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"
37 /***********************************************************/
38 /* bottom halves (can be seen as timers which expire ASAP) */
40 /* QEMUBH::flags values */
42 /* Already enqueued and waiting for aio_bh_poll() */
43 BH_PENDING
= (1 << 0),
45 /* Invoke the callback */
46 BH_SCHEDULED
= (1 << 1),
48 /* Delete without invoking callback */
49 BH_DELETED
= (1 << 2),
51 /* Delete after invoking callback */
52 BH_ONESHOT
= (1 << 3),
54 /* Schedule periodically when the event loop is idle */
63 QSLIST_ENTRY(QEMUBH
) next
;
67 /* Called concurrently from any thread */
68 static void aio_bh_enqueue(QEMUBH
*bh
, unsigned new_flags
)
70 AioContext
*ctx
= bh
->ctx
;
74 * The memory barrier implicit in qatomic_fetch_or makes sure that:
75 * 1. idle & any writes needed by the callback are done before the
76 * locations are read in the aio_bh_poll.
77 * 2. ctx is loaded before the callback has a chance to execute and bh
80 old_flags
= qatomic_fetch_or(&bh
->flags
, BH_PENDING
| new_flags
);
81 if (!(old_flags
& BH_PENDING
)) {
82 QSLIST_INSERT_HEAD_ATOMIC(&ctx
->bh_list
, bh
, next
);
88 /* Only called from aio_bh_poll() and aio_ctx_finalize() */
89 static QEMUBH
*aio_bh_dequeue(BHList
*head
, unsigned *flags
)
91 QEMUBH
*bh
= QSLIST_FIRST_RCU(head
);
97 QSLIST_REMOVE_HEAD(head
, next
);
100 * The qatomic_and is paired with aio_bh_enqueue(). The implicit memory
101 * barrier ensures that the callback sees all writes done by the scheduling
102 * thread. It also ensures that the scheduling thread sees the cleared
103 * flag before bh->cb has run, and thus will call aio_notify again if
106 *flags
= qatomic_fetch_and(&bh
->flags
,
107 ~(BH_PENDING
| BH_SCHEDULED
| BH_IDLE
));
111 void aio_bh_schedule_oneshot_full(AioContext
*ctx
, QEMUBHFunc
*cb
,
112 void *opaque
, const char *name
)
115 bh
= g_new(QEMUBH
, 1);
122 aio_bh_enqueue(bh
, BH_SCHEDULED
| BH_ONESHOT
);
125 QEMUBH
*aio_bh_new_full(AioContext
*ctx
, QEMUBHFunc
*cb
, void *opaque
,
129 bh
= g_new(QEMUBH
, 1);
139 void aio_bh_call(QEMUBH
*bh
)
144 /* Multiple occurrences of aio_bh_poll cannot be called concurrently. */
145 int aio_bh_poll(AioContext
*ctx
)
151 QSLIST_MOVE_ATOMIC(&slice
.bh_list
, &ctx
->bh_list
);
152 QSIMPLEQ_INSERT_TAIL(&ctx
->bh_slice_list
, &slice
, next
);
154 while ((s
= QSIMPLEQ_FIRST(&ctx
->bh_slice_list
))) {
158 bh
= aio_bh_dequeue(&s
->bh_list
, &flags
);
160 QSIMPLEQ_REMOVE_HEAD(&ctx
->bh_slice_list
, next
);
164 if ((flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
165 /* Idle BHs don't count as progress */
166 if (!(flags
& BH_IDLE
)) {
171 if (flags
& (BH_DELETED
| BH_ONESHOT
)) {
179 void qemu_bh_schedule_idle(QEMUBH
*bh
)
181 aio_bh_enqueue(bh
, BH_SCHEDULED
| BH_IDLE
);
184 void qemu_bh_schedule(QEMUBH
*bh
)
186 aio_bh_enqueue(bh
, BH_SCHEDULED
);
189 /* This func is async.
191 void qemu_bh_cancel(QEMUBH
*bh
)
193 qatomic_and(&bh
->flags
, ~BH_SCHEDULED
);
196 /* This func is async.The bottom half will do the delete action at the finial
199 void qemu_bh_delete(QEMUBH
*bh
)
201 aio_bh_enqueue(bh
, BH_DELETED
);
204 static int64_t aio_compute_bh_timeout(BHList
*head
, int timeout
)
208 QSLIST_FOREACH_RCU(bh
, head
, next
) {
209 if ((bh
->flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
210 if (bh
->flags
& BH_IDLE
) {
211 /* idle bottom halves will be polled at least
215 /* non-idle bottom halves will be executed
226 aio_compute_timeout(AioContext
*ctx
)
232 timeout
= aio_compute_bh_timeout(&ctx
->bh_list
, timeout
);
237 QSIMPLEQ_FOREACH(s
, &ctx
->bh_slice_list
, next
) {
238 timeout
= aio_compute_bh_timeout(&s
->bh_list
, timeout
);
244 deadline
= timerlistgroup_deadline_ns(&ctx
->tlg
);
248 return qemu_soonest_timeout(timeout
, deadline
);
253 aio_ctx_prepare(GSource
*source
, gint
*timeout
)
255 AioContext
*ctx
= (AioContext
*) source
;
257 qatomic_set(&ctx
->notify_me
, qatomic_read(&ctx
->notify_me
) | 1);
260 * Write ctx->notify_me before computing the timeout
261 * (reading bottom half flags, etc.). Pairs with
262 * smp_mb in aio_notify().
266 /* We assume there is no timeout already supplied */
267 *timeout
= qemu_timeout_ns_to_ms(aio_compute_timeout(ctx
));
269 if (aio_prepare(ctx
)) {
273 return *timeout
== 0;
277 aio_ctx_check(GSource
*source
)
279 AioContext
*ctx
= (AioContext
*) source
;
283 /* Finish computing the timeout before clearing the flag. */
284 qatomic_store_release(&ctx
->notify_me
, qatomic_read(&ctx
->notify_me
) & ~1);
285 aio_notify_accept(ctx
);
287 QSLIST_FOREACH_RCU(bh
, &ctx
->bh_list
, next
) {
288 if ((bh
->flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
293 QSIMPLEQ_FOREACH(s
, &ctx
->bh_slice_list
, next
) {
294 QSLIST_FOREACH_RCU(bh
, &s
->bh_list
, next
) {
295 if ((bh
->flags
& (BH_SCHEDULED
| BH_DELETED
)) == BH_SCHEDULED
) {
300 return aio_pending(ctx
) || (timerlistgroup_deadline_ns(&ctx
->tlg
) == 0);
304 aio_ctx_dispatch(GSource
*source
,
305 GSourceFunc callback
,
308 AioContext
*ctx
= (AioContext
*) source
;
310 assert(callback
== NULL
);
316 aio_ctx_finalize(GSource
*source
)
318 AioContext
*ctx
= (AioContext
*) source
;
322 thread_pool_free(ctx
->thread_pool
);
324 #ifdef CONFIG_LINUX_AIO
325 if (ctx
->linux_aio
) {
326 laio_detach_aio_context(ctx
->linux_aio
, ctx
);
327 laio_cleanup(ctx
->linux_aio
);
328 ctx
->linux_aio
= NULL
;
332 #ifdef CONFIG_LINUX_IO_URING
333 if (ctx
->linux_io_uring
) {
334 luring_detach_aio_context(ctx
->linux_io_uring
, ctx
);
335 luring_cleanup(ctx
->linux_io_uring
);
336 ctx
->linux_io_uring
= NULL
;
340 assert(QSLIST_EMPTY(&ctx
->scheduled_coroutines
));
341 qemu_bh_delete(ctx
->co_schedule_bh
);
343 /* There must be no aio_bh_poll() calls going on */
344 assert(QSIMPLEQ_EMPTY(&ctx
->bh_slice_list
));
346 while ((bh
= aio_bh_dequeue(&ctx
->bh_list
, &flags
))) {
348 * qemu_bh_delete() must have been called on BHs in this AioContext. In
349 * many cases memory leaks, hangs, or inconsistent state occur when a
350 * BH is leaked because something still expects it to run.
352 * If you hit this, fix the lifecycle of the BH so that
353 * qemu_bh_delete() and any associated cleanup is called before the
354 * AioContext is finalized.
356 if (unlikely(!(flags
& BH_DELETED
))) {
357 fprintf(stderr
, "%s: BH '%s' leaked, aborting...\n",
365 aio_set_event_notifier(ctx
, &ctx
->notifier
, false, NULL
, NULL
);
366 event_notifier_cleanup(&ctx
->notifier
);
367 qemu_rec_mutex_destroy(&ctx
->lock
);
368 qemu_lockcnt_destroy(&ctx
->list_lock
);
369 timerlistgroup_deinit(&ctx
->tlg
);
370 aio_context_destroy(ctx
);
373 static GSourceFuncs aio_source_funcs
= {
380 GSource
*aio_get_g_source(AioContext
*ctx
)
382 aio_context_use_g_source(ctx
);
383 g_source_ref(&ctx
->source
);
387 ThreadPool
*aio_get_thread_pool(AioContext
*ctx
)
389 if (!ctx
->thread_pool
) {
390 ctx
->thread_pool
= thread_pool_new(ctx
);
392 return ctx
->thread_pool
;
395 #ifdef CONFIG_LINUX_AIO
396 LinuxAioState
*aio_setup_linux_aio(AioContext
*ctx
, Error
**errp
)
398 if (!ctx
->linux_aio
) {
399 ctx
->linux_aio
= laio_init(errp
);
400 if (ctx
->linux_aio
) {
401 laio_attach_aio_context(ctx
->linux_aio
, ctx
);
404 return ctx
->linux_aio
;
407 LinuxAioState
*aio_get_linux_aio(AioContext
*ctx
)
409 assert(ctx
->linux_aio
);
410 return ctx
->linux_aio
;
414 #ifdef CONFIG_LINUX_IO_URING
415 LuringState
*aio_setup_linux_io_uring(AioContext
*ctx
, Error
**errp
)
417 if (ctx
->linux_io_uring
) {
418 return ctx
->linux_io_uring
;
421 ctx
->linux_io_uring
= luring_init(errp
);
422 if (!ctx
->linux_io_uring
) {
426 luring_attach_aio_context(ctx
->linux_io_uring
, ctx
);
427 return ctx
->linux_io_uring
;
430 LuringState
*aio_get_linux_io_uring(AioContext
*ctx
)
432 assert(ctx
->linux_io_uring
);
433 return ctx
->linux_io_uring
;
437 void aio_notify(AioContext
*ctx
)
440 * Write e.g. bh->flags before writing ctx->notified. Pairs with smp_mb in
444 qatomic_set(&ctx
->notified
, true);
447 * Write ctx->notified before reading ctx->notify_me. Pairs
448 * with smp_mb in aio_ctx_prepare or aio_poll.
451 if (qatomic_read(&ctx
->notify_me
)) {
452 event_notifier_set(&ctx
->notifier
);
456 void aio_notify_accept(AioContext
*ctx
)
458 qatomic_set(&ctx
->notified
, false);
461 * Write ctx->notified before reading e.g. bh->flags. Pairs with smp_wmb
467 static void aio_timerlist_notify(void *opaque
, QEMUClockType type
)
472 static void aio_context_notifier_cb(EventNotifier
*e
)
474 AioContext
*ctx
= container_of(e
, AioContext
, notifier
);
476 event_notifier_test_and_clear(&ctx
->notifier
);
479 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
480 static bool aio_context_notifier_poll(void *opaque
)
482 EventNotifier
*e
= opaque
;
483 AioContext
*ctx
= container_of(e
, AioContext
, notifier
);
485 return qatomic_read(&ctx
->notified
);
488 static void co_schedule_bh_cb(void *opaque
)
490 AioContext
*ctx
= opaque
;
491 QSLIST_HEAD(, Coroutine
) straight
, reversed
;
493 QSLIST_MOVE_ATOMIC(&reversed
, &ctx
->scheduled_coroutines
);
494 QSLIST_INIT(&straight
);
496 while (!QSLIST_EMPTY(&reversed
)) {
497 Coroutine
*co
= QSLIST_FIRST(&reversed
);
498 QSLIST_REMOVE_HEAD(&reversed
, co_scheduled_next
);
499 QSLIST_INSERT_HEAD(&straight
, co
, co_scheduled_next
);
502 while (!QSLIST_EMPTY(&straight
)) {
503 Coroutine
*co
= QSLIST_FIRST(&straight
);
504 QSLIST_REMOVE_HEAD(&straight
, co_scheduled_next
);
505 trace_aio_co_schedule_bh_cb(ctx
, co
);
506 aio_context_acquire(ctx
);
508 /* Protected by write barrier in qemu_aio_coroutine_enter */
509 qatomic_set(&co
->scheduled
, NULL
);
510 qemu_aio_coroutine_enter(ctx
, co
);
511 aio_context_release(ctx
);
515 AioContext
*aio_context_new(Error
**errp
)
520 ctx
= (AioContext
*) g_source_new(&aio_source_funcs
, sizeof(AioContext
));
521 QSLIST_INIT(&ctx
->bh_list
);
522 QSIMPLEQ_INIT(&ctx
->bh_slice_list
);
523 aio_context_setup(ctx
);
525 ret
= event_notifier_init(&ctx
->notifier
, false);
527 error_setg_errno(errp
, -ret
, "Failed to initialize event notifier");
530 g_source_set_can_recurse(&ctx
->source
, true);
531 qemu_lockcnt_init(&ctx
->list_lock
);
533 ctx
->co_schedule_bh
= aio_bh_new(ctx
, co_schedule_bh_cb
, ctx
);
534 QSLIST_INIT(&ctx
->scheduled_coroutines
);
536 aio_set_event_notifier(ctx
, &ctx
->notifier
,
538 aio_context_notifier_cb
,
539 aio_context_notifier_poll
);
540 #ifdef CONFIG_LINUX_AIO
541 ctx
->linux_aio
= NULL
;
544 #ifdef CONFIG_LINUX_IO_URING
545 ctx
->linux_io_uring
= NULL
;
548 ctx
->thread_pool
= NULL
;
549 qemu_rec_mutex_init(&ctx
->lock
);
550 timerlistgroup_init(&ctx
->tlg
, aio_timerlist_notify
, ctx
);
553 ctx
->poll_max_ns
= 0;
555 ctx
->poll_shrink
= 0;
557 ctx
->aio_max_batch
= 0;
561 g_source_destroy(&ctx
->source
);
565 void aio_co_schedule(AioContext
*ctx
, Coroutine
*co
)
567 trace_aio_co_schedule(ctx
, co
);
568 const char *scheduled
= qatomic_cmpxchg(&co
->scheduled
, NULL
,
573 "%s: Co-routine was already scheduled in '%s'\n",
574 __func__
, scheduled
);
578 /* The coroutine might run and release the last ctx reference before we
579 * invoke qemu_bh_schedule(). Take a reference to keep ctx alive until
582 aio_context_ref(ctx
);
584 QSLIST_INSERT_HEAD_ATOMIC(&ctx
->scheduled_coroutines
,
585 co
, co_scheduled_next
);
586 qemu_bh_schedule(ctx
->co_schedule_bh
);
588 aio_context_unref(ctx
);
591 typedef struct AioCoRescheduleSelf
{
594 } AioCoRescheduleSelf
;
596 static void aio_co_reschedule_self_bh(void *opaque
)
598 AioCoRescheduleSelf
*data
= opaque
;
599 aio_co_schedule(data
->new_ctx
, data
->co
);
602 void coroutine_fn
aio_co_reschedule_self(AioContext
*new_ctx
)
604 AioContext
*old_ctx
= qemu_get_current_aio_context();
606 if (old_ctx
!= new_ctx
) {
607 AioCoRescheduleSelf data
= {
608 .co
= qemu_coroutine_self(),
612 * We can't directly schedule the coroutine in the target context
613 * because this would be racy: The other thread could try to enter the
614 * coroutine before it has yielded in this one.
616 aio_bh_schedule_oneshot(old_ctx
, aio_co_reschedule_self_bh
, &data
);
617 qemu_coroutine_yield();
621 void aio_co_wake(struct Coroutine
*co
)
625 /* Read coroutine before co->ctx. Matches smp_wmb in
626 * qemu_coroutine_enter.
628 smp_read_barrier_depends();
629 ctx
= qatomic_read(&co
->ctx
);
631 aio_co_enter(ctx
, co
);
634 void aio_co_enter(AioContext
*ctx
, struct Coroutine
*co
)
636 if (ctx
!= qemu_get_current_aio_context()) {
637 aio_co_schedule(ctx
, co
);
641 if (qemu_in_coroutine()) {
642 Coroutine
*self
= qemu_coroutine_self();
644 QSIMPLEQ_INSERT_TAIL(&self
->co_queue_wakeup
, co
, co_queue_next
);
646 aio_context_acquire(ctx
);
647 qemu_aio_coroutine_enter(ctx
, co
);
648 aio_context_release(ctx
);
652 void aio_context_ref(AioContext
*ctx
)
654 g_source_ref(&ctx
->source
);
657 void aio_context_unref(AioContext
*ctx
)
659 g_source_unref(&ctx
->source
);
662 void aio_context_acquire(AioContext
*ctx
)
664 qemu_rec_mutex_lock(&ctx
->lock
);
667 void aio_context_release(AioContext
*ctx
)
669 qemu_rec_mutex_unlock(&ctx
->lock
);
672 static __thread AioContext
*my_aiocontext
;
674 AioContext
*qemu_get_current_aio_context(void)
677 return my_aiocontext
;
679 if (qemu_mutex_iothread_locked()) {
680 /* Possibly in a vCPU thread. */
681 return qemu_get_aio_context();
686 void qemu_set_current_aio_context(AioContext
*ctx
)
688 assert(!my_aiocontext
);