michael@0: /*	$OpenBSD: kqueue.c,v 1.5 2002/07/10 14:41:31 art Exp $	*/
michael@0: 
michael@0: /*
michael@0:  * Copyright 2000-2007 Niels Provos <provos@citi.umich.edu>
michael@0:  * Copyright 2007-2012 Niels Provos and Nick Mathewson
michael@0:  *
michael@0:  * Redistribution and use in source and binary forms, with or without
michael@0:  * modification, are permitted provided that the following conditions
michael@0:  * are met:
michael@0:  * 1. Redistributions of source code must retain the above copyright
michael@0:  *    notice, this list of conditions and the following disclaimer.
michael@0:  * 2. Redistributions in binary form must reproduce the above copyright
michael@0:  *    notice, this list of conditions and the following disclaimer in the
michael@0:  *    documentation and/or other materials provided with the distribution.
michael@0:  * 3. The name of the author may not be used to endorse or promote products
michael@0:  *    derived from this software without specific prior written permission.
michael@0:  *
michael@0:  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
michael@0:  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
michael@0:  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
michael@0:  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
michael@0:  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
michael@0:  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
michael@0:  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
michael@0:  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
michael@0:  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
michael@0:  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
michael@0:  */
michael@0: #include "event2/event-config.h"
michael@0: 
michael@0: #define _GNU_SOURCE
michael@0: 
michael@0: #include <sys/types.h>
michael@0: #ifdef _EVENT_HAVE_SYS_TIME_H
michael@0: #include <sys/time.h>
michael@0: #endif
michael@0: #include <sys/queue.h>
michael@0: #include <sys/event.h>
michael@0: #include <signal.h>
michael@0: #include <stdio.h>
michael@0: #include <stdlib.h>
michael@0: #include <string.h>
michael@0: #include <unistd.h>
michael@0: #include <errno.h>
michael@0: #ifdef _EVENT_HAVE_INTTYPES_H
michael@0: #include <inttypes.h>
michael@0: #endif
michael@0: 
michael@0: /* Some platforms apparently define the udata field of struct kevent as
michael@0:  * intptr_t, whereas others define it as void*.  There doesn't seem to be an
michael@0:  * easy way to tell them apart via autoconf, so we need to use OS macros. */
michael@0: #if defined(_EVENT_HAVE_INTTYPES_H) && !defined(__OpenBSD__) && !defined(__FreeBSD__) && !defined(__darwin__) && !defined(__APPLE__)
michael@0: #define PTR_TO_UDATA(x)	((intptr_t)(x))
michael@0: #define INT_TO_UDATA(x) ((intptr_t)(x))
michael@0: #else
michael@0: #define PTR_TO_UDATA(x)	(x)
michael@0: #define INT_TO_UDATA(x) ((void*)(x))
michael@0: #endif
michael@0: 
michael@0: #include "event-internal.h"
michael@0: #include "log-internal.h"
michael@0: #include "evmap-internal.h"
michael@0: #include "event2/thread.h"
michael@0: #include "evthread-internal.h"
michael@0: #include "changelist-internal.h"
michael@0: 
michael@0: #define NEVENT		64
michael@0: 
michael@0: struct kqop {
michael@0: 	struct kevent *changes;
michael@0: 	int changes_size;
michael@0: 
michael@0: 	struct kevent *events;
michael@0: 	int events_size;
michael@0: 	int kq;
michael@0: 	pid_t pid;
michael@0: };
michael@0: 
michael@0: static void kqop_free(struct kqop *kqop);
michael@0: 
michael@0: static void *kq_init(struct event_base *);
michael@0: static int kq_sig_add(struct event_base *, int, short, short, void *);
michael@0: static int kq_sig_del(struct event_base *, int, short, short, void *);
michael@0: static int kq_dispatch(struct event_base *, struct timeval *);
michael@0: static void kq_dealloc(struct event_base *);
michael@0: 
michael@0: const struct eventop kqops = {
michael@0: 	"kqueue",
michael@0: 	kq_init,
michael@0: 	event_changelist_add,
michael@0: 	event_changelist_del,
michael@0: 	kq_dispatch,
michael@0: 	kq_dealloc,
michael@0: 	1 /* need reinit */,
michael@0:     EV_FEATURE_ET|EV_FEATURE_O1|EV_FEATURE_FDS,
michael@0: 	EVENT_CHANGELIST_FDINFO_SIZE
michael@0: };
michael@0: 
michael@0: static const struct eventop kqsigops = {
michael@0: 	"kqueue_signal",
michael@0: 	NULL,
michael@0: 	kq_sig_add,
michael@0: 	kq_sig_del,
michael@0: 	NULL,
michael@0: 	NULL,
michael@0: 	1 /* need reinit */,
michael@0: 	0,
michael@0: 	0
michael@0: };
michael@0: 
michael@0: static void *
michael@0: kq_init(struct event_base *base)
michael@0: {
michael@0: 	int kq = -1;
michael@0: 	struct kqop *kqueueop = NULL;
michael@0: 
michael@0: 	if (!(kqueueop = mm_calloc(1, sizeof(struct kqop))))
michael@0: 		return (NULL);
michael@0: 
michael@0: /* Initialize the kernel queue */
michael@0: 
michael@0: 	if ((kq = kqueue()) == -1) {
michael@0: 		event_warn("kqueue");
michael@0: 		goto err;
michael@0: 	}
michael@0: 
michael@0: 	kqueueop->kq = kq;
michael@0: 
michael@0: 	kqueueop->pid = getpid();
michael@0: 
michael@0: 	/* Initialize fields */
michael@0: 	kqueueop->changes = mm_calloc(NEVENT, sizeof(struct kevent));
michael@0: 	if (kqueueop->changes == NULL)
michael@0: 		goto err;
michael@0: 	kqueueop->events = mm_calloc(NEVENT, sizeof(struct kevent));
michael@0: 	if (kqueueop->events == NULL)
michael@0: 		goto err;
michael@0: 	kqueueop->events_size = kqueueop->changes_size = NEVENT;
michael@0: 
michael@0: 	/* Check for Mac OS X kqueue bug. */
michael@0: 	memset(&kqueueop->changes[0], 0, sizeof kqueueop->changes[0]);
michael@0: 	kqueueop->changes[0].ident = -1;
michael@0: 	kqueueop->changes[0].filter = EVFILT_READ;
michael@0: 	kqueueop->changes[0].flags = EV_ADD;
michael@0: 	/*
michael@0: 	 * If kqueue works, then kevent will succeed, and it will
michael@0: 	 * stick an error in events[0].  If kqueue is broken, then
michael@0: 	 * kevent will fail.
michael@0: 	 */
michael@0: 	if (kevent(kq,
michael@0: 		kqueueop->changes, 1, kqueueop->events, NEVENT, NULL) != 1 ||
michael@0: 	    (int)kqueueop->events[0].ident != -1 ||
michael@0: 	    kqueueop->events[0].flags != EV_ERROR) {
michael@0: 		event_warn("%s: detected broken kqueue; not using.", __func__);
michael@0: 		goto err;
michael@0: 	}
michael@0: 
michael@0: 	base->evsigsel = &kqsigops;
michael@0: 
michael@0: 	return (kqueueop);
michael@0: err:
michael@0: 	if (kqueueop)
michael@0: 		kqop_free(kqueueop);
michael@0: 
michael@0: 	return (NULL);
michael@0: }
michael@0: 
michael@0: #define ADD_UDATA 0x30303
michael@0: 
michael@0: static void
michael@0: kq_setup_kevent(struct kevent *out, evutil_socket_t fd, int filter, short change)
michael@0: {
michael@0: 	memset(out, 0, sizeof(struct kevent));
michael@0: 	out->ident = fd;
michael@0: 	out->filter = filter;
michael@0: 
michael@0: 	if (change & EV_CHANGE_ADD) {
michael@0: 		out->flags = EV_ADD;
michael@0: 		/* We set a magic number here so that we can tell 'add'
michael@0: 		 * errors from 'del' errors. */
michael@0: 		out->udata = INT_TO_UDATA(ADD_UDATA);
michael@0: 		if (change & EV_ET)
michael@0: 			out->flags |= EV_CLEAR;
michael@0: #ifdef NOTE_EOF
michael@0: 		/* Make it behave like select() and poll() */
michael@0: 		if (filter == EVFILT_READ)
michael@0: 			out->fflags = NOTE_EOF;
michael@0: #endif
michael@0: 	} else {
michael@0: 		EVUTIL_ASSERT(change & EV_CHANGE_DEL);
michael@0: 		out->flags = EV_DELETE;
michael@0: 	}
michael@0: }
michael@0: 
michael@0: static int
michael@0: kq_build_changes_list(const struct event_changelist *changelist,
michael@0:     struct kqop *kqop)
michael@0: {
michael@0: 	int i;
michael@0: 	int n_changes = 0;
michael@0: 
michael@0: 	for (i = 0; i < changelist->n_changes; ++i) {
michael@0: 		struct event_change *in_ch = &changelist->changes[i];
michael@0: 		struct kevent *out_ch;
michael@0: 		if (n_changes >= kqop->changes_size - 1) {
michael@0: 			int newsize = kqop->changes_size * 2;
michael@0: 			struct kevent *newchanges;
michael@0: 
michael@0: 			newchanges = mm_realloc(kqop->changes,
michael@0: 			    newsize * sizeof(struct kevent));
michael@0: 			if (newchanges == NULL) {
michael@0: 				event_warn("%s: realloc", __func__);
michael@0: 				return (-1);
michael@0: 			}
michael@0: 			kqop->changes = newchanges;
michael@0: 			kqop->changes_size = newsize;
michael@0: 		}
michael@0: 		if (in_ch->read_change) {
michael@0: 			out_ch = &kqop->changes[n_changes++];
michael@0: 			kq_setup_kevent(out_ch, in_ch->fd, EVFILT_READ,
michael@0: 			    in_ch->read_change);
michael@0: 		}
michael@0: 		if (in_ch->write_change) {
michael@0: 			out_ch = &kqop->changes[n_changes++];
michael@0: 			kq_setup_kevent(out_ch, in_ch->fd, EVFILT_WRITE,
michael@0: 			    in_ch->write_change);
michael@0: 		}
michael@0: 	}
michael@0: 	return n_changes;
michael@0: }
michael@0: 
michael@0: static int
michael@0: kq_grow_events(struct kqop *kqop, size_t new_size)
michael@0: {
michael@0: 	struct kevent *newresult;
michael@0: 
michael@0: 	newresult = mm_realloc(kqop->events,
michael@0: 	    new_size * sizeof(struct kevent));
michael@0: 
michael@0: 	if (newresult) {
michael@0: 		kqop->events = newresult;
michael@0: 		kqop->events_size = new_size;
michael@0: 		return 0;
michael@0: 	} else {
michael@0: 		return -1;
michael@0: 	}
michael@0: }
michael@0: 
michael@0: static int
michael@0: kq_dispatch(struct event_base *base, struct timeval *tv)
michael@0: {
michael@0: 	struct kqop *kqop = base->evbase;
michael@0: 	struct kevent *events = kqop->events;
michael@0: 	struct kevent *changes;
michael@0: 	struct timespec ts, *ts_p = NULL;
michael@0: 	int i, n_changes, res;
michael@0: 
michael@0: 	if (tv != NULL) {
michael@0: 		TIMEVAL_TO_TIMESPEC(tv, &ts);
michael@0: 		ts_p = &ts;
michael@0: 	}
michael@0: 
michael@0: 	/* Build "changes" from "base->changes" */
michael@0: 	EVUTIL_ASSERT(kqop->changes);
michael@0: 	n_changes = kq_build_changes_list(&base->changelist, kqop);
michael@0: 	if (n_changes < 0)
michael@0: 		return -1;
michael@0: 
michael@0: 	event_changelist_remove_all(&base->changelist, base);
michael@0: 
michael@0: 	/* steal the changes array in case some broken code tries to call
michael@0: 	 * dispatch twice at once. */
michael@0: 	changes = kqop->changes;
michael@0: 	kqop->changes = NULL;
michael@0: 
michael@0: 	/* Make sure that 'events' is at least as long as the list of changes:
michael@0: 	 * otherwise errors in the changes can get reported as a -1 return
michael@0: 	 * value from kevent() rather than as EV_ERROR events in the events
michael@0: 	 * array.
michael@0: 	 *
michael@0: 	 * (We could instead handle -1 return values from kevent() by
michael@0: 	 * retrying with a smaller changes array or a larger events array,
michael@0: 	 * but this approach seems less risky for now.)
michael@0: 	 */
michael@0: 	if (kqop->events_size < n_changes) {
michael@0: 		int new_size = kqop->events_size;
michael@0: 		do {
michael@0: 			new_size *= 2;
michael@0: 		} while (new_size < n_changes);
michael@0: 
michael@0: 		kq_grow_events(kqop, new_size);
michael@0: 		events = kqop->events;
michael@0: 	}
michael@0: 
michael@0: 	EVBASE_RELEASE_LOCK(base, th_base_lock);
michael@0: 
michael@0: 	res = kevent(kqop->kq, changes, n_changes,
michael@0: 	    events, kqop->events_size, ts_p);
michael@0: 
michael@0: 	EVBASE_ACQUIRE_LOCK(base, th_base_lock);
michael@0: 
michael@0: 	EVUTIL_ASSERT(kqop->changes == NULL);
michael@0: 	kqop->changes = changes;
michael@0: 
michael@0: 	if (res == -1) {
michael@0: 		if (errno != EINTR) {
michael@0: 			event_warn("kevent");
michael@0: 			return (-1);
michael@0: 		}
michael@0: 
michael@0: 		return (0);
michael@0: 	}
michael@0: 
michael@0: 	event_debug(("%s: kevent reports %d", __func__, res));
michael@0: 
michael@0: 	for (i = 0; i < res; i++) {
michael@0: 		int which = 0;
michael@0: 
michael@0: 		if (events[i].flags & EV_ERROR) {
michael@0: 			switch (events[i].data) {
michael@0: 
michael@0: 			/* Can occur on delete if we are not currently
michael@0: 			 * watching any events on this fd.  That can
michael@0: 			 * happen when the fd was closed and another
michael@0: 			 * file was opened with that fd. */
michael@0: 			case ENOENT:
michael@0: 			/* Can occur for reasons not fully understood
michael@0: 			 * on FreeBSD. */
michael@0: 			case EINVAL:
michael@0: 				continue;
michael@0: 
michael@0: 			/* Can occur on a delete if the fd is closed. */
michael@0: 			case EBADF:
michael@0: 				/* XXXX On NetBSD, we can also get EBADF if we
michael@0: 				 * try to add the write side of a pipe, but
michael@0: 				 * the read side has already been closed.
michael@0: 				 * Other BSDs call this situation 'EPIPE'. It
michael@0: 				 * would be good if we had a way to report
michael@0: 				 * this situation. */
michael@0: 				continue;
michael@0: 			/* These two can occur on an add if the fd was one side
michael@0: 			 * of a pipe, and the other side was closed. */
michael@0: 			case EPERM:
michael@0: 			case EPIPE:
michael@0: 				/* Report read events, if we're listening for
michael@0: 				 * them, so that the user can learn about any
michael@0: 				 * add errors.  (If the operation was a
michael@0: 				 * delete, then udata should be cleared.) */
michael@0: 				if (events[i].udata) {
michael@0: 					/* The operation was an add:
michael@0: 					 * report the error as a read. */
michael@0: 					which |= EV_READ;
michael@0: 					break;
michael@0: 				} else {
michael@0: 					/* The operation was a del:
michael@0: 					 * report nothing. */
michael@0: 					continue;
michael@0: 				}
michael@0: 
michael@0: 			/* Other errors shouldn't occur. */
michael@0: 			default:
michael@0: 				errno = events[i].data;
michael@0: 				return (-1);
michael@0: 			}
michael@0: 		} else if (events[i].filter == EVFILT_READ) {
michael@0: 			which |= EV_READ;
michael@0: 		} else if (events[i].filter == EVFILT_WRITE) {
michael@0: 			which |= EV_WRITE;
michael@0: 		} else if (events[i].filter == EVFILT_SIGNAL) {
michael@0: 			which |= EV_SIGNAL;
michael@0: 		}
michael@0: 
michael@0: 		if (!which)
michael@0: 			continue;
michael@0: 
michael@0: 		if (events[i].filter == EVFILT_SIGNAL) {
michael@0: 			evmap_signal_active(base, events[i].ident, 1);
michael@0: 		} else {
michael@0: 			evmap_io_active(base, events[i].ident, which | EV_ET);
michael@0: 		}
michael@0: 	}
michael@0: 
michael@0: 	if (res == kqop->events_size) {
michael@0: 		/* We used all the events space that we have. Maybe we should
michael@0: 		   make it bigger. */
michael@0: 		kq_grow_events(kqop, kqop->events_size * 2);
michael@0: 	}
michael@0: 
michael@0: 	return (0);
michael@0: }
michael@0: 
michael@0: static void
michael@0: kqop_free(struct kqop *kqop)
michael@0: {
michael@0: 	if (kqop->changes)
michael@0: 		mm_free(kqop->changes);
michael@0: 	if (kqop->events)
michael@0: 		mm_free(kqop->events);
michael@0: 	if (kqop->kq >= 0 && kqop->pid == getpid())
michael@0: 		close(kqop->kq);
michael@0: 	memset(kqop, 0, sizeof(struct kqop));
michael@0: 	mm_free(kqop);
michael@0: }
michael@0: 
michael@0: static void
michael@0: kq_dealloc(struct event_base *base)
michael@0: {
michael@0: 	struct kqop *kqop = base->evbase;
michael@0: 	evsig_dealloc(base);
michael@0: 	kqop_free(kqop);
michael@0: }
michael@0: 
michael@0: /* signal handling */
michael@0: static int
michael@0: kq_sig_add(struct event_base *base, int nsignal, short old, short events, void *p)
michael@0: {
michael@0: 	struct kqop *kqop = base->evbase;
michael@0: 	struct kevent kev;
michael@0: 	struct timespec timeout = { 0, 0 };
michael@0: 	(void)p;
michael@0: 
michael@0: 	EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
michael@0: 
michael@0: 	memset(&kev, 0, sizeof(kev));
michael@0: 	kev.ident = nsignal;
michael@0: 	kev.filter = EVFILT_SIGNAL;
michael@0: 	kev.flags = EV_ADD;
michael@0: 
michael@0: 	/* Be ready for the signal if it is sent any
michael@0: 	 * time between now and the next call to
michael@0: 	 * kq_dispatch. */
michael@0: 	if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
michael@0: 		return (-1);
michael@0: 
michael@0: 	/* Backported from
michael@0: 	 * https://github.com/nmathewson/Libevent/commit/148458e0a1fd25e167aa2ef229d1c9a70b27c3e9 */
michael@0: 	/* We can set the handler for most signals to SIG_IGN and
michael@0: 	 * still have them reported to us in the queue.  However,
michael@0: 	 * if the handler for SIGCHLD is SIG_IGN, the system reaps
michael@0: 	 * zombie processes for us, and we don't get any notification.
michael@0: 	 * This appears to be the only signal with this quirk. */
michael@0: 	if (_evsig_set_handler(base, nsignal,
michael@0: 	                       nsignal == SIGCHLD ? SIG_DFL : SIG_IGN) == -1) {
michael@0: 		return (-1);
michael@0: 	}
michael@0: 
michael@0: 	return (0);
michael@0: }
michael@0: 
michael@0: static int
michael@0: kq_sig_del(struct event_base *base, int nsignal, short old, short events, void *p)
michael@0: {
michael@0: 	struct kqop *kqop = base->evbase;
michael@0: 	struct kevent kev;
michael@0: 
michael@0: 	struct timespec timeout = { 0, 0 };
michael@0: 	(void)p;
michael@0: 
michael@0: 	EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
michael@0: 
michael@0: 	memset(&kev, 0, sizeof(kev));
michael@0: 	kev.ident = nsignal;
michael@0: 	kev.filter = EVFILT_SIGNAL;
michael@0: 	kev.flags = EV_DELETE;
michael@0: 
michael@0: 	/* Because we insert signal events
michael@0: 	 * immediately, we need to delete them
michael@0: 	 * immediately, too */
michael@0: 	if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
michael@0: 		return (-1);
michael@0: 
michael@0: 	if (_evsig_restore_handler(base, nsignal) == -1)
michael@0: 		return (-1);
michael@0: 
michael@0: 	return (0);
michael@0: }