xref: /freebsd/sys/kern/kern_event.c (revision b2d48be1bc7df45ddd13b143a160d0acb5a383c5)
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 void	kqueue_destroy(struct kqueue *kq);
103 static void	kqueue_drain(struct kqueue *kq, struct thread *td);
104 static int	kqueue_expand(struct kqueue *kq, struct filterops *fops,
105 		    uintptr_t ident, int waitok);
106 static void	kqueue_task(void *arg, int pending);
107 static int	kqueue_scan(struct kqueue *kq, int maxevents,
108 		    struct kevent_copyops *k_ops,
109 		    const struct timespec *timeout,
110 		    struct kevent *keva, struct thread *td);
111 static void 	kqueue_wakeup(struct kqueue *kq);
112 static struct filterops *kqueue_fo_find(int filt);
113 static void	kqueue_fo_release(int filt);
114 
115 static fo_ioctl_t	kqueue_ioctl;
116 static fo_poll_t	kqueue_poll;
117 static fo_kqfilter_t	kqueue_kqfilter;
118 static fo_stat_t	kqueue_stat;
119 static fo_close_t	kqueue_close;
120 static fo_fill_kinfo_t	kqueue_fill_kinfo;
121 
122 static struct fileops kqueueops = {
123 	.fo_read = invfo_rdwr,
124 	.fo_write = invfo_rdwr,
125 	.fo_truncate = invfo_truncate,
126 	.fo_ioctl = kqueue_ioctl,
127 	.fo_poll = kqueue_poll,
128 	.fo_kqfilter = kqueue_kqfilter,
129 	.fo_stat = kqueue_stat,
130 	.fo_close = kqueue_close,
131 	.fo_chmod = invfo_chmod,
132 	.fo_chown = invfo_chown,
133 	.fo_sendfile = invfo_sendfile,
134 	.fo_fill_kinfo = kqueue_fill_kinfo,
135 };
136 
137 static int 	knote_attach(struct knote *kn, struct kqueue *kq);
138 static void 	knote_drop(struct knote *kn, struct thread *td);
139 static void 	knote_enqueue(struct knote *kn);
140 static void 	knote_dequeue(struct knote *kn);
141 static void 	knote_init(void);
142 static struct 	knote *knote_alloc(int waitok);
143 static void 	knote_free(struct knote *kn);
144 
145 static void	filt_kqdetach(struct knote *kn);
146 static int	filt_kqueue(struct knote *kn, long hint);
147 static int	filt_procattach(struct knote *kn);
148 static void	filt_procdetach(struct knote *kn);
149 static int	filt_proc(struct knote *kn, long hint);
150 static int	filt_fileattach(struct knote *kn);
151 static void	filt_timerexpire(void *knx);
152 static int	filt_timerattach(struct knote *kn);
153 static void	filt_timerdetach(struct knote *kn);
154 static int	filt_timer(struct knote *kn, long hint);
155 static int	filt_userattach(struct knote *kn);
156 static void	filt_userdetach(struct knote *kn);
157 static int	filt_user(struct knote *kn, long hint);
158 static void	filt_usertouch(struct knote *kn, struct kevent *kev,
159 		    u_long type);
160 
161 static struct filterops file_filtops = {
162 	.f_isfd = 1,
163 	.f_attach = filt_fileattach,
164 };
165 static struct filterops kqread_filtops = {
166 	.f_isfd = 1,
167 	.f_detach = filt_kqdetach,
168 	.f_event = filt_kqueue,
169 };
170 /* XXX - move to kern_proc.c?  */
171 static struct filterops proc_filtops = {
172 	.f_isfd = 0,
173 	.f_attach = filt_procattach,
174 	.f_detach = filt_procdetach,
175 	.f_event = filt_proc,
176 };
177 static struct filterops timer_filtops = {
178 	.f_isfd = 0,
179 	.f_attach = filt_timerattach,
180 	.f_detach = filt_timerdetach,
181 	.f_event = filt_timer,
182 };
183 static struct filterops user_filtops = {
184 	.f_attach = filt_userattach,
185 	.f_detach = filt_userdetach,
186 	.f_event = filt_user,
187 	.f_touch = filt_usertouch,
188 };
189 
190 static uma_zone_t	knote_zone;
191 static atomic_uint	kq_ncallouts = ATOMIC_VAR_INIT(0);
192 static unsigned int 	kq_calloutmax = 4 * 1024;
193 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
194     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
195 
196 /* XXX - ensure not KN_INFLUX?? */
197 #define KNOTE_ACTIVATE(kn, islock) do { 				\
198 	if ((islock))							\
199 		mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED);		\
200 	else								\
201 		KQ_LOCK((kn)->kn_kq);					\
202 	(kn)->kn_status |= KN_ACTIVE;					\
203 	if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)		\
204 		knote_enqueue((kn));					\
205 	if (!(islock))							\
206 		KQ_UNLOCK((kn)->kn_kq);					\
207 } while(0)
208 #define KQ_LOCK(kq) do {						\
209 	mtx_lock(&(kq)->kq_lock);					\
210 } while (0)
211 #define KQ_FLUX_WAKEUP(kq) do {						\
212 	if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) {		\
213 		(kq)->kq_state &= ~KQ_FLUXWAIT;				\
214 		wakeup((kq));						\
215 	}								\
216 } while (0)
217 #define KQ_UNLOCK_FLUX(kq) do {						\
218 	KQ_FLUX_WAKEUP(kq);						\
219 	mtx_unlock(&(kq)->kq_lock);					\
220 } while (0)
221 #define KQ_UNLOCK(kq) do {						\
222 	mtx_unlock(&(kq)->kq_lock);					\
223 } while (0)
224 #define KQ_OWNED(kq) do {						\
225 	mtx_assert(&(kq)->kq_lock, MA_OWNED);				\
226 } while (0)
227 #define KQ_NOTOWNED(kq) do {						\
228 	mtx_assert(&(kq)->kq_lock, MA_NOTOWNED);			\
229 } while (0)
230 #define KN_LIST_LOCK(kn) do {						\
231 	if (kn->kn_knlist != NULL)					\
232 		kn->kn_knlist->kl_lock(kn->kn_knlist->kl_lockarg);	\
233 } while (0)
234 #define KN_LIST_UNLOCK(kn) do {						\
235 	if (kn->kn_knlist != NULL) 					\
236 		kn->kn_knlist->kl_unlock(kn->kn_knlist->kl_lockarg);	\
237 } while (0)
238 #define	KNL_ASSERT_LOCK(knl, islocked) do {				\
239 	if (islocked)							\
240 		KNL_ASSERT_LOCKED(knl);				\
241 	else								\
242 		KNL_ASSERT_UNLOCKED(knl);				\
243 } while (0)
244 #ifdef INVARIANTS
245 #define	KNL_ASSERT_LOCKED(knl) do {					\
246 	knl->kl_assert_locked((knl)->kl_lockarg);			\
247 } while (0)
248 #define	KNL_ASSERT_UNLOCKED(knl) do {					\
249 	knl->kl_assert_unlocked((knl)->kl_lockarg);			\
250 } while (0)
251 #else /* !INVARIANTS */
252 #define	KNL_ASSERT_LOCKED(knl) do {} while(0)
253 #define	KNL_ASSERT_UNLOCKED(knl) do {} while (0)
254 #endif /* INVARIANTS */
255 
256 #ifndef	KN_HASHSIZE
257 #define	KN_HASHSIZE		64		/* XXX should be tunable */
258 #endif
259 
260 #define KN_HASH(val, mask)	(((val) ^ (val >> 8)) & (mask))
261 
262 static int
263 filt_nullattach(struct knote *kn)
264 {
265 
266 	return (ENXIO);
267 };
268 
269 struct filterops null_filtops = {
270 	.f_isfd = 0,
271 	.f_attach = filt_nullattach,
272 };
273 
274 /* XXX - make SYSINIT to add these, and move into respective modules. */
275 extern struct filterops sig_filtops;
276 extern struct filterops fs_filtops;
277 
278 /*
279  * Table for for all system-defined filters.
280  */
281 static struct mtx	filterops_lock;
282 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
283 	MTX_DEF);
284 static struct {
285 	struct filterops *for_fop;
286 	int for_nolock;
287 	int for_refcnt;
288 } sysfilt_ops[EVFILT_SYSCOUNT] = {
289 	{ &file_filtops, 1 },			/* EVFILT_READ */
290 	{ &file_filtops, 1 },			/* EVFILT_WRITE */
291 	{ &null_filtops },			/* EVFILT_AIO */
292 	{ &file_filtops, 1 },			/* EVFILT_VNODE */
293 	{ &proc_filtops, 1 },			/* EVFILT_PROC */
294 	{ &sig_filtops, 1 },			/* EVFILT_SIGNAL */
295 	{ &timer_filtops, 1 },			/* EVFILT_TIMER */
296 	{ &file_filtops, 1 },			/* EVFILT_PROCDESC */
297 	{ &fs_filtops, 1 },			/* EVFILT_FS */
298 	{ &null_filtops },			/* EVFILT_LIO */
299 	{ &user_filtops, 1 },			/* EVFILT_USER */
300 	{ &null_filtops },			/* EVFILT_SENDFILE */
301 };
302 
303 /*
304  * Simple redirection for all cdevsw style objects to call their fo_kqfilter
305  * method.
306  */
307 static int
308 filt_fileattach(struct knote *kn)
309 {
310 
311 	return (fo_kqfilter(kn->kn_fp, kn));
312 }
313 
314 /*ARGSUSED*/
315 static int
316 kqueue_kqfilter(struct file *fp, struct knote *kn)
317 {
318 	struct kqueue *kq = kn->kn_fp->f_data;
319 
320 	if (kn->kn_filter != EVFILT_READ)
321 		return (EINVAL);
322 
323 	kn->kn_status |= KN_KQUEUE;
324 	kn->kn_fop = &kqread_filtops;
325 	knlist_add(&kq->kq_sel.si_note, kn, 0);
326 
327 	return (0);
328 }
329 
330 static void
331 filt_kqdetach(struct knote *kn)
332 {
333 	struct kqueue *kq = kn->kn_fp->f_data;
334 
335 	knlist_remove(&kq->kq_sel.si_note, kn, 0);
336 }
337 
338 /*ARGSUSED*/
339 static int
340 filt_kqueue(struct knote *kn, long hint)
341 {
342 	struct kqueue *kq = kn->kn_fp->f_data;
343 
344 	kn->kn_data = kq->kq_count;
345 	return (kn->kn_data > 0);
346 }
347 
348 /* XXX - move to kern_proc.c?  */
349 static int
350 filt_procattach(struct knote *kn)
351 {
352 	struct proc *p;
353 	int immediate;
354 	int error;
355 
356 	immediate = 0;
357 	p = pfind(kn->kn_id);
358 	if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
359 		p = zpfind(kn->kn_id);
360 		immediate = 1;
361 	} else if (p != NULL && (p->p_flag & P_WEXIT)) {
362 		immediate = 1;
363 	}
364 
365 	if (p == NULL)
366 		return (ESRCH);
367 	if ((error = p_cansee(curthread, p))) {
368 		PROC_UNLOCK(p);
369 		return (error);
370 	}
371 
372 	kn->kn_ptr.p_proc = p;
373 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
374 
375 	/*
376 	 * internal flag indicating registration done by kernel
377 	 */
378 	if (kn->kn_flags & EV_FLAG1) {
379 		kn->kn_data = kn->kn_sdata;		/* ppid */
380 		kn->kn_fflags = NOTE_CHILD;
381 		kn->kn_flags &= ~EV_FLAG1;
382 	}
383 
384 	if (immediate == 0)
385 		knlist_add(&p->p_klist, kn, 1);
386 
387 	/*
388 	 * Immediately activate any exit notes if the target process is a
389 	 * zombie.  This is necessary to handle the case where the target
390 	 * process, e.g. a child, dies before the kevent is registered.
391 	 */
392 	if (immediate && filt_proc(kn, NOTE_EXIT))
393 		KNOTE_ACTIVATE(kn, 0);
394 
395 	PROC_UNLOCK(p);
396 
397 	return (0);
398 }
399 
400 /*
401  * The knote may be attached to a different process, which may exit,
402  * leaving nothing for the knote to be attached to.  So when the process
403  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
404  * it will be deleted when read out.  However, as part of the knote deletion,
405  * this routine is called, so a check is needed to avoid actually performing
406  * a detach, because the original process does not exist any more.
407  */
408 /* XXX - move to kern_proc.c?  */
409 static void
410 filt_procdetach(struct knote *kn)
411 {
412 	struct proc *p;
413 
414 	p = kn->kn_ptr.p_proc;
415 	knlist_remove(&p->p_klist, kn, 0);
416 	kn->kn_ptr.p_proc = NULL;
417 }
418 
419 /* XXX - move to kern_proc.c?  */
420 static int
421 filt_proc(struct knote *kn, long hint)
422 {
423 	struct proc *p;
424 	u_int event;
425 
426 	p = kn->kn_ptr.p_proc;
427 	/* Mask off extra data. */
428 	event = (u_int)hint & NOTE_PCTRLMASK;
429 
430 	/* If the user is interested in this event, record it. */
431 	if (kn->kn_sfflags & event)
432 		kn->kn_fflags |= event;
433 
434 	/* Process is gone, so flag the event as finished. */
435 	if (event == NOTE_EXIT) {
436 		if (!(kn->kn_status & KN_DETACHED))
437 			knlist_remove_inevent(&p->p_klist, kn);
438 		kn->kn_flags |= EV_EOF | EV_ONESHOT;
439 		kn->kn_ptr.p_proc = NULL;
440 		if (kn->kn_fflags & NOTE_EXIT)
441 			kn->kn_data = KW_EXITCODE(p->p_xexit, p->p_xsig);
442 		if (kn->kn_fflags == 0)
443 			kn->kn_flags |= EV_DROP;
444 		return (1);
445 	}
446 
447 	return (kn->kn_fflags != 0);
448 }
449 
450 /*
451  * Called when the process forked. It mostly does the same as the
452  * knote(), activating all knotes registered to be activated when the
453  * process forked. Additionally, for each knote attached to the
454  * parent, check whether user wants to track the new process. If so
455  * attach a new knote to it, and immediately report an event with the
456  * child's pid.
457  */
458 void
459 knote_fork(struct knlist *list, int pid)
460 {
461 	struct kqueue *kq;
462 	struct knote *kn;
463 	struct kevent kev;
464 	int error;
465 
466 	if (list == NULL)
467 		return;
468 	list->kl_lock(list->kl_lockarg);
469 
470 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
471 		/*
472 		 * XXX - Why do we skip the kn if it is _INFLUX?  Does this
473 		 * mean we will not properly wake up some notes?
474 		 */
475 		if ((kn->kn_status & KN_INFLUX) == KN_INFLUX)
476 			continue;
477 		kq = kn->kn_kq;
478 		KQ_LOCK(kq);
479 		if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
480 			KQ_UNLOCK(kq);
481 			continue;
482 		}
483 
484 		/*
485 		 * The same as knote(), activate the event.
486 		 */
487 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
488 			kn->kn_status |= KN_HASKQLOCK;
489 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
490 				KNOTE_ACTIVATE(kn, 1);
491 			kn->kn_status &= ~KN_HASKQLOCK;
492 			KQ_UNLOCK(kq);
493 			continue;
494 		}
495 
496 		/*
497 		 * The NOTE_TRACK case. In addition to the activation
498 		 * of the event, we need to register new event to
499 		 * track the child. Drop the locks in preparation for
500 		 * the call to kqueue_register().
501 		 */
502 		kn->kn_status |= KN_INFLUX;
503 		KQ_UNLOCK(kq);
504 		list->kl_unlock(list->kl_lockarg);
505 
506 		/*
507 		 * Activate existing knote and register a knote with
508 		 * new process.
509 		 */
510 		kev.ident = pid;
511 		kev.filter = kn->kn_filter;
512 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
513 		kev.fflags = kn->kn_sfflags;
514 		kev.data = kn->kn_id;		/* parent */
515 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
516 		error = kqueue_register(kq, &kev, NULL, 0);
517 		if (error)
518 			kn->kn_fflags |= NOTE_TRACKERR;
519 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
520 			KNOTE_ACTIVATE(kn, 0);
521 		KQ_LOCK(kq);
522 		kn->kn_status &= ~KN_INFLUX;
523 		KQ_UNLOCK_FLUX(kq);
524 		list->kl_lock(list->kl_lockarg);
525 	}
526 	list->kl_unlock(list->kl_lockarg);
527 }
528 
529 /*
530  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
531  * interval timer support code.
532  */
533 
534 #define NOTE_TIMER_PRECMASK	(NOTE_SECONDS|NOTE_MSECONDS|NOTE_USECONDS| \
535 				NOTE_NSECONDS)
536 
537 static __inline sbintime_t
538 timer2sbintime(intptr_t data, int flags)
539 {
540 	sbintime_t modifier;
541 
542 	switch (flags & NOTE_TIMER_PRECMASK) {
543 	case NOTE_SECONDS:
544 		modifier = SBT_1S;
545 		break;
546 	case NOTE_MSECONDS: /* FALLTHROUGH */
547 	case 0:
548 		modifier = SBT_1MS;
549 		break;
550 	case NOTE_USECONDS:
551 		modifier = SBT_1US;
552 		break;
553 	case NOTE_NSECONDS:
554 		modifier = SBT_1NS;
555 		break;
556 	default:
557 		return (-1);
558 	}
559 
560 #ifdef __LP64__
561 	if (data > SBT_MAX / modifier)
562 		return (SBT_MAX);
563 #endif
564 	return (modifier * data);
565 }
566 
567 static void
568 filt_timerexpire(void *knx)
569 {
570 	struct callout *calloutp;
571 	struct knote *kn;
572 
573 	kn = knx;
574 	kn->kn_data++;
575 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
576 
577 	if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
578 		calloutp = (struct callout *)kn->kn_hook;
579 		*kn->kn_ptr.p_nexttime += timer2sbintime(kn->kn_sdata,
580 		    kn->kn_sfflags);
581 		callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
582 		    filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
583 	}
584 }
585 
586 /*
587  * data contains amount of time to sleep
588  */
589 static int
590 filt_timerattach(struct knote *kn)
591 {
592 	struct callout *calloutp;
593 	sbintime_t to;
594 	unsigned int ncallouts;
595 
596 	if ((intptr_t)kn->kn_sdata < 0)
597 		return (EINVAL);
598 	if ((intptr_t)kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
599 		kn->kn_sdata = 1;
600 	/* Only precision unit are supported in flags so far */
601 	if (kn->kn_sfflags & ~NOTE_TIMER_PRECMASK)
602 		return (EINVAL);
603 
604 	to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
605 	if (to < 0)
606 		return (EINVAL);
607 
608 	ncallouts = atomic_load_explicit(&kq_ncallouts, memory_order_relaxed);
609 	do {
610 		if (ncallouts >= kq_calloutmax)
611 			return (ENOMEM);
612 	} while (!atomic_compare_exchange_weak_explicit(&kq_ncallouts,
613 	    &ncallouts, ncallouts + 1, memory_order_relaxed,
614 	    memory_order_relaxed));
615 
616 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
617 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
618 	kn->kn_ptr.p_nexttime = malloc(sizeof(sbintime_t), M_KQUEUE, M_WAITOK);
619 	calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
620 	callout_init(calloutp, 1);
621 	kn->kn_hook = calloutp;
622 	*kn->kn_ptr.p_nexttime = to + sbinuptime();
623 	callout_reset_sbt_on(calloutp, *kn->kn_ptr.p_nexttime, 0,
624 	    filt_timerexpire, kn, PCPU_GET(cpuid), C_ABSOLUTE);
625 
626 	return (0);
627 }
628 
629 static void
630 filt_timerdetach(struct knote *kn)
631 {
632 	struct callout *calloutp;
633 	unsigned int old;
634 
635 	calloutp = (struct callout *)kn->kn_hook;
636 	callout_drain(calloutp);
637 	free(calloutp, M_KQUEUE);
638 	free(kn->kn_ptr.p_nexttime, M_KQUEUE);
639 	old = atomic_fetch_sub_explicit(&kq_ncallouts, 1, memory_order_relaxed);
640 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
641 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
642 }
643 
644 static int
645 filt_timer(struct knote *kn, long hint)
646 {
647 
648 	return (kn->kn_data != 0);
649 }
650 
651 static int
652 filt_userattach(struct knote *kn)
653 {
654 
655 	/*
656 	 * EVFILT_USER knotes are not attached to anything in the kernel.
657 	 */
658 	kn->kn_hook = NULL;
659 	if (kn->kn_fflags & NOTE_TRIGGER)
660 		kn->kn_hookid = 1;
661 	else
662 		kn->kn_hookid = 0;
663 	return (0);
664 }
665 
666 static void
667 filt_userdetach(__unused struct knote *kn)
668 {
669 
670 	/*
671 	 * EVFILT_USER knotes are not attached to anything in the kernel.
672 	 */
673 }
674 
675 static int
676 filt_user(struct knote *kn, __unused long hint)
677 {
678 
679 	return (kn->kn_hookid);
680 }
681 
682 static void
683 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
684 {
685 	u_int ffctrl;
686 
687 	switch (type) {
688 	case EVENT_REGISTER:
689 		if (kev->fflags & NOTE_TRIGGER)
690 			kn->kn_hookid = 1;
691 
692 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
693 		kev->fflags &= NOTE_FFLAGSMASK;
694 		switch (ffctrl) {
695 		case NOTE_FFNOP:
696 			break;
697 
698 		case NOTE_FFAND:
699 			kn->kn_sfflags &= kev->fflags;
700 			break;
701 
702 		case NOTE_FFOR:
703 			kn->kn_sfflags |= kev->fflags;
704 			break;
705 
706 		case NOTE_FFCOPY:
707 			kn->kn_sfflags = kev->fflags;
708 			break;
709 
710 		default:
711 			/* XXX Return error? */
712 			break;
713 		}
714 		kn->kn_sdata = kev->data;
715 		if (kev->flags & EV_CLEAR) {
716 			kn->kn_hookid = 0;
717 			kn->kn_data = 0;
718 			kn->kn_fflags = 0;
719 		}
720 		break;
721 
722         case EVENT_PROCESS:
723 		*kev = kn->kn_kevent;
724 		kev->fflags = kn->kn_sfflags;
725 		kev->data = kn->kn_sdata;
726 		if (kn->kn_flags & EV_CLEAR) {
727 			kn->kn_hookid = 0;
728 			kn->kn_data = 0;
729 			kn->kn_fflags = 0;
730 		}
731 		break;
732 
733 	default:
734 		panic("filt_usertouch() - invalid type (%ld)", type);
735 		break;
736 	}
737 }
738 
739 int
740 sys_kqueue(struct thread *td, struct kqueue_args *uap)
741 {
742 
743 	return (kern_kqueue(td, 0, NULL));
744 }
745 
746 static void
747 kqueue_init(struct kqueue *kq)
748 {
749 
750 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
751 	TAILQ_INIT(&kq->kq_head);
752 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
753 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
754 }
755 
756 int
757 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
758 {
759 	struct filedesc *fdp;
760 	struct kqueue *kq;
761 	struct file *fp;
762 	struct ucred *cred;
763 	int fd, error;
764 
765 	fdp = td->td_proc->p_fd;
766 	cred = td->td_ucred;
767 	if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
768 		return (ENOMEM);
769 
770 	error = falloc_caps(td, &fp, &fd, flags, fcaps);
771 	if (error != 0) {
772 		chgkqcnt(cred->cr_ruidinfo, -1, 0);
773 		return (error);
774 	}
775 
776 	/* An extra reference on `fp' has been held for us by falloc(). */
777 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
778 	kqueue_init(kq);
779 	kq->kq_fdp = fdp;
780 	kq->kq_cred = crhold(cred);
781 
782 	FILEDESC_XLOCK(fdp);
783 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
784 	FILEDESC_XUNLOCK(fdp);
785 
786 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
787 	fdrop(fp, td);
788 
789 	td->td_retval[0] = fd;
790 	return (0);
791 }
792 
793 #ifndef _SYS_SYSPROTO_H_
794 struct kevent_args {
795 	int	fd;
796 	const struct kevent *changelist;
797 	int	nchanges;
798 	struct	kevent *eventlist;
799 	int	nevents;
800 	const struct timespec *timeout;
801 };
802 #endif
803 int
804 sys_kevent(struct thread *td, struct kevent_args *uap)
805 {
806 	struct timespec ts, *tsp;
807 	struct kevent_copyops k_ops = { uap,
808 					kevent_copyout,
809 					kevent_copyin};
810 	int error;
811 #ifdef KTRACE
812 	struct uio ktruio;
813 	struct iovec ktriov;
814 	struct uio *ktruioin = NULL;
815 	struct uio *ktruioout = NULL;
816 #endif
817 
818 	if (uap->timeout != NULL) {
819 		error = copyin(uap->timeout, &ts, sizeof(ts));
820 		if (error)
821 			return (error);
822 		tsp = &ts;
823 	} else
824 		tsp = NULL;
825 
826 #ifdef KTRACE
827 	if (KTRPOINT(td, KTR_GENIO)) {
828 		ktriov.iov_base = uap->changelist;
829 		ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
830 		ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
831 		    .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
832 		    .uio_td = td };
833 		ktruioin = cloneuio(&ktruio);
834 		ktriov.iov_base = uap->eventlist;
835 		ktriov.iov_len = uap->nevents * sizeof(struct kevent);
836 		ktruioout = cloneuio(&ktruio);
837 	}
838 #endif
839 
840 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
841 	    &k_ops, tsp);
842 
843 #ifdef KTRACE
844 	if (ktruioin != NULL) {
845 		ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
846 		ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
847 		ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
848 		ktrgenio(uap->fd, UIO_READ, ktruioout, error);
849 	}
850 #endif
851 
852 	return (error);
853 }
854 
855 /*
856  * Copy 'count' items into the destination list pointed to by uap->eventlist.
857  */
858 static int
859 kevent_copyout(void *arg, struct kevent *kevp, int count)
860 {
861 	struct kevent_args *uap;
862 	int error;
863 
864 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
865 	uap = (struct kevent_args *)arg;
866 
867 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
868 	if (error == 0)
869 		uap->eventlist += count;
870 	return (error);
871 }
872 
873 /*
874  * Copy 'count' items from the list pointed to by uap->changelist.
875  */
876 static int
877 kevent_copyin(void *arg, struct kevent *kevp, int count)
878 {
879 	struct kevent_args *uap;
880 	int error;
881 
882 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
883 	uap = (struct kevent_args *)arg;
884 
885 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
886 	if (error == 0)
887 		uap->changelist += count;
888 	return (error);
889 }
890 
891 int
892 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
893     struct kevent_copyops *k_ops, const struct timespec *timeout)
894 {
895 	cap_rights_t rights;
896 	struct file *fp;
897 	int error;
898 
899 	cap_rights_init(&rights);
900 	if (nchanges > 0)
901 		cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
902 	if (nevents > 0)
903 		cap_rights_set(&rights, CAP_KQUEUE_EVENT);
904 	error = fget(td, fd, &rights, &fp);
905 	if (error != 0)
906 		return (error);
907 
908 	error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
909 	fdrop(fp, td);
910 
911 	return (error);
912 }
913 
914 static int
915 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
916     struct kevent_copyops *k_ops, const struct timespec *timeout)
917 {
918 	struct kevent keva[KQ_NEVENTS];
919 	struct kevent *kevp, *changes;
920 	int i, n, nerrors, error;
921 
922 	nerrors = 0;
923 	while (nchanges > 0) {
924 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
925 		error = k_ops->k_copyin(k_ops->arg, keva, n);
926 		if (error)
927 			return (error);
928 		changes = keva;
929 		for (i = 0; i < n; i++) {
930 			kevp = &changes[i];
931 			if (!kevp->filter)
932 				continue;
933 			kevp->flags &= ~EV_SYSFLAGS;
934 			error = kqueue_register(kq, kevp, td, 1);
935 			if (error || (kevp->flags & EV_RECEIPT)) {
936 				if (nevents == 0)
937 					return (error);
938 				kevp->flags = EV_ERROR;
939 				kevp->data = error;
940 				(void)k_ops->k_copyout(k_ops->arg, kevp, 1);
941 				nevents--;
942 				nerrors++;
943 			}
944 		}
945 		nchanges -= n;
946 	}
947 	if (nerrors) {
948 		td->td_retval[0] = nerrors;
949 		return (0);
950 	}
951 
952 	return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
953 }
954 
955 int
956 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
957     struct kevent_copyops *k_ops, const struct timespec *timeout)
958 {
959 	struct kqueue *kq;
960 	int error;
961 
962 	error = kqueue_acquire(fp, &kq);
963 	if (error != 0)
964 		return (error);
965 	error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
966 	kqueue_release(kq, 0);
967 	return (error);
968 }
969 
970 /*
971  * Performs a kevent() call on a temporarily created kqueue. This can be
972  * used to perform one-shot polling, similar to poll() and select().
973  */
974 int
975 kern_kevent_anonymous(struct thread *td, int nevents,
976     struct kevent_copyops *k_ops)
977 {
978 	struct kqueue kq = {};
979 	int error;
980 
981 	kqueue_init(&kq);
982 	kq.kq_refcnt = 1;
983 	error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
984 	kqueue_drain(&kq, td);
985 	kqueue_destroy(&kq);
986 	return (error);
987 }
988 
989 int
990 kqueue_add_filteropts(int filt, struct filterops *filtops)
991 {
992 	int error;
993 
994 	error = 0;
995 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
996 		printf(
997 "trying to add a filterop that is out of range: %d is beyond %d\n",
998 		    ~filt, EVFILT_SYSCOUNT);
999 		return EINVAL;
1000 	}
1001 	mtx_lock(&filterops_lock);
1002 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1003 	    sysfilt_ops[~filt].for_fop != NULL)
1004 		error = EEXIST;
1005 	else {
1006 		sysfilt_ops[~filt].for_fop = filtops;
1007 		sysfilt_ops[~filt].for_refcnt = 0;
1008 	}
1009 	mtx_unlock(&filterops_lock);
1010 
1011 	return (error);
1012 }
1013 
1014 int
1015 kqueue_del_filteropts(int filt)
1016 {
1017 	int error;
1018 
1019 	error = 0;
1020 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1021 		return EINVAL;
1022 
1023 	mtx_lock(&filterops_lock);
1024 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1025 	    sysfilt_ops[~filt].for_fop == NULL)
1026 		error = EINVAL;
1027 	else if (sysfilt_ops[~filt].for_refcnt != 0)
1028 		error = EBUSY;
1029 	else {
1030 		sysfilt_ops[~filt].for_fop = &null_filtops;
1031 		sysfilt_ops[~filt].for_refcnt = 0;
1032 	}
1033 	mtx_unlock(&filterops_lock);
1034 
1035 	return error;
1036 }
1037 
1038 static struct filterops *
1039 kqueue_fo_find(int filt)
1040 {
1041 
1042 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1043 		return NULL;
1044 
1045 	if (sysfilt_ops[~filt].for_nolock)
1046 		return sysfilt_ops[~filt].for_fop;
1047 
1048 	mtx_lock(&filterops_lock);
1049 	sysfilt_ops[~filt].for_refcnt++;
1050 	if (sysfilt_ops[~filt].for_fop == NULL)
1051 		sysfilt_ops[~filt].for_fop = &null_filtops;
1052 	mtx_unlock(&filterops_lock);
1053 
1054 	return sysfilt_ops[~filt].for_fop;
1055 }
1056 
1057 static void
1058 kqueue_fo_release(int filt)
1059 {
1060 
1061 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1062 		return;
1063 
1064 	if (sysfilt_ops[~filt].for_nolock)
1065 		return;
1066 
1067 	mtx_lock(&filterops_lock);
1068 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1069 	    ("filter object refcount not valid on release"));
1070 	sysfilt_ops[~filt].for_refcnt--;
1071 	mtx_unlock(&filterops_lock);
1072 }
1073 
1074 /*
1075  * A ref to kq (obtained via kqueue_acquire) must be held.  waitok will
1076  * influence if memory allocation should wait.  Make sure it is 0 if you
1077  * hold any mutexes.
1078  */
1079 static int
1080 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1081 {
1082 	struct filterops *fops;
1083 	struct file *fp;
1084 	struct knote *kn, *tkn;
1085 	cap_rights_t rights;
1086 	int error, filt, event;
1087 	int haskqglobal, filedesc_unlock;
1088 
1089 	fp = NULL;
1090 	kn = NULL;
1091 	error = 0;
1092 	haskqglobal = 0;
1093 	filedesc_unlock = 0;
1094 
1095 	filt = kev->filter;
1096 	fops = kqueue_fo_find(filt);
1097 	if (fops == NULL)
1098 		return EINVAL;
1099 
1100 	if (kev->flags & EV_ADD) {
1101 		/*
1102 		 * Prevent waiting with locks.  Non-sleepable
1103 		 * allocation failures are handled in the loop, only
1104 		 * if the spare knote appears to be actually required.
1105 		 */
1106 		tkn = knote_alloc(waitok);
1107 	} else {
1108 		tkn = NULL;
1109 	}
1110 
1111 findkn:
1112 	if (fops->f_isfd) {
1113 		KASSERT(td != NULL, ("td is NULL"));
1114 		error = fget(td, kev->ident,
1115 		    cap_rights_init(&rights, CAP_EVENT), &fp);
1116 		if (error)
1117 			goto done;
1118 
1119 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1120 		    kev->ident, 0) != 0) {
1121 			/* try again */
1122 			fdrop(fp, td);
1123 			fp = NULL;
1124 			error = kqueue_expand(kq, fops, kev->ident, waitok);
1125 			if (error)
1126 				goto done;
1127 			goto findkn;
1128 		}
1129 
1130 		if (fp->f_type == DTYPE_KQUEUE) {
1131 			/*
1132 			 * if we add some inteligence about what we are doing,
1133 			 * we should be able to support events on ourselves.
1134 			 * We need to know when we are doing this to prevent
1135 			 * getting both the knlist lock and the kq lock since
1136 			 * they are the same thing.
1137 			 */
1138 			if (fp->f_data == kq) {
1139 				error = EINVAL;
1140 				goto done;
1141 			}
1142 
1143 			/*
1144 			 * Pre-lock the filedesc before the global
1145 			 * lock mutex, see the comment in
1146 			 * kqueue_close().
1147 			 */
1148 			FILEDESC_XLOCK(td->td_proc->p_fd);
1149 			filedesc_unlock = 1;
1150 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1151 		}
1152 
1153 		KQ_LOCK(kq);
1154 		if (kev->ident < kq->kq_knlistsize) {
1155 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1156 				if (kev->filter == kn->kn_filter)
1157 					break;
1158 		}
1159 	} else {
1160 		if ((kev->flags & EV_ADD) == EV_ADD)
1161 			kqueue_expand(kq, fops, kev->ident, waitok);
1162 
1163 		KQ_LOCK(kq);
1164 		if (kq->kq_knhashmask != 0) {
1165 			struct klist *list;
1166 
1167 			list = &kq->kq_knhash[
1168 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1169 			SLIST_FOREACH(kn, list, kn_link)
1170 				if (kev->ident == kn->kn_id &&
1171 				    kev->filter == kn->kn_filter)
1172 					break;
1173 		}
1174 	}
1175 
1176 	/* knote is in the process of changing, wait for it to stablize. */
1177 	if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1178 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1179 		if (filedesc_unlock) {
1180 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1181 			filedesc_unlock = 0;
1182 		}
1183 		kq->kq_state |= KQ_FLUXWAIT;
1184 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1185 		if (fp != NULL) {
1186 			fdrop(fp, td);
1187 			fp = NULL;
1188 		}
1189 		goto findkn;
1190 	}
1191 
1192 	/*
1193 	 * kn now contains the matching knote, or NULL if no match
1194 	 */
1195 	if (kn == NULL) {
1196 		if (kev->flags & EV_ADD) {
1197 			kn = tkn;
1198 			tkn = NULL;
1199 			if (kn == NULL) {
1200 				KQ_UNLOCK(kq);
1201 				error = ENOMEM;
1202 				goto done;
1203 			}
1204 			kn->kn_fp = fp;
1205 			kn->kn_kq = kq;
1206 			kn->kn_fop = fops;
1207 			/*
1208 			 * apply reference counts to knote structure, and
1209 			 * do not release it at the end of this routine.
1210 			 */
1211 			fops = NULL;
1212 			fp = NULL;
1213 
1214 			kn->kn_sfflags = kev->fflags;
1215 			kn->kn_sdata = kev->data;
1216 			kev->fflags = 0;
1217 			kev->data = 0;
1218 			kn->kn_kevent = *kev;
1219 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1220 			    EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1221 			kn->kn_status = KN_INFLUX|KN_DETACHED;
1222 
1223 			error = knote_attach(kn, kq);
1224 			KQ_UNLOCK(kq);
1225 			if (error != 0) {
1226 				tkn = kn;
1227 				goto done;
1228 			}
1229 
1230 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1231 				knote_drop(kn, td);
1232 				goto done;
1233 			}
1234 			KN_LIST_LOCK(kn);
1235 			goto done_ev_add;
1236 		} else {
1237 			/* No matching knote and the EV_ADD flag is not set. */
1238 			KQ_UNLOCK(kq);
1239 			error = ENOENT;
1240 			goto done;
1241 		}
1242 	}
1243 
1244 	if (kev->flags & EV_DELETE) {
1245 		kn->kn_status |= KN_INFLUX;
1246 		KQ_UNLOCK(kq);
1247 		if (!(kn->kn_status & KN_DETACHED))
1248 			kn->kn_fop->f_detach(kn);
1249 		knote_drop(kn, td);
1250 		goto done;
1251 	}
1252 
1253 	if (kev->flags & EV_FORCEONESHOT) {
1254 		kn->kn_flags |= EV_ONESHOT;
1255 		KNOTE_ACTIVATE(kn, 1);
1256 	}
1257 
1258 	/*
1259 	 * The user may change some filter values after the initial EV_ADD,
1260 	 * but doing so will not reset any filter which has already been
1261 	 * triggered.
1262 	 */
1263 	kn->kn_status |= KN_INFLUX | KN_SCAN;
1264 	KQ_UNLOCK(kq);
1265 	KN_LIST_LOCK(kn);
1266 	kn->kn_kevent.udata = kev->udata;
1267 	if (!fops->f_isfd && fops->f_touch != NULL) {
1268 		fops->f_touch(kn, kev, EVENT_REGISTER);
1269 	} else {
1270 		kn->kn_sfflags = kev->fflags;
1271 		kn->kn_sdata = kev->data;
1272 	}
1273 
1274 	/*
1275 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1276 	 * the initial attach event decides that the event is "completed"
1277 	 * already.  i.e. filt_procattach is called on a zombie process.  It
1278 	 * will call filt_proc which will remove it from the list, and NULL
1279 	 * kn_knlist.
1280 	 */
1281 done_ev_add:
1282 	if ((kev->flags & EV_DISABLE) &&
1283 	    ((kn->kn_status & KN_DISABLED) == 0)) {
1284 		kn->kn_status |= KN_DISABLED;
1285 	}
1286 
1287 	if ((kn->kn_status & KN_DISABLED) == 0)
1288 		event = kn->kn_fop->f_event(kn, 0);
1289 	else
1290 		event = 0;
1291 	KQ_LOCK(kq);
1292 	if (event)
1293 		KNOTE_ACTIVATE(kn, 1);
1294 	kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1295 	KN_LIST_UNLOCK(kn);
1296 
1297 	if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1298 		kn->kn_status &= ~KN_DISABLED;
1299 		if ((kn->kn_status & KN_ACTIVE) &&
1300 		    ((kn->kn_status & KN_QUEUED) == 0))
1301 			knote_enqueue(kn);
1302 	}
1303 	KQ_UNLOCK_FLUX(kq);
1304 
1305 done:
1306 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1307 	if (filedesc_unlock)
1308 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1309 	if (fp != NULL)
1310 		fdrop(fp, td);
1311 	knote_free(tkn);
1312 	if (fops != NULL)
1313 		kqueue_fo_release(filt);
1314 	return (error);
1315 }
1316 
1317 static int
1318 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1319 {
1320 	int error;
1321 	struct kqueue *kq;
1322 
1323 	error = 0;
1324 
1325 	kq = fp->f_data;
1326 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1327 		return (EBADF);
1328 	*kqp = kq;
1329 	KQ_LOCK(kq);
1330 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1331 		KQ_UNLOCK(kq);
1332 		return (EBADF);
1333 	}
1334 	kq->kq_refcnt++;
1335 	KQ_UNLOCK(kq);
1336 
1337 	return error;
1338 }
1339 
1340 static void
1341 kqueue_release(struct kqueue *kq, int locked)
1342 {
1343 	if (locked)
1344 		KQ_OWNED(kq);
1345 	else
1346 		KQ_LOCK(kq);
1347 	kq->kq_refcnt--;
1348 	if (kq->kq_refcnt == 1)
1349 		wakeup(&kq->kq_refcnt);
1350 	if (!locked)
1351 		KQ_UNLOCK(kq);
1352 }
1353 
1354 static void
1355 kqueue_schedtask(struct kqueue *kq)
1356 {
1357 
1358 	KQ_OWNED(kq);
1359 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1360 	    ("scheduling kqueue task while draining"));
1361 
1362 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1363 		taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1364 		kq->kq_state |= KQ_TASKSCHED;
1365 	}
1366 }
1367 
1368 /*
1369  * Expand the kq to make sure we have storage for fops/ident pair.
1370  *
1371  * Return 0 on success (or no work necessary), return errno on failure.
1372  *
1373  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1374  * If kqueue_register is called from a non-fd context, there usually/should
1375  * be no locks held.
1376  */
1377 static int
1378 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1379 	int waitok)
1380 {
1381 	struct klist *list, *tmp_knhash, *to_free;
1382 	u_long tmp_knhashmask;
1383 	int size;
1384 	int fd;
1385 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
1386 
1387 	KQ_NOTOWNED(kq);
1388 
1389 	to_free = NULL;
1390 	if (fops->f_isfd) {
1391 		fd = ident;
1392 		if (kq->kq_knlistsize <= fd) {
1393 			size = kq->kq_knlistsize;
1394 			while (size <= fd)
1395 				size += KQEXTENT;
1396 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1397 			if (list == NULL)
1398 				return ENOMEM;
1399 			KQ_LOCK(kq);
1400 			if (kq->kq_knlistsize > fd) {
1401 				to_free = list;
1402 				list = NULL;
1403 			} else {
1404 				if (kq->kq_knlist != NULL) {
1405 					bcopy(kq->kq_knlist, list,
1406 					    kq->kq_knlistsize * sizeof(*list));
1407 					to_free = kq->kq_knlist;
1408 					kq->kq_knlist = NULL;
1409 				}
1410 				bzero((caddr_t)list +
1411 				    kq->kq_knlistsize * sizeof(*list),
1412 				    (size - kq->kq_knlistsize) * sizeof(*list));
1413 				kq->kq_knlistsize = size;
1414 				kq->kq_knlist = list;
1415 			}
1416 			KQ_UNLOCK(kq);
1417 		}
1418 	} else {
1419 		if (kq->kq_knhashmask == 0) {
1420 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1421 			    &tmp_knhashmask);
1422 			if (tmp_knhash == NULL)
1423 				return ENOMEM;
1424 			KQ_LOCK(kq);
1425 			if (kq->kq_knhashmask == 0) {
1426 				kq->kq_knhash = tmp_knhash;
1427 				kq->kq_knhashmask = tmp_knhashmask;
1428 			} else {
1429 				to_free = tmp_knhash;
1430 			}
1431 			KQ_UNLOCK(kq);
1432 		}
1433 	}
1434 	free(to_free, M_KQUEUE);
1435 
1436 	KQ_NOTOWNED(kq);
1437 	return 0;
1438 }
1439 
1440 static void
1441 kqueue_task(void *arg, int pending)
1442 {
1443 	struct kqueue *kq;
1444 	int haskqglobal;
1445 
1446 	haskqglobal = 0;
1447 	kq = arg;
1448 
1449 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1450 	KQ_LOCK(kq);
1451 
1452 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1453 
1454 	kq->kq_state &= ~KQ_TASKSCHED;
1455 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1456 		wakeup(&kq->kq_state);
1457 	}
1458 	KQ_UNLOCK(kq);
1459 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1460 }
1461 
1462 /*
1463  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1464  * We treat KN_MARKER knotes as if they are INFLUX.
1465  */
1466 static int
1467 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1468     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1469 {
1470 	struct kevent *kevp;
1471 	struct knote *kn, *marker;
1472 	sbintime_t asbt, rsbt;
1473 	int count, error, haskqglobal, influx, nkev, touch;
1474 
1475 	count = maxevents;
1476 	nkev = 0;
1477 	error = 0;
1478 	haskqglobal = 0;
1479 
1480 	if (maxevents == 0)
1481 		goto done_nl;
1482 
1483 	rsbt = 0;
1484 	if (tsp != NULL) {
1485 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1486 		    tsp->tv_nsec >= 1000000000) {
1487 			error = EINVAL;
1488 			goto done_nl;
1489 		}
1490 		if (timespecisset(tsp)) {
1491 			if (tsp->tv_sec <= INT32_MAX) {
1492 				rsbt = tstosbt(*tsp);
1493 				if (TIMESEL(&asbt, rsbt))
1494 					asbt += tc_tick_sbt;
1495 				if (asbt <= SBT_MAX - rsbt)
1496 					asbt += rsbt;
1497 				else
1498 					asbt = 0;
1499 				rsbt >>= tc_precexp;
1500 			} else
1501 				asbt = 0;
1502 		} else
1503 			asbt = -1;
1504 	} else
1505 		asbt = 0;
1506 	marker = knote_alloc(1);
1507 	marker->kn_status = KN_MARKER;
1508 	KQ_LOCK(kq);
1509 
1510 retry:
1511 	kevp = keva;
1512 	if (kq->kq_count == 0) {
1513 		if (asbt == -1) {
1514 			error = EWOULDBLOCK;
1515 		} else {
1516 			kq->kq_state |= KQ_SLEEP;
1517 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1518 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1519 		}
1520 		if (error == 0)
1521 			goto retry;
1522 		/* don't restart after signals... */
1523 		if (error == ERESTART)
1524 			error = EINTR;
1525 		else if (error == EWOULDBLOCK)
1526 			error = 0;
1527 		goto done;
1528 	}
1529 
1530 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1531 	influx = 0;
1532 	while (count) {
1533 		KQ_OWNED(kq);
1534 		kn = TAILQ_FIRST(&kq->kq_head);
1535 
1536 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1537 		    (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1538 			if (influx) {
1539 				influx = 0;
1540 				KQ_FLUX_WAKEUP(kq);
1541 			}
1542 			kq->kq_state |= KQ_FLUXWAIT;
1543 			error = msleep(kq, &kq->kq_lock, PSOCK,
1544 			    "kqflxwt", 0);
1545 			continue;
1546 		}
1547 
1548 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1549 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1550 			kn->kn_status &= ~KN_QUEUED;
1551 			kq->kq_count--;
1552 			continue;
1553 		}
1554 		if (kn == marker) {
1555 			KQ_FLUX_WAKEUP(kq);
1556 			if (count == maxevents)
1557 				goto retry;
1558 			goto done;
1559 		}
1560 		KASSERT((kn->kn_status & KN_INFLUX) == 0,
1561 		    ("KN_INFLUX set when not suppose to be"));
1562 
1563 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1564 			kn->kn_status &= ~KN_QUEUED;
1565 			kn->kn_status |= KN_INFLUX;
1566 			kq->kq_count--;
1567 			KQ_UNLOCK(kq);
1568 			/*
1569 			 * We don't need to lock the list since we've marked
1570 			 * it _INFLUX.
1571 			 */
1572 			if (!(kn->kn_status & KN_DETACHED))
1573 				kn->kn_fop->f_detach(kn);
1574 			knote_drop(kn, td);
1575 			KQ_LOCK(kq);
1576 			continue;
1577 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1578 			kn->kn_status &= ~KN_QUEUED;
1579 			kn->kn_status |= KN_INFLUX;
1580 			kq->kq_count--;
1581 			KQ_UNLOCK(kq);
1582 			/*
1583 			 * We don't need to lock the list since we've marked
1584 			 * it _INFLUX.
1585 			 */
1586 			*kevp = kn->kn_kevent;
1587 			if (!(kn->kn_status & KN_DETACHED))
1588 				kn->kn_fop->f_detach(kn);
1589 			knote_drop(kn, td);
1590 			KQ_LOCK(kq);
1591 			kn = NULL;
1592 		} else {
1593 			kn->kn_status |= KN_INFLUX | KN_SCAN;
1594 			KQ_UNLOCK(kq);
1595 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1596 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1597 			KN_LIST_LOCK(kn);
1598 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1599 				KQ_LOCK(kq);
1600 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1601 				kn->kn_status &=
1602 				    ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX |
1603 				    KN_SCAN);
1604 				kq->kq_count--;
1605 				KN_LIST_UNLOCK(kn);
1606 				influx = 1;
1607 				continue;
1608 			}
1609 			touch = (!kn->kn_fop->f_isfd &&
1610 			    kn->kn_fop->f_touch != NULL);
1611 			if (touch)
1612 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1613 			else
1614 				*kevp = kn->kn_kevent;
1615 			KQ_LOCK(kq);
1616 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1617 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1618 				/*
1619 				 * Manually clear knotes who weren't
1620 				 * 'touch'ed.
1621 				 */
1622 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1623 					kn->kn_data = 0;
1624 					kn->kn_fflags = 0;
1625 				}
1626 				if (kn->kn_flags & EV_DISPATCH)
1627 					kn->kn_status |= KN_DISABLED;
1628 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1629 				kq->kq_count--;
1630 			} else
1631 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1632 
1633 			kn->kn_status &= ~(KN_INFLUX | KN_SCAN);
1634 			KN_LIST_UNLOCK(kn);
1635 			influx = 1;
1636 		}
1637 
1638 		/* we are returning a copy to the user */
1639 		kevp++;
1640 		nkev++;
1641 		count--;
1642 
1643 		if (nkev == KQ_NEVENTS) {
1644 			influx = 0;
1645 			KQ_UNLOCK_FLUX(kq);
1646 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1647 			nkev = 0;
1648 			kevp = keva;
1649 			KQ_LOCK(kq);
1650 			if (error)
1651 				break;
1652 		}
1653 	}
1654 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1655 done:
1656 	KQ_OWNED(kq);
1657 	KQ_UNLOCK_FLUX(kq);
1658 	knote_free(marker);
1659 done_nl:
1660 	KQ_NOTOWNED(kq);
1661 	if (nkev != 0)
1662 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1663 	td->td_retval[0] = maxevents - count;
1664 	return (error);
1665 }
1666 
1667 /*ARGSUSED*/
1668 static int
1669 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1670 	struct ucred *active_cred, struct thread *td)
1671 {
1672 	/*
1673 	 * Enabling sigio causes two major problems:
1674 	 * 1) infinite recursion:
1675 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1676 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1677 	 * into itself over and over.  Sending the sigio causes the kqueue
1678 	 * to become ready, which in turn posts sigio again, forever.
1679 	 * Solution: this can be solved by setting a flag in the kqueue that
1680 	 * we have a SIGIO in progress.
1681 	 * 2) locking problems:
1682 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1683 	 * us above the proc and pgrp locks.
1684 	 * Solution: Post a signal using an async mechanism, being sure to
1685 	 * record a generation count in the delivery so that we do not deliver
1686 	 * a signal to the wrong process.
1687 	 *
1688 	 * Note, these two mechanisms are somewhat mutually exclusive!
1689 	 */
1690 #if 0
1691 	struct kqueue *kq;
1692 
1693 	kq = fp->f_data;
1694 	switch (cmd) {
1695 	case FIOASYNC:
1696 		if (*(int *)data) {
1697 			kq->kq_state |= KQ_ASYNC;
1698 		} else {
1699 			kq->kq_state &= ~KQ_ASYNC;
1700 		}
1701 		return (0);
1702 
1703 	case FIOSETOWN:
1704 		return (fsetown(*(int *)data, &kq->kq_sigio));
1705 
1706 	case FIOGETOWN:
1707 		*(int *)data = fgetown(&kq->kq_sigio);
1708 		return (0);
1709 	}
1710 #endif
1711 
1712 	return (ENOTTY);
1713 }
1714 
1715 /*ARGSUSED*/
1716 static int
1717 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1718 	struct thread *td)
1719 {
1720 	struct kqueue *kq;
1721 	int revents = 0;
1722 	int error;
1723 
1724 	if ((error = kqueue_acquire(fp, &kq)))
1725 		return POLLERR;
1726 
1727 	KQ_LOCK(kq);
1728 	if (events & (POLLIN | POLLRDNORM)) {
1729 		if (kq->kq_count) {
1730 			revents |= events & (POLLIN | POLLRDNORM);
1731 		} else {
1732 			selrecord(td, &kq->kq_sel);
1733 			if (SEL_WAITING(&kq->kq_sel))
1734 				kq->kq_state |= KQ_SEL;
1735 		}
1736 	}
1737 	kqueue_release(kq, 1);
1738 	KQ_UNLOCK(kq);
1739 	return (revents);
1740 }
1741 
1742 /*ARGSUSED*/
1743 static int
1744 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1745 	struct thread *td)
1746 {
1747 
1748 	bzero((void *)st, sizeof *st);
1749 	/*
1750 	 * We no longer return kq_count because the unlocked value is useless.
1751 	 * If you spent all this time getting the count, why not spend your
1752 	 * syscall better by calling kevent?
1753 	 *
1754 	 * XXX - This is needed for libc_r.
1755 	 */
1756 	st->st_mode = S_IFIFO;
1757 	return (0);
1758 }
1759 
1760 static void
1761 kqueue_drain(struct kqueue *kq, struct thread *td)
1762 {
1763 	struct knote *kn;
1764 	int i;
1765 
1766 	KQ_LOCK(kq);
1767 
1768 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1769 	    ("kqueue already closing"));
1770 	kq->kq_state |= KQ_CLOSING;
1771 	if (kq->kq_refcnt > 1)
1772 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1773 
1774 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1775 
1776 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
1777 	    ("kqueue's knlist not empty"));
1778 
1779 	for (i = 0; i < kq->kq_knlistsize; i++) {
1780 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1781 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1782 				kq->kq_state |= KQ_FLUXWAIT;
1783 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1784 				continue;
1785 			}
1786 			kn->kn_status |= KN_INFLUX;
1787 			KQ_UNLOCK(kq);
1788 			if (!(kn->kn_status & KN_DETACHED))
1789 				kn->kn_fop->f_detach(kn);
1790 			knote_drop(kn, td);
1791 			KQ_LOCK(kq);
1792 		}
1793 	}
1794 	if (kq->kq_knhashmask != 0) {
1795 		for (i = 0; i <= kq->kq_knhashmask; i++) {
1796 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1797 				if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1798 					kq->kq_state |= KQ_FLUXWAIT;
1799 					msleep(kq, &kq->kq_lock, PSOCK,
1800 					       "kqclo2", 0);
1801 					continue;
1802 				}
1803 				kn->kn_status |= KN_INFLUX;
1804 				KQ_UNLOCK(kq);
1805 				if (!(kn->kn_status & KN_DETACHED))
1806 					kn->kn_fop->f_detach(kn);
1807 				knote_drop(kn, td);
1808 				KQ_LOCK(kq);
1809 			}
1810 		}
1811 	}
1812 
1813 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1814 		kq->kq_state |= KQ_TASKDRAIN;
1815 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1816 	}
1817 
1818 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1819 		selwakeuppri(&kq->kq_sel, PSOCK);
1820 		if (!SEL_WAITING(&kq->kq_sel))
1821 			kq->kq_state &= ~KQ_SEL;
1822 	}
1823 
1824 	KQ_UNLOCK(kq);
1825 }
1826 
1827 static void
1828 kqueue_destroy(struct kqueue *kq)
1829 {
1830 
1831 	KASSERT(kq->kq_fdp == NULL,
1832 	    ("kqueue still attached to a file descriptor"));
1833 	seldrain(&kq->kq_sel);
1834 	knlist_destroy(&kq->kq_sel.si_note);
1835 	mtx_destroy(&kq->kq_lock);
1836 
1837 	if (kq->kq_knhash != NULL)
1838 		free(kq->kq_knhash, M_KQUEUE);
1839 	if (kq->kq_knlist != NULL)
1840 		free(kq->kq_knlist, M_KQUEUE);
1841 
1842 	funsetown(&kq->kq_sigio);
1843 }
1844 
1845 /*ARGSUSED*/
1846 static int
1847 kqueue_close(struct file *fp, struct thread *td)
1848 {
1849 	struct kqueue *kq = fp->f_data;
1850 	struct filedesc *fdp;
1851 	int error;
1852 	int filedesc_unlock;
1853 
1854 	if ((error = kqueue_acquire(fp, &kq)))
1855 		return error;
1856 	kqueue_drain(kq, td);
1857 
1858 	/*
1859 	 * We could be called due to the knote_drop() doing fdrop(),
1860 	 * called from kqueue_register().  In this case the global
1861 	 * lock is owned, and filedesc sx is locked before, to not
1862 	 * take the sleepable lock after non-sleepable.
1863 	 */
1864 	fdp = kq->kq_fdp;
1865 	kq->kq_fdp = NULL;
1866 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1867 		FILEDESC_XLOCK(fdp);
1868 		filedesc_unlock = 1;
1869 	} else
1870 		filedesc_unlock = 0;
1871 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1872 	if (filedesc_unlock)
1873 		FILEDESC_XUNLOCK(fdp);
1874 
1875 	kqueue_destroy(kq);
1876 	chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
1877 	crfree(kq->kq_cred);
1878 	free(kq, M_KQUEUE);
1879 	fp->f_data = NULL;
1880 
1881 	return (0);
1882 }
1883 
1884 static int
1885 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1886 {
1887 
1888 	kif->kf_type = KF_TYPE_KQUEUE;
1889 	return (0);
1890 }
1891 
1892 static void
1893 kqueue_wakeup(struct kqueue *kq)
1894 {
1895 	KQ_OWNED(kq);
1896 
1897 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1898 		kq->kq_state &= ~KQ_SLEEP;
1899 		wakeup(kq);
1900 	}
1901 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1902 		selwakeuppri(&kq->kq_sel, PSOCK);
1903 		if (!SEL_WAITING(&kq->kq_sel))
1904 			kq->kq_state &= ~KQ_SEL;
1905 	}
1906 	if (!knlist_empty(&kq->kq_sel.si_note))
1907 		kqueue_schedtask(kq);
1908 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1909 		pgsigio(&kq->kq_sigio, SIGIO, 0);
1910 	}
1911 }
1912 
1913 /*
1914  * Walk down a list of knotes, activating them if their event has triggered.
1915  *
1916  * There is a possibility to optimize in the case of one kq watching another.
1917  * Instead of scheduling a task to wake it up, you could pass enough state
1918  * down the chain to make up the parent kqueue.  Make this code functional
1919  * first.
1920  */
1921 void
1922 knote(struct knlist *list, long hint, int lockflags)
1923 {
1924 	struct kqueue *kq;
1925 	struct knote *kn, *tkn;
1926 	int error;
1927 
1928 	if (list == NULL)
1929 		return;
1930 
1931 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1932 
1933 	if ((lockflags & KNF_LISTLOCKED) == 0)
1934 		list->kl_lock(list->kl_lockarg);
1935 
1936 	/*
1937 	 * If we unlock the list lock (and set KN_INFLUX), we can
1938 	 * eliminate the kqueue scheduling, but this will introduce
1939 	 * four lock/unlock's for each knote to test.  Also, marker
1940 	 * would be needed to keep iteration position, since filters
1941 	 * or other threads could remove events.
1942 	 */
1943 	SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
1944 		kq = kn->kn_kq;
1945 		KQ_LOCK(kq);
1946 		if ((kn->kn_status & (KN_INFLUX | KN_SCAN)) == KN_INFLUX) {
1947 			/*
1948 			 * Do not process the influx notes, except for
1949 			 * the influx coming from the kq unlock in the
1950 			 * kqueue_scan().  In the later case, we do
1951 			 * not interfere with the scan, since the code
1952 			 * fragment in kqueue_scan() locks the knlist,
1953 			 * and cannot proceed until we finished.
1954 			 */
1955 			KQ_UNLOCK(kq);
1956 		} else if ((lockflags & KNF_NOKQLOCK) != 0) {
1957 			kn->kn_status |= KN_INFLUX;
1958 			KQ_UNLOCK(kq);
1959 			error = kn->kn_fop->f_event(kn, hint);
1960 			KQ_LOCK(kq);
1961 			kn->kn_status &= ~KN_INFLUX;
1962 			if (error)
1963 				KNOTE_ACTIVATE(kn, 1);
1964 			KQ_UNLOCK_FLUX(kq);
1965 		} else {
1966 			kn->kn_status |= KN_HASKQLOCK;
1967 			if (kn->kn_fop->f_event(kn, hint))
1968 				KNOTE_ACTIVATE(kn, 1);
1969 			kn->kn_status &= ~KN_HASKQLOCK;
1970 			KQ_UNLOCK(kq);
1971 		}
1972 	}
1973 	if ((lockflags & KNF_LISTLOCKED) == 0)
1974 		list->kl_unlock(list->kl_lockarg);
1975 }
1976 
1977 /*
1978  * add a knote to a knlist
1979  */
1980 void
1981 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1982 {
1983 	KNL_ASSERT_LOCK(knl, islocked);
1984 	KQ_NOTOWNED(kn->kn_kq);
1985 	KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1986 	    (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1987 	if (!islocked)
1988 		knl->kl_lock(knl->kl_lockarg);
1989 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1990 	if (!islocked)
1991 		knl->kl_unlock(knl->kl_lockarg);
1992 	KQ_LOCK(kn->kn_kq);
1993 	kn->kn_knlist = knl;
1994 	kn->kn_status &= ~KN_DETACHED;
1995 	KQ_UNLOCK(kn->kn_kq);
1996 }
1997 
1998 static void
1999 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
2000 {
2001 	KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
2002 	KNL_ASSERT_LOCK(knl, knlislocked);
2003 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2004 	if (!kqislocked)
2005 		KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
2006     ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
2007 	if (!knlislocked)
2008 		knl->kl_lock(knl->kl_lockarg);
2009 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2010 	kn->kn_knlist = NULL;
2011 	if (!knlislocked)
2012 		knl->kl_unlock(knl->kl_lockarg);
2013 	if (!kqislocked)
2014 		KQ_LOCK(kn->kn_kq);
2015 	kn->kn_status |= KN_DETACHED;
2016 	if (!kqislocked)
2017 		KQ_UNLOCK(kn->kn_kq);
2018 }
2019 
2020 /*
2021  * remove knote from the specified knlist
2022  */
2023 void
2024 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2025 {
2026 
2027 	knlist_remove_kq(knl, kn, islocked, 0);
2028 }
2029 
2030 /*
2031  * remove knote from the specified knlist while in f_event handler.
2032  */
2033 void
2034 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
2035 {
2036 
2037 	knlist_remove_kq(knl, kn, 1,
2038 	    (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
2039 }
2040 
2041 int
2042 knlist_empty(struct knlist *knl)
2043 {
2044 
2045 	KNL_ASSERT_LOCKED(knl);
2046 	return SLIST_EMPTY(&knl->kl_list);
2047 }
2048 
2049 static struct mtx	knlist_lock;
2050 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2051 	MTX_DEF);
2052 static void knlist_mtx_lock(void *arg);
2053 static void knlist_mtx_unlock(void *arg);
2054 
2055 static void
2056 knlist_mtx_lock(void *arg)
2057 {
2058 
2059 	mtx_lock((struct mtx *)arg);
2060 }
2061 
2062 static void
2063 knlist_mtx_unlock(void *arg)
2064 {
2065 
2066 	mtx_unlock((struct mtx *)arg);
2067 }
2068 
2069 static void
2070 knlist_mtx_assert_locked(void *arg)
2071 {
2072 
2073 	mtx_assert((struct mtx *)arg, MA_OWNED);
2074 }
2075 
2076 static void
2077 knlist_mtx_assert_unlocked(void *arg)
2078 {
2079 
2080 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2081 }
2082 
2083 static void
2084 knlist_rw_rlock(void *arg)
2085 {
2086 
2087 	rw_rlock((struct rwlock *)arg);
2088 }
2089 
2090 static void
2091 knlist_rw_runlock(void *arg)
2092 {
2093 
2094 	rw_runlock((struct rwlock *)arg);
2095 }
2096 
2097 static void
2098 knlist_rw_assert_locked(void *arg)
2099 {
2100 
2101 	rw_assert((struct rwlock *)arg, RA_LOCKED);
2102 }
2103 
2104 static void
2105 knlist_rw_assert_unlocked(void *arg)
2106 {
2107 
2108 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2109 }
2110 
2111 void
2112 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2113     void (*kl_unlock)(void *),
2114     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2115 {
2116 
2117 	if (lock == NULL)
2118 		knl->kl_lockarg = &knlist_lock;
2119 	else
2120 		knl->kl_lockarg = lock;
2121 
2122 	if (kl_lock == NULL)
2123 		knl->kl_lock = knlist_mtx_lock;
2124 	else
2125 		knl->kl_lock = kl_lock;
2126 	if (kl_unlock == NULL)
2127 		knl->kl_unlock = knlist_mtx_unlock;
2128 	else
2129 		knl->kl_unlock = kl_unlock;
2130 	if (kl_assert_locked == NULL)
2131 		knl->kl_assert_locked = knlist_mtx_assert_locked;
2132 	else
2133 		knl->kl_assert_locked = kl_assert_locked;
2134 	if (kl_assert_unlocked == NULL)
2135 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2136 	else
2137 		knl->kl_assert_unlocked = kl_assert_unlocked;
2138 
2139 	SLIST_INIT(&knl->kl_list);
2140 }
2141 
2142 void
2143 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2144 {
2145 
2146 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2147 }
2148 
2149 void
2150 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2151 {
2152 
2153 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2154 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2155 }
2156 
2157 void
2158 knlist_destroy(struct knlist *knl)
2159 {
2160 
2161 #ifdef INVARIANTS
2162 	/*
2163 	 * if we run across this error, we need to find the offending
2164 	 * driver and have it call knlist_clear or knlist_delete.
2165 	 */
2166 	if (!SLIST_EMPTY(&knl->kl_list))
2167 		printf("WARNING: destroying knlist w/ knotes on it!\n");
2168 #endif
2169 
2170 	knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2171 	SLIST_INIT(&knl->kl_list);
2172 }
2173 
2174 /*
2175  * Even if we are locked, we may need to drop the lock to allow any influx
2176  * knotes time to "settle".
2177  */
2178 void
2179 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2180 {
2181 	struct knote *kn, *kn2;
2182 	struct kqueue *kq;
2183 
2184 	if (islocked)
2185 		KNL_ASSERT_LOCKED(knl);
2186 	else {
2187 		KNL_ASSERT_UNLOCKED(knl);
2188 again:		/* need to reacquire lock since we have dropped it */
2189 		knl->kl_lock(knl->kl_lockarg);
2190 	}
2191 
2192 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2193 		kq = kn->kn_kq;
2194 		KQ_LOCK(kq);
2195 		if ((kn->kn_status & KN_INFLUX)) {
2196 			KQ_UNLOCK(kq);
2197 			continue;
2198 		}
2199 		knlist_remove_kq(knl, kn, 1, 1);
2200 		if (killkn) {
2201 			kn->kn_status |= KN_INFLUX | KN_DETACHED;
2202 			KQ_UNLOCK(kq);
2203 			knote_drop(kn, td);
2204 		} else {
2205 			/* Make sure cleared knotes disappear soon */
2206 			kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2207 			KQ_UNLOCK(kq);
2208 		}
2209 		kq = NULL;
2210 	}
2211 
2212 	if (!SLIST_EMPTY(&knl->kl_list)) {
2213 		/* there are still KN_INFLUX remaining */
2214 		kn = SLIST_FIRST(&knl->kl_list);
2215 		kq = kn->kn_kq;
2216 		KQ_LOCK(kq);
2217 		KASSERT(kn->kn_status & KN_INFLUX,
2218 		    ("knote removed w/o list lock"));
2219 		knl->kl_unlock(knl->kl_lockarg);
2220 		kq->kq_state |= KQ_FLUXWAIT;
2221 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2222 		kq = NULL;
2223 		goto again;
2224 	}
2225 
2226 	if (islocked)
2227 		KNL_ASSERT_LOCKED(knl);
2228 	else {
2229 		knl->kl_unlock(knl->kl_lockarg);
2230 		KNL_ASSERT_UNLOCKED(knl);
2231 	}
2232 }
2233 
2234 /*
2235  * Remove all knotes referencing a specified fd must be called with FILEDESC
2236  * lock.  This prevents a race where a new fd comes along and occupies the
2237  * entry and we attach a knote to the fd.
2238  */
2239 void
2240 knote_fdclose(struct thread *td, int fd)
2241 {
2242 	struct filedesc *fdp = td->td_proc->p_fd;
2243 	struct kqueue *kq;
2244 	struct knote *kn;
2245 	int influx;
2246 
2247 	FILEDESC_XLOCK_ASSERT(fdp);
2248 
2249 	/*
2250 	 * We shouldn't have to worry about new kevents appearing on fd
2251 	 * since filedesc is locked.
2252 	 */
2253 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2254 		KQ_LOCK(kq);
2255 
2256 again:
2257 		influx = 0;
2258 		while (kq->kq_knlistsize > fd &&
2259 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2260 			if (kn->kn_status & KN_INFLUX) {
2261 				/* someone else might be waiting on our knote */
2262 				if (influx)
2263 					wakeup(kq);
2264 				kq->kq_state |= KQ_FLUXWAIT;
2265 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2266 				goto again;
2267 			}
2268 			kn->kn_status |= KN_INFLUX;
2269 			KQ_UNLOCK(kq);
2270 			if (!(kn->kn_status & KN_DETACHED))
2271 				kn->kn_fop->f_detach(kn);
2272 			knote_drop(kn, td);
2273 			influx = 1;
2274 			KQ_LOCK(kq);
2275 		}
2276 		KQ_UNLOCK_FLUX(kq);
2277 	}
2278 }
2279 
2280 static int
2281 knote_attach(struct knote *kn, struct kqueue *kq)
2282 {
2283 	struct klist *list;
2284 
2285 	KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2286 	KQ_OWNED(kq);
2287 
2288 	if (kn->kn_fop->f_isfd) {
2289 		if (kn->kn_id >= kq->kq_knlistsize)
2290 			return ENOMEM;
2291 		list = &kq->kq_knlist[kn->kn_id];
2292 	} else {
2293 		if (kq->kq_knhash == NULL)
2294 			return ENOMEM;
2295 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2296 	}
2297 
2298 	SLIST_INSERT_HEAD(list, kn, kn_link);
2299 
2300 	return 0;
2301 }
2302 
2303 /*
2304  * knote must already have been detached using the f_detach method.
2305  * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2306  * to prevent other removal.
2307  */
2308 static void
2309 knote_drop(struct knote *kn, struct thread *td)
2310 {
2311 	struct kqueue *kq;
2312 	struct klist *list;
2313 
2314 	kq = kn->kn_kq;
2315 
2316 	KQ_NOTOWNED(kq);
2317 	KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2318 	    ("knote_drop called without KN_INFLUX set in kn_status"));
2319 
2320 	KQ_LOCK(kq);
2321 	if (kn->kn_fop->f_isfd)
2322 		list = &kq->kq_knlist[kn->kn_id];
2323 	else
2324 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2325 
2326 	if (!SLIST_EMPTY(list))
2327 		SLIST_REMOVE(list, kn, knote, kn_link);
2328 	if (kn->kn_status & KN_QUEUED)
2329 		knote_dequeue(kn);
2330 	KQ_UNLOCK_FLUX(kq);
2331 
2332 	if (kn->kn_fop->f_isfd) {
2333 		fdrop(kn->kn_fp, td);
2334 		kn->kn_fp = NULL;
2335 	}
2336 	kqueue_fo_release(kn->kn_kevent.filter);
2337 	kn->kn_fop = NULL;
2338 	knote_free(kn);
2339 }
2340 
2341 static void
2342 knote_enqueue(struct knote *kn)
2343 {
2344 	struct kqueue *kq = kn->kn_kq;
2345 
2346 	KQ_OWNED(kn->kn_kq);
2347 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2348 
2349 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2350 	kn->kn_status |= KN_QUEUED;
2351 	kq->kq_count++;
2352 	kqueue_wakeup(kq);
2353 }
2354 
2355 static void
2356 knote_dequeue(struct knote *kn)
2357 {
2358 	struct kqueue *kq = kn->kn_kq;
2359 
2360 	KQ_OWNED(kn->kn_kq);
2361 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2362 
2363 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2364 	kn->kn_status &= ~KN_QUEUED;
2365 	kq->kq_count--;
2366 }
2367 
2368 static void
2369 knote_init(void)
2370 {
2371 
2372 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2373 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2374 }
2375 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2376 
2377 static struct knote *
2378 knote_alloc(int waitok)
2379 {
2380 
2381 	return (uma_zalloc(knote_zone, (waitok ? M_WAITOK : M_NOWAIT) |
2382 	    M_ZERO));
2383 }
2384 
2385 static void
2386 knote_free(struct knote *kn)
2387 {
2388 
2389 	uma_zfree(knote_zone, kn);
2390 }
2391 
2392 /*
2393  * Register the kev w/ the kq specified by fd.
2394  */
2395 int
2396 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2397 {
2398 	struct kqueue *kq;
2399 	struct file *fp;
2400 	cap_rights_t rights;
2401 	int error;
2402 
2403 	error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2404 	if (error != 0)
2405 		return (error);
2406 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2407 		goto noacquire;
2408 
2409 	error = kqueue_register(kq, kev, td, waitok);
2410 
2411 	kqueue_release(kq, 0);
2412 
2413 noacquire:
2414 	fdrop(fp, td);
2415 
2416 	return error;
2417 }
2418