/* $OpenBSD: kqueue.c,v 1.5 2002/07/10 14:41:31 art Exp $ */ /* * Copyright 2000-2007 Niels Provos * Copyright 2007-2012 Niels Provos and Nick Mathewson * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. The name of the author may not be used to endorse or promote products * derived from this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include "event2/event-config.h" #include "evconfig-private.h" #ifdef EVENT__HAVE_KQUEUE #include #ifdef EVENT__HAVE_SYS_TIME_H #include #endif #include #include #include #include #include #include #include #include #include #ifdef EVENT__HAVE_INTTYPES_H #include #endif /* Some platforms apparently define the udata field of struct kevent as * intptr_t, whereas others define it as void*. There doesn't seem to be an * easy way to tell them apart via autoconf, so we need to use OS macros. */ #if defined(EVENT__HAVE_INTTYPES_H) && !defined(__OpenBSD__) && !defined(__FreeBSD__) && !defined(__darwin__) && !defined(__APPLE__) && !defined(__CloudABI__) #define PTR_TO_UDATA(x) ((intptr_t)(x)) #define INT_TO_UDATA(x) ((intptr_t)(x)) #else #define PTR_TO_UDATA(x) (x) #define INT_TO_UDATA(x) ((void*)(x)) #endif #include "event-internal.h" #include "log-internal.h" #include "evmap-internal.h" #include "event2/thread.h" #include "event2/util.h" #include "evthread-internal.h" #include "changelist-internal.h" #include "kqueue-internal.h" #define NEVENT 64 struct kqop { struct kevent *changes; int changes_size; struct kevent *events; int events_size; int kq; int notify_event_added; pid_t pid; }; static void kqop_free(struct kqop *kqop); static void *kq_init(struct event_base *); static int kq_sig_add(struct event_base *, int, short, short, void *); static int kq_sig_del(struct event_base *, int, short, short, void *); static int kq_dispatch(struct event_base *, struct timeval *); static void kq_dealloc(struct event_base *); const struct eventop kqops = { "kqueue", kq_init, event_changelist_add_, event_changelist_del_, kq_dispatch, kq_dealloc, 1 /* need reinit */, EV_FEATURE_ET|EV_FEATURE_O1|EV_FEATURE_FDS, EVENT_CHANGELIST_FDINFO_SIZE }; static const struct eventop kqsigops = { "kqueue_signal", NULL, kq_sig_add, kq_sig_del, NULL, NULL, 1 /* need reinit */, 0, 0 }; static void * kq_init(struct event_base *base) { int kq = -1; struct kqop *kqueueop = NULL; if (!(kqueueop = mm_calloc(1, sizeof(struct kqop)))) return (NULL); /* Initialize the kernel queue */ if ((kq = kqueue()) == -1) { event_warn("kqueue"); goto err; } kqueueop->kq = kq; kqueueop->pid = getpid(); /* Initialize fields */ kqueueop->changes = mm_calloc(NEVENT, sizeof(struct kevent)); if (kqueueop->changes == NULL) goto err; kqueueop->events = mm_calloc(NEVENT, sizeof(struct kevent)); if (kqueueop->events == NULL) goto err; kqueueop->events_size = kqueueop->changes_size = NEVENT; /* Check for Mac OS X kqueue bug. */ memset(&kqueueop->changes[0], 0, sizeof kqueueop->changes[0]); kqueueop->changes[0].ident = -1; kqueueop->changes[0].filter = EVFILT_READ; kqueueop->changes[0].flags = EV_ADD; /* * If kqueue works, then kevent will succeed, and it will * stick an error in events[0]. If kqueue is broken, then * kevent will fail. */ if (kevent(kq, kqueueop->changes, 1, kqueueop->events, NEVENT, NULL) != 1 || (int)kqueueop->events[0].ident != -1 || !(kqueueop->events[0].flags & EV_ERROR)) { event_warn("%s: detected broken kqueue; not using.", __func__); goto err; } base->evsigsel = &kqsigops; return (kqueueop); err: if (kqueueop) kqop_free(kqueueop); return (NULL); } #define ADD_UDATA 0x30303 static void kq_setup_kevent(struct kevent *out, evutil_socket_t fd, int filter, short change) { memset(out, 0, sizeof(struct kevent)); out->ident = fd; out->filter = filter; if (change & EV_CHANGE_ADD) { out->flags = EV_ADD; /* We set a magic number here so that we can tell 'add' * errors from 'del' errors. */ out->udata = INT_TO_UDATA(ADD_UDATA); if (change & EV_ET) out->flags |= EV_CLEAR; #ifdef NOTE_EOF /* Make it behave like select() and poll() */ if (filter == EVFILT_READ) out->fflags = NOTE_EOF; #endif } else { EVUTIL_ASSERT(change & EV_CHANGE_DEL); out->flags = EV_DELETE; } } static int kq_build_changes_list(const struct event_changelist *changelist, struct kqop *kqop) { int i; int n_changes = 0; for (i = 0; i < changelist->n_changes; ++i) { struct event_change *in_ch = &changelist->changes[i]; struct kevent *out_ch; if (n_changes >= kqop->changes_size - 1) { int newsize; struct kevent *newchanges; if (kqop->changes_size > INT_MAX / 2 || (size_t)kqop->changes_size * 2 > EV_SIZE_MAX / sizeof(struct kevent)) { event_warnx("%s: int overflow", __func__); return (-1); } newsize = kqop->changes_size * 2; newchanges = mm_realloc(kqop->changes, newsize * sizeof(struct kevent)); if (newchanges == NULL) { event_warn("%s: realloc", __func__); return (-1); } kqop->changes = newchanges; kqop->changes_size = newsize; } if (in_ch->read_change) { out_ch = &kqop->changes[n_changes++]; kq_setup_kevent(out_ch, in_ch->fd, EVFILT_READ, in_ch->read_change); } if (in_ch->write_change) { out_ch = &kqop->changes[n_changes++]; kq_setup_kevent(out_ch, in_ch->fd, EVFILT_WRITE, in_ch->write_change); } } return n_changes; } static int kq_grow_events(struct kqop *kqop, size_t new_size) { struct kevent *newresult; newresult = mm_realloc(kqop->events, new_size * sizeof(struct kevent)); if (newresult) { kqop->events = newresult; kqop->events_size = new_size; return 0; } else { return -1; } } 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) { ts.tv_sec = tv->tv_sec; ts.tv_nsec = tv->tv_usec * 1000; 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; #if defined(__FreeBSD__) /* * This currently occurs if an FD is closed * before the EV_DELETE makes it out via kevent(). * The FreeBSD capabilities code sees the blank * capability set and rejects the request to * modify an event. * * To be strictly correct - when an FD is closed, * all the registered events are also removed. * Queuing EV_DELETE to a closed FD is wrong. * The event(s) should just be deleted from * the pending changelist. */ case ENOTCAPABLE: continue; #endif /* 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 void kqop_free(struct kqop *kqop) { if (kqop->changes) mm_free(kqop->changes); if (kqop->events) mm_free(kqop->events); if (kqop->kq >= 0 && kqop->pid == getpid()) close(kqop->kq); memset(kqop, 0, sizeof(struct kqop)); mm_free(kqop); } static void kq_dealloc(struct event_base *base) { struct kqop *kqop = base->evbase; evsig_dealloc_(base); kqop_free(kqop); } /* signal handling */ static int kq_sig_add(struct event_base *base, int nsignal, short old, short events, void *p) { struct kqop *kqop = base->evbase; struct kevent kev; struct timespec timeout = { 0, 0 }; (void)p; EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG); memset(&kev, 0, sizeof(kev)); kev.ident = nsignal; kev.filter = EVFILT_SIGNAL; kev.flags = EV_ADD; /* Be ready for the signal if it is sent any * time between now and the next call to * kq_dispatch. */ if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) return (-1); /* We can set the handler for most signals to SIG_IGN and * still have them reported to us in the queue. However, * if the handler for SIGCHLD is SIG_IGN, the system reaps * zombie processes for us, and we don't get any notification. * This appears to be the only signal with this quirk. */ if (evsig_set_handler_(base, nsignal, nsignal == SIGCHLD ? SIG_DFL : SIG_IGN) == -1) return (-1); return (0); } static int kq_sig_del(struct event_base *base, int nsignal, short old, short events, void *p) { struct kqop *kqop = base->evbase; struct kevent kev; struct timespec timeout = { 0, 0 }; (void)p; EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG); memset(&kev, 0, sizeof(kev)); kev.ident = nsignal; kev.filter = EVFILT_SIGNAL; kev.flags = EV_DELETE; /* Because we insert signal events * immediately, we need to delete them * immediately, too */ if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) return (-1); if (evsig_restore_handler_(base, nsignal) == -1) return (-1); return (0); } /* OSX 10.6 and FreeBSD 8.1 add support for EVFILT_USER, which we can use * to wake up the event loop from another thread. */ /* Magic number we use for our filter ID. */ #define NOTIFY_IDENT 42 int event_kq_add_notify_event_(struct event_base *base) { struct kqop *kqop = base->evbase; #if defined(EVFILT_USER) && defined(NOTE_TRIGGER) struct kevent kev; struct timespec timeout = { 0, 0 }; #endif if (kqop->notify_event_added) return 0; #if defined(EVFILT_USER) && defined(NOTE_TRIGGER) memset(&kev, 0, sizeof(kev)); kev.ident = NOTIFY_IDENT; kev.filter = EVFILT_USER; kev.flags = EV_ADD | EV_CLEAR; if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) { event_warn("kevent: adding EVFILT_USER event"); return -1; } kqop->notify_event_added = 1; return 0; #else return -1; #endif } int event_kq_notify_base_(struct event_base *base) { struct kqop *kqop = base->evbase; #if defined(EVFILT_USER) && defined(NOTE_TRIGGER) struct kevent kev; struct timespec timeout = { 0, 0 }; #endif if (! kqop->notify_event_added) return -1; #if defined(EVFILT_USER) && defined(NOTE_TRIGGER) memset(&kev, 0, sizeof(kev)); kev.ident = NOTIFY_IDENT; kev.filter = EVFILT_USER; kev.fflags = NOTE_TRIGGER; if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) { event_warn("kevent: triggering EVFILT_USER event"); return -1; } return 0; #else return -1; #endif } #endif /* EVENT__HAVE_KQUEUE */