Set last_sent_block
[qemu/ar7.git] / aio-posix.c
blob482b31650203b3ab3e25cc2d652b64274030012c
1 /*
2 * QEMU aio implementation
4 * Copyright IBM, Corp. 2008
6 * Authors:
7 * Anthony Liguori <aliguori@us.ibm.com>
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"
20 #ifdef CONFIG_EPOLL
21 #include <sys/epoll.h>
22 #endif
24 struct AioHandler
26 GPollFD pfd;
27 IOHandler *io_read;
28 IOHandler *io_write;
29 int deleted;
30 void *opaque;
31 bool is_external;
32 QLIST_ENTRY(AioHandler) node;
35 #ifdef CONFIG_EPOLL
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) {
44 return;
46 ctx->epoll_enabled = false;
47 close(ctx->epollfd);
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)
60 AioHandler *node;
61 struct epoll_event event;
63 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
64 int r;
65 if (node->deleted || !node->pfd.events) {
66 continue;
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);
71 if (r) {
72 return false;
75 ctx->epoll_enabled = true;
76 return true;
79 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
81 struct epoll_event event;
82 int r;
84 if (!ctx->epoll_enabled) {
85 return;
87 if (!node->pfd.events) {
88 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_DEL, node->pfd.fd, &event);
89 if (r) {
90 aio_epoll_disable(ctx);
92 } else {
93 event.data.ptr = node;
94 event.events = epoll_events_from_pfd(node->pfd.events);
95 if (is_new) {
96 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
97 if (r) {
98 aio_epoll_disable(ctx);
100 } else {
101 r = epoll_ctl(ctx->epollfd, EPOLL_CTL_MOD, node->pfd.fd, &event);
102 if (r) {
103 aio_epoll_disable(ctx);
109 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
110 unsigned npfd, int64_t timeout)
112 AioHandler *node;
113 int i, ret = 0;
114 struct epoll_event events[128];
116 assert(npfd == 1);
117 assert(pfds[0].fd == ctx->epollfd);
118 if (timeout > 0) {
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]),
124 timeout);
125 if (ret <= 0) {
126 goto out;
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);
137 out:
138 return ret;
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) {
151 return false;
153 if (aio_epoll_enabled(ctx)) {
154 return true;
156 if (npfd >= EPOLL_ENABLE_THRESHOLD) {
157 if (aio_epoll_try_enable(ctx)) {
158 return true;
159 } else {
160 aio_epoll_disable(ctx);
163 return false;
166 #else
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)
175 assert(false);
178 static bool aio_epoll_enabled(AioContext *ctx)
180 return false;
183 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
184 unsigned npfd, int64_t timeout)
186 return false;
189 #endif
191 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
193 AioHandler *node;
195 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
196 if (node->pfd.fd == fd)
197 if (!node->deleted)
198 return node;
201 return NULL;
204 void aio_set_fd_handler(AioContext *ctx,
205 int fd,
206 bool is_external,
207 IOHandler *io_read,
208 IOHandler *io_write,
209 void *opaque)
211 AioHandler *node;
212 bool is_new = false;
213 bool deleted = false;
215 node = find_aio_handler(ctx, fd);
217 /* Are we deleting the fd handler? */
218 if (!io_read && !io_write) {
219 if (node) {
220 g_source_remove_poll(&ctx->source, &node->pfd);
222 /* If the lock is held, just mark the node as deleted */
223 if (ctx->walking_handlers) {
224 node->deleted = 1;
225 node->pfd.revents = 0;
226 } else {
227 /* Otherwise, delete it for real. We can't just mark it as
228 * deleted because deleted nodes are only cleaned up after
229 * releasing the walking_handlers lock.
231 QLIST_REMOVE(node, node);
232 deleted = true;
235 } else {
236 if (node == NULL) {
237 /* Alloc and insert if it's not already there */
238 node = g_new0(AioHandler, 1);
239 node->pfd.fd = fd;
240 QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
242 g_source_add_poll(&ctx->source, &node->pfd);
243 is_new = true;
245 /* Update handler with latest information */
246 node->io_read = io_read;
247 node->io_write = io_write;
248 node->opaque = opaque;
249 node->is_external = is_external;
251 node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
252 node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
255 aio_epoll_update(ctx, node, is_new);
256 aio_notify(ctx);
257 if (deleted) {
258 g_free(node);
262 void aio_set_event_notifier(AioContext *ctx,
263 EventNotifier *notifier,
264 bool is_external,
265 EventNotifierHandler *io_read)
267 aio_set_fd_handler(ctx, event_notifier_get_fd(notifier),
268 is_external, (IOHandler *)io_read, NULL, notifier);
271 bool aio_prepare(AioContext *ctx)
273 return false;
276 bool aio_pending(AioContext *ctx)
278 AioHandler *node;
280 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
281 int revents;
283 revents = node->pfd.revents & node->pfd.events;
284 if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read) {
285 return true;
287 if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write) {
288 return true;
292 return false;
295 bool aio_dispatch(AioContext *ctx)
297 AioHandler *node;
298 bool progress = false;
301 * If there are callbacks left that have been queued, we need to call them.
302 * Do not call select in this case, because it is possible that the caller
303 * does not need a complete flush (as is the case for aio_poll loops).
305 if (aio_bh_poll(ctx)) {
306 progress = true;
310 * We have to walk very carefully in case aio_set_fd_handler is
311 * called while we're walking.
313 node = QLIST_FIRST(&ctx->aio_handlers);
314 while (node) {
315 AioHandler *tmp;
316 int revents;
318 ctx->walking_handlers++;
320 revents = node->pfd.revents & node->pfd.events;
321 node->pfd.revents = 0;
323 if (!node->deleted &&
324 (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
325 node->io_read) {
326 node->io_read(node->opaque);
328 /* aio_notify() does not count as progress */
329 if (node->opaque != &ctx->notifier) {
330 progress = true;
333 if (!node->deleted &&
334 (revents & (G_IO_OUT | G_IO_ERR)) &&
335 node->io_write) {
336 node->io_write(node->opaque);
337 progress = true;
340 tmp = node;
341 node = QLIST_NEXT(node, node);
343 ctx->walking_handlers--;
345 if (!ctx->walking_handlers && tmp->deleted) {
346 QLIST_REMOVE(tmp, node);
347 g_free(tmp);
351 /* Run our timers */
352 progress |= timerlistgroup_run_timers(&ctx->tlg);
354 return progress;
357 /* These thread-local variables are used only in a small part of aio_poll
358 * around the call to the poll() system call. In particular they are not
359 * used while aio_poll is performing callbacks, which makes it much easier
360 * to think about reentrancy!
362 * Stack-allocated arrays would be perfect but they have size limitations;
363 * heap allocation is expensive enough that we want to reuse arrays across
364 * calls to aio_poll(). And because poll() has to be called without holding
365 * any lock, the arrays cannot be stored in AioContext. Thread-local data
366 * has none of the disadvantages of these three options.
368 static __thread GPollFD *pollfds;
369 static __thread AioHandler **nodes;
370 static __thread unsigned npfd, nalloc;
371 static __thread Notifier pollfds_cleanup_notifier;
373 static void pollfds_cleanup(Notifier *n, void *unused)
375 g_assert(npfd == 0);
376 g_free(pollfds);
377 g_free(nodes);
378 nalloc = 0;
381 static void add_pollfd(AioHandler *node)
383 if (npfd == nalloc) {
384 if (nalloc == 0) {
385 pollfds_cleanup_notifier.notify = pollfds_cleanup;
386 qemu_thread_atexit_add(&pollfds_cleanup_notifier);
387 nalloc = 8;
388 } else {
389 g_assert(nalloc <= INT_MAX);
390 nalloc *= 2;
392 pollfds = g_renew(GPollFD, pollfds, nalloc);
393 nodes = g_renew(AioHandler *, nodes, nalloc);
395 nodes[npfd] = node;
396 pollfds[npfd] = (GPollFD) {
397 .fd = node->pfd.fd,
398 .events = node->pfd.events,
400 npfd++;
403 bool aio_poll(AioContext *ctx, bool blocking)
405 AioHandler *node;
406 int i, ret;
407 bool progress;
408 int64_t timeout;
410 aio_context_acquire(ctx);
411 progress = false;
413 /* aio_notify can avoid the expensive event_notifier_set if
414 * everything (file descriptors, bottom halves, timers) will
415 * be re-evaluated before the next blocking poll(). This is
416 * already true when aio_poll is called with blocking == false;
417 * if blocking == true, it is only true after poll() returns,
418 * so disable the optimization now.
420 if (blocking) {
421 atomic_add(&ctx->notify_me, 2);
424 ctx->walking_handlers++;
426 assert(npfd == 0);
428 /* fill pollfds */
429 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
430 if (!node->deleted && node->pfd.events
431 && !aio_epoll_enabled(ctx)
432 && aio_node_check(ctx, node->is_external)) {
433 add_pollfd(node);
437 timeout = blocking ? aio_compute_timeout(ctx) : 0;
439 /* wait until next event */
440 if (timeout) {
441 aio_context_release(ctx);
443 if (aio_epoll_check_poll(ctx, pollfds, npfd, timeout)) {
444 AioHandler epoll_handler;
446 epoll_handler.pfd.fd = ctx->epollfd;
447 epoll_handler.pfd.events = G_IO_IN | G_IO_OUT | G_IO_HUP | G_IO_ERR;
448 npfd = 0;
449 add_pollfd(&epoll_handler);
450 ret = aio_epoll(ctx, pollfds, npfd, timeout);
451 } else {
452 ret = qemu_poll_ns(pollfds, npfd, timeout);
454 if (blocking) {
455 atomic_sub(&ctx->notify_me, 2);
457 if (timeout) {
458 aio_context_acquire(ctx);
461 aio_notify_accept(ctx);
463 /* if we have any readable fds, dispatch event */
464 if (ret > 0) {
465 for (i = 0; i < npfd; i++) {
466 nodes[i]->pfd.revents = pollfds[i].revents;
470 npfd = 0;
471 ctx->walking_handlers--;
473 /* Run dispatch even if there were no readable fds to run timers */
474 if (aio_dispatch(ctx)) {
475 progress = true;
478 aio_context_release(ctx);
480 return progress;
483 void aio_context_setup(AioContext *ctx, Error **errp)
485 #ifdef CONFIG_EPOLL
486 assert(!ctx->epollfd);
487 ctx->epollfd = epoll_create1(EPOLL_CLOEXEC);
488 if (ctx->epollfd == -1) {
489 ctx->epoll_available = false;
490 } else {
491 ctx->epoll_available = true;
493 #endif