2 * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
3 * Copyright 2004 John-Mark Gurney <jmg@FreeBSD.org>
4 * Copyright (c) 2009 Apple, Inc.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
32 #include "opt_ktrace.h"
33 #include "opt_kqueue.h"
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/capsicum.h>
38 #include <sys/kernel.h>
40 #include <sys/mutex.h>
41 #include <sys/rwlock.h>
43 #include <sys/malloc.h>
44 #include <sys/unistd.h>
46 #include <sys/filedesc.h>
47 #include <sys/filio.h>
48 #include <sys/fcntl.h>
49 #include <sys/kthread.h>
50 #include <sys/selinfo.h>
51 #include <sys/queue.h>
52 #include <sys/event.h>
53 #include <sys/eventvar.h>
55 #include <sys/protosw.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sigio.h>
58 #include <sys/signalvar.h>
59 #include <sys/socket.h>
60 #include <sys/socketvar.h>
62 #include <sys/sysctl.h>
63 #include <sys/sysproto.h>
64 #include <sys/syscallsubr.h>
65 #include <sys/taskqueue.h>
69 #include <sys/ktrace.h>
71 #include <machine/atomic.h>
75 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
78 * This lock is used if multiple kq locks are required. This possibly
79 * should be made into a per proc lock.
81 static struct mtx kq_global;
82 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
83 #define KQ_GLOBAL_LOCK(lck, haslck) do { \
88 #define KQ_GLOBAL_UNLOCK(lck, haslck) do { \
94 TASKQUEUE_DEFINE_THREAD(kqueue_ctx);
96 static int kevent_copyout(void *arg, struct kevent *kevp, int count);
97 static int kevent_copyin(void *arg, struct kevent *kevp, int count);
98 static int kqueue_register(struct kqueue *kq, struct kevent *kev,
99 struct thread *td, int waitok);
100 static int kqueue_acquire(struct file *fp, struct kqueue **kqp);
101 static void kqueue_release(struct kqueue *kq, int locked);
102 static void kqueue_destroy(struct kqueue *kq);
103 static void kqueue_drain(struct kqueue *kq, struct thread *td);
104 static int kqueue_expand(struct kqueue *kq, struct filterops *fops,
105 uintptr_t ident, int waitok);
106 static void kqueue_task(void *arg, int pending);
107 static int kqueue_scan(struct kqueue *kq, int maxevents,
108 struct kevent_copyops *k_ops,
109 const struct timespec *timeout,
110 struct kevent *keva, struct thread *td);
111 static void kqueue_wakeup(struct kqueue *kq);
112 static struct filterops *kqueue_fo_find(int filt);
113 static void kqueue_fo_release(int filt);
115 static fo_ioctl_t kqueue_ioctl;
116 static fo_poll_t kqueue_poll;
117 static fo_kqfilter_t kqueue_kqfilter;
118 static fo_stat_t kqueue_stat;
119 static fo_close_t kqueue_close;
120 static fo_fill_kinfo_t kqueue_fill_kinfo;
122 static struct fileops kqueueops = {
123 .fo_read = invfo_rdwr,
124 .fo_write = invfo_rdwr,
125 .fo_truncate = invfo_truncate,
126 .fo_ioctl = kqueue_ioctl,
127 .fo_poll = kqueue_poll,
128 .fo_kqfilter = kqueue_kqfilter,
129 .fo_stat = kqueue_stat,
130 .fo_close = kqueue_close,
131 .fo_chmod = invfo_chmod,
132 .fo_chown = invfo_chown,
133 .fo_sendfile = invfo_sendfile,
134 .fo_fill_kinfo = kqueue_fill_kinfo,
137 static int knote_attach(struct knote *kn, struct kqueue *kq);
138 static void knote_drop(struct knote *kn, struct thread *td);
139 static void knote_enqueue(struct knote *kn);
140 static void knote_dequeue(struct knote *kn);
141 static void knote_init(void);
142 static struct knote *knote_alloc(int waitok);
143 static void knote_free(struct knote *kn);
145 static void filt_kqdetach(struct knote *kn);
146 static int filt_kqueue(struct knote *kn, long hint);
147 static int filt_procattach(struct knote *kn);
148 static void filt_procdetach(struct knote *kn);
149 static int filt_proc(struct knote *kn, long hint);
150 static int filt_fileattach(struct knote *kn);
151 static void filt_timerexpire(void *knx);
152 static int filt_timerattach(struct knote *kn);
153 static void filt_timerdetach(struct knote *kn);
154 static int filt_timer(struct knote *kn, long hint);
155 static int filt_userattach(struct knote *kn);
156 static void filt_userdetach(struct knote *kn);
157 static int filt_user(struct knote *kn, long hint);
158 static void filt_usertouch(struct knote *kn, struct kevent *kev,
161 static struct filterops file_filtops = {
163 .f_attach = filt_fileattach,
165 static struct filterops kqread_filtops = {
167 .f_detach = filt_kqdetach,
168 .f_event = filt_kqueue,
170 /* XXX - move to kern_proc.c? */
171 static struct filterops proc_filtops = {
173 .f_attach = filt_procattach,
174 .f_detach = filt_procdetach,
175 .f_event = filt_proc,
177 static struct filterops timer_filtops = {
179 .f_attach = filt_timerattach,
180 .f_detach = filt_timerdetach,
181 .f_event = filt_timer,
183 static struct filterops user_filtops = {
184 .f_attach = filt_userattach,
185 .f_detach = filt_userdetach,
186 .f_event = filt_user,
187 .f_touch = filt_usertouch,
190 static uma_zone_t knote_zone;
191 static unsigned int kq_ncallouts = 0;
192 static unsigned int kq_calloutmax = 4 * 1024;
193 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
194 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
196 /* XXX - ensure not KN_INFLUX?? */
197 #define KNOTE_ACTIVATE(kn, islock) do { \
199 mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED); \
201 KQ_LOCK((kn)->kn_kq); \
202 (kn)->kn_status |= KN_ACTIVE; \
203 if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \
204 knote_enqueue((kn)); \
206 KQ_UNLOCK((kn)->kn_kq); \
208 #define KQ_LOCK(kq) do { \
209 mtx_lock(&(kq)->kq_lock); \
211 #define KQ_FLUX_WAKEUP(kq) do { \
212 if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) { \
213 (kq)->kq_state &= ~KQ_FLUXWAIT; \
217 #define KQ_UNLOCK_FLUX(kq) do { \
218 KQ_FLUX_WAKEUP(kq); \
219 mtx_unlock(&(kq)->kq_lock); \
221 #define KQ_UNLOCK(kq) do { \
222 mtx_unlock(&(kq)->kq_lock); \
224 #define KQ_OWNED(kq) do { \
225 mtx_assert(&(kq)->kq_lock, MA_OWNED); \
227 #define KQ_NOTOWNED(kq) do { \
228 mtx_assert(&(kq)->kq_lock, MA_NOTOWNED); \
231 static struct knlist *
232 kn_list_lock(struct knote *kn)
238 knl->kl_lock(knl->kl_lockarg);
243 kn_list_unlock(struct knlist *knl)
249 do_free = knl->kl_autodestroy && knlist_empty(knl);
250 knl->kl_unlock(knl->kl_lockarg);
257 #define KNL_ASSERT_LOCK(knl, islocked) do { \
259 KNL_ASSERT_LOCKED(knl); \
261 KNL_ASSERT_UNLOCKED(knl); \
264 #define KNL_ASSERT_LOCKED(knl) do { \
265 knl->kl_assert_locked((knl)->kl_lockarg); \
267 #define KNL_ASSERT_UNLOCKED(knl) do { \
268 knl->kl_assert_unlocked((knl)->kl_lockarg); \
270 #else /* !INVARIANTS */
271 #define KNL_ASSERT_LOCKED(knl) do {} while(0)
272 #define KNL_ASSERT_UNLOCKED(knl) do {} while (0)
273 #endif /* INVARIANTS */
276 #define KN_HASHSIZE 64 /* XXX should be tunable */
279 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask))
282 filt_nullattach(struct knote *kn)
288 struct filterops null_filtops = {
290 .f_attach = filt_nullattach,
293 /* XXX - make SYSINIT to add these, and move into respective modules. */
294 extern struct filterops sig_filtops;
295 extern struct filterops fs_filtops;
298 * Table for for all system-defined filters.
300 static struct mtx filterops_lock;
301 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
304 struct filterops *for_fop;
307 } sysfilt_ops[EVFILT_SYSCOUNT] = {
308 { &file_filtops, 1 }, /* EVFILT_READ */
309 { &file_filtops, 1 }, /* EVFILT_WRITE */
310 { &null_filtops }, /* EVFILT_AIO */
311 { &file_filtops, 1 }, /* EVFILT_VNODE */
312 { &proc_filtops, 1 }, /* EVFILT_PROC */
313 { &sig_filtops, 1 }, /* EVFILT_SIGNAL */
314 { &timer_filtops, 1 }, /* EVFILT_TIMER */
315 { &file_filtops, 1 }, /* EVFILT_PROCDESC */
316 { &fs_filtops, 1 }, /* EVFILT_FS */
317 { &null_filtops }, /* EVFILT_LIO */
318 { &user_filtops, 1 }, /* EVFILT_USER */
319 { &null_filtops }, /* EVFILT_SENDFILE */
323 * Simple redirection for all cdevsw style objects to call their fo_kqfilter
327 filt_fileattach(struct knote *kn)
330 return (fo_kqfilter(kn->kn_fp, kn));
335 kqueue_kqfilter(struct file *fp, struct knote *kn)
337 struct kqueue *kq = kn->kn_fp->f_data;
339 if (kn->kn_filter != EVFILT_READ)
342 kn->kn_status |= KN_KQUEUE;
343 kn->kn_fop = &kqread_filtops;
344 knlist_add(&kq->kq_sel.si_note, kn, 0);
350 filt_kqdetach(struct knote *kn)
352 struct kqueue *kq = kn->kn_fp->f_data;
354 knlist_remove(&kq->kq_sel.si_note, kn, 0);
359 filt_kqueue(struct knote *kn, long hint)
361 struct kqueue *kq = kn->kn_fp->f_data;
363 kn->kn_data = kq->kq_count;
364 return (kn->kn_data > 0);
367 /* XXX - move to kern_proc.c? */
369 filt_procattach(struct knote *kn)
373 bool exiting, immediate;
375 exiting = immediate = false;
376 p = pfind(kn->kn_id);
377 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
378 p = zpfind(kn->kn_id);
380 } else if (p != NULL && (p->p_flag & P_WEXIT)) {
386 if ((error = p_cansee(curthread, p))) {
391 kn->kn_ptr.p_proc = p;
392 kn->kn_flags |= EV_CLEAR; /* automatically set */
395 * Internal flag indicating registration done by kernel for the
396 * purposes of getting a NOTE_CHILD notification.
398 if (kn->kn_flags & EV_FLAG2) {
399 kn->kn_flags &= ~EV_FLAG2;
400 kn->kn_data = kn->kn_sdata; /* ppid */
401 kn->kn_fflags = NOTE_CHILD;
402 kn->kn_sfflags &= ~(NOTE_EXIT | NOTE_EXEC | NOTE_FORK);
403 immediate = true; /* Force immediate activation of child note. */
406 * Internal flag indicating registration done by kernel (for other than
409 if (kn->kn_flags & EV_FLAG1) {
410 kn->kn_flags &= ~EV_FLAG1;
413 knlist_add(p->p_klist, kn, 1);
416 * Immediately activate any child notes or, in the case of a zombie
417 * target process, exit notes. The latter is necessary to handle the
418 * case where the target process, e.g. a child, dies before the kevent
421 if (immediate || (exiting && filt_proc(kn, NOTE_EXIT)))
422 KNOTE_ACTIVATE(kn, 0);
430 * The knote may be attached to a different process, which may exit,
431 * leaving nothing for the knote to be attached to. So when the process
432 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
433 * it will be deleted when read out. However, as part of the knote deletion,
434 * this routine is called, so a check is needed to avoid actually performing
435 * a detach, because the original process does not exist any more.
437 /* XXX - move to kern_proc.c? */
439 filt_procdetach(struct knote *kn)
442 knlist_remove(kn->kn_knlist, kn, 0);
443 kn->kn_ptr.p_proc = NULL;
446 /* XXX - move to kern_proc.c? */
448 filt_proc(struct knote *kn, long hint)
453 p = kn->kn_ptr.p_proc;
454 if (p == NULL) /* already activated, from attach filter */
457 /* Mask off extra data. */
458 event = (u_int)hint & NOTE_PCTRLMASK;
460 /* If the user is interested in this event, record it. */
461 if (kn->kn_sfflags & event)
462 kn->kn_fflags |= event;
464 /* Process is gone, so flag the event as finished. */
465 if (event == NOTE_EXIT) {
466 kn->kn_flags |= EV_EOF | EV_ONESHOT;
467 kn->kn_ptr.p_proc = NULL;
468 if (kn->kn_fflags & NOTE_EXIT)
469 kn->kn_data = KW_EXITCODE(p->p_xexit, p->p_xsig);
470 if (kn->kn_fflags == 0)
471 kn->kn_flags |= EV_DROP;
475 return (kn->kn_fflags != 0);
479 * Called when the process forked. It mostly does the same as the
480 * knote(), activating all knotes registered to be activated when the
481 * process forked. Additionally, for each knote attached to the
482 * parent, check whether user wants to track the new process. If so
483 * attach a new knote to it, and immediately report an event with the
487 knote_fork(struct knlist *list, int pid)
496 list->kl_lock(list->kl_lockarg);
498 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
501 if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
507 * The same as knote(), activate the event.
509 if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
510 kn->kn_status |= KN_HASKQLOCK;
511 if (kn->kn_fop->f_event(kn, NOTE_FORK))
512 KNOTE_ACTIVATE(kn, 1);
513 kn->kn_status &= ~KN_HASKQLOCK;
519 * The NOTE_TRACK case. In addition to the activation
520 * of the event, we need to register new events to
521 * track the child. Drop the locks in preparation for
522 * the call to kqueue_register().
524 kn->kn_status |= KN_INFLUX;
526 list->kl_unlock(list->kl_lockarg);
529 * Activate existing knote and register tracking knotes with
532 * First register a knote to get just the child notice. This
533 * must be a separate note from a potential NOTE_EXIT
534 * notification since both NOTE_CHILD and NOTE_EXIT are defined
535 * to use the data field (in conflicting ways).
538 kev.filter = kn->kn_filter;
539 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_ONESHOT |
541 kev.fflags = kn->kn_sfflags;
542 kev.data = kn->kn_id; /* parent */
543 kev.udata = kn->kn_kevent.udata;/* preserve udata */
544 error = kqueue_register(kq, &kev, NULL, 0);
546 kn->kn_fflags |= NOTE_TRACKERR;
549 * Then register another knote to track other potential events
550 * from the new process.
553 kev.filter = kn->kn_filter;
554 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
555 kev.fflags = kn->kn_sfflags;
556 kev.data = kn->kn_id; /* parent */
557 kev.udata = kn->kn_kevent.udata;/* preserve udata */
558 error = kqueue_register(kq, &kev, NULL, 0);
560 kn->kn_fflags |= NOTE_TRACKERR;
561 if (kn->kn_fop->f_event(kn, NOTE_FORK))
562 KNOTE_ACTIVATE(kn, 0);
564 kn->kn_status &= ~KN_INFLUX;
566 list->kl_lock(list->kl_lockarg);
568 list->kl_unlock(list->kl_lockarg);
572 * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
573 * interval timer support code.
576 #define NOTE_TIMER_PRECMASK \
577 (NOTE_SECONDS | NOTE_MSECONDS | NOTE_USECONDS | NOTE_NSECONDS)
580 timer2sbintime(intptr_t data, int flags)
585 * Macros for converting to the fractional second portion of an
586 * sbintime_t using 64bit multiplication to improve precision.
588 #define NS_TO_SBT(ns) (((ns) * (((uint64_t)1 << 63) / 500000000)) >> 32)
589 #define US_TO_SBT(us) (((us) * (((uint64_t)1 << 63) / 500000)) >> 32)
590 #define MS_TO_SBT(ms) (((ms) * (((uint64_t)1 << 63) / 500)) >> 32)
591 switch (flags & NOTE_TIMER_PRECMASK) {
594 if (data > (SBT_MAX / SBT_1S))
597 return ((sbintime_t)data << 32);
598 case NOTE_MSECONDS: /* FALLTHROUGH */
603 if (secs > (SBT_MAX / SBT_1S))
606 return (secs << 32 | MS_TO_SBT(data % 1000));
608 return (MS_TO_SBT(data));
610 if (data >= 1000000) {
611 secs = data / 1000000;
613 if (secs > (SBT_MAX / SBT_1S))
616 return (secs << 32 | US_TO_SBT(data % 1000000));
618 return (US_TO_SBT(data));
620 if (data >= 1000000000) {
621 secs = data / 1000000000;
623 if (secs > (SBT_MAX / SBT_1S))
626 return (secs << 32 | US_TO_SBT(data % 1000000000));
628 return (NS_TO_SBT(data));
635 struct kq_timer_cb_data {
637 sbintime_t next; /* next timer event fires at */
638 sbintime_t to; /* precalculated timer period */
642 filt_timerexpire(void *knx)
645 struct kq_timer_cb_data *kc;
649 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */
651 if ((kn->kn_flags & EV_ONESHOT) != 0)
656 callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
657 PCPU_GET(cpuid), C_ABSOLUTE);
661 * data contains amount of time to sleep
664 filt_timerattach(struct knote *kn)
666 struct kq_timer_cb_data *kc;
668 unsigned int ncallouts;
670 if (kn->kn_sdata < 0)
672 if (kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
674 /* Only precision unit are supported in flags so far */
675 if ((kn->kn_sfflags & ~NOTE_TIMER_PRECMASK) != 0)
678 to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
683 ncallouts = kq_ncallouts;
684 if (ncallouts >= kq_calloutmax)
686 } while (!atomic_cmpset_int(&kq_ncallouts, ncallouts, ncallouts + 1));
688 kn->kn_flags |= EV_CLEAR; /* automatically set */
689 kn->kn_status &= ~KN_DETACHED; /* knlist_add clears it */
690 kn->kn_ptr.p_v = kc = malloc(sizeof(*kc), M_KQUEUE, M_WAITOK);
691 callout_init(&kc->c, 1);
692 kc->next = to + sbinuptime();
694 callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
695 PCPU_GET(cpuid), C_ABSOLUTE);
701 filt_timerdetach(struct knote *kn)
703 struct kq_timer_cb_data *kc;
707 callout_drain(&kc->c);
709 old = atomic_fetchadd_int(&kq_ncallouts, -1);
710 KASSERT(old > 0, ("Number of callouts cannot become negative"));
711 kn->kn_status |= KN_DETACHED; /* knlist_remove sets it */
715 filt_timer(struct knote *kn, long hint)
718 return (kn->kn_data != 0);
722 filt_userattach(struct knote *kn)
726 * EVFILT_USER knotes are not attached to anything in the kernel.
729 if (kn->kn_fflags & NOTE_TRIGGER)
737 filt_userdetach(__unused struct knote *kn)
741 * EVFILT_USER knotes are not attached to anything in the kernel.
746 filt_user(struct knote *kn, __unused long hint)
749 return (kn->kn_hookid);
753 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
759 if (kev->fflags & NOTE_TRIGGER)
762 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
763 kev->fflags &= NOTE_FFLAGSMASK;
769 kn->kn_sfflags &= kev->fflags;
773 kn->kn_sfflags |= kev->fflags;
777 kn->kn_sfflags = kev->fflags;
781 /* XXX Return error? */
784 kn->kn_sdata = kev->data;
785 if (kev->flags & EV_CLEAR) {
793 *kev = kn->kn_kevent;
794 kev->fflags = kn->kn_sfflags;
795 kev->data = kn->kn_sdata;
796 if (kn->kn_flags & EV_CLEAR) {
804 panic("filt_usertouch() - invalid type (%ld)", type);
810 sys_kqueue(struct thread *td, struct kqueue_args *uap)
813 return (kern_kqueue(td, 0, NULL));
817 kqueue_init(struct kqueue *kq)
820 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
821 TAILQ_INIT(&kq->kq_head);
822 knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
823 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
827 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
829 struct filedesc *fdp;
835 fdp = td->td_proc->p_fd;
837 if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
840 error = falloc_caps(td, &fp, &fd, flags, fcaps);
842 chgkqcnt(cred->cr_ruidinfo, -1, 0);
846 /* An extra reference on `fp' has been held for us by falloc(). */
847 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
850 kq->kq_cred = crhold(cred);
853 TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
854 FILEDESC_XUNLOCK(fdp);
856 finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
859 td->td_retval[0] = fd;
863 #ifndef _SYS_SYSPROTO_H_
866 const struct kevent *changelist;
868 struct kevent *eventlist;
870 const struct timespec *timeout;
874 sys_kevent(struct thread *td, struct kevent_args *uap)
876 struct timespec ts, *tsp;
877 struct kevent_copyops k_ops = {
879 .k_copyout = kevent_copyout,
880 .k_copyin = kevent_copyin,
883 struct kevent *eventlist = uap->eventlist;
887 if (uap->timeout != NULL) {
888 error = copyin(uap->timeout, &ts, sizeof(ts));
896 if (KTRPOINT(td, KTR_STRUCT_ARRAY))
897 ktrstructarray("kevent", UIO_USERSPACE, uap->changelist,
898 uap->nchanges, sizeof(struct kevent));
901 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
905 if (error == 0 && KTRPOINT(td, KTR_STRUCT_ARRAY))
906 ktrstructarray("kevent", UIO_USERSPACE, eventlist,
907 td->td_retval[0], sizeof(struct kevent));
914 * Copy 'count' items into the destination list pointed to by uap->eventlist.
917 kevent_copyout(void *arg, struct kevent *kevp, int count)
919 struct kevent_args *uap;
922 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
923 uap = (struct kevent_args *)arg;
925 error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
927 uap->eventlist += count;
932 * Copy 'count' items from the list pointed to by uap->changelist.
935 kevent_copyin(void *arg, struct kevent *kevp, int count)
937 struct kevent_args *uap;
940 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
941 uap = (struct kevent_args *)arg;
943 error = copyin(uap->changelist, kevp, count * sizeof *kevp);
945 uap->changelist += count;
950 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
951 struct kevent_copyops *k_ops, const struct timespec *timeout)
957 cap_rights_init(&rights);
959 cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
961 cap_rights_set(&rights, CAP_KQUEUE_EVENT);
962 error = fget(td, fd, &rights, &fp);
966 error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
973 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
974 struct kevent_copyops *k_ops, const struct timespec *timeout)
976 struct kevent keva[KQ_NEVENTS];
977 struct kevent *kevp, *changes;
978 int i, n, nerrors, error;
981 while (nchanges > 0) {
982 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
983 error = k_ops->k_copyin(k_ops->arg, keva, n);
987 for (i = 0; i < n; i++) {
991 kevp->flags &= ~EV_SYSFLAGS;
992 error = kqueue_register(kq, kevp, td, 1);
993 if (error || (kevp->flags & EV_RECEIPT)) {
996 kevp->flags = EV_ERROR;
998 (void)k_ops->k_copyout(k_ops->arg, kevp, 1);
1006 td->td_retval[0] = nerrors;
1010 return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
1014 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
1015 struct kevent_copyops *k_ops, const struct timespec *timeout)
1020 error = kqueue_acquire(fp, &kq);
1023 error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
1024 kqueue_release(kq, 0);
1029 * Performs a kevent() call on a temporarily created kqueue. This can be
1030 * used to perform one-shot polling, similar to poll() and select().
1033 kern_kevent_anonymous(struct thread *td, int nevents,
1034 struct kevent_copyops *k_ops)
1036 struct kqueue kq = {};
1041 error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
1042 kqueue_drain(&kq, td);
1043 kqueue_destroy(&kq);
1048 kqueue_add_filteropts(int filt, struct filterops *filtops)
1053 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1055 "trying to add a filterop that is out of range: %d is beyond %d\n",
1056 ~filt, EVFILT_SYSCOUNT);
1059 mtx_lock(&filterops_lock);
1060 if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1061 sysfilt_ops[~filt].for_fop != NULL)
1064 sysfilt_ops[~filt].for_fop = filtops;
1065 sysfilt_ops[~filt].for_refcnt = 0;
1067 mtx_unlock(&filterops_lock);
1073 kqueue_del_filteropts(int filt)
1078 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1081 mtx_lock(&filterops_lock);
1082 if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1083 sysfilt_ops[~filt].for_fop == NULL)
1085 else if (sysfilt_ops[~filt].for_refcnt != 0)
1088 sysfilt_ops[~filt].for_fop = &null_filtops;
1089 sysfilt_ops[~filt].for_refcnt = 0;
1091 mtx_unlock(&filterops_lock);
1096 static struct filterops *
1097 kqueue_fo_find(int filt)
1100 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1103 if (sysfilt_ops[~filt].for_nolock)
1104 return sysfilt_ops[~filt].for_fop;
1106 mtx_lock(&filterops_lock);
1107 sysfilt_ops[~filt].for_refcnt++;
1108 if (sysfilt_ops[~filt].for_fop == NULL)
1109 sysfilt_ops[~filt].for_fop = &null_filtops;
1110 mtx_unlock(&filterops_lock);
1112 return sysfilt_ops[~filt].for_fop;
1116 kqueue_fo_release(int filt)
1119 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1122 if (sysfilt_ops[~filt].for_nolock)
1125 mtx_lock(&filterops_lock);
1126 KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1127 ("filter object refcount not valid on release"));
1128 sysfilt_ops[~filt].for_refcnt--;
1129 mtx_unlock(&filterops_lock);
1133 * A ref to kq (obtained via kqueue_acquire) must be held. waitok will
1134 * influence if memory allocation should wait. Make sure it is 0 if you
1138 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1140 struct filterops *fops;
1142 struct knote *kn, *tkn;
1144 cap_rights_t rights;
1145 int error, filt, event;
1146 int haskqglobal, filedesc_unlock;
1148 if ((kev->flags & (EV_ENABLE | EV_DISABLE)) == (EV_ENABLE | EV_DISABLE))
1156 filedesc_unlock = 0;
1159 fops = kqueue_fo_find(filt);
1163 if (kev->flags & EV_ADD) {
1165 * Prevent waiting with locks. Non-sleepable
1166 * allocation failures are handled in the loop, only
1167 * if the spare knote appears to be actually required.
1169 tkn = knote_alloc(waitok);
1176 KASSERT(td != NULL, ("td is NULL"));
1177 if (kev->ident > INT_MAX)
1180 error = fget(td, kev->ident,
1181 cap_rights_init(&rights, CAP_EVENT), &fp);
1185 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1186 kev->ident, 0) != 0) {
1190 error = kqueue_expand(kq, fops, kev->ident, waitok);
1196 if (fp->f_type == DTYPE_KQUEUE) {
1198 * If we add some intelligence about what we are doing,
1199 * we should be able to support events on ourselves.
1200 * We need to know when we are doing this to prevent
1201 * getting both the knlist lock and the kq lock since
1202 * they are the same thing.
1204 if (fp->f_data == kq) {
1210 * Pre-lock the filedesc before the global
1211 * lock mutex, see the comment in
1214 FILEDESC_XLOCK(td->td_proc->p_fd);
1215 filedesc_unlock = 1;
1216 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1220 if (kev->ident < kq->kq_knlistsize) {
1221 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1222 if (kev->filter == kn->kn_filter)
1226 if ((kev->flags & EV_ADD) == EV_ADD)
1227 kqueue_expand(kq, fops, kev->ident, waitok);
1232 * If possible, find an existing knote to use for this kevent.
1234 if (kev->filter == EVFILT_PROC &&
1235 (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1236 /* This is an internal creation of a process tracking
1237 * note. Don't attempt to coalesce this with an
1241 } else if (kq->kq_knhashmask != 0) {
1244 list = &kq->kq_knhash[
1245 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1246 SLIST_FOREACH(kn, list, kn_link)
1247 if (kev->ident == kn->kn_id &&
1248 kev->filter == kn->kn_filter)
1253 /* knote is in the process of changing, wait for it to stabilize. */
1254 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1255 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1256 if (filedesc_unlock) {
1257 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1258 filedesc_unlock = 0;
1260 kq->kq_state |= KQ_FLUXWAIT;
1261 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1270 * kn now contains the matching knote, or NULL if no match
1273 if (kev->flags & EV_ADD) {
1285 * apply reference counts to knote structure, and
1286 * do not release it at the end of this routine.
1291 kn->kn_sfflags = kev->fflags;
1292 kn->kn_sdata = kev->data;
1295 kn->kn_kevent = *kev;
1296 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1297 EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1298 kn->kn_status = KN_INFLUX|KN_DETACHED;
1299 if ((kev->flags & EV_DISABLE) != 0)
1300 kn->kn_status |= KN_DISABLED;
1302 error = knote_attach(kn, kq);
1309 if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1313 knl = kn_list_lock(kn);
1316 /* No matching knote and the EV_ADD flag is not set. */
1323 if (kev->flags & EV_DELETE) {
1324 kn->kn_status |= KN_INFLUX;
1326 if (!(kn->kn_status & KN_DETACHED))
1327 kn->kn_fop->f_detach(kn);
1332 if (kev->flags & EV_FORCEONESHOT) {
1333 kn->kn_flags |= EV_ONESHOT;
1334 KNOTE_ACTIVATE(kn, 1);
1337 if ((kev->flags & EV_ENABLE) != 0)
1338 kn->kn_status &= ~KN_DISABLED;
1339 else if ((kev->flags & EV_DISABLE) != 0)
1340 kn->kn_status |= KN_DISABLED;
1343 * The user may change some filter values after the initial EV_ADD,
1344 * but doing so will not reset any filter which has already been
1347 kn->kn_status |= KN_INFLUX | KN_SCAN;
1349 knl = kn_list_lock(kn);
1350 kn->kn_kevent.udata = kev->udata;
1351 if (!fops->f_isfd && fops->f_touch != NULL) {
1352 fops->f_touch(kn, kev, EVENT_REGISTER);
1354 kn->kn_sfflags = kev->fflags;
1355 kn->kn_sdata = kev->data;
1360 * We can get here with kn->kn_knlist == NULL. This can happen when
1361 * the initial attach event decides that the event is "completed"
1362 * already, e.g., filt_procattach() is called on a zombie process. It
1363 * will call filt_proc() which will remove it from the list, and NULL
1366 * KN_DISABLED will be stable while the knote is in flux, so the
1367 * unlocked read will not race with an update.
1369 if ((kn->kn_status & KN_DISABLED) == 0)
1370 event = kn->kn_fop->f_event(kn, 0);
1376 kn->kn_status |= KN_ACTIVE;
1377 if ((kn->kn_status & (KN_ACTIVE | KN_DISABLED | KN_QUEUED)) ==
1380 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1381 kn_list_unlock(knl);
1385 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1386 if (filedesc_unlock)
1387 FILEDESC_XUNLOCK(td->td_proc->p_fd);
1392 kqueue_fo_release(filt);
1397 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1405 if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1409 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1420 kqueue_release(struct kqueue *kq, int locked)
1427 if (kq->kq_refcnt == 1)
1428 wakeup(&kq->kq_refcnt);
1434 kqueue_schedtask(struct kqueue *kq)
1438 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1439 ("scheduling kqueue task while draining"));
1441 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1442 taskqueue_enqueue(taskqueue_kqueue_ctx, &kq->kq_task);
1443 kq->kq_state |= KQ_TASKSCHED;
1448 * Expand the kq to make sure we have storage for fops/ident pair.
1450 * Return 0 on success (or no work necessary), return errno on failure.
1452 * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1453 * If kqueue_register is called from a non-fd context, there usually/should
1457 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1460 struct klist *list, *tmp_knhash, *to_free;
1461 u_long tmp_knhashmask;
1464 int mflag = waitok ? M_WAITOK : M_NOWAIT;
1471 if (kq->kq_knlistsize <= fd) {
1472 size = kq->kq_knlistsize;
1475 list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1479 if (kq->kq_knlistsize > fd) {
1483 if (kq->kq_knlist != NULL) {
1484 bcopy(kq->kq_knlist, list,
1485 kq->kq_knlistsize * sizeof(*list));
1486 to_free = kq->kq_knlist;
1487 kq->kq_knlist = NULL;
1489 bzero((caddr_t)list +
1490 kq->kq_knlistsize * sizeof(*list),
1491 (size - kq->kq_knlistsize) * sizeof(*list));
1492 kq->kq_knlistsize = size;
1493 kq->kq_knlist = list;
1498 if (kq->kq_knhashmask == 0) {
1499 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1501 if (tmp_knhash == NULL)
1504 if (kq->kq_knhashmask == 0) {
1505 kq->kq_knhash = tmp_knhash;
1506 kq->kq_knhashmask = tmp_knhashmask;
1508 to_free = tmp_knhash;
1513 free(to_free, M_KQUEUE);
1520 kqueue_task(void *arg, int pending)
1528 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1531 KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1533 kq->kq_state &= ~KQ_TASKSCHED;
1534 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1535 wakeup(&kq->kq_state);
1538 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1542 * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1543 * We treat KN_MARKER knotes as if they are INFLUX.
1546 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1547 const struct timespec *tsp, struct kevent *keva, struct thread *td)
1549 struct kevent *kevp;
1550 struct knote *kn, *marker;
1552 sbintime_t asbt, rsbt;
1553 int count, error, haskqglobal, influx, nkev, touch;
1565 if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1566 tsp->tv_nsec >= 1000000000) {
1570 if (timespecisset(tsp)) {
1571 if (tsp->tv_sec <= INT32_MAX) {
1572 rsbt = tstosbt(*tsp);
1573 if (TIMESEL(&asbt, rsbt))
1574 asbt += tc_tick_sbt;
1575 if (asbt <= SBT_MAX - rsbt)
1579 rsbt >>= tc_precexp;
1586 marker = knote_alloc(1);
1587 marker->kn_status = KN_MARKER;
1592 if (kq->kq_count == 0) {
1594 error = EWOULDBLOCK;
1596 kq->kq_state |= KQ_SLEEP;
1597 error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1598 "kqread", asbt, rsbt, C_ABSOLUTE);
1602 /* don't restart after signals... */
1603 if (error == ERESTART)
1605 else if (error == EWOULDBLOCK)
1610 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1614 kn = TAILQ_FIRST(&kq->kq_head);
1616 if ((kn->kn_status == KN_MARKER && kn != marker) ||
1617 (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1622 kq->kq_state |= KQ_FLUXWAIT;
1623 error = msleep(kq, &kq->kq_lock, PSOCK,
1628 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1629 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1630 kn->kn_status &= ~KN_QUEUED;
1636 if (count == maxevents)
1640 KASSERT((kn->kn_status & KN_INFLUX) == 0,
1641 ("KN_INFLUX set when not suppose to be"));
1643 if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1644 kn->kn_status &= ~KN_QUEUED;
1645 kn->kn_status |= KN_INFLUX;
1649 * We don't need to lock the list since we've marked
1652 if (!(kn->kn_status & KN_DETACHED))
1653 kn->kn_fop->f_detach(kn);
1657 } else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1658 kn->kn_status &= ~KN_QUEUED;
1659 kn->kn_status |= KN_INFLUX;
1663 * We don't need to lock the list since we've marked
1666 *kevp = kn->kn_kevent;
1667 if (!(kn->kn_status & KN_DETACHED))
1668 kn->kn_fop->f_detach(kn);
1673 kn->kn_status |= KN_INFLUX | KN_SCAN;
1675 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1676 KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1677 knl = kn_list_lock(kn);
1678 if (kn->kn_fop->f_event(kn, 0) == 0) {
1680 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1682 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1685 kn_list_unlock(knl);
1689 touch = (!kn->kn_fop->f_isfd &&
1690 kn->kn_fop->f_touch != NULL);
1692 kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1694 *kevp = kn->kn_kevent;
1696 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1697 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1699 * Manually clear knotes who weren't
1702 if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1706 if (kn->kn_flags & EV_DISPATCH)
1707 kn->kn_status |= KN_DISABLED;
1708 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1711 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1713 kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1714 kn_list_unlock(knl);
1718 /* we are returning a copy to the user */
1723 if (nkev == KQ_NEVENTS) {
1726 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1734 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1742 error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1743 td->td_retval[0] = maxevents - count;
1749 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1750 struct ucred *active_cred, struct thread *td)
1753 * Enabling sigio causes two major problems:
1754 * 1) infinite recursion:
1755 * Synopsys: kevent is being used to track signals and have FIOASYNC
1756 * set. On receipt of a signal this will cause a kqueue to recurse
1757 * into itself over and over. Sending the sigio causes the kqueue
1758 * to become ready, which in turn posts sigio again, forever.
1759 * Solution: this can be solved by setting a flag in the kqueue that
1760 * we have a SIGIO in progress.
1761 * 2) locking problems:
1762 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1763 * us above the proc and pgrp locks.
1764 * Solution: Post a signal using an async mechanism, being sure to
1765 * record a generation count in the delivery so that we do not deliver
1766 * a signal to the wrong process.
1768 * Note, these two mechanisms are somewhat mutually exclusive!
1777 kq->kq_state |= KQ_ASYNC;
1779 kq->kq_state &= ~KQ_ASYNC;
1784 return (fsetown(*(int *)data, &kq->kq_sigio));
1787 *(int *)data = fgetown(&kq->kq_sigio);
1797 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1804 if ((error = kqueue_acquire(fp, &kq)))
1808 if (events & (POLLIN | POLLRDNORM)) {
1810 revents |= events & (POLLIN | POLLRDNORM);
1812 selrecord(td, &kq->kq_sel);
1813 if (SEL_WAITING(&kq->kq_sel))
1814 kq->kq_state |= KQ_SEL;
1817 kqueue_release(kq, 1);
1824 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1828 bzero((void *)st, sizeof *st);
1830 * We no longer return kq_count because the unlocked value is useless.
1831 * If you spent all this time getting the count, why not spend your
1832 * syscall better by calling kevent?
1834 * XXX - This is needed for libc_r.
1836 st->st_mode = S_IFIFO;
1841 kqueue_drain(struct kqueue *kq, struct thread *td)
1848 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1849 ("kqueue already closing"));
1850 kq->kq_state |= KQ_CLOSING;
1851 if (kq->kq_refcnt > 1)
1852 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1854 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1856 KASSERT(knlist_empty(&kq->kq_sel.si_note),
1857 ("kqueue's knlist not empty"));
1859 for (i = 0; i < kq->kq_knlistsize; i++) {
1860 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1861 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1862 kq->kq_state |= KQ_FLUXWAIT;
1863 msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1866 kn->kn_status |= KN_INFLUX;
1868 if (!(kn->kn_status & KN_DETACHED))
1869 kn->kn_fop->f_detach(kn);
1874 if (kq->kq_knhashmask != 0) {
1875 for (i = 0; i <= kq->kq_knhashmask; i++) {
1876 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1877 if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1878 kq->kq_state |= KQ_FLUXWAIT;
1879 msleep(kq, &kq->kq_lock, PSOCK,
1883 kn->kn_status |= KN_INFLUX;
1885 if (!(kn->kn_status & KN_DETACHED))
1886 kn->kn_fop->f_detach(kn);
1893 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1894 kq->kq_state |= KQ_TASKDRAIN;
1895 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1898 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1899 selwakeuppri(&kq->kq_sel, PSOCK);
1900 if (!SEL_WAITING(&kq->kq_sel))
1901 kq->kq_state &= ~KQ_SEL;
1908 kqueue_destroy(struct kqueue *kq)
1911 KASSERT(kq->kq_fdp == NULL,
1912 ("kqueue still attached to a file descriptor"));
1913 seldrain(&kq->kq_sel);
1914 knlist_destroy(&kq->kq_sel.si_note);
1915 mtx_destroy(&kq->kq_lock);
1917 if (kq->kq_knhash != NULL)
1918 free(kq->kq_knhash, M_KQUEUE);
1919 if (kq->kq_knlist != NULL)
1920 free(kq->kq_knlist, M_KQUEUE);
1922 funsetown(&kq->kq_sigio);
1927 kqueue_close(struct file *fp, struct thread *td)
1929 struct kqueue *kq = fp->f_data;
1930 struct filedesc *fdp;
1932 int filedesc_unlock;
1934 if ((error = kqueue_acquire(fp, &kq)))
1936 kqueue_drain(kq, td);
1939 * We could be called due to the knote_drop() doing fdrop(),
1940 * called from kqueue_register(). In this case the global
1941 * lock is owned, and filedesc sx is locked before, to not
1942 * take the sleepable lock after non-sleepable.
1946 if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1947 FILEDESC_XLOCK(fdp);
1948 filedesc_unlock = 1;
1950 filedesc_unlock = 0;
1951 TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1952 if (filedesc_unlock)
1953 FILEDESC_XUNLOCK(fdp);
1956 chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
1957 crfree(kq->kq_cred);
1965 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1968 kif->kf_type = KF_TYPE_KQUEUE;
1973 kqueue_wakeup(struct kqueue *kq)
1977 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1978 kq->kq_state &= ~KQ_SLEEP;
1981 if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1982 selwakeuppri(&kq->kq_sel, PSOCK);
1983 if (!SEL_WAITING(&kq->kq_sel))
1984 kq->kq_state &= ~KQ_SEL;
1986 if (!knlist_empty(&kq->kq_sel.si_note))
1987 kqueue_schedtask(kq);
1988 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1989 pgsigio(&kq->kq_sigio, SIGIO, 0);
1994 * Walk down a list of knotes, activating them if their event has triggered.
1996 * There is a possibility to optimize in the case of one kq watching another.
1997 * Instead of scheduling a task to wake it up, you could pass enough state
1998 * down the chain to make up the parent kqueue. Make this code functional
2002 knote(struct knlist *list, long hint, int lockflags)
2005 struct knote *kn, *tkn;
2012 KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
2014 if ((lockflags & KNF_LISTLOCKED) == 0)
2015 list->kl_lock(list->kl_lockarg);
2018 * If we unlock the list lock (and set KN_INFLUX), we can
2019 * eliminate the kqueue scheduling, but this will introduce
2020 * four lock/unlock's for each knote to test. Also, marker
2021 * would be needed to keep iteration position, since filters
2022 * or other threads could remove events.
2024 SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
2027 if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
2029 * Do not process the influx notes, except for
2030 * the influx coming from the kq unlock in the
2031 * kqueue_scan(). In the later case, we do
2032 * not interfere with the scan, since the code
2033 * fragment in kqueue_scan() locks the knlist,
2034 * and cannot proceed until we finished.
2037 } else if ((lockflags & KNF_NOKQLOCK) != 0) {
2038 own_influx = (kn->kn_status & KN_INFLUX) == 0;
2040 kn->kn_status |= KN_INFLUX;
2042 error = kn->kn_fop->f_event(kn, hint);
2045 kn->kn_status &= ~KN_INFLUX;
2047 KNOTE_ACTIVATE(kn, 1);
2050 kn->kn_status |= KN_HASKQLOCK;
2051 if (kn->kn_fop->f_event(kn, hint))
2052 KNOTE_ACTIVATE(kn, 1);
2053 kn->kn_status &= ~KN_HASKQLOCK;
2057 if ((lockflags & KNF_LISTLOCKED) == 0)
2058 list->kl_unlock(list->kl_lockarg);
2062 * add a knote to a knlist
2065 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2067 KNL_ASSERT_LOCK(knl, islocked);
2068 KQ_NOTOWNED(kn->kn_kq);
2069 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
2070 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
2072 knl->kl_lock(knl->kl_lockarg);
2073 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2075 knl->kl_unlock(knl->kl_lockarg);
2077 kn->kn_knlist = knl;
2078 kn->kn_status &= ~KN_DETACHED;
2079 KQ_UNLOCK(kn->kn_kq);
2083 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked,
2086 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
2087 KNL_ASSERT_LOCK(knl, knlislocked);
2088 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2090 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
2091 ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
2093 knl->kl_lock(knl->kl_lockarg);
2094 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2095 kn->kn_knlist = NULL;
2097 kn_list_unlock(knl);
2100 kn->kn_status |= KN_DETACHED;
2102 KQ_UNLOCK(kn->kn_kq);
2106 * remove knote from the specified knlist
2109 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2112 knlist_remove_kq(knl, kn, islocked, 0);
2116 knlist_empty(struct knlist *knl)
2119 KNL_ASSERT_LOCKED(knl);
2120 return (SLIST_EMPTY(&knl->kl_list));
2123 static struct mtx knlist_lock;
2124 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2126 static void knlist_mtx_lock(void *arg);
2127 static void knlist_mtx_unlock(void *arg);
2130 knlist_mtx_lock(void *arg)
2133 mtx_lock((struct mtx *)arg);
2137 knlist_mtx_unlock(void *arg)
2140 mtx_unlock((struct mtx *)arg);
2144 knlist_mtx_assert_locked(void *arg)
2147 mtx_assert((struct mtx *)arg, MA_OWNED);
2151 knlist_mtx_assert_unlocked(void *arg)
2154 mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2158 knlist_rw_rlock(void *arg)
2161 rw_rlock((struct rwlock *)arg);
2165 knlist_rw_runlock(void *arg)
2168 rw_runlock((struct rwlock *)arg);
2172 knlist_rw_assert_locked(void *arg)
2175 rw_assert((struct rwlock *)arg, RA_LOCKED);
2179 knlist_rw_assert_unlocked(void *arg)
2182 rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2186 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2187 void (*kl_unlock)(void *),
2188 void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2192 knl->kl_lockarg = &knlist_lock;
2194 knl->kl_lockarg = lock;
2196 if (kl_lock == NULL)
2197 knl->kl_lock = knlist_mtx_lock;
2199 knl->kl_lock = kl_lock;
2200 if (kl_unlock == NULL)
2201 knl->kl_unlock = knlist_mtx_unlock;
2203 knl->kl_unlock = kl_unlock;
2204 if (kl_assert_locked == NULL)
2205 knl->kl_assert_locked = knlist_mtx_assert_locked;
2207 knl->kl_assert_locked = kl_assert_locked;
2208 if (kl_assert_unlocked == NULL)
2209 knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2211 knl->kl_assert_unlocked = kl_assert_unlocked;
2213 knl->kl_autodestroy = 0;
2214 SLIST_INIT(&knl->kl_list);
2218 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2221 knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2225 knlist_alloc(struct mtx *lock)
2229 knl = malloc(sizeof(struct knlist), M_KQUEUE, M_WAITOK);
2230 knlist_init_mtx(knl, lock);
2235 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2238 knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2239 knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2243 knlist_destroy(struct knlist *knl)
2246 KASSERT(KNLIST_EMPTY(knl),
2247 ("destroying knlist %p with knotes on it", knl));
2251 knlist_detach(struct knlist *knl)
2254 KNL_ASSERT_LOCKED(knl);
2255 knl->kl_autodestroy = 1;
2256 if (knlist_empty(knl)) {
2257 knlist_destroy(knl);
2258 free(knl, M_KQUEUE);
2263 * Even if we are locked, we may need to drop the lock to allow any influx
2264 * knotes time to "settle".
2267 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2269 struct knote *kn, *kn2;
2272 KASSERT(!knl->kl_autodestroy, ("cleardel for autodestroy %p", knl));
2274 KNL_ASSERT_LOCKED(knl);
2276 KNL_ASSERT_UNLOCKED(knl);
2277 again: /* need to reacquire lock since we have dropped it */
2278 knl->kl_lock(knl->kl_lockarg);
2281 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2284 if ((kn->kn_status & KN_INFLUX)) {
2288 knlist_remove_kq(knl, kn, 1, 1);
2290 kn->kn_status |= KN_INFLUX | KN_DETACHED;
2294 /* Make sure cleared knotes disappear soon */
2295 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2301 if (!SLIST_EMPTY(&knl->kl_list)) {
2302 /* there are still KN_INFLUX remaining */
2303 kn = SLIST_FIRST(&knl->kl_list);
2306 KASSERT(kn->kn_status & KN_INFLUX,
2307 ("knote removed w/o list lock"));
2308 knl->kl_unlock(knl->kl_lockarg);
2309 kq->kq_state |= KQ_FLUXWAIT;
2310 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2316 KNL_ASSERT_LOCKED(knl);
2318 knl->kl_unlock(knl->kl_lockarg);
2319 KNL_ASSERT_UNLOCKED(knl);
2324 * Remove all knotes referencing a specified fd must be called with FILEDESC
2325 * lock. This prevents a race where a new fd comes along and occupies the
2326 * entry and we attach a knote to the fd.
2329 knote_fdclose(struct thread *td, int fd)
2331 struct filedesc *fdp = td->td_proc->p_fd;
2336 FILEDESC_XLOCK_ASSERT(fdp);
2339 * We shouldn't have to worry about new kevents appearing on fd
2340 * since filedesc is locked.
2342 TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2347 while (kq->kq_knlistsize > fd &&
2348 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2349 if (kn->kn_status & KN_INFLUX) {
2350 /* someone else might be waiting on our knote */
2353 kq->kq_state |= KQ_FLUXWAIT;
2354 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2357 kn->kn_status |= KN_INFLUX;
2359 if (!(kn->kn_status & KN_DETACHED))
2360 kn->kn_fop->f_detach(kn);
2370 knote_attach(struct knote *kn, struct kqueue *kq)
2374 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2377 if (kn->kn_fop->f_isfd) {
2378 if (kn->kn_id >= kq->kq_knlistsize)
2380 list = &kq->kq_knlist[kn->kn_id];
2382 if (kq->kq_knhash == NULL)
2384 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2386 SLIST_INSERT_HEAD(list, kn, kn_link);
2391 * knote must already have been detached using the f_detach method.
2392 * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2393 * to prevent other removal.
2396 knote_drop(struct knote *kn, struct thread *td)
2404 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2405 ("knote_drop called without KN_INFLUX set in kn_status"));
2408 if (kn->kn_fop->f_isfd)
2409 list = &kq->kq_knlist[kn->kn_id];
2411 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2413 if (!SLIST_EMPTY(list))
2414 SLIST_REMOVE(list, kn, knote, kn_link);
2415 if (kn->kn_status & KN_QUEUED)
2419 if (kn->kn_fop->f_isfd) {
2420 fdrop(kn->kn_fp, td);
2423 kqueue_fo_release(kn->kn_kevent.filter);
2429 knote_enqueue(struct knote *kn)
2431 struct kqueue *kq = kn->kn_kq;
2433 KQ_OWNED(kn->kn_kq);
2434 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2436 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2437 kn->kn_status |= KN_QUEUED;
2443 knote_dequeue(struct knote *kn)
2445 struct kqueue *kq = kn->kn_kq;
2447 KQ_OWNED(kn->kn_kq);
2448 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2450 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2451 kn->kn_status &= ~KN_QUEUED;
2459 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2460 NULL, NULL, UMA_ALIGN_PTR, 0);
2462 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2464 static struct knote *
2465 knote_alloc(int waitok)
2468 return (uma_zalloc(knote_zone, (waitok ? M_WAITOK : M_NOWAIT) |
2473 knote_free(struct knote *kn)
2476 uma_zfree(knote_zone, kn);
2480 * Register the kev w/ the kq specified by fd.
2483 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2487 cap_rights_t rights;
2490 error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2493 if ((error = kqueue_acquire(fp, &kq)) != 0)
2496 error = kqueue_register(kq, kev, td, waitok);
2497 kqueue_release(kq, 0);