Debian Codegraph @

src/backend/utils/activity/pgstat_io.c

raw
1 /* -------------------------------------------------------------------------
2 *
3 * pgstat_io.c
4 * Implementation of IO statistics.
5 *
6 * This file contains the implementation of IO statistics. It is kept separate
7 * from pgstat.c to enforce the line between the statistics access / storage
8 * implementation and the details about individual types of statistics.
9 *
10 * Copyright (c) 2021-2025, PostgreSQL Global Development Group
11 *
12 * IDENTIFICATION
13 * src/backend/utils/activity/pgstat_io.c
14 * -------------------------------------------------------------------------
15 */
16
17 #include "postgres.h"
18
19 #include "executor/instrument.h"
20 #include "storage/bufmgr.h"
21 #include "utils/pgstat_internal.h"
22
23 static PgStat_PendingIO PendingIOStats;
24 static bool have_iostats = false;
25
26 /*
27 * Check that stats have not been counted for any combination of IOObject,
28 * IOContext, and IOOp which are not tracked for the passed-in BackendType. If
29 * stats are tracked for this combination and IO times are non-zero, counts
30 * should be non-zero.
31 *
32 * The passed-in PgStat_BktypeIO must contain stats from the BackendType
33 * specified by the second parameter. Caller is responsible for locking the
34 * passed-in PgStat_BktypeIO, if needed.
35 */
36 bool
37 pgstat_bktype_io_stats_valid(PgStat_BktypeIO *backend_io,
38 BackendType bktype)
39 {
40 for (int io_object = 0; io_object < IOOBJECT_NUM_TYPES; io_object++)
41 {
42 for (int io_context = 0; io_context < IOCONTEXT_NUM_TYPES; io_context++)
43 {
44 for (int io_op = 0; io_op < IOOP_NUM_TYPES; io_op++)
45 {
46 /* we do track it */
47 if (pgstat_tracks_io_op(bktype, io_object, io_context, io_op))
48 {
49 /* ensure that if IO times are non-zero, counts are > 0 */
50 if (backend_io->times[io_object][io_context][io_op] != 0 &&
51 backend_io->counts[io_object][io_context][io_op] <= 0)
52 return false;
53
54 continue;
55 }
56
57 /* we don't track it, and it is not 0 */
58 if (backend_io->counts[io_object][io_context][io_op] != 0)
59 return false;
60 }
61 }
62 }
63
64 return true;
65 }
66
67 void
68 pgstat_count_io_op(IOObject io_object, IOContext io_context, IOOp io_op,
69 uint32 cnt, uint64 bytes)
70 {
71 Assert((unsigned int) io_object < IOOBJECT_NUM_TYPES);
72 Assert((unsigned int) io_context < IOCONTEXT_NUM_TYPES);
73 Assert(pgstat_is_ioop_tracked_in_bytes(io_op) || bytes == 0);
74 Assert(pgstat_tracks_io_op(MyBackendType, io_object, io_context, io_op));
75
76 PendingIOStats.counts[io_object][io_context][io_op] += cnt;
77 PendingIOStats.bytes[io_object][io_context][io_op] += bytes;
78
79 /* Add the per-backend counts */
80 pgstat_count_backend_io_op(io_object, io_context, io_op, cnt, bytes);
81
82 have_iostats = true;
83 pgstat_report_fixed = true;
84 }
85
86 /*
87 * Initialize the internal timing for an IO operation, depending on an
88 * IO timing GUC.
89 */
90 instr_time
91 pgstat_prepare_io_time(bool track_io_guc)
92 {
93 instr_time io_start;
94
95 if (track_io_guc)
96 INSTR_TIME_SET_CURRENT(io_start);
97 else
98 {
99 /*
100 * There is no need to set io_start when an IO timing GUC is disabled.
101 * Initialize it to zero to avoid compiler warnings and to let
102 * pgstat_count_io_op_time() know that timings should be ignored.
103 */
104 INSTR_TIME_SET_ZERO(io_start);
105 }
106
107 return io_start;
108 }
109
110 /*
111 * Like pgstat_count_io_op() except it also accumulates time.
112 *
113 * The calls related to pgstat_count_buffer_*() are for pgstat_database. As
114 * pg_stat_database only counts block read and write times, these are done for
115 * IOOP_READ, IOOP_WRITE and IOOP_EXTEND.
116 *
117 * pgBufferUsage is used for EXPLAIN. pgBufferUsage has write and read stats
118 * for shared, local and temporary blocks. pg_stat_io does not track the
119 * activity of temporary blocks, so these are ignored here.
120 */
121 void
122 pgstat_count_io_op_time(IOObject io_object, IOContext io_context, IOOp io_op,
123 instr_time start_time, uint32 cnt, uint64 bytes)
124 {
125 if (!INSTR_TIME_IS_ZERO(start_time))
126 {
127 instr_time io_time;
128
129 INSTR_TIME_SET_CURRENT(io_time);
130 INSTR_TIME_SUBTRACT(io_time, start_time);
131
132 if (io_object != IOOBJECT_WAL)
133 {
134 if (io_op == IOOP_WRITE || io_op == IOOP_EXTEND)
135 {
136 pgstat_count_buffer_write_time(INSTR_TIME_GET_MICROSEC(io_time));
137 if (io_object == IOOBJECT_RELATION)
138 INSTR_TIME_ADD(pgBufferUsage.shared_blk_write_time, io_time);
139 else if (io_object == IOOBJECT_TEMP_RELATION)
140 INSTR_TIME_ADD(pgBufferUsage.local_blk_write_time, io_time);
141 }
142 else if (io_op == IOOP_READ)
143 {
144 pgstat_count_buffer_read_time(INSTR_TIME_GET_MICROSEC(io_time));
145 if (io_object == IOOBJECT_RELATION)
146 INSTR_TIME_ADD(pgBufferUsage.shared_blk_read_time, io_time);
147 else if (io_object == IOOBJECT_TEMP_RELATION)
148 INSTR_TIME_ADD(pgBufferUsage.local_blk_read_time, io_time);
149 }
150 }
151
152 INSTR_TIME_ADD(PendingIOStats.pending_times[io_object][io_context][io_op],
153 io_time);
154
155 /* Add the per-backend count */
156 pgstat_count_backend_io_op_time(io_object, io_context, io_op,
157 io_time);
158 }
159
160 pgstat_count_io_op(io_object, io_context, io_op, cnt, bytes);
161 }
162
163 PgStat_IO *
164 pgstat_fetch_stat_io(void)
165 {
166 pgstat_snapshot_fixed(PGSTAT_KIND_IO);
167
168 return &pgStatLocal.snapshot.io;
169 }
170
171 /*
172 * Simpler wrapper of pgstat_io_flush_cb()
173 */
174 void
175 pgstat_flush_io(bool nowait)
176 {
177 (void) pgstat_io_flush_cb(nowait);
178 }
179
180 /*
181 * Flush out locally pending IO statistics
182 *
183 * If no stats have been recorded, this function returns false.
184 *
185 * If nowait is true, this function returns true if the lock could not be
186 * acquired. Otherwise, return false.
187 */
188 bool
189 pgstat_io_flush_cb(bool nowait)
190 {
191 LWLock *bktype_lock;
192 PgStat_BktypeIO *bktype_shstats;
193
194 if (!have_iostats)
195 return false;
196
197 bktype_lock = &pgStatLocal.shmem->io.locks[MyBackendType];
198 bktype_shstats =
199 &pgStatLocal.shmem->io.stats.stats[MyBackendType];
200
201 if (!nowait)
202 LWLockAcquire(bktype_lock, LW_EXCLUSIVE);
203 else if (!LWLockConditionalAcquire(bktype_lock, LW_EXCLUSIVE))
204 return true;
205
206 for (int io_object = 0; io_object < IOOBJECT_NUM_TYPES; io_object++)
207 {
208 for (int io_context = 0; io_context < IOCONTEXT_NUM_TYPES; io_context++)
209 {
210 for (int io_op = 0; io_op < IOOP_NUM_TYPES; io_op++)
211 {
212 instr_time time;
213
214 bktype_shstats->counts[io_object][io_context][io_op] +=
215 PendingIOStats.counts[io_object][io_context][io_op];
216
217 bktype_shstats->bytes[io_object][io_context][io_op] +=
218 PendingIOStats.bytes[io_object][io_context][io_op];
219
220 time = PendingIOStats.pending_times[io_object][io_context][io_op];
221
222 bktype_shstats->times[io_object][io_context][io_op] +=
223 INSTR_TIME_GET_MICROSEC(time);
224 }
225 }
226 }
227
228 Assert(pgstat_bktype_io_stats_valid(bktype_shstats, MyBackendType));
229
230 LWLockRelease(bktype_lock);
231
232 memset(&PendingIOStats, 0, sizeof(PendingIOStats));
233
234 have_iostats = false;
235
236 return false;
237 }
238
239 const char *
240 pgstat_get_io_context_name(IOContext io_context)
241 {
242 switch (io_context)
243 {
244 case IOCONTEXT_BULKREAD:
245 return "bulkread";
246 case IOCONTEXT_BULKWRITE:
247 return "bulkwrite";
248 case IOCONTEXT_INIT:
249 return "init";
250 case IOCONTEXT_NORMAL:
251 return "normal";
252 case IOCONTEXT_VACUUM:
253 return "vacuum";
254 }
255
256 elog(ERROR, "unrecognized IOContext value: %d", io_context);
257 pg_unreachable();
258 }
259
260 const char *
261 pgstat_get_io_object_name(IOObject io_object)
262 {
263 switch (io_object)
264 {
265 case IOOBJECT_RELATION:
266 return "relation";
267 case IOOBJECT_TEMP_RELATION:
268 return "temp relation";
269 case IOOBJECT_WAL:
270 return "wal";
271 }
272
273 elog(ERROR, "unrecognized IOObject value: %d", io_object);
274 pg_unreachable();
275 }
276
277 void
278 pgstat_io_init_shmem_cb(void *stats)
279 {
280 PgStatShared_IO *stat_shmem = (PgStatShared_IO *) stats;
281
282 for (int i = 0; i < BACKEND_NUM_TYPES; i++)
283 LWLockInitialize(&stat_shmem->locks[i], LWTRANCHE_PGSTATS_DATA);
284 }
285
286 void
287 pgstat_io_reset_all_cb(TimestampTz ts)
288 {
289 for (int i = 0; i < BACKEND_NUM_TYPES; i++)
290 {
291 LWLock *bktype_lock = &pgStatLocal.shmem->io.locks[i];
292 PgStat_BktypeIO *bktype_shstats = &pgStatLocal.shmem->io.stats.stats[i];
293
294 LWLockAcquire(bktype_lock, LW_EXCLUSIVE);
295
296 /*
297 * Use the lock in the first BackendType's PgStat_BktypeIO to protect
298 * the reset timestamp as well.
299 */
300 if (i == 0)
301 pgStatLocal.shmem->io.stats.stat_reset_timestamp = ts;
302
303 memset(bktype_shstats, 0, sizeof(*bktype_shstats));
304 LWLockRelease(bktype_lock);
305 }
306 }
307
308 void
309 pgstat_io_snapshot_cb(void)
310 {
311 for (int i = 0; i < BACKEND_NUM_TYPES; i++)
312 {
313 LWLock *bktype_lock = &pgStatLocal.shmem->io.locks[i];
314 PgStat_BktypeIO *bktype_shstats = &pgStatLocal.shmem->io.stats.stats[i];
315 PgStat_BktypeIO *bktype_snap = &pgStatLocal.snapshot.io.stats[i];
316
317 LWLockAcquire(bktype_lock, LW_SHARED);
318
319 /*
320 * Use the lock in the first BackendType's PgStat_BktypeIO to protect
321 * the reset timestamp as well.
322 */
323 if (i == 0)
324 pgStatLocal.snapshot.io.stat_reset_timestamp =
325 pgStatLocal.shmem->io.stats.stat_reset_timestamp;
326
327 /* using struct assignment due to better type safety */
328 *bktype_snap = *bktype_shstats;
329 LWLockRelease(bktype_lock);
330 }
331 }
332
333 /*
334 * IO statistics are not collected for all BackendTypes.
335 *
336 * The following BackendTypes do not participate in the cumulative stats
337 * subsystem or do not perform IO on which we currently track:
338 * - Dead-end backend because it is not connected to shared memory and
339 * doesn't do any IO
340 * - Syslogger because it is not connected to shared memory
341 * - Archiver because most relevant archiving IO is delegated to a
342 * specialized command or module
343 *
344 * Function returns true if BackendType participates in the cumulative stats
345 * subsystem for IO and false if it does not.
346 *
347 * When adding a new BackendType, also consider adding relevant restrictions to
348 * pgstat_tracks_io_object() and pgstat_tracks_io_op().
349 */
350 bool
351 pgstat_tracks_io_bktype(BackendType bktype)
352 {
353 /*
354 * List every type so that new backend types trigger a warning about
355 * needing to adjust this switch.
356 */
357 switch (bktype)
358 {
359 case B_INVALID:
360 case B_DEAD_END_BACKEND:
361 case B_ARCHIVER:
362 case B_LOGGER:
363 return false;
364
365 case B_AUTOVAC_LAUNCHER:
366 case B_AUTOVAC_WORKER:
367 case B_BACKEND:
368 case B_BG_WORKER:
369 case B_BG_WRITER:
370 case B_CHECKPOINTER:
371 case B_IO_WORKER:
372 case B_SLOTSYNC_WORKER:
373 case B_STANDALONE_BACKEND:
374 case B_STARTUP:
375 case B_WAL_RECEIVER:
376 case B_WAL_SENDER:
377 case B_WAL_SUMMARIZER:
378 case B_WAL_WRITER:
379 return true;
380 }
381
382 return false;
383 }
384
385 /*
386 * Some BackendTypes do not perform IO on certain IOObjects or in certain
387 * IOContexts. Some IOObjects are never operated on in some IOContexts. Check
388 * that the given BackendType is expected to do IO in the given IOContext and
389 * on the given IOObject and that the given IOObject is expected to be operated
390 * on in the given IOContext.
391 */
392 bool
393 pgstat_tracks_io_object(BackendType bktype, IOObject io_object,
394 IOContext io_context)
395 {
396 bool no_temp_rel;
397
398 /*
399 * Some BackendTypes should never track IO statistics.
400 */
401 if (!pgstat_tracks_io_bktype(bktype))
402 return false;
403
404 /*
405 * Currently, IO on IOOBJECT_WAL objects can only occur in the
406 * IOCONTEXT_NORMAL and IOCONTEXT_INIT IOContexts.
407 */
408 if (io_object == IOOBJECT_WAL &&
409 (io_context != IOCONTEXT_NORMAL &&
410 io_context != IOCONTEXT_INIT))
411 return false;
412
413 /*
414 * Currently, IO on temporary relations can only occur in the
415 * IOCONTEXT_NORMAL IOContext.
416 */
417 if (io_context != IOCONTEXT_NORMAL &&
418 io_object == IOOBJECT_TEMP_RELATION)
419 return false;
420
421 /*
422 * In core Postgres, only regular backends and WAL Sender processes
423 * executing queries will use local buffers and operate on temporary
424 * relations. Parallel workers will not use local buffers (see
425 * InitLocalBuffers()); however, extensions leveraging background workers
426 * have no such limitation, so track IO on IOOBJECT_TEMP_RELATION for
427 * BackendType B_BG_WORKER.
428 */
429 no_temp_rel = bktype == B_AUTOVAC_LAUNCHER || bktype == B_BG_WRITER ||
430 bktype == B_CHECKPOINTER || bktype == B_AUTOVAC_WORKER ||
431 bktype == B_STANDALONE_BACKEND || bktype == B_STARTUP ||
432 bktype == B_WAL_SUMMARIZER || bktype == B_WAL_WRITER ||
433 bktype == B_WAL_RECEIVER;
434
435 if (no_temp_rel && io_context == IOCONTEXT_NORMAL &&
436 io_object == IOOBJECT_TEMP_RELATION)
437 return false;
438
439 /*
440 * Some BackendTypes only perform IO under IOOBJECT_WAL, hence exclude all
441 * rows for all the other objects for these.
442 */
443 if ((bktype == B_WAL_SUMMARIZER || bktype == B_WAL_RECEIVER ||
444 bktype == B_WAL_WRITER) && io_object != IOOBJECT_WAL)
445 return false;
446
447 /*
448 * Some BackendTypes do not currently perform any IO in certain
449 * IOContexts, and, while it may not be inherently incorrect for them to
450 * do so, excluding those rows from the view makes the view easier to use.
451 */
452 if ((bktype == B_CHECKPOINTER || bktype == B_BG_WRITER) &&
453 (io_context == IOCONTEXT_BULKREAD ||
454 io_context == IOCONTEXT_BULKWRITE ||
455 io_context == IOCONTEXT_VACUUM))
456 return false;
457
458 if (bktype == B_AUTOVAC_LAUNCHER && io_context == IOCONTEXT_VACUUM)
459 return false;
460
461 if ((bktype == B_AUTOVAC_WORKER || bktype == B_AUTOVAC_LAUNCHER) &&
462 io_context == IOCONTEXT_BULKWRITE)
463 return false;
464
465 return true;
466 }
467
468 /*
469 * Some BackendTypes will never do certain IOOps and some IOOps should not
470 * occur in certain IOContexts or on certain IOObjects. Check that the given
471 * IOOp is valid for the given BackendType in the given IOContext and on the
472 * given IOObject. Note that there are currently no cases of an IOOp being
473 * invalid for a particular BackendType only within a certain IOContext and/or
474 * only on a certain IOObject.
475 */
476 bool
477 pgstat_tracks_io_op(BackendType bktype, IOObject io_object,
478 IOContext io_context, IOOp io_op)
479 {
480 bool strategy_io_context;
481
482 /* if (io_context, io_object) will never collect stats, we're done */
483 if (!pgstat_tracks_io_object(bktype, io_object, io_context))
484 return false;
485
486 /*
487 * Some BackendTypes will not do certain IOOps.
488 */
489 if (bktype == B_BG_WRITER &&
490 (io_op == IOOP_READ || io_op == IOOP_EVICT || io_op == IOOP_HIT))
491 return false;
492
493 if (bktype == B_CHECKPOINTER &&
494 ((io_object != IOOBJECT_WAL && io_op == IOOP_READ) ||
495 (io_op == IOOP_EVICT || io_op == IOOP_HIT)))
496 return false;
497
498 if ((bktype == B_AUTOVAC_LAUNCHER || bktype == B_BG_WRITER ||
499 bktype == B_CHECKPOINTER) && io_op == IOOP_EXTEND)
500 return false;
501
502 /*
503 * Some BackendTypes do not perform reads with IOOBJECT_WAL.
504 */
505 if (io_object == IOOBJECT_WAL && io_op == IOOP_READ &&
506 (bktype == B_WAL_RECEIVER || bktype == B_BG_WRITER ||
507 bktype == B_AUTOVAC_LAUNCHER || bktype == B_AUTOVAC_WORKER ||
508 bktype == B_WAL_WRITER))
509 return false;
510
511 /*
512 * Temporary tables are not logged and thus do not require fsync'ing.
513 * Writeback is not requested for temporary tables.
514 */
515 if (io_object == IOOBJECT_TEMP_RELATION &&
516 (io_op == IOOP_FSYNC || io_op == IOOP_WRITEBACK))
517 return false;
518
519 /*
520 * Some IOOps are not valid in certain IOContexts and some IOOps are only
521 * valid in certain contexts.
522 */
523 if (io_context == IOCONTEXT_BULKREAD && io_op == IOOP_EXTEND)
524 return false;
525
526 strategy_io_context = io_context == IOCONTEXT_BULKREAD ||
527 io_context == IOCONTEXT_BULKWRITE || io_context == IOCONTEXT_VACUUM;
528
529 /*
530 * IOOP_REUSE is only relevant when a BufferAccessStrategy is in use.
531 */
532 if (!strategy_io_context && io_op == IOOP_REUSE)
533 return false;
534
535 /*
536 * IOOBJECT_WAL IOObject will not do certain IOOps depending on IOContext.
537 */
538 if (io_object == IOOBJECT_WAL && io_context == IOCONTEXT_INIT &&
539 !(io_op == IOOP_WRITE || io_op == IOOP_FSYNC))
540 return false;
541
542 if (io_object == IOOBJECT_WAL && io_context == IOCONTEXT_NORMAL &&
543 !(io_op == IOOP_WRITE || io_op == IOOP_READ || io_op == IOOP_FSYNC))
544 return false;
545
546 /*
547 * IOOP_FSYNC IOOps done by a backend using a BufferAccessStrategy are
548 * counted in the IOCONTEXT_NORMAL IOContext. See comment in
549 * register_dirty_segment() for more details.
550 */
551 if (strategy_io_context && io_op == IOOP_FSYNC)
552 return false;
553
554
555 return true;
556 }