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/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>
68 #include <machine/atomic.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 unsigned int kq_ncallouts = 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 for the
368 * purposes of getting a NOTE_CHILD notification.
370 if (kn->kn_flags & EV_FLAG2) {
371 kn->kn_flags &= ~EV_FLAG2;
372 kn->kn_data = kn->kn_sdata; /* ppid */
373 kn->kn_fflags = NOTE_CHILD;
374 kn->kn_sfflags &= ~NOTE_EXIT;
375 immediate = 1; /* Force immediate activation of child note. */
378 * Internal flag indicating registration done by kernel (for other than
381 if (kn->kn_flags & EV_FLAG1) {
382 kn->kn_flags &= ~EV_FLAG1;
386 knlist_add(&p->p_klist, kn, 1);
389 * Immediately activate any child notes or, in the case of a zombie
390 * target process, exit notes. The latter is necessary to handle the
391 * case where the target process, e.g. a child, dies before the kevent
394 if (immediate && filt_proc(kn, NOTE_EXIT))
395 KNOTE_ACTIVATE(kn, 0);
403 * The knote may be attached to a different process, which may exit,
404 * leaving nothing for the knote to be attached to. So when the process
405 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
406 * it will be deleted when read out. However, as part of the knote deletion,
407 * this routine is called, so a check is needed to avoid actually performing
408 * a detach, because the original process does not exist any more.
410 /* XXX - move to kern_proc.c? */
412 filt_procdetach(struct knote *kn)
416 p = kn->kn_ptr.p_proc;
417 knlist_remove(&p->p_klist, kn, 0);
418 kn->kn_ptr.p_proc = NULL;
421 /* XXX - move to kern_proc.c? */
423 filt_proc(struct knote *kn, long hint)
425 struct proc *p = kn->kn_ptr.p_proc;
429 * mask off extra data
431 event = (u_int)hint & NOTE_PCTRLMASK;
434 * if the user is interested in this event, record it.
436 if (kn->kn_sfflags & event)
437 kn->kn_fflags |= event;
440 * process is gone, so flag the event as finished.
442 if (event == NOTE_EXIT) {
443 if (!(kn->kn_status & KN_DETACHED))
444 knlist_remove_inevent(&p->p_klist, kn);
445 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
446 kn->kn_ptr.p_proc = NULL;
447 if (kn->kn_fflags & NOTE_EXIT)
448 kn->kn_data = p->p_xstat;
449 if (kn->kn_fflags == 0)
450 kn->kn_flags |= EV_DROP;
454 return (kn->kn_fflags != 0);
458 * Called when the process forked. It mostly does the same as the
459 * knote(), activating all knotes registered to be activated when the
460 * process forked. Additionally, for each knote attached to the
461 * parent, check whether user wants to track the new process. If so
462 * attach a new knote to it, and immediately report an event with the
466 knote_fork(struct knlist *list, int pid)
475 list->kl_lock(list->kl_lockarg);
477 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
478 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX)
482 if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
488 * The same as knote(), activate the event.
490 if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
491 kn->kn_status |= KN_HASKQLOCK;
492 if (kn->kn_fop->f_event(kn, NOTE_FORK))
493 KNOTE_ACTIVATE(kn, 1);
494 kn->kn_status &= ~KN_HASKQLOCK;
500 * The NOTE_TRACK case. In addition to the activation
501 * of the event, we need to register new events to
502 * track the child. Drop the locks in preparation for
503 * the call to kqueue_register().
505 kn->kn_status |= KN_INFLUX;
507 list->kl_unlock(list->kl_lockarg);
510 * Activate existing knote and register tracking knotes with
513 * First register a knote to get just the child notice. This
514 * must be a separate note from a potential NOTE_EXIT
515 * notification since both NOTE_CHILD and NOTE_EXIT are defined
516 * to use the data field (in conflicting ways).
519 kev.filter = kn->kn_filter;
520 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_ONESHOT | EV_FLAG2;
521 kev.fflags = kn->kn_sfflags;
522 kev.data = kn->kn_id; /* parent */
523 kev.udata = kn->kn_kevent.udata;/* preserve udata */
524 error = kqueue_register(kq, &kev, NULL, 0);
526 kn->kn_fflags |= NOTE_TRACKERR;
529 * Then register another knote to track other potential events
530 * from the new process.
533 kev.filter = kn->kn_filter;
534 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
535 kev.fflags = kn->kn_sfflags;
536 kev.data = kn->kn_id; /* parent */
537 kev.udata = kn->kn_kevent.udata;/* preserve udata */
538 error = kqueue_register(kq, &kev, NULL, 0);
540 kn->kn_fflags |= NOTE_TRACKERR;
541 if (kn->kn_fop->f_event(kn, NOTE_FORK))
542 KNOTE_ACTIVATE(kn, 0);
544 kn->kn_status &= ~KN_INFLUX;
546 list->kl_lock(list->kl_lockarg);
548 list->kl_unlock(list->kl_lockarg);
552 * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
553 * interval timer support code.
556 #define NOTE_TIMER_PRECMASK \
557 (NOTE_SECONDS | NOTE_MSECONDS | NOTE_USECONDS | NOTE_NSECONDS)
560 timer2sbintime(intptr_t data, int flags)
565 * Macros for converting to the fractional second portion of an
566 * sbintime_t using 64bit multiplication to improve precision.
568 #define NS_TO_SBT(ns) (((ns) * (((uint64_t)1 << 63) / 500000000)) >> 32)
569 #define US_TO_SBT(us) (((us) * (((uint64_t)1 << 63) / 500000)) >> 32)
570 #define MS_TO_SBT(ms) (((ms) * (((uint64_t)1 << 63) / 500)) >> 32)
571 switch (flags & NOTE_TIMER_PRECMASK) {
574 if (data > (SBT_MAX / SBT_1S))
577 return ((sbintime_t)data << 32);
578 case NOTE_MSECONDS: /* FALLTHROUGH */
583 if (secs > (SBT_MAX / SBT_1S))
586 return (secs << 32 | MS_TO_SBT(data % 1000));
588 return (MS_TO_SBT(data));
590 if (data >= 1000000) {
591 secs = data / 1000000;
593 if (secs > (SBT_MAX / SBT_1S))
596 return (secs << 32 | US_TO_SBT(data % 1000000));
598 return (US_TO_SBT(data));
600 if (data >= 1000000000) {
601 secs = data / 1000000000;
603 if (secs > (SBT_MAX / SBT_1S))
606 return (secs << 32 | US_TO_SBT(data % 1000000000));
608 return (NS_TO_SBT(data));
615 struct kq_timer_cb_data {
617 sbintime_t next; /* next timer event fires at */
618 sbintime_t to; /* precalculated timer period */
622 filt_timerexpire(void *knx)
625 struct kq_timer_cb_data *kc;
629 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */
631 if ((kn->kn_flags & EV_ONESHOT) != 0)
636 callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
637 PCPU_GET(cpuid), C_ABSOLUTE);
641 * data contains amount of time to sleep
644 filt_timerattach(struct knote *kn)
646 struct kq_timer_cb_data *kc;
648 unsigned int ncallouts;
650 if (kn->kn_sdata < 0)
652 if (kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
654 /* Only precision unit are supported in flags so far */
655 if ((kn->kn_sfflags & ~NOTE_TIMER_PRECMASK) != 0)
658 to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
663 ncallouts = kq_ncallouts;
664 if (ncallouts >= kq_calloutmax)
666 } while (!atomic_cmpset_int(&kq_ncallouts, ncallouts, ncallouts + 1));
668 kn->kn_flags |= EV_CLEAR; /* automatically set */
669 kn->kn_status &= ~KN_DETACHED; /* knlist_add clears it */
670 kn->kn_ptr.p_v = kc = malloc(sizeof(*kc), M_KQUEUE, M_WAITOK);
671 callout_init(&kc->c, 1);
672 kc->next = to + sbinuptime();
674 callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
675 PCPU_GET(cpuid), C_ABSOLUTE);
681 filt_timerdetach(struct knote *kn)
683 struct kq_timer_cb_data *kc;
687 callout_drain(&kc->c);
689 old = atomic_fetchadd_int(&kq_ncallouts, -1);
690 KASSERT(old > 0, ("Number of callouts cannot become negative"));
691 kn->kn_status |= KN_DETACHED; /* knlist_remove sets it */
695 filt_timer(struct knote *kn, long hint)
698 return (kn->kn_data != 0);
702 filt_userattach(struct knote *kn)
706 * EVFILT_USER knotes are not attached to anything in the kernel.
709 if (kn->kn_fflags & NOTE_TRIGGER)
717 filt_userdetach(__unused struct knote *kn)
721 * EVFILT_USER knotes are not attached to anything in the kernel.
726 filt_user(struct knote *kn, __unused long hint)
729 return (kn->kn_hookid);
733 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
739 if (kev->fflags & NOTE_TRIGGER)
742 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
743 kev->fflags &= NOTE_FFLAGSMASK;
749 kn->kn_sfflags &= kev->fflags;
753 kn->kn_sfflags |= kev->fflags;
757 kn->kn_sfflags = kev->fflags;
761 /* XXX Return error? */
764 kn->kn_sdata = kev->data;
765 if (kev->flags & EV_CLEAR) {
773 *kev = kn->kn_kevent;
774 kev->fflags = kn->kn_sfflags;
775 kev->data = kn->kn_sdata;
776 if (kn->kn_flags & EV_CLEAR) {
784 panic("filt_usertouch() - invalid type (%ld)", type);
790 sys_kqueue(struct thread *td, struct kqueue_args *uap)
793 return (kern_kqueue(td, 0));
797 kern_kqueue(struct thread *td, int flags)
799 struct filedesc *fdp;
804 fdp = td->td_proc->p_fd;
805 error = falloc(td, &fp, &fd, flags);
809 /* An extra reference on `fp' has been held for us by falloc(). */
810 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
811 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
812 TAILQ_INIT(&kq->kq_head);
814 knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
815 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
818 TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
819 FILEDESC_XUNLOCK(fdp);
821 finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
824 td->td_retval[0] = fd;
831 kev_iovlen(int n, u_int kgio)
834 if (n < 0 || n >= kgio / sizeof(struct kevent))
836 return (n * sizeof(struct kevent));
840 #ifndef _SYS_SYSPROTO_H_
843 const struct kevent *changelist;
845 struct kevent *eventlist;
847 const struct timespec *timeout;
851 sys_kevent(struct thread *td, struct kevent_args *uap)
853 struct timespec ts, *tsp;
854 struct kevent_copyops k_ops = { uap,
861 struct uio *ktruioin = NULL;
862 struct uio *ktruioout = NULL;
866 if (uap->timeout != NULL) {
867 error = copyin(uap->timeout, &ts, sizeof(ts));
875 if (KTRPOINT(td, KTR_GENIO)) {
876 kgio = ktr_geniosize;
877 ktriov.iov_base = uap->changelist;
878 ktriov.iov_len = kev_iovlen(uap->nchanges, kgio);
879 ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
880 .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
882 ktruioin = cloneuio(&ktruio);
883 ktriov.iov_base = uap->eventlist;
884 ktriov.iov_len = kev_iovlen(uap->nevents, kgio);
885 ktriov.iov_len = uap->nevents * sizeof(struct kevent);
886 ktruioout = cloneuio(&ktruio);
890 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
894 if (ktruioin != NULL) {
895 ktruioin->uio_resid = kev_iovlen(uap->nchanges, kgio);
896 ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
897 ktruioout->uio_resid = kev_iovlen(td->td_retval[0], kgio);
898 ktrgenio(uap->fd, UIO_READ, ktruioout, error);
906 * Copy 'count' items into the destination list pointed to by uap->eventlist.
909 kevent_copyout(void *arg, struct kevent *kevp, int count)
911 struct kevent_args *uap;
914 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
915 uap = (struct kevent_args *)arg;
917 error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
919 uap->eventlist += count;
924 * Copy 'count' items from the list pointed to by uap->changelist.
927 kevent_copyin(void *arg, struct kevent *kevp, int count)
929 struct kevent_args *uap;
932 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
933 uap = (struct kevent_args *)arg;
935 error = copyin(uap->changelist, kevp, count * sizeof *kevp);
937 uap->changelist += count;
942 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
943 struct kevent_copyops *k_ops, const struct timespec *timeout)
949 cap_rights_init(&rights);
951 cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
953 cap_rights_set(&rights, CAP_KQUEUE_EVENT);
954 error = fget(td, fd, &rights, &fp);
958 error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
965 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
966 struct kevent_copyops *k_ops, const struct timespec *timeout)
968 struct kevent keva[KQ_NEVENTS];
969 struct kevent *kevp, *changes;
971 int i, n, nerrors, error;
973 error = kqueue_acquire(fp, &kq);
979 while (nchanges > 0) {
980 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
981 error = k_ops->k_copyin(k_ops->arg, keva, n);
985 for (i = 0; i < n; i++) {
989 kevp->flags &= ~EV_SYSFLAGS;
990 error = kqueue_register(kq, kevp, td, 1);
991 if (error || (kevp->flags & EV_RECEIPT)) {
993 kevp->flags = EV_ERROR;
995 (void) k_ops->k_copyout(k_ops->arg,
1007 td->td_retval[0] = nerrors;
1012 error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
1014 kqueue_release(kq, 0);
1019 kqueue_add_filteropts(int filt, struct filterops *filtops)
1024 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1026 "trying to add a filterop that is out of range: %d is beyond %d\n",
1027 ~filt, EVFILT_SYSCOUNT);
1030 mtx_lock(&filterops_lock);
1031 if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1032 sysfilt_ops[~filt].for_fop != NULL)
1035 sysfilt_ops[~filt].for_fop = filtops;
1036 sysfilt_ops[~filt].for_refcnt = 0;
1038 mtx_unlock(&filterops_lock);
1044 kqueue_del_filteropts(int filt)
1049 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1052 mtx_lock(&filterops_lock);
1053 if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1054 sysfilt_ops[~filt].for_fop == NULL)
1056 else if (sysfilt_ops[~filt].for_refcnt != 0)
1059 sysfilt_ops[~filt].for_fop = &null_filtops;
1060 sysfilt_ops[~filt].for_refcnt = 0;
1062 mtx_unlock(&filterops_lock);
1067 static struct filterops *
1068 kqueue_fo_find(int filt)
1071 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1074 mtx_lock(&filterops_lock);
1075 sysfilt_ops[~filt].for_refcnt++;
1076 if (sysfilt_ops[~filt].for_fop == NULL)
1077 sysfilt_ops[~filt].for_fop = &null_filtops;
1078 mtx_unlock(&filterops_lock);
1080 return sysfilt_ops[~filt].for_fop;
1084 kqueue_fo_release(int filt)
1087 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1090 mtx_lock(&filterops_lock);
1091 KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1092 ("filter object refcount not valid on release"));
1093 sysfilt_ops[~filt].for_refcnt--;
1094 mtx_unlock(&filterops_lock);
1098 * A ref to kq (obtained via kqueue_acquire) must be held. waitok will
1099 * influence if memory allocation should wait. Make sure it is 0 if you
1103 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1105 struct filterops *fops;
1107 struct knote *kn, *tkn;
1108 cap_rights_t rights;
1109 int error, filt, event;
1110 int haskqglobal, filedesc_unlock;
1116 filedesc_unlock = 0;
1119 fops = kqueue_fo_find(filt);
1123 tkn = knote_alloc(waitok); /* prevent waiting with locks */
1127 KASSERT(td != NULL, ("td is NULL"));
1128 if (kev->ident > INT_MAX)
1131 error = fget(td, kev->ident,
1132 cap_rights_init(&rights, CAP_EVENT), &fp);
1136 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1137 kev->ident, 0) != 0) {
1141 error = kqueue_expand(kq, fops, kev->ident, waitok);
1147 if (fp->f_type == DTYPE_KQUEUE) {
1149 * If we add some intelligence about what we are doing,
1150 * we should be able to support events on ourselves.
1151 * We need to know when we are doing this to prevent
1152 * getting both the knlist lock and the kq lock since
1153 * they are the same thing.
1155 if (fp->f_data == kq) {
1161 * Pre-lock the filedesc before the global
1162 * lock mutex, see the comment in
1165 FILEDESC_XLOCK(td->td_proc->p_fd);
1166 filedesc_unlock = 1;
1167 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1171 if (kev->ident < kq->kq_knlistsize) {
1172 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1173 if (kev->filter == kn->kn_filter)
1177 if ((kev->flags & EV_ADD) == EV_ADD)
1178 kqueue_expand(kq, fops, kev->ident, waitok);
1183 * If possible, find an existing knote to use for this kevent.
1185 if (kev->filter == EVFILT_PROC &&
1186 (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1187 /* This is an internal creation of a process tracking
1188 * note. Don't attempt to coalesce this with an
1192 } else if (kq->kq_knhashmask != 0) {
1195 list = &kq->kq_knhash[
1196 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1197 SLIST_FOREACH(kn, list, kn_link)
1198 if (kev->ident == kn->kn_id &&
1199 kev->filter == kn->kn_filter)
1204 /* knote is in the process of changing, wait for it to stabilize. */
1205 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1206 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1207 if (filedesc_unlock) {
1208 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1209 filedesc_unlock = 0;
1211 kq->kq_state |= KQ_FLUXWAIT;
1212 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1221 * kn now contains the matching knote, or NULL if no match
1224 if (kev->flags & EV_ADD) {
1236 * apply reference counts to knote structure, and
1237 * do not release it at the end of this routine.
1242 kn->kn_sfflags = kev->fflags;
1243 kn->kn_sdata = kev->data;
1246 kn->kn_kevent = *kev;
1247 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1248 EV_ENABLE | EV_DISABLE);
1249 kn->kn_status = KN_INFLUX|KN_DETACHED;
1251 error = knote_attach(kn, kq);
1258 if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1265 /* No matching knote and the EV_ADD flag is not set. */
1272 if (kev->flags & EV_DELETE) {
1273 kn->kn_status |= KN_INFLUX;
1275 if (!(kn->kn_status & KN_DETACHED))
1276 kn->kn_fop->f_detach(kn);
1282 * The user may change some filter values after the initial EV_ADD,
1283 * but doing so will not reset any filter which has already been
1286 kn->kn_status |= KN_INFLUX | KN_SCAN;
1289 kn->kn_kevent.udata = kev->udata;
1290 if (!fops->f_isfd && fops->f_touch != NULL) {
1291 fops->f_touch(kn, kev, EVENT_REGISTER);
1293 kn->kn_sfflags = kev->fflags;
1294 kn->kn_sdata = kev->data;
1298 * We can get here with kn->kn_knlist == NULL. This can happen when
1299 * the initial attach event decides that the event is "completed"
1300 * already. i.e. filt_procattach is called on a zombie process. It
1301 * will call filt_proc which will remove it from the list, and NULL
1305 event = kn->kn_fop->f_event(kn, 0);
1308 KNOTE_ACTIVATE(kn, 1);
1309 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1312 if ((kev->flags & EV_DISABLE) &&
1313 ((kn->kn_status & KN_DISABLED) == 0)) {
1314 kn->kn_status |= KN_DISABLED;
1317 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1318 kn->kn_status &= ~KN_DISABLED;
1319 if ((kn->kn_status & KN_ACTIVE) &&
1320 ((kn->kn_status & KN_QUEUED) == 0))
1326 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1327 if (filedesc_unlock)
1328 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1334 kqueue_fo_release(filt);
1339 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1347 if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1351 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1362 kqueue_release(struct kqueue *kq, int locked)
1369 if (kq->kq_refcnt == 1)
1370 wakeup(&kq->kq_refcnt);
1376 kqueue_schedtask(struct kqueue *kq)
1380 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1381 ("scheduling kqueue task while draining"));
1383 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1384 taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1385 kq->kq_state |= KQ_TASKSCHED;
1390 * Expand the kq to make sure we have storage for fops/ident pair.
1392 * Return 0 on success (or no work necessary), return errno on failure.
1394 * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1395 * If kqueue_register is called from a non-fd context, there usually/should
1399 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1402 struct klist *list, *tmp_knhash, *to_free;
1403 u_long tmp_knhashmask;
1406 int mflag = waitok ? M_WAITOK : M_NOWAIT;
1413 if (kq->kq_knlistsize <= fd) {
1414 size = kq->kq_knlistsize;
1417 list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1421 if (kq->kq_knlistsize > fd) {
1425 if (kq->kq_knlist != NULL) {
1426 bcopy(kq->kq_knlist, list,
1427 kq->kq_knlistsize * sizeof(*list));
1428 to_free = kq->kq_knlist;
1429 kq->kq_knlist = NULL;
1431 bzero((caddr_t)list +
1432 kq->kq_knlistsize * sizeof(*list),
1433 (size - kq->kq_knlistsize) * sizeof(*list));
1434 kq->kq_knlistsize = size;
1435 kq->kq_knlist = list;
1440 if (kq->kq_knhashmask == 0) {
1441 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1443 if (tmp_knhash == NULL)
1446 if (kq->kq_knhashmask == 0) {
1447 kq->kq_knhash = tmp_knhash;
1448 kq->kq_knhashmask = tmp_knhashmask;
1450 to_free = tmp_knhash;
1455 free(to_free, M_KQUEUE);
1462 kqueue_task(void *arg, int pending)
1470 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1473 KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1475 kq->kq_state &= ~KQ_TASKSCHED;
1476 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1477 wakeup(&kq->kq_state);
1480 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1484 * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1485 * We treat KN_MARKER knotes as if they are INFLUX.
1488 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1489 const struct timespec *tsp, struct kevent *keva, struct thread *td)
1491 struct kevent *kevp;
1492 struct knote *kn, *marker;
1493 sbintime_t asbt, rsbt;
1494 int count, error, haskqglobal, influx, nkev, touch;
1506 if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1507 tsp->tv_nsec >= 1000000000) {
1511 if (timespecisset(tsp)) {
1512 if (tsp->tv_sec <= INT32_MAX) {
1513 rsbt = tstosbt(*tsp);
1514 if (TIMESEL(&asbt, rsbt))
1515 asbt += tc_tick_sbt;
1516 if (asbt <= SBT_MAX - rsbt)
1520 rsbt >>= tc_precexp;
1527 marker = knote_alloc(1);
1528 if (marker == NULL) {
1532 marker->kn_status = KN_MARKER;
1537 if (kq->kq_count == 0) {
1539 error = EWOULDBLOCK;
1541 kq->kq_state |= KQ_SLEEP;
1542 error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1543 "kqread", asbt, rsbt, C_ABSOLUTE);
1547 /* don't restart after signals... */
1548 if (error == ERESTART)
1550 else if (error == EWOULDBLOCK)
1555 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1559 kn = TAILQ_FIRST(&kq->kq_head);
1561 if ((kn->kn_status == KN_MARKER && kn != marker) ||
1562 (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1567 kq->kq_state |= KQ_FLUXWAIT;
1568 error = msleep(kq, &kq->kq_lock, PSOCK,
1573 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1574 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1575 kn->kn_status &= ~KN_QUEUED;
1581 if (count == maxevents)
1585 KASSERT((kn->kn_status & KN_INFLUX) == 0,
1586 ("KN_INFLUX set when not suppose to be"));
1588 if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1589 kn->kn_status &= ~KN_QUEUED;
1590 kn->kn_status |= KN_INFLUX;
1594 * We don't need to lock the list since we've marked
1597 if (!(kn->kn_status & KN_DETACHED))
1598 kn->kn_fop->f_detach(kn);
1602 } else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1603 kn->kn_status &= ~KN_QUEUED;
1604 kn->kn_status |= KN_INFLUX;
1608 * We don't need to lock the list since we've marked
1611 *kevp = kn->kn_kevent;
1612 if (!(kn->kn_status & KN_DETACHED))
1613 kn->kn_fop->f_detach(kn);
1618 kn->kn_status |= KN_INFLUX | KN_SCAN;
1620 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1621 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1623 if (kn->kn_fop->f_event(kn, 0) == 0) {
1625 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1627 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1634 touch = (!kn->kn_fop->f_isfd &&
1635 kn->kn_fop->f_touch != NULL);
1637 kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1639 *kevp = kn->kn_kevent;
1641 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1642 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1644 * Manually clear knotes who weren't
1647 if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1651 if (kn->kn_flags & EV_DISPATCH)
1652 kn->kn_status |= KN_DISABLED;
1653 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1656 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1658 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1663 /* we are returning a copy to the user */
1668 if (nkev == KQ_NEVENTS) {
1671 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1679 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1687 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1688 td->td_retval[0] = maxevents - count;
1694 * This could be expanded to call kqueue_scan, if desired.
1698 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
1699 int flags, struct thread *td)
1706 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1707 int flags, struct thread *td)
1714 kqueue_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1723 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1724 struct ucred *active_cred, struct thread *td)
1727 * Enabling sigio causes two major problems:
1728 * 1) infinite recursion:
1729 * Synopsys: kevent is being used to track signals and have FIOASYNC
1730 * set. On receipt of a signal this will cause a kqueue to recurse
1731 * into itself over and over. Sending the sigio causes the kqueue
1732 * to become ready, which in turn posts sigio again, forever.
1733 * Solution: this can be solved by setting a flag in the kqueue that
1734 * we have a SIGIO in progress.
1735 * 2) locking problems:
1736 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1737 * us above the proc and pgrp locks.
1738 * Solution: Post a signal using an async mechanism, being sure to
1739 * record a generation count in the delivery so that we do not deliver
1740 * a signal to the wrong process.
1742 * Note, these two mechanisms are somewhat mutually exclusive!
1751 kq->kq_state |= KQ_ASYNC;
1753 kq->kq_state &= ~KQ_ASYNC;
1758 return (fsetown(*(int *)data, &kq->kq_sigio));
1761 *(int *)data = fgetown(&kq->kq_sigio);
1771 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1778 if ((error = kqueue_acquire(fp, &kq)))
1782 if (events & (POLLIN | POLLRDNORM)) {
1784 revents |= events & (POLLIN | POLLRDNORM);
1786 selrecord(td, &kq->kq_sel);
1787 if (SEL_WAITING(&kq->kq_sel))
1788 kq->kq_state |= KQ_SEL;
1791 kqueue_release(kq, 1);
1798 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1802 bzero((void *)st, sizeof *st);
1804 * We no longer return kq_count because the unlocked value is useless.
1805 * If you spent all this time getting the count, why not spend your
1806 * syscall better by calling kevent?
1808 * XXX - This is needed for libc_r.
1810 st->st_mode = S_IFIFO;
1816 kqueue_close(struct file *fp, struct thread *td)
1818 struct kqueue *kq = fp->f_data;
1819 struct filedesc *fdp;
1823 int filedesc_unlock;
1825 if ((error = kqueue_acquire(fp, &kq)))
1828 filedesc_unlock = 0;
1831 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1832 ("kqueue already closing"));
1833 kq->kq_state |= KQ_CLOSING;
1834 if (kq->kq_refcnt > 1)
1835 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1837 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1840 KASSERT(knlist_empty(&kq->kq_sel.si_note),
1841 ("kqueue's knlist not empty"));
1843 for (i = 0; i < kq->kq_knlistsize; i++) {
1844 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1845 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1846 kq->kq_state |= KQ_FLUXWAIT;
1847 msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1850 kn->kn_status |= KN_INFLUX;
1852 if (!(kn->kn_status & KN_DETACHED))
1853 kn->kn_fop->f_detach(kn);
1858 if (kq->kq_knhashmask != 0) {
1859 for (i = 0; i <= kq->kq_knhashmask; i++) {
1860 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1861 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1862 kq->kq_state |= KQ_FLUXWAIT;
1863 msleep(kq, &kq->kq_lock, PSOCK,
1867 kn->kn_status |= KN_INFLUX;
1869 if (!(kn->kn_status & KN_DETACHED))
1870 kn->kn_fop->f_detach(kn);
1877 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1878 kq->kq_state |= KQ_TASKDRAIN;
1879 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1882 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1883 selwakeuppri(&kq->kq_sel, PSOCK);
1884 if (!SEL_WAITING(&kq->kq_sel))
1885 kq->kq_state &= ~KQ_SEL;
1891 * We could be called due to the knote_drop() doing fdrop(),
1892 * called from kqueue_register(). In this case the global
1893 * lock is owned, and filedesc sx is locked before, to not
1894 * take the sleepable lock after non-sleepable.
1896 if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1897 FILEDESC_XLOCK(fdp);
1898 filedesc_unlock = 1;
1900 filedesc_unlock = 0;
1901 TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1902 if (filedesc_unlock)
1903 FILEDESC_XUNLOCK(fdp);
1905 seldrain(&kq->kq_sel);
1906 knlist_destroy(&kq->kq_sel.si_note);
1907 mtx_destroy(&kq->kq_lock);
1910 if (kq->kq_knhash != NULL)
1911 free(kq->kq_knhash, M_KQUEUE);
1912 if (kq->kq_knlist != NULL)
1913 free(kq->kq_knlist, M_KQUEUE);
1915 funsetown(&kq->kq_sigio);
1923 kqueue_wakeup(struct kqueue *kq)
1927 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1928 kq->kq_state &= ~KQ_SLEEP;
1931 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1932 selwakeuppri(&kq->kq_sel, PSOCK);
1933 if (!SEL_WAITING(&kq->kq_sel))
1934 kq->kq_state &= ~KQ_SEL;
1936 if (!knlist_empty(&kq->kq_sel.si_note))
1937 kqueue_schedtask(kq);
1938 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1939 pgsigio(&kq->kq_sigio, SIGIO, 0);
1944 * Walk down a list of knotes, activating them if their event has triggered.
1946 * There is a possibility to optimize in the case of one kq watching another.
1947 * Instead of scheduling a task to wake it up, you could pass enough state
1948 * down the chain to make up the parent kqueue. Make this code functional
1952 knote(struct knlist *list, long hint, int lockflags)
1955 struct knote *kn, *tkn;
1962 KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1964 if ((lockflags & KNF_LISTLOCKED) == 0)
1965 list->kl_lock(list->kl_lockarg);
1968 * If we unlock the list lock (and set KN_INFLUX), we can
1969 * eliminate the kqueue scheduling, but this will introduce
1970 * four lock/unlock's for each knote to test. Also, marker
1971 * would be needed to keep iteration position, since filters
1972 * or other threads could remove events.
1974 SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
1977 if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
1979 * Do not process the influx notes, except for
1980 * the influx coming from the kq unlock in the
1981 * kqueue_scan(). In the later case, we do
1982 * not interfere with the scan, since the code
1983 * fragment in kqueue_scan() locks the knlist,
1984 * and cannot proceed until we finished.
1987 } else if ((lockflags & KNF_NOKQLOCK) != 0) {
1988 own_influx = (kn->kn_status & KN_INFLUX) == 0;
1990 kn->kn_status |= KN_INFLUX;
1992 error = kn->kn_fop->f_event(kn, hint);
1995 kn->kn_status &= ~KN_INFLUX;
1997 KNOTE_ACTIVATE(kn, 1);
2000 kn->kn_status |= KN_HASKQLOCK;
2001 if (kn->kn_fop->f_event(kn, hint))
2002 KNOTE_ACTIVATE(kn, 1);
2003 kn->kn_status &= ~KN_HASKQLOCK;
2007 if ((lockflags & KNF_LISTLOCKED) == 0)
2008 list->kl_unlock(list->kl_lockarg);
2012 * add a knote to a knlist
2015 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2017 KNL_ASSERT_LOCK(knl, islocked);
2018 KQ_NOTOWNED(kn->kn_kq);
2019 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
2020 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
2022 knl->kl_lock(knl->kl_lockarg);
2023 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2025 knl->kl_unlock(knl->kl_lockarg);
2027 kn->kn_knlist = knl;
2028 kn->kn_status &= ~KN_DETACHED;
2029 KQ_UNLOCK(kn->kn_kq);
2033 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
2035 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
2036 KNL_ASSERT_LOCK(knl, knlislocked);
2037 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2039 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
2040 ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
2042 knl->kl_lock(knl->kl_lockarg);
2043 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2044 kn->kn_knlist = NULL;
2046 knl->kl_unlock(knl->kl_lockarg);
2049 kn->kn_status |= KN_DETACHED;
2051 KQ_UNLOCK(kn->kn_kq);
2055 * remove knote from the specified knlist
2058 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2061 knlist_remove_kq(knl, kn, islocked, 0);
2065 * remove knote from the specified knlist while in f_event handler.
2068 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
2071 knlist_remove_kq(knl, kn, 1,
2072 (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
2076 knlist_empty(struct knlist *knl)
2079 KNL_ASSERT_LOCKED(knl);
2080 return (SLIST_EMPTY(&knl->kl_list));
2083 static struct mtx knlist_lock;
2084 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2086 static void knlist_mtx_lock(void *arg);
2087 static void knlist_mtx_unlock(void *arg);
2090 knlist_mtx_lock(void *arg)
2093 mtx_lock((struct mtx *)arg);
2097 knlist_mtx_unlock(void *arg)
2100 mtx_unlock((struct mtx *)arg);
2104 knlist_mtx_assert_locked(void *arg)
2107 mtx_assert((struct mtx *)arg, MA_OWNED);
2111 knlist_mtx_assert_unlocked(void *arg)
2114 mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2118 knlist_rw_rlock(void *arg)
2121 rw_rlock((struct rwlock *)arg);
2125 knlist_rw_runlock(void *arg)
2128 rw_runlock((struct rwlock *)arg);
2132 knlist_rw_assert_locked(void *arg)
2135 rw_assert((struct rwlock *)arg, RA_LOCKED);
2139 knlist_rw_assert_unlocked(void *arg)
2142 rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2146 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2147 void (*kl_unlock)(void *),
2148 void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2152 knl->kl_lockarg = &knlist_lock;
2154 knl->kl_lockarg = lock;
2156 if (kl_lock == NULL)
2157 knl->kl_lock = knlist_mtx_lock;
2159 knl->kl_lock = kl_lock;
2160 if (kl_unlock == NULL)
2161 knl->kl_unlock = knlist_mtx_unlock;
2163 knl->kl_unlock = kl_unlock;
2164 if (kl_assert_locked == NULL)
2165 knl->kl_assert_locked = knlist_mtx_assert_locked;
2167 knl->kl_assert_locked = kl_assert_locked;
2168 if (kl_assert_unlocked == NULL)
2169 knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2171 knl->kl_assert_unlocked = kl_assert_unlocked;
2173 SLIST_INIT(&knl->kl_list);
2177 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2180 knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2184 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2187 knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2188 knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2192 knlist_destroy(struct knlist *knl)
2195 KASSERT(KNLIST_EMPTY(knl),
2196 ("destroying knlist %p with knotes on it", knl));
2200 * Even if we are locked, we may need to drop the lock to allow any influx
2201 * knotes time to "settle".
2204 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2206 struct knote *kn, *kn2;
2210 KNL_ASSERT_LOCKED(knl);
2212 KNL_ASSERT_UNLOCKED(knl);
2213 again: /* need to reacquire lock since we have dropped it */
2214 knl->kl_lock(knl->kl_lockarg);
2217 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2220 if ((kn->kn_status & KN_INFLUX)) {
2224 knlist_remove_kq(knl, kn, 1, 1);
2226 kn->kn_status |= KN_INFLUX | KN_DETACHED;
2230 /* Make sure cleared knotes disappear soon */
2231 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2237 if (!SLIST_EMPTY(&knl->kl_list)) {
2238 /* there are still KN_INFLUX remaining */
2239 kn = SLIST_FIRST(&knl->kl_list);
2242 KASSERT(kn->kn_status & KN_INFLUX,
2243 ("knote removed w/o list lock"));
2244 knl->kl_unlock(knl->kl_lockarg);
2245 kq->kq_state |= KQ_FLUXWAIT;
2246 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2252 KNL_ASSERT_LOCKED(knl);
2254 knl->kl_unlock(knl->kl_lockarg);
2255 KNL_ASSERT_UNLOCKED(knl);
2260 * Remove all knotes referencing a specified fd must be called with FILEDESC
2261 * lock. This prevents a race where a new fd comes along and occupies the
2262 * entry and we attach a knote to the fd.
2265 knote_fdclose(struct thread *td, int fd)
2267 struct filedesc *fdp = td->td_proc->p_fd;
2272 FILEDESC_XLOCK_ASSERT(fdp);
2275 * We shouldn't have to worry about new kevents appearing on fd
2276 * since filedesc is locked.
2278 TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2283 while (kq->kq_knlistsize > fd &&
2284 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2285 if (kn->kn_status & KN_INFLUX) {
2286 /* someone else might be waiting on our knote */
2289 kq->kq_state |= KQ_FLUXWAIT;
2290 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2293 kn->kn_status |= KN_INFLUX;
2295 if (!(kn->kn_status & KN_DETACHED))
2296 kn->kn_fop->f_detach(kn);
2306 knote_attach(struct knote *kn, struct kqueue *kq)
2310 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2313 if (kn->kn_fop->f_isfd) {
2314 if (kn->kn_id >= kq->kq_knlistsize)
2316 list = &kq->kq_knlist[kn->kn_id];
2318 if (kq->kq_knhash == NULL)
2320 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2322 SLIST_INSERT_HEAD(list, kn, kn_link);
2327 * knote must already have been detached using the f_detach method.
2328 * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2329 * to prevent other removal.
2332 knote_drop(struct knote *kn, struct thread *td)
2340 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2341 ("knote_drop called without KN_INFLUX set in kn_status"));
2344 if (kn->kn_fop->f_isfd)
2345 list = &kq->kq_knlist[kn->kn_id];
2347 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2349 if (!SLIST_EMPTY(list))
2350 SLIST_REMOVE(list, kn, knote, kn_link);
2351 if (kn->kn_status & KN_QUEUED)
2355 if (kn->kn_fop->f_isfd) {
2356 fdrop(kn->kn_fp, td);
2359 kqueue_fo_release(kn->kn_kevent.filter);
2365 knote_enqueue(struct knote *kn)
2367 struct kqueue *kq = kn->kn_kq;
2369 KQ_OWNED(kn->kn_kq);
2370 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2372 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2373 kn->kn_status |= KN_QUEUED;
2379 knote_dequeue(struct knote *kn)
2381 struct kqueue *kq = kn->kn_kq;
2383 KQ_OWNED(kn->kn_kq);
2384 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2386 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2387 kn->kn_status &= ~KN_QUEUED;
2395 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2396 NULL, NULL, UMA_ALIGN_PTR, 0);
2398 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2400 static struct knote *
2401 knote_alloc(int waitok)
2403 return ((struct knote *)uma_zalloc(knote_zone,
2404 (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2408 knote_free(struct knote *kn)
2411 uma_zfree(knote_zone, kn);
2415 * Register the kev w/ the kq specified by fd.
2418 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2422 cap_rights_t rights;
2425 error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2428 if ((error = kqueue_acquire(fp, &kq)) != 0)
2431 error = kqueue_register(kq, kev, td, waitok);
2432 kqueue_release(kq, 0);