migration/rdma: Convert qemu_rdma_exchange_send() to Error
[qemu/armbru.git] / util / async.c
blob8f90ddc3047a9f0567da47ab6f89b6190e1d0f9d
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 "block/graph-lock.h"
31 #include "qemu/main-loop.h"
32 #include "qemu/atomic.h"
33 #include "qemu/rcu_queue.h"
34 #include "block/raw-aio.h"
35 #include "qemu/coroutine_int.h"
36 #include "qemu/coroutine-tls.h"
37 #include "sysemu/cpu-timers.h"
38 #include "trace.h"
40 /***********************************************************/
41 /* bottom halves (can be seen as timers which expire ASAP) */
43 /* QEMUBH::flags values */
44 enum {
45 /* Already enqueued and waiting for aio_bh_poll() */
46 BH_PENDING = (1 << 0),
48 /* Invoke the callback */
49 BH_SCHEDULED = (1 << 1),
51 /* Delete without invoking callback */
52 BH_DELETED = (1 << 2),
54 /* Delete after invoking callback */
55 BH_ONESHOT = (1 << 3),
57 /* Schedule periodically when the event loop is idle */
58 BH_IDLE = (1 << 4),
61 struct QEMUBH {
62 AioContext *ctx;
63 const char *name;
64 QEMUBHFunc *cb;
65 void *opaque;
66 QSLIST_ENTRY(QEMUBH) next;
67 unsigned flags;
68 MemReentrancyGuard *reentrancy_guard;
71 /* Called concurrently from any thread */
72 static void aio_bh_enqueue(QEMUBH *bh, unsigned new_flags)
74 AioContext *ctx = bh->ctx;
75 unsigned old_flags;
78 * Synchronizes with atomic_fetch_and() in aio_bh_dequeue(), ensuring that
79 * insertion starts after BH_PENDING is set.
81 old_flags = qatomic_fetch_or(&bh->flags, BH_PENDING | new_flags);
83 if (!(old_flags & BH_PENDING)) {
85 * At this point the bottom half becomes visible to aio_bh_poll().
86 * This insertion thus synchronizes with QSLIST_MOVE_ATOMIC in
87 * aio_bh_poll(), ensuring that:
88 * 1. any writes needed by the callback are visible from the callback
89 * after aio_bh_dequeue() returns bh.
90 * 2. ctx is loaded before the callback has a chance to execute and bh
91 * could be freed.
93 QSLIST_INSERT_HEAD_ATOMIC(&ctx->bh_list, bh, next);
96 aio_notify(ctx);
98 * Workaround for record/replay.
99 * vCPU execution should be suspended when new BH is set.
100 * This is needed to avoid guest timeouts caused
101 * by the long cycles of the execution.
103 icount_notify_exit();
106 /* Only called from aio_bh_poll() and aio_ctx_finalize() */
107 static QEMUBH *aio_bh_dequeue(BHList *head, unsigned *flags)
109 QEMUBH *bh = QSLIST_FIRST_RCU(head);
111 if (!bh) {
112 return NULL;
115 QSLIST_REMOVE_HEAD(head, next);
118 * Synchronizes with qatomic_fetch_or() in aio_bh_enqueue(), ensuring that
119 * the removal finishes before BH_PENDING is reset.
121 *flags = qatomic_fetch_and(&bh->flags,
122 ~(BH_PENDING | BH_SCHEDULED | BH_IDLE));
123 return bh;
126 void aio_bh_schedule_oneshot_full(AioContext *ctx, QEMUBHFunc *cb,
127 void *opaque, const char *name)
129 QEMUBH *bh;
130 bh = g_new(QEMUBH, 1);
131 *bh = (QEMUBH){
132 .ctx = ctx,
133 .cb = cb,
134 .opaque = opaque,
135 .name = name,
137 aio_bh_enqueue(bh, BH_SCHEDULED | BH_ONESHOT);
140 QEMUBH *aio_bh_new_full(AioContext *ctx, QEMUBHFunc *cb, void *opaque,
141 const char *name, MemReentrancyGuard *reentrancy_guard)
143 QEMUBH *bh;
144 bh = g_new(QEMUBH, 1);
145 *bh = (QEMUBH){
146 .ctx = ctx,
147 .cb = cb,
148 .opaque = opaque,
149 .name = name,
150 .reentrancy_guard = reentrancy_guard,
152 return bh;
155 void aio_bh_call(QEMUBH *bh)
157 bool last_engaged_in_io = false;
159 /* Make a copy of the guard-pointer as cb may free the bh */
160 MemReentrancyGuard *reentrancy_guard = bh->reentrancy_guard;
161 if (reentrancy_guard) {
162 last_engaged_in_io = reentrancy_guard->engaged_in_io;
163 if (reentrancy_guard->engaged_in_io) {
164 trace_reentrant_aio(bh->ctx, bh->name);
166 reentrancy_guard->engaged_in_io = true;
169 bh->cb(bh->opaque);
171 if (reentrancy_guard) {
172 reentrancy_guard->engaged_in_io = last_engaged_in_io;
176 /* Multiple occurrences of aio_bh_poll cannot be called concurrently. */
177 int aio_bh_poll(AioContext *ctx)
179 BHListSlice slice;
180 BHListSlice *s;
181 int ret = 0;
183 /* Synchronizes with QSLIST_INSERT_HEAD_ATOMIC in aio_bh_enqueue(). */
184 QSLIST_MOVE_ATOMIC(&slice.bh_list, &ctx->bh_list);
187 * GCC13 [-Werror=dangling-pointer=] complains that the local variable
188 * 'slice' is being stored in the global 'ctx->bh_slice_list' but the
189 * list is emptied before this function returns.
191 #if !defined(__clang__)
192 #pragma GCC diagnostic push
193 #pragma GCC diagnostic ignored "-Wpragmas"
194 #pragma GCC diagnostic ignored "-Wdangling-pointer="
195 #endif
196 QSIMPLEQ_INSERT_TAIL(&ctx->bh_slice_list, &slice, next);
197 #if !defined(__clang__)
198 #pragma GCC diagnostic pop
199 #endif
201 while ((s = QSIMPLEQ_FIRST(&ctx->bh_slice_list))) {
202 QEMUBH *bh;
203 unsigned flags;
205 bh = aio_bh_dequeue(&s->bh_list, &flags);
206 if (!bh) {
207 QSIMPLEQ_REMOVE_HEAD(&ctx->bh_slice_list, next);
208 continue;
211 if ((flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
212 /* Idle BHs don't count as progress */
213 if (!(flags & BH_IDLE)) {
214 ret = 1;
216 aio_bh_call(bh);
218 if (flags & (BH_DELETED | BH_ONESHOT)) {
219 g_free(bh);
223 return ret;
226 void qemu_bh_schedule_idle(QEMUBH *bh)
228 aio_bh_enqueue(bh, BH_SCHEDULED | BH_IDLE);
231 void qemu_bh_schedule(QEMUBH *bh)
233 aio_bh_enqueue(bh, BH_SCHEDULED);
236 /* This func is async.
238 void qemu_bh_cancel(QEMUBH *bh)
240 qatomic_and(&bh->flags, ~BH_SCHEDULED);
243 /* This func is async.The bottom half will do the delete action at the finial
244 * end.
246 void qemu_bh_delete(QEMUBH *bh)
248 aio_bh_enqueue(bh, BH_DELETED);
251 static int64_t aio_compute_bh_timeout(BHList *head, int timeout)
253 QEMUBH *bh;
255 QSLIST_FOREACH_RCU(bh, head, next) {
256 if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
257 if (bh->flags & BH_IDLE) {
258 /* idle bottom halves will be polled at least
259 * every 10ms */
260 timeout = 10000000;
261 } else {
262 /* non-idle bottom halves will be executed
263 * immediately */
264 return 0;
269 return timeout;
272 int64_t
273 aio_compute_timeout(AioContext *ctx)
275 BHListSlice *s;
276 int64_t deadline;
277 int timeout = -1;
279 timeout = aio_compute_bh_timeout(&ctx->bh_list, timeout);
280 if (timeout == 0) {
281 return 0;
284 QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
285 timeout = aio_compute_bh_timeout(&s->bh_list, timeout);
286 if (timeout == 0) {
287 return 0;
291 deadline = timerlistgroup_deadline_ns(&ctx->tlg);
292 if (deadline == 0) {
293 return 0;
294 } else {
295 return qemu_soonest_timeout(timeout, deadline);
299 static gboolean
300 aio_ctx_prepare(GSource *source, gint *timeout)
302 AioContext *ctx = (AioContext *) source;
304 qatomic_set(&ctx->notify_me, qatomic_read(&ctx->notify_me) | 1);
307 * Write ctx->notify_me before computing the timeout
308 * (reading bottom half flags, etc.). Pairs with
309 * smp_mb in aio_notify().
311 smp_mb();
313 /* We assume there is no timeout already supplied */
314 *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
316 if (aio_prepare(ctx)) {
317 *timeout = 0;
320 return *timeout == 0;
323 static gboolean
324 aio_ctx_check(GSource *source)
326 AioContext *ctx = (AioContext *) source;
327 QEMUBH *bh;
328 BHListSlice *s;
330 /* Finish computing the timeout before clearing the flag. */
331 qatomic_store_release(&ctx->notify_me, qatomic_read(&ctx->notify_me) & ~1);
332 aio_notify_accept(ctx);
334 QSLIST_FOREACH_RCU(bh, &ctx->bh_list, next) {
335 if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
336 return true;
340 QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
341 QSLIST_FOREACH_RCU(bh, &s->bh_list, next) {
342 if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
343 return true;
347 return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
350 static gboolean
351 aio_ctx_dispatch(GSource *source,
352 GSourceFunc callback,
353 gpointer user_data)
355 AioContext *ctx = (AioContext *) source;
357 assert(callback == NULL);
358 aio_dispatch(ctx);
359 return true;
362 static void
363 aio_ctx_finalize(GSource *source)
365 AioContext *ctx = (AioContext *) source;
366 QEMUBH *bh;
367 unsigned flags;
369 thread_pool_free(ctx->thread_pool);
371 #ifdef CONFIG_LINUX_AIO
372 if (ctx->linux_aio) {
373 laio_detach_aio_context(ctx->linux_aio, ctx);
374 laio_cleanup(ctx->linux_aio);
375 ctx->linux_aio = NULL;
377 #endif
379 #ifdef CONFIG_LINUX_IO_URING
380 if (ctx->linux_io_uring) {
381 luring_detach_aio_context(ctx->linux_io_uring, ctx);
382 luring_cleanup(ctx->linux_io_uring);
383 ctx->linux_io_uring = NULL;
385 #endif
387 assert(QSLIST_EMPTY(&ctx->scheduled_coroutines));
388 qemu_bh_delete(ctx->co_schedule_bh);
390 /* There must be no aio_bh_poll() calls going on */
391 assert(QSIMPLEQ_EMPTY(&ctx->bh_slice_list));
393 while ((bh = aio_bh_dequeue(&ctx->bh_list, &flags))) {
395 * qemu_bh_delete() must have been called on BHs in this AioContext. In
396 * many cases memory leaks, hangs, or inconsistent state occur when a
397 * BH is leaked because something still expects it to run.
399 * If you hit this, fix the lifecycle of the BH so that
400 * qemu_bh_delete() and any associated cleanup is called before the
401 * AioContext is finalized.
403 if (unlikely(!(flags & BH_DELETED))) {
404 fprintf(stderr, "%s: BH '%s' leaked, aborting...\n",
405 __func__, bh->name);
406 abort();
409 g_free(bh);
412 aio_set_event_notifier(ctx, &ctx->notifier, NULL, NULL, NULL);
413 event_notifier_cleanup(&ctx->notifier);
414 qemu_rec_mutex_destroy(&ctx->lock);
415 qemu_lockcnt_destroy(&ctx->list_lock);
416 timerlistgroup_deinit(&ctx->tlg);
417 unregister_aiocontext(ctx);
418 aio_context_destroy(ctx);
421 static GSourceFuncs aio_source_funcs = {
422 aio_ctx_prepare,
423 aio_ctx_check,
424 aio_ctx_dispatch,
425 aio_ctx_finalize
428 GSource *aio_get_g_source(AioContext *ctx)
430 aio_context_use_g_source(ctx);
431 g_source_ref(&ctx->source);
432 return &ctx->source;
435 ThreadPool *aio_get_thread_pool(AioContext *ctx)
437 if (!ctx->thread_pool) {
438 ctx->thread_pool = thread_pool_new(ctx);
440 return ctx->thread_pool;
443 #ifdef CONFIG_LINUX_AIO
444 LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp)
446 if (!ctx->linux_aio) {
447 ctx->linux_aio = laio_init(errp);
448 if (ctx->linux_aio) {
449 laio_attach_aio_context(ctx->linux_aio, ctx);
452 return ctx->linux_aio;
455 LinuxAioState *aio_get_linux_aio(AioContext *ctx)
457 assert(ctx->linux_aio);
458 return ctx->linux_aio;
460 #endif
462 #ifdef CONFIG_LINUX_IO_URING
463 LuringState *aio_setup_linux_io_uring(AioContext *ctx, Error **errp)
465 if (ctx->linux_io_uring) {
466 return ctx->linux_io_uring;
469 ctx->linux_io_uring = luring_init(errp);
470 if (!ctx->linux_io_uring) {
471 return NULL;
474 luring_attach_aio_context(ctx->linux_io_uring, ctx);
475 return ctx->linux_io_uring;
478 LuringState *aio_get_linux_io_uring(AioContext *ctx)
480 assert(ctx->linux_io_uring);
481 return ctx->linux_io_uring;
483 #endif
485 void aio_notify(AioContext *ctx)
488 * Write e.g. ctx->bh_list before writing ctx->notified. Pairs with
489 * smp_mb() in aio_notify_accept().
491 smp_wmb();
492 qatomic_set(&ctx->notified, true);
495 * Write ctx->notified (and also ctx->bh_list) before reading ctx->notify_me.
496 * Pairs with smp_mb() in aio_ctx_prepare or aio_poll.
498 smp_mb();
499 if (qatomic_read(&ctx->notify_me)) {
500 event_notifier_set(&ctx->notifier);
504 void aio_notify_accept(AioContext *ctx)
506 qatomic_set(&ctx->notified, false);
509 * Order reads of ctx->notified (in aio_context_notifier_poll()) and the
510 * above clearing of ctx->notified before reads of e.g. bh->flags. Pairs
511 * with smp_wmb() in aio_notify.
513 smp_mb();
516 static void aio_timerlist_notify(void *opaque, QEMUClockType type)
518 aio_notify(opaque);
521 static void aio_context_notifier_cb(EventNotifier *e)
523 AioContext *ctx = container_of(e, AioContext, notifier);
525 event_notifier_test_and_clear(&ctx->notifier);
528 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
529 static bool aio_context_notifier_poll(void *opaque)
531 EventNotifier *e = opaque;
532 AioContext *ctx = container_of(e, AioContext, notifier);
535 * No need for load-acquire because we just want to kick the
536 * event loop. aio_notify_accept() takes care of synchronizing
537 * the event loop with the producers.
539 return qatomic_read(&ctx->notified);
542 static void aio_context_notifier_poll_ready(EventNotifier *e)
544 /* Do nothing, we just wanted to kick the event loop */
547 static void co_schedule_bh_cb(void *opaque)
549 AioContext *ctx = opaque;
550 QSLIST_HEAD(, Coroutine) straight, reversed;
552 QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines);
553 QSLIST_INIT(&straight);
555 while (!QSLIST_EMPTY(&reversed)) {
556 Coroutine *co = QSLIST_FIRST(&reversed);
557 QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next);
558 QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next);
561 while (!QSLIST_EMPTY(&straight)) {
562 Coroutine *co = QSLIST_FIRST(&straight);
563 QSLIST_REMOVE_HEAD(&straight, co_scheduled_next);
564 trace_aio_co_schedule_bh_cb(ctx, co);
565 aio_context_acquire(ctx);
567 /* Protected by write barrier in qemu_aio_coroutine_enter */
568 qatomic_set(&co->scheduled, NULL);
569 qemu_aio_coroutine_enter(ctx, co);
570 aio_context_release(ctx);
574 AioContext *aio_context_new(Error **errp)
576 int ret;
577 AioContext *ctx;
579 ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
580 QSLIST_INIT(&ctx->bh_list);
581 QSIMPLEQ_INIT(&ctx->bh_slice_list);
582 aio_context_setup(ctx);
584 ret = event_notifier_init(&ctx->notifier, false);
585 if (ret < 0) {
586 error_setg_errno(errp, -ret, "Failed to initialize event notifier");
587 goto fail;
589 g_source_set_can_recurse(&ctx->source, true);
590 qemu_lockcnt_init(&ctx->list_lock);
592 ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx);
593 QSLIST_INIT(&ctx->scheduled_coroutines);
595 aio_set_event_notifier(ctx, &ctx->notifier,
596 aio_context_notifier_cb,
597 aio_context_notifier_poll,
598 aio_context_notifier_poll_ready);
599 #ifdef CONFIG_LINUX_AIO
600 ctx->linux_aio = NULL;
601 #endif
603 #ifdef CONFIG_LINUX_IO_URING
604 ctx->linux_io_uring = NULL;
605 #endif
607 ctx->thread_pool = NULL;
608 qemu_rec_mutex_init(&ctx->lock);
609 timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
611 ctx->poll_ns = 0;
612 ctx->poll_max_ns = 0;
613 ctx->poll_grow = 0;
614 ctx->poll_shrink = 0;
616 ctx->aio_max_batch = 0;
618 ctx->thread_pool_min = 0;
619 ctx->thread_pool_max = THREAD_POOL_MAX_THREADS_DEFAULT;
621 register_aiocontext(ctx);
623 return ctx;
624 fail:
625 g_source_destroy(&ctx->source);
626 return NULL;
629 void aio_co_schedule(AioContext *ctx, Coroutine *co)
631 trace_aio_co_schedule(ctx, co);
632 const char *scheduled = qatomic_cmpxchg(&co->scheduled, NULL,
633 __func__);
635 if (scheduled) {
636 fprintf(stderr,
637 "%s: Co-routine was already scheduled in '%s'\n",
638 __func__, scheduled);
639 abort();
642 /* The coroutine might run and release the last ctx reference before we
643 * invoke qemu_bh_schedule(). Take a reference to keep ctx alive until
644 * we're done.
646 aio_context_ref(ctx);
648 QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines,
649 co, co_scheduled_next);
650 qemu_bh_schedule(ctx->co_schedule_bh);
652 aio_context_unref(ctx);
655 typedef struct AioCoRescheduleSelf {
656 Coroutine *co;
657 AioContext *new_ctx;
658 } AioCoRescheduleSelf;
660 static void aio_co_reschedule_self_bh(void *opaque)
662 AioCoRescheduleSelf *data = opaque;
663 aio_co_schedule(data->new_ctx, data->co);
666 void coroutine_fn aio_co_reschedule_self(AioContext *new_ctx)
668 AioContext *old_ctx = qemu_get_current_aio_context();
670 if (old_ctx != new_ctx) {
671 AioCoRescheduleSelf data = {
672 .co = qemu_coroutine_self(),
673 .new_ctx = new_ctx,
676 * We can't directly schedule the coroutine in the target context
677 * because this would be racy: The other thread could try to enter the
678 * coroutine before it has yielded in this one.
680 aio_bh_schedule_oneshot(old_ctx, aio_co_reschedule_self_bh, &data);
681 qemu_coroutine_yield();
685 void aio_co_wake(Coroutine *co)
687 AioContext *ctx;
689 /* Read coroutine before co->ctx. Matches smp_wmb in
690 * qemu_coroutine_enter.
692 smp_read_barrier_depends();
693 ctx = qatomic_read(&co->ctx);
695 aio_co_enter(ctx, co);
698 void aio_co_enter(AioContext *ctx, Coroutine *co)
700 if (ctx != qemu_get_current_aio_context()) {
701 aio_co_schedule(ctx, co);
702 return;
705 if (qemu_in_coroutine()) {
706 Coroutine *self = qemu_coroutine_self();
707 assert(self != co);
708 QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next);
709 } else {
710 aio_context_acquire(ctx);
711 qemu_aio_coroutine_enter(ctx, co);
712 aio_context_release(ctx);
716 void aio_context_ref(AioContext *ctx)
718 g_source_ref(&ctx->source);
721 void aio_context_unref(AioContext *ctx)
723 g_source_unref(&ctx->source);
726 void aio_context_acquire(AioContext *ctx)
728 qemu_rec_mutex_lock(&ctx->lock);
731 void aio_context_release(AioContext *ctx)
733 qemu_rec_mutex_unlock(&ctx->lock);
736 QEMU_DEFINE_STATIC_CO_TLS(AioContext *, my_aiocontext)
738 AioContext *qemu_get_current_aio_context(void)
740 AioContext *ctx = get_my_aiocontext();
741 if (ctx) {
742 return ctx;
744 if (qemu_mutex_iothread_locked()) {
745 /* Possibly in a vCPU thread. */
746 return qemu_get_aio_context();
748 return NULL;
751 void qemu_set_current_aio_context(AioContext *ctx)
753 assert(!get_my_aiocontext());
754 set_my_aiocontext(ctx);
757 void aio_context_set_thread_pool_params(AioContext *ctx, int64_t min,
758 int64_t max, Error **errp)
761 if (min > max || !max || min > INT_MAX || max > INT_MAX) {
762 error_setg(errp, "bad thread-pool-min/thread-pool-max values");
763 return;
766 ctx->thread_pool_min = min;
767 ctx->thread_pool_max = max;
769 if (ctx->thread_pool) {
770 thread_pool_update_params(ctx->thread_pool, ctx);