michael@0: /*
michael@0:  * Submitted by David Pacheco (dp.spambait@gmail.com)
michael@0:  *
michael@0:  * Copyright 2006-2007 Niels Provos
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 SUN MICROSYSTEMS, INC. ``AS IS'' AND ANY
michael@0:  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
michael@0:  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
michael@0:  * DISCLAIMED. IN NO EVENT SHALL SUN MICROSYSTEMS, INC. BE LIABLE FOR ANY
michael@0:  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
michael@0:  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
michael@0:  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
michael@0:  * ON ANY 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 THIS
michael@0:  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
michael@0:  */
michael@0: 
michael@0: /*
michael@0:  * Copyright (c) 2007 Sun Microsystems. All rights reserved.
michael@0:  * Use is subject to license terms.
michael@0:  */
michael@0: 
michael@0: /*
michael@0:  * evport.c: event backend using Solaris 10 event ports. See port_create(3C).
michael@0:  * This implementation is loosely modeled after the one used for select(2) (in
michael@0:  * select.c).
michael@0:  *
michael@0:  * The outstanding events are tracked in a data structure called evport_data.
michael@0:  * Each entry in the ed_fds array corresponds to a file descriptor, and contains
michael@0:  * pointers to the read and write events that correspond to that fd. (That is,
michael@0:  * when the file is readable, the "read" event should handle it, etc.)
michael@0:  *
michael@0:  * evport_add and evport_del update this data structure. evport_dispatch uses it
michael@0:  * to determine where to callback when an event occurs (which it gets from
michael@0:  * port_getn).
michael@0:  *
michael@0:  * Helper functions are used: grow() grows the file descriptor array as
michael@0:  * necessary when large fd's come in. reassociate() takes care of maintaining
michael@0:  * the proper file-descriptor/event-port associations.
michael@0:  *
michael@0:  * As in the select(2) implementation, signals are handled by evsignal.
michael@0:  */
michael@0: 
michael@0: #include "event2/event-config.h"
michael@0: 
michael@0: #include <sys/time.h>
michael@0: #include <sys/queue.h>
michael@0: #include <errno.h>
michael@0: #include <poll.h>
michael@0: #include <port.h>
michael@0: #include <signal.h>
michael@0: #include <stdio.h>
michael@0: #include <stdlib.h>
michael@0: #include <string.h>
michael@0: #include <time.h>
michael@0: #include <unistd.h>
michael@0: 
michael@0: #include "event2/thread.h"
michael@0: 
michael@0: #include "evthread-internal.h"
michael@0: #include "event-internal.h"
michael@0: #include "log-internal.h"
michael@0: #include "evsignal-internal.h"
michael@0: #include "evmap-internal.h"
michael@0: 
michael@0: /*
michael@0:  * Default value for ed_nevents, which is the maximum file descriptor number we
michael@0:  * can handle. If an event comes in for a file descriptor F > nevents, we will
michael@0:  * grow the array of file descriptors, doubling its size.
michael@0:  */
michael@0: #define DEFAULT_NFDS	16
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * EVENTS_PER_GETN is the maximum number of events to retrieve from port_getn on
michael@0:  * any particular call. You can speed things up by increasing this, but it will
michael@0:  * (obviously) require more memory.
michael@0:  */
michael@0: #define EVENTS_PER_GETN 8
michael@0: 
michael@0: /*
michael@0:  * Per-file-descriptor information about what events we're subscribed to. These
michael@0:  * fields are NULL if no event is subscribed to either of them.
michael@0:  */
michael@0: 
michael@0: struct fd_info {
michael@0: 	short fdi_what;		/* combinations of EV_READ and EV_WRITE */
michael@0: };
michael@0: 
michael@0: #define FDI_HAS_READ(fdi)  ((fdi)->fdi_what & EV_READ)
michael@0: #define FDI_HAS_WRITE(fdi) ((fdi)->fdi_what & EV_WRITE)
michael@0: #define FDI_HAS_EVENTS(fdi) (FDI_HAS_READ(fdi) || FDI_HAS_WRITE(fdi))
michael@0: #define FDI_TO_SYSEVENTS(fdi) (FDI_HAS_READ(fdi) ? POLLIN : 0) | \
michael@0:     (FDI_HAS_WRITE(fdi) ? POLLOUT : 0)
michael@0: 
michael@0: struct evport_data {
michael@0: 	int		ed_port;	/* event port for system events  */
michael@0: 	int		ed_nevents;	/* number of allocated fdi's	 */
michael@0: 	struct fd_info *ed_fds;		/* allocated fdi table		 */
michael@0: 	/* fdi's that we need to reassoc */
michael@0: 	int ed_pending[EVENTS_PER_GETN]; /* fd's with pending events */
michael@0: };
michael@0: 
michael@0: static void*	evport_init(struct event_base *);
michael@0: static int evport_add(struct event_base *, int fd, short old, short events, void *);
michael@0: static int evport_del(struct event_base *, int fd, short old, short events, void *);
michael@0: static int	evport_dispatch(struct event_base *, struct timeval *);
michael@0: static void	evport_dealloc(struct event_base *);
michael@0: 
michael@0: const struct eventop evportops = {
michael@0: 	"evport",
michael@0: 	evport_init,
michael@0: 	evport_add,
michael@0: 	evport_del,
michael@0: 	evport_dispatch,
michael@0: 	evport_dealloc,
michael@0: 	1, /* need reinit */
michael@0: 	0, /* features */
michael@0: 	0, /* fdinfo length */
michael@0: };
michael@0: 
michael@0: /*
michael@0:  * Initialize the event port implementation.
michael@0:  */
michael@0: 
michael@0: static void*
michael@0: evport_init(struct event_base *base)
michael@0: {
michael@0: 	struct evport_data *evpd;
michael@0: 	int i;
michael@0: 
michael@0: 	if (!(evpd = mm_calloc(1, sizeof(struct evport_data))))
michael@0: 		return (NULL);
michael@0: 
michael@0: 	if ((evpd->ed_port = port_create()) == -1) {
michael@0: 		mm_free(evpd);
michael@0: 		return (NULL);
michael@0: 	}
michael@0: 
michael@0: 	/*
michael@0: 	 * Initialize file descriptor structure
michael@0: 	 */
michael@0: 	evpd->ed_fds = mm_calloc(DEFAULT_NFDS, sizeof(struct fd_info));
michael@0: 	if (evpd->ed_fds == NULL) {
michael@0: 		close(evpd->ed_port);
michael@0: 		mm_free(evpd);
michael@0: 		return (NULL);
michael@0: 	}
michael@0: 	evpd->ed_nevents = DEFAULT_NFDS;
michael@0: 	for (i = 0; i < EVENTS_PER_GETN; i++)
michael@0: 		evpd->ed_pending[i] = -1;
michael@0: 
michael@0: 	evsig_init(base);
michael@0: 
michael@0: 	return (evpd);
michael@0: }
michael@0: 
michael@0: #ifdef CHECK_INVARIANTS
michael@0: /*
michael@0:  * Checks some basic properties about the evport_data structure. Because it
michael@0:  * checks all file descriptors, this function can be expensive when the maximum
michael@0:  * file descriptor ever used is rather large.
michael@0:  */
michael@0: 
michael@0: static void
michael@0: check_evportop(struct evport_data *evpd)
michael@0: {
michael@0: 	EVUTIL_ASSERT(evpd);
michael@0: 	EVUTIL_ASSERT(evpd->ed_nevents > 0);
michael@0: 	EVUTIL_ASSERT(evpd->ed_port > 0);
michael@0: 	EVUTIL_ASSERT(evpd->ed_fds > 0);
michael@0: }
michael@0: 
michael@0: /*
michael@0:  * Verifies very basic integrity of a given port_event.
michael@0:  */
michael@0: static void
michael@0: check_event(port_event_t* pevt)
michael@0: {
michael@0: 	/*
michael@0: 	 * We've only registered for PORT_SOURCE_FD events. The only
michael@0: 	 * other thing we can legitimately receive is PORT_SOURCE_ALERT,
michael@0: 	 * but since we're not using port_alert either, we can assume
michael@0: 	 * PORT_SOURCE_FD.
michael@0: 	 */
michael@0: 	EVUTIL_ASSERT(pevt->portev_source == PORT_SOURCE_FD);
michael@0: 	EVUTIL_ASSERT(pevt->portev_user == NULL);
michael@0: }
michael@0: 
michael@0: #else
michael@0: #define check_evportop(epop)
michael@0: #define check_event(pevt)
michael@0: #endif /* CHECK_INVARIANTS */
michael@0: 
michael@0: /*
michael@0:  * Doubles the size of the allocated file descriptor array.
michael@0:  */
michael@0: static int
michael@0: grow(struct evport_data *epdp, int factor)
michael@0: {
michael@0: 	struct fd_info *tmp;
michael@0: 	int oldsize = epdp->ed_nevents;
michael@0: 	int newsize = factor * oldsize;
michael@0: 	EVUTIL_ASSERT(factor > 1);
michael@0: 
michael@0: 	check_evportop(epdp);
michael@0: 
michael@0: 	tmp = mm_realloc(epdp->ed_fds, sizeof(struct fd_info) * newsize);
michael@0: 	if (NULL == tmp)
michael@0: 		return -1;
michael@0: 	epdp->ed_fds = tmp;
michael@0: 	memset((char*) (epdp->ed_fds + oldsize), 0,
michael@0: 	    (newsize - oldsize)*sizeof(struct fd_info));
michael@0: 	epdp->ed_nevents = newsize;
michael@0: 
michael@0: 	check_evportop(epdp);
michael@0: 
michael@0: 	return 0;
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * (Re)associates the given file descriptor with the event port. The OS events
michael@0:  * are specified (implicitly) from the fd_info struct.
michael@0:  */
michael@0: static int
michael@0: reassociate(struct evport_data *epdp, struct fd_info *fdip, int fd)
michael@0: {
michael@0: 	int sysevents = FDI_TO_SYSEVENTS(fdip);
michael@0: 
michael@0: 	if (sysevents != 0) {
michael@0: 		if (port_associate(epdp->ed_port, PORT_SOURCE_FD,
michael@0: 				   fd, sysevents, NULL) == -1) {
michael@0: 			event_warn("port_associate");
michael@0: 			return (-1);
michael@0: 		}
michael@0: 	}
michael@0: 
michael@0: 	check_evportop(epdp);
michael@0: 
michael@0: 	return (0);
michael@0: }
michael@0: 
michael@0: /*
michael@0:  * Main event loop - polls port_getn for some number of events, and processes
michael@0:  * them.
michael@0:  */
michael@0: 
michael@0: static int
michael@0: evport_dispatch(struct event_base *base, struct timeval *tv)
michael@0: {
michael@0: 	int i, res;
michael@0: 	struct evport_data *epdp = base->evbase;
michael@0: 	port_event_t pevtlist[EVENTS_PER_GETN];
michael@0: 
michael@0: 	/*
michael@0: 	 * port_getn will block until it has at least nevents events. It will
michael@0: 	 * also return how many it's given us (which may be more than we asked
michael@0: 	 * for, as long as it's less than our maximum (EVENTS_PER_GETN)) in
michael@0: 	 * nevents.
michael@0: 	 */
michael@0: 	int nevents = 1;
michael@0: 
michael@0: 	/*
michael@0: 	 * We have to convert a struct timeval to a struct timespec
michael@0: 	 * (only difference is nanoseconds vs. microseconds). If no time-based
michael@0: 	 * events are active, we should wait for I/O (and tv == NULL).
michael@0: 	 */
michael@0: 	struct timespec ts;
michael@0: 	struct timespec *ts_p = NULL;
michael@0: 	if (tv != NULL) {
michael@0: 		ts.tv_sec = tv->tv_sec;
michael@0: 		ts.tv_nsec = tv->tv_usec * 1000;
michael@0: 		ts_p = &ts;
michael@0: 	}
michael@0: 
michael@0: 	/*
michael@0: 	 * Before doing anything else, we need to reassociate the events we hit
michael@0: 	 * last time which need reassociation. See comment at the end of the
michael@0: 	 * loop below.
michael@0: 	 */
michael@0: 	for (i = 0; i < EVENTS_PER_GETN; ++i) {
michael@0: 		struct fd_info *fdi = NULL;
michael@0: 		if (epdp->ed_pending[i] != -1) {
michael@0: 			fdi = &(epdp->ed_fds[epdp->ed_pending[i]]);
michael@0: 		}
michael@0: 
michael@0: 		if (fdi != NULL && FDI_HAS_EVENTS(fdi)) {
michael@0: 			int fd = epdp->ed_pending[i];
michael@0: 			reassociate(epdp, fdi, fd);
michael@0: 			epdp->ed_pending[i] = -1;
michael@0: 		}
michael@0: 	}
michael@0: 
michael@0: 	EVBASE_RELEASE_LOCK(base, th_base_lock);
michael@0: 
michael@0: 	res = port_getn(epdp->ed_port, pevtlist, EVENTS_PER_GETN,
michael@0: 	    (unsigned int *) &nevents, ts_p);
michael@0: 
michael@0: 	EVBASE_ACQUIRE_LOCK(base, th_base_lock);
michael@0: 
michael@0: 	if (res == -1) {
michael@0: 		if (errno == EINTR || errno == EAGAIN) {
michael@0: 			return (0);
michael@0: 		} else if (errno == ETIME) {
michael@0: 			if (nevents == 0)
michael@0: 				return (0);
michael@0: 		} else {
michael@0: 			event_warn("port_getn");
michael@0: 			return (-1);
michael@0: 		}
michael@0: 	}
michael@0: 
michael@0: 	event_debug(("%s: port_getn reports %d events", __func__, nevents));
michael@0: 
michael@0: 	for (i = 0; i < nevents; ++i) {
michael@0: 		struct fd_info *fdi;
michael@0: 		port_event_t *pevt = &pevtlist[i];
michael@0: 		int fd = (int) pevt->portev_object;
michael@0: 
michael@0: 		check_evportop(epdp);
michael@0: 		check_event(pevt);
michael@0: 		epdp->ed_pending[i] = fd;
michael@0: 
michael@0: 		/*
michael@0: 		 * Figure out what kind of event it was
michael@0: 		 * (because we have to pass this to the callback)
michael@0: 		 */
michael@0: 		res = 0;
michael@0: 		if (pevt->portev_events & (POLLERR|POLLHUP)) {
michael@0: 			res = EV_READ | EV_WRITE;
michael@0: 		} else {
michael@0: 			if (pevt->portev_events & POLLIN)
michael@0: 				res |= EV_READ;
michael@0: 			if (pevt->portev_events & POLLOUT)
michael@0: 				res |= EV_WRITE;
michael@0: 		}
michael@0: 
michael@0: 		/*
michael@0: 		 * Check for the error situations or a hangup situation
michael@0: 		 */
michael@0: 		if (pevt->portev_events & (POLLERR|POLLHUP|POLLNVAL))
michael@0: 			res |= EV_READ|EV_WRITE;
michael@0: 
michael@0: 		EVUTIL_ASSERT(epdp->ed_nevents > fd);
michael@0: 		fdi = &(epdp->ed_fds[fd]);
michael@0: 
michael@0: 		evmap_io_active(base, fd, res);
michael@0: 	} /* end of all events gotten */
michael@0: 
michael@0: 	check_evportop(epdp);
michael@0: 
michael@0: 	return (0);
michael@0: }
michael@0: 
michael@0: 
michael@0: /*
michael@0:  * Adds the given event (so that you will be notified when it happens via
michael@0:  * the callback function).
michael@0:  */
michael@0: 
michael@0: static int
michael@0: evport_add(struct event_base *base, int fd, short old, short events, void *p)
michael@0: {
michael@0: 	struct evport_data *evpd = base->evbase;
michael@0: 	struct fd_info *fdi;
michael@0: 	int factor;
michael@0: 	(void)p;
michael@0: 
michael@0: 	check_evportop(evpd);
michael@0: 
michael@0: 	/*
michael@0: 	 * If necessary, grow the file descriptor info table
michael@0: 	 */
michael@0: 
michael@0: 	factor = 1;
michael@0: 	while (fd >= factor * evpd->ed_nevents)
michael@0: 		factor *= 2;
michael@0: 
michael@0: 	if (factor > 1) {
michael@0: 		if (-1 == grow(evpd, factor)) {
michael@0: 			return (-1);
michael@0: 		}
michael@0: 	}
michael@0: 
michael@0: 	fdi = &evpd->ed_fds[fd];
michael@0: 	fdi->fdi_what |= events;
michael@0: 
michael@0: 	return reassociate(evpd, fdi, fd);
michael@0: }
michael@0: 
michael@0: /*
michael@0:  * Removes the given event from the list of events to wait for.
michael@0:  */
michael@0: 
michael@0: static int
michael@0: evport_del(struct event_base *base, int fd, short old, short events, void *p)
michael@0: {
michael@0: 	struct evport_data *evpd = base->evbase;
michael@0: 	struct fd_info *fdi;
michael@0: 	int i;
michael@0: 	int associated = 1;
michael@0: 	(void)p;
michael@0: 
michael@0: 	check_evportop(evpd);
michael@0: 
michael@0: 	if (evpd->ed_nevents < fd) {
michael@0: 		return (-1);
michael@0: 	}
michael@0: 
michael@0: 	for (i = 0; i < EVENTS_PER_GETN; ++i) {
michael@0: 		if (evpd->ed_pending[i] == fd) {
michael@0: 			associated = 0;
michael@0: 			break;
michael@0: 		}
michael@0: 	}
michael@0: 
michael@0: 	fdi = &evpd->ed_fds[fd];
michael@0: 	if (events & EV_READ)
michael@0: 		fdi->fdi_what &= ~EV_READ;
michael@0: 	if (events & EV_WRITE)
michael@0: 		fdi->fdi_what &= ~EV_WRITE;
michael@0: 
michael@0: 	if (associated) {
michael@0: 		if (!FDI_HAS_EVENTS(fdi) &&
michael@0: 		    port_dissociate(evpd->ed_port, PORT_SOURCE_FD, fd) == -1) {
michael@0: 			/*
michael@0: 			 * Ignore EBADFD error the fd could have been closed
michael@0: 			 * before event_del() was called.
michael@0: 			 */
michael@0: 			if (errno != EBADFD) {
michael@0: 				event_warn("port_dissociate");
michael@0: 				return (-1);
michael@0: 			}
michael@0: 		} else {
michael@0: 			if (FDI_HAS_EVENTS(fdi)) {
michael@0: 				return (reassociate(evpd, fdi, fd));
michael@0: 			}
michael@0: 		}
michael@0: 	} else {
michael@0: 		if ((fdi->fdi_what & (EV_READ|EV_WRITE)) == 0) {
michael@0: 			evpd->ed_pending[i] = -1;
michael@0: 		}
michael@0: 	}
michael@0: 	return 0;
michael@0: }
michael@0: 
michael@0: 
michael@0: static void
michael@0: evport_dealloc(struct event_base *base)
michael@0: {
michael@0: 	struct evport_data *evpd = base->evbase;
michael@0: 
michael@0: 	evsig_dealloc(base);
michael@0: 
michael@0: 	close(evpd->ed_port);
michael@0: 
michael@0: 	if (evpd->ed_fds)
michael@0: 		mm_free(evpd->ed_fds);
michael@0: 	mm_free(evpd);
michael@0: }