xref: /dragonfly/sys/kern/kern_event.c (revision 6e278935)
1 /*-
2  * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  * $FreeBSD: src/sys/kern/kern_event.c,v 1.2.2.10 2004/04/04 07:03:14 cperciva Exp $
27  */
28 
29 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/kernel.h>
32 #include <sys/proc.h>
33 #include <sys/malloc.h>
34 #include <sys/unistd.h>
35 #include <sys/file.h>
36 #include <sys/lock.h>
37 #include <sys/fcntl.h>
38 #include <sys/queue.h>
39 #include <sys/event.h>
40 #include <sys/eventvar.h>
41 #include <sys/protosw.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
44 #include <sys/stat.h>
45 #include <sys/sysctl.h>
46 #include <sys/sysproto.h>
47 #include <sys/thread.h>
48 #include <sys/uio.h>
49 #include <sys/signalvar.h>
50 #include <sys/filio.h>
51 #include <sys/ktr.h>
52 
53 #include <sys/thread2.h>
54 #include <sys/file2.h>
55 #include <sys/mplock2.h>
56 
57 /*
58  * Global token for kqueue subsystem
59  */
60 #if 0
61 struct lwkt_token kq_token = LWKT_TOKEN_INITIALIZER(kq_token);
62 SYSCTL_LONG(_lwkt, OID_AUTO, kq_collisions,
63     CTLFLAG_RW, &kq_token.t_collisions, 0,
64     "Collision counter of kq_token");
65 #endif
66 
67 MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
68 
69 struct kevent_copyin_args {
70 	struct kevent_args	*ka;
71 	int			pchanges;
72 };
73 
74 static int	kqueue_sleep(struct kqueue *kq, struct timespec *tsp);
75 static int	kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
76 		    struct knote *marker);
77 static int 	kqueue_read(struct file *fp, struct uio *uio,
78 		    struct ucred *cred, int flags);
79 static int	kqueue_write(struct file *fp, struct uio *uio,
80 		    struct ucred *cred, int flags);
81 static int	kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
82 		    struct ucred *cred, struct sysmsg *msg);
83 static int 	kqueue_kqfilter(struct file *fp, struct knote *kn);
84 static int 	kqueue_stat(struct file *fp, struct stat *st,
85 		    struct ucred *cred);
86 static int 	kqueue_close(struct file *fp);
87 static void	kqueue_wakeup(struct kqueue *kq);
88 static int	filter_attach(struct knote *kn);
89 static int	filter_event(struct knote *kn, long hint);
90 
91 /*
92  * MPSAFE
93  */
94 static struct fileops kqueueops = {
95 	.fo_read = kqueue_read,
96 	.fo_write = kqueue_write,
97 	.fo_ioctl = kqueue_ioctl,
98 	.fo_kqfilter = kqueue_kqfilter,
99 	.fo_stat = kqueue_stat,
100 	.fo_close = kqueue_close,
101 	.fo_shutdown = nofo_shutdown
102 };
103 
104 static void 	knote_attach(struct knote *kn);
105 static void 	knote_drop(struct knote *kn);
106 static void	knote_detach_and_drop(struct knote *kn);
107 static void 	knote_enqueue(struct knote *kn);
108 static void 	knote_dequeue(struct knote *kn);
109 static struct 	knote *knote_alloc(void);
110 static void 	knote_free(struct knote *kn);
111 
112 static void	filt_kqdetach(struct knote *kn);
113 static int	filt_kqueue(struct knote *kn, long hint);
114 static int	filt_procattach(struct knote *kn);
115 static void	filt_procdetach(struct knote *kn);
116 static int	filt_proc(struct knote *kn, long hint);
117 static int	filt_fileattach(struct knote *kn);
118 static void	filt_timerexpire(void *knx);
119 static int	filt_timerattach(struct knote *kn);
120 static void	filt_timerdetach(struct knote *kn);
121 static int	filt_timer(struct knote *kn, long hint);
122 
123 static struct filterops file_filtops =
124 	{ FILTEROP_ISFD, filt_fileattach, NULL, NULL };
125 static struct filterops kqread_filtops =
126 	{ FILTEROP_ISFD, NULL, filt_kqdetach, filt_kqueue };
127 static struct filterops proc_filtops =
128 	{ 0, filt_procattach, filt_procdetach, filt_proc };
129 static struct filterops timer_filtops =
130 	{ 0, filt_timerattach, filt_timerdetach, filt_timer };
131 
132 static int 		kq_ncallouts = 0;
133 static int 		kq_calloutmax = (4 * 1024);
134 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
135     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
136 static int		kq_checkloop = 1000000;
137 SYSCTL_INT(_kern, OID_AUTO, kq_checkloop, CTLFLAG_RW,
138     &kq_checkloop, 0, "Maximum number of callouts allocated for kqueue");
139 
140 #define KNOTE_ACTIVATE(kn) do { 					\
141 	kn->kn_status |= KN_ACTIVE;					\
142 	if ((kn->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)		\
143 		knote_enqueue(kn);					\
144 } while(0)
145 
146 #define	KN_HASHSIZE		64		/* XXX should be tunable */
147 #define KN_HASH(val, mask)	(((val) ^ (val >> 8)) & (mask))
148 
149 extern struct filterops aio_filtops;
150 extern struct filterops sig_filtops;
151 
152 /*
153  * Table for for all system-defined filters.
154  */
155 static struct filterops *sysfilt_ops[] = {
156 	&file_filtops,			/* EVFILT_READ */
157 	&file_filtops,			/* EVFILT_WRITE */
158 	&aio_filtops,			/* EVFILT_AIO */
159 	&file_filtops,			/* EVFILT_VNODE */
160 	&proc_filtops,			/* EVFILT_PROC */
161 	&sig_filtops,			/* EVFILT_SIGNAL */
162 	&timer_filtops,			/* EVFILT_TIMER */
163 	&file_filtops,			/* EVFILT_EXCEPT */
164 };
165 
166 static int
167 filt_fileattach(struct knote *kn)
168 {
169 	return (fo_kqfilter(kn->kn_fp, kn));
170 }
171 
172 /*
173  * MPSAFE
174  */
175 static int
176 kqueue_kqfilter(struct file *fp, struct knote *kn)
177 {
178 	struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
179 
180 	if (kn->kn_filter != EVFILT_READ)
181 		return (EOPNOTSUPP);
182 
183 	kn->kn_fop = &kqread_filtops;
184 	knote_insert(&kq->kq_kqinfo.ki_note, kn);
185 	return (0);
186 }
187 
188 static void
189 filt_kqdetach(struct knote *kn)
190 {
191 	struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
192 
193 	knote_remove(&kq->kq_kqinfo.ki_note, kn);
194 }
195 
196 /*ARGSUSED*/
197 static int
198 filt_kqueue(struct knote *kn, long hint)
199 {
200 	struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
201 
202 	kn->kn_data = kq->kq_count;
203 	return (kn->kn_data > 0);
204 }
205 
206 static int
207 filt_procattach(struct knote *kn)
208 {
209 	struct proc *p;
210 	int immediate;
211 
212 	immediate = 0;
213 	p = pfind(kn->kn_id);
214 	if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
215 		p = zpfind(kn->kn_id);
216 		immediate = 1;
217 	}
218 	if (p == NULL) {
219 		return (ESRCH);
220 	}
221 	if (!PRISON_CHECK(curthread->td_ucred, p->p_ucred)) {
222 		if (p)
223 			PRELE(p);
224 		return (EACCES);
225 	}
226 
227 	lwkt_gettoken(&p->p_token);
228 	kn->kn_ptr.p_proc = p;
229 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
230 
231 	/*
232 	 * internal flag indicating registration done by kernel
233 	 */
234 	if (kn->kn_flags & EV_FLAG1) {
235 		kn->kn_data = kn->kn_sdata;		/* ppid */
236 		kn->kn_fflags = NOTE_CHILD;
237 		kn->kn_flags &= ~EV_FLAG1;
238 	}
239 
240 	knote_insert(&p->p_klist, kn);
241 
242 	/*
243 	 * Immediately activate any exit notes if the target process is a
244 	 * zombie.  This is necessary to handle the case where the target
245 	 * process, e.g. a child, dies before the kevent is negistered.
246 	 */
247 	if (immediate && filt_proc(kn, NOTE_EXIT))
248 		KNOTE_ACTIVATE(kn);
249 	lwkt_reltoken(&p->p_token);
250 	PRELE(p);
251 
252 	return (0);
253 }
254 
255 /*
256  * The knote may be attached to a different process, which may exit,
257  * leaving nothing for the knote to be attached to.  So when the process
258  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
259  * it will be deleted when read out.  However, as part of the knote deletion,
260  * this routine is called, so a check is needed to avoid actually performing
261  * a detach, because the original process does not exist any more.
262  */
263 static void
264 filt_procdetach(struct knote *kn)
265 {
266 	struct proc *p;
267 
268 	if (kn->kn_status & KN_DETACHED)
269 		return;
270 	/* XXX locking? take proc_token here? */
271 	p = kn->kn_ptr.p_proc;
272 	knote_remove(&p->p_klist, kn);
273 }
274 
275 static int
276 filt_proc(struct knote *kn, long hint)
277 {
278 	u_int event;
279 
280 	/*
281 	 * mask off extra data
282 	 */
283 	event = (u_int)hint & NOTE_PCTRLMASK;
284 
285 	/*
286 	 * if the user is interested in this event, record it.
287 	 */
288 	if (kn->kn_sfflags & event)
289 		kn->kn_fflags |= event;
290 
291 	/*
292 	 * Process is gone, so flag the event as finished.  Detach the
293 	 * knote from the process now because the process will be poof,
294 	 * gone later on.
295 	 */
296 	if (event == NOTE_EXIT) {
297 		struct proc *p = kn->kn_ptr.p_proc;
298 		if ((kn->kn_status & KN_DETACHED) == 0) {
299 			PHOLD(p);
300 			knote_remove(&p->p_klist, kn);
301 			kn->kn_status |= KN_DETACHED;
302 			kn->kn_data = p->p_xstat;
303 			kn->kn_ptr.p_proc = NULL;
304 			PRELE(p);
305 		}
306 		kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
307 		return (1);
308 	}
309 
310 	/*
311 	 * process forked, and user wants to track the new process,
312 	 * so attach a new knote to it, and immediately report an
313 	 * event with the parent's pid.
314 	 */
315 	if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) {
316 		struct kevent kev;
317 		int error;
318 
319 		/*
320 		 * register knote with new process.
321 		 */
322 		kev.ident = hint & NOTE_PDATAMASK;	/* pid */
323 		kev.filter = kn->kn_filter;
324 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
325 		kev.fflags = kn->kn_sfflags;
326 		kev.data = kn->kn_id;			/* parent */
327 		kev.udata = kn->kn_kevent.udata;	/* preserve udata */
328 		error = kqueue_register(kn->kn_kq, &kev);
329 		if (error)
330 			kn->kn_fflags |= NOTE_TRACKERR;
331 	}
332 
333 	return (kn->kn_fflags != 0);
334 }
335 
336 /*
337  * The callout interlocks with callout_terminate() but can still
338  * race a deletion so if KN_DELETING is set we just don't touch
339  * the knote.
340  */
341 static void
342 filt_timerexpire(void *knx)
343 {
344 	struct lwkt_token *tok;
345 	struct knote *kn = knx;
346 	struct callout *calloutp;
347 	struct timeval tv;
348 	int tticks;
349 
350 	tok = lwkt_token_pool_lookup(kn->kn_kq);
351 	lwkt_gettoken(tok);
352 	if ((kn->kn_status & KN_DELETING) == 0) {
353 		kn->kn_data++;
354 		KNOTE_ACTIVATE(kn);
355 
356 		if ((kn->kn_flags & EV_ONESHOT) == 0) {
357 			tv.tv_sec = kn->kn_sdata / 1000;
358 			tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
359 			tticks = tvtohz_high(&tv);
360 			calloutp = (struct callout *)kn->kn_hook;
361 			callout_reset(calloutp, tticks, filt_timerexpire, kn);
362 		}
363 	}
364 	lwkt_reltoken(tok);
365 }
366 
367 /*
368  * data contains amount of time to sleep, in milliseconds
369  */
370 static int
371 filt_timerattach(struct knote *kn)
372 {
373 	struct callout *calloutp;
374 	struct timeval tv;
375 	int tticks;
376 
377 	if (kq_ncallouts >= kq_calloutmax) {
378 		kn->kn_hook = NULL;
379 		return (ENOMEM);
380 	}
381 	kq_ncallouts++;
382 
383 	tv.tv_sec = kn->kn_sdata / 1000;
384 	tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
385 	tticks = tvtohz_high(&tv);
386 
387 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
388 	calloutp = kmalloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
389 	callout_init(calloutp);
390 	kn->kn_hook = (caddr_t)calloutp;
391 	callout_reset(calloutp, tticks, filt_timerexpire, kn);
392 
393 	return (0);
394 }
395 
396 /*
397  * This function is called with the knote flagged locked but it is
398  * still possible to race a callout event due to the callback blocking.
399  * We must call callout_terminate() instead of callout_stop() to deal
400  * with the race.
401  */
402 static void
403 filt_timerdetach(struct knote *kn)
404 {
405 	struct callout *calloutp;
406 
407 	calloutp = (struct callout *)kn->kn_hook;
408 	callout_terminate(calloutp);
409 	kfree(calloutp, M_KQUEUE);
410 	kq_ncallouts--;
411 }
412 
413 static int
414 filt_timer(struct knote *kn, long hint)
415 {
416 
417 	return (kn->kn_data != 0);
418 }
419 
420 /*
421  * Acquire a knote, return non-zero on success, 0 on failure.
422  *
423  * If we cannot acquire the knote we sleep and return 0.  The knote
424  * may be stale on return in this case and the caller must restart
425  * whatever loop they are in.
426  *
427  * Related kq token must be held.
428  */
429 static __inline
430 int
431 knote_acquire(struct knote *kn)
432 {
433 	if (kn->kn_status & KN_PROCESSING) {
434 		kn->kn_status |= KN_WAITING | KN_REPROCESS;
435 		tsleep(kn, 0, "kqepts", hz);
436 		/* knote may be stale now */
437 		return(0);
438 	}
439 	kn->kn_status |= KN_PROCESSING;
440 	return(1);
441 }
442 
443 /*
444  * Release an acquired knote, clearing KN_PROCESSING and handling any
445  * KN_REPROCESS events.
446  *
447  * Caller must be holding the related kq token
448  *
449  * Non-zero is returned if the knote is destroyed or detached.
450  */
451 static __inline
452 int
453 knote_release(struct knote *kn)
454 {
455 	while (kn->kn_status & KN_REPROCESS) {
456 		kn->kn_status &= ~KN_REPROCESS;
457 		if (kn->kn_status & KN_WAITING) {
458 			kn->kn_status &= ~KN_WAITING;
459 			wakeup(kn);
460 		}
461 		if (kn->kn_status & KN_DELETING) {
462 			knote_detach_and_drop(kn);
463 			return(1);
464 			/* NOT REACHED */
465 		}
466 		if (filter_event(kn, 0))
467 			KNOTE_ACTIVATE(kn);
468 	}
469 	if (kn->kn_status & KN_DETACHED) {
470 		kn->kn_status &= ~KN_PROCESSING;
471 		return(1);
472 	} else {
473 		kn->kn_status &= ~KN_PROCESSING;
474 		return(0);
475 	}
476 }
477 
478 /*
479  * Initialize a kqueue.
480  *
481  * NOTE: The lwp/proc code initializes a kqueue for select/poll ops.
482  *
483  * MPSAFE
484  */
485 void
486 kqueue_init(struct kqueue *kq, struct filedesc *fdp)
487 {
488 	TAILQ_INIT(&kq->kq_knpend);
489 	TAILQ_INIT(&kq->kq_knlist);
490 	kq->kq_count = 0;
491 	kq->kq_fdp = fdp;
492 	SLIST_INIT(&kq->kq_kqinfo.ki_note);
493 }
494 
495 /*
496  * Terminate a kqueue.  Freeing the actual kq itself is left up to the
497  * caller (it might be embedded in a lwp so we don't do it here).
498  *
499  * The kq's knlist must be completely eradicated so block on any
500  * processing races.
501  */
502 void
503 kqueue_terminate(struct kqueue *kq)
504 {
505 	struct lwkt_token *tok;
506 	struct knote *kn;
507 
508 	tok = lwkt_token_pool_lookup(kq);
509 	lwkt_gettoken(tok);
510 	while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) {
511 		if (knote_acquire(kn))
512 			knote_detach_and_drop(kn);
513 	}
514 	if (kq->kq_knhash) {
515 		kfree(kq->kq_knhash, M_KQUEUE);
516 		kq->kq_knhash = NULL;
517 		kq->kq_knhashmask = 0;
518 	}
519 	lwkt_reltoken(tok);
520 }
521 
522 /*
523  * MPSAFE
524  */
525 int
526 sys_kqueue(struct kqueue_args *uap)
527 {
528 	struct thread *td = curthread;
529 	struct kqueue *kq;
530 	struct file *fp;
531 	int fd, error;
532 
533 	error = falloc(td->td_lwp, &fp, &fd);
534 	if (error)
535 		return (error);
536 	fp->f_flag = FREAD | FWRITE;
537 	fp->f_type = DTYPE_KQUEUE;
538 	fp->f_ops = &kqueueops;
539 
540 	kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO);
541 	kqueue_init(kq, td->td_proc->p_fd);
542 	fp->f_data = kq;
543 
544 	fsetfd(kq->kq_fdp, fp, fd);
545 	uap->sysmsg_result = fd;
546 	fdrop(fp);
547 	return (error);
548 }
549 
550 /*
551  * Copy 'count' items into the destination list pointed to by uap->eventlist.
552  */
553 static int
554 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res)
555 {
556 	struct kevent_copyin_args *kap;
557 	int error;
558 
559 	kap = (struct kevent_copyin_args *)arg;
560 
561 	error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp));
562 	if (error == 0) {
563 		kap->ka->eventlist += count;
564 		*res += count;
565 	} else {
566 		*res = -1;
567 	}
568 
569 	return (error);
570 }
571 
572 /*
573  * Copy at most 'max' items from the list pointed to by kap->changelist,
574  * return number of items in 'events'.
575  */
576 static int
577 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events)
578 {
579 	struct kevent_copyin_args *kap;
580 	int error, count;
581 
582 	kap = (struct kevent_copyin_args *)arg;
583 
584 	count = min(kap->ka->nchanges - kap->pchanges, max);
585 	error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp);
586 	if (error == 0) {
587 		kap->ka->changelist += count;
588 		kap->pchanges += count;
589 		*events = count;
590 	}
591 
592 	return (error);
593 }
594 
595 /*
596  * MPSAFE
597  */
598 int
599 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap,
600 	    k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn,
601 	    struct timespec *tsp_in)
602 {
603 	struct kevent *kevp;
604 	struct timespec *tsp;
605 	int i, n, total, error, nerrors = 0;
606 	int lres;
607 	int limit = kq_checkloop;
608 	struct kevent kev[KQ_NEVENTS];
609 	struct knote marker;
610 	struct lwkt_token *tok;
611 
612 	tsp = tsp_in;
613 	*res = 0;
614 
615 	tok = lwkt_token_pool_lookup(kq);
616 	lwkt_gettoken(tok);
617 	for ( ;; ) {
618 		n = 0;
619 		error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n);
620 		if (error)
621 			goto done;
622 		if (n == 0)
623 			break;
624 		for (i = 0; i < n; i++) {
625 			kevp = &kev[i];
626 			kevp->flags &= ~EV_SYSFLAGS;
627 			error = kqueue_register(kq, kevp);
628 
629 			/*
630 			 * If a registration returns an error we
631 			 * immediately post the error.  The kevent()
632 			 * call itself will fail with the error if
633 			 * no space is available for posting.
634 			 *
635 			 * Such errors normally bypass the timeout/blocking
636 			 * code.  However, if the copyoutfn function refuses
637 			 * to post the error (see sys_poll()), then we
638 			 * ignore it too.
639 			 */
640 			if (error) {
641 				kevp->flags = EV_ERROR;
642 				kevp->data = error;
643 				lres = *res;
644 				kevent_copyoutfn(uap, kevp, 1, res);
645 				if (*res < 0) {
646 					goto done;
647 				} else if (lres != *res) {
648 					nevents--;
649 					nerrors++;
650 				}
651 			}
652 		}
653 	}
654 	if (nerrors) {
655 		error = 0;
656 		goto done;
657 	}
658 
659 	/*
660 	 * Acquire/wait for events - setup timeout
661 	 */
662 	if (tsp != NULL) {
663 		struct timespec ats;
664 
665 		if (tsp->tv_sec || tsp->tv_nsec) {
666 			nanouptime(&ats);
667 			timespecadd(tsp, &ats);		/* tsp = target time */
668 		}
669 	}
670 
671 	/*
672 	 * Loop as required.
673 	 *
674 	 * Collect as many events as we can. Sleeping on successive
675 	 * loops is disabled if copyoutfn has incremented (*res).
676 	 *
677 	 * The loop stops if an error occurs, all events have been
678 	 * scanned (the marker has been reached), or fewer than the
679 	 * maximum number of events is found.
680 	 *
681 	 * The copyoutfn function does not have to increment (*res) in
682 	 * order for the loop to continue.
683 	 *
684 	 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents.
685 	 */
686 	total = 0;
687 	error = 0;
688 	marker.kn_filter = EVFILT_MARKER;
689 	marker.kn_status = KN_PROCESSING;
690 	TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
691 	while ((n = nevents - total) > 0) {
692 		if (n > KQ_NEVENTS)
693 			n = KQ_NEVENTS;
694 
695 		/*
696 		 * If no events are pending sleep until timeout (if any)
697 		 * or an event occurs.
698 		 *
699 		 * After the sleep completes the marker is moved to the
700 		 * end of the list, making any received events available
701 		 * to our scan.
702 		 */
703 		if (kq->kq_count == 0 && *res == 0) {
704 			error = kqueue_sleep(kq, tsp);
705 			if (error)
706 				break;
707 
708 			TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
709 			TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
710 		}
711 
712 		/*
713 		 * Process all received events
714 		 * Account for all non-spurious events in our total
715 		 */
716 		i = kqueue_scan(kq, kev, n, &marker);
717 		if (i) {
718 			lres = *res;
719 			error = kevent_copyoutfn(uap, kev, i, res);
720 			total += *res - lres;
721 			if (error)
722 				break;
723 		}
724 		if (limit && --limit == 0)
725 			panic("kqueue: checkloop failed i=%d", i);
726 
727 		/*
728 		 * Normally when fewer events are returned than requested
729 		 * we can stop.  However, if only spurious events were
730 		 * collected the copyout will not bump (*res) and we have
731 		 * to continue.
732 		 */
733 		if (i < n && *res)
734 			break;
735 
736 		/*
737 		 * Deal with an edge case where spurious events can cause
738 		 * a loop to occur without moving the marker.  This can
739 		 * prevent kqueue_scan() from picking up new events which
740 		 * race us.  We must be sure to move the marker for this
741 		 * case.
742 		 *
743 		 * NOTE: We do not want to move the marker if events
744 		 *	 were scanned because normal kqueue operations
745 		 *	 may reactivate events.  Moving the marker in
746 		 *	 that case could result in duplicates for the
747 		 *	 same event.
748 		 */
749 		if (i == 0) {
750 			TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
751 			TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
752 		}
753 	}
754 	TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
755 
756 	/* Timeouts do not return EWOULDBLOCK. */
757 	if (error == EWOULDBLOCK)
758 		error = 0;
759 
760 done:
761 	lwkt_reltoken(tok);
762 	return (error);
763 }
764 
765 /*
766  * MPALMOSTSAFE
767  */
768 int
769 sys_kevent(struct kevent_args *uap)
770 {
771 	struct thread *td = curthread;
772 	struct proc *p = td->td_proc;
773 	struct timespec ts, *tsp;
774 	struct kqueue *kq;
775 	struct file *fp = NULL;
776 	struct kevent_copyin_args *kap, ka;
777 	int error;
778 
779 	if (uap->timeout) {
780 		error = copyin(uap->timeout, &ts, sizeof(ts));
781 		if (error)
782 			return (error);
783 		tsp = &ts;
784 	} else {
785 		tsp = NULL;
786 	}
787 
788 	fp = holdfp(p->p_fd, uap->fd, -1);
789 	if (fp == NULL)
790 		return (EBADF);
791 	if (fp->f_type != DTYPE_KQUEUE) {
792 		fdrop(fp);
793 		return (EBADF);
794 	}
795 
796 	kq = (struct kqueue *)fp->f_data;
797 
798 	kap = &ka;
799 	kap->ka = uap;
800 	kap->pchanges = 0;
801 
802 	error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap,
803 			    kevent_copyin, kevent_copyout, tsp);
804 
805 	fdrop(fp);
806 
807 	return (error);
808 }
809 
810 /*
811  * Caller must be holding the kq token
812  */
813 int
814 kqueue_register(struct kqueue *kq, struct kevent *kev)
815 {
816 	struct lwkt_token *tok;
817 	struct filedesc *fdp = kq->kq_fdp;
818 	struct filterops *fops;
819 	struct file *fp = NULL;
820 	struct knote *kn = NULL;
821 	int error = 0;
822 
823 	if (kev->filter < 0) {
824 		if (kev->filter + EVFILT_SYSCOUNT < 0)
825 			return (EINVAL);
826 		fops = sysfilt_ops[~kev->filter];	/* to 0-base index */
827 	} else {
828 		/*
829 		 * XXX
830 		 * filter attach routine is responsible for insuring that
831 		 * the identifier can be attached to it.
832 		 */
833 		kprintf("unknown filter: %d\n", kev->filter);
834 		return (EINVAL);
835 	}
836 
837 	tok = lwkt_token_pool_lookup(kq);
838 	lwkt_gettoken(tok);
839 	if (fops->f_flags & FILTEROP_ISFD) {
840 		/* validate descriptor */
841 		fp = holdfp(fdp, kev->ident, -1);
842 		if (fp == NULL) {
843 			lwkt_reltoken(tok);
844 			return (EBADF);
845 		}
846 		lwkt_getpooltoken(&fp->f_klist);
847 again1:
848 		SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
849 			if (kn->kn_kq == kq &&
850 			    kn->kn_filter == kev->filter &&
851 			    kn->kn_id == kev->ident) {
852 				if (knote_acquire(kn) == 0)
853 					goto again1;
854 				break;
855 			}
856 		}
857 		lwkt_relpooltoken(&fp->f_klist);
858 	} else {
859 		if (kq->kq_knhashmask) {
860 			struct klist *list;
861 
862 			list = &kq->kq_knhash[
863 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
864 			lwkt_getpooltoken(list);
865 again2:
866 			SLIST_FOREACH(kn, list, kn_link) {
867 				if (kn->kn_id == kev->ident &&
868 				    kn->kn_filter == kev->filter) {
869 					if (knote_acquire(kn) == 0)
870 						goto again2;
871 					break;
872 				}
873 			}
874 			lwkt_relpooltoken(list);
875 		}
876 	}
877 
878 	/*
879 	 * NOTE: At this point if kn is non-NULL we will have acquired
880 	 *	 it and set KN_PROCESSING.
881 	 */
882 	if (kn == NULL && ((kev->flags & EV_ADD) == 0)) {
883 		error = ENOENT;
884 		goto done;
885 	}
886 
887 	/*
888 	 * kn now contains the matching knote, or NULL if no match
889 	 */
890 	if (kev->flags & EV_ADD) {
891 		if (kn == NULL) {
892 			kn = knote_alloc();
893 			if (kn == NULL) {
894 				error = ENOMEM;
895 				goto done;
896 			}
897 			kn->kn_fp = fp;
898 			kn->kn_kq = kq;
899 			kn->kn_fop = fops;
900 
901 			/*
902 			 * apply reference count to knote structure, and
903 			 * do not release it at the end of this routine.
904 			 */
905 			fp = NULL;
906 
907 			kn->kn_sfflags = kev->fflags;
908 			kn->kn_sdata = kev->data;
909 			kev->fflags = 0;
910 			kev->data = 0;
911 			kn->kn_kevent = *kev;
912 
913 			/*
914 			 * KN_PROCESSING prevents the knote from getting
915 			 * ripped out from under us while we are trying
916 			 * to attach it, in case the attach blocks.
917 			 */
918 			kn->kn_status = KN_PROCESSING;
919 			knote_attach(kn);
920 			if ((error = filter_attach(kn)) != 0) {
921 				kn->kn_status |= KN_DELETING | KN_REPROCESS;
922 				knote_drop(kn);
923 				goto done;
924 			}
925 
926 			/*
927 			 * Interlock against close races which either tried
928 			 * to remove our knote while we were blocked or missed
929 			 * it entirely prior to our attachment.  We do not
930 			 * want to end up with a knote on a closed descriptor.
931 			 */
932 			if ((fops->f_flags & FILTEROP_ISFD) &&
933 			    checkfdclosed(fdp, kev->ident, kn->kn_fp)) {
934 				kn->kn_status |= KN_DELETING | KN_REPROCESS;
935 			}
936 		} else {
937 			/*
938 			 * The user may change some filter values after the
939 			 * initial EV_ADD, but doing so will not reset any
940 			 * filter which have already been triggered.
941 			 */
942 			KKASSERT(kn->kn_status & KN_PROCESSING);
943 			kn->kn_sfflags = kev->fflags;
944 			kn->kn_sdata = kev->data;
945 			kn->kn_kevent.udata = kev->udata;
946 		}
947 
948 		/*
949 		 * Execute the filter event to immediately activate the
950 		 * knote if necessary.  If reprocessing events are pending
951 		 * due to blocking above we do not run the filter here
952 		 * but instead let knote_release() do it.  Otherwise we
953 		 * might run the filter on a deleted event.
954 		 */
955 		if ((kn->kn_status & KN_REPROCESS) == 0) {
956 			if (filter_event(kn, 0))
957 				KNOTE_ACTIVATE(kn);
958 		}
959 	} else if (kev->flags & EV_DELETE) {
960 		/*
961 		 * Delete the existing knote
962 		 */
963 		knote_detach_and_drop(kn);
964 		goto done;
965 	}
966 
967 	/*
968 	 * Disablement does not deactivate a knote here.
969 	 */
970 	if ((kev->flags & EV_DISABLE) &&
971 	    ((kn->kn_status & KN_DISABLED) == 0)) {
972 		kn->kn_status |= KN_DISABLED;
973 	}
974 
975 	/*
976 	 * Re-enablement may have to immediately enqueue an active knote.
977 	 */
978 	if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
979 		kn->kn_status &= ~KN_DISABLED;
980 		if ((kn->kn_status & KN_ACTIVE) &&
981 		    ((kn->kn_status & KN_QUEUED) == 0)) {
982 			knote_enqueue(kn);
983 		}
984 	}
985 
986 	/*
987 	 * Handle any required reprocessing
988 	 */
989 	knote_release(kn);
990 	/* kn may be invalid now */
991 
992 done:
993 	lwkt_reltoken(tok);
994 	if (fp != NULL)
995 		fdrop(fp);
996 	return (error);
997 }
998 
999 /*
1000  * Block as necessary until the target time is reached.
1001  * If tsp is NULL we block indefinitely.  If tsp->ts_secs/nsecs are both
1002  * 0 we do not block at all.
1003  *
1004  * Caller must be holding the kq token.
1005  */
1006 static int
1007 kqueue_sleep(struct kqueue *kq, struct timespec *tsp)
1008 {
1009 	int error = 0;
1010 
1011 	if (tsp == NULL) {
1012 		kq->kq_state |= KQ_SLEEP;
1013 		error = tsleep(kq, PCATCH, "kqread", 0);
1014 	} else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) {
1015 		error = EWOULDBLOCK;
1016 	} else {
1017 		struct timespec ats;
1018 		struct timespec atx = *tsp;
1019 		int timeout;
1020 
1021 		nanouptime(&ats);
1022 		timespecsub(&atx, &ats);
1023 		if (ats.tv_sec < 0) {
1024 			error = EWOULDBLOCK;
1025 		} else {
1026 			timeout = atx.tv_sec > 24 * 60 * 60 ?
1027 				24 * 60 * 60 * hz : tstohz_high(&atx);
1028 			kq->kq_state |= KQ_SLEEP;
1029 			error = tsleep(kq, PCATCH, "kqread", timeout);
1030 		}
1031 	}
1032 
1033 	/* don't restart after signals... */
1034 	if (error == ERESTART)
1035 		return (EINTR);
1036 
1037 	return (error);
1038 }
1039 
1040 /*
1041  * Scan the kqueue, return the number of active events placed in kevp up
1042  * to count.
1043  *
1044  * Continuous mode events may get recycled, do not continue scanning past
1045  * marker unless no events have been collected.
1046  *
1047  * Caller must be holding the kq token
1048  */
1049 static int
1050 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
1051             struct knote *marker)
1052 {
1053         struct knote *kn, local_marker;
1054         int total;
1055 
1056         total = 0;
1057 	local_marker.kn_filter = EVFILT_MARKER;
1058 	local_marker.kn_status = KN_PROCESSING;
1059 
1060 	/*
1061 	 * Collect events.
1062 	 */
1063 	TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe);
1064 	while (count) {
1065 		kn = TAILQ_NEXT(&local_marker, kn_tqe);
1066 		if (kn->kn_filter == EVFILT_MARKER) {
1067 			/* Marker reached, we are done */
1068 			if (kn == marker)
1069 				break;
1070 
1071 			/* Move local marker past some other threads marker */
1072 			kn = TAILQ_NEXT(kn, kn_tqe);
1073 			TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1074 			TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe);
1075 			continue;
1076 		}
1077 
1078 		/*
1079 		 * We can't skip a knote undergoing processing, otherwise
1080 		 * we risk not returning it when the user process expects
1081 		 * it should be returned.  Sleep and retry.
1082 		 */
1083 		if (knote_acquire(kn) == 0)
1084 			continue;
1085 
1086 		/*
1087 		 * Remove the event for processing.
1088 		 *
1089 		 * WARNING!  We must leave KN_QUEUED set to prevent the
1090 		 *	     event from being KNOTE_ACTIVATE()d while
1091 		 *	     the queue state is in limbo, in case we
1092 		 *	     block.
1093 		 *
1094 		 * WARNING!  We must set KN_PROCESSING to avoid races
1095 		 *	     against deletion or another thread's
1096 		 *	     processing.
1097 		 */
1098 		TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1099 		kq->kq_count--;
1100 
1101 		/*
1102 		 * We have to deal with an extremely important race against
1103 		 * file descriptor close()s here.  The file descriptor can
1104 		 * disappear MPSAFE, and there is a small window of
1105 		 * opportunity between that and the call to knote_fdclose().
1106 		 *
1107 		 * If we hit that window here while doselect or dopoll is
1108 		 * trying to delete a spurious event they will not be able
1109 		 * to match up the event against a knote and will go haywire.
1110 		 */
1111 		if ((kn->kn_fop->f_flags & FILTEROP_ISFD) &&
1112 		    checkfdclosed(kq->kq_fdp, kn->kn_kevent.ident, kn->kn_fp)) {
1113 			kn->kn_status |= KN_DELETING | KN_REPROCESS;
1114 		}
1115 
1116 		if (kn->kn_status & KN_DISABLED) {
1117 			/*
1118 			 * If disabled we ensure the event is not queued
1119 			 * but leave its active bit set.  On re-enablement
1120 			 * the event may be immediately triggered.
1121 			 */
1122 			kn->kn_status &= ~KN_QUEUED;
1123 		} else if ((kn->kn_flags & EV_ONESHOT) == 0 &&
1124 			   (kn->kn_status & KN_DELETING) == 0 &&
1125 			   filter_event(kn, 0) == 0) {
1126 			/*
1127 			 * If not running in one-shot mode and the event
1128 			 * is no longer present we ensure it is removed
1129 			 * from the queue and ignore it.
1130 			 */
1131 			kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1132 		} else {
1133 			/*
1134 			 * Post the event
1135 			 */
1136 			*kevp++ = kn->kn_kevent;
1137 			++total;
1138 			--count;
1139 
1140 			if (kn->kn_flags & EV_ONESHOT) {
1141 				kn->kn_status &= ~KN_QUEUED;
1142 				kn->kn_status |= KN_DELETING | KN_REPROCESS;
1143 			} else if (kn->kn_flags & EV_CLEAR) {
1144 				kn->kn_data = 0;
1145 				kn->kn_fflags = 0;
1146 				kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1147 			} else {
1148 				TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1149 				kq->kq_count++;
1150 			}
1151 		}
1152 
1153 		/*
1154 		 * Handle any post-processing states
1155 		 */
1156 		knote_release(kn);
1157 	}
1158 	TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1159 
1160 	return (total);
1161 }
1162 
1163 /*
1164  * XXX
1165  * This could be expanded to call kqueue_scan, if desired.
1166  *
1167  * MPSAFE
1168  */
1169 static int
1170 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1171 {
1172 	return (ENXIO);
1173 }
1174 
1175 /*
1176  * MPSAFE
1177  */
1178 static int
1179 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1180 {
1181 	return (ENXIO);
1182 }
1183 
1184 /*
1185  * MPALMOSTSAFE
1186  */
1187 static int
1188 kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
1189 	     struct ucred *cred, struct sysmsg *msg)
1190 {
1191 	struct lwkt_token *tok;
1192 	struct kqueue *kq;
1193 	int error;
1194 
1195 	kq = (struct kqueue *)fp->f_data;
1196 	tok = lwkt_token_pool_lookup(kq);
1197 	lwkt_gettoken(tok);
1198 
1199 	switch(com) {
1200 	case FIOASYNC:
1201 		if (*(int *)data)
1202 			kq->kq_state |= KQ_ASYNC;
1203 		else
1204 			kq->kq_state &= ~KQ_ASYNC;
1205 		error = 0;
1206 		break;
1207 	case FIOSETOWN:
1208 		error = fsetown(*(int *)data, &kq->kq_sigio);
1209 		break;
1210 	default:
1211 		error = ENOTTY;
1212 		break;
1213 	}
1214 	lwkt_reltoken(tok);
1215 	return (error);
1216 }
1217 
1218 /*
1219  * MPSAFE
1220  */
1221 static int
1222 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred)
1223 {
1224 	struct kqueue *kq = (struct kqueue *)fp->f_data;
1225 
1226 	bzero((void *)st, sizeof(*st));
1227 	st->st_size = kq->kq_count;
1228 	st->st_blksize = sizeof(struct kevent);
1229 	st->st_mode = S_IFIFO;
1230 	return (0);
1231 }
1232 
1233 /*
1234  * MPSAFE
1235  */
1236 static int
1237 kqueue_close(struct file *fp)
1238 {
1239 	struct kqueue *kq = (struct kqueue *)fp->f_data;
1240 
1241 	kqueue_terminate(kq);
1242 
1243 	fp->f_data = NULL;
1244 	funsetown(&kq->kq_sigio);
1245 
1246 	kfree(kq, M_KQUEUE);
1247 	return (0);
1248 }
1249 
1250 static void
1251 kqueue_wakeup(struct kqueue *kq)
1252 {
1253 	if (kq->kq_state & KQ_SLEEP) {
1254 		kq->kq_state &= ~KQ_SLEEP;
1255 		wakeup(kq);
1256 	}
1257 	KNOTE(&kq->kq_kqinfo.ki_note, 0);
1258 }
1259 
1260 /*
1261  * Calls filterops f_attach function, acquiring mplock if filter is not
1262  * marked as FILTEROP_MPSAFE.
1263  *
1264  * Caller must be holding the related kq token
1265  */
1266 static int
1267 filter_attach(struct knote *kn)
1268 {
1269 	int ret;
1270 
1271 	if (!(kn->kn_fop->f_flags & FILTEROP_MPSAFE)) {
1272 		get_mplock();
1273 		ret = kn->kn_fop->f_attach(kn);
1274 		rel_mplock();
1275 	} else {
1276 		ret = kn->kn_fop->f_attach(kn);
1277 	}
1278 
1279 	return (ret);
1280 }
1281 
1282 /*
1283  * Detach the knote and drop it, destroying the knote.
1284  *
1285  * Calls filterops f_detach function, acquiring mplock if filter is not
1286  * marked as FILTEROP_MPSAFE.
1287  *
1288  * Caller must be holding the related kq token
1289  */
1290 static void
1291 knote_detach_and_drop(struct knote *kn)
1292 {
1293 	kn->kn_status |= KN_DELETING | KN_REPROCESS;
1294 	if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1295 		kn->kn_fop->f_detach(kn);
1296 	} else {
1297 		get_mplock();
1298 		kn->kn_fop->f_detach(kn);
1299 		rel_mplock();
1300 	}
1301 	knote_drop(kn);
1302 }
1303 
1304 /*
1305  * Calls filterops f_event function, acquiring mplock if filter is not
1306  * marked as FILTEROP_MPSAFE.
1307  *
1308  * If the knote is in the middle of being created or deleted we cannot
1309  * safely call the filter op.
1310  *
1311  * Caller must be holding the related kq token
1312  */
1313 static int
1314 filter_event(struct knote *kn, long hint)
1315 {
1316 	int ret;
1317 
1318 	if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1319 		ret = kn->kn_fop->f_event(kn, hint);
1320 	} else {
1321 		get_mplock();
1322 		ret = kn->kn_fop->f_event(kn, hint);
1323 		rel_mplock();
1324 	}
1325 	return (ret);
1326 }
1327 
1328 /*
1329  * Walk down a list of knotes, activating them if their event has triggered.
1330  *
1331  * If we encounter any knotes which are undergoing processing we just mark
1332  * them for reprocessing and do not try to [re]activate the knote.  However,
1333  * if a hint is being passed we have to wait and that makes things a bit
1334  * sticky.
1335  */
1336 void
1337 knote(struct klist *list, long hint)
1338 {
1339 	struct kqueue *kq;
1340 	struct knote *kn;
1341 	struct knote *kntmp;
1342 
1343 	lwkt_getpooltoken(list);
1344 restart:
1345 	SLIST_FOREACH(kn, list, kn_next) {
1346 		kq = kn->kn_kq;
1347 		lwkt_getpooltoken(kq);
1348 
1349 		/* temporary verification hack */
1350 		SLIST_FOREACH(kntmp, list, kn_next) {
1351 			if (kn == kntmp)
1352 				break;
1353 		}
1354 		if (kn != kntmp || kn->kn_kq != kq) {
1355 			lwkt_relpooltoken(kq);
1356 			goto restart;
1357 		}
1358 
1359 		if (kn->kn_status & KN_PROCESSING) {
1360 			/*
1361 			 * Someone else is processing the knote, ask the
1362 			 * other thread to reprocess it and don't mess
1363 			 * with it otherwise.
1364 			 */
1365 			if (hint == 0) {
1366 				kn->kn_status |= KN_REPROCESS;
1367 				lwkt_relpooltoken(kq);
1368 				continue;
1369 			}
1370 
1371 			/*
1372 			 * If the hint is non-zero we have to wait or risk
1373 			 * losing the state the caller is trying to update.
1374 			 *
1375 			 * XXX This is a real problem, certain process
1376 			 *     and signal filters will bump kn_data for
1377 			 *     already-processed notes more than once if
1378 			 *     we restart the list scan.  FIXME.
1379 			 */
1380 			kn->kn_status |= KN_WAITING | KN_REPROCESS;
1381 			tsleep(kn, 0, "knotec", hz);
1382 			lwkt_relpooltoken(kq);
1383 			goto restart;
1384 		}
1385 
1386 		/*
1387 		 * Become the reprocessing master ourselves.
1388 		 *
1389 		 * If hint is non-zer running the event is mandatory
1390 		 * when not deleting so do it whether reprocessing is
1391 		 * set or not.
1392 		 */
1393 		kn->kn_status |= KN_PROCESSING;
1394 		if ((kn->kn_status & KN_DELETING) == 0) {
1395 			if (filter_event(kn, hint))
1396 				KNOTE_ACTIVATE(kn);
1397 		}
1398 		if (knote_release(kn)) {
1399 			lwkt_relpooltoken(kq);
1400 			goto restart;
1401 		}
1402 		lwkt_relpooltoken(kq);
1403 	}
1404 	lwkt_relpooltoken(list);
1405 }
1406 
1407 /*
1408  * Insert knote at head of klist.
1409  *
1410  * This function may only be called via a filter function and thus
1411  * kq_token should already be held and marked for processing.
1412  */
1413 void
1414 knote_insert(struct klist *klist, struct knote *kn)
1415 {
1416 	lwkt_getpooltoken(klist);
1417 	KKASSERT(kn->kn_status & KN_PROCESSING);
1418 	SLIST_INSERT_HEAD(klist, kn, kn_next);
1419 	lwkt_relpooltoken(klist);
1420 }
1421 
1422 /*
1423  * Remove knote from a klist
1424  *
1425  * This function may only be called via a filter function and thus
1426  * kq_token should already be held and marked for processing.
1427  */
1428 void
1429 knote_remove(struct klist *klist, struct knote *kn)
1430 {
1431 	lwkt_getpooltoken(klist);
1432 	KKASSERT(kn->kn_status & KN_PROCESSING);
1433 	SLIST_REMOVE(klist, kn, knote, kn_next);
1434 	lwkt_relpooltoken(klist);
1435 }
1436 
1437 #if 0
1438 /*
1439  * Remove all knotes from a specified klist
1440  *
1441  * Only called from aio.
1442  */
1443 void
1444 knote_empty(struct klist *list)
1445 {
1446 	struct knote *kn;
1447 
1448 	lwkt_gettoken(&kq_token);
1449 	while ((kn = SLIST_FIRST(list)) != NULL) {
1450 		if (knote_acquire(kn))
1451 			knote_detach_and_drop(kn);
1452 	}
1453 	lwkt_reltoken(&kq_token);
1454 }
1455 #endif
1456 
1457 void
1458 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst,
1459 		    struct filterops *ops, void *hook)
1460 {
1461 	struct kqueue *kq;
1462 	struct knote *kn;
1463 
1464 	lwkt_getpooltoken(&src->ki_note);
1465 	lwkt_getpooltoken(&dst->ki_note);
1466 	while ((kn = SLIST_FIRST(&src->ki_note)) != NULL) {
1467 		kq = kn->kn_kq;
1468 		lwkt_getpooltoken(kq);
1469 		if (SLIST_FIRST(&src->ki_note) != kn || kn->kn_kq != kq) {
1470 			lwkt_relpooltoken(kq);
1471 			continue;
1472 		}
1473 		if (knote_acquire(kn)) {
1474 			knote_remove(&src->ki_note, kn);
1475 			kn->kn_fop = ops;
1476 			kn->kn_hook = hook;
1477 			knote_insert(&dst->ki_note, kn);
1478 			knote_release(kn);
1479 			/* kn may be invalid now */
1480 		}
1481 		lwkt_relpooltoken(kq);
1482 	}
1483 	lwkt_relpooltoken(&dst->ki_note);
1484 	lwkt_relpooltoken(&src->ki_note);
1485 }
1486 
1487 /*
1488  * Remove all knotes referencing a specified fd
1489  */
1490 void
1491 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd)
1492 {
1493 	struct kqueue *kq;
1494 	struct knote *kn;
1495 	struct knote *kntmp;
1496 
1497 	lwkt_getpooltoken(&fp->f_klist);
1498 restart:
1499 	SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
1500 		if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) {
1501 			kq = kn->kn_kq;
1502 			lwkt_getpooltoken(kq);
1503 
1504 			/* temporary verification hack */
1505 			SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) {
1506 				if (kn == kntmp)
1507 					break;
1508 			}
1509 			if (kn != kntmp || kn->kn_kq->kq_fdp != fdp ||
1510 			    kn->kn_id != fd || kn->kn_kq != kq) {
1511 				lwkt_relpooltoken(kq);
1512 				goto restart;
1513 			}
1514 			if (knote_acquire(kn))
1515 				knote_detach_and_drop(kn);
1516 			lwkt_relpooltoken(kq);
1517 			goto restart;
1518 		}
1519 	}
1520 	lwkt_relpooltoken(&fp->f_klist);
1521 }
1522 
1523 /*
1524  * Low level attach function.
1525  *
1526  * The knote should already be marked for processing.
1527  * Caller must hold the related kq token.
1528  */
1529 static void
1530 knote_attach(struct knote *kn)
1531 {
1532 	struct klist *list;
1533 	struct kqueue *kq = kn->kn_kq;
1534 
1535 	if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1536 		KKASSERT(kn->kn_fp);
1537 		list = &kn->kn_fp->f_klist;
1538 	} else {
1539 		if (kq->kq_knhashmask == 0)
1540 			kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1541 						 &kq->kq_knhashmask);
1542 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1543 	}
1544 	lwkt_getpooltoken(list);
1545 	SLIST_INSERT_HEAD(list, kn, kn_link);
1546 	TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink);
1547 	lwkt_relpooltoken(list);
1548 }
1549 
1550 /*
1551  * Low level drop function.
1552  *
1553  * The knote should already be marked for processing.
1554  * Caller must hold the related kq token.
1555  */
1556 static void
1557 knote_drop(struct knote *kn)
1558 {
1559 	struct kqueue *kq;
1560 	struct klist *list;
1561 
1562 	kq = kn->kn_kq;
1563 
1564 	if (kn->kn_fop->f_flags & FILTEROP_ISFD)
1565 		list = &kn->kn_fp->f_klist;
1566 	else
1567 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1568 
1569 	lwkt_getpooltoken(list);
1570 	SLIST_REMOVE(list, kn, knote, kn_link);
1571 	TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink);
1572 	if (kn->kn_status & KN_QUEUED)
1573 		knote_dequeue(kn);
1574 	if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1575 		fdrop(kn->kn_fp);
1576 		kn->kn_fp = NULL;
1577 	}
1578 	knote_free(kn);
1579 	lwkt_relpooltoken(list);
1580 }
1581 
1582 /*
1583  * Low level enqueue function.
1584  *
1585  * The knote should already be marked for processing.
1586  * Caller must be holding the kq token
1587  */
1588 static void
1589 knote_enqueue(struct knote *kn)
1590 {
1591 	struct kqueue *kq = kn->kn_kq;
1592 
1593 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
1594 	TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1595 	kn->kn_status |= KN_QUEUED;
1596 	++kq->kq_count;
1597 
1598 	/*
1599 	 * Send SIGIO on request (typically set up as a mailbox signal)
1600 	 */
1601 	if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1)
1602 		pgsigio(kq->kq_sigio, SIGIO, 0);
1603 
1604 	kqueue_wakeup(kq);
1605 }
1606 
1607 /*
1608  * Low level dequeue function.
1609  *
1610  * The knote should already be marked for processing.
1611  * Caller must be holding the kq token
1612  */
1613 static void
1614 knote_dequeue(struct knote *kn)
1615 {
1616 	struct kqueue *kq = kn->kn_kq;
1617 
1618 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
1619 	TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1620 	kn->kn_status &= ~KN_QUEUED;
1621 	kq->kq_count--;
1622 }
1623 
1624 static struct knote *
1625 knote_alloc(void)
1626 {
1627 	return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK);
1628 }
1629 
1630 static void
1631 knote_free(struct knote *kn)
1632 {
1633 	kfree(kn, M_KQUEUE);
1634 }
1635