Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * vacuum.c
4 : * The postgres vacuum cleaner.
5 : *
6 : * This file includes (a) control and dispatch code for VACUUM and ANALYZE
7 : * commands, (b) code to compute various vacuum thresholds, and (c) index
8 : * vacuum code.
9 : *
10 : * VACUUM for heap AM is implemented in vacuumlazy.c, parallel vacuum in
11 : * vacuumparallel.c, ANALYZE in analyze.c, and VACUUM FULL is a variant of
12 : * CLUSTER, handled in cluster.c.
13 : *
14 : *
15 : * Portions Copyright (c) 1996-2026, 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/cluster.h"
42 : #include "commands/defrem.h"
43 : #include "commands/progress.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);
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 1222 : check_vacuum_buffer_usage_limit(int *newval, void **extra,
141 : GucSource source)
142 : {
143 : /* Value upper and lower hard limits are inclusive */
144 1222 : if (*newval == 0 || (*newval >= MIN_BAS_VAC_RING_SIZE_KB &&
145 1222 : *newval <= MAX_BAS_VAC_RING_SIZE_KB))
146 1222 : 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 8674 : ExecVacuum(ParseState *pstate, VacuumStmt *vacstmt, bool isTopLevel)
164 : {
165 : VacuumParams params;
166 8674 : BufferAccessStrategy bstrategy = NULL;
167 8674 : bool verbose = false;
168 8674 : bool skip_locked = false;
169 8674 : bool analyze = false;
170 8674 : bool freeze = false;
171 8674 : bool full = false;
172 8674 : bool disable_page_skipping = false;
173 8674 : bool process_main = true;
174 8674 : bool process_toast = true;
175 : int ring_size;
176 8674 : bool skip_database_stats = false;
177 8674 : bool only_database_stats = false;
178 : MemoryContext vac_context;
179 : ListCell *lc;
180 :
181 : /* index_cleanup and truncate values unspecified for now */
182 8674 : params.index_cleanup = VACOPTVALUE_UNSPECIFIED;
183 8674 : params.truncate = VACOPTVALUE_UNSPECIFIED;
184 :
185 : /* By default parallel vacuum is enabled */
186 8674 : params.nworkers = 0;
187 :
188 : /* Will be set later if we recurse to a TOAST table. */
189 8674 : 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 8674 : ring_size = -1;
196 :
197 : /* Parse options list */
198 18227 : foreach(lc, vacstmt->options)
199 : {
200 9577 : DefElem *opt = (DefElem *) lfirst(lc);
201 :
202 : /* Parse common options for VACUUM and ANALYZE */
203 9577 : if (strcmp(opt->defname, "verbose") == 0)
204 25 : verbose = defGetBoolean(opt);
205 9552 : else if (strcmp(opt->defname, "skip_locked") == 0)
206 182 : skip_locked = defGetBoolean(opt);
207 9370 : 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 9334 : 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 9330 : else if (strcmp(opt->defname, "analyze") == 0)
242 1834 : analyze = defGetBoolean(opt);
243 7496 : else if (strcmp(opt->defname, "freeze") == 0)
244 1885 : 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 8650 : params.options =
306 8650 : (vacstmt->is_vacuumcmd ? VACOPT_VACUUM : VACOPT_ANALYZE) |
307 8650 : (verbose ? VACOPT_VERBOSE : 0) |
308 8650 : (skip_locked ? VACOPT_SKIP_LOCKED : 0) |
309 8650 : (analyze ? VACOPT_ANALYZE : 0) |
310 8650 : (freeze ? VACOPT_FREEZE : 0) |
311 8650 : (full ? VACOPT_FULL : 0) |
312 8650 : (disable_page_skipping ? VACOPT_DISABLE_PAGE_SKIPPING : 0) |
313 8650 : (process_main ? VACOPT_PROCESS_MAIN : 0) |
314 8650 : (process_toast ? VACOPT_PROCESS_TOAST : 0) |
315 8650 : (skip_database_stats ? VACOPT_SKIP_DATABASE_STATS : 0) |
316 8650 : (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 8650 : 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 8646 : 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 8642 : if (!(params.options & VACOPT_ANALYZE))
343 : {
344 7799 : foreach(lc, vacstmt->rels)
345 : {
346 3847 : VacuumRelation *vrel = lfirst_node(VacuumRelation, lc);
347 :
348 3847 : 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 8638 : 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 8638 : 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 8634 : 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 8630 : if (params.options & VACOPT_FREEZE)
395 : {
396 1885 : params.freeze_min_age = 0;
397 1885 : params.freeze_table_age = 0;
398 1885 : params.multixact_freeze_min_age = 0;
399 1885 : params.multixact_freeze_table_age = 0;
400 : }
401 : else
402 : {
403 6745 : params.freeze_min_age = -1;
404 6745 : params.freeze_table_age = -1;
405 6745 : params.multixact_freeze_min_age = -1;
406 6745 : params.multixact_freeze_table_age = -1;
407 : }
408 :
409 : /* user-invoked vacuum is never "for wraparound" */
410 8630 : 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 8630 : params.log_vacuum_min_duration = -1;
417 8630 : params.log_analyze_min_duration = -1;
418 :
419 : /*
420 : * Later, in vacuum_rel(), we check if a reloption override was specified.
421 : */
422 8630 : 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 8630 : 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 8630 : if ((params.options & (VACOPT_ONLY_DATABASE_STATS |
442 285 : VACOPT_FULL)) == 0 ||
443 285 : (params.options & VACOPT_ANALYZE) != 0)
444 : {
445 :
446 8349 : 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 8349 : if (ring_size == -1)
457 8329 : ring_size = VacuumBufferUsageLimit;
458 :
459 8349 : bstrategy = GetAccessStrategyWithSize(BAS_VACUUM, ring_size);
460 :
461 8349 : MemoryContextSwitchTo(old_context);
462 : }
463 :
464 : /* Now go through the common routine */
465 8630 : vacuum(vacstmt->rels, params, bstrategy, vac_context, isTopLevel);
466 :
467 : /* Finally, clean up the vacuum memory context */
468 8543 : MemoryContextDelete(vac_context);
469 8543 : }
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 109159 : 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 109159 : 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 109159 : if (params.options & VACOPT_VACUUM)
514 : {
515 106112 : PreventInTransactionBlock(isTopLevel, stmttype);
516 106099 : in_outer_xact = false;
517 : }
518 : else
519 3047 : 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 109146 : 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 109138 : if (params.options & VACOPT_ONLY_DATABASE_STATS)
537 : {
538 : /* We don't process any tables in this case */
539 : Assert(relations == NIL);
540 : }
541 109063 : else if (relations != NIL)
542 : {
543 108950 : List *newrels = NIL;
544 : ListCell *lc;
545 :
546 217981 : foreach(lc, relations)
547 : {
548 109055 : VacuumRelation *vrel = lfirst_node(VacuumRelation, lc);
549 : List *sublist;
550 : MemoryContext old_context;
551 :
552 109055 : sublist = expand_vacuum_rel(vrel, vac_context, params.options);
553 109031 : old_context = MemoryContextSwitchTo(vac_context);
554 109031 : newrels = list_concat(newrels, sublist);
555 109031 : MemoryContextSwitchTo(old_context);
556 : }
557 108926 : relations = newrels;
558 : }
559 : else
560 113 : 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 109114 : if (params.options & VACOPT_VACUUM)
577 106091 : use_own_xacts = true;
578 : else
579 : {
580 : Assert(params.options & VACOPT_ANALYZE);
581 3023 : if (AmAutoVacuumWorkerProcess())
582 195 : use_own_xacts = true;
583 2828 : else if (in_outer_xact)
584 164 : use_own_xacts = false;
585 2664 : else if (list_length(relations) > 1)
586 518 : use_own_xacts = true;
587 : else
588 2146 : 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 109114 : if (use_own_xacts)
600 : {
601 : Assert(!in_outer_xact);
602 :
603 : /* ActiveSnapshot is not set by autovacuum */
604 106804 : if (ActiveSnapshotSet())
605 6275 : PopActiveSnapshot();
606 :
607 : /* matches the StartTransaction in PostgresMain() */
608 106804 : CommitTransactionCommand();
609 : }
610 :
611 : /* Turn vacuum cost accounting on or off, and set/clear in_vacuum */
612 109114 : PG_TRY();
613 : {
614 : ListCell *cur;
615 :
616 109114 : in_vacuum = true;
617 109114 : VacuumFailsafeActive = false;
618 109114 : VacuumUpdateCosts();
619 109114 : VacuumCostBalance = 0;
620 109114 : VacuumCostBalanceLocal = 0;
621 109114 : VacuumSharedCostBalance = NULL;
622 109114 : VacuumActiveNWorkers = NULL;
623 :
624 : /*
625 : * Loop to process each selected relation.
626 : */
627 228066 : foreach(cur, relations)
628 : {
629 118994 : VacuumRelation *vrel = lfirst_node(VacuumRelation, cur);
630 :
631 118994 : if (params.options & VACOPT_VACUUM)
632 : {
633 111185 : if (!vacuum_rel(vrel->oid, vrel->relation, params, bstrategy))
634 62 : continue;
635 : }
636 :
637 118928 : if (params.options & VACOPT_ANALYZE)
638 : {
639 : /*
640 : * If using separate xacts, start one for analyze. Otherwise,
641 : * we can use the outer transaction.
642 : */
643 9818 : if (use_own_xacts)
644 : {
645 7524 : StartTransactionCommand();
646 : /* functions in indexes may want a snapshot set */
647 7524 : PushActiveSnapshot(GetTransactionSnapshot());
648 : }
649 :
650 9818 : analyze_rel(vrel->oid, vrel->relation, params,
651 : vrel->va_cols, in_outer_xact, bstrategy);
652 :
653 9780 : if (use_own_xacts)
654 : {
655 7499 : PopActiveSnapshot();
656 : /* standard_ProcessUtility() does CCI if !use_own_xacts */
657 7499 : CommandCounterIncrement();
658 7499 : CommitTransactionCommand();
659 : }
660 : else
661 : {
662 : /*
663 : * If we're not using separate xacts, better separate the
664 : * ANALYZE actions with CCIs. This avoids trouble if user
665 : * says "ANALYZE t, t".
666 : */
667 2281 : CommandCounterIncrement();
668 : }
669 : }
670 :
671 : /*
672 : * Ensure VacuumFailsafeActive has been reset before vacuuming the
673 : * next relation.
674 : */
675 118890 : VacuumFailsafeActive = false;
676 : }
677 : }
678 42 : PG_FINALLY();
679 : {
680 109114 : in_vacuum = false;
681 109114 : VacuumCostActive = false;
682 109114 : VacuumFailsafeActive = false;
683 109114 : VacuumCostBalance = 0;
684 : }
685 109114 : PG_END_TRY();
686 :
687 : /*
688 : * Finish up processing.
689 : */
690 109072 : if (use_own_xacts)
691 : {
692 : /* here, we are not in a transaction */
693 :
694 : /*
695 : * This matches the CommitTransaction waiting for us in
696 : * PostgresMain().
697 : */
698 106775 : StartTransactionCommand();
699 : }
700 :
701 109072 : if ((params.options & VACOPT_VACUUM) &&
702 106070 : !(params.options & VACOPT_SKIP_DATABASE_STATS))
703 : {
704 : /*
705 : * Update pg_database.datfrozenxid, and truncate pg_xact if possible.
706 : */
707 1114 : vac_update_datfrozenxid();
708 : }
709 :
710 109072 : }
711 :
712 : /*
713 : * Check if the current user has privileges to vacuum or analyze the relation.
714 : * If not, issue a WARNING log message and return false to let the caller
715 : * decide what to do with this relation. This routine is used to decide if a
716 : * relation can be processed for VACUUM or ANALYZE.
717 : */
718 : bool
719 145655 : vacuum_is_permitted_for_relation(Oid relid, Form_pg_class reltuple,
720 : bits32 options)
721 : {
722 : char *relname;
723 :
724 : Assert((options & (VACOPT_VACUUM | VACOPT_ANALYZE)) != 0);
725 :
726 : /*----------
727 : * A role has privileges to vacuum or analyze the relation if any of the
728 : * following are true:
729 : * - the role owns the current database and the relation is not shared
730 : * - the role has the MAINTAIN privilege on the relation
731 : *----------
732 : */
733 145655 : if ((object_ownercheck(DatabaseRelationId, MyDatabaseId, GetUserId()) &&
734 166675 : !reltuple->relisshared) ||
735 23476 : pg_class_aclcheck(relid, GetUserId(), ACL_MAINTAIN) == ACLCHECK_OK)
736 143511 : return true;
737 :
738 2144 : relname = NameStr(reltuple->relname);
739 :
740 2144 : if ((options & VACOPT_VACUUM) != 0)
741 : {
742 148 : ereport(WARNING,
743 : (errmsg("permission denied to vacuum \"%s\", skipping it",
744 : relname)));
745 :
746 : /*
747 : * For VACUUM ANALYZE, both logs could show up, but just generate
748 : * information for VACUUM as that would be the first one to be
749 : * processed.
750 : */
751 148 : return false;
752 : }
753 :
754 1996 : if ((options & VACOPT_ANALYZE) != 0)
755 1996 : ereport(WARNING,
756 : (errmsg("permission denied to analyze \"%s\", skipping it",
757 : relname)));
758 :
759 1996 : return false;
760 : }
761 :
762 :
763 : /*
764 : * vacuum_open_relation
765 : *
766 : * This routine is used for attempting to open and lock a relation which
767 : * is going to be vacuumed or analyzed. If the relation cannot be opened
768 : * or locked, a log is emitted if possible.
769 : */
770 : Relation
771 126557 : vacuum_open_relation(Oid relid, RangeVar *relation, bits32 options,
772 : bool verbose, LOCKMODE lmode)
773 : {
774 : Relation rel;
775 126557 : bool rel_lock = true;
776 : int elevel;
777 :
778 : Assert((options & (VACOPT_VACUUM | VACOPT_ANALYZE)) != 0);
779 :
780 : /*
781 : * Open the relation and get the appropriate lock on it.
782 : *
783 : * There's a race condition here: the relation may have gone away since
784 : * the last time we saw it. If so, we don't need to vacuum or analyze it.
785 : *
786 : * If we've been asked not to wait for the relation lock, acquire it first
787 : * in non-blocking mode, before calling try_relation_open().
788 : */
789 126557 : if (!(options & VACOPT_SKIP_LOCKED))
790 125841 : rel = try_relation_open(relid, lmode);
791 716 : else if (ConditionalLockRelationOid(relid, lmode))
792 704 : rel = try_relation_open(relid, NoLock);
793 : else
794 : {
795 12 : rel = NULL;
796 12 : rel_lock = false;
797 : }
798 :
799 : /* if relation is opened, leave */
800 126557 : if (rel)
801 126539 : return rel;
802 :
803 : /*
804 : * Relation could not be opened, hence generate if possible a log
805 : * informing on the situation.
806 : *
807 : * If the RangeVar is not defined, we do not have enough information to
808 : * provide a meaningful log statement. Chances are that the caller has
809 : * intentionally not provided this information so that this logging is
810 : * skipped, anyway.
811 : */
812 18 : if (relation == NULL)
813 9 : return NULL;
814 :
815 : /*
816 : * Determine the log level.
817 : *
818 : * For manual VACUUM or ANALYZE, we emit a WARNING to match the log
819 : * statements in the permission checks; otherwise, only log if the caller
820 : * so requested.
821 : */
822 9 : if (!AmAutoVacuumWorkerProcess())
823 7 : elevel = WARNING;
824 2 : else if (verbose)
825 2 : elevel = LOG;
826 : else
827 0 : return NULL;
828 :
829 9 : if ((options & VACOPT_VACUUM) != 0)
830 : {
831 5 : if (!rel_lock)
832 3 : ereport(elevel,
833 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
834 : errmsg("skipping vacuum of \"%s\" --- lock not available",
835 : relation->relname)));
836 : else
837 2 : ereport(elevel,
838 : (errcode(ERRCODE_UNDEFINED_TABLE),
839 : errmsg("skipping vacuum of \"%s\" --- relation no longer exists",
840 : relation->relname)));
841 :
842 : /*
843 : * For VACUUM ANALYZE, both logs could show up, but just generate
844 : * information for VACUUM as that would be the first one to be
845 : * processed.
846 : */
847 5 : return NULL;
848 : }
849 :
850 4 : if ((options & VACOPT_ANALYZE) != 0)
851 : {
852 4 : if (!rel_lock)
853 3 : ereport(elevel,
854 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
855 : errmsg("skipping analyze of \"%s\" --- lock not available",
856 : relation->relname)));
857 : else
858 1 : ereport(elevel,
859 : (errcode(ERRCODE_UNDEFINED_TABLE),
860 : errmsg("skipping analyze of \"%s\" --- relation no longer exists",
861 : relation->relname)));
862 : }
863 :
864 4 : return NULL;
865 : }
866 :
867 :
868 : /*
869 : * Given a VacuumRelation, fill in the table OID if it wasn't specified,
870 : * and optionally add VacuumRelations for partitions or inheritance children.
871 : *
872 : * If a VacuumRelation does not have an OID supplied and is a partitioned
873 : * table, an extra entry will be added to the output for each partition.
874 : * Presently, only autovacuum supplies OIDs when calling vacuum(), and
875 : * it does not want us to expand partitioned tables.
876 : *
877 : * We take care not to modify the input data structure, but instead build
878 : * new VacuumRelation(s) to return. (But note that they will reference
879 : * unmodified parts of the input, eg column lists.) New data structures
880 : * are made in vac_context.
881 : */
882 : static List *
883 109055 : expand_vacuum_rel(VacuumRelation *vrel, MemoryContext vac_context,
884 : int options)
885 : {
886 109055 : List *vacrels = NIL;
887 : MemoryContext oldcontext;
888 :
889 : /* If caller supplied OID, there's nothing we need do here. */
890 109055 : if (OidIsValid(vrel->oid))
891 : {
892 100529 : oldcontext = MemoryContextSwitchTo(vac_context);
893 100529 : vacrels = lappend(vacrels, vrel);
894 100529 : MemoryContextSwitchTo(oldcontext);
895 : }
896 : else
897 : {
898 : /*
899 : * Process a specific relation, and possibly partitions or child
900 : * tables thereof.
901 : */
902 : Oid relid;
903 : HeapTuple tuple;
904 : Form_pg_class classForm;
905 : bool include_children;
906 : bool is_partitioned_table;
907 : int rvr_opts;
908 :
909 : /*
910 : * Since autovacuum workers supply OIDs when calling vacuum(), no
911 : * autovacuum worker should reach this code.
912 : */
913 : Assert(!AmAutoVacuumWorkerProcess());
914 :
915 : /*
916 : * We transiently take AccessShareLock to protect the syscache lookup
917 : * below, as well as find_all_inheritors's expectation that the caller
918 : * holds some lock on the starting relation.
919 : */
920 8526 : rvr_opts = (options & VACOPT_SKIP_LOCKED) ? RVR_SKIP_LOCKED : 0;
921 8526 : relid = RangeVarGetRelidExtended(vrel->relation,
922 : AccessShareLock,
923 : rvr_opts,
924 : NULL, NULL);
925 :
926 : /*
927 : * If the lock is unavailable, emit the same log statement that
928 : * vacuum_rel() and analyze_rel() would.
929 : */
930 8502 : if (!OidIsValid(relid))
931 : {
932 4 : if (options & VACOPT_VACUUM)
933 3 : ereport(WARNING,
934 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
935 : errmsg("skipping vacuum of \"%s\" --- lock not available",
936 : vrel->relation->relname)));
937 : else
938 1 : ereport(WARNING,
939 : (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
940 : errmsg("skipping analyze of \"%s\" --- lock not available",
941 : vrel->relation->relname)));
942 4 : return vacrels;
943 : }
944 :
945 : /*
946 : * To check whether the relation is a partitioned table and its
947 : * ownership, fetch its syscache entry.
948 : */
949 8498 : tuple = SearchSysCache1(RELOID, ObjectIdGetDatum(relid));
950 8498 : if (!HeapTupleIsValid(tuple))
951 0 : elog(ERROR, "cache lookup failed for relation %u", relid);
952 8498 : classForm = (Form_pg_class) GETSTRUCT(tuple);
953 :
954 : /*
955 : * Make a returnable VacuumRelation for this rel if the user has the
956 : * required privileges.
957 : */
958 8498 : if (vacuum_is_permitted_for_relation(relid, classForm, options))
959 : {
960 8346 : oldcontext = MemoryContextSwitchTo(vac_context);
961 8346 : vacrels = lappend(vacrels, makeVacuumRelation(vrel->relation,
962 : relid,
963 : vrel->va_cols));
964 8346 : MemoryContextSwitchTo(oldcontext);
965 : }
966 :
967 : /*
968 : * Vacuuming a partitioned table with ONLY will not do anything since
969 : * the partitioned table itself is empty. Issue a warning if the user
970 : * requests this.
971 : */
972 8498 : include_children = vrel->relation->inh;
973 8498 : is_partitioned_table = (classForm->relkind == RELKIND_PARTITIONED_TABLE);
974 8498 : if ((options & VACOPT_VACUUM) && is_partitioned_table && !include_children)
975 4 : ereport(WARNING,
976 : (errmsg("VACUUM ONLY of partitioned table \"%s\" has no effect",
977 : vrel->relation->relname)));
978 :
979 8498 : ReleaseSysCache(tuple);
980 :
981 : /*
982 : * Unless the user has specified ONLY, make relation list entries for
983 : * its partitions or inheritance child tables. Note that the list
984 : * returned by find_all_inheritors() includes the passed-in OID, so we
985 : * have to skip that. There's no point in taking locks on the
986 : * individual partitions or child tables yet, and doing so would just
987 : * add unnecessary deadlock risk. For this last reason, we do not yet
988 : * check the ownership of the partitions/tables, which get added to
989 : * the list to process. Ownership will be checked later on anyway.
990 : */
991 8498 : if (include_children)
992 : {
993 8478 : List *part_oids = find_all_inheritors(relid, NoLock, NULL);
994 : ListCell *part_lc;
995 :
996 18409 : foreach(part_lc, part_oids)
997 : {
998 9931 : Oid part_oid = lfirst_oid(part_lc);
999 :
1000 9931 : if (part_oid == relid)
1001 8478 : continue; /* ignore original table */
1002 :
1003 : /*
1004 : * We omit a RangeVar since it wouldn't be appropriate to
1005 : * complain about failure to open one of these relations
1006 : * later.
1007 : */
1008 1453 : oldcontext = MemoryContextSwitchTo(vac_context);
1009 1453 : vacrels = lappend(vacrels, makeVacuumRelation(NULL,
1010 : part_oid,
1011 : vrel->va_cols));
1012 1453 : MemoryContextSwitchTo(oldcontext);
1013 : }
1014 : }
1015 :
1016 : /*
1017 : * Release lock again. This means that by the time we actually try to
1018 : * process the table, it might be gone or renamed. In the former case
1019 : * we'll silently ignore it; in the latter case we'll process it
1020 : * anyway, but we must beware that the RangeVar doesn't necessarily
1021 : * identify it anymore. This isn't ideal, perhaps, but there's little
1022 : * practical alternative, since we're typically going to commit this
1023 : * transaction and begin a new one between now and then. Moreover,
1024 : * holding locks on multiple relations would create significant risk
1025 : * of deadlock.
1026 : */
1027 8498 : UnlockRelationOid(relid, AccessShareLock);
1028 : }
1029 :
1030 109027 : return vacrels;
1031 : }
1032 :
1033 : /*
1034 : * Construct a list of VacuumRelations for all vacuumable rels in
1035 : * the current database. The list is built in vac_context.
1036 : */
1037 : static List *
1038 113 : get_all_vacuum_rels(MemoryContext vac_context, int options)
1039 : {
1040 113 : List *vacrels = NIL;
1041 : Relation pgclass;
1042 : TableScanDesc scan;
1043 : HeapTuple tuple;
1044 :
1045 113 : pgclass = table_open(RelationRelationId, AccessShareLock);
1046 :
1047 113 : scan = table_beginscan_catalog(pgclass, 0, NULL);
1048 :
1049 56375 : while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
1050 : {
1051 56262 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(tuple);
1052 : MemoryContext oldcontext;
1053 56262 : Oid relid = classForm->oid;
1054 :
1055 : /*
1056 : * We include partitioned tables here; depending on which operation is
1057 : * to be performed, caller will decide whether to process or ignore
1058 : * them.
1059 : */
1060 56262 : if (classForm->relkind != RELKIND_RELATION &&
1061 45785 : classForm->relkind != RELKIND_MATVIEW &&
1062 45753 : classForm->relkind != RELKIND_PARTITIONED_TABLE)
1063 45644 : continue;
1064 :
1065 : /* check permissions of relation */
1066 10618 : if (!vacuum_is_permitted_for_relation(relid, classForm, options))
1067 1920 : continue;
1068 :
1069 : /*
1070 : * Build VacuumRelation(s) specifying the table OIDs to be processed.
1071 : * We omit a RangeVar since it wouldn't be appropriate to complain
1072 : * about failure to open one of these relations later.
1073 : */
1074 8698 : oldcontext = MemoryContextSwitchTo(vac_context);
1075 8698 : vacrels = lappend(vacrels, makeVacuumRelation(NULL,
1076 : relid,
1077 : NIL));
1078 8698 : MemoryContextSwitchTo(oldcontext);
1079 : }
1080 :
1081 113 : table_endscan(scan);
1082 113 : table_close(pgclass, AccessShareLock);
1083 :
1084 113 : return vacrels;
1085 : }
1086 :
1087 : /*
1088 : * vacuum_get_cutoffs() -- compute OldestXmin and freeze cutoff points
1089 : *
1090 : * The target relation and VACUUM parameters are our inputs.
1091 : *
1092 : * Output parameters are the cutoffs that VACUUM caller should use.
1093 : *
1094 : * Return value indicates if vacuumlazy.c caller should make its VACUUM
1095 : * operation aggressive. An aggressive VACUUM must advance relfrozenxid up to
1096 : * FreezeLimit (at a minimum), and relminmxid up to MultiXactCutoff (at a
1097 : * minimum).
1098 : */
1099 : bool
1100 116628 : vacuum_get_cutoffs(Relation rel, const VacuumParams params,
1101 : struct VacuumCutoffs *cutoffs)
1102 : {
1103 : int freeze_min_age,
1104 : multixact_freeze_min_age,
1105 : freeze_table_age,
1106 : multixact_freeze_table_age,
1107 : effective_multixact_freeze_max_age;
1108 : TransactionId nextXID,
1109 : safeOldestXmin,
1110 : aggressiveXIDCutoff;
1111 : MultiXactId nextMXID,
1112 : safeOldestMxact,
1113 : aggressiveMXIDCutoff;
1114 :
1115 : /* Use mutable copies of freeze age parameters */
1116 116628 : freeze_min_age = params.freeze_min_age;
1117 116628 : multixact_freeze_min_age = params.multixact_freeze_min_age;
1118 116628 : freeze_table_age = params.freeze_table_age;
1119 116628 : multixact_freeze_table_age = params.multixact_freeze_table_age;
1120 :
1121 : /* Set pg_class fields in cutoffs */
1122 116628 : cutoffs->relfrozenxid = rel->rd_rel->relfrozenxid;
1123 116628 : cutoffs->relminmxid = rel->rd_rel->relminmxid;
1124 :
1125 : /*
1126 : * Acquire OldestXmin.
1127 : *
1128 : * We can always ignore processes running lazy vacuum. This is because we
1129 : * use these values only for deciding which tuples we must keep in the
1130 : * tables. Since lazy vacuum doesn't write its XID anywhere (usually no
1131 : * XID assigned), it's safe to ignore it. In theory it could be
1132 : * problematic to ignore lazy vacuums in a full vacuum, but keep in mind
1133 : * that only one vacuum process can be working on a particular table at
1134 : * any time, and that each vacuum is always an independent transaction.
1135 : */
1136 116628 : cutoffs->OldestXmin = GetOldestNonRemovableTransactionId(rel);
1137 :
1138 : Assert(TransactionIdIsNormal(cutoffs->OldestXmin));
1139 :
1140 : /* Acquire OldestMxact */
1141 116628 : cutoffs->OldestMxact = GetOldestMultiXactId();
1142 : Assert(MultiXactIdIsValid(cutoffs->OldestMxact));
1143 :
1144 : /* Acquire next XID/next MXID values used to apply age-based settings */
1145 116628 : nextXID = ReadNextTransactionId();
1146 116628 : nextMXID = ReadNextMultiXactId();
1147 :
1148 : /*
1149 : * Also compute the multixact age for which freezing is urgent. This is
1150 : * normally autovacuum_multixact_freeze_max_age, but may be less if
1151 : * multixact members are bloated.
1152 : */
1153 116628 : effective_multixact_freeze_max_age = MultiXactMemberFreezeThreshold();
1154 :
1155 : /*
1156 : * Almost ready to set freeze output parameters; check if OldestXmin or
1157 : * OldestMxact are held back to an unsafe degree before we start on that
1158 : */
1159 116628 : safeOldestXmin = nextXID - autovacuum_freeze_max_age;
1160 116628 : if (!TransactionIdIsNormal(safeOldestXmin))
1161 0 : safeOldestXmin = FirstNormalTransactionId;
1162 116628 : safeOldestMxact = nextMXID - effective_multixact_freeze_max_age;
1163 116628 : if (safeOldestMxact < FirstMultiXactId)
1164 0 : safeOldestMxact = FirstMultiXactId;
1165 116628 : if (TransactionIdPrecedes(cutoffs->OldestXmin, safeOldestXmin))
1166 81621 : ereport(WARNING,
1167 : (errmsg("cutoff for removing and freezing tuples is far in the past"),
1168 : errhint("Close open transactions soon to avoid wraparound problems.\n"
1169 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1170 116628 : if (MultiXactIdPrecedes(cutoffs->OldestMxact, safeOldestMxact))
1171 0 : ereport(WARNING,
1172 : (errmsg("cutoff for freezing multixacts is far in the past"),
1173 : errhint("Close open transactions soon to avoid wraparound problems.\n"
1174 : "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
1175 :
1176 : /*
1177 : * Determine the minimum freeze age to use: as specified by the caller, or
1178 : * vacuum_freeze_min_age, but in any case not more than half
1179 : * autovacuum_freeze_max_age, so that autovacuums to prevent XID
1180 : * wraparound won't occur too frequently.
1181 : */
1182 116628 : if (freeze_min_age < 0)
1183 7008 : freeze_min_age = vacuum_freeze_min_age;
1184 116628 : freeze_min_age = Min(freeze_min_age, autovacuum_freeze_max_age / 2);
1185 : Assert(freeze_min_age >= 0);
1186 :
1187 : /* Compute FreezeLimit, being careful to generate a normal XID */
1188 116628 : cutoffs->FreezeLimit = nextXID - freeze_min_age;
1189 116628 : if (!TransactionIdIsNormal(cutoffs->FreezeLimit))
1190 0 : cutoffs->FreezeLimit = FirstNormalTransactionId;
1191 : /* FreezeLimit must always be <= OldestXmin */
1192 116628 : if (TransactionIdPrecedes(cutoffs->OldestXmin, cutoffs->FreezeLimit))
1193 95035 : cutoffs->FreezeLimit = cutoffs->OldestXmin;
1194 :
1195 : /*
1196 : * Determine the minimum multixact freeze age to use: as specified by
1197 : * caller, or vacuum_multixact_freeze_min_age, but in any case not more
1198 : * than half effective_multixact_freeze_max_age, so that autovacuums to
1199 : * prevent MultiXact wraparound won't occur too frequently.
1200 : */
1201 116628 : if (multixact_freeze_min_age < 0)
1202 7008 : multixact_freeze_min_age = vacuum_multixact_freeze_min_age;
1203 116628 : multixact_freeze_min_age = Min(multixact_freeze_min_age,
1204 : effective_multixact_freeze_max_age / 2);
1205 : Assert(multixact_freeze_min_age >= 0);
1206 :
1207 : /* Compute MultiXactCutoff, being careful to generate a valid value */
1208 116628 : cutoffs->MultiXactCutoff = nextMXID - multixact_freeze_min_age;
1209 116628 : if (cutoffs->MultiXactCutoff < FirstMultiXactId)
1210 0 : cutoffs->MultiXactCutoff = FirstMultiXactId;
1211 : /* MultiXactCutoff must always be <= OldestMxact */
1212 116628 : if (MultiXactIdPrecedes(cutoffs->OldestMxact, cutoffs->MultiXactCutoff))
1213 3 : cutoffs->MultiXactCutoff = cutoffs->OldestMxact;
1214 :
1215 : /*
1216 : * Finally, figure out if caller needs to do an aggressive VACUUM or not.
1217 : *
1218 : * Determine the table freeze age to use: as specified by the caller, or
1219 : * the value of the vacuum_freeze_table_age GUC, but in any case not more
1220 : * than autovacuum_freeze_max_age * 0.95, so that if you have e.g nightly
1221 : * VACUUM schedule, the nightly VACUUM gets a chance to freeze XIDs before
1222 : * anti-wraparound autovacuum is launched.
1223 : */
1224 116628 : if (freeze_table_age < 0)
1225 7008 : freeze_table_age = vacuum_freeze_table_age;
1226 116628 : freeze_table_age = Min(freeze_table_age, autovacuum_freeze_max_age * 0.95);
1227 : Assert(freeze_table_age >= 0);
1228 116628 : aggressiveXIDCutoff = nextXID - freeze_table_age;
1229 116628 : if (!TransactionIdIsNormal(aggressiveXIDCutoff))
1230 0 : aggressiveXIDCutoff = FirstNormalTransactionId;
1231 116628 : if (TransactionIdPrecedesOrEquals(cutoffs->relfrozenxid,
1232 : aggressiveXIDCutoff))
1233 109614 : return true;
1234 :
1235 : /*
1236 : * Similar to the above, determine the table freeze age to use for
1237 : * multixacts: as specified by the caller, or the value of the
1238 : * vacuum_multixact_freeze_table_age GUC, but in any case not more than
1239 : * effective_multixact_freeze_max_age * 0.95, so that if you have e.g.
1240 : * nightly VACUUM schedule, the nightly VACUUM gets a chance to freeze
1241 : * multixacts before anti-wraparound autovacuum is launched.
1242 : */
1243 7014 : if (multixact_freeze_table_age < 0)
1244 6892 : multixact_freeze_table_age = vacuum_multixact_freeze_table_age;
1245 7014 : multixact_freeze_table_age =
1246 7014 : Min(multixact_freeze_table_age,
1247 : effective_multixact_freeze_max_age * 0.95);
1248 : Assert(multixact_freeze_table_age >= 0);
1249 7014 : aggressiveMXIDCutoff = nextMXID - multixact_freeze_table_age;
1250 7014 : if (aggressiveMXIDCutoff < FirstMultiXactId)
1251 0 : aggressiveMXIDCutoff = FirstMultiXactId;
1252 7014 : if (MultiXactIdPrecedesOrEquals(cutoffs->relminmxid,
1253 : aggressiveMXIDCutoff))
1254 0 : return true;
1255 :
1256 : /* Non-aggressive VACUUM */
1257 7014 : return false;
1258 : }
1259 :
1260 : /*
1261 : * vacuum_xid_failsafe_check() -- Used by VACUUM's wraparound failsafe
1262 : * mechanism to determine if its table's relfrozenxid and relminmxid are now
1263 : * dangerously far in the past.
1264 : *
1265 : * When we return true, VACUUM caller triggers the failsafe.
1266 : */
1267 : bool
1268 118416 : vacuum_xid_failsafe_check(const struct VacuumCutoffs *cutoffs)
1269 : {
1270 118416 : TransactionId relfrozenxid = cutoffs->relfrozenxid;
1271 118416 : MultiXactId relminmxid = cutoffs->relminmxid;
1272 : TransactionId xid_skip_limit;
1273 : MultiXactId multi_skip_limit;
1274 : int skip_index_vacuum;
1275 :
1276 : Assert(TransactionIdIsNormal(relfrozenxid));
1277 : Assert(MultiXactIdIsValid(relminmxid));
1278 :
1279 : /*
1280 : * Determine the index skipping age to use. In any case no less than
1281 : * autovacuum_freeze_max_age * 1.05.
1282 : */
1283 118416 : skip_index_vacuum = Max(vacuum_failsafe_age, autovacuum_freeze_max_age * 1.05);
1284 :
1285 118416 : xid_skip_limit = ReadNextTransactionId() - skip_index_vacuum;
1286 118416 : if (!TransactionIdIsNormal(xid_skip_limit))
1287 0 : xid_skip_limit = FirstNormalTransactionId;
1288 :
1289 118416 : if (TransactionIdPrecedes(relfrozenxid, xid_skip_limit))
1290 : {
1291 : /* The table's relfrozenxid is too old */
1292 22806 : return true;
1293 : }
1294 :
1295 : /*
1296 : * Similar to above, determine the index skipping age to use for
1297 : * multixact. In any case no less than autovacuum_multixact_freeze_max_age *
1298 : * 1.05.
1299 : */
1300 95610 : skip_index_vacuum = Max(vacuum_multixact_failsafe_age,
1301 : autovacuum_multixact_freeze_max_age * 1.05);
1302 :
1303 95610 : multi_skip_limit = ReadNextMultiXactId() - skip_index_vacuum;
1304 95610 : if (multi_skip_limit < FirstMultiXactId)
1305 0 : multi_skip_limit = FirstMultiXactId;
1306 :
1307 95610 : if (MultiXactIdPrecedes(relminmxid, multi_skip_limit))
1308 : {
1309 : /* The table's relminmxid is too old */
1310 0 : return true;
1311 : }
1312 :
1313 95610 : return false;
1314 : }
1315 :
1316 : /*
1317 : * vac_estimate_reltuples() -- estimate the new value for pg_class.reltuples
1318 : *
1319 : * If we scanned the whole relation then we should just use the count of
1320 : * live tuples seen; but if we did not, we should not blindly extrapolate
1321 : * from that number, since VACUUM may have scanned a quite nonrandom
1322 : * subset of the table. When we have only partial information, we take
1323 : * the old value of pg_class.reltuples/pg_class.relpages as a measurement
1324 : * of the tuple density in the unscanned pages.
1325 : *
1326 : * Note: scanned_tuples should count only *live* tuples, since
1327 : * pg_class.reltuples is defined that way.
1328 : */
1329 : double
1330 116240 : vac_estimate_reltuples(Relation relation,
1331 : BlockNumber total_pages,
1332 : BlockNumber scanned_pages,
1333 : double scanned_tuples)
1334 : {
1335 116240 : BlockNumber old_rel_pages = relation->rd_rel->relpages;
1336 116240 : double old_rel_tuples = relation->rd_rel->reltuples;
1337 : double old_density;
1338 : double unscanned_pages;
1339 : double total_tuples;
1340 :
1341 : /* If we did scan the whole table, just use the count as-is */
1342 116240 : if (scanned_pages >= total_pages)
1343 112391 : return scanned_tuples;
1344 :
1345 : /*
1346 : * When successive VACUUM commands scan the same few pages again and
1347 : * again, without anything from the table really changing, there is a risk
1348 : * that our beliefs about tuple density will gradually become distorted.
1349 : * This might be caused by vacuumlazy.c implementation details, such as
1350 : * its tendency to always scan the last heap page. Handle that here.
1351 : *
1352 : * If the relation is _exactly_ the same size according to the existing
1353 : * pg_class entry, and only a few of its pages (less than 2%) were
1354 : * scanned, keep the existing value of reltuples. Also keep the existing
1355 : * value when only a subset of rel's pages <= a single page were scanned.
1356 : *
1357 : * (Note: we might be returning -1 here.)
1358 : */
1359 3849 : if (old_rel_pages == total_pages &&
1360 3832 : scanned_pages < (double) total_pages * 0.02)
1361 2785 : return old_rel_tuples;
1362 1064 : if (scanned_pages <= 1)
1363 958 : return old_rel_tuples;
1364 :
1365 : /*
1366 : * If old density is unknown, we can't do much except scale up
1367 : * scanned_tuples to match total_pages.
1368 : */
1369 106 : if (old_rel_tuples < 0 || old_rel_pages == 0)
1370 1 : return floor((scanned_tuples / scanned_pages) * total_pages + 0.5);
1371 :
1372 : /*
1373 : * Okay, we've covered the corner cases. The normal calculation is to
1374 : * convert the old measurement to a density (tuples per page), then
1375 : * estimate the number of tuples in the unscanned pages using that figure,
1376 : * and finally add on the number of tuples in the scanned pages.
1377 : */
1378 105 : old_density = old_rel_tuples / old_rel_pages;
1379 105 : unscanned_pages = (double) total_pages - (double) scanned_pages;
1380 105 : total_tuples = old_density * unscanned_pages + scanned_tuples;
1381 105 : return floor(total_tuples + 0.5);
1382 : }
1383 :
1384 :
1385 : /*
1386 : * vac_update_relstats() -- update statistics for one relation
1387 : *
1388 : * Update the whole-relation statistics that are kept in its pg_class
1389 : * row. There are additional stats that will be updated if we are
1390 : * doing ANALYZE, but we always update these stats. This routine works
1391 : * for both index and heap relation entries in pg_class.
1392 : *
1393 : * We violate transaction semantics here by overwriting the rel's
1394 : * existing pg_class tuple with the new values. This is reasonably
1395 : * safe as long as we're sure that the new values are correct whether or
1396 : * not this transaction commits. The reason for doing this is that if
1397 : * we updated these tuples in the usual way, vacuuming pg_class itself
1398 : * wouldn't work very well --- by the time we got done with a vacuum
1399 : * cycle, most of the tuples in pg_class would've been obsoleted. Of
1400 : * course, this only works for fixed-size not-null columns, but these are.
1401 : *
1402 : * Another reason for doing it this way is that when we are in a lazy
1403 : * VACUUM and have PROC_IN_VACUUM set, we mustn't do any regular updates.
1404 : * Somebody vacuuming pg_class might think they could delete a tuple
1405 : * marked with xmin = our xid.
1406 : *
1407 : * In addition to fundamentally nontransactional statistics such as
1408 : * relpages and relallvisible, we try to maintain certain lazily-updated
1409 : * DDL flags such as relhasindex, by clearing them if no longer correct.
1410 : * It's safe to do this in VACUUM, which can't run in parallel with
1411 : * CREATE INDEX/RULE/TRIGGER and can't be part of a transaction block.
1412 : * However, it's *not* safe to do it in an ANALYZE that's within an
1413 : * outer transaction, because for example the current transaction might
1414 : * have dropped the last index; then we'd think relhasindex should be
1415 : * cleared, but if the transaction later rolls back this would be wrong.
1416 : * So we refrain from updating the DDL flags if we're inside an outer
1417 : * transaction. This is OK since postponing the flag maintenance is
1418 : * always allowable.
1419 : *
1420 : * Note: num_tuples should count only *live* tuples, since
1421 : * pg_class.reltuples is defined that way.
1422 : *
1423 : * This routine is shared by VACUUM and ANALYZE.
1424 : */
1425 : void
1426 140775 : vac_update_relstats(Relation relation,
1427 : BlockNumber num_pages, double num_tuples,
1428 : BlockNumber num_all_visible_pages,
1429 : BlockNumber num_all_frozen_pages,
1430 : bool hasindex, TransactionId frozenxid,
1431 : MultiXactId minmulti,
1432 : bool *frozenxid_updated, bool *minmulti_updated,
1433 : bool in_outer_xact)
1434 : {
1435 140775 : Oid relid = RelationGetRelid(relation);
1436 : Relation rd;
1437 : ScanKeyData key[1];
1438 : HeapTuple ctup;
1439 : void *inplace_state;
1440 : Form_pg_class pgcform;
1441 : bool dirty,
1442 : futurexid,
1443 : futuremxid;
1444 : TransactionId oldfrozenxid;
1445 : MultiXactId oldminmulti;
1446 :
1447 140775 : rd = table_open(RelationRelationId, RowExclusiveLock);
1448 :
1449 : /* Fetch a copy of the tuple to scribble on */
1450 140775 : ScanKeyInit(&key[0],
1451 : Anum_pg_class_oid,
1452 : BTEqualStrategyNumber, F_OIDEQ,
1453 : ObjectIdGetDatum(relid));
1454 140775 : systable_inplace_update_begin(rd, ClassOidIndexId, true,
1455 : NULL, 1, key, &ctup, &inplace_state);
1456 140775 : if (!HeapTupleIsValid(ctup))
1457 0 : elog(ERROR, "pg_class entry for relid %u vanished during vacuuming",
1458 : relid);
1459 140775 : pgcform = (Form_pg_class) GETSTRUCT(ctup);
1460 :
1461 : /* Apply statistical updates, if any, to copied tuple */
1462 :
1463 140775 : dirty = false;
1464 140775 : if (pgcform->relpages != (int32) num_pages)
1465 : {
1466 5443 : pgcform->relpages = (int32) num_pages;
1467 5443 : dirty = true;
1468 : }
1469 140775 : if (pgcform->reltuples != (float4) num_tuples)
1470 : {
1471 11666 : pgcform->reltuples = (float4) num_tuples;
1472 11666 : dirty = true;
1473 : }
1474 140775 : if (pgcform->relallvisible != (int32) num_all_visible_pages)
1475 : {
1476 3241 : pgcform->relallvisible = (int32) num_all_visible_pages;
1477 3241 : dirty = true;
1478 : }
1479 140775 : if (pgcform->relallfrozen != (int32) num_all_frozen_pages)
1480 : {
1481 2763 : pgcform->relallfrozen = (int32) num_all_frozen_pages;
1482 2763 : dirty = true;
1483 : }
1484 :
1485 : /* Apply DDL updates, but not inside an outer transaction (see above) */
1486 :
1487 140775 : if (!in_outer_xact)
1488 : {
1489 : /*
1490 : * If we didn't find any indexes, reset relhasindex.
1491 : */
1492 140493 : if (pgcform->relhasindex && !hasindex)
1493 : {
1494 14 : pgcform->relhasindex = false;
1495 14 : dirty = true;
1496 : }
1497 :
1498 : /* We also clear relhasrules and relhastriggers if needed */
1499 140493 : if (pgcform->relhasrules && relation->rd_rules == NULL)
1500 : {
1501 0 : pgcform->relhasrules = false;
1502 0 : dirty = true;
1503 : }
1504 140493 : if (pgcform->relhastriggers && relation->trigdesc == NULL)
1505 : {
1506 4 : pgcform->relhastriggers = false;
1507 4 : dirty = true;
1508 : }
1509 : }
1510 :
1511 : /*
1512 : * Update relfrozenxid, unless caller passed InvalidTransactionId
1513 : * indicating it has no new data.
1514 : *
1515 : * Ordinarily, we don't let relfrozenxid go backwards. However, if the
1516 : * stored relfrozenxid is "in the future" then it seems best to assume
1517 : * it's corrupt, and overwrite with the oldest remaining XID in the table.
1518 : * This should match vac_update_datfrozenxid() concerning what we consider
1519 : * to be "in the future".
1520 : */
1521 140775 : oldfrozenxid = pgcform->relfrozenxid;
1522 140775 : futurexid = false;
1523 140775 : if (frozenxid_updated)
1524 116238 : *frozenxid_updated = false;
1525 140775 : if (TransactionIdIsNormal(frozenxid) && oldfrozenxid != frozenxid)
1526 : {
1527 33158 : bool update = false;
1528 :
1529 33158 : if (TransactionIdPrecedes(oldfrozenxid, frozenxid))
1530 33075 : update = true;
1531 83 : else if (TransactionIdPrecedes(ReadNextTransactionId(), oldfrozenxid))
1532 0 : futurexid = update = true;
1533 :
1534 33158 : if (update)
1535 : {
1536 33075 : pgcform->relfrozenxid = frozenxid;
1537 33075 : dirty = true;
1538 33075 : if (frozenxid_updated)
1539 33075 : *frozenxid_updated = true;
1540 : }
1541 : }
1542 :
1543 : /* Similarly for relminmxid */
1544 140775 : oldminmulti = pgcform->relminmxid;
1545 140775 : futuremxid = false;
1546 140775 : if (minmulti_updated)
1547 116238 : *minmulti_updated = false;
1548 140775 : if (MultiXactIdIsValid(minmulti) && oldminmulti != minmulti)
1549 : {
1550 240 : bool update = false;
1551 :
1552 240 : if (MultiXactIdPrecedes(oldminmulti, minmulti))
1553 240 : update = true;
1554 0 : else if (MultiXactIdPrecedes(ReadNextMultiXactId(), oldminmulti))
1555 0 : futuremxid = update = true;
1556 :
1557 240 : if (update)
1558 : {
1559 240 : pgcform->relminmxid = minmulti;
1560 240 : dirty = true;
1561 240 : if (minmulti_updated)
1562 240 : *minmulti_updated = true;
1563 : }
1564 : }
1565 :
1566 : /* If anything changed, write out the tuple. */
1567 140775 : if (dirty)
1568 41236 : systable_inplace_update_finish(inplace_state, ctup);
1569 : else
1570 99539 : systable_inplace_update_cancel(inplace_state);
1571 :
1572 140775 : table_close(rd, RowExclusiveLock);
1573 :
1574 140775 : if (futurexid)
1575 0 : ereport(WARNING,
1576 : (errcode(ERRCODE_DATA_CORRUPTED),
1577 : errmsg_internal("overwrote invalid relfrozenxid value %u with new value %u for table \"%s\"",
1578 : oldfrozenxid, frozenxid,
1579 : RelationGetRelationName(relation))));
1580 140775 : if (futuremxid)
1581 0 : ereport(WARNING,
1582 : (errcode(ERRCODE_DATA_CORRUPTED),
1583 : errmsg_internal("overwrote invalid relminmxid value %u with new value %u for table \"%s\"",
1584 : oldminmulti, minmulti,
1585 : RelationGetRelationName(relation))));
1586 140775 : }
1587 :
1588 :
1589 : /*
1590 : * vac_update_datfrozenxid() -- update pg_database.datfrozenxid for our DB
1591 : *
1592 : * Update pg_database's datfrozenxid entry for our database to be the
1593 : * minimum of the pg_class.relfrozenxid values.
1594 : *
1595 : * Similarly, update our datminmxid to be the minimum of the
1596 : * pg_class.relminmxid values.
1597 : *
1598 : * If we are able to advance either pg_database value, also try to
1599 : * truncate pg_xact and pg_multixact.
1600 : *
1601 : * We violate transaction semantics here by overwriting the database's
1602 : * existing pg_database tuple with the new values. This is reasonably
1603 : * safe since the new values are correct whether or not this transaction
1604 : * commits. As with vac_update_relstats, this avoids leaving dead tuples
1605 : * behind after a VACUUM.
1606 : */
1607 : void
1608 2621 : vac_update_datfrozenxid(void)
1609 : {
1610 : HeapTuple tuple;
1611 : Form_pg_database dbform;
1612 : Relation relation;
1613 : SysScanDesc scan;
1614 : HeapTuple classTup;
1615 : TransactionId newFrozenXid;
1616 : MultiXactId newMinMulti;
1617 : TransactionId lastSaneFrozenXid;
1618 : MultiXactId lastSaneMinMulti;
1619 2621 : bool bogus = false;
1620 2621 : bool dirty = false;
1621 : ScanKeyData key[1];
1622 : void *inplace_state;
1623 :
1624 : /*
1625 : * Restrict this task to one backend per database. This avoids race
1626 : * conditions that would move datfrozenxid or datminmxid backward. It
1627 : * avoids calling vac_truncate_clog() with a datfrozenxid preceding a
1628 : * datfrozenxid passed to an earlier vac_truncate_clog() call.
1629 : */
1630 2621 : LockDatabaseFrozenIds(ExclusiveLock);
1631 :
1632 : /*
1633 : * Initialize the "min" calculation with
1634 : * GetOldestNonRemovableTransactionId(), which is a reasonable
1635 : * approximation to the minimum relfrozenxid for not-yet-committed
1636 : * pg_class entries for new tables; see AddNewRelationTuple(). So we
1637 : * cannot produce a wrong minimum by starting with this.
1638 : */
1639 2621 : newFrozenXid = GetOldestNonRemovableTransactionId(NULL);
1640 :
1641 : /*
1642 : * Similarly, initialize the MultiXact "min" with the value that would be
1643 : * used on pg_class for new tables. See AddNewRelationTuple().
1644 : */
1645 2621 : newMinMulti = GetOldestMultiXactId();
1646 :
1647 : /*
1648 : * Identify the latest relfrozenxid and relminmxid values that we could
1649 : * validly see during the scan. These are conservative values, but it's
1650 : * not really worth trying to be more exact.
1651 : */
1652 2621 : lastSaneFrozenXid = ReadNextTransactionId();
1653 2621 : lastSaneMinMulti = ReadNextMultiXactId();
1654 :
1655 : /*
1656 : * We must seqscan pg_class to find the minimum Xid, because there is no
1657 : * index that can help us here.
1658 : *
1659 : * See vac_truncate_clog() for the race condition to prevent.
1660 : */
1661 2621 : relation = table_open(RelationRelationId, AccessShareLock);
1662 :
1663 2621 : scan = systable_beginscan(relation, InvalidOid, false,
1664 : NULL, 0, NULL);
1665 :
1666 1508583 : while ((classTup = systable_getnext(scan)) != NULL)
1667 : {
1668 1505962 : Form_pg_class classForm = (Form_pg_class) GETSTRUCT(classTup);
1669 1505962 : volatile TransactionId *relfrozenxid_p = &classForm->relfrozenxid;
1670 1505962 : volatile TransactionId *relminmxid_p = &classForm->relminmxid;
1671 1505962 : TransactionId relfrozenxid = *relfrozenxid_p;
1672 1505962 : TransactionId relminmxid = *relminmxid_p;
1673 :
1674 : /*
1675 : * Only consider relations able to hold unfrozen XIDs (anything else
1676 : * should have InvalidTransactionId in relfrozenxid anyway).
1677 : */
1678 1505962 : if (classForm->relkind != RELKIND_RELATION &&
1679 1186463 : classForm->relkind != RELKIND_MATVIEW &&
1680 1184862 : classForm->relkind != RELKIND_TOASTVALUE)
1681 : {
1682 : Assert(!TransactionIdIsValid(relfrozenxid));
1683 : Assert(!MultiXactIdIsValid(relminmxid));
1684 1022417 : continue;
1685 : }
1686 :
1687 : /*
1688 : * Some table AMs might not need per-relation xid / multixid horizons.
1689 : * It therefore seems reasonable to allow relfrozenxid and relminmxid
1690 : * to not be set (i.e. set to their respective Invalid*Id)
1691 : * independently. Thus validate and compute horizon for each only if
1692 : * set.
1693 : *
1694 : * If things are working properly, no relation should have a
1695 : * relfrozenxid or relminmxid that is "in the future". However, such
1696 : * cases have been known to arise due to bugs in pg_upgrade. If we
1697 : * see any entries that are "in the future", chicken out and don't do
1698 : * anything. This ensures we won't truncate clog & multixact SLRUs
1699 : * before those relations have been scanned and cleaned up.
1700 : */
1701 :
1702 483545 : if (TransactionIdIsValid(relfrozenxid))
1703 : {
1704 : Assert(TransactionIdIsNormal(relfrozenxid));
1705 :
1706 : /* check for values in the future */
1707 483545 : if (TransactionIdPrecedes(lastSaneFrozenXid, relfrozenxid))
1708 : {
1709 0 : bogus = true;
1710 0 : break;
1711 : }
1712 :
1713 : /* determine new horizon */
1714 483545 : if (TransactionIdPrecedes(relfrozenxid, newFrozenXid))
1715 2751 : newFrozenXid = relfrozenxid;
1716 : }
1717 :
1718 483545 : if (MultiXactIdIsValid(relminmxid))
1719 : {
1720 : /* check for values in the future */
1721 483545 : if (MultiXactIdPrecedes(lastSaneMinMulti, relminmxid))
1722 : {
1723 0 : bogus = true;
1724 0 : break;
1725 : }
1726 :
1727 : /* determine new horizon */
1728 483545 : if (MultiXactIdPrecedes(relminmxid, newMinMulti))
1729 210 : newMinMulti = relminmxid;
1730 : }
1731 : }
1732 :
1733 : /* we're done with pg_class */
1734 2621 : systable_endscan(scan);
1735 2621 : table_close(relation, AccessShareLock);
1736 :
1737 : /* chicken out if bogus data found */
1738 2621 : if (bogus)
1739 0 : return;
1740 :
1741 : Assert(TransactionIdIsNormal(newFrozenXid));
1742 : Assert(MultiXactIdIsValid(newMinMulti));
1743 :
1744 : /* Now fetch the pg_database tuple we need to update. */
1745 2621 : relation = table_open(DatabaseRelationId, RowExclusiveLock);
1746 :
1747 : /*
1748 : * Fetch a copy of the tuple to scribble on. We could check the syscache
1749 : * tuple first. If that concluded !dirty, we'd avoid waiting on
1750 : * concurrent heap_update() and would avoid exclusive-locking the buffer.
1751 : * For now, don't optimize that.
1752 : */
1753 2621 : ScanKeyInit(&key[0],
1754 : Anum_pg_database_oid,
1755 : BTEqualStrategyNumber, F_OIDEQ,
1756 : ObjectIdGetDatum(MyDatabaseId));
1757 :
1758 2621 : systable_inplace_update_begin(relation, DatabaseOidIndexId, true,
1759 : NULL, 1, key, &tuple, &inplace_state);
1760 :
1761 2621 : if (!HeapTupleIsValid(tuple))
1762 0 : elog(ERROR, "could not find tuple for database %u", MyDatabaseId);
1763 :
1764 2621 : dbform = (Form_pg_database) GETSTRUCT(tuple);
1765 :
1766 : /*
1767 : * As in vac_update_relstats(), we ordinarily don't want to let
1768 : * datfrozenxid go backward; but if it's "in the future" then it must be
1769 : * corrupt and it seems best to overwrite it.
1770 : */
1771 2965 : if (dbform->datfrozenxid != newFrozenXid &&
1772 344 : (TransactionIdPrecedes(dbform->datfrozenxid, newFrozenXid) ||
1773 0 : TransactionIdPrecedes(lastSaneFrozenXid, dbform->datfrozenxid)))
1774 : {
1775 344 : dbform->datfrozenxid = newFrozenXid;
1776 344 : dirty = true;
1777 : }
1778 : else
1779 2277 : newFrozenXid = dbform->datfrozenxid;
1780 :
1781 : /* Ditto for datminmxid */
1782 2622 : if (dbform->datminmxid != newMinMulti &&
1783 1 : (MultiXactIdPrecedes(dbform->datminmxid, newMinMulti) ||
1784 0 : MultiXactIdPrecedes(lastSaneMinMulti, dbform->datminmxid)))
1785 : {
1786 1 : dbform->datminmxid = newMinMulti;
1787 1 : dirty = true;
1788 : }
1789 : else
1790 2620 : newMinMulti = dbform->datminmxid;
1791 :
1792 2621 : if (dirty)
1793 344 : systable_inplace_update_finish(inplace_state, tuple);
1794 : else
1795 2277 : systable_inplace_update_cancel(inplace_state);
1796 :
1797 2621 : heap_freetuple(tuple);
1798 2621 : table_close(relation, RowExclusiveLock);
1799 :
1800 : /*
1801 : * If we were able to advance datfrozenxid or datminmxid, see if we can
1802 : * truncate pg_xact and/or pg_multixact. Also do it if the shared
1803 : * XID-wrap-limit info is stale, since this action will update that too.
1804 : */
1805 2621 : if (dirty || ForceTransactionIdLimitUpdate())
1806 1078 : vac_truncate_clog(newFrozenXid, newMinMulti,
1807 : lastSaneFrozenXid, lastSaneMinMulti);
1808 : }
1809 :
1810 :
1811 : /*
1812 : * vac_truncate_clog() -- attempt to truncate the commit log
1813 : *
1814 : * Scan pg_database to determine the system-wide oldest datfrozenxid,
1815 : * and use it to truncate the transaction commit log (pg_xact).
1816 : * Also update the XID wrap limit info maintained by varsup.c.
1817 : * Likewise for datminmxid.
1818 : *
1819 : * The passed frozenXID and minMulti are the updated values for my own
1820 : * pg_database entry. They're used to initialize the "min" calculations.
1821 : * The caller also passes the "last sane" XID and MXID, since it has
1822 : * those at hand already.
1823 : *
1824 : * This routine is only invoked when we've managed to change our
1825 : * DB's datfrozenxid/datminmxid values, or we found that the shared
1826 : * XID-wrap-limit info is stale.
1827 : */
1828 : static void
1829 1078 : vac_truncate_clog(TransactionId frozenXID,
1830 : MultiXactId minMulti,
1831 : TransactionId lastSaneFrozenXid,
1832 : MultiXactId lastSaneMinMulti)
1833 : {
1834 1078 : TransactionId nextXID = ReadNextTransactionId();
1835 : Relation relation;
1836 : TableScanDesc scan;
1837 : HeapTuple tuple;
1838 : Oid oldestxid_datoid;
1839 : Oid minmulti_datoid;
1840 1078 : bool bogus = false;
1841 1078 : bool frozenAlreadyWrapped = false;
1842 :
1843 : /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */
1844 1078 : LWLockAcquire(WrapLimitsVacuumLock, LW_EXCLUSIVE);
1845 :
1846 : /* init oldest datoids to sync with my frozenXID/minMulti values */
1847 1078 : oldestxid_datoid = MyDatabaseId;
1848 1078 : minmulti_datoid = MyDatabaseId;
1849 :
1850 : /*
1851 : * Scan pg_database to compute the minimum datfrozenxid/datminmxid
1852 : *
1853 : * Since vac_update_datfrozenxid updates datfrozenxid/datminmxid in-place,
1854 : * the values could change while we look at them. Fetch each one just
1855 : * once to ensure sane behavior of the comparison logic. (Here, as in
1856 : * many other places, we assume that fetching or updating an XID in shared
1857 : * storage is atomic.)
1858 : *
1859 : * Note: we need not worry about a race condition with new entries being
1860 : * inserted by CREATE DATABASE. Any such entry will have a copy of some
1861 : * existing DB's datfrozenxid, and that source DB cannot be ours because
1862 : * of the interlock against copying a DB containing an active backend.
1863 : * Hence the new entry will not reduce the minimum. Also, if two VACUUMs
1864 : * concurrently modify the datfrozenxid's of different databases, the
1865 : * worst possible outcome is that pg_xact is not truncated as aggressively
1866 : * as it could be.
1867 : */
1868 1078 : relation = table_open(DatabaseRelationId, AccessShareLock);
1869 :
1870 1078 : scan = table_beginscan_catalog(relation, 0, NULL);
1871 :
1872 4224 : while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
1873 : {
1874 3146 : Form_pg_database dbform = (Form_pg_database) GETSTRUCT(tuple);
1875 3146 : volatile TransactionId *datfrozenxid_p = &dbform->datfrozenxid;
1876 3146 : volatile TransactionId *datminmxid_p = &dbform->datminmxid;
1877 3146 : TransactionId datfrozenxid = *datfrozenxid_p;
1878 3146 : TransactionId datminmxid = *datminmxid_p;
1879 :
1880 : Assert(TransactionIdIsNormal(datfrozenxid));
1881 : Assert(MultiXactIdIsValid(datminmxid));
1882 :
1883 : /*
1884 : * If database is in the process of getting dropped, or has been
1885 : * interrupted while doing so, no connections to it are possible
1886 : * anymore. Therefore we don't need to take it into account here.
1887 : * Which is good, because it can't be processed by autovacuum either.
1888 : */
1889 3146 : if (database_is_invalid_form((Form_pg_database) dbform))
1890 : {
1891 1 : elog(DEBUG2,
1892 : "skipping invalid database \"%s\" while computing relfrozenxid",
1893 : NameStr(dbform->datname));
1894 1 : continue;
1895 : }
1896 :
1897 : /*
1898 : * If things are working properly, no database should have a
1899 : * datfrozenxid or datminmxid that is "in the future". However, such
1900 : * cases have been known to arise due to bugs in pg_upgrade. If we
1901 : * see any entries that are "in the future", chicken out and don't do
1902 : * anything. This ensures we won't truncate clog before those
1903 : * databases have been scanned and cleaned up. (We will issue the
1904 : * "already wrapped" warning if appropriate, though.)
1905 : */
1906 6290 : if (TransactionIdPrecedes(lastSaneFrozenXid, datfrozenxid) ||
1907 3145 : MultiXactIdPrecedes(lastSaneMinMulti, datminmxid))
1908 0 : bogus = true;
1909 :
1910 3145 : if (TransactionIdPrecedes(nextXID, datfrozenxid))
1911 0 : frozenAlreadyWrapped = true;
1912 3145 : else if (TransactionIdPrecedes(datfrozenxid, frozenXID))
1913 : {
1914 273 : frozenXID = datfrozenxid;
1915 273 : oldestxid_datoid = dbform->oid;
1916 : }
1917 :
1918 3145 : if (MultiXactIdPrecedes(datminmxid, minMulti))
1919 : {
1920 2 : minMulti = datminmxid;
1921 2 : minmulti_datoid = dbform->oid;
1922 : }
1923 : }
1924 :
1925 1078 : table_endscan(scan);
1926 :
1927 1078 : table_close(relation, AccessShareLock);
1928 :
1929 : /*
1930 : * Do not truncate CLOG if we seem to have suffered wraparound already;
1931 : * the computed minimum XID might be bogus. This case should now be
1932 : * impossible due to the defenses in GetNewTransactionId, but we keep the
1933 : * test anyway.
1934 : */
1935 1078 : if (frozenAlreadyWrapped)
1936 : {
1937 0 : ereport(WARNING,
1938 : (errmsg("some databases have not been vacuumed in over 2 billion transactions"),
1939 : errdetail("You might have already suffered transaction-wraparound data loss.")));
1940 0 : LWLockRelease(WrapLimitsVacuumLock);
1941 0 : return;
1942 : }
1943 :
1944 : /* chicken out if data is bogus in any other way */
1945 1078 : if (bogus)
1946 : {
1947 0 : LWLockRelease(WrapLimitsVacuumLock);
1948 0 : return;
1949 : }
1950 :
1951 : /*
1952 : * Freeze any old transaction IDs in the async notification queue before
1953 : * CLOG truncation.
1954 : */
1955 1078 : AsyncNotifyFreezeXids(frozenXID);
1956 :
1957 : /*
1958 : * Advance the oldest value for commit timestamps before truncating, so
1959 : * that if a user requests a timestamp for a transaction we're truncating
1960 : * away right after this point, they get NULL instead of an ugly "file not
1961 : * found" error from slru.c. This doesn't matter for xact/multixact
1962 : * because they are not subject to arbitrary lookups from users.
1963 : */
1964 1078 : AdvanceOldestCommitTsXid(frozenXID);
1965 :
1966 : /*
1967 : * Truncate CLOG, multixact and CommitTs to the oldest computed value.
1968 : */
1969 1078 : TruncateCLOG(frozenXID, oldestxid_datoid);
1970 1078 : TruncateCommitTs(frozenXID);
1971 1078 : TruncateMultiXact(minMulti, minmulti_datoid);
1972 :
1973 : /*
1974 : * Update the wrap limit for GetNewTransactionId and creation of new
1975 : * MultiXactIds. Note: these functions will also signal the postmaster
1976 : * for an(other) autovac cycle if needed. XXX should we avoid possibly
1977 : * signaling twice?
1978 : */
1979 1078 : SetTransactionIdLimit(frozenXID, oldestxid_datoid);
1980 1078 : SetMultiXactIdLimit(minMulti, minmulti_datoid);
1981 :
1982 1078 : LWLockRelease(WrapLimitsVacuumLock);
1983 : }
1984 :
1985 :
1986 : /*
1987 : * vacuum_rel() -- vacuum one heap relation
1988 : *
1989 : * relid identifies the relation to vacuum. If relation is supplied,
1990 : * use the name therein for reporting any failure to open/lock the rel;
1991 : * do not use it once we've successfully opened the rel, since it might
1992 : * be stale.
1993 : *
1994 : * Returns true if it's okay to proceed with a requested ANALYZE
1995 : * operation on this table.
1996 : *
1997 : * Doing one heap at a time incurs extra overhead, since we need to
1998 : * check that the heap exists again just before we vacuum it. The
1999 : * reason that we do this is so that vacuuming can be spread across
2000 : * many small transactions. Otherwise, two-phase locking would require
2001 : * us to lock the entire database during one pass of the vacuum cleaner.
2002 : *
2003 : * At entry and exit, we are not inside a transaction.
2004 : */
2005 : static bool
2006 116731 : vacuum_rel(Oid relid, RangeVar *relation, VacuumParams params,
2007 : BufferAccessStrategy bstrategy)
2008 : {
2009 : LOCKMODE lmode;
2010 : Relation rel;
2011 : LockRelId lockrelid;
2012 : Oid priv_relid;
2013 : Oid toast_relid;
2014 : Oid save_userid;
2015 : int save_sec_context;
2016 : int save_nestlevel;
2017 : VacuumParams toast_vacuum_params;
2018 :
2019 : /*
2020 : * This function scribbles on the parameters, so make a copy early to
2021 : * avoid affecting the TOAST table (if we do end up recursing to it).
2022 : */
2023 116731 : memcpy(&toast_vacuum_params, ¶ms, sizeof(VacuumParams));
2024 :
2025 : /* Begin a transaction for vacuuming this relation */
2026 116731 : StartTransactionCommand();
2027 :
2028 116731 : if (!(params.options & VACOPT_FULL))
2029 : {
2030 : /*
2031 : * In lazy vacuum, we can set the PROC_IN_VACUUM flag, which lets
2032 : * other concurrent VACUUMs know that they can ignore this one while
2033 : * determining their OldestXmin. (The reason we don't set it during a
2034 : * full VACUUM is exactly that we may have to run user-defined
2035 : * functions for functional indexes, and we want to make sure that if
2036 : * they use the snapshot set above, any tuples it requires can't get
2037 : * removed from other tables. An index function that depends on the
2038 : * contents of other tables is arguably broken, but we won't break it
2039 : * here by violating transaction semantics.)
2040 : *
2041 : * We also set the VACUUM_FOR_WRAPAROUND flag, which is passed down by
2042 : * autovacuum; it's used to avoid canceling a vacuum that was invoked
2043 : * in an emergency.
2044 : *
2045 : * Note: these flags remain set until CommitTransaction or
2046 : * AbortTransaction. We don't want to clear them until we reset
2047 : * MyProc->xid/xmin, otherwise GetOldestNonRemovableTransactionId()
2048 : * might appear to go backwards, which is probably Not Good. (We also
2049 : * set PROC_IN_VACUUM *before* taking our own snapshot, so that our
2050 : * xmin doesn't become visible ahead of setting the flag.)
2051 : */
2052 116488 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2053 116488 : MyProc->statusFlags |= PROC_IN_VACUUM;
2054 116488 : if (params.is_wraparound)
2055 100187 : MyProc->statusFlags |= PROC_VACUUM_FOR_WRAPAROUND;
2056 116488 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2057 116488 : LWLockRelease(ProcArrayLock);
2058 : }
2059 :
2060 : /*
2061 : * Need to acquire a snapshot to prevent pg_subtrans from being truncated,
2062 : * cutoff xids in local memory wrapping around, and to have updated xmin
2063 : * horizons.
2064 : */
2065 116731 : PushActiveSnapshot(GetTransactionSnapshot());
2066 :
2067 : /*
2068 : * Check for user-requested abort. Note we want this to be inside a
2069 : * transaction, so xact.c doesn't issue useless WARNING.
2070 : */
2071 116731 : CHECK_FOR_INTERRUPTS();
2072 :
2073 : /*
2074 : * Determine the type of lock we want --- hard exclusive lock for a FULL
2075 : * vacuum, but just ShareUpdateExclusiveLock for concurrent vacuum. Either
2076 : * way, we can be sure that no other backend is vacuuming the same table.
2077 : */
2078 233462 : lmode = (params.options & VACOPT_FULL) ?
2079 116731 : AccessExclusiveLock : ShareUpdateExclusiveLock;
2080 :
2081 : /* open the relation and get the appropriate lock on it */
2082 116731 : rel = vacuum_open_relation(relid, relation, params.options,
2083 116731 : params.log_vacuum_min_duration >= 0, lmode);
2084 :
2085 : /* leave if relation could not be opened or locked */
2086 116731 : if (!rel)
2087 : {
2088 12 : PopActiveSnapshot();
2089 12 : CommitTransactionCommand();
2090 12 : return false;
2091 : }
2092 :
2093 : /*
2094 : * When recursing to a TOAST table, check privileges on the parent. NB:
2095 : * This is only safe to do because we hold a session lock on the main
2096 : * relation that prevents concurrent deletion.
2097 : */
2098 116719 : if (OidIsValid(params.toast_parent))
2099 5546 : priv_relid = params.toast_parent;
2100 : else
2101 111173 : priv_relid = RelationGetRelid(rel);
2102 :
2103 : /*
2104 : * Check if relation needs to be skipped based on privileges. This check
2105 : * happens also when building the relation list to vacuum for a manual
2106 : * operation, and needs to be done additionally here as VACUUM could
2107 : * happen across multiple transactions where privileges could have changed
2108 : * in-between. Make sure to only generate logs for VACUUM in this case.
2109 : */
2110 116719 : if (!vacuum_is_permitted_for_relation(priv_relid,
2111 : rel->rd_rel,
2112 116719 : params.options & ~VACOPT_ANALYZE))
2113 : {
2114 48 : relation_close(rel, lmode);
2115 48 : PopActiveSnapshot();
2116 48 : CommitTransactionCommand();
2117 48 : return false;
2118 : }
2119 :
2120 : /*
2121 : * Check that it's of a vacuumable relkind.
2122 : */
2123 116671 : if (rel->rd_rel->relkind != RELKIND_RELATION &&
2124 41877 : rel->rd_rel->relkind != RELKIND_MATVIEW &&
2125 41872 : rel->rd_rel->relkind != RELKIND_TOASTVALUE &&
2126 125 : rel->rd_rel->relkind != RELKIND_PARTITIONED_TABLE)
2127 : {
2128 1 : ereport(WARNING,
2129 : (errmsg("skipping \"%s\" --- cannot vacuum non-tables or special system tables",
2130 : RelationGetRelationName(rel))));
2131 1 : relation_close(rel, lmode);
2132 1 : PopActiveSnapshot();
2133 1 : CommitTransactionCommand();
2134 1 : return false;
2135 : }
2136 :
2137 : /*
2138 : * Silently ignore tables that are temp tables of other backends ---
2139 : * trying to vacuum these will lead to great unhappiness, since their
2140 : * contents are probably not up-to-date on disk. (We don't throw a
2141 : * warning here; it would just lead to chatter during a database-wide
2142 : * VACUUM.)
2143 : */
2144 116670 : if (RELATION_IS_OTHER_TEMP(rel))
2145 : {
2146 1 : relation_close(rel, lmode);
2147 1 : PopActiveSnapshot();
2148 1 : CommitTransactionCommand();
2149 1 : return false;
2150 : }
2151 :
2152 : /*
2153 : * Silently ignore partitioned tables as there is no work to be done. The
2154 : * useful work is on their child partitions, which have been queued up for
2155 : * us separately.
2156 : */
2157 116669 : if (rel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
2158 : {
2159 124 : relation_close(rel, lmode);
2160 124 : PopActiveSnapshot();
2161 124 : CommitTransactionCommand();
2162 : /* It's OK to proceed with ANALYZE on this table */
2163 124 : return true;
2164 : }
2165 :
2166 : /*
2167 : * Get a session-level lock too. This will protect our access to the
2168 : * relation across multiple transactions, so that we can vacuum the
2169 : * relation's TOAST table (if any) secure in the knowledge that no one is
2170 : * deleting the parent relation.
2171 : *
2172 : * NOTE: this cannot block, even if someone else is waiting for access,
2173 : * because the lock manager knows that both lock requests are from the
2174 : * same process.
2175 : */
2176 116545 : lockrelid = rel->rd_lockInfo.lockRelId;
2177 116545 : LockRelationIdForSession(&lockrelid, lmode);
2178 :
2179 : /*
2180 : * Set index_cleanup option based on index_cleanup reloption if it wasn't
2181 : * specified in VACUUM command, or when running in an autovacuum worker
2182 : */
2183 116545 : if (params.index_cleanup == VACOPTVALUE_UNSPECIFIED)
2184 : {
2185 : StdRdOptIndexCleanup vacuum_index_cleanup;
2186 :
2187 116402 : if (rel->rd_options == NULL)
2188 115156 : vacuum_index_cleanup = STDRD_OPTION_VACUUM_INDEX_CLEANUP_AUTO;
2189 : else
2190 1246 : vacuum_index_cleanup =
2191 1246 : ((StdRdOptions *) rel->rd_options)->vacuum_index_cleanup;
2192 :
2193 116402 : if (vacuum_index_cleanup == STDRD_OPTION_VACUUM_INDEX_CLEANUP_AUTO)
2194 116374 : params.index_cleanup = VACOPTVALUE_AUTO;
2195 28 : else if (vacuum_index_cleanup == STDRD_OPTION_VACUUM_INDEX_CLEANUP_ON)
2196 14 : params.index_cleanup = VACOPTVALUE_ENABLED;
2197 : else
2198 : {
2199 : Assert(vacuum_index_cleanup ==
2200 : STDRD_OPTION_VACUUM_INDEX_CLEANUP_OFF);
2201 14 : params.index_cleanup = VACOPTVALUE_DISABLED;
2202 : }
2203 : }
2204 :
2205 : #ifdef USE_INJECTION_POINTS
2206 116545 : if (params.index_cleanup == VACOPTVALUE_AUTO)
2207 116378 : INJECTION_POINT("vacuum-index-cleanup-auto", NULL);
2208 167 : else if (params.index_cleanup == VACOPTVALUE_DISABLED)
2209 144 : INJECTION_POINT("vacuum-index-cleanup-disabled", NULL);
2210 23 : else if (params.index_cleanup == VACOPTVALUE_ENABLED)
2211 23 : INJECTION_POINT("vacuum-index-cleanup-enabled", NULL);
2212 : #endif
2213 :
2214 : /*
2215 : * Check if the vacuum_max_eager_freeze_failure_rate table storage
2216 : * parameter was specified. This overrides the GUC value.
2217 : */
2218 116545 : if (rel->rd_options != NULL &&
2219 1254 : ((StdRdOptions *) rel->rd_options)->vacuum_max_eager_freeze_failure_rate >= 0)
2220 0 : params.max_eager_freeze_failure_rate =
2221 0 : ((StdRdOptions *) rel->rd_options)->vacuum_max_eager_freeze_failure_rate;
2222 :
2223 : /*
2224 : * Set truncate option based on truncate reloption or GUC if it wasn't
2225 : * specified in VACUUM command, or when running in an autovacuum worker
2226 : */
2227 116545 : if (params.truncate == VACOPTVALUE_UNSPECIFIED)
2228 : {
2229 116406 : StdRdOptions *opts = (StdRdOptions *) rel->rd_options;
2230 :
2231 116406 : if (opts && opts->vacuum_truncate != PG_TERNARY_UNSET)
2232 : {
2233 20 : if (opts->vacuum_truncate == PG_TERNARY_TRUE)
2234 6 : params.truncate = VACOPTVALUE_ENABLED;
2235 : else
2236 14 : params.truncate = VACOPTVALUE_DISABLED;
2237 : }
2238 116386 : else if (vacuum_truncate)
2239 116372 : params.truncate = VACOPTVALUE_ENABLED;
2240 : else
2241 14 : params.truncate = VACOPTVALUE_DISABLED;
2242 : }
2243 :
2244 : #ifdef USE_INJECTION_POINTS
2245 116545 : if (params.truncate == VACOPTVALUE_AUTO)
2246 0 : INJECTION_POINT("vacuum-truncate-auto", NULL);
2247 116545 : else if (params.truncate == VACOPTVALUE_DISABLED)
2248 166 : INJECTION_POINT("vacuum-truncate-disabled", NULL);
2249 116379 : else if (params.truncate == VACOPTVALUE_ENABLED)
2250 116379 : INJECTION_POINT("vacuum-truncate-enabled", NULL);
2251 : #endif
2252 :
2253 : /*
2254 : * Remember the relation's TOAST relation for later, if the caller asked
2255 : * us to process it. In VACUUM FULL, though, the toast table is
2256 : * automatically rebuilt by cluster_rel so we shouldn't recurse to it,
2257 : * unless PROCESS_MAIN is disabled.
2258 : */
2259 116545 : if ((params.options & VACOPT_PROCESS_TOAST) != 0 &&
2260 16130 : ((params.options & VACOPT_FULL) == 0 ||
2261 226 : (params.options & VACOPT_PROCESS_MAIN) == 0))
2262 15908 : toast_relid = rel->rd_rel->reltoastrelid;
2263 : else
2264 100637 : toast_relid = InvalidOid;
2265 :
2266 : /*
2267 : * Switch to the table owner's userid, so that any index functions are run
2268 : * as that user. Also lock down security-restricted operations and
2269 : * arrange to make GUC variable changes local to this command. (This is
2270 : * unnecessary, but harmless, for lazy VACUUM.)
2271 : */
2272 116545 : GetUserIdAndSecContext(&save_userid, &save_sec_context);
2273 116545 : SetUserIdAndSecContext(rel->rd_rel->relowner,
2274 : save_sec_context | SECURITY_RESTRICTED_OPERATION);
2275 116545 : save_nestlevel = NewGUCNestLevel();
2276 116545 : RestrictSearchPath();
2277 :
2278 : /*
2279 : * If PROCESS_MAIN is set (the default), it's time to vacuum the main
2280 : * relation. Otherwise, we can skip this part. If processing the TOAST
2281 : * table is required (e.g., PROCESS_TOAST is set), we force PROCESS_MAIN
2282 : * to be set when we recurse to the TOAST table.
2283 : */
2284 116545 : if (params.options & VACOPT_PROCESS_MAIN)
2285 : {
2286 : /*
2287 : * Do the actual work --- either FULL or "lazy" vacuum
2288 : */
2289 116460 : if (params.options & VACOPT_FULL)
2290 : {
2291 222 : ClusterParams cluster_params = {0};
2292 :
2293 222 : if ((params.options & VACOPT_VERBOSE) != 0)
2294 1 : cluster_params.options |= CLUOPT_VERBOSE;
2295 :
2296 : /* VACUUM FULL is a variant of REPACK; see cluster.c */
2297 222 : cluster_rel(REPACK_COMMAND_VACUUMFULL, rel, InvalidOid,
2298 : &cluster_params);
2299 : /* cluster_rel closes the relation, but keeps lock */
2300 :
2301 218 : rel = NULL;
2302 : }
2303 : else
2304 116238 : table_relation_vacuum(rel, params, bstrategy);
2305 : }
2306 :
2307 : /* Roll back any GUC changes executed by index functions */
2308 116541 : AtEOXact_GUC(false, save_nestlevel);
2309 :
2310 : /* Restore userid and security context */
2311 116541 : SetUserIdAndSecContext(save_userid, save_sec_context);
2312 :
2313 : /* all done with this class, but hold lock until commit */
2314 116541 : if (rel)
2315 116323 : relation_close(rel, NoLock);
2316 :
2317 : /*
2318 : * Complete the transaction and free all temporary memory used.
2319 : */
2320 116541 : PopActiveSnapshot();
2321 116541 : CommitTransactionCommand();
2322 :
2323 : /*
2324 : * If the relation has a secondary toast rel, vacuum that too while we
2325 : * still hold the session lock on the main table. Note however that
2326 : * "analyze" will not get done on the toast table. This is good, because
2327 : * the toaster always uses hardcoded index access and statistics are
2328 : * totally unimportant for toast relations.
2329 : */
2330 116541 : if (toast_relid != InvalidOid)
2331 : {
2332 : /*
2333 : * Force VACOPT_PROCESS_MAIN so vacuum_rel() processes it. Likewise,
2334 : * set toast_parent so that the privilege checks are done on the main
2335 : * relation. NB: This is only safe to do because we hold a session
2336 : * lock on the main relation that prevents concurrent deletion.
2337 : */
2338 5546 : toast_vacuum_params.options |= VACOPT_PROCESS_MAIN;
2339 5546 : toast_vacuum_params.toast_parent = relid;
2340 :
2341 5546 : vacuum_rel(toast_relid, NULL, toast_vacuum_params, bstrategy);
2342 : }
2343 :
2344 : /*
2345 : * Now release the session-level lock on the main table.
2346 : */
2347 116541 : UnlockRelationIdForSession(&lockrelid, lmode);
2348 :
2349 : /* Report that we really did it. */
2350 116541 : return true;
2351 : }
2352 :
2353 :
2354 : /*
2355 : * Open all the vacuumable indexes of the given relation, obtaining the
2356 : * specified kind of lock on each. Return an array of Relation pointers for
2357 : * the indexes into *Irel, and the number of indexes into *nindexes.
2358 : *
2359 : * We consider an index vacuumable if it is marked insertable (indisready).
2360 : * If it isn't, probably a CREATE INDEX CONCURRENTLY command failed early in
2361 : * execution, and what we have is too corrupt to be processable. We will
2362 : * vacuum even if the index isn't indisvalid; this is important because in a
2363 : * unique index, uniqueness checks will be performed anyway and had better not
2364 : * hit dangling index pointers.
2365 : */
2366 : void
2367 125455 : vac_open_indexes(Relation relation, LOCKMODE lockmode,
2368 : int *nindexes, Relation **Irel)
2369 : {
2370 : List *indexoidlist;
2371 : ListCell *indexoidscan;
2372 : int i;
2373 :
2374 : Assert(lockmode != NoLock);
2375 :
2376 125455 : indexoidlist = RelationGetIndexList(relation);
2377 :
2378 : /* allocate enough memory for all indexes */
2379 125455 : i = list_length(indexoidlist);
2380 :
2381 125455 : if (i > 0)
2382 117836 : *Irel = (Relation *) palloc(i * sizeof(Relation));
2383 : else
2384 7619 : *Irel = NULL;
2385 :
2386 : /* collect just the ready indexes */
2387 125455 : i = 0;
2388 315038 : foreach(indexoidscan, indexoidlist)
2389 : {
2390 189583 : Oid indexoid = lfirst_oid(indexoidscan);
2391 : Relation indrel;
2392 :
2393 189583 : indrel = index_open(indexoid, lockmode);
2394 189583 : if (indrel->rd_index->indisready)
2395 189583 : (*Irel)[i++] = indrel;
2396 : else
2397 0 : index_close(indrel, lockmode);
2398 : }
2399 :
2400 125455 : *nindexes = i;
2401 :
2402 125455 : list_free(indexoidlist);
2403 125455 : }
2404 :
2405 : /*
2406 : * Release the resources acquired by vac_open_indexes. Optionally release
2407 : * the locks (say NoLock to keep 'em).
2408 : */
2409 : void
2410 126040 : vac_close_indexes(int nindexes, Relation *Irel, LOCKMODE lockmode)
2411 : {
2412 126040 : if (Irel == NULL)
2413 8208 : return;
2414 :
2415 307407 : while (nindexes--)
2416 : {
2417 189575 : Relation ind = Irel[nindexes];
2418 :
2419 189575 : index_close(ind, lockmode);
2420 : }
2421 117832 : pfree(Irel);
2422 : }
2423 :
2424 : /*
2425 : * vacuum_delay_point --- check for interrupts and cost-based delay.
2426 : *
2427 : * This should be called in each major loop of VACUUM processing,
2428 : * typically once per page processed.
2429 : */
2430 : void
2431 50271890 : vacuum_delay_point(bool is_analyze)
2432 : {
2433 50271890 : double msec = 0;
2434 :
2435 : /* Always check for interrupts */
2436 50271890 : CHECK_FOR_INTERRUPTS();
2437 :
2438 50271890 : if (InterruptPending ||
2439 50271890 : (!VacuumCostActive && !ConfigReloadPending))
2440 44329081 : return;
2441 :
2442 : /*
2443 : * Autovacuum workers should reload the configuration file if requested.
2444 : * This allows changes to [autovacuum_]vacuum_cost_limit and
2445 : * [autovacuum_]vacuum_cost_delay to take effect while a table is being
2446 : * vacuumed or analyzed.
2447 : */
2448 5942809 : if (ConfigReloadPending && AmAutoVacuumWorkerProcess())
2449 : {
2450 0 : ConfigReloadPending = false;
2451 0 : ProcessConfigFile(PGC_SIGHUP);
2452 0 : VacuumUpdateCosts();
2453 : }
2454 :
2455 : /*
2456 : * If we disabled cost-based delays after reloading the config file,
2457 : * return.
2458 : */
2459 5942809 : if (!VacuumCostActive)
2460 0 : return;
2461 :
2462 : /*
2463 : * For parallel vacuum, the delay is computed based on the shared cost
2464 : * balance. See compute_parallel_delay.
2465 : */
2466 5942809 : if (VacuumSharedCostBalance != NULL)
2467 0 : msec = compute_parallel_delay();
2468 5942809 : else if (VacuumCostBalance >= vacuum_cost_limit)
2469 2862 : msec = vacuum_cost_delay * VacuumCostBalance / vacuum_cost_limit;
2470 :
2471 : /* Nap if appropriate */
2472 5942809 : if (msec > 0)
2473 : {
2474 : instr_time delay_start;
2475 :
2476 2862 : if (msec > vacuum_cost_delay * 4)
2477 7 : msec = vacuum_cost_delay * 4;
2478 :
2479 2862 : if (track_cost_delay_timing)
2480 0 : INSTR_TIME_SET_CURRENT(delay_start);
2481 :
2482 2862 : pgstat_report_wait_start(WAIT_EVENT_VACUUM_DELAY);
2483 2862 : pg_usleep(msec * 1000);
2484 2862 : pgstat_report_wait_end();
2485 :
2486 2862 : if (track_cost_delay_timing)
2487 : {
2488 : instr_time delay_end;
2489 : instr_time delay;
2490 :
2491 0 : INSTR_TIME_SET_CURRENT(delay_end);
2492 0 : INSTR_TIME_SET_ZERO(delay);
2493 0 : INSTR_TIME_ACCUM_DIFF(delay, delay_end, delay_start);
2494 :
2495 : /*
2496 : * For parallel workers, we only report the delay time every once
2497 : * in a while to avoid overloading the leader with messages and
2498 : * interrupts.
2499 : */
2500 0 : if (IsParallelWorker())
2501 : {
2502 : static instr_time last_report_time;
2503 : instr_time time_since_last_report;
2504 :
2505 : Assert(!is_analyze);
2506 :
2507 : /* Accumulate the delay time */
2508 0 : parallel_vacuum_worker_delay_ns += INSTR_TIME_GET_NANOSEC(delay);
2509 :
2510 : /* Calculate interval since last report */
2511 0 : INSTR_TIME_SET_ZERO(time_since_last_report);
2512 0 : INSTR_TIME_ACCUM_DIFF(time_since_last_report, delay_end, last_report_time);
2513 :
2514 : /* If we haven't reported in a while, do so now */
2515 0 : if (INSTR_TIME_GET_NANOSEC(time_since_last_report) >=
2516 : PARALLEL_VACUUM_DELAY_REPORT_INTERVAL_NS)
2517 : {
2518 0 : pgstat_progress_parallel_incr_param(PROGRESS_VACUUM_DELAY_TIME,
2519 : parallel_vacuum_worker_delay_ns);
2520 :
2521 : /* Reset variables */
2522 0 : last_report_time = delay_end;
2523 0 : parallel_vacuum_worker_delay_ns = 0;
2524 : }
2525 : }
2526 0 : else if (is_analyze)
2527 0 : pgstat_progress_incr_param(PROGRESS_ANALYZE_DELAY_TIME,
2528 0 : INSTR_TIME_GET_NANOSEC(delay));
2529 : else
2530 0 : pgstat_progress_incr_param(PROGRESS_VACUUM_DELAY_TIME,
2531 0 : INSTR_TIME_GET_NANOSEC(delay));
2532 : }
2533 :
2534 : /*
2535 : * We don't want to ignore postmaster death during very long vacuums
2536 : * with vacuum_cost_delay configured. We can't use the usual
2537 : * WaitLatch() approach here because we want microsecond-based sleep
2538 : * durations above.
2539 : */
2540 2862 : if (IsUnderPostmaster && !PostmasterIsAlive())
2541 0 : exit(1);
2542 :
2543 2862 : VacuumCostBalance = 0;
2544 :
2545 : /*
2546 : * Balance and update limit values for autovacuum workers. We must do
2547 : * this periodically, as the number of workers across which we are
2548 : * balancing the limit may have changed.
2549 : *
2550 : * TODO: There may be better criteria for determining when to do this
2551 : * besides "check after napping".
2552 : */
2553 2862 : AutoVacuumUpdateCostLimit();
2554 :
2555 : /* Might have gotten an interrupt while sleeping */
2556 2862 : CHECK_FOR_INTERRUPTS();
2557 : }
2558 : }
2559 :
2560 : /*
2561 : * Computes the vacuum delay for parallel workers.
2562 : *
2563 : * The basic idea of a cost-based delay for parallel vacuum is to allow each
2564 : * worker to sleep in proportion to the share of work it's done. We achieve this
2565 : * by allowing all parallel vacuum workers including the leader process to
2566 : * have a shared view of cost related parameters (mainly VacuumCostBalance).
2567 : * We allow each worker to update it as and when it has incurred any cost and
2568 : * then based on that decide whether it needs to sleep. We compute the time
2569 : * to sleep for a worker based on the cost it has incurred
2570 : * (VacuumCostBalanceLocal) and then reduce the VacuumSharedCostBalance by
2571 : * that amount. This avoids putting to sleep those workers which have done less
2572 : * I/O than other workers and therefore ensure that workers
2573 : * which are doing more I/O got throttled more.
2574 : *
2575 : * We allow a worker to sleep only if it has performed I/O above a certain
2576 : * threshold, which is calculated based on the number of active workers
2577 : * (VacuumActiveNWorkers), and the overall cost balance is more than
2578 : * VacuumCostLimit set by the system. Testing reveals that we achieve
2579 : * the required throttling if we force a worker that has done more than 50%
2580 : * of its share of work to sleep.
2581 : */
2582 : static double
2583 0 : compute_parallel_delay(void)
2584 : {
2585 0 : double msec = 0;
2586 : uint32 shared_balance;
2587 : int nworkers;
2588 :
2589 : /* Parallel vacuum must be active */
2590 : Assert(VacuumSharedCostBalance);
2591 :
2592 0 : nworkers = pg_atomic_read_u32(VacuumActiveNWorkers);
2593 :
2594 : /* At least count itself */
2595 : Assert(nworkers >= 1);
2596 :
2597 : /* Update the shared cost balance value atomically */
2598 0 : shared_balance = pg_atomic_add_fetch_u32(VacuumSharedCostBalance, VacuumCostBalance);
2599 :
2600 : /* Compute the total local balance for the current worker */
2601 0 : VacuumCostBalanceLocal += VacuumCostBalance;
2602 :
2603 0 : if ((shared_balance >= vacuum_cost_limit) &&
2604 0 : (VacuumCostBalanceLocal > 0.5 * ((double) vacuum_cost_limit / nworkers)))
2605 : {
2606 : /* Compute sleep time based on the local cost balance */
2607 0 : msec = vacuum_cost_delay * VacuumCostBalanceLocal / vacuum_cost_limit;
2608 0 : pg_atomic_sub_fetch_u32(VacuumSharedCostBalance, VacuumCostBalanceLocal);
2609 0 : VacuumCostBalanceLocal = 0;
2610 : }
2611 :
2612 : /*
2613 : * Reset the local balance as we accumulated it into the shared value.
2614 : */
2615 0 : VacuumCostBalance = 0;
2616 :
2617 0 : return msec;
2618 : }
2619 :
2620 : /*
2621 : * A wrapper function of defGetBoolean().
2622 : *
2623 : * This function returns VACOPTVALUE_ENABLED and VACOPTVALUE_DISABLED instead
2624 : * of true and false.
2625 : */
2626 : static VacOptValue
2627 182 : get_vacoptval_from_boolean(DefElem *def)
2628 : {
2629 182 : return defGetBoolean(def) ? VACOPTVALUE_ENABLED : VACOPTVALUE_DISABLED;
2630 : }
2631 :
2632 : /*
2633 : * vac_bulkdel_one_index() -- bulk-deletion for index relation.
2634 : *
2635 : * Returns bulk delete stats derived from input stats
2636 : */
2637 : IndexBulkDeleteResult *
2638 1451 : vac_bulkdel_one_index(IndexVacuumInfo *ivinfo, IndexBulkDeleteResult *istat,
2639 : TidStore *dead_items, VacDeadItemsInfo *dead_items_info)
2640 : {
2641 : /* Do bulk deletion */
2642 1451 : istat = index_bulk_delete(ivinfo, istat, vac_tid_reaped,
2643 : dead_items);
2644 :
2645 1451 : ereport(ivinfo->message_level,
2646 : (errmsg("scanned index \"%s\" to remove %" PRId64 " row versions",
2647 : RelationGetRelationName(ivinfo->index),
2648 : dead_items_info->num_items)));
2649 :
2650 1451 : return istat;
2651 : }
2652 :
2653 : /*
2654 : * vac_cleanup_one_index() -- do post-vacuum cleanup for index relation.
2655 : *
2656 : * Returns bulk delete stats derived from input stats
2657 : */
2658 : IndexBulkDeleteResult *
2659 141206 : vac_cleanup_one_index(IndexVacuumInfo *ivinfo, IndexBulkDeleteResult *istat)
2660 : {
2661 141206 : istat = index_vacuum_cleanup(ivinfo, istat);
2662 :
2663 141206 : if (istat)
2664 1603 : ereport(ivinfo->message_level,
2665 : (errmsg("index \"%s\" now contains %.0f row versions in %u pages",
2666 : RelationGetRelationName(ivinfo->index),
2667 : istat->num_index_tuples,
2668 : istat->num_pages),
2669 : errdetail("%.0f index row versions were removed.\n"
2670 : "%u index pages were newly deleted.\n"
2671 : "%u index pages are currently deleted, of which %u are currently reusable.",
2672 : istat->tuples_removed,
2673 : istat->pages_newly_deleted,
2674 : istat->pages_deleted, istat->pages_free)));
2675 :
2676 141206 : return istat;
2677 : }
2678 :
2679 : /*
2680 : * vac_tid_reaped() -- is a particular tid deletable?
2681 : *
2682 : * This has the right signature to be an IndexBulkDeleteCallback.
2683 : */
2684 : static bool
2685 3899973 : vac_tid_reaped(ItemPointer itemptr, void *state)
2686 : {
2687 3899973 : TidStore *dead_items = (TidStore *) state;
2688 :
2689 3899973 : return TidStoreIsMember(dead_items, itemptr);
2690 : }
|
