xref: /freebsd/sys/kern/kern_event.c (revision 9eb3f14324736fe9895cf83ab17756dfea8f7d8e)
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 	/* Mask off extra data. */
455 	event = (u_int)hint & NOTE_PCTRLMASK;
456 
457 	/* If the user is interested in this event, record it. */
458 	if (kn->kn_sfflags & event)
459 		kn->kn_fflags |= event;
460 
461 	/* Process is gone, so flag the event as finished. */
462 	if (event == NOTE_EXIT) {
463 		kn->kn_flags |= EV_EOF | EV_ONESHOT;
464 		kn->kn_ptr.p_proc = NULL;
465 		if (kn->kn_fflags & NOTE_EXIT)
466 			kn->kn_data = KW_EXITCODE(p->p_xexit, p->p_xsig);
467 		if (kn->kn_fflags == 0)
468 			kn->kn_flags |= EV_DROP;
469 		return (1);
470 	}
471 
472 	return (kn->kn_fflags != 0);
473 }
474 
475 /*
476  * Called when the process forked. It mostly does the same as the
477  * knote(), activating all knotes registered to be activated when the
478  * process forked. Additionally, for each knote attached to the
479  * parent, check whether user wants to track the new process. If so
480  * attach a new knote to it, and immediately report an event with the
481  * child's pid.
482  */
483 void
484 knote_fork(struct knlist *list, int pid)
485 {
486 	struct kqueue *kq;
487 	struct knote *kn;
488 	struct kevent kev;
489 	int error;
490 
491 	if (list == NULL)
492 		return;
493 	list->kl_lock(list->kl_lockarg);
494 
495 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
496 		kq = kn->kn_kq;
497 		KQ_LOCK(kq);
498 		if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
499 			KQ_UNLOCK(kq);
500 			continue;
501 		}
502 
503 		/*
504 		 * The same as knote(), activate the event.
505 		 */
506 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
507 			kn->kn_status |= KN_HASKQLOCK;
508 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
509 				KNOTE_ACTIVATE(kn, 1);
510 			kn->kn_status &= ~KN_HASKQLOCK;
511 			KQ_UNLOCK(kq);
512 			continue;
513 		}
514 
515 		/*
516 		 * The NOTE_TRACK case. In addition to the activation
517 		 * of the event, we need to register new events to
518 		 * track the child. Drop the locks in preparation for
519 		 * the call to kqueue_register().
520 		 */
521 		kn->kn_status |= KN_INFLUX;
522 		KQ_UNLOCK(kq);
523 		list->kl_unlock(list->kl_lockarg);
524 
525 		/*
526 		 * Activate existing knote and register tracking knotes with
527 		 * new process.
528 		 *
529 		 * First register a knote to get just the child notice. This
530 		 * must be a separate note from a potential NOTE_EXIT
531 		 * notification since both NOTE_CHILD and NOTE_EXIT are defined
532 		 * to use the data field (in conflicting ways).
533 		 */
534 		kev.ident = pid;
535 		kev.filter = kn->kn_filter;
536 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_ONESHOT |
537 		    EV_FLAG2;
538 		kev.fflags = kn->kn_sfflags;
539 		kev.data = kn->kn_id;		/* parent */
540 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
541 		error = kqueue_register(kq, &kev, NULL, 0);
542 		if (error)
543 			kn->kn_fflags |= NOTE_TRACKERR;
544 
545 		/*
546 		 * Then register another knote to track other potential events
547 		 * from the new process.
548 		 */
549 		kev.ident = pid;
550 		kev.filter = kn->kn_filter;
551 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
552 		kev.fflags = kn->kn_sfflags;
553 		kev.data = kn->kn_id;		/* parent */
554 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
555 		error = kqueue_register(kq, &kev, NULL, 0);
556 		if (error)
557 			kn->kn_fflags |= NOTE_TRACKERR;
558 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
559 			KNOTE_ACTIVATE(kn, 0);
560 		KQ_LOCK(kq);
561 		kn->kn_status &= ~KN_INFLUX;
562 		KQ_UNLOCK_FLUX(kq);
563 		list->kl_lock(list->kl_lockarg);
564 	}
565 	list->kl_unlock(list->kl_lockarg);
566 }
567 
568 /*
569  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
570  * interval timer support code.
571  */
572 
573 #define NOTE_TIMER_PRECMASK	(NOTE_SECONDS|NOTE_MSECONDS|NOTE_USECONDS| \
574 				NOTE_NSECONDS)
575 
576 static sbintime_t
577 timer2sbintime(intptr_t data, int flags)
578 {
579 
580         /*
581          * Macros for converting to the fractional second portion of an
582          * sbintime_t using 64bit multiplication to improve precision.
583          */
584 #define NS_TO_SBT(ns) (((ns) * (((uint64_t)1 << 63) / 500000000)) >> 32)
585 #define US_TO_SBT(us) (((us) * (((uint64_t)1 << 63) / 500000)) >> 32)
586 #define MS_TO_SBT(ms) (((ms) * (((uint64_t)1 << 63) / 500)) >> 32)
587 	switch (flags & NOTE_TIMER_PRECMASK) {
588 	case NOTE_SECONDS:
589 #ifdef __LP64__
590 		if (data > (SBT_MAX / SBT_1S))
591 			return SBT_MAX;
592 #endif
593 		return ((sbintime_t)data << 32);
594 	case NOTE_MSECONDS: /* FALLTHROUGH */
595 	case 0:
596 		if (data >= 1000) {
597 			int64_t secs = data / 1000;
598 #ifdef __LP64__
599 			if (secs > (SBT_MAX / SBT_1S))
600 				return SBT_MAX;
601 #endif
602 			return (secs << 32 | MS_TO_SBT(data % 1000));
603 		}
604 		return MS_TO_SBT(data);
605 	case NOTE_USECONDS:
606 		if (data >= 1000000) {
607 			int64_t secs = data / 1000000;
608 #ifdef __LP64__
609 			if (secs > (SBT_MAX / SBT_1S))
610 				return SBT_MAX;
611 #endif
612 			return (secs << 32 | US_TO_SBT(data % 1000000));
613 		}
614 		return US_TO_SBT(data);
615 	case NOTE_NSECONDS:
616 		if (data >= 1000000000) {
617 			int64_t secs = data / 1000000000;
618 #ifdef __LP64__
619 			if (secs > (SBT_MAX / SBT_1S))
620 				return SBT_MAX;
621 #endif
622 			return (secs << 32 | US_TO_SBT(data % 1000000000));
623 		}
624 		return NS_TO_SBT(data);
625 	default:
626 		break;
627 	}
628 	return (-1);
629 }
630 
631 static void
632 filt_timerexpire(void *knx)
633 {
634 	struct callout *calloutp;
635 	struct knote *kn;
636 
637 	kn = knx;
638 	kn->kn_data++;
639 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
640 
641 	if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
642 		calloutp = (struct callout *)kn->kn_hook;
643 		*kn->kn_ptr.p_nexttime += timer2sbintime(kn->kn_sdata,
644 		    kn->kn_sfflags);
645 		callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
646 		    filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
647 	}
648 }
649 
650 /*
651  * data contains amount of time to sleep
652  */
653 static int
654 filt_timerattach(struct knote *kn)
655 {
656 	struct callout *calloutp;
657 	sbintime_t to;
658 	unsigned int ncallouts;
659 
660 	if ((intptr_t)kn->kn_sdata < 0)
661 		return (EINVAL);
662 	if ((intptr_t)kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
663 		kn->kn_sdata = 1;
664 	/* Only precision unit are supported in flags so far */
665 	if (kn->kn_sfflags & ~NOTE_TIMER_PRECMASK)
666 		return (EINVAL);
667 
668 	to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
669 	if (to < 0)
670 		return (EINVAL);
671 
672 	ncallouts = atomic_load_explicit(&kq_ncallouts, memory_order_relaxed);
673 	do {
674 		if (ncallouts >= kq_calloutmax)
675 			return (ENOMEM);
676 	} while (!atomic_compare_exchange_weak_explicit(&kq_ncallouts,
677 	    &ncallouts, ncallouts + 1, memory_order_relaxed,
678 	    memory_order_relaxed));
679 
680 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
681 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
682 	kn->kn_ptr.p_nexttime = malloc(sizeof(sbintime_t), M_KQUEUE, M_WAITOK);
683 	calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
684 	callout_init(calloutp, 1);
685 	kn->kn_hook = calloutp;
686 	*kn->kn_ptr.p_nexttime = to + sbinuptime();
687 	callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
688 	    filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
689 
690 	return (0);
691 }
692 
693 static void
694 filt_timerdetach(struct knote *kn)
695 {
696 	struct callout *calloutp;
697 	unsigned int old;
698 
699 	calloutp = (struct callout *)kn->kn_hook;
700 	callout_drain(calloutp);
701 	free(calloutp, M_KQUEUE);
702 	free(kn->kn_ptr.p_nexttime, M_KQUEUE);
703 	old = atomic_fetch_sub_explicit(&kq_ncallouts, 1, memory_order_relaxed);
704 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
705 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
706 }
707 
708 static int
709 filt_timer(struct knote *kn, long hint)
710 {
711 
712 	return (kn->kn_data != 0);
713 }
714 
715 static int
716 filt_userattach(struct knote *kn)
717 {
718 
719 	/*
720 	 * EVFILT_USER knotes are not attached to anything in the kernel.
721 	 */
722 	kn->kn_hook = NULL;
723 	if (kn->kn_fflags & NOTE_TRIGGER)
724 		kn->kn_hookid = 1;
725 	else
726 		kn->kn_hookid = 0;
727 	return (0);
728 }
729 
730 static void
731 filt_userdetach(__unused struct knote *kn)
732 {
733 
734 	/*
735 	 * EVFILT_USER knotes are not attached to anything in the kernel.
736 	 */
737 }
738 
739 static int
740 filt_user(struct knote *kn, __unused long hint)
741 {
742 
743 	return (kn->kn_hookid);
744 }
745 
746 static void
747 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
748 {
749 	u_int ffctrl;
750 
751 	switch (type) {
752 	case EVENT_REGISTER:
753 		if (kev->fflags & NOTE_TRIGGER)
754 			kn->kn_hookid = 1;
755 
756 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
757 		kev->fflags &= NOTE_FFLAGSMASK;
758 		switch (ffctrl) {
759 		case NOTE_FFNOP:
760 			break;
761 
762 		case NOTE_FFAND:
763 			kn->kn_sfflags &= kev->fflags;
764 			break;
765 
766 		case NOTE_FFOR:
767 			kn->kn_sfflags |= kev->fflags;
768 			break;
769 
770 		case NOTE_FFCOPY:
771 			kn->kn_sfflags = kev->fflags;
772 			break;
773 
774 		default:
775 			/* XXX Return error? */
776 			break;
777 		}
778 		kn->kn_sdata = kev->data;
779 		if (kev->flags & EV_CLEAR) {
780 			kn->kn_hookid = 0;
781 			kn->kn_data = 0;
782 			kn->kn_fflags = 0;
783 		}
784 		break;
785 
786         case EVENT_PROCESS:
787 		*kev = kn->kn_kevent;
788 		kev->fflags = kn->kn_sfflags;
789 		kev->data = kn->kn_sdata;
790 		if (kn->kn_flags & EV_CLEAR) {
791 			kn->kn_hookid = 0;
792 			kn->kn_data = 0;
793 			kn->kn_fflags = 0;
794 		}
795 		break;
796 
797 	default:
798 		panic("filt_usertouch() - invalid type (%ld)", type);
799 		break;
800 	}
801 }
802 
803 int
804 sys_kqueue(struct thread *td, struct kqueue_args *uap)
805 {
806 
807 	return (kern_kqueue(td, 0, NULL));
808 }
809 
810 static void
811 kqueue_init(struct kqueue *kq)
812 {
813 
814 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
815 	TAILQ_INIT(&kq->kq_head);
816 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
817 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
818 }
819 
820 int
821 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
822 {
823 	struct filedesc *fdp;
824 	struct kqueue *kq;
825 	struct file *fp;
826 	struct ucred *cred;
827 	int fd, error;
828 
829 	fdp = td->td_proc->p_fd;
830 	cred = td->td_ucred;
831 	if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
832 		return (ENOMEM);
833 
834 	error = falloc_caps(td, &fp, &fd, flags, fcaps);
835 	if (error != 0) {
836 		chgkqcnt(cred->cr_ruidinfo, -1, 0);
837 		return (error);
838 	}
839 
840 	/* An extra reference on `fp' has been held for us by falloc(). */
841 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
842 	kqueue_init(kq);
843 	kq->kq_fdp = fdp;
844 	kq->kq_cred = crhold(cred);
845 
846 	FILEDESC_XLOCK(fdp);
847 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
848 	FILEDESC_XUNLOCK(fdp);
849 
850 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
851 	fdrop(fp, td);
852 
853 	td->td_retval[0] = fd;
854 	return (0);
855 }
856 
857 #ifndef _SYS_SYSPROTO_H_
858 struct kevent_args {
859 	int	fd;
860 	const struct kevent *changelist;
861 	int	nchanges;
862 	struct	kevent *eventlist;
863 	int	nevents;
864 	const struct timespec *timeout;
865 };
866 #endif
867 int
868 sys_kevent(struct thread *td, struct kevent_args *uap)
869 {
870 	struct timespec ts, *tsp;
871 	struct kevent_copyops k_ops = { uap,
872 					kevent_copyout,
873 					kevent_copyin};
874 	int error;
875 #ifdef KTRACE
876 	struct uio ktruio;
877 	struct iovec ktriov;
878 	struct uio *ktruioin = NULL;
879 	struct uio *ktruioout = NULL;
880 #endif
881 
882 	if (uap->timeout != NULL) {
883 		error = copyin(uap->timeout, &ts, sizeof(ts));
884 		if (error)
885 			return (error);
886 		tsp = &ts;
887 	} else
888 		tsp = NULL;
889 
890 #ifdef KTRACE
891 	if (KTRPOINT(td, KTR_GENIO)) {
892 		ktriov.iov_base = uap->changelist;
893 		ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
894 		ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
895 		    .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
896 		    .uio_td = td };
897 		ktruioin = cloneuio(&ktruio);
898 		ktriov.iov_base = uap->eventlist;
899 		ktriov.iov_len = uap->nevents * sizeof(struct kevent);
900 		ktruioout = cloneuio(&ktruio);
901 	}
902 #endif
903 
904 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
905 	    &k_ops, tsp);
906 
907 #ifdef KTRACE
908 	if (ktruioin != NULL) {
909 		ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
910 		ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
911 		ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
912 		ktrgenio(uap->fd, UIO_READ, ktruioout, error);
913 	}
914 #endif
915 
916 	return (error);
917 }
918 
919 /*
920  * Copy 'count' items into the destination list pointed to by uap->eventlist.
921  */
922 static int
923 kevent_copyout(void *arg, struct kevent *kevp, int count)
924 {
925 	struct kevent_args *uap;
926 	int error;
927 
928 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
929 	uap = (struct kevent_args *)arg;
930 
931 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
932 	if (error == 0)
933 		uap->eventlist += count;
934 	return (error);
935 }
936 
937 /*
938  * Copy 'count' items from the list pointed to by uap->changelist.
939  */
940 static int
941 kevent_copyin(void *arg, struct kevent *kevp, int count)
942 {
943 	struct kevent_args *uap;
944 	int error;
945 
946 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
947 	uap = (struct kevent_args *)arg;
948 
949 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
950 	if (error == 0)
951 		uap->changelist += count;
952 	return (error);
953 }
954 
955 int
956 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
957     struct kevent_copyops *k_ops, const struct timespec *timeout)
958 {
959 	cap_rights_t rights;
960 	struct file *fp;
961 	int error;
962 
963 	cap_rights_init(&rights);
964 	if (nchanges > 0)
965 		cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
966 	if (nevents > 0)
967 		cap_rights_set(&rights, CAP_KQUEUE_EVENT);
968 	error = fget(td, fd, &rights, &fp);
969 	if (error != 0)
970 		return (error);
971 
972 	error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
973 	fdrop(fp, td);
974 
975 	return (error);
976 }
977 
978 static int
979 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
980     struct kevent_copyops *k_ops, const struct timespec *timeout)
981 {
982 	struct kevent keva[KQ_NEVENTS];
983 	struct kevent *kevp, *changes;
984 	int i, n, nerrors, error;
985 
986 	nerrors = 0;
987 	while (nchanges > 0) {
988 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
989 		error = k_ops->k_copyin(k_ops->arg, keva, n);
990 		if (error)
991 			return (error);
992 		changes = keva;
993 		for (i = 0; i < n; i++) {
994 			kevp = &changes[i];
995 			if (!kevp->filter)
996 				continue;
997 			kevp->flags &= ~EV_SYSFLAGS;
998 			error = kqueue_register(kq, kevp, td, 1);
999 			if (error || (kevp->flags & EV_RECEIPT)) {
1000 				if (nevents == 0)
1001 					return (error);
1002 				kevp->flags = EV_ERROR;
1003 				kevp->data = error;
1004 				(void)k_ops->k_copyout(k_ops->arg, kevp, 1);
1005 				nevents--;
1006 				nerrors++;
1007 			}
1008 		}
1009 		nchanges -= n;
1010 	}
1011 	if (nerrors) {
1012 		td->td_retval[0] = nerrors;
1013 		return (0);
1014 	}
1015 
1016 	return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
1017 }
1018 
1019 int
1020 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
1021     struct kevent_copyops *k_ops, const struct timespec *timeout)
1022 {
1023 	struct kqueue *kq;
1024 	int error;
1025 
1026 	error = kqueue_acquire(fp, &kq);
1027 	if (error != 0)
1028 		return (error);
1029 	error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
1030 	kqueue_release(kq, 0);
1031 	return (error);
1032 }
1033 
1034 /*
1035  * Performs a kevent() call on a temporarily created kqueue. This can be
1036  * used to perform one-shot polling, similar to poll() and select().
1037  */
1038 int
1039 kern_kevent_anonymous(struct thread *td, int nevents,
1040     struct kevent_copyops *k_ops)
1041 {
1042 	struct kqueue kq = {};
1043 	int error;
1044 
1045 	kqueue_init(&kq);
1046 	kq.kq_refcnt = 1;
1047 	error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
1048 	kqueue_drain(&kq, td);
1049 	kqueue_destroy(&kq);
1050 	return (error);
1051 }
1052 
1053 int
1054 kqueue_add_filteropts(int filt, struct filterops *filtops)
1055 {
1056 	int error;
1057 
1058 	error = 0;
1059 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1060 		printf(
1061 "trying to add a filterop that is out of range: %d is beyond %d\n",
1062 		    ~filt, EVFILT_SYSCOUNT);
1063 		return EINVAL;
1064 	}
1065 	mtx_lock(&filterops_lock);
1066 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1067 	    sysfilt_ops[~filt].for_fop != NULL)
1068 		error = EEXIST;
1069 	else {
1070 		sysfilt_ops[~filt].for_fop = filtops;
1071 		sysfilt_ops[~filt].for_refcnt = 0;
1072 	}
1073 	mtx_unlock(&filterops_lock);
1074 
1075 	return (error);
1076 }
1077 
1078 int
1079 kqueue_del_filteropts(int filt)
1080 {
1081 	int error;
1082 
1083 	error = 0;
1084 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1085 		return EINVAL;
1086 
1087 	mtx_lock(&filterops_lock);
1088 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1089 	    sysfilt_ops[~filt].for_fop == NULL)
1090 		error = EINVAL;
1091 	else if (sysfilt_ops[~filt].for_refcnt != 0)
1092 		error = EBUSY;
1093 	else {
1094 		sysfilt_ops[~filt].for_fop = &null_filtops;
1095 		sysfilt_ops[~filt].for_refcnt = 0;
1096 	}
1097 	mtx_unlock(&filterops_lock);
1098 
1099 	return error;
1100 }
1101 
1102 static struct filterops *
1103 kqueue_fo_find(int filt)
1104 {
1105 
1106 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1107 		return NULL;
1108 
1109 	if (sysfilt_ops[~filt].for_nolock)
1110 		return sysfilt_ops[~filt].for_fop;
1111 
1112 	mtx_lock(&filterops_lock);
1113 	sysfilt_ops[~filt].for_refcnt++;
1114 	if (sysfilt_ops[~filt].for_fop == NULL)
1115 		sysfilt_ops[~filt].for_fop = &null_filtops;
1116 	mtx_unlock(&filterops_lock);
1117 
1118 	return sysfilt_ops[~filt].for_fop;
1119 }
1120 
1121 static void
1122 kqueue_fo_release(int filt)
1123 {
1124 
1125 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1126 		return;
1127 
1128 	if (sysfilt_ops[~filt].for_nolock)
1129 		return;
1130 
1131 	mtx_lock(&filterops_lock);
1132 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1133 	    ("filter object refcount not valid on release"));
1134 	sysfilt_ops[~filt].for_refcnt--;
1135 	mtx_unlock(&filterops_lock);
1136 }
1137 
1138 /*
1139  * A ref to kq (obtained via kqueue_acquire) must be held.  waitok will
1140  * influence if memory allocation should wait.  Make sure it is 0 if you
1141  * hold any mutexes.
1142  */
1143 static int
1144 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1145 {
1146 	struct filterops *fops;
1147 	struct file *fp;
1148 	struct knote *kn, *tkn;
1149 	struct knlist *knl;
1150 	cap_rights_t rights;
1151 	int error, filt, event;
1152 	int haskqglobal, filedesc_unlock;
1153 
1154 	if ((kev->flags & (EV_ENABLE | EV_DISABLE)) == (EV_ENABLE | EV_DISABLE))
1155 		return (EINVAL);
1156 
1157 	fp = NULL;
1158 	kn = NULL;
1159 	knl = NULL;
1160 	error = 0;
1161 	haskqglobal = 0;
1162 	filedesc_unlock = 0;
1163 
1164 	filt = kev->filter;
1165 	fops = kqueue_fo_find(filt);
1166 	if (fops == NULL)
1167 		return EINVAL;
1168 
1169 	if (kev->flags & EV_ADD) {
1170 		/*
1171 		 * Prevent waiting with locks.  Non-sleepable
1172 		 * allocation failures are handled in the loop, only
1173 		 * if the spare knote appears to be actually required.
1174 		 */
1175 		tkn = knote_alloc(waitok);
1176 	} else {
1177 		tkn = NULL;
1178 	}
1179 
1180 findkn:
1181 	if (fops->f_isfd) {
1182 		KASSERT(td != NULL, ("td is NULL"));
1183 		error = fget(td, kev->ident,
1184 		    cap_rights_init(&rights, CAP_EVENT), &fp);
1185 		if (error)
1186 			goto done;
1187 
1188 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1189 		    kev->ident, 0) != 0) {
1190 			/* try again */
1191 			fdrop(fp, td);
1192 			fp = NULL;
1193 			error = kqueue_expand(kq, fops, kev->ident, waitok);
1194 			if (error)
1195 				goto done;
1196 			goto findkn;
1197 		}
1198 
1199 		if (fp->f_type == DTYPE_KQUEUE) {
1200 			/*
1201 			 * If we add some intelligence about what we are doing,
1202 			 * we should be able to support events on ourselves.
1203 			 * We need to know when we are doing this to prevent
1204 			 * getting both the knlist lock and the kq lock since
1205 			 * they are the same thing.
1206 			 */
1207 			if (fp->f_data == kq) {
1208 				error = EINVAL;
1209 				goto done;
1210 			}
1211 
1212 			/*
1213 			 * Pre-lock the filedesc before the global
1214 			 * lock mutex, see the comment in
1215 			 * kqueue_close().
1216 			 */
1217 			FILEDESC_XLOCK(td->td_proc->p_fd);
1218 			filedesc_unlock = 1;
1219 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1220 		}
1221 
1222 		KQ_LOCK(kq);
1223 		if (kev->ident < kq->kq_knlistsize) {
1224 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1225 				if (kev->filter == kn->kn_filter)
1226 					break;
1227 		}
1228 	} else {
1229 		if ((kev->flags & EV_ADD) == EV_ADD)
1230 			kqueue_expand(kq, fops, kev->ident, waitok);
1231 
1232 		KQ_LOCK(kq);
1233 
1234 		/*
1235 		 * If possible, find an existing knote to use for this kevent.
1236 		 */
1237 		if (kev->filter == EVFILT_PROC &&
1238 		    (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1239 			/* This is an internal creation of a process tracking
1240 			 * note. Don't attempt to coalesce this with an
1241 			 * existing note.
1242 			 */
1243 			;
1244 		} else if (kq->kq_knhashmask != 0) {
1245 			struct klist *list;
1246 
1247 			list = &kq->kq_knhash[
1248 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1249 			SLIST_FOREACH(kn, list, kn_link)
1250 				if (kev->ident == kn->kn_id &&
1251 				    kev->filter == kn->kn_filter)
1252 					break;
1253 		}
1254 	}
1255 
1256 	/* knote is in the process of changing, wait for it to stabilize. */
1257 	if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1258 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1259 		if (filedesc_unlock) {
1260 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1261 			filedesc_unlock = 0;
1262 		}
1263 		kq->kq_state |= KQ_FLUXWAIT;
1264 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1265 		if (fp != NULL) {
1266 			fdrop(fp, td);
1267 			fp = NULL;
1268 		}
1269 		goto findkn;
1270 	}
1271 
1272 	/*
1273 	 * kn now contains the matching knote, or NULL if no match
1274 	 */
1275 	if (kn == NULL) {
1276 		if (kev->flags & EV_ADD) {
1277 			kn = tkn;
1278 			tkn = NULL;
1279 			if (kn == NULL) {
1280 				KQ_UNLOCK(kq);
1281 				error = ENOMEM;
1282 				goto done;
1283 			}
1284 			kn->kn_fp = fp;
1285 			kn->kn_kq = kq;
1286 			kn->kn_fop = fops;
1287 			/*
1288 			 * apply reference counts to knote structure, and
1289 			 * do not release it at the end of this routine.
1290 			 */
1291 			fops = NULL;
1292 			fp = NULL;
1293 
1294 			kn->kn_sfflags = kev->fflags;
1295 			kn->kn_sdata = kev->data;
1296 			kev->fflags = 0;
1297 			kev->data = 0;
1298 			kn->kn_kevent = *kev;
1299 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1300 			    EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1301 			kn->kn_status = KN_INFLUX|KN_DETACHED;
1302 
1303 			error = knote_attach(kn, kq);
1304 			KQ_UNLOCK(kq);
1305 			if (error != 0) {
1306 				tkn = kn;
1307 				goto done;
1308 			}
1309 
1310 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1311 				knote_drop(kn, td);
1312 				goto done;
1313 			}
1314 			knl = kn_list_lock(kn);
1315 			goto done_ev_add;
1316 		} else {
1317 			/* No matching knote and the EV_ADD flag is not set. */
1318 			KQ_UNLOCK(kq);
1319 			error = ENOENT;
1320 			goto done;
1321 		}
1322 	}
1323 
1324 	if (kev->flags & EV_DELETE) {
1325 		kn->kn_status |= KN_INFLUX;
1326 		KQ_UNLOCK(kq);
1327 		if (!(kn->kn_status & KN_DETACHED))
1328 			kn->kn_fop->f_detach(kn);
1329 		knote_drop(kn, td);
1330 		goto done;
1331 	}
1332 
1333 	if (kev->flags & EV_FORCEONESHOT) {
1334 		kn->kn_flags |= EV_ONESHOT;
1335 		KNOTE_ACTIVATE(kn, 1);
1336 	}
1337 
1338 	/*
1339 	 * The user may change some filter values after the initial EV_ADD,
1340 	 * but doing so will not reset any filter which has already been
1341 	 * triggered.
1342 	 */
1343 	kn->kn_status |= KN_INFLUX | KN_SCAN;
1344 	KQ_UNLOCK(kq);
1345 	knl = kn_list_lock(kn);
1346 	kn->kn_kevent.udata = kev->udata;
1347 	if (!fops->f_isfd && fops->f_touch != NULL) {
1348 		fops->f_touch(kn, kev, EVENT_REGISTER);
1349 	} else {
1350 		kn->kn_sfflags = kev->fflags;
1351 		kn->kn_sdata = kev->data;
1352 	}
1353 
1354 	/*
1355 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1356 	 * the initial attach event decides that the event is "completed"
1357 	 * already.  i.e. filt_procattach is called on a zombie process.  It
1358 	 * will call filt_proc which will remove it from the list, and NULL
1359 	 * kn_knlist.
1360 	 */
1361 done_ev_add:
1362 	if ((kev->flags & EV_ENABLE) != 0)
1363 		kn->kn_status &= ~KN_DISABLED;
1364 	else if ((kev->flags & EV_DISABLE) != 0)
1365 		kn->kn_status |= KN_DISABLED;
1366 
1367 	if ((kn->kn_status & KN_DISABLED) == 0)
1368 		event = kn->kn_fop->f_event(kn, 0);
1369 	else
1370 		event = 0;
1371 
1372 	KQ_LOCK(kq);
1373 	if (event)
1374 		kn->kn_status |= KN_ACTIVE;
1375 	if ((kn->kn_status & (KN_ACTIVE | KN_DISABLED | KN_QUEUED)) ==
1376 	    KN_ACTIVE)
1377 		knote_enqueue(kn);
1378 	kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1379 	kn_list_unlock(knl);
1380 	KQ_UNLOCK_FLUX(kq);
1381 
1382 done:
1383 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1384 	if (filedesc_unlock)
1385 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1386 	if (fp != NULL)
1387 		fdrop(fp, td);
1388 	knote_free(tkn);
1389 	if (fops != NULL)
1390 		kqueue_fo_release(filt);
1391 	return (error);
1392 }
1393 
1394 static int
1395 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1396 {
1397 	int error;
1398 	struct kqueue *kq;
1399 
1400 	error = 0;
1401 
1402 	kq = fp->f_data;
1403 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1404 		return (EBADF);
1405 	*kqp = kq;
1406 	KQ_LOCK(kq);
1407 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1408 		KQ_UNLOCK(kq);
1409 		return (EBADF);
1410 	}
1411 	kq->kq_refcnt++;
1412 	KQ_UNLOCK(kq);
1413 
1414 	return error;
1415 }
1416 
1417 static void
1418 kqueue_release(struct kqueue *kq, int locked)
1419 {
1420 	if (locked)
1421 		KQ_OWNED(kq);
1422 	else
1423 		KQ_LOCK(kq);
1424 	kq->kq_refcnt--;
1425 	if (kq->kq_refcnt == 1)
1426 		wakeup(&kq->kq_refcnt);
1427 	if (!locked)
1428 		KQ_UNLOCK(kq);
1429 }
1430 
1431 static void
1432 kqueue_schedtask(struct kqueue *kq)
1433 {
1434 
1435 	KQ_OWNED(kq);
1436 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1437 	    ("scheduling kqueue task while draining"));
1438 
1439 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1440 		taskqueue_enqueue(taskqueue_kqueue_ctx, &kq->kq_task);
1441 		kq->kq_state |= KQ_TASKSCHED;
1442 	}
1443 }
1444 
1445 /*
1446  * Expand the kq to make sure we have storage for fops/ident pair.
1447  *
1448  * Return 0 on success (or no work necessary), return errno on failure.
1449  *
1450  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1451  * If kqueue_register is called from a non-fd context, there usually/should
1452  * be no locks held.
1453  */
1454 static int
1455 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1456 	int waitok)
1457 {
1458 	struct klist *list, *tmp_knhash, *to_free;
1459 	u_long tmp_knhashmask;
1460 	int size;
1461 	int fd;
1462 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
1463 
1464 	KQ_NOTOWNED(kq);
1465 
1466 	to_free = NULL;
1467 	if (fops->f_isfd) {
1468 		fd = ident;
1469 		if (kq->kq_knlistsize <= fd) {
1470 			size = kq->kq_knlistsize;
1471 			while (size <= fd)
1472 				size += KQEXTENT;
1473 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1474 			if (list == NULL)
1475 				return ENOMEM;
1476 			KQ_LOCK(kq);
1477 			if (kq->kq_knlistsize > fd) {
1478 				to_free = list;
1479 				list = NULL;
1480 			} else {
1481 				if (kq->kq_knlist != NULL) {
1482 					bcopy(kq->kq_knlist, list,
1483 					    kq->kq_knlistsize * sizeof(*list));
1484 					to_free = kq->kq_knlist;
1485 					kq->kq_knlist = NULL;
1486 				}
1487 				bzero((caddr_t)list +
1488 				    kq->kq_knlistsize * sizeof(*list),
1489 				    (size - kq->kq_knlistsize) * sizeof(*list));
1490 				kq->kq_knlistsize = size;
1491 				kq->kq_knlist = list;
1492 			}
1493 			KQ_UNLOCK(kq);
1494 		}
1495 	} else {
1496 		if (kq->kq_knhashmask == 0) {
1497 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1498 			    &tmp_knhashmask);
1499 			if (tmp_knhash == NULL)
1500 				return ENOMEM;
1501 			KQ_LOCK(kq);
1502 			if (kq->kq_knhashmask == 0) {
1503 				kq->kq_knhash = tmp_knhash;
1504 				kq->kq_knhashmask = tmp_knhashmask;
1505 			} else {
1506 				to_free = tmp_knhash;
1507 			}
1508 			KQ_UNLOCK(kq);
1509 		}
1510 	}
1511 	free(to_free, M_KQUEUE);
1512 
1513 	KQ_NOTOWNED(kq);
1514 	return 0;
1515 }
1516 
1517 static void
1518 kqueue_task(void *arg, int pending)
1519 {
1520 	struct kqueue *kq;
1521 	int haskqglobal;
1522 
1523 	haskqglobal = 0;
1524 	kq = arg;
1525 
1526 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1527 	KQ_LOCK(kq);
1528 
1529 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1530 
1531 	kq->kq_state &= ~KQ_TASKSCHED;
1532 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1533 		wakeup(&kq->kq_state);
1534 	}
1535 	KQ_UNLOCK(kq);
1536 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1537 }
1538 
1539 /*
1540  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1541  * We treat KN_MARKER knotes as if they are INFLUX.
1542  */
1543 static int
1544 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1545     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1546 {
1547 	struct kevent *kevp;
1548 	struct knote *kn, *marker;
1549 	struct knlist *knl;
1550 	sbintime_t asbt, rsbt;
1551 	int count, error, haskqglobal, influx, nkev, touch;
1552 
1553 	count = maxevents;
1554 	nkev = 0;
1555 	error = 0;
1556 	haskqglobal = 0;
1557 
1558 	if (maxevents == 0)
1559 		goto done_nl;
1560 
1561 	rsbt = 0;
1562 	if (tsp != NULL) {
1563 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1564 		    tsp->tv_nsec >= 1000000000) {
1565 			error = EINVAL;
1566 			goto done_nl;
1567 		}
1568 		if (timespecisset(tsp)) {
1569 			if (tsp->tv_sec <= INT32_MAX) {
1570 				rsbt = tstosbt(*tsp);
1571 				if (TIMESEL(&asbt, rsbt))
1572 					asbt += tc_tick_sbt;
1573 				if (asbt <= SBT_MAX - rsbt)
1574 					asbt += rsbt;
1575 				else
1576 					asbt = 0;
1577 				rsbt >>= tc_precexp;
1578 			} else
1579 				asbt = 0;
1580 		} else
1581 			asbt = -1;
1582 	} else
1583 		asbt = 0;
1584 	marker = knote_alloc(1);
1585 	marker->kn_status = KN_MARKER;
1586 	KQ_LOCK(kq);
1587 
1588 retry:
1589 	kevp = keva;
1590 	if (kq->kq_count == 0) {
1591 		if (asbt == -1) {
1592 			error = EWOULDBLOCK;
1593 		} else {
1594 			kq->kq_state |= KQ_SLEEP;
1595 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1596 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1597 		}
1598 		if (error == 0)
1599 			goto retry;
1600 		/* don't restart after signals... */
1601 		if (error == ERESTART)
1602 			error = EINTR;
1603 		else if (error == EWOULDBLOCK)
1604 			error = 0;
1605 		goto done;
1606 	}
1607 
1608 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1609 	influx = 0;
1610 	while (count) {
1611 		KQ_OWNED(kq);
1612 		kn = TAILQ_FIRST(&kq->kq_head);
1613 
1614 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1615 		    (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1616 			if (influx) {
1617 				influx = 0;
1618 				KQ_FLUX_WAKEUP(kq);
1619 			}
1620 			kq->kq_state |= KQ_FLUXWAIT;
1621 			error = msleep(kq, &kq->kq_lock, PSOCK,
1622 			    "kqflxwt", 0);
1623 			continue;
1624 		}
1625 
1626 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1627 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1628 			kn->kn_status &= ~KN_QUEUED;
1629 			kq->kq_count--;
1630 			continue;
1631 		}
1632 		if (kn == marker) {
1633 			KQ_FLUX_WAKEUP(kq);
1634 			if (count == maxevents)
1635 				goto retry;
1636 			goto done;
1637 		}
1638 		KASSERT((kn->kn_status & KN_INFLUX) == 0,
1639 		    ("KN_INFLUX set when not suppose to be"));
1640 
1641 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1642 			kn->kn_status &= ~KN_QUEUED;
1643 			kn->kn_status |= KN_INFLUX;
1644 			kq->kq_count--;
1645 			KQ_UNLOCK(kq);
1646 			/*
1647 			 * We don't need to lock the list since we've marked
1648 			 * it _INFLUX.
1649 			 */
1650 			if (!(kn->kn_status & KN_DETACHED))
1651 				kn->kn_fop->f_detach(kn);
1652 			knote_drop(kn, td);
1653 			KQ_LOCK(kq);
1654 			continue;
1655 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1656 			kn->kn_status &= ~KN_QUEUED;
1657 			kn->kn_status |= KN_INFLUX;
1658 			kq->kq_count--;
1659 			KQ_UNLOCK(kq);
1660 			/*
1661 			 * We don't need to lock the list since we've marked
1662 			 * it _INFLUX.
1663 			 */
1664 			*kevp = kn->kn_kevent;
1665 			if (!(kn->kn_status & KN_DETACHED))
1666 				kn->kn_fop->f_detach(kn);
1667 			knote_drop(kn, td);
1668 			KQ_LOCK(kq);
1669 			kn = NULL;
1670 		} else {
1671 			kn->kn_status |= KN_INFLUX | KN_SCAN;
1672 			KQ_UNLOCK(kq);
1673 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1674 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1675 			knl = kn_list_lock(kn);
1676 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1677 				KQ_LOCK(kq);
1678 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1679 				kn->kn_status &=
1680 				    ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1681 				    KN_SCAN);
1682 				kq->kq_count--;
1683 				kn_list_unlock(knl);
1684 				influx = 1;
1685 				continue;
1686 			}
1687 			touch = (!kn->kn_fop->f_isfd &&
1688 			    kn->kn_fop->f_touch != NULL);
1689 			if (touch)
1690 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1691 			else
1692 				*kevp = kn->kn_kevent;
1693 			KQ_LOCK(kq);
1694 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1695 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1696 				/*
1697 				 * Manually clear knotes who weren't
1698 				 * 'touch'ed.
1699 				 */
1700 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1701 					kn->kn_data = 0;
1702 					kn->kn_fflags = 0;
1703 				}
1704 				if (kn->kn_flags & EV_DISPATCH)
1705 					kn->kn_status |= KN_DISABLED;
1706 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1707 				kq->kq_count--;
1708 			} else
1709 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1710 
1711 			kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1712 			kn_list_unlock(knl);
1713 			influx = 1;
1714 		}
1715 
1716 		/* we are returning a copy to the user */
1717 		kevp++;
1718 		nkev++;
1719 		count--;
1720 
1721 		if (nkev == KQ_NEVENTS) {
1722 			influx = 0;
1723 			KQ_UNLOCK_FLUX(kq);
1724 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1725 			nkev = 0;
1726 			kevp = keva;
1727 			KQ_LOCK(kq);
1728 			if (error)
1729 				break;
1730 		}
1731 	}
1732 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1733 done:
1734 	KQ_OWNED(kq);
1735 	KQ_UNLOCK_FLUX(kq);
1736 	knote_free(marker);
1737 done_nl:
1738 	KQ_NOTOWNED(kq);
1739 	if (nkev != 0)
1740 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1741 	td->td_retval[0] = maxevents - count;
1742 	return (error);
1743 }
1744 
1745 /*ARGSUSED*/
1746 static int
1747 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1748 	struct ucred *active_cred, struct thread *td)
1749 {
1750 	/*
1751 	 * Enabling sigio causes two major problems:
1752 	 * 1) infinite recursion:
1753 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1754 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1755 	 * into itself over and over.  Sending the sigio causes the kqueue
1756 	 * to become ready, which in turn posts sigio again, forever.
1757 	 * Solution: this can be solved by setting a flag in the kqueue that
1758 	 * we have a SIGIO in progress.
1759 	 * 2) locking problems:
1760 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1761 	 * us above the proc and pgrp locks.
1762 	 * Solution: Post a signal using an async mechanism, being sure to
1763 	 * record a generation count in the delivery so that we do not deliver
1764 	 * a signal to the wrong process.
1765 	 *
1766 	 * Note, these two mechanisms are somewhat mutually exclusive!
1767 	 */
1768 #if 0
1769 	struct kqueue *kq;
1770 
1771 	kq = fp->f_data;
1772 	switch (cmd) {
1773 	case FIOASYNC:
1774 		if (*(int *)data) {
1775 			kq->kq_state |= KQ_ASYNC;
1776 		} else {
1777 			kq->kq_state &= ~KQ_ASYNC;
1778 		}
1779 		return (0);
1780 
1781 	case FIOSETOWN:
1782 		return (fsetown(*(int *)data, &kq->kq_sigio));
1783 
1784 	case FIOGETOWN:
1785 		*(int *)data = fgetown(&kq->kq_sigio);
1786 		return (0);
1787 	}
1788 #endif
1789 
1790 	return (ENOTTY);
1791 }
1792 
1793 /*ARGSUSED*/
1794 static int
1795 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1796 	struct thread *td)
1797 {
1798 	struct kqueue *kq;
1799 	int revents = 0;
1800 	int error;
1801 
1802 	if ((error = kqueue_acquire(fp, &kq)))
1803 		return POLLERR;
1804 
1805 	KQ_LOCK(kq);
1806 	if (events & (POLLIN | POLLRDNORM)) {
1807 		if (kq->kq_count) {
1808 			revents |= events & (POLLIN | POLLRDNORM);
1809 		} else {
1810 			selrecord(td, &kq->kq_sel);
1811 			if (SEL_WAITING(&kq->kq_sel))
1812 				kq->kq_state |= KQ_SEL;
1813 		}
1814 	}
1815 	kqueue_release(kq, 1);
1816 	KQ_UNLOCK(kq);
1817 	return (revents);
1818 }
1819 
1820 /*ARGSUSED*/
1821 static int
1822 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1823 	struct thread *td)
1824 {
1825 
1826 	bzero((void *)st, sizeof *st);
1827 	/*
1828 	 * We no longer return kq_count because the unlocked value is useless.
1829 	 * If you spent all this time getting the count, why not spend your
1830 	 * syscall better by calling kevent?
1831 	 *
1832 	 * XXX - This is needed for libc_r.
1833 	 */
1834 	st->st_mode = S_IFIFO;
1835 	return (0);
1836 }
1837 
1838 static void
1839 kqueue_drain(struct kqueue *kq, struct thread *td)
1840 {
1841 	struct knote *kn;
1842 	int i;
1843 
1844 	KQ_LOCK(kq);
1845 
1846 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1847 	    ("kqueue already closing"));
1848 	kq->kq_state |= KQ_CLOSING;
1849 	if (kq->kq_refcnt > 1)
1850 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1851 
1852 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1853 
1854 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
1855 	    ("kqueue's knlist not empty"));
1856 
1857 	for (i = 0; i < kq->kq_knlistsize; i++) {
1858 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1859 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1860 				kq->kq_state |= KQ_FLUXWAIT;
1861 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1862 				continue;
1863 			}
1864 			kn->kn_status |= KN_INFLUX;
1865 			KQ_UNLOCK(kq);
1866 			if (!(kn->kn_status & KN_DETACHED))
1867 				kn->kn_fop->f_detach(kn);
1868 			knote_drop(kn, td);
1869 			KQ_LOCK(kq);
1870 		}
1871 	}
1872 	if (kq->kq_knhashmask != 0) {
1873 		for (i = 0; i <= kq->kq_knhashmask; i++) {
1874 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1875 				if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1876 					kq->kq_state |= KQ_FLUXWAIT;
1877 					msleep(kq, &kq->kq_lock, PSOCK,
1878 					       "kqclo2", 0);
1879 					continue;
1880 				}
1881 				kn->kn_status |= KN_INFLUX;
1882 				KQ_UNLOCK(kq);
1883 				if (!(kn->kn_status & KN_DETACHED))
1884 					kn->kn_fop->f_detach(kn);
1885 				knote_drop(kn, td);
1886 				KQ_LOCK(kq);
1887 			}
1888 		}
1889 	}
1890 
1891 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1892 		kq->kq_state |= KQ_TASKDRAIN;
1893 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1894 	}
1895 
1896 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1897 		selwakeuppri(&kq->kq_sel, PSOCK);
1898 		if (!SEL_WAITING(&kq->kq_sel))
1899 			kq->kq_state &= ~KQ_SEL;
1900 	}
1901 
1902 	KQ_UNLOCK(kq);
1903 }
1904 
1905 static void
1906 kqueue_destroy(struct kqueue *kq)
1907 {
1908 
1909 	KASSERT(kq->kq_fdp == NULL,
1910 	    ("kqueue still attached to a file descriptor"));
1911 	seldrain(&kq->kq_sel);
1912 	knlist_destroy(&kq->kq_sel.si_note);
1913 	mtx_destroy(&kq->kq_lock);
1914 
1915 	if (kq->kq_knhash != NULL)
1916 		free(kq->kq_knhash, M_KQUEUE);
1917 	if (kq->kq_knlist != NULL)
1918 		free(kq->kq_knlist, M_KQUEUE);
1919 
1920 	funsetown(&kq->kq_sigio);
1921 }
1922 
1923 /*ARGSUSED*/
1924 static int
1925 kqueue_close(struct file *fp, struct thread *td)
1926 {
1927 	struct kqueue *kq = fp->f_data;
1928 	struct filedesc *fdp;
1929 	int error;
1930 	int filedesc_unlock;
1931 
1932 	if ((error = kqueue_acquire(fp, &kq)))
1933 		return error;
1934 	kqueue_drain(kq, td);
1935 
1936 	/*
1937 	 * We could be called due to the knote_drop() doing fdrop(),
1938 	 * called from kqueue_register().  In this case the global
1939 	 * lock is owned, and filedesc sx is locked before, to not
1940 	 * take the sleepable lock after non-sleepable.
1941 	 */
1942 	fdp = kq->kq_fdp;
1943 	kq->kq_fdp = NULL;
1944 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1945 		FILEDESC_XLOCK(fdp);
1946 		filedesc_unlock = 1;
1947 	} else
1948 		filedesc_unlock = 0;
1949 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1950 	if (filedesc_unlock)
1951 		FILEDESC_XUNLOCK(fdp);
1952 
1953 	kqueue_destroy(kq);
1954 	chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
1955 	crfree(kq->kq_cred);
1956 	free(kq, M_KQUEUE);
1957 	fp->f_data = NULL;
1958 
1959 	return (0);
1960 }
1961 
1962 static int
1963 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1964 {
1965 
1966 	kif->kf_type = KF_TYPE_KQUEUE;
1967 	return (0);
1968 }
1969 
1970 static void
1971 kqueue_wakeup(struct kqueue *kq)
1972 {
1973 	KQ_OWNED(kq);
1974 
1975 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1976 		kq->kq_state &= ~KQ_SLEEP;
1977 		wakeup(kq);
1978 	}
1979 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1980 		selwakeuppri(&kq->kq_sel, PSOCK);
1981 		if (!SEL_WAITING(&kq->kq_sel))
1982 			kq->kq_state &= ~KQ_SEL;
1983 	}
1984 	if (!knlist_empty(&kq->kq_sel.si_note))
1985 		kqueue_schedtask(kq);
1986 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1987 		pgsigio(&kq->kq_sigio, SIGIO, 0);
1988 	}
1989 }
1990 
1991 /*
1992  * Walk down a list of knotes, activating them if their event has triggered.
1993  *
1994  * There is a possibility to optimize in the case of one kq watching another.
1995  * Instead of scheduling a task to wake it up, you could pass enough state
1996  * down the chain to make up the parent kqueue.  Make this code functional
1997  * first.
1998  */
1999 void
2000 knote(struct knlist *list, long hint, int lockflags)
2001 {
2002 	struct kqueue *kq;
2003 	struct knote *kn, *tkn;
2004 	int error;
2005 
2006 	if (list == NULL)
2007 		return;
2008 
2009 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
2010 
2011 	if ((lockflags & KNF_LISTLOCKED) == 0)
2012 		list->kl_lock(list->kl_lockarg);
2013 
2014 	/*
2015 	 * If we unlock the list lock (and set KN_INFLUX), we can
2016 	 * eliminate the kqueue scheduling, but this will introduce
2017 	 * four lock/unlock's for each knote to test.  Also, marker
2018 	 * would be needed to keep iteration position, since filters
2019 	 * or other threads could remove events.
2020 	 */
2021 	SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
2022 		kq = kn->kn_kq;
2023 		KQ_LOCK(kq);
2024 		if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
2025 			/*
2026 			 * Do not process the influx notes, except for
2027 			 * the influx coming from the kq unlock in the
2028 			 * kqueue_scan().  In the later case, we do
2029 			 * not interfere with the scan, since the code
2030 			 * fragment in kqueue_scan() locks the knlist,
2031 			 * and cannot proceed until we finished.
2032 			 */
2033 			KQ_UNLOCK(kq);
2034 		} else if ((lockflags & KNF_NOKQLOCK) != 0) {
2035 			kn->kn_status |= KN_INFLUX;
2036 			KQ_UNLOCK(kq);
2037 			error = kn->kn_fop->f_event(kn, hint);
2038 			KQ_LOCK(kq);
2039 			kn->kn_status &= ~KN_INFLUX;
2040 			if (error)
2041 				KNOTE_ACTIVATE(kn, 1);
2042 			KQ_UNLOCK_FLUX(kq);
2043 		} else {
2044 			kn->kn_status |= KN_HASKQLOCK;
2045 			if (kn->kn_fop->f_event(kn, hint))
2046 				KNOTE_ACTIVATE(kn, 1);
2047 			kn->kn_status &= ~KN_HASKQLOCK;
2048 			KQ_UNLOCK(kq);
2049 		}
2050 	}
2051 	if ((lockflags & KNF_LISTLOCKED) == 0)
2052 		list->kl_unlock(list->kl_lockarg);
2053 }
2054 
2055 /*
2056  * add a knote to a knlist
2057  */
2058 void
2059 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2060 {
2061 	KNL_ASSERT_LOCK(knl, islocked);
2062 	KQ_NOTOWNED(kn->kn_kq);
2063 	KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
2064 	    (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
2065 	if (!islocked)
2066 		knl->kl_lock(knl->kl_lockarg);
2067 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2068 	if (!islocked)
2069 		knl->kl_unlock(knl->kl_lockarg);
2070 	KQ_LOCK(kn->kn_kq);
2071 	kn->kn_knlist = knl;
2072 	kn->kn_status &= ~KN_DETACHED;
2073 	KQ_UNLOCK(kn->kn_kq);
2074 }
2075 
2076 static void
2077 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked,
2078     int kqislocked)
2079 {
2080 	KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
2081 	KNL_ASSERT_LOCK(knl, knlislocked);
2082 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2083 	if (!kqislocked)
2084 		KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
2085     ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
2086 	if (!knlislocked)
2087 		knl->kl_lock(knl->kl_lockarg);
2088 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2089 	kn->kn_knlist = NULL;
2090 	if (!knlislocked)
2091 		kn_list_unlock(knl);
2092 	if (!kqislocked)
2093 		KQ_LOCK(kn->kn_kq);
2094 	kn->kn_status |= KN_DETACHED;
2095 	if (!kqislocked)
2096 		KQ_UNLOCK(kn->kn_kq);
2097 }
2098 
2099 /*
2100  * remove knote from the specified knlist
2101  */
2102 void
2103 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2104 {
2105 
2106 	knlist_remove_kq(knl, kn, islocked, 0);
2107 }
2108 
2109 int
2110 knlist_empty(struct knlist *knl)
2111 {
2112 
2113 	KNL_ASSERT_LOCKED(knl);
2114 	return SLIST_EMPTY(&knl->kl_list);
2115 }
2116 
2117 static struct mtx	knlist_lock;
2118 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2119 	MTX_DEF);
2120 static void knlist_mtx_lock(void *arg);
2121 static void knlist_mtx_unlock(void *arg);
2122 
2123 static void
2124 knlist_mtx_lock(void *arg)
2125 {
2126 
2127 	mtx_lock((struct mtx *)arg);
2128 }
2129 
2130 static void
2131 knlist_mtx_unlock(void *arg)
2132 {
2133 
2134 	mtx_unlock((struct mtx *)arg);
2135 }
2136 
2137 static void
2138 knlist_mtx_assert_locked(void *arg)
2139 {
2140 
2141 	mtx_assert((struct mtx *)arg, MA_OWNED);
2142 }
2143 
2144 static void
2145 knlist_mtx_assert_unlocked(void *arg)
2146 {
2147 
2148 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2149 }
2150 
2151 static void
2152 knlist_rw_rlock(void *arg)
2153 {
2154 
2155 	rw_rlock((struct rwlock *)arg);
2156 }
2157 
2158 static void
2159 knlist_rw_runlock(void *arg)
2160 {
2161 
2162 	rw_runlock((struct rwlock *)arg);
2163 }
2164 
2165 static void
2166 knlist_rw_assert_locked(void *arg)
2167 {
2168 
2169 	rw_assert((struct rwlock *)arg, RA_LOCKED);
2170 }
2171 
2172 static void
2173 knlist_rw_assert_unlocked(void *arg)
2174 {
2175 
2176 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2177 }
2178 
2179 void
2180 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2181     void (*kl_unlock)(void *),
2182     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2183 {
2184 
2185 	if (lock == NULL)
2186 		knl->kl_lockarg = &knlist_lock;
2187 	else
2188 		knl->kl_lockarg = lock;
2189 
2190 	if (kl_lock == NULL)
2191 		knl->kl_lock = knlist_mtx_lock;
2192 	else
2193 		knl->kl_lock = kl_lock;
2194 	if (kl_unlock == NULL)
2195 		knl->kl_unlock = knlist_mtx_unlock;
2196 	else
2197 		knl->kl_unlock = kl_unlock;
2198 	if (kl_assert_locked == NULL)
2199 		knl->kl_assert_locked = knlist_mtx_assert_locked;
2200 	else
2201 		knl->kl_assert_locked = kl_assert_locked;
2202 	if (kl_assert_unlocked == NULL)
2203 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2204 	else
2205 		knl->kl_assert_unlocked = kl_assert_unlocked;
2206 
2207 	knl->kl_autodestroy = 0;
2208 	SLIST_INIT(&knl->kl_list);
2209 }
2210 
2211 void
2212 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2213 {
2214 
2215 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2216 }
2217 
2218 struct knlist *
2219 knlist_alloc(struct mtx *lock)
2220 {
2221 	struct knlist *knl;
2222 
2223 	knl = malloc(sizeof(struct knlist), M_KQUEUE, M_WAITOK);
2224 	knlist_init_mtx(knl, lock);
2225 	return (knl);
2226 }
2227 
2228 void
2229 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2230 {
2231 
2232 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2233 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2234 }
2235 
2236 void
2237 knlist_destroy(struct knlist *knl)
2238 {
2239 
2240 #ifdef INVARIANTS
2241 	/*
2242 	 * if we run across this error, we need to find the offending
2243 	 * driver and have it call knlist_clear or knlist_delete.
2244 	 */
2245 	if (!SLIST_EMPTY(&knl->kl_list))
2246 		printf("WARNING: destroying knlist w/ knotes on it!\n");
2247 #endif
2248 
2249 	knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2250 	SLIST_INIT(&knl->kl_list);
2251 }
2252 
2253 void
2254 knlist_detach(struct knlist *knl)
2255 {
2256 
2257 	KNL_ASSERT_LOCKED(knl);
2258 	knl->kl_autodestroy = 1;
2259 	if (knlist_empty(knl)) {
2260 		knlist_destroy(knl);
2261 		free(knl, M_KQUEUE);
2262 	}
2263 }
2264 
2265 /*
2266  * Even if we are locked, we may need to drop the lock to allow any influx
2267  * knotes time to "settle".
2268  */
2269 void
2270 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2271 {
2272 	struct knote *kn, *kn2;
2273 	struct kqueue *kq;
2274 
2275 	KASSERT(!knl->kl_autodestroy, ("cleardel for autodestroy %p", knl));
2276 	if (islocked)
2277 		KNL_ASSERT_LOCKED(knl);
2278 	else {
2279 		KNL_ASSERT_UNLOCKED(knl);
2280 again:		/* need to reacquire lock since we have dropped it */
2281 		knl->kl_lock(knl->kl_lockarg);
2282 	}
2283 
2284 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2285 		kq = kn->kn_kq;
2286 		KQ_LOCK(kq);
2287 		if ((kn->kn_status & KN_INFLUX)) {
2288 			KQ_UNLOCK(kq);
2289 			continue;
2290 		}
2291 		knlist_remove_kq(knl, kn, 1, 1);
2292 		if (killkn) {
2293 			kn->kn_status |= KN_INFLUX | KN_DETACHED;
2294 			KQ_UNLOCK(kq);
2295 			knote_drop(kn, td);
2296 		} else {
2297 			/* Make sure cleared knotes disappear soon */
2298 			kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2299 			KQ_UNLOCK(kq);
2300 		}
2301 		kq = NULL;
2302 	}
2303 
2304 	if (!SLIST_EMPTY(&knl->kl_list)) {
2305 		/* there are still KN_INFLUX remaining */
2306 		kn = SLIST_FIRST(&knl->kl_list);
2307 		kq = kn->kn_kq;
2308 		KQ_LOCK(kq);
2309 		KASSERT(kn->kn_status & KN_INFLUX,
2310 		    ("knote removed w/o list lock"));
2311 		knl->kl_unlock(knl->kl_lockarg);
2312 		kq->kq_state |= KQ_FLUXWAIT;
2313 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2314 		kq = NULL;
2315 		goto again;
2316 	}
2317 
2318 	if (islocked)
2319 		KNL_ASSERT_LOCKED(knl);
2320 	else {
2321 		knl->kl_unlock(knl->kl_lockarg);
2322 		KNL_ASSERT_UNLOCKED(knl);
2323 	}
2324 }
2325 
2326 /*
2327  * Remove all knotes referencing a specified fd must be called with FILEDESC
2328  * lock.  This prevents a race where a new fd comes along and occupies the
2329  * entry and we attach a knote to the fd.
2330  */
2331 void
2332 knote_fdclose(struct thread *td, int fd)
2333 {
2334 	struct filedesc *fdp = td->td_proc->p_fd;
2335 	struct kqueue *kq;
2336 	struct knote *kn;
2337 	int influx;
2338 
2339 	FILEDESC_XLOCK_ASSERT(fdp);
2340 
2341 	/*
2342 	 * We shouldn't have to worry about new kevents appearing on fd
2343 	 * since filedesc is locked.
2344 	 */
2345 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2346 		KQ_LOCK(kq);
2347 
2348 again:
2349 		influx = 0;
2350 		while (kq->kq_knlistsize > fd &&
2351 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2352 			if (kn->kn_status & KN_INFLUX) {
2353 				/* someone else might be waiting on our knote */
2354 				if (influx)
2355 					wakeup(kq);
2356 				kq->kq_state |= KQ_FLUXWAIT;
2357 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2358 				goto again;
2359 			}
2360 			kn->kn_status |= KN_INFLUX;
2361 			KQ_UNLOCK(kq);
2362 			if (!(kn->kn_status & KN_DETACHED))
2363 				kn->kn_fop->f_detach(kn);
2364 			knote_drop(kn, td);
2365 			influx = 1;
2366 			KQ_LOCK(kq);
2367 		}
2368 		KQ_UNLOCK_FLUX(kq);
2369 	}
2370 }
2371 
2372 static int
2373 knote_attach(struct knote *kn, struct kqueue *kq)
2374 {
2375 	struct klist *list;
2376 
2377 	KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2378 	KQ_OWNED(kq);
2379 
2380 	if (kn->kn_fop->f_isfd) {
2381 		if (kn->kn_id >= kq->kq_knlistsize)
2382 			return ENOMEM;
2383 		list = &kq->kq_knlist[kn->kn_id];
2384 	} else {
2385 		if (kq->kq_knhash == NULL)
2386 			return ENOMEM;
2387 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2388 	}
2389 
2390 	SLIST_INSERT_HEAD(list, kn, kn_link);
2391 
2392 	return 0;
2393 }
2394 
2395 /*
2396  * knote must already have been detached using the f_detach method.
2397  * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2398  * to prevent other removal.
2399  */
2400 static void
2401 knote_drop(struct knote *kn, struct thread *td)
2402 {
2403 	struct kqueue *kq;
2404 	struct klist *list;
2405 
2406 	kq = kn->kn_kq;
2407 
2408 	KQ_NOTOWNED(kq);
2409 	KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2410 	    ("knote_drop called without KN_INFLUX set in kn_status"));
2411 
2412 	KQ_LOCK(kq);
2413 	if (kn->kn_fop->f_isfd)
2414 		list = &kq->kq_knlist[kn->kn_id];
2415 	else
2416 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2417 
2418 	if (!SLIST_EMPTY(list))
2419 		SLIST_REMOVE(list, kn, knote, kn_link);
2420 	if (kn->kn_status & KN_QUEUED)
2421 		knote_dequeue(kn);
2422 	KQ_UNLOCK_FLUX(kq);
2423 
2424 	if (kn->kn_fop->f_isfd) {
2425 		fdrop(kn->kn_fp, td);
2426 		kn->kn_fp = NULL;
2427 	}
2428 	kqueue_fo_release(kn->kn_kevent.filter);
2429 	kn->kn_fop = NULL;
2430 	knote_free(kn);
2431 }
2432 
2433 static void
2434 knote_enqueue(struct knote *kn)
2435 {
2436 	struct kqueue *kq = kn->kn_kq;
2437 
2438 	KQ_OWNED(kn->kn_kq);
2439 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2440 
2441 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2442 	kn->kn_status |= KN_QUEUED;
2443 	kq->kq_count++;
2444 	kqueue_wakeup(kq);
2445 }
2446 
2447 static void
2448 knote_dequeue(struct knote *kn)
2449 {
2450 	struct kqueue *kq = kn->kn_kq;
2451 
2452 	KQ_OWNED(kn->kn_kq);
2453 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2454 
2455 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2456 	kn->kn_status &= ~KN_QUEUED;
2457 	kq->kq_count--;
2458 }
2459 
2460 static void
2461 knote_init(void)
2462 {
2463 
2464 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2465 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2466 }
2467 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2468 
2469 static struct knote *
2470 knote_alloc(int waitok)
2471 {
2472 
2473 	return (uma_zalloc(knote_zone, (waitok ? M_WAITOK : M_NOWAIT) |
2474 	    M_ZERO));
2475 }
2476 
2477 static void
2478 knote_free(struct knote *kn)
2479 {
2480 
2481 	uma_zfree(knote_zone, kn);
2482 }
2483 
2484 /*
2485  * Register the kev w/ the kq specified by fd.
2486  */
2487 int
2488 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2489 {
2490 	struct kqueue *kq;
2491 	struct file *fp;
2492 	cap_rights_t rights;
2493 	int error;
2494 
2495 	error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2496 	if (error != 0)
2497 		return (error);
2498 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2499 		goto noacquire;
2500 
2501 	error = kqueue_register(kq, kev, td, waitok);
2502 
2503 	kqueue_release(kq, 0);
2504 
2505 noacquire:
2506 	fdrop(fp, td);
2507 
2508 	return error;
2509 }
2510