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)
735 return (kern_kqueue(td, 0));
739 kern_kqueue(struct thread *td, int flags)
741 struct filedesc *fdp;
746 fdp = td->td_proc->p_fd;
747 error = falloc(td, &fp, &fd, flags);
751 /* An extra reference on `fp' has been held for us by falloc(). */
752 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
753 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
754 TAILQ_INIT(&kq->kq_head);
756 knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
757 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
760 TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
761 FILEDESC_XUNLOCK(fdp);
763 finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
766 td->td_retval[0] = fd;
771 #ifndef _SYS_SYSPROTO_H_
774 const struct kevent *changelist;
776 struct kevent *eventlist;
778 const struct timespec *timeout;
782 sys_kevent(struct thread *td, struct kevent_args *uap)
784 struct timespec ts, *tsp;
785 struct kevent_copyops k_ops = { uap,
792 struct uio *ktruioin = NULL;
793 struct uio *ktruioout = NULL;
796 if (uap->timeout != NULL) {
797 error = copyin(uap->timeout, &ts, sizeof(ts));
805 if (KTRPOINT(td, KTR_GENIO)) {
806 ktriov.iov_base = uap->changelist;
807 ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
808 ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
809 .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
811 ktruioin = cloneuio(&ktruio);
812 ktriov.iov_base = uap->eventlist;
813 ktriov.iov_len = uap->nevents * sizeof(struct kevent);
814 ktruioout = cloneuio(&ktruio);
818 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
822 if (ktruioin != NULL) {
823 ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
824 ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
825 ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
826 ktrgenio(uap->fd, UIO_READ, ktruioout, error);
834 * Copy 'count' items into the destination list pointed to by uap->eventlist.
837 kevent_copyout(void *arg, struct kevent *kevp, int count)
839 struct kevent_args *uap;
842 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
843 uap = (struct kevent_args *)arg;
845 error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
847 uap->eventlist += count;
852 * Copy 'count' items from the list pointed to by uap->changelist.
855 kevent_copyin(void *arg, struct kevent *kevp, int count)
857 struct kevent_args *uap;
860 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
861 uap = (struct kevent_args *)arg;
863 error = copyin(uap->changelist, kevp, count * sizeof *kevp);
865 uap->changelist += count;
870 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
871 struct kevent_copyops *k_ops, const struct timespec *timeout)
877 cap_rights_init(&rights);
879 cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
881 cap_rights_set(&rights, CAP_KQUEUE_EVENT);
882 error = fget(td, fd, &rights, &fp);
886 error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
893 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
894 struct kevent_copyops *k_ops, const struct timespec *timeout)
896 struct kevent keva[KQ_NEVENTS];
897 struct kevent *kevp, *changes;
899 int i, n, nerrors, error;
901 error = kqueue_acquire(fp, &kq);
907 while (nchanges > 0) {
908 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
909 error = k_ops->k_copyin(k_ops->arg, keva, n);
913 for (i = 0; i < n; i++) {
917 kevp->flags &= ~EV_SYSFLAGS;
918 error = kqueue_register(kq, kevp, td, 1);
919 if (error || (kevp->flags & EV_RECEIPT)) {
921 kevp->flags = EV_ERROR;
923 (void) k_ops->k_copyout(k_ops->arg,
935 td->td_retval[0] = nerrors;
940 error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
942 kqueue_release(kq, 0);
947 kqueue_add_filteropts(int filt, struct filterops *filtops)
952 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
954 "trying to add a filterop that is out of range: %d is beyond %d\n",
955 ~filt, EVFILT_SYSCOUNT);
958 mtx_lock(&filterops_lock);
959 if (sysfilt_ops[~filt].for_fop != &null_filtops &&
960 sysfilt_ops[~filt].for_fop != NULL)
963 sysfilt_ops[~filt].for_fop = filtops;
964 sysfilt_ops[~filt].for_refcnt = 0;
966 mtx_unlock(&filterops_lock);
972 kqueue_del_filteropts(int filt)
977 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
980 mtx_lock(&filterops_lock);
981 if (sysfilt_ops[~filt].for_fop == &null_filtops ||
982 sysfilt_ops[~filt].for_fop == NULL)
984 else if (sysfilt_ops[~filt].for_refcnt != 0)
987 sysfilt_ops[~filt].for_fop = &null_filtops;
988 sysfilt_ops[~filt].for_refcnt = 0;
990 mtx_unlock(&filterops_lock);
995 static struct filterops *
996 kqueue_fo_find(int filt)
999 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1002 mtx_lock(&filterops_lock);
1003 sysfilt_ops[~filt].for_refcnt++;
1004 if (sysfilt_ops[~filt].for_fop == NULL)
1005 sysfilt_ops[~filt].for_fop = &null_filtops;
1006 mtx_unlock(&filterops_lock);
1008 return sysfilt_ops[~filt].for_fop;
1012 kqueue_fo_release(int filt)
1015 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1018 mtx_lock(&filterops_lock);
1019 KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1020 ("filter object refcount not valid on release"));
1021 sysfilt_ops[~filt].for_refcnt--;
1022 mtx_unlock(&filterops_lock);
1026 * A ref to kq (obtained via kqueue_acquire) must be held. waitok will
1027 * influence if memory allocation should wait. Make sure it is 0 if you
1031 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1033 struct filterops *fops;
1035 struct knote *kn, *tkn;
1036 cap_rights_t rights;
1037 int error, filt, event;
1038 int haskqglobal, filedesc_unlock;
1044 filedesc_unlock = 0;
1047 fops = kqueue_fo_find(filt);
1051 tkn = knote_alloc(waitok); /* prevent waiting with locks */
1055 KASSERT(td != NULL, ("td is NULL"));
1056 error = fget(td, kev->ident,
1057 cap_rights_init(&rights, CAP_EVENT), &fp);
1061 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1062 kev->ident, 0) != 0) {
1066 error = kqueue_expand(kq, fops, kev->ident, waitok);
1072 if (fp->f_type == DTYPE_KQUEUE) {
1074 * if we add some inteligence about what we are doing,
1075 * we should be able to support events on ourselves.
1076 * We need to know when we are doing this to prevent
1077 * getting both the knlist lock and the kq lock since
1078 * they are the same thing.
1080 if (fp->f_data == kq) {
1086 * Pre-lock the filedesc before the global
1087 * lock mutex, see the comment in
1090 FILEDESC_XLOCK(td->td_proc->p_fd);
1091 filedesc_unlock = 1;
1092 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1096 if (kev->ident < kq->kq_knlistsize) {
1097 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1098 if (kev->filter == kn->kn_filter)
1102 if ((kev->flags & EV_ADD) == EV_ADD)
1103 kqueue_expand(kq, fops, kev->ident, waitok);
1106 if (kq->kq_knhashmask != 0) {
1109 list = &kq->kq_knhash[
1110 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1111 SLIST_FOREACH(kn, list, kn_link)
1112 if (kev->ident == kn->kn_id &&
1113 kev->filter == kn->kn_filter)
1118 /* knote is in the process of changing, wait for it to stablize. */
1119 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1120 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1121 if (filedesc_unlock) {
1122 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1123 filedesc_unlock = 0;
1125 kq->kq_state |= KQ_FLUXWAIT;
1126 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1135 * kn now contains the matching knote, or NULL if no match
1138 if (kev->flags & EV_ADD) {
1150 * apply reference counts to knote structure, and
1151 * do not release it at the end of this routine.
1156 kn->kn_sfflags = kev->fflags;
1157 kn->kn_sdata = kev->data;
1160 kn->kn_kevent = *kev;
1161 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1162 EV_ENABLE | EV_DISABLE);
1163 kn->kn_status = KN_INFLUX|KN_DETACHED;
1165 error = knote_attach(kn, kq);
1172 if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1179 /* No matching knote and the EV_ADD flag is not set. */
1186 if (kev->flags & EV_DELETE) {
1187 kn->kn_status |= KN_INFLUX;
1189 if (!(kn->kn_status & KN_DETACHED))
1190 kn->kn_fop->f_detach(kn);
1196 * The user may change some filter values after the initial EV_ADD,
1197 * but doing so will not reset any filter which has already been
1200 kn->kn_status |= KN_INFLUX | KN_SCAN;
1203 kn->kn_kevent.udata = kev->udata;
1204 if (!fops->f_isfd && fops->f_touch != NULL) {
1205 fops->f_touch(kn, kev, EVENT_REGISTER);
1207 kn->kn_sfflags = kev->fflags;
1208 kn->kn_sdata = kev->data;
1212 * We can get here with kn->kn_knlist == NULL. This can happen when
1213 * the initial attach event decides that the event is "completed"
1214 * already. i.e. filt_procattach is called on a zombie process. It
1215 * will call filt_proc which will remove it from the list, and NULL
1219 event = kn->kn_fop->f_event(kn, 0);
1222 KNOTE_ACTIVATE(kn, 1);
1223 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1226 if ((kev->flags & EV_DISABLE) &&
1227 ((kn->kn_status & KN_DISABLED) == 0)) {
1228 kn->kn_status |= KN_DISABLED;
1231 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1232 kn->kn_status &= ~KN_DISABLED;
1233 if ((kn->kn_status & KN_ACTIVE) &&
1234 ((kn->kn_status & KN_QUEUED) == 0))
1240 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1241 if (filedesc_unlock)
1242 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1248 kqueue_fo_release(filt);
1253 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1261 if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1265 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1276 kqueue_release(struct kqueue *kq, int locked)
1283 if (kq->kq_refcnt == 1)
1284 wakeup(&kq->kq_refcnt);
1290 kqueue_schedtask(struct kqueue *kq)
1294 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1295 ("scheduling kqueue task while draining"));
1297 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1298 taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1299 kq->kq_state |= KQ_TASKSCHED;
1304 * Expand the kq to make sure we have storage for fops/ident pair.
1306 * Return 0 on success (or no work necessary), return errno on failure.
1308 * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1309 * If kqueue_register is called from a non-fd context, there usually/should
1313 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1316 struct klist *list, *tmp_knhash, *to_free;
1317 u_long tmp_knhashmask;
1320 int mflag = waitok ? M_WAITOK : M_NOWAIT;
1327 if (kq->kq_knlistsize <= fd) {
1328 size = kq->kq_knlistsize;
1331 list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1335 if (kq->kq_knlistsize > fd) {
1339 if (kq->kq_knlist != NULL) {
1340 bcopy(kq->kq_knlist, list,
1341 kq->kq_knlistsize * sizeof(*list));
1342 to_free = kq->kq_knlist;
1343 kq->kq_knlist = NULL;
1345 bzero((caddr_t)list +
1346 kq->kq_knlistsize * sizeof(*list),
1347 (size - kq->kq_knlistsize) * sizeof(*list));
1348 kq->kq_knlistsize = size;
1349 kq->kq_knlist = list;
1354 if (kq->kq_knhashmask == 0) {
1355 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1357 if (tmp_knhash == NULL)
1360 if (kq->kq_knhashmask == 0) {
1361 kq->kq_knhash = tmp_knhash;
1362 kq->kq_knhashmask = tmp_knhashmask;
1364 to_free = tmp_knhash;
1369 free(to_free, M_KQUEUE);
1376 kqueue_task(void *arg, int pending)
1384 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1387 KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1389 kq->kq_state &= ~KQ_TASKSCHED;
1390 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1391 wakeup(&kq->kq_state);
1394 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1398 * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1399 * We treat KN_MARKER knotes as if they are INFLUX.
1402 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1403 const struct timespec *tsp, struct kevent *keva, struct thread *td)
1405 struct kevent *kevp;
1406 struct knote *kn, *marker;
1407 sbintime_t asbt, rsbt;
1408 int count, error, haskqglobal, influx, nkev, touch;
1420 if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1421 tsp->tv_nsec >= 1000000000) {
1425 if (timespecisset(tsp)) {
1426 if (tsp->tv_sec <= INT32_MAX) {
1427 rsbt = tstosbt(*tsp);
1428 if (TIMESEL(&asbt, rsbt))
1429 asbt += tc_tick_sbt;
1430 if (asbt <= INT64_MAX - rsbt)
1434 rsbt >>= tc_precexp;
1441 marker = knote_alloc(1);
1442 if (marker == NULL) {
1446 marker->kn_status = KN_MARKER;
1451 if (kq->kq_count == 0) {
1453 error = EWOULDBLOCK;
1455 kq->kq_state |= KQ_SLEEP;
1456 error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1457 "kqread", asbt, rsbt, C_ABSOLUTE);
1461 /* don't restart after signals... */
1462 if (error == ERESTART)
1464 else if (error == EWOULDBLOCK)
1469 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1473 kn = TAILQ_FIRST(&kq->kq_head);
1475 if ((kn->kn_status == KN_MARKER && kn != marker) ||
1476 (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1481 kq->kq_state |= KQ_FLUXWAIT;
1482 error = msleep(kq, &kq->kq_lock, PSOCK,
1487 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1488 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1489 kn->kn_status &= ~KN_QUEUED;
1495 if (count == maxevents)
1499 KASSERT((kn->kn_status & KN_INFLUX) == 0,
1500 ("KN_INFLUX set when not suppose to be"));
1502 if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1503 kn->kn_status &= ~KN_QUEUED;
1504 kn->kn_status |= KN_INFLUX;
1508 * We don't need to lock the list since we've marked
1511 if (!(kn->kn_status & KN_DETACHED))
1512 kn->kn_fop->f_detach(kn);
1516 } else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1517 kn->kn_status &= ~KN_QUEUED;
1518 kn->kn_status |= KN_INFLUX;
1522 * We don't need to lock the list since we've marked
1525 *kevp = kn->kn_kevent;
1526 if (!(kn->kn_status & KN_DETACHED))
1527 kn->kn_fop->f_detach(kn);
1532 kn->kn_status |= KN_INFLUX | KN_SCAN;
1534 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1535 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1537 if (kn->kn_fop->f_event(kn, 0) == 0) {
1539 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1541 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1548 touch = (!kn->kn_fop->f_isfd &&
1549 kn->kn_fop->f_touch != NULL);
1551 kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1553 *kevp = kn->kn_kevent;
1555 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1556 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1558 * Manually clear knotes who weren't
1561 if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1565 if (kn->kn_flags & EV_DISPATCH)
1566 kn->kn_status |= KN_DISABLED;
1567 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1570 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1572 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1577 /* we are returning a copy to the user */
1582 if (nkev == KQ_NEVENTS) {
1585 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1593 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1601 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1602 td->td_retval[0] = maxevents - count;
1608 * This could be expanded to call kqueue_scan, if desired.
1612 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
1613 int flags, struct thread *td)
1620 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1621 int flags, struct thread *td)
1628 kqueue_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1637 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1638 struct ucred *active_cred, struct thread *td)
1641 * Enabling sigio causes two major problems:
1642 * 1) infinite recursion:
1643 * Synopsys: kevent is being used to track signals and have FIOASYNC
1644 * set. On receipt of a signal this will cause a kqueue to recurse
1645 * into itself over and over. Sending the sigio causes the kqueue
1646 * to become ready, which in turn posts sigio again, forever.
1647 * Solution: this can be solved by setting a flag in the kqueue that
1648 * we have a SIGIO in progress.
1649 * 2) locking problems:
1650 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1651 * us above the proc and pgrp locks.
1652 * Solution: Post a signal using an async mechanism, being sure to
1653 * record a generation count in the delivery so that we do not deliver
1654 * a signal to the wrong process.
1656 * Note, these two mechanisms are somewhat mutually exclusive!
1665 kq->kq_state |= KQ_ASYNC;
1667 kq->kq_state &= ~KQ_ASYNC;
1672 return (fsetown(*(int *)data, &kq->kq_sigio));
1675 *(int *)data = fgetown(&kq->kq_sigio);
1685 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1692 if ((error = kqueue_acquire(fp, &kq)))
1696 if (events & (POLLIN | POLLRDNORM)) {
1698 revents |= events & (POLLIN | POLLRDNORM);
1700 selrecord(td, &kq->kq_sel);
1701 if (SEL_WAITING(&kq->kq_sel))
1702 kq->kq_state |= KQ_SEL;
1705 kqueue_release(kq, 1);
1712 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1716 bzero((void *)st, sizeof *st);
1718 * We no longer return kq_count because the unlocked value is useless.
1719 * If you spent all this time getting the count, why not spend your
1720 * syscall better by calling kevent?
1722 * XXX - This is needed for libc_r.
1724 st->st_mode = S_IFIFO;
1730 kqueue_close(struct file *fp, struct thread *td)
1732 struct kqueue *kq = fp->f_data;
1733 struct filedesc *fdp;
1737 int filedesc_unlock;
1739 if ((error = kqueue_acquire(fp, &kq)))
1742 filedesc_unlock = 0;
1745 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1746 ("kqueue already closing"));
1747 kq->kq_state |= KQ_CLOSING;
1748 if (kq->kq_refcnt > 1)
1749 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1751 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1754 KASSERT(knlist_empty(&kq->kq_sel.si_note),
1755 ("kqueue's knlist not empty"));
1757 for (i = 0; i < kq->kq_knlistsize; i++) {
1758 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1759 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1760 kq->kq_state |= KQ_FLUXWAIT;
1761 msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1764 kn->kn_status |= KN_INFLUX;
1766 if (!(kn->kn_status & KN_DETACHED))
1767 kn->kn_fop->f_detach(kn);
1772 if (kq->kq_knhashmask != 0) {
1773 for (i = 0; i <= kq->kq_knhashmask; i++) {
1774 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1775 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1776 kq->kq_state |= KQ_FLUXWAIT;
1777 msleep(kq, &kq->kq_lock, PSOCK,
1781 kn->kn_status |= KN_INFLUX;
1783 if (!(kn->kn_status & KN_DETACHED))
1784 kn->kn_fop->f_detach(kn);
1791 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1792 kq->kq_state |= KQ_TASKDRAIN;
1793 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
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;
1805 * We could be called due to the knote_drop() doing fdrop(),
1806 * called from kqueue_register(). In this case the global
1807 * lock is owned, and filedesc sx is locked before, to not
1808 * take the sleepable lock after non-sleepable.
1810 if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1811 FILEDESC_XLOCK(fdp);
1812 filedesc_unlock = 1;
1814 filedesc_unlock = 0;
1815 TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1816 if (filedesc_unlock)
1817 FILEDESC_XUNLOCK(fdp);
1819 seldrain(&kq->kq_sel);
1820 knlist_destroy(&kq->kq_sel.si_note);
1821 mtx_destroy(&kq->kq_lock);
1824 if (kq->kq_knhash != NULL)
1825 free(kq->kq_knhash, M_KQUEUE);
1826 if (kq->kq_knlist != NULL)
1827 free(kq->kq_knlist, M_KQUEUE);
1829 funsetown(&kq->kq_sigio);
1837 kqueue_wakeup(struct kqueue *kq)
1841 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1842 kq->kq_state &= ~KQ_SLEEP;
1845 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1846 selwakeuppri(&kq->kq_sel, PSOCK);
1847 if (!SEL_WAITING(&kq->kq_sel))
1848 kq->kq_state &= ~KQ_SEL;
1850 if (!knlist_empty(&kq->kq_sel.si_note))
1851 kqueue_schedtask(kq);
1852 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1853 pgsigio(&kq->kq_sigio, SIGIO, 0);
1858 * Walk down a list of knotes, activating them if their event has triggered.
1860 * There is a possibility to optimize in the case of one kq watching another.
1861 * Instead of scheduling a task to wake it up, you could pass enough state
1862 * down the chain to make up the parent kqueue. Make this code functional
1866 knote(struct knlist *list, long hint, int lockflags)
1869 struct knote *kn, *tkn;
1875 KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1877 if ((lockflags & KNF_LISTLOCKED) == 0)
1878 list->kl_lock(list->kl_lockarg);
1881 * If we unlock the list lock (and set KN_INFLUX), we can
1882 * eliminate the kqueue scheduling, but this will introduce
1883 * four lock/unlock's for each knote to test. Also, marker
1884 * would be needed to keep iteration position, since filters
1885 * or other threads could remove events.
1887 SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
1890 if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
1892 * Do not process the influx notes, except for
1893 * the influx coming from the kq unlock in the
1894 * kqueue_scan(). In the later case, we do
1895 * not interfere with the scan, since the code
1896 * fragment in kqueue_scan() locks the knlist,
1897 * and cannot proceed until we finished.
1900 } else if ((lockflags & KNF_NOKQLOCK) != 0) {
1901 kn->kn_status |= KN_INFLUX;
1903 error = kn->kn_fop->f_event(kn, hint);
1905 kn->kn_status &= ~KN_INFLUX;
1907 KNOTE_ACTIVATE(kn, 1);
1910 kn->kn_status |= KN_HASKQLOCK;
1911 if (kn->kn_fop->f_event(kn, hint))
1912 KNOTE_ACTIVATE(kn, 1);
1913 kn->kn_status &= ~KN_HASKQLOCK;
1917 if ((lockflags & KNF_LISTLOCKED) == 0)
1918 list->kl_unlock(list->kl_lockarg);
1922 * add a knote to a knlist
1925 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1927 KNL_ASSERT_LOCK(knl, islocked);
1928 KQ_NOTOWNED(kn->kn_kq);
1929 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1930 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1932 knl->kl_lock(knl->kl_lockarg);
1933 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1935 knl->kl_unlock(knl->kl_lockarg);
1937 kn->kn_knlist = knl;
1938 kn->kn_status &= ~KN_DETACHED;
1939 KQ_UNLOCK(kn->kn_kq);
1943 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1945 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1946 KNL_ASSERT_LOCK(knl, knlislocked);
1947 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1949 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1950 ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1952 knl->kl_lock(knl->kl_lockarg);
1953 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1954 kn->kn_knlist = NULL;
1956 knl->kl_unlock(knl->kl_lockarg);
1959 kn->kn_status |= KN_DETACHED;
1961 KQ_UNLOCK(kn->kn_kq);
1965 * remove knote from the specified knlist
1968 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1971 knlist_remove_kq(knl, kn, islocked, 0);
1975 * remove knote from the specified knlist while in f_event handler.
1978 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1981 knlist_remove_kq(knl, kn, 1,
1982 (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1986 knlist_empty(struct knlist *knl)
1989 KNL_ASSERT_LOCKED(knl);
1990 return SLIST_EMPTY(&knl->kl_list);
1993 static struct mtx knlist_lock;
1994 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
1996 static void knlist_mtx_lock(void *arg);
1997 static void knlist_mtx_unlock(void *arg);
2000 knlist_mtx_lock(void *arg)
2003 mtx_lock((struct mtx *)arg);
2007 knlist_mtx_unlock(void *arg)
2010 mtx_unlock((struct mtx *)arg);
2014 knlist_mtx_assert_locked(void *arg)
2017 mtx_assert((struct mtx *)arg, MA_OWNED);
2021 knlist_mtx_assert_unlocked(void *arg)
2024 mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2028 knlist_rw_rlock(void *arg)
2031 rw_rlock((struct rwlock *)arg);
2035 knlist_rw_runlock(void *arg)
2038 rw_runlock((struct rwlock *)arg);
2042 knlist_rw_assert_locked(void *arg)
2045 rw_assert((struct rwlock *)arg, RA_LOCKED);
2049 knlist_rw_assert_unlocked(void *arg)
2052 rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2056 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2057 void (*kl_unlock)(void *),
2058 void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2062 knl->kl_lockarg = &knlist_lock;
2064 knl->kl_lockarg = lock;
2066 if (kl_lock == NULL)
2067 knl->kl_lock = knlist_mtx_lock;
2069 knl->kl_lock = kl_lock;
2070 if (kl_unlock == NULL)
2071 knl->kl_unlock = knlist_mtx_unlock;
2073 knl->kl_unlock = kl_unlock;
2074 if (kl_assert_locked == NULL)
2075 knl->kl_assert_locked = knlist_mtx_assert_locked;
2077 knl->kl_assert_locked = kl_assert_locked;
2078 if (kl_assert_unlocked == NULL)
2079 knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2081 knl->kl_assert_unlocked = kl_assert_unlocked;
2083 SLIST_INIT(&knl->kl_list);
2087 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2090 knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2094 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2097 knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2098 knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2102 knlist_destroy(struct knlist *knl)
2107 * if we run across this error, we need to find the offending
2108 * driver and have it call knlist_clear or knlist_delete.
2110 if (!SLIST_EMPTY(&knl->kl_list))
2111 printf("WARNING: destroying knlist w/ knotes on it!\n");
2114 knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2115 SLIST_INIT(&knl->kl_list);
2119 * Even if we are locked, we may need to drop the lock to allow any influx
2120 * knotes time to "settle".
2123 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2125 struct knote *kn, *kn2;
2129 KNL_ASSERT_LOCKED(knl);
2131 KNL_ASSERT_UNLOCKED(knl);
2132 again: /* need to reacquire lock since we have dropped it */
2133 knl->kl_lock(knl->kl_lockarg);
2136 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2139 if ((kn->kn_status & KN_INFLUX)) {
2143 knlist_remove_kq(knl, kn, 1, 1);
2145 kn->kn_status |= KN_INFLUX | KN_DETACHED;
2149 /* Make sure cleared knotes disappear soon */
2150 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2156 if (!SLIST_EMPTY(&knl->kl_list)) {
2157 /* there are still KN_INFLUX remaining */
2158 kn = SLIST_FIRST(&knl->kl_list);
2161 KASSERT(kn->kn_status & KN_INFLUX,
2162 ("knote removed w/o list lock"));
2163 knl->kl_unlock(knl->kl_lockarg);
2164 kq->kq_state |= KQ_FLUXWAIT;
2165 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2171 KNL_ASSERT_LOCKED(knl);
2173 knl->kl_unlock(knl->kl_lockarg);
2174 KNL_ASSERT_UNLOCKED(knl);
2179 * Remove all knotes referencing a specified fd must be called with FILEDESC
2180 * lock. This prevents a race where a new fd comes along and occupies the
2181 * entry and we attach a knote to the fd.
2184 knote_fdclose(struct thread *td, int fd)
2186 struct filedesc *fdp = td->td_proc->p_fd;
2191 FILEDESC_XLOCK_ASSERT(fdp);
2194 * We shouldn't have to worry about new kevents appearing on fd
2195 * since filedesc is locked.
2197 TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2202 while (kq->kq_knlistsize > fd &&
2203 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2204 if (kn->kn_status & KN_INFLUX) {
2205 /* someone else might be waiting on our knote */
2208 kq->kq_state |= KQ_FLUXWAIT;
2209 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2212 kn->kn_status |= KN_INFLUX;
2214 if (!(kn->kn_status & KN_DETACHED))
2215 kn->kn_fop->f_detach(kn);
2225 knote_attach(struct knote *kn, struct kqueue *kq)
2229 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2232 if (kn->kn_fop->f_isfd) {
2233 if (kn->kn_id >= kq->kq_knlistsize)
2235 list = &kq->kq_knlist[kn->kn_id];
2237 if (kq->kq_knhash == NULL)
2239 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2242 SLIST_INSERT_HEAD(list, kn, kn_link);
2248 * knote must already have been detached using the f_detach method.
2249 * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2250 * to prevent other removal.
2253 knote_drop(struct knote *kn, struct thread *td)
2261 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2262 ("knote_drop called without KN_INFLUX set in kn_status"));
2265 if (kn->kn_fop->f_isfd)
2266 list = &kq->kq_knlist[kn->kn_id];
2268 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2270 if (!SLIST_EMPTY(list))
2271 SLIST_REMOVE(list, kn, knote, kn_link);
2272 if (kn->kn_status & KN_QUEUED)
2276 if (kn->kn_fop->f_isfd) {
2277 fdrop(kn->kn_fp, td);
2280 kqueue_fo_release(kn->kn_kevent.filter);
2286 knote_enqueue(struct knote *kn)
2288 struct kqueue *kq = kn->kn_kq;
2290 KQ_OWNED(kn->kn_kq);
2291 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2293 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2294 kn->kn_status |= KN_QUEUED;
2300 knote_dequeue(struct knote *kn)
2302 struct kqueue *kq = kn->kn_kq;
2304 KQ_OWNED(kn->kn_kq);
2305 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2307 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2308 kn->kn_status &= ~KN_QUEUED;
2316 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2317 NULL, NULL, UMA_ALIGN_PTR, 0);
2319 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2321 static struct knote *
2322 knote_alloc(int waitok)
2324 return ((struct knote *)uma_zalloc(knote_zone,
2325 (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2329 knote_free(struct knote *kn)
2332 uma_zfree(knote_zone, kn);
2336 * Register the kev w/ the kq specified by fd.
2339 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2343 cap_rights_t rights;
2346 error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2349 if ((error = kqueue_acquire(fp, &kq)) != 0)
2352 error = kqueue_register(kq, kev, td, waitok);
2354 kqueue_release(kq, 0);