| 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 |
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 |
sz = 256 * 1024; |
| 119 |
Assert(dsa_minimum_size() <= sz); |
| 120 |
return MAXALIGN(sz); |
| 121 |
} |
| 122 |
|
| 123 |
/* |
| 124 |
* Compute shared memory space needed for cumulative statistics |
| 125 |
*/ |
| 126 |
Size |
| 127 |
StatsShmemSize(void) |
| 128 |
{ |
| 129 |
Size sz; |
| 130 |
|
| 131 |
sz = MAXALIGN(sizeof(PgStat_ShmemControl)); |
| 132 |
sz = add_size(sz, pgstat_dsa_init_size()); |
| 133 |
|
| 134 |
/* Add shared memory for all the custom fixed-numbered statistics */ |
| 135 |
for (PgStat_Kind kind = PGSTAT_KIND_CUSTOM_MIN; kind <= PGSTAT_KIND_CUSTOM_MAX; kind++) |
| 136 |
{ |
| 137 |
const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind); |
| 138 |
|
| 139 |
if (!kind_info) |
| 140 |
continue; |
| 141 |
if (!kind_info->fixed_amount) |
| 142 |
continue; |
| 143 |
|
| 144 |
Assert(kind_info->shared_size != 0); |
| 145 |
sz = add_size(sz, MAXALIGN(kind_info->shared_size)); |
| 146 |
} |
| 147 |
|
| 148 |
return sz; |
| 149 |
} |
| 150 |
|
| 151 |
/* |
| 152 |
* Initialize cumulative statistics system during startup |
| 153 |
*/ |
| 154 |
void |
| 155 |
StatsShmemInit(void) |
| 156 |
{ |
| 157 |
bool found; |
| 158 |
Size sz; |
| 159 |
|
| 160 |
sz = StatsShmemSize(); |
| 161 |
pgStatLocal.shmem = (PgStat_ShmemControl *) |
| 162 |
ShmemInitStruct("Shared Memory Stats", sz, &found); |
| 163 |
|
| 164 |
if (!IsUnderPostmaster) |
| 165 |
{ |
| 166 |
dsa_area *dsa; |
| 167 |
dshash_table *dsh; |
| 168 |
PgStat_ShmemControl *ctl = pgStatLocal.shmem; |
| 169 |
char *p = (char *) ctl; |
| 170 |
|
| 171 |
Assert(!found); |
| 172 |
|
| 173 |
/* the allocation of pgStatLocal.shmem itself */ |
| 174 |
p += MAXALIGN(sizeof(PgStat_ShmemControl)); |
| 175 |
|
| 176 |
/* |
| 177 |
* Create a small dsa allocation in plain shared memory. This is |
| 178 |
* required because postmaster cannot use dsm segments. It also |
| 179 |
* provides a small efficiency win. |
| 180 |
*/ |
| 181 |
ctl->raw_dsa_area = p; |
| 182 |
p += MAXALIGN(pgstat_dsa_init_size()); |
| 183 |
dsa = dsa_create_in_place(ctl->raw_dsa_area, |
| 184 |
pgstat_dsa_init_size(), |
| 185 |
LWTRANCHE_PGSTATS_DSA, NULL); |
| 186 |
dsa_pin(dsa); |
| 187 |
|
| 188 |
/* |
| 189 |
* To ensure dshash is created in "plain" shared memory, temporarily |
| 190 |
* limit size of dsa to the initial size of the dsa. |
| 191 |
*/ |
| 192 |
dsa_set_size_limit(dsa, pgstat_dsa_init_size()); |
| 193 |
|
| 194 |
/* |
| 195 |
* With the limit in place, create the dshash table. XXX: It'd be nice |
| 196 |
* if there were dshash_create_in_place(). |
| 197 |
*/ |
| 198 |
dsh = dshash_create(dsa, &dsh_params, NULL); |
| 199 |
ctl->hash_handle = dshash_get_hash_table_handle(dsh); |
| 200 |
|
| 201 |
/* lift limit set above */ |
| 202 |
dsa_set_size_limit(dsa, -1); |
| 203 |
|
| 204 |
/* |
| 205 |
* Postmaster will never access these again, thus free the local |
| 206 |
* dsa/dshash references. |
| 207 |
*/ |
| 208 |
dshash_detach(dsh); |
| 209 |
dsa_detach(dsa); |
| 210 |
|
| 211 |
pg_atomic_init_u64(&ctl->gc_request_count, 1); |
| 212 |
|
| 213 |
/* initialize fixed-numbered stats */ |
| 214 |
for (PgStat_Kind kind = PGSTAT_KIND_MIN; kind <= PGSTAT_KIND_MAX; kind++) |
| 215 |
{ |
| 216 |
const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind); |
| 217 |
char *ptr; |
| 218 |
|
| 219 |
if (!kind_info || !kind_info->fixed_amount) |
| 220 |
continue; |
| 221 |
|
| 222 |
if (pgstat_is_kind_builtin(kind)) |
| 223 |
ptr = ((char *) ctl) + kind_info->shared_ctl_off; |
| 224 |
else |
| 225 |
{ |
| 226 |
int idx = kind - PGSTAT_KIND_CUSTOM_MIN; |
| 227 |
|
| 228 |
Assert(kind_info->shared_size != 0); |
| 229 |
ctl->custom_data[idx] = p; |
| 230 |
p += MAXALIGN(kind_info->shared_size); |
| 231 |
ptr = ctl->custom_data[idx]; |
| 232 |
} |
| 233 |
|
| 234 |
kind_info->init_shmem_cb(ptr); |
| 235 |
} |
| 236 |
} |
| 237 |
else |
| 238 |
{ |
| 239 |
Assert(found); |
| 240 |
} |
| 241 |
} |
| 242 |
|
| 243 |
void |
| 244 |
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 |
oldcontext = MemoryContextSwitchTo(TopMemoryContext); |
| 252 |
|
| 253 |
pgStatLocal.dsa = dsa_attach_in_place(pgStatLocal.shmem->raw_dsa_area, |
| 254 |
NULL); |
| 255 |
dsa_pin_mapping(pgStatLocal.dsa); |
| 256 |
|
| 257 |
pgStatLocal.shared_hash = dshash_attach(pgStatLocal.dsa, &dsh_params, |
| 258 |
pgStatLocal.shmem->hash_handle, |
| 259 |
NULL); |
| 260 |
|
| 261 |
MemoryContextSwitchTo(oldcontext); |
| 262 |
} |
| 263 |
|
| 264 |
void |
| 265 |
pgstat_detach_shmem(void) |
| 266 |
{ |
| 267 |
Assert(pgStatLocal.dsa); |
| 268 |
|
| 269 |
/* we shouldn't leave references to shared stats */ |
| 270 |
pgstat_release_all_entry_refs(false); |
| 271 |
|
| 272 |
dshash_detach(pgStatLocal.shared_hash); |
| 273 |
pgStatLocal.shared_hash = NULL; |
| 274 |
|
| 275 |
dsa_detach(pgStatLocal.dsa); |
| 276 |
|
| 277 |
/* |
| 278 |
* dsa_detach() does not decrement the DSA reference count as no segment |
| 279 |
* was provided to dsa_attach_in_place(), causing no cleanup callbacks to |
| 280 |
* be registered. Hence, release it manually now. |
| 281 |
*/ |
| 282 |
dsa_release_in_place(pgStatLocal.shmem->raw_dsa_area); |
| 283 |
|
| 284 |
pgStatLocal.dsa = NULL; |
| 285 |
} |
| 286 |
|
| 287 |
|
| 288 |
/* ------------------------------------------------------------ |
| 289 |
* Maintenance of shared memory stats entries |
| 290 |
* ------------------------------------------------------------ |
| 291 |
*/ |
| 292 |
|
| 293 |
/* |
| 294 |
* Initialize entry newly-created. |
| 295 |
* |
| 296 |
* Returns NULL in the event of an allocation failure, so as callers can |
| 297 |
* take cleanup actions as the entry initialized is already inserted in the |
| 298 |
* shared hashtable. |
| 299 |
*/ |
| 300 |
PgStatShared_Common * |
| 301 |
pgstat_init_entry(PgStat_Kind kind, |
| 302 |
PgStatShared_HashEntry *shhashent) |
| 303 |
{ |
| 304 |
/* Create new stats entry. */ |
| 305 |
dsa_pointer chunk; |
| 306 |
PgStatShared_Common *shheader; |
| 307 |
|
| 308 |
/* |
| 309 |
* Initialize refcount to 1, marking it as valid / not dropped. The entry |
| 310 |
* can't be freed before the initialization because it can't be found as |
| 311 |
* long as we hold the dshash partition lock. Caller needs to increase |
| 312 |
* further if a longer lived reference is needed. |
| 313 |
*/ |
| 314 |
pg_atomic_init_u32(&shhashent->refcount, 1); |
| 315 |
|
| 316 |
/* |
| 317 |
* Initialize "generation" to 0, as freshly created. |
| 318 |
*/ |
| 319 |
pg_atomic_init_u32(&shhashent->generation, 0); |
| 320 |
shhashent->dropped = false; |
| 321 |
|
| 322 |
chunk = dsa_allocate_extended(pgStatLocal.dsa, |
| 323 |
pgstat_get_kind_info(kind)->shared_size, |
| 324 |
DSA_ALLOC_ZERO | DSA_ALLOC_NO_OOM); |
| 325 |
if (chunk == InvalidDsaPointer) |
| 326 |
return NULL; |
| 327 |
|
| 328 |
shheader = dsa_get_address(pgStatLocal.dsa, chunk); |
| 329 |
shheader->magic = 0xdeadbeef; |
| 330 |
|
| 331 |
/* Link the new entry from the hash entry. */ |
| 332 |
shhashent->body = chunk; |
| 333 |
|
| 334 |
LWLockInitialize(&shheader->lock, LWTRANCHE_PGSTATS_DATA); |
| 335 |
|
| 336 |
return shheader; |
| 337 |
} |
| 338 |
|
| 339 |
static PgStatShared_Common * |
| 340 |
pgstat_reinit_entry(PgStat_Kind kind, PgStatShared_HashEntry *shhashent) |
| 341 |
{ |
| 342 |
PgStatShared_Common *shheader; |
| 343 |
|
| 344 |
shheader = dsa_get_address(pgStatLocal.dsa, shhashent->body); |
| 345 |
|
| 346 |
/* mark as not dropped anymore */ |
| 347 |
pg_atomic_fetch_add_u32(&shhashent->refcount, 1); |
| 348 |
|
| 349 |
/* |
| 350 |
* Increment "generation", to let any backend with local references know |
| 351 |
* that what they point to is outdated. |
| 352 |
*/ |
| 353 |
pg_atomic_fetch_add_u32(&shhashent->generation, 1); |
| 354 |
shhashent->dropped = false; |
| 355 |
|
| 356 |
/* reinitialize content */ |
| 357 |
Assert(shheader->magic == 0xdeadbeef); |
| 358 |
memset(pgstat_get_entry_data(kind, shheader), 0, |
| 359 |
pgstat_get_entry_len(kind)); |
| 360 |
|
| 361 |
return shheader; |
| 362 |
} |
| 363 |
|
| 364 |
static void |
| 365 |
pgstat_setup_shared_refs(void) |
| 366 |
{ |
| 367 |
if (likely(pgStatEntryRefHash != NULL)) |
| 368 |
return; |
| 369 |
|
| 370 |
pgStatEntryRefHash = |
| 371 |
pgstat_entry_ref_hash_create(pgStatEntryRefHashContext, |
| 372 |
PGSTAT_ENTRY_REF_HASH_SIZE, NULL); |
| 373 |
pgStatSharedRefAge = pg_atomic_read_u64(&pgStatLocal.shmem->gc_request_count); |
| 374 |
Assert(pgStatSharedRefAge != 0); |
| 375 |
} |
| 376 |
|
| 377 |
/* |
| 378 |
* Helper function for pgstat_get_entry_ref(). |
| 379 |
*/ |
| 380 |
static void |
| 381 |
pgstat_acquire_entry_ref(PgStat_EntryRef *entry_ref, |
| 382 |
PgStatShared_HashEntry *shhashent, |
| 383 |
PgStatShared_Common *shheader) |
| 384 |
{ |
| 385 |
Assert(shheader->magic == 0xdeadbeef); |
| 386 |
Assert(pg_atomic_read_u32(&shhashent->refcount) > 0); |
| 387 |
|
| 388 |
pg_atomic_fetch_add_u32(&shhashent->refcount, 1); |
| 389 |
|
| 390 |
dshash_release_lock(pgStatLocal.shared_hash, shhashent); |
| 391 |
|
| 392 |
entry_ref->shared_stats = shheader; |
| 393 |
entry_ref->shared_entry = shhashent; |
| 394 |
entry_ref->generation = pg_atomic_read_u32(&shhashent->generation); |
| 395 |
} |
| 396 |
|
| 397 |
/* |
| 398 |
* Helper function for pgstat_get_entry_ref(). |
| 399 |
*/ |
| 400 |
static bool |
| 401 |
pgstat_get_entry_ref_cached(PgStat_HashKey key, PgStat_EntryRef **entry_ref_p) |
| 402 |
{ |
| 403 |
bool found; |
| 404 |
PgStat_EntryRefHashEntry *cache_entry; |
| 405 |
|
| 406 |
/* |
| 407 |
* We immediately insert a cache entry, because it avoids 1) multiple |
| 408 |
* hashtable lookups in case of a cache miss 2) having to deal with |
| 409 |
* out-of-memory errors after incrementing PgStatShared_Common->refcount. |
| 410 |
*/ |
| 411 |
|
| 412 |
cache_entry = pgstat_entry_ref_hash_insert(pgStatEntryRefHash, key, &found); |
| 413 |
|
| 414 |
if (!found || !cache_entry->entry_ref) |
| 415 |
{ |
| 416 |
PgStat_EntryRef *entry_ref; |
| 417 |
|
| 418 |
cache_entry->entry_ref = entry_ref = |
| 419 |
MemoryContextAlloc(pgStatSharedRefContext, |
| 420 |
sizeof(PgStat_EntryRef)); |
| 421 |
entry_ref->shared_stats = NULL; |
| 422 |
entry_ref->shared_entry = NULL; |
| 423 |
entry_ref->pending = NULL; |
| 424 |
|
| 425 |
found = false; |
| 426 |
} |
| 427 |
else if (cache_entry->entry_ref->shared_stats == NULL) |
| 428 |
{ |
| 429 |
Assert(cache_entry->entry_ref->pending == NULL); |
| 430 |
found = false; |
| 431 |
} |
| 432 |
else |
| 433 |
{ |
| 434 |
PgStat_EntryRef *entry_ref PG_USED_FOR_ASSERTS_ONLY; |
| 435 |
|
| 436 |
entry_ref = cache_entry->entry_ref; |
| 437 |
Assert(entry_ref->shared_entry != NULL); |
| 438 |
Assert(entry_ref->shared_stats != NULL); |
| 439 |
|
| 440 |
Assert(entry_ref->shared_stats->magic == 0xdeadbeef); |
| 441 |
/* should have at least our reference */ |
| 442 |
Assert(pg_atomic_read_u32(&entry_ref->shared_entry->refcount) > 0); |
| 443 |
} |
| 444 |
|
| 445 |
*entry_ref_p = cache_entry->entry_ref; |
| 446 |
return found; |
| 447 |
} |
| 448 |
|
| 449 |
/* |
| 450 |
* Get a shared stats reference. If create is true, the shared stats object is |
| 451 |
* created if it does not exist. |
| 452 |
* |
| 453 |
* When create is true, and created_entry is non-NULL, it'll be set to true |
| 454 |
* if the entry is newly created, false otherwise. |
| 455 |
*/ |
| 456 |
PgStat_EntryRef * |
| 457 |
pgstat_get_entry_ref(PgStat_Kind kind, Oid dboid, uint64 objid, bool create, |
| 458 |
bool *created_entry) |
| 459 |
{ |
| 460 |
PgStat_HashKey key; |
| 461 |
PgStatShared_HashEntry *shhashent; |
| 462 |
PgStatShared_Common *shheader = NULL; |
| 463 |
PgStat_EntryRef *entry_ref; |
| 464 |
|
| 465 |
/* clear padding */ |
| 466 |
memset(&key, 0, sizeof(struct PgStat_HashKey)); |
| 467 |
|
| 468 |
key.kind = kind; |
| 469 |
key.dboid = dboid; |
| 470 |
key.objid = objid; |
| 471 |
|
| 472 |
/* |
| 473 |
* passing in created_entry only makes sense if we possibly could create |
| 474 |
* entry. |
| 475 |
*/ |
| 476 |
Assert(create || created_entry == NULL); |
| 477 |
pgstat_assert_is_up(); |
| 478 |
Assert(pgStatLocal.shared_hash != NULL); |
| 479 |
Assert(!pgStatLocal.shmem->is_shutdown); |
| 480 |
|
| 481 |
pgstat_setup_memcxt(); |
| 482 |
pgstat_setup_shared_refs(); |
| 483 |
|
| 484 |
if (created_entry != NULL) |
| 485 |
*created_entry = false; |
| 486 |
|
| 487 |
/* |
| 488 |
* Check if other backends dropped stats that could not be deleted because |
| 489 |
* somebody held references to it. If so, check this backend's references. |
| 490 |
* This is not expected to happen often. The location of the check is a |
| 491 |
* bit random, but this is a relatively frequently called path, so better |
| 492 |
* than most. |
| 493 |
*/ |
| 494 |
if (pgstat_need_entry_refs_gc()) |
| 495 |
pgstat_gc_entry_refs(); |
| 496 |
|
| 497 |
/* |
| 498 |
* First check the lookup cache hashtable in local memory. If we find a |
| 499 |
* match here we can avoid taking locks / causing contention. |
| 500 |
*/ |
| 501 |
if (pgstat_get_entry_ref_cached(key, &entry_ref)) |
| 502 |
return entry_ref; |
| 503 |
|
| 504 |
Assert(entry_ref != NULL); |
| 505 |
|
| 506 |
/* |
| 507 |
* Do a lookup in the hash table first - it's quite likely that the entry |
| 508 |
* already exists, and that way we only need a shared lock. |
| 509 |
*/ |
| 510 |
shhashent = dshash_find(pgStatLocal.shared_hash, &key, false); |
| 511 |
|
| 512 |
if (create && !shhashent) |
| 513 |
{ |
| 514 |
bool shfound; |
| 515 |
|
| 516 |
/* |
| 517 |
* It's possible that somebody created the entry since the above |
| 518 |
* lookup. If so, fall through to the same path as if we'd have if it |
| 519 |
* already had been created before the dshash_find() calls. |
| 520 |
*/ |
| 521 |
shhashent = dshash_find_or_insert(pgStatLocal.shared_hash, &key, &shfound); |
| 522 |
if (!shfound) |
| 523 |
{ |
| 524 |
shheader = pgstat_init_entry(kind, shhashent); |
| 525 |
if (shheader == NULL) |
| 526 |
{ |
| 527 |
/* |
| 528 |
* Failed the allocation of a new entry, so clean up both the |
| 529 |
* local reference and the shared hashtable before giving up. |
| 530 |
* Clean the local state first, since releasing the dshash |
| 531 |
* lock can process a pending interrupt. |
| 532 |
*/ |
| 533 |
pgstat_release_entry_ref(key, entry_ref, false); |
| 534 |
dshash_delete_entry(pgStatLocal.shared_hash, shhashent); |
| 535 |
|
| 536 |
ereport(ERROR, |
| 537 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 538 |
errmsg("out of memory"), |
| 539 |
errdetail("Failed while allocating entry %u/%u/%" PRIu64 ".", |
| 540 |
key.kind, key.dboid, key.objid))); |
| 541 |
} |
| 542 |
pgstat_acquire_entry_ref(entry_ref, shhashent, shheader); |
| 543 |
|
| 544 |
if (created_entry != NULL) |
| 545 |
*created_entry = true; |
| 546 |
|
| 547 |
return entry_ref; |
| 548 |
} |
| 549 |
} |
| 550 |
|
| 551 |
if (!shhashent) |
| 552 |
{ |
| 553 |
/* |
| 554 |
* If we're not creating, delete the reference again. In all |
| 555 |
* likelihood it's just a stats lookup - no point wasting memory for a |
| 556 |
* shared ref to nothing... |
| 557 |
*/ |
| 558 |
pgstat_release_entry_ref(key, entry_ref, false); |
| 559 |
|
| 560 |
return NULL; |
| 561 |
} |
| 562 |
else |
| 563 |
{ |
| 564 |
/* |
| 565 |
* Can get here either because dshash_find() found a match, or if |
| 566 |
* dshash_find_or_insert() found a concurrently inserted entry. |
| 567 |
*/ |
| 568 |
|
| 569 |
if (shhashent->dropped && create) |
| 570 |
{ |
| 571 |
/* |
| 572 |
* There are legitimate cases where the old stats entry might not |
| 573 |
* yet have been dropped by the time it's reused. The most obvious |
| 574 |
* case are replication slot stats, where a new slot can be |
| 575 |
* created with the same index just after dropping. But oid |
| 576 |
* wraparound can lead to other cases as well. We just reset the |
| 577 |
* stats to their plain state, while incrementing its "generation" |
| 578 |
* in the shared entry for any remaining local references. |
| 579 |
*/ |
| 580 |
shheader = pgstat_reinit_entry(kind, shhashent); |
| 581 |
pgstat_acquire_entry_ref(entry_ref, shhashent, shheader); |
| 582 |
|
| 583 |
if (created_entry != NULL) |
| 584 |
*created_entry = true; |
| 585 |
|
| 586 |
return entry_ref; |
| 587 |
} |
| 588 |
else if (shhashent->dropped) |
| 589 |
{ |
| 590 |
dshash_release_lock(pgStatLocal.shared_hash, shhashent); |
| 591 |
pgstat_release_entry_ref(key, entry_ref, false); |
| 592 |
|
| 593 |
return NULL; |
| 594 |
} |
| 595 |
else |
| 596 |
{ |
| 597 |
shheader = dsa_get_address(pgStatLocal.dsa, shhashent->body); |
| 598 |
pgstat_acquire_entry_ref(entry_ref, shhashent, shheader); |
| 599 |
|
| 600 |
return entry_ref; |
| 601 |
} |
| 602 |
} |
| 603 |
} |
| 604 |
|
| 605 |
static void |
| 606 |
pgstat_release_entry_ref(PgStat_HashKey key, PgStat_EntryRef *entry_ref, |
| 607 |
bool discard_pending) |
| 608 |
{ |
| 609 |
if (entry_ref && entry_ref->pending) |
| 610 |
{ |
| 611 |
if (discard_pending) |
| 612 |
pgstat_delete_pending_entry(entry_ref); |
| 613 |
else |
| 614 |
elog(ERROR, "releasing ref with pending data"); |
| 615 |
} |
| 616 |
|
| 617 |
if (entry_ref && entry_ref->shared_stats) |
| 618 |
{ |
| 619 |
Assert(entry_ref->shared_stats->magic == 0xdeadbeef); |
| 620 |
Assert(entry_ref->pending == NULL); |
| 621 |
|
| 622 |
/* |
| 623 |
* This can't race with another backend looking up the stats entry and |
| 624 |
* increasing the refcount because it is not "legal" to create |
| 625 |
* additional references to dropped entries. |
| 626 |
*/ |
| 627 |
if (pg_atomic_fetch_sub_u32(&entry_ref->shared_entry->refcount, 1) == 1) |
| 628 |
{ |
| 629 |
PgStatShared_HashEntry *shent; |
| 630 |
|
| 631 |
/* |
| 632 |
* We're the last referrer to this entry, try to drop the shared |
| 633 |
* entry. |
| 634 |
*/ |
| 635 |
|
| 636 |
/* only dropped entries can reach a 0 refcount */ |
| 637 |
Assert(entry_ref->shared_entry->dropped); |
| 638 |
|
| 639 |
shent = dshash_find(pgStatLocal.shared_hash, |
| 640 |
&entry_ref->shared_entry->key, |
| 641 |
true); |
| 642 |
if (!shent) |
| 643 |
elog(ERROR, "could not find just referenced shared stats entry"); |
| 644 |
|
| 645 |
/* |
| 646 |
* This entry may have been reinitialized while trying to release |
| 647 |
* it, so double-check that it has not been reused while holding a |
| 648 |
* lock on its shared entry. |
| 649 |
*/ |
| 650 |
if (pg_atomic_read_u32(&entry_ref->shared_entry->generation) == |
| 651 |
entry_ref->generation) |
| 652 |
{ |
| 653 |
/* Same "generation", so we're OK with the removal */ |
| 654 |
Assert(pg_atomic_read_u32(&entry_ref->shared_entry->refcount) == 0); |
| 655 |
Assert(entry_ref->shared_entry == shent); |
| 656 |
pgstat_free_entry(shent, NULL); |
| 657 |
} |
| 658 |
else |
| 659 |
{ |
| 660 |
/* |
| 661 |
* Shared stats entry has been reinitialized, so do not drop |
| 662 |
* its shared entry, only release its lock. |
| 663 |
*/ |
| 664 |
dshash_release_lock(pgStatLocal.shared_hash, shent); |
| 665 |
} |
| 666 |
} |
| 667 |
} |
| 668 |
|
| 669 |
if (!pgstat_entry_ref_hash_delete(pgStatEntryRefHash, key)) |
| 670 |
elog(ERROR, "entry ref vanished before deletion"); |
| 671 |
|
| 672 |
if (entry_ref) |
| 673 |
pfree(entry_ref); |
| 674 |
} |
| 675 |
|
| 676 |
/* |
| 677 |
* Acquire exclusive lock on the entry. |
| 678 |
* |
| 679 |
* If nowait is true, it's just a conditional acquire, and the result |
| 680 |
* *must* be checked to verify success. |
| 681 |
* If nowait is false, waits as necessary, always returning true. |
| 682 |
*/ |
| 683 |
bool |
| 684 |
pgstat_lock_entry(PgStat_EntryRef *entry_ref, bool nowait) |
| 685 |
{ |
| 686 |
LWLock *lock = &entry_ref->shared_stats->lock; |
| 687 |
|
| 688 |
if (nowait) |
| 689 |
return LWLockConditionalAcquire(lock, LW_EXCLUSIVE); |
| 690 |
|
| 691 |
LWLockAcquire(lock, LW_EXCLUSIVE); |
| 692 |
return true; |
| 693 |
} |
| 694 |
|
| 695 |
/* |
| 696 |
* Acquire shared lock on the entry. |
| 697 |
* |
| 698 |
* Separate from pgstat_lock_entry() as most callers will need to lock |
| 699 |
* exclusively. The wait semantics are identical. |
| 700 |
*/ |
| 701 |
bool |
| 702 |
pgstat_lock_entry_shared(PgStat_EntryRef *entry_ref, bool nowait) |
| 703 |
{ |
| 704 |
LWLock *lock = &entry_ref->shared_stats->lock; |
| 705 |
|
| 706 |
if (nowait) |
| 707 |
return LWLockConditionalAcquire(lock, LW_SHARED); |
| 708 |
|
| 709 |
LWLockAcquire(lock, LW_SHARED); |
| 710 |
return true; |
| 711 |
} |
| 712 |
|
| 713 |
void |
| 714 |
pgstat_unlock_entry(PgStat_EntryRef *entry_ref) |
| 715 |
{ |
| 716 |
LWLockRelease(&entry_ref->shared_stats->lock); |
| 717 |
} |
| 718 |
|
| 719 |
/* |
| 720 |
* Helper function to fetch and lock shared stats. |
| 721 |
*/ |
| 722 |
PgStat_EntryRef * |
| 723 |
pgstat_get_entry_ref_locked(PgStat_Kind kind, Oid dboid, uint64 objid, |
| 724 |
bool nowait) |
| 725 |
{ |
| 726 |
PgStat_EntryRef *entry_ref; |
| 727 |
|
| 728 |
/* find shared table stats entry corresponding to the local entry */ |
| 729 |
entry_ref = pgstat_get_entry_ref(kind, dboid, objid, true, NULL); |
| 730 |
|
| 731 |
/* lock the shared entry to protect the content, skip if failed */ |
| 732 |
if (!pgstat_lock_entry(entry_ref, nowait)) |
| 733 |
return NULL; |
| 734 |
|
| 735 |
return entry_ref; |
| 736 |
} |
| 737 |
|
| 738 |
void |
| 739 |
pgstat_request_entry_refs_gc(void) |
| 740 |
{ |
| 741 |
pg_atomic_fetch_add_u64(&pgStatLocal.shmem->gc_request_count, 1); |
| 742 |
} |
| 743 |
|
| 744 |
static bool |
| 745 |
pgstat_need_entry_refs_gc(void) |
| 746 |
{ |
| 747 |
uint64 curage; |
| 748 |
|
| 749 |
if (!pgStatEntryRefHash) |
| 750 |
return false; |
| 751 |
|
| 752 |
/* should have been initialized when creating pgStatEntryRefHash */ |
| 753 |
Assert(pgStatSharedRefAge != 0); |
| 754 |
|
| 755 |
curage = pg_atomic_read_u64(&pgStatLocal.shmem->gc_request_count); |
| 756 |
|
| 757 |
return pgStatSharedRefAge != curage; |
| 758 |
} |
| 759 |
|
| 760 |
static void |
| 761 |
pgstat_gc_entry_refs(void) |
| 762 |
{ |
| 763 |
pgstat_entry_ref_hash_iterator i; |
| 764 |
PgStat_EntryRefHashEntry *ent; |
| 765 |
uint64 curage; |
| 766 |
|
| 767 |
curage = pg_atomic_read_u64(&pgStatLocal.shmem->gc_request_count); |
| 768 |
Assert(curage != 0); |
| 769 |
|
| 770 |
/* |
| 771 |
* Some entries have been dropped or reinitialized. Invalidate cache |
| 772 |
* pointer to them. |
| 773 |
*/ |
| 774 |
pgstat_entry_ref_hash_start_iterate(pgStatEntryRefHash, &i); |
| 775 |
while ((ent = pgstat_entry_ref_hash_iterate(pgStatEntryRefHash, &i)) != NULL) |
| 776 |
{ |
| 777 |
PgStat_EntryRef *entry_ref = ent->entry_ref; |
| 778 |
|
| 779 |
Assert(!entry_ref->shared_stats || |
| 780 |
entry_ref->shared_stats->magic == 0xdeadbeef); |
| 781 |
|
| 782 |
/* |
| 783 |
* "generation" checks for the case of entries being reinitialized, |
| 784 |
* and "dropped" for the case where these are.. dropped. |
| 785 |
*/ |
| 786 |
if (!entry_ref->shared_entry->dropped && |
| 787 |
pg_atomic_read_u32(&entry_ref->shared_entry->generation) == |
| 788 |
entry_ref->generation) |
| 789 |
continue; |
| 790 |
|
| 791 |
/* cannot gc shared ref that has pending data */ |
| 792 |
if (entry_ref->pending != NULL) |
| 793 |
continue; |
| 794 |
|
| 795 |
pgstat_release_entry_ref(ent->key, entry_ref, false); |
| 796 |
} |
| 797 |
|
| 798 |
pgStatSharedRefAge = curage; |
| 799 |
} |
| 800 |
|
| 801 |
static void |
| 802 |
pgstat_release_matching_entry_refs(bool discard_pending, ReleaseMatchCB match, |
| 803 |
Datum match_data) |
| 804 |
{ |
| 805 |
pgstat_entry_ref_hash_iterator i; |
| 806 |
PgStat_EntryRefHashEntry *ent; |
| 807 |
|
| 808 |
if (pgStatEntryRefHash == NULL) |
| 809 |
return; |
| 810 |
|
| 811 |
pgstat_entry_ref_hash_start_iterate(pgStatEntryRefHash, &i); |
| 812 |
|
| 813 |
while ((ent = pgstat_entry_ref_hash_iterate(pgStatEntryRefHash, &i)) |
| 814 |
!= NULL) |
| 815 |
{ |
| 816 |
Assert(ent->entry_ref != NULL); |
| 817 |
|
| 818 |
if (match && !match(ent, match_data)) |
| 819 |
continue; |
| 820 |
|
| 821 |
pgstat_release_entry_ref(ent->key, ent->entry_ref, discard_pending); |
| 822 |
} |
| 823 |
} |
| 824 |
|
| 825 |
/* |
| 826 |
* Release all local references to shared stats entries. |
| 827 |
* |
| 828 |
* When a process exits it cannot do so while still holding references onto |
| 829 |
* stats entries, otherwise the shared stats entries could never be freed. |
| 830 |
*/ |
| 831 |
static void |
| 832 |
pgstat_release_all_entry_refs(bool discard_pending) |
| 833 |
{ |
| 834 |
if (pgStatEntryRefHash == NULL) |
| 835 |
return; |
| 836 |
|
| 837 |
pgstat_release_matching_entry_refs(discard_pending, NULL, 0); |
| 838 |
Assert(pgStatEntryRefHash->members == 0); |
| 839 |
pgstat_entry_ref_hash_destroy(pgStatEntryRefHash); |
| 840 |
pgStatEntryRefHash = NULL; |
| 841 |
} |
| 842 |
|
| 843 |
static bool |
| 844 |
match_db(PgStat_EntryRefHashEntry *ent, Datum match_data) |
| 845 |
{ |
| 846 |
Oid dboid = DatumGetObjectId(match_data); |
| 847 |
|
| 848 |
return ent->key.dboid == dboid; |
| 849 |
} |
| 850 |
|
| 851 |
static void |
| 852 |
pgstat_release_db_entry_refs(Oid dboid) |
| 853 |
{ |
| 854 |
pgstat_release_matching_entry_refs( /* discard pending = */ true, |
| 855 |
match_db, |
| 856 |
ObjectIdGetDatum(dboid)); |
| 857 |
} |
| 858 |
|
| 859 |
|
| 860 |
/* ------------------------------------------------------------ |
| 861 |
* Dropping and resetting of stats entries |
| 862 |
* ------------------------------------------------------------ |
| 863 |
*/ |
| 864 |
|
| 865 |
static void |
| 866 |
pgstat_free_entry(PgStatShared_HashEntry *shent, dshash_seq_status *hstat) |
| 867 |
{ |
| 868 |
dsa_pointer pdsa; |
| 869 |
|
| 870 |
/* |
| 871 |
* Fetch dsa pointer before deleting entry - that way we can free the |
| 872 |
* memory after releasing the lock. |
| 873 |
*/ |
| 874 |
pdsa = shent->body; |
| 875 |
|
| 876 |
if (!hstat) |
| 877 |
dshash_delete_entry(pgStatLocal.shared_hash, shent); |
| 878 |
else |
| 879 |
dshash_delete_current(hstat); |
| 880 |
|
| 881 |
dsa_free(pgStatLocal.dsa, pdsa); |
| 882 |
} |
| 883 |
|
| 884 |
/* |
| 885 |
* Helper for both pgstat_drop_database_and_contents() and |
| 886 |
* pgstat_drop_entry_ext(). If hstat is non-null delete the shared entry using |
| 887 |
* dshash_delete_current(), otherwise use dshash_delete_entry(). In either |
| 888 |
* case the entry needs to be already locked. |
| 889 |
*/ |
| 890 |
static bool |
| 891 |
pgstat_drop_entry_internal(PgStatShared_HashEntry *shent, |
| 892 |
dshash_seq_status *hstat) |
| 893 |
{ |
| 894 |
Assert(shent->body != InvalidDsaPointer); |
| 895 |
|
| 896 |
/* should already have released local reference */ |
| 897 |
if (pgStatEntryRefHash) |
| 898 |
Assert(!pgstat_entry_ref_hash_lookup(pgStatEntryRefHash, shent->key)); |
| 899 |
|
| 900 |
/* |
| 901 |
* Signal that the entry is dropped - this will eventually cause other |
| 902 |
* backends to release their references. |
| 903 |
*/ |
| 904 |
Assert(!shent->dropped); |
| 905 |
shent->dropped = true; |
| 906 |
|
| 907 |
/* release refcount marking entry as not dropped */ |
| 908 |
if (pg_atomic_sub_fetch_u32(&shent->refcount, 1) == 0) |
| 909 |
{ |
| 910 |
pgstat_free_entry(shent, hstat); |
| 911 |
return true; |
| 912 |
} |
| 913 |
else |
| 914 |
{ |
| 915 |
if (!hstat) |
| 916 |
dshash_release_lock(pgStatLocal.shared_hash, shent); |
| 917 |
return false; |
| 918 |
} |
| 919 |
} |
| 920 |
|
| 921 |
/* |
| 922 |
* Drop stats for the database and all the objects inside that database. |
| 923 |
*/ |
| 924 |
static void |
| 925 |
pgstat_drop_database_and_contents(Oid dboid) |
| 926 |
{ |
| 927 |
dshash_seq_status hstat; |
| 928 |
PgStatShared_HashEntry *p; |
| 929 |
uint64 not_freed_count = 0; |
| 930 |
|
| 931 |
Assert(OidIsValid(dboid)); |
| 932 |
|
| 933 |
Assert(pgStatLocal.shared_hash != NULL); |
| 934 |
|
| 935 |
/* |
| 936 |
* This backend might very well be the only backend holding a reference to |
| 937 |
* about-to-be-dropped entries. Ensure that we're not preventing it from |
| 938 |
* being cleaned up till later. |
| 939 |
* |
| 940 |
* Doing this separately from the dshash iteration below avoids having to |
| 941 |
* do so while holding a partition lock on the shared hashtable. |
| 942 |
*/ |
| 943 |
pgstat_release_db_entry_refs(dboid); |
| 944 |
|
| 945 |
/* some of the dshash entries are to be removed, take exclusive lock. */ |
| 946 |
dshash_seq_init(&hstat, pgStatLocal.shared_hash, true); |
| 947 |
while ((p = dshash_seq_next(&hstat)) != NULL) |
| 948 |
{ |
| 949 |
if (p->dropped) |
| 950 |
continue; |
| 951 |
|
| 952 |
if (p->key.dboid != dboid) |
| 953 |
continue; |
| 954 |
|
| 955 |
if (!pgstat_drop_entry_internal(p, &hstat)) |
| 956 |
{ |
| 957 |
/* |
| 958 |
* Even statistics for a dropped database might currently be |
| 959 |
* accessed (consider e.g. database stats for pg_stat_database). |
| 960 |
*/ |
| 961 |
not_freed_count++; |
| 962 |
} |
| 963 |
} |
| 964 |
dshash_seq_term(&hstat); |
| 965 |
|
| 966 |
/* |
| 967 |
* If some of the stats data could not be freed, signal the reference |
| 968 |
* holders to run garbage collection of their cached pgStatLocal.shmem. |
| 969 |
*/ |
| 970 |
if (not_freed_count > 0) |
| 971 |
pgstat_request_entry_refs_gc(); |
| 972 |
} |
| 973 |
|
| 974 |
/* |
| 975 |
* ABI-preserving wrapper around pgstat_drop_entry_ext(). |
| 976 |
* |
| 977 |
* The original routine introduced in v15 did not include "missing_ok". |
| 978 |
*/ |
| 979 |
bool |
| 980 |
pgstat_drop_entry(PgStat_Kind kind, Oid dboid, uint64 objid) |
| 981 |
{ |
| 982 |
return pgstat_drop_entry_ext(kind, dboid, objid, false); |
| 983 |
} |
| 984 |
|
| 985 |
/* |
| 986 |
* Drop a single stats entry. |
| 987 |
* |
| 988 |
* This routine returns false if the stats entry of the dropped object could |
| 989 |
* not be freed, true otherwise. |
| 990 |
* |
| 991 |
* If missing_ok is true, skip entries that have been concurrently dropped. |
| 992 |
* |
| 993 |
* The callers of this function should call pgstat_request_entry_refs_gc() |
| 994 |
* if the stats entry could not be freed, to ensure that this entry's memory |
| 995 |
* can be reclaimed later by a different backend calling |
| 996 |
* pgstat_gc_entry_refs(). |
| 997 |
*/ |
| 998 |
bool |
| 999 |
pgstat_drop_entry_ext(PgStat_Kind kind, Oid dboid, uint64 objid, |
| 1000 |
bool missing_ok) |
| 1001 |
{ |
| 1002 |
PgStat_HashKey key; |
| 1003 |
PgStatShared_HashEntry *shent; |
| 1004 |
bool freed = true; |
| 1005 |
|
| 1006 |
/* clear padding */ |
| 1007 |
memset(&key, 0, sizeof(struct PgStat_HashKey)); |
| 1008 |
|
| 1009 |
key.kind = kind; |
| 1010 |
key.dboid = dboid; |
| 1011 |
key.objid = objid; |
| 1012 |
|
| 1013 |
/* delete local reference */ |
| 1014 |
if (pgStatEntryRefHash) |
| 1015 |
{ |
| 1016 |
PgStat_EntryRefHashEntry *lohashent = |
| 1017 |
pgstat_entry_ref_hash_lookup(pgStatEntryRefHash, key); |
| 1018 |
|
| 1019 |
if (lohashent) |
| 1020 |
pgstat_release_entry_ref(lohashent->key, lohashent->entry_ref, |
| 1021 |
true); |
| 1022 |
} |
| 1023 |
|
| 1024 |
/* mark entry in shared hashtable as deleted, drop if possible */ |
| 1025 |
shent = dshash_find(pgStatLocal.shared_hash, &key, true); |
| 1026 |
if (shent) |
| 1027 |
{ |
| 1028 |
if (shent->dropped) |
| 1029 |
{ |
| 1030 |
if (!missing_ok) |
| 1031 |
elog(ERROR, |
| 1032 |
"trying to drop stats entry already dropped: kind=%s dboid=%u objid=%" PRIu64 " refcount=%u generation=%u", |
| 1033 |
pgstat_get_kind_info(shent->key.kind)->name, |
| 1034 |
shent->key.dboid, |
| 1035 |
shent->key.objid, |
| 1036 |
pg_atomic_read_u32(&shent->refcount), |
| 1037 |
pg_atomic_read_u32(&shent->generation)); |
| 1038 |
dshash_release_lock(pgStatLocal.shared_hash, shent); |
| 1039 |
return true; |
| 1040 |
} |
| 1041 |
|
| 1042 |
freed = pgstat_drop_entry_internal(shent, NULL); |
| 1043 |
|
| 1044 |
/* |
| 1045 |
* Database stats contain other stats. Drop those as well when |
| 1046 |
* dropping the database. XXX: Perhaps this should be done in a |
| 1047 |
* slightly more principled way? But not obvious what that'd look |
| 1048 |
* like, and so far this is the only case... |
| 1049 |
*/ |
| 1050 |
if (key.kind == PGSTAT_KIND_DATABASE) |
| 1051 |
pgstat_drop_database_and_contents(key.dboid); |
| 1052 |
} |
| 1053 |
|
| 1054 |
return freed; |
| 1055 |
} |
| 1056 |
|
| 1057 |
/* |
| 1058 |
* Scan through the shared hashtable of stats, dropping statistics if |
| 1059 |
* approved by the optional do_drop() function. |
| 1060 |
*/ |
| 1061 |
void |
| 1062 |
pgstat_drop_matching_entries(bool (*do_drop) (PgStatShared_HashEntry *, Datum), |
| 1063 |
Datum match_data) |
| 1064 |
{ |
| 1065 |
dshash_seq_status hstat; |
| 1066 |
PgStatShared_HashEntry *ps; |
| 1067 |
uint64 not_freed_count = 0; |
| 1068 |
|
| 1069 |
/* entries are removed, take an exclusive lock */ |
| 1070 |
dshash_seq_init(&hstat, pgStatLocal.shared_hash, true); |
| 1071 |
while ((ps = dshash_seq_next(&hstat)) != NULL) |
| 1072 |
{ |
| 1073 |
if (ps->dropped) |
| 1074 |
continue; |
| 1075 |
|
| 1076 |
if (do_drop != NULL && !do_drop(ps, match_data)) |
| 1077 |
continue; |
| 1078 |
|
| 1079 |
/* delete local reference */ |
| 1080 |
if (pgStatEntryRefHash) |
| 1081 |
{ |
| 1082 |
PgStat_EntryRefHashEntry *lohashent = |
| 1083 |
pgstat_entry_ref_hash_lookup(pgStatEntryRefHash, ps->key); |
| 1084 |
|
| 1085 |
if (lohashent) |
| 1086 |
pgstat_release_entry_ref(lohashent->key, lohashent->entry_ref, |
| 1087 |
true); |
| 1088 |
} |
| 1089 |
|
| 1090 |
if (!pgstat_drop_entry_internal(ps, &hstat)) |
| 1091 |
not_freed_count++; |
| 1092 |
} |
| 1093 |
dshash_seq_term(&hstat); |
| 1094 |
|
| 1095 |
if (not_freed_count > 0) |
| 1096 |
pgstat_request_entry_refs_gc(); |
| 1097 |
} |
| 1098 |
|
| 1099 |
/* |
| 1100 |
* Scan through the shared hashtable of stats and drop all entries. |
| 1101 |
*/ |
| 1102 |
void |
| 1103 |
pgstat_drop_all_entries(void) |
| 1104 |
{ |
| 1105 |
pgstat_drop_matching_entries(NULL, 0); |
| 1106 |
} |
| 1107 |
|
| 1108 |
static void |
| 1109 |
shared_stat_reset_contents(PgStat_Kind kind, PgStatShared_Common *header, |
| 1110 |
TimestampTz ts) |
| 1111 |
{ |
| 1112 |
const PgStat_KindInfo *kind_info = pgstat_get_kind_info(kind); |
| 1113 |
|
| 1114 |
memset(pgstat_get_entry_data(kind, header), 0, |
| 1115 |
pgstat_get_entry_len(kind)); |
| 1116 |
|
| 1117 |
if (kind_info->reset_timestamp_cb) |
| 1118 |
kind_info->reset_timestamp_cb(header, ts); |
| 1119 |
} |
| 1120 |
|
| 1121 |
/* |
| 1122 |
* Reset one variable-numbered stats entry. |
| 1123 |
*/ |
| 1124 |
void |
| 1125 |
pgstat_reset_entry(PgStat_Kind kind, Oid dboid, uint64 objid, TimestampTz ts) |
| 1126 |
{ |
| 1127 |
PgStat_EntryRef *entry_ref; |
| 1128 |
|
| 1129 |
Assert(!pgstat_get_kind_info(kind)->fixed_amount); |
| 1130 |
|
| 1131 |
entry_ref = pgstat_get_entry_ref(kind, dboid, objid, false, NULL); |
| 1132 |
if (!entry_ref || entry_ref->shared_entry->dropped) |
| 1133 |
return; |
| 1134 |
|
| 1135 |
(void) pgstat_lock_entry(entry_ref, false); |
| 1136 |
shared_stat_reset_contents(kind, entry_ref->shared_stats, ts); |
| 1137 |
pgstat_unlock_entry(entry_ref); |
| 1138 |
} |
| 1139 |
|
| 1140 |
/* |
| 1141 |
* Scan through the shared hashtable of stats, resetting statistics if |
| 1142 |
* approved by the provided do_reset() function. |
| 1143 |
*/ |
| 1144 |
void |
| 1145 |
pgstat_reset_matching_entries(bool (*do_reset) (PgStatShared_HashEntry *, Datum), |
| 1146 |
Datum match_data, TimestampTz ts) |
| 1147 |
{ |
| 1148 |
dshash_seq_status hstat; |
| 1149 |
PgStatShared_HashEntry *p; |
| 1150 |
|
| 1151 |
/* dshash entry is not modified, take shared lock */ |
| 1152 |
dshash_seq_init(&hstat, pgStatLocal.shared_hash, false); |
| 1153 |
while ((p = dshash_seq_next(&hstat)) != NULL) |
| 1154 |
{ |
| 1155 |
PgStatShared_Common *header; |
| 1156 |
|
| 1157 |
if (p->dropped) |
| 1158 |
continue; |
| 1159 |
|
| 1160 |
if (!do_reset(p, match_data)) |
| 1161 |
continue; |
| 1162 |
|
| 1163 |
header = dsa_get_address(pgStatLocal.dsa, p->body); |
| 1164 |
|
| 1165 |
LWLockAcquire(&header->lock, LW_EXCLUSIVE); |
| 1166 |
|
| 1167 |
shared_stat_reset_contents(p->key.kind, header, ts); |
| 1168 |
|
| 1169 |
LWLockRelease(&header->lock); |
| 1170 |
} |
| 1171 |
dshash_seq_term(&hstat); |
| 1172 |
} |
| 1173 |
|
| 1174 |
static bool |
| 1175 |
match_kind(PgStatShared_HashEntry *p, Datum match_data) |
| 1176 |
{ |
| 1177 |
return p->key.kind == DatumGetInt32(match_data); |
| 1178 |
} |
| 1179 |
|
| 1180 |
void |
| 1181 |
pgstat_reset_entries_of_kind(PgStat_Kind kind, TimestampTz ts) |
| 1182 |
{ |
| 1183 |
pgstat_reset_matching_entries(match_kind, Int32GetDatum(kind), ts); |
| 1184 |
} |
| 1185 |
|
| 1186 |
static void |
| 1187 |
pgstat_setup_memcxt(void) |
| 1188 |
{ |
| 1189 |
if (unlikely(!pgStatSharedRefContext)) |
| 1190 |
pgStatSharedRefContext = |
| 1191 |
AllocSetContextCreate(TopMemoryContext, |
| 1192 |
"PgStat Shared Ref", |
| 1193 |
ALLOCSET_SMALL_SIZES); |
| 1194 |
if (unlikely(!pgStatEntryRefHashContext)) |
| 1195 |
pgStatEntryRefHashContext = |
| 1196 |
AllocSetContextCreate(TopMemoryContext, |
| 1197 |
"PgStat Shared Ref Hash", |
| 1198 |
ALLOCSET_SMALL_SIZES); |
| 1199 |
} |