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 "qemu-common.h"
29 #include "block/aio.h"
30 #include "block/thread-pool.h"
31 #include "qemu/main-loop.h"
32 #include "qemu/atomic.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) */
50 void aio_bh_schedule_oneshot(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
53 bh = g_new(QEMUBH, 1);
59 qemu_lockcnt_lock(&ctx->list_lock);
60 bh->next = ctx->first_bh;
63 /* Make sure that the members are ready before putting bh into list */
66 qemu_lockcnt_unlock(&ctx->list_lock);
70 QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
73 bh = g_new(QEMUBH, 1);
79 qemu_lockcnt_lock(&ctx->list_lock);
80 bh->next = ctx->first_bh;
81 /* Make sure that the members are ready before putting bh into list */
84 qemu_lockcnt_unlock(&ctx->list_lock);
88 void aio_bh_call(QEMUBH *bh)
93 /* Multiple occurrences of aio_bh_poll cannot be called concurrently.
94 * The count in ctx->list_lock is incremented before the call, and is
95 * not affected by the call.
97 int aio_bh_poll(AioContext *ctx)
99 QEMUBH *bh, **bhp, *next;
101 bool deleted = false;
104 for (bh = atomic_rcu_read(&ctx->first_bh); bh; bh = next) {
105 next = atomic_rcu_read(&bh->next);
106 /* The atomic_xchg is paired with the one in qemu_bh_schedule. The
107 * implicit memory barrier ensures that the callback sees all writes
108 * done by the scheduling thread. It also ensures that the scheduling
109 * thread sees the zero before bh->cb has run, and thus will call
110 * aio_notify again if necessary.
112 if (atomic_xchg(&bh->scheduled, 0)) {
113 /* Idle BHs don't count as progress */
125 /* remove deleted bhs */
130 if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
131 bhp = &ctx->first_bh;
134 if (bh->deleted && !bh->scheduled) {
141 qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
146 void qemu_bh_schedule_idle(QEMUBH *bh)
149 /* Make sure that idle & any writes needed by the callback are done
150 * before the locations are read in the aio_bh_poll.
152 atomic_mb_set(&bh->scheduled, 1);
155 void qemu_bh_schedule(QEMUBH *bh)
161 /* The memory barrier implicit in atomic_xchg makes sure that:
162 * 1. idle & any writes needed by the callback are done before the
163 * locations are read in the aio_bh_poll.
164 * 2. ctx is loaded before scheduled is set and the callback has a chance
167 if (atomic_xchg(&bh->scheduled, 1) == 0) {
173 /* This func is async.
175 void qemu_bh_cancel(QEMUBH *bh)
180 /* This func is async.The bottom half will do the delete action at the finial
183 void qemu_bh_delete(QEMUBH *bh)
190 aio_compute_timeout(AioContext *ctx)
196 for (bh = atomic_rcu_read(&ctx->first_bh); bh;
197 bh = atomic_rcu_read(&bh->next)) {
200 /* idle bottom halves will be polled at least
204 /* non-idle bottom halves will be executed
211 deadline = timerlistgroup_deadline_ns(&ctx->tlg);
215 return qemu_soonest_timeout(timeout, deadline);
220 aio_ctx_prepare(GSource *source, gint *timeout)
222 AioContext *ctx = (AioContext *) source;
224 atomic_or(&ctx->notify_me, 1);
226 /* We assume there is no timeout already supplied */
227 *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
229 if (aio_prepare(ctx)) {
233 return *timeout == 0;
237 aio_ctx_check(GSource *source)
239 AioContext *ctx = (AioContext *) source;
242 atomic_and(&ctx->notify_me, ~1);
243 aio_notify_accept(ctx);
245 for (bh = ctx->first_bh; bh; bh = bh->next) {
250 return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
254 aio_ctx_dispatch(GSource *source,
255 GSourceFunc callback,
258 AioContext *ctx = (AioContext *) source;
260 assert(callback == NULL);
266 aio_ctx_finalize(GSource *source)
268 AioContext *ctx = (AioContext *) source;
270 thread_pool_free(ctx->thread_pool);
272 #ifdef CONFIG_LINUX_AIO
273 if (ctx->linux_aio) {
274 laio_detach_aio_context(ctx->linux_aio, ctx);
275 laio_cleanup(ctx->linux_aio);
276 ctx->linux_aio = NULL;
280 assert(QSLIST_EMPTY(&ctx->scheduled_coroutines));
281 qemu_bh_delete(ctx->co_schedule_bh);
283 qemu_lockcnt_lock(&ctx->list_lock);
284 assert(!qemu_lockcnt_count(&ctx->list_lock));
285 while (ctx->first_bh) {
286 QEMUBH *next = ctx->first_bh->next;
288 /* qemu_bh_delete() must have been called on BHs in this AioContext */
289 assert(ctx->first_bh->deleted);
291 g_free(ctx->first_bh);
292 ctx->first_bh = next;
294 qemu_lockcnt_unlock(&ctx->list_lock);
296 aio_set_event_notifier(ctx, &ctx->notifier, false, NULL, NULL);
297 event_notifier_cleanup(&ctx->notifier);
298 qemu_rec_mutex_destroy(&ctx->lock);
299 qemu_lockcnt_destroy(&ctx->list_lock);
300 timerlistgroup_deinit(&ctx->tlg);
303 static GSourceFuncs aio_source_funcs = {
310 GSource *aio_get_g_source(AioContext *ctx)
312 g_source_ref(&ctx->source);
316 ThreadPool *aio_get_thread_pool(AioContext *ctx)
318 if (!ctx->thread_pool) {
319 ctx->thread_pool = thread_pool_new(ctx);
321 return ctx->thread_pool;
324 #ifdef CONFIG_LINUX_AIO
325 LinuxAioState *aio_get_linux_aio(AioContext *ctx)
327 if (!ctx->linux_aio) {
328 ctx->linux_aio = laio_init();
329 laio_attach_aio_context(ctx->linux_aio, ctx);
331 return ctx->linux_aio;
335 void aio_notify(AioContext *ctx)
337 /* Write e.g. bh->scheduled before reading ctx->notify_me. Pairs
338 * with atomic_or in aio_ctx_prepare or atomic_add in aio_poll.
341 if (ctx->notify_me) {
342 event_notifier_set(&ctx->notifier);
343 atomic_mb_set(&ctx->notified, true);
347 void aio_notify_accept(AioContext *ctx)
349 if (atomic_xchg(&ctx->notified, false)) {
350 event_notifier_test_and_clear(&ctx->notifier);
354 static void aio_timerlist_notify(void *opaque, QEMUClockType type)
359 static void event_notifier_dummy_cb(EventNotifier *e)
363 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
364 static bool event_notifier_poll(void *opaque)
366 EventNotifier *e = opaque;
367 AioContext *ctx = container_of(e, AioContext, notifier);
369 return atomic_read(&ctx->notified);
372 static void co_schedule_bh_cb(void *opaque)
374 AioContext *ctx = opaque;
375 QSLIST_HEAD(, Coroutine) straight, reversed;
377 QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines);
378 QSLIST_INIT(&straight);
380 while (!QSLIST_EMPTY(&reversed)) {
381 Coroutine *co = QSLIST_FIRST(&reversed);
382 QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next);
383 QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next);
386 while (!QSLIST_EMPTY(&straight)) {
387 Coroutine *co = QSLIST_FIRST(&straight);
388 QSLIST_REMOVE_HEAD(&straight, co_scheduled_next);
389 trace_aio_co_schedule_bh_cb(ctx, co);
390 aio_context_acquire(ctx);
391 qemu_coroutine_enter(co);
392 aio_context_release(ctx);
396 AioContext *aio_context_new(Error **errp)
401 ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
402 aio_context_setup(ctx);
404 ret = event_notifier_init(&ctx->notifier, false);
406 error_setg_errno(errp, -ret, "Failed to initialize event notifier");
409 g_source_set_can_recurse(&ctx->source, true);
410 qemu_lockcnt_init(&ctx->list_lock);
412 ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx);
413 QSLIST_INIT(&ctx->scheduled_coroutines);
415 aio_set_event_notifier(ctx, &ctx->notifier,
417 (EventNotifierHandler *)
418 event_notifier_dummy_cb,
419 event_notifier_poll);
420 #ifdef CONFIG_LINUX_AIO
421 ctx->linux_aio = NULL;
423 ctx->thread_pool = NULL;
424 qemu_rec_mutex_init(&ctx->lock);
425 timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
428 ctx->poll_max_ns = 0;
430 ctx->poll_shrink = 0;
434 g_source_destroy(&ctx->source);
438 void aio_co_schedule(AioContext *ctx, Coroutine *co)
440 trace_aio_co_schedule(ctx, co);
441 QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines,
442 co, co_scheduled_next);
443 qemu_bh_schedule(ctx->co_schedule_bh);
446 void aio_co_wake(struct Coroutine *co)
450 /* Read coroutine before co->ctx. Matches smp_wmb in
451 * qemu_coroutine_enter.
453 smp_read_barrier_depends();
454 ctx = atomic_read(&co->ctx);
456 aio_co_enter(ctx, co);
459 void aio_co_enter(AioContext *ctx, struct Coroutine *co)
461 if (ctx != qemu_get_current_aio_context()) {
462 aio_co_schedule(ctx, co);
466 if (qemu_in_coroutine()) {
467 Coroutine *self = qemu_coroutine_self();
469 QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next);
471 aio_context_acquire(ctx);
472 qemu_aio_coroutine_enter(ctx, co);
473 aio_context_release(ctx);
477 void aio_context_ref(AioContext *ctx)
479 g_source_ref(&ctx->source);
482 void aio_context_unref(AioContext *ctx)
484 g_source_unref(&ctx->source);
487 void aio_context_acquire(AioContext *ctx)
489 qemu_rec_mutex_lock(&ctx->lock);
492 void aio_context_release(AioContext *ctx)
494 qemu_rec_mutex_unlock(&ctx->lock);