xref: /freebsd/sys/kern/kern_event.c (revision e8d8bef961a50d4dc22501cde4fb9fb0be1b2532)
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/limits.h>
46 #include <sys/lock.h>
47 #include <sys/mutex.h>
48 #include <sys/rwlock.h>
49 #include <sys/proc.h>
50 #include <sys/malloc.h>
51 #include <sys/unistd.h>
52 #include <sys/file.h>
53 #include <sys/filedesc.h>
54 #include <sys/filio.h>
55 #include <sys/fcntl.h>
56 #include <sys/kthread.h>
57 #include <sys/selinfo.h>
58 #include <sys/queue.h>
59 #include <sys/event.h>
60 #include <sys/eventvar.h>
61 #include <sys/poll.h>
62 #include <sys/protosw.h>
63 #include <sys/resourcevar.h>
64 #include <sys/sigio.h>
65 #include <sys/signalvar.h>
66 #include <sys/socket.h>
67 #include <sys/socketvar.h>
68 #include <sys/stat.h>
69 #include <sys/sysctl.h>
70 #include <sys/sysproto.h>
71 #include <sys/syscallsubr.h>
72 #include <sys/taskqueue.h>
73 #include <sys/uio.h>
74 #include <sys/user.h>
75 #ifdef KTRACE
76 #include <sys/ktrace.h>
77 #endif
78 #include <machine/atomic.h>
79 
80 #include <vm/uma.h>
81 
82 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
83 
84 /*
85  * This lock is used if multiple kq locks are required.  This possibly
86  * should be made into a per proc lock.
87  */
88 static struct mtx	kq_global;
89 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
90 #define KQ_GLOBAL_LOCK(lck, haslck)	do {	\
91 	if (!haslck)				\
92 		mtx_lock(lck);			\
93 	haslck = 1;				\
94 } while (0)
95 #define KQ_GLOBAL_UNLOCK(lck, haslck)	do {	\
96 	if (haslck)				\
97 		mtx_unlock(lck);			\
98 	haslck = 0;				\
99 } while (0)
100 
101 TASKQUEUE_DEFINE_THREAD(kqueue_ctx);
102 
103 static int	kevent_copyout(void *arg, struct kevent *kevp, int count);
104 static int	kevent_copyin(void *arg, struct kevent *kevp, int count);
105 static int	kqueue_register(struct kqueue *kq, struct kevent *kev,
106 		    struct thread *td, int mflag);
107 static int	kqueue_acquire(struct file *fp, struct kqueue **kqp);
108 static void	kqueue_release(struct kqueue *kq, int locked);
109 static void	kqueue_destroy(struct kqueue *kq);
110 static void	kqueue_drain(struct kqueue *kq, struct thread *td);
111 static int	kqueue_expand(struct kqueue *kq, struct filterops *fops,
112 		    uintptr_t ident, int mflag);
113 static void	kqueue_task(void *arg, int pending);
114 static int	kqueue_scan(struct kqueue *kq, int maxevents,
115 		    struct kevent_copyops *k_ops,
116 		    const struct timespec *timeout,
117 		    struct kevent *keva, struct thread *td);
118 static void 	kqueue_wakeup(struct kqueue *kq);
119 static struct filterops *kqueue_fo_find(int filt);
120 static void	kqueue_fo_release(int filt);
121 struct g_kevent_args;
122 static int	kern_kevent_generic(struct thread *td,
123 		    struct g_kevent_args *uap,
124 		    struct kevent_copyops *k_ops, const char *struct_name);
125 
126 static fo_ioctl_t	kqueue_ioctl;
127 static fo_poll_t	kqueue_poll;
128 static fo_kqfilter_t	kqueue_kqfilter;
129 static fo_stat_t	kqueue_stat;
130 static fo_close_t	kqueue_close;
131 static fo_fill_kinfo_t	kqueue_fill_kinfo;
132 
133 static struct fileops kqueueops = {
134 	.fo_read = invfo_rdwr,
135 	.fo_write = invfo_rdwr,
136 	.fo_truncate = invfo_truncate,
137 	.fo_ioctl = kqueue_ioctl,
138 	.fo_poll = kqueue_poll,
139 	.fo_kqfilter = kqueue_kqfilter,
140 	.fo_stat = kqueue_stat,
141 	.fo_close = kqueue_close,
142 	.fo_chmod = invfo_chmod,
143 	.fo_chown = invfo_chown,
144 	.fo_sendfile = invfo_sendfile,
145 	.fo_fill_kinfo = kqueue_fill_kinfo,
146 };
147 
148 static int 	knote_attach(struct knote *kn, struct kqueue *kq);
149 static void 	knote_drop(struct knote *kn, struct thread *td);
150 static void 	knote_drop_detached(struct knote *kn, struct thread *td);
151 static void 	knote_enqueue(struct knote *kn);
152 static void 	knote_dequeue(struct knote *kn);
153 static void 	knote_init(void);
154 static struct 	knote *knote_alloc(int mflag);
155 static void 	knote_free(struct knote *kn);
156 
157 static void	filt_kqdetach(struct knote *kn);
158 static int	filt_kqueue(struct knote *kn, long hint);
159 static int	filt_procattach(struct knote *kn);
160 static void	filt_procdetach(struct knote *kn);
161 static int	filt_proc(struct knote *kn, long hint);
162 static int	filt_fileattach(struct knote *kn);
163 static void	filt_timerexpire(void *knx);
164 static void	filt_timerexpire_l(struct knote *kn, bool proc_locked);
165 static int	filt_timerattach(struct knote *kn);
166 static void	filt_timerdetach(struct knote *kn);
167 static void	filt_timerstart(struct knote *kn, sbintime_t to);
168 static void	filt_timertouch(struct knote *kn, struct kevent *kev,
169 		    u_long type);
170 static int	filt_timervalidate(struct knote *kn, sbintime_t *to);
171 static int	filt_timer(struct knote *kn, long hint);
172 static int	filt_userattach(struct knote *kn);
173 static void	filt_userdetach(struct knote *kn);
174 static int	filt_user(struct knote *kn, long hint);
175 static void	filt_usertouch(struct knote *kn, struct kevent *kev,
176 		    u_long type);
177 
178 static struct filterops file_filtops = {
179 	.f_isfd = 1,
180 	.f_attach = filt_fileattach,
181 };
182 static struct filterops kqread_filtops = {
183 	.f_isfd = 1,
184 	.f_detach = filt_kqdetach,
185 	.f_event = filt_kqueue,
186 };
187 /* XXX - move to kern_proc.c?  */
188 static struct filterops proc_filtops = {
189 	.f_isfd = 0,
190 	.f_attach = filt_procattach,
191 	.f_detach = filt_procdetach,
192 	.f_event = filt_proc,
193 };
194 static struct filterops timer_filtops = {
195 	.f_isfd = 0,
196 	.f_attach = filt_timerattach,
197 	.f_detach = filt_timerdetach,
198 	.f_event = filt_timer,
199 	.f_touch = filt_timertouch,
200 };
201 static struct filterops user_filtops = {
202 	.f_attach = filt_userattach,
203 	.f_detach = filt_userdetach,
204 	.f_event = filt_user,
205 	.f_touch = filt_usertouch,
206 };
207 
208 static uma_zone_t	knote_zone;
209 static unsigned int __exclusive_cache_line	kq_ncallouts;
210 static unsigned int 	kq_calloutmax = 4 * 1024;
211 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
212     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
213 
214 /* XXX - ensure not influx ? */
215 #define KNOTE_ACTIVATE(kn, islock) do { 				\
216 	if ((islock))							\
217 		mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED);		\
218 	else								\
219 		KQ_LOCK((kn)->kn_kq);					\
220 	(kn)->kn_status |= KN_ACTIVE;					\
221 	if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)		\
222 		knote_enqueue((kn));					\
223 	if (!(islock))							\
224 		KQ_UNLOCK((kn)->kn_kq);					\
225 } while (0)
226 #define KQ_LOCK(kq) do {						\
227 	mtx_lock(&(kq)->kq_lock);					\
228 } while (0)
229 #define KQ_FLUX_WAKEUP(kq) do {						\
230 	if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) {		\
231 		(kq)->kq_state &= ~KQ_FLUXWAIT;				\
232 		wakeup((kq));						\
233 	}								\
234 } while (0)
235 #define KQ_UNLOCK_FLUX(kq) do {						\
236 	KQ_FLUX_WAKEUP(kq);						\
237 	mtx_unlock(&(kq)->kq_lock);					\
238 } while (0)
239 #define KQ_UNLOCK(kq) do {						\
240 	mtx_unlock(&(kq)->kq_lock);					\
241 } while (0)
242 #define KQ_OWNED(kq) do {						\
243 	mtx_assert(&(kq)->kq_lock, MA_OWNED);				\
244 } while (0)
245 #define KQ_NOTOWNED(kq) do {						\
246 	mtx_assert(&(kq)->kq_lock, MA_NOTOWNED);			\
247 } while (0)
248 
249 static struct knlist *
250 kn_list_lock(struct knote *kn)
251 {
252 	struct knlist *knl;
253 
254 	knl = kn->kn_knlist;
255 	if (knl != NULL)
256 		knl->kl_lock(knl->kl_lockarg);
257 	return (knl);
258 }
259 
260 static void
261 kn_list_unlock(struct knlist *knl)
262 {
263 	bool do_free;
264 
265 	if (knl == NULL)
266 		return;
267 	do_free = knl->kl_autodestroy && knlist_empty(knl);
268 	knl->kl_unlock(knl->kl_lockarg);
269 	if (do_free) {
270 		knlist_destroy(knl);
271 		free(knl, M_KQUEUE);
272 	}
273 }
274 
275 static bool
276 kn_in_flux(struct knote *kn)
277 {
278 
279 	return (kn->kn_influx > 0);
280 }
281 
282 static void
283 kn_enter_flux(struct knote *kn)
284 {
285 
286 	KQ_OWNED(kn->kn_kq);
287 	MPASS(kn->kn_influx < INT_MAX);
288 	kn->kn_influx++;
289 }
290 
291 static bool
292 kn_leave_flux(struct knote *kn)
293 {
294 
295 	KQ_OWNED(kn->kn_kq);
296 	MPASS(kn->kn_influx > 0);
297 	kn->kn_influx--;
298 	return (kn->kn_influx == 0);
299 }
300 
301 #define	KNL_ASSERT_LOCK(knl, islocked) do {				\
302 	if (islocked)							\
303 		KNL_ASSERT_LOCKED(knl);				\
304 	else								\
305 		KNL_ASSERT_UNLOCKED(knl);				\
306 } while (0)
307 #ifdef INVARIANTS
308 #define	KNL_ASSERT_LOCKED(knl) do {					\
309 	knl->kl_assert_lock((knl)->kl_lockarg, LA_LOCKED);		\
310 } while (0)
311 #define	KNL_ASSERT_UNLOCKED(knl) do {					\
312 	knl->kl_assert_lock((knl)->kl_lockarg, LA_UNLOCKED);		\
313 } while (0)
314 #else /* !INVARIANTS */
315 #define	KNL_ASSERT_LOCKED(knl) do {} while (0)
316 #define	KNL_ASSERT_UNLOCKED(knl) do {} while (0)
317 #endif /* INVARIANTS */
318 
319 #ifndef	KN_HASHSIZE
320 #define	KN_HASHSIZE		64		/* XXX should be tunable */
321 #endif
322 
323 #define KN_HASH(val, mask)	(((val) ^ (val >> 8)) & (mask))
324 
325 static int
326 filt_nullattach(struct knote *kn)
327 {
328 
329 	return (ENXIO);
330 };
331 
332 struct filterops null_filtops = {
333 	.f_isfd = 0,
334 	.f_attach = filt_nullattach,
335 };
336 
337 /* XXX - make SYSINIT to add these, and move into respective modules. */
338 extern struct filterops sig_filtops;
339 extern struct filterops fs_filtops;
340 
341 /*
342  * Table for for all system-defined filters.
343  */
344 static struct mtx	filterops_lock;
345 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
346 	MTX_DEF);
347 static struct {
348 	struct filterops *for_fop;
349 	int for_nolock;
350 	int for_refcnt;
351 } sysfilt_ops[EVFILT_SYSCOUNT] = {
352 	{ &file_filtops, 1 },			/* EVFILT_READ */
353 	{ &file_filtops, 1 },			/* EVFILT_WRITE */
354 	{ &null_filtops },			/* EVFILT_AIO */
355 	{ &file_filtops, 1 },			/* EVFILT_VNODE */
356 	{ &proc_filtops, 1 },			/* EVFILT_PROC */
357 	{ &sig_filtops, 1 },			/* EVFILT_SIGNAL */
358 	{ &timer_filtops, 1 },			/* EVFILT_TIMER */
359 	{ &file_filtops, 1 },			/* EVFILT_PROCDESC */
360 	{ &fs_filtops, 1 },			/* EVFILT_FS */
361 	{ &null_filtops },			/* EVFILT_LIO */
362 	{ &user_filtops, 1 },			/* EVFILT_USER */
363 	{ &null_filtops },			/* EVFILT_SENDFILE */
364 	{ &file_filtops, 1 },                   /* EVFILT_EMPTY */
365 };
366 
367 /*
368  * Simple redirection for all cdevsw style objects to call their fo_kqfilter
369  * method.
370  */
371 static int
372 filt_fileattach(struct knote *kn)
373 {
374 
375 	return (fo_kqfilter(kn->kn_fp, kn));
376 }
377 
378 /*ARGSUSED*/
379 static int
380 kqueue_kqfilter(struct file *fp, struct knote *kn)
381 {
382 	struct kqueue *kq = kn->kn_fp->f_data;
383 
384 	if (kn->kn_filter != EVFILT_READ)
385 		return (EINVAL);
386 
387 	kn->kn_status |= KN_KQUEUE;
388 	kn->kn_fop = &kqread_filtops;
389 	knlist_add(&kq->kq_sel.si_note, kn, 0);
390 
391 	return (0);
392 }
393 
394 static void
395 filt_kqdetach(struct knote *kn)
396 {
397 	struct kqueue *kq = kn->kn_fp->f_data;
398 
399 	knlist_remove(&kq->kq_sel.si_note, kn, 0);
400 }
401 
402 /*ARGSUSED*/
403 static int
404 filt_kqueue(struct knote *kn, long hint)
405 {
406 	struct kqueue *kq = kn->kn_fp->f_data;
407 
408 	kn->kn_data = kq->kq_count;
409 	return (kn->kn_data > 0);
410 }
411 
412 /* XXX - move to kern_proc.c?  */
413 static int
414 filt_procattach(struct knote *kn)
415 {
416 	struct proc *p;
417 	int error;
418 	bool exiting, immediate;
419 
420 	exiting = immediate = false;
421 	if (kn->kn_sfflags & NOTE_EXIT)
422 		p = pfind_any(kn->kn_id);
423 	else
424 		p = pfind(kn->kn_id);
425 	if (p == NULL)
426 		return (ESRCH);
427 	if (p->p_flag & P_WEXIT)
428 		exiting = true;
429 
430 	if ((error = p_cansee(curthread, p))) {
431 		PROC_UNLOCK(p);
432 		return (error);
433 	}
434 
435 	kn->kn_ptr.p_proc = p;
436 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
437 
438 	/*
439 	 * Internal flag indicating registration done by kernel for the
440 	 * purposes of getting a NOTE_CHILD notification.
441 	 */
442 	if (kn->kn_flags & EV_FLAG2) {
443 		kn->kn_flags &= ~EV_FLAG2;
444 		kn->kn_data = kn->kn_sdata;		/* ppid */
445 		kn->kn_fflags = NOTE_CHILD;
446 		kn->kn_sfflags &= ~(NOTE_EXIT | NOTE_EXEC | NOTE_FORK);
447 		immediate = true; /* Force immediate activation of child note. */
448 	}
449 	/*
450 	 * Internal flag indicating registration done by kernel (for other than
451 	 * NOTE_CHILD).
452 	 */
453 	if (kn->kn_flags & EV_FLAG1) {
454 		kn->kn_flags &= ~EV_FLAG1;
455 	}
456 
457 	knlist_add(p->p_klist, kn, 1);
458 
459 	/*
460 	 * Immediately activate any child notes or, in the case of a zombie
461 	 * target process, exit notes.  The latter is necessary to handle the
462 	 * case where the target process, e.g. a child, dies before the kevent
463 	 * is registered.
464 	 */
465 	if (immediate || (exiting && filt_proc(kn, NOTE_EXIT)))
466 		KNOTE_ACTIVATE(kn, 0);
467 
468 	PROC_UNLOCK(p);
469 
470 	return (0);
471 }
472 
473 /*
474  * The knote may be attached to a different process, which may exit,
475  * leaving nothing for the knote to be attached to.  So when the process
476  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
477  * it will be deleted when read out.  However, as part of the knote deletion,
478  * this routine is called, so a check is needed to avoid actually performing
479  * a detach, because the original process does not exist any more.
480  */
481 /* XXX - move to kern_proc.c?  */
482 static void
483 filt_procdetach(struct knote *kn)
484 {
485 
486 	knlist_remove(kn->kn_knlist, kn, 0);
487 	kn->kn_ptr.p_proc = NULL;
488 }
489 
490 /* XXX - move to kern_proc.c?  */
491 static int
492 filt_proc(struct knote *kn, long hint)
493 {
494 	struct proc *p;
495 	u_int event;
496 
497 	p = kn->kn_ptr.p_proc;
498 	if (p == NULL) /* already activated, from attach filter */
499 		return (0);
500 
501 	/* Mask off extra data. */
502 	event = (u_int)hint & NOTE_PCTRLMASK;
503 
504 	/* If the user is interested in this event, record it. */
505 	if (kn->kn_sfflags & event)
506 		kn->kn_fflags |= event;
507 
508 	/* Process is gone, so flag the event as finished. */
509 	if (event == NOTE_EXIT) {
510 		kn->kn_flags |= EV_EOF | EV_ONESHOT;
511 		kn->kn_ptr.p_proc = NULL;
512 		if (kn->kn_fflags & NOTE_EXIT)
513 			kn->kn_data = KW_EXITCODE(p->p_xexit, p->p_xsig);
514 		if (kn->kn_fflags == 0)
515 			kn->kn_flags |= EV_DROP;
516 		return (1);
517 	}
518 
519 	return (kn->kn_fflags != 0);
520 }
521 
522 /*
523  * Called when the process forked. It mostly does the same as the
524  * knote(), activating all knotes registered to be activated when the
525  * process forked. Additionally, for each knote attached to the
526  * parent, check whether user wants to track the new process. If so
527  * attach a new knote to it, and immediately report an event with the
528  * child's pid.
529  */
530 void
531 knote_fork(struct knlist *list, int pid)
532 {
533 	struct kqueue *kq;
534 	struct knote *kn;
535 	struct kevent kev;
536 	int error;
537 
538 	MPASS(list != NULL);
539 	KNL_ASSERT_LOCKED(list);
540 	if (SLIST_EMPTY(&list->kl_list))
541 		return;
542 
543 	memset(&kev, 0, sizeof(kev));
544 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
545 		kq = kn->kn_kq;
546 		KQ_LOCK(kq);
547 		if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
548 			KQ_UNLOCK(kq);
549 			continue;
550 		}
551 
552 		/*
553 		 * The same as knote(), activate the event.
554 		 */
555 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
556 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
557 				KNOTE_ACTIVATE(kn, 1);
558 			KQ_UNLOCK(kq);
559 			continue;
560 		}
561 
562 		/*
563 		 * The NOTE_TRACK case. In addition to the activation
564 		 * of the event, we need to register new events to
565 		 * track the child. Drop the locks in preparation for
566 		 * the call to kqueue_register().
567 		 */
568 		kn_enter_flux(kn);
569 		KQ_UNLOCK(kq);
570 		list->kl_unlock(list->kl_lockarg);
571 
572 		/*
573 		 * Activate existing knote and register tracking knotes with
574 		 * new process.
575 		 *
576 		 * First register a knote to get just the child notice. This
577 		 * must be a separate note from a potential NOTE_EXIT
578 		 * notification since both NOTE_CHILD and NOTE_EXIT are defined
579 		 * to use the data field (in conflicting ways).
580 		 */
581 		kev.ident = pid;
582 		kev.filter = kn->kn_filter;
583 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_ONESHOT |
584 		    EV_FLAG2;
585 		kev.fflags = kn->kn_sfflags;
586 		kev.data = kn->kn_id;		/* parent */
587 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
588 		error = kqueue_register(kq, &kev, NULL, M_NOWAIT);
589 		if (error)
590 			kn->kn_fflags |= NOTE_TRACKERR;
591 
592 		/*
593 		 * Then register another knote to track other potential events
594 		 * from the new process.
595 		 */
596 		kev.ident = pid;
597 		kev.filter = kn->kn_filter;
598 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
599 		kev.fflags = kn->kn_sfflags;
600 		kev.data = kn->kn_id;		/* parent */
601 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
602 		error = kqueue_register(kq, &kev, NULL, M_NOWAIT);
603 		if (error)
604 			kn->kn_fflags |= NOTE_TRACKERR;
605 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
606 			KNOTE_ACTIVATE(kn, 0);
607 		list->kl_lock(list->kl_lockarg);
608 		KQ_LOCK(kq);
609 		kn_leave_flux(kn);
610 		KQ_UNLOCK_FLUX(kq);
611 	}
612 }
613 
614 /*
615  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
616  * interval timer support code.
617  */
618 
619 #define NOTE_TIMER_PRECMASK						\
620     (NOTE_SECONDS | NOTE_MSECONDS | NOTE_USECONDS | NOTE_NSECONDS)
621 
622 static sbintime_t
623 timer2sbintime(int64_t data, int flags)
624 {
625 	int64_t secs;
626 
627         /*
628          * Macros for converting to the fractional second portion of an
629          * sbintime_t using 64bit multiplication to improve precision.
630          */
631 #define NS_TO_SBT(ns) (((ns) * (((uint64_t)1 << 63) / 500000000)) >> 32)
632 #define US_TO_SBT(us) (((us) * (((uint64_t)1 << 63) / 500000)) >> 32)
633 #define MS_TO_SBT(ms) (((ms) * (((uint64_t)1 << 63) / 500)) >> 32)
634 	switch (flags & NOTE_TIMER_PRECMASK) {
635 	case NOTE_SECONDS:
636 #ifdef __LP64__
637 		if (data > (SBT_MAX / SBT_1S))
638 			return (SBT_MAX);
639 #endif
640 		return ((sbintime_t)data << 32);
641 	case NOTE_MSECONDS: /* FALLTHROUGH */
642 	case 0:
643 		if (data >= 1000) {
644 			secs = data / 1000;
645 #ifdef __LP64__
646 			if (secs > (SBT_MAX / SBT_1S))
647 				return (SBT_MAX);
648 #endif
649 			return (secs << 32 | MS_TO_SBT(data % 1000));
650 		}
651 		return (MS_TO_SBT(data));
652 	case NOTE_USECONDS:
653 		if (data >= 1000000) {
654 			secs = data / 1000000;
655 #ifdef __LP64__
656 			if (secs > (SBT_MAX / SBT_1S))
657 				return (SBT_MAX);
658 #endif
659 			return (secs << 32 | US_TO_SBT(data % 1000000));
660 		}
661 		return (US_TO_SBT(data));
662 	case NOTE_NSECONDS:
663 		if (data >= 1000000000) {
664 			secs = data / 1000000000;
665 #ifdef __LP64__
666 			if (secs > (SBT_MAX / SBT_1S))
667 				return (SBT_MAX);
668 #endif
669 			return (secs << 32 | NS_TO_SBT(data % 1000000000));
670 		}
671 		return (NS_TO_SBT(data));
672 	default:
673 		break;
674 	}
675 	return (-1);
676 }
677 
678 struct kq_timer_cb_data {
679 	struct callout c;
680 	struct proc *p;
681 	struct knote *kn;
682 	int cpuid;
683 	int flags;
684 	TAILQ_ENTRY(kq_timer_cb_data) link;
685 	sbintime_t next;	/* next timer event fires at */
686 	sbintime_t to;		/* precalculated timer period, 0 for abs */
687 };
688 
689 #define	KQ_TIMER_CB_ENQUEUED	0x01
690 
691 static void
692 kqtimer_sched_callout(struct kq_timer_cb_data *kc)
693 {
694 	callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kc->kn,
695 	    kc->cpuid, C_ABSOLUTE);
696 }
697 
698 void
699 kqtimer_proc_continue(struct proc *p)
700 {
701 	struct kq_timer_cb_data *kc, *kc1;
702 	struct bintime bt;
703 	sbintime_t now;
704 
705 	PROC_LOCK_ASSERT(p, MA_OWNED);
706 
707 	getboottimebin(&bt);
708 	now = bttosbt(bt);
709 
710 	TAILQ_FOREACH_SAFE(kc, &p->p_kqtim_stop, link, kc1) {
711 		TAILQ_REMOVE(&p->p_kqtim_stop, kc, link);
712 		kc->flags &= ~KQ_TIMER_CB_ENQUEUED;
713 		if (kc->next <= now)
714 			filt_timerexpire_l(kc->kn, true);
715 		else
716 			kqtimer_sched_callout(kc);
717 	}
718 }
719 
720 static void
721 filt_timerexpire_l(struct knote *kn, bool proc_locked)
722 {
723 	struct kq_timer_cb_data *kc;
724 	struct proc *p;
725 	uint64_t delta;
726 	sbintime_t now;
727 
728 	kc = kn->kn_ptr.p_v;
729 
730 	if ((kn->kn_flags & EV_ONESHOT) != 0 || kc->to == 0) {
731 		kn->kn_data++;
732 		KNOTE_ACTIVATE(kn, 0);
733 		return;
734 	}
735 
736 	now = sbinuptime();
737 	if (now >= kc->next) {
738 		delta = (now - kc->next) / kc->to;
739 		if (delta == 0)
740 			delta = 1;
741 		kn->kn_data += delta;
742 		kc->next += (delta + 1) * kc->to;
743 		if (now >= kc->next)	/* overflow */
744 			kc->next = now + kc->to;
745 		KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
746 	}
747 
748 	/*
749 	 * Initial check for stopped kc->p is racy.  It is fine to
750 	 * miss the set of the stop flags, at worst we would schedule
751 	 * one more callout.  On the other hand, it is not fine to not
752 	 * schedule when we we missed clearing of the flags, we
753 	 * recheck them under the lock and observe consistent state.
754 	 */
755 	p = kc->p;
756 	if (P_SHOULDSTOP(p) || P_KILLED(p)) {
757 		if (!proc_locked)
758 			PROC_LOCK(p);
759 		if (P_SHOULDSTOP(p) || P_KILLED(p)) {
760 			if ((kc->flags & KQ_TIMER_CB_ENQUEUED) == 0) {
761 				kc->flags |= KQ_TIMER_CB_ENQUEUED;
762 				TAILQ_INSERT_TAIL(&p->p_kqtim_stop, kc, link);
763 			}
764 			if (!proc_locked)
765 				PROC_UNLOCK(p);
766 			return;
767 		}
768 		if (!proc_locked)
769 			PROC_UNLOCK(p);
770 	}
771 	kqtimer_sched_callout(kc);
772 }
773 
774 static void
775 filt_timerexpire(void *knx)
776 {
777 	filt_timerexpire_l(knx, false);
778 }
779 
780 /*
781  * data contains amount of time to sleep
782  */
783 static int
784 filt_timervalidate(struct knote *kn, sbintime_t *to)
785 {
786 	struct bintime bt;
787 	sbintime_t sbt;
788 
789 	if (kn->kn_sdata < 0)
790 		return (EINVAL);
791 	if (kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
792 		kn->kn_sdata = 1;
793 	/*
794 	 * The only fflags values supported are the timer unit
795 	 * (precision) and the absolute time indicator.
796 	 */
797 	if ((kn->kn_sfflags & ~(NOTE_TIMER_PRECMASK | NOTE_ABSTIME)) != 0)
798 		return (EINVAL);
799 
800 	*to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
801 	if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
802 		getboottimebin(&bt);
803 		sbt = bttosbt(bt);
804 		*to -= sbt;
805 	}
806 	if (*to < 0)
807 		return (EINVAL);
808 	return (0);
809 }
810 
811 static int
812 filt_timerattach(struct knote *kn)
813 {
814 	struct kq_timer_cb_data *kc;
815 	sbintime_t to;
816 	int error;
817 
818 	error = filt_timervalidate(kn, &to);
819 	if (error != 0)
820 		return (error);
821 
822 	if (atomic_fetchadd_int(&kq_ncallouts, 1) + 1 > kq_calloutmax) {
823 		atomic_subtract_int(&kq_ncallouts, 1);
824 		return (ENOMEM);
825 	}
826 
827 	if ((kn->kn_sfflags & NOTE_ABSTIME) == 0)
828 		kn->kn_flags |= EV_CLEAR;	/* automatically set */
829 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
830 	kn->kn_ptr.p_v = kc = malloc(sizeof(*kc), M_KQUEUE, M_WAITOK);
831 	kc->kn = kn;
832 	kc->p = curproc;
833 	kc->cpuid = PCPU_GET(cpuid);
834 	kc->flags = 0;
835 	callout_init(&kc->c, 1);
836 	filt_timerstart(kn, to);
837 
838 	return (0);
839 }
840 
841 static void
842 filt_timerstart(struct knote *kn, sbintime_t to)
843 {
844 	struct kq_timer_cb_data *kc;
845 
846 	kc = kn->kn_ptr.p_v;
847 	if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
848 		kc->next = to;
849 		kc->to = 0;
850 	} else {
851 		kc->next = to + sbinuptime();
852 		kc->to = to;
853 	}
854 	kqtimer_sched_callout(kc);
855 }
856 
857 static void
858 filt_timerdetach(struct knote *kn)
859 {
860 	struct kq_timer_cb_data *kc;
861 	unsigned int old __unused;
862 
863 	kc = kn->kn_ptr.p_v;
864 	callout_drain(&kc->c);
865 	if ((kc->flags & KQ_TIMER_CB_ENQUEUED) != 0) {
866 		PROC_LOCK(kc->p);
867 		TAILQ_REMOVE(&kc->p->p_kqtim_stop, kc, link);
868 		PROC_UNLOCK(kc->p);
869 	}
870 	free(kc, M_KQUEUE);
871 	old = atomic_fetchadd_int(&kq_ncallouts, -1);
872 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
873 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
874 }
875 
876 static void
877 filt_timertouch(struct knote *kn, struct kevent *kev, u_long type)
878 {
879 	struct kq_timer_cb_data *kc;
880 	struct kqueue *kq;
881 	sbintime_t to;
882 	int error;
883 
884 	switch (type) {
885 	case EVENT_REGISTER:
886 		/* Handle re-added timers that update data/fflags */
887 		if (kev->flags & EV_ADD) {
888 			kc = kn->kn_ptr.p_v;
889 
890 			/* Drain any existing callout. */
891 			callout_drain(&kc->c);
892 
893 			/* Throw away any existing undelivered record
894 			 * of the timer expiration. This is done under
895 			 * the presumption that if a process is
896 			 * re-adding this timer with new parameters,
897 			 * it is no longer interested in what may have
898 			 * happened under the old parameters. If it is
899 			 * interested, it can wait for the expiration,
900 			 * delete the old timer definition, and then
901 			 * add the new one.
902 			 *
903 			 * This has to be done while the kq is locked:
904 			 *   - if enqueued, dequeue
905 			 *   - make it no longer active
906 			 *   - clear the count of expiration events
907 			 */
908 			kq = kn->kn_kq;
909 			KQ_LOCK(kq);
910 			if (kn->kn_status & KN_QUEUED)
911 				knote_dequeue(kn);
912 
913 			kn->kn_status &= ~KN_ACTIVE;
914 			kn->kn_data = 0;
915 			KQ_UNLOCK(kq);
916 
917 			/* Reschedule timer based on new data/fflags */
918 			kn->kn_sfflags = kev->fflags;
919 			kn->kn_sdata = kev->data;
920 			error = filt_timervalidate(kn, &to);
921 			if (error != 0) {
922 			  	kn->kn_flags |= EV_ERROR;
923 				kn->kn_data = error;
924 			} else
925 			  	filt_timerstart(kn, to);
926 		}
927 		break;
928 
929         case EVENT_PROCESS:
930 		*kev = kn->kn_kevent;
931 		if (kn->kn_flags & EV_CLEAR) {
932 			kn->kn_data = 0;
933 			kn->kn_fflags = 0;
934 		}
935 		break;
936 
937 	default:
938 		panic("filt_timertouch() - invalid type (%ld)", type);
939 		break;
940 	}
941 }
942 
943 static int
944 filt_timer(struct knote *kn, long hint)
945 {
946 
947 	return (kn->kn_data != 0);
948 }
949 
950 static int
951 filt_userattach(struct knote *kn)
952 {
953 
954 	/*
955 	 * EVFILT_USER knotes are not attached to anything in the kernel.
956 	 */
957 	kn->kn_hook = NULL;
958 	if (kn->kn_fflags & NOTE_TRIGGER)
959 		kn->kn_hookid = 1;
960 	else
961 		kn->kn_hookid = 0;
962 	return (0);
963 }
964 
965 static void
966 filt_userdetach(__unused struct knote *kn)
967 {
968 
969 	/*
970 	 * EVFILT_USER knotes are not attached to anything in the kernel.
971 	 */
972 }
973 
974 static int
975 filt_user(struct knote *kn, __unused long hint)
976 {
977 
978 	return (kn->kn_hookid);
979 }
980 
981 static void
982 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
983 {
984 	u_int ffctrl;
985 
986 	switch (type) {
987 	case EVENT_REGISTER:
988 		if (kev->fflags & NOTE_TRIGGER)
989 			kn->kn_hookid = 1;
990 
991 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
992 		kev->fflags &= NOTE_FFLAGSMASK;
993 		switch (ffctrl) {
994 		case NOTE_FFNOP:
995 			break;
996 
997 		case NOTE_FFAND:
998 			kn->kn_sfflags &= kev->fflags;
999 			break;
1000 
1001 		case NOTE_FFOR:
1002 			kn->kn_sfflags |= kev->fflags;
1003 			break;
1004 
1005 		case NOTE_FFCOPY:
1006 			kn->kn_sfflags = kev->fflags;
1007 			break;
1008 
1009 		default:
1010 			/* XXX Return error? */
1011 			break;
1012 		}
1013 		kn->kn_sdata = kev->data;
1014 		if (kev->flags & EV_CLEAR) {
1015 			kn->kn_hookid = 0;
1016 			kn->kn_data = 0;
1017 			kn->kn_fflags = 0;
1018 		}
1019 		break;
1020 
1021         case EVENT_PROCESS:
1022 		*kev = kn->kn_kevent;
1023 		kev->fflags = kn->kn_sfflags;
1024 		kev->data = kn->kn_sdata;
1025 		if (kn->kn_flags & EV_CLEAR) {
1026 			kn->kn_hookid = 0;
1027 			kn->kn_data = 0;
1028 			kn->kn_fflags = 0;
1029 		}
1030 		break;
1031 
1032 	default:
1033 		panic("filt_usertouch() - invalid type (%ld)", type);
1034 		break;
1035 	}
1036 }
1037 
1038 int
1039 sys_kqueue(struct thread *td, struct kqueue_args *uap)
1040 {
1041 
1042 	return (kern_kqueue(td, 0, NULL));
1043 }
1044 
1045 static void
1046 kqueue_init(struct kqueue *kq)
1047 {
1048 
1049 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
1050 	TAILQ_INIT(&kq->kq_head);
1051 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
1052 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
1053 }
1054 
1055 int
1056 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
1057 {
1058 	struct filedesc *fdp;
1059 	struct kqueue *kq;
1060 	struct file *fp;
1061 	struct ucred *cred;
1062 	int fd, error;
1063 
1064 	fdp = td->td_proc->p_fd;
1065 	cred = td->td_ucred;
1066 	if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
1067 		return (ENOMEM);
1068 
1069 	error = falloc_caps(td, &fp, &fd, flags, fcaps);
1070 	if (error != 0) {
1071 		chgkqcnt(cred->cr_ruidinfo, -1, 0);
1072 		return (error);
1073 	}
1074 
1075 	/* An extra reference on `fp' has been held for us by falloc(). */
1076 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
1077 	kqueue_init(kq);
1078 	kq->kq_fdp = fdp;
1079 	kq->kq_cred = crhold(cred);
1080 
1081 	FILEDESC_XLOCK(fdp);
1082 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
1083 	FILEDESC_XUNLOCK(fdp);
1084 
1085 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
1086 	fdrop(fp, td);
1087 
1088 	td->td_retval[0] = fd;
1089 	return (0);
1090 }
1091 
1092 struct g_kevent_args {
1093 	int	fd;
1094 	void	*changelist;
1095 	int	nchanges;
1096 	void	*eventlist;
1097 	int	nevents;
1098 	const struct timespec *timeout;
1099 };
1100 
1101 int
1102 sys_kevent(struct thread *td, struct kevent_args *uap)
1103 {
1104 	struct kevent_copyops k_ops = {
1105 		.arg = uap,
1106 		.k_copyout = kevent_copyout,
1107 		.k_copyin = kevent_copyin,
1108 		.kevent_size = sizeof(struct kevent),
1109 	};
1110 	struct g_kevent_args gk_args = {
1111 		.fd = uap->fd,
1112 		.changelist = uap->changelist,
1113 		.nchanges = uap->nchanges,
1114 		.eventlist = uap->eventlist,
1115 		.nevents = uap->nevents,
1116 		.timeout = uap->timeout,
1117 	};
1118 
1119 	return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent"));
1120 }
1121 
1122 static int
1123 kern_kevent_generic(struct thread *td, struct g_kevent_args *uap,
1124     struct kevent_copyops *k_ops, const char *struct_name)
1125 {
1126 	struct timespec ts, *tsp;
1127 #ifdef KTRACE
1128 	struct kevent *eventlist = uap->eventlist;
1129 #endif
1130 	int error;
1131 
1132 	if (uap->timeout != NULL) {
1133 		error = copyin(uap->timeout, &ts, sizeof(ts));
1134 		if (error)
1135 			return (error);
1136 		tsp = &ts;
1137 	} else
1138 		tsp = NULL;
1139 
1140 #ifdef KTRACE
1141 	if (KTRPOINT(td, KTR_STRUCT_ARRAY))
1142 		ktrstructarray(struct_name, UIO_USERSPACE, uap->changelist,
1143 		    uap->nchanges, k_ops->kevent_size);
1144 #endif
1145 
1146 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
1147 	    k_ops, tsp);
1148 
1149 #ifdef KTRACE
1150 	if (error == 0 && KTRPOINT(td, KTR_STRUCT_ARRAY))
1151 		ktrstructarray(struct_name, UIO_USERSPACE, eventlist,
1152 		    td->td_retval[0], k_ops->kevent_size);
1153 #endif
1154 
1155 	return (error);
1156 }
1157 
1158 /*
1159  * Copy 'count' items into the destination list pointed to by uap->eventlist.
1160  */
1161 static int
1162 kevent_copyout(void *arg, struct kevent *kevp, int count)
1163 {
1164 	struct kevent_args *uap;
1165 	int error;
1166 
1167 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1168 	uap = (struct kevent_args *)arg;
1169 
1170 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
1171 	if (error == 0)
1172 		uap->eventlist += count;
1173 	return (error);
1174 }
1175 
1176 /*
1177  * Copy 'count' items from the list pointed to by uap->changelist.
1178  */
1179 static int
1180 kevent_copyin(void *arg, struct kevent *kevp, int count)
1181 {
1182 	struct kevent_args *uap;
1183 	int error;
1184 
1185 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1186 	uap = (struct kevent_args *)arg;
1187 
1188 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
1189 	if (error == 0)
1190 		uap->changelist += count;
1191 	return (error);
1192 }
1193 
1194 #ifdef COMPAT_FREEBSD11
1195 static int
1196 kevent11_copyout(void *arg, struct kevent *kevp, int count)
1197 {
1198 	struct freebsd11_kevent_args *uap;
1199 	struct kevent_freebsd11 kev11;
1200 	int error, i;
1201 
1202 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1203 	uap = (struct freebsd11_kevent_args *)arg;
1204 
1205 	for (i = 0; i < count; i++) {
1206 		kev11.ident = kevp->ident;
1207 		kev11.filter = kevp->filter;
1208 		kev11.flags = kevp->flags;
1209 		kev11.fflags = kevp->fflags;
1210 		kev11.data = kevp->data;
1211 		kev11.udata = kevp->udata;
1212 		error = copyout(&kev11, uap->eventlist, sizeof(kev11));
1213 		if (error != 0)
1214 			break;
1215 		uap->eventlist++;
1216 		kevp++;
1217 	}
1218 	return (error);
1219 }
1220 
1221 /*
1222  * Copy 'count' items from the list pointed to by uap->changelist.
1223  */
1224 static int
1225 kevent11_copyin(void *arg, struct kevent *kevp, int count)
1226 {
1227 	struct freebsd11_kevent_args *uap;
1228 	struct kevent_freebsd11 kev11;
1229 	int error, i;
1230 
1231 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1232 	uap = (struct freebsd11_kevent_args *)arg;
1233 
1234 	for (i = 0; i < count; i++) {
1235 		error = copyin(uap->changelist, &kev11, sizeof(kev11));
1236 		if (error != 0)
1237 			break;
1238 		kevp->ident = kev11.ident;
1239 		kevp->filter = kev11.filter;
1240 		kevp->flags = kev11.flags;
1241 		kevp->fflags = kev11.fflags;
1242 		kevp->data = (uintptr_t)kev11.data;
1243 		kevp->udata = kev11.udata;
1244 		bzero(&kevp->ext, sizeof(kevp->ext));
1245 		uap->changelist++;
1246 		kevp++;
1247 	}
1248 	return (error);
1249 }
1250 
1251 int
1252 freebsd11_kevent(struct thread *td, struct freebsd11_kevent_args *uap)
1253 {
1254 	struct kevent_copyops k_ops = {
1255 		.arg = uap,
1256 		.k_copyout = kevent11_copyout,
1257 		.k_copyin = kevent11_copyin,
1258 		.kevent_size = sizeof(struct kevent_freebsd11),
1259 	};
1260 	struct g_kevent_args gk_args = {
1261 		.fd = uap->fd,
1262 		.changelist = uap->changelist,
1263 		.nchanges = uap->nchanges,
1264 		.eventlist = uap->eventlist,
1265 		.nevents = uap->nevents,
1266 		.timeout = uap->timeout,
1267 	};
1268 
1269 	return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent_freebsd11"));
1270 }
1271 #endif
1272 
1273 int
1274 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
1275     struct kevent_copyops *k_ops, const struct timespec *timeout)
1276 {
1277 	cap_rights_t rights;
1278 	struct file *fp;
1279 	int error;
1280 
1281 	cap_rights_init_zero(&rights);
1282 	if (nchanges > 0)
1283 		cap_rights_set_one(&rights, CAP_KQUEUE_CHANGE);
1284 	if (nevents > 0)
1285 		cap_rights_set_one(&rights, CAP_KQUEUE_EVENT);
1286 	error = fget(td, fd, &rights, &fp);
1287 	if (error != 0)
1288 		return (error);
1289 
1290 	error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
1291 	fdrop(fp, td);
1292 
1293 	return (error);
1294 }
1295 
1296 static int
1297 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
1298     struct kevent_copyops *k_ops, const struct timespec *timeout)
1299 {
1300 	struct kevent keva[KQ_NEVENTS];
1301 	struct kevent *kevp, *changes;
1302 	int i, n, nerrors, error;
1303 
1304 	if (nchanges < 0)
1305 		return (EINVAL);
1306 
1307 	nerrors = 0;
1308 	while (nchanges > 0) {
1309 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
1310 		error = k_ops->k_copyin(k_ops->arg, keva, n);
1311 		if (error)
1312 			return (error);
1313 		changes = keva;
1314 		for (i = 0; i < n; i++) {
1315 			kevp = &changes[i];
1316 			if (!kevp->filter)
1317 				continue;
1318 			kevp->flags &= ~EV_SYSFLAGS;
1319 			error = kqueue_register(kq, kevp, td, M_WAITOK);
1320 			if (error || (kevp->flags & EV_RECEIPT)) {
1321 				if (nevents == 0)
1322 					return (error);
1323 				kevp->flags = EV_ERROR;
1324 				kevp->data = error;
1325 				(void)k_ops->k_copyout(k_ops->arg, kevp, 1);
1326 				nevents--;
1327 				nerrors++;
1328 			}
1329 		}
1330 		nchanges -= n;
1331 	}
1332 	if (nerrors) {
1333 		td->td_retval[0] = nerrors;
1334 		return (0);
1335 	}
1336 
1337 	return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
1338 }
1339 
1340 int
1341 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
1342     struct kevent_copyops *k_ops, const struct timespec *timeout)
1343 {
1344 	struct kqueue *kq;
1345 	int error;
1346 
1347 	error = kqueue_acquire(fp, &kq);
1348 	if (error != 0)
1349 		return (error);
1350 	error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
1351 	kqueue_release(kq, 0);
1352 	return (error);
1353 }
1354 
1355 /*
1356  * Performs a kevent() call on a temporarily created kqueue. This can be
1357  * used to perform one-shot polling, similar to poll() and select().
1358  */
1359 int
1360 kern_kevent_anonymous(struct thread *td, int nevents,
1361     struct kevent_copyops *k_ops)
1362 {
1363 	struct kqueue kq = {};
1364 	int error;
1365 
1366 	kqueue_init(&kq);
1367 	kq.kq_refcnt = 1;
1368 	error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
1369 	kqueue_drain(&kq, td);
1370 	kqueue_destroy(&kq);
1371 	return (error);
1372 }
1373 
1374 int
1375 kqueue_add_filteropts(int filt, struct filterops *filtops)
1376 {
1377 	int error;
1378 
1379 	error = 0;
1380 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1381 		printf(
1382 "trying to add a filterop that is out of range: %d is beyond %d\n",
1383 		    ~filt, EVFILT_SYSCOUNT);
1384 		return EINVAL;
1385 	}
1386 	mtx_lock(&filterops_lock);
1387 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1388 	    sysfilt_ops[~filt].for_fop != NULL)
1389 		error = EEXIST;
1390 	else {
1391 		sysfilt_ops[~filt].for_fop = filtops;
1392 		sysfilt_ops[~filt].for_refcnt = 0;
1393 	}
1394 	mtx_unlock(&filterops_lock);
1395 
1396 	return (error);
1397 }
1398 
1399 int
1400 kqueue_del_filteropts(int filt)
1401 {
1402 	int error;
1403 
1404 	error = 0;
1405 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1406 		return EINVAL;
1407 
1408 	mtx_lock(&filterops_lock);
1409 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1410 	    sysfilt_ops[~filt].for_fop == NULL)
1411 		error = EINVAL;
1412 	else if (sysfilt_ops[~filt].for_refcnt != 0)
1413 		error = EBUSY;
1414 	else {
1415 		sysfilt_ops[~filt].for_fop = &null_filtops;
1416 		sysfilt_ops[~filt].for_refcnt = 0;
1417 	}
1418 	mtx_unlock(&filterops_lock);
1419 
1420 	return error;
1421 }
1422 
1423 static struct filterops *
1424 kqueue_fo_find(int filt)
1425 {
1426 
1427 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1428 		return NULL;
1429 
1430 	if (sysfilt_ops[~filt].for_nolock)
1431 		return sysfilt_ops[~filt].for_fop;
1432 
1433 	mtx_lock(&filterops_lock);
1434 	sysfilt_ops[~filt].for_refcnt++;
1435 	if (sysfilt_ops[~filt].for_fop == NULL)
1436 		sysfilt_ops[~filt].for_fop = &null_filtops;
1437 	mtx_unlock(&filterops_lock);
1438 
1439 	return sysfilt_ops[~filt].for_fop;
1440 }
1441 
1442 static void
1443 kqueue_fo_release(int filt)
1444 {
1445 
1446 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1447 		return;
1448 
1449 	if (sysfilt_ops[~filt].for_nolock)
1450 		return;
1451 
1452 	mtx_lock(&filterops_lock);
1453 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1454 	    ("filter object refcount not valid on release"));
1455 	sysfilt_ops[~filt].for_refcnt--;
1456 	mtx_unlock(&filterops_lock);
1457 }
1458 
1459 /*
1460  * A ref to kq (obtained via kqueue_acquire) must be held.
1461  */
1462 static int
1463 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td,
1464     int mflag)
1465 {
1466 	struct filterops *fops;
1467 	struct file *fp;
1468 	struct knote *kn, *tkn;
1469 	struct knlist *knl;
1470 	int error, filt, event;
1471 	int haskqglobal, filedesc_unlock;
1472 
1473 	if ((kev->flags & (EV_ENABLE | EV_DISABLE)) == (EV_ENABLE | EV_DISABLE))
1474 		return (EINVAL);
1475 
1476 	fp = NULL;
1477 	kn = NULL;
1478 	knl = NULL;
1479 	error = 0;
1480 	haskqglobal = 0;
1481 	filedesc_unlock = 0;
1482 
1483 	filt = kev->filter;
1484 	fops = kqueue_fo_find(filt);
1485 	if (fops == NULL)
1486 		return EINVAL;
1487 
1488 	if (kev->flags & EV_ADD) {
1489 		/*
1490 		 * Prevent waiting with locks.  Non-sleepable
1491 		 * allocation failures are handled in the loop, only
1492 		 * if the spare knote appears to be actually required.
1493 		 */
1494 		tkn = knote_alloc(mflag);
1495 	} else {
1496 		tkn = NULL;
1497 	}
1498 
1499 findkn:
1500 	if (fops->f_isfd) {
1501 		KASSERT(td != NULL, ("td is NULL"));
1502 		if (kev->ident > INT_MAX)
1503 			error = EBADF;
1504 		else
1505 			error = fget(td, kev->ident, &cap_event_rights, &fp);
1506 		if (error)
1507 			goto done;
1508 
1509 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1510 		    kev->ident, M_NOWAIT) != 0) {
1511 			/* try again */
1512 			fdrop(fp, td);
1513 			fp = NULL;
1514 			error = kqueue_expand(kq, fops, kev->ident, mflag);
1515 			if (error)
1516 				goto done;
1517 			goto findkn;
1518 		}
1519 
1520 		if (fp->f_type == DTYPE_KQUEUE) {
1521 			/*
1522 			 * If we add some intelligence about what we are doing,
1523 			 * we should be able to support events on ourselves.
1524 			 * We need to know when we are doing this to prevent
1525 			 * getting both the knlist lock and the kq lock since
1526 			 * they are the same thing.
1527 			 */
1528 			if (fp->f_data == kq) {
1529 				error = EINVAL;
1530 				goto done;
1531 			}
1532 
1533 			/*
1534 			 * Pre-lock the filedesc before the global
1535 			 * lock mutex, see the comment in
1536 			 * kqueue_close().
1537 			 */
1538 			FILEDESC_XLOCK(td->td_proc->p_fd);
1539 			filedesc_unlock = 1;
1540 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1541 		}
1542 
1543 		KQ_LOCK(kq);
1544 		if (kev->ident < kq->kq_knlistsize) {
1545 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1546 				if (kev->filter == kn->kn_filter)
1547 					break;
1548 		}
1549 	} else {
1550 		if ((kev->flags & EV_ADD) == EV_ADD) {
1551 			error = kqueue_expand(kq, fops, kev->ident, mflag);
1552 			if (error != 0)
1553 				goto done;
1554 		}
1555 
1556 		KQ_LOCK(kq);
1557 
1558 		/*
1559 		 * If possible, find an existing knote to use for this kevent.
1560 		 */
1561 		if (kev->filter == EVFILT_PROC &&
1562 		    (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1563 			/* This is an internal creation of a process tracking
1564 			 * note. Don't attempt to coalesce this with an
1565 			 * existing note.
1566 			 */
1567 			;
1568 		} else if (kq->kq_knhashmask != 0) {
1569 			struct klist *list;
1570 
1571 			list = &kq->kq_knhash[
1572 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1573 			SLIST_FOREACH(kn, list, kn_link)
1574 				if (kev->ident == kn->kn_id &&
1575 				    kev->filter == kn->kn_filter)
1576 					break;
1577 		}
1578 	}
1579 
1580 	/* knote is in the process of changing, wait for it to stabilize. */
1581 	if (kn != NULL && kn_in_flux(kn)) {
1582 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1583 		if (filedesc_unlock) {
1584 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1585 			filedesc_unlock = 0;
1586 		}
1587 		kq->kq_state |= KQ_FLUXWAIT;
1588 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1589 		if (fp != NULL) {
1590 			fdrop(fp, td);
1591 			fp = NULL;
1592 		}
1593 		goto findkn;
1594 	}
1595 
1596 	/*
1597 	 * kn now contains the matching knote, or NULL if no match
1598 	 */
1599 	if (kn == NULL) {
1600 		if (kev->flags & EV_ADD) {
1601 			kn = tkn;
1602 			tkn = NULL;
1603 			if (kn == NULL) {
1604 				KQ_UNLOCK(kq);
1605 				error = ENOMEM;
1606 				goto done;
1607 			}
1608 			kn->kn_fp = fp;
1609 			kn->kn_kq = kq;
1610 			kn->kn_fop = fops;
1611 			/*
1612 			 * apply reference counts to knote structure, and
1613 			 * do not release it at the end of this routine.
1614 			 */
1615 			fops = NULL;
1616 			fp = NULL;
1617 
1618 			kn->kn_sfflags = kev->fflags;
1619 			kn->kn_sdata = kev->data;
1620 			kev->fflags = 0;
1621 			kev->data = 0;
1622 			kn->kn_kevent = *kev;
1623 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1624 			    EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1625 			kn->kn_status = KN_DETACHED;
1626 			if ((kev->flags & EV_DISABLE) != 0)
1627 				kn->kn_status |= KN_DISABLED;
1628 			kn_enter_flux(kn);
1629 
1630 			error = knote_attach(kn, kq);
1631 			KQ_UNLOCK(kq);
1632 			if (error != 0) {
1633 				tkn = kn;
1634 				goto done;
1635 			}
1636 
1637 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1638 				knote_drop_detached(kn, td);
1639 				goto done;
1640 			}
1641 			knl = kn_list_lock(kn);
1642 			goto done_ev_add;
1643 		} else {
1644 			/* No matching knote and the EV_ADD flag is not set. */
1645 			KQ_UNLOCK(kq);
1646 			error = ENOENT;
1647 			goto done;
1648 		}
1649 	}
1650 
1651 	if (kev->flags & EV_DELETE) {
1652 		kn_enter_flux(kn);
1653 		KQ_UNLOCK(kq);
1654 		knote_drop(kn, td);
1655 		goto done;
1656 	}
1657 
1658 	if (kev->flags & EV_FORCEONESHOT) {
1659 		kn->kn_flags |= EV_ONESHOT;
1660 		KNOTE_ACTIVATE(kn, 1);
1661 	}
1662 
1663 	if ((kev->flags & EV_ENABLE) != 0)
1664 		kn->kn_status &= ~KN_DISABLED;
1665 	else if ((kev->flags & EV_DISABLE) != 0)
1666 		kn->kn_status |= KN_DISABLED;
1667 
1668 	/*
1669 	 * The user may change some filter values after the initial EV_ADD,
1670 	 * but doing so will not reset any filter which has already been
1671 	 * triggered.
1672 	 */
1673 	kn->kn_status |= KN_SCAN;
1674 	kn_enter_flux(kn);
1675 	KQ_UNLOCK(kq);
1676 	knl = kn_list_lock(kn);
1677 	kn->kn_kevent.udata = kev->udata;
1678 	if (!fops->f_isfd && fops->f_touch != NULL) {
1679 		fops->f_touch(kn, kev, EVENT_REGISTER);
1680 	} else {
1681 		kn->kn_sfflags = kev->fflags;
1682 		kn->kn_sdata = kev->data;
1683 	}
1684 
1685 done_ev_add:
1686 	/*
1687 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1688 	 * the initial attach event decides that the event is "completed"
1689 	 * already, e.g., filt_procattach() is called on a zombie process.  It
1690 	 * will call filt_proc() which will remove it from the list, and NULL
1691 	 * kn_knlist.
1692 	 *
1693 	 * KN_DISABLED will be stable while the knote is in flux, so the
1694 	 * unlocked read will not race with an update.
1695 	 */
1696 	if ((kn->kn_status & KN_DISABLED) == 0)
1697 		event = kn->kn_fop->f_event(kn, 0);
1698 	else
1699 		event = 0;
1700 
1701 	KQ_LOCK(kq);
1702 	if (event)
1703 		kn->kn_status |= KN_ACTIVE;
1704 	if ((kn->kn_status & (KN_ACTIVE | KN_DISABLED | KN_QUEUED)) ==
1705 	    KN_ACTIVE)
1706 		knote_enqueue(kn);
1707 	kn->kn_status &= ~KN_SCAN;
1708 	kn_leave_flux(kn);
1709 	kn_list_unlock(knl);
1710 	KQ_UNLOCK_FLUX(kq);
1711 
1712 done:
1713 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1714 	if (filedesc_unlock)
1715 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1716 	if (fp != NULL)
1717 		fdrop(fp, td);
1718 	knote_free(tkn);
1719 	if (fops != NULL)
1720 		kqueue_fo_release(filt);
1721 	return (error);
1722 }
1723 
1724 static int
1725 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1726 {
1727 	int error;
1728 	struct kqueue *kq;
1729 
1730 	error = 0;
1731 
1732 	kq = fp->f_data;
1733 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1734 		return (EBADF);
1735 	*kqp = kq;
1736 	KQ_LOCK(kq);
1737 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1738 		KQ_UNLOCK(kq);
1739 		return (EBADF);
1740 	}
1741 	kq->kq_refcnt++;
1742 	KQ_UNLOCK(kq);
1743 
1744 	return error;
1745 }
1746 
1747 static void
1748 kqueue_release(struct kqueue *kq, int locked)
1749 {
1750 	if (locked)
1751 		KQ_OWNED(kq);
1752 	else
1753 		KQ_LOCK(kq);
1754 	kq->kq_refcnt--;
1755 	if (kq->kq_refcnt == 1)
1756 		wakeup(&kq->kq_refcnt);
1757 	if (!locked)
1758 		KQ_UNLOCK(kq);
1759 }
1760 
1761 static void
1762 kqueue_schedtask(struct kqueue *kq)
1763 {
1764 
1765 	KQ_OWNED(kq);
1766 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1767 	    ("scheduling kqueue task while draining"));
1768 
1769 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1770 		taskqueue_enqueue(taskqueue_kqueue_ctx, &kq->kq_task);
1771 		kq->kq_state |= KQ_TASKSCHED;
1772 	}
1773 }
1774 
1775 /*
1776  * Expand the kq to make sure we have storage for fops/ident pair.
1777  *
1778  * Return 0 on success (or no work necessary), return errno on failure.
1779  */
1780 static int
1781 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1782     int mflag)
1783 {
1784 	struct klist *list, *tmp_knhash, *to_free;
1785 	u_long tmp_knhashmask;
1786 	int error, fd, size;
1787 
1788 	KQ_NOTOWNED(kq);
1789 
1790 	error = 0;
1791 	to_free = NULL;
1792 	if (fops->f_isfd) {
1793 		fd = ident;
1794 		if (kq->kq_knlistsize <= fd) {
1795 			size = kq->kq_knlistsize;
1796 			while (size <= fd)
1797 				size += KQEXTENT;
1798 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1799 			if (list == NULL)
1800 				return ENOMEM;
1801 			KQ_LOCK(kq);
1802 			if ((kq->kq_state & KQ_CLOSING) != 0) {
1803 				to_free = list;
1804 				error = EBADF;
1805 			} else if (kq->kq_knlistsize > fd) {
1806 				to_free = list;
1807 			} else {
1808 				if (kq->kq_knlist != NULL) {
1809 					bcopy(kq->kq_knlist, list,
1810 					    kq->kq_knlistsize * sizeof(*list));
1811 					to_free = kq->kq_knlist;
1812 					kq->kq_knlist = NULL;
1813 				}
1814 				bzero((caddr_t)list +
1815 				    kq->kq_knlistsize * sizeof(*list),
1816 				    (size - kq->kq_knlistsize) * sizeof(*list));
1817 				kq->kq_knlistsize = size;
1818 				kq->kq_knlist = list;
1819 			}
1820 			KQ_UNLOCK(kq);
1821 		}
1822 	} else {
1823 		if (kq->kq_knhashmask == 0) {
1824 			tmp_knhash = hashinit_flags(KN_HASHSIZE, M_KQUEUE,
1825 			    &tmp_knhashmask, (mflag & M_WAITOK) != 0 ?
1826 			    HASH_WAITOK : HASH_NOWAIT);
1827 			if (tmp_knhash == NULL)
1828 				return (ENOMEM);
1829 			KQ_LOCK(kq);
1830 			if ((kq->kq_state & KQ_CLOSING) != 0) {
1831 				to_free = tmp_knhash;
1832 				error = EBADF;
1833 			} else if (kq->kq_knhashmask == 0) {
1834 				kq->kq_knhash = tmp_knhash;
1835 				kq->kq_knhashmask = tmp_knhashmask;
1836 			} else {
1837 				to_free = tmp_knhash;
1838 			}
1839 			KQ_UNLOCK(kq);
1840 		}
1841 	}
1842 	free(to_free, M_KQUEUE);
1843 
1844 	KQ_NOTOWNED(kq);
1845 	return (error);
1846 }
1847 
1848 static void
1849 kqueue_task(void *arg, int pending)
1850 {
1851 	struct kqueue *kq;
1852 	int haskqglobal;
1853 
1854 	haskqglobal = 0;
1855 	kq = arg;
1856 
1857 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1858 	KQ_LOCK(kq);
1859 
1860 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1861 
1862 	kq->kq_state &= ~KQ_TASKSCHED;
1863 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1864 		wakeup(&kq->kq_state);
1865 	}
1866 	KQ_UNLOCK(kq);
1867 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1868 }
1869 
1870 /*
1871  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1872  * We treat KN_MARKER knotes as if they are in flux.
1873  */
1874 static int
1875 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1876     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1877 {
1878 	struct kevent *kevp;
1879 	struct knote *kn, *marker;
1880 	struct knlist *knl;
1881 	sbintime_t asbt, rsbt;
1882 	int count, error, haskqglobal, influx, nkev, touch;
1883 
1884 	count = maxevents;
1885 	nkev = 0;
1886 	error = 0;
1887 	haskqglobal = 0;
1888 
1889 	if (maxevents == 0)
1890 		goto done_nl;
1891 	if (maxevents < 0) {
1892 		error = EINVAL;
1893 		goto done_nl;
1894 	}
1895 
1896 	rsbt = 0;
1897 	if (tsp != NULL) {
1898 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1899 		    tsp->tv_nsec >= 1000000000) {
1900 			error = EINVAL;
1901 			goto done_nl;
1902 		}
1903 		if (timespecisset(tsp)) {
1904 			if (tsp->tv_sec <= INT32_MAX) {
1905 				rsbt = tstosbt(*tsp);
1906 				if (TIMESEL(&asbt, rsbt))
1907 					asbt += tc_tick_sbt;
1908 				if (asbt <= SBT_MAX - rsbt)
1909 					asbt += rsbt;
1910 				else
1911 					asbt = 0;
1912 				rsbt >>= tc_precexp;
1913 			} else
1914 				asbt = 0;
1915 		} else
1916 			asbt = -1;
1917 	} else
1918 		asbt = 0;
1919 	marker = knote_alloc(M_WAITOK);
1920 	marker->kn_status = KN_MARKER;
1921 	KQ_LOCK(kq);
1922 
1923 retry:
1924 	kevp = keva;
1925 	if (kq->kq_count == 0) {
1926 		if (asbt == -1) {
1927 			error = EWOULDBLOCK;
1928 		} else {
1929 			kq->kq_state |= KQ_SLEEP;
1930 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1931 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1932 		}
1933 		if (error == 0)
1934 			goto retry;
1935 		/* don't restart after signals... */
1936 		if (error == ERESTART)
1937 			error = EINTR;
1938 		else if (error == EWOULDBLOCK)
1939 			error = 0;
1940 		goto done;
1941 	}
1942 
1943 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1944 	influx = 0;
1945 	while (count) {
1946 		KQ_OWNED(kq);
1947 		kn = TAILQ_FIRST(&kq->kq_head);
1948 
1949 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1950 		    kn_in_flux(kn)) {
1951 			if (influx) {
1952 				influx = 0;
1953 				KQ_FLUX_WAKEUP(kq);
1954 			}
1955 			kq->kq_state |= KQ_FLUXWAIT;
1956 			error = msleep(kq, &kq->kq_lock, PSOCK,
1957 			    "kqflxwt", 0);
1958 			continue;
1959 		}
1960 
1961 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1962 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1963 			kn->kn_status &= ~KN_QUEUED;
1964 			kq->kq_count--;
1965 			continue;
1966 		}
1967 		if (kn == marker) {
1968 			KQ_FLUX_WAKEUP(kq);
1969 			if (count == maxevents)
1970 				goto retry;
1971 			goto done;
1972 		}
1973 		KASSERT(!kn_in_flux(kn),
1974 		    ("knote %p is unexpectedly in flux", kn));
1975 
1976 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1977 			kn->kn_status &= ~KN_QUEUED;
1978 			kn_enter_flux(kn);
1979 			kq->kq_count--;
1980 			KQ_UNLOCK(kq);
1981 			/*
1982 			 * We don't need to lock the list since we've
1983 			 * marked it as in flux.
1984 			 */
1985 			knote_drop(kn, td);
1986 			KQ_LOCK(kq);
1987 			continue;
1988 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1989 			kn->kn_status &= ~KN_QUEUED;
1990 			kn_enter_flux(kn);
1991 			kq->kq_count--;
1992 			KQ_UNLOCK(kq);
1993 			/*
1994 			 * We don't need to lock the list since we've
1995 			 * marked the knote as being in flux.
1996 			 */
1997 			*kevp = kn->kn_kevent;
1998 			knote_drop(kn, td);
1999 			KQ_LOCK(kq);
2000 			kn = NULL;
2001 		} else {
2002 			kn->kn_status |= KN_SCAN;
2003 			kn_enter_flux(kn);
2004 			KQ_UNLOCK(kq);
2005 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
2006 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
2007 			knl = kn_list_lock(kn);
2008 			if (kn->kn_fop->f_event(kn, 0) == 0) {
2009 				KQ_LOCK(kq);
2010 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
2011 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE |
2012 				    KN_SCAN);
2013 				kn_leave_flux(kn);
2014 				kq->kq_count--;
2015 				kn_list_unlock(knl);
2016 				influx = 1;
2017 				continue;
2018 			}
2019 			touch = (!kn->kn_fop->f_isfd &&
2020 			    kn->kn_fop->f_touch != NULL);
2021 			if (touch)
2022 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
2023 			else
2024 				*kevp = kn->kn_kevent;
2025 			KQ_LOCK(kq);
2026 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
2027 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
2028 				/*
2029 				 * Manually clear knotes who weren't
2030 				 * 'touch'ed.
2031 				 */
2032 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
2033 					kn->kn_data = 0;
2034 					kn->kn_fflags = 0;
2035 				}
2036 				if (kn->kn_flags & EV_DISPATCH)
2037 					kn->kn_status |= KN_DISABLED;
2038 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
2039 				kq->kq_count--;
2040 			} else
2041 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2042 
2043 			kn->kn_status &= ~KN_SCAN;
2044 			kn_leave_flux(kn);
2045 			kn_list_unlock(knl);
2046 			influx = 1;
2047 		}
2048 
2049 		/* we are returning a copy to the user */
2050 		kevp++;
2051 		nkev++;
2052 		count--;
2053 
2054 		if (nkev == KQ_NEVENTS) {
2055 			influx = 0;
2056 			KQ_UNLOCK_FLUX(kq);
2057 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
2058 			nkev = 0;
2059 			kevp = keva;
2060 			KQ_LOCK(kq);
2061 			if (error)
2062 				break;
2063 		}
2064 	}
2065 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
2066 done:
2067 	KQ_OWNED(kq);
2068 	KQ_UNLOCK_FLUX(kq);
2069 	knote_free(marker);
2070 done_nl:
2071 	KQ_NOTOWNED(kq);
2072 	if (nkev != 0)
2073 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
2074 	td->td_retval[0] = maxevents - count;
2075 	return (error);
2076 }
2077 
2078 /*ARGSUSED*/
2079 static int
2080 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
2081 	struct ucred *active_cred, struct thread *td)
2082 {
2083 	/*
2084 	 * Enabling sigio causes two major problems:
2085 	 * 1) infinite recursion:
2086 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
2087 	 * set.  On receipt of a signal this will cause a kqueue to recurse
2088 	 * into itself over and over.  Sending the sigio causes the kqueue
2089 	 * to become ready, which in turn posts sigio again, forever.
2090 	 * Solution: this can be solved by setting a flag in the kqueue that
2091 	 * we have a SIGIO in progress.
2092 	 * 2) locking problems:
2093 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
2094 	 * us above the proc and pgrp locks.
2095 	 * Solution: Post a signal using an async mechanism, being sure to
2096 	 * record a generation count in the delivery so that we do not deliver
2097 	 * a signal to the wrong process.
2098 	 *
2099 	 * Note, these two mechanisms are somewhat mutually exclusive!
2100 	 */
2101 #if 0
2102 	struct kqueue *kq;
2103 
2104 	kq = fp->f_data;
2105 	switch (cmd) {
2106 	case FIOASYNC:
2107 		if (*(int *)data) {
2108 			kq->kq_state |= KQ_ASYNC;
2109 		} else {
2110 			kq->kq_state &= ~KQ_ASYNC;
2111 		}
2112 		return (0);
2113 
2114 	case FIOSETOWN:
2115 		return (fsetown(*(int *)data, &kq->kq_sigio));
2116 
2117 	case FIOGETOWN:
2118 		*(int *)data = fgetown(&kq->kq_sigio);
2119 		return (0);
2120 	}
2121 #endif
2122 
2123 	return (ENOTTY);
2124 }
2125 
2126 /*ARGSUSED*/
2127 static int
2128 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
2129 	struct thread *td)
2130 {
2131 	struct kqueue *kq;
2132 	int revents = 0;
2133 	int error;
2134 
2135 	if ((error = kqueue_acquire(fp, &kq)))
2136 		return POLLERR;
2137 
2138 	KQ_LOCK(kq);
2139 	if (events & (POLLIN | POLLRDNORM)) {
2140 		if (kq->kq_count) {
2141 			revents |= events & (POLLIN | POLLRDNORM);
2142 		} else {
2143 			selrecord(td, &kq->kq_sel);
2144 			if (SEL_WAITING(&kq->kq_sel))
2145 				kq->kq_state |= KQ_SEL;
2146 		}
2147 	}
2148 	kqueue_release(kq, 1);
2149 	KQ_UNLOCK(kq);
2150 	return (revents);
2151 }
2152 
2153 /*ARGSUSED*/
2154 static int
2155 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
2156 	struct thread *td)
2157 {
2158 
2159 	bzero((void *)st, sizeof *st);
2160 	/*
2161 	 * We no longer return kq_count because the unlocked value is useless.
2162 	 * If you spent all this time getting the count, why not spend your
2163 	 * syscall better by calling kevent?
2164 	 *
2165 	 * XXX - This is needed for libc_r.
2166 	 */
2167 	st->st_mode = S_IFIFO;
2168 	return (0);
2169 }
2170 
2171 static void
2172 kqueue_drain(struct kqueue *kq, struct thread *td)
2173 {
2174 	struct knote *kn;
2175 	int i;
2176 
2177 	KQ_LOCK(kq);
2178 
2179 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
2180 	    ("kqueue already closing"));
2181 	kq->kq_state |= KQ_CLOSING;
2182 	if (kq->kq_refcnt > 1)
2183 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
2184 
2185 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
2186 
2187 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
2188 	    ("kqueue's knlist not empty"));
2189 
2190 	for (i = 0; i < kq->kq_knlistsize; i++) {
2191 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
2192 			if (kn_in_flux(kn)) {
2193 				kq->kq_state |= KQ_FLUXWAIT;
2194 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
2195 				continue;
2196 			}
2197 			kn_enter_flux(kn);
2198 			KQ_UNLOCK(kq);
2199 			knote_drop(kn, td);
2200 			KQ_LOCK(kq);
2201 		}
2202 	}
2203 	if (kq->kq_knhashmask != 0) {
2204 		for (i = 0; i <= kq->kq_knhashmask; i++) {
2205 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
2206 				if (kn_in_flux(kn)) {
2207 					kq->kq_state |= KQ_FLUXWAIT;
2208 					msleep(kq, &kq->kq_lock, PSOCK,
2209 					       "kqclo2", 0);
2210 					continue;
2211 				}
2212 				kn_enter_flux(kn);
2213 				KQ_UNLOCK(kq);
2214 				knote_drop(kn, td);
2215 				KQ_LOCK(kq);
2216 			}
2217 		}
2218 	}
2219 
2220 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
2221 		kq->kq_state |= KQ_TASKDRAIN;
2222 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
2223 	}
2224 
2225 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2226 		selwakeuppri(&kq->kq_sel, PSOCK);
2227 		if (!SEL_WAITING(&kq->kq_sel))
2228 			kq->kq_state &= ~KQ_SEL;
2229 	}
2230 
2231 	KQ_UNLOCK(kq);
2232 }
2233 
2234 static void
2235 kqueue_destroy(struct kqueue *kq)
2236 {
2237 
2238 	KASSERT(kq->kq_fdp == NULL,
2239 	    ("kqueue still attached to a file descriptor"));
2240 	seldrain(&kq->kq_sel);
2241 	knlist_destroy(&kq->kq_sel.si_note);
2242 	mtx_destroy(&kq->kq_lock);
2243 
2244 	if (kq->kq_knhash != NULL)
2245 		free(kq->kq_knhash, M_KQUEUE);
2246 	if (kq->kq_knlist != NULL)
2247 		free(kq->kq_knlist, M_KQUEUE);
2248 
2249 	funsetown(&kq->kq_sigio);
2250 }
2251 
2252 /*ARGSUSED*/
2253 static int
2254 kqueue_close(struct file *fp, struct thread *td)
2255 {
2256 	struct kqueue *kq = fp->f_data;
2257 	struct filedesc *fdp;
2258 	int error;
2259 	int filedesc_unlock;
2260 
2261 	if ((error = kqueue_acquire(fp, &kq)))
2262 		return error;
2263 	kqueue_drain(kq, td);
2264 
2265 	/*
2266 	 * We could be called due to the knote_drop() doing fdrop(),
2267 	 * called from kqueue_register().  In this case the global
2268 	 * lock is owned, and filedesc sx is locked before, to not
2269 	 * take the sleepable lock after non-sleepable.
2270 	 */
2271 	fdp = kq->kq_fdp;
2272 	kq->kq_fdp = NULL;
2273 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
2274 		FILEDESC_XLOCK(fdp);
2275 		filedesc_unlock = 1;
2276 	} else
2277 		filedesc_unlock = 0;
2278 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
2279 	if (filedesc_unlock)
2280 		FILEDESC_XUNLOCK(fdp);
2281 
2282 	kqueue_destroy(kq);
2283 	chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
2284 	crfree(kq->kq_cred);
2285 	free(kq, M_KQUEUE);
2286 	fp->f_data = NULL;
2287 
2288 	return (0);
2289 }
2290 
2291 static int
2292 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2293 {
2294 
2295 	kif->kf_type = KF_TYPE_KQUEUE;
2296 	return (0);
2297 }
2298 
2299 static void
2300 kqueue_wakeup(struct kqueue *kq)
2301 {
2302 	KQ_OWNED(kq);
2303 
2304 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
2305 		kq->kq_state &= ~KQ_SLEEP;
2306 		wakeup(kq);
2307 	}
2308 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2309 		selwakeuppri(&kq->kq_sel, PSOCK);
2310 		if (!SEL_WAITING(&kq->kq_sel))
2311 			kq->kq_state &= ~KQ_SEL;
2312 	}
2313 	if (!knlist_empty(&kq->kq_sel.si_note))
2314 		kqueue_schedtask(kq);
2315 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
2316 		pgsigio(&kq->kq_sigio, SIGIO, 0);
2317 	}
2318 }
2319 
2320 /*
2321  * Walk down a list of knotes, activating them if their event has triggered.
2322  *
2323  * There is a possibility to optimize in the case of one kq watching another.
2324  * Instead of scheduling a task to wake it up, you could pass enough state
2325  * down the chain to make up the parent kqueue.  Make this code functional
2326  * first.
2327  */
2328 void
2329 knote(struct knlist *list, long hint, int lockflags)
2330 {
2331 	struct kqueue *kq;
2332 	struct knote *kn, *tkn;
2333 	int error;
2334 
2335 	if (list == NULL)
2336 		return;
2337 
2338 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
2339 
2340 	if ((lockflags & KNF_LISTLOCKED) == 0)
2341 		list->kl_lock(list->kl_lockarg);
2342 
2343 	/*
2344 	 * If we unlock the list lock (and enter influx), we can
2345 	 * eliminate the kqueue scheduling, but this will introduce
2346 	 * four lock/unlock's for each knote to test.  Also, marker
2347 	 * would be needed to keep iteration position, since filters
2348 	 * or other threads could remove events.
2349 	 */
2350 	SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
2351 		kq = kn->kn_kq;
2352 		KQ_LOCK(kq);
2353 		if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
2354 			/*
2355 			 * Do not process the influx notes, except for
2356 			 * the influx coming from the kq unlock in the
2357 			 * kqueue_scan().  In the later case, we do
2358 			 * not interfere with the scan, since the code
2359 			 * fragment in kqueue_scan() locks the knlist,
2360 			 * and cannot proceed until we finished.
2361 			 */
2362 			KQ_UNLOCK(kq);
2363 		} else if ((lockflags & KNF_NOKQLOCK) != 0) {
2364 			kn_enter_flux(kn);
2365 			KQ_UNLOCK(kq);
2366 			error = kn->kn_fop->f_event(kn, hint);
2367 			KQ_LOCK(kq);
2368 			kn_leave_flux(kn);
2369 			if (error)
2370 				KNOTE_ACTIVATE(kn, 1);
2371 			KQ_UNLOCK_FLUX(kq);
2372 		} else {
2373 			if (kn->kn_fop->f_event(kn, hint))
2374 				KNOTE_ACTIVATE(kn, 1);
2375 			KQ_UNLOCK(kq);
2376 		}
2377 	}
2378 	if ((lockflags & KNF_LISTLOCKED) == 0)
2379 		list->kl_unlock(list->kl_lockarg);
2380 }
2381 
2382 /*
2383  * add a knote to a knlist
2384  */
2385 void
2386 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2387 {
2388 
2389 	KNL_ASSERT_LOCK(knl, islocked);
2390 	KQ_NOTOWNED(kn->kn_kq);
2391 	KASSERT(kn_in_flux(kn), ("knote %p not in flux", kn));
2392 	KASSERT((kn->kn_status & KN_DETACHED) != 0,
2393 	    ("knote %p was not detached", kn));
2394 	if (!islocked)
2395 		knl->kl_lock(knl->kl_lockarg);
2396 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2397 	if (!islocked)
2398 		knl->kl_unlock(knl->kl_lockarg);
2399 	KQ_LOCK(kn->kn_kq);
2400 	kn->kn_knlist = knl;
2401 	kn->kn_status &= ~KN_DETACHED;
2402 	KQ_UNLOCK(kn->kn_kq);
2403 }
2404 
2405 static void
2406 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked,
2407     int kqislocked)
2408 {
2409 
2410 	KASSERT(!kqislocked || knlislocked, ("kq locked w/o knl locked"));
2411 	KNL_ASSERT_LOCK(knl, knlislocked);
2412 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2413 	KASSERT(kqislocked || kn_in_flux(kn), ("knote %p not in flux", kn));
2414 	KASSERT((kn->kn_status & KN_DETACHED) == 0,
2415 	    ("knote %p was already detached", kn));
2416 	if (!knlislocked)
2417 		knl->kl_lock(knl->kl_lockarg);
2418 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2419 	kn->kn_knlist = NULL;
2420 	if (!knlislocked)
2421 		kn_list_unlock(knl);
2422 	if (!kqislocked)
2423 		KQ_LOCK(kn->kn_kq);
2424 	kn->kn_status |= KN_DETACHED;
2425 	if (!kqislocked)
2426 		KQ_UNLOCK(kn->kn_kq);
2427 }
2428 
2429 /*
2430  * remove knote from the specified knlist
2431  */
2432 void
2433 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2434 {
2435 
2436 	knlist_remove_kq(knl, kn, islocked, 0);
2437 }
2438 
2439 int
2440 knlist_empty(struct knlist *knl)
2441 {
2442 
2443 	KNL_ASSERT_LOCKED(knl);
2444 	return (SLIST_EMPTY(&knl->kl_list));
2445 }
2446 
2447 static struct mtx knlist_lock;
2448 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2449     MTX_DEF);
2450 static void knlist_mtx_lock(void *arg);
2451 static void knlist_mtx_unlock(void *arg);
2452 
2453 static void
2454 knlist_mtx_lock(void *arg)
2455 {
2456 
2457 	mtx_lock((struct mtx *)arg);
2458 }
2459 
2460 static void
2461 knlist_mtx_unlock(void *arg)
2462 {
2463 
2464 	mtx_unlock((struct mtx *)arg);
2465 }
2466 
2467 static void
2468 knlist_mtx_assert_lock(void *arg, int what)
2469 {
2470 
2471 	if (what == LA_LOCKED)
2472 		mtx_assert((struct mtx *)arg, MA_OWNED);
2473 	else
2474 		mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2475 }
2476 
2477 static void
2478 knlist_rw_rlock(void *arg)
2479 {
2480 
2481 	rw_rlock((struct rwlock *)arg);
2482 }
2483 
2484 static void
2485 knlist_rw_runlock(void *arg)
2486 {
2487 
2488 	rw_runlock((struct rwlock *)arg);
2489 }
2490 
2491 static void
2492 knlist_rw_assert_lock(void *arg, int what)
2493 {
2494 
2495 	if (what == LA_LOCKED)
2496 		rw_assert((struct rwlock *)arg, RA_LOCKED);
2497 	else
2498 		rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2499 }
2500 
2501 void
2502 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2503     void (*kl_unlock)(void *),
2504     void (*kl_assert_lock)(void *, int))
2505 {
2506 
2507 	if (lock == NULL)
2508 		knl->kl_lockarg = &knlist_lock;
2509 	else
2510 		knl->kl_lockarg = lock;
2511 
2512 	if (kl_lock == NULL)
2513 		knl->kl_lock = knlist_mtx_lock;
2514 	else
2515 		knl->kl_lock = kl_lock;
2516 	if (kl_unlock == NULL)
2517 		knl->kl_unlock = knlist_mtx_unlock;
2518 	else
2519 		knl->kl_unlock = kl_unlock;
2520 	if (kl_assert_lock == NULL)
2521 		knl->kl_assert_lock = knlist_mtx_assert_lock;
2522 	else
2523 		knl->kl_assert_lock = kl_assert_lock;
2524 
2525 	knl->kl_autodestroy = 0;
2526 	SLIST_INIT(&knl->kl_list);
2527 }
2528 
2529 void
2530 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2531 {
2532 
2533 	knlist_init(knl, lock, NULL, NULL, NULL);
2534 }
2535 
2536 struct knlist *
2537 knlist_alloc(struct mtx *lock)
2538 {
2539 	struct knlist *knl;
2540 
2541 	knl = malloc(sizeof(struct knlist), M_KQUEUE, M_WAITOK);
2542 	knlist_init_mtx(knl, lock);
2543 	return (knl);
2544 }
2545 
2546 void
2547 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2548 {
2549 
2550 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2551 	    knlist_rw_assert_lock);
2552 }
2553 
2554 void
2555 knlist_destroy(struct knlist *knl)
2556 {
2557 
2558 	KASSERT(KNLIST_EMPTY(knl),
2559 	    ("destroying knlist %p with knotes on it", knl));
2560 }
2561 
2562 void
2563 knlist_detach(struct knlist *knl)
2564 {
2565 
2566 	KNL_ASSERT_LOCKED(knl);
2567 	knl->kl_autodestroy = 1;
2568 	if (knlist_empty(knl)) {
2569 		knlist_destroy(knl);
2570 		free(knl, M_KQUEUE);
2571 	}
2572 }
2573 
2574 /*
2575  * Even if we are locked, we may need to drop the lock to allow any influx
2576  * knotes time to "settle".
2577  */
2578 void
2579 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2580 {
2581 	struct knote *kn, *kn2;
2582 	struct kqueue *kq;
2583 
2584 	KASSERT(!knl->kl_autodestroy, ("cleardel for autodestroy %p", knl));
2585 	if (islocked)
2586 		KNL_ASSERT_LOCKED(knl);
2587 	else {
2588 		KNL_ASSERT_UNLOCKED(knl);
2589 again:		/* need to reacquire lock since we have dropped it */
2590 		knl->kl_lock(knl->kl_lockarg);
2591 	}
2592 
2593 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2594 		kq = kn->kn_kq;
2595 		KQ_LOCK(kq);
2596 		if (kn_in_flux(kn)) {
2597 			KQ_UNLOCK(kq);
2598 			continue;
2599 		}
2600 		knlist_remove_kq(knl, kn, 1, 1);
2601 		if (killkn) {
2602 			kn_enter_flux(kn);
2603 			KQ_UNLOCK(kq);
2604 			knote_drop_detached(kn, td);
2605 		} else {
2606 			/* Make sure cleared knotes disappear soon */
2607 			kn->kn_flags |= EV_EOF | EV_ONESHOT;
2608 			KQ_UNLOCK(kq);
2609 		}
2610 		kq = NULL;
2611 	}
2612 
2613 	if (!SLIST_EMPTY(&knl->kl_list)) {
2614 		/* there are still in flux knotes remaining */
2615 		kn = SLIST_FIRST(&knl->kl_list);
2616 		kq = kn->kn_kq;
2617 		KQ_LOCK(kq);
2618 		KASSERT(kn_in_flux(kn), ("knote removed w/o list lock"));
2619 		knl->kl_unlock(knl->kl_lockarg);
2620 		kq->kq_state |= KQ_FLUXWAIT;
2621 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2622 		kq = NULL;
2623 		goto again;
2624 	}
2625 
2626 	if (islocked)
2627 		KNL_ASSERT_LOCKED(knl);
2628 	else {
2629 		knl->kl_unlock(knl->kl_lockarg);
2630 		KNL_ASSERT_UNLOCKED(knl);
2631 	}
2632 }
2633 
2634 /*
2635  * Remove all knotes referencing a specified fd must be called with FILEDESC
2636  * lock.  This prevents a race where a new fd comes along and occupies the
2637  * entry and we attach a knote to the fd.
2638  */
2639 void
2640 knote_fdclose(struct thread *td, int fd)
2641 {
2642 	struct filedesc *fdp = td->td_proc->p_fd;
2643 	struct kqueue *kq;
2644 	struct knote *kn;
2645 	int influx;
2646 
2647 	FILEDESC_XLOCK_ASSERT(fdp);
2648 
2649 	/*
2650 	 * We shouldn't have to worry about new kevents appearing on fd
2651 	 * since filedesc is locked.
2652 	 */
2653 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2654 		KQ_LOCK(kq);
2655 
2656 again:
2657 		influx = 0;
2658 		while (kq->kq_knlistsize > fd &&
2659 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2660 			if (kn_in_flux(kn)) {
2661 				/* someone else might be waiting on our knote */
2662 				if (influx)
2663 					wakeup(kq);
2664 				kq->kq_state |= KQ_FLUXWAIT;
2665 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2666 				goto again;
2667 			}
2668 			kn_enter_flux(kn);
2669 			KQ_UNLOCK(kq);
2670 			influx = 1;
2671 			knote_drop(kn, td);
2672 			KQ_LOCK(kq);
2673 		}
2674 		KQ_UNLOCK_FLUX(kq);
2675 	}
2676 }
2677 
2678 static int
2679 knote_attach(struct knote *kn, struct kqueue *kq)
2680 {
2681 	struct klist *list;
2682 
2683 	KASSERT(kn_in_flux(kn), ("knote %p not marked influx", kn));
2684 	KQ_OWNED(kq);
2685 
2686 	if ((kq->kq_state & KQ_CLOSING) != 0)
2687 		return (EBADF);
2688 	if (kn->kn_fop->f_isfd) {
2689 		if (kn->kn_id >= kq->kq_knlistsize)
2690 			return (ENOMEM);
2691 		list = &kq->kq_knlist[kn->kn_id];
2692 	} else {
2693 		if (kq->kq_knhash == NULL)
2694 			return (ENOMEM);
2695 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2696 	}
2697 	SLIST_INSERT_HEAD(list, kn, kn_link);
2698 	return (0);
2699 }
2700 
2701 static void
2702 knote_drop(struct knote *kn, struct thread *td)
2703 {
2704 
2705 	if ((kn->kn_status & KN_DETACHED) == 0)
2706 		kn->kn_fop->f_detach(kn);
2707 	knote_drop_detached(kn, td);
2708 }
2709 
2710 static void
2711 knote_drop_detached(struct knote *kn, struct thread *td)
2712 {
2713 	struct kqueue *kq;
2714 	struct klist *list;
2715 
2716 	kq = kn->kn_kq;
2717 
2718 	KASSERT((kn->kn_status & KN_DETACHED) != 0,
2719 	    ("knote %p still attached", kn));
2720 	KQ_NOTOWNED(kq);
2721 
2722 	KQ_LOCK(kq);
2723 	KASSERT(kn->kn_influx == 1,
2724 	    ("knote_drop called on %p with influx %d", kn, kn->kn_influx));
2725 
2726 	if (kn->kn_fop->f_isfd)
2727 		list = &kq->kq_knlist[kn->kn_id];
2728 	else
2729 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2730 
2731 	if (!SLIST_EMPTY(list))
2732 		SLIST_REMOVE(list, kn, knote, kn_link);
2733 	if (kn->kn_status & KN_QUEUED)
2734 		knote_dequeue(kn);
2735 	KQ_UNLOCK_FLUX(kq);
2736 
2737 	if (kn->kn_fop->f_isfd) {
2738 		fdrop(kn->kn_fp, td);
2739 		kn->kn_fp = NULL;
2740 	}
2741 	kqueue_fo_release(kn->kn_kevent.filter);
2742 	kn->kn_fop = NULL;
2743 	knote_free(kn);
2744 }
2745 
2746 static void
2747 knote_enqueue(struct knote *kn)
2748 {
2749 	struct kqueue *kq = kn->kn_kq;
2750 
2751 	KQ_OWNED(kn->kn_kq);
2752 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2753 
2754 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2755 	kn->kn_status |= KN_QUEUED;
2756 	kq->kq_count++;
2757 	kqueue_wakeup(kq);
2758 }
2759 
2760 static void
2761 knote_dequeue(struct knote *kn)
2762 {
2763 	struct kqueue *kq = kn->kn_kq;
2764 
2765 	KQ_OWNED(kn->kn_kq);
2766 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2767 
2768 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2769 	kn->kn_status &= ~KN_QUEUED;
2770 	kq->kq_count--;
2771 }
2772 
2773 static void
2774 knote_init(void)
2775 {
2776 
2777 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2778 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2779 }
2780 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2781 
2782 static struct knote *
2783 knote_alloc(int mflag)
2784 {
2785 
2786 	return (uma_zalloc(knote_zone, mflag | M_ZERO));
2787 }
2788 
2789 static void
2790 knote_free(struct knote *kn)
2791 {
2792 
2793 	uma_zfree(knote_zone, kn);
2794 }
2795 
2796 /*
2797  * Register the kev w/ the kq specified by fd.
2798  */
2799 int
2800 kqfd_register(int fd, struct kevent *kev, struct thread *td, int mflag)
2801 {
2802 	struct kqueue *kq;
2803 	struct file *fp;
2804 	cap_rights_t rights;
2805 	int error;
2806 
2807 	error = fget(td, fd, cap_rights_init_one(&rights, CAP_KQUEUE_CHANGE),
2808 	    &fp);
2809 	if (error != 0)
2810 		return (error);
2811 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2812 		goto noacquire;
2813 
2814 	error = kqueue_register(kq, kev, td, mflag);
2815 	kqueue_release(kq, 0);
2816 
2817 noacquire:
2818 	fdrop(fp, td);
2819 	return (error);
2820 }
2821