xref: /freebsd/sys/kern/kern_event.c (revision 0183e0151669735d62584fbba9125ed90716af5e)
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/queue.h>
52 #include <sys/event.h>
53 #include <sys/eventvar.h>
54 #include <sys/poll.h>
55 #include <sys/protosw.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sigio.h>
58 #include <sys/signalvar.h>
59 #include <sys/socket.h>
60 #include <sys/socketvar.h>
61 #include <sys/stat.h>
62 #include <sys/sysctl.h>
63 #include <sys/sysproto.h>
64 #include <sys/syscallsubr.h>
65 #include <sys/taskqueue.h>
66 #include <sys/uio.h>
67 #include <sys/user.h>
68 #ifdef KTRACE
69 #include <sys/ktrace.h>
70 #endif
71 #include <machine/atomic.h>
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_ctx);
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_drop_detached(struct knote *kn, struct thread *td);
140 static void 	knote_enqueue(struct knote *kn);
141 static void 	knote_dequeue(struct knote *kn);
142 static void 	knote_init(void);
143 static struct 	knote *knote_alloc(int waitok);
144 static void 	knote_free(struct knote *kn);
145 
146 static void	filt_kqdetach(struct knote *kn);
147 static int	filt_kqueue(struct knote *kn, long hint);
148 static int	filt_procattach(struct knote *kn);
149 static void	filt_procdetach(struct knote *kn);
150 static int	filt_proc(struct knote *kn, long hint);
151 static int	filt_fileattach(struct knote *kn);
152 static void	filt_timerexpire(void *knx);
153 static int	filt_timerattach(struct knote *kn);
154 static void	filt_timerdetach(struct knote *kn);
155 static int	filt_timer(struct knote *kn, long hint);
156 static int	filt_userattach(struct knote *kn);
157 static void	filt_userdetach(struct knote *kn);
158 static int	filt_user(struct knote *kn, long hint);
159 static void	filt_usertouch(struct knote *kn, struct kevent *kev,
160 		    u_long type);
161 
162 static struct filterops file_filtops = {
163 	.f_isfd = 1,
164 	.f_attach = filt_fileattach,
165 };
166 static struct filterops kqread_filtops = {
167 	.f_isfd = 1,
168 	.f_detach = filt_kqdetach,
169 	.f_event = filt_kqueue,
170 };
171 /* XXX - move to kern_proc.c?  */
172 static struct filterops proc_filtops = {
173 	.f_isfd = 0,
174 	.f_attach = filt_procattach,
175 	.f_detach = filt_procdetach,
176 	.f_event = filt_proc,
177 };
178 static struct filterops timer_filtops = {
179 	.f_isfd = 0,
180 	.f_attach = filt_timerattach,
181 	.f_detach = filt_timerdetach,
182 	.f_event = filt_timer,
183 };
184 static struct filterops user_filtops = {
185 	.f_attach = filt_userattach,
186 	.f_detach = filt_userdetach,
187 	.f_event = filt_user,
188 	.f_touch = filt_usertouch,
189 };
190 
191 static uma_zone_t	knote_zone;
192 static unsigned int	kq_ncallouts = 0;
193 static unsigned int 	kq_calloutmax = 4 * 1024;
194 SYSCTL_UINT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
195     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
196 
197 /* XXX - ensure not influx ? */
198 #define KNOTE_ACTIVATE(kn, islock) do { 				\
199 	if ((islock))							\
200 		mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED);		\
201 	else								\
202 		KQ_LOCK((kn)->kn_kq);					\
203 	(kn)->kn_status |= KN_ACTIVE;					\
204 	if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)		\
205 		knote_enqueue((kn));					\
206 	if (!(islock))							\
207 		KQ_UNLOCK((kn)->kn_kq);					\
208 } while(0)
209 #define KQ_LOCK(kq) do {						\
210 	mtx_lock(&(kq)->kq_lock);					\
211 } while (0)
212 #define KQ_FLUX_WAKEUP(kq) do {						\
213 	if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) {		\
214 		(kq)->kq_state &= ~KQ_FLUXWAIT;				\
215 		wakeup((kq));						\
216 	}								\
217 } while (0)
218 #define KQ_UNLOCK_FLUX(kq) do {						\
219 	KQ_FLUX_WAKEUP(kq);						\
220 	mtx_unlock(&(kq)->kq_lock);					\
221 } while (0)
222 #define KQ_UNLOCK(kq) do {						\
223 	mtx_unlock(&(kq)->kq_lock);					\
224 } while (0)
225 #define KQ_OWNED(kq) do {						\
226 	mtx_assert(&(kq)->kq_lock, MA_OWNED);				\
227 } while (0)
228 #define KQ_NOTOWNED(kq) do {						\
229 	mtx_assert(&(kq)->kq_lock, MA_NOTOWNED);			\
230 } while (0)
231 
232 static struct knlist *
233 kn_list_lock(struct knote *kn)
234 {
235 	struct knlist *knl;
236 
237 	knl = kn->kn_knlist;
238 	if (knl != NULL)
239 		knl->kl_lock(knl->kl_lockarg);
240 	return (knl);
241 }
242 
243 static void
244 kn_list_unlock(struct knlist *knl)
245 {
246 	bool do_free;
247 
248 	if (knl == NULL)
249 		return;
250 	do_free = knl->kl_autodestroy && knlist_empty(knl);
251 	knl->kl_unlock(knl->kl_lockarg);
252 	if (do_free) {
253 		knlist_destroy(knl);
254 		free(knl, M_KQUEUE);
255 	}
256 }
257 
258 static bool
259 kn_in_flux(struct knote *kn)
260 {
261 
262 	return (kn->kn_influx > 0);
263 }
264 
265 static void
266 kn_enter_flux(struct knote *kn)
267 {
268 
269 	KQ_OWNED(kn->kn_kq);
270 	MPASS(kn->kn_influx < INT_MAX);
271 	kn->kn_influx++;
272 }
273 
274 static bool
275 kn_leave_flux(struct knote *kn)
276 {
277 
278 	KQ_OWNED(kn->kn_kq);
279 	MPASS(kn->kn_influx > 0);
280 	kn->kn_influx--;
281 	return (kn->kn_influx == 0);
282 }
283 
284 #define	KNL_ASSERT_LOCK(knl, islocked) do {				\
285 	if (islocked)							\
286 		KNL_ASSERT_LOCKED(knl);				\
287 	else								\
288 		KNL_ASSERT_UNLOCKED(knl);				\
289 } while (0)
290 #ifdef INVARIANTS
291 #define	KNL_ASSERT_LOCKED(knl) do {					\
292 	knl->kl_assert_locked((knl)->kl_lockarg);			\
293 } while (0)
294 #define	KNL_ASSERT_UNLOCKED(knl) do {					\
295 	knl->kl_assert_unlocked((knl)->kl_lockarg);			\
296 } while (0)
297 #else /* !INVARIANTS */
298 #define	KNL_ASSERT_LOCKED(knl) do {} while(0)
299 #define	KNL_ASSERT_UNLOCKED(knl) do {} while (0)
300 #endif /* INVARIANTS */
301 
302 #ifndef	KN_HASHSIZE
303 #define	KN_HASHSIZE		64		/* XXX should be tunable */
304 #endif
305 
306 #define KN_HASH(val, mask)	(((val) ^ (val >> 8)) & (mask))
307 
308 static int
309 filt_nullattach(struct knote *kn)
310 {
311 
312 	return (ENXIO);
313 };
314 
315 struct filterops null_filtops = {
316 	.f_isfd = 0,
317 	.f_attach = filt_nullattach,
318 };
319 
320 /* XXX - make SYSINIT to add these, and move into respective modules. */
321 extern struct filterops sig_filtops;
322 extern struct filterops fs_filtops;
323 
324 /*
325  * Table for for all system-defined filters.
326  */
327 static struct mtx	filterops_lock;
328 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops",
329 	MTX_DEF);
330 static struct {
331 	struct filterops *for_fop;
332 	int for_nolock;
333 	int for_refcnt;
334 } sysfilt_ops[EVFILT_SYSCOUNT] = {
335 	{ &file_filtops, 1 },			/* EVFILT_READ */
336 	{ &file_filtops, 1 },			/* EVFILT_WRITE */
337 	{ &null_filtops },			/* EVFILT_AIO */
338 	{ &file_filtops, 1 },			/* EVFILT_VNODE */
339 	{ &proc_filtops, 1 },			/* EVFILT_PROC */
340 	{ &sig_filtops, 1 },			/* EVFILT_SIGNAL */
341 	{ &timer_filtops, 1 },			/* EVFILT_TIMER */
342 	{ &file_filtops, 1 },			/* EVFILT_PROCDESC */
343 	{ &fs_filtops, 1 },			/* EVFILT_FS */
344 	{ &null_filtops },			/* EVFILT_LIO */
345 	{ &user_filtops, 1 },			/* EVFILT_USER */
346 	{ &null_filtops },			/* EVFILT_SENDFILE */
347 	{ &file_filtops, 1 },                   /* EVFILT_EMPTY */
348 };
349 
350 /*
351  * Simple redirection for all cdevsw style objects to call their fo_kqfilter
352  * method.
353  */
354 static int
355 filt_fileattach(struct knote *kn)
356 {
357 
358 	return (fo_kqfilter(kn->kn_fp, kn));
359 }
360 
361 /*ARGSUSED*/
362 static int
363 kqueue_kqfilter(struct file *fp, struct knote *kn)
364 {
365 	struct kqueue *kq = kn->kn_fp->f_data;
366 
367 	if (kn->kn_filter != EVFILT_READ)
368 		return (EINVAL);
369 
370 	kn->kn_status |= KN_KQUEUE;
371 	kn->kn_fop = &kqread_filtops;
372 	knlist_add(&kq->kq_sel.si_note, kn, 0);
373 
374 	return (0);
375 }
376 
377 static void
378 filt_kqdetach(struct knote *kn)
379 {
380 	struct kqueue *kq = kn->kn_fp->f_data;
381 
382 	knlist_remove(&kq->kq_sel.si_note, kn, 0);
383 }
384 
385 /*ARGSUSED*/
386 static int
387 filt_kqueue(struct knote *kn, long hint)
388 {
389 	struct kqueue *kq = kn->kn_fp->f_data;
390 
391 	kn->kn_data = kq->kq_count;
392 	return (kn->kn_data > 0);
393 }
394 
395 /* XXX - move to kern_proc.c?  */
396 static int
397 filt_procattach(struct knote *kn)
398 {
399 	struct proc *p;
400 	int error;
401 	bool exiting, immediate;
402 
403 	exiting = immediate = false;
404 	p = pfind(kn->kn_id);
405 	if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
406 		p = zpfind(kn->kn_id);
407 		exiting = true;
408 	} else if (p != NULL && (p->p_flag & P_WEXIT)) {
409 		exiting = true;
410 	}
411 
412 	if (p == NULL)
413 		return (ESRCH);
414 	if ((error = p_cansee(curthread, p))) {
415 		PROC_UNLOCK(p);
416 		return (error);
417 	}
418 
419 	kn->kn_ptr.p_proc = p;
420 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
421 
422 	/*
423 	 * Internal flag indicating registration done by kernel for the
424 	 * purposes of getting a NOTE_CHILD notification.
425 	 */
426 	if (kn->kn_flags & EV_FLAG2) {
427 		kn->kn_flags &= ~EV_FLAG2;
428 		kn->kn_data = kn->kn_sdata;		/* ppid */
429 		kn->kn_fflags = NOTE_CHILD;
430 		kn->kn_sfflags &= ~(NOTE_EXIT | NOTE_EXEC | NOTE_FORK);
431 		immediate = true; /* Force immediate activation of child note. */
432 	}
433 	/*
434 	 * Internal flag indicating registration done by kernel (for other than
435 	 * NOTE_CHILD).
436 	 */
437 	if (kn->kn_flags & EV_FLAG1) {
438 		kn->kn_flags &= ~EV_FLAG1;
439 	}
440 
441 	knlist_add(p->p_klist, kn, 1);
442 
443 	/*
444 	 * Immediately activate any child notes or, in the case of a zombie
445 	 * target process, exit notes.  The latter is necessary to handle the
446 	 * case where the target process, e.g. a child, dies before the kevent
447 	 * is registered.
448 	 */
449 	if (immediate || (exiting && filt_proc(kn, NOTE_EXIT)))
450 		KNOTE_ACTIVATE(kn, 0);
451 
452 	PROC_UNLOCK(p);
453 
454 	return (0);
455 }
456 
457 /*
458  * The knote may be attached to a different process, which may exit,
459  * leaving nothing for the knote to be attached to.  So when the process
460  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
461  * it will be deleted when read out.  However, as part of the knote deletion,
462  * this routine is called, so a check is needed to avoid actually performing
463  * a detach, because the original process does not exist any more.
464  */
465 /* XXX - move to kern_proc.c?  */
466 static void
467 filt_procdetach(struct knote *kn)
468 {
469 
470 	knlist_remove(kn->kn_knlist, kn, 0);
471 	kn->kn_ptr.p_proc = NULL;
472 }
473 
474 /* XXX - move to kern_proc.c?  */
475 static int
476 filt_proc(struct knote *kn, long hint)
477 {
478 	struct proc *p;
479 	u_int event;
480 
481 	p = kn->kn_ptr.p_proc;
482 	if (p == NULL) /* already activated, from attach filter */
483 		return (0);
484 
485 	/* Mask off extra data. */
486 	event = (u_int)hint & NOTE_PCTRLMASK;
487 
488 	/* If the user is interested in this event, record it. */
489 	if (kn->kn_sfflags & event)
490 		kn->kn_fflags |= event;
491 
492 	/* Process is gone, so flag the event as finished. */
493 	if (event == NOTE_EXIT) {
494 		kn->kn_flags |= EV_EOF | EV_ONESHOT;
495 		kn->kn_ptr.p_proc = NULL;
496 		if (kn->kn_fflags & NOTE_EXIT)
497 			kn->kn_data = KW_EXITCODE(p->p_xexit, p->p_xsig);
498 		if (kn->kn_fflags == 0)
499 			kn->kn_flags |= EV_DROP;
500 		return (1);
501 	}
502 
503 	return (kn->kn_fflags != 0);
504 }
505 
506 /*
507  * Called when the process forked. It mostly does the same as the
508  * knote(), activating all knotes registered to be activated when the
509  * process forked. Additionally, for each knote attached to the
510  * parent, check whether user wants to track the new process. If so
511  * attach a new knote to it, and immediately report an event with the
512  * child's pid.
513  */
514 void
515 knote_fork(struct knlist *list, int pid)
516 {
517 	struct kqueue *kq;
518 	struct knote *kn;
519 	struct kevent kev;
520 	int error;
521 
522 	if (list == NULL)
523 		return;
524 	list->kl_lock(list->kl_lockarg);
525 
526 	SLIST_FOREACH(kn, &list->kl_list, kn_selnext) {
527 		kq = kn->kn_kq;
528 		KQ_LOCK(kq);
529 		if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
530 			KQ_UNLOCK(kq);
531 			continue;
532 		}
533 
534 		/*
535 		 * The same as knote(), activate the event.
536 		 */
537 		if ((kn->kn_sfflags & NOTE_TRACK) == 0) {
538 			kn->kn_status |= KN_HASKQLOCK;
539 			if (kn->kn_fop->f_event(kn, NOTE_FORK))
540 				KNOTE_ACTIVATE(kn, 1);
541 			kn->kn_status &= ~KN_HASKQLOCK;
542 			KQ_UNLOCK(kq);
543 			continue;
544 		}
545 
546 		/*
547 		 * The NOTE_TRACK case. In addition to the activation
548 		 * of the event, we need to register new events to
549 		 * track the child. Drop the locks in preparation for
550 		 * the call to kqueue_register().
551 		 */
552 		kn_enter_flux(kn);
553 		KQ_UNLOCK(kq);
554 		list->kl_unlock(list->kl_lockarg);
555 
556 		/*
557 		 * Activate existing knote and register tracking knotes with
558 		 * new process.
559 		 *
560 		 * First register a knote to get just the child notice. This
561 		 * must be a separate note from a potential NOTE_EXIT
562 		 * notification since both NOTE_CHILD and NOTE_EXIT are defined
563 		 * to use the data field (in conflicting ways).
564 		 */
565 		kev.ident = pid;
566 		kev.filter = kn->kn_filter;
567 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_ONESHOT |
568 		    EV_FLAG2;
569 		kev.fflags = kn->kn_sfflags;
570 		kev.data = kn->kn_id;		/* parent */
571 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
572 		error = kqueue_register(kq, &kev, NULL, 0);
573 		if (error)
574 			kn->kn_fflags |= NOTE_TRACKERR;
575 
576 		/*
577 		 * Then register another knote to track other potential events
578 		 * from the new process.
579 		 */
580 		kev.ident = pid;
581 		kev.filter = kn->kn_filter;
582 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
583 		kev.fflags = kn->kn_sfflags;
584 		kev.data = kn->kn_id;		/* parent */
585 		kev.udata = kn->kn_kevent.udata;/* preserve udata */
586 		error = kqueue_register(kq, &kev, NULL, 0);
587 		if (error)
588 			kn->kn_fflags |= NOTE_TRACKERR;
589 		if (kn->kn_fop->f_event(kn, NOTE_FORK))
590 			KNOTE_ACTIVATE(kn, 0);
591 		KQ_LOCK(kq);
592 		kn_leave_flux(kn);
593 		KQ_UNLOCK_FLUX(kq);
594 		list->kl_lock(list->kl_lockarg);
595 	}
596 	list->kl_unlock(list->kl_lockarg);
597 }
598 
599 /*
600  * XXX: EVFILT_TIMER should perhaps live in kern_time.c beside the
601  * interval timer support code.
602  */
603 
604 #define NOTE_TIMER_PRECMASK	(NOTE_SECONDS|NOTE_MSECONDS|NOTE_USECONDS| \
605 				NOTE_NSECONDS)
606 
607 static sbintime_t
608 timer2sbintime(intptr_t data, int flags)
609 {
610 
611         /*
612          * Macros for converting to the fractional second portion of an
613          * sbintime_t using 64bit multiplication to improve precision.
614          */
615 #define NS_TO_SBT(ns) (((ns) * (((uint64_t)1 << 63) / 500000000)) >> 32)
616 #define US_TO_SBT(us) (((us) * (((uint64_t)1 << 63) / 500000)) >> 32)
617 #define MS_TO_SBT(ms) (((ms) * (((uint64_t)1 << 63) / 500)) >> 32)
618 	switch (flags & NOTE_TIMER_PRECMASK) {
619 	case NOTE_SECONDS:
620 #ifdef __LP64__
621 		if (data > (SBT_MAX / SBT_1S))
622 			return (SBT_MAX);
623 #endif
624 		return ((sbintime_t)data << 32);
625 	case NOTE_MSECONDS: /* FALLTHROUGH */
626 	case 0:
627 		if (data >= 1000) {
628 			int64_t secs = data / 1000;
629 #ifdef __LP64__
630 			if (secs > (SBT_MAX / SBT_1S))
631 				return (SBT_MAX);
632 #endif
633 			return (secs << 32 | MS_TO_SBT(data % 1000));
634 		}
635 		return MS_TO_SBT(data);
636 	case NOTE_USECONDS:
637 		if (data >= 1000000) {
638 			int64_t secs = data / 1000000;
639 #ifdef __LP64__
640 			if (secs > (SBT_MAX / SBT_1S))
641 				return (SBT_MAX);
642 #endif
643 			return (secs << 32 | US_TO_SBT(data % 1000000));
644 		}
645 		return US_TO_SBT(data);
646 	case NOTE_NSECONDS:
647 		if (data >= 1000000000) {
648 			int64_t secs = data / 1000000000;
649 #ifdef __LP64__
650 			if (secs > (SBT_MAX / SBT_1S))
651 				return (SBT_MAX);
652 #endif
653 			return (secs << 32 | US_TO_SBT(data % 1000000000));
654 		}
655 		return (NS_TO_SBT(data));
656 	default:
657 		break;
658 	}
659 	return (-1);
660 }
661 
662 struct kq_timer_cb_data {
663 	struct callout c;
664 	sbintime_t next;	/* next timer event fires at */
665 	sbintime_t to;		/* precalculated timer period */
666 };
667 
668 static void
669 filt_timerexpire(void *knx)
670 {
671 	struct knote *kn;
672 	struct kq_timer_cb_data *kc;
673 
674 	kn = knx;
675 	kn->kn_data++;
676 	KNOTE_ACTIVATE(kn, 0);	/* XXX - handle locking */
677 
678 	if ((kn->kn_flags & EV_ONESHOT) != 0)
679 		return;
680 
681 	kc = kn->kn_ptr.p_v;
682 	kc->next += kc->to;
683 	callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
684 	    PCPU_GET(cpuid), C_ABSOLUTE);
685 }
686 
687 /*
688  * data contains amount of time to sleep
689  */
690 static int
691 filt_timerattach(struct knote *kn)
692 {
693 	struct kq_timer_cb_data *kc;
694 	sbintime_t to;
695 	unsigned int ncallouts;
696 
697 	if (kn->kn_sdata < 0)
698 		return (EINVAL);
699 	if (kn->kn_sdata == 0 && (kn->kn_flags & EV_ONESHOT) == 0)
700 		kn->kn_sdata = 1;
701 	/* Only precision unit are supported in flags so far */
702 	if ((kn->kn_sfflags & ~NOTE_TIMER_PRECMASK) != 0)
703 		return (EINVAL);
704 
705 	to = timer2sbintime(kn->kn_sdata, kn->kn_sfflags);
706 	if (to < 0)
707 		return (EINVAL);
708 
709 	do {
710 		ncallouts = kq_ncallouts;
711 		if (ncallouts >= kq_calloutmax)
712 			return (ENOMEM);
713 	} while (!atomic_cmpset_int(&kq_ncallouts, ncallouts, ncallouts + 1));
714 
715 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
716 	kn->kn_status &= ~KN_DETACHED;		/* knlist_add clears it */
717 	kn->kn_ptr.p_v = kc = malloc(sizeof(*kc), M_KQUEUE, M_WAITOK);
718 	callout_init(&kc->c, 1);
719 	kc->next = to + sbinuptime();
720 	kc->to = to;
721 	callout_reset_sbt_on(&kc->c, kc->next, 0, filt_timerexpire, kn,
722 	    PCPU_GET(cpuid), C_ABSOLUTE);
723 
724 	return (0);
725 }
726 
727 static void
728 filt_timerdetach(struct knote *kn)
729 {
730 	struct kq_timer_cb_data *kc;
731 	unsigned int old;
732 
733 	kc = kn->kn_ptr.p_v;
734 	callout_drain(&kc->c);
735 	free(kc, M_KQUEUE);
736 	old = atomic_fetchadd_int(&kq_ncallouts, -1);
737 	KASSERT(old > 0, ("Number of callouts cannot become negative"));
738 	kn->kn_status |= KN_DETACHED;	/* knlist_remove sets it */
739 }
740 
741 static int
742 filt_timer(struct knote *kn, long hint)
743 {
744 
745 	return (kn->kn_data != 0);
746 }
747 
748 static int
749 filt_userattach(struct knote *kn)
750 {
751 
752 	/*
753 	 * EVFILT_USER knotes are not attached to anything in the kernel.
754 	 */
755 	kn->kn_hook = NULL;
756 	if (kn->kn_fflags & NOTE_TRIGGER)
757 		kn->kn_hookid = 1;
758 	else
759 		kn->kn_hookid = 0;
760 	return (0);
761 }
762 
763 static void
764 filt_userdetach(__unused struct knote *kn)
765 {
766 
767 	/*
768 	 * EVFILT_USER knotes are not attached to anything in the kernel.
769 	 */
770 }
771 
772 static int
773 filt_user(struct knote *kn, __unused long hint)
774 {
775 
776 	return (kn->kn_hookid);
777 }
778 
779 static void
780 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
781 {
782 	u_int ffctrl;
783 
784 	switch (type) {
785 	case EVENT_REGISTER:
786 		if (kev->fflags & NOTE_TRIGGER)
787 			kn->kn_hookid = 1;
788 
789 		ffctrl = kev->fflags & NOTE_FFCTRLMASK;
790 		kev->fflags &= NOTE_FFLAGSMASK;
791 		switch (ffctrl) {
792 		case NOTE_FFNOP:
793 			break;
794 
795 		case NOTE_FFAND:
796 			kn->kn_sfflags &= kev->fflags;
797 			break;
798 
799 		case NOTE_FFOR:
800 			kn->kn_sfflags |= kev->fflags;
801 			break;
802 
803 		case NOTE_FFCOPY:
804 			kn->kn_sfflags = kev->fflags;
805 			break;
806 
807 		default:
808 			/* XXX Return error? */
809 			break;
810 		}
811 		kn->kn_sdata = kev->data;
812 		if (kev->flags & EV_CLEAR) {
813 			kn->kn_hookid = 0;
814 			kn->kn_data = 0;
815 			kn->kn_fflags = 0;
816 		}
817 		break;
818 
819         case EVENT_PROCESS:
820 		*kev = kn->kn_kevent;
821 		kev->fflags = kn->kn_sfflags;
822 		kev->data = kn->kn_sdata;
823 		if (kn->kn_flags & EV_CLEAR) {
824 			kn->kn_hookid = 0;
825 			kn->kn_data = 0;
826 			kn->kn_fflags = 0;
827 		}
828 		break;
829 
830 	default:
831 		panic("filt_usertouch() - invalid type (%ld)", type);
832 		break;
833 	}
834 }
835 
836 int
837 sys_kqueue(struct thread *td, struct kqueue_args *uap)
838 {
839 
840 	return (kern_kqueue(td, 0, NULL));
841 }
842 
843 static void
844 kqueue_init(struct kqueue *kq)
845 {
846 
847 	mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK);
848 	TAILQ_INIT(&kq->kq_head);
849 	knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock);
850 	TASK_INIT(&kq->kq_task, 0, kqueue_task, kq);
851 }
852 
853 int
854 kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps)
855 {
856 	struct filedesc *fdp;
857 	struct kqueue *kq;
858 	struct file *fp;
859 	struct ucred *cred;
860 	int fd, error;
861 
862 	fdp = td->td_proc->p_fd;
863 	cred = td->td_ucred;
864 	if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES)))
865 		return (ENOMEM);
866 
867 	error = falloc_caps(td, &fp, &fd, flags, fcaps);
868 	if (error != 0) {
869 		chgkqcnt(cred->cr_ruidinfo, -1, 0);
870 		return (error);
871 	}
872 
873 	/* An extra reference on `fp' has been held for us by falloc(). */
874 	kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO);
875 	kqueue_init(kq);
876 	kq->kq_fdp = fdp;
877 	kq->kq_cred = crhold(cred);
878 
879 	FILEDESC_XLOCK(fdp);
880 	TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list);
881 	FILEDESC_XUNLOCK(fdp);
882 
883 	finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops);
884 	fdrop(fp, td);
885 
886 	td->td_retval[0] = fd;
887 	return (0);
888 }
889 
890 #ifdef KTRACE
891 static size_t
892 kev_iovlen(int n, u_int kgio)
893 {
894 
895 	if (n < 0 || n >= kgio / sizeof(struct kevent))
896 		return (kgio);
897 	return (n * sizeof(struct kevent));
898 }
899 #endif
900 
901 #ifndef _SYS_SYSPROTO_H_
902 struct kevent_args {
903 	int	fd;
904 	const struct kevent *changelist;
905 	int	nchanges;
906 	struct	kevent *eventlist;
907 	int	nevents;
908 	const struct timespec *timeout;
909 };
910 #endif
911 int
912 sys_kevent(struct thread *td, struct kevent_args *uap)
913 {
914 	struct timespec ts, *tsp;
915 	struct kevent_copyops k_ops = {
916 		.arg = uap,
917 		.k_copyout = kevent_copyout,
918 		.k_copyin = kevent_copyin,
919 	};
920 	int error;
921 #ifdef KTRACE
922 	struct uio ktruio;
923 	struct iovec ktriov;
924 	struct uio *ktruioin = NULL;
925 	struct uio *ktruioout = NULL;
926 	u_int kgio;
927 #endif
928 
929 	if (uap->timeout != NULL) {
930 		error = copyin(uap->timeout, &ts, sizeof(ts));
931 		if (error)
932 			return (error);
933 		tsp = &ts;
934 	} else
935 		tsp = NULL;
936 
937 #ifdef KTRACE
938 	if (KTRPOINT(td, KTR_GENIO)) {
939 		kgio = ktr_geniosize;
940 		ktriov.iov_base = uap->changelist;
941 		ktriov.iov_len = kev_iovlen(uap->nchanges, kgio);
942 		ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1,
943 		    .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ,
944 		    .uio_td = td };
945 		ktruioin = cloneuio(&ktruio);
946 		ktriov.iov_base = uap->eventlist;
947 		ktriov.iov_len = kev_iovlen(uap->nevents, kgio);
948 		ktriov.iov_len = uap->nevents * sizeof(struct kevent);
949 		ktruioout = cloneuio(&ktruio);
950 	}
951 #endif
952 
953 	error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents,
954 	    &k_ops, tsp);
955 
956 #ifdef KTRACE
957 	if (ktruioin != NULL) {
958 		ktruioin->uio_resid = kev_iovlen(uap->nchanges, kgio);
959 		ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0);
960 		ktruioout->uio_resid = kev_iovlen(td->td_retval[0], kgio);
961 		ktrgenio(uap->fd, UIO_READ, ktruioout, error);
962 	}
963 #endif
964 
965 	return (error);
966 }
967 
968 /*
969  * Copy 'count' items into the destination list pointed to by uap->eventlist.
970  */
971 static int
972 kevent_copyout(void *arg, struct kevent *kevp, int count)
973 {
974 	struct kevent_args *uap;
975 	int error;
976 
977 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
978 	uap = (struct kevent_args *)arg;
979 
980 	error = copyout(kevp, uap->eventlist, count * sizeof *kevp);
981 	if (error == 0)
982 		uap->eventlist += count;
983 	return (error);
984 }
985 
986 /*
987  * Copy 'count' items from the list pointed to by uap->changelist.
988  */
989 static int
990 kevent_copyin(void *arg, struct kevent *kevp, int count)
991 {
992 	struct kevent_args *uap;
993 	int error;
994 
995 	KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count));
996 	uap = (struct kevent_args *)arg;
997 
998 	error = copyin(uap->changelist, kevp, count * sizeof *kevp);
999 	if (error == 0)
1000 		uap->changelist += count;
1001 	return (error);
1002 }
1003 
1004 int
1005 kern_kevent(struct thread *td, int fd, int nchanges, int nevents,
1006     struct kevent_copyops *k_ops, const struct timespec *timeout)
1007 {
1008 	cap_rights_t rights;
1009 	struct file *fp;
1010 	int error;
1011 
1012 	cap_rights_init(&rights);
1013 	if (nchanges > 0)
1014 		cap_rights_set(&rights, CAP_KQUEUE_CHANGE);
1015 	if (nevents > 0)
1016 		cap_rights_set(&rights, CAP_KQUEUE_EVENT);
1017 	error = fget(td, fd, &rights, &fp);
1018 	if (error != 0)
1019 		return (error);
1020 
1021 	error = kern_kevent_fp(td, fp, nchanges, nevents, k_ops, timeout);
1022 	fdrop(fp, td);
1023 
1024 	return (error);
1025 }
1026 
1027 static int
1028 kqueue_kevent(struct kqueue *kq, struct thread *td, int nchanges, int nevents,
1029     struct kevent_copyops *k_ops, const struct timespec *timeout)
1030 {
1031 	struct kevent keva[KQ_NEVENTS];
1032 	struct kevent *kevp, *changes;
1033 	int i, n, nerrors, error;
1034 
1035 	nerrors = 0;
1036 	while (nchanges > 0) {
1037 		n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges;
1038 		error = k_ops->k_copyin(k_ops->arg, keva, n);
1039 		if (error)
1040 			return (error);
1041 		changes = keva;
1042 		for (i = 0; i < n; i++) {
1043 			kevp = &changes[i];
1044 			if (!kevp->filter)
1045 				continue;
1046 			kevp->flags &= ~EV_SYSFLAGS;
1047 			error = kqueue_register(kq, kevp, td, 1);
1048 			if (error || (kevp->flags & EV_RECEIPT)) {
1049 				if (nevents == 0)
1050 					return (error);
1051 				kevp->flags = EV_ERROR;
1052 				kevp->data = error;
1053 				(void)k_ops->k_copyout(k_ops->arg, kevp, 1);
1054 				nevents--;
1055 				nerrors++;
1056 			}
1057 		}
1058 		nchanges -= n;
1059 	}
1060 	if (nerrors) {
1061 		td->td_retval[0] = nerrors;
1062 		return (0);
1063 	}
1064 
1065 	return (kqueue_scan(kq, nevents, k_ops, timeout, keva, td));
1066 }
1067 
1068 int
1069 kern_kevent_fp(struct thread *td, struct file *fp, int nchanges, int nevents,
1070     struct kevent_copyops *k_ops, const struct timespec *timeout)
1071 {
1072 	struct kqueue *kq;
1073 	int error;
1074 
1075 	error = kqueue_acquire(fp, &kq);
1076 	if (error != 0)
1077 		return (error);
1078 	error = kqueue_kevent(kq, td, nchanges, nevents, k_ops, timeout);
1079 	kqueue_release(kq, 0);
1080 	return (error);
1081 }
1082 
1083 /*
1084  * Performs a kevent() call on a temporarily created kqueue. This can be
1085  * used to perform one-shot polling, similar to poll() and select().
1086  */
1087 int
1088 kern_kevent_anonymous(struct thread *td, int nevents,
1089     struct kevent_copyops *k_ops)
1090 {
1091 	struct kqueue kq = {};
1092 	int error;
1093 
1094 	kqueue_init(&kq);
1095 	kq.kq_refcnt = 1;
1096 	error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL);
1097 	kqueue_drain(&kq, td);
1098 	kqueue_destroy(&kq);
1099 	return (error);
1100 }
1101 
1102 int
1103 kqueue_add_filteropts(int filt, struct filterops *filtops)
1104 {
1105 	int error;
1106 
1107 	error = 0;
1108 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) {
1109 		printf(
1110 "trying to add a filterop that is out of range: %d is beyond %d\n",
1111 		    ~filt, EVFILT_SYSCOUNT);
1112 		return EINVAL;
1113 	}
1114 	mtx_lock(&filterops_lock);
1115 	if (sysfilt_ops[~filt].for_fop != &null_filtops &&
1116 	    sysfilt_ops[~filt].for_fop != NULL)
1117 		error = EEXIST;
1118 	else {
1119 		sysfilt_ops[~filt].for_fop = filtops;
1120 		sysfilt_ops[~filt].for_refcnt = 0;
1121 	}
1122 	mtx_unlock(&filterops_lock);
1123 
1124 	return (error);
1125 }
1126 
1127 int
1128 kqueue_del_filteropts(int filt)
1129 {
1130 	int error;
1131 
1132 	error = 0;
1133 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1134 		return EINVAL;
1135 
1136 	mtx_lock(&filterops_lock);
1137 	if (sysfilt_ops[~filt].for_fop == &null_filtops ||
1138 	    sysfilt_ops[~filt].for_fop == NULL)
1139 		error = EINVAL;
1140 	else if (sysfilt_ops[~filt].for_refcnt != 0)
1141 		error = EBUSY;
1142 	else {
1143 		sysfilt_ops[~filt].for_fop = &null_filtops;
1144 		sysfilt_ops[~filt].for_refcnt = 0;
1145 	}
1146 	mtx_unlock(&filterops_lock);
1147 
1148 	return error;
1149 }
1150 
1151 static struct filterops *
1152 kqueue_fo_find(int filt)
1153 {
1154 
1155 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1156 		return NULL;
1157 
1158 	if (sysfilt_ops[~filt].for_nolock)
1159 		return sysfilt_ops[~filt].for_fop;
1160 
1161 	mtx_lock(&filterops_lock);
1162 	sysfilt_ops[~filt].for_refcnt++;
1163 	if (sysfilt_ops[~filt].for_fop == NULL)
1164 		sysfilt_ops[~filt].for_fop = &null_filtops;
1165 	mtx_unlock(&filterops_lock);
1166 
1167 	return sysfilt_ops[~filt].for_fop;
1168 }
1169 
1170 static void
1171 kqueue_fo_release(int filt)
1172 {
1173 
1174 	if (filt > 0 || filt + EVFILT_SYSCOUNT < 0)
1175 		return;
1176 
1177 	if (sysfilt_ops[~filt].for_nolock)
1178 		return;
1179 
1180 	mtx_lock(&filterops_lock);
1181 	KASSERT(sysfilt_ops[~filt].for_refcnt > 0,
1182 	    ("filter object refcount not valid on release"));
1183 	sysfilt_ops[~filt].for_refcnt--;
1184 	mtx_unlock(&filterops_lock);
1185 }
1186 
1187 /*
1188  * A ref to kq (obtained via kqueue_acquire) must be held.  waitok will
1189  * influence if memory allocation should wait.  Make sure it is 0 if you
1190  * hold any mutexes.
1191  */
1192 static int
1193 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok)
1194 {
1195 	struct filterops *fops;
1196 	struct file *fp;
1197 	struct knote *kn, *tkn;
1198 	struct knlist *knl;
1199 	cap_rights_t rights;
1200 	int error, filt, event;
1201 	int haskqglobal, filedesc_unlock;
1202 
1203 	if ((kev->flags & (EV_ENABLE | EV_DISABLE)) == (EV_ENABLE | EV_DISABLE))
1204 		return (EINVAL);
1205 
1206 	fp = NULL;
1207 	kn = NULL;
1208 	knl = NULL;
1209 	error = 0;
1210 	haskqglobal = 0;
1211 	filedesc_unlock = 0;
1212 
1213 	filt = kev->filter;
1214 	fops = kqueue_fo_find(filt);
1215 	if (fops == NULL)
1216 		return EINVAL;
1217 
1218 	if (kev->flags & EV_ADD) {
1219 		/*
1220 		 * Prevent waiting with locks.  Non-sleepable
1221 		 * allocation failures are handled in the loop, only
1222 		 * if the spare knote appears to be actually required.
1223 		 */
1224 		tkn = knote_alloc(waitok);
1225 	} else {
1226 		tkn = NULL;
1227 	}
1228 
1229 findkn:
1230 	if (fops->f_isfd) {
1231 		KASSERT(td != NULL, ("td is NULL"));
1232 		if (kev->ident > INT_MAX)
1233 			error = EBADF;
1234 		else
1235 			error = fget(td, kev->ident,
1236 			    cap_rights_init(&rights, CAP_EVENT), &fp);
1237 		if (error)
1238 			goto done;
1239 
1240 		if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops,
1241 		    kev->ident, 0) != 0) {
1242 			/* try again */
1243 			fdrop(fp, td);
1244 			fp = NULL;
1245 			error = kqueue_expand(kq, fops, kev->ident, waitok);
1246 			if (error)
1247 				goto done;
1248 			goto findkn;
1249 		}
1250 
1251 		if (fp->f_type == DTYPE_KQUEUE) {
1252 			/*
1253 			 * If we add some intelligence about what we are doing,
1254 			 * we should be able to support events on ourselves.
1255 			 * We need to know when we are doing this to prevent
1256 			 * getting both the knlist lock and the kq lock since
1257 			 * they are the same thing.
1258 			 */
1259 			if (fp->f_data == kq) {
1260 				error = EINVAL;
1261 				goto done;
1262 			}
1263 
1264 			/*
1265 			 * Pre-lock the filedesc before the global
1266 			 * lock mutex, see the comment in
1267 			 * kqueue_close().
1268 			 */
1269 			FILEDESC_XLOCK(td->td_proc->p_fd);
1270 			filedesc_unlock = 1;
1271 			KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1272 		}
1273 
1274 		KQ_LOCK(kq);
1275 		if (kev->ident < kq->kq_knlistsize) {
1276 			SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link)
1277 				if (kev->filter == kn->kn_filter)
1278 					break;
1279 		}
1280 	} else {
1281 		if ((kev->flags & EV_ADD) == EV_ADD)
1282 			kqueue_expand(kq, fops, kev->ident, waitok);
1283 
1284 		KQ_LOCK(kq);
1285 
1286 		/*
1287 		 * If possible, find an existing knote to use for this kevent.
1288 		 */
1289 		if (kev->filter == EVFILT_PROC &&
1290 		    (kev->flags & (EV_FLAG1 | EV_FLAG2)) != 0) {
1291 			/* This is an internal creation of a process tracking
1292 			 * note. Don't attempt to coalesce this with an
1293 			 * existing note.
1294 			 */
1295 			;
1296 		} else if (kq->kq_knhashmask != 0) {
1297 			struct klist *list;
1298 
1299 			list = &kq->kq_knhash[
1300 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1301 			SLIST_FOREACH(kn, list, kn_link)
1302 				if (kev->ident == kn->kn_id &&
1303 				    kev->filter == kn->kn_filter)
1304 					break;
1305 		}
1306 	}
1307 
1308 	/* knote is in the process of changing, wait for it to stabilize. */
1309 	if (kn != NULL && kn_in_flux(kn)) {
1310 		KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1311 		if (filedesc_unlock) {
1312 			FILEDESC_XUNLOCK(td->td_proc->p_fd);
1313 			filedesc_unlock = 0;
1314 		}
1315 		kq->kq_state |= KQ_FLUXWAIT;
1316 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0);
1317 		if (fp != NULL) {
1318 			fdrop(fp, td);
1319 			fp = NULL;
1320 		}
1321 		goto findkn;
1322 	}
1323 
1324 	/*
1325 	 * kn now contains the matching knote, or NULL if no match
1326 	 */
1327 	if (kn == NULL) {
1328 		if (kev->flags & EV_ADD) {
1329 			kn = tkn;
1330 			tkn = NULL;
1331 			if (kn == NULL) {
1332 				KQ_UNLOCK(kq);
1333 				error = ENOMEM;
1334 				goto done;
1335 			}
1336 			kn->kn_fp = fp;
1337 			kn->kn_kq = kq;
1338 			kn->kn_fop = fops;
1339 			/*
1340 			 * apply reference counts to knote structure, and
1341 			 * do not release it at the end of this routine.
1342 			 */
1343 			fops = NULL;
1344 			fp = NULL;
1345 
1346 			kn->kn_sfflags = kev->fflags;
1347 			kn->kn_sdata = kev->data;
1348 			kev->fflags = 0;
1349 			kev->data = 0;
1350 			kn->kn_kevent = *kev;
1351 			kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE |
1352 			    EV_ENABLE | EV_DISABLE | EV_FORCEONESHOT);
1353 			kn->kn_status = KN_DETACHED;
1354 			kn_enter_flux(kn);
1355 
1356 			error = knote_attach(kn, kq);
1357 			KQ_UNLOCK(kq);
1358 			if (error != 0) {
1359 				tkn = kn;
1360 				goto done;
1361 			}
1362 
1363 			if ((error = kn->kn_fop->f_attach(kn)) != 0) {
1364 				knote_drop_detached(kn, td);
1365 				goto done;
1366 			}
1367 			knl = kn_list_lock(kn);
1368 			goto done_ev_add;
1369 		} else {
1370 			/* No matching knote and the EV_ADD flag is not set. */
1371 			KQ_UNLOCK(kq);
1372 			error = ENOENT;
1373 			goto done;
1374 		}
1375 	}
1376 
1377 	if (kev->flags & EV_DELETE) {
1378 		kn_enter_flux(kn);
1379 		KQ_UNLOCK(kq);
1380 		knote_drop(kn, td);
1381 		goto done;
1382 	}
1383 
1384 	if (kev->flags & EV_FORCEONESHOT) {
1385 		kn->kn_flags |= EV_ONESHOT;
1386 		KNOTE_ACTIVATE(kn, 1);
1387 	}
1388 
1389 	/*
1390 	 * The user may change some filter values after the initial EV_ADD,
1391 	 * but doing so will not reset any filter which has already been
1392 	 * triggered.
1393 	 */
1394 	kn->kn_status |= KN_SCAN;
1395 	kn_enter_flux(kn);
1396 	KQ_UNLOCK(kq);
1397 	knl = kn_list_lock(kn);
1398 	kn->kn_kevent.udata = kev->udata;
1399 	if (!fops->f_isfd && fops->f_touch != NULL) {
1400 		fops->f_touch(kn, kev, EVENT_REGISTER);
1401 	} else {
1402 		kn->kn_sfflags = kev->fflags;
1403 		kn->kn_sdata = kev->data;
1404 	}
1405 
1406 	/*
1407 	 * We can get here with kn->kn_knlist == NULL.  This can happen when
1408 	 * the initial attach event decides that the event is "completed"
1409 	 * already.  i.e. filt_procattach is called on a zombie process.  It
1410 	 * will call filt_proc which will remove it from the list, and NULL
1411 	 * kn_knlist.
1412 	 */
1413 done_ev_add:
1414 	if ((kev->flags & EV_ENABLE) != 0)
1415 		kn->kn_status &= ~KN_DISABLED;
1416 	else if ((kev->flags & EV_DISABLE) != 0)
1417 		kn->kn_status |= KN_DISABLED;
1418 
1419 	if ((kn->kn_status & KN_DISABLED) == 0)
1420 		event = kn->kn_fop->f_event(kn, 0);
1421 	else
1422 		event = 0;
1423 
1424 	KQ_LOCK(kq);
1425 	if (event)
1426 		kn->kn_status |= KN_ACTIVE;
1427 	if ((kn->kn_status & (KN_ACTIVE | KN_DISABLED | KN_QUEUED)) ==
1428 	    KN_ACTIVE)
1429 		knote_enqueue(kn);
1430 	kn->kn_status &= ~KN_SCAN;
1431 	kn_leave_flux(kn);
1432 	kn_list_unlock(knl);
1433 	KQ_UNLOCK_FLUX(kq);
1434 
1435 done:
1436 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1437 	if (filedesc_unlock)
1438 		FILEDESC_XUNLOCK(td->td_proc->p_fd);
1439 	if (fp != NULL)
1440 		fdrop(fp, td);
1441 	knote_free(tkn);
1442 	if (fops != NULL)
1443 		kqueue_fo_release(filt);
1444 	return (error);
1445 }
1446 
1447 static int
1448 kqueue_acquire(struct file *fp, struct kqueue **kqp)
1449 {
1450 	int error;
1451 	struct kqueue *kq;
1452 
1453 	error = 0;
1454 
1455 	kq = fp->f_data;
1456 	if (fp->f_type != DTYPE_KQUEUE || kq == NULL)
1457 		return (EBADF);
1458 	*kqp = kq;
1459 	KQ_LOCK(kq);
1460 	if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) {
1461 		KQ_UNLOCK(kq);
1462 		return (EBADF);
1463 	}
1464 	kq->kq_refcnt++;
1465 	KQ_UNLOCK(kq);
1466 
1467 	return error;
1468 }
1469 
1470 static void
1471 kqueue_release(struct kqueue *kq, int locked)
1472 {
1473 	if (locked)
1474 		KQ_OWNED(kq);
1475 	else
1476 		KQ_LOCK(kq);
1477 	kq->kq_refcnt--;
1478 	if (kq->kq_refcnt == 1)
1479 		wakeup(&kq->kq_refcnt);
1480 	if (!locked)
1481 		KQ_UNLOCK(kq);
1482 }
1483 
1484 static void
1485 kqueue_schedtask(struct kqueue *kq)
1486 {
1487 
1488 	KQ_OWNED(kq);
1489 	KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN),
1490 	    ("scheduling kqueue task while draining"));
1491 
1492 	if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) {
1493 		taskqueue_enqueue(taskqueue_kqueue_ctx, &kq->kq_task);
1494 		kq->kq_state |= KQ_TASKSCHED;
1495 	}
1496 }
1497 
1498 /*
1499  * Expand the kq to make sure we have storage for fops/ident pair.
1500  *
1501  * Return 0 on success (or no work necessary), return errno on failure.
1502  *
1503  * Not calling hashinit w/ waitok (proper malloc flag) should be safe.
1504  * If kqueue_register is called from a non-fd context, there usually/should
1505  * be no locks held.
1506  */
1507 static int
1508 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident,
1509 	int waitok)
1510 {
1511 	struct klist *list, *tmp_knhash, *to_free;
1512 	u_long tmp_knhashmask;
1513 	int size;
1514 	int fd;
1515 	int mflag = waitok ? M_WAITOK : M_NOWAIT;
1516 
1517 	KQ_NOTOWNED(kq);
1518 
1519 	to_free = NULL;
1520 	if (fops->f_isfd) {
1521 		fd = ident;
1522 		if (kq->kq_knlistsize <= fd) {
1523 			size = kq->kq_knlistsize;
1524 			while (size <= fd)
1525 				size += KQEXTENT;
1526 			list = malloc(size * sizeof(*list), M_KQUEUE, mflag);
1527 			if (list == NULL)
1528 				return ENOMEM;
1529 			KQ_LOCK(kq);
1530 			if (kq->kq_knlistsize > fd) {
1531 				to_free = list;
1532 				list = NULL;
1533 			} else {
1534 				if (kq->kq_knlist != NULL) {
1535 					bcopy(kq->kq_knlist, list,
1536 					    kq->kq_knlistsize * sizeof(*list));
1537 					to_free = kq->kq_knlist;
1538 					kq->kq_knlist = NULL;
1539 				}
1540 				bzero((caddr_t)list +
1541 				    kq->kq_knlistsize * sizeof(*list),
1542 				    (size - kq->kq_knlistsize) * sizeof(*list));
1543 				kq->kq_knlistsize = size;
1544 				kq->kq_knlist = list;
1545 			}
1546 			KQ_UNLOCK(kq);
1547 		}
1548 	} else {
1549 		if (kq->kq_knhashmask == 0) {
1550 			tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1551 			    &tmp_knhashmask);
1552 			if (tmp_knhash == NULL)
1553 				return ENOMEM;
1554 			KQ_LOCK(kq);
1555 			if (kq->kq_knhashmask == 0) {
1556 				kq->kq_knhash = tmp_knhash;
1557 				kq->kq_knhashmask = tmp_knhashmask;
1558 			} else {
1559 				to_free = tmp_knhash;
1560 			}
1561 			KQ_UNLOCK(kq);
1562 		}
1563 	}
1564 	free(to_free, M_KQUEUE);
1565 
1566 	KQ_NOTOWNED(kq);
1567 	return 0;
1568 }
1569 
1570 static void
1571 kqueue_task(void *arg, int pending)
1572 {
1573 	struct kqueue *kq;
1574 	int haskqglobal;
1575 
1576 	haskqglobal = 0;
1577 	kq = arg;
1578 
1579 	KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1580 	KQ_LOCK(kq);
1581 
1582 	KNOTE_LOCKED(&kq->kq_sel.si_note, 0);
1583 
1584 	kq->kq_state &= ~KQ_TASKSCHED;
1585 	if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) {
1586 		wakeup(&kq->kq_state);
1587 	}
1588 	KQ_UNLOCK(kq);
1589 	KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1590 }
1591 
1592 /*
1593  * Scan, update kn_data (if not ONESHOT), and copyout triggered events.
1594  * We treat KN_MARKER knotes as if they are in flux.
1595  */
1596 static int
1597 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops,
1598     const struct timespec *tsp, struct kevent *keva, struct thread *td)
1599 {
1600 	struct kevent *kevp;
1601 	struct knote *kn, *marker;
1602 	struct knlist *knl;
1603 	sbintime_t asbt, rsbt;
1604 	int count, error, haskqglobal, influx, nkev, touch;
1605 
1606 	count = maxevents;
1607 	nkev = 0;
1608 	error = 0;
1609 	haskqglobal = 0;
1610 
1611 	if (maxevents == 0)
1612 		goto done_nl;
1613 
1614 	rsbt = 0;
1615 	if (tsp != NULL) {
1616 		if (tsp->tv_sec < 0 || tsp->tv_nsec < 0 ||
1617 		    tsp->tv_nsec >= 1000000000) {
1618 			error = EINVAL;
1619 			goto done_nl;
1620 		}
1621 		if (timespecisset(tsp)) {
1622 			if (tsp->tv_sec <= INT32_MAX) {
1623 				rsbt = tstosbt(*tsp);
1624 				if (TIMESEL(&asbt, rsbt))
1625 					asbt += tc_tick_sbt;
1626 				if (asbt <= SBT_MAX - rsbt)
1627 					asbt += rsbt;
1628 				else
1629 					asbt = 0;
1630 				rsbt >>= tc_precexp;
1631 			} else
1632 				asbt = 0;
1633 		} else
1634 			asbt = -1;
1635 	} else
1636 		asbt = 0;
1637 	marker = knote_alloc(1);
1638 	marker->kn_status = KN_MARKER;
1639 	KQ_LOCK(kq);
1640 
1641 retry:
1642 	kevp = keva;
1643 	if (kq->kq_count == 0) {
1644 		if (asbt == -1) {
1645 			error = EWOULDBLOCK;
1646 		} else {
1647 			kq->kq_state |= KQ_SLEEP;
1648 			error = msleep_sbt(kq, &kq->kq_lock, PSOCK | PCATCH,
1649 			    "kqread", asbt, rsbt, C_ABSOLUTE);
1650 		}
1651 		if (error == 0)
1652 			goto retry;
1653 		/* don't restart after signals... */
1654 		if (error == ERESTART)
1655 			error = EINTR;
1656 		else if (error == EWOULDBLOCK)
1657 			error = 0;
1658 		goto done;
1659 	}
1660 
1661 	TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe);
1662 	influx = 0;
1663 	while (count) {
1664 		KQ_OWNED(kq);
1665 		kn = TAILQ_FIRST(&kq->kq_head);
1666 
1667 		if ((kn->kn_status == KN_MARKER && kn != marker) ||
1668 		    kn_in_flux(kn)) {
1669 			if (influx) {
1670 				influx = 0;
1671 				KQ_FLUX_WAKEUP(kq);
1672 			}
1673 			kq->kq_state |= KQ_FLUXWAIT;
1674 			error = msleep(kq, &kq->kq_lock, PSOCK,
1675 			    "kqflxwt", 0);
1676 			continue;
1677 		}
1678 
1679 		TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
1680 		if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) {
1681 			kn->kn_status &= ~KN_QUEUED;
1682 			kq->kq_count--;
1683 			continue;
1684 		}
1685 		if (kn == marker) {
1686 			KQ_FLUX_WAKEUP(kq);
1687 			if (count == maxevents)
1688 				goto retry;
1689 			goto done;
1690 		}
1691 		KASSERT(!kn_in_flux(kn),
1692 		    ("knote %p is unexpectedly in flux", kn));
1693 
1694 		if ((kn->kn_flags & EV_DROP) == EV_DROP) {
1695 			kn->kn_status &= ~KN_QUEUED;
1696 			kn_enter_flux(kn);
1697 			kq->kq_count--;
1698 			KQ_UNLOCK(kq);
1699 			/*
1700 			 * We don't need to lock the list since we've
1701 			 * marked it as in flux.
1702 			 */
1703 			knote_drop(kn, td);
1704 			KQ_LOCK(kq);
1705 			continue;
1706 		} else if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) {
1707 			kn->kn_status &= ~KN_QUEUED;
1708 			kn_enter_flux(kn);
1709 			kq->kq_count--;
1710 			KQ_UNLOCK(kq);
1711 			/*
1712 			 * We don't need to lock the list since we've
1713 			 * marked the knote as being in flux.
1714 			 */
1715 			*kevp = kn->kn_kevent;
1716 			knote_drop(kn, td);
1717 			KQ_LOCK(kq);
1718 			kn = NULL;
1719 		} else {
1720 			kn->kn_status |= KN_SCAN;
1721 			kn_enter_flux(kn);
1722 			KQ_UNLOCK(kq);
1723 			if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE)
1724 				KQ_GLOBAL_LOCK(&kq_global, haskqglobal);
1725 			knl = kn_list_lock(kn);
1726 			if (kn->kn_fop->f_event(kn, 0) == 0) {
1727 				KQ_LOCK(kq);
1728 				KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1729 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE |
1730 				    KN_SCAN);
1731 				kn_leave_flux(kn);
1732 				kq->kq_count--;
1733 				kn_list_unlock(knl);
1734 				influx = 1;
1735 				continue;
1736 			}
1737 			touch = (!kn->kn_fop->f_isfd &&
1738 			    kn->kn_fop->f_touch != NULL);
1739 			if (touch)
1740 				kn->kn_fop->f_touch(kn, kevp, EVENT_PROCESS);
1741 			else
1742 				*kevp = kn->kn_kevent;
1743 			KQ_LOCK(kq);
1744 			KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal);
1745 			if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1746 				/*
1747 				 * Manually clear knotes who weren't
1748 				 * 'touch'ed.
1749 				 */
1750 				if (touch == 0 && kn->kn_flags & EV_CLEAR) {
1751 					kn->kn_data = 0;
1752 					kn->kn_fflags = 0;
1753 				}
1754 				if (kn->kn_flags & EV_DISPATCH)
1755 					kn->kn_status |= KN_DISABLED;
1756 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1757 				kq->kq_count--;
1758 			} else
1759 				TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
1760 
1761 			kn->kn_status &= ~KN_SCAN;
1762 			kn_leave_flux(kn);
1763 			kn_list_unlock(knl);
1764 			influx = 1;
1765 		}
1766 
1767 		/* we are returning a copy to the user */
1768 		kevp++;
1769 		nkev++;
1770 		count--;
1771 
1772 		if (nkev == KQ_NEVENTS) {
1773 			influx = 0;
1774 			KQ_UNLOCK_FLUX(kq);
1775 			error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1776 			nkev = 0;
1777 			kevp = keva;
1778 			KQ_LOCK(kq);
1779 			if (error)
1780 				break;
1781 		}
1782 	}
1783 	TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe);
1784 done:
1785 	KQ_OWNED(kq);
1786 	KQ_UNLOCK_FLUX(kq);
1787 	knote_free(marker);
1788 done_nl:
1789 	KQ_NOTOWNED(kq);
1790 	if (nkev != 0)
1791 		error = k_ops->k_copyout(k_ops->arg, keva, nkev);
1792 	td->td_retval[0] = maxevents - count;
1793 	return (error);
1794 }
1795 
1796 /*ARGSUSED*/
1797 static int
1798 kqueue_ioctl(struct file *fp, u_long cmd, void *data,
1799 	struct ucred *active_cred, struct thread *td)
1800 {
1801 	/*
1802 	 * Enabling sigio causes two major problems:
1803 	 * 1) infinite recursion:
1804 	 * Synopsys: kevent is being used to track signals and have FIOASYNC
1805 	 * set.  On receipt of a signal this will cause a kqueue to recurse
1806 	 * into itself over and over.  Sending the sigio causes the kqueue
1807 	 * to become ready, which in turn posts sigio again, forever.
1808 	 * Solution: this can be solved by setting a flag in the kqueue that
1809 	 * we have a SIGIO in progress.
1810 	 * 2) locking problems:
1811 	 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts
1812 	 * us above the proc and pgrp locks.
1813 	 * Solution: Post a signal using an async mechanism, being sure to
1814 	 * record a generation count in the delivery so that we do not deliver
1815 	 * a signal to the wrong process.
1816 	 *
1817 	 * Note, these two mechanisms are somewhat mutually exclusive!
1818 	 */
1819 #if 0
1820 	struct kqueue *kq;
1821 
1822 	kq = fp->f_data;
1823 	switch (cmd) {
1824 	case FIOASYNC:
1825 		if (*(int *)data) {
1826 			kq->kq_state |= KQ_ASYNC;
1827 		} else {
1828 			kq->kq_state &= ~KQ_ASYNC;
1829 		}
1830 		return (0);
1831 
1832 	case FIOSETOWN:
1833 		return (fsetown(*(int *)data, &kq->kq_sigio));
1834 
1835 	case FIOGETOWN:
1836 		*(int *)data = fgetown(&kq->kq_sigio);
1837 		return (0);
1838 	}
1839 #endif
1840 
1841 	return (ENOTTY);
1842 }
1843 
1844 /*ARGSUSED*/
1845 static int
1846 kqueue_poll(struct file *fp, int events, struct ucred *active_cred,
1847 	struct thread *td)
1848 {
1849 	struct kqueue *kq;
1850 	int revents = 0;
1851 	int error;
1852 
1853 	if ((error = kqueue_acquire(fp, &kq)))
1854 		return POLLERR;
1855 
1856 	KQ_LOCK(kq);
1857 	if (events & (POLLIN | POLLRDNORM)) {
1858 		if (kq->kq_count) {
1859 			revents |= events & (POLLIN | POLLRDNORM);
1860 		} else {
1861 			selrecord(td, &kq->kq_sel);
1862 			if (SEL_WAITING(&kq->kq_sel))
1863 				kq->kq_state |= KQ_SEL;
1864 		}
1865 	}
1866 	kqueue_release(kq, 1);
1867 	KQ_UNLOCK(kq);
1868 	return (revents);
1869 }
1870 
1871 /*ARGSUSED*/
1872 static int
1873 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred,
1874 	struct thread *td)
1875 {
1876 
1877 	bzero((void *)st, sizeof *st);
1878 	/*
1879 	 * We no longer return kq_count because the unlocked value is useless.
1880 	 * If you spent all this time getting the count, why not spend your
1881 	 * syscall better by calling kevent?
1882 	 *
1883 	 * XXX - This is needed for libc_r.
1884 	 */
1885 	st->st_mode = S_IFIFO;
1886 	return (0);
1887 }
1888 
1889 static void
1890 kqueue_drain(struct kqueue *kq, struct thread *td)
1891 {
1892 	struct knote *kn;
1893 	int i;
1894 
1895 	KQ_LOCK(kq);
1896 
1897 	KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING,
1898 	    ("kqueue already closing"));
1899 	kq->kq_state |= KQ_CLOSING;
1900 	if (kq->kq_refcnt > 1)
1901 		msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0);
1902 
1903 	KASSERT(kq->kq_refcnt == 1, ("other refs are out there!"));
1904 
1905 	KASSERT(knlist_empty(&kq->kq_sel.si_note),
1906 	    ("kqueue's knlist not empty"));
1907 
1908 	for (i = 0; i < kq->kq_knlistsize; i++) {
1909 		while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) {
1910 			if (kn_in_flux(kn)) {
1911 				kq->kq_state |= KQ_FLUXWAIT;
1912 				msleep(kq, &kq->kq_lock, PSOCK, "kqclo1", 0);
1913 				continue;
1914 			}
1915 			kn_enter_flux(kn);
1916 			KQ_UNLOCK(kq);
1917 			knote_drop(kn, td);
1918 			KQ_LOCK(kq);
1919 		}
1920 	}
1921 	if (kq->kq_knhashmask != 0) {
1922 		for (i = 0; i <= kq->kq_knhashmask; i++) {
1923 			while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) {
1924 				if (kn_in_flux(kn)) {
1925 					kq->kq_state |= KQ_FLUXWAIT;
1926 					msleep(kq, &kq->kq_lock, PSOCK,
1927 					       "kqclo2", 0);
1928 					continue;
1929 				}
1930 				kn_enter_flux(kn);
1931 				KQ_UNLOCK(kq);
1932 				knote_drop(kn, td);
1933 				KQ_LOCK(kq);
1934 			}
1935 		}
1936 	}
1937 
1938 	if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) {
1939 		kq->kq_state |= KQ_TASKDRAIN;
1940 		msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0);
1941 	}
1942 
1943 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
1944 		selwakeuppri(&kq->kq_sel, PSOCK);
1945 		if (!SEL_WAITING(&kq->kq_sel))
1946 			kq->kq_state &= ~KQ_SEL;
1947 	}
1948 
1949 	KQ_UNLOCK(kq);
1950 }
1951 
1952 static void
1953 kqueue_destroy(struct kqueue *kq)
1954 {
1955 
1956 	KASSERT(kq->kq_fdp == NULL,
1957 	    ("kqueue still attached to a file descriptor"));
1958 	seldrain(&kq->kq_sel);
1959 	knlist_destroy(&kq->kq_sel.si_note);
1960 	mtx_destroy(&kq->kq_lock);
1961 
1962 	if (kq->kq_knhash != NULL)
1963 		free(kq->kq_knhash, M_KQUEUE);
1964 	if (kq->kq_knlist != NULL)
1965 		free(kq->kq_knlist, M_KQUEUE);
1966 
1967 	funsetown(&kq->kq_sigio);
1968 }
1969 
1970 /*ARGSUSED*/
1971 static int
1972 kqueue_close(struct file *fp, struct thread *td)
1973 {
1974 	struct kqueue *kq = fp->f_data;
1975 	struct filedesc *fdp;
1976 	int error;
1977 	int filedesc_unlock;
1978 
1979 	if ((error = kqueue_acquire(fp, &kq)))
1980 		return error;
1981 	kqueue_drain(kq, td);
1982 
1983 	/*
1984 	 * We could be called due to the knote_drop() doing fdrop(),
1985 	 * called from kqueue_register().  In this case the global
1986 	 * lock is owned, and filedesc sx is locked before, to not
1987 	 * take the sleepable lock after non-sleepable.
1988 	 */
1989 	fdp = kq->kq_fdp;
1990 	kq->kq_fdp = NULL;
1991 	if (!sx_xlocked(FILEDESC_LOCK(fdp))) {
1992 		FILEDESC_XLOCK(fdp);
1993 		filedesc_unlock = 1;
1994 	} else
1995 		filedesc_unlock = 0;
1996 	TAILQ_REMOVE(&fdp->fd_kqlist, kq, kq_list);
1997 	if (filedesc_unlock)
1998 		FILEDESC_XUNLOCK(fdp);
1999 
2000 	kqueue_destroy(kq);
2001 	chgkqcnt(kq->kq_cred->cr_ruidinfo, -1, 0);
2002 	crfree(kq->kq_cred);
2003 	free(kq, M_KQUEUE);
2004 	fp->f_data = NULL;
2005 
2006 	return (0);
2007 }
2008 
2009 static int
2010 kqueue_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2011 {
2012 
2013 	kif->kf_type = KF_TYPE_KQUEUE;
2014 	return (0);
2015 }
2016 
2017 static void
2018 kqueue_wakeup(struct kqueue *kq)
2019 {
2020 	KQ_OWNED(kq);
2021 
2022 	if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) {
2023 		kq->kq_state &= ~KQ_SLEEP;
2024 		wakeup(kq);
2025 	}
2026 	if ((kq->kq_state & KQ_SEL) == KQ_SEL) {
2027 		selwakeuppri(&kq->kq_sel, PSOCK);
2028 		if (!SEL_WAITING(&kq->kq_sel))
2029 			kq->kq_state &= ~KQ_SEL;
2030 	}
2031 	if (!knlist_empty(&kq->kq_sel.si_note))
2032 		kqueue_schedtask(kq);
2033 	if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) {
2034 		pgsigio(&kq->kq_sigio, SIGIO, 0);
2035 	}
2036 }
2037 
2038 /*
2039  * Walk down a list of knotes, activating them if their event has triggered.
2040  *
2041  * There is a possibility to optimize in the case of one kq watching another.
2042  * Instead of scheduling a task to wake it up, you could pass enough state
2043  * down the chain to make up the parent kqueue.  Make this code functional
2044  * first.
2045  */
2046 void
2047 knote(struct knlist *list, long hint, int lockflags)
2048 {
2049 	struct kqueue *kq;
2050 	struct knote *kn, *tkn;
2051 	int error;
2052 
2053 	if (list == NULL)
2054 		return;
2055 
2056 	KNL_ASSERT_LOCK(list, lockflags & KNF_LISTLOCKED);
2057 
2058 	if ((lockflags & KNF_LISTLOCKED) == 0)
2059 		list->kl_lock(list->kl_lockarg);
2060 
2061 	/*
2062 	 * If we unlock the list lock (and enter influx), we can
2063 	 * eliminate the kqueue scheduling, but this will introduce
2064 	 * four lock/unlock's for each knote to test.  Also, marker
2065 	 * would be needed to keep iteration position, since filters
2066 	 * or other threads could remove events.
2067 	 */
2068 	SLIST_FOREACH_SAFE(kn, &list->kl_list, kn_selnext, tkn) {
2069 		kq = kn->kn_kq;
2070 		KQ_LOCK(kq);
2071 		if (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0) {
2072 			/*
2073 			 * Do not process the influx notes, except for
2074 			 * the influx coming from the kq unlock in the
2075 			 * kqueue_scan().  In the later case, we do
2076 			 * not interfere with the scan, since the code
2077 			 * fragment in kqueue_scan() locks the knlist,
2078 			 * and cannot proceed until we finished.
2079 			 */
2080 			KQ_UNLOCK(kq);
2081 		} else if ((lockflags & KNF_NOKQLOCK) != 0) {
2082 			kn_enter_flux(kn);
2083 			KQ_UNLOCK(kq);
2084 			error = kn->kn_fop->f_event(kn, hint);
2085 			KQ_LOCK(kq);
2086 			kn_leave_flux(kn);
2087 			if (error)
2088 				KNOTE_ACTIVATE(kn, 1);
2089 			KQ_UNLOCK_FLUX(kq);
2090 		} else {
2091 			kn->kn_status |= KN_HASKQLOCK;
2092 			if (kn->kn_fop->f_event(kn, hint))
2093 				KNOTE_ACTIVATE(kn, 1);
2094 			kn->kn_status &= ~KN_HASKQLOCK;
2095 			KQ_UNLOCK(kq);
2096 		}
2097 	}
2098 	if ((lockflags & KNF_LISTLOCKED) == 0)
2099 		list->kl_unlock(list->kl_lockarg);
2100 }
2101 
2102 /*
2103  * add a knote to a knlist
2104  */
2105 void
2106 knlist_add(struct knlist *knl, struct knote *kn, int islocked)
2107 {
2108 
2109 	KNL_ASSERT_LOCK(knl, islocked);
2110 	KQ_NOTOWNED(kn->kn_kq);
2111 	KASSERT(kn_in_flux(kn), ("knote %p not in flux", kn));
2112 	KASSERT((kn->kn_status & KN_DETACHED) != 0,
2113 	    ("knote %p was not detached", kn));
2114 	if (!islocked)
2115 		knl->kl_lock(knl->kl_lockarg);
2116 	SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext);
2117 	if (!islocked)
2118 		knl->kl_unlock(knl->kl_lockarg);
2119 	KQ_LOCK(kn->kn_kq);
2120 	kn->kn_knlist = knl;
2121 	kn->kn_status &= ~KN_DETACHED;
2122 	KQ_UNLOCK(kn->kn_kq);
2123 }
2124 
2125 static void
2126 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked,
2127     int kqislocked)
2128 {
2129 
2130 	KASSERT(!kqislocked || knlislocked, ("kq locked w/o knl locked"));
2131 	KNL_ASSERT_LOCK(knl, knlislocked);
2132 	mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED);
2133 	KASSERT(kqislocked || kn_in_flux(kn), ("knote %p not in flux", kn));
2134 	KASSERT((kn->kn_status & KN_DETACHED) == 0,
2135 	    ("knote %p was already detached", kn));
2136 	if (!knlislocked)
2137 		knl->kl_lock(knl->kl_lockarg);
2138 	SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext);
2139 	kn->kn_knlist = NULL;
2140 	if (!knlislocked)
2141 		kn_list_unlock(knl);
2142 	if (!kqislocked)
2143 		KQ_LOCK(kn->kn_kq);
2144 	kn->kn_status |= KN_DETACHED;
2145 	if (!kqislocked)
2146 		KQ_UNLOCK(kn->kn_kq);
2147 }
2148 
2149 /*
2150  * remove knote from the specified knlist
2151  */
2152 void
2153 knlist_remove(struct knlist *knl, struct knote *kn, int islocked)
2154 {
2155 
2156 	knlist_remove_kq(knl, kn, islocked, 0);
2157 }
2158 
2159 int
2160 knlist_empty(struct knlist *knl)
2161 {
2162 
2163 	KNL_ASSERT_LOCKED(knl);
2164 	return (SLIST_EMPTY(&knl->kl_list));
2165 }
2166 
2167 static struct mtx knlist_lock;
2168 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects",
2169     MTX_DEF);
2170 static void knlist_mtx_lock(void *arg);
2171 static void knlist_mtx_unlock(void *arg);
2172 
2173 static void
2174 knlist_mtx_lock(void *arg)
2175 {
2176 
2177 	mtx_lock((struct mtx *)arg);
2178 }
2179 
2180 static void
2181 knlist_mtx_unlock(void *arg)
2182 {
2183 
2184 	mtx_unlock((struct mtx *)arg);
2185 }
2186 
2187 static void
2188 knlist_mtx_assert_locked(void *arg)
2189 {
2190 
2191 	mtx_assert((struct mtx *)arg, MA_OWNED);
2192 }
2193 
2194 static void
2195 knlist_mtx_assert_unlocked(void *arg)
2196 {
2197 
2198 	mtx_assert((struct mtx *)arg, MA_NOTOWNED);
2199 }
2200 
2201 static void
2202 knlist_rw_rlock(void *arg)
2203 {
2204 
2205 	rw_rlock((struct rwlock *)arg);
2206 }
2207 
2208 static void
2209 knlist_rw_runlock(void *arg)
2210 {
2211 
2212 	rw_runlock((struct rwlock *)arg);
2213 }
2214 
2215 static void
2216 knlist_rw_assert_locked(void *arg)
2217 {
2218 
2219 	rw_assert((struct rwlock *)arg, RA_LOCKED);
2220 }
2221 
2222 static void
2223 knlist_rw_assert_unlocked(void *arg)
2224 {
2225 
2226 	rw_assert((struct rwlock *)arg, RA_UNLOCKED);
2227 }
2228 
2229 void
2230 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *),
2231     void (*kl_unlock)(void *),
2232     void (*kl_assert_locked)(void *), void (*kl_assert_unlocked)(void *))
2233 {
2234 
2235 	if (lock == NULL)
2236 		knl->kl_lockarg = &knlist_lock;
2237 	else
2238 		knl->kl_lockarg = lock;
2239 
2240 	if (kl_lock == NULL)
2241 		knl->kl_lock = knlist_mtx_lock;
2242 	else
2243 		knl->kl_lock = kl_lock;
2244 	if (kl_unlock == NULL)
2245 		knl->kl_unlock = knlist_mtx_unlock;
2246 	else
2247 		knl->kl_unlock = kl_unlock;
2248 	if (kl_assert_locked == NULL)
2249 		knl->kl_assert_locked = knlist_mtx_assert_locked;
2250 	else
2251 		knl->kl_assert_locked = kl_assert_locked;
2252 	if (kl_assert_unlocked == NULL)
2253 		knl->kl_assert_unlocked = knlist_mtx_assert_unlocked;
2254 	else
2255 		knl->kl_assert_unlocked = kl_assert_unlocked;
2256 
2257 	knl->kl_autodestroy = 0;
2258 	SLIST_INIT(&knl->kl_list);
2259 }
2260 
2261 void
2262 knlist_init_mtx(struct knlist *knl, struct mtx *lock)
2263 {
2264 
2265 	knlist_init(knl, lock, NULL, NULL, NULL, NULL);
2266 }
2267 
2268 struct knlist *
2269 knlist_alloc(struct mtx *lock)
2270 {
2271 	struct knlist *knl;
2272 
2273 	knl = malloc(sizeof(struct knlist), M_KQUEUE, M_WAITOK);
2274 	knlist_init_mtx(knl, lock);
2275 	return (knl);
2276 }
2277 
2278 void
2279 knlist_init_rw_reader(struct knlist *knl, struct rwlock *lock)
2280 {
2281 
2282 	knlist_init(knl, lock, knlist_rw_rlock, knlist_rw_runlock,
2283 	    knlist_rw_assert_locked, knlist_rw_assert_unlocked);
2284 }
2285 
2286 void
2287 knlist_destroy(struct knlist *knl)
2288 {
2289 
2290 	KASSERT(KNLIST_EMPTY(knl),
2291 	    ("destroying knlist %p with knotes on it", knl));
2292 }
2293 
2294 void
2295 knlist_detach(struct knlist *knl)
2296 {
2297 
2298 	KNL_ASSERT_LOCKED(knl);
2299 	knl->kl_autodestroy = 1;
2300 	if (knlist_empty(knl)) {
2301 		knlist_destroy(knl);
2302 		free(knl, M_KQUEUE);
2303 	}
2304 }
2305 
2306 /*
2307  * Even if we are locked, we may need to drop the lock to allow any influx
2308  * knotes time to "settle".
2309  */
2310 void
2311 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn)
2312 {
2313 	struct knote *kn, *kn2;
2314 	struct kqueue *kq;
2315 
2316 	KASSERT(!knl->kl_autodestroy, ("cleardel for autodestroy %p", knl));
2317 	if (islocked)
2318 		KNL_ASSERT_LOCKED(knl);
2319 	else {
2320 		KNL_ASSERT_UNLOCKED(knl);
2321 again:		/* need to reacquire lock since we have dropped it */
2322 		knl->kl_lock(knl->kl_lockarg);
2323 	}
2324 
2325 	SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) {
2326 		kq = kn->kn_kq;
2327 		KQ_LOCK(kq);
2328 		if (kn_in_flux(kn)) {
2329 			KQ_UNLOCK(kq);
2330 			continue;
2331 		}
2332 		knlist_remove_kq(knl, kn, 1, 1);
2333 		if (killkn) {
2334 			kn_enter_flux(kn);
2335 			KQ_UNLOCK(kq);
2336 			knote_drop_detached(kn, td);
2337 		} else {
2338 			/* Make sure cleared knotes disappear soon */
2339 			kn->kn_flags |= EV_EOF | EV_ONESHOT;
2340 			KQ_UNLOCK(kq);
2341 		}
2342 		kq = NULL;
2343 	}
2344 
2345 	if (!SLIST_EMPTY(&knl->kl_list)) {
2346 		/* there are still in flux knotes remaining */
2347 		kn = SLIST_FIRST(&knl->kl_list);
2348 		kq = kn->kn_kq;
2349 		KQ_LOCK(kq);
2350 		KASSERT(kn_in_flux(kn), ("knote removed w/o list lock"));
2351 		knl->kl_unlock(knl->kl_lockarg);
2352 		kq->kq_state |= KQ_FLUXWAIT;
2353 		msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0);
2354 		kq = NULL;
2355 		goto again;
2356 	}
2357 
2358 	if (islocked)
2359 		KNL_ASSERT_LOCKED(knl);
2360 	else {
2361 		knl->kl_unlock(knl->kl_lockarg);
2362 		KNL_ASSERT_UNLOCKED(knl);
2363 	}
2364 }
2365 
2366 /*
2367  * Remove all knotes referencing a specified fd must be called with FILEDESC
2368  * lock.  This prevents a race where a new fd comes along and occupies the
2369  * entry and we attach a knote to the fd.
2370  */
2371 void
2372 knote_fdclose(struct thread *td, int fd)
2373 {
2374 	struct filedesc *fdp = td->td_proc->p_fd;
2375 	struct kqueue *kq;
2376 	struct knote *kn;
2377 	int influx;
2378 
2379 	FILEDESC_XLOCK_ASSERT(fdp);
2380 
2381 	/*
2382 	 * We shouldn't have to worry about new kevents appearing on fd
2383 	 * since filedesc is locked.
2384 	 */
2385 	TAILQ_FOREACH(kq, &fdp->fd_kqlist, kq_list) {
2386 		KQ_LOCK(kq);
2387 
2388 again:
2389 		influx = 0;
2390 		while (kq->kq_knlistsize > fd &&
2391 		    (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) {
2392 			if (kn_in_flux(kn)) {
2393 				/* someone else might be waiting on our knote */
2394 				if (influx)
2395 					wakeup(kq);
2396 				kq->kq_state |= KQ_FLUXWAIT;
2397 				msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0);
2398 				goto again;
2399 			}
2400 			kn_enter_flux(kn);
2401 			KQ_UNLOCK(kq);
2402 			influx = 1;
2403 			knote_drop(kn, td);
2404 			KQ_LOCK(kq);
2405 		}
2406 		KQ_UNLOCK_FLUX(kq);
2407 	}
2408 }
2409 
2410 static int
2411 knote_attach(struct knote *kn, struct kqueue *kq)
2412 {
2413 	struct klist *list;
2414 
2415 	KASSERT(kn_in_flux(kn), ("knote %p not marked influx", kn));
2416 	KQ_OWNED(kq);
2417 
2418 	if (kn->kn_fop->f_isfd) {
2419 		if (kn->kn_id >= kq->kq_knlistsize)
2420 			return (ENOMEM);
2421 		list = &kq->kq_knlist[kn->kn_id];
2422 	} else {
2423 		if (kq->kq_knhash == NULL)
2424 			return (ENOMEM);
2425 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2426 	}
2427 	SLIST_INSERT_HEAD(list, kn, kn_link);
2428 	return (0);
2429 }
2430 
2431 static void
2432 knote_drop(struct knote *kn, struct thread *td)
2433 {
2434 
2435 	if ((kn->kn_status & KN_DETACHED) == 0)
2436 		kn->kn_fop->f_detach(kn);
2437 	knote_drop_detached(kn, td);
2438 }
2439 
2440 static void
2441 knote_drop_detached(struct knote *kn, struct thread *td)
2442 {
2443 	struct kqueue *kq;
2444 	struct klist *list;
2445 
2446 	kq = kn->kn_kq;
2447 
2448 	KASSERT((kn->kn_status & KN_DETACHED) != 0,
2449 	    ("knote %p still attached", kn));
2450 	KQ_NOTOWNED(kq);
2451 
2452 	KQ_LOCK(kq);
2453 	KASSERT(kn->kn_influx == 1,
2454 	    ("knote_drop called on %p with influx %d", kn, kn->kn_influx));
2455 
2456 	if (kn->kn_fop->f_isfd)
2457 		list = &kq->kq_knlist[kn->kn_id];
2458 	else
2459 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
2460 
2461 	if (!SLIST_EMPTY(list))
2462 		SLIST_REMOVE(list, kn, knote, kn_link);
2463 	if (kn->kn_status & KN_QUEUED)
2464 		knote_dequeue(kn);
2465 	KQ_UNLOCK_FLUX(kq);
2466 
2467 	if (kn->kn_fop->f_isfd) {
2468 		fdrop(kn->kn_fp, td);
2469 		kn->kn_fp = NULL;
2470 	}
2471 	kqueue_fo_release(kn->kn_kevent.filter);
2472 	kn->kn_fop = NULL;
2473 	knote_free(kn);
2474 }
2475 
2476 static void
2477 knote_enqueue(struct knote *kn)
2478 {
2479 	struct kqueue *kq = kn->kn_kq;
2480 
2481 	KQ_OWNED(kn->kn_kq);
2482 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
2483 
2484 	TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe);
2485 	kn->kn_status |= KN_QUEUED;
2486 	kq->kq_count++;
2487 	kqueue_wakeup(kq);
2488 }
2489 
2490 static void
2491 knote_dequeue(struct knote *kn)
2492 {
2493 	struct kqueue *kq = kn->kn_kq;
2494 
2495 	KQ_OWNED(kn->kn_kq);
2496 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
2497 
2498 	TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe);
2499 	kn->kn_status &= ~KN_QUEUED;
2500 	kq->kq_count--;
2501 }
2502 
2503 static void
2504 knote_init(void)
2505 {
2506 
2507 	knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL,
2508 	    NULL, NULL, UMA_ALIGN_PTR, 0);
2509 }
2510 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL);
2511 
2512 static struct knote *
2513 knote_alloc(int waitok)
2514 {
2515 
2516 	return (uma_zalloc(knote_zone, (waitok ? M_WAITOK : M_NOWAIT) |
2517 	    M_ZERO));
2518 }
2519 
2520 static void
2521 knote_free(struct knote *kn)
2522 {
2523 
2524 	uma_zfree(knote_zone, kn);
2525 }
2526 
2527 /*
2528  * Register the kev w/ the kq specified by fd.
2529  */
2530 int
2531 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok)
2532 {
2533 	struct kqueue *kq;
2534 	struct file *fp;
2535 	cap_rights_t rights;
2536 	int error;
2537 
2538 	error = fget(td, fd, cap_rights_init(&rights, CAP_KQUEUE_CHANGE), &fp);
2539 	if (error != 0)
2540 		return (error);
2541 	if ((error = kqueue_acquire(fp, &kq)) != 0)
2542 		goto noacquire;
2543 
2544 	error = kqueue_register(kq, kev, td, waitok);
2545 	kqueue_release(kq, 0);
2546 
2547 noacquire:
2548 	fdrop(fp, td);
2549 	return (error);
2550 }
2551