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>
45 #include <sys/limits.h>
47 #include <sys/mutex.h>
48 #include <sys/rwlock.h>
50 #include <sys/malloc.h>
51 #include <sys/unistd.h>
53 #include <sys/filedesc.h>
54 #include <sys/filio.h>
55 #include <sys/fcntl.h>
56 #include <sys/kthread.h>
57 #include <sys/selinfo.h>
58 #include <sys/queue.h>
59 #include <sys/event.h>
60 #include <sys/eventvar.h>
62 #include <sys/protosw.h>
63 #include <sys/resourcevar.h>
64 #include <sys/sigio.h>
65 #include <sys/signalvar.h>
66 #include <sys/socket.h>
67 #include <sys/socketvar.h>
69 #include <sys/sysctl.h>
70 #include <sys/sysproto.h>
71 #include <sys/syscallsubr.h>
72 #include <sys/taskqueue.h>
76 #include <sys/ktrace.h>
78 #include <machine/atomic.h>
82 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
85 * This lock is used if multiple kq locks are required. This possibly
86 * should be made into a per proc lock.
88 static struct mtx kq_global;
89 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
90 #define KQ_GLOBAL_LOCK(lck, haslck) do { \
95 #define KQ_GLOBAL_UNLOCK(lck, haslck) do { \
101 TASKQUEUE_DEFINE_THREAD(kqueue_ctx);
103 static int kevent_copyout(void *arg, struct kevent *kevp, int count);
104 static int kevent_copyin(void *arg, struct kevent *kevp, int count);
105 static int kqueue_register(struct kqueue *kq, struct kevent *kev,
106 struct thread *td, int mflag);
107 static int kqueue_acquire(struct file *fp, struct kqueue **kqp);
108 static void kqueue_release(struct kqueue *kq, int locked);
109 static void kqueue_destroy(struct kqueue *kq);
110 static void kqueue_drain(struct kqueue *kq, struct thread *td);
111 static int kqueue_expand(struct kqueue *kq, struct filterops *fops,
112 uintptr_t ident, int mflag);
113 static void kqueue_task(void *arg, int pending);
114 static int kqueue_scan(struct kqueue *kq, int maxevents,
115 struct kevent_copyops *k_ops,
116 const struct timespec *timeout,
117 struct kevent *keva, struct thread *td);
118 static void kqueue_wakeup(struct kqueue *kq);
119 static struct filterops *kqueue_fo_find(int filt);
120 static void kqueue_fo_release(int filt);
121 struct g_kevent_args;
122 static int kern_kevent_generic(struct thread *td,
123 struct g_kevent_args *uap,
124 struct kevent_copyops *k_ops, const char *struct_name);
126 static fo_ioctl_t kqueue_ioctl;
127 static fo_poll_t kqueue_poll;
128 static fo_kqfilter_t kqueue_kqfilter;
129 static fo_stat_t kqueue_stat;
130 static fo_close_t kqueue_close;
131 static fo_fill_kinfo_t kqueue_fill_kinfo;
133 static struct fileops kqueueops = {
134 .fo_read = invfo_rdwr,
135 .fo_write = invfo_rdwr,
136 .fo_truncate = invfo_truncate,
137 .fo_ioctl = kqueue_ioctl,
138 .fo_poll = kqueue_poll,
139 .fo_kqfilter = kqueue_kqfilter,
140 .fo_stat = kqueue_stat,
141 .fo_close = kqueue_close,
142 .fo_chmod = invfo_chmod,
143 .fo_chown = invfo_chown,
144 .fo_sendfile = invfo_sendfile,
145 .fo_fill_kinfo = kqueue_fill_kinfo,
148 static int knote_attach(struct knote *kn, struct kqueue *kq);
149 static void knote_drop(struct knote *kn, struct thread *td);
150 static void knote_drop_detached(struct knote *kn, struct thread *td);
151 static void knote_enqueue(struct knote *kn);
152 static void knote_dequeue(struct knote *kn);
153 static void knote_init(void);
154 static struct knote *knote_alloc(int mflag);
155 static void knote_free(struct knote *kn);
157 static void filt_kqdetach(struct knote *kn);
158 static int filt_kqueue(struct knote *kn, long hint);
159 static int filt_procattach(struct knote *kn);
160 static void filt_procdetach(struct knote *kn);
161 static int filt_proc(struct knote *kn, long hint);
162 static int filt_fileattach(struct knote *kn);
163 static void filt_timerexpire(void *knx);
164 static void filt_timerexpire_l(struct knote *kn, bool proc_locked);
165 static int filt_timerattach(struct knote *kn);
166 static void filt_timerdetach(struct knote *kn);
167 static void filt_timerstart(struct knote *kn, sbintime_t to);
168 static void filt_timertouch(struct knote *kn, struct kevent *kev,
170 static int filt_timervalidate(struct knote *kn, sbintime_t *to);
171 static int filt_timer(struct knote *kn, long hint);
172 static int filt_userattach(struct knote *kn);
173 static void filt_userdetach(struct knote *kn);
174 static int filt_user(struct knote *kn, long hint);
175 static void filt_usertouch(struct knote *kn, struct kevent *kev,
178 static struct filterops file_filtops = {
180 .f_attach = filt_fileattach,
182 static struct filterops kqread_filtops = {
184 .f_detach = filt_kqdetach,
185 .f_event = filt_kqueue,
187 /* XXX - move to kern_proc.c? */
188 static struct filterops proc_filtops = {
190 .f_attach = filt_procattach,
191 .f_detach = filt_procdetach,
192 .f_event = filt_proc,
194 static struct filterops timer_filtops = {
196 .f_attach = filt_timerattach,
197 .f_detach = filt_timerdetach,
198 .f_event = filt_timer,
199 .f_touch = filt_timertouch,
201 static struct filterops user_filtops = {
202 .f_attach = filt_userattach,
203 .f_detach = filt_userdetach,
204 .f_event = filt_user,
205 .f_touch = filt_usertouch,
208 static uma_zone_t knote_zone;
209 static unsigned int __exclusive_cache_line kq_ncallouts;
210 static unsigned int kq_calloutmax = 4 * 1024;
211 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
212 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
214 /* XXX - ensure not influx ? */
215 #define KNOTE_ACTIVATE(kn, islock) do { \
217 mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED); \
219 KQ_LOCK((kn)->kn_kq); \
220 (kn)->kn_status |= KN_ACTIVE; \
221 if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \
222 knote_enqueue((kn)); \
224 KQ_UNLOCK((kn)->kn_kq); \
226 #define KQ_LOCK(kq) do { \
227 mtx_lock(&(kq)->kq_lock); \
229 #define KQ_FLUX_WAKEUP(kq) do { \
230 if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) { \
231 (kq)->kq_state &= ~KQ_FLUXWAIT; \
235 #define KQ_UNLOCK_FLUX(kq) do { \
236 KQ_FLUX_WAKEUP(kq); \
237 mtx_unlock(&(kq)->kq_lock); \
239 #define KQ_UNLOCK(kq) do { \
240 mtx_unlock(&(kq)->kq_lock); \
242 #define KQ_OWNED(kq) do { \
243 mtx_assert(&(kq)->kq_lock, MA_OWNED); \
245 #define KQ_NOTOWNED(kq) do { \
246 mtx_assert(&(kq)->kq_lock, MA_NOTOWNED); \
249 static struct knlist *
250 kn_list_lock(struct knote *kn)
256 knl->kl_lock(knl->kl_lockarg);
261 kn_list_unlock(struct knlist *knl)
267 do_free = knl->kl_autodestroy && knlist_empty(knl);
268 knl->kl_unlock(knl->kl_lockarg);
276 kn_in_flux(struct knote *kn)
279 return (kn->kn_influx > 0);
283 kn_enter_flux(struct knote *kn)
287 MPASS(kn->kn_influx < INT_MAX);
292 kn_leave_flux(struct knote *kn)
296 MPASS(kn->kn_influx > 0);
298 return (kn->kn_influx == 0);
301 #define KNL_ASSERT_LOCK(knl, islocked) do { \
303 KNL_ASSERT_LOCKED(knl); \
305 KNL_ASSERT_UNLOCKED(knl); \
308 #define KNL_ASSERT_LOCKED(knl) do { \
309 knl->kl_assert_lock((knl)->kl_lockarg, LA_LOCKED); \
311 #define KNL_ASSERT_UNLOCKED(knl) do { \
312 knl->kl_assert_lock((knl)->kl_lockarg, LA_UNLOCKED); \
314 #else /* !INVARIANTS */
315 #define KNL_ASSERT_LOCKED(knl) do {} while(0)
316 #define KNL_ASSERT_UNLOCKED(knl) do {} while (0)
317 #endif /* INVARIANTS */
320 #define KN_HASHSIZE 64 /* XXX should be tunable */
323 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask))
326 filt_nullattach(struct knote *kn)
332 struct filterops null_filtops = {
334 .f_attach = filt_nullattach,
337 /* XXX - make SYSINIT to add these, and move into respective modules. */
338 extern struct filterops sig_filtops;
339 extern struct filterops fs_filtops;
342 * Table for for all system-defined filters.
344 static struct mtx filterops_lock;
345 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
348 struct filterops *for_fop;
351 } sysfilt_ops[EVFILT_SYSCOUNT] = {
352 { &file_filtops, 1 }, /* EVFILT_READ */
353 { &file_filtops, 1 }, /* EVFILT_WRITE */
354 { &null_filtops }, /* EVFILT_AIO */
355 { &file_filtops, 1 }, /* EVFILT_VNODE */
356 { &proc_filtops, 1 }, /* EVFILT_PROC */
357 { &sig_filtops, 1 }, /* EVFILT_SIGNAL */
358 { &timer_filtops, 1 }, /* EVFILT_TIMER */
359 { &file_filtops, 1 }, /* EVFILT_PROCDESC */
360 { &fs_filtops, 1 }, /* EVFILT_FS */
361 { &null_filtops }, /* EVFILT_LIO */
362 { &user_filtops, 1 }, /* EVFILT_USER */
363 { &null_filtops }, /* EVFILT_SENDFILE */
364 { &file_filtops, 1 }, /* EVFILT_EMPTY */
368 * Simple redirection for all cdevsw style objects to call their fo_kqfilter
372 filt_fileattach(struct knote *kn)
375 return (fo_kqfilter(kn->kn_fp, kn));
380 kqueue_kqfilter(struct file *fp, struct knote *kn)
382 struct kqueue *kq = kn->kn_fp->f_data;
384 if (kn->kn_filter != EVFILT_READ)
387 kn->kn_status |= KN_KQUEUE;
388 kn->kn_fop = &kqread_filtops;
389 knlist_add(&kq->kq_sel.si_note, kn, 0);
395 filt_kqdetach(struct knote *kn)
397 struct kqueue *kq = kn->kn_fp->f_data;
399 knlist_remove(&kq->kq_sel.si_note, kn, 0);
404 filt_kqueue(struct knote *kn, long hint)
406 struct kqueue *kq = kn->kn_fp->f_data;
408 kn->kn_data = kq->kq_count;
409 return (kn->kn_data > 0);
412 /* XXX - move to kern_proc.c? */
414 filt_procattach(struct knote *kn)
418 bool exiting, immediate;
420 exiting = immediate = false;
421 if (kn->kn_sfflags & NOTE_EXIT)
422 p = pfind_any(kn->kn_id);
424 p = pfind(kn->kn_id);
427 if (p->p_flag & P_WEXIT)
430 if ((error = p_cansee(curthread, p))) {
435 kn->kn_ptr.p_proc = p;
436 kn->kn_flags |= EV_CLEAR; /* automatically set */
439 * Internal flag indicating registration done by kernel for the
440 * purposes of getting a NOTE_CHILD notification.
442 if (kn->kn_flags & EV_FLAG2) {
443 kn->kn_flags &= ~EV_FLAG2;
444 kn->kn_data = kn->kn_sdata; /* ppid */
445 kn->kn_fflags = NOTE_CHILD;
446 kn->kn_sfflags &= ~(NOTE_EXIT | NOTE_EXEC | NOTE_FORK);
447 immediate = true; /* Force immediate activation of child note. */
450 * Internal flag indicating registration done by kernel (for other than
453 if (kn->kn_flags & EV_FLAG1) {
454 kn->kn_flags &= ~EV_FLAG1;
457 knlist_add(p->p_klist, kn, 1);
460 * Immediately activate any child notes or, in the case of a zombie
461 * target process, exit notes. The latter is necessary to handle the
462 * case where the target process, e.g. a child, dies before the kevent
465 if (immediate || (exiting && filt_proc(kn, NOTE_EXIT)))
466 KNOTE_ACTIVATE(kn, 0);
474 * The knote may be attached to a different process, which may exit,
475 * leaving nothing for the knote to be attached to. So when the process
476 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
477 * it will be deleted when read out. However, as part of the knote deletion,
478 * this routine is called, so a check is needed to avoid actually performing
479 * a detach, because the original process does not exist any more.
481 /* XXX - move to kern_proc.c? */
483 filt_procdetach(struct knote *kn)
486 knlist_remove(kn->kn_knlist, kn, 0);
487 kn->kn_ptr.p_proc = NULL;
490 /* XXX - move to kern_proc.c? */
492 filt_proc(struct knote *kn, long hint)
497 p = kn->kn_ptr.p_proc;
498 if (p == NULL) /* already activated, from attach filter */
501 /* Mask off extra data. */
502 event = (u_int)hint & NOTE_PCTRLMASK;
504 /* If the user is interested in this event, record it. */
505 if (kn->kn_sfflags & event)
506 kn->kn_fflags |= event;
508 /* Process is gone, so flag the event as finished. */
509 if (event == NOTE_EXIT) {
510 kn->kn_flags |= EV_EOF | EV_ONESHOT;
511 kn->kn_ptr.p_proc = NULL;
512 if (kn->kn_fflags & NOTE_EXIT)
513 kn->kn_data = KW_EXITCODE(p->p_xexit, p->p_xsig);
514 if (kn->kn_fflags == 0)
515 kn->kn_flags |= EV_DROP;
519 return (kn->kn_fflags != 0);
523 * Called when the process forked. It mostly does the same as the
524 * knote(), activating all knotes registered to be activated when the
525 * process forked. Additionally, for each knote attached to the
526 * parent, check whether user wants to track the new process. If so
527 * attach a new knote to it, and immediately report an event with the
531 knote_fork(struct knlist *list, int pid)
539 KNL_ASSERT_LOCKED(list);
540 if (SLIST_EMPTY(&list->kl_list))
543 memset(&kev, 0, sizeof(kev));
544 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
547 if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
553 * The same as knote(), activate the event.
555 if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
556 if (kn->kn_fop->f_event(kn, NOTE_FORK))
557 KNOTE_ACTIVATE(kn, 1);
563 * The NOTE_TRACK case. In addition to the activation
564 * of the event, we need to register new events to
565 * track the child. Drop the locks in preparation for
566 * the call to kqueue_register().
570 list->kl_unlock(list->kl_lockarg);
573 * Activate existing knote and register tracking knotes with
576 * First register a knote to get just the child notice. This
577 * must be a separate note from a potential NOTE_EXIT
578 * notification since both NOTE_CHILD and NOTE_EXIT are defined
579 * to use the data field (in conflicting ways).
582 kev.filter = kn->kn_filter;
583 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_ONESHOT |
585 kev.fflags = kn->kn_sfflags;
586 kev.data = kn->kn_id; /* parent */
587 kev.udata = kn->kn_kevent.udata;/* preserve udata */
588 error = kqueue_register(kq, &kev, NULL, M_NOWAIT);
590 kn->kn_fflags |= NOTE_TRACKERR;
593 * Then register another knote to track other potential events
594 * from the new process.
597 kev.filter = kn->kn_filter;
598 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
599 kev.fflags = kn->kn_sfflags;
600 kev.data = kn->kn_id; /* parent */
601 kev.udata = kn->kn_kevent.udata;/* preserve udata */
602 error = kqueue_register(kq, &kev, NULL, M_NOWAIT);
604 kn->kn_fflags |= NOTE_TRACKERR;
605 if (kn->kn_fop->f_event(kn, NOTE_FORK))
606 KNOTE_ACTIVATE(kn, 0);
607 list->kl_lock(list->kl_lockarg);
615 * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
616 * interval timer support code.
619 #define NOTE_TIMER_PRECMASK \
620 (NOTE_SECONDS | NOTE_MSECONDS | NOTE_USECONDS | NOTE_NSECONDS)
623 timer2sbintime(int64_t data, int flags)
628 * Macros for converting to the fractional second portion of an
629 * sbintime_t using 64bit multiplication to improve precision.
631 #define NS_TO_SBT(ns) (((ns) * (((uint64_t)1 << 63) / 500000000)) >> 32)
632 #define US_TO_SBT(us) (((us) * (((uint64_t)1 << 63) / 500000)) >> 32)
633 #define MS_TO_SBT(ms) (((ms) * (((uint64_t)1 << 63) / 500)) >> 32)
634 switch (flags & NOTE_TIMER_PRECMASK) {
637 if (data > (SBT_MAX / SBT_1S))
640 return ((sbintime_t)data << 32);
641 case NOTE_MSECONDS: /* FALLTHROUGH */
646 if (secs > (SBT_MAX / SBT_1S))
649 return (secs << 32 | MS_TO_SBT(data % 1000));
651 return (MS_TO_SBT(data));
653 if (data >= 1000000) {
654 secs = data / 1000000;
656 if (secs > (SBT_MAX / SBT_1S))
659 return (secs << 32 | US_TO_SBT(data % 1000000));
661 return (US_TO_SBT(data));
663 if (data >= 1000000000) {
664 secs = data / 1000000000;
666 if (secs > (SBT_MAX / SBT_1S))
669 return (secs << 32 | NS_TO_SBT(data % 1000000000));
671 return (NS_TO_SBT(data));
678 struct kq_timer_cb_data {
684 TAILQ_ENTRY(kq_timer_cb_data) link;
685 sbintime_t next; /* next timer event fires at */
686 sbintime_t to; /* precalculated timer period, 0 for abs */
689 #define KQ_TIMER_CB_ENQUEUED 0x01
692 kqtimer_sched_callout(struct kq_timer_cb_data *kc)
694 callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kc->kn,
695 kc->cpuid, C_ABSOLUTE);
699 kqtimer_proc_continue(struct proc *p)
701 struct kq_timer_cb_data *kc, *kc1;
705 PROC_LOCK_ASSERT(p, MA_OWNED);
710 TAILQ_FOREACH_SAFE(kc, &p->p_kqtim_stop, link, kc1) {
711 TAILQ_REMOVE(&p->p_kqtim_stop, kc, link);
712 kc->flags &= ~KQ_TIMER_CB_ENQUEUED;
714 filt_timerexpire_l(kc->kn, true);
716 kqtimer_sched_callout(kc);
721 filt_timerexpire_l(struct knote *kn, bool proc_locked)
723 struct kq_timer_cb_data *kc;
730 if ((kn->kn_flags & EV_ONESHOT) != 0 || kc->to == 0) {
732 KNOTE_ACTIVATE(kn, 0);
737 if (now >= kc->next) {
738 delta = (now - kc->next) / kc->to;
741 kn->kn_data += delta;
742 kc->next += (delta + 1) * kc->to;
743 if (now >= kc->next) /* overflow */
744 kc->next = now + kc->to;
745 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */
749 * Initial check for stopped kc->p is racy. It is fine to
750 * miss the set of the stop flags, at worst we would schedule
751 * one more callout. On the other hand, it is not fine to not
752 * schedule when we we missed clearing of the flags, we
753 * recheck them under the lock and observe consistent state.
756 if (P_SHOULDSTOP(p) || P_KILLED(p)) {
759 if (P_SHOULDSTOP(p) || P_KILLED(p)) {
760 if ((kc->flags & KQ_TIMER_CB_ENQUEUED) == 0) {
761 kc->flags |= KQ_TIMER_CB_ENQUEUED;
762 TAILQ_INSERT_TAIL(&p->p_kqtim_stop, kc, link);
771 kqtimer_sched_callout(kc);
775 filt_timerexpire(void *knx)
777 filt_timerexpire_l(knx, false);
781 * data contains amount of time to sleep
784 filt_timervalidate(struct knote *kn, sbintime_t *to)
789 if (kn->kn_sdata < 0)
791 if (kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
794 * The only fflags values supported are the timer unit
795 * (precision) and the absolute time indicator.
797 if ((kn->kn_sfflags & ~(NOTE_TIMER_PRECMASK | NOTE_ABSTIME)) != 0)
800 *to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
801 if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
812 filt_timerattach(struct knote *kn)
814 struct kq_timer_cb_data *kc;
818 error = filt_timervalidate(kn, &to);
822 if (atomic_fetchadd_int(&kq_ncallouts, 1) + 1 > kq_calloutmax) {
823 atomic_subtract_int(&kq_ncallouts, 1);
827 if ((kn->kn_sfflags & NOTE_ABSTIME) == 0)
828 kn->kn_flags |= EV_CLEAR; /* automatically set */
829 kn->kn_status &= ~KN_DETACHED; /* knlist_add clears it */
830 kn->kn_ptr.p_v = kc = malloc(sizeof(*kc), M_KQUEUE, M_WAITOK);
833 kc->cpuid = PCPU_GET(cpuid);
835 callout_init(&kc->c, 1);
836 filt_timerstart(kn, to);
842 filt_timerstart(struct knote *kn, sbintime_t to)
844 struct kq_timer_cb_data *kc;
847 if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
851 kc->next = to + sbinuptime();
854 kqtimer_sched_callout(kc);
858 filt_timerdetach(struct knote *kn)
860 struct kq_timer_cb_data *kc;
861 unsigned int old __unused;
864 callout_drain(&kc->c);
865 if ((kc->flags & KQ_TIMER_CB_ENQUEUED) != 0) {
867 TAILQ_REMOVE(&kc->p->p_kqtim_stop, kc, link);
871 old = atomic_fetchadd_int(&kq_ncallouts, -1);
872 KASSERT(old > 0, ("Number of callouts cannot become negative"));
873 kn->kn_status |= KN_DETACHED; /* knlist_remove sets it */
877 filt_timertouch(struct knote *kn, struct kevent *kev, u_long type)
879 struct kq_timer_cb_data *kc;
886 /* Handle re-added timers that update data/fflags */
887 if (kev->flags & EV_ADD) {
890 /* Drain any existing callout. */
891 callout_drain(&kc->c);
893 /* Throw away any existing undelivered record
894 * of the timer expiration. This is done under
895 * the presumption that if a process is
896 * re-adding this timer with new parameters,
897 * it is no longer interested in what may have
898 * happened under the old parameters. If it is
899 * interested, it can wait for the expiration,
900 * delete the old timer definition, and then
903 * This has to be done while the kq is locked:
904 * - if enqueued, dequeue
905 * - make it no longer active
906 * - clear the count of expiration events
910 if (kn->kn_status & KN_QUEUED)
913 kn->kn_status &= ~KN_ACTIVE;
917 /* Reschedule timer based on new data/fflags */
918 kn->kn_sfflags = kev->fflags;
919 kn->kn_sdata = kev->data;
920 error = filt_timervalidate(kn, &to);
922 kn->kn_flags |= EV_ERROR;
925 filt_timerstart(kn, to);
930 *kev = kn->kn_kevent;
931 if (kn->kn_flags & EV_CLEAR) {
938 panic("filt_timertouch() - invalid type (%ld)", type);
944 filt_timer(struct knote *kn, long hint)
947 return (kn->kn_data != 0);
951 filt_userattach(struct knote *kn)
955 * EVFILT_USER knotes are not attached to anything in the kernel.
958 if (kn->kn_fflags & NOTE_TRIGGER)
966 filt_userdetach(__unused struct knote *kn)
970 * EVFILT_USER knotes are not attached to anything in the kernel.
975 filt_user(struct knote *kn, __unused long hint)
978 return (kn->kn_hookid);
982 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
988 if (kev->fflags & NOTE_TRIGGER)
991 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
992 kev->fflags &= NOTE_FFLAGSMASK;
998 kn->kn_sfflags &= kev->fflags;
1002 kn->kn_sfflags |= kev->fflags;
1006 kn->kn_sfflags = kev->fflags;
1010 /* XXX Return error? */
1013 kn->kn_sdata = kev->data;
1014 if (kev->flags & EV_CLEAR) {
1022 *kev = kn->kn_kevent;
1023 kev->fflags = kn->kn_sfflags;
1024 kev->data = kn->kn_sdata;
1025 if (kn->kn_flags & EV_CLEAR) {
1033 panic("filt_usertouch() - invalid type (%ld)", type);
1039 sys_kqueue(struct thread *td, struct kqueue_args *uap)
1042 return (kern_kqueue(td, 0, NULL));
1046 kqueue_init(struct kqueue *kq)
1049 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
1050 TAILQ_INIT(&kq->kq_head);
1051 knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
1052 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
1056 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
1058 struct filedesc *fdp;
1064 fdp = td->td_proc->p_fd;
1065 cred = td->td_ucred;
1066 if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
1069 error = falloc_caps(td, &fp, &fd, flags, fcaps);
1071 chgkqcnt(cred->cr_ruidinfo, -1, 0);
1075 /* An extra reference on `fp' has been held for us by falloc(). */
1076 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
1079 kq->kq_cred = crhold(cred);
1081 FILEDESC_XLOCK(fdp);
1082 TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
1083 FILEDESC_XUNLOCK(fdp);
1085 finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
1088 td->td_retval[0] = fd;
1092 struct g_kevent_args {
1098 const struct timespec *timeout;
1102 sys_kevent(struct thread *td, struct kevent_args *uap)
1104 struct kevent_copyops k_ops = {
1106 .k_copyout = kevent_copyout,
1107 .k_copyin = kevent_copyin,
1108 .kevent_size = sizeof(struct kevent),
1110 struct g_kevent_args gk_args = {
1112 .changelist = uap->changelist,
1113 .nchanges = uap->nchanges,
1114 .eventlist = uap->eventlist,
1115 .nevents = uap->nevents,
1116 .timeout = uap->timeout,
1119 return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent"));
1123 kern_kevent_generic(struct thread *td, struct g_kevent_args *uap,
1124 struct kevent_copyops *k_ops, const char *struct_name)
1126 struct timespec ts, *tsp;
1128 struct kevent *eventlist = uap->eventlist;
1132 if (uap->timeout != NULL) {
1133 error = copyin(uap->timeout, &ts, sizeof(ts));
1141 if (KTRPOINT(td, KTR_STRUCT_ARRAY))
1142 ktrstructarray(struct_name, UIO_USERSPACE, uap->changelist,
1143 uap->nchanges, k_ops->kevent_size);
1146 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
1150 if (error == 0 && KTRPOINT(td, KTR_STRUCT_ARRAY))
1151 ktrstructarray(struct_name, UIO_USERSPACE, eventlist,
1152 td->td_retval[0], k_ops->kevent_size);
1159 * Copy 'count' items into the destination list pointed to by uap->eventlist.
1162 kevent_copyout(void *arg, struct kevent *kevp, int count)
1164 struct kevent_args *uap;
1167 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1168 uap = (struct kevent_args *)arg;
1170 error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
1172 uap->eventlist += count;
1177 * Copy 'count' items from the list pointed to by uap->changelist.
1180 kevent_copyin(void *arg, struct kevent *kevp, int count)
1182 struct kevent_args *uap;
1185 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1186 uap = (struct kevent_args *)arg;
1188 error = copyin(uap->changelist, kevp, count * sizeof *kevp);
1190 uap->changelist += count;
1194 #ifdef COMPAT_FREEBSD11
1196 kevent11_copyout(void *arg, struct kevent *kevp, int count)
1198 struct freebsd11_kevent_args *uap;
1199 struct kevent_freebsd11 kev11;
1202 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1203 uap = (struct freebsd11_kevent_args *)arg;
1205 for (i = 0; i < count; i++) {
1206 kev11.ident = kevp->ident;
1207 kev11.filter = kevp->filter;
1208 kev11.flags = kevp->flags;
1209 kev11.fflags = kevp->fflags;
1210 kev11.data = kevp->data;
1211 kev11.udata = kevp->udata;
1212 error = copyout(&kev11, uap->eventlist, sizeof(kev11));
1222 * Copy 'count' items from the list pointed to by uap->changelist.
1225 kevent11_copyin(void *arg, struct kevent *kevp, int count)
1227 struct freebsd11_kevent_args *uap;
1228 struct kevent_freebsd11 kev11;
1231 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1232 uap = (struct freebsd11_kevent_args *)arg;
1234 for (i = 0; i < count; i++) {
1235 error = copyin(uap->changelist, &kev11, sizeof(kev11));
1238 kevp->ident = kev11.ident;
1239 kevp->filter = kev11.filter;
1240 kevp->flags = kev11.flags;
1241 kevp->fflags = kev11.fflags;
1242 kevp->data = (uintptr_t)kev11.data;
1243 kevp->udata = kev11.udata;
1244 bzero(&kevp->ext, sizeof(kevp->ext));
1252 freebsd11_kevent(struct thread *td, struct freebsd11_kevent_args *uap)
1254 struct kevent_copyops k_ops = {
1256 .k_copyout = kevent11_copyout,
1257 .k_copyin = kevent11_copyin,
1258 .kevent_size = sizeof(struct kevent_freebsd11),
1260 struct g_kevent_args gk_args = {
1262 .changelist = uap->changelist,
1263 .nchanges = uap->nchanges,
1264 .eventlist = uap->eventlist,
1265 .nevents = uap->nevents,
1266 .timeout = uap->timeout,
1269 return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent_freebsd11"));
1274 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
1275 struct kevent_copyops *k_ops, const struct timespec *timeout)
1277 cap_rights_t rights;
1281 cap_rights_init_zero(&rights);
1283 cap_rights_set_one(&rights, CAP_KQUEUE_CHANGE);
1285 cap_rights_set_one(&rights, CAP_KQUEUE_EVENT);
1286 error = fget(td, fd, &rights, &fp);
1290 error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
1297 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
1298 struct kevent_copyops *k_ops, const struct timespec *timeout)
1300 struct kevent keva[KQ_NEVENTS];
1301 struct kevent *kevp, *changes;
1302 int i, n, nerrors, error;
1305 while (nchanges > 0) {
1306 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
1307 error = k_ops->k_copyin(k_ops->arg, keva, n);
1311 for (i = 0; i < n; i++) {
1315 kevp->flags &= ~EV_SYSFLAGS;
1316 error = kqueue_register(kq, kevp, td, M_WAITOK);
1317 if (error || (kevp->flags & EV_RECEIPT)) {
1320 kevp->flags = EV_ERROR;
1322 (void)k_ops->k_copyout(k_ops->arg, kevp, 1);
1330 td->td_retval[0] = nerrors;
1334 return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
1338 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
1339 struct kevent_copyops *k_ops, const struct timespec *timeout)
1344 error = kqueue_acquire(fp, &kq);
1347 error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
1348 kqueue_release(kq, 0);
1353 * Performs a kevent() call on a temporarily created kqueue. This can be
1354 * used to perform one-shot polling, similar to poll() and select().
1357 kern_kevent_anonymous(struct thread *td, int nevents,
1358 struct kevent_copyops *k_ops)
1360 struct kqueue kq = {};
1365 error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
1366 kqueue_drain(&kq, td);
1367 kqueue_destroy(&kq);
1372 kqueue_add_filteropts(int filt, struct filterops *filtops)
1377 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1379 "trying to add a filterop that is out of range: %d is beyond %d\n",
1380 ~filt, EVFILT_SYSCOUNT);
1383 mtx_lock(&filterops_lock);
1384 if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1385 sysfilt_ops[~filt].for_fop != NULL)
1388 sysfilt_ops[~filt].for_fop = filtops;
1389 sysfilt_ops[~filt].for_refcnt = 0;
1391 mtx_unlock(&filterops_lock);
1397 kqueue_del_filteropts(int filt)
1402 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1405 mtx_lock(&filterops_lock);
1406 if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1407 sysfilt_ops[~filt].for_fop == NULL)
1409 else if (sysfilt_ops[~filt].for_refcnt != 0)
1412 sysfilt_ops[~filt].for_fop = &null_filtops;
1413 sysfilt_ops[~filt].for_refcnt = 0;
1415 mtx_unlock(&filterops_lock);
1420 static struct filterops *
1421 kqueue_fo_find(int filt)
1424 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1427 if (sysfilt_ops[~filt].for_nolock)
1428 return sysfilt_ops[~filt].for_fop;
1430 mtx_lock(&filterops_lock);
1431 sysfilt_ops[~filt].for_refcnt++;
1432 if (sysfilt_ops[~filt].for_fop == NULL)
1433 sysfilt_ops[~filt].for_fop = &null_filtops;
1434 mtx_unlock(&filterops_lock);
1436 return sysfilt_ops[~filt].for_fop;
1440 kqueue_fo_release(int filt)
1443 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1446 if (sysfilt_ops[~filt].for_nolock)
1449 mtx_lock(&filterops_lock);
1450 KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1451 ("filter object refcount not valid on release"));
1452 sysfilt_ops[~filt].for_refcnt--;
1453 mtx_unlock(&filterops_lock);
1457 * A ref to kq (obtained via kqueue_acquire) must be held.
1460 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td,
1463 struct filterops *fops;
1465 struct knote *kn, *tkn;
1467 int error, filt, event;
1468 int haskqglobal, filedesc_unlock;
1470 if ((kev->flags & (EV_ENABLE | EV_DISABLE)) == (EV_ENABLE | EV_DISABLE))
1478 filedesc_unlock = 0;
1481 fops = kqueue_fo_find(filt);
1485 if (kev->flags & EV_ADD) {
1487 * Prevent waiting with locks. Non-sleepable
1488 * allocation failures are handled in the loop, only
1489 * if the spare knote appears to be actually required.
1491 tkn = knote_alloc(mflag);
1498 KASSERT(td != NULL, ("td is NULL"));
1499 if (kev->ident > INT_MAX)
1502 error = fget(td, kev->ident, &cap_event_rights, &fp);
1506 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1507 kev->ident, M_NOWAIT) != 0) {
1511 error = kqueue_expand(kq, fops, kev->ident, mflag);
1517 if (fp->f_type == DTYPE_KQUEUE) {
1519 * If we add some intelligence about what we are doing,
1520 * we should be able to support events on ourselves.
1521 * We need to know when we are doing this to prevent
1522 * getting both the knlist lock and the kq lock since
1523 * they are the same thing.
1525 if (fp->f_data == kq) {
1531 * Pre-lock the filedesc before the global
1532 * lock mutex, see the comment in
1535 FILEDESC_XLOCK(td->td_proc->p_fd);
1536 filedesc_unlock = 1;
1537 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1541 if (kev->ident < kq->kq_knlistsize) {
1542 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1543 if (kev->filter == kn->kn_filter)
1547 if ((kev->flags & EV_ADD) == EV_ADD) {
1548 error = kqueue_expand(kq, fops, kev->ident, mflag);
1556 * If possible, find an existing knote to use for this kevent.
1558 if (kev->filter == EVFILT_PROC &&
1559 (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1560 /* This is an internal creation of a process tracking
1561 * note. Don't attempt to coalesce this with an
1565 } else if (kq->kq_knhashmask != 0) {
1568 list = &kq->kq_knhash[
1569 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1570 SLIST_FOREACH(kn, list, kn_link)
1571 if (kev->ident == kn->kn_id &&
1572 kev->filter == kn->kn_filter)
1577 /* knote is in the process of changing, wait for it to stabilize. */
1578 if (kn != NULL && kn_in_flux(kn)) {
1579 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1580 if (filedesc_unlock) {
1581 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1582 filedesc_unlock = 0;
1584 kq->kq_state |= KQ_FLUXWAIT;
1585 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1594 * kn now contains the matching knote, or NULL if no match
1597 if (kev->flags & EV_ADD) {
1609 * apply reference counts to knote structure, and
1610 * do not release it at the end of this routine.
1615 kn->kn_sfflags = kev->fflags;
1616 kn->kn_sdata = kev->data;
1619 kn->kn_kevent = *kev;
1620 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1621 EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1622 kn->kn_status = KN_DETACHED;
1623 if ((kev->flags & EV_DISABLE) != 0)
1624 kn->kn_status |= KN_DISABLED;
1627 error = knote_attach(kn, kq);
1634 if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1635 knote_drop_detached(kn, td);
1638 knl = kn_list_lock(kn);
1641 /* No matching knote and the EV_ADD flag is not set. */
1648 if (kev->flags & EV_DELETE) {
1655 if (kev->flags & EV_FORCEONESHOT) {
1656 kn->kn_flags |= EV_ONESHOT;
1657 KNOTE_ACTIVATE(kn, 1);
1660 if ((kev->flags & EV_ENABLE) != 0)
1661 kn->kn_status &= ~KN_DISABLED;
1662 else if ((kev->flags & EV_DISABLE) != 0)
1663 kn->kn_status |= KN_DISABLED;
1666 * The user may change some filter values after the initial EV_ADD,
1667 * but doing so will not reset any filter which has already been
1670 kn->kn_status |= KN_SCAN;
1673 knl = kn_list_lock(kn);
1674 kn->kn_kevent.udata = kev->udata;
1675 if (!fops->f_isfd && fops->f_touch != NULL) {
1676 fops->f_touch(kn, kev, EVENT_REGISTER);
1678 kn->kn_sfflags = kev->fflags;
1679 kn->kn_sdata = kev->data;
1684 * We can get here with kn->kn_knlist == NULL. This can happen when
1685 * the initial attach event decides that the event is "completed"
1686 * already, e.g., filt_procattach() is called on a zombie process. It
1687 * will call filt_proc() which will remove it from the list, and NULL
1690 * KN_DISABLED will be stable while the knote is in flux, so the
1691 * unlocked read will not race with an update.
1693 if ((kn->kn_status & KN_DISABLED) == 0)
1694 event = kn->kn_fop->f_event(kn, 0);
1700 kn->kn_status |= KN_ACTIVE;
1701 if ((kn->kn_status & (KN_ACTIVE | KN_DISABLED | KN_QUEUED)) ==
1704 kn->kn_status &= ~KN_SCAN;
1706 kn_list_unlock(knl);
1710 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1711 if (filedesc_unlock)
1712 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1717 kqueue_fo_release(filt);
1722 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1730 if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1734 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1745 kqueue_release(struct kqueue *kq, int locked)
1752 if (kq->kq_refcnt == 1)
1753 wakeup(&kq->kq_refcnt);
1759 kqueue_schedtask(struct kqueue *kq)
1763 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1764 ("scheduling kqueue task while draining"));
1766 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1767 taskqueue_enqueue(taskqueue_kqueue_ctx, &kq->kq_task);
1768 kq->kq_state |= KQ_TASKSCHED;
1773 * Expand the kq to make sure we have storage for fops/ident pair.
1775 * Return 0 on success (or no work necessary), return errno on failure.
1778 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1781 struct klist *list, *tmp_knhash, *to_free;
1782 u_long tmp_knhashmask;
1783 int error, fd, size;
1791 if (kq->kq_knlistsize <= fd) {
1792 size = kq->kq_knlistsize;
1795 list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1799 if ((kq->kq_state & KQ_CLOSING) != 0) {
1802 } else if (kq->kq_knlistsize > fd) {
1805 if (kq->kq_knlist != NULL) {
1806 bcopy(kq->kq_knlist, list,
1807 kq->kq_knlistsize * sizeof(*list));
1808 to_free = kq->kq_knlist;
1809 kq->kq_knlist = NULL;
1811 bzero((caddr_t)list +
1812 kq->kq_knlistsize * sizeof(*list),
1813 (size - kq->kq_knlistsize) * sizeof(*list));
1814 kq->kq_knlistsize = size;
1815 kq->kq_knlist = list;
1820 if (kq->kq_knhashmask == 0) {
1821 tmp_knhash = hashinit_flags(KN_HASHSIZE, M_KQUEUE,
1822 &tmp_knhashmask, (mflag & M_WAITOK) != 0 ?
1823 HASH_WAITOK : HASH_NOWAIT);
1824 if (tmp_knhash == NULL)
1827 if ((kq->kq_state & KQ_CLOSING) != 0) {
1828 to_free = tmp_knhash;
1830 } else if (kq->kq_knhashmask == 0) {
1831 kq->kq_knhash = tmp_knhash;
1832 kq->kq_knhashmask = tmp_knhashmask;
1834 to_free = tmp_knhash;
1839 free(to_free, M_KQUEUE);
1846 kqueue_task(void *arg, int pending)
1854 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1857 KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1859 kq->kq_state &= ~KQ_TASKSCHED;
1860 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1861 wakeup(&kq->kq_state);
1864 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1868 * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1869 * We treat KN_MARKER knotes as if they are in flux.
1872 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1873 const struct timespec *tsp, struct kevent *keva, struct thread *td)
1875 struct kevent *kevp;
1876 struct knote *kn, *marker;
1878 sbintime_t asbt, rsbt;
1879 int count, error, haskqglobal, influx, nkev, touch;
1891 if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1892 tsp->tv_nsec >= 1000000000) {
1896 if (timespecisset(tsp)) {
1897 if (tsp->tv_sec <= INT32_MAX) {
1898 rsbt = tstosbt(*tsp);
1899 if (TIMESEL(&asbt, rsbt))
1900 asbt += tc_tick_sbt;
1901 if (asbt <= SBT_MAX - rsbt)
1905 rsbt >>= tc_precexp;
1912 marker = knote_alloc(M_WAITOK);
1913 marker->kn_status = KN_MARKER;
1918 if (kq->kq_count == 0) {
1920 error = EWOULDBLOCK;
1922 kq->kq_state |= KQ_SLEEP;
1923 error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1924 "kqread", asbt, rsbt, C_ABSOLUTE);
1928 /* don't restart after signals... */
1929 if (error == ERESTART)
1931 else if (error == EWOULDBLOCK)
1936 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1940 kn = TAILQ_FIRST(&kq->kq_head);
1942 if ((kn->kn_status == KN_MARKER && kn != marker) ||
1948 kq->kq_state |= KQ_FLUXWAIT;
1949 error = msleep(kq, &kq->kq_lock, PSOCK,
1954 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1955 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1956 kn->kn_status &= ~KN_QUEUED;
1962 if (count == maxevents)
1966 KASSERT(!kn_in_flux(kn),
1967 ("knote %p is unexpectedly in flux", kn));
1969 if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1970 kn->kn_status &= ~KN_QUEUED;
1975 * We don't need to lock the list since we've
1976 * marked it as in flux.
1981 } else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1982 kn->kn_status &= ~KN_QUEUED;
1987 * We don't need to lock the list since we've
1988 * marked the knote as being in flux.
1990 *kevp = kn->kn_kevent;
1995 kn->kn_status |= KN_SCAN;
1998 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1999 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
2000 knl = kn_list_lock(kn);
2001 if (kn->kn_fop->f_event(kn, 0) == 0) {
2003 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
2004 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE |
2008 kn_list_unlock(knl);
2012 touch = (!kn->kn_fop->f_isfd &&
2013 kn->kn_fop->f_touch != NULL);
2015 kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
2017 *kevp = kn->kn_kevent;
2019 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
2020 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
2022 * Manually clear knotes who weren't
2025 if (touch == 0 && kn->kn_flags & EV_CLEAR) {
2029 if (kn->kn_flags & EV_DISPATCH)
2030 kn->kn_status |= KN_DISABLED;
2031 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
2034 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2036 kn->kn_status &= ~KN_SCAN;
2038 kn_list_unlock(knl);
2042 /* we are returning a copy to the user */
2047 if (nkev == KQ_NEVENTS) {
2050 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
2058 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
2066 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
2067 td->td_retval[0] = maxevents - count;
2073 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
2074 struct ucred *active_cred, struct thread *td)
2077 * Enabling sigio causes two major problems:
2078 * 1) infinite recursion:
2079 * Synopsys: kevent is being used to track signals and have FIOASYNC
2080 * set. On receipt of a signal this will cause a kqueue to recurse
2081 * into itself over and over. Sending the sigio causes the kqueue
2082 * to become ready, which in turn posts sigio again, forever.
2083 * Solution: this can be solved by setting a flag in the kqueue that
2084 * we have a SIGIO in progress.
2085 * 2) locking problems:
2086 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
2087 * us above the proc and pgrp locks.
2088 * Solution: Post a signal using an async mechanism, being sure to
2089 * record a generation count in the delivery so that we do not deliver
2090 * a signal to the wrong process.
2092 * Note, these two mechanisms are somewhat mutually exclusive!
2101 kq->kq_state |= KQ_ASYNC;
2103 kq->kq_state &= ~KQ_ASYNC;
2108 return (fsetown(*(int *)data, &kq->kq_sigio));
2111 *(int *)data = fgetown(&kq->kq_sigio);
2121 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
2128 if ((error = kqueue_acquire(fp, &kq)))
2132 if (events & (POLLIN | POLLRDNORM)) {
2134 revents |= events & (POLLIN | POLLRDNORM);
2136 selrecord(td, &kq->kq_sel);
2137 if (SEL_WAITING(&kq->kq_sel))
2138 kq->kq_state |= KQ_SEL;
2141 kqueue_release(kq, 1);
2148 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
2152 bzero((void *)st, sizeof *st);
2154 * We no longer return kq_count because the unlocked value is useless.
2155 * If you spent all this time getting the count, why not spend your
2156 * syscall better by calling kevent?
2158 * XXX - This is needed for libc_r.
2160 st->st_mode = S_IFIFO;
2165 kqueue_drain(struct kqueue *kq, struct thread *td)
2172 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
2173 ("kqueue already closing"));
2174 kq->kq_state |= KQ_CLOSING;
2175 if (kq->kq_refcnt > 1)
2176 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
2178 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
2180 KASSERT(knlist_empty(&kq->kq_sel.si_note),
2181 ("kqueue's knlist not empty"));
2183 for (i = 0; i < kq->kq_knlistsize; i++) {
2184 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
2185 if (kn_in_flux(kn)) {
2186 kq->kq_state |= KQ_FLUXWAIT;
2187 msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
2196 if (kq->kq_knhashmask != 0) {
2197 for (i = 0; i <= kq->kq_knhashmask; i++) {
2198 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
2199 if (kn_in_flux(kn)) {
2200 kq->kq_state |= KQ_FLUXWAIT;
2201 msleep(kq, &kq->kq_lock, PSOCK,
2213 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
2214 kq->kq_state |= KQ_TASKDRAIN;
2215 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
2218 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2219 selwakeuppri(&kq->kq_sel, PSOCK);
2220 if (!SEL_WAITING(&kq->kq_sel))
2221 kq->kq_state &= ~KQ_SEL;
2228 kqueue_destroy(struct kqueue *kq)
2231 KASSERT(kq->kq_fdp == NULL,
2232 ("kqueue still attached to a file descriptor"));
2233 seldrain(&kq->kq_sel);
2234 knlist_destroy(&kq->kq_sel.si_note);
2235 mtx_destroy(&kq->kq_lock);
2237 if (kq->kq_knhash != NULL)
2238 free(kq->kq_knhash, M_KQUEUE);
2239 if (kq->kq_knlist != NULL)
2240 free(kq->kq_knlist, M_KQUEUE);
2242 funsetown(&kq->kq_sigio);
2247 kqueue_close(struct file *fp, struct thread *td)
2249 struct kqueue *kq = fp->f_data;
2250 struct filedesc *fdp;
2252 int filedesc_unlock;
2254 if ((error = kqueue_acquire(fp, &kq)))
2256 kqueue_drain(kq, td);
2259 * We could be called due to the knote_drop() doing fdrop(),
2260 * called from kqueue_register(). In this case the global
2261 * lock is owned, and filedesc sx is locked before, to not
2262 * take the sleepable lock after non-sleepable.
2266 if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
2267 FILEDESC_XLOCK(fdp);
2268 filedesc_unlock = 1;
2270 filedesc_unlock = 0;
2271 TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
2272 if (filedesc_unlock)
2273 FILEDESC_XUNLOCK(fdp);
2276 chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
2277 crfree(kq->kq_cred);
2285 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2288 kif->kf_type = KF_TYPE_KQUEUE;
2293 kqueue_wakeup(struct kqueue *kq)
2297 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
2298 kq->kq_state &= ~KQ_SLEEP;
2301 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2302 selwakeuppri(&kq->kq_sel, PSOCK);
2303 if (!SEL_WAITING(&kq->kq_sel))
2304 kq->kq_state &= ~KQ_SEL;
2306 if (!knlist_empty(&kq->kq_sel.si_note))
2307 kqueue_schedtask(kq);
2308 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
2309 pgsigio(&kq->kq_sigio, SIGIO, 0);
2314 * Walk down a list of knotes, activating them if their event has triggered.
2316 * There is a possibility to optimize in the case of one kq watching another.
2317 * Instead of scheduling a task to wake it up, you could pass enough state
2318 * down the chain to make up the parent kqueue. Make this code functional
2322 knote(struct knlist *list, long hint, int lockflags)
2325 struct knote *kn, *tkn;
2331 KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
2333 if ((lockflags & KNF_LISTLOCKED) == 0)
2334 list->kl_lock(list->kl_lockarg);
2337 * If we unlock the list lock (and enter influx), we can
2338 * eliminate the kqueue scheduling, but this will introduce
2339 * four lock/unlock's for each knote to test. Also, marker
2340 * would be needed to keep iteration position, since filters
2341 * or other threads could remove events.
2343 SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
2346 if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
2348 * Do not process the influx notes, except for
2349 * the influx coming from the kq unlock in the
2350 * kqueue_scan(). In the later case, we do
2351 * not interfere with the scan, since the code
2352 * fragment in kqueue_scan() locks the knlist,
2353 * and cannot proceed until we finished.
2356 } else if ((lockflags & KNF_NOKQLOCK) != 0) {
2359 error = kn->kn_fop->f_event(kn, hint);
2363 KNOTE_ACTIVATE(kn, 1);
2366 if (kn->kn_fop->f_event(kn, hint))
2367 KNOTE_ACTIVATE(kn, 1);
2371 if ((lockflags & KNF_LISTLOCKED) == 0)
2372 list->kl_unlock(list->kl_lockarg);
2376 * add a knote to a knlist
2379 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2382 KNL_ASSERT_LOCK(knl, islocked);
2383 KQ_NOTOWNED(kn->kn_kq);
2384 KASSERT(kn_in_flux(kn), ("knote %p not in flux", kn));
2385 KASSERT((kn->kn_status & KN_DETACHED) != 0,
2386 ("knote %p was not detached", kn));
2388 knl->kl_lock(knl->kl_lockarg);
2389 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2391 knl->kl_unlock(knl->kl_lockarg);
2393 kn->kn_knlist = knl;
2394 kn->kn_status &= ~KN_DETACHED;
2395 KQ_UNLOCK(kn->kn_kq);
2399 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked,
2403 KASSERT(!kqislocked || knlislocked, ("kq locked w/o knl locked"));
2404 KNL_ASSERT_LOCK(knl, knlislocked);
2405 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2406 KASSERT(kqislocked || kn_in_flux(kn), ("knote %p not in flux", kn));
2407 KASSERT((kn->kn_status & KN_DETACHED) == 0,
2408 ("knote %p was already detached", kn));
2410 knl->kl_lock(knl->kl_lockarg);
2411 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2412 kn->kn_knlist = NULL;
2414 kn_list_unlock(knl);
2417 kn->kn_status |= KN_DETACHED;
2419 KQ_UNLOCK(kn->kn_kq);
2423 * remove knote from the specified knlist
2426 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2429 knlist_remove_kq(knl, kn, islocked, 0);
2433 knlist_empty(struct knlist *knl)
2436 KNL_ASSERT_LOCKED(knl);
2437 return (SLIST_EMPTY(&knl->kl_list));
2440 static struct mtx knlist_lock;
2441 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2443 static void knlist_mtx_lock(void *arg);
2444 static void knlist_mtx_unlock(void *arg);
2447 knlist_mtx_lock(void *arg)
2450 mtx_lock((struct mtx *)arg);
2454 knlist_mtx_unlock(void *arg)
2457 mtx_unlock((struct mtx *)arg);
2461 knlist_mtx_assert_lock(void *arg, int what)
2464 if (what == LA_LOCKED)
2465 mtx_assert((struct mtx *)arg, MA_OWNED);
2467 mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2471 knlist_rw_rlock(void *arg)
2474 rw_rlock((struct rwlock *)arg);
2478 knlist_rw_runlock(void *arg)
2481 rw_runlock((struct rwlock *)arg);
2485 knlist_rw_assert_lock(void *arg, int what)
2488 if (what == LA_LOCKED)
2489 rw_assert((struct rwlock *)arg, RA_LOCKED);
2491 rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2495 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2496 void (*kl_unlock)(void *),
2497 void (*kl_assert_lock)(void *, int))
2501 knl->kl_lockarg = &knlist_lock;
2503 knl->kl_lockarg = lock;
2505 if (kl_lock == NULL)
2506 knl->kl_lock = knlist_mtx_lock;
2508 knl->kl_lock = kl_lock;
2509 if (kl_unlock == NULL)
2510 knl->kl_unlock = knlist_mtx_unlock;
2512 knl->kl_unlock = kl_unlock;
2513 if (kl_assert_lock == NULL)
2514 knl->kl_assert_lock = knlist_mtx_assert_lock;
2516 knl->kl_assert_lock = kl_assert_lock;
2518 knl->kl_autodestroy = 0;
2519 SLIST_INIT(&knl->kl_list);
2523 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2526 knlist_init(knl, lock, NULL, NULL, NULL);
2530 knlist_alloc(struct mtx *lock)
2534 knl = malloc(sizeof(struct knlist), M_KQUEUE, M_WAITOK);
2535 knlist_init_mtx(knl, lock);
2540 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2543 knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2544 knlist_rw_assert_lock);
2548 knlist_destroy(struct knlist *knl)
2551 KASSERT(KNLIST_EMPTY(knl),
2552 ("destroying knlist %p with knotes on it", knl));
2556 knlist_detach(struct knlist *knl)
2559 KNL_ASSERT_LOCKED(knl);
2560 knl->kl_autodestroy = 1;
2561 if (knlist_empty(knl)) {
2562 knlist_destroy(knl);
2563 free(knl, M_KQUEUE);
2568 * Even if we are locked, we may need to drop the lock to allow any influx
2569 * knotes time to "settle".
2572 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2574 struct knote *kn, *kn2;
2577 KASSERT(!knl->kl_autodestroy, ("cleardel for autodestroy %p", knl));
2579 KNL_ASSERT_LOCKED(knl);
2581 KNL_ASSERT_UNLOCKED(knl);
2582 again: /* need to reacquire lock since we have dropped it */
2583 knl->kl_lock(knl->kl_lockarg);
2586 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2589 if (kn_in_flux(kn)) {
2593 knlist_remove_kq(knl, kn, 1, 1);
2597 knote_drop_detached(kn, td);
2599 /* Make sure cleared knotes disappear soon */
2600 kn->kn_flags |= EV_EOF | EV_ONESHOT;
2606 if (!SLIST_EMPTY(&knl->kl_list)) {
2607 /* there are still in flux knotes remaining */
2608 kn = SLIST_FIRST(&knl->kl_list);
2611 KASSERT(kn_in_flux(kn), ("knote removed w/o list lock"));
2612 knl->kl_unlock(knl->kl_lockarg);
2613 kq->kq_state |= KQ_FLUXWAIT;
2614 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2620 KNL_ASSERT_LOCKED(knl);
2622 knl->kl_unlock(knl->kl_lockarg);
2623 KNL_ASSERT_UNLOCKED(knl);
2628 * Remove all knotes referencing a specified fd must be called with FILEDESC
2629 * lock. This prevents a race where a new fd comes along and occupies the
2630 * entry and we attach a knote to the fd.
2633 knote_fdclose(struct thread *td, int fd)
2635 struct filedesc *fdp = td->td_proc->p_fd;
2640 FILEDESC_XLOCK_ASSERT(fdp);
2643 * We shouldn't have to worry about new kevents appearing on fd
2644 * since filedesc is locked.
2646 TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2651 while (kq->kq_knlistsize > fd &&
2652 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2653 if (kn_in_flux(kn)) {
2654 /* someone else might be waiting on our knote */
2657 kq->kq_state |= KQ_FLUXWAIT;
2658 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2672 knote_attach(struct knote *kn, struct kqueue *kq)
2676 KASSERT(kn_in_flux(kn), ("knote %p not marked influx", kn));
2679 if ((kq->kq_state & KQ_CLOSING) != 0)
2681 if (kn->kn_fop->f_isfd) {
2682 if (kn->kn_id >= kq->kq_knlistsize)
2684 list = &kq->kq_knlist[kn->kn_id];
2686 if (kq->kq_knhash == NULL)
2688 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2690 SLIST_INSERT_HEAD(list, kn, kn_link);
2695 knote_drop(struct knote *kn, struct thread *td)
2698 if ((kn->kn_status & KN_DETACHED) == 0)
2699 kn->kn_fop->f_detach(kn);
2700 knote_drop_detached(kn, td);
2704 knote_drop_detached(struct knote *kn, struct thread *td)
2711 KASSERT((kn->kn_status & KN_DETACHED) != 0,
2712 ("knote %p still attached", kn));
2716 KASSERT(kn->kn_influx == 1,
2717 ("knote_drop called on %p with influx %d", kn, kn->kn_influx));
2719 if (kn->kn_fop->f_isfd)
2720 list = &kq->kq_knlist[kn->kn_id];
2722 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2724 if (!SLIST_EMPTY(list))
2725 SLIST_REMOVE(list, kn, knote, kn_link);
2726 if (kn->kn_status & KN_QUEUED)
2730 if (kn->kn_fop->f_isfd) {
2731 fdrop(kn->kn_fp, td);
2734 kqueue_fo_release(kn->kn_kevent.filter);
2740 knote_enqueue(struct knote *kn)
2742 struct kqueue *kq = kn->kn_kq;
2744 KQ_OWNED(kn->kn_kq);
2745 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2747 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2748 kn->kn_status |= KN_QUEUED;
2754 knote_dequeue(struct knote *kn)
2756 struct kqueue *kq = kn->kn_kq;
2758 KQ_OWNED(kn->kn_kq);
2759 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2761 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2762 kn->kn_status &= ~KN_QUEUED;
2770 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2771 NULL, NULL, UMA_ALIGN_PTR, 0);
2773 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2775 static struct knote *
2776 knote_alloc(int mflag)
2779 return (uma_zalloc(knote_zone, mflag | M_ZERO));
2783 knote_free(struct knote *kn)
2786 uma_zfree(knote_zone, kn);
2790 * Register the kev w/ the kq specified by fd.
2793 kqfd_register(int fd, struct kevent *kev, struct thread *td, int mflag)
2797 cap_rights_t rights;
2800 error = fget(td, fd, cap_rights_init_one(&rights, CAP_KQUEUE_CHANGE),
2804 if ((error = kqueue_acquire(fp, &kq)) != 0)
2807 error = kqueue_register(kq, kev, td, mflag);
2808 kqueue_release(kq, 0);