1 /* $OpenBSD: kqueue.c,v 1.5 2002/07/10 14:41:31 art Exp $ */
4 * Copyright 2000-2007 Niels Provos <provos@citi.umich.edu>
5 * Copyright 2007-2012 Niels Provos and Nick Mathewson
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. The name of the author may not be used to endorse or promote products
16 * derived from this software without specific prior written permission.
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
27 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 #include "event2/event-config.h"
30 #include "evconfig-private.h"
32 #ifdef EVENT__HAVE_KQUEUE
34 #include <sys/types.h>
35 #ifdef EVENT__HAVE_SYS_TIME_H
38 #include <sys/queue.h>
39 #include <sys/event.h>
46 #ifdef EVENT__HAVE_INTTYPES_H
50 /* Some platforms apparently define the udata field of struct kevent as
51 * intptr_t, whereas others define it as void*. There doesn't seem to be an
52 * easy way to tell them apart via autoconf, so we need to use OS macros. */
53 #if defined(EVENT__HAVE_INTTYPES_H) && !defined(__OpenBSD__) && !defined(__FreeBSD__) && !defined(__darwin__) && !defined(__APPLE__)
54 #define PTR_TO_UDATA(x) ((intptr_t)(x))
55 #define INT_TO_UDATA(x) ((intptr_t)(x))
57 #define PTR_TO_UDATA(x) (x)
58 #define INT_TO_UDATA(x) ((void*)(x))
61 #include "event-internal.h"
62 #include "log-internal.h"
63 #include "evmap-internal.h"
64 #include "event2/thread.h"
65 #include "evthread-internal.h"
66 #include "changelist-internal.h"
68 #include "kqueue-internal.h"
73 struct kevent *changes;
76 struct kevent *events;
79 int notify_event_added;
83 static void kqop_free(struct kqop *kqop);
85 static void *kq_init(struct event_base *);
86 static int kq_sig_add(struct event_base *, int, short, short, void *);
87 static int kq_sig_del(struct event_base *, int, short, short, void *);
88 static int kq_dispatch(struct event_base *, struct timeval *);
89 static void kq_dealloc(struct event_base *);
91 const struct eventop kqops = {
94 event_changelist_add_,
95 event_changelist_del_,
99 EV_FEATURE_ET|EV_FEATURE_O1|EV_FEATURE_FDS,
100 EVENT_CHANGELIST_FDINFO_SIZE
103 static const struct eventop kqsigops = {
116 kq_init(struct event_base *base)
119 struct kqop *kqueueop = NULL;
121 if (!(kqueueop = mm_calloc(1, sizeof(struct kqop))))
124 /* Initialize the kernel queue */
126 if ((kq = kqueue()) == -1) {
127 event_warn("kqueue");
133 kqueueop->pid = getpid();
135 /* Initialize fields */
136 kqueueop->changes = mm_calloc(NEVENT, sizeof(struct kevent));
137 if (kqueueop->changes == NULL)
139 kqueueop->events = mm_calloc(NEVENT, sizeof(struct kevent));
140 if (kqueueop->events == NULL)
142 kqueueop->events_size = kqueueop->changes_size = NEVENT;
144 /* Check for Mac OS X kqueue bug. */
145 memset(&kqueueop->changes[0], 0, sizeof kqueueop->changes[0]);
146 kqueueop->changes[0].ident = -1;
147 kqueueop->changes[0].filter = EVFILT_READ;
148 kqueueop->changes[0].flags = EV_ADD;
150 * If kqueue works, then kevent will succeed, and it will
151 * stick an error in events[0]. If kqueue is broken, then
155 kqueueop->changes, 1, kqueueop->events, NEVENT, NULL) != 1 ||
156 (int)kqueueop->events[0].ident != -1 ||
157 kqueueop->events[0].flags != EV_ERROR) {
158 event_warn("%s: detected broken kqueue; not using.", __func__);
162 base->evsigsel = &kqsigops;
172 #define ADD_UDATA 0x30303
175 kq_setup_kevent(struct kevent *out, evutil_socket_t fd, int filter, short change)
177 memset(out, 0, sizeof(struct kevent));
179 out->filter = filter;
181 if (change & EV_CHANGE_ADD) {
183 /* We set a magic number here so that we can tell 'add'
184 * errors from 'del' errors. */
185 out->udata = INT_TO_UDATA(ADD_UDATA);
187 out->flags |= EV_CLEAR;
189 /* Make it behave like select() and poll() */
190 if (filter == EVFILT_READ)
191 out->fflags = NOTE_EOF;
194 EVUTIL_ASSERT(change & EV_CHANGE_DEL);
195 out->flags = EV_DELETE;
200 kq_build_changes_list(const struct event_changelist *changelist,
206 for (i = 0; i < changelist->n_changes; ++i) {
207 struct event_change *in_ch = &changelist->changes[i];
208 struct kevent *out_ch;
209 if (n_changes >= kqop->changes_size - 1) {
210 int newsize = kqop->changes_size * 2;
211 struct kevent *newchanges;
213 newchanges = mm_realloc(kqop->changes,
214 newsize * sizeof(struct kevent));
215 if (newchanges == NULL) {
216 event_warn("%s: realloc", __func__);
219 kqop->changes = newchanges;
220 kqop->changes_size = newsize;
222 if (in_ch->read_change) {
223 out_ch = &kqop->changes[n_changes++];
224 kq_setup_kevent(out_ch, in_ch->fd, EVFILT_READ,
227 if (in_ch->write_change) {
228 out_ch = &kqop->changes[n_changes++];
229 kq_setup_kevent(out_ch, in_ch->fd, EVFILT_WRITE,
230 in_ch->write_change);
237 kq_grow_events(struct kqop *kqop, size_t new_size)
239 struct kevent *newresult;
241 newresult = mm_realloc(kqop->events,
242 new_size * sizeof(struct kevent));
245 kqop->events = newresult;
246 kqop->events_size = new_size;
254 kq_dispatch(struct event_base *base, struct timeval *tv)
256 struct kqop *kqop = base->evbase;
257 struct kevent *events = kqop->events;
258 struct kevent *changes;
259 struct timespec ts, *ts_p = NULL;
260 int i, n_changes, res;
263 TIMEVAL_TO_TIMESPEC(tv, &ts);
267 /* Build "changes" from "base->changes" */
268 EVUTIL_ASSERT(kqop->changes);
269 n_changes = kq_build_changes_list(&base->changelist, kqop);
273 event_changelist_remove_all_(&base->changelist, base);
275 /* steal the changes array in case some broken code tries to call
276 * dispatch twice at once. */
277 changes = kqop->changes;
278 kqop->changes = NULL;
280 /* Make sure that 'events' is at least as long as the list of changes:
281 * otherwise errors in the changes can get reported as a -1 return
282 * value from kevent() rather than as EV_ERROR events in the events
285 * (We could instead handle -1 return values from kevent() by
286 * retrying with a smaller changes array or a larger events array,
287 * but this approach seems less risky for now.)
289 if (kqop->events_size < n_changes) {
290 int new_size = kqop->events_size;
293 } while (new_size < n_changes);
295 kq_grow_events(kqop, new_size);
296 events = kqop->events;
299 EVBASE_RELEASE_LOCK(base, th_base_lock);
301 res = kevent(kqop->kq, changes, n_changes,
302 events, kqop->events_size, ts_p);
304 EVBASE_ACQUIRE_LOCK(base, th_base_lock);
306 EVUTIL_ASSERT(kqop->changes == NULL);
307 kqop->changes = changes;
310 if (errno != EINTR) {
311 event_warn("kevent");
318 event_debug(("%s: kevent reports %d", __func__, res));
320 for (i = 0; i < res; i++) {
323 if (events[i].flags & EV_ERROR) {
324 switch (events[i].data) {
326 /* Can occur on delete if we are not currently
327 * watching any events on this fd. That can
328 * happen when the fd was closed and another
329 * file was opened with that fd. */
331 /* Can occur for reasons not fully understood
335 #if defined(__FreeBSD__) && defined(ENOTCAPABLE)
337 * This currently occurs if an FD is closed
338 * before the EV_DELETE makes it out via kevent().
339 * The FreeBSD capabilities code sees the blank
340 * capability set and rejects the request to
343 * To be strictly correct - when an FD is closed,
344 * all the registered events are also removed.
345 * Queuing EV_DELETE to a closed FD is wrong.
346 * The event(s) should just be deleted from
347 * the pending changelist.
353 /* Can occur on a delete if the fd is closed. */
355 /* XXXX On NetBSD, we can also get EBADF if we
356 * try to add the write side of a pipe, but
357 * the read side has already been closed.
358 * Other BSDs call this situation 'EPIPE'. It
359 * would be good if we had a way to report
362 /* These two can occur on an add if the fd was one side
363 * of a pipe, and the other side was closed. */
366 /* Report read events, if we're listening for
367 * them, so that the user can learn about any
368 * add errors. (If the operation was a
369 * delete, then udata should be cleared.) */
370 if (events[i].udata) {
371 /* The operation was an add:
372 * report the error as a read. */
376 /* The operation was a del:
381 /* Other errors shouldn't occur. */
383 errno = events[i].data;
386 } else if (events[i].filter == EVFILT_READ) {
388 } else if (events[i].filter == EVFILT_WRITE) {
390 } else if (events[i].filter == EVFILT_SIGNAL) {
393 } else if (events[i].filter == EVFILT_USER) {
394 base->is_notify_pending = 0;
401 if (events[i].filter == EVFILT_SIGNAL) {
402 evmap_signal_active_(base, events[i].ident, 1);
404 evmap_io_active_(base, events[i].ident, which | EV_ET);
408 if (res == kqop->events_size) {
409 /* We used all the events space that we have. Maybe we should
411 kq_grow_events(kqop, kqop->events_size * 2);
418 kqop_free(struct kqop *kqop)
421 mm_free(kqop->changes);
423 mm_free(kqop->events);
424 if (kqop->kq >= 0 && kqop->pid == getpid())
426 memset(kqop, 0, sizeof(struct kqop));
431 kq_dealloc(struct event_base *base)
433 struct kqop *kqop = base->evbase;
434 evsig_dealloc_(base);
438 /* signal handling */
440 kq_sig_add(struct event_base *base, int nsignal, short old, short events, void *p)
442 struct kqop *kqop = base->evbase;
444 struct timespec timeout = { 0, 0 };
447 EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
449 memset(&kev, 0, sizeof(kev));
451 kev.filter = EVFILT_SIGNAL;
454 /* Be ready for the signal if it is sent any
455 * time between now and the next call to
457 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
460 /* We can set the handler for most signals to SIG_IGN and
461 * still have them reported to us in the queue. However,
462 * if the handler for SIGCHLD is SIG_IGN, the system reaps
463 * zombie processes for us, and we don't get any notification.
464 * This appears to be the only signal with this quirk. */
465 if (evsig_set_handler_(base, nsignal,
466 nsignal == SIGCHLD ? SIG_DFL : SIG_IGN) == -1)
473 kq_sig_del(struct event_base *base, int nsignal, short old, short events, void *p)
475 struct kqop *kqop = base->evbase;
478 struct timespec timeout = { 0, 0 };
481 EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
483 memset(&kev, 0, sizeof(kev));
485 kev.filter = EVFILT_SIGNAL;
486 kev.flags = EV_DELETE;
488 /* Because we insert signal events
489 * immediately, we need to delete them
490 * immediately, too */
491 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
494 if (evsig_restore_handler_(base, nsignal) == -1)
501 /* OSX 10.6 and FreeBSD 8.1 add support for EVFILT_USER, which we can use
502 * to wake up the event loop from another thread. */
504 /* Magic number we use for our filter ID. */
505 #define NOTIFY_IDENT 42
508 event_kq_add_notify_event_(struct event_base *base)
510 struct kqop *kqop = base->evbase;
511 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER)
513 struct timespec timeout = { 0, 0 };
516 if (kqop->notify_event_added)
519 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER)
520 memset(&kev, 0, sizeof(kev));
521 kev.ident = NOTIFY_IDENT;
522 kev.filter = EVFILT_USER;
523 kev.flags = EV_ADD | EV_CLEAR;
525 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) {
526 event_warn("kevent: adding EVFILT_USER event");
530 kqop->notify_event_added = 1;
539 event_kq_notify_base_(struct event_base *base)
541 struct kqop *kqop = base->evbase;
542 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER)
544 struct timespec timeout = { 0, 0 };
546 if (! kqop->notify_event_added)
549 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER)
550 memset(&kev, 0, sizeof(kev));
551 kev.ident = NOTIFY_IDENT;
552 kev.filter = EVFILT_USER;
553 kev.fflags = NOTE_TRIGGER;
555 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) {
556 event_warn("kevent: triggering EVFILT_USER event");
566 #endif /* EVENT__HAVE_KQUEUE */