]>
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 <aliguori@us.ibm.com> | |
9 | * Paolo Bonzini <pbonzini@redhat.com> | |
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/osdep.h" | |
19 | #include "qemu-common.h" | |
20 | #include "block/block.h" | |
21 | #include "qemu/queue.h" | |
22 | #include "qemu/sockets.h" | |
23 | ||
24 | struct AioHandler { | |
25 | EventNotifier *e; | |
26 | IOHandler *io_read; | |
27 | IOHandler *io_write; | |
28 | EventNotifierHandler *io_notify; | |
29 | GPollFD pfd; | |
30 | int deleted; | |
31 | void *opaque; | |
32 | bool is_external; | |
33 | QLIST_ENTRY(AioHandler) node; | |
34 | }; | |
35 | ||
36 | void aio_set_fd_handler(AioContext *ctx, | |
37 | int fd, | |
38 | bool is_external, | |
39 | IOHandler *io_read, | |
40 | IOHandler *io_write, | |
41 | void *opaque) | |
42 | { | |
43 | /* fd is a SOCKET in our case */ | |
44 | AioHandler *node; | |
45 | ||
46 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { | |
47 | if (node->pfd.fd == fd && !node->deleted) { | |
48 | break; | |
49 | } | |
50 | } | |
51 | ||
52 | /* Are we deleting the fd handler? */ | |
53 | if (!io_read && !io_write) { | |
54 | if (node) { | |
55 | /* If the lock is held, just mark the node as deleted */ | |
56 | if (ctx->walking_handlers) { | |
57 | node->deleted = 1; | |
58 | node->pfd.revents = 0; | |
59 | } else { | |
60 | /* Otherwise, delete it for real. We can't just mark it as | |
61 | * deleted because deleted nodes are only cleaned up after | |
62 | * releasing the walking_handlers lock. | |
63 | */ | |
64 | QLIST_REMOVE(node, node); | |
65 | g_free(node); | |
66 | } | |
67 | } | |
68 | } else { | |
69 | HANDLE event; | |
70 | ||
71 | if (node == NULL) { | |
72 | /* Alloc and insert if it's not already there */ | |
73 | node = g_new0(AioHandler, 1); | |
74 | node->pfd.fd = fd; | |
75 | QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node); | |
76 | } | |
77 | ||
78 | node->pfd.events = 0; | |
79 | if (node->io_read) { | |
80 | node->pfd.events |= G_IO_IN; | |
81 | } | |
82 | if (node->io_write) { | |
83 | node->pfd.events |= G_IO_OUT; | |
84 | } | |
85 | ||
86 | node->e = &ctx->notifier; | |
87 | ||
88 | /* Update handler with latest information */ | |
89 | node->opaque = opaque; | |
90 | node->io_read = io_read; | |
91 | node->io_write = io_write; | |
92 | node->is_external = is_external; | |
93 | ||
94 | event = event_notifier_get_handle(&ctx->notifier); | |
95 | WSAEventSelect(node->pfd.fd, event, | |
96 | FD_READ | FD_ACCEPT | FD_CLOSE | | |
97 | FD_CONNECT | FD_WRITE | FD_OOB); | |
98 | } | |
99 | ||
100 | aio_notify(ctx); | |
101 | } | |
102 | ||
103 | void aio_set_event_notifier(AioContext *ctx, | |
104 | EventNotifier *e, | |
105 | bool is_external, | |
106 | EventNotifierHandler *io_notify) | |
107 | { | |
108 | AioHandler *node; | |
109 | ||
110 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { | |
111 | if (node->e == e && !node->deleted) { | |
112 | break; | |
113 | } | |
114 | } | |
115 | ||
116 | /* Are we deleting the fd handler? */ | |
117 | if (!io_notify) { | |
118 | if (node) { | |
119 | g_source_remove_poll(&ctx->source, &node->pfd); | |
120 | ||
121 | /* If the lock is held, just mark the node as deleted */ | |
122 | if (ctx->walking_handlers) { | |
123 | node->deleted = 1; | |
124 | node->pfd.revents = 0; | |
125 | } else { | |
126 | /* Otherwise, delete it for real. We can't just mark it as | |
127 | * deleted because deleted nodes are only cleaned up after | |
128 | * releasing the walking_handlers lock. | |
129 | */ | |
130 | QLIST_REMOVE(node, node); | |
131 | g_free(node); | |
132 | } | |
133 | } | |
134 | } else { | |
135 | if (node == NULL) { | |
136 | /* Alloc and insert if it's not already there */ | |
137 | node = g_new0(AioHandler, 1); | |
138 | node->e = e; | |
139 | node->pfd.fd = (uintptr_t)event_notifier_get_handle(e); | |
140 | node->pfd.events = G_IO_IN; | |
141 | node->is_external = is_external; | |
142 | QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node); | |
143 | ||
144 | g_source_add_poll(&ctx->source, &node->pfd); | |
145 | } | |
146 | /* Update handler with latest information */ | |
147 | node->io_notify = io_notify; | |
148 | } | |
149 | ||
150 | aio_notify(ctx); | |
151 | } | |
152 | ||
153 | bool aio_prepare(AioContext *ctx) | |
154 | { | |
155 | static struct timeval tv0; | |
156 | AioHandler *node; | |
157 | bool have_select_revents = false; | |
158 | fd_set rfds, wfds; | |
159 | ||
160 | /* fill fd sets */ | |
161 | FD_ZERO(&rfds); | |
162 | FD_ZERO(&wfds); | |
163 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { | |
164 | if (node->io_read) { | |
165 | FD_SET ((SOCKET)node->pfd.fd, &rfds); | |
166 | } | |
167 | if (node->io_write) { | |
168 | FD_SET ((SOCKET)node->pfd.fd, &wfds); | |
169 | } | |
170 | } | |
171 | ||
172 | if (select(0, &rfds, &wfds, NULL, &tv0) > 0) { | |
173 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { | |
174 | node->pfd.revents = 0; | |
175 | if (FD_ISSET(node->pfd.fd, &rfds)) { | |
176 | node->pfd.revents |= G_IO_IN; | |
177 | have_select_revents = true; | |
178 | } | |
179 | ||
180 | if (FD_ISSET(node->pfd.fd, &wfds)) { | |
181 | node->pfd.revents |= G_IO_OUT; | |
182 | have_select_revents = true; | |
183 | } | |
184 | } | |
185 | } | |
186 | ||
187 | return have_select_revents; | |
188 | } | |
189 | ||
190 | bool aio_pending(AioContext *ctx) | |
191 | { | |
192 | AioHandler *node; | |
193 | ||
194 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { | |
195 | if (node->pfd.revents && node->io_notify) { | |
196 | return true; | |
197 | } | |
198 | ||
199 | if ((node->pfd.revents & G_IO_IN) && node->io_read) { | |
200 | return true; | |
201 | } | |
202 | if ((node->pfd.revents & G_IO_OUT) && node->io_write) { | |
203 | return true; | |
204 | } | |
205 | } | |
206 | ||
207 | return false; | |
208 | } | |
209 | ||
210 | static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) | |
211 | { | |
212 | AioHandler *node; | |
213 | bool progress = false; | |
214 | ||
215 | /* | |
216 | * We have to walk very carefully in case aio_set_fd_handler is | |
217 | * called while we're walking. | |
218 | */ | |
219 | node = QLIST_FIRST(&ctx->aio_handlers); | |
220 | while (node) { | |
221 | AioHandler *tmp; | |
222 | int revents = node->pfd.revents; | |
223 | ||
224 | ctx->walking_handlers++; | |
225 | ||
226 | if (!node->deleted && | |
227 | (revents || event_notifier_get_handle(node->e) == event) && | |
228 | node->io_notify) { | |
229 | node->pfd.revents = 0; | |
230 | node->io_notify(node->e); | |
231 | ||
232 | /* aio_notify() does not count as progress */ | |
233 | if (node->e != &ctx->notifier) { | |
234 | progress = true; | |
235 | } | |
236 | } | |
237 | ||
238 | if (!node->deleted && | |
239 | (node->io_read || node->io_write)) { | |
240 | node->pfd.revents = 0; | |
241 | if ((revents & G_IO_IN) && node->io_read) { | |
242 | node->io_read(node->opaque); | |
243 | progress = true; | |
244 | } | |
245 | if ((revents & G_IO_OUT) && node->io_write) { | |
246 | node->io_write(node->opaque); | |
247 | progress = true; | |
248 | } | |
249 | ||
250 | /* if the next select() will return an event, we have progressed */ | |
251 | if (event == event_notifier_get_handle(&ctx->notifier)) { | |
252 | WSANETWORKEVENTS ev; | |
253 | WSAEnumNetworkEvents(node->pfd.fd, event, &ev); | |
254 | if (ev.lNetworkEvents) { | |
255 | progress = true; | |
256 | } | |
257 | } | |
258 | } | |
259 | ||
260 | tmp = node; | |
261 | node = QLIST_NEXT(node, node); | |
262 | ||
263 | ctx->walking_handlers--; | |
264 | ||
265 | if (!ctx->walking_handlers && tmp->deleted) { | |
266 | QLIST_REMOVE(tmp, node); | |
267 | g_free(tmp); | |
268 | } | |
269 | } | |
270 | ||
271 | return progress; | |
272 | } | |
273 | ||
274 | bool aio_dispatch(AioContext *ctx) | |
275 | { | |
276 | bool progress; | |
277 | ||
278 | progress = aio_bh_poll(ctx); | |
279 | progress |= aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE); | |
280 | progress |= timerlistgroup_run_timers(&ctx->tlg); | |
281 | return progress; | |
282 | } | |
283 | ||
284 | bool aio_poll(AioContext *ctx, bool blocking) | |
285 | { | |
286 | AioHandler *node; | |
287 | HANDLE events[MAXIMUM_WAIT_OBJECTS + 1]; | |
288 | bool progress, have_select_revents, first; | |
289 | int count; | |
290 | int timeout; | |
291 | ||
292 | aio_context_acquire(ctx); | |
293 | progress = false; | |
294 | ||
295 | /* aio_notify can avoid the expensive event_notifier_set if | |
296 | * everything (file descriptors, bottom halves, timers) will | |
297 | * be re-evaluated before the next blocking poll(). This is | |
298 | * already true when aio_poll is called with blocking == false; | |
299 | * if blocking == true, it is only true after poll() returns, | |
300 | * so disable the optimization now. | |
301 | */ | |
302 | if (blocking) { | |
303 | atomic_add(&ctx->notify_me, 2); | |
304 | } | |
305 | ||
306 | have_select_revents = aio_prepare(ctx); | |
307 | ||
308 | ctx->walking_handlers++; | |
309 | ||
310 | /* fill fd sets */ | |
311 | count = 0; | |
312 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { | |
313 | if (!node->deleted && node->io_notify | |
314 | && aio_node_check(ctx, node->is_external)) { | |
315 | events[count++] = event_notifier_get_handle(node->e); | |
316 | } | |
317 | } | |
318 | ||
319 | ctx->walking_handlers--; | |
320 | first = true; | |
321 | ||
322 | /* ctx->notifier is always registered. */ | |
323 | assert(count > 0); | |
324 | ||
325 | /* Multiple iterations, all of them non-blocking except the first, | |
326 | * may be necessary to process all pending events. After the first | |
327 | * WaitForMultipleObjects call ctx->notify_me will be decremented. | |
328 | */ | |
329 | do { | |
330 | HANDLE event; | |
331 | int ret; | |
332 | ||
333 | timeout = blocking && !have_select_revents | |
334 | ? qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)) : 0; | |
335 | if (timeout) { | |
336 | aio_context_release(ctx); | |
337 | } | |
338 | ret = WaitForMultipleObjects(count, events, FALSE, timeout); | |
339 | if (blocking) { | |
340 | assert(first); | |
341 | atomic_sub(&ctx->notify_me, 2); | |
342 | } | |
343 | if (timeout) { | |
344 | aio_context_acquire(ctx); | |
345 | } | |
346 | ||
347 | if (first) { | |
348 | aio_notify_accept(ctx); | |
349 | progress |= aio_bh_poll(ctx); | |
350 | first = false; | |
351 | } | |
352 | ||
353 | /* if we have any signaled events, dispatch event */ | |
354 | event = NULL; | |
355 | if ((DWORD) (ret - WAIT_OBJECT_0) < count) { | |
356 | event = events[ret - WAIT_OBJECT_0]; | |
357 | events[ret - WAIT_OBJECT_0] = events[--count]; | |
358 | } else if (!have_select_revents) { | |
359 | break; | |
360 | } | |
361 | ||
362 | have_select_revents = false; | |
363 | blocking = false; | |
364 | ||
365 | progress |= aio_dispatch_handlers(ctx, event); | |
366 | } while (count > 0); | |
367 | ||
368 | progress |= timerlistgroup_run_timers(&ctx->tlg); | |
369 | ||
370 | aio_context_release(ctx); | |
371 | return progress; | |
372 | } | |
373 | ||
374 | void aio_context_setup(AioContext *ctx) | |
375 | { | |
376 | } |