2 * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * $FreeBSD: src/sys/kern/kern_event.c,v 1.2.2.10 2004/04/04 07:03:14 cperciva Exp $
29 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/kernel.h>
33 #include <sys/malloc.h>
34 #include <sys/unistd.h>
37 #include <sys/fcntl.h>
38 #include <sys/queue.h>
39 #include <sys/event.h>
40 #include <sys/eventvar.h>
41 #include <sys/protosw.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
45 #include <sys/sysctl.h>
46 #include <sys/sysproto.h>
47 #include <sys/thread.h>
49 #include <sys/signalvar.h>
50 #include <sys/filio.h>
53 #include <sys/thread2.h>
54 #include <sys/file2.h>
55 #include <sys/mplock2.h>
57 #define EVENT_REGISTER 1
58 #define EVENT_PROCESS 2
61 * Global token for kqueue subsystem
64 struct lwkt_token kq_token = LWKT_TOKEN_INITIALIZER(kq_token);
65 SYSCTL_LONG(_lwkt, OID_AUTO, kq_collisions,
66 CTLFLAG_RW, &kq_token.t_collisions, 0,
67 "Collision counter of kq_token");
70 MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
72 struct kevent_copyin_args {
73 struct kevent_args *ka;
77 static int kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
78 struct knote *marker);
79 static int kqueue_read(struct file *fp, struct uio *uio,
80 struct ucred *cred, int flags);
81 static int kqueue_write(struct file *fp, struct uio *uio,
82 struct ucred *cred, int flags);
83 static int kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
84 struct ucred *cred, struct sysmsg *msg);
85 static int kqueue_kqfilter(struct file *fp, struct knote *kn);
86 static int kqueue_stat(struct file *fp, struct stat *st,
88 static int kqueue_close(struct file *fp);
89 static void kqueue_wakeup(struct kqueue *kq);
90 static int filter_attach(struct knote *kn);
91 static int filter_event(struct knote *kn, long hint);
96 static struct fileops kqueueops = {
97 .fo_read = kqueue_read,
98 .fo_write = kqueue_write,
99 .fo_ioctl = kqueue_ioctl,
100 .fo_kqfilter = kqueue_kqfilter,
101 .fo_stat = kqueue_stat,
102 .fo_close = kqueue_close,
103 .fo_shutdown = nofo_shutdown
106 static void knote_attach(struct knote *kn);
107 static void knote_drop(struct knote *kn);
108 static void knote_detach_and_drop(struct knote *kn);
109 static void knote_enqueue(struct knote *kn);
110 static void knote_dequeue(struct knote *kn);
111 static struct knote *knote_alloc(void);
112 static void knote_free(struct knote *kn);
114 static void filt_kqdetach(struct knote *kn);
115 static int filt_kqueue(struct knote *kn, long hint);
116 static int filt_procattach(struct knote *kn);
117 static void filt_procdetach(struct knote *kn);
118 static int filt_proc(struct knote *kn, long hint);
119 static int filt_fileattach(struct knote *kn);
120 static void filt_timerexpire(void *knx);
121 static int filt_timerattach(struct knote *kn);
122 static void filt_timerdetach(struct knote *kn);
123 static int filt_timer(struct knote *kn, long hint);
124 static int filt_userattach(struct knote *kn);
125 static void filt_userdetach(struct knote *kn);
126 static int filt_user(struct knote *kn, long hint);
127 static void filt_usertouch(struct knote *kn, struct kevent *kev,
130 static struct filterops file_filtops =
131 { FILTEROP_ISFD | FILTEROP_MPSAFE, filt_fileattach, NULL, NULL };
132 static struct filterops kqread_filtops =
133 { FILTEROP_ISFD | FILTEROP_MPSAFE, NULL, filt_kqdetach, filt_kqueue };
134 static struct filterops proc_filtops =
135 { 0, filt_procattach, filt_procdetach, filt_proc };
136 static struct filterops timer_filtops =
137 { 0, filt_timerattach, filt_timerdetach, filt_timer };
138 static struct filterops user_filtops =
139 { 0, filt_userattach, filt_userdetach, filt_user };
141 static int kq_ncallouts = 0;
142 static int kq_calloutmax = (4 * 1024);
143 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
144 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
145 static int kq_checkloop = 1000000;
146 SYSCTL_INT(_kern, OID_AUTO, kq_checkloop, CTLFLAG_RW,
147 &kq_checkloop, 0, "Maximum number of callouts allocated for kqueue");
149 #define KNOTE_ACTIVATE(kn) do { \
150 kn->kn_status |= KN_ACTIVE; \
151 if ((kn->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \
155 #define KN_HASHSIZE 64 /* XXX should be tunable */
156 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask))
158 extern struct filterops aio_filtops;
159 extern struct filterops sig_filtops;
162 * Table for for all system-defined filters.
164 static struct filterops *sysfilt_ops[] = {
165 &file_filtops, /* EVFILT_READ */
166 &file_filtops, /* EVFILT_WRITE */
167 &aio_filtops, /* EVFILT_AIO */
168 &file_filtops, /* EVFILT_VNODE */
169 &proc_filtops, /* EVFILT_PROC */
170 &sig_filtops, /* EVFILT_SIGNAL */
171 &timer_filtops, /* EVFILT_TIMER */
172 &file_filtops, /* EVFILT_EXCEPT */
173 &user_filtops, /* EVFILT_USER */
177 filt_fileattach(struct knote *kn)
179 return (fo_kqfilter(kn->kn_fp, kn));
186 kqueue_kqfilter(struct file *fp, struct knote *kn)
188 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
190 if (kn->kn_filter != EVFILT_READ)
193 kn->kn_fop = &kqread_filtops;
194 knote_insert(&kq->kq_kqinfo.ki_note, kn);
199 filt_kqdetach(struct knote *kn)
201 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
203 knote_remove(&kq->kq_kqinfo.ki_note, kn);
208 filt_kqueue(struct knote *kn, long hint)
210 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
212 kn->kn_data = kq->kq_count;
213 return (kn->kn_data > 0);
217 filt_procattach(struct knote *kn)
223 p = pfind(kn->kn_id);
224 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
225 p = zpfind(kn->kn_id);
231 if (!PRISON_CHECK(curthread->td_ucred, p->p_ucred)) {
237 lwkt_gettoken(&p->p_token);
238 kn->kn_ptr.p_proc = p;
239 kn->kn_flags |= EV_CLEAR; /* automatically set */
242 * internal flag indicating registration done by kernel
244 if (kn->kn_flags & EV_FLAG1) {
245 kn->kn_data = kn->kn_sdata; /* ppid */
246 kn->kn_fflags = NOTE_CHILD;
247 kn->kn_flags &= ~EV_FLAG1;
250 knote_insert(&p->p_klist, kn);
253 * Immediately activate any exit notes if the target process is a
254 * zombie. This is necessary to handle the case where the target
255 * process, e.g. a child, dies before the kevent is negistered.
257 if (immediate && filt_proc(kn, NOTE_EXIT))
259 lwkt_reltoken(&p->p_token);
266 * The knote may be attached to a different process, which may exit,
267 * leaving nothing for the knote to be attached to. So when the process
268 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
269 * it will be deleted when read out. However, as part of the knote deletion,
270 * this routine is called, so a check is needed to avoid actually performing
271 * a detach, because the original process does not exist any more.
274 filt_procdetach(struct knote *kn)
278 if (kn->kn_status & KN_DETACHED)
280 p = kn->kn_ptr.p_proc;
281 knote_remove(&p->p_klist, kn);
285 filt_proc(struct knote *kn, long hint)
290 * mask off extra data
292 event = (u_int)hint & NOTE_PCTRLMASK;
295 * if the user is interested in this event, record it.
297 if (kn->kn_sfflags & event)
298 kn->kn_fflags |= event;
301 * Process is gone, so flag the event as finished. Detach the
302 * knote from the process now because the process will be poof,
305 if (event == NOTE_EXIT) {
306 struct proc *p = kn->kn_ptr.p_proc;
307 if ((kn->kn_status & KN_DETACHED) == 0) {
309 knote_remove(&p->p_klist, kn);
310 kn->kn_status |= KN_DETACHED;
311 kn->kn_data = p->p_xstat;
312 kn->kn_ptr.p_proc = NULL;
315 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
320 * process forked, and user wants to track the new process,
321 * so attach a new knote to it, and immediately report an
322 * event with the parent's pid.
324 if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) {
329 * register knote with new process.
331 kev.ident = hint & NOTE_PDATAMASK; /* pid */
332 kev.filter = kn->kn_filter;
333 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
334 kev.fflags = kn->kn_sfflags;
335 kev.data = kn->kn_id; /* parent */
336 kev.udata = kn->kn_kevent.udata; /* preserve udata */
337 error = kqueue_register(kn->kn_kq, &kev);
339 kn->kn_fflags |= NOTE_TRACKERR;
342 return (kn->kn_fflags != 0);
346 * The callout interlocks with callout_terminate() but can still
347 * race a deletion so if KN_DELETING is set we just don't touch
351 filt_timerexpire(void *knx)
353 struct lwkt_token *tok;
354 struct knote *kn = knx;
355 struct callout *calloutp;
359 tok = lwkt_token_pool_lookup(kn->kn_kq);
361 if ((kn->kn_status & KN_DELETING) == 0) {
365 if ((kn->kn_flags & EV_ONESHOT) == 0) {
366 tv.tv_sec = kn->kn_sdata / 1000;
367 tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
368 tticks = tvtohz_high(&tv);
369 calloutp = (struct callout *)kn->kn_hook;
370 callout_reset(calloutp, tticks, filt_timerexpire, kn);
377 * data contains amount of time to sleep, in milliseconds
380 filt_timerattach(struct knote *kn)
382 struct callout *calloutp;
386 if (kq_ncallouts >= kq_calloutmax) {
392 tv.tv_sec = kn->kn_sdata / 1000;
393 tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
394 tticks = tvtohz_high(&tv);
396 kn->kn_flags |= EV_CLEAR; /* automatically set */
397 calloutp = kmalloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
398 callout_init(calloutp);
399 kn->kn_hook = (caddr_t)calloutp;
400 callout_reset(calloutp, tticks, filt_timerexpire, kn);
406 * This function is called with the knote flagged locked but it is
407 * still possible to race a callout event due to the callback blocking.
408 * We must call callout_terminate() instead of callout_stop() to deal
412 filt_timerdetach(struct knote *kn)
414 struct callout *calloutp;
416 calloutp = (struct callout *)kn->kn_hook;
417 callout_terminate(calloutp);
418 kfree(calloutp, M_KQUEUE);
423 filt_timer(struct knote *kn, long hint)
426 return (kn->kn_data != 0);
433 filt_userattach(struct knote *kn)
436 if (kn->kn_fflags & NOTE_TRIGGER)
437 kn->kn_ptr.hookid = 1;
439 kn->kn_ptr.hookid = 0;
444 * This function is called with the knote flagged locked but it is
445 * still possible to race a callout event due to the callback blocking.
446 * We must call callout_terminate() instead of callout_stop() to deal
450 filt_userdetach(struct knote *kn)
456 filt_user(struct knote *kn, long hint)
458 return (kn->kn_ptr.hookid);
462 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
468 if (kev->fflags & NOTE_TRIGGER)
469 kn->kn_ptr.hookid = 1;
471 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
472 kev->fflags &= NOTE_FFLAGSMASK;
478 kn->kn_sfflags &= kev->fflags;
482 kn->kn_sfflags |= kev->fflags;
486 kn->kn_sfflags = kev->fflags;
490 /* XXX Return error? */
493 kn->kn_sdata = kev->data;
496 * This is not the correct use of EV_CLEAR in an event
497 * modification, it should have been passed as a NOTE instead.
498 * But we need to maintain compatibility with Apple & FreeBSD.
500 * Note however that EV_CLEAR can still be used when doing
501 * the initial registration of the event and works as expected
502 * (clears the event on reception).
504 if (kev->flags & EV_CLEAR) {
505 kn->kn_ptr.hookid = 0;
512 *kev = kn->kn_kevent;
513 kev->fflags = kn->kn_sfflags;
514 kev->data = kn->kn_sdata;
515 if (kn->kn_flags & EV_CLEAR) {
516 kn->kn_ptr.hookid = 0;
517 /* kn_data, kn_fflags handled by parent */
522 panic("filt_usertouch() - invalid type (%ld)", type);
528 * Acquire a knote, return non-zero on success, 0 on failure.
530 * If we cannot acquire the knote we sleep and return 0. The knote
531 * may be stale on return in this case and the caller must restart
532 * whatever loop they are in.
534 * Related kq token must be held.
538 knote_acquire(struct knote *kn)
540 if (kn->kn_status & KN_PROCESSING) {
541 kn->kn_status |= KN_WAITING | KN_REPROCESS;
542 tsleep(kn, 0, "kqepts", hz);
543 /* knote may be stale now */
546 kn->kn_status |= KN_PROCESSING;
551 * Release an acquired knote, clearing KN_PROCESSING and handling any
552 * KN_REPROCESS events.
554 * Caller must be holding the related kq token
556 * Non-zero is returned if the knote is destroyed or detached.
560 knote_release(struct knote *kn)
562 while (kn->kn_status & KN_REPROCESS) {
563 kn->kn_status &= ~KN_REPROCESS;
564 if (kn->kn_status & KN_WAITING) {
565 kn->kn_status &= ~KN_WAITING;
568 if (kn->kn_status & KN_DELETING) {
569 knote_detach_and_drop(kn);
573 if (filter_event(kn, 0))
576 if (kn->kn_status & KN_DETACHED) {
577 kn->kn_status &= ~KN_PROCESSING;
580 kn->kn_status &= ~KN_PROCESSING;
586 * Initialize a kqueue.
588 * NOTE: The lwp/proc code initializes a kqueue for select/poll ops.
593 kqueue_init(struct kqueue *kq, struct filedesc *fdp)
595 TAILQ_INIT(&kq->kq_knpend);
596 TAILQ_INIT(&kq->kq_knlist);
599 SLIST_INIT(&kq->kq_kqinfo.ki_note);
603 * Terminate a kqueue. Freeing the actual kq itself is left up to the
604 * caller (it might be embedded in a lwp so we don't do it here).
606 * The kq's knlist must be completely eradicated so block on any
610 kqueue_terminate(struct kqueue *kq)
612 struct lwkt_token *tok;
615 tok = lwkt_token_pool_lookup(kq);
617 while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) {
618 if (knote_acquire(kn))
619 knote_detach_and_drop(kn);
622 hashdestroy(kq->kq_knhash, M_KQUEUE, kq->kq_knhashmask);
623 kq->kq_knhash = NULL;
624 kq->kq_knhashmask = 0;
633 sys_kqueue(struct kqueue_args *uap)
635 struct thread *td = curthread;
640 error = falloc(td->td_lwp, &fp, &fd);
643 fp->f_flag = FREAD | FWRITE;
644 fp->f_type = DTYPE_KQUEUE;
645 fp->f_ops = &kqueueops;
647 kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO);
648 kqueue_init(kq, td->td_proc->p_fd);
651 fsetfd(kq->kq_fdp, fp, fd);
652 uap->sysmsg_result = fd;
658 * Copy 'count' items into the destination list pointed to by uap->eventlist.
661 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res)
663 struct kevent_copyin_args *kap;
666 kap = (struct kevent_copyin_args *)arg;
668 error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp));
670 kap->ka->eventlist += count;
680 * Copy at most 'max' items from the list pointed to by kap->changelist,
681 * return number of items in 'events'.
684 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events)
686 struct kevent_copyin_args *kap;
689 kap = (struct kevent_copyin_args *)arg;
691 count = min(kap->ka->nchanges - kap->pchanges, max);
692 error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp);
694 kap->ka->changelist += count;
695 kap->pchanges += count;
706 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap,
707 k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn,
708 struct timespec *tsp_in)
711 struct timespec *tsp, ats;
712 int i, n, total, error, nerrors = 0;
714 int limit = kq_checkloop;
715 struct kevent kev[KQ_NEVENTS];
717 struct lwkt_token *tok;
719 if (tsp_in == NULL || tsp_in->tv_sec || tsp_in->tv_nsec)
720 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC);
727 error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n);
732 for (i = 0; i < n; i++) {
734 kevp->flags &= ~EV_SYSFLAGS;
735 error = kqueue_register(kq, kevp);
738 * If a registration returns an error we
739 * immediately post the error. The kevent()
740 * call itself will fail with the error if
741 * no space is available for posting.
743 * Such errors normally bypass the timeout/blocking
744 * code. However, if the copyoutfn function refuses
745 * to post the error (see sys_poll()), then we
749 kevp->flags = EV_ERROR;
752 kevent_copyoutfn(uap, kevp, 1, res);
755 } else if (lres != *res) {
766 * Acquire/wait for events - setup timeout
769 if (tsp->tv_sec || tsp->tv_nsec) {
771 timespecadd(tsp, &ats); /* tsp = target time */
778 * Collect as many events as we can. Sleeping on successive
779 * loops is disabled if copyoutfn has incremented (*res).
781 * The loop stops if an error occurs, all events have been
782 * scanned (the marker has been reached), or fewer than the
783 * maximum number of events is found.
785 * The copyoutfn function does not have to increment (*res) in
786 * order for the loop to continue.
788 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents.
792 marker.kn_filter = EVFILT_MARKER;
793 marker.kn_status = KN_PROCESSING;
794 tok = lwkt_token_pool_lookup(kq);
796 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
798 while ((n = nevents - total) > 0) {
803 * If no events are pending sleep until timeout (if any)
804 * or an event occurs.
806 * After the sleep completes the marker is moved to the
807 * end of the list, making any received events available
810 if (kq->kq_count == 0 && *res == 0) {
815 } else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) {
819 struct timespec atx = *tsp;
822 timespecsub(&atx, &ats);
823 if (atx.tv_sec < 0) {
827 timeout = atx.tv_sec > 24 * 60 * 60 ?
834 if (kq->kq_count == 0) {
835 kq->kq_state |= KQ_SLEEP;
836 error = tsleep(kq, PCATCH, "kqread", timeout);
838 /* don't restart after signals... */
839 if (error == ERESTART)
846 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
847 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker,
854 * Process all received events
855 * Account for all non-spurious events in our total
857 i = kqueue_scan(kq, kev, n, &marker);
860 error = kevent_copyoutfn(uap, kev, i, res);
861 total += *res - lres;
865 if (limit && --limit == 0)
866 panic("kqueue: checkloop failed i=%d", i);
869 * Normally when fewer events are returned than requested
870 * we can stop. However, if only spurious events were
871 * collected the copyout will not bump (*res) and we have
878 * Deal with an edge case where spurious events can cause
879 * a loop to occur without moving the marker. This can
880 * prevent kqueue_scan() from picking up new events which
881 * race us. We must be sure to move the marker for this
884 * NOTE: We do not want to move the marker if events
885 * were scanned because normal kqueue operations
886 * may reactivate events. Moving the marker in
887 * that case could result in duplicates for the
892 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
893 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
898 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
901 /* Timeouts do not return EWOULDBLOCK. */
902 if (error == EWOULDBLOCK)
911 sys_kevent(struct kevent_args *uap)
913 struct thread *td = curthread;
914 struct proc *p = td->td_proc;
915 struct timespec ts, *tsp;
917 struct file *fp = NULL;
918 struct kevent_copyin_args *kap, ka;
922 error = copyin(uap->timeout, &ts, sizeof(ts));
929 fp = holdfp(p->p_fd, uap->fd, -1);
932 if (fp->f_type != DTYPE_KQUEUE) {
937 kq = (struct kqueue *)fp->f_data;
943 error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap,
944 kevent_copyin, kevent_copyout, tsp);
952 kqueue_register(struct kqueue *kq, struct kevent *kev)
954 struct lwkt_token *tok;
955 struct filedesc *fdp = kq->kq_fdp;
956 struct filterops *fops;
957 struct file *fp = NULL;
958 struct knote *kn = NULL;
961 if (kev->filter < 0) {
962 if (kev->filter + EVFILT_SYSCOUNT < 0)
964 fops = sysfilt_ops[~kev->filter]; /* to 0-base index */
968 * filter attach routine is responsible for insuring that
969 * the identifier can be attached to it.
974 tok = lwkt_token_pool_lookup(kq);
976 if (fops->f_flags & FILTEROP_ISFD) {
977 /* validate descriptor */
978 fp = holdfp(fdp, kev->ident, -1);
983 lwkt_getpooltoken(&fp->f_klist);
985 SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
986 if (kn->kn_kq == kq &&
987 kn->kn_filter == kev->filter &&
988 kn->kn_id == kev->ident) {
989 if (knote_acquire(kn) == 0)
994 lwkt_relpooltoken(&fp->f_klist);
996 if (kq->kq_knhashmask) {
999 list = &kq->kq_knhash[
1000 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1001 lwkt_getpooltoken(list);
1003 SLIST_FOREACH(kn, list, kn_link) {
1004 if (kn->kn_id == kev->ident &&
1005 kn->kn_filter == kev->filter) {
1006 if (knote_acquire(kn) == 0)
1011 lwkt_relpooltoken(list);
1016 * NOTE: At this point if kn is non-NULL we will have acquired
1017 * it and set KN_PROCESSING.
1019 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) {
1025 * kn now contains the matching knote, or NULL if no match
1027 if (kev->flags & EV_ADD) {
1039 * apply reference count to knote structure, and
1040 * do not release it at the end of this routine.
1044 kn->kn_sfflags = kev->fflags;
1045 kn->kn_sdata = kev->data;
1048 kn->kn_kevent = *kev;
1051 * KN_PROCESSING prevents the knote from getting
1052 * ripped out from under us while we are trying
1053 * to attach it, in case the attach blocks.
1055 kn->kn_status = KN_PROCESSING;
1057 if ((error = filter_attach(kn)) != 0) {
1058 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1064 * Interlock against close races which either tried
1065 * to remove our knote while we were blocked or missed
1066 * it entirely prior to our attachment. We do not
1067 * want to end up with a knote on a closed descriptor.
1069 if ((fops->f_flags & FILTEROP_ISFD) &&
1070 checkfdclosed(fdp, kev->ident, kn->kn_fp)) {
1071 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1075 * The user may change some filter values after the
1076 * initial EV_ADD, but doing so will not reset any
1077 * filter which have already been triggered.
1079 KKASSERT(kn->kn_status & KN_PROCESSING);
1080 if (fops == &user_filtops) {
1081 filt_usertouch(kn, kev, EVENT_REGISTER);
1083 kn->kn_sfflags = kev->fflags;
1084 kn->kn_sdata = kev->data;
1085 kn->kn_kevent.udata = kev->udata;
1090 * Execute the filter event to immediately activate the
1091 * knote if necessary. If reprocessing events are pending
1092 * due to blocking above we do not run the filter here
1093 * but instead let knote_release() do it. Otherwise we
1094 * might run the filter on a deleted event.
1096 if ((kn->kn_status & KN_REPROCESS) == 0) {
1097 if (filter_event(kn, 0))
1100 } else if (kev->flags & EV_DELETE) {
1102 * Delete the existing knote
1104 knote_detach_and_drop(kn);
1108 * Modify an existing event.
1110 * The user may change some filter values after the
1111 * initial EV_ADD, but doing so will not reset any
1112 * filter which have already been triggered.
1114 KKASSERT(kn->kn_status & KN_PROCESSING);
1115 if (fops == &user_filtops) {
1116 filt_usertouch(kn, kev, EVENT_REGISTER);
1118 kn->kn_sfflags = kev->fflags;
1119 kn->kn_sdata = kev->data;
1120 kn->kn_kevent.udata = kev->udata;
1124 * Execute the filter event to immediately activate the
1125 * knote if necessary. If reprocessing events are pending
1126 * due to blocking above we do not run the filter here
1127 * but instead let knote_release() do it. Otherwise we
1128 * might run the filter on a deleted event.
1130 if ((kn->kn_status & KN_REPROCESS) == 0) {
1131 if (filter_event(kn, 0))
1137 * Disablement does not deactivate a knote here.
1139 if ((kev->flags & EV_DISABLE) &&
1140 ((kn->kn_status & KN_DISABLED) == 0)) {
1141 kn->kn_status |= KN_DISABLED;
1145 * Re-enablement may have to immediately enqueue an active knote.
1147 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1148 kn->kn_status &= ~KN_DISABLED;
1149 if ((kn->kn_status & KN_ACTIVE) &&
1150 ((kn->kn_status & KN_QUEUED) == 0)) {
1156 * Handle any required reprocessing
1159 /* kn may be invalid now */
1169 * Scan the kqueue, return the number of active events placed in kevp up
1172 * Continuous mode events may get recycled, do not continue scanning past
1173 * marker unless no events have been collected.
1176 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
1177 struct knote *marker)
1179 struct knote *kn, local_marker;
1183 local_marker.kn_filter = EVFILT_MARKER;
1184 local_marker.kn_status = KN_PROCESSING;
1186 lwkt_getpooltoken(kq);
1191 TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe);
1193 kn = TAILQ_NEXT(&local_marker, kn_tqe);
1194 if (kn->kn_filter == EVFILT_MARKER) {
1195 /* Marker reached, we are done */
1199 /* Move local marker past some other threads marker */
1200 kn = TAILQ_NEXT(kn, kn_tqe);
1201 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1202 TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe);
1207 * We can't skip a knote undergoing processing, otherwise
1208 * we risk not returning it when the user process expects
1209 * it should be returned. Sleep and retry.
1211 if (knote_acquire(kn) == 0)
1215 * Remove the event for processing.
1217 * WARNING! We must leave KN_QUEUED set to prevent the
1218 * event from being KNOTE_ACTIVATE()d while
1219 * the queue state is in limbo, in case we
1222 * WARNING! We must set KN_PROCESSING to avoid races
1223 * against deletion or another thread's
1226 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1230 * We have to deal with an extremely important race against
1231 * file descriptor close()s here. The file descriptor can
1232 * disappear MPSAFE, and there is a small window of
1233 * opportunity between that and the call to knote_fdclose().
1235 * If we hit that window here while doselect or dopoll is
1236 * trying to delete a spurious event they will not be able
1237 * to match up the event against a knote and will go haywire.
1239 if ((kn->kn_fop->f_flags & FILTEROP_ISFD) &&
1240 checkfdclosed(kq->kq_fdp, kn->kn_kevent.ident, kn->kn_fp)) {
1241 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1244 if (kn->kn_status & KN_DISABLED) {
1246 * If disabled we ensure the event is not queued
1247 * but leave its active bit set. On re-enablement
1248 * the event may be immediately triggered.
1250 kn->kn_status &= ~KN_QUEUED;
1251 } else if ((kn->kn_flags & EV_ONESHOT) == 0 &&
1252 (kn->kn_status & KN_DELETING) == 0 &&
1253 filter_event(kn, 0) == 0) {
1255 * If not running in one-shot mode and the event
1256 * is no longer present we ensure it is removed
1257 * from the queue and ignore it.
1259 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1264 if (kn->kn_fop == &user_filtops)
1265 filt_usertouch(kn, kevp, EVENT_PROCESS);
1267 *kevp = kn->kn_kevent;
1272 if (kn->kn_flags & EV_ONESHOT) {
1273 kn->kn_status &= ~KN_QUEUED;
1274 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1276 if (kn->kn_flags & EV_CLEAR) {
1279 kn->kn_status &= ~(KN_QUEUED |
1282 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1289 * Handle any post-processing states
1293 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1295 lwkt_relpooltoken(kq);
1301 * This could be expanded to call kqueue_scan, if desired.
1306 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1315 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1324 kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
1325 struct ucred *cred, struct sysmsg *msg)
1327 struct lwkt_token *tok;
1331 kq = (struct kqueue *)fp->f_data;
1332 tok = lwkt_token_pool_lookup(kq);
1338 kq->kq_state |= KQ_ASYNC;
1340 kq->kq_state &= ~KQ_ASYNC;
1344 error = fsetown(*(int *)data, &kq->kq_sigio);
1358 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred)
1360 struct kqueue *kq = (struct kqueue *)fp->f_data;
1362 bzero((void *)st, sizeof(*st));
1363 st->st_size = kq->kq_count;
1364 st->st_blksize = sizeof(struct kevent);
1365 st->st_mode = S_IFIFO;
1373 kqueue_close(struct file *fp)
1375 struct kqueue *kq = (struct kqueue *)fp->f_data;
1377 kqueue_terminate(kq);
1380 funsetown(&kq->kq_sigio);
1382 kfree(kq, M_KQUEUE);
1387 kqueue_wakeup(struct kqueue *kq)
1389 if (kq->kq_state & KQ_SLEEP) {
1390 kq->kq_state &= ~KQ_SLEEP;
1393 KNOTE(&kq->kq_kqinfo.ki_note, 0);
1397 * Calls filterops f_attach function, acquiring mplock if filter is not
1398 * marked as FILTEROP_MPSAFE.
1400 * Caller must be holding the related kq token
1403 filter_attach(struct knote *kn)
1407 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1408 ret = kn->kn_fop->f_attach(kn);
1411 ret = kn->kn_fop->f_attach(kn);
1418 * Detach the knote and drop it, destroying the knote.
1420 * Calls filterops f_detach function, acquiring mplock if filter is not
1421 * marked as FILTEROP_MPSAFE.
1423 * Caller must be holding the related kq token
1426 knote_detach_and_drop(struct knote *kn)
1428 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1429 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1430 kn->kn_fop->f_detach(kn);
1433 kn->kn_fop->f_detach(kn);
1440 * Calls filterops f_event function, acquiring mplock if filter is not
1441 * marked as FILTEROP_MPSAFE.
1443 * If the knote is in the middle of being created or deleted we cannot
1444 * safely call the filter op.
1446 * Caller must be holding the related kq token
1449 filter_event(struct knote *kn, long hint)
1453 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1454 ret = kn->kn_fop->f_event(kn, hint);
1457 ret = kn->kn_fop->f_event(kn, hint);
1464 * Walk down a list of knotes, activating them if their event has triggered.
1466 * If we encounter any knotes which are undergoing processing we just mark
1467 * them for reprocessing and do not try to [re]activate the knote. However,
1468 * if a hint is being passed we have to wait and that makes things a bit
1472 knote(struct klist *list, long hint)
1476 struct knote *kntmp;
1478 lwkt_getpooltoken(list);
1480 SLIST_FOREACH(kn, list, kn_next) {
1482 lwkt_getpooltoken(kq);
1484 /* temporary verification hack */
1485 SLIST_FOREACH(kntmp, list, kn_next) {
1489 if (kn != kntmp || kn->kn_kq != kq) {
1490 lwkt_relpooltoken(kq);
1494 if (kn->kn_status & KN_PROCESSING) {
1496 * Someone else is processing the knote, ask the
1497 * other thread to reprocess it and don't mess
1498 * with it otherwise.
1501 kn->kn_status |= KN_REPROCESS;
1502 lwkt_relpooltoken(kq);
1507 * If the hint is non-zero we have to wait or risk
1508 * losing the state the caller is trying to update.
1510 * XXX This is a real problem, certain process
1511 * and signal filters will bump kn_data for
1512 * already-processed notes more than once if
1513 * we restart the list scan. FIXME.
1515 kn->kn_status |= KN_WAITING | KN_REPROCESS;
1516 tsleep(kn, 0, "knotec", hz);
1517 lwkt_relpooltoken(kq);
1522 * Become the reprocessing master ourselves.
1524 * If hint is non-zero running the event is mandatory
1525 * when not deleting so do it whether reprocessing is
1528 kn->kn_status |= KN_PROCESSING;
1529 if ((kn->kn_status & KN_DELETING) == 0) {
1530 if (filter_event(kn, hint))
1533 if (knote_release(kn)) {
1534 lwkt_relpooltoken(kq);
1537 lwkt_relpooltoken(kq);
1539 lwkt_relpooltoken(list);
1543 * Insert knote at head of klist.
1545 * This function may only be called via a filter function and thus
1546 * kq_token should already be held and marked for processing.
1549 knote_insert(struct klist *klist, struct knote *kn)
1551 lwkt_getpooltoken(klist);
1552 KKASSERT(kn->kn_status & KN_PROCESSING);
1553 SLIST_INSERT_HEAD(klist, kn, kn_next);
1554 lwkt_relpooltoken(klist);
1558 * Remove knote from a klist
1560 * This function may only be called via a filter function and thus
1561 * kq_token should already be held and marked for processing.
1564 knote_remove(struct klist *klist, struct knote *kn)
1566 lwkt_getpooltoken(klist);
1567 KKASSERT(kn->kn_status & KN_PROCESSING);
1568 SLIST_REMOVE(klist, kn, knote, kn_next);
1569 lwkt_relpooltoken(klist);
1574 * Remove all knotes from a specified klist
1576 * Only called from aio.
1579 knote_empty(struct klist *list)
1583 lwkt_gettoken(&kq_token);
1584 while ((kn = SLIST_FIRST(list)) != NULL) {
1585 if (knote_acquire(kn))
1586 knote_detach_and_drop(kn);
1588 lwkt_reltoken(&kq_token);
1593 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst,
1594 struct filterops *ops, void *hook)
1599 lwkt_getpooltoken(&src->ki_note);
1600 lwkt_getpooltoken(&dst->ki_note);
1601 while ((kn = SLIST_FIRST(&src->ki_note)) != NULL) {
1603 lwkt_getpooltoken(kq);
1604 if (SLIST_FIRST(&src->ki_note) != kn || kn->kn_kq != kq) {
1605 lwkt_relpooltoken(kq);
1608 if (knote_acquire(kn)) {
1609 knote_remove(&src->ki_note, kn);
1612 knote_insert(&dst->ki_note, kn);
1614 /* kn may be invalid now */
1616 lwkt_relpooltoken(kq);
1618 lwkt_relpooltoken(&dst->ki_note);
1619 lwkt_relpooltoken(&src->ki_note);
1623 * Remove all knotes referencing a specified fd
1626 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd)
1630 struct knote *kntmp;
1632 lwkt_getpooltoken(&fp->f_klist);
1634 SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
1635 if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) {
1637 lwkt_getpooltoken(kq);
1639 /* temporary verification hack */
1640 SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) {
1644 if (kn != kntmp || kn->kn_kq->kq_fdp != fdp ||
1645 kn->kn_id != fd || kn->kn_kq != kq) {
1646 lwkt_relpooltoken(kq);
1649 if (knote_acquire(kn))
1650 knote_detach_and_drop(kn);
1651 lwkt_relpooltoken(kq);
1655 lwkt_relpooltoken(&fp->f_klist);
1659 * Low level attach function.
1661 * The knote should already be marked for processing.
1662 * Caller must hold the related kq token.
1665 knote_attach(struct knote *kn)
1668 struct kqueue *kq = kn->kn_kq;
1670 if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1671 KKASSERT(kn->kn_fp);
1672 list = &kn->kn_fp->f_klist;
1674 if (kq->kq_knhashmask == 0)
1675 kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1676 &kq->kq_knhashmask);
1677 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1679 lwkt_getpooltoken(list);
1680 SLIST_INSERT_HEAD(list, kn, kn_link);
1681 TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink);
1682 lwkt_relpooltoken(list);
1686 * Low level drop function.
1688 * The knote should already be marked for processing.
1689 * Caller must hold the related kq token.
1692 knote_drop(struct knote *kn)
1699 if (kn->kn_fop->f_flags & FILTEROP_ISFD)
1700 list = &kn->kn_fp->f_klist;
1702 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1704 lwkt_getpooltoken(list);
1705 SLIST_REMOVE(list, kn, knote, kn_link);
1706 TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink);
1707 if (kn->kn_status & KN_QUEUED)
1709 if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1714 lwkt_relpooltoken(list);
1718 * Low level enqueue function.
1720 * The knote should already be marked for processing.
1721 * Caller must be holding the kq token
1724 knote_enqueue(struct knote *kn)
1726 struct kqueue *kq = kn->kn_kq;
1728 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
1729 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1730 kn->kn_status |= KN_QUEUED;
1734 * Send SIGIO on request (typically set up as a mailbox signal)
1736 if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1)
1737 pgsigio(kq->kq_sigio, SIGIO, 0);
1743 * Low level dequeue function.
1745 * The knote should already be marked for processing.
1746 * Caller must be holding the kq token
1749 knote_dequeue(struct knote *kn)
1751 struct kqueue *kq = kn->kn_kq;
1753 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
1754 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1755 kn->kn_status &= ~KN_QUEUED;
1759 static struct knote *
1762 return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK);
1766 knote_free(struct knote *kn)
1768 kfree(kn, M_KQUEUE);