xref: /libevent-2.1.12/devpoll.c (revision fee2c779)
1 /*
2  * Copyright 2000-2009 Niels Provos <[email protected]>
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  * 3. The name of the author may not be used to endorse or promote products
13  *    derived from this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 #include "event-config.h"
27 
28 #include <sys/types.h>
29 #include <sys/resource.h>
30 #ifdef _EVENT_HAVE_SYS_TIME_H
31 #include <sys/time.h>
32 #endif
33 #include <sys/queue.h>
34 #include <sys/devpoll.h>
35 #include <signal.h>
36 #include <stdio.h>
37 #include <stdlib.h>
38 #include <string.h>
39 #include <unistd.h>
40 #include <fcntl.h>
41 #include <errno.h>
42 
43 #include "event2/event.h"
44 #include "event2/event_struct.h"
45 #include "event2/thread.h"
46 #include "event-internal.h"
47 #include "evsignal-internal.h"
48 #include "log-internal.h"
49 #include "evmap-internal.h"
50 #include "evthread-internal.h"
51 
52 struct devpollop {
53 	struct pollfd *events;
54 	int nevents;
55 	int dpfd;
56 	struct pollfd *changes;
57 	int nchanges;
58 };
59 
60 static void *devpoll_init	(struct event_base *);
61 static int devpoll_add(struct event_base *, int fd, short old, short events, void *);
62 static int devpoll_del(struct event_base *, int fd, short old, short events, void *);
63 static int devpoll_dispatch	(struct event_base *, struct timeval *);
64 static void devpoll_dealloc	(struct event_base *);
65 
66 const struct eventop devpollops = {
67 	"devpoll",
68 	devpoll_init,
69 	devpoll_add,
70 	devpoll_del,
71 	devpoll_dispatch,
72 	devpoll_dealloc,
73 	1, /* need reinit */
74 	EV_FEATURE_FDS|EV_FEATURE_O1,
75 	0
76 };
77 
78 #define NEVENT	32000
79 
80 static int
81 devpoll_commit(struct devpollop *devpollop)
82 {
83 	/*
84 	 * Due to a bug in Solaris, we have to use pwrite with an offset of 0.
85 	 * Write is limited to 2GB of data, until it will fail.
86 	 */
87 	if (pwrite(devpollop->dpfd, devpollop->changes,
88 		sizeof(struct pollfd) * devpollop->nchanges, 0) == -1)
89 		return(-1);
90 
91 	devpollop->nchanges = 0;
92 	return(0);
93 }
94 
95 static int
96 devpoll_queue(struct devpollop *devpollop, int fd, int events) {
97 	struct pollfd *pfd;
98 
99 	if (devpollop->nchanges >= devpollop->nevents) {
100 		/*
101 		 * Change buffer is full, must commit it to /dev/poll before
102 		 * adding more
103 		 */
104 		if (devpoll_commit(devpollop) != 0)
105 			return(-1);
106 	}
107 
108 	pfd = &devpollop->changes[devpollop->nchanges++];
109 	pfd->fd = fd;
110 	pfd->events = events;
111 	pfd->revents = 0;
112 
113 	return(0);
114 }
115 
116 static void *
117 devpoll_init(struct event_base *base)
118 {
119 	int dpfd, nfiles = NEVENT;
120 	struct rlimit rl;
121 	struct devpollop *devpollop;
122 
123 	if (!(devpollop = mm_calloc(1, sizeof(struct devpollop))))
124 		return (NULL);
125 
126 	if (getrlimit(RLIMIT_NOFILE, &rl) == 0 &&
127 	    rl.rlim_cur != RLIM_INFINITY)
128 		nfiles = rl.rlim_cur;
129 
130 	/* Initialize the kernel queue */
131 	if ((dpfd = open("/dev/poll", O_RDWR)) == -1) {
132                 event_warn("open: /dev/poll");
133 		mm_free(devpollop);
134 		return (NULL);
135 	}
136 
137 	devpollop->dpfd = dpfd;
138 
139 	/* Initialize fields */
140 	/* FIXME: allocating 'nfiles' worth of space here can be
141 	 * expensive and unnecessary.  See how epoll.c does it instead. */
142 	devpollop->events = mm_calloc(nfiles, sizeof(struct pollfd));
143 	if (devpollop->events == NULL) {
144 		mm_free(devpollop);
145 		close(dpfd);
146 		return (NULL);
147 	}
148 	devpollop->nevents = nfiles;
149 
150 	devpollop->changes = mm_calloc(nfiles, sizeof(struct pollfd));
151 	if (devpollop->changes == NULL) {
152 		mm_free(devpollop->events);
153 		mm_free(devpollop);
154 		close(dpfd);
155 		return (NULL);
156 	}
157 
158 	evsig_init(base);
159 
160 	return (devpollop);
161 }
162 
163 static int
164 devpoll_dispatch(struct event_base *base, struct timeval *tv)
165 {
166 	struct devpollop *devpollop = base->evbase;
167 	struct pollfd *events = devpollop->events;
168 	struct dvpoll dvp;
169 	int i, res, timeout = -1;
170 
171 	if (devpollop->nchanges)
172 		devpoll_commit(devpollop);
173 
174 	if (tv != NULL)
175 		timeout = tv->tv_sec * 1000 + (tv->tv_usec + 999) / 1000;
176 
177 	dvp.dp_fds = devpollop->events;
178 	dvp.dp_nfds = devpollop->nevents;
179 	dvp.dp_timeout = timeout;
180 
181 	EVBASE_RELEASE_LOCK(base, th_base_lock);
182 
183 	res = ioctl(devpollop->dpfd, DP_POLL, &dvp);
184 
185 	EVBASE_ACQUIRE_LOCK(base, th_base_lock);
186 
187 	if (res == -1) {
188 		if (errno != EINTR) {
189 			event_warn("ioctl: DP_POLL");
190 			return (-1);
191 		}
192 
193 		evsig_process(base);
194 		return (0);
195 	} else if (base->sig.evsig_caught) {
196 		evsig_process(base);
197 	}
198 
199 	event_debug(("%s: devpoll_wait reports %d", __func__, res));
200 
201 	for (i = 0; i < res; i++) {
202 		int which = 0;
203 		int what = events[i].revents;
204 
205                 if (what & POLLHUP)
206                         what |= POLLIN | POLLOUT;
207                 else if (what & POLLERR)
208                         what |= POLLIN | POLLOUT;
209 
210 		if (what & POLLIN)
211 			which |= EV_READ;
212 		if (what & POLLOUT)
213 			which |= EV_WRITE;
214 
215 		if (!which)
216 			continue;
217 
218 		/* XXX(niels): not sure if this works for devpoll */
219 		evmap_io_active(base, events[i].fd, which);
220 	}
221 
222 	return (0);
223 }
224 
225 
226 static int
227 devpoll_add(struct event_base *base, int fd, short old, short events, void *p)
228 {
229 	struct devpollop *devpollop = base->evbase;
230 	int res;
231 	(void)p;
232 
233 	/*
234 	 * It's not necessary to OR the existing read/write events that we
235 	 * are currently interested in with the new event we are adding.
236 	 * The /dev/poll driver ORs any new events with the existing events
237 	 * that it has cached for the fd.
238 	 */
239 
240 	res = 0;
241 	if (events & EV_READ)
242 		res |= POLLIN;
243 	if (events & EV_WRITE)
244 		res |= POLLOUT;
245 
246 	if (devpoll_queue(devpollop, fd, res) != 0)
247 		return(-1);
248 
249 	return (0);
250 }
251 
252 static int
253 devpoll_del(struct event_base *base, int fd, short old, short events, void *p)
254 {
255 	struct devpollop *devpollop = base->evbase;
256 	int res;
257 	(void)p;
258 
259 	res = 0;
260 	if (events & EV_READ)
261 		res |= POLLIN;
262 	if (events & EV_WRITE)
263 		res |= POLLOUT;
264 
265 	/*
266 	 * The only way to remove an fd from the /dev/poll monitored set is
267 	 * to use POLLREMOVE by itself.  This removes ALL events for the fd
268 	 * provided so if we care about two events and are only removing one
269 	 * we must re-add the other event after POLLREMOVE.
270 	 */
271 
272 	if (devpoll_queue(devpollop, fd, POLLREMOVE) != 0)
273 		return(-1);
274 
275 	if ((res & (POLLIN|POLLOUT)) != (POLLIN|POLLOUT)) {
276 		/*
277 		 * We're not deleting all events, so we must resubmit the
278 		 * event that we are still interested in if one exists.
279 		 */
280 
281 		if ((res & POLLIN) && (old & EV_WRITE)) {
282 			/* Deleting read, still care about write */
283 			devpoll_queue(devpollop, fd, POLLOUT);
284 		} else if ((res & POLLOUT) && (old & EV_READ)) {
285 			/* Deleting write, still care about read */
286 			devpoll_queue(devpollop, fd, POLLIN);
287 		}
288 	}
289 
290 	return (0);
291 }
292 
293 static void
294 devpoll_dealloc(struct event_base *base)
295 {
296 	struct devpollop *devpollop = base->evbase;
297 
298 	evsig_dealloc(base);
299 	if (devpollop->events)
300 		mm_free(devpollop->events);
301 	if (devpollop->changes)
302 		mm_free(devpollop->changes);
303 	if (devpollop->dpfd >= 0)
304 		close(devpollop->dpfd);
305 
306 	memset(devpollop, 0, sizeof(struct devpollop));
307 	mm_free(devpollop);
308 }
309