xref: /dflybsd-src/sys/kern/kern_event.c (revision d4b8aec4bb44a374c3e91969c1a7df6569da7be3)
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  * $DragonFly: src/sys/kern/kern_event.c,v 1.33 2007/02/03 17:05:57 corecode Exp $
28  */
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
33 #include <sys/proc.h>
34 #include <sys/malloc.h>
35 #include <sys/unistd.h>
36 #include <sys/file.h>
37 #include <sys/lock.h>
38 #include <sys/fcntl.h>
39 #include <sys/queue.h>
40 #include <sys/event.h>
41 #include <sys/eventvar.h>
42 #include <sys/protosw.h>
43 #include <sys/socket.h>
44 #include <sys/socketvar.h>
45 #include <sys/stat.h>
46 #include <sys/sysctl.h>
47 #include <sys/sysproto.h>
48 #include <sys/thread.h>
49 #include <sys/uio.h>
50 #include <sys/signalvar.h>
51 #include <sys/filio.h>
52 #include <sys/ktr.h>
53 
54 #include <sys/thread2.h>
55 #include <sys/file2.h>
56 #include <sys/mplock2.h>
57 
58 #include <vm/vm_zone.h>
59 
60 /*
61  * Global token for kqueue subsystem
62  */
63 struct lwkt_token kq_token = LWKT_TOKEN_UP_INITIALIZER;
64 
65 MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
66 
67 struct kevent_copyin_args {
68 	struct kevent_args	*ka;
69 	int			pchanges;
70 };
71 
72 static int	kqueue_sleep(struct kqueue *kq, struct timespec *tsp);
73 static int	kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
74 		    struct knote *marker);
75 static int 	kqueue_read(struct file *fp, struct uio *uio,
76 		    struct ucred *cred, int flags);
77 static int	kqueue_write(struct file *fp, struct uio *uio,
78 		    struct ucred *cred, int flags);
79 static int	kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
80 		    struct ucred *cred, struct sysmsg *msg);
81 static int 	kqueue_kqfilter(struct file *fp, struct knote *kn);
82 static int 	kqueue_stat(struct file *fp, struct stat *st,
83 		    struct ucred *cred);
84 static int 	kqueue_close(struct file *fp);
85 static void	kqueue_wakeup(struct kqueue *kq);
86 static int	filter_attach(struct knote *kn);
87 static int	filter_event(struct knote *kn, long hint);
88 
89 /*
90  * MPSAFE
91  */
92 static struct fileops kqueueops = {
93 	.fo_read = kqueue_read,
94 	.fo_write = kqueue_write,
95 	.fo_ioctl = kqueue_ioctl,
96 	.fo_kqfilter = kqueue_kqfilter,
97 	.fo_stat = kqueue_stat,
98 	.fo_close = kqueue_close,
99 	.fo_shutdown = nofo_shutdown
100 };
101 
102 static void 	knote_attach(struct knote *kn);
103 static void 	knote_drop(struct knote *kn);
104 static void	knote_detach_and_drop(struct knote *kn);
105 static void 	knote_enqueue(struct knote *kn);
106 static void 	knote_dequeue(struct knote *kn);
107 static void 	knote_init(void);
108 static struct 	knote *knote_alloc(void);
109 static void 	knote_free(struct knote *kn);
110 
111 static void	filt_kqdetach(struct knote *kn);
112 static int	filt_kqueue(struct knote *kn, long hint);
113 static int	filt_procattach(struct knote *kn);
114 static void	filt_procdetach(struct knote *kn);
115 static int	filt_proc(struct knote *kn, long hint);
116 static int	filt_fileattach(struct knote *kn);
117 static void	filt_timerexpire(void *knx);
118 static int	filt_timerattach(struct knote *kn);
119 static void	filt_timerdetach(struct knote *kn);
120 static int	filt_timer(struct knote *kn, long hint);
121 
122 static struct filterops file_filtops =
123 	{ FILTEROP_ISFD, filt_fileattach, NULL, NULL };
124 static struct filterops kqread_filtops =
125 	{ FILTEROP_ISFD, NULL, filt_kqdetach, filt_kqueue };
126 static struct filterops proc_filtops =
127 	{ 0, filt_procattach, filt_procdetach, filt_proc };
128 static struct filterops timer_filtops =
129 	{ 0, filt_timerattach, filt_timerdetach, filt_timer };
130 
131 static vm_zone_t	knote_zone;
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 	lwkt_gettoken(&proc_token);
214 	p = pfind(kn->kn_id);
215 	if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
216 		p = zpfind(kn->kn_id);
217 		immediate = 1;
218 	}
219 	if (p == NULL) {
220 		lwkt_reltoken(&proc_token);
221 		return (ESRCH);
222 	}
223 	if (!PRISON_CHECK(curthread->td_ucred, p->p_ucred)) {
224 		lwkt_reltoken(&proc_token);
225 		return (EACCES);
226 	}
227 
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(&proc_token);
250 
251 	return (0);
252 }
253 
254 /*
255  * The knote may be attached to a different process, which may exit,
256  * leaving nothing for the knote to be attached to.  So when the process
257  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
258  * it will be deleted when read out.  However, as part of the knote deletion,
259  * this routine is called, so a check is needed to avoid actually performing
260  * a detach, because the original process does not exist any more.
261  */
262 static void
263 filt_procdetach(struct knote *kn)
264 {
265 	struct proc *p;
266 
267 	if (kn->kn_status & KN_DETACHED)
268 		return;
269 	/* XXX locking? take proc_token here? */
270 	p = kn->kn_ptr.p_proc;
271 	knote_remove(&p->p_klist, kn);
272 }
273 
274 static int
275 filt_proc(struct knote *kn, long hint)
276 {
277 	u_int event;
278 
279 	/*
280 	 * mask off extra data
281 	 */
282 	event = (u_int)hint & NOTE_PCTRLMASK;
283 
284 	/*
285 	 * if the user is interested in this event, record it.
286 	 */
287 	if (kn->kn_sfflags & event)
288 		kn->kn_fflags |= event;
289 
290 	/*
291 	 * Process is gone, so flag the event as finished.  Detach the
292 	 * knote from the process now because the process will be poof,
293 	 * gone later on.
294 	 */
295 	if (event == NOTE_EXIT) {
296 		struct proc *p = kn->kn_ptr.p_proc;
297 		if ((kn->kn_status & KN_DETACHED) == 0) {
298 			knote_remove(&p->p_klist, kn);
299 			kn->kn_status |= KN_DETACHED;
300 			kn->kn_data = p->p_xstat;
301 			kn->kn_ptr.p_proc = NULL;
302 		}
303 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
304 		return (1);
305 	}
306 
307 	/*
308 	 * process forked, and user wants to track the new process,
309 	 * so attach a new knote to it, and immediately report an
310 	 * event with the parent's pid.
311 	 */
312 	if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) {
313 		struct kevent kev;
314 		int error;
315 
316 		/*
317 		 * register knote with new process.
318 		 */
319 		kev.ident = hint & NOTE_PDATAMASK;	/* pid */
320 		kev.filter = kn->kn_filter;
321 		kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
322 		kev.fflags = kn->kn_sfflags;
323 		kev.data = kn->kn_id;			/* parent */
324 		kev.udata = kn->kn_kevent.udata;	/* preserve udata */
325 		error = kqueue_register(kn->kn_kq, &kev);
326 		if (error)
327 			kn->kn_fflags |= NOTE_TRACKERR;
328 	}
329 
330 	return (kn->kn_fflags != 0);
331 }
332 
333 static void
334 filt_timerexpire(void *knx)
335 {
336 	struct knote *kn = knx;
337 	struct callout *calloutp;
338 	struct timeval tv;
339 	int tticks;
340 
341 	kn->kn_data++;
342 	KNOTE_ACTIVATE(kn);
343 
344 	if ((kn->kn_flags & EV_ONESHOT) == 0) {
345 		tv.tv_sec = kn->kn_sdata / 1000;
346 		tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
347 		tticks = tvtohz_high(&tv);
348 		calloutp = (struct callout *)kn->kn_hook;
349 		callout_reset(calloutp, tticks, filt_timerexpire, kn);
350 	}
351 }
352 
353 /*
354  * data contains amount of time to sleep, in milliseconds
355  */
356 static int
357 filt_timerattach(struct knote *kn)
358 {
359 	struct callout *calloutp;
360 	struct timeval tv;
361 	int tticks;
362 
363 	if (kq_ncallouts >= kq_calloutmax)
364 		return (ENOMEM);
365 	kq_ncallouts++;
366 
367 	tv.tv_sec = kn->kn_sdata / 1000;
368 	tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
369 	tticks = tvtohz_high(&tv);
370 
371 	kn->kn_flags |= EV_CLEAR;		/* automatically set */
372 	MALLOC(calloutp, struct callout *, sizeof(*calloutp),
373 	    M_KQUEUE, M_WAITOK);
374 	callout_init(calloutp);
375 	kn->kn_hook = (caddr_t)calloutp;
376 	callout_reset(calloutp, tticks, filt_timerexpire, kn);
377 
378 	return (0);
379 }
380 
381 static void
382 filt_timerdetach(struct knote *kn)
383 {
384 	struct callout *calloutp;
385 
386 	calloutp = (struct callout *)kn->kn_hook;
387 	callout_stop(calloutp);
388 	FREE(calloutp, M_KQUEUE);
389 	kq_ncallouts--;
390 }
391 
392 static int
393 filt_timer(struct knote *kn, long hint)
394 {
395 
396 	return (kn->kn_data != 0);
397 }
398 
399 /*
400  * Initialize a kqueue.
401  *
402  * NOTE: The lwp/proc code initializes a kqueue for select/poll ops.
403  *
404  * MPSAFE
405  */
406 void
407 kqueue_init(struct kqueue *kq, struct filedesc *fdp)
408 {
409 	TAILQ_INIT(&kq->kq_knpend);
410 	TAILQ_INIT(&kq->kq_knlist);
411 	kq->kq_count = 0;
412 	kq->kq_fdp = fdp;
413 	SLIST_INIT(&kq->kq_kqinfo.ki_note);
414 }
415 
416 /*
417  * Terminate a kqueue.  Freeing the actual kq itself is left up to the
418  * caller (it might be embedded in a lwp so we don't do it here).
419  */
420 void
421 kqueue_terminate(struct kqueue *kq)
422 {
423 	struct knote *kn;
424 
425 	lwkt_gettoken(&kq_token);
426 	while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL)
427 		knote_detach_and_drop(kn);
428 
429 	if (kq->kq_knhash) {
430 		kfree(kq->kq_knhash, M_KQUEUE);
431 		kq->kq_knhash = NULL;
432 		kq->kq_knhashmask = 0;
433 	}
434 	lwkt_reltoken(&kq_token);
435 }
436 
437 /*
438  * MPSAFE
439  */
440 int
441 sys_kqueue(struct kqueue_args *uap)
442 {
443 	struct thread *td = curthread;
444 	struct kqueue *kq;
445 	struct file *fp;
446 	int fd, error;
447 
448 	error = falloc(td->td_lwp, &fp, &fd);
449 	if (error)
450 		return (error);
451 	fp->f_flag = FREAD | FWRITE;
452 	fp->f_type = DTYPE_KQUEUE;
453 	fp->f_ops = &kqueueops;
454 
455 	kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO);
456 	kqueue_init(kq, td->td_proc->p_fd);
457 	fp->f_data = kq;
458 
459 	fsetfd(kq->kq_fdp, fp, fd);
460 	uap->sysmsg_result = fd;
461 	fdrop(fp);
462 	return (error);
463 }
464 
465 /*
466  * Copy 'count' items into the destination list pointed to by uap->eventlist.
467  */
468 static int
469 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res)
470 {
471 	struct kevent_copyin_args *kap;
472 	int error;
473 
474 	kap = (struct kevent_copyin_args *)arg;
475 
476 	error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp));
477 	if (error == 0) {
478 		kap->ka->eventlist += count;
479 		*res += count;
480 	} else {
481 		*res = -1;
482 	}
483 
484 	return (error);
485 }
486 
487 /*
488  * Copy at most 'max' items from the list pointed to by kap->changelist,
489  * return number of items in 'events'.
490  */
491 static int
492 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events)
493 {
494 	struct kevent_copyin_args *kap;
495 	int error, count;
496 
497 	kap = (struct kevent_copyin_args *)arg;
498 
499 	count = min(kap->ka->nchanges - kap->pchanges, max);
500 	error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp);
501 	if (error == 0) {
502 		kap->ka->changelist += count;
503 		kap->pchanges += count;
504 		*events = count;
505 	}
506 
507 	return (error);
508 }
509 
510 /*
511  * MPSAFE
512  */
513 int
514 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap,
515 	    k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn,
516 	    struct timespec *tsp_in)
517 {
518 	struct kevent *kevp;
519 	struct timespec *tsp;
520 	int i, n, total, error, nerrors = 0;
521 	int lres;
522 	int limit = kq_checkloop;
523 	struct kevent kev[KQ_NEVENTS];
524 	struct knote marker;
525 
526 	tsp = tsp_in;
527 	*res = 0;
528 
529 	lwkt_gettoken(&kq_token);
530 	for ( ;; ) {
531 		n = 0;
532 		error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n);
533 		if (error)
534 			goto done;
535 		if (n == 0)
536 			break;
537 		for (i = 0; i < n; i++) {
538 			kevp = &kev[i];
539 			kevp->flags &= ~EV_SYSFLAGS;
540 			error = kqueue_register(kq, kevp);
541 
542 			/*
543 			 * If a registration returns an error we
544 			 * immediately post the error.  The kevent()
545 			 * call itself will fail with the error if
546 			 * no space is available for posting.
547 			 *
548 			 * Such errors normally bypass the timeout/blocking
549 			 * code.  However, if the copyoutfn function refuses
550 			 * to post the error (see sys_poll()), then we
551 			 * ignore it too.
552 			 */
553 			if (error) {
554 				kevp->flags = EV_ERROR;
555 				kevp->data = error;
556 				lres = *res;
557 				kevent_copyoutfn(uap, kevp, 1, res);
558 				if (lres != *res) {
559 					nevents--;
560 					nerrors++;
561 				}
562 			}
563 		}
564 	}
565 	if (nerrors) {
566 		error = 0;
567 		goto done;
568 	}
569 
570 	/*
571 	 * Acquire/wait for events - setup timeout
572 	 */
573 	if (tsp != NULL) {
574 		struct timespec ats;
575 
576 		if (tsp->tv_sec || tsp->tv_nsec) {
577 			nanouptime(&ats);
578 			timespecadd(tsp, &ats);		/* tsp = target time */
579 		}
580 	}
581 
582 	/*
583 	 * Loop as required.
584 	 *
585 	 * Collect as many events as we can. Sleeping on successive
586 	 * loops is disabled if copyoutfn has incremented (*res).
587 	 *
588 	 * The loop stops if an error occurs, all events have been
589 	 * scanned (the marker has been reached), or fewer than the
590 	 * maximum number of events is found.
591 	 *
592 	 * The copyoutfn function does not have to increment (*res) in
593 	 * order for the loop to continue.
594 	 *
595 	 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents.
596 	 */
597 	total = 0;
598 	error = 0;
599 	marker.kn_filter = EVFILT_MARKER;
600 	TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
601 	while ((n = nevents - total) > 0) {
602 		if (n > KQ_NEVENTS)
603 			n = KQ_NEVENTS;
604 
605 		/*
606 		 * If no events are pending sleep until timeout (if any)
607 		 * or an event occurs.
608 		 *
609 		 * After the sleep completes the marker is moved to the
610 		 * end of the list, making any received events available
611 		 * to our scan.
612 		 */
613 		if (kq->kq_count == 0 && *res == 0) {
614 			error = kqueue_sleep(kq, tsp);
615 			if (error)
616 				break;
617 
618 			TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
619 			TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
620 		}
621 
622 		/*
623 		 * Process all received events
624 		 * Account for all non-spurious events in our total
625 		 */
626 		i = kqueue_scan(kq, kev, n, &marker);
627 		if (i) {
628 			lres = *res;
629 			error = kevent_copyoutfn(uap, kev, i, res);
630 			total += *res - lres;
631 			if (error)
632 				break;
633 		}
634 		if (limit && --limit == 0)
635 			panic("kqueue: checkloop failed i=%d", i);
636 
637 		/*
638 		 * Normally when fewer events are returned than requested
639 		 * we can stop.  However, if only spurious events were
640 		 * collected the copyout will not bump (*res) and we have
641 		 * to continue.
642 		 */
643 		if (i < n && *res)
644 			break;
645 
646 		/*
647 		 * Deal with an edge case where spurious events can cause
648 		 * a loop to occur without moving the marker.  This can
649 		 * prevent kqueue_scan() from picking up new events which
650 		 * race us.  We must be sure to move the marker for this
651 		 * case.
652 		 *
653 		 * NOTE: We do not want to move the marker if events
654 		 *	 were scanned because normal kqueue operations
655 		 *	 may reactivate events.  Moving the marker in
656 		 *	 that case could result in duplicates for the
657 		 *	 same event.
658 		 */
659 		if (i == 0) {
660 			TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
661 			TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
662 		}
663 	}
664 	TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
665 
666 	/* Timeouts do not return EWOULDBLOCK. */
667 	if (error == EWOULDBLOCK)
668 		error = 0;
669 
670 done:
671 	lwkt_reltoken(&kq_token);
672 	return (error);
673 }
674 
675 /*
676  * MPALMOSTSAFE
677  */
678 int
679 sys_kevent(struct kevent_args *uap)
680 {
681 	struct thread *td = curthread;
682 	struct proc *p = td->td_proc;
683 	struct timespec ts, *tsp;
684 	struct kqueue *kq;
685 	struct file *fp = NULL;
686 	struct kevent_copyin_args *kap, ka;
687 	int error;
688 
689 	if (uap->timeout) {
690 		error = copyin(uap->timeout, &ts, sizeof(ts));
691 		if (error)
692 			return (error);
693 		tsp = &ts;
694 	} else {
695 		tsp = NULL;
696 	}
697 
698 	fp = holdfp(p->p_fd, uap->fd, -1);
699 	if (fp == NULL)
700 		return (EBADF);
701 	if (fp->f_type != DTYPE_KQUEUE) {
702 		fdrop(fp);
703 		return (EBADF);
704 	}
705 
706 	kq = (struct kqueue *)fp->f_data;
707 
708 	kap = &ka;
709 	kap->ka = uap;
710 	kap->pchanges = 0;
711 
712 	error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap,
713 			    kevent_copyin, kevent_copyout, tsp);
714 
715 	fdrop(fp);
716 
717 	return (error);
718 }
719 
720 int
721 kqueue_register(struct kqueue *kq, struct kevent *kev)
722 {
723 	struct filedesc *fdp = kq->kq_fdp;
724 	struct filterops *fops;
725 	struct file *fp = NULL;
726 	struct knote *kn = NULL;
727 	int error = 0;
728 
729 	if (kev->filter < 0) {
730 		if (kev->filter + EVFILT_SYSCOUNT < 0)
731 			return (EINVAL);
732 		fops = sysfilt_ops[~kev->filter];	/* to 0-base index */
733 	} else {
734 		/*
735 		 * XXX
736 		 * filter attach routine is responsible for insuring that
737 		 * the identifier can be attached to it.
738 		 */
739 		kprintf("unknown filter: %d\n", kev->filter);
740 		return (EINVAL);
741 	}
742 
743 	lwkt_gettoken(&kq_token);
744 	if (fops->f_flags & FILTEROP_ISFD) {
745 		/* validate descriptor */
746 		fp = holdfp(fdp, kev->ident, -1);
747 		if (fp == NULL) {
748 			lwkt_reltoken(&kq_token);
749 			return (EBADF);
750 		}
751 
752 		SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
753 			if (kn->kn_kq == kq &&
754 			    kn->kn_filter == kev->filter &&
755 			    kn->kn_id == kev->ident) {
756 				break;
757 			}
758 		}
759 	} else {
760 		if (kq->kq_knhashmask) {
761 			struct klist *list;
762 
763 			list = &kq->kq_knhash[
764 			    KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
765 			SLIST_FOREACH(kn, list, kn_link) {
766 				if (kn->kn_id == kev->ident &&
767 				    kn->kn_filter == kev->filter)
768 					break;
769 			}
770 		}
771 	}
772 
773 	if (kn == NULL && ((kev->flags & EV_ADD) == 0)) {
774 		error = ENOENT;
775 		goto done;
776 	}
777 
778 	/*
779 	 * kn now contains the matching knote, or NULL if no match
780 	 */
781 	if (kev->flags & EV_ADD) {
782 		if (kn == NULL) {
783 			kn = knote_alloc();
784 			if (kn == NULL) {
785 				error = ENOMEM;
786 				goto done;
787 			}
788 			kn->kn_fp = fp;
789 			kn->kn_kq = kq;
790 			kn->kn_fop = fops;
791 
792 			/*
793 			 * apply reference count to knote structure, and
794 			 * do not release it at the end of this routine.
795 			 */
796 			fp = NULL;
797 
798 			kn->kn_sfflags = kev->fflags;
799 			kn->kn_sdata = kev->data;
800 			kev->fflags = 0;
801 			kev->data = 0;
802 			kn->kn_kevent = *kev;
803 
804 			/*
805 			 * Interlock against creation/deletion races due
806 			 * to f_attach() blocking.  knote_attach() will set
807 			 * KN_CREATING.
808 			 */
809 			knote_attach(kn);
810 			if ((error = filter_attach(kn)) != 0) {
811 				kn->kn_status |= KN_DELETING;
812 				knote_drop(kn);
813 				goto done;
814 			}
815 			kn->kn_status &= ~KN_CREATING;
816 
817 			/*
818 			 * Interlock against close races which remove our
819 			 * knotes.  We do not want to end up with a knote
820 			 * on a closed descriptor.
821 			 */
822 			if ((fops->f_flags & FILTEROP_ISFD) &&
823 			    (error = checkfdclosed(fdp, kev->ident, kn->kn_fp)) != 0) {
824 				knote_detach_and_drop(kn);
825 				goto done;
826 			}
827 		} else {
828 			/*
829 			 * The user may change some filter values after the
830 			 * initial EV_ADD, but doing so will not reset any
831 			 * filter which have already been triggered.
832 			 */
833 			kn->kn_sfflags = kev->fflags;
834 			kn->kn_sdata = kev->data;
835 			kn->kn_kevent.udata = kev->udata;
836 		}
837 
838 		if (filter_event(kn, 0))
839 			KNOTE_ACTIVATE(kn);
840 	} else if (kev->flags & EV_DELETE) {
841 		knote_detach_and_drop(kn);
842 		goto done;
843 	}
844 
845 	if ((kev->flags & EV_DISABLE) &&
846 	    ((kn->kn_status & KN_DISABLED) == 0)) {
847 		kn->kn_status |= KN_DISABLED;
848 	}
849 
850 	if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
851 		kn->kn_status &= ~KN_DISABLED;
852 		if ((kn->kn_status & KN_ACTIVE) &&
853 		    ((kn->kn_status & KN_QUEUED) == 0))
854 			knote_enqueue(kn);
855 	}
856 
857 done:
858 	lwkt_reltoken(&kq_token);
859 	if (fp != NULL)
860 		fdrop(fp);
861 	return (error);
862 }
863 
864 /*
865  * Block as necessary until the target time is reached.
866  * If tsp is NULL we block indefinitely.  If tsp->ts_secs/nsecs are both
867  * 0 we do not block at all.
868  */
869 static int
870 kqueue_sleep(struct kqueue *kq, struct timespec *tsp)
871 {
872 	int error = 0;
873 
874 	if (tsp == NULL) {
875 		kq->kq_state |= KQ_SLEEP;
876 		error = tsleep(kq, PCATCH, "kqread", 0);
877 	} else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) {
878 		error = EWOULDBLOCK;
879 	} else {
880 		struct timespec ats;
881 		struct timespec atx = *tsp;
882 		int timeout;
883 
884 		nanouptime(&ats);
885 		timespecsub(&atx, &ats);
886 		if (ats.tv_sec < 0) {
887 			error = EWOULDBLOCK;
888 		} else {
889 			timeout = atx.tv_sec > 24 * 60 * 60 ?
890 				24 * 60 * 60 * hz : tstohz_high(&atx);
891 			kq->kq_state |= KQ_SLEEP;
892 			error = tsleep(kq, PCATCH, "kqread", timeout);
893 		}
894 	}
895 
896 	/* don't restart after signals... */
897 	if (error == ERESTART)
898 		return (EINTR);
899 
900 	return (error);
901 }
902 
903 /*
904  * Scan the kqueue, return the number of active events placed in kevp up
905  * to count.
906  *
907  * Continuous mode events may get recycled, do not continue scanning past
908  * marker unless no events have been collected.
909  */
910 static int
911 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
912             struct knote *marker)
913 {
914         struct knote *kn, local_marker;
915         int total;
916 
917         total = 0;
918 	local_marker.kn_filter = EVFILT_MARKER;
919 
920 	/*
921 	 * Collect events.
922 	 */
923 	TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe);
924 	while (count) {
925 		kn = TAILQ_NEXT(&local_marker, kn_tqe);
926 		if (kn->kn_filter == EVFILT_MARKER) {
927 			/* Marker reached, we are done */
928 			if (kn == marker)
929 				break;
930 
931 			/* Move local marker past some other threads marker */
932 			kn = TAILQ_NEXT(kn, kn_tqe);
933 			TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
934 			TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe);
935 			continue;
936 		}
937 
938 		/*
939 		 * Remove the event for processing.
940 		 *
941 		 * WARNING!  We must leave KN_QUEUED set to prevent the
942 		 *	     event from being KNOTE()d again while we
943 		 *	     potentially block in the filter function.
944 		 *
945 		 *	     This protects the knote from everything except
946 		 *	     getting dropped.
947 		 *
948 		 * WARNING!  KN_PROCESSING is meant to handle any cases
949 		 *	     that leaving KN_QUEUED set does not.
950 		 */
951 		TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
952 		kq->kq_count--;
953 		kn->kn_status |= KN_PROCESSING;
954 
955 		/*
956 		 * Even though close/dup2 will clean out pending knotes this
957 		 * code is MPSAFE and it is possible to race a close inbetween
958 		 * the removal of its descriptor and the clearing out of the
959 		 * knote(s).
960 		 *
961 		 * In this case we must ensure that the knote is not queued
962 		 * to knpend or we risk an infinite kernel loop calling
963 		 * kscan, because the select/poll code will not be able to
964 		 * delete the event.
965 		 */
966 		if ((kn->kn_fop->f_flags & FILTEROP_ISFD) &&
967 		    checkfdclosed(kq->kq_fdp, kn->kn_kevent.ident, kn->kn_fp)) {
968 			kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE |
969 					   KN_PROCESSING);
970 			continue;
971 		}
972 
973 		/*
974 		 * If disabled we ensure the event is not queued but leave
975 		 * its active bit set.  On re-enablement the event may be
976 		 * immediately triggered.
977 		 */
978 		if (kn->kn_status & KN_DISABLED) {
979 			kn->kn_status &= ~(KN_QUEUED | KN_PROCESSING);
980 			continue;
981 		}
982 
983 		/*
984 		 * If not running in one-shot mode and the event is no
985 		 * longer present we ensure it is removed from the queue and
986 		 * ignore it.
987 		 */
988 		if ((kn->kn_flags & EV_ONESHOT) == 0 &&
989 		    filter_event(kn, 0) == 0) {
990 			kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE |
991 					   KN_PROCESSING);
992 			continue;
993 		}
994 
995 		*kevp++ = kn->kn_kevent;
996 		++total;
997 		--count;
998 
999 		/*
1000 		 * Post-event action on the note
1001 		 */
1002 		if (kn->kn_flags & EV_ONESHOT) {
1003 			kn->kn_status &= ~(KN_QUEUED | KN_PROCESSING);
1004 			knote_detach_and_drop(kn);
1005 		} else if (kn->kn_flags & EV_CLEAR) {
1006 			kn->kn_data = 0;
1007 			kn->kn_fflags = 0;
1008 			kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE |
1009 					   KN_PROCESSING);
1010 		} else {
1011 			TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1012 			kq->kq_count++;
1013 			kn->kn_status &= ~KN_PROCESSING;
1014 		}
1015 	}
1016 	TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1017 
1018 	return (total);
1019 }
1020 
1021 /*
1022  * XXX
1023  * This could be expanded to call kqueue_scan, if desired.
1024  *
1025  * MPSAFE
1026  */
1027 static int
1028 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1029 {
1030 	return (ENXIO);
1031 }
1032 
1033 /*
1034  * MPSAFE
1035  */
1036 static int
1037 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1038 {
1039 	return (ENXIO);
1040 }
1041 
1042 /*
1043  * MPALMOSTSAFE
1044  */
1045 static int
1046 kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
1047 	     struct ucred *cred, struct sysmsg *msg)
1048 {
1049 	struct kqueue *kq;
1050 	int error;
1051 
1052 	lwkt_gettoken(&kq_token);
1053 	kq = (struct kqueue *)fp->f_data;
1054 
1055 	switch(com) {
1056 	case FIOASYNC:
1057 		if (*(int *)data)
1058 			kq->kq_state |= KQ_ASYNC;
1059 		else
1060 			kq->kq_state &= ~KQ_ASYNC;
1061 		error = 0;
1062 		break;
1063 	case FIOSETOWN:
1064 		error = fsetown(*(int *)data, &kq->kq_sigio);
1065 		break;
1066 	default:
1067 		error = ENOTTY;
1068 		break;
1069 	}
1070 	lwkt_reltoken(&kq_token);
1071 	return (error);
1072 }
1073 
1074 /*
1075  * MPSAFE
1076  */
1077 static int
1078 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred)
1079 {
1080 	struct kqueue *kq = (struct kqueue *)fp->f_data;
1081 
1082 	bzero((void *)st, sizeof(*st));
1083 	st->st_size = kq->kq_count;
1084 	st->st_blksize = sizeof(struct kevent);
1085 	st->st_mode = S_IFIFO;
1086 	return (0);
1087 }
1088 
1089 /*
1090  * MPSAFE
1091  */
1092 static int
1093 kqueue_close(struct file *fp)
1094 {
1095 	struct kqueue *kq = (struct kqueue *)fp->f_data;
1096 
1097 	kqueue_terminate(kq);
1098 
1099 	fp->f_data = NULL;
1100 	funsetown(kq->kq_sigio);
1101 
1102 	kfree(kq, M_KQUEUE);
1103 	return (0);
1104 }
1105 
1106 static void
1107 kqueue_wakeup(struct kqueue *kq)
1108 {
1109 	if (kq->kq_state & KQ_SLEEP) {
1110 		kq->kq_state &= ~KQ_SLEEP;
1111 		wakeup(kq);
1112 	}
1113 	KNOTE(&kq->kq_kqinfo.ki_note, 0);
1114 }
1115 
1116 /*
1117  * Calls filterops f_attach function, acquiring mplock if filter is not
1118  * marked as FILTEROP_MPSAFE.
1119  */
1120 static int
1121 filter_attach(struct knote *kn)
1122 {
1123 	int ret;
1124 
1125 	if (!(kn->kn_fop->f_flags & FILTEROP_MPSAFE)) {
1126 		get_mplock();
1127 		ret = kn->kn_fop->f_attach(kn);
1128 		rel_mplock();
1129 	} else {
1130 		ret = kn->kn_fop->f_attach(kn);
1131 	}
1132 
1133 	return (ret);
1134 }
1135 
1136 /*
1137  * Detach the knote and drop it, destroying the knote.
1138  *
1139  * Calls filterops f_detach function, acquiring mplock if filter is not
1140  * marked as FILTEROP_MPSAFE.
1141  *
1142  * This can race due to the MP lock and/or locks acquired by f_detach,
1143  * so we interlock with KN_DELETING.  It is also possible to race
1144  * a create for the same reason if userland tries to delete the knote
1145  * before the create is complete.
1146  */
1147 static void
1148 knote_detach_and_drop(struct knote *kn)
1149 {
1150 	if (kn->kn_status & (KN_CREATING | KN_DELETING))
1151 		return;
1152 	kn->kn_status |= KN_DELETING;
1153 
1154 	if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1155 		kn->kn_fop->f_detach(kn);
1156 	} else {
1157 		get_mplock();
1158 		kn->kn_fop->f_detach(kn);
1159 		rel_mplock();
1160 	}
1161 	knote_drop(kn);
1162 }
1163 
1164 /*
1165  * Calls filterops f_event function, acquiring mplock if filter is not
1166  * marked as FILTEROP_MPSAFE.
1167  *
1168  * If the knote is in the middle of being created or deleted we cannot
1169  * safely call the filter op.
1170  */
1171 static int
1172 filter_event(struct knote *kn, long hint)
1173 {
1174 	int ret;
1175 
1176 	if (kn->kn_status & (KN_CREATING | KN_DELETING))
1177 		return(0);
1178 
1179 	if (!(kn->kn_fop->f_flags & FILTEROP_MPSAFE)) {
1180 		get_mplock();
1181 		ret = kn->kn_fop->f_event(kn, hint);
1182 		rel_mplock();
1183 	} else {
1184 		ret = kn->kn_fop->f_event(kn, hint);
1185 	}
1186 
1187 	return (ret);
1188 }
1189 
1190 /*
1191  * walk down a list of knotes, activating them if their event has triggered.
1192  */
1193 void
1194 knote(struct klist *list, long hint)
1195 {
1196 	struct knote *kn;
1197 
1198 	lwkt_gettoken(&kq_token);
1199 	SLIST_FOREACH(kn, list, kn_next) {
1200 		if (filter_event(kn, hint))
1201 			KNOTE_ACTIVATE(kn);
1202 	}
1203 	lwkt_reltoken(&kq_token);
1204 }
1205 
1206 /*
1207  * insert knote at head of klist
1208  *
1209  * Requires: kq_token
1210  */
1211 void
1212 knote_insert(struct klist *klist, struct knote *kn)
1213 {
1214 	lwkt_gettoken(&kq_token);
1215 	SLIST_INSERT_HEAD(klist, kn, kn_next);
1216 	lwkt_reltoken(&kq_token);
1217 }
1218 
1219 /*
1220  * remove knote from a klist
1221  *
1222  * Requires: kq_token
1223  */
1224 void
1225 knote_remove(struct klist *klist, struct knote *kn)
1226 {
1227 	lwkt_gettoken(&kq_token);
1228 	SLIST_REMOVE(klist, kn, knote, kn_next);
1229 	lwkt_reltoken(&kq_token);
1230 }
1231 
1232 /*
1233  * remove all knotes from a specified klist
1234  */
1235 void
1236 knote_empty(struct klist *list)
1237 {
1238 	struct knote *kn;
1239 
1240 	lwkt_gettoken(&kq_token);
1241 	while ((kn = SLIST_FIRST(list)) != NULL)
1242 		knote_detach_and_drop(kn);
1243 	lwkt_reltoken(&kq_token);
1244 }
1245 
1246 /*
1247  * remove all knotes referencing a specified fd
1248  */
1249 void
1250 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd)
1251 {
1252 	struct knote *kn;
1253 
1254 	lwkt_gettoken(&kq_token);
1255 restart:
1256 	SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
1257 		if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) {
1258 			knote_detach_and_drop(kn);
1259 			goto restart;
1260 		}
1261 	}
1262 	lwkt_reltoken(&kq_token);
1263 }
1264 
1265 static void
1266 knote_attach(struct knote *kn)
1267 {
1268 	struct klist *list;
1269 	struct kqueue *kq = kn->kn_kq;
1270 
1271 	if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1272 		KKASSERT(kn->kn_fp);
1273 		list = &kn->kn_fp->f_klist;
1274 	} else {
1275 		if (kq->kq_knhashmask == 0)
1276 			kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1277 						 &kq->kq_knhashmask);
1278 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1279 	}
1280 	SLIST_INSERT_HEAD(list, kn, kn_link);
1281 	TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink);
1282 	kn->kn_status = KN_CREATING;
1283 }
1284 
1285 static void
1286 knote_drop(struct knote *kn)
1287 {
1288 	struct kqueue *kq;
1289 	struct klist *list;
1290 
1291 	kq = kn->kn_kq;
1292 
1293 	if (kn->kn_fop->f_flags & FILTEROP_ISFD)
1294 		list = &kn->kn_fp->f_klist;
1295 	else
1296 		list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1297 
1298 	SLIST_REMOVE(list, kn, knote, kn_link);
1299 	TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink);
1300 	if (kn->kn_status & KN_QUEUED)
1301 		knote_dequeue(kn);
1302 	if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1303 		fdrop(kn->kn_fp);
1304 		kn->kn_fp = NULL;
1305 	}
1306 	knote_free(kn);
1307 }
1308 
1309 static void
1310 knote_enqueue(struct knote *kn)
1311 {
1312 	struct kqueue *kq = kn->kn_kq;
1313 
1314 	KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
1315 
1316 	TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1317 	kn->kn_status |= KN_QUEUED;
1318 	++kq->kq_count;
1319 
1320 	/*
1321 	 * Send SIGIO on request (typically set up as a mailbox signal)
1322 	 */
1323 	if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1)
1324 		pgsigio(kq->kq_sigio, SIGIO, 0);
1325 
1326 	kqueue_wakeup(kq);
1327 }
1328 
1329 static void
1330 knote_dequeue(struct knote *kn)
1331 {
1332 	struct kqueue *kq = kn->kn_kq;
1333 
1334 	KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
1335 	KKASSERT((kn->kn_status & KN_PROCESSING) == 0);
1336 
1337 	TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1338 	kn->kn_status &= ~KN_QUEUED;
1339 	kq->kq_count--;
1340 }
1341 
1342 static void
1343 knote_init(void)
1344 {
1345 	knote_zone = zinit("KNOTE", sizeof(struct knote), 0, 0, 1);
1346 }
1347 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL)
1348 
1349 static struct knote *
1350 knote_alloc(void)
1351 {
1352 	return ((struct knote *)zalloc(knote_zone));
1353 }
1354 
1355 static void
1356 knote_free(struct knote *kn)
1357 {
1358 	zfree(knote_zone, kn);
1359 }
1360