xref: /qemu/util/aio-posix.c (revision e3a6e0da)
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 "qemu/rcu.h"
19 #include "qemu/rcu_queue.h"
20 #include "qemu/sockets.h"
21 #include "qemu/cutils.h"
22 #include "trace.h"
23 #include "aio-posix.h"
24 
25 /* Stop userspace polling on a handler if it isn't active for some time */
26 #define POLL_IDLE_INTERVAL_NS (7 * NANOSECONDS_PER_SECOND)
27 
28 bool aio_poll_disabled(AioContext *ctx)
29 {
30     return atomic_read(&ctx->poll_disable_cnt);
31 }
32 
33 void aio_add_ready_handler(AioHandlerList *ready_list,
34                            AioHandler *node,
35                            int revents)
36 {
37     QLIST_SAFE_REMOVE(node, node_ready); /* remove from nested parent's list */
38     node->pfd.revents = revents;
39     QLIST_INSERT_HEAD(ready_list, node, node_ready);
40 }
41 
42 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
43 {
44     AioHandler *node;
45 
46     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
47         if (node->pfd.fd == fd) {
48             if (!QLIST_IS_INSERTED(node, node_deleted)) {
49                 return node;
50             }
51         }
52     }
53 
54     return NULL;
55 }
56 
57 static bool aio_remove_fd_handler(AioContext *ctx, AioHandler *node)
58 {
59     /* If the GSource is in the process of being destroyed then
60      * g_source_remove_poll() causes an assertion failure.  Skip
61      * removal in that case, because glib cleans up its state during
62      * destruction anyway.
63      */
64     if (!g_source_is_destroyed(&ctx->source)) {
65         g_source_remove_poll(&ctx->source, &node->pfd);
66     }
67 
68     node->pfd.revents = 0;
69 
70     /* If the fd monitor has already marked it deleted, leave it alone */
71     if (QLIST_IS_INSERTED(node, node_deleted)) {
72         return false;
73     }
74 
75     /* If a read is in progress, just mark the node as deleted */
76     if (qemu_lockcnt_count(&ctx->list_lock)) {
77         QLIST_INSERT_HEAD_RCU(&ctx->deleted_aio_handlers, node, node_deleted);
78         return false;
79     }
80     /* Otherwise, delete it for real.  We can't just mark it as
81      * deleted because deleted nodes are only cleaned up while
82      * no one is walking the handlers list.
83      */
84     QLIST_SAFE_REMOVE(node, node_poll);
85     QLIST_REMOVE(node, node);
86     return true;
87 }
88 
89 void aio_set_fd_handler(AioContext *ctx,
90                         int fd,
91                         bool is_external,
92                         IOHandler *io_read,
93                         IOHandler *io_write,
94                         AioPollFn *io_poll,
95                         void *opaque)
96 {
97     AioHandler *node;
98     AioHandler *new_node = NULL;
99     bool is_new = false;
100     bool deleted = false;
101     int poll_disable_change;
102 
103     qemu_lockcnt_lock(&ctx->list_lock);
104 
105     node = find_aio_handler(ctx, fd);
106 
107     /* Are we deleting the fd handler? */
108     if (!io_read && !io_write && !io_poll) {
109         if (node == NULL) {
110             qemu_lockcnt_unlock(&ctx->list_lock);
111             return;
112         }
113         /* Clean events in order to unregister fd from the ctx epoll. */
114         node->pfd.events = 0;
115 
116         poll_disable_change = -!node->io_poll;
117     } else {
118         poll_disable_change = !io_poll - (node && !node->io_poll);
119         if (node == NULL) {
120             is_new = true;
121         }
122         /* Alloc and insert if it's not already there */
123         new_node = g_new0(AioHandler, 1);
124 
125         /* Update handler with latest information */
126         new_node->io_read = io_read;
127         new_node->io_write = io_write;
128         new_node->io_poll = io_poll;
129         new_node->opaque = opaque;
130         new_node->is_external = is_external;
131 
132         if (is_new) {
133             new_node->pfd.fd = fd;
134         } else {
135             new_node->pfd = node->pfd;
136         }
137         g_source_add_poll(&ctx->source, &new_node->pfd);
138 
139         new_node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
140         new_node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
141 
142         QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, new_node, node);
143     }
144 
145     /* No need to order poll_disable_cnt writes against other updates;
146      * the counter is only used to avoid wasting time and latency on
147      * iterated polling when the system call will be ultimately necessary.
148      * Changing handlers is a rare event, and a little wasted polling until
149      * the aio_notify below is not an issue.
150      */
151     atomic_set(&ctx->poll_disable_cnt,
152                atomic_read(&ctx->poll_disable_cnt) + poll_disable_change);
153 
154     ctx->fdmon_ops->update(ctx, node, new_node);
155     if (node) {
156         deleted = aio_remove_fd_handler(ctx, node);
157     }
158     qemu_lockcnt_unlock(&ctx->list_lock);
159     aio_notify(ctx);
160 
161     if (deleted) {
162         g_free(node);
163     }
164 }
165 
166 void aio_set_fd_poll(AioContext *ctx, int fd,
167                      IOHandler *io_poll_begin,
168                      IOHandler *io_poll_end)
169 {
170     AioHandler *node = find_aio_handler(ctx, fd);
171 
172     if (!node) {
173         return;
174     }
175 
176     node->io_poll_begin = io_poll_begin;
177     node->io_poll_end = io_poll_end;
178 }
179 
180 void aio_set_event_notifier(AioContext *ctx,
181                             EventNotifier *notifier,
182                             bool is_external,
183                             EventNotifierHandler *io_read,
184                             AioPollFn *io_poll)
185 {
186     aio_set_fd_handler(ctx, event_notifier_get_fd(notifier), is_external,
187                        (IOHandler *)io_read, NULL, io_poll, notifier);
188 }
189 
190 void aio_set_event_notifier_poll(AioContext *ctx,
191                                  EventNotifier *notifier,
192                                  EventNotifierHandler *io_poll_begin,
193                                  EventNotifierHandler *io_poll_end)
194 {
195     aio_set_fd_poll(ctx, event_notifier_get_fd(notifier),
196                     (IOHandler *)io_poll_begin,
197                     (IOHandler *)io_poll_end);
198 }
199 
200 static bool poll_set_started(AioContext *ctx, bool started)
201 {
202     AioHandler *node;
203     bool progress = false;
204 
205     if (started == ctx->poll_started) {
206         return false;
207     }
208 
209     ctx->poll_started = started;
210 
211     qemu_lockcnt_inc(&ctx->list_lock);
212     QLIST_FOREACH(node, &ctx->poll_aio_handlers, node_poll) {
213         IOHandler *fn;
214 
215         if (QLIST_IS_INSERTED(node, node_deleted)) {
216             continue;
217         }
218 
219         if (started) {
220             fn = node->io_poll_begin;
221         } else {
222             fn = node->io_poll_end;
223         }
224 
225         if (fn) {
226             fn(node->opaque);
227         }
228 
229         /* Poll one last time in case ->io_poll_end() raced with the event */
230         if (!started) {
231             progress = node->io_poll(node->opaque) || progress;
232         }
233     }
234     qemu_lockcnt_dec(&ctx->list_lock);
235 
236     return progress;
237 }
238 
239 
240 bool aio_prepare(AioContext *ctx)
241 {
242     /* Poll mode cannot be used with glib's event loop, disable it. */
243     poll_set_started(ctx, false);
244 
245     return false;
246 }
247 
248 bool aio_pending(AioContext *ctx)
249 {
250     AioHandler *node;
251     bool result = false;
252 
253     /*
254      * We have to walk very carefully in case aio_set_fd_handler is
255      * called while we're walking.
256      */
257     qemu_lockcnt_inc(&ctx->list_lock);
258 
259     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
260         int revents;
261 
262         revents = node->pfd.revents & node->pfd.events;
263         if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read &&
264             aio_node_check(ctx, node->is_external)) {
265             result = true;
266             break;
267         }
268         if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write &&
269             aio_node_check(ctx, node->is_external)) {
270             result = true;
271             break;
272         }
273     }
274     qemu_lockcnt_dec(&ctx->list_lock);
275 
276     return result;
277 }
278 
279 static void aio_free_deleted_handlers(AioContext *ctx)
280 {
281     AioHandler *node;
282 
283     if (QLIST_EMPTY_RCU(&ctx->deleted_aio_handlers)) {
284         return;
285     }
286     if (!qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
287         return; /* we are nested, let the parent do the freeing */
288     }
289 
290     while ((node = QLIST_FIRST_RCU(&ctx->deleted_aio_handlers))) {
291         QLIST_REMOVE(node, node);
292         QLIST_REMOVE(node, node_deleted);
293         QLIST_SAFE_REMOVE(node, node_poll);
294         g_free(node);
295     }
296 
297     qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
298 }
299 
300 static bool aio_dispatch_handler(AioContext *ctx, AioHandler *node)
301 {
302     bool progress = false;
303     int revents;
304 
305     revents = node->pfd.revents & node->pfd.events;
306     node->pfd.revents = 0;
307 
308     /*
309      * Start polling AioHandlers when they become ready because activity is
310      * likely to continue.  Note that starvation is theoretically possible when
311      * fdmon_supports_polling(), but only until the fd fires for the first
312      * time.
313      */
314     if (!QLIST_IS_INSERTED(node, node_deleted) &&
315         !QLIST_IS_INSERTED(node, node_poll) &&
316         node->io_poll) {
317         trace_poll_add(ctx, node, node->pfd.fd, revents);
318         if (ctx->poll_started && node->io_poll_begin) {
319             node->io_poll_begin(node->opaque);
320         }
321         QLIST_INSERT_HEAD(&ctx->poll_aio_handlers, node, node_poll);
322     }
323 
324     if (!QLIST_IS_INSERTED(node, node_deleted) &&
325         (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
326         aio_node_check(ctx, node->is_external) &&
327         node->io_read) {
328         node->io_read(node->opaque);
329 
330         /* aio_notify() does not count as progress */
331         if (node->opaque != &ctx->notifier) {
332             progress = true;
333         }
334     }
335     if (!QLIST_IS_INSERTED(node, node_deleted) &&
336         (revents & (G_IO_OUT | G_IO_ERR)) &&
337         aio_node_check(ctx, node->is_external) &&
338         node->io_write) {
339         node->io_write(node->opaque);
340         progress = true;
341     }
342 
343     return progress;
344 }
345 
346 /*
347  * If we have a list of ready handlers then this is more efficient than
348  * scanning all handlers with aio_dispatch_handlers().
349  */
350 static bool aio_dispatch_ready_handlers(AioContext *ctx,
351                                         AioHandlerList *ready_list)
352 {
353     bool progress = false;
354     AioHandler *node;
355 
356     while ((node = QLIST_FIRST(ready_list))) {
357         QLIST_REMOVE(node, node_ready);
358         progress = aio_dispatch_handler(ctx, node) || progress;
359     }
360 
361     return progress;
362 }
363 
364 /* Slower than aio_dispatch_ready_handlers() but only used via glib */
365 static bool aio_dispatch_handlers(AioContext *ctx)
366 {
367     AioHandler *node, *tmp;
368     bool progress = false;
369 
370     QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) {
371         progress = aio_dispatch_handler(ctx, node) || progress;
372     }
373 
374     return progress;
375 }
376 
377 void aio_dispatch(AioContext *ctx)
378 {
379     qemu_lockcnt_inc(&ctx->list_lock);
380     aio_bh_poll(ctx);
381     aio_dispatch_handlers(ctx);
382     aio_free_deleted_handlers(ctx);
383     qemu_lockcnt_dec(&ctx->list_lock);
384 
385     timerlistgroup_run_timers(&ctx->tlg);
386 }
387 
388 static bool run_poll_handlers_once(AioContext *ctx,
389                                    int64_t now,
390                                    int64_t *timeout)
391 {
392     bool progress = false;
393     AioHandler *node;
394     AioHandler *tmp;
395 
396     QLIST_FOREACH_SAFE(node, &ctx->poll_aio_handlers, node_poll, tmp) {
397         if (aio_node_check(ctx, node->is_external) &&
398             node->io_poll(node->opaque)) {
399             node->poll_idle_timeout = now + POLL_IDLE_INTERVAL_NS;
400 
401             /*
402              * Polling was successful, exit try_poll_mode immediately
403              * to adjust the next polling time.
404              */
405             *timeout = 0;
406             if (node->opaque != &ctx->notifier) {
407                 progress = true;
408             }
409         }
410 
411         /* Caller handles freeing deleted nodes.  Don't do it here. */
412     }
413 
414     return progress;
415 }
416 
417 static bool fdmon_supports_polling(AioContext *ctx)
418 {
419     return ctx->fdmon_ops->need_wait != aio_poll_disabled;
420 }
421 
422 static bool remove_idle_poll_handlers(AioContext *ctx, int64_t now)
423 {
424     AioHandler *node;
425     AioHandler *tmp;
426     bool progress = false;
427 
428     /*
429      * File descriptor monitoring implementations without userspace polling
430      * support suffer from starvation when a subset of handlers is polled
431      * because fds will not be processed in a timely fashion.  Don't remove
432      * idle poll handlers.
433      */
434     if (!fdmon_supports_polling(ctx)) {
435         return false;
436     }
437 
438     QLIST_FOREACH_SAFE(node, &ctx->poll_aio_handlers, node_poll, tmp) {
439         if (node->poll_idle_timeout == 0LL) {
440             node->poll_idle_timeout = now + POLL_IDLE_INTERVAL_NS;
441         } else if (now >= node->poll_idle_timeout) {
442             trace_poll_remove(ctx, node, node->pfd.fd);
443             node->poll_idle_timeout = 0LL;
444             QLIST_SAFE_REMOVE(node, node_poll);
445             if (ctx->poll_started && node->io_poll_end) {
446                 node->io_poll_end(node->opaque);
447 
448                 /*
449                  * Final poll in case ->io_poll_end() races with an event.
450                  * Nevermind about re-adding the handler in the rare case where
451                  * this causes progress.
452                  */
453                 progress = node->io_poll(node->opaque) || progress;
454             }
455         }
456     }
457 
458     return progress;
459 }
460 
461 /* run_poll_handlers:
462  * @ctx: the AioContext
463  * @max_ns: maximum time to poll for, in nanoseconds
464  *
465  * Polls for a given time.
466  *
467  * Note that the caller must have incremented ctx->list_lock.
468  *
469  * Returns: true if progress was made, false otherwise
470  */
471 static bool run_poll_handlers(AioContext *ctx, int64_t max_ns, int64_t *timeout)
472 {
473     bool progress;
474     int64_t start_time, elapsed_time;
475 
476     assert(qemu_lockcnt_count(&ctx->list_lock) > 0);
477 
478     trace_run_poll_handlers_begin(ctx, max_ns, *timeout);
479 
480     /*
481      * Optimization: ->io_poll() handlers often contain RCU read critical
482      * sections and we therefore see many rcu_read_lock() -> rcu_read_unlock()
483      * -> rcu_read_lock() -> ... sequences with expensive memory
484      * synchronization primitives.  Make the entire polling loop an RCU
485      * critical section because nested rcu_read_lock()/rcu_read_unlock() calls
486      * are cheap.
487      */
488     RCU_READ_LOCK_GUARD();
489 
490     start_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
491     do {
492         progress = run_poll_handlers_once(ctx, start_time, timeout);
493         elapsed_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start_time;
494         max_ns = qemu_soonest_timeout(*timeout, max_ns);
495         assert(!(max_ns && progress));
496     } while (elapsed_time < max_ns && !ctx->fdmon_ops->need_wait(ctx));
497 
498     if (remove_idle_poll_handlers(ctx, start_time + elapsed_time)) {
499         *timeout = 0;
500         progress = true;
501     }
502 
503     /* If time has passed with no successful polling, adjust *timeout to
504      * keep the same ending time.
505      */
506     if (*timeout != -1) {
507         *timeout -= MIN(*timeout, elapsed_time);
508     }
509 
510     trace_run_poll_handlers_end(ctx, progress, *timeout);
511     return progress;
512 }
513 
514 /* try_poll_mode:
515  * @ctx: the AioContext
516  * @timeout: timeout for blocking wait, computed by the caller and updated if
517  *    polling succeeds.
518  *
519  * Note that the caller must have incremented ctx->list_lock.
520  *
521  * Returns: true if progress was made, false otherwise
522  */
523 static bool try_poll_mode(AioContext *ctx, int64_t *timeout)
524 {
525     int64_t max_ns;
526 
527     if (QLIST_EMPTY_RCU(&ctx->poll_aio_handlers)) {
528         return false;
529     }
530 
531     max_ns = qemu_soonest_timeout(*timeout, ctx->poll_ns);
532     if (max_ns && !ctx->fdmon_ops->need_wait(ctx)) {
533         poll_set_started(ctx, true);
534 
535         if (run_poll_handlers(ctx, max_ns, timeout)) {
536             return true;
537         }
538     }
539 
540     if (poll_set_started(ctx, false)) {
541         *timeout = 0;
542         return true;
543     }
544 
545     return false;
546 }
547 
548 bool aio_poll(AioContext *ctx, bool blocking)
549 {
550     AioHandlerList ready_list = QLIST_HEAD_INITIALIZER(ready_list);
551     int ret = 0;
552     bool progress;
553     bool use_notify_me;
554     int64_t timeout;
555     int64_t start = 0;
556 
557     /*
558      * There cannot be two concurrent aio_poll calls for the same AioContext (or
559      * an aio_poll concurrent with a GSource prepare/check/dispatch callback).
560      * We rely on this below to avoid slow locked accesses to ctx->notify_me.
561      */
562     assert(in_aio_context_home_thread(ctx));
563 
564     qemu_lockcnt_inc(&ctx->list_lock);
565 
566     if (ctx->poll_max_ns) {
567         start = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
568     }
569 
570     timeout = blocking ? aio_compute_timeout(ctx) : 0;
571     progress = try_poll_mode(ctx, &timeout);
572     assert(!(timeout && progress));
573 
574     /*
575      * aio_notify can avoid the expensive event_notifier_set if
576      * everything (file descriptors, bottom halves, timers) will
577      * be re-evaluated before the next blocking poll().  This is
578      * already true when aio_poll is called with blocking == false;
579      * if blocking == true, it is only true after poll() returns,
580      * so disable the optimization now.
581      */
582     use_notify_me = timeout != 0;
583     if (use_notify_me) {
584         atomic_set(&ctx->notify_me, atomic_read(&ctx->notify_me) + 2);
585         /*
586          * Write ctx->notify_me before reading ctx->notified.  Pairs with
587          * smp_mb in aio_notify().
588          */
589         smp_mb();
590 
591         /* Don't block if aio_notify() was called */
592         if (atomic_read(&ctx->notified)) {
593             timeout = 0;
594         }
595     }
596 
597     /* If polling is allowed, non-blocking aio_poll does not need the
598      * system call---a single round of run_poll_handlers_once suffices.
599      */
600     if (timeout || ctx->fdmon_ops->need_wait(ctx)) {
601         ret = ctx->fdmon_ops->wait(ctx, &ready_list, timeout);
602     }
603 
604     if (use_notify_me) {
605         /* Finish the poll before clearing the flag.  */
606         atomic_store_release(&ctx->notify_me,
607                              atomic_read(&ctx->notify_me) - 2);
608     }
609 
610     aio_notify_accept(ctx);
611 
612     /* Adjust polling time */
613     if (ctx->poll_max_ns) {
614         int64_t block_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start;
615 
616         if (block_ns <= ctx->poll_ns) {
617             /* This is the sweet spot, no adjustment needed */
618         } else if (block_ns > ctx->poll_max_ns) {
619             /* We'd have to poll for too long, poll less */
620             int64_t old = ctx->poll_ns;
621 
622             if (ctx->poll_shrink) {
623                 ctx->poll_ns /= ctx->poll_shrink;
624             } else {
625                 ctx->poll_ns = 0;
626             }
627 
628             trace_poll_shrink(ctx, old, ctx->poll_ns);
629         } else if (ctx->poll_ns < ctx->poll_max_ns &&
630                    block_ns < ctx->poll_max_ns) {
631             /* There is room to grow, poll longer */
632             int64_t old = ctx->poll_ns;
633             int64_t grow = ctx->poll_grow;
634 
635             if (grow == 0) {
636                 grow = 2;
637             }
638 
639             if (ctx->poll_ns) {
640                 ctx->poll_ns *= grow;
641             } else {
642                 ctx->poll_ns = 4000; /* start polling at 4 microseconds */
643             }
644 
645             if (ctx->poll_ns > ctx->poll_max_ns) {
646                 ctx->poll_ns = ctx->poll_max_ns;
647             }
648 
649             trace_poll_grow(ctx, old, ctx->poll_ns);
650         }
651     }
652 
653     progress |= aio_bh_poll(ctx);
654 
655     if (ret > 0) {
656         progress |= aio_dispatch_ready_handlers(ctx, &ready_list);
657     }
658 
659     aio_free_deleted_handlers(ctx);
660 
661     qemu_lockcnt_dec(&ctx->list_lock);
662 
663     progress |= timerlistgroup_run_timers(&ctx->tlg);
664 
665     return progress;
666 }
667 
668 void aio_context_setup(AioContext *ctx)
669 {
670     ctx->fdmon_ops = &fdmon_poll_ops;
671     ctx->epollfd = -1;
672 
673     /* Use the fastest fd monitoring implementation if available */
674     if (fdmon_io_uring_setup(ctx)) {
675         return;
676     }
677 
678     fdmon_epoll_setup(ctx);
679 }
680 
681 void aio_context_destroy(AioContext *ctx)
682 {
683     fdmon_io_uring_destroy(ctx);
684     fdmon_epoll_disable(ctx);
685     aio_free_deleted_handlers(ctx);
686 }
687 
688 void aio_context_use_g_source(AioContext *ctx)
689 {
690     /*
691      * Disable io_uring when the glib main loop is used because it doesn't
692      * support mixed glib/aio_poll() usage. It relies on aio_poll() being
693      * called regularly so that changes to the monitored file descriptors are
694      * submitted, otherwise a list of pending fd handlers builds up.
695      */
696     fdmon_io_uring_destroy(ctx);
697     aio_free_deleted_handlers(ctx);
698 }
699 
700 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns,
701                                  int64_t grow, int64_t shrink, Error **errp)
702 {
703     /* No thread synchronization here, it doesn't matter if an incorrect value
704      * is used once.
705      */
706     ctx->poll_max_ns = max_ns;
707     ctx->poll_ns = 0;
708     ctx->poll_grow = grow;
709     ctx->poll_shrink = shrink;
710 
711     aio_notify(ctx);
712 }
713