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 */
62 QSLIST_ENTRY(QEMUBH) next;
66 /* Called concurrently from any thread */
67 static void aio_bh_enqueue(QEMUBH *bh, unsigned new_flags)
69 AioContext *ctx = bh->ctx;
73 * The memory barrier implicit in atomic_fetch_or makes sure that:
74 * 1. idle & any writes needed by the callback are done before the
75 * locations are read in the aio_bh_poll.
76 * 2. ctx is loaded before the callback has a chance to execute and bh
79 old_flags = atomic_fetch_or(&bh->flags, BH_PENDING | new_flags);
80 if (!(old_flags & BH_PENDING)) {
81 QSLIST_INSERT_HEAD_ATOMIC(&ctx->bh_list, bh, next);
87 /* Only called from aio_bh_poll() and aio_ctx_finalize() */
88 static QEMUBH *aio_bh_dequeue(BHList *head, unsigned *flags)
90 QEMUBH *bh = QSLIST_FIRST_RCU(head);
96 QSLIST_REMOVE_HEAD(head, next);
99 * The atomic_and is paired with aio_bh_enqueue(). The implicit memory
100 * barrier ensures that the callback sees all writes done by the scheduling
101 * thread. It also ensures that the scheduling thread sees the cleared
102 * flag before bh->cb has run, and thus will call aio_notify again if
105 *flags = atomic_fetch_and(&bh->flags,
106 ~(BH_PENDING | BH_SCHEDULED | BH_IDLE));
110 void aio_bh_schedule_oneshot(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
113 bh = g_new(QEMUBH, 1);
119 aio_bh_enqueue(bh, BH_SCHEDULED | BH_ONESHOT);
122 QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
125 bh = g_new(QEMUBH, 1);
134 void aio_bh_call(QEMUBH *bh)
139 /* Multiple occurrences of aio_bh_poll cannot be called concurrently. */
140 int aio_bh_poll(AioContext *ctx)
146 QSLIST_MOVE_ATOMIC(&slice.bh_list, &ctx->bh_list);
147 QSIMPLEQ_INSERT_TAIL(&ctx->bh_slice_list, &slice, next);
149 while ((s = QSIMPLEQ_FIRST(&ctx->bh_slice_list))) {
153 bh = aio_bh_dequeue(&s->bh_list, &flags);
155 QSIMPLEQ_REMOVE_HEAD(&ctx->bh_slice_list, next);
159 if ((flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
160 /* Idle BHs don't count as progress */
161 if (!(flags & BH_IDLE)) {
166 if (flags & (BH_DELETED | BH_ONESHOT)) {
174 void qemu_bh_schedule_idle(QEMUBH *bh)
176 aio_bh_enqueue(bh, BH_SCHEDULED | BH_IDLE);
179 void qemu_bh_schedule(QEMUBH *bh)
181 aio_bh_enqueue(bh, BH_SCHEDULED);
184 /* This func is async.
186 void qemu_bh_cancel(QEMUBH *bh)
188 atomic_and(&bh->flags, ~BH_SCHEDULED);
191 /* This func is async.The bottom half will do the delete action at the finial
194 void qemu_bh_delete(QEMUBH *bh)
196 aio_bh_enqueue(bh, BH_DELETED);
199 static int64_t aio_compute_bh_timeout(BHList *head, int timeout)
203 QSLIST_FOREACH_RCU(bh, head, next) {
204 if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
205 if (bh->flags & BH_IDLE) {
206 /* idle bottom halves will be polled at least
210 /* non-idle bottom halves will be executed
221 aio_compute_timeout(AioContext *ctx)
227 timeout = aio_compute_bh_timeout(&ctx->bh_list, timeout);
232 QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
233 timeout = aio_compute_bh_timeout(&s->bh_list, timeout);
239 deadline = timerlistgroup_deadline_ns(&ctx->tlg);
243 return qemu_soonest_timeout(timeout, deadline);
248 aio_ctx_prepare(GSource *source, gint *timeout)
250 AioContext *ctx = (AioContext *) source;
252 atomic_set(&ctx->notify_me, atomic_read(&ctx->notify_me) | 1);
255 * Write ctx->notify_me before computing the timeout
256 * (reading bottom half flags, etc.). Pairs with
257 * smp_mb in aio_notify().
261 /* We assume there is no timeout already supplied */
262 *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
264 if (aio_prepare(ctx)) {
268 return *timeout == 0;
272 aio_ctx_check(GSource *source)
274 AioContext *ctx = (AioContext *) source;
278 /* Finish computing the timeout before clearing the flag. */
279 atomic_store_release(&ctx->notify_me, atomic_read(&ctx->notify_me) & ~1);
280 aio_notify_accept(ctx);
282 QSLIST_FOREACH_RCU(bh, &ctx->bh_list, next) {
283 if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
288 QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
289 QSLIST_FOREACH_RCU(bh, &s->bh_list, next) {
290 if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
295 return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
299 aio_ctx_dispatch(GSource *source,
300 GSourceFunc callback,
303 AioContext *ctx = (AioContext *) source;
305 assert(callback == NULL);
311 aio_ctx_finalize(GSource *source)
313 AioContext *ctx = (AioContext *) source;
317 thread_pool_free(ctx->thread_pool);
319 #ifdef CONFIG_LINUX_AIO
320 if (ctx->linux_aio) {
321 laio_detach_aio_context(ctx->linux_aio, ctx);
322 laio_cleanup(ctx->linux_aio);
323 ctx->linux_aio = NULL;
327 #ifdef CONFIG_LINUX_IO_URING
328 if (ctx->linux_io_uring) {
329 luring_detach_aio_context(ctx->linux_io_uring, ctx);
330 luring_cleanup(ctx->linux_io_uring);
331 ctx->linux_io_uring = NULL;
335 assert(QSLIST_EMPTY(&ctx->scheduled_coroutines));
336 qemu_bh_delete(ctx->co_schedule_bh);
338 /* There must be no aio_bh_poll() calls going on */
339 assert(QSIMPLEQ_EMPTY(&ctx->bh_slice_list));
341 while ((bh = aio_bh_dequeue(&ctx->bh_list, &flags))) {
342 /* qemu_bh_delete() must have been called on BHs in this AioContext */
343 assert(flags & BH_DELETED);
348 aio_set_event_notifier(ctx, &ctx->notifier, false, NULL, NULL);
349 event_notifier_cleanup(&ctx->notifier);
350 qemu_rec_mutex_destroy(&ctx->lock);
351 qemu_lockcnt_destroy(&ctx->list_lock);
352 timerlistgroup_deinit(&ctx->tlg);
353 aio_context_destroy(ctx);
356 static GSourceFuncs aio_source_funcs = {
363 GSource *aio_get_g_source(AioContext *ctx)
365 g_source_ref(&ctx->source);
369 ThreadPool *aio_get_thread_pool(AioContext *ctx)
371 if (!ctx->thread_pool) {
372 ctx->thread_pool = thread_pool_new(ctx);
374 return ctx->thread_pool;
377 #ifdef CONFIG_LINUX_AIO
378 LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp)
380 if (!ctx->linux_aio) {
381 ctx->linux_aio = laio_init(errp);
382 if (ctx->linux_aio) {
383 laio_attach_aio_context(ctx->linux_aio, ctx);
386 return ctx->linux_aio;
389 LinuxAioState *aio_get_linux_aio(AioContext *ctx)
391 assert(ctx->linux_aio);
392 return ctx->linux_aio;
396 #ifdef CONFIG_LINUX_IO_URING
397 LuringState *aio_setup_linux_io_uring(AioContext *ctx, Error **errp)
399 if (ctx->linux_io_uring) {
400 return ctx->linux_io_uring;
403 ctx->linux_io_uring = luring_init(errp);
404 if (!ctx->linux_io_uring) {
408 luring_attach_aio_context(ctx->linux_io_uring, ctx);
409 return ctx->linux_io_uring;
412 LuringState *aio_get_linux_io_uring(AioContext *ctx)
414 assert(ctx->linux_io_uring);
415 return ctx->linux_io_uring;
419 void aio_notify(AioContext *ctx)
421 /* Write e.g. bh->scheduled before reading ctx->notify_me. Pairs
422 * with smp_mb in aio_ctx_prepare or aio_poll.
425 if (atomic_read(&ctx->notify_me)) {
426 event_notifier_set(&ctx->notifier);
427 atomic_mb_set(&ctx->notified, true);
431 void aio_notify_accept(AioContext *ctx)
433 if (atomic_xchg(&ctx->notified, false)
438 event_notifier_test_and_clear(&ctx->notifier);
442 static void aio_timerlist_notify(void *opaque, QEMUClockType type)
447 static void event_notifier_dummy_cb(EventNotifier *e)
451 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
452 static bool event_notifier_poll(void *opaque)
454 EventNotifier *e = opaque;
455 AioContext *ctx = container_of(e, AioContext, notifier);
457 return atomic_read(&ctx->notified);
460 static void co_schedule_bh_cb(void *opaque)
462 AioContext *ctx = opaque;
463 QSLIST_HEAD(, Coroutine) straight, reversed;
465 QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines);
466 QSLIST_INIT(&straight);
468 while (!QSLIST_EMPTY(&reversed)) {
469 Coroutine *co = QSLIST_FIRST(&reversed);
470 QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next);
471 QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next);
474 while (!QSLIST_EMPTY(&straight)) {
475 Coroutine *co = QSLIST_FIRST(&straight);
476 QSLIST_REMOVE_HEAD(&straight, co_scheduled_next);
477 trace_aio_co_schedule_bh_cb(ctx, co);
478 aio_context_acquire(ctx);
480 /* Protected by write barrier in qemu_aio_coroutine_enter */
481 atomic_set(&co->scheduled, NULL);
482 qemu_aio_coroutine_enter(ctx, co);
483 aio_context_release(ctx);
487 AioContext *aio_context_new(Error **errp)
492 ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
493 QSLIST_INIT(&ctx->bh_list);
494 QSIMPLEQ_INIT(&ctx->bh_slice_list);
495 aio_context_setup(ctx);
497 ret = event_notifier_init(&ctx->notifier, false);
499 error_setg_errno(errp, -ret, "Failed to initialize event notifier");
502 g_source_set_can_recurse(&ctx->source, true);
503 qemu_lockcnt_init(&ctx->list_lock);
505 ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx);
506 QSLIST_INIT(&ctx->scheduled_coroutines);
508 aio_set_event_notifier(ctx, &ctx->notifier,
510 event_notifier_dummy_cb,
511 event_notifier_poll);
512 #ifdef CONFIG_LINUX_AIO
513 ctx->linux_aio = NULL;
516 #ifdef CONFIG_LINUX_IO_URING
517 ctx->linux_io_uring = NULL;
520 ctx->thread_pool = NULL;
521 qemu_rec_mutex_init(&ctx->lock);
522 timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
525 ctx->poll_max_ns = 0;
527 ctx->poll_shrink = 0;
531 g_source_destroy(&ctx->source);
535 void aio_co_schedule(AioContext *ctx, Coroutine *co)
537 trace_aio_co_schedule(ctx, co);
538 const char *scheduled = atomic_cmpxchg(&co->scheduled, NULL,
543 "%s: Co-routine was already scheduled in '%s'\n",
544 __func__, scheduled);
548 /* The coroutine might run and release the last ctx reference before we
549 * invoke qemu_bh_schedule(). Take a reference to keep ctx alive until
552 aio_context_ref(ctx);
554 QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines,
555 co, co_scheduled_next);
556 qemu_bh_schedule(ctx->co_schedule_bh);
558 aio_context_unref(ctx);
561 void aio_co_wake(struct Coroutine *co)
565 /* Read coroutine before co->ctx. Matches smp_wmb in
566 * qemu_coroutine_enter.
568 smp_read_barrier_depends();
569 ctx = atomic_read(&co->ctx);
571 aio_co_enter(ctx, co);
574 void aio_co_enter(AioContext *ctx, struct Coroutine *co)
576 if (ctx != qemu_get_current_aio_context()) {
577 aio_co_schedule(ctx, co);
581 if (qemu_in_coroutine()) {
582 Coroutine *self = qemu_coroutine_self();
584 QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next);
586 aio_context_acquire(ctx);
587 qemu_aio_coroutine_enter(ctx, co);
588 aio_context_release(ctx);
592 void aio_context_ref(AioContext *ctx)
594 g_source_ref(&ctx->source);
597 void aio_context_unref(AioContext *ctx)
599 g_source_unref(&ctx->source);
602 void aio_context_acquire(AioContext *ctx)
604 qemu_rec_mutex_lock(&ctx->lock);
607 void aio_context_release(AioContext *ctx)
609 qemu_rec_mutex_unlock(&ctx->lock);