xref: /freebsd/sys/kern/kern_event.c (revision 63cbe8d1d95f97e93929ec66f1138693d08dd9f6)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
5  * Copyright 2004 John-Mark Gurney <jmg@FreeBSD.org>
6  * Copyright (c) 2009 Apple, Inc.
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include "opt_ktrace.h"
35 #include "opt_kqueue.h"
36 
37 #ifdef COMPAT_FREEBSD11
38 #define	_WANT_FREEBSD11_KEVENT
39 #endif
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/capsicum.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/rwlock.h>
48 #include <sys/proc.h>
49 #include <sys/malloc.h>
50 #include <sys/unistd.h>
51 #include <sys/file.h>
52 #include <sys/filedesc.h>
53 #include <sys/filio.h>
54 #include <sys/fcntl.h>
55 #include <sys/kthread.h>
56 #include <sys/selinfo.h>
57 #include <sys/queue.h>
58 #include <sys/event.h>
59 #include <sys/eventvar.h>
60 #include <sys/poll.h>
61 #include <sys/protosw.h>
62 #include <sys/resourcevar.h>
63 #include <sys/sigio.h>
64 #include <sys/signalvar.h>
65 #include <sys/socket.h>
66 #include <sys/socketvar.h>
67 #include <sys/stat.h>
68 #include <sys/sysctl.h>
69 #include <sys/sysproto.h>
70 #include <sys/syscallsubr.h>
71 #include <sys/taskqueue.h>
72 #include <sys/uio.h>
73 #include <sys/user.h>
74 #ifdef KTRACE
75 #include <sys/ktrace.h>
76 #endif
77 #include <machine/atomic.h>
78 
79 #include <vm/uma.h>
80 
81 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
82 
83 /*
84  * This lock is used if multiple kq locks are required.  This possibly
85  * should be made into a per proc lock.
86  */
87 static struct mtx	kq_global;
88 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
89 #define KQ_GLOBAL_LOCK(lck, haslck)	do {	\
90 	if (!haslck)				\
91 		mtx_lock(lck);			\
92 	haslck = 1;				\
93 } while (0)
94 #define KQ_GLOBAL_UNLOCK(lck, haslck)	do {	\
95 	if (haslck)				\
96 		mtx_unlock(lck);			\
97 	haslck = 0;				\
98 } while (0)
99 
100 TASKQUEUE_DEFINE_THREAD(kqueue_ctx);
101 
102 static int	kevent_copyout(void *arg, struct kevent *kevp, int count);
103 static int	kevent_copyin(void *arg, struct kevent *kevp, int count);
104 static int	kqueue_register(struct kqueue *kq, struct kevent *kev,
105 		    struct thread *td, int waitok);
106 static int	kqueue_acquire(struct file *fp, struct kqueue **kqp);
107 static void	kqueue_release(struct kqueue *kq, int locked);
108 static void	kqueue_destroy(struct kqueue *kq);
109 static void	kqueue_drain(struct kqueue *kq, struct thread *td);
110 static int	kqueue_expand(struct kqueue *kq, struct filterops *fops,
111 		    uintptr_t ident, int waitok);
112 static void	kqueue_task(void *arg, int pending);
113 static int	kqueue_scan(struct kqueue *kq, int maxevents,
114 		    struct kevent_copyops *k_ops,
115 		    const struct timespec *timeout,
116 		    struct kevent *keva, struct thread *td);
117 static void 	kqueue_wakeup(struct kqueue *kq);
118 static struct filterops *kqueue_fo_find(int filt);
119 static void	kqueue_fo_release(int filt);
120 struct g_kevent_args;
121 static int	kern_kevent_generic(struct thread *td,
122 		    struct g_kevent_args *uap,
123 		    struct kevent_copyops *k_ops, const char *struct_name);
124 
125 static fo_ioctl_t	kqueue_ioctl;
126 static fo_poll_t	kqueue_poll;
127 static fo_kqfilter_t	kqueue_kqfilter;
128 static fo_stat_t	kqueue_stat;
129 static fo_close_t	kqueue_close;
130 static fo_fill_kinfo_t	kqueue_fill_kinfo;
131 
132 static struct fileops kqueueops = {
133 	.fo_read = invfo_rdwr,
134 	.fo_write = invfo_rdwr,
135 	.fo_truncate = invfo_truncate,
136 	.fo_ioctl = kqueue_ioctl,
137 	.fo_poll = kqueue_poll,
138 	.fo_kqfilter = kqueue_kqfilter,
139 	.fo_stat = kqueue_stat,
140 	.fo_close = kqueue_close,
141 	.fo_chmod = invfo_chmod,
142 	.fo_chown = invfo_chown,
143 	.fo_sendfile = invfo_sendfile,
144 	.fo_fill_kinfo = kqueue_fill_kinfo,
145 };
146 
147 static int 	knote_attach(struct knote *kn, struct kqueue *kq);
148 static void 	knote_drop(struct knote *kn, struct thread *td);
149 static void 	knote_drop_detached(struct knote *kn, struct thread *td);
150 static void 	knote_enqueue(struct knote *kn);
151 static void 	knote_dequeue(struct knote *kn);
152 static void 	knote_init(void);
153 static struct 	knote *knote_alloc(int waitok);
154 static void 	knote_free(struct knote *kn);
155 
156 static void	filt_kqdetach(struct knote *kn);
157 static int	filt_kqueue(struct knote *kn, long hint);
158 static int	filt_procattach(struct knote *kn);
159 static void	filt_procdetach(struct knote *kn);
160 static int	filt_proc(struct knote *kn, long hint);
161 static int	filt_fileattach(struct knote *kn);
162 static void	filt_timerexpire(void *knx);
163 static int	filt_timerattach(struct knote *kn);
164 static void	filt_timerdetach(struct knote *kn);
165 static void	filt_timerstart(struct knote *kn, sbintime_t to);
166 static void	filt_timertouch(struct knote *kn, struct kevent *kev,
167 		    u_long type);
168 static int	filt_timervalidate(struct knote *kn, sbintime_t *to);
169 static int	filt_timer(struct knote *kn, long hint);
170 static int	filt_userattach(struct knote *kn);
171 static void	filt_userdetach(struct knote *kn);
172 static int	filt_user(struct knote *kn, long hint);
173 static void	filt_usertouch(struct knote *kn, struct kevent *kev,
174 		    u_long type);
175 
176 static struct filterops file_filtops = {
177 	.f_isfd = 1,
178 	.f_attach = filt_fileattach,
179 };
180 static struct filterops kqread_filtops = {
181 	.f_isfd = 1,
182 	.f_detach = filt_kqdetach,
183 	.f_event = filt_kqueue,
184 };
185 /* XXX - move to kern_proc.c?  */
186 static struct filterops proc_filtops = {
187 	.f_isfd = 0,
188 	.f_attach = filt_procattach,
189 	.f_detach = filt_procdetach,
190 	.f_event = filt_proc,
191 };
192 static struct filterops timer_filtops = {
193 	.f_isfd = 0,
194 	.f_attach = filt_timerattach,
195 	.f_detach = filt_timerdetach,
196 	.f_event = filt_timer,
197 	.f_touch = filt_timertouch,
198 };
199 static struct filterops user_filtops = {
200 	.f_attach = filt_userattach,
201 	.f_detach = filt_userdetach,
202 	.f_event = filt_user,
203 	.f_touch = filt_usertouch,
204 };
205 
206 static uma_zone_t	knote_zone;
207 static unsigned int	kq_ncallouts = 0;
208 static unsigned int 	kq_calloutmax = 4 * 1024;
209 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
210     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
211 
212 /* XXX - ensure not influx ? */
213 #define KNOTE_ACTIVATE(kn, islock) do { 				\
214 	if ((islock))							\
215 		mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED);		\
216 	else								\
217 		KQ_LOCK((kn)->kn_kq);					\
218 	(kn)->kn_status |= KN_ACTIVE;					\
219 	if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)		\
220 		knote_enqueue((kn));					\
221 	if (!(islock))							\
222 		KQ_UNLOCK((kn)->kn_kq);					\
223 } while(0)
224 #define KQ_LOCK(kq) do {						\
225 	mtx_lock(&(kq)->kq_lock);					\
226 } while (0)
227 #define KQ_FLUX_WAKEUP(kq) do {						\
228 	if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) {		\
229 		(kq)->kq_state &= ~KQ_FLUXWAIT;				\
230 		wakeup((kq));						\
231 	}								\
232 } while (0)
233 #define KQ_UNLOCK_FLUX(kq) do {						\
234 	KQ_FLUX_WAKEUP(kq);						\
235 	mtx_unlock(&(kq)->kq_lock);					\
236 } while (0)
237 #define KQ_UNLOCK(kq) do {						\
238 	mtx_unlock(&(kq)->kq_lock);					\
239 } while (0)
240 #define KQ_OWNED(kq) do {						\
241 	mtx_assert(&(kq)->kq_lock, MA_OWNED);				\
242 } while (0)
243 #define KQ_NOTOWNED(kq) do {						\
244 	mtx_assert(&(kq)->kq_lock, MA_NOTOWNED);			\
245 } while (0)
246 
247 static struct knlist *
248 kn_list_lock(struct knote *kn)
249 {
250 	struct knlist *knl;
251 
252 	knl = kn->kn_knlist;
253 	if (knl != NULL)
254 		knl->kl_lock(knl->kl_lockarg);
255 	return (knl);
256 }
257 
258 static void
259 kn_list_unlock(struct knlist *knl)
260 {
261 	bool do_free;
262 
263 	if (knl == NULL)
264 		return;
265 	do_free = knl->kl_autodestroy && knlist_empty(knl);
266 	knl->kl_unlock(knl->kl_lockarg);
267 	if (do_free) {
268 		knlist_destroy(knl);
269 		free(knl, M_KQUEUE);
270 	}
271 }
272 
273 static bool
274 kn_in_flux(struct knote *kn)
275 {
276 
277 	return (kn->kn_influx > 0);
278 }
279 
280 static void
281 kn_enter_flux(struct knote *kn)
282 {
283 
284 	KQ_OWNED(kn->kn_kq);
285 	MPASS(kn->kn_influx < INT_MAX);
286 	kn->kn_influx++;
287 }
288 
289 static bool
290 kn_leave_flux(struct knote *kn)
291 {
292 
293 	KQ_OWNED(kn->kn_kq);
294 	MPASS(kn->kn_influx > 0);
295 	kn->kn_influx--;
296 	return (kn->kn_influx == 0);
297 }
298 
299 #define	KNL_ASSERT_LOCK(knl, islocked) do {				\
300 	if (islocked)							\
301 		KNL_ASSERT_LOCKED(knl);				\
302 	else								\
303 		KNL_ASSERT_UNLOCKED(knl);				\
304 } while (0)
305 #ifdef INVARIANTS
306 #define	KNL_ASSERT_LOCKED(knl) do {					\
307 	knl->kl_assert_locked((knl)->kl_lockarg);			\
308 } while (0)
309 #define	KNL_ASSERT_UNLOCKED(knl) do {					\
310 	knl->kl_assert_unlocked((knl)->kl_lockarg);			\
311 } while (0)
312 #else /* !INVARIANTS */
313 #define	KNL_ASSERT_LOCKED(knl) do {} while(0)
314 #define	KNL_ASSERT_UNLOCKED(knl) do {} while (0)
315 #endif /* INVARIANTS */
316 
317 #ifndef	KN_HASHSIZE
318 #define	KN_HASHSIZE		64		/* XXX should be tunable */
319 #endif
320 
321 #define KN_HASH(val, mask)	(((val) ^ (val >> 8)) & (mask))
322 
323 static int
324 filt_nullattach(struct knote *kn)
325 {
326 
327 	return (ENXIO);
328 };
329 
330 struct filterops null_filtops = {
331 	.f_isfd = 0,
332 	.f_attach = filt_nullattach,
333 };
334 
335 /* XXX - make SYSINIT to add these, and move into respective modules. */
336 extern struct filterops sig_filtops;
337 extern struct filterops fs_filtops;
338 
339 /*
340  * Table for for all system-defined filters.
341  */
342 static struct mtx	filterops_lock;
343 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
344 	MTX_DEF);
345 static struct {
346 	struct filterops *for_fop;
347 	int for_nolock;
348 	int for_refcnt;
349 } sysfilt_ops[EVFILT_SYSCOUNT] = {
350 	{ &file_filtops, 1 },			/* EVFILT_READ */
351 	{ &file_filtops, 1 },			/* EVFILT_WRITE */
352 	{ &null_filtops },			/* EVFILT_AIO */
353 	{ &file_filtops, 1 },			/* EVFILT_VNODE */
354 	{ &proc_filtops, 1 },			/* EVFILT_PROC */
355 	{ &sig_filtops, 1 },			/* EVFILT_SIGNAL */
356 	{ &timer_filtops, 1 },			/* EVFILT_TIMER */
357 	{ &file_filtops, 1 },			/* EVFILT_PROCDESC */
358 	{ &fs_filtops, 1 },			/* EVFILT_FS */
359 	{ &null_filtops },			/* EVFILT_LIO */
360 	{ &user_filtops, 1 },			/* EVFILT_USER */
361 	{ &null_filtops },			/* EVFILT_SENDFILE */
362 	{ &file_filtops, 1 },                   /* EVFILT_EMPTY */
363 };
364 
365 /*
366  * Simple redirection for all cdevsw style objects to call their fo_kqfilter
367  * method.
368  */
369 static int
370 filt_fileattach(struct knote *kn)
371 {
372 
373 	return (fo_kqfilter(kn->kn_fp, kn));
374 }
375 
376 /*ARGSUSED*/
377 static int
378 kqueue_kqfilter(struct file *fp, struct knote *kn)
379 {
380 	struct kqueue *kq = kn->kn_fp->f_data;
381 
382 	if (kn->kn_filter != EVFILT_READ)
383 		return (EINVAL);
384 
385 	kn->kn_status |= KN_KQUEUE;
386 	kn->kn_fop = &kqread_filtops;
387 	knlist_add(&kq->kq_sel.si_note, kn, 0);
388 
389 	return (0);
390 }
391 
392 static void
393 filt_kqdetach(struct knote *kn)
394 {
395 	struct kqueue *kq = kn->kn_fp->f_data;
396 
397 	knlist_remove(&kq->kq_sel.si_note, kn, 0);
398 }
399 
400 /*ARGSUSED*/
401 static int
402 filt_kqueue(struct knote *kn, long hint)
403 {
404 	struct kqueue *kq = kn->kn_fp->f_data;
405 
406 	kn->kn_data = kq->kq_count;
407 	return (kn->kn_data > 0);
408 }
409 
410 /* XXX - move to kern_proc.c?  */
411 static int
412 filt_procattach(struct knote *kn)
413 {
414 	struct proc *p;
415 	int error;
416 	bool exiting, immediate;
417 
418 	exiting = immediate = false;
419 	if (kn->kn_sfflags & NOTE_EXIT)
420 		p = pfind_any(kn->kn_id);
421 	else
422 		p = pfind(kn->kn_id);
423 	if (p == NULL)
424 		return (ESRCH);
425 	if (p->p_flag & P_WEXIT)
426 		exiting = true;
427 
428 	if ((error = p_cansee(curthread, p))) {
429 		PROC_UNLOCK(p);
430 		return (error);
431 	}
432 
433 	kn->kn_ptr.p_proc = p;
434 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
435 
436 	/*
437 	 * Internal flag indicating registration done by kernel for the
438 	 * purposes of getting a NOTE_CHILD notification.
439 	 */
440 	if (kn->kn_flags & EV_FLAG2) {
441 		kn->kn_flags &= ~EV_FLAG2;
442 		kn->kn_data = kn->kn_sdata;		/* ppid */
443 		kn->kn_fflags = NOTE_CHILD;
444 		kn->kn_sfflags &= ~(NOTE_EXIT | NOTE_EXEC | NOTE_FORK);
445 		immediate = true; /* Force immediate activation of child note. */
446 	}
447 	/*
448 	 * Internal flag indicating registration done by kernel (for other than
449 	 * NOTE_CHILD).
450 	 */
451 	if (kn->kn_flags & EV_FLAG1) {
452 		kn->kn_flags &= ~EV_FLAG1;
453 	}
454 
455 	knlist_add(p->p_klist, kn, 1);
456 
457 	/*
458 	 * Immediately activate any child notes or, in the case of a zombie
459 	 * target process, exit notes.  The latter is necessary to handle the
460 	 * case where the target process, e.g. a child, dies before the kevent
461 	 * is registered.
462 	 */
463 	if (immediate || (exiting && filt_proc(kn, NOTE_EXIT)))
464 		KNOTE_ACTIVATE(kn, 0);
465 
466 	PROC_UNLOCK(p);
467 
468 	return (0);
469 }
470 
471 /*
472  * The knote may be attached to a different process, which may exit,
473  * leaving nothing for the knote to be attached to.  So when the process
474  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
475  * it will be deleted when read out.  However, as part of the knote deletion,
476  * this routine is called, so a check is needed to avoid actually performing
477  * a detach, because the original process does not exist any more.
478  */
479 /* XXX - move to kern_proc.c?  */
480 static void
481 filt_procdetach(struct knote *kn)
482 {
483 
484 	knlist_remove(kn->kn_knlist, kn, 0);
485 	kn->kn_ptr.p_proc = NULL;
486 }
487 
488 /* XXX - move to kern_proc.c?  */
489 static int
490 filt_proc(struct knote *kn, long hint)
491 {
492 	struct proc *p;
493 	u_int event;
494 
495 	p = kn->kn_ptr.p_proc;
496 	if (p == NULL) /* already activated, from attach filter */
497 		return (0);
498 
499 	/* Mask off extra data. */
500 	event = (u_int)hint & NOTE_PCTRLMASK;
501 
502 	/* If the user is interested in this event, record it. */
503 	if (kn->kn_sfflags & event)
504 		kn->kn_fflags |= event;
505 
506 	/* Process is gone, so flag the event as finished. */
507 	if (event == NOTE_EXIT) {
508 		kn->kn_flags |= EV_EOF | EV_ONESHOT;
509 		kn->kn_ptr.p_proc = NULL;
510 		if (kn->kn_fflags & NOTE_EXIT)
511 			kn->kn_data = KW_EXITCODE(p->p_xexit, p->p_xsig);
512 		if (kn->kn_fflags == 0)
513 			kn->kn_flags |= EV_DROP;
514 		return (1);
515 	}
516 
517 	return (kn->kn_fflags != 0);
518 }
519 
520 /*
521  * Called when the process forked. It mostly does the same as the
522  * knote(), activating all knotes registered to be activated when the
523  * process forked. Additionally, for each knote attached to the
524  * parent, check whether user wants to track the new process. If so
525  * attach a new knote to it, and immediately report an event with the
526  * child's pid.
527  */
528 void
529 knote_fork(struct knlist *list, int pid)
530 {
531 	struct kqueue *kq;
532 	struct knote *kn;
533 	struct kevent kev;
534 	int error;
535 
536 	if (list == NULL)
537 		return;
538 	list->kl_lock(list->kl_lockarg);
539 
540 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
541 		kq = kn->kn_kq;
542 		KQ_LOCK(kq);
543 		if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
544 			KQ_UNLOCK(kq);
545 			continue;
546 		}
547 
548 		/*
549 		 * The same as knote(), activate the event.
550 		 */
551 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
552 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
553 				KNOTE_ACTIVATE(kn, 1);
554 			KQ_UNLOCK(kq);
555 			continue;
556 		}
557 
558 		/*
559 		 * The NOTE_TRACK case. In addition to the activation
560 		 * of the event, we need to register new events to
561 		 * track the child. Drop the locks in preparation for
562 		 * the call to kqueue_register().
563 		 */
564 		kn_enter_flux(kn);
565 		KQ_UNLOCK(kq);
566 		list->kl_unlock(list->kl_lockarg);
567 
568 		/*
569 		 * Activate existing knote and register tracking knotes with
570 		 * new process.
571 		 *
572 		 * First register a knote to get just the child notice. This
573 		 * must be a separate note from a potential NOTE_EXIT
574 		 * notification since both NOTE_CHILD and NOTE_EXIT are defined
575 		 * to use the data field (in conflicting ways).
576 		 */
577 		kev.ident = pid;
578 		kev.filter = kn->kn_filter;
579 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_ONESHOT |
580 		    EV_FLAG2;
581 		kev.fflags = kn->kn_sfflags;
582 		kev.data = kn->kn_id;		/* parent */
583 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
584 		error = kqueue_register(kq, &kev, NULL, 0);
585 		if (error)
586 			kn->kn_fflags |= NOTE_TRACKERR;
587 
588 		/*
589 		 * Then register another knote to track other potential events
590 		 * from the new process.
591 		 */
592 		kev.ident = pid;
593 		kev.filter = kn->kn_filter;
594 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
595 		kev.fflags = kn->kn_sfflags;
596 		kev.data = kn->kn_id;		/* parent */
597 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
598 		error = kqueue_register(kq, &kev, NULL, 0);
599 		if (error)
600 			kn->kn_fflags |= NOTE_TRACKERR;
601 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
602 			KNOTE_ACTIVATE(kn, 0);
603 		KQ_LOCK(kq);
604 		kn_leave_flux(kn);
605 		KQ_UNLOCK_FLUX(kq);
606 		list->kl_lock(list->kl_lockarg);
607 	}
608 	list->kl_unlock(list->kl_lockarg);
609 }
610 
611 /*
612  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
613  * interval timer support code.
614  */
615 
616 #define NOTE_TIMER_PRECMASK						\
617     (NOTE_SECONDS | NOTE_MSECONDS | NOTE_USECONDS | NOTE_NSECONDS)
618 
619 static sbintime_t
620 timer2sbintime(intptr_t data, int flags)
621 {
622 	int64_t secs;
623 
624         /*
625          * Macros for converting to the fractional second portion of an
626          * sbintime_t using 64bit multiplication to improve precision.
627          */
628 #define NS_TO_SBT(ns) (((ns) * (((uint64_t)1 << 63) / 500000000)) >> 32)
629 #define US_TO_SBT(us) (((us) * (((uint64_t)1 << 63) / 500000)) >> 32)
630 #define MS_TO_SBT(ms) (((ms) * (((uint64_t)1 << 63) / 500)) >> 32)
631 	switch (flags & NOTE_TIMER_PRECMASK) {
632 	case NOTE_SECONDS:
633 #ifdef __LP64__
634 		if (data > (SBT_MAX / SBT_1S))
635 			return (SBT_MAX);
636 #endif
637 		return ((sbintime_t)data << 32);
638 	case NOTE_MSECONDS: /* FALLTHROUGH */
639 	case 0:
640 		if (data >= 1000) {
641 			secs = data / 1000;
642 #ifdef __LP64__
643 			if (secs > (SBT_MAX / SBT_1S))
644 				return (SBT_MAX);
645 #endif
646 			return (secs << 32 | MS_TO_SBT(data % 1000));
647 		}
648 		return (MS_TO_SBT(data));
649 	case NOTE_USECONDS:
650 		if (data >= 1000000) {
651 			secs = data / 1000000;
652 #ifdef __LP64__
653 			if (secs > (SBT_MAX / SBT_1S))
654 				return (SBT_MAX);
655 #endif
656 			return (secs << 32 | US_TO_SBT(data % 1000000));
657 		}
658 		return (US_TO_SBT(data));
659 	case NOTE_NSECONDS:
660 		if (data >= 1000000000) {
661 			secs = data / 1000000000;
662 #ifdef __LP64__
663 			if (secs > (SBT_MAX / SBT_1S))
664 				return (SBT_MAX);
665 #endif
666 			return (secs << 32 | US_TO_SBT(data % 1000000000));
667 		}
668 		return (NS_TO_SBT(data));
669 	default:
670 		break;
671 	}
672 	return (-1);
673 }
674 
675 struct kq_timer_cb_data {
676 	struct callout c;
677 	sbintime_t next;	/* next timer event fires at */
678 	sbintime_t to;		/* precalculated timer period, 0 for abs */
679 };
680 
681 static void
682 filt_timerexpire(void *knx)
683 {
684 	struct knote *kn;
685 	struct kq_timer_cb_data *kc;
686 
687 	kn = knx;
688 	kn->kn_data++;
689 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
690 
691 	if ((kn->kn_flags & EV_ONESHOT) != 0)
692 		return;
693 	kc = kn->kn_ptr.p_v;
694 	if (kc->to == 0)
695 		return;
696 	kc->next += kc->to;
697 	callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
698 	    PCPU_GET(cpuid), C_ABSOLUTE);
699 }
700 
701 /*
702  * data contains amount of time to sleep
703  */
704 static int
705 filt_timervalidate(struct knote *kn, sbintime_t *to)
706 {
707 	struct bintime bt;
708 	sbintime_t sbt;
709 
710 	if (kn->kn_sdata < 0)
711 		return (EINVAL);
712 	if (kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
713 		kn->kn_sdata = 1;
714 	/*
715 	 * The only fflags values supported are the timer unit
716 	 * (precision) and the absolute time indicator.
717 	 */
718 	if ((kn->kn_sfflags & ~(NOTE_TIMER_PRECMASK | NOTE_ABSTIME)) != 0)
719 		return (EINVAL);
720 
721 	*to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
722 	if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
723 		getboottimebin(&bt);
724 		sbt = bttosbt(bt);
725 		*to -= sbt;
726 	}
727 	if (*to < 0)
728 		return (EINVAL);
729 	return (0);
730 }
731 
732 static int
733 filt_timerattach(struct knote *kn)
734 {
735 	struct kq_timer_cb_data *kc;
736 	sbintime_t to;
737 	unsigned int ncallouts;
738 	int error;
739 
740 	error = filt_timervalidate(kn, &to);
741 	if (error != 0)
742 		return (error);
743 
744 	do {
745 		ncallouts = kq_ncallouts;
746 		if (ncallouts >= kq_calloutmax)
747 			return (ENOMEM);
748 	} while (!atomic_cmpset_int(&kq_ncallouts, ncallouts, ncallouts + 1));
749 
750 	if ((kn->kn_sfflags & NOTE_ABSTIME) == 0)
751 		kn->kn_flags |= EV_CLEAR;	/* automatically set */
752 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
753 	kn->kn_ptr.p_v = kc = malloc(sizeof(*kc), M_KQUEUE, M_WAITOK);
754 	callout_init(&kc->c, 1);
755 	filt_timerstart(kn, to);
756 
757 	return (0);
758 }
759 
760 static void
761 filt_timerstart(struct knote *kn, sbintime_t to)
762 {
763 	struct kq_timer_cb_data *kc;
764 
765 	kc = kn->kn_ptr.p_v;
766 	if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
767 		kc->next = to;
768 		kc->to = 0;
769 	} else {
770 		kc->next = to + sbinuptime();
771 		kc->to = to;
772 	}
773 	callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
774 	    PCPU_GET(cpuid), C_ABSOLUTE);
775 }
776 
777 static void
778 filt_timerdetach(struct knote *kn)
779 {
780 	struct kq_timer_cb_data *kc;
781 	unsigned int old __unused;
782 
783 	kc = kn->kn_ptr.p_v;
784 	callout_drain(&kc->c);
785 	free(kc, M_KQUEUE);
786 	old = atomic_fetchadd_int(&kq_ncallouts, -1);
787 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
788 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
789 }
790 
791 static void
792 filt_timertouch(struct knote *kn, struct kevent *kev, u_long type)
793 {
794 	struct kq_timer_cb_data *kc;
795 	struct kqueue *kq;
796 	sbintime_t to;
797 	int error;
798 
799 	switch (type) {
800 	case EVENT_REGISTER:
801 		/* Handle re-added timers that update data/fflags */
802 		if (kev->flags & EV_ADD) {
803 			kc = kn->kn_ptr.p_v;
804 
805 			/* Drain any existing callout. */
806 			callout_drain(&kc->c);
807 
808 			/* Throw away any existing undelivered record
809 			 * of the timer expiration. This is done under
810 			 * the presumption that if a process is
811 			 * re-adding this timer with new parameters,
812 			 * it is no longer interested in what may have
813 			 * happened under the old parameters. If it is
814 			 * interested, it can wait for the expiration,
815 			 * delete the old timer definition, and then
816 			 * add the new one.
817 			 *
818 			 * This has to be done while the kq is locked:
819 			 *   - if enqueued, dequeue
820 			 *   - make it no longer active
821 			 *   - clear the count of expiration events
822 			 */
823 			kq = kn->kn_kq;
824 			KQ_LOCK(kq);
825 			if (kn->kn_status & KN_QUEUED)
826 				knote_dequeue(kn);
827 
828 			kn->kn_status &= ~KN_ACTIVE;
829 			kn->kn_data = 0;
830 			KQ_UNLOCK(kq);
831 
832 			/* Reschedule timer based on new data/fflags */
833 			kn->kn_sfflags = kev->fflags;
834 			kn->kn_sdata = kev->data;
835 			error = filt_timervalidate(kn, &to);
836 			if (error != 0) {
837 			  	kn->kn_flags |= EV_ERROR;
838 				kn->kn_data = error;
839 			} else
840 			  	filt_timerstart(kn, to);
841 		}
842 		break;
843 
844         case EVENT_PROCESS:
845 		*kev = kn->kn_kevent;
846 		if (kn->kn_flags & EV_CLEAR) {
847 			kn->kn_data = 0;
848 			kn->kn_fflags = 0;
849 		}
850 		break;
851 
852 	default:
853 		panic("filt_timertouch() - invalid type (%ld)", type);
854 		break;
855 	}
856 }
857 
858 static int
859 filt_timer(struct knote *kn, long hint)
860 {
861 
862 	return (kn->kn_data != 0);
863 }
864 
865 static int
866 filt_userattach(struct knote *kn)
867 {
868 
869 	/*
870 	 * EVFILT_USER knotes are not attached to anything in the kernel.
871 	 */
872 	kn->kn_hook = NULL;
873 	if (kn->kn_fflags & NOTE_TRIGGER)
874 		kn->kn_hookid = 1;
875 	else
876 		kn->kn_hookid = 0;
877 	return (0);
878 }
879 
880 static void
881 filt_userdetach(__unused struct knote *kn)
882 {
883 
884 	/*
885 	 * EVFILT_USER knotes are not attached to anything in the kernel.
886 	 */
887 }
888 
889 static int
890 filt_user(struct knote *kn, __unused long hint)
891 {
892 
893 	return (kn->kn_hookid);
894 }
895 
896 static void
897 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
898 {
899 	u_int ffctrl;
900 
901 	switch (type) {
902 	case EVENT_REGISTER:
903 		if (kev->fflags & NOTE_TRIGGER)
904 			kn->kn_hookid = 1;
905 
906 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
907 		kev->fflags &= NOTE_FFLAGSMASK;
908 		switch (ffctrl) {
909 		case NOTE_FFNOP:
910 			break;
911 
912 		case NOTE_FFAND:
913 			kn->kn_sfflags &= kev->fflags;
914 			break;
915 
916 		case NOTE_FFOR:
917 			kn->kn_sfflags |= kev->fflags;
918 			break;
919 
920 		case NOTE_FFCOPY:
921 			kn->kn_sfflags = kev->fflags;
922 			break;
923 
924 		default:
925 			/* XXX Return error? */
926 			break;
927 		}
928 		kn->kn_sdata = kev->data;
929 		if (kev->flags & EV_CLEAR) {
930 			kn->kn_hookid = 0;
931 			kn->kn_data = 0;
932 			kn->kn_fflags = 0;
933 		}
934 		break;
935 
936         case EVENT_PROCESS:
937 		*kev = kn->kn_kevent;
938 		kev->fflags = kn->kn_sfflags;
939 		kev->data = kn->kn_sdata;
940 		if (kn->kn_flags & EV_CLEAR) {
941 			kn->kn_hookid = 0;
942 			kn->kn_data = 0;
943 			kn->kn_fflags = 0;
944 		}
945 		break;
946 
947 	default:
948 		panic("filt_usertouch() - invalid type (%ld)", type);
949 		break;
950 	}
951 }
952 
953 int
954 sys_kqueue(struct thread *td, struct kqueue_args *uap)
955 {
956 
957 	return (kern_kqueue(td, 0, NULL));
958 }
959 
960 static void
961 kqueue_init(struct kqueue *kq)
962 {
963 
964 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
965 	TAILQ_INIT(&kq->kq_head);
966 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
967 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
968 }
969 
970 int
971 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
972 {
973 	struct filedesc *fdp;
974 	struct kqueue *kq;
975 	struct file *fp;
976 	struct ucred *cred;
977 	int fd, error;
978 
979 	fdp = td->td_proc->p_fd;
980 	cred = td->td_ucred;
981 	if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
982 		return (ENOMEM);
983 
984 	error = falloc_caps(td, &fp, &fd, flags, fcaps);
985 	if (error != 0) {
986 		chgkqcnt(cred->cr_ruidinfo, -1, 0);
987 		return (error);
988 	}
989 
990 	/* An extra reference on `fp' has been held for us by falloc(). */
991 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
992 	kqueue_init(kq);
993 	kq->kq_fdp = fdp;
994 	kq->kq_cred = crhold(cred);
995 
996 	FILEDESC_XLOCK(fdp);
997 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
998 	FILEDESC_XUNLOCK(fdp);
999 
1000 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
1001 	fdrop(fp, td);
1002 
1003 	td->td_retval[0] = fd;
1004 	return (0);
1005 }
1006 
1007 struct g_kevent_args {
1008 	int	fd;
1009 	void	*changelist;
1010 	int	nchanges;
1011 	void	*eventlist;
1012 	int	nevents;
1013 	const struct timespec *timeout;
1014 };
1015 
1016 int
1017 sys_kevent(struct thread *td, struct kevent_args *uap)
1018 {
1019 	struct kevent_copyops k_ops = {
1020 		.arg = uap,
1021 		.k_copyout = kevent_copyout,
1022 		.k_copyin = kevent_copyin,
1023 		.kevent_size = sizeof(struct kevent),
1024 	};
1025 	struct g_kevent_args gk_args = {
1026 		.fd = uap->fd,
1027 		.changelist = uap->changelist,
1028 		.nchanges = uap->nchanges,
1029 		.eventlist = uap->eventlist,
1030 		.nevents = uap->nevents,
1031 		.timeout = uap->timeout,
1032 	};
1033 
1034 	return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent"));
1035 }
1036 
1037 static int
1038 kern_kevent_generic(struct thread *td, struct g_kevent_args *uap,
1039     struct kevent_copyops *k_ops, const char *struct_name)
1040 {
1041 	struct timespec ts, *tsp;
1042 #ifdef KTRACE
1043 	struct kevent *eventlist = uap->eventlist;
1044 #endif
1045 	int error;
1046 
1047 	if (uap->timeout != NULL) {
1048 		error = copyin(uap->timeout, &ts, sizeof(ts));
1049 		if (error)
1050 			return (error);
1051 		tsp = &ts;
1052 	} else
1053 		tsp = NULL;
1054 
1055 #ifdef KTRACE
1056 	if (KTRPOINT(td, KTR_STRUCT_ARRAY))
1057 		ktrstructarray(struct_name, UIO_USERSPACE, uap->changelist,
1058 		    uap->nchanges, k_ops->kevent_size);
1059 #endif
1060 
1061 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
1062 	    k_ops, tsp);
1063 
1064 #ifdef KTRACE
1065 	if (error == 0 && KTRPOINT(td, KTR_STRUCT_ARRAY))
1066 		ktrstructarray(struct_name, UIO_USERSPACE, eventlist,
1067 		    td->td_retval[0], k_ops->kevent_size);
1068 #endif
1069 
1070 	return (error);
1071 }
1072 
1073 /*
1074  * Copy 'count' items into the destination list pointed to by uap->eventlist.
1075  */
1076 static int
1077 kevent_copyout(void *arg, struct kevent *kevp, int count)
1078 {
1079 	struct kevent_args *uap;
1080 	int error;
1081 
1082 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1083 	uap = (struct kevent_args *)arg;
1084 
1085 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
1086 	if (error == 0)
1087 		uap->eventlist += count;
1088 	return (error);
1089 }
1090 
1091 /*
1092  * Copy 'count' items from the list pointed to by uap->changelist.
1093  */
1094 static int
1095 kevent_copyin(void *arg, struct kevent *kevp, int count)
1096 {
1097 	struct kevent_args *uap;
1098 	int error;
1099 
1100 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1101 	uap = (struct kevent_args *)arg;
1102 
1103 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
1104 	if (error == 0)
1105 		uap->changelist += count;
1106 	return (error);
1107 }
1108 
1109 #ifdef COMPAT_FREEBSD11
1110 static int
1111 kevent11_copyout(void *arg, struct kevent *kevp, int count)
1112 {
1113 	struct freebsd11_kevent_args *uap;
1114 	struct kevent_freebsd11 kev11;
1115 	int error, i;
1116 
1117 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1118 	uap = (struct freebsd11_kevent_args *)arg;
1119 
1120 	for (i = 0; i < count; i++) {
1121 		kev11.ident = kevp->ident;
1122 		kev11.filter = kevp->filter;
1123 		kev11.flags = kevp->flags;
1124 		kev11.fflags = kevp->fflags;
1125 		kev11.data = kevp->data;
1126 		kev11.udata = kevp->udata;
1127 		error = copyout(&kev11, uap->eventlist, sizeof(kev11));
1128 		if (error != 0)
1129 			break;
1130 		uap->eventlist++;
1131 		kevp++;
1132 	}
1133 	return (error);
1134 }
1135 
1136 /*
1137  * Copy 'count' items from the list pointed to by uap->changelist.
1138  */
1139 static int
1140 kevent11_copyin(void *arg, struct kevent *kevp, int count)
1141 {
1142 	struct freebsd11_kevent_args *uap;
1143 	struct kevent_freebsd11 kev11;
1144 	int error, i;
1145 
1146 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1147 	uap = (struct freebsd11_kevent_args *)arg;
1148 
1149 	for (i = 0; i < count; i++) {
1150 		error = copyin(uap->changelist, &kev11, sizeof(kev11));
1151 		if (error != 0)
1152 			break;
1153 		kevp->ident = kev11.ident;
1154 		kevp->filter = kev11.filter;
1155 		kevp->flags = kev11.flags;
1156 		kevp->fflags = kev11.fflags;
1157 		kevp->data = (uintptr_t)kev11.data;
1158 		kevp->udata = kev11.udata;
1159 		bzero(&kevp->ext, sizeof(kevp->ext));
1160 		uap->changelist++;
1161 		kevp++;
1162 	}
1163 	return (error);
1164 }
1165 
1166 int
1167 freebsd11_kevent(struct thread *td, struct freebsd11_kevent_args *uap)
1168 {
1169 	struct kevent_copyops k_ops = {
1170 		.arg = uap,
1171 		.k_copyout = kevent11_copyout,
1172 		.k_copyin = kevent11_copyin,
1173 		.kevent_size = sizeof(struct kevent_freebsd11),
1174 	};
1175 	struct g_kevent_args gk_args = {
1176 		.fd = uap->fd,
1177 		.changelist = uap->changelist,
1178 		.nchanges = uap->nchanges,
1179 		.eventlist = uap->eventlist,
1180 		.nevents = uap->nevents,
1181 		.timeout = uap->timeout,
1182 	};
1183 
1184 	return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent_freebsd11"));
1185 }
1186 #endif
1187 
1188 int
1189 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
1190     struct kevent_copyops *k_ops, const struct timespec *timeout)
1191 {
1192 	cap_rights_t rights;
1193 	struct file *fp;
1194 	int error;
1195 
1196 	cap_rights_init(&rights);
1197 	if (nchanges > 0)
1198 		cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
1199 	if (nevents > 0)
1200 		cap_rights_set(&rights, CAP_KQUEUE_EVENT);
1201 	error = fget(td, fd, &rights, &fp);
1202 	if (error != 0)
1203 		return (error);
1204 
1205 	error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
1206 	fdrop(fp, td);
1207 
1208 	return (error);
1209 }
1210 
1211 static int
1212 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
1213     struct kevent_copyops *k_ops, const struct timespec *timeout)
1214 {
1215 	struct kevent keva[KQ_NEVENTS];
1216 	struct kevent *kevp, *changes;
1217 	int i, n, nerrors, error;
1218 
1219 	nerrors = 0;
1220 	while (nchanges > 0) {
1221 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
1222 		error = k_ops->k_copyin(k_ops->arg, keva, n);
1223 		if (error)
1224 			return (error);
1225 		changes = keva;
1226 		for (i = 0; i < n; i++) {
1227 			kevp = &changes[i];
1228 			if (!kevp->filter)
1229 				continue;
1230 			kevp->flags &= ~EV_SYSFLAGS;
1231 			error = kqueue_register(kq, kevp, td, 1);
1232 			if (error || (kevp->flags & EV_RECEIPT)) {
1233 				if (nevents == 0)
1234 					return (error);
1235 				kevp->flags = EV_ERROR;
1236 				kevp->data = error;
1237 				(void)k_ops->k_copyout(k_ops->arg, kevp, 1);
1238 				nevents--;
1239 				nerrors++;
1240 			}
1241 		}
1242 		nchanges -= n;
1243 	}
1244 	if (nerrors) {
1245 		td->td_retval[0] = nerrors;
1246 		return (0);
1247 	}
1248 
1249 	return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
1250 }
1251 
1252 int
1253 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
1254     struct kevent_copyops *k_ops, const struct timespec *timeout)
1255 {
1256 	struct kqueue *kq;
1257 	int error;
1258 
1259 	error = kqueue_acquire(fp, &kq);
1260 	if (error != 0)
1261 		return (error);
1262 	error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
1263 	kqueue_release(kq, 0);
1264 	return (error);
1265 }
1266 
1267 /*
1268  * Performs a kevent() call on a temporarily created kqueue. This can be
1269  * used to perform one-shot polling, similar to poll() and select().
1270  */
1271 int
1272 kern_kevent_anonymous(struct thread *td, int nevents,
1273     struct kevent_copyops *k_ops)
1274 {
1275 	struct kqueue kq = {};
1276 	int error;
1277 
1278 	kqueue_init(&kq);
1279 	kq.kq_refcnt = 1;
1280 	error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
1281 	kqueue_drain(&kq, td);
1282 	kqueue_destroy(&kq);
1283 	return (error);
1284 }
1285 
1286 int
1287 kqueue_add_filteropts(int filt, struct filterops *filtops)
1288 {
1289 	int error;
1290 
1291 	error = 0;
1292 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1293 		printf(
1294 "trying to add a filterop that is out of range: %d is beyond %d\n",
1295 		    ~filt, EVFILT_SYSCOUNT);
1296 		return EINVAL;
1297 	}
1298 	mtx_lock(&filterops_lock);
1299 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1300 	    sysfilt_ops[~filt].for_fop != NULL)
1301 		error = EEXIST;
1302 	else {
1303 		sysfilt_ops[~filt].for_fop = filtops;
1304 		sysfilt_ops[~filt].for_refcnt = 0;
1305 	}
1306 	mtx_unlock(&filterops_lock);
1307 
1308 	return (error);
1309 }
1310 
1311 int
1312 kqueue_del_filteropts(int filt)
1313 {
1314 	int error;
1315 
1316 	error = 0;
1317 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1318 		return EINVAL;
1319 
1320 	mtx_lock(&filterops_lock);
1321 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1322 	    sysfilt_ops[~filt].for_fop == NULL)
1323 		error = EINVAL;
1324 	else if (sysfilt_ops[~filt].for_refcnt != 0)
1325 		error = EBUSY;
1326 	else {
1327 		sysfilt_ops[~filt].for_fop = &null_filtops;
1328 		sysfilt_ops[~filt].for_refcnt = 0;
1329 	}
1330 	mtx_unlock(&filterops_lock);
1331 
1332 	return error;
1333 }
1334 
1335 static struct filterops *
1336 kqueue_fo_find(int filt)
1337 {
1338 
1339 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1340 		return NULL;
1341 
1342 	if (sysfilt_ops[~filt].for_nolock)
1343 		return sysfilt_ops[~filt].for_fop;
1344 
1345 	mtx_lock(&filterops_lock);
1346 	sysfilt_ops[~filt].for_refcnt++;
1347 	if (sysfilt_ops[~filt].for_fop == NULL)
1348 		sysfilt_ops[~filt].for_fop = &null_filtops;
1349 	mtx_unlock(&filterops_lock);
1350 
1351 	return sysfilt_ops[~filt].for_fop;
1352 }
1353 
1354 static void
1355 kqueue_fo_release(int filt)
1356 {
1357 
1358 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1359 		return;
1360 
1361 	if (sysfilt_ops[~filt].for_nolock)
1362 		return;
1363 
1364 	mtx_lock(&filterops_lock);
1365 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1366 	    ("filter object refcount not valid on release"));
1367 	sysfilt_ops[~filt].for_refcnt--;
1368 	mtx_unlock(&filterops_lock);
1369 }
1370 
1371 /*
1372  * A ref to kq (obtained via kqueue_acquire) must be held.  waitok will
1373  * influence if memory allocation should wait.  Make sure it is 0 if you
1374  * hold any mutexes.
1375  */
1376 static int
1377 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1378 {
1379 	struct filterops *fops;
1380 	struct file *fp;
1381 	struct knote *kn, *tkn;
1382 	struct knlist *knl;
1383 	int error, filt, event;
1384 	int haskqglobal, filedesc_unlock;
1385 
1386 	if ((kev->flags & (EV_ENABLE | EV_DISABLE)) == (EV_ENABLE | EV_DISABLE))
1387 		return (EINVAL);
1388 
1389 	fp = NULL;
1390 	kn = NULL;
1391 	knl = NULL;
1392 	error = 0;
1393 	haskqglobal = 0;
1394 	filedesc_unlock = 0;
1395 
1396 	filt = kev->filter;
1397 	fops = kqueue_fo_find(filt);
1398 	if (fops == NULL)
1399 		return EINVAL;
1400 
1401 	if (kev->flags & EV_ADD) {
1402 		/*
1403 		 * Prevent waiting with locks.  Non-sleepable
1404 		 * allocation failures are handled in the loop, only
1405 		 * if the spare knote appears to be actually required.
1406 		 */
1407 		tkn = knote_alloc(waitok);
1408 	} else {
1409 		tkn = NULL;
1410 	}
1411 
1412 findkn:
1413 	if (fops->f_isfd) {
1414 		KASSERT(td != NULL, ("td is NULL"));
1415 		if (kev->ident > INT_MAX)
1416 			error = EBADF;
1417 		else
1418 			error = fget(td, kev->ident, &cap_event_rights, &fp);
1419 		if (error)
1420 			goto done;
1421 
1422 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1423 		    kev->ident, 0) != 0) {
1424 			/* try again */
1425 			fdrop(fp, td);
1426 			fp = NULL;
1427 			error = kqueue_expand(kq, fops, kev->ident, waitok);
1428 			if (error)
1429 				goto done;
1430 			goto findkn;
1431 		}
1432 
1433 		if (fp->f_type == DTYPE_KQUEUE) {
1434 			/*
1435 			 * If we add some intelligence about what we are doing,
1436 			 * we should be able to support events on ourselves.
1437 			 * We need to know when we are doing this to prevent
1438 			 * getting both the knlist lock and the kq lock since
1439 			 * they are the same thing.
1440 			 */
1441 			if (fp->f_data == kq) {
1442 				error = EINVAL;
1443 				goto done;
1444 			}
1445 
1446 			/*
1447 			 * Pre-lock the filedesc before the global
1448 			 * lock mutex, see the comment in
1449 			 * kqueue_close().
1450 			 */
1451 			FILEDESC_XLOCK(td->td_proc->p_fd);
1452 			filedesc_unlock = 1;
1453 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1454 		}
1455 
1456 		KQ_LOCK(kq);
1457 		if (kev->ident < kq->kq_knlistsize) {
1458 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1459 				if (kev->filter == kn->kn_filter)
1460 					break;
1461 		}
1462 	} else {
1463 		if ((kev->flags & EV_ADD) == EV_ADD)
1464 			kqueue_expand(kq, fops, kev->ident, waitok);
1465 
1466 		KQ_LOCK(kq);
1467 
1468 		/*
1469 		 * If possible, find an existing knote to use for this kevent.
1470 		 */
1471 		if (kev->filter == EVFILT_PROC &&
1472 		    (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1473 			/* This is an internal creation of a process tracking
1474 			 * note. Don't attempt to coalesce this with an
1475 			 * existing note.
1476 			 */
1477 			;
1478 		} else if (kq->kq_knhashmask != 0) {
1479 			struct klist *list;
1480 
1481 			list = &kq->kq_knhash[
1482 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1483 			SLIST_FOREACH(kn, list, kn_link)
1484 				if (kev->ident == kn->kn_id &&
1485 				    kev->filter == kn->kn_filter)
1486 					break;
1487 		}
1488 	}
1489 
1490 	/* knote is in the process of changing, wait for it to stabilize. */
1491 	if (kn != NULL && kn_in_flux(kn)) {
1492 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1493 		if (filedesc_unlock) {
1494 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1495 			filedesc_unlock = 0;
1496 		}
1497 		kq->kq_state |= KQ_FLUXWAIT;
1498 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1499 		if (fp != NULL) {
1500 			fdrop(fp, td);
1501 			fp = NULL;
1502 		}
1503 		goto findkn;
1504 	}
1505 
1506 	/*
1507 	 * kn now contains the matching knote, or NULL if no match
1508 	 */
1509 	if (kn == NULL) {
1510 		if (kev->flags & EV_ADD) {
1511 			kn = tkn;
1512 			tkn = NULL;
1513 			if (kn == NULL) {
1514 				KQ_UNLOCK(kq);
1515 				error = ENOMEM;
1516 				goto done;
1517 			}
1518 			kn->kn_fp = fp;
1519 			kn->kn_kq = kq;
1520 			kn->kn_fop = fops;
1521 			/*
1522 			 * apply reference counts to knote structure, and
1523 			 * do not release it at the end of this routine.
1524 			 */
1525 			fops = NULL;
1526 			fp = NULL;
1527 
1528 			kn->kn_sfflags = kev->fflags;
1529 			kn->kn_sdata = kev->data;
1530 			kev->fflags = 0;
1531 			kev->data = 0;
1532 			kn->kn_kevent = *kev;
1533 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1534 			    EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1535 			kn->kn_status = KN_DETACHED;
1536 			if ((kev->flags & EV_DISABLE) != 0)
1537 				kn->kn_status |= KN_DISABLED;
1538 			kn_enter_flux(kn);
1539 
1540 			error = knote_attach(kn, kq);
1541 			KQ_UNLOCK(kq);
1542 			if (error != 0) {
1543 				tkn = kn;
1544 				goto done;
1545 			}
1546 
1547 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1548 				knote_drop_detached(kn, td);
1549 				goto done;
1550 			}
1551 			knl = kn_list_lock(kn);
1552 			goto done_ev_add;
1553 		} else {
1554 			/* No matching knote and the EV_ADD flag is not set. */
1555 			KQ_UNLOCK(kq);
1556 			error = ENOENT;
1557 			goto done;
1558 		}
1559 	}
1560 
1561 	if (kev->flags & EV_DELETE) {
1562 		kn_enter_flux(kn);
1563 		KQ_UNLOCK(kq);
1564 		knote_drop(kn, td);
1565 		goto done;
1566 	}
1567 
1568 	if (kev->flags & EV_FORCEONESHOT) {
1569 		kn->kn_flags |= EV_ONESHOT;
1570 		KNOTE_ACTIVATE(kn, 1);
1571 	}
1572 
1573 	if ((kev->flags & EV_ENABLE) != 0)
1574 		kn->kn_status &= ~KN_DISABLED;
1575 	else if ((kev->flags & EV_DISABLE) != 0)
1576 		kn->kn_status |= KN_DISABLED;
1577 
1578 	/*
1579 	 * The user may change some filter values after the initial EV_ADD,
1580 	 * but doing so will not reset any filter which has already been
1581 	 * triggered.
1582 	 */
1583 	kn->kn_status |= KN_SCAN;
1584 	kn_enter_flux(kn);
1585 	KQ_UNLOCK(kq);
1586 	knl = kn_list_lock(kn);
1587 	kn->kn_kevent.udata = kev->udata;
1588 	if (!fops->f_isfd && fops->f_touch != NULL) {
1589 		fops->f_touch(kn, kev, EVENT_REGISTER);
1590 	} else {
1591 		kn->kn_sfflags = kev->fflags;
1592 		kn->kn_sdata = kev->data;
1593 	}
1594 
1595 done_ev_add:
1596 	/*
1597 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1598 	 * the initial attach event decides that the event is "completed"
1599 	 * already, e.g., filt_procattach() is called on a zombie process.  It
1600 	 * will call filt_proc() which will remove it from the list, and NULL
1601 	 * kn_knlist.
1602 	 *
1603 	 * KN_DISABLED will be stable while the knote is in flux, so the
1604 	 * unlocked read will not race with an update.
1605 	 */
1606 	if ((kn->kn_status & KN_DISABLED) == 0)
1607 		event = kn->kn_fop->f_event(kn, 0);
1608 	else
1609 		event = 0;
1610 
1611 	KQ_LOCK(kq);
1612 	if (event)
1613 		kn->kn_status |= KN_ACTIVE;
1614 	if ((kn->kn_status & (KN_ACTIVE | KN_DISABLED | KN_QUEUED)) ==
1615 	    KN_ACTIVE)
1616 		knote_enqueue(kn);
1617 	kn->kn_status &= ~KN_SCAN;
1618 	kn_leave_flux(kn);
1619 	kn_list_unlock(knl);
1620 	KQ_UNLOCK_FLUX(kq);
1621 
1622 done:
1623 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1624 	if (filedesc_unlock)
1625 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1626 	if (fp != NULL)
1627 		fdrop(fp, td);
1628 	knote_free(tkn);
1629 	if (fops != NULL)
1630 		kqueue_fo_release(filt);
1631 	return (error);
1632 }
1633 
1634 static int
1635 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1636 {
1637 	int error;
1638 	struct kqueue *kq;
1639 
1640 	error = 0;
1641 
1642 	kq = fp->f_data;
1643 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1644 		return (EBADF);
1645 	*kqp = kq;
1646 	KQ_LOCK(kq);
1647 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1648 		KQ_UNLOCK(kq);
1649 		return (EBADF);
1650 	}
1651 	kq->kq_refcnt++;
1652 	KQ_UNLOCK(kq);
1653 
1654 	return error;
1655 }
1656 
1657 static void
1658 kqueue_release(struct kqueue *kq, int locked)
1659 {
1660 	if (locked)
1661 		KQ_OWNED(kq);
1662 	else
1663 		KQ_LOCK(kq);
1664 	kq->kq_refcnt--;
1665 	if (kq->kq_refcnt == 1)
1666 		wakeup(&kq->kq_refcnt);
1667 	if (!locked)
1668 		KQ_UNLOCK(kq);
1669 }
1670 
1671 static void
1672 kqueue_schedtask(struct kqueue *kq)
1673 {
1674 
1675 	KQ_OWNED(kq);
1676 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1677 	    ("scheduling kqueue task while draining"));
1678 
1679 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1680 		taskqueue_enqueue(taskqueue_kqueue_ctx, &kq->kq_task);
1681 		kq->kq_state |= KQ_TASKSCHED;
1682 	}
1683 }
1684 
1685 /*
1686  * Expand the kq to make sure we have storage for fops/ident pair.
1687  *
1688  * Return 0 on success (or no work necessary), return errno on failure.
1689  *
1690  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1691  * If kqueue_register is called from a non-fd context, there usually/should
1692  * be no locks held.
1693  */
1694 static int
1695 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1696 	int waitok)
1697 {
1698 	struct klist *list, *tmp_knhash, *to_free;
1699 	u_long tmp_knhashmask;
1700 	int size;
1701 	int fd;
1702 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
1703 
1704 	KQ_NOTOWNED(kq);
1705 
1706 	to_free = NULL;
1707 	if (fops->f_isfd) {
1708 		fd = ident;
1709 		if (kq->kq_knlistsize <= fd) {
1710 			size = kq->kq_knlistsize;
1711 			while (size <= fd)
1712 				size += KQEXTENT;
1713 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1714 			if (list == NULL)
1715 				return ENOMEM;
1716 			KQ_LOCK(kq);
1717 			if (kq->kq_knlistsize > fd) {
1718 				to_free = list;
1719 				list = NULL;
1720 			} else {
1721 				if (kq->kq_knlist != NULL) {
1722 					bcopy(kq->kq_knlist, list,
1723 					    kq->kq_knlistsize * sizeof(*list));
1724 					to_free = kq->kq_knlist;
1725 					kq->kq_knlist = NULL;
1726 				}
1727 				bzero((caddr_t)list +
1728 				    kq->kq_knlistsize * sizeof(*list),
1729 				    (size - kq->kq_knlistsize) * sizeof(*list));
1730 				kq->kq_knlistsize = size;
1731 				kq->kq_knlist = list;
1732 			}
1733 			KQ_UNLOCK(kq);
1734 		}
1735 	} else {
1736 		if (kq->kq_knhashmask == 0) {
1737 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1738 			    &tmp_knhashmask);
1739 			if (tmp_knhash == NULL)
1740 				return ENOMEM;
1741 			KQ_LOCK(kq);
1742 			if (kq->kq_knhashmask == 0) {
1743 				kq->kq_knhash = tmp_knhash;
1744 				kq->kq_knhashmask = tmp_knhashmask;
1745 			} else {
1746 				to_free = tmp_knhash;
1747 			}
1748 			KQ_UNLOCK(kq);
1749 		}
1750 	}
1751 	free(to_free, M_KQUEUE);
1752 
1753 	KQ_NOTOWNED(kq);
1754 	return 0;
1755 }
1756 
1757 static void
1758 kqueue_task(void *arg, int pending)
1759 {
1760 	struct kqueue *kq;
1761 	int haskqglobal;
1762 
1763 	haskqglobal = 0;
1764 	kq = arg;
1765 
1766 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1767 	KQ_LOCK(kq);
1768 
1769 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1770 
1771 	kq->kq_state &= ~KQ_TASKSCHED;
1772 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1773 		wakeup(&kq->kq_state);
1774 	}
1775 	KQ_UNLOCK(kq);
1776 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1777 }
1778 
1779 /*
1780  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1781  * We treat KN_MARKER knotes as if they are in flux.
1782  */
1783 static int
1784 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1785     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1786 {
1787 	struct kevent *kevp;
1788 	struct knote *kn, *marker;
1789 	struct knlist *knl;
1790 	sbintime_t asbt, rsbt;
1791 	int count, error, haskqglobal, influx, nkev, touch;
1792 
1793 	count = maxevents;
1794 	nkev = 0;
1795 	error = 0;
1796 	haskqglobal = 0;
1797 
1798 	if (maxevents == 0)
1799 		goto done_nl;
1800 
1801 	rsbt = 0;
1802 	if (tsp != NULL) {
1803 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1804 		    tsp->tv_nsec >= 1000000000) {
1805 			error = EINVAL;
1806 			goto done_nl;
1807 		}
1808 		if (timespecisset(tsp)) {
1809 			if (tsp->tv_sec <= INT32_MAX) {
1810 				rsbt = tstosbt(*tsp);
1811 				if (TIMESEL(&asbt, rsbt))
1812 					asbt += tc_tick_sbt;
1813 				if (asbt <= SBT_MAX - rsbt)
1814 					asbt += rsbt;
1815 				else
1816 					asbt = 0;
1817 				rsbt >>= tc_precexp;
1818 			} else
1819 				asbt = 0;
1820 		} else
1821 			asbt = -1;
1822 	} else
1823 		asbt = 0;
1824 	marker = knote_alloc(1);
1825 	marker->kn_status = KN_MARKER;
1826 	KQ_LOCK(kq);
1827 
1828 retry:
1829 	kevp = keva;
1830 	if (kq->kq_count == 0) {
1831 		if (asbt == -1) {
1832 			error = EWOULDBLOCK;
1833 		} else {
1834 			kq->kq_state |= KQ_SLEEP;
1835 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1836 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1837 		}
1838 		if (error == 0)
1839 			goto retry;
1840 		/* don't restart after signals... */
1841 		if (error == ERESTART)
1842 			error = EINTR;
1843 		else if (error == EWOULDBLOCK)
1844 			error = 0;
1845 		goto done;
1846 	}
1847 
1848 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1849 	influx = 0;
1850 	while (count) {
1851 		KQ_OWNED(kq);
1852 		kn = TAILQ_FIRST(&kq->kq_head);
1853 
1854 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1855 		    kn_in_flux(kn)) {
1856 			if (influx) {
1857 				influx = 0;
1858 				KQ_FLUX_WAKEUP(kq);
1859 			}
1860 			kq->kq_state |= KQ_FLUXWAIT;
1861 			error = msleep(kq, &kq->kq_lock, PSOCK,
1862 			    "kqflxwt", 0);
1863 			continue;
1864 		}
1865 
1866 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1867 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1868 			kn->kn_status &= ~KN_QUEUED;
1869 			kq->kq_count--;
1870 			continue;
1871 		}
1872 		if (kn == marker) {
1873 			KQ_FLUX_WAKEUP(kq);
1874 			if (count == maxevents)
1875 				goto retry;
1876 			goto done;
1877 		}
1878 		KASSERT(!kn_in_flux(kn),
1879 		    ("knote %p is unexpectedly in flux", kn));
1880 
1881 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1882 			kn->kn_status &= ~KN_QUEUED;
1883 			kn_enter_flux(kn);
1884 			kq->kq_count--;
1885 			KQ_UNLOCK(kq);
1886 			/*
1887 			 * We don't need to lock the list since we've
1888 			 * marked it as in flux.
1889 			 */
1890 			knote_drop(kn, td);
1891 			KQ_LOCK(kq);
1892 			continue;
1893 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1894 			kn->kn_status &= ~KN_QUEUED;
1895 			kn_enter_flux(kn);
1896 			kq->kq_count--;
1897 			KQ_UNLOCK(kq);
1898 			/*
1899 			 * We don't need to lock the list since we've
1900 			 * marked the knote as being in flux.
1901 			 */
1902 			*kevp = kn->kn_kevent;
1903 			knote_drop(kn, td);
1904 			KQ_LOCK(kq);
1905 			kn = NULL;
1906 		} else {
1907 			kn->kn_status |= KN_SCAN;
1908 			kn_enter_flux(kn);
1909 			KQ_UNLOCK(kq);
1910 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1911 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1912 			knl = kn_list_lock(kn);
1913 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1914 				KQ_LOCK(kq);
1915 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1916 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE |
1917 				    KN_SCAN);
1918 				kn_leave_flux(kn);
1919 				kq->kq_count--;
1920 				kn_list_unlock(knl);
1921 				influx = 1;
1922 				continue;
1923 			}
1924 			touch = (!kn->kn_fop->f_isfd &&
1925 			    kn->kn_fop->f_touch != NULL);
1926 			if (touch)
1927 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1928 			else
1929 				*kevp = kn->kn_kevent;
1930 			KQ_LOCK(kq);
1931 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1932 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1933 				/*
1934 				 * Manually clear knotes who weren't
1935 				 * 'touch'ed.
1936 				 */
1937 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1938 					kn->kn_data = 0;
1939 					kn->kn_fflags = 0;
1940 				}
1941 				if (kn->kn_flags & EV_DISPATCH)
1942 					kn->kn_status |= KN_DISABLED;
1943 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1944 				kq->kq_count--;
1945 			} else
1946 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1947 
1948 			kn->kn_status &= ~KN_SCAN;
1949 			kn_leave_flux(kn);
1950 			kn_list_unlock(knl);
1951 			influx = 1;
1952 		}
1953 
1954 		/* we are returning a copy to the user */
1955 		kevp++;
1956 		nkev++;
1957 		count--;
1958 
1959 		if (nkev == KQ_NEVENTS) {
1960 			influx = 0;
1961 			KQ_UNLOCK_FLUX(kq);
1962 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1963 			nkev = 0;
1964 			kevp = keva;
1965 			KQ_LOCK(kq);
1966 			if (error)
1967 				break;
1968 		}
1969 	}
1970 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1971 done:
1972 	KQ_OWNED(kq);
1973 	KQ_UNLOCK_FLUX(kq);
1974 	knote_free(marker);
1975 done_nl:
1976 	KQ_NOTOWNED(kq);
1977 	if (nkev != 0)
1978 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1979 	td->td_retval[0] = maxevents - count;
1980 	return (error);
1981 }
1982 
1983 /*ARGSUSED*/
1984 static int
1985 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1986 	struct ucred *active_cred, struct thread *td)
1987 {
1988 	/*
1989 	 * Enabling sigio causes two major problems:
1990 	 * 1) infinite recursion:
1991 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1992 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1993 	 * into itself over and over.  Sending the sigio causes the kqueue
1994 	 * to become ready, which in turn posts sigio again, forever.
1995 	 * Solution: this can be solved by setting a flag in the kqueue that
1996 	 * we have a SIGIO in progress.
1997 	 * 2) locking problems:
1998 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1999 	 * us above the proc and pgrp locks.
2000 	 * Solution: Post a signal using an async mechanism, being sure to
2001 	 * record a generation count in the delivery so that we do not deliver
2002 	 * a signal to the wrong process.
2003 	 *
2004 	 * Note, these two mechanisms are somewhat mutually exclusive!
2005 	 */
2006 #if 0
2007 	struct kqueue *kq;
2008 
2009 	kq = fp->f_data;
2010 	switch (cmd) {
2011 	case FIOASYNC:
2012 		if (*(int *)data) {
2013 			kq->kq_state |= KQ_ASYNC;
2014 		} else {
2015 			kq->kq_state &= ~KQ_ASYNC;
2016 		}
2017 		return (0);
2018 
2019 	case FIOSETOWN:
2020 		return (fsetown(*(int *)data, &kq->kq_sigio));
2021 
2022 	case FIOGETOWN:
2023 		*(int *)data = fgetown(&kq->kq_sigio);
2024 		return (0);
2025 	}
2026 #endif
2027 
2028 	return (ENOTTY);
2029 }
2030 
2031 /*ARGSUSED*/
2032 static int
2033 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
2034 	struct thread *td)
2035 {
2036 	struct kqueue *kq;
2037 	int revents = 0;
2038 	int error;
2039 
2040 	if ((error = kqueue_acquire(fp, &kq)))
2041 		return POLLERR;
2042 
2043 	KQ_LOCK(kq);
2044 	if (events & (POLLIN | POLLRDNORM)) {
2045 		if (kq->kq_count) {
2046 			revents |= events & (POLLIN | POLLRDNORM);
2047 		} else {
2048 			selrecord(td, &kq->kq_sel);
2049 			if (SEL_WAITING(&kq->kq_sel))
2050 				kq->kq_state |= KQ_SEL;
2051 		}
2052 	}
2053 	kqueue_release(kq, 1);
2054 	KQ_UNLOCK(kq);
2055 	return (revents);
2056 }
2057 
2058 /*ARGSUSED*/
2059 static int
2060 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
2061 	struct thread *td)
2062 {
2063 
2064 	bzero((void *)st, sizeof *st);
2065 	/*
2066 	 * We no longer return kq_count because the unlocked value is useless.
2067 	 * If you spent all this time getting the count, why not spend your
2068 	 * syscall better by calling kevent?
2069 	 *
2070 	 * XXX - This is needed for libc_r.
2071 	 */
2072 	st->st_mode = S_IFIFO;
2073 	return (0);
2074 }
2075 
2076 static void
2077 kqueue_drain(struct kqueue *kq, struct thread *td)
2078 {
2079 	struct knote *kn;
2080 	int i;
2081 
2082 	KQ_LOCK(kq);
2083 
2084 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
2085 	    ("kqueue already closing"));
2086 	kq->kq_state |= KQ_CLOSING;
2087 	if (kq->kq_refcnt > 1)
2088 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
2089 
2090 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
2091 
2092 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
2093 	    ("kqueue's knlist not empty"));
2094 
2095 	for (i = 0; i < kq->kq_knlistsize; i++) {
2096 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
2097 			if (kn_in_flux(kn)) {
2098 				kq->kq_state |= KQ_FLUXWAIT;
2099 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
2100 				continue;
2101 			}
2102 			kn_enter_flux(kn);
2103 			KQ_UNLOCK(kq);
2104 			knote_drop(kn, td);
2105 			KQ_LOCK(kq);
2106 		}
2107 	}
2108 	if (kq->kq_knhashmask != 0) {
2109 		for (i = 0; i <= kq->kq_knhashmask; i++) {
2110 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
2111 				if (kn_in_flux(kn)) {
2112 					kq->kq_state |= KQ_FLUXWAIT;
2113 					msleep(kq, &kq->kq_lock, PSOCK,
2114 					       "kqclo2", 0);
2115 					continue;
2116 				}
2117 				kn_enter_flux(kn);
2118 				KQ_UNLOCK(kq);
2119 				knote_drop(kn, td);
2120 				KQ_LOCK(kq);
2121 			}
2122 		}
2123 	}
2124 
2125 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
2126 		kq->kq_state |= KQ_TASKDRAIN;
2127 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
2128 	}
2129 
2130 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2131 		selwakeuppri(&kq->kq_sel, PSOCK);
2132 		if (!SEL_WAITING(&kq->kq_sel))
2133 			kq->kq_state &= ~KQ_SEL;
2134 	}
2135 
2136 	KQ_UNLOCK(kq);
2137 }
2138 
2139 static void
2140 kqueue_destroy(struct kqueue *kq)
2141 {
2142 
2143 	KASSERT(kq->kq_fdp == NULL,
2144 	    ("kqueue still attached to a file descriptor"));
2145 	seldrain(&kq->kq_sel);
2146 	knlist_destroy(&kq->kq_sel.si_note);
2147 	mtx_destroy(&kq->kq_lock);
2148 
2149 	if (kq->kq_knhash != NULL)
2150 		free(kq->kq_knhash, M_KQUEUE);
2151 	if (kq->kq_knlist != NULL)
2152 		free(kq->kq_knlist, M_KQUEUE);
2153 
2154 	funsetown(&kq->kq_sigio);
2155 }
2156 
2157 /*ARGSUSED*/
2158 static int
2159 kqueue_close(struct file *fp, struct thread *td)
2160 {
2161 	struct kqueue *kq = fp->f_data;
2162 	struct filedesc *fdp;
2163 	int error;
2164 	int filedesc_unlock;
2165 
2166 	if ((error = kqueue_acquire(fp, &kq)))
2167 		return error;
2168 	kqueue_drain(kq, td);
2169 
2170 	/*
2171 	 * We could be called due to the knote_drop() doing fdrop(),
2172 	 * called from kqueue_register().  In this case the global
2173 	 * lock is owned, and filedesc sx is locked before, to not
2174 	 * take the sleepable lock after non-sleepable.
2175 	 */
2176 	fdp = kq->kq_fdp;
2177 	kq->kq_fdp = NULL;
2178 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
2179 		FILEDESC_XLOCK(fdp);
2180 		filedesc_unlock = 1;
2181 	} else
2182 		filedesc_unlock = 0;
2183 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
2184 	if (filedesc_unlock)
2185 		FILEDESC_XUNLOCK(fdp);
2186 
2187 	kqueue_destroy(kq);
2188 	chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
2189 	crfree(kq->kq_cred);
2190 	free(kq, M_KQUEUE);
2191 	fp->f_data = NULL;
2192 
2193 	return (0);
2194 }
2195 
2196 static int
2197 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2198 {
2199 
2200 	kif->kf_type = KF_TYPE_KQUEUE;
2201 	return (0);
2202 }
2203 
2204 static void
2205 kqueue_wakeup(struct kqueue *kq)
2206 {
2207 	KQ_OWNED(kq);
2208 
2209 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
2210 		kq->kq_state &= ~KQ_SLEEP;
2211 		wakeup(kq);
2212 	}
2213 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2214 		selwakeuppri(&kq->kq_sel, PSOCK);
2215 		if (!SEL_WAITING(&kq->kq_sel))
2216 			kq->kq_state &= ~KQ_SEL;
2217 	}
2218 	if (!knlist_empty(&kq->kq_sel.si_note))
2219 		kqueue_schedtask(kq);
2220 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
2221 		pgsigio(&kq->kq_sigio, SIGIO, 0);
2222 	}
2223 }
2224 
2225 /*
2226  * Walk down a list of knotes, activating them if their event has triggered.
2227  *
2228  * There is a possibility to optimize in the case of one kq watching another.
2229  * Instead of scheduling a task to wake it up, you could pass enough state
2230  * down the chain to make up the parent kqueue.  Make this code functional
2231  * first.
2232  */
2233 void
2234 knote(struct knlist *list, long hint, int lockflags)
2235 {
2236 	struct kqueue *kq;
2237 	struct knote *kn, *tkn;
2238 	int error;
2239 
2240 	if (list == NULL)
2241 		return;
2242 
2243 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
2244 
2245 	if ((lockflags & KNF_LISTLOCKED) == 0)
2246 		list->kl_lock(list->kl_lockarg);
2247 
2248 	/*
2249 	 * If we unlock the list lock (and enter influx), we can
2250 	 * eliminate the kqueue scheduling, but this will introduce
2251 	 * four lock/unlock's for each knote to test.  Also, marker
2252 	 * would be needed to keep iteration position, since filters
2253 	 * or other threads could remove events.
2254 	 */
2255 	SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
2256 		kq = kn->kn_kq;
2257 		KQ_LOCK(kq);
2258 		if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
2259 			/*
2260 			 * Do not process the influx notes, except for
2261 			 * the influx coming from the kq unlock in the
2262 			 * kqueue_scan().  In the later case, we do
2263 			 * not interfere with the scan, since the code
2264 			 * fragment in kqueue_scan() locks the knlist,
2265 			 * and cannot proceed until we finished.
2266 			 */
2267 			KQ_UNLOCK(kq);
2268 		} else if ((lockflags & KNF_NOKQLOCK) != 0) {
2269 			kn_enter_flux(kn);
2270 			KQ_UNLOCK(kq);
2271 			error = kn->kn_fop->f_event(kn, hint);
2272 			KQ_LOCK(kq);
2273 			kn_leave_flux(kn);
2274 			if (error)
2275 				KNOTE_ACTIVATE(kn, 1);
2276 			KQ_UNLOCK_FLUX(kq);
2277 		} else {
2278 			if (kn->kn_fop->f_event(kn, hint))
2279 				KNOTE_ACTIVATE(kn, 1);
2280 			KQ_UNLOCK(kq);
2281 		}
2282 	}
2283 	if ((lockflags & KNF_LISTLOCKED) == 0)
2284 		list->kl_unlock(list->kl_lockarg);
2285 }
2286 
2287 /*
2288  * add a knote to a knlist
2289  */
2290 void
2291 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2292 {
2293 
2294 	KNL_ASSERT_LOCK(knl, islocked);
2295 	KQ_NOTOWNED(kn->kn_kq);
2296 	KASSERT(kn_in_flux(kn), ("knote %p not in flux", kn));
2297 	KASSERT((kn->kn_status & KN_DETACHED) != 0,
2298 	    ("knote %p was not detached", kn));
2299 	if (!islocked)
2300 		knl->kl_lock(knl->kl_lockarg);
2301 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2302 	if (!islocked)
2303 		knl->kl_unlock(knl->kl_lockarg);
2304 	KQ_LOCK(kn->kn_kq);
2305 	kn->kn_knlist = knl;
2306 	kn->kn_status &= ~KN_DETACHED;
2307 	KQ_UNLOCK(kn->kn_kq);
2308 }
2309 
2310 static void
2311 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked,
2312     int kqislocked)
2313 {
2314 
2315 	KASSERT(!kqislocked || knlislocked, ("kq locked w/o knl locked"));
2316 	KNL_ASSERT_LOCK(knl, knlislocked);
2317 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2318 	KASSERT(kqislocked || kn_in_flux(kn), ("knote %p not in flux", kn));
2319 	KASSERT((kn->kn_status & KN_DETACHED) == 0,
2320 	    ("knote %p was already detached", kn));
2321 	if (!knlislocked)
2322 		knl->kl_lock(knl->kl_lockarg);
2323 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2324 	kn->kn_knlist = NULL;
2325 	if (!knlislocked)
2326 		kn_list_unlock(knl);
2327 	if (!kqislocked)
2328 		KQ_LOCK(kn->kn_kq);
2329 	kn->kn_status |= KN_DETACHED;
2330 	if (!kqislocked)
2331 		KQ_UNLOCK(kn->kn_kq);
2332 }
2333 
2334 /*
2335  * remove knote from the specified knlist
2336  */
2337 void
2338 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2339 {
2340 
2341 	knlist_remove_kq(knl, kn, islocked, 0);
2342 }
2343 
2344 int
2345 knlist_empty(struct knlist *knl)
2346 {
2347 
2348 	KNL_ASSERT_LOCKED(knl);
2349 	return (SLIST_EMPTY(&knl->kl_list));
2350 }
2351 
2352 static struct mtx knlist_lock;
2353 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2354     MTX_DEF);
2355 static void knlist_mtx_lock(void *arg);
2356 static void knlist_mtx_unlock(void *arg);
2357 
2358 static void
2359 knlist_mtx_lock(void *arg)
2360 {
2361 
2362 	mtx_lock((struct mtx *)arg);
2363 }
2364 
2365 static void
2366 knlist_mtx_unlock(void *arg)
2367 {
2368 
2369 	mtx_unlock((struct mtx *)arg);
2370 }
2371 
2372 static void
2373 knlist_mtx_assert_locked(void *arg)
2374 {
2375 
2376 	mtx_assert((struct mtx *)arg, MA_OWNED);
2377 }
2378 
2379 static void
2380 knlist_mtx_assert_unlocked(void *arg)
2381 {
2382 
2383 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2384 }
2385 
2386 static void
2387 knlist_rw_rlock(void *arg)
2388 {
2389 
2390 	rw_rlock((struct rwlock *)arg);
2391 }
2392 
2393 static void
2394 knlist_rw_runlock(void *arg)
2395 {
2396 
2397 	rw_runlock((struct rwlock *)arg);
2398 }
2399 
2400 static void
2401 knlist_rw_assert_locked(void *arg)
2402 {
2403 
2404 	rw_assert((struct rwlock *)arg, RA_LOCKED);
2405 }
2406 
2407 static void
2408 knlist_rw_assert_unlocked(void *arg)
2409 {
2410 
2411 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2412 }
2413 
2414 void
2415 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2416     void (*kl_unlock)(void *),
2417     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2418 {
2419 
2420 	if (lock == NULL)
2421 		knl->kl_lockarg = &knlist_lock;
2422 	else
2423 		knl->kl_lockarg = lock;
2424 
2425 	if (kl_lock == NULL)
2426 		knl->kl_lock = knlist_mtx_lock;
2427 	else
2428 		knl->kl_lock = kl_lock;
2429 	if (kl_unlock == NULL)
2430 		knl->kl_unlock = knlist_mtx_unlock;
2431 	else
2432 		knl->kl_unlock = kl_unlock;
2433 	if (kl_assert_locked == NULL)
2434 		knl->kl_assert_locked = knlist_mtx_assert_locked;
2435 	else
2436 		knl->kl_assert_locked = kl_assert_locked;
2437 	if (kl_assert_unlocked == NULL)
2438 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2439 	else
2440 		knl->kl_assert_unlocked = kl_assert_unlocked;
2441 
2442 	knl->kl_autodestroy = 0;
2443 	SLIST_INIT(&knl->kl_list);
2444 }
2445 
2446 void
2447 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2448 {
2449 
2450 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2451 }
2452 
2453 struct knlist *
2454 knlist_alloc(struct mtx *lock)
2455 {
2456 	struct knlist *knl;
2457 
2458 	knl = malloc(sizeof(struct knlist), M_KQUEUE, M_WAITOK);
2459 	knlist_init_mtx(knl, lock);
2460 	return (knl);
2461 }
2462 
2463 void
2464 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2465 {
2466 
2467 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2468 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2469 }
2470 
2471 void
2472 knlist_destroy(struct knlist *knl)
2473 {
2474 
2475 	KASSERT(KNLIST_EMPTY(knl),
2476 	    ("destroying knlist %p with knotes on it", knl));
2477 }
2478 
2479 void
2480 knlist_detach(struct knlist *knl)
2481 {
2482 
2483 	KNL_ASSERT_LOCKED(knl);
2484 	knl->kl_autodestroy = 1;
2485 	if (knlist_empty(knl)) {
2486 		knlist_destroy(knl);
2487 		free(knl, M_KQUEUE);
2488 	}
2489 }
2490 
2491 /*
2492  * Even if we are locked, we may need to drop the lock to allow any influx
2493  * knotes time to "settle".
2494  */
2495 void
2496 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2497 {
2498 	struct knote *kn, *kn2;
2499 	struct kqueue *kq;
2500 
2501 	KASSERT(!knl->kl_autodestroy, ("cleardel for autodestroy %p", knl));
2502 	if (islocked)
2503 		KNL_ASSERT_LOCKED(knl);
2504 	else {
2505 		KNL_ASSERT_UNLOCKED(knl);
2506 again:		/* need to reacquire lock since we have dropped it */
2507 		knl->kl_lock(knl->kl_lockarg);
2508 	}
2509 
2510 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2511 		kq = kn->kn_kq;
2512 		KQ_LOCK(kq);
2513 		if (kn_in_flux(kn)) {
2514 			KQ_UNLOCK(kq);
2515 			continue;
2516 		}
2517 		knlist_remove_kq(knl, kn, 1, 1);
2518 		if (killkn) {
2519 			kn_enter_flux(kn);
2520 			KQ_UNLOCK(kq);
2521 			knote_drop_detached(kn, td);
2522 		} else {
2523 			/* Make sure cleared knotes disappear soon */
2524 			kn->kn_flags |= EV_EOF | EV_ONESHOT;
2525 			KQ_UNLOCK(kq);
2526 		}
2527 		kq = NULL;
2528 	}
2529 
2530 	if (!SLIST_EMPTY(&knl->kl_list)) {
2531 		/* there are still in flux knotes remaining */
2532 		kn = SLIST_FIRST(&knl->kl_list);
2533 		kq = kn->kn_kq;
2534 		KQ_LOCK(kq);
2535 		KASSERT(kn_in_flux(kn), ("knote removed w/o list lock"));
2536 		knl->kl_unlock(knl->kl_lockarg);
2537 		kq->kq_state |= KQ_FLUXWAIT;
2538 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2539 		kq = NULL;
2540 		goto again;
2541 	}
2542 
2543 	if (islocked)
2544 		KNL_ASSERT_LOCKED(knl);
2545 	else {
2546 		knl->kl_unlock(knl->kl_lockarg);
2547 		KNL_ASSERT_UNLOCKED(knl);
2548 	}
2549 }
2550 
2551 /*
2552  * Remove all knotes referencing a specified fd must be called with FILEDESC
2553  * lock.  This prevents a race where a new fd comes along and occupies the
2554  * entry and we attach a knote to the fd.
2555  */
2556 void
2557 knote_fdclose(struct thread *td, int fd)
2558 {
2559 	struct filedesc *fdp = td->td_proc->p_fd;
2560 	struct kqueue *kq;
2561 	struct knote *kn;
2562 	int influx;
2563 
2564 	FILEDESC_XLOCK_ASSERT(fdp);
2565 
2566 	/*
2567 	 * We shouldn't have to worry about new kevents appearing on fd
2568 	 * since filedesc is locked.
2569 	 */
2570 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2571 		KQ_LOCK(kq);
2572 
2573 again:
2574 		influx = 0;
2575 		while (kq->kq_knlistsize > fd &&
2576 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2577 			if (kn_in_flux(kn)) {
2578 				/* someone else might be waiting on our knote */
2579 				if (influx)
2580 					wakeup(kq);
2581 				kq->kq_state |= KQ_FLUXWAIT;
2582 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2583 				goto again;
2584 			}
2585 			kn_enter_flux(kn);
2586 			KQ_UNLOCK(kq);
2587 			influx = 1;
2588 			knote_drop(kn, td);
2589 			KQ_LOCK(kq);
2590 		}
2591 		KQ_UNLOCK_FLUX(kq);
2592 	}
2593 }
2594 
2595 static int
2596 knote_attach(struct knote *kn, struct kqueue *kq)
2597 {
2598 	struct klist *list;
2599 
2600 	KASSERT(kn_in_flux(kn), ("knote %p not marked influx", kn));
2601 	KQ_OWNED(kq);
2602 
2603 	if (kn->kn_fop->f_isfd) {
2604 		if (kn->kn_id >= kq->kq_knlistsize)
2605 			return (ENOMEM);
2606 		list = &kq->kq_knlist[kn->kn_id];
2607 	} else {
2608 		if (kq->kq_knhash == NULL)
2609 			return (ENOMEM);
2610 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2611 	}
2612 	SLIST_INSERT_HEAD(list, kn, kn_link);
2613 	return (0);
2614 }
2615 
2616 static void
2617 knote_drop(struct knote *kn, struct thread *td)
2618 {
2619 
2620 	if ((kn->kn_status & KN_DETACHED) == 0)
2621 		kn->kn_fop->f_detach(kn);
2622 	knote_drop_detached(kn, td);
2623 }
2624 
2625 static void
2626 knote_drop_detached(struct knote *kn, struct thread *td)
2627 {
2628 	struct kqueue *kq;
2629 	struct klist *list;
2630 
2631 	kq = kn->kn_kq;
2632 
2633 	KASSERT((kn->kn_status & KN_DETACHED) != 0,
2634 	    ("knote %p still attached", kn));
2635 	KQ_NOTOWNED(kq);
2636 
2637 	KQ_LOCK(kq);
2638 	KASSERT(kn->kn_influx == 1,
2639 	    ("knote_drop called on %p with influx %d", kn, kn->kn_influx));
2640 
2641 	if (kn->kn_fop->f_isfd)
2642 		list = &kq->kq_knlist[kn->kn_id];
2643 	else
2644 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2645 
2646 	if (!SLIST_EMPTY(list))
2647 		SLIST_REMOVE(list, kn, knote, kn_link);
2648 	if (kn->kn_status & KN_QUEUED)
2649 		knote_dequeue(kn);
2650 	KQ_UNLOCK_FLUX(kq);
2651 
2652 	if (kn->kn_fop->f_isfd) {
2653 		fdrop(kn->kn_fp, td);
2654 		kn->kn_fp = NULL;
2655 	}
2656 	kqueue_fo_release(kn->kn_kevent.filter);
2657 	kn->kn_fop = NULL;
2658 	knote_free(kn);
2659 }
2660 
2661 static void
2662 knote_enqueue(struct knote *kn)
2663 {
2664 	struct kqueue *kq = kn->kn_kq;
2665 
2666 	KQ_OWNED(kn->kn_kq);
2667 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2668 
2669 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2670 	kn->kn_status |= KN_QUEUED;
2671 	kq->kq_count++;
2672 	kqueue_wakeup(kq);
2673 }
2674 
2675 static void
2676 knote_dequeue(struct knote *kn)
2677 {
2678 	struct kqueue *kq = kn->kn_kq;
2679 
2680 	KQ_OWNED(kn->kn_kq);
2681 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2682 
2683 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2684 	kn->kn_status &= ~KN_QUEUED;
2685 	kq->kq_count--;
2686 }
2687 
2688 static void
2689 knote_init(void)
2690 {
2691 
2692 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2693 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2694 }
2695 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2696 
2697 static struct knote *
2698 knote_alloc(int waitok)
2699 {
2700 
2701 	return (uma_zalloc(knote_zone, (waitok ? M_WAITOK : M_NOWAIT) |
2702 	    M_ZERO));
2703 }
2704 
2705 static void
2706 knote_free(struct knote *kn)
2707 {
2708 
2709 	uma_zfree(knote_zone, kn);
2710 }
2711 
2712 /*
2713  * Register the kev w/ the kq specified by fd.
2714  */
2715 int
2716 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2717 {
2718 	struct kqueue *kq;
2719 	struct file *fp;
2720 	cap_rights_t rights;
2721 	int error;
2722 
2723 	error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2724 	if (error != 0)
2725 		return (error);
2726 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2727 		goto noacquire;
2728 
2729 	error = kqueue_register(kq, kev, td, waitok);
2730 	kqueue_release(kq, 0);
2731 
2732 noacquire:
2733 	fdrop(fp, td);
2734 	return (error);
2735 }
2736