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>
58 * Global token for kqueue subsystem
61 struct lwkt_token kq_token = LWKT_TOKEN_INITIALIZER(kq_token);
62 SYSCTL_LONG(_lwkt, OID_AUTO, kq_collisions,
63 CTLFLAG_RW, &kq_token.t_collisions, 0,
64 "Collision counter of kq_token");
67 MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
69 struct kevent_copyin_args {
70 struct kevent_args *ka;
74 static int kqueue_sleep(struct kqueue *kq, struct timespec *tsp);
75 static int kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
76 struct knote *marker);
77 static int kqueue_read(struct file *fp, struct uio *uio,
78 struct ucred *cred, int flags);
79 static int kqueue_write(struct file *fp, struct uio *uio,
80 struct ucred *cred, int flags);
81 static int kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
82 struct ucred *cred, struct sysmsg *msg);
83 static int kqueue_kqfilter(struct file *fp, struct knote *kn);
84 static int kqueue_stat(struct file *fp, struct stat *st,
86 static int kqueue_close(struct file *fp);
87 static void kqueue_wakeup(struct kqueue *kq);
88 static int filter_attach(struct knote *kn);
89 static int filter_event(struct knote *kn, long hint);
94 static struct fileops kqueueops = {
95 .fo_read = kqueue_read,
96 .fo_write = kqueue_write,
97 .fo_ioctl = kqueue_ioctl,
98 .fo_kqfilter = kqueue_kqfilter,
99 .fo_stat = kqueue_stat,
100 .fo_close = kqueue_close,
101 .fo_shutdown = nofo_shutdown
104 static void knote_attach(struct knote *kn);
105 static void knote_drop(struct knote *kn);
106 static void knote_detach_and_drop(struct knote *kn);
107 static void knote_enqueue(struct knote *kn);
108 static void knote_dequeue(struct knote *kn);
109 static struct knote *knote_alloc(void);
110 static void knote_free(struct knote *kn);
112 static void filt_kqdetach(struct knote *kn);
113 static int filt_kqueue(struct knote *kn, long hint);
114 static int filt_procattach(struct knote *kn);
115 static void filt_procdetach(struct knote *kn);
116 static int filt_proc(struct knote *kn, long hint);
117 static int filt_fileattach(struct knote *kn);
118 static void filt_timerexpire(void *knx);
119 static int filt_timerattach(struct knote *kn);
120 static void filt_timerdetach(struct knote *kn);
121 static int filt_timer(struct knote *kn, long hint);
123 static struct filterops file_filtops =
124 { FILTEROP_ISFD, filt_fileattach, NULL, NULL };
125 static struct filterops kqread_filtops =
126 { FILTEROP_ISFD, NULL, filt_kqdetach, filt_kqueue };
127 static struct filterops proc_filtops =
128 { 0, filt_procattach, filt_procdetach, filt_proc };
129 static struct filterops timer_filtops =
130 { 0, filt_timerattach, filt_timerdetach, filt_timer };
132 static int kq_ncallouts = 0;
133 static int kq_calloutmax = (4 * 1024);
134 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
135 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
136 static int kq_checkloop = 1000000;
137 SYSCTL_INT(_kern, OID_AUTO, kq_checkloop, CTLFLAG_RW,
138 &kq_checkloop, 0, "Maximum number of callouts allocated for kqueue");
140 #define KNOTE_ACTIVATE(kn) do { \
141 kn->kn_status |= KN_ACTIVE; \
142 if ((kn->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \
146 #define KN_HASHSIZE 64 /* XXX should be tunable */
147 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask))
149 extern struct filterops aio_filtops;
150 extern struct filterops sig_filtops;
153 * Table for for all system-defined filters.
155 static struct filterops *sysfilt_ops[] = {
156 &file_filtops, /* EVFILT_READ */
157 &file_filtops, /* EVFILT_WRITE */
158 &aio_filtops, /* EVFILT_AIO */
159 &file_filtops, /* EVFILT_VNODE */
160 &proc_filtops, /* EVFILT_PROC */
161 &sig_filtops, /* EVFILT_SIGNAL */
162 &timer_filtops, /* EVFILT_TIMER */
163 &file_filtops, /* EVFILT_EXCEPT */
167 filt_fileattach(struct knote *kn)
169 return (fo_kqfilter(kn->kn_fp, kn));
176 kqueue_kqfilter(struct file *fp, struct knote *kn)
178 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
180 if (kn->kn_filter != EVFILT_READ)
183 kn->kn_fop = &kqread_filtops;
184 knote_insert(&kq->kq_kqinfo.ki_note, kn);
189 filt_kqdetach(struct knote *kn)
191 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
193 knote_remove(&kq->kq_kqinfo.ki_note, kn);
198 filt_kqueue(struct knote *kn, long hint)
200 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
202 kn->kn_data = kq->kq_count;
203 return (kn->kn_data > 0);
207 filt_procattach(struct knote *kn)
213 p = pfind(kn->kn_id);
214 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
215 p = zpfind(kn->kn_id);
221 if (!PRISON_CHECK(curthread->td_ucred, p->p_ucred)) {
227 lwkt_gettoken(&p->p_token);
228 kn->kn_ptr.p_proc = p;
229 kn->kn_flags |= EV_CLEAR; /* automatically set */
232 * internal flag indicating registration done by kernel
234 if (kn->kn_flags & EV_FLAG1) {
235 kn->kn_data = kn->kn_sdata; /* ppid */
236 kn->kn_fflags = NOTE_CHILD;
237 kn->kn_flags &= ~EV_FLAG1;
240 knote_insert(&p->p_klist, kn);
243 * Immediately activate any exit notes if the target process is a
244 * zombie. This is necessary to handle the case where the target
245 * process, e.g. a child, dies before the kevent is negistered.
247 if (immediate && filt_proc(kn, NOTE_EXIT))
249 lwkt_reltoken(&p->p_token);
256 * The knote may be attached to a different process, which may exit,
257 * leaving nothing for the knote to be attached to. So when the process
258 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
259 * it will be deleted when read out. However, as part of the knote deletion,
260 * this routine is called, so a check is needed to avoid actually performing
261 * a detach, because the original process does not exist any more.
264 filt_procdetach(struct knote *kn)
268 if (kn->kn_status & KN_DETACHED)
270 p = kn->kn_ptr.p_proc;
271 knote_remove(&p->p_klist, kn);
275 filt_proc(struct knote *kn, long hint)
280 * mask off extra data
282 event = (u_int)hint & NOTE_PCTRLMASK;
285 * if the user is interested in this event, record it.
287 if (kn->kn_sfflags & event)
288 kn->kn_fflags |= event;
291 * Process is gone, so flag the event as finished. Detach the
292 * knote from the process now because the process will be poof,
295 if (event == NOTE_EXIT) {
296 struct proc *p = kn->kn_ptr.p_proc;
297 if ((kn->kn_status & KN_DETACHED) == 0) {
299 knote_remove(&p->p_klist, kn);
300 kn->kn_status |= KN_DETACHED;
301 kn->kn_data = p->p_xstat;
302 kn->kn_ptr.p_proc = NULL;
305 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
310 * process forked, and user wants to track the new process,
311 * so attach a new knote to it, and immediately report an
312 * event with the parent's pid.
314 if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) {
319 * register knote with new process.
321 kev.ident = hint & NOTE_PDATAMASK; /* pid */
322 kev.filter = kn->kn_filter;
323 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
324 kev.fflags = kn->kn_sfflags;
325 kev.data = kn->kn_id; /* parent */
326 kev.udata = kn->kn_kevent.udata; /* preserve udata */
327 error = kqueue_register(kn->kn_kq, &kev);
329 kn->kn_fflags |= NOTE_TRACKERR;
332 return (kn->kn_fflags != 0);
336 * The callout interlocks with callout_terminate() but can still
337 * race a deletion so if KN_DELETING is set we just don't touch
341 filt_timerexpire(void *knx)
343 struct lwkt_token *tok;
344 struct knote *kn = knx;
345 struct callout *calloutp;
349 tok = lwkt_token_pool_lookup(kn->kn_kq);
351 if ((kn->kn_status & KN_DELETING) == 0) {
355 if ((kn->kn_flags & EV_ONESHOT) == 0) {
356 tv.tv_sec = kn->kn_sdata / 1000;
357 tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
358 tticks = tvtohz_high(&tv);
359 calloutp = (struct callout *)kn->kn_hook;
360 callout_reset(calloutp, tticks, filt_timerexpire, kn);
367 * data contains amount of time to sleep, in milliseconds
370 filt_timerattach(struct knote *kn)
372 struct callout *calloutp;
376 if (kq_ncallouts >= kq_calloutmax) {
382 tv.tv_sec = kn->kn_sdata / 1000;
383 tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
384 tticks = tvtohz_high(&tv);
386 kn->kn_flags |= EV_CLEAR; /* automatically set */
387 calloutp = kmalloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
388 callout_init(calloutp);
389 kn->kn_hook = (caddr_t)calloutp;
390 callout_reset(calloutp, tticks, filt_timerexpire, kn);
396 * This function is called with the knote flagged locked but it is
397 * still possible to race a callout event due to the callback blocking.
398 * We must call callout_terminate() instead of callout_stop() to deal
402 filt_timerdetach(struct knote *kn)
404 struct callout *calloutp;
406 calloutp = (struct callout *)kn->kn_hook;
407 callout_terminate(calloutp);
408 kfree(calloutp, M_KQUEUE);
413 filt_timer(struct knote *kn, long hint)
416 return (kn->kn_data != 0);
420 * Acquire a knote, return non-zero on success, 0 on failure.
422 * If we cannot acquire the knote we sleep and return 0. The knote
423 * may be stale on return in this case and the caller must restart
424 * whatever loop they are in.
426 * Related kq token must be held.
430 knote_acquire(struct knote *kn)
432 if (kn->kn_status & KN_PROCESSING) {
433 kn->kn_status |= KN_WAITING | KN_REPROCESS;
434 tsleep(kn, 0, "kqepts", hz);
435 /* knote may be stale now */
438 kn->kn_status |= KN_PROCESSING;
443 * Release an acquired knote, clearing KN_PROCESSING and handling any
444 * KN_REPROCESS events.
446 * Caller must be holding the related kq token
448 * Non-zero is returned if the knote is destroyed or detached.
452 knote_release(struct knote *kn)
454 while (kn->kn_status & KN_REPROCESS) {
455 kn->kn_status &= ~KN_REPROCESS;
456 if (kn->kn_status & KN_WAITING) {
457 kn->kn_status &= ~KN_WAITING;
460 if (kn->kn_status & KN_DELETING) {
461 knote_detach_and_drop(kn);
465 if (filter_event(kn, 0))
468 if (kn->kn_status & KN_DETACHED) {
469 kn->kn_status &= ~KN_PROCESSING;
472 kn->kn_status &= ~KN_PROCESSING;
478 * Initialize a kqueue.
480 * NOTE: The lwp/proc code initializes a kqueue for select/poll ops.
485 kqueue_init(struct kqueue *kq, struct filedesc *fdp)
487 TAILQ_INIT(&kq->kq_knpend);
488 TAILQ_INIT(&kq->kq_knlist);
491 SLIST_INIT(&kq->kq_kqinfo.ki_note);
495 * Terminate a kqueue. Freeing the actual kq itself is left up to the
496 * caller (it might be embedded in a lwp so we don't do it here).
498 * The kq's knlist must be completely eradicated so block on any
502 kqueue_terminate(struct kqueue *kq)
504 struct lwkt_token *tok;
507 tok = lwkt_token_pool_lookup(kq);
509 while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) {
510 if (knote_acquire(kn))
511 knote_detach_and_drop(kn);
514 hashdestroy(kq->kq_knhash, M_KQUEUE, kq->kq_knhashmask);
515 kq->kq_knhash = NULL;
516 kq->kq_knhashmask = 0;
525 sys_kqueue(struct kqueue_args *uap)
527 struct thread *td = curthread;
532 error = falloc(td->td_lwp, &fp, &fd);
535 fp->f_flag = FREAD | FWRITE;
536 fp->f_type = DTYPE_KQUEUE;
537 fp->f_ops = &kqueueops;
539 kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO);
540 kqueue_init(kq, td->td_proc->p_fd);
543 fsetfd(kq->kq_fdp, fp, fd);
544 uap->sysmsg_result = fd;
550 * Copy 'count' items into the destination list pointed to by uap->eventlist.
553 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res)
555 struct kevent_copyin_args *kap;
558 kap = (struct kevent_copyin_args *)arg;
560 error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp));
562 kap->ka->eventlist += count;
572 * Copy at most 'max' items from the list pointed to by kap->changelist,
573 * return number of items in 'events'.
576 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events)
578 struct kevent_copyin_args *kap;
581 kap = (struct kevent_copyin_args *)arg;
583 count = min(kap->ka->nchanges - kap->pchanges, max);
584 error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp);
586 kap->ka->changelist += count;
587 kap->pchanges += count;
598 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap,
599 k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn,
600 struct timespec *tsp_in)
603 struct timespec *tsp;
604 int i, n, total, error, nerrors = 0;
606 int limit = kq_checkloop;
607 struct kevent kev[KQ_NEVENTS];
609 struct lwkt_token *tok;
611 if (tsp_in == NULL || tsp_in->tv_sec || tsp_in->tv_nsec)
612 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC);
618 tok = lwkt_token_pool_lookup(kq);
622 error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n);
627 for (i = 0; i < n; i++) {
629 kevp->flags &= ~EV_SYSFLAGS;
630 error = kqueue_register(kq, kevp);
633 * If a registration returns an error we
634 * immediately post the error. The kevent()
635 * call itself will fail with the error if
636 * no space is available for posting.
638 * Such errors normally bypass the timeout/blocking
639 * code. However, if the copyoutfn function refuses
640 * to post the error (see sys_poll()), then we
644 kevp->flags = EV_ERROR;
647 kevent_copyoutfn(uap, kevp, 1, res);
650 } else if (lres != *res) {
663 * Acquire/wait for events - setup timeout
668 if (tsp->tv_sec || tsp->tv_nsec) {
670 timespecadd(tsp, &ats); /* tsp = target time */
677 * Collect as many events as we can. Sleeping on successive
678 * loops is disabled if copyoutfn has incremented (*res).
680 * The loop stops if an error occurs, all events have been
681 * scanned (the marker has been reached), or fewer than the
682 * maximum number of events is found.
684 * The copyoutfn function does not have to increment (*res) in
685 * order for the loop to continue.
687 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents.
691 marker.kn_filter = EVFILT_MARKER;
692 marker.kn_status = KN_PROCESSING;
693 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
694 while ((n = nevents - total) > 0) {
699 * If no events are pending sleep until timeout (if any)
700 * or an event occurs.
702 * After the sleep completes the marker is moved to the
703 * end of the list, making any received events available
706 if (kq->kq_count == 0 && *res == 0) {
707 error = kqueue_sleep(kq, tsp);
711 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
712 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
716 * Process all received events
717 * Account for all non-spurious events in our total
719 i = kqueue_scan(kq, kev, n, &marker);
722 error = kevent_copyoutfn(uap, kev, i, res);
723 total += *res - lres;
727 if (limit && --limit == 0)
728 panic("kqueue: checkloop failed i=%d", i);
731 * Normally when fewer events are returned than requested
732 * we can stop. However, if only spurious events were
733 * collected the copyout will not bump (*res) and we have
740 * Deal with an edge case where spurious events can cause
741 * a loop to occur without moving the marker. This can
742 * prevent kqueue_scan() from picking up new events which
743 * race us. We must be sure to move the marker for this
746 * NOTE: We do not want to move the marker if events
747 * were scanned because normal kqueue operations
748 * may reactivate events. Moving the marker in
749 * that case could result in duplicates for the
753 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
754 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
757 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
759 /* Timeouts do not return EWOULDBLOCK. */
760 if (error == EWOULDBLOCK)
772 sys_kevent(struct kevent_args *uap)
774 struct thread *td = curthread;
775 struct proc *p = td->td_proc;
776 struct timespec ts, *tsp;
778 struct file *fp = NULL;
779 struct kevent_copyin_args *kap, ka;
783 error = copyin(uap->timeout, &ts, sizeof(ts));
790 fp = holdfp(p->p_fd, uap->fd, -1);
793 if (fp->f_type != DTYPE_KQUEUE) {
798 kq = (struct kqueue *)fp->f_data;
804 error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap,
805 kevent_copyin, kevent_copyout, tsp);
813 * Caller must be holding the kq token
816 kqueue_register(struct kqueue *kq, struct kevent *kev)
818 struct lwkt_token *tok;
819 struct filedesc *fdp = kq->kq_fdp;
820 struct filterops *fops;
821 struct file *fp = NULL;
822 struct knote *kn = NULL;
825 if (kev->filter < 0) {
826 if (kev->filter + EVFILT_SYSCOUNT < 0)
828 fops = sysfilt_ops[~kev->filter]; /* to 0-base index */
832 * filter attach routine is responsible for insuring that
833 * the identifier can be attached to it.
835 kprintf("unknown filter: %d\n", kev->filter);
839 tok = lwkt_token_pool_lookup(kq);
841 if (fops->f_flags & FILTEROP_ISFD) {
842 /* validate descriptor */
843 fp = holdfp(fdp, kev->ident, -1);
848 lwkt_getpooltoken(&fp->f_klist);
850 SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
851 if (kn->kn_kq == kq &&
852 kn->kn_filter == kev->filter &&
853 kn->kn_id == kev->ident) {
854 if (knote_acquire(kn) == 0)
859 lwkt_relpooltoken(&fp->f_klist);
861 if (kq->kq_knhashmask) {
864 list = &kq->kq_knhash[
865 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
866 lwkt_getpooltoken(list);
868 SLIST_FOREACH(kn, list, kn_link) {
869 if (kn->kn_id == kev->ident &&
870 kn->kn_filter == kev->filter) {
871 if (knote_acquire(kn) == 0)
876 lwkt_relpooltoken(list);
881 * NOTE: At this point if kn is non-NULL we will have acquired
882 * it and set KN_PROCESSING.
884 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) {
890 * kn now contains the matching knote, or NULL if no match
892 if (kev->flags & EV_ADD) {
904 * apply reference count to knote structure, and
905 * do not release it at the end of this routine.
909 kn->kn_sfflags = kev->fflags;
910 kn->kn_sdata = kev->data;
913 kn->kn_kevent = *kev;
916 * KN_PROCESSING prevents the knote from getting
917 * ripped out from under us while we are trying
918 * to attach it, in case the attach blocks.
920 kn->kn_status = KN_PROCESSING;
922 if ((error = filter_attach(kn)) != 0) {
923 kn->kn_status |= KN_DELETING | KN_REPROCESS;
929 * Interlock against close races which either tried
930 * to remove our knote while we were blocked or missed
931 * it entirely prior to our attachment. We do not
932 * want to end up with a knote on a closed descriptor.
934 if ((fops->f_flags & FILTEROP_ISFD) &&
935 checkfdclosed(fdp, kev->ident, kn->kn_fp)) {
936 kn->kn_status |= KN_DELETING | KN_REPROCESS;
940 * The user may change some filter values after the
941 * initial EV_ADD, but doing so will not reset any
942 * filter which have already been triggered.
944 KKASSERT(kn->kn_status & KN_PROCESSING);
945 kn->kn_sfflags = kev->fflags;
946 kn->kn_sdata = kev->data;
947 kn->kn_kevent.udata = kev->udata;
951 * Execute the filter event to immediately activate the
952 * knote if necessary. If reprocessing events are pending
953 * due to blocking above we do not run the filter here
954 * but instead let knote_release() do it. Otherwise we
955 * might run the filter on a deleted event.
957 if ((kn->kn_status & KN_REPROCESS) == 0) {
958 if (filter_event(kn, 0))
961 } else if (kev->flags & EV_DELETE) {
963 * Delete the existing knote
965 knote_detach_and_drop(kn);
970 * Disablement does not deactivate a knote here.
972 if ((kev->flags & EV_DISABLE) &&
973 ((kn->kn_status & KN_DISABLED) == 0)) {
974 kn->kn_status |= KN_DISABLED;
978 * Re-enablement may have to immediately enqueue an active knote.
980 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
981 kn->kn_status &= ~KN_DISABLED;
982 if ((kn->kn_status & KN_ACTIVE) &&
983 ((kn->kn_status & KN_QUEUED) == 0)) {
989 * Handle any required reprocessing
992 /* kn may be invalid now */
1002 * Block as necessary until the target time is reached.
1003 * If tsp is NULL we block indefinitely. If tsp->ts_secs/nsecs are both
1004 * 0 we do not block at all.
1006 * Caller must be holding the kq token.
1009 kqueue_sleep(struct kqueue *kq, struct timespec *tsp)
1014 kq->kq_state |= KQ_SLEEP;
1015 error = tsleep(kq, PCATCH, "kqread", 0);
1016 } else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) {
1017 error = EWOULDBLOCK;
1019 struct timespec ats;
1020 struct timespec atx = *tsp;
1023 getnanouptime(&ats);
1024 timespecsub(&atx, &ats);
1025 if (ats.tv_sec < 0) {
1026 error = EWOULDBLOCK;
1028 timeout = atx.tv_sec > 24 * 60 * 60 ?
1029 24 * 60 * 60 * hz : tstohz_high(&atx);
1030 kq->kq_state |= KQ_SLEEP;
1031 error = tsleep(kq, PCATCH, "kqread", timeout);
1035 /* don't restart after signals... */
1036 if (error == ERESTART)
1043 * Scan the kqueue, return the number of active events placed in kevp up
1046 * Continuous mode events may get recycled, do not continue scanning past
1047 * marker unless no events have been collected.
1049 * Caller must be holding the kq token
1052 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
1053 struct knote *marker)
1055 struct knote *kn, local_marker;
1059 local_marker.kn_filter = EVFILT_MARKER;
1060 local_marker.kn_status = KN_PROCESSING;
1065 TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe);
1067 kn = TAILQ_NEXT(&local_marker, kn_tqe);
1068 if (kn->kn_filter == EVFILT_MARKER) {
1069 /* Marker reached, we are done */
1073 /* Move local marker past some other threads marker */
1074 kn = TAILQ_NEXT(kn, kn_tqe);
1075 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1076 TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe);
1081 * We can't skip a knote undergoing processing, otherwise
1082 * we risk not returning it when the user process expects
1083 * it should be returned. Sleep and retry.
1085 if (knote_acquire(kn) == 0)
1089 * Remove the event for processing.
1091 * WARNING! We must leave KN_QUEUED set to prevent the
1092 * event from being KNOTE_ACTIVATE()d while
1093 * the queue state is in limbo, in case we
1096 * WARNING! We must set KN_PROCESSING to avoid races
1097 * against deletion or another thread's
1100 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1104 * We have to deal with an extremely important race against
1105 * file descriptor close()s here. The file descriptor can
1106 * disappear MPSAFE, and there is a small window of
1107 * opportunity between that and the call to knote_fdclose().
1109 * If we hit that window here while doselect or dopoll is
1110 * trying to delete a spurious event they will not be able
1111 * to match up the event against a knote and will go haywire.
1113 if ((kn->kn_fop->f_flags & FILTEROP_ISFD) &&
1114 checkfdclosed(kq->kq_fdp, kn->kn_kevent.ident, kn->kn_fp)) {
1115 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1118 if (kn->kn_status & KN_DISABLED) {
1120 * If disabled we ensure the event is not queued
1121 * but leave its active bit set. On re-enablement
1122 * the event may be immediately triggered.
1124 kn->kn_status &= ~KN_QUEUED;
1125 } else if ((kn->kn_flags & EV_ONESHOT) == 0 &&
1126 (kn->kn_status & KN_DELETING) == 0 &&
1127 filter_event(kn, 0) == 0) {
1129 * If not running in one-shot mode and the event
1130 * is no longer present we ensure it is removed
1131 * from the queue and ignore it.
1133 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1138 *kevp++ = kn->kn_kevent;
1142 if (kn->kn_flags & EV_ONESHOT) {
1143 kn->kn_status &= ~KN_QUEUED;
1144 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1145 } else if (kn->kn_flags & EV_CLEAR) {
1148 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1150 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1156 * Handle any post-processing states
1160 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1167 * This could be expanded to call kqueue_scan, if desired.
1172 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1181 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1190 kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
1191 struct ucred *cred, struct sysmsg *msg)
1193 struct lwkt_token *tok;
1197 kq = (struct kqueue *)fp->f_data;
1198 tok = lwkt_token_pool_lookup(kq);
1204 kq->kq_state |= KQ_ASYNC;
1206 kq->kq_state &= ~KQ_ASYNC;
1210 error = fsetown(*(int *)data, &kq->kq_sigio);
1224 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred)
1226 struct kqueue *kq = (struct kqueue *)fp->f_data;
1228 bzero((void *)st, sizeof(*st));
1229 st->st_size = kq->kq_count;
1230 st->st_blksize = sizeof(struct kevent);
1231 st->st_mode = S_IFIFO;
1239 kqueue_close(struct file *fp)
1241 struct kqueue *kq = (struct kqueue *)fp->f_data;
1243 kqueue_terminate(kq);
1246 funsetown(&kq->kq_sigio);
1248 kfree(kq, M_KQUEUE);
1253 kqueue_wakeup(struct kqueue *kq)
1255 if (kq->kq_state & KQ_SLEEP) {
1256 kq->kq_state &= ~KQ_SLEEP;
1259 KNOTE(&kq->kq_kqinfo.ki_note, 0);
1263 * Calls filterops f_attach function, acquiring mplock if filter is not
1264 * marked as FILTEROP_MPSAFE.
1266 * Caller must be holding the related kq token
1269 filter_attach(struct knote *kn)
1273 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1274 ret = kn->kn_fop->f_attach(kn);
1277 ret = kn->kn_fop->f_attach(kn);
1284 * Detach the knote and drop it, destroying the knote.
1286 * Calls filterops f_detach function, acquiring mplock if filter is not
1287 * marked as FILTEROP_MPSAFE.
1289 * Caller must be holding the related kq token
1292 knote_detach_and_drop(struct knote *kn)
1294 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1295 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1296 kn->kn_fop->f_detach(kn);
1299 kn->kn_fop->f_detach(kn);
1306 * Calls filterops f_event function, acquiring mplock if filter is not
1307 * marked as FILTEROP_MPSAFE.
1309 * If the knote is in the middle of being created or deleted we cannot
1310 * safely call the filter op.
1312 * Caller must be holding the related kq token
1315 filter_event(struct knote *kn, long hint)
1319 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1320 ret = kn->kn_fop->f_event(kn, hint);
1323 ret = kn->kn_fop->f_event(kn, hint);
1330 * Walk down a list of knotes, activating them if their event has triggered.
1332 * If we encounter any knotes which are undergoing processing we just mark
1333 * them for reprocessing and do not try to [re]activate the knote. However,
1334 * if a hint is being passed we have to wait and that makes things a bit
1338 knote(struct klist *list, long hint)
1342 struct knote *kntmp;
1344 lwkt_getpooltoken(list);
1346 SLIST_FOREACH(kn, list, kn_next) {
1348 lwkt_getpooltoken(kq);
1350 /* temporary verification hack */
1351 SLIST_FOREACH(kntmp, list, kn_next) {
1355 if (kn != kntmp || kn->kn_kq != kq) {
1356 lwkt_relpooltoken(kq);
1360 if (kn->kn_status & KN_PROCESSING) {
1362 * Someone else is processing the knote, ask the
1363 * other thread to reprocess it and don't mess
1364 * with it otherwise.
1367 kn->kn_status |= KN_REPROCESS;
1368 lwkt_relpooltoken(kq);
1373 * If the hint is non-zero we have to wait or risk
1374 * losing the state the caller is trying to update.
1376 * XXX This is a real problem, certain process
1377 * and signal filters will bump kn_data for
1378 * already-processed notes more than once if
1379 * we restart the list scan. FIXME.
1381 kn->kn_status |= KN_WAITING | KN_REPROCESS;
1382 tsleep(kn, 0, "knotec", hz);
1383 lwkt_relpooltoken(kq);
1388 * Become the reprocessing master ourselves.
1390 * If hint is non-zer running the event is mandatory
1391 * when not deleting so do it whether reprocessing is
1394 kn->kn_status |= KN_PROCESSING;
1395 if ((kn->kn_status & KN_DELETING) == 0) {
1396 if (filter_event(kn, hint))
1399 if (knote_release(kn)) {
1400 lwkt_relpooltoken(kq);
1403 lwkt_relpooltoken(kq);
1405 lwkt_relpooltoken(list);
1409 * Insert knote at head of klist.
1411 * This function may only be called via a filter function and thus
1412 * kq_token should already be held and marked for processing.
1415 knote_insert(struct klist *klist, struct knote *kn)
1417 lwkt_getpooltoken(klist);
1418 KKASSERT(kn->kn_status & KN_PROCESSING);
1419 SLIST_INSERT_HEAD(klist, kn, kn_next);
1420 lwkt_relpooltoken(klist);
1424 * Remove knote from a klist
1426 * This function may only be called via a filter function and thus
1427 * kq_token should already be held and marked for processing.
1430 knote_remove(struct klist *klist, struct knote *kn)
1432 lwkt_getpooltoken(klist);
1433 KKASSERT(kn->kn_status & KN_PROCESSING);
1434 SLIST_REMOVE(klist, kn, knote, kn_next);
1435 lwkt_relpooltoken(klist);
1440 * Remove all knotes from a specified klist
1442 * Only called from aio.
1445 knote_empty(struct klist *list)
1449 lwkt_gettoken(&kq_token);
1450 while ((kn = SLIST_FIRST(list)) != NULL) {
1451 if (knote_acquire(kn))
1452 knote_detach_and_drop(kn);
1454 lwkt_reltoken(&kq_token);
1459 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst,
1460 struct filterops *ops, void *hook)
1465 lwkt_getpooltoken(&src->ki_note);
1466 lwkt_getpooltoken(&dst->ki_note);
1467 while ((kn = SLIST_FIRST(&src->ki_note)) != NULL) {
1469 lwkt_getpooltoken(kq);
1470 if (SLIST_FIRST(&src->ki_note) != kn || kn->kn_kq != kq) {
1471 lwkt_relpooltoken(kq);
1474 if (knote_acquire(kn)) {
1475 knote_remove(&src->ki_note, kn);
1478 knote_insert(&dst->ki_note, kn);
1480 /* kn may be invalid now */
1482 lwkt_relpooltoken(kq);
1484 lwkt_relpooltoken(&dst->ki_note);
1485 lwkt_relpooltoken(&src->ki_note);
1489 * Remove all knotes referencing a specified fd
1492 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd)
1496 struct knote *kntmp;
1498 lwkt_getpooltoken(&fp->f_klist);
1500 SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
1501 if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) {
1503 lwkt_getpooltoken(kq);
1505 /* temporary verification hack */
1506 SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) {
1510 if (kn != kntmp || kn->kn_kq->kq_fdp != fdp ||
1511 kn->kn_id != fd || kn->kn_kq != kq) {
1512 lwkt_relpooltoken(kq);
1515 if (knote_acquire(kn))
1516 knote_detach_and_drop(kn);
1517 lwkt_relpooltoken(kq);
1521 lwkt_relpooltoken(&fp->f_klist);
1525 * Low level attach function.
1527 * The knote should already be marked for processing.
1528 * Caller must hold the related kq token.
1531 knote_attach(struct knote *kn)
1534 struct kqueue *kq = kn->kn_kq;
1536 if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1537 KKASSERT(kn->kn_fp);
1538 list = &kn->kn_fp->f_klist;
1540 if (kq->kq_knhashmask == 0)
1541 kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1542 &kq->kq_knhashmask);
1543 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1545 lwkt_getpooltoken(list);
1546 SLIST_INSERT_HEAD(list, kn, kn_link);
1547 TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink);
1548 lwkt_relpooltoken(list);
1552 * Low level drop function.
1554 * The knote should already be marked for processing.
1555 * Caller must hold the related kq token.
1558 knote_drop(struct knote *kn)
1565 if (kn->kn_fop->f_flags & FILTEROP_ISFD)
1566 list = &kn->kn_fp->f_klist;
1568 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1570 lwkt_getpooltoken(list);
1571 SLIST_REMOVE(list, kn, knote, kn_link);
1572 TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink);
1573 if (kn->kn_status & KN_QUEUED)
1575 if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1580 lwkt_relpooltoken(list);
1584 * Low level enqueue function.
1586 * The knote should already be marked for processing.
1587 * Caller must be holding the kq token
1590 knote_enqueue(struct knote *kn)
1592 struct kqueue *kq = kn->kn_kq;
1594 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
1595 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1596 kn->kn_status |= KN_QUEUED;
1600 * Send SIGIO on request (typically set up as a mailbox signal)
1602 if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1)
1603 pgsigio(kq->kq_sigio, SIGIO, 0);
1609 * Low level dequeue function.
1611 * The knote should already be marked for processing.
1612 * Caller must be holding the kq token
1615 knote_dequeue(struct knote *kn)
1617 struct kqueue *kq = kn->kn_kq;
1619 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
1620 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1621 kn->kn_status &= ~KN_QUEUED;
1625 static struct knote *
1628 return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK);
1632 knote_free(struct knote *kn)
1634 kfree(kn, M_KQUEUE);