xref: /freebsd/sys/kern/kern_event.c (revision 13ea0450a9c8742119d36f3bf8f47accdce46e54)
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 mflag);
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 mflag);
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 mflag);
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 	MPASS(list != NULL);
537 	KNL_ASSERT_LOCKED(list);
538 	if (SLIST_EMPTY(&list->kl_list))
539 		return;
540 
541 	memset(&kev, 0, sizeof(kev));
542 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
543 		kq = kn->kn_kq;
544 		KQ_LOCK(kq);
545 		if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
546 			KQ_UNLOCK(kq);
547 			continue;
548 		}
549 
550 		/*
551 		 * The same as knote(), activate the event.
552 		 */
553 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
554 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
555 				KNOTE_ACTIVATE(kn, 1);
556 			KQ_UNLOCK(kq);
557 			continue;
558 		}
559 
560 		/*
561 		 * The NOTE_TRACK case. In addition to the activation
562 		 * of the event, we need to register new events to
563 		 * track the child. Drop the locks in preparation for
564 		 * the call to kqueue_register().
565 		 */
566 		kn_enter_flux(kn);
567 		KQ_UNLOCK(kq);
568 		list->kl_unlock(list->kl_lockarg);
569 
570 		/*
571 		 * Activate existing knote and register tracking knotes with
572 		 * new process.
573 		 *
574 		 * First register a knote to get just the child notice. This
575 		 * must be a separate note from a potential NOTE_EXIT
576 		 * notification since both NOTE_CHILD and NOTE_EXIT are defined
577 		 * to use the data field (in conflicting ways).
578 		 */
579 		kev.ident = pid;
580 		kev.filter = kn->kn_filter;
581 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_ONESHOT |
582 		    EV_FLAG2;
583 		kev.fflags = kn->kn_sfflags;
584 		kev.data = kn->kn_id;		/* parent */
585 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
586 		error = kqueue_register(kq, &kev, NULL, M_NOWAIT);
587 		if (error)
588 			kn->kn_fflags |= NOTE_TRACKERR;
589 
590 		/*
591 		 * Then register another knote to track other potential events
592 		 * from the new process.
593 		 */
594 		kev.ident = pid;
595 		kev.filter = kn->kn_filter;
596 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
597 		kev.fflags = kn->kn_sfflags;
598 		kev.data = kn->kn_id;		/* parent */
599 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
600 		error = kqueue_register(kq, &kev, NULL, M_NOWAIT);
601 		if (error)
602 			kn->kn_fflags |= NOTE_TRACKERR;
603 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
604 			KNOTE_ACTIVATE(kn, 0);
605 		list->kl_lock(list->kl_lockarg);
606 		KQ_LOCK(kq);
607 		kn_leave_flux(kn);
608 		KQ_UNLOCK_FLUX(kq);
609 	}
610 }
611 
612 /*
613  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
614  * interval timer support code.
615  */
616 
617 #define NOTE_TIMER_PRECMASK						\
618     (NOTE_SECONDS | NOTE_MSECONDS | NOTE_USECONDS | NOTE_NSECONDS)
619 
620 static sbintime_t
621 timer2sbintime(intptr_t data, int flags)
622 {
623 	int64_t secs;
624 
625         /*
626          * Macros for converting to the fractional second portion of an
627          * sbintime_t using 64bit multiplication to improve precision.
628          */
629 #define NS_TO_SBT(ns) (((ns) * (((uint64_t)1 << 63) / 500000000)) >> 32)
630 #define US_TO_SBT(us) (((us) * (((uint64_t)1 << 63) / 500000)) >> 32)
631 #define MS_TO_SBT(ms) (((ms) * (((uint64_t)1 << 63) / 500)) >> 32)
632 	switch (flags & NOTE_TIMER_PRECMASK) {
633 	case NOTE_SECONDS:
634 #ifdef __LP64__
635 		if (data > (SBT_MAX / SBT_1S))
636 			return (SBT_MAX);
637 #endif
638 		return ((sbintime_t)data << 32);
639 	case NOTE_MSECONDS: /* FALLTHROUGH */
640 	case 0:
641 		if (data >= 1000) {
642 			secs = data / 1000;
643 #ifdef __LP64__
644 			if (secs > (SBT_MAX / SBT_1S))
645 				return (SBT_MAX);
646 #endif
647 			return (secs << 32 | MS_TO_SBT(data % 1000));
648 		}
649 		return (MS_TO_SBT(data));
650 	case NOTE_USECONDS:
651 		if (data >= 1000000) {
652 			secs = data / 1000000;
653 #ifdef __LP64__
654 			if (secs > (SBT_MAX / SBT_1S))
655 				return (SBT_MAX);
656 #endif
657 			return (secs << 32 | US_TO_SBT(data % 1000000));
658 		}
659 		return (US_TO_SBT(data));
660 	case NOTE_NSECONDS:
661 		if (data >= 1000000000) {
662 			secs = data / 1000000000;
663 #ifdef __LP64__
664 			if (secs > (SBT_MAX / SBT_1S))
665 				return (SBT_MAX);
666 #endif
667 			return (secs << 32 | US_TO_SBT(data % 1000000000));
668 		}
669 		return (NS_TO_SBT(data));
670 	default:
671 		break;
672 	}
673 	return (-1);
674 }
675 
676 struct kq_timer_cb_data {
677 	struct callout c;
678 	sbintime_t next;	/* next timer event fires at */
679 	sbintime_t to;		/* precalculated timer period, 0 for abs */
680 };
681 
682 static void
683 filt_timerexpire(void *knx)
684 {
685 	struct knote *kn;
686 	struct kq_timer_cb_data *kc;
687 
688 	kn = knx;
689 	kn->kn_data++;
690 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
691 
692 	if ((kn->kn_flags & EV_ONESHOT) != 0)
693 		return;
694 	kc = kn->kn_ptr.p_v;
695 	if (kc->to == 0)
696 		return;
697 	kc->next += kc->to;
698 	callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
699 	    PCPU_GET(cpuid), C_ABSOLUTE);
700 }
701 
702 /*
703  * data contains amount of time to sleep
704  */
705 static int
706 filt_timervalidate(struct knote *kn, sbintime_t *to)
707 {
708 	struct bintime bt;
709 	sbintime_t sbt;
710 
711 	if (kn->kn_sdata < 0)
712 		return (EINVAL);
713 	if (kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
714 		kn->kn_sdata = 1;
715 	/*
716 	 * The only fflags values supported are the timer unit
717 	 * (precision) and the absolute time indicator.
718 	 */
719 	if ((kn->kn_sfflags & ~(NOTE_TIMER_PRECMASK | NOTE_ABSTIME)) != 0)
720 		return (EINVAL);
721 
722 	*to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
723 	if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
724 		getboottimebin(&bt);
725 		sbt = bttosbt(bt);
726 		*to -= sbt;
727 	}
728 	if (*to < 0)
729 		return (EINVAL);
730 	return (0);
731 }
732 
733 static int
734 filt_timerattach(struct knote *kn)
735 {
736 	struct kq_timer_cb_data *kc;
737 	sbintime_t to;
738 	unsigned int ncallouts;
739 	int error;
740 
741 	error = filt_timervalidate(kn, &to);
742 	if (error != 0)
743 		return (error);
744 
745 	do {
746 		ncallouts = kq_ncallouts;
747 		if (ncallouts >= kq_calloutmax)
748 			return (ENOMEM);
749 	} while (!atomic_cmpset_int(&kq_ncallouts, ncallouts, ncallouts + 1));
750 
751 	if ((kn->kn_sfflags & NOTE_ABSTIME) == 0)
752 		kn->kn_flags |= EV_CLEAR;	/* automatically set */
753 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
754 	kn->kn_ptr.p_v = kc = malloc(sizeof(*kc), M_KQUEUE, M_WAITOK);
755 	callout_init(&kc->c, 1);
756 	filt_timerstart(kn, to);
757 
758 	return (0);
759 }
760 
761 static void
762 filt_timerstart(struct knote *kn, sbintime_t to)
763 {
764 	struct kq_timer_cb_data *kc;
765 
766 	kc = kn->kn_ptr.p_v;
767 	if ((kn->kn_sfflags & NOTE_ABSTIME) != 0) {
768 		kc->next = to;
769 		kc->to = 0;
770 	} else {
771 		kc->next = to + sbinuptime();
772 		kc->to = to;
773 	}
774 	callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
775 	    PCPU_GET(cpuid), C_ABSOLUTE);
776 }
777 
778 static void
779 filt_timerdetach(struct knote *kn)
780 {
781 	struct kq_timer_cb_data *kc;
782 	unsigned int old __unused;
783 
784 	kc = kn->kn_ptr.p_v;
785 	callout_drain(&kc->c);
786 	free(kc, M_KQUEUE);
787 	old = atomic_fetchadd_int(&kq_ncallouts, -1);
788 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
789 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
790 }
791 
792 static void
793 filt_timertouch(struct knote *kn, struct kevent *kev, u_long type)
794 {
795 	struct kq_timer_cb_data *kc;
796 	struct kqueue *kq;
797 	sbintime_t to;
798 	int error;
799 
800 	switch (type) {
801 	case EVENT_REGISTER:
802 		/* Handle re-added timers that update data/fflags */
803 		if (kev->flags & EV_ADD) {
804 			kc = kn->kn_ptr.p_v;
805 
806 			/* Drain any existing callout. */
807 			callout_drain(&kc->c);
808 
809 			/* Throw away any existing undelivered record
810 			 * of the timer expiration. This is done under
811 			 * the presumption that if a process is
812 			 * re-adding this timer with new parameters,
813 			 * it is no longer interested in what may have
814 			 * happened under the old parameters. If it is
815 			 * interested, it can wait for the expiration,
816 			 * delete the old timer definition, and then
817 			 * add the new one.
818 			 *
819 			 * This has to be done while the kq is locked:
820 			 *   - if enqueued, dequeue
821 			 *   - make it no longer active
822 			 *   - clear the count of expiration events
823 			 */
824 			kq = kn->kn_kq;
825 			KQ_LOCK(kq);
826 			if (kn->kn_status & KN_QUEUED)
827 				knote_dequeue(kn);
828 
829 			kn->kn_status &= ~KN_ACTIVE;
830 			kn->kn_data = 0;
831 			KQ_UNLOCK(kq);
832 
833 			/* Reschedule timer based on new data/fflags */
834 			kn->kn_sfflags = kev->fflags;
835 			kn->kn_sdata = kev->data;
836 			error = filt_timervalidate(kn, &to);
837 			if (error != 0) {
838 			  	kn->kn_flags |= EV_ERROR;
839 				kn->kn_data = error;
840 			} else
841 			  	filt_timerstart(kn, to);
842 		}
843 		break;
844 
845         case EVENT_PROCESS:
846 		*kev = kn->kn_kevent;
847 		if (kn->kn_flags & EV_CLEAR) {
848 			kn->kn_data = 0;
849 			kn->kn_fflags = 0;
850 		}
851 		break;
852 
853 	default:
854 		panic("filt_timertouch() - invalid type (%ld)", type);
855 		break;
856 	}
857 }
858 
859 static int
860 filt_timer(struct knote *kn, long hint)
861 {
862 
863 	return (kn->kn_data != 0);
864 }
865 
866 static int
867 filt_userattach(struct knote *kn)
868 {
869 
870 	/*
871 	 * EVFILT_USER knotes are not attached to anything in the kernel.
872 	 */
873 	kn->kn_hook = NULL;
874 	if (kn->kn_fflags & NOTE_TRIGGER)
875 		kn->kn_hookid = 1;
876 	else
877 		kn->kn_hookid = 0;
878 	return (0);
879 }
880 
881 static void
882 filt_userdetach(__unused struct knote *kn)
883 {
884 
885 	/*
886 	 * EVFILT_USER knotes are not attached to anything in the kernel.
887 	 */
888 }
889 
890 static int
891 filt_user(struct knote *kn, __unused long hint)
892 {
893 
894 	return (kn->kn_hookid);
895 }
896 
897 static void
898 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
899 {
900 	u_int ffctrl;
901 
902 	switch (type) {
903 	case EVENT_REGISTER:
904 		if (kev->fflags & NOTE_TRIGGER)
905 			kn->kn_hookid = 1;
906 
907 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
908 		kev->fflags &= NOTE_FFLAGSMASK;
909 		switch (ffctrl) {
910 		case NOTE_FFNOP:
911 			break;
912 
913 		case NOTE_FFAND:
914 			kn->kn_sfflags &= kev->fflags;
915 			break;
916 
917 		case NOTE_FFOR:
918 			kn->kn_sfflags |= kev->fflags;
919 			break;
920 
921 		case NOTE_FFCOPY:
922 			kn->kn_sfflags = kev->fflags;
923 			break;
924 
925 		default:
926 			/* XXX Return error? */
927 			break;
928 		}
929 		kn->kn_sdata = kev->data;
930 		if (kev->flags & EV_CLEAR) {
931 			kn->kn_hookid = 0;
932 			kn->kn_data = 0;
933 			kn->kn_fflags = 0;
934 		}
935 		break;
936 
937         case EVENT_PROCESS:
938 		*kev = kn->kn_kevent;
939 		kev->fflags = kn->kn_sfflags;
940 		kev->data = kn->kn_sdata;
941 		if (kn->kn_flags & EV_CLEAR) {
942 			kn->kn_hookid = 0;
943 			kn->kn_data = 0;
944 			kn->kn_fflags = 0;
945 		}
946 		break;
947 
948 	default:
949 		panic("filt_usertouch() - invalid type (%ld)", type);
950 		break;
951 	}
952 }
953 
954 int
955 sys_kqueue(struct thread *td, struct kqueue_args *uap)
956 {
957 
958 	return (kern_kqueue(td, 0, NULL));
959 }
960 
961 static void
962 kqueue_init(struct kqueue *kq)
963 {
964 
965 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
966 	TAILQ_INIT(&kq->kq_head);
967 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
968 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
969 }
970 
971 int
972 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
973 {
974 	struct filedesc *fdp;
975 	struct kqueue *kq;
976 	struct file *fp;
977 	struct ucred *cred;
978 	int fd, error;
979 
980 	fdp = td->td_proc->p_fd;
981 	cred = td->td_ucred;
982 	if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
983 		return (ENOMEM);
984 
985 	error = falloc_caps(td, &fp, &fd, flags, fcaps);
986 	if (error != 0) {
987 		chgkqcnt(cred->cr_ruidinfo, -1, 0);
988 		return (error);
989 	}
990 
991 	/* An extra reference on `fp' has been held for us by falloc(). */
992 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
993 	kqueue_init(kq);
994 	kq->kq_fdp = fdp;
995 	kq->kq_cred = crhold(cred);
996 
997 	FILEDESC_XLOCK(fdp);
998 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
999 	FILEDESC_XUNLOCK(fdp);
1000 
1001 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
1002 	fdrop(fp, td);
1003 
1004 	td->td_retval[0] = fd;
1005 	return (0);
1006 }
1007 
1008 struct g_kevent_args {
1009 	int	fd;
1010 	void	*changelist;
1011 	int	nchanges;
1012 	void	*eventlist;
1013 	int	nevents;
1014 	const struct timespec *timeout;
1015 };
1016 
1017 int
1018 sys_kevent(struct thread *td, struct kevent_args *uap)
1019 {
1020 	struct kevent_copyops k_ops = {
1021 		.arg = uap,
1022 		.k_copyout = kevent_copyout,
1023 		.k_copyin = kevent_copyin,
1024 		.kevent_size = sizeof(struct kevent),
1025 	};
1026 	struct g_kevent_args gk_args = {
1027 		.fd = uap->fd,
1028 		.changelist = uap->changelist,
1029 		.nchanges = uap->nchanges,
1030 		.eventlist = uap->eventlist,
1031 		.nevents = uap->nevents,
1032 		.timeout = uap->timeout,
1033 	};
1034 
1035 	return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent"));
1036 }
1037 
1038 static int
1039 kern_kevent_generic(struct thread *td, struct g_kevent_args *uap,
1040     struct kevent_copyops *k_ops, const char *struct_name)
1041 {
1042 	struct timespec ts, *tsp;
1043 #ifdef KTRACE
1044 	struct kevent *eventlist = uap->eventlist;
1045 #endif
1046 	int error;
1047 
1048 	if (uap->timeout != NULL) {
1049 		error = copyin(uap->timeout, &ts, sizeof(ts));
1050 		if (error)
1051 			return (error);
1052 		tsp = &ts;
1053 	} else
1054 		tsp = NULL;
1055 
1056 #ifdef KTRACE
1057 	if (KTRPOINT(td, KTR_STRUCT_ARRAY))
1058 		ktrstructarray(struct_name, UIO_USERSPACE, uap->changelist,
1059 		    uap->nchanges, k_ops->kevent_size);
1060 #endif
1061 
1062 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
1063 	    k_ops, tsp);
1064 
1065 #ifdef KTRACE
1066 	if (error == 0 && KTRPOINT(td, KTR_STRUCT_ARRAY))
1067 		ktrstructarray(struct_name, UIO_USERSPACE, eventlist,
1068 		    td->td_retval[0], k_ops->kevent_size);
1069 #endif
1070 
1071 	return (error);
1072 }
1073 
1074 /*
1075  * Copy 'count' items into the destination list pointed to by uap->eventlist.
1076  */
1077 static int
1078 kevent_copyout(void *arg, struct kevent *kevp, int count)
1079 {
1080 	struct kevent_args *uap;
1081 	int error;
1082 
1083 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1084 	uap = (struct kevent_args *)arg;
1085 
1086 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
1087 	if (error == 0)
1088 		uap->eventlist += count;
1089 	return (error);
1090 }
1091 
1092 /*
1093  * Copy 'count' items from the list pointed to by uap->changelist.
1094  */
1095 static int
1096 kevent_copyin(void *arg, struct kevent *kevp, int count)
1097 {
1098 	struct kevent_args *uap;
1099 	int error;
1100 
1101 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1102 	uap = (struct kevent_args *)arg;
1103 
1104 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
1105 	if (error == 0)
1106 		uap->changelist += count;
1107 	return (error);
1108 }
1109 
1110 #ifdef COMPAT_FREEBSD11
1111 static int
1112 kevent11_copyout(void *arg, struct kevent *kevp, int count)
1113 {
1114 	struct freebsd11_kevent_args *uap;
1115 	struct kevent_freebsd11 kev11;
1116 	int error, i;
1117 
1118 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1119 	uap = (struct freebsd11_kevent_args *)arg;
1120 
1121 	for (i = 0; i < count; i++) {
1122 		kev11.ident = kevp->ident;
1123 		kev11.filter = kevp->filter;
1124 		kev11.flags = kevp->flags;
1125 		kev11.fflags = kevp->fflags;
1126 		kev11.data = kevp->data;
1127 		kev11.udata = kevp->udata;
1128 		error = copyout(&kev11, uap->eventlist, sizeof(kev11));
1129 		if (error != 0)
1130 			break;
1131 		uap->eventlist++;
1132 		kevp++;
1133 	}
1134 	return (error);
1135 }
1136 
1137 /*
1138  * Copy 'count' items from the list pointed to by uap->changelist.
1139  */
1140 static int
1141 kevent11_copyin(void *arg, struct kevent *kevp, int count)
1142 {
1143 	struct freebsd11_kevent_args *uap;
1144 	struct kevent_freebsd11 kev11;
1145 	int error, i;
1146 
1147 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
1148 	uap = (struct freebsd11_kevent_args *)arg;
1149 
1150 	for (i = 0; i < count; i++) {
1151 		error = copyin(uap->changelist, &kev11, sizeof(kev11));
1152 		if (error != 0)
1153 			break;
1154 		kevp->ident = kev11.ident;
1155 		kevp->filter = kev11.filter;
1156 		kevp->flags = kev11.flags;
1157 		kevp->fflags = kev11.fflags;
1158 		kevp->data = (uintptr_t)kev11.data;
1159 		kevp->udata = kev11.udata;
1160 		bzero(&kevp->ext, sizeof(kevp->ext));
1161 		uap->changelist++;
1162 		kevp++;
1163 	}
1164 	return (error);
1165 }
1166 
1167 int
1168 freebsd11_kevent(struct thread *td, struct freebsd11_kevent_args *uap)
1169 {
1170 	struct kevent_copyops k_ops = {
1171 		.arg = uap,
1172 		.k_copyout = kevent11_copyout,
1173 		.k_copyin = kevent11_copyin,
1174 		.kevent_size = sizeof(struct kevent_freebsd11),
1175 	};
1176 	struct g_kevent_args gk_args = {
1177 		.fd = uap->fd,
1178 		.changelist = uap->changelist,
1179 		.nchanges = uap->nchanges,
1180 		.eventlist = uap->eventlist,
1181 		.nevents = uap->nevents,
1182 		.timeout = uap->timeout,
1183 	};
1184 
1185 	return (kern_kevent_generic(td, &gk_args, &k_ops, "kevent_freebsd11"));
1186 }
1187 #endif
1188 
1189 int
1190 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
1191     struct kevent_copyops *k_ops, const struct timespec *timeout)
1192 {
1193 	cap_rights_t rights;
1194 	struct file *fp;
1195 	int error;
1196 
1197 	cap_rights_init(&rights);
1198 	if (nchanges > 0)
1199 		cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
1200 	if (nevents > 0)
1201 		cap_rights_set(&rights, CAP_KQUEUE_EVENT);
1202 	error = fget(td, fd, &rights, &fp);
1203 	if (error != 0)
1204 		return (error);
1205 
1206 	error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
1207 	fdrop(fp, td);
1208 
1209 	return (error);
1210 }
1211 
1212 static int
1213 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
1214     struct kevent_copyops *k_ops, const struct timespec *timeout)
1215 {
1216 	struct kevent keva[KQ_NEVENTS];
1217 	struct kevent *kevp, *changes;
1218 	int i, n, nerrors, error;
1219 
1220 	nerrors = 0;
1221 	while (nchanges > 0) {
1222 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
1223 		error = k_ops->k_copyin(k_ops->arg, keva, n);
1224 		if (error)
1225 			return (error);
1226 		changes = keva;
1227 		for (i = 0; i < n; i++) {
1228 			kevp = &changes[i];
1229 			if (!kevp->filter)
1230 				continue;
1231 			kevp->flags &= ~EV_SYSFLAGS;
1232 			error = kqueue_register(kq, kevp, td, M_WAITOK);
1233 			if (error || (kevp->flags & EV_RECEIPT)) {
1234 				if (nevents == 0)
1235 					return (error);
1236 				kevp->flags = EV_ERROR;
1237 				kevp->data = error;
1238 				(void)k_ops->k_copyout(k_ops->arg, kevp, 1);
1239 				nevents--;
1240 				nerrors++;
1241 			}
1242 		}
1243 		nchanges -= n;
1244 	}
1245 	if (nerrors) {
1246 		td->td_retval[0] = nerrors;
1247 		return (0);
1248 	}
1249 
1250 	return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
1251 }
1252 
1253 int
1254 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
1255     struct kevent_copyops *k_ops, const struct timespec *timeout)
1256 {
1257 	struct kqueue *kq;
1258 	int error;
1259 
1260 	error = kqueue_acquire(fp, &kq);
1261 	if (error != 0)
1262 		return (error);
1263 	error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
1264 	kqueue_release(kq, 0);
1265 	return (error);
1266 }
1267 
1268 /*
1269  * Performs a kevent() call on a temporarily created kqueue. This can be
1270  * used to perform one-shot polling, similar to poll() and select().
1271  */
1272 int
1273 kern_kevent_anonymous(struct thread *td, int nevents,
1274     struct kevent_copyops *k_ops)
1275 {
1276 	struct kqueue kq = {};
1277 	int error;
1278 
1279 	kqueue_init(&kq);
1280 	kq.kq_refcnt = 1;
1281 	error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
1282 	kqueue_drain(&kq, td);
1283 	kqueue_destroy(&kq);
1284 	return (error);
1285 }
1286 
1287 int
1288 kqueue_add_filteropts(int filt, struct filterops *filtops)
1289 {
1290 	int error;
1291 
1292 	error = 0;
1293 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1294 		printf(
1295 "trying to add a filterop that is out of range: %d is beyond %d\n",
1296 		    ~filt, EVFILT_SYSCOUNT);
1297 		return EINVAL;
1298 	}
1299 	mtx_lock(&filterops_lock);
1300 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1301 	    sysfilt_ops[~filt].for_fop != NULL)
1302 		error = EEXIST;
1303 	else {
1304 		sysfilt_ops[~filt].for_fop = filtops;
1305 		sysfilt_ops[~filt].for_refcnt = 0;
1306 	}
1307 	mtx_unlock(&filterops_lock);
1308 
1309 	return (error);
1310 }
1311 
1312 int
1313 kqueue_del_filteropts(int filt)
1314 {
1315 	int error;
1316 
1317 	error = 0;
1318 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1319 		return EINVAL;
1320 
1321 	mtx_lock(&filterops_lock);
1322 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1323 	    sysfilt_ops[~filt].for_fop == NULL)
1324 		error = EINVAL;
1325 	else if (sysfilt_ops[~filt].for_refcnt != 0)
1326 		error = EBUSY;
1327 	else {
1328 		sysfilt_ops[~filt].for_fop = &null_filtops;
1329 		sysfilt_ops[~filt].for_refcnt = 0;
1330 	}
1331 	mtx_unlock(&filterops_lock);
1332 
1333 	return error;
1334 }
1335 
1336 static struct filterops *
1337 kqueue_fo_find(int filt)
1338 {
1339 
1340 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1341 		return NULL;
1342 
1343 	if (sysfilt_ops[~filt].for_nolock)
1344 		return sysfilt_ops[~filt].for_fop;
1345 
1346 	mtx_lock(&filterops_lock);
1347 	sysfilt_ops[~filt].for_refcnt++;
1348 	if (sysfilt_ops[~filt].for_fop == NULL)
1349 		sysfilt_ops[~filt].for_fop = &null_filtops;
1350 	mtx_unlock(&filterops_lock);
1351 
1352 	return sysfilt_ops[~filt].for_fop;
1353 }
1354 
1355 static void
1356 kqueue_fo_release(int filt)
1357 {
1358 
1359 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1360 		return;
1361 
1362 	if (sysfilt_ops[~filt].for_nolock)
1363 		return;
1364 
1365 	mtx_lock(&filterops_lock);
1366 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1367 	    ("filter object refcount not valid on release"));
1368 	sysfilt_ops[~filt].for_refcnt--;
1369 	mtx_unlock(&filterops_lock);
1370 }
1371 
1372 /*
1373  * A ref to kq (obtained via kqueue_acquire) must be held.
1374  */
1375 static int
1376 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td,
1377     int mflag)
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(mflag);
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, M_NOWAIT) != 0) {
1424 			/* try again */
1425 			fdrop(fp, td);
1426 			fp = NULL;
1427 			error = kqueue_expand(kq, fops, kev->ident, mflag);
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 			error = kqueue_expand(kq, fops, kev->ident, mflag);
1465 			if (error != 0)
1466 				goto done;
1467 		}
1468 
1469 		KQ_LOCK(kq);
1470 
1471 		/*
1472 		 * If possible, find an existing knote to use for this kevent.
1473 		 */
1474 		if (kev->filter == EVFILT_PROC &&
1475 		    (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1476 			/* This is an internal creation of a process tracking
1477 			 * note. Don't attempt to coalesce this with an
1478 			 * existing note.
1479 			 */
1480 			;
1481 		} else if (kq->kq_knhashmask != 0) {
1482 			struct klist *list;
1483 
1484 			list = &kq->kq_knhash[
1485 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1486 			SLIST_FOREACH(kn, list, kn_link)
1487 				if (kev->ident == kn->kn_id &&
1488 				    kev->filter == kn->kn_filter)
1489 					break;
1490 		}
1491 	}
1492 
1493 	/* knote is in the process of changing, wait for it to stabilize. */
1494 	if (kn != NULL && kn_in_flux(kn)) {
1495 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1496 		if (filedesc_unlock) {
1497 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1498 			filedesc_unlock = 0;
1499 		}
1500 		kq->kq_state |= KQ_FLUXWAIT;
1501 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1502 		if (fp != NULL) {
1503 			fdrop(fp, td);
1504 			fp = NULL;
1505 		}
1506 		goto findkn;
1507 	}
1508 
1509 	/*
1510 	 * kn now contains the matching knote, or NULL if no match
1511 	 */
1512 	if (kn == NULL) {
1513 		if (kev->flags & EV_ADD) {
1514 			kn = tkn;
1515 			tkn = NULL;
1516 			if (kn == NULL) {
1517 				KQ_UNLOCK(kq);
1518 				error = ENOMEM;
1519 				goto done;
1520 			}
1521 			kn->kn_fp = fp;
1522 			kn->kn_kq = kq;
1523 			kn->kn_fop = fops;
1524 			/*
1525 			 * apply reference counts to knote structure, and
1526 			 * do not release it at the end of this routine.
1527 			 */
1528 			fops = NULL;
1529 			fp = NULL;
1530 
1531 			kn->kn_sfflags = kev->fflags;
1532 			kn->kn_sdata = kev->data;
1533 			kev->fflags = 0;
1534 			kev->data = 0;
1535 			kn->kn_kevent = *kev;
1536 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1537 			    EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1538 			kn->kn_status = KN_DETACHED;
1539 			if ((kev->flags & EV_DISABLE) != 0)
1540 				kn->kn_status |= KN_DISABLED;
1541 			kn_enter_flux(kn);
1542 
1543 			error = knote_attach(kn, kq);
1544 			KQ_UNLOCK(kq);
1545 			if (error != 0) {
1546 				tkn = kn;
1547 				goto done;
1548 			}
1549 
1550 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1551 				knote_drop_detached(kn, td);
1552 				goto done;
1553 			}
1554 			knl = kn_list_lock(kn);
1555 			goto done_ev_add;
1556 		} else {
1557 			/* No matching knote and the EV_ADD flag is not set. */
1558 			KQ_UNLOCK(kq);
1559 			error = ENOENT;
1560 			goto done;
1561 		}
1562 	}
1563 
1564 	if (kev->flags & EV_DELETE) {
1565 		kn_enter_flux(kn);
1566 		KQ_UNLOCK(kq);
1567 		knote_drop(kn, td);
1568 		goto done;
1569 	}
1570 
1571 	if (kev->flags & EV_FORCEONESHOT) {
1572 		kn->kn_flags |= EV_ONESHOT;
1573 		KNOTE_ACTIVATE(kn, 1);
1574 	}
1575 
1576 	if ((kev->flags & EV_ENABLE) != 0)
1577 		kn->kn_status &= ~KN_DISABLED;
1578 	else if ((kev->flags & EV_DISABLE) != 0)
1579 		kn->kn_status |= KN_DISABLED;
1580 
1581 	/*
1582 	 * The user may change some filter values after the initial EV_ADD,
1583 	 * but doing so will not reset any filter which has already been
1584 	 * triggered.
1585 	 */
1586 	kn->kn_status |= KN_SCAN;
1587 	kn_enter_flux(kn);
1588 	KQ_UNLOCK(kq);
1589 	knl = kn_list_lock(kn);
1590 	kn->kn_kevent.udata = kev->udata;
1591 	if (!fops->f_isfd && fops->f_touch != NULL) {
1592 		fops->f_touch(kn, kev, EVENT_REGISTER);
1593 	} else {
1594 		kn->kn_sfflags = kev->fflags;
1595 		kn->kn_sdata = kev->data;
1596 	}
1597 
1598 done_ev_add:
1599 	/*
1600 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1601 	 * the initial attach event decides that the event is "completed"
1602 	 * already, e.g., filt_procattach() is called on a zombie process.  It
1603 	 * will call filt_proc() which will remove it from the list, and NULL
1604 	 * kn_knlist.
1605 	 *
1606 	 * KN_DISABLED will be stable while the knote is in flux, so the
1607 	 * unlocked read will not race with an update.
1608 	 */
1609 	if ((kn->kn_status & KN_DISABLED) == 0)
1610 		event = kn->kn_fop->f_event(kn, 0);
1611 	else
1612 		event = 0;
1613 
1614 	KQ_LOCK(kq);
1615 	if (event)
1616 		kn->kn_status |= KN_ACTIVE;
1617 	if ((kn->kn_status & (KN_ACTIVE | KN_DISABLED | KN_QUEUED)) ==
1618 	    KN_ACTIVE)
1619 		knote_enqueue(kn);
1620 	kn->kn_status &= ~KN_SCAN;
1621 	kn_leave_flux(kn);
1622 	kn_list_unlock(knl);
1623 	KQ_UNLOCK_FLUX(kq);
1624 
1625 done:
1626 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1627 	if (filedesc_unlock)
1628 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1629 	if (fp != NULL)
1630 		fdrop(fp, td);
1631 	knote_free(tkn);
1632 	if (fops != NULL)
1633 		kqueue_fo_release(filt);
1634 	return (error);
1635 }
1636 
1637 static int
1638 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1639 {
1640 	int error;
1641 	struct kqueue *kq;
1642 
1643 	error = 0;
1644 
1645 	kq = fp->f_data;
1646 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1647 		return (EBADF);
1648 	*kqp = kq;
1649 	KQ_LOCK(kq);
1650 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1651 		KQ_UNLOCK(kq);
1652 		return (EBADF);
1653 	}
1654 	kq->kq_refcnt++;
1655 	KQ_UNLOCK(kq);
1656 
1657 	return error;
1658 }
1659 
1660 static void
1661 kqueue_release(struct kqueue *kq, int locked)
1662 {
1663 	if (locked)
1664 		KQ_OWNED(kq);
1665 	else
1666 		KQ_LOCK(kq);
1667 	kq->kq_refcnt--;
1668 	if (kq->kq_refcnt == 1)
1669 		wakeup(&kq->kq_refcnt);
1670 	if (!locked)
1671 		KQ_UNLOCK(kq);
1672 }
1673 
1674 static void
1675 kqueue_schedtask(struct kqueue *kq)
1676 {
1677 
1678 	KQ_OWNED(kq);
1679 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1680 	    ("scheduling kqueue task while draining"));
1681 
1682 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1683 		taskqueue_enqueue(taskqueue_kqueue_ctx, &kq->kq_task);
1684 		kq->kq_state |= KQ_TASKSCHED;
1685 	}
1686 }
1687 
1688 /*
1689  * Expand the kq to make sure we have storage for fops/ident pair.
1690  *
1691  * Return 0 on success (or no work necessary), return errno on failure.
1692  */
1693 static int
1694 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1695     int mflag)
1696 {
1697 	struct klist *list, *tmp_knhash, *to_free;
1698 	u_long tmp_knhashmask;
1699 	int error, fd, size;
1700 
1701 	KQ_NOTOWNED(kq);
1702 
1703 	error = 0;
1704 	to_free = NULL;
1705 	if (fops->f_isfd) {
1706 		fd = ident;
1707 		if (kq->kq_knlistsize <= fd) {
1708 			size = kq->kq_knlistsize;
1709 			while (size <= fd)
1710 				size += KQEXTENT;
1711 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1712 			if (list == NULL)
1713 				return ENOMEM;
1714 			KQ_LOCK(kq);
1715 			if ((kq->kq_state & KQ_CLOSING) != 0) {
1716 				to_free = list;
1717 				error = EBADF;
1718 			} else if (kq->kq_knlistsize > fd) {
1719 				to_free = list;
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_flags(KN_HASHSIZE, M_KQUEUE,
1738 			    &tmp_knhashmask, (mflag & M_WAITOK) != 0 ?
1739 			    HASH_WAITOK : HASH_NOWAIT);
1740 			if (tmp_knhash == NULL)
1741 				return (ENOMEM);
1742 			KQ_LOCK(kq);
1743 			if ((kq->kq_state & KQ_CLOSING) != 0) {
1744 				to_free = tmp_knhash;
1745 				error = EBADF;
1746 			} else if (kq->kq_knhashmask == 0) {
1747 				kq->kq_knhash = tmp_knhash;
1748 				kq->kq_knhashmask = tmp_knhashmask;
1749 			} else {
1750 				to_free = tmp_knhash;
1751 			}
1752 			KQ_UNLOCK(kq);
1753 		}
1754 	}
1755 	free(to_free, M_KQUEUE);
1756 
1757 	KQ_NOTOWNED(kq);
1758 	return (error);
1759 }
1760 
1761 static void
1762 kqueue_task(void *arg, int pending)
1763 {
1764 	struct kqueue *kq;
1765 	int haskqglobal;
1766 
1767 	haskqglobal = 0;
1768 	kq = arg;
1769 
1770 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1771 	KQ_LOCK(kq);
1772 
1773 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1774 
1775 	kq->kq_state &= ~KQ_TASKSCHED;
1776 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1777 		wakeup(&kq->kq_state);
1778 	}
1779 	KQ_UNLOCK(kq);
1780 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1781 }
1782 
1783 /*
1784  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1785  * We treat KN_MARKER knotes as if they are in flux.
1786  */
1787 static int
1788 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1789     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1790 {
1791 	struct kevent *kevp;
1792 	struct knote *kn, *marker;
1793 	struct knlist *knl;
1794 	sbintime_t asbt, rsbt;
1795 	int count, error, haskqglobal, influx, nkev, touch;
1796 
1797 	count = maxevents;
1798 	nkev = 0;
1799 	error = 0;
1800 	haskqglobal = 0;
1801 
1802 	if (maxevents == 0)
1803 		goto done_nl;
1804 
1805 	rsbt = 0;
1806 	if (tsp != NULL) {
1807 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1808 		    tsp->tv_nsec >= 1000000000) {
1809 			error = EINVAL;
1810 			goto done_nl;
1811 		}
1812 		if (timespecisset(tsp)) {
1813 			if (tsp->tv_sec <= INT32_MAX) {
1814 				rsbt = tstosbt(*tsp);
1815 				if (TIMESEL(&asbt, rsbt))
1816 					asbt += tc_tick_sbt;
1817 				if (asbt <= SBT_MAX - rsbt)
1818 					asbt += rsbt;
1819 				else
1820 					asbt = 0;
1821 				rsbt >>= tc_precexp;
1822 			} else
1823 				asbt = 0;
1824 		} else
1825 			asbt = -1;
1826 	} else
1827 		asbt = 0;
1828 	marker = knote_alloc(M_WAITOK);
1829 	marker->kn_status = KN_MARKER;
1830 	KQ_LOCK(kq);
1831 
1832 retry:
1833 	kevp = keva;
1834 	if (kq->kq_count == 0) {
1835 		if (asbt == -1) {
1836 			error = EWOULDBLOCK;
1837 		} else {
1838 			kq->kq_state |= KQ_SLEEP;
1839 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1840 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1841 		}
1842 		if (error == 0)
1843 			goto retry;
1844 		/* don't restart after signals... */
1845 		if (error == ERESTART)
1846 			error = EINTR;
1847 		else if (error == EWOULDBLOCK)
1848 			error = 0;
1849 		goto done;
1850 	}
1851 
1852 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1853 	influx = 0;
1854 	while (count) {
1855 		KQ_OWNED(kq);
1856 		kn = TAILQ_FIRST(&kq->kq_head);
1857 
1858 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1859 		    kn_in_flux(kn)) {
1860 			if (influx) {
1861 				influx = 0;
1862 				KQ_FLUX_WAKEUP(kq);
1863 			}
1864 			kq->kq_state |= KQ_FLUXWAIT;
1865 			error = msleep(kq, &kq->kq_lock, PSOCK,
1866 			    "kqflxwt", 0);
1867 			continue;
1868 		}
1869 
1870 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1871 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1872 			kn->kn_status &= ~KN_QUEUED;
1873 			kq->kq_count--;
1874 			continue;
1875 		}
1876 		if (kn == marker) {
1877 			KQ_FLUX_WAKEUP(kq);
1878 			if (count == maxevents)
1879 				goto retry;
1880 			goto done;
1881 		}
1882 		KASSERT(!kn_in_flux(kn),
1883 		    ("knote %p is unexpectedly in flux", kn));
1884 
1885 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1886 			kn->kn_status &= ~KN_QUEUED;
1887 			kn_enter_flux(kn);
1888 			kq->kq_count--;
1889 			KQ_UNLOCK(kq);
1890 			/*
1891 			 * We don't need to lock the list since we've
1892 			 * marked it as in flux.
1893 			 */
1894 			knote_drop(kn, td);
1895 			KQ_LOCK(kq);
1896 			continue;
1897 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1898 			kn->kn_status &= ~KN_QUEUED;
1899 			kn_enter_flux(kn);
1900 			kq->kq_count--;
1901 			KQ_UNLOCK(kq);
1902 			/*
1903 			 * We don't need to lock the list since we've
1904 			 * marked the knote as being in flux.
1905 			 */
1906 			*kevp = kn->kn_kevent;
1907 			knote_drop(kn, td);
1908 			KQ_LOCK(kq);
1909 			kn = NULL;
1910 		} else {
1911 			kn->kn_status |= KN_SCAN;
1912 			kn_enter_flux(kn);
1913 			KQ_UNLOCK(kq);
1914 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1915 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1916 			knl = kn_list_lock(kn);
1917 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1918 				KQ_LOCK(kq);
1919 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1920 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE |
1921 				    KN_SCAN);
1922 				kn_leave_flux(kn);
1923 				kq->kq_count--;
1924 				kn_list_unlock(knl);
1925 				influx = 1;
1926 				continue;
1927 			}
1928 			touch = (!kn->kn_fop->f_isfd &&
1929 			    kn->kn_fop->f_touch != NULL);
1930 			if (touch)
1931 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1932 			else
1933 				*kevp = kn->kn_kevent;
1934 			KQ_LOCK(kq);
1935 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1936 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1937 				/*
1938 				 * Manually clear knotes who weren't
1939 				 * 'touch'ed.
1940 				 */
1941 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1942 					kn->kn_data = 0;
1943 					kn->kn_fflags = 0;
1944 				}
1945 				if (kn->kn_flags & EV_DISPATCH)
1946 					kn->kn_status |= KN_DISABLED;
1947 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1948 				kq->kq_count--;
1949 			} else
1950 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1951 
1952 			kn->kn_status &= ~KN_SCAN;
1953 			kn_leave_flux(kn);
1954 			kn_list_unlock(knl);
1955 			influx = 1;
1956 		}
1957 
1958 		/* we are returning a copy to the user */
1959 		kevp++;
1960 		nkev++;
1961 		count--;
1962 
1963 		if (nkev == KQ_NEVENTS) {
1964 			influx = 0;
1965 			KQ_UNLOCK_FLUX(kq);
1966 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1967 			nkev = 0;
1968 			kevp = keva;
1969 			KQ_LOCK(kq);
1970 			if (error)
1971 				break;
1972 		}
1973 	}
1974 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1975 done:
1976 	KQ_OWNED(kq);
1977 	KQ_UNLOCK_FLUX(kq);
1978 	knote_free(marker);
1979 done_nl:
1980 	KQ_NOTOWNED(kq);
1981 	if (nkev != 0)
1982 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1983 	td->td_retval[0] = maxevents - count;
1984 	return (error);
1985 }
1986 
1987 /*ARGSUSED*/
1988 static int
1989 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1990 	struct ucred *active_cred, struct thread *td)
1991 {
1992 	/*
1993 	 * Enabling sigio causes two major problems:
1994 	 * 1) infinite recursion:
1995 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1996 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1997 	 * into itself over and over.  Sending the sigio causes the kqueue
1998 	 * to become ready, which in turn posts sigio again, forever.
1999 	 * Solution: this can be solved by setting a flag in the kqueue that
2000 	 * we have a SIGIO in progress.
2001 	 * 2) locking problems:
2002 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
2003 	 * us above the proc and pgrp locks.
2004 	 * Solution: Post a signal using an async mechanism, being sure to
2005 	 * record a generation count in the delivery so that we do not deliver
2006 	 * a signal to the wrong process.
2007 	 *
2008 	 * Note, these two mechanisms are somewhat mutually exclusive!
2009 	 */
2010 #if 0
2011 	struct kqueue *kq;
2012 
2013 	kq = fp->f_data;
2014 	switch (cmd) {
2015 	case FIOASYNC:
2016 		if (*(int *)data) {
2017 			kq->kq_state |= KQ_ASYNC;
2018 		} else {
2019 			kq->kq_state &= ~KQ_ASYNC;
2020 		}
2021 		return (0);
2022 
2023 	case FIOSETOWN:
2024 		return (fsetown(*(int *)data, &kq->kq_sigio));
2025 
2026 	case FIOGETOWN:
2027 		*(int *)data = fgetown(&kq->kq_sigio);
2028 		return (0);
2029 	}
2030 #endif
2031 
2032 	return (ENOTTY);
2033 }
2034 
2035 /*ARGSUSED*/
2036 static int
2037 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
2038 	struct thread *td)
2039 {
2040 	struct kqueue *kq;
2041 	int revents = 0;
2042 	int error;
2043 
2044 	if ((error = kqueue_acquire(fp, &kq)))
2045 		return POLLERR;
2046 
2047 	KQ_LOCK(kq);
2048 	if (events & (POLLIN | POLLRDNORM)) {
2049 		if (kq->kq_count) {
2050 			revents |= events & (POLLIN | POLLRDNORM);
2051 		} else {
2052 			selrecord(td, &kq->kq_sel);
2053 			if (SEL_WAITING(&kq->kq_sel))
2054 				kq->kq_state |= KQ_SEL;
2055 		}
2056 	}
2057 	kqueue_release(kq, 1);
2058 	KQ_UNLOCK(kq);
2059 	return (revents);
2060 }
2061 
2062 /*ARGSUSED*/
2063 static int
2064 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
2065 	struct thread *td)
2066 {
2067 
2068 	bzero((void *)st, sizeof *st);
2069 	/*
2070 	 * We no longer return kq_count because the unlocked value is useless.
2071 	 * If you spent all this time getting the count, why not spend your
2072 	 * syscall better by calling kevent?
2073 	 *
2074 	 * XXX - This is needed for libc_r.
2075 	 */
2076 	st->st_mode = S_IFIFO;
2077 	return (0);
2078 }
2079 
2080 static void
2081 kqueue_drain(struct kqueue *kq, struct thread *td)
2082 {
2083 	struct knote *kn;
2084 	int i;
2085 
2086 	KQ_LOCK(kq);
2087 
2088 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
2089 	    ("kqueue already closing"));
2090 	kq->kq_state |= KQ_CLOSING;
2091 	if (kq->kq_refcnt > 1)
2092 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
2093 
2094 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
2095 
2096 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
2097 	    ("kqueue's knlist not empty"));
2098 
2099 	for (i = 0; i < kq->kq_knlistsize; i++) {
2100 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
2101 			if (kn_in_flux(kn)) {
2102 				kq->kq_state |= KQ_FLUXWAIT;
2103 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
2104 				continue;
2105 			}
2106 			kn_enter_flux(kn);
2107 			KQ_UNLOCK(kq);
2108 			knote_drop(kn, td);
2109 			KQ_LOCK(kq);
2110 		}
2111 	}
2112 	if (kq->kq_knhashmask != 0) {
2113 		for (i = 0; i <= kq->kq_knhashmask; i++) {
2114 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
2115 				if (kn_in_flux(kn)) {
2116 					kq->kq_state |= KQ_FLUXWAIT;
2117 					msleep(kq, &kq->kq_lock, PSOCK,
2118 					       "kqclo2", 0);
2119 					continue;
2120 				}
2121 				kn_enter_flux(kn);
2122 				KQ_UNLOCK(kq);
2123 				knote_drop(kn, td);
2124 				KQ_LOCK(kq);
2125 			}
2126 		}
2127 	}
2128 
2129 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
2130 		kq->kq_state |= KQ_TASKDRAIN;
2131 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
2132 	}
2133 
2134 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2135 		selwakeuppri(&kq->kq_sel, PSOCK);
2136 		if (!SEL_WAITING(&kq->kq_sel))
2137 			kq->kq_state &= ~KQ_SEL;
2138 	}
2139 
2140 	KQ_UNLOCK(kq);
2141 }
2142 
2143 static void
2144 kqueue_destroy(struct kqueue *kq)
2145 {
2146 
2147 	KASSERT(kq->kq_fdp == NULL,
2148 	    ("kqueue still attached to a file descriptor"));
2149 	seldrain(&kq->kq_sel);
2150 	knlist_destroy(&kq->kq_sel.si_note);
2151 	mtx_destroy(&kq->kq_lock);
2152 
2153 	if (kq->kq_knhash != NULL)
2154 		free(kq->kq_knhash, M_KQUEUE);
2155 	if (kq->kq_knlist != NULL)
2156 		free(kq->kq_knlist, M_KQUEUE);
2157 
2158 	funsetown(&kq->kq_sigio);
2159 }
2160 
2161 /*ARGSUSED*/
2162 static int
2163 kqueue_close(struct file *fp, struct thread *td)
2164 {
2165 	struct kqueue *kq = fp->f_data;
2166 	struct filedesc *fdp;
2167 	int error;
2168 	int filedesc_unlock;
2169 
2170 	if ((error = kqueue_acquire(fp, &kq)))
2171 		return error;
2172 	kqueue_drain(kq, td);
2173 
2174 	/*
2175 	 * We could be called due to the knote_drop() doing fdrop(),
2176 	 * called from kqueue_register().  In this case the global
2177 	 * lock is owned, and filedesc sx is locked before, to not
2178 	 * take the sleepable lock after non-sleepable.
2179 	 */
2180 	fdp = kq->kq_fdp;
2181 	kq->kq_fdp = NULL;
2182 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
2183 		FILEDESC_XLOCK(fdp);
2184 		filedesc_unlock = 1;
2185 	} else
2186 		filedesc_unlock = 0;
2187 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
2188 	if (filedesc_unlock)
2189 		FILEDESC_XUNLOCK(fdp);
2190 
2191 	kqueue_destroy(kq);
2192 	chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
2193 	crfree(kq->kq_cred);
2194 	free(kq, M_KQUEUE);
2195 	fp->f_data = NULL;
2196 
2197 	return (0);
2198 }
2199 
2200 static int
2201 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2202 {
2203 
2204 	kif->kf_type = KF_TYPE_KQUEUE;
2205 	return (0);
2206 }
2207 
2208 static void
2209 kqueue_wakeup(struct kqueue *kq)
2210 {
2211 	KQ_OWNED(kq);
2212 
2213 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
2214 		kq->kq_state &= ~KQ_SLEEP;
2215 		wakeup(kq);
2216 	}
2217 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2218 		selwakeuppri(&kq->kq_sel, PSOCK);
2219 		if (!SEL_WAITING(&kq->kq_sel))
2220 			kq->kq_state &= ~KQ_SEL;
2221 	}
2222 	if (!knlist_empty(&kq->kq_sel.si_note))
2223 		kqueue_schedtask(kq);
2224 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
2225 		pgsigio(&kq->kq_sigio, SIGIO, 0);
2226 	}
2227 }
2228 
2229 /*
2230  * Walk down a list of knotes, activating them if their event has triggered.
2231  *
2232  * There is a possibility to optimize in the case of one kq watching another.
2233  * Instead of scheduling a task to wake it up, you could pass enough state
2234  * down the chain to make up the parent kqueue.  Make this code functional
2235  * first.
2236  */
2237 void
2238 knote(struct knlist *list, long hint, int lockflags)
2239 {
2240 	struct kqueue *kq;
2241 	struct knote *kn, *tkn;
2242 	int error;
2243 
2244 	if (list == NULL)
2245 		return;
2246 
2247 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
2248 
2249 	if ((lockflags & KNF_LISTLOCKED) == 0)
2250 		list->kl_lock(list->kl_lockarg);
2251 
2252 	/*
2253 	 * If we unlock the list lock (and enter influx), we can
2254 	 * eliminate the kqueue scheduling, but this will introduce
2255 	 * four lock/unlock's for each knote to test.  Also, marker
2256 	 * would be needed to keep iteration position, since filters
2257 	 * or other threads could remove events.
2258 	 */
2259 	SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
2260 		kq = kn->kn_kq;
2261 		KQ_LOCK(kq);
2262 		if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
2263 			/*
2264 			 * Do not process the influx notes, except for
2265 			 * the influx coming from the kq unlock in the
2266 			 * kqueue_scan().  In the later case, we do
2267 			 * not interfere with the scan, since the code
2268 			 * fragment in kqueue_scan() locks the knlist,
2269 			 * and cannot proceed until we finished.
2270 			 */
2271 			KQ_UNLOCK(kq);
2272 		} else if ((lockflags & KNF_NOKQLOCK) != 0) {
2273 			kn_enter_flux(kn);
2274 			KQ_UNLOCK(kq);
2275 			error = kn->kn_fop->f_event(kn, hint);
2276 			KQ_LOCK(kq);
2277 			kn_leave_flux(kn);
2278 			if (error)
2279 				KNOTE_ACTIVATE(kn, 1);
2280 			KQ_UNLOCK_FLUX(kq);
2281 		} else {
2282 			if (kn->kn_fop->f_event(kn, hint))
2283 				KNOTE_ACTIVATE(kn, 1);
2284 			KQ_UNLOCK(kq);
2285 		}
2286 	}
2287 	if ((lockflags & KNF_LISTLOCKED) == 0)
2288 		list->kl_unlock(list->kl_lockarg);
2289 }
2290 
2291 /*
2292  * add a knote to a knlist
2293  */
2294 void
2295 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2296 {
2297 
2298 	KNL_ASSERT_LOCK(knl, islocked);
2299 	KQ_NOTOWNED(kn->kn_kq);
2300 	KASSERT(kn_in_flux(kn), ("knote %p not in flux", kn));
2301 	KASSERT((kn->kn_status & KN_DETACHED) != 0,
2302 	    ("knote %p was not detached", kn));
2303 	if (!islocked)
2304 		knl->kl_lock(knl->kl_lockarg);
2305 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2306 	if (!islocked)
2307 		knl->kl_unlock(knl->kl_lockarg);
2308 	KQ_LOCK(kn->kn_kq);
2309 	kn->kn_knlist = knl;
2310 	kn->kn_status &= ~KN_DETACHED;
2311 	KQ_UNLOCK(kn->kn_kq);
2312 }
2313 
2314 static void
2315 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked,
2316     int kqislocked)
2317 {
2318 
2319 	KASSERT(!kqislocked || knlislocked, ("kq locked w/o knl locked"));
2320 	KNL_ASSERT_LOCK(knl, knlislocked);
2321 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2322 	KASSERT(kqislocked || kn_in_flux(kn), ("knote %p not in flux", kn));
2323 	KASSERT((kn->kn_status & KN_DETACHED) == 0,
2324 	    ("knote %p was already detached", kn));
2325 	if (!knlislocked)
2326 		knl->kl_lock(knl->kl_lockarg);
2327 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2328 	kn->kn_knlist = NULL;
2329 	if (!knlislocked)
2330 		kn_list_unlock(knl);
2331 	if (!kqislocked)
2332 		KQ_LOCK(kn->kn_kq);
2333 	kn->kn_status |= KN_DETACHED;
2334 	if (!kqislocked)
2335 		KQ_UNLOCK(kn->kn_kq);
2336 }
2337 
2338 /*
2339  * remove knote from the specified knlist
2340  */
2341 void
2342 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2343 {
2344 
2345 	knlist_remove_kq(knl, kn, islocked, 0);
2346 }
2347 
2348 int
2349 knlist_empty(struct knlist *knl)
2350 {
2351 
2352 	KNL_ASSERT_LOCKED(knl);
2353 	return (SLIST_EMPTY(&knl->kl_list));
2354 }
2355 
2356 static struct mtx knlist_lock;
2357 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2358     MTX_DEF);
2359 static void knlist_mtx_lock(void *arg);
2360 static void knlist_mtx_unlock(void *arg);
2361 
2362 static void
2363 knlist_mtx_lock(void *arg)
2364 {
2365 
2366 	mtx_lock((struct mtx *)arg);
2367 }
2368 
2369 static void
2370 knlist_mtx_unlock(void *arg)
2371 {
2372 
2373 	mtx_unlock((struct mtx *)arg);
2374 }
2375 
2376 static void
2377 knlist_mtx_assert_locked(void *arg)
2378 {
2379 
2380 	mtx_assert((struct mtx *)arg, MA_OWNED);
2381 }
2382 
2383 static void
2384 knlist_mtx_assert_unlocked(void *arg)
2385 {
2386 
2387 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2388 }
2389 
2390 static void
2391 knlist_rw_rlock(void *arg)
2392 {
2393 
2394 	rw_rlock((struct rwlock *)arg);
2395 }
2396 
2397 static void
2398 knlist_rw_runlock(void *arg)
2399 {
2400 
2401 	rw_runlock((struct rwlock *)arg);
2402 }
2403 
2404 static void
2405 knlist_rw_assert_locked(void *arg)
2406 {
2407 
2408 	rw_assert((struct rwlock *)arg, RA_LOCKED);
2409 }
2410 
2411 static void
2412 knlist_rw_assert_unlocked(void *arg)
2413 {
2414 
2415 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2416 }
2417 
2418 void
2419 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2420     void (*kl_unlock)(void *),
2421     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2422 {
2423 
2424 	if (lock == NULL)
2425 		knl->kl_lockarg = &knlist_lock;
2426 	else
2427 		knl->kl_lockarg = lock;
2428 
2429 	if (kl_lock == NULL)
2430 		knl->kl_lock = knlist_mtx_lock;
2431 	else
2432 		knl->kl_lock = kl_lock;
2433 	if (kl_unlock == NULL)
2434 		knl->kl_unlock = knlist_mtx_unlock;
2435 	else
2436 		knl->kl_unlock = kl_unlock;
2437 	if (kl_assert_locked == NULL)
2438 		knl->kl_assert_locked = knlist_mtx_assert_locked;
2439 	else
2440 		knl->kl_assert_locked = kl_assert_locked;
2441 	if (kl_assert_unlocked == NULL)
2442 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2443 	else
2444 		knl->kl_assert_unlocked = kl_assert_unlocked;
2445 
2446 	knl->kl_autodestroy = 0;
2447 	SLIST_INIT(&knl->kl_list);
2448 }
2449 
2450 void
2451 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2452 {
2453 
2454 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2455 }
2456 
2457 struct knlist *
2458 knlist_alloc(struct mtx *lock)
2459 {
2460 	struct knlist *knl;
2461 
2462 	knl = malloc(sizeof(struct knlist), M_KQUEUE, M_WAITOK);
2463 	knlist_init_mtx(knl, lock);
2464 	return (knl);
2465 }
2466 
2467 void
2468 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2469 {
2470 
2471 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2472 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2473 }
2474 
2475 void
2476 knlist_destroy(struct knlist *knl)
2477 {
2478 
2479 	KASSERT(KNLIST_EMPTY(knl),
2480 	    ("destroying knlist %p with knotes on it", knl));
2481 }
2482 
2483 void
2484 knlist_detach(struct knlist *knl)
2485 {
2486 
2487 	KNL_ASSERT_LOCKED(knl);
2488 	knl->kl_autodestroy = 1;
2489 	if (knlist_empty(knl)) {
2490 		knlist_destroy(knl);
2491 		free(knl, M_KQUEUE);
2492 	}
2493 }
2494 
2495 /*
2496  * Even if we are locked, we may need to drop the lock to allow any influx
2497  * knotes time to "settle".
2498  */
2499 void
2500 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2501 {
2502 	struct knote *kn, *kn2;
2503 	struct kqueue *kq;
2504 
2505 	KASSERT(!knl->kl_autodestroy, ("cleardel for autodestroy %p", knl));
2506 	if (islocked)
2507 		KNL_ASSERT_LOCKED(knl);
2508 	else {
2509 		KNL_ASSERT_UNLOCKED(knl);
2510 again:		/* need to reacquire lock since we have dropped it */
2511 		knl->kl_lock(knl->kl_lockarg);
2512 	}
2513 
2514 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2515 		kq = kn->kn_kq;
2516 		KQ_LOCK(kq);
2517 		if (kn_in_flux(kn)) {
2518 			KQ_UNLOCK(kq);
2519 			continue;
2520 		}
2521 		knlist_remove_kq(knl, kn, 1, 1);
2522 		if (killkn) {
2523 			kn_enter_flux(kn);
2524 			KQ_UNLOCK(kq);
2525 			knote_drop_detached(kn, td);
2526 		} else {
2527 			/* Make sure cleared knotes disappear soon */
2528 			kn->kn_flags |= EV_EOF | EV_ONESHOT;
2529 			KQ_UNLOCK(kq);
2530 		}
2531 		kq = NULL;
2532 	}
2533 
2534 	if (!SLIST_EMPTY(&knl->kl_list)) {
2535 		/* there are still in flux knotes remaining */
2536 		kn = SLIST_FIRST(&knl->kl_list);
2537 		kq = kn->kn_kq;
2538 		KQ_LOCK(kq);
2539 		KASSERT(kn_in_flux(kn), ("knote removed w/o list lock"));
2540 		knl->kl_unlock(knl->kl_lockarg);
2541 		kq->kq_state |= KQ_FLUXWAIT;
2542 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2543 		kq = NULL;
2544 		goto again;
2545 	}
2546 
2547 	if (islocked)
2548 		KNL_ASSERT_LOCKED(knl);
2549 	else {
2550 		knl->kl_unlock(knl->kl_lockarg);
2551 		KNL_ASSERT_UNLOCKED(knl);
2552 	}
2553 }
2554 
2555 /*
2556  * Remove all knotes referencing a specified fd must be called with FILEDESC
2557  * lock.  This prevents a race where a new fd comes along and occupies the
2558  * entry and we attach a knote to the fd.
2559  */
2560 void
2561 knote_fdclose(struct thread *td, int fd)
2562 {
2563 	struct filedesc *fdp = td->td_proc->p_fd;
2564 	struct kqueue *kq;
2565 	struct knote *kn;
2566 	int influx;
2567 
2568 	FILEDESC_XLOCK_ASSERT(fdp);
2569 
2570 	/*
2571 	 * We shouldn't have to worry about new kevents appearing on fd
2572 	 * since filedesc is locked.
2573 	 */
2574 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2575 		KQ_LOCK(kq);
2576 
2577 again:
2578 		influx = 0;
2579 		while (kq->kq_knlistsize > fd &&
2580 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2581 			if (kn_in_flux(kn)) {
2582 				/* someone else might be waiting on our knote */
2583 				if (influx)
2584 					wakeup(kq);
2585 				kq->kq_state |= KQ_FLUXWAIT;
2586 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2587 				goto again;
2588 			}
2589 			kn_enter_flux(kn);
2590 			KQ_UNLOCK(kq);
2591 			influx = 1;
2592 			knote_drop(kn, td);
2593 			KQ_LOCK(kq);
2594 		}
2595 		KQ_UNLOCK_FLUX(kq);
2596 	}
2597 }
2598 
2599 static int
2600 knote_attach(struct knote *kn, struct kqueue *kq)
2601 {
2602 	struct klist *list;
2603 
2604 	KASSERT(kn_in_flux(kn), ("knote %p not marked influx", kn));
2605 	KQ_OWNED(kq);
2606 
2607 	if ((kq->kq_state & KQ_CLOSING) != 0)
2608 		return (EBADF);
2609 	if (kn->kn_fop->f_isfd) {
2610 		if (kn->kn_id >= kq->kq_knlistsize)
2611 			return (ENOMEM);
2612 		list = &kq->kq_knlist[kn->kn_id];
2613 	} else {
2614 		if (kq->kq_knhash == NULL)
2615 			return (ENOMEM);
2616 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2617 	}
2618 	SLIST_INSERT_HEAD(list, kn, kn_link);
2619 	return (0);
2620 }
2621 
2622 static void
2623 knote_drop(struct knote *kn, struct thread *td)
2624 {
2625 
2626 	if ((kn->kn_status & KN_DETACHED) == 0)
2627 		kn->kn_fop->f_detach(kn);
2628 	knote_drop_detached(kn, td);
2629 }
2630 
2631 static void
2632 knote_drop_detached(struct knote *kn, struct thread *td)
2633 {
2634 	struct kqueue *kq;
2635 	struct klist *list;
2636 
2637 	kq = kn->kn_kq;
2638 
2639 	KASSERT((kn->kn_status & KN_DETACHED) != 0,
2640 	    ("knote %p still attached", kn));
2641 	KQ_NOTOWNED(kq);
2642 
2643 	KQ_LOCK(kq);
2644 	KASSERT(kn->kn_influx == 1,
2645 	    ("knote_drop called on %p with influx %d", kn, kn->kn_influx));
2646 
2647 	if (kn->kn_fop->f_isfd)
2648 		list = &kq->kq_knlist[kn->kn_id];
2649 	else
2650 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2651 
2652 	if (!SLIST_EMPTY(list))
2653 		SLIST_REMOVE(list, kn, knote, kn_link);
2654 	if (kn->kn_status & KN_QUEUED)
2655 		knote_dequeue(kn);
2656 	KQ_UNLOCK_FLUX(kq);
2657 
2658 	if (kn->kn_fop->f_isfd) {
2659 		fdrop(kn->kn_fp, td);
2660 		kn->kn_fp = NULL;
2661 	}
2662 	kqueue_fo_release(kn->kn_kevent.filter);
2663 	kn->kn_fop = NULL;
2664 	knote_free(kn);
2665 }
2666 
2667 static void
2668 knote_enqueue(struct knote *kn)
2669 {
2670 	struct kqueue *kq = kn->kn_kq;
2671 
2672 	KQ_OWNED(kn->kn_kq);
2673 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2674 
2675 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2676 	kn->kn_status |= KN_QUEUED;
2677 	kq->kq_count++;
2678 	kqueue_wakeup(kq);
2679 }
2680 
2681 static void
2682 knote_dequeue(struct knote *kn)
2683 {
2684 	struct kqueue *kq = kn->kn_kq;
2685 
2686 	KQ_OWNED(kn->kn_kq);
2687 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2688 
2689 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2690 	kn->kn_status &= ~KN_QUEUED;
2691 	kq->kq_count--;
2692 }
2693 
2694 static void
2695 knote_init(void)
2696 {
2697 
2698 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2699 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2700 }
2701 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2702 
2703 static struct knote *
2704 knote_alloc(int mflag)
2705 {
2706 
2707 	return (uma_zalloc(knote_zone, mflag | M_ZERO));
2708 }
2709 
2710 static void
2711 knote_free(struct knote *kn)
2712 {
2713 
2714 	uma_zfree(knote_zone, kn);
2715 }
2716 
2717 /*
2718  * Register the kev w/ the kq specified by fd.
2719  */
2720 int
2721 kqfd_register(int fd, struct kevent *kev, struct thread *td, int mflag)
2722 {
2723 	struct kqueue *kq;
2724 	struct file *fp;
2725 	cap_rights_t rights;
2726 	int error;
2727 
2728 	error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2729 	if (error != 0)
2730 		return (error);
2731 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2732 		goto noacquire;
2733 
2734 	error = kqueue_register(kq, kev, td, mflag);
2735 	kqueue_release(kq, 0);
2736 
2737 noacquire:
2738 	fdrop(fp, td);
2739 	return (error);
2740 }
2741