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 reload the pgstats data just before vacuuming each |
48 | * table, to avoid vacuuming a table that was just finished being vacuumed by |
49 | * another worker and thus is no longer noted in shared memory. However, |
50 | * there is a window (caused by pgstat delay) on which a worker may choose a |
51 | * table that was already vacuumed; this is a bug in the current design. |
52 | * |
53 | * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group |
54 | * Portions Copyright (c) 1994, Regents of the University of California |
55 | * |
56 | * |
57 | * IDENTIFICATION |
58 | * src/backend/postmaster/autovacuum.c |
59 | * |
60 | *------------------------------------------------------------------------- |
61 | */ |
62 | #include "postgres.h" |
63 | |
64 | #include <signal.h> |
65 | #include <sys/time.h> |
66 | #include <unistd.h> |
67 | |
68 | #include "access/heapam.h" |
69 | #include "access/htup_details.h" |
70 | #include "access/multixact.h" |
71 | #include "access/reloptions.h" |
72 | #include "access/tableam.h" |
73 | #include "access/transam.h" |
74 | #include "access/xact.h" |
75 | #include "catalog/dependency.h" |
76 | #include "catalog/namespace.h" |
77 | #include "catalog/pg_database.h" |
78 | #include "commands/dbcommands.h" |
79 | #include "commands/vacuum.h" |
80 | #include "lib/ilist.h" |
81 | #include "libpq/pqsignal.h" |
82 | #include "miscadmin.h" |
83 | #include "nodes/makefuncs.h" |
84 | #include "pgstat.h" |
85 | #include "postmaster/autovacuum.h" |
86 | #include "postmaster/fork_process.h" |
87 | #include "postmaster/postmaster.h" |
88 | #include "storage/bufmgr.h" |
89 | #include "storage/ipc.h" |
90 | #include "storage/latch.h" |
91 | #include "storage/lmgr.h" |
92 | #include "storage/pmsignal.h" |
93 | #include "storage/proc.h" |
94 | #include "storage/procsignal.h" |
95 | #include "storage/sinvaladt.h" |
96 | #include "storage/smgr.h" |
97 | #include "tcop/tcopprot.h" |
98 | #include "utils/fmgroids.h" |
99 | #include "utils/fmgrprotos.h" |
100 | #include "utils/lsyscache.h" |
101 | #include "utils/memutils.h" |
102 | #include "utils/ps_status.h" |
103 | #include "utils/rel.h" |
104 | #include "utils/snapmgr.h" |
105 | #include "utils/syscache.h" |
106 | #include "utils/timeout.h" |
107 | #include "utils/timestamp.h" |
108 | |
109 | |
110 | /* |
111 | * GUC parameters |
112 | */ |
113 | bool autovacuum_start_daemon = false; |
114 | int autovacuum_max_workers; |
115 | int autovacuum_work_mem = -1; |
116 | int autovacuum_naptime; |
117 | int autovacuum_vac_thresh; |
118 | double autovacuum_vac_scale; |
119 | int autovacuum_anl_thresh; |
120 | double autovacuum_anl_scale; |
121 | int autovacuum_freeze_max_age; |
122 | int autovacuum_multixact_freeze_max_age; |
123 | |
124 | double autovacuum_vac_cost_delay; |
125 | int autovacuum_vac_cost_limit; |
126 | |
127 | int Log_autovacuum_min_duration = -1; |
128 | |
129 | /* how long to keep pgstat data in the launcher, in milliseconds */ |
130 | #define STATS_READ_DELAY 1000 |
131 | |
132 | /* the minimum allowed time between two awakenings of the launcher */ |
133 | #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */ |
134 | #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */ |
135 | |
136 | /* Flags to tell if we are in an autovacuum process */ |
137 | static bool am_autovacuum_launcher = false; |
138 | static bool am_autovacuum_worker = false; |
139 | |
140 | /* Flags set by signal handlers */ |
141 | static volatile sig_atomic_t got_SIGHUP = false; |
142 | static volatile sig_atomic_t got_SIGUSR2 = false; |
143 | static volatile sig_atomic_t got_SIGTERM = false; |
144 | |
145 | /* Comparison points for determining whether freeze_max_age is exceeded */ |
146 | static TransactionId recentXid; |
147 | static MultiXactId recentMulti; |
148 | |
149 | /* Default freeze ages to use for autovacuum (varies by database) */ |
150 | static int default_freeze_min_age; |
151 | static int default_freeze_table_age; |
152 | static int default_multixact_freeze_min_age; |
153 | static int default_multixact_freeze_table_age; |
154 | |
155 | /* Memory context for long-lived data */ |
156 | static MemoryContext AutovacMemCxt; |
157 | |
158 | /* struct to keep track of databases in launcher */ |
159 | typedef struct avl_dbase |
160 | { |
161 | Oid adl_datid; /* hash key -- must be first */ |
162 | TimestampTz adl_next_worker; |
163 | int adl_score; |
164 | dlist_node adl_node; |
165 | } avl_dbase; |
166 | |
167 | /* struct to keep track of databases in worker */ |
168 | typedef struct avw_dbase |
169 | { |
170 | Oid adw_datid; |
171 | char *adw_name; |
172 | TransactionId adw_frozenxid; |
173 | MultiXactId adw_minmulti; |
174 | PgStat_StatDBEntry *adw_entry; |
175 | } avw_dbase; |
176 | |
177 | /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */ |
178 | typedef struct av_relation |
179 | { |
180 | Oid ar_toastrelid; /* hash key - must be first */ |
181 | Oid ar_relid; |
182 | bool ar_hasrelopts; |
183 | AutoVacOpts ar_reloptions; /* copy of AutoVacOpts from the main table's |
184 | * reloptions, or NULL if none */ |
185 | } av_relation; |
186 | |
187 | /* struct to keep track of tables to vacuum and/or analyze, after rechecking */ |
188 | typedef struct autovac_table |
189 | { |
190 | Oid at_relid; |
191 | VacuumParams at_params; |
192 | double at_vacuum_cost_delay; |
193 | int at_vacuum_cost_limit; |
194 | bool at_dobalance; |
195 | bool at_sharedrel; |
196 | char *at_relname; |
197 | char *at_nspname; |
198 | char *at_datname; |
199 | } autovac_table; |
200 | |
201 | /*------------- |
202 | * This struct holds information about a single worker's whereabouts. We keep |
203 | * an array of these in shared memory, sized according to |
204 | * autovacuum_max_workers. |
205 | * |
206 | * wi_links entry into free list or running list |
207 | * wi_dboid OID of the database this worker is supposed to work on |
208 | * wi_tableoid OID of the table currently being vacuumed, if any |
209 | * wi_sharedrel flag indicating whether table is marked relisshared |
210 | * wi_proc pointer to PGPROC of the running worker, NULL if not started |
211 | * wi_launchtime Time at which this worker was launched |
212 | * wi_cost_* Vacuum cost-based delay parameters current in this worker |
213 | * |
214 | * All fields are protected by AutovacuumLock, except for wi_tableoid and |
215 | * wi_sharedrel which are protected by AutovacuumScheduleLock (note these |
216 | * two fields are read-only for everyone except that worker itself). |
217 | *------------- |
218 | */ |
219 | typedef struct WorkerInfoData |
220 | { |
221 | dlist_node wi_links; |
222 | Oid wi_dboid; |
223 | Oid wi_tableoid; |
224 | PGPROC *wi_proc; |
225 | TimestampTz wi_launchtime; |
226 | bool wi_dobalance; |
227 | bool wi_sharedrel; |
228 | double wi_cost_delay; |
229 | int wi_cost_limit; |
230 | int wi_cost_limit_base; |
231 | } WorkerInfoData; |
232 | |
233 | typedef struct WorkerInfoData *WorkerInfo; |
234 | |
235 | /* |
236 | * Possible signals received by the launcher from remote processes. These are |
237 | * stored atomically in shared memory so that other processes can set them |
238 | * without locking. |
239 | */ |
240 | typedef enum |
241 | { |
242 | AutoVacForkFailed, /* failed trying to start a worker */ |
243 | AutoVacRebalance, /* rebalance the cost limits */ |
244 | AutoVacNumSignals /* must be last */ |
245 | } AutoVacuumSignal; |
246 | |
247 | /* |
248 | * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This |
249 | * list is mostly protected by AutovacuumLock, except that if an item is |
250 | * marked 'active' other processes must not modify the work-identifying |
251 | * members. |
252 | */ |
253 | typedef struct AutoVacuumWorkItem |
254 | { |
255 | AutoVacuumWorkItemType avw_type; |
256 | bool avw_used; /* below data is valid */ |
257 | bool avw_active; /* being processed */ |
258 | Oid avw_database; |
259 | Oid avw_relation; |
260 | BlockNumber avw_blockNumber; |
261 | } AutoVacuumWorkItem; |
262 | |
263 | #define NUM_WORKITEMS 256 |
264 | |
265 | /*------------- |
266 | * The main autovacuum shmem struct. On shared memory we store this main |
267 | * struct and the array of WorkerInfo structs. This struct keeps: |
268 | * |
269 | * av_signal set by other processes to indicate various conditions |
270 | * av_launcherpid the PID of the autovacuum launcher |
271 | * av_freeWorkers the WorkerInfo freelist |
272 | * av_runningWorkers the WorkerInfo non-free queue |
273 | * av_startingWorker pointer to WorkerInfo currently being started (cleared by |
274 | * the worker itself as soon as it's up and running) |
275 | * av_workItems work item array |
276 | * |
277 | * This struct is protected by AutovacuumLock, except for av_signal and parts |
278 | * of the worker list (see above). |
279 | *------------- |
280 | */ |
281 | typedef struct |
282 | { |
283 | sig_atomic_t av_signal[AutoVacNumSignals]; |
284 | pid_t av_launcherpid; |
285 | dlist_head av_freeWorkers; |
286 | dlist_head av_runningWorkers; |
287 | WorkerInfo av_startingWorker; |
288 | AutoVacuumWorkItem av_workItems[NUM_WORKITEMS]; |
289 | } AutoVacuumShmemStruct; |
290 | |
291 | static AutoVacuumShmemStruct *AutoVacuumShmem; |
292 | |
293 | /* |
294 | * the database list (of avl_dbase elements) in the launcher, and the context |
295 | * that contains it |
296 | */ |
297 | static dlist_head DatabaseList = DLIST_STATIC_INIT(DatabaseList); |
298 | static MemoryContext DatabaseListCxt = NULL; |
299 | |
300 | /* Pointer to my own WorkerInfo, valid on each worker */ |
301 | static WorkerInfo MyWorkerInfo = NULL; |
302 | |
303 | /* PID of launcher, valid only in worker while shutting down */ |
304 | int AutovacuumLauncherPid = 0; |
305 | |
306 | #ifdef EXEC_BACKEND |
307 | static pid_t avlauncher_forkexec(void); |
308 | static pid_t avworker_forkexec(void); |
309 | #endif |
310 | NON_EXEC_STATIC void AutoVacWorkerMain(int argc, char *argv[]) pg_attribute_noreturn(); |
311 | NON_EXEC_STATIC void AutoVacLauncherMain(int argc, char *argv[]) pg_attribute_noreturn(); |
312 | |
313 | static Oid do_start_worker(void); |
314 | static void launcher_determine_sleep(bool canlaunch, bool recursing, |
315 | struct timeval *nap); |
316 | static void launch_worker(TimestampTz now); |
317 | static List *get_database_list(void); |
318 | static void rebuild_database_list(Oid newdb); |
319 | static int db_comparator(const void *a, const void *b); |
320 | static void autovac_balance_cost(void); |
321 | |
322 | static void do_autovacuum(void); |
323 | static void FreeWorkerInfo(int code, Datum arg); |
324 | |
325 | static autovac_table *table_recheck_autovac(Oid relid, HTAB *table_toast_map, |
326 | TupleDesc pg_class_desc, |
327 | int effective_multixact_freeze_max_age); |
328 | static void relation_needs_vacanalyze(Oid relid, AutoVacOpts *relopts, |
329 | Form_pg_class classForm, |
330 | PgStat_StatTabEntry *tabentry, |
331 | int effective_multixact_freeze_max_age, |
332 | bool *dovacuum, bool *doanalyze, bool *wraparound); |
333 | |
334 | static void autovacuum_do_vac_analyze(autovac_table *tab, |
335 | BufferAccessStrategy bstrategy); |
336 | static AutoVacOpts *extract_autovac_opts(HeapTuple tup, |
337 | TupleDesc pg_class_desc); |
338 | static PgStat_StatTabEntry *get_pgstat_tabentry_relid(Oid relid, bool isshared, |
339 | PgStat_StatDBEntry *shared, |
340 | PgStat_StatDBEntry *dbentry); |
341 | static void perform_work_item(AutoVacuumWorkItem *workitem); |
342 | static void autovac_report_activity(autovac_table *tab); |
343 | static void autovac_report_workitem(AutoVacuumWorkItem *workitem, |
344 | const char *nspname, const char *relname); |
345 | static void av_sighup_handler(SIGNAL_ARGS); |
346 | static void avl_sigusr2_handler(SIGNAL_ARGS); |
347 | static void avl_sigterm_handler(SIGNAL_ARGS); |
348 | static void autovac_refresh_stats(void); |
349 | |
350 | |
351 | |
352 | /******************************************************************** |
353 | * AUTOVACUUM LAUNCHER CODE |
354 | ********************************************************************/ |
355 | |
356 | #ifdef EXEC_BACKEND |
357 | /* |
358 | * forkexec routine for the autovacuum launcher process. |
359 | * |
360 | * Format up the arglist, then fork and exec. |
361 | */ |
362 | static pid_t |
363 | avlauncher_forkexec(void) |
364 | { |
365 | char *av[10]; |
366 | int ac = 0; |
367 | |
368 | av[ac++] = "postgres" ; |
369 | av[ac++] = "--forkavlauncher" ; |
370 | av[ac++] = NULL; /* filled in by postmaster_forkexec */ |
371 | av[ac] = NULL; |
372 | |
373 | Assert(ac < lengthof(av)); |
374 | |
375 | return postmaster_forkexec(ac, av); |
376 | } |
377 | |
378 | /* |
379 | * We need this set from the outside, before InitProcess is called |
380 | */ |
381 | void |
382 | AutovacuumLauncherIAm(void) |
383 | { |
384 | am_autovacuum_launcher = true; |
385 | } |
386 | #endif |
387 | |
388 | /* |
389 | * Main entry point for autovacuum launcher process, to be called from the |
390 | * postmaster. |
391 | */ |
392 | int |
393 | StartAutoVacLauncher(void) |
394 | { |
395 | pid_t AutoVacPID; |
396 | |
397 | #ifdef EXEC_BACKEND |
398 | switch ((AutoVacPID = avlauncher_forkexec())) |
399 | #else |
400 | switch ((AutoVacPID = fork_process())) |
401 | #endif |
402 | { |
403 | case -1: |
404 | ereport(LOG, |
405 | (errmsg("could not fork autovacuum launcher process: %m" ))); |
406 | return 0; |
407 | |
408 | #ifndef EXEC_BACKEND |
409 | case 0: |
410 | /* in postmaster child ... */ |
411 | InitPostmasterChild(); |
412 | |
413 | /* Close the postmaster's sockets */ |
414 | ClosePostmasterPorts(false); |
415 | |
416 | AutoVacLauncherMain(0, NULL); |
417 | break; |
418 | #endif |
419 | default: |
420 | return (int) AutoVacPID; |
421 | } |
422 | |
423 | /* shouldn't get here */ |
424 | return 0; |
425 | } |
426 | |
427 | /* |
428 | * Main loop for the autovacuum launcher process. |
429 | */ |
430 | NON_EXEC_STATIC void |
431 | AutoVacLauncherMain(int argc, char *argv[]) |
432 | { |
433 | sigjmp_buf local_sigjmp_buf; |
434 | |
435 | am_autovacuum_launcher = true; |
436 | |
437 | /* Identify myself via ps */ |
438 | init_ps_display(pgstat_get_backend_desc(B_AUTOVAC_LAUNCHER), "" , "" , "" ); |
439 | |
440 | ereport(DEBUG1, |
441 | (errmsg("autovacuum launcher started" ))); |
442 | |
443 | if (PostAuthDelay) |
444 | pg_usleep(PostAuthDelay * 1000000L); |
445 | |
446 | SetProcessingMode(InitProcessing); |
447 | |
448 | /* |
449 | * Set up signal handlers. We operate on databases much like a regular |
450 | * backend, so we use the same signal handling. See equivalent code in |
451 | * tcop/postgres.c. |
452 | */ |
453 | pqsignal(SIGHUP, av_sighup_handler); |
454 | pqsignal(SIGINT, StatementCancelHandler); |
455 | pqsignal(SIGTERM, avl_sigterm_handler); |
456 | |
457 | pqsignal(SIGQUIT, quickdie); |
458 | InitializeTimeouts(); /* establishes SIGALRM handler */ |
459 | |
460 | pqsignal(SIGPIPE, SIG_IGN); |
461 | pqsignal(SIGUSR1, procsignal_sigusr1_handler); |
462 | pqsignal(SIGUSR2, avl_sigusr2_handler); |
463 | pqsignal(SIGFPE, FloatExceptionHandler); |
464 | pqsignal(SIGCHLD, SIG_DFL); |
465 | |
466 | /* Early initialization */ |
467 | BaseInit(); |
468 | |
469 | /* |
470 | * Create a per-backend PGPROC struct in shared memory, except in the |
471 | * EXEC_BACKEND case where this was done in SubPostmasterMain. We must do |
472 | * this before we can use LWLocks (and in the EXEC_BACKEND case we already |
473 | * had to do some stuff with LWLocks). |
474 | */ |
475 | #ifndef EXEC_BACKEND |
476 | InitProcess(); |
477 | #endif |
478 | |
479 | InitPostgres(NULL, InvalidOid, NULL, InvalidOid, NULL, false); |
480 | |
481 | SetProcessingMode(NormalProcessing); |
482 | |
483 | /* |
484 | * Create a memory context that we will do all our work in. We do this so |
485 | * that we can reset the context during error recovery and thereby avoid |
486 | * possible memory leaks. |
487 | */ |
488 | AutovacMemCxt = AllocSetContextCreate(TopMemoryContext, |
489 | "Autovacuum Launcher" , |
490 | ALLOCSET_DEFAULT_SIZES); |
491 | MemoryContextSwitchTo(AutovacMemCxt); |
492 | |
493 | /* |
494 | * If an exception is encountered, processing resumes here. |
495 | * |
496 | * This code is a stripped down version of PostgresMain error recovery. |
497 | */ |
498 | if (sigsetjmp(local_sigjmp_buf, 1) != 0) |
499 | { |
500 | /* since not using PG_TRY, must reset error stack by hand */ |
501 | error_context_stack = NULL; |
502 | |
503 | /* Prevents interrupts while cleaning up */ |
504 | HOLD_INTERRUPTS(); |
505 | |
506 | /* Forget any pending QueryCancel or timeout request */ |
507 | disable_all_timeouts(false); |
508 | QueryCancelPending = false; /* second to avoid race condition */ |
509 | |
510 | /* Report the error to the server log */ |
511 | EmitErrorReport(); |
512 | |
513 | /* Abort the current transaction in order to recover */ |
514 | AbortCurrentTransaction(); |
515 | |
516 | /* |
517 | * Release any other resources, for the case where we were not in a |
518 | * transaction. |
519 | */ |
520 | LWLockReleaseAll(); |
521 | pgstat_report_wait_end(); |
522 | AbortBufferIO(); |
523 | UnlockBuffers(); |
524 | /* this is probably dead code, but let's be safe: */ |
525 | if (AuxProcessResourceOwner) |
526 | ReleaseAuxProcessResources(false); |
527 | AtEOXact_Buffers(false); |
528 | AtEOXact_SMgr(); |
529 | AtEOXact_Files(false); |
530 | AtEOXact_HashTables(false); |
531 | |
532 | /* |
533 | * Now return to normal top-level context and clear ErrorContext for |
534 | * next time. |
535 | */ |
536 | MemoryContextSwitchTo(AutovacMemCxt); |
537 | FlushErrorState(); |
538 | |
539 | /* Flush any leaked data in the top-level context */ |
540 | MemoryContextResetAndDeleteChildren(AutovacMemCxt); |
541 | |
542 | /* don't leave dangling pointers to freed memory */ |
543 | DatabaseListCxt = NULL; |
544 | dlist_init(&DatabaseList); |
545 | |
546 | /* |
547 | * Make sure pgstat also considers our stat data as gone. Note: we |
548 | * mustn't use autovac_refresh_stats here. |
549 | */ |
550 | pgstat_clear_snapshot(); |
551 | |
552 | /* Now we can allow interrupts again */ |
553 | RESUME_INTERRUPTS(); |
554 | |
555 | /* if in shutdown mode, no need for anything further; just go away */ |
556 | if (got_SIGTERM) |
557 | goto shutdown; |
558 | |
559 | /* |
560 | * Sleep at least 1 second after any error. We don't want to be |
561 | * filling the error logs as fast as we can. |
562 | */ |
563 | pg_usleep(1000000L); |
564 | } |
565 | |
566 | /* We can now handle ereport(ERROR) */ |
567 | PG_exception_stack = &local_sigjmp_buf; |
568 | |
569 | /* must unblock signals before calling rebuild_database_list */ |
570 | PG_SETMASK(&UnBlockSig); |
571 | |
572 | /* |
573 | * Set always-secure search path. Launcher doesn't connect to a database, |
574 | * so this has no effect. |
575 | */ |
576 | SetConfigOption("search_path" , "" , PGC_SUSET, PGC_S_OVERRIDE); |
577 | |
578 | /* |
579 | * Force zero_damaged_pages OFF in the autovac process, even if it is set |
580 | * in postgresql.conf. We don't really want such a dangerous option being |
581 | * applied non-interactively. |
582 | */ |
583 | SetConfigOption("zero_damaged_pages" , "false" , PGC_SUSET, PGC_S_OVERRIDE); |
584 | |
585 | /* |
586 | * Force settable timeouts off to avoid letting these settings prevent |
587 | * regular maintenance from being executed. |
588 | */ |
589 | SetConfigOption("statement_timeout" , "0" , PGC_SUSET, PGC_S_OVERRIDE); |
590 | SetConfigOption("lock_timeout" , "0" , PGC_SUSET, PGC_S_OVERRIDE); |
591 | SetConfigOption("idle_in_transaction_session_timeout" , "0" , |
592 | PGC_SUSET, PGC_S_OVERRIDE); |
593 | |
594 | /* |
595 | * Force default_transaction_isolation to READ COMMITTED. We don't want |
596 | * to pay the overhead of serializable mode, nor add any risk of causing |
597 | * deadlocks or delaying other transactions. |
598 | */ |
599 | SetConfigOption("default_transaction_isolation" , "read committed" , |
600 | PGC_SUSET, PGC_S_OVERRIDE); |
601 | |
602 | /* |
603 | * In emergency mode, just start a worker (unless shutdown was requested) |
604 | * and go away. |
605 | */ |
606 | if (!AutoVacuumingActive()) |
607 | { |
608 | if (!got_SIGTERM) |
609 | do_start_worker(); |
610 | proc_exit(0); /* done */ |
611 | } |
612 | |
613 | AutoVacuumShmem->av_launcherpid = MyProcPid; |
614 | |
615 | /* |
616 | * Create the initial database list. The invariant we want this list to |
617 | * keep is that it's ordered by decreasing next_time. As soon as an entry |
618 | * is updated to a higher time, it will be moved to the front (which is |
619 | * correct because the only operation is to add autovacuum_naptime to the |
620 | * entry, and time always increases). |
621 | */ |
622 | rebuild_database_list(InvalidOid); |
623 | |
624 | /* loop until shutdown request */ |
625 | while (!got_SIGTERM) |
626 | { |
627 | struct timeval nap; |
628 | TimestampTz current_time = 0; |
629 | bool can_launch; |
630 | |
631 | /* |
632 | * This loop is a bit different from the normal use of WaitLatch, |
633 | * because we'd like to sleep before the first launch of a child |
634 | * process. So it's WaitLatch, then ResetLatch, then check for |
635 | * wakening conditions. |
636 | */ |
637 | |
638 | launcher_determine_sleep(!dlist_is_empty(&AutoVacuumShmem->av_freeWorkers), |
639 | false, &nap); |
640 | |
641 | /* |
642 | * Wait until naptime expires or we get some type of signal (all the |
643 | * signal handlers will wake us by calling SetLatch). |
644 | */ |
645 | (void) WaitLatch(MyLatch, |
646 | WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH, |
647 | (nap.tv_sec * 1000L) + (nap.tv_usec / 1000L), |
648 | WAIT_EVENT_AUTOVACUUM_MAIN); |
649 | |
650 | ResetLatch(MyLatch); |
651 | |
652 | /* Process sinval catchup interrupts that happened while sleeping */ |
653 | ProcessCatchupInterrupt(); |
654 | |
655 | /* the normal shutdown case */ |
656 | if (got_SIGTERM) |
657 | break; |
658 | |
659 | if (got_SIGHUP) |
660 | { |
661 | got_SIGHUP = false; |
662 | ProcessConfigFile(PGC_SIGHUP); |
663 | |
664 | /* shutdown requested in config file? */ |
665 | if (!AutoVacuumingActive()) |
666 | break; |
667 | |
668 | /* rebalance in case the default cost parameters changed */ |
669 | LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE); |
670 | autovac_balance_cost(); |
671 | LWLockRelease(AutovacuumLock); |
672 | |
673 | /* rebuild the list in case the naptime changed */ |
674 | rebuild_database_list(InvalidOid); |
675 | } |
676 | |
677 | /* |
678 | * a worker finished, or postmaster signalled failure to start a |
679 | * worker |
680 | */ |
681 | if (got_SIGUSR2) |
682 | { |
683 | got_SIGUSR2 = false; |
684 | |
685 | /* rebalance cost limits, if needed */ |
686 | if (AutoVacuumShmem->av_signal[AutoVacRebalance]) |
687 | { |
688 | LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE); |
689 | AutoVacuumShmem->av_signal[AutoVacRebalance] = false; |
690 | autovac_balance_cost(); |
691 | LWLockRelease(AutovacuumLock); |
692 | } |
693 | |
694 | if (AutoVacuumShmem->av_signal[AutoVacForkFailed]) |
695 | { |
696 | /* |
697 | * If the postmaster failed to start a new worker, we sleep |
698 | * for a little while and resend the signal. The new worker's |
699 | * state is still in memory, so this is sufficient. After |
700 | * that, we restart the main loop. |
701 | * |
702 | * XXX should we put a limit to the number of times we retry? |
703 | * I don't think it makes much sense, because a future start |
704 | * of a worker will continue to fail in the same way. |
705 | */ |
706 | AutoVacuumShmem->av_signal[AutoVacForkFailed] = false; |
707 | pg_usleep(1000000L); /* 1s */ |
708 | SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER); |
709 | continue; |
710 | } |
711 | } |
712 | |
713 | /* |
714 | * There are some conditions that we need to check before trying to |
715 | * start a worker. First, we need to make sure that there is a worker |
716 | * slot available. Second, we need to make sure that no other worker |
717 | * failed while starting up. |
718 | */ |
719 | |
720 | current_time = GetCurrentTimestamp(); |
721 | LWLockAcquire(AutovacuumLock, LW_SHARED); |
722 | |
723 | can_launch = !dlist_is_empty(&AutoVacuumShmem->av_freeWorkers); |
724 | |
725 | if (AutoVacuumShmem->av_startingWorker != NULL) |
726 | { |
727 | int waittime; |
728 | WorkerInfo worker = AutoVacuumShmem->av_startingWorker; |
729 | |
730 | /* |
731 | * We can't launch another worker when another one is still |
732 | * starting up (or failed while doing so), so just sleep for a bit |
733 | * more; that worker will wake us up again as soon as it's ready. |
734 | * We will only wait autovacuum_naptime seconds (up to a maximum |
735 | * of 60 seconds) for this to happen however. Note that failure |
736 | * to connect to a particular database is not a problem here, |
737 | * because the worker removes itself from the startingWorker |
738 | * pointer before trying to connect. Problems detected by the |
739 | * postmaster (like fork() failure) are also reported and handled |
740 | * differently. The only problems that may cause this code to |
741 | * fire are errors in the earlier sections of AutoVacWorkerMain, |
742 | * before the worker removes the WorkerInfo from the |
743 | * startingWorker pointer. |
744 | */ |
745 | waittime = Min(autovacuum_naptime, 60) * 1000; |
746 | if (TimestampDifferenceExceeds(worker->wi_launchtime, current_time, |
747 | waittime)) |
748 | { |
749 | LWLockRelease(AutovacuumLock); |
750 | LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE); |
751 | |
752 | /* |
753 | * No other process can put a worker in starting mode, so if |
754 | * startingWorker is still INVALID after exchanging our lock, |
755 | * we assume it's the same one we saw above (so we don't |
756 | * recheck the launch time). |
757 | */ |
758 | if (AutoVacuumShmem->av_startingWorker != NULL) |
759 | { |
760 | worker = AutoVacuumShmem->av_startingWorker; |
761 | worker->wi_dboid = InvalidOid; |
762 | worker->wi_tableoid = InvalidOid; |
763 | worker->wi_sharedrel = false; |
764 | worker->wi_proc = NULL; |
765 | worker->wi_launchtime = 0; |
766 | dlist_push_head(&AutoVacuumShmem->av_freeWorkers, |
767 | &worker->wi_links); |
768 | AutoVacuumShmem->av_startingWorker = NULL; |
769 | elog(WARNING, "worker took too long to start; canceled" ); |
770 | } |
771 | } |
772 | else |
773 | can_launch = false; |
774 | } |
775 | LWLockRelease(AutovacuumLock); /* either shared or exclusive */ |
776 | |
777 | /* if we can't do anything, just go back to sleep */ |
778 | if (!can_launch) |
779 | continue; |
780 | |
781 | /* We're OK to start a new worker */ |
782 | |
783 | if (dlist_is_empty(&DatabaseList)) |
784 | { |
785 | /* |
786 | * Special case when the list is empty: start a worker right away. |
787 | * This covers the initial case, when no database is in pgstats |
788 | * (thus the list is empty). Note that the constraints in |
789 | * launcher_determine_sleep keep us from starting workers too |
790 | * quickly (at most once every autovacuum_naptime when the list is |
791 | * empty). |
792 | */ |
793 | launch_worker(current_time); |
794 | } |
795 | else |
796 | { |
797 | /* |
798 | * because rebuild_database_list constructs a list with most |
799 | * distant adl_next_worker first, we obtain our database from the |
800 | * tail of the list. |
801 | */ |
802 | avl_dbase *avdb; |
803 | |
804 | avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList); |
805 | |
806 | /* |
807 | * launch a worker if next_worker is right now or it is in the |
808 | * past |
809 | */ |
810 | if (TimestampDifferenceExceeds(avdb->adl_next_worker, |
811 | current_time, 0)) |
812 | launch_worker(current_time); |
813 | } |
814 | } |
815 | |
816 | /* Normal exit from the autovac launcher is here */ |
817 | shutdown: |
818 | ereport(DEBUG1, |
819 | (errmsg("autovacuum launcher shutting down" ))); |
820 | AutoVacuumShmem->av_launcherpid = 0; |
821 | |
822 | proc_exit(0); /* done */ |
823 | } |
824 | |
825 | /* |
826 | * Determine the time to sleep, based on the database list. |
827 | * |
828 | * The "canlaunch" parameter indicates whether we can start a worker right now, |
829 | * for example due to the workers being all busy. If this is false, we will |
830 | * cause a long sleep, which will be interrupted when a worker exits. |
831 | */ |
832 | static void |
833 | launcher_determine_sleep(bool canlaunch, bool recursing, struct timeval *nap) |
834 | { |
835 | /* |
836 | * We sleep until the next scheduled vacuum. We trust that when the |
837 | * database list was built, care was taken so that no entries have times |
838 | * in the past; if the first entry has too close a next_worker value, or a |
839 | * time in the past, we will sleep a small nominal time. |
840 | */ |
841 | if (!canlaunch) |
842 | { |
843 | nap->tv_sec = autovacuum_naptime; |
844 | nap->tv_usec = 0; |
845 | } |
846 | else if (!dlist_is_empty(&DatabaseList)) |
847 | { |
848 | TimestampTz current_time = GetCurrentTimestamp(); |
849 | TimestampTz next_wakeup; |
850 | avl_dbase *avdb; |
851 | long secs; |
852 | int usecs; |
853 | |
854 | avdb = dlist_tail_element(avl_dbase, adl_node, &DatabaseList); |
855 | |
856 | next_wakeup = avdb->adl_next_worker; |
857 | TimestampDifference(current_time, next_wakeup, &secs, &usecs); |
858 | |
859 | nap->tv_sec = secs; |
860 | nap->tv_usec = usecs; |
861 | } |
862 | else |
863 | { |
864 | /* list is empty, sleep for whole autovacuum_naptime seconds */ |
865 | nap->tv_sec = autovacuum_naptime; |
866 | nap->tv_usec = 0; |
867 | } |
868 | |
869 | /* |
870 | * If the result is exactly zero, it means a database had an entry with |
871 | * time in the past. Rebuild the list so that the databases are evenly |
872 | * distributed again, and recalculate the time to sleep. This can happen |
873 | * if there are more tables needing vacuum than workers, and they all take |
874 | * longer to vacuum than autovacuum_naptime. |
875 | * |
876 | * We only recurse once. rebuild_database_list should always return times |
877 | * in the future, but it seems best not to trust too much on that. |
878 | */ |
879 | if (nap->tv_sec == 0 && nap->tv_usec == 0 && !recursing) |
880 | { |
881 | rebuild_database_list(InvalidOid); |
882 | launcher_determine_sleep(canlaunch, true, nap); |
883 | return; |
884 | } |
885 | |
886 | /* The smallest time we'll allow the launcher to sleep. */ |
887 | if (nap->tv_sec <= 0 && nap->tv_usec <= MIN_AUTOVAC_SLEEPTIME * 1000) |
888 | { |
889 | nap->tv_sec = 0; |
890 | nap->tv_usec = MIN_AUTOVAC_SLEEPTIME * 1000; |
891 | } |
892 | |
893 | /* |
894 | * If the sleep time is too large, clamp it to an arbitrary maximum (plus |
895 | * any fractional seconds, for simplicity). This avoids an essentially |
896 | * infinite sleep in strange cases like the system clock going backwards a |
897 | * few years. |
898 | */ |
899 | if (nap->tv_sec > MAX_AUTOVAC_SLEEPTIME) |
900 | nap->tv_sec = MAX_AUTOVAC_SLEEPTIME; |
901 | } |
902 | |
903 | /* |
904 | * Build an updated DatabaseList. It must only contain databases that appear |
905 | * in pgstats, and must be sorted by next_worker from highest to lowest, |
906 | * distributed regularly across the next autovacuum_naptime interval. |
907 | * |
908 | * Receives the Oid of the database that made this list be generated (we call |
909 | * this the "new" database, because when the database was already present on |
910 | * the list, we expect that this function is not called at all). The |
911 | * preexisting list, if any, will be used to preserve the order of the |
912 | * databases in the autovacuum_naptime period. The new database is put at the |
913 | * end of the interval. The actual values are not saved, which should not be |
914 | * much of a problem. |
915 | */ |
916 | static void |
917 | rebuild_database_list(Oid newdb) |
918 | { |
919 | List *dblist; |
920 | ListCell *cell; |
921 | MemoryContext newcxt; |
922 | MemoryContext oldcxt; |
923 | MemoryContext tmpcxt; |
924 | HASHCTL hctl; |
925 | int score; |
926 | int nelems; |
927 | HTAB *dbhash; |
928 | dlist_iter iter; |
929 | |
930 | /* use fresh stats */ |
931 | autovac_refresh_stats(); |
932 | |
933 | newcxt = AllocSetContextCreate(AutovacMemCxt, |
934 | "AV dblist" , |
935 | ALLOCSET_DEFAULT_SIZES); |
936 | tmpcxt = AllocSetContextCreate(newcxt, |
937 | "tmp AV dblist" , |
938 | ALLOCSET_DEFAULT_SIZES); |
939 | oldcxt = MemoryContextSwitchTo(tmpcxt); |
940 | |
941 | /* |
942 | * Implementing this is not as simple as it sounds, because we need to put |
943 | * the new database at the end of the list; next the databases that were |
944 | * already on the list, and finally (at the tail of the list) all the |
945 | * other databases that are not on the existing list. |
946 | * |
947 | * To do this, we build an empty hash table of scored databases. We will |
948 | * start with the lowest score (zero) for the new database, then |
949 | * increasing scores for the databases in the existing list, in order, and |
950 | * lastly increasing scores for all databases gotten via |
951 | * get_database_list() that are not already on the hash. |
952 | * |
953 | * Then we will put all the hash elements into an array, sort the array by |
954 | * score, and finally put the array elements into the new doubly linked |
955 | * list. |
956 | */ |
957 | hctl.keysize = sizeof(Oid); |
958 | hctl.entrysize = sizeof(avl_dbase); |
959 | hctl.hcxt = tmpcxt; |
960 | dbhash = hash_create("db hash" , 20, &hctl, /* magic number here FIXME */ |
961 | HASH_ELEM | HASH_BLOBS | HASH_CONTEXT); |
962 | |
963 | /* start by inserting the new database */ |
964 | score = 0; |
965 | if (OidIsValid(newdb)) |
966 | { |
967 | avl_dbase *db; |
968 | PgStat_StatDBEntry *entry; |
969 | |
970 | /* only consider this database if it has a pgstat entry */ |
971 | entry = pgstat_fetch_stat_dbentry(newdb); |
972 | if (entry != NULL) |
973 | { |
974 | /* we assume it isn't found because the hash was just created */ |
975 | db = hash_search(dbhash, &newdb, HASH_ENTER, NULL); |
976 | |
977 | /* hash_search already filled in the key */ |
978 | db->adl_score = score++; |
979 | /* next_worker is filled in later */ |
980 | } |
981 | } |
982 | |
983 | /* Now insert the databases from the existing list */ |
984 | dlist_foreach(iter, &DatabaseList) |
985 | { |
986 | avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur); |
987 | avl_dbase *db; |
988 | bool found; |
989 | PgStat_StatDBEntry *entry; |
990 | |
991 | /* |
992 | * skip databases with no stat entries -- in particular, this gets rid |
993 | * of dropped databases |
994 | */ |
995 | entry = pgstat_fetch_stat_dbentry(avdb->adl_datid); |
996 | if (entry == NULL) |
997 | continue; |
998 | |
999 | db = hash_search(dbhash, &(avdb->adl_datid), HASH_ENTER, &found); |
1000 | |
1001 | if (!found) |
1002 | { |
1003 | /* hash_search already filled in the key */ |
1004 | db->adl_score = score++; |
1005 | /* next_worker is filled in later */ |
1006 | } |
1007 | } |
1008 | |
1009 | /* finally, insert all qualifying databases not previously inserted */ |
1010 | dblist = get_database_list(); |
1011 | foreach(cell, dblist) |
1012 | { |
1013 | avw_dbase *avdb = lfirst(cell); |
1014 | avl_dbase *db; |
1015 | bool found; |
1016 | PgStat_StatDBEntry *entry; |
1017 | |
1018 | /* only consider databases with a pgstat entry */ |
1019 | entry = pgstat_fetch_stat_dbentry(avdb->adw_datid); |
1020 | if (entry == NULL) |
1021 | continue; |
1022 | |
1023 | db = hash_search(dbhash, &(avdb->adw_datid), HASH_ENTER, &found); |
1024 | /* only update the score if the database was not already on the hash */ |
1025 | if (!found) |
1026 | { |
1027 | /* hash_search already filled in the key */ |
1028 | db->adl_score = score++; |
1029 | /* next_worker is filled in later */ |
1030 | } |
1031 | } |
1032 | nelems = score; |
1033 | |
1034 | /* from here on, the allocated memory belongs to the new list */ |
1035 | MemoryContextSwitchTo(newcxt); |
1036 | dlist_init(&DatabaseList); |
1037 | |
1038 | if (nelems > 0) |
1039 | { |
1040 | TimestampTz current_time; |
1041 | int millis_increment; |
1042 | avl_dbase *dbary; |
1043 | avl_dbase *db; |
1044 | HASH_SEQ_STATUS seq; |
1045 | int i; |
1046 | |
1047 | /* put all the hash elements into an array */ |
1048 | dbary = palloc(nelems * sizeof(avl_dbase)); |
1049 | |
1050 | i = 0; |
1051 | hash_seq_init(&seq, dbhash); |
1052 | while ((db = hash_seq_search(&seq)) != NULL) |
1053 | memcpy(&(dbary[i++]), db, sizeof(avl_dbase)); |
1054 | |
1055 | /* sort the array */ |
1056 | qsort(dbary, nelems, sizeof(avl_dbase), db_comparator); |
1057 | |
1058 | /* |
1059 | * Determine the time interval between databases in the schedule. If |
1060 | * we see that the configured naptime would take us to sleep times |
1061 | * lower than our min sleep time (which launcher_determine_sleep is |
1062 | * coded not to allow), silently use a larger naptime (but don't touch |
1063 | * the GUC variable). |
1064 | */ |
1065 | millis_increment = 1000.0 * autovacuum_naptime / nelems; |
1066 | if (millis_increment <= MIN_AUTOVAC_SLEEPTIME) |
1067 | millis_increment = MIN_AUTOVAC_SLEEPTIME * 1.1; |
1068 | |
1069 | current_time = GetCurrentTimestamp(); |
1070 | |
1071 | /* |
1072 | * move the elements from the array into the dllist, setting the |
1073 | * next_worker while walking the array |
1074 | */ |
1075 | for (i = 0; i < nelems; i++) |
1076 | { |
1077 | avl_dbase *db = &(dbary[i]); |
1078 | |
1079 | current_time = TimestampTzPlusMilliseconds(current_time, |
1080 | millis_increment); |
1081 | db->adl_next_worker = current_time; |
1082 | |
1083 | /* later elements should go closer to the head of the list */ |
1084 | dlist_push_head(&DatabaseList, &db->adl_node); |
1085 | } |
1086 | } |
1087 | |
1088 | /* all done, clean up memory */ |
1089 | if (DatabaseListCxt != NULL) |
1090 | MemoryContextDelete(DatabaseListCxt); |
1091 | MemoryContextDelete(tmpcxt); |
1092 | DatabaseListCxt = newcxt; |
1093 | MemoryContextSwitchTo(oldcxt); |
1094 | } |
1095 | |
1096 | /* qsort comparator for avl_dbase, using adl_score */ |
1097 | static int |
1098 | db_comparator(const void *a, const void *b) |
1099 | { |
1100 | if (((const avl_dbase *) a)->adl_score == ((const avl_dbase *) b)->adl_score) |
1101 | return 0; |
1102 | else |
1103 | return (((const avl_dbase *) a)->adl_score < ((const avl_dbase *) b)->adl_score) ? 1 : -1; |
1104 | } |
1105 | |
1106 | /* |
1107 | * do_start_worker |
1108 | * |
1109 | * Bare-bones procedure for starting an autovacuum worker from the launcher. |
1110 | * It determines what database to work on, sets up shared memory stuff and |
1111 | * signals postmaster to start the worker. It fails gracefully if invoked when |
1112 | * autovacuum_workers are already active. |
1113 | * |
1114 | * Return value is the OID of the database that the worker is going to process, |
1115 | * or InvalidOid if no worker was actually started. |
1116 | */ |
1117 | static Oid |
1118 | do_start_worker(void) |
1119 | { |
1120 | List *dblist; |
1121 | ListCell *cell; |
1122 | TransactionId xidForceLimit; |
1123 | MultiXactId multiForceLimit; |
1124 | bool for_xid_wrap; |
1125 | bool for_multi_wrap; |
1126 | avw_dbase *avdb; |
1127 | TimestampTz current_time; |
1128 | bool skipit = false; |
1129 | Oid retval = InvalidOid; |
1130 | MemoryContext tmpcxt, |
1131 | oldcxt; |
1132 | |
1133 | /* return quickly when there are no free workers */ |
1134 | LWLockAcquire(AutovacuumLock, LW_SHARED); |
1135 | if (dlist_is_empty(&AutoVacuumShmem->av_freeWorkers)) |
1136 | { |
1137 | LWLockRelease(AutovacuumLock); |
1138 | return InvalidOid; |
1139 | } |
1140 | LWLockRelease(AutovacuumLock); |
1141 | |
1142 | /* |
1143 | * Create and switch to a temporary context to avoid leaking the memory |
1144 | * allocated for the database list. |
1145 | */ |
1146 | tmpcxt = AllocSetContextCreate(CurrentMemoryContext, |
1147 | "Start worker tmp cxt" , |
1148 | ALLOCSET_DEFAULT_SIZES); |
1149 | oldcxt = MemoryContextSwitchTo(tmpcxt); |
1150 | |
1151 | /* use fresh stats */ |
1152 | autovac_refresh_stats(); |
1153 | |
1154 | /* Get a list of databases */ |
1155 | dblist = get_database_list(); |
1156 | |
1157 | /* |
1158 | * Determine the oldest datfrozenxid/relfrozenxid that we will allow to |
1159 | * pass without forcing a vacuum. (This limit can be tightened for |
1160 | * particular tables, but not loosened.) |
1161 | */ |
1162 | recentXid = ReadNewTransactionId(); |
1163 | xidForceLimit = recentXid - autovacuum_freeze_max_age; |
1164 | /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */ |
1165 | /* this can cause the limit to go backwards by 3, but that's OK */ |
1166 | if (xidForceLimit < FirstNormalTransactionId) |
1167 | xidForceLimit -= FirstNormalTransactionId; |
1168 | |
1169 | /* Also determine the oldest datminmxid we will consider. */ |
1170 | recentMulti = ReadNextMultiXactId(); |
1171 | multiForceLimit = recentMulti - MultiXactMemberFreezeThreshold(); |
1172 | if (multiForceLimit < FirstMultiXactId) |
1173 | multiForceLimit -= FirstMultiXactId; |
1174 | |
1175 | /* |
1176 | * Choose a database to connect to. We pick the database that was least |
1177 | * recently auto-vacuumed, or one that needs vacuuming to prevent Xid |
1178 | * wraparound-related data loss. If any db at risk of Xid wraparound is |
1179 | * found, we pick the one with oldest datfrozenxid, independently of |
1180 | * autovacuum times; similarly we pick the one with the oldest datminmxid |
1181 | * if any is in MultiXactId wraparound. Note that those in Xid wraparound |
1182 | * danger are given more priority than those in multi wraparound danger. |
1183 | * |
1184 | * Note that a database with no stats entry is not considered, except for |
1185 | * Xid wraparound purposes. The theory is that if no one has ever |
1186 | * connected to it since the stats were last initialized, it doesn't need |
1187 | * vacuuming. |
1188 | * |
1189 | * XXX This could be improved if we had more info about whether it needs |
1190 | * vacuuming before connecting to it. Perhaps look through the pgstats |
1191 | * data for the database's tables? One idea is to keep track of the |
1192 | * number of new and dead tuples per database in pgstats. However it |
1193 | * isn't clear how to construct a metric that measures that and not cause |
1194 | * starvation for less busy databases. |
1195 | */ |
1196 | avdb = NULL; |
1197 | for_xid_wrap = false; |
1198 | for_multi_wrap = false; |
1199 | current_time = GetCurrentTimestamp(); |
1200 | foreach(cell, dblist) |
1201 | { |
1202 | avw_dbase *tmp = lfirst(cell); |
1203 | dlist_iter iter; |
1204 | |
1205 | /* Check to see if this one is at risk of wraparound */ |
1206 | if (TransactionIdPrecedes(tmp->adw_frozenxid, xidForceLimit)) |
1207 | { |
1208 | if (avdb == NULL || |
1209 | TransactionIdPrecedes(tmp->adw_frozenxid, |
1210 | avdb->adw_frozenxid)) |
1211 | avdb = tmp; |
1212 | for_xid_wrap = true; |
1213 | continue; |
1214 | } |
1215 | else if (for_xid_wrap) |
1216 | continue; /* ignore not-at-risk DBs */ |
1217 | else if (MultiXactIdPrecedes(tmp->adw_minmulti, multiForceLimit)) |
1218 | { |
1219 | if (avdb == NULL || |
1220 | MultiXactIdPrecedes(tmp->adw_minmulti, avdb->adw_minmulti)) |
1221 | avdb = tmp; |
1222 | for_multi_wrap = true; |
1223 | continue; |
1224 | } |
1225 | else if (for_multi_wrap) |
1226 | continue; /* ignore not-at-risk DBs */ |
1227 | |
1228 | /* Find pgstat entry if any */ |
1229 | tmp->adw_entry = pgstat_fetch_stat_dbentry(tmp->adw_datid); |
1230 | |
1231 | /* |
1232 | * Skip a database with no pgstat entry; it means it hasn't seen any |
1233 | * activity. |
1234 | */ |
1235 | if (!tmp->adw_entry) |
1236 | continue; |
1237 | |
1238 | /* |
1239 | * Also, skip a database that appears on the database list as having |
1240 | * been processed recently (less than autovacuum_naptime seconds ago). |
1241 | * We do this so that we don't select a database which we just |
1242 | * selected, but that pgstat hasn't gotten around to updating the last |
1243 | * autovacuum time yet. |
1244 | */ |
1245 | skipit = false; |
1246 | |
1247 | dlist_reverse_foreach(iter, &DatabaseList) |
1248 | { |
1249 | avl_dbase *dbp = dlist_container(avl_dbase, adl_node, iter.cur); |
1250 | |
1251 | if (dbp->adl_datid == tmp->adw_datid) |
1252 | { |
1253 | /* |
1254 | * Skip this database if its next_worker value falls between |
1255 | * the current time and the current time plus naptime. |
1256 | */ |
1257 | if (!TimestampDifferenceExceeds(dbp->adl_next_worker, |
1258 | current_time, 0) && |
1259 | !TimestampDifferenceExceeds(current_time, |
1260 | dbp->adl_next_worker, |
1261 | autovacuum_naptime * 1000)) |
1262 | skipit = true; |
1263 | |
1264 | break; |
1265 | } |
1266 | } |
1267 | if (skipit) |
1268 | continue; |
1269 | |
1270 | /* |
1271 | * Remember the db with oldest autovac time. (If we are here, both |
1272 | * tmp->entry and db->entry must be non-null.) |
1273 | */ |
1274 | if (avdb == NULL || |
1275 | tmp->adw_entry->last_autovac_time < avdb->adw_entry->last_autovac_time) |
1276 | avdb = tmp; |
1277 | } |
1278 | |
1279 | /* Found a database -- process it */ |
1280 | if (avdb != NULL) |
1281 | { |
1282 | WorkerInfo worker; |
1283 | dlist_node *wptr; |
1284 | |
1285 | LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE); |
1286 | |
1287 | /* |
1288 | * Get a worker entry from the freelist. We checked above, so there |
1289 | * really should be a free slot. |
1290 | */ |
1291 | wptr = dlist_pop_head_node(&AutoVacuumShmem->av_freeWorkers); |
1292 | |
1293 | worker = dlist_container(WorkerInfoData, wi_links, wptr); |
1294 | worker->wi_dboid = avdb->adw_datid; |
1295 | worker->wi_proc = NULL; |
1296 | worker->wi_launchtime = GetCurrentTimestamp(); |
1297 | |
1298 | AutoVacuumShmem->av_startingWorker = worker; |
1299 | |
1300 | LWLockRelease(AutovacuumLock); |
1301 | |
1302 | SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER); |
1303 | |
1304 | retval = avdb->adw_datid; |
1305 | } |
1306 | else if (skipit) |
1307 | { |
1308 | /* |
1309 | * If we skipped all databases on the list, rebuild it, because it |
1310 | * probably contains a dropped database. |
1311 | */ |
1312 | rebuild_database_list(InvalidOid); |
1313 | } |
1314 | |
1315 | MemoryContextSwitchTo(oldcxt); |
1316 | MemoryContextDelete(tmpcxt); |
1317 | |
1318 | return retval; |
1319 | } |
1320 | |
1321 | /* |
1322 | * launch_worker |
1323 | * |
1324 | * Wrapper for starting a worker from the launcher. Besides actually starting |
1325 | * it, update the database list to reflect the next time that another one will |
1326 | * need to be started on the selected database. The actual database choice is |
1327 | * left to do_start_worker. |
1328 | * |
1329 | * This routine is also expected to insert an entry into the database list if |
1330 | * the selected database was previously absent from the list. |
1331 | */ |
1332 | static void |
1333 | launch_worker(TimestampTz now) |
1334 | { |
1335 | Oid dbid; |
1336 | dlist_iter iter; |
1337 | |
1338 | dbid = do_start_worker(); |
1339 | if (OidIsValid(dbid)) |
1340 | { |
1341 | bool found = false; |
1342 | |
1343 | /* |
1344 | * Walk the database list and update the corresponding entry. If the |
1345 | * database is not on the list, we'll recreate the list. |
1346 | */ |
1347 | dlist_foreach(iter, &DatabaseList) |
1348 | { |
1349 | avl_dbase *avdb = dlist_container(avl_dbase, adl_node, iter.cur); |
1350 | |
1351 | if (avdb->adl_datid == dbid) |
1352 | { |
1353 | found = true; |
1354 | |
1355 | /* |
1356 | * add autovacuum_naptime seconds to the current time, and use |
1357 | * that as the new "next_worker" field for this database. |
1358 | */ |
1359 | avdb->adl_next_worker = |
1360 | TimestampTzPlusMilliseconds(now, autovacuum_naptime * 1000); |
1361 | |
1362 | dlist_move_head(&DatabaseList, iter.cur); |
1363 | break; |
1364 | } |
1365 | } |
1366 | |
1367 | /* |
1368 | * If the database was not present in the database list, we rebuild |
1369 | * the list. It's possible that the database does not get into the |
1370 | * list anyway, for example if it's a database that doesn't have a |
1371 | * pgstat entry, but this is not a problem because we don't want to |
1372 | * schedule workers regularly into those in any case. |
1373 | */ |
1374 | if (!found) |
1375 | rebuild_database_list(dbid); |
1376 | } |
1377 | } |
1378 | |
1379 | /* |
1380 | * Called from postmaster to signal a failure to fork a process to become |
1381 | * worker. The postmaster should kill(SIGUSR2) the launcher shortly |
1382 | * after calling this function. |
1383 | */ |
1384 | void |
1385 | AutoVacWorkerFailed(void) |
1386 | { |
1387 | AutoVacuumShmem->av_signal[AutoVacForkFailed] = true; |
1388 | } |
1389 | |
1390 | /* SIGHUP: set flag to re-read config file at next convenient time */ |
1391 | static void |
1392 | av_sighup_handler(SIGNAL_ARGS) |
1393 | { |
1394 | int save_errno = errno; |
1395 | |
1396 | got_SIGHUP = true; |
1397 | SetLatch(MyLatch); |
1398 | |
1399 | errno = save_errno; |
1400 | } |
1401 | |
1402 | /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */ |
1403 | static void |
1404 | avl_sigusr2_handler(SIGNAL_ARGS) |
1405 | { |
1406 | int save_errno = errno; |
1407 | |
1408 | got_SIGUSR2 = true; |
1409 | SetLatch(MyLatch); |
1410 | |
1411 | errno = save_errno; |
1412 | } |
1413 | |
1414 | /* SIGTERM: time to die */ |
1415 | static void |
1416 | avl_sigterm_handler(SIGNAL_ARGS) |
1417 | { |
1418 | int save_errno = errno; |
1419 | |
1420 | got_SIGTERM = true; |
1421 | SetLatch(MyLatch); |
1422 | |
1423 | errno = save_errno; |
1424 | } |
1425 | |
1426 | |
1427 | /******************************************************************** |
1428 | * AUTOVACUUM WORKER CODE |
1429 | ********************************************************************/ |
1430 | |
1431 | #ifdef EXEC_BACKEND |
1432 | /* |
1433 | * forkexec routines for the autovacuum worker. |
1434 | * |
1435 | * Format up the arglist, then fork and exec. |
1436 | */ |
1437 | static pid_t |
1438 | avworker_forkexec(void) |
1439 | { |
1440 | char *av[10]; |
1441 | int ac = 0; |
1442 | |
1443 | av[ac++] = "postgres" ; |
1444 | av[ac++] = "--forkavworker" ; |
1445 | av[ac++] = NULL; /* filled in by postmaster_forkexec */ |
1446 | av[ac] = NULL; |
1447 | |
1448 | Assert(ac < lengthof(av)); |
1449 | |
1450 | return postmaster_forkexec(ac, av); |
1451 | } |
1452 | |
1453 | /* |
1454 | * We need this set from the outside, before InitProcess is called |
1455 | */ |
1456 | void |
1457 | AutovacuumWorkerIAm(void) |
1458 | { |
1459 | am_autovacuum_worker = true; |
1460 | } |
1461 | #endif |
1462 | |
1463 | /* |
1464 | * Main entry point for autovacuum worker process. |
1465 | * |
1466 | * This code is heavily based on pgarch.c, q.v. |
1467 | */ |
1468 | int |
1469 | StartAutoVacWorker(void) |
1470 | { |
1471 | pid_t worker_pid; |
1472 | |
1473 | #ifdef EXEC_BACKEND |
1474 | switch ((worker_pid = avworker_forkexec())) |
1475 | #else |
1476 | switch ((worker_pid = fork_process())) |
1477 | #endif |
1478 | { |
1479 | case -1: |
1480 | ereport(LOG, |
1481 | (errmsg("could not fork autovacuum worker process: %m" ))); |
1482 | return 0; |
1483 | |
1484 | #ifndef EXEC_BACKEND |
1485 | case 0: |
1486 | /* in postmaster child ... */ |
1487 | InitPostmasterChild(); |
1488 | |
1489 | /* Close the postmaster's sockets */ |
1490 | ClosePostmasterPorts(false); |
1491 | |
1492 | AutoVacWorkerMain(0, NULL); |
1493 | break; |
1494 | #endif |
1495 | default: |
1496 | return (int) worker_pid; |
1497 | } |
1498 | |
1499 | /* shouldn't get here */ |
1500 | return 0; |
1501 | } |
1502 | |
1503 | /* |
1504 | * AutoVacWorkerMain |
1505 | */ |
1506 | NON_EXEC_STATIC void |
1507 | AutoVacWorkerMain(int argc, char *argv[]) |
1508 | { |
1509 | sigjmp_buf local_sigjmp_buf; |
1510 | Oid dbid; |
1511 | |
1512 | am_autovacuum_worker = true; |
1513 | |
1514 | /* Identify myself via ps */ |
1515 | init_ps_display(pgstat_get_backend_desc(B_AUTOVAC_WORKER), "" , "" , "" ); |
1516 | |
1517 | SetProcessingMode(InitProcessing); |
1518 | |
1519 | /* |
1520 | * Set up signal handlers. We operate on databases much like a regular |
1521 | * backend, so we use the same signal handling. See equivalent code in |
1522 | * tcop/postgres.c. |
1523 | */ |
1524 | pqsignal(SIGHUP, av_sighup_handler); |
1525 | |
1526 | /* |
1527 | * SIGINT is used to signal canceling the current table's vacuum; SIGTERM |
1528 | * means abort and exit cleanly, and SIGQUIT means abandon ship. |
1529 | */ |
1530 | pqsignal(SIGINT, StatementCancelHandler); |
1531 | pqsignal(SIGTERM, die); |
1532 | pqsignal(SIGQUIT, quickdie); |
1533 | InitializeTimeouts(); /* establishes SIGALRM handler */ |
1534 | |
1535 | pqsignal(SIGPIPE, SIG_IGN); |
1536 | pqsignal(SIGUSR1, procsignal_sigusr1_handler); |
1537 | pqsignal(SIGUSR2, SIG_IGN); |
1538 | pqsignal(SIGFPE, FloatExceptionHandler); |
1539 | pqsignal(SIGCHLD, SIG_DFL); |
1540 | |
1541 | /* Early initialization */ |
1542 | BaseInit(); |
1543 | |
1544 | /* |
1545 | * Create a per-backend PGPROC struct in shared memory, except in the |
1546 | * EXEC_BACKEND case where this was done in SubPostmasterMain. We must do |
1547 | * this before we can use LWLocks (and in the EXEC_BACKEND case we already |
1548 | * had to do some stuff with LWLocks). |
1549 | */ |
1550 | #ifndef EXEC_BACKEND |
1551 | InitProcess(); |
1552 | #endif |
1553 | |
1554 | /* |
1555 | * If an exception is encountered, processing resumes here. |
1556 | * |
1557 | * See notes in postgres.c about the design of this coding. |
1558 | */ |
1559 | if (sigsetjmp(local_sigjmp_buf, 1) != 0) |
1560 | { |
1561 | /* Prevents interrupts while cleaning up */ |
1562 | HOLD_INTERRUPTS(); |
1563 | |
1564 | /* Report the error to the server log */ |
1565 | EmitErrorReport(); |
1566 | |
1567 | /* |
1568 | * We can now go away. Note that because we called InitProcess, a |
1569 | * callback was registered to do ProcKill, which will clean up |
1570 | * necessary state. |
1571 | */ |
1572 | proc_exit(0); |
1573 | } |
1574 | |
1575 | /* We can now handle ereport(ERROR) */ |
1576 | PG_exception_stack = &local_sigjmp_buf; |
1577 | |
1578 | PG_SETMASK(&UnBlockSig); |
1579 | |
1580 | /* |
1581 | * Set always-secure search path, so malicious users can't redirect user |
1582 | * code (e.g. pg_index.indexprs). (That code runs in a |
1583 | * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not |
1584 | * take control of the entire autovacuum worker in any case.) |
1585 | */ |
1586 | SetConfigOption("search_path" , "" , PGC_SUSET, PGC_S_OVERRIDE); |
1587 | |
1588 | /* |
1589 | * Force zero_damaged_pages OFF in the autovac process, even if it is set |
1590 | * in postgresql.conf. We don't really want such a dangerous option being |
1591 | * applied non-interactively. |
1592 | */ |
1593 | SetConfigOption("zero_damaged_pages" , "false" , PGC_SUSET, PGC_S_OVERRIDE); |
1594 | |
1595 | /* |
1596 | * Force settable timeouts off to avoid letting these settings prevent |
1597 | * regular maintenance from being executed. |
1598 | */ |
1599 | SetConfigOption("statement_timeout" , "0" , PGC_SUSET, PGC_S_OVERRIDE); |
1600 | SetConfigOption("lock_timeout" , "0" , PGC_SUSET, PGC_S_OVERRIDE); |
1601 | SetConfigOption("idle_in_transaction_session_timeout" , "0" , |
1602 | PGC_SUSET, PGC_S_OVERRIDE); |
1603 | |
1604 | /* |
1605 | * Force default_transaction_isolation to READ COMMITTED. We don't want |
1606 | * to pay the overhead of serializable mode, nor add any risk of causing |
1607 | * deadlocks or delaying other transactions. |
1608 | */ |
1609 | SetConfigOption("default_transaction_isolation" , "read committed" , |
1610 | PGC_SUSET, PGC_S_OVERRIDE); |
1611 | |
1612 | /* |
1613 | * Force synchronous replication off to allow regular maintenance even if |
1614 | * we are waiting for standbys to connect. This is important to ensure we |
1615 | * aren't blocked from performing anti-wraparound tasks. |
1616 | */ |
1617 | if (synchronous_commit > SYNCHRONOUS_COMMIT_LOCAL_FLUSH) |
1618 | SetConfigOption("synchronous_commit" , "local" , |
1619 | PGC_SUSET, PGC_S_OVERRIDE); |
1620 | |
1621 | /* |
1622 | * Get the info about the database we're going to work on. |
1623 | */ |
1624 | LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE); |
1625 | |
1626 | /* |
1627 | * beware of startingWorker being INVALID; this should normally not |
1628 | * happen, but if a worker fails after forking and before this, the |
1629 | * launcher might have decided to remove it from the queue and start |
1630 | * again. |
1631 | */ |
1632 | if (AutoVacuumShmem->av_startingWorker != NULL) |
1633 | { |
1634 | MyWorkerInfo = AutoVacuumShmem->av_startingWorker; |
1635 | dbid = MyWorkerInfo->wi_dboid; |
1636 | MyWorkerInfo->wi_proc = MyProc; |
1637 | |
1638 | /* insert into the running list */ |
1639 | dlist_push_head(&AutoVacuumShmem->av_runningWorkers, |
1640 | &MyWorkerInfo->wi_links); |
1641 | |
1642 | /* |
1643 | * remove from the "starting" pointer, so that the launcher can start |
1644 | * a new worker if required |
1645 | */ |
1646 | AutoVacuumShmem->av_startingWorker = NULL; |
1647 | LWLockRelease(AutovacuumLock); |
1648 | |
1649 | on_shmem_exit(FreeWorkerInfo, 0); |
1650 | |
1651 | /* wake up the launcher */ |
1652 | if (AutoVacuumShmem->av_launcherpid != 0) |
1653 | kill(AutoVacuumShmem->av_launcherpid, SIGUSR2); |
1654 | } |
1655 | else |
1656 | { |
1657 | /* no worker entry for me, go away */ |
1658 | elog(WARNING, "autovacuum worker started without a worker entry" ); |
1659 | dbid = InvalidOid; |
1660 | LWLockRelease(AutovacuumLock); |
1661 | } |
1662 | |
1663 | if (OidIsValid(dbid)) |
1664 | { |
1665 | char dbname[NAMEDATALEN]; |
1666 | |
1667 | /* |
1668 | * Report autovac startup to the stats collector. We deliberately do |
1669 | * this before InitPostgres, so that the last_autovac_time will get |
1670 | * updated even if the connection attempt fails. This is to prevent |
1671 | * autovac from getting "stuck" repeatedly selecting an unopenable |
1672 | * database, rather than making any progress on stuff it can connect |
1673 | * to. |
1674 | */ |
1675 | pgstat_report_autovac(dbid); |
1676 | |
1677 | /* |
1678 | * Connect to the selected database |
1679 | * |
1680 | * Note: if we have selected a just-deleted database (due to using |
1681 | * stale stats info), we'll fail and exit here. |
1682 | */ |
1683 | InitPostgres(NULL, dbid, NULL, InvalidOid, dbname, false); |
1684 | SetProcessingMode(NormalProcessing); |
1685 | set_ps_display(dbname, false); |
1686 | ereport(DEBUG1, |
1687 | (errmsg("autovacuum: processing database \"%s\"" , dbname))); |
1688 | |
1689 | if (PostAuthDelay) |
1690 | pg_usleep(PostAuthDelay * 1000000L); |
1691 | |
1692 | /* And do an appropriate amount of work */ |
1693 | recentXid = ReadNewTransactionId(); |
1694 | recentMulti = ReadNextMultiXactId(); |
1695 | do_autovacuum(); |
1696 | } |
1697 | |
1698 | /* |
1699 | * The launcher will be notified of my death in ProcKill, *if* we managed |
1700 | * to get a worker slot at all |
1701 | */ |
1702 | |
1703 | /* All done, go away */ |
1704 | proc_exit(0); |
1705 | } |
1706 | |
1707 | /* |
1708 | * Return a WorkerInfo to the free list |
1709 | */ |
1710 | static void |
1711 | FreeWorkerInfo(int code, Datum arg) |
1712 | { |
1713 | if (MyWorkerInfo != NULL) |
1714 | { |
1715 | LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE); |
1716 | |
1717 | /* |
1718 | * Wake the launcher up so that he can launch a new worker immediately |
1719 | * if required. We only save the launcher's PID in local memory here; |
1720 | * the actual signal will be sent when the PGPROC is recycled. Note |
1721 | * that we always do this, so that the launcher can rebalance the cost |
1722 | * limit setting of the remaining workers. |
1723 | * |
1724 | * We somewhat ignore the risk that the launcher changes its PID |
1725 | * between us reading it and the actual kill; we expect ProcKill to be |
1726 | * called shortly after us, and we assume that PIDs are not reused too |
1727 | * quickly after a process exits. |
1728 | */ |
1729 | AutovacuumLauncherPid = AutoVacuumShmem->av_launcherpid; |
1730 | |
1731 | dlist_delete(&MyWorkerInfo->wi_links); |
1732 | MyWorkerInfo->wi_dboid = InvalidOid; |
1733 | MyWorkerInfo->wi_tableoid = InvalidOid; |
1734 | MyWorkerInfo->wi_sharedrel = false; |
1735 | MyWorkerInfo->wi_proc = NULL; |
1736 | MyWorkerInfo->wi_launchtime = 0; |
1737 | MyWorkerInfo->wi_dobalance = false; |
1738 | MyWorkerInfo->wi_cost_delay = 0; |
1739 | MyWorkerInfo->wi_cost_limit = 0; |
1740 | MyWorkerInfo->wi_cost_limit_base = 0; |
1741 | dlist_push_head(&AutoVacuumShmem->av_freeWorkers, |
1742 | &MyWorkerInfo->wi_links); |
1743 | /* not mine anymore */ |
1744 | MyWorkerInfo = NULL; |
1745 | |
1746 | /* |
1747 | * now that we're inactive, cause a rebalancing of the surviving |
1748 | * workers |
1749 | */ |
1750 | AutoVacuumShmem->av_signal[AutoVacRebalance] = true; |
1751 | LWLockRelease(AutovacuumLock); |
1752 | } |
1753 | } |
1754 | |
1755 | /* |
1756 | * Update the cost-based delay parameters, so that multiple workers consume |
1757 | * each a fraction of the total available I/O. |
1758 | */ |
1759 | void |
1760 | AutoVacuumUpdateDelay(void) |
1761 | { |
1762 | if (MyWorkerInfo) |
1763 | { |
1764 | VacuumCostDelay = MyWorkerInfo->wi_cost_delay; |
1765 | VacuumCostLimit = MyWorkerInfo->wi_cost_limit; |
1766 | } |
1767 | } |
1768 | |
1769 | /* |
1770 | * autovac_balance_cost |
1771 | * Recalculate the cost limit setting for each active worker. |
1772 | * |
1773 | * Caller must hold the AutovacuumLock in exclusive mode. |
1774 | */ |
1775 | static void |
1776 | autovac_balance_cost(void) |
1777 | { |
1778 | /* |
1779 | * The idea here is that we ration out I/O equally. The amount of I/O |
1780 | * that a worker can consume is determined by cost_limit/cost_delay, so we |
1781 | * try to equalize those ratios rather than the raw limit settings. |
1782 | * |
1783 | * note: in cost_limit, zero also means use value from elsewhere, because |
1784 | * zero is not a valid value. |
1785 | */ |
1786 | int vac_cost_limit = (autovacuum_vac_cost_limit > 0 ? |
1787 | autovacuum_vac_cost_limit : VacuumCostLimit); |
1788 | double vac_cost_delay = (autovacuum_vac_cost_delay >= 0 ? |
1789 | autovacuum_vac_cost_delay : VacuumCostDelay); |
1790 | double cost_total; |
1791 | double cost_avail; |
1792 | dlist_iter iter; |
1793 | |
1794 | /* not set? nothing to do */ |
1795 | if (vac_cost_limit <= 0 || vac_cost_delay <= 0) |
1796 | return; |
1797 | |
1798 | /* calculate the total base cost limit of participating active workers */ |
1799 | cost_total = 0.0; |
1800 | dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers) |
1801 | { |
1802 | WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur); |
1803 | |
1804 | if (worker->wi_proc != NULL && |
1805 | worker->wi_dobalance && |
1806 | worker->wi_cost_limit_base > 0 && worker->wi_cost_delay > 0) |
1807 | cost_total += |
1808 | (double) worker->wi_cost_limit_base / worker->wi_cost_delay; |
1809 | } |
1810 | |
1811 | /* there are no cost limits -- nothing to do */ |
1812 | if (cost_total <= 0) |
1813 | return; |
1814 | |
1815 | /* |
1816 | * Adjust cost limit of each active worker to balance the total of cost |
1817 | * limit to autovacuum_vacuum_cost_limit. |
1818 | */ |
1819 | cost_avail = (double) vac_cost_limit / vac_cost_delay; |
1820 | dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers) |
1821 | { |
1822 | WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur); |
1823 | |
1824 | if (worker->wi_proc != NULL && |
1825 | worker->wi_dobalance && |
1826 | worker->wi_cost_limit_base > 0 && worker->wi_cost_delay > 0) |
1827 | { |
1828 | int limit = (int) |
1829 | (cost_avail * worker->wi_cost_limit_base / cost_total); |
1830 | |
1831 | /* |
1832 | * We put a lower bound of 1 on the cost_limit, to avoid division- |
1833 | * by-zero in the vacuum code. Also, in case of roundoff trouble |
1834 | * in these calculations, let's be sure we don't ever set |
1835 | * cost_limit to more than the base value. |
1836 | */ |
1837 | worker->wi_cost_limit = Max(Min(limit, |
1838 | worker->wi_cost_limit_base), |
1839 | 1); |
1840 | } |
1841 | |
1842 | if (worker->wi_proc != NULL) |
1843 | elog(DEBUG2, "autovac_balance_cost(pid=%u db=%u, rel=%u, dobalance=%s cost_limit=%d, cost_limit_base=%d, cost_delay=%g)" , |
1844 | worker->wi_proc->pid, worker->wi_dboid, worker->wi_tableoid, |
1845 | worker->wi_dobalance ? "yes" : "no" , |
1846 | worker->wi_cost_limit, worker->wi_cost_limit_base, |
1847 | worker->wi_cost_delay); |
1848 | } |
1849 | } |
1850 | |
1851 | /* |
1852 | * get_database_list |
1853 | * Return a list of all databases found in pg_database. |
1854 | * |
1855 | * The list and associated data is allocated in the caller's memory context, |
1856 | * which is in charge of ensuring that it's properly cleaned up afterwards. |
1857 | * |
1858 | * Note: this is the only function in which the autovacuum launcher uses a |
1859 | * transaction. Although we aren't attached to any particular database and |
1860 | * therefore can't access most catalogs, we do have enough infrastructure |
1861 | * to do a seqscan on pg_database. |
1862 | */ |
1863 | static List * |
1864 | get_database_list(void) |
1865 | { |
1866 | List *dblist = NIL; |
1867 | Relation rel; |
1868 | TableScanDesc scan; |
1869 | HeapTuple tup; |
1870 | MemoryContext resultcxt; |
1871 | |
1872 | /* This is the context that we will allocate our output data in */ |
1873 | resultcxt = CurrentMemoryContext; |
1874 | |
1875 | /* |
1876 | * Start a transaction so we can access pg_database, and get a snapshot. |
1877 | * We don't have a use for the snapshot itself, but we're interested in |
1878 | * the secondary effect that it sets RecentGlobalXmin. (This is critical |
1879 | * for anything that reads heap pages, because HOT may decide to prune |
1880 | * them even if the process doesn't attempt to modify any tuples.) |
1881 | */ |
1882 | StartTransactionCommand(); |
1883 | (void) GetTransactionSnapshot(); |
1884 | |
1885 | rel = table_open(DatabaseRelationId, AccessShareLock); |
1886 | scan = table_beginscan_catalog(rel, 0, NULL); |
1887 | |
1888 | while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection))) |
1889 | { |
1890 | Form_pg_database pgdatabase = (Form_pg_database) GETSTRUCT(tup); |
1891 | avw_dbase *avdb; |
1892 | MemoryContext oldcxt; |
1893 | |
1894 | /* |
1895 | * Allocate our results in the caller's context, not the |
1896 | * transaction's. We do this inside the loop, and restore the original |
1897 | * context at the end, so that leaky things like heap_getnext() are |
1898 | * not called in a potentially long-lived context. |
1899 | */ |
1900 | oldcxt = MemoryContextSwitchTo(resultcxt); |
1901 | |
1902 | avdb = (avw_dbase *) palloc(sizeof(avw_dbase)); |
1903 | |
1904 | avdb->adw_datid = pgdatabase->oid; |
1905 | avdb->adw_name = pstrdup(NameStr(pgdatabase->datname)); |
1906 | avdb->adw_frozenxid = pgdatabase->datfrozenxid; |
1907 | avdb->adw_minmulti = pgdatabase->datminmxid; |
1908 | /* this gets set later: */ |
1909 | avdb->adw_entry = NULL; |
1910 | |
1911 | dblist = lappend(dblist, avdb); |
1912 | MemoryContextSwitchTo(oldcxt); |
1913 | } |
1914 | |
1915 | table_endscan(scan); |
1916 | table_close(rel, AccessShareLock); |
1917 | |
1918 | CommitTransactionCommand(); |
1919 | |
1920 | return dblist; |
1921 | } |
1922 | |
1923 | /* |
1924 | * Process a database table-by-table |
1925 | * |
1926 | * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in |
1927 | * order not to ignore shutdown commands for too long. |
1928 | */ |
1929 | static void |
1930 | do_autovacuum(void) |
1931 | { |
1932 | Relation classRel; |
1933 | HeapTuple tuple; |
1934 | TableScanDesc relScan; |
1935 | Form_pg_database dbForm; |
1936 | List *table_oids = NIL; |
1937 | List *orphan_oids = NIL; |
1938 | HASHCTL ctl; |
1939 | HTAB *table_toast_map; |
1940 | ListCell *volatile cell; |
1941 | PgStat_StatDBEntry *shared; |
1942 | PgStat_StatDBEntry *dbentry; |
1943 | BufferAccessStrategy bstrategy; |
1944 | ScanKeyData key; |
1945 | TupleDesc pg_class_desc; |
1946 | int effective_multixact_freeze_max_age; |
1947 | bool did_vacuum = false; |
1948 | bool found_concurrent_worker = false; |
1949 | int i; |
1950 | |
1951 | /* |
1952 | * StartTransactionCommand and CommitTransactionCommand will automatically |
1953 | * switch to other contexts. We need this one to keep the list of |
1954 | * relations to vacuum/analyze across transactions. |
1955 | */ |
1956 | AutovacMemCxt = AllocSetContextCreate(TopMemoryContext, |
1957 | "AV worker" , |
1958 | ALLOCSET_DEFAULT_SIZES); |
1959 | MemoryContextSwitchTo(AutovacMemCxt); |
1960 | |
1961 | /* |
1962 | * may be NULL if we couldn't find an entry (only happens if we are |
1963 | * forcing a vacuum for anti-wrap purposes). |
1964 | */ |
1965 | dbentry = pgstat_fetch_stat_dbentry(MyDatabaseId); |
1966 | |
1967 | /* Start a transaction so our commands have one to play into. */ |
1968 | StartTransactionCommand(); |
1969 | |
1970 | /* |
1971 | * Clean up any dead statistics collector entries for this DB. We always |
1972 | * want to do this exactly once per DB-processing cycle, even if we find |
1973 | * nothing worth vacuuming in the database. |
1974 | */ |
1975 | pgstat_vacuum_stat(); |
1976 | |
1977 | /* |
1978 | * Compute the multixact age for which freezing is urgent. This is |
1979 | * normally autovacuum_multixact_freeze_max_age, but may be less if we are |
1980 | * short of multixact member space. |
1981 | */ |
1982 | effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold(); |
1983 | |
1984 | /* |
1985 | * Find the pg_database entry and select the default freeze ages. We use |
1986 | * zero in template and nonconnectable databases, else the system-wide |
1987 | * default. |
1988 | */ |
1989 | tuple = SearchSysCache1(DATABASEOID, ObjectIdGetDatum(MyDatabaseId)); |
1990 | if (!HeapTupleIsValid(tuple)) |
1991 | elog(ERROR, "cache lookup failed for database %u" , MyDatabaseId); |
1992 | dbForm = (Form_pg_database) GETSTRUCT(tuple); |
1993 | |
1994 | if (dbForm->datistemplate || !dbForm->datallowconn) |
1995 | { |
1996 | default_freeze_min_age = 0; |
1997 | default_freeze_table_age = 0; |
1998 | default_multixact_freeze_min_age = 0; |
1999 | default_multixact_freeze_table_age = 0; |
2000 | } |
2001 | else |
2002 | { |
2003 | default_freeze_min_age = vacuum_freeze_min_age; |
2004 | default_freeze_table_age = vacuum_freeze_table_age; |
2005 | default_multixact_freeze_min_age = vacuum_multixact_freeze_min_age; |
2006 | default_multixact_freeze_table_age = vacuum_multixact_freeze_table_age; |
2007 | } |
2008 | |
2009 | ReleaseSysCache(tuple); |
2010 | |
2011 | /* StartTransactionCommand changed elsewhere */ |
2012 | MemoryContextSwitchTo(AutovacMemCxt); |
2013 | |
2014 | /* The database hash where pgstat keeps shared relations */ |
2015 | shared = pgstat_fetch_stat_dbentry(InvalidOid); |
2016 | |
2017 | classRel = table_open(RelationRelationId, AccessShareLock); |
2018 | |
2019 | /* create a copy so we can use it after closing pg_class */ |
2020 | pg_class_desc = CreateTupleDescCopy(RelationGetDescr(classRel)); |
2021 | |
2022 | /* create hash table for toast <-> main relid mapping */ |
2023 | MemSet(&ctl, 0, sizeof(ctl)); |
2024 | ctl.keysize = sizeof(Oid); |
2025 | ctl.entrysize = sizeof(av_relation); |
2026 | |
2027 | table_toast_map = hash_create("TOAST to main relid map" , |
2028 | 100, |
2029 | &ctl, |
2030 | HASH_ELEM | HASH_BLOBS); |
2031 | |
2032 | /* |
2033 | * Scan pg_class to determine which tables to vacuum. |
2034 | * |
2035 | * We do this in two passes: on the first one we collect the list of plain |
2036 | * relations and materialized views, and on the second one we collect |
2037 | * TOAST tables. The reason for doing the second pass is that during it we |
2038 | * want to use the main relation's pg_class.reloptions entry if the TOAST |
2039 | * table does not have any, and we cannot obtain it unless we know |
2040 | * beforehand what's the main table OID. |
2041 | * |
2042 | * We need to check TOAST tables separately because in cases with short, |
2043 | * wide tables there might be proportionally much more activity in the |
2044 | * TOAST table than in its parent. |
2045 | */ |
2046 | relScan = table_beginscan_catalog(classRel, 0, NULL); |
2047 | |
2048 | /* |
2049 | * On the first pass, we collect main tables to vacuum, and also the main |
2050 | * table relid to TOAST relid mapping. |
2051 | */ |
2052 | while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL) |
2053 | { |
2054 | Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple); |
2055 | PgStat_StatTabEntry *tabentry; |
2056 | AutoVacOpts *relopts; |
2057 | Oid relid; |
2058 | bool dovacuum; |
2059 | bool doanalyze; |
2060 | bool wraparound; |
2061 | |
2062 | if (classForm->relkind != RELKIND_RELATION && |
2063 | classForm->relkind != RELKIND_MATVIEW) |
2064 | continue; |
2065 | |
2066 | relid = classForm->oid; |
2067 | |
2068 | /* |
2069 | * Check if it is a temp table (presumably, of some other backend's). |
2070 | * We cannot safely process other backends' temp tables. |
2071 | */ |
2072 | if (classForm->relpersistence == RELPERSISTENCE_TEMP) |
2073 | { |
2074 | /* |
2075 | * We just ignore it if the owning backend is still active and |
2076 | * using the temporary schema. |
2077 | */ |
2078 | if (!isTempNamespaceInUse(classForm->relnamespace)) |
2079 | { |
2080 | /* |
2081 | * The table seems to be orphaned -- although it might be that |
2082 | * the owning backend has already deleted it and exited; our |
2083 | * pg_class scan snapshot is not necessarily up-to-date |
2084 | * anymore, so we could be looking at a committed-dead entry. |
2085 | * Remember it so we can try to delete it later. |
2086 | */ |
2087 | orphan_oids = lappend_oid(orphan_oids, relid); |
2088 | } |
2089 | continue; |
2090 | } |
2091 | |
2092 | /* Fetch reloptions and the pgstat entry for this table */ |
2093 | relopts = extract_autovac_opts(tuple, pg_class_desc); |
2094 | tabentry = get_pgstat_tabentry_relid(relid, classForm->relisshared, |
2095 | shared, dbentry); |
2096 | |
2097 | /* Check if it needs vacuum or analyze */ |
2098 | relation_needs_vacanalyze(relid, relopts, classForm, tabentry, |
2099 | effective_multixact_freeze_max_age, |
2100 | &dovacuum, &doanalyze, &wraparound); |
2101 | |
2102 | /* Relations that need work are added to table_oids */ |
2103 | if (dovacuum || doanalyze) |
2104 | table_oids = lappend_oid(table_oids, relid); |
2105 | |
2106 | /* |
2107 | * Remember TOAST associations for the second pass. Note: we must do |
2108 | * this whether or not the table is going to be vacuumed, because we |
2109 | * don't automatically vacuum toast tables along the parent table. |
2110 | */ |
2111 | if (OidIsValid(classForm->reltoastrelid)) |
2112 | { |
2113 | av_relation *hentry; |
2114 | bool found; |
2115 | |
2116 | hentry = hash_search(table_toast_map, |
2117 | &classForm->reltoastrelid, |
2118 | HASH_ENTER, &found); |
2119 | |
2120 | if (!found) |
2121 | { |
2122 | /* hash_search already filled in the key */ |
2123 | hentry->ar_relid = relid; |
2124 | hentry->ar_hasrelopts = false; |
2125 | if (relopts != NULL) |
2126 | { |
2127 | hentry->ar_hasrelopts = true; |
2128 | memcpy(&hentry->ar_reloptions, relopts, |
2129 | sizeof(AutoVacOpts)); |
2130 | } |
2131 | } |
2132 | } |
2133 | } |
2134 | |
2135 | table_endscan(relScan); |
2136 | |
2137 | /* second pass: check TOAST tables */ |
2138 | ScanKeyInit(&key, |
2139 | Anum_pg_class_relkind, |
2140 | BTEqualStrategyNumber, F_CHAREQ, |
2141 | CharGetDatum(RELKIND_TOASTVALUE)); |
2142 | |
2143 | relScan = table_beginscan_catalog(classRel, 1, &key); |
2144 | while ((tuple = heap_getnext(relScan, ForwardScanDirection)) != NULL) |
2145 | { |
2146 | Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple); |
2147 | PgStat_StatTabEntry *tabentry; |
2148 | Oid relid; |
2149 | AutoVacOpts *relopts = NULL; |
2150 | bool dovacuum; |
2151 | bool doanalyze; |
2152 | bool wraparound; |
2153 | |
2154 | /* |
2155 | * We cannot safely process other backends' temp tables, so skip 'em. |
2156 | */ |
2157 | if (classForm->relpersistence == RELPERSISTENCE_TEMP) |
2158 | continue; |
2159 | |
2160 | relid = classForm->oid; |
2161 | |
2162 | /* |
2163 | * fetch reloptions -- if this toast table does not have them, try the |
2164 | * main rel |
2165 | */ |
2166 | relopts = extract_autovac_opts(tuple, pg_class_desc); |
2167 | if (relopts == NULL) |
2168 | { |
2169 | av_relation *hentry; |
2170 | bool found; |
2171 | |
2172 | hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found); |
2173 | if (found && hentry->ar_hasrelopts) |
2174 | relopts = &hentry->ar_reloptions; |
2175 | } |
2176 | |
2177 | /* Fetch the pgstat entry for this table */ |
2178 | tabentry = get_pgstat_tabentry_relid(relid, classForm->relisshared, |
2179 | shared, dbentry); |
2180 | |
2181 | relation_needs_vacanalyze(relid, relopts, classForm, tabentry, |
2182 | effective_multixact_freeze_max_age, |
2183 | &dovacuum, &doanalyze, &wraparound); |
2184 | |
2185 | /* ignore analyze for toast tables */ |
2186 | if (dovacuum) |
2187 | table_oids = lappend_oid(table_oids, relid); |
2188 | } |
2189 | |
2190 | table_endscan(relScan); |
2191 | table_close(classRel, AccessShareLock); |
2192 | |
2193 | /* |
2194 | * Recheck orphan temporary tables, and if they still seem orphaned, drop |
2195 | * them. We'll eat a transaction per dropped table, which might seem |
2196 | * excessive, but we should only need to do anything as a result of a |
2197 | * previous backend crash, so this should not happen often enough to |
2198 | * justify "optimizing". Using separate transactions ensures that we |
2199 | * don't bloat the lock table if there are many temp tables to be dropped, |
2200 | * and it ensures that we don't lose work if a deletion attempt fails. |
2201 | */ |
2202 | foreach(cell, orphan_oids) |
2203 | { |
2204 | Oid relid = lfirst_oid(cell); |
2205 | Form_pg_class classForm; |
2206 | ObjectAddress object; |
2207 | |
2208 | /* |
2209 | * Check for user-requested abort. |
2210 | */ |
2211 | CHECK_FOR_INTERRUPTS(); |
2212 | |
2213 | /* |
2214 | * Try to lock the table. If we can't get the lock immediately, |
2215 | * somebody else is using (or dropping) the table, so it's not our |
2216 | * concern anymore. Having the lock prevents race conditions below. |
2217 | */ |
2218 | if (!ConditionalLockRelationOid(relid, AccessExclusiveLock)) |
2219 | continue; |
2220 | |
2221 | /* |
2222 | * Re-fetch the pg_class tuple and re-check whether it still seems to |
2223 | * be an orphaned temp table. If it's not there or no longer the same |
2224 | * relation, ignore it. |
2225 | */ |
2226 | tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid)); |
2227 | if (!HeapTupleIsValid(tuple)) |
2228 | { |
2229 | /* be sure to drop useless lock so we don't bloat lock table */ |
2230 | UnlockRelationOid(relid, AccessExclusiveLock); |
2231 | continue; |
2232 | } |
2233 | classForm = (Form_pg_class) GETSTRUCT(tuple); |
2234 | |
2235 | /* |
2236 | * Make all the same tests made in the loop above. In event of OID |
2237 | * counter wraparound, the pg_class entry we have now might be |
2238 | * completely unrelated to the one we saw before. |
2239 | */ |
2240 | if (!((classForm->relkind == RELKIND_RELATION || |
2241 | classForm->relkind == RELKIND_MATVIEW) && |
2242 | classForm->relpersistence == RELPERSISTENCE_TEMP)) |
2243 | { |
2244 | UnlockRelationOid(relid, AccessExclusiveLock); |
2245 | continue; |
2246 | } |
2247 | |
2248 | if (isTempNamespaceInUse(classForm->relnamespace)) |
2249 | { |
2250 | UnlockRelationOid(relid, AccessExclusiveLock); |
2251 | continue; |
2252 | } |
2253 | |
2254 | /* OK, let's delete it */ |
2255 | ereport(LOG, |
2256 | (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"" , |
2257 | get_database_name(MyDatabaseId), |
2258 | get_namespace_name(classForm->relnamespace), |
2259 | NameStr(classForm->relname)))); |
2260 | |
2261 | object.classId = RelationRelationId; |
2262 | object.objectId = relid; |
2263 | object.objectSubId = 0; |
2264 | performDeletion(&object, DROP_CASCADE, |
2265 | PERFORM_DELETION_INTERNAL | |
2266 | PERFORM_DELETION_QUIETLY | |
2267 | PERFORM_DELETION_SKIP_EXTENSIONS); |
2268 | |
2269 | /* |
2270 | * To commit the deletion, end current transaction and start a new |
2271 | * one. Note this also releases the lock we took. |
2272 | */ |
2273 | CommitTransactionCommand(); |
2274 | StartTransactionCommand(); |
2275 | |
2276 | /* StartTransactionCommand changed current memory context */ |
2277 | MemoryContextSwitchTo(AutovacMemCxt); |
2278 | } |
2279 | |
2280 | /* |
2281 | * Create a buffer access strategy object for VACUUM to use. We want to |
2282 | * use the same one across all the vacuum operations we perform, since the |
2283 | * point is for VACUUM not to blow out the shared cache. |
2284 | */ |
2285 | bstrategy = GetAccessStrategy(BAS_VACUUM); |
2286 | |
2287 | /* |
2288 | * create a memory context to act as fake PortalContext, so that the |
2289 | * contexts created in the vacuum code are cleaned up for each table. |
2290 | */ |
2291 | PortalContext = AllocSetContextCreate(AutovacMemCxt, |
2292 | "Autovacuum Portal" , |
2293 | ALLOCSET_DEFAULT_SIZES); |
2294 | |
2295 | /* |
2296 | * Perform operations on collected tables. |
2297 | */ |
2298 | foreach(cell, table_oids) |
2299 | { |
2300 | Oid relid = lfirst_oid(cell); |
2301 | HeapTuple classTup; |
2302 | autovac_table *tab; |
2303 | bool isshared; |
2304 | bool skipit; |
2305 | double stdVacuumCostDelay; |
2306 | int stdVacuumCostLimit; |
2307 | dlist_iter iter; |
2308 | |
2309 | CHECK_FOR_INTERRUPTS(); |
2310 | |
2311 | /* |
2312 | * Check for config changes before processing each collected table. |
2313 | */ |
2314 | if (got_SIGHUP) |
2315 | { |
2316 | got_SIGHUP = false; |
2317 | ProcessConfigFile(PGC_SIGHUP); |
2318 | |
2319 | /* |
2320 | * You might be tempted to bail out if we see autovacuum is now |
2321 | * disabled. Must resist that temptation -- this might be a |
2322 | * for-wraparound emergency worker, in which case that would be |
2323 | * entirely inappropriate. |
2324 | */ |
2325 | } |
2326 | |
2327 | /* |
2328 | * Find out whether the table is shared or not. (It's slightly |
2329 | * annoying to fetch the syscache entry just for this, but in typical |
2330 | * cases it adds little cost because table_recheck_autovac would |
2331 | * refetch the entry anyway. We could buy that back by copying the |
2332 | * tuple here and passing it to table_recheck_autovac, but that |
2333 | * increases the odds of that function working with stale data.) |
2334 | */ |
2335 | classTup = SearchSysCache1(RELOID, ObjectIdGetDatum(relid)); |
2336 | if (!HeapTupleIsValid(classTup)) |
2337 | continue; /* somebody deleted the rel, forget it */ |
2338 | isshared = ((Form_pg_class) GETSTRUCT(classTup))->relisshared; |
2339 | ReleaseSysCache(classTup); |
2340 | |
2341 | /* |
2342 | * Hold schedule lock from here until we've claimed the table. We |
2343 | * also need the AutovacuumLock to walk the worker array, but that one |
2344 | * can just be a shared lock. |
2345 | */ |
2346 | LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE); |
2347 | LWLockAcquire(AutovacuumLock, LW_SHARED); |
2348 | |
2349 | /* |
2350 | * Check whether the table is being vacuumed concurrently by another |
2351 | * worker. |
2352 | */ |
2353 | skipit = false; |
2354 | dlist_foreach(iter, &AutoVacuumShmem->av_runningWorkers) |
2355 | { |
2356 | WorkerInfo worker = dlist_container(WorkerInfoData, wi_links, iter.cur); |
2357 | |
2358 | /* ignore myself */ |
2359 | if (worker == MyWorkerInfo) |
2360 | continue; |
2361 | |
2362 | /* ignore workers in other databases (unless table is shared) */ |
2363 | if (!worker->wi_sharedrel && worker->wi_dboid != MyDatabaseId) |
2364 | continue; |
2365 | |
2366 | if (worker->wi_tableoid == relid) |
2367 | { |
2368 | skipit = true; |
2369 | found_concurrent_worker = true; |
2370 | break; |
2371 | } |
2372 | } |
2373 | LWLockRelease(AutovacuumLock); |
2374 | if (skipit) |
2375 | { |
2376 | LWLockRelease(AutovacuumScheduleLock); |
2377 | continue; |
2378 | } |
2379 | |
2380 | /* |
2381 | * Store the table's OID in shared memory before releasing the |
2382 | * schedule lock, so that other workers don't try to vacuum it |
2383 | * concurrently. (We claim it here so as not to hold |
2384 | * AutovacuumScheduleLock while rechecking the stats.) |
2385 | */ |
2386 | MyWorkerInfo->wi_tableoid = relid; |
2387 | MyWorkerInfo->wi_sharedrel = isshared; |
2388 | LWLockRelease(AutovacuumScheduleLock); |
2389 | |
2390 | /* |
2391 | * Check whether pgstat data still says we need to vacuum this table. |
2392 | * It could have changed if something else processed the table while |
2393 | * we weren't looking. |
2394 | * |
2395 | * Note: we have a special case in pgstat code to ensure that the |
2396 | * stats we read are as up-to-date as possible, to avoid the problem |
2397 | * that somebody just finished vacuuming this table. The window to |
2398 | * the race condition is not closed but it is very small. |
2399 | */ |
2400 | MemoryContextSwitchTo(AutovacMemCxt); |
2401 | tab = table_recheck_autovac(relid, table_toast_map, pg_class_desc, |
2402 | effective_multixact_freeze_max_age); |
2403 | if (tab == NULL) |
2404 | { |
2405 | /* someone else vacuumed the table, or it went away */ |
2406 | LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE); |
2407 | MyWorkerInfo->wi_tableoid = InvalidOid; |
2408 | MyWorkerInfo->wi_sharedrel = false; |
2409 | LWLockRelease(AutovacuumScheduleLock); |
2410 | continue; |
2411 | } |
2412 | |
2413 | /* |
2414 | * Remember the prevailing values of the vacuum cost GUCs. We have to |
2415 | * restore these at the bottom of the loop, else we'll compute wrong |
2416 | * values in the next iteration of autovac_balance_cost(). |
2417 | */ |
2418 | stdVacuumCostDelay = VacuumCostDelay; |
2419 | stdVacuumCostLimit = VacuumCostLimit; |
2420 | |
2421 | /* Must hold AutovacuumLock while mucking with cost balance info */ |
2422 | LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE); |
2423 | |
2424 | /* advertise my cost delay parameters for the balancing algorithm */ |
2425 | MyWorkerInfo->wi_dobalance = tab->at_dobalance; |
2426 | MyWorkerInfo->wi_cost_delay = tab->at_vacuum_cost_delay; |
2427 | MyWorkerInfo->wi_cost_limit = tab->at_vacuum_cost_limit; |
2428 | MyWorkerInfo->wi_cost_limit_base = tab->at_vacuum_cost_limit; |
2429 | |
2430 | /* do a balance */ |
2431 | autovac_balance_cost(); |
2432 | |
2433 | /* set the active cost parameters from the result of that */ |
2434 | AutoVacuumUpdateDelay(); |
2435 | |
2436 | /* done */ |
2437 | LWLockRelease(AutovacuumLock); |
2438 | |
2439 | /* clean up memory before each iteration */ |
2440 | MemoryContextResetAndDeleteChildren(PortalContext); |
2441 | |
2442 | /* |
2443 | * Save the relation name for a possible error message, to avoid a |
2444 | * catalog lookup in case of an error. If any of these return NULL, |
2445 | * then the relation has been dropped since last we checked; skip it. |
2446 | * Note: they must live in a long-lived memory context because we call |
2447 | * vacuum and analyze in different transactions. |
2448 | */ |
2449 | |
2450 | tab->at_relname = get_rel_name(tab->at_relid); |
2451 | tab->at_nspname = get_namespace_name(get_rel_namespace(tab->at_relid)); |
2452 | tab->at_datname = get_database_name(MyDatabaseId); |
2453 | if (!tab->at_relname || !tab->at_nspname || !tab->at_datname) |
2454 | goto deleted; |
2455 | |
2456 | /* |
2457 | * We will abort vacuuming the current table if something errors out, |
2458 | * and continue with the next one in schedule; in particular, this |
2459 | * happens if we are interrupted with SIGINT. |
2460 | */ |
2461 | PG_TRY(); |
2462 | { |
2463 | /* Use PortalContext for any per-table allocations */ |
2464 | MemoryContextSwitchTo(PortalContext); |
2465 | |
2466 | /* have at it */ |
2467 | autovacuum_do_vac_analyze(tab, bstrategy); |
2468 | |
2469 | /* |
2470 | * Clear a possible query-cancel signal, to avoid a late reaction |
2471 | * to an automatically-sent signal because of vacuuming the |
2472 | * current table (we're done with it, so it would make no sense to |
2473 | * cancel at this point.) |
2474 | */ |
2475 | QueryCancelPending = false; |
2476 | } |
2477 | PG_CATCH(); |
2478 | { |
2479 | /* |
2480 | * Abort the transaction, start a new one, and proceed with the |
2481 | * next table in our list. |
2482 | */ |
2483 | HOLD_INTERRUPTS(); |
2484 | if (tab->at_params.options & VACOPT_VACUUM) |
2485 | errcontext("automatic vacuum of table \"%s.%s.%s\"" , |
2486 | tab->at_datname, tab->at_nspname, tab->at_relname); |
2487 | else |
2488 | errcontext("automatic analyze of table \"%s.%s.%s\"" , |
2489 | tab->at_datname, tab->at_nspname, tab->at_relname); |
2490 | EmitErrorReport(); |
2491 | |
2492 | /* this resets the PGXACT flags too */ |
2493 | AbortOutOfAnyTransaction(); |
2494 | FlushErrorState(); |
2495 | MemoryContextResetAndDeleteChildren(PortalContext); |
2496 | |
2497 | /* restart our transaction for the following operations */ |
2498 | StartTransactionCommand(); |
2499 | RESUME_INTERRUPTS(); |
2500 | } |
2501 | PG_END_TRY(); |
2502 | |
2503 | /* Make sure we're back in AutovacMemCxt */ |
2504 | MemoryContextSwitchTo(AutovacMemCxt); |
2505 | |
2506 | did_vacuum = true; |
2507 | |
2508 | /* the PGXACT flags are reset at the next end of transaction */ |
2509 | |
2510 | /* be tidy */ |
2511 | deleted: |
2512 | if (tab->at_datname != NULL) |
2513 | pfree(tab->at_datname); |
2514 | if (tab->at_nspname != NULL) |
2515 | pfree(tab->at_nspname); |
2516 | if (tab->at_relname != NULL) |
2517 | pfree(tab->at_relname); |
2518 | pfree(tab); |
2519 | |
2520 | /* |
2521 | * Remove my info from shared memory. We could, but intentionally |
2522 | * don't, clear wi_cost_limit and friends --- this is on the |
2523 | * assumption that we probably have more to do with similar cost |
2524 | * settings, so we don't want to give up our share of I/O for a very |
2525 | * short interval and thereby thrash the global balance. |
2526 | */ |
2527 | LWLockAcquire(AutovacuumScheduleLock, LW_EXCLUSIVE); |
2528 | MyWorkerInfo->wi_tableoid = InvalidOid; |
2529 | MyWorkerInfo->wi_sharedrel = false; |
2530 | LWLockRelease(AutovacuumScheduleLock); |
2531 | |
2532 | /* restore vacuum cost GUCs for the next iteration */ |
2533 | VacuumCostDelay = stdVacuumCostDelay; |
2534 | VacuumCostLimit = stdVacuumCostLimit; |
2535 | } |
2536 | |
2537 | /* |
2538 | * Perform additional work items, as requested by backends. |
2539 | */ |
2540 | LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE); |
2541 | for (i = 0; i < NUM_WORKITEMS; i++) |
2542 | { |
2543 | AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i]; |
2544 | |
2545 | if (!workitem->avw_used) |
2546 | continue; |
2547 | if (workitem->avw_active) |
2548 | continue; |
2549 | if (workitem->avw_database != MyDatabaseId) |
2550 | continue; |
2551 | |
2552 | /* claim this one, and release lock while performing it */ |
2553 | workitem->avw_active = true; |
2554 | LWLockRelease(AutovacuumLock); |
2555 | |
2556 | perform_work_item(workitem); |
2557 | |
2558 | /* |
2559 | * Check for config changes before acquiring lock for further jobs. |
2560 | */ |
2561 | CHECK_FOR_INTERRUPTS(); |
2562 | if (got_SIGHUP) |
2563 | { |
2564 | got_SIGHUP = false; |
2565 | ProcessConfigFile(PGC_SIGHUP); |
2566 | } |
2567 | |
2568 | LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE); |
2569 | |
2570 | /* and mark it done */ |
2571 | workitem->avw_active = false; |
2572 | workitem->avw_used = false; |
2573 | } |
2574 | LWLockRelease(AutovacuumLock); |
2575 | |
2576 | /* |
2577 | * We leak table_toast_map here (among other things), but since we're |
2578 | * going away soon, it's not a problem. |
2579 | */ |
2580 | |
2581 | /* |
2582 | * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We |
2583 | * only need to do this once, not after each table. |
2584 | * |
2585 | * Even if we didn't vacuum anything, it may still be important to do |
2586 | * this, because one indirect effect of vac_update_datfrozenxid() is to |
2587 | * update ShmemVariableCache->xidVacLimit. That might need to be done |
2588 | * even if we haven't vacuumed anything, because relations with older |
2589 | * relfrozenxid values or other databases with older datfrozenxid values |
2590 | * might have been dropped, allowing xidVacLimit to advance. |
2591 | * |
2592 | * However, it's also important not to do this blindly in all cases, |
2593 | * because when autovacuum=off this will restart the autovacuum launcher. |
2594 | * If we're not careful, an infinite loop can result, where workers find |
2595 | * no work to do and restart the launcher, which starts another worker in |
2596 | * the same database that finds no work to do. To prevent that, we skip |
2597 | * this if (1) we found no work to do and (2) we skipped at least one |
2598 | * table due to concurrent autovacuum activity. In that case, the other |
2599 | * worker has already done it, or will do so when it finishes. |
2600 | */ |
2601 | if (did_vacuum || !found_concurrent_worker) |
2602 | vac_update_datfrozenxid(); |
2603 | |
2604 | /* Finally close out the last transaction. */ |
2605 | CommitTransactionCommand(); |
2606 | } |
2607 | |
2608 | /* |
2609 | * Execute a previously registered work item. |
2610 | */ |
2611 | static void |
2612 | perform_work_item(AutoVacuumWorkItem *workitem) |
2613 | { |
2614 | char *cur_datname = NULL; |
2615 | char *cur_nspname = NULL; |
2616 | char *cur_relname = NULL; |
2617 | |
2618 | /* |
2619 | * Note we do not store table info in MyWorkerInfo, since this is not |
2620 | * vacuuming proper. |
2621 | */ |
2622 | |
2623 | /* |
2624 | * Save the relation name for a possible error message, to avoid a catalog |
2625 | * lookup in case of an error. If any of these return NULL, then the |
2626 | * relation has been dropped since last we checked; skip it. |
2627 | */ |
2628 | Assert(CurrentMemoryContext == AutovacMemCxt); |
2629 | |
2630 | cur_relname = get_rel_name(workitem->avw_relation); |
2631 | cur_nspname = get_namespace_name(get_rel_namespace(workitem->avw_relation)); |
2632 | cur_datname = get_database_name(MyDatabaseId); |
2633 | if (!cur_relname || !cur_nspname || !cur_datname) |
2634 | goto deleted2; |
2635 | |
2636 | autovac_report_workitem(workitem, cur_nspname, cur_relname); |
2637 | |
2638 | /* clean up memory before each work item */ |
2639 | MemoryContextResetAndDeleteChildren(PortalContext); |
2640 | |
2641 | /* |
2642 | * We will abort the current work item if something errors out, and |
2643 | * continue with the next one; in particular, this happens if we are |
2644 | * interrupted with SIGINT. Note that this means that the work item list |
2645 | * can be lossy. |
2646 | */ |
2647 | PG_TRY(); |
2648 | { |
2649 | /* Use PortalContext for any per-work-item allocations */ |
2650 | MemoryContextSwitchTo(PortalContext); |
2651 | |
2652 | /* have at it */ |
2653 | switch (workitem->avw_type) |
2654 | { |
2655 | case AVW_BRINSummarizeRange: |
2656 | DirectFunctionCall2(brin_summarize_range, |
2657 | ObjectIdGetDatum(workitem->avw_relation), |
2658 | Int64GetDatum((int64) workitem->avw_blockNumber)); |
2659 | break; |
2660 | default: |
2661 | elog(WARNING, "unrecognized work item found: type %d" , |
2662 | workitem->avw_type); |
2663 | break; |
2664 | } |
2665 | |
2666 | /* |
2667 | * Clear a possible query-cancel signal, to avoid a late reaction to |
2668 | * an automatically-sent signal because of vacuuming the current table |
2669 | * (we're done with it, so it would make no sense to cancel at this |
2670 | * point.) |
2671 | */ |
2672 | QueryCancelPending = false; |
2673 | } |
2674 | PG_CATCH(); |
2675 | { |
2676 | /* |
2677 | * Abort the transaction, start a new one, and proceed with the next |
2678 | * table in our list. |
2679 | */ |
2680 | HOLD_INTERRUPTS(); |
2681 | errcontext("processing work entry for relation \"%s.%s.%s\"" , |
2682 | cur_datname, cur_nspname, cur_relname); |
2683 | EmitErrorReport(); |
2684 | |
2685 | /* this resets the PGXACT flags too */ |
2686 | AbortOutOfAnyTransaction(); |
2687 | FlushErrorState(); |
2688 | MemoryContextResetAndDeleteChildren(PortalContext); |
2689 | |
2690 | /* restart our transaction for the following operations */ |
2691 | StartTransactionCommand(); |
2692 | RESUME_INTERRUPTS(); |
2693 | } |
2694 | PG_END_TRY(); |
2695 | |
2696 | /* Make sure we're back in AutovacMemCxt */ |
2697 | MemoryContextSwitchTo(AutovacMemCxt); |
2698 | |
2699 | /* We intentionally do not set did_vacuum here */ |
2700 | |
2701 | /* be tidy */ |
2702 | deleted2: |
2703 | if (cur_datname) |
2704 | pfree(cur_datname); |
2705 | if (cur_nspname) |
2706 | pfree(cur_nspname); |
2707 | if (cur_relname) |
2708 | pfree(cur_relname); |
2709 | } |
2710 | |
2711 | /* |
2712 | * extract_autovac_opts |
2713 | * |
2714 | * Given a relation's pg_class tuple, return the AutoVacOpts portion of |
2715 | * reloptions, if set; otherwise, return NULL. |
2716 | */ |
2717 | static AutoVacOpts * |
2718 | (HeapTuple tup, TupleDesc pg_class_desc) |
2719 | { |
2720 | bytea *relopts; |
2721 | AutoVacOpts *av; |
2722 | |
2723 | Assert(((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_RELATION || |
2724 | ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_MATVIEW || |
2725 | ((Form_pg_class) GETSTRUCT(tup))->relkind == RELKIND_TOASTVALUE); |
2726 | |
2727 | relopts = extractRelOptions(tup, pg_class_desc, NULL); |
2728 | if (relopts == NULL) |
2729 | return NULL; |
2730 | |
2731 | av = palloc(sizeof(AutoVacOpts)); |
2732 | memcpy(av, &(((StdRdOptions *) relopts)->autovacuum), sizeof(AutoVacOpts)); |
2733 | pfree(relopts); |
2734 | |
2735 | return av; |
2736 | } |
2737 | |
2738 | /* |
2739 | * get_pgstat_tabentry_relid |
2740 | * |
2741 | * Fetch the pgstat entry of a table, either local to a database or shared. |
2742 | */ |
2743 | static PgStat_StatTabEntry * |
2744 | get_pgstat_tabentry_relid(Oid relid, bool isshared, PgStat_StatDBEntry *shared, |
2745 | PgStat_StatDBEntry *dbentry) |
2746 | { |
2747 | PgStat_StatTabEntry *tabentry = NULL; |
2748 | |
2749 | if (isshared) |
2750 | { |
2751 | if (PointerIsValid(shared)) |
2752 | tabentry = hash_search(shared->tables, &relid, |
2753 | HASH_FIND, NULL); |
2754 | } |
2755 | else if (PointerIsValid(dbentry)) |
2756 | tabentry = hash_search(dbentry->tables, &relid, |
2757 | HASH_FIND, NULL); |
2758 | |
2759 | return tabentry; |
2760 | } |
2761 | |
2762 | /* |
2763 | * table_recheck_autovac |
2764 | * |
2765 | * Recheck whether a table still needs vacuum or analyze. Return value is a |
2766 | * valid autovac_table pointer if it does, NULL otherwise. |
2767 | * |
2768 | * Note that the returned autovac_table does not have the name fields set. |
2769 | */ |
2770 | static autovac_table * |
2771 | table_recheck_autovac(Oid relid, HTAB *table_toast_map, |
2772 | TupleDesc pg_class_desc, |
2773 | int effective_multixact_freeze_max_age) |
2774 | { |
2775 | Form_pg_class classForm; |
2776 | HeapTuple classTup; |
2777 | bool dovacuum; |
2778 | bool doanalyze; |
2779 | autovac_table *tab = NULL; |
2780 | PgStat_StatTabEntry *tabentry; |
2781 | PgStat_StatDBEntry *shared; |
2782 | PgStat_StatDBEntry *dbentry; |
2783 | bool wraparound; |
2784 | AutoVacOpts *avopts; |
2785 | |
2786 | /* use fresh stats */ |
2787 | autovac_refresh_stats(); |
2788 | |
2789 | shared = pgstat_fetch_stat_dbentry(InvalidOid); |
2790 | dbentry = pgstat_fetch_stat_dbentry(MyDatabaseId); |
2791 | |
2792 | /* fetch the relation's relcache entry */ |
2793 | classTup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid)); |
2794 | if (!HeapTupleIsValid(classTup)) |
2795 | return NULL; |
2796 | classForm = (Form_pg_class) GETSTRUCT(classTup); |
2797 | |
2798 | /* |
2799 | * Get the applicable reloptions. If it is a TOAST table, try to get the |
2800 | * main table reloptions if the toast table itself doesn't have. |
2801 | */ |
2802 | avopts = extract_autovac_opts(classTup, pg_class_desc); |
2803 | if (classForm->relkind == RELKIND_TOASTVALUE && |
2804 | avopts == NULL && table_toast_map != NULL) |
2805 | { |
2806 | av_relation *hentry; |
2807 | bool found; |
2808 | |
2809 | hentry = hash_search(table_toast_map, &relid, HASH_FIND, &found); |
2810 | if (found && hentry->ar_hasrelopts) |
2811 | avopts = &hentry->ar_reloptions; |
2812 | } |
2813 | |
2814 | /* fetch the pgstat table entry */ |
2815 | tabentry = get_pgstat_tabentry_relid(relid, classForm->relisshared, |
2816 | shared, dbentry); |
2817 | |
2818 | relation_needs_vacanalyze(relid, avopts, classForm, tabentry, |
2819 | effective_multixact_freeze_max_age, |
2820 | &dovacuum, &doanalyze, &wraparound); |
2821 | |
2822 | /* ignore ANALYZE for toast tables */ |
2823 | if (classForm->relkind == RELKIND_TOASTVALUE) |
2824 | doanalyze = false; |
2825 | |
2826 | /* OK, it needs something done */ |
2827 | if (doanalyze || dovacuum) |
2828 | { |
2829 | int freeze_min_age; |
2830 | int freeze_table_age; |
2831 | int multixact_freeze_min_age; |
2832 | int multixact_freeze_table_age; |
2833 | int vac_cost_limit; |
2834 | double vac_cost_delay; |
2835 | int log_min_duration; |
2836 | |
2837 | /* |
2838 | * Calculate the vacuum cost parameters and the freeze ages. If there |
2839 | * are options set in pg_class.reloptions, use them; in the case of a |
2840 | * toast table, try the main table too. Otherwise use the GUC |
2841 | * defaults, autovacuum's own first and plain vacuum second. |
2842 | */ |
2843 | |
2844 | /* -1 in autovac setting means use plain vacuum_cost_delay */ |
2845 | vac_cost_delay = (avopts && avopts->vacuum_cost_delay >= 0) |
2846 | ? avopts->vacuum_cost_delay |
2847 | : (autovacuum_vac_cost_delay >= 0) |
2848 | ? autovacuum_vac_cost_delay |
2849 | : VacuumCostDelay; |
2850 | |
2851 | /* 0 or -1 in autovac setting means use plain vacuum_cost_limit */ |
2852 | vac_cost_limit = (avopts && avopts->vacuum_cost_limit > 0) |
2853 | ? avopts->vacuum_cost_limit |
2854 | : (autovacuum_vac_cost_limit > 0) |
2855 | ? autovacuum_vac_cost_limit |
2856 | : VacuumCostLimit; |
2857 | |
2858 | /* -1 in autovac setting means use log_autovacuum_min_duration */ |
2859 | log_min_duration = (avopts && avopts->log_min_duration >= 0) |
2860 | ? avopts->log_min_duration |
2861 | : Log_autovacuum_min_duration; |
2862 | |
2863 | /* these do not have autovacuum-specific settings */ |
2864 | freeze_min_age = (avopts && avopts->freeze_min_age >= 0) |
2865 | ? avopts->freeze_min_age |
2866 | : default_freeze_min_age; |
2867 | |
2868 | freeze_table_age = (avopts && avopts->freeze_table_age >= 0) |
2869 | ? avopts->freeze_table_age |
2870 | : default_freeze_table_age; |
2871 | |
2872 | multixact_freeze_min_age = (avopts && |
2873 | avopts->multixact_freeze_min_age >= 0) |
2874 | ? avopts->multixact_freeze_min_age |
2875 | : default_multixact_freeze_min_age; |
2876 | |
2877 | multixact_freeze_table_age = (avopts && |
2878 | avopts->multixact_freeze_table_age >= 0) |
2879 | ? avopts->multixact_freeze_table_age |
2880 | : default_multixact_freeze_table_age; |
2881 | |
2882 | tab = palloc(sizeof(autovac_table)); |
2883 | tab->at_relid = relid; |
2884 | tab->at_sharedrel = classForm->relisshared; |
2885 | tab->at_params.options = VACOPT_SKIPTOAST | |
2886 | (dovacuum ? VACOPT_VACUUM : 0) | |
2887 | (doanalyze ? VACOPT_ANALYZE : 0) | |
2888 | (!wraparound ? VACOPT_SKIP_LOCKED : 0); |
2889 | tab->at_params.index_cleanup = VACOPT_TERNARY_DEFAULT; |
2890 | tab->at_params.truncate = VACOPT_TERNARY_DEFAULT; |
2891 | tab->at_params.freeze_min_age = freeze_min_age; |
2892 | tab->at_params.freeze_table_age = freeze_table_age; |
2893 | tab->at_params.multixact_freeze_min_age = multixact_freeze_min_age; |
2894 | tab->at_params.multixact_freeze_table_age = multixact_freeze_table_age; |
2895 | tab->at_params.is_wraparound = wraparound; |
2896 | tab->at_params.log_min_duration = log_min_duration; |
2897 | tab->at_vacuum_cost_limit = vac_cost_limit; |
2898 | tab->at_vacuum_cost_delay = vac_cost_delay; |
2899 | tab->at_relname = NULL; |
2900 | tab->at_nspname = NULL; |
2901 | tab->at_datname = NULL; |
2902 | |
2903 | /* |
2904 | * If any of the cost delay parameters has been set individually for |
2905 | * this table, disable the balancing algorithm. |
2906 | */ |
2907 | tab->at_dobalance = |
2908 | !(avopts && (avopts->vacuum_cost_limit > 0 || |
2909 | avopts->vacuum_cost_delay > 0)); |
2910 | } |
2911 | |
2912 | heap_freetuple(classTup); |
2913 | |
2914 | return tab; |
2915 | } |
2916 | |
2917 | /* |
2918 | * relation_needs_vacanalyze |
2919 | * |
2920 | * Check whether a relation needs to be vacuumed or analyzed; return each into |
2921 | * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is |
2922 | * being forced because of Xid or multixact wraparound. |
2923 | * |
2924 | * relopts is a pointer to the AutoVacOpts options (either for itself in the |
2925 | * case of a plain table, or for either itself or its parent table in the case |
2926 | * of a TOAST table), NULL if none; tabentry is the pgstats entry, which can be |
2927 | * NULL. |
2928 | * |
2929 | * A table needs to be vacuumed if the number of dead tuples exceeds a |
2930 | * threshold. This threshold is calculated as |
2931 | * |
2932 | * threshold = vac_base_thresh + vac_scale_factor * reltuples |
2933 | * |
2934 | * For analyze, the analysis done is that the number of tuples inserted, |
2935 | * deleted and updated since the last analyze exceeds a threshold calculated |
2936 | * in the same fashion as above. Note that the collector actually stores |
2937 | * the number of tuples (both live and dead) that there were as of the last |
2938 | * analyze. This is asymmetric to the VACUUM case. |
2939 | * |
2940 | * We also force vacuum if the table's relfrozenxid is more than freeze_max_age |
2941 | * transactions back, and if its relminmxid is more than |
2942 | * multixact_freeze_max_age multixacts back. |
2943 | * |
2944 | * A table whose autovacuum_enabled option is false is |
2945 | * automatically skipped (unless we have to vacuum it due to freeze_max_age). |
2946 | * Thus autovacuum can be disabled for specific tables. Also, when the stats |
2947 | * collector does not have data about a table, it will be skipped. |
2948 | * |
2949 | * A table whose vac_base_thresh value is < 0 takes the base value from the |
2950 | * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor |
2951 | * value < 0 is substituted with the value of |
2952 | * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze. |
2953 | */ |
2954 | static void |
2955 | relation_needs_vacanalyze(Oid relid, |
2956 | AutoVacOpts *relopts, |
2957 | Form_pg_class classForm, |
2958 | PgStat_StatTabEntry *tabentry, |
2959 | int effective_multixact_freeze_max_age, |
2960 | /* output params below */ |
2961 | bool *dovacuum, |
2962 | bool *doanalyze, |
2963 | bool *wraparound) |
2964 | { |
2965 | bool force_vacuum; |
2966 | bool av_enabled; |
2967 | float4 reltuples; /* pg_class.reltuples */ |
2968 | |
2969 | /* constants from reloptions or GUC variables */ |
2970 | int vac_base_thresh, |
2971 | anl_base_thresh; |
2972 | float4 vac_scale_factor, |
2973 | anl_scale_factor; |
2974 | |
2975 | /* thresholds calculated from above constants */ |
2976 | float4 vacthresh, |
2977 | anlthresh; |
2978 | |
2979 | /* number of vacuum (resp. analyze) tuples at this time */ |
2980 | float4 vactuples, |
2981 | anltuples; |
2982 | |
2983 | /* freeze parameters */ |
2984 | int freeze_max_age; |
2985 | int multixact_freeze_max_age; |
2986 | TransactionId xidForceLimit; |
2987 | MultiXactId multiForceLimit; |
2988 | |
2989 | AssertArg(classForm != NULL); |
2990 | AssertArg(OidIsValid(relid)); |
2991 | |
2992 | /* |
2993 | * Determine vacuum/analyze equation parameters. We have two possible |
2994 | * sources: the passed reloptions (which could be a main table or a toast |
2995 | * table), or the autovacuum GUC variables. |
2996 | */ |
2997 | |
2998 | /* -1 in autovac setting means use plain vacuum_scale_factor */ |
2999 | vac_scale_factor = (relopts && relopts->vacuum_scale_factor >= 0) |
3000 | ? relopts->vacuum_scale_factor |
3001 | : autovacuum_vac_scale; |
3002 | |
3003 | vac_base_thresh = (relopts && relopts->vacuum_threshold >= 0) |
3004 | ? relopts->vacuum_threshold |
3005 | : autovacuum_vac_thresh; |
3006 | |
3007 | anl_scale_factor = (relopts && relopts->analyze_scale_factor >= 0) |
3008 | ? relopts->analyze_scale_factor |
3009 | : autovacuum_anl_scale; |
3010 | |
3011 | anl_base_thresh = (relopts && relopts->analyze_threshold >= 0) |
3012 | ? relopts->analyze_threshold |
3013 | : autovacuum_anl_thresh; |
3014 | |
3015 | freeze_max_age = (relopts && relopts->freeze_max_age >= 0) |
3016 | ? Min(relopts->freeze_max_age, autovacuum_freeze_max_age) |
3017 | : autovacuum_freeze_max_age; |
3018 | |
3019 | multixact_freeze_max_age = (relopts && relopts->multixact_freeze_max_age >= 0) |
3020 | ? Min(relopts->multixact_freeze_max_age, effective_multixact_freeze_max_age) |
3021 | : effective_multixact_freeze_max_age; |
3022 | |
3023 | av_enabled = (relopts ? relopts->enabled : true); |
3024 | |
3025 | /* Force vacuum if table is at risk of wraparound */ |
3026 | xidForceLimit = recentXid - freeze_max_age; |
3027 | if (xidForceLimit < FirstNormalTransactionId) |
3028 | xidForceLimit -= FirstNormalTransactionId; |
3029 | force_vacuum = (TransactionIdIsNormal(classForm->relfrozenxid) && |
3030 | TransactionIdPrecedes(classForm->relfrozenxid, |
3031 | xidForceLimit)); |
3032 | if (!force_vacuum) |
3033 | { |
3034 | multiForceLimit = recentMulti - multixact_freeze_max_age; |
3035 | if (multiForceLimit < FirstMultiXactId) |
3036 | multiForceLimit -= FirstMultiXactId; |
3037 | force_vacuum = MultiXactIdIsValid(classForm->relminmxid) && |
3038 | MultiXactIdPrecedes(classForm->relminmxid, multiForceLimit); |
3039 | } |
3040 | *wraparound = force_vacuum; |
3041 | |
3042 | /* User disabled it in pg_class.reloptions? (But ignore if at risk) */ |
3043 | if (!av_enabled && !force_vacuum) |
3044 | { |
3045 | *doanalyze = false; |
3046 | *dovacuum = false; |
3047 | return; |
3048 | } |
3049 | |
3050 | /* |
3051 | * If we found the table in the stats hash, and autovacuum is currently |
3052 | * enabled, make a threshold-based decision whether to vacuum and/or |
3053 | * analyze. If autovacuum is currently disabled, we must be here for |
3054 | * anti-wraparound vacuuming only, so don't vacuum (or analyze) anything |
3055 | * that's not being forced. |
3056 | */ |
3057 | if (PointerIsValid(tabentry) && AutoVacuumingActive()) |
3058 | { |
3059 | reltuples = classForm->reltuples; |
3060 | vactuples = tabentry->n_dead_tuples; |
3061 | anltuples = tabentry->changes_since_analyze; |
3062 | |
3063 | vacthresh = (float4) vac_base_thresh + vac_scale_factor * reltuples; |
3064 | anlthresh = (float4) anl_base_thresh + anl_scale_factor * reltuples; |
3065 | |
3066 | /* |
3067 | * Note that we don't need to take special consideration for stat |
3068 | * reset, because if that happens, the last vacuum and analyze counts |
3069 | * will be reset too. |
3070 | */ |
3071 | elog(DEBUG3, "%s: vac: %.0f (threshold %.0f), anl: %.0f (threshold %.0f)" , |
3072 | NameStr(classForm->relname), |
3073 | vactuples, vacthresh, anltuples, anlthresh); |
3074 | |
3075 | /* Determine if this table needs vacuum or analyze. */ |
3076 | *dovacuum = force_vacuum || (vactuples > vacthresh); |
3077 | *doanalyze = (anltuples > anlthresh); |
3078 | } |
3079 | else |
3080 | { |
3081 | /* |
3082 | * Skip a table not found in stat hash, unless we have to force vacuum |
3083 | * for anti-wrap purposes. If it's not acted upon, there's no need to |
3084 | * vacuum it. |
3085 | */ |
3086 | *dovacuum = force_vacuum; |
3087 | *doanalyze = false; |
3088 | } |
3089 | |
3090 | /* ANALYZE refuses to work with pg_statistic */ |
3091 | if (relid == StatisticRelationId) |
3092 | *doanalyze = false; |
3093 | } |
3094 | |
3095 | /* |
3096 | * autovacuum_do_vac_analyze |
3097 | * Vacuum and/or analyze the specified table |
3098 | */ |
3099 | static void |
3100 | autovacuum_do_vac_analyze(autovac_table *tab, BufferAccessStrategy bstrategy) |
3101 | { |
3102 | RangeVar *rangevar; |
3103 | VacuumRelation *rel; |
3104 | List *rel_list; |
3105 | |
3106 | /* Let pgstat know what we're doing */ |
3107 | autovac_report_activity(tab); |
3108 | |
3109 | /* Set up one VacuumRelation target, identified by OID, for vacuum() */ |
3110 | rangevar = makeRangeVar(tab->at_nspname, tab->at_relname, -1); |
3111 | rel = makeVacuumRelation(rangevar, tab->at_relid, NIL); |
3112 | rel_list = list_make1(rel); |
3113 | |
3114 | vacuum(rel_list, &tab->at_params, bstrategy, true); |
3115 | } |
3116 | |
3117 | /* |
3118 | * autovac_report_activity |
3119 | * Report to pgstat what autovacuum is doing |
3120 | * |
3121 | * We send a SQL string corresponding to what the user would see if the |
3122 | * equivalent command was to be issued manually. |
3123 | * |
3124 | * Note we assume that we are going to report the next command as soon as we're |
3125 | * done with the current one, and exit right after the last one, so we don't |
3126 | * bother to report "<IDLE>" or some such. |
3127 | */ |
3128 | static void |
3129 | autovac_report_activity(autovac_table *tab) |
3130 | { |
3131 | #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56) |
3132 | char activity[MAX_AUTOVAC_ACTIV_LEN]; |
3133 | int len; |
3134 | |
3135 | /* Report the command and possible options */ |
3136 | if (tab->at_params.options & VACOPT_VACUUM) |
3137 | snprintf(activity, MAX_AUTOVAC_ACTIV_LEN, |
3138 | "autovacuum: VACUUM%s" , |
3139 | tab->at_params.options & VACOPT_ANALYZE ? " ANALYZE" : "" ); |
3140 | else |
3141 | snprintf(activity, MAX_AUTOVAC_ACTIV_LEN, |
3142 | "autovacuum: ANALYZE" ); |
3143 | |
3144 | /* |
3145 | * Report the qualified name of the relation. |
3146 | */ |
3147 | len = strlen(activity); |
3148 | |
3149 | snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len, |
3150 | " %s.%s%s" , tab->at_nspname, tab->at_relname, |
3151 | tab->at_params.is_wraparound ? " (to prevent wraparound)" : "" ); |
3152 | |
3153 | /* Set statement_timestamp() to current time for pg_stat_activity */ |
3154 | SetCurrentStatementStartTimestamp(); |
3155 | |
3156 | pgstat_report_activity(STATE_RUNNING, activity); |
3157 | } |
3158 | |
3159 | /* |
3160 | * autovac_report_workitem |
3161 | * Report to pgstat that autovacuum is processing a work item |
3162 | */ |
3163 | static void |
3164 | autovac_report_workitem(AutoVacuumWorkItem *workitem, |
3165 | const char *nspname, const char *relname) |
3166 | { |
3167 | char activity[MAX_AUTOVAC_ACTIV_LEN + 12 + 2]; |
3168 | char blk[12 + 2]; |
3169 | int len; |
3170 | |
3171 | switch (workitem->avw_type) |
3172 | { |
3173 | case AVW_BRINSummarizeRange: |
3174 | snprintf(activity, MAX_AUTOVAC_ACTIV_LEN, |
3175 | "autovacuum: BRIN summarize" ); |
3176 | break; |
3177 | } |
3178 | |
3179 | /* |
3180 | * Report the qualified name of the relation, and the block number if any |
3181 | */ |
3182 | len = strlen(activity); |
3183 | |
3184 | if (BlockNumberIsValid(workitem->avw_blockNumber)) |
3185 | snprintf(blk, sizeof(blk), " %u" , workitem->avw_blockNumber); |
3186 | else |
3187 | blk[0] = '\0'; |
3188 | |
3189 | snprintf(activity + len, MAX_AUTOVAC_ACTIV_LEN - len, |
3190 | " %s.%s%s" , nspname, relname, blk); |
3191 | |
3192 | /* Set statement_timestamp() to current time for pg_stat_activity */ |
3193 | SetCurrentStatementStartTimestamp(); |
3194 | |
3195 | pgstat_report_activity(STATE_RUNNING, activity); |
3196 | } |
3197 | |
3198 | /* |
3199 | * AutoVacuumingActive |
3200 | * Check GUC vars and report whether the autovacuum process should be |
3201 | * running. |
3202 | */ |
3203 | bool |
3204 | AutoVacuumingActive(void) |
3205 | { |
3206 | if (!autovacuum_start_daemon || !pgstat_track_counts) |
3207 | return false; |
3208 | return true; |
3209 | } |
3210 | |
3211 | /* |
3212 | * Request one work item to the next autovacuum run processing our database. |
3213 | * Return false if the request can't be recorded. |
3214 | */ |
3215 | bool |
3216 | AutoVacuumRequestWork(AutoVacuumWorkItemType type, Oid relationId, |
3217 | BlockNumber blkno) |
3218 | { |
3219 | int i; |
3220 | bool result = false; |
3221 | |
3222 | LWLockAcquire(AutovacuumLock, LW_EXCLUSIVE); |
3223 | |
3224 | /* |
3225 | * Locate an unused work item and fill it with the given data. |
3226 | */ |
3227 | for (i = 0; i < NUM_WORKITEMS; i++) |
3228 | { |
3229 | AutoVacuumWorkItem *workitem = &AutoVacuumShmem->av_workItems[i]; |
3230 | |
3231 | if (workitem->avw_used) |
3232 | continue; |
3233 | |
3234 | workitem->avw_used = true; |
3235 | workitem->avw_active = false; |
3236 | workitem->avw_type = type; |
3237 | workitem->avw_database = MyDatabaseId; |
3238 | workitem->avw_relation = relationId; |
3239 | workitem->avw_blockNumber = blkno; |
3240 | result = true; |
3241 | |
3242 | /* done */ |
3243 | break; |
3244 | } |
3245 | |
3246 | LWLockRelease(AutovacuumLock); |
3247 | |
3248 | return result; |
3249 | } |
3250 | |
3251 | /* |
3252 | * autovac_init |
3253 | * This is called at postmaster initialization. |
3254 | * |
3255 | * All we do here is annoy the user if he got it wrong. |
3256 | */ |
3257 | void |
3258 | autovac_init(void) |
3259 | { |
3260 | if (autovacuum_start_daemon && !pgstat_track_counts) |
3261 | ereport(WARNING, |
3262 | (errmsg("autovacuum not started because of misconfiguration" ), |
3263 | errhint("Enable the \"track_counts\" option." ))); |
3264 | } |
3265 | |
3266 | /* |
3267 | * IsAutoVacuum functions |
3268 | * Return whether this is either a launcher autovacuum process or a worker |
3269 | * process. |
3270 | */ |
3271 | bool |
3272 | IsAutoVacuumLauncherProcess(void) |
3273 | { |
3274 | return am_autovacuum_launcher; |
3275 | } |
3276 | |
3277 | bool |
3278 | IsAutoVacuumWorkerProcess(void) |
3279 | { |
3280 | return am_autovacuum_worker; |
3281 | } |
3282 | |
3283 | |
3284 | /* |
3285 | * AutoVacuumShmemSize |
3286 | * Compute space needed for autovacuum-related shared memory |
3287 | */ |
3288 | Size |
3289 | AutoVacuumShmemSize(void) |
3290 | { |
3291 | Size size; |
3292 | |
3293 | /* |
3294 | * Need the fixed struct and the array of WorkerInfoData. |
3295 | */ |
3296 | size = sizeof(AutoVacuumShmemStruct); |
3297 | size = MAXALIGN(size); |
3298 | size = add_size(size, mul_size(autovacuum_max_workers, |
3299 | sizeof(WorkerInfoData))); |
3300 | return size; |
3301 | } |
3302 | |
3303 | /* |
3304 | * AutoVacuumShmemInit |
3305 | * Allocate and initialize autovacuum-related shared memory |
3306 | */ |
3307 | void |
3308 | AutoVacuumShmemInit(void) |
3309 | { |
3310 | bool found; |
3311 | |
3312 | AutoVacuumShmem = (AutoVacuumShmemStruct *) |
3313 | ShmemInitStruct("AutoVacuum Data" , |
3314 | AutoVacuumShmemSize(), |
3315 | &found); |
3316 | |
3317 | if (!IsUnderPostmaster) |
3318 | { |
3319 | WorkerInfo worker; |
3320 | int i; |
3321 | |
3322 | Assert(!found); |
3323 | |
3324 | AutoVacuumShmem->av_launcherpid = 0; |
3325 | dlist_init(&AutoVacuumShmem->av_freeWorkers); |
3326 | dlist_init(&AutoVacuumShmem->av_runningWorkers); |
3327 | AutoVacuumShmem->av_startingWorker = NULL; |
3328 | memset(AutoVacuumShmem->av_workItems, 0, |
3329 | sizeof(AutoVacuumWorkItem) * NUM_WORKITEMS); |
3330 | |
3331 | worker = (WorkerInfo) ((char *) AutoVacuumShmem + |
3332 | MAXALIGN(sizeof(AutoVacuumShmemStruct))); |
3333 | |
3334 | /* initialize the WorkerInfo free list */ |
3335 | for (i = 0; i < autovacuum_max_workers; i++) |
3336 | dlist_push_head(&AutoVacuumShmem->av_freeWorkers, |
3337 | &worker[i].wi_links); |
3338 | } |
3339 | else |
3340 | Assert(found); |
3341 | } |
3342 | |
3343 | /* |
3344 | * autovac_refresh_stats |
3345 | * Refresh pgstats data for an autovacuum process |
3346 | * |
3347 | * Cause the next pgstats read operation to obtain fresh data, but throttle |
3348 | * such refreshing in the autovacuum launcher. This is mostly to avoid |
3349 | * rereading the pgstats files too many times in quick succession when there |
3350 | * are many databases. |
3351 | * |
3352 | * Note: we avoid throttling in the autovac worker, as it would be |
3353 | * counterproductive in the recheck logic. |
3354 | */ |
3355 | static void |
3356 | autovac_refresh_stats(void) |
3357 | { |
3358 | if (IsAutoVacuumLauncherProcess()) |
3359 | { |
3360 | static TimestampTz last_read = 0; |
3361 | TimestampTz current_time; |
3362 | |
3363 | current_time = GetCurrentTimestamp(); |
3364 | |
3365 | if (!TimestampDifferenceExceeds(last_read, current_time, |
3366 | STATS_READ_DELAY)) |
3367 | return; |
3368 | |
3369 | last_read = current_time; |
3370 | } |
3371 | |
3372 | pgstat_clear_snapshot(); |
3373 | } |
3374 | |