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/vacuum.h"
43 : #include "miscadmin.h"
44 : #include "nodes/makefuncs.h"
45 : #include "pgstat.h"
46 : #include "postmaster/autovacuum.h"
47 : #include "postmaster/bgworker_internals.h"
48 : #include "postmaster/interrupt.h"
49 : #include "storage/bufmgr.h"
50 : #include "storage/lmgr.h"
51 : #include "storage/pmsignal.h"
52 : #include "storage/proc.h"
53 : #include "storage/procarray.h"
54 : #include "utils/acl.h"
55 : #include "utils/fmgroids.h"
56 : #include "utils/guc.h"
57 : #include "utils/guc_hooks.h"
58 : #include "utils/memutils.h"
59 : #include "utils/snapmgr.h"
60 : #include "utils/syscache.h"
61 :
62 :
63 : /*
64 : * GUC parameters
65 : */
66 : int vacuum_freeze_min_age;
67 : int vacuum_freeze_table_age;
68 : int vacuum_multixact_freeze_min_age;
69 : int vacuum_multixact_freeze_table_age;
70 : int vacuum_failsafe_age;
71 : int vacuum_multixact_failsafe_age;
72 :
73 : /*
74 : * Variables for cost-based vacuum delay. The defaults differ between
75 : * autovacuum and vacuum. They should be set with the appropriate GUC value in
76 : * vacuum code. They are initialized here to the defaults for client backends
77 : * executing VACUUM or ANALYZE.
78 : */
79 : double vacuum_cost_delay = 0;
80 : int vacuum_cost_limit = 200;
81 :
82 : /*
83 : * VacuumFailsafeActive is a defined as a global so that we can determine
84 : * whether or not to re-enable cost-based vacuum delay when vacuuming a table.
85 : * If failsafe mode has been engaged, we will not re-enable cost-based delay
86 : * for the table until after vacuuming has completed, regardless of other
87 : * settings.
88 : *
89 : * Only VACUUM code should inspect this variable and only table access methods
90 : * should set it to true. In Table AM-agnostic VACUUM code, this variable is
91 : * inspected to determine whether or not to allow cost-based delays. Table AMs
92 : * are free to set it if they desire this behavior, but it is false by default
93 : * and reset to false in between vacuuming each relation.
94 : */
95 : bool VacuumFailsafeActive = false;
96 :
97 : /*
98 : * Variables for cost-based parallel vacuum. See comments atop
99 : * compute_parallel_delay to understand how it works.
100 : */
101 : pg_atomic_uint32 *VacuumSharedCostBalance = NULL;
102 : pg_atomic_uint32 *VacuumActiveNWorkers = NULL;
103 : int VacuumCostBalanceLocal = 0;
104 :
105 : /* non-export function prototypes */
106 : static List *expand_vacuum_rel(VacuumRelation *vrel,
107 : MemoryContext vac_context, int options);
108 : static List *get_all_vacuum_rels(MemoryContext vac_context, int options);
109 : static void vac_truncate_clog(TransactionId frozenXID,
110 : MultiXactId minMulti,
111 : TransactionId lastSaneFrozenXid,
112 : MultiXactId lastSaneMinMulti);
113 : static bool vacuum_rel(Oid relid, RangeVar *relation, VacuumParams *params,
114 : BufferAccessStrategy bstrategy);
115 : static double compute_parallel_delay(void);
116 : static VacOptValue get_vacoptval_from_boolean(DefElem *def);
117 : static bool vac_tid_reaped(ItemPointer itemptr, void *state);
118 :
119 : /*
120 : * GUC check function to ensure GUC value specified is within the allowable
121 : * range.
122 : */
123 : bool
124 1982 : check_vacuum_buffer_usage_limit(int *newval, void **extra,
125 : GucSource source)
126 : {
127 : /* Value upper and lower hard limits are inclusive */
128 1982 : if (*newval == 0 || (*newval >= MIN_BAS_VAC_RING_SIZE_KB &&
129 1982 : *newval <= MAX_BAS_VAC_RING_SIZE_KB))
130 1982 : return true;
131 :
132 : /* Value does not fall within any allowable range */
133 0 : GUC_check_errdetail("\"%s\" must be 0 or between %d kB and %d kB.",
134 : "vacuum_buffer_usage_limit",
135 : MIN_BAS_VAC_RING_SIZE_KB, MAX_BAS_VAC_RING_SIZE_KB);
136 :
137 0 : return false;
138 : }
139 :
140 : /*
141 : * Primary entry point for manual VACUUM and ANALYZE commands
142 : *
143 : * This is mainly a preparation wrapper for the real operations that will
144 : * happen in vacuum().
145 : */
146 : void
147 10734 : ExecVacuum(ParseState *pstate, VacuumStmt *vacstmt, bool isTopLevel)
148 : {
149 : VacuumParams params;
150 10734 : BufferAccessStrategy bstrategy = NULL;
151 10734 : bool verbose = false;
152 10734 : bool skip_locked = false;
153 10734 : bool analyze = false;
154 10734 : bool freeze = false;
155 10734 : bool full = false;
156 10734 : bool disable_page_skipping = false;
157 10734 : bool process_main = true;
158 10734 : bool process_toast = true;
159 : int ring_size;
160 10734 : bool skip_database_stats = false;
161 10734 : bool only_database_stats = false;
162 : MemoryContext vac_context;
163 : ListCell *lc;
164 :
165 : /* index_cleanup and truncate values unspecified for now */
166 10734 : params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
167 10734 : params.truncate = VACOPTVALUE_UNSPECIFIED;
168 :
169 : /* By default parallel vacuum is enabled */
170 10734 : params.nworkers = 0;
171 :
172 : /* Will be set later if we recurse to a TOAST table. */
173 10734 : params.toast_parent = InvalidOid;
174 :
175 : /*
176 : * Set this to an invalid value so it is clear whether or not a
177 : * BUFFER_USAGE_LIMIT was specified when making the access strategy.
178 : */
179 10734 : ring_size = -1;
180 :
181 : /* Parse options list */
182 19712 : foreach(lc, vacstmt->options)
183 : {
184 9014 : DefElem *opt = (DefElem *) lfirst(lc);
185 :
186 : /* Parse common options for VACUUM and ANALYZE */
187 9014 : if (strcmp(opt->defname, "verbose") == 0)
188 38 : verbose = defGetBoolean(opt);
189 8976 : else if (strcmp(opt->defname, "skip_locked") == 0)
190 334 : skip_locked = defGetBoolean(opt);
191 8642 : else if (strcmp(opt->defname, "buffer_usage_limit") == 0)
192 : {
193 : const char *hintmsg;
194 : int result;
195 : char *vac_buffer_size;
196 :
197 54 : vac_buffer_size = defGetString(opt);
198 :
199 : /*
200 : * Check that the specified value is valid and the size falls
201 : * within the hard upper and lower limits if it is not 0.
202 : */
203 54 : if (!parse_int(vac_buffer_size, &result, GUC_UNIT_KB, &hintmsg) ||
204 48 : (result != 0 &&
205 36 : (result < MIN_BAS_VAC_RING_SIZE_KB || result > MAX_BAS_VAC_RING_SIZE_KB)))
206 : {
207 18 : ereport(ERROR,
208 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
209 : errmsg("BUFFER_USAGE_LIMIT option must be 0 or between %d kB and %d kB",
210 : MIN_BAS_VAC_RING_SIZE_KB, MAX_BAS_VAC_RING_SIZE_KB),
211 : hintmsg ? errhint("%s", _(hintmsg)) : 0));
212 : }
213 :
214 36 : ring_size = result;
215 : }
216 8588 : else if (!vacstmt->is_vacuumcmd)
217 6 : ereport(ERROR,
218 : (errcode(ERRCODE_SYNTAX_ERROR),
219 : errmsg("unrecognized ANALYZE option \"%s\"", opt->defname),
220 : parser_errposition(pstate, opt->location)));
221 :
222 : /* Parse options available on VACUUM */
223 8582 : else if (strcmp(opt->defname, "analyze") == 0)
224 1376 : analyze = defGetBoolean(opt);
225 7206 : else if (strcmp(opt->defname, "freeze") == 0)
226 1162 : freeze = defGetBoolean(opt);
227 6044 : else if (strcmp(opt->defname, "full") == 0)
228 370 : full = defGetBoolean(opt);
229 5674 : else if (strcmp(opt->defname, "disable_page_skipping") == 0)
230 184 : disable_page_skipping = defGetBoolean(opt);
231 5490 : else if (strcmp(opt->defname, "index_cleanup") == 0)
232 : {
233 : /* Interpret no string as the default, which is 'auto' */
234 174 : if (!opt->arg)
235 0 : params.index_cleanup = VACOPTVALUE_AUTO;
236 : else
237 : {
238 174 : char *sval = defGetString(opt);
239 :
240 : /* Try matching on 'auto' string, or fall back on boolean */
241 174 : if (pg_strcasecmp(sval, "auto") == 0)
242 6 : params.index_cleanup = VACOPTVALUE_AUTO;
243 : else
244 168 : params.index_cleanup = get_vacoptval_from_boolean(opt);
245 : }
246 : }
247 5316 : else if (strcmp(opt->defname, "process_main") == 0)
248 154 : process_main = defGetBoolean(opt);
249 5162 : else if (strcmp(opt->defname, "process_toast") == 0)
250 160 : process_toast = defGetBoolean(opt);
251 5002 : else if (strcmp(opt->defname, "truncate") == 0)
252 148 : params.truncate = get_vacoptval_from_boolean(opt);
253 4854 : else if (strcmp(opt->defname, "parallel") == 0)
254 : {
255 338 : if (opt->arg == NULL)
256 : {
257 6 : ereport(ERROR,
258 : (errcode(ERRCODE_SYNTAX_ERROR),
259 : errmsg("parallel option requires a value between 0 and %d",
260 : MAX_PARALLEL_WORKER_LIMIT),
261 : parser_errposition(pstate, opt->location)));
262 : }
263 : else
264 : {
265 : int nworkers;
266 :
267 332 : nworkers = defGetInt32(opt);
268 332 : if (nworkers < 0 || nworkers > MAX_PARALLEL_WORKER_LIMIT)
269 6 : ereport(ERROR,
270 : (errcode(ERRCODE_SYNTAX_ERROR),
271 : errmsg("parallel workers for vacuum must be between 0 and %d",
272 : MAX_PARALLEL_WORKER_LIMIT),
273 : parser_errposition(pstate, opt->location)));
274 :
275 : /*
276 : * Disable parallel vacuum, if user has specified parallel
277 : * degree as zero.
278 : */
279 326 : if (nworkers == 0)
280 154 : params.nworkers = -1;
281 : else
282 172 : params.nworkers = nworkers;
283 : }
284 : }
285 4516 : else if (strcmp(opt->defname, "skip_database_stats") == 0)
286 4398 : skip_database_stats = defGetBoolean(opt);
287 118 : else if (strcmp(opt->defname, "only_database_stats") == 0)
288 118 : only_database_stats = defGetBoolean(opt);
289 : else
290 0 : ereport(ERROR,
291 : (errcode(ERRCODE_SYNTAX_ERROR),
292 : errmsg("unrecognized VACUUM option \"%s\"", opt->defname),
293 : parser_errposition(pstate, opt->location)));
294 : }
295 :
296 : /* Set vacuum options */
297 10698 : params.options =
298 10698 : (vacstmt->is_vacuumcmd ? VACOPT_VACUUM : VACOPT_ANALYZE) |
299 10698 : (verbose ? VACOPT_VERBOSE : 0) |
300 10698 : (skip_locked ? VACOPT_SKIP_LOCKED : 0) |
301 10698 : (analyze ? VACOPT_ANALYZE : 0) |
302 10698 : (freeze ? VACOPT_FREEZE : 0) |
303 10698 : (full ? VACOPT_FULL : 0) |
304 10698 : (disable_page_skipping ? VACOPT_DISABLE_PAGE_SKIPPING : 0) |
305 10698 : (process_main ? VACOPT_PROCESS_MAIN : 0) |
306 10698 : (process_toast ? VACOPT_PROCESS_TOAST : 0) |
307 10698 : (skip_database_stats ? VACOPT_SKIP_DATABASE_STATS : 0) |
308 10698 : (only_database_stats ? VACOPT_ONLY_DATABASE_STATS : 0);
309 :
310 : /* sanity checks on options */
311 : Assert(params.options & (VACOPT_VACUUM | VACOPT_ANALYZE));
312 : Assert((params.options & VACOPT_VACUUM) ||
313 : !(params.options & (VACOPT_FULL | VACOPT_FREEZE)));
314 :
315 10698 : if ((params.options & VACOPT_FULL) && params.nworkers > 0)
316 6 : ereport(ERROR,
317 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
318 : errmsg("VACUUM FULL cannot be performed in parallel")));
319 :
320 : /*
321 : * BUFFER_USAGE_LIMIT does nothing for VACUUM (FULL) so just raise an
322 : * ERROR for that case. VACUUM (FULL, ANALYZE) does make use of it, so
323 : * we'll permit that.
324 : */
325 10692 : if (ring_size != -1 && (params.options & VACOPT_FULL) &&
326 6 : !(params.options & VACOPT_ANALYZE))
327 6 : ereport(ERROR,
328 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
329 : errmsg("BUFFER_USAGE_LIMIT cannot be specified for VACUUM FULL")));
330 :
331 : /*
332 : * Make sure VACOPT_ANALYZE is specified if any column lists are present.
333 : */
334 10686 : if (!(params.options & VACOPT_ANALYZE))
335 : {
336 9324 : foreach(lc, vacstmt->rels)
337 : {
338 4576 : VacuumRelation *vrel = lfirst_node(VacuumRelation, lc);
339 :
340 4576 : if (vrel->va_cols != NIL)
341 6 : ereport(ERROR,
342 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
343 : errmsg("ANALYZE option must be specified when a column list is provided")));
344 : }
345 : }
346 :
347 :
348 : /*
349 : * Sanity check DISABLE_PAGE_SKIPPING option.
350 : */
351 10680 : if ((params.options & VACOPT_FULL) != 0 &&
352 346 : (params.options & VACOPT_DISABLE_PAGE_SKIPPING) != 0)
353 0 : ereport(ERROR,
354 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
355 : errmsg("VACUUM option DISABLE_PAGE_SKIPPING cannot be used with FULL")));
356 :
357 : /* sanity check for PROCESS_TOAST */
358 10680 : if ((params.options & VACOPT_FULL) != 0 &&
359 346 : (params.options & VACOPT_PROCESS_TOAST) == 0)
360 6 : ereport(ERROR,
361 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
362 : errmsg("PROCESS_TOAST required with VACUUM FULL")));
363 :
364 : /* sanity check for ONLY_DATABASE_STATS */
365 10674 : if (params.options & VACOPT_ONLY_DATABASE_STATS)
366 : {
367 : Assert(params.options & VACOPT_VACUUM);
368 118 : if (vacstmt->rels != NIL)
369 6 : ereport(ERROR,
370 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
371 : errmsg("ONLY_DATABASE_STATS cannot be specified with a list of tables")));
372 : /* don't require people to turn off PROCESS_TOAST/MAIN explicitly */
373 112 : if (params.options & ~(VACOPT_VACUUM |
374 : VACOPT_VERBOSE |
375 : VACOPT_PROCESS_MAIN |
376 : VACOPT_PROCESS_TOAST |
377 : VACOPT_ONLY_DATABASE_STATS))
378 0 : ereport(ERROR,
379 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
380 : errmsg("ONLY_DATABASE_STATS cannot be specified with other VACUUM options")));
381 : }
382 :
383 : /*
384 : * All freeze ages are zero if the FREEZE option is given; otherwise pass
385 : * them as -1 which means to use the default values.
386 : */
387 10668 : if (params.options & VACOPT_FREEZE)
388 : {
389 1162 : params.freeze_min_age = 0;
390 1162 : params.freeze_table_age = 0;
391 1162 : params.multixact_freeze_min_age = 0;
392 1162 : params.multixact_freeze_table_age = 0;
393 : }
394 : else
395 : {
396 9506 : params.freeze_min_age = -1;
397 9506 : params.freeze_table_age = -1;
398 9506 : params.multixact_freeze_min_age = -1;
399 9506 : params.multixact_freeze_table_age = -1;
400 : }
401 :
402 : /* user-invoked vacuum is never "for wraparound" */
403 10668 : params.is_wraparound = false;
404 :
405 : /* user-invoked vacuum uses VACOPT_VERBOSE instead of log_min_duration */
406 10668 : params.log_min_duration = -1;
407 :
408 : /*
409 : * Create special memory context for cross-transaction storage.
410 : *
411 : * Since it is a child of PortalContext, it will go away eventually even
412 : * if we suffer an error; there's no need for special abort cleanup logic.
413 : */
414 10668 : vac_context = AllocSetContextCreate(PortalContext,
415 : "Vacuum",
416 : ALLOCSET_DEFAULT_SIZES);
417 :
418 : /*
419 : * Make a buffer strategy object in the cross-transaction memory context.
420 : * We needn't bother making this for VACUUM (FULL) or VACUUM
421 : * (ONLY_DATABASE_STATS) as they'll not make use of it. VACUUM (FULL,
422 : * ANALYZE) is possible, so we'd better ensure that we make a strategy
423 : * when we see ANALYZE.
424 : */
425 10668 : if ((params.options & (VACOPT_ONLY_DATABASE_STATS |
426 452 : VACOPT_FULL)) == 0 ||
427 452 : (params.options & VACOPT_ANALYZE) != 0)
428 : {
429 :
430 10222 : MemoryContext old_context = MemoryContextSwitchTo(vac_context);
431 :
432 : Assert(ring_size >= -1);
433 :
434 : /*
435 : * If BUFFER_USAGE_LIMIT was specified by the VACUUM or ANALYZE
436 : * command, it overrides the value of VacuumBufferUsageLimit. Either
437 : * value may be 0, in which case GetAccessStrategyWithSize() will
438 : * return NULL, effectively allowing full use of shared buffers.
439 : */
440 10222 : if (ring_size == -1)
441 10192 : ring_size = VacuumBufferUsageLimit;
442 :
443 10222 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, ring_size);
444 :
445 10222 : MemoryContextSwitchTo(old_context);
446 : }
447 :
448 : /* Now go through the common routine */
449 10668 : vacuum(vacstmt->rels, ¶ms, bstrategy, vac_context, isTopLevel);
450 :
451 : /* Finally, clean up the vacuum memory context */
452 10542 : MemoryContextDelete(vac_context);
453 10542 : }
454 :
455 : /*
456 : * Internal entry point for autovacuum and the VACUUM / ANALYZE commands.
457 : *
458 : * relations, if not NIL, is a list of VacuumRelation to process; otherwise,
459 : * we process all relevant tables in the database. For each VacuumRelation,
460 : * if a valid OID is supplied, the table with that OID is what to process;
461 : * otherwise, the VacuumRelation's RangeVar indicates what to process.
462 : *
463 : * params contains a set of parameters that can be used to customize the
464 : * behavior.
465 : *
466 : * bstrategy may be passed in as NULL when the caller does not want to
467 : * restrict the number of shared_buffers that VACUUM / ANALYZE can use,
468 : * otherwise, the caller must build a BufferAccessStrategy with the number of
469 : * shared_buffers that VACUUM / ANALYZE should try to limit themselves to
470 : * using.
471 : *
472 : * isTopLevel should be passed down from ProcessUtility.
473 : *
474 : * It is the caller's responsibility that all parameters are allocated in a
475 : * memory context that will not disappear at transaction commit.
476 : */
477 : void
478 104490 : vacuum(List *relations, VacuumParams *params, BufferAccessStrategy bstrategy,
479 : MemoryContext vac_context, bool isTopLevel)
480 : {
481 : static bool in_vacuum = false;
482 :
483 : const char *stmttype;
484 : volatile bool in_outer_xact,
485 : use_own_xacts;
486 :
487 : Assert(params != NULL);
488 :
489 104490 : stmttype = (params->options & VACOPT_VACUUM) ? "VACUUM" : "ANALYZE";
490 :
491 : /*
492 : * We cannot run VACUUM inside a user transaction block; if we were inside
493 : * a transaction, then our commit- and start-transaction-command calls
494 : * would not have the intended effect! There are numerous other subtle
495 : * dependencies on this, too.
496 : *
497 : * ANALYZE (without VACUUM) can run either way.
498 : */
499 104490 : if (params->options & VACOPT_VACUUM)
500 : {
501 99754 : PreventInTransactionBlock(isTopLevel, stmttype);
502 99740 : in_outer_xact = false;
503 : }
504 : else
505 4736 : in_outer_xact = IsInTransactionBlock(isTopLevel);
506 :
507 : /*
508 : * Check for and disallow recursive calls. This could happen when VACUUM
509 : * FULL or ANALYZE calls a hostile index expression that itself calls
510 : * ANALYZE.
511 : */
512 104476 : if (in_vacuum)
513 12 : ereport(ERROR,
514 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
515 : errmsg("%s cannot be executed from VACUUM or ANALYZE",
516 : stmttype)));
517 :
518 : /*
519 : * Build list of relation(s) to process, putting any new data in
520 : * vac_context for safekeeping.
521 : */
522 104464 : if (params->options & VACOPT_ONLY_DATABASE_STATS)
523 : {
524 : /* We don't process any tables in this case */
525 : Assert(relations == NIL);
526 : }
527 104352 : else if (relations != NIL)
528 : {
529 104160 : List *newrels = NIL;
530 : ListCell *lc;
531 :
532 208408 : foreach(lc, relations)
533 : {
534 104284 : VacuumRelation *vrel = lfirst_node(VacuumRelation, lc);
535 : List *sublist;
536 : MemoryContext old_context;
537 :
538 104284 : sublist = expand_vacuum_rel(vrel, vac_context, params->options);
539 104248 : old_context = MemoryContextSwitchTo(vac_context);
540 104248 : newrels = list_concat(newrels, sublist);
541 104248 : MemoryContextSwitchTo(old_context);
542 : }
543 104124 : relations = newrels;
544 : }
545 : else
546 192 : relations = get_all_vacuum_rels(vac_context, params->options);
547 :
548 : /*
549 : * Decide whether we need to start/commit our own transactions.
550 : *
551 : * For VACUUM (with or without ANALYZE): always do so, so that we can
552 : * release locks as soon as possible. (We could possibly use the outer
553 : * transaction for a one-table VACUUM, but handling TOAST tables would be
554 : * problematic.)
555 : *
556 : * For ANALYZE (no VACUUM): if inside a transaction block, we cannot
557 : * start/commit our own transactions. Also, there's no need to do so if
558 : * only processing one relation. For multiple relations when not within a
559 : * transaction block, and also in an autovacuum worker, use own
560 : * transactions so we can release locks sooner.
561 : */
562 104428 : if (params->options & VACOPT_VACUUM)
563 99728 : use_own_xacts = true;
564 : else
565 : {
566 : Assert(params->options & VACOPT_ANALYZE);
567 4700 : if (AmAutoVacuumWorkerProcess())
568 180 : use_own_xacts = true;
569 4520 : else if (in_outer_xact)
570 238 : use_own_xacts = false;
571 4282 : else if (list_length(relations) > 1)
572 722 : use_own_xacts = true;
573 : else
574 3560 : use_own_xacts = false;
575 : }
576 :
577 : /*
578 : * vacuum_rel expects to be entered with no transaction active; it will
579 : * start and commit its own transaction. But we are called by an SQL
580 : * command, and so we are executing inside a transaction already. We
581 : * commit the transaction started in PostgresMain() here, and start
582 : * another one before exiting to match the commit waiting for us back in
583 : * PostgresMain().
584 : */
585 104428 : if (use_own_xacts)
586 : {
587 : Assert(!in_outer_xact);
588 :
589 : /* ActiveSnapshot is not set by autovacuum */
590 100630 : if (ActiveSnapshotSet())
591 6808 : PopActiveSnapshot();
592 :
593 : /* matches the StartTransaction in PostgresMain() */
594 100630 : CommitTransactionCommand();
595 : }
596 :
597 : /* Turn vacuum cost accounting on or off, and set/clear in_vacuum */
598 104428 : PG_TRY();
599 : {
600 : ListCell *cur;
601 :
602 104428 : in_vacuum = true;
603 104428 : VacuumFailsafeActive = false;
604 104428 : VacuumUpdateCosts();
605 104428 : VacuumCostBalance = 0;
606 104428 : VacuumCostBalanceLocal = 0;
607 104428 : VacuumSharedCostBalance = NULL;
608 104428 : VacuumActiveNWorkers = NULL;
609 :
610 : /*
611 : * Loop to process each selected relation.
612 : */
613 224336 : foreach(cur, relations)
614 : {
615 119974 : VacuumRelation *vrel = lfirst_node(VacuumRelation, cur);
616 :
617 119974 : if (params->options & VACOPT_VACUUM)
618 : {
619 107804 : if (!vacuum_rel(vrel->oid, vrel->relation, params, bstrategy))
620 100 : continue;
621 : }
622 :
623 119868 : if (params->options & VACOPT_ANALYZE)
624 : {
625 : /*
626 : * If using separate xacts, start one for analyze. Otherwise,
627 : * we can use the outer transaction.
628 : */
629 13784 : if (use_own_xacts)
630 : {
631 10036 : StartTransactionCommand();
632 : /* functions in indexes may want a snapshot set */
633 10036 : PushActiveSnapshot(GetTransactionSnapshot());
634 : }
635 :
636 13784 : analyze_rel(vrel->oid, vrel->relation, params,
637 : vrel->va_cols, in_outer_xact, bstrategy);
638 :
639 13724 : if (use_own_xacts)
640 : {
641 9996 : PopActiveSnapshot();
642 9996 : CommitTransactionCommand();
643 : }
644 : else
645 : {
646 : /*
647 : * If we're not using separate xacts, better separate the
648 : * ANALYZE actions with CCIs. This avoids trouble if user
649 : * says "ANALYZE t, t".
650 : */
651 3728 : CommandCounterIncrement();
652 : }
653 : }
654 :
655 : /*
656 : * Ensure VacuumFailsafeActive has been reset before vacuuming the
657 : * next relation.
658 : */
659 119808 : VacuumFailsafeActive = false;
660 : }
661 : }
662 64 : PG_FINALLY();
663 : {
664 104426 : in_vacuum = false;
665 104426 : VacuumCostActive = false;
666 104426 : VacuumFailsafeActive = false;
667 104426 : VacuumCostBalance = 0;
668 : }
669 104426 : PG_END_TRY();
670 :
671 : /*
672 : * Finish up processing.
673 : */
674 104362 : if (use_own_xacts)
675 : {
676 : /* here, we are not in a transaction */
677 :
678 : /*
679 : * This matches the CommitTransaction waiting for us in
680 : * PostgresMain().
681 : */
682 100584 : StartTransactionCommand();
683 : }
684 :
685 104362 : if ((params->options & VACOPT_VACUUM) &&
686 99694 : !(params->options & VACOPT_SKIP_DATABASE_STATS))
687 : {
688 : /*
689 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible.
690 : */
691 1658 : vac_update_datfrozenxid();
692 : }
693 :
694 104362 : }
695 :
696 : /*
697 : * Check if the current user has privileges to vacuum or analyze the relation.
698 : * If not, issue a WARNING log message and return false to let the caller
699 : * decide what to do with this relation. This routine is used to decide if a
700 : * relation can be processed for VACUUM or ANALYZE.
701 : */
702 : bool
703 153360 : vacuum_is_permitted_for_relation(Oid relid, Form_pg_class reltuple,
704 : bits32 options)
705 : {
706 : char *relname;
707 :
708 : Assert((options & (VACOPT_VACUUM | VACOPT_ANALYZE)) != 0);
709 :
710 : /*----------
711 : * A role has privileges to vacuum or analyze the relation if any of the
712 : * following are true:
713 : * - the role owns the current database and the relation is not shared
714 : * - the role has the MAINTAIN privilege on the relation
715 : *----------
716 : */
717 153360 : if ((object_ownercheck(DatabaseRelationId, MyDatabaseId, GetUserId()) &&
718 177642 : !reltuple->relisshared) ||
719 25056 : pg_class_aclcheck(relid, GetUserId(), ACL_MAINTAIN) == ACLCHECK_OK)
720 153020 : return true;
721 :
722 340 : relname = NameStr(reltuple->relname);
723 :
724 340 : if ((options & VACOPT_VACUUM) != 0)
725 : {
726 224 : ereport(WARNING,
727 : (errmsg("permission denied to vacuum \"%s\", skipping it",
728 : relname)));
729 :
730 : /*
731 : * For VACUUM ANALYZE, both logs could show up, but just generate
732 : * information for VACUUM as that would be the first one to be
733 : * processed.
734 : */
735 224 : return false;
736 : }
737 :
738 116 : if ((options & VACOPT_ANALYZE) != 0)
739 116 : ereport(WARNING,
740 : (errmsg("permission denied to analyze \"%s\", skipping it",
741 : relname)));
742 :
743 116 : return false;
744 : }
745 :
746 :
747 : /*
748 : * vacuum_open_relation
749 : *
750 : * This routine is used for attempting to open and lock a relation which
751 : * is going to be vacuumed or analyzed. If the relation cannot be opened
752 : * or locked, a log is emitted if possible.
753 : */
754 : Relation
755 128996 : vacuum_open_relation(Oid relid, RangeVar *relation, bits32 options,
756 : bool verbose, LOCKMODE lmode)
757 : {
758 : Relation rel;
759 128996 : bool rel_lock = true;
760 : int elevel;
761 :
762 : Assert((options & (VACOPT_VACUUM | VACOPT_ANALYZE)) != 0);
763 :
764 : /*
765 : * Open the relation and get the appropriate lock on it.
766 : *
767 : * There's a race condition here: the relation may have gone away since
768 : * the last time we saw it. If so, we don't need to vacuum or analyze it.
769 : *
770 : * If we've been asked not to wait for the relation lock, acquire it first
771 : * in non-blocking mode, before calling try_relation_open().
772 : */
773 128996 : if (!(options & VACOPT_SKIP_LOCKED))
774 128066 : rel = try_relation_open(relid, lmode);
775 930 : else if (ConditionalLockRelationOid(relid, lmode))
776 910 : rel = try_relation_open(relid, NoLock);
777 : else
778 : {
779 20 : rel = NULL;
780 20 : rel_lock = false;
781 : }
782 :
783 : /* if relation is opened, leave */
784 128996 : if (rel)
785 128964 : return rel;
786 :
787 : /*
788 : * Relation could not be opened, hence generate if possible a log
789 : * informing on the situation.
790 : *
791 : * If the RangeVar is not defined, we do not have enough information to
792 : * provide a meaningful log statement. Chances are that the caller has
793 : * intentionally not provided this information so that this logging is
794 : * skipped, anyway.
795 : */
796 32 : if (relation == NULL)
797 18 : return NULL;
798 :
799 : /*
800 : * Determine the log level.
801 : *
802 : * For manual VACUUM or ANALYZE, we emit a WARNING to match the log
803 : * statements in the permission checks; otherwise, only log if the caller
804 : * so requested.
805 : */
806 14 : if (!AmAutoVacuumWorkerProcess())
807 14 : elevel = WARNING;
808 0 : else if (verbose)
809 0 : elevel = LOG;
810 : else
811 0 : return NULL;
812 :
813 14 : if ((options & VACOPT_VACUUM) != 0)
814 : {
815 10 : if (!rel_lock)
816 6 : ereport(elevel,
817 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
818 : errmsg("skipping vacuum of \"%s\" --- lock not available",
819 : relation->relname)));
820 : else
821 4 : ereport(elevel,
822 : (errcode(ERRCODE_UNDEFINED_TABLE),
823 : errmsg("skipping vacuum of \"%s\" --- relation no longer exists",
824 : relation->relname)));
825 :
826 : /*
827 : * For VACUUM ANALYZE, both logs could show up, but just generate
828 : * information for VACUUM as that would be the first one to be
829 : * processed.
830 : */
831 10 : return NULL;
832 : }
833 :
834 4 : if ((options & VACOPT_ANALYZE) != 0)
835 : {
836 4 : if (!rel_lock)
837 2 : ereport(elevel,
838 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
839 : errmsg("skipping analyze of \"%s\" --- lock not available",
840 : relation->relname)));
841 : else
842 2 : ereport(elevel,
843 : (errcode(ERRCODE_UNDEFINED_TABLE),
844 : errmsg("skipping analyze of \"%s\" --- relation no longer exists",
845 : relation->relname)));
846 : }
847 :
848 4 : return NULL;
849 : }
850 :
851 :
852 : /*
853 : * Given a VacuumRelation, fill in the table OID if it wasn't specified,
854 : * and optionally add VacuumRelations for partitions or inheritance children.
855 : *
856 : * If a VacuumRelation does not have an OID supplied and is a partitioned
857 : * table, an extra entry will be added to the output for each partition.
858 : * Presently, only autovacuum supplies OIDs when calling vacuum(), and
859 : * it does not want us to expand partitioned tables.
860 : *
861 : * We take care not to modify the input data structure, but instead build
862 : * new VacuumRelation(s) to return. (But note that they will reference
863 : * unmodified parts of the input, eg column lists.) New data structures
864 : * are made in vac_context.
865 : */
866 : static List *
867 104284 : expand_vacuum_rel(VacuumRelation *vrel, MemoryContext vac_context,
868 : int options)
869 : {
870 104284 : List *vacrels = NIL;
871 : MemoryContext oldcontext;
872 :
873 : /* If caller supplied OID, there's nothing we need do here. */
874 104284 : if (OidIsValid(vrel->oid))
875 : {
876 93822 : oldcontext = MemoryContextSwitchTo(vac_context);
877 93822 : vacrels = lappend(vacrels, vrel);
878 93822 : MemoryContextSwitchTo(oldcontext);
879 : }
880 : else
881 : {
882 : /*
883 : * Process a specific relation, and possibly partitions or child
884 : * tables thereof.
885 : */
886 : Oid relid;
887 : HeapTuple tuple;
888 : Form_pg_class classForm;
889 : bool include_children;
890 : bool is_partitioned_table;
891 : int rvr_opts;
892 :
893 : /*
894 : * Since autovacuum workers supply OIDs when calling vacuum(), no
895 : * autovacuum worker should reach this code.
896 : */
897 : Assert(!AmAutoVacuumWorkerProcess());
898 :
899 : /*
900 : * We transiently take AccessShareLock to protect the syscache lookup
901 : * below, as well as find_all_inheritors's expectation that the caller
902 : * holds some lock on the starting relation.
903 : */
904 10462 : rvr_opts = (options & VACOPT_SKIP_LOCKED) ? RVR_SKIP_LOCKED : 0;
905 10462 : relid = RangeVarGetRelidExtended(vrel->relation,
906 : AccessShareLock,
907 : rvr_opts,
908 : NULL, NULL);
909 :
910 : /*
911 : * If the lock is unavailable, emit the same log statement that
912 : * vacuum_rel() and analyze_rel() would.
913 : */
914 10426 : if (!OidIsValid(relid))
915 : {
916 8 : if (options & VACOPT_VACUUM)
917 6 : ereport(WARNING,
918 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
919 : errmsg("skipping vacuum of \"%s\" --- lock not available",
920 : vrel->relation->relname)));
921 : else
922 2 : ereport(WARNING,
923 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
924 : errmsg("skipping analyze of \"%s\" --- lock not available",
925 : vrel->relation->relname)));
926 8 : return vacrels;
927 : }
928 :
929 : /*
930 : * To check whether the relation is a partitioned table and its
931 : * ownership, fetch its syscache entry.
932 : */
933 10418 : tuple = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
934 10418 : if (!HeapTupleIsValid(tuple))
935 0 : elog(ERROR, "cache lookup failed for relation %u", relid);
936 10418 : classForm = (Form_pg_class) GETSTRUCT(tuple);
937 :
938 : /*
939 : * Make a returnable VacuumRelation for this rel if the user has the
940 : * required privileges.
941 : */
942 10418 : if (vacuum_is_permitted_for_relation(relid, classForm, options))
943 : {
944 10186 : oldcontext = MemoryContextSwitchTo(vac_context);
945 10186 : vacrels = lappend(vacrels, makeVacuumRelation(vrel->relation,
946 : relid,
947 : vrel->va_cols));
948 10186 : MemoryContextSwitchTo(oldcontext);
949 : }
950 :
951 : /*
952 : * Vacuuming a partitioned table with ONLY will not do anything since
953 : * the partitioned table itself is empty. Issue a warning if the user
954 : * requests this.
955 : */
956 10418 : include_children = vrel->relation->inh;
957 10418 : is_partitioned_table = (classForm->relkind == RELKIND_PARTITIONED_TABLE);
958 10418 : if ((options & VACOPT_VACUUM) && is_partitioned_table && !include_children)
959 6 : ereport(WARNING,
960 : (errmsg("VACUUM ONLY of partitioned table \"%s\" has no effect",
961 : vrel->relation->relname)));
962 :
963 10418 : ReleaseSysCache(tuple);
964 :
965 : /*
966 : * Unless the user has specified ONLY, make relation list entries for
967 : * its partitions or inheritance child tables. Note that the list
968 : * returned by find_all_inheritors() includes the passed-in OID, so we
969 : * have to skip that. There's no point in taking locks on the
970 : * individual partitions or child tables yet, and doing so would just
971 : * add unnecessary deadlock risk. For this last reason, we do not yet
972 : * check the ownership of the partitions/tables, which get added to
973 : * the list to process. Ownership will be checked later on anyway.
974 : */
975 10418 : if (include_children)
976 : {
977 10388 : List *part_oids = find_all_inheritors(relid, NoLock, NULL);
978 : ListCell *part_lc;
979 :
980 22812 : foreach(part_lc, part_oids)
981 : {
982 12424 : Oid part_oid = lfirst_oid(part_lc);
983 :
984 12424 : if (part_oid == relid)
985 10388 : continue; /* ignore original table */
986 :
987 : /*
988 : * We omit a RangeVar since it wouldn't be appropriate to
989 : * complain about failure to open one of these relations
990 : * later.
991 : */
992 2036 : oldcontext = MemoryContextSwitchTo(vac_context);
993 2036 : vacrels = lappend(vacrels, makeVacuumRelation(NULL,
994 : part_oid,
995 : vrel->va_cols));
996 2036 : MemoryContextSwitchTo(oldcontext);
997 : }
998 : }
999 :
1000 : /*
1001 : * Release lock again. This means that by the time we actually try to
1002 : * process the table, it might be gone or renamed. In the former case
1003 : * we'll silently ignore it; in the latter case we'll process it
1004 : * anyway, but we must beware that the RangeVar doesn't necessarily
1005 : * identify it anymore. This isn't ideal, perhaps, but there's little
1006 : * practical alternative, since we're typically going to commit this
1007 : * transaction and begin a new one between now and then. Moreover,
1008 : * holding locks on multiple relations would create significant risk
1009 : * of deadlock.
1010 : */
1011 10418 : UnlockRelationOid(relid, AccessShareLock);
1012 : }
1013 :
1014 104240 : return vacrels;
1015 : }
1016 :
1017 : /*
1018 : * Construct a list of VacuumRelations for all vacuumable rels in
1019 : * the current database. The list is built in vac_context.
1020 : */
1021 : static List *
1022 192 : get_all_vacuum_rels(MemoryContext vac_context, int options)
1023 : {
1024 192 : List *vacrels = NIL;
1025 : Relation pgclass;
1026 : TableScanDesc scan;
1027 : HeapTuple tuple;
1028 :
1029 192 : pgclass = table_open(RelationRelationId, AccessShareLock);
1030 :
1031 192 : scan = table_beginscan_catalog(pgclass, 0, NULL);
1032 :
1033 82014 : while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
1034 : {
1035 81822 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
1036 : MemoryContext oldcontext;
1037 81822 : Oid relid = classForm->oid;
1038 :
1039 : /*
1040 : * We include partitioned tables here; depending on which operation is
1041 : * to be performed, caller will decide whether to process or ignore
1042 : * them.
1043 : */
1044 81822 : if (classForm->relkind != RELKIND_RELATION &&
1045 67912 : classForm->relkind != RELKIND_MATVIEW &&
1046 67906 : classForm->relkind != RELKIND_PARTITIONED_TABLE)
1047 67844 : continue;
1048 :
1049 : /* check permissions of relation */
1050 13978 : if (!vacuum_is_permitted_for_relation(relid, classForm, options))
1051 0 : continue;
1052 :
1053 : /*
1054 : * Build VacuumRelation(s) specifying the table OIDs to be processed.
1055 : * We omit a RangeVar since it wouldn't be appropriate to complain
1056 : * about failure to open one of these relations later.
1057 : */
1058 13978 : oldcontext = MemoryContextSwitchTo(vac_context);
1059 13978 : vacrels = lappend(vacrels, makeVacuumRelation(NULL,
1060 : relid,
1061 : NIL));
1062 13978 : MemoryContextSwitchTo(oldcontext);
1063 : }
1064 :
1065 192 : table_endscan(scan);
1066 192 : table_close(pgclass, AccessShareLock);
1067 :
1068 192 : return vacrels;
1069 : }
1070 :
1071 : /*
1072 : * vacuum_get_cutoffs() -- compute OldestXmin and freeze cutoff points
1073 : *
1074 : * The target relation and VACUUM parameters are our inputs.
1075 : *
1076 : * Output parameters are the cutoffs that VACUUM caller should use.
1077 : *
1078 : * Return value indicates if vacuumlazy.c caller should make its VACUUM
1079 : * operation aggressive. An aggressive VACUUM must advance relfrozenxid up to
1080 : * FreezeLimit (at a minimum), and relminmxid up to MultiXactCutoff (at a
1081 : * minimum).
1082 : */
1083 : bool
1084 114968 : vacuum_get_cutoffs(Relation rel, const VacuumParams *params,
1085 : struct VacuumCutoffs *cutoffs)
1086 : {
1087 : int freeze_min_age,
1088 : multixact_freeze_min_age,
1089 : freeze_table_age,
1090 : multixact_freeze_table_age,
1091 : effective_multixact_freeze_max_age;
1092 : TransactionId nextXID,
1093 : safeOldestXmin,
1094 : aggressiveXIDCutoff;
1095 : MultiXactId nextMXID,
1096 : safeOldestMxact,
1097 : aggressiveMXIDCutoff;
1098 :
1099 : /* Use mutable copies of freeze age parameters */
1100 114968 : freeze_min_age = params->freeze_min_age;
1101 114968 : multixact_freeze_min_age = params->multixact_freeze_min_age;
1102 114968 : freeze_table_age = params->freeze_table_age;
1103 114968 : multixact_freeze_table_age = params->multixact_freeze_table_age;
1104 :
1105 : /* Set pg_class fields in cutoffs */
1106 114968 : cutoffs->relfrozenxid = rel->rd_rel->relfrozenxid;
1107 114968 : cutoffs->relminmxid = rel->rd_rel->relminmxid;
1108 :
1109 : /*
1110 : * Acquire OldestXmin.
1111 : *
1112 : * We can always ignore processes running lazy vacuum. This is because we
1113 : * use these values only for deciding which tuples we must keep in the
1114 : * tables. Since lazy vacuum doesn't write its XID anywhere (usually no
1115 : * XID assigned), it's safe to ignore it. In theory it could be
1116 : * problematic to ignore lazy vacuums in a full vacuum, but keep in mind
1117 : * that only one vacuum process can be working on a particular table at
1118 : * any time, and that each vacuum is always an independent transaction.
1119 : */
1120 114968 : cutoffs->OldestXmin = GetOldestNonRemovableTransactionId(rel);
1121 :
1122 : Assert(TransactionIdIsNormal(cutoffs->OldestXmin));
1123 :
1124 : /* Acquire OldestMxact */
1125 114968 : cutoffs->OldestMxact = GetOldestMultiXactId();
1126 : Assert(MultiXactIdIsValid(cutoffs->OldestMxact));
1127 :
1128 : /* Acquire next XID/next MXID values used to apply age-based settings */
1129 114968 : nextXID = ReadNextTransactionId();
1130 114968 : nextMXID = ReadNextMultiXactId();
1131 :
1132 : /*
1133 : * Also compute the multixact age for which freezing is urgent. This is
1134 : * normally autovacuum_multixact_freeze_max_age, but may be less if we are
1135 : * short of multixact member space.
1136 : */
1137 114968 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1138 :
1139 : /*
1140 : * Almost ready to set freeze output parameters; check if OldestXmin or
1141 : * OldestMxact are held back to an unsafe degree before we start on that
1142 : */
1143 114968 : safeOldestXmin = nextXID - autovacuum_freeze_max_age;
1144 114968 : if (!TransactionIdIsNormal(safeOldestXmin))
1145 0 : safeOldestXmin = FirstNormalTransactionId;
1146 114968 : safeOldestMxact = nextMXID - effective_multixact_freeze_max_age;
1147 114968 : if (safeOldestMxact < FirstMultiXactId)
1148 0 : safeOldestMxact = FirstMultiXactId;
1149 114968 : if (TransactionIdPrecedes(cutoffs->OldestXmin, safeOldestXmin))
1150 57162 : ereport(WARNING,
1151 : (errmsg("cutoff for removing and freezing tuples is far in the past"),
1152 : errhint("Close open transactions soon to avoid wraparound problems.\n"
1153 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1154 114968 : if (MultiXactIdPrecedes(cutoffs->OldestMxact, safeOldestMxact))
1155 0 : ereport(WARNING,
1156 : (errmsg("cutoff for freezing multixacts is far in the past"),
1157 : errhint("Close open transactions soon to avoid wraparound problems.\n"
1158 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1159 :
1160 : /*
1161 : * Determine the minimum freeze age to use: as specified by the caller, or
1162 : * vacuum_freeze_min_age, but in any case not more than half
1163 : * autovacuum_freeze_max_age, so that autovacuums to prevent XID
1164 : * wraparound won't occur too frequently.
1165 : */
1166 114968 : if (freeze_min_age < 0)
1167 8942 : freeze_min_age = vacuum_freeze_min_age;
1168 114968 : freeze_min_age = Min(freeze_min_age, autovacuum_freeze_max_age / 2);
1169 : Assert(freeze_min_age >= 0);
1170 :
1171 : /* Compute FreezeLimit, being careful to generate a normal XID */
1172 114968 : cutoffs->FreezeLimit = nextXID - freeze_min_age;
1173 114968 : if (!TransactionIdIsNormal(cutoffs->FreezeLimit))
1174 0 : cutoffs->FreezeLimit = FirstNormalTransactionId;
1175 : /* FreezeLimit must always be <= OldestXmin */
1176 114968 : if (TransactionIdPrecedes(cutoffs->OldestXmin, cutoffs->FreezeLimit))
1177 81838 : cutoffs->FreezeLimit = cutoffs->OldestXmin;
1178 :
1179 : /*
1180 : * Determine the minimum multixact freeze age to use: as specified by
1181 : * caller, or vacuum_multixact_freeze_min_age, but in any case not more
1182 : * than half effective_multixact_freeze_max_age, so that autovacuums to
1183 : * prevent MultiXact wraparound won't occur too frequently.
1184 : */
1185 114968 : if (multixact_freeze_min_age < 0)
1186 8942 : multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1187 114968 : multixact_freeze_min_age = Min(multixact_freeze_min_age,
1188 : effective_multixact_freeze_max_age / 2);
1189 : Assert(multixact_freeze_min_age >= 0);
1190 :
1191 : /* Compute MultiXactCutoff, being careful to generate a valid value */
1192 114968 : cutoffs->MultiXactCutoff = nextMXID - multixact_freeze_min_age;
1193 114968 : if (cutoffs->MultiXactCutoff < FirstMultiXactId)
1194 0 : cutoffs->MultiXactCutoff = FirstMultiXactId;
1195 : /* MultiXactCutoff must always be <= OldestMxact */
1196 114968 : if (MultiXactIdPrecedes(cutoffs->OldestMxact, cutoffs->MultiXactCutoff))
1197 4 : cutoffs->MultiXactCutoff = cutoffs->OldestMxact;
1198 :
1199 : /*
1200 : * Finally, figure out if caller needs to do an aggressive VACUUM or not.
1201 : *
1202 : * Determine the table freeze age to use: as specified by the caller, or
1203 : * the value of the vacuum_freeze_table_age GUC, but in any case not more
1204 : * than autovacuum_freeze_max_age * 0.95, so that if you have e.g nightly
1205 : * VACUUM schedule, the nightly VACUUM gets a chance to freeze XIDs before
1206 : * anti-wraparound autovacuum is launched.
1207 : */
1208 114968 : if (freeze_table_age < 0)
1209 8942 : freeze_table_age = vacuum_freeze_table_age;
1210 114968 : freeze_table_age = Min(freeze_table_age, autovacuum_freeze_max_age * 0.95);
1211 : Assert(freeze_table_age >= 0);
1212 114968 : aggressiveXIDCutoff = nextXID - freeze_table_age;
1213 114968 : if (!TransactionIdIsNormal(aggressiveXIDCutoff))
1214 0 : aggressiveXIDCutoff = FirstNormalTransactionId;
1215 114968 : if (TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid,
1216 : aggressiveXIDCutoff))
1217 106118 : return true;
1218 :
1219 : /*
1220 : * Similar to the above, determine the table freeze age to use for
1221 : * multixacts: as specified by the caller, or the value of the
1222 : * vacuum_multixact_freeze_table_age GUC, but in any case not more than
1223 : * effective_multixact_freeze_max_age * 0.95, so that if you have e.g.
1224 : * nightly VACUUM schedule, the nightly VACUUM gets a chance to freeze
1225 : * multixacts before anti-wraparound autovacuum is launched.
1226 : */
1227 8850 : if (multixact_freeze_table_age < 0)
1228 8722 : multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1229 8850 : multixact_freeze_table_age =
1230 8850 : Min(multixact_freeze_table_age,
1231 : effective_multixact_freeze_max_age * 0.95);
1232 : Assert(multixact_freeze_table_age >= 0);
1233 8850 : aggressiveMXIDCutoff = nextMXID - multixact_freeze_table_age;
1234 8850 : if (aggressiveMXIDCutoff < FirstMultiXactId)
1235 0 : aggressiveMXIDCutoff = FirstMultiXactId;
1236 8850 : if (MultiXactIdPrecedesOrEquals(cutoffs->relminmxid,
1237 : aggressiveMXIDCutoff))
1238 0 : return true;
1239 :
1240 : /* Non-aggressive VACUUM */
1241 8850 : return false;
1242 : }
1243 :
1244 : /*
1245 : * vacuum_xid_failsafe_check() -- Used by VACUUM's wraparound failsafe
1246 : * mechanism to determine if its table's relfrozenxid and relminmxid are now
1247 : * dangerously far in the past.
1248 : *
1249 : * When we return true, VACUUM caller triggers the failsafe.
1250 : */
1251 : bool
1252 117504 : vacuum_xid_failsafe_check(const struct VacuumCutoffs *cutoffs)
1253 : {
1254 117504 : TransactionId relfrozenxid = cutoffs->relfrozenxid;
1255 117504 : MultiXactId relminmxid = cutoffs->relminmxid;
1256 : TransactionId xid_skip_limit;
1257 : MultiXactId multi_skip_limit;
1258 : int skip_index_vacuum;
1259 :
1260 : Assert(TransactionIdIsNormal(relfrozenxid));
1261 : Assert(MultiXactIdIsValid(relminmxid));
1262 :
1263 : /*
1264 : * Determine the index skipping age to use. In any case no less than
1265 : * autovacuum_freeze_max_age * 1.05.
1266 : */
1267 117504 : skip_index_vacuum = Max(vacuum_failsafe_age, autovacuum_freeze_max_age * 1.05);
1268 :
1269 117504 : xid_skip_limit = ReadNextTransactionId() - skip_index_vacuum;
1270 117504 : if (!TransactionIdIsNormal(xid_skip_limit))
1271 0 : xid_skip_limit = FirstNormalTransactionId;
1272 :
1273 117504 : if (TransactionIdPrecedes(relfrozenxid, xid_skip_limit))
1274 : {
1275 : /* The table's relfrozenxid is too old */
1276 19278 : return true;
1277 : }
1278 :
1279 : /*
1280 : * Similar to above, determine the index skipping age to use for
1281 : * multixact. In any case no less than autovacuum_multixact_freeze_max_age *
1282 : * 1.05.
1283 : */
1284 98226 : skip_index_vacuum = Max(vacuum_multixact_failsafe_age,
1285 : autovacuum_multixact_freeze_max_age * 1.05);
1286 :
1287 98226 : multi_skip_limit = ReadNextMultiXactId() - skip_index_vacuum;
1288 98226 : if (multi_skip_limit < FirstMultiXactId)
1289 0 : multi_skip_limit = FirstMultiXactId;
1290 :
1291 98226 : if (MultiXactIdPrecedes(relminmxid, multi_skip_limit))
1292 : {
1293 : /* The table's relminmxid is too old */
1294 0 : return true;
1295 : }
1296 :
1297 98226 : return false;
1298 : }
1299 :
1300 : /*
1301 : * vac_estimate_reltuples() -- estimate the new value for pg_class.reltuples
1302 : *
1303 : * If we scanned the whole relation then we should just use the count of
1304 : * live tuples seen; but if we did not, we should not blindly extrapolate
1305 : * from that number, since VACUUM may have scanned a quite nonrandom
1306 : * subset of the table. When we have only partial information, we take
1307 : * the old value of pg_class.reltuples/pg_class.relpages as a measurement
1308 : * of the tuple density in the unscanned pages.
1309 : *
1310 : * Note: scanned_tuples should count only *live* tuples, since
1311 : * pg_class.reltuples is defined that way.
1312 : */
1313 : double
1314 114426 : vac_estimate_reltuples(Relation relation,
1315 : BlockNumber total_pages,
1316 : BlockNumber scanned_pages,
1317 : double scanned_tuples)
1318 : {
1319 114426 : BlockNumber old_rel_pages = relation->rd_rel->relpages;
1320 114426 : double old_rel_tuples = relation->rd_rel->reltuples;
1321 : double old_density;
1322 : double unscanned_pages;
1323 : double total_tuples;
1324 :
1325 : /* If we did scan the whole table, just use the count as-is */
1326 114426 : if (scanned_pages >= total_pages)
1327 110488 : return scanned_tuples;
1328 :
1329 : /*
1330 : * When successive VACUUM commands scan the same few pages again and
1331 : * again, without anything from the table really changing, there is a risk
1332 : * that our beliefs about tuple density will gradually become distorted.
1333 : * This might be caused by vacuumlazy.c implementation details, such as
1334 : * its tendency to always scan the last heap page. Handle that here.
1335 : *
1336 : * If the relation is _exactly_ the same size according to the existing
1337 : * pg_class entry, and only a few of its pages (less than 2%) were
1338 : * scanned, keep the existing value of reltuples. Also keep the existing
1339 : * value when only a subset of rel's pages <= a single page were scanned.
1340 : *
1341 : * (Note: we might be returning -1 here.)
1342 : */
1343 3938 : if (old_rel_pages == total_pages &&
1344 3908 : scanned_pages < (double) total_pages * 0.02)
1345 2732 : return old_rel_tuples;
1346 1206 : if (scanned_pages <= 1)
1347 948 : return old_rel_tuples;
1348 :
1349 : /*
1350 : * If old density is unknown, we can't do much except scale up
1351 : * scanned_tuples to match total_pages.
1352 : */
1353 258 : if (old_rel_tuples < 0 || old_rel_pages == 0)
1354 2 : return floor((scanned_tuples / scanned_pages) * total_pages + 0.5);
1355 :
1356 : /*
1357 : * Okay, we've covered the corner cases. The normal calculation is to
1358 : * convert the old measurement to a density (tuples per page), then
1359 : * estimate the number of tuples in the unscanned pages using that figure,
1360 : * and finally add on the number of tuples in the scanned pages.
1361 : */
1362 256 : old_density = old_rel_tuples / old_rel_pages;
1363 256 : unscanned_pages = (double) total_pages - (double) scanned_pages;
1364 256 : total_tuples = old_density * unscanned_pages + scanned_tuples;
1365 256 : return floor(total_tuples + 0.5);
1366 : }
1367 :
1368 :
1369 : /*
1370 : * vac_update_relstats() -- update statistics for one relation
1371 : *
1372 : * Update the whole-relation statistics that are kept in its pg_class
1373 : * row. There are additional stats that will be updated if we are
1374 : * doing ANALYZE, but we always update these stats. This routine works
1375 : * for both index and heap relation entries in pg_class.
1376 : *
1377 : * We violate transaction semantics here by overwriting the rel's
1378 : * existing pg_class tuple with the new values. This is reasonably
1379 : * safe as long as we're sure that the new values are correct whether or
1380 : * not this transaction commits. The reason for doing this is that if
1381 : * we updated these tuples in the usual way, vacuuming pg_class itself
1382 : * wouldn't work very well --- by the time we got done with a vacuum
1383 : * cycle, most of the tuples in pg_class would've been obsoleted. Of
1384 : * course, this only works for fixed-size not-null columns, but these are.
1385 : *
1386 : * Another reason for doing it this way is that when we are in a lazy
1387 : * VACUUM and have PROC_IN_VACUUM set, we mustn't do any regular updates.
1388 : * Somebody vacuuming pg_class might think they could delete a tuple
1389 : * marked with xmin = our xid.
1390 : *
1391 : * In addition to fundamentally nontransactional statistics such as
1392 : * relpages and relallvisible, we try to maintain certain lazily-updated
1393 : * DDL flags such as relhasindex, by clearing them if no longer correct.
1394 : * It's safe to do this in VACUUM, which can't run in parallel with
1395 : * CREATE INDEX/RULE/TRIGGER and can't be part of a transaction block.
1396 : * However, it's *not* safe to do it in an ANALYZE that's within an
1397 : * outer transaction, because for example the current transaction might
1398 : * have dropped the last index; then we'd think relhasindex should be
1399 : * cleared, but if the transaction later rolls back this would be wrong.
1400 : * So we refrain from updating the DDL flags if we're inside an outer
1401 : * transaction. This is OK since postponing the flag maintenance is
1402 : * always allowable.
1403 : *
1404 : * Note: num_tuples should count only *live* tuples, since
1405 : * pg_class.reltuples is defined that way.
1406 : *
1407 : * This routine is shared by VACUUM and ANALYZE.
1408 : */
1409 : void
1410 148820 : vac_update_relstats(Relation relation,
1411 : BlockNumber num_pages, double num_tuples,
1412 : BlockNumber num_all_visible_pages,
1413 : bool hasindex, TransactionId frozenxid,
1414 : MultiXactId minmulti,
1415 : bool *frozenxid_updated, bool *minmulti_updated,
1416 : bool in_outer_xact)
1417 : {
1418 148820 : Oid relid = RelationGetRelid(relation);
1419 : Relation rd;
1420 : ScanKeyData key[1];
1421 : HeapTuple ctup;
1422 : void *inplace_state;
1423 : Form_pg_class pgcform;
1424 : bool dirty,
1425 : futurexid,
1426 : futuremxid;
1427 : TransactionId oldfrozenxid;
1428 : MultiXactId oldminmulti;
1429 :
1430 148820 : rd = table_open(RelationRelationId, RowExclusiveLock);
1431 :
1432 : /* Fetch a copy of the tuple to scribble on */
1433 148820 : ScanKeyInit(&key[0],
1434 : Anum_pg_class_oid,
1435 : BTEqualStrategyNumber, F_OIDEQ,
1436 : ObjectIdGetDatum(relid));
1437 148820 : systable_inplace_update_begin(rd, ClassOidIndexId, true,
1438 : NULL, 1, key, &ctup, &inplace_state);
1439 148820 : if (!HeapTupleIsValid(ctup))
1440 0 : elog(ERROR, "pg_class entry for relid %u vanished during vacuuming",
1441 : relid);
1442 148820 : pgcform = (Form_pg_class) GETSTRUCT(ctup);
1443 :
1444 : /* Apply statistical updates, if any, to copied tuple */
1445 :
1446 148820 : dirty = false;
1447 148820 : if (pgcform->relpages != (int32) num_pages)
1448 : {
1449 8486 : pgcform->relpages = (int32) num_pages;
1450 8486 : dirty = true;
1451 : }
1452 148820 : if (pgcform->reltuples != (float4) num_tuples)
1453 : {
1454 18040 : pgcform->reltuples = (float4) num_tuples;
1455 18040 : dirty = true;
1456 : }
1457 148820 : if (pgcform->relallvisible != (int32) num_all_visible_pages)
1458 : {
1459 5138 : pgcform->relallvisible = (int32) num_all_visible_pages;
1460 5138 : dirty = true;
1461 : }
1462 :
1463 : /* Apply DDL updates, but not inside an outer transaction (see above) */
1464 :
1465 148820 : if (!in_outer_xact)
1466 : {
1467 : /*
1468 : * If we didn't find any indexes, reset relhasindex.
1469 : */
1470 148506 : if (pgcform->relhasindex && !hasindex)
1471 : {
1472 20 : pgcform->relhasindex = false;
1473 20 : dirty = true;
1474 : }
1475 :
1476 : /* We also clear relhasrules and relhastriggers if needed */
1477 148506 : if (pgcform->relhasrules && relation->rd_rules == NULL)
1478 : {
1479 0 : pgcform->relhasrules = false;
1480 0 : dirty = true;
1481 : }
1482 148506 : if (pgcform->relhastriggers && relation->trigdesc == NULL)
1483 : {
1484 6 : pgcform->relhastriggers = false;
1485 6 : dirty = true;
1486 : }
1487 : }
1488 :
1489 : /*
1490 : * Update relfrozenxid, unless caller passed InvalidTransactionId
1491 : * indicating it has no new data.
1492 : *
1493 : * Ordinarily, we don't let relfrozenxid go backwards. However, if the
1494 : * stored relfrozenxid is "in the future" then it seems best to assume
1495 : * it's corrupt, and overwrite with the oldest remaining XID in the table.
1496 : * This should match vac_update_datfrozenxid() concerning what we consider
1497 : * to be "in the future".
1498 : */
1499 148820 : oldfrozenxid = pgcform->relfrozenxid;
1500 148820 : futurexid = false;
1501 148820 : if (frozenxid_updated)
1502 114422 : *frozenxid_updated = false;
1503 148820 : if (TransactionIdIsNormal(frozenxid) && oldfrozenxid != frozenxid)
1504 : {
1505 55410 : bool update = false;
1506 :
1507 55410 : if (TransactionIdPrecedes(oldfrozenxid, frozenxid))
1508 55342 : update = true;
1509 68 : else if (TransactionIdPrecedes(ReadNextTransactionId(), oldfrozenxid))
1510 0 : futurexid = update = true;
1511 :
1512 55410 : if (update)
1513 : {
1514 55342 : pgcform->relfrozenxid = frozenxid;
1515 55342 : dirty = true;
1516 55342 : if (frozenxid_updated)
1517 55342 : *frozenxid_updated = true;
1518 : }
1519 : }
1520 :
1521 : /* Similarly for relminmxid */
1522 148820 : oldminmulti = pgcform->relminmxid;
1523 148820 : futuremxid = false;
1524 148820 : if (minmulti_updated)
1525 114422 : *minmulti_updated = false;
1526 148820 : if (MultiXactIdIsValid(minmulti) && oldminmulti != minmulti)
1527 : {
1528 358 : bool update = false;
1529 :
1530 358 : if (MultiXactIdPrecedes(oldminmulti, minmulti))
1531 358 : update = true;
1532 0 : else if (MultiXactIdPrecedes(ReadNextMultiXactId(), oldminmulti))
1533 0 : futuremxid = update = true;
1534 :
1535 358 : if (update)
1536 : {
1537 358 : pgcform->relminmxid = minmulti;
1538 358 : dirty = true;
1539 358 : if (minmulti_updated)
1540 358 : *minmulti_updated = true;
1541 : }
1542 : }
1543 :
1544 : /* If anything changed, write out the tuple. */
1545 148820 : if (dirty)
1546 67974 : systable_inplace_update_finish(inplace_state, ctup);
1547 : else
1548 80846 : systable_inplace_update_cancel(inplace_state);
1549 :
1550 148820 : table_close(rd, RowExclusiveLock);
1551 :
1552 148820 : if (futurexid)
1553 0 : ereport(WARNING,
1554 : (errcode(ERRCODE_DATA_CORRUPTED),
1555 : errmsg_internal("overwrote invalid relfrozenxid value %u with new value %u for table \"%s\"",
1556 : oldfrozenxid, frozenxid,
1557 : RelationGetRelationName(relation))));
1558 148820 : if (futuremxid)
1559 0 : ereport(WARNING,
1560 : (errcode(ERRCODE_DATA_CORRUPTED),
1561 : errmsg_internal("overwrote invalid relminmxid value %u with new value %u for table \"%s\"",
1562 : oldminmulti, minmulti,
1563 : RelationGetRelationName(relation))));
1564 148820 : }
1565 :
1566 :
1567 : /*
1568 : * vac_update_datfrozenxid() -- update pg_database.datfrozenxid for our DB
1569 : *
1570 : * Update pg_database's datfrozenxid entry for our database to be the
1571 : * minimum of the pg_class.relfrozenxid values.
1572 : *
1573 : * Similarly, update our datminmxid to be the minimum of the
1574 : * pg_class.relminmxid values.
1575 : *
1576 : * If we are able to advance either pg_database value, also try to
1577 : * truncate pg_xact and pg_multixact.
1578 : *
1579 : * We violate transaction semantics here by overwriting the database's
1580 : * existing pg_database tuple with the new values. This is reasonably
1581 : * safe since the new values are correct whether or not this transaction
1582 : * commits. As with vac_update_relstats, this avoids leaving dead tuples
1583 : * behind after a VACUUM.
1584 : */
1585 : void
1586 3922 : vac_update_datfrozenxid(void)
1587 : {
1588 : HeapTuple tuple;
1589 : Form_pg_database dbform;
1590 : Relation relation;
1591 : SysScanDesc scan;
1592 : HeapTuple classTup;
1593 : TransactionId newFrozenXid;
1594 : MultiXactId newMinMulti;
1595 : TransactionId lastSaneFrozenXid;
1596 : MultiXactId lastSaneMinMulti;
1597 3922 : bool bogus = false;
1598 3922 : bool dirty = false;
1599 : ScanKeyData key[1];
1600 : void *inplace_state;
1601 :
1602 : /*
1603 : * Restrict this task to one backend per database. This avoids race
1604 : * conditions that would move datfrozenxid or datminmxid backward. It
1605 : * avoids calling vac_truncate_clog() with a datfrozenxid preceding a
1606 : * datfrozenxid passed to an earlier vac_truncate_clog() call.
1607 : */
1608 3922 : LockDatabaseFrozenIds(ExclusiveLock);
1609 :
1610 : /*
1611 : * Initialize the "min" calculation with
1612 : * GetOldestNonRemovableTransactionId(), which is a reasonable
1613 : * approximation to the minimum relfrozenxid for not-yet-committed
1614 : * pg_class entries for new tables; see AddNewRelationTuple(). So we
1615 : * cannot produce a wrong minimum by starting with this.
1616 : */
1617 3922 : newFrozenXid = GetOldestNonRemovableTransactionId(NULL);
1618 :
1619 : /*
1620 : * Similarly, initialize the MultiXact "min" with the value that would be
1621 : * used on pg_class for new tables. See AddNewRelationTuple().
1622 : */
1623 3922 : newMinMulti = GetOldestMultiXactId();
1624 :
1625 : /*
1626 : * Identify the latest relfrozenxid and relminmxid values that we could
1627 : * validly see during the scan. These are conservative values, but it's
1628 : * not really worth trying to be more exact.
1629 : */
1630 3922 : lastSaneFrozenXid = ReadNextTransactionId();
1631 3922 : lastSaneMinMulti = ReadNextMultiXactId();
1632 :
1633 : /*
1634 : * We must seqscan pg_class to find the minimum Xid, because there is no
1635 : * index that can help us here.
1636 : *
1637 : * See vac_truncate_clog() for the race condition to prevent.
1638 : */
1639 3922 : relation = table_open(RelationRelationId, AccessShareLock);
1640 :
1641 3922 : scan = systable_beginscan(relation, InvalidOid, false,
1642 : NULL, 0, NULL);
1643 :
1644 2014034 : while ((classTup = systable_getnext(scan)) != NULL)
1645 : {
1646 2010112 : volatile FormData_pg_class *classForm = (Form_pg_class) GETSTRUCT(classTup);
1647 2010112 : TransactionId relfrozenxid = classForm->relfrozenxid;
1648 2010112 : TransactionId relminmxid = classForm->relminmxid;
1649 :
1650 : /*
1651 : * Only consider relations able to hold unfrozen XIDs (anything else
1652 : * should have InvalidTransactionId in relfrozenxid anyway).
1653 : */
1654 2010112 : if (classForm->relkind != RELKIND_RELATION &&
1655 1595152 : classForm->relkind != RELKIND_MATVIEW &&
1656 1593054 : classForm->relkind != RELKIND_TOASTVALUE)
1657 : {
1658 : Assert(!TransactionIdIsValid(relfrozenxid));
1659 : Assert(!MultiXactIdIsValid(relminmxid));
1660 1372120 : continue;
1661 : }
1662 :
1663 : /*
1664 : * Some table AMs might not need per-relation xid / multixid horizons.
1665 : * It therefore seems reasonable to allow relfrozenxid and relminmxid
1666 : * to not be set (i.e. set to their respective Invalid*Id)
1667 : * independently. Thus validate and compute horizon for each only if
1668 : * set.
1669 : *
1670 : * If things are working properly, no relation should have a
1671 : * relfrozenxid or relminmxid that is "in the future". However, such
1672 : * cases have been known to arise due to bugs in pg_upgrade. If we
1673 : * see any entries that are "in the future", chicken out and don't do
1674 : * anything. This ensures we won't truncate clog & multixact SLRUs
1675 : * before those relations have been scanned and cleaned up.
1676 : */
1677 :
1678 637992 : if (TransactionIdIsValid(relfrozenxid))
1679 : {
1680 : Assert(TransactionIdIsNormal(relfrozenxid));
1681 :
1682 : /* check for values in the future */
1683 637992 : if (TransactionIdPrecedes(lastSaneFrozenXid, relfrozenxid))
1684 : {
1685 0 : bogus = true;
1686 0 : break;
1687 : }
1688 :
1689 : /* determine new horizon */
1690 637992 : if (TransactionIdPrecedes(relfrozenxid, newFrozenXid))
1691 3958 : newFrozenXid = relfrozenxid;
1692 : }
1693 :
1694 637992 : if (MultiXactIdIsValid(relminmxid))
1695 : {
1696 : /* check for values in the future */
1697 637992 : if (MultiXactIdPrecedes(lastSaneMinMulti, relminmxid))
1698 : {
1699 0 : bogus = true;
1700 0 : break;
1701 : }
1702 :
1703 : /* determine new horizon */
1704 637992 : if (MultiXactIdPrecedes(relminmxid, newMinMulti))
1705 200 : newMinMulti = relminmxid;
1706 : }
1707 : }
1708 :
1709 : /* we're done with pg_class */
1710 3922 : systable_endscan(scan);
1711 3922 : table_close(relation, AccessShareLock);
1712 :
1713 : /* chicken out if bogus data found */
1714 3922 : if (bogus)
1715 0 : return;
1716 :
1717 : Assert(TransactionIdIsNormal(newFrozenXid));
1718 : Assert(MultiXactIdIsValid(newMinMulti));
1719 :
1720 : /* Now fetch the pg_database tuple we need to update. */
1721 3922 : relation = table_open(DatabaseRelationId, RowExclusiveLock);
1722 :
1723 : /*
1724 : * Fetch a copy of the tuple to scribble on. We could check the syscache
1725 : * tuple first. If that concluded !dirty, we'd avoid waiting on
1726 : * concurrent heap_update() and would avoid exclusive-locking the buffer.
1727 : * For now, don't optimize that.
1728 : */
1729 3922 : ScanKeyInit(&key[0],
1730 : Anum_pg_database_oid,
1731 : BTEqualStrategyNumber, F_OIDEQ,
1732 : ObjectIdGetDatum(MyDatabaseId));
1733 :
1734 3922 : systable_inplace_update_begin(relation, DatabaseOidIndexId, true,
1735 : NULL, 1, key, &tuple, &inplace_state);
1736 :
1737 3922 : if (!HeapTupleIsValid(tuple))
1738 0 : elog(ERROR, "could not find tuple for database %u", MyDatabaseId);
1739 :
1740 3922 : dbform = (Form_pg_database) GETSTRUCT(tuple);
1741 :
1742 : /*
1743 : * As in vac_update_relstats(), we ordinarily don't want to let
1744 : * datfrozenxid go backward; but if it's "in the future" then it must be
1745 : * corrupt and it seems best to overwrite it.
1746 : */
1747 4440 : if (dbform->datfrozenxid != newFrozenXid &&
1748 518 : (TransactionIdPrecedes(dbform->datfrozenxid, newFrozenXid) ||
1749 0 : TransactionIdPrecedes(lastSaneFrozenXid, dbform->datfrozenxid)))
1750 : {
1751 518 : dbform->datfrozenxid = newFrozenXid;
1752 518 : dirty = true;
1753 : }
1754 : else
1755 3404 : newFrozenXid = dbform->datfrozenxid;
1756 :
1757 : /* Ditto for datminmxid */
1758 3924 : if (dbform->datminmxid != newMinMulti &&
1759 2 : (MultiXactIdPrecedes(dbform->datminmxid, newMinMulti) ||
1760 0 : MultiXactIdPrecedes(lastSaneMinMulti, dbform->datminmxid)))
1761 : {
1762 2 : dbform->datminmxid = newMinMulti;
1763 2 : dirty = true;
1764 : }
1765 : else
1766 3920 : newMinMulti = dbform->datminmxid;
1767 :
1768 3922 : if (dirty)
1769 518 : systable_inplace_update_finish(inplace_state, tuple);
1770 : else
1771 3404 : systable_inplace_update_cancel(inplace_state);
1772 :
1773 3922 : heap_freetuple(tuple);
1774 3922 : table_close(relation, RowExclusiveLock);
1775 :
1776 : /*
1777 : * If we were able to advance datfrozenxid or datminmxid, see if we can
1778 : * truncate pg_xact and/or pg_multixact. Also do it if the shared
1779 : * XID-wrap-limit info is stale, since this action will update that too.
1780 : */
1781 3922 : if (dirty || ForceTransactionIdLimitUpdate())
1782 1062 : vac_truncate_clog(newFrozenXid, newMinMulti,
1783 : lastSaneFrozenXid, lastSaneMinMulti);
1784 : }
1785 :
1786 :
1787 : /*
1788 : * vac_truncate_clog() -- attempt to truncate the commit log
1789 : *
1790 : * Scan pg_database to determine the system-wide oldest datfrozenxid,
1791 : * and use it to truncate the transaction commit log (pg_xact).
1792 : * Also update the XID wrap limit info maintained by varsup.c.
1793 : * Likewise for datminmxid.
1794 : *
1795 : * The passed frozenXID and minMulti are the updated values for my own
1796 : * pg_database entry. They're used to initialize the "min" calculations.
1797 : * The caller also passes the "last sane" XID and MXID, since it has
1798 : * those at hand already.
1799 : *
1800 : * This routine is only invoked when we've managed to change our
1801 : * DB's datfrozenxid/datminmxid values, or we found that the shared
1802 : * XID-wrap-limit info is stale.
1803 : */
1804 : static void
1805 1062 : vac_truncate_clog(TransactionId frozenXID,
1806 : MultiXactId minMulti,
1807 : TransactionId lastSaneFrozenXid,
1808 : MultiXactId lastSaneMinMulti)
1809 : {
1810 1062 : TransactionId nextXID = ReadNextTransactionId();
1811 : Relation relation;
1812 : TableScanDesc scan;
1813 : HeapTuple tuple;
1814 : Oid oldestxid_datoid;
1815 : Oid minmulti_datoid;
1816 1062 : bool bogus = false;
1817 1062 : bool frozenAlreadyWrapped = false;
1818 :
1819 : /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */
1820 1062 : LWLockAcquire(WrapLimitsVacuumLock, LW_EXCLUSIVE);
1821 :
1822 : /* init oldest datoids to sync with my frozenXID/minMulti values */
1823 1062 : oldestxid_datoid = MyDatabaseId;
1824 1062 : minmulti_datoid = MyDatabaseId;
1825 :
1826 : /*
1827 : * Scan pg_database to compute the minimum datfrozenxid/datminmxid
1828 : *
1829 : * Since vac_update_datfrozenxid updates datfrozenxid/datminmxid in-place,
1830 : * the values could change while we look at them. Fetch each one just
1831 : * once to ensure sane behavior of the comparison logic. (Here, as in
1832 : * many other places, we assume that fetching or updating an XID in shared
1833 : * storage is atomic.)
1834 : *
1835 : * Note: we need not worry about a race condition with new entries being
1836 : * inserted by CREATE DATABASE. Any such entry will have a copy of some
1837 : * existing DB's datfrozenxid, and that source DB cannot be ours because
1838 : * of the interlock against copying a DB containing an active backend.
1839 : * Hence the new entry will not reduce the minimum. Also, if two VACUUMs
1840 : * concurrently modify the datfrozenxid's of different databases, the
1841 : * worst possible outcome is that pg_xact is not truncated as aggressively
1842 : * as it could be.
1843 : */
1844 1062 : relation = table_open(DatabaseRelationId, AccessShareLock);
1845 :
1846 1062 : scan = table_beginscan_catalog(relation, 0, NULL);
1847 :
1848 4094 : while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
1849 : {
1850 3032 : volatile FormData_pg_database *dbform = (Form_pg_database) GETSTRUCT(tuple);
1851 3032 : TransactionId datfrozenxid = dbform->datfrozenxid;
1852 3032 : TransactionId datminmxid = dbform->datminmxid;
1853 :
1854 : Assert(TransactionIdIsNormal(datfrozenxid));
1855 : Assert(MultiXactIdIsValid(datminmxid));
1856 :
1857 : /*
1858 : * If database is in the process of getting dropped, or has been
1859 : * interrupted while doing so, no connections to it are possible
1860 : * anymore. Therefore we don't need to take it into account here.
1861 : * Which is good, because it can't be processed by autovacuum either.
1862 : */
1863 3032 : if (database_is_invalid_form((Form_pg_database) dbform))
1864 : {
1865 2 : elog(DEBUG2,
1866 : "skipping invalid database \"%s\" while computing relfrozenxid",
1867 : NameStr(dbform->datname));
1868 2 : continue;
1869 : }
1870 :
1871 : /*
1872 : * If things are working properly, no database should have a
1873 : * datfrozenxid or datminmxid that is "in the future". However, such
1874 : * cases have been known to arise due to bugs in pg_upgrade. If we
1875 : * see any entries that are "in the future", chicken out and don't do
1876 : * anything. This ensures we won't truncate clog before those
1877 : * databases have been scanned and cleaned up. (We will issue the
1878 : * "already wrapped" warning if appropriate, though.)
1879 : */
1880 6060 : if (TransactionIdPrecedes(lastSaneFrozenXid, datfrozenxid) ||
1881 3030 : MultiXactIdPrecedes(lastSaneMinMulti, datminmxid))
1882 0 : bogus = true;
1883 :
1884 3030 : if (TransactionIdPrecedes(nextXID, datfrozenxid))
1885 0 : frozenAlreadyWrapped = true;
1886 3030 : else if (TransactionIdPrecedes(datfrozenxid, frozenXID))
1887 : {
1888 510 : frozenXID = datfrozenxid;
1889 510 : oldestxid_datoid = dbform->oid;
1890 : }
1891 :
1892 3030 : if (MultiXactIdPrecedes(datminmxid, minMulti))
1893 : {
1894 4 : minMulti = datminmxid;
1895 4 : minmulti_datoid = dbform->oid;
1896 : }
1897 : }
1898 :
1899 1062 : table_endscan(scan);
1900 :
1901 1062 : table_close(relation, AccessShareLock);
1902 :
1903 : /*
1904 : * Do not truncate CLOG if we seem to have suffered wraparound already;
1905 : * the computed minimum XID might be bogus. This case should now be
1906 : * impossible due to the defenses in GetNewTransactionId, but we keep the
1907 : * test anyway.
1908 : */
1909 1062 : if (frozenAlreadyWrapped)
1910 : {
1911 0 : ereport(WARNING,
1912 : (errmsg("some databases have not been vacuumed in over 2 billion transactions"),
1913 : errdetail("You might have already suffered transaction-wraparound data loss.")));
1914 0 : LWLockRelease(WrapLimitsVacuumLock);
1915 0 : return;
1916 : }
1917 :
1918 : /* chicken out if data is bogus in any other way */
1919 1062 : if (bogus)
1920 : {
1921 0 : LWLockRelease(WrapLimitsVacuumLock);
1922 0 : return;
1923 : }
1924 :
1925 : /*
1926 : * Advance the oldest value for commit timestamps before truncating, so
1927 : * that if a user requests a timestamp for a transaction we're truncating
1928 : * away right after this point, they get NULL instead of an ugly "file not
1929 : * found" error from slru.c. This doesn't matter for xact/multixact
1930 : * because they are not subject to arbitrary lookups from users.
1931 : */
1932 1062 : AdvanceOldestCommitTsXid(frozenXID);
1933 :
1934 : /*
1935 : * Truncate CLOG, multixact and CommitTs to the oldest computed value.
1936 : */
1937 1062 : TruncateCLOG(frozenXID, oldestxid_datoid);
1938 1062 : TruncateCommitTs(frozenXID);
1939 1062 : TruncateMultiXact(minMulti, minmulti_datoid);
1940 :
1941 : /*
1942 : * Update the wrap limit for GetNewTransactionId and creation of new
1943 : * MultiXactIds. Note: these functions will also signal the postmaster
1944 : * for an(other) autovac cycle if needed. XXX should we avoid possibly
1945 : * signaling twice?
1946 : */
1947 1062 : SetTransactionIdLimit(frozenXID, oldestxid_datoid);
1948 1062 : SetMultiXactIdLimit(minMulti, minmulti_datoid, false);
1949 :
1950 1062 : LWLockRelease(WrapLimitsVacuumLock);
1951 : }
1952 :
1953 :
1954 : /*
1955 : * vacuum_rel() -- vacuum one heap relation
1956 : *
1957 : * relid identifies the relation to vacuum. If relation is supplied,
1958 : * use the name therein for reporting any failure to open/lock the rel;
1959 : * do not use it once we've successfully opened the rel, since it might
1960 : * be stale.
1961 : *
1962 : * Returns true if it's okay to proceed with a requested ANALYZE
1963 : * operation on this table.
1964 : *
1965 : * Doing one heap at a time incurs extra overhead, since we need to
1966 : * check that the heap exists again just before we vacuum it. The
1967 : * reason that we do this is so that vacuuming can be spread across
1968 : * many small transactions. Otherwise, two-phase locking would require
1969 : * us to lock the entire database during one pass of the vacuum cleaner.
1970 : *
1971 : * At entry and exit, we are not inside a transaction.
1972 : */
1973 : static bool
1974 115212 : vacuum_rel(Oid relid, RangeVar *relation, VacuumParams *params,
1975 : BufferAccessStrategy bstrategy)
1976 : {
1977 : LOCKMODE lmode;
1978 : Relation rel;
1979 : LockRelId lockrelid;
1980 : Oid priv_relid;
1981 : Oid toast_relid;
1982 : Oid save_userid;
1983 : int save_sec_context;
1984 : int save_nestlevel;
1985 :
1986 : Assert(params != NULL);
1987 :
1988 : /* Begin a transaction for vacuuming this relation */
1989 115212 : StartTransactionCommand();
1990 :
1991 115212 : if (!(params->options & VACOPT_FULL))
1992 : {
1993 : /*
1994 : * In lazy vacuum, we can set the PROC_IN_VACUUM flag, which lets
1995 : * other concurrent VACUUMs know that they can ignore this one while
1996 : * determining their OldestXmin. (The reason we don't set it during a
1997 : * full VACUUM is exactly that we may have to run user-defined
1998 : * functions for functional indexes, and we want to make sure that if
1999 : * they use the snapshot set above, any tuples it requires can't get
2000 : * removed from other tables. An index function that depends on the
2001 : * contents of other tables is arguably broken, but we won't break it
2002 : * here by violating transaction semantics.)
2003 : *
2004 : * We also set the VACUUM_FOR_WRAPAROUND flag, which is passed down by
2005 : * autovacuum; it's used to avoid canceling a vacuum that was invoked
2006 : * in an emergency.
2007 : *
2008 : * Note: these flags remain set until CommitTransaction or
2009 : * AbortTransaction. We don't want to clear them until we reset
2010 : * MyProc->xid/xmin, otherwise GetOldestNonRemovableTransactionId()
2011 : * might appear to go backwards, which is probably Not Good. (We also
2012 : * set PROC_IN_VACUUM *before* taking our own snapshot, so that our
2013 : * xmin doesn't become visible ahead of setting the flag.)
2014 : */
2015 114820 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2016 114820 : MyProc->statusFlags |= PROC_IN_VACUUM;
2017 114820 : if (params->is_wraparound)
2018 93490 : MyProc->statusFlags |= PROC_VACUUM_FOR_WRAPAROUND;
2019 114820 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2020 114820 : LWLockRelease(ProcArrayLock);
2021 : }
2022 :
2023 : /*
2024 : * Need to acquire a snapshot to prevent pg_subtrans from being truncated,
2025 : * cutoff xids in local memory wrapping around, and to have updated xmin
2026 : * horizons.
2027 : */
2028 115212 : PushActiveSnapshot(GetTransactionSnapshot());
2029 :
2030 : /*
2031 : * Check for user-requested abort. Note we want this to be inside a
2032 : * transaction, so xact.c doesn't issue useless WARNING.
2033 : */
2034 115212 : CHECK_FOR_INTERRUPTS();
2035 :
2036 : /*
2037 : * Determine the type of lock we want --- hard exclusive lock for a FULL
2038 : * vacuum, but just ShareUpdateExclusiveLock for concurrent vacuum. Either
2039 : * way, we can be sure that no other backend is vacuuming the same table.
2040 : */
2041 230424 : lmode = (params->options & VACOPT_FULL) ?
2042 115212 : AccessExclusiveLock : ShareUpdateExclusiveLock;
2043 :
2044 : /* open the relation and get the appropriate lock on it */
2045 115212 : rel = vacuum_open_relation(relid, relation, params->options,
2046 115212 : params->log_min_duration >= 0, lmode);
2047 :
2048 : /* leave if relation could not be opened or locked */
2049 115212 : if (!rel)
2050 : {
2051 24 : PopActiveSnapshot();
2052 24 : CommitTransactionCommand();
2053 24 : return false;
2054 : }
2055 :
2056 : /*
2057 : * When recursing to a TOAST table, check privileges on the parent. NB:
2058 : * This is only safe to do because we hold a session lock on the main
2059 : * relation that prevents concurrent deletion.
2060 : */
2061 115188 : if (OidIsValid(params->toast_parent))
2062 7408 : priv_relid = params->toast_parent;
2063 : else
2064 107780 : priv_relid = RelationGetRelid(rel);
2065 :
2066 : /*
2067 : * Check if relation needs to be skipped based on privileges. This check
2068 : * happens also when building the relation list to vacuum for a manual
2069 : * operation, and needs to be done additionally here as VACUUM could
2070 : * happen across multiple transactions where privileges could have changed
2071 : * in-between. Make sure to only generate logs for VACUUM in this case.
2072 : */
2073 115188 : if (!vacuum_is_permitted_for_relation(priv_relid,
2074 : rel->rd_rel,
2075 115188 : params->options & ~VACOPT_ANALYZE))
2076 : {
2077 72 : relation_close(rel, lmode);
2078 72 : PopActiveSnapshot();
2079 72 : CommitTransactionCommand();
2080 72 : return false;
2081 : }
2082 :
2083 : /*
2084 : * Check that it's of a vacuumable relkind.
2085 : */
2086 115116 : if (rel->rd_rel->relkind != RELKIND_RELATION &&
2087 42410 : rel->rd_rel->relkind != RELKIND_MATVIEW &&
2088 42402 : rel->rd_rel->relkind != RELKIND_TOASTVALUE &&
2089 180 : rel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
2090 : {
2091 2 : ereport(WARNING,
2092 : (errmsg("skipping \"%s\" --- cannot vacuum non-tables or special system tables",
2093 : RelationGetRelationName(rel))));
2094 2 : relation_close(rel, lmode);
2095 2 : PopActiveSnapshot();
2096 2 : CommitTransactionCommand();
2097 2 : return false;
2098 : }
2099 :
2100 : /*
2101 : * Silently ignore tables that are temp tables of other backends ---
2102 : * trying to vacuum these will lead to great unhappiness, since their
2103 : * contents are probably not up-to-date on disk. (We don't throw a
2104 : * warning here; it would just lead to chatter during a database-wide
2105 : * VACUUM.)
2106 : */
2107 115114 : if (RELATION_IS_OTHER_TEMP(rel))
2108 : {
2109 2 : relation_close(rel, lmode);
2110 2 : PopActiveSnapshot();
2111 2 : CommitTransactionCommand();
2112 2 : return false;
2113 : }
2114 :
2115 : /*
2116 : * Silently ignore partitioned tables as there is no work to be done. The
2117 : * useful work is on their child partitions, which have been queued up for
2118 : * us separately.
2119 : */
2120 115112 : if (rel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
2121 : {
2122 178 : relation_close(rel, lmode);
2123 178 : PopActiveSnapshot();
2124 178 : CommitTransactionCommand();
2125 : /* It's OK to proceed with ANALYZE on this table */
2126 178 : return true;
2127 : }
2128 :
2129 : /*
2130 : * Get a session-level lock too. This will protect our access to the
2131 : * relation across multiple transactions, so that we can vacuum the
2132 : * relation's TOAST table (if any) secure in the knowledge that no one is
2133 : * deleting the parent relation.
2134 : *
2135 : * NOTE: this cannot block, even if someone else is waiting for access,
2136 : * because the lock manager knows that both lock requests are from the
2137 : * same process.
2138 : */
2139 114934 : lockrelid = rel->rd_lockInfo.lockRelId;
2140 114934 : LockRelationIdForSession(&lockrelid, lmode);
2141 :
2142 : /*
2143 : * Set index_cleanup option based on index_cleanup reloption if it wasn't
2144 : * specified in VACUUM command, or when running in an autovacuum worker
2145 : */
2146 114934 : if (params->index_cleanup == VACOPTVALUE_UNSPECIFIED)
2147 : {
2148 : StdRdOptIndexCleanup vacuum_index_cleanup;
2149 :
2150 99276 : if (rel->rd_options == NULL)
2151 97746 : vacuum_index_cleanup = STDRD_OPTION_VACUUM_INDEX_CLEANUP_AUTO;
2152 : else
2153 1530 : vacuum_index_cleanup =
2154 1530 : ((StdRdOptions *) rel->rd_options)->vacuum_index_cleanup;
2155 :
2156 99276 : if (vacuum_index_cleanup == STDRD_OPTION_VACUUM_INDEX_CLEANUP_AUTO)
2157 99252 : params->index_cleanup = VACOPTVALUE_AUTO;
2158 24 : else if (vacuum_index_cleanup == STDRD_OPTION_VACUUM_INDEX_CLEANUP_ON)
2159 12 : params->index_cleanup = VACOPTVALUE_ENABLED;
2160 : else
2161 : {
2162 : Assert(vacuum_index_cleanup ==
2163 : STDRD_OPTION_VACUUM_INDEX_CLEANUP_OFF);
2164 12 : params->index_cleanup = VACOPTVALUE_DISABLED;
2165 : }
2166 : }
2167 :
2168 : /*
2169 : * Set truncate option based on truncate reloption if it wasn't specified
2170 : * in VACUUM command, or when running in an autovacuum worker
2171 : */
2172 114934 : if (params->truncate == VACOPTVALUE_UNSPECIFIED)
2173 : {
2174 99302 : if (rel->rd_options == NULL ||
2175 1530 : ((StdRdOptions *) rel->rd_options)->vacuum_truncate)
2176 99296 : params->truncate = VACOPTVALUE_ENABLED;
2177 : else
2178 6 : params->truncate = VACOPTVALUE_DISABLED;
2179 : }
2180 :
2181 : /*
2182 : * Remember the relation's TOAST relation for later, if the caller asked
2183 : * us to process it. In VACUUM FULL, though, the toast table is
2184 : * automatically rebuilt by cluster_rel so we shouldn't recurse to it,
2185 : * unless PROCESS_MAIN is disabled.
2186 : */
2187 114934 : if ((params->options & VACOPT_PROCESS_TOAST) != 0 &&
2188 21144 : ((params->options & VACOPT_FULL) == 0 ||
2189 364 : (params->options & VACOPT_PROCESS_MAIN) == 0))
2190 20786 : toast_relid = rel->rd_rel->reltoastrelid;
2191 : else
2192 94148 : toast_relid = InvalidOid;
2193 :
2194 : /*
2195 : * Switch to the table owner's userid, so that any index functions are run
2196 : * as that user. Also lock down security-restricted operations and
2197 : * arrange to make GUC variable changes local to this command. (This is
2198 : * unnecessary, but harmless, for lazy VACUUM.)
2199 : */
2200 114934 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
2201 114934 : SetUserIdAndSecContext(rel->rd_rel->relowner,
2202 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
2203 114934 : save_nestlevel = NewGUCNestLevel();
2204 114934 : RestrictSearchPath();
2205 :
2206 : /*
2207 : * If PROCESS_MAIN is set (the default), it's time to vacuum the main
2208 : * relation. Otherwise, we can skip this part. If processing the TOAST
2209 : * table is required (e.g., PROCESS_TOAST is set), we force PROCESS_MAIN
2210 : * to be set when we recurse to the TOAST table.
2211 : */
2212 114934 : if (params->options & VACOPT_PROCESS_MAIN)
2213 : {
2214 : /*
2215 : * Do the actual work --- either FULL or "lazy" vacuum
2216 : */
2217 114780 : if (params->options & VACOPT_FULL)
2218 : {
2219 358 : ClusterParams cluster_params = {0};
2220 :
2221 358 : if ((params->options & VACOPT_VERBOSE) != 0)
2222 2 : cluster_params.options |= CLUOPT_VERBOSE;
2223 :
2224 : /* VACUUM FULL is now a variant of CLUSTER; see cluster.c */
2225 358 : cluster_rel(rel, InvalidOid, &cluster_params);
2226 : /* cluster_rel closes the relation, but keeps lock */
2227 :
2228 352 : rel = NULL;
2229 : }
2230 : else
2231 114422 : table_relation_vacuum(rel, params, bstrategy);
2232 : }
2233 :
2234 : /* Roll back any GUC changes executed by index functions */
2235 114928 : AtEOXact_GUC(false, save_nestlevel);
2236 :
2237 : /* Restore userid and security context */
2238 114928 : SetUserIdAndSecContext(save_userid, save_sec_context);
2239 :
2240 : /* all done with this class, but hold lock until commit */
2241 114928 : if (rel)
2242 114576 : relation_close(rel, NoLock);
2243 :
2244 : /*
2245 : * Complete the transaction and free all temporary memory used.
2246 : */
2247 114928 : PopActiveSnapshot();
2248 114928 : CommitTransactionCommand();
2249 :
2250 : /*
2251 : * If the relation has a secondary toast rel, vacuum that too while we
2252 : * still hold the session lock on the main table. Note however that
2253 : * "analyze" will not get done on the toast table. This is good, because
2254 : * the toaster always uses hardcoded index access and statistics are
2255 : * totally unimportant for toast relations.
2256 : */
2257 114928 : if (toast_relid != InvalidOid)
2258 : {
2259 : VacuumParams toast_vacuum_params;
2260 :
2261 : /*
2262 : * Force VACOPT_PROCESS_MAIN so vacuum_rel() processes it. Likewise,
2263 : * set toast_parent so that the privilege checks are done on the main
2264 : * relation. NB: This is only safe to do because we hold a session
2265 : * lock on the main relation that prevents concurrent deletion.
2266 : */
2267 7408 : memcpy(&toast_vacuum_params, params, sizeof(VacuumParams));
2268 7408 : toast_vacuum_params.options |= VACOPT_PROCESS_MAIN;
2269 7408 : toast_vacuum_params.toast_parent = relid;
2270 :
2271 7408 : vacuum_rel(toast_relid, NULL, &toast_vacuum_params, bstrategy);
2272 : }
2273 :
2274 : /*
2275 : * Now release the session-level lock on the main table.
2276 : */
2277 114928 : UnlockRelationIdForSession(&lockrelid, lmode);
2278 :
2279 : /* Report that we really did it. */
2280 114928 : return true;
2281 : }
2282 :
2283 :
2284 : /*
2285 : * Open all the vacuumable indexes of the given relation, obtaining the
2286 : * specified kind of lock on each. Return an array of Relation pointers for
2287 : * the indexes into *Irel, and the number of indexes into *nindexes.
2288 : *
2289 : * We consider an index vacuumable if it is marked insertable (indisready).
2290 : * If it isn't, probably a CREATE INDEX CONCURRENTLY command failed early in
2291 : * execution, and what we have is too corrupt to be processable. We will
2292 : * vacuum even if the index isn't indisvalid; this is important because in a
2293 : * unique index, uniqueness checks will be performed anyway and had better not
2294 : * hit dangling index pointers.
2295 : */
2296 : void
2297 127324 : vac_open_indexes(Relation relation, LOCKMODE lockmode,
2298 : int *nindexes, Relation **Irel)
2299 : {
2300 : List *indexoidlist;
2301 : ListCell *indexoidscan;
2302 : int i;
2303 :
2304 : Assert(lockmode != NoLock);
2305 :
2306 127324 : indexoidlist = RelationGetIndexList(relation);
2307 :
2308 : /* allocate enough memory for all indexes */
2309 127324 : i = list_length(indexoidlist);
2310 :
2311 127324 : if (i > 0)
2312 118752 : *Irel = (Relation *) palloc(i * sizeof(Relation));
2313 : else
2314 8572 : *Irel = NULL;
2315 :
2316 : /* collect just the ready indexes */
2317 127324 : i = 0;
2318 316116 : foreach(indexoidscan, indexoidlist)
2319 : {
2320 188792 : Oid indexoid = lfirst_oid(indexoidscan);
2321 : Relation indrel;
2322 :
2323 188792 : indrel = index_open(indexoid, lockmode);
2324 188792 : if (indrel->rd_index->indisready)
2325 188792 : (*Irel)[i++] = indrel;
2326 : else
2327 0 : index_close(indrel, lockmode);
2328 : }
2329 :
2330 127324 : *nindexes = i;
2331 :
2332 127324 : list_free(indexoidlist);
2333 127324 : }
2334 :
2335 : /*
2336 : * Release the resources acquired by vac_open_indexes. Optionally release
2337 : * the locks (say NoLock to keep 'em).
2338 : */
2339 : void
2340 128108 : vac_close_indexes(int nindexes, Relation *Irel, LOCKMODE lockmode)
2341 : {
2342 128108 : if (Irel == NULL)
2343 9364 : return;
2344 :
2345 307520 : while (nindexes--)
2346 : {
2347 188776 : Relation ind = Irel[nindexes];
2348 :
2349 188776 : index_close(ind, lockmode);
2350 : }
2351 118744 : pfree(Irel);
2352 : }
2353 :
2354 : /*
2355 : * vacuum_delay_point --- check for interrupts and cost-based delay.
2356 : *
2357 : * This should be called in each major loop of VACUUM processing,
2358 : * typically once per page processed.
2359 : */
2360 : void
2361 70316542 : vacuum_delay_point(void)
2362 : {
2363 70316542 : double msec = 0;
2364 :
2365 : /* Always check for interrupts */
2366 70316542 : CHECK_FOR_INTERRUPTS();
2367 :
2368 70316542 : if (InterruptPending ||
2369 70316542 : (!VacuumCostActive && !ConfigReloadPending))
2370 63397674 : return;
2371 :
2372 : /*
2373 : * Autovacuum workers should reload the configuration file if requested.
2374 : * This allows changes to [autovacuum_]vacuum_cost_limit and
2375 : * [autovacuum_]vacuum_cost_delay to take effect while a table is being
2376 : * vacuumed or analyzed.
2377 : */
2378 6918868 : if (ConfigReloadPending && AmAutoVacuumWorkerProcess())
2379 : {
2380 0 : ConfigReloadPending = false;
2381 0 : ProcessConfigFile(PGC_SIGHUP);
2382 0 : VacuumUpdateCosts();
2383 : }
2384 :
2385 : /*
2386 : * If we disabled cost-based delays after reloading the config file,
2387 : * return.
2388 : */
2389 6918868 : if (!VacuumCostActive)
2390 0 : return;
2391 :
2392 : /*
2393 : * For parallel vacuum, the delay is computed based on the shared cost
2394 : * balance. See compute_parallel_delay.
2395 : */
2396 6918868 : if (VacuumSharedCostBalance != NULL)
2397 0 : msec = compute_parallel_delay();
2398 6918868 : else if (VacuumCostBalance >= vacuum_cost_limit)
2399 4062 : msec = vacuum_cost_delay * VacuumCostBalance / vacuum_cost_limit;
2400 :
2401 : /* Nap if appropriate */
2402 6918868 : if (msec > 0)
2403 : {
2404 4062 : if (msec > vacuum_cost_delay * 4)
2405 10 : msec = vacuum_cost_delay * 4;
2406 :
2407 4062 : pgstat_report_wait_start(WAIT_EVENT_VACUUM_DELAY);
2408 4062 : pg_usleep(msec * 1000);
2409 4062 : pgstat_report_wait_end();
2410 :
2411 : /*
2412 : * We don't want to ignore postmaster death during very long vacuums
2413 : * with vacuum_cost_delay configured. We can't use the usual
2414 : * WaitLatch() approach here because we want microsecond-based sleep
2415 : * durations above.
2416 : */
2417 4062 : if (IsUnderPostmaster && !PostmasterIsAlive())
2418 0 : exit(1);
2419 :
2420 4062 : VacuumCostBalance = 0;
2421 :
2422 : /*
2423 : * Balance and update limit values for autovacuum workers. We must do
2424 : * this periodically, as the number of workers across which we are
2425 : * balancing the limit may have changed.
2426 : *
2427 : * TODO: There may be better criteria for determining when to do this
2428 : * besides "check after napping".
2429 : */
2430 4062 : AutoVacuumUpdateCostLimit();
2431 :
2432 : /* Might have gotten an interrupt while sleeping */
2433 4062 : CHECK_FOR_INTERRUPTS();
2434 : }
2435 : }
2436 :
2437 : /*
2438 : * Computes the vacuum delay for parallel workers.
2439 : *
2440 : * The basic idea of a cost-based delay for parallel vacuum is to allow each
2441 : * worker to sleep in proportion to the share of work it's done. We achieve this
2442 : * by allowing all parallel vacuum workers including the leader process to
2443 : * have a shared view of cost related parameters (mainly VacuumCostBalance).
2444 : * We allow each worker to update it as and when it has incurred any cost and
2445 : * then based on that decide whether it needs to sleep. We compute the time
2446 : * to sleep for a worker based on the cost it has incurred
2447 : * (VacuumCostBalanceLocal) and then reduce the VacuumSharedCostBalance by
2448 : * that amount. This avoids putting to sleep those workers which have done less
2449 : * I/O than other workers and therefore ensure that workers
2450 : * which are doing more I/O got throttled more.
2451 : *
2452 : * We allow a worker to sleep only if it has performed I/O above a certain
2453 : * threshold, which is calculated based on the number of active workers
2454 : * (VacuumActiveNWorkers), and the overall cost balance is more than
2455 : * VacuumCostLimit set by the system. Testing reveals that we achieve
2456 : * the required throttling if we force a worker that has done more than 50%
2457 : * of its share of work to sleep.
2458 : */
2459 : static double
2460 0 : compute_parallel_delay(void)
2461 : {
2462 0 : double msec = 0;
2463 : uint32 shared_balance;
2464 : int nworkers;
2465 :
2466 : /* Parallel vacuum must be active */
2467 : Assert(VacuumSharedCostBalance);
2468 :
2469 0 : nworkers = pg_atomic_read_u32(VacuumActiveNWorkers);
2470 :
2471 : /* At least count itself */
2472 : Assert(nworkers >= 1);
2473 :
2474 : /* Update the shared cost balance value atomically */
2475 0 : shared_balance = pg_atomic_add_fetch_u32(VacuumSharedCostBalance, VacuumCostBalance);
2476 :
2477 : /* Compute the total local balance for the current worker */
2478 0 : VacuumCostBalanceLocal += VacuumCostBalance;
2479 :
2480 0 : if ((shared_balance >= vacuum_cost_limit) &&
2481 0 : (VacuumCostBalanceLocal > 0.5 * ((double) vacuum_cost_limit / nworkers)))
2482 : {
2483 : /* Compute sleep time based on the local cost balance */
2484 0 : msec = vacuum_cost_delay * VacuumCostBalanceLocal / vacuum_cost_limit;
2485 0 : pg_atomic_sub_fetch_u32(VacuumSharedCostBalance, VacuumCostBalanceLocal);
2486 0 : VacuumCostBalanceLocal = 0;
2487 : }
2488 :
2489 : /*
2490 : * Reset the local balance as we accumulated it into the shared value.
2491 : */
2492 0 : VacuumCostBalance = 0;
2493 :
2494 0 : return msec;
2495 : }
2496 :
2497 : /*
2498 : * A wrapper function of defGetBoolean().
2499 : *
2500 : * This function returns VACOPTVALUE_ENABLED and VACOPTVALUE_DISABLED instead
2501 : * of true and false.
2502 : */
2503 : static VacOptValue
2504 316 : get_vacoptval_from_boolean(DefElem *def)
2505 : {
2506 316 : return defGetBoolean(def) ? VACOPTVALUE_ENABLED : VACOPTVALUE_DISABLED;
2507 : }
2508 :
2509 : /*
2510 : * vac_bulkdel_one_index() -- bulk-deletion for index relation.
2511 : *
2512 : * Returns bulk delete stats derived from input stats
2513 : */
2514 : IndexBulkDeleteResult *
2515 2058 : vac_bulkdel_one_index(IndexVacuumInfo *ivinfo, IndexBulkDeleteResult *istat,
2516 : TidStore *dead_items, VacDeadItemsInfo *dead_items_info)
2517 : {
2518 : /* Do bulk deletion */
2519 2058 : istat = index_bulk_delete(ivinfo, istat, vac_tid_reaped,
2520 : dead_items);
2521 :
2522 2058 : ereport(ivinfo->message_level,
2523 : (errmsg("scanned index \"%s\" to remove %lld row versions",
2524 : RelationGetRelationName(ivinfo->index),
2525 : (long long) dead_items_info->num_items)));
2526 :
2527 2058 : return istat;
2528 : }
2529 :
2530 : /*
2531 : * vac_cleanup_one_index() -- do post-vacuum cleanup for index relation.
2532 : *
2533 : * Returns bulk delete stats derived from input stats
2534 : */
2535 : IndexBulkDeleteResult *
2536 140800 : vac_cleanup_one_index(IndexVacuumInfo *ivinfo, IndexBulkDeleteResult *istat)
2537 : {
2538 140800 : istat = index_vacuum_cleanup(ivinfo, istat);
2539 :
2540 140800 : if (istat)
2541 2316 : ereport(ivinfo->message_level,
2542 : (errmsg("index \"%s\" now contains %.0f row versions in %u pages",
2543 : RelationGetRelationName(ivinfo->index),
2544 : istat->num_index_tuples,
2545 : istat->num_pages),
2546 : errdetail("%.0f index row versions were removed.\n"
2547 : "%u index pages were newly deleted.\n"
2548 : "%u index pages are currently deleted, of which %u are currently reusable.",
2549 : istat->tuples_removed,
2550 : istat->pages_newly_deleted,
2551 : istat->pages_deleted, istat->pages_free)));
2552 :
2553 140800 : return istat;
2554 : }
2555 :
2556 : /*
2557 : * vac_tid_reaped() -- is a particular tid deletable?
2558 : *
2559 : * This has the right signature to be an IndexBulkDeleteCallback.
2560 : */
2561 : static bool
2562 6032946 : vac_tid_reaped(ItemPointer itemptr, void *state)
2563 : {
2564 6032946 : TidStore *dead_items = (TidStore *) state;
2565 :
2566 6032946 : return TidStoreIsMember(dead_items, itemptr);
2567 : }
|
