2 * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
3 * Copyright 2004 John-Mark Gurney <jmg@FreeBSD.org>
4 * Copyright (c) 2009 Apple, Inc.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
32 #include "opt_ktrace.h"
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
38 #include <sys/mutex.h>
40 #include <sys/malloc.h>
41 #include <sys/unistd.h>
43 #include <sys/filedesc.h>
44 #include <sys/filio.h>
45 #include <sys/fcntl.h>
46 #include <sys/kthread.h>
47 #include <sys/selinfo.h>
48 #include <sys/queue.h>
49 #include <sys/event.h>
50 #include <sys/eventvar.h>
52 #include <sys/protosw.h>
53 #include <sys/sigio.h>
54 #include <sys/signalvar.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
58 #include <sys/sysctl.h>
59 #include <sys/sysproto.h>
60 #include <sys/syscallsubr.h>
61 #include <sys/taskqueue.h>
64 #include <sys/ktrace.h>
69 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
72 * This lock is used if multiple kq locks are required. This possibly
73 * should be made into a per proc lock.
75 static struct mtx kq_global;
76 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
77 #define KQ_GLOBAL_LOCK(lck, haslck) do { \
82 #define KQ_GLOBAL_UNLOCK(lck, haslck) do { \
88 TASKQUEUE_DEFINE_THREAD(kqueue);
90 static int kevent_copyout(void *arg, struct kevent *kevp, int count);
91 static int kevent_copyin(void *arg, struct kevent *kevp, int count);
92 static int kqueue_register(struct kqueue *kq, struct kevent *kev,
93 struct thread *td, int waitok);
94 static int kqueue_acquire(struct file *fp, struct kqueue **kqp);
95 static void kqueue_release(struct kqueue *kq, int locked);
96 static int kqueue_expand(struct kqueue *kq, struct filterops *fops,
97 uintptr_t ident, int waitok);
98 static void kqueue_task(void *arg, int pending);
99 static int kqueue_scan(struct kqueue *kq, int maxevents,
100 struct kevent_copyops *k_ops,
101 const struct timespec *timeout,
102 struct kevent *keva, struct thread *td);
103 static void kqueue_wakeup(struct kqueue *kq);
104 static struct filterops *kqueue_fo_find(int filt);
105 static void kqueue_fo_release(int filt);
107 static fo_rdwr_t kqueue_read;
108 static fo_rdwr_t kqueue_write;
109 static fo_truncate_t kqueue_truncate;
110 static fo_ioctl_t kqueue_ioctl;
111 static fo_poll_t kqueue_poll;
112 static fo_kqfilter_t kqueue_kqfilter;
113 static fo_stat_t kqueue_stat;
114 static fo_close_t kqueue_close;
116 static struct fileops kqueueops = {
117 .fo_read = kqueue_read,
118 .fo_write = kqueue_write,
119 .fo_truncate = kqueue_truncate,
120 .fo_ioctl = kqueue_ioctl,
121 .fo_poll = kqueue_poll,
122 .fo_kqfilter = kqueue_kqfilter,
123 .fo_stat = kqueue_stat,
124 .fo_close = kqueue_close,
127 static int knote_attach(struct knote *kn, struct kqueue *kq);
128 static void knote_drop(struct knote *kn, struct thread *td);
129 static void knote_enqueue(struct knote *kn);
130 static void knote_dequeue(struct knote *kn);
131 static void knote_init(void);
132 static struct knote *knote_alloc(int waitok);
133 static void knote_free(struct knote *kn);
135 static void filt_kqdetach(struct knote *kn);
136 static int filt_kqueue(struct knote *kn, long hint);
137 static int filt_procattach(struct knote *kn);
138 static void filt_procdetach(struct knote *kn);
139 static int filt_proc(struct knote *kn, long hint);
140 static int filt_fileattach(struct knote *kn);
141 static void filt_timerexpire(void *knx);
142 static int filt_timerattach(struct knote *kn);
143 static void filt_timerdetach(struct knote *kn);
144 static int filt_timer(struct knote *kn, long hint);
145 static int filt_userattach(struct knote *kn);
146 static void filt_userdetach(struct knote *kn);
147 static int filt_user(struct knote *kn, long hint);
148 static void filt_usertouch(struct knote *kn, struct kevent *kev,
151 static struct filterops file_filtops =
152 { 1, filt_fileattach, NULL, NULL };
153 static struct filterops kqread_filtops =
154 { 1, NULL, filt_kqdetach, filt_kqueue };
155 /* XXX - move to kern_proc.c? */
156 static struct filterops proc_filtops =
157 { 0, filt_procattach, filt_procdetach, filt_proc };
158 static struct filterops timer_filtops =
159 { 0, filt_timerattach, filt_timerdetach, filt_timer };
160 static struct filterops user_filtops = {
161 .f_attach = filt_userattach,
162 .f_detach = filt_userdetach,
163 .f_event = filt_user,
164 .f_touch = filt_usertouch,
167 static uma_zone_t knote_zone;
168 static int kq_ncallouts = 0;
169 static int kq_calloutmax = (4 * 1024);
170 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
171 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
173 /* XXX - ensure not KN_INFLUX?? */
174 #define KNOTE_ACTIVATE(kn, islock) do { \
176 mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED); \
178 KQ_LOCK((kn)->kn_kq); \
179 (kn)->kn_status |= KN_ACTIVE; \
180 if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \
181 knote_enqueue((kn)); \
183 KQ_UNLOCK((kn)->kn_kq); \
185 #define KQ_LOCK(kq) do { \
186 mtx_lock(&(kq)->kq_lock); \
188 #define KQ_FLUX_WAKEUP(kq) do { \
189 if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) { \
190 (kq)->kq_state &= ~KQ_FLUXWAIT; \
194 #define KQ_UNLOCK_FLUX(kq) do { \
195 KQ_FLUX_WAKEUP(kq); \
196 mtx_unlock(&(kq)->kq_lock); \
198 #define KQ_UNLOCK(kq) do { \
199 mtx_unlock(&(kq)->kq_lock); \
201 #define KQ_OWNED(kq) do { \
202 mtx_assert(&(kq)->kq_lock, MA_OWNED); \
204 #define KQ_NOTOWNED(kq) do { \
205 mtx_assert(&(kq)->kq_lock, MA_NOTOWNED); \
207 #define KN_LIST_LOCK(kn) do { \
208 if (kn->kn_knlist != NULL) \
209 kn->kn_knlist->kl_lock(kn->kn_knlist->kl_lockarg); \
211 #define KN_LIST_UNLOCK(kn) do { \
212 if (kn->kn_knlist != NULL) \
213 kn->kn_knlist->kl_unlock(kn->kn_knlist->kl_lockarg); \
215 #define KNL_ASSERT_LOCK(knl, islocked) do { \
217 KNL_ASSERT_LOCKED(knl); \
219 KNL_ASSERT_UNLOCKED(knl); \
222 #define KNL_ASSERT_LOCKED(knl) do { \
223 knl->kl_assert_locked((knl)->kl_lockarg); \
225 #define KNL_ASSERT_UNLOCKED(knl) do { \
226 knl->kl_assert_unlocked((knl)->kl_lockarg); \
228 #else /* !INVARIANTS */
229 #define KNL_ASSERT_LOCKED(knl) do {} while(0)
230 #define KNL_ASSERT_UNLOCKED(knl) do {} while (0)
231 #endif /* INVARIANTS */
233 #define KN_HASHSIZE 64 /* XXX should be tunable */
234 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask))
237 filt_nullattach(struct knote *kn)
243 struct filterops null_filtops =
244 { 0, filt_nullattach, NULL, NULL };
246 /* XXX - make SYSINIT to add these, and move into respective modules. */
247 extern struct filterops sig_filtops;
248 extern struct filterops fs_filtops;
251 * Table for for all system-defined filters.
253 static struct mtx filterops_lock;
254 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
257 struct filterops *for_fop;
259 } sysfilt_ops[EVFILT_SYSCOUNT] = {
260 { &file_filtops }, /* EVFILT_READ */
261 { &file_filtops }, /* EVFILT_WRITE */
262 { &null_filtops }, /* EVFILT_AIO */
263 { &file_filtops }, /* EVFILT_VNODE */
264 { &proc_filtops }, /* EVFILT_PROC */
265 { &sig_filtops }, /* EVFILT_SIGNAL */
266 { &timer_filtops }, /* EVFILT_TIMER */
267 { &null_filtops }, /* former EVFILT_NETDEV */
268 { &fs_filtops }, /* EVFILT_FS */
269 { &null_filtops }, /* EVFILT_LIO */
270 { &user_filtops }, /* EVFILT_USER */
274 * Simple redirection for all cdevsw style objects to call their fo_kqfilter
278 filt_fileattach(struct knote *kn)
281 return (fo_kqfilter(kn->kn_fp, kn));
286 kqueue_kqfilter(struct file *fp, struct knote *kn)
288 struct kqueue *kq = kn->kn_fp->f_data;
290 if (kn->kn_filter != EVFILT_READ)
293 kn->kn_status |= KN_KQUEUE;
294 kn->kn_fop = &kqread_filtops;
295 knlist_add(&kq->kq_sel.si_note, kn, 0);
301 filt_kqdetach(struct knote *kn)
303 struct kqueue *kq = kn->kn_fp->f_data;
305 knlist_remove(&kq->kq_sel.si_note, kn, 0);
310 filt_kqueue(struct knote *kn, long hint)
312 struct kqueue *kq = kn->kn_fp->f_data;
314 kn->kn_data = kq->kq_count;
315 return (kn->kn_data > 0);
318 /* XXX - move to kern_proc.c? */
320 filt_procattach(struct knote *kn)
327 p = pfind(kn->kn_id);
328 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
329 p = zpfind(kn->kn_id);
331 } else if (p != NULL && (p->p_flag & P_WEXIT)) {
337 if ((error = p_cansee(curthread, p))) {
342 kn->kn_ptr.p_proc = p;
343 kn->kn_flags |= EV_CLEAR; /* automatically set */
346 * internal flag indicating registration done by kernel
348 if (kn->kn_flags & EV_FLAG1) {
349 kn->kn_data = kn->kn_sdata; /* ppid */
350 kn->kn_fflags = NOTE_CHILD;
351 kn->kn_flags &= ~EV_FLAG1;
355 knlist_add(&p->p_klist, kn, 1);
358 * Immediately activate any exit notes if the target process is a
359 * zombie. This is necessary to handle the case where the target
360 * process, e.g. a child, dies before the kevent is registered.
362 if (immediate && filt_proc(kn, NOTE_EXIT))
363 KNOTE_ACTIVATE(kn, 0);
371 * The knote may be attached to a different process, which may exit,
372 * leaving nothing for the knote to be attached to. So when the process
373 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
374 * it will be deleted when read out. However, as part of the knote deletion,
375 * this routine is called, so a check is needed to avoid actually performing
376 * a detach, because the original process does not exist any more.
378 /* XXX - move to kern_proc.c? */
380 filt_procdetach(struct knote *kn)
384 p = kn->kn_ptr.p_proc;
385 knlist_remove(&p->p_klist, kn, 0);
386 kn->kn_ptr.p_proc = NULL;
389 /* XXX - move to kern_proc.c? */
391 filt_proc(struct knote *kn, long hint)
393 struct proc *p = kn->kn_ptr.p_proc;
397 * mask off extra data
399 event = (u_int)hint & NOTE_PCTRLMASK;
402 * if the user is interested in this event, record it.
404 if (kn->kn_sfflags & event)
405 kn->kn_fflags |= event;
408 * process is gone, so flag the event as finished.
410 if (event == NOTE_EXIT) {
411 if (!(kn->kn_status & KN_DETACHED))
412 knlist_remove_inevent(&p->p_klist, kn);
413 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
414 kn->kn_ptr.p_proc = NULL;
415 if (kn->kn_fflags & NOTE_EXIT)
416 kn->kn_data = p->p_xstat;
417 if (kn->kn_fflags == 0)
418 kn->kn_flags |= EV_DROP;
422 return (kn->kn_fflags != 0);
426 * Called when the process forked. It mostly does the same as the
427 * knote(), activating all knotes registered to be activated when the
428 * process forked. Additionally, for each knote attached to the
429 * parent, check whether user wants to track the new process. If so
430 * attach a new knote to it, and immediately report an event with the
434 knote_fork(struct knlist *list, int pid)
443 list->kl_lock(list->kl_lockarg);
445 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
446 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX)
450 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
456 * The same as knote(), activate the event.
458 if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
459 kn->kn_status |= KN_HASKQLOCK;
460 if (kn->kn_fop->f_event(kn, NOTE_FORK))
461 KNOTE_ACTIVATE(kn, 1);
462 kn->kn_status &= ~KN_HASKQLOCK;
468 * The NOTE_TRACK case. In addition to the activation
469 * of the event, we need to register new event to
470 * track the child. Drop the locks in preparation for
471 * the call to kqueue_register().
473 kn->kn_status |= KN_INFLUX;
475 list->kl_unlock(list->kl_lockarg);
478 * Activate existing knote and register a knote with
482 kev.filter = kn->kn_filter;
483 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
484 kev.fflags = kn->kn_sfflags;
485 kev.data = kn->kn_id; /* parent */
486 kev.udata = kn->kn_kevent.udata;/* preserve udata */
487 error = kqueue_register(kq, &kev, NULL, 0);
489 kn->kn_fflags |= NOTE_TRACKERR;
490 if (kn->kn_fop->f_event(kn, NOTE_FORK))
491 KNOTE_ACTIVATE(kn, 0);
493 kn->kn_status &= ~KN_INFLUX;
495 list->kl_lock(list->kl_lockarg);
497 list->kl_unlock(list->kl_lockarg);
501 * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
502 * interval timer support code.
505 timertoticks(intptr_t data)
510 tv.tv_sec = data / 1000;
511 tv.tv_usec = (data % 1000) * 1000;
512 tticks = tvtohz(&tv);
518 filt_timerexpire(void *knx)
520 struct knote *kn = knx;
521 struct callout *calloutp;
524 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */
527 * timertoticks() uses tvtohz() which always adds 1 to allow
528 * for the time until the next clock interrupt being strictly
529 * less than 1 clock tick. We don't want that here since we
530 * want to appear to be in sync with the clock interrupt even
531 * when we're delayed.
533 if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
534 calloutp = (struct callout *)kn->kn_hook;
535 callout_reset_curcpu(calloutp, timertoticks(kn->kn_sdata) - 1,
536 filt_timerexpire, kn);
541 * data contains amount of time to sleep, in milliseconds
544 filt_timerattach(struct knote *kn)
546 struct callout *calloutp;
548 atomic_add_int(&kq_ncallouts, 1);
550 if (kq_ncallouts >= kq_calloutmax) {
551 atomic_add_int(&kq_ncallouts, -1);
555 kn->kn_flags |= EV_CLEAR; /* automatically set */
556 kn->kn_status &= ~KN_DETACHED; /* knlist_add usually sets it */
557 calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
558 callout_init(calloutp, CALLOUT_MPSAFE);
559 kn->kn_hook = calloutp;
560 callout_reset_curcpu(calloutp, timertoticks(kn->kn_sdata),
561 filt_timerexpire, kn);
567 filt_timerdetach(struct knote *kn)
569 struct callout *calloutp;
571 calloutp = (struct callout *)kn->kn_hook;
572 callout_drain(calloutp);
573 free(calloutp, M_KQUEUE);
574 atomic_add_int(&kq_ncallouts, -1);
575 kn->kn_status |= KN_DETACHED; /* knlist_remove usually clears it */
579 filt_timer(struct knote *kn, long hint)
582 return (kn->kn_data != 0);
586 filt_userattach(struct knote *kn)
590 * EVFILT_USER knotes are not attached to anything in the kernel.
593 if (kn->kn_fflags & NOTE_TRIGGER)
601 filt_userdetach(__unused struct knote *kn)
605 * EVFILT_USER knotes are not attached to anything in the kernel.
610 filt_user(struct knote *kn, __unused long hint)
613 return (kn->kn_hookid);
617 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
623 if (kev->fflags & NOTE_TRIGGER)
626 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
627 kev->fflags &= NOTE_FFLAGSMASK;
633 kn->kn_sfflags &= kev->fflags;
637 kn->kn_sfflags |= kev->fflags;
641 kn->kn_sfflags = kev->fflags;
645 /* XXX Return error? */
648 kn->kn_sdata = kev->data;
649 if (kev->flags & EV_CLEAR) {
657 *kev = kn->kn_kevent;
658 kev->fflags = kn->kn_sfflags;
659 kev->data = kn->kn_sdata;
660 if (kn->kn_flags & EV_CLEAR) {
668 panic("filt_usertouch() - invalid type (%ld)", type);
674 kqueue(struct thread *td, struct kqueue_args *uap)
676 struct filedesc *fdp;
681 fdp = td->td_proc->p_fd;
682 error = falloc(td, &fp, &fd);
686 /* An extra reference on `nfp' has been held for us by falloc(). */
687 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
688 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
689 TAILQ_INIT(&kq->kq_head);
691 knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
692 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
695 SLIST_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
696 FILEDESC_XUNLOCK(fdp);
698 finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
701 td->td_retval[0] = fd;
706 #ifndef _SYS_SYSPROTO_H_
709 const struct kevent *changelist;
711 struct kevent *eventlist;
713 const struct timespec *timeout;
717 kevent(struct thread *td, struct kevent_args *uap)
719 struct timespec ts, *tsp;
720 struct kevent_copyops k_ops = { uap,
727 struct uio *ktruioin = NULL;
728 struct uio *ktruioout = NULL;
731 if (uap->timeout != NULL) {
732 error = copyin(uap->timeout, &ts, sizeof(ts));
740 if (KTRPOINT(td, KTR_GENIO)) {
741 ktriov.iov_base = uap->changelist;
742 ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
743 ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
744 .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
746 ktruioin = cloneuio(&ktruio);
747 ktriov.iov_base = uap->eventlist;
748 ktriov.iov_len = uap->nevents * sizeof(struct kevent);
749 ktruioout = cloneuio(&ktruio);
753 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
757 if (ktruioin != NULL) {
758 ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
759 ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
760 ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
761 ktrgenio(uap->fd, UIO_READ, ktruioout, error);
769 * Copy 'count' items into the destination list pointed to by uap->eventlist.
772 kevent_copyout(void *arg, struct kevent *kevp, int count)
774 struct kevent_args *uap;
777 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
778 uap = (struct kevent_args *)arg;
780 error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
782 uap->eventlist += count;
787 * Copy 'count' items from the list pointed to by uap->changelist.
790 kevent_copyin(void *arg, struct kevent *kevp, int count)
792 struct kevent_args *uap;
795 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
796 uap = (struct kevent_args *)arg;
798 error = copyin(uap->changelist, kevp, count * sizeof *kevp);
800 uap->changelist += count;
805 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
806 struct kevent_copyops *k_ops, const struct timespec *timeout)
808 struct kevent keva[KQ_NEVENTS];
809 struct kevent *kevp, *changes;
812 int i, n, nerrors, error;
814 if ((error = fget(td, fd, &fp)) != 0)
816 if ((error = kqueue_acquire(fp, &kq)) != 0)
821 while (nchanges > 0) {
822 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
823 error = k_ops->k_copyin(k_ops->arg, keva, n);
827 for (i = 0; i < n; i++) {
831 kevp->flags &= ~EV_SYSFLAGS;
832 error = kqueue_register(kq, kevp, td, 1);
833 if (error || (kevp->flags & EV_RECEIPT)) {
835 kevp->flags = EV_ERROR;
837 (void) k_ops->k_copyout(k_ops->arg,
849 td->td_retval[0] = nerrors;
854 error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
856 kqueue_release(kq, 0);
863 kqueue_add_filteropts(int filt, struct filterops *filtops)
868 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
870 "trying to add a filterop that is out of range: %d is beyond %d\n",
871 ~filt, EVFILT_SYSCOUNT);
874 mtx_lock(&filterops_lock);
875 if (sysfilt_ops[~filt].for_fop != &null_filtops &&
876 sysfilt_ops[~filt].for_fop != NULL)
879 sysfilt_ops[~filt].for_fop = filtops;
880 sysfilt_ops[~filt].for_refcnt = 0;
882 mtx_unlock(&filterops_lock);
888 kqueue_del_filteropts(int filt)
893 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
896 mtx_lock(&filterops_lock);
897 if (sysfilt_ops[~filt].for_fop == &null_filtops ||
898 sysfilt_ops[~filt].for_fop == NULL)
900 else if (sysfilt_ops[~filt].for_refcnt != 0)
903 sysfilt_ops[~filt].for_fop = &null_filtops;
904 sysfilt_ops[~filt].for_refcnt = 0;
906 mtx_unlock(&filterops_lock);
911 static struct filterops *
912 kqueue_fo_find(int filt)
915 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
918 mtx_lock(&filterops_lock);
919 sysfilt_ops[~filt].for_refcnt++;
920 if (sysfilt_ops[~filt].for_fop == NULL)
921 sysfilt_ops[~filt].for_fop = &null_filtops;
922 mtx_unlock(&filterops_lock);
924 return sysfilt_ops[~filt].for_fop;
928 kqueue_fo_release(int filt)
931 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
934 mtx_lock(&filterops_lock);
935 KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
936 ("filter object refcount not valid on release"));
937 sysfilt_ops[~filt].for_refcnt--;
938 mtx_unlock(&filterops_lock);
942 * A ref to kq (obtained via kqueue_acquire) must be held. waitok will
943 * influence if memory allocation should wait. Make sure it is 0 if you
947 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
949 struct filterops *fops;
951 struct knote *kn, *tkn;
952 int error, filt, event;
961 fops = kqueue_fo_find(filt);
965 tkn = knote_alloc(waitok); /* prevent waiting with locks */
969 KASSERT(td != NULL, ("td is NULL"));
970 error = fget(td, kev->ident, &fp);
974 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
975 kev->ident, 0) != 0) {
979 error = kqueue_expand(kq, fops, kev->ident, waitok);
985 if (fp->f_type == DTYPE_KQUEUE) {
987 * if we add some inteligence about what we are doing,
988 * we should be able to support events on ourselves.
989 * We need to know when we are doing this to prevent
990 * getting both the knlist lock and the kq lock since
991 * they are the same thing.
993 if (fp->f_data == kq) {
998 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1002 if (kev->ident < kq->kq_knlistsize) {
1003 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1004 if (kev->filter == kn->kn_filter)
1008 if ((kev->flags & EV_ADD) == EV_ADD)
1009 kqueue_expand(kq, fops, kev->ident, waitok);
1012 if (kq->kq_knhashmask != 0) {
1015 list = &kq->kq_knhash[
1016 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1017 SLIST_FOREACH(kn, list, kn_link)
1018 if (kev->ident == kn->kn_id &&
1019 kev->filter == kn->kn_filter)
1024 /* knote is in the process of changing, wait for it to stablize. */
1025 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1026 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1027 kq->kq_state |= KQ_FLUXWAIT;
1028 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1037 * kn now contains the matching knote, or NULL if no match
1040 if (kev->flags & EV_ADD) {
1052 * apply reference counts to knote structure, and
1053 * do not release it at the end of this routine.
1058 kn->kn_sfflags = kev->fflags;
1059 kn->kn_sdata = kev->data;
1062 kn->kn_kevent = *kev;
1063 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1064 EV_ENABLE | EV_DISABLE);
1065 kn->kn_status = KN_INFLUX|KN_DETACHED;
1067 error = knote_attach(kn, kq);
1074 if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1081 /* No matching knote and the EV_ADD flag is not set. */
1088 if (kev->flags & EV_DELETE) {
1089 kn->kn_status |= KN_INFLUX;
1091 if (!(kn->kn_status & KN_DETACHED))
1092 kn->kn_fop->f_detach(kn);
1098 * The user may change some filter values after the initial EV_ADD,
1099 * but doing so will not reset any filter which has already been
1102 kn->kn_status |= KN_INFLUX;
1105 kn->kn_kevent.udata = kev->udata;
1106 if (!fops->f_isfd && fops->f_touch != NULL) {
1107 fops->f_touch(kn, kev, EVENT_REGISTER);
1109 kn->kn_sfflags = kev->fflags;
1110 kn->kn_sdata = kev->data;
1114 * We can get here with kn->kn_knlist == NULL. This can happen when
1115 * the initial attach event decides that the event is "completed"
1116 * already. i.e. filt_procattach is called on a zombie process. It
1117 * will call filt_proc which will remove it from the list, and NULL
1121 event = kn->kn_fop->f_event(kn, 0);
1124 KNOTE_ACTIVATE(kn, 1);
1125 kn->kn_status &= ~KN_INFLUX;
1128 if ((kev->flags & EV_DISABLE) &&
1129 ((kn->kn_status & KN_DISABLED) == 0)) {
1130 kn->kn_status |= KN_DISABLED;
1133 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1134 kn->kn_status &= ~KN_DISABLED;
1135 if ((kn->kn_status & KN_ACTIVE) &&
1136 ((kn->kn_status & KN_QUEUED) == 0))
1142 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1148 kqueue_fo_release(filt);
1153 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1161 if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1165 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1176 kqueue_release(struct kqueue *kq, int locked)
1183 if (kq->kq_refcnt == 1)
1184 wakeup(&kq->kq_refcnt);
1190 kqueue_schedtask(struct kqueue *kq)
1194 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1195 ("scheduling kqueue task while draining"));
1197 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1198 taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1199 kq->kq_state |= KQ_TASKSCHED;
1204 * Expand the kq to make sure we have storage for fops/ident pair.
1206 * Return 0 on success (or no work necessary), return errno on failure.
1208 * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1209 * If kqueue_register is called from a non-fd context, there usually/should
1213 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1216 struct klist *list, *tmp_knhash, *to_free;
1217 u_long tmp_knhashmask;
1220 int mflag = waitok ? M_WAITOK : M_NOWAIT;
1227 if (kq->kq_knlistsize <= fd) {
1228 size = kq->kq_knlistsize;
1231 list = malloc(size * sizeof list, M_KQUEUE, mflag);
1235 if (kq->kq_knlistsize > fd) {
1239 if (kq->kq_knlist != NULL) {
1240 bcopy(kq->kq_knlist, list,
1241 kq->kq_knlistsize * sizeof list);
1242 to_free = kq->kq_knlist;
1243 kq->kq_knlist = NULL;
1245 bzero((caddr_t)list +
1246 kq->kq_knlistsize * sizeof list,
1247 (size - kq->kq_knlistsize) * sizeof list);
1248 kq->kq_knlistsize = size;
1249 kq->kq_knlist = list;
1254 if (kq->kq_knhashmask == 0) {
1255 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1257 if (tmp_knhash == NULL)
1260 if (kq->kq_knhashmask == 0) {
1261 kq->kq_knhash = tmp_knhash;
1262 kq->kq_knhashmask = tmp_knhashmask;
1264 to_free = tmp_knhash;
1269 free(to_free, M_KQUEUE);
1276 kqueue_task(void *arg, int pending)
1284 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1287 KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1289 kq->kq_state &= ~KQ_TASKSCHED;
1290 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1291 wakeup(&kq->kq_state);
1294 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1298 * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1299 * We treat KN_MARKER knotes as if they are INFLUX.
1302 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1303 const struct timespec *tsp, struct kevent *keva, struct thread *td)
1305 struct kevent *kevp;
1306 struct timeval atv, rtv, ttv;
1307 struct knote *kn, *marker;
1308 int count, timeout, nkev, error, influx;
1309 int haskqglobal, touch;
1320 TIMESPEC_TO_TIMEVAL(&atv, tsp);
1321 if (itimerfix(&atv)) {
1325 if (tsp->tv_sec == 0 && tsp->tv_nsec == 0)
1328 timeout = atv.tv_sec > 24 * 60 * 60 ?
1329 24 * 60 * 60 * hz : tvtohz(&atv);
1330 getmicrouptime(&rtv);
1331 timevaladd(&atv, &rtv);
1337 marker = knote_alloc(1);
1338 if (marker == NULL) {
1342 marker->kn_status = KN_MARKER;
1347 if (atv.tv_sec || atv.tv_usec) {
1348 getmicrouptime(&rtv);
1349 if (timevalcmp(&rtv, &atv, >=))
1352 timevalsub(&ttv, &rtv);
1353 timeout = ttv.tv_sec > 24 * 60 * 60 ?
1354 24 * 60 * 60 * hz : tvtohz(&ttv);
1359 if (kq->kq_count == 0) {
1361 error = EWOULDBLOCK;
1363 kq->kq_state |= KQ_SLEEP;
1364 error = msleep(kq, &kq->kq_lock, PSOCK | PCATCH,
1369 /* don't restart after signals... */
1370 if (error == ERESTART)
1372 else if (error == EWOULDBLOCK)
1377 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1381 kn = TAILQ_FIRST(&kq->kq_head);
1383 if ((kn->kn_status == KN_MARKER && kn != marker) ||
1384 (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1389 kq->kq_state |= KQ_FLUXWAIT;
1390 error = msleep(kq, &kq->kq_lock, PSOCK,
1395 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1396 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1397 kn->kn_status &= ~KN_QUEUED;
1403 if (count == maxevents)
1407 KASSERT((kn->kn_status & KN_INFLUX) == 0,
1408 ("KN_INFLUX set when not suppose to be"));
1410 if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1411 kn->kn_status &= ~KN_QUEUED;
1412 kn->kn_status |= KN_INFLUX;
1416 * We don't need to lock the list since we've marked
1419 if (!(kn->kn_status & KN_DETACHED))
1420 kn->kn_fop->f_detach(kn);
1424 } else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1425 kn->kn_status &= ~KN_QUEUED;
1426 kn->kn_status |= KN_INFLUX;
1430 * We don't need to lock the list since we've marked
1433 *kevp = kn->kn_kevent;
1434 if (!(kn->kn_status & KN_DETACHED))
1435 kn->kn_fop->f_detach(kn);
1440 kn->kn_status |= KN_INFLUX;
1442 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1443 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1445 if (kn->kn_fop->f_event(kn, 0) == 0) {
1447 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1449 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX);
1455 touch = (!kn->kn_fop->f_isfd &&
1456 kn->kn_fop->f_touch != NULL);
1458 kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1460 *kevp = kn->kn_kevent;
1462 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1463 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1465 * Manually clear knotes who weren't
1468 if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1472 if (kn->kn_flags & EV_DISPATCH)
1473 kn->kn_status |= KN_DISABLED;
1474 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1477 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1479 kn->kn_status &= ~(KN_INFLUX);
1484 /* we are returning a copy to the user */
1489 if (nkev == KQ_NEVENTS) {
1492 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1500 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1508 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1509 td->td_retval[0] = maxevents - count;
1515 * This could be expanded to call kqueue_scan, if desired.
1519 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
1520 int flags, struct thread *td)
1527 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1528 int flags, struct thread *td)
1535 kqueue_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1544 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1545 struct ucred *active_cred, struct thread *td)
1548 * Enabling sigio causes two major problems:
1549 * 1) infinite recursion:
1550 * Synopsys: kevent is being used to track signals and have FIOASYNC
1551 * set. On receipt of a signal this will cause a kqueue to recurse
1552 * into itself over and over. Sending the sigio causes the kqueue
1553 * to become ready, which in turn posts sigio again, forever.
1554 * Solution: this can be solved by setting a flag in the kqueue that
1555 * we have a SIGIO in progress.
1556 * 2) locking problems:
1557 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1558 * us above the proc and pgrp locks.
1559 * Solution: Post a signal using an async mechanism, being sure to
1560 * record a generation count in the delivery so that we do not deliver
1561 * a signal to the wrong process.
1563 * Note, these two mechanisms are somewhat mutually exclusive!
1572 kq->kq_state |= KQ_ASYNC;
1574 kq->kq_state &= ~KQ_ASYNC;
1579 return (fsetown(*(int *)data, &kq->kq_sigio));
1582 *(int *)data = fgetown(&kq->kq_sigio);
1592 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1599 if ((error = kqueue_acquire(fp, &kq)))
1603 if (events & (POLLIN | POLLRDNORM)) {
1605 revents |= events & (POLLIN | POLLRDNORM);
1607 selrecord(td, &kq->kq_sel);
1608 if (SEL_WAITING(&kq->kq_sel))
1609 kq->kq_state |= KQ_SEL;
1612 kqueue_release(kq, 1);
1619 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1623 bzero((void *)st, sizeof *st);
1625 * We no longer return kq_count because the unlocked value is useless.
1626 * If you spent all this time getting the count, why not spend your
1627 * syscall better by calling kevent?
1629 * XXX - This is needed for libc_r.
1631 st->st_mode = S_IFIFO;
1637 kqueue_close(struct file *fp, struct thread *td)
1639 struct kqueue *kq = fp->f_data;
1640 struct filedesc *fdp;
1645 if ((error = kqueue_acquire(fp, &kq)))
1650 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1651 ("kqueue already closing"));
1652 kq->kq_state |= KQ_CLOSING;
1653 if (kq->kq_refcnt > 1)
1654 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1656 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1659 KASSERT(knlist_empty(&kq->kq_sel.si_note),
1660 ("kqueue's knlist not empty"));
1662 for (i = 0; i < kq->kq_knlistsize; i++) {
1663 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1664 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1665 kq->kq_state |= KQ_FLUXWAIT;
1666 msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1669 kn->kn_status |= KN_INFLUX;
1671 if (!(kn->kn_status & KN_DETACHED))
1672 kn->kn_fop->f_detach(kn);
1677 if (kq->kq_knhashmask != 0) {
1678 for (i = 0; i <= kq->kq_knhashmask; i++) {
1679 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1680 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1681 kq->kq_state |= KQ_FLUXWAIT;
1682 msleep(kq, &kq->kq_lock, PSOCK,
1686 kn->kn_status |= KN_INFLUX;
1688 if (!(kn->kn_status & KN_DETACHED))
1689 kn->kn_fop->f_detach(kn);
1696 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1697 kq->kq_state |= KQ_TASKDRAIN;
1698 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1701 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1702 selwakeuppri(&kq->kq_sel, PSOCK);
1703 if (!SEL_WAITING(&kq->kq_sel))
1704 kq->kq_state &= ~KQ_SEL;
1709 FILEDESC_XLOCK(fdp);
1710 SLIST_REMOVE(&fdp->fd_kqlist, kq, kqueue, kq_list);
1711 FILEDESC_XUNLOCK(fdp);
1713 seldrain(&kq->kq_sel);
1714 knlist_destroy(&kq->kq_sel.si_note);
1715 mtx_destroy(&kq->kq_lock);
1718 if (kq->kq_knhash != NULL)
1719 free(kq->kq_knhash, M_KQUEUE);
1720 if (kq->kq_knlist != NULL)
1721 free(kq->kq_knlist, M_KQUEUE);
1723 funsetown(&kq->kq_sigio);
1731 kqueue_wakeup(struct kqueue *kq)
1735 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1736 kq->kq_state &= ~KQ_SLEEP;
1739 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1740 selwakeuppri(&kq->kq_sel, PSOCK);
1741 if (!SEL_WAITING(&kq->kq_sel))
1742 kq->kq_state &= ~KQ_SEL;
1744 if (!knlist_empty(&kq->kq_sel.si_note))
1745 kqueue_schedtask(kq);
1746 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1747 pgsigio(&kq->kq_sigio, SIGIO, 0);
1752 * Walk down a list of knotes, activating them if their event has triggered.
1754 * There is a possibility to optimize in the case of one kq watching another.
1755 * Instead of scheduling a task to wake it up, you could pass enough state
1756 * down the chain to make up the parent kqueue. Make this code functional
1760 knote(struct knlist *list, long hint, int lockflags)
1769 KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1771 if ((lockflags & KNF_LISTLOCKED) == 0)
1772 list->kl_lock(list->kl_lockarg);
1775 * If we unlock the list lock (and set KN_INFLUX), we can eliminate
1776 * the kqueue scheduling, but this will introduce four
1777 * lock/unlock's for each knote to test. If we do, continue to use
1778 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is
1779 * only safe if you want to remove the current item, which we are
1782 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
1784 if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) {
1786 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1788 } else if ((lockflags & KNF_NOKQLOCK) != 0) {
1789 kn->kn_status |= KN_INFLUX;
1791 error = kn->kn_fop->f_event(kn, hint);
1793 kn->kn_status &= ~KN_INFLUX;
1795 KNOTE_ACTIVATE(kn, 1);
1798 kn->kn_status |= KN_HASKQLOCK;
1799 if (kn->kn_fop->f_event(kn, hint))
1800 KNOTE_ACTIVATE(kn, 1);
1801 kn->kn_status &= ~KN_HASKQLOCK;
1807 if ((lockflags & KNF_LISTLOCKED) == 0)
1808 list->kl_unlock(list->kl_lockarg);
1812 * add a knote to a knlist
1815 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1817 KNL_ASSERT_LOCK(knl, islocked);
1818 KQ_NOTOWNED(kn->kn_kq);
1819 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1820 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1822 knl->kl_lock(knl->kl_lockarg);
1823 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1825 knl->kl_unlock(knl->kl_lockarg);
1827 kn->kn_knlist = knl;
1828 kn->kn_status &= ~KN_DETACHED;
1829 KQ_UNLOCK(kn->kn_kq);
1833 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1835 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1836 KNL_ASSERT_LOCK(knl, knlislocked);
1837 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1839 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1840 ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1842 knl->kl_lock(knl->kl_lockarg);
1843 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1844 kn->kn_knlist = NULL;
1846 knl->kl_unlock(knl->kl_lockarg);
1849 kn->kn_status |= KN_DETACHED;
1851 KQ_UNLOCK(kn->kn_kq);
1855 * remove all knotes from a specified klist
1858 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1861 knlist_remove_kq(knl, kn, islocked, 0);
1865 * remove knote from a specified klist while in f_event handler.
1868 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1871 knlist_remove_kq(knl, kn, 1,
1872 (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1876 knlist_empty(struct knlist *knl)
1878 KNL_ASSERT_LOCKED(knl);
1879 return SLIST_EMPTY(&knl->kl_list);
1882 static struct mtx knlist_lock;
1883 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
1885 static void knlist_mtx_lock(void *arg);
1886 static void knlist_mtx_unlock(void *arg);
1889 knlist_mtx_lock(void *arg)
1891 mtx_lock((struct mtx *)arg);
1895 knlist_mtx_unlock(void *arg)
1897 mtx_unlock((struct mtx *)arg);
1901 knlist_mtx_assert_locked(void *arg)
1903 mtx_assert((struct mtx *)arg, MA_OWNED);
1907 knlist_mtx_assert_unlocked(void *arg)
1909 mtx_assert((struct mtx *)arg, MA_NOTOWNED);
1913 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
1914 void (*kl_unlock)(void *),
1915 void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
1919 knl->kl_lockarg = &knlist_lock;
1921 knl->kl_lockarg = lock;
1923 if (kl_lock == NULL)
1924 knl->kl_lock = knlist_mtx_lock;
1926 knl->kl_lock = kl_lock;
1927 if (kl_unlock == NULL)
1928 knl->kl_unlock = knlist_mtx_unlock;
1930 knl->kl_unlock = kl_unlock;
1931 if (kl_assert_locked == NULL)
1932 knl->kl_assert_locked = knlist_mtx_assert_locked;
1934 knl->kl_assert_locked = kl_assert_locked;
1935 if (kl_assert_unlocked == NULL)
1936 knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
1938 knl->kl_assert_unlocked = kl_assert_unlocked;
1940 SLIST_INIT(&knl->kl_list);
1944 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
1947 knlist_init(knl, lock, NULL, NULL, NULL, NULL);
1951 knlist_destroy(struct knlist *knl)
1956 * if we run across this error, we need to find the offending
1957 * driver and have it call knlist_clear.
1959 if (!SLIST_EMPTY(&knl->kl_list))
1960 printf("WARNING: destroying knlist w/ knotes on it!\n");
1963 knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
1964 SLIST_INIT(&knl->kl_list);
1968 * Even if we are locked, we may need to drop the lock to allow any influx
1969 * knotes time to "settle".
1972 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
1974 struct knote *kn, *kn2;
1978 KNL_ASSERT_LOCKED(knl);
1980 KNL_ASSERT_UNLOCKED(knl);
1981 again: /* need to reacquire lock since we have dropped it */
1982 knl->kl_lock(knl->kl_lockarg);
1985 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
1988 if ((kn->kn_status & KN_INFLUX)) {
1992 knlist_remove_kq(knl, kn, 1, 1);
1994 kn->kn_status |= KN_INFLUX | KN_DETACHED;
1998 /* Make sure cleared knotes disappear soon */
1999 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2005 if (!SLIST_EMPTY(&knl->kl_list)) {
2006 /* there are still KN_INFLUX remaining */
2007 kn = SLIST_FIRST(&knl->kl_list);
2010 KASSERT(kn->kn_status & KN_INFLUX,
2011 ("knote removed w/o list lock"));
2012 knl->kl_unlock(knl->kl_lockarg);
2013 kq->kq_state |= KQ_FLUXWAIT;
2014 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2020 KNL_ASSERT_LOCKED(knl);
2022 knl->kl_unlock(knl->kl_lockarg);
2023 KNL_ASSERT_UNLOCKED(knl);
2028 * Remove all knotes referencing a specified fd must be called with FILEDESC
2029 * lock. This prevents a race where a new fd comes along and occupies the
2030 * entry and we attach a knote to the fd.
2033 knote_fdclose(struct thread *td, int fd)
2035 struct filedesc *fdp = td->td_proc->p_fd;
2040 FILEDESC_XLOCK_ASSERT(fdp);
2043 * We shouldn't have to worry about new kevents appearing on fd
2044 * since filedesc is locked.
2046 SLIST_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2051 while (kq->kq_knlistsize > fd &&
2052 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2053 if (kn->kn_status & KN_INFLUX) {
2054 /* someone else might be waiting on our knote */
2057 kq->kq_state |= KQ_FLUXWAIT;
2058 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2061 kn->kn_status |= KN_INFLUX;
2063 if (!(kn->kn_status & KN_DETACHED))
2064 kn->kn_fop->f_detach(kn);
2074 knote_attach(struct knote *kn, struct kqueue *kq)
2078 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2081 if (kn->kn_fop->f_isfd) {
2082 if (kn->kn_id >= kq->kq_knlistsize)
2084 list = &kq->kq_knlist[kn->kn_id];
2086 if (kq->kq_knhash == NULL)
2088 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2091 SLIST_INSERT_HEAD(list, kn, kn_link);
2097 * knote must already have been detached using the f_detach method.
2098 * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2099 * to prevent other removal.
2102 knote_drop(struct knote *kn, struct thread *td)
2110 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2111 ("knote_drop called without KN_INFLUX set in kn_status"));
2114 if (kn->kn_fop->f_isfd)
2115 list = &kq->kq_knlist[kn->kn_id];
2117 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2119 if (!SLIST_EMPTY(list))
2120 SLIST_REMOVE(list, kn, knote, kn_link);
2121 if (kn->kn_status & KN_QUEUED)
2125 if (kn->kn_fop->f_isfd) {
2126 fdrop(kn->kn_fp, td);
2129 kqueue_fo_release(kn->kn_kevent.filter);
2135 knote_enqueue(struct knote *kn)
2137 struct kqueue *kq = kn->kn_kq;
2139 KQ_OWNED(kn->kn_kq);
2140 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2142 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2143 kn->kn_status |= KN_QUEUED;
2149 knote_dequeue(struct knote *kn)
2151 struct kqueue *kq = kn->kn_kq;
2153 KQ_OWNED(kn->kn_kq);
2154 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2156 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2157 kn->kn_status &= ~KN_QUEUED;
2165 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2166 NULL, NULL, UMA_ALIGN_PTR, 0);
2168 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2170 static struct knote *
2171 knote_alloc(int waitok)
2173 return ((struct knote *)uma_zalloc(knote_zone,
2174 (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2178 knote_free(struct knote *kn)
2181 uma_zfree(knote_zone, kn);
2185 * Register the kev w/ the kq specified by fd.
2188 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2194 if ((error = fget(td, fd, &fp)) != 0)
2196 if ((error = kqueue_acquire(fp, &kq)) != 0)
2199 error = kqueue_register(kq, kev, td, waitok);
2201 kqueue_release(kq, 0);