Line data Source code
1 : /* ----------
2 : * pgstat.c
3 : * Infrastructure for the cumulative statistics system.
4 : *
5 : * The cumulative statistics system accumulates statistics for different kinds
6 : * of objects. Some kinds of statistics are collected for a fixed number of
7 : * objects (most commonly 1), e.g., checkpointer statistics. Other kinds of
8 : * statistics are collected for a varying number of objects
9 : * (e.g. relations). See PgStat_KindInfo for a list of currently handled
10 : * statistics.
11 : *
12 : * Statistics are loaded from the filesystem during startup (by the startup
13 : * process), unless preceded by a crash, in which case all stats are
14 : * discarded. They are written out by the checkpointer process just before
15 : * shutting down, except when shutting down in immediate mode.
16 : *
17 : * Fixed-numbered stats are stored in plain (non-dynamic) shared memory.
18 : *
19 : * Statistics for variable-numbered objects are stored in dynamic shared
20 : * memory and can be found via a dshash hashtable. The statistics counters are
21 : * not part of the dshash entry (PgStatShared_HashEntry) directly, but are
22 : * separately allocated (PgStatShared_HashEntry->body). The separate
23 : * allocation allows different kinds of statistics to be stored in the same
24 : * hashtable without wasting space in PgStatShared_HashEntry.
25 : *
26 : * Variable-numbered stats are addressed by PgStat_HashKey while running. It
27 : * is not possible to have statistics for an object that cannot be addressed
28 : * that way at runtime. A wider identifier can be used when serializing to
29 : * disk (used for replication slot stats).
30 : *
31 : * To avoid contention on the shared hashtable, each backend has a
32 : * backend-local hashtable (pgStatEntryRefHash) in front of the shared
33 : * hashtable, containing references (PgStat_EntryRef) to shared hashtable
34 : * entries. The shared hashtable only needs to be accessed when no prior
35 : * reference is found in the local hashtable. Besides pointing to the
36 : * shared hashtable entry (PgStatShared_HashEntry) PgStat_EntryRef also
37 : * contains a pointer to the shared statistics data, as a process-local
38 : * address, to reduce access costs.
39 : *
40 : * The names for structs stored in shared memory are prefixed with
41 : * PgStatShared instead of PgStat. Each stats entry in shared memory is
42 : * protected by a dedicated lwlock.
43 : *
44 : * Most stats updates are first accumulated locally in each process as pending
45 : * entries, then later flushed to shared memory (just after commit, or by
46 : * idle-timeout). This practically eliminates contention on individual stats
47 : * entries. For most kinds of variable-numbered pending stats data is stored
48 : * in PgStat_EntryRef->pending. All entries with pending data are in the
49 : * pgStatPending list. Pending statistics updates are flushed out by
50 : * pgstat_report_stat().
51 : *
52 : * The behavior of different kinds of statistics is determined by the kind's
53 : * entry in pgstat_kind_infos, see PgStat_KindInfo for details.
54 : *
55 : * The consistency of read accesses to statistics can be configured using the
56 : * stats_fetch_consistency GUC (see config.sgml and monitoring.sgml for the
57 : * settings). When using PGSTAT_FETCH_CONSISTENCY_CACHE or
58 : * PGSTAT_FETCH_CONSISTENCY_SNAPSHOT statistics are stored in
59 : * pgStatLocal.snapshot.
60 : *
61 : * To keep things manageable, stats handling is split across several
62 : * files. Infrastructure pieces are in:
63 : * - pgstat.c - this file, to tie it all together
64 : * - pgstat_shmem.c - nearly everything dealing with shared memory, including
65 : * the maintenance of hashtable entries
66 : * - pgstat_xact.c - transactional integration, including the transactional
67 : * creation and dropping of stats entries
68 : *
69 : * Each statistics kind is handled in a dedicated file:
70 : * - pgstat_archiver.c
71 : * - pgstat_bgwriter.c
72 : * - pgstat_checkpointer.c
73 : * - pgstat_database.c
74 : * - pgstat_function.c
75 : * - pgstat_io.c
76 : * - pgstat_relation.c
77 : * - pgstat_replslot.c
78 : * - pgstat_slru.c
79 : * - pgstat_subscription.c
80 : * - pgstat_wal.c
81 : *
82 : * Whenever possible infrastructure files should not contain code related to
83 : * specific kinds of stats.
84 : *
85 : *
86 : * Copyright (c) 2001-2023, PostgreSQL Global Development Group
87 : *
88 : * IDENTIFICATION
89 : * src/backend/utils/activity/pgstat.c
90 : * ----------
91 : */
92 : #include "postgres.h"
93 :
94 : #include <unistd.h>
95 :
96 : #include "access/transam.h"
97 : #include "access/xact.h"
98 : #include "lib/dshash.h"
99 : #include "pgstat.h"
100 : #include "port/atomics.h"
101 : #include "storage/fd.h"
102 : #include "storage/ipc.h"
103 : #include "storage/lwlock.h"
104 : #include "storage/pg_shmem.h"
105 : #include "storage/shmem.h"
106 : #include "utils/guc_hooks.h"
107 : #include "utils/memutils.h"
108 : #include "utils/pgstat_internal.h"
109 : #include "utils/timestamp.h"
110 :
111 :
112 : /* ----------
113 : * Timer definitions.
114 : *
115 : * In milliseconds.
116 : * ----------
117 : */
118 :
119 : /* minimum interval non-forced stats flushes.*/
120 : #define PGSTAT_MIN_INTERVAL 1000
121 : /* how long until to block flushing pending stats updates */
122 : #define PGSTAT_MAX_INTERVAL 60000
123 : /* when to call pgstat_report_stat() again, even when idle */
124 : #define PGSTAT_IDLE_INTERVAL 10000
125 :
126 : /* ----------
127 : * Initial size hints for the hash tables used in statistics.
128 : * ----------
129 : */
130 :
131 : #define PGSTAT_SNAPSHOT_HASH_SIZE 512
132 :
133 :
134 : /* hash table for statistics snapshots entry */
135 : typedef struct PgStat_SnapshotEntry
136 : {
137 : PgStat_HashKey key;
138 : char status; /* for simplehash use */
139 : void *data; /* the stats data itself */
140 : } PgStat_SnapshotEntry;
141 :
142 :
143 : /* ----------
144 : * Backend-local Hash Table Definitions
145 : * ----------
146 : */
147 :
148 : /* for stats snapshot entries */
149 : #define SH_PREFIX pgstat_snapshot
150 : #define SH_ELEMENT_TYPE PgStat_SnapshotEntry
151 : #define SH_KEY_TYPE PgStat_HashKey
152 : #define SH_KEY key
153 : #define SH_HASH_KEY(tb, key) \
154 : pgstat_hash_hash_key(&key, sizeof(PgStat_HashKey), NULL)
155 : #define SH_EQUAL(tb, a, b) \
156 : pgstat_cmp_hash_key(&a, &b, sizeof(PgStat_HashKey), NULL) == 0
157 : #define SH_SCOPE static inline
158 : #define SH_DEFINE
159 : #define SH_DECLARE
160 : #include "lib/simplehash.h"
161 :
162 :
163 : /* ----------
164 : * Local function forward declarations
165 : * ----------
166 : */
167 :
168 : static void pgstat_write_statsfile(void);
169 : static void pgstat_read_statsfile(void);
170 :
171 : static void pgstat_reset_after_failure(void);
172 :
173 : static bool pgstat_flush_pending_entries(bool nowait);
174 :
175 : static void pgstat_prep_snapshot(void);
176 : static void pgstat_build_snapshot(void);
177 : static void pgstat_build_snapshot_fixed(PgStat_Kind kind);
178 :
179 : static inline bool pgstat_is_kind_valid(int ikind);
180 :
181 :
182 : /* ----------
183 : * GUC parameters
184 : * ----------
185 : */
186 :
187 : bool pgstat_track_counts = false;
188 : int pgstat_fetch_consistency = PGSTAT_FETCH_CONSISTENCY_CACHE;
189 :
190 :
191 : /* ----------
192 : * state shared with pgstat_*.c
193 : * ----------
194 : */
195 :
196 : PgStat_LocalState pgStatLocal;
197 :
198 :
199 : /* ----------
200 : * Local data
201 : *
202 : * NB: There should be only variables related to stats infrastructure here,
203 : * not for specific kinds of stats.
204 : * ----------
205 : */
206 :
207 : /*
208 : * Memory contexts containing the pgStatEntryRefHash table, the
209 : * pgStatSharedRef entries, and pending data respectively. Mostly to make it
210 : * easier to track / attribute memory usage.
211 : */
212 :
213 : static MemoryContext pgStatPendingContext = NULL;
214 :
215 : /*
216 : * Backend local list of PgStat_EntryRef with unflushed pending stats.
217 : *
218 : * Newly pending entries should only ever be added to the end of the list,
219 : * otherwise pgstat_flush_pending_entries() might not see them immediately.
220 : */
221 : static dlist_head pgStatPending = DLIST_STATIC_INIT(pgStatPending);
222 :
223 :
224 : /*
225 : * Force the next stats flush to happen regardless of
226 : * PGSTAT_MIN_INTERVAL. Useful in test scripts.
227 : */
228 : static bool pgStatForceNextFlush = false;
229 :
230 : /*
231 : * Force-clear existing snapshot before next use when stats_fetch_consistency
232 : * is changed.
233 : */
234 : static bool force_stats_snapshot_clear = false;
235 :
236 :
237 : /*
238 : * For assertions that check pgstat is not used before initialization / after
239 : * shutdown.
240 : */
241 : #ifdef USE_ASSERT_CHECKING
242 : static bool pgstat_is_initialized = false;
243 : static bool pgstat_is_shutdown = false;
244 : #endif
245 :
246 :
247 : /*
248 : * The different kinds of statistics.
249 : *
250 : * If reasonably possible, handling specific to one kind of stats should go
251 : * through this abstraction, rather than making more of pgstat.c aware.
252 : *
253 : * See comments for struct PgStat_KindInfo for details about the individual
254 : * fields.
255 : *
256 : * XXX: It'd be nicer to define this outside of this file. But there doesn't
257 : * seem to be a great way of doing that, given the split across multiple
258 : * files.
259 : */
260 : static const PgStat_KindInfo pgstat_kind_infos[PGSTAT_NUM_KINDS] = {
261 :
262 : /* stats kinds for variable-numbered objects */
263 :
264 : [PGSTAT_KIND_DATABASE] = {
265 : .name = "database",
266 :
267 : .fixed_amount = false,
268 : /* so pg_stat_database entries can be seen in all databases */
269 : .accessed_across_databases = true,
270 :
271 : .shared_size = sizeof(PgStatShared_Database),
272 : .shared_data_off = offsetof(PgStatShared_Database, stats),
273 : .shared_data_len = sizeof(((PgStatShared_Database *) 0)->stats),
274 : .pending_size = sizeof(PgStat_StatDBEntry),
275 :
276 : .flush_pending_cb = pgstat_database_flush_cb,
277 : .reset_timestamp_cb = pgstat_database_reset_timestamp_cb,
278 : },
279 :
280 : [PGSTAT_KIND_RELATION] = {
281 : .name = "relation",
282 :
283 : .fixed_amount = false,
284 :
285 : .shared_size = sizeof(PgStatShared_Relation),
286 : .shared_data_off = offsetof(PgStatShared_Relation, stats),
287 : .shared_data_len = sizeof(((PgStatShared_Relation *) 0)->stats),
288 : .pending_size = sizeof(PgStat_TableStatus),
289 :
290 : .flush_pending_cb = pgstat_relation_flush_cb,
291 : .delete_pending_cb = pgstat_relation_delete_pending_cb,
292 : },
293 :
294 : [PGSTAT_KIND_FUNCTION] = {
295 : .name = "function",
296 :
297 : .fixed_amount = false,
298 :
299 : .shared_size = sizeof(PgStatShared_Function),
300 : .shared_data_off = offsetof(PgStatShared_Function, stats),
301 : .shared_data_len = sizeof(((PgStatShared_Function *) 0)->stats),
302 : .pending_size = sizeof(PgStat_FunctionCounts),
303 :
304 : .flush_pending_cb = pgstat_function_flush_cb,
305 : },
306 :
307 : [PGSTAT_KIND_REPLSLOT] = {
308 : .name = "replslot",
309 :
310 : .fixed_amount = false,
311 :
312 : .accessed_across_databases = true,
313 : .named_on_disk = true,
314 :
315 : .shared_size = sizeof(PgStatShared_ReplSlot),
316 : .shared_data_off = offsetof(PgStatShared_ReplSlot, stats),
317 : .shared_data_len = sizeof(((PgStatShared_ReplSlot *) 0)->stats),
318 :
319 : .reset_timestamp_cb = pgstat_replslot_reset_timestamp_cb,
320 : .to_serialized_name = pgstat_replslot_to_serialized_name_cb,
321 : .from_serialized_name = pgstat_replslot_from_serialized_name_cb,
322 : },
323 :
324 : [PGSTAT_KIND_SUBSCRIPTION] = {
325 : .name = "subscription",
326 :
327 : .fixed_amount = false,
328 : /* so pg_stat_subscription_stats entries can be seen in all databases */
329 : .accessed_across_databases = true,
330 :
331 : .shared_size = sizeof(PgStatShared_Subscription),
332 : .shared_data_off = offsetof(PgStatShared_Subscription, stats),
333 : .shared_data_len = sizeof(((PgStatShared_Subscription *) 0)->stats),
334 : .pending_size = sizeof(PgStat_BackendSubEntry),
335 :
336 : .flush_pending_cb = pgstat_subscription_flush_cb,
337 : .reset_timestamp_cb = pgstat_subscription_reset_timestamp_cb,
338 : },
339 :
340 :
341 : /* stats for fixed-numbered (mostly 1) objects */
342 :
343 : [PGSTAT_KIND_ARCHIVER] = {
344 : .name = "archiver",
345 :
346 : .fixed_amount = true,
347 :
348 : .reset_all_cb = pgstat_archiver_reset_all_cb,
349 : .snapshot_cb = pgstat_archiver_snapshot_cb,
350 : },
351 :
352 : [PGSTAT_KIND_BGWRITER] = {
353 : .name = "bgwriter",
354 :
355 : .fixed_amount = true,
356 :
357 : .reset_all_cb = pgstat_bgwriter_reset_all_cb,
358 : .snapshot_cb = pgstat_bgwriter_snapshot_cb,
359 : },
360 :
361 : [PGSTAT_KIND_CHECKPOINTER] = {
362 : .name = "checkpointer",
363 :
364 : .fixed_amount = true,
365 :
366 : .reset_all_cb = pgstat_checkpointer_reset_all_cb,
367 : .snapshot_cb = pgstat_checkpointer_snapshot_cb,
368 : },
369 :
370 : [PGSTAT_KIND_IO] = {
371 : .name = "io",
372 :
373 : .fixed_amount = true,
374 :
375 : .reset_all_cb = pgstat_io_reset_all_cb,
376 : .snapshot_cb = pgstat_io_snapshot_cb,
377 : },
378 :
379 : [PGSTAT_KIND_SLRU] = {
380 : .name = "slru",
381 :
382 : .fixed_amount = true,
383 :
384 : .reset_all_cb = pgstat_slru_reset_all_cb,
385 : .snapshot_cb = pgstat_slru_snapshot_cb,
386 : },
387 :
388 : [PGSTAT_KIND_WAL] = {
389 : .name = "wal",
390 :
391 : .fixed_amount = true,
392 :
393 : .reset_all_cb = pgstat_wal_reset_all_cb,
394 : .snapshot_cb = pgstat_wal_snapshot_cb,
395 : },
396 : };
397 :
398 :
399 : /* ------------------------------------------------------------
400 : * Functions managing the state of the stats system for all backends.
401 : * ------------------------------------------------------------
402 : */
403 :
404 : /*
405 : * Read on-disk stats into memory at server start.
406 : *
407 : * Should only be called by the startup process or in single user mode.
408 : */
409 : void
410 2066 : pgstat_restore_stats(void)
411 : {
412 2066 : pgstat_read_statsfile();
413 2066 : }
414 :
415 : /*
416 : * Remove the stats file. This is currently used only if WAL recovery is
417 : * needed after a crash.
418 : *
419 : * Should only be called by the startup process or in single user mode.
420 : */
421 : void
422 272 : pgstat_discard_stats(void)
423 : {
424 : int ret;
425 :
426 : /* NB: this needs to be done even in single user mode */
427 :
428 272 : ret = unlink(PGSTAT_STAT_PERMANENT_FILENAME);
429 272 : if (ret != 0)
430 : {
431 270 : if (errno == ENOENT)
432 270 : elog(DEBUG2,
433 : "didn't need to unlink permanent stats file \"%s\" - didn't exist",
434 : PGSTAT_STAT_PERMANENT_FILENAME);
435 : else
436 0 : ereport(LOG,
437 : (errcode_for_file_access(),
438 : errmsg("could not unlink permanent statistics file \"%s\": %m",
439 : PGSTAT_STAT_PERMANENT_FILENAME)));
440 : }
441 : else
442 : {
443 2 : ereport(DEBUG2,
444 : (errcode_for_file_access(),
445 : errmsg_internal("unlinked permanent statistics file \"%s\"",
446 : PGSTAT_STAT_PERMANENT_FILENAME)));
447 : }
448 :
449 : /*
450 : * Reset stats contents. This will set reset timestamps of fixed-numbered
451 : * stats to the current time (no variable stats exist).
452 : */
453 272 : pgstat_reset_after_failure();
454 272 : }
455 :
456 : /*
457 : * pgstat_before_server_shutdown() needs to be called by exactly one process
458 : * during regular server shutdowns. Otherwise all stats will be lost.
459 : *
460 : * We currently only write out stats for proc_exit(0). We might want to change
461 : * that at some point... But right now pgstat_discard_stats() would be called
462 : * during the start after a disorderly shutdown, anyway.
463 : */
464 : void
465 1924 : pgstat_before_server_shutdown(int code, Datum arg)
466 : {
467 : Assert(pgStatLocal.shmem != NULL);
468 : Assert(!pgStatLocal.shmem->is_shutdown);
469 :
470 : /*
471 : * Stats should only be reported after pgstat_initialize() and before
472 : * pgstat_shutdown(). This is a convenient point to catch most violations
473 : * of this rule.
474 : */
475 : Assert(pgstat_is_initialized && !pgstat_is_shutdown);
476 :
477 : /* flush out our own pending changes before writing out */
478 1924 : pgstat_report_stat(true);
479 :
480 : /*
481 : * Only write out file during normal shutdown. Don't even signal that
482 : * we've shutdown during irregular shutdowns, because the shutdown
483 : * sequence isn't coordinated to ensure this backend shuts down last.
484 : */
485 1924 : if (code == 0)
486 : {
487 1914 : pgStatLocal.shmem->is_shutdown = true;
488 1914 : pgstat_write_statsfile();
489 : }
490 1924 : }
491 :
492 :
493 : /* ------------------------------------------------------------
494 : * Backend initialization / shutdown functions
495 : * ------------------------------------------------------------
496 : */
497 :
498 : /*
499 : * Shut down a single backend's statistics reporting at process exit.
500 : *
501 : * Flush out any remaining statistics counts. Without this, operations
502 : * triggered during backend exit (such as temp table deletions) won't be
503 : * counted.
504 : */
505 : static void
506 27400 : pgstat_shutdown_hook(int code, Datum arg)
507 : {
508 : Assert(!pgstat_is_shutdown);
509 : Assert(IsUnderPostmaster || !IsPostmasterEnvironment);
510 :
511 : /*
512 : * If we got as far as discovering our own database ID, we can flush out
513 : * what we did so far. Otherwise, we'd be reporting an invalid database
514 : * ID, so forget it. (This means that accesses to pg_database during
515 : * failed backend starts might never get counted.)
516 : */
517 27400 : if (OidIsValid(MyDatabaseId))
518 21848 : pgstat_report_disconnect(MyDatabaseId);
519 :
520 27400 : pgstat_report_stat(true);
521 :
522 : /* there shouldn't be any pending changes left */
523 : Assert(dlist_is_empty(&pgStatPending));
524 27400 : dlist_init(&pgStatPending);
525 :
526 27400 : pgstat_detach_shmem();
527 :
528 : #ifdef USE_ASSERT_CHECKING
529 : pgstat_is_shutdown = true;
530 : #endif
531 27400 : }
532 :
533 : /*
534 : * Initialize pgstats state, and set up our on-proc-exit hook. Called from
535 : * BaseInit().
536 : *
537 : * NOTE: MyDatabaseId isn't set yet; so the shutdown hook has to be careful.
538 : */
539 : void
540 27400 : pgstat_initialize(void)
541 : {
542 : Assert(!pgstat_is_initialized);
543 :
544 27400 : pgstat_attach_shmem();
545 :
546 27400 : pgstat_init_wal();
547 :
548 : /* Set up a process-exit hook to clean up */
549 27400 : before_shmem_exit(pgstat_shutdown_hook, 0);
550 :
551 : #ifdef USE_ASSERT_CHECKING
552 : pgstat_is_initialized = true;
553 : #endif
554 27400 : }
555 :
556 :
557 : /* ------------------------------------------------------------
558 : * Public functions used by backends follow
559 : * ------------------------------------------------------------
560 : */
561 :
562 : /*
563 : * Must be called by processes that performs DML: tcop/postgres.c, logical
564 : * receiver processes, SPI worker, etc. to flush pending statistics updates to
565 : * shared memory.
566 : *
567 : * Unless called with 'force', pending stats updates are flushed happen once
568 : * per PGSTAT_MIN_INTERVAL (1000ms). When not forced, stats flushes do not
569 : * block on lock acquisition, except if stats updates have been pending for
570 : * longer than PGSTAT_MAX_INTERVAL (60000ms).
571 : *
572 : * Whenever pending stats updates remain at the end of pgstat_report_stat() a
573 : * suggested idle timeout is returned. Currently this is always
574 : * PGSTAT_IDLE_INTERVAL (10000ms). Callers can use the returned time to set up
575 : * a timeout after which to call pgstat_report_stat(true), but are not
576 : * required to do so.
577 : *
578 : * Note that this is called only when not within a transaction, so it is fair
579 : * to use transaction stop time as an approximation of current time.
580 : */
581 : long
582 848572 : pgstat_report_stat(bool force)
583 : {
584 : static TimestampTz pending_since = 0;
585 : static TimestampTz last_flush = 0;
586 : bool partial_flush;
587 : TimestampTz now;
588 : bool nowait;
589 :
590 : pgstat_assert_is_up();
591 : Assert(!IsTransactionOrTransactionBlock());
592 :
593 : /* "absorb" the forced flush even if there's nothing to flush */
594 848572 : if (pgStatForceNextFlush)
595 : {
596 402 : force = true;
597 402 : pgStatForceNextFlush = false;
598 : }
599 :
600 : /* Don't expend a clock check if nothing to do */
601 848572 : if (dlist_is_empty(&pgStatPending) &&
602 13822 : !have_iostats &&
603 13582 : !have_slrustats &&
604 11994 : !pgstat_have_pending_wal())
605 : {
606 : Assert(pending_since == 0);
607 11908 : return 0;
608 : }
609 :
610 : /*
611 : * There should never be stats to report once stats are shut down. Can't
612 : * assert that before the checks above, as there is an unconditional
613 : * pgstat_report_stat() call in pgstat_shutdown_hook() - which at least
614 : * the process that ran pgstat_before_server_shutdown() will still call.
615 : */
616 : Assert(!pgStatLocal.shmem->is_shutdown);
617 :
618 836664 : now = GetCurrentTransactionStopTimestamp();
619 :
620 836664 : if (!force)
621 : {
622 1580030 : if (pending_since > 0 &&
623 770758 : TimestampDifferenceExceeds(pending_since, now, PGSTAT_MAX_INTERVAL))
624 : {
625 : /* don't keep pending updates longer than PGSTAT_MAX_INTERVAL */
626 0 : force = true;
627 : }
628 809272 : else if (last_flush > 0 &&
629 790602 : !TimestampDifferenceExceeds(last_flush, now, PGSTAT_MIN_INTERVAL))
630 : {
631 : /* don't flush too frequently */
632 789112 : if (pending_since == 0)
633 19806 : pending_since = now;
634 :
635 789112 : return PGSTAT_IDLE_INTERVAL;
636 : }
637 : }
638 :
639 47552 : pgstat_update_dbstats(now);
640 :
641 : /* don't wait for lock acquisition when !force */
642 47552 : nowait = !force;
643 :
644 47552 : partial_flush = false;
645 :
646 : /* flush database / relation / function / ... stats */
647 47552 : partial_flush |= pgstat_flush_pending_entries(nowait);
648 :
649 : /* flush IO stats */
650 47552 : partial_flush |= pgstat_flush_io(nowait);
651 :
652 : /* flush wal stats */
653 47552 : partial_flush |= pgstat_flush_wal(nowait);
654 :
655 : /* flush SLRU stats */
656 47552 : partial_flush |= pgstat_slru_flush(nowait);
657 :
658 47552 : last_flush = now;
659 :
660 : /*
661 : * If some of the pending stats could not be flushed due to lock
662 : * contention, let the caller know when to retry.
663 : */
664 47552 : if (partial_flush)
665 : {
666 : /* force should have prevented us from getting here */
667 : Assert(!force);
668 :
669 : /* remember since when stats have been pending */
670 0 : if (pending_since == 0)
671 0 : pending_since = now;
672 :
673 0 : return PGSTAT_IDLE_INTERVAL;
674 : }
675 :
676 47552 : pending_since = 0;
677 :
678 47552 : return 0;
679 : }
680 :
681 : /*
682 : * Force locally pending stats to be flushed during the next
683 : * pgstat_report_stat() call. This is useful for writing tests.
684 : */
685 : void
686 402 : pgstat_force_next_flush(void)
687 : {
688 402 : pgStatForceNextFlush = true;
689 402 : }
690 :
691 : /*
692 : * Only for use by pgstat_reset_counters()
693 : */
694 : static bool
695 21044 : match_db_entries(PgStatShared_HashEntry *entry, Datum match_data)
696 : {
697 21044 : return entry->key.dboid == DatumGetObjectId(MyDatabaseId);
698 : }
699 :
700 : /*
701 : * Reset counters for our database.
702 : *
703 : * Permission checking for this function is managed through the normal
704 : * GRANT system.
705 : */
706 : void
707 26 : pgstat_reset_counters(void)
708 : {
709 26 : TimestampTz ts = GetCurrentTimestamp();
710 :
711 26 : pgstat_reset_matching_entries(match_db_entries,
712 : ObjectIdGetDatum(MyDatabaseId),
713 : ts);
714 26 : }
715 :
716 : /*
717 : * Reset a single variable-numbered entry.
718 : *
719 : * If the stats kind is within a database, also reset the database's
720 : * stat_reset_timestamp.
721 : *
722 : * Permission checking for this function is managed through the normal
723 : * GRANT system.
724 : */
725 : void
726 32 : pgstat_reset(PgStat_Kind kind, Oid dboid, Oid objoid)
727 : {
728 32 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
729 32 : TimestampTz ts = GetCurrentTimestamp();
730 :
731 : /* not needed atm, and doesn't make sense with the current signature */
732 : Assert(!pgstat_get_kind_info(kind)->fixed_amount);
733 :
734 : /* reset the "single counter" */
735 32 : pgstat_reset_entry(kind, dboid, objoid, ts);
736 :
737 32 : if (!kind_info->accessed_across_databases)
738 10 : pgstat_reset_database_timestamp(dboid, ts);
739 32 : }
740 :
741 : /*
742 : * Reset stats for all entries of a kind.
743 : *
744 : * Permission checking for this function is managed through the normal
745 : * GRANT system.
746 : */
747 : void
748 50 : pgstat_reset_of_kind(PgStat_Kind kind)
749 : {
750 50 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
751 50 : TimestampTz ts = GetCurrentTimestamp();
752 :
753 50 : if (kind_info->fixed_amount)
754 42 : kind_info->reset_all_cb(ts);
755 : else
756 8 : pgstat_reset_entries_of_kind(kind, ts);
757 50 : }
758 :
759 :
760 : /* ------------------------------------------------------------
761 : * Fetching of stats
762 : * ------------------------------------------------------------
763 : */
764 :
765 : /*
766 : * Discard any data collected in the current transaction. Any subsequent
767 : * request will cause new snapshots to be read.
768 : *
769 : * This is also invoked during transaction commit or abort to discard
770 : * the no-longer-wanted snapshot. Updates of stats_fetch_consistency can
771 : * cause this routine to be called.
772 : */
773 : void
774 976336 : pgstat_clear_snapshot(void)
775 : {
776 : pgstat_assert_is_up();
777 :
778 976336 : memset(&pgStatLocal.snapshot.fixed_valid, 0,
779 : sizeof(pgStatLocal.snapshot.fixed_valid));
780 976336 : pgStatLocal.snapshot.stats = NULL;
781 976336 : pgStatLocal.snapshot.mode = PGSTAT_FETCH_CONSISTENCY_NONE;
782 :
783 : /* Release memory, if any was allocated */
784 976336 : if (pgStatLocal.snapshot.context)
785 : {
786 804 : MemoryContextDelete(pgStatLocal.snapshot.context);
787 :
788 : /* Reset variables */
789 804 : pgStatLocal.snapshot.context = NULL;
790 : }
791 :
792 : /*
793 : * Historically the backend_status.c facilities lived in this file, and
794 : * were reset with the same function. For now keep it that way, and
795 : * forward the reset request.
796 : */
797 976336 : pgstat_clear_backend_activity_snapshot();
798 :
799 : /* Reset this flag, as it may be possible that a cleanup was forced. */
800 976336 : force_stats_snapshot_clear = false;
801 976336 : }
802 :
803 : void *
804 16422 : pgstat_fetch_entry(PgStat_Kind kind, Oid dboid, Oid objoid)
805 : {
806 : PgStat_HashKey key;
807 : PgStat_EntryRef *entry_ref;
808 : void *stats_data;
809 16422 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
810 :
811 : /* should be called from backends */
812 : Assert(IsUnderPostmaster || !IsPostmasterEnvironment);
813 : Assert(!kind_info->fixed_amount);
814 :
815 16422 : pgstat_prep_snapshot();
816 :
817 16422 : key.kind = kind;
818 16422 : key.dboid = dboid;
819 16422 : key.objoid = objoid;
820 :
821 : /* if we need to build a full snapshot, do so */
822 16422 : if (pgstat_fetch_consistency == PGSTAT_FETCH_CONSISTENCY_SNAPSHOT)
823 460 : pgstat_build_snapshot();
824 :
825 : /* if caching is desired, look up in cache */
826 16422 : if (pgstat_fetch_consistency > PGSTAT_FETCH_CONSISTENCY_NONE)
827 : {
828 8988 : PgStat_SnapshotEntry *entry = NULL;
829 :
830 8988 : entry = pgstat_snapshot_lookup(pgStatLocal.snapshot.stats, key);
831 :
832 8988 : if (entry)
833 524 : return entry->data;
834 :
835 : /*
836 : * If we built a full snapshot and the key is not in
837 : * pgStatLocal.snapshot.stats, there are no matching stats.
838 : */
839 8464 : if (pgstat_fetch_consistency == PGSTAT_FETCH_CONSISTENCY_SNAPSHOT)
840 28 : return NULL;
841 : }
842 :
843 15870 : pgStatLocal.snapshot.mode = pgstat_fetch_consistency;
844 :
845 15870 : entry_ref = pgstat_get_entry_ref(kind, dboid, objoid, false, NULL);
846 :
847 15870 : if (entry_ref == NULL || entry_ref->shared_entry->dropped)
848 : {
849 : /* create empty entry when using PGSTAT_FETCH_CONSISTENCY_CACHE */
850 4502 : if (pgstat_fetch_consistency == PGSTAT_FETCH_CONSISTENCY_CACHE)
851 : {
852 1694 : PgStat_SnapshotEntry *entry = NULL;
853 : bool found;
854 :
855 1694 : entry = pgstat_snapshot_insert(pgStatLocal.snapshot.stats, key, &found);
856 : Assert(!found);
857 1694 : entry->data = NULL;
858 : }
859 4502 : return NULL;
860 : }
861 :
862 : /*
863 : * Allocate in caller's context for PGSTAT_FETCH_CONSISTENCY_NONE,
864 : * otherwise we could quickly end up with a fair bit of memory used due to
865 : * repeated accesses.
866 : */
867 11368 : if (pgstat_fetch_consistency == PGSTAT_FETCH_CONSISTENCY_NONE)
868 4626 : stats_data = palloc(kind_info->shared_data_len);
869 : else
870 6742 : stats_data = MemoryContextAlloc(pgStatLocal.snapshot.context,
871 6742 : kind_info->shared_data_len);
872 :
873 11368 : pgstat_lock_entry_shared(entry_ref, false);
874 22736 : memcpy(stats_data,
875 11368 : pgstat_get_entry_data(kind, entry_ref->shared_stats),
876 11368 : kind_info->shared_data_len);
877 11368 : pgstat_unlock_entry(entry_ref);
878 :
879 11368 : if (pgstat_fetch_consistency > PGSTAT_FETCH_CONSISTENCY_NONE)
880 : {
881 6742 : PgStat_SnapshotEntry *entry = NULL;
882 : bool found;
883 :
884 6742 : entry = pgstat_snapshot_insert(pgStatLocal.snapshot.stats, key, &found);
885 6742 : entry->data = stats_data;
886 : }
887 :
888 11368 : return stats_data;
889 : }
890 :
891 : /*
892 : * If a stats snapshot has been taken, return the timestamp at which that was
893 : * done, and set *have_snapshot to true. Otherwise *have_snapshot is set to
894 : * false.
895 : */
896 : TimestampTz
897 60 : pgstat_get_stat_snapshot_timestamp(bool *have_snapshot)
898 : {
899 60 : if (force_stats_snapshot_clear)
900 18 : pgstat_clear_snapshot();
901 :
902 60 : if (pgStatLocal.snapshot.mode == PGSTAT_FETCH_CONSISTENCY_SNAPSHOT)
903 : {
904 24 : *have_snapshot = true;
905 24 : return pgStatLocal.snapshot.snapshot_timestamp;
906 : }
907 :
908 36 : *have_snapshot = false;
909 :
910 36 : return 0;
911 : }
912 :
913 : bool
914 158 : pgstat_have_entry(PgStat_Kind kind, Oid dboid, Oid objoid)
915 : {
916 : /* fixed-numbered stats always exist */
917 158 : if (pgstat_get_kind_info(kind)->fixed_amount)
918 12 : return true;
919 :
920 146 : return pgstat_get_entry_ref(kind, dboid, objoid, false, NULL) != NULL;
921 : }
922 :
923 : /*
924 : * Ensure snapshot for fixed-numbered 'kind' exists.
925 : *
926 : * Typically used by the pgstat_fetch_* functions for a kind of stats, before
927 : * massaging the data into the desired format.
928 : */
929 : void
930 410 : pgstat_snapshot_fixed(PgStat_Kind kind)
931 : {
932 : Assert(pgstat_is_kind_valid(kind));
933 : Assert(pgstat_get_kind_info(kind)->fixed_amount);
934 :
935 410 : if (pgstat_fetch_consistency == PGSTAT_FETCH_CONSISTENCY_SNAPSHOT)
936 24 : pgstat_build_snapshot();
937 : else
938 386 : pgstat_build_snapshot_fixed(kind);
939 :
940 : Assert(pgStatLocal.snapshot.fixed_valid[kind]);
941 410 : }
942 :
943 : static void
944 16474 : pgstat_prep_snapshot(void)
945 : {
946 16474 : if (force_stats_snapshot_clear)
947 18 : pgstat_clear_snapshot();
948 :
949 16474 : if (pgstat_fetch_consistency == PGSTAT_FETCH_CONSISTENCY_NONE ||
950 9040 : pgStatLocal.snapshot.stats != NULL)
951 15670 : return;
952 :
953 804 : if (!pgStatLocal.snapshot.context)
954 804 : pgStatLocal.snapshot.context = AllocSetContextCreate(TopMemoryContext,
955 : "PgStat Snapshot",
956 : ALLOCSET_SMALL_SIZES);
957 :
958 804 : pgStatLocal.snapshot.stats =
959 804 : pgstat_snapshot_create(pgStatLocal.snapshot.context,
960 : PGSTAT_SNAPSHOT_HASH_SIZE,
961 : NULL);
962 : }
963 :
964 : static void
965 484 : pgstat_build_snapshot(void)
966 : {
967 : dshash_seq_status hstat;
968 : PgStatShared_HashEntry *p;
969 :
970 : /* should only be called when we need a snapshot */
971 : Assert(pgstat_fetch_consistency == PGSTAT_FETCH_CONSISTENCY_SNAPSHOT);
972 :
973 : /* snapshot already built */
974 484 : if (pgStatLocal.snapshot.mode == PGSTAT_FETCH_CONSISTENCY_SNAPSHOT)
975 432 : return;
976 :
977 52 : pgstat_prep_snapshot();
978 :
979 : Assert(pgStatLocal.snapshot.stats->members == 0);
980 :
981 52 : pgStatLocal.snapshot.snapshot_timestamp = GetCurrentTimestamp();
982 :
983 : /*
984 : * Snapshot all variable stats.
985 : */
986 52 : dshash_seq_init(&hstat, pgStatLocal.shared_hash, false);
987 47286 : while ((p = dshash_seq_next(&hstat)) != NULL)
988 : {
989 47234 : PgStat_Kind kind = p->key.kind;
990 47234 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
991 : bool found;
992 : PgStat_SnapshotEntry *entry;
993 : PgStatShared_Common *stats_data;
994 :
995 : /*
996 : * Check if the stats object should be included in the snapshot.
997 : * Unless the stats kind can be accessed from all databases (e.g.,
998 : * database stats themselves), we only include stats for the current
999 : * database or objects not associated with a database (e.g. shared
1000 : * relations).
1001 : */
1002 47234 : if (p->key.dboid != MyDatabaseId &&
1003 15540 : p->key.dboid != InvalidOid &&
1004 12624 : !kind_info->accessed_across_databases)
1005 12724 : continue;
1006 :
1007 34714 : if (p->dropped)
1008 204 : continue;
1009 :
1010 : Assert(pg_atomic_read_u32(&p->refcount) > 0);
1011 :
1012 34510 : stats_data = dsa_get_address(pgStatLocal.dsa, p->body);
1013 : Assert(stats_data);
1014 :
1015 34510 : entry = pgstat_snapshot_insert(pgStatLocal.snapshot.stats, p->key, &found);
1016 : Assert(!found);
1017 :
1018 69020 : entry->data = MemoryContextAlloc(pgStatLocal.snapshot.context,
1019 34510 : kind_info->shared_size);
1020 :
1021 : /*
1022 : * Acquire the LWLock directly instead of using
1023 : * pg_stat_lock_entry_shared() which requires a reference.
1024 : */
1025 34510 : LWLockAcquire(&stats_data->lock, LW_SHARED);
1026 69020 : memcpy(entry->data,
1027 34510 : pgstat_get_entry_data(kind, stats_data),
1028 34510 : kind_info->shared_size);
1029 34510 : LWLockRelease(&stats_data->lock);
1030 : }
1031 52 : dshash_seq_term(&hstat);
1032 :
1033 : /*
1034 : * Build snapshot of all fixed-numbered stats.
1035 : */
1036 624 : for (int kind = PGSTAT_KIND_FIRST_VALID; kind <= PGSTAT_KIND_LAST; kind++)
1037 : {
1038 572 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
1039 :
1040 572 : if (!kind_info->fixed_amount)
1041 : {
1042 : Assert(kind_info->snapshot_cb == NULL);
1043 260 : continue;
1044 : }
1045 :
1046 312 : pgstat_build_snapshot_fixed(kind);
1047 : }
1048 :
1049 52 : pgStatLocal.snapshot.mode = PGSTAT_FETCH_CONSISTENCY_SNAPSHOT;
1050 : }
1051 :
1052 : static void
1053 12182 : pgstat_build_snapshot_fixed(PgStat_Kind kind)
1054 : {
1055 12182 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
1056 :
1057 : Assert(kind_info->fixed_amount);
1058 : Assert(kind_info->snapshot_cb != NULL);
1059 :
1060 12182 : if (pgstat_fetch_consistency == PGSTAT_FETCH_CONSISTENCY_NONE)
1061 : {
1062 : /* rebuild every time */
1063 11514 : pgStatLocal.snapshot.fixed_valid[kind] = false;
1064 : }
1065 668 : else if (pgStatLocal.snapshot.fixed_valid[kind])
1066 : {
1067 : /* in snapshot mode we shouldn't get called again */
1068 : Assert(pgstat_fetch_consistency == PGSTAT_FETCH_CONSISTENCY_CACHE);
1069 12 : return;
1070 : }
1071 :
1072 : Assert(!pgStatLocal.snapshot.fixed_valid[kind]);
1073 :
1074 12170 : kind_info->snapshot_cb();
1075 :
1076 : Assert(!pgStatLocal.snapshot.fixed_valid[kind]);
1077 12170 : pgStatLocal.snapshot.fixed_valid[kind] = true;
1078 : }
1079 :
1080 :
1081 : /* ------------------------------------------------------------
1082 : * Backend-local pending stats infrastructure
1083 : * ------------------------------------------------------------
1084 : */
1085 :
1086 : /*
1087 : * Returns the appropriate PgStat_EntryRef, preparing it to receive pending
1088 : * stats if not already done.
1089 : *
1090 : * If created_entry is non-NULL, it'll be set to true if the entry is newly
1091 : * created, false otherwise.
1092 : */
1093 : PgStat_EntryRef *
1094 2862882 : pgstat_prep_pending_entry(PgStat_Kind kind, Oid dboid, Oid objoid, bool *created_entry)
1095 : {
1096 : PgStat_EntryRef *entry_ref;
1097 :
1098 : /* need to be able to flush out */
1099 : Assert(pgstat_get_kind_info(kind)->flush_pending_cb != NULL);
1100 :
1101 2862882 : if (unlikely(!pgStatPendingContext))
1102 : {
1103 25720 : pgStatPendingContext =
1104 25720 : AllocSetContextCreate(TopMemoryContext,
1105 : "PgStat Pending",
1106 : ALLOCSET_SMALL_SIZES);
1107 : }
1108 :
1109 2862882 : entry_ref = pgstat_get_entry_ref(kind, dboid, objoid,
1110 : true, created_entry);
1111 :
1112 2862882 : if (entry_ref->pending == NULL)
1113 : {
1114 1449542 : size_t entrysize = pgstat_get_kind_info(kind)->pending_size;
1115 :
1116 : Assert(entrysize != (size_t) -1);
1117 :
1118 1449542 : entry_ref->pending = MemoryContextAllocZero(pgStatPendingContext, entrysize);
1119 1449542 : dlist_push_tail(&pgStatPending, &entry_ref->pending_node);
1120 : }
1121 :
1122 2862882 : return entry_ref;
1123 : }
1124 :
1125 : /*
1126 : * Return an existing stats entry, or NULL.
1127 : *
1128 : * This should only be used for helper function for pgstatfuncs.c - outside of
1129 : * that it shouldn't be needed.
1130 : */
1131 : PgStat_EntryRef *
1132 84 : pgstat_fetch_pending_entry(PgStat_Kind kind, Oid dboid, Oid objoid)
1133 : {
1134 : PgStat_EntryRef *entry_ref;
1135 :
1136 84 : entry_ref = pgstat_get_entry_ref(kind, dboid, objoid, false, NULL);
1137 :
1138 84 : if (entry_ref == NULL || entry_ref->pending == NULL)
1139 30 : return NULL;
1140 :
1141 54 : return entry_ref;
1142 : }
1143 :
1144 : void
1145 1449542 : pgstat_delete_pending_entry(PgStat_EntryRef *entry_ref)
1146 : {
1147 1449542 : PgStat_Kind kind = entry_ref->shared_entry->key.kind;
1148 1449542 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
1149 1449542 : void *pending_data = entry_ref->pending;
1150 :
1151 : Assert(pending_data != NULL);
1152 : /* !fixed_amount stats should be handled explicitly */
1153 : Assert(!pgstat_get_kind_info(kind)->fixed_amount);
1154 :
1155 1449542 : if (kind_info->delete_pending_cb)
1156 1368572 : kind_info->delete_pending_cb(entry_ref);
1157 :
1158 1449542 : pfree(pending_data);
1159 1449542 : entry_ref->pending = NULL;
1160 :
1161 1449542 : dlist_delete(&entry_ref->pending_node);
1162 1449542 : }
1163 :
1164 : /*
1165 : * Flush out pending stats for database objects (databases, relations,
1166 : * functions).
1167 : */
1168 : static bool
1169 47552 : pgstat_flush_pending_entries(bool nowait)
1170 : {
1171 47552 : bool have_pending = false;
1172 47552 : dlist_node *cur = NULL;
1173 :
1174 : /*
1175 : * Need to be a bit careful iterating over the list of pending entries.
1176 : * Processing a pending entry may queue further pending entries to the end
1177 : * of the list that we want to process, so a simple iteration won't do.
1178 : * Further complicating matters is that we want to delete the current
1179 : * entry in each iteration from the list if we flushed successfully.
1180 : *
1181 : * So we just keep track of the next pointer in each loop iteration.
1182 : */
1183 47552 : if (!dlist_is_empty(&pgStatPending))
1184 47552 : cur = dlist_head_node(&pgStatPending);
1185 :
1186 1441608 : while (cur)
1187 : {
1188 1394056 : PgStat_EntryRef *entry_ref =
1189 1394056 : dlist_container(PgStat_EntryRef, pending_node, cur);
1190 1394056 : PgStat_HashKey key = entry_ref->shared_entry->key;
1191 1394056 : PgStat_Kind kind = key.kind;
1192 1394056 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
1193 : bool did_flush;
1194 : dlist_node *next;
1195 :
1196 : Assert(!kind_info->fixed_amount);
1197 : Assert(kind_info->flush_pending_cb != NULL);
1198 :
1199 : /* flush the stats, if possible */
1200 1394056 : did_flush = kind_info->flush_pending_cb(entry_ref, nowait);
1201 :
1202 : Assert(did_flush || nowait);
1203 :
1204 : /* determine next entry, before deleting the pending entry */
1205 1394056 : if (dlist_has_next(&pgStatPending, cur))
1206 1346504 : next = dlist_next_node(&pgStatPending, cur);
1207 : else
1208 47552 : next = NULL;
1209 :
1210 : /* if successfully flushed, remove entry */
1211 1394056 : if (did_flush)
1212 1394056 : pgstat_delete_pending_entry(entry_ref);
1213 : else
1214 0 : have_pending = true;
1215 :
1216 1394056 : cur = next;
1217 : }
1218 :
1219 : Assert(dlist_is_empty(&pgStatPending) == !have_pending);
1220 :
1221 47552 : return have_pending;
1222 : }
1223 :
1224 :
1225 : /* ------------------------------------------------------------
1226 : * Helper / infrastructure functions
1227 : * ------------------------------------------------------------
1228 : */
1229 :
1230 : PgStat_Kind
1231 164 : pgstat_get_kind_from_str(char *kind_str)
1232 : {
1233 466 : for (int kind = PGSTAT_KIND_FIRST_VALID; kind <= PGSTAT_KIND_LAST; kind++)
1234 : {
1235 460 : if (pg_strcasecmp(kind_str, pgstat_kind_infos[kind].name) == 0)
1236 158 : return kind;
1237 : }
1238 :
1239 6 : ereport(ERROR,
1240 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1241 : errmsg("invalid statistics kind: \"%s\"", kind_str)));
1242 : return PGSTAT_KIND_DATABASE; /* avoid compiler warnings */
1243 : }
1244 :
1245 : static inline bool
1246 302814 : pgstat_is_kind_valid(int ikind)
1247 : {
1248 302814 : return ikind >= PGSTAT_KIND_FIRST_VALID && ikind <= PGSTAT_KIND_LAST;
1249 : }
1250 :
1251 : const PgStat_KindInfo *
1252 6995022 : pgstat_get_kind_info(PgStat_Kind kind)
1253 : {
1254 : Assert(pgstat_is_kind_valid(kind));
1255 :
1256 6995022 : return &pgstat_kind_infos[kind];
1257 : }
1258 :
1259 : /*
1260 : * Stats should only be reported after pgstat_initialize() and before
1261 : * pgstat_shutdown(). This check is put in a few central places to catch
1262 : * violations of this rule more easily.
1263 : */
1264 : #ifdef USE_ASSERT_CHECKING
1265 : void
1266 : pgstat_assert_is_up(void)
1267 : {
1268 : Assert(pgstat_is_initialized && !pgstat_is_shutdown);
1269 : }
1270 : #endif
1271 :
1272 :
1273 : /* ------------------------------------------------------------
1274 : * reading and writing of on-disk stats file
1275 : * ------------------------------------------------------------
1276 : */
1277 :
1278 : /* helpers for pgstat_write_statsfile() */
1279 : static void
1280 893248 : write_chunk(FILE *fpout, void *ptr, size_t len)
1281 : {
1282 : int rc;
1283 :
1284 893248 : rc = fwrite(ptr, len, 1, fpout);
1285 :
1286 : /* we'll check for errors with ferror once at the end */
1287 : (void) rc;
1288 893248 : }
1289 :
1290 : #define write_chunk_s(fpout, ptr) write_chunk(fpout, ptr, sizeof(*ptr))
1291 :
1292 : /*
1293 : * This function is called in the last process that is accessing the shared
1294 : * stats so locking is not required.
1295 : */
1296 : static void
1297 1914 : pgstat_write_statsfile(void)
1298 : {
1299 : FILE *fpout;
1300 : int32 format_id;
1301 1914 : const char *tmpfile = PGSTAT_STAT_PERMANENT_TMPFILE;
1302 1914 : const char *statfile = PGSTAT_STAT_PERMANENT_FILENAME;
1303 : dshash_seq_status hstat;
1304 : PgStatShared_HashEntry *ps;
1305 :
1306 : pgstat_assert_is_up();
1307 :
1308 : /* we're shutting down, so it's ok to just override this */
1309 1914 : pgstat_fetch_consistency = PGSTAT_FETCH_CONSISTENCY_NONE;
1310 :
1311 1914 : elog(DEBUG2, "writing stats file \"%s\"", statfile);
1312 :
1313 : /*
1314 : * Open the statistics temp file to write out the current values.
1315 : */
1316 1914 : fpout = AllocateFile(tmpfile, PG_BINARY_W);
1317 1914 : if (fpout == NULL)
1318 : {
1319 0 : ereport(LOG,
1320 : (errcode_for_file_access(),
1321 : errmsg("could not open temporary statistics file \"%s\": %m",
1322 : tmpfile)));
1323 0 : return;
1324 : }
1325 :
1326 : /*
1327 : * Write the file header --- currently just a format ID.
1328 : */
1329 1914 : format_id = PGSTAT_FILE_FORMAT_ID;
1330 1914 : write_chunk_s(fpout, &format_id);
1331 :
1332 : /*
1333 : * XXX: The following could now be generalized to just iterate over
1334 : * pgstat_kind_infos instead of knowing about the different kinds of
1335 : * stats.
1336 : */
1337 :
1338 : /*
1339 : * Write archiver stats struct
1340 : */
1341 1914 : pgstat_build_snapshot_fixed(PGSTAT_KIND_ARCHIVER);
1342 1914 : write_chunk_s(fpout, &pgStatLocal.snapshot.archiver);
1343 :
1344 : /*
1345 : * Write bgwriter stats struct
1346 : */
1347 1914 : pgstat_build_snapshot_fixed(PGSTAT_KIND_BGWRITER);
1348 1914 : write_chunk_s(fpout, &pgStatLocal.snapshot.bgwriter);
1349 :
1350 : /*
1351 : * Write checkpointer stats struct
1352 : */
1353 1914 : pgstat_build_snapshot_fixed(PGSTAT_KIND_CHECKPOINTER);
1354 1914 : write_chunk_s(fpout, &pgStatLocal.snapshot.checkpointer);
1355 :
1356 : /*
1357 : * Write IO stats struct
1358 : */
1359 1914 : pgstat_build_snapshot_fixed(PGSTAT_KIND_IO);
1360 1914 : write_chunk_s(fpout, &pgStatLocal.snapshot.io);
1361 :
1362 : /*
1363 : * Write SLRU stats struct
1364 : */
1365 1914 : pgstat_build_snapshot_fixed(PGSTAT_KIND_SLRU);
1366 1914 : write_chunk_s(fpout, &pgStatLocal.snapshot.slru);
1367 :
1368 : /*
1369 : * Write WAL stats struct
1370 : */
1371 1914 : pgstat_build_snapshot_fixed(PGSTAT_KIND_WAL);
1372 1914 : write_chunk_s(fpout, &pgStatLocal.snapshot.wal);
1373 :
1374 : /*
1375 : * Walk through the stats entries
1376 : */
1377 1914 : dshash_seq_init(&hstat, pgStatLocal.shared_hash, false);
1378 441796 : while ((ps = dshash_seq_next(&hstat)) != NULL)
1379 : {
1380 : PgStatShared_Common *shstats;
1381 439882 : const PgStat_KindInfo *kind_info = NULL;
1382 :
1383 439882 : CHECK_FOR_INTERRUPTS();
1384 :
1385 : /* we may have some "dropped" entries not yet removed, skip them */
1386 : Assert(!ps->dropped);
1387 439882 : if (ps->dropped)
1388 0 : continue;
1389 :
1390 439882 : shstats = (PgStatShared_Common *) dsa_get_address(pgStatLocal.dsa, ps->body);
1391 :
1392 439882 : kind_info = pgstat_get_kind_info(ps->key.kind);
1393 :
1394 : /* if not dropped the valid-entry refcount should exist */
1395 : Assert(pg_atomic_read_u32(&ps->refcount) > 0);
1396 :
1397 439882 : if (!kind_info->to_serialized_name)
1398 : {
1399 : /* normal stats entry, identified by PgStat_HashKey */
1400 439796 : fputc('S', fpout);
1401 439796 : write_chunk_s(fpout, &ps->key);
1402 : }
1403 : else
1404 : {
1405 : /* stats entry identified by name on disk (e.g. slots) */
1406 : NameData name;
1407 :
1408 86 : kind_info->to_serialized_name(&ps->key, shstats, &name);
1409 :
1410 86 : fputc('N', fpout);
1411 86 : write_chunk_s(fpout, &ps->key.kind);
1412 86 : write_chunk_s(fpout, &name);
1413 : }
1414 :
1415 : /* Write except the header part of the entry */
1416 439882 : write_chunk(fpout,
1417 : pgstat_get_entry_data(ps->key.kind, shstats),
1418 : pgstat_get_entry_len(ps->key.kind));
1419 : }
1420 1914 : dshash_seq_term(&hstat);
1421 :
1422 : /*
1423 : * No more output to be done. Close the temp file and replace the old
1424 : * pgstat.stat with it. The ferror() check replaces testing for error
1425 : * after each individual fputc or fwrite (in write_chunk()) above.
1426 : */
1427 1914 : fputc('E', fpout);
1428 :
1429 1914 : if (ferror(fpout))
1430 : {
1431 0 : ereport(LOG,
1432 : (errcode_for_file_access(),
1433 : errmsg("could not write temporary statistics file \"%s\": %m",
1434 : tmpfile)));
1435 0 : FreeFile(fpout);
1436 0 : unlink(tmpfile);
1437 : }
1438 1914 : else if (FreeFile(fpout) < 0)
1439 : {
1440 0 : ereport(LOG,
1441 : (errcode_for_file_access(),
1442 : errmsg("could not close temporary statistics file \"%s\": %m",
1443 : tmpfile)));
1444 0 : unlink(tmpfile);
1445 : }
1446 1914 : else if (rename(tmpfile, statfile) < 0)
1447 : {
1448 0 : ereport(LOG,
1449 : (errcode_for_file_access(),
1450 : errmsg("could not rename temporary statistics file \"%s\" to \"%s\": %m",
1451 : tmpfile, statfile)));
1452 0 : unlink(tmpfile);
1453 : }
1454 : }
1455 :
1456 : /* helpers for pgstat_read_statsfile() */
1457 : static bool
1458 615868 : read_chunk(FILE *fpin, void *ptr, size_t len)
1459 : {
1460 615868 : return fread(ptr, 1, len, fpin) == len;
1461 : }
1462 :
1463 : #define read_chunk_s(fpin, ptr) read_chunk(fpin, ptr, sizeof(*ptr))
1464 :
1465 : /*
1466 : * Reads in existing statistics file into the shared stats hash.
1467 : *
1468 : * This function is called in the only process that is accessing the shared
1469 : * stats so locking is not required.
1470 : */
1471 : static void
1472 2066 : pgstat_read_statsfile(void)
1473 : {
1474 : FILE *fpin;
1475 : int32 format_id;
1476 : bool found;
1477 2066 : const char *statfile = PGSTAT_STAT_PERMANENT_FILENAME;
1478 2066 : PgStat_ShmemControl *shmem = pgStatLocal.shmem;
1479 :
1480 : /* shouldn't be called from postmaster */
1481 : Assert(IsUnderPostmaster || !IsPostmasterEnvironment);
1482 :
1483 2066 : elog(DEBUG2, "reading stats file \"%s\"", statfile);
1484 :
1485 : /*
1486 : * Try to open the stats file. If it doesn't exist, the backends simply
1487 : * returns zero for anything and statistics simply starts from scratch
1488 : * with empty counters.
1489 : *
1490 : * ENOENT is a possibility if stats collection was previously disabled or
1491 : * has not yet written the stats file for the first time. Any other
1492 : * failure condition is suspicious.
1493 : */
1494 2066 : if ((fpin = AllocateFile(statfile, PG_BINARY_R)) == NULL)
1495 : {
1496 606 : if (errno != ENOENT)
1497 0 : ereport(LOG,
1498 : (errcode_for_file_access(),
1499 : errmsg("could not open statistics file \"%s\": %m",
1500 : statfile)));
1501 606 : pgstat_reset_after_failure();
1502 606 : return;
1503 : }
1504 :
1505 : /*
1506 : * Verify it's of the expected format.
1507 : */
1508 1460 : if (!read_chunk_s(fpin, &format_id) ||
1509 1460 : format_id != PGSTAT_FILE_FORMAT_ID)
1510 2 : goto error;
1511 :
1512 : /*
1513 : * XXX: The following could now be generalized to just iterate over
1514 : * pgstat_kind_infos instead of knowing about the different kinds of
1515 : * stats.
1516 : */
1517 :
1518 : /*
1519 : * Read archiver stats struct
1520 : */
1521 1458 : if (!read_chunk_s(fpin, &shmem->archiver.stats))
1522 0 : goto error;
1523 :
1524 : /*
1525 : * Read bgwriter stats struct
1526 : */
1527 1458 : if (!read_chunk_s(fpin, &shmem->bgwriter.stats))
1528 0 : goto error;
1529 :
1530 : /*
1531 : * Read checkpointer stats struct
1532 : */
1533 1458 : if (!read_chunk_s(fpin, &shmem->checkpointer.stats))
1534 0 : goto error;
1535 :
1536 : /*
1537 : * Read IO stats struct
1538 : */
1539 1458 : if (!read_chunk_s(fpin, &shmem->io.stats))
1540 0 : goto error;
1541 :
1542 : /*
1543 : * Read SLRU stats struct
1544 : */
1545 1458 : if (!read_chunk_s(fpin, &shmem->slru.stats))
1546 0 : goto error;
1547 :
1548 : /*
1549 : * Read WAL stats struct
1550 : */
1551 1458 : if (!read_chunk_s(fpin, &shmem->wal.stats))
1552 0 : goto error;
1553 :
1554 : /*
1555 : * We found an existing statistics file. Read it and put all the hash
1556 : * table entries into place.
1557 : */
1558 : for (;;)
1559 302814 : {
1560 304272 : int t = fgetc(fpin);
1561 :
1562 304272 : switch (t)
1563 : {
1564 302814 : case 'S':
1565 : case 'N':
1566 : {
1567 : PgStat_HashKey key;
1568 : PgStatShared_HashEntry *p;
1569 : PgStatShared_Common *header;
1570 :
1571 302814 : CHECK_FOR_INTERRUPTS();
1572 :
1573 302814 : if (t == 'S')
1574 : {
1575 : /* normal stats entry, identified by PgStat_HashKey */
1576 302780 : if (!read_chunk_s(fpin, &key))
1577 0 : goto error;
1578 :
1579 302780 : if (!pgstat_is_kind_valid(key.kind))
1580 0 : goto error;
1581 : }
1582 : else
1583 : {
1584 : /* stats entry identified by name on disk (e.g. slots) */
1585 34 : const PgStat_KindInfo *kind_info = NULL;
1586 : PgStat_Kind kind;
1587 : NameData name;
1588 :
1589 34 : if (!read_chunk_s(fpin, &kind))
1590 0 : goto error;
1591 34 : if (!read_chunk_s(fpin, &name))
1592 0 : goto error;
1593 34 : if (!pgstat_is_kind_valid(kind))
1594 0 : goto error;
1595 :
1596 34 : kind_info = pgstat_get_kind_info(kind);
1597 :
1598 34 : if (!kind_info->from_serialized_name)
1599 0 : goto error;
1600 :
1601 34 : if (!kind_info->from_serialized_name(&name, &key))
1602 : {
1603 : /* skip over data for entry we don't care about */
1604 2 : if (fseek(fpin, pgstat_get_entry_len(kind), SEEK_CUR) != 0)
1605 0 : goto error;
1606 :
1607 2 : continue;
1608 : }
1609 :
1610 : Assert(key.kind == kind);
1611 : }
1612 :
1613 : /*
1614 : * This intentionally doesn't use pgstat_get_entry_ref() -
1615 : * putting all stats into checkpointer's
1616 : * pgStatEntryRefHash would be wasted effort and memory.
1617 : */
1618 302812 : p = dshash_find_or_insert(pgStatLocal.shared_hash, &key, &found);
1619 :
1620 : /* don't allow duplicate entries */
1621 302812 : if (found)
1622 : {
1623 0 : dshash_release_lock(pgStatLocal.shared_hash, p);
1624 0 : elog(WARNING, "found duplicate stats entry %d/%u/%u",
1625 : key.kind, key.dboid, key.objoid);
1626 0 : goto error;
1627 : }
1628 :
1629 302812 : header = pgstat_init_entry(key.kind, p);
1630 302812 : dshash_release_lock(pgStatLocal.shared_hash, p);
1631 :
1632 302812 : if (!read_chunk(fpin,
1633 : pgstat_get_entry_data(key.kind, header),
1634 : pgstat_get_entry_len(key.kind)))
1635 0 : goto error;
1636 :
1637 302812 : break;
1638 : }
1639 1458 : case 'E':
1640 : /* check that 'E' actually signals end of file */
1641 1458 : if (fgetc(fpin) != EOF)
1642 2 : goto error;
1643 :
1644 1456 : goto done;
1645 :
1646 0 : default:
1647 0 : goto error;
1648 : }
1649 : }
1650 :
1651 1460 : done:
1652 1460 : FreeFile(fpin);
1653 :
1654 1460 : elog(DEBUG2, "removing permanent stats file \"%s\"", statfile);
1655 1460 : unlink(statfile);
1656 :
1657 1460 : return;
1658 :
1659 4 : error:
1660 4 : ereport(LOG,
1661 : (errmsg("corrupted statistics file \"%s\"", statfile)));
1662 :
1663 4 : pgstat_reset_after_failure();
1664 :
1665 4 : goto done;
1666 : }
1667 :
1668 : /*
1669 : * Helper to reset / drop stats after a crash or after restoring stats from
1670 : * disk failed, potentially after already loading parts.
1671 : */
1672 : static void
1673 882 : pgstat_reset_after_failure(void)
1674 : {
1675 882 : TimestampTz ts = GetCurrentTimestamp();
1676 :
1677 : /* reset fixed-numbered stats */
1678 10584 : for (int kind = PGSTAT_KIND_FIRST_VALID; kind <= PGSTAT_KIND_LAST; kind++)
1679 : {
1680 9702 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
1681 :
1682 9702 : if (!kind_info->fixed_amount)
1683 4410 : continue;
1684 :
1685 5292 : kind_info->reset_all_cb(ts);
1686 : }
1687 :
1688 : /* and drop variable-numbered ones */
1689 882 : pgstat_drop_all_entries();
1690 882 : }
1691 :
1692 : /*
1693 : * GUC assign_hook for stats_fetch_consistency.
1694 : */
1695 : void
1696 4646 : assign_stats_fetch_consistency(int newval, void *extra)
1697 : {
1698 : /*
1699 : * Changing this value in a transaction may cause snapshot state
1700 : * inconsistencies, so force a clear of the current snapshot on the next
1701 : * snapshot build attempt.
1702 : */
1703 4646 : if (pgstat_fetch_consistency != newval)
1704 734 : force_stats_snapshot_clear = true;
1705 4646 : }
|