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 #ifdef CONFIG_EPOLL_CREATE1
22 #include <sys/epoll.h>
33 QLIST_ENTRY(AioHandler) node;
36 #ifdef CONFIG_EPOLL_CREATE1
38 /* The fd number threashold to switch to epoll */
39 #define EPOLL_ENABLE_THRESHOLD 64
41 static void aio_epoll_disable(AioContext *ctx)
43 ctx->epoll_available = false;
44 if (!ctx->epoll_enabled) {
47 ctx->epoll_enabled = false;
51 static inline int epoll_events_from_pfd(int pfd_events)
53 return (pfd_events & G_IO_IN ? EPOLLIN : 0) |
54 (pfd_events & G_IO_OUT ? EPOLLOUT : 0) |
55 (pfd_events & G_IO_HUP ? EPOLLHUP : 0) |
56 (pfd_events & G_IO_ERR ? EPOLLERR : 0);
59 static bool aio_epoll_try_enable(AioContext *ctx)
62 struct epoll_event event;
64 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
66 if (node->deleted || !node->pfd.events) {
69 event.events = epoll_events_from_pfd(node->pfd.events);
70 event.data.ptr = node;
71 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
76 ctx->epoll_enabled = true;
80 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
82 struct epoll_event event;
85 if (!ctx->epoll_enabled) {
88 if (!node->pfd.events) {
89 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_DEL, node->pfd.fd, &event);
91 aio_epoll_disable(ctx);
94 event.data.ptr = node;
95 event.events = epoll_events_from_pfd(node->pfd.events);
97 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
99 aio_epoll_disable(ctx);
102 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_MOD, node->pfd.fd, &event);
104 aio_epoll_disable(ctx);
110 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
111 unsigned npfd, int64_t timeout)
115 struct epoll_event events[128];
118 assert(pfds[0].fd == ctx->epollfd);
120 ret = qemu_poll_ns(pfds, npfd, timeout);
122 if (timeout <= 0 || ret > 0) {
123 ret = epoll_wait(ctx->epollfd, events,
124 sizeof(events) / sizeof(events[0]),
129 for (i = 0; i < ret; i++) {
130 int ev = events[i].events;
131 node = events[i].data.ptr;
132 node->pfd.revents = (ev & EPOLLIN ? G_IO_IN : 0) |
133 (ev & EPOLLOUT ? G_IO_OUT : 0) |
134 (ev & EPOLLHUP ? G_IO_HUP : 0) |
135 (ev & EPOLLERR ? G_IO_ERR : 0);
142 static bool aio_epoll_enabled(AioContext *ctx)
144 /* Fall back to ppoll when external clients are disabled. */
145 return !aio_external_disabled(ctx) && ctx->epoll_enabled;
148 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
149 unsigned npfd, int64_t timeout)
151 if (!ctx->epoll_available) {
154 if (aio_epoll_enabled(ctx)) {
157 if (npfd >= EPOLL_ENABLE_THRESHOLD) {
158 if (aio_epoll_try_enable(ctx)) {
161 aio_epoll_disable(ctx);
169 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
173 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
174 unsigned npfd, int64_t timeout)
179 static bool aio_epoll_enabled(AioContext *ctx)
184 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
185 unsigned npfd, int64_t timeout)
192 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
196 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
197 if (node->pfd.fd == fd)
205 void aio_set_fd_handler(AioContext *ctx,
214 bool deleted = false;
216 node = find_aio_handler(ctx, fd);
218 /* Are we deleting the fd handler? */
219 if (!io_read && !io_write) {
221 g_source_remove_poll(&ctx->source, &node->pfd);
223 /* If the lock is held, just mark the node as deleted */
224 if (ctx->walking_handlers) {
226 node->pfd.revents = 0;
228 /* Otherwise, delete it for real. We can't just mark it as
229 * deleted because deleted nodes are only cleaned up after
230 * releasing the walking_handlers lock.
232 QLIST_REMOVE(node, node);
238 /* Alloc and insert if it's not already there */
239 node = g_new0(AioHandler, 1);
241 QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
243 g_source_add_poll(&ctx->source, &node->pfd);
246 /* Update handler with latest information */
247 node->io_read = io_read;
248 node->io_write = io_write;
249 node->opaque = opaque;
250 node->is_external = is_external;
252 node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
253 node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
256 aio_epoll_update(ctx, node, is_new);
263 void aio_set_event_notifier(AioContext *ctx,
264 EventNotifier *notifier,
266 EventNotifierHandler *io_read)
268 aio_set_fd_handler(ctx, event_notifier_get_fd(notifier),
269 is_external, (IOHandler *)io_read, NULL, notifier);
272 bool aio_prepare(AioContext *ctx)
277 bool aio_pending(AioContext *ctx)
281 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
284 revents = node->pfd.revents & node->pfd.events;
285 if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read) {
288 if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write) {
296 bool aio_dispatch(AioContext *ctx)
299 bool progress = false;
302 * If there are callbacks left that have been queued, we need to call them.
303 * Do not call select in this case, because it is possible that the caller
304 * does not need a complete flush (as is the case for aio_poll loops).
306 if (aio_bh_poll(ctx)) {
311 * We have to walk very carefully in case aio_set_fd_handler is
312 * called while we're walking.
314 node = QLIST_FIRST(&ctx->aio_handlers);
319 ctx->walking_handlers++;
321 revents = node->pfd.revents & node->pfd.events;
322 node->pfd.revents = 0;
324 if (!node->deleted &&
325 (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
327 node->io_read(node->opaque);
329 /* aio_notify() does not count as progress */
330 if (node->opaque != &ctx->notifier) {
334 if (!node->deleted &&
335 (revents & (G_IO_OUT | G_IO_ERR)) &&
337 node->io_write(node->opaque);
342 node = QLIST_NEXT(node, node);
344 ctx->walking_handlers--;
346 if (!ctx->walking_handlers && tmp->deleted) {
347 QLIST_REMOVE(tmp, node);
353 progress |= timerlistgroup_run_timers(&ctx->tlg);
358 /* These thread-local variables are used only in a small part of aio_poll
359 * around the call to the poll() system call. In particular they are not
360 * used while aio_poll is performing callbacks, which makes it much easier
361 * to think about reentrancy!
363 * Stack-allocated arrays would be perfect but they have size limitations;
364 * heap allocation is expensive enough that we want to reuse arrays across
365 * calls to aio_poll(). And because poll() has to be called without holding
366 * any lock, the arrays cannot be stored in AioContext. Thread-local data
367 * has none of the disadvantages of these three options.
369 static __thread GPollFD *pollfds;
370 static __thread AioHandler **nodes;
371 static __thread unsigned npfd, nalloc;
372 static __thread Notifier pollfds_cleanup_notifier;
374 static void pollfds_cleanup(Notifier *n, void *unused)
382 static void add_pollfd(AioHandler *node)
384 if (npfd == nalloc) {
386 pollfds_cleanup_notifier.notify = pollfds_cleanup;
387 qemu_thread_atexit_add(&pollfds_cleanup_notifier);
390 g_assert(nalloc <= INT_MAX);
393 pollfds = g_renew(GPollFD, pollfds, nalloc);
394 nodes = g_renew(AioHandler *, nodes, nalloc);
397 pollfds[npfd] = (GPollFD) {
399 .events = node->pfd.events,
404 bool aio_poll(AioContext *ctx, bool blocking)
411 aio_context_acquire(ctx);
414 /* aio_notify can avoid the expensive event_notifier_set if
415 * everything (file descriptors, bottom halves, timers) will
416 * be re-evaluated before the next blocking poll(). This is
417 * already true when aio_poll is called with blocking == false;
418 * if blocking == true, it is only true after poll() returns,
419 * so disable the optimization now.
422 atomic_add(&ctx->notify_me, 2);
425 ctx->walking_handlers++;
430 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
431 if (!node->deleted && node->pfd.events
432 && !aio_epoll_enabled(ctx)
433 && aio_node_check(ctx, node->is_external)) {
438 timeout = blocking ? aio_compute_timeout(ctx) : 0;
440 /* wait until next event */
442 aio_context_release(ctx);
444 if (aio_epoll_check_poll(ctx, pollfds, npfd, timeout)) {
445 AioHandler epoll_handler;
447 epoll_handler.pfd.fd = ctx->epollfd;
448 epoll_handler.pfd.events = G_IO_IN | G_IO_OUT | G_IO_HUP | G_IO_ERR;
450 add_pollfd(&epoll_handler);
451 ret = aio_epoll(ctx, pollfds, npfd, timeout);
453 ret = qemu_poll_ns(pollfds, npfd, timeout);
456 atomic_sub(&ctx->notify_me, 2);
459 aio_context_acquire(ctx);
462 aio_notify_accept(ctx);
464 /* if we have any readable fds, dispatch event */
466 for (i = 0; i < npfd; i++) {
467 nodes[i]->pfd.revents = pollfds[i].revents;
472 ctx->walking_handlers--;
474 /* Run dispatch even if there were no readable fds to run timers */
475 if (aio_dispatch(ctx)) {
479 aio_context_release(ctx);
484 void aio_context_setup(AioContext *ctx, Error **errp)
486 #ifdef CONFIG_EPOLL_CREATE1
487 assert(!ctx->epollfd);
488 ctx->epollfd = epoll_create1(EPOLL_CLOEXEC);
489 if (ctx->epollfd == -1) {
490 ctx->epoll_available = false;
492 ctx->epoll_available = true;