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)
177 atomic_mb_set(&bh->scheduled, 0);
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);
301 aio_context_destroy(ctx);
304 static GSourceFuncs aio_source_funcs = {
311 GSource *aio_get_g_source(AioContext *ctx)
313 g_source_ref(&ctx->source);
317 ThreadPool *aio_get_thread_pool(AioContext *ctx)
319 if (!ctx->thread_pool) {
320 ctx->thread_pool = thread_pool_new(ctx);
322 return ctx->thread_pool;
325 #ifdef CONFIG_LINUX_AIO
326 LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp)
328 if (!ctx->linux_aio) {
329 ctx->linux_aio = laio_init(errp);
330 if (ctx->linux_aio) {
331 laio_attach_aio_context(ctx->linux_aio, ctx);
334 return ctx->linux_aio;
337 LinuxAioState *aio_get_linux_aio(AioContext *ctx)
339 assert(ctx->linux_aio);
340 return ctx->linux_aio;
344 void aio_notify(AioContext *ctx)
346 /* Write e.g. bh->scheduled before reading ctx->notify_me. Pairs
347 * with atomic_or in aio_ctx_prepare or atomic_add in aio_poll.
350 if (ctx->notify_me) {
351 event_notifier_set(&ctx->notifier);
352 atomic_mb_set(&ctx->notified, true);
356 void aio_notify_accept(AioContext *ctx)
358 if (atomic_xchg(&ctx->notified, false)) {
359 event_notifier_test_and_clear(&ctx->notifier);
363 static void aio_timerlist_notify(void *opaque, QEMUClockType type)
368 static void event_notifier_dummy_cb(EventNotifier *e)
372 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
373 static bool event_notifier_poll(void *opaque)
375 EventNotifier *e = opaque;
376 AioContext *ctx = container_of(e, AioContext, notifier);
378 return atomic_read(&ctx->notified);
381 static void co_schedule_bh_cb(void *opaque)
383 AioContext *ctx = opaque;
384 QSLIST_HEAD(, Coroutine) straight, reversed;
386 QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines);
387 QSLIST_INIT(&straight);
389 while (!QSLIST_EMPTY(&reversed)) {
390 Coroutine *co = QSLIST_FIRST(&reversed);
391 QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next);
392 QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next);
395 while (!QSLIST_EMPTY(&straight)) {
396 Coroutine *co = QSLIST_FIRST(&straight);
397 QSLIST_REMOVE_HEAD(&straight, co_scheduled_next);
398 trace_aio_co_schedule_bh_cb(ctx, co);
399 aio_context_acquire(ctx);
401 /* Protected by write barrier in qemu_aio_coroutine_enter */
402 atomic_set(&co->scheduled, NULL);
403 qemu_aio_coroutine_enter(ctx, co);
404 aio_context_release(ctx);
408 AioContext *aio_context_new(Error **errp)
413 ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
414 aio_context_setup(ctx);
416 ret = event_notifier_init(&ctx->notifier, false);
418 error_setg_errno(errp, -ret, "Failed to initialize event notifier");
421 g_source_set_can_recurse(&ctx->source, true);
422 qemu_lockcnt_init(&ctx->list_lock);
424 ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx);
425 QSLIST_INIT(&ctx->scheduled_coroutines);
427 aio_set_event_notifier(ctx, &ctx->notifier,
429 (EventNotifierHandler *)
430 event_notifier_dummy_cb,
431 event_notifier_poll);
432 #ifdef CONFIG_LINUX_AIO
433 ctx->linux_aio = NULL;
435 ctx->thread_pool = NULL;
436 qemu_rec_mutex_init(&ctx->lock);
437 timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
440 ctx->poll_max_ns = 0;
442 ctx->poll_shrink = 0;
446 g_source_destroy(&ctx->source);
450 void aio_co_schedule(AioContext *ctx, Coroutine *co)
452 trace_aio_co_schedule(ctx, co);
453 const char *scheduled = atomic_cmpxchg(&co->scheduled, NULL,
458 "%s: Co-routine was already scheduled in '%s'\n",
459 __func__, scheduled);
463 QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines,
464 co, co_scheduled_next);
465 qemu_bh_schedule(ctx->co_schedule_bh);
468 void aio_co_wake(struct Coroutine *co)
472 /* Read coroutine before co->ctx. Matches smp_wmb in
473 * qemu_coroutine_enter.
475 smp_read_barrier_depends();
476 ctx = atomic_read(&co->ctx);
478 aio_co_enter(ctx, co);
481 void aio_co_enter(AioContext *ctx, struct Coroutine *co)
483 if (ctx != qemu_get_current_aio_context()) {
484 aio_co_schedule(ctx, co);
488 if (qemu_in_coroutine()) {
489 Coroutine *self = qemu_coroutine_self();
491 QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next);
493 aio_context_acquire(ctx);
494 qemu_aio_coroutine_enter(ctx, co);
495 aio_context_release(ctx);
499 void aio_context_ref(AioContext *ctx)
501 g_source_ref(&ctx->source);
504 void aio_context_unref(AioContext *ctx)
506 g_source_unref(&ctx->source);
509 void aio_context_acquire(AioContext *ctx)
511 qemu_rec_mutex_lock(&ctx->lock);
514 void aio_context_release(AioContext *ctx)
516 qemu_rec_mutex_unlock(&ctx->lock);