xref: /freebsd/sys/kern/kern_event.c (revision 2d0d168606f477d58d1a9aa72f94915da2f1733a)
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 	.fo_sendfile = invfo_sendfile,
131 };
132 
133 static int 	knote_attach(struct knote *kn, struct kqueue *kq);
134 static void 	knote_drop(struct knote *kn, struct thread *td);
135 static void 	knote_enqueue(struct knote *kn);
136 static void 	knote_dequeue(struct knote *kn);
137 static void 	knote_init(void);
138 static struct 	knote *knote_alloc(int waitok);
139 static void 	knote_free(struct knote *kn);
140 
141 static void	filt_kqdetach(struct knote *kn);
142 static int	filt_kqueue(struct knote *kn, long hint);
143 static int	filt_procattach(struct knote *kn);
144 static void	filt_procdetach(struct knote *kn);
145 static int	filt_proc(struct knote *kn, long hint);
146 static int	filt_fileattach(struct knote *kn);
147 static void	filt_timerexpire(void *knx);
148 static int	filt_timerattach(struct knote *kn);
149 static void	filt_timerdetach(struct knote *kn);
150 static int	filt_timer(struct knote *kn, long hint);
151 static int	filt_userattach(struct knote *kn);
152 static void	filt_userdetach(struct knote *kn);
153 static int	filt_user(struct knote *kn, long hint);
154 static void	filt_usertouch(struct knote *kn, struct kevent *kev,
155 		    u_long type);
156 
157 static struct filterops file_filtops = {
158 	.f_isfd = 1,
159 	.f_attach = filt_fileattach,
160 };
161 static struct filterops kqread_filtops = {
162 	.f_isfd = 1,
163 	.f_detach = filt_kqdetach,
164 	.f_event = filt_kqueue,
165 };
166 /* XXX - move to kern_proc.c?  */
167 static struct filterops proc_filtops = {
168 	.f_isfd = 0,
169 	.f_attach = filt_procattach,
170 	.f_detach = filt_procdetach,
171 	.f_event = filt_proc,
172 };
173 static struct filterops timer_filtops = {
174 	.f_isfd = 0,
175 	.f_attach = filt_timerattach,
176 	.f_detach = filt_timerdetach,
177 	.f_event = filt_timer,
178 };
179 static struct filterops user_filtops = {
180 	.f_attach = filt_userattach,
181 	.f_detach = filt_userdetach,
182 	.f_event = filt_user,
183 	.f_touch = filt_usertouch,
184 };
185 
186 static uma_zone_t	knote_zone;
187 static atomic_uint	kq_ncallouts = ATOMIC_VAR_INIT(0);
188 static unsigned int 	kq_calloutmax = 4 * 1024;
189 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
190     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
191 
192 /* XXX - ensure not KN_INFLUX?? */
193 #define KNOTE_ACTIVATE(kn, islock) do { 				\
194 	if ((islock))							\
195 		mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED);		\
196 	else								\
197 		KQ_LOCK((kn)->kn_kq);					\
198 	(kn)->kn_status |= KN_ACTIVE;					\
199 	if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)		\
200 		knote_enqueue((kn));					\
201 	if (!(islock))							\
202 		KQ_UNLOCK((kn)->kn_kq);					\
203 } while(0)
204 #define KQ_LOCK(kq) do {						\
205 	mtx_lock(&(kq)->kq_lock);					\
206 } while (0)
207 #define KQ_FLUX_WAKEUP(kq) do {						\
208 	if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) {		\
209 		(kq)->kq_state &= ~KQ_FLUXWAIT;				\
210 		wakeup((kq));						\
211 	}								\
212 } while (0)
213 #define KQ_UNLOCK_FLUX(kq) do {						\
214 	KQ_FLUX_WAKEUP(kq);						\
215 	mtx_unlock(&(kq)->kq_lock);					\
216 } while (0)
217 #define KQ_UNLOCK(kq) do {						\
218 	mtx_unlock(&(kq)->kq_lock);					\
219 } while (0)
220 #define KQ_OWNED(kq) do {						\
221 	mtx_assert(&(kq)->kq_lock, MA_OWNED);				\
222 } while (0)
223 #define KQ_NOTOWNED(kq) do {						\
224 	mtx_assert(&(kq)->kq_lock, MA_NOTOWNED);			\
225 } while (0)
226 #define KN_LIST_LOCK(kn) do {						\
227 	if (kn->kn_knlist != NULL)					\
228 		kn->kn_knlist->kl_lock(kn->kn_knlist->kl_lockarg);	\
229 } while (0)
230 #define KN_LIST_UNLOCK(kn) do {						\
231 	if (kn->kn_knlist != NULL) 					\
232 		kn->kn_knlist->kl_unlock(kn->kn_knlist->kl_lockarg);	\
233 } while (0)
234 #define	KNL_ASSERT_LOCK(knl, islocked) do {				\
235 	if (islocked)							\
236 		KNL_ASSERT_LOCKED(knl);				\
237 	else								\
238 		KNL_ASSERT_UNLOCKED(knl);				\
239 } while (0)
240 #ifdef INVARIANTS
241 #define	KNL_ASSERT_LOCKED(knl) do {					\
242 	knl->kl_assert_locked((knl)->kl_lockarg);			\
243 } while (0)
244 #define	KNL_ASSERT_UNLOCKED(knl) do {					\
245 	knl->kl_assert_unlocked((knl)->kl_lockarg);			\
246 } while (0)
247 #else /* !INVARIANTS */
248 #define	KNL_ASSERT_LOCKED(knl) do {} while(0)
249 #define	KNL_ASSERT_UNLOCKED(knl) do {} while (0)
250 #endif /* INVARIANTS */
251 
252 #define	KN_HASHSIZE		64		/* XXX should be tunable */
253 #define KN_HASH(val, mask)	(((val) ^ (val >> 8)) & (mask))
254 
255 static int
256 filt_nullattach(struct knote *kn)
257 {
258 
259 	return (ENXIO);
260 };
261 
262 struct filterops null_filtops = {
263 	.f_isfd = 0,
264 	.f_attach = filt_nullattach,
265 };
266 
267 /* XXX - make SYSINIT to add these, and move into respective modules. */
268 extern struct filterops sig_filtops;
269 extern struct filterops fs_filtops;
270 
271 /*
272  * Table for for all system-defined filters.
273  */
274 static struct mtx	filterops_lock;
275 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
276 	MTX_DEF);
277 static struct {
278 	struct filterops *for_fop;
279 	int for_refcnt;
280 } sysfilt_ops[EVFILT_SYSCOUNT] = {
281 	{ &file_filtops },			/* EVFILT_READ */
282 	{ &file_filtops },			/* EVFILT_WRITE */
283 	{ &null_filtops },			/* EVFILT_AIO */
284 	{ &file_filtops },			/* EVFILT_VNODE */
285 	{ &proc_filtops },			/* EVFILT_PROC */
286 	{ &sig_filtops },			/* EVFILT_SIGNAL */
287 	{ &timer_filtops },			/* EVFILT_TIMER */
288 	{ &null_filtops },			/* former EVFILT_NETDEV */
289 	{ &fs_filtops },			/* EVFILT_FS */
290 	{ &null_filtops },			/* EVFILT_LIO */
291 	{ &user_filtops },			/* EVFILT_USER */
292 };
293 
294 /*
295  * Simple redirection for all cdevsw style objects to call their fo_kqfilter
296  * method.
297  */
298 static int
299 filt_fileattach(struct knote *kn)
300 {
301 
302 	return (fo_kqfilter(kn->kn_fp, kn));
303 }
304 
305 /*ARGSUSED*/
306 static int
307 kqueue_kqfilter(struct file *fp, struct knote *kn)
308 {
309 	struct kqueue *kq = kn->kn_fp->f_data;
310 
311 	if (kn->kn_filter != EVFILT_READ)
312 		return (EINVAL);
313 
314 	kn->kn_status |= KN_KQUEUE;
315 	kn->kn_fop = &kqread_filtops;
316 	knlist_add(&kq->kq_sel.si_note, kn, 0);
317 
318 	return (0);
319 }
320 
321 static void
322 filt_kqdetach(struct knote *kn)
323 {
324 	struct kqueue *kq = kn->kn_fp->f_data;
325 
326 	knlist_remove(&kq->kq_sel.si_note, kn, 0);
327 }
328 
329 /*ARGSUSED*/
330 static int
331 filt_kqueue(struct knote *kn, long hint)
332 {
333 	struct kqueue *kq = kn->kn_fp->f_data;
334 
335 	kn->kn_data = kq->kq_count;
336 	return (kn->kn_data > 0);
337 }
338 
339 /* XXX - move to kern_proc.c?  */
340 static int
341 filt_procattach(struct knote *kn)
342 {
343 	struct proc *p;
344 	int immediate;
345 	int error;
346 
347 	immediate = 0;
348 	p = pfind(kn->kn_id);
349 	if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
350 		p = zpfind(kn->kn_id);
351 		immediate = 1;
352 	} else if (p != NULL && (p->p_flag & P_WEXIT)) {
353 		immediate = 1;
354 	}
355 
356 	if (p == NULL)
357 		return (ESRCH);
358 	if ((error = p_cansee(curthread, p))) {
359 		PROC_UNLOCK(p);
360 		return (error);
361 	}
362 
363 	kn->kn_ptr.p_proc = p;
364 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
365 
366 	/*
367 	 * internal flag indicating registration done by kernel
368 	 */
369 	if (kn->kn_flags & EV_FLAG1) {
370 		kn->kn_data = kn->kn_sdata;		/* ppid */
371 		kn->kn_fflags = NOTE_CHILD;
372 		kn->kn_flags &= ~EV_FLAG1;
373 	}
374 
375 	if (immediate == 0)
376 		knlist_add(&p->p_klist, kn, 1);
377 
378 	/*
379 	 * Immediately activate any exit notes if the target process is a
380 	 * zombie.  This is necessary to handle the case where the target
381 	 * process, e.g. a child, dies before the kevent is registered.
382 	 */
383 	if (immediate && filt_proc(kn, NOTE_EXIT))
384 		KNOTE_ACTIVATE(kn, 0);
385 
386 	PROC_UNLOCK(p);
387 
388 	return (0);
389 }
390 
391 /*
392  * The knote may be attached to a different process, which may exit,
393  * leaving nothing for the knote to be attached to.  So when the process
394  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
395  * it will be deleted when read out.  However, as part of the knote deletion,
396  * this routine is called, so a check is needed to avoid actually performing
397  * a detach, because the original process does not exist any more.
398  */
399 /* XXX - move to kern_proc.c?  */
400 static void
401 filt_procdetach(struct knote *kn)
402 {
403 	struct proc *p;
404 
405 	p = kn->kn_ptr.p_proc;
406 	knlist_remove(&p->p_klist, kn, 0);
407 	kn->kn_ptr.p_proc = NULL;
408 }
409 
410 /* XXX - move to kern_proc.c?  */
411 static int
412 filt_proc(struct knote *kn, long hint)
413 {
414 	struct proc *p = kn->kn_ptr.p_proc;
415 	u_int event;
416 
417 	/*
418 	 * mask off extra data
419 	 */
420 	event = (u_int)hint & NOTE_PCTRLMASK;
421 
422 	/*
423 	 * if the user is interested in this event, record it.
424 	 */
425 	if (kn->kn_sfflags & event)
426 		kn->kn_fflags |= event;
427 
428 	/*
429 	 * process is gone, so flag the event as finished.
430 	 */
431 	if (event == NOTE_EXIT) {
432 		if (!(kn->kn_status & KN_DETACHED))
433 			knlist_remove_inevent(&p->p_klist, kn);
434 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
435 		kn->kn_ptr.p_proc = NULL;
436 		if (kn->kn_fflags & NOTE_EXIT)
437 			kn->kn_data = p->p_xstat;
438 		if (kn->kn_fflags == 0)
439 			kn->kn_flags |= EV_DROP;
440 		return (1);
441 	}
442 
443 	return (kn->kn_fflags != 0);
444 }
445 
446 /*
447  * Called when the process forked. It mostly does the same as the
448  * knote(), activating all knotes registered to be activated when the
449  * process forked. Additionally, for each knote attached to the
450  * parent, check whether user wants to track the new process. If so
451  * attach a new knote to it, and immediately report an event with the
452  * child's pid.
453  */
454 void
455 knote_fork(struct knlist *list, int pid)
456 {
457 	struct kqueue *kq;
458 	struct knote *kn;
459 	struct kevent kev;
460 	int error;
461 
462 	if (list == NULL)
463 		return;
464 	list->kl_lock(list->kl_lockarg);
465 
466 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
467 		if ((kn->kn_status & KN_INFLUX) == KN_INFLUX)
468 			continue;
469 		kq = kn->kn_kq;
470 		KQ_LOCK(kq);
471 		if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
472 			KQ_UNLOCK(kq);
473 			continue;
474 		}
475 
476 		/*
477 		 * The same as knote(), activate the event.
478 		 */
479 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
480 			kn->kn_status |= KN_HASKQLOCK;
481 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
482 				KNOTE_ACTIVATE(kn, 1);
483 			kn->kn_status &= ~KN_HASKQLOCK;
484 			KQ_UNLOCK(kq);
485 			continue;
486 		}
487 
488 		/*
489 		 * The NOTE_TRACK case. In addition to the activation
490 		 * of the event, we need to register new event to
491 		 * track the child. Drop the locks in preparation for
492 		 * the call to kqueue_register().
493 		 */
494 		kn->kn_status |= KN_INFLUX;
495 		KQ_UNLOCK(kq);
496 		list->kl_unlock(list->kl_lockarg);
497 
498 		/*
499 		 * Activate existing knote and register a knote with
500 		 * new process.
501 		 */
502 		kev.ident = pid;
503 		kev.filter = kn->kn_filter;
504 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
505 		kev.fflags = kn->kn_sfflags;
506 		kev.data = kn->kn_id;		/* parent */
507 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
508 		error = kqueue_register(kq, &kev, NULL, 0);
509 		if (error)
510 			kn->kn_fflags |= NOTE_TRACKERR;
511 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
512 			KNOTE_ACTIVATE(kn, 0);
513 		KQ_LOCK(kq);
514 		kn->kn_status &= ~KN_INFLUX;
515 		KQ_UNLOCK_FLUX(kq);
516 		list->kl_lock(list->kl_lockarg);
517 	}
518 	list->kl_unlock(list->kl_lockarg);
519 }
520 
521 /*
522  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
523  * interval timer support code.
524  */
525 static __inline sbintime_t
526 timer2sbintime(intptr_t data)
527 {
528 
529 	return (SBT_1MS * data);
530 }
531 
532 static void
533 filt_timerexpire(void *knx)
534 {
535 	struct callout *calloutp;
536 	struct knote *kn;
537 
538 	kn = knx;
539 	kn->kn_data++;
540 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
541 
542 	if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
543 		calloutp = (struct callout *)kn->kn_hook;
544 		callout_reset_sbt_on(calloutp,
545 		    timer2sbintime(kn->kn_sdata), 0 /* 1ms? */,
546 		    filt_timerexpire, kn, PCPU_GET(cpuid), 0);
547 	}
548 }
549 
550 /*
551  * data contains amount of time to sleep, in milliseconds
552  */
553 static int
554 filt_timerattach(struct knote *kn)
555 {
556 	struct callout *calloutp;
557 	sbintime_t to;
558 	unsigned int ncallouts;
559 
560 	if ((intptr_t)kn->kn_sdata < 0)
561 		return (EINVAL);
562 	if ((intptr_t)kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
563 		kn->kn_sdata = 1;
564 	to = timer2sbintime(kn->kn_sdata);
565 	if (to < 0)
566 		return (EINVAL);
567 
568 	ncallouts = atomic_load_explicit(&kq_ncallouts, memory_order_relaxed);
569 	do {
570 		if (ncallouts >= kq_calloutmax)
571 			return (ENOMEM);
572 	} while (!atomic_compare_exchange_weak_explicit(&kq_ncallouts,
573 	    &ncallouts, ncallouts + 1, memory_order_relaxed,
574 	    memory_order_relaxed));
575 
576 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
577 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
578 	calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
579 	callout_init(calloutp, CALLOUT_MPSAFE);
580 	kn->kn_hook = calloutp;
581 	callout_reset_sbt_on(calloutp, to, 0 /* 1ms? */,
582 	    filt_timerexpire, kn, PCPU_GET(cpuid), 0);
583 
584 	return (0);
585 }
586 
587 static void
588 filt_timerdetach(struct knote *kn)
589 {
590 	struct callout *calloutp;
591 	unsigned int old;
592 
593 	calloutp = (struct callout *)kn->kn_hook;
594 	callout_drain(calloutp);
595 	free(calloutp, M_KQUEUE);
596 	old = atomic_fetch_sub_explicit(&kq_ncallouts, 1, memory_order_relaxed);
597 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
598 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
599 }
600 
601 static int
602 filt_timer(struct knote *kn, long hint)
603 {
604 
605 	return (kn->kn_data != 0);
606 }
607 
608 static int
609 filt_userattach(struct knote *kn)
610 {
611 
612 	/*
613 	 * EVFILT_USER knotes are not attached to anything in the kernel.
614 	 */
615 	kn->kn_hook = NULL;
616 	if (kn->kn_fflags & NOTE_TRIGGER)
617 		kn->kn_hookid = 1;
618 	else
619 		kn->kn_hookid = 0;
620 	return (0);
621 }
622 
623 static void
624 filt_userdetach(__unused struct knote *kn)
625 {
626 
627 	/*
628 	 * EVFILT_USER knotes are not attached to anything in the kernel.
629 	 */
630 }
631 
632 static int
633 filt_user(struct knote *kn, __unused long hint)
634 {
635 
636 	return (kn->kn_hookid);
637 }
638 
639 static void
640 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
641 {
642 	u_int ffctrl;
643 
644 	switch (type) {
645 	case EVENT_REGISTER:
646 		if (kev->fflags & NOTE_TRIGGER)
647 			kn->kn_hookid = 1;
648 
649 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
650 		kev->fflags &= NOTE_FFLAGSMASK;
651 		switch (ffctrl) {
652 		case NOTE_FFNOP:
653 			break;
654 
655 		case NOTE_FFAND:
656 			kn->kn_sfflags &= kev->fflags;
657 			break;
658 
659 		case NOTE_FFOR:
660 			kn->kn_sfflags |= kev->fflags;
661 			break;
662 
663 		case NOTE_FFCOPY:
664 			kn->kn_sfflags = kev->fflags;
665 			break;
666 
667 		default:
668 			/* XXX Return error? */
669 			break;
670 		}
671 		kn->kn_sdata = kev->data;
672 		if (kev->flags & EV_CLEAR) {
673 			kn->kn_hookid = 0;
674 			kn->kn_data = 0;
675 			kn->kn_fflags = 0;
676 		}
677 		break;
678 
679         case EVENT_PROCESS:
680 		*kev = kn->kn_kevent;
681 		kev->fflags = kn->kn_sfflags;
682 		kev->data = kn->kn_sdata;
683 		if (kn->kn_flags & EV_CLEAR) {
684 			kn->kn_hookid = 0;
685 			kn->kn_data = 0;
686 			kn->kn_fflags = 0;
687 		}
688 		break;
689 
690 	default:
691 		panic("filt_usertouch() - invalid type (%ld)", type);
692 		break;
693 	}
694 }
695 
696 int
697 sys_kqueue(struct thread *td, struct kqueue_args *uap)
698 {
699 	struct filedesc *fdp;
700 	struct kqueue *kq;
701 	struct file *fp;
702 	int fd, error;
703 
704 	fdp = td->td_proc->p_fd;
705 	error = falloc(td, &fp, &fd, 0);
706 	if (error)
707 		goto done2;
708 
709 	/* An extra reference on `fp' has been held for us by falloc(). */
710 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
711 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
712 	TAILQ_INIT(&kq->kq_head);
713 	kq->kq_fdp = fdp;
714 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
715 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
716 
717 	FILEDESC_XLOCK(fdp);
718 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
719 	FILEDESC_XUNLOCK(fdp);
720 
721 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
722 	fdrop(fp, td);
723 
724 	td->td_retval[0] = fd;
725 done2:
726 	return (error);
727 }
728 
729 #ifndef _SYS_SYSPROTO_H_
730 struct kevent_args {
731 	int	fd;
732 	const struct kevent *changelist;
733 	int	nchanges;
734 	struct	kevent *eventlist;
735 	int	nevents;
736 	const struct timespec *timeout;
737 };
738 #endif
739 int
740 sys_kevent(struct thread *td, struct kevent_args *uap)
741 {
742 	struct timespec ts, *tsp;
743 	struct kevent_copyops k_ops = { uap,
744 					kevent_copyout,
745 					kevent_copyin};
746 	int error;
747 #ifdef KTRACE
748 	struct uio ktruio;
749 	struct iovec ktriov;
750 	struct uio *ktruioin = NULL;
751 	struct uio *ktruioout = NULL;
752 #endif
753 
754 	if (uap->timeout != NULL) {
755 		error = copyin(uap->timeout, &ts, sizeof(ts));
756 		if (error)
757 			return (error);
758 		tsp = &ts;
759 	} else
760 		tsp = NULL;
761 
762 #ifdef KTRACE
763 	if (KTRPOINT(td, KTR_GENIO)) {
764 		ktriov.iov_base = uap->changelist;
765 		ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
766 		ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
767 		    .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
768 		    .uio_td = td };
769 		ktruioin = cloneuio(&ktruio);
770 		ktriov.iov_base = uap->eventlist;
771 		ktriov.iov_len = uap->nevents * sizeof(struct kevent);
772 		ktruioout = cloneuio(&ktruio);
773 	}
774 #endif
775 
776 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
777 	    &k_ops, tsp);
778 
779 #ifdef KTRACE
780 	if (ktruioin != NULL) {
781 		ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
782 		ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
783 		ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
784 		ktrgenio(uap->fd, UIO_READ, ktruioout, error);
785 	}
786 #endif
787 
788 	return (error);
789 }
790 
791 /*
792  * Copy 'count' items into the destination list pointed to by uap->eventlist.
793  */
794 static int
795 kevent_copyout(void *arg, struct kevent *kevp, int count)
796 {
797 	struct kevent_args *uap;
798 	int error;
799 
800 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
801 	uap = (struct kevent_args *)arg;
802 
803 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
804 	if (error == 0)
805 		uap->eventlist += count;
806 	return (error);
807 }
808 
809 /*
810  * Copy 'count' items from the list pointed to by uap->changelist.
811  */
812 static int
813 kevent_copyin(void *arg, struct kevent *kevp, int count)
814 {
815 	struct kevent_args *uap;
816 	int error;
817 
818 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
819 	uap = (struct kevent_args *)arg;
820 
821 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
822 	if (error == 0)
823 		uap->changelist += count;
824 	return (error);
825 }
826 
827 int
828 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
829     struct kevent_copyops *k_ops, const struct timespec *timeout)
830 {
831 	struct kevent keva[KQ_NEVENTS];
832 	struct kevent *kevp, *changes;
833 	struct kqueue *kq;
834 	struct file *fp;
835 	cap_rights_t rights;
836 	int i, n, nerrors, error;
837 
838 	error = fget(td, fd, cap_rights_init(&rights, CAP_POST_EVENT), &fp);
839 	if (error != 0)
840 		return (error);
841 	if ((error = kqueue_acquire(fp, &kq)) != 0)
842 		goto done_norel;
843 
844 	nerrors = 0;
845 
846 	while (nchanges > 0) {
847 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
848 		error = k_ops->k_copyin(k_ops->arg, keva, n);
849 		if (error)
850 			goto done;
851 		changes = keva;
852 		for (i = 0; i < n; i++) {
853 			kevp = &changes[i];
854 			if (!kevp->filter)
855 				continue;
856 			kevp->flags &= ~EV_SYSFLAGS;
857 			error = kqueue_register(kq, kevp, td, 1);
858 			if (error || (kevp->flags & EV_RECEIPT)) {
859 				if (nevents != 0) {
860 					kevp->flags = EV_ERROR;
861 					kevp->data = error;
862 					(void) k_ops->k_copyout(k_ops->arg,
863 					    kevp, 1);
864 					nevents--;
865 					nerrors++;
866 				} else {
867 					goto done;
868 				}
869 			}
870 		}
871 		nchanges -= n;
872 	}
873 	if (nerrors) {
874 		td->td_retval[0] = nerrors;
875 		error = 0;
876 		goto done;
877 	}
878 
879 	error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
880 done:
881 	kqueue_release(kq, 0);
882 done_norel:
883 	fdrop(fp, td);
884 	return (error);
885 }
886 
887 int
888 kqueue_add_filteropts(int filt, struct filterops *filtops)
889 {
890 	int error;
891 
892 	error = 0;
893 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
894 		printf(
895 "trying to add a filterop that is out of range: %d is beyond %d\n",
896 		    ~filt, EVFILT_SYSCOUNT);
897 		return EINVAL;
898 	}
899 	mtx_lock(&filterops_lock);
900 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
901 	    sysfilt_ops[~filt].for_fop != NULL)
902 		error = EEXIST;
903 	else {
904 		sysfilt_ops[~filt].for_fop = filtops;
905 		sysfilt_ops[~filt].for_refcnt = 0;
906 	}
907 	mtx_unlock(&filterops_lock);
908 
909 	return (error);
910 }
911 
912 int
913 kqueue_del_filteropts(int filt)
914 {
915 	int error;
916 
917 	error = 0;
918 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
919 		return EINVAL;
920 
921 	mtx_lock(&filterops_lock);
922 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
923 	    sysfilt_ops[~filt].for_fop == NULL)
924 		error = EINVAL;
925 	else if (sysfilt_ops[~filt].for_refcnt != 0)
926 		error = EBUSY;
927 	else {
928 		sysfilt_ops[~filt].for_fop = &null_filtops;
929 		sysfilt_ops[~filt].for_refcnt = 0;
930 	}
931 	mtx_unlock(&filterops_lock);
932 
933 	return error;
934 }
935 
936 static struct filterops *
937 kqueue_fo_find(int filt)
938 {
939 
940 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
941 		return NULL;
942 
943 	mtx_lock(&filterops_lock);
944 	sysfilt_ops[~filt].for_refcnt++;
945 	if (sysfilt_ops[~filt].for_fop == NULL)
946 		sysfilt_ops[~filt].for_fop = &null_filtops;
947 	mtx_unlock(&filterops_lock);
948 
949 	return sysfilt_ops[~filt].for_fop;
950 }
951 
952 static void
953 kqueue_fo_release(int filt)
954 {
955 
956 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
957 		return;
958 
959 	mtx_lock(&filterops_lock);
960 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
961 	    ("filter object refcount not valid on release"));
962 	sysfilt_ops[~filt].for_refcnt--;
963 	mtx_unlock(&filterops_lock);
964 }
965 
966 /*
967  * A ref to kq (obtained via kqueue_acquire) must be held.  waitok will
968  * influence if memory allocation should wait.  Make sure it is 0 if you
969  * hold any mutexes.
970  */
971 static int
972 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
973 {
974 	struct filterops *fops;
975 	struct file *fp;
976 	struct knote *kn, *tkn;
977 	cap_rights_t rights;
978 	int error, filt, event;
979 	int haskqglobal, filedesc_unlock;
980 
981 	fp = NULL;
982 	kn = NULL;
983 	error = 0;
984 	haskqglobal = 0;
985 	filedesc_unlock = 0;
986 
987 	filt = kev->filter;
988 	fops = kqueue_fo_find(filt);
989 	if (fops == NULL)
990 		return EINVAL;
991 
992 	tkn = knote_alloc(waitok);		/* prevent waiting with locks */
993 
994 findkn:
995 	if (fops->f_isfd) {
996 		KASSERT(td != NULL, ("td is NULL"));
997 		error = fget(td, kev->ident,
998 		    cap_rights_init(&rights, CAP_POLL_EVENT), &fp);
999 		if (error)
1000 			goto done;
1001 
1002 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1003 		    kev->ident, 0) != 0) {
1004 			/* try again */
1005 			fdrop(fp, td);
1006 			fp = NULL;
1007 			error = kqueue_expand(kq, fops, kev->ident, waitok);
1008 			if (error)
1009 				goto done;
1010 			goto findkn;
1011 		}
1012 
1013 		if (fp->f_type == DTYPE_KQUEUE) {
1014 			/*
1015 			 * if we add some inteligence about what we are doing,
1016 			 * we should be able to support events on ourselves.
1017 			 * We need to know when we are doing this to prevent
1018 			 * getting both the knlist lock and the kq lock since
1019 			 * they are the same thing.
1020 			 */
1021 			if (fp->f_data == kq) {
1022 				error = EINVAL;
1023 				goto done;
1024 			}
1025 
1026 			/*
1027 			 * Pre-lock the filedesc before the global
1028 			 * lock mutex, see the comment in
1029 			 * kqueue_close().
1030 			 */
1031 			FILEDESC_XLOCK(td->td_proc->p_fd);
1032 			filedesc_unlock = 1;
1033 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1034 		}
1035 
1036 		KQ_LOCK(kq);
1037 		if (kev->ident < kq->kq_knlistsize) {
1038 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1039 				if (kev->filter == kn->kn_filter)
1040 					break;
1041 		}
1042 	} else {
1043 		if ((kev->flags & EV_ADD) == EV_ADD)
1044 			kqueue_expand(kq, fops, kev->ident, waitok);
1045 
1046 		KQ_LOCK(kq);
1047 		if (kq->kq_knhashmask != 0) {
1048 			struct klist *list;
1049 
1050 			list = &kq->kq_knhash[
1051 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1052 			SLIST_FOREACH(kn, list, kn_link)
1053 				if (kev->ident == kn->kn_id &&
1054 				    kev->filter == kn->kn_filter)
1055 					break;
1056 		}
1057 	}
1058 
1059 	/* knote is in the process of changing, wait for it to stablize. */
1060 	if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1061 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1062 		if (filedesc_unlock) {
1063 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1064 			filedesc_unlock = 0;
1065 		}
1066 		kq->kq_state |= KQ_FLUXWAIT;
1067 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1068 		if (fp != NULL) {
1069 			fdrop(fp, td);
1070 			fp = NULL;
1071 		}
1072 		goto findkn;
1073 	}
1074 
1075 	/*
1076 	 * kn now contains the matching knote, or NULL if no match
1077 	 */
1078 	if (kn == NULL) {
1079 		if (kev->flags & EV_ADD) {
1080 			kn = tkn;
1081 			tkn = NULL;
1082 			if (kn == NULL) {
1083 				KQ_UNLOCK(kq);
1084 				error = ENOMEM;
1085 				goto done;
1086 			}
1087 			kn->kn_fp = fp;
1088 			kn->kn_kq = kq;
1089 			kn->kn_fop = fops;
1090 			/*
1091 			 * apply reference counts to knote structure, and
1092 			 * do not release it at the end of this routine.
1093 			 */
1094 			fops = NULL;
1095 			fp = NULL;
1096 
1097 			kn->kn_sfflags = kev->fflags;
1098 			kn->kn_sdata = kev->data;
1099 			kev->fflags = 0;
1100 			kev->data = 0;
1101 			kn->kn_kevent = *kev;
1102 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1103 			    EV_ENABLE | EV_DISABLE);
1104 			kn->kn_status = KN_INFLUX|KN_DETACHED;
1105 
1106 			error = knote_attach(kn, kq);
1107 			KQ_UNLOCK(kq);
1108 			if (error != 0) {
1109 				tkn = kn;
1110 				goto done;
1111 			}
1112 
1113 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1114 				knote_drop(kn, td);
1115 				goto done;
1116 			}
1117 			KN_LIST_LOCK(kn);
1118 			goto done_ev_add;
1119 		} else {
1120 			/* No matching knote and the EV_ADD flag is not set. */
1121 			KQ_UNLOCK(kq);
1122 			error = ENOENT;
1123 			goto done;
1124 		}
1125 	}
1126 
1127 	if (kev->flags & EV_DELETE) {
1128 		kn->kn_status |= KN_INFLUX;
1129 		KQ_UNLOCK(kq);
1130 		if (!(kn->kn_status & KN_DETACHED))
1131 			kn->kn_fop->f_detach(kn);
1132 		knote_drop(kn, td);
1133 		goto done;
1134 	}
1135 
1136 	/*
1137 	 * The user may change some filter values after the initial EV_ADD,
1138 	 * but doing so will not reset any filter which has already been
1139 	 * triggered.
1140 	 */
1141 	kn->kn_status |= KN_INFLUX;
1142 	KQ_UNLOCK(kq);
1143 	KN_LIST_LOCK(kn);
1144 	kn->kn_kevent.udata = kev->udata;
1145 	if (!fops->f_isfd && fops->f_touch != NULL) {
1146 		fops->f_touch(kn, kev, EVENT_REGISTER);
1147 	} else {
1148 		kn->kn_sfflags = kev->fflags;
1149 		kn->kn_sdata = kev->data;
1150 	}
1151 
1152 	/*
1153 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1154 	 * the initial attach event decides that the event is "completed"
1155 	 * already.  i.e. filt_procattach is called on a zombie process.  It
1156 	 * will call filt_proc which will remove it from the list, and NULL
1157 	 * kn_knlist.
1158 	 */
1159 done_ev_add:
1160 	event = kn->kn_fop->f_event(kn, 0);
1161 	KQ_LOCK(kq);
1162 	if (event)
1163 		KNOTE_ACTIVATE(kn, 1);
1164 	kn->kn_status &= ~KN_INFLUX;
1165 	KN_LIST_UNLOCK(kn);
1166 
1167 	if ((kev->flags & EV_DISABLE) &&
1168 	    ((kn->kn_status & KN_DISABLED) == 0)) {
1169 		kn->kn_status |= KN_DISABLED;
1170 	}
1171 
1172 	if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1173 		kn->kn_status &= ~KN_DISABLED;
1174 		if ((kn->kn_status & KN_ACTIVE) &&
1175 		    ((kn->kn_status & KN_QUEUED) == 0))
1176 			knote_enqueue(kn);
1177 	}
1178 	KQ_UNLOCK_FLUX(kq);
1179 
1180 done:
1181 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1182 	if (filedesc_unlock)
1183 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1184 	if (fp != NULL)
1185 		fdrop(fp, td);
1186 	if (tkn != NULL)
1187 		knote_free(tkn);
1188 	if (fops != NULL)
1189 		kqueue_fo_release(filt);
1190 	return (error);
1191 }
1192 
1193 static int
1194 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1195 {
1196 	int error;
1197 	struct kqueue *kq;
1198 
1199 	error = 0;
1200 
1201 	kq = fp->f_data;
1202 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1203 		return (EBADF);
1204 	*kqp = kq;
1205 	KQ_LOCK(kq);
1206 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1207 		KQ_UNLOCK(kq);
1208 		return (EBADF);
1209 	}
1210 	kq->kq_refcnt++;
1211 	KQ_UNLOCK(kq);
1212 
1213 	return error;
1214 }
1215 
1216 static void
1217 kqueue_release(struct kqueue *kq, int locked)
1218 {
1219 	if (locked)
1220 		KQ_OWNED(kq);
1221 	else
1222 		KQ_LOCK(kq);
1223 	kq->kq_refcnt--;
1224 	if (kq->kq_refcnt == 1)
1225 		wakeup(&kq->kq_refcnt);
1226 	if (!locked)
1227 		KQ_UNLOCK(kq);
1228 }
1229 
1230 static void
1231 kqueue_schedtask(struct kqueue *kq)
1232 {
1233 
1234 	KQ_OWNED(kq);
1235 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1236 	    ("scheduling kqueue task while draining"));
1237 
1238 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1239 		taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1240 		kq->kq_state |= KQ_TASKSCHED;
1241 	}
1242 }
1243 
1244 /*
1245  * Expand the kq to make sure we have storage for fops/ident pair.
1246  *
1247  * Return 0 on success (or no work necessary), return errno on failure.
1248  *
1249  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1250  * If kqueue_register is called from a non-fd context, there usually/should
1251  * be no locks held.
1252  */
1253 static int
1254 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1255 	int waitok)
1256 {
1257 	struct klist *list, *tmp_knhash, *to_free;
1258 	u_long tmp_knhashmask;
1259 	int size;
1260 	int fd;
1261 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
1262 
1263 	KQ_NOTOWNED(kq);
1264 
1265 	to_free = NULL;
1266 	if (fops->f_isfd) {
1267 		fd = ident;
1268 		if (kq->kq_knlistsize <= fd) {
1269 			size = kq->kq_knlistsize;
1270 			while (size <= fd)
1271 				size += KQEXTENT;
1272 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1273 			if (list == NULL)
1274 				return ENOMEM;
1275 			KQ_LOCK(kq);
1276 			if (kq->kq_knlistsize > fd) {
1277 				to_free = list;
1278 				list = NULL;
1279 			} else {
1280 				if (kq->kq_knlist != NULL) {
1281 					bcopy(kq->kq_knlist, list,
1282 					    kq->kq_knlistsize * sizeof(*list));
1283 					to_free = kq->kq_knlist;
1284 					kq->kq_knlist = NULL;
1285 				}
1286 				bzero((caddr_t)list +
1287 				    kq->kq_knlistsize * sizeof(*list),
1288 				    (size - kq->kq_knlistsize) * sizeof(*list));
1289 				kq->kq_knlistsize = size;
1290 				kq->kq_knlist = list;
1291 			}
1292 			KQ_UNLOCK(kq);
1293 		}
1294 	} else {
1295 		if (kq->kq_knhashmask == 0) {
1296 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1297 			    &tmp_knhashmask);
1298 			if (tmp_knhash == NULL)
1299 				return ENOMEM;
1300 			KQ_LOCK(kq);
1301 			if (kq->kq_knhashmask == 0) {
1302 				kq->kq_knhash = tmp_knhash;
1303 				kq->kq_knhashmask = tmp_knhashmask;
1304 			} else {
1305 				to_free = tmp_knhash;
1306 			}
1307 			KQ_UNLOCK(kq);
1308 		}
1309 	}
1310 	free(to_free, M_KQUEUE);
1311 
1312 	KQ_NOTOWNED(kq);
1313 	return 0;
1314 }
1315 
1316 static void
1317 kqueue_task(void *arg, int pending)
1318 {
1319 	struct kqueue *kq;
1320 	int haskqglobal;
1321 
1322 	haskqglobal = 0;
1323 	kq = arg;
1324 
1325 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1326 	KQ_LOCK(kq);
1327 
1328 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1329 
1330 	kq->kq_state &= ~KQ_TASKSCHED;
1331 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1332 		wakeup(&kq->kq_state);
1333 	}
1334 	KQ_UNLOCK(kq);
1335 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1336 }
1337 
1338 /*
1339  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1340  * We treat KN_MARKER knotes as if they are INFLUX.
1341  */
1342 static int
1343 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1344     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1345 {
1346 	struct kevent *kevp;
1347 	struct knote *kn, *marker;
1348 	sbintime_t asbt, rsbt;
1349 	int count, error, haskqglobal, influx, nkev, touch;
1350 
1351 	count = maxevents;
1352 	nkev = 0;
1353 	error = 0;
1354 	haskqglobal = 0;
1355 
1356 	if (maxevents == 0)
1357 		goto done_nl;
1358 
1359 	rsbt = 0;
1360 	if (tsp != NULL) {
1361 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1362 		    tsp->tv_nsec >= 1000000000) {
1363 			error = EINVAL;
1364 			goto done_nl;
1365 		}
1366 		if (timespecisset(tsp)) {
1367 			if (tsp->tv_sec <= INT32_MAX) {
1368 				rsbt = tstosbt(*tsp);
1369 				if (TIMESEL(&asbt, rsbt))
1370 					asbt += tc_tick_sbt;
1371 				if (asbt <= INT64_MAX - rsbt)
1372 					asbt += rsbt;
1373 				else
1374 					asbt = 0;
1375 				rsbt >>= tc_precexp;
1376 			} else
1377 				asbt = 0;
1378 		} else
1379 			asbt = -1;
1380 	} else
1381 		asbt = 0;
1382 	marker = knote_alloc(1);
1383 	if (marker == NULL) {
1384 		error = ENOMEM;
1385 		goto done_nl;
1386 	}
1387 	marker->kn_status = KN_MARKER;
1388 	KQ_LOCK(kq);
1389 
1390 retry:
1391 	kevp = keva;
1392 	if (kq->kq_count == 0) {
1393 		if (asbt == -1) {
1394 			error = EWOULDBLOCK;
1395 		} else {
1396 			kq->kq_state |= KQ_SLEEP;
1397 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1398 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1399 		}
1400 		if (error == 0)
1401 			goto retry;
1402 		/* don't restart after signals... */
1403 		if (error == ERESTART)
1404 			error = EINTR;
1405 		else if (error == EWOULDBLOCK)
1406 			error = 0;
1407 		goto done;
1408 	}
1409 
1410 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1411 	influx = 0;
1412 	while (count) {
1413 		KQ_OWNED(kq);
1414 		kn = TAILQ_FIRST(&kq->kq_head);
1415 
1416 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1417 		    (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1418 			if (influx) {
1419 				influx = 0;
1420 				KQ_FLUX_WAKEUP(kq);
1421 			}
1422 			kq->kq_state |= KQ_FLUXWAIT;
1423 			error = msleep(kq, &kq->kq_lock, PSOCK,
1424 			    "kqflxwt", 0);
1425 			continue;
1426 		}
1427 
1428 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1429 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1430 			kn->kn_status &= ~KN_QUEUED;
1431 			kq->kq_count--;
1432 			continue;
1433 		}
1434 		if (kn == marker) {
1435 			KQ_FLUX_WAKEUP(kq);
1436 			if (count == maxevents)
1437 				goto retry;
1438 			goto done;
1439 		}
1440 		KASSERT((kn->kn_status & KN_INFLUX) == 0,
1441 		    ("KN_INFLUX set when not suppose to be"));
1442 
1443 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1444 			kn->kn_status &= ~KN_QUEUED;
1445 			kn->kn_status |= KN_INFLUX;
1446 			kq->kq_count--;
1447 			KQ_UNLOCK(kq);
1448 			/*
1449 			 * We don't need to lock the list since we've marked
1450 			 * it _INFLUX.
1451 			 */
1452 			if (!(kn->kn_status & KN_DETACHED))
1453 				kn->kn_fop->f_detach(kn);
1454 			knote_drop(kn, td);
1455 			KQ_LOCK(kq);
1456 			continue;
1457 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1458 			kn->kn_status &= ~KN_QUEUED;
1459 			kn->kn_status |= KN_INFLUX;
1460 			kq->kq_count--;
1461 			KQ_UNLOCK(kq);
1462 			/*
1463 			 * We don't need to lock the list since we've marked
1464 			 * it _INFLUX.
1465 			 */
1466 			*kevp = kn->kn_kevent;
1467 			if (!(kn->kn_status & KN_DETACHED))
1468 				kn->kn_fop->f_detach(kn);
1469 			knote_drop(kn, td);
1470 			KQ_LOCK(kq);
1471 			kn = NULL;
1472 		} else {
1473 			kn->kn_status |= KN_INFLUX;
1474 			KQ_UNLOCK(kq);
1475 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1476 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1477 			KN_LIST_LOCK(kn);
1478 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1479 				KQ_LOCK(kq);
1480 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1481 				kn->kn_status &=
1482 				    ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX);
1483 				kq->kq_count--;
1484 				KN_LIST_UNLOCK(kn);
1485 				influx = 1;
1486 				continue;
1487 			}
1488 			touch = (!kn->kn_fop->f_isfd &&
1489 			    kn->kn_fop->f_touch != NULL);
1490 			if (touch)
1491 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1492 			else
1493 				*kevp = kn->kn_kevent;
1494 			KQ_LOCK(kq);
1495 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1496 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1497 				/*
1498 				 * Manually clear knotes who weren't
1499 				 * 'touch'ed.
1500 				 */
1501 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1502 					kn->kn_data = 0;
1503 					kn->kn_fflags = 0;
1504 				}
1505 				if (kn->kn_flags & EV_DISPATCH)
1506 					kn->kn_status |= KN_DISABLED;
1507 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1508 				kq->kq_count--;
1509 			} else
1510 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1511 
1512 			kn->kn_status &= ~(KN_INFLUX);
1513 			KN_LIST_UNLOCK(kn);
1514 			influx = 1;
1515 		}
1516 
1517 		/* we are returning a copy to the user */
1518 		kevp++;
1519 		nkev++;
1520 		count--;
1521 
1522 		if (nkev == KQ_NEVENTS) {
1523 			influx = 0;
1524 			KQ_UNLOCK_FLUX(kq);
1525 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1526 			nkev = 0;
1527 			kevp = keva;
1528 			KQ_LOCK(kq);
1529 			if (error)
1530 				break;
1531 		}
1532 	}
1533 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1534 done:
1535 	KQ_OWNED(kq);
1536 	KQ_UNLOCK_FLUX(kq);
1537 	knote_free(marker);
1538 done_nl:
1539 	KQ_NOTOWNED(kq);
1540 	if (nkev != 0)
1541 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1542 	td->td_retval[0] = maxevents - count;
1543 	return (error);
1544 }
1545 
1546 /*
1547  * XXX
1548  * This could be expanded to call kqueue_scan, if desired.
1549  */
1550 /*ARGSUSED*/
1551 static int
1552 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
1553 	int flags, struct thread *td)
1554 {
1555 	return (ENXIO);
1556 }
1557 
1558 /*ARGSUSED*/
1559 static int
1560 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1561 	 int flags, struct thread *td)
1562 {
1563 	return (ENXIO);
1564 }
1565 
1566 /*ARGSUSED*/
1567 static int
1568 kqueue_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1569 	struct thread *td)
1570 {
1571 
1572 	return (EINVAL);
1573 }
1574 
1575 /*ARGSUSED*/
1576 static int
1577 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1578 	struct ucred *active_cred, struct thread *td)
1579 {
1580 	/*
1581 	 * Enabling sigio causes two major problems:
1582 	 * 1) infinite recursion:
1583 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1584 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1585 	 * into itself over and over.  Sending the sigio causes the kqueue
1586 	 * to become ready, which in turn posts sigio again, forever.
1587 	 * Solution: this can be solved by setting a flag in the kqueue that
1588 	 * we have a SIGIO in progress.
1589 	 * 2) locking problems:
1590 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1591 	 * us above the proc and pgrp locks.
1592 	 * Solution: Post a signal using an async mechanism, being sure to
1593 	 * record a generation count in the delivery so that we do not deliver
1594 	 * a signal to the wrong process.
1595 	 *
1596 	 * Note, these two mechanisms are somewhat mutually exclusive!
1597 	 */
1598 #if 0
1599 	struct kqueue *kq;
1600 
1601 	kq = fp->f_data;
1602 	switch (cmd) {
1603 	case FIOASYNC:
1604 		if (*(int *)data) {
1605 			kq->kq_state |= KQ_ASYNC;
1606 		} else {
1607 			kq->kq_state &= ~KQ_ASYNC;
1608 		}
1609 		return (0);
1610 
1611 	case FIOSETOWN:
1612 		return (fsetown(*(int *)data, &kq->kq_sigio));
1613 
1614 	case FIOGETOWN:
1615 		*(int *)data = fgetown(&kq->kq_sigio);
1616 		return (0);
1617 	}
1618 #endif
1619 
1620 	return (ENOTTY);
1621 }
1622 
1623 /*ARGSUSED*/
1624 static int
1625 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1626 	struct thread *td)
1627 {
1628 	struct kqueue *kq;
1629 	int revents = 0;
1630 	int error;
1631 
1632 	if ((error = kqueue_acquire(fp, &kq)))
1633 		return POLLERR;
1634 
1635 	KQ_LOCK(kq);
1636 	if (events & (POLLIN | POLLRDNORM)) {
1637 		if (kq->kq_count) {
1638 			revents |= events & (POLLIN | POLLRDNORM);
1639 		} else {
1640 			selrecord(td, &kq->kq_sel);
1641 			if (SEL_WAITING(&kq->kq_sel))
1642 				kq->kq_state |= KQ_SEL;
1643 		}
1644 	}
1645 	kqueue_release(kq, 1);
1646 	KQ_UNLOCK(kq);
1647 	return (revents);
1648 }
1649 
1650 /*ARGSUSED*/
1651 static int
1652 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1653 	struct thread *td)
1654 {
1655 
1656 	bzero((void *)st, sizeof *st);
1657 	/*
1658 	 * We no longer return kq_count because the unlocked value is useless.
1659 	 * If you spent all this time getting the count, why not spend your
1660 	 * syscall better by calling kevent?
1661 	 *
1662 	 * XXX - This is needed for libc_r.
1663 	 */
1664 	st->st_mode = S_IFIFO;
1665 	return (0);
1666 }
1667 
1668 /*ARGSUSED*/
1669 static int
1670 kqueue_close(struct file *fp, struct thread *td)
1671 {
1672 	struct kqueue *kq = fp->f_data;
1673 	struct filedesc *fdp;
1674 	struct knote *kn;
1675 	int i;
1676 	int error;
1677 	int filedesc_unlock;
1678 
1679 	if ((error = kqueue_acquire(fp, &kq)))
1680 		return error;
1681 
1682 	filedesc_unlock = 0;
1683 	KQ_LOCK(kq);
1684 
1685 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1686 	    ("kqueue already closing"));
1687 	kq->kq_state |= KQ_CLOSING;
1688 	if (kq->kq_refcnt > 1)
1689 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1690 
1691 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1692 	fdp = kq->kq_fdp;
1693 
1694 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
1695 	    ("kqueue's knlist not empty"));
1696 
1697 	for (i = 0; i < kq->kq_knlistsize; i++) {
1698 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1699 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1700 				kq->kq_state |= KQ_FLUXWAIT;
1701 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1702 				continue;
1703 			}
1704 			kn->kn_status |= KN_INFLUX;
1705 			KQ_UNLOCK(kq);
1706 			if (!(kn->kn_status & KN_DETACHED))
1707 				kn->kn_fop->f_detach(kn);
1708 			knote_drop(kn, td);
1709 			KQ_LOCK(kq);
1710 		}
1711 	}
1712 	if (kq->kq_knhashmask != 0) {
1713 		for (i = 0; i <= kq->kq_knhashmask; i++) {
1714 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1715 				if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1716 					kq->kq_state |= KQ_FLUXWAIT;
1717 					msleep(kq, &kq->kq_lock, PSOCK,
1718 					       "kqclo2", 0);
1719 					continue;
1720 				}
1721 				kn->kn_status |= KN_INFLUX;
1722 				KQ_UNLOCK(kq);
1723 				if (!(kn->kn_status & KN_DETACHED))
1724 					kn->kn_fop->f_detach(kn);
1725 				knote_drop(kn, td);
1726 				KQ_LOCK(kq);
1727 			}
1728 		}
1729 	}
1730 
1731 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1732 		kq->kq_state |= KQ_TASKDRAIN;
1733 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1734 	}
1735 
1736 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1737 		selwakeuppri(&kq->kq_sel, PSOCK);
1738 		if (!SEL_WAITING(&kq->kq_sel))
1739 			kq->kq_state &= ~KQ_SEL;
1740 	}
1741 
1742 	KQ_UNLOCK(kq);
1743 
1744 	/*
1745 	 * We could be called due to the knote_drop() doing fdrop(),
1746 	 * called from kqueue_register().  In this case the global
1747 	 * lock is owned, and filedesc sx is locked before, to not
1748 	 * take the sleepable lock after non-sleepable.
1749 	 */
1750 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1751 		FILEDESC_XLOCK(fdp);
1752 		filedesc_unlock = 1;
1753 	} else
1754 		filedesc_unlock = 0;
1755 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1756 	if (filedesc_unlock)
1757 		FILEDESC_XUNLOCK(fdp);
1758 
1759 	seldrain(&kq->kq_sel);
1760 	knlist_destroy(&kq->kq_sel.si_note);
1761 	mtx_destroy(&kq->kq_lock);
1762 	kq->kq_fdp = NULL;
1763 
1764 	if (kq->kq_knhash != NULL)
1765 		free(kq->kq_knhash, M_KQUEUE);
1766 	if (kq->kq_knlist != NULL)
1767 		free(kq->kq_knlist, M_KQUEUE);
1768 
1769 	funsetown(&kq->kq_sigio);
1770 	free(kq, M_KQUEUE);
1771 	fp->f_data = NULL;
1772 
1773 	return (0);
1774 }
1775 
1776 static void
1777 kqueue_wakeup(struct kqueue *kq)
1778 {
1779 	KQ_OWNED(kq);
1780 
1781 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1782 		kq->kq_state &= ~KQ_SLEEP;
1783 		wakeup(kq);
1784 	}
1785 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1786 		selwakeuppri(&kq->kq_sel, PSOCK);
1787 		if (!SEL_WAITING(&kq->kq_sel))
1788 			kq->kq_state &= ~KQ_SEL;
1789 	}
1790 	if (!knlist_empty(&kq->kq_sel.si_note))
1791 		kqueue_schedtask(kq);
1792 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1793 		pgsigio(&kq->kq_sigio, SIGIO, 0);
1794 	}
1795 }
1796 
1797 /*
1798  * Walk down a list of knotes, activating them if their event has triggered.
1799  *
1800  * There is a possibility to optimize in the case of one kq watching another.
1801  * Instead of scheduling a task to wake it up, you could pass enough state
1802  * down the chain to make up the parent kqueue.  Make this code functional
1803  * first.
1804  */
1805 void
1806 knote(struct knlist *list, long hint, int lockflags)
1807 {
1808 	struct kqueue *kq;
1809 	struct knote *kn;
1810 	int error;
1811 
1812 	if (list == NULL)
1813 		return;
1814 
1815 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1816 
1817 	if ((lockflags & KNF_LISTLOCKED) == 0)
1818 		list->kl_lock(list->kl_lockarg);
1819 
1820 	/*
1821 	 * If we unlock the list lock (and set KN_INFLUX), we can eliminate
1822 	 * the kqueue scheduling, but this will introduce four
1823 	 * lock/unlock's for each knote to test.  If we do, continue to use
1824 	 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is
1825 	 * only safe if you want to remove the current item, which we are
1826 	 * not doing.
1827 	 */
1828 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
1829 		kq = kn->kn_kq;
1830 		if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) {
1831 			KQ_LOCK(kq);
1832 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1833 				KQ_UNLOCK(kq);
1834 			} else if ((lockflags & KNF_NOKQLOCK) != 0) {
1835 				kn->kn_status |= KN_INFLUX;
1836 				KQ_UNLOCK(kq);
1837 				error = kn->kn_fop->f_event(kn, hint);
1838 				KQ_LOCK(kq);
1839 				kn->kn_status &= ~KN_INFLUX;
1840 				if (error)
1841 					KNOTE_ACTIVATE(kn, 1);
1842 				KQ_UNLOCK_FLUX(kq);
1843 			} else {
1844 				kn->kn_status |= KN_HASKQLOCK;
1845 				if (kn->kn_fop->f_event(kn, hint))
1846 					KNOTE_ACTIVATE(kn, 1);
1847 				kn->kn_status &= ~KN_HASKQLOCK;
1848 				KQ_UNLOCK(kq);
1849 			}
1850 		}
1851 		kq = NULL;
1852 	}
1853 	if ((lockflags & KNF_LISTLOCKED) == 0)
1854 		list->kl_unlock(list->kl_lockarg);
1855 }
1856 
1857 /*
1858  * add a knote to a knlist
1859  */
1860 void
1861 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1862 {
1863 	KNL_ASSERT_LOCK(knl, islocked);
1864 	KQ_NOTOWNED(kn->kn_kq);
1865 	KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1866 	    (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1867 	if (!islocked)
1868 		knl->kl_lock(knl->kl_lockarg);
1869 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1870 	if (!islocked)
1871 		knl->kl_unlock(knl->kl_lockarg);
1872 	KQ_LOCK(kn->kn_kq);
1873 	kn->kn_knlist = knl;
1874 	kn->kn_status &= ~KN_DETACHED;
1875 	KQ_UNLOCK(kn->kn_kq);
1876 }
1877 
1878 static void
1879 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1880 {
1881 	KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1882 	KNL_ASSERT_LOCK(knl, knlislocked);
1883 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1884 	if (!kqislocked)
1885 		KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1886     ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1887 	if (!knlislocked)
1888 		knl->kl_lock(knl->kl_lockarg);
1889 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1890 	kn->kn_knlist = NULL;
1891 	if (!knlislocked)
1892 		knl->kl_unlock(knl->kl_lockarg);
1893 	if (!kqislocked)
1894 		KQ_LOCK(kn->kn_kq);
1895 	kn->kn_status |= KN_DETACHED;
1896 	if (!kqislocked)
1897 		KQ_UNLOCK(kn->kn_kq);
1898 }
1899 
1900 /*
1901  * remove knote from the specified knlist
1902  */
1903 void
1904 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1905 {
1906 
1907 	knlist_remove_kq(knl, kn, islocked, 0);
1908 }
1909 
1910 /*
1911  * remove knote from the specified knlist while in f_event handler.
1912  */
1913 void
1914 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1915 {
1916 
1917 	knlist_remove_kq(knl, kn, 1,
1918 	    (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1919 }
1920 
1921 int
1922 knlist_empty(struct knlist *knl)
1923 {
1924 
1925 	KNL_ASSERT_LOCKED(knl);
1926 	return SLIST_EMPTY(&knl->kl_list);
1927 }
1928 
1929 static struct mtx	knlist_lock;
1930 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
1931 	MTX_DEF);
1932 static void knlist_mtx_lock(void *arg);
1933 static void knlist_mtx_unlock(void *arg);
1934 
1935 static void
1936 knlist_mtx_lock(void *arg)
1937 {
1938 
1939 	mtx_lock((struct mtx *)arg);
1940 }
1941 
1942 static void
1943 knlist_mtx_unlock(void *arg)
1944 {
1945 
1946 	mtx_unlock((struct mtx *)arg);
1947 }
1948 
1949 static void
1950 knlist_mtx_assert_locked(void *arg)
1951 {
1952 
1953 	mtx_assert((struct mtx *)arg, MA_OWNED);
1954 }
1955 
1956 static void
1957 knlist_mtx_assert_unlocked(void *arg)
1958 {
1959 
1960 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
1961 }
1962 
1963 static void
1964 knlist_rw_rlock(void *arg)
1965 {
1966 
1967 	rw_rlock((struct rwlock *)arg);
1968 }
1969 
1970 static void
1971 knlist_rw_runlock(void *arg)
1972 {
1973 
1974 	rw_runlock((struct rwlock *)arg);
1975 }
1976 
1977 static void
1978 knlist_rw_assert_locked(void *arg)
1979 {
1980 
1981 	rw_assert((struct rwlock *)arg, RA_LOCKED);
1982 }
1983 
1984 static void
1985 knlist_rw_assert_unlocked(void *arg)
1986 {
1987 
1988 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
1989 }
1990 
1991 void
1992 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
1993     void (*kl_unlock)(void *),
1994     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
1995 {
1996 
1997 	if (lock == NULL)
1998 		knl->kl_lockarg = &knlist_lock;
1999 	else
2000 		knl->kl_lockarg = lock;
2001 
2002 	if (kl_lock == NULL)
2003 		knl->kl_lock = knlist_mtx_lock;
2004 	else
2005 		knl->kl_lock = kl_lock;
2006 	if (kl_unlock == NULL)
2007 		knl->kl_unlock = knlist_mtx_unlock;
2008 	else
2009 		knl->kl_unlock = kl_unlock;
2010 	if (kl_assert_locked == NULL)
2011 		knl->kl_assert_locked = knlist_mtx_assert_locked;
2012 	else
2013 		knl->kl_assert_locked = kl_assert_locked;
2014 	if (kl_assert_unlocked == NULL)
2015 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2016 	else
2017 		knl->kl_assert_unlocked = kl_assert_unlocked;
2018 
2019 	SLIST_INIT(&knl->kl_list);
2020 }
2021 
2022 void
2023 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2024 {
2025 
2026 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2027 }
2028 
2029 void
2030 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2031 {
2032 
2033 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2034 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2035 }
2036 
2037 void
2038 knlist_destroy(struct knlist *knl)
2039 {
2040 
2041 #ifdef INVARIANTS
2042 	/*
2043 	 * if we run across this error, we need to find the offending
2044 	 * driver and have it call knlist_clear or knlist_delete.
2045 	 */
2046 	if (!SLIST_EMPTY(&knl->kl_list))
2047 		printf("WARNING: destroying knlist w/ knotes on it!\n");
2048 #endif
2049 
2050 	knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2051 	SLIST_INIT(&knl->kl_list);
2052 }
2053 
2054 /*
2055  * Even if we are locked, we may need to drop the lock to allow any influx
2056  * knotes time to "settle".
2057  */
2058 void
2059 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2060 {
2061 	struct knote *kn, *kn2;
2062 	struct kqueue *kq;
2063 
2064 	if (islocked)
2065 		KNL_ASSERT_LOCKED(knl);
2066 	else {
2067 		KNL_ASSERT_UNLOCKED(knl);
2068 again:		/* need to reacquire lock since we have dropped it */
2069 		knl->kl_lock(knl->kl_lockarg);
2070 	}
2071 
2072 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2073 		kq = kn->kn_kq;
2074 		KQ_LOCK(kq);
2075 		if ((kn->kn_status & KN_INFLUX)) {
2076 			KQ_UNLOCK(kq);
2077 			continue;
2078 		}
2079 		knlist_remove_kq(knl, kn, 1, 1);
2080 		if (killkn) {
2081 			kn->kn_status |= KN_INFLUX | KN_DETACHED;
2082 			KQ_UNLOCK(kq);
2083 			knote_drop(kn, td);
2084 		} else {
2085 			/* Make sure cleared knotes disappear soon */
2086 			kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2087 			KQ_UNLOCK(kq);
2088 		}
2089 		kq = NULL;
2090 	}
2091 
2092 	if (!SLIST_EMPTY(&knl->kl_list)) {
2093 		/* there are still KN_INFLUX remaining */
2094 		kn = SLIST_FIRST(&knl->kl_list);
2095 		kq = kn->kn_kq;
2096 		KQ_LOCK(kq);
2097 		KASSERT(kn->kn_status & KN_INFLUX,
2098 		    ("knote removed w/o list lock"));
2099 		knl->kl_unlock(knl->kl_lockarg);
2100 		kq->kq_state |= KQ_FLUXWAIT;
2101 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2102 		kq = NULL;
2103 		goto again;
2104 	}
2105 
2106 	if (islocked)
2107 		KNL_ASSERT_LOCKED(knl);
2108 	else {
2109 		knl->kl_unlock(knl->kl_lockarg);
2110 		KNL_ASSERT_UNLOCKED(knl);
2111 	}
2112 }
2113 
2114 /*
2115  * Remove all knotes referencing a specified fd must be called with FILEDESC
2116  * lock.  This prevents a race where a new fd comes along and occupies the
2117  * entry and we attach a knote to the fd.
2118  */
2119 void
2120 knote_fdclose(struct thread *td, int fd)
2121 {
2122 	struct filedesc *fdp = td->td_proc->p_fd;
2123 	struct kqueue *kq;
2124 	struct knote *kn;
2125 	int influx;
2126 
2127 	FILEDESC_XLOCK_ASSERT(fdp);
2128 
2129 	/*
2130 	 * We shouldn't have to worry about new kevents appearing on fd
2131 	 * since filedesc is locked.
2132 	 */
2133 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2134 		KQ_LOCK(kq);
2135 
2136 again:
2137 		influx = 0;
2138 		while (kq->kq_knlistsize > fd &&
2139 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2140 			if (kn->kn_status & KN_INFLUX) {
2141 				/* someone else might be waiting on our knote */
2142 				if (influx)
2143 					wakeup(kq);
2144 				kq->kq_state |= KQ_FLUXWAIT;
2145 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2146 				goto again;
2147 			}
2148 			kn->kn_status |= KN_INFLUX;
2149 			KQ_UNLOCK(kq);
2150 			if (!(kn->kn_status & KN_DETACHED))
2151 				kn->kn_fop->f_detach(kn);
2152 			knote_drop(kn, td);
2153 			influx = 1;
2154 			KQ_LOCK(kq);
2155 		}
2156 		KQ_UNLOCK_FLUX(kq);
2157 	}
2158 }
2159 
2160 static int
2161 knote_attach(struct knote *kn, struct kqueue *kq)
2162 {
2163 	struct klist *list;
2164 
2165 	KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2166 	KQ_OWNED(kq);
2167 
2168 	if (kn->kn_fop->f_isfd) {
2169 		if (kn->kn_id >= kq->kq_knlistsize)
2170 			return ENOMEM;
2171 		list = &kq->kq_knlist[kn->kn_id];
2172 	} else {
2173 		if (kq->kq_knhash == NULL)
2174 			return ENOMEM;
2175 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2176 	}
2177 
2178 	SLIST_INSERT_HEAD(list, kn, kn_link);
2179 
2180 	return 0;
2181 }
2182 
2183 /*
2184  * knote must already have been detached using the f_detach method.
2185  * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2186  * to prevent other removal.
2187  */
2188 static void
2189 knote_drop(struct knote *kn, struct thread *td)
2190 {
2191 	struct kqueue *kq;
2192 	struct klist *list;
2193 
2194 	kq = kn->kn_kq;
2195 
2196 	KQ_NOTOWNED(kq);
2197 	KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2198 	    ("knote_drop called without KN_INFLUX set in kn_status"));
2199 
2200 	KQ_LOCK(kq);
2201 	if (kn->kn_fop->f_isfd)
2202 		list = &kq->kq_knlist[kn->kn_id];
2203 	else
2204 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2205 
2206 	if (!SLIST_EMPTY(list))
2207 		SLIST_REMOVE(list, kn, knote, kn_link);
2208 	if (kn->kn_status & KN_QUEUED)
2209 		knote_dequeue(kn);
2210 	KQ_UNLOCK_FLUX(kq);
2211 
2212 	if (kn->kn_fop->f_isfd) {
2213 		fdrop(kn->kn_fp, td);
2214 		kn->kn_fp = NULL;
2215 	}
2216 	kqueue_fo_release(kn->kn_kevent.filter);
2217 	kn->kn_fop = NULL;
2218 	knote_free(kn);
2219 }
2220 
2221 static void
2222 knote_enqueue(struct knote *kn)
2223 {
2224 	struct kqueue *kq = kn->kn_kq;
2225 
2226 	KQ_OWNED(kn->kn_kq);
2227 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2228 
2229 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2230 	kn->kn_status |= KN_QUEUED;
2231 	kq->kq_count++;
2232 	kqueue_wakeup(kq);
2233 }
2234 
2235 static void
2236 knote_dequeue(struct knote *kn)
2237 {
2238 	struct kqueue *kq = kn->kn_kq;
2239 
2240 	KQ_OWNED(kn->kn_kq);
2241 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2242 
2243 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2244 	kn->kn_status &= ~KN_QUEUED;
2245 	kq->kq_count--;
2246 }
2247 
2248 static void
2249 knote_init(void)
2250 {
2251 
2252 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2253 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2254 }
2255 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2256 
2257 static struct knote *
2258 knote_alloc(int waitok)
2259 {
2260 	return ((struct knote *)uma_zalloc(knote_zone,
2261 	    (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2262 }
2263 
2264 static void
2265 knote_free(struct knote *kn)
2266 {
2267 	if (kn != NULL)
2268 		uma_zfree(knote_zone, kn);
2269 }
2270 
2271 /*
2272  * Register the kev w/ the kq specified by fd.
2273  */
2274 int
2275 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2276 {
2277 	struct kqueue *kq;
2278 	struct file *fp;
2279 	cap_rights_t rights;
2280 	int error;
2281 
2282 	error = fget(td, fd, cap_rights_init(&rights, CAP_POST_EVENT), &fp);
2283 	if (error != 0)
2284 		return (error);
2285 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2286 		goto noacquire;
2287 
2288 	error = kqueue_register(kq, kev, td, waitok);
2289 
2290 	kqueue_release(kq, 0);
2291 
2292 noacquire:
2293 	fdrop(fp, td);
2294 
2295 	return error;
2296 }
2297