xref: /freebsd/sys/kern/kern_event.c (revision 529a53abe2287eae08a3af62749273df775254e9)
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/capability.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 #ifdef KTRACE
69 #include <sys/ktrace.h>
70 #endif
71 
72 #include <vm/uma.h>
73 
74 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
75 
76 /*
77  * This lock is used if multiple kq locks are required.  This possibly
78  * should be made into a per proc lock.
79  */
80 static struct mtx	kq_global;
81 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF);
82 #define KQ_GLOBAL_LOCK(lck, haslck)	do {	\
83 	if (!haslck)				\
84 		mtx_lock(lck);			\
85 	haslck = 1;				\
86 } while (0)
87 #define KQ_GLOBAL_UNLOCK(lck, haslck)	do {	\
88 	if (haslck)				\
89 		mtx_unlock(lck);			\
90 	haslck = 0;				\
91 } while (0)
92 
93 TASKQUEUE_DEFINE_THREAD(kqueue);
94 
95 static int	kevent_copyout(void *arg, struct kevent *kevp, int count);
96 static int	kevent_copyin(void *arg, struct kevent *kevp, int count);
97 static int	kqueue_register(struct kqueue *kq, struct kevent *kev,
98 		    struct thread *td, int waitok);
99 static int	kqueue_acquire(struct file *fp, struct kqueue **kqp);
100 static void	kqueue_release(struct kqueue *kq, int locked);
101 static int	kqueue_expand(struct kqueue *kq, struct filterops *fops,
102 		    uintptr_t ident, int waitok);
103 static void	kqueue_task(void *arg, int pending);
104 static int	kqueue_scan(struct kqueue *kq, int maxevents,
105 		    struct kevent_copyops *k_ops,
106 		    const struct timespec *timeout,
107 		    struct kevent *keva, struct thread *td);
108 static void 	kqueue_wakeup(struct kqueue *kq);
109 static struct filterops *kqueue_fo_find(int filt);
110 static void	kqueue_fo_release(int filt);
111 
112 static fo_rdwr_t	kqueue_read;
113 static fo_rdwr_t	kqueue_write;
114 static fo_truncate_t	kqueue_truncate;
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 
121 static struct fileops kqueueops = {
122 	.fo_read = kqueue_read,
123 	.fo_write = kqueue_write,
124 	.fo_truncate = kqueue_truncate,
125 	.fo_ioctl = kqueue_ioctl,
126 	.fo_poll = kqueue_poll,
127 	.fo_kqfilter = kqueue_kqfilter,
128 	.fo_stat = kqueue_stat,
129 	.fo_close = kqueue_close,
130 	.fo_chmod = invfo_chmod,
131 	.fo_chown = invfo_chown,
132 	.fo_sendfile = invfo_sendfile,
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 	{ &null_filtops },			/* former EVFILT_NETDEV */
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 = kn->kn_ptr.p_proc;
421 	u_int event;
422 
423 	/*
424 	 * mask off extra data
425 	 */
426 	event = (u_int)hint & NOTE_PCTRLMASK;
427 
428 	/*
429 	 * if the user is interested in this event, record it.
430 	 */
431 	if (kn->kn_sfflags & event)
432 		kn->kn_fflags |= event;
433 
434 	/*
435 	 * process is gone, so flag the event as finished.
436 	 */
437 	if (event == NOTE_EXIT) {
438 		if (!(kn->kn_status & KN_DETACHED))
439 			knlist_remove_inevent(&p->p_klist, kn);
440 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
441 		kn->kn_ptr.p_proc = NULL;
442 		if (kn->kn_fflags & NOTE_EXIT)
443 			kn->kn_data = p->p_xstat;
444 		if (kn->kn_fflags == 0)
445 			kn->kn_flags |= EV_DROP;
446 		return (1);
447 	}
448 
449 	return (kn->kn_fflags != 0);
450 }
451 
452 /*
453  * Called when the process forked. It mostly does the same as the
454  * knote(), activating all knotes registered to be activated when the
455  * process forked. Additionally, for each knote attached to the
456  * parent, check whether user wants to track the new process. If so
457  * attach a new knote to it, and immediately report an event with the
458  * child's pid.
459  */
460 void
461 knote_fork(struct knlist *list, int pid)
462 {
463 	struct kqueue *kq;
464 	struct knote *kn;
465 	struct kevent kev;
466 	int error;
467 
468 	if (list == NULL)
469 		return;
470 	list->kl_lock(list->kl_lockarg);
471 
472 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
473 		if ((kn->kn_status & KN_INFLUX) == KN_INFLUX)
474 			continue;
475 		kq = kn->kn_kq;
476 		KQ_LOCK(kq);
477 		if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
478 			KQ_UNLOCK(kq);
479 			continue;
480 		}
481 
482 		/*
483 		 * The same as knote(), activate the event.
484 		 */
485 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
486 			kn->kn_status |= KN_HASKQLOCK;
487 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
488 				KNOTE_ACTIVATE(kn, 1);
489 			kn->kn_status &= ~KN_HASKQLOCK;
490 			KQ_UNLOCK(kq);
491 			continue;
492 		}
493 
494 		/*
495 		 * The NOTE_TRACK case. In addition to the activation
496 		 * of the event, we need to register new event to
497 		 * track the child. Drop the locks in preparation for
498 		 * the call to kqueue_register().
499 		 */
500 		kn->kn_status |= KN_INFLUX;
501 		KQ_UNLOCK(kq);
502 		list->kl_unlock(list->kl_lockarg);
503 
504 		/*
505 		 * Activate existing knote and register a knote with
506 		 * new process.
507 		 */
508 		kev.ident = pid;
509 		kev.filter = kn->kn_filter;
510 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
511 		kev.fflags = kn->kn_sfflags;
512 		kev.data = kn->kn_id;		/* parent */
513 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
514 		error = kqueue_register(kq, &kev, NULL, 0);
515 		if (error)
516 			kn->kn_fflags |= NOTE_TRACKERR;
517 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
518 			KNOTE_ACTIVATE(kn, 0);
519 		KQ_LOCK(kq);
520 		kn->kn_status &= ~KN_INFLUX;
521 		KQ_UNLOCK_FLUX(kq);
522 		list->kl_lock(list->kl_lockarg);
523 	}
524 	list->kl_unlock(list->kl_lockarg);
525 }
526 
527 /*
528  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
529  * interval timer support code.
530  */
531 static __inline sbintime_t
532 timer2sbintime(intptr_t data)
533 {
534 
535 #ifdef __LP64__
536 	if (data > INT64_MAX / SBT_1MS)
537 		return INT64_MAX;
538 #endif
539 	return (SBT_1MS * data);
540 }
541 
542 static void
543 filt_timerexpire(void *knx)
544 {
545 	struct callout *calloutp;
546 	struct knote *kn;
547 
548 	kn = knx;
549 	kn->kn_data++;
550 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
551 
552 	if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) {
553 		calloutp = (struct callout *)kn->kn_hook;
554 		callout_reset_sbt_on(calloutp,
555 		    timer2sbintime(kn->kn_sdata), 0 /* 1ms? */,
556 		    filt_timerexpire, kn, PCPU_GET(cpuid), 0);
557 	}
558 }
559 
560 /*
561  * data contains amount of time to sleep, in milliseconds
562  */
563 static int
564 filt_timerattach(struct knote *kn)
565 {
566 	struct callout *calloutp;
567 	sbintime_t to;
568 	unsigned int ncallouts;
569 
570 	if ((intptr_t)kn->kn_sdata < 0)
571 		return (EINVAL);
572 	if ((intptr_t)kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
573 		kn->kn_sdata = 1;
574 	to = timer2sbintime(kn->kn_sdata);
575 	if (to < 0)
576 		return (EINVAL);
577 
578 	ncallouts = atomic_load_explicit(&kq_ncallouts, memory_order_relaxed);
579 	do {
580 		if (ncallouts >= kq_calloutmax)
581 			return (ENOMEM);
582 	} while (!atomic_compare_exchange_weak_explicit(&kq_ncallouts,
583 	    &ncallouts, ncallouts + 1, memory_order_relaxed,
584 	    memory_order_relaxed));
585 
586 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
587 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
588 	calloutp = malloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
589 	callout_init(calloutp, CALLOUT_MPSAFE);
590 	kn->kn_hook = calloutp;
591 	callout_reset_sbt_on(calloutp, to, 0 /* 1ms? */,
592 	    filt_timerexpire, kn, PCPU_GET(cpuid), 0);
593 
594 	return (0);
595 }
596 
597 static void
598 filt_timerdetach(struct knote *kn)
599 {
600 	struct callout *calloutp;
601 	unsigned int old;
602 
603 	calloutp = (struct callout *)kn->kn_hook;
604 	callout_drain(calloutp);
605 	free(calloutp, M_KQUEUE);
606 	old = atomic_fetch_sub_explicit(&kq_ncallouts, 1, memory_order_relaxed);
607 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
608 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
609 }
610 
611 static int
612 filt_timer(struct knote *kn, long hint)
613 {
614 
615 	return (kn->kn_data != 0);
616 }
617 
618 static int
619 filt_userattach(struct knote *kn)
620 {
621 
622 	/*
623 	 * EVFILT_USER knotes are not attached to anything in the kernel.
624 	 */
625 	kn->kn_hook = NULL;
626 	if (kn->kn_fflags & NOTE_TRIGGER)
627 		kn->kn_hookid = 1;
628 	else
629 		kn->kn_hookid = 0;
630 	return (0);
631 }
632 
633 static void
634 filt_userdetach(__unused struct knote *kn)
635 {
636 
637 	/*
638 	 * EVFILT_USER knotes are not attached to anything in the kernel.
639 	 */
640 }
641 
642 static int
643 filt_user(struct knote *kn, __unused long hint)
644 {
645 
646 	return (kn->kn_hookid);
647 }
648 
649 static void
650 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
651 {
652 	u_int ffctrl;
653 
654 	switch (type) {
655 	case EVENT_REGISTER:
656 		if (kev->fflags & NOTE_TRIGGER)
657 			kn->kn_hookid = 1;
658 
659 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
660 		kev->fflags &= NOTE_FFLAGSMASK;
661 		switch (ffctrl) {
662 		case NOTE_FFNOP:
663 			break;
664 
665 		case NOTE_FFAND:
666 			kn->kn_sfflags &= kev->fflags;
667 			break;
668 
669 		case NOTE_FFOR:
670 			kn->kn_sfflags |= kev->fflags;
671 			break;
672 
673 		case NOTE_FFCOPY:
674 			kn->kn_sfflags = kev->fflags;
675 			break;
676 
677 		default:
678 			/* XXX Return error? */
679 			break;
680 		}
681 		kn->kn_sdata = kev->data;
682 		if (kev->flags & EV_CLEAR) {
683 			kn->kn_hookid = 0;
684 			kn->kn_data = 0;
685 			kn->kn_fflags = 0;
686 		}
687 		break;
688 
689         case EVENT_PROCESS:
690 		*kev = kn->kn_kevent;
691 		kev->fflags = kn->kn_sfflags;
692 		kev->data = kn->kn_sdata;
693 		if (kn->kn_flags & EV_CLEAR) {
694 			kn->kn_hookid = 0;
695 			kn->kn_data = 0;
696 			kn->kn_fflags = 0;
697 		}
698 		break;
699 
700 	default:
701 		panic("filt_usertouch() - invalid type (%ld)", type);
702 		break;
703 	}
704 }
705 
706 int
707 sys_kqueue(struct thread *td, struct kqueue_args *uap)
708 {
709 	struct filedesc *fdp;
710 	struct kqueue *kq;
711 	struct file *fp;
712 	struct proc *p;
713 	struct ucred *cred;
714 	int fd, error;
715 
716 	p = td->td_proc;
717 	cred = td->td_ucred;
718 	crhold(cred);
719 	PROC_LOCK(p);
720 	if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td->td_proc,
721 	    RLIMIT_KQUEUES))) {
722 		PROC_UNLOCK(p);
723 		crfree(cred);
724 		return (ENOMEM);
725 	}
726 	PROC_UNLOCK(p);
727 
728 	fdp = p->p_fd;
729 	error = falloc(td, &fp, &fd, 0);
730 	if (error)
731 		goto done2;
732 
733 	/* An extra reference on `fp' has been held for us by falloc(). */
734 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
735 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK);
736 	TAILQ_INIT(&kq->kq_head);
737 	kq->kq_fdp = fdp;
738 	kq->kq_cred = cred;
739 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
740 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
741 
742 	FILEDESC_XLOCK(fdp);
743 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
744 	FILEDESC_XUNLOCK(fdp);
745 
746 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
747 	fdrop(fp, td);
748 
749 	td->td_retval[0] = fd;
750 done2:
751 	if (error != 0) {
752 		chgkqcnt(cred->cr_ruidinfo, -1, 0);
753 		crfree(cred);
754 	}
755 	return (error);
756 }
757 
758 #ifndef _SYS_SYSPROTO_H_
759 struct kevent_args {
760 	int	fd;
761 	const struct kevent *changelist;
762 	int	nchanges;
763 	struct	kevent *eventlist;
764 	int	nevents;
765 	const struct timespec *timeout;
766 };
767 #endif
768 int
769 sys_kevent(struct thread *td, struct kevent_args *uap)
770 {
771 	struct timespec ts, *tsp;
772 	struct kevent_copyops k_ops = { uap,
773 					kevent_copyout,
774 					kevent_copyin};
775 	int error;
776 #ifdef KTRACE
777 	struct uio ktruio;
778 	struct iovec ktriov;
779 	struct uio *ktruioin = NULL;
780 	struct uio *ktruioout = NULL;
781 #endif
782 
783 	if (uap->timeout != NULL) {
784 		error = copyin(uap->timeout, &ts, sizeof(ts));
785 		if (error)
786 			return (error);
787 		tsp = &ts;
788 	} else
789 		tsp = NULL;
790 
791 #ifdef KTRACE
792 	if (KTRPOINT(td, KTR_GENIO)) {
793 		ktriov.iov_base = uap->changelist;
794 		ktriov.iov_len = uap->nchanges * sizeof(struct kevent);
795 		ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
796 		    .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
797 		    .uio_td = td };
798 		ktruioin = cloneuio(&ktruio);
799 		ktriov.iov_base = uap->eventlist;
800 		ktriov.iov_len = uap->nevents * sizeof(struct kevent);
801 		ktruioout = cloneuio(&ktruio);
802 	}
803 #endif
804 
805 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
806 	    &k_ops, tsp);
807 
808 #ifdef KTRACE
809 	if (ktruioin != NULL) {
810 		ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent);
811 		ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
812 		ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent);
813 		ktrgenio(uap->fd, UIO_READ, ktruioout, error);
814 	}
815 #endif
816 
817 	return (error);
818 }
819 
820 /*
821  * Copy 'count' items into the destination list pointed to by uap->eventlist.
822  */
823 static int
824 kevent_copyout(void *arg, struct kevent *kevp, int count)
825 {
826 	struct kevent_args *uap;
827 	int error;
828 
829 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
830 	uap = (struct kevent_args *)arg;
831 
832 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
833 	if (error == 0)
834 		uap->eventlist += count;
835 	return (error);
836 }
837 
838 /*
839  * Copy 'count' items from the list pointed to by uap->changelist.
840  */
841 static int
842 kevent_copyin(void *arg, struct kevent *kevp, int count)
843 {
844 	struct kevent_args *uap;
845 	int error;
846 
847 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
848 	uap = (struct kevent_args *)arg;
849 
850 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
851 	if (error == 0)
852 		uap->changelist += count;
853 	return (error);
854 }
855 
856 int
857 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
858     struct kevent_copyops *k_ops, const struct timespec *timeout)
859 {
860 	struct kevent keva[KQ_NEVENTS];
861 	struct kevent *kevp, *changes;
862 	struct kqueue *kq;
863 	struct file *fp;
864 	cap_rights_t rights;
865 	int i, n, nerrors, error;
866 
867 	cap_rights_init(&rights);
868 	if (nchanges > 0)
869 		cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
870 	if (nevents > 0)
871 		cap_rights_set(&rights, CAP_KQUEUE_EVENT);
872 	error = fget(td, fd, &rights, &fp);
873 	if (error != 0)
874 		return (error);
875 
876 	error = kqueue_acquire(fp, &kq);
877 	if (error != 0)
878 		goto done_norel;
879 
880 	nerrors = 0;
881 
882 	while (nchanges > 0) {
883 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
884 		error = k_ops->k_copyin(k_ops->arg, keva, n);
885 		if (error)
886 			goto done;
887 		changes = keva;
888 		for (i = 0; i < n; i++) {
889 			kevp = &changes[i];
890 			if (!kevp->filter)
891 				continue;
892 			kevp->flags &= ~EV_SYSFLAGS;
893 			error = kqueue_register(kq, kevp, td, 1);
894 			if (error || (kevp->flags & EV_RECEIPT)) {
895 				if (nevents != 0) {
896 					kevp->flags = EV_ERROR;
897 					kevp->data = error;
898 					(void) k_ops->k_copyout(k_ops->arg,
899 					    kevp, 1);
900 					nevents--;
901 					nerrors++;
902 				} else {
903 					goto done;
904 				}
905 			}
906 		}
907 		nchanges -= n;
908 	}
909 	if (nerrors) {
910 		td->td_retval[0] = nerrors;
911 		error = 0;
912 		goto done;
913 	}
914 
915 	error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td);
916 done:
917 	kqueue_release(kq, 0);
918 done_norel:
919 	fdrop(fp, td);
920 	return (error);
921 }
922 
923 int
924 kqueue_add_filteropts(int filt, struct filterops *filtops)
925 {
926 	int error;
927 
928 	error = 0;
929 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
930 		printf(
931 "trying to add a filterop that is out of range: %d is beyond %d\n",
932 		    ~filt, EVFILT_SYSCOUNT);
933 		return EINVAL;
934 	}
935 	mtx_lock(&filterops_lock);
936 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
937 	    sysfilt_ops[~filt].for_fop != NULL)
938 		error = EEXIST;
939 	else {
940 		sysfilt_ops[~filt].for_fop = filtops;
941 		sysfilt_ops[~filt].for_refcnt = 0;
942 	}
943 	mtx_unlock(&filterops_lock);
944 
945 	return (error);
946 }
947 
948 int
949 kqueue_del_filteropts(int filt)
950 {
951 	int error;
952 
953 	error = 0;
954 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
955 		return EINVAL;
956 
957 	mtx_lock(&filterops_lock);
958 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
959 	    sysfilt_ops[~filt].for_fop == NULL)
960 		error = EINVAL;
961 	else if (sysfilt_ops[~filt].for_refcnt != 0)
962 		error = EBUSY;
963 	else {
964 		sysfilt_ops[~filt].for_fop = &null_filtops;
965 		sysfilt_ops[~filt].for_refcnt = 0;
966 	}
967 	mtx_unlock(&filterops_lock);
968 
969 	return error;
970 }
971 
972 static struct filterops *
973 kqueue_fo_find(int filt)
974 {
975 
976 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
977 		return NULL;
978 
979 	mtx_lock(&filterops_lock);
980 	sysfilt_ops[~filt].for_refcnt++;
981 	if (sysfilt_ops[~filt].for_fop == NULL)
982 		sysfilt_ops[~filt].for_fop = &null_filtops;
983 	mtx_unlock(&filterops_lock);
984 
985 	return sysfilt_ops[~filt].for_fop;
986 }
987 
988 static void
989 kqueue_fo_release(int filt)
990 {
991 
992 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
993 		return;
994 
995 	mtx_lock(&filterops_lock);
996 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
997 	    ("filter object refcount not valid on release"));
998 	sysfilt_ops[~filt].for_refcnt--;
999 	mtx_unlock(&filterops_lock);
1000 }
1001 
1002 /*
1003  * A ref to kq (obtained via kqueue_acquire) must be held.  waitok will
1004  * influence if memory allocation should wait.  Make sure it is 0 if you
1005  * hold any mutexes.
1006  */
1007 static int
1008 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1009 {
1010 	struct filterops *fops;
1011 	struct file *fp;
1012 	struct knote *kn, *tkn;
1013 	cap_rights_t rights;
1014 	int error, filt, event;
1015 	int haskqglobal, filedesc_unlock;
1016 
1017 	fp = NULL;
1018 	kn = NULL;
1019 	error = 0;
1020 	haskqglobal = 0;
1021 	filedesc_unlock = 0;
1022 
1023 	filt = kev->filter;
1024 	fops = kqueue_fo_find(filt);
1025 	if (fops == NULL)
1026 		return EINVAL;
1027 
1028 	tkn = knote_alloc(waitok);		/* prevent waiting with locks */
1029 
1030 findkn:
1031 	if (fops->f_isfd) {
1032 		KASSERT(td != NULL, ("td is NULL"));
1033 		error = fget(td, kev->ident,
1034 		    cap_rights_init(&rights, CAP_EVENT), &fp);
1035 		if (error)
1036 			goto done;
1037 
1038 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1039 		    kev->ident, 0) != 0) {
1040 			/* try again */
1041 			fdrop(fp, td);
1042 			fp = NULL;
1043 			error = kqueue_expand(kq, fops, kev->ident, waitok);
1044 			if (error)
1045 				goto done;
1046 			goto findkn;
1047 		}
1048 
1049 		if (fp->f_type == DTYPE_KQUEUE) {
1050 			/*
1051 			 * if we add some inteligence about what we are doing,
1052 			 * we should be able to support events on ourselves.
1053 			 * We need to know when we are doing this to prevent
1054 			 * getting both the knlist lock and the kq lock since
1055 			 * they are the same thing.
1056 			 */
1057 			if (fp->f_data == kq) {
1058 				error = EINVAL;
1059 				goto done;
1060 			}
1061 
1062 			/*
1063 			 * Pre-lock the filedesc before the global
1064 			 * lock mutex, see the comment in
1065 			 * kqueue_close().
1066 			 */
1067 			FILEDESC_XLOCK(td->td_proc->p_fd);
1068 			filedesc_unlock = 1;
1069 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1070 		}
1071 
1072 		KQ_LOCK(kq);
1073 		if (kev->ident < kq->kq_knlistsize) {
1074 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1075 				if (kev->filter == kn->kn_filter)
1076 					break;
1077 		}
1078 	} else {
1079 		if ((kev->flags & EV_ADD) == EV_ADD)
1080 			kqueue_expand(kq, fops, kev->ident, waitok);
1081 
1082 		KQ_LOCK(kq);
1083 		if (kq->kq_knhashmask != 0) {
1084 			struct klist *list;
1085 
1086 			list = &kq->kq_knhash[
1087 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1088 			SLIST_FOREACH(kn, list, kn_link)
1089 				if (kev->ident == kn->kn_id &&
1090 				    kev->filter == kn->kn_filter)
1091 					break;
1092 		}
1093 	}
1094 
1095 	/* knote is in the process of changing, wait for it to stablize. */
1096 	if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1097 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1098 		if (filedesc_unlock) {
1099 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1100 			filedesc_unlock = 0;
1101 		}
1102 		kq->kq_state |= KQ_FLUXWAIT;
1103 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1104 		if (fp != NULL) {
1105 			fdrop(fp, td);
1106 			fp = NULL;
1107 		}
1108 		goto findkn;
1109 	}
1110 
1111 	/*
1112 	 * kn now contains the matching knote, or NULL if no match
1113 	 */
1114 	if (kn == NULL) {
1115 		if (kev->flags & EV_ADD) {
1116 			kn = tkn;
1117 			tkn = NULL;
1118 			if (kn == NULL) {
1119 				KQ_UNLOCK(kq);
1120 				error = ENOMEM;
1121 				goto done;
1122 			}
1123 			kn->kn_fp = fp;
1124 			kn->kn_kq = kq;
1125 			kn->kn_fop = fops;
1126 			/*
1127 			 * apply reference counts to knote structure, and
1128 			 * do not release it at the end of this routine.
1129 			 */
1130 			fops = NULL;
1131 			fp = NULL;
1132 
1133 			kn->kn_sfflags = kev->fflags;
1134 			kn->kn_sdata = kev->data;
1135 			kev->fflags = 0;
1136 			kev->data = 0;
1137 			kn->kn_kevent = *kev;
1138 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1139 			    EV_ENABLE | EV_DISABLE);
1140 			kn->kn_status = KN_INFLUX|KN_DETACHED;
1141 
1142 			error = knote_attach(kn, kq);
1143 			KQ_UNLOCK(kq);
1144 			if (error != 0) {
1145 				tkn = kn;
1146 				goto done;
1147 			}
1148 
1149 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1150 				knote_drop(kn, td);
1151 				goto done;
1152 			}
1153 			KN_LIST_LOCK(kn);
1154 			goto done_ev_add;
1155 		} else {
1156 			/* No matching knote and the EV_ADD flag is not set. */
1157 			KQ_UNLOCK(kq);
1158 			error = ENOENT;
1159 			goto done;
1160 		}
1161 	}
1162 
1163 	if (kev->flags & EV_DELETE) {
1164 		kn->kn_status |= KN_INFLUX;
1165 		KQ_UNLOCK(kq);
1166 		if (!(kn->kn_status & KN_DETACHED))
1167 			kn->kn_fop->f_detach(kn);
1168 		knote_drop(kn, td);
1169 		goto done;
1170 	}
1171 
1172 	/*
1173 	 * The user may change some filter values after the initial EV_ADD,
1174 	 * but doing so will not reset any filter which has already been
1175 	 * triggered.
1176 	 */
1177 	kn->kn_status |= KN_INFLUX;
1178 	KQ_UNLOCK(kq);
1179 	KN_LIST_LOCK(kn);
1180 	kn->kn_kevent.udata = kev->udata;
1181 	if (!fops->f_isfd && fops->f_touch != NULL) {
1182 		fops->f_touch(kn, kev, EVENT_REGISTER);
1183 	} else {
1184 		kn->kn_sfflags = kev->fflags;
1185 		kn->kn_sdata = kev->data;
1186 	}
1187 
1188 	/*
1189 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1190 	 * the initial attach event decides that the event is "completed"
1191 	 * already.  i.e. filt_procattach is called on a zombie process.  It
1192 	 * will call filt_proc which will remove it from the list, and NULL
1193 	 * kn_knlist.
1194 	 */
1195 done_ev_add:
1196 	event = kn->kn_fop->f_event(kn, 0);
1197 	KQ_LOCK(kq);
1198 	if (event)
1199 		KNOTE_ACTIVATE(kn, 1);
1200 	kn->kn_status &= ~KN_INFLUX;
1201 	KN_LIST_UNLOCK(kn);
1202 
1203 	if ((kev->flags & EV_DISABLE) &&
1204 	    ((kn->kn_status & KN_DISABLED) == 0)) {
1205 		kn->kn_status |= KN_DISABLED;
1206 	}
1207 
1208 	if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1209 		kn->kn_status &= ~KN_DISABLED;
1210 		if ((kn->kn_status & KN_ACTIVE) &&
1211 		    ((kn->kn_status & KN_QUEUED) == 0))
1212 			knote_enqueue(kn);
1213 	}
1214 	KQ_UNLOCK_FLUX(kq);
1215 
1216 done:
1217 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1218 	if (filedesc_unlock)
1219 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1220 	if (fp != NULL)
1221 		fdrop(fp, td);
1222 	if (tkn != NULL)
1223 		knote_free(tkn);
1224 	if (fops != NULL)
1225 		kqueue_fo_release(filt);
1226 	return (error);
1227 }
1228 
1229 static int
1230 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1231 {
1232 	int error;
1233 	struct kqueue *kq;
1234 
1235 	error = 0;
1236 
1237 	kq = fp->f_data;
1238 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1239 		return (EBADF);
1240 	*kqp = kq;
1241 	KQ_LOCK(kq);
1242 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1243 		KQ_UNLOCK(kq);
1244 		return (EBADF);
1245 	}
1246 	kq->kq_refcnt++;
1247 	KQ_UNLOCK(kq);
1248 
1249 	return error;
1250 }
1251 
1252 static void
1253 kqueue_release(struct kqueue *kq, int locked)
1254 {
1255 	if (locked)
1256 		KQ_OWNED(kq);
1257 	else
1258 		KQ_LOCK(kq);
1259 	kq->kq_refcnt--;
1260 	if (kq->kq_refcnt == 1)
1261 		wakeup(&kq->kq_refcnt);
1262 	if (!locked)
1263 		KQ_UNLOCK(kq);
1264 }
1265 
1266 static void
1267 kqueue_schedtask(struct kqueue *kq)
1268 {
1269 
1270 	KQ_OWNED(kq);
1271 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1272 	    ("scheduling kqueue task while draining"));
1273 
1274 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1275 		taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task);
1276 		kq->kq_state |= KQ_TASKSCHED;
1277 	}
1278 }
1279 
1280 /*
1281  * Expand the kq to make sure we have storage for fops/ident pair.
1282  *
1283  * Return 0 on success (or no work necessary), return errno on failure.
1284  *
1285  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1286  * If kqueue_register is called from a non-fd context, there usually/should
1287  * be no locks held.
1288  */
1289 static int
1290 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1291 	int waitok)
1292 {
1293 	struct klist *list, *tmp_knhash, *to_free;
1294 	u_long tmp_knhashmask;
1295 	int size;
1296 	int fd;
1297 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
1298 
1299 	KQ_NOTOWNED(kq);
1300 
1301 	to_free = NULL;
1302 	if (fops->f_isfd) {
1303 		fd = ident;
1304 		if (kq->kq_knlistsize <= fd) {
1305 			size = kq->kq_knlistsize;
1306 			while (size <= fd)
1307 				size += KQEXTENT;
1308 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1309 			if (list == NULL)
1310 				return ENOMEM;
1311 			KQ_LOCK(kq);
1312 			if (kq->kq_knlistsize > fd) {
1313 				to_free = list;
1314 				list = NULL;
1315 			} else {
1316 				if (kq->kq_knlist != NULL) {
1317 					bcopy(kq->kq_knlist, list,
1318 					    kq->kq_knlistsize * sizeof(*list));
1319 					to_free = kq->kq_knlist;
1320 					kq->kq_knlist = NULL;
1321 				}
1322 				bzero((caddr_t)list +
1323 				    kq->kq_knlistsize * sizeof(*list),
1324 				    (size - kq->kq_knlistsize) * sizeof(*list));
1325 				kq->kq_knlistsize = size;
1326 				kq->kq_knlist = list;
1327 			}
1328 			KQ_UNLOCK(kq);
1329 		}
1330 	} else {
1331 		if (kq->kq_knhashmask == 0) {
1332 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1333 			    &tmp_knhashmask);
1334 			if (tmp_knhash == NULL)
1335 				return ENOMEM;
1336 			KQ_LOCK(kq);
1337 			if (kq->kq_knhashmask == 0) {
1338 				kq->kq_knhash = tmp_knhash;
1339 				kq->kq_knhashmask = tmp_knhashmask;
1340 			} else {
1341 				to_free = tmp_knhash;
1342 			}
1343 			KQ_UNLOCK(kq);
1344 		}
1345 	}
1346 	free(to_free, M_KQUEUE);
1347 
1348 	KQ_NOTOWNED(kq);
1349 	return 0;
1350 }
1351 
1352 static void
1353 kqueue_task(void *arg, int pending)
1354 {
1355 	struct kqueue *kq;
1356 	int haskqglobal;
1357 
1358 	haskqglobal = 0;
1359 	kq = arg;
1360 
1361 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1362 	KQ_LOCK(kq);
1363 
1364 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1365 
1366 	kq->kq_state &= ~KQ_TASKSCHED;
1367 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1368 		wakeup(&kq->kq_state);
1369 	}
1370 	KQ_UNLOCK(kq);
1371 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1372 }
1373 
1374 /*
1375  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1376  * We treat KN_MARKER knotes as if they are INFLUX.
1377  */
1378 static int
1379 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1380     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1381 {
1382 	struct kevent *kevp;
1383 	struct knote *kn, *marker;
1384 	sbintime_t asbt, rsbt;
1385 	int count, error, haskqglobal, influx, nkev, touch;
1386 
1387 	count = maxevents;
1388 	nkev = 0;
1389 	error = 0;
1390 	haskqglobal = 0;
1391 
1392 	if (maxevents == 0)
1393 		goto done_nl;
1394 
1395 	rsbt = 0;
1396 	if (tsp != NULL) {
1397 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1398 		    tsp->tv_nsec >= 1000000000) {
1399 			error = EINVAL;
1400 			goto done_nl;
1401 		}
1402 		if (timespecisset(tsp)) {
1403 			if (tsp->tv_sec <= INT32_MAX) {
1404 				rsbt = tstosbt(*tsp);
1405 				if (TIMESEL(&asbt, rsbt))
1406 					asbt += tc_tick_sbt;
1407 				if (asbt <= INT64_MAX - rsbt)
1408 					asbt += rsbt;
1409 				else
1410 					asbt = 0;
1411 				rsbt >>= tc_precexp;
1412 			} else
1413 				asbt = 0;
1414 		} else
1415 			asbt = -1;
1416 	} else
1417 		asbt = 0;
1418 	marker = knote_alloc(1);
1419 	if (marker == NULL) {
1420 		error = ENOMEM;
1421 		goto done_nl;
1422 	}
1423 	marker->kn_status = KN_MARKER;
1424 	KQ_LOCK(kq);
1425 
1426 retry:
1427 	kevp = keva;
1428 	if (kq->kq_count == 0) {
1429 		if (asbt == -1) {
1430 			error = EWOULDBLOCK;
1431 		} else {
1432 			kq->kq_state |= KQ_SLEEP;
1433 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1434 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1435 		}
1436 		if (error == 0)
1437 			goto retry;
1438 		/* don't restart after signals... */
1439 		if (error == ERESTART)
1440 			error = EINTR;
1441 		else if (error == EWOULDBLOCK)
1442 			error = 0;
1443 		goto done;
1444 	}
1445 
1446 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1447 	influx = 0;
1448 	while (count) {
1449 		KQ_OWNED(kq);
1450 		kn = TAILQ_FIRST(&kq->kq_head);
1451 
1452 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1453 		    (kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1454 			if (influx) {
1455 				influx = 0;
1456 				KQ_FLUX_WAKEUP(kq);
1457 			}
1458 			kq->kq_state |= KQ_FLUXWAIT;
1459 			error = msleep(kq, &kq->kq_lock, PSOCK,
1460 			    "kqflxwt", 0);
1461 			continue;
1462 		}
1463 
1464 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1465 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1466 			kn->kn_status &= ~KN_QUEUED;
1467 			kq->kq_count--;
1468 			continue;
1469 		}
1470 		if (kn == marker) {
1471 			KQ_FLUX_WAKEUP(kq);
1472 			if (count == maxevents)
1473 				goto retry;
1474 			goto done;
1475 		}
1476 		KASSERT((kn->kn_status & KN_INFLUX) == 0,
1477 		    ("KN_INFLUX set when not suppose to be"));
1478 
1479 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1480 			kn->kn_status &= ~KN_QUEUED;
1481 			kn->kn_status |= KN_INFLUX;
1482 			kq->kq_count--;
1483 			KQ_UNLOCK(kq);
1484 			/*
1485 			 * We don't need to lock the list since we've marked
1486 			 * it _INFLUX.
1487 			 */
1488 			if (!(kn->kn_status & KN_DETACHED))
1489 				kn->kn_fop->f_detach(kn);
1490 			knote_drop(kn, td);
1491 			KQ_LOCK(kq);
1492 			continue;
1493 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1494 			kn->kn_status &= ~KN_QUEUED;
1495 			kn->kn_status |= KN_INFLUX;
1496 			kq->kq_count--;
1497 			KQ_UNLOCK(kq);
1498 			/*
1499 			 * We don't need to lock the list since we've marked
1500 			 * it _INFLUX.
1501 			 */
1502 			*kevp = kn->kn_kevent;
1503 			if (!(kn->kn_status & KN_DETACHED))
1504 				kn->kn_fop->f_detach(kn);
1505 			knote_drop(kn, td);
1506 			KQ_LOCK(kq);
1507 			kn = NULL;
1508 		} else {
1509 			kn->kn_status |= KN_INFLUX;
1510 			KQ_UNLOCK(kq);
1511 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1512 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1513 			KN_LIST_LOCK(kn);
1514 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1515 				KQ_LOCK(kq);
1516 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1517 				kn->kn_status &=
1518 				    ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX);
1519 				kq->kq_count--;
1520 				KN_LIST_UNLOCK(kn);
1521 				influx = 1;
1522 				continue;
1523 			}
1524 			touch = (!kn->kn_fop->f_isfd &&
1525 			    kn->kn_fop->f_touch != NULL);
1526 			if (touch)
1527 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1528 			else
1529 				*kevp = kn->kn_kevent;
1530 			KQ_LOCK(kq);
1531 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1532 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1533 				/*
1534 				 * Manually clear knotes who weren't
1535 				 * 'touch'ed.
1536 				 */
1537 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1538 					kn->kn_data = 0;
1539 					kn->kn_fflags = 0;
1540 				}
1541 				if (kn->kn_flags & EV_DISPATCH)
1542 					kn->kn_status |= KN_DISABLED;
1543 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1544 				kq->kq_count--;
1545 			} else
1546 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1547 
1548 			kn->kn_status &= ~(KN_INFLUX);
1549 			KN_LIST_UNLOCK(kn);
1550 			influx = 1;
1551 		}
1552 
1553 		/* we are returning a copy to the user */
1554 		kevp++;
1555 		nkev++;
1556 		count--;
1557 
1558 		if (nkev == KQ_NEVENTS) {
1559 			influx = 0;
1560 			KQ_UNLOCK_FLUX(kq);
1561 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1562 			nkev = 0;
1563 			kevp = keva;
1564 			KQ_LOCK(kq);
1565 			if (error)
1566 				break;
1567 		}
1568 	}
1569 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1570 done:
1571 	KQ_OWNED(kq);
1572 	KQ_UNLOCK_FLUX(kq);
1573 	knote_free(marker);
1574 done_nl:
1575 	KQ_NOTOWNED(kq);
1576 	if (nkev != 0)
1577 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1578 	td->td_retval[0] = maxevents - count;
1579 	return (error);
1580 }
1581 
1582 /*
1583  * XXX
1584  * This could be expanded to call kqueue_scan, if desired.
1585  */
1586 /*ARGSUSED*/
1587 static int
1588 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
1589 	int flags, struct thread *td)
1590 {
1591 	return (ENXIO);
1592 }
1593 
1594 /*ARGSUSED*/
1595 static int
1596 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1597 	 int flags, struct thread *td)
1598 {
1599 	return (ENXIO);
1600 }
1601 
1602 /*ARGSUSED*/
1603 static int
1604 kqueue_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1605 	struct thread *td)
1606 {
1607 
1608 	return (EINVAL);
1609 }
1610 
1611 /*ARGSUSED*/
1612 static int
1613 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1614 	struct ucred *active_cred, struct thread *td)
1615 {
1616 	/*
1617 	 * Enabling sigio causes two major problems:
1618 	 * 1) infinite recursion:
1619 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1620 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1621 	 * into itself over and over.  Sending the sigio causes the kqueue
1622 	 * to become ready, which in turn posts sigio again, forever.
1623 	 * Solution: this can be solved by setting a flag in the kqueue that
1624 	 * we have a SIGIO in progress.
1625 	 * 2) locking problems:
1626 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1627 	 * us above the proc and pgrp locks.
1628 	 * Solution: Post a signal using an async mechanism, being sure to
1629 	 * record a generation count in the delivery so that we do not deliver
1630 	 * a signal to the wrong process.
1631 	 *
1632 	 * Note, these two mechanisms are somewhat mutually exclusive!
1633 	 */
1634 #if 0
1635 	struct kqueue *kq;
1636 
1637 	kq = fp->f_data;
1638 	switch (cmd) {
1639 	case FIOASYNC:
1640 		if (*(int *)data) {
1641 			kq->kq_state |= KQ_ASYNC;
1642 		} else {
1643 			kq->kq_state &= ~KQ_ASYNC;
1644 		}
1645 		return (0);
1646 
1647 	case FIOSETOWN:
1648 		return (fsetown(*(int *)data, &kq->kq_sigio));
1649 
1650 	case FIOGETOWN:
1651 		*(int *)data = fgetown(&kq->kq_sigio);
1652 		return (0);
1653 	}
1654 #endif
1655 
1656 	return (ENOTTY);
1657 }
1658 
1659 /*ARGSUSED*/
1660 static int
1661 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1662 	struct thread *td)
1663 {
1664 	struct kqueue *kq;
1665 	int revents = 0;
1666 	int error;
1667 
1668 	if ((error = kqueue_acquire(fp, &kq)))
1669 		return POLLERR;
1670 
1671 	KQ_LOCK(kq);
1672 	if (events & (POLLIN | POLLRDNORM)) {
1673 		if (kq->kq_count) {
1674 			revents |= events & (POLLIN | POLLRDNORM);
1675 		} else {
1676 			selrecord(td, &kq->kq_sel);
1677 			if (SEL_WAITING(&kq->kq_sel))
1678 				kq->kq_state |= KQ_SEL;
1679 		}
1680 	}
1681 	kqueue_release(kq, 1);
1682 	KQ_UNLOCK(kq);
1683 	return (revents);
1684 }
1685 
1686 /*ARGSUSED*/
1687 static int
1688 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1689 	struct thread *td)
1690 {
1691 
1692 	bzero((void *)st, sizeof *st);
1693 	/*
1694 	 * We no longer return kq_count because the unlocked value is useless.
1695 	 * If you spent all this time getting the count, why not spend your
1696 	 * syscall better by calling kevent?
1697 	 *
1698 	 * XXX - This is needed for libc_r.
1699 	 */
1700 	st->st_mode = S_IFIFO;
1701 	return (0);
1702 }
1703 
1704 /*ARGSUSED*/
1705 static int
1706 kqueue_close(struct file *fp, struct thread *td)
1707 {
1708 	struct kqueue *kq = fp->f_data;
1709 	struct filedesc *fdp;
1710 	struct knote *kn;
1711 	int i;
1712 	int error;
1713 	int filedesc_unlock;
1714 
1715 	if ((error = kqueue_acquire(fp, &kq)))
1716 		return error;
1717 
1718 	filedesc_unlock = 0;
1719 	KQ_LOCK(kq);
1720 
1721 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1722 	    ("kqueue already closing"));
1723 	kq->kq_state |= KQ_CLOSING;
1724 	if (kq->kq_refcnt > 1)
1725 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1726 
1727 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1728 	fdp = kq->kq_fdp;
1729 
1730 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
1731 	    ("kqueue's knlist not empty"));
1732 
1733 	for (i = 0; i < kq->kq_knlistsize; i++) {
1734 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1735 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1736 				kq->kq_state |= KQ_FLUXWAIT;
1737 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1738 				continue;
1739 			}
1740 			kn->kn_status |= KN_INFLUX;
1741 			KQ_UNLOCK(kq);
1742 			if (!(kn->kn_status & KN_DETACHED))
1743 				kn->kn_fop->f_detach(kn);
1744 			knote_drop(kn, td);
1745 			KQ_LOCK(kq);
1746 		}
1747 	}
1748 	if (kq->kq_knhashmask != 0) {
1749 		for (i = 0; i <= kq->kq_knhashmask; i++) {
1750 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1751 				if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1752 					kq->kq_state |= KQ_FLUXWAIT;
1753 					msleep(kq, &kq->kq_lock, PSOCK,
1754 					       "kqclo2", 0);
1755 					continue;
1756 				}
1757 				kn->kn_status |= KN_INFLUX;
1758 				KQ_UNLOCK(kq);
1759 				if (!(kn->kn_status & KN_DETACHED))
1760 					kn->kn_fop->f_detach(kn);
1761 				knote_drop(kn, td);
1762 				KQ_LOCK(kq);
1763 			}
1764 		}
1765 	}
1766 
1767 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1768 		kq->kq_state |= KQ_TASKDRAIN;
1769 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1770 	}
1771 
1772 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1773 		selwakeuppri(&kq->kq_sel, PSOCK);
1774 		if (!SEL_WAITING(&kq->kq_sel))
1775 			kq->kq_state &= ~KQ_SEL;
1776 	}
1777 
1778 	KQ_UNLOCK(kq);
1779 
1780 	/*
1781 	 * We could be called due to the knote_drop() doing fdrop(),
1782 	 * called from kqueue_register().  In this case the global
1783 	 * lock is owned, and filedesc sx is locked before, to not
1784 	 * take the sleepable lock after non-sleepable.
1785 	 */
1786 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1787 		FILEDESC_XLOCK(fdp);
1788 		filedesc_unlock = 1;
1789 	} else
1790 		filedesc_unlock = 0;
1791 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1792 	if (filedesc_unlock)
1793 		FILEDESC_XUNLOCK(fdp);
1794 
1795 	seldrain(&kq->kq_sel);
1796 	knlist_destroy(&kq->kq_sel.si_note);
1797 	mtx_destroy(&kq->kq_lock);
1798 	kq->kq_fdp = NULL;
1799 
1800 	if (kq->kq_knhash != NULL)
1801 		free(kq->kq_knhash, M_KQUEUE);
1802 	if (kq->kq_knlist != NULL)
1803 		free(kq->kq_knlist, M_KQUEUE);
1804 
1805 	funsetown(&kq->kq_sigio);
1806 	chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
1807 	crfree(kq->kq_cred);
1808 	free(kq, M_KQUEUE);
1809 	fp->f_data = NULL;
1810 
1811 	return (0);
1812 }
1813 
1814 static void
1815 kqueue_wakeup(struct kqueue *kq)
1816 {
1817 	KQ_OWNED(kq);
1818 
1819 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
1820 		kq->kq_state &= ~KQ_SLEEP;
1821 		wakeup(kq);
1822 	}
1823 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1824 		selwakeuppri(&kq->kq_sel, PSOCK);
1825 		if (!SEL_WAITING(&kq->kq_sel))
1826 			kq->kq_state &= ~KQ_SEL;
1827 	}
1828 	if (!knlist_empty(&kq->kq_sel.si_note))
1829 		kqueue_schedtask(kq);
1830 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
1831 		pgsigio(&kq->kq_sigio, SIGIO, 0);
1832 	}
1833 }
1834 
1835 /*
1836  * Walk down a list of knotes, activating them if their event has triggered.
1837  *
1838  * There is a possibility to optimize in the case of one kq watching another.
1839  * Instead of scheduling a task to wake it up, you could pass enough state
1840  * down the chain to make up the parent kqueue.  Make this code functional
1841  * first.
1842  */
1843 void
1844 knote(struct knlist *list, long hint, int lockflags)
1845 {
1846 	struct kqueue *kq;
1847 	struct knote *kn;
1848 	int error;
1849 
1850 	if (list == NULL)
1851 		return;
1852 
1853 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
1854 
1855 	if ((lockflags & KNF_LISTLOCKED) == 0)
1856 		list->kl_lock(list->kl_lockarg);
1857 
1858 	/*
1859 	 * If we unlock the list lock (and set KN_INFLUX), we can eliminate
1860 	 * the kqueue scheduling, but this will introduce four
1861 	 * lock/unlock's for each knote to test.  If we do, continue to use
1862 	 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is
1863 	 * only safe if you want to remove the current item, which we are
1864 	 * not doing.
1865 	 */
1866 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
1867 		kq = kn->kn_kq;
1868 		if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) {
1869 			KQ_LOCK(kq);
1870 			if ((kn->kn_status & KN_INFLUX) == KN_INFLUX) {
1871 				KQ_UNLOCK(kq);
1872 			} else if ((lockflags & KNF_NOKQLOCK) != 0) {
1873 				kn->kn_status |= KN_INFLUX;
1874 				KQ_UNLOCK(kq);
1875 				error = kn->kn_fop->f_event(kn, hint);
1876 				KQ_LOCK(kq);
1877 				kn->kn_status &= ~KN_INFLUX;
1878 				if (error)
1879 					KNOTE_ACTIVATE(kn, 1);
1880 				KQ_UNLOCK_FLUX(kq);
1881 			} else {
1882 				kn->kn_status |= KN_HASKQLOCK;
1883 				if (kn->kn_fop->f_event(kn, hint))
1884 					KNOTE_ACTIVATE(kn, 1);
1885 				kn->kn_status &= ~KN_HASKQLOCK;
1886 				KQ_UNLOCK(kq);
1887 			}
1888 		}
1889 		kq = NULL;
1890 	}
1891 	if ((lockflags & KNF_LISTLOCKED) == 0)
1892 		list->kl_unlock(list->kl_lockarg);
1893 }
1894 
1895 /*
1896  * add a knote to a knlist
1897  */
1898 void
1899 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
1900 {
1901 	KNL_ASSERT_LOCK(knl, islocked);
1902 	KQ_NOTOWNED(kn->kn_kq);
1903 	KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) ==
1904 	    (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED"));
1905 	if (!islocked)
1906 		knl->kl_lock(knl->kl_lockarg);
1907 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
1908 	if (!islocked)
1909 		knl->kl_unlock(knl->kl_lockarg);
1910 	KQ_LOCK(kn->kn_kq);
1911 	kn->kn_knlist = knl;
1912 	kn->kn_status &= ~KN_DETACHED;
1913 	KQ_UNLOCK(kn->kn_kq);
1914 }
1915 
1916 static void
1917 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked)
1918 {
1919 	KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked"));
1920 	KNL_ASSERT_LOCK(knl, knlislocked);
1921 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
1922 	if (!kqislocked)
1923 		KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX,
1924     ("knlist_remove called w/o knote being KN_INFLUX or already removed"));
1925 	if (!knlislocked)
1926 		knl->kl_lock(knl->kl_lockarg);
1927 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
1928 	kn->kn_knlist = NULL;
1929 	if (!knlislocked)
1930 		knl->kl_unlock(knl->kl_lockarg);
1931 	if (!kqislocked)
1932 		KQ_LOCK(kn->kn_kq);
1933 	kn->kn_status |= KN_DETACHED;
1934 	if (!kqislocked)
1935 		KQ_UNLOCK(kn->kn_kq);
1936 }
1937 
1938 /*
1939  * remove knote from the specified knlist
1940  */
1941 void
1942 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
1943 {
1944 
1945 	knlist_remove_kq(knl, kn, islocked, 0);
1946 }
1947 
1948 /*
1949  * remove knote from the specified knlist while in f_event handler.
1950  */
1951 void
1952 knlist_remove_inevent(struct knlist *knl, struct knote *kn)
1953 {
1954 
1955 	knlist_remove_kq(knl, kn, 1,
1956 	    (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK);
1957 }
1958 
1959 int
1960 knlist_empty(struct knlist *knl)
1961 {
1962 
1963 	KNL_ASSERT_LOCKED(knl);
1964 	return SLIST_EMPTY(&knl->kl_list);
1965 }
1966 
1967 static struct mtx	knlist_lock;
1968 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
1969 	MTX_DEF);
1970 static void knlist_mtx_lock(void *arg);
1971 static void knlist_mtx_unlock(void *arg);
1972 
1973 static void
1974 knlist_mtx_lock(void *arg)
1975 {
1976 
1977 	mtx_lock((struct mtx *)arg);
1978 }
1979 
1980 static void
1981 knlist_mtx_unlock(void *arg)
1982 {
1983 
1984 	mtx_unlock((struct mtx *)arg);
1985 }
1986 
1987 static void
1988 knlist_mtx_assert_locked(void *arg)
1989 {
1990 
1991 	mtx_assert((struct mtx *)arg, MA_OWNED);
1992 }
1993 
1994 static void
1995 knlist_mtx_assert_unlocked(void *arg)
1996 {
1997 
1998 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
1999 }
2000 
2001 static void
2002 knlist_rw_rlock(void *arg)
2003 {
2004 
2005 	rw_rlock((struct rwlock *)arg);
2006 }
2007 
2008 static void
2009 knlist_rw_runlock(void *arg)
2010 {
2011 
2012 	rw_runlock((struct rwlock *)arg);
2013 }
2014 
2015 static void
2016 knlist_rw_assert_locked(void *arg)
2017 {
2018 
2019 	rw_assert((struct rwlock *)arg, RA_LOCKED);
2020 }
2021 
2022 static void
2023 knlist_rw_assert_unlocked(void *arg)
2024 {
2025 
2026 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2027 }
2028 
2029 void
2030 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2031     void (*kl_unlock)(void *),
2032     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2033 {
2034 
2035 	if (lock == NULL)
2036 		knl->kl_lockarg = &knlist_lock;
2037 	else
2038 		knl->kl_lockarg = lock;
2039 
2040 	if (kl_lock == NULL)
2041 		knl->kl_lock = knlist_mtx_lock;
2042 	else
2043 		knl->kl_lock = kl_lock;
2044 	if (kl_unlock == NULL)
2045 		knl->kl_unlock = knlist_mtx_unlock;
2046 	else
2047 		knl->kl_unlock = kl_unlock;
2048 	if (kl_assert_locked == NULL)
2049 		knl->kl_assert_locked = knlist_mtx_assert_locked;
2050 	else
2051 		knl->kl_assert_locked = kl_assert_locked;
2052 	if (kl_assert_unlocked == NULL)
2053 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2054 	else
2055 		knl->kl_assert_unlocked = kl_assert_unlocked;
2056 
2057 	SLIST_INIT(&knl->kl_list);
2058 }
2059 
2060 void
2061 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2062 {
2063 
2064 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2065 }
2066 
2067 void
2068 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2069 {
2070 
2071 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2072 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2073 }
2074 
2075 void
2076 knlist_destroy(struct knlist *knl)
2077 {
2078 
2079 #ifdef INVARIANTS
2080 	/*
2081 	 * if we run across this error, we need to find the offending
2082 	 * driver and have it call knlist_clear or knlist_delete.
2083 	 */
2084 	if (!SLIST_EMPTY(&knl->kl_list))
2085 		printf("WARNING: destroying knlist w/ knotes on it!\n");
2086 #endif
2087 
2088 	knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL;
2089 	SLIST_INIT(&knl->kl_list);
2090 }
2091 
2092 /*
2093  * Even if we are locked, we may need to drop the lock to allow any influx
2094  * knotes time to "settle".
2095  */
2096 void
2097 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2098 {
2099 	struct knote *kn, *kn2;
2100 	struct kqueue *kq;
2101 
2102 	if (islocked)
2103 		KNL_ASSERT_LOCKED(knl);
2104 	else {
2105 		KNL_ASSERT_UNLOCKED(knl);
2106 again:		/* need to reacquire lock since we have dropped it */
2107 		knl->kl_lock(knl->kl_lockarg);
2108 	}
2109 
2110 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2111 		kq = kn->kn_kq;
2112 		KQ_LOCK(kq);
2113 		if ((kn->kn_status & KN_INFLUX)) {
2114 			KQ_UNLOCK(kq);
2115 			continue;
2116 		}
2117 		knlist_remove_kq(knl, kn, 1, 1);
2118 		if (killkn) {
2119 			kn->kn_status |= KN_INFLUX | KN_DETACHED;
2120 			KQ_UNLOCK(kq);
2121 			knote_drop(kn, td);
2122 		} else {
2123 			/* Make sure cleared knotes disappear soon */
2124 			kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2125 			KQ_UNLOCK(kq);
2126 		}
2127 		kq = NULL;
2128 	}
2129 
2130 	if (!SLIST_EMPTY(&knl->kl_list)) {
2131 		/* there are still KN_INFLUX remaining */
2132 		kn = SLIST_FIRST(&knl->kl_list);
2133 		kq = kn->kn_kq;
2134 		KQ_LOCK(kq);
2135 		KASSERT(kn->kn_status & KN_INFLUX,
2136 		    ("knote removed w/o list lock"));
2137 		knl->kl_unlock(knl->kl_lockarg);
2138 		kq->kq_state |= KQ_FLUXWAIT;
2139 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2140 		kq = NULL;
2141 		goto again;
2142 	}
2143 
2144 	if (islocked)
2145 		KNL_ASSERT_LOCKED(knl);
2146 	else {
2147 		knl->kl_unlock(knl->kl_lockarg);
2148 		KNL_ASSERT_UNLOCKED(knl);
2149 	}
2150 }
2151 
2152 /*
2153  * Remove all knotes referencing a specified fd must be called with FILEDESC
2154  * lock.  This prevents a race where a new fd comes along and occupies the
2155  * entry and we attach a knote to the fd.
2156  */
2157 void
2158 knote_fdclose(struct thread *td, int fd)
2159 {
2160 	struct filedesc *fdp = td->td_proc->p_fd;
2161 	struct kqueue *kq;
2162 	struct knote *kn;
2163 	int influx;
2164 
2165 	FILEDESC_XLOCK_ASSERT(fdp);
2166 
2167 	/*
2168 	 * We shouldn't have to worry about new kevents appearing on fd
2169 	 * since filedesc is locked.
2170 	 */
2171 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2172 		KQ_LOCK(kq);
2173 
2174 again:
2175 		influx = 0;
2176 		while (kq->kq_knlistsize > fd &&
2177 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2178 			if (kn->kn_status & KN_INFLUX) {
2179 				/* someone else might be waiting on our knote */
2180 				if (influx)
2181 					wakeup(kq);
2182 				kq->kq_state |= KQ_FLUXWAIT;
2183 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2184 				goto again;
2185 			}
2186 			kn->kn_status |= KN_INFLUX;
2187 			KQ_UNLOCK(kq);
2188 			if (!(kn->kn_status & KN_DETACHED))
2189 				kn->kn_fop->f_detach(kn);
2190 			knote_drop(kn, td);
2191 			influx = 1;
2192 			KQ_LOCK(kq);
2193 		}
2194 		KQ_UNLOCK_FLUX(kq);
2195 	}
2196 }
2197 
2198 static int
2199 knote_attach(struct knote *kn, struct kqueue *kq)
2200 {
2201 	struct klist *list;
2202 
2203 	KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX"));
2204 	KQ_OWNED(kq);
2205 
2206 	if (kn->kn_fop->f_isfd) {
2207 		if (kn->kn_id >= kq->kq_knlistsize)
2208 			return ENOMEM;
2209 		list = &kq->kq_knlist[kn->kn_id];
2210 	} else {
2211 		if (kq->kq_knhash == NULL)
2212 			return ENOMEM;
2213 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2214 	}
2215 
2216 	SLIST_INSERT_HEAD(list, kn, kn_link);
2217 
2218 	return 0;
2219 }
2220 
2221 /*
2222  * knote must already have been detached using the f_detach method.
2223  * no lock need to be held, it is assumed that the KN_INFLUX flag is set
2224  * to prevent other removal.
2225  */
2226 static void
2227 knote_drop(struct knote *kn, struct thread *td)
2228 {
2229 	struct kqueue *kq;
2230 	struct klist *list;
2231 
2232 	kq = kn->kn_kq;
2233 
2234 	KQ_NOTOWNED(kq);
2235 	KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX,
2236 	    ("knote_drop called without KN_INFLUX set in kn_status"));
2237 
2238 	KQ_LOCK(kq);
2239 	if (kn->kn_fop->f_isfd)
2240 		list = &kq->kq_knlist[kn->kn_id];
2241 	else
2242 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2243 
2244 	if (!SLIST_EMPTY(list))
2245 		SLIST_REMOVE(list, kn, knote, kn_link);
2246 	if (kn->kn_status & KN_QUEUED)
2247 		knote_dequeue(kn);
2248 	KQ_UNLOCK_FLUX(kq);
2249 
2250 	if (kn->kn_fop->f_isfd) {
2251 		fdrop(kn->kn_fp, td);
2252 		kn->kn_fp = NULL;
2253 	}
2254 	kqueue_fo_release(kn->kn_kevent.filter);
2255 	kn->kn_fop = NULL;
2256 	knote_free(kn);
2257 }
2258 
2259 static void
2260 knote_enqueue(struct knote *kn)
2261 {
2262 	struct kqueue *kq = kn->kn_kq;
2263 
2264 	KQ_OWNED(kn->kn_kq);
2265 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2266 
2267 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2268 	kn->kn_status |= KN_QUEUED;
2269 	kq->kq_count++;
2270 	kqueue_wakeup(kq);
2271 }
2272 
2273 static void
2274 knote_dequeue(struct knote *kn)
2275 {
2276 	struct kqueue *kq = kn->kn_kq;
2277 
2278 	KQ_OWNED(kn->kn_kq);
2279 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2280 
2281 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2282 	kn->kn_status &= ~KN_QUEUED;
2283 	kq->kq_count--;
2284 }
2285 
2286 static void
2287 knote_init(void)
2288 {
2289 
2290 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2291 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2292 }
2293 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2294 
2295 static struct knote *
2296 knote_alloc(int waitok)
2297 {
2298 	return ((struct knote *)uma_zalloc(knote_zone,
2299 	    (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO));
2300 }
2301 
2302 static void
2303 knote_free(struct knote *kn)
2304 {
2305 	if (kn != NULL)
2306 		uma_zfree(knote_zone, kn);
2307 }
2308 
2309 /*
2310  * Register the kev w/ the kq specified by fd.
2311  */
2312 int
2313 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2314 {
2315 	struct kqueue *kq;
2316 	struct file *fp;
2317 	cap_rights_t rights;
2318 	int error;
2319 
2320 	error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2321 	if (error != 0)
2322 		return (error);
2323 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2324 		goto noacquire;
2325 
2326 	error = kqueue_register(kq, kev, td, waitok);
2327 
2328 	kqueue_release(kq, 0);
2329 
2330 noacquire:
2331 	fdrop(fp, td);
2332 
2333 	return error;
2334 }
2335