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/capability.h>
37 #include <sys/kernel.h>
39 #include <sys/mutex.h>
40 #include <sys/rwlock.h>
42 #include <sys/malloc.h>
43 #include <sys/unistd.h>
45 #include <sys/filedesc.h>
46 #include <sys/filio.h>
47 #include <sys/fcntl.h>
48 #include <sys/kthread.h>
49 #include <sys/selinfo.h>
50 #include <sys/stdatomic.h>
51 #include <sys/queue.h>
52 #include <sys/event.h>
53 #include <sys/eventvar.h>
55 #include <sys/protosw.h>
56 #include <sys/sigio.h>
57 #include <sys/signalvar.h>
58 #include <sys/socket.h>
59 #include <sys/socketvar.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/syscallsubr.h>
64 #include <sys/taskqueue.h>
67 #include <sys/ktrace.h>
72 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
75 * This lock is used if multiple kq locks are required. This possibly
76 * should be made into a per proc lock.
78 static struct mtx kq_global;
79 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
80 #define KQ_GLOBAL_LOCK(lck, haslck) do { \
85 #define KQ_GLOBAL_UNLOCK(lck, haslck) do { \
91 TASKQUEUE_DEFINE_THREAD(kqueue);
93 static int kevent_copyout(void *arg, struct kevent *kevp, int count);
94 static int kevent_copyin(void *arg, struct kevent *kevp, int count);
95 static int kqueue_register(struct kqueue *kq, struct kevent *kev,
96 struct thread *td, int waitok);
97 static int kqueue_acquire(struct file *fp, struct kqueue **kqp);
98 static void kqueue_release(struct kqueue *kq, int locked);
99 static int kqueue_expand(struct kqueue *kq, struct filterops *fops,
100 uintptr_t ident, int waitok);
101 static void kqueue_task(void *arg, int pending);
102 static int kqueue_scan(struct kqueue *kq, int maxevents,
103 struct kevent_copyops *k_ops,
104 const struct timespec *timeout,
105 struct kevent *keva, struct thread *td);
106 static void kqueue_wakeup(struct kqueue *kq);
107 static struct filterops *kqueue_fo_find(int filt);
108 static void kqueue_fo_release(int filt);
110 static fo_rdwr_t kqueue_read;
111 static fo_rdwr_t kqueue_write;
112 static fo_truncate_t kqueue_truncate;
113 static fo_ioctl_t kqueue_ioctl;
114 static fo_poll_t kqueue_poll;
115 static fo_kqfilter_t kqueue_kqfilter;
116 static fo_stat_t kqueue_stat;
117 static fo_close_t kqueue_close;
119 static struct fileops kqueueops = {
120 .fo_read = kqueue_read,
121 .fo_write = kqueue_write,
122 .fo_truncate = kqueue_truncate,
123 .fo_ioctl = kqueue_ioctl,
124 .fo_poll = kqueue_poll,
125 .fo_kqfilter = kqueue_kqfilter,
126 .fo_stat = kqueue_stat,
127 .fo_close = kqueue_close,
128 .fo_chmod = invfo_chmod,
129 .fo_chown = invfo_chown,
130 .fo_sendfile = invfo_sendfile,
133 static int knote_attach(struct knote *kn, struct kqueue *kq);
134 static void knote_drop(struct knote *kn, struct thread *td);
135 static void knote_enqueue(struct knote *kn);
136 static void knote_dequeue(struct knote *kn);
137 static void knote_init(void);
138 static struct knote *knote_alloc(int waitok);
139 static void knote_free(struct knote *kn);
141 static void filt_kqdetach(struct knote *kn);
142 static int filt_kqueue(struct knote *kn, long hint);
143 static int filt_procattach(struct knote *kn);
144 static void filt_procdetach(struct knote *kn);
145 static int filt_proc(struct knote *kn, long hint);
146 static int filt_fileattach(struct knote *kn);
147 static void filt_timerexpire(void *knx);
148 static int filt_timerattach(struct knote *kn);
149 static void filt_timerdetach(struct knote *kn);
150 static int filt_timer(struct knote *kn, long hint);
151 static int filt_userattach(struct knote *kn);
152 static void filt_userdetach(struct knote *kn);
153 static int filt_user(struct knote *kn, long hint);
154 static void filt_usertouch(struct knote *kn, struct kevent *kev,
157 static struct filterops file_filtops = {
159 .f_attach = filt_fileattach,
161 static struct filterops kqread_filtops = {
163 .f_detach = filt_kqdetach,
164 .f_event = filt_kqueue,
166 /* XXX - move to kern_proc.c? */
167 static struct filterops proc_filtops = {
169 .f_attach = filt_procattach,
170 .f_detach = filt_procdetach,
171 .f_event = filt_proc,
173 static struct filterops timer_filtops = {
175 .f_attach = filt_timerattach,
176 .f_detach = filt_timerdetach,
177 .f_event = filt_timer,
179 static struct filterops user_filtops = {
180 .f_attach = filt_userattach,
181 .f_detach = filt_userdetach,
182 .f_event = filt_user,
183 .f_touch = filt_usertouch,
186 static uma_zone_t knote_zone;
187 static atomic_uint kq_ncallouts = ATOMIC_VAR_INIT(0);
188 static unsigned int kq_calloutmax = 4 * 1024;
189 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
190 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
192 /* XXX - ensure not KN_INFLUX?? */
193 #define KNOTE_ACTIVATE(kn, islock) do { \
195 mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED); \
197 KQ_LOCK((kn)->kn_kq); \
198 (kn)->kn_status |= KN_ACTIVE; \
199 if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \
200 knote_enqueue((kn)); \
202 KQ_UNLOCK((kn)->kn_kq); \
204 #define KQ_LOCK(kq) do { \
205 mtx_lock(&(kq)->kq_lock); \
207 #define KQ_FLUX_WAKEUP(kq) do { \
208 if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) { \
209 (kq)->kq_state &= ~KQ_FLUXWAIT; \
213 #define KQ_UNLOCK_FLUX(kq) do { \
214 KQ_FLUX_WAKEUP(kq); \
215 mtx_unlock(&(kq)->kq_lock); \
217 #define KQ_UNLOCK(kq) do { \
218 mtx_unlock(&(kq)->kq_lock); \
220 #define KQ_OWNED(kq) do { \
221 mtx_assert(&(kq)->kq_lock, MA_OWNED); \
223 #define KQ_NOTOWNED(kq) do { \
224 mtx_assert(&(kq)->kq_lock, MA_NOTOWNED); \
226 #define KN_LIST_LOCK(kn) do { \
227 if (kn->kn_knlist != NULL) \
228 kn->kn_knlist->kl_lock(kn->kn_knlist->kl_lockarg); \
230 #define KN_LIST_UNLOCK(kn) do { \
231 if (kn->kn_knlist != NULL) \
232 kn->kn_knlist->kl_unlock(kn->kn_knlist->kl_lockarg); \
234 #define KNL_ASSERT_LOCK(knl, islocked) do { \
236 KNL_ASSERT_LOCKED(knl); \
238 KNL_ASSERT_UNLOCKED(knl); \
241 #define KNL_ASSERT_LOCKED(knl) do { \
242 knl->kl_assert_locked((knl)->kl_lockarg); \
244 #define KNL_ASSERT_UNLOCKED(knl) do { \
245 knl->kl_assert_unlocked((knl)->kl_lockarg); \
247 #else /* !INVARIANTS */
248 #define KNL_ASSERT_LOCKED(knl) do {} while(0)
249 #define KNL_ASSERT_UNLOCKED(knl) do {} while (0)
250 #endif /* INVARIANTS */
252 #define KN_HASHSIZE 64 /* XXX should be tunable */
253 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask))
256 filt_nullattach(struct knote *kn)
262 struct filterops null_filtops = {
264 .f_attach = filt_nullattach,
267 /* XXX - make SYSINIT to add these, and move into respective modules. */
268 extern struct filterops sig_filtops;
269 extern struct filterops fs_filtops;
272 * Table for for all system-defined filters.
274 static struct mtx filterops_lock;
275 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
278 struct filterops *for_fop;
280 } sysfilt_ops[EVFILT_SYSCOUNT] = {
281 { &file_filtops }, /* EVFILT_READ */
282 { &file_filtops }, /* EVFILT_WRITE */
283 { &null_filtops }, /* EVFILT_AIO */
284 { &file_filtops }, /* EVFILT_VNODE */
285 { &proc_filtops }, /* EVFILT_PROC */
286 { &sig_filtops }, /* EVFILT_SIGNAL */
287 { &timer_filtops }, /* EVFILT_TIMER */
288 { &null_filtops }, /* former EVFILT_NETDEV */
289 { &fs_filtops }, /* EVFILT_FS */
290 { &null_filtops }, /* EVFILT_LIO */
291 { &user_filtops }, /* EVFILT_USER */
295 * Simple redirection for all cdevsw style objects to call their fo_kqfilter
299 filt_fileattach(struct knote *kn)
302 return (fo_kqfilter(kn->kn_fp, kn));
307 kqueue_kqfilter(struct file *fp, struct knote *kn)
309 struct kqueue *kq = kn->kn_fp->f_data;
311 if (kn->kn_filter != EVFILT_READ)
314 kn->kn_status |= KN_KQUEUE;
315 kn->kn_fop = &kqread_filtops;
316 knlist_add(&kq->kq_sel.si_note, kn, 0);
322 filt_kqdetach(struct knote *kn)
324 struct kqueue *kq = kn->kn_fp->f_data;
326 knlist_remove(&kq->kq_sel.si_note, kn, 0);
331 filt_kqueue(struct knote *kn, long hint)
333 struct kqueue *kq = kn->kn_fp->f_data;
335 kn->kn_data = kq->kq_count;
336 return (kn->kn_data > 0);
339 /* XXX - move to kern_proc.c? */
341 filt_procattach(struct knote *kn)
348 p = pfind(kn->kn_id);
349 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
350 p = zpfind(kn->kn_id);
352 } else if (p != NULL && (p->p_flag & P_WEXIT)) {
358 if ((error = p_cansee(curthread, p))) {
363 kn->kn_ptr.p_proc = p;
364 kn->kn_flags |= EV_CLEAR; /* automatically set */
367 * internal flag indicating registration done by kernel
369 if (kn->kn_flags & EV_FLAG1) {
370 kn->kn_data = kn->kn_sdata; /* ppid */
371 kn->kn_fflags = NOTE_CHILD;
372 kn->kn_flags &= ~EV_FLAG1;
376 knlist_add(&p->p_klist, kn, 1);
379 * Immediately activate any exit notes if the target process is a
380 * zombie. This is necessary to handle the case where the target
381 * process, e.g. a child, dies before the kevent is registered.
383 if (immediate && filt_proc(kn, NOTE_EXIT))
384 KNOTE_ACTIVATE(kn, 0);
392 * The knote may be attached to a different process, which may exit,
393 * leaving nothing for the knote to be attached to. So when the process
394 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
395 * it will be deleted when read out. However, as part of the knote deletion,
396 * this routine is called, so a check is needed to avoid actually performing
397 * a detach, because the original process does not exist any more.
399 /* XXX - move to kern_proc.c? */
401 filt_procdetach(struct knote *kn)
405 p = kn->kn_ptr.p_proc;
406 knlist_remove(&p->p_klist, kn, 0);
407 kn->kn_ptr.p_proc = NULL;
410 /* XXX - move to kern_proc.c? */
412 filt_proc(struct knote *kn, long hint)
414 struct proc *p = kn->kn_ptr.p_proc;
418 * mask off extra data
420 event = (u_int)hint & NOTE_PCTRLMASK;
423 * if the user is interested in this event, record it.
425 if (kn->kn_sfflags & event)
426 kn->kn_fflags |= event;
429 * process is gone, so flag the event as finished.
431 if (event == NOTE_EXIT) {
432 if (!(kn->kn_status & KN_DETACHED))
433 knlist_remove_inevent(&p->p_klist, kn);
434 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
435 kn->kn_ptr.p_proc = NULL;
436 if (kn->kn_fflags & NOTE_EXIT)
437 kn->kn_data = p->p_xstat;
438 if (kn->kn_fflags == 0)
439 kn->kn_flags |= EV_DROP;
443 return (kn->kn_fflags != 0);
447 * Called when the process forked. It mostly does the same as the
448 * knote(), activating all knotes registered to be activated when the
449 * process forked. Additionally, for each knote attached to the
450 * parent, check whether user wants to track the new process. If so
451 * attach a new knote to it, and immediately report an event with the
455 knote_fork(struct knlist *list, int pid)
464 list->kl_lock(list->kl_lockarg);
466 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
467 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX)
471 if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
477 * The same as knote(), activate the event.
479 if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
480 kn->kn_status |= KN_HASKQLOCK;
481 if (kn->kn_fop->f_event(kn, NOTE_FORK))
482 KNOTE_ACTIVATE(kn, 1);
483 kn->kn_status &= ~KN_HASKQLOCK;
489 * The NOTE_TRACK case. In addition to the activation
490 * of the event, we need to register new event to
491 * track the child. Drop the locks in preparation for
492 * the call to kqueue_register().
494 kn->kn_status |= KN_INFLUX;
496 list->kl_unlock(list->kl_lockarg);
499 * Activate existing knote and register a knote with
503 kev.filter = kn->kn_filter;
504 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
505 kev.fflags = kn->kn_sfflags;
506 kev.data = kn->kn_id; /* parent */
507 kev.udata = kn->kn_kevent.udata;/* preserve udata */
508 error = kqueue_register(kq, &kev, NULL, 0);
510 kn->kn_fflags |= NOTE_TRACKERR;
511 if (kn->kn_fop->f_event(kn, NOTE_FORK))
512 KNOTE_ACTIVATE(kn, 0);
514 kn->kn_status &= ~KN_INFLUX;
516 list->kl_lock(list->kl_lockarg);
518 list->kl_unlock(list->kl_lockarg);
522 * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
523 * interval timer support code.
525 static __inline sbintime_t
526 timer2sbintime(intptr_t data)
529 return (SBT_1MS * data);
533 filt_timerexpire(void *knx)
535 struct callout *calloutp;
540 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */
542 if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
543 calloutp = (struct callout *)kn->kn_hook;
544 *kn->kn_ptr.p_nexttime += timer2sbintime(kn->kn_sdata);
545 callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
546 filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
551 * data contains amount of time to sleep, in milliseconds
554 filt_timerattach(struct knote *kn)
556 struct callout *calloutp;
558 unsigned int ncallouts;
560 if ((intptr_t)kn->kn_sdata < 0)
562 if ((intptr_t)kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
564 to = timer2sbintime(kn->kn_sdata);
568 ncallouts = atomic_load_explicit(&kq_ncallouts, memory_order_relaxed);
570 if (ncallouts >= kq_calloutmax)
572 } while (!atomic_compare_exchange_weak_explicit(&kq_ncallouts,
573 &ncallouts, ncallouts + 1, memory_order_relaxed,
574 memory_order_relaxed));
576 kn->kn_flags |= EV_CLEAR; /* automatically set */
577 kn->kn_status &= ~KN_DETACHED; /* knlist_add clears it */
578 kn->kn_ptr.p_nexttime = malloc(sizeof(sbintime_t), M_KQUEUE, M_WAITOK);
579 calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
580 callout_init(calloutp, CALLOUT_MPSAFE);
581 kn->kn_hook = calloutp;
582 *kn->kn_ptr.p_nexttime = to + sbinuptime();
583 callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
584 filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
590 filt_timerdetach(struct knote *kn)
592 struct callout *calloutp;
595 calloutp = (struct callout *)kn->kn_hook;
596 callout_drain(calloutp);
597 free(calloutp, M_KQUEUE);
598 free(kn->kn_ptr.p_nexttime, M_KQUEUE);
599 old = atomic_fetch_sub_explicit(&kq_ncallouts, 1, memory_order_relaxed);
600 KASSERT(old > 0, ("Number of callouts cannot become negative"));
601 kn->kn_status |= KN_DETACHED; /* knlist_remove sets it */
605 filt_timer(struct knote *kn, long hint)
608 return (kn->kn_data != 0);
612 filt_userattach(struct knote *kn)
616 * EVFILT_USER knotes are not attached to anything in the kernel.
619 if (kn->kn_fflags & NOTE_TRIGGER)
627 filt_userdetach(__unused struct knote *kn)
631 * EVFILT_USER knotes are not attached to anything in the kernel.
636 filt_user(struct knote *kn, __unused long hint)
639 return (kn->kn_hookid);
643 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
649 if (kev->fflags & NOTE_TRIGGER)
652 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
653 kev->fflags &= NOTE_FFLAGSMASK;
659 kn->kn_sfflags &= kev->fflags;
663 kn->kn_sfflags |= kev->fflags;
667 kn->kn_sfflags = kev->fflags;
671 /* XXX Return error? */
674 kn->kn_sdata = kev->data;
675 if (kev->flags & EV_CLEAR) {
683 *kev = kn->kn_kevent;
684 kev->fflags = kn->kn_sfflags;
685 kev->data = kn->kn_sdata;
686 if (kn->kn_flags & EV_CLEAR) {
694 panic("filt_usertouch() - invalid type (%ld)", type);
700 sys_kqueue(struct thread *td, struct kqueue_args *uap)
702 struct filedesc *fdp;
707 fdp = td->td_proc->p_fd;
708 error = falloc(td, &fp, &fd, 0);
712 /* An extra reference on `fp' has been held for us by falloc(). */
713 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
714 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
715 TAILQ_INIT(&kq->kq_head);
717 knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
718 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
721 TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
722 FILEDESC_XUNLOCK(fdp);
724 finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
727 td->td_retval[0] = fd;
732 #ifndef _SYS_SYSPROTO_H_
735 const struct kevent *changelist;
737 struct kevent *eventlist;
739 const struct timespec *timeout;
743 sys_kevent(struct thread *td, struct kevent_args *uap)
745 struct timespec ts, *tsp;
746 struct kevent_copyops k_ops = { uap,
753 struct uio *ktruioin = NULL;
754 struct uio *ktruioout = NULL;
757 if (uap->timeout != NULL) {
758 error = copyin(uap->timeout, &ts, sizeof(ts));
766 if (KTRPOINT(td, KTR_GENIO)) {
767 ktriov.iov_base = uap->changelist;
768 ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
769 ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
770 .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
772 ktruioin = cloneuio(&ktruio);
773 ktriov.iov_base = uap->eventlist;
774 ktriov.iov_len = uap->nevents * sizeof(struct kevent);
775 ktruioout = cloneuio(&ktruio);
779 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
783 if (ktruioin != NULL) {
784 ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
785 ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
786 ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
787 ktrgenio(uap->fd, UIO_READ, ktruioout, error);
795 * Copy 'count' items into the destination list pointed to by uap->eventlist.
798 kevent_copyout(void *arg, struct kevent *kevp, int count)
800 struct kevent_args *uap;
803 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
804 uap = (struct kevent_args *)arg;
806 error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
808 uap->eventlist += count;
813 * Copy 'count' items from the list pointed to by uap->changelist.
816 kevent_copyin(void *arg, struct kevent *kevp, int count)
818 struct kevent_args *uap;
821 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
822 uap = (struct kevent_args *)arg;
824 error = copyin(uap->changelist, kevp, count * sizeof *kevp);
826 uap->changelist += count;
831 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
832 struct kevent_copyops *k_ops, const struct timespec *timeout)
834 struct kevent keva[KQ_NEVENTS];
835 struct kevent *kevp, *changes;
839 int i, n, nerrors, error;
841 cap_rights_init(&rights);
843 cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
845 cap_rights_set(&rights, CAP_KQUEUE_EVENT);
846 error = fget(td, fd, &rights, &fp);
850 error = kqueue_acquire(fp, &kq);
856 while (nchanges > 0) {
857 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
858 error = k_ops->k_copyin(k_ops->arg, keva, n);
862 for (i = 0; i < n; i++) {
866 kevp->flags &= ~EV_SYSFLAGS;
867 error = kqueue_register(kq, kevp, td, 1);
868 if (error || (kevp->flags & EV_RECEIPT)) {
870 kevp->flags = EV_ERROR;
872 (void) k_ops->k_copyout(k_ops->arg,
884 td->td_retval[0] = nerrors;
889 error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
891 kqueue_release(kq, 0);
898 kqueue_add_filteropts(int filt, struct filterops *filtops)
903 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
905 "trying to add a filterop that is out of range: %d is beyond %d\n",
906 ~filt, EVFILT_SYSCOUNT);
909 mtx_lock(&filterops_lock);
910 if (sysfilt_ops[~filt].for_fop != &null_filtops &&
911 sysfilt_ops[~filt].for_fop != NULL)
914 sysfilt_ops[~filt].for_fop = filtops;
915 sysfilt_ops[~filt].for_refcnt = 0;
917 mtx_unlock(&filterops_lock);
923 kqueue_del_filteropts(int filt)
928 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
931 mtx_lock(&filterops_lock);
932 if (sysfilt_ops[~filt].for_fop == &null_filtops ||
933 sysfilt_ops[~filt].for_fop == NULL)
935 else if (sysfilt_ops[~filt].for_refcnt != 0)
938 sysfilt_ops[~filt].for_fop = &null_filtops;
939 sysfilt_ops[~filt].for_refcnt = 0;
941 mtx_unlock(&filterops_lock);
946 static struct filterops *
947 kqueue_fo_find(int filt)
950 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
953 mtx_lock(&filterops_lock);
954 sysfilt_ops[~filt].for_refcnt++;
955 if (sysfilt_ops[~filt].for_fop == NULL)
956 sysfilt_ops[~filt].for_fop = &null_filtops;
957 mtx_unlock(&filterops_lock);
959 return sysfilt_ops[~filt].for_fop;
963 kqueue_fo_release(int filt)
966 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
969 mtx_lock(&filterops_lock);
970 KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
971 ("filter object refcount not valid on release"));
972 sysfilt_ops[~filt].for_refcnt--;
973 mtx_unlock(&filterops_lock);
977 * A ref to kq (obtained via kqueue_acquire) must be held. waitok will
978 * influence if memory allocation should wait. Make sure it is 0 if you
982 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
984 struct filterops *fops;
986 struct knote *kn, *tkn;
988 int error, filt, event;
989 int haskqglobal, filedesc_unlock;
998 fops = kqueue_fo_find(filt);
1002 tkn = knote_alloc(waitok); /* prevent waiting with locks */
1006 KASSERT(td != NULL, ("td is NULL"));
1007 error = fget(td, kev->ident,
1008 cap_rights_init(&rights, CAP_EVENT), &fp);
1012 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1013 kev->ident, 0) != 0) {
1017 error = kqueue_expand(kq, fops, kev->ident, waitok);
1023 if (fp->f_type == DTYPE_KQUEUE) {
1025 * if we add some inteligence about what we are doing,
1026 * we should be able to support events on ourselves.
1027 * We need to know when we are doing this to prevent
1028 * getting both the knlist lock and the kq lock since
1029 * they are the same thing.
1031 if (fp->f_data == kq) {
1037 * Pre-lock the filedesc before the global
1038 * lock mutex, see the comment in
1041 FILEDESC_XLOCK(td->td_proc->p_fd);
1042 filedesc_unlock = 1;
1043 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1047 if (kev->ident < kq->kq_knlistsize) {
1048 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1049 if (kev->filter == kn->kn_filter)
1053 if ((kev->flags & EV_ADD) == EV_ADD)
1054 kqueue_expand(kq, fops, kev->ident, waitok);
1057 if (kq->kq_knhashmask != 0) {
1060 list = &kq->kq_knhash[
1061 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1062 SLIST_FOREACH(kn, list, kn_link)
1063 if (kev->ident == kn->kn_id &&
1064 kev->filter == kn->kn_filter)
1069 /* knote is in the process of changing, wait for it to stablize. */
1070 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1071 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1072 if (filedesc_unlock) {
1073 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1074 filedesc_unlock = 0;
1076 kq->kq_state |= KQ_FLUXWAIT;
1077 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1086 * kn now contains the matching knote, or NULL if no match
1089 if (kev->flags & EV_ADD) {
1101 * apply reference counts to knote structure, and
1102 * do not release it at the end of this routine.
1107 kn->kn_sfflags = kev->fflags;
1108 kn->kn_sdata = kev->data;
1111 kn->kn_kevent = *kev;
1112 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1113 EV_ENABLE | EV_DISABLE);
1114 kn->kn_status = KN_INFLUX|KN_DETACHED;
1116 error = knote_attach(kn, kq);
1123 if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1130 /* No matching knote and the EV_ADD flag is not set. */
1137 if (kev->flags & EV_DELETE) {
1138 kn->kn_status |= KN_INFLUX;
1140 if (!(kn->kn_status & KN_DETACHED))
1141 kn->kn_fop->f_detach(kn);
1147 * The user may change some filter values after the initial EV_ADD,
1148 * but doing so will not reset any filter which has already been
1151 kn->kn_status |= KN_INFLUX | KN_SCAN;
1154 kn->kn_kevent.udata = kev->udata;
1155 if (!fops->f_isfd && fops->f_touch != NULL) {
1156 fops->f_touch(kn, kev, EVENT_REGISTER);
1158 kn->kn_sfflags = kev->fflags;
1159 kn->kn_sdata = kev->data;
1163 * We can get here with kn->kn_knlist == NULL. This can happen when
1164 * the initial attach event decides that the event is "completed"
1165 * already. i.e. filt_procattach is called on a zombie process. It
1166 * will call filt_proc which will remove it from the list, and NULL
1170 event = kn->kn_fop->f_event(kn, 0);
1173 KNOTE_ACTIVATE(kn, 1);
1174 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1177 if ((kev->flags & EV_DISABLE) &&
1178 ((kn->kn_status & KN_DISABLED) == 0)) {
1179 kn->kn_status |= KN_DISABLED;
1182 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1183 kn->kn_status &= ~KN_DISABLED;
1184 if ((kn->kn_status & KN_ACTIVE) &&
1185 ((kn->kn_status & KN_QUEUED) == 0))
1191 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1192 if (filedesc_unlock)
1193 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1199 kqueue_fo_release(filt);
1204 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1212 if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1216 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1227 kqueue_release(struct kqueue *kq, int locked)
1234 if (kq->kq_refcnt == 1)
1235 wakeup(&kq->kq_refcnt);
1241 kqueue_schedtask(struct kqueue *kq)
1245 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1246 ("scheduling kqueue task while draining"));
1248 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1249 taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1250 kq->kq_state |= KQ_TASKSCHED;
1255 * Expand the kq to make sure we have storage for fops/ident pair.
1257 * Return 0 on success (or no work necessary), return errno on failure.
1259 * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1260 * If kqueue_register is called from a non-fd context, there usually/should
1264 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1267 struct klist *list, *tmp_knhash, *to_free;
1268 u_long tmp_knhashmask;
1271 int mflag = waitok ? M_WAITOK : M_NOWAIT;
1278 if (kq->kq_knlistsize <= fd) {
1279 size = kq->kq_knlistsize;
1282 list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1286 if (kq->kq_knlistsize > fd) {
1290 if (kq->kq_knlist != NULL) {
1291 bcopy(kq->kq_knlist, list,
1292 kq->kq_knlistsize * sizeof(*list));
1293 to_free = kq->kq_knlist;
1294 kq->kq_knlist = NULL;
1296 bzero((caddr_t)list +
1297 kq->kq_knlistsize * sizeof(*list),
1298 (size - kq->kq_knlistsize) * sizeof(*list));
1299 kq->kq_knlistsize = size;
1300 kq->kq_knlist = list;
1305 if (kq->kq_knhashmask == 0) {
1306 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1308 if (tmp_knhash == NULL)
1311 if (kq->kq_knhashmask == 0) {
1312 kq->kq_knhash = tmp_knhash;
1313 kq->kq_knhashmask = tmp_knhashmask;
1315 to_free = tmp_knhash;
1320 free(to_free, M_KQUEUE);
1327 kqueue_task(void *arg, int pending)
1335 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1338 KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1340 kq->kq_state &= ~KQ_TASKSCHED;
1341 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1342 wakeup(&kq->kq_state);
1345 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1349 * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1350 * We treat KN_MARKER knotes as if they are INFLUX.
1353 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1354 const struct timespec *tsp, struct kevent *keva, struct thread *td)
1356 struct kevent *kevp;
1357 struct knote *kn, *marker;
1358 sbintime_t asbt, rsbt;
1359 int count, error, haskqglobal, influx, nkev, touch;
1371 if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1372 tsp->tv_nsec >= 1000000000) {
1376 if (timespecisset(tsp)) {
1377 if (tsp->tv_sec <= INT32_MAX) {
1378 rsbt = tstosbt(*tsp);
1379 if (TIMESEL(&asbt, rsbt))
1380 asbt += tc_tick_sbt;
1381 if (asbt <= INT64_MAX - rsbt)
1385 rsbt >>= tc_precexp;
1392 marker = knote_alloc(1);
1393 if (marker == NULL) {
1397 marker->kn_status = KN_MARKER;
1402 if (kq->kq_count == 0) {
1404 error = EWOULDBLOCK;
1406 kq->kq_state |= KQ_SLEEP;
1407 error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1408 "kqread", asbt, rsbt, C_ABSOLUTE);
1412 /* don't restart after signals... */
1413 if (error == ERESTART)
1415 else if (error == EWOULDBLOCK)
1420 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1424 kn = TAILQ_FIRST(&kq->kq_head);
1426 if ((kn->kn_status == KN_MARKER && kn != marker) ||
1427 (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1432 kq->kq_state |= KQ_FLUXWAIT;
1433 error = msleep(kq, &kq->kq_lock, PSOCK,
1438 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1439 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1440 kn->kn_status &= ~KN_QUEUED;
1446 if (count == maxevents)
1450 KASSERT((kn->kn_status & KN_INFLUX) == 0,
1451 ("KN_INFLUX set when not suppose to be"));
1453 if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1454 kn->kn_status &= ~KN_QUEUED;
1455 kn->kn_status |= KN_INFLUX;
1459 * We don't need to lock the list since we've marked
1462 if (!(kn->kn_status & KN_DETACHED))
1463 kn->kn_fop->f_detach(kn);
1467 } else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1468 kn->kn_status &= ~KN_QUEUED;
1469 kn->kn_status |= KN_INFLUX;
1473 * We don't need to lock the list since we've marked
1476 *kevp = kn->kn_kevent;
1477 if (!(kn->kn_status & KN_DETACHED))
1478 kn->kn_fop->f_detach(kn);
1483 kn->kn_status |= KN_INFLUX | KN_SCAN;
1485 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1486 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1488 if (kn->kn_fop->f_event(kn, 0) == 0) {
1490 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1492 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1499 touch = (!kn->kn_fop->f_isfd &&
1500 kn->kn_fop->f_touch != NULL);
1502 kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1504 *kevp = kn->kn_kevent;
1506 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1507 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1509 * Manually clear knotes who weren't
1512 if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1516 if (kn->kn_flags & EV_DISPATCH)
1517 kn->kn_status |= KN_DISABLED;
1518 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1521 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1523 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1528 /* we are returning a copy to the user */
1533 if (nkev == KQ_NEVENTS) {
1536 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1544 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1552 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1553 td->td_retval[0] = maxevents - count;
1559 * This could be expanded to call kqueue_scan, if desired.
1563 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
1564 int flags, struct thread *td)
1571 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1572 int flags, struct thread *td)
1579 kqueue_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1588 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1589 struct ucred *active_cred, struct thread *td)
1592 * Enabling sigio causes two major problems:
1593 * 1) infinite recursion:
1594 * Synopsys: kevent is being used to track signals and have FIOASYNC
1595 * set. On receipt of a signal this will cause a kqueue to recurse
1596 * into itself over and over. Sending the sigio causes the kqueue
1597 * to become ready, which in turn posts sigio again, forever.
1598 * Solution: this can be solved by setting a flag in the kqueue that
1599 * we have a SIGIO in progress.
1600 * 2) locking problems:
1601 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1602 * us above the proc and pgrp locks.
1603 * Solution: Post a signal using an async mechanism, being sure to
1604 * record a generation count in the delivery so that we do not deliver
1605 * a signal to the wrong process.
1607 * Note, these two mechanisms are somewhat mutually exclusive!
1616 kq->kq_state |= KQ_ASYNC;
1618 kq->kq_state &= ~KQ_ASYNC;
1623 return (fsetown(*(int *)data, &kq->kq_sigio));
1626 *(int *)data = fgetown(&kq->kq_sigio);
1636 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1643 if ((error = kqueue_acquire(fp, &kq)))
1647 if (events & (POLLIN | POLLRDNORM)) {
1649 revents |= events & (POLLIN | POLLRDNORM);
1651 selrecord(td, &kq->kq_sel);
1652 if (SEL_WAITING(&kq->kq_sel))
1653 kq->kq_state |= KQ_SEL;
1656 kqueue_release(kq, 1);
1663 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1667 bzero((void *)st, sizeof *st);
1669 * We no longer return kq_count because the unlocked value is useless.
1670 * If you spent all this time getting the count, why not spend your
1671 * syscall better by calling kevent?
1673 * XXX - This is needed for libc_r.
1675 st->st_mode = S_IFIFO;
1681 kqueue_close(struct file *fp, struct thread *td)
1683 struct kqueue *kq = fp->f_data;
1684 struct filedesc *fdp;
1688 int filedesc_unlock;
1690 if ((error = kqueue_acquire(fp, &kq)))
1693 filedesc_unlock = 0;
1696 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1697 ("kqueue already closing"));
1698 kq->kq_state |= KQ_CLOSING;
1699 if (kq->kq_refcnt > 1)
1700 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1702 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1705 KASSERT(knlist_empty(&kq->kq_sel.si_note),
1706 ("kqueue's knlist not empty"));
1708 for (i = 0; i < kq->kq_knlistsize; i++) {
1709 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1710 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1711 kq->kq_state |= KQ_FLUXWAIT;
1712 msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1715 kn->kn_status |= KN_INFLUX;
1717 if (!(kn->kn_status & KN_DETACHED))
1718 kn->kn_fop->f_detach(kn);
1723 if (kq->kq_knhashmask != 0) {
1724 for (i = 0; i <= kq->kq_knhashmask; i++) {
1725 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1726 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1727 kq->kq_state |= KQ_FLUXWAIT;
1728 msleep(kq, &kq->kq_lock, PSOCK,
1732 kn->kn_status |= KN_INFLUX;
1734 if (!(kn->kn_status & KN_DETACHED))
1735 kn->kn_fop->f_detach(kn);
1742 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1743 kq->kq_state |= KQ_TASKDRAIN;
1744 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1747 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1748 selwakeuppri(&kq->kq_sel, PSOCK);
1749 if (!SEL_WAITING(&kq->kq_sel))
1750 kq->kq_state &= ~KQ_SEL;
1756 * We could be called due to the knote_drop() doing fdrop(),
1757 * called from kqueue_register(). In this case the global
1758 * lock is owned, and filedesc sx is locked before, to not
1759 * take the sleepable lock after non-sleepable.
1761 if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1762 FILEDESC_XLOCK(fdp);
1763 filedesc_unlock = 1;
1765 filedesc_unlock = 0;
1766 TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1767 if (filedesc_unlock)
1768 FILEDESC_XUNLOCK(fdp);
1770 seldrain(&kq->kq_sel);
1771 knlist_destroy(&kq->kq_sel.si_note);
1772 mtx_destroy(&kq->kq_lock);
1775 if (kq->kq_knhash != NULL)
1776 free(kq->kq_knhash, M_KQUEUE);
1777 if (kq->kq_knlist != NULL)
1778 free(kq->kq_knlist, M_KQUEUE);
1780 funsetown(&kq->kq_sigio);
1788 kqueue_wakeup(struct kqueue *kq)
1792 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1793 kq->kq_state &= ~KQ_SLEEP;
1796 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1797 selwakeuppri(&kq->kq_sel, PSOCK);
1798 if (!SEL_WAITING(&kq->kq_sel))
1799 kq->kq_state &= ~KQ_SEL;
1801 if (!knlist_empty(&kq->kq_sel.si_note))
1802 kqueue_schedtask(kq);
1803 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1804 pgsigio(&kq->kq_sigio, SIGIO, 0);
1809 * Walk down a list of knotes, activating them if their event has triggered.
1811 * There is a possibility to optimize in the case of one kq watching another.
1812 * Instead of scheduling a task to wake it up, you could pass enough state
1813 * down the chain to make up the parent kqueue. Make this code functional
1817 knote(struct knlist *list, long hint, int lockflags)
1826 KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1828 if ((lockflags & KNF_LISTLOCKED) == 0)
1829 list->kl_lock(list->kl_lockarg);
1832 * If we unlock the list lock (and set KN_INFLUX), we can eliminate
1833 * the kqueue scheduling, but this will introduce four
1834 * lock/unlock's for each knote to test. If we do, continue to use
1835 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is
1836 * only safe if you want to remove the current item, which we are
1839 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
1842 if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
1844 * Do not process the influx notes, except for
1845 * the influx coming from the kq unlock in the
1846 * kqueue_scan(). In the later case, we do
1847 * not interfere with the scan, since the code
1848 * fragment in kqueue_scan() locks the knlist,
1849 * and cannot proceed until we finished.
1852 } else if ((lockflags & KNF_NOKQLOCK) != 0) {
1853 kn->kn_status |= KN_INFLUX;
1855 error = kn->kn_fop->f_event(kn, hint);
1857 kn->kn_status &= ~KN_INFLUX;
1859 KNOTE_ACTIVATE(kn, 1);
1862 kn->kn_status |= KN_HASKQLOCK;
1863 if (kn->kn_fop->f_event(kn, hint))
1864 KNOTE_ACTIVATE(kn, 1);
1865 kn->kn_status &= ~KN_HASKQLOCK;
1869 if ((lockflags & KNF_LISTLOCKED) == 0)
1870 list->kl_unlock(list->kl_lockarg);
1874 * add a knote to a knlist
1877 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1879 KNL_ASSERT_LOCK(knl, islocked);
1880 KQ_NOTOWNED(kn->kn_kq);
1881 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1882 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1884 knl->kl_lock(knl->kl_lockarg);
1885 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1887 knl->kl_unlock(knl->kl_lockarg);
1889 kn->kn_knlist = knl;
1890 kn->kn_status &= ~KN_DETACHED;
1891 KQ_UNLOCK(kn->kn_kq);
1895 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1897 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1898 KNL_ASSERT_LOCK(knl, knlislocked);
1899 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1901 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1902 ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1904 knl->kl_lock(knl->kl_lockarg);
1905 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1906 kn->kn_knlist = NULL;
1908 knl->kl_unlock(knl->kl_lockarg);
1911 kn->kn_status |= KN_DETACHED;
1913 KQ_UNLOCK(kn->kn_kq);
1917 * remove knote from the specified knlist
1920 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1923 knlist_remove_kq(knl, kn, islocked, 0);
1927 * remove knote from the specified knlist while in f_event handler.
1930 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1933 knlist_remove_kq(knl, kn, 1,
1934 (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1938 knlist_empty(struct knlist *knl)
1941 KNL_ASSERT_LOCKED(knl);
1942 return SLIST_EMPTY(&knl->kl_list);
1945 static struct mtx knlist_lock;
1946 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
1948 static void knlist_mtx_lock(void *arg);
1949 static void knlist_mtx_unlock(void *arg);
1952 knlist_mtx_lock(void *arg)
1955 mtx_lock((struct mtx *)arg);
1959 knlist_mtx_unlock(void *arg)
1962 mtx_unlock((struct mtx *)arg);
1966 knlist_mtx_assert_locked(void *arg)
1969 mtx_assert((struct mtx *)arg, MA_OWNED);
1973 knlist_mtx_assert_unlocked(void *arg)
1976 mtx_assert((struct mtx *)arg, MA_NOTOWNED);
1980 knlist_rw_rlock(void *arg)
1983 rw_rlock((struct rwlock *)arg);
1987 knlist_rw_runlock(void *arg)
1990 rw_runlock((struct rwlock *)arg);
1994 knlist_rw_assert_locked(void *arg)
1997 rw_assert((struct rwlock *)arg, RA_LOCKED);
2001 knlist_rw_assert_unlocked(void *arg)
2004 rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2008 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2009 void (*kl_unlock)(void *),
2010 void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2014 knl->kl_lockarg = &knlist_lock;
2016 knl->kl_lockarg = lock;
2018 if (kl_lock == NULL)
2019 knl->kl_lock = knlist_mtx_lock;
2021 knl->kl_lock = kl_lock;
2022 if (kl_unlock == NULL)
2023 knl->kl_unlock = knlist_mtx_unlock;
2025 knl->kl_unlock = kl_unlock;
2026 if (kl_assert_locked == NULL)
2027 knl->kl_assert_locked = knlist_mtx_assert_locked;
2029 knl->kl_assert_locked = kl_assert_locked;
2030 if (kl_assert_unlocked == NULL)
2031 knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2033 knl->kl_assert_unlocked = kl_assert_unlocked;
2035 SLIST_INIT(&knl->kl_list);
2039 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2042 knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2046 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2049 knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2050 knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2054 knlist_destroy(struct knlist *knl)
2059 * if we run across this error, we need to find the offending
2060 * driver and have it call knlist_clear or knlist_delete.
2062 if (!SLIST_EMPTY(&knl->kl_list))
2063 printf("WARNING: destroying knlist w/ knotes on it!\n");
2066 knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2067 SLIST_INIT(&knl->kl_list);
2071 * Even if we are locked, we may need to drop the lock to allow any influx
2072 * knotes time to "settle".
2075 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2077 struct knote *kn, *kn2;
2081 KNL_ASSERT_LOCKED(knl);
2083 KNL_ASSERT_UNLOCKED(knl);
2084 again: /* need to reacquire lock since we have dropped it */
2085 knl->kl_lock(knl->kl_lockarg);
2088 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2091 if ((kn->kn_status & KN_INFLUX)) {
2095 knlist_remove_kq(knl, kn, 1, 1);
2097 kn->kn_status |= KN_INFLUX | KN_DETACHED;
2101 /* Make sure cleared knotes disappear soon */
2102 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2108 if (!SLIST_EMPTY(&knl->kl_list)) {
2109 /* there are still KN_INFLUX remaining */
2110 kn = SLIST_FIRST(&knl->kl_list);
2113 KASSERT(kn->kn_status & KN_INFLUX,
2114 ("knote removed w/o list lock"));
2115 knl->kl_unlock(knl->kl_lockarg);
2116 kq->kq_state |= KQ_FLUXWAIT;
2117 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2123 KNL_ASSERT_LOCKED(knl);
2125 knl->kl_unlock(knl->kl_lockarg);
2126 KNL_ASSERT_UNLOCKED(knl);
2131 * Remove all knotes referencing a specified fd must be called with FILEDESC
2132 * lock. This prevents a race where a new fd comes along and occupies the
2133 * entry and we attach a knote to the fd.
2136 knote_fdclose(struct thread *td, int fd)
2138 struct filedesc *fdp = td->td_proc->p_fd;
2143 FILEDESC_XLOCK_ASSERT(fdp);
2146 * We shouldn't have to worry about new kevents appearing on fd
2147 * since filedesc is locked.
2149 TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2154 while (kq->kq_knlistsize > fd &&
2155 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2156 if (kn->kn_status & KN_INFLUX) {
2157 /* someone else might be waiting on our knote */
2160 kq->kq_state |= KQ_FLUXWAIT;
2161 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2164 kn->kn_status |= KN_INFLUX;
2166 if (!(kn->kn_status & KN_DETACHED))
2167 kn->kn_fop->f_detach(kn);
2177 knote_attach(struct knote *kn, struct kqueue *kq)
2181 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2184 if (kn->kn_fop->f_isfd) {
2185 if (kn->kn_id >= kq->kq_knlistsize)
2187 list = &kq->kq_knlist[kn->kn_id];
2189 if (kq->kq_knhash == NULL)
2191 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2194 SLIST_INSERT_HEAD(list, kn, kn_link);
2200 * knote must already have been detached using the f_detach method.
2201 * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2202 * to prevent other removal.
2205 knote_drop(struct knote *kn, struct thread *td)
2213 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2214 ("knote_drop called without KN_INFLUX set in kn_status"));
2217 if (kn->kn_fop->f_isfd)
2218 list = &kq->kq_knlist[kn->kn_id];
2220 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2222 if (!SLIST_EMPTY(list))
2223 SLIST_REMOVE(list, kn, knote, kn_link);
2224 if (kn->kn_status & KN_QUEUED)
2228 if (kn->kn_fop->f_isfd) {
2229 fdrop(kn->kn_fp, td);
2232 kqueue_fo_release(kn->kn_kevent.filter);
2238 knote_enqueue(struct knote *kn)
2240 struct kqueue *kq = kn->kn_kq;
2242 KQ_OWNED(kn->kn_kq);
2243 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2245 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2246 kn->kn_status |= KN_QUEUED;
2252 knote_dequeue(struct knote *kn)
2254 struct kqueue *kq = kn->kn_kq;
2256 KQ_OWNED(kn->kn_kq);
2257 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2259 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2260 kn->kn_status &= ~KN_QUEUED;
2268 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2269 NULL, NULL, UMA_ALIGN_PTR, 0);
2271 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2273 static struct knote *
2274 knote_alloc(int waitok)
2276 return ((struct knote *)uma_zalloc(knote_zone,
2277 (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2281 knote_free(struct knote *kn)
2284 uma_zfree(knote_zone, kn);
2288 * Register the kev w/ the kq specified by fd.
2291 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2295 cap_rights_t rights;
2298 error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2301 if ((error = kqueue_acquire(fp, &kq)) != 0)
2304 error = kqueue_register(kq, kev, td, waitok);
2306 kqueue_release(kq, 0);