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 "block/graph-lock.h"
31 #include "qemu/main-loop.h"
32 #include "qemu/atomic.h"
33 #include "qemu/rcu_queue.h"
34 #include "block/raw-aio.h"
35 #include "qemu/coroutine_int.h"
36 #include "qemu/coroutine-tls.h"
37 #include "sysemu/cpu-timers.h"
40 /***********************************************************/
41 /* bottom halves (can be seen as timers which expire ASAP) */
43 /* QEMUBH::flags values */
45 /* Already enqueued and waiting for aio_bh_poll() */
46 BH_PENDING
= (1 << 0),
48 /* Invoke the callback */
49 BH_SCHEDULED
= (1 << 1),
51 /* Delete without invoking callback */
52 BH_DELETED
= (1 << 2),
54 /* Delete after invoking callback */
55 BH_ONESHOT
= (1 << 3),
57 /* Schedule periodically when the event loop is idle */
66 QSLIST_ENTRY(QEMUBH
) next
;
70 /* Called concurrently from any thread */
71 static void aio_bh_enqueue(QEMUBH
*bh
, unsigned new_flags
)
73 AioContext
*ctx
= bh
->ctx
;
77 * The memory barrier implicit in qatomic_fetch_or makes sure that:
78 * 1. idle & any writes needed by the callback are done before the
79 * locations are read in the aio_bh_poll.
80 * 2. ctx is loaded before the callback has a chance to execute and bh
83 old_flags
= qatomic_fetch_or(&bh
->flags
, BH_PENDING
| new_flags
);
84 if (!(old_flags
& BH_PENDING
)) {
85 QSLIST_INSERT_HEAD_ATOMIC(&ctx
->bh_list
, bh
, next
);
90 * Workaround for record/replay.
91 * vCPU execution should be suspended when new BH is set.
92 * This is needed to avoid guest timeouts caused
93 * by the long cycles of the execution.
98 /* Only called from aio_bh_poll() and aio_ctx_finalize() */
99 static QEMUBH
*aio_bh_dequeue(BHList
*head
, unsigned *flags
)
101 QEMUBH
*bh
= QSLIST_FIRST_RCU(head
);
107 QSLIST_REMOVE_HEAD(head
, next
);
110 * The qatomic_and is paired with aio_bh_enqueue(). The implicit memory
111 * barrier ensures that the callback sees all writes done by the scheduling
112 * thread. It also ensures that the scheduling thread sees the cleared
113 * flag before bh->cb has run, and thus will call aio_notify again if
116 *flags
= qatomic_fetch_and(&bh
->flags
,
117 ~(BH_PENDING
| BH_SCHEDULED
| BH_IDLE
));
121 void aio_bh_schedule_oneshot_full(AioContext
*ctx
, QEMUBHFunc
*cb
,
122 void *opaque
, const char *name
)
125 bh
= g_new(QEMUBH
, 1);
132 aio_bh_enqueue(bh
, BH_SCHEDULED
| BH_ONESHOT
);
135 QEMUBH
*aio_bh_new_full(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
,
139 bh
= g_new(QEMUBH
, 1);
149 void aio_bh_call(QEMUBH
*bh
)
154 /* Multiple occurrences of aio_bh_poll cannot be called concurrently. */
155 int aio_bh_poll(AioContext
*ctx
)
161 QSLIST_MOVE_ATOMIC(&slice
.bh_list
, &ctx
->bh_list
);
162 QSIMPLEQ_INSERT_TAIL(&ctx
->bh_slice_list
, &slice
, next
);
164 while ((s
= QSIMPLEQ_FIRST(&ctx
->bh_slice_list
))) {
168 bh
= aio_bh_dequeue(&s
->bh_list
, &flags
);
170 QSIMPLEQ_REMOVE_HEAD(&ctx
->bh_slice_list
, next
);
174 if ((flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
175 /* Idle BHs don't count as progress */
176 if (!(flags
& BH_IDLE
)) {
181 if (flags
& (BH_DELETED
| BH_ONESHOT
)) {
189 void qemu_bh_schedule_idle(QEMUBH
*bh
)
191 aio_bh_enqueue(bh
, BH_SCHEDULED
| BH_IDLE
);
194 void qemu_bh_schedule(QEMUBH
*bh
)
196 aio_bh_enqueue(bh
, BH_SCHEDULED
);
199 /* This func is async.
201 void qemu_bh_cancel(QEMUBH
*bh
)
203 qatomic_and(&bh
->flags
, ~BH_SCHEDULED
);
206 /* This func is async.The bottom half will do the delete action at the finial
209 void qemu_bh_delete(QEMUBH
*bh
)
211 aio_bh_enqueue(bh
, BH_DELETED
);
214 static int64_t aio_compute_bh_timeout(BHList
*head
, int timeout
)
218 QSLIST_FOREACH_RCU(bh
, head
, next
) {
219 if ((bh
->flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
220 if (bh
->flags
& BH_IDLE
) {
221 /* idle bottom halves will be polled at least
225 /* non-idle bottom halves will be executed
236 aio_compute_timeout(AioContext
*ctx
)
242 timeout
= aio_compute_bh_timeout(&ctx
->bh_list
, timeout
);
247 QSIMPLEQ_FOREACH(s
, &ctx
->bh_slice_list
, next
) {
248 timeout
= aio_compute_bh_timeout(&s
->bh_list
, timeout
);
254 deadline
= timerlistgroup_deadline_ns(&ctx
->tlg
);
258 return qemu_soonest_timeout(timeout
, deadline
);
263 aio_ctx_prepare(GSource
*source
, gint
*timeout
)
265 AioContext
*ctx
= (AioContext
*) source
;
267 qatomic_set(&ctx
->notify_me
, qatomic_read(&ctx
->notify_me
) | 1);
270 * Write ctx->notify_me before computing the timeout
271 * (reading bottom half flags, etc.). Pairs with
272 * smp_mb in aio_notify().
276 /* We assume there is no timeout already supplied */
277 *timeout
= qemu_timeout_ns_to_ms(aio_compute_timeout(ctx
));
279 if (aio_prepare(ctx
)) {
283 return *timeout
== 0;
287 aio_ctx_check(GSource
*source
)
289 AioContext
*ctx
= (AioContext
*) source
;
293 /* Finish computing the timeout before clearing the flag. */
294 qatomic_store_release(&ctx
->notify_me
, qatomic_read(&ctx
->notify_me
) & ~1);
295 aio_notify_accept(ctx
);
297 QSLIST_FOREACH_RCU(bh
, &ctx
->bh_list
, next
) {
298 if ((bh
->flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
303 QSIMPLEQ_FOREACH(s
, &ctx
->bh_slice_list
, next
) {
304 QSLIST_FOREACH_RCU(bh
, &s
->bh_list
, next
) {
305 if ((bh
->flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
310 return aio_pending(ctx
) || (timerlistgroup_deadline_ns(&ctx
->tlg
) == 0);
314 aio_ctx_dispatch(GSource
*source
,
315 GSourceFunc callback
,
318 AioContext
*ctx
= (AioContext
*) source
;
320 assert(callback
== NULL
);
326 aio_ctx_finalize(GSource
*source
)
328 AioContext
*ctx
= (AioContext
*) source
;
332 thread_pool_free(ctx
->thread_pool
);
334 #ifdef CONFIG_LINUX_AIO
335 if (ctx
->linux_aio
) {
336 laio_detach_aio_context(ctx
->linux_aio
, ctx
);
337 laio_cleanup(ctx
->linux_aio
);
338 ctx
->linux_aio
= NULL
;
342 #ifdef CONFIG_LINUX_IO_URING
343 if (ctx
->linux_io_uring
) {
344 luring_detach_aio_context(ctx
->linux_io_uring
, ctx
);
345 luring_cleanup(ctx
->linux_io_uring
);
346 ctx
->linux_io_uring
= NULL
;
350 assert(QSLIST_EMPTY(&ctx
->scheduled_coroutines
));
351 qemu_bh_delete(ctx
->co_schedule_bh
);
353 /* There must be no aio_bh_poll() calls going on */
354 assert(QSIMPLEQ_EMPTY(&ctx
->bh_slice_list
));
356 while ((bh
= aio_bh_dequeue(&ctx
->bh_list
, &flags
))) {
358 * qemu_bh_delete() must have been called on BHs in this AioContext. In
359 * many cases memory leaks, hangs, or inconsistent state occur when a
360 * BH is leaked because something still expects it to run.
362 * If you hit this, fix the lifecycle of the BH so that
363 * qemu_bh_delete() and any associated cleanup is called before the
364 * AioContext is finalized.
366 if (unlikely(!(flags
& BH_DELETED
))) {
367 fprintf(stderr
, "%s: BH '%s' leaked, aborting...\n",
375 aio_set_event_notifier(ctx
, &ctx
->notifier
, false, NULL
, NULL
, NULL
);
376 event_notifier_cleanup(&ctx
->notifier
);
377 qemu_rec_mutex_destroy(&ctx
->lock
);
378 qemu_lockcnt_destroy(&ctx
->list_lock
);
379 timerlistgroup_deinit(&ctx
->tlg
);
380 unregister_aiocontext(ctx
);
381 aio_context_destroy(ctx
);
384 static GSourceFuncs aio_source_funcs
= {
391 GSource
*aio_get_g_source(AioContext
*ctx
)
393 aio_context_use_g_source(ctx
);
394 g_source_ref(&ctx
->source
);
398 ThreadPool
*aio_get_thread_pool(AioContext
*ctx
)
400 if (!ctx
->thread_pool
) {
401 ctx
->thread_pool
= thread_pool_new(ctx
);
403 return ctx
->thread_pool
;
406 #ifdef CONFIG_LINUX_AIO
407 LinuxAioState
*aio_setup_linux_aio(AioContext
*ctx
, Error
**errp
)
409 if (!ctx
->linux_aio
) {
410 ctx
->linux_aio
= laio_init(errp
);
411 if (ctx
->linux_aio
) {
412 laio_attach_aio_context(ctx
->linux_aio
, ctx
);
415 return ctx
->linux_aio
;
418 LinuxAioState
*aio_get_linux_aio(AioContext
*ctx
)
420 assert(ctx
->linux_aio
);
421 return ctx
->linux_aio
;
425 #ifdef CONFIG_LINUX_IO_URING
426 LuringState
*aio_setup_linux_io_uring(AioContext
*ctx
, Error
**errp
)
428 if (ctx
->linux_io_uring
) {
429 return ctx
->linux_io_uring
;
432 ctx
->linux_io_uring
= luring_init(errp
);
433 if (!ctx
->linux_io_uring
) {
437 luring_attach_aio_context(ctx
->linux_io_uring
, ctx
);
438 return ctx
->linux_io_uring
;
441 LuringState
*aio_get_linux_io_uring(AioContext
*ctx
)
443 assert(ctx
->linux_io_uring
);
444 return ctx
->linux_io_uring
;
448 void aio_notify(AioContext
*ctx
)
451 * Write e.g. bh->flags before writing ctx->notified. Pairs with smp_mb in
455 qatomic_set(&ctx
->notified
, true);
458 * Write ctx->notified before reading ctx->notify_me. Pairs
459 * with smp_mb in aio_ctx_prepare or aio_poll.
462 if (qatomic_read(&ctx
->notify_me
)) {
463 event_notifier_set(&ctx
->notifier
);
467 void aio_notify_accept(AioContext
*ctx
)
469 qatomic_set(&ctx
->notified
, false);
472 * Write ctx->notified before reading e.g. bh->flags. Pairs with smp_wmb
478 static void aio_timerlist_notify(void *opaque
, QEMUClockType type
)
483 static void aio_context_notifier_cb(EventNotifier
*e
)
485 AioContext
*ctx
= container_of(e
, AioContext
, notifier
);
487 event_notifier_test_and_clear(&ctx
->notifier
);
490 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
491 static bool aio_context_notifier_poll(void *opaque
)
493 EventNotifier
*e
= opaque
;
494 AioContext
*ctx
= container_of(e
, AioContext
, notifier
);
496 return qatomic_read(&ctx
->notified
);
499 static void aio_context_notifier_poll_ready(EventNotifier
*e
)
501 /* Do nothing, we just wanted to kick the event loop */
504 static void co_schedule_bh_cb(void *opaque
)
506 AioContext
*ctx
= opaque
;
507 QSLIST_HEAD(, Coroutine
) straight
, reversed
;
509 QSLIST_MOVE_ATOMIC(&reversed
, &ctx
->scheduled_coroutines
);
510 QSLIST_INIT(&straight
);
512 while (!QSLIST_EMPTY(&reversed
)) {
513 Coroutine
*co
= QSLIST_FIRST(&reversed
);
514 QSLIST_REMOVE_HEAD(&reversed
, co_scheduled_next
);
515 QSLIST_INSERT_HEAD(&straight
, co
, co_scheduled_next
);
518 while (!QSLIST_EMPTY(&straight
)) {
519 Coroutine
*co
= QSLIST_FIRST(&straight
);
520 QSLIST_REMOVE_HEAD(&straight
, co_scheduled_next
);
521 trace_aio_co_schedule_bh_cb(ctx
, co
);
522 aio_context_acquire(ctx
);
524 /* Protected by write barrier in qemu_aio_coroutine_enter */
525 qatomic_set(&co
->scheduled
, NULL
);
526 qemu_aio_coroutine_enter(ctx
, co
);
527 aio_context_release(ctx
);
531 AioContext
*aio_context_new(Error
**errp
)
536 ctx
= (AioContext
*) g_source_new(&aio_source_funcs
, sizeof(AioContext
));
537 QSLIST_INIT(&ctx
->bh_list
);
538 QSIMPLEQ_INIT(&ctx
->bh_slice_list
);
539 aio_context_setup(ctx
);
541 ret
= event_notifier_init(&ctx
->notifier
, false);
543 error_setg_errno(errp
, -ret
, "Failed to initialize event notifier");
546 g_source_set_can_recurse(&ctx
->source
, true);
547 qemu_lockcnt_init(&ctx
->list_lock
);
549 ctx
->co_schedule_bh
= aio_bh_new(ctx
, co_schedule_bh_cb
, ctx
);
550 QSLIST_INIT(&ctx
->scheduled_coroutines
);
552 aio_set_event_notifier(ctx
, &ctx
->notifier
,
554 aio_context_notifier_cb
,
555 aio_context_notifier_poll
,
556 aio_context_notifier_poll_ready
);
557 #ifdef CONFIG_LINUX_AIO
558 ctx
->linux_aio
= NULL
;
561 #ifdef CONFIG_LINUX_IO_URING
562 ctx
->linux_io_uring
= NULL
;
565 ctx
->thread_pool
= NULL
;
566 qemu_rec_mutex_init(&ctx
->lock
);
567 timerlistgroup_init(&ctx
->tlg
, aio_timerlist_notify
, ctx
);
570 ctx
->poll_max_ns
= 0;
572 ctx
->poll_shrink
= 0;
574 ctx
->aio_max_batch
= 0;
576 ctx
->thread_pool_min
= 0;
577 ctx
->thread_pool_max
= THREAD_POOL_MAX_THREADS_DEFAULT
;
579 register_aiocontext(ctx
);
583 g_source_destroy(&ctx
->source
);
587 void aio_co_schedule(AioContext
*ctx
, Coroutine
*co
)
589 trace_aio_co_schedule(ctx
, co
);
590 const char *scheduled
= qatomic_cmpxchg(&co
->scheduled
, NULL
,
595 "%s: Co-routine was already scheduled in '%s'\n",
596 __func__
, scheduled
);
600 /* The coroutine might run and release the last ctx reference before we
601 * invoke qemu_bh_schedule(). Take a reference to keep ctx alive until
604 aio_context_ref(ctx
);
606 QSLIST_INSERT_HEAD_ATOMIC(&ctx
->scheduled_coroutines
,
607 co
, co_scheduled_next
);
608 qemu_bh_schedule(ctx
->co_schedule_bh
);
610 aio_context_unref(ctx
);
613 typedef struct AioCoRescheduleSelf
{
616 } AioCoRescheduleSelf
;
618 static void aio_co_reschedule_self_bh(void *opaque
)
620 AioCoRescheduleSelf
*data
= opaque
;
621 aio_co_schedule(data
->new_ctx
, data
->co
);
624 void coroutine_fn
aio_co_reschedule_self(AioContext
*new_ctx
)
626 AioContext
*old_ctx
= qemu_get_current_aio_context();
628 if (old_ctx
!= new_ctx
) {
629 AioCoRescheduleSelf data
= {
630 .co
= qemu_coroutine_self(),
634 * We can't directly schedule the coroutine in the target context
635 * because this would be racy: The other thread could try to enter the
636 * coroutine before it has yielded in this one.
638 aio_bh_schedule_oneshot(old_ctx
, aio_co_reschedule_self_bh
, &data
);
639 qemu_coroutine_yield();
643 void aio_co_wake(Coroutine
*co
)
647 /* Read coroutine before co->ctx. Matches smp_wmb in
648 * qemu_coroutine_enter.
650 smp_read_barrier_depends();
651 ctx
= qatomic_read(&co
->ctx
);
653 aio_co_enter(ctx
, co
);
656 void aio_co_enter(AioContext
*ctx
, Coroutine
*co
)
658 if (ctx
!= qemu_get_current_aio_context()) {
659 aio_co_schedule(ctx
, co
);
663 if (qemu_in_coroutine()) {
664 Coroutine
*self
= qemu_coroutine_self();
666 QSIMPLEQ_INSERT_TAIL(&self
->co_queue_wakeup
, co
, co_queue_next
);
668 aio_context_acquire(ctx
);
669 qemu_aio_coroutine_enter(ctx
, co
);
670 aio_context_release(ctx
);
674 void aio_context_ref(AioContext
*ctx
)
676 g_source_ref(&ctx
->source
);
679 void aio_context_unref(AioContext
*ctx
)
681 g_source_unref(&ctx
->source
);
684 void aio_context_acquire(AioContext
*ctx
)
686 qemu_rec_mutex_lock(&ctx
->lock
);
689 void aio_context_release(AioContext
*ctx
)
691 qemu_rec_mutex_unlock(&ctx
->lock
);
694 QEMU_DEFINE_STATIC_CO_TLS(AioContext
*, my_aiocontext
)
696 AioContext
*qemu_get_current_aio_context(void)
698 AioContext
*ctx
= get_my_aiocontext();
702 if (qemu_mutex_iothread_locked()) {
703 /* Possibly in a vCPU thread. */
704 return qemu_get_aio_context();
709 void qemu_set_current_aio_context(AioContext
*ctx
)
711 assert(!get_my_aiocontext());
712 set_my_aiocontext(ctx
);
715 void aio_context_set_thread_pool_params(AioContext
*ctx
, int64_t min
,
716 int64_t max
, Error
**errp
)
719 if (min
> max
|| !max
|| min
> INT_MAX
|| max
> INT_MAX
) {
720 error_setg(errp
, "bad thread-pool-min/thread-pool-max values");
724 ctx
->thread_pool_min
= min
;
725 ctx
->thread_pool_max
= max
;
727 if (ctx
->thread_pool
) {
728 thread_pool_update_params(ctx
->thread_pool
, ctx
);