 : {
1212 4236 : avw_dbase *tmp = lfirst(cell);
1213 : dlist_iter iter;
1214 :
1215 : /* Check to see if this one is at risk of wraparound */
1216 4236 : if (TransactionIdPrecedes(tmp->adw_frozenxid, xidForceLimit))
1217 : {
1218 4015 : if (avdb == NULL ||
1219 1590 : TransactionIdPrecedes(tmp->adw_frozenxid,
1220 : avdb->adw_frozenxid))
1221 907 : avdb = tmp;
1222 2425 : for_xid_wrap = true;
1223 3307 : continue;
1224 : }
1225 1811 : else if (for_xid_wrap)
1226 85 : continue; /* ignore not-at-risk DBs */
1227 1726 : else if (MultiXactIdPrecedes(tmp->adw_minmulti, multiForceLimit))
1228 : {
1229 0 : if (avdb == NULL ||
1230 0 : MultiXactIdPrecedes(tmp->adw_minmulti, avdb->adw_minmulti))
1231 0 : avdb = tmp;
1232 0 : for_multi_wrap = true;
1233 0 : continue;
1234 : }
1235 1726 : else if (for_multi_wrap)
1236 0 : continue; /* ignore not-at-risk DBs */
1237 :
1238 : /* Find pgstat entry if any */
1239 1726 : tmp->adw_entry = pgstat_fetch_stat_dbentry(tmp->adw_datid);
1240 :
1241 : /*
1242 : * Skip a database with no pgstat entry; it means it hasn't seen any
1243 : * activity.
1244 : */
1245 1726 : if (!tmp->adw_entry)
1246 73 : continue;
1247 :
1248 : /*
1249 : * Also, skip a database that appears on the database list as having
1250 : * been processed recently (less than autovacuum_naptime seconds ago).
1251 : * We do this so that we don't select a database which we just
1252 : * selected, but that pgstat hasn't gotten around to updating the last
1253 : * autovacuum time yet.
1254 : */
1255 1653 : skipit = false;
1256 :
1257 3377 : dlist_reverse_foreach(iter, &DatabaseList)
1258 : {
1259 3352 : avl_dbase *dbp = dlist_container(avl_dbase, adl_node, iter.cur);
1260 :
1261 3352 : if (dbp->adl_datid == tmp->adw_datid)
1262 : {
1263 : /*
1264 : * Skip this database if its next_worker value falls between
1265 : * the current time and the current time plus naptime.
1266 : */
1267 1628 : if (!TimestampDifferenceExceeds(dbp->adl_next_worker,
1268 724 : current_time, 0) &&
1269 724 : !TimestampDifferenceExceeds(current_time,
1270 : dbp->adl_next_worker,
1271 : autovacuum_naptime * 1000))
1272 724 : skipit = true;
1273 :
1274 1628 : break;
1275 : }
1276 : }
1277 1653 : if (skipit)
1278 724 : continue;
1279 :
1280 : /*
1281 : * Remember the db with oldest autovac time. (If we are here, both
1282 : * tmp->entry and db->entry must be non-null.)
1283 : */
1284 929 : if (avdb == NULL ||
1285 370 : tmp->adw_entry->last_autovac_time < avdb->adw_entry->last_autovac_time)
1286 690 : avdb = tmp;
1287 : }
1288 :
1289 : /* Found a database -- process it */
1290 1398 : if (avdb != NULL)
1291 : {
1292 : WorkerInfo worker;
1293 : dlist_node *wptr;
1294 :
1295 1394 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1296 :
1297 : /*
1298 : * Get a worker entry from the freelist. We checked above, so there
1299 : * really should be a free slot.
1300 : */
1301 1394 : wptr = dclist_pop_head_node(&AutoVacuumShmem->av_freeWorkers);
1302 :
1303 1394 : worker = dlist_container(WorkerInfoData, wi_links, wptr);
1304 1394 : worker->wi_dboid = avdb->adw_datid;
1305 1394 : worker->wi_proc = NULL;
1306 1394 : worker->wi_launchtime = GetCurrentTimestamp();
1307 :
1308 1394 : AutoVacuumShmem->av_startingWorker = worker;
1309 :
1310 1394 : LWLockRelease(AutovacuumLock);
1311 :
1312 1394 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
1313 :
1314 1394 : retval = avdb->adw_datid;
1315 : }
1316 4 : else if (skipit)
1317 : {
1318 : /*
1319 : * If we skipped all databases on the list, rebuild it, because it
1320 : * probably contains a dropped database.
1321 : */
1322 0 : rebuild_database_list(InvalidOid);
1323 : }
1324 :
1325 1398 : MemoryContextSwitchTo(oldcxt);
1326 1398 : MemoryContextDelete(tmpcxt);
1327 :
1328 1398 : return retval;
1329 : }
1330 :
1331 : /*
1332 : * launch_worker
1333 : *
1334 : * Wrapper for starting a worker from the launcher. Besides actually starting
1335 : * it, update the database list to reflect the next time that another one will
1336 : * need to be started on the selected database. The actual database choice is
1337 : * left to do_start_worker.
1338 : *
1339 : * This routine is also expected to insert an entry into the database list if
1340 : * the selected database was previously absent from the list.
1341 : */
1342 : static void
1343 1398 : launch_worker(TimestampTz now)
1344 : {
1345 : Oid dbid;
1346 : dlist_iter iter;
1347 :
1348 1398 : dbid = do_start_worker();
1349 1398 : if (OidIsValid(dbid))
1350 : {
1351 1394 : bool found = false;
1352 :
1353 : /*
1354 : * Walk the database list and update the corresponding entry. If the
1355 : * database is not on the list, we'll recreate the list.
1356 : */
1357 2763 : dlist_foreach(iter, &DatabaseList)
1358 : {
1359 2749 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1360 :
1361 2749 : if (avdb->adl_datid == dbid)
1362 : {
1363 1380 : found = true;
1364 :
1365 : /*
1366 : * add autovacuum_naptime seconds to the current time, and use
1367 : * that as the new "next_worker" field for this database.
1368 : */
1369 1380 : avdb->adl_next_worker =
1370 1380 : TimestampTzPlusMilliseconds(now, autovacuum_naptime * 1000);
1371 :
1372 1380 : dlist_move_head(&DatabaseList, iter.cur);
1373 1380 : break;
1374 : }
1375 : }
1376 :
1377 : /*
1378 : * If the database was not present in the database list, we rebuild
1379 : * the list. It's possible that the database does not get into the
1380 : * list anyway, for example if it's a database that doesn't have a
1381 : * pgstat entry, but this is not a problem because we don't want to
1382 : * schedule workers regularly into those in any case.
1383 : */
1384 1394 : if (!found)
1385 14 : rebuild_database_list(dbid);
1386 : }
1387 1398 : }
1388 :
1389 : /*
1390 : * Called from postmaster to signal a failure to fork a process to become
1391 : * worker. The postmaster should kill(SIGUSR2) the launcher shortly
1392 : * after calling this function.
1393 : */
1394 : void
1395 0 : AutoVacWorkerFailed(void)
1396 : {
1397 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = true;
1398 0 : }
1399 :
1400 : /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */
1401 : static void
1402 2784 : avl_sigusr2_handler(SIGNAL_ARGS)
1403 : {
1404 2784 : got_SIGUSR2 = true;
1405 2784 : SetLatch(MyLatch);
1406 2784 : }
1407 :
1408 :
1409 : /********************************************************************
1410 : * AUTOVACUUM WORKER CODE
1411 : ********************************************************************/
1412 :
1413 : /*
1414 : * Main entry point for autovacuum worker processes.
1415 : */
1416 : void
1417 1401 : AutoVacWorkerMain(const void *startup_data, size_t startup_data_len)
1418 : {
1419 : sigjmp_buf local_sigjmp_buf;
1420 : Oid dbid;
1421 :
1422 : Assert(startup_data_len == 0);
1423 :
1424 : /* Release postmaster's working memory context */
1425 1401 : if (PostmasterContext)
1426 : {
1427 1401 : MemoryContextDelete(PostmasterContext);
1428 1401 : PostmasterContext = NULL;
1429 : }
1430 :
1431 1401 : init_ps_display(NULL);
1432 :
1433 : Assert(GetProcessingMode() == InitProcessing);
1434 :
1435 : /*
1436 : * Set up signal handlers. We operate on databases much like a regular
1437 : * backend, so we use the same signal handling. See equivalent code in
1438 : * tcop/postgres.c.
1439 : */
1440 1401 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1441 :
1442 : /*
1443 : * SIGINT is used to signal canceling the current table's vacuum; SIGTERM
1444 : * means abort and exit cleanly, and SIGQUIT means abandon ship.
1445 : */
1446 1401 : pqsignal(SIGINT, StatementCancelHandler);
1447 1401 : pqsignal(SIGTERM, die);
1448 : /* SIGQUIT handler was already set up by InitPostmasterChild */
1449 :
1450 1401 : InitializeTimeouts(); /* establishes SIGALRM handler */
1451 :
1452 1401 : pqsignal(SIGPIPE, SIG_IGN);
1453 1401 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1454 1401 : pqsignal(SIGUSR2, SIG_IGN);
1455 1401 : pqsignal(SIGFPE, FloatExceptionHandler);
1456 1401 : pqsignal(SIGCHLD, SIG_DFL);
1457 :
1458 : /*
1459 : * Create a per-backend PGPROC struct in shared memory. We must do this
1460 : * before we can use LWLocks or access any shared memory.
1461 : */
1462 1401 : InitProcess();
1463 :
1464 : /* Early initialization */
1465 1401 : BaseInit();
1466 :
1467 : /*
1468 : * If an exception is encountered, processing resumes here.
1469 : *
1470 : * Unlike most auxiliary processes, we don't attempt to continue
1471 : * processing after an error; we just clean up and exit. The autovac
1472 : * launcher is responsible for spawning another worker later.
1473 : *
1474 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
1475 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
1476 : * signals other than SIGQUIT will be blocked until we exit. It might
1477 : * seem that this policy makes the HOLD_INTERRUPTS() call redundant, but
1478 : * it is not since InterruptPending might be set already.
1479 : */
1480 1401 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1481 : {
1482 : /* since not using PG_TRY, must reset error stack by hand */
1483 0 : error_context_stack = NULL;
1484 :
1485 : /* Prevents interrupts while cleaning up */
1486 0 : HOLD_INTERRUPTS();
1487 :
1488 : /* Report the error to the server log */
1489 0 : EmitErrorReport();
1490 :
1491 : /*
1492 : * We can now go away. Note that because we called InitProcess, a
1493 : * callback was registered to do ProcKill, which will clean up
1494 : * necessary state.
1495 : */
1496 0 : proc_exit(0);
1497 : }
1498 :
1499 : /* We can now handle ereport(ERROR) */
1500 1401 : PG_exception_stack = &local_sigjmp_buf;
1501 :
1502 1401 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
1503 :
1504 : /*
1505 : * Set always-secure search path, so malicious users can't redirect user
1506 : * code (e.g. pg_index.indexprs). (That code runs in a
1507 : * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not
1508 : * take control of the entire autovacuum worker in any case.)
1509 : */
1510 1401 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1511 :
1512 : /*
1513 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
1514 : * in postgresql.conf. We don't really want such a dangerous option being
1515 : * applied non-interactively.
1516 : */
1517 1401 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
1518 :
1519 : /*
1520 : * Force settable timeouts off to avoid letting these settings prevent
1521 : * regular maintenance from being executed.
1522 : */
1523 1401 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1524 1401 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1525 1401 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1526 1401 : SetConfigOption("idle_in_transaction_session_timeout", "0",
1527 : PGC_SUSET, PGC_S_OVERRIDE);
1528 :
1529 : /*
1530 : * Force default_transaction_isolation to READ COMMITTED. We don't want
1531 : * to pay the overhead of serializable mode, nor add any risk of causing
1532 : * deadlocks or delaying other transactions.
1533 : */
1534 1401 : SetConfigOption("default_transaction_isolation", "read committed",
1535 : PGC_SUSET, PGC_S_OVERRIDE);
1536 :
1537 : /*
1538 : * Force synchronous replication off to allow regular maintenance even if
1539 : * we are waiting for standbys to connect. This is important to ensure we
1540 : * aren't blocked from performing anti-wraparound tasks.
1541 : */
1542 1401 : if (synchronous_commit > SYNCHRONOUS_COMMIT_LOCAL_FLUSH)
1543 1401 : SetConfigOption("synchronous_commit", "local",
1544 : PGC_SUSET, PGC_S_OVERRIDE);
1545 :
1546 : /*
1547 : * Even when system is configured to use a different fetch consistency,
1548 : * for autovac we always want fresh stats.
1549 : */
1550 1401 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
1551 :
1552 : /*
1553 : * Get the info about the database we're going to work on.
1554 : */
1555 1401 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1556 :
1557 : /*
1558 : * beware of startingWorker being INVALID; this should normally not
1559 : * happen, but if a worker fails after forking and before this, the
1560 : * launcher might have decided to remove it from the queue and start
1561 : * again.
1562 : */
1563 1401 : if (AutoVacuumShmem->av_startingWorker != NULL)
1564 : {
1565 1401 : MyWorkerInfo = AutoVacuumShmem->av_startingWorker;
1566 1401 : dbid = MyWorkerInfo->wi_dboid;
1567 1401 : MyWorkerInfo->wi_proc = MyProc;
1568 :
1569 : /* insert into the running list */
1570 1401 : dlist_push_head(&AutoVacuumShmem->av_runningWorkers,
1571 1401 : &MyWorkerInfo->wi_links);
1572 :
1573 : /*
1574 : * remove from the "starting" pointer, so that the launcher can start
1575 : * a new worker if required
1576 : */
1577 1401 : AutoVacuumShmem->av_startingWorker = NULL;
1578 1401 : LWLockRelease(AutovacuumLock);
1579 :
1580 1401 : on_shmem_exit(FreeWorkerInfo, 0);
1581 :
1582 : /* wake up the launcher */
1583 1401 : if (AutoVacuumShmem->av_launcherpid != 0)
1584 1401 : kill(AutoVacuumShmem->av_launcherpid, SIGUSR2);
1585 : }
1586 : else
1587 : {
1588 : /* no worker entry for me, go away */
1589 0 : elog(WARNING, "autovacuum worker started without a worker entry");
1590 0 : dbid = InvalidOid;
1591 0 : LWLockRelease(AutovacuumLock);
1592 : }
1593 :
1594 1401 : if (OidIsValid(dbid))
1595 : {
1596 : char dbname[NAMEDATALEN];
1597 :
1598 : /*
1599 : * Report autovac startup to the cumulative stats system. We
1600 : * deliberately do this before InitPostgres, so that the
1601 : * last_autovac_time will get updated even if the connection attempt
1602 : * fails. This is to prevent autovac from getting "stuck" repeatedly
1603 : * selecting an unopenable database, rather than making any progress
1604 : * on stuff it can connect to.
1605 : */
1606 1401 : pgstat_report_autovac(dbid);
1607 :
1608 : /*
1609 : * Connect to the selected database, specifying no particular user,
1610 : * and ignoring datallowconn. Collect the database's name for
1611 : * display.
1612 : *
1613 : * Note: if we have selected a just-deleted database (due to using
1614 : * stale stats info), we'll fail and exit here.
1615 : */
1616 1401 : InitPostgres(NULL, dbid, NULL, InvalidOid,
1617 : INIT_PG_OVERRIDE_ALLOW_CONNS,
1618 : dbname);
1619 1401 : SetProcessingMode(NormalProcessing);
1620 1401 : set_ps_display(dbname);
1621 1401 : ereport(DEBUG1,
1622 : (errmsg_internal("autovacuum: processing database \"%s\"", dbname)));
1623 :
1624 1401 : if (PostAuthDelay)
1625 0 : pg_usleep(PostAuthDelay * 1000000L);
1626 :
1627 : /* And do an appropriate amount of work */
1628 1401 : recentXid = ReadNextTransactionId();
1629 1401 : recentMulti = ReadNextMultiXactId();
1630 1401 : do_autovacuum();
1631 : }
1632 :
1633 : /* All done, go away */
1634 1399 : proc_exit(0);
1635 : }
1636 :
1637 : /*
1638 : * Return a WorkerInfo to the free list
1639 : */
1640 : static void
1641 1401 : FreeWorkerInfo(int code, Datum arg)
1642 : {
1643 1401 : if (MyWorkerInfo != NULL)
1644 : {
1645 1401 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1646 :
1647 1401 : dlist_delete(&MyWorkerInfo->wi_links);
1648 1401 : MyWorkerInfo->wi_dboid = InvalidOid;
1649 1401 : MyWorkerInfo->wi_tableoid = InvalidOid;
1650 1401 : MyWorkerInfo->wi_sharedrel = false;
1651 1401 : MyWorkerInfo->wi_proc = NULL;
1652 1401 : MyWorkerInfo->wi_launchtime = 0;
1653 1401 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
1654 1401 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
1655 1401 : &MyWorkerInfo->wi_links);
1656 : /* not mine anymore */
1657 1401 : MyWorkerInfo = NULL;
1658 :
1659 : /*
1660 : * now that we're inactive, cause a rebalancing of the surviving
1661 : * workers
1662 : */
1663 1401 : AutoVacuumShmem->av_signal[AutoVacRebalance] = true;
1664 1401 : LWLockRelease(AutovacuumLock);
1665 : }
1666 1401 : }
1667 :
1668 : /*
1669 : * Update vacuum cost-based delay-related parameters for autovacuum workers and
1670 : * backends executing VACUUM or ANALYZE using the value of relevant GUCs and
1671 : * global state. This must be called during setup for vacuum and after every
1672 : * config reload to ensure up-to-date values.
1673 : */
1674 : void
1675 207252 : VacuumUpdateCosts(void)
1676 : {
1677 207252 : if (MyWorkerInfo)
1678 : {
1679 198624 : if (av_storage_param_cost_delay >= 0)
1680 0 : vacuum_cost_delay = av_storage_param_cost_delay;
1681 198624 : else if (autovacuum_vac_cost_delay >= 0)
1682 198624 : vacuum_cost_delay = autovacuum_vac_cost_delay;
1683 : else
1684 : /* fall back to VacuumCostDelay */
1685 0 : vacuum_cost_delay = VacuumCostDelay;
1686 :
1687 198624 : AutoVacuumUpdateCostLimit();
1688 : }
1689 : else
1690 : {
1691 : /* Must be explicit VACUUM or ANALYZE */
1692 8628 : vacuum_cost_delay = VacuumCostDelay;
1693 8628 : vacuum_cost_limit = VacuumCostLimit;
1694 : }
1695 :
1696 : /*
1697 : * If configuration changes are allowed to impact VacuumCostActive, make
1698 : * sure it is updated.
1699 : */
1700 207252 : if (VacuumFailsafeActive)
1701 : Assert(!VacuumCostActive);
1702 207252 : else if (vacuum_cost_delay > 0)
1703 198624 : VacuumCostActive = true;
1704 : else
1705 : {
1706 8628 : VacuumCostActive = false;
1707 8628 : VacuumCostBalance = 0;
1708 : }
1709 :
1710 : /*
1711 : * Since the cost logging requires a lock, avoid rendering the log message
1712 : * in case we are using a message level where the log wouldn't be emitted.
1713 : */
1714 207252 : if (MyWorkerInfo && message_level_is_interesting(DEBUG2))
1715 : {
1716 : Oid dboid,
1717 : tableoid;
1718 :
1719 : Assert(!LWLockHeldByMe(AutovacuumLock));
1720 :
1721 0 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1722 0 : dboid = MyWorkerInfo->wi_dboid;
1723 0 : tableoid = MyWorkerInfo->wi_tableoid;
1724 0 : LWLockRelease(AutovacuumLock);
1725 :
1726 0 : elog(DEBUG2,
1727 : "Autovacuum VacuumUpdateCosts(db=%u, rel=%u, dobalance=%s, cost_limit=%d, cost_delay=%g active=%s failsafe=%s)",
1728 : dboid, tableoid, pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance) ? "no" : "yes",
1729 : vacuum_cost_limit, vacuum_cost_delay,
1730 : vacuum_cost_delay > 0 ? "yes" : "no",
1731 : VacuumFailsafeActive ? "yes" : "no");
1732 : }
1733 207252 : }
1734 :
1735 : /*
1736 : * Update vacuum_cost_limit with the correct value for an autovacuum worker,
1737 : * given the value of other relevant cost limit parameters and the number of
1738 : * workers across which the limit must be balanced. Autovacuum workers must
1739 : * call this regularly in case av_nworkersForBalance has been updated by
1740 : * another worker or by the autovacuum launcher. They must also call it after a
1741 : * config reload.
1742 : */
1743 : void
1744 201802 : AutoVacuumUpdateCostLimit(void)
1745 : {
1746 201802 : if (!MyWorkerInfo)
1747 0 : return;
1748 :
1749 : /*
1750 : * note: in cost_limit, zero also means use value from elsewhere, because
1751 : * zero is not a valid value.
1752 : */
1753 :
1754 201802 : if (av_storage_param_cost_limit > 0)
1755 0 : vacuum_cost_limit = av_storage_param_cost_limit;
1756 : else
1757 : {
1758 : int nworkers_for_balance;
1759 :
1760 201802 : if (autovacuum_vac_cost_limit > 0)
1761 0 : vacuum_cost_limit = autovacuum_vac_cost_limit;
1762 : else
1763 201802 : vacuum_cost_limit = VacuumCostLimit;
1764 :
1765 : /* Only balance limit if no cost-related storage parameters specified */
1766 201802 : if (pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance))
1767 0 : return;
1768 :
1769 : Assert(vacuum_cost_limit > 0);
1770 :
1771 201802 : nworkers_for_balance = pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1772 :
1773 : /* There is at least 1 autovac worker (this worker) */
1774 201802 : if (nworkers_for_balance <= 0)
1775 0 : elog(ERROR, "nworkers_for_balance must be > 0");
1776 :
1777 201802 : vacuum_cost_limit = Max(vacuum_cost_limit / nworkers_for_balance, 1);
1778 : }
1779 : }
1780 :
1781 : /*
1782 : * autovac_recalculate_workers_for_balance
1783 : * Recalculate the number of workers to consider, given cost-related
1784 : * storage parameters and the current number of active workers.
1785 : *
1786 : * Caller must hold the AutovacuumLock in at least shared mode to access
1787 : * worker->wi_proc.
1788 : */
1789 : static void
1790 100673 : autovac_recalculate_workers_for_balance(void)
1791 : {
1792 : dlist_iter iter;
1793 : int orig_nworkers_for_balance;
1794 100673 : int nworkers_for_balance = 0;
1795 :
1796 : Assert(LWLockHeldByMe(AutovacuumLock));
1797 :
1798 100673 : orig_nworkers_for_balance =
1799 100673 : pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1800 :
1801 321374 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1802 : {
1803 220701 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1804 :
1805 441402 : if (worker->wi_proc == NULL ||
1806 220701 : pg_atomic_unlocked_test_flag(&worker->wi_dobalance))
1807 3156 : continue;
1808 :
1809 217545 : nworkers_for_balance++;
1810 : }
1811 :
1812 100673 : if (nworkers_for_balance != orig_nworkers_for_balance)
1813 1885 : pg_atomic_write_u32(&AutoVacuumShmem->av_nworkersForBalance,
1814 : nworkers_for_balance);
1815 100673 : }
1816 :
1817 : /*
1818 : * get_database_list
1819 : * Return a list of all databases found in pg_database.
1820 : *
1821 : * The list and associated data is allocated in the caller's memory context,
1822 : * which is in charge of ensuring that it's properly cleaned up afterwards.
1823 : *
1824 : * Note: this is the only function in which the autovacuum launcher uses a
1825 : * transaction. Although we aren't attached to any particular database and
1826 : * therefore can't access most catalogs, we do have enough infrastructure
1827 : * to do a seqscan on pg_database.
1828 : */
1829 : static List *
1830 1982 : get_database_list(void)
1831 : {
1832 1982 : List *dblist = NIL;
1833 : Relation rel;
1834 : TableScanDesc scan;
1835 : HeapTuple tup;
1836 : MemoryContext resultcxt;
1837 :
1838 : /* This is the context that we will allocate our output data in */
1839 1982 : resultcxt = CurrentMemoryContext;
1840 :
1841 : /*
1842 : * Start a transaction so we can access pg_database.
1843 : */
1844 1982 : StartTransactionCommand();
1845 :
1846 1982 : rel = table_open(DatabaseRelationId, AccessShareLock);
1847 1982 : scan = table_beginscan_catalog(rel, 0, NULL);
1848 :
1849 8282 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1850 : {
1851 6300 : Form_pg_database pgdatabase = (Form_pg_database) GETSTRUCT(tup);
1852 : avw_dbase *avdb;
1853 : MemoryContext oldcxt;
1854 :
1855 : /*
1856 : * If database has partially been dropped, we can't, nor need to,
1857 : * vacuum it.
1858 : */
1859 6300 : if (database_is_invalid_form(pgdatabase))
1860 : {
1861 2 : elog(DEBUG2,
1862 : "autovacuum: skipping invalid database \"%s\"",
1863 : NameStr(pgdatabase->datname));
1864 2 : continue;
1865 : }
1866 :
1867 : /*
1868 : * Allocate our results in the caller's context, not the
1869 : * transaction's. We do this inside the loop, and restore the original
1870 : * context at the end, so that leaky things like heap_getnext() are
1871 : * not called in a potentially long-lived context.
1872 : */
1873 6298 : oldcxt = MemoryContextSwitchTo(resultcxt);
1874 :
1875 6298 : avdb = palloc_object(avw_dbase);
1876 :
1877 6298 : avdb->adw_datid = pgdatabase->oid;
1878 6298 : avdb->adw_name = pstrdup(NameStr(pgdatabase->datname));
1879 6298 : avdb->adw_frozenxid = pgdatabase->datfrozenxid;
1880 6298 : avdb->adw_minmulti = pgdatabase->datminmxid;
1881 : /* this gets set later: */
1882 6298 : avdb->adw_entry = NULL;
1883 :
1884 6298 : dblist = lappend(dblist, avdb);
1885 6298 : MemoryContextSwitchTo(oldcxt);
1886 : }
1887 :
1888 1982 : table_endscan(scan);
1889 1982 : table_close(rel, AccessShareLock);
1890 :
1891 1982 : CommitTransactionCommand();
1892 :
1893 : /* Be sure to restore caller's memory context */
1894 1982 : MemoryContextSwitchTo(resultcxt);
1895 :
1896 1982 : return dblist;
1897 : }
1898 :
1899 : /*
1900 : * List comparator for TableToProcess. Note that this sorts the tables based
1901 : * on their scores in descending order.
1902 : */
1903 : static int
1904 101935 : TableToProcessComparator(const ListCell *a, const ListCell *b)
1905 : {
1906 101935 : TableToProcess *t1 = (TableToProcess *) lfirst(a);
1907 101935 : TableToProcess *t2 = (TableToProcess *) lfirst(b);
1908 :
1909 101935 : return (t2->score < t1->score) ? -1 : (t2->score > t1->score) ? 1 : 0;
1910 : }
1911 :
1912 : /*
1913 : * Process a database table-by-table
1914 : *
1915 : * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in
1916 : * order not to ignore shutdown commands for too long.
1917 : */
1918 : static void
1919 1401 : do_autovacuum(void)
1920 : {
1921 : Relation classRel;
1922 : HeapTuple tuple;
1923 : TableScanDesc relScan;
1924 : Form_pg_database dbForm;
1925 1401 : List *tables_to_process = NIL;
1926 1401 : List *orphan_oids = NIL;
1927 : HASHCTL ctl;
1928 : HTAB *table_toast_map;
1929 : ListCell *volatile cell;
1930 : BufferAccessStrategy bstrategy;
1931 : ScanKeyData key;
1932 : TupleDesc pg_class_desc;
1933 : int effective_multixact_freeze_max_age;
1934 1401 : bool did_vacuum = false;
1935 1401 : bool found_concurrent_worker = false;
1936 : int i;
1937 :
1938 : /*
1939 : * StartTransactionCommand and CommitTransactionCommand will automatically
1940 : * switch to other contexts. We need this one to keep the list of
1941 : * relations to vacuum/analyze across transactions.
1942 : */
1943 1401 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
1944 : "Autovacuum worker",
1945 : ALLOCSET_DEFAULT_SIZES);
1946 1401 : MemoryContextSwitchTo(AutovacMemCxt);
1947 :
1948 : /* Start a transaction so our commands have one to play into. */
1949 1401 : StartTransactionCommand();
1950 :
1951 : /*
1952 : * This injection point is put in a transaction block to work with a wait
1953 : * that uses a condition variable.
1954 : */
1955 1401 : INJECTION_POINT("autovacuum-worker-start", NULL);
1956 :
1957 : /*
1958 : * Compute the multixact age for which freezing is urgent. This is
1959 : * normally autovacuum_multixact_freeze_max_age, but may be less if
1960 : * multixact members are bloated.
1961 : */
1962 1400 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1963 :
1964 : /*
1965 : * Find the pg_database entry and select the default freeze ages. We use
1966 : * zero in template and nonconnectable databases, else the system-wide
1967 : * default.
1968 : */
1969 1400 : tuple = SearchSysCache1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
1970 1400 : if (!HeapTupleIsValid(tuple))
1971 0 : elog(ERROR, "cache lookup failed for database %u", MyDatabaseId);
1972 1400 : dbForm = (Form_pg_database) GETSTRUCT(tuple);
1973 :
1974 1400 : if (dbForm->datistemplate || !dbForm->datallowconn)
1975 : {
1976 440 : default_freeze_min_age = 0;
1977 440 : default_freeze_table_age = 0;
1978 440 : default_multixact_freeze_min_age = 0;
1979 440 : default_multixact_freeze_table_age = 0;
1980 : }
1981 : else
1982 : {
1983 960 : default_freeze_min_age = vacuum_freeze_min_age;
1984 960 : default_freeze_table_age = vacuum_freeze_table_age;
1985 960 : default_multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1986 960 : default_multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1987 : }
1988 :
1989 1400 : ReleaseSysCache(tuple);
1990 :
1991 : /* StartTransactionCommand changed elsewhere */
1992 1400 : MemoryContextSwitchTo(AutovacMemCxt);
1993 :
1994 1400 : classRel = table_open(RelationRelationId, AccessShareLock);
1995 :
1996 : /* create a copy so we can use it after closing pg_class */
1997 1400 : pg_class_desc = CreateTupleDescCopy(RelationGetDescr(classRel));
1998 :
1999 : /* create hash table for toast <-> main relid mapping */
2000 1400 : ctl.keysize = sizeof(Oid);
2001 1400 : ctl.entrysize = sizeof(av_relation);
2002 :
2003 1400 : table_toast_map = hash_create("TOAST to main relid map",
2004 : 100,
2005 : &ctl,
2006 : HASH_ELEM | HASH_BLOBS);
2007 :
2008 : /*
2009 : * Scan pg_class to determine which tables to vacuum.
2010 : *
2011 : * We do this in two passes: on the first one we collect the list of plain
2012 : * relations and materialized views, and on the second one we collect
2013 : * TOAST tables. The reason for doing the second pass is that during it we
2014 : * want to use the main relation's pg_class.reloptions entry if the TOAST
2015 : * table does not have any, and we cannot obtain it unless we know
2016 : * beforehand what's the main table OID.
2017 : *
2018 : * We need to check TOAST tables separately because in cases with short,
2019 : * wide tables there might be proportionally much more activity in the
2020 : * TOAST table than in its parent.
2021 : */
2022 1400 : relScan = table_beginscan_catalog(classRel, 0, NULL);
2023 :
2024 : /*
2025 : * On the first pass, we collect main tables to vacuum, and also the main
2026 : * table relid to TOAST relid mapping.
2027 : */
2028 641156 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2029 : {
2030 639756 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2031 : PgStat_StatTabEntry *tabentry;
2032 : AutoVacOpts *relopts;
2033 : Oid relid;
2034 : bool dovacuum;
2035 : bool doanalyze;
2036 : bool wraparound;
2037 : AutoVacuumScores scores;
2038 :
2039 639756 : if (classForm->relkind != RELKIND_RELATION &&
2040 534206 : classForm->relkind != RELKIND_MATVIEW)
2041 534184 : continue;
2042 :
2043 105588 : relid = classForm->oid;
2044 :
2045 : /*
2046 : * Check if it is a temp table (presumably, of some other backend's).
2047 : * We cannot safely process other backends' temp tables.
2048 : */
2049 105588 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2050 : {
2051 : /*
2052 : * We just ignore it if the owning backend is still active and
2053 : * using the temporary schema. Also, for safety, ignore it if the
2054 : * namespace doesn't exist or isn't a temp namespace after all.
2055 : */
2056 16 : if (checkTempNamespaceStatus(classForm->relnamespace) == TEMP_NAMESPACE_IDLE)
2057 : {
2058 : /*
2059 : * The table seems to be orphaned -- although it might be that
2060 : * the owning backend has already deleted it and exited; our
2061 : * pg_class scan snapshot is not necessarily up-to-date
2062 : * anymore, so we could be looking at a committed-dead entry.
2063 : * Remember it so we can try to delete it later.
2064 : */
2065 0 : orphan_oids = lappend_oid(orphan_oids, relid);
2066 : }
2067 16 : continue;
2068 : }
2069 :
2070 : /* Fetch reloptions and the pgstat entry for this table */
2071 105572 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2072 105572 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2073 : relid);
2074 :
2075 : /* Check if it needs vacuum or analyze */
2076 105572 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2077 : effective_multixact_freeze_max_age,
2078 : &dovacuum, &doanalyze, &wraparound,
2079 : &scores);
2080 :
2081 : /* Relations that need work are added to tables_to_process */
2082 105572 : if (dovacuum || doanalyze)
2083 : {
2084 64019 : TableToProcess *table = palloc_object(TableToProcess);
2085 :
2086 64019 : table->oid = relid;
2087 64019 : table->score = scores.max;
2088 64019 : tables_to_process = lappend(tables_to_process, table);
2089 : }
2090 :
2091 : /*
2092 : * Remember TOAST associations for the second pass. Note: we must do
2093 : * this whether or not the table is going to be vacuumed, because we
2094 : * don't automatically vacuum toast tables along the parent table.
2095 : */
2096 105572 : if (OidIsValid(classForm->reltoastrelid))
2097 : {
2098 : av_relation *hentry;
2099 : bool found;
2100 :
2101 118832 : hentry = hash_search(table_toast_map,
2102 59416 : &classForm->reltoastrelid,
2103 : HASH_ENTER, &found);
2104 :
2105 59416 : if (!found)
2106 : {
2107 : /* hash_search already filled in the key */
2108 59416 : hentry->ar_relid = relid;
2109 59416 : hentry->ar_hasrelopts = false;
2110 59416 : if (relopts != NULL)
2111 : {
2112 1245 : hentry->ar_hasrelopts = true;
2113 1245 : memcpy(&hentry->ar_reloptions, relopts,
2114 : sizeof(AutoVacOpts));
2115 : }
2116 : }
2117 : }
2118 :
2119 : /* Release stuff to avoid per-relation leakage */
2120 105572 : if (relopts)
2121 1283 : pfree(relopts);
2122 105572 : if (tabentry)
2123 103149 : pfree(tabentry);
2124 : }
2125 :
2126 1400 : table_endscan(relScan);
2127 :
2128 : /* second pass: check TOAST tables */
2129 1400 : ScanKeyInit(&key,
2130 : Anum_pg_class_relkind,
2131 : BTEqualStrategyNumber, F_CHAREQ,
2132 : CharGetDatum(RELKIND_TOASTVALUE));
2133 :
2134 1400 : relScan = table_beginscan_catalog(classRel, 1, &key);
2135 60820 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2136 : {
2137 59420 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2138 : PgStat_StatTabEntry *tabentry;
2139 : Oid relid;
2140 : AutoVacOpts *relopts;
2141 59420 : bool free_relopts = false;
2142 : bool dovacuum;
2143 : bool doanalyze;
2144 : bool wraparound;
2145 : AutoVacuumScores scores;
2146 :
2147 : /*
2148 : * We cannot safely process other backends' temp tables, so skip 'em.
2149 : */
2150 59420 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2151 4 : continue;
2152 :
2153 59416 : relid = classForm->oid;
2154 :
2155 : /*
2156 : * fetch reloptions -- if this toast table does not have them, try the
2157 : * main rel
2158 : */
2159 59416 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2160 59416 : if (relopts)
2161 1 : free_relopts = true;
2162 : else
2163 : {
2164 : av_relation *hentry;
2165 : bool found;
2166 :
2167 59415 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2168 59415 : if (found && hentry->ar_hasrelopts)
2169 1244 : relopts = &hentry->ar_reloptions;
2170 : }
2171 :
2172 : /* Fetch the pgstat entry for this table */
2173 59416 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2174 : relid);
2175 :
2176 59416 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2177 : effective_multixact_freeze_max_age,
2178 : &dovacuum, &doanalyze, &wraparound,
2179 : &scores);
2180 :
2181 : /* ignore analyze for toast tables */
2182 59416 : if (dovacuum)
2183 : {
2184 35969 : TableToProcess *table = palloc_object(TableToProcess);
2185 :
2186 35969 : table->oid = relid;
2187 35969 : table->score = scores.max;
2188 35969 : tables_to_process = lappend(tables_to_process, table);
2189 : }
2190 :
2191 : /* Release stuff to avoid leakage */
2192 59416 : if (free_relopts)
2193 1 : pfree(relopts);
2194 59416 : if (tabentry)
2195 57788 : pfree(tabentry);
2196 : }
2197 :
2198 1400 : table_endscan(relScan);
2199 1400 : table_close(classRel, AccessShareLock);
2200 :
2201 : /*
2202 : * Recheck orphan temporary tables, and if they still seem orphaned, drop
2203 : * them. We'll eat a transaction per dropped table, which might seem
2204 : * excessive, but we should only need to do anything as a result of a
2205 : * previous backend crash, so this should not happen often enough to
2206 : * justify "optimizing". Using separate transactions ensures that we
2207 : * don't bloat the lock table if there are many temp tables to be dropped,
2208 : * and it ensures that we don't lose work if a deletion attempt fails.
2209 : */
2210 1400 : foreach(cell, orphan_oids)
2211 : {
2212 0 : Oid relid = lfirst_oid(cell);
2213 : Form_pg_class classForm;
2214 : ObjectAddress object;
2215 :
2216 : /*
2217 : * Check for user-requested abort.
2218 : */
2219 0 : CHECK_FOR_INTERRUPTS();
2220 :
2221 : /*
2222 : * Try to lock the table. If we can't get the lock immediately,
2223 : * somebody else is using (or dropping) the table, so it's not our
2224 : * concern anymore. Having the lock prevents race conditions below.
2225 : */
2226 0 : if (!ConditionalLockRelationOid(relid, AccessExclusiveLock))
2227 0 : continue;
2228 :
2229 : /*
2230 : * Re-fetch the pg_class tuple and re-check whether it still seems to
2231 : * be an orphaned temp table. If it's not there or no longer the same
2232 : * relation, ignore it.
2233 : */
2234 0 : tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2235 0 : if (!HeapTupleIsValid(tuple))
2236 : {
2237 : /* be sure to drop useless lock so we don't bloat lock table */
2238 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2239 0 : continue;
2240 : }
2241 0 : classForm = (Form_pg_class) GETSTRUCT(tuple);
2242 :
2243 : /*
2244 : * Make all the same tests made in the loop above. In event of OID
2245 : * counter wraparound, the pg_class entry we have now might be
2246 : * completely unrelated to the one we saw before.
2247 : */
2248 0 : if (!((classForm->relkind == RELKIND_RELATION ||
2249 0 : classForm->relkind == RELKIND_MATVIEW) &&
2250 0 : classForm->relpersistence == RELPERSISTENCE_TEMP))
2251 : {
2252 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2253 0 : continue;
2254 : }
2255 :
2256 0 : if (checkTempNamespaceStatus(classForm->relnamespace) != TEMP_NAMESPACE_IDLE)
2257 : {
2258 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2259 0 : continue;
2260 : }
2261 :
2262 : /*
2263 : * Try to lock the temp namespace, too. Even though we have lock on
2264 : * the table itself, there's a risk of deadlock against an incoming
2265 : * backend trying to clean out the temp namespace, in case this table
2266 : * has dependencies (such as sequences) that the backend's
2267 : * performDeletion call might visit in a different order. If we can
2268 : * get AccessShareLock on the namespace, that's sufficient to ensure
2269 : * we're not running concurrently with RemoveTempRelations. If we
2270 : * can't, back off and let RemoveTempRelations do its thing.
2271 : */
2272 0 : if (!ConditionalLockDatabaseObject(NamespaceRelationId,
2273 : classForm->relnamespace, 0,
2274 : AccessShareLock))
2275 : {
2276 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2277 0 : continue;
2278 : }
2279 :
2280 : /* OK, let's delete it */
2281 0 : ereport(LOG,
2282 : (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"",
2283 : get_database_name(MyDatabaseId),
2284 : get_namespace_name(classForm->relnamespace),
2285 : NameStr(classForm->relname))));
2286 :
2287 : /*
2288 : * Deletion might involve TOAST table access, so ensure we have a
2289 : * valid snapshot.
2290 : */
2291 0 : PushActiveSnapshot(GetTransactionSnapshot());
2292 :
2293 0 : object.classId = RelationRelationId;
2294 0 : object.objectId = relid;
2295 0 : object.objectSubId = 0;
2296 0 : performDeletion(&object, DROP_CASCADE,
2297 : PERFORM_DELETION_INTERNAL |
2298 : PERFORM_DELETION_QUIETLY |
2299 : PERFORM_DELETION_SKIP_EXTENSIONS);
2300 :
2301 : /*
2302 : * To commit the deletion, end current transaction and start a new
2303 : * one. Note this also releases the locks we took.
2304 : */
2305 0 : PopActiveSnapshot();
2306 0 : CommitTransactionCommand();
2307 0 : StartTransactionCommand();
2308 :
2309 : /* StartTransactionCommand changed current memory context */
2310 0 : MemoryContextSwitchTo(AutovacMemCxt);
2311 : }
2312 :
2313 : /*
2314 : * In case list_sort() would modify the list even when all the scores are
2315 : * 0.0, skip sorting if all the weight parameters are set to 0.0. This is
2316 : * probably not necessary, but we want to ensure folks have a guaranteed
2317 : * escape hatch from the scoring system.
2318 : */
2319 1400 : if (autovacuum_freeze_score_weight != 0.0 ||
2320 0 : autovacuum_multixact_freeze_score_weight != 0.0 ||
2321 0 : autovacuum_vacuum_score_weight != 0.0 ||
2322 0 : autovacuum_vacuum_insert_score_weight != 0.0 ||
2323 0 : autovacuum_analyze_score_weight != 0.0)
2324 1400 : list_sort(tables_to_process, TableToProcessComparator);
2325 :
2326 : /*
2327 : * Optionally, create a buffer access strategy object for VACUUM to use.
2328 : * We use the same BufferAccessStrategy object for all tables VACUUMed by
2329 : * this worker to prevent autovacuum from blowing out shared buffers.
2330 : *
2331 : * VacuumBufferUsageLimit being set to 0 results in
2332 : * GetAccessStrategyWithSize returning NULL, effectively meaning we can
2333 : * use up to all of shared buffers.
2334 : *
2335 : * If we later enter failsafe mode on any of the tables being vacuumed, we
2336 : * will cease use of the BufferAccessStrategy only for that table.
2337 : *
2338 : * XXX should we consider adding code to adjust the size of this if
2339 : * VacuumBufferUsageLimit changes?
2340 : */
2341 1400 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, VacuumBufferUsageLimit);
2342 :
2343 : /*
2344 : * create a memory context to act as fake PortalContext, so that the
2345 : * contexts created in the vacuum code are cleaned up for each table.
2346 : */
2347 1400 : PortalContext = AllocSetContextCreate(AutovacMemCxt,
2348 : "Autovacuum Portal",
2349 : ALLOCSET_DEFAULT_SIZES);
2350 :
2351 : /*
2352 : * Perform operations on collected tables.
2353 : */
2354 102779 : foreach_ptr(TableToProcess, table, tables_to_process)
2355 : {
2356 99981 : Oid relid = table->oid;
2357 : HeapTuple classTup;
2358 : autovac_table *tab;
2359 : bool isshared;
2360 : bool skipit;
2361 : dlist_iter iter;
2362 :
2363 99981 : CHECK_FOR_INTERRUPTS();
2364 :
2365 : /*
2366 : * Check for config changes before processing each collected table.
2367 : */
2368 99981 : if (ConfigReloadPending)
2369 : {
2370 0 : ConfigReloadPending = false;
2371 0 : ProcessConfigFile(PGC_SIGHUP);
2372 :
2373 : /*
2374 : * You might be tempted to bail out if we see autovacuum is now
2375 : * disabled. Must resist that temptation -- this might be a
2376 : * for-wraparound emergency worker, in which case that would be
2377 : * entirely inappropriate.
2378 : */
2379 : }
2380 :
2381 : /*
2382 : * Find out whether the table is shared or not. (It's slightly
2383 : * annoying to fetch the syscache entry just for this, but in typical
2384 : * cases it adds little cost because table_recheck_autovac would
2385 : * refetch the entry anyway. We could buy that back by copying the
2386 : * tuple here and passing it to table_recheck_autovac, but that
2387 : * increases the odds of that function working with stale data.)
2388 : */
2389 99981 : classTup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
2390 99981 : if (!HeapTupleIsValid(classTup))
2391 669 : continue; /* somebody deleted the rel, forget it */
2392 99981 : isshared = ((Form_pg_class) GETSTRUCT(classTup))->relisshared;
2393 99981 : ReleaseSysCache(classTup);
2394 :
2395 : /*
2396 : * Hold schedule lock from here until we've claimed the table. We
2397 : * also need the AutovacuumLock to walk the worker array, but that one
2398 : * can just be a shared lock.
2399 : */
2400 99981 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2401 99981 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2402 :
2403 : /*
2404 : * Check whether the table is being vacuumed concurrently by another
2405 : * worker.
2406 : */
2407 99981 : skipit = false;
2408 320347 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
2409 : {
2410 221010 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
2411 :
2412 : /* ignore myself */
2413 221010 : if (worker == MyWorkerInfo)
2414 99709 : continue;
2415 :
2416 : /* ignore workers in other databases (unless table is shared) */
2417 121301 : if (!worker->wi_sharedrel && worker->wi_dboid != MyDatabaseId)
2418 0 : continue;
2419 :
2420 121301 : if (worker->wi_tableoid == relid)
2421 : {
2422 644 : skipit = true;
2423 644 : found_concurrent_worker = true;
2424 644 : break;
2425 : }
2426 : }
2427 99981 : LWLockRelease(AutovacuumLock);
2428 99981 : if (skipit)
2429 : {
2430 644 : LWLockRelease(AutovacuumScheduleLock);
2431 644 : continue;
2432 : }
2433 :
2434 : /*
2435 : * Store the table's OID in shared memory before releasing the
2436 : * schedule lock, so that other workers don't try to vacuum it
2437 : * concurrently. (We claim it here so as not to hold
2438 : * AutovacuumScheduleLock while rechecking the stats.)
2439 : */
2440 99337 : MyWorkerInfo->wi_tableoid = relid;
2441 99337 : MyWorkerInfo->wi_sharedrel = isshared;
2442 99337 : LWLockRelease(AutovacuumScheduleLock);
2443 :
2444 : /*
2445 : * Check whether pgstat data still says we need to vacuum this table.
2446 : * It could have changed if something else processed the table while
2447 : * we weren't looking. This doesn't entirely close the race condition,
2448 : * but it is very small.
2449 : */
2450 99337 : MemoryContextSwitchTo(AutovacMemCxt);
2451 99337 : tab = table_recheck_autovac(relid, table_toast_map, pg_class_desc,
2452 : effective_multixact_freeze_max_age);
2453 99337 : if (tab == NULL)
2454 : {
2455 : /* someone else vacuumed the table, or it went away */
2456 25 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2457 25 : MyWorkerInfo->wi_tableoid = InvalidOid;
2458 25 : MyWorkerInfo->wi_sharedrel = false;
2459 25 : LWLockRelease(AutovacuumScheduleLock);
2460 25 : continue;
2461 : }
2462 :
2463 : /*
2464 : * Save the cost-related storage parameter values in global variables
2465 : * for reference when updating vacuum_cost_delay and vacuum_cost_limit
2466 : * during vacuuming this table.
2467 : */
2468 99312 : av_storage_param_cost_delay = tab->at_storage_param_vac_cost_delay;
2469 99312 : av_storage_param_cost_limit = tab->at_storage_param_vac_cost_limit;
2470 :
2471 : /*
2472 : * We only expect this worker to ever set the flag, so don't bother
2473 : * checking the return value. We shouldn't have to retry.
2474 : */
2475 99312 : if (tab->at_dobalance)
2476 99312 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2477 : else
2478 0 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
2479 :
2480 99312 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2481 99312 : autovac_recalculate_workers_for_balance();
2482 99312 : LWLockRelease(AutovacuumLock);
2483 :
2484 : /*
2485 : * We wait until this point to update cost delay and cost limit
2486 : * values, even though we reloaded the configuration file above, so
2487 : * that we can take into account the cost-related storage parameters.
2488 : */
2489 99312 : VacuumUpdateCosts();
2490 :
2491 :
2492 : /* clean up memory before each iteration */
2493 99312 : MemoryContextReset(PortalContext);
2494 :
2495 : /*
2496 : * Save the relation name for a possible error message, to avoid a
2497 : * catalog lookup in case of an error. If any of these return NULL,
2498 : * then the relation has been dropped since last we checked; skip it.
2499 : * Note: they must live in a long-lived memory context because we call
2500 : * vacuum and analyze in different transactions.
2501 : */
2502 :
2503 99312 : tab->at_relname = get_rel_name(tab->at_relid);
2504 99312 : tab->at_nspname = get_namespace_name(get_rel_namespace(tab->at_relid));
2505 99312 : tab->at_datname = get_database_name(MyDatabaseId);
2506 99312 : if (!tab->at_relname || !tab->at_nspname || !tab->at_datname)
2507 0 : goto deleted;
2508 :
2509 : /*
2510 : * We will abort vacuuming the current table if something errors out,
2511 : * and continue with the next one in schedule; in particular, this
2512 : * happens if we are interrupted with SIGINT.
2513 : */
2514 99312 : PG_TRY();
2515 : {
2516 : /* Use PortalContext for any per-table allocations */
2517 99312 : MemoryContextSwitchTo(PortalContext);
2518 :
2519 : /* have at it */
2520 99312 : autovacuum_do_vac_analyze(tab, bstrategy);
2521 :
2522 : /*
2523 : * Clear a possible query-cancel signal, to avoid a late reaction
2524 : * to an automatically-sent signal because of vacuuming the
2525 : * current table (we're done with it, so it would make no sense to
2526 : * cancel at this point.)
2527 : */
2528 99311 : QueryCancelPending = false;
2529 : }
2530 0 : PG_CATCH();
2531 : {
2532 : /*
2533 : * Abort the transaction, start a new one, and proceed with the
2534 : * next table in our list.
2535 : */
2536 0 : HOLD_INTERRUPTS();
2537 0 : if (tab->at_params.options & VACOPT_VACUUM)
2538 0 : errcontext("automatic vacuum of table \"%s.%s.%s\"",
2539 : tab->at_datname, tab->at_nspname, tab->at_relname);
2540 : else
2541 0 : errcontext("automatic analyze of table \"%s.%s.%s\"",
2542 : tab->at_datname, tab->at_nspname, tab->at_relname);
2543 0 : EmitErrorReport();
2544 :
2545 : /* this resets ProcGlobal->statusFlags[i] too */
2546 0 : AbortOutOfAnyTransaction();
2547 0 : FlushErrorState();
2548 0 : MemoryContextReset(PortalContext);
2549 :
2550 : /* restart our transaction for the following operations */
2551 0 : StartTransactionCommand();
2552 0 : RESUME_INTERRUPTS();
2553 : }
2554 99311 : PG_END_TRY();
2555 :
2556 : /* Make sure we're back in AutovacMemCxt */
2557 99311 : MemoryContextSwitchTo(AutovacMemCxt);
2558 :
2559 99311 : did_vacuum = true;
2560 :
2561 : /* ProcGlobal->statusFlags[i] are reset at the next end of xact */
2562 :
2563 : /* be tidy */
2564 99311 : deleted:
2565 99311 : if (tab->at_datname != NULL)
2566 99311 : pfree(tab->at_datname);
2567 99311 : if (tab->at_nspname != NULL)
2568 99311 : pfree(tab->at_nspname);
2569 99311 : if (tab->at_relname != NULL)
2570 99311 : pfree(tab->at_relname);
2571 99311 : pfree(tab);
2572 :
2573 : /*
2574 : * Remove my info from shared memory. We set wi_dobalance on the
2575 : * assumption that we are more likely than not to vacuum a table with
2576 : * no cost-related storage parameters next, so we want to claim our
2577 : * share of I/O as soon as possible to avoid thrashing the global
2578 : * balance.
2579 : */
2580 99311 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2581 99311 : MyWorkerInfo->wi_tableoid = InvalidOid;
2582 99311 : MyWorkerInfo->wi_sharedrel = false;
2583 99311 : LWLockRelease(AutovacuumScheduleLock);
2584 99311 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2585 : }
2586 :
2587 1399 : list_free_deep(tables_to_process);
2588 :
2589 : /*
2590 : * Perform additional work items, as requested by backends.
2591 : */
2592 1399 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2593 359543 : for (i = 0; i < NUM_WORKITEMS; i++)
2594 : {
2595 358144 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
2596 :
2597 358144 : if (!workitem->avw_used)
2598 358138 : continue;
2599 6 : if (workitem->avw_active)
2600 0 : continue;
2601 6 : if (workitem->avw_database != MyDatabaseId)
2602 0 : continue;
2603 :
2604 : /* claim this one, and release lock while performing it */
2605 6 : workitem->avw_active = true;
2606 6 : LWLockRelease(AutovacuumLock);
2607 :
2608 6 : PushActiveSnapshot(GetTransactionSnapshot());
2609 6 : perform_work_item(workitem);
2610 6 : if (ActiveSnapshotSet()) /* transaction could have aborted */
2611 6 : PopActiveSnapshot();
2612 :
2613 : /*
2614 : * Check for config changes before acquiring lock for further jobs.
2615 : */
2616 6 : CHECK_FOR_INTERRUPTS();
2617 6 : if (ConfigReloadPending)
2618 : {
2619 0 : ConfigReloadPending = false;
2620 0 : ProcessConfigFile(PGC_SIGHUP);
2621 0 : VacuumUpdateCosts();
2622 : }
2623 :
2624 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2625 :
2626 : /* and mark it done */
2627 6 : workitem->avw_active = false;
2628 6 : workitem->avw_used = false;
2629 : }
2630 1399 : LWLockRelease(AutovacuumLock);
2631 :
2632 : /*
2633 : * We leak table_toast_map here (among other things), but since we're
2634 : * going away soon, it's not a problem normally. But when using Valgrind,
2635 : * release some stuff to reduce complaints about leaked storage.
2636 : */
2637 : #ifdef USE_VALGRIND
2638 : hash_destroy(table_toast_map);
2639 : FreeTupleDesc(pg_class_desc);
2640 : if (bstrategy)
2641 : pfree(bstrategy);
2642 : #endif
2643 :
2644 : /* Run the rest in xact context, mainly to avoid Valgrind leak warnings */
2645 1399 : MemoryContextSwitchTo(TopTransactionContext);
2646 :
2647 : /*
2648 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We
2649 : * only need to do this once, not after each table.
2650 : *
2651 : * Even if we didn't vacuum anything, it may still be important to do
2652 : * this, because one indirect effect of vac_update_datfrozenxid() is to
2653 : * update TransamVariables->xidVacLimit. That might need to be done even
2654 : * if we haven't vacuumed anything, because relations with older
2655 : * relfrozenxid values or other databases with older datfrozenxid values
2656 : * might have been dropped, allowing xidVacLimit to advance.
2657 : *
2658 : * However, it's also important not to do this blindly in all cases,
2659 : * because when autovacuum=off this will restart the autovacuum launcher.
2660 : * If we're not careful, an infinite loop can result, where workers find
2661 : * no work to do and restart the launcher, which starts another worker in
2662 : * the same database that finds no work to do. To prevent that, we skip
2663 : * this if (1) we found no work to do and (2) we skipped at least one
2664 : * table due to concurrent autovacuum activity. In that case, the other
2665 : * worker has already done it, or will do so when it finishes.
2666 : */
2667 1399 : if (did_vacuum || !found_concurrent_worker)
2668 1399 : vac_update_datfrozenxid();
2669 :
2670 : /* Finally close out the last transaction. */
2671 1399 : CommitTransactionCommand();
2672 1399 : }
2673 :
2674 : /*
2675 : * Execute a previously registered work item.
2676 : */
2677 : static void
2678 6 : perform_work_item(AutoVacuumWorkItem *workitem)
2679 : {
2680 6 : char *cur_datname = NULL;
2681 6 : char *cur_nspname = NULL;
2682 6 : char *cur_relname = NULL;
2683 :
2684 : /*
2685 : * Note we do not store table info in MyWorkerInfo, since this is not
2686 : * vacuuming proper.
2687 : */
2688 :
2689 : /*
2690 : * Save the relation name for a possible error message, to avoid a catalog
2691 : * lookup in case of an error. If any of these return NULL, then the
2692 : * relation has been dropped since last we checked; skip it.
2693 : */
2694 : Assert(CurrentMemoryContext == AutovacMemCxt);
2695 :
2696 6 : cur_relname = get_rel_name(workitem->avw_relation);
2697 6 : cur_nspname = get_namespace_name(get_rel_namespace(workitem->avw_relation));
2698 6 : cur_datname = get_database_name(MyDatabaseId);
2699 6 : if (!cur_relname || !cur_nspname || !cur_datname)
2700 0 : goto deleted2;
2701 :
2702 6 : autovac_report_workitem(workitem, cur_nspname, cur_relname);
2703 :
2704 : /* clean up memory before each work item */
2705 6 : MemoryContextReset(PortalContext);
2706 :
2707 : /*
2708 : * We will abort the current work item if something errors out, and
2709 : * continue with the next one; in particular, this happens if we are
2710 : * interrupted with SIGINT. Note that this means that the work item list
2711 : * can be lossy.
2712 : */
2713 6 : PG_TRY();
2714 : {
2715 : /* Use PortalContext for any per-work-item allocations */
2716 6 : MemoryContextSwitchTo(PortalContext);
2717 :
2718 : /*
2719 : * Have at it. Functions called here are responsible for any required
2720 : * user switch and sandbox.
2721 : */
2722 6 : switch (workitem->avw_type)
2723 : {
2724 6 : case AVW_BRINSummarizeRange:
2725 6 : DirectFunctionCall2(brin_summarize_range,
2726 : ObjectIdGetDatum(workitem->avw_relation),
2727 : Int64GetDatum((int64) workitem->avw_blockNumber));
2728 6 : break;
2729 0 : default:
2730 0 : elog(WARNING, "unrecognized work item found: type %d",
2731 : workitem->avw_type);
2732 0 : break;
2733 : }
2734 :
2735 : /*
2736 : * Clear a possible query-cancel signal, to avoid a late reaction to
2737 : * an automatically-sent signal because of vacuuming the current table
2738 : * (we're done with it, so it would make no sense to cancel at this
2739 : * point.)
2740 : */
2741 6 : QueryCancelPending = false;
2742 : }
2743 0 : PG_CATCH();
2744 : {
2745 : /*
2746 : * Abort the transaction, start a new one, and proceed with the next
2747 : * table in our list.
2748 : */
2749 0 : HOLD_INTERRUPTS();
2750 0 : errcontext("processing work entry for relation \"%s.%s.%s\"",
2751 : cur_datname, cur_nspname, cur_relname);
2752 0 : EmitErrorReport();
2753 :
2754 : /* this resets ProcGlobal->statusFlags[i] too */
2755 0 : AbortOutOfAnyTransaction();
2756 0 : FlushErrorState();
2757 0 : MemoryContextReset(PortalContext);
2758 :
2759 : /* restart our transaction for the following operations */
2760 0 : StartTransactionCommand();
2761 0 : RESUME_INTERRUPTS();
2762 : }
2763 6 : PG_END_TRY();
2764 :
2765 : /* Make sure we're back in AutovacMemCxt */
2766 6 : MemoryContextSwitchTo(AutovacMemCxt);
2767 :
2768 : /* We intentionally do not set did_vacuum here */
2769 :
2770 : /* be tidy */
2771 6 : deleted2:
2772 6 : if (cur_datname)
2773 6 : pfree(cur_datname);
2774 6 : if (cur_nspname)
2775 6 : pfree(cur_nspname);
2776 6 : if (cur_relname)
2777 6 : pfree(cur_relname);
2778 6 : }
2779 :
2780 : /*
2781 : * extract_autovac_opts
2782 : *
2783 : * Given a relation's pg_class tuple, return a palloc'd copy of the
2784 : * AutoVacOpts portion of reloptions, if set; otherwise, return NULL.
2785 : *
2786 : * Note: callers do not have a relation lock on the table at this point,
2787 : * so the table could have been dropped, and its catalog rows gone, after
2788 : * we acquired the pg_class row. If pg_class had a TOAST table, this would
2789 : * be a risk; fortunately, it doesn't.
2790 : */
2791 : static AutoVacOpts *
2792 264325 : extract_autovac_opts(HeapTuple tup, TupleDesc pg_class_desc)
2793 : {
2794 : bytea *relopts;
2795 : AutoVacOpts *av;
2796 :
2797 : Assert(((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_RELATION ||
2798 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_MATVIEW ||
2799 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_TOASTVALUE);
2800 :
2801 264325 : relopts = extractRelOptions(tup, pg_class_desc, NULL);
2802 264325 : if (relopts == NULL)
2803 262068 : return NULL;
2804 :
2805 2257 : av = palloc_object(AutoVacOpts);
2806 2257 : memcpy(av, &(((StdRdOptions *) relopts)->autovacuum), sizeof(AutoVacOpts));
2807 2257 : pfree(relopts);
2808 :
2809 2257 : return av;
2810 : }
2811 :
2812 :
2813 : /*
2814 : * table_recheck_autovac
2815 : *
2816 : * Recheck whether a table still needs vacuum or analyze. Return value is a
2817 : * valid autovac_table pointer if it does, NULL otherwise.
2818 : *
2819 : * Note that the returned autovac_table does not have the name fields set.
2820 : */
2821 : static autovac_table *
2822 99337 : table_recheck_autovac(Oid relid, HTAB *table_toast_map,
2823 : TupleDesc pg_class_desc,
2824 : int effective_multixact_freeze_max_age)
2825 : {
2826 : Form_pg_class classForm;
2827 : HeapTuple classTup;
2828 : bool dovacuum;
2829 : bool doanalyze;
2830 99337 : autovac_table *tab = NULL;
2831 : bool wraparound;
2832 : AutoVacOpts *avopts;
2833 99337 : bool free_avopts = false;
2834 :
2835 : /* fetch the relation's relcache entry */
2836 99337 : classTup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2837 99337 : if (!HeapTupleIsValid(classTup))
2838 0 : return NULL;
2839 99337 : classForm = (Form_pg_class) GETSTRUCT(classTup);
2840 :
2841 : /*
2842 : * Get the applicable reloptions. If it is a TOAST table, try to get the
2843 : * main table reloptions if the toast table itself doesn't have.
2844 : */
2845 99337 : avopts = extract_autovac_opts(classTup, pg_class_desc);
2846 99337 : if (avopts)
2847 973 : free_avopts = true;
2848 98364 : else if (classForm->relkind == RELKIND_TOASTVALUE &&
2849 : table_toast_map != NULL)
2850 : {
2851 : av_relation *hentry;
2852 : bool found;
2853 :
2854 35688 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2855 35688 : if (found && hentry->ar_hasrelopts)
2856 950 : avopts = &hentry->ar_reloptions;
2857 : }
2858 :
2859 99337 : recheck_relation_needs_vacanalyze(relid, avopts, classForm,
2860 : effective_multixact_freeze_max_age,
2861 : &dovacuum, &doanalyze, &wraparound);
2862 :
2863 : /* OK, it needs something done */
2864 99337 : if (doanalyze || dovacuum)
2865 : {
2866 : int freeze_min_age;
2867 : int freeze_table_age;
2868 : int multixact_freeze_min_age;
2869 : int multixact_freeze_table_age;
2870 : int log_vacuum_min_duration;
2871 : int log_analyze_min_duration;
2872 :
2873 : /*
2874 : * Calculate the vacuum cost parameters and the freeze ages. If there
2875 : * are options set in pg_class.reloptions, use them; in the case of a
2876 : * toast table, try the main table too. Otherwise use the GUC
2877 : * defaults, autovacuum's own first and plain vacuum second.
2878 : */
2879 :
2880 : /* -1 in autovac setting means use log_autovacuum_min_duration */
2881 1923 : log_vacuum_min_duration = (avopts && avopts->log_vacuum_min_duration >= 0)
2882 : ? avopts->log_vacuum_min_duration
2883 101235 : : Log_autovacuum_min_duration;
2884 :
2885 : /* -1 in autovac setting means use log_autoanalyze_min_duration */
2886 1923 : log_analyze_min_duration = (avopts && avopts->log_analyze_min_duration >= 0)
2887 : ? avopts->log_analyze_min_duration
2888 101235 : : Log_autoanalyze_min_duration;
2889 :
2890 : /* these do not have autovacuum-specific settings */
2891 1923 : freeze_min_age = (avopts && avopts->freeze_min_age >= 0)
2892 : ? avopts->freeze_min_age
2893 101235 : : default_freeze_min_age;
2894 :
2895 1923 : freeze_table_age = (avopts && avopts->freeze_table_age >= 0)
2896 : ? avopts->freeze_table_age
2897 101235 : : default_freeze_table_age;
2898 :
2899 101235 : multixact_freeze_min_age = (avopts &&
2900 1923 : avopts->multixact_freeze_min_age >= 0)
2901 : ? avopts->multixact_freeze_min_age
2902 101235 : : default_multixact_freeze_min_age;
2903 :
2904 101235 : multixact_freeze_table_age = (avopts &&
2905 1923 : avopts->multixact_freeze_table_age >= 0)
2906 : ? avopts->multixact_freeze_table_age
2907 101235 : : default_multixact_freeze_table_age;
2908 :
2909 99312 : tab = palloc_object(autovac_table);
2910 99312 : tab->at_relid = relid;
2911 :
2912 : /*
2913 : * Select VACUUM options. Note we don't say VACOPT_PROCESS_TOAST, so
2914 : * that vacuum() skips toast relations. Also note we tell vacuum() to
2915 : * skip vac_update_datfrozenxid(); we'll do that separately.
2916 : */
2917 99312 : tab->at_params.options =
2918 99312 : (dovacuum ? (VACOPT_VACUUM |
2919 : VACOPT_PROCESS_MAIN |
2920 99312 : VACOPT_SKIP_DATABASE_STATS) : 0) |
2921 99312 : (doanalyze ? VACOPT_ANALYZE : 0) |
2922 99312 : (!wraparound ? VACOPT_SKIP_LOCKED : 0);
2923 :
2924 : /*
2925 : * index_cleanup and truncate are unspecified at first in autovacuum.
2926 : * They will be filled in with usable values using their reloptions
2927 : * (or reloption defaults) later.
2928 : */
2929 99312 : tab->at_params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
2930 99312 : tab->at_params.truncate = VACOPTVALUE_UNSPECIFIED;
2931 : /* As of now, we don't support parallel vacuum for autovacuum */
2932 99312 : tab->at_params.nworkers = -1;
2933 99312 : tab->at_params.freeze_min_age = freeze_min_age;
2934 99312 : tab->at_params.freeze_table_age = freeze_table_age;
2935 99312 : tab->at_params.multixact_freeze_min_age = multixact_freeze_min_age;
2936 99312 : tab->at_params.multixact_freeze_table_age = multixact_freeze_table_age;
2937 99312 : tab->at_params.is_wraparound = wraparound;
2938 99312 : tab->at_params.log_vacuum_min_duration = log_vacuum_min_duration;
2939 99312 : tab->at_params.log_analyze_min_duration = log_analyze_min_duration;
2940 99312 : tab->at_params.toast_parent = InvalidOid;
2941 :
2942 : /*
2943 : * Later, in vacuum_rel(), we check reloptions for any
2944 : * vacuum_max_eager_freeze_failure_rate override.
2945 : */
2946 99312 : tab->at_params.max_eager_freeze_failure_rate = vacuum_max_eager_freeze_failure_rate;
2947 99312 : tab->at_storage_param_vac_cost_limit = avopts ?
2948 99312 : avopts->vacuum_cost_limit : 0;
2949 99312 : tab->at_storage_param_vac_cost_delay = avopts ?
2950 99312 : avopts->vacuum_cost_delay : -1;
2951 99312 : tab->at_relname = NULL;
2952 99312 : tab->at_nspname = NULL;
2953 99312 : tab->at_datname = NULL;
2954 :
2955 : /*
2956 : * If any of the cost delay parameters has been set individually for
2957 : * this table, disable the balancing algorithm.
2958 : */
2959 99312 : tab->at_dobalance =
2960 101235 : !(avopts && (avopts->vacuum_cost_limit > 0 ||
2961 101235 : avopts->vacuum_cost_delay >= 0));
2962 : }
2963 :
2964 99337 : if (free_avopts)
2965 973 : pfree(avopts);
2966 99337 : heap_freetuple(classTup);
2967 99337 : return tab;
2968 : }
2969 :
2970 : /*
2971 : * recheck_relation_needs_vacanalyze
2972 : *
2973 : * Subroutine for table_recheck_autovac.
2974 : *
2975 : * Fetch the pgstat of a relation and recheck whether a relation
2976 : * needs to be vacuumed or analyzed.
2977 : */
2978 : static void
2979 99337 : recheck_relation_needs_vacanalyze(Oid relid,
2980 : AutoVacOpts *avopts,
2981 : Form_pg_class classForm,
2982 : int effective_multixact_freeze_max_age,
2983 : bool *dovacuum,
2984 : bool *doanalyze,
2985 : bool *wraparound)
2986 : {
2987 : PgStat_StatTabEntry *tabentry;
2988 : AutoVacuumScores scores;
2989 :
2990 : /* fetch the pgstat table entry */
2991 99337 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2992 : relid);
2993 :
2994 99337 : relation_needs_vacanalyze(relid, avopts, classForm, tabentry,
2995 : effective_multixact_freeze_max_age,
2996 : dovacuum, doanalyze, wraparound,
2997 : &scores);
2998 :
2999 : /* Release tabentry to avoid leakage */
3000 99337 : if (tabentry)
3001 98832 : pfree(tabentry);
3002 99337 : }
3003 :
3004 : /*
3005 : * relation_needs_vacanalyze
3006 : *
3007 : * Check whether a relation needs to be vacuumed or analyzed; return each into
3008 : * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is
3009 : * being forced because of Xid or multixact wraparound.
3010 : *
3011 : * relopts is a pointer to the AutoVacOpts options (either for itself in the
3012 : * case of a plain table, or for either itself or its parent table in the case
3013 : * of a TOAST table), NULL if none; tabentry is the pgstats entry, which can be
3014 : * NULL.
3015 : *
3016 : * A table needs to be vacuumed if the number of dead tuples exceeds a
3017 : * threshold. This threshold is calculated as
3018 : *
3019 : * threshold = vac_base_thresh + vac_scale_factor * reltuples
3020 : * if (threshold > vac_max_thresh)
3021 : * threshold = vac_max_thresh;
3022 : *
3023 : * For analyze, the analysis done is that the number of tuples inserted,
3024 : * deleted and updated since the last analyze exceeds a threshold calculated
3025 : * in the same fashion as above. Note that the cumulative stats system stores
3026 : * the number of tuples (both live and dead) that there were as of the last
3027 : * analyze. This is asymmetric to the VACUUM case.
3028 : *
3029 : * We also force vacuum if the table's relfrozenxid is more than freeze_max_age
3030 : * transactions back, and if its relminmxid is more than
3031 : * multixact_freeze_max_age multixacts back.
3032 : *
3033 : * A table whose autovacuum_enabled option is false is
3034 : * automatically skipped (unless we have to vacuum it due to freeze_max_age).
3035 : * Thus autovacuum can be disabled for specific tables. Also, when the cumulative
3036 : * stats system does not have data about a table, it will be skipped.
3037 : *
3038 : * A table whose vac_base_thresh value is < 0 takes the base value from the
3039 : * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor
3040 : * value < 0 is substituted with the value of
3041 : * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze.
3042 : *
3043 : * This function also returns scores that can be used to sort the list of
3044 : * tables to process. The idea is to have autovacuum prioritize tables that
3045 : * are furthest beyond their thresholds (e.g., a table nearing transaction ID
3046 : * wraparound should be vacuumed first). This prioritization scheme is
3047 : * certainly far from perfect; there are simply too many possibilities for any
3048 : * scoring technique to work across all workloads, and the situation might
3049 : * change significantly between the time we calculate the score and the time
3050 : * that autovacuum processes it. However, we have attempted to develop
3051 : * something that is expected to work for a large portion of workloads with
3052 : * reasonable parameter settings.
3053 : *
3054 : * The autovacuum table score is calculated as the maximum of the ratios of
3055 : * each of the table's relevant values to its threshold. For example, if the
3056 : * number of inserted tuples is 100, and the insert threshold for the table is
3057 : * 80, the insert score is 1.25. If all other scores are below that value, the
3058 : * returned score will be 1.25. The other criteria considered for the score
3059 : * are the table ages (both relfrozenxid and relminmxid) compared to the
3060 : * corresponding freeze-max-age setting, the number of updated/deleted tuples
3061 : * compared to the vacuum threshold, and the number of inserted/updated/deleted
3062 : * tuples compared to the analyze threshold.
3063 : *
3064 : * One exception to the previous paragraph is for tables nearing wraparound,
3065 : * i.e., those that have surpassed the effective failsafe ages. In that case,
3066 : * the relfrozen/relminmxid-based score is scaled aggressively so that the
3067 : * table has a decent chance of sorting to the front of the list.
3068 : *
3069 : * To adjust how strongly each component contributes to the score, the
3070 : * following parameters can be adjusted from their default of 1.0 to anywhere
3071 : * between 0.0 and 10.0 (inclusive). Setting all of these to 0.0 restores
3072 : * pre-v19 prioritization behavior:
3073 : *
3074 : * autovacuum_freeze_score_weight
3075 : * autovacuum_multixact_freeze_score_weight
3076 : * autovacuum_vacuum_score_weight
3077 : * autovacuum_vacuum_insert_score_weight
3078 : * autovacuum_analyze_score_weight
3079 : *
3080 : * The autovacuum table score is returned in scores->max. The component scores
3081 : * are also returned in the "scores" argument via the other members of the
3082 : * AutoVacuumScores struct.
3083 : */
3084 : static void
3085 264325 : relation_needs_vacanalyze(Oid relid,
3086 : AutoVacOpts *relopts,
3087 : Form_pg_class classForm,
3088 : PgStat_StatTabEntry *tabentry,
3089 : int effective_multixact_freeze_max_age,
3090 : /* output params below */
3091 : bool *dovacuum,
3092 : bool *doanalyze,
3093 : bool *wraparound,
3094 : AutoVacuumScores *scores)
3095 : {
3096 : bool force_vacuum;
3097 : bool av_enabled;
3098 :
3099 : /* constants from reloptions or GUC variables */
3100 : int vac_base_thresh,
3101 : vac_max_thresh,
3102 : vac_ins_base_thresh,
3103 : anl_base_thresh;
3104 : float4 vac_scale_factor,
3105 : vac_ins_scale_factor,
3106 : anl_scale_factor;
3107 :
3108 : /* thresholds calculated from above constants */
3109 : float4 vacthresh,
3110 : vacinsthresh,
3111 : anlthresh;
3112 :
3113 : /* number of vacuum (resp. analyze) tuples at this time */
3114 : float4 vactuples,
3115 : instuples,
3116 : anltuples;
3117 :
3118 : /* freeze parameters */
3119 : int freeze_max_age;
3120 : int multixact_freeze_max_age;
3121 : TransactionId xidForceLimit;
3122 : TransactionId relfrozenxid;
3123 : MultiXactId relminmxid;
3124 : MultiXactId multiForceLimit;
3125 :
3126 : Assert(classForm != NULL);
3127 : Assert(OidIsValid(relid));
3128 :
3129 264325 : memset(scores, 0, sizeof(AutoVacuumScores));
3130 264325 : *dovacuum = false;
3131 264325 : *doanalyze = false;
3132 :
3133 : /*
3134 : * Determine vacuum/analyze equation parameters. We have two possible
3135 : * sources: the passed reloptions (which could be a main table or a toast
3136 : * table), or the autovacuum GUC variables.
3137 : */
3138 :
3139 : /* -1 in autovac setting means use plain vacuum_scale_factor */
3140 4451 : vac_scale_factor = (relopts && relopts->vacuum_scale_factor >= 0)
3141 0 : ? relopts->vacuum_scale_factor
3142 268776 : : autovacuum_vac_scale;
3143 :
3144 4451 : vac_base_thresh = (relopts && relopts->vacuum_threshold >= 0)
3145 : ? relopts->vacuum_threshold
3146 268776 : : autovacuum_vac_thresh;
3147 :
3148 : /* -1 is used to disable max threshold */
3149 4451 : vac_max_thresh = (relopts && relopts->vacuum_max_threshold >= -1)
3150 : ? relopts->vacuum_max_threshold
3151 268776 : : autovacuum_vac_max_thresh;
3152 :
3153 4451 : vac_ins_scale_factor = (relopts && relopts->vacuum_ins_scale_factor >= 0)
3154 0 : ? relopts->vacuum_ins_scale_factor
3155 268776 : : autovacuum_vac_ins_scale;
3156 :
3157 : /* -1 is used to disable insert vacuums */
3158 4451 : vac_ins_base_thresh = (relopts && relopts->vacuum_ins_threshold >= -1)
3159 : ? relopts->vacuum_ins_threshold
3160 268776 : : autovacuum_vac_ins_thresh;
3161 :
3162 4451 : anl_scale_factor = (relopts && relopts->analyze_scale_factor >= 0)
3163 0 : ? relopts->analyze_scale_factor
3164 268776 : : autovacuum_anl_scale;
3165 :
3166 4451 : anl_base_thresh = (relopts && relopts->analyze_threshold >= 0)
3167 : ? relopts->analyze_threshold
3168 268776 : : autovacuum_anl_thresh;
3169 :
3170 4451 : freeze_max_age = (relopts && relopts->freeze_max_age >= 0)
3171 0 : ? Min(relopts->freeze_max_age, autovacuum_freeze_max_age)
3172 268776 : : autovacuum_freeze_max_age;
3173 :
3174 4451 : multixact_freeze_max_age = (relopts && relopts->multixact_freeze_max_age >= 0)
3175 0 : ? Min(relopts->multixact_freeze_max_age, effective_multixact_freeze_max_age)
3176 268776 : : effective_multixact_freeze_max_age;
3177 :
3178 264325 : av_enabled = (relopts ? relopts->enabled : true);
3179 :
3180 264325 : relfrozenxid = classForm->relfrozenxid;
3181 264325 : relminmxid = classForm->relminmxid;
3182 :
3183 : /* Force vacuum if table is at risk of wraparound */
3184 264325 : xidForceLimit = recentXid - freeze_max_age;
3185 264325 : if (xidForceLimit < FirstNormalTransactionId)
3186 0 : xidForceLimit -= FirstNormalTransactionId;
3187 528650 : force_vacuum = (TransactionIdIsNormal(relfrozenxid) &&
3188 264325 : TransactionIdPrecedes(relfrozenxid, xidForceLimit));
3189 264325 : if (!force_vacuum)
3190 : {
3191 65838 : multiForceLimit = recentMulti - multixact_freeze_max_age;
3192 65838 : if (multiForceLimit < FirstMultiXactId)
3193 0 : multiForceLimit -= FirstMultiXactId;
3194 131676 : force_vacuum = MultiXactIdIsValid(relminmxid) &&
3195 65838 : MultiXactIdPrecedes(relminmxid, multiForceLimit);
3196 : }
3197 264325 : *wraparound = force_vacuum;
3198 :
3199 : /* Update the score. */
3200 264325 : if (force_vacuum)
3201 : {
3202 : uint32 xid_age;
3203 : uint32 mxid_age;
3204 : int effective_xid_failsafe_age;
3205 : int effective_mxid_failsafe_age;
3206 :
3207 : /*
3208 : * To calculate the (M)XID age portion of the score, divide the age by
3209 : * its respective *_freeze_max_age parameter.
3210 : */
3211 198487 : xid_age = TransactionIdIsNormal(relfrozenxid) ? recentXid - relfrozenxid : 0;
3212 198487 : mxid_age = MultiXactIdIsValid(relminmxid) ? recentMulti - relminmxid : 0;
3213 :
3214 198487 : scores->xid = (double) xid_age / freeze_max_age;
3215 198487 : scores->mxid = (double) mxid_age / multixact_freeze_max_age;
3216 :
3217 : /*
3218 : * To ensure tables are given increased priority once they begin
3219 : * approaching wraparound, we scale the score aggressively if the ages
3220 : * surpass vacuum_failsafe_age or vacuum_multixact_failsafe_age.
3221 : *
3222 : * As in vacuum_xid_failsafe_check(), the effective failsafe age is no
3223 : * less than 105% the value of the respective *_freeze_max_age
3224 : * parameter. Note that per-table settings could result in a low
3225 : * score even if the table surpasses the failsafe settings. However,
3226 : * this is a strange enough corner case that we don't bother trying to
3227 : * handle it.
3228 : *
3229 : * We further adjust the effective failsafe ages with the weight
3230 : * parameters so that increasing them lowers the ages at which we
3231 : * begin scaling aggressively.
3232 : */
3233 198487 : effective_xid_failsafe_age = Max(vacuum_failsafe_age,
3234 : autovacuum_freeze_max_age * 1.05);
3235 198487 : effective_mxid_failsafe_age = Max(vacuum_multixact_failsafe_age,
3236 : autovacuum_multixact_freeze_max_age * 1.05);
3237 :
3238 198487 : if (autovacuum_freeze_score_weight > 1.0)
3239 0 : effective_xid_failsafe_age /= autovacuum_freeze_score_weight;
3240 198487 : if (autovacuum_multixact_freeze_score_weight > 1.0)
3241 0 : effective_mxid_failsafe_age /= autovacuum_multixact_freeze_score_weight;
3242 :
3243 198487 : if (xid_age >= effective_xid_failsafe_age)
3244 46608 : scores->xid = pow(scores->xid, Max(1.0, (double) xid_age / 100000000));
3245 198487 : if (mxid_age >= effective_mxid_failsafe_age)
3246 0 : scores->mxid = pow(scores->mxid, Max(1.0, (double) mxid_age / 100000000));
3247 :
3248 198487 : scores->xid *= autovacuum_freeze_score_weight;
3249 198487 : scores->mxid *= autovacuum_multixact_freeze_score_weight;
3250 :
3251 198487 : scores->max = Max(scores->xid, scores->mxid);
3252 198487 : *dovacuum = true;
3253 : }
3254 :
3255 : /* User disabled it in pg_class.reloptions? (But ignore if at risk) */
3256 264325 : if (!av_enabled && !force_vacuum)
3257 567 : return;
3258 :
3259 : /*
3260 : * If we found stats for the table, and autovacuum is currently enabled,
3261 : * make a threshold-based decision whether to vacuum and/or analyze. If
3262 : * autovacuum is currently disabled, we must be here for anti-wraparound
3263 : * vacuuming only, so don't vacuum (or analyze) anything that's not being
3264 : * forced.
3265 : */
3266 263758 : if (tabentry && AutoVacuumingActive())
3267 : {
3268 259202 : float4 pcnt_unfrozen = 1;
3269 259202 : float4 reltuples = classForm->reltuples;
3270 259202 : int32 relpages = classForm->relpages;
3271 259202 : int32 relallfrozen = classForm->relallfrozen;
3272 :
3273 259202 : vactuples = tabentry->dead_tuples;
3274 259202 : instuples = tabentry->ins_since_vacuum;
3275 259202 : anltuples = tabentry->mod_since_analyze;
3276 :
3277 : /* If the table hasn't yet been vacuumed, take reltuples as zero */
3278 259202 : if (reltuples < 0)
3279 1537 : reltuples = 0;
3280 :
3281 : /*
3282 : * If we have data for relallfrozen, calculate the unfrozen percentage
3283 : * of the table to modify insert scale factor. This helps us decide
3284 : * whether or not to vacuum an insert-heavy table based on the number
3285 : * of inserts to the more "active" part of the table.
3286 : */
3287 259202 : if (relpages > 0 && relallfrozen > 0)
3288 : {
3289 : /*
3290 : * It could be the stats were updated manually and relallfrozen >
3291 : * relpages. Clamp relallfrozen to relpages to avoid nonsensical
3292 : * calculations.
3293 : */
3294 94469 : relallfrozen = Min(relallfrozen, relpages);
3295 94469 : pcnt_unfrozen = 1 - ((float4) relallfrozen / relpages);
3296 : }
3297 :
3298 259202 : vacthresh = (float4) vac_base_thresh + vac_scale_factor * reltuples;
3299 259202 : if (vac_max_thresh >= 0 && vacthresh > (float4) vac_max_thresh)
3300 0 : vacthresh = (float4) vac_max_thresh;
3301 :
3302 259202 : vacinsthresh = (float4) vac_ins_base_thresh +
3303 259202 : vac_ins_scale_factor * reltuples * pcnt_unfrozen;
3304 259202 : anlthresh = (float4) anl_base_thresh + anl_scale_factor * reltuples;
3305 :
3306 : /*
3307 : * Determine if this table needs vacuum, and update the score if it
3308 : * does.
3309 : */
3310 259202 : if (vactuples > vacthresh)
3311 : {
3312 241 : scores->vac = (double) vactuples / Max(vacthresh, 1);
3313 241 : scores->vac *= autovacuum_vacuum_score_weight;
3314 241 : scores->max = Max(scores->max, scores->vac);
3315 241 : *dovacuum = true;
3316 : }
3317 :
3318 259202 : if (vac_ins_base_thresh >= 0 && instuples > vacinsthresh)
3319 : {
3320 196 : scores->vac_ins = (double) instuples / Max(vacinsthresh, 1);
3321 196 : scores->vac_ins *= autovacuum_vacuum_insert_score_weight;
3322 196 : scores->max = Max(scores->max, scores->vac_ins);
3323 196 : *dovacuum = true;
3324 : }
3325 :
3326 : /*
3327 : * Determine if this table needs analyze, and update the score if it
3328 : * does. Note that we don't analyze TOAST tables and pg_statistic.
3329 : */
3330 259202 : if (anltuples > anlthresh &&
3331 1090 : relid != StatisticRelationId &&
3332 1090 : classForm->relkind != RELKIND_TOASTVALUE)
3333 : {
3334 786 : scores->anl = (double) anltuples / Max(anlthresh, 1);
3335 786 : scores->anl *= autovacuum_analyze_score_weight;
3336 786 : scores->max = Max(scores->max, scores->anl);
3337 786 : *doanalyze = true;
3338 : }
3339 :
3340 259202 : if (vac_ins_base_thresh >= 0)
3341 259202 : elog(DEBUG3, "%s: vac: %.0f (thresh %.0f, score %.2f), ins: %.0f (thresh %.0f, score %.2f), anl: %.0f (thresh %.0f, score %.2f), xid score: %.2f, mxid score: %.2f",
3342 : NameStr(classForm->relname),
3343 : vactuples, vacthresh, scores->vac,
3344 : instuples, vacinsthresh, scores->vac_ins,
3345 : anltuples, anlthresh, scores->anl,
3346 : scores->xid, scores->mxid);
3347 : else
3348 0 : elog(DEBUG3, "%s: vac: %.0f (thresh %.0f, score %.2f), ins: (disabled), anl: %.0f (thresh %.0f, score %.2f), xid score: %.2f, mxid score: %.2f",
3349 : NameStr(classForm->relname),
3350 : vactuples, vacthresh, scores->vac,
3351 : anltuples, anlthresh, scores->anl,
3352 : scores->xid, scores->mxid);
3353 : }
3354 : }
3355 :
3356 : /*
3357 : * autovacuum_do_vac_analyze
3358 : * Vacuum and/or analyze the specified table
3359 : *
3360 : * We expect the caller to have switched into a memory context that won't
3361 : * disappear at transaction commit.
3362 : */
3363 : static void
3364 99312 : autovacuum_do_vac_analyze(autovac_table *tab, BufferAccessStrategy bstrategy)
3365 : {
3366 : RangeVar *rangevar;
3367 : VacuumRelation *rel;
3368 : List *rel_list;
3369 : MemoryContext vac_context;
3370 : MemoryContext old_context;
3371 :
3372 : /* Let pgstat know what we're doing */
3373 99312 : autovac_report_activity(tab);
3374 :
3375 : /* Create a context that vacuum() can use as cross-transaction storage */
3376 99312 : vac_context = AllocSetContextCreate(CurrentMemoryContext,
3377 : "Vacuum",
3378 : ALLOCSET_DEFAULT_SIZES);
3379 :
3380 : /* Set up one VacuumRelation target, identified by OID, for vacuum() */
3381 99312 : old_context = MemoryContextSwitchTo(vac_context);
3382 99312 : rangevar = makeRangeVar(tab->at_nspname, tab->at_relname, -1);
3383 99312 : rel = makeVacuumRelation(rangevar, tab->at_relid, NIL);
3384 99312 : rel_list = list_make1(rel);
3385 99312 : MemoryContextSwitchTo(old_context);
3386 :
3387 99312 : vacuum(rel_list, &tab->at_params, bstrategy, vac_context, true);
3388 :
3389 99311 : MemoryContextDelete(vac_context);
3390 99311 : }
3391 :
3392 : /*
3393 : * autovac_report_activity
3394 : * Report to pgstat what autovacuum is doing
3395 : *
3396 : * We send a SQL string corresponding to what the user would see if the
3397 : * equivalent command was to be issued manually.
3398 : *
3399 : * Note we assume that we are going to report the next command as soon as we're
3400 : * done with the current one, and exit right after the last one, so we don't
3401 : * bother to report "<IDLE>" or some such.
3402 : */
3403 : static void
3404 99312 : autovac_report_activity(autovac_table *tab)
3405 : {
3406 : #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56)
3407 : char activity[MAX_AUTOVAC_ACTIV_LEN];
3408 : int len;
3409 :
3410 : /* Report the command and possible options */
3411 99312 : if (tab->at_params.options & VACOPT_VACUUM)
3412 99103 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3413 : "autovacuum: VACUUM%s",
3414 99103 : tab->at_params.options & VACOPT_ANALYZE ? " ANALYZE" : "");
3415 : else
3416 209 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3417 : "autovacuum: ANALYZE");
3418 :
3419 : /*
3420 : * Report the qualified name of the relation.
3421 : */
3422 99312 : len = strlen(activity);
3423 :
3424 99312 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3425 : " %s.%s%s", tab->at_nspname, tab->at_relname,
3426 99312 : tab->at_params.is_wraparound ? " (to prevent wraparound)" : "");
3427 :
3428 : /* Set statement_timestamp() to current time for pg_stat_activity */
3429 99312 : SetCurrentStatementStartTimestamp();
3430 :
3431 99312 : pgstat_report_activity(STATE_RUNNING, activity);
3432 99312 : }
3433 :
3434 : /*
3435 : * autovac_report_workitem
3436 : * Report to pgstat that autovacuum is processing a work item
3437 : */
3438 : static void
3439 6 : autovac_report_workitem(AutoVacuumWorkItem *workitem,
3440 : const char *nspname, const char *relname)
3441 : {
3442 : char activity[MAX_AUTOVAC_ACTIV_LEN + 12 + 2];
3443 : char blk[12 + 2];
3444 : int len;
3445 :
3446 6 : switch (workitem->avw_type)
3447 : {
3448 6 : case AVW_BRINSummarizeRange:
3449 6 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3450 : "autovacuum: BRIN summarize");
3451 6 : break;
3452 : }
3453 :
3454 : /*
3455 : * Report the qualified name of the relation, and the block number if any
3456 : */
3457 6 : len = strlen(activity);
3458 :
3459 6 : if (BlockNumberIsValid(workitem->avw_blockNumber))
3460 6 : snprintf(blk, sizeof(blk), " %u", workitem->avw_blockNumber);
3461 : else
3462 0 : blk[0] = '\0';
3463 :
3464 6 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3465 : " %s.%s%s", nspname, relname, blk);
3466 :
3467 : /* Set statement_timestamp() to current time for pg_stat_activity */
3468 6 : SetCurrentStatementStartTimestamp();
3469 :
3470 6 : pgstat_report_activity(STATE_RUNNING, activity);
3471 6 : }
3472 :
3473 : /*
3474 : * AutoVacuumingActive
3475 : * Check GUC vars and report whether the autovacuum process should be
3476 : * running.
3477 : */
3478 : bool
3479 303847 : AutoVacuumingActive(void)
3480 : {
3481 303847 : if (!autovacuum_start_daemon || !pgstat_track_counts)
3482 5225 : return false;
3483 298622 : return true;
3484 : }
3485 :
3486 : /*
3487 : * Request one work item to the next autovacuum run processing our database.
3488 : * Return false if the request can't be recorded.
3489 : */
3490 : bool
3491 6 : AutoVacuumRequestWork(AutoVacuumWorkItemType type, Oid relationId,
3492 : BlockNumber blkno)
3493 : {
3494 : int i;
3495 6 : bool result = false;
3496 :
3497 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
3498 :
3499 : /*
3500 : * Locate an unused work item and fill it with the given data.
3501 : */
3502 21 : for (i = 0; i < NUM_WORKITEMS; i++)
3503 : {
3504 21 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
3505 :
3506 21 : if (workitem->avw_used)
3507 15 : continue;
3508 :
3509 6 : workitem->avw_used = true;
3510 6 : workitem->avw_active = false;
3511 6 : workitem->avw_type = type;
3512 6 : workitem->avw_database = MyDatabaseId;
3513 6 : workitem->avw_relation = relationId;
3514 6 : workitem->avw_blockNumber = blkno;
3515 6 : result = true;
3516 :
3517 : /* done */
3518 6 : break;
3519 : }
3520 :
3521 6 : LWLockRelease(AutovacuumLock);
3522 :
3523 6 : return result;
3524 : }
3525 :
3526 : /*
3527 : * autovac_init
3528 : * This is called at postmaster initialization.
3529 : *
3530 : * All we do here is annoy the user if he got it wrong.
3531 : */
3532 : void
3533 957 : autovac_init(void)
3534 : {
3535 957 : if (!autovacuum_start_daemon)
3536 124 : return;
3537 833 : else if (!pgstat_track_counts)
3538 0 : ereport(WARNING,
3539 : (errmsg("autovacuum not started because of misconfiguration"),
3540 : errhint("Enable the \"track_counts\" option.")));
3541 : else
3542 833 : check_av_worker_gucs();
3543 : }
3544 :
3545 : /*
3546 : * AutoVacuumShmemSize
3547 : * Compute space needed for autovacuum-related shared memory
3548 : */
3549 : Size
3550 3402 : AutoVacuumShmemSize(void)
3551 : {
3552 : Size size;
3553 :
3554 : /*
3555 : * Need the fixed struct and the array of WorkerInfoData.
3556 : */
3557 3402 : size = sizeof(AutoVacuumShmemStruct);
3558 3402 : size = MAXALIGN(size);
3559 3402 : size = add_size(size, mul_size(autovacuum_worker_slots,
3560 : sizeof(WorkerInfoData)));
3561 3402 : return size;
3562 : }
3563 :
3564 : /*
3565 : * AutoVacuumShmemInit
3566 : * Allocate and initialize autovacuum-related shared memory
3567 : */
3568 : void
3569 1185 : AutoVacuumShmemInit(void)
3570 : {
3571 : bool found;
3572 :
3573 1185 : AutoVacuumShmem = (AutoVacuumShmemStruct *)
3574 1185 : ShmemInitStruct("AutoVacuum Data",
3575 : AutoVacuumShmemSize(),
3576 : &found);
3577 :
3578 1185 : if (!IsUnderPostmaster)
3579 : {
3580 : WorkerInfo worker;
3581 : int i;
3582 :
3583 : Assert(!found);
3584 :
3585 1185 : AutoVacuumShmem->av_launcherpid = 0;
3586 1185 : dclist_init(&AutoVacuumShmem->av_freeWorkers);
3587 1185 : dlist_init(&AutoVacuumShmem->av_runningWorkers);
3588 1185 : AutoVacuumShmem->av_startingWorker = NULL;
3589 1185 : memset(AutoVacuumShmem->av_workItems, 0,
3590 : sizeof(AutoVacuumWorkItem) * NUM_WORKITEMS);
3591 :
3592 1185 : worker = (WorkerInfo) ((char *) AutoVacuumShmem +
3593 : MAXALIGN(sizeof(AutoVacuumShmemStruct)));
3594 :
3595 : /* initialize the WorkerInfo free list */
3596 14308 : for (i = 0; i < autovacuum_worker_slots; i++)
3597 : {
3598 13123 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
3599 13123 : &worker[i].wi_links);
3600 13123 : pg_atomic_init_flag(&worker[i].wi_dobalance);
3601 : }
3602 :
3603 1185 : pg_atomic_init_u32(&AutoVacuumShmem->av_nworkersForBalance, 0);
3604 :
3605 : }
3606 : else
3607 : Assert(found);
3608 1185 : }
3609 :
3610 : /*
3611 : * GUC check_hook for autovacuum_work_mem
3612 : */
3613 : bool
3614 1231 : check_autovacuum_work_mem(int *newval, void **extra, GucSource source)
3615 : {
3616 : /*
3617 : * -1 indicates fallback.
3618 : *
3619 : * If we haven't yet changed the boot_val default of -1, just let it be.
3620 : * Autovacuum will look to maintenance_work_mem instead.
3621 : */
3622 1231 : if (*newval == -1)
3623 1229 : return true;
3624 :
3625 : /*
3626 : * We clamp manually-set values to at least 64kB. Since
3627 : * maintenance_work_mem is always set to at least this value, do the same
3628 : * here.
3629 : */
3630 2 : if (*newval < 64)
3631 2 : *newval = 64;
3632 :
3633 2 : return true;
3634 : }
3635 :
3636 : /*
3637 : * Returns whether there is a free autovacuum worker slot available.
3638 : */
3639 : static bool
3640 11253 : av_worker_available(void)
3641 : {
3642 : int free_slots;
3643 : int reserved_slots;
3644 :
3645 11253 : free_slots = dclist_count(&AutoVacuumShmem->av_freeWorkers);
3646 :
3647 11253 : reserved_slots = autovacuum_worker_slots - autovacuum_max_workers;
3648 11253 : reserved_slots = Max(0, reserved_slots);
3649 :
3650 11253 : return free_slots > reserved_slots;
3651 : }
3652 :
3653 : /*
3654 : * Emits a WARNING if autovacuum_worker_slots < autovacuum_max_workers.
3655 : */
3656 : static void
3657 833 : check_av_worker_gucs(void)
3658 : {
3659 833 : if (autovacuum_worker_slots < autovacuum_max_workers)
3660 0 : ereport(WARNING,
3661 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3662 : errmsg("\"autovacuum_max_workers\" (%d) should be less than or equal to \"autovacuum_worker_slots\" (%d)",
3663 : autovacuum_max_workers, autovacuum_worker_slots),
3664 : errdetail("The server will only start up to \"autovacuum_worker_slots\" (%d) autovacuum workers at a given time.",
3665 : autovacuum_worker_slots)));
3666 833 : }
|
