2 * Data plane event loop
4 * Copyright (c) 2003-2008 Fabrice Bellard
5 * Copyright (c) 2009-2017 QEMU contributors
7 * Permission is hereby granted, free of charge, to any person obtaining a copy
8 * of this software and associated documentation files (the "Software"), to deal
9 * in the Software without restriction, including without limitation the rights
10 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11 * copies of the Software, and to permit persons to whom the Software is
12 * furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice shall be included in
15 * all copies or substantial portions of the Software.
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
26 #include "qemu/osdep.h"
27 #include "qapi/error.h"
28 #include "block/aio.h"
29 #include "block/thread-pool.h"
30 #include "qemu/main-loop.h"
31 #include "qemu/atomic.h"
32 #include "qemu/rcu_queue.h"
33 #include "block/raw-aio.h"
34 #include "qemu/coroutine_int.h"
35 #include "qemu/coroutine-tls.h"
38 /***********************************************************/
39 /* bottom halves (can be seen as timers which expire ASAP) */
41 /* QEMUBH::flags values */
43 /* Already enqueued and waiting for aio_bh_poll() */
44 BH_PENDING
= (1 << 0),
46 /* Invoke the callback */
47 BH_SCHEDULED
= (1 << 1),
49 /* Delete without invoking callback */
50 BH_DELETED
= (1 << 2),
52 /* Delete after invoking callback */
53 BH_ONESHOT
= (1 << 3),
55 /* Schedule periodically when the event loop is idle */
64 QSLIST_ENTRY(QEMUBH
) next
;
68 /* Called concurrently from any thread */
69 static void aio_bh_enqueue(QEMUBH
*bh
, unsigned new_flags
)
71 AioContext
*ctx
= bh
->ctx
;
75 * The memory barrier implicit in qatomic_fetch_or makes sure that:
76 * 1. idle & any writes needed by the callback are done before the
77 * locations are read in the aio_bh_poll.
78 * 2. ctx is loaded before the callback has a chance to execute and bh
81 old_flags
= qatomic_fetch_or(&bh
->flags
, BH_PENDING
| new_flags
);
82 if (!(old_flags
& BH_PENDING
)) {
83 QSLIST_INSERT_HEAD_ATOMIC(&ctx
->bh_list
, bh
, next
);
89 /* Only called from aio_bh_poll() and aio_ctx_finalize() */
90 static QEMUBH
*aio_bh_dequeue(BHList
*head
, unsigned *flags
)
92 QEMUBH
*bh
= QSLIST_FIRST_RCU(head
);
98 QSLIST_REMOVE_HEAD(head
, next
);
101 * The qatomic_and is paired with aio_bh_enqueue(). The implicit memory
102 * barrier ensures that the callback sees all writes done by the scheduling
103 * thread. It also ensures that the scheduling thread sees the cleared
104 * flag before bh->cb has run, and thus will call aio_notify again if
107 *flags
= qatomic_fetch_and(&bh
->flags
,
108 ~(BH_PENDING
| BH_SCHEDULED
| BH_IDLE
));
112 void aio_bh_schedule_oneshot_full(AioContext
*ctx
, QEMUBHFunc
*cb
,
113 void *opaque
, const char *name
)
116 bh
= g_new(QEMUBH
, 1);
123 aio_bh_enqueue(bh
, BH_SCHEDULED
| BH_ONESHOT
);
126 QEMUBH
*aio_bh_new_full(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
,
130 bh
= g_new(QEMUBH
, 1);
140 void aio_bh_call(QEMUBH
*bh
)
145 /* Multiple occurrences of aio_bh_poll cannot be called concurrently. */
146 int aio_bh_poll(AioContext
*ctx
)
152 QSLIST_MOVE_ATOMIC(&slice
.bh_list
, &ctx
->bh_list
);
153 QSIMPLEQ_INSERT_TAIL(&ctx
->bh_slice_list
, &slice
, next
);
155 while ((s
= QSIMPLEQ_FIRST(&ctx
->bh_slice_list
))) {
159 bh
= aio_bh_dequeue(&s
->bh_list
, &flags
);
161 QSIMPLEQ_REMOVE_HEAD(&ctx
->bh_slice_list
, next
);
165 if ((flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
166 /* Idle BHs don't count as progress */
167 if (!(flags
& BH_IDLE
)) {
172 if (flags
& (BH_DELETED
| BH_ONESHOT
)) {
180 void qemu_bh_schedule_idle(QEMUBH
*bh
)
182 aio_bh_enqueue(bh
, BH_SCHEDULED
| BH_IDLE
);
185 void qemu_bh_schedule(QEMUBH
*bh
)
187 aio_bh_enqueue(bh
, BH_SCHEDULED
);
190 /* This func is async.
192 void qemu_bh_cancel(QEMUBH
*bh
)
194 qatomic_and(&bh
->flags
, ~BH_SCHEDULED
);
197 /* This func is async.The bottom half will do the delete action at the finial
200 void qemu_bh_delete(QEMUBH
*bh
)
202 aio_bh_enqueue(bh
, BH_DELETED
);
205 static int64_t aio_compute_bh_timeout(BHList
*head
, int timeout
)
209 QSLIST_FOREACH_RCU(bh
, head
, next
) {
210 if ((bh
->flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
211 if (bh
->flags
& BH_IDLE
) {
212 /* idle bottom halves will be polled at least
216 /* non-idle bottom halves will be executed
227 aio_compute_timeout(AioContext
*ctx
)
233 timeout
= aio_compute_bh_timeout(&ctx
->bh_list
, timeout
);
238 QSIMPLEQ_FOREACH(s
, &ctx
->bh_slice_list
, next
) {
239 timeout
= aio_compute_bh_timeout(&s
->bh_list
, timeout
);
245 deadline
= timerlistgroup_deadline_ns(&ctx
->tlg
);
249 return qemu_soonest_timeout(timeout
, deadline
);
254 aio_ctx_prepare(GSource
*source
, gint
*timeout
)
256 AioContext
*ctx
= (AioContext
*) source
;
258 qatomic_set(&ctx
->notify_me
, qatomic_read(&ctx
->notify_me
) | 1);
261 * Write ctx->notify_me before computing the timeout
262 * (reading bottom half flags, etc.). Pairs with
263 * smp_mb in aio_notify().
267 /* We assume there is no timeout already supplied */
268 *timeout
= qemu_timeout_ns_to_ms(aio_compute_timeout(ctx
));
270 if (aio_prepare(ctx
)) {
274 return *timeout
== 0;
278 aio_ctx_check(GSource
*source
)
280 AioContext
*ctx
= (AioContext
*) source
;
284 /* Finish computing the timeout before clearing the flag. */
285 qatomic_store_release(&ctx
->notify_me
, qatomic_read(&ctx
->notify_me
) & ~1);
286 aio_notify_accept(ctx
);
288 QSLIST_FOREACH_RCU(bh
, &ctx
->bh_list
, next
) {
289 if ((bh
->flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
294 QSIMPLEQ_FOREACH(s
, &ctx
->bh_slice_list
, next
) {
295 QSLIST_FOREACH_RCU(bh
, &s
->bh_list
, next
) {
296 if ((bh
->flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
301 return aio_pending(ctx
) || (timerlistgroup_deadline_ns(&ctx
->tlg
) == 0);
305 aio_ctx_dispatch(GSource
*source
,
306 GSourceFunc callback
,
309 AioContext
*ctx
= (AioContext
*) source
;
311 assert(callback
== NULL
);
317 aio_ctx_finalize(GSource
*source
)
319 AioContext
*ctx
= (AioContext
*) source
;
323 thread_pool_free(ctx
->thread_pool
);
325 #ifdef CONFIG_LINUX_AIO
326 if (ctx
->linux_aio
) {
327 laio_detach_aio_context(ctx
->linux_aio
, ctx
);
328 laio_cleanup(ctx
->linux_aio
);
329 ctx
->linux_aio
= NULL
;
333 #ifdef CONFIG_LINUX_IO_URING
334 if (ctx
->linux_io_uring
) {
335 luring_detach_aio_context(ctx
->linux_io_uring
, ctx
);
336 luring_cleanup(ctx
->linux_io_uring
);
337 ctx
->linux_io_uring
= NULL
;
341 assert(QSLIST_EMPTY(&ctx
->scheduled_coroutines
));
342 qemu_bh_delete(ctx
->co_schedule_bh
);
344 /* There must be no aio_bh_poll() calls going on */
345 assert(QSIMPLEQ_EMPTY(&ctx
->bh_slice_list
));
347 while ((bh
= aio_bh_dequeue(&ctx
->bh_list
, &flags
))) {
349 * qemu_bh_delete() must have been called on BHs in this AioContext. In
350 * many cases memory leaks, hangs, or inconsistent state occur when a
351 * BH is leaked because something still expects it to run.
353 * If you hit this, fix the lifecycle of the BH so that
354 * qemu_bh_delete() and any associated cleanup is called before the
355 * AioContext is finalized.
357 if (unlikely(!(flags
& BH_DELETED
))) {
358 fprintf(stderr
, "%s: BH '%s' leaked, aborting...\n",
366 aio_set_event_notifier(ctx
, &ctx
->notifier
, false, NULL
, NULL
, NULL
);
367 event_notifier_cleanup(&ctx
->notifier
);
368 qemu_rec_mutex_destroy(&ctx
->lock
);
369 qemu_lockcnt_destroy(&ctx
->list_lock
);
370 timerlistgroup_deinit(&ctx
->tlg
);
371 aio_context_destroy(ctx
);
374 static GSourceFuncs aio_source_funcs
= {
381 GSource
*aio_get_g_source(AioContext
*ctx
)
383 aio_context_use_g_source(ctx
);
384 g_source_ref(&ctx
->source
);
388 ThreadPool
*aio_get_thread_pool(AioContext
*ctx
)
390 if (!ctx
->thread_pool
) {
391 ctx
->thread_pool
= thread_pool_new(ctx
);
393 return ctx
->thread_pool
;
396 #ifdef CONFIG_LINUX_AIO
397 LinuxAioState
*aio_setup_linux_aio(AioContext
*ctx
, Error
**errp
)
399 if (!ctx
->linux_aio
) {
400 ctx
->linux_aio
= laio_init(errp
);
401 if (ctx
->linux_aio
) {
402 laio_attach_aio_context(ctx
->linux_aio
, ctx
);
405 return ctx
->linux_aio
;
408 LinuxAioState
*aio_get_linux_aio(AioContext
*ctx
)
410 assert(ctx
->linux_aio
);
411 return ctx
->linux_aio
;
415 #ifdef CONFIG_LINUX_IO_URING
416 LuringState
*aio_setup_linux_io_uring(AioContext
*ctx
, Error
**errp
)
418 if (ctx
->linux_io_uring
) {
419 return ctx
->linux_io_uring
;
422 ctx
->linux_io_uring
= luring_init(errp
);
423 if (!ctx
->linux_io_uring
) {
427 luring_attach_aio_context(ctx
->linux_io_uring
, ctx
);
428 return ctx
->linux_io_uring
;
431 LuringState
*aio_get_linux_io_uring(AioContext
*ctx
)
433 assert(ctx
->linux_io_uring
);
434 return ctx
->linux_io_uring
;
438 void aio_notify(AioContext
*ctx
)
441 * Write e.g. bh->flags before writing ctx->notified. Pairs with smp_mb in
445 qatomic_set(&ctx
->notified
, true);
448 * Write ctx->notified before reading ctx->notify_me. Pairs
449 * with smp_mb in aio_ctx_prepare or aio_poll.
452 if (qatomic_read(&ctx
->notify_me
)) {
453 event_notifier_set(&ctx
->notifier
);
457 void aio_notify_accept(AioContext
*ctx
)
459 qatomic_set(&ctx
->notified
, false);
462 * Write ctx->notified before reading e.g. bh->flags. Pairs with smp_wmb
468 static void aio_timerlist_notify(void *opaque
, QEMUClockType type
)
473 static void aio_context_notifier_cb(EventNotifier
*e
)
475 AioContext
*ctx
= container_of(e
, AioContext
, notifier
);
477 event_notifier_test_and_clear(&ctx
->notifier
);
480 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
481 static bool aio_context_notifier_poll(void *opaque
)
483 EventNotifier
*e
= opaque
;
484 AioContext
*ctx
= container_of(e
, AioContext
, notifier
);
486 return qatomic_read(&ctx
->notified
);
489 static void aio_context_notifier_poll_ready(EventNotifier
*e
)
491 /* Do nothing, we just wanted to kick the event loop */
494 static void co_schedule_bh_cb(void *opaque
)
496 AioContext
*ctx
= opaque
;
497 QSLIST_HEAD(, Coroutine
) straight
, reversed
;
499 QSLIST_MOVE_ATOMIC(&reversed
, &ctx
->scheduled_coroutines
);
500 QSLIST_INIT(&straight
);
502 while (!QSLIST_EMPTY(&reversed
)) {
503 Coroutine
*co
= QSLIST_FIRST(&reversed
);
504 QSLIST_REMOVE_HEAD(&reversed
, co_scheduled_next
);
505 QSLIST_INSERT_HEAD(&straight
, co
, co_scheduled_next
);
508 while (!QSLIST_EMPTY(&straight
)) {
509 Coroutine
*co
= QSLIST_FIRST(&straight
);
510 QSLIST_REMOVE_HEAD(&straight
, co_scheduled_next
);
511 trace_aio_co_schedule_bh_cb(ctx
, co
);
512 aio_context_acquire(ctx
);
514 /* Protected by write barrier in qemu_aio_coroutine_enter */
515 qatomic_set(&co
->scheduled
, NULL
);
516 qemu_aio_coroutine_enter(ctx
, co
);
517 aio_context_release(ctx
);
521 AioContext
*aio_context_new(Error
**errp
)
526 ctx
= (AioContext
*) g_source_new(&aio_source_funcs
, sizeof(AioContext
));
527 QSLIST_INIT(&ctx
->bh_list
);
528 QSIMPLEQ_INIT(&ctx
->bh_slice_list
);
529 aio_context_setup(ctx
);
531 ret
= event_notifier_init(&ctx
->notifier
, false);
533 error_setg_errno(errp
, -ret
, "Failed to initialize event notifier");
536 g_source_set_can_recurse(&ctx
->source
, true);
537 qemu_lockcnt_init(&ctx
->list_lock
);
539 ctx
->co_schedule_bh
= aio_bh_new(ctx
, co_schedule_bh_cb
, ctx
);
540 QSLIST_INIT(&ctx
->scheduled_coroutines
);
542 aio_set_event_notifier(ctx
, &ctx
->notifier
,
544 aio_context_notifier_cb
,
545 aio_context_notifier_poll
,
546 aio_context_notifier_poll_ready
);
547 #ifdef CONFIG_LINUX_AIO
548 ctx
->linux_aio
= NULL
;
551 #ifdef CONFIG_LINUX_IO_URING
552 ctx
->linux_io_uring
= NULL
;
555 ctx
->thread_pool
= NULL
;
556 qemu_rec_mutex_init(&ctx
->lock
);
557 timerlistgroup_init(&ctx
->tlg
, aio_timerlist_notify
, ctx
);
560 ctx
->poll_max_ns
= 0;
562 ctx
->poll_shrink
= 0;
564 ctx
->aio_max_batch
= 0;
568 g_source_destroy(&ctx
->source
);
572 void aio_co_schedule(AioContext
*ctx
, Coroutine
*co
)
574 trace_aio_co_schedule(ctx
, co
);
575 const char *scheduled
= qatomic_cmpxchg(&co
->scheduled
, NULL
,
580 "%s: Co-routine was already scheduled in '%s'\n",
581 __func__
, scheduled
);
585 /* The coroutine might run and release the last ctx reference before we
586 * invoke qemu_bh_schedule(). Take a reference to keep ctx alive until
589 aio_context_ref(ctx
);
591 QSLIST_INSERT_HEAD_ATOMIC(&ctx
->scheduled_coroutines
,
592 co
, co_scheduled_next
);
593 qemu_bh_schedule(ctx
->co_schedule_bh
);
595 aio_context_unref(ctx
);
598 typedef struct AioCoRescheduleSelf
{
601 } AioCoRescheduleSelf
;
603 static void aio_co_reschedule_self_bh(void *opaque
)
605 AioCoRescheduleSelf
*data
= opaque
;
606 aio_co_schedule(data
->new_ctx
, data
->co
);
609 void coroutine_fn
aio_co_reschedule_self(AioContext
*new_ctx
)
611 AioContext
*old_ctx
= qemu_get_current_aio_context();
613 if (old_ctx
!= new_ctx
) {
614 AioCoRescheduleSelf data
= {
615 .co
= qemu_coroutine_self(),
619 * We can't directly schedule the coroutine in the target context
620 * because this would be racy: The other thread could try to enter the
621 * coroutine before it has yielded in this one.
623 aio_bh_schedule_oneshot(old_ctx
, aio_co_reschedule_self_bh
, &data
);
624 qemu_coroutine_yield();
628 void aio_co_wake(struct Coroutine
*co
)
632 /* Read coroutine before co->ctx. Matches smp_wmb in
633 * qemu_coroutine_enter.
635 smp_read_barrier_depends();
636 ctx
= qatomic_read(&co
->ctx
);
638 aio_co_enter(ctx
, co
);
641 void aio_co_enter(AioContext
*ctx
, struct Coroutine
*co
)
643 if (ctx
!= qemu_get_current_aio_context()) {
644 aio_co_schedule(ctx
, co
);
648 if (qemu_in_coroutine()) {
649 Coroutine
*self
= qemu_coroutine_self();
651 QSIMPLEQ_INSERT_TAIL(&self
->co_queue_wakeup
, co
, co_queue_next
);
653 aio_context_acquire(ctx
);
654 qemu_aio_coroutine_enter(ctx
, co
);
655 aio_context_release(ctx
);
659 void aio_context_ref(AioContext
*ctx
)
661 g_source_ref(&ctx
->source
);
664 void aio_context_unref(AioContext
*ctx
)
666 g_source_unref(&ctx
->source
);
669 void aio_context_acquire(AioContext
*ctx
)
671 qemu_rec_mutex_lock(&ctx
->lock
);
674 void aio_context_release(AioContext
*ctx
)
676 qemu_rec_mutex_unlock(&ctx
->lock
);
679 QEMU_DEFINE_STATIC_CO_TLS(AioContext
*, my_aiocontext
)
681 AioContext
*qemu_get_current_aio_context(void)
683 AioContext
*ctx
= get_my_aiocontext();
687 if (qemu_mutex_iothread_locked()) {
688 /* Possibly in a vCPU thread. */
689 return qemu_get_aio_context();
694 void qemu_set_current_aio_context(AioContext
*ctx
)
696 assert(!get_my_aiocontext());
697 set_my_aiocontext(ctx
);