2 * QEMU aio implementation
4 * Copyright IBM Corp., 2008
5 * Copyright Red Hat Inc., 2012
11 * This work is licensed under the terms of the GNU GPL, version 2. See
12 * the COPYING file in the top-level directory.
14 * Contributions after 2012-01-13 are licensed under the terms of the
15 * GNU GPL, version 2 or (at your option) any later version.
18 #include "qemu/osdep.h"
19 #include "qemu-common.h"
20 #include "block/block.h"
21 #include "qemu/queue.h"
22 #include "qemu/sockets.h"
23 #include "qapi/error.h"
24 #include "qemu/rcu_queue.h"
30 EventNotifierHandler *io_notify;
35 QLIST_ENTRY(AioHandler) node;
38 void aio_set_fd_handler(AioContext *ctx,
46 /* fd is a SOCKET in our case */
49 qemu_lockcnt_lock(&ctx->list_lock);
50 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
51 if (node->pfd.fd == fd && !node->deleted) {
56 /* Are we deleting the fd handler? */
57 if (!io_read && !io_write) {
59 /* If aio_poll is in progress, just mark the node as deleted */
60 if (qemu_lockcnt_count(&ctx->list_lock)) {
62 node->pfd.revents = 0;
64 /* Otherwise, delete it for real. We can't just mark it as
65 * deleted because deleted nodes are only cleaned up after
66 * releasing the list_lock.
68 QLIST_REMOVE(node, node);
76 /* Alloc and insert if it's not already there */
77 node = g_new0(AioHandler, 1);
79 QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node);
84 node->pfd.events |= G_IO_IN;
87 node->pfd.events |= G_IO_OUT;
90 node->e = &ctx->notifier;
92 /* Update handler with latest information */
93 node->opaque = opaque;
94 node->io_read = io_read;
95 node->io_write = io_write;
96 node->is_external = is_external;
98 event = event_notifier_get_handle(&ctx->notifier);
99 WSAEventSelect(node->pfd.fd, event,
100 FD_READ | FD_ACCEPT | FD_CLOSE |
101 FD_CONNECT | FD_WRITE | FD_OOB);
104 qemu_lockcnt_unlock(&ctx->list_lock);
108 void aio_set_fd_poll(AioContext *ctx, int fd,
109 IOHandler *io_poll_begin,
110 IOHandler *io_poll_end)
112 /* Not implemented */
115 void aio_set_event_notifier(AioContext *ctx,
118 EventNotifierHandler *io_notify,
123 qemu_lockcnt_lock(&ctx->list_lock);
124 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
125 if (node->e == e && !node->deleted) {
130 /* Are we deleting the fd handler? */
133 g_source_remove_poll(&ctx->source, &node->pfd);
135 /* aio_poll is in progress, just mark the node as deleted */
136 if (qemu_lockcnt_count(&ctx->list_lock)) {
138 node->pfd.revents = 0;
140 /* Otherwise, delete it for real. We can't just mark it as
141 * deleted because deleted nodes are only cleaned up after
142 * releasing the list_lock.
144 QLIST_REMOVE(node, node);
150 /* Alloc and insert if it's not already there */
151 node = g_new0(AioHandler, 1);
153 node->pfd.fd = (uintptr_t)event_notifier_get_handle(e);
154 node->pfd.events = G_IO_IN;
155 node->is_external = is_external;
156 QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node);
158 g_source_add_poll(&ctx->source, &node->pfd);
160 /* Update handler with latest information */
161 node->io_notify = io_notify;
164 qemu_lockcnt_unlock(&ctx->list_lock);
168 void aio_set_event_notifier_poll(AioContext *ctx,
169 EventNotifier *notifier,
170 EventNotifierHandler *io_poll_begin,
171 EventNotifierHandler *io_poll_end)
173 /* Not implemented */
176 bool aio_prepare(AioContext *ctx)
178 static struct timeval tv0;
180 bool have_select_revents = false;
184 * We have to walk very carefully in case aio_set_fd_handler is
185 * called while we're walking.
187 qemu_lockcnt_inc(&ctx->list_lock);
192 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
194 FD_SET ((SOCKET)node->pfd.fd, &rfds);
196 if (node->io_write) {
197 FD_SET ((SOCKET)node->pfd.fd, &wfds);
201 if (select(0, &rfds, &wfds, NULL, &tv0) > 0) {
202 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
203 node->pfd.revents = 0;
204 if (FD_ISSET(node->pfd.fd, &rfds)) {
205 node->pfd.revents |= G_IO_IN;
206 have_select_revents = true;
209 if (FD_ISSET(node->pfd.fd, &wfds)) {
210 node->pfd.revents |= G_IO_OUT;
211 have_select_revents = true;
216 qemu_lockcnt_dec(&ctx->list_lock);
217 return have_select_revents;
220 bool aio_pending(AioContext *ctx)
226 * We have to walk very carefully in case aio_set_fd_handler is
227 * called while we're walking.
229 qemu_lockcnt_inc(&ctx->list_lock);
230 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
231 if (node->pfd.revents && node->io_notify) {
236 if ((node->pfd.revents & G_IO_IN) && node->io_read) {
240 if ((node->pfd.revents & G_IO_OUT) && node->io_write) {
246 qemu_lockcnt_dec(&ctx->list_lock);
250 static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event)
253 bool progress = false;
256 qemu_lockcnt_inc(&ctx->list_lock);
259 * We have to walk very carefully in case aio_set_fd_handler is
260 * called while we're walking.
262 QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) {
263 int revents = node->pfd.revents;
265 if (!node->deleted &&
266 (revents || event_notifier_get_handle(node->e) == event) &&
268 node->pfd.revents = 0;
269 node->io_notify(node->e);
271 /* aio_notify() does not count as progress */
272 if (node->e != &ctx->notifier) {
277 if (!node->deleted &&
278 (node->io_read || node->io_write)) {
279 node->pfd.revents = 0;
280 if ((revents & G_IO_IN) && node->io_read) {
281 node->io_read(node->opaque);
284 if ((revents & G_IO_OUT) && node->io_write) {
285 node->io_write(node->opaque);
289 /* if the next select() will return an event, we have progressed */
290 if (event == event_notifier_get_handle(&ctx->notifier)) {
292 WSAEnumNetworkEvents(node->pfd.fd, event, &ev);
293 if (ev.lNetworkEvents) {
300 if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
301 QLIST_REMOVE(node, node);
303 qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
308 qemu_lockcnt_dec(&ctx->list_lock);
312 void aio_dispatch(AioContext *ctx)
315 aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE);
316 timerlistgroup_run_timers(&ctx->tlg);
319 bool aio_poll(AioContext *ctx, bool blocking)
322 HANDLE events[MAXIMUM_WAIT_OBJECTS + 1];
323 bool progress, have_select_revents, first;
329 /* aio_notify can avoid the expensive event_notifier_set if
330 * everything (file descriptors, bottom halves, timers) will
331 * be re-evaluated before the next blocking poll(). This is
332 * already true when aio_poll is called with blocking == false;
333 * if blocking == true, it is only true after poll() returns,
334 * so disable the optimization now.
337 atomic_add(&ctx->notify_me, 2);
340 qemu_lockcnt_inc(&ctx->list_lock);
341 have_select_revents = aio_prepare(ctx);
345 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
346 if (!node->deleted && node->io_notify
347 && aio_node_check(ctx, node->is_external)) {
348 events[count++] = event_notifier_get_handle(node->e);
352 qemu_lockcnt_dec(&ctx->list_lock);
355 /* ctx->notifier is always registered. */
358 /* Multiple iterations, all of them non-blocking except the first,
359 * may be necessary to process all pending events. After the first
360 * WaitForMultipleObjects call ctx->notify_me will be decremented.
366 timeout = blocking && !have_select_revents
367 ? qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)) : 0;
368 ret = WaitForMultipleObjects(count, events, FALSE, timeout);
371 atomic_sub(&ctx->notify_me, 2);
375 aio_notify_accept(ctx);
376 progress |= aio_bh_poll(ctx);
380 /* if we have any signaled events, dispatch event */
382 if ((DWORD) (ret - WAIT_OBJECT_0) < count) {
383 event = events[ret - WAIT_OBJECT_0];
384 events[ret - WAIT_OBJECT_0] = events[--count];
385 } else if (!have_select_revents) {
389 have_select_revents = false;
392 progress |= aio_dispatch_handlers(ctx, event);
395 progress |= timerlistgroup_run_timers(&ctx->tlg);
399 void aio_context_setup(AioContext *ctx)
403 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns,
404 int64_t grow, int64_t shrink, Error **errp)
406 error_setg(errp, "AioContext polling is not implemented on Windows");