xref: /freebsd/sys/kern/kern_event.c (revision 3fc36ee018bb836bd1796067cf4ef8683f166ebc)
1 /*-
2  * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
3  * Copyright 2004 John-Mark Gurney <jmg@FreeBSD.org>
4  * Copyright (c) 2009 Apple, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include "opt_ktrace.h"
33 #include "opt_kqueue.h"
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/capsicum.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/rwlock.h>
42 #include <sys/proc.h>
43 #include <sys/malloc.h>
44 #include <sys/unistd.h>
45 #include <sys/file.h>
46 #include <sys/filedesc.h>
47 #include <sys/filio.h>
48 #include <sys/fcntl.h>
49 #include <sys/kthread.h>
50 #include <sys/selinfo.h>
51 #include <sys/stdatomic.h>
52 #include <sys/queue.h>
53 #include <sys/event.h>
54 #include <sys/eventvar.h>
55 #include <sys/poll.h>
56 #include <sys/protosw.h>
57 #include <sys/resourcevar.h>
58 #include <sys/sigio.h>
59 #include <sys/signalvar.h>
60 #include <sys/socket.h>
61 #include <sys/socketvar.h>
62 #include <sys/stat.h>
63 #include <sys/sysctl.h>
64 #include <sys/sysproto.h>
65 #include <sys/syscallsubr.h>
66 #include <sys/taskqueue.h>
67 #include <sys/uio.h>
68 #include <sys/user.h>
69 #ifdef KTRACE
70 #include <sys/ktrace.h>
71 #endif
72 
73 #include <vm/uma.h>
74 
75 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
76 
77 /*
78  * This lock is used if multiple kq locks are required.  This possibly
79  * should be made into a per proc lock.
80  */
81 static struct mtx	kq_global;
82 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
83 #define KQ_GLOBAL_LOCK(lck, haslck)	do {	\
84 	if (!haslck)				\
85 		mtx_lock(lck);			\
86 	haslck = 1;				\
87 } while (0)
88 #define KQ_GLOBAL_UNLOCK(lck, haslck)	do {	\
89 	if (haslck)				\
90 		mtx_unlock(lck);			\
91 	haslck = 0;				\
92 } while (0)
93 
94 TASKQUEUE_DEFINE_THREAD(kqueue_ctx);
95 
96 static int	kevent_copyout(void *arg, struct kevent *kevp, int count);
97 static int	kevent_copyin(void *arg, struct kevent *kevp, int count);
98 static int	kqueue_register(struct kqueue *kq, struct kevent *kev,
99 		    struct thread *td, int waitok);
100 static int	kqueue_acquire(struct file *fp, struct kqueue **kqp);
101 static void	kqueue_release(struct kqueue *kq, int locked);
102 static void	kqueue_destroy(struct kqueue *kq);
103 static void	kqueue_drain(struct kqueue *kq, struct thread *td);
104 static int	kqueue_expand(struct kqueue *kq, struct filterops *fops,
105 		    uintptr_t ident, int waitok);
106 static void	kqueue_task(void *arg, int pending);
107 static int	kqueue_scan(struct kqueue *kq, int maxevents,
108 		    struct kevent_copyops *k_ops,
109 		    const struct timespec *timeout,
110 		    struct kevent *keva, struct thread *td);
111 static void 	kqueue_wakeup(struct kqueue *kq);
112 static struct filterops *kqueue_fo_find(int filt);
113 static void	kqueue_fo_release(int filt);
114 
115 static fo_ioctl_t	kqueue_ioctl;
116 static fo_poll_t	kqueue_poll;
117 static fo_kqfilter_t	kqueue_kqfilter;
118 static fo_stat_t	kqueue_stat;
119 static fo_close_t	kqueue_close;
120 static fo_fill_kinfo_t	kqueue_fill_kinfo;
121 
122 static struct fileops kqueueops = {
123 	.fo_read = invfo_rdwr,
124 	.fo_write = invfo_rdwr,
125 	.fo_truncate = invfo_truncate,
126 	.fo_ioctl = kqueue_ioctl,
127 	.fo_poll = kqueue_poll,
128 	.fo_kqfilter = kqueue_kqfilter,
129 	.fo_stat = kqueue_stat,
130 	.fo_close = kqueue_close,
131 	.fo_chmod = invfo_chmod,
132 	.fo_chown = invfo_chown,
133 	.fo_sendfile = invfo_sendfile,
134 	.fo_fill_kinfo = kqueue_fill_kinfo,
135 };
136 
137 static int 	knote_attach(struct knote *kn, struct kqueue *kq);
138 static void 	knote_drop(struct knote *kn, struct thread *td);
139 static void 	knote_enqueue(struct knote *kn);
140 static void 	knote_dequeue(struct knote *kn);
141 static void 	knote_init(void);
142 static struct 	knote *knote_alloc(int waitok);
143 static void 	knote_free(struct knote *kn);
144 
145 static void	filt_kqdetach(struct knote *kn);
146 static int	filt_kqueue(struct knote *kn, long hint);
147 static int	filt_procattach(struct knote *kn);
148 static void	filt_procdetach(struct knote *kn);
149 static int	filt_proc(struct knote *kn, long hint);
150 static int	filt_fileattach(struct knote *kn);
151 static void	filt_timerexpire(void *knx);
152 static int	filt_timerattach(struct knote *kn);
153 static void	filt_timerdetach(struct knote *kn);
154 static int	filt_timer(struct knote *kn, long hint);
155 static int	filt_userattach(struct knote *kn);
156 static void	filt_userdetach(struct knote *kn);
157 static int	filt_user(struct knote *kn, long hint);
158 static void	filt_usertouch(struct knote *kn, struct kevent *kev,
159 		    u_long type);
160 
161 static struct filterops file_filtops = {
162 	.f_isfd = 1,
163 	.f_attach = filt_fileattach,
164 };
165 static struct filterops kqread_filtops = {
166 	.f_isfd = 1,
167 	.f_detach = filt_kqdetach,
168 	.f_event = filt_kqueue,
169 };
170 /* XXX - move to kern_proc.c?  */
171 static struct filterops proc_filtops = {
172 	.f_isfd = 0,
173 	.f_attach = filt_procattach,
174 	.f_detach = filt_procdetach,
175 	.f_event = filt_proc,
176 };
177 static struct filterops timer_filtops = {
178 	.f_isfd = 0,
179 	.f_attach = filt_timerattach,
180 	.f_detach = filt_timerdetach,
181 	.f_event = filt_timer,
182 };
183 static struct filterops user_filtops = {
184 	.f_attach = filt_userattach,
185 	.f_detach = filt_userdetach,
186 	.f_event = filt_user,
187 	.f_touch = filt_usertouch,
188 };
189 
190 static uma_zone_t	knote_zone;
191 static atomic_uint	kq_ncallouts = ATOMIC_VAR_INIT(0);
192 static unsigned int 	kq_calloutmax = 4 * 1024;
193 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
194     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
195 
196 /* XXX - ensure not KN_INFLUX?? */
197 #define KNOTE_ACTIVATE(kn, islock) do { 				\
198 	if ((islock))							\
199 		mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED);		\
200 	else								\
201 		KQ_LOCK((kn)->kn_kq);					\
202 	(kn)->kn_status |= KN_ACTIVE;					\
203 	if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)		\
204 		knote_enqueue((kn));					\
205 	if (!(islock))							\
206 		KQ_UNLOCK((kn)->kn_kq);					\
207 } while(0)
208 #define KQ_LOCK(kq) do {						\
209 	mtx_lock(&(kq)->kq_lock);					\
210 } while (0)
211 #define KQ_FLUX_WAKEUP(kq) do {						\
212 	if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) {		\
213 		(kq)->kq_state &= ~KQ_FLUXWAIT;				\
214 		wakeup((kq));						\
215 	}								\
216 } while (0)
217 #define KQ_UNLOCK_FLUX(kq) do {						\
218 	KQ_FLUX_WAKEUP(kq);						\
219 	mtx_unlock(&(kq)->kq_lock);					\
220 } while (0)
221 #define KQ_UNLOCK(kq) do {						\
222 	mtx_unlock(&(kq)->kq_lock);					\
223 } while (0)
224 #define KQ_OWNED(kq) do {						\
225 	mtx_assert(&(kq)->kq_lock, MA_OWNED);				\
226 } while (0)
227 #define KQ_NOTOWNED(kq) do {						\
228 	mtx_assert(&(kq)->kq_lock, MA_NOTOWNED);			\
229 } while (0)
230 
231 static struct knlist *
232 kn_list_lock(struct knote *kn)
233 {
234 	struct knlist *knl;
235 
236 	knl = kn->kn_knlist;
237 	if (knl != NULL)
238 		knl->kl_lock(knl->kl_lockarg);
239 	return (knl);
240 }
241 
242 static void
243 kn_list_unlock(struct knlist *knl)
244 {
245 	bool do_free;
246 
247 	if (knl == NULL)
248 		return;
249 	do_free = knl->kl_autodestroy && knlist_empty(knl);
250 	knl->kl_unlock(knl->kl_lockarg);
251 	if (do_free) {
252 		knlist_destroy(knl);
253 		free(knl, M_KQUEUE);
254 	}
255 }
256 
257 #define	KNL_ASSERT_LOCK(knl, islocked) do {				\
258 	if (islocked)							\
259 		KNL_ASSERT_LOCKED(knl);				\
260 	else								\
261 		KNL_ASSERT_UNLOCKED(knl);				\
262 } while (0)
263 #ifdef INVARIANTS
264 #define	KNL_ASSERT_LOCKED(knl) do {					\
265 	knl->kl_assert_locked((knl)->kl_lockarg);			\
266 } while (0)
267 #define	KNL_ASSERT_UNLOCKED(knl) do {					\
268 	knl->kl_assert_unlocked((knl)->kl_lockarg);			\
269 } while (0)
270 #else /* !INVARIANTS */
271 #define	KNL_ASSERT_LOCKED(knl) do {} while(0)
272 #define	KNL_ASSERT_UNLOCKED(knl) do {} while (0)
273 #endif /* INVARIANTS */
274 
275 #ifndef	KN_HASHSIZE
276 #define	KN_HASHSIZE		64		/* XXX should be tunable */
277 #endif
278 
279 #define KN_HASH(val, mask)	(((val) ^ (val >> 8)) & (mask))
280 
281 static int
282 filt_nullattach(struct knote *kn)
283 {
284 
285 	return (ENXIO);
286 };
287 
288 struct filterops null_filtops = {
289 	.f_isfd = 0,
290 	.f_attach = filt_nullattach,
291 };
292 
293 /* XXX - make SYSINIT to add these, and move into respective modules. */
294 extern struct filterops sig_filtops;
295 extern struct filterops fs_filtops;
296 
297 /*
298  * Table for for all system-defined filters.
299  */
300 static struct mtx	filterops_lock;
301 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
302 	MTX_DEF);
303 static struct {
304 	struct filterops *for_fop;
305 	int for_nolock;
306 	int for_refcnt;
307 } sysfilt_ops[EVFILT_SYSCOUNT] = {
308 	{ &file_filtops, 1 },			/* EVFILT_READ */
309 	{ &file_filtops, 1 },			/* EVFILT_WRITE */
310 	{ &null_filtops },			/* EVFILT_AIO */
311 	{ &file_filtops, 1 },			/* EVFILT_VNODE */
312 	{ &proc_filtops, 1 },			/* EVFILT_PROC */
313 	{ &sig_filtops, 1 },			/* EVFILT_SIGNAL */
314 	{ &timer_filtops, 1 },			/* EVFILT_TIMER */
315 	{ &file_filtops, 1 },			/* EVFILT_PROCDESC */
316 	{ &fs_filtops, 1 },			/* EVFILT_FS */
317 	{ &null_filtops },			/* EVFILT_LIO */
318 	{ &user_filtops, 1 },			/* EVFILT_USER */
319 	{ &null_filtops },			/* EVFILT_SENDFILE */
320 };
321 
322 /*
323  * Simple redirection for all cdevsw style objects to call their fo_kqfilter
324  * method.
325  */
326 static int
327 filt_fileattach(struct knote *kn)
328 {
329 
330 	return (fo_kqfilter(kn->kn_fp, kn));
331 }
332 
333 /*ARGSUSED*/
334 static int
335 kqueue_kqfilter(struct file *fp, struct knote *kn)
336 {
337 	struct kqueue *kq = kn->kn_fp->f_data;
338 
339 	if (kn->kn_filter != EVFILT_READ)
340 		return (EINVAL);
341 
342 	kn->kn_status |= KN_KQUEUE;
343 	kn->kn_fop = &kqread_filtops;
344 	knlist_add(&kq->kq_sel.si_note, kn, 0);
345 
346 	return (0);
347 }
348 
349 static void
350 filt_kqdetach(struct knote *kn)
351 {
352 	struct kqueue *kq = kn->kn_fp->f_data;
353 
354 	knlist_remove(&kq->kq_sel.si_note, kn, 0);
355 }
356 
357 /*ARGSUSED*/
358 static int
359 filt_kqueue(struct knote *kn, long hint)
360 {
361 	struct kqueue *kq = kn->kn_fp->f_data;
362 
363 	kn->kn_data = kq->kq_count;
364 	return (kn->kn_data > 0);
365 }
366 
367 /* XXX - move to kern_proc.c?  */
368 static int
369 filt_procattach(struct knote *kn)
370 {
371 	struct proc *p;
372 	int error;
373 	bool exiting, immediate;
374 
375 	exiting = immediate = false;
376 	p = pfind(kn->kn_id);
377 	if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
378 		p = zpfind(kn->kn_id);
379 		exiting = true;
380 	} else if (p != NULL && (p->p_flag & P_WEXIT)) {
381 		exiting = true;
382 	}
383 
384 	if (p == NULL)
385 		return (ESRCH);
386 	if ((error = p_cansee(curthread, p))) {
387 		PROC_UNLOCK(p);
388 		return (error);
389 	}
390 
391 	kn->kn_ptr.p_proc = p;
392 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
393 
394 	/*
395 	 * Internal flag indicating registration done by kernel for the
396 	 * purposes of getting a NOTE_CHILD notification.
397 	 */
398 	if (kn->kn_flags & EV_FLAG2) {
399 		kn->kn_flags &= ~EV_FLAG2;
400 		kn->kn_data = kn->kn_sdata;		/* ppid */
401 		kn->kn_fflags = NOTE_CHILD;
402 		kn->kn_sfflags &= ~(NOTE_EXIT | NOTE_EXEC | NOTE_FORK);
403 		immediate = true; /* Force immediate activation of child note. */
404 	}
405 	/*
406 	 * Internal flag indicating registration done by kernel (for other than
407 	 * NOTE_CHILD).
408 	 */
409 	if (kn->kn_flags & EV_FLAG1) {
410 		kn->kn_flags &= ~EV_FLAG1;
411 	}
412 
413 	knlist_add(p->p_klist, kn, 1);
414 
415 	/*
416 	 * Immediately activate any child notes or, in the case of a zombie
417 	 * target process, exit notes.  The latter is necessary to handle the
418 	 * case where the target process, e.g. a child, dies before the kevent
419 	 * is registered.
420 	 */
421 	if (immediate || (exiting && filt_proc(kn, NOTE_EXIT)))
422 		KNOTE_ACTIVATE(kn, 0);
423 
424 	PROC_UNLOCK(p);
425 
426 	return (0);
427 }
428 
429 /*
430  * The knote may be attached to a different process, which may exit,
431  * leaving nothing for the knote to be attached to.  So when the process
432  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
433  * it will be deleted when read out.  However, as part of the knote deletion,
434  * this routine is called, so a check is needed to avoid actually performing
435  * a detach, because the original process does not exist any more.
436  */
437 /* XXX - move to kern_proc.c?  */
438 static void
439 filt_procdetach(struct knote *kn)
440 {
441 
442 	knlist_remove(kn->kn_knlist, kn, 0);
443 	kn->kn_ptr.p_proc = NULL;
444 }
445 
446 /* XXX - move to kern_proc.c?  */
447 static int
448 filt_proc(struct knote *kn, long hint)
449 {
450 	struct proc *p;
451 	u_int event;
452 
453 	p = kn->kn_ptr.p_proc;
454 	if (p == NULL) /* already activated, from attach filter */
455 		return (0);
456 
457 	/* Mask off extra data. */
458 	event = (u_int)hint & NOTE_PCTRLMASK;
459 
460 	/* If the user is interested in this event, record it. */
461 	if (kn->kn_sfflags & event)
462 		kn->kn_fflags |= event;
463 
464 	/* Process is gone, so flag the event as finished. */
465 	if (event == NOTE_EXIT) {
466 		kn->kn_flags |= EV_EOF | EV_ONESHOT;
467 		kn->kn_ptr.p_proc = NULL;
468 		if (kn->kn_fflags & NOTE_EXIT)
469 			kn->kn_data = KW_EXITCODE(p->p_xexit, p->p_xsig);
470 		if (kn->kn_fflags == 0)
471 			kn->kn_flags |= EV_DROP;
472 		return (1);
473 	}
474 
475 	return (kn->kn_fflags != 0);
476 }
477 
478 /*
479  * Called when the process forked. It mostly does the same as the
480  * knote(), activating all knotes registered to be activated when the
481  * process forked. Additionally, for each knote attached to the
482  * parent, check whether user wants to track the new process. If so
483  * attach a new knote to it, and immediately report an event with the
484  * child's pid.
485  */
486 void
487 knote_fork(struct knlist *list, int pid)
488 {
489 	struct kqueue *kq;
490 	struct knote *kn;
491 	struct kevent kev;
492 	int error;
493 
494 	if (list == NULL)
495 		return;
496 	list->kl_lock(list->kl_lockarg);
497 
498 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
499 		kq = kn->kn_kq;
500 		KQ_LOCK(kq);
501 		if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
502 			KQ_UNLOCK(kq);
503 			continue;
504 		}
505 
506 		/*
507 		 * The same as knote(), activate the event.
508 		 */
509 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
510 			kn->kn_status |= KN_HASKQLOCK;
511 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
512 				KNOTE_ACTIVATE(kn, 1);
513 			kn->kn_status &= ~KN_HASKQLOCK;
514 			KQ_UNLOCK(kq);
515 			continue;
516 		}
517 
518 		/*
519 		 * The NOTE_TRACK case. In addition to the activation
520 		 * of the event, we need to register new events to
521 		 * track the child. Drop the locks in preparation for
522 		 * the call to kqueue_register().
523 		 */
524 		kn->kn_status |= KN_INFLUX;
525 		KQ_UNLOCK(kq);
526 		list->kl_unlock(list->kl_lockarg);
527 
528 		/*
529 		 * Activate existing knote and register tracking knotes with
530 		 * new process.
531 		 *
532 		 * First register a knote to get just the child notice. This
533 		 * must be a separate note from a potential NOTE_EXIT
534 		 * notification since both NOTE_CHILD and NOTE_EXIT are defined
535 		 * to use the data field (in conflicting ways).
536 		 */
537 		kev.ident = pid;
538 		kev.filter = kn->kn_filter;
539 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_ONESHOT |
540 		    EV_FLAG2;
541 		kev.fflags = kn->kn_sfflags;
542 		kev.data = kn->kn_id;		/* parent */
543 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
544 		error = kqueue_register(kq, &kev, NULL, 0);
545 		if (error)
546 			kn->kn_fflags |= NOTE_TRACKERR;
547 
548 		/*
549 		 * Then register another knote to track other potential events
550 		 * from the new process.
551 		 */
552 		kev.ident = pid;
553 		kev.filter = kn->kn_filter;
554 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
555 		kev.fflags = kn->kn_sfflags;
556 		kev.data = kn->kn_id;		/* parent */
557 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
558 		error = kqueue_register(kq, &kev, NULL, 0);
559 		if (error)
560 			kn->kn_fflags |= NOTE_TRACKERR;
561 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
562 			KNOTE_ACTIVATE(kn, 0);
563 		KQ_LOCK(kq);
564 		kn->kn_status &= ~KN_INFLUX;
565 		KQ_UNLOCK_FLUX(kq);
566 		list->kl_lock(list->kl_lockarg);
567 	}
568 	list->kl_unlock(list->kl_lockarg);
569 }
570 
571 /*
572  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
573  * interval timer support code.
574  */
575 
576 #define NOTE_TIMER_PRECMASK	(NOTE_SECONDS|NOTE_MSECONDS|NOTE_USECONDS| \
577 				NOTE_NSECONDS)
578 
579 static sbintime_t
580 timer2sbintime(intptr_t data, int flags)
581 {
582 
583         /*
584          * Macros for converting to the fractional second portion of an
585          * sbintime_t using 64bit multiplication to improve precision.
586          */
587 #define NS_TO_SBT(ns) (((ns) * (((uint64_t)1 << 63) / 500000000)) >> 32)
588 #define US_TO_SBT(us) (((us) * (((uint64_t)1 << 63) / 500000)) >> 32)
589 #define MS_TO_SBT(ms) (((ms) * (((uint64_t)1 << 63) / 500)) >> 32)
590 	switch (flags & NOTE_TIMER_PRECMASK) {
591 	case NOTE_SECONDS:
592 #ifdef __LP64__
593 		if (data > (SBT_MAX / SBT_1S))
594 			return SBT_MAX;
595 #endif
596 		return ((sbintime_t)data << 32);
597 	case NOTE_MSECONDS: /* FALLTHROUGH */
598 	case 0:
599 		if (data >= 1000) {
600 			int64_t secs = data / 1000;
601 #ifdef __LP64__
602 			if (secs > (SBT_MAX / SBT_1S))
603 				return SBT_MAX;
604 #endif
605 			return (secs << 32 | MS_TO_SBT(data % 1000));
606 		}
607 		return MS_TO_SBT(data);
608 	case NOTE_USECONDS:
609 		if (data >= 1000000) {
610 			int64_t secs = data / 1000000;
611 #ifdef __LP64__
612 			if (secs > (SBT_MAX / SBT_1S))
613 				return SBT_MAX;
614 #endif
615 			return (secs << 32 | US_TO_SBT(data % 1000000));
616 		}
617 		return US_TO_SBT(data);
618 	case NOTE_NSECONDS:
619 		if (data >= 1000000000) {
620 			int64_t secs = data / 1000000000;
621 #ifdef __LP64__
622 			if (secs > (SBT_MAX / SBT_1S))
623 				return SBT_MAX;
624 #endif
625 			return (secs << 32 | US_TO_SBT(data % 1000000000));
626 		}
627 		return NS_TO_SBT(data);
628 	default:
629 		break;
630 	}
631 	return (-1);
632 }
633 
634 static void
635 filt_timerexpire(void *knx)
636 {
637 	struct callout *calloutp;
638 	struct knote *kn;
639 
640 	kn = knx;
641 	kn->kn_data++;
642 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
643 
644 	if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
645 		calloutp = (struct callout *)kn->kn_hook;
646 		*kn->kn_ptr.p_nexttime += timer2sbintime(kn->kn_sdata,
647 		    kn->kn_sfflags);
648 		callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
649 		    filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
650 	}
651 }
652 
653 /*
654  * data contains amount of time to sleep
655  */
656 static int
657 filt_timerattach(struct knote *kn)
658 {
659 	struct callout *calloutp;
660 	sbintime_t to;
661 	unsigned int ncallouts;
662 
663 	if ((intptr_t)kn->kn_sdata < 0)
664 		return (EINVAL);
665 	if ((intptr_t)kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
666 		kn->kn_sdata = 1;
667 	/* Only precision unit are supported in flags so far */
668 	if (kn->kn_sfflags & ~NOTE_TIMER_PRECMASK)
669 		return (EINVAL);
670 
671 	to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
672 	if (to < 0)
673 		return (EINVAL);
674 
675 	ncallouts = atomic_load_explicit(&kq_ncallouts, memory_order_relaxed);
676 	do {
677 		if (ncallouts >= kq_calloutmax)
678 			return (ENOMEM);
679 	} while (!atomic_compare_exchange_weak_explicit(&kq_ncallouts,
680 	    &ncallouts, ncallouts + 1, memory_order_relaxed,
681 	    memory_order_relaxed));
682 
683 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
684 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
685 	kn->kn_ptr.p_nexttime = malloc(sizeof(sbintime_t), M_KQUEUE, M_WAITOK);
686 	calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
687 	callout_init(calloutp, 1);
688 	kn->kn_hook = calloutp;
689 	*kn->kn_ptr.p_nexttime = to + sbinuptime();
690 	callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
691 	    filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
692 
693 	return (0);
694 }
695 
696 static void
697 filt_timerdetach(struct knote *kn)
698 {
699 	struct callout *calloutp;
700 	unsigned int old;
701 
702 	calloutp = (struct callout *)kn->kn_hook;
703 	callout_drain(calloutp);
704 	free(calloutp, M_KQUEUE);
705 	free(kn->kn_ptr.p_nexttime, M_KQUEUE);
706 	old = atomic_fetch_sub_explicit(&kq_ncallouts, 1, memory_order_relaxed);
707 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
708 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
709 }
710 
711 static int
712 filt_timer(struct knote *kn, long hint)
713 {
714 
715 	return (kn->kn_data != 0);
716 }
717 
718 static int
719 filt_userattach(struct knote *kn)
720 {
721 
722 	/*
723 	 * EVFILT_USER knotes are not attached to anything in the kernel.
724 	 */
725 	kn->kn_hook = NULL;
726 	if (kn->kn_fflags & NOTE_TRIGGER)
727 		kn->kn_hookid = 1;
728 	else
729 		kn->kn_hookid = 0;
730 	return (0);
731 }
732 
733 static void
734 filt_userdetach(__unused struct knote *kn)
735 {
736 
737 	/*
738 	 * EVFILT_USER knotes are not attached to anything in the kernel.
739 	 */
740 }
741 
742 static int
743 filt_user(struct knote *kn, __unused long hint)
744 {
745 
746 	return (kn->kn_hookid);
747 }
748 
749 static void
750 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
751 {
752 	u_int ffctrl;
753 
754 	switch (type) {
755 	case EVENT_REGISTER:
756 		if (kev->fflags & NOTE_TRIGGER)
757 			kn->kn_hookid = 1;
758 
759 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
760 		kev->fflags &= NOTE_FFLAGSMASK;
761 		switch (ffctrl) {
762 		case NOTE_FFNOP:
763 			break;
764 
765 		case NOTE_FFAND:
766 			kn->kn_sfflags &= kev->fflags;
767 			break;
768 
769 		case NOTE_FFOR:
770 			kn->kn_sfflags |= kev->fflags;
771 			break;
772 
773 		case NOTE_FFCOPY:
774 			kn->kn_sfflags = kev->fflags;
775 			break;
776 
777 		default:
778 			/* XXX Return error? */
779 			break;
780 		}
781 		kn->kn_sdata = kev->data;
782 		if (kev->flags & EV_CLEAR) {
783 			kn->kn_hookid = 0;
784 			kn->kn_data = 0;
785 			kn->kn_fflags = 0;
786 		}
787 		break;
788 
789         case EVENT_PROCESS:
790 		*kev = kn->kn_kevent;
791 		kev->fflags = kn->kn_sfflags;
792 		kev->data = kn->kn_sdata;
793 		if (kn->kn_flags & EV_CLEAR) {
794 			kn->kn_hookid = 0;
795 			kn->kn_data = 0;
796 			kn->kn_fflags = 0;
797 		}
798 		break;
799 
800 	default:
801 		panic("filt_usertouch() - invalid type (%ld)", type);
802 		break;
803 	}
804 }
805 
806 int
807 sys_kqueue(struct thread *td, struct kqueue_args *uap)
808 {
809 
810 	return (kern_kqueue(td, 0, NULL));
811 }
812 
813 static void
814 kqueue_init(struct kqueue *kq)
815 {
816 
817 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
818 	TAILQ_INIT(&kq->kq_head);
819 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
820 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
821 }
822 
823 int
824 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
825 {
826 	struct filedesc *fdp;
827 	struct kqueue *kq;
828 	struct file *fp;
829 	struct ucred *cred;
830 	int fd, error;
831 
832 	fdp = td->td_proc->p_fd;
833 	cred = td->td_ucred;
834 	if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
835 		return (ENOMEM);
836 
837 	error = falloc_caps(td, &fp, &fd, flags, fcaps);
838 	if (error != 0) {
839 		chgkqcnt(cred->cr_ruidinfo, -1, 0);
840 		return (error);
841 	}
842 
843 	/* An extra reference on `fp' has been held for us by falloc(). */
844 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
845 	kqueue_init(kq);
846 	kq->kq_fdp = fdp;
847 	kq->kq_cred = crhold(cred);
848 
849 	FILEDESC_XLOCK(fdp);
850 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
851 	FILEDESC_XUNLOCK(fdp);
852 
853 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
854 	fdrop(fp, td);
855 
856 	td->td_retval[0] = fd;
857 	return (0);
858 }
859 
860 #ifndef _SYS_SYSPROTO_H_
861 struct kevent_args {
862 	int	fd;
863 	const struct kevent *changelist;
864 	int	nchanges;
865 	struct	kevent *eventlist;
866 	int	nevents;
867 	const struct timespec *timeout;
868 };
869 #endif
870 int
871 sys_kevent(struct thread *td, struct kevent_args *uap)
872 {
873 	struct timespec ts, *tsp;
874 	struct kevent_copyops k_ops = { uap,
875 					kevent_copyout,
876 					kevent_copyin};
877 	int error;
878 #ifdef KTRACE
879 	struct uio ktruio;
880 	struct iovec ktriov;
881 	struct uio *ktruioin = NULL;
882 	struct uio *ktruioout = NULL;
883 #endif
884 
885 	if (uap->timeout != NULL) {
886 		error = copyin(uap->timeout, &ts, sizeof(ts));
887 		if (error)
888 			return (error);
889 		tsp = &ts;
890 	} else
891 		tsp = NULL;
892 
893 #ifdef KTRACE
894 	if (KTRPOINT(td, KTR_GENIO)) {
895 		ktriov.iov_base = uap->changelist;
896 		ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
897 		ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
898 		    .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
899 		    .uio_td = td };
900 		ktruioin = cloneuio(&ktruio);
901 		ktriov.iov_base = uap->eventlist;
902 		ktriov.iov_len = uap->nevents * sizeof(struct kevent);
903 		ktruioout = cloneuio(&ktruio);
904 	}
905 #endif
906 
907 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
908 	    &k_ops, tsp);
909 
910 #ifdef KTRACE
911 	if (ktruioin != NULL) {
912 		ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
913 		ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
914 		ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
915 		ktrgenio(uap->fd, UIO_READ, ktruioout, error);
916 	}
917 #endif
918 
919 	return (error);
920 }
921 
922 /*
923  * Copy 'count' items into the destination list pointed to by uap->eventlist.
924  */
925 static int
926 kevent_copyout(void *arg, struct kevent *kevp, int count)
927 {
928 	struct kevent_args *uap;
929 	int error;
930 
931 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
932 	uap = (struct kevent_args *)arg;
933 
934 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
935 	if (error == 0)
936 		uap->eventlist += count;
937 	return (error);
938 }
939 
940 /*
941  * Copy 'count' items from the list pointed to by uap->changelist.
942  */
943 static int
944 kevent_copyin(void *arg, struct kevent *kevp, int count)
945 {
946 	struct kevent_args *uap;
947 	int error;
948 
949 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
950 	uap = (struct kevent_args *)arg;
951 
952 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
953 	if (error == 0)
954 		uap->changelist += count;
955 	return (error);
956 }
957 
958 int
959 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
960     struct kevent_copyops *k_ops, const struct timespec *timeout)
961 {
962 	cap_rights_t rights;
963 	struct file *fp;
964 	int error;
965 
966 	cap_rights_init(&rights);
967 	if (nchanges > 0)
968 		cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
969 	if (nevents > 0)
970 		cap_rights_set(&rights, CAP_KQUEUE_EVENT);
971 	error = fget(td, fd, &rights, &fp);
972 	if (error != 0)
973 		return (error);
974 
975 	error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
976 	fdrop(fp, td);
977 
978 	return (error);
979 }
980 
981 static int
982 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
983     struct kevent_copyops *k_ops, const struct timespec *timeout)
984 {
985 	struct kevent keva[KQ_NEVENTS];
986 	struct kevent *kevp, *changes;
987 	int i, n, nerrors, error;
988 
989 	nerrors = 0;
990 	while (nchanges > 0) {
991 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
992 		error = k_ops->k_copyin(k_ops->arg, keva, n);
993 		if (error)
994 			return (error);
995 		changes = keva;
996 		for (i = 0; i < n; i++) {
997 			kevp = &changes[i];
998 			if (!kevp->filter)
999 				continue;
1000 			kevp->flags &= ~EV_SYSFLAGS;
1001 			error = kqueue_register(kq, kevp, td, 1);
1002 			if (error || (kevp->flags & EV_RECEIPT)) {
1003 				if (nevents == 0)
1004 					return (error);
1005 				kevp->flags = EV_ERROR;
1006 				kevp->data = error;
1007 				(void)k_ops->k_copyout(k_ops->arg, kevp, 1);
1008 				nevents--;
1009 				nerrors++;
1010 			}
1011 		}
1012 		nchanges -= n;
1013 	}
1014 	if (nerrors) {
1015 		td->td_retval[0] = nerrors;
1016 		return (0);
1017 	}
1018 
1019 	return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
1020 }
1021 
1022 int
1023 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
1024     struct kevent_copyops *k_ops, const struct timespec *timeout)
1025 {
1026 	struct kqueue *kq;
1027 	int error;
1028 
1029 	error = kqueue_acquire(fp, &kq);
1030 	if (error != 0)
1031 		return (error);
1032 	error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
1033 	kqueue_release(kq, 0);
1034 	return (error);
1035 }
1036 
1037 /*
1038  * Performs a kevent() call on a temporarily created kqueue. This can be
1039  * used to perform one-shot polling, similar to poll() and select().
1040  */
1041 int
1042 kern_kevent_anonymous(struct thread *td, int nevents,
1043     struct kevent_copyops *k_ops)
1044 {
1045 	struct kqueue kq = {};
1046 	int error;
1047 
1048 	kqueue_init(&kq);
1049 	kq.kq_refcnt = 1;
1050 	error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
1051 	kqueue_drain(&kq, td);
1052 	kqueue_destroy(&kq);
1053 	return (error);
1054 }
1055 
1056 int
1057 kqueue_add_filteropts(int filt, struct filterops *filtops)
1058 {
1059 	int error;
1060 
1061 	error = 0;
1062 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1063 		printf(
1064 "trying to add a filterop that is out of range: %d is beyond %d\n",
1065 		    ~filt, EVFILT_SYSCOUNT);
1066 		return EINVAL;
1067 	}
1068 	mtx_lock(&filterops_lock);
1069 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1070 	    sysfilt_ops[~filt].for_fop != NULL)
1071 		error = EEXIST;
1072 	else {
1073 		sysfilt_ops[~filt].for_fop = filtops;
1074 		sysfilt_ops[~filt].for_refcnt = 0;
1075 	}
1076 	mtx_unlock(&filterops_lock);
1077 
1078 	return (error);
1079 }
1080 
1081 int
1082 kqueue_del_filteropts(int filt)
1083 {
1084 	int error;
1085 
1086 	error = 0;
1087 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1088 		return EINVAL;
1089 
1090 	mtx_lock(&filterops_lock);
1091 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1092 	    sysfilt_ops[~filt].for_fop == NULL)
1093 		error = EINVAL;
1094 	else if (sysfilt_ops[~filt].for_refcnt != 0)
1095 		error = EBUSY;
1096 	else {
1097 		sysfilt_ops[~filt].for_fop = &null_filtops;
1098 		sysfilt_ops[~filt].for_refcnt = 0;
1099 	}
1100 	mtx_unlock(&filterops_lock);
1101 
1102 	return error;
1103 }
1104 
1105 static struct filterops *
1106 kqueue_fo_find(int filt)
1107 {
1108 
1109 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1110 		return NULL;
1111 
1112 	if (sysfilt_ops[~filt].for_nolock)
1113 		return sysfilt_ops[~filt].for_fop;
1114 
1115 	mtx_lock(&filterops_lock);
1116 	sysfilt_ops[~filt].for_refcnt++;
1117 	if (sysfilt_ops[~filt].for_fop == NULL)
1118 		sysfilt_ops[~filt].for_fop = &null_filtops;
1119 	mtx_unlock(&filterops_lock);
1120 
1121 	return sysfilt_ops[~filt].for_fop;
1122 }
1123 
1124 static void
1125 kqueue_fo_release(int filt)
1126 {
1127 
1128 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1129 		return;
1130 
1131 	if (sysfilt_ops[~filt].for_nolock)
1132 		return;
1133 
1134 	mtx_lock(&filterops_lock);
1135 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1136 	    ("filter object refcount not valid on release"));
1137 	sysfilt_ops[~filt].for_refcnt--;
1138 	mtx_unlock(&filterops_lock);
1139 }
1140 
1141 /*
1142  * A ref to kq (obtained via kqueue_acquire) must be held.  waitok will
1143  * influence if memory allocation should wait.  Make sure it is 0 if you
1144  * hold any mutexes.
1145  */
1146 static int
1147 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1148 {
1149 	struct filterops *fops;
1150 	struct file *fp;
1151 	struct knote *kn, *tkn;
1152 	struct knlist *knl;
1153 	cap_rights_t rights;
1154 	int error, filt, event;
1155 	int haskqglobal, filedesc_unlock;
1156 
1157 	if ((kev->flags & (EV_ENABLE | EV_DISABLE)) == (EV_ENABLE | EV_DISABLE))
1158 		return (EINVAL);
1159 
1160 	fp = NULL;
1161 	kn = NULL;
1162 	knl = NULL;
1163 	error = 0;
1164 	haskqglobal = 0;
1165 	filedesc_unlock = 0;
1166 
1167 	filt = kev->filter;
1168 	fops = kqueue_fo_find(filt);
1169 	if (fops == NULL)
1170 		return EINVAL;
1171 
1172 	if (kev->flags & EV_ADD) {
1173 		/*
1174 		 * Prevent waiting with locks.  Non-sleepable
1175 		 * allocation failures are handled in the loop, only
1176 		 * if the spare knote appears to be actually required.
1177 		 */
1178 		tkn = knote_alloc(waitok);
1179 	} else {
1180 		tkn = NULL;
1181 	}
1182 
1183 findkn:
1184 	if (fops->f_isfd) {
1185 		KASSERT(td != NULL, ("td is NULL"));
1186 		if (kev->ident > INT_MAX)
1187 			error = EBADF;
1188 		else
1189 			error = fget(td, kev->ident,
1190 			    cap_rights_init(&rights, CAP_EVENT), &fp);
1191 		if (error)
1192 			goto done;
1193 
1194 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1195 		    kev->ident, 0) != 0) {
1196 			/* try again */
1197 			fdrop(fp, td);
1198 			fp = NULL;
1199 			error = kqueue_expand(kq, fops, kev->ident, waitok);
1200 			if (error)
1201 				goto done;
1202 			goto findkn;
1203 		}
1204 
1205 		if (fp->f_type == DTYPE_KQUEUE) {
1206 			/*
1207 			 * If we add some intelligence about what we are doing,
1208 			 * we should be able to support events on ourselves.
1209 			 * We need to know when we are doing this to prevent
1210 			 * getting both the knlist lock and the kq lock since
1211 			 * they are the same thing.
1212 			 */
1213 			if (fp->f_data == kq) {
1214 				error = EINVAL;
1215 				goto done;
1216 			}
1217 
1218 			/*
1219 			 * Pre-lock the filedesc before the global
1220 			 * lock mutex, see the comment in
1221 			 * kqueue_close().
1222 			 */
1223 			FILEDESC_XLOCK(td->td_proc->p_fd);
1224 			filedesc_unlock = 1;
1225 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1226 		}
1227 
1228 		KQ_LOCK(kq);
1229 		if (kev->ident < kq->kq_knlistsize) {
1230 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1231 				if (kev->filter == kn->kn_filter)
1232 					break;
1233 		}
1234 	} else {
1235 		if ((kev->flags & EV_ADD) == EV_ADD)
1236 			kqueue_expand(kq, fops, kev->ident, waitok);
1237 
1238 		KQ_LOCK(kq);
1239 
1240 		/*
1241 		 * If possible, find an existing knote to use for this kevent.
1242 		 */
1243 		if (kev->filter == EVFILT_PROC &&
1244 		    (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1245 			/* This is an internal creation of a process tracking
1246 			 * note. Don't attempt to coalesce this with an
1247 			 * existing note.
1248 			 */
1249 			;
1250 		} else if (kq->kq_knhashmask != 0) {
1251 			struct klist *list;
1252 
1253 			list = &kq->kq_knhash[
1254 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1255 			SLIST_FOREACH(kn, list, kn_link)
1256 				if (kev->ident == kn->kn_id &&
1257 				    kev->filter == kn->kn_filter)
1258 					break;
1259 		}
1260 	}
1261 
1262 	/* knote is in the process of changing, wait for it to stabilize. */
1263 	if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1264 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1265 		if (filedesc_unlock) {
1266 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1267 			filedesc_unlock = 0;
1268 		}
1269 		kq->kq_state |= KQ_FLUXWAIT;
1270 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1271 		if (fp != NULL) {
1272 			fdrop(fp, td);
1273 			fp = NULL;
1274 		}
1275 		goto findkn;
1276 	}
1277 
1278 	/*
1279 	 * kn now contains the matching knote, or NULL if no match
1280 	 */
1281 	if (kn == NULL) {
1282 		if (kev->flags & EV_ADD) {
1283 			kn = tkn;
1284 			tkn = NULL;
1285 			if (kn == NULL) {
1286 				KQ_UNLOCK(kq);
1287 				error = ENOMEM;
1288 				goto done;
1289 			}
1290 			kn->kn_fp = fp;
1291 			kn->kn_kq = kq;
1292 			kn->kn_fop = fops;
1293 			/*
1294 			 * apply reference counts to knote structure, and
1295 			 * do not release it at the end of this routine.
1296 			 */
1297 			fops = NULL;
1298 			fp = NULL;
1299 
1300 			kn->kn_sfflags = kev->fflags;
1301 			kn->kn_sdata = kev->data;
1302 			kev->fflags = 0;
1303 			kev->data = 0;
1304 			kn->kn_kevent = *kev;
1305 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1306 			    EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1307 			kn->kn_status = KN_INFLUX|KN_DETACHED;
1308 
1309 			error = knote_attach(kn, kq);
1310 			KQ_UNLOCK(kq);
1311 			if (error != 0) {
1312 				tkn = kn;
1313 				goto done;
1314 			}
1315 
1316 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1317 				knote_drop(kn, td);
1318 				goto done;
1319 			}
1320 			knl = kn_list_lock(kn);
1321 			goto done_ev_add;
1322 		} else {
1323 			/* No matching knote and the EV_ADD flag is not set. */
1324 			KQ_UNLOCK(kq);
1325 			error = ENOENT;
1326 			goto done;
1327 		}
1328 	}
1329 
1330 	if (kev->flags & EV_DELETE) {
1331 		kn->kn_status |= KN_INFLUX;
1332 		KQ_UNLOCK(kq);
1333 		if (!(kn->kn_status & KN_DETACHED))
1334 			kn->kn_fop->f_detach(kn);
1335 		knote_drop(kn, td);
1336 		goto done;
1337 	}
1338 
1339 	if (kev->flags & EV_FORCEONESHOT) {
1340 		kn->kn_flags |= EV_ONESHOT;
1341 		KNOTE_ACTIVATE(kn, 1);
1342 	}
1343 
1344 	/*
1345 	 * The user may change some filter values after the initial EV_ADD,
1346 	 * but doing so will not reset any filter which has already been
1347 	 * triggered.
1348 	 */
1349 	kn->kn_status |= KN_INFLUX | KN_SCAN;
1350 	KQ_UNLOCK(kq);
1351 	knl = kn_list_lock(kn);
1352 	kn->kn_kevent.udata = kev->udata;
1353 	if (!fops->f_isfd && fops->f_touch != NULL) {
1354 		fops->f_touch(kn, kev, EVENT_REGISTER);
1355 	} else {
1356 		kn->kn_sfflags = kev->fflags;
1357 		kn->kn_sdata = kev->data;
1358 	}
1359 
1360 	/*
1361 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1362 	 * the initial attach event decides that the event is "completed"
1363 	 * already.  i.e. filt_procattach is called on a zombie process.  It
1364 	 * will call filt_proc which will remove it from the list, and NULL
1365 	 * kn_knlist.
1366 	 */
1367 done_ev_add:
1368 	if ((kev->flags & EV_ENABLE) != 0)
1369 		kn->kn_status &= ~KN_DISABLED;
1370 	else if ((kev->flags & EV_DISABLE) != 0)
1371 		kn->kn_status |= KN_DISABLED;
1372 
1373 	if ((kn->kn_status & KN_DISABLED) == 0)
1374 		event = kn->kn_fop->f_event(kn, 0);
1375 	else
1376 		event = 0;
1377 
1378 	KQ_LOCK(kq);
1379 	if (event)
1380 		kn->kn_status |= KN_ACTIVE;
1381 	if ((kn->kn_status & (KN_ACTIVE | KN_DISABLED | KN_QUEUED)) ==
1382 	    KN_ACTIVE)
1383 		knote_enqueue(kn);
1384 	kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1385 	kn_list_unlock(knl);
1386 	KQ_UNLOCK_FLUX(kq);
1387 
1388 done:
1389 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1390 	if (filedesc_unlock)
1391 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1392 	if (fp != NULL)
1393 		fdrop(fp, td);
1394 	knote_free(tkn);
1395 	if (fops != NULL)
1396 		kqueue_fo_release(filt);
1397 	return (error);
1398 }
1399 
1400 static int
1401 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1402 {
1403 	int error;
1404 	struct kqueue *kq;
1405 
1406 	error = 0;
1407 
1408 	kq = fp->f_data;
1409 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1410 		return (EBADF);
1411 	*kqp = kq;
1412 	KQ_LOCK(kq);
1413 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1414 		KQ_UNLOCK(kq);
1415 		return (EBADF);
1416 	}
1417 	kq->kq_refcnt++;
1418 	KQ_UNLOCK(kq);
1419 
1420 	return error;
1421 }
1422 
1423 static void
1424 kqueue_release(struct kqueue *kq, int locked)
1425 {
1426 	if (locked)
1427 		KQ_OWNED(kq);
1428 	else
1429 		KQ_LOCK(kq);
1430 	kq->kq_refcnt--;
1431 	if (kq->kq_refcnt == 1)
1432 		wakeup(&kq->kq_refcnt);
1433 	if (!locked)
1434 		KQ_UNLOCK(kq);
1435 }
1436 
1437 static void
1438 kqueue_schedtask(struct kqueue *kq)
1439 {
1440 
1441 	KQ_OWNED(kq);
1442 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1443 	    ("scheduling kqueue task while draining"));
1444 
1445 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1446 		taskqueue_enqueue(taskqueue_kqueue_ctx, &kq->kq_task);
1447 		kq->kq_state |= KQ_TASKSCHED;
1448 	}
1449 }
1450 
1451 /*
1452  * Expand the kq to make sure we have storage for fops/ident pair.
1453  *
1454  * Return 0 on success (or no work necessary), return errno on failure.
1455  *
1456  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1457  * If kqueue_register is called from a non-fd context, there usually/should
1458  * be no locks held.
1459  */
1460 static int
1461 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1462 	int waitok)
1463 {
1464 	struct klist *list, *tmp_knhash, *to_free;
1465 	u_long tmp_knhashmask;
1466 	int size;
1467 	int fd;
1468 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
1469 
1470 	KQ_NOTOWNED(kq);
1471 
1472 	to_free = NULL;
1473 	if (fops->f_isfd) {
1474 		fd = ident;
1475 		if (kq->kq_knlistsize <= fd) {
1476 			size = kq->kq_knlistsize;
1477 			while (size <= fd)
1478 				size += KQEXTENT;
1479 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1480 			if (list == NULL)
1481 				return ENOMEM;
1482 			KQ_LOCK(kq);
1483 			if (kq->kq_knlistsize > fd) {
1484 				to_free = list;
1485 				list = NULL;
1486 			} else {
1487 				if (kq->kq_knlist != NULL) {
1488 					bcopy(kq->kq_knlist, list,
1489 					    kq->kq_knlistsize * sizeof(*list));
1490 					to_free = kq->kq_knlist;
1491 					kq->kq_knlist = NULL;
1492 				}
1493 				bzero((caddr_t)list +
1494 				    kq->kq_knlistsize * sizeof(*list),
1495 				    (size - kq->kq_knlistsize) * sizeof(*list));
1496 				kq->kq_knlistsize = size;
1497 				kq->kq_knlist = list;
1498 			}
1499 			KQ_UNLOCK(kq);
1500 		}
1501 	} else {
1502 		if (kq->kq_knhashmask == 0) {
1503 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1504 			    &tmp_knhashmask);
1505 			if (tmp_knhash == NULL)
1506 				return ENOMEM;
1507 			KQ_LOCK(kq);
1508 			if (kq->kq_knhashmask == 0) {
1509 				kq->kq_knhash = tmp_knhash;
1510 				kq->kq_knhashmask = tmp_knhashmask;
1511 			} else {
1512 				to_free = tmp_knhash;
1513 			}
1514 			KQ_UNLOCK(kq);
1515 		}
1516 	}
1517 	free(to_free, M_KQUEUE);
1518 
1519 	KQ_NOTOWNED(kq);
1520 	return 0;
1521 }
1522 
1523 static void
1524 kqueue_task(void *arg, int pending)
1525 {
1526 	struct kqueue *kq;
1527 	int haskqglobal;
1528 
1529 	haskqglobal = 0;
1530 	kq = arg;
1531 
1532 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1533 	KQ_LOCK(kq);
1534 
1535 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1536 
1537 	kq->kq_state &= ~KQ_TASKSCHED;
1538 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1539 		wakeup(&kq->kq_state);
1540 	}
1541 	KQ_UNLOCK(kq);
1542 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1543 }
1544 
1545 /*
1546  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1547  * We treat KN_MARKER knotes as if they are INFLUX.
1548  */
1549 static int
1550 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1551     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1552 {
1553 	struct kevent *kevp;
1554 	struct knote *kn, *marker;
1555 	struct knlist *knl;
1556 	sbintime_t asbt, rsbt;
1557 	int count, error, haskqglobal, influx, nkev, touch;
1558 
1559 	count = maxevents;
1560 	nkev = 0;
1561 	error = 0;
1562 	haskqglobal = 0;
1563 
1564 	if (maxevents == 0)
1565 		goto done_nl;
1566 
1567 	rsbt = 0;
1568 	if (tsp != NULL) {
1569 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1570 		    tsp->tv_nsec >= 1000000000) {
1571 			error = EINVAL;
1572 			goto done_nl;
1573 		}
1574 		if (timespecisset(tsp)) {
1575 			if (tsp->tv_sec <= INT32_MAX) {
1576 				rsbt = tstosbt(*tsp);
1577 				if (TIMESEL(&asbt, rsbt))
1578 					asbt += tc_tick_sbt;
1579 				if (asbt <= SBT_MAX - rsbt)
1580 					asbt += rsbt;
1581 				else
1582 					asbt = 0;
1583 				rsbt >>= tc_precexp;
1584 			} else
1585 				asbt = 0;
1586 		} else
1587 			asbt = -1;
1588 	} else
1589 		asbt = 0;
1590 	marker = knote_alloc(1);
1591 	marker->kn_status = KN_MARKER;
1592 	KQ_LOCK(kq);
1593 
1594 retry:
1595 	kevp = keva;
1596 	if (kq->kq_count == 0) {
1597 		if (asbt == -1) {
1598 			error = EWOULDBLOCK;
1599 		} else {
1600 			kq->kq_state |= KQ_SLEEP;
1601 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1602 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1603 		}
1604 		if (error == 0)
1605 			goto retry;
1606 		/* don't restart after signals... */
1607 		if (error == ERESTART)
1608 			error = EINTR;
1609 		else if (error == EWOULDBLOCK)
1610 			error = 0;
1611 		goto done;
1612 	}
1613 
1614 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1615 	influx = 0;
1616 	while (count) {
1617 		KQ_OWNED(kq);
1618 		kn = TAILQ_FIRST(&kq->kq_head);
1619 
1620 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1621 		    (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1622 			if (influx) {
1623 				influx = 0;
1624 				KQ_FLUX_WAKEUP(kq);
1625 			}
1626 			kq->kq_state |= KQ_FLUXWAIT;
1627 			error = msleep(kq, &kq->kq_lock, PSOCK,
1628 			    "kqflxwt", 0);
1629 			continue;
1630 		}
1631 
1632 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1633 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1634 			kn->kn_status &= ~KN_QUEUED;
1635 			kq->kq_count--;
1636 			continue;
1637 		}
1638 		if (kn == marker) {
1639 			KQ_FLUX_WAKEUP(kq);
1640 			if (count == maxevents)
1641 				goto retry;
1642 			goto done;
1643 		}
1644 		KASSERT((kn->kn_status & KN_INFLUX) == 0,
1645 		    ("KN_INFLUX set when not suppose to be"));
1646 
1647 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1648 			kn->kn_status &= ~KN_QUEUED;
1649 			kn->kn_status |= KN_INFLUX;
1650 			kq->kq_count--;
1651 			KQ_UNLOCK(kq);
1652 			/*
1653 			 * We don't need to lock the list since we've marked
1654 			 * it _INFLUX.
1655 			 */
1656 			if (!(kn->kn_status & KN_DETACHED))
1657 				kn->kn_fop->f_detach(kn);
1658 			knote_drop(kn, td);
1659 			KQ_LOCK(kq);
1660 			continue;
1661 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1662 			kn->kn_status &= ~KN_QUEUED;
1663 			kn->kn_status |= KN_INFLUX;
1664 			kq->kq_count--;
1665 			KQ_UNLOCK(kq);
1666 			/*
1667 			 * We don't need to lock the list since we've marked
1668 			 * it _INFLUX.
1669 			 */
1670 			*kevp = kn->kn_kevent;
1671 			if (!(kn->kn_status & KN_DETACHED))
1672 				kn->kn_fop->f_detach(kn);
1673 			knote_drop(kn, td);
1674 			KQ_LOCK(kq);
1675 			kn = NULL;
1676 		} else {
1677 			kn->kn_status |= KN_INFLUX | KN_SCAN;
1678 			KQ_UNLOCK(kq);
1679 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1680 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1681 			knl = kn_list_lock(kn);
1682 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1683 				KQ_LOCK(kq);
1684 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1685 				kn->kn_status &=
1686 				    ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1687 				    KN_SCAN);
1688 				kq->kq_count--;
1689 				kn_list_unlock(knl);
1690 				influx = 1;
1691 				continue;
1692 			}
1693 			touch = (!kn->kn_fop->f_isfd &&
1694 			    kn->kn_fop->f_touch != NULL);
1695 			if (touch)
1696 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1697 			else
1698 				*kevp = kn->kn_kevent;
1699 			KQ_LOCK(kq);
1700 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1701 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1702 				/*
1703 				 * Manually clear knotes who weren't
1704 				 * 'touch'ed.
1705 				 */
1706 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1707 					kn->kn_data = 0;
1708 					kn->kn_fflags = 0;
1709 				}
1710 				if (kn->kn_flags & EV_DISPATCH)
1711 					kn->kn_status |= KN_DISABLED;
1712 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1713 				kq->kq_count--;
1714 			} else
1715 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1716 
1717 			kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1718 			kn_list_unlock(knl);
1719 			influx = 1;
1720 		}
1721 
1722 		/* we are returning a copy to the user */
1723 		kevp++;
1724 		nkev++;
1725 		count--;
1726 
1727 		if (nkev == KQ_NEVENTS) {
1728 			influx = 0;
1729 			KQ_UNLOCK_FLUX(kq);
1730 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1731 			nkev = 0;
1732 			kevp = keva;
1733 			KQ_LOCK(kq);
1734 			if (error)
1735 				break;
1736 		}
1737 	}
1738 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1739 done:
1740 	KQ_OWNED(kq);
1741 	KQ_UNLOCK_FLUX(kq);
1742 	knote_free(marker);
1743 done_nl:
1744 	KQ_NOTOWNED(kq);
1745 	if (nkev != 0)
1746 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1747 	td->td_retval[0] = maxevents - count;
1748 	return (error);
1749 }
1750 
1751 /*ARGSUSED*/
1752 static int
1753 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1754 	struct ucred *active_cred, struct thread *td)
1755 {
1756 	/*
1757 	 * Enabling sigio causes two major problems:
1758 	 * 1) infinite recursion:
1759 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1760 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1761 	 * into itself over and over.  Sending the sigio causes the kqueue
1762 	 * to become ready, which in turn posts sigio again, forever.
1763 	 * Solution: this can be solved by setting a flag in the kqueue that
1764 	 * we have a SIGIO in progress.
1765 	 * 2) locking problems:
1766 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1767 	 * us above the proc and pgrp locks.
1768 	 * Solution: Post a signal using an async mechanism, being sure to
1769 	 * record a generation count in the delivery so that we do not deliver
1770 	 * a signal to the wrong process.
1771 	 *
1772 	 * Note, these two mechanisms are somewhat mutually exclusive!
1773 	 */
1774 #if 0
1775 	struct kqueue *kq;
1776 
1777 	kq = fp->f_data;
1778 	switch (cmd) {
1779 	case FIOASYNC:
1780 		if (*(int *)data) {
1781 			kq->kq_state |= KQ_ASYNC;
1782 		} else {
1783 			kq->kq_state &= ~KQ_ASYNC;
1784 		}
1785 		return (0);
1786 
1787 	case FIOSETOWN:
1788 		return (fsetown(*(int *)data, &kq->kq_sigio));
1789 
1790 	case FIOGETOWN:
1791 		*(int *)data = fgetown(&kq->kq_sigio);
1792 		return (0);
1793 	}
1794 #endif
1795 
1796 	return (ENOTTY);
1797 }
1798 
1799 /*ARGSUSED*/
1800 static int
1801 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1802 	struct thread *td)
1803 {
1804 	struct kqueue *kq;
1805 	int revents = 0;
1806 	int error;
1807 
1808 	if ((error = kqueue_acquire(fp, &kq)))
1809 		return POLLERR;
1810 
1811 	KQ_LOCK(kq);
1812 	if (events & (POLLIN | POLLRDNORM)) {
1813 		if (kq->kq_count) {
1814 			revents |= events & (POLLIN | POLLRDNORM);
1815 		} else {
1816 			selrecord(td, &kq->kq_sel);
1817 			if (SEL_WAITING(&kq->kq_sel))
1818 				kq->kq_state |= KQ_SEL;
1819 		}
1820 	}
1821 	kqueue_release(kq, 1);
1822 	KQ_UNLOCK(kq);
1823 	return (revents);
1824 }
1825 
1826 /*ARGSUSED*/
1827 static int
1828 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1829 	struct thread *td)
1830 {
1831 
1832 	bzero((void *)st, sizeof *st);
1833 	/*
1834 	 * We no longer return kq_count because the unlocked value is useless.
1835 	 * If you spent all this time getting the count, why not spend your
1836 	 * syscall better by calling kevent?
1837 	 *
1838 	 * XXX - This is needed for libc_r.
1839 	 */
1840 	st->st_mode = S_IFIFO;
1841 	return (0);
1842 }
1843 
1844 static void
1845 kqueue_drain(struct kqueue *kq, struct thread *td)
1846 {
1847 	struct knote *kn;
1848 	int i;
1849 
1850 	KQ_LOCK(kq);
1851 
1852 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1853 	    ("kqueue already closing"));
1854 	kq->kq_state |= KQ_CLOSING;
1855 	if (kq->kq_refcnt > 1)
1856 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1857 
1858 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1859 
1860 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
1861 	    ("kqueue's knlist not empty"));
1862 
1863 	for (i = 0; i < kq->kq_knlistsize; i++) {
1864 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1865 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1866 				kq->kq_state |= KQ_FLUXWAIT;
1867 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1868 				continue;
1869 			}
1870 			kn->kn_status |= KN_INFLUX;
1871 			KQ_UNLOCK(kq);
1872 			if (!(kn->kn_status & KN_DETACHED))
1873 				kn->kn_fop->f_detach(kn);
1874 			knote_drop(kn, td);
1875 			KQ_LOCK(kq);
1876 		}
1877 	}
1878 	if (kq->kq_knhashmask != 0) {
1879 		for (i = 0; i <= kq->kq_knhashmask; i++) {
1880 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1881 				if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1882 					kq->kq_state |= KQ_FLUXWAIT;
1883 					msleep(kq, &kq->kq_lock, PSOCK,
1884 					       "kqclo2", 0);
1885 					continue;
1886 				}
1887 				kn->kn_status |= KN_INFLUX;
1888 				KQ_UNLOCK(kq);
1889 				if (!(kn->kn_status & KN_DETACHED))
1890 					kn->kn_fop->f_detach(kn);
1891 				knote_drop(kn, td);
1892 				KQ_LOCK(kq);
1893 			}
1894 		}
1895 	}
1896 
1897 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1898 		kq->kq_state |= KQ_TASKDRAIN;
1899 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1900 	}
1901 
1902 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1903 		selwakeuppri(&kq->kq_sel, PSOCK);
1904 		if (!SEL_WAITING(&kq->kq_sel))
1905 			kq->kq_state &= ~KQ_SEL;
1906 	}
1907 
1908 	KQ_UNLOCK(kq);
1909 }
1910 
1911 static void
1912 kqueue_destroy(struct kqueue *kq)
1913 {
1914 
1915 	KASSERT(kq->kq_fdp == NULL,
1916 	    ("kqueue still attached to a file descriptor"));
1917 	seldrain(&kq->kq_sel);
1918 	knlist_destroy(&kq->kq_sel.si_note);
1919 	mtx_destroy(&kq->kq_lock);
1920 
1921 	if (kq->kq_knhash != NULL)
1922 		free(kq->kq_knhash, M_KQUEUE);
1923 	if (kq->kq_knlist != NULL)
1924 		free(kq->kq_knlist, M_KQUEUE);
1925 
1926 	funsetown(&kq->kq_sigio);
1927 }
1928 
1929 /*ARGSUSED*/
1930 static int
1931 kqueue_close(struct file *fp, struct thread *td)
1932 {
1933 	struct kqueue *kq = fp->f_data;
1934 	struct filedesc *fdp;
1935 	int error;
1936 	int filedesc_unlock;
1937 
1938 	if ((error = kqueue_acquire(fp, &kq)))
1939 		return error;
1940 	kqueue_drain(kq, td);
1941 
1942 	/*
1943 	 * We could be called due to the knote_drop() doing fdrop(),
1944 	 * called from kqueue_register().  In this case the global
1945 	 * lock is owned, and filedesc sx is locked before, to not
1946 	 * take the sleepable lock after non-sleepable.
1947 	 */
1948 	fdp = kq->kq_fdp;
1949 	kq->kq_fdp = NULL;
1950 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1951 		FILEDESC_XLOCK(fdp);
1952 		filedesc_unlock = 1;
1953 	} else
1954 		filedesc_unlock = 0;
1955 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1956 	if (filedesc_unlock)
1957 		FILEDESC_XUNLOCK(fdp);
1958 
1959 	kqueue_destroy(kq);
1960 	chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
1961 	crfree(kq->kq_cred);
1962 	free(kq, M_KQUEUE);
1963 	fp->f_data = NULL;
1964 
1965 	return (0);
1966 }
1967 
1968 static int
1969 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1970 {
1971 
1972 	kif->kf_type = KF_TYPE_KQUEUE;
1973 	return (0);
1974 }
1975 
1976 static void
1977 kqueue_wakeup(struct kqueue *kq)
1978 {
1979 	KQ_OWNED(kq);
1980 
1981 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1982 		kq->kq_state &= ~KQ_SLEEP;
1983 		wakeup(kq);
1984 	}
1985 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1986 		selwakeuppri(&kq->kq_sel, PSOCK);
1987 		if (!SEL_WAITING(&kq->kq_sel))
1988 			kq->kq_state &= ~KQ_SEL;
1989 	}
1990 	if (!knlist_empty(&kq->kq_sel.si_note))
1991 		kqueue_schedtask(kq);
1992 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1993 		pgsigio(&kq->kq_sigio, SIGIO, 0);
1994 	}
1995 }
1996 
1997 /*
1998  * Walk down a list of knotes, activating them if their event has triggered.
1999  *
2000  * There is a possibility to optimize in the case of one kq watching another.
2001  * Instead of scheduling a task to wake it up, you could pass enough state
2002  * down the chain to make up the parent kqueue.  Make this code functional
2003  * first.
2004  */
2005 void
2006 knote(struct knlist *list, long hint, int lockflags)
2007 {
2008 	struct kqueue *kq;
2009 	struct knote *kn, *tkn;
2010 	int error;
2011 
2012 	if (list == NULL)
2013 		return;
2014 
2015 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
2016 
2017 	if ((lockflags & KNF_LISTLOCKED) == 0)
2018 		list->kl_lock(list->kl_lockarg);
2019 
2020 	/*
2021 	 * If we unlock the list lock (and set KN_INFLUX), we can
2022 	 * eliminate the kqueue scheduling, but this will introduce
2023 	 * four lock/unlock's for each knote to test.  Also, marker
2024 	 * would be needed to keep iteration position, since filters
2025 	 * or other threads could remove events.
2026 	 */
2027 	SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
2028 		kq = kn->kn_kq;
2029 		KQ_LOCK(kq);
2030 		if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
2031 			/*
2032 			 * Do not process the influx notes, except for
2033 			 * the influx coming from the kq unlock in the
2034 			 * kqueue_scan().  In the later case, we do
2035 			 * not interfere with the scan, since the code
2036 			 * fragment in kqueue_scan() locks the knlist,
2037 			 * and cannot proceed until we finished.
2038 			 */
2039 			KQ_UNLOCK(kq);
2040 		} else if ((lockflags & KNF_NOKQLOCK) != 0) {
2041 			kn->kn_status |= KN_INFLUX;
2042 			KQ_UNLOCK(kq);
2043 			error = kn->kn_fop->f_event(kn, hint);
2044 			KQ_LOCK(kq);
2045 			kn->kn_status &= ~KN_INFLUX;
2046 			if (error)
2047 				KNOTE_ACTIVATE(kn, 1);
2048 			KQ_UNLOCK_FLUX(kq);
2049 		} else {
2050 			kn->kn_status |= KN_HASKQLOCK;
2051 			if (kn->kn_fop->f_event(kn, hint))
2052 				KNOTE_ACTIVATE(kn, 1);
2053 			kn->kn_status &= ~KN_HASKQLOCK;
2054 			KQ_UNLOCK(kq);
2055 		}
2056 	}
2057 	if ((lockflags & KNF_LISTLOCKED) == 0)
2058 		list->kl_unlock(list->kl_lockarg);
2059 }
2060 
2061 /*
2062  * add a knote to a knlist
2063  */
2064 void
2065 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2066 {
2067 	KNL_ASSERT_LOCK(knl, islocked);
2068 	KQ_NOTOWNED(kn->kn_kq);
2069 	KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
2070 	    (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
2071 	if (!islocked)
2072 		knl->kl_lock(knl->kl_lockarg);
2073 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2074 	if (!islocked)
2075 		knl->kl_unlock(knl->kl_lockarg);
2076 	KQ_LOCK(kn->kn_kq);
2077 	kn->kn_knlist = knl;
2078 	kn->kn_status &= ~KN_DETACHED;
2079 	KQ_UNLOCK(kn->kn_kq);
2080 }
2081 
2082 static void
2083 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked,
2084     int kqislocked)
2085 {
2086 	KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
2087 	KNL_ASSERT_LOCK(knl, knlislocked);
2088 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2089 	if (!kqislocked)
2090 		KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
2091     ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
2092 	if (!knlislocked)
2093 		knl->kl_lock(knl->kl_lockarg);
2094 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2095 	kn->kn_knlist = NULL;
2096 	if (!knlislocked)
2097 		kn_list_unlock(knl);
2098 	if (!kqislocked)
2099 		KQ_LOCK(kn->kn_kq);
2100 	kn->kn_status |= KN_DETACHED;
2101 	if (!kqislocked)
2102 		KQ_UNLOCK(kn->kn_kq);
2103 }
2104 
2105 /*
2106  * remove knote from the specified knlist
2107  */
2108 void
2109 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2110 {
2111 
2112 	knlist_remove_kq(knl, kn, islocked, 0);
2113 }
2114 
2115 int
2116 knlist_empty(struct knlist *knl)
2117 {
2118 
2119 	KNL_ASSERT_LOCKED(knl);
2120 	return SLIST_EMPTY(&knl->kl_list);
2121 }
2122 
2123 static struct mtx	knlist_lock;
2124 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2125 	MTX_DEF);
2126 static void knlist_mtx_lock(void *arg);
2127 static void knlist_mtx_unlock(void *arg);
2128 
2129 static void
2130 knlist_mtx_lock(void *arg)
2131 {
2132 
2133 	mtx_lock((struct mtx *)arg);
2134 }
2135 
2136 static void
2137 knlist_mtx_unlock(void *arg)
2138 {
2139 
2140 	mtx_unlock((struct mtx *)arg);
2141 }
2142 
2143 static void
2144 knlist_mtx_assert_locked(void *arg)
2145 {
2146 
2147 	mtx_assert((struct mtx *)arg, MA_OWNED);
2148 }
2149 
2150 static void
2151 knlist_mtx_assert_unlocked(void *arg)
2152 {
2153 
2154 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2155 }
2156 
2157 static void
2158 knlist_rw_rlock(void *arg)
2159 {
2160 
2161 	rw_rlock((struct rwlock *)arg);
2162 }
2163 
2164 static void
2165 knlist_rw_runlock(void *arg)
2166 {
2167 
2168 	rw_runlock((struct rwlock *)arg);
2169 }
2170 
2171 static void
2172 knlist_rw_assert_locked(void *arg)
2173 {
2174 
2175 	rw_assert((struct rwlock *)arg, RA_LOCKED);
2176 }
2177 
2178 static void
2179 knlist_rw_assert_unlocked(void *arg)
2180 {
2181 
2182 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2183 }
2184 
2185 void
2186 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2187     void (*kl_unlock)(void *),
2188     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2189 {
2190 
2191 	if (lock == NULL)
2192 		knl->kl_lockarg = &knlist_lock;
2193 	else
2194 		knl->kl_lockarg = lock;
2195 
2196 	if (kl_lock == NULL)
2197 		knl->kl_lock = knlist_mtx_lock;
2198 	else
2199 		knl->kl_lock = kl_lock;
2200 	if (kl_unlock == NULL)
2201 		knl->kl_unlock = knlist_mtx_unlock;
2202 	else
2203 		knl->kl_unlock = kl_unlock;
2204 	if (kl_assert_locked == NULL)
2205 		knl->kl_assert_locked = knlist_mtx_assert_locked;
2206 	else
2207 		knl->kl_assert_locked = kl_assert_locked;
2208 	if (kl_assert_unlocked == NULL)
2209 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2210 	else
2211 		knl->kl_assert_unlocked = kl_assert_unlocked;
2212 
2213 	knl->kl_autodestroy = 0;
2214 	SLIST_INIT(&knl->kl_list);
2215 }
2216 
2217 void
2218 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2219 {
2220 
2221 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2222 }
2223 
2224 struct knlist *
2225 knlist_alloc(struct mtx *lock)
2226 {
2227 	struct knlist *knl;
2228 
2229 	knl = malloc(sizeof(struct knlist), M_KQUEUE, M_WAITOK);
2230 	knlist_init_mtx(knl, lock);
2231 	return (knl);
2232 }
2233 
2234 void
2235 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2236 {
2237 
2238 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2239 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2240 }
2241 
2242 void
2243 knlist_destroy(struct knlist *knl)
2244 {
2245 
2246 #ifdef INVARIANTS
2247 	/*
2248 	 * if we run across this error, we need to find the offending
2249 	 * driver and have it call knlist_clear or knlist_delete.
2250 	 */
2251 	if (!SLIST_EMPTY(&knl->kl_list))
2252 		printf("WARNING: destroying knlist w/ knotes on it!\n");
2253 #endif
2254 
2255 	knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2256 	SLIST_INIT(&knl->kl_list);
2257 }
2258 
2259 void
2260 knlist_detach(struct knlist *knl)
2261 {
2262 
2263 	KNL_ASSERT_LOCKED(knl);
2264 	knl->kl_autodestroy = 1;
2265 	if (knlist_empty(knl)) {
2266 		knlist_destroy(knl);
2267 		free(knl, M_KQUEUE);
2268 	}
2269 }
2270 
2271 /*
2272  * Even if we are locked, we may need to drop the lock to allow any influx
2273  * knotes time to "settle".
2274  */
2275 void
2276 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2277 {
2278 	struct knote *kn, *kn2;
2279 	struct kqueue *kq;
2280 
2281 	KASSERT(!knl->kl_autodestroy, ("cleardel for autodestroy %p", knl));
2282 	if (islocked)
2283 		KNL_ASSERT_LOCKED(knl);
2284 	else {
2285 		KNL_ASSERT_UNLOCKED(knl);
2286 again:		/* need to reacquire lock since we have dropped it */
2287 		knl->kl_lock(knl->kl_lockarg);
2288 	}
2289 
2290 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2291 		kq = kn->kn_kq;
2292 		KQ_LOCK(kq);
2293 		if ((kn->kn_status & KN_INFLUX)) {
2294 			KQ_UNLOCK(kq);
2295 			continue;
2296 		}
2297 		knlist_remove_kq(knl, kn, 1, 1);
2298 		if (killkn) {
2299 			kn->kn_status |= KN_INFLUX | KN_DETACHED;
2300 			KQ_UNLOCK(kq);
2301 			knote_drop(kn, td);
2302 		} else {
2303 			/* Make sure cleared knotes disappear soon */
2304 			kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2305 			KQ_UNLOCK(kq);
2306 		}
2307 		kq = NULL;
2308 	}
2309 
2310 	if (!SLIST_EMPTY(&knl->kl_list)) {
2311 		/* there are still KN_INFLUX remaining */
2312 		kn = SLIST_FIRST(&knl->kl_list);
2313 		kq = kn->kn_kq;
2314 		KQ_LOCK(kq);
2315 		KASSERT(kn->kn_status & KN_INFLUX,
2316 		    ("knote removed w/o list lock"));
2317 		knl->kl_unlock(knl->kl_lockarg);
2318 		kq->kq_state |= KQ_FLUXWAIT;
2319 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2320 		kq = NULL;
2321 		goto again;
2322 	}
2323 
2324 	if (islocked)
2325 		KNL_ASSERT_LOCKED(knl);
2326 	else {
2327 		knl->kl_unlock(knl->kl_lockarg);
2328 		KNL_ASSERT_UNLOCKED(knl);
2329 	}
2330 }
2331 
2332 /*
2333  * Remove all knotes referencing a specified fd must be called with FILEDESC
2334  * lock.  This prevents a race where a new fd comes along and occupies the
2335  * entry and we attach a knote to the fd.
2336  */
2337 void
2338 knote_fdclose(struct thread *td, int fd)
2339 {
2340 	struct filedesc *fdp = td->td_proc->p_fd;
2341 	struct kqueue *kq;
2342 	struct knote *kn;
2343 	int influx;
2344 
2345 	FILEDESC_XLOCK_ASSERT(fdp);
2346 
2347 	/*
2348 	 * We shouldn't have to worry about new kevents appearing on fd
2349 	 * since filedesc is locked.
2350 	 */
2351 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2352 		KQ_LOCK(kq);
2353 
2354 again:
2355 		influx = 0;
2356 		while (kq->kq_knlistsize > fd &&
2357 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2358 			if (kn->kn_status & KN_INFLUX) {
2359 				/* someone else might be waiting on our knote */
2360 				if (influx)
2361 					wakeup(kq);
2362 				kq->kq_state |= KQ_FLUXWAIT;
2363 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2364 				goto again;
2365 			}
2366 			kn->kn_status |= KN_INFLUX;
2367 			KQ_UNLOCK(kq);
2368 			if (!(kn->kn_status & KN_DETACHED))
2369 				kn->kn_fop->f_detach(kn);
2370 			knote_drop(kn, td);
2371 			influx = 1;
2372 			KQ_LOCK(kq);
2373 		}
2374 		KQ_UNLOCK_FLUX(kq);
2375 	}
2376 }
2377 
2378 static int
2379 knote_attach(struct knote *kn, struct kqueue *kq)
2380 {
2381 	struct klist *list;
2382 
2383 	KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2384 	KQ_OWNED(kq);
2385 
2386 	if (kn->kn_fop->f_isfd) {
2387 		if (kn->kn_id >= kq->kq_knlistsize)
2388 			return ENOMEM;
2389 		list = &kq->kq_knlist[kn->kn_id];
2390 	} else {
2391 		if (kq->kq_knhash == NULL)
2392 			return ENOMEM;
2393 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2394 	}
2395 
2396 	SLIST_INSERT_HEAD(list, kn, kn_link);
2397 
2398 	return 0;
2399 }
2400 
2401 /*
2402  * knote must already have been detached using the f_detach method.
2403  * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2404  * to prevent other removal.
2405  */
2406 static void
2407 knote_drop(struct knote *kn, struct thread *td)
2408 {
2409 	struct kqueue *kq;
2410 	struct klist *list;
2411 
2412 	kq = kn->kn_kq;
2413 
2414 	KQ_NOTOWNED(kq);
2415 	KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2416 	    ("knote_drop called without KN_INFLUX set in kn_status"));
2417 
2418 	KQ_LOCK(kq);
2419 	if (kn->kn_fop->f_isfd)
2420 		list = &kq->kq_knlist[kn->kn_id];
2421 	else
2422 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2423 
2424 	if (!SLIST_EMPTY(list))
2425 		SLIST_REMOVE(list, kn, knote, kn_link);
2426 	if (kn->kn_status & KN_QUEUED)
2427 		knote_dequeue(kn);
2428 	KQ_UNLOCK_FLUX(kq);
2429 
2430 	if (kn->kn_fop->f_isfd) {
2431 		fdrop(kn->kn_fp, td);
2432 		kn->kn_fp = NULL;
2433 	}
2434 	kqueue_fo_release(kn->kn_kevent.filter);
2435 	kn->kn_fop = NULL;
2436 	knote_free(kn);
2437 }
2438 
2439 static void
2440 knote_enqueue(struct knote *kn)
2441 {
2442 	struct kqueue *kq = kn->kn_kq;
2443 
2444 	KQ_OWNED(kn->kn_kq);
2445 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2446 
2447 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2448 	kn->kn_status |= KN_QUEUED;
2449 	kq->kq_count++;
2450 	kqueue_wakeup(kq);
2451 }
2452 
2453 static void
2454 knote_dequeue(struct knote *kn)
2455 {
2456 	struct kqueue *kq = kn->kn_kq;
2457 
2458 	KQ_OWNED(kn->kn_kq);
2459 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2460 
2461 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2462 	kn->kn_status &= ~KN_QUEUED;
2463 	kq->kq_count--;
2464 }
2465 
2466 static void
2467 knote_init(void)
2468 {
2469 
2470 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2471 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2472 }
2473 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2474 
2475 static struct knote *
2476 knote_alloc(int waitok)
2477 {
2478 
2479 	return (uma_zalloc(knote_zone, (waitok ? M_WAITOK : M_NOWAIT) |
2480 	    M_ZERO));
2481 }
2482 
2483 static void
2484 knote_free(struct knote *kn)
2485 {
2486 
2487 	uma_zfree(knote_zone, kn);
2488 }
2489 
2490 /*
2491  * Register the kev w/ the kq specified by fd.
2492  */
2493 int
2494 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2495 {
2496 	struct kqueue *kq;
2497 	struct file *fp;
2498 	cap_rights_t rights;
2499 	int error;
2500 
2501 	error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2502 	if (error != 0)
2503 		return (error);
2504 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2505 		goto noacquire;
2506 
2507 	error = kqueue_register(kq, kev, td, waitok);
2508 
2509 	kqueue_release(kq, 0);
2510 
2511 noacquire:
2512 	fdrop(fp, td);
2513 
2514 	return error;
2515 }
2516