xref: /freebsd/sys/kern/kern_event.c (revision 3823d5e198425b4f5e5a80267d195769d1063773)
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_refcnt;
285 } sysfilt_ops[EVFILT_SYSCOUNT] = {
286 	{ &file_filtops },			/* EVFILT_READ */
287 	{ &file_filtops },			/* EVFILT_WRITE */
288 	{ &null_filtops },			/* EVFILT_AIO */
289 	{ &file_filtops },			/* EVFILT_VNODE */
290 	{ &proc_filtops },			/* EVFILT_PROC */
291 	{ &sig_filtops },			/* EVFILT_SIGNAL */
292 	{ &timer_filtops },			/* EVFILT_TIMER */
293 	{ &file_filtops },			/* EVFILT_PROCDESC */
294 	{ &fs_filtops },			/* EVFILT_FS */
295 	{ &null_filtops },			/* EVFILT_LIO */
296 	{ &user_filtops },			/* EVFILT_USER */
297 	{ &null_filtops },			/* EVFILT_SENDFILE */
298 };
299 
300 /*
301  * Simple redirection for all cdevsw style objects to call their fo_kqfilter
302  * method.
303  */
304 static int
305 filt_fileattach(struct knote *kn)
306 {
307 
308 	return (fo_kqfilter(kn->kn_fp, kn));
309 }
310 
311 /*ARGSUSED*/
312 static int
313 kqueue_kqfilter(struct file *fp, struct knote *kn)
314 {
315 	struct kqueue *kq = kn->kn_fp->f_data;
316 
317 	if (kn->kn_filter != EVFILT_READ)
318 		return (EINVAL);
319 
320 	kn->kn_status |= KN_KQUEUE;
321 	kn->kn_fop = &kqread_filtops;
322 	knlist_add(&kq->kq_sel.si_note, kn, 0);
323 
324 	return (0);
325 }
326 
327 static void
328 filt_kqdetach(struct knote *kn)
329 {
330 	struct kqueue *kq = kn->kn_fp->f_data;
331 
332 	knlist_remove(&kq->kq_sel.si_note, kn, 0);
333 }
334 
335 /*ARGSUSED*/
336 static int
337 filt_kqueue(struct knote *kn, long hint)
338 {
339 	struct kqueue *kq = kn->kn_fp->f_data;
340 
341 	kn->kn_data = kq->kq_count;
342 	return (kn->kn_data > 0);
343 }
344 
345 /* XXX - move to kern_proc.c?  */
346 static int
347 filt_procattach(struct knote *kn)
348 {
349 	struct proc *p;
350 	int immediate;
351 	int error;
352 
353 	immediate = 0;
354 	p = pfind(kn->kn_id);
355 	if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
356 		p = zpfind(kn->kn_id);
357 		immediate = 1;
358 	} else if (p != NULL && (p->p_flag & P_WEXIT)) {
359 		immediate = 1;
360 	}
361 
362 	if (p == NULL)
363 		return (ESRCH);
364 	if ((error = p_cansee(curthread, p))) {
365 		PROC_UNLOCK(p);
366 		return (error);
367 	}
368 
369 	kn->kn_ptr.p_proc = p;
370 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
371 
372 	/*
373 	 * internal flag indicating registration done by kernel
374 	 */
375 	if (kn->kn_flags & EV_FLAG1) {
376 		kn->kn_data = kn->kn_sdata;		/* ppid */
377 		kn->kn_fflags = NOTE_CHILD;
378 		kn->kn_flags &= ~EV_FLAG1;
379 	}
380 
381 	if (immediate == 0)
382 		knlist_add(&p->p_klist, kn, 1);
383 
384 	/*
385 	 * Immediately activate any exit notes if the target process is a
386 	 * zombie.  This is necessary to handle the case where the target
387 	 * process, e.g. a child, dies before the kevent is registered.
388 	 */
389 	if (immediate && filt_proc(kn, NOTE_EXIT))
390 		KNOTE_ACTIVATE(kn, 0);
391 
392 	PROC_UNLOCK(p);
393 
394 	return (0);
395 }
396 
397 /*
398  * The knote may be attached to a different process, which may exit,
399  * leaving nothing for the knote to be attached to.  So when the process
400  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
401  * it will be deleted when read out.  However, as part of the knote deletion,
402  * this routine is called, so a check is needed to avoid actually performing
403  * a detach, because the original process does not exist any more.
404  */
405 /* XXX - move to kern_proc.c?  */
406 static void
407 filt_procdetach(struct knote *kn)
408 {
409 	struct proc *p;
410 
411 	p = kn->kn_ptr.p_proc;
412 	knlist_remove(&p->p_klist, kn, 0);
413 	kn->kn_ptr.p_proc = NULL;
414 }
415 
416 /* XXX - move to kern_proc.c?  */
417 static int
418 filt_proc(struct knote *kn, long hint)
419 {
420 	struct proc *p;
421 	u_int event;
422 
423 	p = kn->kn_ptr.p_proc;
424 	/* Mask off extra data. */
425 	event = (u_int)hint & NOTE_PCTRLMASK;
426 
427 	/* If the user is interested in this event, record it. */
428 	if (kn->kn_sfflags & event)
429 		kn->kn_fflags |= event;
430 
431 	/* Process is gone, so flag the event as finished. */
432 	if (event == NOTE_EXIT) {
433 		if (!(kn->kn_status & KN_DETACHED))
434 			knlist_remove_inevent(&p->p_klist, kn);
435 		kn->kn_flags |= EV_EOF | EV_ONESHOT;
436 		kn->kn_ptr.p_proc = NULL;
437 		if (kn->kn_fflags & NOTE_EXIT)
438 			kn->kn_data = p->p_xstat;
439 		if (kn->kn_fflags == 0)
440 			kn->kn_flags |= EV_DROP;
441 		return (1);
442 	}
443 
444 	return (kn->kn_fflags != 0);
445 }
446 
447 /*
448  * Called when the process forked. It mostly does the same as the
449  * knote(), activating all knotes registered to be activated when the
450  * process forked. Additionally, for each knote attached to the
451  * parent, check whether user wants to track the new process. If so
452  * attach a new knote to it, and immediately report an event with the
453  * child's pid.
454  */
455 void
456 knote_fork(struct knlist *list, int pid)
457 {
458 	struct kqueue *kq;
459 	struct knote *kn;
460 	struct kevent kev;
461 	int error;
462 
463 	if (list == NULL)
464 		return;
465 	list->kl_lock(list->kl_lockarg);
466 
467 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
468 		if ((kn->kn_status & KN_INFLUX) == KN_INFLUX)
469 			continue;
470 		kq = kn->kn_kq;
471 		KQ_LOCK(kq);
472 		if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
473 			KQ_UNLOCK(kq);
474 			continue;
475 		}
476 
477 		/*
478 		 * The same as knote(), activate the event.
479 		 */
480 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
481 			kn->kn_status |= KN_HASKQLOCK;
482 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
483 				KNOTE_ACTIVATE(kn, 1);
484 			kn->kn_status &= ~KN_HASKQLOCK;
485 			KQ_UNLOCK(kq);
486 			continue;
487 		}
488 
489 		/*
490 		 * The NOTE_TRACK case. In addition to the activation
491 		 * of the event, we need to register new event to
492 		 * track the child. Drop the locks in preparation for
493 		 * the call to kqueue_register().
494 		 */
495 		kn->kn_status |= KN_INFLUX;
496 		KQ_UNLOCK(kq);
497 		list->kl_unlock(list->kl_lockarg);
498 
499 		/*
500 		 * Activate existing knote and register a knote with
501 		 * new process.
502 		 */
503 		kev.ident = pid;
504 		kev.filter = kn->kn_filter;
505 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
506 		kev.fflags = kn->kn_sfflags;
507 		kev.data = kn->kn_id;		/* parent */
508 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
509 		error = kqueue_register(kq, &kev, NULL, 0);
510 		if (error)
511 			kn->kn_fflags |= NOTE_TRACKERR;
512 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
513 			KNOTE_ACTIVATE(kn, 0);
514 		KQ_LOCK(kq);
515 		kn->kn_status &= ~KN_INFLUX;
516 		KQ_UNLOCK_FLUX(kq);
517 		list->kl_lock(list->kl_lockarg);
518 	}
519 	list->kl_unlock(list->kl_lockarg);
520 }
521 
522 /*
523  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
524  * interval timer support code.
525  */
526 
527 #define NOTE_TIMER_PRECMASK	(NOTE_SECONDS|NOTE_MSECONDS|NOTE_USECONDS| \
528 				NOTE_NSECONDS)
529 
530 static __inline sbintime_t
531 timer2sbintime(intptr_t data, int flags)
532 {
533 	sbintime_t modifier;
534 
535 	switch (flags & NOTE_TIMER_PRECMASK) {
536 	case NOTE_SECONDS:
537 		modifier = SBT_1S;
538 		break;
539 	case NOTE_MSECONDS: /* FALLTHROUGH */
540 	case 0:
541 		modifier = SBT_1MS;
542 		break;
543 	case NOTE_USECONDS:
544 		modifier = SBT_1US;
545 		break;
546 	case NOTE_NSECONDS:
547 		modifier = SBT_1NS;
548 		break;
549 	default:
550 		return (-1);
551 	}
552 
553 #ifdef __LP64__
554 	if (data > SBT_MAX / modifier)
555 		return (SBT_MAX);
556 #endif
557 	return (modifier * data);
558 }
559 
560 static void
561 filt_timerexpire(void *knx)
562 {
563 	struct callout *calloutp;
564 	struct knote *kn;
565 
566 	kn = knx;
567 	kn->kn_data++;
568 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
569 
570 	if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
571 		calloutp = (struct callout *)kn->kn_hook;
572 		*kn->kn_ptr.p_nexttime += timer2sbintime(kn->kn_sdata,
573 		    kn->kn_sfflags);
574 		callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
575 		    filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
576 	}
577 }
578 
579 /*
580  * data contains amount of time to sleep
581  */
582 static int
583 filt_timerattach(struct knote *kn)
584 {
585 	struct callout *calloutp;
586 	sbintime_t to;
587 	unsigned int ncallouts;
588 
589 	if ((intptr_t)kn->kn_sdata < 0)
590 		return (EINVAL);
591 	if ((intptr_t)kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
592 		kn->kn_sdata = 1;
593 	/* Only precision unit are supported in flags so far */
594 	if (kn->kn_sfflags & ~NOTE_TIMER_PRECMASK)
595 		return (EINVAL);
596 
597 	to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
598 	if (to < 0)
599 		return (EINVAL);
600 
601 	ncallouts = atomic_load_explicit(&kq_ncallouts, memory_order_relaxed);
602 	do {
603 		if (ncallouts >= kq_calloutmax)
604 			return (ENOMEM);
605 	} while (!atomic_compare_exchange_weak_explicit(&kq_ncallouts,
606 	    &ncallouts, ncallouts + 1, memory_order_relaxed,
607 	    memory_order_relaxed));
608 
609 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
610 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
611 	kn->kn_ptr.p_nexttime = malloc(sizeof(sbintime_t), M_KQUEUE, M_WAITOK);
612 	calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
613 	callout_init(calloutp, CALLOUT_MPSAFE);
614 	kn->kn_hook = calloutp;
615 	*kn->kn_ptr.p_nexttime = to + sbinuptime();
616 	callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
617 	    filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
618 
619 	return (0);
620 }
621 
622 static void
623 filt_timerdetach(struct knote *kn)
624 {
625 	struct callout *calloutp;
626 	unsigned int old;
627 
628 	calloutp = (struct callout *)kn->kn_hook;
629 	callout_drain(calloutp);
630 	free(calloutp, M_KQUEUE);
631 	free(kn->kn_ptr.p_nexttime, M_KQUEUE);
632 	old = atomic_fetch_sub_explicit(&kq_ncallouts, 1, memory_order_relaxed);
633 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
634 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
635 }
636 
637 static int
638 filt_timer(struct knote *kn, long hint)
639 {
640 
641 	return (kn->kn_data != 0);
642 }
643 
644 static int
645 filt_userattach(struct knote *kn)
646 {
647 
648 	/*
649 	 * EVFILT_USER knotes are not attached to anything in the kernel.
650 	 */
651 	kn->kn_hook = NULL;
652 	if (kn->kn_fflags & NOTE_TRIGGER)
653 		kn->kn_hookid = 1;
654 	else
655 		kn->kn_hookid = 0;
656 	return (0);
657 }
658 
659 static void
660 filt_userdetach(__unused struct knote *kn)
661 {
662 
663 	/*
664 	 * EVFILT_USER knotes are not attached to anything in the kernel.
665 	 */
666 }
667 
668 static int
669 filt_user(struct knote *kn, __unused long hint)
670 {
671 
672 	return (kn->kn_hookid);
673 }
674 
675 static void
676 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
677 {
678 	u_int ffctrl;
679 
680 	switch (type) {
681 	case EVENT_REGISTER:
682 		if (kev->fflags & NOTE_TRIGGER)
683 			kn->kn_hookid = 1;
684 
685 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
686 		kev->fflags &= NOTE_FFLAGSMASK;
687 		switch (ffctrl) {
688 		case NOTE_FFNOP:
689 			break;
690 
691 		case NOTE_FFAND:
692 			kn->kn_sfflags &= kev->fflags;
693 			break;
694 
695 		case NOTE_FFOR:
696 			kn->kn_sfflags |= kev->fflags;
697 			break;
698 
699 		case NOTE_FFCOPY:
700 			kn->kn_sfflags = kev->fflags;
701 			break;
702 
703 		default:
704 			/* XXX Return error? */
705 			break;
706 		}
707 		kn->kn_sdata = kev->data;
708 		if (kev->flags & EV_CLEAR) {
709 			kn->kn_hookid = 0;
710 			kn->kn_data = 0;
711 			kn->kn_fflags = 0;
712 		}
713 		break;
714 
715         case EVENT_PROCESS:
716 		*kev = kn->kn_kevent;
717 		kev->fflags = kn->kn_sfflags;
718 		kev->data = kn->kn_sdata;
719 		if (kn->kn_flags & EV_CLEAR) {
720 			kn->kn_hookid = 0;
721 			kn->kn_data = 0;
722 			kn->kn_fflags = 0;
723 		}
724 		break;
725 
726 	default:
727 		panic("filt_usertouch() - invalid type (%ld)", type);
728 		break;
729 	}
730 }
731 
732 int
733 sys_kqueue(struct thread *td, struct kqueue_args *uap)
734 {
735 	struct filedesc *fdp;
736 	struct kqueue *kq;
737 	struct file *fp;
738 	struct proc *p;
739 	struct ucred *cred;
740 	int fd, error;
741 
742 	p = td->td_proc;
743 	cred = td->td_ucred;
744 	crhold(cred);
745 	PROC_LOCK(p);
746 	if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td->td_proc,
747 	    RLIMIT_KQUEUES))) {
748 		PROC_UNLOCK(p);
749 		crfree(cred);
750 		return (ENOMEM);
751 	}
752 	PROC_UNLOCK(p);
753 
754 	fdp = p->p_fd;
755 	error = falloc(td, &fp, &fd, 0);
756 	if (error)
757 		goto done2;
758 
759 	/* An extra reference on `fp' has been held for us by falloc(). */
760 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
761 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
762 	TAILQ_INIT(&kq->kq_head);
763 	kq->kq_fdp = fdp;
764 	kq->kq_cred = cred;
765 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
766 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
767 
768 	FILEDESC_XLOCK(fdp);
769 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
770 	FILEDESC_XUNLOCK(fdp);
771 
772 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
773 	fdrop(fp, td);
774 
775 	td->td_retval[0] = fd;
776 done2:
777 	if (error != 0) {
778 		chgkqcnt(cred->cr_ruidinfo, -1, 0);
779 		crfree(cred);
780 	}
781 	return (error);
782 }
783 
784 #ifndef _SYS_SYSPROTO_H_
785 struct kevent_args {
786 	int	fd;
787 	const struct kevent *changelist;
788 	int	nchanges;
789 	struct	kevent *eventlist;
790 	int	nevents;
791 	const struct timespec *timeout;
792 };
793 #endif
794 int
795 sys_kevent(struct thread *td, struct kevent_args *uap)
796 {
797 	struct timespec ts, *tsp;
798 	struct kevent_copyops k_ops = { uap,
799 					kevent_copyout,
800 					kevent_copyin};
801 	int error;
802 #ifdef KTRACE
803 	struct uio ktruio;
804 	struct iovec ktriov;
805 	struct uio *ktruioin = NULL;
806 	struct uio *ktruioout = NULL;
807 #endif
808 
809 	if (uap->timeout != NULL) {
810 		error = copyin(uap->timeout, &ts, sizeof(ts));
811 		if (error)
812 			return (error);
813 		tsp = &ts;
814 	} else
815 		tsp = NULL;
816 
817 #ifdef KTRACE
818 	if (KTRPOINT(td, KTR_GENIO)) {
819 		ktriov.iov_base = uap->changelist;
820 		ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
821 		ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
822 		    .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
823 		    .uio_td = td };
824 		ktruioin = cloneuio(&ktruio);
825 		ktriov.iov_base = uap->eventlist;
826 		ktriov.iov_len = uap->nevents * sizeof(struct kevent);
827 		ktruioout = cloneuio(&ktruio);
828 	}
829 #endif
830 
831 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
832 	    &k_ops, tsp);
833 
834 #ifdef KTRACE
835 	if (ktruioin != NULL) {
836 		ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
837 		ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
838 		ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
839 		ktrgenio(uap->fd, UIO_READ, ktruioout, error);
840 	}
841 #endif
842 
843 	return (error);
844 }
845 
846 /*
847  * Copy 'count' items into the destination list pointed to by uap->eventlist.
848  */
849 static int
850 kevent_copyout(void *arg, struct kevent *kevp, int count)
851 {
852 	struct kevent_args *uap;
853 	int error;
854 
855 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
856 	uap = (struct kevent_args *)arg;
857 
858 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
859 	if (error == 0)
860 		uap->eventlist += count;
861 	return (error);
862 }
863 
864 /*
865  * Copy 'count' items from the list pointed to by uap->changelist.
866  */
867 static int
868 kevent_copyin(void *arg, struct kevent *kevp, int count)
869 {
870 	struct kevent_args *uap;
871 	int error;
872 
873 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
874 	uap = (struct kevent_args *)arg;
875 
876 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
877 	if (error == 0)
878 		uap->changelist += count;
879 	return (error);
880 }
881 
882 int
883 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
884     struct kevent_copyops *k_ops, const struct timespec *timeout)
885 {
886 	struct kevent keva[KQ_NEVENTS];
887 	struct kevent *kevp, *changes;
888 	struct kqueue *kq;
889 	struct file *fp;
890 	cap_rights_t rights;
891 	int i, n, nerrors, error;
892 
893 	cap_rights_init(&rights);
894 	if (nchanges > 0)
895 		cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
896 	if (nevents > 0)
897 		cap_rights_set(&rights, CAP_KQUEUE_EVENT);
898 	error = fget(td, fd, &rights, &fp);
899 	if (error != 0)
900 		return (error);
901 
902 	error = kqueue_acquire(fp, &kq);
903 	if (error != 0)
904 		goto done_norel;
905 
906 	nerrors = 0;
907 
908 	while (nchanges > 0) {
909 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
910 		error = k_ops->k_copyin(k_ops->arg, keva, n);
911 		if (error)
912 			goto done;
913 		changes = keva;
914 		for (i = 0; i < n; i++) {
915 			kevp = &changes[i];
916 			if (!kevp->filter)
917 				continue;
918 			kevp->flags &= ~EV_SYSFLAGS;
919 			error = kqueue_register(kq, kevp, td, 1);
920 			if (error || (kevp->flags & EV_RECEIPT)) {
921 				if (nevents != 0) {
922 					kevp->flags = EV_ERROR;
923 					kevp->data = error;
924 					(void) k_ops->k_copyout(k_ops->arg,
925 					    kevp, 1);
926 					nevents--;
927 					nerrors++;
928 				} else {
929 					goto done;
930 				}
931 			}
932 		}
933 		nchanges -= n;
934 	}
935 	if (nerrors) {
936 		td->td_retval[0] = nerrors;
937 		error = 0;
938 		goto done;
939 	}
940 
941 	error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
942 done:
943 	kqueue_release(kq, 0);
944 done_norel:
945 	fdrop(fp, td);
946 	return (error);
947 }
948 
949 int
950 kqueue_add_filteropts(int filt, struct filterops *filtops)
951 {
952 	int error;
953 
954 	error = 0;
955 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
956 		printf(
957 "trying to add a filterop that is out of range: %d is beyond %d\n",
958 		    ~filt, EVFILT_SYSCOUNT);
959 		return EINVAL;
960 	}
961 	mtx_lock(&filterops_lock);
962 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
963 	    sysfilt_ops[~filt].for_fop != NULL)
964 		error = EEXIST;
965 	else {
966 		sysfilt_ops[~filt].for_fop = filtops;
967 		sysfilt_ops[~filt].for_refcnt = 0;
968 	}
969 	mtx_unlock(&filterops_lock);
970 
971 	return (error);
972 }
973 
974 int
975 kqueue_del_filteropts(int filt)
976 {
977 	int error;
978 
979 	error = 0;
980 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
981 		return EINVAL;
982 
983 	mtx_lock(&filterops_lock);
984 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
985 	    sysfilt_ops[~filt].for_fop == NULL)
986 		error = EINVAL;
987 	else if (sysfilt_ops[~filt].for_refcnt != 0)
988 		error = EBUSY;
989 	else {
990 		sysfilt_ops[~filt].for_fop = &null_filtops;
991 		sysfilt_ops[~filt].for_refcnt = 0;
992 	}
993 	mtx_unlock(&filterops_lock);
994 
995 	return error;
996 }
997 
998 static struct filterops *
999 kqueue_fo_find(int filt)
1000 {
1001 
1002 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1003 		return NULL;
1004 
1005 	mtx_lock(&filterops_lock);
1006 	sysfilt_ops[~filt].for_refcnt++;
1007 	if (sysfilt_ops[~filt].for_fop == NULL)
1008 		sysfilt_ops[~filt].for_fop = &null_filtops;
1009 	mtx_unlock(&filterops_lock);
1010 
1011 	return sysfilt_ops[~filt].for_fop;
1012 }
1013 
1014 static void
1015 kqueue_fo_release(int filt)
1016 {
1017 
1018 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1019 		return;
1020 
1021 	mtx_lock(&filterops_lock);
1022 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1023 	    ("filter object refcount not valid on release"));
1024 	sysfilt_ops[~filt].for_refcnt--;
1025 	mtx_unlock(&filterops_lock);
1026 }
1027 
1028 /*
1029  * A ref to kq (obtained via kqueue_acquire) must be held.  waitok will
1030  * influence if memory allocation should wait.  Make sure it is 0 if you
1031  * hold any mutexes.
1032  */
1033 static int
1034 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1035 {
1036 	struct filterops *fops;
1037 	struct file *fp;
1038 	struct knote *kn, *tkn;
1039 	cap_rights_t rights;
1040 	int error, filt, event;
1041 	int haskqglobal, filedesc_unlock;
1042 
1043 	fp = NULL;
1044 	kn = NULL;
1045 	error = 0;
1046 	haskqglobal = 0;
1047 	filedesc_unlock = 0;
1048 
1049 	filt = kev->filter;
1050 	fops = kqueue_fo_find(filt);
1051 	if (fops == NULL)
1052 		return EINVAL;
1053 
1054 	tkn = knote_alloc(waitok);		/* prevent waiting with locks */
1055 
1056 findkn:
1057 	if (fops->f_isfd) {
1058 		KASSERT(td != NULL, ("td is NULL"));
1059 		error = fget(td, kev->ident,
1060 		    cap_rights_init(&rights, CAP_EVENT), &fp);
1061 		if (error)
1062 			goto done;
1063 
1064 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1065 		    kev->ident, 0) != 0) {
1066 			/* try again */
1067 			fdrop(fp, td);
1068 			fp = NULL;
1069 			error = kqueue_expand(kq, fops, kev->ident, waitok);
1070 			if (error)
1071 				goto done;
1072 			goto findkn;
1073 		}
1074 
1075 		if (fp->f_type == DTYPE_KQUEUE) {
1076 			/*
1077 			 * if we add some inteligence about what we are doing,
1078 			 * we should be able to support events on ourselves.
1079 			 * We need to know when we are doing this to prevent
1080 			 * getting both the knlist lock and the kq lock since
1081 			 * they are the same thing.
1082 			 */
1083 			if (fp->f_data == kq) {
1084 				error = EINVAL;
1085 				goto done;
1086 			}
1087 
1088 			/*
1089 			 * Pre-lock the filedesc before the global
1090 			 * lock mutex, see the comment in
1091 			 * kqueue_close().
1092 			 */
1093 			FILEDESC_XLOCK(td->td_proc->p_fd);
1094 			filedesc_unlock = 1;
1095 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1096 		}
1097 
1098 		KQ_LOCK(kq);
1099 		if (kev->ident < kq->kq_knlistsize) {
1100 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1101 				if (kev->filter == kn->kn_filter)
1102 					break;
1103 		}
1104 	} else {
1105 		if ((kev->flags & EV_ADD) == EV_ADD)
1106 			kqueue_expand(kq, fops, kev->ident, waitok);
1107 
1108 		KQ_LOCK(kq);
1109 		if (kq->kq_knhashmask != 0) {
1110 			struct klist *list;
1111 
1112 			list = &kq->kq_knhash[
1113 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1114 			SLIST_FOREACH(kn, list, kn_link)
1115 				if (kev->ident == kn->kn_id &&
1116 				    kev->filter == kn->kn_filter)
1117 					break;
1118 		}
1119 	}
1120 
1121 	/* knote is in the process of changing, wait for it to stablize. */
1122 	if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1123 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1124 		if (filedesc_unlock) {
1125 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1126 			filedesc_unlock = 0;
1127 		}
1128 		kq->kq_state |= KQ_FLUXWAIT;
1129 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1130 		if (fp != NULL) {
1131 			fdrop(fp, td);
1132 			fp = NULL;
1133 		}
1134 		goto findkn;
1135 	}
1136 
1137 	/*
1138 	 * kn now contains the matching knote, or NULL if no match
1139 	 */
1140 	if (kn == NULL) {
1141 		if (kev->flags & EV_ADD) {
1142 			kn = tkn;
1143 			tkn = NULL;
1144 			if (kn == NULL) {
1145 				KQ_UNLOCK(kq);
1146 				error = ENOMEM;
1147 				goto done;
1148 			}
1149 			kn->kn_fp = fp;
1150 			kn->kn_kq = kq;
1151 			kn->kn_fop = fops;
1152 			/*
1153 			 * apply reference counts to knote structure, and
1154 			 * do not release it at the end of this routine.
1155 			 */
1156 			fops = NULL;
1157 			fp = NULL;
1158 
1159 			kn->kn_sfflags = kev->fflags;
1160 			kn->kn_sdata = kev->data;
1161 			kev->fflags = 0;
1162 			kev->data = 0;
1163 			kn->kn_kevent = *kev;
1164 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1165 			    EV_ENABLE | EV_DISABLE);
1166 			kn->kn_status = KN_INFLUX|KN_DETACHED;
1167 
1168 			error = knote_attach(kn, kq);
1169 			KQ_UNLOCK(kq);
1170 			if (error != 0) {
1171 				tkn = kn;
1172 				goto done;
1173 			}
1174 
1175 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1176 				knote_drop(kn, td);
1177 				goto done;
1178 			}
1179 			KN_LIST_LOCK(kn);
1180 			goto done_ev_add;
1181 		} else {
1182 			/* No matching knote and the EV_ADD flag is not set. */
1183 			KQ_UNLOCK(kq);
1184 			error = ENOENT;
1185 			goto done;
1186 		}
1187 	}
1188 
1189 	if (kev->flags & EV_DELETE) {
1190 		kn->kn_status |= KN_INFLUX;
1191 		KQ_UNLOCK(kq);
1192 		if (!(kn->kn_status & KN_DETACHED))
1193 			kn->kn_fop->f_detach(kn);
1194 		knote_drop(kn, td);
1195 		goto done;
1196 	}
1197 
1198 	/*
1199 	 * The user may change some filter values after the initial EV_ADD,
1200 	 * but doing so will not reset any filter which has already been
1201 	 * triggered.
1202 	 */
1203 	kn->kn_status |= KN_INFLUX | KN_SCAN;
1204 	KQ_UNLOCK(kq);
1205 	KN_LIST_LOCK(kn);
1206 	kn->kn_kevent.udata = kev->udata;
1207 	if (!fops->f_isfd && fops->f_touch != NULL) {
1208 		fops->f_touch(kn, kev, EVENT_REGISTER);
1209 	} else {
1210 		kn->kn_sfflags = kev->fflags;
1211 		kn->kn_sdata = kev->data;
1212 	}
1213 
1214 	/*
1215 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1216 	 * the initial attach event decides that the event is "completed"
1217 	 * already.  i.e. filt_procattach is called on a zombie process.  It
1218 	 * will call filt_proc which will remove it from the list, and NULL
1219 	 * kn_knlist.
1220 	 */
1221 done_ev_add:
1222 	event = kn->kn_fop->f_event(kn, 0);
1223 	KQ_LOCK(kq);
1224 	if (event)
1225 		KNOTE_ACTIVATE(kn, 1);
1226 	kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1227 	KN_LIST_UNLOCK(kn);
1228 
1229 	if ((kev->flags & EV_DISABLE) &&
1230 	    ((kn->kn_status & KN_DISABLED) == 0)) {
1231 		kn->kn_status |= KN_DISABLED;
1232 	}
1233 
1234 	if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1235 		kn->kn_status &= ~KN_DISABLED;
1236 		if ((kn->kn_status & KN_ACTIVE) &&
1237 		    ((kn->kn_status & KN_QUEUED) == 0))
1238 			knote_enqueue(kn);
1239 	}
1240 	KQ_UNLOCK_FLUX(kq);
1241 
1242 done:
1243 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1244 	if (filedesc_unlock)
1245 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1246 	if (fp != NULL)
1247 		fdrop(fp, td);
1248 	if (tkn != NULL)
1249 		knote_free(tkn);
1250 	if (fops != NULL)
1251 		kqueue_fo_release(filt);
1252 	return (error);
1253 }
1254 
1255 static int
1256 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1257 {
1258 	int error;
1259 	struct kqueue *kq;
1260 
1261 	error = 0;
1262 
1263 	kq = fp->f_data;
1264 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1265 		return (EBADF);
1266 	*kqp = kq;
1267 	KQ_LOCK(kq);
1268 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1269 		KQ_UNLOCK(kq);
1270 		return (EBADF);
1271 	}
1272 	kq->kq_refcnt++;
1273 	KQ_UNLOCK(kq);
1274 
1275 	return error;
1276 }
1277 
1278 static void
1279 kqueue_release(struct kqueue *kq, int locked)
1280 {
1281 	if (locked)
1282 		KQ_OWNED(kq);
1283 	else
1284 		KQ_LOCK(kq);
1285 	kq->kq_refcnt--;
1286 	if (kq->kq_refcnt == 1)
1287 		wakeup(&kq->kq_refcnt);
1288 	if (!locked)
1289 		KQ_UNLOCK(kq);
1290 }
1291 
1292 static void
1293 kqueue_schedtask(struct kqueue *kq)
1294 {
1295 
1296 	KQ_OWNED(kq);
1297 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1298 	    ("scheduling kqueue task while draining"));
1299 
1300 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1301 		taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1302 		kq->kq_state |= KQ_TASKSCHED;
1303 	}
1304 }
1305 
1306 /*
1307  * Expand the kq to make sure we have storage for fops/ident pair.
1308  *
1309  * Return 0 on success (or no work necessary), return errno on failure.
1310  *
1311  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1312  * If kqueue_register is called from a non-fd context, there usually/should
1313  * be no locks held.
1314  */
1315 static int
1316 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1317 	int waitok)
1318 {
1319 	struct klist *list, *tmp_knhash, *to_free;
1320 	u_long tmp_knhashmask;
1321 	int size;
1322 	int fd;
1323 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
1324 
1325 	KQ_NOTOWNED(kq);
1326 
1327 	to_free = NULL;
1328 	if (fops->f_isfd) {
1329 		fd = ident;
1330 		if (kq->kq_knlistsize <= fd) {
1331 			size = kq->kq_knlistsize;
1332 			while (size <= fd)
1333 				size += KQEXTENT;
1334 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1335 			if (list == NULL)
1336 				return ENOMEM;
1337 			KQ_LOCK(kq);
1338 			if (kq->kq_knlistsize > fd) {
1339 				to_free = list;
1340 				list = NULL;
1341 			} else {
1342 				if (kq->kq_knlist != NULL) {
1343 					bcopy(kq->kq_knlist, list,
1344 					    kq->kq_knlistsize * sizeof(*list));
1345 					to_free = kq->kq_knlist;
1346 					kq->kq_knlist = NULL;
1347 				}
1348 				bzero((caddr_t)list +
1349 				    kq->kq_knlistsize * sizeof(*list),
1350 				    (size - kq->kq_knlistsize) * sizeof(*list));
1351 				kq->kq_knlistsize = size;
1352 				kq->kq_knlist = list;
1353 			}
1354 			KQ_UNLOCK(kq);
1355 		}
1356 	} else {
1357 		if (kq->kq_knhashmask == 0) {
1358 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1359 			    &tmp_knhashmask);
1360 			if (tmp_knhash == NULL)
1361 				return ENOMEM;
1362 			KQ_LOCK(kq);
1363 			if (kq->kq_knhashmask == 0) {
1364 				kq->kq_knhash = tmp_knhash;
1365 				kq->kq_knhashmask = tmp_knhashmask;
1366 			} else {
1367 				to_free = tmp_knhash;
1368 			}
1369 			KQ_UNLOCK(kq);
1370 		}
1371 	}
1372 	free(to_free, M_KQUEUE);
1373 
1374 	KQ_NOTOWNED(kq);
1375 	return 0;
1376 }
1377 
1378 static void
1379 kqueue_task(void *arg, int pending)
1380 {
1381 	struct kqueue *kq;
1382 	int haskqglobal;
1383 
1384 	haskqglobal = 0;
1385 	kq = arg;
1386 
1387 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1388 	KQ_LOCK(kq);
1389 
1390 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1391 
1392 	kq->kq_state &= ~KQ_TASKSCHED;
1393 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1394 		wakeup(&kq->kq_state);
1395 	}
1396 	KQ_UNLOCK(kq);
1397 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1398 }
1399 
1400 /*
1401  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1402  * We treat KN_MARKER knotes as if they are INFLUX.
1403  */
1404 static int
1405 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1406     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1407 {
1408 	struct kevent *kevp;
1409 	struct knote *kn, *marker;
1410 	sbintime_t asbt, rsbt;
1411 	int count, error, haskqglobal, influx, nkev, touch;
1412 
1413 	count = maxevents;
1414 	nkev = 0;
1415 	error = 0;
1416 	haskqglobal = 0;
1417 
1418 	if (maxevents == 0)
1419 		goto done_nl;
1420 
1421 	rsbt = 0;
1422 	if (tsp != NULL) {
1423 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1424 		    tsp->tv_nsec >= 1000000000) {
1425 			error = EINVAL;
1426 			goto done_nl;
1427 		}
1428 		if (timespecisset(tsp)) {
1429 			if (tsp->tv_sec <= INT32_MAX) {
1430 				rsbt = tstosbt(*tsp);
1431 				if (TIMESEL(&asbt, rsbt))
1432 					asbt += tc_tick_sbt;
1433 				if (asbt <= SBT_MAX - rsbt)
1434 					asbt += rsbt;
1435 				else
1436 					asbt = 0;
1437 				rsbt >>= tc_precexp;
1438 			} else
1439 				asbt = 0;
1440 		} else
1441 			asbt = -1;
1442 	} else
1443 		asbt = 0;
1444 	marker = knote_alloc(1);
1445 	if (marker == NULL) {
1446 		error = ENOMEM;
1447 		goto done_nl;
1448 	}
1449 	marker->kn_status = KN_MARKER;
1450 	KQ_LOCK(kq);
1451 
1452 retry:
1453 	kevp = keva;
1454 	if (kq->kq_count == 0) {
1455 		if (asbt == -1) {
1456 			error = EWOULDBLOCK;
1457 		} else {
1458 			kq->kq_state |= KQ_SLEEP;
1459 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1460 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1461 		}
1462 		if (error == 0)
1463 			goto retry;
1464 		/* don't restart after signals... */
1465 		if (error == ERESTART)
1466 			error = EINTR;
1467 		else if (error == EWOULDBLOCK)
1468 			error = 0;
1469 		goto done;
1470 	}
1471 
1472 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1473 	influx = 0;
1474 	while (count) {
1475 		KQ_OWNED(kq);
1476 		kn = TAILQ_FIRST(&kq->kq_head);
1477 
1478 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1479 		    (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1480 			if (influx) {
1481 				influx = 0;
1482 				KQ_FLUX_WAKEUP(kq);
1483 			}
1484 			kq->kq_state |= KQ_FLUXWAIT;
1485 			error = msleep(kq, &kq->kq_lock, PSOCK,
1486 			    "kqflxwt", 0);
1487 			continue;
1488 		}
1489 
1490 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1491 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1492 			kn->kn_status &= ~KN_QUEUED;
1493 			kq->kq_count--;
1494 			continue;
1495 		}
1496 		if (kn == marker) {
1497 			KQ_FLUX_WAKEUP(kq);
1498 			if (count == maxevents)
1499 				goto retry;
1500 			goto done;
1501 		}
1502 		KASSERT((kn->kn_status & KN_INFLUX) == 0,
1503 		    ("KN_INFLUX set when not suppose to be"));
1504 
1505 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1506 			kn->kn_status &= ~KN_QUEUED;
1507 			kn->kn_status |= KN_INFLUX;
1508 			kq->kq_count--;
1509 			KQ_UNLOCK(kq);
1510 			/*
1511 			 * We don't need to lock the list since we've marked
1512 			 * it _INFLUX.
1513 			 */
1514 			if (!(kn->kn_status & KN_DETACHED))
1515 				kn->kn_fop->f_detach(kn);
1516 			knote_drop(kn, td);
1517 			KQ_LOCK(kq);
1518 			continue;
1519 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1520 			kn->kn_status &= ~KN_QUEUED;
1521 			kn->kn_status |= KN_INFLUX;
1522 			kq->kq_count--;
1523 			KQ_UNLOCK(kq);
1524 			/*
1525 			 * We don't need to lock the list since we've marked
1526 			 * it _INFLUX.
1527 			 */
1528 			*kevp = kn->kn_kevent;
1529 			if (!(kn->kn_status & KN_DETACHED))
1530 				kn->kn_fop->f_detach(kn);
1531 			knote_drop(kn, td);
1532 			KQ_LOCK(kq);
1533 			kn = NULL;
1534 		} else {
1535 			kn->kn_status |= KN_INFLUX | KN_SCAN;
1536 			KQ_UNLOCK(kq);
1537 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1538 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1539 			KN_LIST_LOCK(kn);
1540 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1541 				KQ_LOCK(kq);
1542 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1543 				kn->kn_status &=
1544 				    ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1545 				    KN_SCAN);
1546 				kq->kq_count--;
1547 				KN_LIST_UNLOCK(kn);
1548 				influx = 1;
1549 				continue;
1550 			}
1551 			touch = (!kn->kn_fop->f_isfd &&
1552 			    kn->kn_fop->f_touch != NULL);
1553 			if (touch)
1554 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1555 			else
1556 				*kevp = kn->kn_kevent;
1557 			KQ_LOCK(kq);
1558 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1559 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1560 				/*
1561 				 * Manually clear knotes who weren't
1562 				 * 'touch'ed.
1563 				 */
1564 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1565 					kn->kn_data = 0;
1566 					kn->kn_fflags = 0;
1567 				}
1568 				if (kn->kn_flags & EV_DISPATCH)
1569 					kn->kn_status |= KN_DISABLED;
1570 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1571 				kq->kq_count--;
1572 			} else
1573 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1574 
1575 			kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1576 			KN_LIST_UNLOCK(kn);
1577 			influx = 1;
1578 		}
1579 
1580 		/* we are returning a copy to the user */
1581 		kevp++;
1582 		nkev++;
1583 		count--;
1584 
1585 		if (nkev == KQ_NEVENTS) {
1586 			influx = 0;
1587 			KQ_UNLOCK_FLUX(kq);
1588 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1589 			nkev = 0;
1590 			kevp = keva;
1591 			KQ_LOCK(kq);
1592 			if (error)
1593 				break;
1594 		}
1595 	}
1596 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1597 done:
1598 	KQ_OWNED(kq);
1599 	KQ_UNLOCK_FLUX(kq);
1600 	knote_free(marker);
1601 done_nl:
1602 	KQ_NOTOWNED(kq);
1603 	if (nkev != 0)
1604 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1605 	td->td_retval[0] = maxevents - count;
1606 	return (error);
1607 }
1608 
1609 /*ARGSUSED*/
1610 static int
1611 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1612 	struct ucred *active_cred, struct thread *td)
1613 {
1614 	/*
1615 	 * Enabling sigio causes two major problems:
1616 	 * 1) infinite recursion:
1617 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1618 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1619 	 * into itself over and over.  Sending the sigio causes the kqueue
1620 	 * to become ready, which in turn posts sigio again, forever.
1621 	 * Solution: this can be solved by setting a flag in the kqueue that
1622 	 * we have a SIGIO in progress.
1623 	 * 2) locking problems:
1624 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1625 	 * us above the proc and pgrp locks.
1626 	 * Solution: Post a signal using an async mechanism, being sure to
1627 	 * record a generation count in the delivery so that we do not deliver
1628 	 * a signal to the wrong process.
1629 	 *
1630 	 * Note, these two mechanisms are somewhat mutually exclusive!
1631 	 */
1632 #if 0
1633 	struct kqueue *kq;
1634 
1635 	kq = fp->f_data;
1636 	switch (cmd) {
1637 	case FIOASYNC:
1638 		if (*(int *)data) {
1639 			kq->kq_state |= KQ_ASYNC;
1640 		} else {
1641 			kq->kq_state &= ~KQ_ASYNC;
1642 		}
1643 		return (0);
1644 
1645 	case FIOSETOWN:
1646 		return (fsetown(*(int *)data, &kq->kq_sigio));
1647 
1648 	case FIOGETOWN:
1649 		*(int *)data = fgetown(&kq->kq_sigio);
1650 		return (0);
1651 	}
1652 #endif
1653 
1654 	return (ENOTTY);
1655 }
1656 
1657 /*ARGSUSED*/
1658 static int
1659 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1660 	struct thread *td)
1661 {
1662 	struct kqueue *kq;
1663 	int revents = 0;
1664 	int error;
1665 
1666 	if ((error = kqueue_acquire(fp, &kq)))
1667 		return POLLERR;
1668 
1669 	KQ_LOCK(kq);
1670 	if (events & (POLLIN | POLLRDNORM)) {
1671 		if (kq->kq_count) {
1672 			revents |= events & (POLLIN | POLLRDNORM);
1673 		} else {
1674 			selrecord(td, &kq->kq_sel);
1675 			if (SEL_WAITING(&kq->kq_sel))
1676 				kq->kq_state |= KQ_SEL;
1677 		}
1678 	}
1679 	kqueue_release(kq, 1);
1680 	KQ_UNLOCK(kq);
1681 	return (revents);
1682 }
1683 
1684 /*ARGSUSED*/
1685 static int
1686 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1687 	struct thread *td)
1688 {
1689 
1690 	bzero((void *)st, sizeof *st);
1691 	/*
1692 	 * We no longer return kq_count because the unlocked value is useless.
1693 	 * If you spent all this time getting the count, why not spend your
1694 	 * syscall better by calling kevent?
1695 	 *
1696 	 * XXX - This is needed for libc_r.
1697 	 */
1698 	st->st_mode = S_IFIFO;
1699 	return (0);
1700 }
1701 
1702 /*ARGSUSED*/
1703 static int
1704 kqueue_close(struct file *fp, struct thread *td)
1705 {
1706 	struct kqueue *kq = fp->f_data;
1707 	struct filedesc *fdp;
1708 	struct knote *kn;
1709 	int i;
1710 	int error;
1711 	int filedesc_unlock;
1712 
1713 	if ((error = kqueue_acquire(fp, &kq)))
1714 		return error;
1715 
1716 	filedesc_unlock = 0;
1717 	KQ_LOCK(kq);
1718 
1719 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1720 	    ("kqueue already closing"));
1721 	kq->kq_state |= KQ_CLOSING;
1722 	if (kq->kq_refcnt > 1)
1723 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1724 
1725 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1726 	fdp = kq->kq_fdp;
1727 
1728 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
1729 	    ("kqueue's knlist not empty"));
1730 
1731 	for (i = 0; i < kq->kq_knlistsize; i++) {
1732 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1733 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1734 				kq->kq_state |= KQ_FLUXWAIT;
1735 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1736 				continue;
1737 			}
1738 			kn->kn_status |= KN_INFLUX;
1739 			KQ_UNLOCK(kq);
1740 			if (!(kn->kn_status & KN_DETACHED))
1741 				kn->kn_fop->f_detach(kn);
1742 			knote_drop(kn, td);
1743 			KQ_LOCK(kq);
1744 		}
1745 	}
1746 	if (kq->kq_knhashmask != 0) {
1747 		for (i = 0; i <= kq->kq_knhashmask; i++) {
1748 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1749 				if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1750 					kq->kq_state |= KQ_FLUXWAIT;
1751 					msleep(kq, &kq->kq_lock, PSOCK,
1752 					       "kqclo2", 0);
1753 					continue;
1754 				}
1755 				kn->kn_status |= KN_INFLUX;
1756 				KQ_UNLOCK(kq);
1757 				if (!(kn->kn_status & KN_DETACHED))
1758 					kn->kn_fop->f_detach(kn);
1759 				knote_drop(kn, td);
1760 				KQ_LOCK(kq);
1761 			}
1762 		}
1763 	}
1764 
1765 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1766 		kq->kq_state |= KQ_TASKDRAIN;
1767 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1768 	}
1769 
1770 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1771 		selwakeuppri(&kq->kq_sel, PSOCK);
1772 		if (!SEL_WAITING(&kq->kq_sel))
1773 			kq->kq_state &= ~KQ_SEL;
1774 	}
1775 
1776 	KQ_UNLOCK(kq);
1777 
1778 	/*
1779 	 * We could be called due to the knote_drop() doing fdrop(),
1780 	 * called from kqueue_register().  In this case the global
1781 	 * lock is owned, and filedesc sx is locked before, to not
1782 	 * take the sleepable lock after non-sleepable.
1783 	 */
1784 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1785 		FILEDESC_XLOCK(fdp);
1786 		filedesc_unlock = 1;
1787 	} else
1788 		filedesc_unlock = 0;
1789 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1790 	if (filedesc_unlock)
1791 		FILEDESC_XUNLOCK(fdp);
1792 
1793 	seldrain(&kq->kq_sel);
1794 	knlist_destroy(&kq->kq_sel.si_note);
1795 	mtx_destroy(&kq->kq_lock);
1796 	kq->kq_fdp = NULL;
1797 
1798 	if (kq->kq_knhash != NULL)
1799 		free(kq->kq_knhash, M_KQUEUE);
1800 	if (kq->kq_knlist != NULL)
1801 		free(kq->kq_knlist, M_KQUEUE);
1802 
1803 	funsetown(&kq->kq_sigio);
1804 	chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
1805 	crfree(kq->kq_cred);
1806 	free(kq, M_KQUEUE);
1807 	fp->f_data = NULL;
1808 
1809 	return (0);
1810 }
1811 
1812 static int
1813 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1814 {
1815 
1816 	kif->kf_type = KF_TYPE_KQUEUE;
1817 	return (0);
1818 }
1819 
1820 static void
1821 kqueue_wakeup(struct kqueue *kq)
1822 {
1823 	KQ_OWNED(kq);
1824 
1825 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1826 		kq->kq_state &= ~KQ_SLEEP;
1827 		wakeup(kq);
1828 	}
1829 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1830 		selwakeuppri(&kq->kq_sel, PSOCK);
1831 		if (!SEL_WAITING(&kq->kq_sel))
1832 			kq->kq_state &= ~KQ_SEL;
1833 	}
1834 	if (!knlist_empty(&kq->kq_sel.si_note))
1835 		kqueue_schedtask(kq);
1836 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1837 		pgsigio(&kq->kq_sigio, SIGIO, 0);
1838 	}
1839 }
1840 
1841 /*
1842  * Walk down a list of knotes, activating them if their event has triggered.
1843  *
1844  * There is a possibility to optimize in the case of one kq watching another.
1845  * Instead of scheduling a task to wake it up, you could pass enough state
1846  * down the chain to make up the parent kqueue.  Make this code functional
1847  * first.
1848  */
1849 void
1850 knote(struct knlist *list, long hint, int lockflags)
1851 {
1852 	struct kqueue *kq;
1853 	struct knote *kn;
1854 	int error;
1855 
1856 	if (list == NULL)
1857 		return;
1858 
1859 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1860 
1861 	if ((lockflags & KNF_LISTLOCKED) == 0)
1862 		list->kl_lock(list->kl_lockarg);
1863 
1864 	/*
1865 	 * If we unlock the list lock (and set KN_INFLUX), we can eliminate
1866 	 * the kqueue scheduling, but this will introduce four
1867 	 * lock/unlock's for each knote to test.  If we do, continue to use
1868 	 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is
1869 	 * only safe if you want to remove the current item, which we are
1870 	 * not doing.
1871 	 */
1872 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
1873 		kq = kn->kn_kq;
1874 		KQ_LOCK(kq);
1875 		if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
1876 			/*
1877 			 * Do not process the influx notes, except for
1878 			 * the influx coming from the kq unlock in the
1879 			 * kqueue_scan().  In the later case, we do
1880 			 * not interfere with the scan, since the code
1881 			 * fragment in kqueue_scan() locks the knlist,
1882 			 * and cannot proceed until we finished.
1883 			 */
1884 			KQ_UNLOCK(kq);
1885 		} else if ((lockflags & KNF_NOKQLOCK) != 0) {
1886 			kn->kn_status |= KN_INFLUX;
1887 			KQ_UNLOCK(kq);
1888 			error = kn->kn_fop->f_event(kn, hint);
1889 			KQ_LOCK(kq);
1890 			kn->kn_status &= ~KN_INFLUX;
1891 			if (error)
1892 				KNOTE_ACTIVATE(kn, 1);
1893 			KQ_UNLOCK_FLUX(kq);
1894 		} else {
1895 			kn->kn_status |= KN_HASKQLOCK;
1896 			if (kn->kn_fop->f_event(kn, hint))
1897 				KNOTE_ACTIVATE(kn, 1);
1898 			kn->kn_status &= ~KN_HASKQLOCK;
1899 			KQ_UNLOCK(kq);
1900 		}
1901 	}
1902 	if ((lockflags & KNF_LISTLOCKED) == 0)
1903 		list->kl_unlock(list->kl_lockarg);
1904 }
1905 
1906 /*
1907  * add a knote to a knlist
1908  */
1909 void
1910 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1911 {
1912 	KNL_ASSERT_LOCK(knl, islocked);
1913 	KQ_NOTOWNED(kn->kn_kq);
1914 	KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1915 	    (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1916 	if (!islocked)
1917 		knl->kl_lock(knl->kl_lockarg);
1918 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1919 	if (!islocked)
1920 		knl->kl_unlock(knl->kl_lockarg);
1921 	KQ_LOCK(kn->kn_kq);
1922 	kn->kn_knlist = knl;
1923 	kn->kn_status &= ~KN_DETACHED;
1924 	KQ_UNLOCK(kn->kn_kq);
1925 }
1926 
1927 static void
1928 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1929 {
1930 	KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1931 	KNL_ASSERT_LOCK(knl, knlislocked);
1932 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1933 	if (!kqislocked)
1934 		KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1935     ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1936 	if (!knlislocked)
1937 		knl->kl_lock(knl->kl_lockarg);
1938 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1939 	kn->kn_knlist = NULL;
1940 	if (!knlislocked)
1941 		knl->kl_unlock(knl->kl_lockarg);
1942 	if (!kqislocked)
1943 		KQ_LOCK(kn->kn_kq);
1944 	kn->kn_status |= KN_DETACHED;
1945 	if (!kqislocked)
1946 		KQ_UNLOCK(kn->kn_kq);
1947 }
1948 
1949 /*
1950  * remove knote from the specified knlist
1951  */
1952 void
1953 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1954 {
1955 
1956 	knlist_remove_kq(knl, kn, islocked, 0);
1957 }
1958 
1959 /*
1960  * remove knote from the specified knlist while in f_event handler.
1961  */
1962 void
1963 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1964 {
1965 
1966 	knlist_remove_kq(knl, kn, 1,
1967 	    (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1968 }
1969 
1970 int
1971 knlist_empty(struct knlist *knl)
1972 {
1973 
1974 	KNL_ASSERT_LOCKED(knl);
1975 	return SLIST_EMPTY(&knl->kl_list);
1976 }
1977 
1978 static struct mtx	knlist_lock;
1979 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
1980 	MTX_DEF);
1981 static void knlist_mtx_lock(void *arg);
1982 static void knlist_mtx_unlock(void *arg);
1983 
1984 static void
1985 knlist_mtx_lock(void *arg)
1986 {
1987 
1988 	mtx_lock((struct mtx *)arg);
1989 }
1990 
1991 static void
1992 knlist_mtx_unlock(void *arg)
1993 {
1994 
1995 	mtx_unlock((struct mtx *)arg);
1996 }
1997 
1998 static void
1999 knlist_mtx_assert_locked(void *arg)
2000 {
2001 
2002 	mtx_assert((struct mtx *)arg, MA_OWNED);
2003 }
2004 
2005 static void
2006 knlist_mtx_assert_unlocked(void *arg)
2007 {
2008 
2009 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2010 }
2011 
2012 static void
2013 knlist_rw_rlock(void *arg)
2014 {
2015 
2016 	rw_rlock((struct rwlock *)arg);
2017 }
2018 
2019 static void
2020 knlist_rw_runlock(void *arg)
2021 {
2022 
2023 	rw_runlock((struct rwlock *)arg);
2024 }
2025 
2026 static void
2027 knlist_rw_assert_locked(void *arg)
2028 {
2029 
2030 	rw_assert((struct rwlock *)arg, RA_LOCKED);
2031 }
2032 
2033 static void
2034 knlist_rw_assert_unlocked(void *arg)
2035 {
2036 
2037 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2038 }
2039 
2040 void
2041 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2042     void (*kl_unlock)(void *),
2043     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2044 {
2045 
2046 	if (lock == NULL)
2047 		knl->kl_lockarg = &knlist_lock;
2048 	else
2049 		knl->kl_lockarg = lock;
2050 
2051 	if (kl_lock == NULL)
2052 		knl->kl_lock = knlist_mtx_lock;
2053 	else
2054 		knl->kl_lock = kl_lock;
2055 	if (kl_unlock == NULL)
2056 		knl->kl_unlock = knlist_mtx_unlock;
2057 	else
2058 		knl->kl_unlock = kl_unlock;
2059 	if (kl_assert_locked == NULL)
2060 		knl->kl_assert_locked = knlist_mtx_assert_locked;
2061 	else
2062 		knl->kl_assert_locked = kl_assert_locked;
2063 	if (kl_assert_unlocked == NULL)
2064 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2065 	else
2066 		knl->kl_assert_unlocked = kl_assert_unlocked;
2067 
2068 	SLIST_INIT(&knl->kl_list);
2069 }
2070 
2071 void
2072 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2073 {
2074 
2075 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2076 }
2077 
2078 void
2079 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2080 {
2081 
2082 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2083 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2084 }
2085 
2086 void
2087 knlist_destroy(struct knlist *knl)
2088 {
2089 
2090 #ifdef INVARIANTS
2091 	/*
2092 	 * if we run across this error, we need to find the offending
2093 	 * driver and have it call knlist_clear or knlist_delete.
2094 	 */
2095 	if (!SLIST_EMPTY(&knl->kl_list))
2096 		printf("WARNING: destroying knlist w/ knotes on it!\n");
2097 #endif
2098 
2099 	knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2100 	SLIST_INIT(&knl->kl_list);
2101 }
2102 
2103 /*
2104  * Even if we are locked, we may need to drop the lock to allow any influx
2105  * knotes time to "settle".
2106  */
2107 void
2108 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2109 {
2110 	struct knote *kn, *kn2;
2111 	struct kqueue *kq;
2112 
2113 	if (islocked)
2114 		KNL_ASSERT_LOCKED(knl);
2115 	else {
2116 		KNL_ASSERT_UNLOCKED(knl);
2117 again:		/* need to reacquire lock since we have dropped it */
2118 		knl->kl_lock(knl->kl_lockarg);
2119 	}
2120 
2121 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2122 		kq = kn->kn_kq;
2123 		KQ_LOCK(kq);
2124 		if ((kn->kn_status & KN_INFLUX)) {
2125 			KQ_UNLOCK(kq);
2126 			continue;
2127 		}
2128 		knlist_remove_kq(knl, kn, 1, 1);
2129 		if (killkn) {
2130 			kn->kn_status |= KN_INFLUX | KN_DETACHED;
2131 			KQ_UNLOCK(kq);
2132 			knote_drop(kn, td);
2133 		} else {
2134 			/* Make sure cleared knotes disappear soon */
2135 			kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2136 			KQ_UNLOCK(kq);
2137 		}
2138 		kq = NULL;
2139 	}
2140 
2141 	if (!SLIST_EMPTY(&knl->kl_list)) {
2142 		/* there are still KN_INFLUX remaining */
2143 		kn = SLIST_FIRST(&knl->kl_list);
2144 		kq = kn->kn_kq;
2145 		KQ_LOCK(kq);
2146 		KASSERT(kn->kn_status & KN_INFLUX,
2147 		    ("knote removed w/o list lock"));
2148 		knl->kl_unlock(knl->kl_lockarg);
2149 		kq->kq_state |= KQ_FLUXWAIT;
2150 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2151 		kq = NULL;
2152 		goto again;
2153 	}
2154 
2155 	if (islocked)
2156 		KNL_ASSERT_LOCKED(knl);
2157 	else {
2158 		knl->kl_unlock(knl->kl_lockarg);
2159 		KNL_ASSERT_UNLOCKED(knl);
2160 	}
2161 }
2162 
2163 /*
2164  * Remove all knotes referencing a specified fd must be called with FILEDESC
2165  * lock.  This prevents a race where a new fd comes along and occupies the
2166  * entry and we attach a knote to the fd.
2167  */
2168 void
2169 knote_fdclose(struct thread *td, int fd)
2170 {
2171 	struct filedesc *fdp = td->td_proc->p_fd;
2172 	struct kqueue *kq;
2173 	struct knote *kn;
2174 	int influx;
2175 
2176 	FILEDESC_XLOCK_ASSERT(fdp);
2177 
2178 	/*
2179 	 * We shouldn't have to worry about new kevents appearing on fd
2180 	 * since filedesc is locked.
2181 	 */
2182 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2183 		KQ_LOCK(kq);
2184 
2185 again:
2186 		influx = 0;
2187 		while (kq->kq_knlistsize > fd &&
2188 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2189 			if (kn->kn_status & KN_INFLUX) {
2190 				/* someone else might be waiting on our knote */
2191 				if (influx)
2192 					wakeup(kq);
2193 				kq->kq_state |= KQ_FLUXWAIT;
2194 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2195 				goto again;
2196 			}
2197 			kn->kn_status |= KN_INFLUX;
2198 			KQ_UNLOCK(kq);
2199 			if (!(kn->kn_status & KN_DETACHED))
2200 				kn->kn_fop->f_detach(kn);
2201 			knote_drop(kn, td);
2202 			influx = 1;
2203 			KQ_LOCK(kq);
2204 		}
2205 		KQ_UNLOCK_FLUX(kq);
2206 	}
2207 }
2208 
2209 static int
2210 knote_attach(struct knote *kn, struct kqueue *kq)
2211 {
2212 	struct klist *list;
2213 
2214 	KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2215 	KQ_OWNED(kq);
2216 
2217 	if (kn->kn_fop->f_isfd) {
2218 		if (kn->kn_id >= kq->kq_knlistsize)
2219 			return ENOMEM;
2220 		list = &kq->kq_knlist[kn->kn_id];
2221 	} else {
2222 		if (kq->kq_knhash == NULL)
2223 			return ENOMEM;
2224 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2225 	}
2226 
2227 	SLIST_INSERT_HEAD(list, kn, kn_link);
2228 
2229 	return 0;
2230 }
2231 
2232 /*
2233  * knote must already have been detached using the f_detach method.
2234  * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2235  * to prevent other removal.
2236  */
2237 static void
2238 knote_drop(struct knote *kn, struct thread *td)
2239 {
2240 	struct kqueue *kq;
2241 	struct klist *list;
2242 
2243 	kq = kn->kn_kq;
2244 
2245 	KQ_NOTOWNED(kq);
2246 	KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2247 	    ("knote_drop called without KN_INFLUX set in kn_status"));
2248 
2249 	KQ_LOCK(kq);
2250 	if (kn->kn_fop->f_isfd)
2251 		list = &kq->kq_knlist[kn->kn_id];
2252 	else
2253 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2254 
2255 	if (!SLIST_EMPTY(list))
2256 		SLIST_REMOVE(list, kn, knote, kn_link);
2257 	if (kn->kn_status & KN_QUEUED)
2258 		knote_dequeue(kn);
2259 	KQ_UNLOCK_FLUX(kq);
2260 
2261 	if (kn->kn_fop->f_isfd) {
2262 		fdrop(kn->kn_fp, td);
2263 		kn->kn_fp = NULL;
2264 	}
2265 	kqueue_fo_release(kn->kn_kevent.filter);
2266 	kn->kn_fop = NULL;
2267 	knote_free(kn);
2268 }
2269 
2270 static void
2271 knote_enqueue(struct knote *kn)
2272 {
2273 	struct kqueue *kq = kn->kn_kq;
2274 
2275 	KQ_OWNED(kn->kn_kq);
2276 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2277 
2278 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2279 	kn->kn_status |= KN_QUEUED;
2280 	kq->kq_count++;
2281 	kqueue_wakeup(kq);
2282 }
2283 
2284 static void
2285 knote_dequeue(struct knote *kn)
2286 {
2287 	struct kqueue *kq = kn->kn_kq;
2288 
2289 	KQ_OWNED(kn->kn_kq);
2290 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2291 
2292 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2293 	kn->kn_status &= ~KN_QUEUED;
2294 	kq->kq_count--;
2295 }
2296 
2297 static void
2298 knote_init(void)
2299 {
2300 
2301 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2302 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2303 }
2304 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2305 
2306 static struct knote *
2307 knote_alloc(int waitok)
2308 {
2309 	return ((struct knote *)uma_zalloc(knote_zone,
2310 	    (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2311 }
2312 
2313 static void
2314 knote_free(struct knote *kn)
2315 {
2316 	if (kn != NULL)
2317 		uma_zfree(knote_zone, kn);
2318 }
2319 
2320 /*
2321  * Register the kev w/ the kq specified by fd.
2322  */
2323 int
2324 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2325 {
2326 	struct kqueue *kq;
2327 	struct file *fp;
2328 	cap_rights_t rights;
2329 	int error;
2330 
2331 	error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2332 	if (error != 0)
2333 		return (error);
2334 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2335 		goto noacquire;
2336 
2337 	error = kqueue_register(kq, kev, td, waitok);
2338 
2339 	kqueue_release(kq, 0);
2340 
2341 noacquire:
2342 	fdrop(fp, td);
2343 
2344 	return error;
2345 }
2346