xref: /freebsd/sys/kern/kern_event.c (revision 6af83ee0d2941d18880b6aaa2b4facd1d30c6106)
1 /*-
2  * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
3  * Copyright 2004 John-Mark Gurney <jmg@FreeBSD.org>
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/kernel.h>
34 #include <sys/lock.h>
35 #include <sys/mutex.h>
36 #include <sys/proc.h>
37 #include <sys/malloc.h>
38 #include <sys/unistd.h>
39 #include <sys/file.h>
40 #include <sys/filedesc.h>
41 #include <sys/filio.h>
42 #include <sys/fcntl.h>
43 #include <sys/kthread.h>
44 #include <sys/selinfo.h>
45 #include <sys/queue.h>
46 #include <sys/event.h>
47 #include <sys/eventvar.h>
48 #include <sys/poll.h>
49 #include <sys/protosw.h>
50 #include <sys/sigio.h>
51 #include <sys/signalvar.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/stat.h>
55 #include <sys/sysctl.h>
56 #include <sys/sysproto.h>
57 #include <sys/taskqueue.h>
58 #include <sys/uio.h>
59 
60 #include <vm/uma.h>
61 
62 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
63 
64 /*
65  * This lock is used if multiple kq locks are required.  This possibly
66  * should be made into a per proc lock.
67  */
68 static struct mtx	kq_global;
69 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
70 #define KQ_GLOBAL_LOCK(lck, haslck)	do {	\
71 	if (!haslck)				\
72 		mtx_lock(lck);			\
73 	haslck = 1;				\
74 } while (0)
75 #define KQ_GLOBAL_UNLOCK(lck, haslck)	do {	\
76 	if (haslck)				\
77 		mtx_unlock(lck);			\
78 	haslck = 0;				\
79 } while (0)
80 
81 TASKQUEUE_DEFINE_THREAD(kqueue);
82 
83 static int	kqueue_aquire(struct file *fp, struct kqueue **kqp);
84 static void	kqueue_release(struct kqueue *kq, int locked);
85 static int	kqueue_expand(struct kqueue *kq, struct filterops *fops,
86 		    uintptr_t ident, int waitok);
87 static void	kqueue_task(void *arg, int pending);
88 static int	kqueue_scan(struct kqueue *kq, int maxevents,
89 		    struct kevent *ulistp, const struct timespec *timeout,
90 		    struct kevent *keva, struct thread *td);
91 static void 	kqueue_wakeup(struct kqueue *kq);
92 static struct filterops *kqueue_fo_find(int filt);
93 static void	kqueue_fo_release(int filt);
94 
95 static fo_rdwr_t	kqueue_read;
96 static fo_rdwr_t	kqueue_write;
97 static fo_ioctl_t	kqueue_ioctl;
98 static fo_poll_t	kqueue_poll;
99 static fo_kqfilter_t	kqueue_kqfilter;
100 static fo_stat_t	kqueue_stat;
101 static fo_close_t	kqueue_close;
102 
103 static struct fileops kqueueops = {
104 	.fo_read = kqueue_read,
105 	.fo_write = kqueue_write,
106 	.fo_ioctl = kqueue_ioctl,
107 	.fo_poll = kqueue_poll,
108 	.fo_kqfilter = kqueue_kqfilter,
109 	.fo_stat = kqueue_stat,
110 	.fo_close = kqueue_close,
111 };
112 
113 static int 	knote_attach(struct knote *kn, struct kqueue *kq);
114 static void 	knote_drop(struct knote *kn, struct thread *td);
115 static void 	knote_enqueue(struct knote *kn);
116 static void 	knote_dequeue(struct knote *kn);
117 static void 	knote_init(void);
118 static struct 	knote *knote_alloc(int waitok);
119 static void 	knote_free(struct knote *kn);
120 
121 static void	filt_kqdetach(struct knote *kn);
122 static int	filt_kqueue(struct knote *kn, long hint);
123 static int	filt_procattach(struct knote *kn);
124 static void	filt_procdetach(struct knote *kn);
125 static int	filt_proc(struct knote *kn, long hint);
126 static int	filt_fileattach(struct knote *kn);
127 static void	filt_timerexpire(void *knx);
128 static int	filt_timerattach(struct knote *kn);
129 static void	filt_timerdetach(struct knote *kn);
130 static int	filt_timer(struct knote *kn, long hint);
131 
132 static struct filterops file_filtops =
133 	{ 1, filt_fileattach, NULL, NULL };
134 static struct filterops kqread_filtops =
135 	{ 1, NULL, filt_kqdetach, filt_kqueue };
136 /* XXX - move to kern_proc.c?  */
137 static struct filterops proc_filtops =
138 	{ 0, filt_procattach, filt_procdetach, filt_proc };
139 static struct filterops timer_filtops =
140 	{ 0, filt_timerattach, filt_timerdetach, filt_timer };
141 
142 static uma_zone_t	knote_zone;
143 static int 		kq_ncallouts = 0;
144 static int 		kq_calloutmax = (4 * 1024);
145 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
146     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
147 
148 /* XXX - ensure not KN_INFLUX?? */
149 #define KNOTE_ACTIVATE(kn, islock) do { 				\
150 	if ((islock))							\
151 		mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED);		\
152 	else								\
153 		KQ_LOCK((kn)->kn_kq);					\
154 	(kn)->kn_status |= KN_ACTIVE;					\
155 	if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)		\
156 		knote_enqueue((kn));					\
157 	if (!(islock))							\
158 		KQ_UNLOCK((kn)->kn_kq);					\
159 } while(0)
160 #define KQ_LOCK(kq) do {						\
161 	mtx_lock(&(kq)->kq_lock);					\
162 } while (0)
163 #define KQ_FLUX_WAKEUP(kq) do {						\
164 	if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) {		\
165 		(kq)->kq_state &= ~KQ_FLUXWAIT;				\
166 		wakeup((kq));						\
167 	}								\
168 } while (0)
169 #define KQ_UNLOCK_FLUX(kq) do {						\
170 	KQ_FLUX_WAKEUP(kq);						\
171 	mtx_unlock(&(kq)->kq_lock);					\
172 } while (0)
173 #define KQ_UNLOCK(kq) do {						\
174 	mtx_unlock(&(kq)->kq_lock);					\
175 } while (0)
176 #define KQ_OWNED(kq) do {						\
177 	mtx_assert(&(kq)->kq_lock, MA_OWNED);				\
178 } while (0)
179 #define KQ_NOTOWNED(kq) do {						\
180 	mtx_assert(&(kq)->kq_lock, MA_NOTOWNED);			\
181 } while (0)
182 #define KN_LIST_LOCK(kn) do {						\
183 	if (kn->kn_knlist != NULL)					\
184 		mtx_lock(kn->kn_knlist->kl_lock);			\
185 } while (0)
186 #define KN_LIST_UNLOCK(kn) do {						\
187 	if (kn->kn_knlist != NULL)					\
188 		mtx_unlock(kn->kn_knlist->kl_lock);			\
189 } while (0)
190 
191 #define	KN_HASHSIZE		64		/* XXX should be tunable */
192 #define KN_HASH(val, mask)	(((val) ^ (val >> 8)) & (mask))
193 
194 static int
195 filt_nullattach(struct knote *kn)
196 {
197 
198 	return (ENXIO);
199 };
200 
201 struct filterops null_filtops =
202 	{ 0, filt_nullattach, NULL, NULL };
203 
204 /* XXX - make SYSINIT to add these, and move into respective modules. */
205 extern struct filterops sig_filtops;
206 extern struct filterops fs_filtops;
207 
208 /*
209  * Table for for all system-defined filters.
210  */
211 static struct mtx	filterops_lock;
212 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
213 	MTX_DEF);
214 static struct {
215 	struct filterops *for_fop;
216 	int for_refcnt;
217 } sysfilt_ops[EVFILT_SYSCOUNT] = {
218 	{ &file_filtops },			/* EVFILT_READ */
219 	{ &file_filtops },			/* EVFILT_WRITE */
220 	{ &null_filtops },			/* EVFILT_AIO */
221 	{ &file_filtops },			/* EVFILT_VNODE */
222 	{ &proc_filtops },			/* EVFILT_PROC */
223 	{ &sig_filtops },			/* EVFILT_SIGNAL */
224 	{ &timer_filtops },			/* EVFILT_TIMER */
225 	{ &file_filtops },			/* EVFILT_NETDEV */
226 	{ &fs_filtops },			/* EVFILT_FS */
227 };
228 
229 /*
230  * Simple redirection for all cdevsw style objects to call their fo_kqfilter
231  * method.
232  */
233 static int
234 filt_fileattach(struct knote *kn)
235 {
236 
237 	return (fo_kqfilter(kn->kn_fp, kn));
238 }
239 
240 /*ARGSUSED*/
241 static int
242 kqueue_kqfilter(struct file *fp, struct knote *kn)
243 {
244 	struct kqueue *kq = kn->kn_fp->f_data;
245 
246 	if (kn->kn_filter != EVFILT_READ)
247 		return (EINVAL);
248 
249 	kn->kn_status |= KN_KQUEUE;
250 	kn->kn_fop = &kqread_filtops;
251 	knlist_add(&kq->kq_sel.si_note, kn, 0);
252 
253 	return (0);
254 }
255 
256 static void
257 filt_kqdetach(struct knote *kn)
258 {
259 	struct kqueue *kq = kn->kn_fp->f_data;
260 
261 	knlist_remove(&kq->kq_sel.si_note, kn, 0);
262 }
263 
264 /*ARGSUSED*/
265 static int
266 filt_kqueue(struct knote *kn, long hint)
267 {
268 	struct kqueue *kq = kn->kn_fp->f_data;
269 
270 	kn->kn_data = kq->kq_count;
271 	return (kn->kn_data > 0);
272 }
273 
274 /* XXX - move to kern_proc.c?  */
275 static int
276 filt_procattach(struct knote *kn)
277 {
278 	struct proc *p;
279 	int immediate;
280 	int error;
281 
282 	immediate = 0;
283 	p = pfind(kn->kn_id);
284 	if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
285 		p = zpfind(kn->kn_id);
286 		immediate = 1;
287 	} else if (p != NULL && (p->p_flag & P_WEXIT)) {
288 		immediate = 1;
289 	}
290 
291 	if (p == NULL)
292 		return (ESRCH);
293 	if ((error = p_cansee(curthread, p)))
294 		return (error);
295 
296 	kn->kn_ptr.p_proc = p;
297 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
298 
299 	/*
300 	 * internal flag indicating registration done by kernel
301 	 */
302 	if (kn->kn_flags & EV_FLAG1) {
303 		kn->kn_data = kn->kn_sdata;		/* ppid */
304 		kn->kn_fflags = NOTE_CHILD;
305 		kn->kn_flags &= ~EV_FLAG1;
306 	}
307 
308 	if (immediate == 0)
309 		knlist_add(&p->p_klist, kn, 1);
310 
311 	/*
312 	 * Immediately activate any exit notes if the target process is a
313 	 * zombie.  This is necessary to handle the case where the target
314 	 * process, e.g. a child, dies before the kevent is registered.
315 	 */
316 	if (immediate && filt_proc(kn, NOTE_EXIT))
317 		KNOTE_ACTIVATE(kn, 0);
318 
319 	PROC_UNLOCK(p);
320 
321 	return (0);
322 }
323 
324 /*
325  * The knote may be attached to a different process, which may exit,
326  * leaving nothing for the knote to be attached to.  So when the process
327  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
328  * it will be deleted when read out.  However, as part of the knote deletion,
329  * this routine is called, so a check is needed to avoid actually performing
330  * a detach, because the original process does not exist any more.
331  */
332 /* XXX - move to kern_proc.c?  */
333 static void
334 filt_procdetach(struct knote *kn)
335 {
336 	struct proc *p;
337 
338 	p = kn->kn_ptr.p_proc;
339 	knlist_remove(&p->p_klist, kn, 0);
340 	kn->kn_ptr.p_proc = NULL;
341 }
342 
343 /* XXX - move to kern_proc.c?  */
344 static int
345 filt_proc(struct knote *kn, long hint)
346 {
347 	struct proc *p = kn->kn_ptr.p_proc;
348 	u_int event;
349 
350 	/*
351 	 * mask off extra data
352 	 */
353 	event = (u_int)hint & NOTE_PCTRLMASK;
354 
355 	/*
356 	 * if the user is interested in this event, record it.
357 	 */
358 	if (kn->kn_sfflags & event)
359 		kn->kn_fflags |= event;
360 
361 	/*
362 	 * process is gone, so flag the event as finished.
363 	 */
364 	if (event == NOTE_EXIT) {
365 		if (!(kn->kn_status & KN_DETACHED))
366 			knlist_remove_inevent(&p->p_klist, kn);
367 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
368 		kn->kn_ptr.p_proc = NULL;
369 		return (1);
370 	}
371 
372 	/*
373 	 * process forked, and user wants to track the new process,
374 	 * so attach a new knote to it, and immediately report an
375 	 * event with the parent's pid.
376 	 */
377 	if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) {
378 		struct kevent kev;
379 		int error;
380 
381 		/*
382 		 * register knote with new process.
383 		 */
384 		kev.ident = hint & NOTE_PDATAMASK;	/* pid */
385 		kev.filter = kn->kn_filter;
386 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
387 		kev.fflags = kn->kn_sfflags;
388 		kev.data = kn->kn_id;			/* parent */
389 		kev.udata = kn->kn_kevent.udata;	/* preserve udata */
390 		error = kqueue_register(kn->kn_kq, &kev, NULL, 0);
391 		if (error)
392 			kn->kn_fflags |= NOTE_TRACKERR;
393 	}
394 
395 	return (kn->kn_fflags != 0);
396 }
397 
398 static int
399 timertoticks(intptr_t data)
400 {
401 	struct timeval tv;
402 	int tticks;
403 
404 	tv.tv_sec = data / 1000;
405 	tv.tv_usec = (data % 1000) * 1000;
406 	tticks = tvtohz(&tv);
407 
408 	return tticks;
409 }
410 
411 /* XXX - move to kern_timeout.c? */
412 static void
413 filt_timerexpire(void *knx)
414 {
415 	struct knote *kn = knx;
416 	struct callout *calloutp;
417 
418 	kn->kn_data++;
419 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
420 
421 	if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
422 		calloutp = (struct callout *)kn->kn_hook;
423 		callout_reset(calloutp, timertoticks(kn->kn_sdata),
424 		    filt_timerexpire, kn);
425 	}
426 }
427 
428 /*
429  * data contains amount of time to sleep, in milliseconds
430  */
431 /* XXX - move to kern_timeout.c? */
432 static int
433 filt_timerattach(struct knote *kn)
434 {
435 	struct callout *calloutp;
436 
437 	atomic_add_int(&kq_ncallouts, 1);
438 
439 	if (kq_ncallouts >= kq_calloutmax) {
440 		atomic_add_int(&kq_ncallouts, -1);
441 		return (ENOMEM);
442 	}
443 
444 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
445 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add usually sets it */
446 	MALLOC(calloutp, struct callout *, sizeof(*calloutp),
447 	    M_KQUEUE, M_WAITOK);
448 	callout_init(calloutp, 1);
449 	kn->kn_hook = calloutp;
450 	callout_reset(calloutp, timertoticks(kn->kn_sdata), filt_timerexpire,
451 	    kn);
452 
453 	return (0);
454 }
455 
456 /* XXX - move to kern_timeout.c? */
457 static void
458 filt_timerdetach(struct knote *kn)
459 {
460 	struct callout *calloutp;
461 
462 	calloutp = (struct callout *)kn->kn_hook;
463 	callout_drain(calloutp);
464 	FREE(calloutp, M_KQUEUE);
465 	atomic_add_int(&kq_ncallouts, -1);
466 	kn->kn_status |= KN_DETACHED;	/* knlist_remove usually clears it */
467 }
468 
469 /* XXX - move to kern_timeout.c? */
470 static int
471 filt_timer(struct knote *kn, long hint)
472 {
473 
474 	return (kn->kn_data != 0);
475 }
476 
477 /*
478  * MPSAFE
479  */
480 int
481 kqueue(struct thread *td, struct kqueue_args *uap)
482 {
483 	struct filedesc *fdp;
484 	struct kqueue *kq;
485 	struct file *fp;
486 	int fd, error;
487 
488 	fdp = td->td_proc->p_fd;
489 	error = falloc(td, &fp, &fd);
490 	if (error)
491 		goto done2;
492 
493 	/* An extra reference on `nfp' has been held for us by falloc(). */
494 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
495 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
496 	TAILQ_INIT(&kq->kq_head);
497 	kq->kq_fdp = fdp;
498 	knlist_init(&kq->kq_sel.si_note, &kq->kq_lock);
499 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
500 
501 	FILEDESC_LOCK_FAST(fdp);
502 	SLIST_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
503 	FILEDESC_UNLOCK_FAST(fdp);
504 
505 	FILE_LOCK(fp);
506 	fp->f_flag = FREAD | FWRITE;
507 	fp->f_type = DTYPE_KQUEUE;
508 	fp->f_ops = &kqueueops;
509 	fp->f_data = kq;
510 	FILE_UNLOCK(fp);
511 	fdrop(fp, td);
512 
513 	td->td_retval[0] = fd;
514 done2:
515 	return (error);
516 }
517 
518 #ifndef _SYS_SYSPROTO_H_
519 struct kevent_args {
520 	int	fd;
521 	const struct kevent *changelist;
522 	int	nchanges;
523 	struct	kevent *eventlist;
524 	int	nevents;
525 	const struct timespec *timeout;
526 };
527 #endif
528 /*
529  * MPSAFE
530  */
531 int
532 kevent(struct thread *td, struct kevent_args *uap)
533 {
534 	struct kevent keva[KQ_NEVENTS];
535 	struct kevent *kevp;
536 	struct kqueue *kq;
537 	struct file *fp;
538 	struct timespec ts;
539 	int i, n, nerrors, error;
540 
541 	if ((error = fget(td, uap->fd, &fp)) != 0)
542 		return (error);
543 	if ((error = kqueue_aquire(fp, &kq)) != 0)
544 		goto done_norel;
545 
546 	if (uap->timeout != NULL) {
547 		error = copyin(uap->timeout, &ts, sizeof(ts));
548 		if (error)
549 			goto done;
550 		uap->timeout = &ts;
551 	}
552 
553 	nerrors = 0;
554 
555 	while (uap->nchanges > 0) {
556 		n = uap->nchanges > KQ_NEVENTS ? KQ_NEVENTS : uap->nchanges;
557 		error = copyin(uap->changelist, keva,
558 		    n * sizeof *keva);
559 		if (error)
560 			goto done;
561 		for (i = 0; i < n; i++) {
562 			kevp = &keva[i];
563 			kevp->flags &= ~EV_SYSFLAGS;
564 			error = kqueue_register(kq, kevp, td, 1);
565 			if (error) {
566 				if (uap->nevents != 0) {
567 					kevp->flags = EV_ERROR;
568 					kevp->data = error;
569 					(void) copyout(kevp,
570 					    uap->eventlist,
571 					    sizeof(*kevp));
572 					uap->eventlist++;
573 					uap->nevents--;
574 					nerrors++;
575 				} else {
576 					goto done;
577 				}
578 			}
579 		}
580 		uap->nchanges -= n;
581 		uap->changelist += n;
582 	}
583 	if (nerrors) {
584 		td->td_retval[0] = nerrors;
585 		error = 0;
586 		goto done;
587 	}
588 
589 	error = kqueue_scan(kq, uap->nevents, uap->eventlist, uap->timeout,
590 	    keva, td);
591 done:
592 	kqueue_release(kq, 0);
593 done_norel:
594 	if (fp != NULL)
595 		fdrop(fp, td);
596 	return (error);
597 }
598 
599 int
600 kqueue_add_filteropts(int filt, struct filterops *filtops)
601 {
602 	int error;
603 
604 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
605 		printf(
606 "trying to add a filterop that is out of range: %d is beyond %d\n",
607 		    ~filt, EVFILT_SYSCOUNT);
608 		return EINVAL;
609 	}
610 	mtx_lock(&filterops_lock);
611 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
612 	    sysfilt_ops[~filt].for_fop != NULL)
613 		error = EEXIST;
614 	else {
615 		sysfilt_ops[~filt].for_fop = filtops;
616 		sysfilt_ops[~filt].for_refcnt = 0;
617 	}
618 	mtx_unlock(&filterops_lock);
619 
620 	return (0);
621 }
622 
623 int
624 kqueue_del_filteropts(int filt)
625 {
626 	int error;
627 
628 	error = 0;
629 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
630 		return EINVAL;
631 
632 	mtx_lock(&filterops_lock);
633 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
634 	    sysfilt_ops[~filt].for_fop == NULL)
635 		error = EINVAL;
636 	else if (sysfilt_ops[~filt].for_refcnt != 0)
637 		error = EBUSY;
638 	else {
639 		sysfilt_ops[~filt].for_fop = &null_filtops;
640 		sysfilt_ops[~filt].for_refcnt = 0;
641 	}
642 	mtx_unlock(&filterops_lock);
643 
644 	return error;
645 }
646 
647 static struct filterops *
648 kqueue_fo_find(int filt)
649 {
650 
651 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
652 		return NULL;
653 
654 	mtx_lock(&filterops_lock);
655 	sysfilt_ops[~filt].for_refcnt++;
656 	if (sysfilt_ops[~filt].for_fop == NULL)
657 		sysfilt_ops[~filt].for_fop = &null_filtops;
658 	mtx_unlock(&filterops_lock);
659 
660 	return sysfilt_ops[~filt].for_fop;
661 }
662 
663 static void
664 kqueue_fo_release(int filt)
665 {
666 
667 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
668 		return;
669 
670 	mtx_lock(&filterops_lock);
671 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
672 	    ("filter object refcount not valid on release"));
673 	sysfilt_ops[~filt].for_refcnt--;
674 	mtx_unlock(&filterops_lock);
675 }
676 
677 /*
678  * A ref to kq (obtained via kqueue_aquire) should be held.  waitok will
679  * influence if memory allocation should wait.  Make sure it is 0 if you
680  * hold any mutexes.
681  */
682 int
683 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
684 {
685 	struct filedesc *fdp;
686 	struct filterops *fops;
687 	struct file *fp;
688 	struct knote *kn, *tkn;
689 	int error, filt, event;
690 	int haskqglobal;
691 	int fd;
692 
693 	fdp = NULL;
694 	fp = NULL;
695 	kn = NULL;
696 	error = 0;
697 	haskqglobal = 0;
698 
699 	filt = kev->filter;
700 	fops = kqueue_fo_find(filt);
701 	if (fops == NULL)
702 		return EINVAL;
703 
704 	tkn = knote_alloc(waitok);		/* prevent waiting with locks */
705 
706 findkn:
707 	if (fops->f_isfd) {
708 		KASSERT(td != NULL, ("td is NULL"));
709 		fdp = td->td_proc->p_fd;
710 		FILEDESC_LOCK(fdp);
711 		/* validate descriptor */
712 		fd = kev->ident;
713 		if (fd < 0 || fd >= fdp->fd_nfiles ||
714 		    (fp = fdp->fd_ofiles[fd]) == NULL) {
715 			FILEDESC_UNLOCK(fdp);
716 			error = EBADF;
717 			goto done;
718 		}
719 		fhold(fp);
720 
721 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
722 		    kev->ident, 0) != 0) {
723 			/* unlock and try again */
724 			FILEDESC_UNLOCK(fdp);
725 			fdrop(fp, td);
726 			fp = NULL;
727 			error = kqueue_expand(kq, fops, kev->ident, waitok);
728 			if (error)
729 				goto done;
730 			goto findkn;
731 		}
732 
733 		if (fp->f_type == DTYPE_KQUEUE) {
734 			/*
735 			 * if we add some inteligence about what we are doing,
736 			 * we should be able to support events on ourselves.
737 			 * We need to know when we are doing this to prevent
738 			 * getting both the knlist lock and the kq lock since
739 			 * they are the same thing.
740 			 */
741 			if (fp->f_data == kq) {
742 				FILEDESC_UNLOCK(fdp);
743 				error = EINVAL;
744 				goto done_noglobal;
745 			}
746 
747 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
748 		}
749 
750 		FILEDESC_UNLOCK(fdp);
751 		KQ_LOCK(kq);
752 		if (kev->ident < kq->kq_knlistsize) {
753 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
754 				if (kev->filter == kn->kn_filter)
755 					break;
756 		}
757 	} else {
758 		if ((kev->flags & EV_ADD) == EV_ADD)
759 			kqueue_expand(kq, fops, kev->ident, waitok);
760 
761 		KQ_LOCK(kq);
762 		if (kq->kq_knhashmask != 0) {
763 			struct klist *list;
764 
765 			list = &kq->kq_knhash[
766 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
767 			SLIST_FOREACH(kn, list, kn_link)
768 				if (kev->ident == kn->kn_id &&
769 				    kev->filter == kn->kn_filter)
770 					break;
771 		}
772 	}
773 
774 	/* knote is in the process of changing, wait for it to stablize. */
775 	if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
776 		if (fp != NULL) {
777 			fdrop(fp, td);
778 			fp = NULL;
779 		}
780 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
781 		kq->kq_state |= KQ_FLUXWAIT;
782 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
783 		goto findkn;
784 	}
785 
786 	if (kn == NULL && ((kev->flags & EV_ADD) == 0)) {
787 		KQ_UNLOCK(kq);
788 		error = ENOENT;
789 		goto done;
790 	}
791 
792 	/*
793 	 * kn now contains the matching knote, or NULL if no match
794 	 */
795 	if (kev->flags & EV_ADD) {
796 		if (kn == NULL) {
797 			kn = tkn;
798 			tkn = NULL;
799 			if (kn == NULL) {
800 				error = ENOMEM;
801 				goto done;
802 			}
803 			kn->kn_fp = fp;
804 			kn->kn_kq = kq;
805 			kn->kn_fop = fops;
806 			/*
807 			 * apply reference counts to knote structure, and
808 			 * do not release it at the end of this routine.
809 			 */
810 			fops = NULL;
811 			fp = NULL;
812 
813 			kn->kn_sfflags = kev->fflags;
814 			kn->kn_sdata = kev->data;
815 			kev->fflags = 0;
816 			kev->data = 0;
817 			kn->kn_kevent = *kev;
818 			kn->kn_status = KN_INFLUX|KN_DETACHED;
819 
820 			error = knote_attach(kn, kq);
821 			KQ_UNLOCK(kq);
822 			if (error != 0) {
823 				tkn = kn;
824 				goto done;
825 			}
826 
827 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
828 				knote_drop(kn, td);
829 				goto done;
830 			}
831 			KN_LIST_LOCK(kn);
832 		} else {
833 			/*
834 			 * The user may change some filter values after the
835 			 * initial EV_ADD, but doing so will not reset any
836 			 * filter which has already been triggered.
837 			 */
838 			kn->kn_status |= KN_INFLUX;
839 			KQ_UNLOCK(kq);
840 			KN_LIST_LOCK(kn);
841 			kn->kn_sfflags = kev->fflags;
842 			kn->kn_sdata = kev->data;
843 			kn->kn_kevent.udata = kev->udata;
844 		}
845 
846 		/*
847 		 * We can get here with kn->kn_knlist == NULL.
848 		 * This can happen when the initial attach event decides that
849 		 * the event is "completed" already.  i.e. filt_procattach
850 		 * is called on a zombie process.  It will call filt_proc
851 		 * which will remove it from the list, and NULL kn_knlist.
852 		 */
853 		event = kn->kn_fop->f_event(kn, 0);
854 		KN_LIST_UNLOCK(kn);
855 		KQ_LOCK(kq);
856 		if (event)
857 			KNOTE_ACTIVATE(kn, 1);
858 		kn->kn_status &= ~KN_INFLUX;
859 	} else if (kev->flags & EV_DELETE) {
860 		kn->kn_status |= KN_INFLUX;
861 		KQ_UNLOCK(kq);
862 		if (!(kn->kn_status & KN_DETACHED))
863 			kn->kn_fop->f_detach(kn);
864 		knote_drop(kn, td);
865 		goto done;
866 	}
867 
868 	if ((kev->flags & EV_DISABLE) &&
869 	    ((kn->kn_status & KN_DISABLED) == 0)) {
870 		kn->kn_status |= KN_DISABLED;
871 	}
872 
873 	if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
874 		kn->kn_status &= ~KN_DISABLED;
875 		if ((kn->kn_status & KN_ACTIVE) &&
876 		    ((kn->kn_status & KN_QUEUED) == 0))
877 			knote_enqueue(kn);
878 	}
879 	KQ_UNLOCK_FLUX(kq);
880 
881 done:
882 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
883 done_noglobal:
884 	if (fp != NULL)
885 		fdrop(fp, td);
886 	if (tkn != NULL)
887 		knote_free(tkn);
888 	if (fops != NULL)
889 		kqueue_fo_release(filt);
890 	return (error);
891 }
892 
893 static int
894 kqueue_aquire(struct file *fp, struct kqueue **kqp)
895 {
896 	int error;
897 	struct kqueue *kq;
898 
899 	error = 0;
900 
901 	FILE_LOCK(fp);
902 	do {
903 		kq = fp->f_data;
904 		if (fp->f_type != DTYPE_KQUEUE || kq == NULL) {
905 			error = EBADF;
906 			break;
907 		}
908 		*kqp = kq;
909 		KQ_LOCK(kq);
910 		if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
911 			KQ_UNLOCK(kq);
912 			error = EBADF;
913 			break;
914 		}
915 		kq->kq_refcnt++;
916 		KQ_UNLOCK(kq);
917 	} while (0);
918 	FILE_UNLOCK(fp);
919 
920 	return error;
921 }
922 
923 static void
924 kqueue_release(struct kqueue *kq, int locked)
925 {
926 	if (locked)
927 		KQ_OWNED(kq);
928 	else
929 		KQ_LOCK(kq);
930 	kq->kq_refcnt--;
931 	if (kq->kq_refcnt == 1)
932 		wakeup(&kq->kq_refcnt);
933 	if (!locked)
934 		KQ_UNLOCK(kq);
935 }
936 
937 static void
938 kqueue_schedtask(struct kqueue *kq)
939 {
940 
941 	KQ_OWNED(kq);
942 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
943 	    ("scheduling kqueue task while draining"));
944 
945 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
946 		taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
947 		kq->kq_state |= KQ_TASKSCHED;
948 	}
949 }
950 
951 /*
952  * Expand the kq to make sure we have storage for fops/ident pair.
953  *
954  * Return 0 on success (or no work necessary), return errno on failure.
955  *
956  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
957  * If kqueue_register is called from a non-fd context, there usually/should
958  * be no locks held.
959  */
960 static int
961 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
962 	int waitok)
963 {
964 	struct klist *list, *tmp_knhash;
965 	u_long tmp_knhashmask;
966 	int size;
967 	int fd;
968 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
969 
970 	KQ_NOTOWNED(kq);
971 
972 	if (fops->f_isfd) {
973 		fd = ident;
974 		if (kq->kq_knlistsize <= fd) {
975 			size = kq->kq_knlistsize;
976 			while (size <= fd)
977 				size += KQEXTENT;
978 			MALLOC(list, struct klist *,
979 			    size * sizeof list, M_KQUEUE, mflag);
980 			if (list == NULL)
981 				return ENOMEM;
982 			KQ_LOCK(kq);
983 			if (kq->kq_knlistsize > fd) {
984 				FREE(list, M_KQUEUE);
985 				list = NULL;
986 			} else {
987 				if (kq->kq_knlist != NULL) {
988 					bcopy(kq->kq_knlist, list,
989 					    kq->kq_knlistsize * sizeof list);
990 					FREE(kq->kq_knlist, M_KQUEUE);
991 					kq->kq_knlist = NULL;
992 				}
993 				bzero((caddr_t)list +
994 				    kq->kq_knlistsize * sizeof list,
995 				    (size - kq->kq_knlistsize) * sizeof list);
996 				kq->kq_knlistsize = size;
997 				kq->kq_knlist = list;
998 			}
999 			KQ_UNLOCK(kq);
1000 		}
1001 	} else {
1002 		if (kq->kq_knhashmask == 0) {
1003 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1004 			    &tmp_knhashmask);
1005 			if (tmp_knhash == NULL)
1006 				return ENOMEM;
1007 			KQ_LOCK(kq);
1008 			if (kq->kq_knhashmask == 0) {
1009 				kq->kq_knhash = tmp_knhash;
1010 				kq->kq_knhashmask = tmp_knhashmask;
1011 			} else {
1012 				free(tmp_knhash, M_KQUEUE);
1013 			}
1014 			KQ_UNLOCK(kq);
1015 		}
1016 	}
1017 
1018 	KQ_NOTOWNED(kq);
1019 	return 0;
1020 }
1021 
1022 static void
1023 kqueue_task(void *arg, int pending)
1024 {
1025 	struct kqueue *kq;
1026 	int haskqglobal;
1027 
1028 	haskqglobal = 0;
1029 	kq = arg;
1030 
1031 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1032 	KQ_LOCK(kq);
1033 
1034 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1035 
1036 	kq->kq_state &= ~KQ_TASKSCHED;
1037 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1038 		wakeup(&kq->kq_state);
1039 	}
1040 	KQ_UNLOCK(kq);
1041 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1042 }
1043 
1044 /*
1045  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1046  * We treat KN_MARKER knotes as if they are INFLUX.
1047  */
1048 static int
1049 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent *ulistp,
1050 	const struct timespec *tsp, struct kevent *keva, struct thread *td)
1051 {
1052 	struct kevent *kevp;
1053 	struct timeval atv, rtv, ttv;
1054 	struct knote *kn, *marker;
1055 	int count, timeout, nkev, error;
1056 	int haskqglobal;
1057 
1058 	count = maxevents;
1059 	nkev = 0;
1060 	error = 0;
1061 	haskqglobal = 0;
1062 
1063 	if (maxevents == 0)
1064 		goto done_nl;
1065 
1066 	if (tsp != NULL) {
1067 		TIMESPEC_TO_TIMEVAL(&atv, tsp);
1068 		if (itimerfix(&atv)) {
1069 			error = EINVAL;
1070 			goto done_nl;
1071 		}
1072 		if (tsp->tv_sec == 0 && tsp->tv_nsec == 0)
1073 			timeout = -1;
1074 		else
1075 			timeout = atv.tv_sec > 24 * 60 * 60 ?
1076 			    24 * 60 * 60 * hz : tvtohz(&atv);
1077 		getmicrouptime(&rtv);
1078 		timevaladd(&atv, &rtv);
1079 	} else {
1080 		atv.tv_sec = 0;
1081 		atv.tv_usec = 0;
1082 		timeout = 0;
1083 	}
1084 	marker = knote_alloc(1);
1085 	if (marker == NULL) {
1086 		error = ENOMEM;
1087 		goto done_nl;
1088 	}
1089 	marker->kn_status = KN_MARKER;
1090 	KQ_LOCK(kq);
1091 	goto start;
1092 
1093 retry:
1094 	if (atv.tv_sec || atv.tv_usec) {
1095 		getmicrouptime(&rtv);
1096 		if (timevalcmp(&rtv, &atv, >=))
1097 			goto done;
1098 		ttv = atv;
1099 		timevalsub(&ttv, &rtv);
1100 		timeout = ttv.tv_sec > 24 * 60 * 60 ?
1101 			24 * 60 * 60 * hz : tvtohz(&ttv);
1102 	}
1103 
1104 start:
1105 	kevp = keva;
1106 	if (kq->kq_count == 0) {
1107 		if (timeout < 0) {
1108 			error = EWOULDBLOCK;
1109 		} else {
1110 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1111 			kq->kq_state |= KQ_SLEEP;
1112 			error = msleep(kq, &kq->kq_lock, PSOCK | PCATCH,
1113 			    "kqread", timeout);
1114 		}
1115 		if (error == 0)
1116 			goto retry;
1117 		/* don't restart after signals... */
1118 		if (error == ERESTART)
1119 			error = EINTR;
1120 		else if (error == EWOULDBLOCK)
1121 			error = 0;
1122 		goto done;
1123 	}
1124 
1125 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1126 	while (count) {
1127 		KQ_OWNED(kq);
1128 		kn = TAILQ_FIRST(&kq->kq_head);
1129 
1130 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1131 		    (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1132 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1133 			kq->kq_state |= KQ_FLUXWAIT;
1134 			error = msleep(kq, &kq->kq_lock, PSOCK,
1135 			    "kqflxwt", 0);
1136 			continue;
1137 		}
1138 
1139 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1140 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1141 			kn->kn_status &= ~KN_QUEUED;
1142 			kq->kq_count--;
1143 			continue;
1144 		}
1145 		if (kn == marker) {
1146 			KQ_FLUX_WAKEUP(kq);
1147 			if (count == maxevents)
1148 				goto retry;
1149 			goto done;
1150 		}
1151 		KASSERT((kn->kn_status & KN_INFLUX) == 0,
1152 		    ("KN_INFLUX set when not suppose to be"));
1153 
1154 		if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1155 			kn->kn_status &= ~KN_QUEUED;
1156 			kn->kn_status |= KN_INFLUX;
1157 			kq->kq_count--;
1158 			KQ_UNLOCK(kq);
1159 			/*
1160 			 * We don't need to lock the list since we've marked
1161 			 * it _INFLUX.
1162 			 */
1163 			*kevp = kn->kn_kevent;
1164 			if (!(kn->kn_status & KN_DETACHED))
1165 				kn->kn_fop->f_detach(kn);
1166 			knote_drop(kn, td);
1167 			KQ_LOCK(kq);
1168 			kn = NULL;
1169 		} else {
1170 			kn->kn_status |= KN_INFLUX;
1171 			KQ_UNLOCK(kq);
1172 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1173 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1174 			KN_LIST_LOCK(kn);
1175 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1176 				KN_LIST_UNLOCK(kn);
1177 				KQ_LOCK(kq);
1178 				kn->kn_status &=
1179 				    ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX);
1180 				kq->kq_count--;
1181 				continue;
1182 			}
1183 			*kevp = kn->kn_kevent;
1184 			KQ_LOCK(kq);
1185 			if (kn->kn_flags & EV_CLEAR) {
1186 				kn->kn_data = 0;
1187 				kn->kn_fflags = 0;
1188 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1189 				kq->kq_count--;
1190 			} else
1191 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1192 			KN_LIST_UNLOCK(kn);
1193 			kn->kn_status &= ~(KN_INFLUX);
1194 		}
1195 
1196 		/* we are returning a copy to the user */
1197 		kevp++;
1198 		nkev++;
1199 		count--;
1200 
1201 		if (nkev == KQ_NEVENTS) {
1202 			KQ_UNLOCK_FLUX(kq);
1203 			error = copyout(keva, ulistp, sizeof *keva * nkev);
1204 			ulistp += nkev;
1205 			nkev = 0;
1206 			kevp = keva;
1207 			KQ_LOCK(kq);
1208 			if (error)
1209 				break;
1210 		}
1211 	}
1212 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1213 done:
1214 	KQ_OWNED(kq);
1215 	KQ_UNLOCK_FLUX(kq);
1216 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1217 	knote_free(marker);
1218 done_nl:
1219 	KQ_NOTOWNED(kq);
1220 	if (nkev != 0)
1221 		error = copyout(keva, ulistp, sizeof *keva * nkev);
1222 	td->td_retval[0] = maxevents - count;
1223 	return (error);
1224 }
1225 
1226 /*
1227  * XXX
1228  * This could be expanded to call kqueue_scan, if desired.
1229  */
1230 /*ARGSUSED*/
1231 static int
1232 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
1233 	int flags, struct thread *td)
1234 {
1235 	return (ENXIO);
1236 }
1237 
1238 /*ARGSUSED*/
1239 static int
1240 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1241 	 int flags, struct thread *td)
1242 {
1243 	return (ENXIO);
1244 }
1245 
1246 /*ARGSUSED*/
1247 static int
1248 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1249 	struct ucred *active_cred, struct thread *td)
1250 {
1251 	/*
1252 	 * Enabling sigio causes two major problems:
1253 	 * 1) infinite recursion:
1254 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1255 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1256 	 * into itself over and over.  Sending the sigio causes the kqueue
1257 	 * to become ready, which in turn posts sigio again, forever.
1258 	 * Solution: this can be solved by setting a flag in the kqueue that
1259 	 * we have a SIGIO in progress.
1260 	 * 2) locking problems:
1261 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1262 	 * us above the proc and pgrp locks.
1263 	 * Solution: Post a signal using an async mechanism, being sure to
1264 	 * record a generation count in the delivery so that we do not deliver
1265 	 * a signal to the wrong process.
1266 	 *
1267 	 * Note, these two mechanisms are somewhat mutually exclusive!
1268 	 */
1269 #if 0
1270 	struct kqueue *kq;
1271 
1272 	kq = fp->f_data;
1273 	switch (cmd) {
1274 	case FIOASYNC:
1275 		if (*(int *)data) {
1276 			kq->kq_state |= KQ_ASYNC;
1277 		} else {
1278 			kq->kq_state &= ~KQ_ASYNC;
1279 		}
1280 		return (0);
1281 
1282 	case FIOSETOWN:
1283 		return (fsetown(*(int *)data, &kq->kq_sigio));
1284 
1285 	case FIOGETOWN:
1286 		*(int *)data = fgetown(&kq->kq_sigio);
1287 		return (0);
1288 	}
1289 #endif
1290 
1291 	return (ENOTTY);
1292 }
1293 
1294 /*ARGSUSED*/
1295 static int
1296 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1297 	struct thread *td)
1298 {
1299 	struct kqueue *kq;
1300 	int revents = 0;
1301 	int error;
1302 
1303 	if ((error = kqueue_aquire(fp, &kq)))
1304 		return POLLERR;
1305 
1306 	KQ_LOCK(kq);
1307 	if (events & (POLLIN | POLLRDNORM)) {
1308 		if (kq->kq_count) {
1309 			revents |= events & (POLLIN | POLLRDNORM);
1310 		} else {
1311 			selrecord(td, &kq->kq_sel);
1312 			kq->kq_state |= KQ_SEL;
1313 		}
1314 	}
1315 	kqueue_release(kq, 1);
1316 	KQ_UNLOCK(kq);
1317 	return (revents);
1318 }
1319 
1320 /*ARGSUSED*/
1321 static int
1322 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1323 	struct thread *td)
1324 {
1325 
1326 	return (ENXIO);
1327 }
1328 
1329 /*ARGSUSED*/
1330 static int
1331 kqueue_close(struct file *fp, struct thread *td)
1332 {
1333 	struct kqueue *kq = fp->f_data;
1334 	struct filedesc *fdp;
1335 	struct knote *kn;
1336 	int i;
1337 	int error;
1338 
1339 	if ((error = kqueue_aquire(fp, &kq)))
1340 		return error;
1341 
1342 	KQ_LOCK(kq);
1343 
1344 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1345 	    ("kqueue already closing"));
1346 	kq->kq_state |= KQ_CLOSING;
1347 	if (kq->kq_refcnt > 1)
1348 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1349 
1350 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1351 	fdp = kq->kq_fdp;
1352 
1353 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
1354 	    ("kqueue's knlist not empty"));
1355 
1356 	for (i = 0; i < kq->kq_knlistsize; i++) {
1357 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1358 			KASSERT((kn->kn_status & KN_INFLUX) == 0,
1359 			    ("KN_INFLUX set when not suppose to be"));
1360 			kn->kn_status |= KN_INFLUX;
1361 			KQ_UNLOCK(kq);
1362 			if (!(kn->kn_status & KN_DETACHED))
1363 				kn->kn_fop->f_detach(kn);
1364 			knote_drop(kn, td);
1365 			KQ_LOCK(kq);
1366 		}
1367 	}
1368 	if (kq->kq_knhashmask != 0) {
1369 		for (i = 0; i <= kq->kq_knhashmask; i++) {
1370 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1371 				KASSERT((kn->kn_status & KN_INFLUX) == 0,
1372 				    ("KN_INFLUX set when not suppose to be"));
1373 				kn->kn_status |= KN_INFLUX;
1374 				KQ_UNLOCK(kq);
1375 				if (!(kn->kn_status & KN_DETACHED))
1376 					kn->kn_fop->f_detach(kn);
1377 				knote_drop(kn, td);
1378 				KQ_LOCK(kq);
1379 			}
1380 		}
1381 	}
1382 
1383 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1384 		kq->kq_state |= KQ_TASKDRAIN;
1385 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1386 	}
1387 
1388 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1389 		kq->kq_state &= ~KQ_SEL;
1390 		selwakeuppri(&kq->kq_sel, PSOCK);
1391 	}
1392 
1393 	KQ_UNLOCK(kq);
1394 
1395 	FILEDESC_LOCK_FAST(fdp);
1396 	SLIST_REMOVE(&fdp->fd_kqlist, kq, kqueue, kq_list);
1397 	FILEDESC_UNLOCK_FAST(fdp);
1398 
1399 	knlist_destroy(&kq->kq_sel.si_note);
1400 	mtx_destroy(&kq->kq_lock);
1401 	kq->kq_fdp = NULL;
1402 
1403 	if (kq->kq_knhash != NULL)
1404 		free(kq->kq_knhash, M_KQUEUE);
1405 	if (kq->kq_knlist != NULL)
1406 		free(kq->kq_knlist, M_KQUEUE);
1407 
1408 	funsetown(&kq->kq_sigio);
1409 	free(kq, M_KQUEUE);
1410 	fp->f_data = NULL;
1411 
1412 	return (0);
1413 }
1414 
1415 static void
1416 kqueue_wakeup(struct kqueue *kq)
1417 {
1418 	KQ_OWNED(kq);
1419 
1420 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1421 		kq->kq_state &= ~KQ_SLEEP;
1422 		wakeup(kq);
1423 	}
1424 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1425 		kq->kq_state &= ~KQ_SEL;
1426 		selwakeuppri(&kq->kq_sel, PSOCK);
1427 	}
1428 	if (!knlist_empty(&kq->kq_sel.si_note))
1429 		kqueue_schedtask(kq);
1430 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1431 		pgsigio(&kq->kq_sigio, SIGIO, 0);
1432 	}
1433 }
1434 
1435 /*
1436  * Walk down a list of knotes, activating them if their event has triggered.
1437  *
1438  * There is a possibility to optimize in the case of one kq watching another.
1439  * Instead of scheduling a task to wake it up, you could pass enough state
1440  * down the chain to make up the parent kqueue.  Make this code functional
1441  * first.
1442  */
1443 void
1444 knote(struct knlist *list, long hint, int islocked)
1445 {
1446 	struct kqueue *kq;
1447 	struct knote *kn;
1448 
1449 	if (list == NULL)
1450 		return;
1451 
1452 	mtx_assert(list->kl_lock, islocked ? MA_OWNED : MA_NOTOWNED);
1453 	if (!islocked)
1454 		mtx_lock(list->kl_lock);
1455 	/*
1456 	 * If we unlock the list lock (and set KN_INFLUX), we can eliminate
1457 	 * the kqueue scheduling, but this will introduce four
1458 	 * lock/unlock's for each knote to test.  If we do, continue to use
1459 	 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is
1460 	 * only safe if you want to remove the current item, which we are
1461 	 * not doing.
1462 	 */
1463 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
1464 		kq = kn->kn_kq;
1465 		if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) {
1466 			KQ_LOCK(kq);
1467 			if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) {
1468 				kn->kn_status |= KN_HASKQLOCK;
1469 				if (kn->kn_fop->f_event(kn, hint))
1470 					KNOTE_ACTIVATE(kn, 1);
1471 				kn->kn_status &= ~KN_HASKQLOCK;
1472 			}
1473 			KQ_UNLOCK(kq);
1474 		}
1475 		kq = NULL;
1476 	}
1477 	if (!islocked)
1478 		mtx_unlock(list->kl_lock);
1479 }
1480 
1481 /*
1482  * add a knote to a knlist
1483  */
1484 void
1485 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1486 {
1487 	mtx_assert(knl->kl_lock, islocked ? MA_OWNED : MA_NOTOWNED);
1488 	KQ_NOTOWNED(kn->kn_kq);
1489 	KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1490 	    (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1491 	if (!islocked)
1492 		mtx_lock(knl->kl_lock);
1493 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1494 	if (!islocked)
1495 		mtx_unlock(knl->kl_lock);
1496 	KQ_LOCK(kn->kn_kq);
1497 	kn->kn_knlist = knl;
1498 	kn->kn_status &= ~KN_DETACHED;
1499 	KQ_UNLOCK(kn->kn_kq);
1500 }
1501 
1502 static void
1503 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1504 {
1505 	KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1506 	mtx_assert(knl->kl_lock, knlislocked ? MA_OWNED : MA_NOTOWNED);
1507 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1508 	if (!kqislocked)
1509 		KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1510     ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1511 	if (!knlislocked)
1512 		mtx_lock(knl->kl_lock);
1513 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1514 	kn->kn_knlist = NULL;
1515 	if (!knlislocked)
1516 		mtx_unlock(knl->kl_lock);
1517 	if (!kqislocked)
1518 		KQ_LOCK(kn->kn_kq);
1519 	kn->kn_status |= KN_DETACHED;
1520 	if (!kqislocked)
1521 		KQ_UNLOCK(kn->kn_kq);
1522 }
1523 
1524 /*
1525  * remove all knotes from a specified klist
1526  */
1527 void
1528 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1529 {
1530 
1531 	knlist_remove_kq(knl, kn, islocked, 0);
1532 }
1533 
1534 /*
1535  * remove knote from a specified klist while in f_event handler.
1536  */
1537 void
1538 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1539 {
1540 
1541 	knlist_remove_kq(knl, kn, 1,
1542 	    (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1543 }
1544 
1545 int
1546 knlist_empty(struct knlist *knl)
1547 {
1548 
1549 	mtx_assert(knl->kl_lock, MA_OWNED);
1550 	return SLIST_EMPTY(&knl->kl_list);
1551 }
1552 
1553 static struct mtx	knlist_lock;
1554 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
1555 	MTX_DEF);
1556 
1557 void
1558 knlist_init(struct knlist *knl, struct mtx *mtx)
1559 {
1560 
1561 	if (mtx == NULL)
1562 		knl->kl_lock = &knlist_lock;
1563 	else
1564 		knl->kl_lock = mtx;
1565 
1566 	SLIST_INIT(&knl->kl_list);
1567 }
1568 
1569 void
1570 knlist_destroy(struct knlist *knl)
1571 {
1572 
1573 #ifdef INVARIANTS
1574 	/*
1575 	 * if we run across this error, we need to find the offending
1576 	 * driver and have it call knlist_clear.
1577 	 */
1578 	if (!SLIST_EMPTY(&knl->kl_list))
1579 		printf("WARNING: destroying knlist w/ knotes on it!\n");
1580 #endif
1581 
1582 	knl->kl_lock = NULL;
1583 	SLIST_INIT(&knl->kl_list);
1584 }
1585 
1586 /*
1587  * Even if we are locked, we may need to drop the lock to allow any influx
1588  * knotes time to "settle".
1589  */
1590 void
1591 knlist_clear(struct knlist *knl, int islocked)
1592 {
1593 	struct knote *kn;
1594 	struct kqueue *kq;
1595 
1596 	if (islocked)
1597 		mtx_assert(knl->kl_lock, MA_OWNED);
1598 	else {
1599 		mtx_assert(knl->kl_lock, MA_NOTOWNED);
1600 again:		/* need to reaquire lock since we have dropped it */
1601 		mtx_lock(knl->kl_lock);
1602 	}
1603 
1604 	SLIST_FOREACH(kn, &knl->kl_list, kn_selnext) {
1605 		kq = kn->kn_kq;
1606 		KQ_LOCK(kq);
1607 		if ((kn->kn_status & KN_INFLUX) &&
1608 		    (kn->kn_status & KN_DETACHED) != KN_DETACHED) {
1609 			KQ_UNLOCK(kq);
1610 			continue;
1611 		}
1612 		/* Make sure cleared knotes disappear soon */
1613 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
1614 		knlist_remove_kq(knl, kn, 1, 1);
1615 		KQ_UNLOCK(kq);
1616 		kq = NULL;
1617 	}
1618 
1619 	if (!SLIST_EMPTY(&knl->kl_list)) {
1620 		/* there are still KN_INFLUX remaining */
1621 		kn = SLIST_FIRST(&knl->kl_list);
1622 		kq = kn->kn_kq;
1623 		KQ_LOCK(kq);
1624 		KASSERT(kn->kn_status & KN_INFLUX,
1625 		    ("knote removed w/o list lock"));
1626 		mtx_unlock(knl->kl_lock);
1627 		kq->kq_state |= KQ_FLUXWAIT;
1628 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
1629 		kq = NULL;
1630 		goto again;
1631 	}
1632 
1633 	SLIST_INIT(&knl->kl_list);
1634 
1635 	if (islocked)
1636 		mtx_assert(knl->kl_lock, MA_OWNED);
1637 	else {
1638 		mtx_unlock(knl->kl_lock);
1639 		mtx_assert(knl->kl_lock, MA_NOTOWNED);
1640 	}
1641 }
1642 
1643 /*
1644  * remove all knotes referencing a specified fd
1645  * must be called with FILEDESC lock.  This prevents a race where a new fd
1646  * comes along and occupies the entry and we attach a knote to the fd.
1647  */
1648 void
1649 knote_fdclose(struct thread *td, int fd)
1650 {
1651 	struct filedesc *fdp = td->td_proc->p_fd;
1652 	struct kqueue *kq;
1653 	struct knote *kn;
1654 	int influx;
1655 
1656 	FILEDESC_LOCK_ASSERT(fdp, MA_OWNED);
1657 
1658 	/*
1659 	 * We shouldn't have to worry about new kevents appearing on fd
1660 	 * since filedesc is locked.
1661 	 */
1662 	SLIST_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
1663 		KQ_LOCK(kq);
1664 
1665 again:
1666 		influx = 0;
1667 		while (kq->kq_knlistsize > fd &&
1668 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
1669 			if (kn->kn_status & KN_INFLUX) {
1670 				/* someone else might be waiting on our knote */
1671 				if (influx)
1672 					wakeup(kq);
1673 				kq->kq_state |= KQ_FLUXWAIT;
1674 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
1675 				goto again;
1676 			}
1677 			kn->kn_status |= KN_INFLUX;
1678 			KQ_UNLOCK(kq);
1679 			if (!(kn->kn_status & KN_DETACHED))
1680 				kn->kn_fop->f_detach(kn);
1681 			knote_drop(kn, td);
1682 			influx = 1;
1683 			KQ_LOCK(kq);
1684 		}
1685 		KQ_UNLOCK_FLUX(kq);
1686 	}
1687 }
1688 
1689 static int
1690 knote_attach(struct knote *kn, struct kqueue *kq)
1691 {
1692 	struct klist *list;
1693 
1694 	KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
1695 	KQ_OWNED(kq);
1696 
1697 	if (kn->kn_fop->f_isfd) {
1698 		if (kn->kn_id >= kq->kq_knlistsize)
1699 			return ENOMEM;
1700 		list = &kq->kq_knlist[kn->kn_id];
1701 	} else {
1702 		if (kq->kq_knhash == NULL)
1703 			return ENOMEM;
1704 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1705 	}
1706 
1707 	SLIST_INSERT_HEAD(list, kn, kn_link);
1708 
1709 	return 0;
1710 }
1711 
1712 /*
1713  * knote must already have been detatched using the f_detach method.
1714  * no lock need to be held, it is assumed that the KN_INFLUX flag is set
1715  * to prevent other removal.
1716  */
1717 static void
1718 knote_drop(struct knote *kn, struct thread *td)
1719 {
1720 	struct kqueue *kq;
1721 	struct klist *list;
1722 
1723 	kq = kn->kn_kq;
1724 
1725 	KQ_NOTOWNED(kq);
1726 	KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
1727 	    ("knote_drop called without KN_INFLUX set in kn_status"));
1728 
1729 	KQ_LOCK(kq);
1730 	if (kn->kn_fop->f_isfd)
1731 		list = &kq->kq_knlist[kn->kn_id];
1732 	else
1733 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1734 
1735 	SLIST_REMOVE(list, kn, knote, kn_link);
1736 	if (kn->kn_status & KN_QUEUED)
1737 		knote_dequeue(kn);
1738 	KQ_UNLOCK_FLUX(kq);
1739 
1740 	if (kn->kn_fop->f_isfd) {
1741 		fdrop(kn->kn_fp, td);
1742 		kn->kn_fp = NULL;
1743 	}
1744 	kqueue_fo_release(kn->kn_kevent.filter);
1745 	kn->kn_fop = NULL;
1746 	knote_free(kn);
1747 }
1748 
1749 static void
1750 knote_enqueue(struct knote *kn)
1751 {
1752 	struct kqueue *kq = kn->kn_kq;
1753 
1754 	KQ_OWNED(kn->kn_kq);
1755 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
1756 
1757 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1758 	kn->kn_status |= KN_QUEUED;
1759 	kq->kq_count++;
1760 	kqueue_wakeup(kq);
1761 }
1762 
1763 static void
1764 knote_dequeue(struct knote *kn)
1765 {
1766 	struct kqueue *kq = kn->kn_kq;
1767 
1768 	KQ_OWNED(kn->kn_kq);
1769 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
1770 
1771 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1772 	kn->kn_status &= ~KN_QUEUED;
1773 	kq->kq_count--;
1774 }
1775 
1776 static void
1777 knote_init(void)
1778 {
1779 
1780 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
1781 	    NULL, NULL, UMA_ALIGN_PTR, 0);
1782 }
1783 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL)
1784 
1785 static struct knote *
1786 knote_alloc(int waitok)
1787 {
1788 	return ((struct knote *)uma_zalloc(knote_zone,
1789 	    (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
1790 }
1791 
1792 static void
1793 knote_free(struct knote *kn)
1794 {
1795 	if (kn != NULL)
1796 		uma_zfree(knote_zone, kn);
1797 }
1798