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 "funcapi.h"
84 : #include "lib/ilist.h"
85 : #include "libpq/pqsignal.h"
86 : #include "miscadmin.h"
87 : #include "nodes/makefuncs.h"
88 : #include "pgstat.h"
89 : #include "postmaster/autovacuum.h"
90 : #include "postmaster/interrupt.h"
91 : #include "postmaster/postmaster.h"
92 : #include "storage/aio_subsys.h"
93 : #include "storage/bufmgr.h"
94 : #include "storage/ipc.h"
95 : #include "storage/fd.h"
96 : #include "storage/latch.h"
97 : #include "storage/lmgr.h"
98 : #include "storage/pmsignal.h"
99 : #include "storage/proc.h"
100 : #include "storage/procsignal.h"
101 : #include "storage/smgr.h"
102 : #include "storage/subsystems.h"
103 : #include "tcop/tcopprot.h"
104 : #include "utils/fmgroids.h"
105 : #include "utils/fmgrprotos.h"
106 : #include "utils/guc_hooks.h"
107 : #include "utils/injection_point.h"
108 : #include "utils/lsyscache.h"
109 : #include "utils/memutils.h"
110 : #include "utils/ps_status.h"
111 : #include "utils/rel.h"
112 : #include "utils/snapmgr.h"
113 : #include "utils/syscache.h"
114 : #include "utils/timeout.h"
115 : #include "utils/timestamp.h"
116 : #include "utils/tuplestore.h"
117 : #include "utils/wait_event.h"
118 :
119 :
120 : /*
121 : * GUC parameters
122 : */
123 : bool autovacuum_start_daemon = false;
124 : int autovacuum_worker_slots;
125 : int autovacuum_max_workers;
126 : int autovacuum_work_mem = -1;
127 : int autovacuum_naptime;
128 : int autovacuum_vac_thresh;
129 : int autovacuum_vac_max_thresh;
130 : double autovacuum_vac_scale;
131 : int autovacuum_vac_ins_thresh;
132 : double autovacuum_vac_ins_scale;
133 : int autovacuum_anl_thresh;
134 : double autovacuum_anl_scale;
135 : int autovacuum_freeze_max_age;
136 : int autovacuum_multixact_freeze_max_age;
137 : double autovacuum_freeze_score_weight = 1.0;
138 : double autovacuum_multixact_freeze_score_weight = 1.0;
139 : double autovacuum_vacuum_score_weight = 1.0;
140 : double autovacuum_vacuum_insert_score_weight = 1.0;
141 : double autovacuum_analyze_score_weight = 1.0;
142 : double autovacuum_vac_cost_delay;
143 : int autovacuum_vac_cost_limit;
144 :
145 : int Log_autovacuum_min_duration = 600000;
146 : int Log_autoanalyze_min_duration = 600000;
147 :
148 : /* the minimum allowed time between two awakenings of the launcher */
149 : #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */
150 : #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */
151 :
152 : /*
153 : * Variables to save the cost-related storage parameters for the current
154 : * relation being vacuumed by this autovacuum worker. Using these, we can
155 : * ensure we don't overwrite the values of vacuum_cost_delay and
156 : * vacuum_cost_limit after reloading the configuration file. They are
157 : * initialized to "invalid" values to indicate that no cost-related storage
158 : * parameters were specified and will be set in do_autovacuum() after checking
159 : * the storage parameters in table_recheck_autovac().
160 : */
161 : static double av_storage_param_cost_delay = -1;
162 : static int av_storage_param_cost_limit = -1;
163 :
164 : /* Flags set by signal handlers */
165 : static volatile sig_atomic_t got_SIGUSR2 = false;
166 :
167 : /* Comparison points for determining whether freeze_max_age is exceeded */
168 : static TransactionId recentXid;
169 : static MultiXactId recentMulti;
170 :
171 : /* Default freeze ages to use for autovacuum (varies by database) */
172 : static int default_freeze_min_age;
173 : static int default_freeze_table_age;
174 : static int default_multixact_freeze_min_age;
175 : static int default_multixact_freeze_table_age;
176 :
177 : /* Memory context for long-lived data */
178 : static MemoryContext AutovacMemCxt;
179 :
180 : /* struct to keep track of databases in launcher */
181 : typedef struct avl_dbase
182 : {
183 : Oid adl_datid; /* hash key -- must be first */
184 : TimestampTz adl_next_worker;
185 : int adl_score;
186 : dlist_node adl_node;
187 : } avl_dbase;
188 :
189 : /* struct to keep track of databases in worker */
190 : typedef struct avw_dbase
191 : {
192 : Oid adw_datid;
193 : char *adw_name;
194 : TransactionId adw_frozenxid;
195 : MultiXactId adw_minmulti;
196 : PgStat_StatDBEntry *adw_entry;
197 : } avw_dbase;
198 :
199 : /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */
200 : typedef struct av_relation
201 : {
202 : Oid ar_toastrelid; /* hash key - must be first */
203 : Oid ar_relid;
204 : bool ar_hasrelopts;
205 : AutoVacOpts ar_reloptions; /* copy of AutoVacOpts from the main table's
206 : * reloptions, or NULL if none */
207 : } av_relation;
208 :
209 : /* struct to keep track of tables to vacuum and/or analyze, after rechecking */
210 : typedef struct autovac_table
211 : {
212 : Oid at_relid;
213 : VacuumParams at_params;
214 : double at_storage_param_vac_cost_delay;
215 : int at_storage_param_vac_cost_limit;
216 : bool at_dobalance;
217 : char *at_relname;
218 : char *at_nspname;
219 : char *at_datname;
220 : } autovac_table;
221 :
222 : /*-------------
223 : * This struct holds information about a single worker's whereabouts. We keep
224 : * an array of these in shared memory, sized according to
225 : * autovacuum_worker_slots.
226 : *
227 : * wi_links entry into free list or running list
228 : * wi_dboid OID of the database this worker is supposed to work on
229 : * wi_tableoid OID of the table currently being vacuumed, if any
230 : * wi_sharedrel flag indicating whether table is marked relisshared
231 : * wi_proc pointer to PGPROC of the running worker, NULL if not started
232 : * wi_launchtime Time at which this worker was launched
233 : * wi_dobalance Whether this worker should be included in balance calculations
234 : *
235 : * All fields are protected by AutovacuumLock, except for wi_tableoid and
236 : * wi_sharedrel which are protected by AutovacuumScheduleLock (note these
237 : * two fields are read-only for everyone except that worker itself).
238 : *-------------
239 : */
240 : typedef struct WorkerInfoData
241 : {
242 : dlist_node wi_links;
243 : Oid wi_dboid;
244 : Oid wi_tableoid;
245 : PGPROC *wi_proc;
246 : TimestampTz wi_launchtime;
247 : pg_atomic_flag wi_dobalance;
248 : bool wi_sharedrel;
249 : } WorkerInfoData;
250 :
251 : typedef struct WorkerInfoData *WorkerInfo;
252 :
253 : /*
254 : * Possible signals received by the launcher from remote processes. These are
255 : * stored atomically in shared memory so that other processes can set them
256 : * without locking.
257 : */
258 : typedef enum
259 : {
260 : AutoVacForkFailed, /* failed trying to start a worker */
261 : AutoVacRebalance, /* rebalance the cost limits */
262 : } AutoVacuumSignal;
263 :
264 : #define AutoVacNumSignals (AutoVacRebalance + 1)
265 :
266 : /*
267 : * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This
268 : * list is mostly protected by AutovacuumLock, except that if an item is
269 : * marked 'active' other processes must not modify the work-identifying
270 : * members.
271 : */
272 : typedef struct AutoVacuumWorkItem
273 : {
274 : AutoVacuumWorkItemType avw_type;
275 : bool avw_used; /* below data is valid */
276 : bool avw_active; /* being processed */
277 : Oid avw_database;
278 : Oid avw_relation;
279 : BlockNumber avw_blockNumber;
280 : } AutoVacuumWorkItem;
281 :
282 : #define NUM_WORKITEMS 256
283 :
284 : /*-------------
285 : * The main autovacuum shmem struct. On shared memory we store this main
286 : * struct and the array of WorkerInfo structs. This struct keeps:
287 : *
288 : * av_signal set by other processes to indicate various conditions
289 : * av_launcherpid the PID of the autovacuum launcher
290 : * av_freeWorkers the WorkerInfo freelist
291 : * av_runningWorkers the WorkerInfo non-free queue
292 : * av_startingWorker pointer to WorkerInfo currently being started (cleared by
293 : * the worker itself as soon as it's up and running)
294 : * av_workItems work item array
295 : * av_nworkersForBalance the number of autovacuum workers to use when
296 : * calculating the per worker cost limit
297 : *
298 : * This struct is protected by AutovacuumLock, except for av_signal and parts
299 : * of the worker list (see above).
300 : *-------------
301 : */
302 : typedef struct
303 : {
304 : sig_atomic_t av_signal[AutoVacNumSignals];
305 : pid_t av_launcherpid;
306 : dclist_head av_freeWorkers;
307 : dlist_head av_runningWorkers;
308 : WorkerInfo av_startingWorker;
309 : AutoVacuumWorkItem av_workItems[NUM_WORKITEMS];
310 : pg_atomic_uint32 av_nworkersForBalance;
311 : } AutoVacuumShmemStruct;
312 :
313 : static AutoVacuumShmemStruct *AutoVacuumShmem;
314 :
315 : static void AutoVacuumShmemRequest(void *arg);
316 : static void AutoVacuumShmemInit(void *arg);
317 :
318 : const ShmemCallbacks AutoVacuumShmemCallbacks = {
319 : .request_fn = AutoVacuumShmemRequest,
320 : .init_fn = AutoVacuumShmemInit,
321 : };
322 :
323 : /*
324 : * the database list (of avl_dbase elements) in the launcher, and the context
325 : * that contains it
326 : */
327 : static dlist_head DatabaseList = DLIST_STATIC_INIT(DatabaseList);
328 : static MemoryContext DatabaseListCxt = NULL;
329 :
330 : /*
331 : * This struct is used by relation_needs_vacanalyze() to return the table's
332 : * score (i.e., the maximum of the component scores) as well as the component
333 : * scores themselves.
334 : */
335 : typedef struct
336 : {
337 : double max; /* maximum of all values below */
338 : double xid; /* transaction ID component */
339 : double mxid; /* multixact ID component */
340 : double vac; /* vacuum component */
341 : double vac_ins; /* vacuum insert component */
342 : double anl; /* analyze component */
343 : } AutoVacuumScores;
344 :
345 : /*
346 : * This struct is used to track and sort the list of tables to process.
347 : */
348 : typedef struct
349 : {
350 : Oid oid;
351 : double score;
352 : } TableToProcess;
353 :
354 : /*
355 : * Dummy pointer to persuade Valgrind that we've not leaked the array of
356 : * avl_dbase structs. Make it global to ensure the compiler doesn't
357 : * optimize it away.
358 : */
359 : #ifdef USE_VALGRIND
360 : extern avl_dbase *avl_dbase_array;
361 : avl_dbase *avl_dbase_array;
362 : #endif
363 :
364 : /* Pointer to my own WorkerInfo, valid on each worker */
365 : static WorkerInfo MyWorkerInfo = NULL;
366 :
367 : static Oid do_start_worker(void);
368 : static void ProcessAutoVacLauncherInterrupts(void);
369 : pg_noreturn static void AutoVacLauncherShutdown(void);
370 : static void launcher_determine_sleep(bool canlaunch, bool recursing,
371 : struct timeval *nap);
372 : static void launch_worker(TimestampTz now);
373 : static List *get_database_list(void);
374 : static void rebuild_database_list(Oid newdb);
375 : static int db_comparator(const void *a, const void *b);
376 : static void autovac_recalculate_workers_for_balance(void);
377 :
378 : static void do_autovacuum(void);
379 : static void FreeWorkerInfo(int code, Datum arg);
380 :
381 : static autovac_table *table_recheck_autovac(Oid relid, HTAB *table_toast_map,
382 : TupleDesc pg_class_desc,
383 : int effective_multixact_freeze_max_age);
384 : static void relation_needs_vacanalyze(Oid relid, AutoVacOpts *relopts,
385 : Form_pg_class classForm,
386 : int effective_multixact_freeze_max_age,
387 : int elevel,
388 : bool *dovacuum, bool *doanalyze, bool *wraparound,
389 : AutoVacuumScores *scores);
390 :
391 : static void autovacuum_do_vac_analyze(autovac_table *tab,
392 : BufferAccessStrategy bstrategy);
393 : static AutoVacOpts *extract_autovac_opts(HeapTuple tup,
394 : TupleDesc pg_class_desc);
395 : static void perform_work_item(AutoVacuumWorkItem *workitem);
396 : static void autovac_report_activity(autovac_table *tab);
397 : static void autovac_report_workitem(AutoVacuumWorkItem *workitem,
398 : const char *nspname, const char *relname);
399 : static void avl_sigusr2_handler(SIGNAL_ARGS);
400 : static bool av_worker_available(void);
401 : static void check_av_worker_gucs(void);
402 :
403 :
404 :
405 : /********************************************************************
406 : * AUTOVACUUM LAUNCHER CODE
407 : ********************************************************************/
408 :
409 : /*
410 : * Main entry point for the autovacuum launcher process.
411 : */
412 : void
413 482 : AutoVacLauncherMain(const void *startup_data, size_t startup_data_len)
414 : {
415 : sigjmp_buf local_sigjmp_buf;
416 :
417 : Assert(startup_data_len == 0);
418 :
419 : /* Release postmaster's working memory context */
420 482 : if (PostmasterContext)
421 : {
422 482 : MemoryContextDelete(PostmasterContext);
423 482 : PostmasterContext = NULL;
424 : }
425 :
426 482 : init_ps_display(NULL);
427 :
428 482 : ereport(DEBUG1,
429 : (errmsg_internal("autovacuum launcher started")));
430 :
431 482 : if (PostAuthDelay)
432 0 : pg_usleep(PostAuthDelay * 1000000L);
433 :
434 : Assert(GetProcessingMode() == InitProcessing);
435 :
436 : /*
437 : * Set up signal handlers. We operate on databases much like a regular
438 : * backend, so we use the same signal handling. See equivalent code in
439 : * tcop/postgres.c.
440 : */
441 482 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
442 482 : pqsignal(SIGINT, StatementCancelHandler);
443 482 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
444 : /* SIGQUIT handler was already set up by InitPostmasterChild */
445 :
446 482 : InitializeTimeouts(); /* establishes SIGALRM handler */
447 :
448 482 : pqsignal(SIGPIPE, PG_SIG_IGN);
449 482 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
450 482 : pqsignal(SIGUSR2, avl_sigusr2_handler);
451 482 : pqsignal(SIGFPE, FloatExceptionHandler);
452 482 : pqsignal(SIGCHLD, PG_SIG_DFL);
453 :
454 : /*
455 : * Create a per-backend PGPROC struct in shared memory. We must do this
456 : * before we can use LWLocks or access any shared memory.
457 : */
458 482 : InitProcess();
459 :
460 : /* Early initialization */
461 482 : BaseInit();
462 :
463 482 : InitPostgres(NULL, InvalidOid, NULL, InvalidOid, 0, NULL);
464 :
465 482 : SetProcessingMode(NormalProcessing);
466 :
467 : /*
468 : * Create a memory context that we will do all our work in. We do this so
469 : * that we can reset the context during error recovery and thereby avoid
470 : * possible memory leaks.
471 : */
472 482 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
473 : "Autovacuum Launcher",
474 : ALLOCSET_DEFAULT_SIZES);
475 482 : MemoryContextSwitchTo(AutovacMemCxt);
476 :
477 : /*
478 : * If an exception is encountered, processing resumes here.
479 : *
480 : * This code is a stripped down version of PostgresMain error recovery.
481 : *
482 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
483 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
484 : * signals other than SIGQUIT will be blocked until we complete error
485 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
486 : * call redundant, but it is not since InterruptPending might be set
487 : * already.
488 : */
489 482 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
490 : {
491 : /* since not using PG_TRY, must reset error stack by hand */
492 0 : error_context_stack = NULL;
493 :
494 : /* Prevents interrupts while cleaning up */
495 0 : HOLD_INTERRUPTS();
496 :
497 : /* Forget any pending QueryCancel or timeout request */
498 0 : disable_all_timeouts(false);
499 0 : QueryCancelPending = false; /* second to avoid race condition */
500 :
501 : /* Report the error to the server log */
502 0 : EmitErrorReport();
503 :
504 : /* Abort the current transaction in order to recover */
505 0 : AbortCurrentTransaction();
506 :
507 : /*
508 : * Release any other resources, for the case where we were not in a
509 : * transaction.
510 : */
511 0 : LWLockReleaseAll();
512 0 : pgstat_report_wait_end();
513 0 : pgaio_error_cleanup();
514 0 : UnlockBuffers();
515 : /* this is probably dead code, but let's be safe: */
516 0 : if (AuxProcessResourceOwner)
517 0 : ReleaseAuxProcessResources(false);
518 0 : AtEOXact_Buffers(false);
519 0 : AtEOXact_SMgr();
520 0 : AtEOXact_Files(false);
521 0 : AtEOXact_HashTables(false);
522 :
523 : /*
524 : * Now return to normal top-level context and clear ErrorContext for
525 : * next time.
526 : */
527 0 : MemoryContextSwitchTo(AutovacMemCxt);
528 0 : FlushErrorState();
529 :
530 : /* Flush any leaked data in the top-level context */
531 0 : MemoryContextReset(AutovacMemCxt);
532 :
533 : /* don't leave dangling pointers to freed memory */
534 0 : DatabaseListCxt = NULL;
535 0 : dlist_init(&DatabaseList);
536 :
537 : /* Now we can allow interrupts again */
538 0 : RESUME_INTERRUPTS();
539 :
540 : /* if in shutdown mode, no need for anything further; just go away */
541 0 : if (ShutdownRequestPending)
542 0 : AutoVacLauncherShutdown();
543 :
544 : /*
545 : * Sleep at least 1 second after any error. We don't want to be
546 : * filling the error logs as fast as we can.
547 : */
548 0 : pg_usleep(1000000L);
549 : }
550 :
551 : /* We can now handle ereport(ERROR) */
552 482 : PG_exception_stack = &local_sigjmp_buf;
553 :
554 : /* must unblock signals before calling rebuild_database_list */
555 482 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
556 :
557 : /*
558 : * Set always-secure search path. Launcher doesn't connect to a database,
559 : * so this has no effect.
560 : */
561 482 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
562 :
563 : /*
564 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
565 : * in postgresql.conf. We don't really want such a dangerous option being
566 : * applied non-interactively.
567 : */
568 482 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
569 :
570 : /*
571 : * Force settable timeouts off to avoid letting these settings prevent
572 : * regular maintenance from being executed.
573 : */
574 482 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
575 482 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
576 482 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
577 482 : SetConfigOption("idle_in_transaction_session_timeout", "0",
578 : PGC_SUSET, PGC_S_OVERRIDE);
579 :
580 : /*
581 : * Force default_transaction_isolation to READ COMMITTED. We don't want
582 : * to pay the overhead of serializable mode, nor add any risk of causing
583 : * deadlocks or delaying other transactions.
584 : */
585 482 : SetConfigOption("default_transaction_isolation", "read committed",
586 : PGC_SUSET, PGC_S_OVERRIDE);
587 :
588 : /*
589 : * Even when system is configured to use a different fetch consistency,
590 : * for autovac we always want fresh stats.
591 : */
592 482 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
593 :
594 : /*
595 : * In emergency mode, just start a worker (unless shutdown was requested)
596 : * and go away.
597 : */
598 482 : if (!AutoVacuumingActive())
599 : {
600 0 : if (!ShutdownRequestPending)
601 0 : do_start_worker();
602 0 : proc_exit(0); /* done */
603 : }
604 :
605 482 : AutoVacuumShmem->av_launcherpid = MyProcPid;
606 :
607 : /*
608 : * Create the initial database list. The invariant we want this list to
609 : * keep is that it's ordered by decreasing next_worker. As soon as an
610 : * entry is updated to a higher time, it will be moved to the front (which
611 : * is correct because the only operation is to add autovacuum_naptime to
612 : * the entry, and time always increases).
613 : */
614 482 : rebuild_database_list(InvalidOid);
615 :
616 : /* loop until shutdown request */
617 5294 : while (!ShutdownRequestPending)
618 : {
619 : struct timeval nap;
620 5293 : TimestampTz current_time = 0;
621 : bool can_launch;
622 :
623 : /*
624 : * This loop is a bit different from the normal use of WaitLatch,
625 : * because we'd like to sleep before the first launch of a child
626 : * process. So it's WaitLatch, then ResetLatch, then check for
627 : * wakening conditions.
628 : */
629 :
630 5293 : launcher_determine_sleep(av_worker_available(), false, &nap);
631 :
632 : /*
633 : * Wait until naptime expires or we get some type of signal (all the
634 : * signal handlers will wake us by calling SetLatch).
635 : */
636 5293 : (void) WaitLatch(MyLatch,
637 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
638 5293 : (nap.tv_sec * 1000L) + (nap.tv_usec / 1000L),
639 : WAIT_EVENT_AUTOVACUUM_MAIN);
640 :
641 5290 : ResetLatch(MyLatch);
642 :
643 5290 : ProcessAutoVacLauncherInterrupts();
644 :
645 : /*
646 : * a worker finished, or postmaster signaled failure to start a worker
647 : */
648 4812 : if (got_SIGUSR2)
649 : {
650 2831 : got_SIGUSR2 = false;
651 :
652 : /* rebalance cost limits, if needed */
653 2831 : if (AutoVacuumShmem->av_signal[AutoVacRebalance])
654 : {
655 1380 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
656 1380 : AutoVacuumShmem->av_signal[AutoVacRebalance] = false;
657 1380 : autovac_recalculate_workers_for_balance();
658 1380 : LWLockRelease(AutovacuumLock);
659 : }
660 :
661 2831 : if (AutoVacuumShmem->av_signal[AutoVacForkFailed])
662 : {
663 : /*
664 : * If the postmaster failed to start a new worker, we sleep
665 : * for a little while and resend the signal. The new worker's
666 : * state is still in memory, so this is sufficient. After
667 : * that, we restart the main loop.
668 : *
669 : * XXX should we put a limit to the number of times we retry?
670 : * I don't think it makes much sense, because a future start
671 : * of a worker will continue to fail in the same way.
672 : */
673 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = false;
674 0 : pg_usleep(1000000L); /* 1s */
675 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
676 1512 : continue;
677 : }
678 : }
679 :
680 : /*
681 : * There are some conditions that we need to check before trying to
682 : * start a worker. First, we need to make sure that there is a worker
683 : * slot available. Second, we need to make sure that no other worker
684 : * failed while starting up.
685 : */
686 :
687 4812 : current_time = GetCurrentTimestamp();
688 4812 : LWLockAcquire(AutovacuumLock, LW_SHARED);
689 :
690 4812 : can_launch = av_worker_available();
691 :
692 4812 : if (AutoVacuumShmem->av_startingWorker != NULL)
693 : {
694 : int waittime;
695 19 : WorkerInfo worker = AutoVacuumShmem->av_startingWorker;
696 :
697 : /*
698 : * We can't launch another worker when another one is still
699 : * starting up (or failed while doing so), so just sleep for a bit
700 : * more; that worker will wake us up again as soon as it's ready.
701 : * We will only wait autovacuum_naptime seconds (up to a maximum
702 : * of 60 seconds) for this to happen however. Note that failure
703 : * to connect to a particular database is not a problem here,
704 : * because the worker removes itself from the startingWorker
705 : * pointer before trying to connect. Problems detected by the
706 : * postmaster (like fork() failure) are also reported and handled
707 : * differently. The only problems that may cause this code to
708 : * fire are errors in the earlier sections of AutoVacWorkerMain,
709 : * before the worker removes the WorkerInfo from the
710 : * startingWorker pointer.
711 : */
712 19 : waittime = Min(autovacuum_naptime, 60) * 1000;
713 19 : if (TimestampDifferenceExceeds(worker->wi_launchtime, current_time,
714 : waittime))
715 : {
716 0 : LWLockRelease(AutovacuumLock);
717 0 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
718 :
719 : /*
720 : * No other process can put a worker in starting mode, so if
721 : * startingWorker is still INVALID after exchanging our lock,
722 : * we assume it's the same one we saw above (so we don't
723 : * recheck the launch time).
724 : */
725 0 : if (AutoVacuumShmem->av_startingWorker != NULL)
726 : {
727 0 : worker = AutoVacuumShmem->av_startingWorker;
728 0 : worker->wi_dboid = InvalidOid;
729 0 : worker->wi_tableoid = InvalidOid;
730 0 : worker->wi_sharedrel = false;
731 0 : worker->wi_proc = NULL;
732 0 : worker->wi_launchtime = 0;
733 0 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
734 : &worker->wi_links);
735 0 : AutoVacuumShmem->av_startingWorker = NULL;
736 0 : ereport(WARNING,
737 : errmsg("autovacuum worker took too long to start; canceled"));
738 : }
739 : }
740 : else
741 19 : can_launch = false;
742 : }
743 4812 : LWLockRelease(AutovacuumLock); /* either shared or exclusive */
744 :
745 : /* if we can't do anything, just go back to sleep */
746 4812 : if (!can_launch)
747 1512 : continue;
748 :
749 : /* We're OK to start a new worker */
750 :
751 3300 : if (dlist_is_empty(&DatabaseList))
752 : {
753 : /*
754 : * Special case when the list is empty: start a worker right away.
755 : * This covers the initial case, when no database is in pgstats
756 : * (thus the list is empty). Note that the constraints in
757 : * launcher_determine_sleep keep us from starting workers too
758 : * quickly (at most once every autovacuum_naptime when the list is
759 : * empty).
760 : */
761 1 : launch_worker(current_time);
762 : }
763 : else
764 : {
765 : /*
766 : * because rebuild_database_list constructs a list with most
767 : * distant adl_next_worker first, we obtain our database from the
768 : * tail of the list.
769 : */
770 : avl_dbase *avdb;
771 :
772 3299 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
773 :
774 : /*
775 : * launch a worker if next_worker is right now or it is in the
776 : * past
777 : */
778 3299 : if (TimestampDifferenceExceeds(avdb->adl_next_worker,
779 : current_time, 0))
780 1418 : launch_worker(current_time);
781 : }
782 : }
783 :
784 1 : AutoVacLauncherShutdown();
785 : }
786 :
787 : /*
788 : * Process any new interrupts.
789 : */
790 : static void
791 5290 : ProcessAutoVacLauncherInterrupts(void)
792 : {
793 : /* the normal shutdown case */
794 5290 : if (ShutdownRequestPending)
795 477 : AutoVacLauncherShutdown();
796 :
797 4813 : if (ConfigReloadPending)
798 : {
799 52 : int autovacuum_max_workers_prev = autovacuum_max_workers;
800 :
801 52 : ConfigReloadPending = false;
802 52 : ProcessConfigFile(PGC_SIGHUP);
803 :
804 : /* shutdown requested in config file? */
805 52 : if (!AutoVacuumingActive())
806 1 : AutoVacLauncherShutdown();
807 :
808 : /*
809 : * If autovacuum_max_workers changed, emit a WARNING if
810 : * autovacuum_worker_slots < autovacuum_max_workers. If it didn't
811 : * change, skip this to avoid too many repeated log messages.
812 : */
813 51 : if (autovacuum_max_workers_prev != autovacuum_max_workers)
814 0 : check_av_worker_gucs();
815 :
816 : /* rebuild the list in case the naptime changed */
817 51 : rebuild_database_list(InvalidOid);
818 : }
819 :
820 : /* Process barrier events */
821 4812 : if (ProcSignalBarrierPending)
822 73 : ProcessProcSignalBarrier();
823 :
824 : /* Perform logging of memory contexts of this process */
825 4812 : if (LogMemoryContextPending)
826 0 : ProcessLogMemoryContextInterrupt();
827 :
828 : /* Process sinval catchup interrupts that happened while sleeping */
829 4812 : ProcessCatchupInterrupt();
830 4812 : }
831 :
832 : /*
833 : * Perform a normal exit from the autovac launcher.
834 : */
835 : static void
836 479 : AutoVacLauncherShutdown(void)
837 : {
838 479 : ereport(DEBUG1,
839 : (errmsg_internal("autovacuum launcher shutting down")));
840 479 : AutoVacuumShmem->av_launcherpid = 0;
841 :
842 479 : proc_exit(0); /* done */
843 : }
844 :
845 : /*
846 : * Determine the time to sleep, based on the database list.
847 : *
848 : * The "canlaunch" parameter indicates whether we can start a worker right now,
849 : * for example due to the workers being all busy. If this is false, we will
850 : * cause a long sleep, which will be interrupted when a worker exits.
851 : */
852 : static void
853 5365 : launcher_determine_sleep(bool canlaunch, bool recursing, struct timeval *nap)
854 : {
855 : /*
856 : * We sleep until the next scheduled vacuum. We trust that when the
857 : * database list was built, care was taken so that no entries have times
858 : * in the past; if the first entry has too close a next_worker value, or a
859 : * time in the past, we will sleep a small nominal time.
860 : */
861 5365 : if (!canlaunch)
862 : {
863 2774 : nap->tv_sec = autovacuum_naptime;
864 2774 : nap->tv_usec = 0;
865 : }
866 2591 : else if (!dlist_is_empty(&DatabaseList))
867 : {
868 2569 : TimestampTz current_time = GetCurrentTimestamp();
869 : TimestampTz next_wakeup;
870 : avl_dbase *avdb;
871 : long secs;
872 : int usecs;
873 :
874 2569 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
875 :
876 2569 : next_wakeup = avdb->adl_next_worker;
877 2569 : TimestampDifference(current_time, next_wakeup, &secs, &usecs);
878 :
879 2569 : nap->tv_sec = secs;
880 2569 : nap->tv_usec = usecs;
881 : }
882 : else
883 : {
884 : /* list is empty, sleep for whole autovacuum_naptime seconds */
885 22 : nap->tv_sec = autovacuum_naptime;
886 22 : nap->tv_usec = 0;
887 : }
888 :
889 : /*
890 : * If the result is exactly zero, it means a database had an entry with
891 : * time in the past. Rebuild the list so that the databases are evenly
892 : * distributed again, and recalculate the time to sleep. This can happen
893 : * if there are more tables needing vacuum than workers, and they all take
894 : * longer to vacuum than autovacuum_naptime.
895 : *
896 : * We only recurse once. rebuild_database_list should always return times
897 : * in the future, but it seems best not to trust too much on that.
898 : */
899 5365 : if (nap->tv_sec == 0 && nap->tv_usec == 0 && !recursing)
900 : {
901 72 : rebuild_database_list(InvalidOid);
902 72 : launcher_determine_sleep(canlaunch, true, nap);
903 72 : return;
904 : }
905 :
906 : /* The smallest time we'll allow the launcher to sleep. */
907 5293 : if (nap->tv_sec <= 0 && nap->tv_usec <= MIN_AUTOVAC_SLEEPTIME * 1000)
908 : {
909 211 : nap->tv_sec = 0;
910 211 : nap->tv_usec = MIN_AUTOVAC_SLEEPTIME * 1000;
911 : }
912 :
913 : /*
914 : * If the sleep time is too large, clamp it to an arbitrary maximum (plus
915 : * any fractional seconds, for simplicity). This avoids an essentially
916 : * infinite sleep in strange cases like the system clock going backwards a
917 : * few years.
918 : */
919 5293 : if (nap->tv_sec > MAX_AUTOVAC_SLEEPTIME)
920 9 : nap->tv_sec = MAX_AUTOVAC_SLEEPTIME;
921 : }
922 :
923 : /*
924 : * Build an updated DatabaseList. It must only contain databases that appear
925 : * in pgstats, and must be sorted by next_worker from highest to lowest,
926 : * distributed regularly across the next autovacuum_naptime interval.
927 : *
928 : * Receives the Oid of the database that made this list be generated (we call
929 : * this the "new" database, because when the database was already present on
930 : * the list, we expect that this function is not called at all). The
931 : * preexisting list, if any, will be used to preserve the order of the
932 : * databases in the autovacuum_naptime period. The new database is put at the
933 : * end of the interval. The actual values are not saved, which should not be
934 : * much of a problem.
935 : */
936 : static void
937 622 : rebuild_database_list(Oid newdb)
938 : {
939 : List *dblist;
940 : ListCell *cell;
941 : MemoryContext newcxt;
942 : MemoryContext oldcxt;
943 : MemoryContext tmpcxt;
944 : HASHCTL hctl;
945 : int score;
946 : int nelems;
947 : HTAB *dbhash;
948 : dlist_iter iter;
949 :
950 622 : newcxt = AllocSetContextCreate(AutovacMemCxt,
951 : "Autovacuum database list",
952 : ALLOCSET_DEFAULT_SIZES);
953 622 : tmpcxt = AllocSetContextCreate(newcxt,
954 : "Autovacuum database list (tmp)",
955 : ALLOCSET_DEFAULT_SIZES);
956 622 : oldcxt = MemoryContextSwitchTo(tmpcxt);
957 :
958 : /*
959 : * Implementing this is not as simple as it sounds, because we need to put
960 : * the new database at the end of the list; next the databases that were
961 : * already on the list, and finally (at the tail of the list) all the
962 : * other databases that are not on the existing list.
963 : *
964 : * To do this, we build an empty hash table of scored databases. We will
965 : * start with the lowest score (zero) for the new database, then
966 : * increasing scores for the databases in the existing list, in order, and
967 : * lastly increasing scores for all databases gotten via
968 : * get_database_list() that are not already on the hash.
969 : *
970 : * Then we will put all the hash elements into an array, sort the array by
971 : * score, and finally put the array elements into the new doubly linked
972 : * list.
973 : */
974 622 : hctl.keysize = sizeof(Oid);
975 622 : hctl.entrysize = sizeof(avl_dbase);
976 622 : hctl.hcxt = tmpcxt;
977 622 : dbhash = hash_create("autovacuum db hash", 20, &hctl, /* magic number here
978 : * FIXME */
979 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
980 :
981 : /* start by inserting the new database */
982 622 : score = 0;
983 622 : if (OidIsValid(newdb))
984 : {
985 : avl_dbase *db;
986 : PgStat_StatDBEntry *entry;
987 :
988 : /* only consider this database if it has a pgstat entry */
989 17 : entry = pgstat_fetch_stat_dbentry(newdb);
990 17 : if (entry != NULL)
991 : {
992 : /* we assume it isn't found because the hash was just created */
993 14 : db = hash_search(dbhash, &newdb, HASH_ENTER, NULL);
994 :
995 : /* hash_search already filled in the key */
996 14 : db->adl_score = score++;
997 : /* next_worker is filled in later */
998 : }
999 : }
1000 :
1001 : /* Now insert the databases from the existing list */
1002 955 : dlist_foreach(iter, &DatabaseList)
1003 : {
1004 333 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1005 : avl_dbase *db;
1006 : bool found;
1007 : PgStat_StatDBEntry *entry;
1008 :
1009 : /*
1010 : * skip databases with no stat entries -- in particular, this gets rid
1011 : * of dropped databases
1012 : */
1013 333 : entry = pgstat_fetch_stat_dbentry(avdb->adl_datid);
1014 333 : if (entry == NULL)
1015 0 : continue;
1016 :
1017 333 : db = hash_search(dbhash, &(avdb->adl_datid), HASH_ENTER, &found);
1018 :
1019 333 : if (!found)
1020 : {
1021 : /* hash_search already filled in the key */
1022 333 : db->adl_score = score++;
1023 : /* next_worker is filled in later */
1024 : }
1025 : }
1026 :
1027 : /* finally, insert all qualifying databases not previously inserted */
1028 622 : dblist = get_database_list();
1029 2805 : foreach(cell, dblist)
1030 : {
1031 2183 : avw_dbase *avdb = lfirst(cell);
1032 : avl_dbase *db;
1033 : bool found;
1034 : PgStat_StatDBEntry *entry;
1035 :
1036 : /* only consider databases with a pgstat entry */
1037 2183 : entry = pgstat_fetch_stat_dbentry(avdb->adw_datid);
1038 2183 : if (entry == NULL)
1039 1118 : continue;
1040 :
1041 1065 : db = hash_search(dbhash, &(avdb->adw_datid), HASH_ENTER, &found);
1042 : /* only update the score if the database was not already on the hash */
1043 1065 : if (!found)
1044 : {
1045 : /* hash_search already filled in the key */
1046 718 : db->adl_score = score++;
1047 : /* next_worker is filled in later */
1048 : }
1049 : }
1050 622 : nelems = score;
1051 :
1052 : /* from here on, the allocated memory belongs to the new list */
1053 622 : MemoryContextSwitchTo(newcxt);
1054 622 : dlist_init(&DatabaseList);
1055 :
1056 622 : if (nelems > 0)
1057 : {
1058 : TimestampTz current_time;
1059 : int millis_increment;
1060 : avl_dbase *dbary;
1061 : avl_dbase *db;
1062 : HASH_SEQ_STATUS seq;
1063 : int i;
1064 :
1065 : /* put all the hash elements into an array */
1066 601 : dbary = palloc(nelems * sizeof(avl_dbase));
1067 : /* keep Valgrind quiet */
1068 : #ifdef USE_VALGRIND
1069 : avl_dbase_array = dbary;
1070 : #endif
1071 :
1072 601 : i = 0;
1073 601 : hash_seq_init(&seq, dbhash);
1074 1666 : while ((db = hash_seq_search(&seq)) != NULL)
1075 1065 : memcpy(&(dbary[i++]), db, sizeof(avl_dbase));
1076 :
1077 : /* sort the array */
1078 601 : qsort(dbary, nelems, sizeof(avl_dbase), db_comparator);
1079 :
1080 : /*
1081 : * Determine the time interval between databases in the schedule. If
1082 : * we see that the configured naptime would take us to sleep times
1083 : * lower than our min sleep time (which launcher_determine_sleep is
1084 : * coded not to allow), silently use a larger naptime (but don't touch
1085 : * the GUC variable).
1086 : */
1087 601 : millis_increment = 1000.0 * autovacuum_naptime / nelems;
1088 601 : if (millis_increment <= MIN_AUTOVAC_SLEEPTIME)
1089 0 : millis_increment = MIN_AUTOVAC_SLEEPTIME * 1.1;
1090 :
1091 601 : current_time = GetCurrentTimestamp();
1092 :
1093 : /*
1094 : * move the elements from the array into the dlist, setting the
1095 : * next_worker while walking the array
1096 : */
1097 1666 : for (i = 0; i < nelems; i++)
1098 : {
1099 1065 : db = &(dbary[i]);
1100 :
1101 1065 : current_time = TimestampTzPlusMilliseconds(current_time,
1102 : millis_increment);
1103 1065 : db->adl_next_worker = current_time;
1104 :
1105 : /* later elements should go closer to the head of the list */
1106 1065 : dlist_push_head(&DatabaseList, &db->adl_node);
1107 : }
1108 : }
1109 :
1110 : /* all done, clean up memory */
1111 622 : if (DatabaseListCxt != NULL)
1112 140 : MemoryContextDelete(DatabaseListCxt);
1113 622 : MemoryContextDelete(tmpcxt);
1114 622 : DatabaseListCxt = newcxt;
1115 622 : MemoryContextSwitchTo(oldcxt);
1116 622 : }
1117 :
1118 : /* qsort comparator for avl_dbase, using adl_score */
1119 : static int
1120 697 : db_comparator(const void *a, const void *b)
1121 : {
1122 1394 : return pg_cmp_s32(((const avl_dbase *) a)->adl_score,
1123 697 : ((const avl_dbase *) b)->adl_score);
1124 : }
1125 :
1126 : /*
1127 : * do_start_worker
1128 : *
1129 : * Bare-bones procedure for starting an autovacuum worker from the launcher.
1130 : * It determines what database to work on, sets up shared memory stuff and
1131 : * signals postmaster to start the worker. It fails gracefully if invoked when
1132 : * autovacuum_workers are already active.
1133 : *
1134 : * Return value is the OID of the database that the worker is going to process,
1135 : * or InvalidOid if no worker was actually started.
1136 : */
1137 : static Oid
1138 1419 : do_start_worker(void)
1139 : {
1140 : List *dblist;
1141 : ListCell *cell;
1142 : TransactionId xidForceLimit;
1143 : MultiXactId multiForceLimit;
1144 : bool for_xid_wrap;
1145 : bool for_multi_wrap;
1146 : avw_dbase *avdb;
1147 : TimestampTz current_time;
1148 1419 : bool skipit = false;
1149 1419 : Oid retval = InvalidOid;
1150 : MemoryContext tmpcxt,
1151 : oldcxt;
1152 :
1153 : /* return quickly when there are no free workers */
1154 1419 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1155 1419 : if (!av_worker_available())
1156 : {
1157 0 : LWLockRelease(AutovacuumLock);
1158 0 : return InvalidOid;
1159 : }
1160 1419 : LWLockRelease(AutovacuumLock);
1161 :
1162 : /*
1163 : * Create and switch to a temporary context to avoid leaking the memory
1164 : * allocated for the database list.
1165 : */
1166 1419 : tmpcxt = AllocSetContextCreate(CurrentMemoryContext,
1167 : "Autovacuum start worker (tmp)",
1168 : ALLOCSET_DEFAULT_SIZES);
1169 1419 : oldcxt = MemoryContextSwitchTo(tmpcxt);
1170 :
1171 : /* Get a list of databases */
1172 1419 : dblist = get_database_list();
1173 :
1174 : /*
1175 : * Determine the oldest datfrozenxid/relfrozenxid that we will allow to
1176 : * pass without forcing a vacuum. (This limit can be tightened for
1177 : * particular tables, but not loosened.)
1178 : */
1179 1419 : recentXid = ReadNextTransactionId();
1180 1419 : xidForceLimit = recentXid - autovacuum_freeze_max_age;
1181 : /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */
1182 : /* this can cause the limit to go backwards by 3, but that's OK */
1183 1419 : if (xidForceLimit < FirstNormalTransactionId)
1184 0 : xidForceLimit -= FirstNormalTransactionId;
1185 :
1186 : /* Also determine the oldest datminmxid we will consider. */
1187 1419 : recentMulti = ReadNextMultiXactId();
1188 1419 : multiForceLimit = recentMulti - MultiXactMemberFreezeThreshold();
1189 1419 : if (multiForceLimit < FirstMultiXactId)
1190 0 : multiForceLimit -= FirstMultiXactId;
1191 :
1192 : /*
1193 : * Choose a database to connect to. We pick the database that was least
1194 : * recently auto-vacuumed, or one that needs vacuuming to prevent Xid
1195 : * wraparound-related data loss. If any db at risk of Xid wraparound is
1196 : * found, we pick the one with oldest datfrozenxid, independently of
1197 : * autovacuum times; similarly we pick the one with the oldest datminmxid
1198 : * if any is in MultiXactId wraparound. Note that those in Xid wraparound
1199 : * danger are given more priority than those in multi wraparound danger.
1200 : *
1201 : * Note that a database with no stats entry is not considered, except for
1202 : * Xid wraparound purposes. The theory is that if no one has ever
1203 : * connected to it since the stats were last initialized, it doesn't need
1204 : * vacuuming.
1205 : *
1206 : * XXX This could be improved if we had more info about whether it needs
1207 : * vacuuming before connecting to it. Perhaps look through the pgstats
1208 : * data for the database's tables? One idea is to keep track of the
1209 : * number of new and dead tuples per database in pgstats. However it
1210 : * isn't clear how to construct a metric that measures that and not cause
1211 : * starvation for less busy databases.
1212 : */
1213 1419 : avdb = NULL;
1214 1419 : for_xid_wrap = false;
1215 1419 : for_multi_wrap = false;
1216 1419 : current_time = GetCurrentTimestamp();
1217 5741 : foreach(cell, dblist)
1218 : {
1219 4322 : avw_dbase *tmp = lfirst(cell);
1220 : dlist_iter iter;
1221 :
1222 : /* Check to see if this one is at risk of wraparound */
1223 4322 : if (TransactionIdPrecedes(tmp->adw_frozenxid, xidForceLimit))
1224 : {
1225 3995 : if (avdb == NULL ||
1226 1589 : TransactionIdPrecedes(tmp->adw_frozenxid,
1227 : avdb->adw_frozenxid))
1228 899 : avdb = tmp;
1229 2406 : for_xid_wrap = true;
1230 3336 : continue;
1231 : }
1232 1916 : else if (for_xid_wrap)
1233 56 : continue; /* ignore not-at-risk DBs */
1234 1860 : else if (MultiXactIdPrecedes(tmp->adw_minmulti, multiForceLimit))
1235 : {
1236 0 : if (avdb == NULL ||
1237 0 : MultiXactIdPrecedes(tmp->adw_minmulti, avdb->adw_minmulti))
1238 0 : avdb = tmp;
1239 0 : for_multi_wrap = true;
1240 0 : continue;
1241 : }
1242 1860 : else if (for_multi_wrap)
1243 0 : continue; /* ignore not-at-risk DBs */
1244 :
1245 : /* Find pgstat entry if any */
1246 1860 : tmp->adw_entry = pgstat_fetch_stat_dbentry(tmp->adw_datid);
1247 :
1248 : /*
1249 : * Skip a database with no pgstat entry; it means it hasn't seen any
1250 : * activity.
1251 : */
1252 1860 : if (!tmp->adw_entry)
1253 65 : continue;
1254 :
1255 : /*
1256 : * Also, skip a database that appears on the database list as having
1257 : * been processed recently (less than autovacuum_naptime seconds ago).
1258 : * We do this so that we don't select a database which we just
1259 : * selected, but that pgstat hasn't gotten around to updating the last
1260 : * autovacuum time yet.
1261 : */
1262 1795 : skipit = false;
1263 :
1264 3733 : dlist_reverse_foreach(iter, &DatabaseList)
1265 : {
1266 3699 : avl_dbase *dbp = dlist_container(avl_dbase, adl_node, iter.cur);
1267 :
1268 3699 : if (dbp->adl_datid == tmp->adw_datid)
1269 : {
1270 : /*
1271 : * Skip this database if its next_worker value falls between
1272 : * the current time and the current time plus naptime.
1273 : */
1274 1761 : if (!TimestampDifferenceExceeds(dbp->adl_next_worker,
1275 809 : current_time, 0) &&
1276 809 : !TimestampDifferenceExceeds(current_time,
1277 : dbp->adl_next_worker,
1278 : autovacuum_naptime * 1000))
1279 809 : skipit = true;
1280 :
1281 1761 : break;
1282 : }
1283 : }
1284 1795 : if (skipit)
1285 809 : continue;
1286 :
1287 : /*
1288 : * Remember the db with oldest autovac time. (If we are here, both
1289 : * tmp->entry and db->entry must be non-null.)
1290 : */
1291 986 : if (avdb == NULL ||
1292 385 : tmp->adw_entry->last_autovac_time < avdb->adw_entry->last_autovac_time)
1293 719 : avdb = tmp;
1294 : }
1295 :
1296 : /* Found a database -- process it */
1297 1419 : if (avdb != NULL)
1298 : {
1299 : WorkerInfo worker;
1300 : dlist_node *wptr;
1301 :
1302 1418 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1303 :
1304 : /*
1305 : * Get a worker entry from the freelist. We checked above, so there
1306 : * really should be a free slot.
1307 : */
1308 1418 : wptr = dclist_pop_head_node(&AutoVacuumShmem->av_freeWorkers);
1309 :
1310 1418 : worker = dlist_container(WorkerInfoData, wi_links, wptr);
1311 1418 : worker->wi_dboid = avdb->adw_datid;
1312 1418 : worker->wi_proc = NULL;
1313 1418 : worker->wi_launchtime = GetCurrentTimestamp();
1314 :
1315 1418 : AutoVacuumShmem->av_startingWorker = worker;
1316 :
1317 1418 : LWLockRelease(AutovacuumLock);
1318 :
1319 1418 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
1320 :
1321 1418 : retval = avdb->adw_datid;
1322 : }
1323 1 : else if (skipit)
1324 : {
1325 : /*
1326 : * If we skipped all databases on the list, rebuild it, because it
1327 : * probably contains a dropped database.
1328 : */
1329 0 : rebuild_database_list(InvalidOid);
1330 : }
1331 :
1332 1419 : MemoryContextSwitchTo(oldcxt);
1333 1419 : MemoryContextDelete(tmpcxt);
1334 :
1335 1419 : return retval;
1336 : }
1337 :
1338 : /*
1339 : * launch_worker
1340 : *
1341 : * Wrapper for starting a worker from the launcher. Besides actually starting
1342 : * it, update the database list to reflect the next time that another one will
1343 : * need to be started on the selected database. The actual database choice is
1344 : * left to do_start_worker.
1345 : *
1346 : * This routine is also expected to insert an entry into the database list if
1347 : * the selected database was previously absent from the list.
1348 : */
1349 : static void
1350 1419 : launch_worker(TimestampTz now)
1351 : {
1352 : Oid dbid;
1353 : dlist_iter iter;
1354 :
1355 1419 : dbid = do_start_worker();
1356 1419 : if (OidIsValid(dbid))
1357 : {
1358 1418 : bool found = false;
1359 :
1360 : /*
1361 : * Walk the database list and update the corresponding entry. If the
1362 : * database is not on the list, we'll recreate the list.
1363 : */
1364 2891 : dlist_foreach(iter, &DatabaseList)
1365 : {
1366 2874 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1367 :
1368 2874 : if (avdb->adl_datid == dbid)
1369 : {
1370 1401 : found = true;
1371 :
1372 : /*
1373 : * add autovacuum_naptime seconds to the current time, and use
1374 : * that as the new "next_worker" field for this database.
1375 : */
1376 1401 : avdb->adl_next_worker =
1377 1401 : TimestampTzPlusMilliseconds(now, autovacuum_naptime * 1000);
1378 :
1379 1401 : dlist_move_head(&DatabaseList, iter.cur);
1380 1401 : break;
1381 : }
1382 : }
1383 :
1384 : /*
1385 : * If the database was not present in the database list, we rebuild
1386 : * the list. It's possible that the database does not get into the
1387 : * list anyway, for example if it's a database that doesn't have a
1388 : * pgstat entry, but this is not a problem because we don't want to
1389 : * schedule workers regularly into those in any case.
1390 : */
1391 1418 : if (!found)
1392 17 : rebuild_database_list(dbid);
1393 : }
1394 1419 : }
1395 :
1396 : /*
1397 : * Called from postmaster to signal a failure to fork a process to become
1398 : * worker. The postmaster should kill(SIGUSR2) the launcher shortly
1399 : * after calling this function.
1400 : */
1401 : void
1402 0 : AutoVacWorkerFailed(void)
1403 : {
1404 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = true;
1405 0 : }
1406 :
1407 : /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */
1408 : static void
1409 2832 : avl_sigusr2_handler(SIGNAL_ARGS)
1410 : {
1411 2832 : got_SIGUSR2 = true;
1412 2832 : SetLatch(MyLatch);
1413 2832 : }
1414 :
1415 :
1416 : /********************************************************************
1417 : * AUTOVACUUM WORKER CODE
1418 : ********************************************************************/
1419 :
1420 : /*
1421 : * Main entry point for autovacuum worker processes.
1422 : */
1423 : void
1424 1425 : AutoVacWorkerMain(const void *startup_data, size_t startup_data_len)
1425 : {
1426 : sigjmp_buf local_sigjmp_buf;
1427 : Oid dbid;
1428 :
1429 : Assert(startup_data_len == 0);
1430 :
1431 : /* Release postmaster's working memory context */
1432 1425 : if (PostmasterContext)
1433 : {
1434 1425 : MemoryContextDelete(PostmasterContext);
1435 1425 : PostmasterContext = NULL;
1436 : }
1437 :
1438 1425 : init_ps_display(NULL);
1439 :
1440 : Assert(GetProcessingMode() == InitProcessing);
1441 :
1442 : /*
1443 : * Set up signal handlers. We operate on databases much like a regular
1444 : * backend, so we use the same signal handling. See equivalent code in
1445 : * tcop/postgres.c.
1446 : */
1447 1425 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1448 :
1449 : /*
1450 : * SIGINT is used to signal canceling the current table's vacuum; SIGTERM
1451 : * means abort and exit cleanly, and SIGQUIT means abandon ship.
1452 : */
1453 1425 : pqsignal(SIGINT, StatementCancelHandler);
1454 1425 : pqsignal(SIGTERM, die);
1455 : /* SIGQUIT handler was already set up by InitPostmasterChild */
1456 :
1457 1425 : InitializeTimeouts(); /* establishes SIGALRM handler */
1458 :
1459 1425 : pqsignal(SIGPIPE, PG_SIG_IGN);
1460 1425 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1461 1425 : pqsignal(SIGUSR2, PG_SIG_IGN);
1462 1425 : pqsignal(SIGFPE, FloatExceptionHandler);
1463 1425 : pqsignal(SIGCHLD, PG_SIG_DFL);
1464 :
1465 : /*
1466 : * Create a per-backend PGPROC struct in shared memory. We must do this
1467 : * before we can use LWLocks or access any shared memory.
1468 : */
1469 1425 : InitProcess();
1470 :
1471 : /* Early initialization */
1472 1425 : BaseInit();
1473 :
1474 : /*
1475 : * If an exception is encountered, processing resumes here.
1476 : *
1477 : * Unlike most auxiliary processes, we don't attempt to continue
1478 : * processing after an error; we just clean up and exit. The autovac
1479 : * launcher is responsible for spawning another worker later.
1480 : *
1481 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
1482 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
1483 : * signals other than SIGQUIT will be blocked until we exit. It might
1484 : * seem that this policy makes the HOLD_INTERRUPTS() call redundant, but
1485 : * it is not since InterruptPending might be set already.
1486 : */
1487 1425 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1488 : {
1489 : /* since not using PG_TRY, must reset error stack by hand */
1490 0 : error_context_stack = NULL;
1491 :
1492 : /* Prevents interrupts while cleaning up */
1493 0 : HOLD_INTERRUPTS();
1494 :
1495 : /* Report the error to the server log */
1496 0 : EmitErrorReport();
1497 :
1498 : /*
1499 : * We can now go away. Note that because we called InitProcess, a
1500 : * callback was registered to do ProcKill, which will clean up
1501 : * necessary state.
1502 : */
1503 0 : proc_exit(0);
1504 : }
1505 :
1506 : /* We can now handle ereport(ERROR) */
1507 1425 : PG_exception_stack = &local_sigjmp_buf;
1508 :
1509 1425 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
1510 :
1511 : /*
1512 : * Set always-secure search path, so malicious users can't redirect user
1513 : * code (e.g. pg_index.indexprs). (That code runs in a
1514 : * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not
1515 : * take control of the entire autovacuum worker in any case.)
1516 : */
1517 1425 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1518 :
1519 : /*
1520 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
1521 : * in postgresql.conf. We don't really want such a dangerous option being
1522 : * applied non-interactively.
1523 : */
1524 1425 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
1525 :
1526 : /*
1527 : * Force settable timeouts off to avoid letting these settings prevent
1528 : * regular maintenance from being executed.
1529 : */
1530 1425 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1531 1425 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1532 1425 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1533 1425 : SetConfigOption("idle_in_transaction_session_timeout", "0",
1534 : PGC_SUSET, PGC_S_OVERRIDE);
1535 :
1536 : /*
1537 : * Force default_transaction_isolation to READ COMMITTED. We don't want
1538 : * to pay the overhead of serializable mode, nor add any risk of causing
1539 : * deadlocks or delaying other transactions.
1540 : */
1541 1425 : SetConfigOption("default_transaction_isolation", "read committed",
1542 : PGC_SUSET, PGC_S_OVERRIDE);
1543 :
1544 : /*
1545 : * Force synchronous replication off to allow regular maintenance even if
1546 : * we are waiting for standbys to connect. This is important to ensure we
1547 : * aren't blocked from performing anti-wraparound tasks.
1548 : */
1549 1425 : if (synchronous_commit > SYNCHRONOUS_COMMIT_LOCAL_FLUSH)
1550 1425 : SetConfigOption("synchronous_commit", "local",
1551 : PGC_SUSET, PGC_S_OVERRIDE);
1552 :
1553 : /*
1554 : * Even when system is configured to use a different fetch consistency,
1555 : * for autovac we always want fresh stats.
1556 : */
1557 1425 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
1558 :
1559 : /*
1560 : * Get the info about the database we're going to work on.
1561 : */
1562 1425 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1563 :
1564 : /*
1565 : * beware of startingWorker being INVALID; this should normally not
1566 : * happen, but if a worker fails after forking and before this, the
1567 : * launcher might have decided to remove it from the queue and start
1568 : * again.
1569 : */
1570 1425 : if (AutoVacuumShmem->av_startingWorker != NULL)
1571 : {
1572 1425 : MyWorkerInfo = AutoVacuumShmem->av_startingWorker;
1573 1425 : dbid = MyWorkerInfo->wi_dboid;
1574 1425 : MyWorkerInfo->wi_proc = MyProc;
1575 :
1576 : /* insert into the running list */
1577 1425 : dlist_push_head(&AutoVacuumShmem->av_runningWorkers,
1578 1425 : &MyWorkerInfo->wi_links);
1579 :
1580 : /*
1581 : * remove from the "starting" pointer, so that the launcher can start
1582 : * a new worker if required
1583 : */
1584 1425 : AutoVacuumShmem->av_startingWorker = NULL;
1585 1425 : LWLockRelease(AutovacuumLock);
1586 :
1587 1425 : on_shmem_exit(FreeWorkerInfo, 0);
1588 :
1589 : /* wake up the launcher */
1590 1425 : if (AutoVacuumShmem->av_launcherpid != 0)
1591 1425 : kill(AutoVacuumShmem->av_launcherpid, SIGUSR2);
1592 : }
1593 : else
1594 : {
1595 : /* no worker entry for me, go away */
1596 0 : elog(WARNING, "autovacuum worker started without a worker entry");
1597 0 : dbid = InvalidOid;
1598 0 : LWLockRelease(AutovacuumLock);
1599 : }
1600 :
1601 1425 : if (OidIsValid(dbid))
1602 : {
1603 : char dbname[NAMEDATALEN];
1604 :
1605 : /*
1606 : * Report autovac startup to the cumulative stats system. We
1607 : * deliberately do this before InitPostgres, so that the
1608 : * last_autovac_time will get updated even if the connection attempt
1609 : * fails. This is to prevent autovac from getting "stuck" repeatedly
1610 : * selecting an unopenable database, rather than making any progress
1611 : * on stuff it can connect to.
1612 : */
1613 1425 : pgstat_report_autovac(dbid);
1614 :
1615 : /*
1616 : * Connect to the selected database, specifying no particular user,
1617 : * and ignoring datallowconn. Collect the database's name for
1618 : * display.
1619 : *
1620 : * Note: if we have selected a just-deleted database (due to using
1621 : * stale stats info), we'll fail and exit here.
1622 : */
1623 1425 : InitPostgres(NULL, dbid, NULL, InvalidOid,
1624 : INIT_PG_OVERRIDE_ALLOW_CONNS,
1625 : dbname);
1626 1425 : SetProcessingMode(NormalProcessing);
1627 1425 : set_ps_display(dbname);
1628 1425 : ereport(DEBUG1,
1629 : (errmsg_internal("autovacuum: processing database \"%s\"", dbname)));
1630 :
1631 1425 : if (PostAuthDelay)
1632 0 : pg_usleep(PostAuthDelay * 1000000L);
1633 :
1634 : /* And do an appropriate amount of work */
1635 1425 : recentXid = ReadNextTransactionId();
1636 1425 : recentMulti = ReadNextMultiXactId();
1637 1425 : do_autovacuum();
1638 : }
1639 :
1640 : /* All done, go away */
1641 1423 : proc_exit(0);
1642 : }
1643 :
1644 : /*
1645 : * Return a WorkerInfo to the free list
1646 : */
1647 : static void
1648 1425 : FreeWorkerInfo(int code, Datum arg)
1649 : {
1650 1425 : if (MyWorkerInfo != NULL)
1651 : {
1652 1425 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1653 :
1654 1425 : dlist_delete(&MyWorkerInfo->wi_links);
1655 1425 : MyWorkerInfo->wi_dboid = InvalidOid;
1656 1425 : MyWorkerInfo->wi_tableoid = InvalidOid;
1657 1425 : MyWorkerInfo->wi_sharedrel = false;
1658 1425 : MyWorkerInfo->wi_proc = NULL;
1659 1425 : MyWorkerInfo->wi_launchtime = 0;
1660 1425 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
1661 1425 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
1662 1425 : &MyWorkerInfo->wi_links);
1663 : /* not mine anymore */
1664 1425 : MyWorkerInfo = NULL;
1665 :
1666 : /*
1667 : * now that we're inactive, cause a rebalancing of the surviving
1668 : * workers
1669 : */
1670 1425 : AutoVacuumShmem->av_signal[AutoVacRebalance] = true;
1671 1425 : LWLockRelease(AutovacuumLock);
1672 : }
1673 1425 : }
1674 :
1675 : /*
1676 : * Update vacuum cost-based delay-related parameters for autovacuum workers and
1677 : * backends executing VACUUM or ANALYZE using the value of relevant GUCs and
1678 : * global state. This must be called during setup for vacuum and after every
1679 : * config reload to ensure up-to-date values.
1680 : */
1681 : void
1682 205670 : VacuumUpdateCosts(void)
1683 : {
1684 205670 : if (MyWorkerInfo)
1685 : {
1686 196989 : if (av_storage_param_cost_delay >= 0)
1687 0 : vacuum_cost_delay = av_storage_param_cost_delay;
1688 196989 : else if (autovacuum_vac_cost_delay >= 0)
1689 196989 : vacuum_cost_delay = autovacuum_vac_cost_delay;
1690 : else
1691 : /* fall back to VacuumCostDelay */
1692 0 : vacuum_cost_delay = VacuumCostDelay;
1693 :
1694 196989 : AutoVacuumUpdateCostLimit();
1695 : }
1696 : else
1697 : {
1698 : /* Must be explicit VACUUM or ANALYZE or parallel autovacuum worker */
1699 8681 : vacuum_cost_delay = VacuumCostDelay;
1700 8681 : vacuum_cost_limit = VacuumCostLimit;
1701 : }
1702 :
1703 : /*
1704 : * If configuration changes are allowed to impact VacuumCostActive, make
1705 : * sure it is updated.
1706 : */
1707 205670 : if (VacuumFailsafeActive)
1708 : Assert(!VacuumCostActive);
1709 205670 : else if (vacuum_cost_delay > 0)
1710 196992 : VacuumCostActive = true;
1711 : else
1712 : {
1713 8678 : VacuumCostActive = false;
1714 8678 : VacuumCostBalance = 0;
1715 : }
1716 :
1717 : /*
1718 : * Since the cost logging requires a lock, avoid rendering the log message
1719 : * in case we are using a message level where the log wouldn't be emitted.
1720 : */
1721 205670 : if (MyWorkerInfo && message_level_is_interesting(DEBUG2))
1722 : {
1723 : Oid dboid,
1724 : tableoid;
1725 :
1726 : Assert(!LWLockHeldByMe(AutovacuumLock));
1727 :
1728 5 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1729 5 : dboid = MyWorkerInfo->wi_dboid;
1730 5 : tableoid = MyWorkerInfo->wi_tableoid;
1731 5 : LWLockRelease(AutovacuumLock);
1732 :
1733 5 : elog(DEBUG2,
1734 : "Autovacuum VacuumUpdateCosts(db=%u, rel=%u, dobalance=%s, cost_limit=%d, cost_delay=%g active=%s failsafe=%s)",
1735 : dboid, tableoid, pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance) ? "no" : "yes",
1736 : vacuum_cost_limit, vacuum_cost_delay,
1737 : vacuum_cost_delay > 0 ? "yes" : "no",
1738 : VacuumFailsafeActive ? "yes" : "no");
1739 : }
1740 205670 : }
1741 :
1742 : /*
1743 : * Update vacuum_cost_limit with the correct value for an autovacuum worker,
1744 : * given the value of other relevant cost limit parameters and the number of
1745 : * workers across which the limit must be balanced. Autovacuum workers must
1746 : * call this regularly in case av_nworkersForBalance has been updated by
1747 : * another worker or by the autovacuum launcher. They must also call it after a
1748 : * config reload.
1749 : */
1750 : void
1751 200124 : AutoVacuumUpdateCostLimit(void)
1752 : {
1753 200124 : if (!MyWorkerInfo)
1754 5 : return;
1755 :
1756 : /*
1757 : * note: in cost_limit, zero also means use value from elsewhere, because
1758 : * zero is not a valid value.
1759 : */
1760 :
1761 200119 : if (av_storage_param_cost_limit > 0)
1762 0 : vacuum_cost_limit = av_storage_param_cost_limit;
1763 : else
1764 : {
1765 : int nworkers_for_balance;
1766 :
1767 200119 : if (autovacuum_vac_cost_limit > 0)
1768 1 : vacuum_cost_limit = autovacuum_vac_cost_limit;
1769 : else
1770 200118 : vacuum_cost_limit = VacuumCostLimit;
1771 :
1772 : /* Only balance limit if no cost-related storage parameters specified */
1773 200119 : if (pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance))
1774 0 : return;
1775 :
1776 : Assert(vacuum_cost_limit > 0);
1777 :
1778 200119 : nworkers_for_balance = pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1779 :
1780 : /* There is at least 1 autovac worker (this worker) */
1781 200119 : if (nworkers_for_balance <= 0)
1782 0 : elog(ERROR, "nworkers_for_balance must be > 0");
1783 :
1784 200119 : vacuum_cost_limit = Max(vacuum_cost_limit / nworkers_for_balance, 1);
1785 : }
1786 : }
1787 :
1788 : /*
1789 : * autovac_recalculate_workers_for_balance
1790 : * Recalculate the number of workers to consider, given cost-related
1791 : * storage parameters and the current number of active workers.
1792 : *
1793 : * Caller must hold the AutovacuumLock in at least shared mode to access
1794 : * worker->wi_proc.
1795 : */
1796 : static void
1797 99874 : autovac_recalculate_workers_for_balance(void)
1798 : {
1799 : dlist_iter iter;
1800 : int orig_nworkers_for_balance;
1801 99874 : int nworkers_for_balance = 0;
1802 :
1803 : Assert(LWLockHeldByMe(AutovacuumLock));
1804 :
1805 99874 : orig_nworkers_for_balance =
1806 99874 : pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1807 :
1808 317575 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1809 : {
1810 217701 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1811 :
1812 435402 : if (worker->wi_proc == NULL ||
1813 217701 : pg_atomic_unlocked_test_flag(&worker->wi_dobalance))
1814 2911 : continue;
1815 :
1816 214790 : nworkers_for_balance++;
1817 : }
1818 :
1819 99874 : if (nworkers_for_balance != orig_nworkers_for_balance)
1820 1862 : pg_atomic_write_u32(&AutoVacuumShmem->av_nworkersForBalance,
1821 : nworkers_for_balance);
1822 99874 : }
1823 :
1824 : /*
1825 : * get_database_list
1826 : * Return a list of all databases found in pg_database.
1827 : *
1828 : * The list and associated data is allocated in the caller's memory context,
1829 : * which is in charge of ensuring that it's properly cleaned up afterwards.
1830 : *
1831 : * Note: this is the only function in which the autovacuum launcher uses a
1832 : * transaction. Although we aren't attached to any particular database and
1833 : * therefore can't access most catalogs, we do have enough infrastructure
1834 : * to do a seqscan on pg_database.
1835 : */
1836 : static List *
1837 2041 : get_database_list(void)
1838 : {
1839 2041 : List *dblist = NIL;
1840 : Relation rel;
1841 : TableScanDesc scan;
1842 : HeapTuple tup;
1843 : MemoryContext resultcxt;
1844 :
1845 : /* This is the context that we will allocate our output data in */
1846 2041 : resultcxt = CurrentMemoryContext;
1847 :
1848 : /*
1849 : * Start a transaction so we can access pg_database.
1850 : */
1851 2041 : StartTransactionCommand();
1852 :
1853 2041 : rel = table_open(DatabaseRelationId, AccessShareLock);
1854 2041 : scan = table_beginscan_catalog(rel, 0, NULL);
1855 :
1856 8551 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1857 : {
1858 6510 : Form_pg_database pgdatabase = (Form_pg_database) GETSTRUCT(tup);
1859 : avw_dbase *avdb;
1860 : MemoryContext oldcxt;
1861 :
1862 : /*
1863 : * If database has partially been dropped, we can't, nor need to,
1864 : * vacuum it.
1865 : */
1866 6510 : if (database_is_invalid_form(pgdatabase))
1867 : {
1868 5 : elog(DEBUG2,
1869 : "autovacuum: skipping invalid database \"%s\"",
1870 : NameStr(pgdatabase->datname));
1871 5 : continue;
1872 : }
1873 :
1874 : /*
1875 : * Allocate our results in the caller's context, not the
1876 : * transaction's. We do this inside the loop, and restore the original
1877 : * context at the end, so that leaky things like heap_getnext() are
1878 : * not called in a potentially long-lived context.
1879 : */
1880 6505 : oldcxt = MemoryContextSwitchTo(resultcxt);
1881 :
1882 6505 : avdb = palloc_object(avw_dbase);
1883 :
1884 6505 : avdb->adw_datid = pgdatabase->oid;
1885 6505 : avdb->adw_name = pstrdup(NameStr(pgdatabase->datname));
1886 6505 : avdb->adw_frozenxid = pgdatabase->datfrozenxid;
1887 6505 : avdb->adw_minmulti = pgdatabase->datminmxid;
1888 : /* this gets set later: */
1889 6505 : avdb->adw_entry = NULL;
1890 :
1891 6505 : dblist = lappend(dblist, avdb);
1892 6505 : MemoryContextSwitchTo(oldcxt);
1893 : }
1894 :
1895 2041 : table_endscan(scan);
1896 2041 : table_close(rel, AccessShareLock);
1897 :
1898 2041 : CommitTransactionCommand();
1899 :
1900 : /* Be sure to restore caller's memory context */
1901 2041 : MemoryContextSwitchTo(resultcxt);
1902 :
1903 2041 : return dblist;
1904 : }
1905 :
1906 : /*
1907 : * List comparator for TableToProcess. Note that this sorts the tables based
1908 : * on their scores in descending order.
1909 : */
1910 : static int
1911 101797 : TableToProcessComparator(const ListCell *a, const ListCell *b)
1912 : {
1913 101797 : TableToProcess *t1 = (TableToProcess *) lfirst(a);
1914 101797 : TableToProcess *t2 = (TableToProcess *) lfirst(b);
1915 :
1916 101797 : return (t2->score < t1->score) ? -1 : (t2->score > t1->score) ? 1 : 0;
1917 : }
1918 :
1919 : /*
1920 : * Process a database table-by-table
1921 : *
1922 : * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in
1923 : * order not to ignore shutdown commands for too long.
1924 : */
1925 : static void
1926 1425 : do_autovacuum(void)
1927 : {
1928 : Relation classRel;
1929 : HeapTuple tuple;
1930 : TableScanDesc relScan;
1931 : Form_pg_database dbForm;
1932 1425 : List *tables_to_process = NIL;
1933 1425 : List *orphan_oids = NIL;
1934 : HASHCTL ctl;
1935 : HTAB *table_toast_map;
1936 : ListCell *volatile cell;
1937 : BufferAccessStrategy bstrategy;
1938 : ScanKeyData key;
1939 : TupleDesc pg_class_desc;
1940 : int effective_multixact_freeze_max_age;
1941 1425 : bool did_vacuum = false;
1942 1425 : bool found_concurrent_worker = false;
1943 : int i;
1944 :
1945 : /*
1946 : * StartTransactionCommand and CommitTransactionCommand will automatically
1947 : * switch to other contexts. We need this one to keep the list of
1948 : * relations to vacuum/analyze across transactions.
1949 : */
1950 1425 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
1951 : "Autovacuum worker",
1952 : ALLOCSET_DEFAULT_SIZES);
1953 1425 : MemoryContextSwitchTo(AutovacMemCxt);
1954 :
1955 : /* Start a transaction so our commands have one to play into. */
1956 1425 : StartTransactionCommand();
1957 :
1958 : /*
1959 : * This injection point is put in a transaction block to work with a wait
1960 : * that uses a condition variable.
1961 : */
1962 1425 : INJECTION_POINT("autovacuum-worker-start", NULL);
1963 :
1964 : /*
1965 : * Compute the multixact age for which freezing is urgent. This is
1966 : * normally autovacuum_multixact_freeze_max_age, but may be less if
1967 : * multixact members are bloated.
1968 : */
1969 1424 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1970 :
1971 : /*
1972 : * Find the pg_database entry and select the default freeze ages. We use
1973 : * zero in template and nonconnectable databases, else the system-wide
1974 : * default.
1975 : */
1976 1424 : tuple = SearchSysCache1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
1977 1424 : if (!HeapTupleIsValid(tuple))
1978 0 : elog(ERROR, "cache lookup failed for database %u", MyDatabaseId);
1979 1424 : dbForm = (Form_pg_database) GETSTRUCT(tuple);
1980 :
1981 1424 : if (dbForm->datistemplate || !dbForm->datallowconn)
1982 : {
1983 462 : default_freeze_min_age = 0;
1984 462 : default_freeze_table_age = 0;
1985 462 : default_multixact_freeze_min_age = 0;
1986 462 : default_multixact_freeze_table_age = 0;
1987 : }
1988 : else
1989 : {
1990 962 : default_freeze_min_age = vacuum_freeze_min_age;
1991 962 : default_freeze_table_age = vacuum_freeze_table_age;
1992 962 : default_multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1993 962 : default_multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1994 : }
1995 :
1996 1424 : ReleaseSysCache(tuple);
1997 :
1998 : /* StartTransactionCommand changed elsewhere */
1999 1424 : MemoryContextSwitchTo(AutovacMemCxt);
2000 :
2001 1424 : classRel = table_open(RelationRelationId, AccessShareLock);
2002 :
2003 : /* create a copy so we can use it after closing pg_class */
2004 1424 : pg_class_desc = CreateTupleDescCopy(RelationGetDescr(classRel));
2005 :
2006 : /* create hash table for toast <-> main relid mapping */
2007 1424 : ctl.keysize = sizeof(Oid);
2008 1424 : ctl.entrysize = sizeof(av_relation);
2009 :
2010 1424 : table_toast_map = hash_create("TOAST to main relid map",
2011 : 100,
2012 : &ctl,
2013 : HASH_ELEM | HASH_BLOBS);
2014 :
2015 : /*
2016 : * Scan pg_class to determine which tables to vacuum.
2017 : *
2018 : * We do this in two passes: on the first one we collect the list of plain
2019 : * relations and materialized views, and on the second one we collect
2020 : * TOAST tables. The reason for doing the second pass is that during it we
2021 : * want to use the main relation's pg_class.reloptions entry if the TOAST
2022 : * table does not have any, and we cannot obtain it unless we know
2023 : * beforehand what's the main table OID.
2024 : *
2025 : * We need to check TOAST tables separately because in cases with short,
2026 : * wide tables there might be proportionally much more activity in the
2027 : * TOAST table than in its parent.
2028 : */
2029 1424 : relScan = table_beginscan_catalog(classRel, 0, NULL);
2030 :
2031 : /*
2032 : * On the first pass, we collect main tables to vacuum, and also the main
2033 : * table relid to TOAST relid mapping.
2034 : */
2035 655501 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2036 : {
2037 654077 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2038 : AutoVacOpts *relopts;
2039 : Oid relid;
2040 : bool dovacuum;
2041 : bool doanalyze;
2042 : bool wraparound;
2043 : AutoVacuumScores scores;
2044 :
2045 654077 : if (classForm->relkind != RELKIND_RELATION &&
2046 546553 : classForm->relkind != RELKIND_MATVIEW)
2047 546512 : continue;
2048 :
2049 107571 : relid = classForm->oid;
2050 :
2051 : /*
2052 : * Check if it is a temp table (presumably, of some other backend's).
2053 : * We cannot safely process other backends' temp tables.
2054 : */
2055 107571 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2056 : {
2057 : /*
2058 : * We just ignore it if the owning backend is still active and
2059 : * using the temporary schema. Also, for safety, ignore it if the
2060 : * namespace doesn't exist or isn't a temp namespace after all.
2061 : */
2062 6 : if (checkTempNamespaceStatus(classForm->relnamespace) == TEMP_NAMESPACE_IDLE)
2063 : {
2064 : /*
2065 : * The table seems to be orphaned -- although it might be that
2066 : * the owning backend has already deleted it and exited; our
2067 : * pg_class scan snapshot is not necessarily up-to-date
2068 : * anymore, so we could be looking at a committed-dead entry.
2069 : * Remember it so we can try to delete it later.
2070 : */
2071 0 : orphan_oids = lappend_oid(orphan_oids, relid);
2072 : }
2073 6 : continue;
2074 : }
2075 :
2076 : /* Fetch reloptions and the pgstat entry for this table */
2077 107565 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2078 :
2079 : /* Check if it needs vacuum or analyze */
2080 107565 : relation_needs_vacanalyze(relid, relopts, classForm,
2081 : effective_multixact_freeze_max_age,
2082 : DEBUG3,
2083 : &dovacuum, &doanalyze, &wraparound,
2084 : &scores);
2085 :
2086 : /* Relations that need work are added to tables_to_process */
2087 107565 : if (dovacuum || doanalyze)
2088 : {
2089 63356 : TableToProcess *table = palloc_object(TableToProcess);
2090 :
2091 63356 : table->oid = relid;
2092 63356 : table->score = scores.max;
2093 63356 : tables_to_process = lappend(tables_to_process, table);
2094 : }
2095 :
2096 : /*
2097 : * Remember TOAST associations for the second pass. Note: we must do
2098 : * this whether or not the table is going to be vacuumed, because we
2099 : * don't automatically vacuum toast tables along the parent table.
2100 : */
2101 107565 : if (OidIsValid(classForm->reltoastrelid))
2102 : {
2103 : av_relation *hentry;
2104 : bool found;
2105 :
2106 121010 : hentry = hash_search(table_toast_map,
2107 60505 : &classForm->reltoastrelid,
2108 : HASH_ENTER, &found);
2109 :
2110 60505 : if (!found)
2111 : {
2112 : /* hash_search already filled in the key */
2113 60505 : hentry->ar_relid = relid;
2114 60505 : hentry->ar_hasrelopts = false;
2115 60505 : if (relopts != NULL)
2116 : {
2117 1244 : hentry->ar_hasrelopts = true;
2118 1244 : memcpy(&hentry->ar_reloptions, relopts,
2119 : sizeof(AutoVacOpts));
2120 : }
2121 : }
2122 : }
2123 :
2124 : /* Release stuff to avoid per-relation leakage */
2125 107565 : if (relopts)
2126 1289 : pfree(relopts);
2127 : }
2128 :
2129 1424 : table_endscan(relScan);
2130 :
2131 : /* second pass: check TOAST tables */
2132 1424 : ScanKeyInit(&key,
2133 : Anum_pg_class_relkind,
2134 : BTEqualStrategyNumber, F_CHAREQ,
2135 : CharGetDatum(RELKIND_TOASTVALUE));
2136 :
2137 1424 : relScan = table_beginscan_catalog(classRel, 1, &key);
2138 61930 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2139 : {
2140 60506 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2141 : Oid relid;
2142 : AutoVacOpts *relopts;
2143 60506 : bool free_relopts = false;
2144 : bool dovacuum;
2145 : bool doanalyze;
2146 : bool wraparound;
2147 : AutoVacuumScores scores;
2148 :
2149 : /*
2150 : * We cannot safely process other backends' temp tables, so skip 'em.
2151 : */
2152 60506 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2153 1 : continue;
2154 :
2155 60505 : relid = classForm->oid;
2156 :
2157 : /*
2158 : * fetch reloptions -- if this toast table does not have them, try the
2159 : * main rel
2160 : */
2161 60505 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2162 60505 : if (relopts)
2163 2 : free_relopts = true;
2164 : else
2165 : {
2166 : av_relation *hentry;
2167 : bool found;
2168 :
2169 60503 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2170 60503 : if (found && hentry->ar_hasrelopts)
2171 1242 : relopts = &hentry->ar_reloptions;
2172 : }
2173 :
2174 60505 : relation_needs_vacanalyze(relid, relopts, classForm,
2175 : effective_multixact_freeze_max_age,
2176 : DEBUG3,
2177 : &dovacuum, &doanalyze, &wraparound,
2178 : &scores);
2179 :
2180 : /* ignore analyze for toast tables */
2181 60505 : if (dovacuum)
2182 : {
2183 35674 : TableToProcess *table = palloc_object(TableToProcess);
2184 :
2185 35674 : table->oid = relid;
2186 35674 : table->score = scores.max;
2187 35674 : tables_to_process = lappend(tables_to_process, table);
2188 : }
2189 :
2190 : /* Release stuff to avoid leakage */
2191 60505 : if (free_relopts)
2192 2 : pfree(relopts);
2193 : }
2194 :
2195 1424 : table_endscan(relScan);
2196 1424 : table_close(classRel, AccessShareLock);
2197 :
2198 : /*
2199 : * Recheck orphan temporary tables, and if they still seem orphaned, drop
2200 : * them. We'll eat a transaction per dropped table, which might seem
2201 : * excessive, but we should only need to do anything as a result of a
2202 : * previous backend crash, so this should not happen often enough to
2203 : * justify "optimizing". Using separate transactions ensures that we
2204 : * don't bloat the lock table if there are many temp tables to be dropped,
2205 : * and it ensures that we don't lose work if a deletion attempt fails.
2206 : */
2207 1424 : foreach(cell, orphan_oids)
2208 : {
2209 0 : Oid relid = lfirst_oid(cell);
2210 : Form_pg_class classForm;
2211 : ObjectAddress object;
2212 :
2213 : /*
2214 : * Check for user-requested abort.
2215 : */
2216 0 : CHECK_FOR_INTERRUPTS();
2217 :
2218 : /*
2219 : * Try to lock the table. If we can't get the lock immediately,
2220 : * somebody else is using (or dropping) the table, so it's not our
2221 : * concern anymore. Having the lock prevents race conditions below.
2222 : */
2223 0 : if (!ConditionalLockRelationOid(relid, AccessExclusiveLock))
2224 0 : continue;
2225 :
2226 : /*
2227 : * Re-fetch the pg_class tuple and re-check whether it still seems to
2228 : * be an orphaned temp table. If it's not there or no longer the same
2229 : * relation, ignore it.
2230 : */
2231 0 : tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2232 0 : if (!HeapTupleIsValid(tuple))
2233 : {
2234 : /* be sure to drop useless lock so we don't bloat lock table */
2235 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2236 0 : continue;
2237 : }
2238 0 : classForm = (Form_pg_class) GETSTRUCT(tuple);
2239 :
2240 : /*
2241 : * Make all the same tests made in the loop above. In event of OID
2242 : * counter wraparound, the pg_class entry we have now might be
2243 : * completely unrelated to the one we saw before.
2244 : */
2245 0 : if (!((classForm->relkind == RELKIND_RELATION ||
2246 0 : classForm->relkind == RELKIND_MATVIEW) &&
2247 0 : classForm->relpersistence == RELPERSISTENCE_TEMP))
2248 : {
2249 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2250 0 : continue;
2251 : }
2252 :
2253 0 : if (checkTempNamespaceStatus(classForm->relnamespace) != TEMP_NAMESPACE_IDLE)
2254 : {
2255 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2256 0 : continue;
2257 : }
2258 :
2259 : /*
2260 : * Try to lock the temp namespace, too. Even though we have lock on
2261 : * the table itself, there's a risk of deadlock against an incoming
2262 : * backend trying to clean out the temp namespace, in case this table
2263 : * has dependencies (such as sequences) that the backend's
2264 : * performDeletion call might visit in a different order. If we can
2265 : * get AccessShareLock on the namespace, that's sufficient to ensure
2266 : * we're not running concurrently with RemoveTempRelations. If we
2267 : * can't, back off and let RemoveTempRelations do its thing.
2268 : */
2269 0 : if (!ConditionalLockDatabaseObject(NamespaceRelationId,
2270 : classForm->relnamespace, 0,
2271 : AccessShareLock))
2272 : {
2273 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2274 0 : continue;
2275 : }
2276 :
2277 : /* OK, let's delete it */
2278 0 : ereport(LOG,
2279 : (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"",
2280 : get_database_name(MyDatabaseId),
2281 : get_namespace_name(classForm->relnamespace),
2282 : NameStr(classForm->relname))));
2283 :
2284 : /*
2285 : * Deletion might involve TOAST table access, so ensure we have a
2286 : * valid snapshot.
2287 : */
2288 0 : PushActiveSnapshot(GetTransactionSnapshot());
2289 :
2290 0 : object.classId = RelationRelationId;
2291 0 : object.objectId = relid;
2292 0 : object.objectSubId = 0;
2293 0 : performDeletion(&object, DROP_CASCADE,
2294 : PERFORM_DELETION_INTERNAL |
2295 : PERFORM_DELETION_QUIETLY |
2296 : PERFORM_DELETION_SKIP_EXTENSIONS);
2297 :
2298 : /*
2299 : * To commit the deletion, end current transaction and start a new
2300 : * one. Note this also releases the locks we took.
2301 : */
2302 0 : PopActiveSnapshot();
2303 0 : CommitTransactionCommand();
2304 0 : StartTransactionCommand();
2305 :
2306 : /* StartTransactionCommand changed current memory context */
2307 0 : MemoryContextSwitchTo(AutovacMemCxt);
2308 : }
2309 :
2310 : /*
2311 : * In case list_sort() would modify the list even when all the scores are
2312 : * 0.0, skip sorting if all the weight parameters are set to 0.0. This is
2313 : * probably not necessary, but we want to ensure folks have a guaranteed
2314 : * escape hatch from the scoring system.
2315 : */
2316 1424 : if (autovacuum_freeze_score_weight != 0.0 ||
2317 0 : autovacuum_multixact_freeze_score_weight != 0.0 ||
2318 0 : autovacuum_vacuum_score_weight != 0.0 ||
2319 0 : autovacuum_vacuum_insert_score_weight != 0.0 ||
2320 0 : autovacuum_analyze_score_weight != 0.0)
2321 1424 : list_sort(tables_to_process, TableToProcessComparator);
2322 :
2323 : /*
2324 : * Optionally, create a buffer access strategy object for VACUUM to use.
2325 : * We use the same BufferAccessStrategy object for all tables VACUUMed by
2326 : * this worker to prevent autovacuum from blowing out shared buffers.
2327 : *
2328 : * VacuumBufferUsageLimit being set to 0 results in
2329 : * GetAccessStrategyWithSize returning NULL, effectively meaning we can
2330 : * use up to all of shared buffers.
2331 : *
2332 : * If we later enter failsafe mode on any of the tables being vacuumed, we
2333 : * will cease use of the BufferAccessStrategy only for that table.
2334 : *
2335 : * XXX should we consider adding code to adjust the size of this if
2336 : * VacuumBufferUsageLimit changes?
2337 : */
2338 1424 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, VacuumBufferUsageLimit);
2339 :
2340 : /*
2341 : * create a memory context to act as fake PortalContext, so that the
2342 : * contexts created in the vacuum code are cleaned up for each table.
2343 : */
2344 1424 : PortalContext = AllocSetContextCreate(AutovacMemCxt,
2345 : "Autovacuum Portal",
2346 : ALLOCSET_DEFAULT_SIZES);
2347 :
2348 : /*
2349 : * Perform operations on collected tables.
2350 : */
2351 101865 : foreach_ptr(TableToProcess, table, tables_to_process)
2352 : {
2353 99019 : Oid relid = table->oid;
2354 : HeapTuple classTup;
2355 : autovac_table *tab;
2356 : bool isshared;
2357 : bool skipit;
2358 : dlist_iter iter;
2359 :
2360 99019 : CHECK_FOR_INTERRUPTS();
2361 :
2362 : /*
2363 : * Check for config changes before processing each collected table.
2364 : */
2365 99019 : if (ConfigReloadPending)
2366 : {
2367 0 : ConfigReloadPending = false;
2368 0 : ProcessConfigFile(PGC_SIGHUP);
2369 :
2370 : /*
2371 : * You might be tempted to bail out if we see autovacuum is now
2372 : * disabled. Must resist that temptation -- this might be a
2373 : * for-wraparound emergency worker, in which case that would be
2374 : * entirely inappropriate.
2375 : */
2376 : }
2377 :
2378 : /*
2379 : * Find out whether the table is shared or not. (It's slightly
2380 : * annoying to fetch the syscache entry just for this, but in typical
2381 : * cases it adds little cost because table_recheck_autovac would
2382 : * refetch the entry anyway. We could buy that back by copying the
2383 : * tuple here and passing it to table_recheck_autovac, but that
2384 : * increases the odds of that function working with stale data.)
2385 : */
2386 99019 : classTup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
2387 99019 : if (!HeapTupleIsValid(classTup))
2388 525 : continue; /* somebody deleted the rel, forget it */
2389 99019 : isshared = ((Form_pg_class) GETSTRUCT(classTup))->relisshared;
2390 99019 : ReleaseSysCache(classTup);
2391 :
2392 : /*
2393 : * Hold schedule lock from here until we've claimed the table. We
2394 : * also need the AutovacuumLock to walk the worker array, but that one
2395 : * can just be a shared lock.
2396 : */
2397 99019 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2398 99019 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2399 :
2400 : /*
2401 : * Check whether the table is being vacuumed concurrently by another
2402 : * worker.
2403 : */
2404 99019 : skipit = false;
2405 316279 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
2406 : {
2407 217782 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
2408 :
2409 : /* ignore myself */
2410 217782 : if (worker == MyWorkerInfo)
2411 98790 : continue;
2412 :
2413 : /* ignore workers in other databases (unless table is shared) */
2414 118992 : if (!worker->wi_sharedrel && worker->wi_dboid != MyDatabaseId)
2415 0 : continue;
2416 :
2417 118992 : if (worker->wi_tableoid == relid)
2418 : {
2419 522 : skipit = true;
2420 522 : found_concurrent_worker = true;
2421 522 : break;
2422 : }
2423 : }
2424 99019 : LWLockRelease(AutovacuumLock);
2425 99019 : if (skipit)
2426 : {
2427 522 : LWLockRelease(AutovacuumScheduleLock);
2428 522 : continue;
2429 : }
2430 :
2431 : /*
2432 : * Store the table's OID in shared memory before releasing the
2433 : * schedule lock, so that other workers don't try to vacuum it
2434 : * concurrently. (We claim it here so as not to hold
2435 : * AutovacuumScheduleLock while rechecking the stats.)
2436 : */
2437 98497 : MyWorkerInfo->wi_tableoid = relid;
2438 98497 : MyWorkerInfo->wi_sharedrel = isshared;
2439 98497 : LWLockRelease(AutovacuumScheduleLock);
2440 :
2441 : /*
2442 : * Check whether pgstat data still says we need to vacuum this table.
2443 : * It could have changed if something else processed the table while
2444 : * we weren't looking. This doesn't entirely close the race condition,
2445 : * but it is very small.
2446 : */
2447 98497 : MemoryContextSwitchTo(AutovacMemCxt);
2448 98497 : tab = table_recheck_autovac(relid, table_toast_map, pg_class_desc,
2449 : effective_multixact_freeze_max_age);
2450 98497 : if (tab == NULL)
2451 : {
2452 : /* someone else vacuumed the table, or it went away */
2453 3 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2454 3 : MyWorkerInfo->wi_tableoid = InvalidOid;
2455 3 : MyWorkerInfo->wi_sharedrel = false;
2456 3 : LWLockRelease(AutovacuumScheduleLock);
2457 3 : continue;
2458 : }
2459 :
2460 : /*
2461 : * Save the cost-related storage parameter values in global variables
2462 : * for reference when updating vacuum_cost_delay and vacuum_cost_limit
2463 : * during vacuuming this table.
2464 : */
2465 98494 : av_storage_param_cost_delay = tab->at_storage_param_vac_cost_delay;
2466 98494 : av_storage_param_cost_limit = tab->at_storage_param_vac_cost_limit;
2467 :
2468 : /*
2469 : * We only expect this worker to ever set the flag, so don't bother
2470 : * checking the return value. We shouldn't have to retry.
2471 : */
2472 98494 : if (tab->at_dobalance)
2473 98494 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2474 : else
2475 0 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
2476 :
2477 98494 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2478 98494 : autovac_recalculate_workers_for_balance();
2479 98494 : LWLockRelease(AutovacuumLock);
2480 :
2481 : /*
2482 : * We wait until this point to update cost delay and cost limit
2483 : * values, even though we reloaded the configuration file above, so
2484 : * that we can take into account the cost-related storage parameters.
2485 : */
2486 98494 : VacuumUpdateCosts();
2487 :
2488 :
2489 : /* clean up memory before each iteration */
2490 98494 : MemoryContextReset(PortalContext);
2491 :
2492 : /*
2493 : * Save the relation name for a possible error message, to avoid a
2494 : * catalog lookup in case of an error. If any of these return NULL,
2495 : * then the relation has been dropped since last we checked; skip it.
2496 : * Note: they must live in a long-lived memory context because we call
2497 : * vacuum and analyze in different transactions.
2498 : */
2499 :
2500 98494 : tab->at_relname = get_rel_name(tab->at_relid);
2501 98494 : tab->at_nspname = get_namespace_name(get_rel_namespace(tab->at_relid));
2502 98494 : tab->at_datname = get_database_name(MyDatabaseId);
2503 98494 : if (!tab->at_relname || !tab->at_nspname || !tab->at_datname)
2504 0 : goto deleted;
2505 :
2506 : /*
2507 : * We will abort vacuuming the current table if something errors out,
2508 : * and continue with the next one in schedule; in particular, this
2509 : * happens if we are interrupted with SIGINT.
2510 : */
2511 98494 : PG_TRY();
2512 : {
2513 : /* Use PortalContext for any per-table allocations */
2514 98494 : MemoryContextSwitchTo(PortalContext);
2515 :
2516 : /* have at it */
2517 98494 : autovacuum_do_vac_analyze(tab, bstrategy);
2518 :
2519 : /*
2520 : * Clear a possible query-cancel signal, to avoid a late reaction
2521 : * to an automatically-sent signal because of vacuuming the
2522 : * current table (we're done with it, so it would make no sense to
2523 : * cancel at this point.)
2524 : */
2525 98493 : QueryCancelPending = false;
2526 : }
2527 0 : PG_CATCH();
2528 : {
2529 : /*
2530 : * Abort the transaction, start a new one, and proceed with the
2531 : * next table in our list.
2532 : */
2533 0 : HOLD_INTERRUPTS();
2534 0 : if (tab->at_params.options & VACOPT_VACUUM)
2535 0 : errcontext("automatic vacuum of table \"%s.%s.%s\"",
2536 : tab->at_datname, tab->at_nspname, tab->at_relname);
2537 : else
2538 0 : errcontext("automatic analyze of table \"%s.%s.%s\"",
2539 : tab->at_datname, tab->at_nspname, tab->at_relname);
2540 0 : EmitErrorReport();
2541 :
2542 : /* this resets ProcGlobal->statusFlags[i] too */
2543 0 : AbortOutOfAnyTransaction();
2544 0 : FlushErrorState();
2545 0 : MemoryContextReset(PortalContext);
2546 :
2547 : /* restart our transaction for the following operations */
2548 0 : StartTransactionCommand();
2549 0 : RESUME_INTERRUPTS();
2550 : }
2551 98493 : PG_END_TRY();
2552 :
2553 : /* Make sure we're back in AutovacMemCxt */
2554 98493 : MemoryContextSwitchTo(AutovacMemCxt);
2555 :
2556 98493 : did_vacuum = true;
2557 :
2558 : /* ProcGlobal->statusFlags[i] are reset at the next end of xact */
2559 :
2560 : /* be tidy */
2561 98493 : deleted:
2562 98493 : if (tab->at_datname != NULL)
2563 98493 : pfree(tab->at_datname);
2564 98493 : if (tab->at_nspname != NULL)
2565 98493 : pfree(tab->at_nspname);
2566 98493 : if (tab->at_relname != NULL)
2567 98493 : pfree(tab->at_relname);
2568 98493 : pfree(tab);
2569 :
2570 : /*
2571 : * Remove my info from shared memory. We set wi_dobalance on the
2572 : * assumption that we are more likely than not to vacuum a table with
2573 : * no cost-related storage parameters next, so we want to claim our
2574 : * share of I/O as soon as possible to avoid thrashing the global
2575 : * balance.
2576 : */
2577 98493 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2578 98493 : MyWorkerInfo->wi_tableoid = InvalidOid;
2579 98493 : MyWorkerInfo->wi_sharedrel = false;
2580 98493 : LWLockRelease(AutovacuumScheduleLock);
2581 98493 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2582 : }
2583 :
2584 1423 : list_free_deep(tables_to_process);
2585 :
2586 : /*
2587 : * Perform additional work items, as requested by backends.
2588 : */
2589 1423 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2590 365711 : for (i = 0; i < NUM_WORKITEMS; i++)
2591 : {
2592 364288 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
2593 :
2594 364288 : if (!workitem->avw_used)
2595 364282 : continue;
2596 6 : if (workitem->avw_active)
2597 0 : continue;
2598 6 : if (workitem->avw_database != MyDatabaseId)
2599 0 : continue;
2600 :
2601 : /* claim this one, and release lock while performing it */
2602 6 : workitem->avw_active = true;
2603 6 : LWLockRelease(AutovacuumLock);
2604 :
2605 6 : PushActiveSnapshot(GetTransactionSnapshot());
2606 6 : perform_work_item(workitem);
2607 6 : if (ActiveSnapshotSet()) /* transaction could have aborted */
2608 6 : PopActiveSnapshot();
2609 :
2610 : /*
2611 : * Check for config changes before acquiring lock for further jobs.
2612 : */
2613 6 : CHECK_FOR_INTERRUPTS();
2614 6 : if (ConfigReloadPending)
2615 : {
2616 0 : ConfigReloadPending = false;
2617 0 : ProcessConfigFile(PGC_SIGHUP);
2618 0 : VacuumUpdateCosts();
2619 : }
2620 :
2621 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2622 :
2623 : /* and mark it done */
2624 6 : workitem->avw_active = false;
2625 6 : workitem->avw_used = false;
2626 : }
2627 1423 : LWLockRelease(AutovacuumLock);
2628 :
2629 : /*
2630 : * We leak table_toast_map here (among other things), but since we're
2631 : * going away soon, it's not a problem normally. But when using Valgrind,
2632 : * release some stuff to reduce complaints about leaked storage.
2633 : */
2634 : #ifdef USE_VALGRIND
2635 : hash_destroy(table_toast_map);
2636 : FreeTupleDesc(pg_class_desc);
2637 : if (bstrategy)
2638 : pfree(bstrategy);
2639 : #endif
2640 :
2641 : /* Run the rest in xact context, mainly to avoid Valgrind leak warnings */
2642 1423 : MemoryContextSwitchTo(TopTransactionContext);
2643 :
2644 : /*
2645 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We
2646 : * only need to do this once, not after each table.
2647 : *
2648 : * Even if we didn't vacuum anything, it may still be important to do
2649 : * this, because one indirect effect of vac_update_datfrozenxid() is to
2650 : * update TransamVariables->xidVacLimit. That might need to be done even
2651 : * if we haven't vacuumed anything, because relations with older
2652 : * relfrozenxid values or other databases with older datfrozenxid values
2653 : * might have been dropped, allowing xidVacLimit to advance.
2654 : *
2655 : * However, it's also important not to do this blindly in all cases,
2656 : * because when autovacuum=off this will restart the autovacuum launcher.
2657 : * If we're not careful, an infinite loop can result, where workers find
2658 : * no work to do and restart the launcher, which starts another worker in
2659 : * the same database that finds no work to do. To prevent that, we skip
2660 : * this if (1) we found no work to do and (2) we skipped at least one
2661 : * table due to concurrent autovacuum activity. In that case, the other
2662 : * worker has already done it, or will do so when it finishes.
2663 : */
2664 1423 : if (did_vacuum || !found_concurrent_worker)
2665 1423 : vac_update_datfrozenxid();
2666 :
2667 : /* Finally close out the last transaction. */
2668 1423 : CommitTransactionCommand();
2669 1423 : }
2670 :
2671 : /*
2672 : * Execute a previously registered work item.
2673 : */
2674 : static void
2675 6 : perform_work_item(AutoVacuumWorkItem *workitem)
2676 : {
2677 6 : char *cur_datname = NULL;
2678 6 : char *cur_nspname = NULL;
2679 6 : char *cur_relname = NULL;
2680 :
2681 : /*
2682 : * Note we do not store table info in MyWorkerInfo, since this is not
2683 : * vacuuming proper.
2684 : */
2685 :
2686 : /*
2687 : * Save the relation name for a possible error message, to avoid a catalog
2688 : * lookup in case of an error. If any of these return NULL, then the
2689 : * relation has been dropped since last we checked; skip it.
2690 : */
2691 : Assert(CurrentMemoryContext == AutovacMemCxt);
2692 :
2693 6 : cur_relname = get_rel_name(workitem->avw_relation);
2694 6 : cur_nspname = get_namespace_name(get_rel_namespace(workitem->avw_relation));
2695 6 : cur_datname = get_database_name(MyDatabaseId);
2696 6 : if (!cur_relname || !cur_nspname || !cur_datname)
2697 0 : goto deleted2;
2698 :
2699 6 : autovac_report_workitem(workitem, cur_nspname, cur_relname);
2700 :
2701 : /* clean up memory before each work item */
2702 6 : MemoryContextReset(PortalContext);
2703 :
2704 : /*
2705 : * We will abort the current work item if something errors out, and
2706 : * continue with the next one; in particular, this happens if we are
2707 : * interrupted with SIGINT. Note that this means that the work item list
2708 : * can be lossy.
2709 : */
2710 6 : PG_TRY();
2711 : {
2712 : /* Use PortalContext for any per-work-item allocations */
2713 6 : MemoryContextSwitchTo(PortalContext);
2714 :
2715 : /*
2716 : * Have at it. Functions called here are responsible for any required
2717 : * user switch and sandbox.
2718 : */
2719 6 : switch (workitem->avw_type)
2720 : {
2721 6 : case AVW_BRINSummarizeRange:
2722 6 : DirectFunctionCall2(brin_summarize_range,
2723 : ObjectIdGetDatum(workitem->avw_relation),
2724 : Int64GetDatum((int64) workitem->avw_blockNumber));
2725 6 : break;
2726 0 : default:
2727 0 : elog(WARNING, "unrecognized work item found: type %d",
2728 : workitem->avw_type);
2729 0 : break;
2730 : }
2731 :
2732 : /*
2733 : * Clear a possible query-cancel signal, to avoid a late reaction to
2734 : * an automatically-sent signal because of vacuuming the current table
2735 : * (we're done with it, so it would make no sense to cancel at this
2736 : * point.)
2737 : */
2738 6 : QueryCancelPending = false;
2739 : }
2740 0 : PG_CATCH();
2741 : {
2742 : /*
2743 : * Abort the transaction, start a new one, and proceed with the next
2744 : * table in our list.
2745 : */
2746 0 : HOLD_INTERRUPTS();
2747 0 : errcontext("processing work entry for relation \"%s.%s.%s\"",
2748 : cur_datname, cur_nspname, cur_relname);
2749 0 : EmitErrorReport();
2750 :
2751 : /* this resets ProcGlobal->statusFlags[i] too */
2752 0 : AbortOutOfAnyTransaction();
2753 0 : FlushErrorState();
2754 0 : MemoryContextReset(PortalContext);
2755 :
2756 : /* restart our transaction for the following operations */
2757 0 : StartTransactionCommand();
2758 0 : RESUME_INTERRUPTS();
2759 : }
2760 6 : PG_END_TRY();
2761 :
2762 : /* Make sure we're back in AutovacMemCxt */
2763 6 : MemoryContextSwitchTo(AutovacMemCxt);
2764 :
2765 : /* We intentionally do not set did_vacuum here */
2766 :
2767 : /* be tidy */
2768 6 : deleted2:
2769 6 : if (cur_datname)
2770 6 : pfree(cur_datname);
2771 6 : if (cur_nspname)
2772 6 : pfree(cur_nspname);
2773 6 : if (cur_relname)
2774 6 : pfree(cur_relname);
2775 6 : }
2776 :
2777 : /*
2778 : * extract_autovac_opts
2779 : *
2780 : * Given a relation's pg_class tuple, return a palloc'd copy of the
2781 : * AutoVacOpts portion of reloptions, if set; otherwise, return NULL.
2782 : *
2783 : * Note: callers do not have a relation lock on the table at this point,
2784 : * so the table could have been dropped, and its catalog rows gone, after
2785 : * we acquired the pg_class row. If pg_class had a TOAST table, this would
2786 : * be a risk; fortunately, it doesn't.
2787 : */
2788 : static AutoVacOpts *
2789 266567 : extract_autovac_opts(HeapTuple tup, TupleDesc pg_class_desc)
2790 : {
2791 : bytea *relopts;
2792 : AutoVacOpts *av;
2793 :
2794 : Assert(((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_RELATION ||
2795 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_MATVIEW ||
2796 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_TOASTVALUE);
2797 :
2798 266567 : relopts = extractRelOptions(tup, pg_class_desc, NULL);
2799 266567 : if (relopts == NULL)
2800 264307 : return NULL;
2801 :
2802 2260 : av = palloc_object(AutoVacOpts);
2803 2260 : memcpy(av, &(((StdRdOptions *) relopts)->autovacuum), sizeof(AutoVacOpts));
2804 2260 : pfree(relopts);
2805 :
2806 2260 : return av;
2807 : }
2808 :
2809 :
2810 : /*
2811 : * table_recheck_autovac
2812 : *
2813 : * Recheck whether a table still needs vacuum or analyze. Return value is a
2814 : * valid autovac_table pointer if it does, NULL otherwise.
2815 : *
2816 : * Note that the returned autovac_table does not have the name fields set.
2817 : */
2818 : static autovac_table *
2819 98497 : table_recheck_autovac(Oid relid, HTAB *table_toast_map,
2820 : TupleDesc pg_class_desc,
2821 : int effective_multixact_freeze_max_age)
2822 : {
2823 : Form_pg_class classForm;
2824 : HeapTuple classTup;
2825 : bool dovacuum;
2826 : bool doanalyze;
2827 98497 : autovac_table *tab = NULL;
2828 : bool wraparound;
2829 : AutoVacOpts *avopts;
2830 98497 : bool free_avopts = false;
2831 : AutoVacuumScores scores;
2832 :
2833 : /* fetch the relation's relcache entry */
2834 98497 : classTup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2835 98497 : if (!HeapTupleIsValid(classTup))
2836 0 : return NULL;
2837 98497 : classForm = (Form_pg_class) GETSTRUCT(classTup);
2838 :
2839 : /*
2840 : * Get the applicable reloptions. If it is a TOAST table, try to get the
2841 : * main table reloptions if the toast table itself doesn't have.
2842 : */
2843 98497 : avopts = extract_autovac_opts(classTup, pg_class_desc);
2844 98497 : if (avopts)
2845 969 : free_avopts = true;
2846 97528 : else if (classForm->relkind == RELKIND_TOASTVALUE &&
2847 : table_toast_map != NULL)
2848 : {
2849 : av_relation *hentry;
2850 : bool found;
2851 :
2852 35418 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2853 35418 : if (found && hentry->ar_hasrelopts)
2854 951 : avopts = &hentry->ar_reloptions;
2855 : }
2856 :
2857 98497 : relation_needs_vacanalyze(relid, avopts, classForm,
2858 : effective_multixact_freeze_max_age,
2859 : DEBUG3,
2860 : &dovacuum, &doanalyze, &wraparound,
2861 : &scores);
2862 :
2863 : /* OK, it needs something done */
2864 98497 : 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 1920 : log_vacuum_min_duration = (avopts && avopts->log_vacuum_min_duration >= 0)
2882 : ? avopts->log_vacuum_min_duration
2883 100414 : : Log_autovacuum_min_duration;
2884 :
2885 : /* -1 in autovac setting means use log_autoanalyze_min_duration */
2886 1920 : log_analyze_min_duration = (avopts && avopts->log_analyze_min_duration >= 0)
2887 : ? avopts->log_analyze_min_duration
2888 100414 : : Log_autoanalyze_min_duration;
2889 :
2890 : /* these do not have autovacuum-specific settings */
2891 1920 : freeze_min_age = (avopts && avopts->freeze_min_age >= 0)
2892 : ? avopts->freeze_min_age
2893 100414 : : default_freeze_min_age;
2894 :
2895 1920 : freeze_table_age = (avopts && avopts->freeze_table_age >= 0)
2896 : ? avopts->freeze_table_age
2897 100414 : : default_freeze_table_age;
2898 :
2899 100414 : multixact_freeze_min_age = (avopts &&
2900 1920 : avopts->multixact_freeze_min_age >= 0)
2901 : ? avopts->multixact_freeze_min_age
2902 100414 : : default_multixact_freeze_min_age;
2903 :
2904 100414 : multixact_freeze_table_age = (avopts &&
2905 1920 : avopts->multixact_freeze_table_age >= 0)
2906 : ? avopts->multixact_freeze_table_age
2907 100414 : : default_multixact_freeze_table_age;
2908 :
2909 98494 : tab = palloc_object(autovac_table);
2910 98494 : 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 98494 : tab->at_params.options =
2918 98494 : (dovacuum ? (VACOPT_VACUUM |
2919 : VACOPT_PROCESS_MAIN |
2920 98494 : VACOPT_SKIP_DATABASE_STATS) : 0) |
2921 98494 : (doanalyze ? VACOPT_ANALYZE : 0) |
2922 98494 : (!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 98494 : tab->at_params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
2930 98494 : tab->at_params.truncate = VACOPTVALUE_UNSPECIFIED;
2931 98494 : tab->at_params.freeze_min_age = freeze_min_age;
2932 98494 : tab->at_params.freeze_table_age = freeze_table_age;
2933 98494 : tab->at_params.multixact_freeze_min_age = multixact_freeze_min_age;
2934 98494 : tab->at_params.multixact_freeze_table_age = multixact_freeze_table_age;
2935 98494 : tab->at_params.is_wraparound = wraparound;
2936 98494 : tab->at_params.log_vacuum_min_duration = log_vacuum_min_duration;
2937 98494 : tab->at_params.log_analyze_min_duration = log_analyze_min_duration;
2938 98494 : tab->at_params.toast_parent = InvalidOid;
2939 :
2940 : /* Determine the number of parallel vacuum workers to use */
2941 98494 : tab->at_params.nworkers = 0;
2942 98494 : if (avopts)
2943 : {
2944 1920 : if (avopts->autovacuum_parallel_workers == 0)
2945 : {
2946 : /*
2947 : * Disable parallel vacuum, if the reloption sets the parallel
2948 : * degree as zero.
2949 : */
2950 0 : tab->at_params.nworkers = -1;
2951 : }
2952 1920 : else if (avopts->autovacuum_parallel_workers > 0)
2953 1 : tab->at_params.nworkers = avopts->autovacuum_parallel_workers;
2954 :
2955 : /*
2956 : * autovacuum_parallel_workers == -1 falls through, keep
2957 : * nworkers=0
2958 : */
2959 : }
2960 :
2961 : /*
2962 : * Later, in vacuum_rel(), we check reloptions for any
2963 : * vacuum_max_eager_freeze_failure_rate override.
2964 : */
2965 98494 : tab->at_params.max_eager_freeze_failure_rate = vacuum_max_eager_freeze_failure_rate;
2966 98494 : tab->at_storage_param_vac_cost_limit = avopts ?
2967 98494 : avopts->vacuum_cost_limit : 0;
2968 98494 : tab->at_storage_param_vac_cost_delay = avopts ?
2969 98494 : avopts->vacuum_cost_delay : -1;
2970 98494 : tab->at_relname = NULL;
2971 98494 : tab->at_nspname = NULL;
2972 98494 : tab->at_datname = NULL;
2973 :
2974 : /*
2975 : * If any of the cost delay parameters has been set individually for
2976 : * this table, disable the balancing algorithm.
2977 : */
2978 98494 : tab->at_dobalance =
2979 100414 : !(avopts && (avopts->vacuum_cost_limit > 0 ||
2980 100414 : avopts->vacuum_cost_delay >= 0));
2981 : }
2982 :
2983 98497 : if (free_avopts)
2984 969 : pfree(avopts);
2985 98497 : heap_freetuple(classTup);
2986 98497 : return tab;
2987 : }
2988 :
2989 : /*
2990 : * relation_needs_vacanalyze
2991 : *
2992 : * Check whether a relation needs to be vacuumed or analyzed; return each into
2993 : * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is
2994 : * being forced because of Xid or multixact wraparound.
2995 : *
2996 : * relopts is a pointer to the AutoVacOpts options (either for itself in the
2997 : * case of a plain table, or for either itself or its parent table in the case
2998 : * of a TOAST table), NULL if none.
2999 : *
3000 : * A table needs to be vacuumed if the number of dead tuples exceeds a
3001 : * threshold. This threshold is calculated as
3002 : *
3003 : * threshold = vac_base_thresh + vac_scale_factor * reltuples
3004 : * if (threshold > vac_max_thresh)
3005 : * threshold = vac_max_thresh;
3006 : *
3007 : * For analyze, the analysis done is that the number of tuples inserted,
3008 : * deleted and updated since the last analyze exceeds a threshold calculated
3009 : * in the same fashion as above. Note that the cumulative stats system stores
3010 : * the number of tuples (both live and dead) that there were as of the last
3011 : * analyze. This is asymmetric to the VACUUM case.
3012 : *
3013 : * We also force vacuum if the table's relfrozenxid is more than freeze_max_age
3014 : * transactions back, and if its relminmxid is more than
3015 : * multixact_freeze_max_age multixacts back.
3016 : *
3017 : * A table whose autovacuum_enabled option is false is
3018 : * automatically skipped (unless we have to vacuum it due to freeze_max_age).
3019 : * Thus autovacuum can be disabled for specific tables. Also, when the cumulative
3020 : * stats system does not have data about a table, it will be skipped.
3021 : *
3022 : * A table whose vac_base_thresh value is < 0 takes the base value from the
3023 : * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor
3024 : * value < 0 is substituted with the value of
3025 : * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze.
3026 : *
3027 : * This function also returns scores that can be used to sort the list of
3028 : * tables to process. The idea is to have autovacuum prioritize tables that
3029 : * are furthest beyond their thresholds (e.g., a table nearing transaction ID
3030 : * wraparound should be vacuumed first). This prioritization scheme is
3031 : * certainly far from perfect; there are simply too many possibilities for any
3032 : * scoring technique to work across all workloads, and the situation might
3033 : * change significantly between the time we calculate the score and the time
3034 : * that autovacuum processes it. However, we have attempted to develop
3035 : * something that is expected to work for a large portion of workloads with
3036 : * reasonable parameter settings.
3037 : *
3038 : * The autovacuum table score is calculated as the maximum of the ratios of
3039 : * each of the table's relevant values to its threshold. For example, if the
3040 : * number of inserted tuples is 100, and the insert threshold for the table is
3041 : * 80, the insert score is 1.25. If all other scores are below that value, the
3042 : * returned score will be 1.25. The other criteria considered for the score
3043 : * are the table ages (both relfrozenxid and relminmxid) compared to the
3044 : * corresponding freeze-max-age setting, the number of updated/deleted tuples
3045 : * compared to the vacuum threshold, and the number of inserted/updated/deleted
3046 : * tuples compared to the analyze threshold.
3047 : *
3048 : * One exception to the previous paragraph is for tables nearing wraparound,
3049 : * i.e., those that have surpassed the effective failsafe ages. In that case,
3050 : * the relfrozen/relminmxid-based score is scaled aggressively so that the
3051 : * table has a decent chance of sorting to the front of the list.
3052 : *
3053 : * To adjust how strongly each component contributes to the score, the
3054 : * following parameters can be adjusted from their default of 1.0 to anywhere
3055 : * between 0.0 and 10.0 (inclusive). Setting all of these to 0.0 restores
3056 : * pre-v19 prioritization behavior:
3057 : *
3058 : * autovacuum_freeze_score_weight
3059 : * autovacuum_multixact_freeze_score_weight
3060 : * autovacuum_vacuum_score_weight
3061 : * autovacuum_vacuum_insert_score_weight
3062 : * autovacuum_analyze_score_weight
3063 : *
3064 : * The autovacuum table score is returned in scores->max. The component scores
3065 : * are also returned in the "scores" argument via the other members of the
3066 : * AutoVacuumScores struct.
3067 : */
3068 : static void
3069 266567 : relation_needs_vacanalyze(Oid relid,
3070 : AutoVacOpts *relopts,
3071 : Form_pg_class classForm,
3072 : int effective_multixact_freeze_max_age,
3073 : int elevel,
3074 : /* output params below */
3075 : bool *dovacuum,
3076 : bool *doanalyze,
3077 : bool *wraparound,
3078 : AutoVacuumScores *scores)
3079 : {
3080 : PgStat_StatTabEntry *tabentry;
3081 : bool force_vacuum;
3082 : bool av_enabled;
3083 266567 : bool may_free = false;
3084 :
3085 : /* constants from reloptions or GUC variables */
3086 : int vac_base_thresh,
3087 : vac_max_thresh,
3088 : vac_ins_base_thresh,
3089 : anl_base_thresh;
3090 : float4 vac_scale_factor,
3091 : vac_ins_scale_factor,
3092 : anl_scale_factor;
3093 :
3094 : /* thresholds calculated from above constants */
3095 : float4 vacthresh,
3096 : vacinsthresh,
3097 : anlthresh;
3098 :
3099 : /* number of vacuum (resp. analyze) tuples at this time */
3100 : float4 vactuples,
3101 : instuples,
3102 : anltuples;
3103 :
3104 : /* freeze parameters */
3105 : int freeze_max_age;
3106 : int multixact_freeze_max_age;
3107 : TransactionId xidForceLimit;
3108 : TransactionId relfrozenxid;
3109 : MultiXactId relminmxid;
3110 : MultiXactId multiForceLimit;
3111 : uint32 xid_age;
3112 : uint32 mxid_age;
3113 : int effective_xid_failsafe_age;
3114 : int effective_mxid_failsafe_age;
3115 :
3116 266567 : float4 pcnt_unfrozen = 1;
3117 266567 : float4 reltuples = classForm->reltuples;
3118 266567 : int32 relpages = classForm->relpages;
3119 266567 : int32 relallfrozen = classForm->relallfrozen;
3120 :
3121 : Assert(classForm != NULL);
3122 : Assert(OidIsValid(relid));
3123 :
3124 266567 : memset(scores, 0, sizeof(AutoVacuumScores));
3125 266567 : *dovacuum = false;
3126 266567 : *doanalyze = false;
3127 :
3128 : /*
3129 : * Determine vacuum/analyze equation parameters. We have two possible
3130 : * sources: the passed reloptions (which could be a main table or a toast
3131 : * table), or the autovacuum GUC variables.
3132 : */
3133 :
3134 : /* -1 in autovac setting means use plain vacuum_scale_factor */
3135 4453 : vac_scale_factor = (relopts && relopts->vacuum_scale_factor >= 0)
3136 0 : ? relopts->vacuum_scale_factor
3137 271020 : : autovacuum_vac_scale;
3138 :
3139 4453 : vac_base_thresh = (relopts && relopts->vacuum_threshold >= 0)
3140 : ? relopts->vacuum_threshold
3141 271020 : : autovacuum_vac_thresh;
3142 :
3143 : /* -1 is used to disable max threshold */
3144 4453 : vac_max_thresh = (relopts && relopts->vacuum_max_threshold >= -1)
3145 : ? relopts->vacuum_max_threshold
3146 271020 : : autovacuum_vac_max_thresh;
3147 :
3148 4453 : vac_ins_scale_factor = (relopts && relopts->vacuum_ins_scale_factor >= 0)
3149 0 : ? relopts->vacuum_ins_scale_factor
3150 271020 : : autovacuum_vac_ins_scale;
3151 :
3152 : /* -1 is used to disable insert vacuums */
3153 4453 : vac_ins_base_thresh = (relopts && relopts->vacuum_ins_threshold >= -1)
3154 : ? relopts->vacuum_ins_threshold
3155 271020 : : autovacuum_vac_ins_thresh;
3156 :
3157 4453 : anl_scale_factor = (relopts && relopts->analyze_scale_factor >= 0)
3158 0 : ? relopts->analyze_scale_factor
3159 271020 : : autovacuum_anl_scale;
3160 :
3161 4453 : anl_base_thresh = (relopts && relopts->analyze_threshold >= 0)
3162 : ? relopts->analyze_threshold
3163 271020 : : autovacuum_anl_thresh;
3164 :
3165 4453 : freeze_max_age = (relopts && relopts->freeze_max_age >= 0)
3166 0 : ? Min(relopts->freeze_max_age, autovacuum_freeze_max_age)
3167 271020 : : autovacuum_freeze_max_age;
3168 :
3169 4453 : multixact_freeze_max_age = (relopts && relopts->multixact_freeze_max_age >= 0)
3170 0 : ? Min(relopts->multixact_freeze_max_age, effective_multixact_freeze_max_age)
3171 271020 : : effective_multixact_freeze_max_age;
3172 :
3173 266567 : av_enabled = (relopts ? relopts->enabled : true);
3174 266567 : av_enabled &= AutoVacuumingActive();
3175 :
3176 266567 : relfrozenxid = classForm->relfrozenxid;
3177 266567 : relminmxid = classForm->relminmxid;
3178 :
3179 : /* Force vacuum if table is at risk of wraparound */
3180 266567 : xidForceLimit = recentXid - freeze_max_age;
3181 266567 : if (xidForceLimit < FirstNormalTransactionId)
3182 0 : xidForceLimit -= FirstNormalTransactionId;
3183 533134 : force_vacuum = (TransactionIdIsNormal(relfrozenxid) &&
3184 266567 : TransactionIdPrecedes(relfrozenxid, xidForceLimit));
3185 266567 : if (!force_vacuum)
3186 : {
3187 69862 : multiForceLimit = recentMulti - multixact_freeze_max_age;
3188 69862 : if (multiForceLimit < FirstMultiXactId)
3189 0 : multiForceLimit -= FirstMultiXactId;
3190 139724 : force_vacuum = MultiXactIdIsValid(relminmxid) &&
3191 69862 : MultiXactIdPrecedes(relminmxid, multiForceLimit);
3192 : }
3193 266567 : *wraparound = force_vacuum;
3194 :
3195 : /*
3196 : * To calculate the (M)XID age portion of the score, divide the age by its
3197 : * respective *_freeze_max_age parameter.
3198 : */
3199 266567 : xid_age = TransactionIdIsNormal(relfrozenxid) ? recentXid - relfrozenxid : 0;
3200 266567 : mxid_age = MultiXactIdIsValid(relminmxid) ? recentMulti - relminmxid : 0;
3201 :
3202 266567 : scores->xid = (double) xid_age / freeze_max_age;
3203 266567 : scores->mxid = (double) mxid_age / multixact_freeze_max_age;
3204 :
3205 : /*
3206 : * To ensure tables are given increased priority once they begin
3207 : * approaching wraparound, we scale the score aggressively if the ages
3208 : * surpass vacuum_failsafe_age or vacuum_multixact_failsafe_age.
3209 : *
3210 : * As in vacuum_xid_failsafe_check(), the effective failsafe age is no
3211 : * less than 105% the value of the respective *_freeze_max_age parameter.
3212 : * Note that per-table settings could result in a low score even if the
3213 : * table surpasses the failsafe settings. However, this is a strange
3214 : * enough corner case that we don't bother trying to handle it.
3215 : *
3216 : * We further adjust the effective failsafe ages with the weight
3217 : * parameters so that increasing them lowers the ages at which we begin
3218 : * scaling aggressively.
3219 : */
3220 266567 : effective_xid_failsafe_age = Max(vacuum_failsafe_age,
3221 : autovacuum_freeze_max_age * 1.05);
3222 266567 : effective_mxid_failsafe_age = Max(vacuum_multixact_failsafe_age,
3223 : autovacuum_multixact_freeze_max_age * 1.05);
3224 :
3225 266567 : if (autovacuum_freeze_score_weight > 1.0)
3226 0 : effective_xid_failsafe_age /= autovacuum_freeze_score_weight;
3227 266567 : if (autovacuum_multixact_freeze_score_weight > 1.0)
3228 0 : effective_mxid_failsafe_age /= autovacuum_multixact_freeze_score_weight;
3229 :
3230 266567 : if (xid_age >= effective_xid_failsafe_age)
3231 47420 : scores->xid = pow(scores->xid, Max(1.0, (double) xid_age / 100000000));
3232 266567 : if (mxid_age >= effective_mxid_failsafe_age)
3233 0 : scores->mxid = pow(scores->mxid, Max(1.0, (double) mxid_age / 100000000));
3234 :
3235 266567 : scores->xid *= autovacuum_freeze_score_weight;
3236 266567 : scores->mxid *= autovacuum_multixact_freeze_score_weight;
3237 :
3238 266567 : scores->max = Max(scores->xid, scores->mxid);
3239 266567 : if (force_vacuum)
3240 196705 : *dovacuum = true;
3241 :
3242 : /*
3243 : * If we found stats for the table, and autovacuum is currently enabled,
3244 : * make a threshold-based decision whether to vacuum and/or analyze. If
3245 : * autovacuum is currently disabled, we must be here for anti-wraparound
3246 : * vacuuming only, so don't vacuum (or analyze) anything that's not being
3247 : * forced.
3248 : */
3249 266567 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
3250 : relid, &may_free);
3251 266567 : if (!tabentry)
3252 4804 : return;
3253 :
3254 261763 : vactuples = tabentry->dead_tuples;
3255 261763 : instuples = tabentry->ins_since_vacuum;
3256 261763 : anltuples = tabentry->mod_since_analyze;
3257 :
3258 : /* If the table hasn't yet been vacuumed, take reltuples as zero */
3259 261763 : if (reltuples < 0)
3260 1950 : reltuples = 0;
3261 :
3262 : /*
3263 : * If we have data for relallfrozen, calculate the unfrozen percentage of
3264 : * the table to modify insert scale factor. This helps us decide whether
3265 : * or not to vacuum an insert-heavy table based on the number of inserts
3266 : * to the more "active" part of the table.
3267 : */
3268 261763 : if (relpages > 0 && relallfrozen > 0)
3269 : {
3270 : /*
3271 : * It could be the stats were updated manually and relallfrozen >
3272 : * relpages. Clamp relallfrozen to relpages to avoid nonsensical
3273 : * calculations.
3274 : */
3275 95281 : relallfrozen = Min(relallfrozen, relpages);
3276 95281 : pcnt_unfrozen = 1 - ((float4) relallfrozen / relpages);
3277 : }
3278 :
3279 261763 : vacthresh = (float4) vac_base_thresh + vac_scale_factor * reltuples;
3280 261763 : if (vac_max_thresh >= 0 && vacthresh > (float4) vac_max_thresh)
3281 0 : vacthresh = (float4) vac_max_thresh;
3282 :
3283 261763 : vacinsthresh = (float4) vac_ins_base_thresh +
3284 261763 : vac_ins_scale_factor * reltuples * pcnt_unfrozen;
3285 261763 : anlthresh = (float4) anl_base_thresh + anl_scale_factor * reltuples;
3286 :
3287 : /* Determine if this table needs vacuum, and update the score. */
3288 261763 : scores->vac = (double) vactuples / Max(vacthresh, 1);
3289 261763 : scores->vac *= autovacuum_vacuum_score_weight;
3290 261763 : scores->max = Max(scores->max, scores->vac);
3291 261763 : if (av_enabled && vactuples > vacthresh)
3292 219 : *dovacuum = true;
3293 :
3294 261763 : if (vac_ins_base_thresh >= 0)
3295 : {
3296 261763 : scores->vac_ins = (double) instuples / Max(vacinsthresh, 1);
3297 261763 : scores->vac_ins *= autovacuum_vacuum_insert_score_weight;
3298 261763 : scores->max = Max(scores->max, scores->vac_ins);
3299 261763 : if (av_enabled && instuples > vacinsthresh)
3300 172 : *dovacuum = true;
3301 : }
3302 :
3303 : /*
3304 : * Determine if this table needs analyze, and update the score. Note that
3305 : * we don't analyze TOAST tables and pg_statistic.
3306 : */
3307 261763 : if (relid != StatisticRelationId &&
3308 259526 : classForm->relkind != RELKIND_TOASTVALUE)
3309 : {
3310 165590 : scores->anl = (double) anltuples / Max(anlthresh, 1);
3311 165590 : scores->anl *= autovacuum_analyze_score_weight;
3312 165590 : scores->max = Max(scores->max, scores->anl);
3313 165590 : if (av_enabled && anltuples > anlthresh)
3314 769 : *doanalyze = true;
3315 : }
3316 :
3317 261763 : if (vac_ins_base_thresh >= 0)
3318 261763 : elog(elevel, "%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",
3319 : NameStr(classForm->relname),
3320 : vactuples, vacthresh, scores->vac,
3321 : instuples, vacinsthresh, scores->vac_ins,
3322 : anltuples, anlthresh, scores->anl,
3323 : scores->xid, scores->mxid);
3324 : else
3325 0 : elog(elevel, "%s: vac: %.0f (thresh %.0f, score %.2f), ins: (disabled), anl: %.0f (thresh %.0f, score %.2f), xid score: %.2f, mxid score: %.2f",
3326 : NameStr(classForm->relname),
3327 : vactuples, vacthresh, scores->vac,
3328 : anltuples, anlthresh, scores->anl,
3329 : scores->xid, scores->mxid);
3330 :
3331 : /* Avoid leaking pgstat entries until the end of autovacuum. */
3332 261763 : if (may_free)
3333 261763 : pfree(tabentry);
3334 : }
3335 :
3336 : /*
3337 : * autovacuum_do_vac_analyze
3338 : * Vacuum and/or analyze the specified table
3339 : *
3340 : * We expect the caller to have switched into a memory context that won't
3341 : * disappear at transaction commit.
3342 : */
3343 : static void
3344 98494 : autovacuum_do_vac_analyze(autovac_table *tab, BufferAccessStrategy bstrategy)
3345 : {
3346 : RangeVar *rangevar;
3347 : VacuumRelation *rel;
3348 : List *rel_list;
3349 : MemoryContext vac_context;
3350 : MemoryContext old_context;
3351 :
3352 : /* Let pgstat know what we're doing */
3353 98494 : autovac_report_activity(tab);
3354 :
3355 : /* Create a context that vacuum() can use as cross-transaction storage */
3356 98494 : vac_context = AllocSetContextCreate(CurrentMemoryContext,
3357 : "Vacuum",
3358 : ALLOCSET_DEFAULT_SIZES);
3359 :
3360 : /* Set up one VacuumRelation target, identified by OID, for vacuum() */
3361 98494 : old_context = MemoryContextSwitchTo(vac_context);
3362 98494 : rangevar = makeRangeVar(tab->at_nspname, tab->at_relname, -1);
3363 98494 : rel = makeVacuumRelation(rangevar, tab->at_relid, NIL);
3364 98494 : rel_list = list_make1(rel);
3365 98494 : MemoryContextSwitchTo(old_context);
3366 :
3367 98494 : vacuum(rel_list, &tab->at_params, bstrategy, vac_context, true);
3368 :
3369 98493 : MemoryContextDelete(vac_context);
3370 98493 : }
3371 :
3372 : /*
3373 : * autovac_report_activity
3374 : * Report to pgstat what autovacuum is doing
3375 : *
3376 : * We send a SQL string corresponding to what the user would see if the
3377 : * equivalent command was to be issued manually.
3378 : *
3379 : * Note we assume that we are going to report the next command as soon as we're
3380 : * done with the current one, and exit right after the last one, so we don't
3381 : * bother to report "<IDLE>" or some such.
3382 : */
3383 : static void
3384 98494 : autovac_report_activity(autovac_table *tab)
3385 : {
3386 : #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56)
3387 : char activity[MAX_AUTOVAC_ACTIV_LEN];
3388 : int len;
3389 :
3390 : /* Report the command and possible options */
3391 98494 : if (tab->at_params.options & VACOPT_VACUUM)
3392 98260 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3393 : "autovacuum: VACUUM%s",
3394 98260 : tab->at_params.options & VACOPT_ANALYZE ? " ANALYZE" : "");
3395 : else
3396 234 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3397 : "autovacuum: ANALYZE");
3398 :
3399 : /*
3400 : * Report the qualified name of the relation.
3401 : */
3402 98494 : len = strlen(activity);
3403 :
3404 98494 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3405 : " %s.%s%s", tab->at_nspname, tab->at_relname,
3406 98494 : tab->at_params.is_wraparound ? " (to prevent wraparound)" : "");
3407 :
3408 : /* Set statement_timestamp() to current time for pg_stat_activity */
3409 98494 : SetCurrentStatementStartTimestamp();
3410 :
3411 98494 : pgstat_report_activity(STATE_RUNNING, activity);
3412 98494 : }
3413 :
3414 : /*
3415 : * autovac_report_workitem
3416 : * Report to pgstat that autovacuum is processing a work item
3417 : */
3418 : static void
3419 6 : autovac_report_workitem(AutoVacuumWorkItem *workitem,
3420 : const char *nspname, const char *relname)
3421 : {
3422 : char activity[MAX_AUTOVAC_ACTIV_LEN + 12 + 2];
3423 : char blk[12 + 2];
3424 : int len;
3425 :
3426 6 : switch (workitem->avw_type)
3427 : {
3428 6 : case AVW_BRINSummarizeRange:
3429 6 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3430 : "autovacuum: BRIN summarize");
3431 6 : break;
3432 : }
3433 :
3434 : /*
3435 : * Report the qualified name of the relation, and the block number if any
3436 : */
3437 6 : len = strlen(activity);
3438 :
3439 6 : if (BlockNumberIsValid(workitem->avw_blockNumber))
3440 6 : snprintf(blk, sizeof(blk), " %u", workitem->avw_blockNumber);
3441 : else
3442 0 : blk[0] = '\0';
3443 :
3444 6 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3445 : " %s.%s%s", nspname, relname, blk);
3446 :
3447 : /* Set statement_timestamp() to current time for pg_stat_activity */
3448 6 : SetCurrentStatementStartTimestamp();
3449 :
3450 6 : pgstat_report_activity(STATE_RUNNING, activity);
3451 6 : }
3452 :
3453 : /*
3454 : * AutoVacuumingActive
3455 : * Check GUC vars and report whether the autovacuum process should be
3456 : * running.
3457 : */
3458 : bool
3459 312815 : AutoVacuumingActive(void)
3460 : {
3461 312815 : if (!autovacuum_start_daemon || !pgstat_track_counts)
3462 5209 : return false;
3463 307606 : return true;
3464 : }
3465 :
3466 : /*
3467 : * Request one work item to the next autovacuum run processing our database.
3468 : * Return false if the request can't be recorded.
3469 : */
3470 : bool
3471 6 : AutoVacuumRequestWork(AutoVacuumWorkItemType type, Oid relationId,
3472 : BlockNumber blkno)
3473 : {
3474 : int i;
3475 6 : bool result = false;
3476 :
3477 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
3478 :
3479 : /*
3480 : * Locate an unused work item and fill it with the given data.
3481 : */
3482 21 : for (i = 0; i < NUM_WORKITEMS; i++)
3483 : {
3484 21 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
3485 :
3486 21 : if (workitem->avw_used)
3487 15 : continue;
3488 :
3489 6 : workitem->avw_used = true;
3490 6 : workitem->avw_active = false;
3491 6 : workitem->avw_type = type;
3492 6 : workitem->avw_database = MyDatabaseId;
3493 6 : workitem->avw_relation = relationId;
3494 6 : workitem->avw_blockNumber = blkno;
3495 6 : result = true;
3496 :
3497 : /* done */
3498 6 : break;
3499 : }
3500 :
3501 6 : LWLockRelease(AutovacuumLock);
3502 :
3503 6 : return result;
3504 : }
3505 :
3506 : /*
3507 : * autovac_init
3508 : * This is called at postmaster initialization.
3509 : *
3510 : * All we do here is annoy the user if he got it wrong.
3511 : */
3512 : void
3513 983 : autovac_init(void)
3514 : {
3515 983 : if (!autovacuum_start_daemon)
3516 124 : return;
3517 859 : else if (!pgstat_track_counts)
3518 0 : ereport(WARNING,
3519 : (errmsg("autovacuum not started because of misconfiguration"),
3520 : errhint("Enable the \"track_counts\" option.")));
3521 : else
3522 859 : check_av_worker_gucs();
3523 : }
3524 :
3525 : /*
3526 : * AutoVacuumShmemRequest
3527 : * Register shared memory space needed for autovacuum
3528 : */
3529 : static void
3530 1238 : AutoVacuumShmemRequest(void *arg)
3531 : {
3532 : Size size;
3533 :
3534 : /*
3535 : * Need the fixed struct and the array of WorkerInfoData.
3536 : */
3537 1238 : size = sizeof(AutoVacuumShmemStruct);
3538 1238 : size = MAXALIGN(size);
3539 1238 : size = add_size(size, mul_size(autovacuum_worker_slots,
3540 : sizeof(WorkerInfoData)));
3541 :
3542 1238 : ShmemRequestStruct(.name = "AutoVacuum Data",
3543 : .size = size,
3544 : .ptr = (void **) &AutoVacuumShmem,
3545 : );
3546 1238 : }
3547 :
3548 : /*
3549 : * AutoVacuumShmemInit
3550 : * Initialize autovacuum-related shared memory
3551 : */
3552 : static void
3553 1235 : AutoVacuumShmemInit(void *arg)
3554 : {
3555 : WorkerInfo worker;
3556 :
3557 1235 : AutoVacuumShmem->av_launcherpid = 0;
3558 1235 : dclist_init(&AutoVacuumShmem->av_freeWorkers);
3559 1235 : dlist_init(&AutoVacuumShmem->av_runningWorkers);
3560 1235 : AutoVacuumShmem->av_startingWorker = NULL;
3561 1235 : memset(AutoVacuumShmem->av_workItems, 0,
3562 : sizeof(AutoVacuumWorkItem) * NUM_WORKITEMS);
3563 :
3564 1235 : worker = (WorkerInfo) ((char *) AutoVacuumShmem +
3565 : MAXALIGN(sizeof(AutoVacuumShmemStruct)));
3566 :
3567 : /* initialize the WorkerInfo free list */
3568 15026 : for (int i = 0; i < autovacuum_worker_slots; i++)
3569 : {
3570 13791 : dclist_push_head(&AutoVacuumShmem->av_freeWorkers,
3571 13791 : &worker[i].wi_links);
3572 13791 : pg_atomic_init_flag(&worker[i].wi_dobalance);
3573 : }
3574 :
3575 1235 : pg_atomic_init_u32(&AutoVacuumShmem->av_nworkersForBalance, 0);
3576 1235 : }
3577 :
3578 : /*
3579 : * GUC check_hook for autovacuum_work_mem
3580 : */
3581 : bool
3582 1281 : check_autovacuum_work_mem(int *newval, void **extra, GucSource source)
3583 : {
3584 : /*
3585 : * -1 indicates fallback.
3586 : *
3587 : * If we haven't yet changed the boot_val default of -1, just let it be.
3588 : * Autovacuum will look to maintenance_work_mem instead.
3589 : */
3590 1281 : if (*newval == -1)
3591 1279 : return true;
3592 :
3593 : /*
3594 : * We clamp manually-set values to at least 64kB. Since
3595 : * maintenance_work_mem is always set to at least this value, do the same
3596 : * here.
3597 : */
3598 2 : if (*newval < 64)
3599 2 : *newval = 64;
3600 :
3601 2 : return true;
3602 : }
3603 :
3604 : /*
3605 : * Returns whether there is a free autovacuum worker slot available.
3606 : */
3607 : static bool
3608 11524 : av_worker_available(void)
3609 : {
3610 : int free_slots;
3611 : int reserved_slots;
3612 :
3613 11524 : free_slots = dclist_count(&AutoVacuumShmem->av_freeWorkers);
3614 :
3615 11524 : reserved_slots = autovacuum_worker_slots - autovacuum_max_workers;
3616 11524 : reserved_slots = Max(0, reserved_slots);
3617 :
3618 11524 : return free_slots > reserved_slots;
3619 : }
3620 :
3621 : /*
3622 : * Emits a WARNING if autovacuum_worker_slots < autovacuum_max_workers.
3623 : */
3624 : static void
3625 859 : check_av_worker_gucs(void)
3626 : {
3627 859 : if (autovacuum_worker_slots < autovacuum_max_workers)
3628 0 : ereport(WARNING,
3629 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
3630 : errmsg("\"autovacuum_max_workers\" (%d) should be less than or equal to \"autovacuum_worker_slots\" (%d)",
3631 : autovacuum_max_workers, autovacuum_worker_slots),
3632 : errdetail("The server will only start up to \"autovacuum_worker_slots\" (%d) autovacuum workers at a given time.",
3633 : autovacuum_worker_slots)));
3634 859 : }
3635 :
3636 : /*
3637 : * pg_stat_get_autovacuum_scores
3638 : *
3639 : * Returns current autovacuum scores for all relevant tables in the current
3640 : * database.
3641 : */
3642 : Datum
3643 0 : pg_stat_get_autovacuum_scores(PG_FUNCTION_ARGS)
3644 : {
3645 : int effective_multixact_freeze_max_age;
3646 : Relation rel;
3647 : TableScanDesc scan;
3648 : HeapTuple tup;
3649 0 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
3650 :
3651 0 : InitMaterializedSRF(fcinfo, 0);
3652 :
3653 : /* some prerequisite initialization */
3654 0 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
3655 0 : recentXid = ReadNextTransactionId();
3656 0 : recentMulti = ReadNextMultiXactId();
3657 :
3658 : /* scan pg_class */
3659 0 : rel = table_open(RelationRelationId, AccessShareLock);
3660 0 : scan = table_beginscan_catalog(rel, 0, NULL);
3661 0 : while ((tup = heap_getnext(scan, ForwardScanDirection)) != NULL)
3662 : {
3663 0 : Form_pg_class form = (Form_pg_class) GETSTRUCT(tup);
3664 : AutoVacOpts *avopts;
3665 : bool dovacuum;
3666 : bool doanalyze;
3667 : bool wraparound;
3668 : AutoVacuumScores scores;
3669 : Datum vals[10];
3670 0 : bool nulls[10] = {false};
3671 :
3672 : /* skip ineligible entries */
3673 0 : if (form->relkind != RELKIND_RELATION &&
3674 0 : form->relkind != RELKIND_MATVIEW &&
3675 0 : form->relkind != RELKIND_TOASTVALUE)
3676 0 : continue;
3677 0 : if (form->relpersistence == RELPERSISTENCE_TEMP)
3678 0 : continue;
3679 :
3680 0 : avopts = extract_autovac_opts(tup, RelationGetDescr(rel));
3681 0 : relation_needs_vacanalyze(form->oid, avopts, form,
3682 : effective_multixact_freeze_max_age,
3683 : LOG_NEVER,
3684 : &dovacuum, &doanalyze, &wraparound,
3685 : &scores);
3686 0 : if (avopts)
3687 0 : pfree(avopts);
3688 :
3689 0 : vals[0] = ObjectIdGetDatum(form->oid);
3690 0 : vals[1] = Float8GetDatum(scores.max);
3691 0 : vals[2] = Float8GetDatum(scores.xid);
3692 0 : vals[3] = Float8GetDatum(scores.mxid);
3693 0 : vals[4] = Float8GetDatum(scores.vac);
3694 0 : vals[5] = Float8GetDatum(scores.vac_ins);
3695 0 : vals[6] = Float8GetDatum(scores.anl);
3696 0 : vals[7] = BoolGetDatum(dovacuum);
3697 0 : vals[8] = BoolGetDatum(doanalyze);
3698 0 : vals[9] = BoolGetDatum(wraparound);
3699 :
3700 0 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, vals, nulls);
3701 : }
3702 0 : table_endscan(scan);
3703 0 : table_close(rel, AccessShareLock);
3704 :
3705 0 : return (Datum) 0;
3706 : }
|
