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 { FILTEROP_MPSAFE, filt_timerattach, filt_timerdetach, filt_timer };
138 static struct filterops user_filtops =
139 { FILTEROP_MPSAFE, 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 loops for kqueue scan");
148 static int kq_wakeup_one = 1;
149 SYSCTL_INT(_kern, OID_AUTO, kq_wakeup_one, CTLFLAG_RW,
150 &kq_wakeup_one, 0, "Wakeup only one kqueue scanner");
152 #define KNOTE_ACTIVATE(kn) do { \
153 kn->kn_status |= KN_ACTIVE; \
154 if ((kn->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \
158 #define KN_HASHSIZE 64 /* XXX should be tunable */
159 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask))
161 extern struct filterops aio_filtops;
162 extern struct filterops sig_filtops;
165 * Table for for all system-defined filters.
167 static struct filterops *sysfilt_ops[] = {
168 &file_filtops, /* EVFILT_READ */
169 &file_filtops, /* EVFILT_WRITE */
170 &aio_filtops, /* EVFILT_AIO */
171 &file_filtops, /* EVFILT_VNODE */
172 &proc_filtops, /* EVFILT_PROC */
173 &sig_filtops, /* EVFILT_SIGNAL */
174 &timer_filtops, /* EVFILT_TIMER */
175 &file_filtops, /* EVFILT_EXCEPT */
176 &user_filtops, /* EVFILT_USER */
180 filt_fileattach(struct knote *kn)
182 return (fo_kqfilter(kn->kn_fp, kn));
189 kqueue_kqfilter(struct file *fp, struct knote *kn)
191 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
193 if (kn->kn_filter != EVFILT_READ)
196 kn->kn_fop = &kqread_filtops;
197 knote_insert(&kq->kq_kqinfo.ki_note, kn);
202 filt_kqdetach(struct knote *kn)
204 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
206 knote_remove(&kq->kq_kqinfo.ki_note, kn);
211 filt_kqueue(struct knote *kn, long hint)
213 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
215 kn->kn_data = kq->kq_count;
216 return (kn->kn_data > 0);
220 filt_procattach(struct knote *kn)
226 p = pfind(kn->kn_id);
227 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
228 p = zpfind(kn->kn_id);
234 if (!PRISON_CHECK(curthread->td_ucred, p->p_ucred)) {
240 lwkt_gettoken(&p->p_token);
241 kn->kn_ptr.p_proc = p;
242 kn->kn_flags |= EV_CLEAR; /* automatically set */
245 * internal flag indicating registration done by kernel
247 if (kn->kn_flags & EV_FLAG1) {
248 kn->kn_data = kn->kn_sdata; /* ppid */
249 kn->kn_fflags = NOTE_CHILD;
250 kn->kn_flags &= ~EV_FLAG1;
253 knote_insert(&p->p_klist, kn);
256 * Immediately activate any exit notes if the target process is a
257 * zombie. This is necessary to handle the case where the target
258 * process, e.g. a child, dies before the kevent is negistered.
260 if (immediate && filt_proc(kn, NOTE_EXIT))
262 lwkt_reltoken(&p->p_token);
269 * The knote may be attached to a different process, which may exit,
270 * leaving nothing for the knote to be attached to. So when the process
271 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
272 * it will be deleted when read out. However, as part of the knote deletion,
273 * this routine is called, so a check is needed to avoid actually performing
274 * a detach, because the original process does not exist any more.
277 filt_procdetach(struct knote *kn)
281 if (kn->kn_status & KN_DETACHED)
283 p = kn->kn_ptr.p_proc;
284 knote_remove(&p->p_klist, kn);
288 filt_proc(struct knote *kn, long hint)
293 * mask off extra data
295 event = (u_int)hint & NOTE_PCTRLMASK;
298 * if the user is interested in this event, record it.
300 if (kn->kn_sfflags & event)
301 kn->kn_fflags |= event;
304 * Process is gone, so flag the event as finished. Detach the
305 * knote from the process now because the process will be poof,
308 if (event == NOTE_EXIT) {
309 struct proc *p = kn->kn_ptr.p_proc;
310 if ((kn->kn_status & KN_DETACHED) == 0) {
312 knote_remove(&p->p_klist, kn);
313 kn->kn_status |= KN_DETACHED;
314 kn->kn_data = p->p_xstat;
315 kn->kn_ptr.p_proc = NULL;
318 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
323 * process forked, and user wants to track the new process,
324 * so attach a new knote to it, and immediately report an
325 * event with the parent's pid.
327 if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) {
332 * register knote with new process.
334 kev.ident = hint & NOTE_PDATAMASK; /* pid */
335 kev.filter = kn->kn_filter;
336 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
337 kev.fflags = kn->kn_sfflags;
338 kev.data = kn->kn_id; /* parent */
339 kev.udata = kn->kn_kevent.udata; /* preserve udata */
340 error = kqueue_register(kn->kn_kq, &kev);
342 kn->kn_fflags |= NOTE_TRACKERR;
345 return (kn->kn_fflags != 0);
349 * The callout interlocks with callout_terminate() but can still
350 * race a deletion so if KN_DELETING is set we just don't touch
354 filt_timerexpire(void *knx)
356 struct lwkt_token *tok;
357 struct knote *kn = knx;
358 struct callout *calloutp;
362 tok = lwkt_token_pool_lookup(kn->kn_kq);
364 if ((kn->kn_status & KN_DELETING) == 0) {
368 if ((kn->kn_flags & EV_ONESHOT) == 0) {
369 tv.tv_sec = kn->kn_sdata / 1000;
370 tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
371 tticks = tvtohz_high(&tv);
372 calloutp = (struct callout *)kn->kn_hook;
373 callout_reset(calloutp, tticks, filt_timerexpire, kn);
380 * data contains amount of time to sleep, in milliseconds
383 filt_timerattach(struct knote *kn)
385 struct callout *calloutp;
390 prev_ncallouts = atomic_fetchadd_int(&kq_ncallouts, 1);
391 if (prev_ncallouts >= kq_calloutmax) {
392 atomic_subtract_int(&kq_ncallouts, 1);
397 tv.tv_sec = kn->kn_sdata / 1000;
398 tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
399 tticks = tvtohz_high(&tv);
401 kn->kn_flags |= EV_CLEAR; /* automatically set */
402 calloutp = kmalloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
403 callout_init_mp(calloutp);
404 kn->kn_hook = (caddr_t)calloutp;
405 callout_reset(calloutp, tticks, filt_timerexpire, kn);
411 * This function is called with the knote flagged locked but it is
412 * still possible to race a callout event due to the callback blocking.
413 * We must call callout_terminate() instead of callout_stop() to deal
417 filt_timerdetach(struct knote *kn)
419 struct callout *calloutp;
421 calloutp = (struct callout *)kn->kn_hook;
422 callout_terminate(calloutp);
423 kfree(calloutp, M_KQUEUE);
424 atomic_subtract_int(&kq_ncallouts, 1);
428 filt_timer(struct knote *kn, long hint)
431 return (kn->kn_data != 0);
438 filt_userattach(struct knote *kn)
441 if (kn->kn_fflags & NOTE_TRIGGER)
442 kn->kn_ptr.hookid = 1;
444 kn->kn_ptr.hookid = 0;
449 filt_userdetach(struct knote *kn)
455 filt_user(struct knote *kn, long hint)
457 return (kn->kn_ptr.hookid);
461 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
467 if (kev->fflags & NOTE_TRIGGER)
468 kn->kn_ptr.hookid = 1;
470 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
471 kev->fflags &= NOTE_FFLAGSMASK;
477 kn->kn_sfflags &= kev->fflags;
481 kn->kn_sfflags |= kev->fflags;
485 kn->kn_sfflags = kev->fflags;
489 /* XXX Return error? */
492 kn->kn_sdata = kev->data;
495 * This is not the correct use of EV_CLEAR in an event
496 * modification, it should have been passed as a NOTE instead.
497 * But we need to maintain compatibility with Apple & FreeBSD.
499 * Note however that EV_CLEAR can still be used when doing
500 * the initial registration of the event and works as expected
501 * (clears the event on reception).
503 if (kev->flags & EV_CLEAR) {
504 kn->kn_ptr.hookid = 0;
511 *kev = kn->kn_kevent;
512 kev->fflags = kn->kn_sfflags;
513 kev->data = kn->kn_sdata;
514 if (kn->kn_flags & EV_CLEAR) {
515 kn->kn_ptr.hookid = 0;
516 /* kn_data, kn_fflags handled by parent */
521 panic("filt_usertouch() - invalid type (%ld)", type);
527 * Acquire a knote, return non-zero on success, 0 on failure.
529 * If we cannot acquire the knote we sleep and return 0. The knote
530 * may be stale on return in this case and the caller must restart
531 * whatever loop they are in.
533 * Related kq token must be held.
536 knote_acquire(struct knote *kn)
538 if (kn->kn_status & KN_PROCESSING) {
539 kn->kn_status |= KN_WAITING | KN_REPROCESS;
540 tsleep(kn, 0, "kqepts", hz);
541 /* knote may be stale now */
544 kn->kn_status |= KN_PROCESSING;
549 * Release an acquired knote, clearing KN_PROCESSING and handling any
550 * KN_REPROCESS events.
552 * Caller must be holding the related kq token
554 * Non-zero is returned if the knote is destroyed or detached.
557 knote_release(struct knote *kn)
559 while (kn->kn_status & KN_REPROCESS) {
560 kn->kn_status &= ~KN_REPROCESS;
561 if (kn->kn_status & KN_WAITING) {
562 kn->kn_status &= ~KN_WAITING;
565 if (kn->kn_status & KN_DELETING) {
566 knote_detach_and_drop(kn);
570 if (filter_event(kn, 0))
573 kn->kn_status &= ~KN_PROCESSING;
574 if (kn->kn_status & KN_DETACHED)
581 * Initialize a kqueue.
583 * NOTE: The lwp/proc code initializes a kqueue for select/poll ops.
588 kqueue_init(struct kqueue *kq, struct filedesc *fdp)
590 TAILQ_INIT(&kq->kq_knpend);
591 TAILQ_INIT(&kq->kq_knlist);
594 SLIST_INIT(&kq->kq_kqinfo.ki_note);
598 * Terminate a kqueue. Freeing the actual kq itself is left up to the
599 * caller (it might be embedded in a lwp so we don't do it here).
601 * The kq's knlist must be completely eradicated so block on any
605 kqueue_terminate(struct kqueue *kq)
607 struct lwkt_token *tok;
610 tok = lwkt_token_pool_lookup(kq);
612 while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) {
613 if (knote_acquire(kn))
614 knote_detach_and_drop(kn);
619 hashdestroy(kq->kq_knhash, M_KQUEUE, kq->kq_knhashmask);
620 kq->kq_knhash = NULL;
621 kq->kq_knhashmask = 0;
629 sys_kqueue(struct kqueue_args *uap)
631 struct thread *td = curthread;
636 error = falloc(td->td_lwp, &fp, &fd);
639 fp->f_flag = FREAD | FWRITE;
640 fp->f_type = DTYPE_KQUEUE;
641 fp->f_ops = &kqueueops;
643 kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO);
644 kqueue_init(kq, td->td_proc->p_fd);
647 fsetfd(kq->kq_fdp, fp, fd);
648 uap->sysmsg_result = fd;
654 * Copy 'count' items into the destination list pointed to by uap->eventlist.
657 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res)
659 struct kevent_copyin_args *kap;
662 kap = (struct kevent_copyin_args *)arg;
664 error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp));
666 kap->ka->eventlist += count;
676 * Copy at most 'max' items from the list pointed to by kap->changelist,
677 * return number of items in 'events'.
680 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events)
682 struct kevent_copyin_args *kap;
685 kap = (struct kevent_copyin_args *)arg;
687 count = min(kap->ka->nchanges - kap->pchanges, max);
688 error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp);
690 kap->ka->changelist += count;
691 kap->pchanges += count;
702 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap,
703 k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn,
704 struct timespec *tsp_in)
707 struct timespec *tsp, ats;
708 int i, n, total, error, nerrors = 0;
710 int limit = kq_checkloop;
711 struct kevent kev[KQ_NEVENTS];
713 struct lwkt_token *tok;
715 if (tsp_in == NULL || tsp_in->tv_sec || tsp_in->tv_nsec)
716 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC);
723 error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n);
728 for (i = 0; i < n; i++) {
730 kevp->flags &= ~EV_SYSFLAGS;
731 error = kqueue_register(kq, kevp);
734 * If a registration returns an error we
735 * immediately post the error. The kevent()
736 * call itself will fail with the error if
737 * no space is available for posting.
739 * Such errors normally bypass the timeout/blocking
740 * code. However, if the copyoutfn function refuses
741 * to post the error (see sys_poll()), then we
744 if (error || (kevp->flags & EV_RECEIPT)) {
745 kevp->flags = EV_ERROR;
748 kevent_copyoutfn(uap, kevp, 1, res);
751 } else if (lres != *res) {
762 * Acquire/wait for events - setup timeout
765 if (tsp->tv_sec || tsp->tv_nsec) {
767 timespecadd(tsp, &ats); /* tsp = target time */
774 * Collect as many events as we can. Sleeping on successive
775 * loops is disabled if copyoutfn has incremented (*res).
777 * The loop stops if an error occurs, all events have been
778 * scanned (the marker has been reached), or fewer than the
779 * maximum number of events is found.
781 * The copyoutfn function does not have to increment (*res) in
782 * order for the loop to continue.
784 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents.
788 marker.kn_filter = EVFILT_MARKER;
789 marker.kn_status = KN_PROCESSING;
790 tok = lwkt_token_pool_lookup(kq);
792 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
794 while ((n = nevents - total) > 0) {
799 * If no events are pending sleep until timeout (if any)
800 * or an event occurs.
802 * After the sleep completes the marker is moved to the
803 * end of the list, making any received events available
806 if (kq->kq_count == 0 && *res == 0) {
811 } else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) {
815 struct timespec atx = *tsp;
818 timespecsub(&atx, &ats);
819 if (atx.tv_sec < 0) {
823 timeout = atx.tv_sec > 24 * 60 * 60 ?
830 if (kq->kq_count == 0) {
831 kq->kq_state |= KQ_SLEEP;
832 error = tsleep(kq, PCATCH, "kqread", timeout);
834 /* don't restart after signals... */
835 if (error == ERESTART)
842 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
843 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker,
850 * Process all received events
851 * Account for all non-spurious events in our total
853 i = kqueue_scan(kq, kev, n, &marker);
856 error = kevent_copyoutfn(uap, kev, i, res);
857 total += *res - lres;
861 if (limit && --limit == 0)
862 panic("kqueue: checkloop failed i=%d", i);
865 * Normally when fewer events are returned than requested
866 * we can stop. However, if only spurious events were
867 * collected the copyout will not bump (*res) and we have
874 * Deal with an edge case where spurious events can cause
875 * a loop to occur without moving the marker. This can
876 * prevent kqueue_scan() from picking up new events which
877 * race us. We must be sure to move the marker for this
880 * NOTE: We do not want to move the marker if events
881 * were scanned because normal kqueue operations
882 * may reactivate events. Moving the marker in
883 * that case could result in duplicates for the
888 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
889 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
894 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
897 /* Timeouts do not return EWOULDBLOCK. */
898 if (error == EWOULDBLOCK)
907 sys_kevent(struct kevent_args *uap)
909 struct thread *td = curthread;
910 struct proc *p = td->td_proc;
911 struct timespec ts, *tsp;
913 struct file *fp = NULL;
914 struct kevent_copyin_args *kap, ka;
918 error = copyin(uap->timeout, &ts, sizeof(ts));
925 fp = holdfp(p->p_fd, uap->fd, -1);
928 if (fp->f_type != DTYPE_KQUEUE) {
933 kq = (struct kqueue *)fp->f_data;
939 error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap,
940 kevent_copyin, kevent_copyout, tsp);
948 kqueue_register(struct kqueue *kq, struct kevent *kev)
950 struct lwkt_token *tok;
951 struct filedesc *fdp = kq->kq_fdp;
952 struct filterops *fops;
953 struct file *fp = NULL;
954 struct knote *kn = NULL;
957 if (kev->filter < 0) {
958 if (kev->filter + EVFILT_SYSCOUNT < 0)
960 fops = sysfilt_ops[~kev->filter]; /* to 0-base index */
964 * filter attach routine is responsible for insuring that
965 * the identifier can be attached to it.
970 tok = lwkt_token_pool_lookup(kq);
972 if (fops->f_flags & FILTEROP_ISFD) {
973 /* validate descriptor */
974 fp = holdfp(fdp, kev->ident, -1);
979 lwkt_getpooltoken(&fp->f_klist);
981 SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
982 if (kn->kn_kq == kq &&
983 kn->kn_filter == kev->filter &&
984 kn->kn_id == kev->ident) {
985 if (knote_acquire(kn) == 0)
990 lwkt_relpooltoken(&fp->f_klist);
992 if (kq->kq_knhashmask) {
995 list = &kq->kq_knhash[
996 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
997 lwkt_getpooltoken(list);
999 SLIST_FOREACH(kn, list, kn_link) {
1000 if (kn->kn_id == kev->ident &&
1001 kn->kn_filter == kev->filter) {
1002 if (knote_acquire(kn) == 0)
1007 lwkt_relpooltoken(list);
1012 * NOTE: At this point if kn is non-NULL we will have acquired
1013 * it and set KN_PROCESSING.
1015 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) {
1021 * kn now contains the matching knote, or NULL if no match
1023 if (kev->flags & EV_ADD) {
1031 * apply reference count to knote structure, and
1032 * do not release it at the end of this routine.
1036 kn->kn_sfflags = kev->fflags;
1037 kn->kn_sdata = kev->data;
1040 kn->kn_kevent = *kev;
1043 * KN_PROCESSING prevents the knote from getting
1044 * ripped out from under us while we are trying
1045 * to attach it, in case the attach blocks.
1047 kn->kn_status = KN_PROCESSING;
1049 if ((error = filter_attach(kn)) != 0) {
1050 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1056 * Interlock against close races which either tried
1057 * to remove our knote while we were blocked or missed
1058 * it entirely prior to our attachment. We do not
1059 * want to end up with a knote on a closed descriptor.
1061 if ((fops->f_flags & FILTEROP_ISFD) &&
1062 checkfdclosed(fdp, kev->ident, kn->kn_fp)) {
1063 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1067 * The user may change some filter values after the
1068 * initial EV_ADD, but doing so will not reset any
1069 * filter which have already been triggered.
1071 KKASSERT(kn->kn_status & KN_PROCESSING);
1072 if (fops == &user_filtops) {
1073 filt_usertouch(kn, kev, EVENT_REGISTER);
1075 kn->kn_sfflags = kev->fflags;
1076 kn->kn_sdata = kev->data;
1077 kn->kn_kevent.udata = kev->udata;
1082 * Execute the filter event to immediately activate the
1083 * knote if necessary. If reprocessing events are pending
1084 * due to blocking above we do not run the filter here
1085 * but instead let knote_release() do it. Otherwise we
1086 * might run the filter on a deleted event.
1088 if ((kn->kn_status & KN_REPROCESS) == 0) {
1089 if (filter_event(kn, 0))
1092 } else if (kev->flags & EV_DELETE) {
1094 * Delete the existing knote
1096 knote_detach_and_drop(kn);
1100 * Modify an existing event.
1102 * The user may change some filter values after the
1103 * initial EV_ADD, but doing so will not reset any
1104 * filter which have already been triggered.
1106 KKASSERT(kn->kn_status & KN_PROCESSING);
1107 if (fops == &user_filtops) {
1108 filt_usertouch(kn, kev, EVENT_REGISTER);
1110 kn->kn_sfflags = kev->fflags;
1111 kn->kn_sdata = kev->data;
1112 kn->kn_kevent.udata = kev->udata;
1116 * Execute the filter event to immediately activate the
1117 * knote if necessary. If reprocessing events are pending
1118 * due to blocking above we do not run the filter here
1119 * but instead let knote_release() do it. Otherwise we
1120 * might run the filter on a deleted event.
1122 if ((kn->kn_status & KN_REPROCESS) == 0) {
1123 if (filter_event(kn, 0))
1129 * Disablement does not deactivate a knote here.
1131 if ((kev->flags & EV_DISABLE) &&
1132 ((kn->kn_status & KN_DISABLED) == 0)) {
1133 kn->kn_status |= KN_DISABLED;
1137 * Re-enablement may have to immediately enqueue an active knote.
1139 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1140 kn->kn_status &= ~KN_DISABLED;
1141 if ((kn->kn_status & KN_ACTIVE) &&
1142 ((kn->kn_status & KN_QUEUED) == 0)) {
1148 * Handle any required reprocessing
1151 /* kn may be invalid now */
1161 * Scan the kqueue, return the number of active events placed in kevp up
1164 * Continuous mode events may get recycled, do not continue scanning past
1165 * marker unless no events have been collected.
1168 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
1169 struct knote *marker)
1171 struct knote *kn, local_marker;
1175 local_marker.kn_filter = EVFILT_MARKER;
1176 local_marker.kn_status = KN_PROCESSING;
1178 lwkt_getpooltoken(kq);
1183 TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe);
1185 kn = TAILQ_NEXT(&local_marker, kn_tqe);
1186 if (kn->kn_filter == EVFILT_MARKER) {
1187 /* Marker reached, we are done */
1191 /* Move local marker past some other threads marker */
1192 kn = TAILQ_NEXT(kn, kn_tqe);
1193 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1194 TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe);
1199 * We can't skip a knote undergoing processing, otherwise
1200 * we risk not returning it when the user process expects
1201 * it should be returned. Sleep and retry.
1203 if (knote_acquire(kn) == 0)
1207 * Remove the event for processing.
1209 * WARNING! We must leave KN_QUEUED set to prevent the
1210 * event from being KNOTE_ACTIVATE()d while
1211 * the queue state is in limbo, in case we
1214 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1218 * We have to deal with an extremely important race against
1219 * file descriptor close()s here. The file descriptor can
1220 * disappear MPSAFE, and there is a small window of
1221 * opportunity between that and the call to knote_fdclose().
1223 * If we hit that window here while doselect or dopoll is
1224 * trying to delete a spurious event they will not be able
1225 * to match up the event against a knote and will go haywire.
1227 if ((kn->kn_fop->f_flags & FILTEROP_ISFD) &&
1228 checkfdclosed(kq->kq_fdp, kn->kn_kevent.ident, kn->kn_fp)) {
1229 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1232 if (kn->kn_status & KN_DISABLED) {
1234 * If disabled we ensure the event is not queued
1235 * but leave its active bit set. On re-enablement
1236 * the event may be immediately triggered.
1238 kn->kn_status &= ~KN_QUEUED;
1239 } else if ((kn->kn_flags & EV_ONESHOT) == 0 &&
1240 (kn->kn_status & KN_DELETING) == 0 &&
1241 filter_event(kn, 0) == 0) {
1243 * If not running in one-shot mode and the event
1244 * is no longer present we ensure it is removed
1245 * from the queue and ignore it.
1247 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1252 if (kn->kn_fop == &user_filtops)
1253 filt_usertouch(kn, kevp, EVENT_PROCESS);
1255 *kevp = kn->kn_kevent;
1260 if (kn->kn_flags & EV_ONESHOT) {
1261 kn->kn_status &= ~KN_QUEUED;
1262 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1264 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1265 if (kn->kn_flags & EV_CLEAR) {
1269 if (kn->kn_flags & EV_DISPATCH) {
1270 kn->kn_status |= KN_DISABLED;
1272 kn->kn_status &= ~(KN_QUEUED |
1275 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1282 * Handle any post-processing states
1286 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1288 lwkt_relpooltoken(kq);
1294 * This could be expanded to call kqueue_scan, if desired.
1299 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1308 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1317 kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
1318 struct ucred *cred, struct sysmsg *msg)
1320 struct lwkt_token *tok;
1324 kq = (struct kqueue *)fp->f_data;
1325 tok = lwkt_token_pool_lookup(kq);
1331 kq->kq_state |= KQ_ASYNC;
1333 kq->kq_state &= ~KQ_ASYNC;
1337 error = fsetown(*(int *)data, &kq->kq_sigio);
1351 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred)
1353 struct kqueue *kq = (struct kqueue *)fp->f_data;
1355 bzero((void *)st, sizeof(*st));
1356 st->st_size = kq->kq_count;
1357 st->st_blksize = sizeof(struct kevent);
1358 st->st_mode = S_IFIFO;
1366 kqueue_close(struct file *fp)
1368 struct kqueue *kq = (struct kqueue *)fp->f_data;
1370 kqueue_terminate(kq);
1373 funsetown(&kq->kq_sigio);
1375 kfree(kq, M_KQUEUE);
1380 kqueue_wakeup(struct kqueue *kq)
1382 if (kq->kq_state & KQ_SLEEP) {
1383 kq->kq_state &= ~KQ_SLEEP;
1389 KNOTE(&kq->kq_kqinfo.ki_note, 0);
1393 * Calls filterops f_attach function, acquiring mplock if filter is not
1394 * marked as FILTEROP_MPSAFE.
1396 * Caller must be holding the related kq token
1399 filter_attach(struct knote *kn)
1403 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1404 ret = kn->kn_fop->f_attach(kn);
1407 ret = kn->kn_fop->f_attach(kn);
1414 * Detach the knote and drop it, destroying the knote.
1416 * Calls filterops f_detach function, acquiring mplock if filter is not
1417 * marked as FILTEROP_MPSAFE.
1419 * Caller must be holding the related kq token
1422 knote_detach_and_drop(struct knote *kn)
1424 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1425 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1426 kn->kn_fop->f_detach(kn);
1429 kn->kn_fop->f_detach(kn);
1436 * Calls filterops f_event function, acquiring mplock if filter is not
1437 * marked as FILTEROP_MPSAFE.
1439 * If the knote is in the middle of being created or deleted we cannot
1440 * safely call the filter op.
1442 * Caller must be holding the related kq token
1445 filter_event(struct knote *kn, long hint)
1449 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1450 ret = kn->kn_fop->f_event(kn, hint);
1453 ret = kn->kn_fop->f_event(kn, hint);
1460 * Walk down a list of knotes, activating them if their event has triggered.
1462 * If we encounter any knotes which are undergoing processing we just mark
1463 * them for reprocessing and do not try to [re]activate the knote. However,
1464 * if a hint is being passed we have to wait and that makes things a bit
1468 knote(struct klist *list, long hint)
1472 struct knote *kntmp;
1474 lwkt_getpooltoken(list);
1476 SLIST_FOREACH(kn, list, kn_next) {
1478 lwkt_getpooltoken(kq);
1480 /* temporary verification hack */
1481 SLIST_FOREACH(kntmp, list, kn_next) {
1485 if (kn != kntmp || kn->kn_kq != kq) {
1486 lwkt_relpooltoken(kq);
1490 if (kn->kn_status & KN_PROCESSING) {
1492 * Someone else is processing the knote, ask the
1493 * other thread to reprocess it and don't mess
1494 * with it otherwise.
1497 kn->kn_status |= KN_REPROCESS;
1498 lwkt_relpooltoken(kq);
1503 * If the hint is non-zero we have to wait or risk
1504 * losing the state the caller is trying to update.
1506 * XXX This is a real problem, certain process
1507 * and signal filters will bump kn_data for
1508 * already-processed notes more than once if
1509 * we restart the list scan. FIXME.
1511 kn->kn_status |= KN_WAITING | KN_REPROCESS;
1512 tsleep(kn, 0, "knotec", hz);
1513 lwkt_relpooltoken(kq);
1518 * Become the reprocessing master ourselves.
1520 * If hint is non-zero running the event is mandatory
1521 * when not deleting so do it whether reprocessing is
1524 kn->kn_status |= KN_PROCESSING;
1525 if ((kn->kn_status & KN_DELETING) == 0) {
1526 if (filter_event(kn, hint))
1529 if (knote_release(kn)) {
1530 lwkt_relpooltoken(kq);
1533 lwkt_relpooltoken(kq);
1535 lwkt_relpooltoken(list);
1539 * Insert knote at head of klist.
1541 * This function may only be called via a filter function and thus
1542 * kq_token should already be held and marked for processing.
1545 knote_insert(struct klist *klist, struct knote *kn)
1547 lwkt_getpooltoken(klist);
1548 KKASSERT(kn->kn_status & KN_PROCESSING);
1549 SLIST_INSERT_HEAD(klist, kn, kn_next);
1550 lwkt_relpooltoken(klist);
1554 * Remove knote from a klist
1556 * This function may only be called via a filter function and thus
1557 * kq_token should already be held and marked for processing.
1560 knote_remove(struct klist *klist, struct knote *kn)
1562 lwkt_getpooltoken(klist);
1563 KKASSERT(kn->kn_status & KN_PROCESSING);
1564 SLIST_REMOVE(klist, kn, knote, kn_next);
1565 lwkt_relpooltoken(klist);
1570 * Remove all knotes from a specified klist
1572 * Only called from aio.
1575 knote_empty(struct klist *list)
1579 lwkt_gettoken(&kq_token);
1580 while ((kn = SLIST_FIRST(list)) != NULL) {
1581 if (knote_acquire(kn))
1582 knote_detach_and_drop(kn);
1584 lwkt_reltoken(&kq_token);
1589 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst,
1590 struct filterops *ops, void *hook)
1595 lwkt_getpooltoken(&src->ki_note);
1596 lwkt_getpooltoken(&dst->ki_note);
1597 while ((kn = SLIST_FIRST(&src->ki_note)) != NULL) {
1599 lwkt_getpooltoken(kq);
1600 if (SLIST_FIRST(&src->ki_note) != kn || kn->kn_kq != kq) {
1601 lwkt_relpooltoken(kq);
1604 if (knote_acquire(kn)) {
1605 knote_remove(&src->ki_note, kn);
1608 knote_insert(&dst->ki_note, kn);
1610 /* kn may be invalid now */
1612 lwkt_relpooltoken(kq);
1614 lwkt_relpooltoken(&dst->ki_note);
1615 lwkt_relpooltoken(&src->ki_note);
1619 * Remove all knotes referencing a specified fd
1622 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd)
1626 struct knote *kntmp;
1628 lwkt_getpooltoken(&fp->f_klist);
1630 SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
1631 if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) {
1633 lwkt_getpooltoken(kq);
1635 /* temporary verification hack */
1636 SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) {
1640 if (kn != kntmp || kn->kn_kq->kq_fdp != fdp ||
1641 kn->kn_id != fd || kn->kn_kq != kq) {
1642 lwkt_relpooltoken(kq);
1645 if (knote_acquire(kn))
1646 knote_detach_and_drop(kn);
1647 lwkt_relpooltoken(kq);
1651 lwkt_relpooltoken(&fp->f_klist);
1655 * Low level attach function.
1657 * The knote should already be marked for processing.
1658 * Caller must hold the related kq token.
1661 knote_attach(struct knote *kn)
1664 struct kqueue *kq = kn->kn_kq;
1666 if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1667 KKASSERT(kn->kn_fp);
1668 list = &kn->kn_fp->f_klist;
1670 if (kq->kq_knhashmask == 0)
1671 kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1672 &kq->kq_knhashmask);
1673 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1675 lwkt_getpooltoken(list);
1676 SLIST_INSERT_HEAD(list, kn, kn_link);
1677 lwkt_relpooltoken(list);
1678 TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink);
1682 * Low level drop function.
1684 * The knote should already be marked for processing.
1685 * Caller must hold the related kq token.
1688 knote_drop(struct knote *kn)
1695 if (kn->kn_fop->f_flags & FILTEROP_ISFD)
1696 list = &kn->kn_fp->f_klist;
1698 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1700 lwkt_getpooltoken(list);
1701 SLIST_REMOVE(list, kn, knote, kn_link);
1702 lwkt_relpooltoken(list);
1703 TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink);
1704 if (kn->kn_status & KN_QUEUED)
1706 if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1714 * Low level enqueue function.
1716 * The knote should already be marked for processing.
1717 * Caller must be holding the kq token
1720 knote_enqueue(struct knote *kn)
1722 struct kqueue *kq = kn->kn_kq;
1724 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
1725 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1726 kn->kn_status |= KN_QUEUED;
1730 * Send SIGIO on request (typically set up as a mailbox signal)
1732 if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1)
1733 pgsigio(kq->kq_sigio, SIGIO, 0);
1739 * Low level dequeue function.
1741 * The knote should already be marked for processing.
1742 * Caller must be holding the kq token
1745 knote_dequeue(struct knote *kn)
1747 struct kqueue *kq = kn->kn_kq;
1749 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
1750 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1751 kn->kn_status &= ~KN_QUEUED;
1755 static struct knote *
1758 return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK);
1762 knote_free(struct knote *kn)
1764 kfree(kn, M_KQUEUE);