static int
kq_dispatch(struct event_base *base, struct timeval *tv)
{
struct kqop *kqop = base->evbase;
struct kevent *events = kqop->events;
struct kevent *changes;
struct timespec ts, *ts_p = NULL;
int i, n_changes, res;
if (tv != NULL) {
TIMEVAL_TO_TIMESPEC(tv, &ts);
ts_p = &ts;
}
/* Build "changes" from "base->changes" */
EVUTIL_ASSERT(kqop->changes);
n_changes = kq_build_changes_list(&base->changelist, kqop);
if (n_changes < 0)
return -1;
event_changelist_remove_all_(&base->changelist, base);
/* steal the changes array in case some broken code tries to call
* dispatch twice at once. */
changes = kqop->changes;
kqop->changes = NULL;
/* Make sure that 'events' is at least as long as the list of changes:
* otherwise errors in the changes can get reported as a -1 return
* value from kevent() rather than as EV_ERROR events in the events
* array.
*
* (We could instead handle -1 return values from kevent() by
* retrying with a smaller changes array or a larger events array,
* but this approach seems less risky for now.)
*/
if (kqop->events_size < n_changes) {
int new_size = kqop->events_size;
do {
new_size *= 2;
} while (new_size < n_changes);
kq_grow_events(kqop, new_size);
events = kqop->events;
}
EVBASE_RELEASE_LOCK(base, th_base_lock);
res = kevent(kqop->kq, changes, n_changes,
events, kqop->events_size, ts_p);
EVBASE_ACQUIRE_LOCK(base, th_base_lock);
EVUTIL_ASSERT(kqop->changes == NULL);
kqop->changes = changes;
if (res == -1) {
if (errno != EINTR) {
event_warn("kevent");
return (-1);
}
return (0);
}
event_debug(("%s: kevent reports %d", __func__, res));
for (i = 0; i < res; i++) {
int which = 0;
if (events[i].flags & EV_ERROR) {
switch (events[i].data) {
/* Can occur on delete if we are not currently
* watching any events on this fd. That can
* happen when the fd was closed and another
* file was opened with that fd. */
case ENOENT:
/* Can occur for reasons not fully understood
* on FreeBSD. */
case EINVAL:
continue;
/* Can occur on a delete if the fd is closed. */
case EBADF:
/* XXXX On NetBSD, we can also get EBADF if we
* try to add the write side of a pipe, but
* the read side has already been closed.
* Other BSDs call this situation 'EPIPE'. It
* would be good if we had a way to report
* this situation. */
continue;
/* These two can occur on an add if the fd was one side
* of a pipe, and the other side was closed. */
case EPERM:
case EPIPE:
/* Report read events, if we're listening for
* them, so that the user can learn about any
* add errors. (If the operation was a
* delete, then udata should be cleared.) */
if (events[i].udata) {
/* The operation was an add:
* report the error as a read. */
which |= EV_READ;
break;
} else {
/* The operation was a del:
* report nothing. */
continue;
}
/* Other errors shouldn't occur. */
default:
errno = events[i].data;
return (-1);
}
} else if (events[i].filter == EVFILT_READ) {
which |= EV_READ;
} else if (events[i].filter == EVFILT_WRITE) {
which |= EV_WRITE;
} else if (events[i].filter == EVFILT_SIGNAL) {
which |= EV_SIGNAL;
#ifdef EVFILT_USER
} else if (events[i].filter == EVFILT_USER) {
base->is_notify_pending = 0;
#endif
}
if (!which)
continue;
if (events[i].filter == EVFILT_SIGNAL) {
evmap_signal_active_(base, events[i].ident, 1);
} else {
evmap_io_active_(base, events[i].ident, which | EV_ET);
}
}
if (res == kqop->events_size) {
/* We used all the events space that we have. Maybe we should
make it bigger. */
kq_grow_events(kqop, kqop->events_size * 2);
}
return (0);
}
static int
epoll_dispatch(struct event_base *base, struct timeval *tv)
{
struct epollop *epollop = base->evbase;
struct epoll_event *events = epollop->events;
int i, res;
long timeout = -1;
#ifdef USING_TIMERFD
if (epollop->timerfd >= 0) {
struct itimerspec is;
is.it_interval.tv_sec = 0;
is.it_interval.tv_nsec = 0;
if (tv == NULL) {
/* No timeout; disarm the timer. */
is.it_value.tv_sec = 0;
is.it_value.tv_nsec = 0;
} else {
if (tv->tv_sec == 0 && tv->tv_usec == 0) {
/* we need to exit immediately; timerfd can't
* do that. */
timeout = 0;
}
is.it_value.tv_sec = tv->tv_sec;
is.it_value.tv_nsec = tv->tv_usec * 1000;
}
/* TODO: we could avoid unnecessary syscalls here by only
calling timerfd_settime when the top timeout changes, or
when we're called with a different timeval.
*/
if (timerfd_settime(epollop->timerfd, 0, &is, NULL) < 0) {
event_warn("timerfd_settime");
}
} else
#endif
if (tv != NULL) {
timeout = evutil_tv_to_msec_(tv);
if (timeout < 0 || timeout > MAX_EPOLL_TIMEOUT_MSEC) {
/* Linux kernels can wait forever if the timeout is
* too big; see comment on MAX_EPOLL_TIMEOUT_MSEC. */
timeout = MAX_EPOLL_TIMEOUT_MSEC;
}
}
epoll_apply_changes(base);
event_changelist_remove_all_(&base->changelist, base);
EVBASE_RELEASE_LOCK(base, th_base_lock);
res = epoll_wait(epollop->epfd, events, epollop->nevents, timeout);
EVBASE_ACQUIRE_LOCK(base, th_base_lock);
if (res == -1) {
if (errno != EINTR) {
event_warn("epoll_wait");
return (-1);
}
return (0);
}
event_debug(("%s: epoll_wait reports %d", __func__, res));
EVUTIL_ASSERT(res <= epollop->nevents);
for (i = 0; i < res; i++) {
int what = events[i].events;
short ev = 0;
#ifdef USING_TIMERFD
if (events[i].data.fd == epollop->timerfd)
continue;
#endif
if (what & (EPOLLHUP|EPOLLERR)) {
ev = EV_READ | EV_WRITE;
} else {
if (what & EPOLLIN)
ev |= EV_READ;
if (what & EPOLLOUT)
ev |= EV_WRITE;
if (what & EPOLLRDHUP)
ev |= EV_CLOSED;
}
if (!ev)
continue;
evmap_io_active_(base, events[i].data.fd, ev | EV_ET);
}
if (res == epollop->nevents && epollop->nevents < MAX_NEVENT) {
/* We used all of the event space this time. We should
be ready for more events next time. */
int new_nevents = epollop->nevents * 2;
struct epoll_event *new_events;
new_events = mm_realloc(epollop->events,
new_nevents * sizeof(struct epoll_event));
if (new_events) {
epollop->events = new_events;
epollop->nevents = new_nevents;
}
}
return (0);
}
