Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * aio.c
4 : * AIO - Core Logic
5 : *
6 : * For documentation about how AIO works on a higher level, including a
7 : * schematic example, see README.md.
8 : *
9 : *
10 : * AIO is a complicated subsystem. To keep things navigable, it is split
11 : * across a number of files:
12 : *
13 : * - method_*.c - different ways of executing AIO (e.g. worker process)
14 : *
15 : * - aio_target.c - IO on different kinds of targets
16 : *
17 : * - aio_io.c - method-independent code for specific IO ops (e.g. readv)
18 : *
19 : * - aio_callback.c - callbacks at IO operation lifecycle events
20 : *
21 : * - aio_init.c - per-server and per-backend initialization
22 : *
23 : * - aio.c - all other topics
24 : *
25 : * - read_stream.c - helper for reading buffered relation data
26 : *
27 : * - README.md - higher-level overview over AIO
28 : *
29 : *
30 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
31 : * Portions Copyright (c) 1994, Regents of the University of California
32 : *
33 : * IDENTIFICATION
34 : * src/backend/storage/aio/aio.c
35 : *
36 : *-------------------------------------------------------------------------
37 : */
38 :
39 : #include "postgres.h"
40 :
41 : #include "lib/ilist.h"
42 : #include "miscadmin.h"
43 : #include "port/atomics.h"
44 : #include "storage/aio.h"
45 : #include "storage/aio_internal.h"
46 : #include "storage/aio_subsys.h"
47 : #include "utils/guc.h"
48 : #include "utils/guc_hooks.h"
49 : #include "utils/resowner.h"
50 : #include "utils/wait_event_types.h"
51 :
52 : #ifdef USE_INJECTION_POINTS
53 : #include "utils/injection_point.h"
54 : #endif
55 :
56 :
57 : static inline void pgaio_io_update_state(PgAioHandle *ioh, PgAioHandleState new_state);
58 : static void pgaio_io_reclaim(PgAioHandle *ioh);
59 : static void pgaio_io_resowner_register(PgAioHandle *ioh);
60 : static void pgaio_io_wait_for_free(void);
61 : static PgAioHandle *pgaio_io_from_wref(PgAioWaitRef *iow, uint64 *ref_generation);
62 : static const char *pgaio_io_state_get_name(PgAioHandleState s);
63 : static void pgaio_io_wait(PgAioHandle *ioh, uint64 ref_generation);
64 :
65 :
66 : /* Options for io_method. */
67 : const struct config_enum_entry io_method_options[] = {
68 : {"sync", IOMETHOD_SYNC, false},
69 : {"worker", IOMETHOD_WORKER, false},
70 : #ifdef IOMETHOD_IO_URING_ENABLED
71 : {"io_uring", IOMETHOD_IO_URING, false},
72 : #endif
73 : {NULL, 0, false}
74 : };
75 :
76 : /* GUCs */
77 : int io_method = DEFAULT_IO_METHOD;
78 : int io_max_concurrency = -1;
79 :
80 : /* global control for AIO */
81 : PgAioCtl *pgaio_ctl;
82 :
83 : /* current backend's per-backend state */
84 : PgAioBackend *pgaio_my_backend;
85 :
86 :
87 : static const IoMethodOps *const pgaio_method_ops_table[] = {
88 : [IOMETHOD_SYNC] = &pgaio_sync_ops,
89 : [IOMETHOD_WORKER] = &pgaio_worker_ops,
90 : #ifdef IOMETHOD_IO_URING_ENABLED
91 : [IOMETHOD_IO_URING] = &pgaio_uring_ops,
92 : #endif
93 : };
94 :
95 : /* callbacks for the configured io_method, set by assign_io_method */
96 : const IoMethodOps *pgaio_method_ops;
97 :
98 :
99 : /*
100 : * Currently there's no infrastructure to pass arguments to injection points,
101 : * so we instead set this up for the duration of the injection point
102 : * invocation. See pgaio_io_call_inj().
103 : */
104 : #ifdef USE_INJECTION_POINTS
105 : static PgAioHandle *pgaio_inj_cur_handle;
106 : #endif
107 :
108 :
109 :
110 : /* --------------------------------------------------------------------------------
111 : * Public Functions related to PgAioHandle
112 : * --------------------------------------------------------------------------------
113 : */
114 :
115 : /*
116 : * Acquire an AioHandle, waiting for IO completion if necessary.
117 : *
118 : * Each backend can only have one AIO handle that has been "handed out" to
119 : * code, but not yet submitted or released. This restriction is necessary to
120 : * ensure that it is possible for code to wait for an unused handle by waiting
121 : * for in-flight IO to complete. There is a limited number of handles in each
122 : * backend, if multiple handles could be handed out without being submitted,
123 : * waiting for all in-flight IO to complete would not guarantee that handles
124 : * free up.
125 : *
126 : * It is cheap to acquire an IO handle, unless all handles are in use. In that
127 : * case this function waits for the oldest IO to complete. If that is not
128 : * desirable, use pgaio_io_acquire_nb().
129 : *
130 : * If a handle was acquired but then does not turn out to be needed,
131 : * e.g. because pgaio_io_acquire() is called before starting an IO in a
132 : * critical section, the handle needs to be released with pgaio_io_release().
133 : *
134 : *
135 : * To react to the completion of the IO as soon as it is known to have
136 : * completed, callbacks can be registered with pgaio_io_register_callbacks().
137 : *
138 : * To actually execute IO using the returned handle, the pgaio_io_start_*()
139 : * family of functions is used. In many cases the pgaio_io_start_*() call will
140 : * not be done directly by code that acquired the handle, but by lower level
141 : * code that gets passed the handle. E.g. if code in bufmgr.c wants to perform
142 : * AIO, it typically will pass the handle to smgr.c, which will pass it on to
143 : * md.c, on to fd.c, which then finally calls pgaio_io_start_*(). This
144 : * forwarding allows the various layers to react to the IO's completion by
145 : * registering callbacks. These callbacks in turn can translate a lower
146 : * layer's result into a result understandable by a higher layer.
147 : *
148 : * During pgaio_io_start_*() the IO is staged (i.e. prepared for execution but
149 : * not submitted to the kernel). Unless in batchmode
150 : * (c.f. pgaio_enter_batchmode()), the IO will also get submitted for
151 : * execution. Note that, whether in batchmode or not, the IO might even
152 : * complete before the functions return.
153 : *
154 : * After pgaio_io_start_*() the AioHandle is "consumed" and may not be
155 : * referenced by the IO issuing code. To e.g. wait for IO, references to the
156 : * IO can be established with pgaio_io_get_wref() *before* pgaio_io_start_*()
157 : * is called. pgaio_wref_wait() can be used to wait for the IO to complete.
158 : *
159 : *
160 : * To know if the IO [partially] succeeded or failed, a PgAioReturn * can be
161 : * passed to pgaio_io_acquire(). Once the issuing backend has called
162 : * pgaio_wref_wait(), the PgAioReturn contains information about whether the
163 : * operation succeeded and details about the first failure, if any. The error
164 : * can be raised / logged with pgaio_result_report().
165 : *
166 : * The lifetime of the memory pointed to be *ret needs to be at least as long
167 : * as the passed in resowner. If the resowner releases resources before the IO
168 : * completes (typically due to an error), the reference to *ret will be
169 : * cleared. In case of resowner cleanup *ret will not be updated with the
170 : * results of the IO operation.
171 : */
172 : PgAioHandle *
173 10432 : pgaio_io_acquire(struct ResourceOwnerData *resowner, PgAioReturn *ret)
174 : {
175 : PgAioHandle *h;
176 :
177 : while (true)
178 : {
179 10432 : h = pgaio_io_acquire_nb(resowner, ret);
180 :
181 10428 : if (h != NULL)
182 5312 : return h;
183 :
184 : /*
185 : * Evidently all handles by this backend are in use. Just wait for
186 : * some to complete.
187 : */
188 5116 : pgaio_io_wait_for_free();
189 : }
190 : }
191 :
192 : /*
193 : * Acquire an AioHandle, returning NULL if no handles are free.
194 : *
195 : * See pgaio_io_acquire(). The only difference is that this function will return
196 : * NULL if there are no idle handles, instead of blocking.
197 : */
198 : PgAioHandle *
199 2480958 : pgaio_io_acquire_nb(struct ResourceOwnerData *resowner, PgAioReturn *ret)
200 : {
201 2480958 : if (pgaio_my_backend->num_staged_ios >= PGAIO_SUBMIT_BATCH_SIZE)
202 : {
203 : Assert(pgaio_my_backend->num_staged_ios == PGAIO_SUBMIT_BATCH_SIZE);
204 0 : pgaio_submit_staged();
205 : }
206 :
207 2480958 : if (pgaio_my_backend->handed_out_io)
208 4 : elog(ERROR, "API violation: Only one IO can be handed out");
209 :
210 2480954 : if (!dclist_is_empty(&pgaio_my_backend->idle_ios))
211 : {
212 2470722 : dlist_node *ion = dclist_pop_head_node(&pgaio_my_backend->idle_ios);
213 2470722 : PgAioHandle *ioh = dclist_container(PgAioHandle, node, ion);
214 :
215 : Assert(ioh->state == PGAIO_HS_IDLE);
216 : Assert(ioh->owner_procno == MyProcNumber);
217 :
218 2470722 : pgaio_io_update_state(ioh, PGAIO_HS_HANDED_OUT);
219 2470722 : pgaio_my_backend->handed_out_io = ioh;
220 :
221 2470722 : if (resowner)
222 2470722 : pgaio_io_resowner_register(ioh);
223 :
224 2470722 : if (ret)
225 : {
226 2470670 : ioh->report_return = ret;
227 2470670 : ret->result.status = PGAIO_RS_UNKNOWN;
228 : }
229 :
230 2470722 : return ioh;
231 : }
232 :
233 10232 : return NULL;
234 : }
235 :
236 : /*
237 : * Release IO handle that turned out to not be required.
238 : *
239 : * See pgaio_io_acquire() for more details.
240 : */
241 : void
242 5016 : pgaio_io_release(PgAioHandle *ioh)
243 : {
244 5016 : if (ioh == pgaio_my_backend->handed_out_io)
245 : {
246 : Assert(ioh->state == PGAIO_HS_HANDED_OUT);
247 : Assert(ioh->resowner);
248 :
249 5012 : pgaio_my_backend->handed_out_io = NULL;
250 5012 : pgaio_io_reclaim(ioh);
251 : }
252 : else
253 : {
254 4 : elog(ERROR, "release in unexpected state");
255 : }
256 5012 : }
257 :
258 : /*
259 : * Release IO handle during resource owner cleanup.
260 : */
261 : void
262 86 : pgaio_io_release_resowner(dlist_node *ioh_node, bool on_error)
263 : {
264 86 : PgAioHandle *ioh = dlist_container(PgAioHandle, resowner_node, ioh_node);
265 :
266 : Assert(ioh->resowner);
267 :
268 86 : ResourceOwnerForgetAioHandle(ioh->resowner, &ioh->resowner_node);
269 86 : ioh->resowner = NULL;
270 :
271 86 : switch (ioh->state)
272 : {
273 0 : case PGAIO_HS_IDLE:
274 0 : elog(ERROR, "unexpected");
275 : break;
276 66 : case PGAIO_HS_HANDED_OUT:
277 : Assert(ioh == pgaio_my_backend->handed_out_io || pgaio_my_backend->handed_out_io == NULL);
278 :
279 66 : if (ioh == pgaio_my_backend->handed_out_io)
280 : {
281 66 : pgaio_my_backend->handed_out_io = NULL;
282 66 : if (!on_error)
283 20 : elog(WARNING, "leaked AIO handle");
284 : }
285 :
286 66 : pgaio_io_reclaim(ioh);
287 66 : break;
288 0 : case PGAIO_HS_DEFINED:
289 : case PGAIO_HS_STAGED:
290 0 : if (!on_error)
291 0 : elog(WARNING, "AIO handle was not submitted");
292 0 : pgaio_submit_staged();
293 0 : break;
294 20 : case PGAIO_HS_SUBMITTED:
295 : case PGAIO_HS_COMPLETED_IO:
296 : case PGAIO_HS_COMPLETED_SHARED:
297 : case PGAIO_HS_COMPLETED_LOCAL:
298 : /* this is expected to happen */
299 20 : break;
300 : }
301 :
302 : /*
303 : * Need to unregister the reporting of the IO's result, the memory it's
304 : * referencing likely has gone away.
305 : */
306 86 : if (ioh->report_return)
307 20 : ioh->report_return = NULL;
308 86 : }
309 :
310 : /*
311 : * Add a [set of] flags to the IO.
312 : *
313 : * Note that this combines flags with already set flags, rather than set flags
314 : * to explicitly the passed in parameters. This is to allow multiple callsites
315 : * to set flags.
316 : */
317 : void
318 4928446 : pgaio_io_set_flag(PgAioHandle *ioh, PgAioHandleFlags flag)
319 : {
320 : Assert(ioh->state == PGAIO_HS_HANDED_OUT);
321 :
322 4928446 : ioh->flags |= flag;
323 4928446 : }
324 :
325 : /*
326 : * Returns an ID uniquely identifying the IO handle. This is only really
327 : * useful for logging, as handles are reused across multiple IOs.
328 : */
329 : int
330 1209862 : pgaio_io_get_id(PgAioHandle *ioh)
331 : {
332 : Assert(ioh >= pgaio_ctl->io_handles &&
333 : ioh < (pgaio_ctl->io_handles + pgaio_ctl->io_handle_count));
334 1209862 : return ioh - pgaio_ctl->io_handles;
335 : }
336 :
337 : /*
338 : * Return the ProcNumber for the process that can use an IO handle. The
339 : * mapping from IO handles to PGPROCs is static, therefore this even works
340 : * when the corresponding PGPROC is not in use.
341 : */
342 : ProcNumber
343 0 : pgaio_io_get_owner(PgAioHandle *ioh)
344 : {
345 0 : return ioh->owner_procno;
346 : }
347 :
348 : /*
349 : * Return a wait reference for the IO. Only wait references can be used to
350 : * wait for an IOs completion, as handles themselves can be reused after
351 : * completion. See also the comment above pgaio_io_acquire().
352 : */
353 : void
354 4931318 : pgaio_io_get_wref(PgAioHandle *ioh, PgAioWaitRef *iow)
355 : {
356 : Assert(ioh->state == PGAIO_HS_HANDED_OUT ||
357 : ioh->state == PGAIO_HS_DEFINED ||
358 : ioh->state == PGAIO_HS_STAGED);
359 : Assert(ioh->generation != 0);
360 :
361 4931318 : iow->aio_index = ioh - pgaio_ctl->io_handles;
362 4931318 : iow->generation_upper = (uint32) (ioh->generation >> 32);
363 4931318 : iow->generation_lower = (uint32) ioh->generation;
364 4931318 : }
365 :
366 :
367 :
368 : /* --------------------------------------------------------------------------------
369 : * Internal Functions related to PgAioHandle
370 : * --------------------------------------------------------------------------------
371 : */
372 :
373 : static inline void
374 19314784 : pgaio_io_update_state(PgAioHandle *ioh, PgAioHandleState new_state)
375 : {
376 19314784 : pgaio_debug_io(DEBUG5, ioh,
377 : "updating state to %s",
378 : pgaio_io_state_get_name(new_state));
379 :
380 : /*
381 : * Ensure the changes signified by the new state are visible before the
382 : * new state becomes visible.
383 : */
384 19314784 : pg_write_barrier();
385 :
386 19314784 : ioh->state = new_state;
387 19314784 : }
388 :
389 : static void
390 2470722 : pgaio_io_resowner_register(PgAioHandle *ioh)
391 : {
392 : Assert(!ioh->resowner);
393 : Assert(CurrentResourceOwner);
394 :
395 2470722 : ResourceOwnerRememberAioHandle(CurrentResourceOwner, &ioh->resowner_node);
396 2470722 : ioh->resowner = CurrentResourceOwner;
397 2470722 : }
398 :
399 : /*
400 : * Stage IO for execution and, if appropriate, submit it immediately.
401 : *
402 : * Should only be called from pgaio_io_start_*().
403 : */
404 : void
405 2465644 : pgaio_io_stage(PgAioHandle *ioh, PgAioOp op)
406 : {
407 : bool needs_synchronous;
408 :
409 : Assert(ioh->state == PGAIO_HS_HANDED_OUT);
410 : Assert(pgaio_my_backend->handed_out_io == ioh);
411 : Assert(pgaio_io_has_target(ioh));
412 :
413 2465644 : ioh->op = op;
414 2465644 : ioh->result = 0;
415 :
416 2465644 : pgaio_io_update_state(ioh, PGAIO_HS_DEFINED);
417 :
418 : /* allow a new IO to be staged */
419 2465644 : pgaio_my_backend->handed_out_io = NULL;
420 :
421 2465644 : pgaio_io_call_stage(ioh);
422 :
423 2465644 : pgaio_io_update_state(ioh, PGAIO_HS_STAGED);
424 :
425 : /*
426 : * Synchronous execution has to be executed, well, synchronously, so check
427 : * that first.
428 : */
429 2465644 : needs_synchronous = pgaio_io_needs_synchronous_execution(ioh);
430 :
431 2465644 : pgaio_debug_io(DEBUG3, ioh,
432 : "staged (synchronous: %d, in_batch: %d)",
433 : needs_synchronous, pgaio_my_backend->in_batchmode);
434 :
435 2465644 : if (!needs_synchronous)
436 : {
437 1152156 : pgaio_my_backend->staged_ios[pgaio_my_backend->num_staged_ios++] = ioh;
438 : Assert(pgaio_my_backend->num_staged_ios <= PGAIO_SUBMIT_BATCH_SIZE);
439 :
440 : /*
441 : * Unless code explicitly opted into batching IOs, submit the IO
442 : * immediately.
443 : */
444 1152156 : if (!pgaio_my_backend->in_batchmode)
445 49078 : pgaio_submit_staged();
446 : }
447 : else
448 : {
449 1313488 : pgaio_io_prepare_submit(ioh);
450 1313488 : pgaio_io_perform_synchronously(ioh);
451 : }
452 2465644 : }
453 :
454 : bool
455 2465644 : pgaio_io_needs_synchronous_execution(PgAioHandle *ioh)
456 : {
457 : /*
458 : * If the caller said to execute the IO synchronously, do so.
459 : *
460 : * XXX: We could optimize the logic when to execute synchronously by first
461 : * checking if there are other IOs in flight and only synchronously
462 : * executing if not. Unclear whether that'll be sufficiently common to be
463 : * worth worrying about.
464 : */
465 2465644 : if (ioh->flags & PGAIO_HF_SYNCHRONOUS)
466 1305444 : return true;
467 :
468 : /* Check if the IO method requires synchronous execution of IO */
469 1160200 : if (pgaio_method_ops->needs_synchronous_execution)
470 1160200 : return pgaio_method_ops->needs_synchronous_execution(ioh);
471 :
472 0 : return false;
473 : }
474 :
475 : /*
476 : * Handle IO being processed by IO method.
477 : *
478 : * Should be called by IO methods / synchronous IO execution, just before the
479 : * IO is performed.
480 : */
481 : void
482 2465644 : pgaio_io_prepare_submit(PgAioHandle *ioh)
483 : {
484 2465644 : pgaio_io_update_state(ioh, PGAIO_HS_SUBMITTED);
485 :
486 2465644 : dclist_push_tail(&pgaio_my_backend->in_flight_ios, &ioh->node);
487 2465644 : }
488 :
489 : /*
490 : * Handle IO getting completed by a method.
491 : *
492 : * Should be called by IO methods / synchronous IO execution, just after the
493 : * IO has been performed.
494 : *
495 : * Expects to be called in a critical section. We expect IOs to be usable for
496 : * WAL etc, which requires being able to execute completion callbacks in a
497 : * critical section.
498 : */
499 : void
500 2255382 : pgaio_io_process_completion(PgAioHandle *ioh, int result)
501 : {
502 : Assert(ioh->state == PGAIO_HS_SUBMITTED);
503 :
504 : Assert(CritSectionCount > 0);
505 :
506 2255382 : ioh->result = result;
507 :
508 2255382 : pgaio_io_update_state(ioh, PGAIO_HS_COMPLETED_IO);
509 :
510 2255382 : pgaio_io_call_inj(ioh, "aio-process-completion-before-shared");
511 :
512 2255382 : pgaio_io_call_complete_shared(ioh);
513 :
514 2255382 : pgaio_io_update_state(ioh, PGAIO_HS_COMPLETED_SHARED);
515 :
516 : /* condition variable broadcast ensures state is visible before wakeup */
517 2255382 : ConditionVariableBroadcast(&ioh->cv);
518 :
519 : /* contains call to pgaio_io_call_complete_local() */
520 2255382 : if (ioh->owner_procno == MyProcNumber)
521 1313488 : pgaio_io_reclaim(ioh);
522 2255382 : }
523 :
524 : /*
525 : * Has the IO completed and thus the IO handle been reused?
526 : *
527 : * This is useful when waiting for IO completion at a low level (e.g. in an IO
528 : * method's ->wait_one() callback).
529 : */
530 : bool
531 3593678 : pgaio_io_was_recycled(PgAioHandle *ioh, uint64 ref_generation, PgAioHandleState *state)
532 : {
533 3593678 : *state = ioh->state;
534 3593678 : pg_read_barrier();
535 :
536 3593678 : return ioh->generation != ref_generation;
537 : }
538 :
539 : /*
540 : * Wait for IO to complete. External code should never use this, outside of
541 : * the AIO subsystem waits are only allowed via pgaio_wref_wait().
542 : */
543 : static void
544 490082 : pgaio_io_wait(PgAioHandle *ioh, uint64 ref_generation)
545 : {
546 : PgAioHandleState state;
547 : bool am_owner;
548 :
549 490082 : am_owner = ioh->owner_procno == MyProcNumber;
550 :
551 490082 : if (pgaio_io_was_recycled(ioh, ref_generation, &state))
552 68 : return;
553 :
554 490014 : if (am_owner)
555 : {
556 485598 : if (state != PGAIO_HS_SUBMITTED
557 117596 : && state != PGAIO_HS_COMPLETED_IO
558 492 : && state != PGAIO_HS_COMPLETED_SHARED
559 0 : && state != PGAIO_HS_COMPLETED_LOCAL)
560 : {
561 0 : elog(PANIC, "waiting for own IO in wrong state: %d",
562 : state);
563 : }
564 : }
565 :
566 : while (true)
567 : {
568 979284 : if (pgaio_io_was_recycled(ioh, ref_generation, &state))
569 2468 : return;
570 :
571 976816 : switch (state)
572 : {
573 0 : case PGAIO_HS_IDLE:
574 : case PGAIO_HS_HANDED_OUT:
575 0 : elog(ERROR, "IO in wrong state: %d", state);
576 : break;
577 :
578 370622 : case PGAIO_HS_SUBMITTED:
579 :
580 : /*
581 : * If we need to wait via the IO method, do so now. Don't
582 : * check via the IO method if the issuing backend is executing
583 : * the IO synchronously.
584 : */
585 370622 : if (pgaio_method_ops->wait_one && !(ioh->flags & PGAIO_HF_SYNCHRONOUS))
586 : {
587 0 : pgaio_method_ops->wait_one(ioh, ref_generation);
588 0 : continue;
589 : }
590 : /* fallthrough */
591 :
592 : /* waiting for owner to submit */
593 : case PGAIO_HS_DEFINED:
594 : case PGAIO_HS_STAGED:
595 : /* waiting for reaper to complete */
596 : /* fallthrough */
597 : case PGAIO_HS_COMPLETED_IO:
598 : /* shouldn't be able to hit this otherwise */
599 : Assert(IsUnderPostmaster);
600 : /* ensure we're going to get woken up */
601 489270 : ConditionVariablePrepareToSleep(&ioh->cv);
602 :
603 977322 : while (!pgaio_io_was_recycled(ioh, ref_generation, &state))
604 : {
605 974862 : if (state == PGAIO_HS_COMPLETED_SHARED ||
606 488074 : state == PGAIO_HS_COMPLETED_LOCAL)
607 : break;
608 488052 : ConditionVariableSleep(&ioh->cv, WAIT_EVENT_AIO_IO_COMPLETION);
609 : }
610 :
611 489270 : ConditionVariableCancelSleep();
612 489270 : break;
613 :
614 487546 : case PGAIO_HS_COMPLETED_SHARED:
615 : case PGAIO_HS_COMPLETED_LOCAL:
616 : /* see above */
617 487546 : if (am_owner)
618 485598 : pgaio_io_reclaim(ioh);
619 487546 : return;
620 : }
621 489270 : }
622 : }
623 :
624 : /*
625 : * Make IO handle ready to be reused after IO has completed or after the
626 : * handle has been released without being used.
627 : */
628 : static void
629 2470722 : pgaio_io_reclaim(PgAioHandle *ioh)
630 : {
631 : /* This is only ok if it's our IO */
632 : Assert(ioh->owner_procno == MyProcNumber);
633 : Assert(ioh->state != PGAIO_HS_IDLE);
634 :
635 : /*
636 : * It's a bit ugly, but right now the easiest place to put the execution
637 : * of local completion callbacks is this function, as we need to execute
638 : * local callbacks just before reclaiming at multiple callsites.
639 : */
640 2470722 : if (ioh->state == PGAIO_HS_COMPLETED_SHARED)
641 : {
642 : PgAioResult local_result;
643 :
644 2465644 : local_result = pgaio_io_call_complete_local(ioh);
645 2465644 : pgaio_io_update_state(ioh, PGAIO_HS_COMPLETED_LOCAL);
646 :
647 2465644 : if (ioh->report_return)
648 : {
649 2465624 : ioh->report_return->result = local_result;
650 2465624 : ioh->report_return->target_data = ioh->target_data;
651 : }
652 : }
653 :
654 2470722 : pgaio_debug_io(DEBUG4, ioh,
655 : "reclaiming: distilled_result: (status %s, id %u, error_data %d), raw_result: %d",
656 : pgaio_result_status_string(ioh->distilled_result.status),
657 : ioh->distilled_result.id,
658 : ioh->distilled_result.error_data,
659 : ioh->result);
660 :
661 : /* if the IO has been defined, it's on the in-flight list, remove */
662 2470722 : if (ioh->state != PGAIO_HS_HANDED_OUT)
663 2465644 : dclist_delete_from(&pgaio_my_backend->in_flight_ios, &ioh->node);
664 :
665 2470722 : if (ioh->resowner)
666 : {
667 2470636 : ResourceOwnerForgetAioHandle(ioh->resowner, &ioh->resowner_node);
668 2470636 : ioh->resowner = NULL;
669 : }
670 :
671 : Assert(!ioh->resowner);
672 :
673 2470722 : ioh->op = PGAIO_OP_INVALID;
674 2470722 : ioh->target = PGAIO_TID_INVALID;
675 2470722 : ioh->flags = 0;
676 2470722 : ioh->num_callbacks = 0;
677 2470722 : ioh->handle_data_len = 0;
678 2470722 : ioh->report_return = NULL;
679 2470722 : ioh->result = 0;
680 2470722 : ioh->distilled_result.status = PGAIO_RS_UNKNOWN;
681 :
682 : /* XXX: the barrier is probably superfluous */
683 2470722 : pg_write_barrier();
684 2470722 : ioh->generation++;
685 :
686 2470722 : pgaio_io_update_state(ioh, PGAIO_HS_IDLE);
687 :
688 : /*
689 : * We push the IO to the head of the idle IO list, that seems more cache
690 : * efficient in cases where only a few IOs are used.
691 : */
692 2470722 : dclist_push_head(&pgaio_my_backend->idle_ios, &ioh->node);
693 2470722 : }
694 :
695 : /*
696 : * Wait for an IO handle to become usable.
697 : *
698 : * This only really is useful for pgaio_io_acquire().
699 : */
700 : static void
701 5116 : pgaio_io_wait_for_free(void)
702 : {
703 5116 : int reclaimed = 0;
704 :
705 5116 : pgaio_debug(DEBUG2, "waiting for self with %d pending",
706 : pgaio_my_backend->num_staged_ios);
707 :
708 : /*
709 : * First check if any of our IOs actually have completed - when using
710 : * worker, that'll often be the case. We could do so as part of the loop
711 : * below, but that'd potentially lead us to wait for some IO submitted
712 : * before.
713 : */
714 10232 : for (int i = 0; i < io_max_concurrency; i++)
715 : {
716 5116 : PgAioHandle *ioh = &pgaio_ctl->io_handles[pgaio_my_backend->io_handle_off + i];
717 :
718 5116 : if (ioh->state == PGAIO_HS_COMPLETED_SHARED)
719 : {
720 4332 : pgaio_io_reclaim(ioh);
721 4332 : reclaimed++;
722 : }
723 : }
724 :
725 5116 : if (reclaimed > 0)
726 4332 : return;
727 :
728 : /*
729 : * If we have any unsubmitted IOs, submit them now. We'll start waiting in
730 : * a second, so it's better they're in flight. This also addresses the
731 : * edge-case that all IOs are unsubmitted.
732 : */
733 784 : if (pgaio_my_backend->num_staged_ios > 0)
734 0 : pgaio_submit_staged();
735 :
736 784 : if (dclist_count(&pgaio_my_backend->in_flight_ios) == 0)
737 0 : elog(ERROR, "no free IOs despite no in-flight IOs");
738 :
739 : /*
740 : * Wait for the oldest in-flight IO to complete.
741 : *
742 : * XXX: Reusing the general IO wait is suboptimal, we don't need to wait
743 : * for that specific IO to complete, we just need *any* IO to complete.
744 : */
745 : {
746 784 : PgAioHandle *ioh = dclist_head_element(PgAioHandle, node,
747 : &pgaio_my_backend->in_flight_ios);
748 :
749 784 : switch (ioh->state)
750 : {
751 : /* should not be in in-flight list */
752 0 : case PGAIO_HS_IDLE:
753 : case PGAIO_HS_DEFINED:
754 : case PGAIO_HS_HANDED_OUT:
755 : case PGAIO_HS_STAGED:
756 : case PGAIO_HS_COMPLETED_LOCAL:
757 0 : elog(ERROR, "shouldn't get here with io:%d in state %d",
758 : pgaio_io_get_id(ioh), ioh->state);
759 : break;
760 :
761 778 : case PGAIO_HS_COMPLETED_IO:
762 : case PGAIO_HS_SUBMITTED:
763 778 : pgaio_debug_io(DEBUG2, ioh,
764 : "waiting for free io with %d in flight",
765 : dclist_count(&pgaio_my_backend->in_flight_ios));
766 :
767 : /*
768 : * In a more general case this would be racy, because the
769 : * generation could increase after we read ioh->state above.
770 : * But we are only looking at IOs by the current backend and
771 : * the IO can only be recycled by this backend.
772 : */
773 778 : pgaio_io_wait(ioh, ioh->generation);
774 778 : break;
775 :
776 6 : case PGAIO_HS_COMPLETED_SHARED:
777 : /* it's possible that another backend just finished this IO */
778 6 : pgaio_io_reclaim(ioh);
779 6 : break;
780 : }
781 :
782 784 : if (dclist_count(&pgaio_my_backend->idle_ios) == 0)
783 0 : elog(PANIC, "no idle IO after waiting for IO to terminate");
784 784 : return;
785 : }
786 : }
787 :
788 : /*
789 : * Internal - code outside of AIO should never need this and it'd be hard for
790 : * such code to be safe.
791 : */
792 : static PgAioHandle *
793 1636276 : pgaio_io_from_wref(PgAioWaitRef *iow, uint64 *ref_generation)
794 : {
795 : PgAioHandle *ioh;
796 :
797 : Assert(iow->aio_index < pgaio_ctl->io_handle_count);
798 :
799 1636276 : ioh = &pgaio_ctl->io_handles[iow->aio_index];
800 :
801 1636276 : *ref_generation = ((uint64) iow->generation_upper) << 32 |
802 1636276 : iow->generation_lower;
803 :
804 : Assert(*ref_generation != 0);
805 :
806 1636276 : return ioh;
807 : }
808 :
809 : static const char *
810 14270 : pgaio_io_state_get_name(PgAioHandleState s)
811 : {
812 : #define PGAIO_HS_TOSTR_CASE(sym) case PGAIO_HS_##sym: return #sym
813 14270 : switch (s)
814 : {
815 0 : PGAIO_HS_TOSTR_CASE(IDLE);
816 4756 : PGAIO_HS_TOSTR_CASE(HANDED_OUT);
817 2378 : PGAIO_HS_TOSTR_CASE(DEFINED);
818 2378 : PGAIO_HS_TOSTR_CASE(STAGED);
819 0 : PGAIO_HS_TOSTR_CASE(SUBMITTED);
820 2380 : PGAIO_HS_TOSTR_CASE(COMPLETED_IO);
821 2378 : PGAIO_HS_TOSTR_CASE(COMPLETED_SHARED);
822 0 : PGAIO_HS_TOSTR_CASE(COMPLETED_LOCAL);
823 : }
824 : #undef PGAIO_HS_TOSTR_CASE
825 :
826 0 : return NULL; /* silence compiler */
827 : }
828 :
829 : const char *
830 14270 : pgaio_io_get_state_name(PgAioHandle *ioh)
831 : {
832 14270 : return pgaio_io_state_get_name(ioh->state);
833 : }
834 :
835 : const char *
836 4756 : pgaio_result_status_string(PgAioResultStatus rs)
837 : {
838 4756 : switch (rs)
839 : {
840 0 : case PGAIO_RS_UNKNOWN:
841 0 : return "UNKNOWN";
842 4396 : case PGAIO_RS_OK:
843 4396 : return "OK";
844 136 : case PGAIO_RS_WARNING:
845 136 : return "WARNING";
846 40 : case PGAIO_RS_PARTIAL:
847 40 : return "PARTIAL";
848 184 : case PGAIO_RS_ERROR:
849 184 : return "ERROR";
850 : }
851 :
852 0 : return NULL; /* silence compiler */
853 : }
854 :
855 :
856 :
857 : /* --------------------------------------------------------------------------------
858 : * Functions primarily related to IO Wait References
859 : * --------------------------------------------------------------------------------
860 : */
861 :
862 : /*
863 : * Mark a wait reference as invalid
864 : */
865 : void
866 25718598 : pgaio_wref_clear(PgAioWaitRef *iow)
867 : {
868 25718598 : iow->aio_index = PG_UINT32_MAX;
869 25718598 : }
870 :
871 : /* Is the wait reference valid? */
872 : bool
873 5036166 : pgaio_wref_valid(PgAioWaitRef *iow)
874 : {
875 5036166 : return iow->aio_index != PG_UINT32_MAX;
876 : }
877 :
878 : /*
879 : * Similar to pgaio_io_get_id(), just for wait references.
880 : */
881 : int
882 0 : pgaio_wref_get_id(PgAioWaitRef *iow)
883 : {
884 : Assert(pgaio_wref_valid(iow));
885 0 : return iow->aio_index;
886 : }
887 :
888 : /*
889 : * Wait for the IO to have completed. Can be called in any process, not just
890 : * in the issuing backend.
891 : */
892 : void
893 489286 : pgaio_wref_wait(PgAioWaitRef *iow)
894 : {
895 : uint64 ref_generation;
896 : PgAioHandle *ioh;
897 :
898 489286 : ioh = pgaio_io_from_wref(iow, &ref_generation);
899 :
900 489286 : pgaio_io_wait(ioh, ref_generation);
901 489286 : }
902 :
903 : /*
904 : * Check if the referenced IO completed, without blocking.
905 : */
906 : bool
907 1146990 : pgaio_wref_check_done(PgAioWaitRef *iow)
908 : {
909 : uint64 ref_generation;
910 : PgAioHandleState state;
911 : bool am_owner;
912 : PgAioHandle *ioh;
913 :
914 1146990 : ioh = pgaio_io_from_wref(iow, &ref_generation);
915 :
916 1146990 : if (pgaio_io_was_recycled(ioh, ref_generation, &state))
917 0 : return true;
918 :
919 1146990 : if (state == PGAIO_HS_IDLE)
920 0 : return true;
921 :
922 1146990 : am_owner = ioh->owner_procno == MyProcNumber;
923 :
924 1146990 : if (state == PGAIO_HS_COMPLETED_SHARED ||
925 484770 : state == PGAIO_HS_COMPLETED_LOCAL)
926 : {
927 662220 : if (am_owner)
928 662220 : pgaio_io_reclaim(ioh);
929 662220 : return true;
930 : }
931 :
932 : /*
933 : * XXX: It likely would be worth checking in with the io method, to give
934 : * the IO method a chance to check if there are completion events queued.
935 : */
936 :
937 484770 : return false;
938 : }
939 :
940 :
941 :
942 : /* --------------------------------------------------------------------------------
943 : * Actions on multiple IOs.
944 : * --------------------------------------------------------------------------------
945 : */
946 :
947 : /*
948 : * Submit IOs in batches going forward.
949 : *
950 : * Submitting multiple IOs at once can be substantially faster than doing so
951 : * one-by-one. At the same time, submitting multiple IOs at once requires more
952 : * care to avoid deadlocks.
953 : *
954 : * Consider backend A staging an IO for buffer 1 and then trying to start IO
955 : * on buffer 2, while backend B does the inverse. If A submitted the IO before
956 : * moving on to buffer 2, this works just fine, B will wait for the IO to
957 : * complete. But if batching were used, each backend will wait for IO that has
958 : * not yet been submitted to complete, i.e. forever.
959 : *
960 : * End batch submission mode with pgaio_exit_batchmode(). (Throwing errors is
961 : * allowed; error recovery will end the batch.)
962 : *
963 : * To avoid deadlocks, code needs to ensure that it will not wait for another
964 : * backend while there is unsubmitted IO. E.g. by using conditional lock
965 : * acquisition when acquiring buffer locks. To check if there currently are
966 : * staged IOs, call pgaio_have_staged() and to submit all staged IOs call
967 : * pgaio_submit_staged().
968 : *
969 : * It is not allowed to enter batchmode while already in batchmode, it's
970 : * unlikely to ever be needed, as code needs to be explicitly aware of being
971 : * called in batchmode, to avoid the deadlock risks explained above.
972 : *
973 : * Note that IOs may get submitted before pgaio_exit_batchmode() is called,
974 : * e.g. because too many IOs have been staged or because pgaio_submit_staged()
975 : * was called.
976 : */
977 : void
978 5413964 : pgaio_enter_batchmode(void)
979 : {
980 5413964 : if (pgaio_my_backend->in_batchmode)
981 0 : elog(ERROR, "starting batch while batch already in progress");
982 5413964 : pgaio_my_backend->in_batchmode = true;
983 5413964 : }
984 :
985 : /*
986 : * Stop submitting IOs in batches.
987 : */
988 : void
989 5413944 : pgaio_exit_batchmode(void)
990 : {
991 : Assert(pgaio_my_backend->in_batchmode);
992 :
993 5413944 : pgaio_submit_staged();
994 5413944 : pgaio_my_backend->in_batchmode = false;
995 5413944 : }
996 :
997 : /*
998 : * Are there staged but unsubmitted IOs?
999 : *
1000 : * See comment above pgaio_enter_batchmode() for why code may need to check if
1001 : * there is IO in that state.
1002 : */
1003 : bool
1004 2470526 : pgaio_have_staged(void)
1005 : {
1006 : Assert(pgaio_my_backend->in_batchmode ||
1007 : pgaio_my_backend->num_staged_ios == 0);
1008 2470526 : return pgaio_my_backend->num_staged_ios > 0;
1009 : }
1010 :
1011 : /*
1012 : * Submit all staged but not yet submitted IOs.
1013 : *
1014 : * Unless in batch mode, this never needs to be called, as IOs get submitted
1015 : * as soon as possible. While in batchmode pgaio_submit_staged() can be called
1016 : * before waiting on another backend, to avoid the risk of deadlocks. See
1017 : * pgaio_enter_batchmode().
1018 : */
1019 : void
1020 21967064 : pgaio_submit_staged(void)
1021 : {
1022 21967064 : int total_submitted = 0;
1023 : int did_submit;
1024 :
1025 21967064 : if (pgaio_my_backend->num_staged_ios == 0)
1026 20816054 : return;
1027 :
1028 :
1029 1151010 : START_CRIT_SECTION();
1030 :
1031 1151010 : did_submit = pgaio_method_ops->submit(pgaio_my_backend->num_staged_ios,
1032 1151010 : pgaio_my_backend->staged_ios);
1033 :
1034 1151010 : END_CRIT_SECTION();
1035 :
1036 1151010 : total_submitted += did_submit;
1037 :
1038 : Assert(total_submitted == did_submit);
1039 :
1040 1151010 : pgaio_my_backend->num_staged_ios = 0;
1041 :
1042 1151010 : pgaio_debug(DEBUG4,
1043 : "aio: submitted %d IOs",
1044 : total_submitted);
1045 : }
1046 :
1047 :
1048 :
1049 : /* --------------------------------------------------------------------------------
1050 : * Other
1051 : * --------------------------------------------------------------------------------
1052 : */
1053 :
1054 :
1055 : /*
1056 : * Perform AIO related cleanup after an error.
1057 : *
1058 : * This should be called early in the error recovery paths, as later steps may
1059 : * need to issue AIO (e.g. to record a transaction abort WAL record).
1060 : */
1061 : void
1062 58292 : pgaio_error_cleanup(void)
1063 : {
1064 : /*
1065 : * It is possible that code errored out after pgaio_enter_batchmode() but
1066 : * before pgaio_exit_batchmode() was called. In that case we need to
1067 : * submit the IO now.
1068 : */
1069 58292 : if (pgaio_my_backend->in_batchmode)
1070 : {
1071 20 : pgaio_my_backend->in_batchmode = false;
1072 :
1073 20 : pgaio_submit_staged();
1074 : }
1075 :
1076 : /*
1077 : * As we aren't in batchmode, there shouldn't be any unsubmitted IOs.
1078 : */
1079 : Assert(pgaio_my_backend->num_staged_ios == 0);
1080 58292 : }
1081 :
1082 : /*
1083 : * Perform AIO related checks at (sub-)transactional boundaries.
1084 : *
1085 : * This should be called late during (sub-)transactional commit/abort, after
1086 : * all steps that might need to perform AIO, so that we can verify that the
1087 : * AIO subsystem is in a valid state at the end of a transaction.
1088 : */
1089 : void
1090 915656 : AtEOXact_Aio(bool is_commit)
1091 : {
1092 : /*
1093 : * We should never be in batch mode at transactional boundaries. In case
1094 : * an error was thrown while in batch mode, pgaio_error_cleanup() should
1095 : * have exited batchmode.
1096 : *
1097 : * In case we are in batchmode somehow, make sure to submit all staged
1098 : * IOs, other backends may need them to complete to continue.
1099 : */
1100 915656 : if (pgaio_my_backend->in_batchmode)
1101 : {
1102 8 : pgaio_error_cleanup();
1103 8 : elog(WARNING, "open AIO batch at end of (sub-)transaction");
1104 : }
1105 :
1106 : /*
1107 : * As we aren't in batchmode, there shouldn't be any unsubmitted IOs.
1108 : */
1109 : Assert(pgaio_my_backend->num_staged_ios == 0);
1110 915656 : }
1111 :
1112 : /*
1113 : * Need to submit staged but not yet submitted IOs using the fd, otherwise
1114 : * the IO would end up targeting something bogus.
1115 : */
1116 : void
1117 16512764 : pgaio_closing_fd(int fd)
1118 : {
1119 : /*
1120 : * Might be called before AIO is initialized or in a subprocess that
1121 : * doesn't use AIO.
1122 : */
1123 16512764 : if (!pgaio_my_backend)
1124 13858 : return;
1125 :
1126 : /*
1127 : * For now just submit all staged IOs - we could be more selective, but
1128 : * it's probably not worth it.
1129 : */
1130 16498906 : pgaio_submit_staged();
1131 :
1132 : /*
1133 : * If requested by the IO method, wait for all IOs that use the
1134 : * to-be-closed FD.
1135 : */
1136 16498906 : if (pgaio_method_ops->wait_on_fd_before_close)
1137 : {
1138 : /*
1139 : * As waiting for one IO to complete may complete multiple IOs, we
1140 : * can't just use a mutable list iterator. The maximum number of
1141 : * in-flight IOs is fairly small, so just restart the loop after
1142 : * waiting for an IO.
1143 : */
1144 0 : while (!dclist_is_empty(&pgaio_my_backend->in_flight_ios))
1145 : {
1146 : dlist_iter iter;
1147 0 : PgAioHandle *ioh = NULL;
1148 :
1149 0 : dclist_foreach(iter, &pgaio_my_backend->in_flight_ios)
1150 : {
1151 0 : ioh = dclist_container(PgAioHandle, node, iter.cur);
1152 :
1153 0 : if (pgaio_io_uses_fd(ioh, fd))
1154 0 : break;
1155 : else
1156 0 : ioh = NULL;
1157 : }
1158 :
1159 0 : if (!ioh)
1160 0 : break;
1161 :
1162 : /* see comment in pgaio_io_wait_for_free() about raciness */
1163 0 : pgaio_io_wait(ioh, ioh->generation);
1164 : }
1165 : }
1166 : }
1167 :
1168 : /*
1169 : * Registered as before_shmem_exit() callback in pgaio_init_backend()
1170 : */
1171 : void
1172 40082 : pgaio_shutdown(int code, Datum arg)
1173 : {
1174 : Assert(pgaio_my_backend);
1175 : Assert(!pgaio_my_backend->handed_out_io);
1176 :
1177 : /* first clean up resources as we would at a transaction boundary */
1178 40082 : AtEOXact_Aio(code == 0);
1179 :
1180 : /*
1181 : * Before exiting, make sure that all IOs are finished. That has two main
1182 : * purposes:
1183 : *
1184 : * - Some kernel-level AIO mechanisms don't deal well with the issuer of
1185 : * an AIO exiting before IO completed
1186 : *
1187 : * - It'd be confusing to see partially finished IOs in stats views etc
1188 : */
1189 40100 : while (!dclist_is_empty(&pgaio_my_backend->in_flight_ios))
1190 : {
1191 18 : PgAioHandle *ioh = dclist_head_element(PgAioHandle, node, &pgaio_my_backend->in_flight_ios);
1192 :
1193 : /* see comment in pgaio_io_wait_for_free() about raciness */
1194 18 : pgaio_io_wait(ioh, ioh->generation);
1195 : }
1196 :
1197 40082 : pgaio_my_backend = NULL;
1198 40082 : }
1199 :
1200 : void
1201 2190 : assign_io_method(int newval, void *extra)
1202 : {
1203 : Assert(pgaio_method_ops_table[newval] != NULL);
1204 : Assert(newval < lengthof(io_method_options));
1205 :
1206 2190 : pgaio_method_ops = pgaio_method_ops_table[newval];
1207 2190 : }
1208 :
1209 : bool
1210 4264 : check_io_max_concurrency(int *newval, void **extra, GucSource source)
1211 : {
1212 4264 : if (*newval == -1)
1213 : {
1214 : /*
1215 : * Auto-tuning will be applied later during startup, as auto-tuning
1216 : * depends on the value of various GUCs.
1217 : */
1218 2168 : return true;
1219 : }
1220 2096 : else if (*newval == 0)
1221 : {
1222 0 : GUC_check_errdetail("Only -1 or values bigger than 0 are valid.");
1223 0 : return false;
1224 : }
1225 :
1226 2096 : return true;
1227 : }
1228 :
1229 :
1230 :
1231 : /* --------------------------------------------------------------------------------
1232 : * Injection point support
1233 : * --------------------------------------------------------------------------------
1234 : */
1235 :
1236 : #ifdef USE_INJECTION_POINTS
1237 :
1238 : /*
1239 : * Call injection point with support for pgaio_inj_io_get().
1240 : */
1241 : void
1242 3197276 : pgaio_io_call_inj(PgAioHandle *ioh, const char *injection_point)
1243 : {
1244 3197276 : pgaio_inj_cur_handle = ioh;
1245 :
1246 3197276 : PG_TRY();
1247 : {
1248 3197276 : InjectionPointCached(injection_point);
1249 : }
1250 2 : PG_FINALLY();
1251 : {
1252 3197276 : pgaio_inj_cur_handle = NULL;
1253 : }
1254 3197276 : PG_END_TRY();
1255 3197274 : }
1256 :
1257 : /*
1258 : * Return IO associated with injection point invocation. This is only needed
1259 : * as injection points currently don't support arguments.
1260 : */
1261 : PgAioHandle *
1262 96 : pgaio_inj_io_get(void)
1263 : {
1264 96 : return pgaio_inj_cur_handle;
1265 : }
1266 :
1267 : #endif
|
