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");
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;
387 prev_ncallouts = atomic_fetchadd_int(&kq_ncallouts, 1);
388 if (prev_ncallouts >= kq_calloutmax) {
389 atomic_subtract_int(&kq_ncallouts, 1);
394 tv.tv_sec = kn->kn_sdata / 1000;
395 tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
396 tticks = tvtohz_high(&tv);
398 kn->kn_flags |= EV_CLEAR; /* automatically set */
399 calloutp = kmalloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
400 callout_init_mp(calloutp);
401 kn->kn_hook = (caddr_t)calloutp;
402 callout_reset(calloutp, tticks, filt_timerexpire, kn);
408 * This function is called with the knote flagged locked but it is
409 * still possible to race a callout event due to the callback blocking.
410 * We must call callout_terminate() instead of callout_stop() to deal
414 filt_timerdetach(struct knote *kn)
416 struct callout *calloutp;
418 calloutp = (struct callout *)kn->kn_hook;
419 callout_terminate(calloutp);
420 kfree(calloutp, M_KQUEUE);
421 atomic_subtract_int(&kq_ncallouts, 1);
425 filt_timer(struct knote *kn, long hint)
428 return (kn->kn_data != 0);
435 filt_userattach(struct knote *kn)
438 if (kn->kn_fflags & NOTE_TRIGGER)
439 kn->kn_ptr.hookid = 1;
441 kn->kn_ptr.hookid = 0;
446 filt_userdetach(struct knote *kn)
452 filt_user(struct knote *kn, long hint)
454 return (kn->kn_ptr.hookid);
458 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
464 if (kev->fflags & NOTE_TRIGGER)
465 kn->kn_ptr.hookid = 1;
467 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
468 kev->fflags &= NOTE_FFLAGSMASK;
474 kn->kn_sfflags &= kev->fflags;
478 kn->kn_sfflags |= kev->fflags;
482 kn->kn_sfflags = kev->fflags;
486 /* XXX Return error? */
489 kn->kn_sdata = kev->data;
492 * This is not the correct use of EV_CLEAR in an event
493 * modification, it should have been passed as a NOTE instead.
494 * But we need to maintain compatibility with Apple & FreeBSD.
496 * Note however that EV_CLEAR can still be used when doing
497 * the initial registration of the event and works as expected
498 * (clears the event on reception).
500 if (kev->flags & EV_CLEAR) {
501 kn->kn_ptr.hookid = 0;
508 *kev = kn->kn_kevent;
509 kev->fflags = kn->kn_sfflags;
510 kev->data = kn->kn_sdata;
511 if (kn->kn_flags & EV_CLEAR) {
512 kn->kn_ptr.hookid = 0;
513 /* kn_data, kn_fflags handled by parent */
518 panic("filt_usertouch() - invalid type (%ld)", type);
524 * Acquire a knote, return non-zero on success, 0 on failure.
526 * If we cannot acquire the knote we sleep and return 0. The knote
527 * may be stale on return in this case and the caller must restart
528 * whatever loop they are in.
530 * Related kq token must be held.
533 knote_acquire(struct knote *kn)
535 if (kn->kn_status & KN_PROCESSING) {
536 kn->kn_status |= KN_WAITING | KN_REPROCESS;
537 tsleep(kn, 0, "kqepts", hz);
538 /* knote may be stale now */
541 kn->kn_status |= KN_PROCESSING;
546 * Release an acquired knote, clearing KN_PROCESSING and handling any
547 * KN_REPROCESS events.
549 * Caller must be holding the related kq token
551 * Non-zero is returned if the knote is destroyed or detached.
554 knote_release(struct knote *kn)
556 while (kn->kn_status & KN_REPROCESS) {
557 kn->kn_status &= ~KN_REPROCESS;
558 if (kn->kn_status & KN_WAITING) {
559 kn->kn_status &= ~KN_WAITING;
562 if (kn->kn_status & KN_DELETING) {
563 knote_detach_and_drop(kn);
567 if (filter_event(kn, 0))
570 kn->kn_status &= ~KN_PROCESSING;
571 if (kn->kn_status & KN_DETACHED)
578 * Initialize a kqueue.
580 * NOTE: The lwp/proc code initializes a kqueue for select/poll ops.
585 kqueue_init(struct kqueue *kq, struct filedesc *fdp)
587 TAILQ_INIT(&kq->kq_knpend);
588 TAILQ_INIT(&kq->kq_knlist);
591 SLIST_INIT(&kq->kq_kqinfo.ki_note);
595 * Terminate a kqueue. Freeing the actual kq itself is left up to the
596 * caller (it might be embedded in a lwp so we don't do it here).
598 * The kq's knlist must be completely eradicated so block on any
602 kqueue_terminate(struct kqueue *kq)
604 struct lwkt_token *tok;
607 tok = lwkt_token_pool_lookup(kq);
609 while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) {
610 if (knote_acquire(kn))
611 knote_detach_and_drop(kn);
616 hashdestroy(kq->kq_knhash, M_KQUEUE, kq->kq_knhashmask);
617 kq->kq_knhash = NULL;
618 kq->kq_knhashmask = 0;
626 sys_kqueue(struct kqueue_args *uap)
628 struct thread *td = curthread;
633 error = falloc(td->td_lwp, &fp, &fd);
636 fp->f_flag = FREAD | FWRITE;
637 fp->f_type = DTYPE_KQUEUE;
638 fp->f_ops = &kqueueops;
640 kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO);
641 kqueue_init(kq, td->td_proc->p_fd);
644 fsetfd(kq->kq_fdp, fp, fd);
645 uap->sysmsg_result = fd;
651 * Copy 'count' items into the destination list pointed to by uap->eventlist.
654 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res)
656 struct kevent_copyin_args *kap;
659 kap = (struct kevent_copyin_args *)arg;
661 error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp));
663 kap->ka->eventlist += count;
673 * Copy at most 'max' items from the list pointed to by kap->changelist,
674 * return number of items in 'events'.
677 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events)
679 struct kevent_copyin_args *kap;
682 kap = (struct kevent_copyin_args *)arg;
684 count = min(kap->ka->nchanges - kap->pchanges, max);
685 error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp);
687 kap->ka->changelist += count;
688 kap->pchanges += count;
699 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap,
700 k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn,
701 struct timespec *tsp_in)
704 struct timespec *tsp, ats;
705 int i, n, total, error, nerrors = 0;
707 int limit = kq_checkloop;
708 struct kevent kev[KQ_NEVENTS];
710 struct lwkt_token *tok;
712 if (tsp_in == NULL || tsp_in->tv_sec || tsp_in->tv_nsec)
713 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC);
720 error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n);
725 for (i = 0; i < n; i++) {
727 kevp->flags &= ~EV_SYSFLAGS;
728 error = kqueue_register(kq, kevp);
731 * If a registration returns an error we
732 * immediately post the error. The kevent()
733 * call itself will fail with the error if
734 * no space is available for posting.
736 * Such errors normally bypass the timeout/blocking
737 * code. However, if the copyoutfn function refuses
738 * to post the error (see sys_poll()), then we
741 if (error || (kevp->flags & EV_RECEIPT)) {
742 kevp->flags = EV_ERROR;
745 kevent_copyoutfn(uap, kevp, 1, res);
748 } else if (lres != *res) {
759 * Acquire/wait for events - setup timeout
762 if (tsp->tv_sec || tsp->tv_nsec) {
764 timespecadd(tsp, &ats); /* tsp = target time */
771 * Collect as many events as we can. Sleeping on successive
772 * loops is disabled if copyoutfn has incremented (*res).
774 * The loop stops if an error occurs, all events have been
775 * scanned (the marker has been reached), or fewer than the
776 * maximum number of events is found.
778 * The copyoutfn function does not have to increment (*res) in
779 * order for the loop to continue.
781 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents.
785 marker.kn_filter = EVFILT_MARKER;
786 marker.kn_status = KN_PROCESSING;
787 tok = lwkt_token_pool_lookup(kq);
789 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
791 while ((n = nevents - total) > 0) {
796 * If no events are pending sleep until timeout (if any)
797 * or an event occurs.
799 * After the sleep completes the marker is moved to the
800 * end of the list, making any received events available
803 if (kq->kq_count == 0 && *res == 0) {
808 } else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) {
812 struct timespec atx = *tsp;
815 timespecsub(&atx, &ats);
816 if (atx.tv_sec < 0) {
820 timeout = atx.tv_sec > 24 * 60 * 60 ?
827 if (kq->kq_count == 0) {
828 kq->kq_state |= KQ_SLEEP;
829 error = tsleep(kq, PCATCH, "kqread", timeout);
831 /* don't restart after signals... */
832 if (error == ERESTART)
839 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
840 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker,
847 * Process all received events
848 * Account for all non-spurious events in our total
850 i = kqueue_scan(kq, kev, n, &marker);
853 error = kevent_copyoutfn(uap, kev, i, res);
854 total += *res - lres;
858 if (limit && --limit == 0)
859 panic("kqueue: checkloop failed i=%d", i);
862 * Normally when fewer events are returned than requested
863 * we can stop. However, if only spurious events were
864 * collected the copyout will not bump (*res) and we have
871 * Deal with an edge case where spurious events can cause
872 * a loop to occur without moving the marker. This can
873 * prevent kqueue_scan() from picking up new events which
874 * race us. We must be sure to move the marker for this
877 * NOTE: We do not want to move the marker if events
878 * were scanned because normal kqueue operations
879 * may reactivate events. Moving the marker in
880 * that case could result in duplicates for the
885 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
886 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
891 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
894 /* Timeouts do not return EWOULDBLOCK. */
895 if (error == EWOULDBLOCK)
904 sys_kevent(struct kevent_args *uap)
906 struct thread *td = curthread;
907 struct proc *p = td->td_proc;
908 struct timespec ts, *tsp;
910 struct file *fp = NULL;
911 struct kevent_copyin_args *kap, ka;
915 error = copyin(uap->timeout, &ts, sizeof(ts));
922 fp = holdfp(p->p_fd, uap->fd, -1);
925 if (fp->f_type != DTYPE_KQUEUE) {
930 kq = (struct kqueue *)fp->f_data;
936 error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap,
937 kevent_copyin, kevent_copyout, tsp);
945 kqueue_register(struct kqueue *kq, struct kevent *kev)
947 struct lwkt_token *tok;
948 struct filedesc *fdp = kq->kq_fdp;
949 struct filterops *fops;
950 struct file *fp = NULL;
951 struct knote *kn = NULL;
954 if (kev->filter < 0) {
955 if (kev->filter + EVFILT_SYSCOUNT < 0)
957 fops = sysfilt_ops[~kev->filter]; /* to 0-base index */
961 * filter attach routine is responsible for insuring that
962 * the identifier can be attached to it.
967 tok = lwkt_token_pool_lookup(kq);
969 if (fops->f_flags & FILTEROP_ISFD) {
970 /* validate descriptor */
971 fp = holdfp(fdp, kev->ident, -1);
976 lwkt_getpooltoken(&fp->f_klist);
978 SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
979 if (kn->kn_kq == kq &&
980 kn->kn_filter == kev->filter &&
981 kn->kn_id == kev->ident) {
982 if (knote_acquire(kn) == 0)
987 lwkt_relpooltoken(&fp->f_klist);
989 if (kq->kq_knhashmask) {
992 list = &kq->kq_knhash[
993 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
994 lwkt_getpooltoken(list);
996 SLIST_FOREACH(kn, list, kn_link) {
997 if (kn->kn_id == kev->ident &&
998 kn->kn_filter == kev->filter) {
999 if (knote_acquire(kn) == 0)
1004 lwkt_relpooltoken(list);
1009 * NOTE: At this point if kn is non-NULL we will have acquired
1010 * it and set KN_PROCESSING.
1012 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) {
1018 * kn now contains the matching knote, or NULL if no match
1020 if (kev->flags & EV_ADD) {
1028 * apply reference count to knote structure, and
1029 * do not release it at the end of this routine.
1033 kn->kn_sfflags = kev->fflags;
1034 kn->kn_sdata = kev->data;
1037 kn->kn_kevent = *kev;
1040 * KN_PROCESSING prevents the knote from getting
1041 * ripped out from under us while we are trying
1042 * to attach it, in case the attach blocks.
1044 kn->kn_status = KN_PROCESSING;
1046 if ((error = filter_attach(kn)) != 0) {
1047 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1053 * Interlock against close races which either tried
1054 * to remove our knote while we were blocked or missed
1055 * it entirely prior to our attachment. We do not
1056 * want to end up with a knote on a closed descriptor.
1058 if ((fops->f_flags & FILTEROP_ISFD) &&
1059 checkfdclosed(fdp, kev->ident, kn->kn_fp)) {
1060 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1064 * The user may change some filter values after the
1065 * initial EV_ADD, but doing so will not reset any
1066 * filter which have already been triggered.
1068 KKASSERT(kn->kn_status & KN_PROCESSING);
1069 if (fops == &user_filtops) {
1070 filt_usertouch(kn, kev, EVENT_REGISTER);
1072 kn->kn_sfflags = kev->fflags;
1073 kn->kn_sdata = kev->data;
1074 kn->kn_kevent.udata = kev->udata;
1079 * Execute the filter event to immediately activate the
1080 * knote if necessary. If reprocessing events are pending
1081 * due to blocking above we do not run the filter here
1082 * but instead let knote_release() do it. Otherwise we
1083 * might run the filter on a deleted event.
1085 if ((kn->kn_status & KN_REPROCESS) == 0) {
1086 if (filter_event(kn, 0))
1089 } else if (kev->flags & EV_DELETE) {
1091 * Delete the existing knote
1093 knote_detach_and_drop(kn);
1097 * Modify an existing event.
1099 * The user may change some filter values after the
1100 * initial EV_ADD, but doing so will not reset any
1101 * filter which have already been triggered.
1103 KKASSERT(kn->kn_status & KN_PROCESSING);
1104 if (fops == &user_filtops) {
1105 filt_usertouch(kn, kev, EVENT_REGISTER);
1107 kn->kn_sfflags = kev->fflags;
1108 kn->kn_sdata = kev->data;
1109 kn->kn_kevent.udata = kev->udata;
1113 * Execute the filter event to immediately activate the
1114 * knote if necessary. If reprocessing events are pending
1115 * due to blocking above we do not run the filter here
1116 * but instead let knote_release() do it. Otherwise we
1117 * might run the filter on a deleted event.
1119 if ((kn->kn_status & KN_REPROCESS) == 0) {
1120 if (filter_event(kn, 0))
1126 * Disablement does not deactivate a knote here.
1128 if ((kev->flags & EV_DISABLE) &&
1129 ((kn->kn_status & KN_DISABLED) == 0)) {
1130 kn->kn_status |= KN_DISABLED;
1134 * Re-enablement may have to immediately enqueue an active knote.
1136 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1137 kn->kn_status &= ~KN_DISABLED;
1138 if ((kn->kn_status & KN_ACTIVE) &&
1139 ((kn->kn_status & KN_QUEUED) == 0)) {
1145 * Handle any required reprocessing
1148 /* kn may be invalid now */
1158 * Scan the kqueue, return the number of active events placed in kevp up
1161 * Continuous mode events may get recycled, do not continue scanning past
1162 * marker unless no events have been collected.
1165 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
1166 struct knote *marker)
1168 struct knote *kn, local_marker;
1172 local_marker.kn_filter = EVFILT_MARKER;
1173 local_marker.kn_status = KN_PROCESSING;
1175 lwkt_getpooltoken(kq);
1180 TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe);
1182 kn = TAILQ_NEXT(&local_marker, kn_tqe);
1183 if (kn->kn_filter == EVFILT_MARKER) {
1184 /* Marker reached, we are done */
1188 /* Move local marker past some other threads marker */
1189 kn = TAILQ_NEXT(kn, kn_tqe);
1190 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1191 TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe);
1196 * We can't skip a knote undergoing processing, otherwise
1197 * we risk not returning it when the user process expects
1198 * it should be returned. Sleep and retry.
1200 if (knote_acquire(kn) == 0)
1204 * Remove the event for processing.
1206 * WARNING! We must leave KN_QUEUED set to prevent the
1207 * event from being KNOTE_ACTIVATE()d while
1208 * the queue state is in limbo, in case we
1211 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1215 * We have to deal with an extremely important race against
1216 * file descriptor close()s here. The file descriptor can
1217 * disappear MPSAFE, and there is a small window of
1218 * opportunity between that and the call to knote_fdclose().
1220 * If we hit that window here while doselect or dopoll is
1221 * trying to delete a spurious event they will not be able
1222 * to match up the event against a knote and will go haywire.
1224 if ((kn->kn_fop->f_flags & FILTEROP_ISFD) &&
1225 checkfdclosed(kq->kq_fdp, kn->kn_kevent.ident, kn->kn_fp)) {
1226 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1229 if (kn->kn_status & KN_DISABLED) {
1231 * If disabled we ensure the event is not queued
1232 * but leave its active bit set. On re-enablement
1233 * the event may be immediately triggered.
1235 kn->kn_status &= ~KN_QUEUED;
1236 } else if ((kn->kn_flags & EV_ONESHOT) == 0 &&
1237 (kn->kn_status & KN_DELETING) == 0 &&
1238 filter_event(kn, 0) == 0) {
1240 * If not running in one-shot mode and the event
1241 * is no longer present we ensure it is removed
1242 * from the queue and ignore it.
1244 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1249 if (kn->kn_fop == &user_filtops)
1250 filt_usertouch(kn, kevp, EVENT_PROCESS);
1252 *kevp = kn->kn_kevent;
1257 if (kn->kn_flags & EV_ONESHOT) {
1258 kn->kn_status &= ~KN_QUEUED;
1259 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1261 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1262 if (kn->kn_flags & EV_CLEAR) {
1266 if (kn->kn_flags & EV_DISPATCH) {
1267 kn->kn_status |= KN_DISABLED;
1269 kn->kn_status &= ~(KN_QUEUED |
1272 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1279 * Handle any post-processing states
1283 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1285 lwkt_relpooltoken(kq);
1291 * This could be expanded to call kqueue_scan, if desired.
1296 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1305 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1314 kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
1315 struct ucred *cred, struct sysmsg *msg)
1317 struct lwkt_token *tok;
1321 kq = (struct kqueue *)fp->f_data;
1322 tok = lwkt_token_pool_lookup(kq);
1328 kq->kq_state |= KQ_ASYNC;
1330 kq->kq_state &= ~KQ_ASYNC;
1334 error = fsetown(*(int *)data, &kq->kq_sigio);
1348 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred)
1350 struct kqueue *kq = (struct kqueue *)fp->f_data;
1352 bzero((void *)st, sizeof(*st));
1353 st->st_size = kq->kq_count;
1354 st->st_blksize = sizeof(struct kevent);
1355 st->st_mode = S_IFIFO;
1363 kqueue_close(struct file *fp)
1365 struct kqueue *kq = (struct kqueue *)fp->f_data;
1367 kqueue_terminate(kq);
1370 funsetown(&kq->kq_sigio);
1372 kfree(kq, M_KQUEUE);
1377 kqueue_wakeup(struct kqueue *kq)
1379 if (kq->kq_state & KQ_SLEEP) {
1380 kq->kq_state &= ~KQ_SLEEP;
1383 KNOTE(&kq->kq_kqinfo.ki_note, 0);
1387 * Calls filterops f_attach function, acquiring mplock if filter is not
1388 * marked as FILTEROP_MPSAFE.
1390 * Caller must be holding the related kq token
1393 filter_attach(struct knote *kn)
1397 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1398 ret = kn->kn_fop->f_attach(kn);
1401 ret = kn->kn_fop->f_attach(kn);
1408 * Detach the knote and drop it, destroying the knote.
1410 * Calls filterops f_detach function, acquiring mplock if filter is not
1411 * marked as FILTEROP_MPSAFE.
1413 * Caller must be holding the related kq token
1416 knote_detach_and_drop(struct knote *kn)
1418 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1419 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1420 kn->kn_fop->f_detach(kn);
1423 kn->kn_fop->f_detach(kn);
1430 * Calls filterops f_event function, acquiring mplock if filter is not
1431 * marked as FILTEROP_MPSAFE.
1433 * If the knote is in the middle of being created or deleted we cannot
1434 * safely call the filter op.
1436 * Caller must be holding the related kq token
1439 filter_event(struct knote *kn, long hint)
1443 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1444 ret = kn->kn_fop->f_event(kn, hint);
1447 ret = kn->kn_fop->f_event(kn, hint);
1454 * Walk down a list of knotes, activating them if their event has triggered.
1456 * If we encounter any knotes which are undergoing processing we just mark
1457 * them for reprocessing and do not try to [re]activate the knote. However,
1458 * if a hint is being passed we have to wait and that makes things a bit
1462 knote(struct klist *list, long hint)
1466 struct knote *kntmp;
1468 lwkt_getpooltoken(list);
1470 SLIST_FOREACH(kn, list, kn_next) {
1472 lwkt_getpooltoken(kq);
1474 /* temporary verification hack */
1475 SLIST_FOREACH(kntmp, list, kn_next) {
1479 if (kn != kntmp || kn->kn_kq != kq) {
1480 lwkt_relpooltoken(kq);
1484 if (kn->kn_status & KN_PROCESSING) {
1486 * Someone else is processing the knote, ask the
1487 * other thread to reprocess it and don't mess
1488 * with it otherwise.
1491 kn->kn_status |= KN_REPROCESS;
1492 lwkt_relpooltoken(kq);
1497 * If the hint is non-zero we have to wait or risk
1498 * losing the state the caller is trying to update.
1500 * XXX This is a real problem, certain process
1501 * and signal filters will bump kn_data for
1502 * already-processed notes more than once if
1503 * we restart the list scan. FIXME.
1505 kn->kn_status |= KN_WAITING | KN_REPROCESS;
1506 tsleep(kn, 0, "knotec", hz);
1507 lwkt_relpooltoken(kq);
1512 * Become the reprocessing master ourselves.
1514 * If hint is non-zero running the event is mandatory
1515 * when not deleting so do it whether reprocessing is
1518 kn->kn_status |= KN_PROCESSING;
1519 if ((kn->kn_status & KN_DELETING) == 0) {
1520 if (filter_event(kn, hint))
1523 if (knote_release(kn)) {
1524 lwkt_relpooltoken(kq);
1527 lwkt_relpooltoken(kq);
1529 lwkt_relpooltoken(list);
1533 * Insert knote at head of klist.
1535 * This function may only be called via a filter function and thus
1536 * kq_token should already be held and marked for processing.
1539 knote_insert(struct klist *klist, struct knote *kn)
1541 lwkt_getpooltoken(klist);
1542 KKASSERT(kn->kn_status & KN_PROCESSING);
1543 SLIST_INSERT_HEAD(klist, kn, kn_next);
1544 lwkt_relpooltoken(klist);
1548 * Remove knote from a klist
1550 * This function may only be called via a filter function and thus
1551 * kq_token should already be held and marked for processing.
1554 knote_remove(struct klist *klist, struct knote *kn)
1556 lwkt_getpooltoken(klist);
1557 KKASSERT(kn->kn_status & KN_PROCESSING);
1558 SLIST_REMOVE(klist, kn, knote, kn_next);
1559 lwkt_relpooltoken(klist);
1564 * Remove all knotes from a specified klist
1566 * Only called from aio.
1569 knote_empty(struct klist *list)
1573 lwkt_gettoken(&kq_token);
1574 while ((kn = SLIST_FIRST(list)) != NULL) {
1575 if (knote_acquire(kn))
1576 knote_detach_and_drop(kn);
1578 lwkt_reltoken(&kq_token);
1583 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst,
1584 struct filterops *ops, void *hook)
1589 lwkt_getpooltoken(&src->ki_note);
1590 lwkt_getpooltoken(&dst->ki_note);
1591 while ((kn = SLIST_FIRST(&src->ki_note)) != NULL) {
1593 lwkt_getpooltoken(kq);
1594 if (SLIST_FIRST(&src->ki_note) != kn || kn->kn_kq != kq) {
1595 lwkt_relpooltoken(kq);
1598 if (knote_acquire(kn)) {
1599 knote_remove(&src->ki_note, kn);
1602 knote_insert(&dst->ki_note, kn);
1604 /* kn may be invalid now */
1606 lwkt_relpooltoken(kq);
1608 lwkt_relpooltoken(&dst->ki_note);
1609 lwkt_relpooltoken(&src->ki_note);
1613 * Remove all knotes referencing a specified fd
1616 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd)
1620 struct knote *kntmp;
1622 lwkt_getpooltoken(&fp->f_klist);
1624 SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
1625 if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) {
1627 lwkt_getpooltoken(kq);
1629 /* temporary verification hack */
1630 SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) {
1634 if (kn != kntmp || kn->kn_kq->kq_fdp != fdp ||
1635 kn->kn_id != fd || kn->kn_kq != kq) {
1636 lwkt_relpooltoken(kq);
1639 if (knote_acquire(kn))
1640 knote_detach_and_drop(kn);
1641 lwkt_relpooltoken(kq);
1645 lwkt_relpooltoken(&fp->f_klist);
1649 * Low level attach function.
1651 * The knote should already be marked for processing.
1652 * Caller must hold the related kq token.
1655 knote_attach(struct knote *kn)
1658 struct kqueue *kq = kn->kn_kq;
1660 if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1661 KKASSERT(kn->kn_fp);
1662 list = &kn->kn_fp->f_klist;
1664 if (kq->kq_knhashmask == 0)
1665 kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1666 &kq->kq_knhashmask);
1667 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1669 lwkt_getpooltoken(list);
1670 SLIST_INSERT_HEAD(list, kn, kn_link);
1671 lwkt_relpooltoken(list);
1672 TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink);
1676 * Low level drop function.
1678 * The knote should already be marked for processing.
1679 * Caller must hold the related kq token.
1682 knote_drop(struct knote *kn)
1689 if (kn->kn_fop->f_flags & FILTEROP_ISFD)
1690 list = &kn->kn_fp->f_klist;
1692 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1694 lwkt_getpooltoken(list);
1695 SLIST_REMOVE(list, kn, knote, kn_link);
1696 lwkt_relpooltoken(list);
1697 TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink);
1698 if (kn->kn_status & KN_QUEUED)
1700 if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1708 * Low level enqueue function.
1710 * The knote should already be marked for processing.
1711 * Caller must be holding the kq token
1714 knote_enqueue(struct knote *kn)
1716 struct kqueue *kq = kn->kn_kq;
1718 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
1719 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1720 kn->kn_status |= KN_QUEUED;
1724 * Send SIGIO on request (typically set up as a mailbox signal)
1726 if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1)
1727 pgsigio(kq->kq_sigio, SIGIO, 0);
1733 * Low level dequeue function.
1735 * The knote should already be marked for processing.
1736 * Caller must be holding the kq token
1739 knote_dequeue(struct knote *kn)
1741 struct kqueue *kq = kn->kn_kq;
1743 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
1744 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1745 kn->kn_status &= ~KN_QUEUED;
1749 static struct knote *
1752 return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK);
1756 knote_free(struct knote *kn)
1758 kfree(kn, M_KQUEUE);