xref: /freebsd/sys/kern/kern_event.c (revision 3b1f7d9e5d6f44b50ff07fde6fd0e1135f213762)
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 
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/capability.h>
37 #include <sys/kernel.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/rwlock.h>
41 #include <sys/proc.h>
42 #include <sys/malloc.h>
43 #include <sys/unistd.h>
44 #include <sys/file.h>
45 #include <sys/filedesc.h>
46 #include <sys/filio.h>
47 #include <sys/fcntl.h>
48 #include <sys/kthread.h>
49 #include <sys/selinfo.h>
50 #include <sys/stdatomic.h>
51 #include <sys/queue.h>
52 #include <sys/event.h>
53 #include <sys/eventvar.h>
54 #include <sys/poll.h>
55 #include <sys/protosw.h>
56 #include <sys/sigio.h>
57 #include <sys/signalvar.h>
58 #include <sys/socket.h>
59 #include <sys/socketvar.h>
60 #include <sys/stat.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/syscallsubr.h>
64 #include <sys/taskqueue.h>
65 #include <sys/uio.h>
66 #ifdef KTRACE
67 #include <sys/ktrace.h>
68 #endif
69 
70 #include <vm/uma.h>
71 
72 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
73 
74 /*
75  * This lock is used if multiple kq locks are required.  This possibly
76  * should be made into a per proc lock.
77  */
78 static struct mtx	kq_global;
79 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
80 #define KQ_GLOBAL_LOCK(lck, haslck)	do {	\
81 	if (!haslck)				\
82 		mtx_lock(lck);			\
83 	haslck = 1;				\
84 } while (0)
85 #define KQ_GLOBAL_UNLOCK(lck, haslck)	do {	\
86 	if (haslck)				\
87 		mtx_unlock(lck);			\
88 	haslck = 0;				\
89 } while (0)
90 
91 TASKQUEUE_DEFINE_THREAD(kqueue);
92 
93 static int	kevent_copyout(void *arg, struct kevent *kevp, int count);
94 static int	kevent_copyin(void *arg, struct kevent *kevp, int count);
95 static int	kqueue_register(struct kqueue *kq, struct kevent *kev,
96 		    struct thread *td, int waitok);
97 static int	kqueue_acquire(struct file *fp, struct kqueue **kqp);
98 static void	kqueue_release(struct kqueue *kq, int locked);
99 static int	kqueue_expand(struct kqueue *kq, struct filterops *fops,
100 		    uintptr_t ident, int waitok);
101 static void	kqueue_task(void *arg, int pending);
102 static int	kqueue_scan(struct kqueue *kq, int maxevents,
103 		    struct kevent_copyops *k_ops,
104 		    const struct timespec *timeout,
105 		    struct kevent *keva, struct thread *td);
106 static void 	kqueue_wakeup(struct kqueue *kq);
107 static struct filterops *kqueue_fo_find(int filt);
108 static void	kqueue_fo_release(int filt);
109 
110 static fo_rdwr_t	kqueue_read;
111 static fo_rdwr_t	kqueue_write;
112 static fo_truncate_t	kqueue_truncate;
113 static fo_ioctl_t	kqueue_ioctl;
114 static fo_poll_t	kqueue_poll;
115 static fo_kqfilter_t	kqueue_kqfilter;
116 static fo_stat_t	kqueue_stat;
117 static fo_close_t	kqueue_close;
118 
119 static struct fileops kqueueops = {
120 	.fo_read = kqueue_read,
121 	.fo_write = kqueue_write,
122 	.fo_truncate = kqueue_truncate,
123 	.fo_ioctl = kqueue_ioctl,
124 	.fo_poll = kqueue_poll,
125 	.fo_kqfilter = kqueue_kqfilter,
126 	.fo_stat = kqueue_stat,
127 	.fo_close = kqueue_close,
128 	.fo_chmod = invfo_chmod,
129 	.fo_chown = invfo_chown,
130 };
131 
132 static int 	knote_attach(struct knote *kn, struct kqueue *kq);
133 static void 	knote_drop(struct knote *kn, struct thread *td);
134 static void 	knote_enqueue(struct knote *kn);
135 static void 	knote_dequeue(struct knote *kn);
136 static void 	knote_init(void);
137 static struct 	knote *knote_alloc(int waitok);
138 static void 	knote_free(struct knote *kn);
139 
140 static void	filt_kqdetach(struct knote *kn);
141 static int	filt_kqueue(struct knote *kn, long hint);
142 static int	filt_procattach(struct knote *kn);
143 static void	filt_procdetach(struct knote *kn);
144 static int	filt_proc(struct knote *kn, long hint);
145 static int	filt_fileattach(struct knote *kn);
146 static void	filt_timerexpire(void *knx);
147 static int	filt_timerattach(struct knote *kn);
148 static void	filt_timerdetach(struct knote *kn);
149 static int	filt_timer(struct knote *kn, long hint);
150 static int	filt_userattach(struct knote *kn);
151 static void	filt_userdetach(struct knote *kn);
152 static int	filt_user(struct knote *kn, long hint);
153 static void	filt_usertouch(struct knote *kn, struct kevent *kev,
154 		    u_long type);
155 
156 static struct filterops file_filtops = {
157 	.f_isfd = 1,
158 	.f_attach = filt_fileattach,
159 };
160 static struct filterops kqread_filtops = {
161 	.f_isfd = 1,
162 	.f_detach = filt_kqdetach,
163 	.f_event = filt_kqueue,
164 };
165 /* XXX - move to kern_proc.c?  */
166 static struct filterops proc_filtops = {
167 	.f_isfd = 0,
168 	.f_attach = filt_procattach,
169 	.f_detach = filt_procdetach,
170 	.f_event = filt_proc,
171 };
172 static struct filterops timer_filtops = {
173 	.f_isfd = 0,
174 	.f_attach = filt_timerattach,
175 	.f_detach = filt_timerdetach,
176 	.f_event = filt_timer,
177 };
178 static struct filterops user_filtops = {
179 	.f_attach = filt_userattach,
180 	.f_detach = filt_userdetach,
181 	.f_event = filt_user,
182 	.f_touch = filt_usertouch,
183 };
184 
185 static uma_zone_t	knote_zone;
186 static atomic_uint	kq_ncallouts = ATOMIC_VAR_INIT(0);
187 static unsigned int 	kq_calloutmax = 4 * 1024;
188 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
189     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
190 
191 /* XXX - ensure not KN_INFLUX?? */
192 #define KNOTE_ACTIVATE(kn, islock) do { 				\
193 	if ((islock))							\
194 		mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED);		\
195 	else								\
196 		KQ_LOCK((kn)->kn_kq);					\
197 	(kn)->kn_status |= KN_ACTIVE;					\
198 	if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)		\
199 		knote_enqueue((kn));					\
200 	if (!(islock))							\
201 		KQ_UNLOCK((kn)->kn_kq);					\
202 } while(0)
203 #define KQ_LOCK(kq) do {						\
204 	mtx_lock(&(kq)->kq_lock);					\
205 } while (0)
206 #define KQ_FLUX_WAKEUP(kq) do {						\
207 	if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) {		\
208 		(kq)->kq_state &= ~KQ_FLUXWAIT;				\
209 		wakeup((kq));						\
210 	}								\
211 } while (0)
212 #define KQ_UNLOCK_FLUX(kq) do {						\
213 	KQ_FLUX_WAKEUP(kq);						\
214 	mtx_unlock(&(kq)->kq_lock);					\
215 } while (0)
216 #define KQ_UNLOCK(kq) do {						\
217 	mtx_unlock(&(kq)->kq_lock);					\
218 } while (0)
219 #define KQ_OWNED(kq) do {						\
220 	mtx_assert(&(kq)->kq_lock, MA_OWNED);				\
221 } while (0)
222 #define KQ_NOTOWNED(kq) do {						\
223 	mtx_assert(&(kq)->kq_lock, MA_NOTOWNED);			\
224 } while (0)
225 #define KN_LIST_LOCK(kn) do {						\
226 	if (kn->kn_knlist != NULL)					\
227 		kn->kn_knlist->kl_lock(kn->kn_knlist->kl_lockarg);	\
228 } while (0)
229 #define KN_LIST_UNLOCK(kn) do {						\
230 	if (kn->kn_knlist != NULL) 					\
231 		kn->kn_knlist->kl_unlock(kn->kn_knlist->kl_lockarg);	\
232 } while (0)
233 #define	KNL_ASSERT_LOCK(knl, islocked) do {				\
234 	if (islocked)							\
235 		KNL_ASSERT_LOCKED(knl);				\
236 	else								\
237 		KNL_ASSERT_UNLOCKED(knl);				\
238 } while (0)
239 #ifdef INVARIANTS
240 #define	KNL_ASSERT_LOCKED(knl) do {					\
241 	knl->kl_assert_locked((knl)->kl_lockarg);			\
242 } while (0)
243 #define	KNL_ASSERT_UNLOCKED(knl) do {					\
244 	knl->kl_assert_unlocked((knl)->kl_lockarg);			\
245 } while (0)
246 #else /* !INVARIANTS */
247 #define	KNL_ASSERT_LOCKED(knl) do {} while(0)
248 #define	KNL_ASSERT_UNLOCKED(knl) do {} while (0)
249 #endif /* INVARIANTS */
250 
251 #define	KN_HASHSIZE		64		/* XXX should be tunable */
252 #define KN_HASH(val, mask)	(((val) ^ (val >> 8)) & (mask))
253 
254 static int
255 filt_nullattach(struct knote *kn)
256 {
257 
258 	return (ENXIO);
259 };
260 
261 struct filterops null_filtops = {
262 	.f_isfd = 0,
263 	.f_attach = filt_nullattach,
264 };
265 
266 /* XXX - make SYSINIT to add these, and move into respective modules. */
267 extern struct filterops sig_filtops;
268 extern struct filterops fs_filtops;
269 
270 /*
271  * Table for for all system-defined filters.
272  */
273 static struct mtx	filterops_lock;
274 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
275 	MTX_DEF);
276 static struct {
277 	struct filterops *for_fop;
278 	int for_refcnt;
279 } sysfilt_ops[EVFILT_SYSCOUNT] = {
280 	{ &file_filtops },			/* EVFILT_READ */
281 	{ &file_filtops },			/* EVFILT_WRITE */
282 	{ &null_filtops },			/* EVFILT_AIO */
283 	{ &file_filtops },			/* EVFILT_VNODE */
284 	{ &proc_filtops },			/* EVFILT_PROC */
285 	{ &sig_filtops },			/* EVFILT_SIGNAL */
286 	{ &timer_filtops },			/* EVFILT_TIMER */
287 	{ &null_filtops },			/* former EVFILT_NETDEV */
288 	{ &fs_filtops },			/* EVFILT_FS */
289 	{ &null_filtops },			/* EVFILT_LIO */
290 	{ &user_filtops },			/* EVFILT_USER */
291 };
292 
293 /*
294  * Simple redirection for all cdevsw style objects to call their fo_kqfilter
295  * method.
296  */
297 static int
298 filt_fileattach(struct knote *kn)
299 {
300 
301 	return (fo_kqfilter(kn->kn_fp, kn));
302 }
303 
304 /*ARGSUSED*/
305 static int
306 kqueue_kqfilter(struct file *fp, struct knote *kn)
307 {
308 	struct kqueue *kq = kn->kn_fp->f_data;
309 
310 	if (kn->kn_filter != EVFILT_READ)
311 		return (EINVAL);
312 
313 	kn->kn_status |= KN_KQUEUE;
314 	kn->kn_fop = &kqread_filtops;
315 	knlist_add(&kq->kq_sel.si_note, kn, 0);
316 
317 	return (0);
318 }
319 
320 static void
321 filt_kqdetach(struct knote *kn)
322 {
323 	struct kqueue *kq = kn->kn_fp->f_data;
324 
325 	knlist_remove(&kq->kq_sel.si_note, kn, 0);
326 }
327 
328 /*ARGSUSED*/
329 static int
330 filt_kqueue(struct knote *kn, long hint)
331 {
332 	struct kqueue *kq = kn->kn_fp->f_data;
333 
334 	kn->kn_data = kq->kq_count;
335 	return (kn->kn_data > 0);
336 }
337 
338 /* XXX - move to kern_proc.c?  */
339 static int
340 filt_procattach(struct knote *kn)
341 {
342 	struct proc *p;
343 	int immediate;
344 	int error;
345 
346 	immediate = 0;
347 	p = pfind(kn->kn_id);
348 	if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
349 		p = zpfind(kn->kn_id);
350 		immediate = 1;
351 	} else if (p != NULL && (p->p_flag & P_WEXIT)) {
352 		immediate = 1;
353 	}
354 
355 	if (p == NULL)
356 		return (ESRCH);
357 	if ((error = p_cansee(curthread, p))) {
358 		PROC_UNLOCK(p);
359 		return (error);
360 	}
361 
362 	kn->kn_ptr.p_proc = p;
363 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
364 
365 	/*
366 	 * internal flag indicating registration done by kernel
367 	 */
368 	if (kn->kn_flags & EV_FLAG1) {
369 		kn->kn_data = kn->kn_sdata;		/* ppid */
370 		kn->kn_fflags = NOTE_CHILD;
371 		kn->kn_flags &= ~EV_FLAG1;
372 	}
373 
374 	if (immediate == 0)
375 		knlist_add(&p->p_klist, kn, 1);
376 
377 	/*
378 	 * Immediately activate any exit notes if the target process is a
379 	 * zombie.  This is necessary to handle the case where the target
380 	 * process, e.g. a child, dies before the kevent is registered.
381 	 */
382 	if (immediate && filt_proc(kn, NOTE_EXIT))
383 		KNOTE_ACTIVATE(kn, 0);
384 
385 	PROC_UNLOCK(p);
386 
387 	return (0);
388 }
389 
390 /*
391  * The knote may be attached to a different process, which may exit,
392  * leaving nothing for the knote to be attached to.  So when the process
393  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
394  * it will be deleted when read out.  However, as part of the knote deletion,
395  * this routine is called, so a check is needed to avoid actually performing
396  * a detach, because the original process does not exist any more.
397  */
398 /* XXX - move to kern_proc.c?  */
399 static void
400 filt_procdetach(struct knote *kn)
401 {
402 	struct proc *p;
403 
404 	p = kn->kn_ptr.p_proc;
405 	knlist_remove(&p->p_klist, kn, 0);
406 	kn->kn_ptr.p_proc = NULL;
407 }
408 
409 /* XXX - move to kern_proc.c?  */
410 static int
411 filt_proc(struct knote *kn, long hint)
412 {
413 	struct proc *p = kn->kn_ptr.p_proc;
414 	u_int event;
415 
416 	/*
417 	 * mask off extra data
418 	 */
419 	event = (u_int)hint & NOTE_PCTRLMASK;
420 
421 	/*
422 	 * if the user is interested in this event, record it.
423 	 */
424 	if (kn->kn_sfflags & event)
425 		kn->kn_fflags |= event;
426 
427 	/*
428 	 * process is gone, so flag the event as finished.
429 	 */
430 	if (event == NOTE_EXIT) {
431 		if (!(kn->kn_status & KN_DETACHED))
432 			knlist_remove_inevent(&p->p_klist, kn);
433 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
434 		kn->kn_data = p->p_xstat;
435 		kn->kn_ptr.p_proc = NULL;
436 		return (1);
437 	}
438 
439 	return (kn->kn_fflags != 0);
440 }
441 
442 /*
443  * Called when the process forked. It mostly does the same as the
444  * knote(), activating all knotes registered to be activated when the
445  * process forked. Additionally, for each knote attached to the
446  * parent, check whether user wants to track the new process. If so
447  * attach a new knote to it, and immediately report an event with the
448  * child's pid.
449  */
450 void
451 knote_fork(struct knlist *list, int pid)
452 {
453 	struct kqueue *kq;
454 	struct knote *kn;
455 	struct kevent kev;
456 	int error;
457 
458 	if (list == NULL)
459 		return;
460 	list->kl_lock(list->kl_lockarg);
461 
462 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
463 		if ((kn->kn_status & KN_INFLUX) == KN_INFLUX)
464 			continue;
465 		kq = kn->kn_kq;
466 		KQ_LOCK(kq);
467 		if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
468 			KQ_UNLOCK(kq);
469 			continue;
470 		}
471 
472 		/*
473 		 * The same as knote(), activate the event.
474 		 */
475 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
476 			kn->kn_status |= KN_HASKQLOCK;
477 			if (kn->kn_fop->f_event(kn, NOTE_FORK | pid))
478 				KNOTE_ACTIVATE(kn, 1);
479 			kn->kn_status &= ~KN_HASKQLOCK;
480 			KQ_UNLOCK(kq);
481 			continue;
482 		}
483 
484 		/*
485 		 * The NOTE_TRACK case. In addition to the activation
486 		 * of the event, we need to register new event to
487 		 * track the child. Drop the locks in preparation for
488 		 * the call to kqueue_register().
489 		 */
490 		kn->kn_status |= KN_INFLUX;
491 		KQ_UNLOCK(kq);
492 		list->kl_unlock(list->kl_lockarg);
493 
494 		/*
495 		 * Activate existing knote and register a knote with
496 		 * new process.
497 		 */
498 		kev.ident = pid;
499 		kev.filter = kn->kn_filter;
500 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
501 		kev.fflags = kn->kn_sfflags;
502 		kev.data = kn->kn_id;		/* parent */
503 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
504 		error = kqueue_register(kq, &kev, NULL, 0);
505 		if (kn->kn_fop->f_event(kn, NOTE_FORK | pid))
506 			KNOTE_ACTIVATE(kn, 0);
507 		if (error)
508 			kn->kn_fflags |= NOTE_TRACKERR;
509 		KQ_LOCK(kq);
510 		kn->kn_status &= ~KN_INFLUX;
511 		KQ_UNLOCK_FLUX(kq);
512 		list->kl_lock(list->kl_lockarg);
513 	}
514 	list->kl_unlock(list->kl_lockarg);
515 }
516 
517 /*
518  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
519  * interval timer support code.
520  */
521 static __inline sbintime_t
522 timer2sbintime(intptr_t data)
523 {
524 
525 	return (SBT_1MS * data);
526 }
527 
528 static void
529 filt_timerexpire(void *knx)
530 {
531 	struct callout *calloutp;
532 	struct knote *kn;
533 
534 	kn = knx;
535 	kn->kn_data++;
536 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
537 
538 	if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
539 		calloutp = (struct callout *)kn->kn_hook;
540 		callout_reset_sbt_on(calloutp,
541 		    timer2sbintime(kn->kn_sdata), 0 /* 1ms? */,
542 		    filt_timerexpire, kn, PCPU_GET(cpuid), 0);
543 	}
544 }
545 
546 /*
547  * data contains amount of time to sleep, in milliseconds
548  */
549 static int
550 filt_timerattach(struct knote *kn)
551 {
552 	struct callout *calloutp;
553 	unsigned int ncallouts;
554 
555 	ncallouts = atomic_load_explicit(&kq_ncallouts, memory_order_relaxed);
556 	do {
557 		if (ncallouts >= kq_calloutmax)
558 			return (ENOMEM);
559 	} while (!atomic_compare_exchange_weak_explicit(&kq_ncallouts,
560 	    &ncallouts, ncallouts + 1, memory_order_relaxed,
561 	    memory_order_relaxed));
562 
563 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
564 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add usually sets it */
565 	calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
566 	callout_init(calloutp, CALLOUT_MPSAFE);
567 	kn->kn_hook = calloutp;
568 	callout_reset_sbt_on(calloutp,
569 	    timer2sbintime(kn->kn_sdata), 0 /* 1ms? */,
570 	    filt_timerexpire, kn, PCPU_GET(cpuid), 0);
571 
572 	return (0);
573 }
574 
575 static void
576 filt_timerdetach(struct knote *kn)
577 {
578 	struct callout *calloutp;
579 	unsigned int old;
580 
581 	calloutp = (struct callout *)kn->kn_hook;
582 	callout_drain(calloutp);
583 	free(calloutp, M_KQUEUE);
584 	old = atomic_fetch_sub_explicit(&kq_ncallouts, 1, memory_order_relaxed);
585 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
586 	kn->kn_status |= KN_DETACHED;	/* knlist_remove usually clears it */
587 }
588 
589 static int
590 filt_timer(struct knote *kn, long hint)
591 {
592 
593 	return (kn->kn_data != 0);
594 }
595 
596 static int
597 filt_userattach(struct knote *kn)
598 {
599 
600 	/*
601 	 * EVFILT_USER knotes are not attached to anything in the kernel.
602 	 */
603 	kn->kn_hook = NULL;
604 	if (kn->kn_fflags & NOTE_TRIGGER)
605 		kn->kn_hookid = 1;
606 	else
607 		kn->kn_hookid = 0;
608 	return (0);
609 }
610 
611 static void
612 filt_userdetach(__unused struct knote *kn)
613 {
614 
615 	/*
616 	 * EVFILT_USER knotes are not attached to anything in the kernel.
617 	 */
618 }
619 
620 static int
621 filt_user(struct knote *kn, __unused long hint)
622 {
623 
624 	return (kn->kn_hookid);
625 }
626 
627 static void
628 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
629 {
630 	u_int ffctrl;
631 
632 	switch (type) {
633 	case EVENT_REGISTER:
634 		if (kev->fflags & NOTE_TRIGGER)
635 			kn->kn_hookid = 1;
636 
637 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
638 		kev->fflags &= NOTE_FFLAGSMASK;
639 		switch (ffctrl) {
640 		case NOTE_FFNOP:
641 			break;
642 
643 		case NOTE_FFAND:
644 			kn->kn_sfflags &= kev->fflags;
645 			break;
646 
647 		case NOTE_FFOR:
648 			kn->kn_sfflags |= kev->fflags;
649 			break;
650 
651 		case NOTE_FFCOPY:
652 			kn->kn_sfflags = kev->fflags;
653 			break;
654 
655 		default:
656 			/* XXX Return error? */
657 			break;
658 		}
659 		kn->kn_sdata = kev->data;
660 		if (kev->flags & EV_CLEAR) {
661 			kn->kn_hookid = 0;
662 			kn->kn_data = 0;
663 			kn->kn_fflags = 0;
664 		}
665 		break;
666 
667         case EVENT_PROCESS:
668 		*kev = kn->kn_kevent;
669 		kev->fflags = kn->kn_sfflags;
670 		kev->data = kn->kn_sdata;
671 		if (kn->kn_flags & EV_CLEAR) {
672 			kn->kn_hookid = 0;
673 			kn->kn_data = 0;
674 			kn->kn_fflags = 0;
675 		}
676 		break;
677 
678 	default:
679 		panic("filt_usertouch() - invalid type (%ld)", type);
680 		break;
681 	}
682 }
683 
684 int
685 sys_kqueue(struct thread *td, struct kqueue_args *uap)
686 {
687 	struct filedesc *fdp;
688 	struct kqueue *kq;
689 	struct file *fp;
690 	int fd, error;
691 
692 	fdp = td->td_proc->p_fd;
693 	error = falloc(td, &fp, &fd, 0);
694 	if (error)
695 		goto done2;
696 
697 	/* An extra reference on `fp' has been held for us by falloc(). */
698 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
699 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
700 	TAILQ_INIT(&kq->kq_head);
701 	kq->kq_fdp = fdp;
702 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
703 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
704 
705 	FILEDESC_XLOCK(fdp);
706 	SLIST_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
707 	FILEDESC_XUNLOCK(fdp);
708 
709 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
710 	fdrop(fp, td);
711 
712 	td->td_retval[0] = fd;
713 done2:
714 	return (error);
715 }
716 
717 #ifndef _SYS_SYSPROTO_H_
718 struct kevent_args {
719 	int	fd;
720 	const struct kevent *changelist;
721 	int	nchanges;
722 	struct	kevent *eventlist;
723 	int	nevents;
724 	const struct timespec *timeout;
725 };
726 #endif
727 int
728 sys_kevent(struct thread *td, struct kevent_args *uap)
729 {
730 	struct timespec ts, *tsp;
731 	struct kevent_copyops k_ops = { uap,
732 					kevent_copyout,
733 					kevent_copyin};
734 	int error;
735 #ifdef KTRACE
736 	struct uio ktruio;
737 	struct iovec ktriov;
738 	struct uio *ktruioin = NULL;
739 	struct uio *ktruioout = NULL;
740 #endif
741 
742 	if (uap->timeout != NULL) {
743 		error = copyin(uap->timeout, &ts, sizeof(ts));
744 		if (error)
745 			return (error);
746 		tsp = &ts;
747 	} else
748 		tsp = NULL;
749 
750 #ifdef KTRACE
751 	if (KTRPOINT(td, KTR_GENIO)) {
752 		ktriov.iov_base = uap->changelist;
753 		ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
754 		ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
755 		    .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
756 		    .uio_td = td };
757 		ktruioin = cloneuio(&ktruio);
758 		ktriov.iov_base = uap->eventlist;
759 		ktriov.iov_len = uap->nevents * sizeof(struct kevent);
760 		ktruioout = cloneuio(&ktruio);
761 	}
762 #endif
763 
764 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
765 	    &k_ops, tsp);
766 
767 #ifdef KTRACE
768 	if (ktruioin != NULL) {
769 		ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
770 		ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
771 		ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
772 		ktrgenio(uap->fd, UIO_READ, ktruioout, error);
773 	}
774 #endif
775 
776 	return (error);
777 }
778 
779 /*
780  * Copy 'count' items into the destination list pointed to by uap->eventlist.
781  */
782 static int
783 kevent_copyout(void *arg, struct kevent *kevp, int count)
784 {
785 	struct kevent_args *uap;
786 	int error;
787 
788 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
789 	uap = (struct kevent_args *)arg;
790 
791 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
792 	if (error == 0)
793 		uap->eventlist += count;
794 	return (error);
795 }
796 
797 /*
798  * Copy 'count' items from the list pointed to by uap->changelist.
799  */
800 static int
801 kevent_copyin(void *arg, struct kevent *kevp, int count)
802 {
803 	struct kevent_args *uap;
804 	int error;
805 
806 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
807 	uap = (struct kevent_args *)arg;
808 
809 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
810 	if (error == 0)
811 		uap->changelist += count;
812 	return (error);
813 }
814 
815 int
816 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
817     struct kevent_copyops *k_ops, const struct timespec *timeout)
818 {
819 	struct kevent keva[KQ_NEVENTS];
820 	struct kevent *kevp, *changes;
821 	struct kqueue *kq;
822 	struct file *fp;
823 	int i, n, nerrors, error;
824 
825 	if ((error = fget(td, fd, CAP_POST_EVENT, &fp)) != 0)
826 		return (error);
827 	if ((error = kqueue_acquire(fp, &kq)) != 0)
828 		goto done_norel;
829 
830 	nerrors = 0;
831 
832 	while (nchanges > 0) {
833 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
834 		error = k_ops->k_copyin(k_ops->arg, keva, n);
835 		if (error)
836 			goto done;
837 		changes = keva;
838 		for (i = 0; i < n; i++) {
839 			kevp = &changes[i];
840 			if (!kevp->filter)
841 				continue;
842 			kevp->flags &= ~EV_SYSFLAGS;
843 			error = kqueue_register(kq, kevp, td, 1);
844 			if (error || (kevp->flags & EV_RECEIPT)) {
845 				if (nevents != 0) {
846 					kevp->flags = EV_ERROR;
847 					kevp->data = error;
848 					(void) k_ops->k_copyout(k_ops->arg,
849 					    kevp, 1);
850 					nevents--;
851 					nerrors++;
852 				} else {
853 					goto done;
854 				}
855 			}
856 		}
857 		nchanges -= n;
858 	}
859 	if (nerrors) {
860 		td->td_retval[0] = nerrors;
861 		error = 0;
862 		goto done;
863 	}
864 
865 	error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
866 done:
867 	kqueue_release(kq, 0);
868 done_norel:
869 	fdrop(fp, td);
870 	return (error);
871 }
872 
873 int
874 kqueue_add_filteropts(int filt, struct filterops *filtops)
875 {
876 	int error;
877 
878 	error = 0;
879 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
880 		printf(
881 "trying to add a filterop that is out of range: %d is beyond %d\n",
882 		    ~filt, EVFILT_SYSCOUNT);
883 		return EINVAL;
884 	}
885 	mtx_lock(&filterops_lock);
886 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
887 	    sysfilt_ops[~filt].for_fop != NULL)
888 		error = EEXIST;
889 	else {
890 		sysfilt_ops[~filt].for_fop = filtops;
891 		sysfilt_ops[~filt].for_refcnt = 0;
892 	}
893 	mtx_unlock(&filterops_lock);
894 
895 	return (error);
896 }
897 
898 int
899 kqueue_del_filteropts(int filt)
900 {
901 	int error;
902 
903 	error = 0;
904 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
905 		return EINVAL;
906 
907 	mtx_lock(&filterops_lock);
908 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
909 	    sysfilt_ops[~filt].for_fop == NULL)
910 		error = EINVAL;
911 	else if (sysfilt_ops[~filt].for_refcnt != 0)
912 		error = EBUSY;
913 	else {
914 		sysfilt_ops[~filt].for_fop = &null_filtops;
915 		sysfilt_ops[~filt].for_refcnt = 0;
916 	}
917 	mtx_unlock(&filterops_lock);
918 
919 	return error;
920 }
921 
922 static struct filterops *
923 kqueue_fo_find(int filt)
924 {
925 
926 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
927 		return NULL;
928 
929 	mtx_lock(&filterops_lock);
930 	sysfilt_ops[~filt].for_refcnt++;
931 	if (sysfilt_ops[~filt].for_fop == NULL)
932 		sysfilt_ops[~filt].for_fop = &null_filtops;
933 	mtx_unlock(&filterops_lock);
934 
935 	return sysfilt_ops[~filt].for_fop;
936 }
937 
938 static void
939 kqueue_fo_release(int filt)
940 {
941 
942 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
943 		return;
944 
945 	mtx_lock(&filterops_lock);
946 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
947 	    ("filter object refcount not valid on release"));
948 	sysfilt_ops[~filt].for_refcnt--;
949 	mtx_unlock(&filterops_lock);
950 }
951 
952 /*
953  * A ref to kq (obtained via kqueue_acquire) must be held.  waitok will
954  * influence if memory allocation should wait.  Make sure it is 0 if you
955  * hold any mutexes.
956  */
957 static int
958 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
959 {
960 	struct filterops *fops;
961 	struct file *fp;
962 	struct knote *kn, *tkn;
963 	int error, filt, event;
964 	int haskqglobal;
965 
966 	fp = NULL;
967 	kn = NULL;
968 	error = 0;
969 	haskqglobal = 0;
970 
971 	filt = kev->filter;
972 	fops = kqueue_fo_find(filt);
973 	if (fops == NULL)
974 		return EINVAL;
975 
976 	tkn = knote_alloc(waitok);		/* prevent waiting with locks */
977 
978 findkn:
979 	if (fops->f_isfd) {
980 		KASSERT(td != NULL, ("td is NULL"));
981 		error = fget(td, kev->ident, CAP_POLL_EVENT, &fp);
982 		if (error)
983 			goto done;
984 
985 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
986 		    kev->ident, 0) != 0) {
987 			/* try again */
988 			fdrop(fp, td);
989 			fp = NULL;
990 			error = kqueue_expand(kq, fops, kev->ident, waitok);
991 			if (error)
992 				goto done;
993 			goto findkn;
994 		}
995 
996 		if (fp->f_type == DTYPE_KQUEUE) {
997 			/*
998 			 * if we add some inteligence about what we are doing,
999 			 * we should be able to support events on ourselves.
1000 			 * We need to know when we are doing this to prevent
1001 			 * getting both the knlist lock and the kq lock since
1002 			 * they are the same thing.
1003 			 */
1004 			if (fp->f_data == kq) {
1005 				error = EINVAL;
1006 				goto done;
1007 			}
1008 
1009 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1010 		}
1011 
1012 		KQ_LOCK(kq);
1013 		if (kev->ident < kq->kq_knlistsize) {
1014 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1015 				if (kev->filter == kn->kn_filter)
1016 					break;
1017 		}
1018 	} else {
1019 		if ((kev->flags & EV_ADD) == EV_ADD)
1020 			kqueue_expand(kq, fops, kev->ident, waitok);
1021 
1022 		KQ_LOCK(kq);
1023 		if (kq->kq_knhashmask != 0) {
1024 			struct klist *list;
1025 
1026 			list = &kq->kq_knhash[
1027 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1028 			SLIST_FOREACH(kn, list, kn_link)
1029 				if (kev->ident == kn->kn_id &&
1030 				    kev->filter == kn->kn_filter)
1031 					break;
1032 		}
1033 	}
1034 
1035 	/* knote is in the process of changing, wait for it to stablize. */
1036 	if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1037 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1038 		kq->kq_state |= KQ_FLUXWAIT;
1039 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1040 		if (fp != NULL) {
1041 			fdrop(fp, td);
1042 			fp = NULL;
1043 		}
1044 		goto findkn;
1045 	}
1046 
1047 	/*
1048 	 * kn now contains the matching knote, or NULL if no match
1049 	 */
1050 	if (kn == NULL) {
1051 		if (kev->flags & EV_ADD) {
1052 			kn = tkn;
1053 			tkn = NULL;
1054 			if (kn == NULL) {
1055 				KQ_UNLOCK(kq);
1056 				error = ENOMEM;
1057 				goto done;
1058 			}
1059 			kn->kn_fp = fp;
1060 			kn->kn_kq = kq;
1061 			kn->kn_fop = fops;
1062 			/*
1063 			 * apply reference counts to knote structure, and
1064 			 * do not release it at the end of this routine.
1065 			 */
1066 			fops = NULL;
1067 			fp = NULL;
1068 
1069 			kn->kn_sfflags = kev->fflags;
1070 			kn->kn_sdata = kev->data;
1071 			kev->fflags = 0;
1072 			kev->data = 0;
1073 			kn->kn_kevent = *kev;
1074 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1075 			    EV_ENABLE | EV_DISABLE);
1076 			kn->kn_status = KN_INFLUX|KN_DETACHED;
1077 
1078 			error = knote_attach(kn, kq);
1079 			KQ_UNLOCK(kq);
1080 			if (error != 0) {
1081 				tkn = kn;
1082 				goto done;
1083 			}
1084 
1085 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1086 				knote_drop(kn, td);
1087 				goto done;
1088 			}
1089 			KN_LIST_LOCK(kn);
1090 			goto done_ev_add;
1091 		} else {
1092 			/* No matching knote and the EV_ADD flag is not set. */
1093 			KQ_UNLOCK(kq);
1094 			error = ENOENT;
1095 			goto done;
1096 		}
1097 	}
1098 
1099 	if (kev->flags & EV_DELETE) {
1100 		kn->kn_status |= KN_INFLUX;
1101 		KQ_UNLOCK(kq);
1102 		if (!(kn->kn_status & KN_DETACHED))
1103 			kn->kn_fop->f_detach(kn);
1104 		knote_drop(kn, td);
1105 		goto done;
1106 	}
1107 
1108 	/*
1109 	 * The user may change some filter values after the initial EV_ADD,
1110 	 * but doing so will not reset any filter which has already been
1111 	 * triggered.
1112 	 */
1113 	kn->kn_status |= KN_INFLUX;
1114 	KQ_UNLOCK(kq);
1115 	KN_LIST_LOCK(kn);
1116 	kn->kn_kevent.udata = kev->udata;
1117 	if (!fops->f_isfd && fops->f_touch != NULL) {
1118 		fops->f_touch(kn, kev, EVENT_REGISTER);
1119 	} else {
1120 		kn->kn_sfflags = kev->fflags;
1121 		kn->kn_sdata = kev->data;
1122 	}
1123 
1124 	/*
1125 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1126 	 * the initial attach event decides that the event is "completed"
1127 	 * already.  i.e. filt_procattach is called on a zombie process.  It
1128 	 * will call filt_proc which will remove it from the list, and NULL
1129 	 * kn_knlist.
1130 	 */
1131 done_ev_add:
1132 	event = kn->kn_fop->f_event(kn, 0);
1133 	KQ_LOCK(kq);
1134 	if (event)
1135 		KNOTE_ACTIVATE(kn, 1);
1136 	kn->kn_status &= ~KN_INFLUX;
1137 	KN_LIST_UNLOCK(kn);
1138 
1139 	if ((kev->flags & EV_DISABLE) &&
1140 	    ((kn->kn_status & KN_DISABLED) == 0)) {
1141 		kn->kn_status |= KN_DISABLED;
1142 	}
1143 
1144 	if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1145 		kn->kn_status &= ~KN_DISABLED;
1146 		if ((kn->kn_status & KN_ACTIVE) &&
1147 		    ((kn->kn_status & KN_QUEUED) == 0))
1148 			knote_enqueue(kn);
1149 	}
1150 	KQ_UNLOCK_FLUX(kq);
1151 
1152 done:
1153 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1154 	if (fp != NULL)
1155 		fdrop(fp, td);
1156 	if (tkn != NULL)
1157 		knote_free(tkn);
1158 	if (fops != NULL)
1159 		kqueue_fo_release(filt);
1160 	return (error);
1161 }
1162 
1163 static int
1164 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1165 {
1166 	int error;
1167 	struct kqueue *kq;
1168 
1169 	error = 0;
1170 
1171 	kq = fp->f_data;
1172 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1173 		return (EBADF);
1174 	*kqp = kq;
1175 	KQ_LOCK(kq);
1176 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1177 		KQ_UNLOCK(kq);
1178 		return (EBADF);
1179 	}
1180 	kq->kq_refcnt++;
1181 	KQ_UNLOCK(kq);
1182 
1183 	return error;
1184 }
1185 
1186 static void
1187 kqueue_release(struct kqueue *kq, int locked)
1188 {
1189 	if (locked)
1190 		KQ_OWNED(kq);
1191 	else
1192 		KQ_LOCK(kq);
1193 	kq->kq_refcnt--;
1194 	if (kq->kq_refcnt == 1)
1195 		wakeup(&kq->kq_refcnt);
1196 	if (!locked)
1197 		KQ_UNLOCK(kq);
1198 }
1199 
1200 static void
1201 kqueue_schedtask(struct kqueue *kq)
1202 {
1203 
1204 	KQ_OWNED(kq);
1205 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1206 	    ("scheduling kqueue task while draining"));
1207 
1208 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1209 		taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1210 		kq->kq_state |= KQ_TASKSCHED;
1211 	}
1212 }
1213 
1214 /*
1215  * Expand the kq to make sure we have storage for fops/ident pair.
1216  *
1217  * Return 0 on success (or no work necessary), return errno on failure.
1218  *
1219  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1220  * If kqueue_register is called from a non-fd context, there usually/should
1221  * be no locks held.
1222  */
1223 static int
1224 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1225 	int waitok)
1226 {
1227 	struct klist *list, *tmp_knhash, *to_free;
1228 	u_long tmp_knhashmask;
1229 	int size;
1230 	int fd;
1231 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
1232 
1233 	KQ_NOTOWNED(kq);
1234 
1235 	to_free = NULL;
1236 	if (fops->f_isfd) {
1237 		fd = ident;
1238 		if (kq->kq_knlistsize <= fd) {
1239 			size = kq->kq_knlistsize;
1240 			while (size <= fd)
1241 				size += KQEXTENT;
1242 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1243 			if (list == NULL)
1244 				return ENOMEM;
1245 			KQ_LOCK(kq);
1246 			if (kq->kq_knlistsize > fd) {
1247 				to_free = list;
1248 				list = NULL;
1249 			} else {
1250 				if (kq->kq_knlist != NULL) {
1251 					bcopy(kq->kq_knlist, list,
1252 					    kq->kq_knlistsize * sizeof(*list));
1253 					to_free = kq->kq_knlist;
1254 					kq->kq_knlist = NULL;
1255 				}
1256 				bzero((caddr_t)list +
1257 				    kq->kq_knlistsize * sizeof(*list),
1258 				    (size - kq->kq_knlistsize) * sizeof(*list));
1259 				kq->kq_knlistsize = size;
1260 				kq->kq_knlist = list;
1261 			}
1262 			KQ_UNLOCK(kq);
1263 		}
1264 	} else {
1265 		if (kq->kq_knhashmask == 0) {
1266 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1267 			    &tmp_knhashmask);
1268 			if (tmp_knhash == NULL)
1269 				return ENOMEM;
1270 			KQ_LOCK(kq);
1271 			if (kq->kq_knhashmask == 0) {
1272 				kq->kq_knhash = tmp_knhash;
1273 				kq->kq_knhashmask = tmp_knhashmask;
1274 			} else {
1275 				to_free = tmp_knhash;
1276 			}
1277 			KQ_UNLOCK(kq);
1278 		}
1279 	}
1280 	free(to_free, M_KQUEUE);
1281 
1282 	KQ_NOTOWNED(kq);
1283 	return 0;
1284 }
1285 
1286 static void
1287 kqueue_task(void *arg, int pending)
1288 {
1289 	struct kqueue *kq;
1290 	int haskqglobal;
1291 
1292 	haskqglobal = 0;
1293 	kq = arg;
1294 
1295 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1296 	KQ_LOCK(kq);
1297 
1298 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1299 
1300 	kq->kq_state &= ~KQ_TASKSCHED;
1301 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1302 		wakeup(&kq->kq_state);
1303 	}
1304 	KQ_UNLOCK(kq);
1305 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1306 }
1307 
1308 /*
1309  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1310  * We treat KN_MARKER knotes as if they are INFLUX.
1311  */
1312 static int
1313 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1314     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1315 {
1316 	struct kevent *kevp;
1317 	struct knote *kn, *marker;
1318 	sbintime_t asbt, rsbt;
1319 	int count, error, haskqglobal, influx, nkev, touch;
1320 
1321 	count = maxevents;
1322 	nkev = 0;
1323 	error = 0;
1324 	haskqglobal = 0;
1325 
1326 	if (maxevents == 0)
1327 		goto done_nl;
1328 
1329 	rsbt = 0;
1330 	if (tsp != NULL) {
1331 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1332 		    tsp->tv_nsec >= 1000000000) {
1333 			error = EINVAL;
1334 			goto done_nl;
1335 		}
1336 		if (timespecisset(tsp)) {
1337 			if (tsp->tv_sec <= INT32_MAX) {
1338 				rsbt = tstosbt(*tsp);
1339 				if (TIMESEL(&asbt, rsbt))
1340 					asbt += tc_tick_sbt;
1341 				if (asbt <= INT64_MAX - rsbt)
1342 					asbt += rsbt;
1343 				else
1344 					asbt = 0;
1345 				rsbt >>= tc_precexp;
1346 			} else
1347 				asbt = 0;
1348 		} else
1349 			asbt = -1;
1350 	} else
1351 		asbt = 0;
1352 	marker = knote_alloc(1);
1353 	if (marker == NULL) {
1354 		error = ENOMEM;
1355 		goto done_nl;
1356 	}
1357 	marker->kn_status = KN_MARKER;
1358 	KQ_LOCK(kq);
1359 
1360 retry:
1361 	kevp = keva;
1362 	if (kq->kq_count == 0) {
1363 		if (asbt == -1) {
1364 			error = EWOULDBLOCK;
1365 		} else {
1366 			kq->kq_state |= KQ_SLEEP;
1367 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1368 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1369 		}
1370 		if (error == 0)
1371 			goto retry;
1372 		/* don't restart after signals... */
1373 		if (error == ERESTART)
1374 			error = EINTR;
1375 		else if (error == EWOULDBLOCK)
1376 			error = 0;
1377 		goto done;
1378 	}
1379 
1380 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1381 	influx = 0;
1382 	while (count) {
1383 		KQ_OWNED(kq);
1384 		kn = TAILQ_FIRST(&kq->kq_head);
1385 
1386 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1387 		    (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1388 			if (influx) {
1389 				influx = 0;
1390 				KQ_FLUX_WAKEUP(kq);
1391 			}
1392 			kq->kq_state |= KQ_FLUXWAIT;
1393 			error = msleep(kq, &kq->kq_lock, PSOCK,
1394 			    "kqflxwt", 0);
1395 			continue;
1396 		}
1397 
1398 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1399 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1400 			kn->kn_status &= ~KN_QUEUED;
1401 			kq->kq_count--;
1402 			continue;
1403 		}
1404 		if (kn == marker) {
1405 			KQ_FLUX_WAKEUP(kq);
1406 			if (count == maxevents)
1407 				goto retry;
1408 			goto done;
1409 		}
1410 		KASSERT((kn->kn_status & KN_INFLUX) == 0,
1411 		    ("KN_INFLUX set when not suppose to be"));
1412 
1413 		if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1414 			kn->kn_status &= ~KN_QUEUED;
1415 			kn->kn_status |= KN_INFLUX;
1416 			kq->kq_count--;
1417 			KQ_UNLOCK(kq);
1418 			/*
1419 			 * We don't need to lock the list since we've marked
1420 			 * it _INFLUX.
1421 			 */
1422 			*kevp = kn->kn_kevent;
1423 			if (!(kn->kn_status & KN_DETACHED))
1424 				kn->kn_fop->f_detach(kn);
1425 			knote_drop(kn, td);
1426 			KQ_LOCK(kq);
1427 			kn = NULL;
1428 		} else {
1429 			kn->kn_status |= KN_INFLUX;
1430 			KQ_UNLOCK(kq);
1431 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1432 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1433 			KN_LIST_LOCK(kn);
1434 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1435 				KQ_LOCK(kq);
1436 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1437 				kn->kn_status &=
1438 				    ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX);
1439 				kq->kq_count--;
1440 				KN_LIST_UNLOCK(kn);
1441 				influx = 1;
1442 				continue;
1443 			}
1444 			touch = (!kn->kn_fop->f_isfd &&
1445 			    kn->kn_fop->f_touch != NULL);
1446 			if (touch)
1447 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1448 			else
1449 				*kevp = kn->kn_kevent;
1450 			KQ_LOCK(kq);
1451 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1452 			if (kn->kn_flags & (EV_CLEAR |  EV_DISPATCH)) {
1453 				/*
1454 				 * Manually clear knotes who weren't
1455 				 * 'touch'ed.
1456 				 */
1457 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1458 					kn->kn_data = 0;
1459 					kn->kn_fflags = 0;
1460 				}
1461 				if (kn->kn_flags & EV_DISPATCH)
1462 					kn->kn_status |= KN_DISABLED;
1463 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1464 				kq->kq_count--;
1465 			} else
1466 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1467 
1468 			kn->kn_status &= ~(KN_INFLUX);
1469 			KN_LIST_UNLOCK(kn);
1470 			influx = 1;
1471 		}
1472 
1473 		/* we are returning a copy to the user */
1474 		kevp++;
1475 		nkev++;
1476 		count--;
1477 
1478 		if (nkev == KQ_NEVENTS) {
1479 			influx = 0;
1480 			KQ_UNLOCK_FLUX(kq);
1481 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1482 			nkev = 0;
1483 			kevp = keva;
1484 			KQ_LOCK(kq);
1485 			if (error)
1486 				break;
1487 		}
1488 	}
1489 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1490 done:
1491 	KQ_OWNED(kq);
1492 	KQ_UNLOCK_FLUX(kq);
1493 	knote_free(marker);
1494 done_nl:
1495 	KQ_NOTOWNED(kq);
1496 	if (nkev != 0)
1497 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1498 	td->td_retval[0] = maxevents - count;
1499 	return (error);
1500 }
1501 
1502 /*
1503  * XXX
1504  * This could be expanded to call kqueue_scan, if desired.
1505  */
1506 /*ARGSUSED*/
1507 static int
1508 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
1509 	int flags, struct thread *td)
1510 {
1511 	return (ENXIO);
1512 }
1513 
1514 /*ARGSUSED*/
1515 static int
1516 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1517 	 int flags, struct thread *td)
1518 {
1519 	return (ENXIO);
1520 }
1521 
1522 /*ARGSUSED*/
1523 static int
1524 kqueue_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1525 	struct thread *td)
1526 {
1527 
1528 	return (EINVAL);
1529 }
1530 
1531 /*ARGSUSED*/
1532 static int
1533 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1534 	struct ucred *active_cred, struct thread *td)
1535 {
1536 	/*
1537 	 * Enabling sigio causes two major problems:
1538 	 * 1) infinite recursion:
1539 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1540 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1541 	 * into itself over and over.  Sending the sigio causes the kqueue
1542 	 * to become ready, which in turn posts sigio again, forever.
1543 	 * Solution: this can be solved by setting a flag in the kqueue that
1544 	 * we have a SIGIO in progress.
1545 	 * 2) locking problems:
1546 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1547 	 * us above the proc and pgrp locks.
1548 	 * Solution: Post a signal using an async mechanism, being sure to
1549 	 * record a generation count in the delivery so that we do not deliver
1550 	 * a signal to the wrong process.
1551 	 *
1552 	 * Note, these two mechanisms are somewhat mutually exclusive!
1553 	 */
1554 #if 0
1555 	struct kqueue *kq;
1556 
1557 	kq = fp->f_data;
1558 	switch (cmd) {
1559 	case FIOASYNC:
1560 		if (*(int *)data) {
1561 			kq->kq_state |= KQ_ASYNC;
1562 		} else {
1563 			kq->kq_state &= ~KQ_ASYNC;
1564 		}
1565 		return (0);
1566 
1567 	case FIOSETOWN:
1568 		return (fsetown(*(int *)data, &kq->kq_sigio));
1569 
1570 	case FIOGETOWN:
1571 		*(int *)data = fgetown(&kq->kq_sigio);
1572 		return (0);
1573 	}
1574 #endif
1575 
1576 	return (ENOTTY);
1577 }
1578 
1579 /*ARGSUSED*/
1580 static int
1581 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1582 	struct thread *td)
1583 {
1584 	struct kqueue *kq;
1585 	int revents = 0;
1586 	int error;
1587 
1588 	if ((error = kqueue_acquire(fp, &kq)))
1589 		return POLLERR;
1590 
1591 	KQ_LOCK(kq);
1592 	if (events & (POLLIN | POLLRDNORM)) {
1593 		if (kq->kq_count) {
1594 			revents |= events & (POLLIN | POLLRDNORM);
1595 		} else {
1596 			selrecord(td, &kq->kq_sel);
1597 			if (SEL_WAITING(&kq->kq_sel))
1598 				kq->kq_state |= KQ_SEL;
1599 		}
1600 	}
1601 	kqueue_release(kq, 1);
1602 	KQ_UNLOCK(kq);
1603 	return (revents);
1604 }
1605 
1606 /*ARGSUSED*/
1607 static int
1608 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1609 	struct thread *td)
1610 {
1611 
1612 	bzero((void *)st, sizeof *st);
1613 	/*
1614 	 * We no longer return kq_count because the unlocked value is useless.
1615 	 * If you spent all this time getting the count, why not spend your
1616 	 * syscall better by calling kevent?
1617 	 *
1618 	 * XXX - This is needed for libc_r.
1619 	 */
1620 	st->st_mode = S_IFIFO;
1621 	return (0);
1622 }
1623 
1624 /*ARGSUSED*/
1625 static int
1626 kqueue_close(struct file *fp, struct thread *td)
1627 {
1628 	struct kqueue *kq = fp->f_data;
1629 	struct filedesc *fdp;
1630 	struct knote *kn;
1631 	int i;
1632 	int error;
1633 
1634 	if ((error = kqueue_acquire(fp, &kq)))
1635 		return error;
1636 
1637 	KQ_LOCK(kq);
1638 
1639 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1640 	    ("kqueue already closing"));
1641 	kq->kq_state |= KQ_CLOSING;
1642 	if (kq->kq_refcnt > 1)
1643 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1644 
1645 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1646 	fdp = kq->kq_fdp;
1647 
1648 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
1649 	    ("kqueue's knlist not empty"));
1650 
1651 	for (i = 0; i < kq->kq_knlistsize; i++) {
1652 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1653 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1654 				kq->kq_state |= KQ_FLUXWAIT;
1655 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1656 				continue;
1657 			}
1658 			kn->kn_status |= KN_INFLUX;
1659 			KQ_UNLOCK(kq);
1660 			if (!(kn->kn_status & KN_DETACHED))
1661 				kn->kn_fop->f_detach(kn);
1662 			knote_drop(kn, td);
1663 			KQ_LOCK(kq);
1664 		}
1665 	}
1666 	if (kq->kq_knhashmask != 0) {
1667 		for (i = 0; i <= kq->kq_knhashmask; i++) {
1668 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1669 				if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1670 					kq->kq_state |= KQ_FLUXWAIT;
1671 					msleep(kq, &kq->kq_lock, PSOCK,
1672 					       "kqclo2", 0);
1673 					continue;
1674 				}
1675 				kn->kn_status |= KN_INFLUX;
1676 				KQ_UNLOCK(kq);
1677 				if (!(kn->kn_status & KN_DETACHED))
1678 					kn->kn_fop->f_detach(kn);
1679 				knote_drop(kn, td);
1680 				KQ_LOCK(kq);
1681 			}
1682 		}
1683 	}
1684 
1685 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1686 		kq->kq_state |= KQ_TASKDRAIN;
1687 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1688 	}
1689 
1690 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1691 		selwakeuppri(&kq->kq_sel, PSOCK);
1692 		if (!SEL_WAITING(&kq->kq_sel))
1693 			kq->kq_state &= ~KQ_SEL;
1694 	}
1695 
1696 	KQ_UNLOCK(kq);
1697 
1698 	FILEDESC_XLOCK(fdp);
1699 	SLIST_REMOVE(&fdp->fd_kqlist, kq, kqueue, kq_list);
1700 	FILEDESC_XUNLOCK(fdp);
1701 
1702 	seldrain(&kq->kq_sel);
1703 	knlist_destroy(&kq->kq_sel.si_note);
1704 	mtx_destroy(&kq->kq_lock);
1705 	kq->kq_fdp = NULL;
1706 
1707 	if (kq->kq_knhash != NULL)
1708 		free(kq->kq_knhash, M_KQUEUE);
1709 	if (kq->kq_knlist != NULL)
1710 		free(kq->kq_knlist, M_KQUEUE);
1711 
1712 	funsetown(&kq->kq_sigio);
1713 	free(kq, M_KQUEUE);
1714 	fp->f_data = NULL;
1715 
1716 	return (0);
1717 }
1718 
1719 static void
1720 kqueue_wakeup(struct kqueue *kq)
1721 {
1722 	KQ_OWNED(kq);
1723 
1724 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1725 		kq->kq_state &= ~KQ_SLEEP;
1726 		wakeup(kq);
1727 	}
1728 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1729 		selwakeuppri(&kq->kq_sel, PSOCK);
1730 		if (!SEL_WAITING(&kq->kq_sel))
1731 			kq->kq_state &= ~KQ_SEL;
1732 	}
1733 	if (!knlist_empty(&kq->kq_sel.si_note))
1734 		kqueue_schedtask(kq);
1735 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1736 		pgsigio(&kq->kq_sigio, SIGIO, 0);
1737 	}
1738 }
1739 
1740 /*
1741  * Walk down a list of knotes, activating them if their event has triggered.
1742  *
1743  * There is a possibility to optimize in the case of one kq watching another.
1744  * Instead of scheduling a task to wake it up, you could pass enough state
1745  * down the chain to make up the parent kqueue.  Make this code functional
1746  * first.
1747  */
1748 void
1749 knote(struct knlist *list, long hint, int lockflags)
1750 {
1751 	struct kqueue *kq;
1752 	struct knote *kn;
1753 	int error;
1754 
1755 	if (list == NULL)
1756 		return;
1757 
1758 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1759 
1760 	if ((lockflags & KNF_LISTLOCKED) == 0)
1761 		list->kl_lock(list->kl_lockarg);
1762 
1763 	/*
1764 	 * If we unlock the list lock (and set KN_INFLUX), we can eliminate
1765 	 * the kqueue scheduling, but this will introduce four
1766 	 * lock/unlock's for each knote to test.  If we do, continue to use
1767 	 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is
1768 	 * only safe if you want to remove the current item, which we are
1769 	 * not doing.
1770 	 */
1771 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
1772 		kq = kn->kn_kq;
1773 		if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) {
1774 			KQ_LOCK(kq);
1775 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1776 				KQ_UNLOCK(kq);
1777 			} else if ((lockflags & KNF_NOKQLOCK) != 0) {
1778 				kn->kn_status |= KN_INFLUX;
1779 				KQ_UNLOCK(kq);
1780 				error = kn->kn_fop->f_event(kn, hint);
1781 				KQ_LOCK(kq);
1782 				kn->kn_status &= ~KN_INFLUX;
1783 				if (error)
1784 					KNOTE_ACTIVATE(kn, 1);
1785 				KQ_UNLOCK_FLUX(kq);
1786 			} else {
1787 				kn->kn_status |= KN_HASKQLOCK;
1788 				if (kn->kn_fop->f_event(kn, hint))
1789 					KNOTE_ACTIVATE(kn, 1);
1790 				kn->kn_status &= ~KN_HASKQLOCK;
1791 				KQ_UNLOCK(kq);
1792 			}
1793 		}
1794 		kq = NULL;
1795 	}
1796 	if ((lockflags & KNF_LISTLOCKED) == 0)
1797 		list->kl_unlock(list->kl_lockarg);
1798 }
1799 
1800 /*
1801  * add a knote to a knlist
1802  */
1803 void
1804 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1805 {
1806 	KNL_ASSERT_LOCK(knl, islocked);
1807 	KQ_NOTOWNED(kn->kn_kq);
1808 	KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1809 	    (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1810 	if (!islocked)
1811 		knl->kl_lock(knl->kl_lockarg);
1812 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1813 	if (!islocked)
1814 		knl->kl_unlock(knl->kl_lockarg);
1815 	KQ_LOCK(kn->kn_kq);
1816 	kn->kn_knlist = knl;
1817 	kn->kn_status &= ~KN_DETACHED;
1818 	KQ_UNLOCK(kn->kn_kq);
1819 }
1820 
1821 static void
1822 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1823 {
1824 	KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1825 	KNL_ASSERT_LOCK(knl, knlislocked);
1826 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1827 	if (!kqislocked)
1828 		KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1829     ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1830 	if (!knlislocked)
1831 		knl->kl_lock(knl->kl_lockarg);
1832 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1833 	kn->kn_knlist = NULL;
1834 	if (!knlislocked)
1835 		knl->kl_unlock(knl->kl_lockarg);
1836 	if (!kqislocked)
1837 		KQ_LOCK(kn->kn_kq);
1838 	kn->kn_status |= KN_DETACHED;
1839 	if (!kqislocked)
1840 		KQ_UNLOCK(kn->kn_kq);
1841 }
1842 
1843 /*
1844  * remove all knotes from a specified klist
1845  */
1846 void
1847 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1848 {
1849 
1850 	knlist_remove_kq(knl, kn, islocked, 0);
1851 }
1852 
1853 /*
1854  * remove knote from a specified klist while in f_event handler.
1855  */
1856 void
1857 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1858 {
1859 
1860 	knlist_remove_kq(knl, kn, 1,
1861 	    (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1862 }
1863 
1864 int
1865 knlist_empty(struct knlist *knl)
1866 {
1867 
1868 	KNL_ASSERT_LOCKED(knl);
1869 	return SLIST_EMPTY(&knl->kl_list);
1870 }
1871 
1872 static struct mtx	knlist_lock;
1873 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
1874 	MTX_DEF);
1875 static void knlist_mtx_lock(void *arg);
1876 static void knlist_mtx_unlock(void *arg);
1877 
1878 static void
1879 knlist_mtx_lock(void *arg)
1880 {
1881 
1882 	mtx_lock((struct mtx *)arg);
1883 }
1884 
1885 static void
1886 knlist_mtx_unlock(void *arg)
1887 {
1888 
1889 	mtx_unlock((struct mtx *)arg);
1890 }
1891 
1892 static void
1893 knlist_mtx_assert_locked(void *arg)
1894 {
1895 
1896 	mtx_assert((struct mtx *)arg, MA_OWNED);
1897 }
1898 
1899 static void
1900 knlist_mtx_assert_unlocked(void *arg)
1901 {
1902 
1903 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
1904 }
1905 
1906 static void
1907 knlist_rw_rlock(void *arg)
1908 {
1909 
1910 	rw_rlock((struct rwlock *)arg);
1911 }
1912 
1913 static void
1914 knlist_rw_runlock(void *arg)
1915 {
1916 
1917 	rw_runlock((struct rwlock *)arg);
1918 }
1919 
1920 static void
1921 knlist_rw_assert_locked(void *arg)
1922 {
1923 
1924 	rw_assert((struct rwlock *)arg, RA_LOCKED);
1925 }
1926 
1927 static void
1928 knlist_rw_assert_unlocked(void *arg)
1929 {
1930 
1931 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
1932 }
1933 
1934 void
1935 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
1936     void (*kl_unlock)(void *),
1937     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
1938 {
1939 
1940 	if (lock == NULL)
1941 		knl->kl_lockarg = &knlist_lock;
1942 	else
1943 		knl->kl_lockarg = lock;
1944 
1945 	if (kl_lock == NULL)
1946 		knl->kl_lock = knlist_mtx_lock;
1947 	else
1948 		knl->kl_lock = kl_lock;
1949 	if (kl_unlock == NULL)
1950 		knl->kl_unlock = knlist_mtx_unlock;
1951 	else
1952 		knl->kl_unlock = kl_unlock;
1953 	if (kl_assert_locked == NULL)
1954 		knl->kl_assert_locked = knlist_mtx_assert_locked;
1955 	else
1956 		knl->kl_assert_locked = kl_assert_locked;
1957 	if (kl_assert_unlocked == NULL)
1958 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
1959 	else
1960 		knl->kl_assert_unlocked = kl_assert_unlocked;
1961 
1962 	SLIST_INIT(&knl->kl_list);
1963 }
1964 
1965 void
1966 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
1967 {
1968 
1969 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
1970 }
1971 
1972 void
1973 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
1974 {
1975 
1976 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
1977 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
1978 }
1979 
1980 void
1981 knlist_destroy(struct knlist *knl)
1982 {
1983 
1984 #ifdef INVARIANTS
1985 	/*
1986 	 * if we run across this error, we need to find the offending
1987 	 * driver and have it call knlist_clear.
1988 	 */
1989 	if (!SLIST_EMPTY(&knl->kl_list))
1990 		printf("WARNING: destroying knlist w/ knotes on it!\n");
1991 #endif
1992 
1993 	knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
1994 	SLIST_INIT(&knl->kl_list);
1995 }
1996 
1997 /*
1998  * Even if we are locked, we may need to drop the lock to allow any influx
1999  * knotes time to "settle".
2000  */
2001 void
2002 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2003 {
2004 	struct knote *kn, *kn2;
2005 	struct kqueue *kq;
2006 
2007 	if (islocked)
2008 		KNL_ASSERT_LOCKED(knl);
2009 	else {
2010 		KNL_ASSERT_UNLOCKED(knl);
2011 again:		/* need to reacquire lock since we have dropped it */
2012 		knl->kl_lock(knl->kl_lockarg);
2013 	}
2014 
2015 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2016 		kq = kn->kn_kq;
2017 		KQ_LOCK(kq);
2018 		if ((kn->kn_status & KN_INFLUX)) {
2019 			KQ_UNLOCK(kq);
2020 			continue;
2021 		}
2022 		knlist_remove_kq(knl, kn, 1, 1);
2023 		if (killkn) {
2024 			kn->kn_status |= KN_INFLUX | KN_DETACHED;
2025 			KQ_UNLOCK(kq);
2026 			knote_drop(kn, td);
2027 		} else {
2028 			/* Make sure cleared knotes disappear soon */
2029 			kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2030 			KQ_UNLOCK(kq);
2031 		}
2032 		kq = NULL;
2033 	}
2034 
2035 	if (!SLIST_EMPTY(&knl->kl_list)) {
2036 		/* there are still KN_INFLUX remaining */
2037 		kn = SLIST_FIRST(&knl->kl_list);
2038 		kq = kn->kn_kq;
2039 		KQ_LOCK(kq);
2040 		KASSERT(kn->kn_status & KN_INFLUX,
2041 		    ("knote removed w/o list lock"));
2042 		knl->kl_unlock(knl->kl_lockarg);
2043 		kq->kq_state |= KQ_FLUXWAIT;
2044 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2045 		kq = NULL;
2046 		goto again;
2047 	}
2048 
2049 	if (islocked)
2050 		KNL_ASSERT_LOCKED(knl);
2051 	else {
2052 		knl->kl_unlock(knl->kl_lockarg);
2053 		KNL_ASSERT_UNLOCKED(knl);
2054 	}
2055 }
2056 
2057 /*
2058  * Remove all knotes referencing a specified fd must be called with FILEDESC
2059  * lock.  This prevents a race where a new fd comes along and occupies the
2060  * entry and we attach a knote to the fd.
2061  */
2062 void
2063 knote_fdclose(struct thread *td, int fd)
2064 {
2065 	struct filedesc *fdp = td->td_proc->p_fd;
2066 	struct kqueue *kq;
2067 	struct knote *kn;
2068 	int influx;
2069 
2070 	FILEDESC_XLOCK_ASSERT(fdp);
2071 
2072 	/*
2073 	 * We shouldn't have to worry about new kevents appearing on fd
2074 	 * since filedesc is locked.
2075 	 */
2076 	SLIST_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2077 		KQ_LOCK(kq);
2078 
2079 again:
2080 		influx = 0;
2081 		while (kq->kq_knlistsize > fd &&
2082 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2083 			if (kn->kn_status & KN_INFLUX) {
2084 				/* someone else might be waiting on our knote */
2085 				if (influx)
2086 					wakeup(kq);
2087 				kq->kq_state |= KQ_FLUXWAIT;
2088 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2089 				goto again;
2090 			}
2091 			kn->kn_status |= KN_INFLUX;
2092 			KQ_UNLOCK(kq);
2093 			if (!(kn->kn_status & KN_DETACHED))
2094 				kn->kn_fop->f_detach(kn);
2095 			knote_drop(kn, td);
2096 			influx = 1;
2097 			KQ_LOCK(kq);
2098 		}
2099 		KQ_UNLOCK_FLUX(kq);
2100 	}
2101 }
2102 
2103 static int
2104 knote_attach(struct knote *kn, struct kqueue *kq)
2105 {
2106 	struct klist *list;
2107 
2108 	KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2109 	KQ_OWNED(kq);
2110 
2111 	if (kn->kn_fop->f_isfd) {
2112 		if (kn->kn_id >= kq->kq_knlistsize)
2113 			return ENOMEM;
2114 		list = &kq->kq_knlist[kn->kn_id];
2115 	} else {
2116 		if (kq->kq_knhash == NULL)
2117 			return ENOMEM;
2118 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2119 	}
2120 
2121 	SLIST_INSERT_HEAD(list, kn, kn_link);
2122 
2123 	return 0;
2124 }
2125 
2126 /*
2127  * knote must already have been detached using the f_detach method.
2128  * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2129  * to prevent other removal.
2130  */
2131 static void
2132 knote_drop(struct knote *kn, struct thread *td)
2133 {
2134 	struct kqueue *kq;
2135 	struct klist *list;
2136 
2137 	kq = kn->kn_kq;
2138 
2139 	KQ_NOTOWNED(kq);
2140 	KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2141 	    ("knote_drop called without KN_INFLUX set in kn_status"));
2142 
2143 	KQ_LOCK(kq);
2144 	if (kn->kn_fop->f_isfd)
2145 		list = &kq->kq_knlist[kn->kn_id];
2146 	else
2147 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2148 
2149 	if (!SLIST_EMPTY(list))
2150 		SLIST_REMOVE(list, kn, knote, kn_link);
2151 	if (kn->kn_status & KN_QUEUED)
2152 		knote_dequeue(kn);
2153 	KQ_UNLOCK_FLUX(kq);
2154 
2155 	if (kn->kn_fop->f_isfd) {
2156 		fdrop(kn->kn_fp, td);
2157 		kn->kn_fp = NULL;
2158 	}
2159 	kqueue_fo_release(kn->kn_kevent.filter);
2160 	kn->kn_fop = NULL;
2161 	knote_free(kn);
2162 }
2163 
2164 static void
2165 knote_enqueue(struct knote *kn)
2166 {
2167 	struct kqueue *kq = kn->kn_kq;
2168 
2169 	KQ_OWNED(kn->kn_kq);
2170 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2171 
2172 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2173 	kn->kn_status |= KN_QUEUED;
2174 	kq->kq_count++;
2175 	kqueue_wakeup(kq);
2176 }
2177 
2178 static void
2179 knote_dequeue(struct knote *kn)
2180 {
2181 	struct kqueue *kq = kn->kn_kq;
2182 
2183 	KQ_OWNED(kn->kn_kq);
2184 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2185 
2186 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2187 	kn->kn_status &= ~KN_QUEUED;
2188 	kq->kq_count--;
2189 }
2190 
2191 static void
2192 knote_init(void)
2193 {
2194 
2195 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2196 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2197 }
2198 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2199 
2200 static struct knote *
2201 knote_alloc(int waitok)
2202 {
2203 	return ((struct knote *)uma_zalloc(knote_zone,
2204 	    (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2205 }
2206 
2207 static void
2208 knote_free(struct knote *kn)
2209 {
2210 	if (kn != NULL)
2211 		uma_zfree(knote_zone, kn);
2212 }
2213 
2214 /*
2215  * Register the kev w/ the kq specified by fd.
2216  */
2217 int
2218 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2219 {
2220 	struct kqueue *kq;
2221 	struct file *fp;
2222 	int error;
2223 
2224 	if ((error = fget(td, fd, CAP_POST_EVENT, &fp)) != 0)
2225 		return (error);
2226 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2227 		goto noacquire;
2228 
2229 	error = kqueue_register(kq, kev, td, waitok);
2230 
2231 	kqueue_release(kq, 0);
2232 
2233 noacquire:
2234 	fdrop(fp, td);
2235 
2236 	return error;
2237 }
2238