xref: /netbsd-src/external/bsd/libevent/dist/kqueue.c (revision bdc22b2e01993381dcefeff2bc9b56ca75a4235c)
1 /*	$NetBSD: kqueue.c,v 1.1.1.3 2017/01/31 21:14:52 christos Exp $	*/
2 /*	$OpenBSD: kqueue.c,v 1.5 2002/07/10 14:41:31 art Exp $	*/
3 
4 /*
5  * Copyright 2000-2007 Niels Provos <provos@citi.umich.edu>
6  * Copyright 2007-2012 Niels Provos and Nick Mathewson
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. The name of the author may not be used to endorse or promote products
17  *    derived from this software without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 #include "event2/event-config.h"
31 #include <sys/cdefs.h>
32 __RCSID("$NetBSD: kqueue.c,v 1.1.1.3 2017/01/31 21:14:52 christos Exp $");
33 #include "evconfig-private.h"
34 
35 #ifdef EVENT__HAVE_KQUEUE
36 
37 #include <sys/types.h>
38 #ifdef EVENT__HAVE_SYS_TIME_H
39 #include <sys/time.h>
40 #endif
41 #include <sys/queue.h>
42 #include <sys/event.h>
43 #include <signal.h>
44 #include <stdio.h>
45 #include <stdlib.h>
46 #include <string.h>
47 #include <unistd.h>
48 #include <errno.h>
49 #ifdef EVENT__HAVE_INTTYPES_H
50 #include <inttypes.h>
51 #endif
52 
53 /* Some platforms apparently define the udata field of struct kevent as
54  * intptr_t, whereas others define it as void*.  There doesn't seem to be an
55  * easy way to tell them apart via autoconf, so we need to use OS macros. */
56 #if defined(EVENT__HAVE_INTTYPES_H) && !defined(__OpenBSD__) && !defined(__FreeBSD__) && !defined(__darwin__) && !defined(__APPLE__) && !defined(__CloudABI__)
57 #define PTR_TO_UDATA(x)	((intptr_t)(x))
58 #define INT_TO_UDATA(x) ((intptr_t)(x))
59 #else
60 #define PTR_TO_UDATA(x)	(x)
61 #define INT_TO_UDATA(x) ((void*)(x))
62 #endif
63 
64 #include "event-internal.h"
65 #include "log-internal.h"
66 #include "evmap-internal.h"
67 #include "event2/thread.h"
68 #include "evthread-internal.h"
69 #include "changelist-internal.h"
70 
71 #include "kqueue-internal.h"
72 
73 #define NEVENT		64
74 
75 struct kqop {
76 	struct kevent *changes;
77 	int changes_size;
78 
79 	struct kevent *events;
80 	int events_size;
81 	int kq;
82 	int notify_event_added;
83 	pid_t pid;
84 };
85 
86 static void kqop_free(struct kqop *kqop);
87 
88 static void *kq_init(struct event_base *);
89 static int kq_sig_add(struct event_base *, int, short, short, void *);
90 static int kq_sig_del(struct event_base *, int, short, short, void *);
91 static int kq_dispatch(struct event_base *, struct timeval *);
92 static void kq_dealloc(struct event_base *);
93 
94 const struct eventop kqops = {
95 	"kqueue",
96 	kq_init,
97 	event_changelist_add_,
98 	event_changelist_del_,
99 	kq_dispatch,
100 	kq_dealloc,
101 	1 /* need reinit */,
102     EV_FEATURE_ET|EV_FEATURE_O1|EV_FEATURE_FDS,
103 	EVENT_CHANGELIST_FDINFO_SIZE
104 };
105 
106 static const struct eventop kqsigops = {
107 	"kqueue_signal",
108 	NULL,
109 	kq_sig_add,
110 	kq_sig_del,
111 	NULL,
112 	NULL,
113 	1 /* need reinit */,
114 	0,
115 	0
116 };
117 
118 static void *
119 kq_init(struct event_base *base)
120 {
121 	int kq = -1;
122 	struct kqop *kqueueop = NULL;
123 
124 	if (!(kqueueop = mm_calloc(1, sizeof(struct kqop))))
125 		return (NULL);
126 
127 /* Initialize the kernel queue */
128 
129 	if ((kq = kqueue()) == -1) {
130 		event_warn("kqueue");
131 		goto err;
132 	}
133 
134 	kqueueop->kq = kq;
135 
136 	kqueueop->pid = getpid();
137 
138 	/* Initialize fields */
139 	kqueueop->changes = mm_calloc(NEVENT, sizeof(struct kevent));
140 	if (kqueueop->changes == NULL)
141 		goto err;
142 	kqueueop->events = mm_calloc(NEVENT, sizeof(struct kevent));
143 	if (kqueueop->events == NULL)
144 		goto err;
145 	kqueueop->events_size = kqueueop->changes_size = NEVENT;
146 
147 	/* Check for Mac OS X kqueue bug. */
148 	memset(&kqueueop->changes[0], 0, sizeof kqueueop->changes[0]);
149 	kqueueop->changes[0].ident = -1;
150 	kqueueop->changes[0].filter = EVFILT_READ;
151 	kqueueop->changes[0].flags = EV_ADD;
152 	/*
153 	 * If kqueue works, then kevent will succeed, and it will
154 	 * stick an error in events[0].  If kqueue is broken, then
155 	 * kevent will fail.
156 	 */
157 	if (kevent(kq,
158 		kqueueop->changes, 1, kqueueop->events, NEVENT, NULL) != 1 ||
159 	    (int)kqueueop->events[0].ident != -1 ||
160 	    !(kqueueop->events[0].flags & EV_ERROR)) {
161 		event_warn("%s: detected broken kqueue; not using.", __func__);
162 		goto err;
163 	}
164 
165 	base->evsigsel = &kqsigops;
166 
167 	return (kqueueop);
168 err:
169 	if (kqueueop)
170 		kqop_free(kqueueop);
171 
172 	return (NULL);
173 }
174 
175 #define ADD_UDATA 0x30303
176 
177 static void
178 kq_setup_kevent(struct kevent *out, evutil_socket_t fd, int filter, short change)
179 {
180 	memset(out, 0, sizeof(struct kevent));
181 	out->ident = fd;
182 	out->filter = filter;
183 
184 	if (change & EV_CHANGE_ADD) {
185 		out->flags = EV_ADD;
186 		/* We set a magic number here so that we can tell 'add'
187 		 * errors from 'del' errors. */
188 		out->udata = INT_TO_UDATA(ADD_UDATA);
189 		if (change & EV_ET)
190 			out->flags |= EV_CLEAR;
191 #ifdef NOTE_EOF
192 		/* Make it behave like select() and poll() */
193 		if (filter == EVFILT_READ)
194 			out->fflags = NOTE_EOF;
195 #endif
196 	} else {
197 		EVUTIL_ASSERT(change & EV_CHANGE_DEL);
198 		out->flags = EV_DELETE;
199 	}
200 }
201 
202 static int
203 kq_build_changes_list(const struct event_changelist *changelist,
204     struct kqop *kqop)
205 {
206 	int i;
207 	int n_changes = 0;
208 
209 	for (i = 0; i < changelist->n_changes; ++i) {
210 		struct event_change *in_ch = &changelist->changes[i];
211 		struct kevent *out_ch;
212 		if (n_changes >= kqop->changes_size - 1) {
213 			int newsize = kqop->changes_size * 2;
214 			struct kevent *newchanges;
215 
216 			newchanges = mm_realloc(kqop->changes,
217 			    newsize * sizeof(struct kevent));
218 			if (newchanges == NULL) {
219 				event_warn("%s: realloc", __func__);
220 				return (-1);
221 			}
222 			kqop->changes = newchanges;
223 			kqop->changes_size = newsize;
224 		}
225 		if (in_ch->read_change) {
226 			out_ch = &kqop->changes[n_changes++];
227 			kq_setup_kevent(out_ch, in_ch->fd, EVFILT_READ,
228 			    in_ch->read_change);
229 		}
230 		if (in_ch->write_change) {
231 			out_ch = &kqop->changes[n_changes++];
232 			kq_setup_kevent(out_ch, in_ch->fd, EVFILT_WRITE,
233 			    in_ch->write_change);
234 		}
235 	}
236 	return n_changes;
237 }
238 
239 static int
240 kq_grow_events(struct kqop *kqop, size_t new_size)
241 {
242 	struct kevent *newresult;
243 
244 	newresult = mm_realloc(kqop->events,
245 	    new_size * sizeof(struct kevent));
246 
247 	if (newresult) {
248 		kqop->events = newresult;
249 		kqop->events_size = new_size;
250 		return 0;
251 	} else {
252 		return -1;
253 	}
254 }
255 
256 static int
257 kq_dispatch(struct event_base *base, struct timeval *tv)
258 {
259 	struct kqop *kqop = base->evbase;
260 	struct kevent *events = kqop->events;
261 	struct kevent *changes;
262 	struct timespec ts, *ts_p = NULL;
263 	int i, n_changes, res;
264 
265 	if (tv != NULL) {
266 		ts.tv_sec = tv->tv_sec;
267 		ts.tv_nsec = tv->tv_usec * 1000;
268 		ts_p = &ts;
269 	}
270 
271 	/* Build "changes" from "base->changes" */
272 	EVUTIL_ASSERT(kqop->changes);
273 	n_changes = kq_build_changes_list(&base->changelist, kqop);
274 	if (n_changes < 0)
275 		return -1;
276 
277 	event_changelist_remove_all_(&base->changelist, base);
278 
279 	/* steal the changes array in case some broken code tries to call
280 	 * dispatch twice at once. */
281 	changes = kqop->changes;
282 	kqop->changes = NULL;
283 
284 	/* Make sure that 'events' is at least as long as the list of changes:
285 	 * otherwise errors in the changes can get reported as a -1 return
286 	 * value from kevent() rather than as EV_ERROR events in the events
287 	 * array.
288 	 *
289 	 * (We could instead handle -1 return values from kevent() by
290 	 * retrying with a smaller changes array or a larger events array,
291 	 * but this approach seems less risky for now.)
292 	 */
293 	if (kqop->events_size < n_changes) {
294 		int new_size = kqop->events_size;
295 		do {
296 			new_size *= 2;
297 		} while (new_size < n_changes);
298 
299 		kq_grow_events(kqop, new_size);
300 		events = kqop->events;
301 	}
302 
303 	EVBASE_RELEASE_LOCK(base, th_base_lock);
304 
305 	res = kevent(kqop->kq, changes, n_changes,
306 	    events, kqop->events_size, ts_p);
307 
308 	EVBASE_ACQUIRE_LOCK(base, th_base_lock);
309 
310 	EVUTIL_ASSERT(kqop->changes == NULL);
311 	kqop->changes = changes;
312 
313 	if (res == -1) {
314 		if (errno != EINTR) {
315 			event_warn("kevent");
316 			return (-1);
317 		}
318 
319 		return (0);
320 	}
321 
322 	event_debug(("%s: kevent reports %d", __func__, res));
323 
324 	for (i = 0; i < res; i++) {
325 		int which = 0;
326 
327 		if (events[i].flags & EV_ERROR) {
328 			switch (events[i].data) {
329 
330 			/* Can occur on delete if we are not currently
331 			 * watching any events on this fd.  That can
332 			 * happen when the fd was closed and another
333 			 * file was opened with that fd. */
334 			case ENOENT:
335 			/* Can occur for reasons not fully understood
336 			 * on FreeBSD. */
337 			case EINVAL:
338 				continue;
339 #if defined(__FreeBSD__)
340 			/*
341 			 * This currently occurs if an FD is closed
342 			 * before the EV_DELETE makes it out via kevent().
343 			 * The FreeBSD capabilities code sees the blank
344 			 * capability set and rejects the request to
345 			 * modify an event.
346 			 *
347 			 * To be strictly correct - when an FD is closed,
348 			 * all the registered events are also removed.
349 			 * Queuing EV_DELETE to a closed FD is wrong.
350 			 * The event(s) should just be deleted from
351 			 * the pending changelist.
352 			 */
353 			case ENOTCAPABLE:
354 				continue;
355 #endif
356 
357 			/* Can occur on a delete if the fd is closed. */
358 			case EBADF:
359 				/* XXXX On NetBSD, we can also get EBADF if we
360 				 * try to add the write side of a pipe, but
361 				 * the read side has already been closed.
362 				 * Other BSDs call this situation 'EPIPE'. It
363 				 * would be good if we had a way to report
364 				 * this situation. */
365 				continue;
366 			/* These two can occur on an add if the fd was one side
367 			 * of a pipe, and the other side was closed. */
368 			case EPERM:
369 			case EPIPE:
370 				/* Report read events, if we're listening for
371 				 * them, so that the user can learn about any
372 				 * add errors.  (If the operation was a
373 				 * delete, then udata should be cleared.) */
374 				if (events[i].udata) {
375 					/* The operation was an add:
376 					 * report the error as a read. */
377 					which |= EV_READ;
378 					break;
379 				} else {
380 					/* The operation was a del:
381 					 * report nothing. */
382 					continue;
383 				}
384 
385 			/* Other errors shouldn't occur. */
386 			default:
387 				errno = events[i].data;
388 				return (-1);
389 			}
390 		} else if (events[i].filter == EVFILT_READ) {
391 			which |= EV_READ;
392 		} else if (events[i].filter == EVFILT_WRITE) {
393 			which |= EV_WRITE;
394 		} else if (events[i].filter == EVFILT_SIGNAL) {
395 			which |= EV_SIGNAL;
396 #ifdef EVFILT_USER
397 		} else if (events[i].filter == EVFILT_USER) {
398 			base->is_notify_pending = 0;
399 #endif
400 		}
401 
402 		if (!which)
403 			continue;
404 
405 		if (events[i].filter == EVFILT_SIGNAL) {
406 			evmap_signal_active_(base, events[i].ident, 1);
407 		} else {
408 			evmap_io_active_(base, events[i].ident, which | EV_ET);
409 		}
410 	}
411 
412 	if (res == kqop->events_size) {
413 		/* We used all the events space that we have. Maybe we should
414 		   make it bigger. */
415 		kq_grow_events(kqop, kqop->events_size * 2);
416 	}
417 
418 	return (0);
419 }
420 
421 static void
422 kqop_free(struct kqop *kqop)
423 {
424 	if (kqop->changes)
425 		mm_free(kqop->changes);
426 	if (kqop->events)
427 		mm_free(kqop->events);
428 	if (kqop->kq >= 0 && kqop->pid == getpid())
429 		close(kqop->kq);
430 	memset(kqop, 0, sizeof(struct kqop));
431 	mm_free(kqop);
432 }
433 
434 static void
435 kq_dealloc(struct event_base *base)
436 {
437 	struct kqop *kqop = base->evbase;
438 	evsig_dealloc_(base);
439 	kqop_free(kqop);
440 }
441 
442 /* signal handling */
443 static int
444 kq_sig_add(struct event_base *base, int nsignal, short old, short events, void *p)
445 {
446 	struct kqop *kqop = base->evbase;
447 	struct kevent kev;
448 	struct timespec timeout = { 0, 0 };
449 	(void)p;
450 
451 	EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
452 
453 	memset(&kev, 0, sizeof(kev));
454 	kev.ident = nsignal;
455 	kev.filter = EVFILT_SIGNAL;
456 	kev.flags = EV_ADD;
457 
458 	/* Be ready for the signal if it is sent any
459 	 * time between now and the next call to
460 	 * kq_dispatch. */
461 	if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
462 		return (-1);
463 
464         /* We can set the handler for most signals to SIG_IGN and
465          * still have them reported to us in the queue.  However,
466          * if the handler for SIGCHLD is SIG_IGN, the system reaps
467          * zombie processes for us, and we don't get any notification.
468          * This appears to be the only signal with this quirk. */
469 	if (evsig_set_handler_(base, nsignal,
470                                nsignal == SIGCHLD ? SIG_DFL : SIG_IGN) == -1)
471 		return (-1);
472 
473 	return (0);
474 }
475 
476 static int
477 kq_sig_del(struct event_base *base, int nsignal, short old, short events, void *p)
478 {
479 	struct kqop *kqop = base->evbase;
480 	struct kevent kev;
481 
482 	struct timespec timeout = { 0, 0 };
483 	(void)p;
484 
485 	EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
486 
487 	memset(&kev, 0, sizeof(kev));
488 	kev.ident = nsignal;
489 	kev.filter = EVFILT_SIGNAL;
490 	kev.flags = EV_DELETE;
491 
492 	/* Because we insert signal events
493 	 * immediately, we need to delete them
494 	 * immediately, too */
495 	if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
496 		return (-1);
497 
498 	if (evsig_restore_handler_(base, nsignal) == -1)
499 		return (-1);
500 
501 	return (0);
502 }
503 
504 
505 /* OSX 10.6 and FreeBSD 8.1 add support for EVFILT_USER, which we can use
506  * to wake up the event loop from another thread. */
507 
508 /* Magic number we use for our filter ID. */
509 #define NOTIFY_IDENT 42
510 
511 int
512 event_kq_add_notify_event_(struct event_base *base)
513 {
514 	struct kqop *kqop = base->evbase;
515 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER)
516 	struct kevent kev;
517 	struct timespec timeout = { 0, 0 };
518 #endif
519 
520 	if (kqop->notify_event_added)
521 		return 0;
522 
523 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER)
524 	memset(&kev, 0, sizeof(kev));
525 	kev.ident = NOTIFY_IDENT;
526 	kev.filter = EVFILT_USER;
527 	kev.flags = EV_ADD | EV_CLEAR;
528 
529 	if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) {
530 		event_warn("kevent: adding EVFILT_USER event");
531 		return -1;
532 	}
533 
534 	kqop->notify_event_added = 1;
535 
536 	return 0;
537 #else
538 	return -1;
539 #endif
540 }
541 
542 int
543 event_kq_notify_base_(struct event_base *base)
544 {
545 	struct kqop *kqop = base->evbase;
546 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER)
547 	struct kevent kev;
548 	struct timespec timeout = { 0, 0 };
549 #endif
550 	if (! kqop->notify_event_added)
551 		return -1;
552 
553 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER)
554 	memset(&kev, 0, sizeof(kev));
555 	kev.ident = NOTIFY_IDENT;
556 	kev.filter = EVFILT_USER;
557 	kev.fflags = NOTE_TRIGGER;
558 
559 	if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) {
560 		event_warn("kevent: triggering EVFILT_USER event");
561 		return -1;
562 	}
563 
564 	return 0;
565 #else
566 	return -1;
567 #endif
568 }
569 
570 #endif /* EVENT__HAVE_KQUEUE */
571