vdpa: Delete duplicated vdpa_feature_bits entry
[qemu.git] / util / async.c
blob63434ddae40ca65fe21718f37a44b4d45c38517f
1 /*
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
23 * THE SOFTWARE.
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"
35 #include "qemu/coroutine-tls.h"
36 #include "sysemu/cpu-timers.h"
37 #include "trace.h"
39 /***********************************************************/
40 /* bottom halves (can be seen as timers which expire ASAP) */
42 /* QEMUBH::flags values */
43 enum {
44 /* Already enqueued and waiting for aio_bh_poll() */
45 BH_PENDING = (1 << 0),
47 /* Invoke the callback */
48 BH_SCHEDULED = (1 << 1),
50 /* Delete without invoking callback */
51 BH_DELETED = (1 << 2),
53 /* Delete after invoking callback */
54 BH_ONESHOT = (1 << 3),
56 /* Schedule periodically when the event loop is idle */
57 BH_IDLE = (1 << 4),
60 struct QEMUBH {
61 AioContext *ctx;
62 const char *name;
63 QEMUBHFunc *cb;
64 void *opaque;
65 QSLIST_ENTRY(QEMUBH) next;
66 unsigned flags;
69 /* Called concurrently from any thread */
70 static void aio_bh_enqueue(QEMUBH *bh, unsigned new_flags)
72 AioContext *ctx = bh->ctx;
73 unsigned old_flags;
76 * The memory barrier implicit in qatomic_fetch_or makes sure that:
77 * 1. idle & any writes needed by the callback are done before the
78 * locations are read in the aio_bh_poll.
79 * 2. ctx is loaded before the callback has a chance to execute and bh
80 * could be freed.
82 old_flags = qatomic_fetch_or(&bh->flags, BH_PENDING | new_flags);
83 if (!(old_flags & BH_PENDING)) {
84 QSLIST_INSERT_HEAD_ATOMIC(&ctx->bh_list, bh, next);
87 aio_notify(ctx);
89 * Workaround for record/replay.
90 * vCPU execution should be suspended when new BH is set.
91 * This is needed to avoid guest timeouts caused
92 * by the long cycles of the execution.
94 icount_notify_exit();
97 /* Only called from aio_bh_poll() and aio_ctx_finalize() */
98 static QEMUBH *aio_bh_dequeue(BHList *head, unsigned *flags)
100 QEMUBH *bh = QSLIST_FIRST_RCU(head);
102 if (!bh) {
103 return NULL;
106 QSLIST_REMOVE_HEAD(head, next);
109 * The qatomic_and is paired with aio_bh_enqueue(). The implicit memory
110 * barrier ensures that the callback sees all writes done by the scheduling
111 * thread. It also ensures that the scheduling thread sees the cleared
112 * flag before bh->cb has run, and thus will call aio_notify again if
113 * necessary.
115 *flags = qatomic_fetch_and(&bh->flags,
116 ~(BH_PENDING | BH_SCHEDULED | BH_IDLE));
117 return bh;
120 void aio_bh_schedule_oneshot_full(AioContext *ctx, QEMUBHFunc *cb,
121 void *opaque, const char *name)
123 QEMUBH *bh;
124 bh = g_new(QEMUBH, 1);
125 *bh = (QEMUBH){
126 .ctx = ctx,
127 .cb = cb,
128 .opaque = opaque,
129 .name = name,
131 aio_bh_enqueue(bh, BH_SCHEDULED | BH_ONESHOT);
134 QEMUBH *aio_bh_new_full(AioContext *ctx, QEMUBHFunc *cb, void *opaque,
135 const char *name)
137 QEMUBH *bh;
138 bh = g_new(QEMUBH, 1);
139 *bh = (QEMUBH){
140 .ctx = ctx,
141 .cb = cb,
142 .opaque = opaque,
143 .name = name,
145 return bh;
148 void aio_bh_call(QEMUBH *bh)
150 bh->cb(bh->opaque);
153 /* Multiple occurrences of aio_bh_poll cannot be called concurrently. */
154 int aio_bh_poll(AioContext *ctx)
156 BHListSlice slice;
157 BHListSlice *s;
158 int ret = 0;
160 QSLIST_MOVE_ATOMIC(&slice.bh_list, &ctx->bh_list);
161 QSIMPLEQ_INSERT_TAIL(&ctx->bh_slice_list, &slice, next);
163 while ((s = QSIMPLEQ_FIRST(&ctx->bh_slice_list))) {
164 QEMUBH *bh;
165 unsigned flags;
167 bh = aio_bh_dequeue(&s->bh_list, &flags);
168 if (!bh) {
169 QSIMPLEQ_REMOVE_HEAD(&ctx->bh_slice_list, next);
170 continue;
173 if ((flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
174 /* Idle BHs don't count as progress */
175 if (!(flags & BH_IDLE)) {
176 ret = 1;
178 aio_bh_call(bh);
180 if (flags & (BH_DELETED | BH_ONESHOT)) {
181 g_free(bh);
185 return ret;
188 void qemu_bh_schedule_idle(QEMUBH *bh)
190 aio_bh_enqueue(bh, BH_SCHEDULED | BH_IDLE);
193 void qemu_bh_schedule(QEMUBH *bh)
195 aio_bh_enqueue(bh, BH_SCHEDULED);
198 /* This func is async.
200 void qemu_bh_cancel(QEMUBH *bh)
202 qatomic_and(&bh->flags, ~BH_SCHEDULED);
205 /* This func is async.The bottom half will do the delete action at the finial
206 * end.
208 void qemu_bh_delete(QEMUBH *bh)
210 aio_bh_enqueue(bh, BH_DELETED);
213 static int64_t aio_compute_bh_timeout(BHList *head, int timeout)
215 QEMUBH *bh;
217 QSLIST_FOREACH_RCU(bh, head, next) {
218 if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
219 if (bh->flags & BH_IDLE) {
220 /* idle bottom halves will be polled at least
221 * every 10ms */
222 timeout = 10000000;
223 } else {
224 /* non-idle bottom halves will be executed
225 * immediately */
226 return 0;
231 return timeout;
234 int64_t
235 aio_compute_timeout(AioContext *ctx)
237 BHListSlice *s;
238 int64_t deadline;
239 int timeout = -1;
241 timeout = aio_compute_bh_timeout(&ctx->bh_list, timeout);
242 if (timeout == 0) {
243 return 0;
246 QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
247 timeout = aio_compute_bh_timeout(&s->bh_list, timeout);
248 if (timeout == 0) {
249 return 0;
253 deadline = timerlistgroup_deadline_ns(&ctx->tlg);
254 if (deadline == 0) {
255 return 0;
256 } else {
257 return qemu_soonest_timeout(timeout, deadline);
261 static gboolean
262 aio_ctx_prepare(GSource *source, gint *timeout)
264 AioContext *ctx = (AioContext *) source;
266 qatomic_set(&ctx->notify_me, qatomic_read(&ctx->notify_me) | 1);
269 * Write ctx->notify_me before computing the timeout
270 * (reading bottom half flags, etc.). Pairs with
271 * smp_mb in aio_notify().
273 smp_mb();
275 /* We assume there is no timeout already supplied */
276 *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
278 if (aio_prepare(ctx)) {
279 *timeout = 0;
282 return *timeout == 0;
285 static gboolean
286 aio_ctx_check(GSource *source)
288 AioContext *ctx = (AioContext *) source;
289 QEMUBH *bh;
290 BHListSlice *s;
292 /* Finish computing the timeout before clearing the flag. */
293 qatomic_store_release(&ctx->notify_me, qatomic_read(&ctx->notify_me) & ~1);
294 aio_notify_accept(ctx);
296 QSLIST_FOREACH_RCU(bh, &ctx->bh_list, next) {
297 if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
298 return true;
302 QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
303 QSLIST_FOREACH_RCU(bh, &s->bh_list, next) {
304 if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
305 return true;
309 return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
312 static gboolean
313 aio_ctx_dispatch(GSource *source,
314 GSourceFunc callback,
315 gpointer user_data)
317 AioContext *ctx = (AioContext *) source;
319 assert(callback == NULL);
320 aio_dispatch(ctx);
321 return true;
324 static void
325 aio_ctx_finalize(GSource *source)
327 AioContext *ctx = (AioContext *) source;
328 QEMUBH *bh;
329 unsigned flags;
331 thread_pool_free(ctx->thread_pool);
333 #ifdef CONFIG_LINUX_AIO
334 if (ctx->linux_aio) {
335 laio_detach_aio_context(ctx->linux_aio, ctx);
336 laio_cleanup(ctx->linux_aio);
337 ctx->linux_aio = NULL;
339 #endif
341 #ifdef CONFIG_LINUX_IO_URING
342 if (ctx->linux_io_uring) {
343 luring_detach_aio_context(ctx->linux_io_uring, ctx);
344 luring_cleanup(ctx->linux_io_uring);
345 ctx->linux_io_uring = NULL;
347 #endif
349 assert(QSLIST_EMPTY(&ctx->scheduled_coroutines));
350 qemu_bh_delete(ctx->co_schedule_bh);
352 /* There must be no aio_bh_poll() calls going on */
353 assert(QSIMPLEQ_EMPTY(&ctx->bh_slice_list));
355 while ((bh = aio_bh_dequeue(&ctx->bh_list, &flags))) {
357 * qemu_bh_delete() must have been called on BHs in this AioContext. In
358 * many cases memory leaks, hangs, or inconsistent state occur when a
359 * BH is leaked because something still expects it to run.
361 * If you hit this, fix the lifecycle of the BH so that
362 * qemu_bh_delete() and any associated cleanup is called before the
363 * AioContext is finalized.
365 if (unlikely(!(flags & BH_DELETED))) {
366 fprintf(stderr, "%s: BH '%s' leaked, aborting...\n",
367 __func__, bh->name);
368 abort();
371 g_free(bh);
374 aio_set_event_notifier(ctx, &ctx->notifier, false, NULL, NULL, NULL);
375 event_notifier_cleanup(&ctx->notifier);
376 qemu_rec_mutex_destroy(&ctx->lock);
377 qemu_lockcnt_destroy(&ctx->list_lock);
378 timerlistgroup_deinit(&ctx->tlg);
379 aio_context_destroy(ctx);
382 static GSourceFuncs aio_source_funcs = {
383 aio_ctx_prepare,
384 aio_ctx_check,
385 aio_ctx_dispatch,
386 aio_ctx_finalize
389 GSource *aio_get_g_source(AioContext *ctx)
391 aio_context_use_g_source(ctx);
392 g_source_ref(&ctx->source);
393 return &ctx->source;
396 ThreadPool *aio_get_thread_pool(AioContext *ctx)
398 if (!ctx->thread_pool) {
399 ctx->thread_pool = thread_pool_new(ctx);
401 return ctx->thread_pool;
404 #ifdef CONFIG_LINUX_AIO
405 LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp)
407 if (!ctx->linux_aio) {
408 ctx->linux_aio = laio_init(errp);
409 if (ctx->linux_aio) {
410 laio_attach_aio_context(ctx->linux_aio, ctx);
413 return ctx->linux_aio;
416 LinuxAioState *aio_get_linux_aio(AioContext *ctx)
418 assert(ctx->linux_aio);
419 return ctx->linux_aio;
421 #endif
423 #ifdef CONFIG_LINUX_IO_URING
424 LuringState *aio_setup_linux_io_uring(AioContext *ctx, Error **errp)
426 if (ctx->linux_io_uring) {
427 return ctx->linux_io_uring;
430 ctx->linux_io_uring = luring_init(errp);
431 if (!ctx->linux_io_uring) {
432 return NULL;
435 luring_attach_aio_context(ctx->linux_io_uring, ctx);
436 return ctx->linux_io_uring;
439 LuringState *aio_get_linux_io_uring(AioContext *ctx)
441 assert(ctx->linux_io_uring);
442 return ctx->linux_io_uring;
444 #endif
446 void aio_notify(AioContext *ctx)
449 * Write e.g. bh->flags before writing ctx->notified. Pairs with smp_mb in
450 * aio_notify_accept.
452 smp_wmb();
453 qatomic_set(&ctx->notified, true);
456 * Write ctx->notified before reading ctx->notify_me. Pairs
457 * with smp_mb in aio_ctx_prepare or aio_poll.
459 smp_mb();
460 if (qatomic_read(&ctx->notify_me)) {
461 event_notifier_set(&ctx->notifier);
465 void aio_notify_accept(AioContext *ctx)
467 qatomic_set(&ctx->notified, false);
470 * Write ctx->notified before reading e.g. bh->flags. Pairs with smp_wmb
471 * in aio_notify.
473 smp_mb();
476 static void aio_timerlist_notify(void *opaque, QEMUClockType type)
478 aio_notify(opaque);
481 static void aio_context_notifier_cb(EventNotifier *e)
483 AioContext *ctx = container_of(e, AioContext, notifier);
485 event_notifier_test_and_clear(&ctx->notifier);
488 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
489 static bool aio_context_notifier_poll(void *opaque)
491 EventNotifier *e = opaque;
492 AioContext *ctx = container_of(e, AioContext, notifier);
494 return qatomic_read(&ctx->notified);
497 static void aio_context_notifier_poll_ready(EventNotifier *e)
499 /* Do nothing, we just wanted to kick the event loop */
502 static void co_schedule_bh_cb(void *opaque)
504 AioContext *ctx = opaque;
505 QSLIST_HEAD(, Coroutine) straight, reversed;
507 QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines);
508 QSLIST_INIT(&straight);
510 while (!QSLIST_EMPTY(&reversed)) {
511 Coroutine *co = QSLIST_FIRST(&reversed);
512 QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next);
513 QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next);
516 while (!QSLIST_EMPTY(&straight)) {
517 Coroutine *co = QSLIST_FIRST(&straight);
518 QSLIST_REMOVE_HEAD(&straight, co_scheduled_next);
519 trace_aio_co_schedule_bh_cb(ctx, co);
520 aio_context_acquire(ctx);
522 /* Protected by write barrier in qemu_aio_coroutine_enter */
523 qatomic_set(&co->scheduled, NULL);
524 qemu_aio_coroutine_enter(ctx, co);
525 aio_context_release(ctx);
529 AioContext *aio_context_new(Error **errp)
531 int ret;
532 AioContext *ctx;
534 ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
535 QSLIST_INIT(&ctx->bh_list);
536 QSIMPLEQ_INIT(&ctx->bh_slice_list);
537 aio_context_setup(ctx);
539 ret = event_notifier_init(&ctx->notifier, false);
540 if (ret < 0) {
541 error_setg_errno(errp, -ret, "Failed to initialize event notifier");
542 goto fail;
544 g_source_set_can_recurse(&ctx->source, true);
545 qemu_lockcnt_init(&ctx->list_lock);
547 ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx);
548 QSLIST_INIT(&ctx->scheduled_coroutines);
550 aio_set_event_notifier(ctx, &ctx->notifier,
551 false,
552 aio_context_notifier_cb,
553 aio_context_notifier_poll,
554 aio_context_notifier_poll_ready);
555 #ifdef CONFIG_LINUX_AIO
556 ctx->linux_aio = NULL;
557 #endif
559 #ifdef CONFIG_LINUX_IO_URING
560 ctx->linux_io_uring = NULL;
561 #endif
563 ctx->thread_pool = NULL;
564 qemu_rec_mutex_init(&ctx->lock);
565 timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
567 ctx->poll_ns = 0;
568 ctx->poll_max_ns = 0;
569 ctx->poll_grow = 0;
570 ctx->poll_shrink = 0;
572 ctx->aio_max_batch = 0;
574 ctx->thread_pool_min = 0;
575 ctx->thread_pool_max = THREAD_POOL_MAX_THREADS_DEFAULT;
577 return ctx;
578 fail:
579 g_source_destroy(&ctx->source);
580 return NULL;
583 void aio_co_schedule(AioContext *ctx, Coroutine *co)
585 trace_aio_co_schedule(ctx, co);
586 const char *scheduled = qatomic_cmpxchg(&co->scheduled, NULL,
587 __func__);
589 if (scheduled) {
590 fprintf(stderr,
591 "%s: Co-routine was already scheduled in '%s'\n",
592 __func__, scheduled);
593 abort();
596 /* The coroutine might run and release the last ctx reference before we
597 * invoke qemu_bh_schedule(). Take a reference to keep ctx alive until
598 * we're done.
600 aio_context_ref(ctx);
602 QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines,
603 co, co_scheduled_next);
604 qemu_bh_schedule(ctx->co_schedule_bh);
606 aio_context_unref(ctx);
609 typedef struct AioCoRescheduleSelf {
610 Coroutine *co;
611 AioContext *new_ctx;
612 } AioCoRescheduleSelf;
614 static void aio_co_reschedule_self_bh(void *opaque)
616 AioCoRescheduleSelf *data = opaque;
617 aio_co_schedule(data->new_ctx, data->co);
620 void coroutine_fn aio_co_reschedule_self(AioContext *new_ctx)
622 AioContext *old_ctx = qemu_get_current_aio_context();
624 if (old_ctx != new_ctx) {
625 AioCoRescheduleSelf data = {
626 .co = qemu_coroutine_self(),
627 .new_ctx = new_ctx,
630 * We can't directly schedule the coroutine in the target context
631 * because this would be racy: The other thread could try to enter the
632 * coroutine before it has yielded in this one.
634 aio_bh_schedule_oneshot(old_ctx, aio_co_reschedule_self_bh, &data);
635 qemu_coroutine_yield();
639 void aio_co_wake(struct Coroutine *co)
641 AioContext *ctx;
643 /* Read coroutine before co->ctx. Matches smp_wmb in
644 * qemu_coroutine_enter.
646 smp_read_barrier_depends();
647 ctx = qatomic_read(&co->ctx);
649 aio_co_enter(ctx, co);
652 void aio_co_enter(AioContext *ctx, struct Coroutine *co)
654 if (ctx != qemu_get_current_aio_context()) {
655 aio_co_schedule(ctx, co);
656 return;
659 if (qemu_in_coroutine()) {
660 Coroutine *self = qemu_coroutine_self();
661 assert(self != co);
662 QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next);
663 } else {
664 aio_context_acquire(ctx);
665 qemu_aio_coroutine_enter(ctx, co);
666 aio_context_release(ctx);
670 void aio_context_ref(AioContext *ctx)
672 g_source_ref(&ctx->source);
675 void aio_context_unref(AioContext *ctx)
677 g_source_unref(&ctx->source);
680 void aio_context_acquire(AioContext *ctx)
682 qemu_rec_mutex_lock(&ctx->lock);
685 void aio_context_release(AioContext *ctx)
687 qemu_rec_mutex_unlock(&ctx->lock);
690 QEMU_DEFINE_STATIC_CO_TLS(AioContext *, my_aiocontext)
692 AioContext *qemu_get_current_aio_context(void)
694 AioContext *ctx = get_my_aiocontext();
695 if (ctx) {
696 return ctx;
698 if (qemu_mutex_iothread_locked()) {
699 /* Possibly in a vCPU thread. */
700 return qemu_get_aio_context();
702 return NULL;
705 void qemu_set_current_aio_context(AioContext *ctx)
707 assert(!get_my_aiocontext());
708 set_my_aiocontext(ctx);
711 void aio_context_set_thread_pool_params(AioContext *ctx, int64_t min,
712 int64_t max, Error **errp)
715 if (min > max || !max || min > INT_MAX || max > INT_MAX) {
716 error_setg(errp, "bad thread-pool-min/thread-pool-max values");
717 return;
720 ctx->thread_pool_min = min;
721 ctx->thread_pool_max = max;
723 if (ctx->thread_pool) {
724 thread_pool_update_params(ctx->thread_pool, ctx);