xref: /qemu/util/async.c (revision 12b35405)
1 /*
2  * Data plane event loop
3  *
4  * Copyright (c) 2003-2008 Fabrice Bellard
5  * Copyright (c) 2009-2017 QEMU contributors
6  *
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:
13  *
14  * The above copyright notice and this permission notice shall be included in
15  * all copies or substantial portions of the Software.
16  *
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.
24  */
25 
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 "trace.h"
36 
37 /***********************************************************/
38 /* bottom halves (can be seen as timers which expire ASAP) */
39 
40 /* QEMUBH::flags values */
41 enum {
42     /* Already enqueued and waiting for aio_bh_poll() */
43     BH_PENDING   = (1 << 0),
44 
45     /* Invoke the callback */
46     BH_SCHEDULED = (1 << 1),
47 
48     /* Delete without invoking callback */
49     BH_DELETED   = (1 << 2),
50 
51     /* Delete after invoking callback */
52     BH_ONESHOT   = (1 << 3),
53 
54     /* Schedule periodically when the event loop is idle */
55     BH_IDLE      = (1 << 4),
56 };
57 
58 struct QEMUBH {
59     AioContext *ctx;
60     QEMUBHFunc *cb;
61     void *opaque;
62     QSLIST_ENTRY(QEMUBH) next;
63     unsigned flags;
64 };
65 
66 /* Called concurrently from any thread */
67 static void aio_bh_enqueue(QEMUBH *bh, unsigned new_flags)
68 {
69     AioContext *ctx = bh->ctx;
70     unsigned old_flags;
71 
72     /*
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
77      *    could be freed.
78      */
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);
82     }
83 
84     aio_notify(ctx);
85 }
86 
87 /* Only called from aio_bh_poll() and aio_ctx_finalize() */
88 static QEMUBH *aio_bh_dequeue(BHList *head, unsigned *flags)
89 {
90     QEMUBH *bh = QSLIST_FIRST_RCU(head);
91 
92     if (!bh) {
93         return NULL;
94     }
95 
96     QSLIST_REMOVE_HEAD(head, next);
97 
98     /*
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
103      * necessary.
104      */
105     *flags = atomic_fetch_and(&bh->flags,
106                               ~(BH_PENDING | BH_SCHEDULED | BH_IDLE));
107     return bh;
108 }
109 
110 void aio_bh_schedule_oneshot(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
111 {
112     QEMUBH *bh;
113     bh = g_new(QEMUBH, 1);
114     *bh = (QEMUBH){
115         .ctx = ctx,
116         .cb = cb,
117         .opaque = opaque,
118     };
119     aio_bh_enqueue(bh, BH_SCHEDULED | BH_ONESHOT);
120 }
121 
122 QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
123 {
124     QEMUBH *bh;
125     bh = g_new(QEMUBH, 1);
126     *bh = (QEMUBH){
127         .ctx = ctx,
128         .cb = cb,
129         .opaque = opaque,
130     };
131     return bh;
132 }
133 
134 void aio_bh_call(QEMUBH *bh)
135 {
136     bh->cb(bh->opaque);
137 }
138 
139 /* Multiple occurrences of aio_bh_poll cannot be called concurrently. */
140 int aio_bh_poll(AioContext *ctx)
141 {
142     BHListSlice slice;
143     BHListSlice *s;
144     int ret = 0;
145 
146     QSLIST_MOVE_ATOMIC(&slice.bh_list, &ctx->bh_list);
147     QSIMPLEQ_INSERT_TAIL(&ctx->bh_slice_list, &slice, next);
148 
149     while ((s = QSIMPLEQ_FIRST(&ctx->bh_slice_list))) {
150         QEMUBH *bh;
151         unsigned flags;
152 
153         bh = aio_bh_dequeue(&s->bh_list, &flags);
154         if (!bh) {
155             QSIMPLEQ_REMOVE_HEAD(&ctx->bh_slice_list, next);
156             continue;
157         }
158 
159         if ((flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
160             /* Idle BHs don't count as progress */
161             if (!(flags & BH_IDLE)) {
162                 ret = 1;
163             }
164             aio_bh_call(bh);
165         }
166         if (flags & (BH_DELETED | BH_ONESHOT)) {
167             g_free(bh);
168         }
169     }
170 
171     return ret;
172 }
173 
174 void qemu_bh_schedule_idle(QEMUBH *bh)
175 {
176     aio_bh_enqueue(bh, BH_SCHEDULED | BH_IDLE);
177 }
178 
179 void qemu_bh_schedule(QEMUBH *bh)
180 {
181     aio_bh_enqueue(bh, BH_SCHEDULED);
182 }
183 
184 /* This func is async.
185  */
186 void qemu_bh_cancel(QEMUBH *bh)
187 {
188     atomic_and(&bh->flags, ~BH_SCHEDULED);
189 }
190 
191 /* This func is async.The bottom half will do the delete action at the finial
192  * end.
193  */
194 void qemu_bh_delete(QEMUBH *bh)
195 {
196     aio_bh_enqueue(bh, BH_DELETED);
197 }
198 
199 static int64_t aio_compute_bh_timeout(BHList *head, int timeout)
200 {
201     QEMUBH *bh;
202 
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
207                  * every 10ms */
208                 timeout = 10000000;
209             } else {
210                 /* non-idle bottom halves will be executed
211                  * immediately */
212                 return 0;
213             }
214         }
215     }
216 
217     return timeout;
218 }
219 
220 int64_t
221 aio_compute_timeout(AioContext *ctx)
222 {
223     BHListSlice *s;
224     int64_t deadline;
225     int timeout = -1;
226 
227     timeout = aio_compute_bh_timeout(&ctx->bh_list, timeout);
228     if (timeout == 0) {
229         return 0;
230     }
231 
232     QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
233         timeout = aio_compute_bh_timeout(&s->bh_list, timeout);
234         if (timeout == 0) {
235             return 0;
236         }
237     }
238 
239     deadline = timerlistgroup_deadline_ns(&ctx->tlg);
240     if (deadline == 0) {
241         return 0;
242     } else {
243         return qemu_soonest_timeout(timeout, deadline);
244     }
245 }
246 
247 static gboolean
248 aio_ctx_prepare(GSource *source, gint    *timeout)
249 {
250     AioContext *ctx = (AioContext *) source;
251 
252     atomic_set(&ctx->notify_me, atomic_read(&ctx->notify_me) | 1);
253 
254     /*
255      * Write ctx->notify_me before computing the timeout
256      * (reading bottom half flags, etc.).  Pairs with
257      * smp_mb in aio_notify().
258      */
259     smp_mb();
260 
261     /* We assume there is no timeout already supplied */
262     *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
263 
264     if (aio_prepare(ctx)) {
265         *timeout = 0;
266     }
267 
268     return *timeout == 0;
269 }
270 
271 static gboolean
272 aio_ctx_check(GSource *source)
273 {
274     AioContext *ctx = (AioContext *) source;
275     QEMUBH *bh;
276     BHListSlice *s;
277 
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);
281 
282     QSLIST_FOREACH_RCU(bh, &ctx->bh_list, next) {
283         if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
284             return true;
285         }
286     }
287 
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) {
291                 return true;
292             }
293         }
294     }
295     return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
296 }
297 
298 static gboolean
299 aio_ctx_dispatch(GSource     *source,
300                  GSourceFunc  callback,
301                  gpointer     user_data)
302 {
303     AioContext *ctx = (AioContext *) source;
304 
305     assert(callback == NULL);
306     aio_dispatch(ctx);
307     return true;
308 }
309 
310 static void
311 aio_ctx_finalize(GSource     *source)
312 {
313     AioContext *ctx = (AioContext *) source;
314     QEMUBH *bh;
315     unsigned flags;
316 
317     thread_pool_free(ctx->thread_pool);
318 
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;
324     }
325 #endif
326 
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;
332     }
333 #endif
334 
335     assert(QSLIST_EMPTY(&ctx->scheduled_coroutines));
336     qemu_bh_delete(ctx->co_schedule_bh);
337 
338     /* There must be no aio_bh_poll() calls going on */
339     assert(QSIMPLEQ_EMPTY(&ctx->bh_slice_list));
340 
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);
344 
345         g_free(bh);
346     }
347 
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);
354 }
355 
356 static GSourceFuncs aio_source_funcs = {
357     aio_ctx_prepare,
358     aio_ctx_check,
359     aio_ctx_dispatch,
360     aio_ctx_finalize
361 };
362 
363 GSource *aio_get_g_source(AioContext *ctx)
364 {
365     aio_context_use_g_source(ctx);
366     g_source_ref(&ctx->source);
367     return &ctx->source;
368 }
369 
370 ThreadPool *aio_get_thread_pool(AioContext *ctx)
371 {
372     if (!ctx->thread_pool) {
373         ctx->thread_pool = thread_pool_new(ctx);
374     }
375     return ctx->thread_pool;
376 }
377 
378 #ifdef CONFIG_LINUX_AIO
379 LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp)
380 {
381     if (!ctx->linux_aio) {
382         ctx->linux_aio = laio_init(errp);
383         if (ctx->linux_aio) {
384             laio_attach_aio_context(ctx->linux_aio, ctx);
385         }
386     }
387     return ctx->linux_aio;
388 }
389 
390 LinuxAioState *aio_get_linux_aio(AioContext *ctx)
391 {
392     assert(ctx->linux_aio);
393     return ctx->linux_aio;
394 }
395 #endif
396 
397 #ifdef CONFIG_LINUX_IO_URING
398 LuringState *aio_setup_linux_io_uring(AioContext *ctx, Error **errp)
399 {
400     if (ctx->linux_io_uring) {
401         return ctx->linux_io_uring;
402     }
403 
404     ctx->linux_io_uring = luring_init(errp);
405     if (!ctx->linux_io_uring) {
406         return NULL;
407     }
408 
409     luring_attach_aio_context(ctx->linux_io_uring, ctx);
410     return ctx->linux_io_uring;
411 }
412 
413 LuringState *aio_get_linux_io_uring(AioContext *ctx)
414 {
415     assert(ctx->linux_io_uring);
416     return ctx->linux_io_uring;
417 }
418 #endif
419 
420 void aio_notify(AioContext *ctx)
421 {
422     /* Write e.g. bh->scheduled before reading ctx->notify_me.  Pairs
423      * with smp_mb in aio_ctx_prepare or aio_poll.
424      */
425     smp_mb();
426     if (atomic_read(&ctx->notify_me)) {
427         event_notifier_set(&ctx->notifier);
428         atomic_mb_set(&ctx->notified, true);
429     }
430 }
431 
432 void aio_notify_accept(AioContext *ctx)
433 {
434     if (atomic_xchg(&ctx->notified, false)
435 #ifdef WIN32
436         || true
437 #endif
438     ) {
439         event_notifier_test_and_clear(&ctx->notifier);
440     }
441 }
442 
443 static void aio_timerlist_notify(void *opaque, QEMUClockType type)
444 {
445     aio_notify(opaque);
446 }
447 
448 static void event_notifier_dummy_cb(EventNotifier *e)
449 {
450 }
451 
452 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
453 static bool event_notifier_poll(void *opaque)
454 {
455     EventNotifier *e = opaque;
456     AioContext *ctx = container_of(e, AioContext, notifier);
457 
458     return atomic_read(&ctx->notified);
459 }
460 
461 static void co_schedule_bh_cb(void *opaque)
462 {
463     AioContext *ctx = opaque;
464     QSLIST_HEAD(, Coroutine) straight, reversed;
465 
466     QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines);
467     QSLIST_INIT(&straight);
468 
469     while (!QSLIST_EMPTY(&reversed)) {
470         Coroutine *co = QSLIST_FIRST(&reversed);
471         QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next);
472         QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next);
473     }
474 
475     while (!QSLIST_EMPTY(&straight)) {
476         Coroutine *co = QSLIST_FIRST(&straight);
477         QSLIST_REMOVE_HEAD(&straight, co_scheduled_next);
478         trace_aio_co_schedule_bh_cb(ctx, co);
479         aio_context_acquire(ctx);
480 
481         /* Protected by write barrier in qemu_aio_coroutine_enter */
482         atomic_set(&co->scheduled, NULL);
483         qemu_aio_coroutine_enter(ctx, co);
484         aio_context_release(ctx);
485     }
486 }
487 
488 AioContext *aio_context_new(Error **errp)
489 {
490     int ret;
491     AioContext *ctx;
492 
493     ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
494     QSLIST_INIT(&ctx->bh_list);
495     QSIMPLEQ_INIT(&ctx->bh_slice_list);
496     aio_context_setup(ctx);
497 
498     ret = event_notifier_init(&ctx->notifier, false);
499     if (ret < 0) {
500         error_setg_errno(errp, -ret, "Failed to initialize event notifier");
501         goto fail;
502     }
503     g_source_set_can_recurse(&ctx->source, true);
504     qemu_lockcnt_init(&ctx->list_lock);
505 
506     ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx);
507     QSLIST_INIT(&ctx->scheduled_coroutines);
508 
509     aio_set_event_notifier(ctx, &ctx->notifier,
510                            false,
511                            event_notifier_dummy_cb,
512                            event_notifier_poll);
513 #ifdef CONFIG_LINUX_AIO
514     ctx->linux_aio = NULL;
515 #endif
516 
517 #ifdef CONFIG_LINUX_IO_URING
518     ctx->linux_io_uring = NULL;
519 #endif
520 
521     ctx->thread_pool = NULL;
522     qemu_rec_mutex_init(&ctx->lock);
523     timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
524 
525     ctx->poll_ns = 0;
526     ctx->poll_max_ns = 0;
527     ctx->poll_grow = 0;
528     ctx->poll_shrink = 0;
529 
530     return ctx;
531 fail:
532     g_source_destroy(&ctx->source);
533     return NULL;
534 }
535 
536 void aio_co_schedule(AioContext *ctx, Coroutine *co)
537 {
538     trace_aio_co_schedule(ctx, co);
539     const char *scheduled = atomic_cmpxchg(&co->scheduled, NULL,
540                                            __func__);
541 
542     if (scheduled) {
543         fprintf(stderr,
544                 "%s: Co-routine was already scheduled in '%s'\n",
545                 __func__, scheduled);
546         abort();
547     }
548 
549     /* The coroutine might run and release the last ctx reference before we
550      * invoke qemu_bh_schedule().  Take a reference to keep ctx alive until
551      * we're done.
552      */
553     aio_context_ref(ctx);
554 
555     QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines,
556                               co, co_scheduled_next);
557     qemu_bh_schedule(ctx->co_schedule_bh);
558 
559     aio_context_unref(ctx);
560 }
561 
562 void aio_co_wake(struct Coroutine *co)
563 {
564     AioContext *ctx;
565 
566     /* Read coroutine before co->ctx.  Matches smp_wmb in
567      * qemu_coroutine_enter.
568      */
569     smp_read_barrier_depends();
570     ctx = atomic_read(&co->ctx);
571 
572     aio_co_enter(ctx, co);
573 }
574 
575 void aio_co_enter(AioContext *ctx, struct Coroutine *co)
576 {
577     if (ctx != qemu_get_current_aio_context()) {
578         aio_co_schedule(ctx, co);
579         return;
580     }
581 
582     if (qemu_in_coroutine()) {
583         Coroutine *self = qemu_coroutine_self();
584         assert(self != co);
585         QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next);
586     } else {
587         aio_context_acquire(ctx);
588         qemu_aio_coroutine_enter(ctx, co);
589         aio_context_release(ctx);
590     }
591 }
592 
593 void aio_context_ref(AioContext *ctx)
594 {
595     g_source_ref(&ctx->source);
596 }
597 
598 void aio_context_unref(AioContext *ctx)
599 {
600     g_source_unref(&ctx->source);
601 }
602 
603 void aio_context_acquire(AioContext *ctx)
604 {
605     qemu_rec_mutex_lock(&ctx->lock);
606 }
607 
608 void aio_context_release(AioContext *ctx)
609 {
610     qemu_rec_mutex_unlock(&ctx->lock);
611 }
612