Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * vacuum.c
4 : * The postgres vacuum cleaner.
5 : *
6 : * This file includes (a) control and dispatch code for VACUUM and ANALYZE
7 : * commands, (b) code to compute various vacuum thresholds, and (c) index
8 : * vacuum code.
9 : *
10 : * VACUUM for heap AM is implemented in vacuumlazy.c, parallel vacuum in
11 : * vacuumparallel.c, ANALYZE in analyze.c, and VACUUM FULL is a variant of
12 : * CLUSTER, handled in cluster.c.
13 : *
14 : *
15 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
16 : * Portions Copyright (c) 1994, Regents of the University of California
17 : *
18 : *
19 : * IDENTIFICATION
20 : * src/backend/commands/vacuum.c
21 : *
22 : *-------------------------------------------------------------------------
23 : */
24 : #include "postgres.h"
25 :
26 : #include <math.h>
27 :
28 : #include "access/clog.h"
29 : #include "access/commit_ts.h"
30 : #include "access/genam.h"
31 : #include "access/heapam.h"
32 : #include "access/htup_details.h"
33 : #include "access/multixact.h"
34 : #include "access/tableam.h"
35 : #include "access/transam.h"
36 : #include "access/xact.h"
37 : #include "catalog/namespace.h"
38 : #include "catalog/pg_database.h"
39 : #include "catalog/pg_inherits.h"
40 : #include "commands/cluster.h"
41 : #include "commands/defrem.h"
42 : #include "commands/progress.h"
43 : #include "commands/vacuum.h"
44 : #include "miscadmin.h"
45 : #include "nodes/makefuncs.h"
46 : #include "pgstat.h"
47 : #include "postmaster/autovacuum.h"
48 : #include "postmaster/bgworker_internals.h"
49 : #include "postmaster/interrupt.h"
50 : #include "storage/bufmgr.h"
51 : #include "storage/lmgr.h"
52 : #include "storage/pmsignal.h"
53 : #include "storage/proc.h"
54 : #include "storage/procarray.h"
55 : #include "utils/acl.h"
56 : #include "utils/fmgroids.h"
57 : #include "utils/guc.h"
58 : #include "utils/guc_hooks.h"
59 : #include "utils/injection_point.h"
60 : #include "utils/memutils.h"
61 : #include "utils/snapmgr.h"
62 : #include "utils/syscache.h"
63 :
64 : /*
65 : * Minimum interval for cost-based vacuum delay reports from a parallel worker.
66 : * This aims to avoid sending too many messages and waking up the leader too
67 : * frequently.
68 : */
69 : #define PARALLEL_VACUUM_DELAY_REPORT_INTERVAL_NS (NS_PER_S)
70 :
71 : /*
72 : * GUC parameters
73 : */
74 : int vacuum_freeze_min_age;
75 : int vacuum_freeze_table_age;
76 : int vacuum_multixact_freeze_min_age;
77 : int vacuum_multixact_freeze_table_age;
78 : int vacuum_failsafe_age;
79 : int vacuum_multixact_failsafe_age;
80 : double vacuum_max_eager_freeze_failure_rate;
81 : bool track_cost_delay_timing;
82 : bool vacuum_truncate;
83 :
84 : /*
85 : * Variables for cost-based vacuum delay. The defaults differ between
86 : * autovacuum and vacuum. They should be set with the appropriate GUC value in
87 : * vacuum code. They are initialized here to the defaults for client backends
88 : * executing VACUUM or ANALYZE.
89 : */
90 : double vacuum_cost_delay = 0;
91 : int vacuum_cost_limit = 200;
92 :
93 : /* Variable for reporting cost-based vacuum delay from parallel workers. */
94 : int64 parallel_vacuum_worker_delay_ns = 0;
95 :
96 : /*
97 : * VacuumFailsafeActive is a defined as a global so that we can determine
98 : * whether or not to re-enable cost-based vacuum delay when vacuuming a table.
99 : * If failsafe mode has been engaged, we will not re-enable cost-based delay
100 : * for the table until after vacuuming has completed, regardless of other
101 : * settings.
102 : *
103 : * Only VACUUM code should inspect this variable and only table access methods
104 : * should set it to true. In Table AM-agnostic VACUUM code, this variable is
105 : * inspected to determine whether or not to allow cost-based delays. Table AMs
106 : * are free to set it if they desire this behavior, but it is false by default
107 : * and reset to false in between vacuuming each relation.
108 : */
109 : bool VacuumFailsafeActive = false;
110 :
111 : /*
112 : * Variables for cost-based parallel vacuum. See comments atop
113 : * compute_parallel_delay to understand how it works.
114 : */
115 : pg_atomic_uint32 *VacuumSharedCostBalance = NULL;
116 : pg_atomic_uint32 *VacuumActiveNWorkers = NULL;
117 : int VacuumCostBalanceLocal = 0;
118 :
119 : /* non-export function prototypes */
120 : static List *expand_vacuum_rel(VacuumRelation *vrel,
121 : MemoryContext vac_context, int options);
122 : static List *get_all_vacuum_rels(MemoryContext vac_context, int options);
123 : static void vac_truncate_clog(TransactionId frozenXID,
124 : MultiXactId minMulti,
125 : TransactionId lastSaneFrozenXid,
126 : MultiXactId lastSaneMinMulti);
127 : static bool vacuum_rel(Oid relid, RangeVar *relation, VacuumParams params,
128 : BufferAccessStrategy bstrategy);
129 : static double compute_parallel_delay(void);
130 : static VacOptValue get_vacoptval_from_boolean(DefElem *def);
131 : static bool vac_tid_reaped(ItemPointer itemptr, void *state);
132 :
133 : /*
134 : * GUC check function to ensure GUC value specified is within the allowable
135 : * range.
136 : */
137 : bool
138 2268 : check_vacuum_buffer_usage_limit(int *newval, void **extra,
139 : GucSource source)
140 : {
141 : /* Value upper and lower hard limits are inclusive */
142 2268 : if (*newval == 0 || (*newval >= MIN_BAS_VAC_RING_SIZE_KB &&
143 2268 : *newval <= MAX_BAS_VAC_RING_SIZE_KB))
144 2268 : return true;
145 :
146 : /* Value does not fall within any allowable range */
147 0 : GUC_check_errdetail("\"%s\" must be 0 or between %d kB and %d kB.",
148 : "vacuum_buffer_usage_limit",
149 : MIN_BAS_VAC_RING_SIZE_KB, MAX_BAS_VAC_RING_SIZE_KB);
150 :
151 0 : return false;
152 : }
153 :
154 : /*
155 : * Primary entry point for manual VACUUM and ANALYZE commands
156 : *
157 : * This is mainly a preparation wrapper for the real operations that will
158 : * happen in vacuum().
159 : */
160 : void
161 13772 : ExecVacuum(ParseState *pstate, VacuumStmt *vacstmt, bool isTopLevel)
162 : {
163 : VacuumParams params;
164 13772 : BufferAccessStrategy bstrategy = NULL;
165 13772 : bool verbose = false;
166 13772 : bool skip_locked = false;
167 13772 : bool analyze = false;
168 13772 : bool freeze = false;
169 13772 : bool full = false;
170 13772 : bool disable_page_skipping = false;
171 13772 : bool process_main = true;
172 13772 : bool process_toast = true;
173 : int ring_size;
174 13772 : bool skip_database_stats = false;
175 13772 : bool only_database_stats = false;
176 : MemoryContext vac_context;
177 : ListCell *lc;
178 :
179 : /* index_cleanup and truncate values unspecified for now */
180 13772 : params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
181 13772 : params.truncate = VACOPTVALUE_UNSPECIFIED;
182 :
183 : /* By default parallel vacuum is enabled */
184 13772 : params.nworkers = 0;
185 :
186 : /* Will be set later if we recurse to a TOAST table. */
187 13772 : params.toast_parent = InvalidOid;
188 :
189 : /*
190 : * Set this to an invalid value so it is clear whether or not a
191 : * BUFFER_USAGE_LIMIT was specified when making the access strategy.
192 : */
193 13772 : ring_size = -1;
194 :
195 : /* Parse options list */
196 28190 : foreach(lc, vacstmt->options)
197 : {
198 14454 : DefElem *opt = (DefElem *) lfirst(lc);
199 :
200 : /* Parse common options for VACUUM and ANALYZE */
201 14454 : if (strcmp(opt->defname, "verbose") == 0)
202 42 : verbose = defGetBoolean(opt);
203 14412 : else if (strcmp(opt->defname, "skip_locked") == 0)
204 334 : skip_locked = defGetBoolean(opt);
205 14078 : else if (strcmp(opt->defname, "buffer_usage_limit") == 0)
206 : {
207 : const char *hintmsg;
208 : int result;
209 : char *vac_buffer_size;
210 :
211 54 : vac_buffer_size = defGetString(opt);
212 :
213 : /*
214 : * Check that the specified value is valid and the size falls
215 : * within the hard upper and lower limits if it is not 0.
216 : */
217 54 : if (!parse_int(vac_buffer_size, &result, GUC_UNIT_KB, &hintmsg) ||
218 48 : (result != 0 &&
219 36 : (result < MIN_BAS_VAC_RING_SIZE_KB || result > MAX_BAS_VAC_RING_SIZE_KB)))
220 : {
221 18 : ereport(ERROR,
222 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
223 : errmsg("%s option must be 0 or between %d kB and %d kB",
224 : "BUFFER_USAGE_LIMIT",
225 : MIN_BAS_VAC_RING_SIZE_KB, MAX_BAS_VAC_RING_SIZE_KB),
226 : hintmsg ? errhint_internal("%s", _(hintmsg)) : 0));
227 : }
228 :
229 36 : ring_size = result;
230 : }
231 14024 : else if (!vacstmt->is_vacuumcmd)
232 6 : ereport(ERROR,
233 : (errcode(ERRCODE_SYNTAX_ERROR),
234 : errmsg("unrecognized ANALYZE option \"%s\"", opt->defname),
235 : parser_errposition(pstate, opt->location)));
236 :
237 : /* Parse options available on VACUUM */
238 14018 : else if (strcmp(opt->defname, "analyze") == 0)
239 2750 : analyze = defGetBoolean(opt);
240 11268 : else if (strcmp(opt->defname, "freeze") == 0)
241 2552 : freeze = defGetBoolean(opt);
242 8716 : else if (strcmp(opt->defname, "full") == 0)
243 392 : full = defGetBoolean(opt);
244 8324 : else if (strcmp(opt->defname, "disable_page_skipping") == 0)
245 214 : disable_page_skipping = defGetBoolean(opt);
246 8110 : else if (strcmp(opt->defname, "index_cleanup") == 0)
247 : {
248 : /* Interpret no string as the default, which is 'auto' */
249 174 : if (!opt->arg)
250 0 : params.index_cleanup = VACOPTVALUE_AUTO;
251 : else
252 : {
253 174 : char *sval = defGetString(opt);
254 :
255 : /* Try matching on 'auto' string, or fall back on boolean */
256 174 : if (pg_strcasecmp(sval, "auto") == 0)
257 6 : params.index_cleanup = VACOPTVALUE_AUTO;
258 : else
259 168 : params.index_cleanup = get_vacoptval_from_boolean(opt);
260 : }
261 : }
262 7936 : else if (strcmp(opt->defname, "process_main") == 0)
263 154 : process_main = defGetBoolean(opt);
264 7782 : else if (strcmp(opt->defname, "process_toast") == 0)
265 160 : process_toast = defGetBoolean(opt);
266 7622 : else if (strcmp(opt->defname, "truncate") == 0)
267 158 : params.truncate = get_vacoptval_from_boolean(opt);
268 7464 : else if (strcmp(opt->defname, "parallel") == 0)
269 : {
270 352 : int nworkers = defGetInt32(opt);
271 :
272 346 : if (nworkers < 0 || nworkers > MAX_PARALLEL_WORKER_LIMIT)
273 6 : ereport(ERROR,
274 : (errcode(ERRCODE_SYNTAX_ERROR),
275 : errmsg("%s option must be between 0 and %d",
276 : "PARALLEL",
277 : MAX_PARALLEL_WORKER_LIMIT),
278 : parser_errposition(pstate, opt->location)));
279 :
280 : /*
281 : * Disable parallel vacuum, if user has specified parallel degree
282 : * as zero.
283 : */
284 340 : if (nworkers == 0)
285 156 : params.nworkers = -1;
286 : else
287 184 : params.nworkers = nworkers;
288 : }
289 7112 : else if (strcmp(opt->defname, "skip_database_stats") == 0)
290 6982 : skip_database_stats = defGetBoolean(opt);
291 130 : else if (strcmp(opt->defname, "only_database_stats") == 0)
292 130 : only_database_stats = defGetBoolean(opt);
293 : else
294 0 : ereport(ERROR,
295 : (errcode(ERRCODE_SYNTAX_ERROR),
296 : errmsg("unrecognized VACUUM option \"%s\"", opt->defname),
297 : parser_errposition(pstate, opt->location)));
298 : }
299 :
300 : /* Set vacuum options */
301 13736 : params.options =
302 13736 : (vacstmt->is_vacuumcmd ? VACOPT_VACUUM : VACOPT_ANALYZE) |
303 13736 : (verbose ? VACOPT_VERBOSE : 0) |
304 13736 : (skip_locked ? VACOPT_SKIP_LOCKED : 0) |
305 13736 : (analyze ? VACOPT_ANALYZE : 0) |
306 13736 : (freeze ? VACOPT_FREEZE : 0) |
307 13736 : (full ? VACOPT_FULL : 0) |
308 13736 : (disable_page_skipping ? VACOPT_DISABLE_PAGE_SKIPPING : 0) |
309 13736 : (process_main ? VACOPT_PROCESS_MAIN : 0) |
310 13736 : (process_toast ? VACOPT_PROCESS_TOAST : 0) |
311 13736 : (skip_database_stats ? VACOPT_SKIP_DATABASE_STATS : 0) |
312 13736 : (only_database_stats ? VACOPT_ONLY_DATABASE_STATS : 0);
313 :
314 : /* sanity checks on options */
315 : Assert(params.options & (VACOPT_VACUUM | VACOPT_ANALYZE));
316 : Assert((params.options & VACOPT_VACUUM) ||
317 : !(params.options & (VACOPT_FULL | VACOPT_FREEZE)));
318 :
319 13736 : if ((params.options & VACOPT_FULL) && params.nworkers > 0)
320 6 : ereport(ERROR,
321 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
322 : errmsg("VACUUM FULL cannot be performed in parallel")));
323 :
324 : /*
325 : * BUFFER_USAGE_LIMIT does nothing for VACUUM (FULL) so just raise an
326 : * ERROR for that case. VACUUM (FULL, ANALYZE) does make use of it, so
327 : * we'll permit that.
328 : */
329 13730 : if (ring_size != -1 && (params.options & VACOPT_FULL) &&
330 6 : !(params.options & VACOPT_ANALYZE))
331 6 : ereport(ERROR,
332 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
333 : errmsg("BUFFER_USAGE_LIMIT cannot be specified for VACUUM FULL")));
334 :
335 : /*
336 : * Make sure VACOPT_ANALYZE is specified if any column lists are present.
337 : */
338 13724 : if (!(params.options & VACOPT_ANALYZE))
339 : {
340 11968 : foreach(lc, vacstmt->rels)
341 : {
342 5888 : VacuumRelation *vrel = lfirst_node(VacuumRelation, lc);
343 :
344 5888 : if (vrel->va_cols != NIL)
345 6 : ereport(ERROR,
346 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
347 : errmsg("ANALYZE option must be specified when a column list is provided")));
348 : }
349 : }
350 :
351 :
352 : /*
353 : * Sanity check DISABLE_PAGE_SKIPPING option.
354 : */
355 13718 : if ((params.options & VACOPT_FULL) != 0 &&
356 368 : (params.options & VACOPT_DISABLE_PAGE_SKIPPING) != 0)
357 0 : ereport(ERROR,
358 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
359 : errmsg("VACUUM option DISABLE_PAGE_SKIPPING cannot be used with FULL")));
360 :
361 : /* sanity check for PROCESS_TOAST */
362 13718 : if ((params.options & VACOPT_FULL) != 0 &&
363 368 : (params.options & VACOPT_PROCESS_TOAST) == 0)
364 6 : ereport(ERROR,
365 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
366 : errmsg("PROCESS_TOAST required with VACUUM FULL")));
367 :
368 : /* sanity check for ONLY_DATABASE_STATS */
369 13712 : if (params.options & VACOPT_ONLY_DATABASE_STATS)
370 : {
371 : Assert(params.options & VACOPT_VACUUM);
372 130 : if (vacstmt->rels != NIL)
373 6 : ereport(ERROR,
374 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
375 : errmsg("ONLY_DATABASE_STATS cannot be specified with a list of tables")));
376 : /* don't require people to turn off PROCESS_TOAST/MAIN explicitly */
377 124 : if (params.options & ~(VACOPT_VACUUM |
378 : VACOPT_VERBOSE |
379 : VACOPT_PROCESS_MAIN |
380 : VACOPT_PROCESS_TOAST |
381 : VACOPT_ONLY_DATABASE_STATS))
382 0 : ereport(ERROR,
383 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
384 : errmsg("ONLY_DATABASE_STATS cannot be specified with other VACUUM options")));
385 : }
386 :
387 : /*
388 : * All freeze ages are zero if the FREEZE option is given; otherwise pass
389 : * them as -1 which means to use the default values.
390 : */
391 13706 : if (params.options & VACOPT_FREEZE)
392 : {
393 2552 : params.freeze_min_age = 0;
394 2552 : params.freeze_table_age = 0;
395 2552 : params.multixact_freeze_min_age = 0;
396 2552 : params.multixact_freeze_table_age = 0;
397 : }
398 : else
399 : {
400 11154 : params.freeze_min_age = -1;
401 11154 : params.freeze_table_age = -1;
402 11154 : params.multixact_freeze_min_age = -1;
403 11154 : params.multixact_freeze_table_age = -1;
404 : }
405 :
406 : /* user-invoked vacuum is never "for wraparound" */
407 13706 : params.is_wraparound = false;
408 :
409 : /*
410 : * user-invoked vacuum uses VACOPT_VERBOSE instead of
411 : * log_vacuum_min_duration and log_analyze_min_duration
412 : */
413 13706 : params.log_vacuum_min_duration = -1;
414 13706 : params.log_analyze_min_duration = -1;
415 :
416 : /*
417 : * Later, in vacuum_rel(), we check if a reloption override was specified.
418 : */
419 13706 : params.max_eager_freeze_failure_rate = vacuum_max_eager_freeze_failure_rate;
420 :
421 : /*
422 : * Create special memory context for cross-transaction storage.
423 : *
424 : * Since it is a child of PortalContext, it will go away eventually even
425 : * if we suffer an error; there's no need for special abort cleanup logic.
426 : */
427 13706 : vac_context = AllocSetContextCreate(PortalContext,
428 : "Vacuum",
429 : ALLOCSET_DEFAULT_SIZES);
430 :
431 : /*
432 : * Make a buffer strategy object in the cross-transaction memory context.
433 : * We needn't bother making this for VACUUM (FULL) or VACUUM
434 : * (ONLY_DATABASE_STATS) as they'll not make use of it. VACUUM (FULL,
435 : * ANALYZE) is possible, so we'd better ensure that we make a strategy
436 : * when we see ANALYZE.
437 : */
438 13706 : if ((params.options & (VACOPT_ONLY_DATABASE_STATS |
439 486 : VACOPT_FULL)) == 0 ||
440 486 : (params.options & VACOPT_ANALYZE) != 0)
441 : {
442 :
443 13226 : MemoryContext old_context = MemoryContextSwitchTo(vac_context);
444 :
445 : Assert(ring_size >= -1);
446 :
447 : /*
448 : * If BUFFER_USAGE_LIMIT was specified by the VACUUM or ANALYZE
449 : * command, it overrides the value of VacuumBufferUsageLimit. Either
450 : * value may be 0, in which case GetAccessStrategyWithSize() will
451 : * return NULL, effectively allowing full use of shared buffers.
452 : */
453 13226 : if (ring_size == -1)
454 13196 : ring_size = VacuumBufferUsageLimit;
455 :
456 13226 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, ring_size);
457 :
458 13226 : MemoryContextSwitchTo(old_context);
459 : }
460 :
461 : /* Now go through the common routine */
462 13706 : vacuum(vacstmt->rels, params, bstrategy, vac_context, isTopLevel);
463 :
464 : /* Finally, clean up the vacuum memory context */
465 13572 : MemoryContextDelete(vac_context);
466 13572 : }
467 :
468 : /*
469 : * Internal entry point for autovacuum and the VACUUM / ANALYZE commands.
470 : *
471 : * relations, if not NIL, is a list of VacuumRelation to process; otherwise,
472 : * we process all relevant tables in the database. For each VacuumRelation,
473 : * if a valid OID is supplied, the table with that OID is what to process;
474 : * otherwise, the VacuumRelation's RangeVar indicates what to process.
475 : *
476 : * params contains a set of parameters that can be used to customize the
477 : * behavior.
478 : *
479 : * bstrategy may be passed in as NULL when the caller does not want to
480 : * restrict the number of shared_buffers that VACUUM / ANALYZE can use,
481 : * otherwise, the caller must build a BufferAccessStrategy with the number of
482 : * shared_buffers that VACUUM / ANALYZE should try to limit themselves to
483 : * using.
484 : *
485 : * isTopLevel should be passed down from ProcessUtility.
486 : *
487 : * It is the caller's responsibility that all parameters are allocated in a
488 : * memory context that will not disappear at transaction commit.
489 : */
490 : void
491 259664 : vacuum(List *relations, const VacuumParams params, BufferAccessStrategy bstrategy,
492 : MemoryContext vac_context, bool isTopLevel)
493 : {
494 : static bool in_vacuum = false;
495 :
496 : const char *stmttype;
497 : volatile bool in_outer_xact,
498 : use_own_xacts;
499 :
500 259664 : stmttype = (params.options & VACOPT_VACUUM) ? "VACUUM" : "ANALYZE";
501 :
502 : /*
503 : * We cannot run VACUUM inside a user transaction block; if we were inside
504 : * a transaction, then our commit- and start-transaction-command calls
505 : * would not have the intended effect! There are numerous other subtle
506 : * dependencies on this, too.
507 : *
508 : * ANALYZE (without VACUUM) can run either way.
509 : */
510 259664 : if (params.options & VACOPT_VACUUM)
511 : {
512 254458 : PreventInTransactionBlock(isTopLevel, stmttype);
513 254438 : in_outer_xact = false;
514 : }
515 : else
516 5206 : in_outer_xact = IsInTransactionBlock(isTopLevel);
517 :
518 : /*
519 : * Check for and disallow recursive calls. This could happen when VACUUM
520 : * FULL or ANALYZE calls a hostile index expression that itself calls
521 : * ANALYZE.
522 : */
523 259644 : if (in_vacuum)
524 12 : ereport(ERROR,
525 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
526 : errmsg("%s cannot be executed from VACUUM or ANALYZE",
527 : stmttype)));
528 :
529 : /*
530 : * Build list of relation(s) to process, putting any new data in
531 : * vac_context for safekeeping.
532 : */
533 259632 : if (params.options & VACOPT_ONLY_DATABASE_STATS)
534 : {
535 : /* We don't process any tables in this case */
536 : Assert(relations == NIL);
537 : }
538 259508 : else if (relations != NIL)
539 : {
540 259290 : List *newrels = NIL;
541 : ListCell *lc;
542 :
543 518684 : foreach(lc, relations)
544 : {
545 259430 : VacuumRelation *vrel = lfirst_node(VacuumRelation, lc);
546 : List *sublist;
547 : MemoryContext old_context;
548 :
549 259430 : sublist = expand_vacuum_rel(vrel, vac_context, params.options);
550 259394 : old_context = MemoryContextSwitchTo(vac_context);
551 259394 : newrels = list_concat(newrels, sublist);
552 259394 : MemoryContextSwitchTo(old_context);
553 : }
554 259254 : relations = newrels;
555 : }
556 : else
557 218 : relations = get_all_vacuum_rels(vac_context, params.options);
558 :
559 : /*
560 : * Decide whether we need to start/commit our own transactions.
561 : *
562 : * For VACUUM (with or without ANALYZE): always do so, so that we can
563 : * release locks as soon as possible. (We could possibly use the outer
564 : * transaction for a one-table VACUUM, but handling TOAST tables would be
565 : * problematic.)
566 : *
567 : * For ANALYZE (no VACUUM): if inside a transaction block, we cannot
568 : * start/commit our own transactions. Also, there's no need to do so if
569 : * only processing one relation. For multiple relations when not within a
570 : * transaction block, and also in an autovacuum worker, use own
571 : * transactions so we can release locks sooner.
572 : */
573 259596 : if (params.options & VACOPT_VACUUM)
574 254426 : use_own_xacts = true;
575 : else
576 : {
577 : Assert(params.options & VACOPT_ANALYZE);
578 5170 : if (AmAutoVacuumWorkerProcess())
579 318 : use_own_xacts = true;
580 4852 : else if (in_outer_xact)
581 238 : use_own_xacts = false;
582 4614 : else if (list_length(relations) > 1)
583 810 : use_own_xacts = true;
584 : else
585 3804 : use_own_xacts = false;
586 : }
587 :
588 : /*
589 : * vacuum_rel expects to be entered with no transaction active; it will
590 : * start and commit its own transaction. But we are called by an SQL
591 : * command, and so we are executing inside a transaction already. We
592 : * commit the transaction started in PostgresMain() here, and start
593 : * another one before exiting to match the commit waiting for us back in
594 : * PostgresMain().
595 : */
596 259596 : if (use_own_xacts)
597 : {
598 : Assert(!in_outer_xact);
599 :
600 : /* ActiveSnapshot is not set by autovacuum */
601 255554 : if (ActiveSnapshotSet())
602 9596 : PopActiveSnapshot();
603 :
604 : /* matches the StartTransaction in PostgresMain() */
605 255554 : CommitTransactionCommand();
606 : }
607 :
608 : /* Turn vacuum cost accounting on or off, and set/clear in_vacuum */
609 259596 : PG_TRY();
610 : {
611 : ListCell *cur;
612 :
613 259596 : in_vacuum = true;
614 259596 : VacuumFailsafeActive = false;
615 259596 : VacuumUpdateCosts();
616 259596 : VacuumCostBalance = 0;
617 259596 : VacuumCostBalanceLocal = 0;
618 259596 : VacuumSharedCostBalance = NULL;
619 259596 : VacuumActiveNWorkers = NULL;
620 :
621 : /*
622 : * Loop to process each selected relation.
623 : */
624 536266 : foreach(cur, relations)
625 : {
626 276736 : VacuumRelation *vrel = lfirst_node(VacuumRelation, cur);
627 :
628 276736 : if (params.options & VACOPT_VACUUM)
629 : {
630 263256 : if (!vacuum_rel(vrel->oid, vrel->relation, params, bstrategy))
631 100 : continue;
632 : }
633 :
634 276628 : if (params.options & VACOPT_ANALYZE)
635 : {
636 : /*
637 : * If using separate xacts, start one for analyze. Otherwise,
638 : * we can use the outer transaction.
639 : */
640 16514 : if (use_own_xacts)
641 : {
642 12522 : StartTransactionCommand();
643 : /* functions in indexes may want a snapshot set */
644 12522 : PushActiveSnapshot(GetTransactionSnapshot());
645 : }
646 :
647 16514 : analyze_rel(vrel->oid, vrel->relation, params,
648 : vrel->va_cols, in_outer_xact, bstrategy);
649 :
650 16456 : if (use_own_xacts)
651 : {
652 12484 : PopActiveSnapshot();
653 : /* standard_ProcessUtility() does CCI if !use_own_xacts */
654 12484 : CommandCounterIncrement();
655 12484 : CommitTransactionCommand();
656 : }
657 : else
658 : {
659 : /*
660 : * If we're not using separate xacts, better separate the
661 : * ANALYZE actions with CCIs. This avoids trouble if user
662 : * says "ANALYZE t, t".
663 : */
664 3972 : CommandCounterIncrement();
665 : }
666 : }
667 :
668 : /*
669 : * Ensure VacuumFailsafeActive has been reset before vacuuming the
670 : * next relation.
671 : */
672 276570 : VacuumFailsafeActive = false;
673 : }
674 : }
675 66 : PG_FINALLY();
676 : {
677 259596 : in_vacuum = false;
678 259596 : VacuumCostActive = false;
679 259596 : VacuumFailsafeActive = false;
680 259596 : VacuumCostBalance = 0;
681 : }
682 259596 : PG_END_TRY();
683 :
684 : /*
685 : * Finish up processing.
686 : */
687 259530 : if (use_own_xacts)
688 : {
689 : /* here, we are not in a transaction */
690 :
691 : /*
692 : * This matches the CommitTransaction waiting for us in
693 : * PostgresMain().
694 : */
695 255508 : StartTransactionCommand();
696 : }
697 :
698 259530 : if ((params.options & VACOPT_VACUUM) &&
699 254392 : !(params.options & VACOPT_SKIP_DATABASE_STATS))
700 : {
701 : /*
702 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible.
703 : */
704 1772 : vac_update_datfrozenxid();
705 : }
706 :
707 259530 : }
708 :
709 : /*
710 : * Check if the current user has privileges to vacuum or analyze the relation.
711 : * If not, issue a WARNING log message and return false to let the caller
712 : * decide what to do with this relation. This routine is used to decide if a
713 : * relation can be processed for VACUUM or ANALYZE.
714 : */
715 : bool
716 320534 : vacuum_is_permitted_for_relation(Oid relid, Form_pg_class reltuple,
717 : bits32 options)
718 : {
719 : char *relname;
720 :
721 : Assert((options & (VACOPT_VACUUM | VACOPT_ANALYZE)) != 0);
722 :
723 : /*----------
724 : * A role has privileges to vacuum or analyze the relation if any of the
725 : * following are true:
726 : * - the role owns the current database and the relation is not shared
727 : * - the role has the MAINTAIN privilege on the relation
728 : *----------
729 : */
730 320534 : if ((object_ownercheck(DatabaseRelationId, MyDatabaseId, GetUserId()) &&
731 370848 : !reltuple->relisshared) ||
732 53884 : pg_class_aclcheck(relid, GetUserId(), ACL_MAINTAIN) == ACLCHECK_OK)
733 317440 : return true;
734 :
735 3094 : relname = NameStr(reltuple->relname);
736 :
737 3094 : if ((options & VACOPT_VACUUM) != 0)
738 : {
739 224 : ereport(WARNING,
740 : (errmsg("permission denied to vacuum \"%s\", skipping it",
741 : relname)));
742 :
743 : /*
744 : * For VACUUM ANALYZE, both logs could show up, but just generate
745 : * information for VACUUM as that would be the first one to be
746 : * processed.
747 : */
748 224 : return false;
749 : }
750 :
751 2870 : if ((options & VACOPT_ANALYZE) != 0)
752 2870 : ereport(WARNING,
753 : (errmsg("permission denied to analyze \"%s\", skipping it",
754 : relname)));
755 :
756 2870 : return false;
757 : }
758 :
759 :
760 : /*
761 : * vacuum_open_relation
762 : *
763 : * This routine is used for attempting to open and lock a relation which
764 : * is going to be vacuumed or analyzed. If the relation cannot be opened
765 : * or locked, a log is emitted if possible.
766 : */
767 : Relation
768 288946 : vacuum_open_relation(Oid relid, RangeVar *relation, bits32 options,
769 : bool verbose, LOCKMODE lmode)
770 : {
771 : Relation rel;
772 288946 : bool rel_lock = true;
773 : int elevel;
774 :
775 : Assert((options & (VACOPT_VACUUM | VACOPT_ANALYZE)) != 0);
776 :
777 : /*
778 : * Open the relation and get the appropriate lock on it.
779 : *
780 : * There's a race condition here: the relation may have gone away since
781 : * the last time we saw it. If so, we don't need to vacuum or analyze it.
782 : *
783 : * If we've been asked not to wait for the relation lock, acquire it first
784 : * in non-blocking mode, before calling try_relation_open().
785 : */
786 288946 : if (!(options & VACOPT_SKIP_LOCKED))
787 287674 : rel = try_relation_open(relid, lmode);
788 1272 : else if (ConditionalLockRelationOid(relid, lmode))
789 1250 : rel = try_relation_open(relid, NoLock);
790 : else
791 : {
792 22 : rel = NULL;
793 22 : rel_lock = false;
794 : }
795 :
796 : /* if relation is opened, leave */
797 288946 : if (rel)
798 288912 : return rel;
799 :
800 : /*
801 : * Relation could not be opened, hence generate if possible a log
802 : * informing on the situation.
803 : *
804 : * If the RangeVar is not defined, we do not have enough information to
805 : * provide a meaningful log statement. Chances are that the caller has
806 : * intentionally not provided this information so that this logging is
807 : * skipped, anyway.
808 : */
809 34 : if (relation == NULL)
810 18 : return NULL;
811 :
812 : /*
813 : * Determine the log level.
814 : *
815 : * For manual VACUUM or ANALYZE, we emit a WARNING to match the log
816 : * statements in the permission checks; otherwise, only log if the caller
817 : * so requested.
818 : */
819 16 : if (!AmAutoVacuumWorkerProcess())
820 14 : elevel = WARNING;
821 2 : else if (verbose)
822 2 : elevel = LOG;
823 : else
824 0 : return NULL;
825 :
826 16 : if ((options & VACOPT_VACUUM) != 0)
827 : {
828 10 : if (!rel_lock)
829 6 : ereport(elevel,
830 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
831 : errmsg("skipping vacuum of \"%s\" --- lock not available",
832 : relation->relname)));
833 : else
834 4 : ereport(elevel,
835 : (errcode(ERRCODE_UNDEFINED_TABLE),
836 : errmsg("skipping vacuum of \"%s\" --- relation no longer exists",
837 : relation->relname)));
838 :
839 : /*
840 : * For VACUUM ANALYZE, both logs could show up, but just generate
841 : * information for VACUUM as that would be the first one to be
842 : * processed.
843 : */
844 10 : return NULL;
845 : }
846 :
847 6 : if ((options & VACOPT_ANALYZE) != 0)
848 : {
849 6 : if (!rel_lock)
850 4 : ereport(elevel,
851 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
852 : errmsg("skipping analyze of \"%s\" --- lock not available",
853 : relation->relname)));
854 : else
855 2 : ereport(elevel,
856 : (errcode(ERRCODE_UNDEFINED_TABLE),
857 : errmsg("skipping analyze of \"%s\" --- relation no longer exists",
858 : relation->relname)));
859 : }
860 :
861 6 : return NULL;
862 : }
863 :
864 :
865 : /*
866 : * Given a VacuumRelation, fill in the table OID if it wasn't specified,
867 : * and optionally add VacuumRelations for partitions or inheritance children.
868 : *
869 : * If a VacuumRelation does not have an OID supplied and is a partitioned
870 : * table, an extra entry will be added to the output for each partition.
871 : * Presently, only autovacuum supplies OIDs when calling vacuum(), and
872 : * it does not want us to expand partitioned tables.
873 : *
874 : * We take care not to modify the input data structure, but instead build
875 : * new VacuumRelation(s) to return. (But note that they will reference
876 : * unmodified parts of the input, eg column lists.) New data structures
877 : * are made in vac_context.
878 : */
879 : static List *
880 259430 : expand_vacuum_rel(VacuumRelation *vrel, MemoryContext vac_context,
881 : int options)
882 : {
883 259430 : List *vacrels = NIL;
884 : MemoryContext oldcontext;
885 :
886 : /* If caller supplied OID, there's nothing we need do here. */
887 259430 : if (OidIsValid(vrel->oid))
888 : {
889 245958 : oldcontext = MemoryContextSwitchTo(vac_context);
890 245958 : vacrels = lappend(vacrels, vrel);
891 245958 : MemoryContextSwitchTo(oldcontext);
892 : }
893 : else
894 : {
895 : /*
896 : * Process a specific relation, and possibly partitions or child
897 : * tables thereof.
898 : */
899 : Oid relid;
900 : HeapTuple tuple;
901 : Form_pg_class classForm;
902 : bool include_children;
903 : bool is_partitioned_table;
904 : int rvr_opts;
905 :
906 : /*
907 : * Since autovacuum workers supply OIDs when calling vacuum(), no
908 : * autovacuum worker should reach this code.
909 : */
910 : Assert(!AmAutoVacuumWorkerProcess());
911 :
912 : /*
913 : * We transiently take AccessShareLock to protect the syscache lookup
914 : * below, as well as find_all_inheritors's expectation that the caller
915 : * holds some lock on the starting relation.
916 : */
917 13472 : rvr_opts = (options & VACOPT_SKIP_LOCKED) ? RVR_SKIP_LOCKED : 0;
918 13472 : relid = RangeVarGetRelidExtended(vrel->relation,
919 : AccessShareLock,
920 : rvr_opts,
921 : NULL, NULL);
922 :
923 : /*
924 : * If the lock is unavailable, emit the same log statement that
925 : * vacuum_rel() and analyze_rel() would.
926 : */
927 13436 : if (!OidIsValid(relid))
928 : {
929 8 : if (options & VACOPT_VACUUM)
930 6 : ereport(WARNING,
931 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
932 : errmsg("skipping vacuum of \"%s\" --- lock not available",
933 : vrel->relation->relname)));
934 : else
935 2 : ereport(WARNING,
936 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
937 : errmsg("skipping analyze of \"%s\" --- lock not available",
938 : vrel->relation->relname)));
939 8 : return vacrels;
940 : }
941 :
942 : /*
943 : * To check whether the relation is a partitioned table and its
944 : * ownership, fetch its syscache entry.
945 : */
946 13428 : tuple = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
947 13428 : if (!HeapTupleIsValid(tuple))
948 0 : elog(ERROR, "cache lookup failed for relation %u", relid);
949 13428 : classForm = (Form_pg_class) GETSTRUCT(tuple);
950 :
951 : /*
952 : * Make a returnable VacuumRelation for this rel if the user has the
953 : * required privileges.
954 : */
955 13428 : if (vacuum_is_permitted_for_relation(relid, classForm, options))
956 : {
957 13196 : oldcontext = MemoryContextSwitchTo(vac_context);
958 13196 : vacrels = lappend(vacrels, makeVacuumRelation(vrel->relation,
959 : relid,
960 : vrel->va_cols));
961 13196 : MemoryContextSwitchTo(oldcontext);
962 : }
963 :
964 : /*
965 : * Vacuuming a partitioned table with ONLY will not do anything since
966 : * the partitioned table itself is empty. Issue a warning if the user
967 : * requests this.
968 : */
969 13428 : include_children = vrel->relation->inh;
970 13428 : is_partitioned_table = (classForm->relkind == RELKIND_PARTITIONED_TABLE);
971 13428 : if ((options & VACOPT_VACUUM) && is_partitioned_table && !include_children)
972 6 : ereport(WARNING,
973 : (errmsg("VACUUM ONLY of partitioned table \"%s\" has no effect",
974 : vrel->relation->relname)));
975 :
976 13428 : ReleaseSysCache(tuple);
977 :
978 : /*
979 : * Unless the user has specified ONLY, make relation list entries for
980 : * its partitions or inheritance child tables. Note that the list
981 : * returned by find_all_inheritors() includes the passed-in OID, so we
982 : * have to skip that. There's no point in taking locks on the
983 : * individual partitions or child tables yet, and doing so would just
984 : * add unnecessary deadlock risk. For this last reason, we do not yet
985 : * check the ownership of the partitions/tables, which get added to
986 : * the list to process. Ownership will be checked later on anyway.
987 : */
988 13428 : if (include_children)
989 : {
990 13398 : List *part_oids = find_all_inheritors(relid, NoLock, NULL);
991 : ListCell *part_lc;
992 :
993 28986 : foreach(part_lc, part_oids)
994 : {
995 15588 : Oid part_oid = lfirst_oid(part_lc);
996 :
997 15588 : if (part_oid == relid)
998 13398 : continue; /* ignore original table */
999 :
1000 : /*
1001 : * We omit a RangeVar since it wouldn't be appropriate to
1002 : * complain about failure to open one of these relations
1003 : * later.
1004 : */
1005 2190 : oldcontext = MemoryContextSwitchTo(vac_context);
1006 2190 : vacrels = lappend(vacrels, makeVacuumRelation(NULL,
1007 : part_oid,
1008 : vrel->va_cols));
1009 2190 : MemoryContextSwitchTo(oldcontext);
1010 : }
1011 : }
1012 :
1013 : /*
1014 : * Release lock again. This means that by the time we actually try to
1015 : * process the table, it might be gone or renamed. In the former case
1016 : * we'll silently ignore it; in the latter case we'll process it
1017 : * anyway, but we must beware that the RangeVar doesn't necessarily
1018 : * identify it anymore. This isn't ideal, perhaps, but there's little
1019 : * practical alternative, since we're typically going to commit this
1020 : * transaction and begin a new one between now and then. Moreover,
1021 : * holding locks on multiple relations would create significant risk
1022 : * of deadlock.
1023 : */
1024 13428 : UnlockRelationOid(relid, AccessShareLock);
1025 : }
1026 :
1027 259386 : return vacrels;
1028 : }
1029 :
1030 : /*
1031 : * Construct a list of VacuumRelations for all vacuumable rels in
1032 : * the current database. The list is built in vac_context.
1033 : */
1034 : static List *
1035 218 : get_all_vacuum_rels(MemoryContext vac_context, int options)
1036 : {
1037 218 : List *vacrels = NIL;
1038 : Relation pgclass;
1039 : TableScanDesc scan;
1040 : HeapTuple tuple;
1041 :
1042 218 : pgclass = table_open(RelationRelationId, AccessShareLock);
1043 :
1044 218 : scan = table_beginscan_catalog(pgclass, 0, NULL);
1045 :
1046 98792 : while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
1047 : {
1048 98574 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
1049 : MemoryContext oldcontext;
1050 98574 : Oid relid = classForm->oid;
1051 :
1052 : /*
1053 : * We include partitioned tables here; depending on which operation is
1054 : * to be performed, caller will decide whether to process or ignore
1055 : * them.
1056 : */
1057 98574 : if (classForm->relkind != RELKIND_RELATION &&
1058 80592 : classForm->relkind != RELKIND_MATVIEW &&
1059 80544 : classForm->relkind != RELKIND_PARTITIONED_TABLE)
1060 80380 : continue;
1061 :
1062 : /* check permissions of relation */
1063 18194 : if (!vacuum_is_permitted_for_relation(relid, classForm, options))
1064 2754 : continue;
1065 :
1066 : /*
1067 : * Build VacuumRelation(s) specifying the table OIDs to be processed.
1068 : * We omit a RangeVar since it wouldn't be appropriate to complain
1069 : * about failure to open one of these relations later.
1070 : */
1071 15440 : oldcontext = MemoryContextSwitchTo(vac_context);
1072 15440 : vacrels = lappend(vacrels, makeVacuumRelation(NULL,
1073 : relid,
1074 : NIL));
1075 15440 : MemoryContextSwitchTo(oldcontext);
1076 : }
1077 :
1078 218 : table_endscan(scan);
1079 218 : table_close(pgclass, AccessShareLock);
1080 :
1081 218 : return vacrels;
1082 : }
1083 :
1084 : /*
1085 : * vacuum_get_cutoffs() -- compute OldestXmin and freeze cutoff points
1086 : *
1087 : * The target relation and VACUUM parameters are our inputs.
1088 : *
1089 : * Output parameters are the cutoffs that VACUUM caller should use.
1090 : *
1091 : * Return value indicates if vacuumlazy.c caller should make its VACUUM
1092 : * operation aggressive. An aggressive VACUUM must advance relfrozenxid up to
1093 : * FreezeLimit (at a minimum), and relminmxid up to MultiXactCutoff (at a
1094 : * minimum).
1095 : */
1096 : bool
1097 272184 : vacuum_get_cutoffs(Relation rel, const VacuumParams params,
1098 : struct VacuumCutoffs *cutoffs)
1099 : {
1100 : int freeze_min_age,
1101 : multixact_freeze_min_age,
1102 : freeze_table_age,
1103 : multixact_freeze_table_age,
1104 : effective_multixact_freeze_max_age;
1105 : TransactionId nextXID,
1106 : safeOldestXmin,
1107 : aggressiveXIDCutoff;
1108 : MultiXactId nextMXID,
1109 : safeOldestMxact,
1110 : aggressiveMXIDCutoff;
1111 :
1112 : /* Use mutable copies of freeze age parameters */
1113 272184 : freeze_min_age = params.freeze_min_age;
1114 272184 : multixact_freeze_min_age = params.multixact_freeze_min_age;
1115 272184 : freeze_table_age = params.freeze_table_age;
1116 272184 : multixact_freeze_table_age = params.multixact_freeze_table_age;
1117 :
1118 : /* Set pg_class fields in cutoffs */
1119 272184 : cutoffs->relfrozenxid = rel->rd_rel->relfrozenxid;
1120 272184 : cutoffs->relminmxid = rel->rd_rel->relminmxid;
1121 :
1122 : /*
1123 : * Acquire OldestXmin.
1124 : *
1125 : * We can always ignore processes running lazy vacuum. This is because we
1126 : * use these values only for deciding which tuples we must keep in the
1127 : * tables. Since lazy vacuum doesn't write its XID anywhere (usually no
1128 : * XID assigned), it's safe to ignore it. In theory it could be
1129 : * problematic to ignore lazy vacuums in a full vacuum, but keep in mind
1130 : * that only one vacuum process can be working on a particular table at
1131 : * any time, and that each vacuum is always an independent transaction.
1132 : */
1133 272184 : cutoffs->OldestXmin = GetOldestNonRemovableTransactionId(rel);
1134 :
1135 : Assert(TransactionIdIsNormal(cutoffs->OldestXmin));
1136 :
1137 : /* Acquire OldestMxact */
1138 272184 : cutoffs->OldestMxact = GetOldestMultiXactId();
1139 : Assert(MultiXactIdIsValid(cutoffs->OldestMxact));
1140 :
1141 : /* Acquire next XID/next MXID values used to apply age-based settings */
1142 272184 : nextXID = ReadNextTransactionId();
1143 272184 : nextMXID = ReadNextMultiXactId();
1144 :
1145 : /*
1146 : * Also compute the multixact age for which freezing is urgent. This is
1147 : * normally autovacuum_multixact_freeze_max_age, but may be less if we are
1148 : * short of multixact member space.
1149 : */
1150 272184 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1151 :
1152 : /*
1153 : * Almost ready to set freeze output parameters; check if OldestXmin or
1154 : * OldestMxact are held back to an unsafe degree before we start on that
1155 : */
1156 272184 : safeOldestXmin = nextXID - autovacuum_freeze_max_age;
1157 272184 : if (!TransactionIdIsNormal(safeOldestXmin))
1158 0 : safeOldestXmin = FirstNormalTransactionId;
1159 272184 : safeOldestMxact = nextMXID - effective_multixact_freeze_max_age;
1160 272184 : if (safeOldestMxact < FirstMultiXactId)
1161 0 : safeOldestMxact = FirstMultiXactId;
1162 272184 : if (TransactionIdPrecedes(cutoffs->OldestXmin, safeOldestXmin))
1163 211592 : ereport(WARNING,
1164 : (errmsg("cutoff for removing and freezing tuples is far in the past"),
1165 : errhint("Close open transactions soon to avoid wraparound problems.\n"
1166 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1167 272184 : if (MultiXactIdPrecedes(cutoffs->OldestMxact, safeOldestMxact))
1168 0 : ereport(WARNING,
1169 : (errmsg("cutoff for freezing multixacts is far in the past"),
1170 : errhint("Close open transactions soon to avoid wraparound problems.\n"
1171 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1172 :
1173 : /*
1174 : * Determine the minimum freeze age to use: as specified by the caller, or
1175 : * vacuum_freeze_min_age, but in any case not more than half
1176 : * autovacuum_freeze_max_age, so that autovacuums to prevent XID
1177 : * wraparound won't occur too frequently.
1178 : */
1179 272184 : if (freeze_min_age < 0)
1180 11006 : freeze_min_age = vacuum_freeze_min_age;
1181 272184 : freeze_min_age = Min(freeze_min_age, autovacuum_freeze_max_age / 2);
1182 : Assert(freeze_min_age >= 0);
1183 :
1184 : /* Compute FreezeLimit, being careful to generate a normal XID */
1185 272184 : cutoffs->FreezeLimit = nextXID - freeze_min_age;
1186 272184 : if (!TransactionIdIsNormal(cutoffs->FreezeLimit))
1187 0 : cutoffs->FreezeLimit = FirstNormalTransactionId;
1188 : /* FreezeLimit must always be <= OldestXmin */
1189 272184 : if (TransactionIdPrecedes(cutoffs->OldestXmin, cutoffs->FreezeLimit))
1190 235408 : cutoffs->FreezeLimit = cutoffs->OldestXmin;
1191 :
1192 : /*
1193 : * Determine the minimum multixact freeze age to use: as specified by
1194 : * caller, or vacuum_multixact_freeze_min_age, but in any case not more
1195 : * than half effective_multixact_freeze_max_age, so that autovacuums to
1196 : * prevent MultiXact wraparound won't occur too frequently.
1197 : */
1198 272184 : if (multixact_freeze_min_age < 0)
1199 11006 : multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1200 272184 : multixact_freeze_min_age = Min(multixact_freeze_min_age,
1201 : effective_multixact_freeze_max_age / 2);
1202 : Assert(multixact_freeze_min_age >= 0);
1203 :
1204 : /* Compute MultiXactCutoff, being careful to generate a valid value */
1205 272184 : cutoffs->MultiXactCutoff = nextMXID - multixact_freeze_min_age;
1206 272184 : if (cutoffs->MultiXactCutoff < FirstMultiXactId)
1207 0 : cutoffs->MultiXactCutoff = FirstMultiXactId;
1208 : /* MultiXactCutoff must always be <= OldestMxact */
1209 272184 : if (MultiXactIdPrecedes(cutoffs->OldestMxact, cutoffs->MultiXactCutoff))
1210 6 : cutoffs->MultiXactCutoff = cutoffs->OldestMxact;
1211 :
1212 : /*
1213 : * Finally, figure out if caller needs to do an aggressive VACUUM or not.
1214 : *
1215 : * Determine the table freeze age to use: as specified by the caller, or
1216 : * the value of the vacuum_freeze_table_age GUC, but in any case not more
1217 : * than autovacuum_freeze_max_age * 0.95, so that if you have e.g nightly
1218 : * VACUUM schedule, the nightly VACUUM gets a chance to freeze XIDs before
1219 : * anti-wraparound autovacuum is launched.
1220 : */
1221 272184 : if (freeze_table_age < 0)
1222 11006 : freeze_table_age = vacuum_freeze_table_age;
1223 272184 : freeze_table_age = Min(freeze_table_age, autovacuum_freeze_max_age * 0.95);
1224 : Assert(freeze_table_age >= 0);
1225 272184 : aggressiveXIDCutoff = nextXID - freeze_table_age;
1226 272184 : if (!TransactionIdIsNormal(aggressiveXIDCutoff))
1227 0 : aggressiveXIDCutoff = FirstNormalTransactionId;
1228 272184 : if (TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid,
1229 : aggressiveXIDCutoff))
1230 261180 : return true;
1231 :
1232 : /*
1233 : * Similar to the above, determine the table freeze age to use for
1234 : * multixacts: as specified by the caller, or the value of the
1235 : * vacuum_multixact_freeze_table_age GUC, but in any case not more than
1236 : * effective_multixact_freeze_max_age * 0.95, so that if you have e.g.
1237 : * nightly VACUUM schedule, the nightly VACUUM gets a chance to freeze
1238 : * multixacts before anti-wraparound autovacuum is launched.
1239 : */
1240 11004 : if (multixact_freeze_table_age < 0)
1241 10788 : multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1242 11004 : multixact_freeze_table_age =
1243 11004 : Min(multixact_freeze_table_age,
1244 : effective_multixact_freeze_max_age * 0.95);
1245 : Assert(multixact_freeze_table_age >= 0);
1246 11004 : aggressiveMXIDCutoff = nextMXID - multixact_freeze_table_age;
1247 11004 : if (aggressiveMXIDCutoff < FirstMultiXactId)
1248 0 : aggressiveMXIDCutoff = FirstMultiXactId;
1249 11004 : if (MultiXactIdPrecedesOrEquals(cutoffs->relminmxid,
1250 : aggressiveMXIDCutoff))
1251 0 : return true;
1252 :
1253 : /* Non-aggressive VACUUM */
1254 11004 : return false;
1255 : }
1256 :
1257 : /*
1258 : * vacuum_xid_failsafe_check() -- Used by VACUUM's wraparound failsafe
1259 : * mechanism to determine if its table's relfrozenxid and relminmxid are now
1260 : * dangerously far in the past.
1261 : *
1262 : * When we return true, VACUUM caller triggers the failsafe.
1263 : */
1264 : bool
1265 275372 : vacuum_xid_failsafe_check(const struct VacuumCutoffs *cutoffs)
1266 : {
1267 275372 : TransactionId relfrozenxid = cutoffs->relfrozenxid;
1268 275372 : MultiXactId relminmxid = cutoffs->relminmxid;
1269 : TransactionId xid_skip_limit;
1270 : MultiXactId multi_skip_limit;
1271 : int skip_index_vacuum;
1272 :
1273 : Assert(TransactionIdIsNormal(relfrozenxid));
1274 : Assert(MultiXactIdIsValid(relminmxid));
1275 :
1276 : /*
1277 : * Determine the index skipping age to use. In any case no less than
1278 : * autovacuum_freeze_max_age * 1.05.
1279 : */
1280 275372 : skip_index_vacuum = Max(vacuum_failsafe_age, autovacuum_freeze_max_age * 1.05);
1281 :
1282 275372 : xid_skip_limit = ReadNextTransactionId() - skip_index_vacuum;
1283 275372 : if (!TransactionIdIsNormal(xid_skip_limit))
1284 0 : xid_skip_limit = FirstNormalTransactionId;
1285 :
1286 275372 : if (TransactionIdPrecedes(relfrozenxid, xid_skip_limit))
1287 : {
1288 : /* The table's relfrozenxid is too old */
1289 72300 : return true;
1290 : }
1291 :
1292 : /*
1293 : * Similar to above, determine the index skipping age to use for
1294 : * multixact. In any case no less than autovacuum_multixact_freeze_max_age *
1295 : * 1.05.
1296 : */
1297 203072 : skip_index_vacuum = Max(vacuum_multixact_failsafe_age,
1298 : autovacuum_multixact_freeze_max_age * 1.05);
1299 :
1300 203072 : multi_skip_limit = ReadNextMultiXactId() - skip_index_vacuum;
1301 203072 : if (multi_skip_limit < FirstMultiXactId)
1302 0 : multi_skip_limit = FirstMultiXactId;
1303 :
1304 203072 : if (MultiXactIdPrecedes(relminmxid, multi_skip_limit))
1305 : {
1306 : /* The table's relminmxid is too old */
1307 0 : return true;
1308 : }
1309 :
1310 203072 : return false;
1311 : }
1312 :
1313 : /*
1314 : * vac_estimate_reltuples() -- estimate the new value for pg_class.reltuples
1315 : *
1316 : * If we scanned the whole relation then we should just use the count of
1317 : * live tuples seen; but if we did not, we should not blindly extrapolate
1318 : * from that number, since VACUUM may have scanned a quite nonrandom
1319 : * subset of the table. When we have only partial information, we take
1320 : * the old value of pg_class.reltuples/pg_class.relpages as a measurement
1321 : * of the tuple density in the unscanned pages.
1322 : *
1323 : * Note: scanned_tuples should count only *live* tuples, since
1324 : * pg_class.reltuples is defined that way.
1325 : */
1326 : double
1327 271616 : vac_estimate_reltuples(Relation relation,
1328 : BlockNumber total_pages,
1329 : BlockNumber scanned_pages,
1330 : double scanned_tuples)
1331 : {
1332 271616 : BlockNumber old_rel_pages = relation->rd_rel->relpages;
1333 271616 : double old_rel_tuples = relation->rd_rel->reltuples;
1334 : double old_density;
1335 : double unscanned_pages;
1336 : double total_tuples;
1337 :
1338 : /* If we did scan the whole table, just use the count as-is */
1339 271616 : if (scanned_pages >= total_pages)
1340 261798 : return scanned_tuples;
1341 :
1342 : /*
1343 : * When successive VACUUM commands scan the same few pages again and
1344 : * again, without anything from the table really changing, there is a risk
1345 : * that our beliefs about tuple density will gradually become distorted.
1346 : * This might be caused by vacuumlazy.c implementation details, such as
1347 : * its tendency to always scan the last heap page. Handle that here.
1348 : *
1349 : * If the relation is _exactly_ the same size according to the existing
1350 : * pg_class entry, and only a few of its pages (less than 2%) were
1351 : * scanned, keep the existing value of reltuples. Also keep the existing
1352 : * value when only a subset of rel's pages <= a single page were scanned.
1353 : *
1354 : * (Note: we might be returning -1 here.)
1355 : */
1356 9818 : if (old_rel_pages == total_pages &&
1357 9790 : scanned_pages < (double) total_pages * 0.02)
1358 7040 : return old_rel_tuples;
1359 2778 : if (scanned_pages <= 1)
1360 2380 : return old_rel_tuples;
1361 :
1362 : /*
1363 : * If old density is unknown, we can't do much except scale up
1364 : * scanned_tuples to match total_pages.
1365 : */
1366 398 : if (old_rel_tuples < 0 || old_rel_pages == 0)
1367 2 : return floor((scanned_tuples / scanned_pages) * total_pages + 0.5);
1368 :
1369 : /*
1370 : * Okay, we've covered the corner cases. The normal calculation is to
1371 : * convert the old measurement to a density (tuples per page), then
1372 : * estimate the number of tuples in the unscanned pages using that figure,
1373 : * and finally add on the number of tuples in the scanned pages.
1374 : */
1375 396 : old_density = old_rel_tuples / old_rel_pages;
1376 396 : unscanned_pages = (double) total_pages - (double) scanned_pages;
1377 396 : total_tuples = old_density * unscanned_pages + scanned_tuples;
1378 396 : return floor(total_tuples + 0.5);
1379 : }
1380 :
1381 :
1382 : /*
1383 : * vac_update_relstats() -- update statistics for one relation
1384 : *
1385 : * Update the whole-relation statistics that are kept in its pg_class
1386 : * row. There are additional stats that will be updated if we are
1387 : * doing ANALYZE, but we always update these stats. This routine works
1388 : * for both index and heap relation entries in pg_class.
1389 : *
1390 : * We violate transaction semantics here by overwriting the rel's
1391 : * existing pg_class tuple with the new values. This is reasonably
1392 : * safe as long as we're sure that the new values are correct whether or
1393 : * not this transaction commits. The reason for doing this is that if
1394 : * we updated these tuples in the usual way, vacuuming pg_class itself
1395 : * wouldn't work very well --- by the time we got done with a vacuum
1396 : * cycle, most of the tuples in pg_class would've been obsoleted. Of
1397 : * course, this only works for fixed-size not-null columns, but these are.
1398 : *
1399 : * Another reason for doing it this way is that when we are in a lazy
1400 : * VACUUM and have PROC_IN_VACUUM set, we mustn't do any regular updates.
1401 : * Somebody vacuuming pg_class might think they could delete a tuple
1402 : * marked with xmin = our xid.
1403 : *
1404 : * In addition to fundamentally nontransactional statistics such as
1405 : * relpages and relallvisible, we try to maintain certain lazily-updated
1406 : * DDL flags such as relhasindex, by clearing them if no longer correct.
1407 : * It's safe to do this in VACUUM, which can't run in parallel with
1408 : * CREATE INDEX/RULE/TRIGGER and can't be part of a transaction block.
1409 : * However, it's *not* safe to do it in an ANALYZE that's within an
1410 : * outer transaction, because for example the current transaction might
1411 : * have dropped the last index; then we'd think relhasindex should be
1412 : * cleared, but if the transaction later rolls back this would be wrong.
1413 : * So we refrain from updating the DDL flags if we're inside an outer
1414 : * transaction. This is OK since postponing the flag maintenance is
1415 : * always allowable.
1416 : *
1417 : * Note: num_tuples should count only *live* tuples, since
1418 : * pg_class.reltuples is defined that way.
1419 : *
1420 : * This routine is shared by VACUUM and ANALYZE.
1421 : */
1422 : void
1423 313456 : vac_update_relstats(Relation relation,
1424 : BlockNumber num_pages, double num_tuples,
1425 : BlockNumber num_all_visible_pages,
1426 : BlockNumber num_all_frozen_pages,
1427 : bool hasindex, TransactionId frozenxid,
1428 : MultiXactId minmulti,
1429 : bool *frozenxid_updated, bool *minmulti_updated,
1430 : bool in_outer_xact)
1431 : {
1432 313456 : Oid relid = RelationGetRelid(relation);
1433 : Relation rd;
1434 : ScanKeyData key[1];
1435 : HeapTuple ctup;
1436 : void *inplace_state;
1437 : Form_pg_class pgcform;
1438 : bool dirty,
1439 : futurexid,
1440 : futuremxid;
1441 : TransactionId oldfrozenxid;
1442 : MultiXactId oldminmulti;
1443 :
1444 313456 : rd = table_open(RelationRelationId, RowExclusiveLock);
1445 :
1446 : /* Fetch a copy of the tuple to scribble on */
1447 313456 : ScanKeyInit(&key[0],
1448 : Anum_pg_class_oid,
1449 : BTEqualStrategyNumber, F_OIDEQ,
1450 : ObjectIdGetDatum(relid));
1451 313456 : systable_inplace_update_begin(rd, ClassOidIndexId, true,
1452 : NULL, 1, key, &ctup, &inplace_state);
1453 313454 : if (!HeapTupleIsValid(ctup))
1454 0 : elog(ERROR, "pg_class entry for relid %u vanished during vacuuming",
1455 : relid);
1456 313454 : pgcform = (Form_pg_class) GETSTRUCT(ctup);
1457 :
1458 : /* Apply statistical updates, if any, to copied tuple */
1459 :
1460 313454 : dirty = false;
1461 313454 : if (pgcform->relpages != (int32) num_pages)
1462 : {
1463 9534 : pgcform->relpages = (int32) num_pages;
1464 9534 : dirty = true;
1465 : }
1466 313454 : if (pgcform->reltuples != (float4) num_tuples)
1467 : {
1468 20426 : pgcform->reltuples = (float4) num_tuples;
1469 20426 : dirty = true;
1470 : }
1471 313454 : if (pgcform->relallvisible != (int32) num_all_visible_pages)
1472 : {
1473 5854 : pgcform->relallvisible = (int32) num_all_visible_pages;
1474 5854 : dirty = true;
1475 : }
1476 313454 : if (pgcform->relallfrozen != (int32) num_all_frozen_pages)
1477 : {
1478 5276 : pgcform->relallfrozen = (int32) num_all_frozen_pages;
1479 5276 : dirty = true;
1480 : }
1481 :
1482 : /* Apply DDL updates, but not inside an outer transaction (see above) */
1483 :
1484 313454 : if (!in_outer_xact)
1485 : {
1486 : /*
1487 : * If we didn't find any indexes, reset relhasindex.
1488 : */
1489 313140 : if (pgcform->relhasindex && !hasindex)
1490 : {
1491 20 : pgcform->relhasindex = false;
1492 20 : dirty = true;
1493 : }
1494 :
1495 : /* We also clear relhasrules and relhastriggers if needed */
1496 313140 : if (pgcform->relhasrules && relation->rd_rules == NULL)
1497 : {
1498 0 : pgcform->relhasrules = false;
1499 0 : dirty = true;
1500 : }
1501 313140 : if (pgcform->relhastriggers && relation->trigdesc == NULL)
1502 : {
1503 6 : pgcform->relhastriggers = false;
1504 6 : dirty = true;
1505 : }
1506 : }
1507 :
1508 : /*
1509 : * Update relfrozenxid, unless caller passed InvalidTransactionId
1510 : * indicating it has no new data.
1511 : *
1512 : * Ordinarily, we don't let relfrozenxid go backwards. However, if the
1513 : * stored relfrozenxid is "in the future" then it seems best to assume
1514 : * it's corrupt, and overwrite with the oldest remaining XID in the table.
1515 : * This should match vac_update_datfrozenxid() concerning what we consider
1516 : * to be "in the future".
1517 : */
1518 313454 : oldfrozenxid = pgcform->relfrozenxid;
1519 313454 : futurexid = false;
1520 313454 : if (frozenxid_updated)
1521 271610 : *frozenxid_updated = false;
1522 313454 : if (TransactionIdIsNormal(frozenxid) && oldfrozenxid != frozenxid)
1523 : {
1524 57814 : bool update = false;
1525 :
1526 57814 : if (TransactionIdPrecedes(oldfrozenxid, frozenxid))
1527 57724 : update = true;
1528 90 : else if (TransactionIdPrecedes(ReadNextTransactionId(), oldfrozenxid))
1529 0 : futurexid = update = true;
1530 :
1531 57814 : if (update)
1532 : {
1533 57724 : pgcform->relfrozenxid = frozenxid;
1534 57724 : dirty = true;
1535 57724 : if (frozenxid_updated)
1536 57724 : *frozenxid_updated = true;
1537 : }
1538 : }
1539 :
1540 : /* Similarly for relminmxid */
1541 313454 : oldminmulti = pgcform->relminmxid;
1542 313454 : futuremxid = false;
1543 313454 : if (minmulti_updated)
1544 271610 : *minmulti_updated = false;
1545 313454 : if (MultiXactIdIsValid(minmulti) && oldminmulti != minmulti)
1546 : {
1547 292 : bool update = false;
1548 :
1549 292 : if (MultiXactIdPrecedes(oldminmulti, minmulti))
1550 292 : update = true;
1551 0 : else if (MultiXactIdPrecedes(ReadNextMultiXactId(), oldminmulti))
1552 0 : futuremxid = update = true;
1553 :
1554 292 : if (update)
1555 : {
1556 292 : pgcform->relminmxid = minmulti;
1557 292 : dirty = true;
1558 292 : if (minmulti_updated)
1559 292 : *minmulti_updated = true;
1560 : }
1561 : }
1562 :
1563 : /* If anything changed, write out the tuple. */
1564 313454 : if (dirty)
1565 72252 : systable_inplace_update_finish(inplace_state, ctup);
1566 : else
1567 241202 : systable_inplace_update_cancel(inplace_state);
1568 :
1569 313454 : table_close(rd, RowExclusiveLock);
1570 :
1571 313454 : if (futurexid)
1572 0 : ereport(WARNING,
1573 : (errcode(ERRCODE_DATA_CORRUPTED),
1574 : errmsg_internal("overwrote invalid relfrozenxid value %u with new value %u for table \"%s\"",
1575 : oldfrozenxid, frozenxid,
1576 : RelationGetRelationName(relation))));
1577 313454 : if (futuremxid)
1578 0 : ereport(WARNING,
1579 : (errcode(ERRCODE_DATA_CORRUPTED),
1580 : errmsg_internal("overwrote invalid relminmxid value %u with new value %u for table \"%s\"",
1581 : oldminmulti, minmulti,
1582 : RelationGetRelationName(relation))));
1583 313454 : }
1584 :
1585 :
1586 : /*
1587 : * vac_update_datfrozenxid() -- update pg_database.datfrozenxid for our DB
1588 : *
1589 : * Update pg_database's datfrozenxid entry for our database to be the
1590 : * minimum of the pg_class.relfrozenxid values.
1591 : *
1592 : * Similarly, update our datminmxid to be the minimum of the
1593 : * pg_class.relminmxid values.
1594 : *
1595 : * If we are able to advance either pg_database value, also try to
1596 : * truncate pg_xact and pg_multixact.
1597 : *
1598 : * We violate transaction semantics here by overwriting the database's
1599 : * existing pg_database tuple with the new values. This is reasonably
1600 : * safe since the new values are correct whether or not this transaction
1601 : * commits. As with vac_update_relstats, this avoids leaving dead tuples
1602 : * behind after a VACUUM.
1603 : */
1604 : void
1605 6244 : vac_update_datfrozenxid(void)
1606 : {
1607 : HeapTuple tuple;
1608 : Form_pg_database dbform;
1609 : Relation relation;
1610 : SysScanDesc scan;
1611 : HeapTuple classTup;
1612 : TransactionId newFrozenXid;
1613 : MultiXactId newMinMulti;
1614 : TransactionId lastSaneFrozenXid;
1615 : MultiXactId lastSaneMinMulti;
1616 6244 : bool bogus = false;
1617 6244 : bool dirty = false;
1618 : ScanKeyData key[1];
1619 : void *inplace_state;
1620 :
1621 : /*
1622 : * Restrict this task to one backend per database. This avoids race
1623 : * conditions that would move datfrozenxid or datminmxid backward. It
1624 : * avoids calling vac_truncate_clog() with a datfrozenxid preceding a
1625 : * datfrozenxid passed to an earlier vac_truncate_clog() call.
1626 : */
1627 6244 : LockDatabaseFrozenIds(ExclusiveLock);
1628 :
1629 : /*
1630 : * Initialize the "min" calculation with
1631 : * GetOldestNonRemovableTransactionId(), which is a reasonable
1632 : * approximation to the minimum relfrozenxid for not-yet-committed
1633 : * pg_class entries for new tables; see AddNewRelationTuple(). So we
1634 : * cannot produce a wrong minimum by starting with this.
1635 : */
1636 6244 : newFrozenXid = GetOldestNonRemovableTransactionId(NULL);
1637 :
1638 : /*
1639 : * Similarly, initialize the MultiXact "min" with the value that would be
1640 : * used on pg_class for new tables. See AddNewRelationTuple().
1641 : */
1642 6244 : newMinMulti = GetOldestMultiXactId();
1643 :
1644 : /*
1645 : * Identify the latest relfrozenxid and relminmxid values that we could
1646 : * validly see during the scan. These are conservative values, but it's
1647 : * not really worth trying to be more exact.
1648 : */
1649 6244 : lastSaneFrozenXid = ReadNextTransactionId();
1650 6244 : lastSaneMinMulti = ReadNextMultiXactId();
1651 :
1652 : /*
1653 : * We must seqscan pg_class to find the minimum Xid, because there is no
1654 : * index that can help us here.
1655 : *
1656 : * See vac_truncate_clog() for the race condition to prevent.
1657 : */
1658 6244 : relation = table_open(RelationRelationId, AccessShareLock);
1659 :
1660 6244 : scan = systable_beginscan(relation, InvalidOid, false,
1661 : NULL, 0, NULL);
1662 :
1663 3066880 : while ((classTup = systable_getnext(scan)) != NULL)
1664 : {
1665 3060636 : volatile FormData_pg_class *classForm = (Form_pg_class) GETSTRUCT(classTup);
1666 3060636 : TransactionId relfrozenxid = classForm->relfrozenxid;
1667 3060636 : TransactionId relminmxid = classForm->relminmxid;
1668 :
1669 : /*
1670 : * Only consider relations able to hold unfrozen XIDs (anything else
1671 : * should have InvalidTransactionId in relfrozenxid anyway).
1672 : */
1673 3060636 : if (classForm->relkind != RELKIND_RELATION &&
1674 2452926 : classForm->relkind != RELKIND_MATVIEW &&
1675 2450604 : classForm->relkind != RELKIND_TOASTVALUE)
1676 : {
1677 : Assert(!TransactionIdIsValid(relfrozenxid));
1678 : Assert(!MultiXactIdIsValid(relminmxid));
1679 2133696 : continue;
1680 : }
1681 :
1682 : /*
1683 : * Some table AMs might not need per-relation xid / multixid horizons.
1684 : * It therefore seems reasonable to allow relfrozenxid and relminmxid
1685 : * to not be set (i.e. set to their respective Invalid*Id)
1686 : * independently. Thus validate and compute horizon for each only if
1687 : * set.
1688 : *
1689 : * If things are working properly, no relation should have a
1690 : * relfrozenxid or relminmxid that is "in the future". However, such
1691 : * cases have been known to arise due to bugs in pg_upgrade. If we
1692 : * see any entries that are "in the future", chicken out and don't do
1693 : * anything. This ensures we won't truncate clog & multixact SLRUs
1694 : * before those relations have been scanned and cleaned up.
1695 : */
1696 :
1697 926940 : if (TransactionIdIsValid(relfrozenxid))
1698 : {
1699 : Assert(TransactionIdIsNormal(relfrozenxid));
1700 :
1701 : /* check for values in the future */
1702 926940 : if (TransactionIdPrecedes(lastSaneFrozenXid, relfrozenxid))
1703 : {
1704 0 : bogus = true;
1705 0 : break;
1706 : }
1707 :
1708 : /* determine new horizon */
1709 926940 : if (TransactionIdPrecedes(relfrozenxid, newFrozenXid))
1710 4876 : newFrozenXid = relfrozenxid;
1711 : }
1712 :
1713 926940 : if (MultiXactIdIsValid(relminmxid))
1714 : {
1715 : /* check for values in the future */
1716 926940 : if (MultiXactIdPrecedes(lastSaneMinMulti, relminmxid))
1717 : {
1718 0 : bogus = true;
1719 0 : break;
1720 : }
1721 :
1722 : /* determine new horizon */
1723 926940 : if (MultiXactIdPrecedes(relminmxid, newMinMulti))
1724 220 : newMinMulti = relminmxid;
1725 : }
1726 : }
1727 :
1728 : /* we're done with pg_class */
1729 6244 : systable_endscan(scan);
1730 6244 : table_close(relation, AccessShareLock);
1731 :
1732 : /* chicken out if bogus data found */
1733 6244 : if (bogus)
1734 0 : return;
1735 :
1736 : Assert(TransactionIdIsNormal(newFrozenXid));
1737 : Assert(MultiXactIdIsValid(newMinMulti));
1738 :
1739 : /* Now fetch the pg_database tuple we need to update. */
1740 6244 : relation = table_open(DatabaseRelationId, RowExclusiveLock);
1741 :
1742 : /*
1743 : * Fetch a copy of the tuple to scribble on. We could check the syscache
1744 : * tuple first. If that concluded !dirty, we'd avoid waiting on
1745 : * concurrent heap_update() and would avoid exclusive-locking the buffer.
1746 : * For now, don't optimize that.
1747 : */
1748 6244 : ScanKeyInit(&key[0],
1749 : Anum_pg_database_oid,
1750 : BTEqualStrategyNumber, F_OIDEQ,
1751 : ObjectIdGetDatum(MyDatabaseId));
1752 :
1753 6244 : systable_inplace_update_begin(relation, DatabaseOidIndexId, true,
1754 : NULL, 1, key, &tuple, &inplace_state);
1755 :
1756 6244 : if (!HeapTupleIsValid(tuple))
1757 0 : elog(ERROR, "could not find tuple for database %u", MyDatabaseId);
1758 :
1759 6244 : dbform = (Form_pg_database) GETSTRUCT(tuple);
1760 :
1761 : /*
1762 : * As in vac_update_relstats(), we ordinarily don't want to let
1763 : * datfrozenxid go backward; but if it's "in the future" then it must be
1764 : * corrupt and it seems best to overwrite it.
1765 : */
1766 6806 : if (dbform->datfrozenxid != newFrozenXid &&
1767 562 : (TransactionIdPrecedes(dbform->datfrozenxid, newFrozenXid) ||
1768 0 : TransactionIdPrecedes(lastSaneFrozenXid, dbform->datfrozenxid)))
1769 : {
1770 562 : dbform->datfrozenxid = newFrozenXid;
1771 562 : dirty = true;
1772 : }
1773 : else
1774 5682 : newFrozenXid = dbform->datfrozenxid;
1775 :
1776 : /* Ditto for datminmxid */
1777 6246 : if (dbform->datminmxid != newMinMulti &&
1778 2 : (MultiXactIdPrecedes(dbform->datminmxid, newMinMulti) ||
1779 0 : MultiXactIdPrecedes(lastSaneMinMulti, dbform->datminmxid)))
1780 : {
1781 2 : dbform->datminmxid = newMinMulti;
1782 2 : dirty = true;
1783 : }
1784 : else
1785 6242 : newMinMulti = dbform->datminmxid;
1786 :
1787 6244 : if (dirty)
1788 562 : systable_inplace_update_finish(inplace_state, tuple);
1789 : else
1790 5682 : systable_inplace_update_cancel(inplace_state);
1791 :
1792 6244 : heap_freetuple(tuple);
1793 6244 : table_close(relation, RowExclusiveLock);
1794 :
1795 : /*
1796 : * If we were able to advance datfrozenxid or datminmxid, see if we can
1797 : * truncate pg_xact and/or pg_multixact. Also do it if the shared
1798 : * XID-wrap-limit info is stale, since this action will update that too.
1799 : */
1800 6244 : if (dirty || ForceTransactionIdLimitUpdate())
1801 2498 : vac_truncate_clog(newFrozenXid, newMinMulti,
1802 : lastSaneFrozenXid, lastSaneMinMulti);
1803 : }
1804 :
1805 :
1806 : /*
1807 : * vac_truncate_clog() -- attempt to truncate the commit log
1808 : *
1809 : * Scan pg_database to determine the system-wide oldest datfrozenxid,
1810 : * and use it to truncate the transaction commit log (pg_xact).
1811 : * Also update the XID wrap limit info maintained by varsup.c.
1812 : * Likewise for datminmxid.
1813 : *
1814 : * The passed frozenXID and minMulti are the updated values for my own
1815 : * pg_database entry. They're used to initialize the "min" calculations.
1816 : * The caller also passes the "last sane" XID and MXID, since it has
1817 : * those at hand already.
1818 : *
1819 : * This routine is only invoked when we've managed to change our
1820 : * DB's datfrozenxid/datminmxid values, or we found that the shared
1821 : * XID-wrap-limit info is stale.
1822 : */
1823 : static void
1824 2498 : vac_truncate_clog(TransactionId frozenXID,
1825 : MultiXactId minMulti,
1826 : TransactionId lastSaneFrozenXid,
1827 : MultiXactId lastSaneMinMulti)
1828 : {
1829 2498 : TransactionId nextXID = ReadNextTransactionId();
1830 : Relation relation;
1831 : TableScanDesc scan;
1832 : HeapTuple tuple;
1833 : Oid oldestxid_datoid;
1834 : Oid minmulti_datoid;
1835 2498 : bool bogus = false;
1836 2498 : bool frozenAlreadyWrapped = false;
1837 :
1838 : /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */
1839 2498 : LWLockAcquire(WrapLimitsVacuumLock, LW_EXCLUSIVE);
1840 :
1841 : /* init oldest datoids to sync with my frozenXID/minMulti values */
1842 2498 : oldestxid_datoid = MyDatabaseId;
1843 2498 : minmulti_datoid = MyDatabaseId;
1844 :
1845 : /*
1846 : * Scan pg_database to compute the minimum datfrozenxid/datminmxid
1847 : *
1848 : * Since vac_update_datfrozenxid updates datfrozenxid/datminmxid in-place,
1849 : * the values could change while we look at them. Fetch each one just
1850 : * once to ensure sane behavior of the comparison logic. (Here, as in
1851 : * many other places, we assume that fetching or updating an XID in shared
1852 : * storage is atomic.)
1853 : *
1854 : * Note: we need not worry about a race condition with new entries being
1855 : * inserted by CREATE DATABASE. Any such entry will have a copy of some
1856 : * existing DB's datfrozenxid, and that source DB cannot be ours because
1857 : * of the interlock against copying a DB containing an active backend.
1858 : * Hence the new entry will not reduce the minimum. Also, if two VACUUMs
1859 : * concurrently modify the datfrozenxid's of different databases, the
1860 : * worst possible outcome is that pg_xact is not truncated as aggressively
1861 : * as it could be.
1862 : */
1863 2498 : relation = table_open(DatabaseRelationId, AccessShareLock);
1864 :
1865 2498 : scan = table_beginscan_catalog(relation, 0, NULL);
1866 :
1867 9818 : while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
1868 : {
1869 7320 : volatile FormData_pg_database *dbform = (Form_pg_database) GETSTRUCT(tuple);
1870 7320 : TransactionId datfrozenxid = dbform->datfrozenxid;
1871 7320 : TransactionId datminmxid = dbform->datminmxid;
1872 :
1873 : Assert(TransactionIdIsNormal(datfrozenxid));
1874 : Assert(MultiXactIdIsValid(datminmxid));
1875 :
1876 : /*
1877 : * If database is in the process of getting dropped, or has been
1878 : * interrupted while doing so, no connections to it are possible
1879 : * anymore. Therefore we don't need to take it into account here.
1880 : * Which is good, because it can't be processed by autovacuum either.
1881 : */
1882 7320 : if (database_is_invalid_form((Form_pg_database) dbform))
1883 : {
1884 2 : elog(DEBUG2,
1885 : "skipping invalid database \"%s\" while computing relfrozenxid",
1886 : NameStr(dbform->datname));
1887 2 : continue;
1888 : }
1889 :
1890 : /*
1891 : * If things are working properly, no database should have a
1892 : * datfrozenxid or datminmxid that is "in the future". However, such
1893 : * cases have been known to arise due to bugs in pg_upgrade. If we
1894 : * see any entries that are "in the future", chicken out and don't do
1895 : * anything. This ensures we won't truncate clog before those
1896 : * databases have been scanned and cleaned up. (We will issue the
1897 : * "already wrapped" warning if appropriate, though.)
1898 : */
1899 14636 : if (TransactionIdPrecedes(lastSaneFrozenXid, datfrozenxid) ||
1900 7318 : MultiXactIdPrecedes(lastSaneMinMulti, datminmxid))
1901 0 : bogus = true;
1902 :
1903 7318 : if (TransactionIdPrecedes(nextXID, datfrozenxid))
1904 0 : frozenAlreadyWrapped = true;
1905 7318 : else if (TransactionIdPrecedes(datfrozenxid, frozenXID))
1906 : {
1907 408 : frozenXID = datfrozenxid;
1908 408 : oldestxid_datoid = dbform->oid;
1909 : }
1910 :
1911 7318 : if (MultiXactIdPrecedes(datminmxid, minMulti))
1912 : {
1913 4 : minMulti = datminmxid;
1914 4 : minmulti_datoid = dbform->oid;
1915 : }
1916 : }
1917 :
1918 2498 : table_endscan(scan);
1919 :
1920 2498 : table_close(relation, AccessShareLock);
1921 :
1922 : /*
1923 : * Do not truncate CLOG if we seem to have suffered wraparound already;
1924 : * the computed minimum XID might be bogus. This case should now be
1925 : * impossible due to the defenses in GetNewTransactionId, but we keep the
1926 : * test anyway.
1927 : */
1928 2498 : if (frozenAlreadyWrapped)
1929 : {
1930 0 : ereport(WARNING,
1931 : (errmsg("some databases have not been vacuumed in over 2 billion transactions"),
1932 : errdetail("You might have already suffered transaction-wraparound data loss.")));
1933 0 : LWLockRelease(WrapLimitsVacuumLock);
1934 0 : return;
1935 : }
1936 :
1937 : /* chicken out if data is bogus in any other way */
1938 2498 : if (bogus)
1939 : {
1940 0 : LWLockRelease(WrapLimitsVacuumLock);
1941 0 : return;
1942 : }
1943 :
1944 : /*
1945 : * Advance the oldest value for commit timestamps before truncating, so
1946 : * that if a user requests a timestamp for a transaction we're truncating
1947 : * away right after this point, they get NULL instead of an ugly "file not
1948 : * found" error from slru.c. This doesn't matter for xact/multixact
1949 : * because they are not subject to arbitrary lookups from users.
1950 : */
1951 2498 : AdvanceOldestCommitTsXid(frozenXID);
1952 :
1953 : /*
1954 : * Truncate CLOG, multixact and CommitTs to the oldest computed value.
1955 : */
1956 2498 : TruncateCLOG(frozenXID, oldestxid_datoid);
1957 2498 : TruncateCommitTs(frozenXID);
1958 2498 : TruncateMultiXact(minMulti, minmulti_datoid);
1959 :
1960 : /*
1961 : * Update the wrap limit for GetNewTransactionId and creation of new
1962 : * MultiXactIds. Note: these functions will also signal the postmaster
1963 : * for an(other) autovac cycle if needed. XXX should we avoid possibly
1964 : * signaling twice?
1965 : */
1966 2498 : SetTransactionIdLimit(frozenXID, oldestxid_datoid);
1967 2498 : SetMultiXactIdLimit(minMulti, minmulti_datoid, false);
1968 :
1969 2498 : LWLockRelease(WrapLimitsVacuumLock);
1970 : }
1971 :
1972 :
1973 : /*
1974 : * vacuum_rel() -- vacuum one heap relation
1975 : *
1976 : * relid identifies the relation to vacuum. If relation is supplied,
1977 : * use the name therein for reporting any failure to open/lock the rel;
1978 : * do not use it once we've successfully opened the rel, since it might
1979 : * be stale.
1980 : *
1981 : * Returns true if it's okay to proceed with a requested ANALYZE
1982 : * operation on this table.
1983 : *
1984 : * Doing one heap at a time incurs extra overhead, since we need to
1985 : * check that the heap exists again just before we vacuum it. The
1986 : * reason that we do this is so that vacuuming can be spread across
1987 : * many small transactions. Otherwise, two-phase locking would require
1988 : * us to lock the entire database during one pass of the vacuum cleaner.
1989 : *
1990 : * At entry and exit, we are not inside a transaction.
1991 : */
1992 : static bool
1993 272432 : vacuum_rel(Oid relid, RangeVar *relation, VacuumParams params,
1994 : BufferAccessStrategy bstrategy)
1995 : {
1996 : LOCKMODE lmode;
1997 : Relation rel;
1998 : LockRelId lockrelid;
1999 : Oid priv_relid;
2000 : Oid toast_relid;
2001 : Oid save_userid;
2002 : int save_sec_context;
2003 : int save_nestlevel;
2004 : VacuumParams toast_vacuum_params;
2005 :
2006 : /*
2007 : * This function scribbles on the parameters, so make a copy early to
2008 : * avoid affecting the TOAST table (if we do end up recursing to it).
2009 : */
2010 272432 : memcpy(&toast_vacuum_params, ¶ms, sizeof(VacuumParams));
2011 :
2012 : /* Begin a transaction for vacuuming this relation */
2013 272432 : StartTransactionCommand();
2014 :
2015 272432 : if (!(params.options & VACOPT_FULL))
2016 : {
2017 : /*
2018 : * In lazy vacuum, we can set the PROC_IN_VACUUM flag, which lets
2019 : * other concurrent VACUUMs know that they can ignore this one while
2020 : * determining their OldestXmin. (The reason we don't set it during a
2021 : * full VACUUM is exactly that we may have to run user-defined
2022 : * functions for functional indexes, and we want to make sure that if
2023 : * they use the snapshot set above, any tuples it requires can't get
2024 : * removed from other tables. An index function that depends on the
2025 : * contents of other tables is arguably broken, but we won't break it
2026 : * here by violating transaction semantics.)
2027 : *
2028 : * We also set the VACUUM_FOR_WRAPAROUND flag, which is passed down by
2029 : * autovacuum; it's used to avoid canceling a vacuum that was invoked
2030 : * in an emergency.
2031 : *
2032 : * Note: these flags remain set until CommitTransaction or
2033 : * AbortTransaction. We don't want to clear them until we reset
2034 : * MyProc->xid/xmin, otherwise GetOldestNonRemovableTransactionId()
2035 : * might appear to go backwards, which is probably Not Good. (We also
2036 : * set PROC_IN_VACUUM *before* taking our own snapshot, so that our
2037 : * xmin doesn't become visible ahead of setting the flag.)
2038 : */
2039 272018 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2040 272018 : MyProc->statusFlags |= PROC_IN_VACUUM;
2041 272018 : if (params.is_wraparound)
2042 245368 : MyProc->statusFlags |= PROC_VACUUM_FOR_WRAPAROUND;
2043 272018 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2044 272018 : LWLockRelease(ProcArrayLock);
2045 : }
2046 :
2047 : /*
2048 : * Need to acquire a snapshot to prevent pg_subtrans from being truncated,
2049 : * cutoff xids in local memory wrapping around, and to have updated xmin
2050 : * horizons.
2051 : */
2052 272432 : PushActiveSnapshot(GetTransactionSnapshot());
2053 :
2054 : /*
2055 : * Check for user-requested abort. Note we want this to be inside a
2056 : * transaction, so xact.c doesn't issue useless WARNING.
2057 : */
2058 272432 : CHECK_FOR_INTERRUPTS();
2059 :
2060 : /*
2061 : * Determine the type of lock we want --- hard exclusive lock for a FULL
2062 : * vacuum, but just ShareUpdateExclusiveLock for concurrent vacuum. Either
2063 : * way, we can be sure that no other backend is vacuuming the same table.
2064 : */
2065 544864 : lmode = (params.options & VACOPT_FULL) ?
2066 272432 : AccessExclusiveLock : ShareUpdateExclusiveLock;
2067 :
2068 : /* open the relation and get the appropriate lock on it */
2069 272432 : rel = vacuum_open_relation(relid, relation, params.options,
2070 272432 : params.log_vacuum_min_duration >= 0, lmode);
2071 :
2072 : /* leave if relation could not be opened or locked */
2073 272432 : if (!rel)
2074 : {
2075 24 : PopActiveSnapshot();
2076 24 : CommitTransactionCommand();
2077 24 : return false;
2078 : }
2079 :
2080 : /*
2081 : * When recursing to a TOAST table, check privileges on the parent. NB:
2082 : * This is only safe to do because we hold a session lock on the main
2083 : * relation that prevents concurrent deletion.
2084 : */
2085 272408 : if (OidIsValid(params.toast_parent))
2086 9176 : priv_relid = params.toast_parent;
2087 : else
2088 263232 : priv_relid = RelationGetRelid(rel);
2089 :
2090 : /*
2091 : * Check if relation needs to be skipped based on privileges. This check
2092 : * happens also when building the relation list to vacuum for a manual
2093 : * operation, and needs to be done additionally here as VACUUM could
2094 : * happen across multiple transactions where privileges could have changed
2095 : * in-between. Make sure to only generate logs for VACUUM in this case.
2096 : */
2097 272408 : if (!vacuum_is_permitted_for_relation(priv_relid,
2098 : rel->rd_rel,
2099 272408 : params.options & ~VACOPT_ANALYZE))
2100 : {
2101 72 : relation_close(rel, lmode);
2102 72 : PopActiveSnapshot();
2103 72 : CommitTransactionCommand();
2104 72 : return false;
2105 : }
2106 :
2107 : /*
2108 : * Check that it's of a vacuumable relkind.
2109 : */
2110 272336 : if (rel->rd_rel->relkind != RELKIND_RELATION &&
2111 99584 : rel->rd_rel->relkind != RELKIND_MATVIEW &&
2112 99576 : rel->rd_rel->relkind != RELKIND_TOASTVALUE &&
2113 188 : rel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
2114 : {
2115 2 : ereport(WARNING,
2116 : (errmsg("skipping \"%s\" --- cannot vacuum non-tables or special system tables",
2117 : RelationGetRelationName(rel))));
2118 2 : relation_close(rel, lmode);
2119 2 : PopActiveSnapshot();
2120 2 : CommitTransactionCommand();
2121 2 : return false;
2122 : }
2123 :
2124 : /*
2125 : * Silently ignore tables that are temp tables of other backends ---
2126 : * trying to vacuum these will lead to great unhappiness, since their
2127 : * contents are probably not up-to-date on disk. (We don't throw a
2128 : * warning here; it would just lead to chatter during a database-wide
2129 : * VACUUM.)
2130 : */
2131 272334 : if (RELATION_IS_OTHER_TEMP(rel))
2132 : {
2133 2 : relation_close(rel, lmode);
2134 2 : PopActiveSnapshot();
2135 2 : CommitTransactionCommand();
2136 2 : return false;
2137 : }
2138 :
2139 : /*
2140 : * Silently ignore partitioned tables as there is no work to be done. The
2141 : * useful work is on their child partitions, which have been queued up for
2142 : * us separately.
2143 : */
2144 272332 : if (rel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
2145 : {
2146 186 : relation_close(rel, lmode);
2147 186 : PopActiveSnapshot();
2148 186 : CommitTransactionCommand();
2149 : /* It's OK to proceed with ANALYZE on this table */
2150 186 : return true;
2151 : }
2152 :
2153 : /*
2154 : * Get a session-level lock too. This will protect our access to the
2155 : * relation across multiple transactions, so that we can vacuum the
2156 : * relation's TOAST table (if any) secure in the knowledge that no one is
2157 : * deleting the parent relation.
2158 : *
2159 : * NOTE: this cannot block, even if someone else is waiting for access,
2160 : * because the lock manager knows that both lock requests are from the
2161 : * same process.
2162 : */
2163 272146 : lockrelid = rel->rd_lockInfo.lockRelId;
2164 272146 : LockRelationIdForSession(&lockrelid, lmode);
2165 :
2166 : /*
2167 : * Set index_cleanup option based on index_cleanup reloption if it wasn't
2168 : * specified in VACUUM command, or when running in an autovacuum worker
2169 : */
2170 272146 : if (params.index_cleanup == VACOPTVALUE_UNSPECIFIED)
2171 : {
2172 : StdRdOptIndexCleanup vacuum_index_cleanup;
2173 :
2174 271888 : if (rel->rd_options == NULL)
2175 268458 : vacuum_index_cleanup = STDRD_OPTION_VACUUM_INDEX_CLEANUP_AUTO;
2176 : else
2177 3430 : vacuum_index_cleanup =
2178 3430 : ((StdRdOptions *) rel->rd_options)->vacuum_index_cleanup;
2179 :
2180 271888 : if (vacuum_index_cleanup == STDRD_OPTION_VACUUM_INDEX_CLEANUP_AUTO)
2181 271844 : params.index_cleanup = VACOPTVALUE_AUTO;
2182 44 : else if (vacuum_index_cleanup == STDRD_OPTION_VACUUM_INDEX_CLEANUP_ON)
2183 22 : params.index_cleanup = VACOPTVALUE_ENABLED;
2184 : else
2185 : {
2186 : Assert(vacuum_index_cleanup ==
2187 : STDRD_OPTION_VACUUM_INDEX_CLEANUP_OFF);
2188 22 : params.index_cleanup = VACOPTVALUE_DISABLED;
2189 : }
2190 : }
2191 :
2192 : #ifdef USE_INJECTION_POINTS
2193 272146 : if (params.index_cleanup == VACOPTVALUE_AUTO)
2194 271850 : INJECTION_POINT("vacuum-index-cleanup-auto", NULL);
2195 296 : else if (params.index_cleanup == VACOPTVALUE_DISABLED)
2196 260 : INJECTION_POINT("vacuum-index-cleanup-disabled", NULL);
2197 36 : else if (params.index_cleanup == VACOPTVALUE_ENABLED)
2198 36 : INJECTION_POINT("vacuum-index-cleanup-enabled", NULL);
2199 : #endif
2200 :
2201 : /*
2202 : * Check if the vacuum_max_eager_freeze_failure_rate table storage
2203 : * parameter was specified. This overrides the GUC value.
2204 : */
2205 272146 : if (rel->rd_options != NULL &&
2206 3442 : ((StdRdOptions *) rel->rd_options)->vacuum_max_eager_freeze_failure_rate >= 0)
2207 0 : params.max_eager_freeze_failure_rate =
2208 0 : ((StdRdOptions *) rel->rd_options)->vacuum_max_eager_freeze_failure_rate;
2209 :
2210 : /*
2211 : * Set truncate option based on truncate reloption or GUC if it wasn't
2212 : * specified in VACUUM command, or when running in an autovacuum worker
2213 : */
2214 272146 : if (params.truncate == VACOPTVALUE_UNSPECIFIED)
2215 : {
2216 271894 : StdRdOptions *opts = (StdRdOptions *) rel->rd_options;
2217 :
2218 271894 : if (opts && opts->vacuum_truncate_set)
2219 : {
2220 32 : if (opts->vacuum_truncate)
2221 10 : params.truncate = VACOPTVALUE_ENABLED;
2222 : else
2223 22 : params.truncate = VACOPTVALUE_DISABLED;
2224 : }
2225 271862 : else if (vacuum_truncate)
2226 271840 : params.truncate = VACOPTVALUE_ENABLED;
2227 : else
2228 22 : params.truncate = VACOPTVALUE_DISABLED;
2229 : }
2230 :
2231 : #ifdef USE_INJECTION_POINTS
2232 272146 : if (params.truncate == VACOPTVALUE_AUTO)
2233 0 : INJECTION_POINT("vacuum-truncate-auto", NULL);
2234 272146 : else if (params.truncate == VACOPTVALUE_DISABLED)
2235 296 : INJECTION_POINT("vacuum-truncate-disabled", NULL);
2236 271850 : else if (params.truncate == VACOPTVALUE_ENABLED)
2237 271850 : INJECTION_POINT("vacuum-truncate-enabled", NULL);
2238 : #endif
2239 :
2240 : /*
2241 : * Remember the relation's TOAST relation for later, if the caller asked
2242 : * us to process it. In VACUUM FULL, though, the toast table is
2243 : * automatically rebuilt by cluster_rel so we shouldn't recurse to it,
2244 : * unless PROCESS_MAIN is disabled.
2245 : */
2246 272146 : if ((params.options & VACOPT_PROCESS_TOAST) != 0 &&
2247 26358 : ((params.options & VACOPT_FULL) == 0 ||
2248 386 : (params.options & VACOPT_PROCESS_MAIN) == 0))
2249 25978 : toast_relid = rel->rd_rel->reltoastrelid;
2250 : else
2251 246168 : toast_relid = InvalidOid;
2252 :
2253 : /*
2254 : * Switch to the table owner's userid, so that any index functions are run
2255 : * as that user. Also lock down security-restricted operations and
2256 : * arrange to make GUC variable changes local to this command. (This is
2257 : * unnecessary, but harmless, for lazy VACUUM.)
2258 : */
2259 272146 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
2260 272146 : SetUserIdAndSecContext(rel->rd_rel->relowner,
2261 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
2262 272146 : save_nestlevel = NewGUCNestLevel();
2263 272146 : RestrictSearchPath();
2264 :
2265 : /*
2266 : * If PROCESS_MAIN is set (the default), it's time to vacuum the main
2267 : * relation. Otherwise, we can skip this part. If processing the TOAST
2268 : * table is required (e.g., PROCESS_TOAST is set), we force PROCESS_MAIN
2269 : * to be set when we recurse to the TOAST table.
2270 : */
2271 272146 : if (params.options & VACOPT_PROCESS_MAIN)
2272 : {
2273 : /*
2274 : * Do the actual work --- either FULL or "lazy" vacuum
2275 : */
2276 271992 : if (params.options & VACOPT_FULL)
2277 : {
2278 380 : ClusterParams cluster_params = {0};
2279 :
2280 380 : if ((params.options & VACOPT_VERBOSE) != 0)
2281 2 : cluster_params.options |= CLUOPT_VERBOSE;
2282 :
2283 : /* VACUUM FULL is now a variant of CLUSTER; see cluster.c */
2284 380 : cluster_rel(rel, InvalidOid, &cluster_params);
2285 : /* cluster_rel closes the relation, but keeps lock */
2286 :
2287 374 : rel = NULL;
2288 : }
2289 : else
2290 271612 : table_relation_vacuum(rel, params, bstrategy);
2291 : }
2292 :
2293 : /* Roll back any GUC changes executed by index functions */
2294 272138 : AtEOXact_GUC(false, save_nestlevel);
2295 :
2296 : /* Restore userid and security context */
2297 272138 : SetUserIdAndSecContext(save_userid, save_sec_context);
2298 :
2299 : /* all done with this class, but hold lock until commit */
2300 272138 : if (rel)
2301 271764 : relation_close(rel, NoLock);
2302 :
2303 : /*
2304 : * Complete the transaction and free all temporary memory used.
2305 : */
2306 272138 : PopActiveSnapshot();
2307 272138 : CommitTransactionCommand();
2308 :
2309 : /*
2310 : * If the relation has a secondary toast rel, vacuum that too while we
2311 : * still hold the session lock on the main table. Note however that
2312 : * "analyze" will not get done on the toast table. This is good, because
2313 : * the toaster always uses hardcoded index access and statistics are
2314 : * totally unimportant for toast relations.
2315 : */
2316 272138 : if (toast_relid != InvalidOid)
2317 : {
2318 : /*
2319 : * Force VACOPT_PROCESS_MAIN so vacuum_rel() processes it. Likewise,
2320 : * set toast_parent so that the privilege checks are done on the main
2321 : * relation. NB: This is only safe to do because we hold a session
2322 : * lock on the main relation that prevents concurrent deletion.
2323 : */
2324 9176 : toast_vacuum_params.options |= VACOPT_PROCESS_MAIN;
2325 9176 : toast_vacuum_params.toast_parent = relid;
2326 :
2327 9176 : vacuum_rel(toast_relid, NULL, toast_vacuum_params, bstrategy);
2328 : }
2329 :
2330 : /*
2331 : * Now release the session-level lock on the main table.
2332 : */
2333 272138 : UnlockRelationIdForSession(&lockrelid, lmode);
2334 :
2335 : /* Report that we really did it. */
2336 272138 : return true;
2337 : }
2338 :
2339 :
2340 : /*
2341 : * Open all the vacuumable indexes of the given relation, obtaining the
2342 : * specified kind of lock on each. Return an array of Relation pointers for
2343 : * the indexes into *Irel, and the number of indexes into *nindexes.
2344 : *
2345 : * We consider an index vacuumable if it is marked insertable (indisready).
2346 : * If it isn't, probably a CREATE INDEX CONCURRENTLY command failed early in
2347 : * execution, and what we have is too corrupt to be processable. We will
2348 : * vacuum even if the index isn't indisvalid; this is important because in a
2349 : * unique index, uniqueness checks will be performed anyway and had better not
2350 : * hit dangling index pointers.
2351 : */
2352 : void
2353 287162 : vac_open_indexes(Relation relation, LOCKMODE lockmode,
2354 : int *nindexes, Relation **Irel)
2355 : {
2356 : List *indexoidlist;
2357 : ListCell *indexoidscan;
2358 : int i;
2359 :
2360 : Assert(lockmode != NoLock);
2361 :
2362 287162 : indexoidlist = RelationGetIndexList(relation);
2363 :
2364 : /* allocate enough memory for all indexes */
2365 287162 : i = list_length(indexoidlist);
2366 :
2367 287162 : if (i > 0)
2368 271060 : *Irel = (Relation *) palloc(i * sizeof(Relation));
2369 : else
2370 16102 : *Irel = NULL;
2371 :
2372 : /* collect just the ready indexes */
2373 287162 : i = 0;
2374 716826 : foreach(indexoidscan, indexoidlist)
2375 : {
2376 429664 : Oid indexoid = lfirst_oid(indexoidscan);
2377 : Relation indrel;
2378 :
2379 429664 : indrel = index_open(indexoid, lockmode);
2380 429664 : if (indrel->rd_index->indisready)
2381 429664 : (*Irel)[i++] = indrel;
2382 : else
2383 0 : index_close(indrel, lockmode);
2384 : }
2385 :
2386 287162 : *nindexes = i;
2387 :
2388 287162 : list_free(indexoidlist);
2389 287162 : }
2390 :
2391 : /*
2392 : * Release the resources acquired by vac_open_indexes. Optionally release
2393 : * the locks (say NoLock to keep 'em).
2394 : */
2395 : void
2396 288034 : vac_close_indexes(int nindexes, Relation *Irel, LOCKMODE lockmode)
2397 : {
2398 288034 : if (Irel == NULL)
2399 16980 : return;
2400 :
2401 700706 : while (nindexes--)
2402 : {
2403 429652 : Relation ind = Irel[nindexes];
2404 :
2405 429652 : index_close(ind, lockmode);
2406 : }
2407 271054 : pfree(Irel);
2408 : }
2409 :
2410 : /*
2411 : * vacuum_delay_point --- check for interrupts and cost-based delay.
2412 : *
2413 : * This should be called in each major loop of VACUUM processing,
2414 : * typically once per page processed.
2415 : */
2416 : void
2417 87418188 : vacuum_delay_point(bool is_analyze)
2418 : {
2419 87418188 : double msec = 0;
2420 :
2421 : /* Always check for interrupts */
2422 87418188 : CHECK_FOR_INTERRUPTS();
2423 :
2424 87418188 : if (InterruptPending ||
2425 87418188 : (!VacuumCostActive && !ConfigReloadPending))
2426 76346144 : return;
2427 :
2428 : /*
2429 : * Autovacuum workers should reload the configuration file if requested.
2430 : * This allows changes to [autovacuum_]vacuum_cost_limit and
2431 : * [autovacuum_]vacuum_cost_delay to take effect while a table is being
2432 : * vacuumed or analyzed.
2433 : */
2434 11072044 : if (ConfigReloadPending && AmAutoVacuumWorkerProcess())
2435 : {
2436 0 : ConfigReloadPending = false;
2437 0 : ProcessConfigFile(PGC_SIGHUP);
2438 0 : VacuumUpdateCosts();
2439 : }
2440 :
2441 : /*
2442 : * If we disabled cost-based delays after reloading the config file,
2443 : * return.
2444 : */
2445 11072044 : if (!VacuumCostActive)
2446 0 : return;
2447 :
2448 : /*
2449 : * For parallel vacuum, the delay is computed based on the shared cost
2450 : * balance. See compute_parallel_delay.
2451 : */
2452 11072044 : if (VacuumSharedCostBalance != NULL)
2453 0 : msec = compute_parallel_delay();
2454 11072044 : else if (VacuumCostBalance >= vacuum_cost_limit)
2455 8970 : msec = vacuum_cost_delay * VacuumCostBalance / vacuum_cost_limit;
2456 :
2457 : /* Nap if appropriate */
2458 11072044 : if (msec > 0)
2459 : {
2460 : instr_time delay_start;
2461 :
2462 8970 : if (msec > vacuum_cost_delay * 4)
2463 14 : msec = vacuum_cost_delay * 4;
2464 :
2465 8970 : if (track_cost_delay_timing)
2466 0 : INSTR_TIME_SET_CURRENT(delay_start);
2467 :
2468 8970 : pgstat_report_wait_start(WAIT_EVENT_VACUUM_DELAY);
2469 8970 : pg_usleep(msec * 1000);
2470 8970 : pgstat_report_wait_end();
2471 :
2472 8970 : if (track_cost_delay_timing)
2473 : {
2474 : instr_time delay_end;
2475 : instr_time delay;
2476 :
2477 0 : INSTR_TIME_SET_CURRENT(delay_end);
2478 0 : INSTR_TIME_SET_ZERO(delay);
2479 0 : INSTR_TIME_ACCUM_DIFF(delay, delay_end, delay_start);
2480 :
2481 : /*
2482 : * For parallel workers, we only report the delay time every once
2483 : * in a while to avoid overloading the leader with messages and
2484 : * interrupts.
2485 : */
2486 0 : if (IsParallelWorker())
2487 : {
2488 : static instr_time last_report_time;
2489 : instr_time time_since_last_report;
2490 :
2491 : Assert(!is_analyze);
2492 :
2493 : /* Accumulate the delay time */
2494 0 : parallel_vacuum_worker_delay_ns += INSTR_TIME_GET_NANOSEC(delay);
2495 :
2496 : /* Calculate interval since last report */
2497 0 : INSTR_TIME_SET_ZERO(time_since_last_report);
2498 0 : INSTR_TIME_ACCUM_DIFF(time_since_last_report, delay_end, last_report_time);
2499 :
2500 : /* If we haven't reported in a while, do so now */
2501 0 : if (INSTR_TIME_GET_NANOSEC(time_since_last_report) >=
2502 : PARALLEL_VACUUM_DELAY_REPORT_INTERVAL_NS)
2503 : {
2504 0 : pgstat_progress_parallel_incr_param(PROGRESS_VACUUM_DELAY_TIME,
2505 : parallel_vacuum_worker_delay_ns);
2506 :
2507 : /* Reset variables */
2508 0 : last_report_time = delay_end;
2509 0 : parallel_vacuum_worker_delay_ns = 0;
2510 : }
2511 : }
2512 0 : else if (is_analyze)
2513 0 : pgstat_progress_incr_param(PROGRESS_ANALYZE_DELAY_TIME,
2514 0 : INSTR_TIME_GET_NANOSEC(delay));
2515 : else
2516 0 : pgstat_progress_incr_param(PROGRESS_VACUUM_DELAY_TIME,
2517 0 : INSTR_TIME_GET_NANOSEC(delay));
2518 : }
2519 :
2520 : /*
2521 : * We don't want to ignore postmaster death during very long vacuums
2522 : * with vacuum_cost_delay configured. We can't use the usual
2523 : * WaitLatch() approach here because we want microsecond-based sleep
2524 : * durations above.
2525 : */
2526 8970 : if (IsUnderPostmaster && !PostmasterIsAlive())
2527 0 : exit(1);
2528 :
2529 8970 : VacuumCostBalance = 0;
2530 :
2531 : /*
2532 : * Balance and update limit values for autovacuum workers. We must do
2533 : * this periodically, as the number of workers across which we are
2534 : * balancing the limit may have changed.
2535 : *
2536 : * TODO: There may be better criteria for determining when to do this
2537 : * besides "check after napping".
2538 : */
2539 8970 : AutoVacuumUpdateCostLimit();
2540 :
2541 : /* Might have gotten an interrupt while sleeping */
2542 8970 : CHECK_FOR_INTERRUPTS();
2543 : }
2544 : }
2545 :
2546 : /*
2547 : * Computes the vacuum delay for parallel workers.
2548 : *
2549 : * The basic idea of a cost-based delay for parallel vacuum is to allow each
2550 : * worker to sleep in proportion to the share of work it's done. We achieve this
2551 : * by allowing all parallel vacuum workers including the leader process to
2552 : * have a shared view of cost related parameters (mainly VacuumCostBalance).
2553 : * We allow each worker to update it as and when it has incurred any cost and
2554 : * then based on that decide whether it needs to sleep. We compute the time
2555 : * to sleep for a worker based on the cost it has incurred
2556 : * (VacuumCostBalanceLocal) and then reduce the VacuumSharedCostBalance by
2557 : * that amount. This avoids putting to sleep those workers which have done less
2558 : * I/O than other workers and therefore ensure that workers
2559 : * which are doing more I/O got throttled more.
2560 : *
2561 : * We allow a worker to sleep only if it has performed I/O above a certain
2562 : * threshold, which is calculated based on the number of active workers
2563 : * (VacuumActiveNWorkers), and the overall cost balance is more than
2564 : * VacuumCostLimit set by the system. Testing reveals that we achieve
2565 : * the required throttling if we force a worker that has done more than 50%
2566 : * of its share of work to sleep.
2567 : */
2568 : static double
2569 0 : compute_parallel_delay(void)
2570 : {
2571 0 : double msec = 0;
2572 : uint32 shared_balance;
2573 : int nworkers;
2574 :
2575 : /* Parallel vacuum must be active */
2576 : Assert(VacuumSharedCostBalance);
2577 :
2578 0 : nworkers = pg_atomic_read_u32(VacuumActiveNWorkers);
2579 :
2580 : /* At least count itself */
2581 : Assert(nworkers >= 1);
2582 :
2583 : /* Update the shared cost balance value atomically */
2584 0 : shared_balance = pg_atomic_add_fetch_u32(VacuumSharedCostBalance, VacuumCostBalance);
2585 :
2586 : /* Compute the total local balance for the current worker */
2587 0 : VacuumCostBalanceLocal += VacuumCostBalance;
2588 :
2589 0 : if ((shared_balance >= vacuum_cost_limit) &&
2590 0 : (VacuumCostBalanceLocal > 0.5 * ((double) vacuum_cost_limit / nworkers)))
2591 : {
2592 : /* Compute sleep time based on the local cost balance */
2593 0 : msec = vacuum_cost_delay * VacuumCostBalanceLocal / vacuum_cost_limit;
2594 0 : pg_atomic_sub_fetch_u32(VacuumSharedCostBalance, VacuumCostBalanceLocal);
2595 0 : VacuumCostBalanceLocal = 0;
2596 : }
2597 :
2598 : /*
2599 : * Reset the local balance as we accumulated it into the shared value.
2600 : */
2601 0 : VacuumCostBalance = 0;
2602 :
2603 0 : return msec;
2604 : }
2605 :
2606 : /*
2607 : * A wrapper function of defGetBoolean().
2608 : *
2609 : * This function returns VACOPTVALUE_ENABLED and VACOPTVALUE_DISABLED instead
2610 : * of true and false.
2611 : */
2612 : static VacOptValue
2613 326 : get_vacoptval_from_boolean(DefElem *def)
2614 : {
2615 326 : return defGetBoolean(def) ? VACOPTVALUE_ENABLED : VACOPTVALUE_DISABLED;
2616 : }
2617 :
2618 : /*
2619 : * vac_bulkdel_one_index() -- bulk-deletion for index relation.
2620 : *
2621 : * Returns bulk delete stats derived from input stats
2622 : */
2623 : IndexBulkDeleteResult *
2624 2512 : vac_bulkdel_one_index(IndexVacuumInfo *ivinfo, IndexBulkDeleteResult *istat,
2625 : TidStore *dead_items, VacDeadItemsInfo *dead_items_info)
2626 : {
2627 : /* Do bulk deletion */
2628 2512 : istat = index_bulk_delete(ivinfo, istat, vac_tid_reaped,
2629 : dead_items);
2630 :
2631 2512 : ereport(ivinfo->message_level,
2632 : (errmsg("scanned index \"%s\" to remove %" PRId64 " row versions",
2633 : RelationGetRelationName(ivinfo->index),
2634 : dead_items_info->num_items)));
2635 :
2636 2512 : return istat;
2637 : }
2638 :
2639 : /*
2640 : * vac_cleanup_one_index() -- do post-vacuum cleanup for index relation.
2641 : *
2642 : * Returns bulk delete stats derived from input stats
2643 : */
2644 : IndexBulkDeleteResult *
2645 297668 : vac_cleanup_one_index(IndexVacuumInfo *ivinfo, IndexBulkDeleteResult *istat)
2646 : {
2647 297668 : istat = index_vacuum_cleanup(ivinfo, istat);
2648 :
2649 297668 : if (istat)
2650 2770 : ereport(ivinfo->message_level,
2651 : (errmsg("index \"%s\" now contains %.0f row versions in %u pages",
2652 : RelationGetRelationName(ivinfo->index),
2653 : istat->num_index_tuples,
2654 : istat->num_pages),
2655 : errdetail("%.0f index row versions were removed.\n"
2656 : "%u index pages were newly deleted.\n"
2657 : "%u index pages are currently deleted, of which %u are currently reusable.",
2658 : istat->tuples_removed,
2659 : istat->pages_newly_deleted,
2660 : istat->pages_deleted, istat->pages_free)));
2661 :
2662 297668 : return istat;
2663 : }
2664 :
2665 : /*
2666 : * vac_tid_reaped() -- is a particular tid deletable?
2667 : *
2668 : * This has the right signature to be an IndexBulkDeleteCallback.
2669 : */
2670 : static bool
2671 6766076 : vac_tid_reaped(ItemPointer itemptr, void *state)
2672 : {
2673 6766076 : TidStore *dead_items = (TidStore *) state;
2674 :
2675 6766076 : return TidStoreIsMember(dead_items, itemptr);
2676 : }
|
