xref: /libevent-2.1.12/devpoll.c (revision 2e36dbe1)
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 #ifdef HAVE_CONFIG_H
27 #include "event-config.h"
28 #endif
29 
30 #include <sys/types.h>
31 #include <sys/resource.h>
32 #ifdef _EVENT_HAVE_SYS_TIME_H
33 #include <sys/time.h>
34 #else
35 #include <sys/_time.h>
36 #endif
37 #include <sys/queue.h>
38 #include <sys/devpoll.h>
39 #include <signal.h>
40 #include <stdio.h>
41 #include <stdlib.h>
42 #include <string.h>
43 #include <unistd.h>
44 #include <fcntl.h>
45 #include <errno.h>
46 
47 #include "event2/event.h"
48 #include "event2/event_struct.h"
49 #include "event-internal.h"
50 #include "evsignal-internal.h"
51 #include "log-internal.h"
52 #include "evmap-internal.h"
53 
54 struct devpollop {
55 	struct pollfd *events;
56 	int nevents;
57 	int dpfd;
58 	struct pollfd *changes;
59 	int nchanges;
60 };
61 
62 static void *devpoll_init	(struct event_base *);
63 static int devpoll_add(struct event_base *, int fd, short old, short events, void *);
64 static int devpoll_del(struct event_base *, int fd, short old, short events, void *);
65 static int devpoll_dispatch	(struct event_base *, struct timeval *);
66 static void devpoll_dealloc	(struct event_base *);
67 
68 const struct eventop devpollops = {
69 	"devpoll",
70 	devpoll_init,
71 	devpoll_add,
72 	devpoll_del,
73 	devpoll_dispatch,
74 	devpoll_dealloc,
75 	1, /* need reinit */
76 	EV_FEATURE_FDS|EV_FEATURE_O1,
77 	0
78 };
79 
80 #define NEVENT	32000
81 
82 static int
83 devpoll_commit(struct devpollop *devpollop)
84 {
85 	/*
86 	 * Due to a bug in Solaris, we have to use pwrite with an offset of 0.
87 	 * Write is limited to 2GB of data, until it will fail.
88 	 */
89 	if (pwrite(devpollop->dpfd, devpollop->changes,
90 		sizeof(struct pollfd) * devpollop->nchanges, 0) == -1)
91 		return(-1);
92 
93 	devpollop->nchanges = 0;
94 	return(0);
95 }
96 
97 static int
98 devpoll_queue(struct devpollop *devpollop, int fd, int events) {
99 	struct pollfd *pfd;
100 
101 	if (devpollop->nchanges >= devpollop->nevents) {
102 		/*
103 		 * Change buffer is full, must commit it to /dev/poll before
104 		 * adding more
105 		 */
106 		if (devpoll_commit(devpollop) != 0)
107 			return(-1);
108 	}
109 
110 	pfd = &devpollop->changes[devpollop->nchanges++];
111 	pfd->fd = fd;
112 	pfd->events = events;
113 	pfd->revents = 0;
114 
115 	return(0);
116 }
117 
118 static void *
119 devpoll_init(struct event_base *base)
120 {
121 	int dpfd, nfiles = NEVENT;
122 	struct rlimit rl;
123 	struct devpollop *devpollop;
124 
125 	if (!(devpollop = mm_calloc(1, sizeof(struct devpollop))))
126 		return (NULL);
127 
128 	if (getrlimit(RLIMIT_NOFILE, &rl) == 0 &&
129 	    rl.rlim_cur != RLIM_INFINITY)
130 		nfiles = rl.rlim_cur;
131 
132 	/* Initialize the kernel queue */
133 	if ((dpfd = open("/dev/poll", O_RDWR)) == -1) {
134                 event_warn("open: /dev/poll");
135 		mm_free(devpollop);
136 		return (NULL);
137 	}
138 
139 	devpollop->dpfd = dpfd;
140 
141 	/* Initialize fields */
142 	/* FIXME: allocating 'nfiles' worth of space here can be
143 	 * expensive and unnecessary.  See how epoll.c does it instead. */
144 	devpollop->events = mm_calloc(nfiles, sizeof(struct pollfd));
145 	if (devpollop->events == NULL) {
146 		mm_free(devpollop);
147 		close(dpfd);
148 		return (NULL);
149 	}
150 	devpollop->nevents = nfiles;
151 
152 	devpollop->changes = mm_calloc(nfiles, sizeof(struct pollfd));
153 	if (devpollop->changes == NULL) {
154 		mm_free(devpollop->events);
155 		mm_free(devpollop);
156 		close(dpfd);
157 		return (NULL);
158 	}
159 
160 	evsig_init(base);
161 
162 	return (devpollop);
163 }
164 
165 static int
166 devpoll_dispatch(struct event_base *base, struct timeval *tv)
167 {
168 	struct devpollop *devpollop = base->evbase;
169 	struct pollfd *events = devpollop->events;
170 	struct dvpoll dvp;
171 	int i, res, timeout = -1;
172 
173 	if (devpollop->nchanges)
174 		devpoll_commit(devpollop);
175 
176 	if (tv != NULL)
177 		timeout = tv->tv_sec * 1000 + (tv->tv_usec + 999) / 1000;
178 
179 	dvp.dp_fds = devpollop->events;
180 	dvp.dp_nfds = devpollop->nevents;
181 	dvp.dp_timeout = timeout;
182 
183 	EVBASE_RELEASE_LOCK(base, EVTHREAD_WRITE, th_base_lock);
184 
185 	res = ioctl(devpollop->dpfd, DP_POLL, &dvp);
186 
187 	EVBASE_ACQUIRE_LOCK(base, EVTHREAD_WRITE, th_base_lock);
188 
189 	if (res == -1) {
190 		if (errno != EINTR) {
191 			event_warn("ioctl: DP_POLL");
192 			return (-1);
193 		}
194 
195 		evsig_process(base);
196 		return (0);
197 	} else if (base->sig.evsig_caught) {
198 		evsig_process(base);
199 	}
200 
201 	event_debug(("%s: devpoll_wait reports %d", __func__, res));
202 
203 	for (i = 0; i < res; i++) {
204 		int which = 0;
205 		int what = events[i].revents;
206 
207                 if (what & POLLHUP)
208                         what |= POLLIN | POLLOUT;
209                 else if (what & POLLERR)
210                         what |= POLLIN | POLLOUT;
211 
212 		if (what & POLLIN)
213 			which |= EV_READ;
214 		if (what & POLLOUT)
215 			which |= EV_WRITE;
216 
217 		if (!which)
218 			continue;
219 
220 		/* XXX(niels): not sure if this works for devpoll */
221 		evmap_io_active(base, events[i].fd, which);
222 	}
223 
224 	return (0);
225 }
226 
227 
228 static int
229 devpoll_add(struct event_base *base, int fd, short old, short events, void *p)
230 {
231 	struct devpollop *devpollop = base->evbase;
232 	int res;
233 	(void)p;
234 
235 	/*
236 	 * It's not necessary to OR the existing read/write events that we
237 	 * are currently interested in with the new event we are adding.
238 	 * The /dev/poll driver ORs any new events with the existing events
239 	 * that it has cached for the fd.
240 	 */
241 
242 	res = 0;
243 	if (events & EV_READ)
244 		res |= POLLIN;
245 	if (events & EV_WRITE)
246 		res |= POLLOUT;
247 
248 	if (devpoll_queue(devpollop, fd, res) != 0)
249 		return(-1);
250 
251 	return (0);
252 }
253 
254 static int
255 devpoll_del(struct event_base *base, int fd, short old, short events, void *p)
256 {
257 	struct devpollop *devpollop = base->evbase;
258 	int res;
259 	(void)p;
260 
261 	res = 0;
262 	if (events & EV_READ)
263 		res |= POLLIN;
264 	if (events & EV_WRITE)
265 		res |= POLLOUT;
266 
267 	/*
268 	 * The only way to remove an fd from the /dev/poll monitored set is
269 	 * to use POLLREMOVE by itself.  This removes ALL events for the fd
270 	 * provided so if we care about two events and are only removing one
271 	 * we must re-add the other event after POLLREMOVE.
272 	 */
273 
274 	if (devpoll_queue(devpollop, fd, POLLREMOVE) != 0)
275 		return(-1);
276 
277 	if ((res & (POLLIN|POLLOUT)) != (POLLIN|POLLOUT)) {
278 		/*
279 		 * We're not deleting all events, so we must resubmit the
280 		 * event that we are still interested in if one exists.
281 		 */
282 
283 		if ((res & POLLIN) && (old & EV_WRITE)) {
284 			/* Deleting read, still care about write */
285 			devpoll_queue(devpollop, fd, POLLOUT);
286 		} else if ((res & POLLOUT) && (old & EV_READ)) {
287 			/* Deleting write, still care about read */
288 			devpoll_queue(devpollop, fd, POLLIN);
289 		}
290 	}
291 
292 	return (0);
293 }
294 
295 static void
296 devpoll_dealloc(struct event_base *base)
297 {
298 	struct devpollop *devpollop = base->evbase;
299 
300 	evsig_dealloc(base);
301 	if (devpollop->events)
302 		mm_free(devpollop->events);
303 	if (devpollop->changes)
304 		mm_free(devpollop->changes);
305 	if (devpollop->dpfd >= 0)
306 		close(devpollop->dpfd);
307 
308 	memset(devpollop, 0, sizeof(struct devpollop));
309 	mm_free(devpollop);
310 }
311