1 /*        $NetBSD: devpoll.c,v 1.1.1.2 2017/01/31 21:14:52 christos Exp $       */
2 /*
3  * Copyright 2000-2009 Niels Provos <provos@citi.umich.edu>
4  * Copyright 2009-2012 Niels Provos and Nick Mathewson
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 3. The name of the author may not be used to endorse or promote products
15  *    derived from this software without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 #include "event2/event-config.h"
29 #include <sys/cdefs.h>
30 __RCSID("$NetBSD: devpoll.c,v 1.1.1.2 2017/01/31 21:14:52 christos Exp $");
31 #include "evconfig-private.h"
32 
33 #ifdef EVENT__HAVE_DEVPOLL
34 
35 #include <sys/types.h>
36 #include <sys/resource.h>
37 #ifdef EVENT__HAVE_SYS_TIME_H
38 #include <sys/time.h>
39 #endif
40 #include <sys/queue.h>
41 #include <sys/devpoll.h>
42 #include <signal.h>
43 #include <stdio.h>
44 #include <stdlib.h>
45 #include <string.h>
46 #include <unistd.h>
47 #include <fcntl.h>
48 #include <errno.h>
49 
50 #include "event2/event.h"
51 #include "event2/event_struct.h"
52 #include "event2/thread.h"
53 #include "event-internal.h"
54 #include "evsignal-internal.h"
55 #include "log-internal.h"
56 #include "evmap-internal.h"
57 #include "evthread-internal.h"
58 
59 struct devpollop {
60           struct pollfd *events;
61           int nevents;
62           int dpfd;
63           struct pollfd *changes;
64           int nchanges;
65 };
66 
67 static void *devpoll_init(struct event_base *);
68 static int devpoll_add(struct event_base *, int fd, short old, short events, void *);
69 static int devpoll_del(struct event_base *, int fd, short old, short events, void *);
70 static int devpoll_dispatch(struct event_base *, struct timeval *);
71 static void devpoll_dealloc(struct event_base *);
72 
73 const struct eventop devpollops = {
74           "devpoll",
75           devpoll_init,
76           devpoll_add,
77           devpoll_del,
78           devpoll_dispatch,
79           devpoll_dealloc,
80           1, /* need reinit */
81           EV_FEATURE_FDS|EV_FEATURE_O1,
82           0
83 };
84 
85 #define NEVENT      32000
86 
87 static int
devpoll_commit(struct devpollop * devpollop)88 devpoll_commit(struct devpollop *devpollop)
89 {
90           /*
91            * Due to a bug in Solaris, we have to use pwrite with an offset of 0.
92            * Write is limited to 2GB of data, until it will fail.
93            */
94           if (pwrite(devpollop->dpfd, devpollop->changes,
95                     sizeof(struct pollfd) * devpollop->nchanges, 0) == -1)
96                     return (-1);
97 
98           devpollop->nchanges = 0;
99           return (0);
100 }
101 
102 static int
devpoll_queue(struct devpollop * devpollop,int fd,int events)103 devpoll_queue(struct devpollop *devpollop, int fd, int events) {
104           struct pollfd *pfd;
105 
106           if (devpollop->nchanges >= devpollop->nevents) {
107                     /*
108                      * Change buffer is full, must commit it to /dev/poll before
109                      * adding more
110                      */
111                     if (devpoll_commit(devpollop) != 0)
112                               return (-1);
113           }
114 
115           pfd = &devpollop->changes[devpollop->nchanges++];
116           pfd->fd = fd;
117           pfd->events = events;
118           pfd->revents = 0;
119 
120           return (0);
121 }
122 
123 static void *
devpoll_init(struct event_base * base)124 devpoll_init(struct event_base *base)
125 {
126           int dpfd, nfiles = NEVENT;
127           struct rlimit rl;
128           struct devpollop *devpollop;
129 
130           if (!(devpollop = mm_calloc(1, sizeof(struct devpollop))))
131                     return (NULL);
132 
133           if (getrlimit(RLIMIT_NOFILE, &rl) == 0 &&
134               rl.rlim_cur != RLIM_INFINITY)
135                     nfiles = rl.rlim_cur;
136 
137           /* Initialize the kernel queue */
138           if ((dpfd = evutil_open_closeonexec_("/dev/poll", O_RDWR, 0)) == -1) {
139                     event_warn("open: /dev/poll");
140                     mm_free(devpollop);
141                     return (NULL);
142           }
143 
144           devpollop->dpfd = dpfd;
145 
146           /* Initialize fields */
147           /* FIXME: allocating 'nfiles' worth of space here can be
148            * expensive and unnecessary.  See how epoll.c does it instead. */
149           devpollop->events = mm_calloc(nfiles, sizeof(struct pollfd));
150           if (devpollop->events == NULL) {
151                     mm_free(devpollop);
152                     close(dpfd);
153                     return (NULL);
154           }
155           devpollop->nevents = nfiles;
156 
157           devpollop->changes = mm_calloc(nfiles, sizeof(struct pollfd));
158           if (devpollop->changes == NULL) {
159                     mm_free(devpollop->events);
160                     mm_free(devpollop);
161                     close(dpfd);
162                     return (NULL);
163           }
164 
165           evsig_init_(base);
166 
167           return (devpollop);
168 }
169 
170 static int
devpoll_dispatch(struct event_base * base,struct timeval * tv)171 devpoll_dispatch(struct event_base *base, struct timeval *tv)
172 {
173           struct devpollop *devpollop = base->evbase;
174           struct pollfd *events = devpollop->events;
175           struct dvpoll dvp;
176           int i, res, timeout = -1;
177 
178           if (devpollop->nchanges)
179                     devpoll_commit(devpollop);
180 
181           if (tv != NULL)
182                     timeout = tv->tv_sec * 1000 + (tv->tv_usec + 999) / 1000;
183 
184           dvp.dp_fds = devpollop->events;
185           dvp.dp_nfds = devpollop->nevents;
186           dvp.dp_timeout = timeout;
187 
188           EVBASE_RELEASE_LOCK(base, th_base_lock);
189 
190           res = ioctl(devpollop->dpfd, DP_POLL, &dvp);
191 
192           EVBASE_ACQUIRE_LOCK(base, th_base_lock);
193 
194           if (res == -1) {
195                     if (errno != EINTR) {
196                               event_warn("ioctl: DP_POLL");
197                               return (-1);
198                     }
199 
200                     return (0);
201           }
202 
203           event_debug(("%s: devpoll_wait reports %d", __func__, res));
204 
205           for (i = 0; i < res; i++) {
206                     int which = 0;
207                     int what = events[i].revents;
208 
209                     if (what & POLLHUP)
210                               what |= POLLIN | POLLOUT;
211                     else if (what & POLLERR)
212                               what |= POLLIN | POLLOUT;
213 
214                     if (what & POLLIN)
215                               which |= EV_READ;
216                     if (what & POLLOUT)
217                               which |= EV_WRITE;
218 
219                     if (!which)
220                               continue;
221 
222                     /* XXX(niels): not sure if this works for devpoll */
223                     evmap_io_active_(base, events[i].fd, which);
224           }
225 
226           return (0);
227 }
228 
229 
230 static int
devpoll_add(struct event_base * base,int fd,short old,short events,void * p)231 devpoll_add(struct event_base *base, int fd, short old, short events, void *p)
232 {
233           struct devpollop *devpollop = base->evbase;
234           int res;
235           (void)p;
236 
237           /*
238            * It's not necessary to OR the existing read/write events that we
239            * are currently interested in with the new event we are adding.
240            * The /dev/poll driver ORs any new events with the existing events
241            * that it has cached for the fd.
242            */
243 
244           res = 0;
245           if (events & EV_READ)
246                     res |= POLLIN;
247           if (events & EV_WRITE)
248                     res |= POLLOUT;
249 
250           if (devpoll_queue(devpollop, fd, res) != 0)
251                     return (-1);
252 
253           return (0);
254 }
255 
256 static int
devpoll_del(struct event_base * base,int fd,short old,short events,void * p)257 devpoll_del(struct event_base *base, int fd, short old, short events, void *p)
258 {
259           struct devpollop *devpollop = base->evbase;
260           int res;
261           (void)p;
262 
263           res = 0;
264           if (events & EV_READ)
265                     res |= POLLIN;
266           if (events & EV_WRITE)
267                     res |= POLLOUT;
268 
269           /*
270            * The only way to remove an fd from the /dev/poll monitored set is
271            * to use POLLREMOVE by itself.  This removes ALL events for the fd
272            * provided so if we care about two events and are only removing one
273            * we must re-add the other event after POLLREMOVE.
274            */
275 
276           if (devpoll_queue(devpollop, fd, POLLREMOVE) != 0)
277                     return (-1);
278 
279           if ((res & (POLLIN|POLLOUT)) != (POLLIN|POLLOUT)) {
280                     /*
281                      * We're not deleting all events, so we must resubmit the
282                      * event that we are still interested in if one exists.
283                      */
284 
285                     if ((res & POLLIN) && (old & EV_WRITE)) {
286                               /* Deleting read, still care about write */
287                               devpoll_queue(devpollop, fd, POLLOUT);
288                     } else if ((res & POLLOUT) && (old & EV_READ)) {
289                               /* Deleting write, still care about read */
290                               devpoll_queue(devpollop, fd, POLLIN);
291                     }
292           }
293 
294           return (0);
295 }
296 
297 static void
devpoll_dealloc(struct event_base * base)298 devpoll_dealloc(struct event_base *base)
299 {
300           struct devpollop *devpollop = base->evbase;
301 
302           evsig_dealloc_(base);
303           if (devpollop->events)
304                     mm_free(devpollop->events);
305           if (devpollop->changes)
306                     mm_free(devpollop->changes);
307           if (devpollop->dpfd >= 0)
308                     close(devpollop->dpfd);
309 
310           memset(devpollop, 0, sizeof(struct devpollop));
311           mm_free(devpollop);
312 }
313 
314 #endif /* EVENT__HAVE_DEVPOLL */
315