| 1 |
/* ------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* pgstat_backend.c |
| 4 |
* Implementation of backend statistics. |
| 5 |
* |
| 6 |
* This file contains the implementation of backend statistics. It is kept |
| 7 |
* separate from pgstat.c to enforce the line between the statistics access / |
| 8 |
* storage implementation and the details about individual types of |
| 9 |
* statistics. |
| 10 |
* |
| 11 |
* This statistics kind uses a proc number as object ID for the hash table |
| 12 |
* of pgstats. Entries are created each time a process is spawned, and are |
| 13 |
* dropped when the process exits. These are not written to the pgstats file |
| 14 |
* on disk. Pending statistics are managed without direct interactions with |
| 15 |
* PgStat_EntryRef->pending, relying on PendingBackendStats instead so as it |
| 16 |
* is possible to report data within critical sections. |
| 17 |
* |
| 18 |
* Copyright (c) 2001-2025, PostgreSQL Global Development Group |
| 19 |
* |
| 20 |
* IDENTIFICATION |
| 21 |
* src/backend/utils/activity/pgstat_backend.c |
| 22 |
* ------------------------------------------------------------------------- |
| 23 |
*/ |
| 24 |
|
| 25 |
#include "postgres.h" |
| 26 |
|
| 27 |
#include "access/xlog.h" |
| 28 |
#include "executor/instrument.h" |
| 29 |
#include "storage/bufmgr.h" |
| 30 |
#include "storage/proc.h" |
| 31 |
#include "storage/procarray.h" |
| 32 |
#include "utils/memutils.h" |
| 33 |
#include "utils/pgstat_internal.h" |
| 34 |
|
| 35 |
/* |
| 36 |
* Backend statistics counts waiting to be flushed out. These counters may be |
| 37 |
* reported within critical sections so we use static memory in order to avoid |
| 38 |
* memory allocation. |
| 39 |
*/ |
| 40 |
static PgStat_BackendPending PendingBackendStats; |
| 41 |
static bool backend_has_iostats = false; |
| 42 |
|
| 43 |
/* |
| 44 |
* WAL usage counters saved from pgWalUsage at the previous call to |
| 45 |
* pgstat_flush_backend(). This is used to calculate how much WAL usage |
| 46 |
* happens between pgstat_flush_backend() calls, by subtracting the |
| 47 |
* previous counters from the current ones. |
| 48 |
*/ |
| 49 |
static WalUsage prevBackendWalUsage; |
| 50 |
|
| 51 |
/* |
| 52 |
* Utility routines to report I/O stats for backends, kept here to avoid |
| 53 |
* exposing PendingBackendStats to the outside world. |
| 54 |
*/ |
| 55 |
void |
| 56 |
pgstat_count_backend_io_op_time(IOObject io_object, IOContext io_context, |
| 57 |
IOOp io_op, instr_time io_time) |
| 58 |
{ |
| 59 |
Assert(track_io_timing || track_wal_io_timing); |
| 60 |
|
| 61 |
if (!pgstat_tracks_backend_bktype(MyBackendType)) |
| 62 |
return; |
| 63 |
|
| 64 |
Assert(pgstat_tracks_io_op(MyBackendType, io_object, io_context, io_op)); |
| 65 |
|
| 66 |
INSTR_TIME_ADD(PendingBackendStats.pending_io.pending_times[io_object][io_context][io_op], |
| 67 |
io_time); |
| 68 |
|
| 69 |
backend_has_iostats = true; |
| 70 |
pgstat_report_fixed = true; |
| 71 |
} |
| 72 |
|
| 73 |
void |
| 74 |
pgstat_count_backend_io_op(IOObject io_object, IOContext io_context, |
| 75 |
IOOp io_op, uint32 cnt, uint64 bytes) |
| 76 |
{ |
| 77 |
if (!pgstat_tracks_backend_bktype(MyBackendType)) |
| 78 |
return; |
| 79 |
|
| 80 |
Assert(pgstat_tracks_io_op(MyBackendType, io_object, io_context, io_op)); |
| 81 |
|
| 82 |
PendingBackendStats.pending_io.counts[io_object][io_context][io_op] += cnt; |
| 83 |
PendingBackendStats.pending_io.bytes[io_object][io_context][io_op] += bytes; |
| 84 |
|
| 85 |
backend_has_iostats = true; |
| 86 |
pgstat_report_fixed = true; |
| 87 |
} |
| 88 |
|
| 89 |
/* |
| 90 |
* Returns statistics of a backend by proc number. |
| 91 |
*/ |
| 92 |
PgStat_Backend * |
| 93 |
pgstat_fetch_stat_backend(ProcNumber procNumber) |
| 94 |
{ |
| 95 |
PgStat_Backend *backend_entry; |
| 96 |
|
| 97 |
backend_entry = (PgStat_Backend *) pgstat_fetch_entry(PGSTAT_KIND_BACKEND, |
| 98 |
InvalidOid, procNumber); |
| 99 |
|
| 100 |
return backend_entry; |
| 101 |
} |
| 102 |
|
| 103 |
/* |
| 104 |
* Returns statistics of a backend by pid. |
| 105 |
* |
| 106 |
* This routine includes sanity checks to ensure that the backend exists and |
| 107 |
* is running. "bktype" can be optionally defined to return the BackendType |
| 108 |
* of the backend whose statistics are returned. |
| 109 |
*/ |
| 110 |
PgStat_Backend * |
| 111 |
pgstat_fetch_stat_backend_by_pid(int pid, BackendType *bktype) |
| 112 |
{ |
| 113 |
PGPROC *proc; |
| 114 |
PgBackendStatus *beentry; |
| 115 |
ProcNumber procNumber; |
| 116 |
PgStat_Backend *backend_stats; |
| 117 |
|
| 118 |
proc = BackendPidGetProc(pid); |
| 119 |
if (bktype) |
| 120 |
*bktype = B_INVALID; |
| 121 |
|
| 122 |
/* this could be an auxiliary process */ |
| 123 |
if (!proc) |
| 124 |
proc = AuxiliaryPidGetProc(pid); |
| 125 |
|
| 126 |
if (!proc) |
| 127 |
return NULL; |
| 128 |
|
| 129 |
procNumber = GetNumberFromPGProc(proc); |
| 130 |
|
| 131 |
beentry = pgstat_get_beentry_by_proc_number(procNumber); |
| 132 |
if (!beentry) |
| 133 |
return NULL; |
| 134 |
|
| 135 |
/* check if the backend type tracks statistics */ |
| 136 |
if (!pgstat_tracks_backend_bktype(beentry->st_backendType)) |
| 137 |
return NULL; |
| 138 |
|
| 139 |
/* if PID does not match, leave */ |
| 140 |
if (beentry->st_procpid != pid) |
| 141 |
return NULL; |
| 142 |
|
| 143 |
if (bktype) |
| 144 |
*bktype = beentry->st_backendType; |
| 145 |
|
| 146 |
/* |
| 147 |
* Retrieve the entry. Note that "beentry" may be freed depending on the |
| 148 |
* value of stats_fetch_consistency, so do not access it from this point. |
| 149 |
*/ |
| 150 |
backend_stats = pgstat_fetch_stat_backend(procNumber); |
| 151 |
if (!backend_stats) |
| 152 |
{ |
| 153 |
if (bktype) |
| 154 |
*bktype = B_INVALID; |
| 155 |
return NULL; |
| 156 |
} |
| 157 |
|
| 158 |
return backend_stats; |
| 159 |
} |
| 160 |
|
| 161 |
/* |
| 162 |
* Flush out locally pending backend IO statistics. Locking is managed |
| 163 |
* by the caller. |
| 164 |
*/ |
| 165 |
static void |
| 166 |
pgstat_flush_backend_entry_io(PgStat_EntryRef *entry_ref) |
| 167 |
{ |
| 168 |
PgStatShared_Backend *shbackendent; |
| 169 |
PgStat_BktypeIO *bktype_shstats; |
| 170 |
PgStat_PendingIO pending_io; |
| 171 |
|
| 172 |
/* |
| 173 |
* This function can be called even if nothing at all has happened for IO |
| 174 |
* statistics. In this case, avoid unnecessarily modifying the stats |
| 175 |
* entry. |
| 176 |
*/ |
| 177 |
if (!backend_has_iostats) |
| 178 |
return; |
| 179 |
|
| 180 |
shbackendent = (PgStatShared_Backend *) entry_ref->shared_stats; |
| 181 |
bktype_shstats = &shbackendent->stats.io_stats; |
| 182 |
pending_io = PendingBackendStats.pending_io; |
| 183 |
|
| 184 |
for (int io_object = 0; io_object < IOOBJECT_NUM_TYPES; io_object++) |
| 185 |
{ |
| 186 |
for (int io_context = 0; io_context < IOCONTEXT_NUM_TYPES; io_context++) |
| 187 |
{ |
| 188 |
for (int io_op = 0; io_op < IOOP_NUM_TYPES; io_op++) |
| 189 |
{ |
| 190 |
instr_time time; |
| 191 |
|
| 192 |
bktype_shstats->counts[io_object][io_context][io_op] += |
| 193 |
pending_io.counts[io_object][io_context][io_op]; |
| 194 |
bktype_shstats->bytes[io_object][io_context][io_op] += |
| 195 |
pending_io.bytes[io_object][io_context][io_op]; |
| 196 |
time = pending_io.pending_times[io_object][io_context][io_op]; |
| 197 |
|
| 198 |
bktype_shstats->times[io_object][io_context][io_op] += |
| 199 |
INSTR_TIME_GET_MICROSEC(time); |
| 200 |
} |
| 201 |
} |
| 202 |
} |
| 203 |
|
| 204 |
/* |
| 205 |
* Clear out the statistics buffer, so it can be re-used. |
| 206 |
*/ |
| 207 |
MemSet(&PendingBackendStats.pending_io, 0, sizeof(PgStat_PendingIO)); |
| 208 |
|
| 209 |
backend_has_iostats = false; |
| 210 |
} |
| 211 |
|
| 212 |
/* |
| 213 |
* To determine whether WAL usage happened. |
| 214 |
*/ |
| 215 |
static inline bool |
| 216 |
pgstat_backend_wal_have_pending(void) |
| 217 |
{ |
| 218 |
return (pgWalUsage.wal_records != prevBackendWalUsage.wal_records); |
| 219 |
} |
| 220 |
|
| 221 |
/* |
| 222 |
* Flush out locally pending backend WAL statistics. Locking is managed |
| 223 |
* by the caller. |
| 224 |
*/ |
| 225 |
static void |
| 226 |
pgstat_flush_backend_entry_wal(PgStat_EntryRef *entry_ref) |
| 227 |
{ |
| 228 |
PgStatShared_Backend *shbackendent; |
| 229 |
PgStat_WalCounters *bktype_shstats; |
| 230 |
WalUsage wal_usage_diff = {0}; |
| 231 |
|
| 232 |
/* |
| 233 |
* This function can be called even if nothing at all has happened for WAL |
| 234 |
* statistics. In this case, avoid unnecessarily modifying the stats |
| 235 |
* entry. |
| 236 |
*/ |
| 237 |
if (!pgstat_backend_wal_have_pending()) |
| 238 |
return; |
| 239 |
|
| 240 |
shbackendent = (PgStatShared_Backend *) entry_ref->shared_stats; |
| 241 |
bktype_shstats = &shbackendent->stats.wal_counters; |
| 242 |
|
| 243 |
/* |
| 244 |
* Calculate how much WAL usage counters were increased by subtracting the |
| 245 |
* previous counters from the current ones. |
| 246 |
*/ |
| 247 |
WalUsageAccumDiff(&wal_usage_diff, &pgWalUsage, &prevBackendWalUsage); |
| 248 |
|
| 249 |
#define WALSTAT_ACC(fld, var_to_add) \ |
| 250 |
(bktype_shstats->fld += var_to_add.fld) |
| 251 |
WALSTAT_ACC(wal_buffers_full, wal_usage_diff); |
| 252 |
WALSTAT_ACC(wal_records, wal_usage_diff); |
| 253 |
WALSTAT_ACC(wal_fpi, wal_usage_diff); |
| 254 |
WALSTAT_ACC(wal_bytes, wal_usage_diff); |
| 255 |
#undef WALSTAT_ACC |
| 256 |
|
| 257 |
/* |
| 258 |
* Save the current counters for the subsequent calculation of WAL usage. |
| 259 |
*/ |
| 260 |
prevBackendWalUsage = pgWalUsage; |
| 261 |
} |
| 262 |
|
| 263 |
/* |
| 264 |
* Flush out locally pending backend statistics |
| 265 |
* |
| 266 |
* "flags" parameter controls which statistics to flush. Returns true |
| 267 |
* if some statistics could not be flushed due to lock contention. |
| 268 |
*/ |
| 269 |
bool |
| 270 |
pgstat_flush_backend(bool nowait, bits32 flags) |
| 271 |
{ |
| 272 |
PgStat_EntryRef *entry_ref; |
| 273 |
bool has_pending_data = false; |
| 274 |
|
| 275 |
if (!pgstat_tracks_backend_bktype(MyBackendType)) |
| 276 |
return false; |
| 277 |
|
| 278 |
/* Some IO data pending? */ |
| 279 |
if ((flags & PGSTAT_BACKEND_FLUSH_IO) && backend_has_iostats) |
| 280 |
has_pending_data = true; |
| 281 |
|
| 282 |
/* Some WAL data pending? */ |
| 283 |
if ((flags & PGSTAT_BACKEND_FLUSH_WAL) && |
| 284 |
pgstat_backend_wal_have_pending()) |
| 285 |
has_pending_data = true; |
| 286 |
|
| 287 |
if (!has_pending_data) |
| 288 |
return false; |
| 289 |
|
| 290 |
entry_ref = pgstat_get_entry_ref_locked(PGSTAT_KIND_BACKEND, InvalidOid, |
| 291 |
MyProcNumber, nowait); |
| 292 |
if (!entry_ref) |
| 293 |
return true; |
| 294 |
|
| 295 |
/* Flush requested statistics */ |
| 296 |
if (flags & PGSTAT_BACKEND_FLUSH_IO) |
| 297 |
pgstat_flush_backend_entry_io(entry_ref); |
| 298 |
|
| 299 |
if (flags & PGSTAT_BACKEND_FLUSH_WAL) |
| 300 |
pgstat_flush_backend_entry_wal(entry_ref); |
| 301 |
|
| 302 |
pgstat_unlock_entry(entry_ref); |
| 303 |
|
| 304 |
return false; |
| 305 |
} |
| 306 |
|
| 307 |
/* |
| 308 |
* Callback to flush out locally pending backend statistics. |
| 309 |
* |
| 310 |
* If some stats could not be flushed due to lock contention, return true. |
| 311 |
*/ |
| 312 |
bool |
| 313 |
pgstat_backend_flush_cb(bool nowait) |
| 314 |
{ |
| 315 |
return pgstat_flush_backend(nowait, PGSTAT_BACKEND_FLUSH_ALL); |
| 316 |
} |
| 317 |
|
| 318 |
/* |
| 319 |
* Create backend statistics entry for proc number. |
| 320 |
*/ |
| 321 |
void |
| 322 |
pgstat_create_backend(ProcNumber procnum) |
| 323 |
{ |
| 324 |
PgStat_EntryRef *entry_ref; |
| 325 |
PgStatShared_Backend *shstatent; |
| 326 |
|
| 327 |
entry_ref = pgstat_get_entry_ref_locked(PGSTAT_KIND_BACKEND, InvalidOid, |
| 328 |
procnum, false); |
| 329 |
shstatent = (PgStatShared_Backend *) entry_ref->shared_stats; |
| 330 |
|
| 331 |
/* |
| 332 |
* NB: need to accept that there might be stats from an older backend, |
| 333 |
* e.g. if we previously used this proc number. |
| 334 |
*/ |
| 335 |
memset(&shstatent->stats, 0, sizeof(shstatent->stats)); |
| 336 |
pgstat_unlock_entry(entry_ref); |
| 337 |
|
| 338 |
MemSet(&PendingBackendStats, 0, sizeof(PgStat_BackendPending)); |
| 339 |
backend_has_iostats = false; |
| 340 |
|
| 341 |
/* |
| 342 |
* Initialize prevBackendWalUsage with pgWalUsage so that |
| 343 |
* pgstat_backend_flush_cb() can calculate how much pgWalUsage counters |
| 344 |
* are increased by subtracting prevBackendWalUsage from pgWalUsage. |
| 345 |
*/ |
| 346 |
prevBackendWalUsage = pgWalUsage; |
| 347 |
} |
| 348 |
|
| 349 |
/* |
| 350 |
* Backend statistics are not collected for all BackendTypes. |
| 351 |
* |
| 352 |
* The following BackendTypes do not participate in the backend stats |
| 353 |
* subsystem: |
| 354 |
* - The same and for the same reasons as in pgstat_tracks_io_bktype(). |
| 355 |
* - B_BG_WRITER, B_CHECKPOINTER, B_STARTUP and B_AUTOVAC_LAUNCHER because their |
| 356 |
* I/O stats are already visible in pg_stat_io and there is only one of those. |
| 357 |
* |
| 358 |
* Function returns true if BackendType participates in the backend stats |
| 359 |
* subsystem and false if it does not. |
| 360 |
* |
| 361 |
* When adding a new BackendType, also consider adding relevant restrictions to |
| 362 |
* pgstat_tracks_io_object() and pgstat_tracks_io_op(). |
| 363 |
*/ |
| 364 |
bool |
| 365 |
pgstat_tracks_backend_bktype(BackendType bktype) |
| 366 |
{ |
| 367 |
/* |
| 368 |
* List every type so that new backend types trigger a warning about |
| 369 |
* needing to adjust this switch. |
| 370 |
*/ |
| 371 |
switch (bktype) |
| 372 |
{ |
| 373 |
case B_INVALID: |
| 374 |
case B_AUTOVAC_LAUNCHER: |
| 375 |
case B_DEAD_END_BACKEND: |
| 376 |
case B_ARCHIVER: |
| 377 |
case B_LOGGER: |
| 378 |
case B_BG_WRITER: |
| 379 |
case B_CHECKPOINTER: |
| 380 |
case B_IO_WORKER: |
| 381 |
case B_STARTUP: |
| 382 |
return false; |
| 383 |
|
| 384 |
case B_AUTOVAC_WORKER: |
| 385 |
case B_BACKEND: |
| 386 |
case B_BG_WORKER: |
| 387 |
case B_STANDALONE_BACKEND: |
| 388 |
case B_SLOTSYNC_WORKER: |
| 389 |
case B_WAL_RECEIVER: |
| 390 |
case B_WAL_SENDER: |
| 391 |
case B_WAL_SUMMARIZER: |
| 392 |
case B_WAL_WRITER: |
| 393 |
return true; |
| 394 |
} |
| 395 |
|
| 396 |
return false; |
| 397 |
} |
| 398 |
|
| 399 |
void |
| 400 |
pgstat_backend_reset_timestamp_cb(PgStatShared_Common *header, TimestampTz ts) |
| 401 |
{ |
| 402 |
((PgStatShared_Backend *) header)->stats.stat_reset_timestamp = ts; |
| 403 |
} |