xref: /qemu/util/async.c (revision a18025f9)
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 "block/raw-aio.h"
33 #include "qemu/coroutine_int.h"
34 #include "trace.h"
35 
36 /***********************************************************/
37 /* bottom halves (can be seen as timers which expire ASAP) */
38 
39 struct QEMUBH {
40     AioContext *ctx;
41     QEMUBHFunc *cb;
42     void *opaque;
43     QEMUBH *next;
44     bool scheduled;
45     bool idle;
46     bool deleted;
47 };
48 
49 void aio_bh_schedule_oneshot(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
50 {
51     QEMUBH *bh;
52     bh = g_new(QEMUBH, 1);
53     *bh = (QEMUBH){
54         .ctx = ctx,
55         .cb = cb,
56         .opaque = opaque,
57     };
58     qemu_lockcnt_lock(&ctx->list_lock);
59     bh->next = ctx->first_bh;
60     bh->scheduled = 1;
61     bh->deleted = 1;
62     /* Make sure that the members are ready before putting bh into list */
63     smp_wmb();
64     ctx->first_bh = bh;
65     qemu_lockcnt_unlock(&ctx->list_lock);
66     aio_notify(ctx);
67 }
68 
69 QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
70 {
71     QEMUBH *bh;
72     bh = g_new(QEMUBH, 1);
73     *bh = (QEMUBH){
74         .ctx = ctx,
75         .cb = cb,
76         .opaque = opaque,
77     };
78     qemu_lockcnt_lock(&ctx->list_lock);
79     bh->next = ctx->first_bh;
80     /* Make sure that the members are ready before putting bh into list */
81     smp_wmb();
82     ctx->first_bh = bh;
83     qemu_lockcnt_unlock(&ctx->list_lock);
84     return bh;
85 }
86 
87 void aio_bh_call(QEMUBH *bh)
88 {
89     bh->cb(bh->opaque);
90 }
91 
92 /* Multiple occurrences of aio_bh_poll cannot be called concurrently.
93  * The count in ctx->list_lock is incremented before the call, and is
94  * not affected by the call.
95  */
96 int aio_bh_poll(AioContext *ctx)
97 {
98     QEMUBH *bh, **bhp, *next;
99     int ret;
100     bool deleted = false;
101 
102     ret = 0;
103     for (bh = atomic_rcu_read(&ctx->first_bh); bh; bh = next) {
104         next = atomic_rcu_read(&bh->next);
105         /* The atomic_xchg is paired with the one in qemu_bh_schedule.  The
106          * implicit memory barrier ensures that the callback sees all writes
107          * done by the scheduling thread.  It also ensures that the scheduling
108          * thread sees the zero before bh->cb has run, and thus will call
109          * aio_notify again if necessary.
110          */
111         if (atomic_xchg(&bh->scheduled, 0)) {
112             /* Idle BHs don't count as progress */
113             if (!bh->idle) {
114                 ret = 1;
115             }
116             bh->idle = 0;
117             aio_bh_call(bh);
118         }
119         if (bh->deleted) {
120             deleted = true;
121         }
122     }
123 
124     /* remove deleted bhs */
125     if (!deleted) {
126         return ret;
127     }
128 
129     if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
130         bhp = &ctx->first_bh;
131         while (*bhp) {
132             bh = *bhp;
133             if (bh->deleted && !bh->scheduled) {
134                 *bhp = bh->next;
135                 g_free(bh);
136             } else {
137                 bhp = &bh->next;
138             }
139         }
140         qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
141     }
142     return ret;
143 }
144 
145 void qemu_bh_schedule_idle(QEMUBH *bh)
146 {
147     bh->idle = 1;
148     /* Make sure that idle & any writes needed by the callback are done
149      * before the locations are read in the aio_bh_poll.
150      */
151     atomic_mb_set(&bh->scheduled, 1);
152 }
153 
154 void qemu_bh_schedule(QEMUBH *bh)
155 {
156     AioContext *ctx;
157 
158     ctx = bh->ctx;
159     bh->idle = 0;
160     /* The memory barrier implicit in atomic_xchg makes sure that:
161      * 1. idle & any writes needed by the callback are done before the
162      *    locations are read in the aio_bh_poll.
163      * 2. ctx is loaded before scheduled is set and the callback has a chance
164      *    to execute.
165      */
166     if (atomic_xchg(&bh->scheduled, 1) == 0) {
167         aio_notify(ctx);
168     }
169 }
170 
171 
172 /* This func is async.
173  */
174 void qemu_bh_cancel(QEMUBH *bh)
175 {
176     atomic_mb_set(&bh->scheduled, 0);
177 }
178 
179 /* This func is async.The bottom half will do the delete action at the finial
180  * end.
181  */
182 void qemu_bh_delete(QEMUBH *bh)
183 {
184     bh->scheduled = 0;
185     bh->deleted = 1;
186 }
187 
188 int64_t
189 aio_compute_timeout(AioContext *ctx)
190 {
191     int64_t deadline;
192     int timeout = -1;
193     QEMUBH *bh;
194 
195     for (bh = atomic_rcu_read(&ctx->first_bh); bh;
196          bh = atomic_rcu_read(&bh->next)) {
197         if (bh->scheduled) {
198             if (bh->idle) {
199                 /* idle bottom halves will be polled at least
200                  * every 10ms */
201                 timeout = 10000000;
202             } else {
203                 /* non-idle bottom halves will be executed
204                  * immediately */
205                 return 0;
206             }
207         }
208     }
209 
210     deadline = timerlistgroup_deadline_ns(&ctx->tlg);
211     if (deadline == 0) {
212         return 0;
213     } else {
214         return qemu_soonest_timeout(timeout, deadline);
215     }
216 }
217 
218 static gboolean
219 aio_ctx_prepare(GSource *source, gint    *timeout)
220 {
221     AioContext *ctx = (AioContext *) source;
222 
223     atomic_or(&ctx->notify_me, 1);
224 
225     /* We assume there is no timeout already supplied */
226     *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
227 
228     if (aio_prepare(ctx)) {
229         *timeout = 0;
230     }
231 
232     return *timeout == 0;
233 }
234 
235 static gboolean
236 aio_ctx_check(GSource *source)
237 {
238     AioContext *ctx = (AioContext *) source;
239     QEMUBH *bh;
240 
241     atomic_and(&ctx->notify_me, ~1);
242     aio_notify_accept(ctx);
243 
244     for (bh = ctx->first_bh; bh; bh = bh->next) {
245         if (bh->scheduled) {
246             return true;
247         }
248     }
249     return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
250 }
251 
252 static gboolean
253 aio_ctx_dispatch(GSource     *source,
254                  GSourceFunc  callback,
255                  gpointer     user_data)
256 {
257     AioContext *ctx = (AioContext *) source;
258 
259     assert(callback == NULL);
260     aio_dispatch(ctx);
261     return true;
262 }
263 
264 static void
265 aio_ctx_finalize(GSource     *source)
266 {
267     AioContext *ctx = (AioContext *) source;
268 
269     thread_pool_free(ctx->thread_pool);
270 
271 #ifdef CONFIG_LINUX_AIO
272     if (ctx->linux_aio) {
273         laio_detach_aio_context(ctx->linux_aio, ctx);
274         laio_cleanup(ctx->linux_aio);
275         ctx->linux_aio = NULL;
276     }
277 #endif
278 
279     assert(QSLIST_EMPTY(&ctx->scheduled_coroutines));
280     qemu_bh_delete(ctx->co_schedule_bh);
281 
282     qemu_lockcnt_lock(&ctx->list_lock);
283     assert(!qemu_lockcnt_count(&ctx->list_lock));
284     while (ctx->first_bh) {
285         QEMUBH *next = ctx->first_bh->next;
286 
287         /* qemu_bh_delete() must have been called on BHs in this AioContext */
288         assert(ctx->first_bh->deleted);
289 
290         g_free(ctx->first_bh);
291         ctx->first_bh = next;
292     }
293     qemu_lockcnt_unlock(&ctx->list_lock);
294 
295     aio_set_event_notifier(ctx, &ctx->notifier, false, NULL, NULL);
296     event_notifier_cleanup(&ctx->notifier);
297     qemu_rec_mutex_destroy(&ctx->lock);
298     qemu_lockcnt_destroy(&ctx->list_lock);
299     timerlistgroup_deinit(&ctx->tlg);
300     aio_context_destroy(ctx);
301 }
302 
303 static GSourceFuncs aio_source_funcs = {
304     aio_ctx_prepare,
305     aio_ctx_check,
306     aio_ctx_dispatch,
307     aio_ctx_finalize
308 };
309 
310 GSource *aio_get_g_source(AioContext *ctx)
311 {
312     g_source_ref(&ctx->source);
313     return &ctx->source;
314 }
315 
316 ThreadPool *aio_get_thread_pool(AioContext *ctx)
317 {
318     if (!ctx->thread_pool) {
319         ctx->thread_pool = thread_pool_new(ctx);
320     }
321     return ctx->thread_pool;
322 }
323 
324 #ifdef CONFIG_LINUX_AIO
325 LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp)
326 {
327     if (!ctx->linux_aio) {
328         ctx->linux_aio = laio_init(errp);
329         if (ctx->linux_aio) {
330             laio_attach_aio_context(ctx->linux_aio, ctx);
331         }
332     }
333     return ctx->linux_aio;
334 }
335 
336 LinuxAioState *aio_get_linux_aio(AioContext *ctx)
337 {
338     assert(ctx->linux_aio);
339     return ctx->linux_aio;
340 }
341 #endif
342 
343 void aio_notify(AioContext *ctx)
344 {
345     /* Write e.g. bh->scheduled before reading ctx->notify_me.  Pairs
346      * with atomic_or in aio_ctx_prepare or atomic_add in aio_poll.
347      */
348     smp_mb();
349     if (ctx->notify_me) {
350         event_notifier_set(&ctx->notifier);
351         atomic_mb_set(&ctx->notified, true);
352     }
353 }
354 
355 void aio_notify_accept(AioContext *ctx)
356 {
357     if (atomic_xchg(&ctx->notified, false)
358 #ifdef WIN32
359         || true
360 #endif
361     ) {
362         event_notifier_test_and_clear(&ctx->notifier);
363     }
364 }
365 
366 static void aio_timerlist_notify(void *opaque, QEMUClockType type)
367 {
368     aio_notify(opaque);
369 }
370 
371 static void event_notifier_dummy_cb(EventNotifier *e)
372 {
373 }
374 
375 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
376 static bool event_notifier_poll(void *opaque)
377 {
378     EventNotifier *e = opaque;
379     AioContext *ctx = container_of(e, AioContext, notifier);
380 
381     return atomic_read(&ctx->notified);
382 }
383 
384 static void co_schedule_bh_cb(void *opaque)
385 {
386     AioContext *ctx = opaque;
387     QSLIST_HEAD(, Coroutine) straight, reversed;
388 
389     QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines);
390     QSLIST_INIT(&straight);
391 
392     while (!QSLIST_EMPTY(&reversed)) {
393         Coroutine *co = QSLIST_FIRST(&reversed);
394         QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next);
395         QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next);
396     }
397 
398     while (!QSLIST_EMPTY(&straight)) {
399         Coroutine *co = QSLIST_FIRST(&straight);
400         QSLIST_REMOVE_HEAD(&straight, co_scheduled_next);
401         trace_aio_co_schedule_bh_cb(ctx, co);
402         aio_context_acquire(ctx);
403 
404         /* Protected by write barrier in qemu_aio_coroutine_enter */
405         atomic_set(&co->scheduled, NULL);
406         qemu_aio_coroutine_enter(ctx, co);
407         aio_context_release(ctx);
408     }
409 }
410 
411 AioContext *aio_context_new(Error **errp)
412 {
413     int ret;
414     AioContext *ctx;
415 
416     ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
417     aio_context_setup(ctx);
418 
419     ret = event_notifier_init(&ctx->notifier, false);
420     if (ret < 0) {
421         error_setg_errno(errp, -ret, "Failed to initialize event notifier");
422         goto fail;
423     }
424     g_source_set_can_recurse(&ctx->source, true);
425     qemu_lockcnt_init(&ctx->list_lock);
426 
427     ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx);
428     QSLIST_INIT(&ctx->scheduled_coroutines);
429 
430     aio_set_event_notifier(ctx, &ctx->notifier,
431                            false,
432                            event_notifier_dummy_cb,
433                            event_notifier_poll);
434 #ifdef CONFIG_LINUX_AIO
435     ctx->linux_aio = NULL;
436 #endif
437     ctx->thread_pool = NULL;
438     qemu_rec_mutex_init(&ctx->lock);
439     timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
440 
441     ctx->poll_ns = 0;
442     ctx->poll_max_ns = 0;
443     ctx->poll_grow = 0;
444     ctx->poll_shrink = 0;
445 
446     return ctx;
447 fail:
448     g_source_destroy(&ctx->source);
449     return NULL;
450 }
451 
452 void aio_co_schedule(AioContext *ctx, Coroutine *co)
453 {
454     trace_aio_co_schedule(ctx, co);
455     const char *scheduled = atomic_cmpxchg(&co->scheduled, NULL,
456                                            __func__);
457 
458     if (scheduled) {
459         fprintf(stderr,
460                 "%s: Co-routine was already scheduled in '%s'\n",
461                 __func__, scheduled);
462         abort();
463     }
464 
465     /* The coroutine might run and release the last ctx reference before we
466      * invoke qemu_bh_schedule().  Take a reference to keep ctx alive until
467      * we're done.
468      */
469     aio_context_ref(ctx);
470 
471     QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines,
472                               co, co_scheduled_next);
473     qemu_bh_schedule(ctx->co_schedule_bh);
474 
475     aio_context_unref(ctx);
476 }
477 
478 void aio_co_wake(struct Coroutine *co)
479 {
480     AioContext *ctx;
481 
482     /* Read coroutine before co->ctx.  Matches smp_wmb in
483      * qemu_coroutine_enter.
484      */
485     smp_read_barrier_depends();
486     ctx = atomic_read(&co->ctx);
487 
488     aio_co_enter(ctx, co);
489 }
490 
491 void aio_co_enter(AioContext *ctx, struct Coroutine *co)
492 {
493     if (ctx != qemu_get_current_aio_context()) {
494         aio_co_schedule(ctx, co);
495         return;
496     }
497 
498     if (qemu_in_coroutine()) {
499         Coroutine *self = qemu_coroutine_self();
500         assert(self != co);
501         QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next);
502     } else {
503         aio_context_acquire(ctx);
504         qemu_aio_coroutine_enter(ctx, co);
505         aio_context_release(ctx);
506     }
507 }
508 
509 void aio_context_ref(AioContext *ctx)
510 {
511     g_source_ref(&ctx->source);
512 }
513 
514 void aio_context_unref(AioContext *ctx)
515 {
516     g_source_unref(&ctx->source);
517 }
518 
519 void aio_context_acquire(AioContext *ctx)
520 {
521     qemu_rec_mutex_lock(&ctx->lock);
522 }
523 
524 void aio_context_release(AioContext *ctx)
525 {
526     qemu_rec_mutex_unlock(&ctx->lock);
527 }
528