xref: /freebsd/sys/kern/kern_event.c (revision 2c30bc1fcfd4b11dc9dc663ec5a5f5c466d00615)
1 /*-
2  * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
3  * Copyright 2004 John-Mark Gurney <jmg@FreeBSD.org>
4  * Copyright (c) 2009 Apple, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include "opt_ktrace.h"
33 #include "opt_kqueue.h"
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/capsicum.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/rwlock.h>
42 #include <sys/proc.h>
43 #include <sys/malloc.h>
44 #include <sys/unistd.h>
45 #include <sys/file.h>
46 #include <sys/filedesc.h>
47 #include <sys/filio.h>
48 #include <sys/fcntl.h>
49 #include <sys/kthread.h>
50 #include <sys/selinfo.h>
51 #include <sys/stdatomic.h>
52 #include <sys/queue.h>
53 #include <sys/event.h>
54 #include <sys/eventvar.h>
55 #include <sys/poll.h>
56 #include <sys/protosw.h>
57 #include <sys/resourcevar.h>
58 #include <sys/sigio.h>
59 #include <sys/signalvar.h>
60 #include <sys/socket.h>
61 #include <sys/socketvar.h>
62 #include <sys/stat.h>
63 #include <sys/sysctl.h>
64 #include <sys/sysproto.h>
65 #include <sys/syscallsubr.h>
66 #include <sys/taskqueue.h>
67 #include <sys/uio.h>
68 #include <sys/user.h>
69 #ifdef KTRACE
70 #include <sys/ktrace.h>
71 #endif
72 
73 #include <vm/uma.h>
74 
75 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
76 
77 /*
78  * This lock is used if multiple kq locks are required.  This possibly
79  * should be made into a per proc lock.
80  */
81 static struct mtx	kq_global;
82 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
83 #define KQ_GLOBAL_LOCK(lck, haslck)	do {	\
84 	if (!haslck)				\
85 		mtx_lock(lck);			\
86 	haslck = 1;				\
87 } while (0)
88 #define KQ_GLOBAL_UNLOCK(lck, haslck)	do {	\
89 	if (haslck)				\
90 		mtx_unlock(lck);			\
91 	haslck = 0;				\
92 } while (0)
93 
94 TASKQUEUE_DEFINE_THREAD(kqueue);
95 
96 static int	kevent_copyout(void *arg, struct kevent *kevp, int count);
97 static int	kevent_copyin(void *arg, struct kevent *kevp, int count);
98 static int	kqueue_register(struct kqueue *kq, struct kevent *kev,
99 		    struct thread *td, int waitok);
100 static int	kqueue_acquire(struct file *fp, struct kqueue **kqp);
101 static void	kqueue_release(struct kqueue *kq, int locked);
102 static int	kqueue_expand(struct kqueue *kq, struct filterops *fops,
103 		    uintptr_t ident, int waitok);
104 static void	kqueue_task(void *arg, int pending);
105 static int	kqueue_scan(struct kqueue *kq, int maxevents,
106 		    struct kevent_copyops *k_ops,
107 		    const struct timespec *timeout,
108 		    struct kevent *keva, struct thread *td);
109 static void 	kqueue_wakeup(struct kqueue *kq);
110 static struct filterops *kqueue_fo_find(int filt);
111 static void	kqueue_fo_release(int filt);
112 
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 static fo_fill_kinfo_t	kqueue_fill_kinfo;
119 
120 static struct fileops kqueueops = {
121 	.fo_read = invfo_rdwr,
122 	.fo_write = invfo_rdwr,
123 	.fo_truncate = invfo_truncate,
124 	.fo_ioctl = kqueue_ioctl,
125 	.fo_poll = kqueue_poll,
126 	.fo_kqfilter = kqueue_kqfilter,
127 	.fo_stat = kqueue_stat,
128 	.fo_close = kqueue_close,
129 	.fo_chmod = invfo_chmod,
130 	.fo_chown = invfo_chown,
131 	.fo_sendfile = invfo_sendfile,
132 	.fo_fill_kinfo = kqueue_fill_kinfo,
133 };
134 
135 static int 	knote_attach(struct knote *kn, struct kqueue *kq);
136 static void 	knote_drop(struct knote *kn, struct thread *td);
137 static void 	knote_enqueue(struct knote *kn);
138 static void 	knote_dequeue(struct knote *kn);
139 static void 	knote_init(void);
140 static struct 	knote *knote_alloc(int waitok);
141 static void 	knote_free(struct knote *kn);
142 
143 static void	filt_kqdetach(struct knote *kn);
144 static int	filt_kqueue(struct knote *kn, long hint);
145 static int	filt_procattach(struct knote *kn);
146 static void	filt_procdetach(struct knote *kn);
147 static int	filt_proc(struct knote *kn, long hint);
148 static int	filt_fileattach(struct knote *kn);
149 static void	filt_timerexpire(void *knx);
150 static int	filt_timerattach(struct knote *kn);
151 static void	filt_timerdetach(struct knote *kn);
152 static int	filt_timer(struct knote *kn, long hint);
153 static int	filt_userattach(struct knote *kn);
154 static void	filt_userdetach(struct knote *kn);
155 static int	filt_user(struct knote *kn, long hint);
156 static void	filt_usertouch(struct knote *kn, struct kevent *kev,
157 		    u_long type);
158 
159 static struct filterops file_filtops = {
160 	.f_isfd = 1,
161 	.f_attach = filt_fileattach,
162 };
163 static struct filterops kqread_filtops = {
164 	.f_isfd = 1,
165 	.f_detach = filt_kqdetach,
166 	.f_event = filt_kqueue,
167 };
168 /* XXX - move to kern_proc.c?  */
169 static struct filterops proc_filtops = {
170 	.f_isfd = 0,
171 	.f_attach = filt_procattach,
172 	.f_detach = filt_procdetach,
173 	.f_event = filt_proc,
174 };
175 static struct filterops timer_filtops = {
176 	.f_isfd = 0,
177 	.f_attach = filt_timerattach,
178 	.f_detach = filt_timerdetach,
179 	.f_event = filt_timer,
180 };
181 static struct filterops user_filtops = {
182 	.f_attach = filt_userattach,
183 	.f_detach = filt_userdetach,
184 	.f_event = filt_user,
185 	.f_touch = filt_usertouch,
186 };
187 
188 static uma_zone_t	knote_zone;
189 static atomic_uint	kq_ncallouts = ATOMIC_VAR_INIT(0);
190 static unsigned int 	kq_calloutmax = 4 * 1024;
191 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
192     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
193 
194 /* XXX - ensure not KN_INFLUX?? */
195 #define KNOTE_ACTIVATE(kn, islock) do { 				\
196 	if ((islock))							\
197 		mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED);		\
198 	else								\
199 		KQ_LOCK((kn)->kn_kq);					\
200 	(kn)->kn_status |= KN_ACTIVE;					\
201 	if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)		\
202 		knote_enqueue((kn));					\
203 	if (!(islock))							\
204 		KQ_UNLOCK((kn)->kn_kq);					\
205 } while(0)
206 #define KQ_LOCK(kq) do {						\
207 	mtx_lock(&(kq)->kq_lock);					\
208 } while (0)
209 #define KQ_FLUX_WAKEUP(kq) do {						\
210 	if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) {		\
211 		(kq)->kq_state &= ~KQ_FLUXWAIT;				\
212 		wakeup((kq));						\
213 	}								\
214 } while (0)
215 #define KQ_UNLOCK_FLUX(kq) do {						\
216 	KQ_FLUX_WAKEUP(kq);						\
217 	mtx_unlock(&(kq)->kq_lock);					\
218 } while (0)
219 #define KQ_UNLOCK(kq) do {						\
220 	mtx_unlock(&(kq)->kq_lock);					\
221 } while (0)
222 #define KQ_OWNED(kq) do {						\
223 	mtx_assert(&(kq)->kq_lock, MA_OWNED);				\
224 } while (0)
225 #define KQ_NOTOWNED(kq) do {						\
226 	mtx_assert(&(kq)->kq_lock, MA_NOTOWNED);			\
227 } while (0)
228 #define KN_LIST_LOCK(kn) do {						\
229 	if (kn->kn_knlist != NULL)					\
230 		kn->kn_knlist->kl_lock(kn->kn_knlist->kl_lockarg);	\
231 } while (0)
232 #define KN_LIST_UNLOCK(kn) do {						\
233 	if (kn->kn_knlist != NULL) 					\
234 		kn->kn_knlist->kl_unlock(kn->kn_knlist->kl_lockarg);	\
235 } while (0)
236 #define	KNL_ASSERT_LOCK(knl, islocked) do {				\
237 	if (islocked)							\
238 		KNL_ASSERT_LOCKED(knl);				\
239 	else								\
240 		KNL_ASSERT_UNLOCKED(knl);				\
241 } while (0)
242 #ifdef INVARIANTS
243 #define	KNL_ASSERT_LOCKED(knl) do {					\
244 	knl->kl_assert_locked((knl)->kl_lockarg);			\
245 } while (0)
246 #define	KNL_ASSERT_UNLOCKED(knl) do {					\
247 	knl->kl_assert_unlocked((knl)->kl_lockarg);			\
248 } while (0)
249 #else /* !INVARIANTS */
250 #define	KNL_ASSERT_LOCKED(knl) do {} while(0)
251 #define	KNL_ASSERT_UNLOCKED(knl) do {} while (0)
252 #endif /* INVARIANTS */
253 
254 #ifndef	KN_HASHSIZE
255 #define	KN_HASHSIZE		64		/* XXX should be tunable */
256 #endif
257 
258 #define KN_HASH(val, mask)	(((val) ^ (val >> 8)) & (mask))
259 
260 static int
261 filt_nullattach(struct knote *kn)
262 {
263 
264 	return (ENXIO);
265 };
266 
267 struct filterops null_filtops = {
268 	.f_isfd = 0,
269 	.f_attach = filt_nullattach,
270 };
271 
272 /* XXX - make SYSINIT to add these, and move into respective modules. */
273 extern struct filterops sig_filtops;
274 extern struct filterops fs_filtops;
275 
276 /*
277  * Table for for all system-defined filters.
278  */
279 static struct mtx	filterops_lock;
280 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
281 	MTX_DEF);
282 static struct {
283 	struct filterops *for_fop;
284 	int for_nolock;
285 	int for_refcnt;
286 } sysfilt_ops[EVFILT_SYSCOUNT] = {
287 	{ &file_filtops, 1 },			/* EVFILT_READ */
288 	{ &file_filtops, 1 },			/* EVFILT_WRITE */
289 	{ &null_filtops },			/* EVFILT_AIO */
290 	{ &file_filtops, 1 },			/* EVFILT_VNODE */
291 	{ &proc_filtops, 1 },			/* EVFILT_PROC */
292 	{ &sig_filtops, 1 },			/* EVFILT_SIGNAL */
293 	{ &timer_filtops, 1 },			/* EVFILT_TIMER */
294 	{ &file_filtops, 1 },			/* EVFILT_PROCDESC */
295 	{ &fs_filtops, 1 },			/* EVFILT_FS */
296 	{ &null_filtops },			/* EVFILT_LIO */
297 	{ &user_filtops, 1 },			/* EVFILT_USER */
298 	{ &null_filtops },			/* EVFILT_SENDFILE */
299 };
300 
301 /*
302  * Simple redirection for all cdevsw style objects to call their fo_kqfilter
303  * method.
304  */
305 static int
306 filt_fileattach(struct knote *kn)
307 {
308 
309 	return (fo_kqfilter(kn->kn_fp, kn));
310 }
311 
312 /*ARGSUSED*/
313 static int
314 kqueue_kqfilter(struct file *fp, struct knote *kn)
315 {
316 	struct kqueue *kq = kn->kn_fp->f_data;
317 
318 	if (kn->kn_filter != EVFILT_READ)
319 		return (EINVAL);
320 
321 	kn->kn_status |= KN_KQUEUE;
322 	kn->kn_fop = &kqread_filtops;
323 	knlist_add(&kq->kq_sel.si_note, kn, 0);
324 
325 	return (0);
326 }
327 
328 static void
329 filt_kqdetach(struct knote *kn)
330 {
331 	struct kqueue *kq = kn->kn_fp->f_data;
332 
333 	knlist_remove(&kq->kq_sel.si_note, kn, 0);
334 }
335 
336 /*ARGSUSED*/
337 static int
338 filt_kqueue(struct knote *kn, long hint)
339 {
340 	struct kqueue *kq = kn->kn_fp->f_data;
341 
342 	kn->kn_data = kq->kq_count;
343 	return (kn->kn_data > 0);
344 }
345 
346 /* XXX - move to kern_proc.c?  */
347 static int
348 filt_procattach(struct knote *kn)
349 {
350 	struct proc *p;
351 	int immediate;
352 	int error;
353 
354 	immediate = 0;
355 	p = pfind(kn->kn_id);
356 	if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
357 		p = zpfind(kn->kn_id);
358 		immediate = 1;
359 	} else if (p != NULL && (p->p_flag & P_WEXIT)) {
360 		immediate = 1;
361 	}
362 
363 	if (p == NULL)
364 		return (ESRCH);
365 	if ((error = p_cansee(curthread, p))) {
366 		PROC_UNLOCK(p);
367 		return (error);
368 	}
369 
370 	kn->kn_ptr.p_proc = p;
371 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
372 
373 	/*
374 	 * internal flag indicating registration done by kernel
375 	 */
376 	if (kn->kn_flags & EV_FLAG1) {
377 		kn->kn_data = kn->kn_sdata;		/* ppid */
378 		kn->kn_fflags = NOTE_CHILD;
379 		kn->kn_flags &= ~EV_FLAG1;
380 	}
381 
382 	if (immediate == 0)
383 		knlist_add(&p->p_klist, kn, 1);
384 
385 	/*
386 	 * Immediately activate any exit notes if the target process is a
387 	 * zombie.  This is necessary to handle the case where the target
388 	 * process, e.g. a child, dies before the kevent is registered.
389 	 */
390 	if (immediate && filt_proc(kn, NOTE_EXIT))
391 		KNOTE_ACTIVATE(kn, 0);
392 
393 	PROC_UNLOCK(p);
394 
395 	return (0);
396 }
397 
398 /*
399  * The knote may be attached to a different process, which may exit,
400  * leaving nothing for the knote to be attached to.  So when the process
401  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
402  * it will be deleted when read out.  However, as part of the knote deletion,
403  * this routine is called, so a check is needed to avoid actually performing
404  * a detach, because the original process does not exist any more.
405  */
406 /* XXX - move to kern_proc.c?  */
407 static void
408 filt_procdetach(struct knote *kn)
409 {
410 	struct proc *p;
411 
412 	p = kn->kn_ptr.p_proc;
413 	knlist_remove(&p->p_klist, kn, 0);
414 	kn->kn_ptr.p_proc = NULL;
415 }
416 
417 /* XXX - move to kern_proc.c?  */
418 static int
419 filt_proc(struct knote *kn, long hint)
420 {
421 	struct proc *p;
422 	u_int event;
423 
424 	p = kn->kn_ptr.p_proc;
425 	/* Mask off extra data. */
426 	event = (u_int)hint & NOTE_PCTRLMASK;
427 
428 	/* If the user is interested in this event, record it. */
429 	if (kn->kn_sfflags & event)
430 		kn->kn_fflags |= event;
431 
432 	/* Process is gone, so flag the event as finished. */
433 	if (event == NOTE_EXIT) {
434 		if (!(kn->kn_status & KN_DETACHED))
435 			knlist_remove_inevent(&p->p_klist, kn);
436 		kn->kn_flags |= EV_EOF | EV_ONESHOT;
437 		kn->kn_ptr.p_proc = NULL;
438 		if (kn->kn_fflags & NOTE_EXIT)
439 			kn->kn_data = p->p_xstat;
440 		if (kn->kn_fflags == 0)
441 			kn->kn_flags |= EV_DROP;
442 		return (1);
443 	}
444 
445 	return (kn->kn_fflags != 0);
446 }
447 
448 /*
449  * Called when the process forked. It mostly does the same as the
450  * knote(), activating all knotes registered to be activated when the
451  * process forked. Additionally, for each knote attached to the
452  * parent, check whether user wants to track the new process. If so
453  * attach a new knote to it, and immediately report an event with the
454  * child's pid.
455  */
456 void
457 knote_fork(struct knlist *list, int pid)
458 {
459 	struct kqueue *kq;
460 	struct knote *kn;
461 	struct kevent kev;
462 	int error;
463 
464 	if (list == NULL)
465 		return;
466 	list->kl_lock(list->kl_lockarg);
467 
468 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
469 		/*
470 		 * XXX - Why do we skip the kn if it is _INFLUX?  Does this
471 		 * mean we will not properly wake up some notes?
472 		 */
473 		if ((kn->kn_status & KN_INFLUX) == KN_INFLUX)
474 			continue;
475 		kq = kn->kn_kq;
476 		KQ_LOCK(kq);
477 		if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
478 			KQ_UNLOCK(kq);
479 			continue;
480 		}
481 
482 		/*
483 		 * The same as knote(), activate the event.
484 		 */
485 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
486 			kn->kn_status |= KN_HASKQLOCK;
487 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
488 				KNOTE_ACTIVATE(kn, 1);
489 			kn->kn_status &= ~KN_HASKQLOCK;
490 			KQ_UNLOCK(kq);
491 			continue;
492 		}
493 
494 		/*
495 		 * The NOTE_TRACK case. In addition to the activation
496 		 * of the event, we need to register new event to
497 		 * track the child. Drop the locks in preparation for
498 		 * the call to kqueue_register().
499 		 */
500 		kn->kn_status |= KN_INFLUX;
501 		KQ_UNLOCK(kq);
502 		list->kl_unlock(list->kl_lockarg);
503 
504 		/*
505 		 * Activate existing knote and register a knote with
506 		 * new process.
507 		 */
508 		kev.ident = pid;
509 		kev.filter = kn->kn_filter;
510 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
511 		kev.fflags = kn->kn_sfflags;
512 		kev.data = kn->kn_id;		/* parent */
513 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
514 		error = kqueue_register(kq, &kev, NULL, 0);
515 		if (error)
516 			kn->kn_fflags |= NOTE_TRACKERR;
517 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
518 			KNOTE_ACTIVATE(kn, 0);
519 		KQ_LOCK(kq);
520 		kn->kn_status &= ~KN_INFLUX;
521 		KQ_UNLOCK_FLUX(kq);
522 		list->kl_lock(list->kl_lockarg);
523 	}
524 	list->kl_unlock(list->kl_lockarg);
525 }
526 
527 /*
528  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
529  * interval timer support code.
530  */
531 
532 #define NOTE_TIMER_PRECMASK	(NOTE_SECONDS|NOTE_MSECONDS|NOTE_USECONDS| \
533 				NOTE_NSECONDS)
534 
535 static __inline sbintime_t
536 timer2sbintime(intptr_t data, int flags)
537 {
538 	sbintime_t modifier;
539 
540 	switch (flags & NOTE_TIMER_PRECMASK) {
541 	case NOTE_SECONDS:
542 		modifier = SBT_1S;
543 		break;
544 	case NOTE_MSECONDS: /* FALLTHROUGH */
545 	case 0:
546 		modifier = SBT_1MS;
547 		break;
548 	case NOTE_USECONDS:
549 		modifier = SBT_1US;
550 		break;
551 	case NOTE_NSECONDS:
552 		modifier = SBT_1NS;
553 		break;
554 	default:
555 		return (-1);
556 	}
557 
558 #ifdef __LP64__
559 	if (data > SBT_MAX / modifier)
560 		return (SBT_MAX);
561 #endif
562 	return (modifier * data);
563 }
564 
565 static void
566 filt_timerexpire(void *knx)
567 {
568 	struct callout *calloutp;
569 	struct knote *kn;
570 
571 	kn = knx;
572 	kn->kn_data++;
573 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
574 
575 	if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
576 		calloutp = (struct callout *)kn->kn_hook;
577 		*kn->kn_ptr.p_nexttime += timer2sbintime(kn->kn_sdata,
578 		    kn->kn_sfflags);
579 		callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
580 		    filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
581 	}
582 }
583 
584 /*
585  * data contains amount of time to sleep
586  */
587 static int
588 filt_timerattach(struct knote *kn)
589 {
590 	struct callout *calloutp;
591 	sbintime_t to;
592 	unsigned int ncallouts;
593 
594 	if ((intptr_t)kn->kn_sdata < 0)
595 		return (EINVAL);
596 	if ((intptr_t)kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
597 		kn->kn_sdata = 1;
598 	/* Only precision unit are supported in flags so far */
599 	if (kn->kn_sfflags & ~NOTE_TIMER_PRECMASK)
600 		return (EINVAL);
601 
602 	to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
603 	if (to < 0)
604 		return (EINVAL);
605 
606 	ncallouts = atomic_load_explicit(&kq_ncallouts, memory_order_relaxed);
607 	do {
608 		if (ncallouts >= kq_calloutmax)
609 			return (ENOMEM);
610 	} while (!atomic_compare_exchange_weak_explicit(&kq_ncallouts,
611 	    &ncallouts, ncallouts + 1, memory_order_relaxed,
612 	    memory_order_relaxed));
613 
614 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
615 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
616 	kn->kn_ptr.p_nexttime = malloc(sizeof(sbintime_t), M_KQUEUE, M_WAITOK);
617 	calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
618 	callout_init(calloutp, CALLOUT_MPSAFE);
619 	kn->kn_hook = calloutp;
620 	*kn->kn_ptr.p_nexttime = to + sbinuptime();
621 	callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
622 	    filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
623 
624 	return (0);
625 }
626 
627 static void
628 filt_timerdetach(struct knote *kn)
629 {
630 	struct callout *calloutp;
631 	unsigned int old;
632 
633 	calloutp = (struct callout *)kn->kn_hook;
634 	callout_drain(calloutp);
635 	free(calloutp, M_KQUEUE);
636 	free(kn->kn_ptr.p_nexttime, M_KQUEUE);
637 	old = atomic_fetch_sub_explicit(&kq_ncallouts, 1, memory_order_relaxed);
638 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
639 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
640 }
641 
642 static int
643 filt_timer(struct knote *kn, long hint)
644 {
645 
646 	return (kn->kn_data != 0);
647 }
648 
649 static int
650 filt_userattach(struct knote *kn)
651 {
652 
653 	/*
654 	 * EVFILT_USER knotes are not attached to anything in the kernel.
655 	 */
656 	kn->kn_hook = NULL;
657 	if (kn->kn_fflags & NOTE_TRIGGER)
658 		kn->kn_hookid = 1;
659 	else
660 		kn->kn_hookid = 0;
661 	return (0);
662 }
663 
664 static void
665 filt_userdetach(__unused struct knote *kn)
666 {
667 
668 	/*
669 	 * EVFILT_USER knotes are not attached to anything in the kernel.
670 	 */
671 }
672 
673 static int
674 filt_user(struct knote *kn, __unused long hint)
675 {
676 
677 	return (kn->kn_hookid);
678 }
679 
680 static void
681 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
682 {
683 	u_int ffctrl;
684 
685 	switch (type) {
686 	case EVENT_REGISTER:
687 		if (kev->fflags & NOTE_TRIGGER)
688 			kn->kn_hookid = 1;
689 
690 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
691 		kev->fflags &= NOTE_FFLAGSMASK;
692 		switch (ffctrl) {
693 		case NOTE_FFNOP:
694 			break;
695 
696 		case NOTE_FFAND:
697 			kn->kn_sfflags &= kev->fflags;
698 			break;
699 
700 		case NOTE_FFOR:
701 			kn->kn_sfflags |= kev->fflags;
702 			break;
703 
704 		case NOTE_FFCOPY:
705 			kn->kn_sfflags = kev->fflags;
706 			break;
707 
708 		default:
709 			/* XXX Return error? */
710 			break;
711 		}
712 		kn->kn_sdata = kev->data;
713 		if (kev->flags & EV_CLEAR) {
714 			kn->kn_hookid = 0;
715 			kn->kn_data = 0;
716 			kn->kn_fflags = 0;
717 		}
718 		break;
719 
720         case EVENT_PROCESS:
721 		*kev = kn->kn_kevent;
722 		kev->fflags = kn->kn_sfflags;
723 		kev->data = kn->kn_sdata;
724 		if (kn->kn_flags & EV_CLEAR) {
725 			kn->kn_hookid = 0;
726 			kn->kn_data = 0;
727 			kn->kn_fflags = 0;
728 		}
729 		break;
730 
731 	default:
732 		panic("filt_usertouch() - invalid type (%ld)", type);
733 		break;
734 	}
735 }
736 
737 int
738 sys_kqueue(struct thread *td, struct kqueue_args *uap)
739 {
740 	struct filedesc *fdp;
741 	struct kqueue *kq;
742 	struct file *fp;
743 	struct proc *p;
744 	struct ucred *cred;
745 	int fd, error;
746 
747 	p = td->td_proc;
748 	cred = td->td_ucred;
749 	crhold(cred);
750 	PROC_LOCK(p);
751 	if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td->td_proc,
752 	    RLIMIT_KQUEUES))) {
753 		PROC_UNLOCK(p);
754 		crfree(cred);
755 		return (ENOMEM);
756 	}
757 	PROC_UNLOCK(p);
758 
759 	fdp = p->p_fd;
760 	error = falloc(td, &fp, &fd, 0);
761 	if (error)
762 		goto done2;
763 
764 	/* An extra reference on `fp' has been held for us by falloc(). */
765 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
766 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
767 	TAILQ_INIT(&kq->kq_head);
768 	kq->kq_fdp = fdp;
769 	kq->kq_cred = cred;
770 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
771 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
772 
773 	FILEDESC_XLOCK(fdp);
774 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
775 	FILEDESC_XUNLOCK(fdp);
776 
777 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
778 	fdrop(fp, td);
779 
780 	td->td_retval[0] = fd;
781 done2:
782 	if (error != 0) {
783 		chgkqcnt(cred->cr_ruidinfo, -1, 0);
784 		crfree(cred);
785 	}
786 	return (error);
787 }
788 
789 #ifndef _SYS_SYSPROTO_H_
790 struct kevent_args {
791 	int	fd;
792 	const struct kevent *changelist;
793 	int	nchanges;
794 	struct	kevent *eventlist;
795 	int	nevents;
796 	const struct timespec *timeout;
797 };
798 #endif
799 int
800 sys_kevent(struct thread *td, struct kevent_args *uap)
801 {
802 	struct timespec ts, *tsp;
803 	struct kevent_copyops k_ops = { uap,
804 					kevent_copyout,
805 					kevent_copyin};
806 	int error;
807 #ifdef KTRACE
808 	struct uio ktruio;
809 	struct iovec ktriov;
810 	struct uio *ktruioin = NULL;
811 	struct uio *ktruioout = NULL;
812 #endif
813 
814 	if (uap->timeout != NULL) {
815 		error = copyin(uap->timeout, &ts, sizeof(ts));
816 		if (error)
817 			return (error);
818 		tsp = &ts;
819 	} else
820 		tsp = NULL;
821 
822 #ifdef KTRACE
823 	if (KTRPOINT(td, KTR_GENIO)) {
824 		ktriov.iov_base = uap->changelist;
825 		ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
826 		ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
827 		    .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
828 		    .uio_td = td };
829 		ktruioin = cloneuio(&ktruio);
830 		ktriov.iov_base = uap->eventlist;
831 		ktriov.iov_len = uap->nevents * sizeof(struct kevent);
832 		ktruioout = cloneuio(&ktruio);
833 	}
834 #endif
835 
836 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
837 	    &k_ops, tsp);
838 
839 #ifdef KTRACE
840 	if (ktruioin != NULL) {
841 		ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
842 		ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
843 		ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
844 		ktrgenio(uap->fd, UIO_READ, ktruioout, error);
845 	}
846 #endif
847 
848 	return (error);
849 }
850 
851 /*
852  * Copy 'count' items into the destination list pointed to by uap->eventlist.
853  */
854 static int
855 kevent_copyout(void *arg, struct kevent *kevp, int count)
856 {
857 	struct kevent_args *uap;
858 	int error;
859 
860 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
861 	uap = (struct kevent_args *)arg;
862 
863 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
864 	if (error == 0)
865 		uap->eventlist += count;
866 	return (error);
867 }
868 
869 /*
870  * Copy 'count' items from the list pointed to by uap->changelist.
871  */
872 static int
873 kevent_copyin(void *arg, struct kevent *kevp, int count)
874 {
875 	struct kevent_args *uap;
876 	int error;
877 
878 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
879 	uap = (struct kevent_args *)arg;
880 
881 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
882 	if (error == 0)
883 		uap->changelist += count;
884 	return (error);
885 }
886 
887 int
888 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
889     struct kevent_copyops *k_ops, const struct timespec *timeout)
890 {
891 	struct kevent keva[KQ_NEVENTS];
892 	struct kevent *kevp, *changes;
893 	struct kqueue *kq;
894 	struct file *fp;
895 	cap_rights_t rights;
896 	int i, n, nerrors, error;
897 
898 	cap_rights_init(&rights);
899 	if (nchanges > 0)
900 		cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
901 	if (nevents > 0)
902 		cap_rights_set(&rights, CAP_KQUEUE_EVENT);
903 	error = fget(td, fd, &rights, &fp);
904 	if (error != 0)
905 		return (error);
906 
907 	error = kqueue_acquire(fp, &kq);
908 	if (error != 0)
909 		goto done_norel;
910 
911 	nerrors = 0;
912 
913 	while (nchanges > 0) {
914 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
915 		error = k_ops->k_copyin(k_ops->arg, keva, n);
916 		if (error)
917 			goto done;
918 		changes = keva;
919 		for (i = 0; i < n; i++) {
920 			kevp = &changes[i];
921 			if (!kevp->filter)
922 				continue;
923 			kevp->flags &= ~EV_SYSFLAGS;
924 			error = kqueue_register(kq, kevp, td, 1);
925 			if (error || (kevp->flags & EV_RECEIPT)) {
926 				if (nevents != 0) {
927 					kevp->flags = EV_ERROR;
928 					kevp->data = error;
929 					(void) k_ops->k_copyout(k_ops->arg,
930 					    kevp, 1);
931 					nevents--;
932 					nerrors++;
933 				} else {
934 					goto done;
935 				}
936 			}
937 		}
938 		nchanges -= n;
939 	}
940 	if (nerrors) {
941 		td->td_retval[0] = nerrors;
942 		error = 0;
943 		goto done;
944 	}
945 
946 	error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
947 done:
948 	kqueue_release(kq, 0);
949 done_norel:
950 	fdrop(fp, td);
951 	return (error);
952 }
953 
954 int
955 kqueue_add_filteropts(int filt, struct filterops *filtops)
956 {
957 	int error;
958 
959 	error = 0;
960 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
961 		printf(
962 "trying to add a filterop that is out of range: %d is beyond %d\n",
963 		    ~filt, EVFILT_SYSCOUNT);
964 		return EINVAL;
965 	}
966 	mtx_lock(&filterops_lock);
967 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
968 	    sysfilt_ops[~filt].for_fop != NULL)
969 		error = EEXIST;
970 	else {
971 		sysfilt_ops[~filt].for_fop = filtops;
972 		sysfilt_ops[~filt].for_refcnt = 0;
973 	}
974 	mtx_unlock(&filterops_lock);
975 
976 	return (error);
977 }
978 
979 int
980 kqueue_del_filteropts(int filt)
981 {
982 	int error;
983 
984 	error = 0;
985 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
986 		return EINVAL;
987 
988 	mtx_lock(&filterops_lock);
989 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
990 	    sysfilt_ops[~filt].for_fop == NULL)
991 		error = EINVAL;
992 	else if (sysfilt_ops[~filt].for_refcnt != 0)
993 		error = EBUSY;
994 	else {
995 		sysfilt_ops[~filt].for_fop = &null_filtops;
996 		sysfilt_ops[~filt].for_refcnt = 0;
997 	}
998 	mtx_unlock(&filterops_lock);
999 
1000 	return error;
1001 }
1002 
1003 static struct filterops *
1004 kqueue_fo_find(int filt)
1005 {
1006 
1007 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1008 		return NULL;
1009 
1010 	if (sysfilt_ops[~filt].for_nolock)
1011 		return sysfilt_ops[~filt].for_fop;
1012 
1013 	mtx_lock(&filterops_lock);
1014 	sysfilt_ops[~filt].for_refcnt++;
1015 	if (sysfilt_ops[~filt].for_fop == NULL)
1016 		sysfilt_ops[~filt].for_fop = &null_filtops;
1017 	mtx_unlock(&filterops_lock);
1018 
1019 	return sysfilt_ops[~filt].for_fop;
1020 }
1021 
1022 static void
1023 kqueue_fo_release(int filt)
1024 {
1025 
1026 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1027 		return;
1028 
1029 	if (sysfilt_ops[~filt].for_nolock)
1030 		return;
1031 
1032 	mtx_lock(&filterops_lock);
1033 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1034 	    ("filter object refcount not valid on release"));
1035 	sysfilt_ops[~filt].for_refcnt--;
1036 	mtx_unlock(&filterops_lock);
1037 }
1038 
1039 /*
1040  * A ref to kq (obtained via kqueue_acquire) must be held.  waitok will
1041  * influence if memory allocation should wait.  Make sure it is 0 if you
1042  * hold any mutexes.
1043  */
1044 static int
1045 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1046 {
1047 	struct filterops *fops;
1048 	struct file *fp;
1049 	struct knote *kn, *tkn;
1050 	cap_rights_t rights;
1051 	int error, filt, event;
1052 	int haskqglobal, filedesc_unlock;
1053 
1054 	fp = NULL;
1055 	kn = NULL;
1056 	error = 0;
1057 	haskqglobal = 0;
1058 	filedesc_unlock = 0;
1059 
1060 	filt = kev->filter;
1061 	fops = kqueue_fo_find(filt);
1062 	if (fops == NULL)
1063 		return EINVAL;
1064 
1065 	if (kev->flags & EV_ADD)
1066 		tkn = knote_alloc(waitok);	/* prevent waiting with locks */
1067 	else
1068 		tkn = NULL;
1069 
1070 findkn:
1071 	if (fops->f_isfd) {
1072 		KASSERT(td != NULL, ("td is NULL"));
1073 		error = fget(td, kev->ident,
1074 		    cap_rights_init(&rights, CAP_EVENT), &fp);
1075 		if (error)
1076 			goto done;
1077 
1078 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1079 		    kev->ident, 0) != 0) {
1080 			/* try again */
1081 			fdrop(fp, td);
1082 			fp = NULL;
1083 			error = kqueue_expand(kq, fops, kev->ident, waitok);
1084 			if (error)
1085 				goto done;
1086 			goto findkn;
1087 		}
1088 
1089 		if (fp->f_type == DTYPE_KQUEUE) {
1090 			/*
1091 			 * if we add some inteligence about what we are doing,
1092 			 * we should be able to support events on ourselves.
1093 			 * We need to know when we are doing this to prevent
1094 			 * getting both the knlist lock and the kq lock since
1095 			 * they are the same thing.
1096 			 */
1097 			if (fp->f_data == kq) {
1098 				error = EINVAL;
1099 				goto done;
1100 			}
1101 
1102 			/*
1103 			 * Pre-lock the filedesc before the global
1104 			 * lock mutex, see the comment in
1105 			 * kqueue_close().
1106 			 */
1107 			FILEDESC_XLOCK(td->td_proc->p_fd);
1108 			filedesc_unlock = 1;
1109 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1110 		}
1111 
1112 		KQ_LOCK(kq);
1113 		if (kev->ident < kq->kq_knlistsize) {
1114 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1115 				if (kev->filter == kn->kn_filter)
1116 					break;
1117 		}
1118 	} else {
1119 		if ((kev->flags & EV_ADD) == EV_ADD)
1120 			kqueue_expand(kq, fops, kev->ident, waitok);
1121 
1122 		KQ_LOCK(kq);
1123 		if (kq->kq_knhashmask != 0) {
1124 			struct klist *list;
1125 
1126 			list = &kq->kq_knhash[
1127 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1128 			SLIST_FOREACH(kn, list, kn_link)
1129 				if (kev->ident == kn->kn_id &&
1130 				    kev->filter == kn->kn_filter)
1131 					break;
1132 		}
1133 	}
1134 
1135 	/* knote is in the process of changing, wait for it to stablize. */
1136 	if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1137 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1138 		if (filedesc_unlock) {
1139 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1140 			filedesc_unlock = 0;
1141 		}
1142 		kq->kq_state |= KQ_FLUXWAIT;
1143 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1144 		if (fp != NULL) {
1145 			fdrop(fp, td);
1146 			fp = NULL;
1147 		}
1148 		goto findkn;
1149 	}
1150 
1151 	/*
1152 	 * kn now contains the matching knote, or NULL if no match
1153 	 */
1154 	if (kn == NULL) {
1155 		if (kev->flags & EV_ADD) {
1156 			kn = tkn;
1157 			tkn = NULL;
1158 			if (kn == NULL) {
1159 				KQ_UNLOCK(kq);
1160 				error = ENOMEM;
1161 				goto done;
1162 			}
1163 			kn->kn_fp = fp;
1164 			kn->kn_kq = kq;
1165 			kn->kn_fop = fops;
1166 			/*
1167 			 * apply reference counts to knote structure, and
1168 			 * do not release it at the end of this routine.
1169 			 */
1170 			fops = NULL;
1171 			fp = NULL;
1172 
1173 			kn->kn_sfflags = kev->fflags;
1174 			kn->kn_sdata = kev->data;
1175 			kev->fflags = 0;
1176 			kev->data = 0;
1177 			kn->kn_kevent = *kev;
1178 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1179 			    EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1180 			kn->kn_status = KN_INFLUX|KN_DETACHED;
1181 
1182 			error = knote_attach(kn, kq);
1183 			KQ_UNLOCK(kq);
1184 			if (error != 0) {
1185 				tkn = kn;
1186 				goto done;
1187 			}
1188 
1189 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1190 				knote_drop(kn, td);
1191 				goto done;
1192 			}
1193 			KN_LIST_LOCK(kn);
1194 			goto done_ev_add;
1195 		} else {
1196 			/* No matching knote and the EV_ADD flag is not set. */
1197 			KQ_UNLOCK(kq);
1198 			error = ENOENT;
1199 			goto done;
1200 		}
1201 	}
1202 
1203 	if (kev->flags & EV_DELETE) {
1204 		kn->kn_status |= KN_INFLUX;
1205 		KQ_UNLOCK(kq);
1206 		if (!(kn->kn_status & KN_DETACHED))
1207 			kn->kn_fop->f_detach(kn);
1208 		knote_drop(kn, td);
1209 		goto done;
1210 	}
1211 
1212 	if (kev->flags & EV_FORCEONESHOT) {
1213 		kn->kn_flags |= EV_ONESHOT;
1214 		KNOTE_ACTIVATE(kn, 1);
1215 	}
1216 
1217 	/*
1218 	 * The user may change some filter values after the initial EV_ADD,
1219 	 * but doing so will not reset any filter which has already been
1220 	 * triggered.
1221 	 */
1222 	kn->kn_status |= KN_INFLUX | KN_SCAN;
1223 	KQ_UNLOCK(kq);
1224 	KN_LIST_LOCK(kn);
1225 	kn->kn_kevent.udata = kev->udata;
1226 	if (!fops->f_isfd && fops->f_touch != NULL) {
1227 		fops->f_touch(kn, kev, EVENT_REGISTER);
1228 	} else {
1229 		kn->kn_sfflags = kev->fflags;
1230 		kn->kn_sdata = kev->data;
1231 	}
1232 
1233 	/*
1234 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1235 	 * the initial attach event decides that the event is "completed"
1236 	 * already.  i.e. filt_procattach is called on a zombie process.  It
1237 	 * will call filt_proc which will remove it from the list, and NULL
1238 	 * kn_knlist.
1239 	 */
1240 done_ev_add:
1241 	if ((kev->flags & EV_DISABLE) &&
1242 	    ((kn->kn_status & KN_DISABLED) == 0)) {
1243 		kn->kn_status |= KN_DISABLED;
1244 	}
1245 
1246 	if ((kn->kn_status & KN_DISABLED) == 0)
1247 		event = kn->kn_fop->f_event(kn, 0);
1248 	else
1249 		event = 0;
1250 	KQ_LOCK(kq);
1251 	if (event)
1252 		KNOTE_ACTIVATE(kn, 1);
1253 	kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1254 	KN_LIST_UNLOCK(kn);
1255 
1256 	if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1257 		kn->kn_status &= ~KN_DISABLED;
1258 		if ((kn->kn_status & KN_ACTIVE) &&
1259 		    ((kn->kn_status & KN_QUEUED) == 0))
1260 			knote_enqueue(kn);
1261 	}
1262 	KQ_UNLOCK_FLUX(kq);
1263 
1264 done:
1265 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1266 	if (filedesc_unlock)
1267 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1268 	if (fp != NULL)
1269 		fdrop(fp, td);
1270 	if (tkn != NULL)
1271 		knote_free(tkn);
1272 	if (fops != NULL)
1273 		kqueue_fo_release(filt);
1274 	return (error);
1275 }
1276 
1277 static int
1278 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1279 {
1280 	int error;
1281 	struct kqueue *kq;
1282 
1283 	error = 0;
1284 
1285 	kq = fp->f_data;
1286 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1287 		return (EBADF);
1288 	*kqp = kq;
1289 	KQ_LOCK(kq);
1290 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1291 		KQ_UNLOCK(kq);
1292 		return (EBADF);
1293 	}
1294 	kq->kq_refcnt++;
1295 	KQ_UNLOCK(kq);
1296 
1297 	return error;
1298 }
1299 
1300 static void
1301 kqueue_release(struct kqueue *kq, int locked)
1302 {
1303 	if (locked)
1304 		KQ_OWNED(kq);
1305 	else
1306 		KQ_LOCK(kq);
1307 	kq->kq_refcnt--;
1308 	if (kq->kq_refcnt == 1)
1309 		wakeup(&kq->kq_refcnt);
1310 	if (!locked)
1311 		KQ_UNLOCK(kq);
1312 }
1313 
1314 static void
1315 kqueue_schedtask(struct kqueue *kq)
1316 {
1317 
1318 	KQ_OWNED(kq);
1319 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1320 	    ("scheduling kqueue task while draining"));
1321 
1322 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1323 		taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1324 		kq->kq_state |= KQ_TASKSCHED;
1325 	}
1326 }
1327 
1328 /*
1329  * Expand the kq to make sure we have storage for fops/ident pair.
1330  *
1331  * Return 0 on success (or no work necessary), return errno on failure.
1332  *
1333  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1334  * If kqueue_register is called from a non-fd context, there usually/should
1335  * be no locks held.
1336  */
1337 static int
1338 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1339 	int waitok)
1340 {
1341 	struct klist *list, *tmp_knhash, *to_free;
1342 	u_long tmp_knhashmask;
1343 	int size;
1344 	int fd;
1345 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
1346 
1347 	KQ_NOTOWNED(kq);
1348 
1349 	to_free = NULL;
1350 	if (fops->f_isfd) {
1351 		fd = ident;
1352 		if (kq->kq_knlistsize <= fd) {
1353 			size = kq->kq_knlistsize;
1354 			while (size <= fd)
1355 				size += KQEXTENT;
1356 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1357 			if (list == NULL)
1358 				return ENOMEM;
1359 			KQ_LOCK(kq);
1360 			if (kq->kq_knlistsize > fd) {
1361 				to_free = list;
1362 				list = NULL;
1363 			} else {
1364 				if (kq->kq_knlist != NULL) {
1365 					bcopy(kq->kq_knlist, list,
1366 					    kq->kq_knlistsize * sizeof(*list));
1367 					to_free = kq->kq_knlist;
1368 					kq->kq_knlist = NULL;
1369 				}
1370 				bzero((caddr_t)list +
1371 				    kq->kq_knlistsize * sizeof(*list),
1372 				    (size - kq->kq_knlistsize) * sizeof(*list));
1373 				kq->kq_knlistsize = size;
1374 				kq->kq_knlist = list;
1375 			}
1376 			KQ_UNLOCK(kq);
1377 		}
1378 	} else {
1379 		if (kq->kq_knhashmask == 0) {
1380 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1381 			    &tmp_knhashmask);
1382 			if (tmp_knhash == NULL)
1383 				return ENOMEM;
1384 			KQ_LOCK(kq);
1385 			if (kq->kq_knhashmask == 0) {
1386 				kq->kq_knhash = tmp_knhash;
1387 				kq->kq_knhashmask = tmp_knhashmask;
1388 			} else {
1389 				to_free = tmp_knhash;
1390 			}
1391 			KQ_UNLOCK(kq);
1392 		}
1393 	}
1394 	free(to_free, M_KQUEUE);
1395 
1396 	KQ_NOTOWNED(kq);
1397 	return 0;
1398 }
1399 
1400 static void
1401 kqueue_task(void *arg, int pending)
1402 {
1403 	struct kqueue *kq;
1404 	int haskqglobal;
1405 
1406 	haskqglobal = 0;
1407 	kq = arg;
1408 
1409 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1410 	KQ_LOCK(kq);
1411 
1412 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1413 
1414 	kq->kq_state &= ~KQ_TASKSCHED;
1415 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1416 		wakeup(&kq->kq_state);
1417 	}
1418 	KQ_UNLOCK(kq);
1419 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1420 }
1421 
1422 /*
1423  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1424  * We treat KN_MARKER knotes as if they are INFLUX.
1425  */
1426 static int
1427 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1428     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1429 {
1430 	struct kevent *kevp;
1431 	struct knote *kn, *marker;
1432 	sbintime_t asbt, rsbt;
1433 	int count, error, haskqglobal, influx, nkev, touch;
1434 
1435 	count = maxevents;
1436 	nkev = 0;
1437 	error = 0;
1438 	haskqglobal = 0;
1439 
1440 	if (maxevents == 0)
1441 		goto done_nl;
1442 
1443 	rsbt = 0;
1444 	if (tsp != NULL) {
1445 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1446 		    tsp->tv_nsec >= 1000000000) {
1447 			error = EINVAL;
1448 			goto done_nl;
1449 		}
1450 		if (timespecisset(tsp)) {
1451 			if (tsp->tv_sec <= INT32_MAX) {
1452 				rsbt = tstosbt(*tsp);
1453 				if (TIMESEL(&asbt, rsbt))
1454 					asbt += tc_tick_sbt;
1455 				if (asbt <= SBT_MAX - rsbt)
1456 					asbt += rsbt;
1457 				else
1458 					asbt = 0;
1459 				rsbt >>= tc_precexp;
1460 			} else
1461 				asbt = 0;
1462 		} else
1463 			asbt = -1;
1464 	} else
1465 		asbt = 0;
1466 	marker = knote_alloc(1);
1467 	if (marker == NULL) {
1468 		error = ENOMEM;
1469 		goto done_nl;
1470 	}
1471 	marker->kn_status = KN_MARKER;
1472 	KQ_LOCK(kq);
1473 
1474 retry:
1475 	kevp = keva;
1476 	if (kq->kq_count == 0) {
1477 		if (asbt == -1) {
1478 			error = EWOULDBLOCK;
1479 		} else {
1480 			kq->kq_state |= KQ_SLEEP;
1481 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1482 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1483 		}
1484 		if (error == 0)
1485 			goto retry;
1486 		/* don't restart after signals... */
1487 		if (error == ERESTART)
1488 			error = EINTR;
1489 		else if (error == EWOULDBLOCK)
1490 			error = 0;
1491 		goto done;
1492 	}
1493 
1494 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1495 	influx = 0;
1496 	while (count) {
1497 		KQ_OWNED(kq);
1498 		kn = TAILQ_FIRST(&kq->kq_head);
1499 
1500 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1501 		    (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1502 			if (influx) {
1503 				influx = 0;
1504 				KQ_FLUX_WAKEUP(kq);
1505 			}
1506 			kq->kq_state |= KQ_FLUXWAIT;
1507 			error = msleep(kq, &kq->kq_lock, PSOCK,
1508 			    "kqflxwt", 0);
1509 			continue;
1510 		}
1511 
1512 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1513 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1514 			kn->kn_status &= ~KN_QUEUED;
1515 			kq->kq_count--;
1516 			continue;
1517 		}
1518 		if (kn == marker) {
1519 			KQ_FLUX_WAKEUP(kq);
1520 			if (count == maxevents)
1521 				goto retry;
1522 			goto done;
1523 		}
1524 		KASSERT((kn->kn_status & KN_INFLUX) == 0,
1525 		    ("KN_INFLUX set when not suppose to be"));
1526 
1527 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1528 			kn->kn_status &= ~KN_QUEUED;
1529 			kn->kn_status |= KN_INFLUX;
1530 			kq->kq_count--;
1531 			KQ_UNLOCK(kq);
1532 			/*
1533 			 * We don't need to lock the list since we've marked
1534 			 * it _INFLUX.
1535 			 */
1536 			if (!(kn->kn_status & KN_DETACHED))
1537 				kn->kn_fop->f_detach(kn);
1538 			knote_drop(kn, td);
1539 			KQ_LOCK(kq);
1540 			continue;
1541 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1542 			kn->kn_status &= ~KN_QUEUED;
1543 			kn->kn_status |= KN_INFLUX;
1544 			kq->kq_count--;
1545 			KQ_UNLOCK(kq);
1546 			/*
1547 			 * We don't need to lock the list since we've marked
1548 			 * it _INFLUX.
1549 			 */
1550 			*kevp = kn->kn_kevent;
1551 			if (!(kn->kn_status & KN_DETACHED))
1552 				kn->kn_fop->f_detach(kn);
1553 			knote_drop(kn, td);
1554 			KQ_LOCK(kq);
1555 			kn = NULL;
1556 		} else {
1557 			kn->kn_status |= KN_INFLUX | KN_SCAN;
1558 			KQ_UNLOCK(kq);
1559 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1560 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1561 			KN_LIST_LOCK(kn);
1562 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1563 				KQ_LOCK(kq);
1564 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1565 				kn->kn_status &=
1566 				    ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1567 				    KN_SCAN);
1568 				kq->kq_count--;
1569 				KN_LIST_UNLOCK(kn);
1570 				influx = 1;
1571 				continue;
1572 			}
1573 			touch = (!kn->kn_fop->f_isfd &&
1574 			    kn->kn_fop->f_touch != NULL);
1575 			if (touch)
1576 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1577 			else
1578 				*kevp = kn->kn_kevent;
1579 			KQ_LOCK(kq);
1580 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1581 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1582 				/*
1583 				 * Manually clear knotes who weren't
1584 				 * 'touch'ed.
1585 				 */
1586 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1587 					kn->kn_data = 0;
1588 					kn->kn_fflags = 0;
1589 				}
1590 				if (kn->kn_flags & EV_DISPATCH)
1591 					kn->kn_status |= KN_DISABLED;
1592 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1593 				kq->kq_count--;
1594 			} else
1595 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1596 
1597 			kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1598 			KN_LIST_UNLOCK(kn);
1599 			influx = 1;
1600 		}
1601 
1602 		/* we are returning a copy to the user */
1603 		kevp++;
1604 		nkev++;
1605 		count--;
1606 
1607 		if (nkev == KQ_NEVENTS) {
1608 			influx = 0;
1609 			KQ_UNLOCK_FLUX(kq);
1610 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1611 			nkev = 0;
1612 			kevp = keva;
1613 			KQ_LOCK(kq);
1614 			if (error)
1615 				break;
1616 		}
1617 	}
1618 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1619 done:
1620 	KQ_OWNED(kq);
1621 	KQ_UNLOCK_FLUX(kq);
1622 	knote_free(marker);
1623 done_nl:
1624 	KQ_NOTOWNED(kq);
1625 	if (nkev != 0)
1626 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1627 	td->td_retval[0] = maxevents - count;
1628 	return (error);
1629 }
1630 
1631 /*ARGSUSED*/
1632 static int
1633 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1634 	struct ucred *active_cred, struct thread *td)
1635 {
1636 	/*
1637 	 * Enabling sigio causes two major problems:
1638 	 * 1) infinite recursion:
1639 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1640 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1641 	 * into itself over and over.  Sending the sigio causes the kqueue
1642 	 * to become ready, which in turn posts sigio again, forever.
1643 	 * Solution: this can be solved by setting a flag in the kqueue that
1644 	 * we have a SIGIO in progress.
1645 	 * 2) locking problems:
1646 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1647 	 * us above the proc and pgrp locks.
1648 	 * Solution: Post a signal using an async mechanism, being sure to
1649 	 * record a generation count in the delivery so that we do not deliver
1650 	 * a signal to the wrong process.
1651 	 *
1652 	 * Note, these two mechanisms are somewhat mutually exclusive!
1653 	 */
1654 #if 0
1655 	struct kqueue *kq;
1656 
1657 	kq = fp->f_data;
1658 	switch (cmd) {
1659 	case FIOASYNC:
1660 		if (*(int *)data) {
1661 			kq->kq_state |= KQ_ASYNC;
1662 		} else {
1663 			kq->kq_state &= ~KQ_ASYNC;
1664 		}
1665 		return (0);
1666 
1667 	case FIOSETOWN:
1668 		return (fsetown(*(int *)data, &kq->kq_sigio));
1669 
1670 	case FIOGETOWN:
1671 		*(int *)data = fgetown(&kq->kq_sigio);
1672 		return (0);
1673 	}
1674 #endif
1675 
1676 	return (ENOTTY);
1677 }
1678 
1679 /*ARGSUSED*/
1680 static int
1681 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1682 	struct thread *td)
1683 {
1684 	struct kqueue *kq;
1685 	int revents = 0;
1686 	int error;
1687 
1688 	if ((error = kqueue_acquire(fp, &kq)))
1689 		return POLLERR;
1690 
1691 	KQ_LOCK(kq);
1692 	if (events & (POLLIN | POLLRDNORM)) {
1693 		if (kq->kq_count) {
1694 			revents |= events & (POLLIN | POLLRDNORM);
1695 		} else {
1696 			selrecord(td, &kq->kq_sel);
1697 			if (SEL_WAITING(&kq->kq_sel))
1698 				kq->kq_state |= KQ_SEL;
1699 		}
1700 	}
1701 	kqueue_release(kq, 1);
1702 	KQ_UNLOCK(kq);
1703 	return (revents);
1704 }
1705 
1706 /*ARGSUSED*/
1707 static int
1708 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1709 	struct thread *td)
1710 {
1711 
1712 	bzero((void *)st, sizeof *st);
1713 	/*
1714 	 * We no longer return kq_count because the unlocked value is useless.
1715 	 * If you spent all this time getting the count, why not spend your
1716 	 * syscall better by calling kevent?
1717 	 *
1718 	 * XXX - This is needed for libc_r.
1719 	 */
1720 	st->st_mode = S_IFIFO;
1721 	return (0);
1722 }
1723 
1724 /*ARGSUSED*/
1725 static int
1726 kqueue_close(struct file *fp, struct thread *td)
1727 {
1728 	struct kqueue *kq = fp->f_data;
1729 	struct filedesc *fdp;
1730 	struct knote *kn;
1731 	int i;
1732 	int error;
1733 	int filedesc_unlock;
1734 
1735 	if ((error = kqueue_acquire(fp, &kq)))
1736 		return error;
1737 
1738 	filedesc_unlock = 0;
1739 	KQ_LOCK(kq);
1740 
1741 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1742 	    ("kqueue already closing"));
1743 	kq->kq_state |= KQ_CLOSING;
1744 	if (kq->kq_refcnt > 1)
1745 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1746 
1747 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1748 	fdp = kq->kq_fdp;
1749 
1750 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
1751 	    ("kqueue's knlist not empty"));
1752 
1753 	for (i = 0; i < kq->kq_knlistsize; i++) {
1754 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1755 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1756 				kq->kq_state |= KQ_FLUXWAIT;
1757 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1758 				continue;
1759 			}
1760 			kn->kn_status |= KN_INFLUX;
1761 			KQ_UNLOCK(kq);
1762 			if (!(kn->kn_status & KN_DETACHED))
1763 				kn->kn_fop->f_detach(kn);
1764 			knote_drop(kn, td);
1765 			KQ_LOCK(kq);
1766 		}
1767 	}
1768 	if (kq->kq_knhashmask != 0) {
1769 		for (i = 0; i <= kq->kq_knhashmask; i++) {
1770 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1771 				if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1772 					kq->kq_state |= KQ_FLUXWAIT;
1773 					msleep(kq, &kq->kq_lock, PSOCK,
1774 					       "kqclo2", 0);
1775 					continue;
1776 				}
1777 				kn->kn_status |= KN_INFLUX;
1778 				KQ_UNLOCK(kq);
1779 				if (!(kn->kn_status & KN_DETACHED))
1780 					kn->kn_fop->f_detach(kn);
1781 				knote_drop(kn, td);
1782 				KQ_LOCK(kq);
1783 			}
1784 		}
1785 	}
1786 
1787 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1788 		kq->kq_state |= KQ_TASKDRAIN;
1789 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1790 	}
1791 
1792 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1793 		selwakeuppri(&kq->kq_sel, PSOCK);
1794 		if (!SEL_WAITING(&kq->kq_sel))
1795 			kq->kq_state &= ~KQ_SEL;
1796 	}
1797 
1798 	KQ_UNLOCK(kq);
1799 
1800 	/*
1801 	 * We could be called due to the knote_drop() doing fdrop(),
1802 	 * called from kqueue_register().  In this case the global
1803 	 * lock is owned, and filedesc sx is locked before, to not
1804 	 * take the sleepable lock after non-sleepable.
1805 	 */
1806 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1807 		FILEDESC_XLOCK(fdp);
1808 		filedesc_unlock = 1;
1809 	} else
1810 		filedesc_unlock = 0;
1811 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1812 	if (filedesc_unlock)
1813 		FILEDESC_XUNLOCK(fdp);
1814 
1815 	seldrain(&kq->kq_sel);
1816 	knlist_destroy(&kq->kq_sel.si_note);
1817 	mtx_destroy(&kq->kq_lock);
1818 	kq->kq_fdp = NULL;
1819 
1820 	if (kq->kq_knhash != NULL)
1821 		free(kq->kq_knhash, M_KQUEUE);
1822 	if (kq->kq_knlist != NULL)
1823 		free(kq->kq_knlist, M_KQUEUE);
1824 
1825 	funsetown(&kq->kq_sigio);
1826 	chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
1827 	crfree(kq->kq_cred);
1828 	free(kq, M_KQUEUE);
1829 	fp->f_data = NULL;
1830 
1831 	return (0);
1832 }
1833 
1834 static int
1835 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1836 {
1837 
1838 	kif->kf_type = KF_TYPE_KQUEUE;
1839 	return (0);
1840 }
1841 
1842 static void
1843 kqueue_wakeup(struct kqueue *kq)
1844 {
1845 	KQ_OWNED(kq);
1846 
1847 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1848 		kq->kq_state &= ~KQ_SLEEP;
1849 		wakeup(kq);
1850 	}
1851 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1852 		selwakeuppri(&kq->kq_sel, PSOCK);
1853 		if (!SEL_WAITING(&kq->kq_sel))
1854 			kq->kq_state &= ~KQ_SEL;
1855 	}
1856 	if (!knlist_empty(&kq->kq_sel.si_note))
1857 		kqueue_schedtask(kq);
1858 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1859 		pgsigio(&kq->kq_sigio, SIGIO, 0);
1860 	}
1861 }
1862 
1863 /*
1864  * Walk down a list of knotes, activating them if their event has triggered.
1865  *
1866  * There is a possibility to optimize in the case of one kq watching another.
1867  * Instead of scheduling a task to wake it up, you could pass enough state
1868  * down the chain to make up the parent kqueue.  Make this code functional
1869  * first.
1870  */
1871 void
1872 knote(struct knlist *list, long hint, int lockflags)
1873 {
1874 	struct kqueue *kq;
1875 	struct knote *kn;
1876 	int error;
1877 
1878 	if (list == NULL)
1879 		return;
1880 
1881 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1882 
1883 	if ((lockflags & KNF_LISTLOCKED) == 0)
1884 		list->kl_lock(list->kl_lockarg);
1885 
1886 	/*
1887 	 * If we unlock the list lock (and set KN_INFLUX), we can eliminate
1888 	 * the kqueue scheduling, but this will introduce four
1889 	 * lock/unlock's for each knote to test.  If we do, continue to use
1890 	 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is
1891 	 * only safe if you want to remove the current item, which we are
1892 	 * not doing.
1893 	 */
1894 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
1895 		kq = kn->kn_kq;
1896 		KQ_LOCK(kq);
1897 		if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
1898 			/*
1899 			 * Do not process the influx notes, except for
1900 			 * the influx coming from the kq unlock in the
1901 			 * kqueue_scan().  In the later case, we do
1902 			 * not interfere with the scan, since the code
1903 			 * fragment in kqueue_scan() locks the knlist,
1904 			 * and cannot proceed until we finished.
1905 			 */
1906 			KQ_UNLOCK(kq);
1907 		} else if ((lockflags & KNF_NOKQLOCK) != 0) {
1908 			kn->kn_status |= KN_INFLUX;
1909 			KQ_UNLOCK(kq);
1910 			error = kn->kn_fop->f_event(kn, hint);
1911 			KQ_LOCK(kq);
1912 			kn->kn_status &= ~KN_INFLUX;
1913 			if (error)
1914 				KNOTE_ACTIVATE(kn, 1);
1915 			KQ_UNLOCK_FLUX(kq);
1916 		} else {
1917 			kn->kn_status |= KN_HASKQLOCK;
1918 			if (kn->kn_fop->f_event(kn, hint))
1919 				KNOTE_ACTIVATE(kn, 1);
1920 			kn->kn_status &= ~KN_HASKQLOCK;
1921 			KQ_UNLOCK(kq);
1922 		}
1923 	}
1924 	if ((lockflags & KNF_LISTLOCKED) == 0)
1925 		list->kl_unlock(list->kl_lockarg);
1926 }
1927 
1928 /*
1929  * add a knote to a knlist
1930  */
1931 void
1932 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1933 {
1934 	KNL_ASSERT_LOCK(knl, islocked);
1935 	KQ_NOTOWNED(kn->kn_kq);
1936 	KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1937 	    (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1938 	if (!islocked)
1939 		knl->kl_lock(knl->kl_lockarg);
1940 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1941 	if (!islocked)
1942 		knl->kl_unlock(knl->kl_lockarg);
1943 	KQ_LOCK(kn->kn_kq);
1944 	kn->kn_knlist = knl;
1945 	kn->kn_status &= ~KN_DETACHED;
1946 	KQ_UNLOCK(kn->kn_kq);
1947 }
1948 
1949 static void
1950 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1951 {
1952 	KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1953 	KNL_ASSERT_LOCK(knl, knlislocked);
1954 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1955 	if (!kqislocked)
1956 		KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1957     ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1958 	if (!knlislocked)
1959 		knl->kl_lock(knl->kl_lockarg);
1960 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1961 	kn->kn_knlist = NULL;
1962 	if (!knlislocked)
1963 		knl->kl_unlock(knl->kl_lockarg);
1964 	if (!kqislocked)
1965 		KQ_LOCK(kn->kn_kq);
1966 	kn->kn_status |= KN_DETACHED;
1967 	if (!kqislocked)
1968 		KQ_UNLOCK(kn->kn_kq);
1969 }
1970 
1971 /*
1972  * remove knote from the specified knlist
1973  */
1974 void
1975 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1976 {
1977 
1978 	knlist_remove_kq(knl, kn, islocked, 0);
1979 }
1980 
1981 /*
1982  * remove knote from the specified knlist while in f_event handler.
1983  */
1984 void
1985 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1986 {
1987 
1988 	knlist_remove_kq(knl, kn, 1,
1989 	    (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1990 }
1991 
1992 int
1993 knlist_empty(struct knlist *knl)
1994 {
1995 
1996 	KNL_ASSERT_LOCKED(knl);
1997 	return SLIST_EMPTY(&knl->kl_list);
1998 }
1999 
2000 static struct mtx	knlist_lock;
2001 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2002 	MTX_DEF);
2003 static void knlist_mtx_lock(void *arg);
2004 static void knlist_mtx_unlock(void *arg);
2005 
2006 static void
2007 knlist_mtx_lock(void *arg)
2008 {
2009 
2010 	mtx_lock((struct mtx *)arg);
2011 }
2012 
2013 static void
2014 knlist_mtx_unlock(void *arg)
2015 {
2016 
2017 	mtx_unlock((struct mtx *)arg);
2018 }
2019 
2020 static void
2021 knlist_mtx_assert_locked(void *arg)
2022 {
2023 
2024 	mtx_assert((struct mtx *)arg, MA_OWNED);
2025 }
2026 
2027 static void
2028 knlist_mtx_assert_unlocked(void *arg)
2029 {
2030 
2031 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2032 }
2033 
2034 static void
2035 knlist_rw_rlock(void *arg)
2036 {
2037 
2038 	rw_rlock((struct rwlock *)arg);
2039 }
2040 
2041 static void
2042 knlist_rw_runlock(void *arg)
2043 {
2044 
2045 	rw_runlock((struct rwlock *)arg);
2046 }
2047 
2048 static void
2049 knlist_rw_assert_locked(void *arg)
2050 {
2051 
2052 	rw_assert((struct rwlock *)arg, RA_LOCKED);
2053 }
2054 
2055 static void
2056 knlist_rw_assert_unlocked(void *arg)
2057 {
2058 
2059 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2060 }
2061 
2062 void
2063 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2064     void (*kl_unlock)(void *),
2065     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2066 {
2067 
2068 	if (lock == NULL)
2069 		knl->kl_lockarg = &knlist_lock;
2070 	else
2071 		knl->kl_lockarg = lock;
2072 
2073 	if (kl_lock == NULL)
2074 		knl->kl_lock = knlist_mtx_lock;
2075 	else
2076 		knl->kl_lock = kl_lock;
2077 	if (kl_unlock == NULL)
2078 		knl->kl_unlock = knlist_mtx_unlock;
2079 	else
2080 		knl->kl_unlock = kl_unlock;
2081 	if (kl_assert_locked == NULL)
2082 		knl->kl_assert_locked = knlist_mtx_assert_locked;
2083 	else
2084 		knl->kl_assert_locked = kl_assert_locked;
2085 	if (kl_assert_unlocked == NULL)
2086 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2087 	else
2088 		knl->kl_assert_unlocked = kl_assert_unlocked;
2089 
2090 	SLIST_INIT(&knl->kl_list);
2091 }
2092 
2093 void
2094 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2095 {
2096 
2097 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2098 }
2099 
2100 void
2101 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2102 {
2103 
2104 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2105 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2106 }
2107 
2108 void
2109 knlist_destroy(struct knlist *knl)
2110 {
2111 
2112 #ifdef INVARIANTS
2113 	/*
2114 	 * if we run across this error, we need to find the offending
2115 	 * driver and have it call knlist_clear or knlist_delete.
2116 	 */
2117 	if (!SLIST_EMPTY(&knl->kl_list))
2118 		printf("WARNING: destroying knlist w/ knotes on it!\n");
2119 #endif
2120 
2121 	knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2122 	SLIST_INIT(&knl->kl_list);
2123 }
2124 
2125 /*
2126  * Even if we are locked, we may need to drop the lock to allow any influx
2127  * knotes time to "settle".
2128  */
2129 void
2130 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2131 {
2132 	struct knote *kn, *kn2;
2133 	struct kqueue *kq;
2134 
2135 	if (islocked)
2136 		KNL_ASSERT_LOCKED(knl);
2137 	else {
2138 		KNL_ASSERT_UNLOCKED(knl);
2139 again:		/* need to reacquire lock since we have dropped it */
2140 		knl->kl_lock(knl->kl_lockarg);
2141 	}
2142 
2143 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2144 		kq = kn->kn_kq;
2145 		KQ_LOCK(kq);
2146 		if ((kn->kn_status & KN_INFLUX)) {
2147 			KQ_UNLOCK(kq);
2148 			continue;
2149 		}
2150 		knlist_remove_kq(knl, kn, 1, 1);
2151 		if (killkn) {
2152 			kn->kn_status |= KN_INFLUX | KN_DETACHED;
2153 			KQ_UNLOCK(kq);
2154 			knote_drop(kn, td);
2155 		} else {
2156 			/* Make sure cleared knotes disappear soon */
2157 			kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2158 			KQ_UNLOCK(kq);
2159 		}
2160 		kq = NULL;
2161 	}
2162 
2163 	if (!SLIST_EMPTY(&knl->kl_list)) {
2164 		/* there are still KN_INFLUX remaining */
2165 		kn = SLIST_FIRST(&knl->kl_list);
2166 		kq = kn->kn_kq;
2167 		KQ_LOCK(kq);
2168 		KASSERT(kn->kn_status & KN_INFLUX,
2169 		    ("knote removed w/o list lock"));
2170 		knl->kl_unlock(knl->kl_lockarg);
2171 		kq->kq_state |= KQ_FLUXWAIT;
2172 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2173 		kq = NULL;
2174 		goto again;
2175 	}
2176 
2177 	if (islocked)
2178 		KNL_ASSERT_LOCKED(knl);
2179 	else {
2180 		knl->kl_unlock(knl->kl_lockarg);
2181 		KNL_ASSERT_UNLOCKED(knl);
2182 	}
2183 }
2184 
2185 /*
2186  * Remove all knotes referencing a specified fd must be called with FILEDESC
2187  * lock.  This prevents a race where a new fd comes along and occupies the
2188  * entry and we attach a knote to the fd.
2189  */
2190 void
2191 knote_fdclose(struct thread *td, int fd)
2192 {
2193 	struct filedesc *fdp = td->td_proc->p_fd;
2194 	struct kqueue *kq;
2195 	struct knote *kn;
2196 	int influx;
2197 
2198 	FILEDESC_XLOCK_ASSERT(fdp);
2199 
2200 	/*
2201 	 * We shouldn't have to worry about new kevents appearing on fd
2202 	 * since filedesc is locked.
2203 	 */
2204 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2205 		KQ_LOCK(kq);
2206 
2207 again:
2208 		influx = 0;
2209 		while (kq->kq_knlistsize > fd &&
2210 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2211 			if (kn->kn_status & KN_INFLUX) {
2212 				/* someone else might be waiting on our knote */
2213 				if (influx)
2214 					wakeup(kq);
2215 				kq->kq_state |= KQ_FLUXWAIT;
2216 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2217 				goto again;
2218 			}
2219 			kn->kn_status |= KN_INFLUX;
2220 			KQ_UNLOCK(kq);
2221 			if (!(kn->kn_status & KN_DETACHED))
2222 				kn->kn_fop->f_detach(kn);
2223 			knote_drop(kn, td);
2224 			influx = 1;
2225 			KQ_LOCK(kq);
2226 		}
2227 		KQ_UNLOCK_FLUX(kq);
2228 	}
2229 }
2230 
2231 static int
2232 knote_attach(struct knote *kn, struct kqueue *kq)
2233 {
2234 	struct klist *list;
2235 
2236 	KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2237 	KQ_OWNED(kq);
2238 
2239 	if (kn->kn_fop->f_isfd) {
2240 		if (kn->kn_id >= kq->kq_knlistsize)
2241 			return ENOMEM;
2242 		list = &kq->kq_knlist[kn->kn_id];
2243 	} else {
2244 		if (kq->kq_knhash == NULL)
2245 			return ENOMEM;
2246 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2247 	}
2248 
2249 	SLIST_INSERT_HEAD(list, kn, kn_link);
2250 
2251 	return 0;
2252 }
2253 
2254 /*
2255  * knote must already have been detached using the f_detach method.
2256  * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2257  * to prevent other removal.
2258  */
2259 static void
2260 knote_drop(struct knote *kn, struct thread *td)
2261 {
2262 	struct kqueue *kq;
2263 	struct klist *list;
2264 
2265 	kq = kn->kn_kq;
2266 
2267 	KQ_NOTOWNED(kq);
2268 	KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2269 	    ("knote_drop called without KN_INFLUX set in kn_status"));
2270 
2271 	KQ_LOCK(kq);
2272 	if (kn->kn_fop->f_isfd)
2273 		list = &kq->kq_knlist[kn->kn_id];
2274 	else
2275 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2276 
2277 	if (!SLIST_EMPTY(list))
2278 		SLIST_REMOVE(list, kn, knote, kn_link);
2279 	if (kn->kn_status & KN_QUEUED)
2280 		knote_dequeue(kn);
2281 	KQ_UNLOCK_FLUX(kq);
2282 
2283 	if (kn->kn_fop->f_isfd) {
2284 		fdrop(kn->kn_fp, td);
2285 		kn->kn_fp = NULL;
2286 	}
2287 	kqueue_fo_release(kn->kn_kevent.filter);
2288 	kn->kn_fop = NULL;
2289 	knote_free(kn);
2290 }
2291 
2292 static void
2293 knote_enqueue(struct knote *kn)
2294 {
2295 	struct kqueue *kq = kn->kn_kq;
2296 
2297 	KQ_OWNED(kn->kn_kq);
2298 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2299 
2300 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2301 	kn->kn_status |= KN_QUEUED;
2302 	kq->kq_count++;
2303 	kqueue_wakeup(kq);
2304 }
2305 
2306 static void
2307 knote_dequeue(struct knote *kn)
2308 {
2309 	struct kqueue *kq = kn->kn_kq;
2310 
2311 	KQ_OWNED(kn->kn_kq);
2312 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2313 
2314 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2315 	kn->kn_status &= ~KN_QUEUED;
2316 	kq->kq_count--;
2317 }
2318 
2319 static void
2320 knote_init(void)
2321 {
2322 
2323 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2324 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2325 }
2326 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2327 
2328 static struct knote *
2329 knote_alloc(int waitok)
2330 {
2331 	return ((struct knote *)uma_zalloc(knote_zone,
2332 	    (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2333 }
2334 
2335 static void
2336 knote_free(struct knote *kn)
2337 {
2338 	if (kn != NULL)
2339 		uma_zfree(knote_zone, kn);
2340 }
2341 
2342 /*
2343  * Register the kev w/ the kq specified by fd.
2344  */
2345 int
2346 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2347 {
2348 	struct kqueue *kq;
2349 	struct file *fp;
2350 	cap_rights_t rights;
2351 	int error;
2352 
2353 	error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2354 	if (error != 0)
2355 		return (error);
2356 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2357 		goto noacquire;
2358 
2359 	error = kqueue_register(kq, kev, td, waitok);
2360 
2361 	kqueue_release(kq, 0);
2362 
2363 noacquire:
2364 	fdrop(fp, td);
2365 
2366 	return error;
2367 }
2368