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 : * REPACK, handled in repack.c.
13 : *
14 : *
15 : * Portions Copyright (c) 1996-2026, 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/async.h"
41 : #include "commands/defrem.h"
42 : #include "commands/progress.h"
43 : #include "commands/repack.h"
44 : #include "commands/vacuum.h"
45 : #include "miscadmin.h"
46 : #include "nodes/makefuncs.h"
47 : #include "pgstat.h"
48 : #include "postmaster/autovacuum.h"
49 : #include "postmaster/bgworker_internals.h"
50 : #include "postmaster/interrupt.h"
51 : #include "storage/bufmgr.h"
52 : #include "storage/lmgr.h"
53 : #include "storage/pmsignal.h"
54 : #include "storage/proc.h"
55 : #include "storage/procarray.h"
56 : #include "utils/acl.h"
57 : #include "utils/fmgroids.h"
58 : #include "utils/guc.h"
59 : #include "utils/guc_hooks.h"
60 : #include "utils/injection_point.h"
61 : #include "utils/memutils.h"
62 : #include "utils/snapmgr.h"
63 : #include "utils/syscache.h"
64 : #include "utils/wait_event.h"
65 :
66 : /*
67 : * Minimum interval for cost-based vacuum delay reports from a parallel worker.
68 : * This aims to avoid sending too many messages and waking up the leader too
69 : * frequently.
70 : */
71 : #define PARALLEL_VACUUM_DELAY_REPORT_INTERVAL_NS (NS_PER_S)
72 :
73 : /*
74 : * GUC parameters
75 : */
76 : int vacuum_freeze_min_age;
77 : int vacuum_freeze_table_age;
78 : int vacuum_multixact_freeze_min_age;
79 : int vacuum_multixact_freeze_table_age;
80 : int vacuum_failsafe_age;
81 : int vacuum_multixact_failsafe_age;
82 : double vacuum_max_eager_freeze_failure_rate;
83 : bool track_cost_delay_timing;
84 : bool vacuum_truncate;
85 :
86 : /*
87 : * Variables for cost-based vacuum delay. The defaults differ between
88 : * autovacuum and vacuum. They should be set with the appropriate GUC value in
89 : * vacuum code. They are initialized here to the defaults for client backends
90 : * executing VACUUM or ANALYZE.
91 : */
92 : double vacuum_cost_delay = 0;
93 : int vacuum_cost_limit = 200;
94 :
95 : /* Variable for reporting cost-based vacuum delay from parallel workers. */
96 : int64 parallel_vacuum_worker_delay_ns = 0;
97 :
98 : /*
99 : * VacuumFailsafeActive is a defined as a global so that we can determine
100 : * whether or not to re-enable cost-based vacuum delay when vacuuming a table.
101 : * If failsafe mode has been engaged, we will not re-enable cost-based delay
102 : * for the table until after vacuuming has completed, regardless of other
103 : * settings.
104 : *
105 : * Only VACUUM code should inspect this variable and only table access methods
106 : * should set it to true. In Table AM-agnostic VACUUM code, this variable is
107 : * inspected to determine whether or not to allow cost-based delays. Table AMs
108 : * are free to set it if they desire this behavior, but it is false by default
109 : * and reset to false in between vacuuming each relation.
110 : */
111 : bool VacuumFailsafeActive = false;
112 :
113 : /*
114 : * Variables for cost-based parallel vacuum. See comments atop
115 : * compute_parallel_delay to understand how it works.
116 : */
117 : pg_atomic_uint32 *VacuumSharedCostBalance = NULL;
118 : pg_atomic_uint32 *VacuumActiveNWorkers = NULL;
119 : int VacuumCostBalanceLocal = 0;
120 :
121 : /* non-export function prototypes */
122 : static List *expand_vacuum_rel(VacuumRelation *vrel,
123 : MemoryContext vac_context, int options);
124 : static List *get_all_vacuum_rels(MemoryContext vac_context, int options);
125 : static void vac_truncate_clog(TransactionId frozenXID,
126 : MultiXactId minMulti,
127 : TransactionId lastSaneFrozenXid,
128 : MultiXactId lastSaneMinMulti);
129 : static bool vacuum_rel(Oid relid, RangeVar *relation, VacuumParams params,
130 : BufferAccessStrategy bstrategy, bool isTopLevel);
131 : static double compute_parallel_delay(void);
132 : static VacOptValue get_vacoptval_from_boolean(DefElem *def);
133 : static bool vac_tid_reaped(ItemPointer itemptr, void *state);
134 :
135 : /*
136 : * GUC check function to ensure GUC value specified is within the allowable
137 : * range.
138 : */
139 : bool
140 1275 : check_vacuum_buffer_usage_limit(int *newval, void **extra,
141 : GucSource source)
142 : {
143 : /* Value upper and lower hard limits are inclusive */
144 1275 : if (*newval == 0 || (*newval >= MIN_BAS_VAC_RING_SIZE_KB &&
145 1275 : *newval <= MAX_BAS_VAC_RING_SIZE_KB))
146 1275 : return true;
147 :
148 : /* Value does not fall within any allowable range */
149 0 : GUC_check_errdetail("\"%s\" must be 0 or between %d kB and %d kB.",
150 : "vacuum_buffer_usage_limit",
151 : MIN_BAS_VAC_RING_SIZE_KB, MAX_BAS_VAC_RING_SIZE_KB);
152 :
153 0 : return false;
154 : }
155 :
156 : /*
157 : * Primary entry point for manual VACUUM and ANALYZE commands
158 : *
159 : * This is mainly a preparation wrapper for the real operations that will
160 : * happen in vacuum().
161 : */
162 : void
163 8704 : ExecVacuum(ParseState *pstate, VacuumStmt *vacstmt, bool isTopLevel)
164 : {
165 : VacuumParams params;
166 8704 : BufferAccessStrategy bstrategy = NULL;
167 8704 : bool verbose = false;
168 8704 : bool skip_locked = false;
169 8704 : bool analyze = false;
170 8704 : bool freeze = false;
171 8704 : bool full = false;
172 8704 : bool disable_page_skipping = false;
173 8704 : bool process_main = true;
174 8704 : bool process_toast = true;
175 : int ring_size;
176 8704 : bool skip_database_stats = false;
177 8704 : bool only_database_stats = false;
178 : MemoryContext vac_context;
179 : ListCell *lc;
180 :
181 : /* index_cleanup and truncate values unspecified for now */
182 8704 : params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
183 8704 : params.truncate = VACOPTVALUE_UNSPECIFIED;
184 :
185 : /* By default parallel vacuum is enabled */
186 8704 : params.nworkers = 0;
187 :
188 : /* Will be set later if we recurse to a TOAST table. */
189 8704 : params.toast_parent = InvalidOid;
190 :
191 : /*
192 : * Set this to an invalid value so it is clear whether or not a
193 : * BUFFER_USAGE_LIMIT was specified when making the access strategy.
194 : */
195 8704 : ring_size = -1;
196 :
197 : /* Parse options list */
198 18266 : foreach(lc, vacstmt->options)
199 : {
200 9586 : DefElem *opt = (DefElem *) lfirst(lc);
201 :
202 : /* Parse common options for VACUUM and ANALYZE */
203 9586 : if (strcmp(opt->defname, "verbose") == 0)
204 25 : verbose = defGetBoolean(opt);
205 9561 : else if (strcmp(opt->defname, "skip_locked") == 0)
206 182 : skip_locked = defGetBoolean(opt);
207 9379 : else if (strcmp(opt->defname, "buffer_usage_limit") == 0)
208 : {
209 : const char *hintmsg;
210 : int result;
211 : char *vac_buffer_size;
212 :
213 36 : vac_buffer_size = defGetString(opt);
214 :
215 : /*
216 : * Check that the specified value is valid and the size falls
217 : * within the hard upper and lower limits if it is not 0.
218 : */
219 36 : if (!parse_int(vac_buffer_size, &result, GUC_UNIT_KB, &hintmsg) ||
220 32 : (result != 0 &&
221 24 : (result < MIN_BAS_VAC_RING_SIZE_KB || result > MAX_BAS_VAC_RING_SIZE_KB)))
222 : {
223 12 : ereport(ERROR,
224 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
225 : errmsg("%s option must be 0 or between %d kB and %d kB",
226 : "BUFFER_USAGE_LIMIT",
227 : MIN_BAS_VAC_RING_SIZE_KB, MAX_BAS_VAC_RING_SIZE_KB),
228 : hintmsg ? errhint_internal("%s", _(hintmsg)) : 0));
229 : }
230 :
231 24 : ring_size = result;
232 : }
233 9343 : else if (!vacstmt->is_vacuumcmd)
234 4 : ereport(ERROR,
235 : (errcode(ERRCODE_SYNTAX_ERROR),
236 : errmsg("unrecognized %s option \"%s\"",
237 : "ANALYZE", opt->defname),
238 : parser_errposition(pstate, opt->location)));
239 :
240 : /* Parse options available on VACUUM */
241 9339 : else if (strcmp(opt->defname, "analyze") == 0)
242 1837 : analyze = defGetBoolean(opt);
243 7502 : else if (strcmp(opt->defname, "freeze") == 0)
244 1891 : freeze = defGetBoolean(opt);
245 5611 : else if (strcmp(opt->defname, "full") == 0)
246 230 : full = defGetBoolean(opt);
247 5381 : else if (strcmp(opt->defname, "disable_page_skipping") == 0)
248 120 : disable_page_skipping = defGetBoolean(opt);
249 5261 : else if (strcmp(opt->defname, "index_cleanup") == 0)
250 : {
251 : /* Interpret no string as the default, which is 'auto' */
252 98 : if (!opt->arg)
253 0 : params.index_cleanup = VACOPTVALUE_AUTO;
254 : else
255 : {
256 98 : char *sval = defGetString(opt);
257 :
258 : /* Try matching on 'auto' string, or fall back on boolean */
259 98 : if (pg_strcasecmp(sval, "auto") == 0)
260 4 : params.index_cleanup = VACOPTVALUE_AUTO;
261 : else
262 94 : params.index_cleanup = get_vacoptval_from_boolean(opt);
263 : }
264 : }
265 5163 : else if (strcmp(opt->defname, "process_main") == 0)
266 85 : process_main = defGetBoolean(opt);
267 5078 : else if (strcmp(opt->defname, "process_toast") == 0)
268 89 : process_toast = defGetBoolean(opt);
269 4989 : else if (strcmp(opt->defname, "truncate") == 0)
270 88 : params.truncate = get_vacoptval_from_boolean(opt);
271 4901 : else if (strcmp(opt->defname, "parallel") == 0)
272 : {
273 199 : int nworkers = defGetInt32(opt);
274 :
275 195 : if (nworkers < 0 || nworkers > MAX_PARALLEL_WORKER_LIMIT)
276 4 : ereport(ERROR,
277 : (errcode(ERRCODE_SYNTAX_ERROR),
278 : errmsg("%s option must be between 0 and %d",
279 : "PARALLEL",
280 : MAX_PARALLEL_WORKER_LIMIT),
281 : parser_errposition(pstate, opt->location)));
282 :
283 : /*
284 : * Disable parallel vacuum, if user has specified parallel degree
285 : * as zero.
286 : */
287 191 : if (nworkers == 0)
288 86 : params.nworkers = -1;
289 : else
290 105 : params.nworkers = nworkers;
291 : }
292 4702 : else if (strcmp(opt->defname, "skip_database_stats") == 0)
293 4623 : skip_database_stats = defGetBoolean(opt);
294 79 : else if (strcmp(opt->defname, "only_database_stats") == 0)
295 79 : only_database_stats = defGetBoolean(opt);
296 : else
297 0 : ereport(ERROR,
298 : (errcode(ERRCODE_SYNTAX_ERROR),
299 : errmsg("unrecognized %s option \"%s\"",
300 : "VACUUM", opt->defname),
301 : parser_errposition(pstate, opt->location)));
302 : }
303 :
304 : /* Set vacuum options */
305 8680 : params.options =
306 8680 : (vacstmt->is_vacuumcmd ? VACOPT_VACUUM : VACOPT_ANALYZE) |
307 8680 : (verbose ? VACOPT_VERBOSE : 0) |
308 8680 : (skip_locked ? VACOPT_SKIP_LOCKED : 0) |
309 8680 : (analyze ? VACOPT_ANALYZE : 0) |
310 8680 : (freeze ? VACOPT_FREEZE : 0) |
311 8680 : (full ? VACOPT_FULL : 0) |
312 8680 : (disable_page_skipping ? VACOPT_DISABLE_PAGE_SKIPPING : 0) |
313 8680 : (process_main ? VACOPT_PROCESS_MAIN : 0) |
314 8680 : (process_toast ? VACOPT_PROCESS_TOAST : 0) |
315 8680 : (skip_database_stats ? VACOPT_SKIP_DATABASE_STATS : 0) |
316 8680 : (only_database_stats ? VACOPT_ONLY_DATABASE_STATS : 0);
317 :
318 : /* sanity checks on options */
319 : Assert(params.options & (VACOPT_VACUUM | VACOPT_ANALYZE));
320 : Assert((params.options & VACOPT_VACUUM) ||
321 : !(params.options & (VACOPT_FULL | VACOPT_FREEZE)));
322 :
323 8680 : if ((params.options & VACOPT_FULL) && params.nworkers > 0)
324 4 : ereport(ERROR,
325 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
326 : errmsg("VACUUM FULL cannot be performed in parallel")));
327 :
328 : /*
329 : * BUFFER_USAGE_LIMIT does nothing for VACUUM (FULL) so just raise an
330 : * ERROR for that case. VACUUM (FULL, ANALYZE) does make use of it, so
331 : * we'll permit that.
332 : */
333 8676 : if (ring_size != -1 && (params.options & VACOPT_FULL) &&
334 4 : !(params.options & VACOPT_ANALYZE))
335 4 : ereport(ERROR,
336 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
337 : errmsg("BUFFER_USAGE_LIMIT cannot be specified for VACUUM FULL")));
338 :
339 : /*
340 : * Make sure VACOPT_ANALYZE is specified if any column lists are present.
341 : */
342 8672 : if (!(params.options & VACOPT_ANALYZE))
343 : {
344 7793 : foreach(lc, vacstmt->rels)
345 : {
346 3841 : VacuumRelation *vrel = lfirst_node(VacuumRelation, lc);
347 :
348 3841 : if (vrel->va_cols != NIL)
349 4 : ereport(ERROR,
350 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
351 : errmsg("ANALYZE option must be specified when a column list is provided")));
352 : }
353 : }
354 :
355 : /*
356 : * Sanity check DISABLE_PAGE_SKIPPING option.
357 : */
358 8668 : if ((params.options & VACOPT_FULL) != 0 &&
359 214 : (params.options & VACOPT_DISABLE_PAGE_SKIPPING) != 0)
360 0 : ereport(ERROR,
361 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
362 : errmsg("VACUUM option DISABLE_PAGE_SKIPPING cannot be used with FULL")));
363 :
364 : /* sanity check for PROCESS_TOAST */
365 8668 : if ((params.options & VACOPT_FULL) != 0 &&
366 214 : (params.options & VACOPT_PROCESS_TOAST) == 0)
367 4 : ereport(ERROR,
368 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
369 : errmsg("PROCESS_TOAST required with VACUUM FULL")));
370 :
371 : /* sanity check for ONLY_DATABASE_STATS */
372 8664 : if (params.options & VACOPT_ONLY_DATABASE_STATS)
373 : {
374 : Assert(params.options & VACOPT_VACUUM);
375 79 : if (vacstmt->rels != NIL)
376 4 : ereport(ERROR,
377 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
378 : errmsg("ONLY_DATABASE_STATS cannot be specified with a list of tables")));
379 : /* don't require people to turn off PROCESS_TOAST/MAIN explicitly */
380 75 : if (params.options & ~(VACOPT_VACUUM |
381 : VACOPT_VERBOSE |
382 : VACOPT_PROCESS_MAIN |
383 : VACOPT_PROCESS_TOAST |
384 : VACOPT_ONLY_DATABASE_STATS))
385 0 : ereport(ERROR,
386 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
387 : errmsg("ONLY_DATABASE_STATS cannot be specified with other VACUUM options")));
388 : }
389 :
390 : /*
391 : * All freeze ages are zero if the FREEZE option is given; otherwise pass
392 : * them as -1 which means to use the default values.
393 : */
394 8660 : if (params.options & VACOPT_FREEZE)
395 : {
396 1891 : params.freeze_min_age = 0;
397 1891 : params.freeze_table_age = 0;
398 1891 : params.multixact_freeze_min_age = 0;
399 1891 : params.multixact_freeze_table_age = 0;
400 : }
401 : else
402 : {
403 6769 : params.freeze_min_age = -1;
404 6769 : params.freeze_table_age = -1;
405 6769 : params.multixact_freeze_min_age = -1;
406 6769 : params.multixact_freeze_table_age = -1;
407 : }
408 :
409 : /* user-invoked vacuum is never "for wraparound" */
410 8660 : params.is_wraparound = false;
411 :
412 : /*
413 : * user-invoked vacuum uses VACOPT_VERBOSE instead of
414 : * log_vacuum_min_duration and log_analyze_min_duration
415 : */
416 8660 : params.log_vacuum_min_duration = -1;
417 8660 : params.log_analyze_min_duration = -1;
418 :
419 : /*
420 : * Later, in vacuum_rel(), we check if a reloption override was specified.
421 : */
422 8660 : params.max_eager_freeze_failure_rate = vacuum_max_eager_freeze_failure_rate;
423 :
424 : /*
425 : * Create special memory context for cross-transaction storage.
426 : *
427 : * Since it is a child of PortalContext, it will go away eventually even
428 : * if we suffer an error; there's no need for special abort cleanup logic.
429 : */
430 8660 : vac_context = AllocSetContextCreate(PortalContext,
431 : "Vacuum",
432 : ALLOCSET_DEFAULT_SIZES);
433 :
434 : /*
435 : * Make a buffer strategy object in the cross-transaction memory context.
436 : * We needn't bother making this for VACUUM (FULL) or VACUUM
437 : * (ONLY_DATABASE_STATS) as they'll not make use of it. VACUUM (FULL,
438 : * ANALYZE) is possible, so we'd better ensure that we make a strategy
439 : * when we see ANALYZE.
440 : */
441 8660 : if ((params.options & (VACOPT_ONLY_DATABASE_STATS |
442 285 : VACOPT_FULL)) == 0 ||
443 285 : (params.options & VACOPT_ANALYZE) != 0)
444 : {
445 :
446 8379 : MemoryContext old_context = MemoryContextSwitchTo(vac_context);
447 :
448 : Assert(ring_size >= -1);
449 :
450 : /*
451 : * If BUFFER_USAGE_LIMIT was specified by the VACUUM or ANALYZE
452 : * command, it overrides the value of VacuumBufferUsageLimit. Either
453 : * value may be 0, in which case GetAccessStrategyWithSize() will
454 : * return NULL, effectively allowing full use of shared buffers.
455 : */
456 8379 : if (ring_size == -1)
457 8359 : ring_size = VacuumBufferUsageLimit;
458 :
459 8379 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, ring_size);
460 :
461 8379 : MemoryContextSwitchTo(old_context);
462 : }
463 :
464 : /* Now go through the common routine */
465 8660 : vacuum(vacstmt->rels, ¶ms, bstrategy, vac_context, isTopLevel);
466 :
467 : /* Finally, clean up the vacuum memory context */
468 8573 : MemoryContextDelete(vac_context);
469 8573 : }
470 :
471 : /*
472 : * Internal entry point for autovacuum and the VACUUM / ANALYZE commands.
473 : *
474 : * relations, if not NIL, is a list of VacuumRelation to process; otherwise,
475 : * we process all relevant tables in the database. For each VacuumRelation,
476 : * if a valid OID is supplied, the table with that OID is what to process;
477 : * otherwise, the VacuumRelation's RangeVar indicates what to process.
478 : *
479 : * params contains a set of parameters that can be used to customize the
480 : * behavior.
481 : *
482 : * bstrategy may be passed in as NULL when the caller does not want to
483 : * restrict the number of shared_buffers that VACUUM / ANALYZE can use,
484 : * otherwise, the caller must build a BufferAccessStrategy with the number of
485 : * shared_buffers that VACUUM / ANALYZE should try to limit themselves to
486 : * using.
487 : *
488 : * isTopLevel should be passed down from ProcessUtility.
489 : *
490 : * It is the caller's responsibility that all parameters are allocated in a
491 : * memory context that will not disappear at transaction commit.
492 : */
493 : void
494 108702 : vacuum(List *relations, const VacuumParams *params, BufferAccessStrategy bstrategy,
495 : MemoryContext vac_context, bool isTopLevel)
496 : {
497 : static bool in_vacuum = false;
498 :
499 : const char *stmttype;
500 : volatile bool in_outer_xact,
501 : use_own_xacts;
502 :
503 108702 : stmttype = (params->options & VACOPT_VACUUM) ? "VACUUM" : "ANALYZE";
504 :
505 : /*
506 : * We cannot run VACUUM inside a user transaction block; if we were inside
507 : * a transaction, then our commit- and start-transaction-command calls
508 : * would not have the intended effect! There are numerous other subtle
509 : * dependencies on this, too.
510 : *
511 : * ANALYZE (without VACUUM) can run either way.
512 : */
513 108702 : if (params->options & VACOPT_VACUUM)
514 : {
515 105654 : PreventInTransactionBlock(isTopLevel, stmttype);
516 105641 : in_outer_xact = false;
517 : }
518 : else
519 3048 : in_outer_xact = IsInTransactionBlock(isTopLevel);
520 :
521 : /*
522 : * Check for and disallow recursive calls. This could happen when VACUUM
523 : * FULL or ANALYZE calls a hostile index expression that itself calls
524 : * ANALYZE.
525 : */
526 108689 : if (in_vacuum)
527 8 : ereport(ERROR,
528 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
529 : errmsg("%s cannot be executed from VACUUM or ANALYZE",
530 : stmttype)));
531 :
532 : /*
533 : * Build list of relation(s) to process, putting any new data in
534 : * vac_context for safekeeping.
535 : */
536 108681 : if (params->options & VACOPT_ONLY_DATABASE_STATS)
537 : {
538 : /* We don't process any tables in this case */
539 : Assert(relations == NIL);
540 : }
541 108606 : else if (relations != NIL)
542 : {
543 108481 : List *newrels = NIL;
544 : ListCell *lc;
545 :
546 217043 : foreach(lc, relations)
547 : {
548 108586 : VacuumRelation *vrel = lfirst_node(VacuumRelation, lc);
549 : List *sublist;
550 : MemoryContext old_context;
551 :
552 108586 : sublist = expand_vacuum_rel(vrel, vac_context, params->options);
553 108562 : old_context = MemoryContextSwitchTo(vac_context);
554 108562 : newrels = list_concat(newrels, sublist);
555 108562 : MemoryContextSwitchTo(old_context);
556 : }
557 108457 : relations = newrels;
558 : }
559 : else
560 125 : relations = get_all_vacuum_rels(vac_context, params->options);
561 :
562 : /*
563 : * Decide whether we need to start/commit our own transactions.
564 : *
565 : * For VACUUM (with or without ANALYZE): always do so, so that we can
566 : * release locks as soon as possible. (We could possibly use the outer
567 : * transaction for a one-table VACUUM, but handling TOAST tables would be
568 : * problematic.)
569 : *
570 : * For ANALYZE (no VACUUM): if inside a transaction block, we cannot
571 : * start/commit our own transactions. Also, there's no need to do so if
572 : * only processing one relation. For multiple relations when not within a
573 : * transaction block, and also in an autovacuum worker, use own
574 : * transactions so we can release locks sooner.
575 : */
576 108657 : if (params->options & VACOPT_VACUUM)
577 105633 : use_own_xacts = true;
578 : else
579 : {
580 : Assert(params->options & VACOPT_ANALYZE);
581 3024 : if (AmAutoVacuumWorkerProcess())
582 169 : use_own_xacts = true;
583 2855 : else if (in_outer_xact)
584 164 : use_own_xacts = false;
585 2691 : else if (list_length(relations) > 1)
586 524 : use_own_xacts = true;
587 : else
588 2167 : use_own_xacts = false;
589 : }
590 :
591 : /*
592 : * vacuum_rel expects to be entered with no transaction active; it will
593 : * start and commit its own transaction. But we are called by an SQL
594 : * command, and so we are executing inside a transaction already. We
595 : * commit the transaction started in PostgresMain() here, and start
596 : * another one before exiting to match the commit waiting for us back in
597 : * PostgresMain().
598 : */
599 108657 : if (use_own_xacts)
600 : {
601 : Assert(!in_outer_xact);
602 :
603 : /* ActiveSnapshot is not set by autovacuum */
604 106326 : if (ActiveSnapshotSet())
605 6284 : PopActiveSnapshot();
606 :
607 : /* matches the StartTransaction in PostgresMain() */
608 106326 : CommitTransactionCommand();
609 : }
610 :
611 : /* Turn vacuum cost accounting on or off, and set/clear in_vacuum */
612 108657 : PG_TRY();
613 : {
614 : ListCell *cur;
615 :
616 108657 : in_vacuum = true;
617 108657 : VacuumFailsafeActive = false;
618 108657 : VacuumUpdateCosts();
619 108657 : VacuumCostBalance = 0;
620 108657 : VacuumCostBalanceLocal = 0;
621 108657 : VacuumSharedCostBalance = NULL;
622 108657 : VacuumActiveNWorkers = NULL;
623 :
624 : /*
625 : * Loop to process each selected relation.
626 : */
627 228024 : foreach(cur, relations)
628 : {
629 119409 : VacuumRelation *vrel = lfirst_node(VacuumRelation, cur);
630 :
631 119409 : if (params->options & VACOPT_VACUUM)
632 : {
633 111167 : if (!vacuum_rel(vrel->oid, vrel->relation, *params, bstrategy,
634 : isTopLevel))
635 62 : continue;
636 : }
637 :
638 119343 : if (params->options & VACOPT_ANALYZE)
639 : {
640 : /*
641 : * If using separate xacts, start one for analyze. Otherwise,
642 : * we can use the outer transaction.
643 : */
644 10218 : if (use_own_xacts)
645 : {
646 7903 : StartTransactionCommand();
647 : /* functions in indexes may want a snapshot set */
648 7903 : PushActiveSnapshot(GetTransactionSnapshot());
649 : }
650 :
651 10218 : analyze_rel(vrel->oid, vrel->relation, params,
652 : vrel->va_cols, in_outer_xact, bstrategy);
653 :
654 10180 : if (use_own_xacts)
655 : {
656 7878 : PopActiveSnapshot();
657 : /* standard_ProcessUtility() does CCI if !use_own_xacts */
658 7878 : CommandCounterIncrement();
659 7878 : CommitTransactionCommand();
660 : }
661 : else
662 : {
663 : /*
664 : * If we're not using separate xacts, better separate the
665 : * ANALYZE actions with CCIs. This avoids trouble if user
666 : * says "ANALYZE t, t".
667 : */
668 2302 : CommandCounterIncrement();
669 : }
670 : }
671 :
672 : /*
673 : * Ensure VacuumFailsafeActive has been reset before vacuuming the
674 : * next relation.
675 : */
676 119305 : VacuumFailsafeActive = false;
677 : }
678 : }
679 42 : PG_FINALLY();
680 : {
681 108657 : in_vacuum = false;
682 108657 : VacuumCostActive = false;
683 108657 : VacuumFailsafeActive = false;
684 108657 : VacuumCostBalance = 0;
685 : }
686 108657 : PG_END_TRY();
687 :
688 : /*
689 : * Finish up processing.
690 : */
691 108615 : if (use_own_xacts)
692 : {
693 : /* here, we are not in a transaction */
694 :
695 : /*
696 : * This matches the CommitTransaction waiting for us in
697 : * PostgresMain().
698 : */
699 106297 : StartTransactionCommand();
700 : }
701 :
702 108615 : if ((params->options & VACOPT_VACUUM) &&
703 105612 : !(params->options & VACOPT_SKIP_DATABASE_STATS))
704 : {
705 : /*
706 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible.
707 : */
708 1117 : vac_update_datfrozenxid();
709 : }
710 :
711 108615 : }
712 :
713 : /*
714 : * Check if the current user has privileges to vacuum or analyze the relation.
715 : * If not, issue a WARNING log message and return false to let the caller
716 : * decide what to do with this relation. This routine is used to decide if a
717 : * relation can be processed for VACUUM or ANALYZE.
718 : */
719 : bool
720 147231 : vacuum_is_permitted_for_relation(Oid relid, Form_pg_class reltuple,
721 : uint32 options)
722 : {
723 : char *relname;
724 :
725 : Assert((options & (VACOPT_VACUUM | VACOPT_ANALYZE)) != 0);
726 :
727 : /*----------
728 : * A role has privileges to vacuum or analyze the relation if any of the
729 : * following are true:
730 : * - the role owns the current database and the relation is not shared
731 : * - the role has the MAINTAIN privilege on the relation
732 : *----------
733 : */
734 147231 : if ((object_ownercheck(DatabaseRelationId, MyDatabaseId, GetUserId()) &&
735 168499 : !reltuple->relisshared) ||
736 23752 : pg_class_aclcheck(relid, GetUserId(), ACL_MAINTAIN) == ACLCHECK_OK)
737 145059 : return true;
738 :
739 2172 : relname = NameStr(reltuple->relname);
740 :
741 2172 : if ((options & VACOPT_VACUUM) != 0)
742 : {
743 148 : ereport(WARNING,
744 : (errmsg("permission denied to vacuum \"%s\", skipping it",
745 : relname)));
746 :
747 : /*
748 : * For VACUUM ANALYZE, both logs could show up, but just generate
749 : * information for VACUUM as that would be the first one to be
750 : * processed.
751 : */
752 148 : return false;
753 : }
754 :
755 2024 : if ((options & VACOPT_ANALYZE) != 0)
756 2024 : ereport(WARNING,
757 : (errmsg("permission denied to analyze \"%s\", skipping it",
758 : relname)));
759 :
760 2024 : return false;
761 : }
762 :
763 :
764 : /*
765 : * vacuum_open_relation
766 : *
767 : * This routine is used for attempting to open and lock a relation which
768 : * is going to be vacuumed or analyzed. If the relation cannot be opened
769 : * or locked, a log is emitted if possible.
770 : */
771 : Relation
772 127201 : vacuum_open_relation(Oid relid, RangeVar *relation, uint32 options,
773 : bool verbose, LOCKMODE lmode)
774 : {
775 : Relation rel;
776 127201 : bool rel_lock = true;
777 : int elevel;
778 :
779 : Assert((options & (VACOPT_VACUUM | VACOPT_ANALYZE)) != 0);
780 :
781 : /*
782 : * Open the relation and get the appropriate lock on it.
783 : *
784 : * There's a race condition here: the relation may have gone away since
785 : * the last time we saw it. If so, we don't need to vacuum or analyze it.
786 : *
787 : * If we've been asked not to wait for the relation lock, acquire it first
788 : * in non-blocking mode, before calling try_relation_open().
789 : */
790 127201 : if (!(options & VACOPT_SKIP_LOCKED))
791 126585 : rel = try_relation_open(relid, lmode);
792 616 : else if (ConditionalLockRelationOid(relid, lmode))
793 606 : rel = try_relation_open(relid, NoLock);
794 : else
795 : {
796 10 : rel = NULL;
797 10 : rel_lock = false;
798 : }
799 :
800 : /* if relation is opened, leave */
801 127201 : if (rel)
802 127185 : return rel;
803 :
804 : /*
805 : * Relation could not be opened, hence generate if possible a log
806 : * informing on the situation.
807 : *
808 : * If the RangeVar is not defined, we do not have enough information to
809 : * provide a meaningful log statement. Chances are that the caller has
810 : * intentionally not provided this information so that this logging is
811 : * skipped, anyway.
812 : */
813 16 : if (relation == NULL)
814 9 : return NULL;
815 :
816 : /*
817 : * Determine the log level.
818 : *
819 : * For manual VACUUM or ANALYZE, we emit a WARNING to match the log
820 : * statements in the permission checks; otherwise, only log if the caller
821 : * so requested.
822 : */
823 7 : if (!AmAutoVacuumWorkerProcess())
824 7 : elevel = WARNING;
825 0 : else if (verbose)
826 0 : elevel = LOG;
827 : else
828 0 : return NULL;
829 :
830 7 : if ((options & VACOPT_VACUUM) != 0)
831 : {
832 5 : if (!rel_lock)
833 3 : ereport(elevel,
834 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
835 : errmsg("skipping vacuum of \"%s\" --- lock not available",
836 : relation->relname)));
837 : else
838 2 : ereport(elevel,
839 : (errcode(ERRCODE_UNDEFINED_TABLE),
840 : errmsg("skipping vacuum of \"%s\" --- relation no longer exists",
841 : relation->relname)));
842 :
843 : /*
844 : * For VACUUM ANALYZE, both logs could show up, but just generate
845 : * information for VACUUM as that would be the first one to be
846 : * processed.
847 : */
848 5 : return NULL;
849 : }
850 :
851 2 : if ((options & VACOPT_ANALYZE) != 0)
852 : {
853 2 : if (!rel_lock)
854 1 : ereport(elevel,
855 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
856 : errmsg("skipping analyze of \"%s\" --- lock not available",
857 : relation->relname)));
858 : else
859 1 : ereport(elevel,
860 : (errcode(ERRCODE_UNDEFINED_TABLE),
861 : errmsg("skipping analyze of \"%s\" --- relation no longer exists",
862 : relation->relname)));
863 : }
864 :
865 2 : return NULL;
866 : }
867 :
868 :
869 : /*
870 : * Given a VacuumRelation, fill in the table OID if it wasn't specified,
871 : * and optionally add VacuumRelations for partitions or inheritance children.
872 : *
873 : * If a VacuumRelation does not have an OID supplied and is a partitioned
874 : * table, an extra entry will be added to the output for each partition.
875 : * Presently, only autovacuum supplies OIDs when calling vacuum(), and
876 : * it does not want us to expand partitioned tables.
877 : *
878 : * We take care not to modify the input data structure, but instead build
879 : * new VacuumRelation(s) to return. (But note that they will reference
880 : * unmodified parts of the input, eg column lists.) New data structures
881 : * are made in vac_context.
882 : */
883 : static List *
884 108586 : expand_vacuum_rel(VacuumRelation *vrel, MemoryContext vac_context,
885 : int options)
886 : {
887 108586 : List *vacrels = NIL;
888 : MemoryContext oldcontext;
889 :
890 : /* If caller supplied OID, there's nothing we need do here. */
891 108586 : if (OidIsValid(vrel->oid))
892 : {
893 100042 : oldcontext = MemoryContextSwitchTo(vac_context);
894 100042 : vacrels = lappend(vacrels, vrel);
895 100042 : MemoryContextSwitchTo(oldcontext);
896 : }
897 : else
898 : {
899 : /*
900 : * Process a specific relation, and possibly partitions or child
901 : * tables thereof.
902 : */
903 : Oid relid;
904 : HeapTuple tuple;
905 : Form_pg_class classForm;
906 : bool include_children;
907 : bool is_partitioned_table;
908 : int rvr_opts;
909 :
910 : /*
911 : * Since autovacuum workers supply OIDs when calling vacuum(), no
912 : * autovacuum worker should reach this code.
913 : */
914 : Assert(!AmAutoVacuumWorkerProcess());
915 :
916 : /*
917 : * We transiently take AccessShareLock to protect the syscache lookup
918 : * below, as well as find_all_inheritors's expectation that the caller
919 : * holds some lock on the starting relation.
920 : */
921 8544 : rvr_opts = (options & VACOPT_SKIP_LOCKED) ? RVR_SKIP_LOCKED : 0;
922 8544 : relid = RangeVarGetRelidExtended(vrel->relation,
923 : AccessShareLock,
924 : rvr_opts,
925 : NULL, NULL);
926 :
927 : /*
928 : * If the lock is unavailable, emit the same log statement that
929 : * vacuum_rel() and analyze_rel() would.
930 : */
931 8520 : if (!OidIsValid(relid))
932 : {
933 4 : if (options & VACOPT_VACUUM)
934 3 : ereport(WARNING,
935 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
936 : errmsg("skipping vacuum of \"%s\" --- lock not available",
937 : vrel->relation->relname)));
938 : else
939 1 : ereport(WARNING,
940 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
941 : errmsg("skipping analyze of \"%s\" --- lock not available",
942 : vrel->relation->relname)));
943 4 : return vacrels;
944 : }
945 :
946 : /*
947 : * To check whether the relation is a partitioned table and its
948 : * ownership, fetch its syscache entry.
949 : */
950 8516 : tuple = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
951 8516 : if (!HeapTupleIsValid(tuple))
952 0 : elog(ERROR, "cache lookup failed for relation %u", relid);
953 8516 : classForm = (Form_pg_class) GETSTRUCT(tuple);
954 :
955 : /*
956 : * Make a returnable VacuumRelation for this rel if the user has the
957 : * required privileges.
958 : */
959 8516 : if (vacuum_is_permitted_for_relation(relid, classForm, options))
960 : {
961 8364 : oldcontext = MemoryContextSwitchTo(vac_context);
962 8364 : vacrels = lappend(vacrels, makeVacuumRelation(vrel->relation,
963 : relid,
964 : vrel->va_cols));
965 8364 : MemoryContextSwitchTo(oldcontext);
966 : }
967 :
968 : /*
969 : * Vacuuming a partitioned table with ONLY will not do anything since
970 : * the partitioned table itself is empty. Issue a warning if the user
971 : * requests this.
972 : */
973 8516 : include_children = vrel->relation->inh;
974 8516 : is_partitioned_table = (classForm->relkind == RELKIND_PARTITIONED_TABLE);
975 8516 : if ((options & VACOPT_VACUUM) && is_partitioned_table && !include_children)
976 4 : ereport(WARNING,
977 : (errmsg("VACUUM ONLY of partitioned table \"%s\" has no effect",
978 : vrel->relation->relname)));
979 :
980 8516 : ReleaseSysCache(tuple);
981 :
982 : /*
983 : * Unless the user has specified ONLY, make relation list entries for
984 : * its partitions or inheritance child tables. Note that the list
985 : * returned by find_all_inheritors() includes the passed-in OID, so we
986 : * have to skip that. There's no point in taking locks on the
987 : * individual partitions or child tables yet, and doing so would just
988 : * add unnecessary deadlock risk. For this last reason, we do not yet
989 : * check the ownership of the partitions/tables, which get added to
990 : * the list to process. Ownership will be checked later on anyway.
991 : */
992 8516 : if (include_children)
993 : {
994 8496 : List *part_oids = find_all_inheritors(relid, NoLock, NULL);
995 : ListCell *part_lc;
996 :
997 18445 : foreach(part_lc, part_oids)
998 : {
999 9949 : Oid part_oid = lfirst_oid(part_lc);
1000 :
1001 9949 : if (part_oid == relid)
1002 8496 : continue; /* ignore original table */
1003 :
1004 : /*
1005 : * We omit a RangeVar since it wouldn't be appropriate to
1006 : * complain about failure to open one of these relations
1007 : * later.
1008 : */
1009 1453 : oldcontext = MemoryContextSwitchTo(vac_context);
1010 1453 : vacrels = lappend(vacrels, makeVacuumRelation(NULL,
1011 : part_oid,
1012 : vrel->va_cols));
1013 1453 : MemoryContextSwitchTo(oldcontext);
1014 : }
1015 : }
1016 :
1017 : /*
1018 : * Release lock again. This means that by the time we actually try to
1019 : * process the table, it might be gone or renamed. In the former case
1020 : * we'll silently ignore it; in the latter case we'll process it
1021 : * anyway, but we must beware that the RangeVar doesn't necessarily
1022 : * identify it anymore. This isn't ideal, perhaps, but there's little
1023 : * practical alternative, since we're typically going to commit this
1024 : * transaction and begin a new one between now and then. Moreover,
1025 : * holding locks on multiple relations would create significant risk
1026 : * of deadlock.
1027 : */
1028 8516 : UnlockRelationOid(relid, AccessShareLock);
1029 : }
1030 :
1031 108558 : return vacrels;
1032 : }
1033 :
1034 : /*
1035 : * Construct a list of VacuumRelations for all vacuumable rels in
1036 : * the current database. The list is built in vac_context.
1037 : */
1038 : static List *
1039 125 : get_all_vacuum_rels(MemoryContext vac_context, int options)
1040 : {
1041 125 : List *vacrels = NIL;
1042 : Relation pgclass;
1043 : TableScanDesc scan;
1044 : HeapTuple tuple;
1045 :
1046 125 : pgclass = table_open(RelationRelationId, AccessShareLock);
1047 :
1048 125 : scan = table_beginscan_catalog(pgclass, 0, NULL);
1049 :
1050 62238 : while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
1051 : {
1052 62113 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
1053 : MemoryContext oldcontext;
1054 62113 : Oid relid = classForm->oid;
1055 :
1056 : /*
1057 : * We include partitioned tables here; depending on which operation is
1058 : * to be performed, caller will decide whether to process or ignore
1059 : * them.
1060 : */
1061 62113 : if (classForm->relkind != RELKIND_RELATION &&
1062 50728 : classForm->relkind != RELKIND_MATVIEW &&
1063 50696 : classForm->relkind != RELKIND_PARTITIONED_TABLE)
1064 50583 : continue;
1065 :
1066 : /* check permissions of relation */
1067 11530 : if (!vacuum_is_permitted_for_relation(relid, classForm, options))
1068 1948 : continue;
1069 :
1070 : /*
1071 : * Build VacuumRelation(s) specifying the table OIDs to be processed.
1072 : * We omit a RangeVar since it wouldn't be appropriate to complain
1073 : * about failure to open one of these relations later.
1074 : */
1075 9582 : oldcontext = MemoryContextSwitchTo(vac_context);
1076 9582 : vacrels = lappend(vacrels, makeVacuumRelation(NULL,
1077 : relid,
1078 : NIL));
1079 9582 : MemoryContextSwitchTo(oldcontext);
1080 : }
1081 :
1082 125 : table_endscan(scan);
1083 125 : table_close(pgclass, AccessShareLock);
1084 :
1085 125 : return vacrels;
1086 : }
1087 :
1088 : /*
1089 : * vacuum_get_cutoffs() -- compute OldestXmin and freeze cutoff points
1090 : *
1091 : * The target relation and VACUUM parameters are our inputs.
1092 : *
1093 : * Output parameters are the cutoffs that VACUUM caller should use.
1094 : *
1095 : * Return value indicates if vacuumlazy.c caller should make its VACUUM
1096 : * operation aggressive. An aggressive VACUUM must advance relfrozenxid up to
1097 : * FreezeLimit (at a minimum), and relminmxid up to MultiXactCutoff (at a
1098 : * minimum).
1099 : */
1100 : bool
1101 116873 : vacuum_get_cutoffs(Relation rel, const VacuumParams *params,
1102 : struct VacuumCutoffs *cutoffs)
1103 : {
1104 : int freeze_min_age,
1105 : multixact_freeze_min_age,
1106 : freeze_table_age,
1107 : multixact_freeze_table_age,
1108 : effective_multixact_freeze_max_age;
1109 : TransactionId nextXID,
1110 : safeOldestXmin,
1111 : aggressiveXIDCutoff;
1112 : MultiXactId nextMXID,
1113 : safeOldestMxact,
1114 : aggressiveMXIDCutoff;
1115 :
1116 : /* Use mutable copies of freeze age parameters */
1117 116873 : freeze_min_age = params->freeze_min_age;
1118 116873 : multixact_freeze_min_age = params->multixact_freeze_min_age;
1119 116873 : freeze_table_age = params->freeze_table_age;
1120 116873 : multixact_freeze_table_age = params->multixact_freeze_table_age;
1121 :
1122 : /* Set pg_class fields in cutoffs */
1123 116873 : cutoffs->relfrozenxid = rel->rd_rel->relfrozenxid;
1124 116873 : cutoffs->relminmxid = rel->rd_rel->relminmxid;
1125 :
1126 : /*
1127 : * Acquire OldestXmin.
1128 : *
1129 : * We can always ignore processes running lazy vacuum. This is because we
1130 : * use these values only for deciding which tuples we must keep in the
1131 : * tables. Since lazy vacuum doesn't write its XID anywhere (usually no
1132 : * XID assigned), it's safe to ignore it. In theory it could be
1133 : * problematic to ignore lazy vacuums in a full vacuum, but keep in mind
1134 : * that only one vacuum process can be working on a particular table at
1135 : * any time, and that each vacuum is always an independent transaction.
1136 : */
1137 116873 : cutoffs->OldestXmin = GetOldestNonRemovableTransactionId(rel);
1138 :
1139 : Assert(TransactionIdIsNormal(cutoffs->OldestXmin));
1140 :
1141 : /* Acquire OldestMxact */
1142 116873 : cutoffs->OldestMxact = GetOldestMultiXactId();
1143 : Assert(MultiXactIdIsValid(cutoffs->OldestMxact));
1144 :
1145 : /* Acquire next XID/next MXID values used to apply age-based settings */
1146 116873 : nextXID = ReadNextTransactionId();
1147 116873 : nextMXID = ReadNextMultiXactId();
1148 :
1149 : /*
1150 : * Also compute the multixact age for which freezing is urgent. This is
1151 : * normally autovacuum_multixact_freeze_max_age, but may be less if
1152 : * multixact members are bloated.
1153 : */
1154 116873 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1155 :
1156 : /*
1157 : * Almost ready to set freeze output parameters; check if OldestXmin or
1158 : * OldestMxact are held back to an unsafe degree before we start on that
1159 : */
1160 116873 : safeOldestXmin = nextXID - autovacuum_freeze_max_age;
1161 116873 : if (!TransactionIdIsNormal(safeOldestXmin))
1162 0 : safeOldestXmin = FirstNormalTransactionId;
1163 116873 : safeOldestMxact = nextMXID - effective_multixact_freeze_max_age;
1164 116873 : if (safeOldestMxact < FirstMultiXactId)
1165 0 : safeOldestMxact = FirstMultiXactId;
1166 116873 : if (TransactionIdPrecedes(cutoffs->OldestXmin, safeOldestXmin))
1167 81239 : ereport(WARNING,
1168 : (errmsg("cutoff for removing and freezing tuples is far in the past"),
1169 : errhint("Close open transactions soon to avoid wraparound problems.\n"
1170 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1171 116873 : if (MultiXactIdPrecedes(cutoffs->OldestMxact, safeOldestMxact))
1172 0 : ereport(WARNING,
1173 : (errmsg("cutoff for freezing multixacts is far in the past"),
1174 : errhint("Close open transactions soon to avoid wraparound problems.\n"
1175 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1176 :
1177 : /*
1178 : * Determine the minimum freeze age to use: as specified by the caller, or
1179 : * vacuum_freeze_min_age, but in any case not more than half
1180 : * autovacuum_freeze_max_age, so that autovacuums to prevent XID
1181 : * wraparound won't occur too frequently.
1182 : */
1183 116873 : if (freeze_min_age < 0)
1184 7029 : freeze_min_age = vacuum_freeze_min_age;
1185 116873 : freeze_min_age = Min(freeze_min_age, autovacuum_freeze_max_age / 2);
1186 : Assert(freeze_min_age >= 0);
1187 :
1188 : /* Compute FreezeLimit, being careful to generate a normal XID */
1189 116873 : cutoffs->FreezeLimit = nextXID - freeze_min_age;
1190 116873 : if (!TransactionIdIsNormal(cutoffs->FreezeLimit))
1191 0 : cutoffs->FreezeLimit = FirstNormalTransactionId;
1192 : /* FreezeLimit must always be <= OldestXmin */
1193 116873 : if (TransactionIdPrecedes(cutoffs->OldestXmin, cutoffs->FreezeLimit))
1194 94068 : cutoffs->FreezeLimit = cutoffs->OldestXmin;
1195 :
1196 : /*
1197 : * Determine the minimum multixact freeze age to use: as specified by
1198 : * caller, or vacuum_multixact_freeze_min_age, but in any case not more
1199 : * than half effective_multixact_freeze_max_age, so that autovacuums to
1200 : * prevent MultiXact wraparound won't occur too frequently.
1201 : */
1202 116873 : if (multixact_freeze_min_age < 0)
1203 7029 : multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1204 116873 : multixact_freeze_min_age = Min(multixact_freeze_min_age,
1205 : effective_multixact_freeze_max_age / 2);
1206 : Assert(multixact_freeze_min_age >= 0);
1207 :
1208 : /* Compute MultiXactCutoff, being careful to generate a valid value */
1209 116873 : cutoffs->MultiXactCutoff = nextMXID - multixact_freeze_min_age;
1210 116873 : if (cutoffs->MultiXactCutoff < FirstMultiXactId)
1211 0 : cutoffs->MultiXactCutoff = FirstMultiXactId;
1212 : /* MultiXactCutoff must always be <= OldestMxact */
1213 116873 : if (MultiXactIdPrecedes(cutoffs->OldestMxact, cutoffs->MultiXactCutoff))
1214 2 : cutoffs->MultiXactCutoff = cutoffs->OldestMxact;
1215 :
1216 : /*
1217 : * Finally, figure out if caller needs to do an aggressive VACUUM or not.
1218 : *
1219 : * Determine the table freeze age to use: as specified by the caller, or
1220 : * the value of the vacuum_freeze_table_age GUC, but in any case not more
1221 : * than autovacuum_freeze_max_age * 0.95, so that if you have e.g nightly
1222 : * VACUUM schedule, the nightly VACUUM gets a chance to freeze XIDs before
1223 : * anti-wraparound autovacuum is launched.
1224 : */
1225 116873 : if (freeze_table_age < 0)
1226 7029 : freeze_table_age = vacuum_freeze_table_age;
1227 116873 : freeze_table_age = Min(freeze_table_age, autovacuum_freeze_max_age * 0.95);
1228 : Assert(freeze_table_age >= 0);
1229 116873 : aggressiveXIDCutoff = nextXID - freeze_table_age;
1230 116873 : if (!TransactionIdIsNormal(aggressiveXIDCutoff))
1231 0 : aggressiveXIDCutoff = FirstNormalTransactionId;
1232 116873 : if (TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid,
1233 : aggressiveXIDCutoff))
1234 109768 : return true;
1235 :
1236 : /*
1237 : * Similar to the above, determine the table freeze age to use for
1238 : * multixacts: as specified by the caller, or the value of the
1239 : * vacuum_multixact_freeze_table_age GUC, but in any case not more than
1240 : * effective_multixact_freeze_max_age * 0.95, so that if you have e.g.
1241 : * nightly VACUUM schedule, the nightly VACUUM gets a chance to freeze
1242 : * multixacts before anti-wraparound autovacuum is launched.
1243 : */
1244 7105 : if (multixact_freeze_table_age < 0)
1245 7029 : multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1246 7105 : multixact_freeze_table_age =
1247 7105 : Min(multixact_freeze_table_age,
1248 : effective_multixact_freeze_max_age * 0.95);
1249 : Assert(multixact_freeze_table_age >= 0);
1250 7105 : aggressiveMXIDCutoff = nextMXID - multixact_freeze_table_age;
1251 7105 : if (aggressiveMXIDCutoff < FirstMultiXactId)
1252 0 : aggressiveMXIDCutoff = FirstMultiXactId;
1253 7105 : if (MultiXactIdPrecedesOrEquals(cutoffs->relminmxid,
1254 : aggressiveMXIDCutoff))
1255 0 : return true;
1256 :
1257 : /* Non-aggressive VACUUM */
1258 7105 : return false;
1259 : }
1260 :
1261 : /*
1262 : * vacuum_xid_failsafe_check() -- Used by VACUUM's wraparound failsafe
1263 : * mechanism to determine if its table's relfrozenxid and relminmxid are now
1264 : * dangerously far in the past.
1265 : *
1266 : * When we return true, VACUUM caller triggers the failsafe.
1267 : */
1268 : bool
1269 118702 : vacuum_xid_failsafe_check(const struct VacuumCutoffs *cutoffs)
1270 : {
1271 118702 : TransactionId relfrozenxid = cutoffs->relfrozenxid;
1272 118702 : MultiXactId relminmxid = cutoffs->relminmxid;
1273 : TransactionId xid_skip_limit;
1274 : MultiXactId multi_skip_limit;
1275 : int skip_index_vacuum;
1276 :
1277 : Assert(TransactionIdIsNormal(relfrozenxid));
1278 : Assert(MultiXactIdIsValid(relminmxid));
1279 :
1280 : /*
1281 : * Determine the index skipping age to use. In any case no less than
1282 : * autovacuum_freeze_max_age * 1.05.
1283 : */
1284 118702 : skip_index_vacuum = Max(vacuum_failsafe_age, autovacuum_freeze_max_age * 1.05);
1285 :
1286 118702 : xid_skip_limit = ReadNextTransactionId() - skip_index_vacuum;
1287 118702 : if (!TransactionIdIsNormal(xid_skip_limit))
1288 0 : xid_skip_limit = FirstNormalTransactionId;
1289 :
1290 118702 : if (TransactionIdPrecedes(relfrozenxid, xid_skip_limit))
1291 : {
1292 : /* The table's relfrozenxid is too old */
1293 23684 : return true;
1294 : }
1295 :
1296 : /*
1297 : * Similar to above, determine the index skipping age to use for
1298 : * multixact. In any case no less than autovacuum_multixact_freeze_max_age *
1299 : * 1.05.
1300 : */
1301 95018 : skip_index_vacuum = Max(vacuum_multixact_failsafe_age,
1302 : autovacuum_multixact_freeze_max_age * 1.05);
1303 :
1304 95018 : multi_skip_limit = ReadNextMultiXactId() - skip_index_vacuum;
1305 95018 : if (multi_skip_limit < FirstMultiXactId)
1306 0 : multi_skip_limit = FirstMultiXactId;
1307 :
1308 95018 : if (MultiXactIdPrecedes(relminmxid, multi_skip_limit))
1309 : {
1310 : /* The table's relminmxid is too old */
1311 0 : return true;
1312 : }
1313 :
1314 95018 : return false;
1315 : }
1316 :
1317 : /*
1318 : * vac_estimate_reltuples() -- estimate the new value for pg_class.reltuples
1319 : *
1320 : * If we scanned the whole relation then we should just use the count of
1321 : * live tuples seen; but if we did not, we should not blindly extrapolate
1322 : * from that number, since VACUUM may have scanned a quite nonrandom
1323 : * subset of the table. When we have only partial information, we take
1324 : * the old value of pg_class.reltuples/pg_class.relpages as a measurement
1325 : * of the tuple density in the unscanned pages.
1326 : *
1327 : * Note: scanned_tuples should count only *live* tuples, since
1328 : * pg_class.reltuples is defined that way.
1329 : */
1330 : double
1331 116484 : vac_estimate_reltuples(Relation relation,
1332 : BlockNumber total_pages,
1333 : BlockNumber scanned_pages,
1334 : double scanned_tuples)
1335 : {
1336 116484 : BlockNumber old_rel_pages = relation->rd_rel->relpages;
1337 116484 : double old_rel_tuples = relation->rd_rel->reltuples;
1338 : double old_density;
1339 : double unscanned_pages;
1340 : double total_tuples;
1341 :
1342 : /* If we did scan the whole table, just use the count as-is */
1343 116484 : if (scanned_pages >= total_pages)
1344 112641 : return scanned_tuples;
1345 :
1346 : /*
1347 : * When successive VACUUM commands scan the same few pages again and
1348 : * again, without anything from the table really changing, there is a risk
1349 : * that our beliefs about tuple density will gradually become distorted.
1350 : * This might be caused by vacuumlazy.c implementation details, such as
1351 : * its tendency to always scan the last heap page. Handle that here.
1352 : *
1353 : * If the relation is _exactly_ the same size according to the existing
1354 : * pg_class entry, and only a few of its pages (less than 2%) were
1355 : * scanned, keep the existing value of reltuples. Also keep the existing
1356 : * value when only a subset of rel's pages <= a single page were scanned.
1357 : *
1358 : * (Note: we might be returning -1 here.)
1359 : */
1360 3843 : if (old_rel_pages == total_pages &&
1361 3821 : scanned_pages < (double) total_pages * 0.02)
1362 2776 : return old_rel_tuples;
1363 1067 : if (scanned_pages <= 1)
1364 955 : return old_rel_tuples;
1365 :
1366 : /*
1367 : * If old density is unknown, we can't do much except scale up
1368 : * scanned_tuples to match total_pages.
1369 : */
1370 112 : if (old_rel_tuples < 0 || old_rel_pages == 0)
1371 4 : return floor((scanned_tuples / scanned_pages) * total_pages + 0.5);
1372 :
1373 : /*
1374 : * Okay, we've covered the corner cases. The normal calculation is to
1375 : * convert the old measurement to a density (tuples per page), then
1376 : * estimate the number of tuples in the unscanned pages using that figure,
1377 : * and finally add on the number of tuples in the scanned pages.
1378 : */
1379 108 : old_density = old_rel_tuples / old_rel_pages;
1380 108 : unscanned_pages = (double) total_pages - (double) scanned_pages;
1381 108 : total_tuples = old_density * unscanned_pages + scanned_tuples;
1382 108 : return floor(total_tuples + 0.5);
1383 : }
1384 :
1385 :
1386 : /*
1387 : * vac_update_relstats() -- update statistics for one relation
1388 : *
1389 : * Update the whole-relation statistics that are kept in its pg_class
1390 : * row. There are additional stats that will be updated if we are
1391 : * doing ANALYZE, but we always update these stats. This routine works
1392 : * for both index and heap relation entries in pg_class.
1393 : *
1394 : * We violate transaction semantics here by overwriting the rel's
1395 : * existing pg_class tuple with the new values. This is reasonably
1396 : * safe as long as we're sure that the new values are correct whether or
1397 : * not this transaction commits. The reason for doing this is that if
1398 : * we updated these tuples in the usual way, vacuuming pg_class itself
1399 : * wouldn't work very well --- by the time we got done with a vacuum
1400 : * cycle, most of the tuples in pg_class would've been obsoleted. Of
1401 : * course, this only works for fixed-size not-null columns, but these are.
1402 : *
1403 : * Another reason for doing it this way is that when we are in a lazy
1404 : * VACUUM and have PROC_IN_VACUUM set, we mustn't do any regular updates.
1405 : * Somebody vacuuming pg_class might think they could delete a tuple
1406 : * marked with xmin = our xid.
1407 : *
1408 : * In addition to fundamentally nontransactional statistics such as
1409 : * relpages and relallvisible, we try to maintain certain lazily-updated
1410 : * DDL flags such as relhasindex, by clearing them if no longer correct.
1411 : * It's safe to do this in VACUUM, which can't run in parallel with
1412 : * CREATE INDEX/RULE/TRIGGER and can't be part of a transaction block.
1413 : * However, it's *not* safe to do it in an ANALYZE that's within an
1414 : * outer transaction, because for example the current transaction might
1415 : * have dropped the last index; then we'd think relhasindex should be
1416 : * cleared, but if the transaction later rolls back this would be wrong.
1417 : * So we refrain from updating the DDL flags if we're inside an outer
1418 : * transaction. This is OK since postponing the flag maintenance is
1419 : * always allowable.
1420 : *
1421 : * Note: num_tuples should count only *live* tuples, since
1422 : * pg_class.reltuples is defined that way.
1423 : *
1424 : * This routine is shared by VACUUM and ANALYZE.
1425 : */
1426 : void
1427 142163 : vac_update_relstats(Relation relation,
1428 : BlockNumber num_pages, double num_tuples,
1429 : BlockNumber num_all_visible_pages,
1430 : BlockNumber num_all_frozen_pages,
1431 : bool hasindex, TransactionId frozenxid,
1432 : MultiXactId minmulti,
1433 : bool *frozenxid_updated, bool *minmulti_updated,
1434 : bool in_outer_xact)
1435 : {
1436 142163 : Oid relid = RelationGetRelid(relation);
1437 : Relation rd;
1438 : ScanKeyData key[1];
1439 : HeapTuple ctup;
1440 : void *inplace_state;
1441 : Form_pg_class pgcform;
1442 : bool dirty,
1443 : futurexid,
1444 : futuremxid;
1445 : TransactionId oldfrozenxid;
1446 : MultiXactId oldminmulti;
1447 :
1448 142163 : rd = table_open(RelationRelationId, RowExclusiveLock);
1449 :
1450 : /* Fetch a copy of the tuple to scribble on */
1451 142163 : ScanKeyInit(&key[0],
1452 : Anum_pg_class_oid,
1453 : BTEqualStrategyNumber, F_OIDEQ,
1454 : ObjectIdGetDatum(relid));
1455 142163 : systable_inplace_update_begin(rd, ClassOidIndexId, true,
1456 : NULL, 1, key, &ctup, &inplace_state);
1457 142163 : if (!HeapTupleIsValid(ctup))
1458 0 : elog(ERROR, "pg_class entry for relid %u vanished during vacuuming",
1459 : relid);
1460 142163 : pgcform = (Form_pg_class) GETSTRUCT(ctup);
1461 :
1462 : /* Apply statistical updates, if any, to copied tuple */
1463 :
1464 142163 : dirty = false;
1465 142163 : if (pgcform->relpages != (int32) num_pages)
1466 : {
1467 5634 : pgcform->relpages = (int32) num_pages;
1468 5634 : dirty = true;
1469 : }
1470 142163 : if (pgcform->reltuples != (float4) num_tuples)
1471 : {
1472 12353 : pgcform->reltuples = (float4) num_tuples;
1473 12353 : dirty = true;
1474 : }
1475 142163 : if (pgcform->relallvisible != (int32) num_all_visible_pages)
1476 : {
1477 3796 : pgcform->relallvisible = (int32) num_all_visible_pages;
1478 3796 : dirty = true;
1479 : }
1480 142163 : if (pgcform->relallfrozen != (int32) num_all_frozen_pages)
1481 : {
1482 3197 : pgcform->relallfrozen = (int32) num_all_frozen_pages;
1483 3197 : dirty = true;
1484 : }
1485 :
1486 : /* Apply DDL updates, but not inside an outer transaction (see above) */
1487 :
1488 142163 : if (!in_outer_xact)
1489 : {
1490 : /*
1491 : * If we didn't find any indexes, reset relhasindex.
1492 : */
1493 141881 : if (pgcform->relhasindex && !hasindex)
1494 : {
1495 15 : pgcform->relhasindex = false;
1496 15 : dirty = true;
1497 : }
1498 :
1499 : /* We also clear relhasrules and relhastriggers if needed */
1500 141881 : if (pgcform->relhasrules && relation->rd_rules == NULL)
1501 : {
1502 0 : pgcform->relhasrules = false;
1503 0 : dirty = true;
1504 : }
1505 141881 : if (pgcform->relhastriggers && relation->trigdesc == NULL)
1506 : {
1507 8 : pgcform->relhastriggers = false;
1508 8 : dirty = true;
1509 : }
1510 : }
1511 :
1512 : /*
1513 : * Update relfrozenxid, unless caller passed InvalidTransactionId
1514 : * indicating it has no new data.
1515 : *
1516 : * Ordinarily, we don't let relfrozenxid go backwards. However, if the
1517 : * stored relfrozenxid is "in the future" then it seems best to assume
1518 : * it's corrupt, and overwrite with the oldest remaining XID in the table.
1519 : * This should match vac_update_datfrozenxid() concerning what we consider
1520 : * to be "in the future".
1521 : */
1522 142163 : oldfrozenxid = pgcform->relfrozenxid;
1523 142163 : futurexid = false;
1524 142163 : if (frozenxid_updated)
1525 116482 : *frozenxid_updated = false;
1526 142163 : if (TransactionIdIsNormal(frozenxid) && oldfrozenxid != frozenxid)
1527 : {
1528 33867 : bool update = false;
1529 :
1530 33867 : if (TransactionIdPrecedes(oldfrozenxid, frozenxid))
1531 33792 : update = true;
1532 75 : else if (TransactionIdPrecedes(ReadNextTransactionId(), oldfrozenxid))
1533 0 : futurexid = update = true;
1534 :
1535 33867 : if (update)
1536 : {
1537 33792 : pgcform->relfrozenxid = frozenxid;
1538 33792 : dirty = true;
1539 33792 : if (frozenxid_updated)
1540 33792 : *frozenxid_updated = true;
1541 : }
1542 : }
1543 :
1544 : /* Similarly for relminmxid */
1545 142163 : oldminmulti = pgcform->relminmxid;
1546 142163 : futuremxid = false;
1547 142163 : if (minmulti_updated)
1548 116482 : *minmulti_updated = false;
1549 142163 : if (MultiXactIdIsValid(minmulti) && oldminmulti != minmulti)
1550 : {
1551 213 : bool update = false;
1552 :
1553 213 : if (MultiXactIdPrecedes(oldminmulti, minmulti))
1554 213 : update = true;
1555 0 : else if (MultiXactIdPrecedes(ReadNextMultiXactId(), oldminmulti))
1556 0 : futuremxid = update = true;
1557 :
1558 213 : if (update)
1559 : {
1560 213 : pgcform->relminmxid = minmulti;
1561 213 : dirty = true;
1562 213 : if (minmulti_updated)
1563 213 : *minmulti_updated = true;
1564 : }
1565 : }
1566 :
1567 : /* If anything changed, write out the tuple. */
1568 142163 : if (dirty)
1569 42473 : systable_inplace_update_finish(inplace_state, ctup);
1570 : else
1571 99690 : systable_inplace_update_cancel(inplace_state);
1572 :
1573 142163 : table_close(rd, RowExclusiveLock);
1574 :
1575 142163 : if (futurexid)
1576 0 : ereport(WARNING,
1577 : (errcode(ERRCODE_DATA_CORRUPTED),
1578 : errmsg_internal("overwrote invalid relfrozenxid value %u with new value %u for table \"%s\"",
1579 : oldfrozenxid, frozenxid,
1580 : RelationGetRelationName(relation))));
1581 142163 : if (futuremxid)
1582 0 : ereport(WARNING,
1583 : (errcode(ERRCODE_DATA_CORRUPTED),
1584 : errmsg_internal("overwrote invalid relminmxid value %u with new value %u for table \"%s\"",
1585 : oldminmulti, minmulti,
1586 : RelationGetRelationName(relation))));
1587 142163 : }
1588 :
1589 :
1590 : /*
1591 : * vac_update_datfrozenxid() -- update pg_database.datfrozenxid for our DB
1592 : *
1593 : * Update pg_database's datfrozenxid entry for our database to be the
1594 : * minimum of the pg_class.relfrozenxid values.
1595 : *
1596 : * Similarly, update our datminmxid to be the minimum of the
1597 : * pg_class.relminmxid values.
1598 : *
1599 : * If we are able to advance either pg_database value, also try to
1600 : * truncate pg_xact and pg_multixact.
1601 : *
1602 : * We violate transaction semantics here by overwriting the database's
1603 : * existing pg_database tuple with the new values. This is reasonably
1604 : * safe since the new values are correct whether or not this transaction
1605 : * commits. As with vac_update_relstats, this avoids leaving dead tuples
1606 : * behind after a VACUUM.
1607 : */
1608 : void
1609 2531 : vac_update_datfrozenxid(void)
1610 : {
1611 : HeapTuple tuple;
1612 : Form_pg_database dbform;
1613 : Relation relation;
1614 : SysScanDesc scan;
1615 : HeapTuple classTup;
1616 : TransactionId newFrozenXid;
1617 : MultiXactId newMinMulti;
1618 : TransactionId lastSaneFrozenXid;
1619 : MultiXactId lastSaneMinMulti;
1620 2531 : bool bogus = false;
1621 2531 : bool dirty = false;
1622 : ScanKeyData key[1];
1623 : void *inplace_state;
1624 :
1625 : /*
1626 : * Restrict this task to one backend per database. This avoids race
1627 : * conditions that would move datfrozenxid or datminmxid backward. It
1628 : * avoids calling vac_truncate_clog() with a datfrozenxid preceding a
1629 : * datfrozenxid passed to an earlier vac_truncate_clog() call.
1630 : */
1631 2531 : LockDatabaseFrozenIds(ExclusiveLock);
1632 :
1633 : /*
1634 : * Initialize the "min" calculation with
1635 : * GetOldestNonRemovableTransactionId(), which is a reasonable
1636 : * approximation to the minimum relfrozenxid for not-yet-committed
1637 : * pg_class entries for new tables; see AddNewRelationTuple(). So we
1638 : * cannot produce a wrong minimum by starting with this.
1639 : */
1640 2531 : newFrozenXid = GetOldestNonRemovableTransactionId(NULL);
1641 :
1642 : /*
1643 : * Similarly, initialize the MultiXact "min" with the value that would be
1644 : * used on pg_class for new tables. See AddNewRelationTuple().
1645 : */
1646 2531 : newMinMulti = GetOldestMultiXactId();
1647 :
1648 : /*
1649 : * Identify the latest relfrozenxid and relminmxid values that we could
1650 : * validly see during the scan. These are conservative values, but it's
1651 : * not really worth trying to be more exact.
1652 : */
1653 2531 : lastSaneFrozenXid = ReadNextTransactionId();
1654 2531 : lastSaneMinMulti = ReadNextMultiXactId();
1655 :
1656 : /*
1657 : * We must seqscan pg_class to find the minimum Xid, because there is no
1658 : * index that can help us here.
1659 : *
1660 : * See vac_truncate_clog() for the race condition to prevent.
1661 : */
1662 2531 : relation = table_open(RelationRelationId, AccessShareLock);
1663 :
1664 2531 : scan = systable_beginscan(relation, InvalidOid, false,
1665 : NULL, 0, NULL);
1666 :
1667 1478387 : while ((classTup = systable_getnext(scan)) != NULL)
1668 : {
1669 1475856 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(classTup);
1670 1475856 : volatile TransactionId *relfrozenxid_p = &classForm->relfrozenxid;
1671 1475856 : volatile TransactionId *relminmxid_p = &classForm->relminmxid;
1672 1475856 : TransactionId relfrozenxid = *relfrozenxid_p;
1673 1475856 : TransactionId relminmxid = *relminmxid_p;
1674 :
1675 : /*
1676 : * Only consider relations able to hold unfrozen XIDs (anything else
1677 : * should have InvalidTransactionId in relfrozenxid anyway).
1678 : */
1679 1475856 : if (classForm->relkind != RELKIND_RELATION &&
1680 1162022 : classForm->relkind != RELKIND_MATVIEW &&
1681 1160420 : classForm->relkind != RELKIND_TOASTVALUE)
1682 : {
1683 : Assert(!TransactionIdIsValid(relfrozenxid));
1684 : Assert(!MultiXactIdIsValid(relminmxid));
1685 1000944 : continue;
1686 : }
1687 :
1688 : /*
1689 : * Some table AMs might not need per-relation xid / multixid horizons.
1690 : * It therefore seems reasonable to allow relfrozenxid and relminmxid
1691 : * to not be set (i.e. set to their respective Invalid*Id)
1692 : * independently. Thus validate and compute horizon for each only if
1693 : * set.
1694 : *
1695 : * If things are working properly, no relation should have a
1696 : * relfrozenxid or relminmxid that is "in the future". However, such
1697 : * cases have been known to arise due to bugs in pg_upgrade. If we
1698 : * see any entries that are "in the future", chicken out and don't do
1699 : * anything. This ensures we won't truncate clog & multixact SLRUs
1700 : * before those relations have been scanned and cleaned up.
1701 : */
1702 :
1703 474912 : if (TransactionIdIsValid(relfrozenxid))
1704 : {
1705 : Assert(TransactionIdIsNormal(relfrozenxid));
1706 :
1707 : /* check for values in the future */
1708 474912 : if (TransactionIdPrecedes(lastSaneFrozenXid, relfrozenxid))
1709 : {
1710 0 : bogus = true;
1711 0 : break;
1712 : }
1713 :
1714 : /* determine new horizon */
1715 474912 : if (TransactionIdPrecedes(relfrozenxid, newFrozenXid))
1716 2491 : newFrozenXid = relfrozenxid;
1717 : }
1718 :
1719 474912 : if (MultiXactIdIsValid(relminmxid))
1720 : {
1721 : /* check for values in the future */
1722 474912 : if (MultiXactIdPrecedes(lastSaneMinMulti, relminmxid))
1723 : {
1724 0 : bogus = true;
1725 0 : break;
1726 : }
1727 :
1728 : /* determine new horizon */
1729 474912 : if (MultiXactIdPrecedes(relminmxid, newMinMulti))
1730 208 : newMinMulti = relminmxid;
1731 : }
1732 : }
1733 :
1734 : /* we're done with pg_class */
1735 2531 : systable_endscan(scan);
1736 2531 : table_close(relation, AccessShareLock);
1737 :
1738 : /* chicken out if bogus data found */
1739 2531 : if (bogus)
1740 0 : return;
1741 :
1742 : Assert(TransactionIdIsNormal(newFrozenXid));
1743 : Assert(MultiXactIdIsValid(newMinMulti));
1744 :
1745 : /* Now fetch the pg_database tuple we need to update. */
1746 2531 : relation = table_open(DatabaseRelationId, RowExclusiveLock);
1747 :
1748 : /*
1749 : * Fetch a copy of the tuple to scribble on. We could check the syscache
1750 : * tuple first. If that concluded !dirty, we'd avoid waiting on
1751 : * concurrent heap_update() and would avoid exclusive-locking the buffer.
1752 : * For now, don't optimize that.
1753 : */
1754 2531 : ScanKeyInit(&key[0],
1755 : Anum_pg_database_oid,
1756 : BTEqualStrategyNumber, F_OIDEQ,
1757 : ObjectIdGetDatum(MyDatabaseId));
1758 :
1759 2531 : systable_inplace_update_begin(relation, DatabaseOidIndexId, true,
1760 : NULL, 1, key, &tuple, &inplace_state);
1761 :
1762 2531 : if (!HeapTupleIsValid(tuple))
1763 0 : elog(ERROR, "could not find tuple for database %u", MyDatabaseId);
1764 :
1765 2531 : dbform = (Form_pg_database) GETSTRUCT(tuple);
1766 :
1767 : /*
1768 : * As in vac_update_relstats(), we ordinarily don't want to let
1769 : * datfrozenxid go backward; but if it's "in the future" then it must be
1770 : * corrupt and it seems best to overwrite it.
1771 : */
1772 2870 : if (dbform->datfrozenxid != newFrozenXid &&
1773 339 : (TransactionIdPrecedes(dbform->datfrozenxid, newFrozenXid) ||
1774 0 : TransactionIdPrecedes(lastSaneFrozenXid, dbform->datfrozenxid)))
1775 : {
1776 339 : dbform->datfrozenxid = newFrozenXid;
1777 339 : dirty = true;
1778 : }
1779 : else
1780 2192 : newFrozenXid = dbform->datfrozenxid;
1781 :
1782 : /* Ditto for datminmxid */
1783 2532 : if (dbform->datminmxid != newMinMulti &&
1784 1 : (MultiXactIdPrecedes(dbform->datminmxid, newMinMulti) ||
1785 0 : MultiXactIdPrecedes(lastSaneMinMulti, dbform->datminmxid)))
1786 : {
1787 1 : dbform->datminmxid = newMinMulti;
1788 1 : dirty = true;
1789 : }
1790 : else
1791 2530 : newMinMulti = dbform->datminmxid;
1792 :
1793 2531 : if (dirty)
1794 339 : systable_inplace_update_finish(inplace_state, tuple);
1795 : else
1796 2192 : systable_inplace_update_cancel(inplace_state);
1797 :
1798 2531 : heap_freetuple(tuple);
1799 2531 : table_close(relation, RowExclusiveLock);
1800 :
1801 : /*
1802 : * If we were able to advance datfrozenxid or datminmxid, see if we can
1803 : * truncate pg_xact and/or pg_multixact. Also do it if the shared
1804 : * XID-wrap-limit info is stale, since this action will update that too.
1805 : */
1806 2531 : if (dirty || ForceTransactionIdLimitUpdate())
1807 1066 : vac_truncate_clog(newFrozenXid, newMinMulti,
1808 : lastSaneFrozenXid, lastSaneMinMulti);
1809 : }
1810 :
1811 :
1812 : /*
1813 : * vac_truncate_clog() -- attempt to truncate the commit log
1814 : *
1815 : * Scan pg_database to determine the system-wide oldest datfrozenxid,
1816 : * and use it to truncate the transaction commit log (pg_xact).
1817 : * Also update the XID wrap limit info maintained by varsup.c.
1818 : * Likewise for datminmxid.
1819 : *
1820 : * The passed frozenXID and minMulti are the updated values for my own
1821 : * pg_database entry. They're used to initialize the "min" calculations.
1822 : * The caller also passes the "last sane" XID and MXID, since it has
1823 : * those at hand already.
1824 : *
1825 : * This routine is only invoked when we've managed to change our
1826 : * DB's datfrozenxid/datminmxid values, or we found that the shared
1827 : * XID-wrap-limit info is stale.
1828 : */
1829 : static void
1830 1066 : vac_truncate_clog(TransactionId frozenXID,
1831 : MultiXactId minMulti,
1832 : TransactionId lastSaneFrozenXid,
1833 : MultiXactId lastSaneMinMulti)
1834 : {
1835 1066 : TransactionId nextXID = ReadNextTransactionId();
1836 : Relation relation;
1837 : TableScanDesc scan;
1838 : HeapTuple tuple;
1839 : Oid oldestxid_datoid;
1840 : Oid minmulti_datoid;
1841 1066 : bool bogus = false;
1842 1066 : bool frozenAlreadyWrapped = false;
1843 :
1844 : /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */
1845 1066 : LWLockAcquire(WrapLimitsVacuumLock, LW_EXCLUSIVE);
1846 :
1847 : /* init oldest datoids to sync with my frozenXID/minMulti values */
1848 1066 : oldestxid_datoid = MyDatabaseId;
1849 1066 : minmulti_datoid = MyDatabaseId;
1850 :
1851 : /*
1852 : * Scan pg_database to compute the minimum datfrozenxid/datminmxid
1853 : *
1854 : * Since vac_update_datfrozenxid updates datfrozenxid/datminmxid in-place,
1855 : * the values could change while we look at them. Fetch each one just
1856 : * once to ensure sane behavior of the comparison logic. (Here, as in
1857 : * many other places, we assume that fetching or updating an XID in shared
1858 : * storage is atomic.)
1859 : *
1860 : * Note: we need not worry about a race condition with new entries being
1861 : * inserted by CREATE DATABASE. Any such entry will have a copy of some
1862 : * existing DB's datfrozenxid, and that source DB cannot be ours because
1863 : * of the interlock against copying a DB containing an active backend.
1864 : * Hence the new entry will not reduce the minimum. Also, if two VACUUMs
1865 : * concurrently modify the datfrozenxid's of different databases, the
1866 : * worst possible outcome is that pg_xact is not truncated as aggressively
1867 : * as it could be.
1868 : */
1869 1066 : relation = table_open(DatabaseRelationId, AccessShareLock);
1870 :
1871 1066 : scan = table_beginscan_catalog(relation, 0, NULL);
1872 :
1873 4166 : while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
1874 : {
1875 3100 : Form_pg_database dbform = (Form_pg_database) GETSTRUCT(tuple);
1876 3100 : volatile TransactionId *datfrozenxid_p = &dbform->datfrozenxid;
1877 3100 : volatile TransactionId *datminmxid_p = &dbform->datminmxid;
1878 3100 : TransactionId datfrozenxid = *datfrozenxid_p;
1879 3100 : TransactionId datminmxid = *datminmxid_p;
1880 :
1881 : Assert(TransactionIdIsNormal(datfrozenxid));
1882 : Assert(MultiXactIdIsValid(datminmxid));
1883 :
1884 : /*
1885 : * If database is in the process of getting dropped, or has been
1886 : * interrupted while doing so, no connections to it are possible
1887 : * anymore. Therefore we don't need to take it into account here.
1888 : * Which is good, because it can't be processed by autovacuum either.
1889 : */
1890 3100 : if (database_is_invalid_form((Form_pg_database) dbform))
1891 : {
1892 3 : elog(DEBUG2,
1893 : "skipping invalid database \"%s\" while computing relfrozenxid",
1894 : NameStr(dbform->datname));
1895 3 : continue;
1896 : }
1897 :
1898 : /*
1899 : * If things are working properly, no database should have a
1900 : * datfrozenxid or datminmxid that is "in the future". However, such
1901 : * cases have been known to arise due to bugs in pg_upgrade. If we
1902 : * see any entries that are "in the future", chicken out and don't do
1903 : * anything. This ensures we won't truncate clog before those
1904 : * databases have been scanned and cleaned up. (We will issue the
1905 : * "already wrapped" warning if appropriate, though.)
1906 : */
1907 6194 : if (TransactionIdPrecedes(lastSaneFrozenXid, datfrozenxid) ||
1908 3097 : MultiXactIdPrecedes(lastSaneMinMulti, datminmxid))
1909 0 : bogus = true;
1910 :
1911 3097 : if (TransactionIdPrecedes(nextXID, datfrozenxid))
1912 0 : frozenAlreadyWrapped = true;
1913 3097 : else if (TransactionIdPrecedes(datfrozenxid, frozenXID))
1914 : {
1915 282 : frozenXID = datfrozenxid;
1916 282 : oldestxid_datoid = dbform->oid;
1917 : }
1918 :
1919 3097 : if (MultiXactIdPrecedes(datminmxid, minMulti))
1920 : {
1921 2 : minMulti = datminmxid;
1922 2 : minmulti_datoid = dbform->oid;
1923 : }
1924 : }
1925 :
1926 1066 : table_endscan(scan);
1927 :
1928 1066 : table_close(relation, AccessShareLock);
1929 :
1930 : /*
1931 : * Do not truncate CLOG if we seem to have suffered wraparound already;
1932 : * the computed minimum XID might be bogus. This case should now be
1933 : * impossible due to the defenses in GetNewTransactionId, but we keep the
1934 : * test anyway.
1935 : */
1936 1066 : if (frozenAlreadyWrapped)
1937 : {
1938 0 : ereport(WARNING,
1939 : (errmsg("some databases have not been vacuumed in over 2 billion transactions"),
1940 : errdetail("You might have already suffered transaction-wraparound data loss.")));
1941 0 : LWLockRelease(WrapLimitsVacuumLock);
1942 0 : return;
1943 : }
1944 :
1945 : /* chicken out if data is bogus in any other way */
1946 1066 : if (bogus)
1947 : {
1948 0 : LWLockRelease(WrapLimitsVacuumLock);
1949 0 : return;
1950 : }
1951 :
1952 : /*
1953 : * Freeze any old transaction IDs in the async notification queue before
1954 : * CLOG truncation.
1955 : */
1956 1066 : AsyncNotifyFreezeXids(frozenXID);
1957 :
1958 : /*
1959 : * Advance the oldest value for commit timestamps before truncating, so
1960 : * that if a user requests a timestamp for a transaction we're truncating
1961 : * away right after this point, they get NULL instead of an ugly "file not
1962 : * found" error from slru.c. This doesn't matter for xact/multixact
1963 : * because they are not subject to arbitrary lookups from users.
1964 : */
1965 1066 : AdvanceOldestCommitTsXid(frozenXID);
1966 :
1967 : /*
1968 : * Truncate CLOG, multixact and CommitTs to the oldest computed value.
1969 : */
1970 1066 : TruncateCLOG(frozenXID, oldestxid_datoid);
1971 1066 : TruncateCommitTs(frozenXID);
1972 1066 : TruncateMultiXact(minMulti, minmulti_datoid);
1973 :
1974 : /*
1975 : * Update the wrap limit for GetNewTransactionId and creation of new
1976 : * MultiXactIds. Note: these functions will also signal the postmaster
1977 : * for an(other) autovac cycle if needed. XXX should we avoid possibly
1978 : * signaling twice?
1979 : */
1980 1066 : SetTransactionIdLimit(frozenXID, oldestxid_datoid);
1981 1066 : SetMultiXactIdLimit(minMulti, minmulti_datoid);
1982 :
1983 1066 : LWLockRelease(WrapLimitsVacuumLock);
1984 : }
1985 :
1986 :
1987 : /*
1988 : * vacuum_rel() -- vacuum one heap relation
1989 : *
1990 : * relid identifies the relation to vacuum. If relation is supplied,
1991 : * use the name therein for reporting any failure to open/lock the rel;
1992 : * do not use it once we've successfully opened the rel, since it might
1993 : * be stale.
1994 : *
1995 : * Returns true if it's okay to proceed with a requested ANALYZE
1996 : * operation on this table.
1997 : *
1998 : * Doing one heap at a time incurs extra overhead, since we need to
1999 : * check that the heap exists again just before we vacuum it. The
2000 : * reason that we do this is so that vacuuming can be spread across
2001 : * many small transactions. Otherwise, two-phase locking would require
2002 : * us to lock the entire database during one pass of the vacuum cleaner.
2003 : *
2004 : * At entry and exit, we are not inside a transaction.
2005 : */
2006 : static bool
2007 116975 : vacuum_rel(Oid relid, RangeVar *relation, VacuumParams params,
2008 : BufferAccessStrategy bstrategy, bool isTopLevel)
2009 : {
2010 : LOCKMODE lmode;
2011 : Relation rel;
2012 : LockRelId lockrelid;
2013 : Oid priv_relid;
2014 : Oid toast_relid;
2015 : Oid save_userid;
2016 : int save_sec_context;
2017 : int save_nestlevel;
2018 : VacuumParams toast_vacuum_params;
2019 :
2020 : /*
2021 : * This function scribbles on the parameters, so make a copy early to
2022 : * avoid affecting the TOAST table (if we do end up recursing to it).
2023 : */
2024 116975 : memcpy(&toast_vacuum_params, ¶ms, sizeof(VacuumParams));
2025 :
2026 : /* Begin a transaction for vacuuming this relation */
2027 116975 : StartTransactionCommand();
2028 :
2029 116975 : if (!(params.options & VACOPT_FULL))
2030 : {
2031 : /*
2032 : * In lazy vacuum, we can set the PROC_IN_VACUUM flag, which lets
2033 : * other concurrent VACUUMs know that they can ignore this one while
2034 : * determining their OldestXmin. (The reason we don't set it during a
2035 : * full VACUUM is exactly that we may have to run user-defined
2036 : * functions for functional indexes, and we want to make sure that if
2037 : * they use the snapshot set above, any tuples it requires can't get
2038 : * removed from other tables. An index function that depends on the
2039 : * contents of other tables is arguably broken, but we won't break it
2040 : * here by violating transaction semantics.)
2041 : *
2042 : * We also set the VACUUM_FOR_WRAPAROUND flag, which is passed down by
2043 : * autovacuum; it's used to avoid canceling a vacuum that was invoked
2044 : * in an emergency.
2045 : *
2046 : * Note: these flags remain set until CommitTransaction or
2047 : * AbortTransaction. We don't want to clear them until we reset
2048 : * MyProc->xid/xmin, otherwise GetOldestNonRemovableTransactionId()
2049 : * might appear to go backwards, which is probably Not Good. (We also
2050 : * set PROC_IN_VACUUM *before* taking our own snapshot, so that our
2051 : * xmin doesn't become visible ahead of setting the flag.)
2052 : */
2053 116732 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2054 116732 : MyProc->statusFlags |= PROC_IN_VACUUM;
2055 116732 : if (params.is_wraparound)
2056 99761 : MyProc->statusFlags |= PROC_VACUUM_FOR_WRAPAROUND;
2057 116732 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2058 116732 : LWLockRelease(ProcArrayLock);
2059 : }
2060 :
2061 : /*
2062 : * Need to acquire a snapshot to prevent pg_subtrans from being truncated,
2063 : * cutoff xids in local memory wrapping around, and to have updated xmin
2064 : * horizons.
2065 : */
2066 116975 : PushActiveSnapshot(GetTransactionSnapshot());
2067 :
2068 : /*
2069 : * Check for user-requested abort. Note we want this to be inside a
2070 : * transaction, so xact.c doesn't issue useless WARNING.
2071 : */
2072 116975 : CHECK_FOR_INTERRUPTS();
2073 :
2074 : /*
2075 : * Determine the type of lock we want --- hard exclusive lock for a FULL
2076 : * vacuum, but just ShareUpdateExclusiveLock for concurrent vacuum. Either
2077 : * way, we can be sure that no other backend is vacuuming the same table.
2078 : */
2079 233950 : lmode = (params.options & VACOPT_FULL) ?
2080 116975 : AccessExclusiveLock : ShareUpdateExclusiveLock;
2081 :
2082 : /* open the relation and get the appropriate lock on it */
2083 116975 : rel = vacuum_open_relation(relid, relation, params.options,
2084 116975 : params.log_vacuum_min_duration >= 0, lmode);
2085 :
2086 : /* leave if relation could not be opened or locked */
2087 116975 : if (!rel)
2088 : {
2089 12 : PopActiveSnapshot();
2090 12 : CommitTransactionCommand();
2091 12 : return false;
2092 : }
2093 :
2094 : /*
2095 : * When recursing to a TOAST table, check privileges on the parent. NB:
2096 : * This is only safe to do because we hold a session lock on the main
2097 : * relation that prevents concurrent deletion.
2098 : */
2099 116963 : if (OidIsValid(params.toast_parent))
2100 5808 : priv_relid = params.toast_parent;
2101 : else
2102 111155 : priv_relid = RelationGetRelid(rel);
2103 :
2104 : /*
2105 : * Check if relation needs to be skipped based on privileges. This check
2106 : * happens also when building the relation list to vacuum for a manual
2107 : * operation, and needs to be done additionally here as VACUUM could
2108 : * happen across multiple transactions where privileges could have changed
2109 : * in-between. Make sure to only generate logs for VACUUM in this case.
2110 : */
2111 116963 : if (!vacuum_is_permitted_for_relation(priv_relid,
2112 : rel->rd_rel,
2113 116963 : params.options & ~VACOPT_ANALYZE))
2114 : {
2115 48 : relation_close(rel, lmode);
2116 48 : PopActiveSnapshot();
2117 48 : CommitTransactionCommand();
2118 48 : return false;
2119 : }
2120 :
2121 : /*
2122 : * Check that it's of a vacuumable relkind.
2123 : */
2124 116915 : if (rel->rd_rel->relkind != RELKIND_RELATION &&
2125 41933 : rel->rd_rel->relkind != RELKIND_MATVIEW &&
2126 41928 : rel->rd_rel->relkind != RELKIND_TOASTVALUE &&
2127 125 : rel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
2128 : {
2129 1 : ereport(WARNING,
2130 : (errmsg("skipping \"%s\" --- cannot vacuum non-tables or special system tables",
2131 : RelationGetRelationName(rel))));
2132 1 : relation_close(rel, lmode);
2133 1 : PopActiveSnapshot();
2134 1 : CommitTransactionCommand();
2135 1 : return false;
2136 : }
2137 :
2138 : /*
2139 : * Silently ignore tables that are temp tables of other backends ---
2140 : * trying to vacuum these will lead to great unhappiness, since their
2141 : * contents are probably not up-to-date on disk. (We don't throw a
2142 : * warning here; it would just lead to chatter during a database-wide
2143 : * VACUUM.)
2144 : */
2145 116914 : if (RELATION_IS_OTHER_TEMP(rel))
2146 : {
2147 1 : relation_close(rel, lmode);
2148 1 : PopActiveSnapshot();
2149 1 : CommitTransactionCommand();
2150 1 : return false;
2151 : }
2152 :
2153 : /*
2154 : * Silently ignore partitioned tables as there is no work to be done. The
2155 : * useful work is on their child partitions, which have been queued up for
2156 : * us separately.
2157 : */
2158 116913 : if (rel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
2159 : {
2160 124 : relation_close(rel, lmode);
2161 124 : PopActiveSnapshot();
2162 124 : CommitTransactionCommand();
2163 : /* It's OK to proceed with ANALYZE on this table */
2164 124 : return true;
2165 : }
2166 :
2167 : /*
2168 : * Get a session-level lock too. This will protect our access to the
2169 : * relation across multiple transactions, so that we can vacuum the
2170 : * relation's TOAST table (if any) secure in the knowledge that no one is
2171 : * deleting the parent relation.
2172 : *
2173 : * NOTE: this cannot block, even if someone else is waiting for access,
2174 : * because the lock manager knows that both lock requests are from the
2175 : * same process.
2176 : */
2177 116789 : lockrelid = rel->rd_lockInfo.lockRelId;
2178 116789 : LockRelationIdForSession(&lockrelid, lmode);
2179 :
2180 : /*
2181 : * Set index_cleanup option based on index_cleanup reloption if it wasn't
2182 : * specified in VACUUM command, or when running in an autovacuum worker
2183 : */
2184 116789 : if (params.index_cleanup == VACOPTVALUE_UNSPECIFIED)
2185 : {
2186 : StdRdOptIndexCleanup vacuum_index_cleanup;
2187 :
2188 116646 : if (rel->rd_options == NULL)
2189 115425 : vacuum_index_cleanup = STDRD_OPTION_VACUUM_INDEX_CLEANUP_AUTO;
2190 : else
2191 1221 : vacuum_index_cleanup =
2192 1221 : ((StdRdOptions *) rel->rd_options)->vacuum_index_cleanup;
2193 :
2194 116646 : if (vacuum_index_cleanup == STDRD_OPTION_VACUUM_INDEX_CLEANUP_AUTO)
2195 116618 : params.index_cleanup = VACOPTVALUE_AUTO;
2196 28 : else if (vacuum_index_cleanup == STDRD_OPTION_VACUUM_INDEX_CLEANUP_ON)
2197 14 : params.index_cleanup = VACOPTVALUE_ENABLED;
2198 : else
2199 : {
2200 : Assert(vacuum_index_cleanup ==
2201 : STDRD_OPTION_VACUUM_INDEX_CLEANUP_OFF);
2202 14 : params.index_cleanup = VACOPTVALUE_DISABLED;
2203 : }
2204 : }
2205 :
2206 : #ifdef USE_INJECTION_POINTS
2207 116789 : if (params.index_cleanup == VACOPTVALUE_AUTO)
2208 116622 : INJECTION_POINT("vacuum-index-cleanup-auto", NULL);
2209 167 : else if (params.index_cleanup == VACOPTVALUE_DISABLED)
2210 144 : INJECTION_POINT("vacuum-index-cleanup-disabled", NULL);
2211 23 : else if (params.index_cleanup == VACOPTVALUE_ENABLED)
2212 23 : INJECTION_POINT("vacuum-index-cleanup-enabled", NULL);
2213 : #endif
2214 :
2215 : /*
2216 : * Check if the vacuum_max_eager_freeze_failure_rate table storage
2217 : * parameter was specified. This overrides the GUC value.
2218 : */
2219 116789 : if (rel->rd_options != NULL &&
2220 1229 : ((StdRdOptions *) rel->rd_options)->vacuum_max_eager_freeze_failure_rate >= 0)
2221 0 : params.max_eager_freeze_failure_rate =
2222 0 : ((StdRdOptions *) rel->rd_options)->vacuum_max_eager_freeze_failure_rate;
2223 :
2224 : /*
2225 : * Set truncate option based on truncate reloption or GUC if it wasn't
2226 : * specified in VACUUM command, or when running in an autovacuum worker
2227 : */
2228 116789 : if (params.truncate == VACOPTVALUE_UNSPECIFIED)
2229 : {
2230 116650 : StdRdOptions *opts = (StdRdOptions *) rel->rd_options;
2231 :
2232 116650 : if (opts && opts->vacuum_truncate != PG_TERNARY_UNSET)
2233 : {
2234 20 : if (opts->vacuum_truncate == PG_TERNARY_TRUE)
2235 6 : params.truncate = VACOPTVALUE_ENABLED;
2236 : else
2237 14 : params.truncate = VACOPTVALUE_DISABLED;
2238 : }
2239 116630 : else if (vacuum_truncate)
2240 116616 : params.truncate = VACOPTVALUE_ENABLED;
2241 : else
2242 14 : params.truncate = VACOPTVALUE_DISABLED;
2243 : }
2244 :
2245 : #ifdef USE_INJECTION_POINTS
2246 116789 : if (params.truncate == VACOPTVALUE_AUTO)
2247 0 : INJECTION_POINT("vacuum-truncate-auto", NULL);
2248 116789 : else if (params.truncate == VACOPTVALUE_DISABLED)
2249 166 : INJECTION_POINT("vacuum-truncate-disabled", NULL);
2250 116623 : else if (params.truncate == VACOPTVALUE_ENABLED)
2251 116623 : INJECTION_POINT("vacuum-truncate-enabled", NULL);
2252 : #endif
2253 :
2254 : /*
2255 : * Remember the relation's TOAST relation for later, if the caller asked
2256 : * us to process it. In VACUUM FULL, though, the toast table is
2257 : * automatically rebuilt by cluster_rel so we shouldn't recurse to it,
2258 : * unless PROCESS_MAIN is disabled.
2259 : */
2260 116789 : if ((params.options & VACOPT_PROCESS_TOAST) != 0 &&
2261 16835 : ((params.options & VACOPT_FULL) == 0 ||
2262 226 : (params.options & VACOPT_PROCESS_MAIN) == 0))
2263 16613 : toast_relid = rel->rd_rel->reltoastrelid;
2264 : else
2265 100176 : toast_relid = InvalidOid;
2266 :
2267 : /*
2268 : * Switch to the table owner's userid, so that any index functions are run
2269 : * as that user. Also lock down security-restricted operations and
2270 : * arrange to make GUC variable changes local to this command. (This is
2271 : * unnecessary, but harmless, for lazy VACUUM.)
2272 : */
2273 116789 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
2274 116789 : SetUserIdAndSecContext(rel->rd_rel->relowner,
2275 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
2276 116789 : save_nestlevel = NewGUCNestLevel();
2277 116789 : RestrictSearchPath();
2278 :
2279 : /*
2280 : * If PROCESS_MAIN is set (the default), it's time to vacuum the main
2281 : * relation. Otherwise, we can skip this part. If processing the TOAST
2282 : * table is required (e.g., PROCESS_TOAST is set), we force PROCESS_MAIN
2283 : * to be set when we recurse to the TOAST table.
2284 : */
2285 116789 : if (params.options & VACOPT_PROCESS_MAIN)
2286 : {
2287 : /*
2288 : * Do the actual work --- either FULL or "lazy" vacuum
2289 : */
2290 116704 : if (params.options & VACOPT_FULL)
2291 : {
2292 222 : ClusterParams cluster_params = {0};
2293 :
2294 222 : if ((params.options & VACOPT_VERBOSE) != 0)
2295 1 : cluster_params.options |= CLUOPT_VERBOSE;
2296 :
2297 : /* VACUUM FULL is a variant of REPACK; see repack.c */
2298 222 : cluster_rel(REPACK_COMMAND_VACUUMFULL, rel, InvalidOid,
2299 : &cluster_params, isTopLevel);
2300 : /* cluster_rel closes the relation, but keeps lock */
2301 :
2302 218 : rel = NULL;
2303 : }
2304 : else
2305 116482 : table_relation_vacuum(rel, ¶ms, bstrategy);
2306 : }
2307 :
2308 : /* Roll back any GUC changes executed by index functions */
2309 116785 : AtEOXact_GUC(false, save_nestlevel);
2310 :
2311 : /* Restore userid and security context */
2312 116785 : SetUserIdAndSecContext(save_userid, save_sec_context);
2313 :
2314 : /* all done with this class, but hold lock until commit */
2315 116785 : if (rel)
2316 116567 : relation_close(rel, NoLock);
2317 :
2318 : /*
2319 : * Complete the transaction and free all temporary memory used.
2320 : */
2321 116785 : PopActiveSnapshot();
2322 116785 : CommitTransactionCommand();
2323 :
2324 : /*
2325 : * If the relation has a secondary toast rel, vacuum that too while we
2326 : * still hold the session lock on the main table. Note however that
2327 : * "analyze" will not get done on the toast table. This is good, because
2328 : * the toaster always uses hardcoded index access and statistics are
2329 : * totally unimportant for toast relations.
2330 : */
2331 116785 : if (toast_relid != InvalidOid)
2332 : {
2333 : /*
2334 : * Force VACOPT_PROCESS_MAIN so vacuum_rel() processes it. Likewise,
2335 : * set toast_parent so that the privilege checks are done on the main
2336 : * relation. NB: This is only safe to do because we hold a session
2337 : * lock on the main relation that prevents concurrent deletion.
2338 : */
2339 5808 : toast_vacuum_params.options |= VACOPT_PROCESS_MAIN;
2340 5808 : toast_vacuum_params.toast_parent = relid;
2341 :
2342 5808 : vacuum_rel(toast_relid, NULL, toast_vacuum_params, bstrategy,
2343 : isTopLevel);
2344 : }
2345 :
2346 : /*
2347 : * Now release the session-level lock on the main table.
2348 : */
2349 116785 : UnlockRelationIdForSession(&lockrelid, lmode);
2350 :
2351 : /* Report that we really did it. */
2352 116785 : return true;
2353 : }
2354 :
2355 :
2356 : /*
2357 : * Open all the vacuumable indexes of the given relation, obtaining the
2358 : * specified kind of lock on each. Return an array of Relation pointers for
2359 : * the indexes into *Irel, and the number of indexes into *nindexes.
2360 : *
2361 : * We consider an index vacuumable if it is marked insertable (indisready).
2362 : * If it isn't, probably a CREATE INDEX CONCURRENTLY command failed early in
2363 : * execution, and what we have is too corrupt to be processable. We will
2364 : * vacuum even if the index isn't indisvalid; this is important because in a
2365 : * unique index, uniqueness checks will be performed anyway and had better not
2366 : * hit dangling index pointers.
2367 : */
2368 : void
2369 126096 : vac_open_indexes(Relation relation, LOCKMODE lockmode,
2370 : int *nindexes, Relation **Irel)
2371 : {
2372 : List *indexoidlist;
2373 : ListCell *indexoidscan;
2374 : int i;
2375 :
2376 : Assert(lockmode != NoLock);
2377 :
2378 126096 : indexoidlist = RelationGetIndexList(relation);
2379 :
2380 : /* allocate enough memory for all indexes */
2381 126096 : i = list_length(indexoidlist);
2382 :
2383 126096 : if (i > 0)
2384 118430 : *Irel = (Relation *) palloc(i * sizeof(Relation));
2385 : else
2386 7666 : *Irel = NULL;
2387 :
2388 : /* collect just the ready indexes */
2389 126096 : i = 0;
2390 316847 : foreach(indexoidscan, indexoidlist)
2391 : {
2392 190751 : Oid indexoid = lfirst_oid(indexoidscan);
2393 : Relation indrel;
2394 :
2395 190751 : indrel = index_open(indexoid, lockmode);
2396 190751 : if (indrel->rd_index->indisready)
2397 190751 : (*Irel)[i++] = indrel;
2398 : else
2399 0 : index_close(indrel, lockmode);
2400 : }
2401 :
2402 126096 : *nindexes = i;
2403 :
2404 126096 : list_free(indexoidlist);
2405 126096 : }
2406 :
2407 : /*
2408 : * Release the resources acquired by vac_open_indexes. Optionally release
2409 : * the locks (say NoLock to keep 'em).
2410 : */
2411 : void
2412 126680 : vac_close_indexes(int nindexes, Relation *Irel, LOCKMODE lockmode)
2413 : {
2414 126680 : if (Irel == NULL)
2415 8254 : return;
2416 :
2417 309169 : while (nindexes--)
2418 : {
2419 190743 : Relation ind = Irel[nindexes];
2420 :
2421 190743 : index_close(ind, lockmode);
2422 : }
2423 118426 : pfree(Irel);
2424 : }
2425 :
2426 : /*
2427 : * vacuum_delay_point --- check for interrupts and cost-based delay.
2428 : *
2429 : * This should be called in each major loop of VACUUM processing,
2430 : * typically once per page processed.
2431 : */
2432 : void
2433 51193451 : vacuum_delay_point(bool is_analyze)
2434 : {
2435 51193451 : double msec = 0;
2436 :
2437 : /* Always check for interrupts */
2438 51193451 : CHECK_FOR_INTERRUPTS();
2439 :
2440 51193451 : if (InterruptPending)
2441 0 : return;
2442 :
2443 51193451 : if (IsParallelWorker())
2444 : {
2445 : /*
2446 : * Update cost-based vacuum delay parameters for a parallel autovacuum
2447 : * worker if any changes are detected. It might enable cost-based
2448 : * delay so it needs to be called before VacuumCostActive check.
2449 : */
2450 256 : parallel_vacuum_update_shared_delay_params();
2451 : }
2452 :
2453 51193451 : if (!VacuumCostActive && !ConfigReloadPending)
2454 46613655 : return;
2455 :
2456 : /*
2457 : * Autovacuum workers should reload the configuration file if requested.
2458 : * This allows changes to [autovacuum_]vacuum_cost_limit and
2459 : * [autovacuum_]vacuum_cost_delay to take effect while a table is being
2460 : * vacuumed or analyzed.
2461 : */
2462 4579796 : if (ConfigReloadPending && AmAutoVacuumWorkerProcess())
2463 : {
2464 1 : ConfigReloadPending = false;
2465 1 : ProcessConfigFile(PGC_SIGHUP);
2466 1 : VacuumUpdateCosts();
2467 :
2468 : /*
2469 : * Propagate cost-based vacuum delay parameters to shared memory if
2470 : * any of them have changed during the config reload.
2471 : */
2472 1 : parallel_vacuum_propagate_shared_delay_params();
2473 : }
2474 :
2475 : /*
2476 : * If we disabled cost-based delays after reloading the config file,
2477 : * return.
2478 : */
2479 4579796 : if (!VacuumCostActive)
2480 0 : return;
2481 :
2482 : /*
2483 : * For parallel vacuum, the delay is computed based on the shared cost
2484 : * balance. See compute_parallel_delay.
2485 : */
2486 4579796 : if (VacuumSharedCostBalance != NULL)
2487 427 : msec = compute_parallel_delay();
2488 4579369 : else if (VacuumCostBalance >= vacuum_cost_limit)
2489 2664 : msec = vacuum_cost_delay * VacuumCostBalance / vacuum_cost_limit;
2490 :
2491 : /* Nap if appropriate */
2492 4579796 : if (msec > 0)
2493 : {
2494 : instr_time delay_start;
2495 :
2496 2685 : if (msec > vacuum_cost_delay * 4)
2497 5 : msec = vacuum_cost_delay * 4;
2498 :
2499 2685 : if (track_cost_delay_timing)
2500 0 : INSTR_TIME_SET_CURRENT(delay_start);
2501 :
2502 2685 : pgstat_report_wait_start(WAIT_EVENT_VACUUM_DELAY);
2503 2685 : pg_usleep(msec * 1000);
2504 2685 : pgstat_report_wait_end();
2505 :
2506 2685 : if (track_cost_delay_timing)
2507 : {
2508 : instr_time delay_end;
2509 : instr_time delay;
2510 :
2511 0 : INSTR_TIME_SET_CURRENT(delay_end);
2512 0 : INSTR_TIME_SET_ZERO(delay);
2513 0 : INSTR_TIME_ACCUM_DIFF(delay, delay_end, delay_start);
2514 :
2515 : /*
2516 : * For parallel workers, we only report the delay time every once
2517 : * in a while to avoid overloading the leader with messages and
2518 : * interrupts.
2519 : */
2520 0 : if (IsParallelWorker())
2521 : {
2522 : static instr_time last_report_time;
2523 : instr_time time_since_last_report;
2524 :
2525 : Assert(!is_analyze);
2526 :
2527 : /* Accumulate the delay time */
2528 0 : parallel_vacuum_worker_delay_ns += INSTR_TIME_GET_NANOSEC(delay);
2529 :
2530 : /* Calculate interval since last report */
2531 0 : INSTR_TIME_SET_ZERO(time_since_last_report);
2532 0 : INSTR_TIME_ACCUM_DIFF(time_since_last_report, delay_end, last_report_time);
2533 :
2534 : /* If we haven't reported in a while, do so now */
2535 0 : if (INSTR_TIME_GET_NANOSEC(time_since_last_report) >=
2536 : PARALLEL_VACUUM_DELAY_REPORT_INTERVAL_NS)
2537 : {
2538 0 : pgstat_progress_parallel_incr_param(PROGRESS_VACUUM_DELAY_TIME,
2539 : parallel_vacuum_worker_delay_ns);
2540 :
2541 : /* Reset variables */
2542 0 : last_report_time = delay_end;
2543 0 : parallel_vacuum_worker_delay_ns = 0;
2544 : }
2545 : }
2546 0 : else if (is_analyze)
2547 0 : pgstat_progress_incr_param(PROGRESS_ANALYZE_DELAY_TIME,
2548 0 : INSTR_TIME_GET_NANOSEC(delay));
2549 : else
2550 0 : pgstat_progress_incr_param(PROGRESS_VACUUM_DELAY_TIME,
2551 0 : INSTR_TIME_GET_NANOSEC(delay));
2552 : }
2553 :
2554 : /*
2555 : * We don't want to ignore postmaster death during very long vacuums
2556 : * with vacuum_cost_delay configured. We can't use the usual
2557 : * WaitLatch() approach here because we want microsecond-based sleep
2558 : * durations above.
2559 : */
2560 2685 : if (IsUnderPostmaster && !PostmasterIsAlive())
2561 0 : exit(1);
2562 :
2563 2685 : VacuumCostBalance = 0;
2564 :
2565 : /*
2566 : * Balance and update limit values for autovacuum workers. We must do
2567 : * this periodically, as the number of workers across which we are
2568 : * balancing the limit may have changed.
2569 : *
2570 : * TODO: There may be better criteria for determining when to do this
2571 : * besides "check after napping".
2572 : */
2573 2685 : AutoVacuumUpdateCostLimit();
2574 :
2575 : /* Might have gotten an interrupt while sleeping */
2576 2685 : CHECK_FOR_INTERRUPTS();
2577 : }
2578 : }
2579 :
2580 : /*
2581 : * Computes the vacuum delay for parallel workers.
2582 : *
2583 : * The basic idea of a cost-based delay for parallel vacuum is to allow each
2584 : * worker to sleep in proportion to the share of work it's done. We achieve this
2585 : * by allowing all parallel vacuum workers including the leader process to
2586 : * have a shared view of cost related parameters (mainly VacuumCostBalance).
2587 : * We allow each worker to update it as and when it has incurred any cost and
2588 : * then based on that decide whether it needs to sleep. We compute the time
2589 : * to sleep for a worker based on the cost it has incurred
2590 : * (VacuumCostBalanceLocal) and then reduce the VacuumSharedCostBalance by
2591 : * that amount. This avoids putting to sleep those workers which have done less
2592 : * I/O than other workers and therefore ensure that workers
2593 : * which are doing more I/O got throttled more.
2594 : *
2595 : * We allow a worker to sleep only if it has performed I/O above a certain
2596 : * threshold, which is calculated based on the number of active workers
2597 : * (VacuumActiveNWorkers), and the overall cost balance is more than
2598 : * VacuumCostLimit set by the system. Testing reveals that we achieve
2599 : * the required throttling if we force a worker that has done more than 50%
2600 : * of its share of work to sleep.
2601 : */
2602 : static double
2603 427 : compute_parallel_delay(void)
2604 : {
2605 427 : double msec = 0;
2606 : uint32 shared_balance;
2607 : int nworkers;
2608 :
2609 : /* Parallel vacuum must be active */
2610 : Assert(VacuumSharedCostBalance);
2611 :
2612 427 : nworkers = pg_atomic_read_u32(VacuumActiveNWorkers);
2613 :
2614 : /* At least count itself */
2615 : Assert(nworkers >= 1);
2616 :
2617 : /* Update the shared cost balance value atomically */
2618 427 : shared_balance = pg_atomic_add_fetch_u32(VacuumSharedCostBalance, VacuumCostBalance);
2619 :
2620 : /* Compute the total local balance for the current worker */
2621 427 : VacuumCostBalanceLocal += VacuumCostBalance;
2622 :
2623 427 : if ((shared_balance >= vacuum_cost_limit) &&
2624 85 : (VacuumCostBalanceLocal > 0.5 * ((double) vacuum_cost_limit / nworkers)))
2625 : {
2626 : /* Compute sleep time based on the local cost balance */
2627 21 : msec = vacuum_cost_delay * VacuumCostBalanceLocal / vacuum_cost_limit;
2628 21 : pg_atomic_sub_fetch_u32(VacuumSharedCostBalance, VacuumCostBalanceLocal);
2629 21 : VacuumCostBalanceLocal = 0;
2630 : }
2631 :
2632 : /*
2633 : * Reset the local balance as we accumulated it into the shared value.
2634 : */
2635 427 : VacuumCostBalance = 0;
2636 :
2637 427 : return msec;
2638 : }
2639 :
2640 : /*
2641 : * A wrapper function of defGetBoolean().
2642 : *
2643 : * This function returns VACOPTVALUE_ENABLED and VACOPTVALUE_DISABLED instead
2644 : * of true and false.
2645 : */
2646 : static VacOptValue
2647 182 : get_vacoptval_from_boolean(DefElem *def)
2648 : {
2649 182 : return defGetBoolean(def) ? VACOPTVALUE_ENABLED : VACOPTVALUE_DISABLED;
2650 : }
2651 :
2652 : /*
2653 : * vac_bulkdel_one_index() -- bulk-deletion for index relation.
2654 : *
2655 : * Returns bulk delete stats derived from input stats
2656 : */
2657 : IndexBulkDeleteResult *
2658 1479 : vac_bulkdel_one_index(IndexVacuumInfo *ivinfo, IndexBulkDeleteResult *istat,
2659 : TidStore *dead_items, VacDeadItemsInfo *dead_items_info)
2660 : {
2661 : /* Do bulk deletion */
2662 1479 : istat = index_bulk_delete(ivinfo, istat, vac_tid_reaped,
2663 : dead_items);
2664 :
2665 1479 : ereport(ivinfo->message_level,
2666 : (errmsg("scanned index \"%s\" to remove %" PRId64 " row versions",
2667 : RelationGetRelationName(ivinfo->index),
2668 : dead_items_info->num_items)));
2669 :
2670 1479 : return istat;
2671 : }
2672 :
2673 : /*
2674 : * vac_cleanup_one_index() -- do post-vacuum cleanup for index relation.
2675 : *
2676 : * Returns bulk delete stats derived from input stats
2677 : */
2678 : IndexBulkDeleteResult *
2679 140310 : vac_cleanup_one_index(IndexVacuumInfo *ivinfo, IndexBulkDeleteResult *istat)
2680 : {
2681 140310 : istat = index_vacuum_cleanup(ivinfo, istat);
2682 :
2683 140310 : if (istat)
2684 1632 : ereport(ivinfo->message_level,
2685 : (errmsg("index \"%s\" now contains %.0f row versions in %u pages",
2686 : RelationGetRelationName(ivinfo->index),
2687 : istat->num_index_tuples,
2688 : istat->num_pages),
2689 : errdetail("%.0f index row versions were removed.\n"
2690 : "%u index pages were newly deleted.\n"
2691 : "%u index pages are currently deleted, of which %u are currently reusable.",
2692 : istat->tuples_removed,
2693 : istat->pages_newly_deleted,
2694 : istat->pages_deleted, istat->pages_free)));
2695 :
2696 140310 : return istat;
2697 : }
2698 :
2699 : /*
2700 : * vac_tid_reaped() -- is a particular tid deletable?
2701 : *
2702 : * This has the right signature to be an IndexBulkDeleteCallback.
2703 : */
2704 : static bool
2705 3918623 : vac_tid_reaped(ItemPointer itemptr, void *state)
2706 : {
2707 3918623 : TidStore *dead_items = (TidStore *) state;
2708 :
2709 3918623 : return TidStoreIsMember(dead_items, itemptr);
2710 : }
|
