Line data Source code
1 : /* -------------------------------------------------------------------------
2 : *
3 : * pgstat_shmem.c
4 : * Storage of stats entries in shared memory
5 : *
6 : * Copyright (c) 2001-2025, PostgreSQL Global Development Group
7 : *
8 : * IDENTIFICATION
9 : * src/backend/utils/activity/pgstat_shmem.c
10 : * -------------------------------------------------------------------------
11 : */
12 :
13 : #include "postgres.h"
14 :
15 : #include "pgstat.h"
16 : #include "storage/shmem.h"
17 : #include "utils/memutils.h"
18 : #include "utils/pgstat_internal.h"
19 :
20 :
21 : #define PGSTAT_ENTRY_REF_HASH_SIZE 128
22 :
23 : /* hash table entry for finding the PgStat_EntryRef for a key */
24 : typedef struct PgStat_EntryRefHashEntry
25 : {
26 : PgStat_HashKey key; /* hash key */
27 : char status; /* for simplehash use */
28 : PgStat_EntryRef *entry_ref;
29 : } PgStat_EntryRefHashEntry;
30 :
31 :
32 : /* for references to shared statistics entries */
33 : #define SH_PREFIX pgstat_entry_ref_hash
34 : #define SH_ELEMENT_TYPE PgStat_EntryRefHashEntry
35 : #define SH_KEY_TYPE PgStat_HashKey
36 : #define SH_KEY key
37 : #define SH_HASH_KEY(tb, key) \
38 : pgstat_hash_hash_key(&key, sizeof(PgStat_HashKey), NULL)
39 : #define SH_EQUAL(tb, a, b) \
40 : pgstat_cmp_hash_key(&a, &b, sizeof(PgStat_HashKey), NULL) == 0
41 : #define SH_SCOPE static inline
42 : #define SH_DEFINE
43 : #define SH_DECLARE
44 : #include "lib/simplehash.h"
45 :
46 :
47 : static void pgstat_drop_database_and_contents(Oid dboid);
48 :
49 : static void pgstat_free_entry(PgStatShared_HashEntry *shent, dshash_seq_status *hstat);
50 :
51 : static void pgstat_release_entry_ref(PgStat_HashKey key, PgStat_EntryRef *entry_ref, bool discard_pending);
52 : static bool pgstat_need_entry_refs_gc(void);
53 : static void pgstat_gc_entry_refs(void);
54 : static void pgstat_release_all_entry_refs(bool discard_pending);
55 : typedef bool (*ReleaseMatchCB) (PgStat_EntryRefHashEntry *, Datum data);
56 : static void pgstat_release_matching_entry_refs(bool discard_pending, ReleaseMatchCB match, Datum match_data);
57 :
58 : static void pgstat_setup_memcxt(void);
59 :
60 :
61 : /* parameter for the shared hash */
62 : static const dshash_parameters dsh_params = {
63 : sizeof(PgStat_HashKey),
64 : sizeof(PgStatShared_HashEntry),
65 : pgstat_cmp_hash_key,
66 : pgstat_hash_hash_key,
67 : dshash_memcpy,
68 : LWTRANCHE_PGSTATS_HASH
69 : };
70 :
71 :
72 : /*
73 : * Backend local references to shared stats entries. If there are pending
74 : * updates to a stats entry, the PgStat_EntryRef is added to the pgStatPending
75 : * list.
76 : *
77 : * When a stats entry is dropped each backend needs to release its reference
78 : * to it before the memory can be released. To trigger that
79 : * pgStatLocal.shmem->gc_request_count is incremented - which each backend
80 : * compares to their copy of pgStatSharedRefAge on a regular basis.
81 : */
82 : static pgstat_entry_ref_hash_hash *pgStatEntryRefHash = NULL;
83 : static int pgStatSharedRefAge = 0; /* cache age of pgStatLocal.shmem */
84 :
85 : /*
86 : * Memory contexts containing the pgStatEntryRefHash table and the
87 : * pgStatSharedRef entries respectively. Kept separate to make it easier to
88 : * track / attribute memory usage.
89 : */
90 : static MemoryContext pgStatSharedRefContext = NULL;
91 : static MemoryContext pgStatEntryRefHashContext = NULL;
92 :
93 :
94 : /* ------------------------------------------------------------
95 : * Public functions called from postmaster follow
96 : * ------------------------------------------------------------
97 : */
98 :
99 : /*
100 : * The size of the shared memory allocation for stats stored in the shared
101 : * stats hash table. This allocation will be done as part of the main shared
102 : * memory, rather than dynamic shared memory, allowing it to be initialized in
103 : * postmaster.
104 : */
105 : static Size
106 11238 : pgstat_dsa_init_size(void)
107 : {
108 : Size sz;
109 :
110 : /*
111 : * The dshash header / initial buckets array needs to fit into "plain"
112 : * shared memory, but it's beneficial to not need dsm segments
113 : * immediately. A size of 256kB seems works well and is not
114 : * disproportional compared to other constant sized shared memory
115 : * allocations. NB: To avoid DSMs further, the user can configure
116 : * min_dynamic_shared_memory.
117 : */
118 11238 : sz = 256 * 1024;
119 : Assert(dsa_minimum_size() <= sz);
120 11238 : return MAXALIGN(sz);
121 : }
122 :
123 : /*
124 : * Compute shared memory space needed for cumulative statistics
125 : */
126 : Size
127 5484 : StatsShmemSize(void)
128 : {
129 : Size sz;
130 :
131 5484 : sz = MAXALIGN(sizeof(PgStat_ShmemControl));
132 5484 : sz = add_size(sz, pgstat_dsa_init_size());
133 :
134 : /* Add shared memory for all the custom fixed-numbered statistics */
135 712920 : for (PgStat_Kind kind = PGSTAT_KIND_CUSTOM_MIN; kind <= PGSTAT_KIND_CUSTOM_MAX; kind++)
136 : {
137 707436 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
138 :
139 707436 : if (!kind_info)
140 707388 : continue;
141 48 : if (!kind_info->fixed_amount)
142 24 : continue;
143 :
144 : Assert(kind_info->shared_size != 0);
145 :
146 24 : sz += MAXALIGN(kind_info->shared_size);
147 : }
148 :
149 5484 : return sz;
150 : }
151 :
152 : /*
153 : * Initialize cumulative statistics system during startup
154 : */
155 : void
156 1918 : StatsShmemInit(void)
157 : {
158 : bool found;
159 : Size sz;
160 :
161 1918 : sz = StatsShmemSize();
162 1918 : pgStatLocal.shmem = (PgStat_ShmemControl *)
163 1918 : ShmemInitStruct("Shared Memory Stats", sz, &found);
164 :
165 1918 : if (!IsUnderPostmaster)
166 : {
167 : dsa_area *dsa;
168 : dshash_table *dsh;
169 1918 : PgStat_ShmemControl *ctl = pgStatLocal.shmem;
170 1918 : char *p = (char *) ctl;
171 :
172 : Assert(!found);
173 :
174 : /* the allocation of pgStatLocal.shmem itself */
175 1918 : p += MAXALIGN(sizeof(PgStat_ShmemControl));
176 :
177 : /*
178 : * Create a small dsa allocation in plain shared memory. This is
179 : * required because postmaster cannot use dsm segments. It also
180 : * provides a small efficiency win.
181 : */
182 1918 : ctl->raw_dsa_area = p;
183 1918 : p += MAXALIGN(pgstat_dsa_init_size());
184 1918 : dsa = dsa_create_in_place(ctl->raw_dsa_area,
185 : pgstat_dsa_init_size(),
186 : LWTRANCHE_PGSTATS_DSA, 0);
187 1918 : dsa_pin(dsa);
188 :
189 : /*
190 : * To ensure dshash is created in "plain" shared memory, temporarily
191 : * limit size of dsa to the initial size of the dsa.
192 : */
193 1918 : dsa_set_size_limit(dsa, pgstat_dsa_init_size());
194 :
195 : /*
196 : * With the limit in place, create the dshash table. XXX: It'd be nice
197 : * if there were dshash_create_in_place().
198 : */
199 1918 : dsh = dshash_create(dsa, &dsh_params, NULL);
200 1918 : ctl->hash_handle = dshash_get_hash_table_handle(dsh);
201 :
202 : /* lift limit set above */
203 1918 : dsa_set_size_limit(dsa, -1);
204 :
205 : /*
206 : * Postmaster will never access these again, thus free the local
207 : * dsa/dshash references.
208 : */
209 1918 : dshash_detach(dsh);
210 1918 : dsa_detach(dsa);
211 :
212 1918 : pg_atomic_init_u64(&ctl->gc_request_count, 1);
213 :
214 : /* initialize fixed-numbered stats */
215 492926 : for (PgStat_Kind kind = PGSTAT_KIND_MIN; kind <= PGSTAT_KIND_MAX; kind++)
216 : {
217 491008 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
218 : char *ptr;
219 :
220 491008 : if (!kind_info || !kind_info->fixed_amount)
221 479492 : continue;
222 :
223 11516 : if (pgstat_is_kind_builtin(kind))
224 11508 : ptr = ((char *) ctl) + kind_info->shared_ctl_off;
225 : else
226 : {
227 8 : int idx = kind - PGSTAT_KIND_CUSTOM_MIN;
228 :
229 : Assert(kind_info->shared_size != 0);
230 8 : ctl->custom_data[idx] = ShmemAlloc(kind_info->shared_size);
231 8 : ptr = ctl->custom_data[idx];
232 : }
233 :
234 11516 : kind_info->init_shmem_cb(ptr);
235 : }
236 : }
237 : else
238 : {
239 : Assert(found);
240 : }
241 1918 : }
242 :
243 : void
244 34724 : pgstat_attach_shmem(void)
245 : {
246 : MemoryContext oldcontext;
247 :
248 : Assert(pgStatLocal.dsa == NULL);
249 :
250 : /* stats shared memory persists for the backend lifetime */
251 34724 : oldcontext = MemoryContextSwitchTo(TopMemoryContext);
252 :
253 34724 : pgStatLocal.dsa = dsa_attach_in_place(pgStatLocal.shmem->raw_dsa_area,
254 : NULL);
255 34724 : dsa_pin_mapping(pgStatLocal.dsa);
256 :
257 69448 : pgStatLocal.shared_hash = dshash_attach(pgStatLocal.dsa, &dsh_params,
258 34724 : pgStatLocal.shmem->hash_handle, 0);
259 :
260 34724 : MemoryContextSwitchTo(oldcontext);
261 34724 : }
262 :
263 : void
264 34724 : pgstat_detach_shmem(void)
265 : {
266 : Assert(pgStatLocal.dsa);
267 :
268 : /* we shouldn't leave references to shared stats */
269 34724 : pgstat_release_all_entry_refs(false);
270 :
271 34724 : dshash_detach(pgStatLocal.shared_hash);
272 34724 : pgStatLocal.shared_hash = NULL;
273 :
274 34724 : dsa_detach(pgStatLocal.dsa);
275 :
276 : /*
277 : * dsa_detach() does not decrement the DSA reference count as no segment
278 : * was provided to dsa_attach_in_place(), causing no cleanup callbacks to
279 : * be registered. Hence, release it manually now.
280 : */
281 34724 : dsa_release_in_place(pgStatLocal.shmem->raw_dsa_area);
282 :
283 34724 : pgStatLocal.dsa = NULL;
284 34724 : }
285 :
286 :
287 : /* ------------------------------------------------------------
288 : * Maintenance of shared memory stats entries
289 : * ------------------------------------------------------------
290 : */
291 :
292 : PgStatShared_Common *
293 534644 : pgstat_init_entry(PgStat_Kind kind,
294 : PgStatShared_HashEntry *shhashent)
295 : {
296 : /* Create new stats entry. */
297 : dsa_pointer chunk;
298 : PgStatShared_Common *shheader;
299 :
300 : /*
301 : * Initialize refcount to 1, marking it as valid / not dropped. The entry
302 : * can't be freed before the initialization because it can't be found as
303 : * long as we hold the dshash partition lock. Caller needs to increase
304 : * further if a longer lived reference is needed.
305 : */
306 534644 : pg_atomic_init_u32(&shhashent->refcount, 1);
307 :
308 : /*
309 : * Initialize "generation" to 0, as freshly created.
310 : */
311 534644 : pg_atomic_init_u32(&shhashent->generation, 0);
312 534644 : shhashent->dropped = false;
313 :
314 534644 : chunk = dsa_allocate0(pgStatLocal.dsa, pgstat_get_kind_info(kind)->shared_size);
315 534644 : shheader = dsa_get_address(pgStatLocal.dsa, chunk);
316 534644 : shheader->magic = 0xdeadbeef;
317 :
318 : /* Link the new entry from the hash entry. */
319 534644 : shhashent->body = chunk;
320 :
321 534644 : LWLockInitialize(&shheader->lock, LWTRANCHE_PGSTATS_DATA);
322 :
323 534644 : return shheader;
324 : }
325 :
326 : static PgStatShared_Common *
327 4 : pgstat_reinit_entry(PgStat_Kind kind, PgStatShared_HashEntry *shhashent)
328 : {
329 : PgStatShared_Common *shheader;
330 :
331 4 : shheader = dsa_get_address(pgStatLocal.dsa, shhashent->body);
332 :
333 : /* mark as not dropped anymore */
334 4 : pg_atomic_fetch_add_u32(&shhashent->refcount, 1);
335 :
336 : /*
337 : * Increment "generation", to let any backend with local references know
338 : * that what they point to is outdated.
339 : */
340 4 : pg_atomic_fetch_add_u32(&shhashent->generation, 1);
341 4 : shhashent->dropped = false;
342 :
343 : /* reinitialize content */
344 : Assert(shheader->magic == 0xdeadbeef);
345 4 : memset(pgstat_get_entry_data(kind, shheader), 0,
346 : pgstat_get_entry_len(kind));
347 :
348 4 : return shheader;
349 : }
350 :
351 : static void
352 105877718 : pgstat_setup_shared_refs(void)
353 : {
354 105877718 : if (likely(pgStatEntryRefHash != NULL))
355 105847510 : return;
356 :
357 30208 : pgStatEntryRefHash =
358 30208 : pgstat_entry_ref_hash_create(pgStatEntryRefHashContext,
359 : PGSTAT_ENTRY_REF_HASH_SIZE, NULL);
360 30208 : pgStatSharedRefAge = pg_atomic_read_u64(&pgStatLocal.shmem->gc_request_count);
361 : Assert(pgStatSharedRefAge != 0);
362 : }
363 :
364 : /*
365 : * Helper function for pgstat_get_entry_ref().
366 : */
367 : static void
368 1610066 : pgstat_acquire_entry_ref(PgStat_EntryRef *entry_ref,
369 : PgStatShared_HashEntry *shhashent,
370 : PgStatShared_Common *shheader)
371 : {
372 : Assert(shheader->magic == 0xdeadbeef);
373 : Assert(pg_atomic_read_u32(&shhashent->refcount) > 0);
374 :
375 1610066 : pg_atomic_fetch_add_u32(&shhashent->refcount, 1);
376 :
377 1610066 : dshash_release_lock(pgStatLocal.shared_hash, shhashent);
378 :
379 1610066 : entry_ref->shared_stats = shheader;
380 1610066 : entry_ref->shared_entry = shhashent;
381 1610066 : entry_ref->generation = pg_atomic_read_u32(&shhashent->generation);
382 1610066 : }
383 :
384 : /*
385 : * Helper function for pgstat_get_entry_ref().
386 : */
387 : static bool
388 105877718 : pgstat_get_entry_ref_cached(PgStat_HashKey key, PgStat_EntryRef **entry_ref_p)
389 : {
390 : bool found;
391 : PgStat_EntryRefHashEntry *cache_entry;
392 :
393 : /*
394 : * We immediately insert a cache entry, because it avoids 1) multiple
395 : * hashtable lookups in case of a cache miss 2) having to deal with
396 : * out-of-memory errors after incrementing PgStatShared_Common->refcount.
397 : */
398 :
399 105877718 : cache_entry = pgstat_entry_ref_hash_insert(pgStatEntryRefHash, key, &found);
400 :
401 105877718 : if (!found || !cache_entry->entry_ref)
402 1760416 : {
403 : PgStat_EntryRef *entry_ref;
404 :
405 1760416 : cache_entry->entry_ref = entry_ref =
406 1760416 : MemoryContextAlloc(pgStatSharedRefContext,
407 : sizeof(PgStat_EntryRef));
408 1760416 : entry_ref->shared_stats = NULL;
409 1760416 : entry_ref->shared_entry = NULL;
410 1760416 : entry_ref->pending = NULL;
411 :
412 1760416 : found = false;
413 : }
414 104117302 : else if (cache_entry->entry_ref->shared_stats == NULL)
415 : {
416 : Assert(cache_entry->entry_ref->pending == NULL);
417 0 : found = false;
418 : }
419 : else
420 : {
421 : PgStat_EntryRef *entry_ref PG_USED_FOR_ASSERTS_ONLY;
422 :
423 104117302 : entry_ref = cache_entry->entry_ref;
424 : Assert(entry_ref->shared_entry != NULL);
425 : Assert(entry_ref->shared_stats != NULL);
426 :
427 : Assert(entry_ref->shared_stats->magic == 0xdeadbeef);
428 : /* should have at least our reference */
429 : Assert(pg_atomic_read_u32(&entry_ref->shared_entry->refcount) > 0);
430 : }
431 :
432 105877718 : *entry_ref_p = cache_entry->entry_ref;
433 105877718 : return found;
434 : }
435 :
436 : /*
437 : * Get a shared stats reference. If create is true, the shared stats object is
438 : * created if it does not exist.
439 : *
440 : * When create is true, and created_entry is non-NULL, it'll be set to true
441 : * if the entry is newly created, false otherwise.
442 : */
443 : PgStat_EntryRef *
444 105877718 : pgstat_get_entry_ref(PgStat_Kind kind, Oid dboid, uint64 objid, bool create,
445 : bool *created_entry)
446 : {
447 : PgStat_HashKey key;
448 : PgStatShared_HashEntry *shhashent;
449 105877718 : PgStatShared_Common *shheader = NULL;
450 : PgStat_EntryRef *entry_ref;
451 :
452 : /* clear padding */
453 105877718 : memset(&key, 0, sizeof(struct PgStat_HashKey));
454 :
455 105877718 : key.kind = kind;
456 105877718 : key.dboid = dboid;
457 105877718 : key.objid = objid;
458 :
459 : /*
460 : * passing in created_entry only makes sense if we possibly could create
461 : * entry.
462 : */
463 : Assert(create || created_entry == NULL);
464 : pgstat_assert_is_up();
465 : Assert(pgStatLocal.shared_hash != NULL);
466 : Assert(!pgStatLocal.shmem->is_shutdown);
467 :
468 105877718 : pgstat_setup_memcxt();
469 105877718 : pgstat_setup_shared_refs();
470 :
471 105877718 : if (created_entry != NULL)
472 214 : *created_entry = false;
473 :
474 : /*
475 : * Check if other backends dropped stats that could not be deleted because
476 : * somebody held references to it. If so, check this backend's references.
477 : * This is not expected to happen often. The location of the check is a
478 : * bit random, but this is a relatively frequently called path, so better
479 : * than most.
480 : */
481 105877718 : if (pgstat_need_entry_refs_gc())
482 11668 : pgstat_gc_entry_refs();
483 :
484 : /*
485 : * First check the lookup cache hashtable in local memory. If we find a
486 : * match here we can avoid taking locks / causing contention.
487 : */
488 105877718 : if (pgstat_get_entry_ref_cached(key, &entry_ref))
489 104117302 : return entry_ref;
490 :
491 : Assert(entry_ref != NULL);
492 :
493 : /*
494 : * Do a lookup in the hash table first - it's quite likely that the entry
495 : * already exists, and that way we only need a shared lock.
496 : */
497 1760416 : shhashent = dshash_find(pgStatLocal.shared_hash, &key, false);
498 :
499 1760416 : if (create && !shhashent)
500 : {
501 : bool shfound;
502 :
503 : /*
504 : * It's possible that somebody created the entry since the above
505 : * lookup. If so, fall through to the same path as if we'd have if it
506 : * already had been created before the dshash_find() calls.
507 : */
508 210830 : shhashent = dshash_find_or_insert(pgStatLocal.shared_hash, &key, &shfound);
509 210830 : if (!shfound)
510 : {
511 210830 : shheader = pgstat_init_entry(kind, shhashent);
512 210830 : pgstat_acquire_entry_ref(entry_ref, shhashent, shheader);
513 :
514 210830 : if (created_entry != NULL)
515 96 : *created_entry = true;
516 :
517 210830 : return entry_ref;
518 : }
519 : }
520 :
521 1549586 : if (!shhashent)
522 : {
523 : /*
524 : * If we're not creating, delete the reference again. In all
525 : * likelihood it's just a stats lookup - no point wasting memory for a
526 : * shared ref to nothing...
527 : */
528 150276 : pgstat_release_entry_ref(key, entry_ref, false);
529 :
530 150276 : return NULL;
531 : }
532 : else
533 : {
534 : /*
535 : * Can get here either because dshash_find() found a match, or if
536 : * dshash_find_or_insert() found a concurrently inserted entry.
537 : */
538 :
539 1399310 : if (shhashent->dropped && create)
540 : {
541 : /*
542 : * There are legitimate cases where the old stats entry might not
543 : * yet have been dropped by the time it's reused. The most obvious
544 : * case are replication slot stats, where a new slot can be
545 : * created with the same index just after dropping. But oid
546 : * wraparound can lead to other cases as well. We just reset the
547 : * stats to their plain state, while incrementing its "generation"
548 : * in the shared entry for any remaining local references.
549 : */
550 4 : shheader = pgstat_reinit_entry(kind, shhashent);
551 4 : pgstat_acquire_entry_ref(entry_ref, shhashent, shheader);
552 :
553 4 : if (created_entry != NULL)
554 0 : *created_entry = true;
555 :
556 4 : return entry_ref;
557 : }
558 1399306 : else if (shhashent->dropped)
559 : {
560 74 : dshash_release_lock(pgStatLocal.shared_hash, shhashent);
561 74 : pgstat_release_entry_ref(key, entry_ref, false);
562 :
563 74 : return NULL;
564 : }
565 : else
566 : {
567 1399232 : shheader = dsa_get_address(pgStatLocal.dsa, shhashent->body);
568 1399232 : pgstat_acquire_entry_ref(entry_ref, shhashent, shheader);
569 :
570 1399232 : return entry_ref;
571 : }
572 : }
573 : }
574 :
575 : static void
576 1760416 : pgstat_release_entry_ref(PgStat_HashKey key, PgStat_EntryRef *entry_ref,
577 : bool discard_pending)
578 : {
579 1760416 : if (entry_ref && entry_ref->pending)
580 : {
581 62240 : if (discard_pending)
582 62240 : pgstat_delete_pending_entry(entry_ref);
583 : else
584 0 : elog(ERROR, "releasing ref with pending data");
585 : }
586 :
587 1760416 : if (entry_ref && entry_ref->shared_stats)
588 : {
589 : Assert(entry_ref->shared_stats->magic == 0xdeadbeef);
590 : Assert(entry_ref->pending == NULL);
591 :
592 : /*
593 : * This can't race with another backend looking up the stats entry and
594 : * increasing the refcount because it is not "legal" to create
595 : * additional references to dropped entries.
596 : */
597 1610066 : if (pg_atomic_fetch_sub_u32(&entry_ref->shared_entry->refcount, 1) == 1)
598 : {
599 : PgStatShared_HashEntry *shent;
600 :
601 : /*
602 : * We're the last referrer to this entry, try to drop the shared
603 : * entry.
604 : */
605 :
606 : /* only dropped entries can reach a 0 refcount */
607 : Assert(entry_ref->shared_entry->dropped);
608 :
609 9352 : shent = dshash_find(pgStatLocal.shared_hash,
610 9352 : &entry_ref->shared_entry->key,
611 : true);
612 9352 : if (!shent)
613 0 : elog(ERROR, "could not find just referenced shared stats entry");
614 :
615 : /*
616 : * This entry may have been reinitialized while trying to release
617 : * it, so double-check that it has not been reused while holding a
618 : * lock on its shared entry.
619 : */
620 9352 : if (pg_atomic_read_u32(&entry_ref->shared_entry->generation) ==
621 9352 : entry_ref->generation)
622 : {
623 : /* Same "generation", so we're OK with the removal */
624 : Assert(pg_atomic_read_u32(&entry_ref->shared_entry->refcount) == 0);
625 : Assert(entry_ref->shared_entry == shent);
626 9352 : pgstat_free_entry(shent, NULL);
627 : }
628 : else
629 : {
630 : /*
631 : * Shared stats entry has been reinitialized, so do not drop
632 : * its shared entry, only release its lock.
633 : */
634 0 : dshash_release_lock(pgStatLocal.shared_hash, shent);
635 : }
636 : }
637 : }
638 :
639 1760416 : if (!pgstat_entry_ref_hash_delete(pgStatEntryRefHash, key))
640 0 : elog(ERROR, "entry ref vanished before deletion");
641 :
642 1760416 : if (entry_ref)
643 1760416 : pfree(entry_ref);
644 1760416 : }
645 :
646 : bool
647 1927698 : pgstat_lock_entry(PgStat_EntryRef *entry_ref, bool nowait)
648 : {
649 1927698 : LWLock *lock = &entry_ref->shared_stats->lock;
650 :
651 1927698 : if (nowait)
652 592050 : return LWLockConditionalAcquire(lock, LW_EXCLUSIVE);
653 :
654 1335648 : LWLockAcquire(lock, LW_EXCLUSIVE);
655 1335648 : return true;
656 : }
657 :
658 : /*
659 : * Separate from pgstat_lock_entry() as most callers will need to lock
660 : * exclusively.
661 : */
662 : bool
663 354756 : pgstat_lock_entry_shared(PgStat_EntryRef *entry_ref, bool nowait)
664 : {
665 354756 : LWLock *lock = &entry_ref->shared_stats->lock;
666 :
667 354756 : if (nowait)
668 0 : return LWLockConditionalAcquire(lock, LW_SHARED);
669 :
670 354756 : LWLockAcquire(lock, LW_SHARED);
671 354756 : return true;
672 : }
673 :
674 : void
675 2282454 : pgstat_unlock_entry(PgStat_EntryRef *entry_ref)
676 : {
677 2282454 : LWLockRelease(&entry_ref->shared_stats->lock);
678 2282454 : }
679 :
680 : /*
681 : * Helper function to fetch and lock shared stats.
682 : */
683 : PgStat_EntryRef *
684 142894 : pgstat_get_entry_ref_locked(PgStat_Kind kind, Oid dboid, uint64 objid,
685 : bool nowait)
686 : {
687 : PgStat_EntryRef *entry_ref;
688 :
689 : /* find shared table stats entry corresponding to the local entry */
690 142894 : entry_ref = pgstat_get_entry_ref(kind, dboid, objid, true, NULL);
691 :
692 : /* lock the shared entry to protect the content, skip if failed */
693 142894 : if (!pgstat_lock_entry(entry_ref, nowait))
694 0 : return NULL;
695 :
696 142894 : return entry_ref;
697 : }
698 :
699 : void
700 3688 : pgstat_request_entry_refs_gc(void)
701 : {
702 3688 : pg_atomic_fetch_add_u64(&pgStatLocal.shmem->gc_request_count, 1);
703 3688 : }
704 :
705 : static bool
706 105877718 : pgstat_need_entry_refs_gc(void)
707 : {
708 : uint64 curage;
709 :
710 105877718 : if (!pgStatEntryRefHash)
711 0 : return false;
712 :
713 : /* should have been initialized when creating pgStatEntryRefHash */
714 : Assert(pgStatSharedRefAge != 0);
715 :
716 105877718 : curage = pg_atomic_read_u64(&pgStatLocal.shmem->gc_request_count);
717 :
718 105877718 : return pgStatSharedRefAge != curage;
719 : }
720 :
721 : static void
722 11668 : pgstat_gc_entry_refs(void)
723 : {
724 : pgstat_entry_ref_hash_iterator i;
725 : PgStat_EntryRefHashEntry *ent;
726 : uint64 curage;
727 :
728 11668 : curage = pg_atomic_read_u64(&pgStatLocal.shmem->gc_request_count);
729 : Assert(curage != 0);
730 :
731 : /*
732 : * Some entries have been dropped or reinitialized. Invalidate cache
733 : * pointer to them.
734 : */
735 11668 : pgstat_entry_ref_hash_start_iterate(pgStatEntryRefHash, &i);
736 870674 : while ((ent = pgstat_entry_ref_hash_iterate(pgStatEntryRefHash, &i)) != NULL)
737 : {
738 859006 : PgStat_EntryRef *entry_ref = ent->entry_ref;
739 :
740 : Assert(!entry_ref->shared_stats ||
741 : entry_ref->shared_stats->magic == 0xdeadbeef);
742 :
743 : /*
744 : * "generation" checks for the case of entries being reinitialized,
745 : * and "dropped" for the case where these are.. dropped.
746 : */
747 859006 : if (!entry_ref->shared_entry->dropped &&
748 621702 : pg_atomic_read_u32(&entry_ref->shared_entry->generation) ==
749 621702 : entry_ref->generation)
750 621698 : continue;
751 :
752 : /* cannot gc shared ref that has pending data */
753 237308 : if (entry_ref->pending != NULL)
754 228226 : continue;
755 :
756 9082 : pgstat_release_entry_ref(ent->key, entry_ref, false);
757 : }
758 :
759 11668 : pgStatSharedRefAge = curage;
760 11668 : }
761 :
762 : static void
763 30254 : pgstat_release_matching_entry_refs(bool discard_pending, ReleaseMatchCB match,
764 : Datum match_data)
765 : {
766 : pgstat_entry_ref_hash_iterator i;
767 : PgStat_EntryRefHashEntry *ent;
768 :
769 30254 : if (pgStatEntryRefHash == NULL)
770 12 : return;
771 :
772 30242 : pgstat_entry_ref_hash_start_iterate(pgStatEntryRefHash, &i);
773 :
774 1539754 : while ((ent = pgstat_entry_ref_hash_iterate(pgStatEntryRefHash, &i))
775 : != NULL)
776 : {
777 : Assert(ent->entry_ref != NULL);
778 :
779 1509512 : if (match && !match(ent, match_data))
780 1106 : continue;
781 :
782 1508406 : pgstat_release_entry_ref(ent->key, ent->entry_ref, discard_pending);
783 : }
784 : }
785 :
786 : /*
787 : * Release all local references to shared stats entries.
788 : *
789 : * When a process exits it cannot do so while still holding references onto
790 : * stats entries, otherwise the shared stats entries could never be freed.
791 : */
792 : static void
793 34724 : pgstat_release_all_entry_refs(bool discard_pending)
794 : {
795 34724 : if (pgStatEntryRefHash == NULL)
796 4516 : return;
797 :
798 30208 : pgstat_release_matching_entry_refs(discard_pending, NULL, 0);
799 : Assert(pgStatEntryRefHash->members == 0);
800 30208 : pgstat_entry_ref_hash_destroy(pgStatEntryRefHash);
801 30208 : pgStatEntryRefHash = NULL;
802 : }
803 :
804 : static bool
805 1106 : match_db(PgStat_EntryRefHashEntry *ent, Datum match_data)
806 : {
807 1106 : Oid dboid = DatumGetObjectId(match_data);
808 :
809 1106 : return ent->key.dboid == dboid;
810 : }
811 :
812 : static void
813 46 : pgstat_release_db_entry_refs(Oid dboid)
814 : {
815 46 : pgstat_release_matching_entry_refs( /* discard pending = */ true,
816 : match_db,
817 : ObjectIdGetDatum(dboid));
818 46 : }
819 :
820 :
821 : /* ------------------------------------------------------------
822 : * Dropping and resetting of stats entries
823 : * ------------------------------------------------------------
824 : */
825 :
826 : static void
827 94756 : pgstat_free_entry(PgStatShared_HashEntry *shent, dshash_seq_status *hstat)
828 : {
829 : dsa_pointer pdsa;
830 :
831 : /*
832 : * Fetch dsa pointer before deleting entry - that way we can free the
833 : * memory after releasing the lock.
834 : */
835 94756 : pdsa = shent->body;
836 :
837 94756 : if (!hstat)
838 92822 : dshash_delete_entry(pgStatLocal.shared_hash, shent);
839 : else
840 1934 : dshash_delete_current(hstat);
841 :
842 94756 : dsa_free(pgStatLocal.dsa, pdsa);
843 94756 : }
844 :
845 : /*
846 : * Helper for both pgstat_drop_database_and_contents() and
847 : * pgstat_drop_entry(). If hstat is non-null delete the shared entry using
848 : * dshash_delete_current(), otherwise use dshash_delete_entry(). In either
849 : * case the entry needs to be already locked.
850 : */
851 : static bool
852 94762 : pgstat_drop_entry_internal(PgStatShared_HashEntry *shent,
853 : dshash_seq_status *hstat)
854 : {
855 : Assert(shent->body != InvalidDsaPointer);
856 :
857 : /* should already have released local reference */
858 94762 : if (pgStatEntryRefHash)
859 : Assert(!pgstat_entry_ref_hash_lookup(pgStatEntryRefHash, shent->key));
860 :
861 : /*
862 : * Signal that the entry is dropped - this will eventually cause other
863 : * backends to release their references.
864 : */
865 94762 : if (shent->dropped)
866 0 : elog(ERROR,
867 : "trying to drop stats entry already dropped: kind=%s dboid=%u objid=%llu refcount=%u",
868 : pgstat_get_kind_info(shent->key.kind)->name,
869 : shent->key.dboid,
870 : (unsigned long long) shent->key.objid,
871 : pg_atomic_read_u32(&shent->refcount));
872 94762 : shent->dropped = true;
873 :
874 : /* release refcount marking entry as not dropped */
875 94762 : if (pg_atomic_sub_fetch_u32(&shent->refcount, 1) == 0)
876 : {
877 85404 : pgstat_free_entry(shent, hstat);
878 85404 : return true;
879 : }
880 : else
881 : {
882 9358 : if (!hstat)
883 9358 : dshash_release_lock(pgStatLocal.shared_hash, shent);
884 9358 : return false;
885 : }
886 : }
887 :
888 : /*
889 : * Drop stats for the database and all the objects inside that database.
890 : */
891 : static void
892 46 : pgstat_drop_database_and_contents(Oid dboid)
893 : {
894 : dshash_seq_status hstat;
895 : PgStatShared_HashEntry *p;
896 46 : uint64 not_freed_count = 0;
897 :
898 : Assert(OidIsValid(dboid));
899 :
900 : Assert(pgStatLocal.shared_hash != NULL);
901 :
902 : /*
903 : * This backend might very well be the only backend holding a reference to
904 : * about-to-be-dropped entries. Ensure that we're not preventing it from
905 : * being cleaned up till later.
906 : *
907 : * Doing this separately from the dshash iteration below avoids having to
908 : * do so while holding a partition lock on the shared hashtable.
909 : */
910 46 : pgstat_release_db_entry_refs(dboid);
911 :
912 : /* some of the dshash entries are to be removed, take exclusive lock. */
913 46 : dshash_seq_init(&hstat, pgStatLocal.shared_hash, true);
914 12180 : while ((p = dshash_seq_next(&hstat)) != NULL)
915 : {
916 12134 : if (p->dropped)
917 2 : continue;
918 :
919 12132 : if (p->key.dboid != dboid)
920 10310 : continue;
921 :
922 1822 : if (!pgstat_drop_entry_internal(p, &hstat))
923 : {
924 : /*
925 : * Even statistics for a dropped database might currently be
926 : * accessed (consider e.g. database stats for pg_stat_database).
927 : */
928 0 : not_freed_count++;
929 : }
930 : }
931 46 : dshash_seq_term(&hstat);
932 :
933 : /*
934 : * If some of the stats data could not be freed, signal the reference
935 : * holders to run garbage collection of their cached pgStatLocal.shmem.
936 : */
937 46 : if (not_freed_count > 0)
938 0 : pgstat_request_entry_refs_gc();
939 46 : }
940 :
941 : /*
942 : * Drop a single stats entry.
943 : *
944 : * This routine returns false if the stats entry of the dropped object could
945 : * not be freed, true otherwise.
946 : *
947 : * The callers of this function should call pgstat_request_entry_refs_gc()
948 : * if the stats entry could not be freed, to ensure that this entry's memory
949 : * can be reclaimed later by a different backend calling
950 : * pgstat_gc_entry_refs().
951 : */
952 : bool
953 132300 : pgstat_drop_entry(PgStat_Kind kind, Oid dboid, uint64 objid)
954 : {
955 : PgStat_HashKey key;
956 : PgStatShared_HashEntry *shent;
957 132300 : bool freed = true;
958 :
959 : /* clear padding */
960 132300 : memset(&key, 0, sizeof(struct PgStat_HashKey));
961 :
962 132300 : key.kind = kind;
963 132300 : key.dboid = dboid;
964 132300 : key.objid = objid;
965 :
966 : /* delete local reference */
967 132300 : if (pgStatEntryRefHash)
968 : {
969 : PgStat_EntryRefHashEntry *lohashent =
970 110158 : pgstat_entry_ref_hash_lookup(pgStatEntryRefHash, key);
971 :
972 110158 : if (lohashent)
973 92578 : pgstat_release_entry_ref(lohashent->key, lohashent->entry_ref,
974 : true);
975 : }
976 :
977 : /* mark entry in shared hashtable as deleted, drop if possible */
978 132300 : shent = dshash_find(pgStatLocal.shared_hash, &key, true);
979 132300 : if (shent)
980 : {
981 92828 : freed = pgstat_drop_entry_internal(shent, NULL);
982 :
983 : /*
984 : * Database stats contain other stats. Drop those as well when
985 : * dropping the database. XXX: Perhaps this should be done in a
986 : * slightly more principled way? But not obvious what that'd look
987 : * like, and so far this is the only case...
988 : */
989 92828 : if (key.kind == PGSTAT_KIND_DATABASE)
990 46 : pgstat_drop_database_and_contents(key.dboid);
991 : }
992 :
993 132300 : return freed;
994 : }
995 :
996 : void
997 458 : pgstat_drop_all_entries(void)
998 : {
999 : dshash_seq_status hstat;
1000 : PgStatShared_HashEntry *ps;
1001 458 : uint64 not_freed_count = 0;
1002 :
1003 458 : dshash_seq_init(&hstat, pgStatLocal.shared_hash, true);
1004 570 : while ((ps = dshash_seq_next(&hstat)) != NULL)
1005 : {
1006 112 : if (ps->dropped)
1007 0 : continue;
1008 :
1009 112 : if (!pgstat_drop_entry_internal(ps, &hstat))
1010 0 : not_freed_count++;
1011 : }
1012 458 : dshash_seq_term(&hstat);
1013 :
1014 458 : if (not_freed_count > 0)
1015 0 : pgstat_request_entry_refs_gc();
1016 458 : }
1017 :
1018 : static void
1019 16984 : shared_stat_reset_contents(PgStat_Kind kind, PgStatShared_Common *header,
1020 : TimestampTz ts)
1021 : {
1022 16984 : const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind);
1023 :
1024 16984 : memset(pgstat_get_entry_data(kind, header), 0,
1025 : pgstat_get_entry_len(kind));
1026 :
1027 16984 : if (kind_info->reset_timestamp_cb)
1028 368 : kind_info->reset_timestamp_cb(header, ts);
1029 16984 : }
1030 :
1031 : /*
1032 : * Reset one variable-numbered stats entry.
1033 : */
1034 : void
1035 340 : pgstat_reset_entry(PgStat_Kind kind, Oid dboid, uint64 objid, TimestampTz ts)
1036 : {
1037 : PgStat_EntryRef *entry_ref;
1038 :
1039 : Assert(!pgstat_get_kind_info(kind)->fixed_amount);
1040 :
1041 340 : entry_ref = pgstat_get_entry_ref(kind, dboid, objid, false, NULL);
1042 340 : if (!entry_ref || entry_ref->shared_entry->dropped)
1043 2 : return;
1044 :
1045 338 : (void) pgstat_lock_entry(entry_ref, false);
1046 338 : shared_stat_reset_contents(kind, entry_ref->shared_stats, ts);
1047 338 : pgstat_unlock_entry(entry_ref);
1048 : }
1049 :
1050 : /*
1051 : * Scan through the shared hashtable of stats, resetting statistics if
1052 : * approved by the provided do_reset() function.
1053 : */
1054 : void
1055 34 : pgstat_reset_matching_entries(bool (*do_reset) (PgStatShared_HashEntry *, Datum),
1056 : Datum match_data, TimestampTz ts)
1057 : {
1058 : dshash_seq_status hstat;
1059 : PgStatShared_HashEntry *p;
1060 :
1061 : /* dshash entry is not modified, take shared lock */
1062 34 : dshash_seq_init(&hstat, pgStatLocal.shared_hash, false);
1063 25046 : while ((p = dshash_seq_next(&hstat)) != NULL)
1064 : {
1065 : PgStatShared_Common *header;
1066 :
1067 25012 : if (p->dropped)
1068 2 : continue;
1069 :
1070 25010 : if (!do_reset(p, match_data))
1071 8364 : continue;
1072 :
1073 16646 : header = dsa_get_address(pgStatLocal.dsa, p->body);
1074 :
1075 16646 : LWLockAcquire(&header->lock, LW_EXCLUSIVE);
1076 :
1077 16646 : shared_stat_reset_contents(p->key.kind, header, ts);
1078 :
1079 16646 : LWLockRelease(&header->lock);
1080 : }
1081 34 : dshash_seq_term(&hstat);
1082 34 : }
1083 :
1084 : static bool
1085 2960 : match_kind(PgStatShared_HashEntry *p, Datum match_data)
1086 : {
1087 2960 : return p->key.kind == DatumGetInt32(match_data);
1088 : }
1089 :
1090 : void
1091 8 : pgstat_reset_entries_of_kind(PgStat_Kind kind, TimestampTz ts)
1092 : {
1093 8 : pgstat_reset_matching_entries(match_kind, Int32GetDatum(kind), ts);
1094 8 : }
1095 :
1096 : static void
1097 105877718 : pgstat_setup_memcxt(void)
1098 : {
1099 105877718 : if (unlikely(!pgStatSharedRefContext))
1100 30208 : pgStatSharedRefContext =
1101 30208 : AllocSetContextCreate(TopMemoryContext,
1102 : "PgStat Shared Ref",
1103 : ALLOCSET_SMALL_SIZES);
1104 105877718 : if (unlikely(!pgStatEntryRefHashContext))
1105 30208 : pgStatEntryRefHashContext =
1106 30208 : AllocSetContextCreate(TopMemoryContext,
1107 : "PgStat Shared Ref Hash",
1108 : ALLOCSET_SMALL_SIZES);
1109 105877718 : }
|
