2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
5 * Copyright 2004 John-Mark Gurney <jmg@FreeBSD.org>
6 * Copyright (c) 2009 Apple, Inc.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
34 #include "opt_ktrace.h"
35 #include "opt_kqueue.h"
37 #ifdef COMPAT_FREEBSD11
38 #define _WANT_FREEBSD11_KEVENT
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/capsicum.h>
44 #include <sys/kernel.h>
46 #include <sys/mutex.h>
47 #include <sys/rwlock.h>
49 #include <sys/malloc.h>
50 #include <sys/unistd.h>
52 #include <sys/filedesc.h>
53 #include <sys/filio.h>
54 #include <sys/fcntl.h>
55 #include <sys/kthread.h>
56 #include <sys/selinfo.h>
57 #include <sys/queue.h>
58 #include <sys/event.h>
59 #include <sys/eventvar.h>
61 #include <sys/protosw.h>
62 #include <sys/resourcevar.h>
63 #include <sys/sigio.h>
64 #include <sys/signalvar.h>
65 #include <sys/socket.h>
66 #include <sys/socketvar.h>
68 #include <sys/sysctl.h>
69 #include <sys/sysproto.h>
70 #include <sys/syscallsubr.h>
71 #include <sys/taskqueue.h>
75 #include <sys/ktrace.h>
77 #include <machine/atomic.h>
81 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
84 * This lock is used if multiple kq locks are required. This possibly
85 * should be made into a per proc lock.
87 static struct mtx kq_global;
88 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
89 #define KQ_GLOBAL_LOCK(lck, haslck) do { \
94 #define KQ_GLOBAL_UNLOCK(lck, haslck) do { \
100 TASKQUEUE_DEFINE_THREAD(kqueue_ctx);
102 static int kevent_copyout(void *arg, struct kevent *kevp, int count);
103 static int kevent_copyin(void *arg, struct kevent *kevp, int count);
104 static int kqueue_register(struct kqueue *kq, struct kevent *kev,
105 struct thread *td, int waitok);
106 static int kqueue_acquire(struct file *fp, struct kqueue **kqp);
107 static void kqueue_release(struct kqueue *kq, int locked);
108 static void kqueue_destroy(struct kqueue *kq);
109 static void kqueue_drain(struct kqueue *kq, struct thread *td);
110 static int kqueue_expand(struct kqueue *kq, struct filterops *fops,
111 uintptr_t ident, int waitok);
112 static void kqueue_task(void *arg, int pending);
113 static int kqueue_scan(struct kqueue *kq, int maxevents,
114 struct kevent_copyops *k_ops,
115 const struct timespec *timeout,
116 struct kevent *keva, struct thread *td);
117 static void kqueue_wakeup(struct kqueue *kq);
118 static struct filterops *kqueue_fo_find(int filt);
119 static void kqueue_fo_release(int filt);
120 struct g_kevent_args;
121 static int kern_kevent_generic(struct thread *td,
122 struct g_kevent_args *uap,
123 struct kevent_copyops *k_ops, const char *struct_name);
125 static fo_ioctl_t kqueue_ioctl;
126 static fo_poll_t kqueue_poll;
127 static fo_kqfilter_t kqueue_kqfilter;
128 static fo_stat_t kqueue_stat;
129 static fo_close_t kqueue_close;
130 static fo_fill_kinfo_t kqueue_fill_kinfo;
132 static struct fileops kqueueops = {
133 .fo_read = invfo_rdwr,
134 .fo_write = invfo_rdwr,
135 .fo_truncate = invfo_truncate,
136 .fo_ioctl = kqueue_ioctl,
137 .fo_poll = kqueue_poll,
138 .fo_kqfilter = kqueue_kqfilter,
139 .fo_stat = kqueue_stat,
140 .fo_close = kqueue_close,
141 .fo_chmod = invfo_chmod,
142 .fo_chown = invfo_chown,
143 .fo_sendfile = invfo_sendfile,
144 .fo_fill_kinfo = kqueue_fill_kinfo,
147 static int knote_attach(struct knote *kn, struct kqueue *kq);
148 static void knote_drop(struct knote *kn, struct thread *td);
149 static void knote_drop_detached(struct knote *kn, struct thread *td);
150 static void knote_enqueue(struct knote *kn);
151 static void knote_dequeue(struct knote *kn);
152 static void knote_init(void);
153 static struct knote *knote_alloc(int waitok);
154 static void knote_free(struct knote *kn);
156 static void filt_kqdetach(struct knote *kn);
157 static int filt_kqueue(struct knote *kn, long hint);
158 static int filt_procattach(struct knote *kn);
159 static void filt_procdetach(struct knote *kn);
160 static int filt_proc(struct knote *kn, long hint);
161 static int filt_fileattach(struct knote *kn);
162 static void filt_timerexpire(void *knx);
163 static int filt_timerattach(struct knote *kn);
164 static void filt_timerdetach(struct knote *kn);
165 static void filt_timerstart(struct knote *kn, sbintime_t to);
166 static void filt_timertouch(struct knote *kn, struct kevent *kev,
168 static int filt_timervalidate(struct knote *kn, sbintime_t *to);
169 static int filt_timer(struct knote *kn, long hint);
170 static int filt_userattach(struct knote *kn);
171 static void filt_userdetach(struct knote *kn);
172 static int filt_user(struct knote *kn, long hint);
173 static void filt_usertouch(struct knote *kn, struct kevent *kev,
176 static struct filterops file_filtops = {
178 .f_attach = filt_fileattach,
180 static struct filterops kqread_filtops = {
182 .f_detach = filt_kqdetach,
183 .f_event = filt_kqueue,
185 /* XXX - move to kern_proc.c? */
186 static struct filterops proc_filtops = {
188 .f_attach = filt_procattach,
189 .f_detach = filt_procdetach,
190 .f_event = filt_proc,
192 static struct filterops timer_filtops = {
194 .f_attach = filt_timerattach,
195 .f_detach = filt_timerdetach,
196 .f_event = filt_timer,
197 .f_touch = filt_timertouch,
199 static struct filterops user_filtops = {
200 .f_attach = filt_userattach,
201 .f_detach = filt_userdetach,
202 .f_event = filt_user,
203 .f_touch = filt_usertouch,
206 static uma_zone_t knote_zone;
207 static unsigned int kq_ncallouts = 0;
208 static unsigned int kq_calloutmax = 4 * 1024;
209 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
210 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
212 /* XXX - ensure not influx ? */
213 #define KNOTE_ACTIVATE(kn, islock) do { \
215 mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED); \
217 KQ_LOCK((kn)->kn_kq); \
218 (kn)->kn_status |= KN_ACTIVE; \
219 if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \
220 knote_enqueue((kn)); \
222 KQ_UNLOCK((kn)->kn_kq); \
224 #define KQ_LOCK(kq) do { \
225 mtx_lock(&(kq)->kq_lock); \
227 #define KQ_FLUX_WAKEUP(kq) do { \
228 if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) { \
229 (kq)->kq_state &= ~KQ_FLUXWAIT; \
233 #define KQ_UNLOCK_FLUX(kq) do { \
234 KQ_FLUX_WAKEUP(kq); \
235 mtx_unlock(&(kq)->kq_lock); \
237 #define KQ_UNLOCK(kq) do { \
238 mtx_unlock(&(kq)->kq_lock); \
240 #define KQ_OWNED(kq) do { \
241 mtx_assert(&(kq)->kq_lock, MA_OWNED); \
243 #define KQ_NOTOWNED(kq) do { \
244 mtx_assert(&(kq)->kq_lock, MA_NOTOWNED); \
247 static struct knlist *
248 kn_list_lock(struct knote *kn)
254 knl->kl_lock(knl->kl_lockarg);
259 kn_list_unlock(struct knlist *knl)
265 do_free = knl->kl_autodestroy && knlist_empty(knl);
266 knl->kl_unlock(knl->kl_lockarg);
274 kn_in_flux(struct knote *kn)
277 return (kn->kn_influx > 0);
281 kn_enter_flux(struct knote *kn)
285 MPASS(kn->kn_influx < INT_MAX);
290 kn_leave_flux(struct knote *kn)
294 MPASS(kn->kn_influx > 0);
296 return (kn->kn_influx == 0);
299 #define KNL_ASSERT_LOCK(knl, islocked) do { \
301 KNL_ASSERT_LOCKED(knl); \
303 KNL_ASSERT_UNLOCKED(knl); \
306 #define KNL_ASSERT_LOCKED(knl) do { \
307 knl->kl_assert_locked((knl)->kl_lockarg); \
309 #define KNL_ASSERT_UNLOCKED(knl) do { \
310 knl->kl_assert_unlocked((knl)->kl_lockarg); \
312 #else /* !INVARIANTS */
313 #define KNL_ASSERT_LOCKED(knl) do {} while(0)
314 #define KNL_ASSERT_UNLOCKED(knl) do {} while (0)
315 #endif /* INVARIANTS */
318 #define KN_HASHSIZE 64 /* XXX should be tunable */
321 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask))
324 filt_nullattach(struct knote *kn)
330 struct filterops null_filtops = {
332 .f_attach = filt_nullattach,
335 /* XXX - make SYSINIT to add these, and move into respective modules. */
336 extern struct filterops sig_filtops;
337 extern struct filterops fs_filtops;
340 * Table for for all system-defined filters.
342 static struct mtx filterops_lock;
343 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
346 struct filterops *for_fop;
349 } sysfilt_ops[EVFILT_SYSCOUNT] = {
350 { &file_filtops, 1 }, /* EVFILT_READ */
351 { &file_filtops, 1 }, /* EVFILT_WRITE */
352 { &null_filtops }, /* EVFILT_AIO */
353 { &file_filtops, 1 }, /* EVFILT_VNODE */
354 { &proc_filtops, 1 }, /* EVFILT_PROC */
355 { &sig_filtops, 1 }, /* EVFILT_SIGNAL */
356 { &timer_filtops, 1 }, /* EVFILT_TIMER */
357 { &file_filtops, 1 }, /* EVFILT_PROCDESC */
358 { &fs_filtops, 1 }, /* EVFILT_FS */
359 { &null_filtops }, /* EVFILT_LIO */
360 { &user_filtops, 1 }, /* EVFILT_USER */
361 { &null_filtops }, /* EVFILT_SENDFILE */
362 { &file_filtops, 1 }, /* EVFILT_EMPTY */
366 * Simple redirection for all cdevsw style objects to call their fo_kqfilter
370 filt_fileattach(struct knote *kn)
373 return (fo_kqfilter(kn->kn_fp, kn));
378 kqueue_kqfilter(struct file *fp, struct knote *kn)
380 struct kqueue *kq = kn->kn_fp->f_data;
382 if (kn->kn_filter != EVFILT_READ)
385 kn->kn_status |= KN_KQUEUE;
386 kn->kn_fop = &kqread_filtops;
387 knlist_add(&kq->kq_sel.si_note, kn, 0);
393 filt_kqdetach(struct knote *kn)
395 struct kqueue *kq = kn->kn_fp->f_data;
397 knlist_remove(&kq->kq_sel.si_note, kn, 0);
402 filt_kqueue(struct knote *kn, long hint)
404 struct kqueue *kq = kn->kn_fp->f_data;
406 kn->kn_data = kq->kq_count;
407 return (kn->kn_data > 0);
410 /* XXX - move to kern_proc.c? */
412 filt_procattach(struct knote *kn)
416 bool exiting, immediate;
418 exiting = immediate = false;
419 if (kn->kn_sfflags & NOTE_EXIT)
420 p = pfind_any(kn->kn_id);
422 p = pfind(kn->kn_id);
425 if (p->p_flag & P_WEXIT)
428 if ((error = p_cansee(curthread, p))) {
433 kn->kn_ptr.p_proc = p;
434 kn->kn_flags |= EV_CLEAR; /* automatically set */
437 * Internal flag indicating registration done by kernel for the
438 * purposes of getting a NOTE_CHILD notification.
440 if (kn->kn_flags & EV_FLAG2) {
441 kn->kn_flags &= ~EV_FLAG2;
442 kn->kn_data = kn->kn_sdata; /* ppid */
443 kn->kn_fflags = NOTE_CHILD;
444 kn->kn_sfflags &= ~(NOTE_EXIT | NOTE_EXEC | NOTE_FORK);
445 immediate = true; /* Force immediate activation of child note. */
448 * Internal flag indicating registration done by kernel (for other than
451 if (kn->kn_flags & EV_FLAG1) {
452 kn->kn_flags &= ~EV_FLAG1;
455 knlist_add(p->p_klist, kn, 1);
458 * Immediately activate any child notes or, in the case of a zombie
459 * target process, exit notes. The latter is necessary to handle the
460 * case where the target process, e.g. a child, dies before the kevent
463 if (immediate || (exiting && filt_proc(kn, NOTE_EXIT)))
464 KNOTE_ACTIVATE(kn, 0);
472 * The knote may be attached to a different process, which may exit,
473 * leaving nothing for the knote to be attached to. So when the process
474 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
475 * it will be deleted when read out. However, as part of the knote deletion,
476 * this routine is called, so a check is needed to avoid actually performing
477 * a detach, because the original process does not exist any more.
479 /* XXX - move to kern_proc.c? */
481 filt_procdetach(struct knote *kn)
484 knlist_remove(kn->kn_knlist, kn, 0);
485 kn->kn_ptr.p_proc = NULL;
488 /* XXX - move to kern_proc.c? */
490 filt_proc(struct knote *kn, long hint)
495 p = kn->kn_ptr.p_proc;
496 if (p == NULL) /* already activated, from attach filter */
499 /* Mask off extra data. */
500 event = (u_int)hint & NOTE_PCTRLMASK;
502 /* If the user is interested in this event, record it. */
503 if (kn->kn_sfflags & event)
504 kn->kn_fflags |= event;
506 /* Process is gone, so flag the event as finished. */
507 if (event == NOTE_EXIT) {
508 kn->kn_flags |= EV_EOF | EV_ONESHOT;
509 kn->kn_ptr.p_proc = NULL;
510 if (kn->kn_fflags & NOTE_EXIT)
511 kn->kn_data = KW_EXITCODE(p->p_xexit, p->p_xsig);
512 if (kn->kn_fflags == 0)
513 kn->kn_flags |= EV_DROP;
517 return (kn->kn_fflags != 0);
521 * Called when the process forked. It mostly does the same as the
522 * knote(), activating all knotes registered to be activated when the
523 * process forked. Additionally, for each knote attached to the
524 * parent, check whether user wants to track the new process. If so
525 * attach a new knote to it, and immediately report an event with the
529 knote_fork(struct knlist *list, int pid)
538 list->kl_lock(list->kl_lockarg);
540 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
543 if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
549 * The same as knote(), activate the event.
551 if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
552 kn->kn_status |= KN_HASKQLOCK;
553 if (kn->kn_fop->f_event(kn, NOTE_FORK))
554 KNOTE_ACTIVATE(kn, 1);
555 kn->kn_status &= ~KN_HASKQLOCK;
561 * The NOTE_TRACK case. In addition to the activation
562 * of the event, we need to register new events to
563 * track the child. Drop the locks in preparation for
564 * the call to kqueue_register().
568 list->kl_unlock(list->kl_lockarg);
571 * Activate existing knote and register tracking knotes with
574 * First register a knote to get just the child notice. This
575 * must be a separate note from a potential NOTE_EXIT
576 * notification since both NOTE_CHILD and NOTE_EXIT are defined
577 * to use the data field (in conflicting ways).
580 kev.filter = kn->kn_filter;
581 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_ONESHOT |
583 kev.fflags = kn->kn_sfflags;
584 kev.data = kn->kn_id; /* parent */
585 kev.udata = kn->kn_kevent.udata;/* preserve udata */
586 error = kqueue_register(kq, &kev, NULL, 0);
588 kn->kn_fflags |= NOTE_TRACKERR;
591 * Then register another knote to track other potential events
592 * from the new process.
595 kev.filter = kn->kn_filter;
596 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
597 kev.fflags = kn->kn_sfflags;
598 kev.data = kn->kn_id; /* parent */
599 kev.udata = kn->kn_kevent.udata;/* preserve udata */
600 error = kqueue_register(kq, &kev, NULL, 0);
602 kn->kn_fflags |= NOTE_TRACKERR;
603 if (kn->kn_fop->f_event(kn, NOTE_FORK))
604 KNOTE_ACTIVATE(kn, 0);
608 list->kl_lock(list->kl_lockarg);
610 list->kl_unlock(list->kl_lockarg);
614 * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
615 * interval timer support code.
618 #define NOTE_TIMER_PRECMASK \
619 (NOTE_SECONDS | NOTE_MSECONDS | NOTE_USECONDS | NOTE_NSECONDS)
622 timer2sbintime(intptr_t data, int flags)
627 * Macros for converting to the fractional second portion of an
628 * sbintime_t using 64bit multiplication to improve precision.
630 #define NS_TO_SBT(ns) (((ns) * (((uint64_t)1 << 63) / 500000000)) >> 32)
631 #define US_TO_SBT(us) (((us) * (((uint64_t)1 << 63) / 500000)) >> 32)
632 #define MS_TO_SBT(ms) (((ms) * (((uint64_t)1 << 63) / 500)) >> 32)
633 switch (flags & NOTE_TIMER_PRECMASK) {
636 if (data > (SBT_MAX / SBT_1S))
639 return ((sbintime_t)data << 32);
640 case NOTE_MSECONDS: /* FALLTHROUGH */
645 if (secs > (SBT_MAX / SBT_1S))
648 return (secs << 32 | MS_TO_SBT(data % 1000));
650 return (MS_TO_SBT(data));
652 if (data >= 1000000) {
653 secs = data / 1000000;
655 if (secs > (SBT_MAX / SBT_1S))
658 return (secs << 32 | US_TO_SBT(data % 1000000));
660 return (US_TO_SBT(data));
662 if (data >= 1000000000) {
663 secs = data / 1000000000;
665 if (secs > (SBT_MAX / SBT_1S))
668 return (secs << 32 | US_TO_SBT(data % 1000000000));
670 return (NS_TO_SBT(data));
677 struct kq_timer_cb_data {
679 sbintime_t next; /* next timer event fires at */
680 sbintime_t to; /* precalculated timer period, 0 for abs */
684 filt_timerexpire(void *knx)
687 struct kq_timer_cb_data *kc;
691 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */
693 if ((kn->kn_flags & EV_ONESHOT) != 0)
699 callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
700 PCPU_GET(cpuid), C_ABSOLUTE);
704 * data contains amount of time to sleep
707 filt_timervalidate(struct knote *kn, sbintime_t *to)
712 if (kn->kn_sdata < 0)
714 if (kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
717 * The only fflags values supported are the timer unit
718 * (precision) and the absolute time indicator.
720 if ((kn->kn_sfflags & ~(NOTE_TIMER_PRECMASK | NOTE_ABSTIME)) != 0)
723 *to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
724 if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
735 filt_timerattach(struct knote *kn)
737 struct kq_timer_cb_data *kc;
739 unsigned int ncallouts;
742 error = filt_timervalidate(kn, &to);
747 ncallouts = kq_ncallouts;
748 if (ncallouts >= kq_calloutmax)
750 } while (!atomic_cmpset_int(&kq_ncallouts, ncallouts, ncallouts + 1));
752 if ((kn->kn_sfflags & NOTE_ABSTIME) == 0)
753 kn->kn_flags |= EV_CLEAR; /* automatically set */
754 kn->kn_status &= ~KN_DETACHED; /* knlist_add clears it */
755 kn->kn_ptr.p_v = kc = malloc(sizeof(*kc), M_KQUEUE, M_WAITOK);
756 callout_init(&kc->c, 1);
757 filt_timerstart(kn, to);
763 filt_timerstart(struct knote *kn, sbintime_t to)
765 struct kq_timer_cb_data *kc;
768 if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
772 kc->next = to + sbinuptime();
775 callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
776 PCPU_GET(cpuid), C_ABSOLUTE);
780 filt_timerdetach(struct knote *kn)
782 struct kq_timer_cb_data *kc;
783 unsigned int old __unused;
786 callout_drain(&kc->c);
788 old = atomic_fetchadd_int(&kq_ncallouts, -1);
789 KASSERT(old > 0, ("Number of callouts cannot become negative"));
790 kn->kn_status |= KN_DETACHED; /* knlist_remove sets it */
794 filt_timertouch(struct knote *kn, struct kevent *kev, u_long type)
796 struct kq_timer_cb_data *kc;
803 /* Handle re-added timers that update data/fflags */
804 if (kev->flags & EV_ADD) {
807 /* Drain any existing callout. */
808 callout_drain(&kc->c);
810 /* Throw away any existing undelivered record
811 * of the timer expiration. This is done under
812 * the presumption that if a process is
813 * re-adding this timer with new parameters,
814 * it is no longer interested in what may have
815 * happened under the old parameters. If it is
816 * interested, it can wait for the expiration,
817 * delete the old timer definition, and then
820 * This has to be done while the kq is locked:
821 * - if enqueued, dequeue
822 * - make it no longer active
823 * - clear the count of expiration events
827 if (kn->kn_status & KN_QUEUED)
830 kn->kn_status &= ~KN_ACTIVE;
834 /* Reschedule timer based on new data/fflags */
835 kn->kn_sfflags = kev->fflags;
836 kn->kn_sdata = kev->data;
837 error = filt_timervalidate(kn, &to);
839 kn->kn_flags |= EV_ERROR;
842 filt_timerstart(kn, to);
847 *kev = kn->kn_kevent;
848 if (kn->kn_flags & EV_CLEAR) {
855 panic("filt_timertouch() - invalid type (%ld)", type);
861 filt_timer(struct knote *kn, long hint)
864 return (kn->kn_data != 0);
868 filt_userattach(struct knote *kn)
872 * EVFILT_USER knotes are not attached to anything in the kernel.
875 if (kn->kn_fflags & NOTE_TRIGGER)
883 filt_userdetach(__unused struct knote *kn)
887 * EVFILT_USER knotes are not attached to anything in the kernel.
892 filt_user(struct knote *kn, __unused long hint)
895 return (kn->kn_hookid);
899 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
905 if (kev->fflags & NOTE_TRIGGER)
908 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
909 kev->fflags &= NOTE_FFLAGSMASK;
915 kn->kn_sfflags &= kev->fflags;
919 kn->kn_sfflags |= kev->fflags;
923 kn->kn_sfflags = kev->fflags;
927 /* XXX Return error? */
930 kn->kn_sdata = kev->data;
931 if (kev->flags & EV_CLEAR) {
939 *kev = kn->kn_kevent;
940 kev->fflags = kn->kn_sfflags;
941 kev->data = kn->kn_sdata;
942 if (kn->kn_flags & EV_CLEAR) {
950 panic("filt_usertouch() - invalid type (%ld)", type);
956 sys_kqueue(struct thread *td, struct kqueue_args *uap)
959 return (kern_kqueue(td, 0, NULL));
963 kqueue_init(struct kqueue *kq)
966 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
967 TAILQ_INIT(&kq->kq_head);
968 knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
969 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
973 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
975 struct filedesc *fdp;
981 fdp = td->td_proc->p_fd;
983 if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
986 error = falloc_caps(td, &fp, &fd, flags, fcaps);
988 chgkqcnt(cred->cr_ruidinfo, -1, 0);
992 /* An extra reference on `fp' has been held for us by falloc(). */
993 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
996 kq->kq_cred = crhold(cred);
999 TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
1000 FILEDESC_XUNLOCK(fdp);
1002 finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
1005 td->td_retval[0] = fd;
1009 struct g_kevent_args {
1015 const struct timespec *timeout;
1019 sys_kevent(struct thread *td, struct kevent_args *uap)
1021 struct kevent_copyops k_ops = {
1023 .k_copyout = kevent_copyout,
1024 .k_copyin = kevent_copyin,
1025 .kevent_size = sizeof(struct kevent),
1027 struct g_kevent_args gk_args = {
1029 .changelist = uap->changelist,
1030 .nchanges = uap->nchanges,
1031 .eventlist = uap->eventlist,
1032 .nevents = uap->nevents,
1033 .timeout = uap->timeout,
1036 return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent"));
1040 kern_kevent_generic(struct thread *td, struct g_kevent_args *uap,
1041 struct kevent_copyops *k_ops, const char *struct_name)
1043 struct timespec ts, *tsp;
1045 struct kevent *eventlist = uap->eventlist;
1049 if (uap->timeout != NULL) {
1050 error = copyin(uap->timeout, &ts, sizeof(ts));
1058 if (KTRPOINT(td, KTR_STRUCT_ARRAY))
1059 ktrstructarray(struct_name, UIO_USERSPACE, uap->changelist,
1060 uap->nchanges, k_ops->kevent_size);
1063 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
1067 if (error == 0 && KTRPOINT(td, KTR_STRUCT_ARRAY))
1068 ktrstructarray(struct_name, UIO_USERSPACE, eventlist,
1069 td->td_retval[0], k_ops->kevent_size);
1076 * Copy 'count' items into the destination list pointed to by uap->eventlist.
1079 kevent_copyout(void *arg, struct kevent *kevp, int count)
1081 struct kevent_args *uap;
1084 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1085 uap = (struct kevent_args *)arg;
1087 error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
1089 uap->eventlist += count;
1094 * Copy 'count' items from the list pointed to by uap->changelist.
1097 kevent_copyin(void *arg, struct kevent *kevp, int count)
1099 struct kevent_args *uap;
1102 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1103 uap = (struct kevent_args *)arg;
1105 error = copyin(uap->changelist, kevp, count * sizeof *kevp);
1107 uap->changelist += count;
1111 #ifdef COMPAT_FREEBSD11
1113 kevent11_copyout(void *arg, struct kevent *kevp, int count)
1115 struct freebsd11_kevent_args *uap;
1116 struct kevent_freebsd11 kev11;
1119 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1120 uap = (struct freebsd11_kevent_args *)arg;
1122 for (i = 0; i < count; i++) {
1123 kev11.ident = kevp->ident;
1124 kev11.filter = kevp->filter;
1125 kev11.flags = kevp->flags;
1126 kev11.fflags = kevp->fflags;
1127 kev11.data = kevp->data;
1128 kev11.udata = kevp->udata;
1129 error = copyout(&kev11, uap->eventlist, sizeof(kev11));
1139 * Copy 'count' items from the list pointed to by uap->changelist.
1142 kevent11_copyin(void *arg, struct kevent *kevp, int count)
1144 struct freebsd11_kevent_args *uap;
1145 struct kevent_freebsd11 kev11;
1148 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1149 uap = (struct freebsd11_kevent_args *)arg;
1151 for (i = 0; i < count; i++) {
1152 error = copyin(uap->changelist, &kev11, sizeof(kev11));
1155 kevp->ident = kev11.ident;
1156 kevp->filter = kev11.filter;
1157 kevp->flags = kev11.flags;
1158 kevp->fflags = kev11.fflags;
1159 kevp->data = (uintptr_t)kev11.data;
1160 kevp->udata = kev11.udata;
1161 bzero(&kevp->ext, sizeof(kevp->ext));
1169 freebsd11_kevent(struct thread *td, struct freebsd11_kevent_args *uap)
1171 struct kevent_copyops k_ops = {
1173 .k_copyout = kevent11_copyout,
1174 .k_copyin = kevent11_copyin,
1175 .kevent_size = sizeof(struct kevent_freebsd11),
1177 struct g_kevent_args gk_args = {
1179 .changelist = uap->changelist,
1180 .nchanges = uap->nchanges,
1181 .eventlist = uap->eventlist,
1182 .nevents = uap->nevents,
1183 .timeout = uap->timeout,
1186 return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent_freebsd11"));
1191 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
1192 struct kevent_copyops *k_ops, const struct timespec *timeout)
1194 cap_rights_t rights;
1198 cap_rights_init(&rights);
1200 cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
1202 cap_rights_set(&rights, CAP_KQUEUE_EVENT);
1203 error = fget(td, fd, &rights, &fp);
1207 error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
1214 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
1215 struct kevent_copyops *k_ops, const struct timespec *timeout)
1217 struct kevent keva[KQ_NEVENTS];
1218 struct kevent *kevp, *changes;
1219 int i, n, nerrors, error;
1222 while (nchanges > 0) {
1223 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
1224 error = k_ops->k_copyin(k_ops->arg, keva, n);
1228 for (i = 0; i < n; i++) {
1232 kevp->flags &= ~EV_SYSFLAGS;
1233 error = kqueue_register(kq, kevp, td, 1);
1234 if (error || (kevp->flags & EV_RECEIPT)) {
1237 kevp->flags = EV_ERROR;
1239 (void)k_ops->k_copyout(k_ops->arg, kevp, 1);
1247 td->td_retval[0] = nerrors;
1251 return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
1255 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
1256 struct kevent_copyops *k_ops, const struct timespec *timeout)
1261 error = kqueue_acquire(fp, &kq);
1264 error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
1265 kqueue_release(kq, 0);
1270 * Performs a kevent() call on a temporarily created kqueue. This can be
1271 * used to perform one-shot polling, similar to poll() and select().
1274 kern_kevent_anonymous(struct thread *td, int nevents,
1275 struct kevent_copyops *k_ops)
1277 struct kqueue kq = {};
1282 error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
1283 kqueue_drain(&kq, td);
1284 kqueue_destroy(&kq);
1289 kqueue_add_filteropts(int filt, struct filterops *filtops)
1294 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1296 "trying to add a filterop that is out of range: %d is beyond %d\n",
1297 ~filt, EVFILT_SYSCOUNT);
1300 mtx_lock(&filterops_lock);
1301 if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1302 sysfilt_ops[~filt].for_fop != NULL)
1305 sysfilt_ops[~filt].for_fop = filtops;
1306 sysfilt_ops[~filt].for_refcnt = 0;
1308 mtx_unlock(&filterops_lock);
1314 kqueue_del_filteropts(int filt)
1319 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1322 mtx_lock(&filterops_lock);
1323 if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1324 sysfilt_ops[~filt].for_fop == NULL)
1326 else if (sysfilt_ops[~filt].for_refcnt != 0)
1329 sysfilt_ops[~filt].for_fop = &null_filtops;
1330 sysfilt_ops[~filt].for_refcnt = 0;
1332 mtx_unlock(&filterops_lock);
1337 static struct filterops *
1338 kqueue_fo_find(int filt)
1341 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1344 if (sysfilt_ops[~filt].for_nolock)
1345 return sysfilt_ops[~filt].for_fop;
1347 mtx_lock(&filterops_lock);
1348 sysfilt_ops[~filt].for_refcnt++;
1349 if (sysfilt_ops[~filt].for_fop == NULL)
1350 sysfilt_ops[~filt].for_fop = &null_filtops;
1351 mtx_unlock(&filterops_lock);
1353 return sysfilt_ops[~filt].for_fop;
1357 kqueue_fo_release(int filt)
1360 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1363 if (sysfilt_ops[~filt].for_nolock)
1366 mtx_lock(&filterops_lock);
1367 KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1368 ("filter object refcount not valid on release"));
1369 sysfilt_ops[~filt].for_refcnt--;
1370 mtx_unlock(&filterops_lock);
1374 * A ref to kq (obtained via kqueue_acquire) must be held. waitok will
1375 * influence if memory allocation should wait. Make sure it is 0 if you
1379 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1381 struct filterops *fops;
1383 struct knote *kn, *tkn;
1385 int error, filt, event;
1386 int haskqglobal, filedesc_unlock;
1388 if ((kev->flags & (EV_ENABLE | EV_DISABLE)) == (EV_ENABLE | EV_DISABLE))
1396 filedesc_unlock = 0;
1399 fops = kqueue_fo_find(filt);
1403 if (kev->flags & EV_ADD) {
1405 * Prevent waiting with locks. Non-sleepable
1406 * allocation failures are handled in the loop, only
1407 * if the spare knote appears to be actually required.
1409 tkn = knote_alloc(waitok);
1416 KASSERT(td != NULL, ("td is NULL"));
1417 if (kev->ident > INT_MAX)
1420 error = fget(td, kev->ident, &cap_event_rights, &fp);
1424 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1425 kev->ident, 0) != 0) {
1429 error = kqueue_expand(kq, fops, kev->ident, waitok);
1435 if (fp->f_type == DTYPE_KQUEUE) {
1437 * If we add some intelligence about what we are doing,
1438 * we should be able to support events on ourselves.
1439 * We need to know when we are doing this to prevent
1440 * getting both the knlist lock and the kq lock since
1441 * they are the same thing.
1443 if (fp->f_data == kq) {
1449 * Pre-lock the filedesc before the global
1450 * lock mutex, see the comment in
1453 FILEDESC_XLOCK(td->td_proc->p_fd);
1454 filedesc_unlock = 1;
1455 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1459 if (kev->ident < kq->kq_knlistsize) {
1460 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1461 if (kev->filter == kn->kn_filter)
1465 if ((kev->flags & EV_ADD) == EV_ADD)
1466 kqueue_expand(kq, fops, kev->ident, waitok);
1471 * If possible, find an existing knote to use for this kevent.
1473 if (kev->filter == EVFILT_PROC &&
1474 (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1475 /* This is an internal creation of a process tracking
1476 * note. Don't attempt to coalesce this with an
1480 } else if (kq->kq_knhashmask != 0) {
1483 list = &kq->kq_knhash[
1484 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1485 SLIST_FOREACH(kn, list, kn_link)
1486 if (kev->ident == kn->kn_id &&
1487 kev->filter == kn->kn_filter)
1492 /* knote is in the process of changing, wait for it to stabilize. */
1493 if (kn != NULL && kn_in_flux(kn)) {
1494 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1495 if (filedesc_unlock) {
1496 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1497 filedesc_unlock = 0;
1499 kq->kq_state |= KQ_FLUXWAIT;
1500 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1509 * kn now contains the matching knote, or NULL if no match
1512 if (kev->flags & EV_ADD) {
1524 * apply reference counts to knote structure, and
1525 * do not release it at the end of this routine.
1530 kn->kn_sfflags = kev->fflags;
1531 kn->kn_sdata = kev->data;
1534 kn->kn_kevent = *kev;
1535 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1536 EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1537 kn->kn_status = KN_DETACHED;
1540 error = knote_attach(kn, kq);
1547 if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1548 knote_drop_detached(kn, td);
1551 knl = kn_list_lock(kn);
1554 /* No matching knote and the EV_ADD flag is not set. */
1561 if (kev->flags & EV_DELETE) {
1568 if (kev->flags & EV_FORCEONESHOT) {
1569 kn->kn_flags |= EV_ONESHOT;
1570 KNOTE_ACTIVATE(kn, 1);
1574 * The user may change some filter values after the initial EV_ADD,
1575 * but doing so will not reset any filter which has already been
1578 kn->kn_status |= KN_SCAN;
1581 knl = kn_list_lock(kn);
1582 kn->kn_kevent.udata = kev->udata;
1583 if (!fops->f_isfd && fops->f_touch != NULL) {
1584 fops->f_touch(kn, kev, EVENT_REGISTER);
1586 kn->kn_sfflags = kev->fflags;
1587 kn->kn_sdata = kev->data;
1591 * We can get here with kn->kn_knlist == NULL. This can happen when
1592 * the initial attach event decides that the event is "completed"
1593 * already. i.e. filt_procattach is called on a zombie process. It
1594 * will call filt_proc which will remove it from the list, and NULL
1598 if ((kev->flags & EV_ENABLE) != 0)
1599 kn->kn_status &= ~KN_DISABLED;
1600 else if ((kev->flags & EV_DISABLE) != 0)
1601 kn->kn_status |= KN_DISABLED;
1603 if ((kn->kn_status & KN_DISABLED) == 0)
1604 event = kn->kn_fop->f_event(kn, 0);
1610 kn->kn_status |= KN_ACTIVE;
1611 if ((kn->kn_status & (KN_ACTIVE | KN_DISABLED | KN_QUEUED)) ==
1614 kn->kn_status &= ~KN_SCAN;
1616 kn_list_unlock(knl);
1620 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1621 if (filedesc_unlock)
1622 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1627 kqueue_fo_release(filt);
1632 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1640 if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1644 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1655 kqueue_release(struct kqueue *kq, int locked)
1662 if (kq->kq_refcnt == 1)
1663 wakeup(&kq->kq_refcnt);
1669 kqueue_schedtask(struct kqueue *kq)
1673 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1674 ("scheduling kqueue task while draining"));
1676 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1677 taskqueue_enqueue(taskqueue_kqueue_ctx, &kq->kq_task);
1678 kq->kq_state |= KQ_TASKSCHED;
1683 * Expand the kq to make sure we have storage for fops/ident pair.
1685 * Return 0 on success (or no work necessary), return errno on failure.
1687 * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1688 * If kqueue_register is called from a non-fd context, there usually/should
1692 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1695 struct klist *list, *tmp_knhash, *to_free;
1696 u_long tmp_knhashmask;
1699 int mflag = waitok ? M_WAITOK : M_NOWAIT;
1706 if (kq->kq_knlistsize <= fd) {
1707 size = kq->kq_knlistsize;
1710 list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1714 if (kq->kq_knlistsize > fd) {
1718 if (kq->kq_knlist != NULL) {
1719 bcopy(kq->kq_knlist, list,
1720 kq->kq_knlistsize * sizeof(*list));
1721 to_free = kq->kq_knlist;
1722 kq->kq_knlist = NULL;
1724 bzero((caddr_t)list +
1725 kq->kq_knlistsize * sizeof(*list),
1726 (size - kq->kq_knlistsize) * sizeof(*list));
1727 kq->kq_knlistsize = size;
1728 kq->kq_knlist = list;
1733 if (kq->kq_knhashmask == 0) {
1734 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1736 if (tmp_knhash == NULL)
1739 if (kq->kq_knhashmask == 0) {
1740 kq->kq_knhash = tmp_knhash;
1741 kq->kq_knhashmask = tmp_knhashmask;
1743 to_free = tmp_knhash;
1748 free(to_free, M_KQUEUE);
1755 kqueue_task(void *arg, int pending)
1763 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1766 KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1768 kq->kq_state &= ~KQ_TASKSCHED;
1769 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1770 wakeup(&kq->kq_state);
1773 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1777 * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1778 * We treat KN_MARKER knotes as if they are in flux.
1781 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1782 const struct timespec *tsp, struct kevent *keva, struct thread *td)
1784 struct kevent *kevp;
1785 struct knote *kn, *marker;
1787 sbintime_t asbt, rsbt;
1788 int count, error, haskqglobal, influx, nkev, touch;
1800 if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1801 tsp->tv_nsec >= 1000000000) {
1805 if (timespecisset(tsp)) {
1806 if (tsp->tv_sec <= INT32_MAX) {
1807 rsbt = tstosbt(*tsp);
1808 if (TIMESEL(&asbt, rsbt))
1809 asbt += tc_tick_sbt;
1810 if (asbt <= SBT_MAX - rsbt)
1814 rsbt >>= tc_precexp;
1821 marker = knote_alloc(1);
1822 marker->kn_status = KN_MARKER;
1827 if (kq->kq_count == 0) {
1829 error = EWOULDBLOCK;
1831 kq->kq_state |= KQ_SLEEP;
1832 error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1833 "kqread", asbt, rsbt, C_ABSOLUTE);
1837 /* don't restart after signals... */
1838 if (error == ERESTART)
1840 else if (error == EWOULDBLOCK)
1845 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1849 kn = TAILQ_FIRST(&kq->kq_head);
1851 if ((kn->kn_status == KN_MARKER && kn != marker) ||
1857 kq->kq_state |= KQ_FLUXWAIT;
1858 error = msleep(kq, &kq->kq_lock, PSOCK,
1863 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1864 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1865 kn->kn_status &= ~KN_QUEUED;
1871 if (count == maxevents)
1875 KASSERT(!kn_in_flux(kn),
1876 ("knote %p is unexpectedly in flux", kn));
1878 if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1879 kn->kn_status &= ~KN_QUEUED;
1884 * We don't need to lock the list since we've
1885 * marked it as in flux.
1890 } else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1891 kn->kn_status &= ~KN_QUEUED;
1896 * We don't need to lock the list since we've
1897 * marked the knote as being in flux.
1899 *kevp = kn->kn_kevent;
1904 kn->kn_status |= KN_SCAN;
1907 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1908 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1909 knl = kn_list_lock(kn);
1910 if (kn->kn_fop->f_event(kn, 0) == 0) {
1912 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1913 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE |
1917 kn_list_unlock(knl);
1921 touch = (!kn->kn_fop->f_isfd &&
1922 kn->kn_fop->f_touch != NULL);
1924 kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1926 *kevp = kn->kn_kevent;
1928 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1929 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1931 * Manually clear knotes who weren't
1934 if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1938 if (kn->kn_flags & EV_DISPATCH)
1939 kn->kn_status |= KN_DISABLED;
1940 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1943 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1945 kn->kn_status &= ~KN_SCAN;
1947 kn_list_unlock(knl);
1951 /* we are returning a copy to the user */
1956 if (nkev == KQ_NEVENTS) {
1959 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1967 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1975 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1976 td->td_retval[0] = maxevents - count;
1982 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1983 struct ucred *active_cred, struct thread *td)
1986 * Enabling sigio causes two major problems:
1987 * 1) infinite recursion:
1988 * Synopsys: kevent is being used to track signals and have FIOASYNC
1989 * set. On receipt of a signal this will cause a kqueue to recurse
1990 * into itself over and over. Sending the sigio causes the kqueue
1991 * to become ready, which in turn posts sigio again, forever.
1992 * Solution: this can be solved by setting a flag in the kqueue that
1993 * we have a SIGIO in progress.
1994 * 2) locking problems:
1995 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1996 * us above the proc and pgrp locks.
1997 * Solution: Post a signal using an async mechanism, being sure to
1998 * record a generation count in the delivery so that we do not deliver
1999 * a signal to the wrong process.
2001 * Note, these two mechanisms are somewhat mutually exclusive!
2010 kq->kq_state |= KQ_ASYNC;
2012 kq->kq_state &= ~KQ_ASYNC;
2017 return (fsetown(*(int *)data, &kq->kq_sigio));
2020 *(int *)data = fgetown(&kq->kq_sigio);
2030 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
2037 if ((error = kqueue_acquire(fp, &kq)))
2041 if (events & (POLLIN | POLLRDNORM)) {
2043 revents |= events & (POLLIN | POLLRDNORM);
2045 selrecord(td, &kq->kq_sel);
2046 if (SEL_WAITING(&kq->kq_sel))
2047 kq->kq_state |= KQ_SEL;
2050 kqueue_release(kq, 1);
2057 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
2061 bzero((void *)st, sizeof *st);
2063 * We no longer return kq_count because the unlocked value is useless.
2064 * If you spent all this time getting the count, why not spend your
2065 * syscall better by calling kevent?
2067 * XXX - This is needed for libc_r.
2069 st->st_mode = S_IFIFO;
2074 kqueue_drain(struct kqueue *kq, struct thread *td)
2081 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
2082 ("kqueue already closing"));
2083 kq->kq_state |= KQ_CLOSING;
2084 if (kq->kq_refcnt > 1)
2085 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
2087 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
2089 KASSERT(knlist_empty(&kq->kq_sel.si_note),
2090 ("kqueue's knlist not empty"));
2092 for (i = 0; i < kq->kq_knlistsize; i++) {
2093 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
2094 if (kn_in_flux(kn)) {
2095 kq->kq_state |= KQ_FLUXWAIT;
2096 msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
2105 if (kq->kq_knhashmask != 0) {
2106 for (i = 0; i <= kq->kq_knhashmask; i++) {
2107 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
2108 if (kn_in_flux(kn)) {
2109 kq->kq_state |= KQ_FLUXWAIT;
2110 msleep(kq, &kq->kq_lock, PSOCK,
2122 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
2123 kq->kq_state |= KQ_TASKDRAIN;
2124 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
2127 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2128 selwakeuppri(&kq->kq_sel, PSOCK);
2129 if (!SEL_WAITING(&kq->kq_sel))
2130 kq->kq_state &= ~KQ_SEL;
2137 kqueue_destroy(struct kqueue *kq)
2140 KASSERT(kq->kq_fdp == NULL,
2141 ("kqueue still attached to a file descriptor"));
2142 seldrain(&kq->kq_sel);
2143 knlist_destroy(&kq->kq_sel.si_note);
2144 mtx_destroy(&kq->kq_lock);
2146 if (kq->kq_knhash != NULL)
2147 free(kq->kq_knhash, M_KQUEUE);
2148 if (kq->kq_knlist != NULL)
2149 free(kq->kq_knlist, M_KQUEUE);
2151 funsetown(&kq->kq_sigio);
2156 kqueue_close(struct file *fp, struct thread *td)
2158 struct kqueue *kq = fp->f_data;
2159 struct filedesc *fdp;
2161 int filedesc_unlock;
2163 if ((error = kqueue_acquire(fp, &kq)))
2165 kqueue_drain(kq, td);
2168 * We could be called due to the knote_drop() doing fdrop(),
2169 * called from kqueue_register(). In this case the global
2170 * lock is owned, and filedesc sx is locked before, to not
2171 * take the sleepable lock after non-sleepable.
2175 if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
2176 FILEDESC_XLOCK(fdp);
2177 filedesc_unlock = 1;
2179 filedesc_unlock = 0;
2180 TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
2181 if (filedesc_unlock)
2182 FILEDESC_XUNLOCK(fdp);
2185 chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
2186 crfree(kq->kq_cred);
2194 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2197 kif->kf_type = KF_TYPE_KQUEUE;
2202 kqueue_wakeup(struct kqueue *kq)
2206 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
2207 kq->kq_state &= ~KQ_SLEEP;
2210 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2211 selwakeuppri(&kq->kq_sel, PSOCK);
2212 if (!SEL_WAITING(&kq->kq_sel))
2213 kq->kq_state &= ~KQ_SEL;
2215 if (!knlist_empty(&kq->kq_sel.si_note))
2216 kqueue_schedtask(kq);
2217 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
2218 pgsigio(&kq->kq_sigio, SIGIO, 0);
2223 * Walk down a list of knotes, activating them if their event has triggered.
2225 * There is a possibility to optimize in the case of one kq watching another.
2226 * Instead of scheduling a task to wake it up, you could pass enough state
2227 * down the chain to make up the parent kqueue. Make this code functional
2231 knote(struct knlist *list, long hint, int lockflags)
2234 struct knote *kn, *tkn;
2240 KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
2242 if ((lockflags & KNF_LISTLOCKED) == 0)
2243 list->kl_lock(list->kl_lockarg);
2246 * If we unlock the list lock (and enter influx), we can
2247 * eliminate the kqueue scheduling, but this will introduce
2248 * four lock/unlock's for each knote to test. Also, marker
2249 * would be needed to keep iteration position, since filters
2250 * or other threads could remove events.
2252 SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
2255 if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
2257 * Do not process the influx notes, except for
2258 * the influx coming from the kq unlock in the
2259 * kqueue_scan(). In the later case, we do
2260 * not interfere with the scan, since the code
2261 * fragment in kqueue_scan() locks the knlist,
2262 * and cannot proceed until we finished.
2265 } else if ((lockflags & KNF_NOKQLOCK) != 0) {
2268 error = kn->kn_fop->f_event(kn, hint);
2272 KNOTE_ACTIVATE(kn, 1);
2275 kn->kn_status |= KN_HASKQLOCK;
2276 if (kn->kn_fop->f_event(kn, hint))
2277 KNOTE_ACTIVATE(kn, 1);
2278 kn->kn_status &= ~KN_HASKQLOCK;
2282 if ((lockflags & KNF_LISTLOCKED) == 0)
2283 list->kl_unlock(list->kl_lockarg);
2287 * add a knote to a knlist
2290 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2293 KNL_ASSERT_LOCK(knl, islocked);
2294 KQ_NOTOWNED(kn->kn_kq);
2295 KASSERT(kn_in_flux(kn), ("knote %p not in flux", kn));
2296 KASSERT((kn->kn_status & KN_DETACHED) != 0,
2297 ("knote %p was not detached", kn));
2299 knl->kl_lock(knl->kl_lockarg);
2300 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2302 knl->kl_unlock(knl->kl_lockarg);
2304 kn->kn_knlist = knl;
2305 kn->kn_status &= ~KN_DETACHED;
2306 KQ_UNLOCK(kn->kn_kq);
2310 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked,
2314 KASSERT(!kqislocked || knlislocked, ("kq locked w/o knl locked"));
2315 KNL_ASSERT_LOCK(knl, knlislocked);
2316 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2317 KASSERT(kqislocked || kn_in_flux(kn), ("knote %p not in flux", kn));
2318 KASSERT((kn->kn_status & KN_DETACHED) == 0,
2319 ("knote %p was already detached", kn));
2321 knl->kl_lock(knl->kl_lockarg);
2322 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2323 kn->kn_knlist = NULL;
2325 kn_list_unlock(knl);
2328 kn->kn_status |= KN_DETACHED;
2330 KQ_UNLOCK(kn->kn_kq);
2334 * remove knote from the specified knlist
2337 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2340 knlist_remove_kq(knl, kn, islocked, 0);
2344 knlist_empty(struct knlist *knl)
2347 KNL_ASSERT_LOCKED(knl);
2348 return (SLIST_EMPTY(&knl->kl_list));
2351 static struct mtx knlist_lock;
2352 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2354 static void knlist_mtx_lock(void *arg);
2355 static void knlist_mtx_unlock(void *arg);
2358 knlist_mtx_lock(void *arg)
2361 mtx_lock((struct mtx *)arg);
2365 knlist_mtx_unlock(void *arg)
2368 mtx_unlock((struct mtx *)arg);
2372 knlist_mtx_assert_locked(void *arg)
2375 mtx_assert((struct mtx *)arg, MA_OWNED);
2379 knlist_mtx_assert_unlocked(void *arg)
2382 mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2386 knlist_rw_rlock(void *arg)
2389 rw_rlock((struct rwlock *)arg);
2393 knlist_rw_runlock(void *arg)
2396 rw_runlock((struct rwlock *)arg);
2400 knlist_rw_assert_locked(void *arg)
2403 rw_assert((struct rwlock *)arg, RA_LOCKED);
2407 knlist_rw_assert_unlocked(void *arg)
2410 rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2414 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2415 void (*kl_unlock)(void *),
2416 void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2420 knl->kl_lockarg = &knlist_lock;
2422 knl->kl_lockarg = lock;
2424 if (kl_lock == NULL)
2425 knl->kl_lock = knlist_mtx_lock;
2427 knl->kl_lock = kl_lock;
2428 if (kl_unlock == NULL)
2429 knl->kl_unlock = knlist_mtx_unlock;
2431 knl->kl_unlock = kl_unlock;
2432 if (kl_assert_locked == NULL)
2433 knl->kl_assert_locked = knlist_mtx_assert_locked;
2435 knl->kl_assert_locked = kl_assert_locked;
2436 if (kl_assert_unlocked == NULL)
2437 knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2439 knl->kl_assert_unlocked = kl_assert_unlocked;
2441 knl->kl_autodestroy = 0;
2442 SLIST_INIT(&knl->kl_list);
2446 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2449 knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2453 knlist_alloc(struct mtx *lock)
2457 knl = malloc(sizeof(struct knlist), M_KQUEUE, M_WAITOK);
2458 knlist_init_mtx(knl, lock);
2463 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2466 knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2467 knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2471 knlist_destroy(struct knlist *knl)
2474 KASSERT(KNLIST_EMPTY(knl),
2475 ("destroying knlist %p with knotes on it", knl));
2479 knlist_detach(struct knlist *knl)
2482 KNL_ASSERT_LOCKED(knl);
2483 knl->kl_autodestroy = 1;
2484 if (knlist_empty(knl)) {
2485 knlist_destroy(knl);
2486 free(knl, M_KQUEUE);
2491 * Even if we are locked, we may need to drop the lock to allow any influx
2492 * knotes time to "settle".
2495 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2497 struct knote *kn, *kn2;
2500 KASSERT(!knl->kl_autodestroy, ("cleardel for autodestroy %p", knl));
2502 KNL_ASSERT_LOCKED(knl);
2504 KNL_ASSERT_UNLOCKED(knl);
2505 again: /* need to reacquire lock since we have dropped it */
2506 knl->kl_lock(knl->kl_lockarg);
2509 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2512 if (kn_in_flux(kn)) {
2516 knlist_remove_kq(knl, kn, 1, 1);
2520 knote_drop_detached(kn, td);
2522 /* Make sure cleared knotes disappear soon */
2523 kn->kn_flags |= EV_EOF | EV_ONESHOT;
2529 if (!SLIST_EMPTY(&knl->kl_list)) {
2530 /* there are still in flux knotes remaining */
2531 kn = SLIST_FIRST(&knl->kl_list);
2534 KASSERT(kn_in_flux(kn), ("knote removed w/o list lock"));
2535 knl->kl_unlock(knl->kl_lockarg);
2536 kq->kq_state |= KQ_FLUXWAIT;
2537 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2543 KNL_ASSERT_LOCKED(knl);
2545 knl->kl_unlock(knl->kl_lockarg);
2546 KNL_ASSERT_UNLOCKED(knl);
2551 * Remove all knotes referencing a specified fd must be called with FILEDESC
2552 * lock. This prevents a race where a new fd comes along and occupies the
2553 * entry and we attach a knote to the fd.
2556 knote_fdclose(struct thread *td, int fd)
2558 struct filedesc *fdp = td->td_proc->p_fd;
2563 FILEDESC_XLOCK_ASSERT(fdp);
2566 * We shouldn't have to worry about new kevents appearing on fd
2567 * since filedesc is locked.
2569 TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2574 while (kq->kq_knlistsize > fd &&
2575 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2576 if (kn_in_flux(kn)) {
2577 /* someone else might be waiting on our knote */
2580 kq->kq_state |= KQ_FLUXWAIT;
2581 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2595 knote_attach(struct knote *kn, struct kqueue *kq)
2599 KASSERT(kn_in_flux(kn), ("knote %p not marked influx", kn));
2602 if (kn->kn_fop->f_isfd) {
2603 if (kn->kn_id >= kq->kq_knlistsize)
2605 list = &kq->kq_knlist[kn->kn_id];
2607 if (kq->kq_knhash == NULL)
2609 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2611 SLIST_INSERT_HEAD(list, kn, kn_link);
2616 knote_drop(struct knote *kn, struct thread *td)
2619 if ((kn->kn_status & KN_DETACHED) == 0)
2620 kn->kn_fop->f_detach(kn);
2621 knote_drop_detached(kn, td);
2625 knote_drop_detached(struct knote *kn, struct thread *td)
2632 KASSERT((kn->kn_status & KN_DETACHED) != 0,
2633 ("knote %p still attached", kn));
2637 KASSERT(kn->kn_influx == 1,
2638 ("knote_drop called on %p with influx %d", kn, kn->kn_influx));
2640 if (kn->kn_fop->f_isfd)
2641 list = &kq->kq_knlist[kn->kn_id];
2643 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2645 if (!SLIST_EMPTY(list))
2646 SLIST_REMOVE(list, kn, knote, kn_link);
2647 if (kn->kn_status & KN_QUEUED)
2651 if (kn->kn_fop->f_isfd) {
2652 fdrop(kn->kn_fp, td);
2655 kqueue_fo_release(kn->kn_kevent.filter);
2661 knote_enqueue(struct knote *kn)
2663 struct kqueue *kq = kn->kn_kq;
2665 KQ_OWNED(kn->kn_kq);
2666 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2668 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2669 kn->kn_status |= KN_QUEUED;
2675 knote_dequeue(struct knote *kn)
2677 struct kqueue *kq = kn->kn_kq;
2679 KQ_OWNED(kn->kn_kq);
2680 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2682 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2683 kn->kn_status &= ~KN_QUEUED;
2691 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2692 NULL, NULL, UMA_ALIGN_PTR, 0);
2694 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2696 static struct knote *
2697 knote_alloc(int waitok)
2700 return (uma_zalloc(knote_zone, (waitok ? M_WAITOK : M_NOWAIT) |
2705 knote_free(struct knote *kn)
2708 uma_zfree(knote_zone, kn);
2712 * Register the kev w/ the kq specified by fd.
2715 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2719 cap_rights_t rights;
2722 error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2725 if ((error = kqueue_acquire(fp, &kq)) != 0)
2728 error = kqueue_register(kq, kev, td, waitok);
2729 kqueue_release(kq, 0);