2 * QEMU aio implementation
4 * Copyright IBM Corp., 2008
5 * Copyright Red Hat Inc., 2012
8 * Anthony Liguori <aliguori@us.ibm.com>
9 * Paolo Bonzini <pbonzini@redhat.com>
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-common.h"
19 #include "block/block.h"
20 #include "qemu/queue.h"
21 #include "qemu/sockets.h"
25 EventNotifierHandler
*io_notify
;
28 QLIST_ENTRY(AioHandler
) node
;
31 void aio_set_event_notifier(AioContext
*ctx
,
33 EventNotifierHandler
*io_notify
)
37 QLIST_FOREACH(node
, &ctx
->aio_handlers
, node
) {
38 if (node
->e
== e
&& !node
->deleted
) {
43 /* Are we deleting the fd handler? */
46 g_source_remove_poll(&ctx
->source
, &node
->pfd
);
48 /* If the lock is held, just mark the node as deleted */
49 if (ctx
->walking_handlers
) {
51 node
->pfd
.revents
= 0;
53 /* Otherwise, delete it for real. We can't just mark it as
54 * deleted because deleted nodes are only cleaned up after
55 * releasing the walking_handlers lock.
57 QLIST_REMOVE(node
, node
);
63 /* Alloc and insert if it's not already there */
64 node
= g_malloc0(sizeof(AioHandler
));
66 node
->pfd
.fd
= (uintptr_t)event_notifier_get_handle(e
);
67 node
->pfd
.events
= G_IO_IN
;
68 QLIST_INSERT_HEAD(&ctx
->aio_handlers
, node
, node
);
70 g_source_add_poll(&ctx
->source
, &node
->pfd
);
72 /* Update handler with latest information */
73 node
->io_notify
= io_notify
;
79 bool aio_pending(AioContext
*ctx
)
83 QLIST_FOREACH(node
, &ctx
->aio_handlers
, node
) {
84 if (node
->pfd
.revents
&& node
->io_notify
) {
92 bool aio_poll(AioContext
*ctx
, bool blocking
)
95 HANDLE events
[MAXIMUM_WAIT_OBJECTS
+ 1];
103 * If there are callbacks left that have been queued, we need to call then.
104 * Do not call select in this case, because it is possible that the caller
105 * does not need a complete flush (as is the case for qemu_aio_wait loops).
107 if (aio_bh_poll(ctx
)) {
113 progress
|= timerlistgroup_run_timers(&ctx
->tlg
);
116 * Then dispatch any pending callbacks from the GSource.
118 * We have to walk very carefully in case qemu_aio_set_fd_handler is
119 * called while we're walking.
121 node
= QLIST_FIRST(&ctx
->aio_handlers
);
125 ctx
->walking_handlers
++;
127 if (node
->pfd
.revents
&& node
->io_notify
) {
128 node
->pfd
.revents
= 0;
129 node
->io_notify(node
->e
);
131 /* aio_notify() does not count as progress */
132 if (node
->e
!= &ctx
->notifier
) {
138 node
= QLIST_NEXT(node
, node
);
140 ctx
->walking_handlers
--;
142 if (!ctx
->walking_handlers
&& tmp
->deleted
) {
143 QLIST_REMOVE(tmp
, node
);
148 if (progress
&& !blocking
) {
152 ctx
->walking_handlers
++;
156 QLIST_FOREACH(node
, &ctx
->aio_handlers
, node
) {
157 if (!node
->deleted
&& node
->io_notify
) {
158 events
[count
++] = event_notifier_get_handle(node
->e
);
162 ctx
->walking_handlers
--;
164 /* wait until next event */
169 qemu_timeout_ns_to_ms(timerlistgroup_deadline_ns(&ctx
->tlg
)) : 0;
170 ret
= WaitForMultipleObjects(count
, events
, FALSE
, timeout
);
172 /* if we have any signaled events, dispatch event */
173 if ((DWORD
) (ret
- WAIT_OBJECT_0
) >= count
) {
179 /* we have to walk very carefully in case
180 * qemu_aio_set_fd_handler is called while we're walking */
181 node
= QLIST_FIRST(&ctx
->aio_handlers
);
185 ctx
->walking_handlers
++;
187 if (!node
->deleted
&&
188 event_notifier_get_handle(node
->e
) == events
[ret
- WAIT_OBJECT_0
] &&
190 node
->io_notify(node
->e
);
192 /* aio_notify() does not count as progress */
193 if (node
->e
!= &ctx
->notifier
) {
199 node
= QLIST_NEXT(node
, node
);
201 ctx
->walking_handlers
--;
203 if (!ctx
->walking_handlers
&& tmp
->deleted
) {
204 QLIST_REMOVE(tmp
, node
);
209 /* Try again, but only call each handler once. */
210 events
[ret
- WAIT_OBJECT_0
] = events
[--count
];
214 /* Run the timers a second time. We do this because otherwise aio_wait
215 * will not note progress - and will stop a drain early - if we have
216 * a timer that was not ready to run entering g_poll but is ready
217 * after g_poll. This will only do anything if a timer has expired.
219 progress
|= timerlistgroup_run_timers(&ctx
->tlg
);