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 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 * 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);
803 if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
806 *to = MAX(0, *to - sbt);
812 filt_timerattach(struct knote *kn)
814 struct kq_timer_cb_data *kc;
819 error = filt_timervalidate(kn, &to);
822 KASSERT(to > 0 || (kn->kn_flags & EV_ONESHOT) != 0 ||
823 (kn->kn_sfflags & NOTE_ABSTIME) != 0,
824 ("%s: periodic timer has a calculated zero timeout", __func__));
826 ("%s: timer has a calculated negative timeout", __func__));
828 if (atomic_fetchadd_int(&kq_ncallouts, 1) + 1 > kq_calloutmax) {
829 atomic_subtract_int(&kq_ncallouts, 1);
833 if ((kn->kn_sfflags & NOTE_ABSTIME) == 0)
834 kn->kn_flags |= EV_CLEAR; /* automatically set */
835 kn->kn_status &= ~KN_DETACHED; /* knlist_add clears it */
836 kn->kn_ptr.p_v = kc = malloc(sizeof(*kc), M_KQUEUE, M_WAITOK);
839 kc->cpuid = PCPU_GET(cpuid);
841 callout_init(&kc->c, 1);
842 filt_timerstart(kn, to);
848 filt_timerstart(struct knote *kn, sbintime_t to)
850 struct kq_timer_cb_data *kc;
853 if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
857 kc->next = to + sbinuptime();
860 kqtimer_sched_callout(kc);
864 filt_timerdetach(struct knote *kn)
866 struct kq_timer_cb_data *kc;
867 unsigned int old __unused;
872 callout_drain(&kc->c);
875 * kqtimer_proc_continue() might have rescheduled this callout.
876 * Double-check, using the process mutex as an interlock.
879 if ((kc->flags & KQ_TIMER_CB_ENQUEUED) != 0) {
880 kc->flags &= ~KQ_TIMER_CB_ENQUEUED;
881 TAILQ_REMOVE(&kc->p->p_kqtim_stop, kc, link);
883 pending = callout_pending(&kc->c);
887 old = atomic_fetchadd_int(&kq_ncallouts, -1);
888 KASSERT(old > 0, ("Number of callouts cannot become negative"));
889 kn->kn_status |= KN_DETACHED; /* knlist_remove sets it */
893 filt_timertouch(struct knote *kn, struct kevent *kev, u_long type)
895 struct kq_timer_cb_data *kc;
902 /* Handle re-added timers that update data/fflags */
903 if (kev->flags & EV_ADD) {
906 /* Drain any existing callout. */
907 callout_drain(&kc->c);
909 /* Throw away any existing undelivered record
910 * of the timer expiration. This is done under
911 * the presumption that if a process is
912 * re-adding this timer with new parameters,
913 * it is no longer interested in what may have
914 * happened under the old parameters. If it is
915 * interested, it can wait for the expiration,
916 * delete the old timer definition, and then
919 * This has to be done while the kq is locked:
920 * - if enqueued, dequeue
921 * - make it no longer active
922 * - clear the count of expiration events
926 if (kn->kn_status & KN_QUEUED)
929 kn->kn_status &= ~KN_ACTIVE;
933 /* Reschedule timer based on new data/fflags */
934 kn->kn_sfflags = kev->fflags;
935 kn->kn_sdata = kev->data;
936 error = filt_timervalidate(kn, &to);
938 kn->kn_flags |= EV_ERROR;
941 filt_timerstart(kn, to);
946 *kev = kn->kn_kevent;
947 if (kn->kn_flags & EV_CLEAR) {
954 panic("filt_timertouch() - invalid type (%ld)", type);
960 filt_timer(struct knote *kn, long hint)
963 return (kn->kn_data != 0);
967 filt_userattach(struct knote *kn)
971 * EVFILT_USER knotes are not attached to anything in the kernel.
974 if (kn->kn_fflags & NOTE_TRIGGER)
982 filt_userdetach(__unused struct knote *kn)
986 * EVFILT_USER knotes are not attached to anything in the kernel.
991 filt_user(struct knote *kn, __unused long hint)
994 return (kn->kn_hookid);
998 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
1003 case EVENT_REGISTER:
1004 if (kev->fflags & NOTE_TRIGGER)
1007 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
1008 kev->fflags &= NOTE_FFLAGSMASK;
1014 kn->kn_sfflags &= kev->fflags;
1018 kn->kn_sfflags |= kev->fflags;
1022 kn->kn_sfflags = kev->fflags;
1026 /* XXX Return error? */
1029 kn->kn_sdata = kev->data;
1030 if (kev->flags & EV_CLEAR) {
1038 *kev = kn->kn_kevent;
1039 kev->fflags = kn->kn_sfflags;
1040 kev->data = kn->kn_sdata;
1041 if (kn->kn_flags & EV_CLEAR) {
1049 panic("filt_usertouch() - invalid type (%ld)", type);
1055 sys_kqueue(struct thread *td, struct kqueue_args *uap)
1058 return (kern_kqueue(td, 0, NULL));
1062 kqueue_init(struct kqueue *kq)
1065 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
1066 TAILQ_INIT(&kq->kq_head);
1067 knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
1068 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
1072 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
1074 struct filedesc *fdp;
1080 fdp = td->td_proc->p_fd;
1081 cred = td->td_ucred;
1082 if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
1085 error = falloc_caps(td, &fp, &fd, flags, fcaps);
1087 chgkqcnt(cred->cr_ruidinfo, -1, 0);
1091 /* An extra reference on `fp' has been held for us by falloc(). */
1092 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
1095 kq->kq_cred = crhold(cred);
1097 FILEDESC_XLOCK(fdp);
1098 TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
1099 FILEDESC_XUNLOCK(fdp);
1101 finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
1104 td->td_retval[0] = fd;
1108 struct g_kevent_args {
1110 const void *changelist;
1114 const struct timespec *timeout;
1118 sys_kevent(struct thread *td, struct kevent_args *uap)
1120 struct kevent_copyops k_ops = {
1122 .k_copyout = kevent_copyout,
1123 .k_copyin = kevent_copyin,
1124 .kevent_size = sizeof(struct kevent),
1126 struct g_kevent_args gk_args = {
1128 .changelist = uap->changelist,
1129 .nchanges = uap->nchanges,
1130 .eventlist = uap->eventlist,
1131 .nevents = uap->nevents,
1132 .timeout = uap->timeout,
1135 return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent"));
1139 kern_kevent_generic(struct thread *td, struct g_kevent_args *uap,
1140 struct kevent_copyops *k_ops, const char *struct_name)
1142 struct timespec ts, *tsp;
1144 struct kevent *eventlist = uap->eventlist;
1148 if (uap->timeout != NULL) {
1149 error = copyin(uap->timeout, &ts, sizeof(ts));
1157 if (KTRPOINT(td, KTR_STRUCT_ARRAY))
1158 ktrstructarray(struct_name, UIO_USERSPACE, uap->changelist,
1159 uap->nchanges, k_ops->kevent_size);
1162 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
1166 if (error == 0 && KTRPOINT(td, KTR_STRUCT_ARRAY))
1167 ktrstructarray(struct_name, UIO_USERSPACE, eventlist,
1168 td->td_retval[0], k_ops->kevent_size);
1175 * Copy 'count' items into the destination list pointed to by uap->eventlist.
1178 kevent_copyout(void *arg, struct kevent *kevp, int count)
1180 struct kevent_args *uap;
1183 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1184 uap = (struct kevent_args *)arg;
1186 error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
1188 uap->eventlist += count;
1193 * Copy 'count' items from the list pointed to by uap->changelist.
1196 kevent_copyin(void *arg, struct kevent *kevp, int count)
1198 struct kevent_args *uap;
1201 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1202 uap = (struct kevent_args *)arg;
1204 error = copyin(uap->changelist, kevp, count * sizeof *kevp);
1206 uap->changelist += count;
1210 #ifdef COMPAT_FREEBSD11
1212 kevent11_copyout(void *arg, struct kevent *kevp, int count)
1214 struct freebsd11_kevent_args *uap;
1215 struct freebsd11_kevent kev11;
1218 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1219 uap = (struct freebsd11_kevent_args *)arg;
1221 for (i = 0; i < count; i++) {
1222 kev11.ident = kevp->ident;
1223 kev11.filter = kevp->filter;
1224 kev11.flags = kevp->flags;
1225 kev11.fflags = kevp->fflags;
1226 kev11.data = kevp->data;
1227 kev11.udata = kevp->udata;
1228 error = copyout(&kev11, uap->eventlist, sizeof(kev11));
1238 * Copy 'count' items from the list pointed to by uap->changelist.
1241 kevent11_copyin(void *arg, struct kevent *kevp, int count)
1243 struct freebsd11_kevent_args *uap;
1244 struct freebsd11_kevent kev11;
1247 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1248 uap = (struct freebsd11_kevent_args *)arg;
1250 for (i = 0; i < count; i++) {
1251 error = copyin(uap->changelist, &kev11, sizeof(kev11));
1254 kevp->ident = kev11.ident;
1255 kevp->filter = kev11.filter;
1256 kevp->flags = kev11.flags;
1257 kevp->fflags = kev11.fflags;
1258 kevp->data = (uintptr_t)kev11.data;
1259 kevp->udata = kev11.udata;
1260 bzero(&kevp->ext, sizeof(kevp->ext));
1268 freebsd11_kevent(struct thread *td, struct freebsd11_kevent_args *uap)
1270 struct kevent_copyops k_ops = {
1272 .k_copyout = kevent11_copyout,
1273 .k_copyin = kevent11_copyin,
1274 .kevent_size = sizeof(struct freebsd11_kevent),
1276 struct g_kevent_args gk_args = {
1278 .changelist = uap->changelist,
1279 .nchanges = uap->nchanges,
1280 .eventlist = uap->eventlist,
1281 .nevents = uap->nevents,
1282 .timeout = uap->timeout,
1285 return (kern_kevent_generic(td, &gk_args, &k_ops, "freebsd11_kevent"));
1290 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
1291 struct kevent_copyops *k_ops, const struct timespec *timeout)
1293 cap_rights_t rights;
1297 cap_rights_init_zero(&rights);
1299 cap_rights_set_one(&rights, CAP_KQUEUE_CHANGE);
1301 cap_rights_set_one(&rights, CAP_KQUEUE_EVENT);
1302 error = fget(td, fd, &rights, &fp);
1306 error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
1313 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
1314 struct kevent_copyops *k_ops, const struct timespec *timeout)
1316 struct kevent keva[KQ_NEVENTS];
1317 struct kevent *kevp, *changes;
1318 int i, n, nerrors, error;
1324 while (nchanges > 0) {
1325 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
1326 error = k_ops->k_copyin(k_ops->arg, keva, n);
1330 for (i = 0; i < n; i++) {
1334 kevp->flags &= ~EV_SYSFLAGS;
1335 error = kqueue_register(kq, kevp, td, M_WAITOK);
1336 if (error || (kevp->flags & EV_RECEIPT)) {
1339 kevp->flags = EV_ERROR;
1341 (void)k_ops->k_copyout(k_ops->arg, kevp, 1);
1349 td->td_retval[0] = nerrors;
1353 return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
1357 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
1358 struct kevent_copyops *k_ops, const struct timespec *timeout)
1363 error = kqueue_acquire(fp, &kq);
1366 error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
1367 kqueue_release(kq, 0);
1372 * Performs a kevent() call on a temporarily created kqueue. This can be
1373 * used to perform one-shot polling, similar to poll() and select().
1376 kern_kevent_anonymous(struct thread *td, int nevents,
1377 struct kevent_copyops *k_ops)
1379 struct kqueue kq = {};
1384 error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
1385 kqueue_drain(&kq, td);
1386 kqueue_destroy(&kq);
1391 kqueue_add_filteropts(int filt, struct filterops *filtops)
1396 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1398 "trying to add a filterop that is out of range: %d is beyond %d\n",
1399 ~filt, EVFILT_SYSCOUNT);
1402 mtx_lock(&filterops_lock);
1403 if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1404 sysfilt_ops[~filt].for_fop != NULL)
1407 sysfilt_ops[~filt].for_fop = filtops;
1408 sysfilt_ops[~filt].for_refcnt = 0;
1410 mtx_unlock(&filterops_lock);
1416 kqueue_del_filteropts(int filt)
1421 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1424 mtx_lock(&filterops_lock);
1425 if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1426 sysfilt_ops[~filt].for_fop == NULL)
1428 else if (sysfilt_ops[~filt].for_refcnt != 0)
1431 sysfilt_ops[~filt].for_fop = &null_filtops;
1432 sysfilt_ops[~filt].for_refcnt = 0;
1434 mtx_unlock(&filterops_lock);
1439 static struct filterops *
1440 kqueue_fo_find(int filt)
1443 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1446 if (sysfilt_ops[~filt].for_nolock)
1447 return sysfilt_ops[~filt].for_fop;
1449 mtx_lock(&filterops_lock);
1450 sysfilt_ops[~filt].for_refcnt++;
1451 if (sysfilt_ops[~filt].for_fop == NULL)
1452 sysfilt_ops[~filt].for_fop = &null_filtops;
1453 mtx_unlock(&filterops_lock);
1455 return sysfilt_ops[~filt].for_fop;
1459 kqueue_fo_release(int filt)
1462 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1465 if (sysfilt_ops[~filt].for_nolock)
1468 mtx_lock(&filterops_lock);
1469 KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1470 ("filter object refcount not valid on release"));
1471 sysfilt_ops[~filt].for_refcnt--;
1472 mtx_unlock(&filterops_lock);
1476 * A ref to kq (obtained via kqueue_acquire) must be held.
1479 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td,
1482 struct filterops *fops;
1484 struct knote *kn, *tkn;
1486 int error, filt, event;
1487 int haskqglobal, filedesc_unlock;
1489 if ((kev->flags & (EV_ENABLE | EV_DISABLE)) == (EV_ENABLE | EV_DISABLE))
1497 filedesc_unlock = 0;
1500 fops = kqueue_fo_find(filt);
1504 if (kev->flags & EV_ADD) {
1505 /* Reject an invalid flag pair early */
1506 if (kev->flags & EV_KEEPUDATA) {
1513 * Prevent waiting with locks. Non-sleepable
1514 * allocation failures are handled in the loop, only
1515 * if the spare knote appears to be actually required.
1517 tkn = knote_alloc(mflag);
1524 KASSERT(td != NULL, ("td is NULL"));
1525 if (kev->ident > INT_MAX)
1528 error = fget(td, kev->ident, &cap_event_rights, &fp);
1532 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1533 kev->ident, M_NOWAIT) != 0) {
1537 error = kqueue_expand(kq, fops, kev->ident, mflag);
1543 if (fp->f_type == DTYPE_KQUEUE) {
1545 * If we add some intelligence about what we are doing,
1546 * we should be able to support events on ourselves.
1547 * We need to know when we are doing this to prevent
1548 * getting both the knlist lock and the kq lock since
1549 * they are the same thing.
1551 if (fp->f_data == kq) {
1557 * Pre-lock the filedesc before the global
1558 * lock mutex, see the comment in
1561 FILEDESC_XLOCK(td->td_proc->p_fd);
1562 filedesc_unlock = 1;
1563 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1567 if (kev->ident < kq->kq_knlistsize) {
1568 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1569 if (kev->filter == kn->kn_filter)
1573 if ((kev->flags & EV_ADD) == EV_ADD) {
1574 error = kqueue_expand(kq, fops, kev->ident, mflag);
1582 * If possible, find an existing knote to use for this kevent.
1584 if (kev->filter == EVFILT_PROC &&
1585 (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1586 /* This is an internal creation of a process tracking
1587 * note. Don't attempt to coalesce this with an
1591 } else if (kq->kq_knhashmask != 0) {
1594 list = &kq->kq_knhash[
1595 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1596 SLIST_FOREACH(kn, list, kn_link)
1597 if (kev->ident == kn->kn_id &&
1598 kev->filter == kn->kn_filter)
1603 /* knote is in the process of changing, wait for it to stabilize. */
1604 if (kn != NULL && kn_in_flux(kn)) {
1605 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1606 if (filedesc_unlock) {
1607 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1608 filedesc_unlock = 0;
1610 kq->kq_state |= KQ_FLUXWAIT;
1611 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1620 * kn now contains the matching knote, or NULL if no match
1623 if (kev->flags & EV_ADD) {
1635 * apply reference counts to knote structure, and
1636 * do not release it at the end of this routine.
1641 kn->kn_sfflags = kev->fflags;
1642 kn->kn_sdata = kev->data;
1645 kn->kn_kevent = *kev;
1646 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1647 EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1648 kn->kn_status = KN_DETACHED;
1649 if ((kev->flags & EV_DISABLE) != 0)
1650 kn->kn_status |= KN_DISABLED;
1653 error = knote_attach(kn, kq);
1660 if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1661 knote_drop_detached(kn, td);
1664 knl = kn_list_lock(kn);
1667 /* No matching knote and the EV_ADD flag is not set. */
1674 if (kev->flags & EV_DELETE) {
1681 if (kev->flags & EV_FORCEONESHOT) {
1682 kn->kn_flags |= EV_ONESHOT;
1683 KNOTE_ACTIVATE(kn, 1);
1686 if ((kev->flags & EV_ENABLE) != 0)
1687 kn->kn_status &= ~KN_DISABLED;
1688 else if ((kev->flags & EV_DISABLE) != 0)
1689 kn->kn_status |= KN_DISABLED;
1692 * The user may change some filter values after the initial EV_ADD,
1693 * but doing so will not reset any filter which has already been
1696 kn->kn_status |= KN_SCAN;
1699 knl = kn_list_lock(kn);
1700 if ((kev->flags & EV_KEEPUDATA) == 0)
1701 kn->kn_kevent.udata = kev->udata;
1702 if (!fops->f_isfd && fops->f_touch != NULL) {
1703 fops->f_touch(kn, kev, EVENT_REGISTER);
1705 kn->kn_sfflags = kev->fflags;
1706 kn->kn_sdata = kev->data;
1711 * We can get here with kn->kn_knlist == NULL. This can happen when
1712 * the initial attach event decides that the event is "completed"
1713 * already, e.g., filt_procattach() is called on a zombie process. It
1714 * will call filt_proc() which will remove it from the list, and NULL
1717 * KN_DISABLED will be stable while the knote is in flux, so the
1718 * unlocked read will not race with an update.
1720 if ((kn->kn_status & KN_DISABLED) == 0)
1721 event = kn->kn_fop->f_event(kn, 0);
1727 kn->kn_status |= KN_ACTIVE;
1728 if ((kn->kn_status & (KN_ACTIVE | KN_DISABLED | KN_QUEUED)) ==
1731 kn->kn_status &= ~KN_SCAN;
1733 kn_list_unlock(knl);
1737 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1738 if (filedesc_unlock)
1739 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1744 kqueue_fo_release(filt);
1749 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1757 if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1761 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1772 kqueue_release(struct kqueue *kq, int locked)
1779 if (kq->kq_refcnt == 1)
1780 wakeup(&kq->kq_refcnt);
1786 kqueue_drain_schedtask(void)
1788 taskqueue_quiesce(taskqueue_kqueue_ctx);
1792 kqueue_schedtask(struct kqueue *kq)
1797 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1798 ("scheduling kqueue task while draining"));
1800 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1801 taskqueue_enqueue(taskqueue_kqueue_ctx, &kq->kq_task);
1802 kq->kq_state |= KQ_TASKSCHED;
1805 td->td_flags |= TDF_ASTPENDING | TDF_KQTICKLED;
1811 * Expand the kq to make sure we have storage for fops/ident pair.
1813 * Return 0 on success (or no work necessary), return errno on failure.
1816 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1819 struct klist *list, *tmp_knhash, *to_free;
1820 u_long tmp_knhashmask;
1821 int error, fd, size;
1829 if (kq->kq_knlistsize <= fd) {
1830 size = kq->kq_knlistsize;
1833 list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1837 if ((kq->kq_state & KQ_CLOSING) != 0) {
1840 } else if (kq->kq_knlistsize > fd) {
1843 if (kq->kq_knlist != NULL) {
1844 bcopy(kq->kq_knlist, list,
1845 kq->kq_knlistsize * sizeof(*list));
1846 to_free = kq->kq_knlist;
1847 kq->kq_knlist = NULL;
1849 bzero((caddr_t)list +
1850 kq->kq_knlistsize * sizeof(*list),
1851 (size - kq->kq_knlistsize) * sizeof(*list));
1852 kq->kq_knlistsize = size;
1853 kq->kq_knlist = list;
1858 if (kq->kq_knhashmask == 0) {
1859 tmp_knhash = hashinit_flags(KN_HASHSIZE, M_KQUEUE,
1860 &tmp_knhashmask, (mflag & M_WAITOK) != 0 ?
1861 HASH_WAITOK : HASH_NOWAIT);
1862 if (tmp_knhash == NULL)
1865 if ((kq->kq_state & KQ_CLOSING) != 0) {
1866 to_free = tmp_knhash;
1868 } else if (kq->kq_knhashmask == 0) {
1869 kq->kq_knhash = tmp_knhash;
1870 kq->kq_knhashmask = tmp_knhashmask;
1872 to_free = tmp_knhash;
1877 free(to_free, M_KQUEUE);
1884 kqueue_task(void *arg, int pending)
1892 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1895 KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1897 kq->kq_state &= ~KQ_TASKSCHED;
1898 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1899 wakeup(&kq->kq_state);
1902 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1906 * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1907 * We treat KN_MARKER knotes as if they are in flux.
1910 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1911 const struct timespec *tsp, struct kevent *keva, struct thread *td)
1913 struct kevent *kevp;
1914 struct knote *kn, *marker;
1916 sbintime_t asbt, rsbt;
1917 int count, error, haskqglobal, influx, nkev, touch;
1926 if (maxevents < 0) {
1933 if (!timespecvalid_interval(tsp)) {
1937 if (timespecisset(tsp)) {
1938 if (tsp->tv_sec <= INT32_MAX) {
1939 rsbt = tstosbt(*tsp);
1940 if (TIMESEL(&asbt, rsbt))
1941 asbt += tc_tick_sbt;
1942 if (asbt <= SBT_MAX - rsbt)
1946 rsbt >>= tc_precexp;
1953 marker = knote_alloc(M_WAITOK);
1954 marker->kn_status = KN_MARKER;
1959 if (kq->kq_count == 0) {
1961 error = EWOULDBLOCK;
1963 kq->kq_state |= KQ_SLEEP;
1964 error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1965 "kqread", asbt, rsbt, C_ABSOLUTE);
1969 /* don't restart after signals... */
1970 if (error == ERESTART)
1972 else if (error == EWOULDBLOCK)
1977 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1981 kn = TAILQ_FIRST(&kq->kq_head);
1983 if ((kn->kn_status == KN_MARKER && kn != marker) ||
1989 kq->kq_state |= KQ_FLUXWAIT;
1990 error = msleep(kq, &kq->kq_lock, PSOCK,
1995 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1996 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1997 kn->kn_status &= ~KN_QUEUED;
2003 if (count == maxevents)
2007 KASSERT(!kn_in_flux(kn),
2008 ("knote %p is unexpectedly in flux", kn));
2010 if ((kn->kn_flags & EV_DROP) == EV_DROP) {
2011 kn->kn_status &= ~KN_QUEUED;
2016 * We don't need to lock the list since we've
2017 * marked it as in flux.
2022 } else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
2023 kn->kn_status &= ~KN_QUEUED;
2028 * We don't need to lock the list since we've
2029 * marked the knote as being in flux.
2031 *kevp = kn->kn_kevent;
2036 kn->kn_status |= KN_SCAN;
2039 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
2040 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
2041 knl = kn_list_lock(kn);
2042 if (kn->kn_fop->f_event(kn, 0) == 0) {
2044 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
2045 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE |
2049 kn_list_unlock(knl);
2053 touch = (!kn->kn_fop->f_isfd &&
2054 kn->kn_fop->f_touch != NULL);
2056 kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
2058 *kevp = kn->kn_kevent;
2060 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
2061 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
2063 * Manually clear knotes who weren't
2066 if (touch == 0 && kn->kn_flags & EV_CLEAR) {
2070 if (kn->kn_flags & EV_DISPATCH)
2071 kn->kn_status |= KN_DISABLED;
2072 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
2075 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2077 kn->kn_status &= ~KN_SCAN;
2079 kn_list_unlock(knl);
2083 /* we are returning a copy to the user */
2088 if (nkev == KQ_NEVENTS) {
2091 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
2099 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
2107 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
2108 td->td_retval[0] = maxevents - count;
2114 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
2115 struct ucred *active_cred, struct thread *td)
2118 * Enabling sigio causes two major problems:
2119 * 1) infinite recursion:
2120 * Synopsys: kevent is being used to track signals and have FIOASYNC
2121 * set. On receipt of a signal this will cause a kqueue to recurse
2122 * into itself over and over. Sending the sigio causes the kqueue
2123 * to become ready, which in turn posts sigio again, forever.
2124 * Solution: this can be solved by setting a flag in the kqueue that
2125 * we have a SIGIO in progress.
2126 * 2) locking problems:
2127 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
2128 * us above the proc and pgrp locks.
2129 * Solution: Post a signal using an async mechanism, being sure to
2130 * record a generation count in the delivery so that we do not deliver
2131 * a signal to the wrong process.
2133 * Note, these two mechanisms are somewhat mutually exclusive!
2142 kq->kq_state |= KQ_ASYNC;
2144 kq->kq_state &= ~KQ_ASYNC;
2149 return (fsetown(*(int *)data, &kq->kq_sigio));
2152 *(int *)data = fgetown(&kq->kq_sigio);
2162 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
2169 if ((error = kqueue_acquire(fp, &kq)))
2173 if (events & (POLLIN | POLLRDNORM)) {
2175 revents |= events & (POLLIN | POLLRDNORM);
2177 selrecord(td, &kq->kq_sel);
2178 if (SEL_WAITING(&kq->kq_sel))
2179 kq->kq_state |= KQ_SEL;
2182 kqueue_release(kq, 1);
2189 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred)
2192 bzero((void *)st, sizeof *st);
2194 * We no longer return kq_count because the unlocked value is useless.
2195 * If you spent all this time getting the count, why not spend your
2196 * syscall better by calling kevent?
2198 * XXX - This is needed for libc_r.
2200 st->st_mode = S_IFIFO;
2205 kqueue_drain(struct kqueue *kq, struct thread *td)
2212 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
2213 ("kqueue already closing"));
2214 kq->kq_state |= KQ_CLOSING;
2215 if (kq->kq_refcnt > 1)
2216 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
2218 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
2220 KASSERT(knlist_empty(&kq->kq_sel.si_note),
2221 ("kqueue's knlist not empty"));
2223 for (i = 0; i < kq->kq_knlistsize; i++) {
2224 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
2225 if (kn_in_flux(kn)) {
2226 kq->kq_state |= KQ_FLUXWAIT;
2227 msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
2236 if (kq->kq_knhashmask != 0) {
2237 for (i = 0; i <= kq->kq_knhashmask; i++) {
2238 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
2239 if (kn_in_flux(kn)) {
2240 kq->kq_state |= KQ_FLUXWAIT;
2241 msleep(kq, &kq->kq_lock, PSOCK,
2253 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
2254 kq->kq_state |= KQ_TASKDRAIN;
2255 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
2258 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2259 selwakeuppri(&kq->kq_sel, PSOCK);
2260 if (!SEL_WAITING(&kq->kq_sel))
2261 kq->kq_state &= ~KQ_SEL;
2268 kqueue_destroy(struct kqueue *kq)
2271 KASSERT(kq->kq_fdp == NULL,
2272 ("kqueue still attached to a file descriptor"));
2273 seldrain(&kq->kq_sel);
2274 knlist_destroy(&kq->kq_sel.si_note);
2275 mtx_destroy(&kq->kq_lock);
2277 if (kq->kq_knhash != NULL)
2278 free(kq->kq_knhash, M_KQUEUE);
2279 if (kq->kq_knlist != NULL)
2280 free(kq->kq_knlist, M_KQUEUE);
2282 funsetown(&kq->kq_sigio);
2287 kqueue_close(struct file *fp, struct thread *td)
2289 struct kqueue *kq = fp->f_data;
2290 struct filedesc *fdp;
2292 int filedesc_unlock;
2294 if ((error = kqueue_acquire(fp, &kq)))
2296 kqueue_drain(kq, td);
2299 * We could be called due to the knote_drop() doing fdrop(),
2300 * called from kqueue_register(). In this case the global
2301 * lock is owned, and filedesc sx is locked before, to not
2302 * take the sleepable lock after non-sleepable.
2306 if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
2307 FILEDESC_XLOCK(fdp);
2308 filedesc_unlock = 1;
2310 filedesc_unlock = 0;
2311 TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
2312 if (filedesc_unlock)
2313 FILEDESC_XUNLOCK(fdp);
2316 chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
2317 crfree(kq->kq_cred);
2325 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2328 kif->kf_type = KF_TYPE_KQUEUE;
2333 kqueue_wakeup(struct kqueue *kq)
2337 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
2338 kq->kq_state &= ~KQ_SLEEP;
2341 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2342 selwakeuppri(&kq->kq_sel, PSOCK);
2343 if (!SEL_WAITING(&kq->kq_sel))
2344 kq->kq_state &= ~KQ_SEL;
2346 if (!knlist_empty(&kq->kq_sel.si_note))
2347 kqueue_schedtask(kq);
2348 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
2349 pgsigio(&kq->kq_sigio, SIGIO, 0);
2354 * Walk down a list of knotes, activating them if their event has triggered.
2356 * There is a possibility to optimize in the case of one kq watching another.
2357 * Instead of scheduling a task to wake it up, you could pass enough state
2358 * down the chain to make up the parent kqueue. Make this code functional
2362 knote(struct knlist *list, long hint, int lockflags)
2365 struct knote *kn, *tkn;
2371 KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
2373 if ((lockflags & KNF_LISTLOCKED) == 0)
2374 list->kl_lock(list->kl_lockarg);
2377 * If we unlock the list lock (and enter influx), we can
2378 * eliminate the kqueue scheduling, but this will introduce
2379 * four lock/unlock's for each knote to test. Also, marker
2380 * would be needed to keep iteration position, since filters
2381 * or other threads could remove events.
2383 SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
2386 if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
2388 * Do not process the influx notes, except for
2389 * the influx coming from the kq unlock in the
2390 * kqueue_scan(). In the later case, we do
2391 * not interfere with the scan, since the code
2392 * fragment in kqueue_scan() locks the knlist,
2393 * and cannot proceed until we finished.
2396 } else if ((lockflags & KNF_NOKQLOCK) != 0) {
2399 error = kn->kn_fop->f_event(kn, hint);
2403 KNOTE_ACTIVATE(kn, 1);
2406 if (kn->kn_fop->f_event(kn, hint))
2407 KNOTE_ACTIVATE(kn, 1);
2411 if ((lockflags & KNF_LISTLOCKED) == 0)
2412 list->kl_unlock(list->kl_lockarg);
2416 * add a knote to a knlist
2419 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2422 KNL_ASSERT_LOCK(knl, islocked);
2423 KQ_NOTOWNED(kn->kn_kq);
2424 KASSERT(kn_in_flux(kn), ("knote %p not in flux", kn));
2425 KASSERT((kn->kn_status & KN_DETACHED) != 0,
2426 ("knote %p was not detached", kn));
2428 knl->kl_lock(knl->kl_lockarg);
2429 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2431 knl->kl_unlock(knl->kl_lockarg);
2433 kn->kn_knlist = knl;
2434 kn->kn_status &= ~KN_DETACHED;
2435 KQ_UNLOCK(kn->kn_kq);
2439 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked,
2443 KASSERT(!kqislocked || knlislocked, ("kq locked w/o knl locked"));
2444 KNL_ASSERT_LOCK(knl, knlislocked);
2445 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2446 KASSERT(kqislocked || kn_in_flux(kn), ("knote %p not in flux", kn));
2447 KASSERT((kn->kn_status & KN_DETACHED) == 0,
2448 ("knote %p was already detached", kn));
2450 knl->kl_lock(knl->kl_lockarg);
2451 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2452 kn->kn_knlist = NULL;
2454 kn_list_unlock(knl);
2457 kn->kn_status |= KN_DETACHED;
2459 KQ_UNLOCK(kn->kn_kq);
2463 * remove knote from the specified knlist
2466 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2469 knlist_remove_kq(knl, kn, islocked, 0);
2473 knlist_empty(struct knlist *knl)
2476 KNL_ASSERT_LOCKED(knl);
2477 return (SLIST_EMPTY(&knl->kl_list));
2480 static struct mtx knlist_lock;
2481 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2483 static void knlist_mtx_lock(void *arg);
2484 static void knlist_mtx_unlock(void *arg);
2487 knlist_mtx_lock(void *arg)
2490 mtx_lock((struct mtx *)arg);
2494 knlist_mtx_unlock(void *arg)
2497 mtx_unlock((struct mtx *)arg);
2501 knlist_mtx_assert_lock(void *arg, int what)
2504 if (what == LA_LOCKED)
2505 mtx_assert((struct mtx *)arg, MA_OWNED);
2507 mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2511 knlist_rw_rlock(void *arg)
2514 rw_rlock((struct rwlock *)arg);
2518 knlist_rw_runlock(void *arg)
2521 rw_runlock((struct rwlock *)arg);
2525 knlist_rw_assert_lock(void *arg, int what)
2528 if (what == LA_LOCKED)
2529 rw_assert((struct rwlock *)arg, RA_LOCKED);
2531 rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2535 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2536 void (*kl_unlock)(void *),
2537 void (*kl_assert_lock)(void *, int))
2541 knl->kl_lockarg = &knlist_lock;
2543 knl->kl_lockarg = lock;
2545 if (kl_lock == NULL)
2546 knl->kl_lock = knlist_mtx_lock;
2548 knl->kl_lock = kl_lock;
2549 if (kl_unlock == NULL)
2550 knl->kl_unlock = knlist_mtx_unlock;
2552 knl->kl_unlock = kl_unlock;
2553 if (kl_assert_lock == NULL)
2554 knl->kl_assert_lock = knlist_mtx_assert_lock;
2556 knl->kl_assert_lock = kl_assert_lock;
2558 knl->kl_autodestroy = 0;
2559 SLIST_INIT(&knl->kl_list);
2563 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2566 knlist_init(knl, lock, NULL, NULL, NULL);
2570 knlist_alloc(struct mtx *lock)
2574 knl = malloc(sizeof(struct knlist), M_KQUEUE, M_WAITOK);
2575 knlist_init_mtx(knl, lock);
2580 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2583 knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2584 knlist_rw_assert_lock);
2588 knlist_destroy(struct knlist *knl)
2591 KASSERT(KNLIST_EMPTY(knl),
2592 ("destroying knlist %p with knotes on it", knl));
2596 knlist_detach(struct knlist *knl)
2599 KNL_ASSERT_LOCKED(knl);
2600 knl->kl_autodestroy = 1;
2601 if (knlist_empty(knl)) {
2602 knlist_destroy(knl);
2603 free(knl, M_KQUEUE);
2608 * Even if we are locked, we may need to drop the lock to allow any influx
2609 * knotes time to "settle".
2612 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2614 struct knote *kn, *kn2;
2617 KASSERT(!knl->kl_autodestroy, ("cleardel for autodestroy %p", knl));
2619 KNL_ASSERT_LOCKED(knl);
2621 KNL_ASSERT_UNLOCKED(knl);
2622 again: /* need to reacquire lock since we have dropped it */
2623 knl->kl_lock(knl->kl_lockarg);
2626 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2629 if (kn_in_flux(kn)) {
2633 knlist_remove_kq(knl, kn, 1, 1);
2637 knote_drop_detached(kn, td);
2639 /* Make sure cleared knotes disappear soon */
2640 kn->kn_flags |= EV_EOF | EV_ONESHOT;
2646 if (!SLIST_EMPTY(&knl->kl_list)) {
2647 /* there are still in flux knotes remaining */
2648 kn = SLIST_FIRST(&knl->kl_list);
2651 KASSERT(kn_in_flux(kn), ("knote removed w/o list lock"));
2652 knl->kl_unlock(knl->kl_lockarg);
2653 kq->kq_state |= KQ_FLUXWAIT;
2654 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2660 KNL_ASSERT_LOCKED(knl);
2662 knl->kl_unlock(knl->kl_lockarg);
2663 KNL_ASSERT_UNLOCKED(knl);
2668 * Remove all knotes referencing a specified fd must be called with FILEDESC
2669 * lock. This prevents a race where a new fd comes along and occupies the
2670 * entry and we attach a knote to the fd.
2673 knote_fdclose(struct thread *td, int fd)
2675 struct filedesc *fdp = td->td_proc->p_fd;
2680 FILEDESC_XLOCK_ASSERT(fdp);
2683 * We shouldn't have to worry about new kevents appearing on fd
2684 * since filedesc is locked.
2686 TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2691 while (kq->kq_knlistsize > fd &&
2692 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2693 if (kn_in_flux(kn)) {
2694 /* someone else might be waiting on our knote */
2697 kq->kq_state |= KQ_FLUXWAIT;
2698 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2712 knote_attach(struct knote *kn, struct kqueue *kq)
2716 KASSERT(kn_in_flux(kn), ("knote %p not marked influx", kn));
2719 if ((kq->kq_state & KQ_CLOSING) != 0)
2721 if (kn->kn_fop->f_isfd) {
2722 if (kn->kn_id >= kq->kq_knlistsize)
2724 list = &kq->kq_knlist[kn->kn_id];
2726 if (kq->kq_knhash == NULL)
2728 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2730 SLIST_INSERT_HEAD(list, kn, kn_link);
2735 knote_drop(struct knote *kn, struct thread *td)
2738 if ((kn->kn_status & KN_DETACHED) == 0)
2739 kn->kn_fop->f_detach(kn);
2740 knote_drop_detached(kn, td);
2744 knote_drop_detached(struct knote *kn, struct thread *td)
2751 KASSERT((kn->kn_status & KN_DETACHED) != 0,
2752 ("knote %p still attached", kn));
2756 KASSERT(kn->kn_influx == 1,
2757 ("knote_drop called on %p with influx %d", kn, kn->kn_influx));
2759 if (kn->kn_fop->f_isfd)
2760 list = &kq->kq_knlist[kn->kn_id];
2762 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2764 if (!SLIST_EMPTY(list))
2765 SLIST_REMOVE(list, kn, knote, kn_link);
2766 if (kn->kn_status & KN_QUEUED)
2770 if (kn->kn_fop->f_isfd) {
2771 fdrop(kn->kn_fp, td);
2774 kqueue_fo_release(kn->kn_kevent.filter);
2780 knote_enqueue(struct knote *kn)
2782 struct kqueue *kq = kn->kn_kq;
2784 KQ_OWNED(kn->kn_kq);
2785 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2787 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2788 kn->kn_status |= KN_QUEUED;
2794 knote_dequeue(struct knote *kn)
2796 struct kqueue *kq = kn->kn_kq;
2798 KQ_OWNED(kn->kn_kq);
2799 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2801 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2802 kn->kn_status &= ~KN_QUEUED;
2810 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2811 NULL, NULL, UMA_ALIGN_PTR, 0);
2813 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2815 static struct knote *
2816 knote_alloc(int mflag)
2819 return (uma_zalloc(knote_zone, mflag | M_ZERO));
2823 knote_free(struct knote *kn)
2826 uma_zfree(knote_zone, kn);
2830 * Register the kev w/ the kq specified by fd.
2833 kqfd_register(int fd, struct kevent *kev, struct thread *td, int mflag)
2837 cap_rights_t rights;
2840 error = fget(td, fd, cap_rights_init_one(&rights, CAP_KQUEUE_CHANGE),
2844 if ((error = kqueue_acquire(fp, &kq)) != 0)
2847 error = kqueue_register(kq, kev, td, mflag);
2848 kqueue_release(kq, 0);