2 * QEMU aio implementation
4 * Copyright IBM, Corp. 2008
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
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.
16 #include "qemu/osdep.h"
17 #include "qemu-common.h"
18 #include "block/block.h"
19 #include "qemu/queue.h"
20 #include "qemu/sockets.h"
21 #include "qemu/cutils.h"
23 #ifdef CONFIG_EPOLL_CREATE1
24 #include <sys/epoll.h>
33 IOHandler *io_poll_begin;
34 IOHandler *io_poll_end;
38 QLIST_ENTRY(AioHandler) node;
41 #ifdef CONFIG_EPOLL_CREATE1
43 /* The fd number threashold to switch to epoll */
44 #define EPOLL_ENABLE_THRESHOLD 64
46 static void aio_epoll_disable(AioContext *ctx)
48 ctx->epoll_available = false;
49 if (!ctx->epoll_enabled) {
52 ctx->epoll_enabled = false;
56 static inline int epoll_events_from_pfd(int pfd_events)
58 return (pfd_events & G_IO_IN ? EPOLLIN : 0) |
59 (pfd_events & G_IO_OUT ? EPOLLOUT : 0) |
60 (pfd_events & G_IO_HUP ? EPOLLHUP : 0) |
61 (pfd_events & G_IO_ERR ? EPOLLERR : 0);
64 static bool aio_epoll_try_enable(AioContext *ctx)
67 struct epoll_event event;
69 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
71 if (node->deleted || !node->pfd.events) {
74 event.events = epoll_events_from_pfd(node->pfd.events);
75 event.data.ptr = node;
76 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
81 ctx->epoll_enabled = true;
85 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
87 struct epoll_event event;
91 if (!ctx->epoll_enabled) {
94 if (!node->pfd.events) {
97 event.data.ptr = node;
98 event.events = epoll_events_from_pfd(node->pfd.events);
99 ctl = is_new ? EPOLL_CTL_ADD : EPOLL_CTL_MOD;
102 r = epoll_ctl(ctx->epollfd, ctl, node->pfd.fd, &event);
104 aio_epoll_disable(ctx);
108 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
109 unsigned npfd, int64_t timeout)
113 struct epoll_event events[128];
116 assert(pfds[0].fd == ctx->epollfd);
118 ret = qemu_poll_ns(pfds, npfd, timeout);
120 if (timeout <= 0 || ret > 0) {
121 ret = epoll_wait(ctx->epollfd, events,
122 sizeof(events) / sizeof(events[0]),
127 for (i = 0; i < ret; i++) {
128 int ev = events[i].events;
129 node = events[i].data.ptr;
130 node->pfd.revents = (ev & EPOLLIN ? G_IO_IN : 0) |
131 (ev & EPOLLOUT ? G_IO_OUT : 0) |
132 (ev & EPOLLHUP ? G_IO_HUP : 0) |
133 (ev & EPOLLERR ? G_IO_ERR : 0);
140 static bool aio_epoll_enabled(AioContext *ctx)
142 /* Fall back to ppoll when external clients are disabled. */
143 return !aio_external_disabled(ctx) && ctx->epoll_enabled;
146 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
147 unsigned npfd, int64_t timeout)
149 if (!ctx->epoll_available) {
152 if (aio_epoll_enabled(ctx)) {
155 if (npfd >= EPOLL_ENABLE_THRESHOLD) {
156 if (aio_epoll_try_enable(ctx)) {
159 aio_epoll_disable(ctx);
167 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
171 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
172 unsigned npfd, int64_t timeout)
177 static bool aio_epoll_enabled(AioContext *ctx)
182 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
183 unsigned npfd, int64_t timeout)
190 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
194 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
195 if (node->pfd.fd == fd)
203 void aio_set_fd_handler(AioContext *ctx,
213 bool deleted = false;
215 node = find_aio_handler(ctx, fd);
217 /* Are we deleting the fd handler? */
218 if (!io_read && !io_write && !io_poll) {
223 g_source_remove_poll(&ctx->source, &node->pfd);
225 /* If the lock is held, just mark the node as deleted */
226 if (ctx->walking_handlers) {
228 node->pfd.revents = 0;
230 /* Otherwise, delete it for real. We can't just mark it as
231 * deleted because deleted nodes are only cleaned up after
232 * releasing the walking_handlers lock.
234 QLIST_REMOVE(node, node);
238 if (!node->io_poll) {
239 ctx->poll_disable_cnt--;
243 /* Alloc and insert if it's not already there */
244 node = g_new0(AioHandler, 1);
246 QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
248 g_source_add_poll(&ctx->source, &node->pfd);
251 ctx->poll_disable_cnt += !io_poll;
253 ctx->poll_disable_cnt += !io_poll - !node->io_poll;
256 /* Update handler with latest information */
257 node->io_read = io_read;
258 node->io_write = io_write;
259 node->io_poll = io_poll;
260 node->opaque = opaque;
261 node->is_external = is_external;
263 node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
264 node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
267 aio_epoll_update(ctx, node, is_new);
275 void aio_set_fd_poll(AioContext *ctx, int fd,
276 IOHandler *io_poll_begin,
277 IOHandler *io_poll_end)
279 AioHandler *node = find_aio_handler(ctx, fd);
285 node->io_poll_begin = io_poll_begin;
286 node->io_poll_end = io_poll_end;
289 void aio_set_event_notifier(AioContext *ctx,
290 EventNotifier *notifier,
292 EventNotifierHandler *io_read,
295 aio_set_fd_handler(ctx, event_notifier_get_fd(notifier), is_external,
296 (IOHandler *)io_read, NULL, io_poll, notifier);
299 void aio_set_event_notifier_poll(AioContext *ctx,
300 EventNotifier *notifier,
301 EventNotifierHandler *io_poll_begin,
302 EventNotifierHandler *io_poll_end)
304 aio_set_fd_poll(ctx, event_notifier_get_fd(notifier),
305 (IOHandler *)io_poll_begin,
306 (IOHandler *)io_poll_end);
309 static void poll_set_started(AioContext *ctx, bool started)
313 if (started == ctx->poll_started) {
317 ctx->poll_started = started;
319 ctx->walking_handlers++;
320 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
328 fn = node->io_poll_begin;
330 fn = node->io_poll_end;
337 ctx->walking_handlers--;
341 bool aio_prepare(AioContext *ctx)
343 /* Poll mode cannot be used with glib's event loop, disable it. */
344 poll_set_started(ctx, false);
349 bool aio_pending(AioContext *ctx)
353 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
356 revents = node->pfd.revents & node->pfd.events;
357 if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read &&
358 aio_node_check(ctx, node->is_external)) {
361 if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write &&
362 aio_node_check(ctx, node->is_external)) {
371 * Note that dispatch_fds == false has the side-effect of post-poning the
372 * freeing of deleted handlers.
374 bool aio_dispatch(AioContext *ctx, bool dispatch_fds)
376 AioHandler *node = NULL;
377 bool progress = false;
380 * If there are callbacks left that have been queued, we need to call them.
381 * Do not call select in this case, because it is possible that the caller
382 * does not need a complete flush (as is the case for aio_poll loops).
384 if (aio_bh_poll(ctx)) {
389 * We have to walk very carefully in case aio_set_fd_handler is
390 * called while we're walking.
393 node = QLIST_FIRST(&ctx->aio_handlers);
399 ctx->walking_handlers++;
401 revents = node->pfd.revents & node->pfd.events;
402 node->pfd.revents = 0;
404 if (!node->deleted &&
405 (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
406 aio_node_check(ctx, node->is_external) &&
408 node->io_read(node->opaque);
410 /* aio_notify() does not count as progress */
411 if (node->opaque != &ctx->notifier) {
415 if (!node->deleted &&
416 (revents & (G_IO_OUT | G_IO_ERR)) &&
417 aio_node_check(ctx, node->is_external) &&
419 node->io_write(node->opaque);
424 node = QLIST_NEXT(node, node);
426 ctx->walking_handlers--;
428 if (!ctx->walking_handlers && tmp->deleted) {
429 QLIST_REMOVE(tmp, node);
435 progress |= timerlistgroup_run_timers(&ctx->tlg);
440 /* These thread-local variables are used only in a small part of aio_poll
441 * around the call to the poll() system call. In particular they are not
442 * used while aio_poll is performing callbacks, which makes it much easier
443 * to think about reentrancy!
445 * Stack-allocated arrays would be perfect but they have size limitations;
446 * heap allocation is expensive enough that we want to reuse arrays across
447 * calls to aio_poll(). And because poll() has to be called without holding
448 * any lock, the arrays cannot be stored in AioContext. Thread-local data
449 * has none of the disadvantages of these three options.
451 static __thread GPollFD *pollfds;
452 static __thread AioHandler **nodes;
453 static __thread unsigned npfd, nalloc;
454 static __thread Notifier pollfds_cleanup_notifier;
456 static void pollfds_cleanup(Notifier *n, void *unused)
464 static void add_pollfd(AioHandler *node)
466 if (npfd == nalloc) {
468 pollfds_cleanup_notifier.notify = pollfds_cleanup;
469 qemu_thread_atexit_add(&pollfds_cleanup_notifier);
472 g_assert(nalloc <= INT_MAX);
475 pollfds = g_renew(GPollFD, pollfds, nalloc);
476 nodes = g_renew(AioHandler *, nodes, nalloc);
479 pollfds[npfd] = (GPollFD) {
481 .events = node->pfd.events,
486 static bool run_poll_handlers_once(AioContext *ctx)
488 bool progress = false;
491 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
492 if (!node->deleted && node->io_poll &&
493 node->io_poll(node->opaque)) {
497 /* Caller handles freeing deleted nodes. Don't do it here. */
503 /* run_poll_handlers:
504 * @ctx: the AioContext
505 * @max_ns: maximum time to poll for, in nanoseconds
507 * Polls for a given time.
509 * Note that ctx->notify_me must be non-zero so this function can detect
512 * Note that the caller must have incremented ctx->walking_handlers.
514 * Returns: true if progress was made, false otherwise
516 static bool run_poll_handlers(AioContext *ctx, int64_t max_ns)
521 assert(ctx->notify_me);
522 assert(ctx->walking_handlers > 0);
523 assert(ctx->poll_disable_cnt == 0);
525 trace_run_poll_handlers_begin(ctx, max_ns);
527 end_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) + max_ns;
530 progress = run_poll_handlers_once(ctx);
531 } while (!progress && qemu_clock_get_ns(QEMU_CLOCK_REALTIME) < end_time);
533 trace_run_poll_handlers_end(ctx, progress);
539 * @ctx: the AioContext
540 * @blocking: busy polling is only attempted when blocking is true
542 * ctx->notify_me must be non-zero so this function can detect aio_notify().
544 * Note that the caller must have incremented ctx->walking_handlers.
546 * Returns: true if progress was made, false otherwise
548 static bool try_poll_mode(AioContext *ctx, bool blocking)
550 if (blocking && ctx->poll_max_ns && ctx->poll_disable_cnt == 0) {
551 /* See qemu_soonest_timeout() uint64_t hack */
552 int64_t max_ns = MIN((uint64_t)aio_compute_timeout(ctx),
553 (uint64_t)ctx->poll_ns);
556 poll_set_started(ctx, true);
558 if (run_poll_handlers(ctx, max_ns)) {
564 poll_set_started(ctx, false);
566 /* Even if we don't run busy polling, try polling once in case it can make
567 * progress and the caller will be able to avoid ppoll(2)/epoll_wait(2).
569 return run_poll_handlers_once(ctx);
572 bool aio_poll(AioContext *ctx, bool blocking)
581 aio_context_acquire(ctx);
584 /* aio_notify can avoid the expensive event_notifier_set if
585 * everything (file descriptors, bottom halves, timers) will
586 * be re-evaluated before the next blocking poll(). This is
587 * already true when aio_poll is called with blocking == false;
588 * if blocking == true, it is only true after poll() returns,
589 * so disable the optimization now.
592 atomic_add(&ctx->notify_me, 2);
595 ctx->walking_handlers++;
597 if (ctx->poll_max_ns) {
598 start = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
601 if (try_poll_mode(ctx, blocking)) {
608 if (!aio_epoll_enabled(ctx)) {
609 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
610 if (!node->deleted && node->pfd.events
611 && aio_node_check(ctx, node->is_external)) {
617 timeout = blocking ? aio_compute_timeout(ctx) : 0;
619 /* wait until next event */
621 aio_context_release(ctx);
623 if (aio_epoll_check_poll(ctx, pollfds, npfd, timeout)) {
624 AioHandler epoll_handler;
626 epoll_handler.pfd.fd = ctx->epollfd;
627 epoll_handler.pfd.events = G_IO_IN | G_IO_OUT | G_IO_HUP | G_IO_ERR;
629 add_pollfd(&epoll_handler);
630 ret = aio_epoll(ctx, pollfds, npfd, timeout);
632 ret = qemu_poll_ns(pollfds, npfd, timeout);
635 aio_context_acquire(ctx);
640 atomic_sub(&ctx->notify_me, 2);
643 /* Adjust polling time */
644 if (ctx->poll_max_ns) {
645 int64_t block_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start;
647 if (block_ns <= ctx->poll_ns) {
648 /* This is the sweet spot, no adjustment needed */
649 } else if (block_ns > ctx->poll_max_ns) {
650 /* We'd have to poll for too long, poll less */
651 int64_t old = ctx->poll_ns;
653 if (ctx->poll_shrink) {
654 ctx->poll_ns /= ctx->poll_shrink;
659 trace_poll_shrink(ctx, old, ctx->poll_ns);
660 } else if (ctx->poll_ns < ctx->poll_max_ns &&
661 block_ns < ctx->poll_max_ns) {
662 /* There is room to grow, poll longer */
663 int64_t old = ctx->poll_ns;
664 int64_t grow = ctx->poll_grow;
671 ctx->poll_ns *= grow;
673 ctx->poll_ns = 4000; /* start polling at 4 microseconds */
676 if (ctx->poll_ns > ctx->poll_max_ns) {
677 ctx->poll_ns = ctx->poll_max_ns;
680 trace_poll_grow(ctx, old, ctx->poll_ns);
684 aio_notify_accept(ctx);
686 /* if we have any readable fds, dispatch event */
688 for (i = 0; i < npfd; i++) {
689 nodes[i]->pfd.revents = pollfds[i].revents;
694 ctx->walking_handlers--;
696 /* Run dispatch even if there were no readable fds to run timers */
697 if (aio_dispatch(ctx, ret > 0)) {
701 aio_context_release(ctx);
706 void aio_context_setup(AioContext *ctx)
708 /* TODO remove this in final patch submission */
709 if (getenv("QEMU_AIO_POLL_MAX_NS")) {
710 fprintf(stderr, "The QEMU_AIO_POLL_MAX_NS environment variable has "
711 "been replaced with -object iothread,poll-max-ns=NUM\n");
715 #ifdef CONFIG_EPOLL_CREATE1
716 assert(!ctx->epollfd);
717 ctx->epollfd = epoll_create1(EPOLL_CLOEXEC);
718 if (ctx->epollfd == -1) {
719 fprintf(stderr, "Failed to create epoll instance: %s", strerror(errno));
720 ctx->epoll_available = false;
722 ctx->epoll_available = true;
727 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns,
728 int64_t grow, int64_t shrink, Error **errp)
730 /* No thread synchronization here, it doesn't matter if an incorrect value
733 ctx->poll_max_ns = max_ns;
735 ctx->poll_grow = grow;
736 ctx->poll_shrink = shrink;