]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * QEMU aio implementation | |
3 | * | |
4 | * Copyright IBM Corp., 2008 | |
5 | * Copyright Red Hat Inc., 2012 | |
6 | * | |
7 | * Authors: | |
8 | * Anthony Liguori <[email protected]> | |
9 | * Paolo Bonzini <[email protected]> | |
10 | * | |
11 | * This work is licensed under the terms of the GNU GPL, version 2. See | |
12 | * the COPYING file in the top-level directory. | |
13 | * | |
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. | |
16 | */ | |
17 | ||
18 | #include "qemu-common.h" | |
19 | #include "block/block.h" | |
20 | #include "qemu/queue.h" | |
21 | #include "qemu/sockets.h" | |
22 | ||
23 | struct AioHandler { | |
24 | EventNotifier *e; | |
25 | EventNotifierHandler *io_notify; | |
26 | GPollFD pfd; | |
27 | int deleted; | |
28 | QLIST_ENTRY(AioHandler) node; | |
29 | }; | |
30 | ||
31 | void aio_set_event_notifier(AioContext *ctx, | |
32 | EventNotifier *e, | |
33 | EventNotifierHandler *io_notify) | |
34 | { | |
35 | AioHandler *node; | |
36 | ||
37 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { | |
38 | if (node->e == e && !node->deleted) { | |
39 | break; | |
40 | } | |
41 | } | |
42 | ||
43 | /* Are we deleting the fd handler? */ | |
44 | if (!io_notify) { | |
45 | if (node) { | |
46 | g_source_remove_poll(&ctx->source, &node->pfd); | |
47 | ||
48 | /* If the lock is held, just mark the node as deleted */ | |
49 | if (ctx->walking_handlers) { | |
50 | node->deleted = 1; | |
51 | node->pfd.revents = 0; | |
52 | } else { | |
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. | |
56 | */ | |
57 | QLIST_REMOVE(node, node); | |
58 | g_free(node); | |
59 | } | |
60 | } | |
61 | } else { | |
62 | if (node == NULL) { | |
63 | /* Alloc and insert if it's not already there */ | |
64 | node = g_malloc0(sizeof(AioHandler)); | |
65 | node->e = e; | |
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); | |
69 | ||
70 | g_source_add_poll(&ctx->source, &node->pfd); | |
71 | } | |
72 | /* Update handler with latest information */ | |
73 | node->io_notify = io_notify; | |
74 | } | |
75 | ||
76 | aio_notify(ctx); | |
77 | } | |
78 | ||
79 | bool aio_pending(AioContext *ctx) | |
80 | { | |
81 | AioHandler *node; | |
82 | ||
83 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { | |
84 | if (node->pfd.revents && node->io_notify) { | |
85 | return true; | |
86 | } | |
87 | } | |
88 | ||
89 | return false; | |
90 | } | |
91 | ||
92 | bool aio_poll(AioContext *ctx, bool blocking) | |
93 | { | |
94 | AioHandler *node; | |
95 | HANDLE events[MAXIMUM_WAIT_OBJECTS + 1]; | |
96 | bool progress; | |
97 | int count; | |
98 | int timeout; | |
99 | ||
100 | progress = false; | |
101 | ||
102 | /* | |
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). | |
106 | */ | |
107 | if (aio_bh_poll(ctx)) { | |
108 | blocking = false; | |
109 | progress = true; | |
110 | } | |
111 | ||
112 | /* Run timers */ | |
113 | progress |= timerlistgroup_run_timers(&ctx->tlg); | |
114 | ||
115 | /* | |
116 | * Then dispatch any pending callbacks from the GSource. | |
117 | * | |
118 | * We have to walk very carefully in case qemu_aio_set_fd_handler is | |
119 | * called while we're walking. | |
120 | */ | |
121 | node = QLIST_FIRST(&ctx->aio_handlers); | |
122 | while (node) { | |
123 | AioHandler *tmp; | |
124 | ||
125 | ctx->walking_handlers++; | |
126 | ||
127 | if (node->pfd.revents && node->io_notify) { | |
128 | node->pfd.revents = 0; | |
129 | node->io_notify(node->e); | |
130 | ||
131 | /* aio_notify() does not count as progress */ | |
132 | if (node->e != &ctx->notifier) { | |
133 | progress = true; | |
134 | } | |
135 | } | |
136 | ||
137 | tmp = node; | |
138 | node = QLIST_NEXT(node, node); | |
139 | ||
140 | ctx->walking_handlers--; | |
141 | ||
142 | if (!ctx->walking_handlers && tmp->deleted) { | |
143 | QLIST_REMOVE(tmp, node); | |
144 | g_free(tmp); | |
145 | } | |
146 | } | |
147 | ||
148 | if (progress && !blocking) { | |
149 | return true; | |
150 | } | |
151 | ||
152 | ctx->walking_handlers++; | |
153 | ||
154 | /* fill fd sets */ | |
155 | count = 0; | |
156 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { | |
157 | if (!node->deleted && node->io_notify) { | |
158 | events[count++] = event_notifier_get_handle(node->e); | |
159 | } | |
160 | } | |
161 | ||
162 | ctx->walking_handlers--; | |
163 | ||
164 | /* early return if we only have the aio_notify() fd */ | |
165 | if (count == 1) { | |
166 | return progress; | |
167 | } | |
168 | ||
169 | /* wait until next event */ | |
170 | while (count > 0) { | |
171 | int ret; | |
172 | ||
173 | timeout = blocking ? | |
174 | qemu_timeout_ns_to_ms(timerlistgroup_deadline_ns(&ctx->tlg)) : 0; | |
175 | ret = WaitForMultipleObjects(count, events, FALSE, timeout); | |
176 | ||
177 | /* if we have any signaled events, dispatch event */ | |
178 | if ((DWORD) (ret - WAIT_OBJECT_0) >= count) { | |
179 | break; | |
180 | } | |
181 | ||
182 | blocking = false; | |
183 | ||
184 | /* we have to walk very carefully in case | |
185 | * qemu_aio_set_fd_handler is called while we're walking */ | |
186 | node = QLIST_FIRST(&ctx->aio_handlers); | |
187 | while (node) { | |
188 | AioHandler *tmp; | |
189 | ||
190 | ctx->walking_handlers++; | |
191 | ||
192 | if (!node->deleted && | |
193 | event_notifier_get_handle(node->e) == events[ret - WAIT_OBJECT_0] && | |
194 | node->io_notify) { | |
195 | node->io_notify(node->e); | |
196 | ||
197 | /* aio_notify() does not count as progress */ | |
198 | if (node->e != &ctx->notifier) { | |
199 | progress = true; | |
200 | } | |
201 | } | |
202 | ||
203 | tmp = node; | |
204 | node = QLIST_NEXT(node, node); | |
205 | ||
206 | ctx->walking_handlers--; | |
207 | ||
208 | if (!ctx->walking_handlers && tmp->deleted) { | |
209 | QLIST_REMOVE(tmp, node); | |
210 | g_free(tmp); | |
211 | } | |
212 | } | |
213 | ||
214 | /* Try again, but only call each handler once. */ | |
215 | events[ret - WAIT_OBJECT_0] = events[--count]; | |
216 | } | |
217 | ||
218 | if (blocking) { | |
219 | /* Run the timers a second time. We do this because otherwise aio_wait | |
220 | * will not note progress - and will stop a drain early - if we have | |
221 | * a timer that was not ready to run entering g_poll but is ready | |
222 | * after g_poll. This will only do anything if a timer has expired. | |
223 | */ | |
224 | progress |= timerlistgroup_run_timers(&ctx->tlg); | |
225 | } | |
226 | ||
227 | return progress; | |
228 | } |