2 * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
3 * Copyright 2004 John-Mark Gurney <jmg@FreeBSD.org>
4 * Copyright (c) 2009 Apple, Inc.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
32 #include "opt_ktrace.h"
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/capability.h>
37 #include <sys/kernel.h>
39 #include <sys/mutex.h>
41 #include <sys/malloc.h>
42 #include <sys/unistd.h>
44 #include <sys/filedesc.h>
45 #include <sys/filio.h>
46 #include <sys/fcntl.h>
47 #include <sys/kthread.h>
48 #include <sys/limits.h>
49 #include <sys/selinfo.h>
50 #include <sys/queue.h>
51 #include <sys/event.h>
52 #include <sys/eventvar.h>
54 #include <sys/protosw.h>
55 #include <sys/sigio.h>
56 #include <sys/signalvar.h>
57 #include <sys/socket.h>
58 #include <sys/socketvar.h>
60 #include <sys/sysctl.h>
61 #include <sys/sysproto.h>
62 #include <sys/syscallsubr.h>
63 #include <sys/taskqueue.h>
66 #include <sys/ktrace.h>
71 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
74 * This lock is used if multiple kq locks are required. This possibly
75 * should be made into a per proc lock.
77 static struct mtx kq_global;
78 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
79 #define KQ_GLOBAL_LOCK(lck, haslck) do { \
84 #define KQ_GLOBAL_UNLOCK(lck, haslck) do { \
90 TASKQUEUE_DEFINE_THREAD(kqueue);
92 static int kevent_copyout(void *arg, struct kevent *kevp, int count);
93 static int kevent_copyin(void *arg, struct kevent *kevp, int count);
94 static int kqueue_register(struct kqueue *kq, struct kevent *kev,
95 struct thread *td, int waitok);
96 static int kqueue_acquire(struct file *fp, struct kqueue **kqp);
97 static void kqueue_release(struct kqueue *kq, int locked);
98 static int kqueue_expand(struct kqueue *kq, struct filterops *fops,
99 uintptr_t ident, int waitok);
100 static void kqueue_task(void *arg, int pending);
101 static int kqueue_scan(struct kqueue *kq, int maxevents,
102 struct kevent_copyops *k_ops,
103 const struct timespec *timeout,
104 struct kevent *keva, struct thread *td);
105 static void kqueue_wakeup(struct kqueue *kq);
106 static struct filterops *kqueue_fo_find(int filt);
107 static void kqueue_fo_release(int filt);
109 static fo_rdwr_t kqueue_read;
110 static fo_rdwr_t kqueue_write;
111 static fo_truncate_t kqueue_truncate;
112 static fo_ioctl_t kqueue_ioctl;
113 static fo_poll_t kqueue_poll;
114 static fo_kqfilter_t kqueue_kqfilter;
115 static fo_stat_t kqueue_stat;
116 static fo_close_t kqueue_close;
118 static struct fileops kqueueops = {
119 .fo_read = kqueue_read,
120 .fo_write = kqueue_write,
121 .fo_truncate = kqueue_truncate,
122 .fo_ioctl = kqueue_ioctl,
123 .fo_poll = kqueue_poll,
124 .fo_kqfilter = kqueue_kqfilter,
125 .fo_stat = kqueue_stat,
126 .fo_close = kqueue_close,
127 .fo_chmod = invfo_chmod,
128 .fo_chown = invfo_chown,
131 static int knote_attach(struct knote *kn, struct kqueue *kq);
132 static void knote_drop(struct knote *kn, struct thread *td);
133 static void knote_enqueue(struct knote *kn);
134 static void knote_dequeue(struct knote *kn);
135 static void knote_init(void);
136 static struct knote *knote_alloc(int waitok);
137 static void knote_free(struct knote *kn);
139 static void filt_kqdetach(struct knote *kn);
140 static int filt_kqueue(struct knote *kn, long hint);
141 static int filt_procattach(struct knote *kn);
142 static void filt_procdetach(struct knote *kn);
143 static int filt_proc(struct knote *kn, long hint);
144 static int filt_fileattach(struct knote *kn);
145 static void filt_timerexpire(void *knx);
146 static int filt_timerattach(struct knote *kn);
147 static void filt_timerdetach(struct knote *kn);
148 static int filt_timer(struct knote *kn, long hint);
149 static int filt_userattach(struct knote *kn);
150 static void filt_userdetach(struct knote *kn);
151 static int filt_user(struct knote *kn, long hint);
152 static void filt_usertouch(struct knote *kn, struct kevent *kev,
155 static struct filterops file_filtops = {
157 .f_attach = filt_fileattach,
159 static struct filterops kqread_filtops = {
161 .f_detach = filt_kqdetach,
162 .f_event = filt_kqueue,
164 /* XXX - move to kern_proc.c? */
165 static struct filterops proc_filtops = {
167 .f_attach = filt_procattach,
168 .f_detach = filt_procdetach,
169 .f_event = filt_proc,
171 static struct filterops timer_filtops = {
173 .f_attach = filt_timerattach,
174 .f_detach = filt_timerdetach,
175 .f_event = filt_timer,
177 static struct filterops user_filtops = {
178 .f_attach = filt_userattach,
179 .f_detach = filt_userdetach,
180 .f_event = filt_user,
181 .f_touch = filt_usertouch,
184 static uma_zone_t knote_zone;
185 static int kq_ncallouts = 0;
186 static int kq_calloutmax = (4 * 1024);
187 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
188 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
190 /* XXX - ensure not KN_INFLUX?? */
191 #define KNOTE_ACTIVATE(kn, islock) do { \
193 mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED); \
195 KQ_LOCK((kn)->kn_kq); \
196 (kn)->kn_status |= KN_ACTIVE; \
197 if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \
198 knote_enqueue((kn)); \
200 KQ_UNLOCK((kn)->kn_kq); \
202 #define KQ_LOCK(kq) do { \
203 mtx_lock(&(kq)->kq_lock); \
205 #define KQ_FLUX_WAKEUP(kq) do { \
206 if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) { \
207 (kq)->kq_state &= ~KQ_FLUXWAIT; \
211 #define KQ_UNLOCK_FLUX(kq) do { \
212 KQ_FLUX_WAKEUP(kq); \
213 mtx_unlock(&(kq)->kq_lock); \
215 #define KQ_UNLOCK(kq) do { \
216 mtx_unlock(&(kq)->kq_lock); \
218 #define KQ_OWNED(kq) do { \
219 mtx_assert(&(kq)->kq_lock, MA_OWNED); \
221 #define KQ_NOTOWNED(kq) do { \
222 mtx_assert(&(kq)->kq_lock, MA_NOTOWNED); \
224 #define KN_LIST_LOCK(kn) do { \
225 if (kn->kn_knlist != NULL) \
226 kn->kn_knlist->kl_lock(kn->kn_knlist->kl_lockarg); \
228 #define KN_LIST_UNLOCK(kn) do { \
229 if (kn->kn_knlist != NULL) \
230 kn->kn_knlist->kl_unlock(kn->kn_knlist->kl_lockarg); \
232 #define KNL_ASSERT_LOCK(knl, islocked) do { \
234 KNL_ASSERT_LOCKED(knl); \
236 KNL_ASSERT_UNLOCKED(knl); \
239 #define KNL_ASSERT_LOCKED(knl) do { \
240 knl->kl_assert_locked((knl)->kl_lockarg); \
242 #define KNL_ASSERT_UNLOCKED(knl) do { \
243 knl->kl_assert_unlocked((knl)->kl_lockarg); \
245 #else /* !INVARIANTS */
246 #define KNL_ASSERT_LOCKED(knl) do {} while(0)
247 #define KNL_ASSERT_UNLOCKED(knl) do {} while (0)
248 #endif /* INVARIANTS */
250 #define KN_HASHSIZE 64 /* XXX should be tunable */
251 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask))
254 filt_nullattach(struct knote *kn)
260 struct filterops null_filtops = {
262 .f_attach = filt_nullattach,
265 /* XXX - make SYSINIT to add these, and move into respective modules. */
266 extern struct filterops sig_filtops;
267 extern struct filterops fs_filtops;
270 * Table for for all system-defined filters.
272 static struct mtx filterops_lock;
273 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
276 struct filterops *for_fop;
278 } sysfilt_ops[EVFILT_SYSCOUNT] = {
279 { &file_filtops }, /* EVFILT_READ */
280 { &file_filtops }, /* EVFILT_WRITE */
281 { &null_filtops }, /* EVFILT_AIO */
282 { &file_filtops }, /* EVFILT_VNODE */
283 { &proc_filtops }, /* EVFILT_PROC */
284 { &sig_filtops }, /* EVFILT_SIGNAL */
285 { &timer_filtops }, /* EVFILT_TIMER */
286 { &null_filtops }, /* former EVFILT_NETDEV */
287 { &fs_filtops }, /* EVFILT_FS */
288 { &null_filtops }, /* EVFILT_LIO */
289 { &user_filtops }, /* EVFILT_USER */
293 * Simple redirection for all cdevsw style objects to call their fo_kqfilter
297 filt_fileattach(struct knote *kn)
300 return (fo_kqfilter(kn->kn_fp, kn));
305 kqueue_kqfilter(struct file *fp, struct knote *kn)
307 struct kqueue *kq = kn->kn_fp->f_data;
309 if (kn->kn_filter != EVFILT_READ)
312 kn->kn_status |= KN_KQUEUE;
313 kn->kn_fop = &kqread_filtops;
314 knlist_add(&kq->kq_sel.si_note, kn, 0);
320 filt_kqdetach(struct knote *kn)
322 struct kqueue *kq = kn->kn_fp->f_data;
324 knlist_remove(&kq->kq_sel.si_note, kn, 0);
329 filt_kqueue(struct knote *kn, long hint)
331 struct kqueue *kq = kn->kn_fp->f_data;
333 kn->kn_data = kq->kq_count;
334 return (kn->kn_data > 0);
337 /* XXX - move to kern_proc.c? */
339 filt_procattach(struct knote *kn)
346 p = pfind(kn->kn_id);
347 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
348 p = zpfind(kn->kn_id);
350 } else if (p != NULL && (p->p_flag & P_WEXIT)) {
356 if ((error = p_cansee(curthread, p))) {
361 kn->kn_ptr.p_proc = p;
362 kn->kn_flags |= EV_CLEAR; /* automatically set */
365 * internal flag indicating registration done by kernel
367 if (kn->kn_flags & EV_FLAG1) {
368 kn->kn_data = kn->kn_sdata; /* ppid */
369 kn->kn_fflags = NOTE_CHILD;
370 kn->kn_flags &= ~EV_FLAG1;
374 knlist_add(&p->p_klist, kn, 1);
377 * Immediately activate any exit notes if the target process is a
378 * zombie. This is necessary to handle the case where the target
379 * process, e.g. a child, dies before the kevent is registered.
381 if (immediate && filt_proc(kn, NOTE_EXIT))
382 KNOTE_ACTIVATE(kn, 0);
390 * The knote may be attached to a different process, which may exit,
391 * leaving nothing for the knote to be attached to. So when the process
392 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
393 * it will be deleted when read out. However, as part of the knote deletion,
394 * this routine is called, so a check is needed to avoid actually performing
395 * a detach, because the original process does not exist any more.
397 /* XXX - move to kern_proc.c? */
399 filt_procdetach(struct knote *kn)
403 p = kn->kn_ptr.p_proc;
404 knlist_remove(&p->p_klist, kn, 0);
405 kn->kn_ptr.p_proc = NULL;
408 /* XXX - move to kern_proc.c? */
410 filt_proc(struct knote *kn, long hint)
412 struct proc *p = kn->kn_ptr.p_proc;
416 * mask off extra data
418 event = (u_int)hint & NOTE_PCTRLMASK;
421 * if the user is interested in this event, record it.
423 if (kn->kn_sfflags & event)
424 kn->kn_fflags |= event;
427 * process is gone, so flag the event as finished.
429 if (event == NOTE_EXIT) {
430 if (!(kn->kn_status & KN_DETACHED))
431 knlist_remove_inevent(&p->p_klist, kn);
432 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
433 kn->kn_ptr.p_proc = NULL;
434 if (kn->kn_fflags & NOTE_EXIT)
435 kn->kn_data = p->p_xstat;
436 if (kn->kn_fflags == 0)
437 kn->kn_flags |= EV_DROP;
441 return (kn->kn_fflags != 0);
445 * Called when the process forked. It mostly does the same as the
446 * knote(), activating all knotes registered to be activated when the
447 * process forked. Additionally, for each knote attached to the
448 * parent, check whether user wants to track the new process. If so
449 * attach a new knote to it, and immediately report an event with the
453 knote_fork(struct knlist *list, int pid)
462 list->kl_lock(list->kl_lockarg);
464 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
465 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX)
469 if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
475 * The same as knote(), activate the event.
477 if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
478 kn->kn_status |= KN_HASKQLOCK;
479 if (kn->kn_fop->f_event(kn, NOTE_FORK))
480 KNOTE_ACTIVATE(kn, 1);
481 kn->kn_status &= ~KN_HASKQLOCK;
487 * The NOTE_TRACK case. In addition to the activation
488 * of the event, we need to register new event to
489 * track the child. Drop the locks in preparation for
490 * the call to kqueue_register().
492 kn->kn_status |= KN_INFLUX;
494 list->kl_unlock(list->kl_lockarg);
497 * Activate existing knote and register a knote with
501 kev.filter = kn->kn_filter;
502 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
503 kev.fflags = kn->kn_sfflags;
504 kev.data = kn->kn_id; /* parent */
505 kev.udata = kn->kn_kevent.udata;/* preserve udata */
506 error = kqueue_register(kq, &kev, NULL, 0);
508 kn->kn_fflags |= NOTE_TRACKERR;
509 if (kn->kn_fop->f_event(kn, NOTE_FORK))
510 KNOTE_ACTIVATE(kn, 0);
512 kn->kn_status &= ~KN_INFLUX;
514 list->kl_lock(list->kl_lockarg);
516 list->kl_unlock(list->kl_lockarg);
520 * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
521 * interval timer support code.
524 timertoticks(intptr_t data)
529 tv.tv_sec = data / 1000;
530 tv.tv_usec = (data % 1000) * 1000;
531 tticks = tvtohz(&tv);
537 filt_timerexpire(void *knx)
539 struct knote *kn = knx;
540 struct callout *calloutp;
543 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */
546 * timertoticks() uses tvtohz() which always adds 1 to allow
547 * for the time until the next clock interrupt being strictly
548 * less than 1 clock tick. We don't want that here since we
549 * want to appear to be in sync with the clock interrupt even
550 * when we're delayed.
552 if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
553 calloutp = (struct callout *)kn->kn_hook;
554 callout_reset_curcpu(calloutp, timertoticks(kn->kn_sdata) - 1,
555 filt_timerexpire, kn);
560 * data contains amount of time to sleep, in milliseconds
563 filt_timerattach(struct knote *kn)
565 struct callout *calloutp;
567 atomic_add_int(&kq_ncallouts, 1);
569 if (kq_ncallouts >= kq_calloutmax) {
570 atomic_add_int(&kq_ncallouts, -1);
574 kn->kn_flags |= EV_CLEAR; /* automatically set */
575 kn->kn_status &= ~KN_DETACHED; /* knlist_add usually sets it */
576 calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
577 callout_init(calloutp, CALLOUT_MPSAFE);
578 kn->kn_hook = calloutp;
579 callout_reset_curcpu(calloutp, timertoticks(kn->kn_sdata),
580 filt_timerexpire, kn);
586 filt_timerdetach(struct knote *kn)
588 struct callout *calloutp;
590 calloutp = (struct callout *)kn->kn_hook;
591 callout_drain(calloutp);
592 free(calloutp, M_KQUEUE);
593 atomic_add_int(&kq_ncallouts, -1);
594 kn->kn_status |= KN_DETACHED; /* knlist_remove usually clears it */
598 filt_timer(struct knote *kn, long hint)
601 return (kn->kn_data != 0);
605 filt_userattach(struct knote *kn)
609 * EVFILT_USER knotes are not attached to anything in the kernel.
612 if (kn->kn_fflags & NOTE_TRIGGER)
620 filt_userdetach(__unused struct knote *kn)
624 * EVFILT_USER knotes are not attached to anything in the kernel.
629 filt_user(struct knote *kn, __unused long hint)
632 return (kn->kn_hookid);
636 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
642 if (kev->fflags & NOTE_TRIGGER)
645 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
646 kev->fflags &= NOTE_FFLAGSMASK;
652 kn->kn_sfflags &= kev->fflags;
656 kn->kn_sfflags |= kev->fflags;
660 kn->kn_sfflags = kev->fflags;
664 /* XXX Return error? */
667 kn->kn_sdata = kev->data;
668 if (kev->flags & EV_CLEAR) {
676 *kev = kn->kn_kevent;
677 kev->fflags = kn->kn_sfflags;
678 kev->data = kn->kn_sdata;
679 if (kn->kn_flags & EV_CLEAR) {
687 panic("filt_usertouch() - invalid type (%ld)", type);
693 sys_kqueue(struct thread *td, struct kqueue_args *uap)
695 struct filedesc *fdp;
700 fdp = td->td_proc->p_fd;
701 error = falloc(td, &fp, &fd, 0);
705 /* An extra reference on `nfp' has been held for us by falloc(). */
706 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
707 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
708 TAILQ_INIT(&kq->kq_head);
710 knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
711 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
714 TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
715 FILEDESC_XUNLOCK(fdp);
717 finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
720 td->td_retval[0] = fd;
725 #ifndef _SYS_SYSPROTO_H_
728 const struct kevent *changelist;
730 struct kevent *eventlist;
732 const struct timespec *timeout;
736 sys_kevent(struct thread *td, struct kevent_args *uap)
738 struct timespec ts, *tsp;
739 struct kevent_copyops k_ops = { uap,
746 struct uio *ktruioin = NULL;
747 struct uio *ktruioout = NULL;
750 if (uap->timeout != NULL) {
751 error = copyin(uap->timeout, &ts, sizeof(ts));
759 if (KTRPOINT(td, KTR_GENIO)) {
760 ktriov.iov_base = uap->changelist;
761 ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
762 ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
763 .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
765 ktruioin = cloneuio(&ktruio);
766 ktriov.iov_base = uap->eventlist;
767 ktriov.iov_len = uap->nevents * sizeof(struct kevent);
768 ktruioout = cloneuio(&ktruio);
772 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
776 if (ktruioin != NULL) {
777 ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
778 ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
779 ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
780 ktrgenio(uap->fd, UIO_READ, ktruioout, error);
788 * Copy 'count' items into the destination list pointed to by uap->eventlist.
791 kevent_copyout(void *arg, struct kevent *kevp, int count)
793 struct kevent_args *uap;
796 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
797 uap = (struct kevent_args *)arg;
799 error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
801 uap->eventlist += count;
806 * Copy 'count' items from the list pointed to by uap->changelist.
809 kevent_copyin(void *arg, struct kevent *kevp, int count)
811 struct kevent_args *uap;
814 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
815 uap = (struct kevent_args *)arg;
817 error = copyin(uap->changelist, kevp, count * sizeof *kevp);
819 uap->changelist += count;
824 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
825 struct kevent_copyops *k_ops, const struct timespec *timeout)
827 struct kevent keva[KQ_NEVENTS];
828 struct kevent *kevp, *changes;
831 int i, n, nerrors, error;
833 if ((error = fget(td, fd, CAP_POST_EVENT, &fp)) != 0)
835 if ((error = kqueue_acquire(fp, &kq)) != 0)
840 while (nchanges > 0) {
841 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
842 error = k_ops->k_copyin(k_ops->arg, keva, n);
846 for (i = 0; i < n; i++) {
850 kevp->flags &= ~EV_SYSFLAGS;
851 error = kqueue_register(kq, kevp, td, 1);
852 if (error || (kevp->flags & EV_RECEIPT)) {
854 kevp->flags = EV_ERROR;
856 (void) k_ops->k_copyout(k_ops->arg,
868 td->td_retval[0] = nerrors;
873 error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
875 kqueue_release(kq, 0);
882 kqueue_add_filteropts(int filt, struct filterops *filtops)
887 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
889 "trying to add a filterop that is out of range: %d is beyond %d\n",
890 ~filt, EVFILT_SYSCOUNT);
893 mtx_lock(&filterops_lock);
894 if (sysfilt_ops[~filt].for_fop != &null_filtops &&
895 sysfilt_ops[~filt].for_fop != NULL)
898 sysfilt_ops[~filt].for_fop = filtops;
899 sysfilt_ops[~filt].for_refcnt = 0;
901 mtx_unlock(&filterops_lock);
907 kqueue_del_filteropts(int filt)
912 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
915 mtx_lock(&filterops_lock);
916 if (sysfilt_ops[~filt].for_fop == &null_filtops ||
917 sysfilt_ops[~filt].for_fop == NULL)
919 else if (sysfilt_ops[~filt].for_refcnt != 0)
922 sysfilt_ops[~filt].for_fop = &null_filtops;
923 sysfilt_ops[~filt].for_refcnt = 0;
925 mtx_unlock(&filterops_lock);
930 static struct filterops *
931 kqueue_fo_find(int filt)
934 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
937 mtx_lock(&filterops_lock);
938 sysfilt_ops[~filt].for_refcnt++;
939 if (sysfilt_ops[~filt].for_fop == NULL)
940 sysfilt_ops[~filt].for_fop = &null_filtops;
941 mtx_unlock(&filterops_lock);
943 return sysfilt_ops[~filt].for_fop;
947 kqueue_fo_release(int filt)
950 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
953 mtx_lock(&filterops_lock);
954 KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
955 ("filter object refcount not valid on release"));
956 sysfilt_ops[~filt].for_refcnt--;
957 mtx_unlock(&filterops_lock);
961 * A ref to kq (obtained via kqueue_acquire) must be held. waitok will
962 * influence if memory allocation should wait. Make sure it is 0 if you
966 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
968 struct filterops *fops;
970 struct knote *kn, *tkn;
971 int error, filt, event;
972 int haskqglobal, filedesc_unlock;
981 fops = kqueue_fo_find(filt);
985 tkn = knote_alloc(waitok); /* prevent waiting with locks */
989 KASSERT(td != NULL, ("td is NULL"));
990 if (kev->ident > INT_MAX)
993 error = fget(td, kev->ident, CAP_POLL_EVENT, &fp);
997 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
998 kev->ident, 0) != 0) {
1002 error = kqueue_expand(kq, fops, kev->ident, waitok);
1008 if (fp->f_type == DTYPE_KQUEUE) {
1010 * if we add some inteligence about what we are doing,
1011 * we should be able to support events on ourselves.
1012 * We need to know when we are doing this to prevent
1013 * getting both the knlist lock and the kq lock since
1014 * they are the same thing.
1016 if (fp->f_data == kq) {
1022 * Pre-lock the filedesc before the global
1023 * lock mutex, see the comment in
1026 FILEDESC_XLOCK(td->td_proc->p_fd);
1027 filedesc_unlock = 1;
1028 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1032 if (kev->ident < kq->kq_knlistsize) {
1033 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1034 if (kev->filter == kn->kn_filter)
1038 if ((kev->flags & EV_ADD) == EV_ADD)
1039 kqueue_expand(kq, fops, kev->ident, waitok);
1042 if (kq->kq_knhashmask != 0) {
1045 list = &kq->kq_knhash[
1046 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1047 SLIST_FOREACH(kn, list, kn_link)
1048 if (kev->ident == kn->kn_id &&
1049 kev->filter == kn->kn_filter)
1054 /* knote is in the process of changing, wait for it to stablize. */
1055 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1056 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1057 if (filedesc_unlock) {
1058 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1059 filedesc_unlock = 0;
1061 kq->kq_state |= KQ_FLUXWAIT;
1062 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1071 * kn now contains the matching knote, or NULL if no match
1074 if (kev->flags & EV_ADD) {
1086 * apply reference counts to knote structure, and
1087 * do not release it at the end of this routine.
1092 kn->kn_sfflags = kev->fflags;
1093 kn->kn_sdata = kev->data;
1096 kn->kn_kevent = *kev;
1097 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1098 EV_ENABLE | EV_DISABLE);
1099 kn->kn_status = KN_INFLUX|KN_DETACHED;
1101 error = knote_attach(kn, kq);
1108 if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1115 /* No matching knote and the EV_ADD flag is not set. */
1122 if (kev->flags & EV_DELETE) {
1123 kn->kn_status |= KN_INFLUX;
1125 if (!(kn->kn_status & KN_DETACHED))
1126 kn->kn_fop->f_detach(kn);
1132 * The user may change some filter values after the initial EV_ADD,
1133 * but doing so will not reset any filter which has already been
1136 kn->kn_status |= KN_INFLUX | KN_SCAN;
1139 kn->kn_kevent.udata = kev->udata;
1140 if (!fops->f_isfd && fops->f_touch != NULL) {
1141 fops->f_touch(kn, kev, EVENT_REGISTER);
1143 kn->kn_sfflags = kev->fflags;
1144 kn->kn_sdata = kev->data;
1148 * We can get here with kn->kn_knlist == NULL. This can happen when
1149 * the initial attach event decides that the event is "completed"
1150 * already. i.e. filt_procattach is called on a zombie process. It
1151 * will call filt_proc which will remove it from the list, and NULL
1155 event = kn->kn_fop->f_event(kn, 0);
1158 KNOTE_ACTIVATE(kn, 1);
1159 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1162 if ((kev->flags & EV_DISABLE) &&
1163 ((kn->kn_status & KN_DISABLED) == 0)) {
1164 kn->kn_status |= KN_DISABLED;
1167 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1168 kn->kn_status &= ~KN_DISABLED;
1169 if ((kn->kn_status & KN_ACTIVE) &&
1170 ((kn->kn_status & KN_QUEUED) == 0))
1176 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1177 if (filedesc_unlock)
1178 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1184 kqueue_fo_release(filt);
1189 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1197 if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1201 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1212 kqueue_release(struct kqueue *kq, int locked)
1219 if (kq->kq_refcnt == 1)
1220 wakeup(&kq->kq_refcnt);
1226 kqueue_schedtask(struct kqueue *kq)
1230 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1231 ("scheduling kqueue task while draining"));
1233 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1234 taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1235 kq->kq_state |= KQ_TASKSCHED;
1240 * Expand the kq to make sure we have storage for fops/ident pair.
1242 * Return 0 on success (or no work necessary), return errno on failure.
1244 * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1245 * If kqueue_register is called from a non-fd context, there usually/should
1249 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1252 struct klist *list, *tmp_knhash, *to_free;
1253 u_long tmp_knhashmask;
1256 int mflag = waitok ? M_WAITOK : M_NOWAIT;
1263 if (kq->kq_knlistsize <= fd) {
1264 size = kq->kq_knlistsize;
1267 list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1271 if (kq->kq_knlistsize > fd) {
1275 if (kq->kq_knlist != NULL) {
1276 bcopy(kq->kq_knlist, list,
1277 kq->kq_knlistsize * sizeof(*list));
1278 to_free = kq->kq_knlist;
1279 kq->kq_knlist = NULL;
1281 bzero((caddr_t)list +
1282 kq->kq_knlistsize * sizeof(*list),
1283 (size - kq->kq_knlistsize) * sizeof(*list));
1284 kq->kq_knlistsize = size;
1285 kq->kq_knlist = list;
1290 if (kq->kq_knhashmask == 0) {
1291 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1293 if (tmp_knhash == NULL)
1296 if (kq->kq_knhashmask == 0) {
1297 kq->kq_knhash = tmp_knhash;
1298 kq->kq_knhashmask = tmp_knhashmask;
1300 to_free = tmp_knhash;
1305 free(to_free, M_KQUEUE);
1312 kqueue_task(void *arg, int pending)
1320 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1323 KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1325 kq->kq_state &= ~KQ_TASKSCHED;
1326 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1327 wakeup(&kq->kq_state);
1330 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1334 * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1335 * We treat KN_MARKER knotes as if they are INFLUX.
1338 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1339 const struct timespec *tsp, struct kevent *keva, struct thread *td)
1341 struct kevent *kevp;
1342 struct timeval atv, rtv, ttv;
1343 struct knote *kn, *marker;
1344 int count, timeout, nkev, error, influx;
1345 int haskqglobal, touch;
1356 TIMESPEC_TO_TIMEVAL(&atv, tsp);
1357 if (itimerfix(&atv)) {
1361 if (tsp->tv_sec == 0 && tsp->tv_nsec == 0)
1364 timeout = atv.tv_sec > 24 * 60 * 60 ?
1365 24 * 60 * 60 * hz : tvtohz(&atv);
1366 getmicrouptime(&rtv);
1367 timevaladd(&atv, &rtv);
1373 marker = knote_alloc(1);
1374 if (marker == NULL) {
1378 marker->kn_status = KN_MARKER;
1383 if (atv.tv_sec || atv.tv_usec) {
1384 getmicrouptime(&rtv);
1385 if (timevalcmp(&rtv, &atv, >=))
1388 timevalsub(&ttv, &rtv);
1389 timeout = ttv.tv_sec > 24 * 60 * 60 ?
1390 24 * 60 * 60 * hz : tvtohz(&ttv);
1395 if (kq->kq_count == 0) {
1397 error = EWOULDBLOCK;
1399 kq->kq_state |= KQ_SLEEP;
1400 error = msleep(kq, &kq->kq_lock, PSOCK | PCATCH,
1405 /* don't restart after signals... */
1406 if (error == ERESTART)
1408 else if (error == EWOULDBLOCK)
1413 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1417 kn = TAILQ_FIRST(&kq->kq_head);
1419 if ((kn->kn_status == KN_MARKER && kn != marker) ||
1420 (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1425 kq->kq_state |= KQ_FLUXWAIT;
1426 error = msleep(kq, &kq->kq_lock, PSOCK,
1431 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1432 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1433 kn->kn_status &= ~KN_QUEUED;
1439 if (count == maxevents)
1443 KASSERT((kn->kn_status & KN_INFLUX) == 0,
1444 ("KN_INFLUX set when not suppose to be"));
1446 if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1447 kn->kn_status &= ~KN_QUEUED;
1448 kn->kn_status |= KN_INFLUX;
1452 * We don't need to lock the list since we've marked
1455 if (!(kn->kn_status & KN_DETACHED))
1456 kn->kn_fop->f_detach(kn);
1460 } else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1461 kn->kn_status &= ~KN_QUEUED;
1462 kn->kn_status |= KN_INFLUX;
1466 * We don't need to lock the list since we've marked
1469 *kevp = kn->kn_kevent;
1470 if (!(kn->kn_status & KN_DETACHED))
1471 kn->kn_fop->f_detach(kn);
1476 kn->kn_status |= KN_INFLUX | KN_SCAN;
1478 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1479 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1481 if (kn->kn_fop->f_event(kn, 0) == 0) {
1483 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1485 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1492 touch = (!kn->kn_fop->f_isfd &&
1493 kn->kn_fop->f_touch != NULL);
1495 kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1497 *kevp = kn->kn_kevent;
1499 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1500 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1502 * Manually clear knotes who weren't
1505 if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1509 if (kn->kn_flags & EV_DISPATCH)
1510 kn->kn_status |= KN_DISABLED;
1511 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1514 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1516 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1521 /* we are returning a copy to the user */
1526 if (nkev == KQ_NEVENTS) {
1529 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1537 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1545 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1546 td->td_retval[0] = maxevents - count;
1552 * This could be expanded to call kqueue_scan, if desired.
1556 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
1557 int flags, struct thread *td)
1564 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1565 int flags, struct thread *td)
1572 kqueue_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1581 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1582 struct ucred *active_cred, struct thread *td)
1585 * Enabling sigio causes two major problems:
1586 * 1) infinite recursion:
1587 * Synopsys: kevent is being used to track signals and have FIOASYNC
1588 * set. On receipt of a signal this will cause a kqueue to recurse
1589 * into itself over and over. Sending the sigio causes the kqueue
1590 * to become ready, which in turn posts sigio again, forever.
1591 * Solution: this can be solved by setting a flag in the kqueue that
1592 * we have a SIGIO in progress.
1593 * 2) locking problems:
1594 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1595 * us above the proc and pgrp locks.
1596 * Solution: Post a signal using an async mechanism, being sure to
1597 * record a generation count in the delivery so that we do not deliver
1598 * a signal to the wrong process.
1600 * Note, these two mechanisms are somewhat mutually exclusive!
1609 kq->kq_state |= KQ_ASYNC;
1611 kq->kq_state &= ~KQ_ASYNC;
1616 return (fsetown(*(int *)data, &kq->kq_sigio));
1619 *(int *)data = fgetown(&kq->kq_sigio);
1629 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1636 if ((error = kqueue_acquire(fp, &kq)))
1640 if (events & (POLLIN | POLLRDNORM)) {
1642 revents |= events & (POLLIN | POLLRDNORM);
1644 selrecord(td, &kq->kq_sel);
1645 if (SEL_WAITING(&kq->kq_sel))
1646 kq->kq_state |= KQ_SEL;
1649 kqueue_release(kq, 1);
1656 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1660 bzero((void *)st, sizeof *st);
1662 * We no longer return kq_count because the unlocked value is useless.
1663 * If you spent all this time getting the count, why not spend your
1664 * syscall better by calling kevent?
1666 * XXX - This is needed for libc_r.
1668 st->st_mode = S_IFIFO;
1674 kqueue_close(struct file *fp, struct thread *td)
1676 struct kqueue *kq = fp->f_data;
1677 struct filedesc *fdp;
1681 int filedesc_unlock;
1683 if ((error = kqueue_acquire(fp, &kq)))
1686 filedesc_unlock = 0;
1689 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1690 ("kqueue already closing"));
1691 kq->kq_state |= KQ_CLOSING;
1692 if (kq->kq_refcnt > 1)
1693 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1695 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1698 KASSERT(knlist_empty(&kq->kq_sel.si_note),
1699 ("kqueue's knlist not empty"));
1701 for (i = 0; i < kq->kq_knlistsize; i++) {
1702 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1703 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1704 kq->kq_state |= KQ_FLUXWAIT;
1705 msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1708 kn->kn_status |= KN_INFLUX;
1710 if (!(kn->kn_status & KN_DETACHED))
1711 kn->kn_fop->f_detach(kn);
1716 if (kq->kq_knhashmask != 0) {
1717 for (i = 0; i <= kq->kq_knhashmask; i++) {
1718 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1719 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1720 kq->kq_state |= KQ_FLUXWAIT;
1721 msleep(kq, &kq->kq_lock, PSOCK,
1725 kn->kn_status |= KN_INFLUX;
1727 if (!(kn->kn_status & KN_DETACHED))
1728 kn->kn_fop->f_detach(kn);
1735 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1736 kq->kq_state |= KQ_TASKDRAIN;
1737 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1740 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1741 selwakeuppri(&kq->kq_sel, PSOCK);
1742 if (!SEL_WAITING(&kq->kq_sel))
1743 kq->kq_state &= ~KQ_SEL;
1749 * We could be called due to the knote_drop() doing fdrop(),
1750 * called from kqueue_register(). In this case the global
1751 * lock is owned, and filedesc sx is locked before, to not
1752 * take the sleepable lock after non-sleepable.
1754 if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1755 FILEDESC_XLOCK(fdp);
1756 filedesc_unlock = 1;
1758 filedesc_unlock = 0;
1759 TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1760 if (filedesc_unlock)
1761 FILEDESC_XUNLOCK(fdp);
1763 seldrain(&kq->kq_sel);
1764 knlist_destroy(&kq->kq_sel.si_note);
1765 mtx_destroy(&kq->kq_lock);
1768 if (kq->kq_knhash != NULL)
1769 free(kq->kq_knhash, M_KQUEUE);
1770 if (kq->kq_knlist != NULL)
1771 free(kq->kq_knlist, M_KQUEUE);
1773 funsetown(&kq->kq_sigio);
1781 kqueue_wakeup(struct kqueue *kq)
1785 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1786 kq->kq_state &= ~KQ_SLEEP;
1789 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1790 selwakeuppri(&kq->kq_sel, PSOCK);
1791 if (!SEL_WAITING(&kq->kq_sel))
1792 kq->kq_state &= ~KQ_SEL;
1794 if (!knlist_empty(&kq->kq_sel.si_note))
1795 kqueue_schedtask(kq);
1796 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1797 pgsigio(&kq->kq_sigio, SIGIO, 0);
1802 * Walk down a list of knotes, activating them if their event has triggered.
1804 * There is a possibility to optimize in the case of one kq watching another.
1805 * Instead of scheduling a task to wake it up, you could pass enough state
1806 * down the chain to make up the parent kqueue. Make this code functional
1810 knote(struct knlist *list, long hint, int lockflags)
1819 KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1821 if ((lockflags & KNF_LISTLOCKED) == 0)
1822 list->kl_lock(list->kl_lockarg);
1825 * If we unlock the list lock (and set KN_INFLUX), we can eliminate
1826 * the kqueue scheduling, but this will introduce four
1827 * lock/unlock's for each knote to test. If we do, continue to use
1828 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is
1829 * only safe if you want to remove the current item, which we are
1832 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
1835 if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
1837 * Do not process the influx notes, except for
1838 * the influx coming from the kq unlock in the
1839 * kqueue_scan(). In the later case, we do
1840 * not interfere with the scan, since the code
1841 * fragment in kqueue_scan() locks the knlist,
1842 * and cannot proceed until we finished.
1845 } else if ((lockflags & KNF_NOKQLOCK) != 0) {
1846 kn->kn_status |= KN_INFLUX;
1848 error = kn->kn_fop->f_event(kn, hint);
1850 kn->kn_status &= ~KN_INFLUX;
1852 KNOTE_ACTIVATE(kn, 1);
1855 kn->kn_status |= KN_HASKQLOCK;
1856 if (kn->kn_fop->f_event(kn, hint))
1857 KNOTE_ACTIVATE(kn, 1);
1858 kn->kn_status &= ~KN_HASKQLOCK;
1862 if ((lockflags & KNF_LISTLOCKED) == 0)
1863 list->kl_unlock(list->kl_lockarg);
1867 * add a knote to a knlist
1870 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1872 KNL_ASSERT_LOCK(knl, islocked);
1873 KQ_NOTOWNED(kn->kn_kq);
1874 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1875 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1877 knl->kl_lock(knl->kl_lockarg);
1878 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1880 knl->kl_unlock(knl->kl_lockarg);
1882 kn->kn_knlist = knl;
1883 kn->kn_status &= ~KN_DETACHED;
1884 KQ_UNLOCK(kn->kn_kq);
1888 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1890 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1891 KNL_ASSERT_LOCK(knl, knlislocked);
1892 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1894 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1895 ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1897 knl->kl_lock(knl->kl_lockarg);
1898 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1899 kn->kn_knlist = NULL;
1901 knl->kl_unlock(knl->kl_lockarg);
1904 kn->kn_status |= KN_DETACHED;
1906 KQ_UNLOCK(kn->kn_kq);
1910 * remove all knotes from a specified klist
1913 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1916 knlist_remove_kq(knl, kn, islocked, 0);
1920 * remove knote from a specified klist while in f_event handler.
1923 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1926 knlist_remove_kq(knl, kn, 1,
1927 (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1931 knlist_empty(struct knlist *knl)
1933 KNL_ASSERT_LOCKED(knl);
1934 return SLIST_EMPTY(&knl->kl_list);
1937 static struct mtx knlist_lock;
1938 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
1940 static void knlist_mtx_lock(void *arg);
1941 static void knlist_mtx_unlock(void *arg);
1944 knlist_mtx_lock(void *arg)
1946 mtx_lock((struct mtx *)arg);
1950 knlist_mtx_unlock(void *arg)
1952 mtx_unlock((struct mtx *)arg);
1956 knlist_mtx_assert_locked(void *arg)
1958 mtx_assert((struct mtx *)arg, MA_OWNED);
1962 knlist_mtx_assert_unlocked(void *arg)
1964 mtx_assert((struct mtx *)arg, MA_NOTOWNED);
1968 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
1969 void (*kl_unlock)(void *),
1970 void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
1974 knl->kl_lockarg = &knlist_lock;
1976 knl->kl_lockarg = lock;
1978 if (kl_lock == NULL)
1979 knl->kl_lock = knlist_mtx_lock;
1981 knl->kl_lock = kl_lock;
1982 if (kl_unlock == NULL)
1983 knl->kl_unlock = knlist_mtx_unlock;
1985 knl->kl_unlock = kl_unlock;
1986 if (kl_assert_locked == NULL)
1987 knl->kl_assert_locked = knlist_mtx_assert_locked;
1989 knl->kl_assert_locked = kl_assert_locked;
1990 if (kl_assert_unlocked == NULL)
1991 knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
1993 knl->kl_assert_unlocked = kl_assert_unlocked;
1995 SLIST_INIT(&knl->kl_list);
1999 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2002 knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2006 knlist_destroy(struct knlist *knl)
2011 * if we run across this error, we need to find the offending
2012 * driver and have it call knlist_clear.
2014 if (!SLIST_EMPTY(&knl->kl_list))
2015 printf("WARNING: destroying knlist w/ knotes on it!\n");
2018 knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2019 SLIST_INIT(&knl->kl_list);
2023 * Even if we are locked, we may need to drop the lock to allow any influx
2024 * knotes time to "settle".
2027 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2029 struct knote *kn, *kn2;
2033 KNL_ASSERT_LOCKED(knl);
2035 KNL_ASSERT_UNLOCKED(knl);
2036 again: /* need to reacquire lock since we have dropped it */
2037 knl->kl_lock(knl->kl_lockarg);
2040 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2043 if ((kn->kn_status & KN_INFLUX)) {
2047 knlist_remove_kq(knl, kn, 1, 1);
2049 kn->kn_status |= KN_INFLUX | KN_DETACHED;
2053 /* Make sure cleared knotes disappear soon */
2054 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2060 if (!SLIST_EMPTY(&knl->kl_list)) {
2061 /* there are still KN_INFLUX remaining */
2062 kn = SLIST_FIRST(&knl->kl_list);
2065 KASSERT(kn->kn_status & KN_INFLUX,
2066 ("knote removed w/o list lock"));
2067 knl->kl_unlock(knl->kl_lockarg);
2068 kq->kq_state |= KQ_FLUXWAIT;
2069 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2075 KNL_ASSERT_LOCKED(knl);
2077 knl->kl_unlock(knl->kl_lockarg);
2078 KNL_ASSERT_UNLOCKED(knl);
2083 * Remove all knotes referencing a specified fd must be called with FILEDESC
2084 * lock. This prevents a race where a new fd comes along and occupies the
2085 * entry and we attach a knote to the fd.
2088 knote_fdclose(struct thread *td, int fd)
2090 struct filedesc *fdp = td->td_proc->p_fd;
2095 FILEDESC_XLOCK_ASSERT(fdp);
2098 * We shouldn't have to worry about new kevents appearing on fd
2099 * since filedesc is locked.
2101 TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2106 while (kq->kq_knlistsize > fd &&
2107 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2108 if (kn->kn_status & KN_INFLUX) {
2109 /* someone else might be waiting on our knote */
2112 kq->kq_state |= KQ_FLUXWAIT;
2113 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2116 kn->kn_status |= KN_INFLUX;
2118 if (!(kn->kn_status & KN_DETACHED))
2119 kn->kn_fop->f_detach(kn);
2129 knote_attach(struct knote *kn, struct kqueue *kq)
2133 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2136 if (kn->kn_fop->f_isfd) {
2137 if (kn->kn_id >= kq->kq_knlistsize)
2139 list = &kq->kq_knlist[kn->kn_id];
2141 if (kq->kq_knhash == NULL)
2143 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2146 SLIST_INSERT_HEAD(list, kn, kn_link);
2152 * knote must already have been detached using the f_detach method.
2153 * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2154 * to prevent other removal.
2157 knote_drop(struct knote *kn, struct thread *td)
2165 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2166 ("knote_drop called without KN_INFLUX set in kn_status"));
2169 if (kn->kn_fop->f_isfd)
2170 list = &kq->kq_knlist[kn->kn_id];
2172 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2174 if (!SLIST_EMPTY(list))
2175 SLIST_REMOVE(list, kn, knote, kn_link);
2176 if (kn->kn_status & KN_QUEUED)
2180 if (kn->kn_fop->f_isfd) {
2181 fdrop(kn->kn_fp, td);
2184 kqueue_fo_release(kn->kn_kevent.filter);
2190 knote_enqueue(struct knote *kn)
2192 struct kqueue *kq = kn->kn_kq;
2194 KQ_OWNED(kn->kn_kq);
2195 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2197 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2198 kn->kn_status |= KN_QUEUED;
2204 knote_dequeue(struct knote *kn)
2206 struct kqueue *kq = kn->kn_kq;
2208 KQ_OWNED(kn->kn_kq);
2209 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2211 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2212 kn->kn_status &= ~KN_QUEUED;
2220 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2221 NULL, NULL, UMA_ALIGN_PTR, 0);
2223 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2225 static struct knote *
2226 knote_alloc(int waitok)
2228 return ((struct knote *)uma_zalloc(knote_zone,
2229 (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2233 knote_free(struct knote *kn)
2236 uma_zfree(knote_zone, kn);
2240 * Register the kev w/ the kq specified by fd.
2243 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2249 if ((error = fget(td, fd, CAP_POST_EVENT, &fp)) != 0)
2251 if ((error = kqueue_acquire(fp, &kq)) != 0)
2254 error = kqueue_register(kq, kev, td, waitok);
2256 kqueue_release(kq, 0);