AioContext: optimize clearing the EventNotifier
[qemu.git] / aio-posix.c
1 /*
2 * QEMU aio implementation
3 *
4 * Copyright IBM, Corp. 2008
5 *
6 * Authors:
7 * Anthony Liguori <aliguori@us.ibm.com>
8 *
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
11 *
12 * Contributions after 2012-01-13 are licensed under the terms of the
13 * GNU GPL, version 2 or (at your option) any later version.
14 */
15
16 #include "qemu-common.h"
17 #include "block/block.h"
18 #include "qemu/queue.h"
19 #include "qemu/sockets.h"
20
21 struct AioHandler
22 {
23 GPollFD pfd;
24 IOHandler *io_read;
25 IOHandler *io_write;
26 int deleted;
27 void *opaque;
28 QLIST_ENTRY(AioHandler) node;
29 };
30
31 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
32 {
33 AioHandler *node;
34
35 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
36 if (node->pfd.fd == fd)
37 if (!node->deleted)
38 return node;
39 }
40
41 return NULL;
42 }
43
44 void aio_set_fd_handler(AioContext *ctx,
45 int fd,
46 IOHandler *io_read,
47 IOHandler *io_write,
48 void *opaque)
49 {
50 AioHandler *node;
51
52 node = find_aio_handler(ctx, fd);
53
54 /* Are we deleting the fd handler? */
55 if (!io_read && !io_write) {
56 if (node) {
57 g_source_remove_poll(&ctx->source, &node->pfd);
58
59 /* If the lock is held, just mark the node as deleted */
60 if (ctx->walking_handlers) {
61 node->deleted = 1;
62 node->pfd.revents = 0;
63 } else {
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 walking_handlers lock.
67 */
68 QLIST_REMOVE(node, node);
69 g_free(node);
70 }
71 }
72 } else {
73 if (node == NULL) {
74 /* Alloc and insert if it's not already there */
75 node = g_new0(AioHandler, 1);
76 node->pfd.fd = fd;
77 QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
78
79 g_source_add_poll(&ctx->source, &node->pfd);
80 }
81 /* Update handler with latest information */
82 node->io_read = io_read;
83 node->io_write = io_write;
84 node->opaque = opaque;
85
86 node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
87 node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
88 }
89
90 aio_notify(ctx);
91 }
92
93 void aio_set_event_notifier(AioContext *ctx,
94 EventNotifier *notifier,
95 EventNotifierHandler *io_read)
96 {
97 aio_set_fd_handler(ctx, event_notifier_get_fd(notifier),
98 (IOHandler *)io_read, NULL, notifier);
99 }
100
101 bool aio_prepare(AioContext *ctx)
102 {
103 return false;
104 }
105
106 bool aio_pending(AioContext *ctx)
107 {
108 AioHandler *node;
109
110 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
111 int revents;
112
113 revents = node->pfd.revents & node->pfd.events;
114 if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read) {
115 return true;
116 }
117 if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write) {
118 return true;
119 }
120 }
121
122 return false;
123 }
124
125 bool aio_dispatch(AioContext *ctx)
126 {
127 AioHandler *node;
128 bool progress = false;
129
130 /*
131 * If there are callbacks left that have been queued, we need to call them.
132 * Do not call select in this case, because it is possible that the caller
133 * does not need a complete flush (as is the case for aio_poll loops).
134 */
135 if (aio_bh_poll(ctx)) {
136 progress = true;
137 }
138
139 /*
140 * We have to walk very carefully in case aio_set_fd_handler is
141 * called while we're walking.
142 */
143 node = QLIST_FIRST(&ctx->aio_handlers);
144 while (node) {
145 AioHandler *tmp;
146 int revents;
147
148 ctx->walking_handlers++;
149
150 revents = node->pfd.revents & node->pfd.events;
151 node->pfd.revents = 0;
152
153 if (!node->deleted &&
154 (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
155 node->io_read) {
156 node->io_read(node->opaque);
157
158 /* aio_notify() does not count as progress */
159 if (node->opaque != &ctx->notifier) {
160 progress = true;
161 }
162 }
163 if (!node->deleted &&
164 (revents & (G_IO_OUT | G_IO_ERR)) &&
165 node->io_write) {
166 node->io_write(node->opaque);
167 progress = true;
168 }
169
170 tmp = node;
171 node = QLIST_NEXT(node, node);
172
173 ctx->walking_handlers--;
174
175 if (!ctx->walking_handlers && tmp->deleted) {
176 QLIST_REMOVE(tmp, node);
177 g_free(tmp);
178 }
179 }
180
181 /* Run our timers */
182 progress |= timerlistgroup_run_timers(&ctx->tlg);
183
184 return progress;
185 }
186
187 /* These thread-local variables are used only in a small part of aio_poll
188 * around the call to the poll() system call. In particular they are not
189 * used while aio_poll is performing callbacks, which makes it much easier
190 * to think about reentrancy!
191 *
192 * Stack-allocated arrays would be perfect but they have size limitations;
193 * heap allocation is expensive enough that we want to reuse arrays across
194 * calls to aio_poll(). And because poll() has to be called without holding
195 * any lock, the arrays cannot be stored in AioContext. Thread-local data
196 * has none of the disadvantages of these three options.
197 */
198 static __thread GPollFD *pollfds;
199 static __thread AioHandler **nodes;
200 static __thread unsigned npfd, nalloc;
201 static __thread Notifier pollfds_cleanup_notifier;
202
203 static void pollfds_cleanup(Notifier *n, void *unused)
204 {
205 g_assert(npfd == 0);
206 g_free(pollfds);
207 g_free(nodes);
208 nalloc = 0;
209 }
210
211 static void add_pollfd(AioHandler *node)
212 {
213 if (npfd == nalloc) {
214 if (nalloc == 0) {
215 pollfds_cleanup_notifier.notify = pollfds_cleanup;
216 qemu_thread_atexit_add(&pollfds_cleanup_notifier);
217 nalloc = 8;
218 } else {
219 g_assert(nalloc <= INT_MAX);
220 nalloc *= 2;
221 }
222 pollfds = g_renew(GPollFD, pollfds, nalloc);
223 nodes = g_renew(AioHandler *, nodes, nalloc);
224 }
225 nodes[npfd] = node;
226 pollfds[npfd] = (GPollFD) {
227 .fd = node->pfd.fd,
228 .events = node->pfd.events,
229 };
230 npfd++;
231 }
232
233 bool aio_poll(AioContext *ctx, bool blocking)
234 {
235 AioHandler *node;
236 int i, ret;
237 bool progress;
238 int64_t timeout;
239
240 aio_context_acquire(ctx);
241 progress = false;
242
243 /* aio_notify can avoid the expensive event_notifier_set if
244 * everything (file descriptors, bottom halves, timers) will
245 * be re-evaluated before the next blocking poll(). This is
246 * already true when aio_poll is called with blocking == false;
247 * if blocking == true, it is only true after poll() returns,
248 * so disable the optimization now.
249 */
250 if (blocking) {
251 atomic_add(&ctx->notify_me, 2);
252 }
253
254 ctx->walking_handlers++;
255
256 assert(npfd == 0);
257
258 /* fill pollfds */
259 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
260 if (!node->deleted && node->pfd.events) {
261 add_pollfd(node);
262 }
263 }
264
265 timeout = blocking ? aio_compute_timeout(ctx) : 0;
266
267 /* wait until next event */
268 if (timeout) {
269 aio_context_release(ctx);
270 }
271 ret = qemu_poll_ns((GPollFD *)pollfds, npfd, timeout);
272 if (blocking) {
273 atomic_sub(&ctx->notify_me, 2);
274 }
275 if (timeout) {
276 aio_context_acquire(ctx);
277 }
278
279 aio_notify_accept(ctx);
280
281 /* if we have any readable fds, dispatch event */
282 if (ret > 0) {
283 for (i = 0; i < npfd; i++) {
284 nodes[i]->pfd.revents = pollfds[i].revents;
285 }
286 }
287
288 npfd = 0;
289 ctx->walking_handlers--;
290
291 /* Run dispatch even if there were no readable fds to run timers */
292 if (aio_dispatch(ctx)) {
293 progress = true;
294 }
295
296 aio_context_release(ctx);
297
298 return progress;
299 }