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/capsicum.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.
526 #define NOTE_TIMER_PRECMASK (NOTE_SECONDS|NOTE_MSECONDS|NOTE_USECONDS| \
529 static __inline sbintime_t
530 timer2sbintime(intptr_t data, int flags)
534 switch (flags & NOTE_TIMER_PRECMASK) {
538 case NOTE_MSECONDS: /* FALLTHROUGH */
553 if (data > SBT_MAX / modifier)
556 return (modifier * data);
560 filt_timerexpire(void *knx)
562 struct callout *calloutp;
567 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */
569 if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
570 calloutp = (struct callout *)kn->kn_hook;
571 *kn->kn_ptr.p_nexttime += timer2sbintime(kn->kn_sdata,
573 callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
574 filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
579 * data contains amount of time to sleep
582 filt_timerattach(struct knote *kn)
584 struct callout *calloutp;
586 unsigned int ncallouts;
588 if ((intptr_t)kn->kn_sdata < 0)
590 if ((intptr_t)kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
592 /* Only precision unit are supported in flags so far */
593 if (kn->kn_sfflags & ~NOTE_TIMER_PRECMASK)
596 to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
600 ncallouts = atomic_load_explicit(&kq_ncallouts, memory_order_relaxed);
602 if (ncallouts >= kq_calloutmax)
604 } while (!atomic_compare_exchange_weak_explicit(&kq_ncallouts,
605 &ncallouts, ncallouts + 1, memory_order_relaxed,
606 memory_order_relaxed));
608 kn->kn_flags |= EV_CLEAR; /* automatically set */
609 kn->kn_status &= ~KN_DETACHED; /* knlist_add clears it */
610 kn->kn_ptr.p_nexttime = malloc(sizeof(sbintime_t), M_KQUEUE, M_WAITOK);
611 calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
612 callout_init(calloutp, CALLOUT_MPSAFE);
613 kn->kn_hook = calloutp;
614 *kn->kn_ptr.p_nexttime = to + sbinuptime();
615 callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
616 filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
622 filt_timerdetach(struct knote *kn)
624 struct callout *calloutp;
627 calloutp = (struct callout *)kn->kn_hook;
628 callout_drain(calloutp);
629 free(calloutp, M_KQUEUE);
630 free(kn->kn_ptr.p_nexttime, M_KQUEUE);
631 old = atomic_fetch_sub_explicit(&kq_ncallouts, 1, memory_order_relaxed);
632 KASSERT(old > 0, ("Number of callouts cannot become negative"));
633 kn->kn_status |= KN_DETACHED; /* knlist_remove sets it */
637 filt_timer(struct knote *kn, long hint)
640 return (kn->kn_data != 0);
644 filt_userattach(struct knote *kn)
648 * EVFILT_USER knotes are not attached to anything in the kernel.
651 if (kn->kn_fflags & NOTE_TRIGGER)
659 filt_userdetach(__unused struct knote *kn)
663 * EVFILT_USER knotes are not attached to anything in the kernel.
668 filt_user(struct knote *kn, __unused long hint)
671 return (kn->kn_hookid);
675 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
681 if (kev->fflags & NOTE_TRIGGER)
684 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
685 kev->fflags &= NOTE_FFLAGSMASK;
691 kn->kn_sfflags &= kev->fflags;
695 kn->kn_sfflags |= kev->fflags;
699 kn->kn_sfflags = kev->fflags;
703 /* XXX Return error? */
706 kn->kn_sdata = kev->data;
707 if (kev->flags & EV_CLEAR) {
715 *kev = kn->kn_kevent;
716 kev->fflags = kn->kn_sfflags;
717 kev->data = kn->kn_sdata;
718 if (kn->kn_flags & EV_CLEAR) {
726 panic("filt_usertouch() - invalid type (%ld)", type);
732 sys_kqueue(struct thread *td, struct kqueue_args *uap)
734 struct filedesc *fdp;
739 fdp = td->td_proc->p_fd;
740 error = falloc(td, &fp, &fd, 0);
744 /* An extra reference on `fp' has been held for us by falloc(). */
745 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
746 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
747 TAILQ_INIT(&kq->kq_head);
749 knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
750 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
753 TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
754 FILEDESC_XUNLOCK(fdp);
756 finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
759 td->td_retval[0] = fd;
764 #ifndef _SYS_SYSPROTO_H_
767 const struct kevent *changelist;
769 struct kevent *eventlist;
771 const struct timespec *timeout;
775 sys_kevent(struct thread *td, struct kevent_args *uap)
777 struct timespec ts, *tsp;
778 struct kevent_copyops k_ops = { uap,
785 struct uio *ktruioin = NULL;
786 struct uio *ktruioout = NULL;
789 if (uap->timeout != NULL) {
790 error = copyin(uap->timeout, &ts, sizeof(ts));
798 if (KTRPOINT(td, KTR_GENIO)) {
799 ktriov.iov_base = uap->changelist;
800 ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
801 ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
802 .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
804 ktruioin = cloneuio(&ktruio);
805 ktriov.iov_base = uap->eventlist;
806 ktriov.iov_len = uap->nevents * sizeof(struct kevent);
807 ktruioout = cloneuio(&ktruio);
811 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
815 if (ktruioin != NULL) {
816 ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
817 ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
818 ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
819 ktrgenio(uap->fd, UIO_READ, ktruioout, error);
827 * Copy 'count' items into the destination list pointed to by uap->eventlist.
830 kevent_copyout(void *arg, struct kevent *kevp, int count)
832 struct kevent_args *uap;
835 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
836 uap = (struct kevent_args *)arg;
838 error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
840 uap->eventlist += count;
845 * Copy 'count' items from the list pointed to by uap->changelist.
848 kevent_copyin(void *arg, struct kevent *kevp, int count)
850 struct kevent_args *uap;
853 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
854 uap = (struct kevent_args *)arg;
856 error = copyin(uap->changelist, kevp, count * sizeof *kevp);
858 uap->changelist += count;
863 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
864 struct kevent_copyops *k_ops, const struct timespec *timeout)
866 struct kevent keva[KQ_NEVENTS];
867 struct kevent *kevp, *changes;
871 int i, n, nerrors, error;
873 cap_rights_init(&rights);
875 cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
877 cap_rights_set(&rights, CAP_KQUEUE_EVENT);
878 error = fget(td, fd, &rights, &fp);
882 error = kqueue_acquire(fp, &kq);
888 while (nchanges > 0) {
889 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
890 error = k_ops->k_copyin(k_ops->arg, keva, n);
894 for (i = 0; i < n; i++) {
898 kevp->flags &= ~EV_SYSFLAGS;
899 error = kqueue_register(kq, kevp, td, 1);
900 if (error || (kevp->flags & EV_RECEIPT)) {
902 kevp->flags = EV_ERROR;
904 (void) k_ops->k_copyout(k_ops->arg,
916 td->td_retval[0] = nerrors;
921 error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
923 kqueue_release(kq, 0);
930 kqueue_add_filteropts(int filt, struct filterops *filtops)
935 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
937 "trying to add a filterop that is out of range: %d is beyond %d\n",
938 ~filt, EVFILT_SYSCOUNT);
941 mtx_lock(&filterops_lock);
942 if (sysfilt_ops[~filt].for_fop != &null_filtops &&
943 sysfilt_ops[~filt].for_fop != NULL)
946 sysfilt_ops[~filt].for_fop = filtops;
947 sysfilt_ops[~filt].for_refcnt = 0;
949 mtx_unlock(&filterops_lock);
955 kqueue_del_filteropts(int filt)
960 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
963 mtx_lock(&filterops_lock);
964 if (sysfilt_ops[~filt].for_fop == &null_filtops ||
965 sysfilt_ops[~filt].for_fop == NULL)
967 else if (sysfilt_ops[~filt].for_refcnt != 0)
970 sysfilt_ops[~filt].for_fop = &null_filtops;
971 sysfilt_ops[~filt].for_refcnt = 0;
973 mtx_unlock(&filterops_lock);
978 static struct filterops *
979 kqueue_fo_find(int filt)
982 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
985 mtx_lock(&filterops_lock);
986 sysfilt_ops[~filt].for_refcnt++;
987 if (sysfilt_ops[~filt].for_fop == NULL)
988 sysfilt_ops[~filt].for_fop = &null_filtops;
989 mtx_unlock(&filterops_lock);
991 return sysfilt_ops[~filt].for_fop;
995 kqueue_fo_release(int filt)
998 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1001 mtx_lock(&filterops_lock);
1002 KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1003 ("filter object refcount not valid on release"));
1004 sysfilt_ops[~filt].for_refcnt--;
1005 mtx_unlock(&filterops_lock);
1009 * A ref to kq (obtained via kqueue_acquire) must be held. waitok will
1010 * influence if memory allocation should wait. Make sure it is 0 if you
1014 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1016 struct filterops *fops;
1018 struct knote *kn, *tkn;
1019 cap_rights_t rights;
1020 int error, filt, event;
1021 int haskqglobal, filedesc_unlock;
1027 filedesc_unlock = 0;
1030 fops = kqueue_fo_find(filt);
1034 tkn = knote_alloc(waitok); /* prevent waiting with locks */
1038 KASSERT(td != NULL, ("td is NULL"));
1039 error = fget(td, kev->ident,
1040 cap_rights_init(&rights, CAP_EVENT), &fp);
1044 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1045 kev->ident, 0) != 0) {
1049 error = kqueue_expand(kq, fops, kev->ident, waitok);
1055 if (fp->f_type == DTYPE_KQUEUE) {
1057 * if we add some inteligence about what we are doing,
1058 * we should be able to support events on ourselves.
1059 * We need to know when we are doing this to prevent
1060 * getting both the knlist lock and the kq lock since
1061 * they are the same thing.
1063 if (fp->f_data == kq) {
1069 * Pre-lock the filedesc before the global
1070 * lock mutex, see the comment in
1073 FILEDESC_XLOCK(td->td_proc->p_fd);
1074 filedesc_unlock = 1;
1075 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1079 if (kev->ident < kq->kq_knlistsize) {
1080 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1081 if (kev->filter == kn->kn_filter)
1085 if ((kev->flags & EV_ADD) == EV_ADD)
1086 kqueue_expand(kq, fops, kev->ident, waitok);
1089 if (kq->kq_knhashmask != 0) {
1092 list = &kq->kq_knhash[
1093 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1094 SLIST_FOREACH(kn, list, kn_link)
1095 if (kev->ident == kn->kn_id &&
1096 kev->filter == kn->kn_filter)
1101 /* knote is in the process of changing, wait for it to stablize. */
1102 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1103 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1104 if (filedesc_unlock) {
1105 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1106 filedesc_unlock = 0;
1108 kq->kq_state |= KQ_FLUXWAIT;
1109 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1118 * kn now contains the matching knote, or NULL if no match
1121 if (kev->flags & EV_ADD) {
1133 * apply reference counts to knote structure, and
1134 * do not release it at the end of this routine.
1139 kn->kn_sfflags = kev->fflags;
1140 kn->kn_sdata = kev->data;
1143 kn->kn_kevent = *kev;
1144 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1145 EV_ENABLE | EV_DISABLE);
1146 kn->kn_status = KN_INFLUX|KN_DETACHED;
1148 error = knote_attach(kn, kq);
1155 if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1162 /* No matching knote and the EV_ADD flag is not set. */
1169 if (kev->flags & EV_DELETE) {
1170 kn->kn_status |= KN_INFLUX;
1172 if (!(kn->kn_status & KN_DETACHED))
1173 kn->kn_fop->f_detach(kn);
1179 * The user may change some filter values after the initial EV_ADD,
1180 * but doing so will not reset any filter which has already been
1183 kn->kn_status |= KN_INFLUX | KN_SCAN;
1186 kn->kn_kevent.udata = kev->udata;
1187 if (!fops->f_isfd && fops->f_touch != NULL) {
1188 fops->f_touch(kn, kev, EVENT_REGISTER);
1190 kn->kn_sfflags = kev->fflags;
1191 kn->kn_sdata = kev->data;
1195 * We can get here with kn->kn_knlist == NULL. This can happen when
1196 * the initial attach event decides that the event is "completed"
1197 * already. i.e. filt_procattach is called on a zombie process. It
1198 * will call filt_proc which will remove it from the list, and NULL
1202 event = kn->kn_fop->f_event(kn, 0);
1205 KNOTE_ACTIVATE(kn, 1);
1206 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1209 if ((kev->flags & EV_DISABLE) &&
1210 ((kn->kn_status & KN_DISABLED) == 0)) {
1211 kn->kn_status |= KN_DISABLED;
1214 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1215 kn->kn_status &= ~KN_DISABLED;
1216 if ((kn->kn_status & KN_ACTIVE) &&
1217 ((kn->kn_status & KN_QUEUED) == 0))
1223 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1224 if (filedesc_unlock)
1225 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1231 kqueue_fo_release(filt);
1236 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1244 if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1248 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1259 kqueue_release(struct kqueue *kq, int locked)
1266 if (kq->kq_refcnt == 1)
1267 wakeup(&kq->kq_refcnt);
1273 kqueue_schedtask(struct kqueue *kq)
1277 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1278 ("scheduling kqueue task while draining"));
1280 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1281 taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1282 kq->kq_state |= KQ_TASKSCHED;
1287 * Expand the kq to make sure we have storage for fops/ident pair.
1289 * Return 0 on success (or no work necessary), return errno on failure.
1291 * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1292 * If kqueue_register is called from a non-fd context, there usually/should
1296 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1299 struct klist *list, *tmp_knhash, *to_free;
1300 u_long tmp_knhashmask;
1303 int mflag = waitok ? M_WAITOK : M_NOWAIT;
1310 if (kq->kq_knlistsize <= fd) {
1311 size = kq->kq_knlistsize;
1314 list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1318 if (kq->kq_knlistsize > fd) {
1322 if (kq->kq_knlist != NULL) {
1323 bcopy(kq->kq_knlist, list,
1324 kq->kq_knlistsize * sizeof(*list));
1325 to_free = kq->kq_knlist;
1326 kq->kq_knlist = NULL;
1328 bzero((caddr_t)list +
1329 kq->kq_knlistsize * sizeof(*list),
1330 (size - kq->kq_knlistsize) * sizeof(*list));
1331 kq->kq_knlistsize = size;
1332 kq->kq_knlist = list;
1337 if (kq->kq_knhashmask == 0) {
1338 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1340 if (tmp_knhash == NULL)
1343 if (kq->kq_knhashmask == 0) {
1344 kq->kq_knhash = tmp_knhash;
1345 kq->kq_knhashmask = tmp_knhashmask;
1347 to_free = tmp_knhash;
1352 free(to_free, M_KQUEUE);
1359 kqueue_task(void *arg, int pending)
1367 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1370 KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1372 kq->kq_state &= ~KQ_TASKSCHED;
1373 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1374 wakeup(&kq->kq_state);
1377 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1381 * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1382 * We treat KN_MARKER knotes as if they are INFLUX.
1385 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1386 const struct timespec *tsp, struct kevent *keva, struct thread *td)
1388 struct kevent *kevp;
1389 struct knote *kn, *marker;
1390 sbintime_t asbt, rsbt;
1391 int count, error, haskqglobal, influx, nkev, touch;
1403 if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1404 tsp->tv_nsec >= 1000000000) {
1408 if (timespecisset(tsp)) {
1409 if (tsp->tv_sec <= INT32_MAX) {
1410 rsbt = tstosbt(*tsp);
1411 if (TIMESEL(&asbt, rsbt))
1412 asbt += tc_tick_sbt;
1413 if (asbt <= INT64_MAX - rsbt)
1417 rsbt >>= tc_precexp;
1424 marker = knote_alloc(1);
1425 if (marker == NULL) {
1429 marker->kn_status = KN_MARKER;
1434 if (kq->kq_count == 0) {
1436 error = EWOULDBLOCK;
1438 kq->kq_state |= KQ_SLEEP;
1439 error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1440 "kqread", asbt, rsbt, C_ABSOLUTE);
1444 /* don't restart after signals... */
1445 if (error == ERESTART)
1447 else if (error == EWOULDBLOCK)
1452 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1456 kn = TAILQ_FIRST(&kq->kq_head);
1458 if ((kn->kn_status == KN_MARKER && kn != marker) ||
1459 (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1464 kq->kq_state |= KQ_FLUXWAIT;
1465 error = msleep(kq, &kq->kq_lock, PSOCK,
1470 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1471 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1472 kn->kn_status &= ~KN_QUEUED;
1478 if (count == maxevents)
1482 KASSERT((kn->kn_status & KN_INFLUX) == 0,
1483 ("KN_INFLUX set when not suppose to be"));
1485 if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1486 kn->kn_status &= ~KN_QUEUED;
1487 kn->kn_status |= KN_INFLUX;
1491 * We don't need to lock the list since we've marked
1494 if (!(kn->kn_status & KN_DETACHED))
1495 kn->kn_fop->f_detach(kn);
1499 } else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1500 kn->kn_status &= ~KN_QUEUED;
1501 kn->kn_status |= KN_INFLUX;
1505 * We don't need to lock the list since we've marked
1508 *kevp = kn->kn_kevent;
1509 if (!(kn->kn_status & KN_DETACHED))
1510 kn->kn_fop->f_detach(kn);
1515 kn->kn_status |= KN_INFLUX | KN_SCAN;
1517 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1518 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1520 if (kn->kn_fop->f_event(kn, 0) == 0) {
1522 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1524 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1531 touch = (!kn->kn_fop->f_isfd &&
1532 kn->kn_fop->f_touch != NULL);
1534 kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1536 *kevp = kn->kn_kevent;
1538 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1539 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1541 * Manually clear knotes who weren't
1544 if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1548 if (kn->kn_flags & EV_DISPATCH)
1549 kn->kn_status |= KN_DISABLED;
1550 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1553 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1555 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1560 /* we are returning a copy to the user */
1565 if (nkev == KQ_NEVENTS) {
1568 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1576 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1584 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1585 td->td_retval[0] = maxevents - count;
1591 * This could be expanded to call kqueue_scan, if desired.
1595 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
1596 int flags, struct thread *td)
1603 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1604 int flags, struct thread *td)
1611 kqueue_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1620 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1621 struct ucred *active_cred, struct thread *td)
1624 * Enabling sigio causes two major problems:
1625 * 1) infinite recursion:
1626 * Synopsys: kevent is being used to track signals and have FIOASYNC
1627 * set. On receipt of a signal this will cause a kqueue to recurse
1628 * into itself over and over. Sending the sigio causes the kqueue
1629 * to become ready, which in turn posts sigio again, forever.
1630 * Solution: this can be solved by setting a flag in the kqueue that
1631 * we have a SIGIO in progress.
1632 * 2) locking problems:
1633 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1634 * us above the proc and pgrp locks.
1635 * Solution: Post a signal using an async mechanism, being sure to
1636 * record a generation count in the delivery so that we do not deliver
1637 * a signal to the wrong process.
1639 * Note, these two mechanisms are somewhat mutually exclusive!
1648 kq->kq_state |= KQ_ASYNC;
1650 kq->kq_state &= ~KQ_ASYNC;
1655 return (fsetown(*(int *)data, &kq->kq_sigio));
1658 *(int *)data = fgetown(&kq->kq_sigio);
1668 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1675 if ((error = kqueue_acquire(fp, &kq)))
1679 if (events & (POLLIN | POLLRDNORM)) {
1681 revents |= events & (POLLIN | POLLRDNORM);
1683 selrecord(td, &kq->kq_sel);
1684 if (SEL_WAITING(&kq->kq_sel))
1685 kq->kq_state |= KQ_SEL;
1688 kqueue_release(kq, 1);
1695 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1699 bzero((void *)st, sizeof *st);
1701 * We no longer return kq_count because the unlocked value is useless.
1702 * If you spent all this time getting the count, why not spend your
1703 * syscall better by calling kevent?
1705 * XXX - This is needed for libc_r.
1707 st->st_mode = S_IFIFO;
1713 kqueue_close(struct file *fp, struct thread *td)
1715 struct kqueue *kq = fp->f_data;
1716 struct filedesc *fdp;
1720 int filedesc_unlock;
1722 if ((error = kqueue_acquire(fp, &kq)))
1725 filedesc_unlock = 0;
1728 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1729 ("kqueue already closing"));
1730 kq->kq_state |= KQ_CLOSING;
1731 if (kq->kq_refcnt > 1)
1732 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1734 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1737 KASSERT(knlist_empty(&kq->kq_sel.si_note),
1738 ("kqueue's knlist not empty"));
1740 for (i = 0; i < kq->kq_knlistsize; i++) {
1741 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1742 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1743 kq->kq_state |= KQ_FLUXWAIT;
1744 msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1747 kn->kn_status |= KN_INFLUX;
1749 if (!(kn->kn_status & KN_DETACHED))
1750 kn->kn_fop->f_detach(kn);
1755 if (kq->kq_knhashmask != 0) {
1756 for (i = 0; i <= kq->kq_knhashmask; i++) {
1757 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1758 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1759 kq->kq_state |= KQ_FLUXWAIT;
1760 msleep(kq, &kq->kq_lock, PSOCK,
1764 kn->kn_status |= KN_INFLUX;
1766 if (!(kn->kn_status & KN_DETACHED))
1767 kn->kn_fop->f_detach(kn);
1774 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1775 kq->kq_state |= KQ_TASKDRAIN;
1776 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1779 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1780 selwakeuppri(&kq->kq_sel, PSOCK);
1781 if (!SEL_WAITING(&kq->kq_sel))
1782 kq->kq_state &= ~KQ_SEL;
1788 * We could be called due to the knote_drop() doing fdrop(),
1789 * called from kqueue_register(). In this case the global
1790 * lock is owned, and filedesc sx is locked before, to not
1791 * take the sleepable lock after non-sleepable.
1793 if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1794 FILEDESC_XLOCK(fdp);
1795 filedesc_unlock = 1;
1797 filedesc_unlock = 0;
1798 TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1799 if (filedesc_unlock)
1800 FILEDESC_XUNLOCK(fdp);
1802 seldrain(&kq->kq_sel);
1803 knlist_destroy(&kq->kq_sel.si_note);
1804 mtx_destroy(&kq->kq_lock);
1807 if (kq->kq_knhash != NULL)
1808 free(kq->kq_knhash, M_KQUEUE);
1809 if (kq->kq_knlist != NULL)
1810 free(kq->kq_knlist, M_KQUEUE);
1812 funsetown(&kq->kq_sigio);
1820 kqueue_wakeup(struct kqueue *kq)
1824 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1825 kq->kq_state &= ~KQ_SLEEP;
1828 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1829 selwakeuppri(&kq->kq_sel, PSOCK);
1830 if (!SEL_WAITING(&kq->kq_sel))
1831 kq->kq_state &= ~KQ_SEL;
1833 if (!knlist_empty(&kq->kq_sel.si_note))
1834 kqueue_schedtask(kq);
1835 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1836 pgsigio(&kq->kq_sigio, SIGIO, 0);
1841 * Walk down a list of knotes, activating them if their event has triggered.
1843 * There is a possibility to optimize in the case of one kq watching another.
1844 * Instead of scheduling a task to wake it up, you could pass enough state
1845 * down the chain to make up the parent kqueue. Make this code functional
1849 knote(struct knlist *list, long hint, int lockflags)
1858 KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1860 if ((lockflags & KNF_LISTLOCKED) == 0)
1861 list->kl_lock(list->kl_lockarg);
1864 * If we unlock the list lock (and set KN_INFLUX), we can eliminate
1865 * the kqueue scheduling, but this will introduce four
1866 * lock/unlock's for each knote to test. If we do, continue to use
1867 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is
1868 * only safe if you want to remove the current item, which we are
1871 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
1874 if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
1876 * Do not process the influx notes, except for
1877 * the influx coming from the kq unlock in the
1878 * kqueue_scan(). In the later case, we do
1879 * not interfere with the scan, since the code
1880 * fragment in kqueue_scan() locks the knlist,
1881 * and cannot proceed until we finished.
1884 } else if ((lockflags & KNF_NOKQLOCK) != 0) {
1885 kn->kn_status |= KN_INFLUX;
1887 error = kn->kn_fop->f_event(kn, hint);
1889 kn->kn_status &= ~KN_INFLUX;
1891 KNOTE_ACTIVATE(kn, 1);
1894 kn->kn_status |= KN_HASKQLOCK;
1895 if (kn->kn_fop->f_event(kn, hint))
1896 KNOTE_ACTIVATE(kn, 1);
1897 kn->kn_status &= ~KN_HASKQLOCK;
1901 if ((lockflags & KNF_LISTLOCKED) == 0)
1902 list->kl_unlock(list->kl_lockarg);
1906 * add a knote to a knlist
1909 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1911 KNL_ASSERT_LOCK(knl, islocked);
1912 KQ_NOTOWNED(kn->kn_kq);
1913 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1914 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1916 knl->kl_lock(knl->kl_lockarg);
1917 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1919 knl->kl_unlock(knl->kl_lockarg);
1921 kn->kn_knlist = knl;
1922 kn->kn_status &= ~KN_DETACHED;
1923 KQ_UNLOCK(kn->kn_kq);
1927 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1929 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1930 KNL_ASSERT_LOCK(knl, knlislocked);
1931 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1933 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1934 ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1936 knl->kl_lock(knl->kl_lockarg);
1937 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1938 kn->kn_knlist = NULL;
1940 knl->kl_unlock(knl->kl_lockarg);
1943 kn->kn_status |= KN_DETACHED;
1945 KQ_UNLOCK(kn->kn_kq);
1949 * remove knote from the specified knlist
1952 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1955 knlist_remove_kq(knl, kn, islocked, 0);
1959 * remove knote from the specified knlist while in f_event handler.
1962 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1965 knlist_remove_kq(knl, kn, 1,
1966 (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1970 knlist_empty(struct knlist *knl)
1973 KNL_ASSERT_LOCKED(knl);
1974 return SLIST_EMPTY(&knl->kl_list);
1977 static struct mtx knlist_lock;
1978 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
1980 static void knlist_mtx_lock(void *arg);
1981 static void knlist_mtx_unlock(void *arg);
1984 knlist_mtx_lock(void *arg)
1987 mtx_lock((struct mtx *)arg);
1991 knlist_mtx_unlock(void *arg)
1994 mtx_unlock((struct mtx *)arg);
1998 knlist_mtx_assert_locked(void *arg)
2001 mtx_assert((struct mtx *)arg, MA_OWNED);
2005 knlist_mtx_assert_unlocked(void *arg)
2008 mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2012 knlist_rw_rlock(void *arg)
2015 rw_rlock((struct rwlock *)arg);
2019 knlist_rw_runlock(void *arg)
2022 rw_runlock((struct rwlock *)arg);
2026 knlist_rw_assert_locked(void *arg)
2029 rw_assert((struct rwlock *)arg, RA_LOCKED);
2033 knlist_rw_assert_unlocked(void *arg)
2036 rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2040 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2041 void (*kl_unlock)(void *),
2042 void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2046 knl->kl_lockarg = &knlist_lock;
2048 knl->kl_lockarg = lock;
2050 if (kl_lock == NULL)
2051 knl->kl_lock = knlist_mtx_lock;
2053 knl->kl_lock = kl_lock;
2054 if (kl_unlock == NULL)
2055 knl->kl_unlock = knlist_mtx_unlock;
2057 knl->kl_unlock = kl_unlock;
2058 if (kl_assert_locked == NULL)
2059 knl->kl_assert_locked = knlist_mtx_assert_locked;
2061 knl->kl_assert_locked = kl_assert_locked;
2062 if (kl_assert_unlocked == NULL)
2063 knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2065 knl->kl_assert_unlocked = kl_assert_unlocked;
2067 SLIST_INIT(&knl->kl_list);
2071 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2074 knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2078 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2081 knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2082 knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2086 knlist_destroy(struct knlist *knl)
2091 * if we run across this error, we need to find the offending
2092 * driver and have it call knlist_clear or knlist_delete.
2094 if (!SLIST_EMPTY(&knl->kl_list))
2095 printf("WARNING: destroying knlist w/ knotes on it!\n");
2098 knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2099 SLIST_INIT(&knl->kl_list);
2103 * Even if we are locked, we may need to drop the lock to allow any influx
2104 * knotes time to "settle".
2107 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2109 struct knote *kn, *kn2;
2113 KNL_ASSERT_LOCKED(knl);
2115 KNL_ASSERT_UNLOCKED(knl);
2116 again: /* need to reacquire lock since we have dropped it */
2117 knl->kl_lock(knl->kl_lockarg);
2120 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2123 if ((kn->kn_status & KN_INFLUX)) {
2127 knlist_remove_kq(knl, kn, 1, 1);
2129 kn->kn_status |= KN_INFLUX | KN_DETACHED;
2133 /* Make sure cleared knotes disappear soon */
2134 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2140 if (!SLIST_EMPTY(&knl->kl_list)) {
2141 /* there are still KN_INFLUX remaining */
2142 kn = SLIST_FIRST(&knl->kl_list);
2145 KASSERT(kn->kn_status & KN_INFLUX,
2146 ("knote removed w/o list lock"));
2147 knl->kl_unlock(knl->kl_lockarg);
2148 kq->kq_state |= KQ_FLUXWAIT;
2149 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2155 KNL_ASSERT_LOCKED(knl);
2157 knl->kl_unlock(knl->kl_lockarg);
2158 KNL_ASSERT_UNLOCKED(knl);
2163 * Remove all knotes referencing a specified fd must be called with FILEDESC
2164 * lock. This prevents a race where a new fd comes along and occupies the
2165 * entry and we attach a knote to the fd.
2168 knote_fdclose(struct thread *td, int fd)
2170 struct filedesc *fdp = td->td_proc->p_fd;
2175 FILEDESC_XLOCK_ASSERT(fdp);
2178 * We shouldn't have to worry about new kevents appearing on fd
2179 * since filedesc is locked.
2181 TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2186 while (kq->kq_knlistsize > fd &&
2187 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2188 if (kn->kn_status & KN_INFLUX) {
2189 /* someone else might be waiting on our knote */
2192 kq->kq_state |= KQ_FLUXWAIT;
2193 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2196 kn->kn_status |= KN_INFLUX;
2198 if (!(kn->kn_status & KN_DETACHED))
2199 kn->kn_fop->f_detach(kn);
2209 knote_attach(struct knote *kn, struct kqueue *kq)
2213 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2216 if (kn->kn_fop->f_isfd) {
2217 if (kn->kn_id >= kq->kq_knlistsize)
2219 list = &kq->kq_knlist[kn->kn_id];
2221 if (kq->kq_knhash == NULL)
2223 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2226 SLIST_INSERT_HEAD(list, kn, kn_link);
2232 * knote must already have been detached using the f_detach method.
2233 * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2234 * to prevent other removal.
2237 knote_drop(struct knote *kn, struct thread *td)
2245 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2246 ("knote_drop called without KN_INFLUX set in kn_status"));
2249 if (kn->kn_fop->f_isfd)
2250 list = &kq->kq_knlist[kn->kn_id];
2252 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2254 if (!SLIST_EMPTY(list))
2255 SLIST_REMOVE(list, kn, knote, kn_link);
2256 if (kn->kn_status & KN_QUEUED)
2260 if (kn->kn_fop->f_isfd) {
2261 fdrop(kn->kn_fp, td);
2264 kqueue_fo_release(kn->kn_kevent.filter);
2270 knote_enqueue(struct knote *kn)
2272 struct kqueue *kq = kn->kn_kq;
2274 KQ_OWNED(kn->kn_kq);
2275 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2277 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2278 kn->kn_status |= KN_QUEUED;
2284 knote_dequeue(struct knote *kn)
2286 struct kqueue *kq = kn->kn_kq;
2288 KQ_OWNED(kn->kn_kq);
2289 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2291 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2292 kn->kn_status &= ~KN_QUEUED;
2300 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2301 NULL, NULL, UMA_ALIGN_PTR, 0);
2303 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2305 static struct knote *
2306 knote_alloc(int waitok)
2308 return ((struct knote *)uma_zalloc(knote_zone,
2309 (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2313 knote_free(struct knote *kn)
2316 uma_zfree(knote_zone, kn);
2320 * Register the kev w/ the kq specified by fd.
2323 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2327 cap_rights_t rights;
2330 error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2333 if ((error = kqueue_acquire(fp, &kq)) != 0)
2336 error = kqueue_register(kq, kev, td, waitok);
2338 kqueue_release(kq, 0);