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) 1996-2024, PostgreSQL Global Development Group
55 : * Portions Copyright (c) 1994, Regents of the University of California
56 : *
57 : *
58 : * IDENTIFICATION
59 : * src/backend/postmaster/autovacuum.c
60 : *
61 : *-------------------------------------------------------------------------
62 : */
63 : #include "postgres.h"
64 :
65 : #include <signal.h>
66 : #include <sys/time.h>
67 : #include <unistd.h>
68 :
69 : #include "access/heapam.h"
70 : #include "access/htup_details.h"
71 : #include "access/multixact.h"
72 : #include "access/reloptions.h"
73 : #include "access/tableam.h"
74 : #include "access/transam.h"
75 : #include "access/xact.h"
76 : #include "catalog/dependency.h"
77 : #include "catalog/namespace.h"
78 : #include "catalog/pg_database.h"
79 : #include "catalog/pg_namespace.h"
80 : #include "commands/dbcommands.h"
81 : #include "commands/vacuum.h"
82 : #include "common/int.h"
83 : #include "lib/ilist.h"
84 : #include "libpq/pqsignal.h"
85 : #include "miscadmin.h"
86 : #include "nodes/makefuncs.h"
87 : #include "pgstat.h"
88 : #include "postmaster/autovacuum.h"
89 : #include "postmaster/interrupt.h"
90 : #include "postmaster/postmaster.h"
91 : #include "storage/bufmgr.h"
92 : #include "storage/ipc.h"
93 : #include "storage/latch.h"
94 : #include "storage/lmgr.h"
95 : #include "storage/pmsignal.h"
96 : #include "storage/proc.h"
97 : #include "storage/procsignal.h"
98 : #include "storage/smgr.h"
99 : #include "tcop/tcopprot.h"
100 : #include "utils/fmgroids.h"
101 : #include "utils/fmgrprotos.h"
102 : #include "utils/guc_hooks.h"
103 : #include "utils/lsyscache.h"
104 : #include "utils/memutils.h"
105 : #include "utils/ps_status.h"
106 : #include "utils/rel.h"
107 : #include "utils/snapmgr.h"
108 : #include "utils/syscache.h"
109 : #include "utils/timeout.h"
110 : #include "utils/timestamp.h"
111 :
112 :
113 : /*
114 : * GUC parameters
115 : */
116 : bool autovacuum_start_daemon = false;
117 : int autovacuum_max_workers;
118 : int autovacuum_work_mem = -1;
119 : int autovacuum_naptime;
120 : int autovacuum_vac_thresh;
121 : double autovacuum_vac_scale;
122 : int autovacuum_vac_ins_thresh;
123 : double autovacuum_vac_ins_scale;
124 : int autovacuum_anl_thresh;
125 : double autovacuum_anl_scale;
126 : int autovacuum_freeze_max_age;
127 : int autovacuum_multixact_freeze_max_age;
128 :
129 : double autovacuum_vac_cost_delay;
130 : int autovacuum_vac_cost_limit;
131 :
132 : int Log_autovacuum_min_duration = 600000;
133 :
134 : /* the minimum allowed time between two awakenings of the launcher */
135 : #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */
136 : #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */
137 :
138 : /*
139 : * Variables to save the cost-related storage parameters for the current
140 : * relation being vacuumed by this autovacuum worker. Using these, we can
141 : * ensure we don't overwrite the values of vacuum_cost_delay and
142 : * vacuum_cost_limit after reloading the configuration file. They are
143 : * initialized to "invalid" values to indicate that no cost-related storage
144 : * parameters were specified and will be set in do_autovacuum() after checking
145 : * the storage parameters in table_recheck_autovac().
146 : */
147 : static double av_storage_param_cost_delay = -1;
148 : static int av_storage_param_cost_limit = -1;
149 :
150 : /* Flags set by signal handlers */
151 : static volatile sig_atomic_t got_SIGUSR2 = false;
152 :
153 : /* Comparison points for determining whether freeze_max_age is exceeded */
154 : static TransactionId recentXid;
155 : static MultiXactId recentMulti;
156 :
157 : /* Default freeze ages to use for autovacuum (varies by database) */
158 : static int default_freeze_min_age;
159 : static int default_freeze_table_age;
160 : static int default_multixact_freeze_min_age;
161 : static int default_multixact_freeze_table_age;
162 :
163 : /* Memory context for long-lived data */
164 : static MemoryContext AutovacMemCxt;
165 :
166 : /* struct to keep track of databases in launcher */
167 : typedef struct avl_dbase
168 : {
169 : Oid adl_datid; /* hash key -- must be first */
170 : TimestampTz adl_next_worker;
171 : int adl_score;
172 : dlist_node adl_node;
173 : } avl_dbase;
174 :
175 : /* struct to keep track of databases in worker */
176 : typedef struct avw_dbase
177 : {
178 : Oid adw_datid;
179 : char *adw_name;
180 : TransactionId adw_frozenxid;
181 : MultiXactId adw_minmulti;
182 : PgStat_StatDBEntry *adw_entry;
183 : } avw_dbase;
184 :
185 : /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */
186 : typedef struct av_relation
187 : {
188 : Oid ar_toastrelid; /* hash key - must be first */
189 : Oid ar_relid;
190 : bool ar_hasrelopts;
191 : AutoVacOpts ar_reloptions; /* copy of AutoVacOpts from the main table's
192 : * reloptions, or NULL if none */
193 : } av_relation;
194 :
195 : /* struct to keep track of tables to vacuum and/or analyze, after rechecking */
196 : typedef struct autovac_table
197 : {
198 : Oid at_relid;
199 : VacuumParams at_params;
200 : double at_storage_param_vac_cost_delay;
201 : int at_storage_param_vac_cost_limit;
202 : bool at_dobalance;
203 : bool at_sharedrel;
204 : char *at_relname;
205 : char *at_nspname;
206 : char *at_datname;
207 : } autovac_table;
208 :
209 : /*-------------
210 : * This struct holds information about a single worker's whereabouts. We keep
211 : * an array of these in shared memory, sized according to
212 : * autovacuum_max_workers.
213 : *
214 : * wi_links entry into free list or running list
215 : * wi_dboid OID of the database this worker is supposed to work on
216 : * wi_tableoid OID of the table currently being vacuumed, if any
217 : * wi_sharedrel flag indicating whether table is marked relisshared
218 : * wi_proc pointer to PGPROC of the running worker, NULL if not started
219 : * wi_launchtime Time at which this worker was launched
220 : * wi_dobalance Whether this worker should be included in balance calculations
221 : *
222 : * All fields are protected by AutovacuumLock, except for wi_tableoid and
223 : * wi_sharedrel which are protected by AutovacuumScheduleLock (note these
224 : * two fields are read-only for everyone except that worker itself).
225 : *-------------
226 : */
227 : typedef struct WorkerInfoData
228 : {
229 : dlist_node wi_links;
230 : Oid wi_dboid;
231 : Oid wi_tableoid;
232 : PGPROC *wi_proc;
233 : TimestampTz wi_launchtime;
234 : pg_atomic_flag wi_dobalance;
235 : bool wi_sharedrel;
236 : } WorkerInfoData;
237 :
238 : typedef struct WorkerInfoData *WorkerInfo;
239 :
240 : /*
241 : * Possible signals received by the launcher from remote processes. These are
242 : * stored atomically in shared memory so that other processes can set them
243 : * without locking.
244 : */
245 : typedef enum
246 : {
247 : AutoVacForkFailed, /* failed trying to start a worker */
248 : AutoVacRebalance, /* rebalance the cost limits */
249 : AutoVacNumSignals, /* must be last */
250 : } AutoVacuumSignal;
251 :
252 : /*
253 : * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This
254 : * list is mostly protected by AutovacuumLock, except that if an item is
255 : * marked 'active' other processes must not modify the work-identifying
256 : * members.
257 : */
258 : typedef struct AutoVacuumWorkItem
259 : {
260 : AutoVacuumWorkItemType avw_type;
261 : bool avw_used; /* below data is valid */
262 : bool avw_active; /* being processed */
263 : Oid avw_database;
264 : Oid avw_relation;
265 : BlockNumber avw_blockNumber;
266 : } AutoVacuumWorkItem;
267 :
268 : #define NUM_WORKITEMS 256
269 :
270 : /*-------------
271 : * The main autovacuum shmem struct. On shared memory we store this main
272 : * struct and the array of WorkerInfo structs. This struct keeps:
273 : *
274 : * av_signal set by other processes to indicate various conditions
275 : * av_launcherpid the PID of the autovacuum launcher
276 : * av_freeWorkers the WorkerInfo freelist
277 : * av_runningWorkers the WorkerInfo non-free queue
278 : * av_startingWorker pointer to WorkerInfo currently being started (cleared by
279 : * the worker itself as soon as it's up and running)
280 : * av_workItems work item array
281 : * av_nworkersForBalance the number of autovacuum workers to use when
282 : * calculating the per worker cost limit
283 : *
284 : * This struct is protected by AutovacuumLock, except for av_signal and parts
285 : * of the worker list (see above).
286 : *-------------
287 : */
288 : typedef struct
289 : {
290 : sig_atomic_t av_signal[AutoVacNumSignals];
291 : pid_t av_launcherpid;
292 : dlist_head av_freeWorkers;
293 : dlist_head av_runningWorkers;
294 : WorkerInfo av_startingWorker;
295 : AutoVacuumWorkItem av_workItems[NUM_WORKITEMS];
296 : pg_atomic_uint32 av_nworkersForBalance;
297 : } AutoVacuumShmemStruct;
298 :
299 : static AutoVacuumShmemStruct *AutoVacuumShmem;
300 :
301 : /*
302 : * the database list (of avl_dbase elements) in the launcher, and the context
303 : * that contains it
304 : */
305 : static dlist_head DatabaseList = DLIST_STATIC_INIT(DatabaseList);
306 : static MemoryContext DatabaseListCxt = NULL;
307 :
308 : /* Pointer to my own WorkerInfo, valid on each worker */
309 : static WorkerInfo MyWorkerInfo = NULL;
310 :
311 : /* PID of launcher, valid only in worker while shutting down */
312 : int AutovacuumLauncherPid = 0;
313 :
314 : static Oid do_start_worker(void);
315 : static void HandleAutoVacLauncherInterrupts(void);
316 : static void AutoVacLauncherShutdown(void) pg_attribute_noreturn();
317 : static void launcher_determine_sleep(bool canlaunch, bool recursing,
318 : struct timeval *nap);
319 : static void launch_worker(TimestampTz now);
320 : static List *get_database_list(void);
321 : static void rebuild_database_list(Oid newdb);
322 : static int db_comparator(const void *a, const void *b);
323 : static void autovac_recalculate_workers_for_balance(void);
324 :
325 : static void do_autovacuum(void);
326 : static void FreeWorkerInfo(int code, Datum arg);
327 :
328 : static autovac_table *table_recheck_autovac(Oid relid, HTAB *table_toast_map,
329 : TupleDesc pg_class_desc,
330 : int effective_multixact_freeze_max_age);
331 : static void recheck_relation_needs_vacanalyze(Oid relid, AutoVacOpts *avopts,
332 : Form_pg_class classForm,
333 : int effective_multixact_freeze_max_age,
334 : bool *dovacuum, bool *doanalyze, bool *wraparound);
335 : static void relation_needs_vacanalyze(Oid relid, AutoVacOpts *relopts,
336 : Form_pg_class classForm,
337 : PgStat_StatTabEntry *tabentry,
338 : int effective_multixact_freeze_max_age,
339 : bool *dovacuum, bool *doanalyze, bool *wraparound);
340 :
341 : static void autovacuum_do_vac_analyze(autovac_table *tab,
342 : BufferAccessStrategy bstrategy);
343 : static AutoVacOpts *extract_autovac_opts(HeapTuple tup,
344 : TupleDesc pg_class_desc);
345 : static void perform_work_item(AutoVacuumWorkItem *workitem);
346 : static void autovac_report_activity(autovac_table *tab);
347 : static void autovac_report_workitem(AutoVacuumWorkItem *workitem,
348 : const char *nspname, const char *relname);
349 : static void avl_sigusr2_handler(SIGNAL_ARGS);
350 :
351 :
352 :
353 : /********************************************************************
354 : * AUTOVACUUM LAUNCHER CODE
355 : ********************************************************************/
356 :
357 : /*
358 : * Main entry point for the autovacuum launcher process.
359 : */
360 : void
361 652 : AutoVacLauncherMain(char *startup_data, size_t startup_data_len)
362 : {
363 : sigjmp_buf local_sigjmp_buf;
364 :
365 : Assert(startup_data_len == 0);
366 :
367 : /* Release postmaster's working memory context */
368 652 : if (PostmasterContext)
369 : {
370 652 : MemoryContextDelete(PostmasterContext);
371 652 : PostmasterContext = NULL;
372 : }
373 :
374 652 : MyBackendType = B_AUTOVAC_LAUNCHER;
375 652 : init_ps_display(NULL);
376 :
377 652 : ereport(DEBUG1,
378 : (errmsg_internal("autovacuum launcher started")));
379 :
380 652 : if (PostAuthDelay)
381 0 : pg_usleep(PostAuthDelay * 1000000L);
382 :
383 652 : SetProcessingMode(InitProcessing);
384 :
385 : /*
386 : * Set up signal handlers. We operate on databases much like a regular
387 : * backend, so we use the same signal handling. See equivalent code in
388 : * tcop/postgres.c.
389 : */
390 652 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
391 652 : pqsignal(SIGINT, StatementCancelHandler);
392 652 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
393 : /* SIGQUIT handler was already set up by InitPostmasterChild */
394 :
395 652 : InitializeTimeouts(); /* establishes SIGALRM handler */
396 :
397 652 : pqsignal(SIGPIPE, SIG_IGN);
398 652 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
399 652 : pqsignal(SIGUSR2, avl_sigusr2_handler);
400 652 : pqsignal(SIGFPE, FloatExceptionHandler);
401 652 : pqsignal(SIGCHLD, SIG_DFL);
402 :
403 : /*
404 : * Create a per-backend PGPROC struct in shared memory. We must do this
405 : * before we can use LWLocks or access any shared memory.
406 : */
407 652 : InitProcess();
408 :
409 : /* Early initialization */
410 652 : BaseInit();
411 :
412 652 : InitPostgres(NULL, InvalidOid, NULL, InvalidOid, 0, NULL);
413 :
414 652 : SetProcessingMode(NormalProcessing);
415 :
416 : /*
417 : * Create a memory context that we will do all our work in. We do this so
418 : * that we can reset the context during error recovery and thereby avoid
419 : * possible memory leaks.
420 : */
421 652 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
422 : "Autovacuum Launcher",
423 : ALLOCSET_DEFAULT_SIZES);
424 652 : MemoryContextSwitchTo(AutovacMemCxt);
425 :
426 : /*
427 : * If an exception is encountered, processing resumes here.
428 : *
429 : * This code is a stripped down version of PostgresMain error recovery.
430 : *
431 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
432 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
433 : * signals other than SIGQUIT will be blocked until we complete error
434 : * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
435 : * call redundant, but it is not since InterruptPending might be set
436 : * already.
437 : */
438 652 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
439 : {
440 : /* since not using PG_TRY, must reset error stack by hand */
441 0 : error_context_stack = NULL;
442 :
443 : /* Prevents interrupts while cleaning up */
444 0 : HOLD_INTERRUPTS();
445 :
446 : /* Forget any pending QueryCancel or timeout request */
447 0 : disable_all_timeouts(false);
448 0 : QueryCancelPending = false; /* second to avoid race condition */
449 :
450 : /* Report the error to the server log */
451 0 : EmitErrorReport();
452 :
453 : /* Abort the current transaction in order to recover */
454 0 : AbortCurrentTransaction();
455 :
456 : /*
457 : * Release any other resources, for the case where we were not in a
458 : * transaction.
459 : */
460 0 : LWLockReleaseAll();
461 0 : pgstat_report_wait_end();
462 0 : UnlockBuffers();
463 : /* this is probably dead code, but let's be safe: */
464 0 : if (AuxProcessResourceOwner)
465 0 : ReleaseAuxProcessResources(false);
466 0 : AtEOXact_Buffers(false);
467 0 : AtEOXact_SMgr();
468 0 : AtEOXact_Files(false);
469 0 : AtEOXact_HashTables(false);
470 :
471 : /*
472 : * Now return to normal top-level context and clear ErrorContext for
473 : * next time.
474 : */
475 0 : MemoryContextSwitchTo(AutovacMemCxt);
476 0 : FlushErrorState();
477 :
478 : /* Flush any leaked data in the top-level context */
479 0 : MemoryContextReset(AutovacMemCxt);
480 :
481 : /* don't leave dangling pointers to freed memory */
482 0 : DatabaseListCxt = NULL;
483 0 : dlist_init(&DatabaseList);
484 :
485 : /* Now we can allow interrupts again */
486 0 : RESUME_INTERRUPTS();
487 :
488 : /* if in shutdown mode, no need for anything further; just go away */
489 0 : if (ShutdownRequestPending)
490 0 : AutoVacLauncherShutdown();
491 :
492 : /*
493 : * Sleep at least 1 second after any error. We don't want to be
494 : * filling the error logs as fast as we can.
495 : */
496 0 : pg_usleep(1000000L);
497 : }
498 :
499 : /* We can now handle ereport(ERROR) */
500 652 : PG_exception_stack = &local_sigjmp_buf;
501 :
502 : /* must unblock signals before calling rebuild_database_list */
503 652 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
504 :
505 : /*
506 : * Set always-secure search path. Launcher doesn't connect to a database,
507 : * so this has no effect.
508 : */
509 652 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
510 :
511 : /*
512 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
513 : * in postgresql.conf. We don't really want such a dangerous option being
514 : * applied non-interactively.
515 : */
516 652 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
517 :
518 : /*
519 : * Force settable timeouts off to avoid letting these settings prevent
520 : * regular maintenance from being executed.
521 : */
522 652 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
523 652 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
524 652 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
525 652 : SetConfigOption("idle_in_transaction_session_timeout", "0",
526 : PGC_SUSET, PGC_S_OVERRIDE);
527 :
528 : /*
529 : * Force default_transaction_isolation to READ COMMITTED. We don't want
530 : * to pay the overhead of serializable mode, nor add any risk of causing
531 : * deadlocks or delaying other transactions.
532 : */
533 652 : SetConfigOption("default_transaction_isolation", "read committed",
534 : PGC_SUSET, PGC_S_OVERRIDE);
535 :
536 : /*
537 : * Even when system is configured to use a different fetch consistency,
538 : * for autovac we always want fresh stats.
539 : */
540 652 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
541 :
542 : /*
543 : * In emergency mode, just start a worker (unless shutdown was requested)
544 : * and go away.
545 : */
546 652 : if (!AutoVacuumingActive())
547 : {
548 0 : if (!ShutdownRequestPending)
549 0 : do_start_worker();
550 0 : proc_exit(0); /* done */
551 : }
552 :
553 652 : AutoVacuumShmem->av_launcherpid = MyProcPid;
554 :
555 : /*
556 : * Create the initial database list. The invariant we want this list to
557 : * keep is that it's ordered by decreasing next_time. As soon as an entry
558 : * is updated to a higher time, it will be moved to the front (which is
559 : * correct because the only operation is to add autovacuum_naptime to the
560 : * entry, and time always increases).
561 : */
562 652 : rebuild_database_list(InvalidOid);
563 :
564 : /* loop until shutdown request */
565 3420 : while (!ShutdownRequestPending)
566 : {
567 : struct timeval nap;
568 3418 : TimestampTz current_time = 0;
569 : bool can_launch;
570 :
571 : /*
572 : * This loop is a bit different from the normal use of WaitLatch,
573 : * because we'd like to sleep before the first launch of a child
574 : * process. So it's WaitLatch, then ResetLatch, then check for
575 : * wakening conditions.
576 : */
577 :
578 3418 : launcher_determine_sleep(!dlist_is_empty(&AutoVacuumShmem->av_freeWorkers),
579 3418 : false, &nap);
580 :
581 : /*
582 : * Wait until naptime expires or we get some type of signal (all the
583 : * signal handlers will wake us by calling SetLatch).
584 : */
585 3418 : (void) WaitLatch(MyLatch,
586 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
587 3418 : (nap.tv_sec * 1000L) + (nap.tv_usec / 1000L),
588 : WAIT_EVENT_AUTOVACUUM_MAIN);
589 :
590 3412 : ResetLatch(MyLatch);
591 :
592 3412 : HandleAutoVacLauncherInterrupts();
593 :
594 : /*
595 : * a worker finished, or postmaster signaled failure to start a worker
596 : */
597 2768 : if (got_SIGUSR2)
598 : {
599 40 : got_SIGUSR2 = false;
600 :
601 : /* rebalance cost limits, if needed */
602 40 : if (AutoVacuumShmem->av_signal[AutoVacRebalance])
603 : {
604 20 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
605 20 : AutoVacuumShmem->av_signal[AutoVacRebalance] = false;
606 20 : autovac_recalculate_workers_for_balance();
607 20 : LWLockRelease(AutovacuumLock);
608 : }
609 :
610 40 : if (AutoVacuumShmem->av_signal[AutoVacForkFailed])
611 : {
612 : /*
613 : * If the postmaster failed to start a new worker, we sleep
614 : * for a little while and resend the signal. The new worker's
615 : * state is still in memory, so this is sufficient. After
616 : * that, we restart the main loop.
617 : *
618 : * XXX should we put a limit to the number of times we retry?
619 : * I don't think it makes much sense, because a future start
620 : * of a worker will continue to fail in the same way.
621 : */
622 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = false;
623 0 : pg_usleep(1000000L); /* 1s */
624 0 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
625 0 : continue;
626 : }
627 : }
628 :
629 : /*
630 : * There are some conditions that we need to check before trying to
631 : * start a worker. First, we need to make sure that there is a worker
632 : * slot available. Second, we need to make sure that no other worker
633 : * failed while starting up.
634 : */
635 :
636 2768 : current_time = GetCurrentTimestamp();
637 2768 : LWLockAcquire(AutovacuumLock, LW_SHARED);
638 :
639 2768 : can_launch = !dlist_is_empty(&AutoVacuumShmem->av_freeWorkers);
640 :
641 2768 : if (AutoVacuumShmem->av_startingWorker != NULL)
642 : {
643 : int waittime;
644 0 : WorkerInfo worker = AutoVacuumShmem->av_startingWorker;
645 :
646 : /*
647 : * We can't launch another worker when another one is still
648 : * starting up (or failed while doing so), so just sleep for a bit
649 : * more; that worker will wake us up again as soon as it's ready.
650 : * We will only wait autovacuum_naptime seconds (up to a maximum
651 : * of 60 seconds) for this to happen however. Note that failure
652 : * to connect to a particular database is not a problem here,
653 : * because the worker removes itself from the startingWorker
654 : * pointer before trying to connect. Problems detected by the
655 : * postmaster (like fork() failure) are also reported and handled
656 : * differently. The only problems that may cause this code to
657 : * fire are errors in the earlier sections of AutoVacWorkerMain,
658 : * before the worker removes the WorkerInfo from the
659 : * startingWorker pointer.
660 : */
661 0 : waittime = Min(autovacuum_naptime, 60) * 1000;
662 0 : if (TimestampDifferenceExceeds(worker->wi_launchtime, current_time,
663 : waittime))
664 : {
665 0 : LWLockRelease(AutovacuumLock);
666 0 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
667 :
668 : /*
669 : * No other process can put a worker in starting mode, so if
670 : * startingWorker is still INVALID after exchanging our lock,
671 : * we assume it's the same one we saw above (so we don't
672 : * recheck the launch time).
673 : */
674 0 : if (AutoVacuumShmem->av_startingWorker != NULL)
675 : {
676 0 : worker = AutoVacuumShmem->av_startingWorker;
677 0 : worker->wi_dboid = InvalidOid;
678 0 : worker->wi_tableoid = InvalidOid;
679 0 : worker->wi_sharedrel = false;
680 0 : worker->wi_proc = NULL;
681 0 : worker->wi_launchtime = 0;
682 0 : dlist_push_head(&AutoVacuumShmem->av_freeWorkers,
683 : &worker->wi_links);
684 0 : AutoVacuumShmem->av_startingWorker = NULL;
685 0 : ereport(WARNING,
686 : errmsg("autovacuum worker took too long to start; canceled"));
687 : }
688 : }
689 : else
690 0 : can_launch = false;
691 : }
692 2768 : LWLockRelease(AutovacuumLock); /* either shared or exclusive */
693 :
694 : /* if we can't do anything, just go back to sleep */
695 2768 : if (!can_launch)
696 0 : continue;
697 :
698 : /* We're OK to start a new worker */
699 :
700 2768 : if (dlist_is_empty(&DatabaseList))
701 : {
702 : /*
703 : * Special case when the list is empty: start a worker right away.
704 : * This covers the initial case, when no database is in pgstats
705 : * (thus the list is empty). Note that the constraints in
706 : * launcher_determine_sleep keep us from starting workers too
707 : * quickly (at most once every autovacuum_naptime when the list is
708 : * empty).
709 : */
710 62 : launch_worker(current_time);
711 : }
712 : else
713 : {
714 : /*
715 : * because rebuild_database_list constructs a list with most
716 : * distant adl_next_worker first, we obtain our database from the
717 : * tail of the list.
718 : */
719 : avl_dbase *avdb;
720 :
721 2706 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
722 :
723 : /*
724 : * launch a worker if next_worker is right now or it is in the
725 : * past
726 : */
727 2706 : if (TimestampDifferenceExceeds(avdb->adl_next_worker,
728 : current_time, 0))
729 20 : launch_worker(current_time);
730 : }
731 : }
732 :
733 2 : AutoVacLauncherShutdown();
734 : }
735 :
736 : /*
737 : * Process any new interrupts.
738 : */
739 : static void
740 3412 : HandleAutoVacLauncherInterrupts(void)
741 : {
742 : /* the normal shutdown case */
743 3412 : if (ShutdownRequestPending)
744 644 : AutoVacLauncherShutdown();
745 :
746 2768 : if (ConfigReloadPending)
747 : {
748 56 : ConfigReloadPending = false;
749 56 : ProcessConfigFile(PGC_SIGHUP);
750 :
751 : /* shutdown requested in config file? */
752 56 : if (!AutoVacuumingActive())
753 0 : AutoVacLauncherShutdown();
754 :
755 : /* rebuild the list in case the naptime changed */
756 56 : rebuild_database_list(InvalidOid);
757 : }
758 :
759 : /* Process barrier events */
760 2768 : if (ProcSignalBarrierPending)
761 42 : ProcessProcSignalBarrier();
762 :
763 : /* Perform logging of memory contexts of this process */
764 2768 : if (LogMemoryContextPending)
765 0 : ProcessLogMemoryContextInterrupt();
766 :
767 : /* Process sinval catchup interrupts that happened while sleeping */
768 2768 : ProcessCatchupInterrupt();
769 2768 : }
770 :
771 : /*
772 : * Perform a normal exit from the autovac launcher.
773 : */
774 : static void
775 646 : AutoVacLauncherShutdown(void)
776 : {
777 646 : ereport(DEBUG1,
778 : (errmsg_internal("autovacuum launcher shutting down")));
779 646 : AutoVacuumShmem->av_launcherpid = 0;
780 :
781 646 : proc_exit(0); /* done */
782 : }
783 :
784 : /*
785 : * Determine the time to sleep, based on the database list.
786 : *
787 : * The "canlaunch" parameter indicates whether we can start a worker right now,
788 : * for example due to the workers being all busy. If this is false, we will
789 : * cause a long sleep, which will be interrupted when a worker exits.
790 : */
791 : static void
792 3418 : launcher_determine_sleep(bool canlaunch, bool recursing, struct timeval *nap)
793 : {
794 : /*
795 : * We sleep until the next scheduled vacuum. We trust that when the
796 : * database list was built, care was taken so that no entries have times
797 : * in the past; if the first entry has too close a next_worker value, or a
798 : * time in the past, we will sleep a small nominal time.
799 : */
800 3418 : if (!canlaunch)
801 : {
802 0 : nap->tv_sec = autovacuum_naptime;
803 0 : nap->tv_usec = 0;
804 : }
805 3418 : else if (!dlist_is_empty(&DatabaseList))
806 : {
807 3294 : TimestampTz current_time = GetCurrentTimestamp();
808 : TimestampTz next_wakeup;
809 : avl_dbase *avdb;
810 : long secs;
811 : int usecs;
812 :
813 3294 : avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList);
814 :
815 3294 : next_wakeup = avdb->adl_next_worker;
816 3294 : TimestampDifference(current_time, next_wakeup, &secs, &usecs);
817 :
818 3294 : nap->tv_sec = secs;
819 3294 : nap->tv_usec = usecs;
820 : }
821 : else
822 : {
823 : /* list is empty, sleep for whole autovacuum_naptime seconds */
824 124 : nap->tv_sec = autovacuum_naptime;
825 124 : nap->tv_usec = 0;
826 : }
827 :
828 : /*
829 : * If the result is exactly zero, it means a database had an entry with
830 : * time in the past. Rebuild the list so that the databases are evenly
831 : * distributed again, and recalculate the time to sleep. This can happen
832 : * if there are more tables needing vacuum than workers, and they all take
833 : * longer to vacuum than autovacuum_naptime.
834 : *
835 : * We only recurse once. rebuild_database_list should always return times
836 : * in the future, but it seems best not to trust too much on that.
837 : */
838 3418 : if (nap->tv_sec == 0 && nap->tv_usec == 0 && !recursing)
839 : {
840 0 : rebuild_database_list(InvalidOid);
841 0 : launcher_determine_sleep(canlaunch, true, nap);
842 0 : return;
843 : }
844 :
845 : /* The smallest time we'll allow the launcher to sleep. */
846 3418 : if (nap->tv_sec <= 0 && nap->tv_usec <= MIN_AUTOVAC_SLEEPTIME * 1000)
847 : {
848 4 : nap->tv_sec = 0;
849 4 : nap->tv_usec = MIN_AUTOVAC_SLEEPTIME * 1000;
850 : }
851 :
852 : /*
853 : * If the sleep time is too large, clamp it to an arbitrary maximum (plus
854 : * any fractional seconds, for simplicity). This avoids an essentially
855 : * infinite sleep in strange cases like the system clock going backwards a
856 : * few years.
857 : */
858 3418 : if (nap->tv_sec > MAX_AUTOVAC_SLEEPTIME)
859 22 : nap->tv_sec = MAX_AUTOVAC_SLEEPTIME;
860 : }
861 :
862 : /*
863 : * Build an updated DatabaseList. It must only contain databases that appear
864 : * in pgstats, and must be sorted by next_worker from highest to lowest,
865 : * distributed regularly across the next autovacuum_naptime interval.
866 : *
867 : * Receives the Oid of the database that made this list be generated (we call
868 : * this the "new" database, because when the database was already present on
869 : * the list, we expect that this function is not called at all). The
870 : * preexisting list, if any, will be used to preserve the order of the
871 : * databases in the autovacuum_naptime period. The new database is put at the
872 : * end of the interval. The actual values are not saved, which should not be
873 : * much of a problem.
874 : */
875 : static void
876 716 : rebuild_database_list(Oid newdb)
877 : {
878 : List *dblist;
879 : ListCell *cell;
880 : MemoryContext newcxt;
881 : MemoryContext oldcxt;
882 : MemoryContext tmpcxt;
883 : HASHCTL hctl;
884 : int score;
885 : int nelems;
886 : HTAB *dbhash;
887 : dlist_iter iter;
888 :
889 716 : newcxt = AllocSetContextCreate(AutovacMemCxt,
890 : "Autovacuum database list",
891 : ALLOCSET_DEFAULT_SIZES);
892 716 : tmpcxt = AllocSetContextCreate(newcxt,
893 : "Autovacuum database list (tmp)",
894 : ALLOCSET_DEFAULT_SIZES);
895 716 : oldcxt = MemoryContextSwitchTo(tmpcxt);
896 :
897 : /*
898 : * Implementing this is not as simple as it sounds, because we need to put
899 : * the new database at the end of the list; next the databases that were
900 : * already on the list, and finally (at the tail of the list) all the
901 : * other databases that are not on the existing list.
902 : *
903 : * To do this, we build an empty hash table of scored databases. We will
904 : * start with the lowest score (zero) for the new database, then
905 : * increasing scores for the databases in the existing list, in order, and
906 : * lastly increasing scores for all databases gotten via
907 : * get_database_list() that are not already on the hash.
908 : *
909 : * Then we will put all the hash elements into an array, sort the array by
910 : * score, and finally put the array elements into the new doubly linked
911 : * list.
912 : */
913 716 : hctl.keysize = sizeof(Oid);
914 716 : hctl.entrysize = sizeof(avl_dbase);
915 716 : hctl.hcxt = tmpcxt;
916 716 : dbhash = hash_create("autovacuum db hash", 20, &hctl, /* magic number here
917 : * FIXME */
918 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
919 :
920 : /* start by inserting the new database */
921 716 : score = 0;
922 716 : if (OidIsValid(newdb))
923 : {
924 : avl_dbase *db;
925 : PgStat_StatDBEntry *entry;
926 :
927 : /* only consider this database if it has a pgstat entry */
928 8 : entry = pgstat_fetch_stat_dbentry(newdb);
929 8 : if (entry != NULL)
930 : {
931 : /* we assume it isn't found because the hash was just created */
932 8 : db = hash_search(dbhash, &newdb, HASH_ENTER, NULL);
933 :
934 : /* hash_search already filled in the key */
935 8 : db->adl_score = score++;
936 : /* next_worker is filled in later */
937 : }
938 : }
939 :
940 : /* Now insert the databases from the existing list */
941 808 : dlist_foreach(iter, &DatabaseList)
942 : {
943 92 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
944 : avl_dbase *db;
945 : bool found;
946 : PgStat_StatDBEntry *entry;
947 :
948 : /*
949 : * skip databases with no stat entries -- in particular, this gets rid
950 : * of dropped databases
951 : */
952 92 : entry = pgstat_fetch_stat_dbentry(avdb->adl_datid);
953 92 : if (entry == NULL)
954 0 : continue;
955 :
956 92 : db = hash_search(dbhash, &(avdb->adl_datid), HASH_ENTER, &found);
957 :
958 92 : if (!found)
959 : {
960 : /* hash_search already filled in the key */
961 92 : db->adl_score = score++;
962 : /* next_worker is filled in later */
963 : }
964 : }
965 :
966 : /* finally, insert all qualifying databases not previously inserted */
967 716 : dblist = get_database_list();
968 3170 : foreach(cell, dblist)
969 : {
970 2454 : avw_dbase *avdb = lfirst(cell);
971 : avl_dbase *db;
972 : bool found;
973 : PgStat_StatDBEntry *entry;
974 :
975 : /* only consider databases with a pgstat entry */
976 2454 : entry = pgstat_fetch_stat_dbentry(avdb->adw_datid);
977 2454 : if (entry == NULL)
978 1510 : continue;
979 :
980 944 : db = hash_search(dbhash, &(avdb->adw_datid), HASH_ENTER, &found);
981 : /* only update the score if the database was not already on the hash */
982 944 : if (!found)
983 : {
984 : /* hash_search already filled in the key */
985 844 : db->adl_score = score++;
986 : /* next_worker is filled in later */
987 : }
988 : }
989 716 : nelems = score;
990 :
991 : /* from here on, the allocated memory belongs to the new list */
992 716 : MemoryContextSwitchTo(newcxt);
993 716 : dlist_init(&DatabaseList);
994 :
995 716 : if (nelems > 0)
996 : {
997 : TimestampTz current_time;
998 : int millis_increment;
999 : avl_dbase *dbary;
1000 : avl_dbase *db;
1001 : HASH_SEQ_STATUS seq;
1002 : int i;
1003 :
1004 : /* put all the hash elements into an array */
1005 652 : dbary = palloc(nelems * sizeof(avl_dbase));
1006 :
1007 652 : i = 0;
1008 652 : hash_seq_init(&seq, dbhash);
1009 1596 : while ((db = hash_seq_search(&seq)) != NULL)
1010 944 : memcpy(&(dbary[i++]), db, sizeof(avl_dbase));
1011 :
1012 : /* sort the array */
1013 652 : qsort(dbary, nelems, sizeof(avl_dbase), db_comparator);
1014 :
1015 : /*
1016 : * Determine the time interval between databases in the schedule. If
1017 : * we see that the configured naptime would take us to sleep times
1018 : * lower than our min sleep time (which launcher_determine_sleep is
1019 : * coded not to allow), silently use a larger naptime (but don't touch
1020 : * the GUC variable).
1021 : */
1022 652 : millis_increment = 1000.0 * autovacuum_naptime / nelems;
1023 652 : if (millis_increment <= MIN_AUTOVAC_SLEEPTIME)
1024 0 : millis_increment = MIN_AUTOVAC_SLEEPTIME * 1.1;
1025 :
1026 652 : current_time = GetCurrentTimestamp();
1027 :
1028 : /*
1029 : * move the elements from the array into the dlist, setting the
1030 : * next_worker while walking the array
1031 : */
1032 1596 : for (i = 0; i < nelems; i++)
1033 : {
1034 944 : db = &(dbary[i]);
1035 :
1036 944 : current_time = TimestampTzPlusMilliseconds(current_time,
1037 : millis_increment);
1038 944 : db->adl_next_worker = current_time;
1039 :
1040 : /* later elements should go closer to the head of the list */
1041 944 : dlist_push_head(&DatabaseList, &db->adl_node);
1042 : }
1043 : }
1044 :
1045 : /* all done, clean up memory */
1046 716 : if (DatabaseListCxt != NULL)
1047 64 : MemoryContextDelete(DatabaseListCxt);
1048 716 : MemoryContextDelete(tmpcxt);
1049 716 : DatabaseListCxt = newcxt;
1050 716 : MemoryContextSwitchTo(oldcxt);
1051 716 : }
1052 :
1053 : /* qsort comparator for avl_dbase, using adl_score */
1054 : static int
1055 424 : db_comparator(const void *a, const void *b)
1056 : {
1057 424 : return pg_cmp_s32(((const avl_dbase *) a)->adl_score,
1058 : ((const avl_dbase *) b)->adl_score);
1059 : }
1060 :
1061 : /*
1062 : * do_start_worker
1063 : *
1064 : * Bare-bones procedure for starting an autovacuum worker from the launcher.
1065 : * It determines what database to work on, sets up shared memory stuff and
1066 : * signals postmaster to start the worker. It fails gracefully if invoked when
1067 : * autovacuum_workers are already active.
1068 : *
1069 : * Return value is the OID of the database that the worker is going to process,
1070 : * or InvalidOid if no worker was actually started.
1071 : */
1072 : static Oid
1073 82 : do_start_worker(void)
1074 : {
1075 : List *dblist;
1076 : ListCell *cell;
1077 : TransactionId xidForceLimit;
1078 : MultiXactId multiForceLimit;
1079 : bool for_xid_wrap;
1080 : bool for_multi_wrap;
1081 : avw_dbase *avdb;
1082 : TimestampTz current_time;
1083 82 : bool skipit = false;
1084 82 : Oid retval = InvalidOid;
1085 : MemoryContext tmpcxt,
1086 : oldcxt;
1087 :
1088 : /* return quickly when there are no free workers */
1089 82 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1090 82 : if (dlist_is_empty(&AutoVacuumShmem->av_freeWorkers))
1091 : {
1092 0 : LWLockRelease(AutovacuumLock);
1093 0 : return InvalidOid;
1094 : }
1095 82 : LWLockRelease(AutovacuumLock);
1096 :
1097 : /*
1098 : * Create and switch to a temporary context to avoid leaking the memory
1099 : * allocated for the database list.
1100 : */
1101 82 : tmpcxt = AllocSetContextCreate(CurrentMemoryContext,
1102 : "Autovacuum start worker (tmp)",
1103 : ALLOCSET_DEFAULT_SIZES);
1104 82 : oldcxt = MemoryContextSwitchTo(tmpcxt);
1105 :
1106 : /* Get a list of databases */
1107 82 : dblist = get_database_list();
1108 :
1109 : /*
1110 : * Determine the oldest datfrozenxid/relfrozenxid that we will allow to
1111 : * pass without forcing a vacuum. (This limit can be tightened for
1112 : * particular tables, but not loosened.)
1113 : */
1114 82 : recentXid = ReadNextTransactionId();
1115 82 : xidForceLimit = recentXid - autovacuum_freeze_max_age;
1116 : /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */
1117 : /* this can cause the limit to go backwards by 3, but that's OK */
1118 82 : if (xidForceLimit < FirstNormalTransactionId)
1119 0 : xidForceLimit -= FirstNormalTransactionId;
1120 :
1121 : /* Also determine the oldest datminmxid we will consider. */
1122 82 : recentMulti = ReadNextMultiXactId();
1123 82 : multiForceLimit = recentMulti - MultiXactMemberFreezeThreshold();
1124 82 : if (multiForceLimit < FirstMultiXactId)
1125 0 : multiForceLimit -= FirstMultiXactId;
1126 :
1127 : /*
1128 : * Choose a database to connect to. We pick the database that was least
1129 : * recently auto-vacuumed, or one that needs vacuuming to prevent Xid
1130 : * wraparound-related data loss. If any db at risk of Xid wraparound is
1131 : * found, we pick the one with oldest datfrozenxid, independently of
1132 : * autovacuum times; similarly we pick the one with the oldest datminmxid
1133 : * if any is in MultiXactId wraparound. Note that those in Xid wraparound
1134 : * danger are given more priority than those in multi wraparound danger.
1135 : *
1136 : * Note that a database with no stats entry is not considered, except for
1137 : * Xid wraparound purposes. The theory is that if no one has ever
1138 : * connected to it since the stats were last initialized, it doesn't need
1139 : * vacuuming.
1140 : *
1141 : * XXX This could be improved if we had more info about whether it needs
1142 : * vacuuming before connecting to it. Perhaps look through the pgstats
1143 : * data for the database's tables? One idea is to keep track of the
1144 : * number of new and dead tuples per database in pgstats. However it
1145 : * isn't clear how to construct a metric that measures that and not cause
1146 : * starvation for less busy databases.
1147 : */
1148 82 : avdb = NULL;
1149 82 : for_xid_wrap = false;
1150 82 : for_multi_wrap = false;
1151 82 : current_time = GetCurrentTimestamp();
1152 380 : foreach(cell, dblist)
1153 : {
1154 298 : avw_dbase *tmp = lfirst(cell);
1155 : dlist_iter iter;
1156 :
1157 : /* Check to see if this one is at risk of wraparound */
1158 298 : if (TransactionIdPrecedes(tmp->adw_frozenxid, xidForceLimit))
1159 : {
1160 0 : if (avdb == NULL ||
1161 0 : TransactionIdPrecedes(tmp->adw_frozenxid,
1162 : avdb->adw_frozenxid))
1163 0 : avdb = tmp;
1164 0 : for_xid_wrap = true;
1165 264 : continue;
1166 : }
1167 298 : else if (for_xid_wrap)
1168 0 : continue; /* ignore not-at-risk DBs */
1169 298 : else if (MultiXactIdPrecedes(tmp->adw_minmulti, multiForceLimit))
1170 : {
1171 0 : if (avdb == NULL ||
1172 0 : MultiXactIdPrecedes(tmp->adw_minmulti, avdb->adw_minmulti))
1173 0 : avdb = tmp;
1174 0 : for_multi_wrap = true;
1175 0 : continue;
1176 : }
1177 298 : else if (for_multi_wrap)
1178 0 : continue; /* ignore not-at-risk DBs */
1179 :
1180 : /* Find pgstat entry if any */
1181 298 : tmp->adw_entry = pgstat_fetch_stat_dbentry(tmp->adw_datid);
1182 :
1183 : /*
1184 : * Skip a database with no pgstat entry; it means it hasn't seen any
1185 : * activity.
1186 : */
1187 298 : if (!tmp->adw_entry)
1188 240 : continue;
1189 :
1190 : /*
1191 : * Also, skip a database that appears on the database list as having
1192 : * been processed recently (less than autovacuum_naptime seconds ago).
1193 : * We do this so that we don't select a database which we just
1194 : * selected, but that pgstat hasn't gotten around to updating the last
1195 : * autovacuum time yet.
1196 : */
1197 58 : skipit = false;
1198 :
1199 108 : dlist_reverse_foreach(iter, &DatabaseList)
1200 : {
1201 94 : avl_dbase *dbp = dlist_container(avl_dbase, adl_node, iter.cur);
1202 :
1203 94 : if (dbp->adl_datid == tmp->adw_datid)
1204 : {
1205 : /*
1206 : * Skip this database if its next_worker value falls between
1207 : * the current time and the current time plus naptime.
1208 : */
1209 44 : if (!TimestampDifferenceExceeds(dbp->adl_next_worker,
1210 24 : current_time, 0) &&
1211 24 : !TimestampDifferenceExceeds(current_time,
1212 : dbp->adl_next_worker,
1213 : autovacuum_naptime * 1000))
1214 24 : skipit = true;
1215 :
1216 44 : break;
1217 : }
1218 : }
1219 58 : if (skipit)
1220 24 : continue;
1221 :
1222 : /*
1223 : * Remember the db with oldest autovac time. (If we are here, both
1224 : * tmp->entry and db->entry must be non-null.)
1225 : */
1226 34 : if (avdb == NULL ||
1227 14 : tmp->adw_entry->last_autovac_time < avdb->adw_entry->last_autovac_time)
1228 20 : avdb = tmp;
1229 : }
1230 :
1231 : /* Found a database -- process it */
1232 82 : if (avdb != NULL)
1233 : {
1234 : WorkerInfo worker;
1235 : dlist_node *wptr;
1236 :
1237 20 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1238 :
1239 : /*
1240 : * Get a worker entry from the freelist. We checked above, so there
1241 : * really should be a free slot.
1242 : */
1243 20 : wptr = dlist_pop_head_node(&AutoVacuumShmem->av_freeWorkers);
1244 :
1245 20 : worker = dlist_container(WorkerInfoData, wi_links, wptr);
1246 20 : worker->wi_dboid = avdb->adw_datid;
1247 20 : worker->wi_proc = NULL;
1248 20 : worker->wi_launchtime = GetCurrentTimestamp();
1249 :
1250 20 : AutoVacuumShmem->av_startingWorker = worker;
1251 :
1252 20 : LWLockRelease(AutovacuumLock);
1253 :
1254 20 : SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER);
1255 :
1256 20 : retval = avdb->adw_datid;
1257 : }
1258 62 : else if (skipit)
1259 : {
1260 : /*
1261 : * If we skipped all databases on the list, rebuild it, because it
1262 : * probably contains a dropped database.
1263 : */
1264 0 : rebuild_database_list(InvalidOid);
1265 : }
1266 :
1267 82 : MemoryContextSwitchTo(oldcxt);
1268 82 : MemoryContextDelete(tmpcxt);
1269 :
1270 82 : return retval;
1271 : }
1272 :
1273 : /*
1274 : * launch_worker
1275 : *
1276 : * Wrapper for starting a worker from the launcher. Besides actually starting
1277 : * it, update the database list to reflect the next time that another one will
1278 : * need to be started on the selected database. The actual database choice is
1279 : * left to do_start_worker.
1280 : *
1281 : * This routine is also expected to insert an entry into the database list if
1282 : * the selected database was previously absent from the list.
1283 : */
1284 : static void
1285 82 : launch_worker(TimestampTz now)
1286 : {
1287 : Oid dbid;
1288 : dlist_iter iter;
1289 :
1290 82 : dbid = do_start_worker();
1291 82 : if (OidIsValid(dbid))
1292 : {
1293 20 : bool found = false;
1294 :
1295 : /*
1296 : * Walk the database list and update the corresponding entry. If the
1297 : * database is not on the list, we'll recreate the list.
1298 : */
1299 52 : dlist_foreach(iter, &DatabaseList)
1300 : {
1301 44 : avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur);
1302 :
1303 44 : if (avdb->adl_datid == dbid)
1304 : {
1305 12 : found = true;
1306 :
1307 : /*
1308 : * add autovacuum_naptime seconds to the current time, and use
1309 : * that as the new "next_worker" field for this database.
1310 : */
1311 12 : avdb->adl_next_worker =
1312 12 : TimestampTzPlusMilliseconds(now, autovacuum_naptime * 1000);
1313 :
1314 12 : dlist_move_head(&DatabaseList, iter.cur);
1315 12 : break;
1316 : }
1317 : }
1318 :
1319 : /*
1320 : * If the database was not present in the database list, we rebuild
1321 : * the list. It's possible that the database does not get into the
1322 : * list anyway, for example if it's a database that doesn't have a
1323 : * pgstat entry, but this is not a problem because we don't want to
1324 : * schedule workers regularly into those in any case.
1325 : */
1326 20 : if (!found)
1327 8 : rebuild_database_list(dbid);
1328 : }
1329 82 : }
1330 :
1331 : /*
1332 : * Called from postmaster to signal a failure to fork a process to become
1333 : * worker. The postmaster should kill(SIGUSR2) the launcher shortly
1334 : * after calling this function.
1335 : */
1336 : void
1337 0 : AutoVacWorkerFailed(void)
1338 : {
1339 0 : AutoVacuumShmem->av_signal[AutoVacForkFailed] = true;
1340 0 : }
1341 :
1342 : /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */
1343 : static void
1344 40 : avl_sigusr2_handler(SIGNAL_ARGS)
1345 : {
1346 40 : got_SIGUSR2 = true;
1347 40 : SetLatch(MyLatch);
1348 40 : }
1349 :
1350 :
1351 : /********************************************************************
1352 : * AUTOVACUUM WORKER CODE
1353 : ********************************************************************/
1354 :
1355 : /*
1356 : * Main entry point for autovacuum worker processes.
1357 : */
1358 : void
1359 22 : AutoVacWorkerMain(char *startup_data, size_t startup_data_len)
1360 : {
1361 : sigjmp_buf local_sigjmp_buf;
1362 : Oid dbid;
1363 :
1364 : Assert(startup_data_len == 0);
1365 :
1366 : /* Release postmaster's working memory context */
1367 22 : if (PostmasterContext)
1368 : {
1369 22 : MemoryContextDelete(PostmasterContext);
1370 22 : PostmasterContext = NULL;
1371 : }
1372 :
1373 22 : MyBackendType = B_AUTOVAC_WORKER;
1374 22 : init_ps_display(NULL);
1375 :
1376 22 : SetProcessingMode(InitProcessing);
1377 :
1378 : /*
1379 : * Set up signal handlers. We operate on databases much like a regular
1380 : * backend, so we use the same signal handling. See equivalent code in
1381 : * tcop/postgres.c.
1382 : */
1383 22 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
1384 :
1385 : /*
1386 : * SIGINT is used to signal canceling the current table's vacuum; SIGTERM
1387 : * means abort and exit cleanly, and SIGQUIT means abandon ship.
1388 : */
1389 22 : pqsignal(SIGINT, StatementCancelHandler);
1390 22 : pqsignal(SIGTERM, die);
1391 : /* SIGQUIT handler was already set up by InitPostmasterChild */
1392 :
1393 22 : InitializeTimeouts(); /* establishes SIGALRM handler */
1394 :
1395 22 : pqsignal(SIGPIPE, SIG_IGN);
1396 22 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
1397 22 : pqsignal(SIGUSR2, SIG_IGN);
1398 22 : pqsignal(SIGFPE, FloatExceptionHandler);
1399 22 : pqsignal(SIGCHLD, SIG_DFL);
1400 :
1401 : /*
1402 : * Create a per-backend PGPROC struct in shared memory. We must do this
1403 : * before we can use LWLocks or access any shared memory.
1404 : */
1405 22 : InitProcess();
1406 :
1407 : /* Early initialization */
1408 22 : BaseInit();
1409 :
1410 : /*
1411 : * If an exception is encountered, processing resumes here.
1412 : *
1413 : * Unlike most auxiliary processes, we don't attempt to continue
1414 : * processing after an error; we just clean up and exit. The autovac
1415 : * launcher is responsible for spawning another worker later.
1416 : *
1417 : * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
1418 : * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
1419 : * signals other than SIGQUIT will be blocked until we exit. It might
1420 : * seem that this policy makes the HOLD_INTERRUPTS() call redundant, but
1421 : * it is not since InterruptPending might be set already.
1422 : */
1423 22 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
1424 : {
1425 : /* since not using PG_TRY, must reset error stack by hand */
1426 0 : error_context_stack = NULL;
1427 :
1428 : /* Prevents interrupts while cleaning up */
1429 0 : HOLD_INTERRUPTS();
1430 :
1431 : /* Report the error to the server log */
1432 0 : EmitErrorReport();
1433 :
1434 : /*
1435 : * We can now go away. Note that because we called InitProcess, a
1436 : * callback was registered to do ProcKill, which will clean up
1437 : * necessary state.
1438 : */
1439 0 : proc_exit(0);
1440 : }
1441 :
1442 : /* We can now handle ereport(ERROR) */
1443 22 : PG_exception_stack = &local_sigjmp_buf;
1444 :
1445 22 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
1446 :
1447 : /*
1448 : * Set always-secure search path, so malicious users can't redirect user
1449 : * code (e.g. pg_index.indexprs). (That code runs in a
1450 : * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not
1451 : * take control of the entire autovacuum worker in any case.)
1452 : */
1453 22 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
1454 :
1455 : /*
1456 : * Force zero_damaged_pages OFF in the autovac process, even if it is set
1457 : * in postgresql.conf. We don't really want such a dangerous option being
1458 : * applied non-interactively.
1459 : */
1460 22 : SetConfigOption("zero_damaged_pages", "false", PGC_SUSET, PGC_S_OVERRIDE);
1461 :
1462 : /*
1463 : * Force settable timeouts off to avoid letting these settings prevent
1464 : * regular maintenance from being executed.
1465 : */
1466 22 : SetConfigOption("statement_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1467 22 : SetConfigOption("transaction_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1468 22 : SetConfigOption("lock_timeout", "0", PGC_SUSET, PGC_S_OVERRIDE);
1469 22 : SetConfigOption("idle_in_transaction_session_timeout", "0",
1470 : PGC_SUSET, PGC_S_OVERRIDE);
1471 :
1472 : /*
1473 : * Force default_transaction_isolation to READ COMMITTED. We don't want
1474 : * to pay the overhead of serializable mode, nor add any risk of causing
1475 : * deadlocks or delaying other transactions.
1476 : */
1477 22 : SetConfigOption("default_transaction_isolation", "read committed",
1478 : PGC_SUSET, PGC_S_OVERRIDE);
1479 :
1480 : /*
1481 : * Force synchronous replication off to allow regular maintenance even if
1482 : * we are waiting for standbys to connect. This is important to ensure we
1483 : * aren't blocked from performing anti-wraparound tasks.
1484 : */
1485 22 : if (synchronous_commit > SYNCHRONOUS_COMMIT_LOCAL_FLUSH)
1486 22 : SetConfigOption("synchronous_commit", "local",
1487 : PGC_SUSET, PGC_S_OVERRIDE);
1488 :
1489 : /*
1490 : * Even when system is configured to use a different fetch consistency,
1491 : * for autovac we always want fresh stats.
1492 : */
1493 22 : SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET, PGC_S_OVERRIDE);
1494 :
1495 : /*
1496 : * Get the info about the database we're going to work on.
1497 : */
1498 22 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1499 :
1500 : /*
1501 : * beware of startingWorker being INVALID; this should normally not
1502 : * happen, but if a worker fails after forking and before this, the
1503 : * launcher might have decided to remove it from the queue and start
1504 : * again.
1505 : */
1506 22 : if (AutoVacuumShmem->av_startingWorker != NULL)
1507 : {
1508 22 : MyWorkerInfo = AutoVacuumShmem->av_startingWorker;
1509 22 : dbid = MyWorkerInfo->wi_dboid;
1510 22 : MyWorkerInfo->wi_proc = MyProc;
1511 :
1512 : /* insert into the running list */
1513 22 : dlist_push_head(&AutoVacuumShmem->av_runningWorkers,
1514 22 : &MyWorkerInfo->wi_links);
1515 :
1516 : /*
1517 : * remove from the "starting" pointer, so that the launcher can start
1518 : * a new worker if required
1519 : */
1520 22 : AutoVacuumShmem->av_startingWorker = NULL;
1521 22 : LWLockRelease(AutovacuumLock);
1522 :
1523 22 : on_shmem_exit(FreeWorkerInfo, 0);
1524 :
1525 : /* wake up the launcher */
1526 22 : if (AutoVacuumShmem->av_launcherpid != 0)
1527 22 : kill(AutoVacuumShmem->av_launcherpid, SIGUSR2);
1528 : }
1529 : else
1530 : {
1531 : /* no worker entry for me, go away */
1532 0 : elog(WARNING, "autovacuum worker started without a worker entry");
1533 0 : dbid = InvalidOid;
1534 0 : LWLockRelease(AutovacuumLock);
1535 : }
1536 :
1537 22 : if (OidIsValid(dbid))
1538 : {
1539 : char dbname[NAMEDATALEN];
1540 :
1541 : /*
1542 : * Report autovac startup to the cumulative stats system. We
1543 : * deliberately do this before InitPostgres, so that the
1544 : * last_autovac_time will get updated even if the connection attempt
1545 : * fails. This is to prevent autovac from getting "stuck" repeatedly
1546 : * selecting an unopenable database, rather than making any progress
1547 : * on stuff it can connect to.
1548 : */
1549 22 : pgstat_report_autovac(dbid);
1550 :
1551 : /*
1552 : * Connect to the selected database, specifying no particular user
1553 : *
1554 : * Note: if we have selected a just-deleted database (due to using
1555 : * stale stats info), we'll fail and exit here.
1556 : */
1557 22 : InitPostgres(NULL, dbid, NULL, InvalidOid, 0, dbname);
1558 22 : SetProcessingMode(NormalProcessing);
1559 22 : set_ps_display(dbname);
1560 22 : ereport(DEBUG1,
1561 : (errmsg_internal("autovacuum: processing database \"%s\"", dbname)));
1562 :
1563 22 : if (PostAuthDelay)
1564 0 : pg_usleep(PostAuthDelay * 1000000L);
1565 :
1566 : /* And do an appropriate amount of work */
1567 22 : recentXid = ReadNextTransactionId();
1568 22 : recentMulti = ReadNextMultiXactId();
1569 22 : do_autovacuum();
1570 : }
1571 :
1572 : /*
1573 : * The launcher will be notified of my death in ProcKill, *if* we managed
1574 : * to get a worker slot at all
1575 : */
1576 :
1577 : /* All done, go away */
1578 22 : proc_exit(0);
1579 : }
1580 :
1581 : /*
1582 : * Return a WorkerInfo to the free list
1583 : */
1584 : static void
1585 22 : FreeWorkerInfo(int code, Datum arg)
1586 : {
1587 22 : if (MyWorkerInfo != NULL)
1588 : {
1589 22 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
1590 :
1591 : /*
1592 : * Wake the launcher up so that he can launch a new worker immediately
1593 : * if required. We only save the launcher's PID in local memory here;
1594 : * the actual signal will be sent when the PGPROC is recycled. Note
1595 : * that we always do this, so that the launcher can rebalance the cost
1596 : * limit setting of the remaining workers.
1597 : *
1598 : * We somewhat ignore the risk that the launcher changes its PID
1599 : * between us reading it and the actual kill; we expect ProcKill to be
1600 : * called shortly after us, and we assume that PIDs are not reused too
1601 : * quickly after a process exits.
1602 : */
1603 22 : AutovacuumLauncherPid = AutoVacuumShmem->av_launcherpid;
1604 :
1605 22 : dlist_delete(&MyWorkerInfo->wi_links);
1606 22 : MyWorkerInfo->wi_dboid = InvalidOid;
1607 22 : MyWorkerInfo->wi_tableoid = InvalidOid;
1608 22 : MyWorkerInfo->wi_sharedrel = false;
1609 22 : MyWorkerInfo->wi_proc = NULL;
1610 22 : MyWorkerInfo->wi_launchtime = 0;
1611 22 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
1612 22 : dlist_push_head(&AutoVacuumShmem->av_freeWorkers,
1613 22 : &MyWorkerInfo->wi_links);
1614 : /* not mine anymore */
1615 22 : MyWorkerInfo = NULL;
1616 :
1617 : /*
1618 : * now that we're inactive, cause a rebalancing of the surviving
1619 : * workers
1620 : */
1621 22 : AutoVacuumShmem->av_signal[AutoVacRebalance] = true;
1622 22 : LWLockRelease(AutovacuumLock);
1623 : }
1624 22 : }
1625 :
1626 : /*
1627 : * Update vacuum cost-based delay-related parameters for autovacuum workers and
1628 : * backends executing VACUUM or ANALYZE using the value of relevant GUCs and
1629 : * global state. This must be called during setup for vacuum and after every
1630 : * config reload to ensure up-to-date values.
1631 : */
1632 : void
1633 10740 : VacuumUpdateCosts(void)
1634 : {
1635 10740 : if (MyWorkerInfo)
1636 : {
1637 416 : if (av_storage_param_cost_delay >= 0)
1638 0 : vacuum_cost_delay = av_storage_param_cost_delay;
1639 416 : else if (autovacuum_vac_cost_delay >= 0)
1640 416 : vacuum_cost_delay = autovacuum_vac_cost_delay;
1641 : else
1642 : /* fall back to VacuumCostDelay */
1643 0 : vacuum_cost_delay = VacuumCostDelay;
1644 :
1645 416 : AutoVacuumUpdateCostLimit();
1646 : }
1647 : else
1648 : {
1649 : /* Must be explicit VACUUM or ANALYZE */
1650 10324 : vacuum_cost_delay = VacuumCostDelay;
1651 10324 : vacuum_cost_limit = VacuumCostLimit;
1652 : }
1653 :
1654 : /*
1655 : * If configuration changes are allowed to impact VacuumCostActive, make
1656 : * sure it is updated.
1657 : */
1658 10740 : if (VacuumFailsafeActive)
1659 : Assert(!VacuumCostActive);
1660 10740 : else if (vacuum_cost_delay > 0)
1661 416 : VacuumCostActive = true;
1662 : else
1663 : {
1664 10324 : VacuumCostActive = false;
1665 10324 : VacuumCostBalance = 0;
1666 : }
1667 :
1668 : /*
1669 : * Since the cost logging requires a lock, avoid rendering the log message
1670 : * in case we are using a message level where the log wouldn't be emitted.
1671 : */
1672 10740 : if (MyWorkerInfo && message_level_is_interesting(DEBUG2))
1673 : {
1674 : Oid dboid,
1675 : tableoid;
1676 :
1677 : Assert(!LWLockHeldByMe(AutovacuumLock));
1678 :
1679 0 : LWLockAcquire(AutovacuumLock, LW_SHARED);
1680 0 : dboid = MyWorkerInfo->wi_dboid;
1681 0 : tableoid = MyWorkerInfo->wi_tableoid;
1682 0 : LWLockRelease(AutovacuumLock);
1683 :
1684 0 : elog(DEBUG2,
1685 : "Autovacuum VacuumUpdateCosts(db=%u, rel=%u, dobalance=%s, cost_limit=%d, cost_delay=%g active=%s failsafe=%s)",
1686 : dboid, tableoid, pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance) ? "no" : "yes",
1687 : vacuum_cost_limit, vacuum_cost_delay,
1688 : vacuum_cost_delay > 0 ? "yes" : "no",
1689 : VacuumFailsafeActive ? "yes" : "no");
1690 : }
1691 10740 : }
1692 :
1693 : /*
1694 : * Update vacuum_cost_limit with the correct value for an autovacuum worker,
1695 : * given the value of other relevant cost limit parameters and the number of
1696 : * workers across which the limit must be balanced. Autovacuum workers must
1697 : * call this regularly in case av_nworkersForBalance has been updated by
1698 : * another worker or by the autovacuum launcher. They must also call it after a
1699 : * config reload.
1700 : */
1701 : void
1702 760 : AutoVacuumUpdateCostLimit(void)
1703 : {
1704 760 : if (!MyWorkerInfo)
1705 0 : return;
1706 :
1707 : /*
1708 : * note: in cost_limit, zero also means use value from elsewhere, because
1709 : * zero is not a valid value.
1710 : */
1711 :
1712 760 : if (av_storage_param_cost_limit > 0)
1713 0 : vacuum_cost_limit = av_storage_param_cost_limit;
1714 : else
1715 : {
1716 : int nworkers_for_balance;
1717 :
1718 760 : if (autovacuum_vac_cost_limit > 0)
1719 0 : vacuum_cost_limit = autovacuum_vac_cost_limit;
1720 : else
1721 760 : vacuum_cost_limit = VacuumCostLimit;
1722 :
1723 : /* Only balance limit if no cost-related storage parameters specified */
1724 760 : if (pg_atomic_unlocked_test_flag(&MyWorkerInfo->wi_dobalance))
1725 0 : return;
1726 :
1727 : Assert(vacuum_cost_limit > 0);
1728 :
1729 760 : nworkers_for_balance = pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1730 :
1731 : /* There is at least 1 autovac worker (this worker) */
1732 760 : if (nworkers_for_balance <= 0)
1733 0 : elog(ERROR, "nworkers_for_balance must be > 0");
1734 :
1735 760 : vacuum_cost_limit = Max(vacuum_cost_limit / nworkers_for_balance, 1);
1736 : }
1737 : }
1738 :
1739 : /*
1740 : * autovac_recalculate_workers_for_balance
1741 : * Recalculate the number of workers to consider, given cost-related
1742 : * storage parameters and the current number of active workers.
1743 : *
1744 : * Caller must hold the AutovacuumLock in at least shared mode to access
1745 : * worker->wi_proc.
1746 : */
1747 : static void
1748 228 : autovac_recalculate_workers_for_balance(void)
1749 : {
1750 : dlist_iter iter;
1751 : int orig_nworkers_for_balance;
1752 228 : int nworkers_for_balance = 0;
1753 :
1754 : Assert(LWLockHeldByMe(AutovacuumLock));
1755 :
1756 228 : orig_nworkers_for_balance =
1757 228 : pg_atomic_read_u32(&AutoVacuumShmem->av_nworkersForBalance);
1758 :
1759 436 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
1760 : {
1761 208 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
1762 :
1763 416 : if (worker->wi_proc == NULL ||
1764 208 : pg_atomic_unlocked_test_flag(&worker->wi_dobalance))
1765 0 : continue;
1766 :
1767 208 : nworkers_for_balance++;
1768 : }
1769 :
1770 228 : if (nworkers_for_balance != orig_nworkers_for_balance)
1771 32 : pg_atomic_write_u32(&AutoVacuumShmem->av_nworkersForBalance,
1772 : nworkers_for_balance);
1773 228 : }
1774 :
1775 : /*
1776 : * get_database_list
1777 : * Return a list of all databases found in pg_database.
1778 : *
1779 : * The list and associated data is allocated in the caller's memory context,
1780 : * which is in charge of ensuring that it's properly cleaned up afterwards.
1781 : *
1782 : * Note: this is the only function in which the autovacuum launcher uses a
1783 : * transaction. Although we aren't attached to any particular database and
1784 : * therefore can't access most catalogs, we do have enough infrastructure
1785 : * to do a seqscan on pg_database.
1786 : */
1787 : static List *
1788 798 : get_database_list(void)
1789 : {
1790 798 : List *dblist = NIL;
1791 : Relation rel;
1792 : TableScanDesc scan;
1793 : HeapTuple tup;
1794 : MemoryContext resultcxt;
1795 :
1796 : /* This is the context that we will allocate our output data in */
1797 798 : resultcxt = CurrentMemoryContext;
1798 :
1799 : /*
1800 : * Start a transaction so we can access pg_database, and get a snapshot.
1801 : * We don't have a use for the snapshot itself, but we're interested in
1802 : * the secondary effect that it sets RecentGlobalXmin. (This is critical
1803 : * for anything that reads heap pages, because HOT may decide to prune
1804 : * them even if the process doesn't attempt to modify any tuples.)
1805 : *
1806 : * FIXME: This comment is inaccurate / the code buggy. A snapshot that is
1807 : * not pushed/active does not reliably prevent HOT pruning (->xmin could
1808 : * e.g. be cleared when cache invalidations are processed).
1809 : */
1810 798 : StartTransactionCommand();
1811 798 : (void) GetTransactionSnapshot();
1812 :
1813 798 : rel = table_open(DatabaseRelationId, AccessShareLock);
1814 798 : scan = table_beginscan_catalog(rel, 0, NULL);
1815 :
1816 3552 : while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
1817 : {
1818 2754 : Form_pg_database pgdatabase = (Form_pg_database) GETSTRUCT(tup);
1819 : avw_dbase *avdb;
1820 : MemoryContext oldcxt;
1821 :
1822 : /*
1823 : * If database has partially been dropped, we can't, nor need to,
1824 : * vacuum it.
1825 : */
1826 2754 : if (database_is_invalid_form(pgdatabase))
1827 : {
1828 2 : elog(DEBUG2,
1829 : "autovacuum: skipping invalid database \"%s\"",
1830 : NameStr(pgdatabase->datname));
1831 2 : continue;
1832 : }
1833 :
1834 : /*
1835 : * Allocate our results in the caller's context, not the
1836 : * transaction's. We do this inside the loop, and restore the original
1837 : * context at the end, so that leaky things like heap_getnext() are
1838 : * not called in a potentially long-lived context.
1839 : */
1840 2752 : oldcxt = MemoryContextSwitchTo(resultcxt);
1841 :
1842 2752 : avdb = (avw_dbase *) palloc(sizeof(avw_dbase));
1843 :
1844 2752 : avdb->adw_datid = pgdatabase->oid;
1845 2752 : avdb->adw_name = pstrdup(NameStr(pgdatabase->datname));
1846 2752 : avdb->adw_frozenxid = pgdatabase->datfrozenxid;
1847 2752 : avdb->adw_minmulti = pgdatabase->datminmxid;
1848 : /* this gets set later: */
1849 2752 : avdb->adw_entry = NULL;
1850 :
1851 2752 : dblist = lappend(dblist, avdb);
1852 2752 : MemoryContextSwitchTo(oldcxt);
1853 : }
1854 :
1855 798 : table_endscan(scan);
1856 798 : table_close(rel, AccessShareLock);
1857 :
1858 798 : CommitTransactionCommand();
1859 :
1860 : /* Be sure to restore caller's memory context */
1861 798 : MemoryContextSwitchTo(resultcxt);
1862 :
1863 798 : return dblist;
1864 : }
1865 :
1866 : /*
1867 : * Process a database table-by-table
1868 : *
1869 : * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in
1870 : * order not to ignore shutdown commands for too long.
1871 : */
1872 : static void
1873 22 : do_autovacuum(void)
1874 : {
1875 : Relation classRel;
1876 : HeapTuple tuple;
1877 : TableScanDesc relScan;
1878 : Form_pg_database dbForm;
1879 22 : List *table_oids = NIL;
1880 22 : List *orphan_oids = NIL;
1881 : HASHCTL ctl;
1882 : HTAB *table_toast_map;
1883 : ListCell *volatile cell;
1884 : BufferAccessStrategy bstrategy;
1885 : ScanKeyData key;
1886 : TupleDesc pg_class_desc;
1887 : int effective_multixact_freeze_max_age;
1888 22 : bool did_vacuum = false;
1889 22 : bool found_concurrent_worker = false;
1890 : int i;
1891 :
1892 : /*
1893 : * StartTransactionCommand and CommitTransactionCommand will automatically
1894 : * switch to other contexts. We need this one to keep the list of
1895 : * relations to vacuum/analyze across transactions.
1896 : */
1897 22 : AutovacMemCxt = AllocSetContextCreate(TopMemoryContext,
1898 : "Autovacuum worker",
1899 : ALLOCSET_DEFAULT_SIZES);
1900 22 : MemoryContextSwitchTo(AutovacMemCxt);
1901 :
1902 : /* Start a transaction so our commands have one to play into. */
1903 22 : StartTransactionCommand();
1904 :
1905 : /*
1906 : * Compute the multixact age for which freezing is urgent. This is
1907 : * normally autovacuum_multixact_freeze_max_age, but may be less if we are
1908 : * short of multixact member space.
1909 : */
1910 22 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1911 :
1912 : /*
1913 : * Find the pg_database entry and select the default freeze ages. We use
1914 : * zero in template and nonconnectable databases, else the system-wide
1915 : * default.
1916 : */
1917 22 : tuple = SearchSysCache1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId));
1918 22 : if (!HeapTupleIsValid(tuple))
1919 0 : elog(ERROR, "cache lookup failed for database %u", MyDatabaseId);
1920 22 : dbForm = (Form_pg_database) GETSTRUCT(tuple);
1921 :
1922 22 : if (dbForm->datistemplate || !dbForm->datallowconn)
1923 : {
1924 4 : default_freeze_min_age = 0;
1925 4 : default_freeze_table_age = 0;
1926 4 : default_multixact_freeze_min_age = 0;
1927 4 : default_multixact_freeze_table_age = 0;
1928 : }
1929 : else
1930 : {
1931 18 : default_freeze_min_age = vacuum_freeze_min_age;
1932 18 : default_freeze_table_age = vacuum_freeze_table_age;
1933 18 : default_multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1934 18 : default_multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1935 : }
1936 :
1937 22 : ReleaseSysCache(tuple);
1938 :
1939 : /* StartTransactionCommand changed elsewhere */
1940 22 : MemoryContextSwitchTo(AutovacMemCxt);
1941 :
1942 22 : classRel = table_open(RelationRelationId, AccessShareLock);
1943 :
1944 : /* create a copy so we can use it after closing pg_class */
1945 22 : pg_class_desc = CreateTupleDescCopy(RelationGetDescr(classRel));
1946 :
1947 : /* create hash table for toast <-> main relid mapping */
1948 22 : ctl.keysize = sizeof(Oid);
1949 22 : ctl.entrysize = sizeof(av_relation);
1950 :
1951 22 : table_toast_map = hash_create("TOAST to main relid map",
1952 : 100,
1953 : &ctl,
1954 : HASH_ELEM | HASH_BLOBS);
1955 :
1956 : /*
1957 : * Scan pg_class to determine which tables to vacuum.
1958 : *
1959 : * We do this in two passes: on the first one we collect the list of plain
1960 : * relations and materialized views, and on the second one we collect
1961 : * TOAST tables. The reason for doing the second pass is that during it we
1962 : * want to use the main relation's pg_class.reloptions entry if the TOAST
1963 : * table does not have any, and we cannot obtain it unless we know
1964 : * beforehand what's the main table OID.
1965 : *
1966 : * We need to check TOAST tables separately because in cases with short,
1967 : * wide tables there might be proportionally much more activity in the
1968 : * TOAST table than in its parent.
1969 : */
1970 22 : relScan = table_beginscan_catalog(classRel, 0, NULL);
1971 :
1972 : /*
1973 : * On the first pass, we collect main tables to vacuum, and also the main
1974 : * table relid to TOAST relid mapping.
1975 : */
1976 11532 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
1977 : {
1978 11510 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
1979 : PgStat_StatTabEntry *tabentry;
1980 : AutoVacOpts *relopts;
1981 : Oid relid;
1982 : bool dovacuum;
1983 : bool doanalyze;
1984 : bool wraparound;
1985 :
1986 11510 : if (classForm->relkind != RELKIND_RELATION &&
1987 9128 : classForm->relkind != RELKIND_MATVIEW)
1988 9128 : continue;
1989 :
1990 2396 : relid = classForm->oid;
1991 :
1992 : /*
1993 : * Check if it is a temp table (presumably, of some other backend's).
1994 : * We cannot safely process other backends' temp tables.
1995 : */
1996 2396 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
1997 : {
1998 : /*
1999 : * We just ignore it if the owning backend is still active and
2000 : * using the temporary schema. Also, for safety, ignore it if the
2001 : * namespace doesn't exist or isn't a temp namespace after all.
2002 : */
2003 14 : if (checkTempNamespaceStatus(classForm->relnamespace) == TEMP_NAMESPACE_IDLE)
2004 : {
2005 : /*
2006 : * The table seems to be orphaned -- although it might be that
2007 : * the owning backend has already deleted it and exited; our
2008 : * pg_class scan snapshot is not necessarily up-to-date
2009 : * anymore, so we could be looking at a committed-dead entry.
2010 : * Remember it so we can try to delete it later.
2011 : */
2012 0 : orphan_oids = lappend_oid(orphan_oids, relid);
2013 : }
2014 14 : continue;
2015 : }
2016 :
2017 : /* Fetch reloptions and the pgstat entry for this table */
2018 2382 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2019 2382 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2020 : relid);
2021 :
2022 : /* Check if it needs vacuum or analyze */
2023 2382 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2024 : effective_multixact_freeze_max_age,
2025 : &dovacuum, &doanalyze, &wraparound);
2026 :
2027 : /* Relations that need work are added to table_oids */
2028 2382 : if (dovacuum || doanalyze)
2029 234 : table_oids = lappend_oid(table_oids, relid);
2030 :
2031 : /*
2032 : * Remember TOAST associations for the second pass. Note: we must do
2033 : * this whether or not the table is going to be vacuumed, because we
2034 : * don't automatically vacuum toast tables along the parent table.
2035 : */
2036 2382 : if (OidIsValid(classForm->reltoastrelid))
2037 : {
2038 : av_relation *hentry;
2039 : bool found;
2040 :
2041 2488 : hentry = hash_search(table_toast_map,
2042 1244 : &classForm->reltoastrelid,
2043 : HASH_ENTER, &found);
2044 :
2045 1244 : if (!found)
2046 : {
2047 : /* hash_search already filled in the key */
2048 1244 : hentry->ar_relid = relid;
2049 1244 : hentry->ar_hasrelopts = false;
2050 1244 : if (relopts != NULL)
2051 : {
2052 18 : hentry->ar_hasrelopts = true;
2053 18 : memcpy(&hentry->ar_reloptions, relopts,
2054 : sizeof(AutoVacOpts));
2055 : }
2056 : }
2057 : }
2058 : }
2059 :
2060 22 : table_endscan(relScan);
2061 :
2062 : /* second pass: check TOAST tables */
2063 22 : ScanKeyInit(&key,
2064 : Anum_pg_class_relkind,
2065 : BTEqualStrategyNumber, F_CHAREQ,
2066 : CharGetDatum(RELKIND_TOASTVALUE));
2067 :
2068 22 : relScan = table_beginscan_catalog(classRel, 1, &key);
2069 1270 : while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL)
2070 : {
2071 1248 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
2072 : PgStat_StatTabEntry *tabentry;
2073 : Oid relid;
2074 1248 : AutoVacOpts *relopts = NULL;
2075 : bool dovacuum;
2076 : bool doanalyze;
2077 : bool wraparound;
2078 :
2079 : /*
2080 : * We cannot safely process other backends' temp tables, so skip 'em.
2081 : */
2082 1248 : if (classForm->relpersistence == RELPERSISTENCE_TEMP)
2083 4 : continue;
2084 :
2085 1244 : relid = classForm->oid;
2086 :
2087 : /*
2088 : * fetch reloptions -- if this toast table does not have them, try the
2089 : * main rel
2090 : */
2091 1244 : relopts = extract_autovac_opts(tuple, pg_class_desc);
2092 1244 : if (relopts == NULL)
2093 : {
2094 : av_relation *hentry;
2095 : bool found;
2096 :
2097 1244 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2098 1244 : if (found && hentry->ar_hasrelopts)
2099 18 : relopts = &hentry->ar_reloptions;
2100 : }
2101 :
2102 : /* Fetch the pgstat entry for this table */
2103 1244 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2104 : relid);
2105 :
2106 1244 : relation_needs_vacanalyze(relid, relopts, classForm, tabentry,
2107 : effective_multixact_freeze_max_age,
2108 : &dovacuum, &doanalyze, &wraparound);
2109 :
2110 : /* ignore analyze for toast tables */
2111 1244 : if (dovacuum)
2112 2 : table_oids = lappend_oid(table_oids, relid);
2113 : }
2114 :
2115 22 : table_endscan(relScan);
2116 22 : table_close(classRel, AccessShareLock);
2117 :
2118 : /*
2119 : * Recheck orphan temporary tables, and if they still seem orphaned, drop
2120 : * them. We'll eat a transaction per dropped table, which might seem
2121 : * excessive, but we should only need to do anything as a result of a
2122 : * previous backend crash, so this should not happen often enough to
2123 : * justify "optimizing". Using separate transactions ensures that we
2124 : * don't bloat the lock table if there are many temp tables to be dropped,
2125 : * and it ensures that we don't lose work if a deletion attempt fails.
2126 : */
2127 22 : foreach(cell, orphan_oids)
2128 : {
2129 0 : Oid relid = lfirst_oid(cell);
2130 : Form_pg_class classForm;
2131 : ObjectAddress object;
2132 :
2133 : /*
2134 : * Check for user-requested abort.
2135 : */
2136 0 : CHECK_FOR_INTERRUPTS();
2137 :
2138 : /*
2139 : * Try to lock the table. If we can't get the lock immediately,
2140 : * somebody else is using (or dropping) the table, so it's not our
2141 : * concern anymore. Having the lock prevents race conditions below.
2142 : */
2143 0 : if (!ConditionalLockRelationOid(relid, AccessExclusiveLock))
2144 0 : continue;
2145 :
2146 : /*
2147 : * Re-fetch the pg_class tuple and re-check whether it still seems to
2148 : * be an orphaned temp table. If it's not there or no longer the same
2149 : * relation, ignore it.
2150 : */
2151 0 : tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2152 0 : if (!HeapTupleIsValid(tuple))
2153 : {
2154 : /* be sure to drop useless lock so we don't bloat lock table */
2155 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2156 0 : continue;
2157 : }
2158 0 : classForm = (Form_pg_class) GETSTRUCT(tuple);
2159 :
2160 : /*
2161 : * Make all the same tests made in the loop above. In event of OID
2162 : * counter wraparound, the pg_class entry we have now might be
2163 : * completely unrelated to the one we saw before.
2164 : */
2165 0 : if (!((classForm->relkind == RELKIND_RELATION ||
2166 0 : classForm->relkind == RELKIND_MATVIEW) &&
2167 0 : classForm->relpersistence == RELPERSISTENCE_TEMP))
2168 : {
2169 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2170 0 : continue;
2171 : }
2172 :
2173 0 : if (checkTempNamespaceStatus(classForm->relnamespace) != TEMP_NAMESPACE_IDLE)
2174 : {
2175 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2176 0 : continue;
2177 : }
2178 :
2179 : /*
2180 : * Try to lock the temp namespace, too. Even though we have lock on
2181 : * the table itself, there's a risk of deadlock against an incoming
2182 : * backend trying to clean out the temp namespace, in case this table
2183 : * has dependencies (such as sequences) that the backend's
2184 : * performDeletion call might visit in a different order. If we can
2185 : * get AccessShareLock on the namespace, that's sufficient to ensure
2186 : * we're not running concurrently with RemoveTempRelations. If we
2187 : * can't, back off and let RemoveTempRelations do its thing.
2188 : */
2189 0 : if (!ConditionalLockDatabaseObject(NamespaceRelationId,
2190 : classForm->relnamespace, 0,
2191 : AccessShareLock))
2192 : {
2193 0 : UnlockRelationOid(relid, AccessExclusiveLock);
2194 0 : continue;
2195 : }
2196 :
2197 : /* OK, let's delete it */
2198 0 : ereport(LOG,
2199 : (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"",
2200 : get_database_name(MyDatabaseId),
2201 : get_namespace_name(classForm->relnamespace),
2202 : NameStr(classForm->relname))));
2203 :
2204 0 : object.classId = RelationRelationId;
2205 0 : object.objectId = relid;
2206 0 : object.objectSubId = 0;
2207 0 : performDeletion(&object, DROP_CASCADE,
2208 : PERFORM_DELETION_INTERNAL |
2209 : PERFORM_DELETION_QUIETLY |
2210 : PERFORM_DELETION_SKIP_EXTENSIONS);
2211 :
2212 : /*
2213 : * To commit the deletion, end current transaction and start a new
2214 : * one. Note this also releases the locks we took.
2215 : */
2216 0 : CommitTransactionCommand();
2217 0 : StartTransactionCommand();
2218 :
2219 : /* StartTransactionCommand changed current memory context */
2220 0 : MemoryContextSwitchTo(AutovacMemCxt);
2221 : }
2222 :
2223 : /*
2224 : * Optionally, create a buffer access strategy object for VACUUM to use.
2225 : * We use the same BufferAccessStrategy object for all tables VACUUMed by
2226 : * this worker to prevent autovacuum from blowing out shared buffers.
2227 : *
2228 : * VacuumBufferUsageLimit being set to 0 results in
2229 : * GetAccessStrategyWithSize returning NULL, effectively meaning we can
2230 : * use up to all of shared buffers.
2231 : *
2232 : * If we later enter failsafe mode on any of the tables being vacuumed, we
2233 : * will cease use of the BufferAccessStrategy only for that table.
2234 : *
2235 : * XXX should we consider adding code to adjust the size of this if
2236 : * VacuumBufferUsageLimit changes?
2237 : */
2238 22 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, VacuumBufferUsageLimit);
2239 :
2240 : /*
2241 : * create a memory context to act as fake PortalContext, so that the
2242 : * contexts created in the vacuum code are cleaned up for each table.
2243 : */
2244 22 : PortalContext = AllocSetContextCreate(AutovacMemCxt,
2245 : "Autovacuum Portal",
2246 : ALLOCSET_DEFAULT_SIZES);
2247 :
2248 : /*
2249 : * Perform operations on collected tables.
2250 : */
2251 258 : foreach(cell, table_oids)
2252 : {
2253 236 : Oid relid = lfirst_oid(cell);
2254 : HeapTuple classTup;
2255 : autovac_table *tab;
2256 : bool isshared;
2257 : bool skipit;
2258 : dlist_iter iter;
2259 :
2260 236 : CHECK_FOR_INTERRUPTS();
2261 :
2262 : /*
2263 : * Check for config changes before processing each collected table.
2264 : */
2265 236 : if (ConfigReloadPending)
2266 : {
2267 0 : ConfigReloadPending = false;
2268 0 : ProcessConfigFile(PGC_SIGHUP);
2269 :
2270 : /*
2271 : * You might be tempted to bail out if we see autovacuum is now
2272 : * disabled. Must resist that temptation -- this might be a
2273 : * for-wraparound emergency worker, in which case that would be
2274 : * entirely inappropriate.
2275 : */
2276 : }
2277 :
2278 : /*
2279 : * Find out whether the table is shared or not. (It's slightly
2280 : * annoying to fetch the syscache entry just for this, but in typical
2281 : * cases it adds little cost because table_recheck_autovac would
2282 : * refetch the entry anyway. We could buy that back by copying the
2283 : * tuple here and passing it to table_recheck_autovac, but that
2284 : * increases the odds of that function working with stale data.)
2285 : */
2286 236 : classTup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
2287 236 : if (!HeapTupleIsValid(classTup))
2288 28 : continue; /* somebody deleted the rel, forget it */
2289 236 : isshared = ((Form_pg_class) GETSTRUCT(classTup))->relisshared;
2290 236 : ReleaseSysCache(classTup);
2291 :
2292 : /*
2293 : * Hold schedule lock from here until we've claimed the table. We
2294 : * also need the AutovacuumLock to walk the worker array, but that one
2295 : * can just be a shared lock.
2296 : */
2297 236 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2298 236 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2299 :
2300 : /*
2301 : * Check whether the table is being vacuumed concurrently by another
2302 : * worker.
2303 : */
2304 236 : skipit = false;
2305 472 : dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers)
2306 : {
2307 236 : WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur);
2308 :
2309 : /* ignore myself */
2310 236 : if (worker == MyWorkerInfo)
2311 236 : continue;
2312 :
2313 : /* ignore workers in other databases (unless table is shared) */
2314 0 : if (!worker->wi_sharedrel && worker->wi_dboid != MyDatabaseId)
2315 0 : continue;
2316 :
2317 0 : if (worker->wi_tableoid == relid)
2318 : {
2319 0 : skipit = true;
2320 0 : found_concurrent_worker = true;
2321 0 : break;
2322 : }
2323 : }
2324 236 : LWLockRelease(AutovacuumLock);
2325 236 : if (skipit)
2326 : {
2327 0 : LWLockRelease(AutovacuumScheduleLock);
2328 0 : continue;
2329 : }
2330 :
2331 : /*
2332 : * Store the table's OID in shared memory before releasing the
2333 : * schedule lock, so that other workers don't try to vacuum it
2334 : * concurrently. (We claim it here so as not to hold
2335 : * AutovacuumScheduleLock while rechecking the stats.)
2336 : */
2337 236 : MyWorkerInfo->wi_tableoid = relid;
2338 236 : MyWorkerInfo->wi_sharedrel = isshared;
2339 236 : LWLockRelease(AutovacuumScheduleLock);
2340 :
2341 : /*
2342 : * Check whether pgstat data still says we need to vacuum this table.
2343 : * It could have changed if something else processed the table while
2344 : * we weren't looking. This doesn't entirely close the race condition,
2345 : * but it is very small.
2346 : */
2347 236 : MemoryContextSwitchTo(AutovacMemCxt);
2348 236 : tab = table_recheck_autovac(relid, table_toast_map, pg_class_desc,
2349 : effective_multixact_freeze_max_age);
2350 236 : if (tab == NULL)
2351 : {
2352 : /* someone else vacuumed the table, or it went away */
2353 28 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2354 28 : MyWorkerInfo->wi_tableoid = InvalidOid;
2355 28 : MyWorkerInfo->wi_sharedrel = false;
2356 28 : LWLockRelease(AutovacuumScheduleLock);
2357 28 : continue;
2358 : }
2359 :
2360 : /*
2361 : * Save the cost-related storage parameter values in global variables
2362 : * for reference when updating vacuum_cost_delay and vacuum_cost_limit
2363 : * during vacuuming this table.
2364 : */
2365 208 : av_storage_param_cost_delay = tab->at_storage_param_vac_cost_delay;
2366 208 : av_storage_param_cost_limit = tab->at_storage_param_vac_cost_limit;
2367 :
2368 : /*
2369 : * We only expect this worker to ever set the flag, so don't bother
2370 : * checking the return value. We shouldn't have to retry.
2371 : */
2372 208 : if (tab->at_dobalance)
2373 208 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2374 : else
2375 0 : pg_atomic_clear_flag(&MyWorkerInfo->wi_dobalance);
2376 :
2377 208 : LWLockAcquire(AutovacuumLock, LW_SHARED);
2378 208 : autovac_recalculate_workers_for_balance();
2379 208 : LWLockRelease(AutovacuumLock);
2380 :
2381 : /*
2382 : * We wait until this point to update cost delay and cost limit
2383 : * values, even though we reloaded the configuration file above, so
2384 : * that we can take into account the cost-related storage parameters.
2385 : */
2386 208 : VacuumUpdateCosts();
2387 :
2388 :
2389 : /* clean up memory before each iteration */
2390 208 : MemoryContextReset(PortalContext);
2391 :
2392 : /*
2393 : * Save the relation name for a possible error message, to avoid a
2394 : * catalog lookup in case of an error. If any of these return NULL,
2395 : * then the relation has been dropped since last we checked; skip it.
2396 : * Note: they must live in a long-lived memory context because we call
2397 : * vacuum and analyze in different transactions.
2398 : */
2399 :
2400 208 : tab->at_relname = get_rel_name(tab->at_relid);
2401 208 : tab->at_nspname = get_namespace_name(get_rel_namespace(tab->at_relid));
2402 208 : tab->at_datname = get_database_name(MyDatabaseId);
2403 208 : if (!tab->at_relname || !tab->at_nspname || !tab->at_datname)
2404 0 : goto deleted;
2405 :
2406 : /*
2407 : * We will abort vacuuming the current table if something errors out,
2408 : * and continue with the next one in schedule; in particular, this
2409 : * happens if we are interrupted with SIGINT.
2410 : */
2411 208 : PG_TRY();
2412 : {
2413 : /* Use PortalContext for any per-table allocations */
2414 208 : MemoryContextSwitchTo(PortalContext);
2415 :
2416 : /* have at it */
2417 208 : autovacuum_do_vac_analyze(tab, bstrategy);
2418 :
2419 : /*
2420 : * Clear a possible query-cancel signal, to avoid a late reaction
2421 : * to an automatically-sent signal because of vacuuming the
2422 : * current table (we're done with it, so it would make no sense to
2423 : * cancel at this point.)
2424 : */
2425 208 : QueryCancelPending = false;
2426 : }
2427 0 : PG_CATCH();
2428 : {
2429 : /*
2430 : * Abort the transaction, start a new one, and proceed with the
2431 : * next table in our list.
2432 : */
2433 0 : HOLD_INTERRUPTS();
2434 0 : if (tab->at_params.options & VACOPT_VACUUM)
2435 0 : errcontext("automatic vacuum of table \"%s.%s.%s\"",
2436 : tab->at_datname, tab->at_nspname, tab->at_relname);
2437 : else
2438 0 : errcontext("automatic analyze of table \"%s.%s.%s\"",
2439 : tab->at_datname, tab->at_nspname, tab->at_relname);
2440 0 : EmitErrorReport();
2441 :
2442 : /* this resets ProcGlobal->statusFlags[i] too */
2443 0 : AbortOutOfAnyTransaction();
2444 0 : FlushErrorState();
2445 0 : MemoryContextReset(PortalContext);
2446 :
2447 : /* restart our transaction for the following operations */
2448 0 : StartTransactionCommand();
2449 0 : RESUME_INTERRUPTS();
2450 : }
2451 208 : PG_END_TRY();
2452 :
2453 : /* Make sure we're back in AutovacMemCxt */
2454 208 : MemoryContextSwitchTo(AutovacMemCxt);
2455 :
2456 208 : did_vacuum = true;
2457 :
2458 : /* ProcGlobal->statusFlags[i] are reset at the next end of xact */
2459 :
2460 : /* be tidy */
2461 208 : deleted:
2462 208 : if (tab->at_datname != NULL)
2463 208 : pfree(tab->at_datname);
2464 208 : if (tab->at_nspname != NULL)
2465 208 : pfree(tab->at_nspname);
2466 208 : if (tab->at_relname != NULL)
2467 208 : pfree(tab->at_relname);
2468 208 : pfree(tab);
2469 :
2470 : /*
2471 : * Remove my info from shared memory. We set wi_dobalance on the
2472 : * assumption that we are more likely than not to vacuum a table with
2473 : * no cost-related storage parameters next, so we want to claim our
2474 : * share of I/O as soon as possible to avoid thrashing the global
2475 : * balance.
2476 : */
2477 208 : LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE);
2478 208 : MyWorkerInfo->wi_tableoid = InvalidOid;
2479 208 : MyWorkerInfo->wi_sharedrel = false;
2480 208 : LWLockRelease(AutovacuumScheduleLock);
2481 208 : pg_atomic_test_set_flag(&MyWorkerInfo->wi_dobalance);
2482 : }
2483 :
2484 : /*
2485 : * Perform additional work items, as requested by backends.
2486 : */
2487 22 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2488 5654 : for (i = 0; i < NUM_WORKITEMS; i++)
2489 : {
2490 5632 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
2491 :
2492 5632 : if (!workitem->avw_used)
2493 5626 : continue;
2494 6 : if (workitem->avw_active)
2495 0 : continue;
2496 6 : if (workitem->avw_database != MyDatabaseId)
2497 0 : continue;
2498 :
2499 : /* claim this one, and release lock while performing it */
2500 6 : workitem->avw_active = true;
2501 6 : LWLockRelease(AutovacuumLock);
2502 :
2503 6 : perform_work_item(workitem);
2504 :
2505 : /*
2506 : * Check for config changes before acquiring lock for further jobs.
2507 : */
2508 6 : CHECK_FOR_INTERRUPTS();
2509 6 : if (ConfigReloadPending)
2510 : {
2511 0 : ConfigReloadPending = false;
2512 0 : ProcessConfigFile(PGC_SIGHUP);
2513 0 : VacuumUpdateCosts();
2514 : }
2515 :
2516 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
2517 :
2518 : /* and mark it done */
2519 6 : workitem->avw_active = false;
2520 6 : workitem->avw_used = false;
2521 : }
2522 22 : LWLockRelease(AutovacuumLock);
2523 :
2524 : /*
2525 : * We leak table_toast_map here (among other things), but since we're
2526 : * going away soon, it's not a problem.
2527 : */
2528 :
2529 : /*
2530 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We
2531 : * only need to do this once, not after each table.
2532 : *
2533 : * Even if we didn't vacuum anything, it may still be important to do
2534 : * this, because one indirect effect of vac_update_datfrozenxid() is to
2535 : * update TransamVariables->xidVacLimit. That might need to be done even
2536 : * if we haven't vacuumed anything, because relations with older
2537 : * relfrozenxid values or other databases with older datfrozenxid values
2538 : * might have been dropped, allowing xidVacLimit to advance.
2539 : *
2540 : * However, it's also important not to do this blindly in all cases,
2541 : * because when autovacuum=off this will restart the autovacuum launcher.
2542 : * If we're not careful, an infinite loop can result, where workers find
2543 : * no work to do and restart the launcher, which starts another worker in
2544 : * the same database that finds no work to do. To prevent that, we skip
2545 : * this if (1) we found no work to do and (2) we skipped at least one
2546 : * table due to concurrent autovacuum activity. In that case, the other
2547 : * worker has already done it, or will do so when it finishes.
2548 : */
2549 22 : if (did_vacuum || !found_concurrent_worker)
2550 22 : vac_update_datfrozenxid();
2551 :
2552 : /* Finally close out the last transaction. */
2553 22 : CommitTransactionCommand();
2554 22 : }
2555 :
2556 : /*
2557 : * Execute a previously registered work item.
2558 : */
2559 : static void
2560 6 : perform_work_item(AutoVacuumWorkItem *workitem)
2561 : {
2562 6 : char *cur_datname = NULL;
2563 6 : char *cur_nspname = NULL;
2564 6 : char *cur_relname = NULL;
2565 :
2566 : /*
2567 : * Note we do not store table info in MyWorkerInfo, since this is not
2568 : * vacuuming proper.
2569 : */
2570 :
2571 : /*
2572 : * Save the relation name for a possible error message, to avoid a catalog
2573 : * lookup in case of an error. If any of these return NULL, then the
2574 : * relation has been dropped since last we checked; skip it.
2575 : */
2576 : Assert(CurrentMemoryContext == AutovacMemCxt);
2577 :
2578 6 : cur_relname = get_rel_name(workitem->avw_relation);
2579 6 : cur_nspname = get_namespace_name(get_rel_namespace(workitem->avw_relation));
2580 6 : cur_datname = get_database_name(MyDatabaseId);
2581 6 : if (!cur_relname || !cur_nspname || !cur_datname)
2582 0 : goto deleted2;
2583 :
2584 6 : autovac_report_workitem(workitem, cur_nspname, cur_relname);
2585 :
2586 : /* clean up memory before each work item */
2587 6 : MemoryContextReset(PortalContext);
2588 :
2589 : /*
2590 : * We will abort the current work item if something errors out, and
2591 : * continue with the next one; in particular, this happens if we are
2592 : * interrupted with SIGINT. Note that this means that the work item list
2593 : * can be lossy.
2594 : */
2595 6 : PG_TRY();
2596 : {
2597 : /* Use PortalContext for any per-work-item allocations */
2598 6 : MemoryContextSwitchTo(PortalContext);
2599 :
2600 : /*
2601 : * Have at it. Functions called here are responsible for any required
2602 : * user switch and sandbox.
2603 : */
2604 6 : switch (workitem->avw_type)
2605 : {
2606 6 : case AVW_BRINSummarizeRange:
2607 6 : DirectFunctionCall2(brin_summarize_range,
2608 : ObjectIdGetDatum(workitem->avw_relation),
2609 : Int64GetDatum((int64) workitem->avw_blockNumber));
2610 6 : break;
2611 0 : default:
2612 0 : elog(WARNING, "unrecognized work item found: type %d",
2613 : workitem->avw_type);
2614 0 : break;
2615 : }
2616 :
2617 : /*
2618 : * Clear a possible query-cancel signal, to avoid a late reaction to
2619 : * an automatically-sent signal because of vacuuming the current table
2620 : * (we're done with it, so it would make no sense to cancel at this
2621 : * point.)
2622 : */
2623 6 : QueryCancelPending = false;
2624 : }
2625 0 : PG_CATCH();
2626 : {
2627 : /*
2628 : * Abort the transaction, start a new one, and proceed with the next
2629 : * table in our list.
2630 : */
2631 0 : HOLD_INTERRUPTS();
2632 0 : errcontext("processing work entry for relation \"%s.%s.%s\"",
2633 : cur_datname, cur_nspname, cur_relname);
2634 0 : EmitErrorReport();
2635 :
2636 : /* this resets ProcGlobal->statusFlags[i] too */
2637 0 : AbortOutOfAnyTransaction();
2638 0 : FlushErrorState();
2639 0 : MemoryContextReset(PortalContext);
2640 :
2641 : /* restart our transaction for the following operations */
2642 0 : StartTransactionCommand();
2643 0 : RESUME_INTERRUPTS();
2644 : }
2645 6 : PG_END_TRY();
2646 :
2647 : /* Make sure we're back in AutovacMemCxt */
2648 6 : MemoryContextSwitchTo(AutovacMemCxt);
2649 :
2650 : /* We intentionally do not set did_vacuum here */
2651 :
2652 : /* be tidy */
2653 6 : deleted2:
2654 6 : if (cur_datname)
2655 6 : pfree(cur_datname);
2656 6 : if (cur_nspname)
2657 6 : pfree(cur_nspname);
2658 6 : if (cur_relname)
2659 6 : pfree(cur_relname);
2660 6 : }
2661 :
2662 : /*
2663 : * extract_autovac_opts
2664 : *
2665 : * Given a relation's pg_class tuple, return the AutoVacOpts portion of
2666 : * reloptions, if set; otherwise, return NULL.
2667 : *
2668 : * Note: callers do not have a relation lock on the table at this point,
2669 : * so the table could have been dropped, and its catalog rows gone, after
2670 : * we acquired the pg_class row. If pg_class had a TOAST table, this would
2671 : * be a risk; fortunately, it doesn't.
2672 : */
2673 : static AutoVacOpts *
2674 3862 : extract_autovac_opts(HeapTuple tup, TupleDesc pg_class_desc)
2675 : {
2676 : bytea *relopts;
2677 : AutoVacOpts *av;
2678 :
2679 : Assert(((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_RELATION ||
2680 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_MATVIEW ||
2681 : ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_TOASTVALUE);
2682 :
2683 3862 : relopts = extractRelOptions(tup, pg_class_desc, NULL);
2684 3862 : if (relopts == NULL)
2685 3820 : return NULL;
2686 :
2687 42 : av = palloc(sizeof(AutoVacOpts));
2688 42 : memcpy(av, &(((StdRdOptions *) relopts)->autovacuum), sizeof(AutoVacOpts));
2689 42 : pfree(relopts);
2690 :
2691 42 : return av;
2692 : }
2693 :
2694 :
2695 : /*
2696 : * table_recheck_autovac
2697 : *
2698 : * Recheck whether a table still needs vacuum or analyze. Return value is a
2699 : * valid autovac_table pointer if it does, NULL otherwise.
2700 : *
2701 : * Note that the returned autovac_table does not have the name fields set.
2702 : */
2703 : static autovac_table *
2704 236 : table_recheck_autovac(Oid relid, HTAB *table_toast_map,
2705 : TupleDesc pg_class_desc,
2706 : int effective_multixact_freeze_max_age)
2707 : {
2708 : Form_pg_class classForm;
2709 : HeapTuple classTup;
2710 : bool dovacuum;
2711 : bool doanalyze;
2712 236 : autovac_table *tab = NULL;
2713 : bool wraparound;
2714 : AutoVacOpts *avopts;
2715 :
2716 : /* fetch the relation's relcache entry */
2717 236 : classTup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid));
2718 236 : if (!HeapTupleIsValid(classTup))
2719 0 : return NULL;
2720 236 : classForm = (Form_pg_class) GETSTRUCT(classTup);
2721 :
2722 : /*
2723 : * Get the applicable reloptions. If it is a TOAST table, try to get the
2724 : * main table reloptions if the toast table itself doesn't have.
2725 : */
2726 236 : avopts = extract_autovac_opts(classTup, pg_class_desc);
2727 236 : if (classForm->relkind == RELKIND_TOASTVALUE &&
2728 2 : avopts == NULL && table_toast_map != NULL)
2729 : {
2730 : av_relation *hentry;
2731 : bool found;
2732 :
2733 2 : hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found);
2734 2 : if (found && hentry->ar_hasrelopts)
2735 0 : avopts = &hentry->ar_reloptions;
2736 : }
2737 :
2738 236 : recheck_relation_needs_vacanalyze(relid, avopts, classForm,
2739 : effective_multixact_freeze_max_age,
2740 : &dovacuum, &doanalyze, &wraparound);
2741 :
2742 : /* OK, it needs something done */
2743 236 : if (doanalyze || dovacuum)
2744 : {
2745 : int freeze_min_age;
2746 : int freeze_table_age;
2747 : int multixact_freeze_min_age;
2748 : int multixact_freeze_table_age;
2749 : int log_min_duration;
2750 :
2751 : /*
2752 : * Calculate the vacuum cost parameters and the freeze ages. If there
2753 : * are options set in pg_class.reloptions, use them; in the case of a
2754 : * toast table, try the main table too. Otherwise use the GUC
2755 : * defaults, autovacuum's own first and plain vacuum second.
2756 : */
2757 :
2758 : /* -1 in autovac setting means use log_autovacuum_min_duration */
2759 2 : log_min_duration = (avopts && avopts->log_min_duration >= 0)
2760 : ? avopts->log_min_duration
2761 210 : : Log_autovacuum_min_duration;
2762 :
2763 : /* these do not have autovacuum-specific settings */
2764 2 : freeze_min_age = (avopts && avopts->freeze_min_age >= 0)
2765 : ? avopts->freeze_min_age
2766 210 : : default_freeze_min_age;
2767 :
2768 2 : freeze_table_age = (avopts && avopts->freeze_table_age >= 0)
2769 : ? avopts->freeze_table_age
2770 210 : : default_freeze_table_age;
2771 :
2772 210 : multixact_freeze_min_age = (avopts &&
2773 2 : avopts->multixact_freeze_min_age >= 0)
2774 : ? avopts->multixact_freeze_min_age
2775 210 : : default_multixact_freeze_min_age;
2776 :
2777 210 : multixact_freeze_table_age = (avopts &&
2778 2 : avopts->multixact_freeze_table_age >= 0)
2779 : ? avopts->multixact_freeze_table_age
2780 210 : : default_multixact_freeze_table_age;
2781 :
2782 208 : tab = palloc(sizeof(autovac_table));
2783 208 : tab->at_relid = relid;
2784 208 : tab->at_sharedrel = classForm->relisshared;
2785 :
2786 : /*
2787 : * Select VACUUM options. Note we don't say VACOPT_PROCESS_TOAST, so
2788 : * that vacuum() skips toast relations. Also note we tell vacuum() to
2789 : * skip vac_update_datfrozenxid(); we'll do that separately.
2790 : */
2791 208 : tab->at_params.options =
2792 208 : (dovacuum ? (VACOPT_VACUUM |
2793 : VACOPT_PROCESS_MAIN |
2794 208 : VACOPT_SKIP_DATABASE_STATS) : 0) |
2795 208 : (doanalyze ? VACOPT_ANALYZE : 0) |
2796 208 : (!wraparound ? VACOPT_SKIP_LOCKED : 0);
2797 :
2798 : /*
2799 : * index_cleanup and truncate are unspecified at first in autovacuum.
2800 : * They will be filled in with usable values using their reloptions
2801 : * (or reloption defaults) later.
2802 : */
2803 208 : tab->at_params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
2804 208 : tab->at_params.truncate = VACOPTVALUE_UNSPECIFIED;
2805 : /* As of now, we don't support parallel vacuum for autovacuum */
2806 208 : tab->at_params.nworkers = -1;
2807 208 : tab->at_params.freeze_min_age = freeze_min_age;
2808 208 : tab->at_params.freeze_table_age = freeze_table_age;
2809 208 : tab->at_params.multixact_freeze_min_age = multixact_freeze_min_age;
2810 208 : tab->at_params.multixact_freeze_table_age = multixact_freeze_table_age;
2811 208 : tab->at_params.is_wraparound = wraparound;
2812 208 : tab->at_params.log_min_duration = log_min_duration;
2813 208 : tab->at_params.toast_parent = InvalidOid;
2814 208 : tab->at_storage_param_vac_cost_limit = avopts ?
2815 208 : avopts->vacuum_cost_limit : 0;
2816 208 : tab->at_storage_param_vac_cost_delay = avopts ?
2817 208 : avopts->vacuum_cost_delay : -1;
2818 208 : tab->at_relname = NULL;
2819 208 : tab->at_nspname = NULL;
2820 208 : tab->at_datname = NULL;
2821 :
2822 : /*
2823 : * If any of the cost delay parameters has been set individually for
2824 : * this table, disable the balancing algorithm.
2825 : */
2826 208 : tab->at_dobalance =
2827 210 : !(avopts && (avopts->vacuum_cost_limit > 0 ||
2828 2 : avopts->vacuum_cost_delay >= 0));
2829 : }
2830 :
2831 236 : heap_freetuple(classTup);
2832 236 : return tab;
2833 : }
2834 :
2835 : /*
2836 : * recheck_relation_needs_vacanalyze
2837 : *
2838 : * Subroutine for table_recheck_autovac.
2839 : *
2840 : * Fetch the pgstat of a relation and recheck whether a relation
2841 : * needs to be vacuumed or analyzed.
2842 : */
2843 : static void
2844 236 : recheck_relation_needs_vacanalyze(Oid relid,
2845 : AutoVacOpts *avopts,
2846 : Form_pg_class classForm,
2847 : int effective_multixact_freeze_max_age,
2848 : bool *dovacuum,
2849 : bool *doanalyze,
2850 : bool *wraparound)
2851 : {
2852 : PgStat_StatTabEntry *tabentry;
2853 :
2854 : /* fetch the pgstat table entry */
2855 236 : tabentry = pgstat_fetch_stat_tabentry_ext(classForm->relisshared,
2856 : relid);
2857 :
2858 236 : relation_needs_vacanalyze(relid, avopts, classForm, tabentry,
2859 : effective_multixact_freeze_max_age,
2860 : dovacuum, doanalyze, wraparound);
2861 :
2862 : /* ignore ANALYZE for toast tables */
2863 236 : if (classForm->relkind == RELKIND_TOASTVALUE)
2864 2 : *doanalyze = false;
2865 236 : }
2866 :
2867 : /*
2868 : * relation_needs_vacanalyze
2869 : *
2870 : * Check whether a relation needs to be vacuumed or analyzed; return each into
2871 : * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is
2872 : * being forced because of Xid or multixact wraparound.
2873 : *
2874 : * relopts is a pointer to the AutoVacOpts options (either for itself in the
2875 : * case of a plain table, or for either itself or its parent table in the case
2876 : * of a TOAST table), NULL if none; tabentry is the pgstats entry, which can be
2877 : * NULL.
2878 : *
2879 : * A table needs to be vacuumed if the number of dead tuples exceeds a
2880 : * threshold. This threshold is calculated as
2881 : *
2882 : * threshold = vac_base_thresh + vac_scale_factor * reltuples
2883 : *
2884 : * For analyze, the analysis done is that the number of tuples inserted,
2885 : * deleted and updated since the last analyze exceeds a threshold calculated
2886 : * in the same fashion as above. Note that the cumulative stats system stores
2887 : * the number of tuples (both live and dead) that there were as of the last
2888 : * analyze. This is asymmetric to the VACUUM case.
2889 : *
2890 : * We also force vacuum if the table's relfrozenxid is more than freeze_max_age
2891 : * transactions back, and if its relminmxid is more than
2892 : * multixact_freeze_max_age multixacts back.
2893 : *
2894 : * A table whose autovacuum_enabled option is false is
2895 : * automatically skipped (unless we have to vacuum it due to freeze_max_age).
2896 : * Thus autovacuum can be disabled for specific tables. Also, when the cumulative
2897 : * stats system does not have data about a table, it will be skipped.
2898 : *
2899 : * A table whose vac_base_thresh value is < 0 takes the base value from the
2900 : * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor
2901 : * value < 0 is substituted with the value of
2902 : * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze.
2903 : */
2904 : static void
2905 3862 : relation_needs_vacanalyze(Oid relid,
2906 : AutoVacOpts *relopts,
2907 : Form_pg_class classForm,
2908 : PgStat_StatTabEntry *tabentry,
2909 : int effective_multixact_freeze_max_age,
2910 : /* output params below */
2911 : bool *dovacuum,
2912 : bool *doanalyze,
2913 : bool *wraparound)
2914 : {
2915 : bool force_vacuum;
2916 : bool av_enabled;
2917 : float4 reltuples; /* pg_class.reltuples */
2918 :
2919 : /* constants from reloptions or GUC variables */
2920 : int vac_base_thresh,
2921 : vac_ins_base_thresh,
2922 : anl_base_thresh;
2923 : float4 vac_scale_factor,
2924 : vac_ins_scale_factor,
2925 : anl_scale_factor;
2926 :
2927 : /* thresholds calculated from above constants */
2928 : float4 vacthresh,
2929 : vacinsthresh,
2930 : anlthresh;
2931 :
2932 : /* number of vacuum (resp. analyze) tuples at this time */
2933 : float4 vactuples,
2934 : instuples,
2935 : anltuples;
2936 :
2937 : /* freeze parameters */
2938 : int freeze_max_age;
2939 : int multixact_freeze_max_age;
2940 : TransactionId xidForceLimit;
2941 : MultiXactId multiForceLimit;
2942 :
2943 : Assert(classForm != NULL);
2944 : Assert(OidIsValid(relid));
2945 :
2946 : /*
2947 : * Determine vacuum/analyze equation parameters. We have two possible
2948 : * sources: the passed reloptions (which could be a main table or a toast
2949 : * table), or the autovacuum GUC variables.
2950 : */
2951 :
2952 : /* -1 in autovac setting means use plain vacuum_scale_factor */
2953 60 : vac_scale_factor = (relopts && relopts->vacuum_scale_factor >= 0)
2954 0 : ? relopts->vacuum_scale_factor
2955 3922 : : autovacuum_vac_scale;
2956 :
2957 60 : vac_base_thresh = (relopts && relopts->vacuum_threshold >= 0)
2958 : ? relopts->vacuum_threshold
2959 3922 : : autovacuum_vac_thresh;
2960 :
2961 60 : vac_ins_scale_factor = (relopts && relopts->vacuum_ins_scale_factor >= 0)
2962 0 : ? relopts->vacuum_ins_scale_factor
2963 3922 : : autovacuum_vac_ins_scale;
2964 :
2965 : /* -1 is used to disable insert vacuums */
2966 60 : vac_ins_base_thresh = (relopts && relopts->vacuum_ins_threshold >= -1)
2967 : ? relopts->vacuum_ins_threshold
2968 3922 : : autovacuum_vac_ins_thresh;
2969 :
2970 60 : anl_scale_factor = (relopts && relopts->analyze_scale_factor >= 0)
2971 0 : ? relopts->analyze_scale_factor
2972 3922 : : autovacuum_anl_scale;
2973 :
2974 60 : anl_base_thresh = (relopts && relopts->analyze_threshold >= 0)
2975 : ? relopts->analyze_threshold
2976 3922 : : autovacuum_anl_thresh;
2977 :
2978 60 : freeze_max_age = (relopts && relopts->freeze_max_age >= 0)
2979 0 : ? Min(relopts->freeze_max_age, autovacuum_freeze_max_age)
2980 3922 : : autovacuum_freeze_max_age;
2981 :
2982 60 : multixact_freeze_max_age = (relopts && relopts->multixact_freeze_max_age >= 0)
2983 0 : ? Min(relopts->multixact_freeze_max_age, effective_multixact_freeze_max_age)
2984 3922 : : effective_multixact_freeze_max_age;
2985 :
2986 3862 : av_enabled = (relopts ? relopts->enabled : true);
2987 :
2988 : /* Force vacuum if table is at risk of wraparound */
2989 3862 : xidForceLimit = recentXid - freeze_max_age;
2990 3862 : if (xidForceLimit < FirstNormalTransactionId)
2991 0 : xidForceLimit -= FirstNormalTransactionId;
2992 7724 : force_vacuum = (TransactionIdIsNormal(classForm->relfrozenxid) &&
2993 3862 : TransactionIdPrecedes(classForm->relfrozenxid,
2994 : xidForceLimit));
2995 3862 : if (!force_vacuum)
2996 : {
2997 3862 : multiForceLimit = recentMulti - multixact_freeze_max_age;
2998 3862 : if (multiForceLimit < FirstMultiXactId)
2999 0 : multiForceLimit -= FirstMultiXactId;
3000 7724 : force_vacuum = MultiXactIdIsValid(classForm->relminmxid) &&
3001 3862 : MultiXactIdPrecedes(classForm->relminmxid, multiForceLimit);
3002 : }
3003 3862 : *wraparound = force_vacuum;
3004 :
3005 : /* User disabled it in pg_class.reloptions? (But ignore if at risk) */
3006 3862 : if (!av_enabled && !force_vacuum)
3007 : {
3008 40 : *doanalyze = false;
3009 40 : *dovacuum = false;
3010 40 : return;
3011 : }
3012 :
3013 : /*
3014 : * If we found stats for the table, and autovacuum is currently enabled,
3015 : * make a threshold-based decision whether to vacuum and/or analyze. If
3016 : * autovacuum is currently disabled, we must be here for anti-wraparound
3017 : * vacuuming only, so don't vacuum (or analyze) anything that's not being
3018 : * forced.
3019 : */
3020 3822 : if (PointerIsValid(tabentry) && AutoVacuumingActive())
3021 : {
3022 2506 : reltuples = classForm->reltuples;
3023 2506 : vactuples = tabentry->dead_tuples;
3024 2506 : instuples = tabentry->ins_since_vacuum;
3025 2506 : anltuples = tabentry->mod_since_analyze;
3026 :
3027 : /* If the table hasn't yet been vacuumed, take reltuples as zero */
3028 2506 : if (reltuples < 0)
3029 670 : reltuples = 0;
3030 :
3031 2506 : vacthresh = (float4) vac_base_thresh + vac_scale_factor * reltuples;
3032 2506 : vacinsthresh = (float4) vac_ins_base_thresh + vac_ins_scale_factor * reltuples;
3033 2506 : anlthresh = (float4) anl_base_thresh + anl_scale_factor * reltuples;
3034 :
3035 : /*
3036 : * Note that we don't need to take special consideration for stat
3037 : * reset, because if that happens, the last vacuum and analyze counts
3038 : * will be reset too.
3039 : */
3040 2506 : if (vac_ins_base_thresh >= 0)
3041 2506 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: %.0f (threshold %.0f), anl: %.0f (threshold %.0f)",
3042 : NameStr(classForm->relname),
3043 : vactuples, vacthresh, instuples, vacinsthresh, anltuples, anlthresh);
3044 : else
3045 0 : elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), ins: (disabled), anl: %.0f (threshold %.0f)",
3046 : NameStr(classForm->relname),
3047 : vactuples, vacthresh, anltuples, anlthresh);
3048 :
3049 : /* Determine if this table needs vacuum or analyze. */
3050 4856 : *dovacuum = force_vacuum || (vactuples > vacthresh) ||
3051 2350 : (vac_ins_base_thresh >= 0 && instuples > vacinsthresh);
3052 2506 : *doanalyze = (anltuples > anlthresh);
3053 : }
3054 : else
3055 : {
3056 : /*
3057 : * Skip a table not found in stat hash, unless we have to force vacuum
3058 : * for anti-wrap purposes. If it's not acted upon, there's no need to
3059 : * vacuum it.
3060 : */
3061 1316 : *dovacuum = force_vacuum;
3062 1316 : *doanalyze = false;
3063 : }
3064 :
3065 : /* ANALYZE refuses to work with pg_statistic */
3066 3822 : if (relid == StatisticRelationId)
3067 24 : *doanalyze = false;
3068 : }
3069 :
3070 : /*
3071 : * autovacuum_do_vac_analyze
3072 : * Vacuum and/or analyze the specified table
3073 : *
3074 : * We expect the caller to have switched into a memory context that won't
3075 : * disappear at transaction commit.
3076 : */
3077 : static void
3078 208 : autovacuum_do_vac_analyze(autovac_table *tab, BufferAccessStrategy bstrategy)
3079 : {
3080 : RangeVar *rangevar;
3081 : VacuumRelation *rel;
3082 : List *rel_list;
3083 : MemoryContext vac_context;
3084 :
3085 : /* Let pgstat know what we're doing */
3086 208 : autovac_report_activity(tab);
3087 :
3088 : /* Set up one VacuumRelation target, identified by OID, for vacuum() */
3089 208 : rangevar = makeRangeVar(tab->at_nspname, tab->at_relname, -1);
3090 208 : rel = makeVacuumRelation(rangevar, tab->at_relid, NIL);
3091 208 : rel_list = list_make1(rel);
3092 :
3093 208 : vac_context = AllocSetContextCreate(CurrentMemoryContext,
3094 : "Vacuum",
3095 : ALLOCSET_DEFAULT_SIZES);
3096 :
3097 208 : vacuum(rel_list, &tab->at_params, bstrategy, vac_context, true);
3098 :
3099 208 : MemoryContextDelete(vac_context);
3100 208 : }
3101 :
3102 : /*
3103 : * autovac_report_activity
3104 : * Report to pgstat what autovacuum is doing
3105 : *
3106 : * We send a SQL string corresponding to what the user would see if the
3107 : * equivalent command was to be issued manually.
3108 : *
3109 : * Note we assume that we are going to report the next command as soon as we're
3110 : * done with the current one, and exit right after the last one, so we don't
3111 : * bother to report "<IDLE>" or some such.
3112 : */
3113 : static void
3114 208 : autovac_report_activity(autovac_table *tab)
3115 : {
3116 : #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56)
3117 : char activity[MAX_AUTOVAC_ACTIV_LEN];
3118 : int len;
3119 :
3120 : /* Report the command and possible options */
3121 208 : if (tab->at_params.options & VACOPT_VACUUM)
3122 100 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3123 : "autovacuum: VACUUM%s",
3124 100 : tab->at_params.options & VACOPT_ANALYZE ? " ANALYZE" : "");
3125 : else
3126 108 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3127 : "autovacuum: ANALYZE");
3128 :
3129 : /*
3130 : * Report the qualified name of the relation.
3131 : */
3132 208 : len = strlen(activity);
3133 :
3134 208 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3135 : " %s.%s%s", tab->at_nspname, tab->at_relname,
3136 208 : tab->at_params.is_wraparound ? " (to prevent wraparound)" : "");
3137 :
3138 : /* Set statement_timestamp() to current time for pg_stat_activity */
3139 208 : SetCurrentStatementStartTimestamp();
3140 :
3141 208 : pgstat_report_activity(STATE_RUNNING, activity);
3142 208 : }
3143 :
3144 : /*
3145 : * autovac_report_workitem
3146 : * Report to pgstat that autovacuum is processing a work item
3147 : */
3148 : static void
3149 6 : autovac_report_workitem(AutoVacuumWorkItem *workitem,
3150 : const char *nspname, const char *relname)
3151 : {
3152 : char activity[MAX_AUTOVAC_ACTIV_LEN + 12 + 2];
3153 : char blk[12 + 2];
3154 : int len;
3155 :
3156 6 : switch (workitem->avw_type)
3157 : {
3158 6 : case AVW_BRINSummarizeRange:
3159 6 : snprintf(activity, MAX_AUTOVAC_ACTIV_LEN,
3160 : "autovacuum: BRIN summarize");
3161 6 : break;
3162 : }
3163 :
3164 : /*
3165 : * Report the qualified name of the relation, and the block number if any
3166 : */
3167 6 : len = strlen(activity);
3168 :
3169 6 : if (BlockNumberIsValid(workitem->avw_blockNumber))
3170 6 : snprintf(blk, sizeof(blk), " %u", workitem->avw_blockNumber);
3171 : else
3172 0 : blk[0] = '\0';
3173 :
3174 6 : snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len,
3175 : " %s.%s%s", nspname, relname, blk);
3176 :
3177 : /* Set statement_timestamp() to current time for pg_stat_activity */
3178 6 : SetCurrentStatementStartTimestamp();
3179 :
3180 6 : pgstat_report_activity(STATE_RUNNING, activity);
3181 6 : }
3182 :
3183 : /*
3184 : * AutoVacuumingActive
3185 : * Check GUC vars and report whether the autovacuum process should be
3186 : * running.
3187 : */
3188 : bool
3189 15206 : AutoVacuumingActive(void)
3190 : {
3191 15206 : if (!autovacuum_start_daemon || !pgstat_track_counts)
3192 5494 : return false;
3193 9712 : return true;
3194 : }
3195 :
3196 : /*
3197 : * Request one work item to the next autovacuum run processing our database.
3198 : * Return false if the request can't be recorded.
3199 : */
3200 : bool
3201 6 : AutoVacuumRequestWork(AutoVacuumWorkItemType type, Oid relationId,
3202 : BlockNumber blkno)
3203 : {
3204 : int i;
3205 6 : bool result = false;
3206 :
3207 6 : LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE);
3208 :
3209 : /*
3210 : * Locate an unused work item and fill it with the given data.
3211 : */
3212 12 : for (i = 0; i < NUM_WORKITEMS; i++)
3213 : {
3214 12 : AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i];
3215 :
3216 12 : if (workitem->avw_used)
3217 6 : continue;
3218 :
3219 6 : workitem->avw_used = true;
3220 6 : workitem->avw_active = false;
3221 6 : workitem->avw_type = type;
3222 6 : workitem->avw_database = MyDatabaseId;
3223 6 : workitem->avw_relation = relationId;
3224 6 : workitem->avw_blockNumber = blkno;
3225 6 : result = true;
3226 :
3227 : /* done */
3228 6 : break;
3229 : }
3230 :
3231 6 : LWLockRelease(AutovacuumLock);
3232 :
3233 6 : return result;
3234 : }
3235 :
3236 : /*
3237 : * autovac_init
3238 : * This is called at postmaster initialization.
3239 : *
3240 : * All we do here is annoy the user if he got it wrong.
3241 : */
3242 : void
3243 1422 : autovac_init(void)
3244 : {
3245 1422 : if (autovacuum_start_daemon && !pgstat_track_counts)
3246 0 : ereport(WARNING,
3247 : (errmsg("autovacuum not started because of misconfiguration"),
3248 : errhint("Enable the \"track_counts\" option.")));
3249 1422 : }
3250 :
3251 : /*
3252 : * AutoVacuumShmemSize
3253 : * Compute space needed for autovacuum-related shared memory
3254 : */
3255 : Size
3256 5066 : AutoVacuumShmemSize(void)
3257 : {
3258 : Size size;
3259 :
3260 : /*
3261 : * Need the fixed struct and the array of WorkerInfoData.
3262 : */
3263 5066 : size = sizeof(AutoVacuumShmemStruct);
3264 5066 : size = MAXALIGN(size);
3265 5066 : size = add_size(size, mul_size(autovacuum_max_workers,
3266 : sizeof(WorkerInfoData)));
3267 5066 : return size;
3268 : }
3269 :
3270 : /*
3271 : * AutoVacuumShmemInit
3272 : * Allocate and initialize autovacuum-related shared memory
3273 : */
3274 : void
3275 1768 : AutoVacuumShmemInit(void)
3276 : {
3277 : bool found;
3278 :
3279 1768 : AutoVacuumShmem = (AutoVacuumShmemStruct *)
3280 1768 : ShmemInitStruct("AutoVacuum Data",
3281 : AutoVacuumShmemSize(),
3282 : &found);
3283 :
3284 1768 : if (!IsUnderPostmaster)
3285 : {
3286 : WorkerInfo worker;
3287 : int i;
3288 :
3289 : Assert(!found);
3290 :
3291 1768 : AutoVacuumShmem->av_launcherpid = 0;
3292 1768 : dlist_init(&AutoVacuumShmem->av_freeWorkers);
3293 1768 : dlist_init(&AutoVacuumShmem->av_runningWorkers);
3294 1768 : AutoVacuumShmem->av_startingWorker = NULL;
3295 1768 : memset(AutoVacuumShmem->av_workItems, 0,
3296 : sizeof(AutoVacuumWorkItem) * NUM_WORKITEMS);
3297 :
3298 1768 : worker = (WorkerInfo) ((char *) AutoVacuumShmem +
3299 : MAXALIGN(sizeof(AutoVacuumShmemStruct)));
3300 :
3301 : /* initialize the WorkerInfo free list */
3302 7072 : for (i = 0; i < autovacuum_max_workers; i++)
3303 : {
3304 5304 : dlist_push_head(&AutoVacuumShmem->av_freeWorkers,
3305 5304 : &worker[i].wi_links);
3306 5304 : pg_atomic_init_flag(&worker[i].wi_dobalance);
3307 : }
3308 :
3309 1768 : pg_atomic_init_u32(&AutoVacuumShmem->av_nworkersForBalance, 0);
3310 :
3311 : }
3312 : else
3313 : Assert(found);
3314 1768 : }
3315 :
3316 : /*
3317 : * GUC check_hook for autovacuum_work_mem
3318 : */
3319 : bool
3320 1834 : check_autovacuum_work_mem(int *newval, void **extra, GucSource source)
3321 : {
3322 : /*
3323 : * -1 indicates fallback.
3324 : *
3325 : * If we haven't yet changed the boot_val default of -1, just let it be.
3326 : * Autovacuum will look to maintenance_work_mem instead.
3327 : */
3328 1834 : if (*newval == -1)
3329 1830 : return true;
3330 :
3331 : /*
3332 : * We clamp manually-set values to at least 1MB. Since
3333 : * maintenance_work_mem is always set to at least this value, do the same
3334 : * here.
3335 : */
3336 4 : if (*newval < 1024)
3337 4 : *newval = 1024;
3338 :
3339 4 : return true;
3340 : }
|