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-common.h"
17 #include "block/block.h"
18 #include "qemu/queue.h"
19 #include "qemu/sockets.h"
21 #include <sys/epoll.h>
32 QLIST_ENTRY(AioHandler) node;
37 /* The fd number threashold to switch to epoll */
38 #define EPOLL_ENABLE_THRESHOLD 64
40 static void aio_epoll_disable(AioContext *ctx)
42 ctx->epoll_available = false;
43 if (!ctx->epoll_enabled) {
46 ctx->epoll_enabled = false;
50 static inline int epoll_events_from_pfd(int pfd_events)
52 return (pfd_events & G_IO_IN ? EPOLLIN : 0) |
53 (pfd_events & G_IO_OUT ? EPOLLOUT : 0) |
54 (pfd_events & G_IO_HUP ? EPOLLHUP : 0) |
55 (pfd_events & G_IO_ERR ? EPOLLERR : 0);
58 static bool aio_epoll_try_enable(AioContext *ctx)
61 struct epoll_event event;
63 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
65 if (node->deleted || !node->pfd.events) {
68 event.events = epoll_events_from_pfd(node->pfd.events);
69 event.data.ptr = node;
70 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
75 ctx->epoll_enabled = true;
79 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
81 struct epoll_event event;
84 if (!ctx->epoll_enabled) {
87 if (!node->pfd.events) {
88 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_DEL, node->pfd.fd, &event);
90 aio_epoll_disable(ctx);
93 event.data.ptr = node;
94 event.events = epoll_events_from_pfd(node->pfd.events);
96 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
98 aio_epoll_disable(ctx);
101 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_MOD, node->pfd.fd, &event);
103 aio_epoll_disable(ctx);
109 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
110 unsigned npfd, int64_t timeout)
114 struct epoll_event events[128];
117 assert(pfds[0].fd == ctx->epollfd);
119 ret = qemu_poll_ns(pfds, npfd, timeout);
121 if (timeout <= 0 || ret > 0) {
122 ret = epoll_wait(ctx->epollfd, events,
123 sizeof(events) / sizeof(events[0]),
128 for (i = 0; i < ret; i++) {
129 int ev = events[i].events;
130 node = events[i].data.ptr;
131 node->pfd.revents = (ev & EPOLLIN ? G_IO_IN : 0) |
132 (ev & EPOLLOUT ? G_IO_OUT : 0) |
133 (ev & EPOLLHUP ? G_IO_HUP : 0) |
134 (ev & EPOLLERR ? G_IO_ERR : 0);
141 static bool aio_epoll_enabled(AioContext *ctx)
143 /* Fall back to ppoll when external clients are disabled. */
144 return !aio_external_disabled(ctx) && ctx->epoll_enabled;
147 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
148 unsigned npfd, int64_t timeout)
150 if (!ctx->epoll_available) {
153 if (aio_epoll_enabled(ctx)) {
156 if (npfd >= EPOLL_ENABLE_THRESHOLD) {
157 if (aio_epoll_try_enable(ctx)) {
160 aio_epoll_disable(ctx);
168 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
172 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
173 unsigned npfd, int64_t timeout)
178 static bool aio_epoll_enabled(AioContext *ctx)
183 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
184 unsigned npfd, int64_t timeout)
191 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
195 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
196 if (node->pfd.fd == fd)
204 void aio_set_fd_handler(AioContext *ctx,
214 node = find_aio_handler(ctx, fd);
216 /* Are we deleting the fd handler? */
217 if (!io_read && !io_write) {
219 g_source_remove_poll(&ctx->source, &node->pfd);
221 /* If the lock is held, just mark the node as deleted */
222 if (ctx->walking_handlers) {
224 node->pfd.revents = 0;
226 /* Otherwise, delete it for real. We can't just mark it as
227 * deleted because deleted nodes are only cleaned up after
228 * releasing the walking_handlers lock.
230 QLIST_REMOVE(node, node);
236 /* Alloc and insert if it's not already there */
237 node = g_new0(AioHandler, 1);
239 QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
241 g_source_add_poll(&ctx->source, &node->pfd);
244 /* Update handler with latest information */
245 node->io_read = io_read;
246 node->io_write = io_write;
247 node->opaque = opaque;
248 node->is_external = is_external;
250 node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
251 node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
254 aio_epoll_update(ctx, node, is_new);
258 void aio_set_event_notifier(AioContext *ctx,
259 EventNotifier *notifier,
261 EventNotifierHandler *io_read)
263 aio_set_fd_handler(ctx, event_notifier_get_fd(notifier),
264 is_external, (IOHandler *)io_read, NULL, notifier);
267 bool aio_prepare(AioContext *ctx)
272 bool aio_pending(AioContext *ctx)
276 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
279 revents = node->pfd.revents & node->pfd.events;
280 if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read) {
283 if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write) {
291 bool aio_dispatch(AioContext *ctx)
294 bool progress = false;
297 * If there are callbacks left that have been queued, we need to call them.
298 * Do not call select in this case, because it is possible that the caller
299 * does not need a complete flush (as is the case for aio_poll loops).
301 if (aio_bh_poll(ctx)) {
306 * We have to walk very carefully in case aio_set_fd_handler is
307 * called while we're walking.
309 node = QLIST_FIRST(&ctx->aio_handlers);
314 ctx->walking_handlers++;
316 revents = node->pfd.revents & node->pfd.events;
317 node->pfd.revents = 0;
319 if (!node->deleted &&
320 (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
322 node->io_read(node->opaque);
324 /* aio_notify() does not count as progress */
325 if (node->opaque != &ctx->notifier) {
329 if (!node->deleted &&
330 (revents & (G_IO_OUT | G_IO_ERR)) &&
332 node->io_write(node->opaque);
337 node = QLIST_NEXT(node, node);
339 ctx->walking_handlers--;
341 if (!ctx->walking_handlers && tmp->deleted) {
342 QLIST_REMOVE(tmp, node);
348 progress |= timerlistgroup_run_timers(&ctx->tlg);
353 /* These thread-local variables are used only in a small part of aio_poll
354 * around the call to the poll() system call. In particular they are not
355 * used while aio_poll is performing callbacks, which makes it much easier
356 * to think about reentrancy!
358 * Stack-allocated arrays would be perfect but they have size limitations;
359 * heap allocation is expensive enough that we want to reuse arrays across
360 * calls to aio_poll(). And because poll() has to be called without holding
361 * any lock, the arrays cannot be stored in AioContext. Thread-local data
362 * has none of the disadvantages of these three options.
364 static __thread GPollFD *pollfds;
365 static __thread AioHandler **nodes;
366 static __thread unsigned npfd, nalloc;
367 static __thread Notifier pollfds_cleanup_notifier;
369 static void pollfds_cleanup(Notifier *n, void *unused)
377 static void add_pollfd(AioHandler *node)
379 if (npfd == nalloc) {
381 pollfds_cleanup_notifier.notify = pollfds_cleanup;
382 qemu_thread_atexit_add(&pollfds_cleanup_notifier);
385 g_assert(nalloc <= INT_MAX);
388 pollfds = g_renew(GPollFD, pollfds, nalloc);
389 nodes = g_renew(AioHandler *, nodes, nalloc);
392 pollfds[npfd] = (GPollFD) {
394 .events = node->pfd.events,
399 bool aio_poll(AioContext *ctx, bool blocking)
406 aio_context_acquire(ctx);
409 /* aio_notify can avoid the expensive event_notifier_set if
410 * everything (file descriptors, bottom halves, timers) will
411 * be re-evaluated before the next blocking poll(). This is
412 * already true when aio_poll is called with blocking == false;
413 * if blocking == true, it is only true after poll() returns,
414 * so disable the optimization now.
417 atomic_add(&ctx->notify_me, 2);
420 ctx->walking_handlers++;
425 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
426 if (!node->deleted && node->pfd.events
427 && !aio_epoll_enabled(ctx)
428 && aio_node_check(ctx, node->is_external)) {
433 timeout = blocking ? aio_compute_timeout(ctx) : 0;
435 /* wait until next event */
437 aio_context_release(ctx);
439 if (aio_epoll_check_poll(ctx, pollfds, npfd, timeout)) {
440 AioHandler epoll_handler;
442 epoll_handler.pfd.fd = ctx->epollfd;
443 epoll_handler.pfd.events = G_IO_IN | G_IO_OUT | G_IO_HUP | G_IO_ERR;
445 add_pollfd(&epoll_handler);
446 ret = aio_epoll(ctx, pollfds, npfd, timeout);
448 ret = qemu_poll_ns(pollfds, npfd, timeout);
451 atomic_sub(&ctx->notify_me, 2);
454 aio_context_acquire(ctx);
457 aio_notify_accept(ctx);
459 /* if we have any readable fds, dispatch event */
461 for (i = 0; i < npfd; i++) {
462 nodes[i]->pfd.revents = pollfds[i].revents;
467 ctx->walking_handlers--;
469 /* Run dispatch even if there were no readable fds to run timers */
470 if (aio_dispatch(ctx)) {
474 aio_context_release(ctx);
479 void aio_context_setup(AioContext *ctx, Error **errp)
482 assert(!ctx->epollfd);
483 ctx->epollfd = epoll_create1(EPOLL_CLOEXEC);
484 if (ctx->epollfd == -1) {
485 ctx->epoll_available = false;
487 ctx->epoll_available = true;