xref: /qemu/util/aio-posix.c (revision 7c1f51bf)
1 /*
2  * QEMU aio implementation
3  *
4  * Copyright IBM, Corp. 2008
5  *
6  * Authors:
7  *  Anthony Liguori   <aliguori@us.ibm.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  *
12  * Contributions after 2012-01-13 are licensed under the terms of the
13  * GNU GPL, version 2 or (at your option) any later version.
14  */
15 
16 #include "qemu/osdep.h"
17 #include "block/block.h"
18 #include "block/thread-pool.h"
19 #include "qemu/main-loop.h"
20 #include "qemu/rcu.h"
21 #include "qemu/rcu_queue.h"
22 #include "qemu/sockets.h"
23 #include "qemu/cutils.h"
24 #include "trace.h"
25 #include "aio-posix.h"
26 
27 /* Stop userspace polling on a handler if it isn't active for some time */
28 #define POLL_IDLE_INTERVAL_NS (7 * NANOSECONDS_PER_SECOND)
29 
30 bool aio_poll_disabled(AioContext *ctx)
31 {
32     return qatomic_read(&ctx->poll_disable_cnt);
33 }
34 
35 void aio_add_ready_handler(AioHandlerList *ready_list,
36                            AioHandler *node,
37                            int revents)
38 {
39     QLIST_SAFE_REMOVE(node, node_ready); /* remove from nested parent's list */
40     node->pfd.revents = revents;
41     QLIST_INSERT_HEAD(ready_list, node, node_ready);
42 }
43 
44 static void aio_add_poll_ready_handler(AioHandlerList *ready_list,
45                                        AioHandler *node)
46 {
47     QLIST_SAFE_REMOVE(node, node_ready); /* remove from nested parent's list */
48     node->poll_ready = true;
49     QLIST_INSERT_HEAD(ready_list, node, node_ready);
50 }
51 
52 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
53 {
54     AioHandler *node;
55 
56     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
57         if (node->pfd.fd == fd) {
58             if (!QLIST_IS_INSERTED(node, node_deleted)) {
59                 return node;
60             }
61         }
62     }
63 
64     return NULL;
65 }
66 
67 static bool aio_remove_fd_handler(AioContext *ctx, AioHandler *node)
68 {
69     /* If the GSource is in the process of being destroyed then
70      * g_source_remove_poll() causes an assertion failure.  Skip
71      * removal in that case, because glib cleans up its state during
72      * destruction anyway.
73      */
74     if (!g_source_is_destroyed(&ctx->source)) {
75         g_source_remove_poll(&ctx->source, &node->pfd);
76     }
77 
78     node->pfd.revents = 0;
79     node->poll_ready = false;
80 
81     /* If the fd monitor has already marked it deleted, leave it alone */
82     if (QLIST_IS_INSERTED(node, node_deleted)) {
83         return false;
84     }
85 
86     /* If a read is in progress, just mark the node as deleted */
87     if (qemu_lockcnt_count(&ctx->list_lock)) {
88         QLIST_INSERT_HEAD_RCU(&ctx->deleted_aio_handlers, node, node_deleted);
89         return false;
90     }
91     /* Otherwise, delete it for real.  We can't just mark it as
92      * deleted because deleted nodes are only cleaned up while
93      * no one is walking the handlers list.
94      */
95     QLIST_SAFE_REMOVE(node, node_poll);
96     QLIST_REMOVE(node, node);
97     return true;
98 }
99 
100 void aio_set_fd_handler(AioContext *ctx,
101                         int fd,
102                         bool is_external,
103                         IOHandler *io_read,
104                         IOHandler *io_write,
105                         AioPollFn *io_poll,
106                         IOHandler *io_poll_ready,
107                         void *opaque)
108 {
109     AioHandler *node;
110     AioHandler *new_node = NULL;
111     bool is_new = false;
112     bool deleted = false;
113     int poll_disable_change;
114 
115     if (io_poll && !io_poll_ready) {
116         io_poll = NULL; /* polling only makes sense if there is a handler */
117     }
118 
119     qemu_lockcnt_lock(&ctx->list_lock);
120 
121     node = find_aio_handler(ctx, fd);
122 
123     /* Are we deleting the fd handler? */
124     if (!io_read && !io_write && !io_poll) {
125         if (node == NULL) {
126             qemu_lockcnt_unlock(&ctx->list_lock);
127             return;
128         }
129         /* Clean events in order to unregister fd from the ctx epoll. */
130         node->pfd.events = 0;
131 
132         poll_disable_change = -!node->io_poll;
133     } else {
134         poll_disable_change = !io_poll - (node && !node->io_poll);
135         if (node == NULL) {
136             is_new = true;
137         }
138         /* Alloc and insert if it's not already there */
139         new_node = g_new0(AioHandler, 1);
140 
141         /* Update handler with latest information */
142         new_node->io_read = io_read;
143         new_node->io_write = io_write;
144         new_node->io_poll = io_poll;
145         new_node->io_poll_ready = io_poll_ready;
146         new_node->opaque = opaque;
147         new_node->is_external = is_external;
148 
149         if (is_new) {
150             new_node->pfd.fd = fd;
151         } else {
152             new_node->pfd = node->pfd;
153         }
154         g_source_add_poll(&ctx->source, &new_node->pfd);
155 
156         new_node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
157         new_node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
158 
159         QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, new_node, node);
160     }
161 
162     /* No need to order poll_disable_cnt writes against other updates;
163      * the counter is only used to avoid wasting time and latency on
164      * iterated polling when the system call will be ultimately necessary.
165      * Changing handlers is a rare event, and a little wasted polling until
166      * the aio_notify below is not an issue.
167      */
168     qatomic_set(&ctx->poll_disable_cnt,
169                qatomic_read(&ctx->poll_disable_cnt) + poll_disable_change);
170 
171     ctx->fdmon_ops->update(ctx, node, new_node);
172     if (node) {
173         deleted = aio_remove_fd_handler(ctx, node);
174     }
175     qemu_lockcnt_unlock(&ctx->list_lock);
176     aio_notify(ctx);
177 
178     if (deleted) {
179         g_free(node);
180     }
181 }
182 
183 static void aio_set_fd_poll(AioContext *ctx, int fd,
184                             IOHandler *io_poll_begin,
185                             IOHandler *io_poll_end)
186 {
187     AioHandler *node = find_aio_handler(ctx, fd);
188 
189     if (!node) {
190         return;
191     }
192 
193     node->io_poll_begin = io_poll_begin;
194     node->io_poll_end = io_poll_end;
195 }
196 
197 void aio_set_event_notifier(AioContext *ctx,
198                             EventNotifier *notifier,
199                             bool is_external,
200                             EventNotifierHandler *io_read,
201                             AioPollFn *io_poll,
202                             EventNotifierHandler *io_poll_ready)
203 {
204     aio_set_fd_handler(ctx, event_notifier_get_fd(notifier), is_external,
205                        (IOHandler *)io_read, NULL, io_poll,
206                        (IOHandler *)io_poll_ready, notifier);
207 }
208 
209 void aio_set_event_notifier_poll(AioContext *ctx,
210                                  EventNotifier *notifier,
211                                  EventNotifierHandler *io_poll_begin,
212                                  EventNotifierHandler *io_poll_end)
213 {
214     aio_set_fd_poll(ctx, event_notifier_get_fd(notifier),
215                     (IOHandler *)io_poll_begin,
216                     (IOHandler *)io_poll_end);
217 }
218 
219 static bool poll_set_started(AioContext *ctx, AioHandlerList *ready_list,
220                              bool started)
221 {
222     AioHandler *node;
223     bool progress = false;
224 
225     if (started == ctx->poll_started) {
226         return false;
227     }
228 
229     ctx->poll_started = started;
230 
231     qemu_lockcnt_inc(&ctx->list_lock);
232     QLIST_FOREACH(node, &ctx->poll_aio_handlers, node_poll) {
233         IOHandler *fn;
234 
235         if (QLIST_IS_INSERTED(node, node_deleted)) {
236             continue;
237         }
238 
239         if (started) {
240             fn = node->io_poll_begin;
241         } else {
242             fn = node->io_poll_end;
243         }
244 
245         if (fn) {
246             fn(node->opaque);
247         }
248 
249         /* Poll one last time in case ->io_poll_end() raced with the event */
250         if (!started && node->io_poll(node->opaque)) {
251             aio_add_poll_ready_handler(ready_list, node);
252             progress = true;
253         }
254     }
255     qemu_lockcnt_dec(&ctx->list_lock);
256 
257     return progress;
258 }
259 
260 
261 bool aio_prepare(AioContext *ctx)
262 {
263     AioHandlerList ready_list = QLIST_HEAD_INITIALIZER(ready_list);
264 
265     /* Poll mode cannot be used with glib's event loop, disable it. */
266     poll_set_started(ctx, &ready_list, false);
267     /* TODO what to do with this list? */
268 
269     return false;
270 }
271 
272 bool aio_pending(AioContext *ctx)
273 {
274     AioHandler *node;
275     bool result = false;
276 
277     /*
278      * We have to walk very carefully in case aio_set_fd_handler is
279      * called while we're walking.
280      */
281     qemu_lockcnt_inc(&ctx->list_lock);
282 
283     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
284         int revents;
285 
286         /* TODO should this check poll ready? */
287         revents = node->pfd.revents & node->pfd.events;
288         if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read &&
289             aio_node_check(ctx, node->is_external)) {
290             result = true;
291             break;
292         }
293         if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write &&
294             aio_node_check(ctx, node->is_external)) {
295             result = true;
296             break;
297         }
298     }
299     qemu_lockcnt_dec(&ctx->list_lock);
300 
301     return result;
302 }
303 
304 static void aio_free_deleted_handlers(AioContext *ctx)
305 {
306     AioHandler *node;
307 
308     if (QLIST_EMPTY_RCU(&ctx->deleted_aio_handlers)) {
309         return;
310     }
311     if (!qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
312         return; /* we are nested, let the parent do the freeing */
313     }
314 
315     while ((node = QLIST_FIRST_RCU(&ctx->deleted_aio_handlers))) {
316         QLIST_REMOVE(node, node);
317         QLIST_REMOVE(node, node_deleted);
318         QLIST_SAFE_REMOVE(node, node_poll);
319         g_free(node);
320     }
321 
322     qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
323 }
324 
325 static bool aio_dispatch_handler(AioContext *ctx, AioHandler *node)
326 {
327     bool progress = false;
328     bool poll_ready;
329     int revents;
330 
331     revents = node->pfd.revents & node->pfd.events;
332     node->pfd.revents = 0;
333 
334     poll_ready = node->poll_ready;
335     node->poll_ready = false;
336 
337     /*
338      * Start polling AioHandlers when they become ready because activity is
339      * likely to continue.  Note that starvation is theoretically possible when
340      * fdmon_supports_polling(), but only until the fd fires for the first
341      * time.
342      */
343     if (!QLIST_IS_INSERTED(node, node_deleted) &&
344         !QLIST_IS_INSERTED(node, node_poll) &&
345         node->io_poll) {
346         trace_poll_add(ctx, node, node->pfd.fd, revents);
347         if (ctx->poll_started && node->io_poll_begin) {
348             node->io_poll_begin(node->opaque);
349         }
350         QLIST_INSERT_HEAD(&ctx->poll_aio_handlers, node, node_poll);
351     }
352     if (!QLIST_IS_INSERTED(node, node_deleted) &&
353         poll_ready && revents == 0 &&
354         aio_node_check(ctx, node->is_external) &&
355         node->io_poll_ready) {
356         /*
357          * Remove temporarily to avoid infinite loops when ->io_poll_ready()
358          * calls aio_poll() before clearing the condition that made the poll
359          * handler become ready.
360          */
361         QLIST_SAFE_REMOVE(node, node_poll);
362 
363         node->io_poll_ready(node->opaque);
364 
365         if (!QLIST_IS_INSERTED(node, node_poll)) {
366             QLIST_INSERT_HEAD(&ctx->poll_aio_handlers, node, node_poll);
367         }
368 
369         /*
370          * Return early since revents was zero. aio_notify() does not count as
371          * progress.
372          */
373         return node->opaque != &ctx->notifier;
374     }
375 
376     if (!QLIST_IS_INSERTED(node, node_deleted) &&
377         (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
378         aio_node_check(ctx, node->is_external) &&
379         node->io_read) {
380         node->io_read(node->opaque);
381 
382         /* aio_notify() does not count as progress */
383         if (node->opaque != &ctx->notifier) {
384             progress = true;
385         }
386     }
387     if (!QLIST_IS_INSERTED(node, node_deleted) &&
388         (revents & (G_IO_OUT | G_IO_ERR)) &&
389         aio_node_check(ctx, node->is_external) &&
390         node->io_write) {
391         node->io_write(node->opaque);
392         progress = true;
393     }
394 
395     return progress;
396 }
397 
398 /*
399  * If we have a list of ready handlers then this is more efficient than
400  * scanning all handlers with aio_dispatch_handlers().
401  */
402 static bool aio_dispatch_ready_handlers(AioContext *ctx,
403                                         AioHandlerList *ready_list)
404 {
405     bool progress = false;
406     AioHandler *node;
407 
408     while ((node = QLIST_FIRST(ready_list))) {
409         QLIST_REMOVE(node, node_ready);
410         progress = aio_dispatch_handler(ctx, node) || progress;
411     }
412 
413     return progress;
414 }
415 
416 /* Slower than aio_dispatch_ready_handlers() but only used via glib */
417 static bool aio_dispatch_handlers(AioContext *ctx)
418 {
419     AioHandler *node, *tmp;
420     bool progress = false;
421 
422     QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) {
423         progress = aio_dispatch_handler(ctx, node) || progress;
424     }
425 
426     return progress;
427 }
428 
429 void aio_dispatch(AioContext *ctx)
430 {
431     qemu_lockcnt_inc(&ctx->list_lock);
432     aio_bh_poll(ctx);
433     aio_dispatch_handlers(ctx);
434     aio_free_deleted_handlers(ctx);
435     qemu_lockcnt_dec(&ctx->list_lock);
436 
437     timerlistgroup_run_timers(&ctx->tlg);
438 }
439 
440 static bool run_poll_handlers_once(AioContext *ctx,
441                                    AioHandlerList *ready_list,
442                                    int64_t now,
443                                    int64_t *timeout)
444 {
445     bool progress = false;
446     AioHandler *node;
447     AioHandler *tmp;
448 
449     QLIST_FOREACH_SAFE(node, &ctx->poll_aio_handlers, node_poll, tmp) {
450         if (aio_node_check(ctx, node->is_external) &&
451             node->io_poll(node->opaque)) {
452             aio_add_poll_ready_handler(ready_list, node);
453 
454             node->poll_idle_timeout = now + POLL_IDLE_INTERVAL_NS;
455 
456             /*
457              * Polling was successful, exit try_poll_mode immediately
458              * to adjust the next polling time.
459              */
460             *timeout = 0;
461             if (node->opaque != &ctx->notifier) {
462                 progress = true;
463             }
464         }
465 
466         /* Caller handles freeing deleted nodes.  Don't do it here. */
467     }
468 
469     return progress;
470 }
471 
472 static bool fdmon_supports_polling(AioContext *ctx)
473 {
474     return ctx->fdmon_ops->need_wait != aio_poll_disabled;
475 }
476 
477 static bool remove_idle_poll_handlers(AioContext *ctx,
478                                       AioHandlerList *ready_list,
479                                       int64_t now)
480 {
481     AioHandler *node;
482     AioHandler *tmp;
483     bool progress = false;
484 
485     /*
486      * File descriptor monitoring implementations without userspace polling
487      * support suffer from starvation when a subset of handlers is polled
488      * because fds will not be processed in a timely fashion.  Don't remove
489      * idle poll handlers.
490      */
491     if (!fdmon_supports_polling(ctx)) {
492         return false;
493     }
494 
495     QLIST_FOREACH_SAFE(node, &ctx->poll_aio_handlers, node_poll, tmp) {
496         if (node->poll_idle_timeout == 0LL) {
497             node->poll_idle_timeout = now + POLL_IDLE_INTERVAL_NS;
498         } else if (now >= node->poll_idle_timeout) {
499             trace_poll_remove(ctx, node, node->pfd.fd);
500             node->poll_idle_timeout = 0LL;
501             QLIST_SAFE_REMOVE(node, node_poll);
502             if (ctx->poll_started && node->io_poll_end) {
503                 node->io_poll_end(node->opaque);
504 
505                 /*
506                  * Final poll in case ->io_poll_end() races with an event.
507                  * Nevermind about re-adding the handler in the rare case where
508                  * this causes progress.
509                  */
510                 if (node->io_poll(node->opaque)) {
511                     aio_add_poll_ready_handler(ready_list, node);
512                     progress = true;
513                 }
514             }
515         }
516     }
517 
518     return progress;
519 }
520 
521 /* run_poll_handlers:
522  * @ctx: the AioContext
523  * @ready_list: the list to place ready handlers on
524  * @max_ns: maximum time to poll for, in nanoseconds
525  *
526  * Polls for a given time.
527  *
528  * Note that the caller must have incremented ctx->list_lock.
529  *
530  * Returns: true if progress was made, false otherwise
531  */
532 static bool run_poll_handlers(AioContext *ctx, AioHandlerList *ready_list,
533                               int64_t max_ns, int64_t *timeout)
534 {
535     bool progress;
536     int64_t start_time, elapsed_time;
537 
538     assert(qemu_lockcnt_count(&ctx->list_lock) > 0);
539 
540     trace_run_poll_handlers_begin(ctx, max_ns, *timeout);
541 
542     /*
543      * Optimization: ->io_poll() handlers often contain RCU read critical
544      * sections and we therefore see many rcu_read_lock() -> rcu_read_unlock()
545      * -> rcu_read_lock() -> ... sequences with expensive memory
546      * synchronization primitives.  Make the entire polling loop an RCU
547      * critical section because nested rcu_read_lock()/rcu_read_unlock() calls
548      * are cheap.
549      */
550     RCU_READ_LOCK_GUARD();
551 
552     start_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
553     do {
554         progress = run_poll_handlers_once(ctx, ready_list,
555                                           start_time, timeout);
556         elapsed_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start_time;
557         max_ns = qemu_soonest_timeout(*timeout, max_ns);
558         assert(!(max_ns && progress));
559     } while (elapsed_time < max_ns && !ctx->fdmon_ops->need_wait(ctx));
560 
561     if (remove_idle_poll_handlers(ctx, ready_list,
562                                   start_time + elapsed_time)) {
563         *timeout = 0;
564         progress = true;
565     }
566 
567     /* If time has passed with no successful polling, adjust *timeout to
568      * keep the same ending time.
569      */
570     if (*timeout != -1) {
571         *timeout -= MIN(*timeout, elapsed_time);
572     }
573 
574     trace_run_poll_handlers_end(ctx, progress, *timeout);
575     return progress;
576 }
577 
578 /* try_poll_mode:
579  * @ctx: the AioContext
580  * @ready_list: list to add handlers that need to be run
581  * @timeout: timeout for blocking wait, computed by the caller and updated if
582  *    polling succeeds.
583  *
584  * Note that the caller must have incremented ctx->list_lock.
585  *
586  * Returns: true if progress was made, false otherwise
587  */
588 static bool try_poll_mode(AioContext *ctx, AioHandlerList *ready_list,
589                           int64_t *timeout)
590 {
591     int64_t max_ns;
592 
593     if (QLIST_EMPTY_RCU(&ctx->poll_aio_handlers)) {
594         return false;
595     }
596 
597     max_ns = qemu_soonest_timeout(*timeout, ctx->poll_ns);
598     if (max_ns && !ctx->fdmon_ops->need_wait(ctx)) {
599         /*
600          * Enable poll mode. It pairs with the poll_set_started() in
601          * aio_poll() which disables poll mode.
602          */
603         poll_set_started(ctx, ready_list, true);
604 
605         if (run_poll_handlers(ctx, ready_list, max_ns, timeout)) {
606             return true;
607         }
608     }
609     return false;
610 }
611 
612 bool aio_poll(AioContext *ctx, bool blocking)
613 {
614     AioHandlerList ready_list = QLIST_HEAD_INITIALIZER(ready_list);
615     bool progress;
616     bool use_notify_me;
617     int64_t timeout;
618     int64_t start = 0;
619 
620     /*
621      * There cannot be two concurrent aio_poll calls for the same AioContext (or
622      * an aio_poll concurrent with a GSource prepare/check/dispatch callback).
623      * We rely on this below to avoid slow locked accesses to ctx->notify_me.
624      *
625      * aio_poll() may only be called in the AioContext's thread. iohandler_ctx
626      * is special in that it runs in the main thread, but that thread's context
627      * is qemu_aio_context.
628      */
629     assert(in_aio_context_home_thread(ctx == iohandler_get_aio_context() ?
630                                       qemu_get_aio_context() : ctx));
631 
632     qemu_lockcnt_inc(&ctx->list_lock);
633 
634     if (ctx->poll_max_ns) {
635         start = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
636     }
637 
638     timeout = blocking ? aio_compute_timeout(ctx) : 0;
639     progress = try_poll_mode(ctx, &ready_list, &timeout);
640     assert(!(timeout && progress));
641 
642     /*
643      * aio_notify can avoid the expensive event_notifier_set if
644      * everything (file descriptors, bottom halves, timers) will
645      * be re-evaluated before the next blocking poll().  This is
646      * already true when aio_poll is called with blocking == false;
647      * if blocking == true, it is only true after poll() returns,
648      * so disable the optimization now.
649      */
650     use_notify_me = timeout != 0;
651     if (use_notify_me) {
652         qatomic_set(&ctx->notify_me, qatomic_read(&ctx->notify_me) + 2);
653         /*
654          * Write ctx->notify_me before reading ctx->notified.  Pairs with
655          * smp_mb in aio_notify().
656          */
657         smp_mb();
658 
659         /* Don't block if aio_notify() was called */
660         if (qatomic_read(&ctx->notified)) {
661             timeout = 0;
662         }
663     }
664 
665     /* If polling is allowed, non-blocking aio_poll does not need the
666      * system call---a single round of run_poll_handlers_once suffices.
667      */
668     if (timeout || ctx->fdmon_ops->need_wait(ctx)) {
669         /*
670          * Disable poll mode. poll mode should be disabled before the call
671          * of ctx->fdmon_ops->wait() so that guest's notification can wake
672          * up IO threads when some work becomes pending. It is essential to
673          * avoid hangs or unnecessary latency.
674          */
675         if (poll_set_started(ctx, &ready_list, false)) {
676             timeout = 0;
677             progress = true;
678         }
679 
680         ctx->fdmon_ops->wait(ctx, &ready_list, timeout);
681     }
682 
683     if (use_notify_me) {
684         /* Finish the poll before clearing the flag.  */
685         qatomic_store_release(&ctx->notify_me,
686                              qatomic_read(&ctx->notify_me) - 2);
687     }
688 
689     aio_notify_accept(ctx);
690 
691     /* Adjust polling time */
692     if (ctx->poll_max_ns) {
693         int64_t block_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start;
694 
695         if (block_ns <= ctx->poll_ns) {
696             /* This is the sweet spot, no adjustment needed */
697         } else if (block_ns > ctx->poll_max_ns) {
698             /* We'd have to poll for too long, poll less */
699             int64_t old = ctx->poll_ns;
700 
701             if (ctx->poll_shrink) {
702                 ctx->poll_ns /= ctx->poll_shrink;
703             } else {
704                 ctx->poll_ns = 0;
705             }
706 
707             trace_poll_shrink(ctx, old, ctx->poll_ns);
708         } else if (ctx->poll_ns < ctx->poll_max_ns &&
709                    block_ns < ctx->poll_max_ns) {
710             /* There is room to grow, poll longer */
711             int64_t old = ctx->poll_ns;
712             int64_t grow = ctx->poll_grow;
713 
714             if (grow == 0) {
715                 grow = 2;
716             }
717 
718             if (ctx->poll_ns) {
719                 ctx->poll_ns *= grow;
720             } else {
721                 ctx->poll_ns = 4000; /* start polling at 4 microseconds */
722             }
723 
724             if (ctx->poll_ns > ctx->poll_max_ns) {
725                 ctx->poll_ns = ctx->poll_max_ns;
726             }
727 
728             trace_poll_grow(ctx, old, ctx->poll_ns);
729         }
730     }
731 
732     progress |= aio_bh_poll(ctx);
733     progress |= aio_dispatch_ready_handlers(ctx, &ready_list);
734 
735     aio_free_deleted_handlers(ctx);
736 
737     qemu_lockcnt_dec(&ctx->list_lock);
738 
739     progress |= timerlistgroup_run_timers(&ctx->tlg);
740 
741     return progress;
742 }
743 
744 void aio_context_setup(AioContext *ctx)
745 {
746     ctx->fdmon_ops = &fdmon_poll_ops;
747     ctx->epollfd = -1;
748 
749     /* Use the fastest fd monitoring implementation if available */
750     if (fdmon_io_uring_setup(ctx)) {
751         return;
752     }
753 
754     fdmon_epoll_setup(ctx);
755 }
756 
757 void aio_context_destroy(AioContext *ctx)
758 {
759     fdmon_io_uring_destroy(ctx);
760     fdmon_epoll_disable(ctx);
761     aio_free_deleted_handlers(ctx);
762 }
763 
764 void aio_context_use_g_source(AioContext *ctx)
765 {
766     /*
767      * Disable io_uring when the glib main loop is used because it doesn't
768      * support mixed glib/aio_poll() usage. It relies on aio_poll() being
769      * called regularly so that changes to the monitored file descriptors are
770      * submitted, otherwise a list of pending fd handlers builds up.
771      */
772     fdmon_io_uring_destroy(ctx);
773     aio_free_deleted_handlers(ctx);
774 }
775 
776 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns,
777                                  int64_t grow, int64_t shrink, Error **errp)
778 {
779     /* No thread synchronization here, it doesn't matter if an incorrect value
780      * is used once.
781      */
782     ctx->poll_max_ns = max_ns;
783     ctx->poll_ns = 0;
784     ctx->poll_grow = grow;
785     ctx->poll_shrink = shrink;
786 
787     aio_notify(ctx);
788 }
789 
790 void aio_context_set_aio_params(AioContext *ctx, int64_t max_batch,
791                                 Error **errp)
792 {
793     /*
794      * No thread synchronization here, it doesn't matter if an incorrect value
795      * is used once.
796      */
797     ctx->aio_max_batch = max_batch;
798 
799     aio_notify(ctx);
800 }
801