michael@0: /*
michael@0:  * Copyright 2000-2009 Niels Provos <provos@citi.umich.edu>
michael@0:  * Copyright 2009-2012 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: #include <sys/types.h>
michael@0: #include <sys/resource.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/devpoll.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 <fcntl.h>
michael@0: #include <errno.h>
michael@0: 
michael@0: #include "event2/event.h"
michael@0: #include "event2/event_struct.h"
michael@0: #include "event2/thread.h"
michael@0: #include "event-internal.h"
michael@0: #include "evsignal-internal.h"
michael@0: #include "log-internal.h"
michael@0: #include "evmap-internal.h"
michael@0: #include "evthread-internal.h"
michael@0: 
michael@0: struct devpollop {
michael@0: 	struct pollfd *events;
michael@0: 	int nevents;
michael@0: 	int dpfd;
michael@0: 	struct pollfd *changes;
michael@0: 	int nchanges;
michael@0: };
michael@0: 
michael@0: static void *devpoll_init(struct event_base *);
michael@0: static int devpoll_add(struct event_base *, int fd, short old, short events, void *);
michael@0: static int devpoll_del(struct event_base *, int fd, short old, short events, void *);
michael@0: static int devpoll_dispatch(struct event_base *, struct timeval *);
michael@0: static void devpoll_dealloc(struct event_base *);
michael@0: 
michael@0: const struct eventop devpollops = {
michael@0: 	"devpoll",
michael@0: 	devpoll_init,
michael@0: 	devpoll_add,
michael@0: 	devpoll_del,
michael@0: 	devpoll_dispatch,
michael@0: 	devpoll_dealloc,
michael@0: 	1, /* need reinit */
michael@0: 	EV_FEATURE_FDS|EV_FEATURE_O1,
michael@0: 	0
michael@0: };
michael@0: 
michael@0: #define NEVENT	32000
michael@0: 
michael@0: static int
michael@0: devpoll_commit(struct devpollop *devpollop)
michael@0: {
michael@0: 	/*
michael@0: 	 * Due to a bug in Solaris, we have to use pwrite with an offset of 0.
michael@0: 	 * Write is limited to 2GB of data, until it will fail.
michael@0: 	 */
michael@0: 	if (pwrite(devpollop->dpfd, devpollop->changes,
michael@0: 		sizeof(struct pollfd) * devpollop->nchanges, 0) == -1)
michael@0: 		return (-1);
michael@0: 
michael@0: 	devpollop->nchanges = 0;
michael@0: 	return (0);
michael@0: }
michael@0: 
michael@0: static int
michael@0: devpoll_queue(struct devpollop *devpollop, int fd, int events) {
michael@0: 	struct pollfd *pfd;
michael@0: 
michael@0: 	if (devpollop->nchanges >= devpollop->nevents) {
michael@0: 		/*
michael@0: 		 * Change buffer is full, must commit it to /dev/poll before
michael@0: 		 * adding more
michael@0: 		 */
michael@0: 		if (devpoll_commit(devpollop) != 0)
michael@0: 			return (-1);
michael@0: 	}
michael@0: 
michael@0: 	pfd = &devpollop->changes[devpollop->nchanges++];
michael@0: 	pfd->fd = fd;
michael@0: 	pfd->events = events;
michael@0: 	pfd->revents = 0;
michael@0: 
michael@0: 	return (0);
michael@0: }
michael@0: 
michael@0: static void *
michael@0: devpoll_init(struct event_base *base)
michael@0: {
michael@0: 	int dpfd, nfiles = NEVENT;
michael@0: 	struct rlimit rl;
michael@0: 	struct devpollop *devpollop;
michael@0: 
michael@0: 	if (!(devpollop = mm_calloc(1, sizeof(struct devpollop))))
michael@0: 		return (NULL);
michael@0: 
michael@0: 	if (getrlimit(RLIMIT_NOFILE, &rl) == 0 &&
michael@0: 	    rl.rlim_cur != RLIM_INFINITY)
michael@0: 		nfiles = rl.rlim_cur;
michael@0: 
michael@0: 	/* Initialize the kernel queue */
michael@0: 	if ((dpfd = evutil_open_closeonexec("/dev/poll", O_RDWR, 0)) == -1) {
michael@0: 		event_warn("open: /dev/poll");
michael@0: 		mm_free(devpollop);
michael@0: 		return (NULL);
michael@0: 	}
michael@0: 
michael@0: 	devpollop->dpfd = dpfd;
michael@0: 
michael@0: 	/* Initialize fields */
michael@0: 	/* FIXME: allocating 'nfiles' worth of space here can be
michael@0: 	 * expensive and unnecessary.  See how epoll.c does it instead. */
michael@0: 	devpollop->events = mm_calloc(nfiles, sizeof(struct pollfd));
michael@0: 	if (devpollop->events == NULL) {
michael@0: 		mm_free(devpollop);
michael@0: 		close(dpfd);
michael@0: 		return (NULL);
michael@0: 	}
michael@0: 	devpollop->nevents = nfiles;
michael@0: 
michael@0: 	devpollop->changes = mm_calloc(nfiles, sizeof(struct pollfd));
michael@0: 	if (devpollop->changes == NULL) {
michael@0: 		mm_free(devpollop->events);
michael@0: 		mm_free(devpollop);
michael@0: 		close(dpfd);
michael@0: 		return (NULL);
michael@0: 	}
michael@0: 
michael@0: 	evsig_init(base);
michael@0: 
michael@0: 	return (devpollop);
michael@0: }
michael@0: 
michael@0: static int
michael@0: devpoll_dispatch(struct event_base *base, struct timeval *tv)
michael@0: {
michael@0: 	struct devpollop *devpollop = base->evbase;
michael@0: 	struct pollfd *events = devpollop->events;
michael@0: 	struct dvpoll dvp;
michael@0: 	int i, res, timeout = -1;
michael@0: 
michael@0: 	if (devpollop->nchanges)
michael@0: 		devpoll_commit(devpollop);
michael@0: 
michael@0: 	if (tv != NULL)
michael@0: 		timeout = tv->tv_sec * 1000 + (tv->tv_usec + 999) / 1000;
michael@0: 
michael@0: 	dvp.dp_fds = devpollop->events;
michael@0: 	dvp.dp_nfds = devpollop->nevents;
michael@0: 	dvp.dp_timeout = timeout;
michael@0: 
michael@0: 	EVBASE_RELEASE_LOCK(base, th_base_lock);
michael@0: 
michael@0: 	res = ioctl(devpollop->dpfd, DP_POLL, &dvp);
michael@0: 
michael@0: 	EVBASE_ACQUIRE_LOCK(base, th_base_lock);
michael@0: 
michael@0: 	if (res == -1) {
michael@0: 		if (errno != EINTR) {
michael@0: 			event_warn("ioctl: DP_POLL");
michael@0: 			return (-1);
michael@0: 		}
michael@0: 
michael@0: 		return (0);
michael@0: 	}
michael@0: 
michael@0: 	event_debug(("%s: devpoll_wait reports %d", __func__, res));
michael@0: 
michael@0: 	for (i = 0; i < res; i++) {
michael@0: 		int which = 0;
michael@0: 		int what = events[i].revents;
michael@0: 
michael@0: 		if (what & POLLHUP)
michael@0: 			what |= POLLIN | POLLOUT;
michael@0: 		else if (what & POLLERR)
michael@0: 			what |= POLLIN | POLLOUT;
michael@0: 
michael@0: 		if (what & POLLIN)
michael@0: 			which |= EV_READ;
michael@0: 		if (what & POLLOUT)
michael@0: 			which |= EV_WRITE;
michael@0: 
michael@0: 		if (!which)
michael@0: 			continue;
michael@0: 
michael@0: 		/* XXX(niels): not sure if this works for devpoll */
michael@0: 		evmap_io_active(base, events[i].fd, which);
michael@0: 	}
michael@0: 
michael@0: 	return (0);
michael@0: }
michael@0: 
michael@0: 
michael@0: static int
michael@0: devpoll_add(struct event_base *base, int fd, short old, short events, void *p)
michael@0: {
michael@0: 	struct devpollop *devpollop = base->evbase;
michael@0: 	int res;
michael@0: 	(void)p;
michael@0: 
michael@0: 	/*
michael@0: 	 * It's not necessary to OR the existing read/write events that we
michael@0: 	 * are currently interested in with the new event we are adding.
michael@0: 	 * The /dev/poll driver ORs any new events with the existing events
michael@0: 	 * that it has cached for the fd.
michael@0: 	 */
michael@0: 
michael@0: 	res = 0;
michael@0: 	if (events & EV_READ)
michael@0: 		res |= POLLIN;
michael@0: 	if (events & EV_WRITE)
michael@0: 		res |= POLLOUT;
michael@0: 
michael@0: 	if (devpoll_queue(devpollop, fd, res) != 0)
michael@0: 		return (-1);
michael@0: 
michael@0: 	return (0);
michael@0: }
michael@0: 
michael@0: static int
michael@0: devpoll_del(struct event_base *base, int fd, short old, short events, void *p)
michael@0: {
michael@0: 	struct devpollop *devpollop = base->evbase;
michael@0: 	int res;
michael@0: 	(void)p;
michael@0: 
michael@0: 	res = 0;
michael@0: 	if (events & EV_READ)
michael@0: 		res |= POLLIN;
michael@0: 	if (events & EV_WRITE)
michael@0: 		res |= POLLOUT;
michael@0: 
michael@0: 	/*
michael@0: 	 * The only way to remove an fd from the /dev/poll monitored set is
michael@0: 	 * to use POLLREMOVE by itself.  This removes ALL events for the fd
michael@0: 	 * provided so if we care about two events and are only removing one
michael@0: 	 * we must re-add the other event after POLLREMOVE.
michael@0: 	 */
michael@0: 
michael@0: 	if (devpoll_queue(devpollop, fd, POLLREMOVE) != 0)
michael@0: 		return (-1);
michael@0: 
michael@0: 	if ((res & (POLLIN|POLLOUT)) != (POLLIN|POLLOUT)) {
michael@0: 		/*
michael@0: 		 * We're not deleting all events, so we must resubmit the
michael@0: 		 * event that we are still interested in if one exists.
michael@0: 		 */
michael@0: 
michael@0: 		if ((res & POLLIN) && (old & EV_WRITE)) {
michael@0: 			/* Deleting read, still care about write */
michael@0: 			devpoll_queue(devpollop, fd, POLLOUT);
michael@0: 		} else if ((res & POLLOUT) && (old & EV_READ)) {
michael@0: 			/* Deleting write, still care about read */
michael@0: 			devpoll_queue(devpollop, fd, POLLIN);
michael@0: 		}
michael@0: 	}
michael@0: 
michael@0: 	return (0);
michael@0: }
michael@0: 
michael@0: static void
michael@0: devpoll_dealloc(struct event_base *base)
michael@0: {
michael@0: 	struct devpollop *devpollop = base->evbase;
michael@0: 
michael@0: 	evsig_dealloc(base);
michael@0: 	if (devpollop->events)
michael@0: 		mm_free(devpollop->events);
michael@0: 	if (devpollop->changes)
michael@0: 		mm_free(devpollop->changes);
michael@0: 	if (devpollop->dpfd >= 0)
michael@0: 		close(devpollop->dpfd);
michael@0: 
michael@0: 	memset(devpollop, 0, sizeof(struct devpollop));
michael@0: 	mm_free(devpollop);
michael@0: }