1/*	$OpenBSD: kqueue.c,v 1.5 2002/07/10 14:41:31 art Exp $	*/
2
3/*
4 * Copyright 2000-2007 Niels Provos <provos@citi.umich.edu>
5 * Copyright 2007-2012 Niels Provos and Nick Mathewson
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
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.
17 *
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.
28 */
29#include "event2/event-config.h"
30
31#define _GNU_SOURCE
32
33#include <sys/types.h>
34#ifdef _EVENT_HAVE_SYS_TIME_H
35#include <sys/time.h>
36#endif
37#include <sys/queue.h>
38#include <sys/event.h>
39#include <signal.h>
40#include <stdio.h>
41#include <stdlib.h>
42#include <string.h>
43#include <unistd.h>
44#include <errno.h>
45#ifdef _EVENT_HAVE_INTTYPES_H
46#include <inttypes.h>
47#endif
48
49/* Some platforms apparently define the udata field of struct kevent as
50 * intptr_t, whereas others define it as void*.  There doesn't seem to be an
51 * easy way to tell them apart via autoconf, so we need to use OS macros. */
52#if defined(_EVENT_HAVE_INTTYPES_H) && !defined(__OpenBSD__) && !defined(__FreeBSD__) && !defined(__darwin__) && !defined(__APPLE__)
53#define PTR_TO_UDATA(x)	((intptr_t)(x))
54#define INT_TO_UDATA(x) ((intptr_t)(x))
55#else
56#define PTR_TO_UDATA(x)	(x)
57#define INT_TO_UDATA(x) ((void*)(x))
58#endif
59
60#include "event-internal.h"
61#include "log-internal.h"
62#include "evmap-internal.h"
63#include "event2/thread.h"
64#include "evthread-internal.h"
65#include "changelist-internal.h"
66
67#define NEVENT		64
68
69struct kqop {
70	struct kevent *changes;
71	int changes_size;
72
73	struct kevent *events;
74	int events_size;
75	int kq;
76	pid_t pid;
77};
78
79static void kqop_free(struct kqop *kqop);
80
81static void *kq_init(struct event_base *);
82static int kq_sig_add(struct event_base *, int, short, short, void *);
83static int kq_sig_del(struct event_base *, int, short, short, void *);
84static int kq_dispatch(struct event_base *, struct timeval *);
85static void kq_dealloc(struct event_base *);
86
87const struct eventop kqops = {
88	"kqueue",
89	kq_init,
90	event_changelist_add,
91	event_changelist_del,
92	kq_dispatch,
93	kq_dealloc,
94	1 /* need reinit */,
95    EV_FEATURE_ET|EV_FEATURE_O1|EV_FEATURE_FDS,
96	EVENT_CHANGELIST_FDINFO_SIZE
97};
98
99static const struct eventop kqsigops = {
100	"kqueue_signal",
101	NULL,
102	kq_sig_add,
103	kq_sig_del,
104	NULL,
105	NULL,
106	1 /* need reinit */,
107	0,
108	0
109};
110
111static void *
112kq_init(struct event_base *base)
113{
114	int kq = -1;
115	struct kqop *kqueueop = NULL;
116
117	if (!(kqueueop = mm_calloc(1, sizeof(struct kqop))))
118		return (NULL);
119
120/* Initialize the kernel queue */
121
122	if ((kq = kqueue()) == -1) {
123		event_warn("kqueue");
124		goto err;
125	}
126
127	kqueueop->kq = kq;
128
129	kqueueop->pid = getpid();
130
131	/* Initialize fields */
132	kqueueop->changes = mm_calloc(NEVENT, sizeof(struct kevent));
133	if (kqueueop->changes == NULL)
134		goto err;
135	kqueueop->events = mm_calloc(NEVENT, sizeof(struct kevent));
136	if (kqueueop->events == NULL)
137		goto err;
138	kqueueop->events_size = kqueueop->changes_size = NEVENT;
139
140	/* Check for Mac OS X kqueue bug. */
141	memset(&kqueueop->changes[0], 0, sizeof kqueueop->changes[0]);
142	kqueueop->changes[0].ident = -1;
143	kqueueop->changes[0].filter = EVFILT_READ;
144	kqueueop->changes[0].flags = EV_ADD;
145	/*
146	 * If kqueue works, then kevent will succeed, and it will
147	 * stick an error in events[0].  If kqueue is broken, then
148	 * kevent will fail.
149	 */
150	if (kevent(kq,
151		kqueueop->changes, 1, kqueueop->events, NEVENT, NULL) != 1 ||
152	    (int)kqueueop->events[0].ident != -1 ||
153	    kqueueop->events[0].flags != EV_ERROR) {
154		event_warn("%s: detected broken kqueue; not using.", __func__);
155		goto err;
156	}
157
158	base->evsigsel = &kqsigops;
159
160	return (kqueueop);
161err:
162	if (kqueueop)
163		kqop_free(kqueueop);
164
165	return (NULL);
166}
167
168static void
169kq_sighandler(int sig)
170{
171	/* Do nothing here */
172}
173
174#define ADD_UDATA 0x30303
175
176static void
177kq_setup_kevent(struct kevent *out, evutil_socket_t fd, int filter, short change)
178{
179	memset(out, 0, sizeof(struct kevent));
180	out->ident = fd;
181	out->filter = filter;
182
183	if (change & EV_CHANGE_ADD) {
184		out->flags = EV_ADD;
185		/* We set a magic number here so that we can tell 'add'
186		 * errors from 'del' errors. */
187		out->udata = INT_TO_UDATA(ADD_UDATA);
188		if (change & EV_ET)
189			out->flags |= EV_CLEAR;
190#ifdef NOTE_EOF
191		/* Make it behave like select() and poll() */
192		if (filter == EVFILT_READ)
193			out->fflags = NOTE_EOF;
194#endif
195	} else {
196		EVUTIL_ASSERT(change & EV_CHANGE_DEL);
197		out->flags = EV_DELETE;
198	}
199}
200
201static int
202kq_build_changes_list(const struct event_changelist *changelist,
203    struct kqop *kqop)
204{
205	int i;
206	int n_changes = 0;
207
208	for (i = 0; i < changelist->n_changes; ++i) {
209		struct event_change *in_ch = &changelist->changes[i];
210		struct kevent *out_ch;
211		if (n_changes >= kqop->changes_size - 1) {
212			int newsize = kqop->changes_size * 2;
213			struct kevent *newchanges;
214
215			newchanges = mm_realloc(kqop->changes,
216			    newsize * sizeof(struct kevent));
217			if (newchanges == NULL) {
218				event_warn("%s: realloc", __func__);
219				return (-1);
220			}
221			kqop->changes = newchanges;
222			kqop->changes_size = newsize;
223		}
224		if (in_ch->read_change) {
225			out_ch = &kqop->changes[n_changes++];
226			kq_setup_kevent(out_ch, in_ch->fd, EVFILT_READ,
227			    in_ch->read_change);
228		}
229		if (in_ch->write_change) {
230			out_ch = &kqop->changes[n_changes++];
231			kq_setup_kevent(out_ch, in_ch->fd, EVFILT_WRITE,
232			    in_ch->write_change);
233		}
234	}
235	return n_changes;
236}
237
238static int
239kq_grow_events(struct kqop *kqop, size_t new_size)
240{
241	struct kevent *newresult;
242
243	newresult = mm_realloc(kqop->events,
244	    new_size * sizeof(struct kevent));
245
246	if (newresult) {
247		kqop->events = newresult;
248		kqop->events_size = new_size;
249		return 0;
250	} else {
251		return -1;
252	}
253}
254
255static int
256kq_dispatch(struct event_base *base, struct timeval *tv)
257{
258	struct kqop *kqop = base->evbase;
259	struct kevent *events = kqop->events;
260	struct kevent *changes;
261	struct timespec ts, *ts_p = NULL;
262	int i, n_changes, res;
263
264	if (tv != NULL) {
265		TIMEVAL_TO_TIMESPEC(tv, &ts);
266		ts_p = &ts;
267	}
268
269	/* Build "changes" from "base->changes" */
270	EVUTIL_ASSERT(kqop->changes);
271	n_changes = kq_build_changes_list(&base->changelist, kqop);
272	if (n_changes < 0)
273		return -1;
274
275	event_changelist_remove_all(&base->changelist, base);
276
277	/* steal the changes array in case some broken code tries to call
278	 * dispatch twice at once. */
279	changes = kqop->changes;
280	kqop->changes = NULL;
281
282	/* Make sure that 'events' is at least as long as the list of changes:
283	 * otherwise errors in the changes can get reported as a -1 return
284	 * value from kevent() rather than as EV_ERROR events in the events
285	 * array.
286	 *
287	 * (We could instead handle -1 return values from kevent() by
288	 * retrying with a smaller changes array or a larger events array,
289	 * but this approach seems less risky for now.)
290	 */
291	if (kqop->events_size < n_changes) {
292		int new_size = kqop->events_size;
293		do {
294			new_size *= 2;
295		} while (new_size < n_changes);
296
297		kq_grow_events(kqop, new_size);
298		events = kqop->events;
299	}
300
301	EVBASE_RELEASE_LOCK(base, th_base_lock);
302
303	res = kevent(kqop->kq, changes, n_changes,
304	    events, kqop->events_size, ts_p);
305
306	EVBASE_ACQUIRE_LOCK(base, th_base_lock);
307
308	EVUTIL_ASSERT(kqop->changes == NULL);
309	kqop->changes = changes;
310
311	if (res == -1) {
312		if (errno != EINTR) {
313			event_warn("kevent");
314			return (-1);
315		}
316
317		return (0);
318	}
319
320	event_debug(("%s: kevent reports %d", __func__, res));
321
322	for (i = 0; i < res; i++) {
323		int which = 0;
324
325		if (events[i].flags & EV_ERROR) {
326			switch (events[i].data) {
327
328			/* Can occur on delete if we are not currently
329			 * watching any events on this fd.  That can
330			 * happen when the fd was closed and another
331			 * file was opened with that fd. */
332			case ENOENT:
333			/* Can occur for reasons not fully understood
334			 * on FreeBSD. */
335			case EINVAL:
336				continue;
337
338			/* Can occur on a delete if the fd is closed. */
339			case EBADF:
340				/* XXXX On NetBSD, we can also get EBADF if we
341				 * try to add the write side of a pipe, but
342				 * the read side has already been closed.
343				 * Other BSDs call this situation 'EPIPE'. It
344				 * would be good if we had a way to report
345				 * this situation. */
346				continue;
347			/* These two can occur on an add if the fd was one side
348			 * of a pipe, and the other side was closed. */
349			case EPERM:
350			case EPIPE:
351				/* Report read events, if we're listening for
352				 * them, so that the user can learn about any
353				 * add errors.  (If the operation was a
354				 * delete, then udata should be cleared.) */
355				if (events[i].udata) {
356					/* The operation was an add:
357					 * report the error as a read. */
358					which |= EV_READ;
359					break;
360				} else {
361					/* The operation was a del:
362					 * report nothing. */
363					continue;
364				}
365
366			/* Other errors shouldn't occur. */
367			default:
368				errno = events[i].data;
369				return (-1);
370			}
371		} else if (events[i].filter == EVFILT_READ) {
372			which |= EV_READ;
373		} else if (events[i].filter == EVFILT_WRITE) {
374			which |= EV_WRITE;
375		} else if (events[i].filter == EVFILT_SIGNAL) {
376			which |= EV_SIGNAL;
377		}
378
379		if (!which)
380			continue;
381
382		if (events[i].filter == EVFILT_SIGNAL) {
383			evmap_signal_active(base, events[i].ident, 1);
384		} else {
385			evmap_io_active(base, events[i].ident, which | EV_ET);
386		}
387	}
388
389	if (res == kqop->events_size) {
390		/* We used all the events space that we have. Maybe we should
391		   make it bigger. */
392		kq_grow_events(kqop, kqop->events_size * 2);
393	}
394
395	return (0);
396}
397
398static void
399kqop_free(struct kqop *kqop)
400{
401	if (kqop->changes)
402		mm_free(kqop->changes);
403	if (kqop->events)
404		mm_free(kqop->events);
405	if (kqop->kq >= 0 && kqop->pid == getpid())
406		close(kqop->kq);
407	memset(kqop, 0, sizeof(struct kqop));
408	mm_free(kqop);
409}
410
411static void
412kq_dealloc(struct event_base *base)
413{
414	struct kqop *kqop = base->evbase;
415	evsig_dealloc(base);
416	kqop_free(kqop);
417}
418
419/* signal handling */
420static int
421kq_sig_add(struct event_base *base, int nsignal, short old, short events, void *p)
422{
423	struct kqop *kqop = base->evbase;
424	struct kevent kev;
425	struct timespec timeout = { 0, 0 };
426	(void)p;
427
428	EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
429
430	memset(&kev, 0, sizeof(kev));
431	kev.ident = nsignal;
432	kev.filter = EVFILT_SIGNAL;
433	kev.flags = EV_ADD;
434
435	/* Be ready for the signal if it is sent any
436	 * time between now and the next call to
437	 * kq_dispatch. */
438	if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
439		return (-1);
440
441	/* XXXX The manpage suggest we could use SIG_IGN instead of a
442	 * do-nothing handler */
443	if (_evsig_set_handler(base, nsignal, kq_sighandler) == -1)
444		return (-1);
445
446	return (0);
447}
448
449static int
450kq_sig_del(struct event_base *base, int nsignal, short old, short events, void *p)
451{
452	struct kqop *kqop = base->evbase;
453	struct kevent kev;
454
455	struct timespec timeout = { 0, 0 };
456	(void)p;
457
458	EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
459
460	memset(&kev, 0, sizeof(kev));
461	kev.ident = nsignal;
462	kev.filter = EVFILT_SIGNAL;
463	kev.flags = EV_DELETE;
464
465	/* Because we insert signal events
466	 * immediately, we need to delete them
467	 * immediately, too */
468	if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
469		return (-1);
470
471	if (_evsig_restore_handler(base, nsignal) == -1)
472		return (-1);
473
474	return (0);
475}
476