xref: /openbmc/qemu/util/aio-win32.c (revision f7214f99)
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 "block/block.h"
20 #include "qemu/main-loop.h"
21 #include "qemu/lockcnt.h"
22 #include "qemu/queue.h"
23 #include "qemu/sockets.h"
24 #include "qapi/error.h"
25 #include "qemu/rcu_queue.h"
26 #include "qemu/error-report.h"
27 
28 struct AioHandler {
29     EventNotifier *e;
30     IOHandler *io_read;
31     IOHandler *io_write;
32     EventNotifierHandler *io_notify;
33     GPollFD pfd;
34     int deleted;
35     void *opaque;
36     QLIST_ENTRY(AioHandler) node;
37 };
38 
39 static void aio_remove_fd_handler(AioContext *ctx, AioHandler *node)
40 {
41     /*
42      * If the GSource is in the process of being destroyed then
43      * g_source_remove_poll() causes an assertion failure.  Skip
44      * removal in that case, because glib cleans up its state during
45      * destruction anyway.
46      */
47     if (!g_source_is_destroyed(&ctx->source)) {
48         g_source_remove_poll(&ctx->source, &node->pfd);
49     }
50 
51     /* If aio_poll is in progress, just mark the node as deleted */
52     if (qemu_lockcnt_count(&ctx->list_lock)) {
53         node->deleted = 1;
54         node->pfd.revents = 0;
55     } else {
56         /* Otherwise, delete it for real.  We can't just mark it as
57          * deleted because deleted nodes are only cleaned up after
58          * releasing the list_lock.
59          */
60         QLIST_REMOVE(node, node);
61         g_free(node);
62     }
63 }
64 
65 void aio_set_fd_handler(AioContext *ctx,
66                         int fd,
67                         IOHandler *io_read,
68                         IOHandler *io_write,
69                         AioPollFn *io_poll,
70                         IOHandler *io_poll_ready,
71                         void *opaque)
72 {
73     AioHandler *old_node;
74     AioHandler *node = NULL;
75     SOCKET s;
76 
77     if (!fd_is_socket(fd)) {
78         error_report("fd=%d is not a socket, AIO implementation is missing", fd);
79         return;
80     }
81 
82     s = _get_osfhandle(fd);
83 
84     qemu_lockcnt_lock(&ctx->list_lock);
85     QLIST_FOREACH(old_node, &ctx->aio_handlers, node) {
86         if (old_node->pfd.fd == s && !old_node->deleted) {
87             break;
88         }
89     }
90 
91     if (io_read || io_write) {
92         HANDLE event;
93         long bitmask = 0;
94 
95         /* Alloc and insert if it's not already there */
96         node = g_new0(AioHandler, 1);
97         node->pfd.fd = s;
98 
99         node->pfd.events = 0;
100         if (node->io_read) {
101             node->pfd.events |= G_IO_IN;
102         }
103         if (node->io_write) {
104             node->pfd.events |= G_IO_OUT;
105         }
106 
107         node->e = &ctx->notifier;
108 
109         /* Update handler with latest information */
110         node->opaque = opaque;
111         node->io_read = io_read;
112         node->io_write = io_write;
113 
114         if (io_read) {
115             bitmask |= FD_READ | FD_ACCEPT | FD_CLOSE;
116         }
117 
118         if (io_write) {
119             bitmask |= FD_WRITE | FD_CONNECT;
120         }
121 
122         QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node);
123         event = event_notifier_get_handle(&ctx->notifier);
124         qemu_socket_select(fd, event, bitmask, NULL);
125     }
126     if (old_node) {
127         aio_remove_fd_handler(ctx, old_node);
128     }
129 
130     qemu_lockcnt_unlock(&ctx->list_lock);
131     aio_notify(ctx);
132 }
133 
134 void aio_set_event_notifier(AioContext *ctx,
135                             EventNotifier *e,
136                             EventNotifierHandler *io_notify,
137                             AioPollFn *io_poll,
138                             EventNotifierHandler *io_poll_ready)
139 {
140     AioHandler *node;
141 
142     qemu_lockcnt_lock(&ctx->list_lock);
143     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
144         if (node->e == e && !node->deleted) {
145             break;
146         }
147     }
148 
149     /* Are we deleting the fd handler? */
150     if (!io_notify) {
151         if (node) {
152             aio_remove_fd_handler(ctx, node);
153         }
154     } else {
155         if (node == NULL) {
156             /* Alloc and insert if it's not already there */
157             node = g_new0(AioHandler, 1);
158             node->e = e;
159             node->pfd.fd = (uintptr_t)event_notifier_get_handle(e);
160             node->pfd.events = G_IO_IN;
161             QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node);
162 
163             g_source_add_poll(&ctx->source, &node->pfd);
164         }
165         /* Update handler with latest information */
166         node->io_notify = io_notify;
167     }
168 
169     qemu_lockcnt_unlock(&ctx->list_lock);
170     aio_notify(ctx);
171 }
172 
173 void aio_set_event_notifier_poll(AioContext *ctx,
174                                  EventNotifier *notifier,
175                                  EventNotifierHandler *io_poll_begin,
176                                  EventNotifierHandler *io_poll_end)
177 {
178     /* Not implemented */
179 }
180 
181 bool aio_prepare(AioContext *ctx)
182 {
183     static struct timeval tv0;
184     AioHandler *node;
185     bool have_select_revents = false;
186     fd_set rfds, wfds;
187 
188     /*
189      * We have to walk very carefully in case aio_set_fd_handler is
190      * called while we're walking.
191      */
192     qemu_lockcnt_inc(&ctx->list_lock);
193 
194     /* fill fd sets */
195     FD_ZERO(&rfds);
196     FD_ZERO(&wfds);
197     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
198         if (node->io_read) {
199             FD_SET ((SOCKET)node->pfd.fd, &rfds);
200         }
201         if (node->io_write) {
202             FD_SET ((SOCKET)node->pfd.fd, &wfds);
203         }
204     }
205 
206     if (select(0, &rfds, &wfds, NULL, &tv0) > 0) {
207         QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
208             node->pfd.revents = 0;
209             if (FD_ISSET(node->pfd.fd, &rfds)) {
210                 node->pfd.revents |= G_IO_IN;
211                 have_select_revents = true;
212             }
213 
214             if (FD_ISSET(node->pfd.fd, &wfds)) {
215                 node->pfd.revents |= G_IO_OUT;
216                 have_select_revents = true;
217             }
218         }
219     }
220 
221     qemu_lockcnt_dec(&ctx->list_lock);
222     return have_select_revents;
223 }
224 
225 bool aio_pending(AioContext *ctx)
226 {
227     AioHandler *node;
228     bool result = false;
229 
230     /*
231      * We have to walk very carefully in case aio_set_fd_handler is
232      * called while we're walking.
233      */
234     qemu_lockcnt_inc(&ctx->list_lock);
235     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
236         if (node->pfd.revents && node->io_notify) {
237             result = true;
238             break;
239         }
240 
241         if ((node->pfd.revents & G_IO_IN) && node->io_read) {
242             result = true;
243             break;
244         }
245         if ((node->pfd.revents & G_IO_OUT) && node->io_write) {
246             result = true;
247             break;
248         }
249     }
250 
251     qemu_lockcnt_dec(&ctx->list_lock);
252     return result;
253 }
254 
255 static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event)
256 {
257     AioHandler *node;
258     bool progress = false;
259     AioHandler *tmp;
260 
261     /*
262      * We have to walk very carefully in case aio_set_fd_handler is
263      * called while we're walking.
264      */
265     QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) {
266         int revents = node->pfd.revents;
267 
268         if (!node->deleted &&
269             (revents || event_notifier_get_handle(node->e) == event) &&
270             node->io_notify) {
271             node->pfd.revents = 0;
272             node->io_notify(node->e);
273 
274             /* aio_notify() does not count as progress */
275             if (node->e != &ctx->notifier) {
276                 progress = true;
277             }
278         }
279 
280         if (!node->deleted &&
281             (node->io_read || node->io_write)) {
282             node->pfd.revents = 0;
283             if ((revents & G_IO_IN) && node->io_read) {
284                 node->io_read(node->opaque);
285                 progress = true;
286             }
287             if ((revents & G_IO_OUT) && node->io_write) {
288                 node->io_write(node->opaque);
289                 progress = true;
290             }
291 
292             /* if the next select() will return an event, we have progressed */
293             if (event == event_notifier_get_handle(&ctx->notifier)) {
294                 WSANETWORKEVENTS ev;
295                 WSAEnumNetworkEvents(node->pfd.fd, event, &ev);
296                 if (ev.lNetworkEvents) {
297                     progress = true;
298                 }
299             }
300         }
301 
302         if (node->deleted) {
303             if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
304                 QLIST_REMOVE(node, node);
305                 g_free(node);
306                 qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
307             }
308         }
309     }
310 
311     return progress;
312 }
313 
314 void aio_dispatch(AioContext *ctx)
315 {
316     qemu_lockcnt_inc(&ctx->list_lock);
317     aio_bh_poll(ctx);
318     aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE);
319     qemu_lockcnt_dec(&ctx->list_lock);
320     timerlistgroup_run_timers(&ctx->tlg);
321 }
322 
323 bool aio_poll(AioContext *ctx, bool blocking)
324 {
325     AioHandler *node;
326     HANDLE events[MAXIMUM_WAIT_OBJECTS];
327     bool progress, have_select_revents, first;
328     unsigned count;
329     int timeout;
330 
331     /*
332      * There cannot be two concurrent aio_poll calls for the same AioContext (or
333      * an aio_poll concurrent with a GSource prepare/check/dispatch callback).
334      * We rely on this below to avoid slow locked accesses to ctx->notify_me.
335      *
336      * aio_poll() may only be called in the AioContext's thread. iohandler_ctx
337      * is special in that it runs in the main thread, but that thread's context
338      * is qemu_aio_context.
339      */
340     assert(in_aio_context_home_thread(ctx == iohandler_get_aio_context() ?
341                                       qemu_get_aio_context() : ctx));
342     progress = false;
343 
344     /* aio_notify can avoid the expensive event_notifier_set if
345      * everything (file descriptors, bottom halves, timers) will
346      * be re-evaluated before the next blocking poll().  This is
347      * already true when aio_poll is called with blocking == false;
348      * if blocking == true, it is only true after poll() returns,
349      * so disable the optimization now.
350      */
351     if (blocking) {
352         qatomic_set(&ctx->notify_me, qatomic_read(&ctx->notify_me) + 2);
353         /*
354          * Write ctx->notify_me before computing the timeout
355          * (reading bottom half flags, etc.).  Pairs with
356          * smp_mb in aio_notify().
357          */
358         smp_mb();
359     }
360 
361     qemu_lockcnt_inc(&ctx->list_lock);
362     have_select_revents = aio_prepare(ctx);
363 
364     /* fill fd sets */
365     count = 0;
366     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
367         if (!node->deleted && node->io_notify) {
368             assert(count < MAXIMUM_WAIT_OBJECTS);
369             events[count++] = event_notifier_get_handle(node->e);
370         }
371     }
372 
373     first = true;
374 
375     /* ctx->notifier is always registered.  */
376     assert(count > 0);
377 
378     /* Multiple iterations, all of them non-blocking except the first,
379      * may be necessary to process all pending events.  After the first
380      * WaitForMultipleObjects call ctx->notify_me will be decremented.
381      */
382     do {
383         HANDLE event;
384         int ret;
385 
386         timeout = blocking && !have_select_revents
387             ? qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)) : 0;
388         ret = WaitForMultipleObjects(count, events, FALSE, timeout);
389         if (blocking) {
390             assert(first);
391             qatomic_store_release(&ctx->notify_me,
392                                   qatomic_read(&ctx->notify_me) - 2);
393             aio_notify_accept(ctx);
394         }
395 
396         if (first) {
397             progress |= aio_bh_poll(ctx);
398             first = false;
399         }
400 
401         /* if we have any signaled events, dispatch event */
402         event = NULL;
403         if ((DWORD) (ret - WAIT_OBJECT_0) < count) {
404             event = events[ret - WAIT_OBJECT_0];
405             events[ret - WAIT_OBJECT_0] = events[--count];
406         } else if (!have_select_revents) {
407             break;
408         }
409 
410         have_select_revents = false;
411         blocking = false;
412 
413         progress |= aio_dispatch_handlers(ctx, event);
414     } while (count > 0);
415 
416     qemu_lockcnt_dec(&ctx->list_lock);
417 
418     progress |= timerlistgroup_run_timers(&ctx->tlg);
419     return progress;
420 }
421 
422 void aio_context_setup(AioContext *ctx)
423 {
424 }
425 
426 void aio_context_destroy(AioContext *ctx)
427 {
428 }
429 
430 void aio_context_use_g_source(AioContext *ctx)
431 {
432 }
433 
434 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns,
435                                  int64_t grow, int64_t shrink, Error **errp)
436 {
437     if (max_ns) {
438         error_setg(errp, "AioContext polling is not implemented on Windows");
439     }
440 }
441 
442 void aio_context_set_aio_params(AioContext *ctx, int64_t max_batch)
443 {
444 }
445